Spectral theory and applications.

An elementary
introductory course. Bucarest Version 2010
Bernard Heler
Universite Paris-Sud,
Departement de Mathematiques,
UMR 8628 du CNRS, Bat. 425,
F-91405 Orsay Cedex, FRANCE
March 26, 2010
Abstract
We intend to present in this course the basic tools in spectral analy-
sis and to illustrate the theory by presenting examples coming from the
Schrodinger operator theory and from various branches of physics : sta-
tistical mechanics, superconductivity, uid mechanics. We also give some
introduction to non self-adjoint operators theory with emphasis on the
role of the pseudo-spectrum. Other examples are treated in the 2008-
2009 version.
Contents
1 Introduction 4
1.1 The free Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 The problem of the boundary . . . . . . . . . . . . . . . . . . . . 6
1.3.1 Ill-posed problems . . . . . . . . . . . . . . . . . . . . . . 6
1.3.2 The periodic problem . . . . . . . . . . . . . . . . . . . . 7
1.3.3 The Dirichlet problem . . . . . . . . . . . . . . . . . . . . 7
1.3.4 The Neumann problem . . . . . . . . . . . . . . . . . . . 7
1.3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Unbounded operators, adjoints, Selfadjoint operators 8
2.1 Unbounded operators . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Adjoints. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Symmetric and selfadjoint operators. . . . . . . . . . . . . . . . . 14
1
3 Representation theorems 16
3.1 Rieszs Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Lax-Milgrams situation. . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 An alternative point of view: \, H, \

. . . . . . . . . . . . . . . . 17
4 Semi-bounded operators and Friedrichs extension. 20
4.1 Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Analysis of the Coulomb case. . . . . . . . . . . . . . . . . . . . . 21
4.3 Friedrichss extension . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5 Compact operators : general properties and examples. 30
5.1 Denition and properties. . . . . . . . . . . . . . . . . . . . . . . 30
5.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 Adjoints and compact operators . . . . . . . . . . . . . . . . . . . 33
5.4 Precompactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 Spectral theory for bounded operators. 36
6.1 Fredholms alternative . . . . . . . . . . . . . . . . . . . . . . . . 36
6.2 Resolvent set for bounded operators . . . . . . . . . . . . . . . . 37
6.3 Spectral theory for compact operators . . . . . . . . . . . . . . . 39
6.4 Spectrum of selfadjoint operators. . . . . . . . . . . . . . . . . . . 41
6.5 Spectral theory for compact selfadjoint operators . . . . . . . . . 42
7 Examples. 44
7.1 The transfer operator. . . . . . . . . . . . . . . . . . . . . . . . . 44
7.1.1 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.1.2 About the physical origin of the problem. . . . . . . . . . 44
7.1.3 Krein-Rutmans Theorem. . . . . . . . . . . . . . . . . . . 48
7.2 The Dirichlet realization, a model of operator with compact re-
solvent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.3 Extension to operators with compact resolvent . . . . . . . . . . 54
7.4 Operators with compact resolvent : the Schrodinger operator in
an unbounded domain. . . . . . . . . . . . . . . . . . . . . . . . . 54
7.5 The Schrodinger operator with magnetic eld . . . . . . . . . . . 56
7.6 Laplace Beltrami operators on a Riemannian compact manifold 58
8 Selfadjoint unbounded operators and spectral theory. 60
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
8.2 Spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8.3 Spectral family and resolution of the identity. . . . . . . . . . . . 64
8.4 The spectral decomposition Theorem. . . . . . . . . . . . . . . . 69
8.5 Applications of the spectral theorem: . . . . . . . . . . . . . . . . 73
8.6 Examples of functions of a selfadjoint operator . . . . . . . . . . 74
8.7 Spectrum and spectral measures . . . . . . . . . . . . . . . . . . 75
2
9 Non-self adjoint operators and c-pseudospectrum 77
9.1 Main denitions and properties . . . . . . . . . . . . . . . . . . . 77
9.2 c-Pseudospectrum : complete analysis of the dierentiation op-
erator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
9.3 Another example of non selfadjoint operator without spectrum . 83
9.4 The non selfadjoint harmonic oscillator . . . . . . . . . . . . . . . 87
9.5 The complex Airy operator in R . . . . . . . . . . . . . . . . . . 88
10 Essentially selfadjoint operators 91
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
10.2 Basic criteria. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
10.3 The Kato -Rellich theorem . . . . . . . . . . . . . . . . . . . . . 94
10.4 Other criteria of selfadjointness for Schrodinger operators . . . . 96
11 Non-selfadjoint case : Maximal accretivity and application to
the Fokker-Planck operator 99
11.1 Accretive operators . . . . . . . . . . . . . . . . . . . . . . . . . . 99
11.2 Application to the Fokker-Planck operator . . . . . . . . . . . . . 99
11.3 Decay of the semi-group and c-pseudospectra . . . . . . . . . . . 101
11.4 Application : The complex Airy operator in R
+
. . . . . . . . . . 102
12 Discrete spectrum, essential spectrum 105
12.1 Discrete spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 105
12.2 Essential spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 105
12.3 Basic examples: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
12.4 On the Schrodinger Operator with constant magnetic eld . . . . 106
12.4.1 Dimension 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 106
12.4.2 The case of R
2
. . . . . . . . . . . . . . . . . . . . . . . . 107
12.4.3 Magnetic Schrodinger operators in dimension 3 . . . . . . 108
12.5 Weyls criterion: . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
13 The max-min principle 112
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
13.2 On positivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
13.3 Variational characterization of the discrete spectrum . . . . . . . 113
13.4 Max-min principle . . . . . . . . . . . . . . . . . . . . . . . . . . 115
13.5 CLR inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
13.6 Essential spectrum and Perssons Theorem . . . . . . . . . . . . 120
14 Exercises and Problems 122
14.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
14.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
3
1 Introduction
Our starting point could be the theory of Hermitian matrices, that is of the
matrices satisfying :

= . If we look for eigenvectors and corresponding

eigenvalues of , that is for pairs (n, ) with n C
k
, n ,= 0 and C such
that n = n, we know that the eigenvalues are real and that one can nd an
orthonormal basis of eigenvectors associated with real eigenvalues.
In order to extend this theory to the case of spaces with innite dimension
(that is replacing the space C
m
by a general Hilbert space H), the rst attempt
consists in developing the theory of compact selfadjoint operators. But it is far
to cover all the interesting cases that are present in Quantum Mechanics. So our
aim is to present a general theory but it is perhaps good to start by looking at
specic operators and to ask naive questions about the existence of pairs (n, )
with n in some suitable domain, n ,= 0 and C such that n = n. We shall
discover in particular that the answer at these questions may depend strongly
on the choice of the domain and on the precise denition of the operator.
1.1 The free Laplacian
The Laplacian has no eigenfunctions in 1
2
, but it has for any R
+
an
eigenfunction in o

(R
m
) (actually in 1

) and for any C an eigenfunction in

T

(R
m
). So what is the right way to extend the theory of Hermitian matrices
on C
k
?
On the other hand, it is easy to produce (take for simplicity : = 1) approximate
eigenfunctions in the form n
n
(r) =
1

n
expir (
xn
2
n
), where is a compactly
supported function of 1
2
-norm equal to 1.
1.2 The harmonic oscillator
As we shall see the harmonic oscillator
H = d
2
,dr
2
+r
2
plays a central role in the theory of quantum mechanics. When looking for
eigenfunctions in o(R), we obtain that there is a sequence of eigenvalues
n
(n N)

n
= (2n 1) .
In particular the fundamental level (in other words the lowest eigenvalue) is

1
= 1
and the splitting between the two rst eigenvalues is 2.
The rst eigenfunction is given by

1
(r) = c
1
exp
r
2
2
(1.2.1)
4
and the other eigenfunctions are obtained by applying the creation operator
1
+
= d,dr +r . (1.2.2)
We observe indeed that
H = 1
+
1

+ 1 , (1.2.3)
where
1

= d,dr +r , (1.2.4)
and has the property
1

1
= 0. (1.2.5)
Note that if n 1
2
is a distributional solution of 1
+
n = 0, then n = 0. Note
also that if n 1
2
is a distributional solution of 1

n = 0, then n = j
1
for
some j R.
The n
th
-eigenfunction is then given by

n
= c
n
(1
+
)
n1

1
.
This can be shown by recursion using the identity
1
+
(H + 2) = H1
+
. (1.2.6)
It is easy to see that
n
(r) = 1
n
(r) exp
x
2
2
where 1
n
(r) is a polynomial of
order n 1. One can also show that the
n
are mutually orthogonal. The
proof of this point is identical to the nite dimensional case, if we observe the
following identity (expressing that H is symmetric) :
< Hn,
L
2=< n, H
L
2 , n o(R), o(R) , (1.2.7)
which is obtained through an integration by parts.
Then it is a standard exercise to show that the family (
n
)
nN
is total and that
we have obtained an orthonormal hilbertian basis of 1
2
, which in some sense
permits to diagonalize the operator H.
Another way to understand the completeness is to show that starting of an eigen-
function n in o

(R) associated with R solution (in the sense of distribution)

of
Hn = n ,
then there exists / N and c
k
,= 0 such that (1

)
k
n = c
k

1
and that the
corresponding is equal to (2/ + 1).
For this proof, we admit that any eigenfunction can be shown to be in o(R) and
use the identity
1

(H 2) = H1

, (1.2.8)
and the inequality
< Hn, n 0 , n o(R) . (1.2.9)
5
This property is called positivity of the operator.
Actually one can show by various ways that
< Hn, n [[n[[
2
, n o(R) . (1.2.10)
One way is to rst establish the Heisenberg Principle
1
:
[[n[[
2
L
2
(R)
2[[rn[[
L
2[[n

[[
L
2 , n o(R) . (1.2.11)
The trick is to observe the identity
1 =
d
dr
r r
d
dr
. (1.2.12)
The inequality (1.2.10) is simply the consequence of the identity
< Hn, n = [[n

[[
2
+[[rn[[
2
, (1.2.13)
which is proved through an integration by parts, and of the application in
(1.2.11) of Cauchy-Schwarz inequality.
Another way is to directly observe the identity
< Hn, n = [[1

n[[
2
+[[n[[
2
, n o . (1.2.14)
1.3 The problem of the boundary
We mainly consider the operator
d
2
dx
2
and look at various problems that can be
asked naively about the existence of eigenfunctions for the problem in 1
2
(]0, 1[).
1.3.1 Ill-posed problems
Look rst at pairs (n, ) H
1
(]0, 1[) C (n ,= 0) such that
dn,dr = n , n(0) = 0 .
It is immediate to see that no such pairs exist. We will come back to this
example later when analyzing non self-adjoint problems.
Look now at pairs (n, ) H
2
(]0, 1[) C (n ,= 0) such that
d
2
n,dr
2
= n .
We can nd for any two linearly independent solutions.
1
Here is a more physical version. If u is normalized by [[u[[
L
2
(R)
= 1, [u[
2
dx denes
a probability measure. One can dene x) =
R
x[u[
2
dx, mean value of the position and
the variance x = (x x))
2
). Similarly, we can consider: Dx) :=
R
(Dxu) u(x)dx and

Dx
:= [[(Dx Dx))u[[
2
. Then (1.2.11) can be extended in the form :
x
Dx

1
4
.
6
1.3.2 The periodic problem
Here we consider pairs (n, ) H
2,per
(]0, 1[) C (n ,= 0) such that
d
2
n,dr
2
= n .
Here
H
2,per
(]0, 1[) = n H
2
(]0, 1[) , n(0) = n(1) and n

(0) = n

(1) .
Here we recall that H
2
(]0, 1[) is included in C
1
([0, 1]) by the Sobolev injection
theorem. It is an easy exercise to show that the pairs are described by two
families

= 4
2
n
2
, n
n
= jcos 2nr , for n N , j R 0 ,

= 4
2
n
2
,
n
= jsin2nr , for n N

, j R 0 .
One observes that = 0 is the lowest eigenvalue and that its multiplicity is one.
This means that the corresponding eigenspace is of dimension one (the other
eigenspaces are of dimension 2). Moreover an eigenfunction in this subspace
never vanishes in ]0, 1[. This is quite evident because n
0
= j ,= 0.
One observes also that one can nd an orthonormal basis in 1
2
(]0, 1[) of eigen-
functions by normalizing the family (cos 2nr (n N), sin2nr (n N

)) or
the family exp2inr (n Z).
We are just recovering the 1
2
-theory of the Fourier series.
1.3.3 The Dirichlet problem
Here we consider pairs (n, ) H
2,D
(]0, 1[) C (n ,= 0) such that d
2
n,dr
2
=
n.
Here
H
2,D
(]0, 1[) = n H
2
(]0, 1[) , n(0) = n(1) = 0 .
It is again an easy exercise to show that the pairs are described by
=
2
n
2
,
n
= jsinnr , for n N

, j R 0 .
One observes that =
2
is the lowest eigenvalue, that its multiplicity is one
(Here all the eigenspaces are one-dimensional) and that an eigenfunction in this
subspace neither vanishes in ]0, 1[.
1.3.4 The Neumann problem
Here we consider pairs
(n, ) H
2,N
(]0, 1[) C (n ,= 0) such that
d
2
n,dr
2
= n .
7
Here
H
2,N
(]0, 1[) = n H
2
(]0, 1[) , n

(0) = n

(1) = 0 .
It is again an easy exercise to show that the pairs are described by
=
2
n
2
,
n
= jcos nr , for n N, j R 0 .
One observes that = 0 is the lowest eigenvalue, that its multiplicity is one
(Here all the eigenspaces are one-dimensional) and that the corresponding eigenspace
is of dimension one and that an eigenfunction in this subspace neither vanishes
in ]0, 1[.
1.3.5 Conclusion
All these examples enter in the so called Sturm-Liouville theory. We have em-
phasized on one property which was always veried in each case: the eigenspace
corresponding to the lowest eigenvalue is one dimensional and one can nd a
strictly positive (in ]0, 1[ or in ] , +[ in the case of the harmonic oscillator)
corresponding eigenfunction. We suggest to the reader to come back at
this introduction after have read the course. He will surely realize
that the theory has permitted to clarify many sometimes badly posed
problems.
2 Unbounded operators, adjoints, Selfadjoint op-
erators
2.1 Unbounded operators
We consider an Hilbert space H. The scalar product will be denoted by :
n, )
H
or more simply by : n, ) when no confusion is possible. We take
the convention that the scalar product is antilinear with respect to the second
argument.
A linear operator (or more simply an operator) T in H is a linear map n Tn
dened on a subspace H
0
of H, denoted by 1(T) and which is called the domain
of T. We shall also denote by 1(T) (or Im T or 1onpc(T)) the range of H
0
by T. We shall say that T is bounded if it is continuous from 1(T) (with the
topology induced by the topology of H) into H. When 1(T) = H, we recover
the notion of linear continuous operators on H. We recall that with
[[T[[
L(H)
= sup
u=0
[[Tn[[
H
[[n[[
H
, (2.1.1)
L(H) is a Banach space. When 1(T) is not equal to H, we shall always assume
that
1(T) is dense in H. (2.1.2)
Note that, if T is bounded, then it admits a unique continuous extension to H.
In this case the generalized notion is not interesting.
8
We are mainly interested in extensions of this theory and would like to consider
unbounded operators.
When using this word, we mean more precisely non necessarily bounded op-
erators.
The point is to nd a natural notion replacing this notion of boundedness. This
is the object of the next denition.
The operator is called closed if the graph G(T) of T is closed in HH. We
recall that
G(T) := (r, j) HH , r 1(T) , j = Tr . (2.1.3)
Equivalently, we can say
Denition 2.1.1. (Closed operators).
Let T be an operator on H with (dense) domain 1(T). We say that T is closed
if the conditions
n
n
1(T),
n
n
n in H,
Tn
n
in H
imply
n 1(T),
= Tn.
Example 2.1.2.
1. T
0
= with 1(T
0
) = C

0
(R
m
) is not closed.
For this, it is enough to consider
2
some n in H
2
(R
m
) and not in C

0
(R
n
)
and to consider a sequence n
n
C

0
such that n
n
n in H
2
. The
sequence (n
n
, n
n
) is contained in G(T
0
) and converges in 1
2
1
2
to
(n, n) which does not belong to G(T
0
).
2. T
1
= with 1(T
1
) = H
2
(R
m
) is closed.
We observe indeed that if n
n
n in 1
2
and (n
n
) in 1
2
then
n = 1
2
. The last step is to observe that this implies that n
H
2
(R
m
) (take the Fourier transform) and (n, n) G(T
1
).
2
We recall that the Sobolev space H
s
(R
m
) is dened as the space
H
s
(R
m
) := u S

(R
m
) [ (1 +[[
2
)
s
2 u L
2
(R
m
) .
Here S

is the set of tempered distributions. H

s
(R
m
) is equipped with the natural Hilbertian
norm :
[[u[[
2
H
s :=
Z
R
m
(1 +[[
2
)
s
[ u()[
2
d .
By Hilbertian norm, we mean that the norm is associated to a scalar product.
When s N, we can also describe H
s
by
H
s
(R
m
) := u L
2
(R
m
) [ D

x
u L
2
, s. t. [[ s .
The natural norm associated with the second denition is equivalent to the rst one.
9
This example suggests another denition.
Denition 2.1.3.
The operator T is called closable if the closure of the graph of T is a graph.
We can then dene the closure T of the operator by a limit procedure via
its graph. We observe indeed that we can consider
1(T) := r H [ j s. t. (r, j) G(T).
For any r 1(T), the assumption that G(T) is a graph says that j is unique.
One can consequently dene T by
Tr = j .
In a more explicit way, the domain of T is the set of the r H such that
r
n
r H and Tr
n
is a Cauchy sequence, and for such r we dene Tr by
Tr = lim
n+
Tr
n
.
Example 2.1.4.
T
0
= with 1(T
0
) = C

0
is closable and is closure is T
1
.
Let us prove it, as an exercise. Let T
0
the closure of T
0
. Let n 1
2
such that
there exists n
n
C

0
such that n
n
n in 1
2
and n
n
in 1
2
. We get
by distribution theory that n 1
2
satises n = 1
2
. By the ellipticity
of the Laplacian (use the Fourier transform), we get that n H
2
. We have
consequently shown that 1(T
0
) H
2
. But C

0
is dense in H
2
and this gives
the inverse inclusion : H
2
1(T
0
). We have consequently,
H
2
= 1(T
1
) = 1(T
0
) ,
and it is then easy to verify that T
1
= T
0
.
These examples lead to a more general question.
Realization of dierential operators as unbounded operators. Let
R
n
and let 1(r, 1
x
) be a partial dierential operator with C

coecients
in . Then the operator 1

dened by
1(1

) = C

0
() , 1

n = 1(r, 1
x
)n, n C

0
() ,
is closable. Here H = 1
2
(). We have indeed
G(1

)

G

:= (n, )) HH [ 1(r, 1
x
)n = ) in T

() .
10
The proof is then actually a simple exercise in distribution theory. This inclusion
shows that G(1

) is a graph. Note that the corresponding operator is dened

as 1

min
with domain
1(1

min
) =
_
n 1
2
() [ a sequence n
n
C

0
() s.t
_
n
n
n in 1
2
()
1(r, 1
x
)n
n
converges in 1
2
()
_
.
The operator 1

min
is then dened for such n by
1

min
n = lim
n+
1(r, 1
x
)n
n
.
Using the theory of distributions, this gives :
1

min
n = 1(r, 1
x
)n .
Note that there exists also a natural closed operator whose graph is

G

and
extending 1

: this is the operator

1

, with domain

:= n 1
2
() , 1(r, 1
x
)n 1
2
() ,
and such that

n = 1(r, 1
x
)n , n

1

,
where the last equality is in the distributional sense. Note that

1

is an exten-
sion of 1

min
in the sense that :

n = 1

min
n , n 1(1

min
) .
Conclusion.
We have associated with a dierential operator 1(r, 1
x
) in an open set three
natural operators. It would be important to know better the connection between
these three operators.
Remark 2.1.5. (Link between continuity and closeness).
If H
0
= H, the closed graph Theorem says that a closed operator T is continuous.
2.2 Adjoints.
When we have an operator T in L(H), it is easy to dene the Hilbertian adjoint
T

by the identity :
T

n, )
H
= n, T)
H
, n H, H . (2.2.1)
The map n, T)
H
denes a continuous antilinear map on H and can be
expressed, using Rieszs Theorem, by the scalar product by an element which is
called T

n. The linearity and the continuity of T

is then easily proved using

(2.2.1).
Let us now give the denition of the adjoint of an unbounded operator.
11
Denition 2.2.1. (Adjoint )
If T is an unbounded operator on H whose domain 1(T) is dense in H, we rst
dene the domain of T

by
1(T

) = n H, 1(T) n, T),
can be extended as an antilinear continuous form on H .
Using the Riesz Theorem, there exists ) H such that
< ) , =< n, T) , n 1(T

), 1(T) .
The uniqueness of ) is a consequence of the density of 1(T) in H and we can
then dene T

n by
T

n = ) .
Remark 2.2.2.
When 1(T) = H and if T is bounded , then we recover as T

the Hilbertian
adjoint.
Example 2.2.3.
T

0
= T
1
.
Let us treat in detail this example. We get
1(T

0
) = n 1
2
[ the map C

0
n, ),
can be extended as an antilinear continuous form on 1
2
.
We observe that
n, )
L
2 =
_
R
m
n()dr = (n)( ) .
The last equality just means that we are considering the distribution (n) on
the test function . The condition appearing in the denition is just that this
distribution is in 1
2
(R
m
). Coming back to the denition of 1(T

0
), we get
1(T

0
) = n 1
2
[ n 1
2
.
But as already seen, this gives
1(T

0
) = H
2
, T

0
n = n , n H
2
.
Proposition 2.2.4. .
T

is a closed operator.
Proof.
Let (
n
) be a sequence in 1(T

) such that
n
in H and T

n
n

in H
for some pair (, n

). We would like to show that (, n

) belongs to the graph

of T

.
For all n 1(T), we have :
Tn, ) = lim
n+
Tn,
n
) = lim
n+
n, T

n
) = n, n

) . (2.2.2)
12
Coming back to the denition of 1(T

), we get from (2.2.2) that 1(T

)
and T

= n

. This means that (, n

) belongs to the graph of T

.
Proposition 2.2.5. .
Let T be an operator in H with domain 1(T). Then the graph G(T

) of T

can
be characterized by
G(T

) = \ ( G(T) )

, (2.2.3)
where \ is the unitary operator dened on HH by
\ n, = , n . (2.2.4)
Proof.
We just observe that for any n 1(T) and (, n

) HH we have the identity

\ (n, Tn) , (, n

))
HH
= Tn, )
H
n, n

)
H
.
The right hand side vanishes for all n 1(T) i 1(T

) and n

= T

, that
is if (, n

) belongs to G(T

). The left hand side vanishes for all n 1(T) i

(, n

) belongs to \ (G(T))

.
Standard Hilbertian analysis, using the continuity of \ and \
1
= \ , then
shows that
\ (G(T))

= \ (G(T))

= \ (G(T))

.
End of proof We have not analyzed till now under which condition the do-
main of the adjoint is dense in H. This is one of the objects of the next theorem.
Theorem 2.2.6.
Let T be a closable operator. Then we have
1. 1(T

) is dense in H,
2. T

:= (T

= T, where we have denoted by T the operator whose graph

is G(T).
Proof.
For the rst point, let us assume that 1(T

) is not dense in H. Then there

exists n ,= 0 such that n is orthogonal to 1(T

).
We consequently get that for any 1(T

), we have
(0, n) , (T

, ))
HH
= 0 .
This shows that (0, n) is orthogonal to \ (G(T

)).
But the previous proposition gives :
\ (G(T)) = G(T

.
We now apply \ to this identity and get, using \
2
= 1,
\
_
G(T

_
= G(T) .
13
But, for any closed subspace / HH, we have
\ (/

) = [\ (/)]

,
as a consequence of the identity
\ (n, ) , (r, j))
HH
= (n, ) , \ (r, j))
HH
.
We nally obtain that (0, n) belongs to the closure of the graph of T, that is
the graph of T because T is closable, and consequently that n = 0. This gives
the contradiction.
For the second point, we rst observe that, 1(T

) being dense in H, we can of

course dene (T

. Using again the proposition and the closeness of T

, we
obtain G(T

) = G(T) and T

= T.
This means more explicitly that
1(T

) = 1(T) , T

n = Tn , n 1(T) .
End of the proof.
2.3 Symmetric and selfadjoint operators.
Denition 2.3.1. (symmetric operators).
We shall say that T : H
0
H is symmetric if it satises
Tn, )
H
= n, T)
H
, n, H
0
.
Example 2.3.2. .
T = with 1(T) = C

0
(R
m
).
If T is symmetric it is easy to see that
1(T) 1(T

) (2.3.1)
and that
Tn = T

n , n 1(T) . (2.3.2)
The two conditions (2.3.1) and (2.3.2) express the property that (T

, 1(T

))
is an extension of (T, 1(T)).
Exercise 2.3.3.
Show that a symmetric operator is closable.
Hint :
Show that, if n
n
is a sequence in 1(T) such that, for some / H, we have
n
n
0 and Tn
n
/, then / = 0.
For a symmetric operator, we have consequently two natural closed exten-
sions:
14
The minimal one denoted by T
min
(or previously T), which is obtained
by taking the operator whose graph is the closure of the graph of T,
The maximal one denoted by T
max
the adjoint of T.
If T
sa
is a selfadjoint extension of T, then T
sa
is automatically an extension of
T
min
and admits
3
as an extension T
max
.
Denition 2.3.4.
We shall say that T is selfadjoint if T

= T, i. e.
1(T) = 1(T

) , and Tn = T

n , n 1(T) .
Starting of a symmetric operator, it is a natural question to ask for the ex-
istence and the uniqueness of a selfadjoint extension. We shall see later that a
natural way is to prove the equality between T
min
and T
max
.
Exercise 2.3.5. Analysis of dierential operators.
Give simple criteria in the case of operators with constant coecients for ob-
taining symmetric operators on C

0
(R
m
). In particular, verify that the operator
1
xj
=
1
i

xj
is symmetric.
Proposition 2.3.6.
A selfadjoint operator is closed.
This is immediate because T

is closed.
Proposition 2.3.7.
Let T be a selfadjoint operator which is invertible. Then T
1
is also selfadjoint.
By invertible, we mean here that T admits an inverse T
1
from 1(T) into
1(T). Let us rst show that 1(T) is dense in H. Let n H such that
< Tn, n
H
= 0, n 1(T).
Coming back to the denition of T

, this implies in particular that n 1(T

)
and T

n = 0. But T is selfadjoint and injective and this implies that n = 0.

We consequently know that 1(T
1
) is dense in H.
Coming back to the analysis of the corresponding graphs it is now easy to show
the second assertion by coming back to the corresponding graphs and by using
Proposition 2.2.5.
Remark 2.3.8.
If T is selfadjoint T +1 is selfadjoint for any real .
3
Use Proposition 2.2.5.
15
3 Representation theorems
We assume that the reader knows about this material but recall it for complete-
ness
4
.
3.1 Rieszs Theorem.
Theorem 3.1.1. (Rieszs Theorem)
Let n 1(n) a linear continuous form on H. Then there exists a unique n H
such that
1(n) = n, n)
H
, n H . (3.1.1)
There is a similar version with antilinear maps :
Theorem 3.1.2.
Let n 1(n) a antilinear continuous form on H. Then there exists a unique
n H such that
1(n) =< n, n
H
, n H . (3.1.2)
3.2 Lax-Milgrams situation.
Let us now consider a continuous sesquilinear form o dened on \ \ :
(n, ) o(n, ) .
We recall that, because of the sesquilinearity, the continuity can be expressed
by the existence of C such that
[o(n, )[ C [[n[[
V
[[[[
V
, n, \ . (3.2.1)
It is immediate to associate, using the Riesz Theorem, a linear map L(\ )
such that
o(n, ) =< n,
V
. (3.2.2)
Denition 3.2.1. (\ -ellipticity)
We shall say that o is \ -elliptic, if there exists 0, such that
[o(n, n)[ [[n[[
2
V
, n \ . (3.2.3)
Theorem 3.2.2. (Lax-Milgrams Theorem)
Let o be a continuous sesquilinear form on \ \ . If o is \ -elliptic, then is
an isomorphism from \ onto \ .
The proof is in three steps.
Step 1 : is injective.
We get indeed from (3.2.3)
[n, n)
V
[ [[n[[
2
V
, n \ . (3.2.4)
4
Here we follow, almost verbatim, the book of D. Huet [Hu].
16
Using Cauchy-Schwarz in the left hand side, we rst get
[[n[[
V
[[n[[
V
[[n[[
2
V
, n \ ,
and consequently
[[n[[
V
[[n[[
V
, n \ . (3.2.5)
This gives clearly the injectivity but actually more.
Step 2 : (\ ) is dense in \ .
Let us consider n \ such that < , n
V
= O , \ . In particular, we
can take = n. This gives o(n, n) = 0 and n = 0 using (3.2.3).
Step 3 :1() := (\ ) is closed in \ .
Let
n
a Cauchy sequence in (\ ) and n
n
the sequence such that n
n
=
n
.
But using (3.2.5), we get that n
n
is a Cauchy sequence which is consequently
convergent to some n \ . But the sequence n
n
tends to n by continuity
and this shows that
n
= n and 1().
Step 4 :
1
is continuous.
The three previous steps show that is bijective. The continuity of
1
is a
consequence of (3.2.5) or of the Banach Theorem.
Remark 3.2.3. .
Let us suppose for simplicity that \ is a real Hilbert space. Using the isomor-
phism 1 between \ and \

given by the Riesz Theorem, one gets also a natural

operator / from \ onto \

such that
o(n, ) = (/n)() , \ . (3.2.6)
We have
/ = 1 .
3.3 An alternative point of view: V, H, V
t
.
We now consider two Hilbert spaces \ and H such that
\ H . (3.3.1)
By this notation of inclusion, we mean also that the injection of \ into H is
continuous or equivalently that there exists a constant C 0 such that, for any
n \ , we have
[[n[[
H
C [[n[[
V
.
We also assume that
\ is dense in H. (3.3.2)
In this case, there exists a natural injection from H into the space \

which is
dened as the space of continuous linear forms on \ . We observe indeed that
17
if / H then \ n < n, /
H
is continuous on \ . So there exists /
h
\

such that
/
h
(n) =< n, /
H
, n \ .
The injectivity is a consequence of the density of \ in H.
We can also associate to the sesquilinear form o an unbounded operator o on
H in the following way.
We rst dene 1(o) by
1(o) = n \ [ o(n, ) is continuous on \ for the topology induced by H.
(3.3.3)
Using again the Riesz Theorem and assumption (3.3.2), this denes on in H by
o(n, ) =< on,
H
, \ . (3.3.4)
Theorem 3.2.2 is completed by
Theorem 3.3.1. .
Under the same assumptions, o is bijective from 1(o) onto H and o
1
L(H).
Moreover 1(o) is dense in H.
Proof.
We rst show that o is injective. This is a consequence of
[[n[[
2
H
C [[n[[
2
V
C[o(n, n)[ = C [ < on, n
H
[ C [[on[[
H
[[n[[
H
, n 1(o) ,
which leads to
[[n[[
H
C [[on[[
H
, n 1(o) . (3.3.5)
We get directly the surjectivity in the following way. If / H and if n \ is
chosen such that
/, )
H
= n, )
V
, \ ,
(which follows from Rieszs Theorem), we can take n =
1
n in \ , which is a
solution of
o(n, ) =< n,
V
.
We then show that n 1(o), using the identity
o(n, ) = /, )
H
, \ ,
and get simultaneously
on = / .
The continuity of o
1
is a consequence of (3.3.5).
Let us show the last statement of the theorem, i.e. the density of 1(o) in H.
Let / H s. t.
n, /)
H
= 0 , n 1(o) .
By the surjectivity of o, there exists 1(o) s. t. :
o = / .
18
We get
< o, n
H
= 0 , n 1(o)
Taking n = and using the \ c||ijticitj, we get that = 0 and consequently
/ = 0.
The hermitian case.
We now consider an hermitian sesquilinear form, that is satisfying
o(n, ) = o(, n) , n, \ . (3.3.6)
This property is transmitted to o in the following way
Theorem 3.3.2.
If o is hermitian and \ -elliptic, we have
1. o is closed;
2. o = o

;
3. 1(o) is dense in \ .
Proof of 2.
We rst observe that the assumption of Hermiticity gives
< on,
H
=< n, o
H
, n 1(o) , 1(o) . (3.3.7)
In other words o is symmetric. This means in particular that
1(o) 1(o

) . (3.3.8)
Let 1(o

). Using the surjectivity of o, there exists

0
1(o) such that
o
0
= o

.
For all n 1(o), we get that
< on,
0

H
=< n, o
0

H
=< n, o

H
=< on,
H
.
Using again the surjectivity of o, we get =
0
1(o). This shows that
1(o) = 1(o

) and o = o

, 1(o).
Proof of 1.
o is closed because o

is closed and o = o

.
Proof of 3.
Let / \ such that
< n, /
V
= 0 , n 1(o) .
Let ) \ such that ) = / ( is an isomorphism from \ onto \ ).
We then have
0 =< n, /
V
=< n, )
V
= < ), n
V
= o(), n) = o(n, )) =< on, )
H
.
Using the surjectivity, we get ) = 0 and consequently / = 0.
19
Remark 3.3.3.
Theorem 3.3.2 gives us a rather easy way to construct selfadjoint operators.
This will be combined with some completion argument in the next Section to get
the Friedrichs extension.
4 Semi-bounded operators and Friedrichs exten-
sion.
4.1 Denition
Denition 4.1.1.
Let T
0
be a symmetric unbounded operator of domain 1(T
0
). We say that T
0
is
semibounded (from below) if there exists a constant C such that
< T
0
n, n
H
C [[n[[
2
H
, n 1(T
0
) . (4.1.1)
Example 4.1.2. (The Schrodinger operator).
We consider on R
m
the operator
1
V
(r, 1
x
) := +\ (r) , (4.1.2)
where \ (r) is a continuous function on R
m
(called the potential) such that there
exists C s.t. :
\ (r) C , r R
m
. (4.1.3)
Then the operator T
0
dened by
1(T
0
) = C

0
(R
m
) and T
0
n = 1
V
(r, 1
x
)n , n 1(T
0
) ,
is a symmetric, semibounded operator.
We have indeed, with H = 1
2
(R
m
),
< 1(r, 1
x
)n, n
H
=
_
R
m
(n +\ n) n dr
=
_
R
m
[n(r)[
2
dr +
_
R
m
\ (r)[n(r)[
2
dr
C [[n[[
2
H
.
(4.1.4)
Exercise 4.1.3.
Let us consider on 1
2
(R
2
; C
2
) the operator

2
j=1

j
1
xj
with domain o(R
2
; C
2
).
Here the
j
are 2 2 Hermitian matrices such that :

j

k
+
k

j
= 2
jk
.
Show that this operator symmetric but not semi-bounded. This operator is
called the Dirac operator. Its domain is H
1
(R
2
, C
2
) and its square is the
Laplacian :
(
2

j=1

j
1
xj
)
2
= () 1
C
2 .
20
4.2 Analysis of the Coulomb case.
There are two important inequalities which are useful when considering the
Coulomb case which plays an important role in atomic Physics. By Coulomb
case, we mean the analysis on R
3
of the operator
o
Z
:=
7
:
,
or of the Klein-Gordon operator
1
Z
:=

+ 1
7
:
.
The operator

+ 1 can easily dened on o(R
3
), using the Fourier trans-
form T, by
T(

+ 1 n)(j) =
_
j
2
+ 1(Tn)(j) .
The rst one is the Hardy Inequality (which can be found for example in the
book of Kato ([Ka], p. 305-307)) :
_
R
3
[r[
2
[n(r)[
2
dr 4
_
R
3
[j[
2
[ n(j)[
2
dj (4.2.1)
and the second one is due to Kato and says that
_
R
3
[r[
1
[n(r)[
2
dr

2
_
R
3
[j[ [ n(j)[
2
dj . (4.2.2)
One proof of the Hardy inequality consists in writing that, for any R and
any n C

0
(R
3
; R), we have :
_
R
3
[n +
r
[r[
2
n[
2
dr 0 . (4.2.3)
This leads to :
_
R
3
[n[
2
+
2
1
[r[
2
[n[
2
dr 2
_
R
3
n
r
[r[
2
n dr .
But an integration by part gives :
2
_
R
3
n
x
|x|
2
ndr =
_
R
3
[n(r)[
2
(div (
x
|x|
2
)) dr
=
_
R
3
[n(r)[
2
(
1
|x|
2
) dr .
Optimizing over leads to =
1
2
and gives then the result.
Remark 4.2.1. .
The same idea works for 3 but fails for = 2. So a good exercise
5
is to
look for substitutes in this case by starting from the inequality :
_
R
3
[n (r)
r
[r[
2
n[
2
dr 0 . (4.2.4)
5
We thank M.J. Esteban for explaining to us the trick.
21
The function (r) can be assumed radial : (r) = p([r[) and the question is to
nd a dierential inequality on p leading at a weaker Hardys type inequality.
One can for example try p(:) = ln(:).
For the proof of Katos inequality, there is a another tricky nice estimate
which, as far as we know, goes back to Hardy and Littlewood. In the case of
the Coulomb potential, we can write, using the explicit computation for the
Fourier-transform of r
1
|x|
:
_
R
3
_
R
3
n(j)
1
|pp

|
2
n(j

)dj dj

=
_
R
3
R
3
n(j)
h(p)
h(p

)

h(p

)
h(p)
1
|pp

|
2
n(j

)dj dj

=
_
R
3
R
3
1
|pp

|
n(j)
h(p)
h(p

)

h(p

)
h(p)
1
|pp

|
n(j

)dj dj

,
where / is a strictly positive measurable function to be determined later.
We then use Cauchy-Schwarz in the last equality in order to get
[
_ _
n(j)
1
|pp

|
2
n(j

)dj dj

_
_
[ n(j)[
2
[
h(p)
h(p

)
[
2 1
|pp

|
2
dj

dj
_1
2

_
_
[ n(j

)[
2
[
h(p

)
h(p)
[
2 1
|pp

|
2
dj

dj
_1
2
=
_
[ n(j)[
2
(
_
[
h(p)
h(p

)
[
2 1
|pp

|
2
dj

) dj
=
_
/(j)
2
[ n(j)[
2
(
_
[
1
h(p

)
[
2 1
|pp

|
2
dj

) dj .
We now write j

[j[ in the integral

_
[
1
h(p

)
[
2 1
|pp

|
2
dj

. We then take
/(j) = [j[ .
The integral becomes
_
[
1
/(j

)
[
2
1
[j j

[
2
dj

= [j[
1
_
1
[

[
2
[

[
2
d

,
with j = [j[.
This is clearly a convergent integral. Moreover, observing the invariance by
rotation, one can show that the integral is independent of S
2
. Hence we
can compute it with = (1, 0, 0).
We nally obtain the existence of an explicit constant C such that
[
_ _
n(j)
1
[j j

[
2
n(j

) dj.dj

[ C
_
R
3
[j[[[ n(j)[
2
dj .
The optimization of the trick leads to C =

2
.
Let us now show how one can use these inequalities.
If we use (4.2.1), we get the semi-boundedness for any 7 0 for the
Schrodinger Coulomb operator (using Cauchy-Schwarz Inequality).
_
R
3
1
:
[n[
2
dr (
_
1
:
2
[n[
2
dr)
1
2
[[n[[ .
22
But we can rewrite the Hardy Inequality in the form
_
R
3
1
:
2
[n[
2
dr 4 < n, n
L
2
(R
3
)
.
So we get, for any c 0,
_
R
3
1
:
[n[
2
dr c < n, n
L
2 +
1
c
[[n[[
2
. (4.2.5)
This leads to :
< o
Z
n, n
L
2 (1 c7) < n, n
7
c
[[n[[
2
.
Taking c =
1
Z
, we have nally shown that
< o
Z
n, n
L
2 7
2
[[n[[
2
. (4.2.6)
Here we are not optimal
6
but there is another way to see that the behavior with respect to 7 is correct.
We just observe some invariance of the model. Let us suppose that we have
proved the inequality for 7 = 1. In order to treat the general case, we make a
change of variable r = j. The operator o
Z
becomes in the new coordinates :

o
Z
=
2
(
y
)
7
j
.
Taking = 7
1
, we obtain

o
Z
= 7
2
_

y

1
j
_
.
The other inequality (4.2.2) is with this respect much better and quite im-
portant for the analysis of the relativistic case. Let us see what we obtain in
the case of Klein-Gordon using Katos Inequality.
We have
< 1
Z
n , n
L
2 (1 7

2
) <

+ 1 n , n
L
2 .
Here the nature of the result is dierent. The proof gives only that 1
Z
is
semibounded if 7
2

. This is actually more than a technical problem !

6
It can be proven (see any standard book in quantum mechanics) that the negative spec-
trum of this operator is discrete and is described by a sequence of eigenvalues tending to 0 :

Z
2
4n
2
with n N

. An eigenfunction related to the lowest eigenvalue

1
4
(Z = 1) is given by
x exp
1
2
[x[. To prove that, one can instead of using Hardy or Kato use the fact that
[[u
x
[x[
u[[
2
0 ,
and then optimize over .
23
4.3 Friedrichss extension
Theorem 4.3.1.
A symmetric semibounded operator T
0
on H (with 1(T
0
) dense in H) admits a
selfadjoint extension.
The extension constructed in the proof is the so-called Friedrichs extension.
The proof can be seen as a variant of Lax-Milgrams Lemma. We can assume
indeed by possibly replacing T
0
by T
0
+
0
1d that T
0
satises
< T
0
n, n
H
[[n[[
2
H
, n 1(T
0
) . (4.3.1)
Let us consider the associated form a priori dened on 1(T
0
) 1(T
0
) :
(n, ) o
0
(n, ) := T
0
n , )
H
. (4.3.2)
The inequality (4.3.1) says that
o
0
(n, n) [[n[[
2
H
, n 1(T
0
) . (4.3.3)
We introduce \ as the completion in H of 1(T
0
) for the norm
n j
0
(n) =
_
o
0
(n, n) .
More concretely n H belongs to \ , if there exists n
n
1(T
0
) such that
n
n
n in H and n
n
is a Cauchy sequence for the norm j
0
.
As a natural norm for \ , we get as a candidate :
[[n[[
V
= lim
n+
j
0
(n
n
) , (4.3.4)
where n
n
is a Cauchy sequence for j
0
tending to n in H.
Let us show that the denition does not depend on the Cauchy sequence. This
is the object of the
Lemma 4.3.2.
Let r
n
a Cauchy sequence in 1(T
0
) for j
0
such that r
n
0 in H. Then
j
0
(r
n
) 0.
Proof of the lemma.
First observe that j
0
(r
n
) is a Cauchy sequence in R
+
and consequently conver-
gent in R
+
.
Let us suppose by contradiction that
j
0
(r
n
) 0 . (4.3.5)
We rst observe that
o
0
(r
n
, r
m
) = o
0
(r
n
, r
n
) +o
0
(r
n
, r
m
r
n
) , (4.3.6)
and that a Cauchy-Schwarz inequality is satised :
[o
0
(r
n
, r
m
r
n
)[
_
o
0
(r
n
, r
n
)
_
o
0
(r
m
r
n
, r
m
r
n
) . (4.3.7)
24
Using also that r
n
is a Cauchy sequence for j
0
and (4.3.6), we obtain that
c 0, s. t. n , : , [o
0
(r
n
, r
m
)
2
[ c . (4.3.8)
We take c =

2
2
and consider the corresponding given by (4.3.8). Coming
back to the denition of o
0
we obtain,
[o
0
(r
n
, r
m
)[ = [ < T
0
r
n
, r
m
[
1
2

2
, n , : . (4.3.9)
But as : +, the left hand side in (4.3.9) tends to 0 because by assumption
r
m
0 and this gives a contradiction.
##.
We now observe that
[[n[[
V
[[n[[
H
, (4.3.10)
as a consequence of (4.3.3) and (4.3.4).
This means that the injection of \ in H is continuous. Note also that \ , which
contains 1(T
0
), is dense in H, by density of 1(T
0
) in H. Moreover, we get a
natural scalar product on \ by extension of o
0
:
< n,
V
:= lim
n+
o
0
(n
n
,
n
) , (4.3.11)
where n
n
and
n
are Cauchy sequences for j
0
tending respectively to n and
in H.
By the second version of the Lax-Milgram Theorem (in the context \, H, \

)
applied with
o(n, ) := < n,
V
,
we get an unbounded selfadjoint operator o on H extending T
0
whose domain
1(o) satises 1(o) \ .
Remark 4.3.3. (Friedrichs extension starting from a sesquilinear form)
One can also start directly from a semi-bounded sesquilinear form o
0
dened on
a dense subspace of H.
As we shall see below, this is actually the right way to proceed for the Neumann
realization of the Laplacian, where we consider on C

() the sequilinear form

(n, ) n, ) .
4.4 Applications
The reader should have some minimal knowledge of Sobolev Spaces and traces
of distributions for reading this subsection (See for example Brezis [Br]).
25
Application 1: The Dirichlet realization.
Let be an open set in R
m
such that is compact and let T
0
be the unbounded
operator dened by
1(T
0
) = C

0
() , T
0
= .
The involved Hilbert space is H = 1
2
(). It is clear that T
0
is symmetric and
positive
7
(hence semi-bounded). Following the previous general construction,
we prefer to consider :

T
0
:= T
0
+1d.
It is easy to see
8
that \ is the closure in H
1
() of C

0
(). This means, at least
if is regular, the space H
1
0
(). The domain of o is then described as
1(o) := n H
1
0
() [ n 1
2
().
o is then the operator ( + 1) acting in the sense of distributions.
When is regular, a standard regularity theorem (see [Lio2], [LiMa]) permits
to show that
1(o) = H
2
() H
1
0
() . (4.4.1)
So we have shown the following theorem
Theorem 4.4.1.
The operator T
1
dened by
1(T
1
) = H
2
() H
1
0
() , T
1
= ,
is selfadjoint and called the Dirichlet realization of in .
We have just to observe that T
1
= o 1 and to use Remark 2.3.8.
Note that T
1
is a selfadjoint extension of T
0
.
Note that by the technique developed in Subsection ??, we have also constructed
another selfadjoint extension of T
0
. So we have constructed, when is relatively
compact, two dierent selfadjoint realizations of T
0
. We say in this case that T
0
is not essentially selfadjoint.
Application 2: The harmonic oscillator.
We can start from
H
0
= +[r[
2
+ 1 ,
with domain
1(H
0
) = C

0
(R
m
) .
7
We shall in fact see later that it is strictly positive.
8
We recall that there are two ways for describing H
1
0
(). In the rst denition we just
take the closure of C

0
() in H
1
().
In the second denition, we describe 1
1
0
() as the subspace in H
1
() of the distributions
whose trace is zero at the boundary. This supposes that the boundary = is regular. In
this case, there exists a unique application
0
continuous from H
1
() onto H
1
2 () extending
the map C

) u u

. It is a standard result (cf Brezis [Br] or Lions-Magenes [LiMa])

that 1
1
0
() = H
1
0
(), when the boundary is regular.
26
Following the scheme of the construction of the Friedrichs extension, we rst
get that
\ = 1
1
(R
m
) := n H
1
(R
m
) [ r
j
n 1
2
(R
m
) , , [1, , :] .
One can indeed rst show that \ 1
1
(R
m
) and then get the equality by proving
the property that C

0
(R
m
) is dense in 1
1
(R
m
). One can then determine the
domain of o as
1(o) = n 1
1
(R
m
) [ ( +[r[
2
+ 1)n 1
2
(R
m
) .
By a regularity theorem (dierential quotients method [LiMa]), one can show
that
1(o) = 1
2
(R
m
) := n H
2
(R
m
) [ r

n 1
2
(R
m
) , s. t. [[ 2.
Application 3 : Schrodinger operator with Coulomb potential
We consequently start from
1(T
0
) = C

0
(R
3
) , T
0
=
1
:
. (4.4.2)
We have seen that T
0
is semibounded and replacing T
0
by T
0
+2, the assumptions
of the proof of Friedrichs extension theorem are satised. We now claim that
\ = H
1
(R
3
) .
Having in mind that C

0
(R
3
) is dense in H
1
(R
3
), we have just to verify that
the norm j
0
and the norm [[ [[
H
1
(R
3
)
are equivalent on C

0
(R
3
). This results
immediately of (4.2.5).
With a little more eort, one gets that 1(o) = H
2
(R
3
).
Application 4 : Neumann problem.
Let be a bounded domain with regular boundary in R
m
. Take H = 1
2
().
Let us consider the sesquilinear form :
o
0
(n, ) =
_

n , )
C
mdr +
_

n dr ,
on C

().
Using Remark 4.3.3 and the density of C

() in H
1
(), we can extend the
sesquilinear form to \ = H
1
() According to the denition of the domain of o,
we observe that, for n 1(o), then it should exist some ) 1
2
() such that,
for all H
1
() :
o(n, ) =
_

)(r)(r)dr . (4.4.3)
Then, one can nd rst show (by taking in (4.4.3) C

0
()) that, in the
sense of distribution,
n +n = ) , (4.4.4)
27
and consequently that :
1(o) \() := n H
1
() [ n 1
2
() . (4.4.5)
But this is not enough for characterizing the domain.
We refer to [Lio1], [Lio2], [LiMa] or better [DaLi], Vol. 4 (p. 1222-1225)) for a
detailed explanation. We rst recall the Green-Riemann Formula :
_

n [ ) =
_

(n) dr +
_

(n,) dj

, (4.4.6)
where dj

is the induced measure on the boundary, which is clearly true for

n H
2
() (or for n C
1
()) and H
1
(). We unfortunately do not know
that \() H
2
() and the inclusion in H
1
() is not sucient for dening
the normal trace. But this formula can be extended in the following way.
We rst observe that, for H
1
0
() and n \, we have :
_

n [ ) =
_

(n) dr . (4.4.7)
This shows that the expression

u
() :=
_

n [ )
_

(n) dr ,
which is well dened for n \ and H
1
() depends only of the restriction
of to .
If
0
C

(), we can then extend

0
inside as a function = 1
0
in
C

() and the distribution

9
is dened as the map
0

u
(1
0
).
One observes also that, when n C
1
() or n H
2
(), the Green-Riemann
formula shows that :

u
(
0
) =
_

(n,)
0
dj

.
So we have found a natural way to extend the notion of trace of the normal
derivative for n \ and we write :

1
n =
u
.
We then conclude (using (4.4.4) and again (4.4.3) this time in full generality)
that :
1(o) = n \() [
1
n = 0 ,
and that
o = + 1 .
The operator o is called the Neumann realization of the Laplacian in 1
2
().
9
By chosing a more specic R continuous from H
1
2 () into H
1
(), we get that u can
be extended as a linear form on H
1
2 ().
28
Remark 4.4.2.
Another standard regularity theorem shows that
1(o) = n H
2
() [
1
n = 0 ,
and the notion of normal trace n
1
n for n H
2
() is more standard
10
.
10
The trace is in H
1
2 ().
29
5 Compact operators : general properties and
examples.
5.1 Denition and properties.
We just recall here very briey the basic properties of compact operators and
their spectral theory. We will emphasize on examples. We refer to the book by
H. Brezis [Br] (Chap. VI).
Let us recall that an operator T from a Banach 1 into a Banach 1 is compact
if the range of the unit ball in 1 by T is relatively compact in 1. We denote by
/(1, 1) the space of compact operators which is a closed subspace in L(1, 1).
There is an alternative equivalent denition in the case when 1 and 1 are
Hilbert spaces by using sequences. The operator is compact if and only if, for
any sequence r
n
which converges weakly in 1, Tr
n
is a strongly convergent
sequence in 1. Here we recall that a sequence is said to be weakly convergent in
H, if, for any j H, the sequence < r
n
, j
H
is convergent. Such a sequence is
bounded (Banach-Steinhauss Theorem) and we recall that, in this case, there
exists a unique r H, such that < r
n
, j
H
< r, j
H
for all j H. In this
case, we write : r
n
j.
Let us recall that when one composes a continuous operator and a compact
operator (in any order) one gets a compact operator. This could be one way to
prove the compactness.
Another ecient way for proving compactness of an operator T is to show that
it is the limit (for the norm convergence) of a sequence of continuous operators
with nite rank (that is whose range is a nite dimensional space). We observe
indeed that a continuous operator with nite rank is clearly a compact operator
(in a nite dimensional space the closed bounded sets are compact).
5.2 Examples
Continuous kernels .
The rst example of this type is the operator T
K
associated to the continuous
kernel 1 on [0, 1] [0, 1].
By this we mean that the operator T
K
is dened by
1 n (T
K
n)(r) =
_
1
0
1(r, j)n(j)dj . (5.2.1)
Here 1 could be the Banach space C
0
([0, 1]) (with the Sup norm) or 1
2
(]0, 1[).
Proposition 5.2.1.
If the kernel 1 is continuous, then the operator T
K
is compact from 1 into 1.
There are two standard proofs for this proposition. The rst one is based on
Ascolis Theorem giving a criterion relating equicontinuity of a subset of con-
tinuous functions on a compact and relatively compact sets in C
0
([0, 1]).
30
The other one is based on the Stone-Weierstrass Theorem permitting to recover
the operator as the limit of a sequence of nite rank operators T
Kn
associated
to kernels 1
n
of the form 1
n
(r, j) =

jn
j=1
)
j,n
(r)p
j,n
(j).
Let us study three other examples.
The rst example comes from statistical mechanics, the second one from the
spectral theory for the Dirichlet realization of the Laplacian and the third one
from Quantum Mechanics.
The transfer operator. The transfer operator is the operator associated
with the kernel
1
t
(r, j) = exp
\ (r)
2
expt[r j[
2
exp
\ (j)
2
, (5.2.2)
where t 0, and \ is a C

(R) function such that

_
R
exp\ (r)dr < + .
The 1
2
- boundedness of operators with integral kernel is proven very often
through the
Lemma 5.2.2. Schurs Lemma
Let K an operator associated with an integral kernel 1, that is a function
(r, j) 1(r, j) on R
m
R
m
, satisfying
`
1
:= sup
xR
m
_
R
m
[1(r, j)[ dj < + ,
`
2
:= sup
yR
m
_
R
m
[1(r, j)[ dr < + .
(5.2.3)
Then K, initially dened for n C

0
(R
m
) by
(1n)(r) =
_
R
m
1(r, j)n(j)dj ,
can be extended as a continuous linear operator in L(1
2
(R
m
)) (still denoted by
K or T
K
), whose norm satises
[[K[[
_
`
1
`
2
. (5.2.4)
Proof:
By the Cauchy-Schwarz inequality, we have
[1n(r)[
2

_
[1(r, j)[[n(j)[
2
dj
_
[1(r, j)[dj `
1
_
[1(r, j)[[n(j)[
2
dj .
Integrating with respect to r and using Fubinis Theorem, we then obtain the
result.
31
In our case, the operator T
K
is actually an Hilbert-Schmidt operator, i.e.
an operator whose integral kernel is in 1
2
(R
m
R
m
) (with : = 1). One can
indeed prove, using Cauchy-Schwarzs inequality, show that :
[1n(r)[
2

_
[n(j)[
2
dj
_
[1(r, j)[
2
dj ,
and one obtains :
[[T
K
[[ [[1[[
L
2
(R
m
R
m
)
. (5.2.5)
It is then easy to show that T
K
is a compact operator. Its kernel 1 is indeed
the limit in 1
2
of a sequence 1
n
such that T
Kn
is of nite rank. If
j
is
an orthonormal basis in 1
2
(R
m
), one rst shows that the basis
k

is an
orthonormal basis in 1
2
(R
m
R
m
). Here we have by denition :
(
k

)(r, j) :=
k
(r)

(j) .
We then obtain
1(r, j) =

k,
c
k,

k
(r)

(j) .
We now introduce
1
n
(r, j) :=

k+n
c
k,

k
(r)

(j) .
It is then easy to see that T
Kn
is of nite rank because its range is included in
the linear space generated by the
k
s (/ = 1, , n). Moreover, we have
lim
n+
[[1 1
n
[[
L
2
(R
m
R
m
)
= 0 .
Coming back to the corresponding operators, we get
lim
n+
[[T
K
T
Kn
[[
L(L
2
(R
m
))
= 0 .
The inverse of the Dirichlet operator .
We come back to the operator o, which was introduced in the study of the
Dirichlet realization. One can show the following
Proposition 5.2.3.
The operator o
1
is compact.
Proof.
The operator o
1
can indeed be considered as the composition of a continuous
operator from 1
2
into \ = H
1
0
() and of the continuous injection of \ into
1
2
(). If is relatively compact, we know (cf [Br]) that we have compact
injection from H
1
() into 1
2
(). Hence the injection of \ into 1
2
is compact
and o
1
is compact. For the continuity result, we observe that, for all n 1(o) :
[[on[[
H
[[n[[
H
on [ n) = o(n, n) [[n[[
2
V
[[n[[
V
[[n[[
H
.
32
This gives, for all n 1(o), the inequality :
[[on[[
H
[[n[[
V
. (5.2.6)
Using the surjectivity of o, we get :
[[o
1
[[
L(H,V )

1

. (5.2.7)
Note that in our example = 1 but that this part of the proof is completely
general.
The inverse of the harmonic oscillator.
The analysis is analogous. We have seen that the Sobolev space H
1
0
(R) has
to be replaced by the space
1
1
(R) := n 1
2
(R) , rn 1
2
and dn,dr 1
2
.
We can then prove, using a standard precompactness criterion, that 1
1
(R) has
compact injection in 1
2
(R). One has in particular to use the inequality :
_
|x|R
[n(r)[
2
dr
1
1
2
[[n[[
2
B
1
(R)
. (5.2.8)
It is very important to realize that the space H
1
(R) is not compactly injected in
1
2
. To understand this point, it is enough to consider the sequence n
n
= (rn)
where is a function in C

0
(R) with norm in 1
2
equal to 1. It is a bounded
sequence in H
1
, which converges weakly in H
1
to 0 and is not convergent in
1
2
(R).
5.3 Adjoints and compact operators
We recall
11
that the adjoint of a bounded operator in the Hilbertian case is
bounded. When 1 and 1 are dierent Hilbert spaces , the Hilbertian adjoint is
dened through the identity :
< Tr, j
F
=< r, T

j
E
, r 1 , j 1 . (5.3.1)
Example 5.3.1.
Let be an open set in R
m
and

the operator of restriction to : 1

2
(R
m
)
n n
/
. Then

is the operator of extension by 0.

Exercise 5.3.2.
Let
0
be the trace operator on r
m
= 0 dened from H
1
(R
m
+
) onto H
1
2
(R
m1
).
Determine the adjoint.
11
We are mainly following Breziss exposition [Br].
33
In an Hilbert space, we have
`

= `

.
In particular, if ` is a closed subspace, we have already used the property
(`

= ` . (5.3.2)
In the case of bounded operators (T L(1, 1)), one gets easily the proper-
ties
(T

) = 1(T)

, (5.3.3)
and
1(T) = ((T

))

. (5.3.4)
Let us also recall the proposition
Proposition 5.3.3.
The adjoint of a compact operator is compact.
5.4 Precompactness
We assume that the reader is aware of the basic results concerning compact sets
in metric spaces. We in particular recall that in a complete metric space 1, an
ecient way to show that a subset ` is relatively compact is to show, that for
any c 0, one can recover ` by a nite family of balls of radius c in 1.
The second standard point is to remember Ascolis Theorem, giving a crite-
rion for a bounded subset in C
0
(1) (1 compact in R
m
) to be relatively compact
in term of uniform equicontinuity. Ascolis Theorem gives in particular :
the compact injection of C
1
(1) into C
0
(1)
the compact injection of H
m
() into C
0
(1), for :
n
2
and with 1 = .
Let us recall nally a general proposition permitting to show that a subset
in 1
2
is relatively compact.
Proposition 5.4.1. Let 1
2
(R
m
). Let us assume that :
1. is bounded in 1
2
(R
m
), that is there exists ` 0 such that :
[[n[[
L
2 ` , n .
2. The expression c(n, 1) :=
_
|x|R
[n(r)[
2
dr tends to zero as 1 +
uniformly with respect to n .
3. For / R
m
, let
h
dened on 1
2
by : (
h
n)(r) = n(r /). Then the
expression (n, /) := [[
h
n n[[
L
2 tends to zero as / 0, uniformly with
respect to .
Then is relatively compact in 1
2
.
34
This proposition can be applied for showing :
the compact injection of H
1
0
() in 1
2
() when is regular and bounded,
the compact injection of H
1
() in 1
2
() when is regular and bounded,
the compact injection of 1
1
(R
m
) in 1
2
(R
m
).
35
6 Spectral theory for bounded operators.
6.1 Fredholms alternative
Let us rst recall Rieszs Theorem.
Theorem 6.1.1.
Let 1 be a normed linear space such that 1
E
is compact then 1 is nite dimen-
sional.
Let us now describe Fredholms alternative.
Theorem 6.1.2.
Let T /(1). Then
(i) (1 T) is nite dimensional.
(ii) 1(1 T) is closed (of nite codimension).
(iii) 1(1 T) = 1 if and only if (1 T) = 0 .
We shall only use this theorem in the Hilbertian framework, so 1 = H, and
we shall prove it for simplicity under the additional assumption that T = T

.
Proof.
We divide the proof in successive steps.
Step 1.
(i) is a consequence of Rieszs Theorem.
Step 2.
Let us show that 1(1 T) is closed.
Let j
n
a sequence in 1(1 T) such that j
n
j in H. We would like to show
that j 1(1 T).
Let r
n
in (1 T)

such that j
n
= r
n
Tr
n
.
Step 2a. Let us rst show the weaker property that the sequence r
n
is
bounded.
Let us indeed suppose that there exists a subsequence r
nj
such that [[r
nj
[[
+. Considering n
nj
= r
nj
,[[r
nj
[[, we observe that
() n
nj
Tn
nj
0 .
The sequence being bounded, we observe that (after possibly extracting a sub-
sequence) one can consider that the sequence n
nj
is weakly convergent. This
implies that Tn
nj
is convergent (T is compact). Using now (), we get the
convergence of n
nj
to n :
n
nj
n , Tn = n , [[n[[ = 1 .
But n (1 T)

, hence we get n = 0 and a contradiction.

Step 2b. We have consequently obtained that the sequence r
n
is bounded.
One can consequently extract a subsequence r
nj
that weakly converges to r

in H. Using the compactness of T, we get Tr

nj
converges strongly to Tr

.
36
Hence the sequence r
nj
tends strongly to j +Tr

.
We have nally
j +Tr

= r

,
and consequently proved that j = r

Tr

.
Step 3.
If (1 T) = 0, then (1 T

) = 0 (here we use for simplication our

additional assumption) and 1(1 T) being closed, we get
1(1 T) = (1 T

= H .
The converse is also immediate as T = T

.
Step 4.
We have
1(1 T)

= (1 T

) = (1 T) .
This shows, according to (i) that 1(1 T) is of nite codimension (second state-
ment of (ii)).
This ends the proof of Fredholms alternative in the particular case that T
is selfadjoint.
Remark 6.1.3.
Under the same asumptions, it is possible to show that
dim (1 T) = dim (1 T

) .
6.2 Resolvent set for bounded operators
In this subsection, 1 could be a Banach on R or C, but we will essentially need
the Hilbertian case in the applications treated here.
Denition 6.2.1. (Resolvent set)
For T L(1), the resolvent set is dened by
(T) = C ; (T 1) is bijective from 1 on 1. (6.2.1)
Note that in this case (T 1)
1
is continuous (Banachs Theorem). It is
easy to see that (T) is an open set in C. If
0
(T), we oberve that
(T ) = (T
0
)
_
1d + (
0
)(T
0
)
1
_
Hence (T ) is invertible if [
0
[ < [[(T
0
)
1
[[
1
. We also get the
following identity for all ,
0
(T) :
(T )
1
(T
0
)
1
= (
0
)(T )
1
(T
0
)
1
. (6.2.2)
37
Denition 6.2.2. (Spectrum)
The spectrum of T, (T), is the complementary set of (T) in C.
Note that (T) is a closed set in C. This is typically the case when T is a
compact injective operator in a Banach space of innite dimension.
We say that is an eigenvalue if (T 1) ,= 0. (T 1) is called the
eigenspace associated with .
Denition 6.2.3. (Point spectrum)
The point spectrum
p
(T) of T is dened as the set of the eigenvalues of T.
Example 6.2.4. (Basic example)
Let H = 1
2
(]0, 1[) and ) C
0
([0, 1]). Let T
f
be the operator of multiplication
by ). Then one has :
(T
f
) = Im ) =: C [ r [0, 1] with )(r) = .

p
(T
f
) = oto()) =: C [ :co:()
1
()) 0 .
For the rst assertion, it is rst immediate to see that if , Im ), then
T
(f)
1 is a continuous inverse of T
f
. On the other side, if = )(r
0
) for
some r
0
]0, 1[ then we have (T
f
)n
n
0 and [[n
n
[[ = 1 for n
n
=
1

n
(
xx0
n
),
where is a C

0
function such that [[[[ = 1. This shows that )(]0, 1[) (T
f
)
and we can conclude by considering the closure.
Note that the point spectrum is not necessarily closed. Note also that one
can have a strict inclusion of the point spectrum in the spectrum as can be
observed in the following example :
Example 6.2.5.
Let us consider 1 = /
2
(N) and let T be the shift operator dened by :
(Tn)
0
= 0 , (Tn)
n
= n
n1
, n 0 ,
where n = (n
0
, , n
n
, ) /
2
(N). Then it is easy to see that T est injective
(so 0 is not an eigenvalue) and is not surjective (so 0 is in the spectrum of T).
As another interesting example, one can consider :
Example 6.2.6.
Let 1 = /
2
(Z, C) and let T be the operator dened by :
(Tn)
n
=
1
2
(n
n1
+n
n+1
) , n Z ,
for n /
2
(Z). Then it is easy to see (by using expansion in Fourier series
12
)
that T has no eigenvalues and its spectrum is [1, +1].
12
Using the isomorphism between
2
(Z; C) and L
2
(S
1
; C), which associates to the sequence
(un)
nZ
the fonction
P
nZ
un exp in, one has to analyze the operator of multiplication by
cos :
f T f = cos f .
It is then analyzed as in example 6.2.4. One concludes that the spectrum of T is [1, +1].
38
Exercise 6.2.7.
Let [0, 1]. Let j and integers which are mutually prime. Analyze the
spectrum

of the operator H

dened on /
2
(Z) by
(H

n)
n
=
1
2
(n
n1
+n
n+1
) + cos 2(
j

n +) n
n
, n Z .
In order to make the analysis easier, one can admit (particular case of the so-
called Floquet theory), that one has =
[0,1]

, where

is the spectrum
of H

reduced to the space of the ns in /

such that n
n+q
= exp2i n
n
, for
n Z.
This operator plays an important role in Solid State Physics and is called the
Harpers operator.
We now replace the rational
p
q
by an irrational number . So we consider the
operator H
,
:=
1
2
(
+1
+
1
) + cos 2( + ) on /
2
(Z, C), where
k
(/ Z)
is the operator dened on /
2
(Z, C) by (
k
n)
n
= n
nk
.
Show that, if , Q, then the spectrum of H
,
is independent of . For this,
one can rst prove that H
,
is unitary equivalent with H
,+k
for any / Z.
Secondly, one can use the density of the set +Z+Z in R. Finally, one can
use the inequality
[[H
,
H
,
[[ 2[

[ .
Proposition 6.2.8.
The spectrum (T) is a compact set included in the ball 1(0, [[T[[).
This proposition is immediate if we observe that (1
T

) is invertible if
[[ [[T[[.
6.3 Spectral theory for compact operators
In the case of a compact operator, one has a more precise description of the
spectrum.
Theorem 6.3.1.
Let T /(1) where 1 is an innite dimensional Banach space. Then
1.
0 (T) .
2.
(T) 0 =
p
(T) 0 .
3. We are in one (and only one) of the following cases
either (T) = 0,
either (T) 0 is nite,
or (T)0 can be described as a sequence of distincts points tending
to 0.
39
4. Each
p
(T) 0 is isolated and dim(T 1) < +.
Proof.
a) If 0 , (T), then T admits a continuous inverse T
1
and the identity, which
can be considered as :
1 = T T
1
,
is a compact operator, as the composition of the compact operator T and the
continuous operator T
1
. Using Rieszs Theorem we get a contradiction in the
case 1 is supposed of innite dimension.
b) The fact that, if ,= 0 and (T), then is an eigenvalue, comes directly
from the Fredholms alternative applied to (1
T

).
c) The last step comes essentially from the following lemma :
Lemma 6.3.2.
Let (
n
)
n1
a sequence of distincts points
n
and
n
(T) 0, for all
n 0.
Then = 0.
Proof.
We just give the proof in the Hilbertian case. For all n 0, let c
n
be a
normalized eigenfunction such that (T
n
)c
n
= 0 and let 1
n
be the vectorial
space spanned by c
1
, c
2
, , c
n
. Let us show that we have a strict inclusion
of 1
n
in 1
n+1
.
We prove this point by recursion. Let us assume the result up to order n 1
and let us show it at order n. If 1
n+1
= 1
n
, then c
n+1
1
n
and we can write
c
n+1
=
n

j=1

j
c
j
.
Let us apply T to this relation. Using the property that c
n+1
is an eigenfunction,
we obtain

n+1
_
_
n

j=1

j
c
j
_
_
=
n

j=1

j
c
j
.
Using the recursion assumption, c
1
, , c
n
is a basis of 1
n
and the
j
be-
ing distincts, we obtain
j
= 0 for all , = 1, , n and a contradiction with
[[c
n+1
[[ = 1.
So we can nd a sequence n
n
such that n
n
1
n
1

n1
and [[n
n
[[ = 1. T being
compact, one would extract a convergent subsequence (still denoted by Tn
n
)
from the sequence (Tn
n
) and, if
n
,= 0, one would also have the conver-
gence of this subsequence (
1
n
Tn
n
) and consequently the Cauchy property.
Let us show that this leads to a contradiction. We remark that
(T
n
)1
n
1
n1
.
40
Let n : 2. We have
[[
Tun
n

Tum
m
[[
2
= [[
(Tn)un
n

(Tm)um
m
+n
n
n
m
[[
2
= [[
(Tn)un
n

(Tm)um
m
n
m
[[
2
+[[n
n
[[
2
[[n
n
[[
2
= 1 .
We can consequently not extract a Cauchy subsequence from the sequence
1
n
Tn
n
. This is in contradiction with the assumption ,= 0.
This ends the proof of the lemma and of the theorem.
We shall now consider the Hilbertian case and see which new properties can
be obtained by using the additional assumption that T is selfadjoint.
6.4 Spectrum of selfadjoint operators.
As T = T

, the spectrum is real. If Im ,= 0, one immediately veries that :

[ Im [ [[n[[
2
[ Im (T )n, n) [[(T )n[[ [[n[[ .
This shows immediately that the map (T ) is injective and with close range.
But the orthogonal of the range of (T ) is the kernel of (T

) which is
reduced to 0. So (T ) is bijective.
This was actually, a consequence of the Lax-Milgram theorem (in the simple
case when \ = H), once we have observed the inequality
[(T )n, n)[ [ Im [ [[n[[
2
.
Using again Lax-Milgrams theorem, we can show
Theorem 6.4.1.
Let T L(H) be a selfadjoint operator. Then the spectrum of T is contained in
[:, `] with : = infTn, n),[[n[[
2
and ` = supTn, n),[[n[[
2
. Moreover : and
` belong to the spectrum of T.
Proof:
We have already mentioned that the spectrum is real. Now if `, we can
apply the Lax-Milgram to the sequilinear-form (n, ) n, ) Tn, ).
Let us now show that ` (T).
We observe that, by Cauchy-Schwarz applied to the scalar product (n, )
`n, ) Tn, ), we have :
[`n Tn, )[ `n Tn, n)
1
2
` T, )
1
2
.
In particular, we get :
[[`n Tn[[
H
[[` T[[
1
2
L(H)
`n Tn, n)
1
2
. (6.4.1)
41
Let (n
n
)
nZ
be a sequence such that [[n
n
[[ = 1 and Tn
n
, n
n
) ` as n +.
By (6.4.1), we get that (T `)n
n
tends to 0 as n +. This implies that
` (T). If not, we would get that n
n
= (` T)
1
((` T)n
n
) tends to 0
in contradiction with [[n
n
[[ = 1.
This theorem admits the following important corollary
Corollary 6.4.2.
Let T L(H) be a selfadjoint operator such that (T) = 0. Then T = 0.
We rst indeed show that : = ` = 0 and consequently that < Tn, n
H
= 0
for all n H. But < Tn, can be written as a linear combination of terms
of the type < Tn, n
H
and this gives the result by taking = Tn.
Another connected property which is useful is the
Proposition 6.4.3.
If T is positive and selfadjoint then [[T[[ = `.
The proof is quite similar. We observe (Cauchy-Schwarz for (n, ) <
Tn, ) that :
[Tn, )[ Tn, n)
1
2
T, )
1
2
.
This implies, using Rieszs Theorem :
[[Tn[[ [[T[[
1
2
(Tn, n))
1
2
.
Coming back to the denition of [[T[[, we obtain :
[[T[[ [[T[[
1
2
`
1
2
,
and the inequality :
[[T[[ ` .
But it is immediate that :
Tn, n) [[T[[ [[n[[
2
.
This gives the converse inequality and the proposition.
6.5 Spectral theory for compact selfadjoint operators
We have a very precise description of the selfadjoint compact operators.
Theorem 6.5.1.
Let H be a separable Hilbert
13
space and T a compact selfadjoint operator. Then
H admits an Hilbertian basis consisting of eigenfunctions of T.
13
that is having a countable dense set.
42
Proof.
Let (
n
)
n1
be the sequence of disjoint eigenvalues of T, except 0. Their exis-
tence comes from Theorem 6.3.1 and we also observe that the eigenvalues are
real.
Let us dene
0
= 0.
We dene 1
0
= (T) and 1
n
= (T
n
1); We know (from Rieszs Theorem)
that
0 < dim 1
n
< + .
Let us show that H is the Hilbertian sum of the (1
n
)
n0
.
(i) The spaces (1
n
) are mutually orthogonal. If n 1
m
and 1
n
with
: ,= n, we have
< Tn,
H
=
m
< n,
H
=< n, T
H
=
n
< n,
H
,
and consequently
< n,
H
= 0 .
(ii) Let 1 be the linear space spanned by the (1
n
)
n0
. Let us verify that 1
is dense in H. It is clear that T1 1 and, using the selfadjoint character of
T, we have also T1

. The operator

T, obtained by restriction of T to
1

, is a compact selfadjoint operator. But one shows easily that (

T) = 0
and consequently

T = 0. But 1

(T) 1 and hence 1

= 0. 1 is
consequently dense in H.
(iii) To end the proof, one chooses in each 1
n
an Hilbertian basis. Taking
the union of these bases, one obtains an Hilbertian basis of H eectively formed
with eigenfunctions of T.
Remark 6.5.2.
If T is a compact selfadjoint operator, we can write any n in the form
n =
+

n=0
n
n
, with n
n
1
n
.
This permits to write
Tn =
+

n=1

n
n
n
.
If, for / N

, we dene T
k
by
T
k
n =
k

n=1

n
n
n
,
we easily see that T
k
is of nite rank and that
[[T T
k
[[ sup
nk+1
[
n
[ .
Hence the operator T appears as the limit in L(H) of the sequence T
k
as /
+.
43
7 Examples.
We go back to our previous examples and analyze their properties.
7.1 The transfer operator.
7.1.1 Compactness
The transfer operator (which was introduced in (5.2.2) ) is compact and admits
consequently a sequence of eigenvalues
n
tending to 0 as n +. Let us
show the
Lemma 7.1.1.
The transfer operator is positive and injective.
Proof of the lemma.
Let n 1
2
(R). We can write
14
, with (r) = exp
V (x)
2
n(r)
< T
K
n, n
H
=
_
R
2
expt[r j[
2
(r)(j)drdj .
Using the properties of the Fourier transform and of the convolution, we deduce
< T
K
n, n
H
= c
t
_
R
exp
[[
2
4t
[

()[
2
d ,
where c
t
0 is a normalization constant.
The spectrum is consequently the union of a sequence of positive eigenvalues
and of its limit 0. T
K
can be diagonalized in an orthonormal basis of eigenfunc-
tions associated with strictly positive eigenvalues. We emphasize that 0 is in
the spectrum but not an eigenvalue.
Theorem 6.4.1 says also that [[T
K
[[ is the largest eigenvalue and is isolated. A
natural question is then to discuss the multiplicity of each eigenvalue, i. e. the
dimension of each associated eigenspace. We shall see later (Krein-Rutmans
Theorem) that the largest eigenvalue is of multiplicity 1.
7.1.2 About the physical origin of the problem.
Our initial problem was to nd a rather general approach for the estimate of
the decay of correlations attached to gaussian like measures of the type
exp(A) dA (7.1.1)
on R
n
with . One proof of this estimate (in the case when has a particular
structure) is based on the analysis of the transfer matrix method, originally
14
If we make only the weak assumption that exp V L
1
(R), it is better to start for the
proof of the positivity by considering us in C

0
(R) and then to treat the general case by using
the density of C

0
in L
2
.
44
due to Kramers-Wannier, that we have already seen for the study of the Ising
model. We present here briey the technique for our toy model. We shall only
consider the case when d = 1 and treat the periodic case, that is the case when
1, , n is a representation of
per
= Z,nZ.
We consider the particular potential

(n)
(A) (A)
1
/
_
_
n

j=1
\ (r
j
) +
[r
j
r
j+1
[
2
4
_
_
(7.1.2)
where we take the convention that r
n+1
= r
1
and where / is possibly a semiclas-
sical parameter which is sometimes chosen equal to one if we are not interested
in the semiclassical aspects. More generally, we could more generally consider
examples of the form:

h
(A) =
1
/
_
_
n

j=1
(\ (r
j
) +1(r
j
, r
j+1
))
_
_
(7.1.3)
where 1 is a symmetric interaction potential on R R. Let us mention how-
ever that the example (7.1.2) appears naturally in quantum eld theory when
the so called lattice approximation is introduced. For this special class of
potentials, we shall demonstrate that the informations given by the transfer
operator method are complementary to the other approachs we have explained
before. We shall present the dictionary between the properties of the measure
/

n
2
exp
h
(A) dA on R
n
and the spectral properties of the transfer operator
K
V
(which is also called Kac operator for some particular models introduced
by M. Kac) whose integral kernel is real and given on R R by
1
V
(r, j) = /

1
2
exp
\ (r)
2/
exp
[r j[
2
4/
exp
\ (j)
2/
. (7.1.4)
By integral kernel (or distribution kernel), we mean
15
a distribution in T

(R
2
)
such that the operator is dened from C

0
(R) into T

(R) by the formula

_
R
(K
V
n)(r)(r) dr =
_
RR
1
V
(r, j) n(r) (j) drdj , n , C

0
(R) .
(7.1.5)
This dictionary permits to obtain interesting connections between estimates
for the quotient j
2
,j
1
of the two rst largest eigenvalues of the transfer operator
and corresponding estimates controlling the speed of convergence of thermody-
namic quantities. In particular this speed of convergence is exponentially rapid
as n +.
We know that when the operator K is compact, symmetric and injective,
then there exists a decreasing (in modulus) sequence j
j
of eigenvalues tending
15
Here we shall always consider much more regular kernels which are in particular continu-
ous. So the notation
R
can be interpreted in the usual sense. In general, this means that the
distribution kernel K
V
is applied to the test function (x, y) u(x) v(y).
45
to 0 and a corresponding sequence of eigenfunctions n
j
which can be normalized
in order to get an orthonormal basis of 1
2
(R). Moreover, the operator becomes
the limit in norm of the family of operators K
(N)
whose corresponding kernel
are dened by
1
(N)
(r, j) =
N

j=1
j
j
n
j
(r) n
j
(j) . (7.1.6)
We recall indeed that
[[KK
(N)
[[ sup
j>N
[j
j
[ . (7.1.7)
The symmetric operators
16
are called trace class if we have in addition the
property that
[[K[[
tr
:=

j
[j
j
[ < + (7.1.8)
In this case, we get that
[[KK
(N)
[[
tr

j>N
[j
j
[ . (7.1.9)
For a trace class symmetric operator, we can in particular dene the trace as
Tr K =

j
j
j
, (7.1.10)
and this operation is continuous on the space of the trace class operators
[ Tr K [ [[K[[
tr
. (7.1.11)
This of course extends the usual notion of trace for matrices. We can actually
compute directly the trace of a trace-class operator in the following way. We
rst observe that
Tr K = lim
N+
Tr K
(N)
. (7.1.12)
Then we observe that
Tr K
(N)
=
N

j=1
j
j
=
_
R
1
(N)
(r, r)dr . (7.1.13)
We consequently obtain that
Tr K = lim
N+
_
R
1
(N)
(r, r)dr . (7.1.14)
16
When K is not symmetric, trace class operators can still be dened by considering

K.
Note that when K is trace class, one can compute the trace by considering any orthonormal
basis e
i
: Tr K =
P
i
Ke
i
, e
i
).
46
If we observe that r

j
[j
j
[[n
j
(r)[
2
is in 1
1
(R), then it is natural
17
to hope
(but this is not trivial!) that r 1(r, r) is in 1
1
and that
Tr K =
_
R
1(r, r) dr . (7.1.15)
Note that it is only when K is positive that the niteness of the right hand side
in (7.1.15) will imply the trace class property.
In the case the operator is Hilbert-Schmidt, that is with a kernel 1 in 1
2
, then
the operator is also compact and we have the identity

j
j
2
j
=
_
RR
[1(r, j)[
2
drdj . (7.1.16)
This gives an easy criterion for verifying the compactness of the operator. Let
us rst look at the thermodynamic limit. This means that we are interested in
the limit lim
n+
1
n
ln
__
R
n
exp(A) dA
_
. We start from the decomposition:
exp(A) = 1
V
(r
1
, r
2
) 1
V
(r
2
, r
3
) 1
V
(r
n1
, r
n
) 1
V
(r
n
, r
1
) (7.1.17)
and we observe that
_
R
n
exp(A) dA =
_
R
1
V,n
(j, j) dj (7.1.18)
where 1
V,n
(r, j) is the distribution kernel of (K
V
)
n
. Our assumption on \
permits to see that (1
V
)
n
is trace class
18
and we rewrite (7.1.18) in the form:
_
R
n
exp(A) dA = Tr [(K
V
)
n
] =

j
j
n
j
, (7.1.19)
where the j
j
are introduced in (7.1.24).
We note also for future use that
1
V,n
(r, j) =

j
j
n
j
n
j
(r) n
j
(j) . (7.1.20)
17
Let us sketch a proof of (7.1.15) in our particular case. K being positive and with an
explicit kernel : exp
V (x)
2
exp [x y[
2
exp
V (y)
2
, one can nd an Hilbert-Schmidt
operator L satisfying L

L = K. The kernel of L is given, for a suitable , by L(x, y) =

c

exp [x y[
2
exp
V (y)
2
. We note indeed that c

exp [x y[
2
is the distribution kernel
of exp t

for a suitable t

> 0.
Then we observe that L is Hilbert-Schmidt and that [[L[[
2
H.S
=
P
j

j
= Tr K. Using the
previously mentioned formula for the Hilbert-Schmidt norm and the property that
K(x, x) =
Z
L

(x, z) L(z, x) dz =
Z
L

(z, x) L(z, x) dz ,
one obtains (7.1.15).
18
See for example [Ro].
47
In particular we get for the thermodynamic limit:
lim
n
ln
_
R
n
exp(A) dA
n
= lnj
1
. (7.1.21)
Moreover the speed of the convergence is easily estimated by:
ln[
ln
_
R
n
exp(A) dA
n
lnj
1
[ = nln(
j
2
j
1
) ln/
2
+ lnn +O(exp
2
n) ,
(7.1.22)
where /
2
is the multiplicity of j
2
.
7.1.3 Krein-Rutmans Theorem.
We observe now that the kernel (r, j) 1
V
(r, j) satises the condition
1
V
(r, j) 0 , r, j R . (7.1.23)
In particular it satises the assumptions of the extended Perron-Frobenius The-
orem also called Krein-Rutmans Theorem and our positive operator K
V
admits
consequently a largest eigenvalue j
1
equal to [[1
V
[[ which is simple and corre-
sponds to a unique strictly positive normalized eigenfunction which we denote
by n
1
. Let j
j
the sequence of eigenvalues that we order as a decreasing sequence
tending to 0:
0 j
j+1
j
j
.... j
2
< j
1
. (7.1.24)
We shall denote by n
j
a corresponding orthonormal basis of eigenfunctions with
1
V
n
j
= j
j
n
j
, [[n
j
[[ = 1 . (7.1.25)
Let us present the statements:
Denition 7.1.2.
Let be a bounded positive operator on a Hilbert space H = 1
2
(A, d) where
(A, ) is a measured space. Then we say that has a strictly positive kernel if,
for each choice of a non negative function H ([[[[ , = 0), we have
0 <
almost everywhere.
It is immediate to see that the transfer operator satises this condition.
The theorem generalizing the Perron-Frobenius Theorem is then the following:
Theorem 7.1.3.
Let be a bounded positive compact symmetric operator on H having a strictly
positive kernel and let
19
[[[[ = be the largest eigenvalue of . Then has
multiplicity 1 and the corresponding eigenfunction n

can be chosen to be a
strictly positive function.
19
See Proposition 6.4.3
48
Proof:
Since maps real functions into real functions, we may assume that n

is real.
We now prove that
n

, n

) [n

[, [n

[).
This is an immediate consequence of the strict positivity of the kernel. We just
write:
n

= n
+

and
[n

[ = n
+

+n

and the above inequality is then a consequence of

n
+

, n

) 0 ,
and
n
+

, n

) 0 .
We then obtain
[[n

[[
2
= n

, n

) [n

[, [n

[) [[[[ [[n

[[
2
= [[n

[[
2
.
This implies
n

, n

) = [n

[, [n

[) .
This equality means
n
+

, n

) +n

, n
+

) = 0 .
We then get a contradiction unless n
+

= 0 or n

= 0. We can then assume

n

0 and the assumption gives again

0 < , n

) = , n

) ,
for any positive . This gives
n

0 o.c .
But
n

=
1
n

and this gives

n

0 o.c .
Finally if there are two linearly independent eigenfunctions

and n

corre-
sponding to , we would obtain the same property for

by considering as
new Hilbert space the orthogonal of n

in H. But it is impossible to have two

orthogonal vectors which are strictly positive.
q.e.d.
Remark 7.1.4.
In the case t = 0, we keep the positivity but lose the injectivity ! The spectrum is
easy to determine. We are indeed dealing with the orthonormal projector asso-
ciated to the function r exp
V (x)
2
. The real number 1 is a simple eigenvalue
and 0 is an eigenvalue whose corresponding eigenspace is innite dimensional.
49
7.2 The Dirichlet realization, a model of operator with
compact resolvent.
We can apply Theorem 6.5.1 to the operator (
D
+1d)
1
. We have seen that
this operator is compact and it is clearly injective (by construction). It was also
seen as a selfadjoint and positive. Moreover, the norm of this operator is less or
equal to 1.
There exists consequently a sequence of distinct eigenvalues j
n
tending to 0
(with 0 < j
n
1) and a corresponding orthonormal basis of eigenfunctions
such that (
D
+1)
1
is diagonalized. If
n,j
(, = 1, . . . , /
n
) is a corresponding
basis of eigenfunctions associated with j
n
, that is, if
(
D
+1)
1

n,j
= j
n

n,j
,
we rst observe that
n,j
1(
D
+1); hence
n,j
H
1
0
() H
2
() (if is
relatively compact with regular boundary) and

n,j
= (
1
j
n
1)
n,j
.
The function
n,j
is consequently an eigenfunction of
D
associated with the
eigenvalue
n
= (
1
n
1).
Let us show, as one easily guesses, that this basis
n,j
permits eectively the
diagonalization of
D
.
Let us indeed consider n =

n,j
n
n,j

n,j
in the domain of
D
. Let us consider
the scalar product <
D
n,
m,

H
. Using the selfadjoint character of
D
,
we get
<
D
n,
m,

H
=< n,
D

m,

H
=
m
n
m,
.
Observing, that 1(
D
) = 1(o
1
), one obtains that the domain of
D
is
characterized by
1(
D
) = n 1
2
[

n,j

2
n
[n
n,j
[
2
< + .
Here we have used the property that for any , we have the identity :

nN
n
n,j

n,j
= o(

nN

n
n
n,j

n,j
) .
We have consequently given a meaning to the following diagonalization formula

D
=

En
, (7.2.1)
where
En
is the orthogonal projector on the eigenspace 1
n
associated with the
eigenvalue
n
.
Let us remark that it results from the property that the sequence j
n
tends to
0 the property that
n
tend to +.
Let us also prove the
50
Lemma 7.2.1.
The lowest eigenvalue of the Dirichlet realization of the Laplacian in a relatively
compact domain is strictly positive :

1
0 . (7.2.2)
Proof.
We know that
1
0; the Dirichlet realization of the Laplacian is indeed posi-
tive. If
1
= 0, a corresponding normalized eigenfunction
1
would satisfy
<
1
,
1
= 0 ,
and consequently

1
= 0 , dans .
This leads rst
1
= Cste in each connected component of but because the
trace of
1
on vanishes (
1
H
1
0
()) implies that
1
= 0. So we get a
contradiction.
Finally it results from standard regularity theorems (See [Br], [LiMa]) that
the eigenfunctions belong (if is regular) to C

().
We now show the
Proposition 7.2.2.
The lowest eigenvalue is simple and one can choose the rst eigenfunction to be
strictly positive.
The natural idea is to apply Krein-Rutmans Theorem to (
D
+1)
1
. One
has to show that this operator is positivity improving. This is indeed the case if
the domain is connected but the proof of this property will not be given.
We will only show that (
D
+ 1)
1
is positivity preserving and observe that
this implies (following the proof of Krein-Rutmans theorem) that if is an
eigenfunction then [[ is an eigenfunction. Then the proof of Proposition 7.2.2
will be completed by using the properties of superharmonic functions.
Lemma 7.2.3.
The operator (
D
+1)
1
is positivity preserving.
The proof is a consequence of the Maximum principle. It is enough to show
that
n +n = ) ,
0
n = 0 and ) 0 a.e ,
implies that n 0 almost everywhere.
We introduce n
+
= max(n, 0) and n

= inf(n, 0). A standard proposition

(see below
20
or the book of Lieb-Loss) shows that n
+
and n

belong to H
1
0
().
Multiplying by n

, we obtain
_
n
+
n

dr [[n

[[
2
[[n

[[
2
=
_

)n

dr 0 ,
20
Recall that [u[ = u
+
+u

and u = u
+
u

.
51
which implies, using the corollary of the next proposition n

= 0.
Proposition 7.2.4.
Suppose that ) 1
1
loc
(R
n
) with ) 1
1
loc
(R
n
). Then also [)[ 1
1
loc
(R
n
) and
with the notation
sign . =
_
z
|z|
, . ,= 0
0, . = 0,
(7.2.3)
we have
[)[(r) = Resign()(r)))(r) almost everywhere . (7.2.4)
In particular,

[)[

[)[ ,
almost everywhere.
Corollary 7.2.5.
Under the assumptions of Proposition 7.2.4, we have if ) is real
)
+
(r) =
1
2
(sign()(r)) + 1))(r) almost everywhere , (7.2.5)
and
)

(r) =
1
2
(sign()(r)) 1))(r) almost everywhere . (7.2.6)
Hence
)
+
(r))

(r) =
1
4
(sign()(r))
2
1)[)(r)[
2
0 almost everywhere . (7.2.7)
Proof of Proposition 7.2.4.
Suppose rst that n C

(R
n
) and dene [.[

=
_
[.[
2
+c
2
c, for . C and
c 0. We observe that
0 [.[

[.[ and lim

0
[.[

= [.[ .
Then [n[

(R
n
) and
[n[

=
Re (nn)
_
[n[
2
+c
2
. (7.2.8)
Let now ) be as in the proposition and dene )

as the convolution
)

= )

52
with

being a standard approximation of the unity for convolution. Explicitly,

we take a C

0
(R
n
) with
0,
_
R
n
(r) dr = 1,
and dene

(r) :=
n
(r,), for r R
n
and 0.
Then )

), [)

[ [)[ and )

) in 1
1
loc
(R
n
) as 0.
Take a test function C

0
(R
n
). We may extract a subsequence
k

kN
(with
k
0 for / ) such that )

k
(r) )(r) for almost every r supp.
We restrict our attention to this subsequence. For simplicity of notation we
omit the / from the notation and write lim
0
instead of lim
k
.
We now calculate, using dominated convergence and (7.2.8),
_
()[)[ dr = lim
0
_
()[)[

dr
= lim
0
lim
0
_
()[)

dr
= lim
0
lim
0
_

Re ()

)
_
[)

[
2
+c
2
dr
Using the pointwise convergence of )

(r) and |)

)|
L
1
(supp )
0, we can
take the limit 0 and get
_
()[)[ dr = lim
0
_

Re ()))
_
[)[
2
+c
2
dr. (7.2.9)
Now, ) 1
1
(R
n
) and
f(x)

|f|
2
+
2
sign)(r) as c 0, so we get (7.2.4) from
(7.2.9) by dominated convergence.
We can now look directly (which is a consequence of this positivity improving
property, but we have not given the proof!) at the property that the rst
eigenfunction does not vanish. The eigenfunction is positive, belongs to C

()
by a regularity theorem and satises
n = n 0 .
Hence n is superharmonic and (see Lieb-Loss) satises the mean value property :
For all j , for all 1 0 such that 1(j, 1) , then
n(j)
1
vol(1(j, 1))
_
B(y,R)
n(.) d. .
Moreover, we know that inf n = 0. Applying this mean value property with j
(if any) such that n(j) = 0, we obtain that n = 0 in 1(j, 1). Using in addition
a connectedness argument, we obtain that in a connected open set n is either
53
identically 0 or strictly positive.
Let us come back to what appears in the proof of Krein-Rutmans theorem.
n
+

and n

are either 0 or strictly positive eigenfunctions and we have also

n
+

, n

) = 0. Hence n
+

or n

should vanish. We can then show the simplicity

of the rst eigenvalue as in the proof of Krein-Rutmans theorem. Hence we
have nally completed the proof of Proposition 7.2.2.
7.3 Extension to operators with compact resolvent
What we have done for the analysis of the Dirichlet realization is indeed quite
general. It can be applied to selfadjoint operators, which are bounded from
below and with compact resolvent.
We show that in this case, there exists an innite sequence (if the Hilbert space
is innite dimensional) of real eigenvalues
n
tending to + such that the cor-
responding eigenspaces are mutually orthogonal, of nite dimension and such
that their corresponding Hilbertian sum is equal to H.
Typically, one can apply the method to the Neumann realization of the Lapla-
cian in a relatively compact domain or to the harmonic oscillator in R
m
.
7.4 Operators with compact resolvent : the Schr odinger
operator in an unbounded domain.
We just recall some criteria of compactness for the resolvent of the Schrodinger
operator 1 = +\ in R
m
in connection with the precompactness criterion.
In the case of the Schrodinger equation on R
m
and if \ is C

and bounded
from below, the domain of the selfadjoint extension is always contained in
Q(1) := H
1
V
(R
m
) = n H
1
(R
m
)[(\ +C)
1
2
n 1
2
(R
m
) .
Q(1) is usually called the form domain of the form
n
_
R
m
[n[
2
dr +
_
R
m
\ (r)[n(r)[
2
dr .
It is then easy to see that, if \ tends to , then the injection of H
1
V
in 1
2
is compact (using a criterion of precompactness). We then obtain, observing
that (1 + 1)
1
is continuous from 1
2
into H
1
V
, that the resolvent (1 +)
1
is
compact for , (1).
In the case of a compact manifold ` and if we consider the Laplace-Beltrami
operator on `, then the compactness of the resolvent is obtain without addi-
tional assumption on \ . The domain of the operator is H
2
(`) and we have
compact injection from H
2
(`) into 1
2
(`).
The condition that \ as [r[ is not a necessary condition. We can
indeed replace it by the weaker sucient condition
Proposition 7.4.1.
Let us assume that the injection of H
1
V
(R
m
) into 1
2
(R
m
) is compact then 1 has
compact resolvent.
54
More concretely, the way to verify this criterion is to show the existence of
a continuous function r (r) tending to as [r[ + such that
H
1
V
(R
m
) 1
2

(R
m
) . (7.4.1)
Of course, the preceding case corresponds to = \ , but, as typical example of
this strategy, we shall show in exercise 7.4.3 that the Schrodinger operator on
R
2
, +r
2
j
2
+ 1, has compact inverse.
On the other hand, the criterion that \ + as [r[ + is not not too far
from optimality.
We can indeed prove
Lemma 7.4.2.
Suppose that \ 0 and that there exists : 0 and a sequence
n
such that
[
n
[ + and such that
sup
nN
sup
xB(n,r)
\ (r) < + . (7.4.2)
Then +\ + 1 is not with compact inverse.
Proof.
Let us consider the sequence

n
(r) = (r
n
) . (7.4.3)
Here is a compactly supported function of 1
2
norm 1 and with support in
1(0, :).
We observe that the
n
are an orthogonal sequence (after possibly extracting
a subsequence for obtaining that the supports of
n
and
n
are disjoint for
n ,= n

) which satises for some constant C

[[
n
[[
2
H
2 +[[\
n
[[
2
L
2 C . (7.4.4)
In particular, there exists C such that :
[[( +\ + 1)
n
[[
L
2 C , n N .
But we can not extract from this sequence a strongly convergent sequence in
1
2
, because
n
is weakly convergent to 0 and [[
n
[[ = 1. So the operator
( +\ +1)
1
can not be a compact operator.
Exercise 7.4.3.
Show that the unbounded operator on 1
2
(R
2
)
1 :=
d
2
dr
2

d
2
dj
2
+r
2
j
2
,
has compact resolvent.
55
Hint.
One can introduce
A
1
=
1
i

x
, A
2
=
1
i

y
, A
3
= rj ,
and show, for , = 1, 2 and for a suitable constant C, the following inequality :
[[(r
2
+j
2
+ 1)

1
4
[A
j
, A
3
]n[[
2
C
_
< 1n, n
L
2
(R
2
)
+[[n[[
2
_
,
for all n C

0
(R
2
).
One can also observe that
i[A
1
, A
3
] = j , i[A
2
, A
3
] = r ,
and then
[[(r
2
+j
2
+ 1)

1
4
[A
1
, A
3
]n[[
2
+[[(r
2
+j
2
+ 1)

1
4
[A
2
, A
3
]n[[
2
+[[n[[
2
[[(r
2
+j
2
+ 1)
1
4
n[[
2
.
For the control of [[(r
2
+j
2
+ 1)

1
4
[A
1
, A
3
]n[[
2
, one can remark that
[[(r
2
+j
2
+ 1)

1
4
[A
1
, A
3
]n[[
2
=
ij
(r
2
+j
2
+ 1)
1
4
n [ (A
1
A
3
A
3
A
1
)n) ,
perform an integration by parts, and control a commutator.
7.5 The Schr odinger operator with magnetic eld
We can consider on R
m
the so-called Schrodinger operator with magnetic eld :
1
A,V
:=
A
+\ , (7.5.1)
where

A
:=
m

j=1
(
1
i

xj

j
(r))
2
. (7.5.2)
Here r

= (
1
(r), ,
n
(r)) is a vector eld on R
m
called the magnetic
potential and \ is called the electric potential. It is easy to see that, when
\ is semi-bounded, the operator is symmetric and semi-bounded on C

0
(R
m
).
We can therefore consider the Friedrichs extension and analyze the property of
this selfadjoint extension.
A general question arises if one can get operator of the type 1
A,V
which are with
compact resolvent if \ = 0. This is the problem which is called the problem of
the magnetic bottle.
The heuristical idea is that the module of the magnetic eld can play in some
sense the role of the electric potential if it does not oscillate too rapidly (: 2).
56
For dening the magnetic eld it is probably easier to consider the magnetic
potential as a one-form

A
=
n

j=1

j
(r)dr
j
.
The magnetic eld is then dened as the two form

B
= d
A
=

j<k
(
xj

k

x
k

j
)dr
j
dr
k
.
The case when : = 2 is particularly simple. In this case,

B
= 1dr
1
dr
2
,
and we can identify
B
with the function
21
r 1(r) = curl (

)(r).
The proof is particularly simple in the case when 1(r) has a constant sign (say
1(r) 0). In this case, we immediately have the inequality
_
R
2
1(r)[n(r)[
2
dr <
A
n, n
L
2 . (7.5.3)
We observe indeed the following identities between operators
1(r) =
1
i
[A
1
, A
2
] ,
A
= A
2
1
+A
2
2
. (7.5.4)
Here
A
1
=
1
i

x1

1
(r) , A
2
=
1
i

x2

2
(r) .
Note also that :

A
n , n) = [[A
1
n[[
2
+[[A
2
n[[
2
,
for all n C

0
(R
2
).
To obtain (7.5.3) is then easy through an integration by parts. One can also
use, introducing 7 = A
1
+iA
2
, the positivity of 7

7 or 77

.
We then easily obtain as in the previous example that the operator is with
compact resolvent if 1(r) +.
As a simple example, one can think of

= (r
2
1
r
2
, +r
1
r
2
2
) ,
which gives
1(r) = r
2
1
+r
2
2
.
Note that the case : = 2 is rather particular and it is more dicult to treat
: 2. We have indeed to introduce partition of unity.
21
Here, for french readers, curl denotes the rotational (in french rotationnel) rot .
57
7.6 Laplace Beltrami operators on a Riemannian compact
manifold
If ` is a compact riemannian manifold, it is well known that in this case one
can canonically dene a measure dj
M
on ` and consequently the Hilbertian
space 1
2
(`). We have also a canonical denition of the gradient. At each
point r of `, we have indeed a scalar product on T
x
` giving an isomorphism
between T
x
` and T

x
`. Using this family of isomorphisms we have a natural
identication between the C

-vector elds on ` and the 1-forms on `. In

this identication, the vector eld gradn associated to a C

function on `
corresponds to the 1-form dn.
Considering on C

(`) C

(`) the sesquilinear form

(n, ) o
0
(n, ) :=
_
M
< gradn(r) , grad(r)
TxM
dj
M
.
There is a natural dierential operator
M
called the Laplace-Beltrami oper-
ator on ` such that
o
0
(n, ) =<
M
n ,
L
2
(M)
.
In this context, it is not dicult to dene the Friedrichs extension and to get a
selfadjoint extension of
M
as a selfadjoint operator on 1
2
(`). The domain
is easily characterized as being H
2
(`), the Sobolev space naturally associated
to 1
2
(`) and one can show that the injection of H
2
(`) into 1
2
(`) is com-
pact because ` is compact. The selfadjoint extension of
M
has compact
resolvent and the general theory can be applied to this example.
The case on the circle S
1
The simplest model is the operator d
2
,d
2
on the circle of radius one whose
spectrum is n
2
, n N. For n 0 the multiplicity is 2. An orthonormal basis
is given by the functions (2)

1
2
expin for n Z. Here the form domain of
the operator is H
1,per
(o
1
) and the domain of the operator is H
2,per
(o
1
). These
spaces have two descriptions. One is to describe these operators as H
1,per
:=
n H
1
(]0, 2[ [ n(0) = n(2) and H
2,per
:= n H
2
(]0, 2[ [ n(0) =
n(2) , n

(0) = n

(2).
The other way is to consider the Fourier coecients of n.
The Fourier coecients of n H
k,per
are in h
k
. Here
h
k
:= n
n
/
2
(Z) [ n
k
n
n
/
2
(Z).
It is then easy to prove the compact injection from H
1,per
in 1
2
(o
1
) or equiva-
lently from h
1
into /
2
.
More generally, elliptic symmetric positive operators of order : 0 admit a
selfadjoint extension with compact resolvent. We refer to the book by Berger-
Gauduchon-Mazet [BGM] for this central subject in Riemannian geometry.
58
The Laplacian on S
2
One can also consider the Laplacian on S
2
. We describe as usual S
2
by the
spherical coordinates, with
r = cos sin, j = sinsin, . = cos , with [, [ , ]0, [ , (7.6.1)
and we add the two poles North and South, corresponding to the two points
(0, 0, 1) and (0, 0, 1).
We are looking for eigenfunctions of the Fiedrichs extension of
L
2
=
1
sin
2

2

1
sin

sin

(7.6.2)
in 1
2
(sind d), satisfying
L
2
Y
m
= /(/ + 1)Y
m
. (7.6.3)
The standard spherical harmonics, corresponding to / 0 and for an integer
: /, . . . , /, are dened by
Y
m
(, ) = c
,m
expi:
1
sin
m

(
1
sin
d
d
)
m
sin
2
, (7.6.4)
where c
,m
is an explicit normalization constant.
For future extensions, we prefer to take this as a denition for : 0 and then
to observe that
Y
,m
= c
,m
Y
,m
. (7.6.5)
For / = 0, we get : = 0 and the constant. For / = 1, we obtain, for : = 1,
the function (, ) sin expi and for : = 1, the function sin expi
and for : = 0 the function cos , which shows that the multiplicity is 3 for the
eigenvalue 2.
To show the completeness it is enough to show that, for given : 0, the
orthogonal family (indexed by / : + N) of functions
,m
() :=
1
sin
m

(
1
sin
d
d
)
m
sin
2
span all 1
2
(]0, [, sind).
For this, we consider C

0
(]0, [) and assume that
_

0
()
,m
() sind = 0 , / :+N .
We would like to deduce that this implies = 0. After a change of variable
t = cos and an integration by parts, we obtain that this problem is equivalent
to the problem to show that, if
_
1
1
(t)((1 t
2
)

)
(m)
(t)dt = 0 , / :+N ,
then = 0.
Observing that the space spanned by the functions (1 t
2
)
m
((1 t
2
)

)
(m)
(which are actually polynomials of exact order /) is the space of all polynomials
we can conclude the completeness.
59
8 Selfadjoint unbounded operators and spectral
theory.
8.1 Introduction
We assume that H is an Hilbert space. Once we have a selfadjoint operator we
can apply the basic spectral decomposition, which we shall now describe without
to give complete proofs. Before to explain the general case, let us come back
to the spectral theorem for compact operators T or operators with compact
resolvent. This will permit us to introduce a new vocabulary.
We have seen that one can obtain a decomposition of H in the form
H =
kN
\
k
, (8.1.1)
such that
Tn
k
=
k
n
k
, if n
k
\
k
. (8.1.2)
Hence we have decomposed H into a direct sum of orthogonal subspaces \
k
in
which the selfadjoint operator T is reduced to multiplication by
k
.
If 1
k
denotes the orthogonal projection operator onto \
k
, we can write
1 =

k
1
k
, (8.1.3)
(the limit is in the strong-convergence sense) and
Tn =

k
1
k
n , n 1(T) . (8.1.4)
Here we recall the denition :
Denition 8.1.1.
An operator 1 L(H) is called an orthogonal projection if 1 = 1

and 1
2
= 1.
If we assume that T is semibounded (with compact resolvent
22
), we can
introduce for any R
G

\
k
, (8.1.5)
and 1

is the orthogonal projection onto G

:
1

1
k
. (8.1.6)
It is easy to see that the function 1

has values in L(H) and satisfy the

following properties :
1

= 1

;
1

= 1
inf(,)
;
22
Note that if (T
0
)
1
is compact for some
0
(T), then it is true for any (T).
60
for all , 1
+0
= 1

;
lim

= 0 , lim
+
1

= 1d .
1

0
All the limits above are in the sense of the strong convergence.
We also observe that
1

k
1

k
0
= 1
k
.
Then in the sense of vectorvalued distributions, we have
d1

k
1
k
, (8.1.7)
where

k
is the Dirac measure at the point
k
. Hence, in the sense of Stieltjes
integrals (this will be explained in more detail below), one can write
r =
_
+

d1

(r) ,
and
T =
_
+

d1

.
This is in this form that we shall generalize the previous formulas to the case
of any selfadjoint operator T.
Functional calculus for operators with compact resolvent :
If ) is a continuous (or piecewise continuous function) one can also dene )(T)
as
)(T) =

k
)(
k
) 1
k
,
as an unbounded operator whose domain is
1()(T)) = r H [

k
[)(
k
)[
2
[[r
k
[[
2
< + ,
where r
k
= 1
k
r.
We can also write )(T) in the form :
< )(T)r, j
H
=
_
R
)() d < 1

r, j ,
where the domain of )(T) is described as
1()(T)) = r H [
_
R
[)()[
2
d < 1

r, r
H
< + .
61
Remark 8.1.2.
There are, for semibounded operators with compact resolvent, two possible con-
ventions for the notation of the eigenvalues. The rst one is to classify them
into an increasing sequence
j
j
j
j+1
counting each eigenvalue according to its multiplicity. The second one is to
describe them as a strictly increasing sequence
k
with
k
eigenvalue of multi-
plicity :
k
.
We now present a list of properties which are easy to verify in this particular
case and which will be still true in the general case.
1. If ) and p coincide on (T), then )(T) = p(T). For any (r, j) H H,
the support of the measure d 1

r, j) is contained
23
in (T).
2. If ) and p are functions on R,
)(T)p(T) = () p)(T) .
In particular, if (T .) is invertible, the inverse is given by )(T) where )
is a continuous function such that )() = ( .)
1
, (T).
3. If ) is bounded, then )(T) is bounded and we have
[[)(T)[[ sup
(T)
[)()[ . (8.1.8)
4. The function ) may be complex. Note that, in this case, we get
)(T)

= )(T) . (8.1.9)
An interesting case is, for . C R, the function ( .)
1
. Then
we get from (8.1.8)
[[(T .)
1
[[
L(H)
[ Im .[
1
. (8.1.10)
5. More generally, this works also for . R (T). We then obtain in this
case the spectral theorem
[[(T .)
1
[[
L(H)
d(., (T))
1
. (8.1.11)
6. If ) C

0
(R), we have :
)(T) =
1

lim
0
+
_
|Imz|>
(

)
.
)(T .)
1
dr.dj . (8.1.12)
23
At the moment, we have not dened (T) when T is unbounded. Think in the case of
operators with compact resolvent of the set of the eigenvalues !
62
Here

) is dened by :

)(r, j) = ()(r) +ij)

(r))(j) ,
where (j) is equal to 1 in a neighborhood of 0 and with compact support.
This formula can be proven using the Green-Formula (rst prove it with
T replaced by the scalar ) or using that :

z
1
.
=
(,0)
,
where
(,0)
is the Dirac measure at (, 0) R
2
and
z
=
1
2
(
x
+ i
y
) .
One should observe that

) is not holomorphic but almost holomorphic
in the sense that :

), . = O(j) ,
as j 0.
Exercise 8.1.3.
Show that, one can also nd, for any 1,

) =

)
N
such that in addition

), . = O(j
N
) ,
as j 0.
8.2 Spectrum.
We now come back to the notion of spectrum that we have only met for bounded
operators.
Denition 8.2.1.
The resolvent set of a closed operator T is the set of the in C such that the
range of (T ) is equal to H and such that (T ) admits a continuous operator
denoted by 1() whose range is included in 1(T) such that :
1()(T ) = 1
D(T)
,
and
(T )1() = 1
H
.
As in the bounded case, we observe that the resolvent set is open. Note also
that the continuity of 1() is actually a consequence of the property that the
graph of 1() is closed (using that T is closed) and that 1() is dened on H.
Denition 8.2.2.
The spectrum of a closed operator T is dened as the complementary set in C
of the resolvent set.
63
It is then rather easy to show that the spectrum (T) is closed in C.
The proof of the fact that the spectrum is contained in R if T is selfadjoint is
very close to the bounded case.
The spectrum (T) is not empty if T is selfadjoint.
The proof is by contradiction. If T has empty spectrum T
1
is a bounded
selfadjoint operator with spectrum equal to 0. We observe indeed that, for
,= 0, the inverse of T
1
is given, if
1
(T), by
1
T(T
1
)
1
.
Hence T
1
should be the 0 operator
24
, which contradicts T T
1
= 1 . This is
no longer true in the non selfadjoint case. At the end of the chapter, we give
some example appearing naturally in various questions in Fluid Mechanics.
8.3 Spectral family and resolution of the identity.
Denition 8.3.1.
A family of orthogonal projectors 1() (or 1

), < < in an Hilbert

space H is called a resolution of the identity (or spectral family) if it satises
the following conditions :

1()1(j) = 1(min(, j)) , (8.3.1)

1() = 0 , 1(+) = 1 , (8.3.2)

where 1() is dened
25
by
1()r = lim

1()r (8.3.3)
for all r in H,

1( + 0) = 1() , (8.3.4)
where 1( + 0) is dened by
1( + 0)r = lim
, >
1(j)r . (8.3.5)
Remark 8.3.2.
We have shown an example of such a family in the previous subsection.
Proposition 8.3.3.
Let 1() be a resolution of identity (=spectral family); then for all r, j H,
the function
< 1()r, j (8.3.6)
24
Here we refer to our analysis of the spectrum of a bounded self-adjoint operator.
25
(8.3.1) gives the existence of the limit (cf also Lemma 8.3.4). The limit in (8.3.3) is taken
in 1. We observe indeed that E()x, x) = [[E()x[[
2
is monotonically increasing.
64
is a function of bounded variation whose total variation
26
satises
\ (r, j) [[r[[ [[j[[ , r, j H . (8.3.7)
Proof.
Let
1
<
2
< <
n
. We rst get from the assumption (8.3.1) that
1
],]
= 1

is an orthogonal projection. From the Cauchy-Schwarz inequality we have

n
j=2
[ < 1
]j1,j]
r, j [ =

n
j=2
[ < 1
]j1,j]
r, 1
]j1,j]
j [

n
j=2
[[1
]j1,j]
r[[ [[1
]j1,j]
j][[

n
j=2
[[1
]j1,j]
r[[
2
_1
2
_

n
j=2
[[1
]j1,j]
j[[
2
_1
2
=
_
[[1
]1,n]
r[[
2
_1
2
_
[[1
]1,n]
j[[
2
_1
2
.
But for : n, we get
[[r[[
2
[[1
]n,m]
r[[
2
=
m1

i=n
[[1
]i,i+1]
r[[
2
. (8.3.8)
We nally obtain that, for any nite sequence
1
<
2
< <
n
, we have
n

j=2
[ < 1
]j1,j]
r, j [ [[r[[ [[j[[ .
This shows the bounded variation and the estimate of the total variation dened
as
\ (r, j) := sup
1, ,n
n

j=2
[ < 1
]j1,j]
r, j [ . (8.3.9)
Hence we have shown that, for all r and j in H, the function 1()r, j)
is with bounded variation and we can then show the existence of 1( + 0) and
1( 0). This is the object of
Lemma 8.3.4.
If 1() is a family of projectors satisfying (8.3.1) and (8.3.2), then, for all
R, the operators
1
+0
= lim
>
1(j) , 1
0
= lim
<
1(j) , (8.3.10)
are well dened when considering the limit for the strong convergence topology.
26
See the denition in (8.3.9)
65
Proof.
Let us show the existence of the left limit. From (8.3.8), we get that, for any
c 0, there exists
0
< such that,

[
0
, [, such that

<

[[1
]

]
r[[
2
c .
It is then easy to show that 1

1
n
r is a Cauchy sequence converging to a limit
and that this limit does not depend on the choice of the sequence tending to .
The proof of the existence of the limit from the right is the same. This ends the
proof of the lemma.
It is then classical (Stieltjes integrals) that one can dene for any continuous
complex valued function )() the integrals
_
b
a
)()d1()r, j) as a limit
27
of Riemann sums.
Proposition 8.3.5.
Let ) be a continuous function on R with complex values and let r H. Then
it is possible to dene for < , the integral
_

)() d1

r
as the strong limit in H of the Riemann sum :

j
)(

j
)(1
j+1
1
j
)r , (8.3.11)
where
=
1
<
2
< <
n
= ,
and

j
]
j
,
j+1
] ,
when max
j
[
j+1

j
[ 0.
The proof is easy using the uniform continuity of ). Note also that the no-
tation could be misleading. May be
_
],]
)()d1

r is less ambiguous.
We now arrive like in the standard theory to the generalized integral.
Denition 8.3.6.
For any given r H and any continuous function ) on R, the integral :
_
+

)()d1

r
is dened as the strong limit in H, if it exists of
_

)()d1

r when
and +.
27
The best is to rst consider the case when x = y and then use a depolarisation formula,
in the same way that, when we have an Hilbertian norm, we can recover the scalar product
from the norm.
66
Remark 8.3.7.
The theory works more generally for any borelian function (cf Reed-Simon,
Vol. 1 [RS-I]). This can be important, because we are in particular interested
in the case when )(t) = 1
],]
(t).
One possibility for the reader who wants to understand how this can be made is
to look at Rudins book [Ru1], which gives the following theorem (Theorem 8.14,
p. 173)
Theorem 8.3.8. .
1. If j is a complex Borel measure on R and if
() )(r) = j(] , r]) , r R ,
then ) is a normalized function with bounded variation (NBV). By NBV
we mean, with bounded variation, but also continuous from the right and
such that lim
x
)(r) = 0.
2. Conversely, to every ) 1\ , there corresponds a unique complex Borel
measure j such that () is satised.
Theorem 8.3.9.
For r given in H and if ) is a complex valued continuous function on R, the
following conditions are equivalent

_
+

)()d1

r exists ; (8.3.12)

_
+

[)()[
2
d[[1

r[[
2
< +; (8.3.13)

j
_
+

)()d(< 1

j, r
H
) (8.3.14)
is a continuous linear form.
Hint for the proof.
a)
(8.3.12) implies (8.3.14) essentially by using repeatedly the Banach-Steinhaus
Theorem (also called Uniform Boundedness Theorem) and the denition of the
integral.
b)
Let us prove that (8.3.14) implies (8.3.13).
67
Let 1 be the linear form appearing in (8.3.14). If we introduce
j =
_

)()d1

r,
then we rst observe (coming back to the Riemann integrals) that
j = 1
],]
j .
It is then not too dicult to show that
1(j) =
_
+

)()d < 1

r, j
=
_
+

)()d < 1

r, 1
],]
j
_
+

)()d < 1
],]
1

r, j
=
_

)()d < 1

r, j
= [[j[[
2
.
Using (8.3.14), we get [[j[[
2
[[1[[ [[j[[ and consequently
[[j[[ [[1[[ . (8.3.15)
Here we observe that the r.h.s. is independent of and .
On the other hand, coming back to Riemann sums, we get
[[j[[
2
=
_

[)()[
2
d[[1

r[[
2
.
We nally obtain
_

[)()[
2
d[[1

r[[
2
[[1[[
2
. (8.3.16)
Hence, taking the limits and +, we obtain (8.3.13).
c)
For the last implication, it is enough to observe that, for

< < <

, we
have
[[
_

)()d1

r
_

)()d1

r[[
2
=
_

[)()[
2
d[[1

r[[
2
+
_

[)()[
2
d[[1

r[[
2
.
Theorem 8.3.10.
Let )() be a real-valued continuous function. Let
1
f
:= r H,
_
+

[)()[
2
d1()r, r) < , .
Then 1
f
is dense in H and we can dene T
f
whose domain is dened by
1(T
f
) = 1
f
,
68
and
T
f
r, j) =
_
+

)()d1()r, j)
for all r in 1(T
f
) and j in H.
The operator T
f
is selfadjoint.
Finally, we have
T
f
1

is an extension of 1

T
f
. (8.3.17)
Proof of the theorem.
From property (8.3.2), we obtain that, for any j in H, there exists a sequence
(
n
,
n
) such that 1
]n,n]
j j as n +.
But 1
],]
j belongs to 1
f
, for any , , and this shows the density of 1
f
in H.
We now observe that ) being real and 1

being symmetric the symmetry is

clear. The selfadjointness is proven by using Theorem 8.3.9.
We observe that, for )
0
= 1, we have T
f0
= 1 and for )
1
() = , we obtain
a selfadjoint operator T
f1
:= T.
In this case we say that
T =
_
+

d1()
is a spectral decomposition of T and we shall note that
[[Tr[[
2
=
_
+

2
d 1()r, r) =
_
+

2
d [[1()r[[
2
,
for r 1(T).
More generally
[[T
f
r[[
2
=
_
+

[)()[
2
d((1()r, r)) =
_
+

[)()[
2
d([[1()r[[
2
)
for r 1(T
f
).
Conclusion. We have consequently seen in this subsection how one can as-
sociate to a spectral family of projectors a selfadjoint operator. We have seen
in the introduction that the converse was true for a compact operator or an
operator with compact resolvent. It remains to prove that this is true in the
general case.
8.4 The spectral decomposition Theorem.
The spectral decomposition Theorem makes explicit that the preceding situation
is actually the general one.
69
Theorem 8.4.1.
Any selfadjoint operator T in an Hilbert space H admits a spectral decomposition
such that
< Tr, j =
_
R
d < 1

r, j
H
,
Tr =
_
R
d(1

r) .
(8.4.1)
Proof.
We shall only give the main points of the proof. We refer to [Hu], [Le-Br] or
[DaLi] for detailed proofs or to [RS-I] for another proof which we describe now
shortly. Another interesting proof is based on Formula 8.1.12 and presented in
the book of Davies [Da].
Step 1.
It is rather natural to imagine that it is essentially enough to treat the case when
T is a bounded selfadjoint operator (or at least a normal bounded operator, that
is satisfying T

T = TT

). If is indeed a general semibounded selfadjoint op-

erator, one can come back to the bounded case by considering (+
0
)
1
, with

0
real, which is bounded and selfadjoint. In the general case
28
, one can con-
sider (+i)
1
.
Step 2.
We analyze rst the spectrum of 1(T) where 1 is a polynomial.
Lemma 8.4.2. .
If 1 is a polynomial, then
(1(T)) = 1() [ (T) (8.4.2)
Proof.
We start from the identity 1(r) 1() = (r )Q

(r) and from the corre-

sponding identity between bounded operators 1(T) 1() = (T )Q

(T).
This permits to construct the inverse of (T ) if one knows the inverse of
1(T) 1().
Conversely, we observe that, if . C and if
j
(.) are the roots of
(1() .), then we can write :
(1(T) .) = c

j
(T
j
(.)) .
This permits to construct the inverse of (1(T) .) if one has the inverses of
(T
j
(.)) (for all ,).
Lemma 8.4.3.
Let T be a bounded self-adjoint operator. Then
[[1(T)[[ = sup
(T)
[1()[ . (8.4.3)
28
Here we recall that an example of operator which is not semibounded is given in Exer-
cise 4.1.3
70
We rst observe that
[[1(T)[[
2
= [[1(T)

1(T)[[
This is the consequence of the general property for bounded linear operators
that :
[[

[[ = [[[[
2
.
We recall that the proof is obtained by observing rst that :
[[

[[ = sup
||x||1,||y||1
[

r, j)[
= sup
x,y ||x||1,||y||1
[r, j)
[[[[
2
,
and secondly that :
[[[[
2
= sup
||x||1
r, r) = sup
||x||1

r, r) [[

[[ .
We then observe that :
[[1(T)[[
2
= [[(

11)(T)[[
= sup
(

PP)(T)
[j[ (using Theorem 6.4.1)
= sup
(T)
[(

11)()[ (using Lemma 8.4.2)
=
_
sup
(T)
[1()[
2
_
.
Step 3.
We have dened a map from the set of polynomials into L(H) dened by
1 (1) = 1(T) , (8.4.4)
which is continuous
[[(1)[[
L(H)
= sup
(T)
[1()[ . (8.4.5)
The set (T) is a compact in R and using the Stone-Weierstrass theorem (which
states the density of the polynomials in C
0
((T))), the map can be uniquely
extended to C
0
((T)). We still denote by this extension. The properties of
are described in the following theorem
Theorem 8.4.4.
Let T be as selfadjoint continuous operator on H. Then there exists a unique
map from C
0
((T)) into L(H) with the following properties :
1.
() +p) = ()) + (p) , ()) = ()) ;
(1) = 1d , (

)) = ())

;
()p) = ()) (p) .
2.
[[())[[ = sup
(T)
[)()[ .
71
3. If ) is dened by )() = , then ()) = T.
4.
(())) = )() [ (T) .
5. If satises T = , then ()) = )().
6. If ) 0, then ()) 0.
All these properties are obtained by showing rst the properties for polyno-
mials 1 and then extending the properties by continuity to continuous functions.
For the last item, note that :
()) = (
_
)) (
_
)) = (
_
))

(
_
)) .
Step 4.
We are now ready to introduce the measures. Let H. Then
) < , )(T)
H
=< , ())
H
is a positive linear functional on C
0
((T)). By measure theory (Riesz Theorem)
(cf Rudin [Ru1]), there exists a unique measure j

on (T), such that

< , )(T)
H
=
_
(T)
)()dj

() . (8.4.6)
This measure is called the spectral measure associated with the vector H.
This measure is a Borel measure. This means that we can extend the map
and (8.4.6) to Borelian functions.
Using the standard Hilbert calculus (that is the link between sesquilinear form
and quadratic forms) we can also construct for any r and j in H a complex
measure dj
x,y
such that
< r, ())j
H
=
_
(T)
)()dj
x,y
() . (8.4.7)
Using the Riesz representation Theorem (Theorem 3.1.1), this gives as, when ) is
bounded, an operator )(T). If ) = 1
],]
, we recover the operator 1

= )(T)
which permits to construct indeed the spectral family announced in Theorem
8.4.1.
Remark 8.4.5.
Modulo some care concerning the domains of the operator, the properties men-
tioned at the rst subsection of this section for operators with compact resolvent
are preserved in the case of an unbounded selfadjoint operator.
72
8.5 Applications of the spectral theorem:
One of the rst applications of the spectral theorem (Property 2.) is the follow-
ing property :
Proposition 8.5.1.
d(, (T))[[r[[ [[(T )r[[ , (8.5.1)
for all r in 1(T).
This proposition is frequently used in the following context. Except very
special cases like the harmonic oscillator, it is usually dicult to get explicitely
the values of the eigenvalues of an operator. One consequently tries to localize
these eigenvalues by using approximations. Let us suppose for example that one
has found
0
and j in 1(1) such that
[[(T
0
)j[[ c (8.5.2)
and [[j[[ = 1 then we deduce the existence of in the spectrum of 1 such that
[
0
[ c.
Standard examples are the case of hermitian matrices or the case of the anhar-
monic oscillator T := /
2 d
2
dx
2
+r
2
+r
4
. In the second case the rst eigenfunction
of the harmonic oscillator /
2 d
2
dx
2
+ r
2
can be used as approximate eigenfunc-
tion j in (8.5.2) with
0
= /. We then nd (8.5.2) with c = O(/
2
).
Another application is, using the property that the spectrum is real, the follow-
ing inequality
[ Im [ [[r[[ [[(T )r[[ (8.5.3)
and this gives an upper bound on the norm of (T )
1
in L(H) by 1,[ Im [.
One can also consider the operator T

= d
2
,dr
2
+ r
2
+ cr
4
. One can show
that near each eigenvalue of the harmonic oscillator (2n +1), then there exists,
when c 0 is small enough, an eigenvalue
n
(c) of T

.
Another good example to analyze is the construction of a sequence of approxi-
mate eigenfunctions considered in Subsection 1.1. From the construction of n
n
such that, with T = ,
[[(T
2
)n
n
[[
L
2
(R
m
)
= O(
1
n
) ,
one obtains that
d((T),
2
)
C
n
, n N .
As n +, we get
2
(T).
It is then easy to show that
(1) = [0, +[.
It is enough to prove indeed, using the Fourier transform, that, for any / 0,
( +/) has an inverse ( +/)
1
sending 1
2
onto H
2
.
Here we have followed in a particular case the proof of the following general
theorem
73
Theorem 8.5.2.
Let T be a selfadjoint operator. Then (1) if and only if there exists a
sequence (n
n
)
nN
, n
n
1(T) such that [[n
n
[[ = 1 and [[(T )n
n
[[ 0 as
n +.
Exercise 8.5.3.
Show the only if , after reading of the proof of Proposition 8.5.4.
Note also the following characterization.
Proposition 8.5.4.
(T) = R, :.t. c 0, 1(] c, +c[) ,= 0 (8.5.4)
The proof uses in one direction the explicit construction of (T )
1
through
Proposition 8.5.1. If and c
0
0 are such 1
]0,+0[
= 0. Then there exists
a continuous function ) on R, such that )(t) = (t )
1
on the support of the
measure d1

. This permits to construct the inverse and to show that is in

the resolvent set of T.
Conversely, let in the set dened by the r.h.s of (8.5.4) later denoted by (T).
For any n N

, let us take r
n
such that [[r
n
[[ = 1 and 1(]
1
n
, +
1
n
[)r
n
=
r
n
. Using property 2. of Theorem 8.4.4, with the function t )
n
(t) = (t
)1
]
1
n
,+
1
n
[
(t), we get :
[[(T )r
n
[[ = [[(T )1(]
1
n
, +
1
n
[)r
n
[[
1
n
[[r
n
[[ =
1
n
.
Applying Proposition 8.5.1, we obtain :
(T) (T) .
8.6 Examples of functions of a selfadjoint operator
We shall for example meet usually in spectral theory the functions
1. ) is the characteristic function of ] , ], Y
],]
; ()) = )(T) is then
()) = 1().
2. ) is the characteristic function of ] , [, Y
],[
; )(T) is then ()) =
1( 0).
3. ) is a compactly supported continuous function. )(T) will be an operator
whose spectrum is localized in the support of ).
4. )
t
() = exp(it) with t real.
)
t
(T) is then a solution of the functional equation
(
t
iT)()(t, T)) = 0
)(0, T) = 1d
We note here that, for all real t, )
t
(T) = exp(itT) is a bounded unitary
operator .
74
5. p
t
() = exp(t) with t real positive. p
t
(T) is then a solution of the
functional equation
(
t
+T)(p(t, T)) = 0 , for t 0
p(0, T) = 1d .
We have discussed in the introduction the case of an operator with compact
resolvent. The other case to understand for the beginner is of course the case
of the free Laplacian on R
n
. Using the Fourier transform T, we get as
unbounded operator the operator of multiplication by
2
. It is not dicult
to dene directly the functional calculus which simply becomes for a borelian
function :
() = T
1
(
2
)T . (8.6.1)
One possibility is to start from ( + 1)
1
, for which this formula is true and
to then follow what was our construction of the functional calculus. Another
possibility is to use the Formula (8.1.12) and to use that (8.6.1) is satised for
( +.)
1
, with . C R.
The spectral family is then dened by
< 1()), p
L
2
(R
m
)
=
_

)() p() d .
8.7 Spectrum and spectral measures
Another interest of the spectral theorem is to permit the study of the dierent
properties of the spectrum according to the nature of the spectral measure and
this leads to the denition of the continuous spectrum and of the pure point
spectrum. Let us briey discuss (without proof) these notions.
Starting of a selfadjoint operator T, one denes H
pp
(pure point subspace) as
the set dened as
H
pp
= H [ ) < )(T),
H
is a pure point measure (8.7.1)
We recall that a measure on A is pure point if
j(A) =

xX
j(r) . (8.7.2)
One can verify that H
pp
is a closed subspace of H and that the corresponding
orthogonal projection
Hpp
satises

Hpp
1(T) 1(T) .
In this case T
/Hpp
is naturally dened as unbounded operator on H
pp
and one
denes the pure point spectrum of T by

pp
(T) = (T
/Hpp
) . (8.7.3)
75
We can similarly dene H
c
(continuous subspace) as the set dened as
H
c
= H [ ) < )(T),
H
is a continuous measure (8.7.4)
We recall that a measure on A is continuous if
j(r)) = 0 , r A . (8.7.5)
One can verify that H
c
is a closed subspace of H and that the corresponding
orthogonal projection
Hc
satises

Hc
1(T) 1(T) .
Moreover, it can be shown (See [RS-I]), that
H = H
pp
H
c
. (8.7.6)
In this case T
/Hc
is naturally dened as unbounded operator on H
c
and one
denes the continuous spectrum of T by

c
(T) = (T
/Hc
) . (8.7.7)
Example 8.7.1.
The spectrum of is continuous.
We observe indeed that
lim
0 , >0
_
|||
2
|
[

)()[
2
d = 0 , ) 1
2
(R
m
) .
One can still rene this discussion by using the natural decomposition of the
measure given by the Radon-Nikodym Theorem. This leads to the notion of
absolutely continuous spectrum and of singularly continuous spectrum.
76
9 Non-self adjoint operators and -pseudospectrum
When the operators are not selfadjoint, one should think that the spectrum is
not the right object because it becomes very unstable by perturbation. It has
been realized in the recent years that a family of sets (parametrized by c 0)
in C called the c-pseudospectrum is the right object for getting this stability.
9.1 Main denitions and properties
Here we follow Chapter 4 in the book by L.N. Trefethen and M. Embree [TrEm].
Denition 9.1.1.
If is a closed operator with dense domain 1() in an Hilbert space H, the
c-pseudospectrum

() of is dened by

() := . C [ [[(.1 )
1
[[
1
c
.
Remark 9.1.2.
In one part of the literature, is replaced by in the above denition. We
have chosen the denition which leads to the simplest equivalent denitions.
We will be interested in this notion in the limit c 0.
We take the convention that [[(.1 )
1
[[ = + if . (), where ()
denotes the spectrum of , so it is clear that we always have :
()

() .
When is selfadjoint (or more generally normal),

() is, by the Spectral

Theorem, given by,

() = . C [ d(., ()) < c .

So this is only in the case of non self-adjoint operators that this new concept (rst
appearing in numerical analysis, see Trefethen [Tr1, Tr2]) becomes interesting.
Although formulated in a rather abstract way, the following (weak version of) a
result by Roch-Silbermann [RoSi] explains rather well to what corresponds the
c-pseudospectrum :
Proposition 9.1.3.

() =
_
{AL(H) s. t. ||A||
L(H)
<}
(+) .
In other words, . is in the c-pseudospectrum of if . is in the spectrum of
some perturbation + of with [[[[ < c. This is indeed a natural notion
thinking of the fact that the models we are analyzing are only approximations
of the real problem and of the fact that the numerical analysis of the model goes
through the analysis of explicitly computable approximated problems. Numer-
ical examples are treated in [Tr2].
77
Proof
Let us rst show the easy part of this characterization of the c-pseudospectrum.
If [[(. )
1
[[
1

, it is clear that for any such that [[[[ < c, (+.)

is invertible. Its inverse is obtined by observing that
(. )
1
(. ) = 1 (. )
1
.
But the left hand side is invertible because
[[(. )
1
[[ [[(. )
1
[[ [[[[ < 1 .
The inverse is consequently given by
(. )
1
=
_
_

j
((. )
1
)
j
_
_
(. )
1
.
The converse is not very dicult. If [[(. )
1
[[
1

, by denition of the
norm, there exists n H such that [[n[[ = 1 and
[[(. )
1
n[[ = j
1
c
.
Let = (. )
1
n. Let () the linear bounded operator such that
()r = j
2
n [r) , r H.
It is clear that is an eigenfunction of + associated to . and that [[[[ =
1

< c.
Hence we have found a perturbation + of such that . (+) and
[[[[ < c.
Another presentation for dening the c-pseudospectrum is to say that .

() if and only if :
either . () or if there exists an c-pseudoeigenfunction that is an n 1()
such that [[n[[ = 1 and [[(. )n[[ < c.
Theorem 9.1.4 (c-Pseudospectrum of the adjoint).
For any closed densely dened operator and any c 0, we have

) =

() ,
where for a subset in C we denote by the set
= . C[ . .
Proof
This is immediate using that, if . , (),
|(. )
1
| = |( .

)
1
| .
78
9.2 -Pseudospectrum : complete analysis of the dieren-
tiation operator.
We consider the operator dened on 1
2
(]0, 1[) by
1() = n H
1
(]0, 1[) , n(1) = 0 ,
and
n = n

, n 1() .
This is clearly a closed operator with dense domain.
The adjoint of is dened on 1
2
(]0, 1[) by
1(

) = n H
1
(]0, 1[) , n(0) = 0 ,
and
n = n

, n 1(

) .
Lemma 9.2.1.
() = and has compact resolvent.
First we can observe that (.) is injective on 1() for any . C. Then
one easily veries that for any . C, the inverse is given by
[(. )
1
)](r) =
_
1
x
exp.(r :) )(:) d: .
It is also clear that this operator is compact.
To analyze the c-pseudospectrum is more interesting. For this we need to
estimate
(.) := |(. )
1
| .
The rst remark is that depends only on Re .. For this we can observe that
the map n expirn is a unitary transform on 1
2
(]0, 1[), which maps 1()
onto 1().
The main result is the following
Theorem 9.2.2.
The function is a subharmonic function which depends only of Re . and
satises
(.)
1
Re .
, for Re . 0 , (9.2.1)
and
(.) =
exp Re .
2 Re .
+O(
1
[ Re .[
) , for Re . < 0 . (9.2.2)
This implies for the c-pseudospectrum of :
79
Corollary 9.2.3.
For c 0, the c-pseudospectrum of is an hyperplane of the form

() = . C[ 1c. < c

(9.2.3)
with
c

_
(lnc) as c , 0
c as c +.
(9.2.4)
Rough estimate from below for Re . < 0
A rst (non optimal) step is to think
29
semi-classical. Let us take . real
and let us consider
r
z
(r) := [2.[
1
2
exp.r.
This function is not in Ker (. ) because it does not satises the boundary
condition. But when . , the boundary condition at r = 1 is almost
satised. Actually
z
lives very close to 0. Moreover, for . < 0,
|
z
|
2
= 1 exp2.
Hence the norm tends to 1 as . .
Let us indeed consider for 0 a cut-o function

such that

= 1 on
[0, 1 ] and = 0 on [1

2
, 1] and let us introduce

z,
(r) =

z
.
We now observe that
z,
1() and that
(. )
z,
(r) =

z
.
The 1
2
norm of the right hand side is exponentially small like exp(1 )..
This shows that, for any 0 there exists .

< 0 such that

1
C

exp(1 ). (.) , for . < .

. (9.2.5)
This is not as good as in the statement of the theorem but this suggests a rather
general point of view. We will complete later the analysis of the behavior of
as Re . .
Rough estimate from above for Re . 0
Here we will try to estimate (.) from above by using an a priori estimate for
(.) (with . real).
For n 1(), we have
(.)n[ n) = .[[n[[
2
+
_
1
0
n

(t) n(t)dt .
29
The semi-classical parameter is h =
1
Re z
80
But by an integration by parts, we observe that
_
1
0
n

(t) n(t) dt =
_
1
0
n(t) n

(t)dt [n(0)[
2
.
Coming back to the previous equality and taking the real part, we obtain
Re (.)n[ n) = .|n|
2
+
1
2
[n(0)[
2
.|n|
2
.
Then we obtain
[[(.)n[[ .[[n[[ , n 1() , (9.2.6)
which implies (9.2.1).
Control of the resolvent using Schurs Lemma
The operator (.)
1
being an operator associated to an integral kernel, one
can analyze what is given by Schurs lemma or by the Hilbert-Schmidt criterion.
The kernel is dened by
1(r, :) =
_
0 for : < r
exp.(r :) for r < :
According to Schurs Lemma, we have to consider sup
x
_
1(r, :)d: and sup
s
_
1(r, :)dr.
If we are interested with the Hilbert-Schmidt norm, we have to compute
_ _
1(r, :)
2
drd:.
All these computations can be done rather explicitely !
For . ,= 0, we have
sup
x
_
1(r, :)d: =
1
.
(1 exp.) ,
and
sup
s
_
1(r, :)dr =
1
.
(1 exp.)
This gives
(.)
1
.
(1 exp.) . (9.2.7)
This is actually an improved version of (9.2.1) for . 0 and for . < 0, it is
better to write it in the form
(.)
1
.
(exp. 1) , (9.2.8)
and to compare it to the lower bound obtained in (9.2.5).
81
A more accurate estimate for . < 0
We can rewrite (. )
1
in the form
(. )
1
= 1
1
1
2
,
with
1
1
(r) :=
_
1
0
exp.(r :) (:), d:
and
1
2
(r) :=
_
x
0
exp.(r :) (:), d:
Observing that 1

2
can be treated as for the proof of (9.2.7), we rst obtain
[[1
2
[[
1
Re .
.
It remains to control [[1
1
[[. This norm can be computed explicitly. We have
indeed
[[1
1
[[ = [[ exp.r[[ [
_
1
0
exp.: (:) d:[
Hence we have just to compute the norm of the linear form

_
1
0
exp.: (:) d:
which is the 1
2
-norm of : exp.:.
This gives
[[1
1
[[ = [[ exp.r[[[[ exp.r[[ =
1
2.
_
(1 c
2z
)(c
2z
1) =
1
2.
c
z
(1c
2z
) .
Combining the estimates of [[1
1
[[ and [[1
2
[[ leads to (9.2.2).
Remark 9.2.4.
One can discretize the preceding problem by considering, for n N

, the matrix

n
= n
1
with
1
= 1 +J where J is the n n matrix such that J
i,j
=
i+1,j
.
One can observe that the spectrum of
n
is n. It is also interesting to analyze
the pseudospectrum.
Remark 9.2.5.
There is a semi-classical version of the pseudospectrum for families
h
. One
can then relate the c appearing in the denition of the c-pseudospectrum with the
parameter / (which could typically be in ]0, /
0
]). For example, we can consider
c(/) = /
N
.
Exercise 9.2.6.
Analyze the pseudospectrum of
d
d
+p() on the circle.
82
9.3 Another example of non selfadjoint operator without
spectrum
We consider the spectrum of the operator
=
d
dr
+r
2
,
on the line.
We can take as domain 1() the space of the n 1
2
(R) such that n 1
2
(R).
We note that C

0
(R) is dense for the graph norm in 1(). Hence is the
closed extension of the dierential operator
d
dx
+ r
2
with domain C

0
(R). We
note that the operator is not selfadjoint. The adjoint is
d
dx
+r
2
.
The two following inequalities can be useful.
Re n[ n) [[rn[[
2
0 . (9.3.1)
This inequality is rst proved for n C

0
(R) and then extended to n 1()
using the density of C

0
(R). A rst consequence is that
[[rn[[
2
[[n[[ [[n[[
1
2
_
[[n[[
2
+[[n[[
2
_
. (9.3.2)
This implies that 1() (with graph norm) has continuous injection in the
weighted space 1
2

with (r) = [r[. Together with the fact that 1() H

1
loc
(R),
this implies that (+1) is invertible and that the inverse is compact.
The second inequality is
[[r
2
n[[
2
+[[n

[[
2
C
_
[[n[[
2
+[[n[[
2
_
. (9.3.3)
To prove it, we observe that
[[n[[
2
= [[n

[[
2
+[[r
2
n[[
2
+ 2 Re n

, r
2
n) .
Using an integration by parts, we get
2 Re n

, r
2
n) = 2rn, n) .
Hence we get (using Cauchy-Schwarz)
[[n[[
2
[[n

[[
2
+[[r
2
n[[
2
2[[rn[[ [[n[[ .
One can then use (9.3.2) to get the conclusion.
(9.3.3) permits to obtain that
1() = n H
1
(R) , r
2
n 1
2
(R) , (9.3.4)
which is not obvious at all.
83
Proposition 9.3.1.
has an inverse, and the inverse is compact. Moreover its spectrum is empty.
We consider on R, the dierential equation
n

+r
2
n = ) . (9.3.5)
For all ) 1
2
(R), let us show that there exists a unique solution in 1
2
(R) de
(9.3.5). An elementary calculus gives
n(r) = exp
1
3
r
3
_
x

exp
1
3
j
3
)(j) dj .
One has to work a little for showing that n 1
2
(this is easier if ) is compactly
supported). If we denote by K the operator which associates to ) the solution
n, the distribution kernel of K is given by
1(r, j) =
_
0 if j r
exp
1
3
(j
3
r
3
) if j < r
.
We note that if there exists an eigenvalue of K ,= 0 and if n

is a corresponding
eigenfunction, then n

satises
n

+r
2
n

=
1

. (9.3.6)
From this we deduce that K has no non zero eigenvalue. One can indeed solve
explicitly (9.3.6) :
n

(r) = C exp
r
3
3
exp
1

r .
It is then easy to see that no one can be in 1
2
(R) when C ,= 0.
To show that K is compact, we can actually show that K is Hilbert-Schmidt.
That is, we will show that 1(r, j) is in 1
2
(R
2
). We have
_
y<x
exp
2
3
(j
3
r
3
) drdj =
_
+

__
x

exp
2
3
(j
3
r
3
) dj
_
dr .
Dividing the domain of integration in two parts, we rst consider :
_
+1
1
__
x

exp
2
3
(j
3
r
3
) dj
_
dr ,
which is bounded from above by
2c
_
1

exp
2
3
j
3
dj .
Then we look at
_
|x|>1
__
x

exp
2
3
(j
3
r
3
) dj
_
dr .
84
Here we observe that
exp
2
3
(j
3
r
3
) = exp
2
3
(j r)(j
2
+jr +r
2
) ,
and that
(j
2
+jr +r
2
)
1
2
(j
2
+r
2
)
1
2
r
2
.
This leads, as j r, to the upper bound la majoration
exp
2
3
(j
3
r
3
) exp
1
3
(j r)r
2
.
and to
_
|x|>1
__
x

exp
2
3
(j
3
r
3
) dj
_
dr 3
_
|x|>1
1
r
2
dr < +.
This implies that the spectrum of K is 0.
Non self-adjoint eects.
We can also try to estimate the solution operator K

corresponding to the
equation
n

+r
2
n = ) . (9.3.7)
It is easy, to see that we can reduce the computation to the case when reel.
We have indeed
K

= expi Im r K
Re
expi Im r .
For < 0, we easily nd, observing that n S(R) and
Re n

+r
2
n [ n)
L
2 = r
2
n [ n) [[n[[
2
,
the estimate
[[K

[[
L(L
2
)

1

.
The case = 0 has been treated before. Without to much eort, we get the
same result for [ Re [ 1.
So we have to consider the case when 1 and control the estimate as +.
Proceeding as in the case = 0, we rst obtain
1

(r, j) =
_
0 si j r
exp
_
1
3
(j
3
r
3
) (j r)
_
si j < r
. (9.3.8)
Again, we see that 1

(r, j) is in 1
2
(R
2
) :
[[1

[[
2
L
2
(R
2
)
=
_
+

__
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr .
85
The proof is similar to the case = 0. We cut the domain of integration.
_
1
14

__
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr ,
is bounded after a new partition of the domain of integration by
4

c
2
3
(1+4
1
2 )
3
_
14

exp
_
2
3
j
3
2j
_
dj 3

c
2
3
(1+2
1
2 )
3
,
and by
_
1
14

_
_
x
14

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr
4

sup
x]14

,1[
_
_
x
14

exp2(j r) dj
_
2

1
2
c
8
3
2
.
Then we consider
_
+1
1
__
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr ,
which is bounded by
2c
2
3
+2
_
14

exp(
2
3
j
3
2j) dj
2c
2
3
+2
Cc
2
3
+2
.
and by
_
+1
1
_
_
x
14

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr
_
+1
1
_
_
x
14

exp2(j r) dj
_
dr

exp(4(1 + 2

))
We look now at
_
1+4

1
__
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr ,
which is similarly controlled by
4

exp2(1 + 4

)
_
14

exp
_
2
3
j
3
2j
_
dj
and by
_
1+4

1
__
x
14

exp2(j r) dj
_
dr .
86
Finally, we have to estimate
_
|x|1+4

__
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr ,
Here we observe that
exp
2
3
(j
3
r
3
) = exp
2
3
(j r)(j
2
+jr +r
2
) ,
and that
(j
2
+jr +r
2
)
1
2
(j
2
+r
2
)
1
2
r
2
.
This leads to the upper bound
exp
2
3
(j
3
r
3
) exp2(j r) exp
1
3
(j r)(r
2
6) .
By simple integration, we get :
_
|x|>1+4

_
_
x

exp
_
2
3
(j
3
r
3
) 2(j r)
_
dj
_
dr
3
_
|x|>1+4

1
x
2
6
dr < +.
The last bound can be controlled independently of Re . Hence, we have nally
proved the existence of C 0 such that, for Re 1, we have
[[K

[[
HS
C[[
C
expC Re
3
2
.
Remark 9.3.2.
Using the Laplace integral method, one can get the asymptotics of [[K

[[
HS
.
Note that [[K

[[
HS
[[K

[[
L(L
2
)
.
One can also nd a lower bound of [[K

[[
L(L
2
)
using quasimodes.
Note that K

is the resolvent of the unbounded operator

d
dx
+r
2
. This operator
is not selfadjoint. It has compact resolvent and empty spectrum. Moreover,
the norm of the resolvent depends only of Re .
Remark 9.3.3.
Using the Fourier transform, on can see that the operator is isospectral to the
complex Airy operator :
1
2
x
+ir, (9.3.9)
This will be analyzed in the next subsection.
9.4 The non selfadjoint harmonic oscillator
Other interesting example to analyze in the same spirit is the complex harmonic
oscillator
H
1
:=
d
2
dr
2
+ir
2
87
(See Davies (1999), [Tr2], [Zw] and references therein). The spectrum can be
seen as the spectrum of
d
2
dx
2
+r
2
rotated by

4
in the complex plane.
Let us sketch how we can guess the result.
If we make a dilation r = j, the operator
d
2
dx
2
+ ir
2
becomes the operator

2 d
2
dy
2
+ij
2
which is unitary equivalent.
We can now consider the family of operator
H

:=
2
d
2
dr
2
+i
2
r
2
,
for in a sector in C.
It can be shown that for in (an open neighborhood of) the sector arg [

8
, 0]
the spectrum is discrete and independent of (this is a mixture of Katos theory
and of the so called Combes-Thomas argument).
Taking = expi

8
, we get
H

8
= expi

4
(
d
2
dr
2
+r
2
) .
Hence :
(H
1
) = expi

4
(2n + 1) , n N . (9.4.1)
Let us give rigorously a part of an alternative proof. If we start from the
basis of eigenfunctions of the standard harmonic oscillator n
n
. It is easy to
see that n
n
(
1
expi

8
r) is an eigenfunction of H

in 1
2
(R). Hence we have
proven one inclusion in (9.4.1).
What is less clear is to show that the family r n
n
(
1
expi

8
r) which is no
more orthogonal generates a dense subspace in 1
2
. This is discussed in Davies
paper.
Remark 9.4.1.
This example shows that when the operators are not selfadjoint many things can
occur. In particular the fact that if . (T), then [[(T .)
1
[[ is controlled by
1
d(,(T)
becomes wrong.
9.5 The complex Airy operator in R
This operator can be dened as the closed extension / of the dierential opera-
tor on C

0
(R) /
+
0
:= 1
2
x
+i r. We observe that / = (/

0
)

with /

0
:= 1
2
x
i r
and that its domain is
1(/) = n H
2
(R) , rn 1
2
(R) .
In particular / has compact resolvent.
It is also easy to see that
Re /n[ n) 0 . (9.5.1)
88
Figure 1: Davies operator: pseudospectra
0 20 40 60 80 100 120 140 160
20
0
20
40
60
80
100
120
dim = 100
7
6
5
4
3
2
1
0
Hence / is the generator of a semi-group o
t
of contraction,
o
t
= expt/. (9.5.2)
Hence all the results of this theory can be applied.
In particular, we have, for Re < 0
[[(/)
1
[[
1
[ Re [
. (9.5.3)
One can also show that the operator is maximally accretive.
A very special property of this operator is that, for any o R,
T
a
/ = (/io)T
a
, (9.5.4)
where T
a
is the translation operator (T
a
n)(r) = n(r o) .
As immediate consequence, we obtain that the spectrum is empty and that the
resolvent of /, which is dened for any C satises
[[(/)
1
[[ = [[(/ Re )
1
[[ . (9.5.5)
The most interesting property is the control of the resolvent for Re 0.
Proposition 9.5.1.
There exist two positive constants C
1
and C
2
, such that
C
1
[ Re [

1
4
exp
4
3
Re
3
2
[[(/)
1
[[ C
2
[ Re [

1
4
exp
4
3
Re
3
2
, (9.5.6)
89
The proof of the (rather standard) upper bound is based on the direct anal-
ysis of the semi-group in the Fourier representation. We note indeed that
T(1
2
x
+i r)T
1
=
2
+
d
d
. (9.5.7)
Then we have
To
t
T
1
= exp(
2
t +t
2

t
3
3
)( t) , (9.5.8)
and this implies immediately
[[o
t
[[ = expmax

(
2
t +t
2

t
3
3
) = exp(
t
3
12
) . (9.5.9)
Then one can get an estimate of the resolvent by using, for C, the formula
(/)
1
=
_
+
0
expt(/) dt . (9.5.10)
For a closed accretive operator, (9.5.10) is standard when Re < 0, but es-
timate (9.5.9) on o
t
gives immediately an holomorphic extension of the right
hand side to the whole space giving for 0 the estimate
[[(/)
1
[[
_
+
0
exp(t
t
3
12
) dt . (9.5.11)
The asymptotic behavior as + of this integral is immediately obtained
by using the Laplace method and the dilation t =
1
2
: in the integral.
In our case, everything can be computed explicitly. We observe that :
_
+
0
exp(
2
t +t
2

t
3
3
)c
t
( t)dt =
_

exp(
:
3

3
3
+( :):) (:) d: ,
which gives eectively the expression of (/)
1
as an operator with integral
kernel given in (9.3.8).
The proof of the lower bound is obtained by constructing quasimodes for the
operator (/ ) in its Fourier representation. We observe (assuming 0),
that
n(; ) := exp
_

3
3
+
2
3

3
2
_
(9.5.12)
is a solution of
(
d
d
+
2
)n(; ) = 0 . (9.5.13)
Multiplying n(; ) by a cut-o function

with support in ]

, +[ and

= 1 on ]

+ 1, +[, we obtain a very good quasimode, concentrated as

+, around

, with an error term giving almost
30
the announced lower
bound for the resolvent.
Of course this is a very special case of a result on the pseudo-spectra but this
leads to an almost optimal result.
30
One should indeed improve the cut-o for getting an optimal result
90
10 Essentially selfadjoint operators
10.1 Introduction
In most of the examples which were presented, the abstract operators are as-
sociated with dierential operators. These dierential operators are naturally
dened on C

0
() or T

(). Most of the time (for suitable potentials increas-

ing slowly at ) they are also dened (when = R
m
) on o(R
m
) or o

(R
m
).
It is important to understand how the abstract point of view can be related
to the PDE point of view. The theory of the essential selfadjointness gives
a clear understanding of the problem. The question is to decide if, starting
from a symmetric operator T, whose domain 1(T) = H
0
is dense in H, there
exists a unique selfadjoint extension T
ext
of T. We recall that it means that
1(T) 1(T
ext
) and T
ext
n = Tn , n 1(T). This leads to
Denition 10.1.1.
A symmetric operator T with domain H
0
is called essentially selfadjoint if its
closure is selfadjoint.
Proposition 10.1.2.
If T is essentially selfadjoint, then its selfadjoint extension
31
is unique.
Indeed suppose that o is a selfadjoint extension of T. Then o is closed and
being an extension of T, is also an extension of its smallest extension T. We
recall from Theorem 2.2.6 that T = T

. Thus, o = o

(T

= T

, and so
o = T

.
Example 10.1.3.
Here we give a list of examples and counter-examples which will be analyzed
later.
1. The dierential operator with domain H
0
= C

0
(R
m
) is essentially
selfadjoint (see later).
2. The dierential operators + [r[
2
with domain H
0
= C

0
(R
m
) or
H
1
= o(R
m
) are essentially selfadjoint and admit consequently a unique
selfadjoint extension. This extension is the same for the two operators.
The domain can be explicitely described as
1
2
(R
m
) =
n 1
2
(R
m
) [ r

x
n 1
2
(R
m
) , , N
m
with[[ +[[ 2 .
3. The dierential operator with domain C

0
() (where is an open
bounded set with smooth boundary) is not essentially selfadjoint. There
exists a lot of selfadjoint extensions related to the choice of a boundary
problem. As we have seen before, we have already met two such exten-
sions :
31
Although, it will not help in this course, note that the converse is true. See for example
the book [Ro], in which the essential selfadjointness is dened dierently.
91
the Dirichlet realization whose domain is the set H
1
0
() H
2
(),
the Neumann realization whose domain is the set :
n H
2
() [ (n,)
/
= 0
4. The Laplace Beltrami operator on a compact manifold ` with domain
C

(`) is essentially selfadjoint on 1

2
(`). The domain of the selfadjoint
extension can be described as H
2
(`) (which can be described using local
charts).
10.2 Basic criteria.
We now give some criteria in order to verify that an operator is essentially
selfadjoint. As already mentioned, one can prove essential selfadjointness by
proving that the minimal closed extension T
min
:= T coincides with T

. One
easily veries (see Proposition 2.2.5 in conjonction with the denition of T) that
T

= T

and we recall that T

= T.
We now observe the
Proposition 10.2.1.
Any closed symmetric extension of T is a restriction of T

.
Proof.
Let o is a closed symmetric extension of T.
We have indeed T o o

and, observing that o

, we consequently get
o T

.
In particular, if T is selfadjoint then T is the unique selfadjoint extension of
T.
We can characterize the selfadjointness through the following general crite-
rion
Theorem 10.2.2.
Let T be a closed symmetric operator. Then the following statements are equiv-
alent :
1. T is selfadjoint.
2. Ker (T

i) = 0;
3. Range (T i) = H.
Proof.
1. implies 2.
This property was already observed (because T = T

and i , (T)).
2. implies 3. .
We rst observe that the property that ker (T

+i) = 0 implies that 1(T i)

is dense in H. Note that the converse is also true. For getting 3., it remains
92
to show that 1(T i) is closed. But, for all in 1(T), we have (using the
symmetry of T)
[[(T i)[[
2
= [[T[[
2
+[[[[
2
. (10.2.1)
If
n
is a sequence in 1(T) such that (T +i)
n
converges to some

, then the
previous identity shows that
n
is a Cauchy sequence, so there exists

such
that
n

in H. But T
n
= (T + i)
n
i
n
is convergent and using that
the graph is closed, we obtain that

1(T) and T

.
3. implies 1. .
Let 1(T

). Let 1(T) such that (T i) = (T

i). T being
symmetric, we have also (T

i)( ) = 0. But, if (T +i) is surjective, then

(T

i) is injective and we get = . This proves that 1(T).

Remark 10.2.3.
Here we have used and proved during the proof of the assertion 2. implies 3.
the following lemma
Lemma 10.2.4.
If T is closed and symmetric, then 1(T i) is closed.
This theorem gives as a corollary a criterion for essential selfadjointness in
the form
Corollary 10.2.5.
Let with domain 1() be a symmetric operator. Then the following are
equivalent
1. is essentially selfadjoint.
2. Ker(

i) = 0.
3. The two spaces 1(i) are dense in H.
We have indeed essentially to apply the previous theorem to , observing in
addition that is symmetric and using Lemma 10.2.4.
Let us here emphasize that in this case, to specify the operator , it is not
necessary to give the exact domain of but a core for that is a subspace 1
such that the closure of
/D
is .
In the same spirit, we have in the semibounded case the following
Theorem 10.2.6.
Let T be a positive, symmetric operator. Then the following statements are
equivalent :
1. T is essentially selfadjoint.
2.
1c: (T

+/) = 0 for some / 0 .

3. Range (T +/) is dense for some / 0.
93
The proof essentially the same as for the previous corollary, if one observes
that if T is positive then the following trivial estimate is a good substitute for
(10.2.1) :
(T +/)n, n) /[[n[[
2
.
Example 10.2.7. (The free Laplacian)
The operator with domain C

0
is essentially selfadjoint. Its selfadjoint
extension is with domain H
2
.
10.3 The Kato -Rellich theorem
We would like to consider the case when 1 = + \ when \ is regular and
tends to 0 as [r[ +. Here \ can be considered as a perturbation of the
Laplacian. One can then apply a general theorem due to Kato-Rellich.
Theorem 10.3.1.
Let be a selfadjoint operator, 1 be a symmetric operator whose domain con-
tains 1(). Let us assume the existence of o and / such that 0 o < 1 and
/ 0 such that
[[1n[[ o [[n[[ +/ [[n[[ (10.3.1)
for all n 1(). Then +1 is selfadjoint on 1().
If is essentially selfadjoint on
32
1 1(), then +1 has the same property.
Proof.
Step 1.
We start from the following identity which only uses that (+1) with domain
1() is symmetric.
[[(+1 i)n[[
2
= [[(+1)n[[
2
+
2
[[n[[
2
, n 1() . (10.3.2)
By the triangle inequality and the symmetry of + 1, we get for a real
0, and for any n 1() :

2[[(+1 i)n[[ [[(+1)n[[ +[[n[[

[[n[[ [[1n[[ +[[n[[
(1 o)[[n[[ + ( /)[[n[[.
(10.3.3)
We now choose /.
Step 2.
Let us show that (+1) with domain 1() is closed. If we start indeed from a
pair (n
n
, )
n
) with n
n
1() and )
n
= (+1)n
n
such that (n
n
, )
n
) (n, )) in
H. From (10.3.3), we get that n
n
is a Cauchy sequence in H. being closed,
we get n 1() and the existence of p such that n
n
p = n.
Now from (10.3.2) and (10.3.1), we get also that 1n
n
is a Cauchy sequence and
32
By this we mean that the closure of A
/D
is A.
94
there exists such that 1n
n
in H.
We claim that 1n = . We have indeed, for any / 1(),
< , /
H
= lim
n+
< 1n
n
, /
H
= lim
n+
< n
n
, 1/
H
=< n, 1/
H
=< 1n, /
H
.
Using the density of 1(), we get = 1n. (We could have also used that 1 is
closable).
We conclude by observing that (+1)n = ) (with ) = p +) as expected.
Step 3.
In order to apply Theorem 10.2.2, we have to show that (+1i) is surjective.
The main element in the proof is the following
Lemma 10.3.2. .
For 0 large enough, we have
[[1(i)
1
[[ < 1 . (10.3.4)
Proof.
We observe that, for n 1(),
[[(i)n[[
2
= [[n[[
2
+
2
[[n[[
2
. (10.3.5)
For n 1(), we have, using two times (10.3.5) and then (10.3.1)
[[1n[[ o[[n[[ +/[[n[[
o[[(+i)n[[ +
b

[[(+i)n[[
(o +
b

)[[(+i)n[[ .
(10.3.6)
It is then enough to choose 0 large enough such that
(o +
/

) < 1 .
Writing
+1 i = [1 +1(i)
1
](i) , (10.3.7)
it is easy to deduce the surjectivity using the lemma and the surjectivity of
(i).
Application.
As an application, let us treat the case of the Schrodinger operator with
Coulomb potential.
Proposition 10.3.3.
The operator
1
|x|
with domain C

0
(R
3
) is essentially selfadjoint.
We recall that the operator is well dened because
1
r
belongs to 1
2
loc
(R
3
).
We rst observe a Sobolev type inequality.
95
Lemma 10.3.4. :
There exists a constant C such that for all n H
2
(R
3
), all o 0 and all r R
3
,
we have
[n(r)[ C(o[[n[[
0
+o
3
[[n[[
0
) .
(Prove rst the inequality
33
for all n with r = 0 and o = 1, then use
translation and dilation.)
In the second step, we show that the potential \ =
1
r
is a perturbation of the
Laplacian.
There exists indeed a constant C such that for all n H
2
(R
3
) and all / 0, we
have
[[\ n[[
0
C(/[[n[[
0
+/
3
[[n[[
0
) .
For this proof we observe that, for any 1 0,
_
\ (r)
2
[n(r)[
2
dr =
_
|x|R
\ (r)
2
[n(r)[
2
dr +
_
|x|R
\ (r)
2
[n(r)[
2
dr ,
and treat the rst term of the right hand side using the Sobolevs type inequality
by
_
|x|R
\ (r)
2
[n(r)[
2
dr ( sup
|x|R
[n(r)[
2
dr)
_
|x|R
\ (r)
2
dr,
and the second term by the trivial estimate
_
|x|R
\ (r)
2
[n(r)[
2
dr ( sup
|x|R
[\ (r)[)
2
_
|x|R
[n(r)[
2
dr .
Using the Sobolev inequality (actually it is enough to take o = 1), we nally
obtain :
[[\ n[[
2
C1([[n[[
2
+[[n[[
2
) +
1
1
2
[[n[[
2
. (10.3.8)
We obtained the expected estimate by considering 1 small enough.
Remark 10.3.5.
We note that the same proof shows that +\ is essentially selfadjoint starting
from C

0
(R
m
) if \ 1
2
+1

, that is if \ = \
1
+\
2
with \
1
1
2
and \
2
1

.
10.4 Other criteria of selfadjointness for Schrodinger op-
erators
We present in this subsection two criteria which are specic of the Schrodinger
case. The rst one seems due to Rellich (See [Sima]) and we present it in the
easy case when the potential is regular. The second one permits to treat singular
potentials and is due to Kato (cf [HiSi] or [Ro]).
The rst theorem is adapted to operators which are already know to be
positive on C

0
(R
m
).
33
In this case, this is just the Sobolevs injection theorem of H
2
(R
3
) into C
0
b
(R
3
), where
C
0
b
(R
3
) is the space of the continuous bounded functions.
96
Theorem 10.4.1.
A Schrodinger operator T = + \ on R
n
associated with a C
0
potential \ ,
which is semibounded on C

0
(R
n
), is essentially selfadjoint. In other words, the
Friedrichs extension is the unique selfadjoint extension starting from C

0
(R
m
).
This theorem is complementary to the second theorem (Theorem 10.4.4)
which will be stated at the end of this subsection because we do not have to
assume the positivity of the potential but only the semi-boundedness of the
operator T.
Proof.
Let T be our operator. Possibly by adding a constant, we can assume that
< Tn, n
H
[[n[[
2
, n C

0
(R
m
) . (10.4.1)
Of course this inequality can be rewritten in the form :
[[n[[
2
+
_
R
m
\ (r)[n(r)[
2
dr [[n[[
2
, n C

0
(R
m
) .
In this form, the inequality can be extended to the elements of H
1
comp
(R
m
),
corresponding to the distributions of H
1
(R
m
) with compact support :
[[n[[
2
+
_
R
m
\ (r)[n(r)[
2
dr [[n[[
2
, n H
1
comp
(R
m
) . (10.4.2)
According to the general criterion of essential selfadjointness (cf Theorem 10.2.6),
it is enough to verify that 1(T) is dense. Let us show this property.
Let ) 1
2
(R
m
), such that
< ), Tn
H
= 0 , n C

0
(R
m
) . (10.4.3)
We have to show that ) = 0.
Because T is real, one can assume that ) is real.
We rst observe that (10.4.3) implies that : ( + \ )) = 0 in T

(R
m
). A
standard regularity theorem for the Laplacian
34
implies that ) H
2
loc
(R
m
).
We now introduce a family of cuto functions
k
by

k
:= (r,/) , / N , (10.4.4)
where is a C

function satisfying 0 1, = 1 on 1(0, 1) and supp

1(0, 2).
For any n C

and any ) H
2
loc
, we have the identity
_
(
k
)) (
k
n)dr +
_

k
(r)
2
\ (r)n(r) )(r) dr
=
_
[(
k
)(r)[
2
n(r))(r) dr +

m
i=1
_
()(
i
n) n(
i
))) (r)
k
(r)(
i

k
)(r) dr
+)(r) , T
2
k
n) .
(10.4.5)
34
This is the property that f L
2
loc
(R
m
), f H
1
loc
(R
m
) implies that f H
1
loc
(R
m
)
together with property that f L
2
loc
(R
m
), f L
2
loc
(R
m
) implies that f H
2
loc
(R
m
) .
97
When ) satises (10.4.3), we get :
_
R
m
(
k
)) (
k
n)dr +
_

k
(r)
2
\ (r)n(r) )(r) dr
=
_
[(
k
)(r)[
2
n(r))(r) dr +

m
i=1
_
()(
i
n) n(
i
))) (r)
k
(r)(
i

k
)(r) dr ,
(10.4.6)
for all n C

(R
m
).
This formula can be extended to functions n H
1
loc
. In particular, we can take
n = ).
We obtain
< (
k
)) , (
k
)) +
_

2
k
\ (r)[)(r)[
2
dr =
_
[
k
[
2
[)(r)[
2
dr . (10.4.7)
Using (10.4.1), (10.4.7) and taking the limit / +, we get
[[)[[
2
= lim
k+
[[
k
)[[
2
limsup
k+
_
< (
k
)) , (
k
)) +
_

2
k
\ (r)[)(r)[
2
dr
_
= limsup
k+
_
)(r)
2
[(
k
)(r)[
2
dr = 0 .
(10.4.8)
This proves the theorem.
Remark 10.4.2.
When \ is C

, we get, in the previous proof, that ) C

and we immediately
can prove (10.4.7) without going through the previous discussion.
Example 10.4.3.
If \ 0 and C

, then the Schrodinger operator + \ with domain

C

0
(R
m
) is essentially selfadjoint. The operator +\ is indeed positive.
If is C

on R
m
, then the operators + [[
2
are essentially
selfadjoint. They are indeed positive on C

0
(R
m
). They can actually be
written in the form

j
7

j
7
j
with 7
j
=
xj

xj
. These operators
appear naturally in statistical mechanics.
Let us now mention, without proof, a quite general theorem due to Kato
(See for example [Ro]).
Theorem 10.4.4.
Let \ in 1
2
loc
(R
m
) such that \ 0 almost everywhere on R
m
. Then + \
with domain C

0
(R
m
) is essentially selfadjoint.
This theorem is based on the so called Katos inequality.
Remark 10.4.5.
This last theorem may be extended to the case of the Schrodinger operator with
magnetic regular potential (See Subsection 7.5).
98
11 Non-selfadjoint case : Maximal accretivity
and application to the Fokker-Planck opera-
tor
11.1 Accretive operators
We collect here some material on accretive operators. The references could be
the books by Dautray-Lions (Vol. 5, Chapter XVII), Reed-Simon or the book
of B. Davies. Let H be a complex (or real) Hilbert space.
Denition 11.1.1.
Let be an unbounded operator in H with domain 1(). We say that is
accretive if
Re r [ r)
H
0 , r 1() . (11.1.1)
Denition 11.1.2.
An accretive operator is maximally accretive if it does not exist an accretive
extension

with strict inclusion of 1() in 1(

).
Proposition 11.1.3.
Let be an accretive operator with domain 1() dense in H. Then is closable
and its closed extension is accretive.
For the analysis of the Fokker-Planck operator, the following criterion, which
extends the standard criterion of essential self-adjointness, will be the most
suitable
Theorem 11.1.4.
For an accretive operator , the following conditions are equivalent
1. is maximally accretive.
2. There exists
0
0 such that

+
0
1 is injective.
3. There exists
1
0 such that the range of +
1
1 is dense in H.
Note that in this case is the innitesimal generator of a contraction
semi-group.
11.2 Application to the Fokker-Planck operator
We would like to show
Proposition 11.2.1.
Let \ be a C

potential on R
n
, then the closure 1 of the Fokker-Planck operator
dened on C

0
(R
2n
) by
1 :=
v
+
1
4
[[
2

n
2
+A
0
, (11.2.1)
99
where
A
0
:= \ (r)
v
+
x
(11.2.2)
is maximally accretive.
Moreover 1

is also maximally accretive.

The idea is to adapt the proof that a semi-bounded Schrodinger operator
with regular potential is essentially self-adjoint on 1
2
(R
n
).
Proof:
We apply the abstract criterion taking H = 1
2
(R
n
R
n
) and = 1. The
operators being real, we can consider everywhere real functions. The accretivity
on C

0
(R
2n
) is clear. We can then consider the closure 1.
Changing 1 in T := 1 + (
n
2
+ 1)1, we would like to show that its range is
dense.
Let ) 1
2
(R
m
), with : = 2n, such that
< ) [ Tn
H
= 0 , n C

0
(R
m
) . (11.2.3)
We have to show that ) = 0.
Because 1 is real, one can assume that ) is real.
We rst observe that (11.2.3) implies that :
(
v
+
2
,4 + 1 A
0
)) = 0 , in T

(R
m
) .
The standard hypoellipticity theorem for the Hormander operators
35
of type 2
implies that ) C

(R
m
).
We now introduce a family of cut-o functions
k
:=
k1,k2
by

k1,k2
(r, ) := (r,/
1
)(,/
2
) , / N
2
, (11.2.4)
where is a C

function satisfying 0 1, = 1 on 1(0, 1) and supp

1(0, 2).
For any n C

0
, we have the identity
_

v
(
k
))
v
(
k
n) drd +
_

k
(r, )
2
(
2
,4 + 1)n(r, ) )(r, ) drd
+
_
)(r, )(A
0
(
2
k
n))(r, ) drd
=
_
[(
v

k
)(r, )[
2
n(r, ))(r, ) drd
+

m
i=1
_
()(
vi
n) n(
vi
))) (r, )
k
(r, )(
vi

k
)(r, ) drd
+)(r, ) [ T
2
k
n) .
(11.2.5)
When ) satises (11.2.3), we get :
_
R
m

v
(
k
))
v
(
k
n) drd +
_

2
k
(
2
,4 + 1)n(r, ) )(r, ) drd
+
_
)(r, )(A
0
(
2
k
n))(r, )drd
=
_
[(
y

k
)(r)[
2
n(r))(r, ) drd
+

m
i=1
_
()(
vi
n) n(
vi
))) (r, )
k
(r, )(
vi

k
)(r, ) drd ,
(11.2.6)
35
These operators are in the form P =
P
k
j=1
X
2
j
+ X
0
+ a(x), where the X
j
are real
vectoreld. If the X
j
together with the brackets [X

, Xm] span at each point x the whole

tangent space, then one can show that the corresponding operator is hypoelliptic. P is said
hypoelliptic if for any u T

and any open set , Pu C

() implies that u C

().
100
for all n C

(R
m
). In particular, we can take n = ).
We obtain
<
v
(
k
)) [
v
(
k
)) +
_

2
k
(
2
,4 + 1)[)(r, )[
2
drd
+
_
)(r, )(A
0
(
2
k
)))(r, ) drd
=
_
[
v

k
[
2
[)(r, )[
2
drd .
(11.2.7)
With an additional integration by part, we get
<
v
(
k
)) [
v
(
k
)) +
_

2
k
(
2
,4 + 1)[)(r, )[
2
drd
+
_

k
)(r, )
2
(A
0

k
)(r, ) drd
=
_
[
v

k
[
2
[)(r, )[
2
drd .
(11.2.8)
This leads to the existence of a constant C such that, for all /,
[[
k
)[[
2
+
1
4
[[
k
)[[
2
C
1
k
2
2
[[)[[
2
+C
1
k1
[[
k
)[[ [[)[[ +C
1
k2
[[\ (r)
k
)[[ [[)[[ .
(11.2.9)
(The constant C will possibly be changed from line to line). This leads to
[[
k
)[[
2
+
1
8
[[
k
)[[
2
C(
1
/
2
2
+
1
/
2
1
)[[)[[
2
+C(/
1
)
1
/
2
[[
k
)[[ [[)[[ , (11.2.10)
where
C(/
1
) = sup
|x|2k1
[
x
\ (r)[
This implies
[[
k
)[[
2
C(

C(/
1
)
/
2
2
+
1
/
2
1
)[[)[[
2
. (11.2.11)
This nally leads to ) = 0. For example, one can take rst the limit
/
2
+, which leads to
[[(
r
/
1
))[[
2

C
/
2
1
[[)[[
2
,
and then the limit /
1
+.
11.3 Decay of the semi-group and -pseudospectra
We recall that for any c 0, we dene the c-pseudospectra by

(/
D
) = C[ [[(/
D
)
1
[[
1
c
, (11.3.1)
with the convention that [[(/
D
)
1
[[ = + if (/
D
).
We have

>0

(/
D
) = (/
D
) . (11.3.2)
101
We dene, for any accretive closed operator, for c 0,

(/) = inf
z(A)
Re . . (11.3.3)
We also dene

0
(/) = lim
t+
1
t
log [[ expt/[[ (11.3.4)

(/) inf
z(A)
Re . . (11.3.5)
Theorem 11.3.1 (Gearhart-Pr uss).
Let / be a densely dened closed operator in an Hilbert space A such that
/ generates a contraction semi-group and let

(/) and
0
(/) denote the
c-pseudospectral abcissa and the growth bound of / respectively. Then
lim
0

(/) =
0
(/) . (11.3.6)
We refer to Engel-Nagel for a proof.
This theorem is interesting because it reduces the question of the decay,
which is basic in the question of the stability to an analysis of the c-spectra of
the operator.
11.4 Application : The complex Airy operator in R
+
Here we mainly describe some results presented in Almog (article in Siam). We
can then associate the Dirichlet realization /
D
of the complex Airy operator
1
2
x
+ir on the half-line, whose domain is
1(/
D
) = n H
1
0
(R
+
), r
1
2
n 1
2
, (1
2
x
+i r)n 1
2
(R
+
) , (11.4.1)
and which is dened (in the sense of distributions) by
/
D
n = (1
2
x
+i r)n. (11.4.2)
Moreover, by construction, we have
Re /
D
n[ n) 0 , n 1(/
D
) . (11.4.3)
Again we have an operator, which is the generator of a semi-group of contraction,
whose adjoint is described by replacing in the previous description (1
2
x
+i r) by
(1
2
x
i r), the operator is injective and as its spectrum contained in Re 0.
Moreover, the operator has compact inverse, hence the spectrum (if any) is
discrete.
Using what is known on the usual Airy operator, Sibuyas theory and a
complex rotation (see alternately what we said for the non self-adjoint harmonic
oscillator), we obtain that the spectrum of /
D
(/
D
) is given by that
(/
D
) =
+
j=1

j
(11.4.4)
102
with

j
= expi

3
j
j
, (11.4.5)
the j
j
s being real zeroes of the Airy function satisfying
0 < j
1
< < j
j
< j
j+1
< . (11.4.6)
It is also shown in Almog that the vector space generated by the correspond-
ing eigenfunctions is dense in 1
2
(R
+
).
We arrive now to the analysis of the properties of the semi-group and the
estimate of the resolvent.
As before, we have, for Re < 0,
[[(/
D
)
1
[[
1
[ Re [
, (11.4.7)
If Im < 0 one gets also a similar inequality, so the main remaining question is
the analysis of the resolvent in the set Re 0 , Im 0, which corresponds
to the numerical range of the symbol.
We recall that for any c 0, we dene the c-pseudospectra by

(/
D
) = C[ [[(/
D
)
1
[[
1
c
, (11.4.8)
with the convention that [[(/
D
)
1
[[ = + if (/
D
).
We have

>0

(/
D
) = (/
D
) . (11.4.9)
We dene, for any accretive closed operator, for c 0,

(/) = inf
z(A)
Re . . (11.4.10)
We also dene

0
(/) = lim
t+
1
t
log [[ expt/[[ (11.4.11)

(/) inf
z(A)
Re . . (11.4.12)
We apply Gearhart-Pr uss theorem to our operator /
D
and our main theorem
is
Theorem 11.4.1.

0
(/
D
) = Re
1
. (11.4.13)
Using the rst eigenfunction it is easy to see that
[[ expt/
D
[[ exp Re
1
t . (11.4.14)
Hence we have immediately
0
0
(/
D
) Re
1
. (11.4.15)
To prove that Re
1

0
(/
D
), it is enough to show the following lemma.
103
Figure 2: Complex Airy operator on the haline : pseudospectra
Lemma 11.4.2.
For any < Re
1
, there exists a constant C such that, for all s.t. Re
[[(/
D
)
1
[[ C . (11.4.16)
Proof : We know that is not in the spectrum. Hence the problem is just
a control of the resolvent as [ Im [ +. The case, when Im < 0 has
already be considered. Hence it remains to control the norm of the resolvent as
Im + and Re [, +].
This is indeed a semi-classical result ! The main idea is that when Im
+, we have to inverse the operator
1
2
x
+i(r Im ) Re .
If we consider the Dirichlet realization in the interval ]0,
Im
2
[ of
1
2
x
+ i(r Im ) Re , it is easy to see that the operator is invertible
by considering the imaginary part of this operator and that this inverse 1
1
()
satises
[[1
1
()[[
2
Im
.
Far from the boundary, we can use the resolvent of the problem on the line for
which we have a uniform control of the norm for Re [, +].
104
12 Discrete spectrum, essential spectrum
12.1 Discrete spectrum
We have already recalled in Proposition 8.5.4 a characterization of the spectrum.
Let us now complete this characterization by introducing dierent spectra.
Denition 12.1.1.
If T is a selfadjoint operator, we shall call discrete spectrum of T the set

disc
(T) = (T) :. t. c 0, dimrange (1(] c, +c[)) < + .
With this new denition, we can say that, for a selfadjoint operator with
compact resolvent, the spectrum is reduced to the discrete spectrum.
For a compact selfadjoint operator, the spectrum is discrete outside 0. We see
in this case that the discrete spectrum is not closed.
Equivalently, let us observe now give another characterization :
Proposition 12.1.2.
Let T be a selfadjoint operator. A real is in the discrete spectrum if and only
if :
is an isolated point in (T) and if is an eigenvalue of nite multiplicity
36
.
Proof.
If
disc
(T), we immediately see that there exists c
0
such that, c such
that 0 < c < c
0
, 1
],+[
becomes a projector independent of c with nite
range. This is actually the projector

= 1
{}
(T) and we observe moreover
1
],+0[
= 0 and 1
]0,[
= 0. This shows that is an isolated point in (T).
Using the spectral representation of T, one immediately get that, if r =

r
(r ,= 0), then r is an eigenfunction of T. Moreover, one easily obtains that
(T ) is invertible on 1(1

). One can indeed nd a continuous bounded

) such that )(T)(T )(1

) = (1

).
Conversely, let be isolated. The previous proof as already shown that in
this case the range of

is an eigenspace. The assumption of nite multiplicity

permits then to conclude.
Remark 12.1.3.
The discrete spectrum is not a closed set ! If we consider in R
3
, the Schrodinger
operator with coulomb potential, the discrete spectrum is a sequence of eigenval-
ues tending to 0 but 0 does not belong to the discrete spectrum.
12.2 Essential spectrum
Denition 12.2.1.
The essential spectrum is the complementary in the spectrum of the discrete
spectrum.
36
that is with corresponding eigenspace of nite dimension.
105
Intuitively, a point of the essential spectrum corresponds
either to a point in the continuous spectrum,
or to a limit point of a sequence of eigenvalues with nite multiplicity,
or to an eigenvalue of innite multiplicity.
The discrete spectrum being composed of isolated points, we get
Proposition 12.2.2.
The essential spectrum of a selfadjoint operator T is closed in R.
12.3 Basic examples:
1. The essential spectrum of a compact selfadjoint operator is reduced to 0.
2. The essential spectrum of an operator with compact resolvent is empty.
3. The Laplacian on R
n
is a selfadjoint operator on 1
2
(R
n
) whose domain
is the Sobolev space H
2
(R
n
). The spectrum is continuous and equal to
R
+
. The essential spectrum is also R
+
and the operator has no discrete
spectrum.
4. The Schrodinger operator with constant magnetic eld (1 ,= 0) in R
2
:
o
B
:= (1
x1

1r
2
2
)
2
+ (1
x2
+
1r
1
2
)
2
, (12.3.1)
with
1
xj
:=
1
i

xj
=
1
i

r
j
.
The spectrum is formed with eigenvalues (2/ +1)[1[ but the spectrum is
not discrete because each eigenvalue is with innite multiplicity. This will
be treated in the next subsection.
12.4 On the Schr odinger Operator with constant mag-
netic eld
The Schrodinger operator with magnetic eld has been briey introduced in
Subsection 7.5. We have in particularly given examples where this operator was
with compact resolvent. In this section, we analyze more in detail the properties
of the Schrodinger operator with constant magnetic eld in dimension 2 and 3.
This appears to play an important role in Superconductivity theory.
12.4.1 Dimension 2
We would like to analyze the spectrum of :
o
B
:= (1
x1

1
2
r
2
)
2
+ (1
x2
+
1
2
r
1
)
2
. (12.4.1)
106
12.4.2 The case of R
2
We rst look at the selfadjoint realization in R
2
. Let us show briey, how one
can analyze its spectrum. We leave as an exercise to show that the spectrum
(or the discrete spectrum) of two selfadjoints operators o and T are the same
if there exists a unitary operator l such that l(o i)
1
l
1
= (T i)
1
. We
note that this implies that l sends the domain of o onto the domain of T.
In order to determine the spectrum of the operator o
B
, we perform a succession
of unitary conjugations. The rst one is called a gauge transformation. We
introduce l
1
on 1
2
(R
2
) dened, for ) 1
2
by
l
1
) = expi1
r
1
r
2
2
) . (12.4.2)
It satises
o
B
l
1
) = l
1
o
1
B
) , ) o(R
2
) , (12.4.3)
with
o
1
B
:= (1
x1
)
2
+ (1
x2
+1r
1
)
2
. (12.4.4)
Remark 12.4.1. .
l
1
is a very special case of what is called a gauge transform. More generally,
we can consider l = expi where is C

. If
A
:=

j
(1
xj

j
)
2
is a
general Schrodinger operator associated with the magnetic potential , then
l
1

A
l =

A
where

= + grad. Here we observe that 1 := rot =
rot

. The associated magnetic eld is unchanged in a gauge transformation.
We are discussing in our example the very special case (but important!) when
the magnetic potential is constant.
We have now to analyze the spectrum of o
1
B
.
Observing that the operator is with constant coecients with respect to the r
2
-
variable, we perform a partial Fourier transform with respect to the r
2
variable
l
2
= T
x22
, (12.4.5)
and get by conjugation, on 1
2
(R
2
x1,2
),
o
2
B
:= (1
x1
)
2
+ (
2
+1r
1
)
2
. (12.4.6)
We now introduce a third unitary transform l
3
(l
3
))(j
1
,
2
) = )(r
1
,
2
) , with j
1
= r
1
+

2
1
, (12.4.7)
and we obtain the operator
o
3
B
:= 1
2
y
+1
2
j
2
, (12.4.8)
operating on 1
2
(R
2
y,2
).
The operator depends only on the j variable. It is easy to nd for this operator
107
an orthonormal basis of eigenvectors. We observe indeed that if ) 1
2
(R
2
),
and if
n
is the (n + 1)-th eigenfunction of the harmonic oscillator, then
(r,
2
) [1[
1
4
)(
2
)
n
([1[
1
2
j)
is an eigenvector corresponding to the eigenvalue (2n+1)[1[. So each eigenspace
has an innite dimension. An orthonormal basis of this eigenspace can be given
by the vectors c
j
(
2
)[1[
1
4

n
([1[
1
2
j) where c
j
(, N) is a basis of 1
2
(R).
We have consequently an empty discrete spectrum. The eigenvalues are usually
called Landau levels.
12.4.3 Magnetic Schrodinger operators in dimension 3
We only consider the Schrodinger operator with constant magnetic eld in R
3
.
After some rotation in R
3
, we arrive to the model :
1(/,

/) = 1
2
x1
+ (1
x2
/r
1
)
2
+1
2
x3
, (12.4.9)
with
/ = [[1[[ .
This time, we can take the partial Fourier transform, with respect to r
2
and r
3
in order to get the operator
1
2
x1
+ (
2
/r
1
)
2
+
2
3
.
When / ,= 0, we can translate in the r
1
variable and get the operator on 1
2
(R
3
)
1
2
y1
+ ([/[j
1
)
2
+
2
3
.
It is then easy to see that the spectrum is [[/[, +[. No point being isolated,
we cannot expect any discrete spectrum.
12.5 Weyls criterion:
We have already mentioned that the essential spectrum is a closed set. In order
to determine the essential spectrum it is useful to have theorems proving the
invariance by perturbation. The following characterization is in this spirit quite
useful.
Theorem 12.5.1.
Let T be a selfadjoint operator. Then belongs to the essential spectrum if and
only if there exists a sequence n
n
in 1(T) with [[n
n
[[ = 1 such that n
n
tends
weakly
37
to 0 and [[(T )n
n
[[ 0 as n .
37
We recall that we say that a sequence un in a separable Hilbert space 1 is weakly con-
vergent if, for any g in 1, < un[g >
H
is convergent. In this case, there exists a unique f such
that < un[g >
H
< f[g >
H
and [[un[[ is a bounded sequence.
108
Let us give a proof of the theorem (we also refer to [Ro], [HiSi] or [RS-IV]).
The sequence appearing in the theorem is called a Weyl sequence. A point
such that there exists an associated Weyl sequence is said to belong to the Weyl
spectrum \(T). Let us show the inclusion
\(T)
ess
(T) . (12.5.1)
We have already seen that
\(T) (T) . (12.5.2)
Let us suppose by contradiction that
disc
(T). Let

:= 1
{}
be the
associated spectral projector. We rst observe that,

being nite range,

hence compact, we have :

n
n
0 H . (12.5.3)
Let us dene
n
n
= (1

)n
n
.
We get [[n
n
[[ +1 and (T )n
n
= (1

)(T )n
n
0.
But (T ) is invertible on 1(1

), so we get n
n
0 and the contradiction.
This shows the announced inclusion (12.5.1).
For the converse, we rst observe that, if
ess
(T) then, for any c 0,
dim1(1
],+[
) = +. Considering a decreasing sequence c
n
such that c
n

0 and lim
n+
c
n
= 0, it is easy to obtain an orthonormal system n
n
such
that n
n
1(1
]n,+n[
). With this last property, we immediately get
38
that
[[(T )n
n
[[ 0.
Corollary 12.5.2.
The operator /
2
+\ with \ a continuous function tending to 0 as [r[
(r R
D
) has R
+
as essential spectrum.
For proving the inclusion of R
+
in the essential spectrum, we can indeed
consider the sequence
n
n
(r) = exp(ir )n
(D/2)
((r 1
n
),n)
with 0 and supported in the ball 1(0, 1) and equal to one on say 1(0,
1
2
).
The sequence 1
n
is chosen such that [1
n
[ (for example [1
n
[ = n
2
) tends to
and such that the support of the n
n
are disjoints.
This is a particular case of a Weyl sequence (called in [HiSi] a Zhislin sequence).
The converse can be obtained by abstract analysis and the fact that we know
that the essential spectrum of is [0, +[. This idea will be formalized
through the following notion of relative compactness.
Denition 12.5.3.
If T is a closed operator with a dense domain 1
T
, we shall say that the operator
\ is relatively compact with respect to T or T-compact if 1
T
1
V
and if the
image by \ of a closed ball in 1
T
(for the graph-norm n
_
[[n[[
2
+[[Tn[[
2
)
is relatively compact in H.
38
See also what we have done for the proof of Proposition 8.5.4.
109
In other words, we shall say that \ is T-compact, if, from each sequence
n
n
in 1
T
which is bounded in H and such that Tn
n
is bounded in H, one can
extract a subsequence n
ni
such that \ n
ni
is convergent in H. Here we recall
(exercise) that when T is closed, 1(T) equipped with the graph norm is an
Hilbert space.
Example 12.5.4.
If \ is the multiplication operator by a continuous function \ tending to 0
then \ is ()-compact. This is a consequence of Proposition 5.4.1 and of the
uniform continuity of \ on each compact.
Weyls Theorem says
Theorem 12.5.5. .
Let T be a selfadjoint operator, and \ be symmetric and T-compact, then T +
\ is selfadjoint and the essential spectrum of T + \ is equal to the essential
spectrum of T.
The rst part can be deduced from what was discussed in the proof of
Kato-Rellichs theorem (See Theorem 10.3.1). We observe indeed the following
variant of Lionss Lemma :
Lemma 12.5.6.
If \ is T-compact and closable, then, for any o 0, there exists / 0 such that
[[\ n[[ o[[Tn[[ +/[[n[[ , n 1(T) . (12.5.4)
Proof of the lemma.
The proof is by contradiction. If (12.5.4) is not true, then there exists o 0
such that, n N

, there exists n
n
1(T) such that
o[[Tn
n
[[ +n[[n
n
[[ < [[\ n
n
[[ . (12.5.5)
Observing that [[\ n
n
[[ , = 0 and that the inequation is homogeneous, we can in
addition assume that n
n
satises the condition :
[[\ n
n
[[ = 1 . (12.5.6)
From these two properties we get that the sequence Tn
n
is bounded and that
n
n
0.
On the other hand, by T-compactness, we can extract a subsequence n
n
k
such
that \ n
n
k
is convergent to with [[[[ = 1. But (0, ) is in the closure of the
graph of \ . But \ being closable, this implies = 0 and a contradiction.
For the second part we can use the Theorem 12.5.1. If we take a Weyls
sequence n
n
such that n
n
0 (weakly) and (T )n
n
0 strongly, let us
consider (T +\ )n
n
. We have simply to show that one can extract a subse-
quence n
n
k
such that (T +\ )n
n
k
0.
110
But Tn
n
is a bounded sequence. By the T-compactness, we can extract a sub-
sequence such that \ n
n
k
converges strongly to some in H. It remains to show
that = 0. But here we can observe that for any ) 1(T), we have
< , ) = lim
k+
< \ n
n
k
, ) = lim
k+
< n
n
k
, \ ) = 0 .
Here we have used the symmetry of \ and the weak convergence of n
n
to 0.
Using the density of 1(T) in H, we obtain = 0.
This shows that a Weyl sequence for T is a Weyl sequence for T + \ . For the
converse, one can intertwine the roles of T and T +\ , once we have shown that
\ is (T + \ )-compact. For this, we can use Lemma 12.5.6, and observe that
the following inequality is true :
[[Tn[[
1
1 o
([[(T +\ )n[[ +/[[n[[) . (12.5.7)
This shows that if n
n
is a sequence such that ([[n
n
[[ +[[(T +\ )n
n
[[) is bounded,
then this sequence has also the property that ([[n
n
[[ +[[Tn
n
[[) is bounded.
111
13 The max-min principle
13.1 Introduction
The max-min principle is an alternative way for describing the lowest part of
the spectrum when it is discrete. It gives also an ecient way to localize these
eigenvalues or to follow their dependence on various parameters.
13.2 On positivity
We rst recall the following denition
Denition 13.2.1.
Let be a symmetric operator. We say that is positive (and we write 0),
if
< n, n 0 , n 1() . (13.2.1)
The following proposition relates the positivity with the spectrum
Proposition 13.2.2.
Let be a selfadjoint operator. Then 0 if and only if () [0, +[.
Proof.
It is clear that if the spectrum is in R
+
, then the operator is positive. This can
be seen for example through the spectral representation :
< n, n =
_
(A)
d[[1

n[[
2
.
Now, if 0, then, for any o 0, +o is invertible. We have indeed
o[[n[[
2
< (+o)n, n [[(+o)n[[ [[n[[ ,
which leads to
o[[n[[ [[(+o)n[[ , n 1() . (13.2.2)
This inequality gives the closed range and the injectivity. being selfadjoint,
we get also from the injectivity, the density of the image of (+o). This shows
that o is not in the spectrum of .
Example 13.2.3.
Let us consider the Schrodinger operator 1 := + \ , with \ C

and
semi-bounded, then
(1) [inf \, +[ . (13.2.3)
112
13.3 Variational characterization of the discrete spectrum
Theorem 13.3.1.
Let be a selfadjoint semibounded operator. Let := inf
ess
() and let us con-
sider ()] , [, described as a sequence (nite or innite) of eigenvalues
that we write in the form

1
<
2
< <
n
.
Then we have

1
= inf
D(A),=0
[[[[
2
< , , (13.3.1)

2
= inf
D(A)K

1
,=0
[[[[
2
< , , (13.3.2)
and, for n 2,

n
= inf
D(A)K

n1
,=0
[[[[
2
< , , (13.3.3)
where
1
j
=
ij
Ker (
i
) .
Proof.
Step 1. Let us start, with the lowest eigenvalue. Let us dene
w
idc/otj
1
()
by
j
1
() := inf
D(A),=0
[[[[
2
< , . (13.3.4)
If
1
is an eigenfunction associated to
1
, we get immediately the inequality
j
1
()
1
() . (13.3.5)
Let us prove the converse inequality. Using the spectral theorem, one get im-
mediately that inf ().
So we get
inf () j
1
() . (13.3.6)
Now, if the spectrum below is not empty, we get

1
() j
1
() .
We have consequently the equality. We have actually a little more.
We have indeed proved that, if j
1
() < , then, the spectrum
below is not empty, and the lowest eigenvalue is j
1
().
113
Step 2. The proof is by recursion, applying Step 1 to
/D(A)K

n1
.
This ends the proof of Theorem 13.3.1.
Example 13.3.2. (Payne-Polya-Weinberger Inequality.)
Let 1 = +\ with \ C

positive and \ + as [r[ +.

Let us assume that \ is even
\ (r) = \ (r) . (13.3.7)
Then
2
satises

2
inf
Q(P), odd
< 1, (13.3.8)
Let n
1
be the rst normalized eigenfunction. We admit that the lowest eigen-
value of the Schrodinger operator is simple (variant of the Krein-Rutmans The-
orem) and that the rst eigenfunction can be chosen strictly positive, with
exponential decay at together with n
1
(this is a consequence of Agmons
inequality [Ag]). Then it is not dicult to verify that n
1
is even. Let us consider

j
:= r
j
n
1
.
j
is in the form domain of 1. We observe that
1(r
j
n
1
) =
1
r
j
n
1
2
j
n
1
.
Taking the scalar product with r
j
n
1
, we then obtain
(
2

1
)[[r
j
n
1
[[
2
2 <
j
n
1
, r
j
n
1

[[n
1
[[
2
1 .
(13.3.9)
We now use the uncertainty principle (1.2.11) and get :
(
2

1
) 4[[
j
n
1
[[
2
. (13.3.10)
On the other hand,
[[n
1
[[
2
+
_
R
m
\ (r)[n
1
(r)[
2
dr =
1
, (13.3.11)
and this gives
[[n
1
[[
2

1
. (13.3.12)
Putting the inequalities (13.3.9) and (13.3.12), we get, summing over ,,

2

1

4
:

1
. (13.3.13)
This inequality is not optimal, in the sense that for : = 1 and \ (r) = r
2
, we
have
2

1
= 2
1
.
114
Example 13.3.3.
Let us consider o
h
:= /
2
+\ on R
m
where \ is a C

potential tending to
0 at and such that inf
xR
m \ (r) < 0.
Then if / 0 is small enough, there exists at least one eigenvalue for o
h
. We
note that the essential spectrum is [0, +[. The proof of the existence of this
eigenvalue is elementary. If r
min
is one point such that \ (r
min
) = inf
x
\ (r), it
is enough to show that, with
h
(r) = exp

h
[rr
min
[
2
, the quotient
<S
h

h
,
h
>
||
h
||
2
tends as / 0 to \ (r
min
) < 0.
13.4 Max-min principle
We now give a more exible criterion for the determination of the bottom of
the spectrum and for the bottom of the essential spectrum. This exibility
comes from the fact that we do not need an explicit knowledge of the various
eigenspaces.
Theorem 13.4.1.
Let H an Hilbert space of innite dimension
39
and be a selfadjoint semi-
bounded operator of domain 1() H. Let us introduce
j
n
() = sup
1,2,...,n1
inf
8
<
:
[span(
1
, . . . ,
n1
)]

;
1() and [[[[ = 1
9
=
;
[ )
H
. (13.4.1)
Then either
(a) j
n
() is the n-th eigenvalue when ordering the eigenvalues in increas-
ing order (and counting the multiplicity) and has a discrete spectrum in
] , j
n
()]
or
(b) j
n
() corresponds to the bottom of the essential spectrum. In this case, we
have j
j
() = j
n
() for all , n.
Remark 13.4.2.
In the case when the operator has compact resolvent, case (b) does not occur and
the supremum in (13.4.1) is a maximum. Similarly the inmum is a minimum.
This explains the traditional terminology Max-Min principle for this theorem.
Proof.
If is a borelian, let 1

be the projection-valued measure for (see Remark

8.3.7).
We rst prove that
dim
_
1onpc1
],a[
_
< n if o < j
n
(). (13.4.2)
dim
_
1onpc1
],a[
_
n if o j
n
() . (13.4.3)
39
In the case of a nite dimensional Hilbert space of dimension d, the minimax principle
holds for n d.
115
Notons que la conjonction de (13.4.2) et (13.4.3) montre que j
n
() est dans
le spectre de .
Step 1: Proof of (13.4.2).
Let o and n be given such that o < j
n
(). Let us prove (13.4.2) by contra-
diction. If it was false, then we would have dim
_
1onpc(1
],a[
)
_
n and
we could nd an n-dimensional space \ 1onpc(1
],a[
). Note now, that
being bounded from below, 1onpc(1
],a[
) is included in 1().
So we can nd an n-dimensional space \ 1(), such that
\ , < , o[[[[
2
. (13.4.4)
But then given any
1
, ,
n1
in H, we can nd \
1
, ,
n1

such that [[[[ = 1 and < , o. Coming back to the denition, this
shows that j
n
() o and a contradiction.
Note that we have proved in this step the following proposition
Proposition 13.4.3.
Suppose that there exists o and an n-dimensional subspace \ 1() such that
(13.4.4) is satised. Then we have the inequality :
j
n
() o . (13.4.5)
Modulo the complete proof of the theorem, we obtain
Corollary 13.4.4.
Under the same assumption as in Proposition 13.4.3, if o is below the bottom
of the essential spectrum of , then has at least n eigenvalues (counted with
multiplicity).
Exercise 13.4.5.
In continuation of Example 13.3.3, show that for any c 0 and any , there
exists /
0
0 such that for / ]0, /
0
], o
h
has at least eigenvalues in
[inf \, inf \ +c]. One can treat rst the case when \ has a unique non degenerate
minimum at 0.
Step 2 : Proof of (13.4.3).
Suppose that (13.4.3) is false. Then dim ( Range (1
],a[
) n 1, so we
can nd (n 1) generators
1
, ,
n1
of this space. Then any 1()
:jon
1
, ,
n1

is in Range (1
[a,+[
), so
< , o[[[[
2
.
Therefore, coming back to the denition of j
n
(), we get j
n
() o in contra-
diction with our initial assumption.
Before to continue the proof, let us emphasize on one point.
116
Remark 13.4.6.
In the denition of j
n
(),
1
, ,
n1
are only assumed to belong to the
Hilbert space H.
Step 3 : j
n
() < +.
First the semi-boundedness from below of gives a uniform lower bound.
Secondly, if j
n
() = +, this would mean by (13.4.2) that :
dim(1onpc(1
],a[
)) < n for all o,
and consequently that H is nite dimensional. This is a contradiction, if H is
innite dimensional. But the nite case is trivial, we have indeed j
n
() [[[[,
in this case.
As the statement of the theorem suggests, there are two cases to consider
and this will be the object of the two next steps.
Step 4.
Let us rst assume (with j
n
= j
n
()) that
dim (1onpc(1
],n+[
)) = + , c 0 . (13.4.6)
We claim that, in this case, we are in the second alternative in the theorem.
Using (13.4.2) and (13.4.6), we get indeed
dim (1onpc(1
]n,n+[
)) = + , c 0 . (13.4.7)
This shows that j
n
()
ess
().
On the other hand, using again (13.4.2), we immediately get that ] , j
n
()[
does not contain any point in the essential spectrum. Thus j
n
() = inf [

ess
().
Let us show now that j
n+1
= j
n
in this case. From the denition of the
j
k
(), it is clear that j
n+1
j
n
, since one can take
n
=
n1
.
But if j
n+1
j
n
, (13.4.2) would also be satised for j
n+1
, and this is in
contradiction with (13.4.6).
Step 5.
Let us now assume that
dim (1onpc(1
],n+0[
)) < , for some c
0
0 . (13.4.8)
Then it is clear, that the spectrum is discrete in ] , j
n
+c
0
[. Then, for c
1
0
small enough,
1onpc(1
],n]
) = 1onpc(1
],n+1[
) ,
and by (13.4.3)
dim (1onpc1
],n]
) n . (13.4.9)
So there are at least n eigenvalues 1
1
1
2
1
n
j
n
for . If 1
n
were
strictly less than j
n
, dim(1onpc1
],En]
) would equal n in contradiction with
117
(13.4.2). Therefore 1
n
= j
n
and j
n
is an eigenvalue.
This ends the proof of Theorem 13.4.1.
A rst natural extension of Theorem 13.4.1 is obtained by
Theorem 13.4.7.
Let be a selfadjoint semibounded operator and Q() its form domain
40
.
Then
j
n
() = sup
1,2,...,n1
inf
8
<
:
[span(
1
, . . . ,
n1
)]

;
Q() and [[[[ = 1
9
=
;
[ )
H
. (13.4.10)
Proof.
Let j
n
be the right hand side of (13.4.10). By imitating the proof of the previous
theorem, we get that each j
n
obeys one of the two conditions. These conditions
determine j
n
and consequently j
n
= j
n
.
One can also note (see Subsection 3.3) that, when constructing the Friedrichs
extension, one has shown that the domain of the Friedrichs extension is dense
in the form domain.
Applications
It is very often useful to apply the max-min principle by taking the mini-
mum over a dense set in Q().
The max-min principle permits to control the continuity of the eigenvalues
with respect to parameters. For example the lowest eigenvalue
1
(c) of

d
2
dx
2
+r
2
+cr
4
increases with respect to c. Show that c
1
(c) is right
continuous on [0, +[. (The reader can admit that the corresponding
eigenfunction is in o(R) for c 0).
The max-min principle permits to give an upperbound on the bottom of
the spectrum and the comparison between the spectrum of two operators.
If 1 in the sense that, Q(1) Q() and
41
< n, n < 1n, n , n Q(1) ,
then

n
()
n
(1) .
Similar conclusions occur if we have 1(1) 1().
40
associated by completion with the form u u[Au)
H
initially dened on D(A).
41
It is enough to verify the inequality on a dense set in Q(B).
118
Example 13.4.8. (Comparison between Dirichlet and Neumann)
Let be a bounded regular connected open set in R
m
. Then the -th eigenvalue
of the Neumann realization of + \ is less or equal to the -th eigenvalue
of the Dirichlet realization. It is indeed enough to observe the inclusion of the
form domains.
Example 13.4.9. (monotonicity with respect to the domain)
Let
1

2
R
m
two bounded regular open sets. Then the n t/ eigenvalue
of the Dirichlet realization of the Schrodinger operator in
2
is less or equal to
the n-th eigenvalue of the Dirichlet realization of the Schrodinger operator in

1
. We observe that we can indeed identify H
1
0
(
1
) with a subspace of H
1
0
(
2
)
by just an extension by 0 in
2

1
.
We then have

n
(
2
) = sup
{1, ,n1L
2
(2)}
inf
8
<
:
H
1
0
(
2
)
< ,
j

L
2
(2) and ||||=1
9
=
;
[[[[
2
L
2
(2)
sup
{1, ,n1L
2
(2)}
inf
8
<
:
H
1
0
(
1
)
< ,
j

L
2
(2) and ||||=1
9
=
;
[[[[
2
L
2
(2)
= sup
{1, ,n1L
2
(2)}
inf
8
<
:
H
1
0
(
1
)
< ,
j

L
2
(1) and ||||=1
9
=
;
[[[[
2
L
2
(1)
= sup
{1, ,n1L
2
(1)}
inf
8
<
:
H
1
0
(
1
)
< ,
j

L
2
(1) and ||||=1
9
=
;
[[[[
2
L
2
(1)
=
n
(
1
) .
Note that this argument is not valid for the Neumann realization.
13.5 CLR inequality
In order to complete the picture, let us mention (confer [RS-IV], p. 101) that,
if : 3, then the following theorem due to Cwickel-Lieb-Rozenbljum is true :
Theorem 13.5.1.
There exists a constant 1
m
, such that, for any \ such that \

1
m
2
, and
if : 3, the number of strictly negative eigenvalues of o
1

is nite and
bounded by

1
m
_
V (x)0
(\ )
m
2
dr . (13.5.1)
This shows that when : 3, we could have examples of negative potentials
\ (which are not identically zero) and such that the corresponding Schrodinger
operator o
1
has no eigenvalues. A sucient condition is indeed
1
m
_
V 0
(\ )
m
2
dr < 1 .
In the other direction, we have
42
the following results.
42
These counterexamples come back (when m = 1 to Avron-Herbst-Simon [AHS] and when
m = 2 to Blanchard-Stubbe [BS]).
119
Proposition 13.5.2.
Let \ be in C

0
(R
m
) (: = 1, 2). If
_
R
m
\ (r) dr < 0 , (13.5.2)
then
Proof.
We just treat the case when \ C

0
(R
m
).
We rst observe that the the essential spectrum is [0, +[. For the proof of the
proposition, it is then enough to nd 1(o
1
) such that
< o
1
,
L
2
(R
m
)
< 0 .
When : = 1, taking
a
= expo[r[, o 0, we nd that
_
R
[

a
(r)[
2
dr = o ,
and
lim
a0
_
R
\ (r)[
a
(r)[
2
dr =
_
R
\ (r) dr < 0 ,
by the dominated convergence Theorem.
When : = 2, we can take
a
(r) = exp
1
2
[r[
a
, o 0, then
_
R
2
[[
a
(r)[[
2
dr =

2
o ,
and
lim
a0
_
R
2
\ (r)[
a
(r)[
2
dr = c

1
2
_
R
2
\ (r) dr < 0 .
13.6 Essential spectrum and Perssons Theorem
We refer to Agmons book [Ag] for details.
Theorem 13.6.1.
Let \ be a real-valued potential in the Kato-Rellich class
43
, and let H = +\
be the corresponding self-adjoint, semibounded Schrodinger operator with do-
main H
2
(R
m
). Then, the bottom of the essential spectrum is given by
inf
ess
(H) = (H) , (13.6.1)
where
(H) := sup
KR
m
_
inf
||||=1
< , H [ C

0
(R
m
/)
_
, (13.6.2)
where the supremum is over all compact subset / R
m
.
43
Cf Theorem 10.3.1.
120
Essentially this is a corollary of Weyls Theorem 12.5.5. We will indeed play
with the fact that
Lemma 13.6.2.

ess
(H) =
ess
(H +\) ,
for any regular potential \ with compact support.
121
14 Exercises and Problems
We present in this section some exercises or problems proposed in the last years.
They sometimes strongly intersect with the course.
14.1 Exercises
Exercise 14.1.1 (a natural problem in Bose-Einstein theory).
Let 0.
Discuss in function of R the semi-boundedness of the operator dened on
S(R
2
) by
H

:=
1
2

x,y
+
1
2

2
:
2
1
z
, (14.1.1)
with
1
z
= i(r
y
j
x
) , (14.1.2)
The answer can be found by showing that

j,k
(r, j) = c

2
(x
2
+y
2
)
(
x
+i
y
)
j
(
x
i
y
)
k
_
c
(x
2
+y
2
)
_
(14.1.3)
where , and / are non-negative integers, is an eigenfunction of H

and of 1
z
.
Exercise 14.1.2. (After Eros, Avron-Seiler-Simon).
Let 1 and Q two selfadjoint projectors in a Hilbert space H. i) Let us assume
that = 1 Q is compact. Show that if ,= 1 is in the spectrum, then
is in the spectrum with the same multiplicity. For this, one can rst show that
with 1 = 1 1 Q,

2
+1
2
= 1 , 1 +1 = 0 .
ii) Assume now that is in addition trace class, that is that the series [j
j
[,
where j
j
are the non zero eigenvalues of , counted with multiplicity. Compute
Tr :=

j
j
j
and show that it is an integer.
Exercise 14.1.3. (Temples inequality).
Let be a selfadjoint operator on an Hilbert space and 1() such that
[[[[ = 1.
Suppose that in some interval ], [, ()], [= and that = [ )
belongs to the interval ], [. Then show that :

c
2

+
c
2

.
with :
c
2
= [[()[[
2
.
As a preliminary result, one can show that ()() and ()()
are positive operators. Then apply the inequalities with .
Show that this inequality is an improvment if c
2
( )( ).
Compare with what is given by the spectral theorem or the minmax principle.
122
Exercise 14.1.4. .
Let (r
1
, r
2
) = (
1
(r
1
, r
2
) ,
2
(r
1
, r
2
)) be a C

vector eld on R
2
. Let \ be
a C

positive function on R
2
.
Let 1 := (1
x1

1
(r
1
, r
2
))
2
+ (1
x2

2
(r
1
, r
2
))
2
+ \ (r) the dierential
operator dened on C

0
(R
2
).
a) Show that 1 admits a selfadjoint extension in 1
2
(R
2
).
b) Show that 1 is essentially selfadjoint.
Exercise 14.1.5. .
We admit the results of Exercise 14.1.4. Show that the selfadjoint extension in
1
2
(R
2
) of
T := (
d
dr
1
ir
2
r
2
1
)
2

d
2
dr
2
2
+r
2
2
,
is with compact resolvent.
Exercise 14.1.6. .
Let \ be a C

positive potential in R
2
. Let us consider, with 1 R 0, the
operator
1 = 1
2
x1
+ (1
x2
+1r
1
)
2
+\ (r) .
a) Recall briey the spectrum of its selfadjoint extension in the case \ = 0.
b) We assume that \ tends to 0 as [r[ +. Determine the essential spectrum
of

1.
Exercise 14.1.7. .
Let 1 be a kernel in o(R
2
) which is strictly positive and symmetric.
a) Show that the associated operator / which is dened on o(R) by
(/n)(r) =
_
R
1(r, j)n(j)dj ,
can be extended as a compact operator on 1
2
(R). b) Let 1 be an open interval
in R and let us denote by /
I
the operator on 1
2
(1) dened by
(/
I
n)(r) =
_
I
1(r, j)n(j)dj .
Let
1
I
be the largest eigenvalue of /
I
. Show that

1
I

1
R
.
Show that we have strict inequality when 1 is not R.
c) Let n
1
be a normalized eigenfunction of / associated with
1
R
. Using its
restriction to 1, show the inequality :

1
R

1
I
(1 [[n
1
[[
L
2
(I
C
)
)
1
.
d) Let 1
n
= [n, n]. Show that
1
In
converges rapidly to
1
R
as n +. More
precisely, show that, for all , N, there exists a constant C
j
such that :
[
1
R

1
In
[ C
j
n
j
, n N

.
123
Exercise 14.1.8. .
Let us consider in =]0, 1[R, a positive C

function \ and let o

0
be the
Schrodinger operator o
0
= +\ dened on C

0
().
(a) Show that o
0
admits a selfadjoint extension on 1
2
(). Let o this extension.
(b) Determine if o is with compact resolvent in the following cases :
1. \ (r) = 0,
2. \ (r) = r
2
1
+r
2
2
,
3. \ (r) = r
2
1
,
4. \ (r) = r
2
2
5. \ (r) = (r
1
r
2
)
2
.
Determine the spectrum in the cases (1) and (4). One can rst determine the
spectrum of the Dirichlet realization (or of Neumann) of d
2
,dr
2
on ]0, 1[.
Exercise 14.1.9. .
We consider in R
2
the operator dened on o(R
2
; C
2
) by
T
0
=
1
1
x1
+
2
1
x2
+
3
.
Here the matrices
j
are hermitian 2 2 matrices such that :

j
+
j

i
= 2
ij
,
and we recall that 1
xj
=
1
i

x
j
for , = 1, 2.
a) Is T
0
symmetric? semi-bounded ? It is suggested to use the Fourier
transform.
b) Compute T
2
0
.
c) Show that T
0
admits a selfadjoint extension T
1
in 1
2
(R
2
; C
2
), and de-
termine its domain.
d) Determine the spectrum of T
1
.
e) We suppose that, for all r R
2
, \ (r) is a 2 2 hermitian matrix, with
bounded C

coecients. Show that T

V
= T
0
+\ admits a selfadjoint extension
and determine its domain.
Exercise 14.1.10. .
Let H
a
be the Dirichlet realization of d
2
,dr
2
+r
2
in ] o, +o[. Show that the
lowest eigenvalue
1
(o) of H
a
is strictly positive, monotonically decreasing as
o + and tend exponentially fast to 1 as o +. Give an estimate as ne
as possible of [
1
(o) 1[.
Exercise 14.1.11. .
We consider on C

0
(R
2
), the operator
1
0
:= (1
x1
r
2
)
2
+ (1
x2
r
1
)
2
.
We recall that 1
xj
=
1
i

xj
.
Show that its natural selfadjoint extension 1 is unitary equivalent to the op-
erator (of domain H
2
). Determine its spectrum and its essential spectrum.
124
Exercise 14.1.12. .
Show that one can associate to the dierential operator on C

0
(R]0, 1[) :
T
0
:= (1
x1
r
2
r
2
1
)
2
+ (1
x2
)
2
,
an unbounded selfadjoint operator T on 1
2
(R]0, 1[) whose spectrum is with
compact resolvent.
Exercise 14.1.13. .
Let be a C
2
- function on R
m
such that [(r)[ + as [r[ + and with
uniformly bounded second derivatives. Let us consider the dierential operator
on C

0
(R
2
) +2 . We consider this operator as an unbounded operator
on H = 1
2
(R
m
, exp2 dr). Show that it admits a selfadjoint extension and
that its spectrum is discrete.
We assume in addition that :
_
R
m
exp2dr < +. Show that its lowest
eigenvalue is simple and determine a corresponding eigenvector.
Exercise 14.1.14. .
We consider in R
3
the dierential operator o
0
:=
1
r
, a priori dened on
C

0
(R
3
).
a) Show that the operator admits a selfadjoint extension o.
b) Show the continuous injection of H
2
(R
3
) into the space of the Holder func-
tions C
s
(R
3
), with : ]0,
1
2
[, and the compact injection for all compact 1 of
C
s
(1) into C
0
(1).
c) Determine the essential spectrum of o. One possibility is to start with the
analysis of o

=

r
where is C

with compact support.

d) Show using the minimax-principle that o has at least one eigenvalue. One
can try to minimize over o n < o
0
n , n),[[n[[
2
with n(r) = expo:.
e) Determine this lowest eigenvalue (using the property that the groundstate
should be radial).
Exercise 14.1.15. .
a) Let p be a continuous function on R such that p(0) = 0. Analyze the conver-
gence of the sequence (p(t)n
n
(t))
n1
in 1
2
(R) where n
n
(t) =

n(nt) and is
a C

function with compact support.

b) let ) C
0
([0, 1]; R). Let T
f
be the multiplication operator by ) dened on
1
2
(]0, 1[): n T
f
n = )n.
Determine the spectrum of T
f
. Discuss in function of ) the possible existence
of eigenvalues. Determine the essential spectrum of T
f
.
Exercise 14.1.16. .
Discuss in function of 0 the possibility of associating to the dierential
operator dene on C

0
(R
3
)
:

a selfadjoint operator on 1
2
(R
3
).
Exercise 14.1.17. .
Let be a non empty open subset in R
d
and let us consider the multiplication
125
operator on 1
2
(R
d
) dened by the multiplication by

where

is equal to 1 in
and 0 outside. Determine the spectrum, the essential spectrum, the discrete
spectrum.
Exercise 14.1.18. .
Show that the spectrum in R
2
of 1 = 1
2
x
+r
2
+1
2
y
is [1, +[.
Exercise 14.1.19. .
Let \ o(R
d
) be positive and let us consider the operator :
T := ( + 1)

1
2
\ ( + 1)

1
2
.
a) Explain how to dene ( + 1)

1
2
, as an operator on 1
2
(R
d
).
b) Show that T is a bounded, selfadjoint, positive, compact operator on 1
2
(R
d
).
c) Discuss the injectivity in function of \ ?
d) Establish a link with the research of pairs (n, j) in H
2
(R
d
) R
+
such that :
( + 1 j\ )n = 0 .
Exercise 14.1.20. .
Let R.
(a) Show that the operator 1

dened on C

0
(R
2
) by
1

:= 1
2
x
+1
2
y
+r
2
j
4
+r
4
j
2
+(r +j)
is semibounded.
One can rst show the inequality :
1
0
n , n) [[rn[[
2
+[[jn[[
2
.
(b) Show that there exists a natural selfadjoint extension of 1

.
(c) What is the corresponding form domain ?
(d) Show that the selfadjoint extension is with compact resolvent.
Exercise 14.1.21. .
Let us consider in R
+
, the Neumann realization in R
+
of 1
0
() := 1
2
t
+(t)
2
,
where is a parameter in R. We would like to nd an upper bound for
0
=
inf

j() where j() is the smallest eigenvalue of 1

0
(). Following the physicist
Kittel, one can proceed by minimizing 1
0
()(; ) [ (; )) over the normalized
functions (t; ) := c

expt
2
( 0). For which value of is this quantity
minimal ?? Deduce the inequality :

0
<
_
1
2

.
Exercise 14.1.22. .
In the same spirit as in the previous exercise. Find an upper bound for the
quartic operator 1
2
t
+
1
4
t
4
.
Using the comparison with an harmonic oscillator 1
2
t
+t
2
+, nd an optimal
126
(with respect to the method) lower bound.
Using the comparison with 1
2
t
+\

(t) where \

(t) = 0 for [t[ and \

(t) =
1
4

4
for [t[ , and optimizing over nd an alternative lower bound (and
compute it with the help of a computer).
Exercise 14.1.23.
44
Let be a bounded regular set in R
d
. Let us denote by
n
(n 1), (resp. j
n
)
the sequence of eigenvalues of the Dirichlet problem for the Laplacian (resp. the
Neumann problem). Show that
j
n+1

n
, n 1 .
One can use the minimax principle and analyze the quantity
_

[n[
2
dr for
n =
n

j=1

D
j
(r) + expi r,
with [[
2
=
n
.
Is the inequality strict ?
We now consider the function
() =
_

expi rdr.
Show that if is balanced (that is if r i r ), then is real.
We denote by () the function
() = inf R
d
, () = 0 .
Show that if
_

n
() 2() then
j
n+2
()
n
() .
This time one can use the minimax principle and analyze the quantity
_

[n[
2
dr
for
n =
n

j=1

D
j
(r) +
1
expi
1
r +
2
expi
2
r,
with [
1
[
2
= [
2
[
2
=
n
and (
1

2
) = 0.
44
This is inspired by a paper of Benguria-Levitin-Parnovski [?] and by discussions with A.
Laptev
127
14.2 Problems
Problem 14.2.1. .
Let us consider in the disk of R
2
:= 1(0, 1) the Dirichlet realization of the
Schrodinger operator
o(/) := +
1
/
2
\ (r) , (14.2.1)
where \ is a C

potential on

satisfying :
\ (r) 0 . (14.2.2)
Here / 0 is a parameter.
a) Show that this operator is with compact resolvent.
b) Let
1
(/) be the lowest eigenvalue of o(/). We would like to analyze the
behavior of
1
(/) as / 0. Show that /
1
(/) is monotonically increasing.
c) Let us assume that \ 0 on

; show that there exists c 0 such that
/
2

1
(/) c . (14.2.3)
d) We assume now that \ = 0 in an open set in . Show that there exists a
constant C 0 such that, for any / 0,

1
(/) C . (14.2.4)
One can use the study of the Dirichlet realization of in .
e) Let us assume that :
\ 0 almost everywhere in . (14.2.5)
Show that, under this assumption :
lim
h0

1
(/) = + . (14.2.6)
One could proceed by contradiction supposing that there exists C such that

1
(/) C , / such that 1 / 0 . (14.2.7)
and establishing the following properties.
For / 0, let us denote by r n
1
(/)(r) an 1
2
-normalized eigenfunction
associted with
1
(/). Show that the family n
1
(/) (0 < / 1) is bounded
in H
1
().
Show the existence of a sequence /
n
(n N) tending to 0 as n + and
n

1
2
() such that
lim
n+
n
1
(/
n
) = n

in 1
2
().
128
Deduce that : _

\ (r) n

(r)
2
dr = 0 .
Deduce that n

= 0 and make explicite the contradiction.

f ) Let us assume that \ (0) = 0; show that there exists a constant C, such
that :

1
(/)
C
/
.
g) Let us assume that \ (r) = O([r[
4
) pr`es de 0. Show that in this case :

1
(/)
C
/
2
3
.
h) We assume that \ (r) [r[
2
near 0; discuss if one can hope a lower bound
in the form

1
(/)
1
C /
.
Justify the answer by illustrating the arguments by examples and counterexam-
ples.
Problem 14.2.2. .
We consider on R and for c 1 := [
1
4
, +[ the operator H

= d
2
,dr
2
+
r
2
+c[r[.
a) Determine the form domain of H

and show that it is independent of c.

b) What is the nature of the spectrum of the associated selfadjoint operator?
c) Let
1
(c) the smallest eigenvalue. Give rough estimates permitting to estimate
from above or below
1
(c) independently of c on every compact interval of 1.
d) Show that, for any compact sub-interval J of 1, there exists a constant C
J
such that, for all c J, any 1
2
-normalized eigenfunction n

of H

associated
with
1
(c) satises :
[[n

[[
B
1
(R)
C
J
.
For this, on can play with : H

, n

)
L
2
(R)
.
e) Show that the lowest eigenvalue is a monotonically increasing sequence of
c 1.
f ) Show that the lowest eigenvalue is a locally Lipschitzian function of c 1.
One can use again the max-min principle.
g) Show that (c) +, as c + and estimate the asymptotic behavior.
h) Discuss the same questions for the case H

= d
2
,dr
2
+r
2
+cr
4
(with c 0).
Problem 14.2.3. .
The aim of this problem is to analyze the spectrum
D
(1) of the Dirichlet real-
ization of the operator 1 := (1
x1

1
2
r
2
)
2
+ (1
x2
+
1
2
r
1
)
2
in R
+
R.
1. Show that one can a priori compare the inmum of the spectrum of 1 in
R
2
and the inmum of
D
(1).
129
2. Compare
D
(1) with the spectrum
D
(Q) of the Dirichlet realization of
Q := 1
2
y1
+ (j
1
j
2
)
2
in R
+
R.
3. We rst consider the following family of Dirichlet problems associated with
the family of dierential operators : H() dened on ]0, +[ by :
H() = 1
2
t
+ (t )
2
.
Compare with the Dirichlet realization of the harmonic oscillator in ]
, +[.
4. Show that the lowest eigenvalue () of H() is a monotonic function of
R.
5. Show that () is a continuous function on R.
6. Analyze the limit of () as .
7. Analyze the limit of () as +.
8. Compute (0). For this, one can compare the spectrum of H(0) with the
spectrum of the harmonic oscillator restricted to the odd functions.
9. Let t n(t; ) the positive 1
2
-normalized eigenfunction associated with
(). Let us admit that this is the restriction to R
+
of a function in o(R).
Let, for R, T

be the distribution in T

(R
+
R) dened by
T

() =
_
+
0
(j
1
, )n

(j
1
)dj
1
.
Compute QT

.
10. By constructing starting from T

a suitable sequence of 1
2
-functions tend-
ing to T

, show that ()
D
(Q).
11. Determine
D
(1).
Problem 14.2.4. .
Let H
a
be the Dirichlet realization of d
2
,dr
2
+r
2
in ] o, +o[.
(a) Briey recall the results concerning the case o = +.
(b) Show that the lowest eigenvalue
1
(o) of H
a
is decreasing for o ]0, +[
and larger than 1.
(c) Show that
1
(o) tends exponentially fast to 1 as o +. One can use a
suitable construction of approximate eigenvectors.
(d) What is the behavior of
1
(o) as o 0. One can use the change of variable
r = oj and analyze the limit lim
a0
o
2

1
(o).
(e) Let j
1
(o) be the smallest eigenvalue of the Neumann realization in ] o, +o[.
Show that j
1
(o)
1
(o).
(f ) Show that, if n
a
is a normalized eigenfunction associated with j
1
(o), then
there exists a constant C such that, for all o 1, we have :
[[rn
a
[[
L
2
(]a,+a[)
C .
130
(g) Show that, for n in C
2
([o, +o]) and in C
2
0
(] o, +o[), we have :

_
+a
a

2
n

(t)n(t)dt =
_
+a
a
[(n)

(t)[
2
dt
_
+a
a

(t)
2
n(t)
2
dt .
(h) Using this identity with n = n
a
, a suitable which should be equal to 1 on
[o + 1, o 1] , the estimate obtained in (f ) and the minimax principle, show
that there exists C such that, for o 1, we have :

1
(o) j
1
(o) +Co
2
.
Deduce the limit of j
1
(o) as o +.
Problem 14.2.5.
We consider, for c R, the dierential operator which is dened on C

0
(R
2
)
by

0,c
= 1
2
x
+ (1
y

1
2
r
2
)
2
+icj ,
and considered as an unbounded operator on 1
2
(R
2
). We recall that 1
x
= i
x
et 1
y
= i
y
. The problem consists in analyzing the spectral properties of
0,c
as a function of c R.
Part I
For R, we consider the unbounded operator on 1
2
(R) h
0
() with domain
C

0
(R) associated with the dierential operator
h
0
() :=
d
2
dr
2
+ (
r
2
2
)
2
.
Ia) Show that, for any , one can construct a selfadjoint extension of h
0
() and
describe its domain.
Ib) Show that the operator has compact resolvent and deduce that the spectrum
consists in a sequence Montrer que loperateur e of eigenvalues
j
() (, N

)
tending to +.
Ic) Show that lim

1
() = +.
Id) Show that
1
() 0.
Ie) Show that
1
() is continuous.
If ) Show that
j
() is monotonically decreasing for < 0.
Ig) We admit that lim
+

1
() = +. Show that
1
() attains its
inmum

for (at least) one point.

Ih) We admit that
1
() is of multiplicity 1, of class C
1
and that one
can, for any , associate with
1
() an eigenfunction n
1
(, ) in o(R) with C
1
dependence of such that [[n
1
[[ = 1 and n
1
0.
Show that

1
() = 2
_
R
(
r
2
2
) n
1
(r, )
2
dr .
131
Deduce that the critical points
c
of
1
satisfy
c
0.
Ii) Show that n
1
(, ) is even.
Ij) Show that, if
c
is a critical point of
1
, then
1(
c
) :=
_
+
0
r(
r
2
2

c
) n
1
(r,
c
)
2
dr 0 ,
Ik) Computing dierently 1(), deduce that

2
c

1
(
c
) .
Ik) Using Gaussian quasimodes for h
1
(0) determine an interval (as good as pos-
sible) in which
1
should have its minimum.
Part II
Here we suppose that c = 0 and we write
0
for
0,0
.
IIa) Show that the operator
0
is symmetric.
IIb) Show that one can construct its Friedrichs extension
Fried
0
, associated with
a sesquilinear form to be dened precisely.
Describe the form domain and the domain of the operator.
IIc) Show that the operator is essentially selfadjoint.
IId) Show that
(
Fried
0
) [

, +[ .
It is suggested to use a partial Fourier transform with respect to j.
IIe) By constructing suitable families of approximate eigenfunctions, show that
(
Fried
0
) = [

, +[
IIf ) Show that
Fried
0
is not with compact resolvent.
Part III
We now suppose that c ,= 0.
IIIa) Show that
Re
0,c
n, n) =
0
n, n) ,
and that
Im
0,c
n, n) = cjn, n) ,
for all n 1(
0,c
).
IIIb) Show that
Re
0,c
n, n)

[[n[[
2
,
for all n 1(
0,c
).
IIIc) Show that
0,c
is closable. We denote by
0,c
its closure. Recall how this
operator is dened and describe its domain.
IIId) Show that, for

,
0,c
+ is injective and with closed range.
IIIe) Show that for

,
0,c
+ has dense range. It is suggest to adapt
132
the proof of IIc.
IIIf ) Show that
(
o,c
) C, Re

.
The aim of questions IIIg to IIIk is to show that
o,c
has, for c ,= 0,
a compact resolvent.
IIIg) Show that, for any compact set 1 R
2
, there exists a constant C
K
such
that, for any n C

0
(R
2
) with compact support in 1, we have
[[n[[
2
H
1
(R
2
)
C
K
_
Re
0,c
n, n) +[[n[[
2
L
2
_
.
IIIh) Show that, for all n C

0
(R
2
), with compact support in j 0 or in
j < 0, we have
_
[j[[n(r, j)[
2
drdj
1
[c[
[ Im
0,c
n, n)[ .
By using a partition of unity, deduce that
[[[j[
1
2
n[[
2
C
_
[[
0
n[[
2
L
2 +[[n[[
2
L
2
_
.
IIIi) Show that, for all n C

0
(R
2
), with compact support in r 0 or in
r < 0, we have
_
[r[[n(r, j)[
2
drdj Re
0,c
n, n)[ .
By using a partition of unity, deduce that there exists a constant C such that,
for all n C

0
(R
2
), we have
[[[r[
1
2
n[[
2
C
_
[[
0
n[[
2
L
2 +[[n[[
2
L
2
_
.
IIIj) Deduce that
0,c
has a compact resolvent.
IIIk) Show that if (n, ) is a spectral pair for
0,c
, then, for all o R, (n
a
,
ico), where n
a
is dened by n
a
(r, j) = n(r, j +o), is a spectral pair.
IIIl) Deduce that the spectrum of
0,c
is empty.
IIIm) Show that
0,c
=

0,c
.
133
References
[Ag] S. Agmon : Lecture on exponential decay of solutions of second order
elliptic equations. Math. Notes, t.29, Princeton University Press.
[A-G] N. I. Akhiezer, I. M. Glazman. Theory of linear operators in Hilbert
space, Pitman, Londres, 1981.
[AHS] J. Avron, I. Herbst, and B. Simon : Schrodinger operators with magnetic
elds I. Duke Math. J. 45, p. 847-883 (1978).
[BGM] M. Berger, P. Gauduchon, and E. Mazet : Spectre dune variete rie-
mannienne, Springer Lecture Notes in Mathematics 194 (1971).
[Br] H. Brezis : Analyse Fonctionnelle. Editions Masson.
[BLP] R. Benguria, M. Lewitin, and L. Parnovski. To be completed.
[BeSt] A. Berno, P. Sternberg : Onset of superconductivity in decreasing elds
for general domains, J. Math. Phys. 39 (1998), p. 1272-1284.
[BS] P. Blanchard, J. Stubbe : Bound states for Schrodinger Hamiltonians :
phase space methods and applications. Rev. in Math. Physics (1997).
[CDV] Y. Colin de Verdi`ere : Spectres de Graphes. Cours Specialises, No 4,
SMF (1998).
[CFKS] H.L. Cycon, R. Froese, W. Kirsch, and B. Simon : Schrodinger opera-
tors. Springer (1987).
[DaHe] M. Dauge, B. Heler : Eigenvalues variation I, Neumann problem for
Sturm-Liouville operators. Journal of Dierential Equations, Vol. 104, n

2,
august 1993, p. 243-262.
[DaLi] R. Dautray, J.-L. Lions : Analyse mathematique et calcul numerique
pour les sciences et les techniques. Tome 2, Chap. VIII, Masson.
[Da] E.B. Davies : Spectral theory and dierential operators. Cambridge studies
in advanced Mathematics.
[Di] J. Dieudonne : Calcul innitesimal. Hermann.
[DiSj] M. Dimassi, J. Sjostrand : Spectral Asymptotics in the Semi-classical
limit. London Mathematical Society, Lecture Note Series 268. Cambridge
University Press 1999.
[GlJa] J. Glimm, A. Jae : Quantum physics (a functional integral point of
view). Springer Verlag, Second edition (1987).
[GrSj] A. Grigis, J. Sjostrand : Microlocal analysis for dierential operators.
An introduction. London Mathematical Society Lecture Note Series, 196.
Cambridge University Press, Cambridge, 1994.
134
[He-I] B. Heler : Theorie spectrale pour des operateurs globalement elliptiques.
Asterisque n
0
112 (1984).
[He-II] B. Heler : Introduction to the semiclassical analysis for the Schrodinger
operator and applications. Springer lecture Notes in Math., n
0
1336 (1988).
[He-III] B. Heler : Semiclassical analysis for Schrodinger operators, Laplace
integrals and transfer operators in large dimension: an introduction. Cours
de DEA, Paris Onze Edition (1995).
[He-IV] B. Heler : Semiclassical analysis and statistical mechanics.
Cours PostDEA. Ecole doctorale de Mathematiques et de Mecanique de
lUniversite Paul Sabatier. December 1998.
[He-V] B. Heler : Bouteilles magnetiques et supraconductivite (dapr`es Heler-
Morame, Lu-Pan et Heler-Pan). Seminaire EDP de lecole Polytechnique
2000-2001.
[HelLaf] B. Heler, O. Latte : On spectral questions around the Rayleigh
equation. To appear in Asymptotic Analysis (2003).
[HeMo1] B. Heler, A. Mohamed : Semiclassical analysis for the ground state
energy of a Schrodinger operator with magnetic wells, Journal of Functional
Analysis 138, N
o
1 (1996), p. 40-81.
[HeMo2] B. Heler, A. Morame : Magnetic bottles in connection with super-
conductivity. Journal of Functional Analysis, Vol. 185, No 2, October,
p. 604-680 (2001).
[HeMo3] B. Heler, A. Morame : Magnetic bottles in connection with super-
conductivity: Case of dimension 3. Proc. Indian Acad. Sci. (Math. Sci.)
Vol. 112, No. 1, February 2002.
[HeMo4] B. Heler, A. Morame : Magnetic bottles for the Neumann problem:
Curvature eects in the case of dimension 3 (general case). Preprint mp arc
02-145 (2002). Annales Ecole Normale Superieure 2004.
[HeSj] B. Heler, J. Sjostrand : Multiple wells in the semiclassical limit I.
Comm. in PDE 9(4), p. 337-408, (1984).
[HiSi] P.D. Hislop, I.M. Sigal : Introduction to Spectral Theory With
applications to Schrodinger Operators Applied Mathematical Sciences
113. Springer (1995).
[Ho] L. Hormander : The analysis of linear partial dierential operators,
Springer Verlag.
[Hu] D. Huet : Decomposition spectrale et operateurs, Presses universitaires
de France (1976).
135
[Ka] T. Kato : Perturbation theory for linear operators, New York, Springer-
Verlag, (1966).
[Le-Br] P. Levy-Bruhl : Introduction `a la theorie spectrale. Editions Dunod
(2003).
[LiLo] E. Lieb, M. Loss : Analysis. Graduate Studies in Mathematics, Vol. 14,
American Mathematical Society.
[Lio1] J.L. Lions : Lecture on elliptic partial dierential equations. Tata insti-
tute of fundamental research. Bombay (1957).
[Lio2] J.L. Lions : Probl`emes aux limites dans les EDP. Seminaire de
Mathematiques superieures de luniversite de Montreal. (1962).
[LiMa] J.L. Lions, E. Magenes : Probl`emes aux limites non-homog`enes. Tome 1.
Editions Dunod.
[RS-I] M. Reed, B. Simon : Methods of modern mathematical Physics. I.
Functional analysis. Academic Press, (1972).
[RS-II] M. Reed, B. Simon : Methods of modern mathematical Physics. II.
Fourier analysis, selfadjointness. Academic Press, (1975).
[RS-III] M. Reed, B. Simon : Methods of modern mathematical Physics. III.
Academic Press, (1976).
[RS-IV] M. Reed, B. Simon : Methods of modern mathematical Physics. IV.
Analysis of operators. Academic Press, (1978).
[Ro] D. Robert : Autour de lapproximation semi-classique. Birkhauser, Vol.68,
Progress in Mathematics, (1987).
[RoSi] S. Roch and B. Silbermann. C

-algebras techniques in numerical analy-

sis. J. Oper. Theory 35, p. 241-280 (1996).
[Ru1] W. Rudin : Real and Complex Analysis. Mc Graw Hill, New York (1974).
[Ru2] W. Rudin : Analyse fonctionnelle. Ediscience international (1997).
[Sib] Y. Sibuya : Global theory of a second order linear ordinary dierential
equation with a polynomial coecient. North holland (1975).
[Sima] C.G. Simader : Essential self-adjointness of Schrodinger operators
bounded from below. Math. Z. 159, p. 47-50 (1978).
[Si] B. Simon : Functional integration and quantum physics. Pure and Applied
mathematics n
0
86, Academic press, New york, (1979).
[Sj] J. Sjostrand. Pseudospectrum for dierential operators. Seminaire `a lEcole
Polytechnique, Exp. No. XVI, Semin. Equ. Deriv. Partielles, Ecole Poly-
tech., Palaiseau (2003).
136
[Tr1] L.N. Trefethen. Pseudospectra of linear operators. Siam Review 39, p. 383-
400 (1997).
[Tr2] L.N. Trefethen. Spectral methods in Mathlab. SIAM Philadelphia (2000).
[TrEm] L.N. Trefethen and M. Embree. Spectra and PseudospectraThe be-
havior of nonnormal matrices and operators. Lecture Notes (Version of 9
November 2004).
[Yo] K. Yosida : Functional Analysis, Grundlehren der mathematischen Wis-
senschaften 123, Springer Verlag, (1980).
[Zw] M. Zworski. A remark on a paper of E.B. Davies. Proc. Amer. Math. Soc.
129 (10), p. 2955-2957 (2001).
137