Introduction to the spectral theory

Lecture notes of the course

given at the University Paris-Sud (Orsay, France)
in September-December 2014

within the master programs

“Partial Differential Equations and Scientific Computing”
(M2 ANEDP) and “Analysis, Arithmetics and Geometry” (M2 AAG)

Lecturer: Konstantin Pankrashkin

Contents
Notation 1

1 Unbounded operators 3
1.1 Closed operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Adjoint operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Operators and quadratic forms 10

2.1 Operators defined by quadratic forms . . . . . . . . . . . . . . . . . . 10
2.2 Semibounded operators and Friedrichs extensions . . . . . . . . . . . 16
2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Spectrum and resolvent 20

3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Basic facts on the spectra of self-adjoint operators . . . . . . . . . . . 24
3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 Spectral theory of compact operators 27

4.1 Fredholm’s alternative and spectra of compact operators . . . . . . . 27
4.2 Integral and Hilbert-Schmidt operators . . . . . . . . . . . . . . . . . 31
4.3 Operators with compact resolvent . . . . . . . . . . . . . . . . . . . . 33
4.4 Schrödinger operators with growing potentials . . . . . . . . . . . . . 35

5 Spectral theorem 37
5.1 Continuous functional calculus . . . . . . . . . . . . . . . . . . . . . . 38
5.2 Borelian functional calculus and L2 representation . . . . . . . . . . . 43

6 Some applications of spectral theorem 51

6.1 Spectral projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Generalized eigenfunctions . . . . . . . . . . . . . . . . . . . . . . . . 54
6.3 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7 Perturbations 58
7.1 Kato-Rellich theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.2 Essential self-adjointness of Schrödinger operators . . . . . . . . . . . 60
7.3 Discrete and essential spectra . . . . . . . . . . . . . . . . . . . . . . 62
7.4 Weyl criterion and relatively compact perturbations . . . . . . . . . . 64
 7.5 Essential spectra for Schrödinger operators . . . . . . . . . . . . . . . 67

8 Variational principle for eigenvalues 69

8.1 Max-min and min-max principles . . . . . . . . . . . . . . . . . . . . 69
8.2 Negative eigenvalues of Schrödinger operators . . . . . . . . . . . . . 71

9 Laplacian eigenvalues for bounded domains 73

9.1 Dirichlet and Neumann eigenvalues . . . . . . . . . . . . . . . . . . . 73
9.2 Weyl asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

References 79
Notation
We list some notations used throughout the text.
The symbol N denotes the set of the natural numbers starting from 0.
If (M, T , µ) is a measure space and f : M → C is a measurable function, then we
denote the essential range and the essential supremum of f w.r.t. the measure µ:
n  o
essµ ran f := z ∈ C : µ m ∈ M : z − f (m) < ε > 0 for all ε > 0 ,
n  o
essµ sup |f | := inf a ∈ R : µ m ∈ M : f (m) > a = 0 .

If the measure µ is obvious in the context, we will omit to indicate it in the notations.
In what follows the term Hilbert space will mean a separable complex Hilbert space.
The symbol H will implicitly denote such a Hilbert space. For two elements x, y ∈
H, then hx, yi will denote the sesquilinear scalar product of x and y. If several
Hilbet spaces are considered in the problem, we will specify the scalar product with
the notation hx, yiH . To respect the convention in quantum mechanics, our scalar
products will always be linear with respect to the second argument, and as antilinear
with respect to the first one:

∀α ∈ C hx, αyi = hαx, yi = αhx, yi.

For example, that the scalar product in the Lebesgue space L2 (R) is defined by
Z
hf, gi = f (x)g(x) dx.
R

If A is a finite or countable set, `2 (A) denotes the vector space of square-summable

functions x : A → C:
2
X
ξ(a) < ∞.
a∈A

Such functions x are sometimes written using subscripts: x(a) = xa , in particular

when A = N or A = Z.
This is a Hilbert space, equipped with the scalar product
X
hx, yi = x(a)y(a).
a∈A

If H and G are two Hilbert spaces, then by L(H, G) and K(H, G) denote the spaces
of bounded linear operators, respectively of compact operators from H and G. Fur-
theremore, L(H) := L(H, H) and K(H) := K(H, H).
If Ω ⊂ Rd is an open set and k ∈ N, then H k (Ω) denotes the kth Sobolev space,
i.e. the space of L2 functions whose partial derivatives up to order k are also in
L2 (Ω). By H0k (Ω) we denote the completion in H k (Ω) of the subspace Cc∞ (Ω) (with
respect to the norm of H k (Ω)). The symbol C k (Ω) denotes the space of functions
on Ω whose partial derivatives up to order k are continuous; in particular, the set

1
of the continuous functions is denoted as C 0 (Ω). This should not be confused with
the notation C0 (Rd ) for the space of continuous functions f : Rd → C vanishing at
infinity: lim|x|→∞ f (x) = 0. The subscript comp on a functional space indicates that
1
its elements have compact supports: for instance Hcomp (Rd ) is the space of functions
1 d
in H (R ) having compact supports.

Recommended books
During the preparation of the notes I used a part of the text by Bernard Helffer which
is available online [9]; an extended version was recently published as a book [10].
Other recommended books are the one E. B. Davies [5] and the book of G. Teschl [17]
(available online).
Additional references for particular topics will be given throughout the text.

2
1 Unbounded operators
1.1 Closed operators
A linear operator T in H is a linear map from a subspace (the domain of T ) D(T ) ⊂
H to H. The range of T is the set ran T := {T x : x ∈ D(T )}. We say that a linear
operator T is bounded if the quantity
kT xk
µ(T ) := sup
x∈D(T ) kxk
x6=0

is finite. In what follows, the word combination “an unbounded operator” should
be understood as “an operator which is not assumed to be bounded”. If D(T ) = H
and T is bounded, we arrive at the notion of a continuous linear operator in H; the
space of such operators is denoted by L(H). This is a Banach space equipped with
the norm kT k := µ(T ).
During the whole course, by introducing a linear operator we always assume that
its domain is dense, if the contrary is not stated explicitly.
If T is a bounded operator in H, it can be uniquely extended to a continuous linear
operator. Let us discuss a similar idea for unbounded operators.
The graph of a linear operator T in H is the set

gr T := (x, T x) : x ∈ D(T ) ⊂ H × H.

For two linear operators T1 and T2 in H we write T1 ⊂ T2 if gr T1 ⊂ gr T2 . I.e.

T1 ⊂ T2 means that D(T1 ) ⊂ D(T2 ) and that T2 x = T1 x for all x ∈ D(T1 ); the
operator T2 is then called an extension of T1 and T1 is called a restriction of T2 .
Definition 1.1 (Closed operator, closable operator).

• A linear operator T in H is called closed if its graph is a closed subspace in

H × H.

• A linear operator T in H is called closable, if the closure gr T of the graph of

T in H × H is still the graph of a certain operator T . This operator T with
gr T = gr T is called the closure of T .

The following proposition is obvious:

Proposition 1.2. A linear operator T in H is closed if and only if the three condi-
tions

• xn ∈ D(T ),

• xn converge to x in H,

• T xn converge to y in H

3
imply the inclusion x ∈ D(T ) and the equality y = T x.

Definition 1.3 (Graph norm). Let T be a linear operator in H. Define on

D(T ) a new
p scalar product
p by hx, yiT = hx, yi + hT x, T yi. The associated norm
kxkT := hx, x, iT = kxk2 + kT xk2 is called the graph norm for T .

The following assertion is also evident.

Proposition 1.4. Let T be a linear operator in H.

• T is closed iff D(T ) is complete in the graph norm.

• If T is closable, then D(T ) is exactly the completion of D(T ) with respect to

the graph norm.

Informally, one could say that D(T ) consists of those x for which there is a unique
candidate for T x if one tries to extend T by density. I.e., a vector x ∈ H belongs to
D(T ) iff:

• there exists a sequence (xn ) ⊂ D(T ) converging to x,

• their exists the limit of T xn ,

• this limit is the same for any sequence xn satisfying the previous two properties.

Let us consider some examples.

Example 1.5 (Bounded linear operators are closed). By the closed graph
theorem, a linear operator T in H with D(T ) = H is closed if and only if it is
bounded. In this course we consider mostly unbounded closed operators.

Example 1.6 (Multiplication operator). Take again H = L2 (Rd ) and pick f ∈

L∞ d
loc (R ). Introduce a linear operator Mf in H as follows:

D(Mf ) = {u ∈ L2 (Rd ) : f u ∈ L2 (Rd )} and Mf u = f u for u ∈ D(Mf ).

It can be easily seen that D(Mf ) equipped

with the graph norm coincides with
2 d 2
the weighted space L R , (1 + |f | )dx , which is complete. This shows that Mf is
closed.
An interested reader may generalize this example by considering multiplications
operators in measure spaces.

Example 1.7 (Laplacians in Rd ). Take H = L2 (Rd ) and consider two operators

in H:

T0 u = −∆u, D(T0 ) = Cc∞ (Rd ),

T1 u = −∆u, D(T1 ) = H 2 (Rd ).

4
We are going to show that T 0 = T1 (it follows that T1 is closed and T0 is not closed).
We prove first the following equality:

H 2 (Rd ) = f ∈ L2 (Rd ) : −∆f ∈ L2 (Rd )


(1.1)

Clearly, H 2 (Rd ) is included into the set on the right-hand side. Now, let f belong
to the set on the right-hand side, we have that fb ∈ L2 (Rd ) and p2 fb ∈ L2 (Rd ). From

p2j + p2k 1 + p2j

pj pk fb ≤ |fb| ≤ p2 |fb|, pj fb ≤ |fb| ≤ (1 + p2 )|fb|
2 2

it follows that pj fb ∈ L2 , pj pk fb ∈ L2 . In summary, ∂ α f ∈ L2 (Rd ) for |α| ≤ 2, which

proves the inclusion f ∈ H 2 (Rd ) and the equality (1.1)
Denote by F : H → H the Fourier transform in L2 (Rd ) and consider the following
operator Tb in H:

D(Tb) = {f ∈ L2 : p 7→ p2 f (p) ∈ L2 }, Tbf (p) = p2 f (p).

Indeed, Tb is closed operator, as this is just a multiplication operator, see Example

1.6. On the other hand, for f ∈ H one has the following equivalence: f ∈ D(T1 ) iff
F f ∈ D(Tb), and in that case F T1 f (p) = TbF f (p). In other words, one can represent

gr T1 = {(F −1 u, F −1 Tbu) : u ∈ D(Tb)} = K(gr Tb),

where K is the linear operator in H × H defined by K(x, y) = (F −1 x, F −1 y). As F

is a unitary operator, so is K, which means, in particular, that K maps closed sets
to closed sets. As gr Tb is closed, the graph gr T1 is also closed, and T1 is a closed
operator.
As we have the inclusion T0 ⊂ T1 and T1 is closed, it follows that T0 is at least
closable, and the domain D(T 0 ) is the completion D(T0 ) in the graph norm of T0
(Proposition 1.4)
or, equivalently, D(T0 ) is the closure of D(T 0 ) in the Hilbert space
D(T1 ), h·, ·iT1 . Actually we have shown above that the graph norm of T1 is equiva-
lent to the norm of H 2 (Rd ), and hence D(T0 ) = Cc∞ (Rd ) is dense in D(T1 ), h·, ·iT1 .

This shows that D(T 0 ) = H 2 (Rd ).

Furthermore, T 0 ⊂ T1 , which means that for any u ∈ D(T 0 ) we have T 0 u = T1 u =
−∆u.

Example 1.8 (Non-closable operator). Take H = L2 (R) and pick a g ∈ H with

g 6= 0. Consider the operator L defined on D(L) = C 0 (R) ∩ L2 (R) by Lf = f (0)g.
Assume that there exists the closure L and let f ∈ D(L). One can find two sequences
(fn ), (gn ) in D(L) such that both converge in the L2 norm to f but such that
fn (0) = 0 and gn (0) = 1 for all n. Then Lfn = 0, Lgn = g for all n, and both
sequences Lfn and Lgn converge, but to different limits. This contradicts to the
closedness of L. Hence L is not closed.

5
Example 1.9 (Partial differential operators). Let Ω be an open subset of Rd
and P (x, Dx ) be a partial differential expression with C ∞ coefficients. Introduce in
H := L2 (Ω) a linear operator P by: D(P ) = Cc∞ (Ω), P u(x) = P (x, Dx )u(x). Like
in the previous example one shows the inclusion

gr P ⊂ (u, f ) ∈ H × H : P (x, Dx )u = f in D0 (Ω) ,



hence gr P is still a graph, and P is closable.

So we see that we naturally associate with the differential expression P (x, Dx ) several
linear operators (besides P ), in particular, the minimal operator Pmin := P , which
is always closed, and the maximal operator Pmax defined by Pmax u = P (x, Dx )u on
the domain 
D(Pmax ) := u ∈ H : P (x, Dx )u ∈ H ,
where P (x, Dx )u is understood in the sense of distributions. Clearly, one always has
the inclusion Pmin ⊂ Pmax , and we saw in example 1.7 that one can have Pmin =
Pmax . But one can easily find examples where this equality does not hold. For
example, for P (x, Dx ) = d/dx and Ω = (0, +∞) we have D(Pmin ) = H01 (0, ∞) and
D(Pmax ) = H 1 (0, ∞). In general, one may expect that Pmin 6= Pmax if Ω has a
boundary.
Such questions become more involved if one studies the partial differential operators
with more singular coefficients (e.g. with coefficients which are not smooth but just
belong to some Lp ). During the course we will see how to deal with some special
classes of such operators.

1.2 Adjoint operators

Recall that for T ∈ L(H) its adjoint T ∗ is defined by the relation

hx, T yi = hT ∗ x, yi for all x, y ∈ H.

The proof of the existence comes from the Riesz representation theorem: for each
x ∈ H the map H 3 y 7→ hx, T yi ∈ C is a continuous linear functional, which means
that there exists a unique vector, denoted by T ∗ x with hx, T yi = hT ∗ x, yi for all
y ∈ H. One can then show that the map x 7→ T ∗ x is linear, and by estimating the
scalar product one shows that T ∗ is also continuous. Let us use the same idea for
unbounded operators.
Definition 1.10 (Adjoint operator). If T be a linear operator in H, then its
adjoint T ∗ is defined as follows. The domain D(T ∗ ) consists of the vectors u ∈ H
for which the map D(T ) 3 v 7→ hu, T vi ∈ C is bounded with respect to the H-norm.
For such u there exists, by the Riesz theorem, a unique vector denoted by T ∗ u such
that hu, T vi = hT ∗ u, vi for all v ∈ D(T ).

We note that the implicit assumption D(T ) = H is important here: if it is not

satisfied, then the value T ∗ u is not uniquely determined, one can add to T ∗ u an
arbitrary vector from D(T )⊥ .

6
Let us give a geometric interpretation of the adjoint operator. Consider a unitary
linear operator
J : H × H → H × H, J(x, y) = (y, −x)
and note that J commutes with the operator of the orthogonal complement in H×H,
i.e. J(V )⊥ = J(V ⊥ ) for any V ⊂ H × H. This will be used several times during the
course.

Proposition 1.11 (Geometric interpretation of the adjoint). Let T be a linear

operator in H. The following two assertions are equivalent:

• u ∈ D(T ∗ ) and f = T ∗ u,

• (u, T ∗ u), J(v, T v) H×H

= 0 for all v ∈ D(T ).

In other words,
gr T ∗ = J(gr T )⊥ . (1.2)

As a simple application we obtain

Proposition 1.12. One has ( T )∗ = T ∗ , and T ∗ is a closed operator.

Proof. Follows from (1.2): the orthogonal complement is always closed, and
J(gr T )⊥ = J(gr T )⊥ .

Up to now we do not know if the domain of the adjoint contains non-zero vectors.
This is discussed in the following proposition.

Proposition 1.13 (Domain of the adjoint). Let T be a closable operator H,

then:

(i) D(T ∗ ) is a dense subspace of H,

(ii) T ∗∗ := (T ∗ )∗ = T .

Proof. The item (ii) easily follows from (i) and (1.2): one applies the same op-
erations again and remark that J 2 = −1 and that taking twice the orthogonal
complement results in taking the closure.
Now let us prove the item (i). Let a vector w ∈ H be orthogonal to D(T ∗ ): hu, wi = 0
for all u ∈ D(T ∗ ). Then one has hJ(u, T ∗ u), (0, w)iH×H ≡ hu, wi + hT ∗ u, 0i = 0 for
all u ∈ D(T ∗ ), which means that (0, w) ∈ J(gr T ∗ )⊥ = gr T . As the operator T is
closable, the set gr T must be a graph, which means that w = 0.

Let us look at some examples.

Example 1.14 (Adjoint for bounded operators). The general definition of the
adjoint operator is compatible with the one for continuous linear operators.

7
Example 1.15 (Laplacians in Rd ). Let us look again at the operators T0 and T1
from example 1.7. We claim that T0∗ = T1 . To see this, let us describe the adjoint
T0∗ using the definition. The domain D(T0∗ ) consists of the functions u ∈ L2 (Rd ) for
which there exists a vector f ∈ L2 (Rd ) such that the equality
Z Z
u(x)(−∆v)(x)dx = f (x)v(x)dx
Rd Rd

holds for all v ∈ D(T0 ) ≡ Cc∞ (Rd ). This means that one should have f = −∆u in
the sense of distributions. Therefore, D(T0∗ ) consists of the functions u ∈ L2 such
that −∆u ∈ L2 . By (1.1) we have u ∈ H 2 (Rd ) = D(T1 ).

Definition 1.16 (Free Laplacian in Rd ). The operator T in L2 (Rd ) defined by

D(T ) = H 2 (Rd ), T u = −∆u,

is called the free Laplacian in Rd . By Example 1.15, is is a self-adjoint operator.

Example 1.17. As an exercise, one can show that for the multiplication operator
Mf from example 1.6 one has (Mf )∗ = Mf .

The following definition introduces two classes of linear operator that will be studied
throughout the course.

Definition 1.18 (Symmetric, self-adjoint, essentially self-adjoint opera-

tors). We say that a linear operator T in H is symmetric (or Hermitian) if

hu, T vi = hT u, vi for all u, v ∈ D(T ),

or, equivalently, if T ⊂ T ∗ . Furthermore:

• T is called self-adjoint if T = T ∗ ,

• T is called essentially self-adjoint if T is self-adjoint.

An important feature of symmetric operators is their closability:

Proposition 1.19. Symmetric operators are closable.

Proof. Indeed for a symmetric operator T we have gr T ⊂ gr T ∗ and, due to the

closedness of T ∗ , gr T ⊂ gr T ∗ .

Example 1.20 (Self-adjoint Laplacian in Rd ). For the laplacian T1 from example

1.7 one has T1 = T1∗ . Indeed, T1 = T0∗ , then T1∗ = T0∗∗ = T0 = T1 .

Example 1.21 (Bounded symmetric operators are self-adjoint). Note that

for T ∈ L(H) the fact to be symmetric is equivalent to the fact to be self-adjoint,
but it is not the case for unbounded operators!

8
Example 1.22 (Self-adjoint multiplication operators). As follows from exam-
ple 1.17, the multiplication operator Mf from example 1.6 is self-adjoint iff f (x) ∈ R
for a.e. x ∈ Rd .

A large class of self-adjoint operators comes from the following proposition.

Proposition 1.23. Let T be an injective self-adjoint operator, then its inverse is
also self-adjoint.

Proof. We show first that D(T −1 ) := ran T is dense in H. Let u ⊥ ran T , then
hu, T vi = 0 for all v ∈ D(T ). This can be rewritten as hu, T vi = h0, vi for all
v ∈ D(T ), which shows that u ∈ D(T ∗ ) and T ∗ u = 0. As T ∗ = T , we have
u ∈ D(T ) and T u = 0. As T in injective, one has u = 0
Now consider the operator S : H × H → H × H given by S(x, y) = (y, x). One has
then gr T −1 = S(gr T ). We conclude the proof by noting that S commutes with J
and with the operation of the orthogonal complement in H × H.

1.3 Exercises
Exercise 1.24. (a) In the Hilbert space H = L2 (0, 1) consider the operator T acting
on the domain D(T ) = H(0, 1) by T f = if 0 .
Is T closed? symmetric? self-adjoint? semibounded from below?
(b) The same questions for the operator T1 given by the same expression but on the
domain D(T1 ) = H01 (0, 1).
Exercise 1.25. (a) Let H1 and H2 be Hilbert spaces. Let A be a linear operator
in H1 , B be a linear operator in H2 . Assume that there exists a unitary operator
U : H2 → H1 such that D(A) = U D(B) and that U ∗ AU f = Bf for all f ∈ D(B);
such A and B are called unitary equivalent.
Let two operators A and B be unitarily equivalent. Show that A is
closed/symmetric/self-adjoint iff B has the same property.
(b) Let (λn ) be an arbitrary sequence of complex numbers, n ∈ N. In the Hilbert
space `2 (N) consider the operator S:

D(S) = (xn ) : there exists N such that xn = 0 for n > N , S(xn ) = (λn xn ).

Describe the closure of S.

(c) Now let H be a separable Hilbert space and T be a linear operator in H with the
following property there exists an orthonormal basis (en )n∈N of H with en ∈ D(T )
and T en = λn en for all n ∈ N, where λn are some complex numbers.

1. Describe the closure T of T . Hint: one may use (a) and (b).

2. Describe the adjoint T ∗ of T .

3. Let all λn be real. Show that the operator T is self-adjoint.

9
Exercise 1.26. Let A and B be self-adjoint operators in a Hilbert space H such
that D(A) ⊂ D(B) and Au = Bu for all u ∈ D(A). Show that D(A) = D(B).
(This property is called the maximality of self-adjoint operators.)

Exercise 1.27. In this exercise, by the sum A + B of a linear operator B with a

continuous operator B; both acting in a Hilbert space H, we mean the operator S
defined by D(S) = D(A), Su := Au + Bu. 5We note that defining the sum of two
operators becomes a non-trivial task if unbounded operators are involved.)
(a) Let A be a closed and B be continuous. Show that A + B is closed.
(b) Assume, in addition, that A is densely defined. Show that (A + B)∗ = A∗ + B ∗ .

Exercise 1.28. Let H := L2 (0, 1). For α ∈ C consider the operator Tα acting as
Tα f = if on the domain
n o
∞


D(Tα ) = f ∈ C [0, 1] : f (1) = αf (0) .

(a) Find the adjoint of Tα .

(b) Find the closure Sα := Tα .
(c) Find all α for which Sα is self-adjoint.

2 Operators and quadratic forms

2.1 Operators defined by quadratic forms
A sesquilinear form t in a Hilbert space H with the domain D(t) ⊂ H is a map
t : H × H ⊃ D(t) × D(t) → C which is linear with respect to the second argument
and is antilinear with respect to the first one. By default we assume that D(t) is
a dense subset of H. (In the literature, one uses also the terms bilinear form and
quadratic form.) We will consider the following classes of sesquilinear forms: a form
t is called

• bounded, if D(t) = H and there exists M > 0 such that t(u, v) ≤ M kuk · kvk
for all u, v ∈ H,

• elliptic, if it is bounded and there exists α > 0 such that t(u, u) ≥ αkuk2 for
all u ∈ H,

• symmetric if t(u, v) = t(u, v) for all u, v ∈ D(t),

• semibounded from below if t is symmetric and for some c ∈ R one has t(u, u) ≥
ckuk2 for all u ∈ D(t); in this case we write t ≥ c;

• positive or non-negative, if one can take c = 0 in the previous item,

• positively definite ot strictly positive, if one can take c > 0.

10
Now let V be a Hilbert space and let t be a bounded sesquilinear form in V. It is
known that there is a uniquely determined operator A ∈ L(V ) such that t(u, v) =
hu, Avi for all u, v ∈ V. Let us recall the following classical result:

Theorem 2.1 (Lax-Milgram theorem). If t is elliptic, then the associated oper-

ator A is an isomorphism of V, that is, A is invertible and A−1 ∈ L(V).

Proof. By assumption, one can find two constants α, C > 0 such that

αkvk2 ≤ t(v, v) ≤ Ckvk2 for all v ∈ V.

This implies αkvk2 ≤ a(v, v) = hv, Avi ≤ kvk · kAvk. Hence,

kAvk ≥ αkvk for all v ∈ V. (2.1)

Step 1. We can see that A is injective, because Av = 0 implies v = 0 by (2.1).

Step 2. Let us show that ran A is closed. Assume that fn ∈ ran A and that fn
converge to f in the norm of V. By the result of step 1 there are uniquely determined
vectors vn ∈ V with fn = Avn . The sequence (fn ) = (Avn ) is convergent and is then
a Cauchy one. By (2.1), the sequence (vn ) is also a Cauchy one and, due to the
completeness of V, converges to some v ∈ V. As A is continuous, Avn converges to
Av. Hence, f = Av, which shows that f ∈ ran A
Step 3. Let us show finally that ran A = V. As we showed already that ran A is
closed, it is sufficient to show that (ran A)⊥ = {0}. Let u ⊥ ran A, then t(u, v) =
hu, Avi = 0 for all v ∈ V . Taking v = u we obtain a(u, u) = 0, and u = 0 by the
ellipticity.

Let us extend the above construction to unbounded operators and forms.

Definition 2.2 (Operator defined by a form).

Let V and t be as in Theorem 2.1. Moreover, assume that V is a dense subset of
another Hilbert space H and that there exists a constant c > 0 such that kukH ≤
ckukV for all u ∈ V. Introduce an operator T defined by t as follows. The domain
D(T ) consists of the vectors v ∈ V for which the map V 3 u 7→ t(u, v) can be
extended to a continuous antilinear map from H to C. By the Riesz theorem, for
such a v there exists a uniquely defined fv ∈ H such that t(u, v) = hu, fv iH for all
u ∈ V, and we set T v := fv .

Note that one can associate an operator T to any sesquilinear form t but the prop-
erties of this operator are quite unpredictable if one does not assume any additional
properties of the form t.

Theorem 2.3. In the situation of definition 2.2 one has

• the domain of T is dense in H,

• T : D(T ) → H is bijective,

11
 • T −1 ∈ L(H).

Proof. Let v ∈ D(T ). Using the V-ellipticity we have the following inequalities:

αkvk2H ≤ αc2 kvk2V ≤ c2 t(v, v) ≤ c2 hv, T viH ≤ c2 kvkH · kT vkH ,

showing that
α
kT vkH ≥ kvkH . (2.2)
c2
We see immediately that T in injective.
Let us show that T is surjective. Let h ∈ H and let A ∈ L(V) be the operator
associated with t. The map V 3 u 7→ hu, hiH is a continuous antilinear map from
V to C, so one can find w ∈ V such that hu, hiH = hu, wiV for all u ∈ V. Denote
v := A−1 w, then hu, hiH = hu, AviV = a(u, v). By definition this means that
v ∈ D(T ) and h = T v.
Hence, T is surjective and injective, and the inverse is bounded by (2.2). It remains
to show that the domain of T is dense in H. Let h ∈ H with hu, hiH = 0 for all
u ∈ D(T ). As T is surjective, there exists v ∈ D(T ) with h = T v. Taking now
u = v we obtain hv, T viH = 0, and the V-ellipticity gives v = 0 and h = 0.

If the form t has some additional properties, then the associated operators T also
enjoys some additional properties.

Theorem 2.4 (Self-adjoint operators defined by forms). Let T be the operator

associated with a symmetric sesqulinear form t in the sense of definition 2.2, then

1. T is a self-adjoint operator in H,

2. D(T ) is a dense subspace of the Hilbert space V.

Proof. For any u, v ∈ D(T ) we have:

hu, T viH = t(u, v) = t(v, u) = hv, T uiH = hT u, viH .

Therefore, T is at least symmetric and T ⊂ T ∗ . Let v ∈ D(T ∗ ). We know from the

previous theorem that T is surjective. This means that we can find v0 ∈ D(T ) such
that T v0 = T ∗ v. Then for all u ∈ D(T ) we have:

hT u, viH = hu, T ∗ viH = hu, T v0 iH = hT u, v0 iH .

As T is surjective, this imples v = v0 and then T = T ∗ .

Let us show the density of D(T ) in V. Let h ∈ V such that hv, hiV = 0 for all
v ∈ D(T ). There exists f ∈ V such that h = Af , where A ∈ L(V) is the opera-
tor associated with t. We have then 0 = hv, hiV = hv, Af iV = t(v, f ) = t(f, v) =
hf, T viH = hT v, f iH . As the vectors T v cover the whole of H as v runs through
D(T ), this imples f = 0 and h = Af = 0, and we see that the orthogonal comple-
ment of D(T ) in V is {0}.

12
An important point in the above consideration is the presence of a certain auxiliary
Hilbert space V. As a set, V coincides with the domain of the form t. This motivates
the following definition:
Definition 2.5 (Closed forms). A sesquilinear form t in a Hilbert space H with
a dense domain D(t) is called closed if the following properties are satisfied:

• t is symmetric,

• t is semibounded from below: t(u, u) ≥ −Ckuk2H for all u ∈ D(t) for some
C ∈ R,

• The domain D(t) equipped with the scalar product

hu, vit := t(u, v) + (C + 1)hu, vi

is a Hilbert space.

The previous considerations imply the following result:

Proposition 2.6 (Operators defined by closed forms).
Let t be a closed sesquilinear form in H, then the associated linear operator T in H
is self-adjoint.


Proof. One simply takes D(t), h·, ·it as V, then t : V × V → C is bounded and
elliptic.
Definition 2.7 (Closable form). Let us introduce another important notion. We
say that a symmetric sesqulinear form t is closable, if there exists a closed form in
H which extends t. The closed sesqulinear form with the above property and with
the minimal domain is called the closure of t and is denoted t.

The following proposition is rather obvious.

Proposition 2.8 (Domain of the closure of a form). If t ≥ −c, c ∈ R, and t
is a closable form, then D(t) is exactly the completion of D(t) with respect to the
scalar product hu, vit := t(u, v) + (c + 1)hu, vi.

It is time to look at examples!

Example 2.9 (Non-closable form). Take H = L2 (R) and consider the form
a(u, v) = u(0)v(0) defined on D(a) = L2 (R) ∩ C 0 (R). This form is densely defined,
symmetric and positive. Let us show that it is not closable. By contradiction,
assume that a is the closure of a. One should then have the following property: if
(un ) is a sequence of vectors from D(a) which is Cauchy with respect to h·, ·ia , then
it converges to some u ∈ D(a) ⊂ H and a(un , un ) converges to a(u, u). But one can
construct two sequences (un ) and (vn ) in D(a) such that

• both converge to u in the L2 -norm,

13
 • un (0) = 1 and vn (0) = 0 for all n.

Then both sequences are a-Cauchy, but the limits of a(un , un ) and a(vn , vn ) are
different. This shows that a cannot exist.

Let us give some “canonical” examples of operators defined by forms.

Example 2.10 (Laplacian). Consider the Hilbert space H = L2 (Rd ) and the form
Z
t(u, v) = ∇u∇v dx, D(t) = H 1 (Rd ),
Rd

which is clearly closed. Let us find the associated operator T . We know from the
very beginning that T is self-adjoint.
Let f ∈ D(T ) and g := T f , then for any u ∈ H 1 (Rd ) we have
Z Z
∇u∇f dx = ug dx.
Rd Rd

In particular, this equality holds for u ∈ Cc∞ (Rd ), which gives

Z Z Z
ug dx = ∇u∇f dx = (−∆u)f dx.
Rd Rd Rd

It follows that g = −∆f in D0 (Rd ). Therefore, for each f ∈ D(T ) we have ∆f ∈

L2 (Rd ), which by (1.1) means that D(T ) ⊂ H 2 (Rd ). In other words, we have T ⊂ T1 ,
where T1 is the free Laplacian in Rd (see Definition 1.16). As both T and T1 are
self-adjoint, we have T = T1 (Exercise 1.26).
Example 2.11 (Neumann boundary condition on the halfline). Take H =
L2 (0, ∞). Consider the form
Z ∞
t(u, v) = u0 (x)v 0 (x)dx, D(t) = H 1 (0, ∞).
0

The form is semibounded below and closed (which is in fact just equivalent to the
completeness of H 1 in the respective Sobolev norm). Let us describe the associated
operator T .
Let v ∈ D(T ), then there exists fv ∈ H such that
Z ∞ Z ∞
0
u0 (x)v (x)dx = u(x)fv (x)dx
0 0

for all u ∈ H 1 . Taking here u ∈ Cc∞ we obtain just the definition of the derivative
in the sense of distributions: fv := −(v 0 )0 = −v 00 . As fv ∈ L2 , the function v must
be in H 2 (0, ∞) and T v = −v 00 .
Now note that for v ∈ H 2 (0, ∞) and u ∈ H 1 (0, ∞) there holds
Z ∞ x=∞
Z ∞
0 0
0
u (x)v (x)dx = u(x)v (x) − u(x)v 00 (x)dx.
0 x=0 0

14
Hence, in order to obtain the requested inequality t(u, v) = hu, T viH the boundary
term must vanish, which gives the additional condition v 0 (0) = 0.
Therefore, the associated operator is TN := T acts as TN v = −v 00 on the domain
D(TN ) = v ∈ H 2 (0, ∞) : v 0 (0) = 0 . It will be referred as the (positive) Laplacian
with the Neumann boundary condition or simply the Neumann Laplacian.
Example 2.12 (Dirichlet boundary condition on the halfline). Take H =
L2 (0, ∞). Consider the form which is a restriction of the previous one,
Z ∞
t0 (u, v) = u0 (x)v 0 (x)dx, D(t0 ) = H01 (0, ∞).
0

The form is still semibounded below and closed (as H01 is still complete with respect
to the H 1 -norm), and the boundary term does not appear when integrating by parts,
which means that the associated operator TD = T acts as TD v = −v 00 on the domain
D(TD ) = H 2 (0, ∞) ∩ H01 (0, ∞) = {v ∈ H 2 (0, ∞) : v(0) = 0}. It will be referred to
as the (positive) Laplacian with the Dirichlet boundary condition or the Dirichlet
Laplacian.
Remark 2.13. In the two previous examples we see an important feature: the fact
that one closed form extends another closed form does not imply the same relation
for the associated operators.
Example 2.14 (Neumann/Dirichlet Laplacians: general case). The two pre-
vious examples can be generalized to the multidimensional case. Let Ω be an open
subset of Rd with a sufficiently regular boundary ∂Ω (for example, a compact Lips-
chitz one). In H = L2 (Ω) consider two sesqulinear forms:
Z
t0 (u, v) = ∇u ∇vdx, D(t0 ) = H01 (Ω),
ZΩ
t(u, v) = ∇u ∇vdx, D(t) = H 1 (Ω).
Ω

Both these forms are closed and semibounded from below, and one can easily show
that the respective operators A and A0 act both as u 7→ −∆u. By a more careful
analysis and, for example, for a smooth ∂Ω, one can show that

D(A0 ) = H 2 (Ω) ∩ H01 (Ω) = {u ∈ H 2 (Ω) : u|∂Ω = 0},

∂u
D(A) = {u ∈ H 2 (Ω) : = 0},
∂n ∂Ω
where n denotes the outward pointing unit normal vector on ∂Ω, and the restrictions
to the boundary should be understood as the respective traces. If the boundary is
not regular, the domains become more complicated, in particular, the domains of A
and A0 are not included in H 2 (Ω), see e.g. the book [7]. Nevertheless, the operator
A0 is called the Dirichlet Laplacian in Ω and A is called the Neumann Laplacian.
Indeed, the whole construction makes sense if the boundary of Ω is non-empty. For
example, A = A0 if Ω = Rd , as H 1 (Rd ) = H01 (Rd ).

15
2.2 Semibounded operators and Friedrichs extensions
Definition 2.15 (Semibounded operator). We say that a symmetric operator
T in H is semibounded from below if there exists a constant C ∈ R such that

hu, T ui ≥ −Chu, ui for all u ∈ D(T ),

and in that case we write T ≥ −C.

Now assume that an operator T is semibounded from below and consider the induced
sesqulinear form t in H,

t(u, v) = hu, T vi, D(t) = D(T ).

Proposition 2.16. The sesqulinear form t is semibounded from below and closable.

Proof. The semiboundedness of t from below follows directly from the analogous
property for T . To show the closability we remark that without loss of generality
one can assume T ≥ 1. By proposition 2.8, the domain V p of the closure of t must be
the completion of D(T ) with respect to the norm p(x) = t(x, x). More concretely,
a vector u ∈ H belongs to V iff there exists a sequence un ∈ D(T ) which is p-
Cauchy such that un converges to u in H. A natural candidate for the norm of u
is p(u) = lim p(un ). Actually we just need to show that this limit is independent
of the choice of the sequence un . Using the standard arguments we are reduced to
prove the following:
Assertion. If (un ) ⊂ D(t) is a p-Cauchy sequence converging to zero in H, then
lim p(un ) = 0.
To prove this assertion we observe first that p(xn ) is a non-negative Cauchy sequence,
and is convergent to some limit α ≥ 0. Suppose by contradiction that α > 0. Now
let us remark that t(un , um ) = t(un , un ) + t(un , um − un ). Moreover, by the Cauchy-
Schwartz inequality for p we have t(un , um − un ) ≤ p(un )p(um − un ). Combining
the two preceding expressions with the fact that un is p-Cauchy, we see that for any
ε > 0 there exists N > 0 such that t(un , um ) − α2 ≤ ε for all n, m > N . Take
ε = α2 /2 and take the associated N , then for n, m > N we have

α2
hun , T um i ≡ t(un , um ) ≥ .
2
On the other hand, the term on the left-hand side goes to 0 as n → ∞. So we obtain
a contradiction, and the assertion is proved.

Definition 2.17 (Friedrichs extensions). Let T be a linear operator in H which

is semibounded from below. Consider the above sesqulinear form t and its closure
t. The self-adjoint operator TF associated with the form t is called the Friedrichs
extension of T .

Corollary 2.18. A semibounded operator always has a self-adjoint extension.

16
Remark 2.19 (Form domain). If T is a self-adjoint operator and is semibounded
from below, then it is the Friedrichs extension of itself. The domain of the associated
form t is usually called the form domain of T and is denoted Q(T ). The form domain
plays an important role in the analysis of self-adjoint operators, see e.g. Section 8.
For f, g ∈ Q(T ) one uses sometimes a slightly abusive writing hf, T gi instead of
t(f, g).
Example 2.20 (Schrödinger operators). A basic example for the Friedrichs
extension is delivered by Schrödinger operators with semibounded potentials. Let
W ∈ L2loc (Rd ) and W ≥ −C, C ∈ R (i.e. W is semibounded from below). In
H = L2 (Rd ) consider the operator T acting as T u(x) = −∆u(x) + W (x)u(x) on the
domain D(T ) = Cc∞ (Rd ). One has clearly T ≥ −C. The Friedrichs extension TF
of T will be called the Schrödinger operator with the potential W . Note that the
sesqulinear form t associated with T is given by
Z Z
t(u, v) = ∇u∇vdx + W uvdx.
Rd Rd

Denote by t the closure of the form t. One can easily show the inclusion
1
p
(Rd ) := u ∈ H 1 (Rd ) : |W |u ∈ L2 (Rd ) .

D(t) ⊂ HW
1
Note that actually we have the equality D(et) = HW (Rd ), see Theorem 8.2.1 in the
book [5] for a rather technical proof, but the inclusion will be sufficient for our
purposes.

Let us extend the above example by including a class of potentials which are not
semibounded from below.
Proposition 2.21 (Hardy inequality). Let d ≥ 3 and u ∈ Cc∞ (Rd ), then
2
(d − 2)2 u(x)
Z Z
2
∇u(x) dx ≥ dx.
Rd 4 Rd |x|2
Proof. For any γ ∈ R one has
Z
xu(x) 2
∇u(x) + γ dx dx ≥ 0,
Rd |x|2
which may be rewritten in the form
2
u(x)
Z 2
Z
2
∇u(x) dx + γ dx
Rd Rd |x|2
Z 
u(x) u(x) 
≥ −γ x · ∇u(x) + x · ∇u(x) dx. (2.3)
Rd |x|2 |x|2
Using the identities
x d−2
∇|u|2 = u ∇u + u∇u, div 2
=
|x| |x|2

17
and the integration by parts we obtain
Z Z
 u(x) u(x)  2 x
x · ∇u(x) 2
+ x · ∇u(x) 2
dx = ∇ u(x) · 2 dx
Rd |x| |x| Rd |x|
2
u(x)
Z Z
2 x
=− u(x) div 2 dx = −(d − 2) dx.
Rd |x| Rd |x|2

Inserting this equality into (2.3) gives

2
u(x)
Z Z
2

∇u(x) dx ≥ γ (d − 2) − γ dx,
Rd Rd |x|2

and in order to optimize the coefficient before the integral on the right-hand side we
take γ = (d − 2)/2.

Note that the integral in the right-hand side of the Hardy inequality is not defined
for d ≤ 2, because the function x 7→ |x|−1 does not belong to L2loc anymore.
2
Corollary 2.22. Let d ≥ 3 and W ∈ L2loc (Rd ) be real-valued with W (x) ≥ (d−2)
4|x|2
,
then the operator T = −∆ + W defined on Cc∞ (Rd ) is semibounded from below and,
hence, has a self-adjoint extension.

Example 2.23 (Coulomb potential). We would like to show that the operator
T = −∆ + q/|x| in L2 (R3 ) is semibounded from below for any real q. For q ≥ 0
we are in the situation of Example 2.20. For q < 0 we are going to use the Hardy
inequality. For any u ∈ Cc∞ (R3 ) and any p ∈ R \ {0} we have:

|u|2 |u|
Z Z
dx = p|u| dx
R3 |x| R3 p|x|
p2 |u|2
Z Z
2 1
≤ |u| dx + 2 dx
2 R3 2p R3 |x|2
p2
Z Z
2 1
≤ |u| dx + 2 |∇u|2 dx,
2 R3 8p R3
and

|u|2 |u|2
Z Z Z Z
2 2
hu, T ui = |∇u| dx + q dx ≥ |∇u| dx − |q| dx
R3 R3 |x| R3 R3 |x|
|q|  |q|p2
 Z Z
2
≥ 1− 2 |∇u| dx − |u|2 dx,
8p R 3 2 R3

and is sufficient to pick any p with 8p2 ≥ |q|.

Therefore, for any q ∈ R the above operator T has a self-adjoint extension (Friedrichs
extension). Actually we will see below that this self-adjoint extension is unique.

18
2.3 Exercises
Exercise 2.24. Show that the following sesquilinear forms t are closed and semi-
bounded from below and describe the associated self-adjoint operators in H (α ∈ R
is a fixed parameter):
Z ∞
2 1
(a) H = L (0, ∞), D(t) = H (0, ∞), t(u, v) = u0 (s)v 0 (s) ds + αu(0)v(0).
0
Z
(b) H = L2 (R), D(t) = H 1 (R), t(u, v) = u0 (s)v 0 (s) ds + αu(0) v(0).
R
Z 1
u0 (s)v 0 (s) ds.
2
 1
(c) H = L (0, 1), D(t) = u ∈ H (0, 1) : u(0) = u(1) , t(u, v) =
0
Exercise 2.25. This exercise shows a possible way of constructing the sum of two
unbounded operators under the assumption that one of them is “smaller” that the
other one. In a sense, we are going to extend the construction of Exercise 1.4.
Let H be a Hilbert space, t be a symmetric sesquilinear form in H which is densely
defined, closed and semibounded below. Let T be the self-adjoint operator in H
associated with t. Let B be a symmetric linear operator in H such that D(t) ⊂ D(B)
and such that there exist α, β > 0 with kBuk2 ≤ αt(u, u) + βkuk2 for all u ∈ D(t).
Consider the operator S on D(S) = D(T ) defined by Su = T u + Bu. We are going
to show that S is self-adjoint.
(a) Consider the sesquilinear form s(u, v) = t(u, v) + hu, Bvi, D(s) = D(t). Show
that s is symmetric, closed and semibounded from below.
(b) Let Se be the operator associated with s. Show that D(S) e = D(T ) and that
Su = T u + Bu for all u ∈ D(T ).
e
(c) Show that S is self-adjoint.
Exercise 2.26. In the examples below the Sobolev embedding theorem and the
previous exercise can be of use.
(a) Let v ∈ L2 (R) be real-valued. Show that the operator A having as domain
D(A) = H 2 (R) and acting by Af (x) = −f 00 (x) + v(x)f (x) is a self-adjoint operator
in L2 (R).
(b) Let v ∈ L2loc (R) be real-valued and 1-periodic, i.e. v(x + 1) = v(x) for all
x ∈ R. Show that the operator A with the domain D(A) = H 2 (R) acting by
Af (x) = −f 00 (x) + v(x)f (x) is self-adjoint.
(c) Let H = L2 (R3 ). Suggest a class of unbounded potentials v : R3 → R such that
the operator A, Af (x) = −∆f (x) + v(x)f (x), with the domain D(A) = H 2 (R3 ) is
self-adjoint in H.
Exercise 2.27. (a) Let H be a Hilbert space and A be a closed densely defined
operator in H (not necessarily symmetric). Consider the operator L given by
Lu = A∗ Au, u ∈ D(L) = u ∈ D(A) : Au ∈ D(A∗ ) .


We will write simply L = A∗ A having in mind the above precise definition. While
the above is a natural definition of the product of two operators, it is not clear if
the domain D(L) is sufficiently large. We are going to study this question.

19
 1. Consider the sesquilinear form b(u, v) = hAu, Avi + hu, vi in H defined on
D(b) = D(A). Show that this form is closed, densely defined and semibounded
from below.
2. Let B be the self-adjoint operator associated with the form b. Find a relation
between L and B and show that L is densely defined, self-adjoint and positive.
3. Let A0 denote the restriction of A to D(L). Show that A0 = A.

(b) A linear operator A acting in a Hilbert space H is called normal if D(A) = D(A∗ )
and kAxk = kA∗ xk for all x ∈ D(A).

1. Show that any normal operator is closed.

2. Let A be a closed operator. Show: A is normal iff A∗ is normal.
3. Let A be a normal operator. Show: hAx, Ayi = hA∗ x, A∗ yi for all x, y ∈
D(A) ≡ D(A∗ ).
4. Let A be a closed operator. Show: A is normal iff AA∗ = A∗ A. Here the
both operators are defined as in (a), the operator AA∗ being understood as
(A∗ )∗ A∗ .

3 Spectrum and resolvent

3.1 Definitions
Actually most definitions of this chapter can be introduced in the Banach spaces,
but we prefer to concentrate on the Hilbertian case.
Definition 3.1 (Resolvent set, spectrum, point spectrum). Let T be a linear
operator in a Hilbert space H. The resolvent set res T consists of the complex z
for which the operator T − z : D(T ) → H is bijective and the inverse (T − z)−1 is
bounded. The spectrum spec T of T is defined by spec T := C \ res T . The point
spectrum specp T is the set of the eigenvalues of T .

Note that very often the resolvent set and the spectrum of T are often denoted by
ρ(T ) and σ(T ), respectively.
Proposition 3.2. If res T 6= ∅, then T is a closed operator.

Proof. Let z ∈ res T , then gr(T − z)−1 is closed by the closed graph theorem, but
then the graph of T − z is also closed, as gr(T − z) and gr(T − z)−1 are isometric in
H × H.
Proposition 3.3. For a closed operator T one has the following equivalence:
(
ker(T − z) = {0},
z ∈ res T iff
ran(T − z) = H.

20
Proof. The ⇒ direction follows from the definition.
Now let T be closed and z ∈ C with ker(T − z) = {0} and ran(T − z) = H. The
inverse (T − z)−1 is then defined everywhere and has a closed graph (as the graph
of T − z is closed), and is then bounded by the closed graph theorem.

Proposition 3.4 (Properties of the resolvent). The set res T is open and the
set spec T is closed. The operator function

res T 3 z 7→ RT (z) := (T − z)−1 ∈ L(H)

called the resolvent of T is holomorphic and satisfies the identities

RT (z1 ) − RT (z2 ) = (z1 − z2 )RT (z1 )RT (z2 ), (3.1)

RT (z1 )RT (z2 ) = RT (z2 )RT (z1 ), (3.2)
d
RT (z) = RT (z)2 (3.3)
dz
for all z, z1 , z2 ∈ res T .

Proof. Let z0 ∈ res H. We have the equality

T − z = (T − z0 ) 1 − (z − z0 )RT (z0 ) .

If |z − z0 | < 1/kRT (z0 )k, then the operator on the right had sinde has a bounded
inverse, which means that z ∈ res T . Moreover, one has the series representation

−1 X
RT (z) = 1 − (z − z0 )RT (z0 ) RT (z0 ) = (z − z0 )j RT (z0 )j+1 , (3.4)
j=0

which shows that RT is holomorphic. The remaining properties can be proved in a

similar way.

3.2 Examples
Let us consider a series of examples showing several situations where an explicit
calculation of the spectrum is possible. We emphasize that the point spectrum is
not the same as the spectrum!

Example 3.5. Consider the multiplication operator Mf from Example 1.6. Recall
that the essential range of a function f is defined by
n  o
ess ran f = λ : µ x : |f (x) − λ| <  > 0 for all  > 0 .

Clearly, this notion makes sense in any measure space. For a continuous function f
and the Lebesgue measure µ, the essential range coincides with the closure of the
usual range.

21
Proposition 3.6 (Spectrum of the multiplication operator). There holds

spec Mf = ess ran f,


specp Mf = λ : µ{x : f (x) = λ} > 0 .

Proof. Let λ ∈ / ess ran f , then the operator M1/(f −λ) is bounded, and one easily
checks that this is the inverse for Mf − λ. On the other hand, let λ ∈ ess ran f . For
any m ∈ N denote
Sem := x : |f (x) − λ| < 2−m


and choose a subset Sm ⊂ Sem of strictly positive but finite measure. If φm is the
characteristic function of Sm , one has
Z
2 2 2 2
(Mf − λ)φm = f (x) − λ φm (x) dx ≤ 2−2m φm ,
Sm

and the operator (Mf − λ)−1 cannot be bounded.

To prove the second assertion we remark that the condition
λ ∈ specp Mf is equiv-
alent to the existence of φ ∈ L2 (Rd ) such that f (x) − λ φ(x) = 0 for a.e. x.
This means that φ(x) = 0 for a.e. x with f (x) 6= λ, and specp Mf = ∅ if
µ{x : f (x) = λ} = 0. On the other hand, if µ{x : f (x) = λ} > 0, one can
choose a subset Σ ⊂ {x : f (x) = λ} of a strictly positive but finite measure, then
the characteristic function φ of Σ is an eigenfunction of Mf corresponding to the
eigenvalue λ.

Example 3.7. It can be shown that the spectrum is invariant under unitary trans-
formations (see Exercise 3.21):

Proposition 3.8 (Spectrum and unitary equivalence). Let two operators A

an B be unitarily equivalent, then spec A = spec B and specp A = specp B.

Example 3.9. Let T be the free Laplacian in Rd (see Definition 1.16). As seen
above, T is unitarily equivalent to the multiplication operator f (p) 7→ p2 f (p) in
L2 (Rd ). By Propositions 3.6 and 3.8 there holds spec T = [0, +∞) and specp T = ∅.

Example 3.10 (Discrete multiplication operator). Take H = `2 (Z). Consider

an aribtrary function a : Z → C, n 7→ an , and the associated operator T :

D(T ) = (ξn ) ∈ `2 (Z) : (an ξn ) ∈ `2 (Z) , (T ξ)n = an ξn .



Similarly to examples 1.6 and 3.6 one can show that T is a closed operator and that

spec T := {an : n ∈ Z}, specp T := {an : n ∈ Z}.

Example 3.11 (Harmonic oscillator). Let H = L2 (R). Consider the operator

T0 = −d2 /dx2 + x2 defined on S(R). We see that this operator is semibounded
from below and denote by T its Friedrichs extension. The operator T is called the
harmonic oscillator ; it is one of the basic models appearing in quantum mechanics.

22
One can easily see that the numbers λn = 2n−1 are eigenvalues of T0 , n ∈ N, and the
associated eigenfunctions φn are given by φn (x) = cn (−d/dx + x)n−1 φ1 (x), φ1 (x) =
c1 exp(−x2 /2), where cn are normalizing constants. It is known that the functions
(φn ) (called Hermite functions) form an orthonormal basis in L2 (R). We remark
that φn ∈ D(T0 ) for all n, hence, T0 is essentially self-adjoint (see Exercise 1.25c).
This means, in particular, that T = T0 .
Furthermore, using the unitary map U : L2 (R) → `2 (N), U f (n) = hφn , f i, one easily
checks that the operator T is unitarily equivalent to the operator of multiplication
by (λn ) in `2 (N), cf. Example 3.10, which gives
spec T = specp T = {2n − 1 : n ∈ N}.
Example 3.12 (A finite-difference operator). Consider again the Hilbert space
H = `2 (Z) and the operator T in H acting as (T u)(n) = u(n − 1) + u(n + 1). Clearly,
T ∈ L(H). To find its spectrum consider the map
X
Φ : `2 (Z) → L2 (0, 1), (Φu)(x) = u(n)e2πinx ,
n∈Z

where the sum on the right hand side should be understood as a series in L2 . It is
known that Φ is a unitary map. On the other hand, for any u ∈ `2 (Z) supported at
a finite number of points we have
X
Φ(T u)(x) = (T u)(n)e2πinx
n
X X
= u(n − 1)e2πinx + u(n + 1)e2πinx
n n
X X
2πi(n+1)x
= u(n)e + u(n)e2πi(n−1)x
n n
X X
=e 2πix
u(n)e 2πinx
+ e−2πix u(n)e2πinx
n n
= 2 cos(2πx)(Φu)(x).
Using the density argument we show that the operator ΦT Φ∗ is exactly the multil-
ication by f (x) = 2 cos(2πx) in the space L2 (0, 1), and its spectrum coincides with
the segment [−2, 2], i.e. with the essential range of f . So we have spec T = [−2, 2]
and specp T = ∅.
Example 3.13 (Empty spectrum). Take H= L2 (0, 1) and consider the operator
T acting as T f = f 0 on the domain D(T ) = f ∈ H 1 (0, 1) : f (0) = 0 . One can
easily see that for any g ∈ L2 (0, 1) and any z ∈ C the equation (T − z)f = g has
the unique solution explicitly given by
Z x
f (x) = ez(x−t) g(t) dt,
0

and the map g 7→ f is bounded in the norm of H. So we obtained res T = C and

spec T = ∅.

23
Example 3.14 (Empty resolvent set). Let us modify the previous example.
Take H = L2 (0, 1) and consider the operator T acting as T f = f 0 on the domain
D(T ) = H 1 (0, 1). Now for any z ∈ C we see that the function φz (x) = ezx belongs
to D(T ) and satisfies (T − z)φz = 0. Therefore, specp T = spec T = C.
As we can see in the two last examples, for general operators one cannot say much
on the location of the spectrum. In what follows we will study mostly self-adjoint
operators, whose spectral theory is now understood much better than for the non-
self-adjoint case.

3.3 Basic facts on the spectra of self-adjoint operators

The following two propositions will be of importance during the whole course.
Proposition 3.15. Let T be a closable operator in a Hilbert space H and z ∈ C,
then
ker(T ∗ − z) = ran(T − z)⊥ , (3.5)
∗ ⊥
ran(T − z) = ker(T − z) . (3.6)
Proof. Note that the second equality can be obtained from the first one by taking
the orthogonal complement in the both parts. Let us prove the first equality. As
D(T ) is dense, the condition f ∈ ker(T ∗ − z) is equivalent to h(T ∗ − z)f, gi = 0 for
all g ∈ D(T ), which can be also rewritten as
hT ∗ f, gi = zhf, gi for all g ∈ D(T ).
By the definition of T ∗ , one has hT ∗ f, gi = hf, T gi and
hf, T gi − zhf, gi ≡ hf, (T − z)gi = 0 for all g ∈ D(T ),
i.e. f ⊥ ran(T − z).
Proposition 3.16 (Spectrum of a self-adjoint operator is real). Let T be a
self-adjoint operator in a Hilbert space H, then spec T ⊂ R, and for any z ∈ C \ R
there holds
1
(T − z)−1 ≤ . (3.7)
|=z|
Proof. Let z ∈ C \ R and u ∈ D(T ). We have
hu, (T − z)ui = hu, T ui − <zhu, ui − i=zhu, ui.
As T is self-adjoint, the number hu, T ui is real. Therefore,
|=z|kuk2 ≤ hu, (T − z)ui ≤ (T − z)u · kuk,
which shows that
(T − z)u ≥ |=z| · kuk. (3.8)
It follows from here that ran(T − z) is closed, that ker(T − z) = {0} and, by
proposition 3.15, than ran(T − z) = H. Therefore, (T − z)−1 ∈ L(H), and the
estimate (3.7) follows from (3.8).

24
The following proposition is of importance when studying bounded operators.

Proposition 3.17 (Spectrum of a continuous operator). Let T ∈ L(H), then

spec T is a non-empty subset of z ∈ C : |z| ≤ kT k .

Proof. Let z ∈ C with |z| > kT k. Represent T − z = −z(1 − T /z). As kT /zk < 1,
the inverse to T − z is defined by the series,
∞
X
−1
(T − z) =− T n z −n−1 .
n=0

and z ∈ res T . This implies the sought inclusion.

Let us show that the spectrum is non-empty. Assume that it is not the case. Then
for any f, g ∈ H the function C 3 z 7→ F (z) := hf, RT (z)gi ∈ C is holomorphic in C
by proposition 3.4. On the other hand, it follows from the above series representation
for the resolvent that for large z the norm of RT (z) tends to zero. It follows that
F (z) → 0 as |z| → ∞ and that F is bounded. By Liouville’s theorem, F is constant,
and, moreover, F (z) = lim|z|→+∞ F (z) = 0. Therefore, hf, RT (z)gi = 0 for all z ∈ C
and f, g ∈ H, which means that RT (z) = 0. This contradicts the definition of the
resolvent and shows that the spectrum of T must be non-empty.

Proposition 3.18 (Location of spectrum of self-adjoint operators). Let T =

T ∗ ∈ L(H). Denote

hu, T ui hu, T ui
m = m(T ) = inf , M = M (T ) = sup ,
u6=0 hu, ui u6=0 hu, ui

then spec T ⊂ [m, M ] and {m, M } ⊂ spec T .

Proof. We proved already that spec T ⊂ R. For λ ∈ (M, +∞) we have

hu, (λ − T )ui ≥ (λ − M )kuk2 ,

and (T − λ)−1 ∈ L(H) by the Lax-Milgram theorem. In the same way one shows
that spec T ∩ (−∞, m) = ∅.
Let us show that M ∈ spec T (for m the proof is similar). Using the Cauchy-Schwarz
inequality for the semi-scalar product (u, v) 7→ hu, (M − T )vi we obtain
2
hu, (M − T )vi ≤ hu, (M − T )ui · hv, (M − T )vi.

Taking the supremum over all u ∈ H with kuk ≤ 1 we arrive at

(M − T )v ≤ kM − T k · v, (M − T )v .

By assumption, one can construct a sequence (un ) with kun k = 1 such that
hun , T un i → M as n → ∞. By the above inequality we have then (M − T )un → 0,
and the operator M − T cannot have bounded inverse. Thus M ∈ spec T .

25
Corollary 3.19. If T = T ∗ ∈ L(H) and spec T = {0}, then T = 0.

Proof. By proposition 3.18 we have m(T ) = M (T ) = 0. This means that hx, T xi =

0 for all x ∈ H, and the polar identity shows that hx, T yi = 0 for all x, y ∈ H.

Let us combine all of the above to show the following fundamental fact.

Theorem 3.20 (Non-emptiness of spectrum). The spectrum of a self-adjoint

operator in a Hilbert space is a non-empty closed subset of the real line.

Proof. In view of the preceding discussion, it remains to show the non-emptyness

of the spectrum. Let T be a self-adjoint operator in a Hilbert space H. By contra-
diction, assume that spec T = ∅. Then, first of all, T −1 ∈ L(H). Let λ ∈ C \ {0}.
One can easily show that the operator
T 1 −1 1 1 1 −1
Lλ := − T− ≡− − 2 T−
λ λ λ λ λ
belongs to L(H) and that (T −1 − λ)Lλ = IdH and Lλ (T −1 − λ) = IdH . Therefore,
λ ∈ res(T −1 ). As λ was an arbitrary non-zero complex number, we have spec(T −1 ) ⊂
{0}. As T −1 is bounded, its spectrum is non-empty, hence, spec T −1 = {0}. On the
other hand, T −1 is self-adjoint by Proposition 1.23, and T −1 = 0 by Corollary 3.19,
which contradicts the definition of the inverse operator.

3.4 Exercises
Exercise 3.21. 1. Let two operators A and B be unitarily equivalent (see Ex-
ercise 1.25). Show that the spec A = spec B and specp A = specp B.

2. Let µ ∈ res A ∩ res B. Show that A and B are unitarily equivalent iff their
resolvents RA (µ) and RB (µ) are unitarily equivalent.

3. Let A be a closed operator. Show that spec A∗ = {z : z ∈ spec A} and that

the resolvent identity RA (z)∗ = RA∗ (z) holds for any z ∈ res A.
R
4. Let k ∈ L1 (R). Consider in L2 (R) the operator A, Af (x) = R k(x−y)f (y) dy.
Show: (i) the operator A is well-defined and bounded, (ii) the spectrum of A
is a connected set.

Exercise 3.22. 1. Let Ω ⊂ Rn be a non-empty open set and let L : Ω → M2 (C)

be a continuous matrix function such that L(x)∗ = L(x) for all x ∈ Ω. Define
an operator A in H = L2 (Ω, C2 ) by
Z
kL(x)f (x)k2C2 dx < +∞ .

Af (x) = L(x)f (x), D(A) = f ∈ H :
Ω

Show that A is self-adjoint and explain how to calculate its spectrum using
the eigenvalues of L(x).

26
 Hint: For each x ∈ Ω, let ξ1 (x) and ξ2 (x) be suitably chosen eigenvectors
of L(x) forming an orthonormal basis of C2 . Consider the map
!
ξ1 (x), f (x) C2
U : H → H, U f (x) =
ξ2 (x), f (x) C2

and the operator M = U AU ∗ .

2. In H = l2 (Z) consider the operator T given by

(
4 if n is even,
T f (n) = f (n − 1) + f (n + 1) + V (n)f (n), V (n) =
−2 if n is odd.

Calculate its spectrum.

Hint: Consider the operators

 
2 2 2 f (2n)
U : l (Z) → l (Z, C ), U f (n) := , n ∈ Z,
f (2n + 1)
X
F : `2 (Z, C2 ) → L2 (0, 1), C2 , (F f )(θ) = f (n)e2πinθ .


n∈Z

Write explicit expressions for the operators S := U T U ∗ and Sb := F SF ∗ and

use the item (1).

Exercise 3.23. Let A be a semibounded from below self-adjoint operator. Show:

hx, Axi
1. inf spec A = inf .
x∈D(A) hx, xi
x6=0

hx, Axi
2. inf spec A = inf , where Q(A) is the form domain of A.
x∈Q(A) hx, xi
x6=0

4 Spectral theory of compact operators

4.1 Fredholm’s alternative and spectra of compact opera-
tors
It is assumed that the the reader already has some knowledge of compact operators.
We recall briefly the key points. Recall first that any Hilbert space is locally compact
in the weak topology, which means that any bounded sequence contains a weakly
convergent subsequence.
A linear operator T acting from a Hilbert space H1 to a Hilbert space H2 is called
compact, if the image of the unit ball in H1 is relatively compact in H2 . We denote

27
by K(H1 , H2 ) the set of all such operators. The definition can also be reformulated
as follows: an operator T : H1 → H2 is compact iff any bounded sequence (xn ) ⊂ H1
has a subsequence (xnk ) such that T xnk converges in H2 .
Recall also that any compact operator is continuous. If A is a continuous operator
and B is a compact one, then the products AB and BA are compact. It is also known
the norm limit of a sequence compact operators is compact, and that any finite-
dimensional operator (i.e. an operator having a finite-dimensional range) is compact.
It is also known that the adjoint of a compact operator is compact (Schauder’s
theorem).
Proposition 4.1. Let T ∈ K(H1 , H2 ), then:

(a) If xn , x ∈ H1 and xn converge weakly to x, then T xn converges to T x in the

norm of H2 ,

(b) ran T and (ker T )⊥ are separable.

Proof. (a) Clearly, it is sufficient to consider the case x = 0.

Let us prove first the following assertion: (A) If (un ) is a sequence which converges
weakly to 0, then it contains a subsequence (unk ) converging to 0 in the norm.
As (un ) is bounded (being weakly convergent) and T is compact, one can extract a
subsequence (unk ) such that T unk converges to some v ∈ H2 . For any g ∈ H2 we
have
hv, gi = limhT unk , gi = hunk , T ∗ gi = 0,
i.e. v = 0, and the assertion (A) is proved.
Now assume by contradiction that T xn does not converge to 0, then there exists a
subsequence (xnk ) and  > 0 with kT xnk k ≥ , which contradicts the assertion (A).
(b) Let (eα )α∈A be a total orthonormal family in (ker T )⊥ . Take any injection
N 3 n 7→ αn ∈ A, then the sequence (eαn ) converges to 0 weakly, and T eαn converges
to 0. It follows that for any  > 0 the set

A := α ∈ A : kT eα k > ε .

is finite, and the representation

[
A= A1/n
n∈N

shows that A is a countable and, finally, that (ker T )⊥ is separable. On the other
hand, the linear hull of the vectors T eα , α ∈ A, is dense in ran T , which means that
ran T is also separable.

The previous proposition shows that it is sufficient to consider compact operators

in separable Hilbert spaces, which is assumed from now on.
Below we will consider some examples, but we prefer to discuss first some basic
questions of the spectral theory.

28
Theorem 4.2 (Fredholm’s alternative). Let T be a compact operator in a Hilbert
space H, then

(a) ker(1 − T ) has a finite dimension,

(b) ran(1 − T ) is closed and of finite codimension,

(c) ran(1 − T ) = H if and only if ker(1 − T ) = {0}.

Proof. We give the proof for the case T = T ∗ only. An interested reader may refer
to Section VI.5 in [12] for the proof of the general case.
To show (a) let us recall the Riesz theorem: if E is a normed vector space such that
the unit ball is relatively compact, then E is finite-dimensional. Let us apply this to
E = ker(1 − T ) with the same norm as in H. For every u ∈ E we have T u = u. As
the unit ball B in E is bounded, it is weakly compact. As T is compact, B = T (B)
is a relatively compact set, which means that E is finite-dimensional.
Let us prove (b). Show first that ran(1 − T ) is closed. Let (yn ) ⊂ ran(1 − T ) such
that yn converges to y in the norm of H. To show that y ∈ ran(1 − T ) we choose
xn ∈ ker(1 − T )⊥ with yn = (1 − T )xn .
We show first that the sequence (xn ) is bounded. Assume by contradiction that it is
not the case, then one can choose a subsequence with norms growing to +∞. To keep
simple notation we denote the subsequence again by xn and denote un := xn /kxn k,
then
(1 − T )xn yn
un − T un = = .
kxn k kxn k
As the norms of yn are bounded, the vectors un −T un converge to 0. As the sequence
un is bounded, one can choose a subsequence unj which is weakly convergent, then
the sequence T unj is convergent with respect to the norm to some u ∈ H due to
the compactness of T . On the other hand, as shown above, unj − T unj converge to
0, which means that u − T u = 0 and u ∈ ker(1 − T ). On the other hand, we have
un ∈ ker(1 − T )⊥ , which means that u ∈ ker(1 − T )⊥ too. This shows that u = 0,
but this contradicts to kun k = 1. This contradiction shows that (xn ) is a bounded
sequence.
As (xn ) is bounded, one can find a subsequence xnj which converges weakly to some
x∞ ∈ H, and then T xnj converge in the norm to T x∞ . Now we have xnj = ynj +T xnj ,
both sequences on the right-hand side are convergent with respect to the norm, so
xnj is also convergent to x∞ in the norm. Finally we obtain x∞ = y + T x∞ , or
y = (1 − T )x∞ , which means that y ∈ ran(1 − T ). So we proved that ran(1 − T ) is
closed.
For our particular case T = T ∗ we have ran(1 − T ) = ker(1 − T )⊥ by Proposition
3.15. Combining this with (a) we complete the proof of (b), and the item (c) is
proved too.

In a sense, the Fredholm alternative show that the operators 1 − T with compact T
behave like operators in finite dimensional spaces. We know that a linear operator

29
in a finite-dimensional space is injective if and only if it is surjective, and we see a
similar feature in the case under consideration. We remark that the Fredholm alter-
native also holds for compact operators in Banach spaces, but we are not discussing
this direction.
Theorem 4.3 (Spectrum of compact operator). Let H be an infinite-
dimensional Hilbert space and T ∈ K(H), then

(a) 0 ∈ spec T ,
(b) spec T \ {0} = specp T \ {0},
(c) we are in one and only one of the following situations:
– spec T \ {0} = ∅,
– spec T \ {0} is a finite set,
– spec T \ {0} is a sequence convergent to 0.
(d) Each λ ∈ spec T \ {0} is isolated (i.e. has a neighbodhood containing no other
values of the spectrum), and dim ker(T − λ) < ∞.

Proof. (a) Assume that 0 ∈ / spec T , then T −1 ∈ L(H), and the operator Id = T −1 T
is compact. This is possible only if H is finite-dimensional.
(b) If λ 6= 0 we have T −λ = −λ(1−T /λ), and by Fredholm alternative the condition
λ ∈ spec T is equivalent to ker(1 − T /λ) ≡ ker(T − λ) 6= {0}.
(c) Here we actually need to prove the following assertion: if (λj ) is a sequence
distinct non-zero eigenvalues of T converging to some λ ∈ C, then λ = 0. For the
proof, assume by contradiction that λ 6= 0. Let (ej ) be the normalized eigenvectors
associated with the eigenvalues λj , T ej = λj ej . One checks that the vectors en are
linearly indepedent. Denote by En the linear subspace spanned by e1 , . . . , en , then
⊥
En ⊂ En+1 and En 6= En+1 . For any n we choose un ∈ En ∩ En−1 with kun k = 1.
As T is compact and (un ) is bounded, one can extract a subsequence unk such that
the sequence (T unk ) converges, and then the sequence T unk /λnk is also convergent.
Let j > k ≥ 2. We can write
2 2
T unj T unk (T − λnj )unj (T − λnk )unk
− = − + unj − unk (4.1)
λnj λnk λnj λnk
Note that for any k we have (T − λnk )Enk ⊂ Enk −1 . On the right-hand side of (4.1)
one has unj ∈ Enj and all the other vectors are in the strictly smaller subspace Enj −1 .
Therefore, unj is orthogonal to the other three vectors, which gives the estimate
2
T unj T unk
− ≥ kunj k2 = 1.
λnj λnk
Therefore, (T unj /λnj ) cannot be a Cauchy sequence, which shows that λ = 0.
The item (d) easily follows from (c) and from the part (a) of the Fredholm alterna-
tive.

30
Finally let us apply all of the above to show the main result on the spectra of of
compact self-adjoint operators.

Theorem 4.4 (Spectrum of compact self-adjoint operator). Let T = T ∗ ∈

K(H), then can construct an orthonormal basis consisting of eigenvectors of T , and
the respective eigenvalues form a real sequence convergent to 0.

Proof. Let (λn )n≥1 be the distinct non-zero eigenvalues of T . As T is self-adjoint,

these eigenvalues are real. Set λ0 = 0, and for n ≥ 0 denote En := ker(T − λn ). One
can easily see that En ⊥ Em for n 6= m. Denote by F the linear hull of ∪n≥0 En . We
are going to show that F is dense in H.
Clearly, we have T (F ) ⊂ F . Due to the self-adjointness of T we also have T (F ⊥ ) ⊂
F ⊥ . Denote by Te the restriction of T to F ⊥ , then Te is compact, self-adjoint, and
its spectrum equals {0}, so Te = 0. But this means that F ⊥ ⊂ ker T = E0 ⊂ F and
shows that F ⊥ = {0}. Therefore F is dense in H.
Now taking an orthonormal basis in each subspace En we obtain an orthonormal
basis in the whole space H.

4.2 Integral and Hilbert-Schmidt operators

An important class of compact operators is delivered by integral operators. For
simplicity we restrict our attention to the case H = L2 (Ω), where Ω ⊂ Rd is an
open set. An interested reader may generalize all the considerations to more general
measure spaces.
Let K ∈ L1loc (Ω × Ω). Consider the operator TK defined by
Z
TK f (x) = K(x, y)f (y) dy (4.2)
Ω

on bounded functions with compact supports. We would like to understand first

under which conditions the expression (4.2) defines a bounded operator in H. A
standard result in this direction is delivered by the following theorem.

Theorem 4.5 (Schur’s test). Assume that

Z Z
M1 = sup K(x, y) dy < ∞, M2 = sup K(x, y) dx < ∞,
x∈Ω Ω y∈Ω Ω

then the above√expression (4.2) defines a continuous linear operator TK with the
norm kTK k ≤ M1 M2 .

31
Proof. We have, using the Cauchy-Schwarz inequality,
Z p p 2
2
|TK u(x)| ≤ |K(x, y)| |K(x, y)| · |u(y)|dy
Z Ω Z
≤ |K(x, y)|dy |K(x, y)| · |u(y)|2 dy
Ω Z Ω

≤ M1 |K(x, y)| · |u(y)|2 dy,

ZΩ Z
2
and kTK uk ≤ M1 K(x, y) |u(y)|2 dydx ≤ M1 M2 kuk2 .
Ω Ω

To obtain a class of compact integral operators we introduce the following notion.

We say that an operator T ∈ L(H) is Hilbert-Schmidt if for some orthonormal basis
(en ) of H the sum X
kT k22 = kT en k2 (4.3)
n

is finite.
Proposition 4.6 (Hilbert-Schmidt norm). Let T be a Hilbert-Schmidt operator,
then the quantity kT k2 (called the Hilbert-Schmidt norm of T ) does not depend on
the choice of the basis, and kT k ≤ kT k2 . Moreover, the adjoint operator T ∗ is also
Hilbert-Schmidt with kT ∗ k2 = kT k2 .

Proof. Let (en ) and (fn ) be two orthonormal bases. Using the Parseval identity we
have
2 2
X XX XX X
kT en k2 = hfm , T en i = hT ∗ fm , en i = kT ∗ fm k2 .
n n m m m m

This shows that the expression (4.3) is independent of the choice of (en ) and that
kT ∗ k2 = kT k2 . To show kT k ≤ kT k2 , let x ∈ H with xn := hen , xi, then
X 2 X 2 X X
kT xk2 = xn T en ≤ |xn |kT en k ≤ |xn |2 kT en k2 = kT k22 kxk2 .
n n n n

Proposition 4.7. Any Hilbert-Schmidt operator is compact.

Proof. For any x ∈ H we have T x = ∞

P
n=1 hen , xiT en . For N ≥ 1 introduce the
operators TN by TN x = N
P
he
n=1 n , xiT e n . One has
N →∞
X
kT − TN k2 ≤ kT − TN k22 = kT en k2 −−−→ 0,
n≥N +1

and T is compact being the norm-limit of the finite-dimensional operators TN .

The following proposition describes the class of integral operators which are Hilbert-
Schmidt and makes a link with integral operators.

32
Proposition 4.8 (Integral Hilbert-Schmidt operators). Let H = L2 (Ω). An
operator T in H is Hilbert-Schmidt iff there exists an integral kernel K ∈ L2 (Ω × Ω)
such that T = TK , cf. Eq. (4.2), and in that case kTK k2 = kKkL2 (Ω×Ω) .

Proof. Let first K ∈ L2 (Ω × Ω). Let us show that the associated operator TK
is Hilbert-Schmidt. Let (en ) be an orthonormal basis in H, then the functions
em,n (x, y) = em (x)en (y) form an orthonormal basis in H ⊗ H ' L2 (Ω × Ω). There
holds
X X 2
X Z Z  2
2
kTK en k = hem , TK en i = em (x) K(x, y)en (y)dy dx
n m,n m,n Ω Ω
X Z Z 2 X 2
= em (x)en (y)K(x, y)dx dy = hem,n , Ki = kKk2L2 (Ω×Ω) .
m,n Ω Ω m,n

Now let T be an arbitrary Hilbert-Schmidt operator in H. We have, for any u ∈ H

and with un := hen , ui,
X X
Tu = hen , uiT en = hen , uihem , T en iem .
n m,n

Take X X
K(x, y) = en (y)hem , T en iem (x) = hem , T en iem,n (x, y).
m,n m,n

One easily checks that K ∈ L2 (Ω × Ω), that T = TK , and that the remaining
properties hold as well.

One can easily see that the operator TK is self-adjoint iff K(x, y) = K(y, x) for a.e.
(x, y) ∈ Ω × Ω. Together the Hilbert-Schmidt condition (Proposition 4.8) this gives
an important class of self-adjoint compact operators to which the previous consid-
erations can be applied. Taking an orthonormal basis consisting of eigenfunctions
we see that a compact self-adjoint operator T is a Hilbert-Schmidt one iff the series
X
λ2n = kT k22
n

is convergent, where λn denote the non-zero eigenvalues of T taking according to

their multiplicities. Moreover, by Proposition 4.8, for T = TK one has the exact
equality (trace formula) X
λ2n = kKk2L2 (Ω×Ω) ,
n

which may be used to estimate the eigenvalues using the integral kernel.

4.3 Operators with compact resolvent

Let us continue the discussion of operators defined by forms, see Section 2.

33
Proposition 4.9. In the situation of Theorem 2.3 one has T −1 ∈ L(H, V).

Proof. For any u ∈ D(T ) we have:

kukH kT ukH ≥ hu, T uiH = |a(u, u)| ≥ αkuk2V ≥ CαkukV kukH ,
i.e. kT ukH ≥ CαkukV and kT −1 ukV ≤ (Cα)−1 kukH .

This gives an important consequence:

Corollary 4.10. In the situation of Theorem 2.3, assume that the embedding j :
V → H is compact, then T −1 is a compact operator.

Proof. Indeed we have T −1 = jL, where L is the operator T −1 viewed as an operator

from H to V. Hence T −1 is compact as a composition of a bounded operator and a
compact one.

The above can be applied to a variety of cases. For example, take the Dirichlet
Laplacian A0 defined in example 2.14. If Ω is relatively compact, then the embedding
of V = H01 (Ω) to H = L2 (Ω) is compact. Therefore, the operator L = (A0 + 1)−1
is compact. Moreover, it is self-adjoint due to the previous considerations. By
Theorem 4.4 there exists an orthonormal basis (en ) of L2 (Ω) such that Len = λn en ,
where λn is a real-valued sequenece converging to 0. By elementary operations,
en ∈ D(A0 ) and A0 en = µn en with
1
µn = − 1.
λn
It is an easy exercise to show that the spectrum of A0 is exactly the union of all the
µn and that µn → +∞ as n → +∞.
The values µn are called the Dirichlet eigenvalues of the domain Ω. It is an important
domain of the modern analysis to study the relations between the geometric and
topological properties of Ω and its Dirichlet eigenvalues.
The preceding example can be easily generalized. More precisely, we say that an
operator A with res A 6= ∅ has a compact resolvent if Rλ (A) is a compact operator
for all λ ∈ res A. One can easily check that it is sufficient to check this property at
a single value λ.
Similar to the preceding constructions one can show:
Proposition 4.11 (Spectra of operators with compact resolvents). Let T be
a self-adjoint operator with a compact resolvent in an infinite-dimensional Hilbert
space, then:
• spec T = specp T ,
• the eigenvalues of T form a sequence converging to ∞.
The proof is completely the same as for the Dirichlet Laplacian if we manage to
show that spec T 6= R. In principle this can be done in a direct way, but we prefer
to show it later using the spectral theorem, see Example 5.23 below.

34
4.4 Schrödinger operators with growing potentials
Let us discuss a particular class of operators with compact resolvents.
Recall the following classical criterion of compactness in L2 (Rd ) (sometimes referred
to as the Riesz-Kolmogorov-Tamarkin criterion):

Proposition 4.12. A subset A ⊂ L2 (Rd ) is relatively compact in L2 (Rd ) if and only

if the following three conditions are satisfied:

(a) A is bounded,

(b) there holds Z

|u(x)|2 dx → 0 as R → ∞
|x|≥R

uniformly for u ∈ A,

(c) kuh −uk → 0 as h → 0 uniformly for u ∈ A. Here, for h ∈ Rd and v ∈ L2 (Rd ),

the symbol vh denote the function defined by vh (x) = v(x + h).

An interested reader may refer to [8] for the proof and various generalizations.
Now let W ∈ L2loc (Rd ) and W ≥ 0. Consider the operator T = −∆ + W defined
as the Friedrichs extension starting from Cc∞ (Rd )and discussed in Example 2.20.
We know already that T is a self-adjoint and semibounded from below operator in
H = L2 (Rd ). We would like to identify a reasonable large class of potentials W for
which T has a compact resolvent.

Theorem 4.13. For r ≥ 0 denote

w(r) := inf W (x).

|x|≥r

If limr→+∞ w(r) = +∞, then the associated Schrödinger operator T = −∆ + W has

a compact resolvent.

Proof. As follows from Example 2.20, it is sufficient to show that the embedding of
1
V = HW (Rd ) to L2√(Rd ) is a compact operator, where V is equipped with the norm
2 2
kukW = kukH 1 + k W ukL2 . Let B be the unit ball in V. We will show that B is
relatively compact in L2 (Rd ) using Proposition 4.12.
The condition (a) holds due to the inequality kukL2 ≤ kukW . The condition (b)
follows from
√
k W uk2L2 kuk2W
Z Z
1
|u(x)|2 dx ≤ W (x)|u(x)|2 ≤ ≤ .
|x|≥R w(R) |x|≥R w(R) w(R)

35
For the condition (c) we have:
Z Z 1 Z
2 d 2
u(x + h) − u(x) dx = u(x + th)dt dx
Rd Rd 0 dt
Z Z 1 2
Z Z 1
2 2
= h · ∇u(x + th)dt dx ≤ h ∇u(x + th) dt dx
Rd 0 Rd 0
Z 1Z
2
≤ h2 ∇u(x + th) dx dt = h2 k∇uk2L2 ≤ h2 kuk2W .
0 Rd

The assumption of Theorem 4.13 is rather easy to check, but this condition in not
optimal one. For example, it is known that the operator −∆ + W with W (x1 , x2 ) =
x21 x22 has a compact resolvent, while the condition cleraly fails.
A rather simple necessary and sufficient condition is known in the one-dimensional
case:

Proposition 4.14 (Molchanov criterium). The operator T = −d2 /dx2 + W has

a compact resolvent iff Z x+δ
lim W (s)ds = +∞
x→∞ x
for any δ > 0.

Necessary and sufficient conditions are also available for the multi-dimensional case,
but their form is much more complicated. An advanced reader may refer to the
paper [14] for the discussion of such questions.

36
5 Spectral theorem
Some points in this section are just sketched to avoid technicalities. A more detailed
presentation can be found in [5, Chapter 2] or in [16, Section 12.7].
The aim of the present section is to define, for a given self-adjoint operator T , the
operators f (T ), where f are sufficiently general functions.
To be provided with a certain motivation, let T be either a compact self-adjoint
operator or a self-adjoint operator with a compact resolvent in a Hilbert space H.
As shown in the previous section, there exists an orthonormal basis (en ) in H and
real numbers λn such that, with
X
Tx = λn hen , xien for all x ∈ D(T ),
n

and the domain D(T ) is characterized by

n o
2
X
D(T ) = x ∈ H : λ2n hen , xi <∞ .
n

For f ∈ C0 (R) one can define an operator f (T ) ∈ L(H) by

X
f (T )x = f (λn )hen , xien .
n

This map f 7→ f (T ) enjoys a number of properties. For example, (f g)(T ) =

f (T )g(T ), f (T ) = f (T )∗ , spec f (T ) = f (spec T ) etc. The existence of such a
construction allows one to write rather explicit expressions for solutions of some
equations. For example, one can easily show that the initial value problem

−ix0 (t) = T x(t), x(0) = y ∈ D(T ), x : R → D(T ),

has a solution that can be written as x(t) = ft (T )y with ft (x) = eitx . Informally
speaking, for a large class of equations involving the operator T one may first assume
that T is a real constant and obtain a formula for the solution, and then one can
give this formula an operator-valued meaning using the above map f 7→ f (T ).
Moreover, if we introduce the map U : H → `2 (N) defined by U x =: (xn ), xn =
hen , xi, then the operator U T U ∗ becomes a multiplication operator (xn ) 7→ (λn xn ),
cf. Example 3.10.
At this point, all the preceding facts are proved for compact self-adjoint operators
and self-adjoint operator with a compact resolvent only. The aim of the present
section is to develop a similar theory for general self-adjoint operators.
To avoid potential misunderstanding let us recall that C0 (R) denotes the class of
the continuous functions f : R → C with lim|x|→+∞ f (x) = 0 equipped with the sup-
norm. This should not be confused with the set C 0 (R) of the continuous functions
on R.

37
5.1 Continuous functional calculus
We say that a function f : C → C belongs to C ∞ (C) if the function of two real
variables R2 3 (x, y) 7→ f (x + iy) ∈ C belongs to C ∞ (R2 ). In the similar way one
defines the classes Cc∞ (C), C k (C) etc. In what follows we always use the notation
z+z z−z
<z =: x, =z =: y for z ∈ C. Using x = and y = , for f ∈ C 1 (C) one
2 2i
defines the derivative
∂ 1 ∂ ∂ 
:= +i
∂z 2 ∂x ∂y
Clearly, ∂g/∂z = 0 if g is a holomorphic function. Recall the Stokes formula written
in this notation: if f ∈ C ∞ (C) and Ω ⊂ C is a domain with a sufficiently regular
boundary, then ZZ I
∂f 1
dx dy = f dz.
Ω ∂z 2i ∂Ω
The following fact is actually known, but is presented in a slightly unusual form.

Lemma 5.1 (Cauchy integral formula). Let f ∈ Cc∞ (C), then for any w ∈ C
we have ZZ
1 ∂f 1
dx dy = f (w).
π C ∂z w − z

Proof. We note first that the singularity 1/z is integrable in two dimensions, and
the integral is well-defined. Let Ω be a large ball containing the support of f and the
point w. For small ε > 0 denote Bε := {z ∈ C : |z − w| ≤ ε}, and set Ωε := Ω \ Bε .
Using the Stokes formula we have:
ZZ ZZ
1 ∂f 1 1 ∂f 1
dx dy = dx dy
π C ∂z w − z π Ω ∂z w − z
ZZ ZZ
1 ∂f 1 1 ∂  1 
= lim dx dy = lim f (z) dx dy
ε→0 π Ωε ∂z w − z ε→0 π Ωε ∂z w−z
I
1 1
= lim f (z) dz
ε→0 2πi ∂Ω
ε
w−z
I I
1 1 1 1
= f (z) dz − lim f (z) dz.
2πi ∂Ω w−z ε→0 2πi |z−w|=ε w−z

The first term on the right-hand side is zero, because f vanishes at the boundary of
Ω. The second term can be calculated explicitly:
2π
iεeit dt
I Z
1 1 1
lim f (z) dz = lim f (w + εeit )
ε→0 2πi |z−w|=ε w−z ε→0 2πi 0 w − (w + εeit )
Z 2π
1
= − lim f (w + εeit )dt = −f (w),
ε→0 2π 0

which gives the result.

38
The main idea of the subsequent presentation is to define the operators f (T ), for
a self-adjoint operator T , using an operator-valued generalization of the Cauchy
integral formula.
Introduce first some notation. For z ∈ C we write
p
hzi := 1 + |z|2 .

For β < 0 denote by Sβ the set of the smooth functions f : R → C satisfying the
estimates
f (n) (x) ≤ cn hxiβ−n

S n ≥ 0 and x ∈ R, where the positive constant cn may depend on f . Set

for any
A := β<0 Sβ ; one can show that A is an alebra. Moreover, if f = P/Q, where P
and Q are polynomials with deg P < deg Q and Q(x) 6= 0 for x ∈ R, then f ∈ A.
For any n ≥ 1 one can introduce the norms on A:
n Z
X
kf kn := f (r) (x) hxir−1 dx.
r=0 R

One can easily see that the above norms on A induce continuous embeddings A →
C0 (R). Moreover, one can prove that Cc∞ (R) is dense in A with respect to any norm
k · kn .
Now let f ∈ C ∞ (R). Pick n ∈ N and a smooth function τ : R → R such that
τ (s) = 1 for |s| < 1 and τ (s) = 0 for |s| > 2. For x, y ∈ R set σ(x, y) := τ (y/hxi).
Define fe ∈ C ∞ (C) by
X n r

(r) (iy)
fe(z) = f (x) σ(x, y).
r=0
r!

Clearly, for x ∈ R we have fe(x) = f (x), so fe is an extension of f . One can check

the following identity:
" n
∂ fe 1 X (r) (iy)r  (iy)n
  1
= f (x) σx + iσy + f (n+1) (x) σ. (5.1)
∂z 2 r=0 r! 2 n!

Now let T be a self-adjoint operator in a Hilbert space H. For f ∈ A define an

operator f (T ) in H by
ZZ
1 ∂ fe
f (T ) := (T − z)−1 dx dy. (5.2)
π C ∂z
This integral expression is called the Helffer-Sjöstrand formula. We need to show
several points: that the integral is well-defined, that it does not depend in the choice
of σ and n etc. This will be done is a series of lemmas.
Note first that, as shown in Proposition 3.16, we have the norm estimate k(T −
z)−1 k ≤ 1/|=z|, and one can see from (5.1) that ∂f e /∂z(x + iy) = O(y n ) for any

39
fixed x, so the subintegral function in (5.2) is locally bounded. By additional tech-
nical efforts one can show that the integral is convergent and defines an continuous
operator with kf (T )k ≤ ckf kn+1 for some c > 0. Using this observation and the
density of Cc∞ (R) in A the most proofs will be provided for f ∈ Cc∞ and extended
to A and larger spaces using the standard density arguments.
Lemma 5.2. If F ∈ Cc∞ (C) and F (z) = O(y 2 ) as y → 0, then
ZZ
1 ∂F
A := (T − z)−1 dx dy = 0.
π C ∂z

Proof. By choosing a sufficiently large N > 0 one may assyme that the support
of F is contained in Ω := {z ∈ C : |x| < N, |y| < N }. For small ε > 0 define
Ωε := {z ∈ C : |x| < N, ε < |y| < N }. Using the Stokes formula we have
ZZ I
1 ∂F −1 1
A = lim (T − z) dx dy = lim F (z) (T − z)−1 dz.
ε→0 π Ωε ∂z ε→0 2πi ∂Ωε

The boundary ∂Ωε consists of eight segments. The integral over the vertical segments
and over the horizontal segments with y = ±N are equal to 0 because the function
F vanishes on these segments. It remains to estimate the integrals over the segments
with y = ±ε. Here we have k(T − z)−1 k ≤ ε−1 and
I 
1 
kAk ≤ lim |F (x + iε)| + |F (x − iε)| ε−1 dx = 0.
ε→0 2π ∂Ω
ε

Corollary 5.3. For f ∈ A the integral in (5.2) is independent of the choice of n ≥ 1

and σ.

Proof. For f ∈ Cc∞ (C) the assertion follows from the definition of fe and Lemma
5.2. This is extended to A using the density arguments.
Lemma 5.4. Let f ∈ Cc∞ (R) with supp f ∩ spec T = ∅, then f (T ) = 0.

Proof. If f ∈ Cc∞ (R), then obviously fe ∈ Cc∞ (C). One can find a finite family
of closed curves γr which do not meet the spectrum of T and enclose a domain U
containing supp fe. Using the Stokes formula we have
ZZ X 1 I
1 ∂ fe −1
f (T ) = (T − z) dx dy = fe(z) (T − z)−1 dz.
π U ∂z r
2πi γr

All the terms in the sum are zero, because fe vanishes on γr .

Lemma 5.5. For f, g ∈ A one has (f g)(T ) = f (T )g(T ).

Proof. By the density arguments is it sufficient to consider the case f, g ∈ Cc∞ (R).
Let K and L be large balls containing the supports of fe and ge respectively. Using
the notation w = u + iv, u, v ∈ R, one can write:
ZZZZ e
1 ∂ f ∂e
g
f (T )g(T ) = 2 (T − z)−1 (T − w)−1 dx dy du dv.
π ∂z ∂w
K×L

40
Using the resolvent identity
1 1
(T − z)−1 (T − w)−1 = (T − w)−1 − (T − z)−1
w−z w−z
we rewrite the preceding integral in the form
ZZ  Z Z ∂ fe 1
1 ∂e
g −1

f (T )g(T ) = 2 (T − w) dx dy du dv
π L ∂w K ∂z w − z
ZZ  Z Z ∂e
1 ∂ fe −1 g 1 
− 2 (T − z) du dv dx dy.
π K ∂z L ∂w w − z

By Lemma 5.1 we have

ZZ ZZ
∂ fe 1 ∂e
g 1
dx dy = πf (w), du dv = −πg(z),
K ∂z w − z L ∂w w − z
and we arrive at
ZZ ZZ
1 ∂eg −1 1 ∂ fe
f (T )g(T ) = fe(w) (T − w) du dv + ge(z) (T − z)−1 dx dy
π L ∂w π K ∂z
ZZ
1 ∂(f ge)
e
= (T − z)−1 dx dy
π K∪L ∂z
∂(fege − ffg)
ZZ f ZZ
1 ∂f g −1 1
= (T − z) dx dy + (T − z)−1 dx dy
π C ∂z π C ∂z
∂(f ge − f g)
ZZ
1 e f
= (f g)(T ) + (T − z)−1 dx dy.
π C ∂z

By direct calculation one can see that (ffg − fege)(z) = O(y 2 ) for small y, and Lemma
5.2 shows that the second integral is zero.

Lemma 5.6. Let w ∈ C \ R. Define a function rw by rw (z) = (z − w)−1 . Then

rw (T ) = (T − w)−1 .

Proof. We provide just the main line of the proof without technical details (they
can be easily recovered). Use first the independence of n and σ. We take n = 1 and
put σ(z) = τ (λy/hxi) where λ > 0 is sufficiently large, to have w ∈
/ supp σ. Without
loss of generality we assume =w > 0. For large m > 0 consider the region

hxi
Ωm := {z ∈ C : |x| < m, < y < 2m}.
m
Using the definition and the Stokes formula we have
ZZ I
1 ∂e
rw −1 1
rw (T ) = lim (T − z) dx dy = lim rew (z) (T − z)−1 dz.
m→∞ π Ωm ∂z m→∞ 2πi ∂Ωm

41
By rather technical explicit estimates (which are omitted here) one can show that
I  
lim rew (z) − rw (z) (T − z)−1 dz = 0.
m→∞ ∂Ωm

and we arrive at
I
1 1
rw (T ) = lim (T − z)−1 dz.
2πi m→∞ ∂Ωm z−w
For sufficiently large m one has the inclusion w ∈ Ωm . For any f, g ∈ H the function
C 3 z 7→ hf, (T − z)−1 gi ∈ C is holomorphic in Ωm , so applying the Cauchy formula,
for large m we have
I
1 1
f, (T − z)−1 g dz = f, (T − w)−1 g ,
2πi ∂Ωm z − w
which shows that rw (T ) = (T − w)−1 .
Lemma 5.7. For any f ∈ A we have:

(a) f (T ) = f (T )∗ ,
(b) f (T ) ≤ kf k∞ .

Proof. The item (a) follows directly from the equalities

∗
(T − z)−1 = (T − z)−1 , fe(z) = fe(z).
p
To show (b), take an arbitrary c > kf k∞ and define g(s) := c − c2 − |f (s)|2 . One
can show that g ∈ A. There holds f f − 2cg + g 2 = 0, and using the preceding
lemmas we obtain f (T )∗ f (T ) − cg(T ) − cg(T )∗ + g(T )∗ g(T ) = 0, and

∗
f (T )∗ f (T ) + c − g(T ) c − g(T ) = c2 .

Let ψ ∈ H. Using the preceding equality we have:

2 2
2
f (T )ψ ≤ f (T )ψ + c − g(T ) ψ
D E D
∗
E
= ψ, f (T )∗ f (T )ψ + ψ, c − g(T ) c − g(T ) ψ
= c2 kψk2 .
As c > kf k∞ was arbitrary, this concludes the proof.

All the preceding lemmas put together lead us to the following fundamental result.
Theorem 5.8 (Spectral theorem, continuous functional calculus). Let T be
a self-adjoint operator in a Hilbert space H. There exists a unique linear map
C0 (R) 3 f 7→ f (T ) ∈ L(H)
with the following properties:

42
 • f 7→ f (T ) is an algebra homomorphism,

• f (T ) = f (T )∗ ,

• kf (T )k ≤ kf k∞ ,

/ R and rw (s) = (s − w)−1 , then rw (T ) = (T − w)−1 ,

• if w ∈

• if supp f does not meet spec T , then f (T ) = 0.

Proof. Existence. If one replaces C0 by A, everything is already proved. But A

is dense in C0 (R) in the sup-norm, because Cc∞ (R) ⊂ A, so we can use the density
argument.
Uniqueness. If we have two such maps, they coincide on the functions f which
are linear combinations of rw , w ∈ C \ R. But such functions are dense in C0 by
the Stone-Weierstrass theorem, so by the density argument both maps coincide on
C0 .
Remark 5.9. • One may wonder why to introduce the class of functions A: one
could just start by Cc∞ which is also dense in C0 . The reason in that we have
no intuition on how the operator f (T ) should look like if f ∈ Cc∞ . On the
other hand, it is naturally expected that for rw (s) = (s − w)−1 we should have
rw (T ) = (T − w)−1 , otherwise there are no reasons why we use the notation
rw (T ). So it is important to have an explicit formula for a sufficiently large
class of functions containing all such rw .

• The approach based on the Helffer-Sjöstrand formula, which is presented here,

is relatively new, and it allows one to consider bounded and unbounded self-
adjoint operators simultaneously. The same results can be obtained by other
methods, starting e.g. with polynomials instead of the resolvents, which is a
more traditional approach, see, for example, Sections VII.1 and VIII.3 in the
book [12].

5.2 Borelian functional calculus and L2 representation

Now we would like to extend the functional calculus to more general functions, not
necessarily continuous and not necessarily vanishing at infinity.
Definition 5.10 (Invariant and cyclic subspaces). Let H be a Hilbert space,
L be a closed linear subspace of H, and T be a self-adjoint linear operator in H.
Let T be bounded. We say that L is an invariant subspace of T (or just T -invariant)
if T (L) ⊂ L. We say that L is a cyclic subspace of T with cyclic vector v if L coincides
with the closed linear hull of all vectors p(T )v, where p are polynomials.
Let T be general. We say that L is an invariant subspace of T (or just T -invariant)
if (T − z)−1 (L) ⊂ L for all z ∈ / R. We say that L is a cyclic subspace of T with
cyclic vector v if L coincides with the closed linear space of all vectors (T − z)−1 v
with z ∈ / R.

43
Clearly, if L is T -invariant, then L⊥ is also T -invariant.

Proposition 5.11. Both definitions of an invariant/cyclic subspace are equivalent

for bounded self-adjoint operators.

Proof. Let T = T ∗ ∈ L(H). We note first that res T is a connected set.

Let a closed subspace L be T -invariant in the sense of the definition for bounded
operators. If z ∈ C and |z| > kT k, then z ∈
/ spec T and
∞
−1
 T −1 X −n−1 n
(T − z) = −z 1 − = z T .
z n=0

If x ∈ L, then T n x ∈ L for any n. As the series on the right hand side converges in
the operator norm sense and as L is closed, (T − z)−1 x belongs to L.
Let us denote W = z ∈ res T : (T − z)−1 (L) ⊂ L . As just shown, W is non-


empty. On the other hand, W is closed in res T in the relative topology: if x ∈ L,

zn ∈ W and zn converge to z ∈ W , then (T − zn )−1 x ∈ L and (T − zn )−1 x converge
to (T − z)−1 x. On the other hand, W is open: if z0 ∈ W and |z − z0 | is sufficiently
small, then X
(T − z)−1 = (z − z0 )n (T − z0 )−n−1 ,
n=0
−1
see (3.4), and (T − z) L ⊂ L. Therefore, W = res T , which shows that L is
T -invariant in the sense of the definition for general operators.
Now let T = T ∗ ∈ L(H), and assume that L is T -invariant in the sense of the
definition for general operators, i.e. (T − z)−1 (L) ⊂ L for any z ∈ / R. Pick any
z ∈/ R and any f ∈ L. We can represent T f = g + h, where g ∈ L and h ∈ L⊥
are uniquely defined vectors. As L⊥ is T -invariant, (T − z)−1 h ⊂ L⊥ . On the other
hand

(T − z)−1 h = (T − z)−1 (T f − g)
= (T − z)−1 (T − z)f + zf − g

= f + (T − z)−1 (zf − g).

As zf −g ∈ L, both vectors on the right-hand side are in L. Therefore, (T −z)−1 h ∈

L, and finally (T − z)−1 h = 0 and h = 0, which shows that T f = g ∈ L. The
equivalence of the two definitions of an invariant subspace is proved.
On the other hand, for both definitions, L is T -cyclic with cyclic vector v iff L is the
smallest T -invariant subspace containing v. Therefore, both definitions of a cyclic
subspace also coincide for bounded self-adjoint operators.

Theorem 5.12 (L2 representation, cyclic case). Let T be a self-adjoint linear

operator in H and let S := spec T . Assume that H is a cyclic subspace for T with
a cyclic vector v, then there exists a bounded measure µ on S with µ(S) ≤ kvk2 and
a unitary map U : H → L2 (S, dµ) with the following properties:

44
 • a vector x ∈ H is in D(T ) iff hU x ∈ L2 (S, dµ), where h is the function on S
given by h(s) = s,

• for any ψ ∈ U D(T ) there holds U T U −1 ψ = hψ.

In other words, T is unitarily equivalent to the operator Mh of the multiplciation by

h in L2 (S, dµ).

Proof. Step 1. Consider the map φ : C0 (R) → C defined by φ(f ) = v, f (T )v .

Let us list the properties of this map:

• φ is linear,

• φ(f ) = φ(f ),

• if f ≥ 0, then φ(f ) ≥ 0. This follows from

p p p 2
φ(f ) = v, f (T )v = v, f (T ) f (T )v = f (T )v .

• φ(f ) ≤ kf k∞ kvk2 .

By the Riesz representation theorem there exists a uniquely defined regular Borel
measure µ such that Z
φ(f ) = f dµ for all f ∈ C0 (R).
R

Moreover, for supp f ∩ S = ∅ we have f (T ) = 0 and φ(f ) = 0, which means that

supp µ ⊂ S, and we can write the above as
Z
v, f (T )v = f dµ for all f ∈ C0 (R). (5.3)
S

Step 2. Consider the map Θ : C0 (R) → L2 (S, dµ) defined by Θf = f . We have

Z
hΘf, Θgi = f g dµ = φ(f g)
S
= v, f (T )∗ g(T )v = f (T )v, g(T )v .

Denote M := f (T )v : f ∈ C0 (R) ⊂ H, then the preceding equality means that
the map
U : H ⊃ M → C0 (R) ⊂ L2 (S, dµ), U f (T )v = f,

is one-to-one and isometric. Moreover, M is dense in H, because v is a cyclic vector.

Furthermore, C0 (R) is a dense subspace of L2 (S, dµ), as µ is regular. Therefore, U
is uniquely extended to a unitary map from H to L2 (S, dµ), and we denote this
extension by the same symbol.

45
Step 3. Let f, fj ∈ C0 (R) and ψj := fj (T )v, j = 1, 2. There holds

ψ1 , f (T )ψ2 = f1 (T )v, f (T )f2 (T )v

= v, (f 1 f f2 )(T )v
Z
= f f 1 f2 dµ
S
= hU ψ1 , Mf U ψ2 i,

where Mf is the operator of the multiplication by f in L2 (S, dµ). In particular, for

any w ∈/ R and rw (s) = (s − w)−1 we obtain U rw (T )U ∗ ξ = rw ξ for all ξ ∈ L2 (S, dµ).
The operator U maps the set ran rw (T ) ≡ D(T ) to the range of Mrw . In other
words, U is a bijection from D(T ) to

ran Mrw = φ ∈ L2 (S, dµ) : x 7→ xφ(x) ∈ L2 (S, dµ) = D(Mh ).



Therefore, if ξ ∈ L2 (S, dµ), then ψ := rw ξ ∈ D(Mh ),

T rw (T )U ∗ ξ = (T − w)rw (T )U ∗ ξ + wrw (T )U ∗ ξ = U ∗ ξ + wrw (T )U ∗ ξ

and, finally,

U T U ∗ ψ = U T U ∗ rw ξ = U T rw (T )U ∗ ξ = U U ∗ ξ + wrw (T )U ∗ ξ

= ξ + wrw ξ = hψ.

Theorem 5.13 (L2 representation). Let T be a self-adjoint operator in a Hilbert

space H with spec T =: S. Then there exists N ⊂ N, a finite measure µ on S × N
and a unitary operator U : H → L2 (S × N, dµ) with the following properties.

• Let h : S × N → R be given by h(s, n) = s. A vector x ∈ H belongs to D(T )

iff hU x ∈ L2 (S × N, dµ),
• for any ψ ∈ U D(T ) there holds U T U −1 ψ = hψ.

Proof. Using the induction one can find N ⊂ N and non-empty closed mutually
orthogonal subspaces Hn ⊂ H with the following properties:
L
• H = n∈N Hn ,
• each Hn is a cyclic subspace of T with cyclic vector vn satisfying kvn k ≤ 2−n .

The restriction Tn of T to Hn is a self-adjoint operator in Hn , and one can apply to all

these operators Theorem 5.12, which gives associated measures µn with µ(S) ≤ 4−n ,
2

Un : Hn → L (S, dµn ). Now one can define a measure µ on S × N
and unitary maps
by µ Ω × {n} = µn (Ω), and a unitary map
M M
U :H≡ Hn → L2 (S × N, dµ) ≡ L2 (S, dµn )
n∈N n∈N

by U (ψn ) = (Un ψn ), and one can easily check that all the properties are verified.

46
Remark 5.14. • The previous theorem shows that any self-adjoint operator
is unitarily equivalent to a multiplication operator in some L2 space, and
this multiplication operator is sometimes called a spectral representation of T .
Clearly, such a representation is not unique, for example, the decomposition
of the Hilbert space in cyclic subspaces is not unique.

• The cardinality of the set N is not invariant. The minimal cardinality among
all possible N is called the spectral multiplicity of T , and it generalizes the
notion of the multiplicity for eigenvalues. Calculating the spectral multiplicity
is a non-trivial problem.

Theorem 5.13 can be used to improve the result of Theorem 5.8. In the rest of the
section we use the function h and the measure µ from Theorem 5.13 without further
specifications.
Introduce the set B∞ consisting of the bounded Borel functions f : R → C. In
B∞
what follows, we say that fn ∈ B∞ converges to f ∈ B∞ and write fn −−→ f if the
following two conditions hold:

• there exists c > 0 such that kfn k∞ ≤ c,

• fn (x) → f (x) for all x.

Definition 5.15 (Strong convergence). Wa say that a sequence An ∈ L(H)

converges strongly to A ∈ L(H) and write A = s − lim An if Ax = lim An x for any
x ∈ H.

Theorem 5.16 (Borel functional calculus). (a) Let T be a self-adjoint oper-

ator in a Hilbert space H. There exists a map B∞ 3 f 7→ f (T ) ∈ L(H)
extending the map from Theorem 5.8 and satisfying the same properties except
that one can improve the estimate kf (T )k ≤ kf k∞ by kf (T )k ≤ kf k∞,T .
∞ B
(b) This extension is unique if we assume that the condition fn −−→ f implies
f (T ) = s − lim fn (T ).

Proof. Consider the map U from Theorem 5.8. Then it is sufficient to define
f (T ) := U ∗ Mf ◦h U , then one routinely check that all the properties hold, and (a) is
proved.
To prove (b) we remark first that the map just defined satisfies the requested con-
B∞
dition: If x ∈ L2 (S, dµ) and fn −−→ f , then fn (h)x converges to f (h)x in the
norm of L2 (S × N, dµ) by the dominated convergence. But this means exactly that
f (T ) = s − lim fn (T ).
On the other hand, C0 (R) is obviously dense in B∞ with respect to the B∞ conver-
gence, which proves the uniqueness of the extension.

We have a series of important corollaries, whose proof is an elementary modification

of the constructions given for the multiplication operator in Example 3.6.

47
Corollary 5.17. • spec T = essµ ran h,

• for any f ∈ B∞ one has spec f (T ) = essµ ran f ◦ h,

• in particular, kf (T )k = essµ sup |f ◦ h|.

Example 5.18. One can also define the operators ϕ(T ) with unbounded functions ϕ
by ϕ(T ) = U ∗ Mϕ◦h U . These operators are in general unbounded, but they are self-
adjoint for real-valued ϕ; this follows from the self-adjointness of the multiplication
operators Mϕ◦h .

Example 5.19. The usual Fourier transform is a classical example of a spectral

representation. For example, Take H = L2 (R) and T = −id/dx with the natural
domain D(T ) = H 1 (R). If F is the Fourier transform, then FT F is exactly the
operator of multiplication x 7→ xf (x), and spec T = R.
In particular, for bounded Borel functions f : R → C one can define the operators
f (T ) by f (T )h = F ∗ Mf F, where Mf is the operator of multiplication by f , i.e. in
general one obtains a pseudodifferential operator.
Let us look at some particular examples. Consider the shift operator A in H which
is defined by Af (x) = f (x + 1). It is a bounded operator, and for any u ∈ S(R)
we have FAF ∗ u(p) = eip u(p). This means that A = eiT , and this gives the relation
spec A = {z : |z| = 1}. On may also look at the operator B defined by
Z x+1
Bf (x) = f (t)dt.
x−1

Using the Fourier transform one can show that B = ϕ(T ), where ϕ(x) = 2 sin x/x
with spec B = ϕ(R).

Example 5.20. For practical computations one does not need to have the canonical
representation from Theorem 5.13 to construct the Borel functional calculus. It is
sufficient to represent T = U ∗ Mf U , where U : H → L2 (X, dµ) and Mf is the
multiplcation operator by some function f . Then for any Borel function ϕ one can
put ϕ(T ) = U ∗ Mϕ◦f U .
For example, for the free Laplacian T in H = L2 (Rd ) the above is realized with
X = Rd and U being the Fourier transform, and with f (p) = p2 . This means that
the operators ϕ(T ) act by
Z
1
ϕ(T )f (x) = d/2
ϕ(p2 )fb(p)eipx dx.
(2π) Rd

For example,
√
Z p
1
−∆ + 1f (x) = d/2
1 + p2 fb(p)eipx dx
(2π) Rd
√
and one can show that D( −∆ + 1) = H 1 (Rd ).

48
Example 5.21. Another classical example is provided by the Fourier series. Take
H = `2 (Zd ) and let a function t : Zd → R satisfy t(−m) = t(m) and t(m) ≤
c1 e−c2 |m| with some c1 , c2 > 0. Define T by
X
T u(m) = t(m − n)u(n).
n∈Zd

One can easily see that T is bounded. If one introduces the unitary map Φ : H →
L2 (Td ), T := R/Z,
X
Φu(x) = e2πimx u(m), mx := m1 x1 + · · · + md xd ,
m∈Zd

then T = Φ∗ Mτ Φ with X
τ (x) = t(m)e2πimx .
m∈Zd

Example 5.22. A less obvious example is given by the Neumann Laplacian TN on

the half-line defined in Example 2.11.
Let T be the free Laplacian in L2 (R). Denote by G := L2p (R) the subspace of L2 (R)
consisting of the even functions. Clearly, G is an invariant subspace for T (the second
derivate of an even function is also an even function), and the restriction of T to G
is a self-adjoint operator; denote this restriction by A. Moreover, G is an invariant
subspace of the Fourier transform F (the Fourier image of an even function is also an
−1/2


even function). Introduce now the a map Φ : L2 (R+ ) → G by Φf
(x) = 2 f |x| .
One checks easily that Φ is unitary and that D(A) = Φ D(TN ) .
So we have TN = Φ∗ AΦ and A = F ∗ M fh F, where M fh is the multiplication by the
2 ∗
function h(p) = p in G. Finally, M fh = ΦMh Φ , where Mh is the multiplication by
2
h in L (R+ ).
At the end of the day we have TN = U ∗ Mh U with U = Φ∗ FΦ, and U is unitary
being a composition of three unitary operators. By direct calculation, for f ∈
L2 (R+ ) ∩ L1 (R+ ) one has
r Z ∞
2
U f (p) = cos(px)f (x) dx.
π 0
This transform U is sometimes called the cos-Fourier transform. Roughly speaking,
U is just the Fourier transform restricted to the even functions together with some
identifications.
An interested reader may adapt the preceding constructions to the Dirichlet Lapla-
cian TD on the half-line, see Example 2.12.

Example 5.23 (Operators with compact resolvents). Let us fill the gap which
was left open in Subsection 4.3. Namely let us show that if a self-adjoint T has a
compact resolvent, then spec T 6= R.

49
Assume that spec T = R and consider the function g given by g(x) = (x − i)−1 .
Then g(T ) = (T − i)−1 is a compact operator, and its spectrum has at most one
accumulation point. On the other hand, using Corollary 5.17 and the continuity of
g one has the equality spec g(T ) = g(spec T ) = g(R), and this set has no isolated
points.

50
6 Some applications of spectral theorem
In this chapter we discuss some direct applications of the spectral theorem to the
estimates of the spectra of self-adjoint operators. We still use without special noti-
fication the measure µ and the function h from Theorem 5.13.
Theorem 6.1 (Distance to spectrum). Let T be a self-adjoint operator in a
Hilbert space H, and 0 6= x ∈ D(T ), then for any λ ∈ C one has the estimate
(T − λ)x
dist(λ, spec T ) ≤ .
kxk
Proof. If λ ∈ spec T , then the left-hand side is zero, and the inequality is valid.
Assume now that λ ∈/ spec T . By Corollary 5.17, one has, with ρ(x) = (x − λ)−1 ,
1
k(T − λ)−1 k = essµ sup |ρ ◦ h| = ,
dist(λ, spec T )
which gives
1
kxk = k(T − λ)−1 (T − λ)xk ≤ k(T − λ)xk.
dist(λ, spec T )
Remark 6.2. The previous theorem is one of the basic tools for the constructing
approximations of the spectrum of the self-adjoint operators. It is important to
understand that the resolvent estimate obtained in Theorem 6.1 uses in an essen-
tial way the self-adjointness of the operator T . For non-self-adjoint operators the
estimate fails even in the finite-dimensional case. For example, take H = C2 and
 
0 1
T = ,
0 0
then spec T = {0}, and for z 6= 0 we have
 
−1 1 z 1
(T − z) =− 2 .
z 0 z
For the vectors e1 = (1, 0) and e2 = (0, 1) one has he1 , (T − z)−1 e2 i = −z −2 , which
shows that the norm of the resolvent near z = 0 is of order z −2 . In the infinite
dimensional-case one can construct examples with k(T − z)−1 k ∼ dist(z, spec T )−n
for any power n.

6.1 Spectral projections

Definition 6.3 (Spectral projection). Let T be a self-adjoint operator in a
Hilbert space H and Ω ⊂ R be a Borel subset. The spectral projection of T on
Ω is the operator ET (Ω) := 1Ω (T ), where 1Ω is the characteristic function of Ω.
This exchange between the index and the argument is due to the fact that the
spectral projections are usually considered as functions of subsets Ω (with a fixed
operator T ).

51
The following proposition summarizes the most important properties of the spectral
projections.
Proposition 6.4. For any self-adjoint operator T acting a in a Hilbert space there
holds:

1. for any Borel subset Ω ⊂ R the associated spectral projection ET (Ω) is an

orthogonal projection commuting with T . In particular, ET (Ω)D(T ) ⊂ D(T ).


2. ET (a, b) = 0 if and only if spec T ∩ (a, b) = ∅.

3. for any λ ∈ R there holds ran ET ({λ}) = ker(T − λ), and f ∈ ker(T − λ) iff
f = ET ({λ})f .


4. spec T = {λ ∈ R : ET (λ − ε, λ + ε) 6= 0 for all ε > 0}.

Proof. To prove (1) we remark that 12Ω = 1Ω and 1Ω = 1Ω , which gives

ET (Ω)ET (Ω) = ET (Ω) and ET (Ω) = ET (Ω)∗ and shows that ET (Ω) is an orthog-
onal projection. To prove the commuting with T we restrict ourselves by consid-
ering T realized as a multiplication operator from Theorem 5.8. Let x ∈ D(T ),
then hx ∈ L2 (S, ×N, µ) and, subsequently, h · 1Ω ◦ h · x ∈ L2 , which means that
1Ω x ∈ D(T ). The commuting follows now from h · 1Ω ◦ h · x = 1Ω ◦ h · h · x.


To prove (2) we note that the condition ET (a, b) = 0 is, by definition, equivalent
to 1(a,b) ◦h = 0 µ-e.a., which in turn means that (a, b)∩essµ ran h = ∅, and it remains
to recall that essµ ran h = spec T , see Corollary 5.17.
The items (3) and (4) are left as elementary exercises.

As an important corollary of the assertion (4) one has the following description of
the spectra of self-adjoint operators, whose proof is another simple exercise.
Corollary 6.5. Let T be self-adjoint, then λ ∈ spec T if and only if there exists a
sequence xn ∈ D(T ) with kxn k ≥ 1 such that (T − λ)xn converge to 0.

One can see from Proposition 6.4 that the spectral projections contains a lot of
useful information about the spectrum. Therefore, it is a good idea to understand
how to calculate them at least for simple sets Ω.
Proposition 6.6 (Spectral projection to a point). For any λ ∈ R there holds

ET ({λ}) = −i s − lim ε(T − λ − iε)−1 .

ε→0+

Proof. For ε > 0 consider the function

iε
fε (x) := − .
x − λ − iε
One has the following properties:

• |fε | ≤ 1,

52
 • fε (λ) = 1,
• if x 6= λ, then fε (x) tends to 0 as ε tends to 0.
B∞
This means that fε −−→ 1{λ} . By Theorem 5.16, ET ({λ}) = s − limε→0+ fε (T ), and
it remains to note that fε (T ) = (T − λ − iε)−1 by Theorem 5.8.
Proposition 6.7 (Stone formula). For a < b one has:
Z b
1

 1
ET (a, b) + ET [a, b] = s − lim =R(λ + iε) dλ.
2 π ε→0+ a
Proof. For ε > 0 consider the function
1 b
Z
1
fε (x) = = dλ.
π a x − λ − iε
By direct computation we have
1 b
 
b−x a−x
Z
ε 1
fε (x) = dλ = arctan − arctan .
π a (λ − x)2 + ε2 π ε ε
Therefore, |fε | ≤ 1, and


 0, x ∈/ [a, b],
 1 
lim fε (x) = 1, x ∈ (a, b), = 1(a,b) (x) + 1[a,b] (x) ,
ε→0+ 2
 1 , x ∈ {a, b},


2
and the rest follows as in the previous proposition.

Finally the following formula can be useful for the computation of spectral projec-
tions on isolated components of the spectrum.
Proposition 6.8 (Spectral projection on isolated part of spectrum). Let
Γ ⊂ C be a smooth closed curve oriented in the anti-clockwise sense which does not
meet spec T , and let Ω be the intersection of the interior of Γ with R, then
I
1
ET (Ω) = (T − z)−1 dz.
2πi Γ
Proof. If x is an intersection point of Γ with R, then, by assumption x ∈
/ essµ ran h.
On the other hand, for x ∈ R \ Γ there holds, using the Cauchy formula,
I (
1 1, x is inside Γ,
(x − z)−1 dz =
2πi Γ 0, x is outside Γ.
Therefore, µ-a.e. one has
I
1
(h − z)−1 dz = 1Ω ◦ h,
2πi Γ

and one can replace h by T using Theorem 5.16.

53
As a final remark we mention that the map Ω 7→ ET (Ω) can be viewed an operator-
valued measure, and one can integrate reasonable scalar function (bounded Borel
ones or even unbounded) with respect to this measure using e.g. the Lebesgue
integral sums. Then one obtains the integral representations,
Z Z
T = λdET (λ), f (T ) = f (λ)dET (λ),
R R

and the associated integral sums can be viewed as certain approximations of the
respective operators.

6.2 Generalized eigenfunctions

Let T be a self-adjoint operator in a Hilbert space H. Let H+ be another Hilbert
space which is continuously and densely embedded in H, i.e. that there exists a
continuous linear operator j : H+ → H whose range is dense. Denote by H− the
space of continuous linear functionals on H+ . The action of h− ∈ H+ on h+ ∈ H+
will be denoted by hh− , h+ i. The space H− has the natural vector structure. There
is a natural linear map j ∗ : H → H− which assign to each h ∈ H the functional j ∗ h
given by hj ∗ h, h+ i = hh, jh+ i. If one introduces the norm in H− by

kh− k− = sup hh− , h+ i ,

kh+ k=1

then j ∗ ∈ L(H, H− ) and, moreover, ker j ∗ = {0}, i.e. j ∗ is an embedding. The triple
H+ ⊂ H ⊂ H− is usually referred to as a Gelfand triple or rigging of H.

Definition 6.9 (Generalized eigenfunction). We say that a vector ψ ∈ H− is

a generalized eigenfunction of T with the generalized eigenvalue λ ∈ R if ψ, (T −
λ)ϕ = 0 for all ϕ ∈ D(T ) ∩ H+ with T ϕ ∈ H+ .

One has the following fundamental result.

Theorem 6.10 (Existence of expansion in generalized eigenfunctions). One

can find a rigging such that:

• the set D := {ϕ ∈ D(T ) ∩ H+ : T ϕ ∈ H+ } is a core of T , e.g. T |D = T ,

• there exists a measure space (M, µ) and a map Φ : M → H− with the following
properties:

– the map H+ 3 h 7→ b h ∈ L2 (M, dµ) defined by b

h(m) := hΦ(m), hi extends
2
to a unitary operator from H to L (M, dµ),
– there exists a measurable function a : M → R such that Φ(m) is a gener-
alized eigenfunction of T with the generalized eigenvalue a(m) for µ-a.e.
m ∈ M.

54
We are not giving any proof here, an interested reader may refer to a good concise
discussion in Supplement S1.2 of the book [2] or to the detailed study in the book
[1].
Example 6.11 (Generalized eigenfunctions of Laplacian). Let H = L2 (Rn )
and T = −∆. One can take H+ = H 2 (Rn ), then H− = H −2 (Rn ). One can easily
show that for any p ∈ R3 the function ψ, ψ(x) = eipx is a generalized eigenfunction
of T with the generalized eigenvalue p2 . The associated map h 7→ b h is the usual
n
Fourier transform, and (M, dµ) is just R with the Lebesgue measure.
Theorem 6.10 just gives a special form of a unitary transform U from Theorem
5.13. Informally speaking, the theorem says that calculating the spectrum is in a
sense equivalent to solving the eigenvalue problem T ψ = λψ but in a certain larger
space H− . On the main difficulties in applying such an approach is that in concrete
examples the spaces H± are described in a rather implicit way, and it difficult to
decide if a given vector/distribution belongs to this space or not. Some particular
cases are indeed well-studied. For example, one has the following nice description of
the spectrum for Schrödinger operators, which we state without proof (see e.g. [4],
Chapter 2):
Theorem 6.12 (Shnol theorem). Let H = L2 (Rn ), V ∈ L2loc , V ≥ 0, T = −∆+V
(we take the operator defined by the Fridrichs extension). Denote by Σ the set of
the real numbers λ for which there exists a non-zero solution u to the differential
equation (−∆ + V )u = λu with the subexponential growth, i.e. such that for any
a > 0 there exists C > 0 such that u(x) ≤ Cea|x| for all x ∈ Rn . Then the spectrum
of T coincides with the closure of Σ.
Note that there are various versions of the above result for differential operators
on manifolds and other related spaces, then the subexponentional growth condi-
tion should be replaced by a suitable relation comparing the growth of generalized
eigenfunctions with the growth of the volume of balls at infinity.
Example 6.13. One can look again at the operator T = −d2 /dx2 in H = L2 (R).
for any λ ∈ R the equation −u00 = λu has two linearly independent solutions. For
λ < 0 all non-zero solutions are exponentially growing for x → +∞ or for x → −∞,
and such values λ do not belong to the spectrum. For λ = 0 one has either a
constant or a linear function, and for λ > 0 the both solutions are bounded, which
gives again spec T = [0, +∞).

6.3 Tensor products

A more detailed discussion of tensor products can be found e.g. in [12, Sections II.4
and VIII.10] or in [17, Sections 1.4 and 4.5].
Let Aj be self-adjoint operators in Hilbert spaces Hj , j = 1, . . . , n. With any
monomial λm mn m1
1 · . . . λn , mj ∈ N, one can associate the operator A1 ⊗ ·An
1 mn
in
H := H1 ⊗ · · · ⊗ Hn defined by
m
(Am mn m1 mn
1 ⊗ · · · ⊗ An )(ψ1 ⊗ · · · ⊗ ψn ) = A1 ψ1 ⊗ · · · ⊗ An ψn ,
1
ψj ∈ D(Aj j ),

55
and then extended by linearity; here the zero powers A0j equal the identity operators
in the respective spaces.
Remark 6.14. For an operator A in a Hilbert space H the domain D(An ) is usually
defined in a recursive way:

D(A0 ) = H and D(An ) = x ∈ D(A) : Ax ∈ D(An−1 ) for n ∈ N.



As an exercise one can show that for a self-adjoint A one has D(An ) = ran RA (z)n
with any z ∈ res A and that D(An ) is dense in H for any n.

Using the above construction one can associate with any real-valued polynomial P
of λ1 , . . . , λn of degree N a linear operator P (A1 , . . . , An ) in H defined on the set H
consisting of the linear combinations of the vectors of the form ψ1 ⊗ · · · ⊗ ψn with
ψj ∈ D(AN j ).

Theorem 6.15 (Spectrum of tensor product). Denote by B the closure of the

above operator P (A1 , . . . , An ), then B is self-adjoint, and

spec B = P (λ1 , . . . , λn ) : λj ∈ spec Aj .

Sketch of the proof. The complete proof involves a number of technicalities, see e.g.
Section III.10 in [12], but the main idea is rather simple. By the spectral theorem, it
is sufficient to consider the case when Aj is the multiplication by a certain function
fj in Hj := L2 (Mj , dµj ). Then

H = L2 (M, dµ), M = M1 × · · · × Mn , µ = µ1 ⊗ · · · ⊗ µn ,

and P (A1 , . . . , An )
acts in H as the multiplication by p, p(x1 , . . . , xn ) =

P f1 (x1 ), . . . , fn (xn ) , and its domain includes at least all the linear combinations
of the functions ψ1 ⊗ · · · ⊗ ψn where ψj are L2 with compact supports. It is a routine
to show that the closure of this operator is just the usual multiplication operator by
p, which gives the sought relation.
Example 6.16 (Laplacian in rectangle). A typical example of the above con-
struction is given by the Laplacians in rectangles. Namely, let a, b > 0 and
Ω = (0, a)×(0, b) ⊂ R2 , H = L2 (Ω), and T be the Dirichlet Laplacian in Ω. One can
show that T can be obtained using the above procedure using the representation

T = La ⊗ 1 + 1 ⊗ Lb ,

where by La we denote the Dirichlet Laplacian in Ha := L2 (0, a), i.e.

La f = −f 00 , D(La ) = H 2 (0, a) ∩ H01 (0, a).

It is known (from the exercises) that the spectrum of La consists of the simple
eigenvalues (πn/a)2 , n ∈ N, with the eigenfunctions x 7→ sin(πnx/a), and this
means that the spectrum of T consists of the eigenvalues
 πm 2  πn 2
λm,n (a, b) = + , m, n ∈ N,
a b
56
and the associated eigenfunctions are the products of the respective eigenfunctions
for La and Lb . The multiplicity of each eigenvalue λ is exactly the number of pairs
(m, n) ∈ N2 for which λ = λm,n .
Note that the closure of the set {λm,n } can be omitted as this is a discrete set.
The same constructions hold for the Neumann Laplacians, one obtains the same
formula for the eigenvalues but now with m, n ∈ N ∪ {0}.

57
7 Perturbations
7.1 Kato-Rellich theorem
We have seen since the beginning of the course that one needs to pay a great attention
to the domains when dealing with unbounded operators. The aim of the present
subsection is to describe some classes of operators in which such problems can be
avoided.

Definition 7.1 (Essentially self-adjoint operator). We say that a linear op-

erator T is essentially self-adjoint on a subspace D ⊂ D(T ) if the closure of the
restriction of T to D is a self-adjoint operator. If the closure of T is self-adjoint,
then we simply say that T is essentially self-adjoint.

Proposition 7.2. An essentially self-adjoint operator has a unique self-adjoint ex-

tension.

Proof. Let T be an essentially self-adjoint operator, and let S be a self-adjoint

extension of T . As S is closed, the inclusion T ⊂ S implies T ⊂ S. On the other
hand, S = S ∗ ⊂ (T )∗ = T (as T is self-adjoint). This shows that S = T .

Theorem 7.3 (Self-adjointness criterion). Let T be a closed symmetric operator

in a Hilbert space H, then the following three assertions are equivalent:

1. T is self-adjoint,

2. ker(T ∗ + i) = ker(T ∗ − i) = {0},

3. ran(T + i) = ran(T − i) = H.

Proof. The implication 1 ⇒ 2 is obvious: a self-adjoint operator cannot have non-

real eigenvalues.
To show the implication 2 ⇒ 3 recall first that ker(T ∗ ± i) = ran(T ∓ i)⊥ . Therefore,
it is sufficient to show that the subspaces ran(T ± i) are closed. For any f ∈ D(T )
we have:
2

(T ± i)f = (T ± i)f, (T ± i)f = hT f, T f i + hf, f i ± i hT f, f i − hf, T f i
= kT f k2 + kf k2 .

Let fn ∈ ran(T ± i) such that fn converge to some f ∈ H. Find ϕn ∈ D(T ) with

fn = (T ± i)ϕn , then due to the preceding equality (ϕn ) and (T ϕn ) are Cauchy
sequences. As T is closed, ϕn converge to some ϕ ∈ D(T ) and T ϕn converge to T ϕ,
and then fn = (T ± i)ϕn converge to (T ± i)ϕ = f and f ∈ ran(T ± i).
It remains to the prove the implication 3 ⇒ 1. Let ϕ ∈ D(T ∗ ). Due to the
surjectivity of T − i one can find ψ ∈ D(T ) with (T − i)ψ = (T ∗ − i)ϕ. As T ⊂ T ∗ ,
we have (T ∗ − i)(ψ − ϕ) = 0. On the other hand, due to ran(T + i) = H we have
ker(T ∗ − i) = 0, which means that ϕ = ψ ∈ D(T ).

58
Note that during the proof we obtained the following simple fact:
Proposition 7.4. Let T be a symmetric operator, then ran(T ± i) = ran(T ± i).

This leads as to the following assertion:

Corollary 7.5 (Essential self-adjointness criterion). Let T be a symmetric
operator in a Hilbert space H, then the following three assertions are equivalent:

1. T is essentially self-adjoint,
2. ker(T ∗ + i) = ker(T ∗ − i) = {0},
3. ran(T + i) and ran(T − i) are dense in H.
Remark 7.6. The above theorem can be modified in several ways. For example, it
still holds if one replaces T ± i by T ± iλ with any λ ∈ R \ {0}. For semibounded
operators we have an alternative version:
Theorem 7.7 (Self-adjointness criterion for semibounded operators). Let
T be a closed symmetric operator in a Hilbert space H and T ≥ 0 and let a > 0,
then the following three assertions are equivalent.

1. T is self-adjoint,
2. ker(T ∗ + a) = {0},
3. ran(T + a) = H.

This is left as an exercise. The analogues of Proposition 7.4 and Corollary 7.5 hold
as well.

Now we would like to apply the above assertions to the study of some perturbations
of self-adjoint operators.
Definition 7.8 (Relative boundedness). Let A be a self-adjoint operator in a
Hilbert space H and B be a linear operator with D(A) ⊂ D(B). Assume that there
exist a, b > 0 such that

kBf k ≤ akAf k + bkf k for all f ∈ D(A),

then B is called relatively bounded with respect to A or, for short, A-bounded. The
infimum of all possible values a is called the relative bound of B with respect to A.
If the relative bound is equal to 0, then B is called infinitesimally small with respect
to A.
Theorem 7.9 (Kato-Rellich). Let A be a self-adjoint operator in H and let B be
a symmetric operator in H which is A-bounded with a relative bound < 1, then the
operator A + B with the domain D(A + B) = D(A) is self-adjoint. Moreover, if A
is essentially self-adjoint on some D ⊂ D(A), then A + B is essentially self-adjoint
on D too.

59
Proof. By assumption, one can find a ∈ (0, 1) and b > 0 such that

kBuk ≤ akAuk + bkuk, for all u ∈ D(A). (7.1)

Step 1. As seen many times, for any λ > 0 one has

2 2
(A + B ± iλ)u = (A + B)u + λ2 kuk2 .

Therefore, for all u ∈ D(A) one can estimate

2 (A + B ± iλ)u ≥ (A + B)u + λkuk ≥ kAuk − kBuk + λkuk

= (1 − a)kAuk + (λ − b)kuk. (7.2)

Let us pick some λ > b.

Step 2. Let us show that A + B with the domain equal to D(A) is a closed operator.
Let (un ) ⊂ D(A) and fn := (A + B)un such that both un and fn converge in H. By
(7.2), Aun is a Cauchy sequence. As A is closed, un converge to som u ∈ D(A) and
Aun converge to Au. By (7.1), Bun is a Cauchy sequence and is hence convergent
to some v ∈ H. Let us show that Bun converge exactly to un (actually this would
follow from the closedness of B, but we did not assume that B is closed or closable!).
Take any h ∈ D(A), then hv, hi = limhBun , hi = limhun , Bhi = hu, Bhi = hBu, hi.
So finally (A + B)un converge to (A + B)u. This shows that A + B is closed.
Step 3. Let us show that the operators A + B ± iλ : D(A) → H are bijective at
least for large λ. As previously, we have k(A ± iλ)uk2 = kAuk2 + λ2 kuk2 . Then
b b

kBuk ≤ akAuk + bkuk ≤ a (A ± iλ)u + (A ± iλ)u = a + (A ± iλ)u .
|λ| |λ|
As a ∈ (0, 1), we can choose λ sufficiently large to have a + b/|λ| < 1. This means
that for such λ we have B(A ± iλ)−1 < 1. Now we can represent
 
−1
A + B ± iλ = 1 + B(A ± iλ) (A ± iλ).

As A is self-adjoint, the operators A ± iλ : D(A) → H are bijections, and 1 + B(A ±

iλ)−1 is a bijection from H to itself. Therefore, A+B ±iλ are bijective, in particular,
ran(A + B ± iλ) = H. By Theorem 7.3 and Remark 7.6, A + B is self-adjoint.
The part concerning the essential self-adjointness is reduced to the proof of the
relation A + B = A + B, which is an elementary exercise.

7.2 Essential self-adjointness of Schrödinger operators

The Kato-Rellich theorem is one of the tools used to simplify the consideration of
the Schrödinger operators.
Theorem 7.10. Let V ∈ Lp (Rd ) + L∞ (Rd ) be real-valued with p = 2 if d ≤ 3 and
p > d/2 if d > 3, then the operator T = −∆ + V with the domain D(T =)H 2 (Rd )
is a self-adjoint operator in L2 (Rd ), and it is essentially self-adjoint on Cc∞ (Rd ).

60
Proof. We give the proof only for the dimension d ≤ 3. For all f ∈ S(Rd ) and
λ > 0 we have the representation
Z
1
f (x) = eipx fb(p) dp
(2π)d/2 Rd
Z
1 1
= d/2 2
(p2 + λ)fb(p)dp
(2π) Rd p + λ
1 1
≤ d/2 2
· (p2 + λ)fb(p)
(2π) p +λ
1 1 
2b

≤ · kp f (p)k + λk bk = aλ k∆f k + bλ kf k
f
(2π)d/2 p2 + λ
with
1 1 λ 1
aλ = , bλ = .
(2π)d/2 p2 + λ (2π)d/2 p2 + λ
By density, for all f ∈ H 2 (Rd ) and all λ > 0 we have

kf k∞ ≤ aλ k∆f k + bλ kf k.

By assumption we can represent V = V1 + V2 with V1 ∈ L2 (Rd ) and V2 ∈ L∞ (Rd ).

Using the preceding estimate we arrive at

kV f k ≤ kV1 f k + kV2 f k ≤ kV1 k2 kf k∞ + kV2 kkf k ≤ e

aλ k∆f k + ebλ kf k, f ∈ H 2 (Rd ),

aλ = kV1 k2 aλ and ebλ = kV1 k2 bλ + kV2 k∞ . As aλ can be made arbitrary small

with e
by a suitable choice of λ, we see that the multiplication operator V is infinitesimally
small with respect to the free Laplacian, and the result follows from the Kato-Rellich
theorem.
The above proof does not work for d ≥ 3 as the function p 7→ (p2 + λ)−1 does not
belong to L2 (Rd ) anymore. The respective parts of argument should be replaced by
suitable Sobolev embedding theorems.
Example 7.11 (Coulomb potential). Consider the three-dimensional case and
the potential V (x) = α/|x|, α ∈ R. For any bounded open set Ω containing the
origin, one has 1Ω V ∈ L2 (R3 ) and (1 − 1Ω )V ∈ L∞ (R3 ), and finally V ∈ L2 (R3 ) +
L∞ (R3 ). This means that the operator T = −∆+α/|x| is self-adjoint on the domain
H 2 (Rd ).

Let us mention some other conditions guaranteeing the essential self-adjointness of

the Schrödinger operators for other types of potentials.
Theorem 7.12. Let H = L2 (Rd ) and let V ∈ C 0 (Rd ) be real-valued such that for
some c ∈ R one has the inequality

hu, (−∆ + V )ui ≥ ckuk2

for all u ∈ Cc∞ (Rd ). Then the operator T = −∆ + V is essentially self-adjoint on

Cc∞ (Rd ).

61
Proof. By adding a constant to the potential V one can assume that T ≥ 1. In
other words, using the integration by parts,
Z Z Z
2 2 2
∇u(x) dx + V (x) u(x) dx ≥ u(x) dx (7.3)
Rd Rd Rd

for all u ∈ Cc∞ (Rd ), and this extends by density at least to all u ∈ Hcomp
1
(Rd ). By
Theorem 7.7 it is sufficient to show that the range of T is dense.
Let f ∈ L2 (Rd ) such that f, (−∆ + V )u = 0 for all u ∈ Cc∞ (Rd ). Note that T
preserve the real-valuedness, and we can suppose without loss of generality that f is
real-valued. We have at least (−∆ + V )f = 0 in the sense of D0 (Rd ), and ∆f = V f .
As V is locally bounded, the function V f is in L2loc (Rd ), and the elliptic regularity
2
gives f ∈ Hloc (Rd ).
Let us pick a real-valued function ϕ ∈ Cc∞ (Rd ) such that ϕ(x) = 1 for |x| ≤ 1,
that ϕ(x) = 0 for |x| ≥ 2 and that 0 ≤ ϕ ≤ 1, and introduce functions ϕn by
1
ϕn (x) = ϕ(x/n). For any u ∈ Hloc (Rd ) we have, by a standard computation:
Z Z
∇(ϕn f )∇(ϕn u)dx + V ϕn f ϕn udx
Rd Rd
Z d Z  ∂u
2
X ∂f  ∂ϕn
= ∇ϕn f u dx + f −u ϕn dx + hf, T ϕ2n ui. (7.4)
Rd j=1 Rd ∂x j ∂x j ∂x j

As ϕ2n u ∈ Cc∞ (Rd ), the last term vanishes. Taking now u = f and using (7.3) we
arrive at Z Z Z
2 2 2 2
∇ϕn f dx ≥ ϕn f dx ≥ ϕ2n f 2 dx,
Rd Rd Ω
where Ω is any ball. As n tends to infinity, the left-hand side goes to 0. On the
other side, the restriction of ϕn f to Ω coincides with f for sufficiently large n, and
this means that f vanishes in Ω. As Ω is arbitrary, f = 0.

Another condition, which complements the preceding theorems, is given without

proof (as it needs some advanced PDE machinery).
Theorem 7.13. Let V ∈ Lloc (Rd ) be non-negative, then the operator −∆ + V is
essentially self-adjoint on Cc∞ (Rd ).

7.3 Discrete and essential spectra

Up to now we just distinguished between the whole spectrum and the point spec-
trum, i.e. the set of the eigenvalues. Let us introduce another classification of
spectra, which is useful when studying various perturbations.
Definition 7.14 (Discrete spectrum, essential spectrum). Let T be a self-
adjoint operator in a Hilbert space H. We define its discrete spectrum specdisc T
by
n
o
specdisc T := λ ∈ spec T : ∃ε > 0 with dim ran ET (λ − ε, λ + ε) < ∞ .

62
The set specess T := spec T \ specdisc T is called the essential spectrum of T .

The following proposition gives an alternative description of the discrete spectrum.

Proposition 7.15. A real λ belongs to specdisc T iff it is an isolated eigenvalue of

T of finite multiplicity.

Proof.
Let λ ∈ specdisc T , then there exists ε0 > 0 such that the operators ET (λ −
ε, λ+ε) do not depend on ε if ε ∈ (0, ε0 ). On the
the other hand, this limit operator


is non-zero, as λ ∈ spec T . This means ET {λ} = s − limε→0+ ET (λ − ε, λ + ε)
6=
0, and λ ∈ spec
p T by Proposition 6.4(3). At the same time, ET (λ − ε0 , λ) =
ET (λ, λ + ε0 ) = 0, and Proposition 6.4(2) show that λ is an isolated point of the
spectrum.
Now let λ be an isolated
eigenvalue of finite

multiplicity. Then there exists ε0 > 0
such that ET (λ−ε0 , λ) = ET (λ, λ+ε0 ) = 0, and dim ran ET ({λ}) = dim ker(T −
λ) < ∞. Therefore,



dim ran ET (λ − ε0 , λ + ε0 ) = dim ran ET (λ − ε0 , λ)


+ dim ran ET (λ, λ + ε0 ) + dim ran ET ({λ}) < ∞.

Therefore, we arrive at the following direct description of the essential spectrum

Proposition 7.16. A value λ ∈ spec T belongs to specess T iff at least one of the
following three conditions holds:

• λ∈
/ specp T ,

• λ is an accumulation point of specp T ,

• dim ker(T − λ) = ∞.

Furthermore, the essential spectrum is a closed set.

Proof. The first part just describes the points of the spectrum which are not isolated
eigenvalues of finite multiplicity.
For the second part we note that specess T is obtained from the closed set spec T by
removing some isolated points. As the removing an isolated point does not change
the property to be closed, specess T is also closed.

Let us list some examples.

Proposition 7.17 (Essential spectrum for compact operators). Let T be a

compact self-adjoint operator in an infinite-dimensional space H, then specess T =
{0}.

63
Proof. By Theorem 4.3, for any ε > 0 the set spec T \ (−ε, ε) consists of a finite

multiplicity, hence we have: specess T \ (−ε, ε) = ∅
number of eigenvalues of finite
and dim
ran ET R \ (−ε, ε) <
∞. On the other hand,
dim H = dim ran ET R \
(−ε, ε) + dim ran ET (−ε, ε) , and dim ran ET (−ε, ε) must be infinite for any
ε > 0, which means that 0 ∈ specess T .
Proposition 7.18 (Essential spectrum of operators with compact resol-
vents). The essential spectrum of a self-adjoint operator is empty if and only if the
operator has a compact resolvent.

Proof is left as an exercise.

Sometimes one uses the following terminology:
Definition 7.19 (Purely discrete spectrum). We say that a self-adjoint operator
T has a purely discrete spectrum in some interval (a, b) if specess T ∩ (a, b) = ∅. If
one has simply specess T = ∅, then we say simply that the spectrum of T is purely
discrete. As follows from the previous proposition, this exactly means that T has a
compact resolvent.
Example 7.20. As seen several times, the free Laplacian in L2 (Rd ) has the spectrum
[0, +∞). This set has no isolated points, so this operator has no discrete spectrum.

The main difference between the discrete and the essential spectra comes from their
behavior with respect to perturbations. This will be discussed in the following
sections.

7.4 Weyl criterion and relatively compact perturbations

Let T be a self-adjoint operator in a Hilbert space H.
The following proposition is an exercise.
Proposition 7.21. Let λ be an isolated eigenvalue of T , then there exists c > 0
such that k(T − λ)uk ≥ ckuk for all u ⊥ ker(T − λ).

The following theorem gives a description of the essential spectrum using approxi-
mating sequences.
Theorem 7.22 (Weyl criterion). The condition λ ∈ specess T is equivalent to the
existence of a sequence (un ) ⊂ D(T ) satisfying the following three properties:

1. kun k ≥ 1,

2. un converge weakly to 0,

3. (T − λ)un converge to 0 in the norm of H.

Such a sequence will be called a singular Weyl sequence for λ. Moreover, as will be
shown in the proof, one can replace the conditions (1) and (2) just by:

64
 1’. un form an orthonormal sequence.

Proof. Denote by W (T ) the set of all real numbers λ for which one can find a
singular Weyl sequence.
Show first the inclusion W (T ) ⊂ specess T . Let λ ∈ W (T ) and let (un ) be an
associated singular Weyl sequence, then we have at least λ ∈ spec T . Assume by
contradiction that λ ∈ specdisc T and denote by Π the orthogonal projection to
ker(T − λ) in H. As Π is a finite-rank operator, it is compact, and the sequence
Πun converge to 0. Therefore, the norms of the vectors wn := (1 − Π)un satisfy
kwn k ≥ 1/2 for large n. On the other hand, the vectors (T −λ)wn = (1−Π)(T −λ)un
converge to 0, which contradicts to Proposition 7.21.


Conversely, if λ ∈ specess T , then dim ran ET (λ − ε, λ + ε) = ∞ for all ε > 0.
In particular, one can find a strictly decreasing to 0 sequence (εn ) with ET (In \
In+1 ) 6= 0, where In := (λ − εn , λ + εn ). Now we can choose un with kun k = 1
and ET (In \ In+1 )un = un . These vectors form an orthonormal sequence and, in
particular, converge weakly to 0. On the other hand,

k(T − λ)un k = k(T − λ)ET (In \ In+1 )un k ≤ εn kun k = εn ,

which shows that the vectors (T − λ)un converge to 0. Therefore, (un ) is a singular
Weyl sequence, and specess T ⊂ W (T ).

The following theorem provides a starting point to the study of perturbations of

self-adjoint operators.

Theorem 7.23 (Stability of essential spectrum). Let A and B be self-adjoint

operators such that for some z ∈ res A ∩ res B the difference of their resolvents
K(z) := (A − z)−1 − (B − z)−1 is a compact operator, then specess A = specess B.

Proof. One can easily see, using the resolvent identities (Proposition 3.4), that
K(z) is compact for all z ∈ res A ∩ res B.
Let λ ∈ specess A and let (un ) be an associated singular Weyl sequence. Without
loss of generality assume that kun k = 1 for all n. We have
 1  1
lim (A − z)−1 − un = lim (A − z)−1 (A − λ)un = 0. (7.5)
λ−z z−λ
On the other hand, as K(z) is compact, the sequence K(z)un converges to 0 with
respect to the norm, and

1  1 
lim (B − λ)(B − z)−1 un = lim (B − z)−1 − un
z−λ λ−z
 1 
= lim (A − z)−1 − un − lim K(z)un = 0.
λ−z
Now denote vn := (B−z)−1 un . The preceding equality shows that (B−λ)vn converge
to 0, and one can easily show that vn converge weakly to 0. It follows again from

65
(7.5) and from the compactness of K(z) that lim kvn k = |λ − z|−1 . Therefore, (vn )
is a singular Weyl sequence for B and λ, and λ ∈ specess B. So we have shown the
inclusion specess A ⊂ specess B. As the participation of A and B is symmetric, we
have also specess A ⊃ specess B.

Let us describe a class of perturbations which can be studied using the preceding
theorem.

Definition 7.24 (Relatively compact operators). Let A be a self-adjoint op-

erator in a Hilbert space H, and let B a closable linear operator in H with
D(A) ⊂ D(B). We say that B is compact with respect to A (or simply A-compact)
if B(A − z)−1 is compact for at least one z ∈ res A. (It follows from the resolvent
identitites that this holds then for all z ∈ res A.

Proposition 7.25. Let B be A-compact, then B is infinitesimally small with respect

to A.

Proof. We show first that

lim B(A − iλ)−1 = 0 (7.6)

λ→+∞

Assume that (7.6) is false. Then one can find a constant α > 0, non-zero vectors un
and a positive sequence λn with lim λn = +∞ such that B(A − iλ)−1 un > αkun k
2
for all n. Set vn := (A−iλ)−1 un . Using kun k2 = (A−iλn )vn = kAvn k2 +λ2n kvn k2
we obtain
kBvn k2 > α2 kAvn k2 + α2 λ2n kvn k2 .
Without loss of generality one may assume the normalization kBvn k = 1, then the
sequence Avn is bounded and vn converge to 0. Let z ∈ res A, then (A − z)vn is
also bounded, one can extract a weakly convergent subsequence (A − z)vnk . Due
to the compactness, the vectors B(A − z)−1 · (A − z)vnk = Bvnk converge to some
w ∈ H with kwk = 1. On the other hand, as shown above, vnk converge to 0, and
the closability of B shows that w = 0. This contradiction shows that (7.6) is true.
Now, for any a > 0 one can find λ > 0 such that kB(A − iλ)−1 uk ≤ akuk for all
u ∈ H. Denoting v := (A − iλ)−1 u and noting that (A − iλ)−1 is a bijection between
H and D(A) we see that

kBvk ≤ ak(A − iλ)vk ≤ akAvk + aλkvk

for all v ∈ D(A). As a > 0 is arbitrary, we get the result.

So a combination of the preceding assertions leads us to the following observation:

Theorem 7.26 (Relatively compact perturbations). Let A be a self-adjoint

operator in a Hilbert space H and let B be symmetric and A-compact, then the
operator A + B with D(A + B) = D(A) is self-adjoint, and the essential spectra of
A and A + B coincide.

66
Proof. The self-adjointness of A + B follows from the Kato-Rellich theorem, and
it remains to show that the difference of the resolvents of A + B and A is compact.
This follows directly from the obvious identity

(A − z)−1 − (A + B − z)−1 = (A + B − z)−1 B(A − z)−1 ,

which holds at least for all z ∈

/ R.

As an exercise one can show the following assertion, which can be useful in some
situations.

Proposition 7.27. Let A be self-adjoint, B be symmetric and A-bounded with a

relative bound < 1, and C be A-compact, then C is also (A + B)-compact.

7.5 Essential spectra for Schrödinger operators

Definition 7.28 (Kato class potential). We say that a measurable function
V : Rd → R belongs to the Kato class if for any ε > 0 one can find real-valued
Vε ∈ Lp (Rd ) and V∞,ε ∈ L∞ (Rd ) such that Vε + V∞,ε = V and kV∞,ε k∞ < ε. Here
p = 2 for d ≤ 3 and p > d/2 for d ≥ 4.

Theorem 7.29. If V is a Kato class potential in Rd , then V is compact with respect

to the free Laplacian T = −∆ in L2 (Rd ), and the essential spectrum of −∆ + V is
equal to [0, ∞).

Proof. We give the proof for d ≤ 3 only. Let F denote the Fourier transform, then
for any f ∈ L2 (Rd ) and z ∈ res T we have

F(T − z)−1 f (p) = (p2 − z)−1 Ff (p).

This means that (T − z)−1 f = gz ? f , where gz is the L2 function with Fgz (p) =
(p2 − z)−1 , and ? stands for the convolution product. In other words,
Z
−1
(T − z) f = gz (x − y)f (y)dy.
Rd

Let ε > 0 and let Vε and V∞,ε be as in Definition 7.28. The operator Vε (T − z)−1 is
an integral one with the integral kernel K(x, y) = Vε (x)gz (x − y), i.e.
Z
−1
Vε (T − z) f (x) = K(x, y)f (y) dy.
Rd

One has
Z Z Z Z
2 2 2
K(x, y) dxdy = Vε (x) dx gz (y) dy
Rd Rd Rd Rd
2 2
= kVε 2 kgz < ∞,

67
which means that Vε (T − z)−1 is a Hilbert-Schmidt operator and, therefore, is com-
pact, see Subsection 4.2. At the same time we have the estimate

kV∞,ε (T − z)−1 k ≤ εk(T − z)−1 k.

Therefore, the operator V (T − z)−1 is compact as it can be represented as the norm

limit of the compact operators Vε (T − z)−1 as ε tends to 0.

Example 7.30 (Coulomb potential). The previous theorem easily applies e.g.
to the operators −∆ + α/|x|. It is sufficient to represent

1 1R (x) 1 − 1R (x)
= + ,
|x| |x| |x|

where 1R is the characteristic function of the ball of radius R > 0 and centered at
the origin with a sufficiently large R. So the essential spectrum of −∆ + α/|x| is
always the same as for the free Laplacian, i.e. [0, +∞).

Another typical application of the Weyl criterion can be illustrated as follows.

Theorem 7.31. Let V ∈ L∞ (Rd ) be real-valued. Assume that there exists α ∈ R

such that the set Ω := x ∈ Rd : V (x) < α has a finite Lebesgue measure, then
−∆ + V has a purely discrete spectrum in (−∞, α).

Proof. Let 1Ω be the characteristic function of Ω. Denote U := (V − α)1Ω and

W := V − U . Then U ∈ Lp (Rd ) for any p ≥ 1 (as U is bounded and supported
by a set of finite measure), in particular, U is of Kato class. At the same time,
W ∈ L∞ (Rd ) and W ≥ α. By Proposition 7.27, U is (−∆ + W )-compact,

specess (−∆ + V ) ∩ (−∞, α) = specess (−∆ + W + U ) ∩ (−∞, α)

= specess (−∆ + W ) ∩ (−∞, α).

On the other hand, specess (−∆ + W ) ⊂ spec(−∆ + W ) ⊂ [α, +∞).

Note that we have no trouble with the domains, as all the operators −∆ + V ,
(−∆ + W ) + U and −∆ + W are defined on the same domain H 2 (Rd ) due to the
boundedness of the potentials.

Remark 7.32. In the physics literature, the situation of Theorem 7.31 is sometimes
referred to as a potential well below α. The same result holds without assumptions
on V outside Ω (i.e. for unbounded potentials), but the proof would then require a
slightly different machinery.

68
8 Variational principle for eigenvalues
8.1 Max-min and min-max principles
Throughout the subsection we denote by T a self-adjoint operator in an infinite-
dimensional Hilbert space H, and we assume that T is semibounded from below. If
specess T = ∅, we denote Σ := +∞, otherwise we put Σ := inf specess T .
Theorem 8.1 (Max-min principle). For n ∈ N introduce the following numbers:
hϕ, T ϕi
µn = µn (T ) = sup inf ,
ψ1 ,...,ψn−1 ∈H ϕ∈D(T ), ϕ6=0 hϕ, ϕi
ϕ⊥ψj , j=1,...,n−1

then we are in one and only one of the following situations:

(a) µn is the nth eigenvalue of T (when ordering all the eigenvalues in the non-
decreasing order according to their multiplicities), and T has a purely discrete
spectrum in (−∞, µn ).
(b) µn = Σ, and µj = µn for all j ≥ n.

Proof. Step 1. Let us prove first two preliminary assertions:

dim ran ET (−∞, a) < n for a < µn , (8.1)


dim ran ET (−∞, a) ≥ n for a > µn . (8.2)


Proof of (8.1). Assume that the assertion is false, then dim ran ET (−∞, a) ≥ n for

some a < µn , and one can find an n−dimensional subspace V ⊂ ran ET (−∞, a) .
As T is semibounded, V ⊂ D(T ). By dimension considerations, for any vectors
ψ1 , . . . , ψn−1 there exists a non-zero vector
ϕ ∈ V orthogonal to all ψj , j = 1, . . . , n−
1, and the inclusion ϕ ∈ ran ET (−∞, a) implies hϕ, T ϕi ≤ ahϕ, ϕi. Therefore, for
any choice of ψj the infimum in the definition of µn is not greater than a, which
gives the inequality µn ≤ a, which contradicts the assumption. Eq. (8.1) is proved.
Proof of (8.2). Again, assume by contradiction
that the assertion is false, then for
some a > µn we have dim ran
ET (−∞, a) ≤ n−1. Let ψ1 , . . .
, ψn−1 be some vectors


spanning ran ET (−∞, a) . Due to the equality ET (−∞, a) + ET [a, +∞)
= Id,
for every ϕ ∈ D(T ) with ϕ ⊥ ψj , j = 1, . . . , n − 1, one has ϕ = ET [a, +∞) ϕ and
hϕ, T ϕi ≥ ahϕ, ϕi, which shows that µn ≥ a. This contradiction proves (8.2).
Step 2. Let us prove that µn < +∞ for any n (note that the equality µn > −∞
follows from the semiboundedness
of T ). Assume that µn = +∞, then, by (8.1),
one has dim ran ET (−∞, a) < n for any a ∈ R, and dim H ≤ n, which contradicts
to the assumption.


Now we have two possibilities:
either dim ran ET (−∞, µn + ε) = ∞ for all ε > 0
or dim ran ET (−∞, µn + ε) < ∞ for some ε > 0. Let us consider them separately.


Step 3. Assume that dim ran ET (−∞, µn + ε) = ∞ for all ε > 0. We are go-
ing to show that the case (b) of the theorem is realized. Due to (8.1), one has

69


dim ran ET (µn − ε, µn + ε) = ∞ for all ε > 0, and µn ∈ specess T . On the other
hand, again by (8.1), specess T ∩ (−∞, µn − ε) = ∅ for all ε > 0, which proves that
µn = Σ. It remains to show that µn+1 = µn . Assume that
µn+1 > µn , then, by (8.1),
for any ε < µn+1 − µn we have dim ran ET (−∞, µn + ε) ≤ n + 1, which contradicts
the assumption. Therefore, µn+1 = µn .


Step 4. Assume now that dim ran ET (−∞, µn + ε) < ∞ for some ε > 0. It
follows directly that the spectrum of T is purely discrete

in (−∞, µn + ε).
More-
over, one can find ε1 >
0 such that ET (−∞, µn ] = ET (−∞, µn
+ ε1 ) . As
dim ran ET (−∞, µn + ε1 ) ≥ n by (8.1), we have dim ran ET (−∞, µn ] ≥ n, which
means that T has at least n eigenvalues λ1 ≤ · · · ≤ λn (counting

with multiplicities)
in (−∞, µn ]. If λn < µn , then dim ran ET − ∞, λn ] ≥ n, which contradicts to
(8.1). This proves the equality µn = λn .
Remark 8.2. Following the convention of Remark 2.19, one may replace the above
definition of the numbers µn by
hϕ, T ψi
µn = sup inf .
ψ1 ,...,ψn−1 ∈H ϕ∈Q(T ), ϕ6=0 hϕ, ϕi
ϕ⊥ψj , j=1,...,n−1

This follows from the fact that D(T ) is dense in Q(T ), see Theorem 2.4 and the
subsequent discussion.
Another elementary observation is given in the following corollary.
Corollary 8.3. If there exists ϕ ∈ Q(T ) with hϕ, T ϕi < Σkϕk2 , then T has at least
one eigenvalue in (−∞, Σ).
Indeed, in this case one has µ1 < Σ, which means that µ1 is an eigenvalue.
By similar considerations one can obtain another variational formula for the eigen-
values, one may refer to Section 4.5 in [5] for its proof:
Theorem 8.4 (Min-max principle). All the assertions of Theorem 8.1 hold with
hϕ, T ϕi hϕ, T ϕi
µn := inf sup = inf sup .
L⊂D(T ) ϕ∈L hϕ, ϕi L⊂Q(T ) ϕ∈L hϕ, ϕi
dim L=n ϕ6=0 dim L=n ϕ6=0

The max-min and min-max principles are powerful tools for the analysis of the
behavior of the eigenvalues with respect to various parameters. As a basic example
we mention the following situation, which will be applied later to some specific
operators:
Definition 8.5. Let A and B be self-adjoint operators in a Hilbert space H, both
semibouneded from below. We write A ≤ B if Q(A) ⊃ Q(B) and hu, Aui ≤ hu, Bui
for all u ∈ Q(B).
As a direct corollary of the max-min principle we obtain:
Corollary 8.6. Let A and B be self-adjoint, and A ≤ B. In addition, assume that A
and B have compact resolvents. If λj (A) and λj (B), j ∈ N, denote their eigenvalues
taken with their multiplicities and enumerated in the non-decreasing order, then
λj (A) ≤ λj (B) for all j ∈ N.

70
8.2 Negative eigenvalues of Schrödinger operators
As seen above in Proposition 7.27, if V is a Kato class potential in Rd , then the
associated Schrödinger operator T = −∆ + V acting in H = L2 (Rd ) has the same
essential spectrum as the free Laplacian, i.e. specess T = [0, +∞) and Σ = 0. In the
present section we would like to discuss the question on the existence of negative
eigenvalues.
We have rather a simple sufficient condition for the one- and two-dimensional cases.

Theorem 8.7. Let d ∈ {1, 2} and V ∈ L∞ (Rd ) ∩ L1 (Rd ) be real-valued such that
Z
V0 := V (x)dx < 0,
Rd

then the associated Schrödinger operator T = −∆ + V has at least one negative

eigenvalue.

Proof. We assumed the boundedness of the potential just to avoid additional tech-
nical issues concerning the domains. It is clear that V ∈ L2 (Rd ), and specess T =
[0, +∞) in virtue of Theorem 7.29. By Corollary 8.3 it is now sufficient to show that
one can find a non-zero ϕ ∈ H 1 (Rd ) with
Z Z
2 2
τ (ϕ) := ∇ϕ(x) dx + V (x) ϕ(x) dx < 0.
Rd Rd

Consider first the case d = 1. Take any ε > 0 and consider the function ϕε given
by ϕε (x) := e−ε|x|/2 . Clearly, ϕε ∈ H 1 (R) for any ε > 0, and the direct computation
shows that
Z Z
0 2 ε 2
ϕε (x) dx = and lim V (x) ϕε (x) dx = V0 < 0.
R 2 ε→0+ Rd

Therefore, for sufficiently small ε one obtains τ (ϕε ) < 0.

ε
Now let d = 2. Take ε > 0 and consider ϕε (x) defined by ϕε (x) = e−|x| /2 . We have

εx|x|ε−2 −|x|ε /2
∇ϕε (x) = − e ,
2
ε2 πε2 ∞ 2ε−1 −rε
Z Z Z
2 2ε−2 −|x|ε
∇ϕε (x) dx = |x| e dx = r e dr
R2 4 R2 2 0
πε ∞ −u
Z
πε
= ue du = ,
2 0 2
and, as previously, Z
2
lim V (x) ϕε (x) dx = V0 < 0,
ε→0+ Rd

and for sufficiently small ε we have again τ (ϕε ) < 0.

71
We see already in the above proof that finding suitable test functions for proving
the existence of eigenvalues may become very tricky and depending on various pa-
rameters. One may easily check that the analog of ϕε for d = 1 does not work for
d = 2 and vice versa. It is a remarkable fact that the analog of Theorem 8.7 does
not hold for the higher dimensions due to the Hardy inequality (Proposition 2.21):

Proposition 8.8. Let d ≥ 3 and let V : Rd → R be bounded with a compact support.

For λ ∈ R consider the Schrödinger operators Tλ := −∆ + λV , then there exists
λ0 > 0 such that spec Tλ = [0, +∞) for all λ ∈ (−λ0 , +∞).

Proof. Due to the compactness of supp V one can find λ0 > 0 in such a way that

(d − 2)2
λ0 V (x) ≤ 2
for all x ∈ Rd .
4|x|

Using the Hardy inequality, for any u ∈ Cc∞ (Rd ) and any λ ∈ (−λ0 , +∞) we have
Z Z
2 2
hu, Tλ ui = ∇u(x) dx + λ V (x) u(x) dx
Rd Z Rd Z
2 2
≥ ∇u(x) dx − λ0 V (x) · u(x) dx
Rd Rd
2
(d − 2)2 u(x)
Z Z
2
≥ ∇u(x) dx − dx ≥ 0.
Rd 4 Rd |x|2

As Tλ is essentially self-adjoint on Cc∞ (Rd ), see Theorem 7.10, this inequality extends
to all u ∈ D(Tλ ), and we obtain Tλ ≥ 0, and this means that spec Tλ ⊂ [0, +∞). On
the other hand, specess Tλ = [0, +∞) as λV is of Kato class (see Theorem 7.29).

72
9 Laplacian eigenvalues for bounded domains
9.1 Dirichlet and Neumann eigenvalues
In this section we discuss some application of the general spectral theory to the
eigenvalues of the Dirichlet and Neumann Laplacians in bounded domains. Let us
recall the setting. Let Ω ⊂ Rd be a bounded open set with a regular boundary
(for example, piecewise smooth and lipschitzian); all the domains appearing in this
section will be supposed to have a regular boundary without further specifications.
Then the embedding of H 1 (Ω) into H := L2 (Ω) is a compact operator. By definition,
the Dirichlet Laplacian TD = −∆D and the Neumann Laplacian TN = −∆N are
the self-adjoint operators in H associated with the sesqulinear forms tD and tN
respectively,
Z
tD (u, v) = ∇u(x) · ∇v(x) dx, D(tD ) = Q(TD ) = H01 (Ω),
ZΩ
tN (u, v) = ∇u(x) · ∇v(x) dx, D(tN ) = Q(TN ) = H 1 (Ω).
Ω

We know that both TD and TN have compact resolvents, and their spectra are purely
D/N D/N
discrete (see Section 4.3). Denote by λj = λj (Ω), j ∈ N, the eigenvalues of
TD/N repeated according to their multiplicities and enumerated in the non-decreasing
order. The eigenvalues are clearly non-negative, and they are usually referred to as
the Dirichlet/Neumann eigenvalues of the domain Ω (the presence of the Laplacian
is assumed implicitly). Let us discuss some basic properties of these eigenvalues.

Proposition 9.1. (a) λN 1 = 0. If Ω is connected, then ker TN is spanned by the

constant function u(x) = 1.

(b) λD
1 > 0.

Proof. (a) Note that u = 1 is clearly an eigenfunction of TN with the eigenvalue 0.

As all the eigenvalues are non-negative, λN1 = 0. Now let u ∈ ker TN , then
Z
2
0 = hu, TN ui = tN (u, u) = ∇u(x) dx,
Ω

which shows that ∇u = 0. Therefore, v is constant on each maximal connected

component of Ω.
(b) We have at least λD D
1 ≥ 0. Assume that λ1 = 0 and let v be an associated
eigenfunction. We have as above ∇v = 0, so v must be constant on each maximal
connected component of Ω. But the restriction of v to the boundary of Ω must
vanish, which gives v = 0.

A direct application of Corollary 8.6 based on the comparison TN ≤ TD gives

Proposition 9.2. For any j ∈ N one has λN D

j (Ω) ≤ λj (Ω).

73
Remark 9.3. Actually a stronger version holds: If the embedding H 1 (Ω) → L2 (Ω)
is compact, then λN D
j+1 (Ω) ≤ λj (Ω) for all j. The proof is, of course, less elementary.

Another important aspect is the dependence of the eigenvalues on the domain.

Proposition 9.4 (Monotonicity with respect to domain, Dirichlet case).

e then λD
Let Ω ⊂ Ω, D
n (Ω) ≤ λn (Ω) for all n ∈ N.
e

Proof. We observe first that if f ∈ H01 (Ω), then its extension fe to Ω

e by zero belongs
to H01 (Ω).
e This allows one to write the following chain of equalities and inequalities:

k∇f k2L2 (Ω)

λD
e
n (Ω) = sup inf
e
e f ∈H01 (Ω),f
ψ1 ,...,ψn−1 ∈L2 (Ω)
e 6=0 kf k2L2 (Ω)
e
hf,ψj iL2 (Ω)
e =0

k∇fek2L2 (Ω)
e k∇f k2L2 (Ω)
≤ sup inf = sup inf
1
e f ∈H0 (Ω),f 6=0
ψ1 ,...,ψn−1 ∈L2 (Ω) kfek2L2 (Ω) 1
e f ∈H0 (Ω),f 6=0
ψ1 ,...,ψn−1 ∈L2 (Ω) kf k2L2 (Ω)
hf,ψj iL2 (Ω) =0
e
hfe,ψj iL2 (Ω)
e =0

k∇f k2L2 (Ω)

= sup inf = λD
n (Ω).
ψ1 ,...,ψn−1 ∈L2 (Ω) f ∈H01 (Ω),f 6=0 kf k2L2 (Ω)
hf,ψj iL2 (Ω) =0

Note that there is no easy generalization of this result to the Neumann case. The
reason can be understood at a certain abstract level. As can be seen from the proof,
for Ω ⊂ Ω e there exists an obvious embedding τ : H01 (Ω) → H01 (Ω)e (extension by
zero) such that kτ uk = kuk for all u ∈ H0 (Ω). If one replaces the spaces H01 by
1

H 1 , then the existence of a bounded embedding and the estimates for its norm in
terms of the two domains become non-trivial. We mention at least one important
case where a kind of the monotonicity can be proved.

Proposition 9.5 (Neumann eigenvalues of composed domains). Let Ω ⊂ Rd

be a bounded open domain with a regular boundary, and let Ω1 and Ω2 be non-
intersecting open subsets of Ω with regular boundaries such that Ω = Ω1 ∪ Ω2 , then
λN N
n (Ω1 ∪ Ω2 ) ≤ λj (Ω) for any j ∈ N.

Proof. Under the assumptions made, any function f ∈ H 1 (Ω) belongs to H 1 (Ω1 ∪

74
Ω2 ), while the spaces L2 (Ω) and L2 (Ω1 ∪ Ω2 ) coincide, and we have

k∇f k2L2 (Ω1 ∪Ω2 )

λN
n (Ω1 ∪ Ω2 ) = sup inf
1
ψ1 ,...,ψn−1 ∈L2 (Ω1 ∪Ω2 ) f ∈H (Ω1 ∪Ω2 ),f 6=0 kf k2L2 (Ω1 ∪Ω2 )
hf,ψj iL2 (Ω ∪Ω ) =0
1 2

k∇f k2L2 (Ω1 ∪Ω2 )

≤ sup inf
ψ1 ,...,ψn−1 ∈L2 (Ω1 ∪Ω2 ) f ∈H 1 (Ω),f 6=0 kf k2L2 (Ω1 ∪Ω2 )
hf,ψj iL2 (Ω ∪Ω ) =0
1 2

k∇f k2L2 (Ω)

= sup inf = λN
n (Ω).
ψ1 ,...,ψn−1 ∈L2 (Ω) f ∈H 1 (Ω),f 6=0 kf k2L2 (Ω)
hf,ψj iL2 (Ω) =0

Remark 9.6. Under the assumptions of proposition 9.5 for any n ∈ N we have
λD D
n (Ω) ≤ λn (Ω1 ∪ Ω2 ), which follows from the inclusion Ω1 ∪ Ω2 ⊂ Ω. Therefore, for
any n ∈ N one has the chain

λN (Ω1 ∪ Ω2 ) ≤ λN (Ω) ≤ λD (Ω) ≤ λD

n (Ω1 ∪ Ω2 ),

and this is the key argument of the so-called Dirichlet-Neumann bracketing which is
used e.g. for estimating the asymptotic behavior of the eigenvalues (see below).

We complete the discussion by proving the continuity of the Dirichlet eigenvalues

with respect to domain.
Proposition 9.7 (ContinuityS with respect to domain, Dirichlet). If Ωj ⊂
Ωj+1 for all j ∈ N, and Ω = ∞ Ω
j=1 j , then λD
n (Ω) = limj→∞ λD
n (Ωj ) for any n ∈ N.

Proof. Let us pick n ∈ N, and let f1 , . . . fn be the mutually orthogonal normalized

eigenfunctions associated with the eigenvalues λD D
1 (Ω), . . . , λn (Ω). If U denotes the
subspace spanned by f1 , . . . , fn , then for any f ∈ U one has the estimate k∇f k2 ≤
λD 2
n (Ω)kf k .
Now take an arbitrary ε > 0. Using the density of Cc∞ (Ω) in H01 (Ω) one can approx-
imate every fj by uj ∈ Cc∞ (Ω) in such a
way that u1 , . . . , un will be linearly inde-
pendent and that k∇uk2L2 (Ω) ≤ λD 2
n (Ω) + ε kukL2 (Ω) for all u from the n-dimensional
subspace V spanned by u1 , . . . , un . Let K ⊂ Ω be a compact subset containing the
supports of all uj and, as a consequence, the supports of all functions from V . One
can find M ∈ N such that K ⊂ Ωm for all m ≥ M , and then for all m ≥ M we
have V ⊂ H01 (Ωm ). Now let ψ1 , . . . , ψn−1 be arbitrary functions from L2 (Ωm ). As V
is n-dimensional, there exists a non-zero v ∈ V which is orthogonal to all ψj . This
means that
k∇uk2L2 (Ωm ) k∇vk2L2 (Ωm ) k∇vk2L2 (Ω)
inf ≤ = ≤ λD
n (Ω) + ε.
u∈H01 (Ωm ),u6=0 kuk2L2 (Ωm ) kvk2L2 (Ωm ) kvk2L2 (Ω)
u⊥ψ1 ,...ψn−1

Due to the arbitrariness of ψj we have λD D

n (Ωm ) ≤ λn (Ω) + ε for all m ≥ M . On the
other hand, λD D
n (Ω) ≤ λn (Ωm ) by monotonicity.

75
9.2 Weyl asymptotics
In this subsection we will discuss some aspects of the asymptotic behavior of the
Laplacian eigenvalues. We introduce the Dirichlet/Neumann counting functions
ND/N (λ, Ω) by
D/N
ND/N (λ, Ω) = the number of j ∈ N for which λj (Ω) ∈ (−∞, λ].

Clearly, ND/N (λ, Ω) is finite for any λ, and it has a jump at each eigenvalue; the
jump is equal to the multiplicity. We emphasize the following obvious properties:

ND (λ, Ω) ≤ NN (λ, Ω) (9.1)

D/N D/N D/N
N (λ, Ω1 ∪ Ω2 ) = N (λ, Ω1 ) + N (λ, Ω2 ) for Ω1 ∩ Ω = ∅. (9.2)
ND (λ, Ω) ≤ ND (Ω)e for Ω ⊂ Ω.e (9.3)

We are going to discuss the following rather general result on the behavior of the
counting functions ND λ → +∞:
Theorem 9.8 (Weyl asymptotics). We have

ND (λ, Ω) ωd
lim d/2
= vol(Ω),
λ→+∞ λ (2π)d

where ωd denotes the volume of the unit ball in Rd .

To keep simple notation we proceed with the proof for the case d = 2 only. Due to
ω2 = π we are reduced to prove
ND (λ, Ω) area(Ω)
lim = . (9.4)
λ→+∞ λ 4π

The proof consists of several steps.

Lemma 9.9. The Weyl asymptotics is valid for rectangles, for both NN and ND .

Proof. Let Ω = (0, a) × (0, b), a, b > 0. As shown in Example 6.16, the Neumann
eigenvalues of Ω are the numbers
 πm 2  πn 2
λ(m, n) := +
a b
with m, n ∈ N0 := {0} ∪ N, and the Dirichlet spectrum consists of the eigenvalues
λ(m, n) with m, n ∈ N. Denote
n
2 x2 y 2 λ o
D(λ) := (x, y) ∈ R : 2 + 2 ≤ 2 , x ≥ 0, y ≥ 0 ,
a b π



then ND (λ, Ω) = #D(λ) ∩ N × N and NN (λ, Ω) = #D(λ) ∩ N0 × N0 , where #
denotes the cardinality.

76
First, counting the points (n, 0) and (0, n) with n ∈ N0 inside D(λ) we obtain the
majoration
a+b
√
NN (λ) − ND (λ) ≤ + 2 λ, λ > 0.
π

the union of the unit cubes [m − 1, m] × [n − 1, n]
At the same time, D(λ) contains
with (m, n) ∈ D(λ) ∩ N × N . As there are exactly ND (λ, Ω) such cubes, we have

λab
ND (λ, Ω) ≤ area D(λ) = .
4π
We also observe that D(λ) is contained in
the union of the unit cubes [m, m + 1] ×
[n, n + 1] with (m, n) ∈ D(λ) ∩ N0 × N0 . As the number of such cubes is exactly
NN (λ, Ω), this gives
λab
NN (λ, Ω) ≥ area D(λ) = .
4π
Putting all together we arrive at
λab a+b
√ λab a+b
√
≤ NN (λ, Ω) ≤ N D (λ, Ω) + +2 λ≤ + + 2 λ,
4π π 4π π
and it remains to recall that area(Ω) = ab.

Definition 9.10 (Domains composed from rectangles). We say that a domain

Ω with a regular boundary is composed from rectangles if there exists a finite family
of non-intersecting open rectangles Ωj , j = 1, . . . , k, with Ω = kj=1 Ωj .
S

Lemma 9.11. The Weyl asymptotics holds for domains composed from rectangles.

Proof. Let Ω be a domain composed from rectangles, ant let Ωj , j = 1, . . . , k, be

the rectangles as in Definition 9.10. Using Remark 9.6 and the equality (9.2) we
obtain the chain

ND (λ, Ω1 ) + · · · + NN (λ, Ωk ) ND (λ, Ω1 ∪ · · · ∪ Ωk ) ND (λ, Ω)

= ≤
λ λ λ
NN (λ, Ω) NN (λ, Ω1 ∪ · · · ∪ Ωk ) NN (λ, Ω1 ) + · · · + ND (λ, Ωk )
≤ ≤ = ,
λ λ λ
and the result is obtained by applying Lemma 9.9 to the quotients ND/N (λ, Ωj )/λ
and by noting that area(Ω) = area(Ω1 ) + · · · + area(Ωk ).

Proof of Theorem 9.8. Let Ω be a domain with a regular boundary. It is a stan-

dard result of the analysis that for any ε > 0 one can find two domains Ωε and Ω
eε
such that:

• both Ωε and Ω
e ε are composed from rectangles,

• Ωε ⊂ Ω ⊂ Ω
e ε,

• area(Ω
e ε \ Ωε ) < ε.

77
Using (9.1) and the monotonicity of the Dirichlet eigenvalues with respect to domain
we have:
ND (λ, Ωε ) ND (λ, Ω) ND (λ, Ωe ε) NN (λ, Ω
e ε)
≤ ≤ ≤ .
λ λ λ λ
By Lemma 9.11, we can find λε > 0 such that

area(Ωε ) − ε ND (λ, Ω) area(Ω

e ε) + ε
≤ ≤ for λ > λε .
4π λ 4π

At the same time, area(Ωε ) ≥ area(Ω) − ε and area(Ω

e ε ) ≤ area(Ω) + ε, so for λ > λε
we have
area(Ω) − 2ε ND (λ, Ω) area(Ω) + 2ε
≤ ≤ ,
4π λ 4π
which gives the sought result.

We note that the Weyl asymptotics also holds for the Neumann Laplacian if the
domain is sufficiently smooth, which can be proved using suitable extension theorem
for Sobolev spaces. The Weyl asymptotics is one of the basic results on the relations
between the Dirichlet/Neumann eigenvalues and the geometric properties of the
domain. It states, in particular, that the spectrum of the domain contains the
information on its dimension and its volume. There are various refinements involving
lower order terms with respect to λ, and the respective coefficients contains some
information on the topology of the domain, on its boundary etc.

78
References
[1] Yu. M. Berezansky: Expansion in eigenfunctions of self-adjoint operators.
AMS, Providence, 1968.

[2] F. A. Berezin, M. A. Shubin: The Schrödinger equation. Springer, 1991.

[3] J. Brüning, V. Geyler, K. Pankrashkin: Spectra of self-adjoint extensions and

applications to solvable Schrödinger operators. Rev. Math. Phys. 20 (2008)
1–70.

[4] H. L. Cycon, R. G. Froese, W. Kirsch, B. Simon: Schrödinger operators. With

applications to quantum mechanics and global geometry. Spronger, 1987.

[5] E. B. Davies: Spectral theory and differential operators. Cambridge University

Press, 1995.

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

[7] P. Grisvard: Singularities in boundary value problems. Recherches en

Mathématiques Appliquées, Vol. 22. Masson, Paris, 1992.

[8] H. Hanche-Olsena, H. Holden: The Kolmogorov-Riesz compactness theorem.

Expositiones Mathematicae, Vol. 28, Issue 4 (2010), pp. 385–394.

[9] B. Helffer: Spectral theory and applications. An elementary introductory

course. Available at http://www.math.u-psud.fr/~helffer/

[10] B. Helffer: Spectral theory and its applications. Cambridge Studies in Applied
Mathematics, 2012.

[11] D. Jerison, C.E. Kenig: Unique continuation and absence of positive eigenval-
ues for Schröodinger operators (with an appendix by E.M. Stein).. Annals of
Math. (2) 121:3 (1985) 463–494.

[12] M. Reed, B. Simon: Methods of modern mathematical physics. Vol. 1. Func-

tional analysis. Academic Press, 1972.

[13] M. Reed, B. Simon: Methods of modern mathematical physics. Vol. 4. Analysis

of operators. Academic Press, 1978.

[14] V. Kondrat’ev, M. Shubin: Discreteness of spectrum for the Schrödinger op-

erators on manifolds of bounded geometry. The Maz’ya anniversary collection,
Vol. 2 (Rostock, 1998), pp. 185–226 (Operator Theory: Advances and Appli-
cations, vol. 110), Birkhäuser, Basel, 1999.

[15] T. Kato: Perturbation theory for linear operators. Springer-Verlag, Berlin etc.,
1966.

79
[16] S. Sternberg: Theory of functions of a real variable. Available at http://www.
math.harvard.edu/~shlomo/docs/Real_Variables.pdf

[17] G. Teschl: Mathematical methods in quantum mechanics. With applications to

Schrödinger operators. AMS, 2009. Available from http://www.mat.univie.
ac.at/~gerald/

80