Hilbert space methods for quantum mechanics

S. Richard

Spring Semester 2016

2
Contents

1 Hilbert space and bounded linear operators 5

1.1 Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Vector-valued functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Bounded linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Special classes of bounded linear operators . . . . . . . . . . . . . . . . 13
1.5 Operator-valued maps . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Unbounded operators 19
2.1 Unbounded, closed, and self-adjoint operators . . . . . . . . . . . . . . 19
2.2 Resolvent and spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Perturbation theory for self-adjoint operators . . . . . . . . . . . . . . . 25

3 Examples 31
3.1 Multiplication and convolution operators . . . . . . . . . . . . . . . . . 31
3.1.1 The harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Schrödinger operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 The hydrogen atom . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 The Weyl calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Schrödinger operators with x12 -potential . . . . . . . . . . . . . . . . . . 39
3.4.1 Two families of Schrödinger operators . . . . . . . . . . . . . . . 40

4 Spectral theory for self-adjoint operators 43

4.1 Stieltjes measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Spectral measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Spectral parts of a self-adjoint operator . . . . . . . . . . . . . . . . . . 51
4.4 The resolvent near the spectrum . . . . . . . . . . . . . . . . . . . . . . 55

5 Scattering theory 63
5.1 Evolution groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 Wave operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 Scattering operator and completeness . . . . . . . . . . . . . . . . . . . 73

3
4 CONTENTS

6 Commutator methods 77
6.1 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Regularity classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.3 Affiliation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.4 Locally smooth operators . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.5 Limiting absorption principle . . . . . . . . . . . . . . . . . . . . . . . 88
6.6 The method of differential inequalities . . . . . . . . . . . . . . . . . . 91
Chapter 1

Hilbert space and bounded linear

operators

This chapter is mainly based on the first two chapters of the book [Amr]. Its content
is quite standard and this theory can be seen as a special instance of bounded linear
operators on more general Banach spaces.

1.1 Hilbert space

Definition 1.1.1. A (complex) Hilbert space H is a vector space on C with a strictly
positive scalar product (or inner product) which is complete for the associated norm1
and which admits a countable orthonormal basis. The scalar product is denoted by ⟨·, ·⟩
and the corresponding norm by ∥ · ∥.
In particular, note that for any f, g, h ∈ H and α ∈ C the following properties hold:
(i) ⟨f, g⟩ = ⟨g, f ⟩,

(ii) ⟨f, g + αh⟩ = ⟨f, g⟩ + α⟨f, h⟩,

(iii) ∥f ∥2 = ⟨f, f ⟩ ≥ 0, and ∥f ∥ = 0 if and only if f = 0.

Note that ⟨g, f ⟩ means the complex conjugate of ⟨g, f ⟩. Note also that the linearity
in the second argument in (ii) is a matter of convention, many authors define the
linearity in the first argument. In (iii) the norm of f is defined in terms of the scalar
product ⟨f, f ⟩. We emphasize that the scalar product can also be defined in terms of
the norm of H, this is the content of the polarisation identity:

4⟨f, g⟩ = ∥f + g∥2 − ∥f − g∥2 − i∥f + ig∥2 + i∥f − ig∥2 . (1.1)

1
Recall that H is said to be complete if any Cauchy sequence in H has a limit in H. More precisely,
{fn }n∈N ⊂ H is a Cauchy sequence if for any ε > 0 there exists N ∈ N such that ∥fn − fm ∥ < ε
for any n, m ≥ N . Then H is complete if for any such sequence there exists f∞ ∈ H such that
limn→∞ ∥fn − f∞ ∥ = 0.

5
6 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

From now on, the symbol H will always denote a Hilbert space.
∑
Examples 1.1.2. (i) H = Cd with ⟨α, β⟩ = dj=1 αj βj for any α, β ∈ Cd ,
∑
(ii) H = l2 (Z) with ⟨a, b⟩ = j∈Z aj bj for any a, b ∈ l2 (Z),
∫
(iii) H = L2 (Rd ) with ⟨f, g⟩ = Rd f (x) g(x) dx for any f, g ∈ L2 (Rd ).
Let us recall some useful inequalities: For any f, g ∈ H one has

|⟨f, g⟩| ≤ ∥f ∥ ∥g∥ Schwarz inequality, (1.2)

∥f + g∥ ≤ ∥f ∥ + ∥g∥ triangle inequality, (1.3)
∥f + g∥ ≤ 2∥f ∥ + 2∥g∥ ,
2 2 2
(1.4)
∥f ∥ − ∥g∥ ≤ ∥f − g∥. (1.5)

The proof of these inequalities is standard and is left as a free exercise, see also [Amr,
p. 3-4]. Let us also recall that f, g ∈ H are said to be orthogonal if ⟨f, g⟩ = 0.
Definition 1.1.3. A sequence {fn }n∈N ⊂ H is strongly convergent to f∞ ∈ H if
limn→∞ ∥fn − f∞ ∥ = 0, or is weakly convergent to f∞ ∈ H if for any g ∈ H one has
limn→∞ ⟨g, fn − f∞ ⟩ = 0. One writes s− limn→∞ fn = f∞ if the sequence is strongly
convergent, and w− limn→∞ fn = f∞ if the sequence is weakly convergent.
Clearly, a strongly convergent sequence is also weakly convergent. The converse is
not true.
Exercise 1.1.4. In the Hilbert space L2 (R), exhibit a sequence which is weakly conver-
gent but not strongly convergent.
Lemma 1.1.5. Consider a sequence {fn }n∈N ⊂ H. One has

s− lim fn = f∞ ⇐⇒ w− lim fn = f∞ and lim ∥fn ∥ = ∥f∞ ∥.

n→∞ n→∞ n→∞

Proof. Assume first that s− limn→∞ fn = f∞ . By the Schwarz inequality one infers that
for any g ∈ H:

|⟨g, fn − f∞ ⟩| ≤ ∥fn − f∞ ∥ ∥g∥ → 0 as n → ∞,

which means that w− limn→∞ fn = f∞ . In addition, by (1.5) one also gets

∥fn ∥ − ∥f∞ ∥ ≤ ∥fn − f∞ ∥ → 0 as n → ∞,

and thus limn→∞ ∥fn ∥ = ∥f∞ ∥.

For the reverse implication, observe first that

∥fn − f∞ ∥2 = ∥fn ∥2 + ∥f∞ ∥2 − ⟨fn , f∞ ⟩ − ⟨f∞ , fn ⟩. (1.6)

If w− limn→∞ fn = f∞ and limn→∞ ∥fn ∥ = ∥f∞ ∥, then the right-hand side of (1.6)
converges to ∥f∞ ∥2 + ∥f∞ ∥2 − ∥f∞ ∥2 − ∥f∞ ∥2 = 0, so that s− limn→∞ fn = f∞ .
1.1. HILBERT SPACE 7

Let us also note that if s− limn→∞ fn = f∞ and s− limn→∞ gn = g∞ then one has

lim ⟨fn , gn ⟩ = ⟨f∞ , g∞ ⟩

n→∞

by a simple application of the Schwarz inequality.

Exercise 1.1.6. Let {en }n∈N be an orthonormal basis of an infinite dimensional Hilbert
space. Show that w− limn→∞ en = 0, but that s− limn→∞ en does not exist.

Exercise 1.1.7. Show that the limit of a strong or a weak Cauchy sequence is unique.
Show also that such a sequence is bounded, i.e. if {fn }n∈N denotes this Cauchy sequence,
then supn∈N ∥fn ∥ < ∞.

For the weak Cauchy sequence, the boundedness can be obtained from the follow-
ing quite general result which will be useful later on. Its proof can be found in [Kat,
Thm. III.1.29]. In the statement, Λ is simply a set.
Theorem 1.1.8 (Uniform boundedness principle). Let {φλ }λ∈Λ be a family of contin-
uous maps2 φλ : H → [0, ∞) satisfying

φλ (f + g) ≤ φλ (f ) + φλ (g) ∀f, g ∈ H.

If the set {φλ (f )}λ∈Λ ⊂ [0, ∞) is bounded for any fixed f ∈ H, then the family {φλ }λ∈Λ
is uniformly bounded, i.e. there exists c > 0 such that supλ φλ (f ) ≤ c for any f ∈ H
with ∥f ∥ = 1.
In the next definition, we introduce the notion of a linear manifold and of a subspace
of a Hilbert space.

Definition 1.1.9. A linear manifold M of a Hilbert space H is a linear subset of H,

or more precisely ∀f, g ∈ M and α ∈ C one has f + αg ∈ M. If M is closed (⇔ any
Cauchy sequence in M converges strongly in M), then M is called a subspace of H.
Note that if M is closed, then M is a Hilbert space in itself, with the scalar product
and norm inherited from H. Be aware that some authors call subspace what we have
defined as a linear manifold, and call closed subspace what we simply call a subspace.
Examples 1.1.10. (i) If f1 , . . . , fn ∈ H, then Vect(f1 , . . . , fn ) is the closed vector
space generated by the linear combinations of f1 , . . . fn . Vect(f1 , . . . , fn ) is a sub-
space.

(ii) If M is a subset of H, then

M⊥ := {f ∈ H | ⟨f, g⟩ = 0, ∀g ∈ M} (1.7)

is a subspace of H.
2
φλ is continuous if φλ (fn ) → φλ (f∞ ) whenever s− limn→∞ fn = f∞ .
8 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

Exercise 1.1.11. Check that in the above example the set M⊥ is a subspace of H.

Exercise 1.1.12. Check that a linear manifold M ⊂ H is dense in H if and only if

M⊥ = {0}.

If M is a subset of H the subspace M⊥ is called the orthocomplement of M in

H. The following result is important in the setting of Hilbert spaces. Its proof is not
complicated but a little bit lengthy, we thus refer to [Amr, Prop. 1.7].

Proposition 1.1.13 (Projection Theorem). Let M be a subspace of a Hilbert space H.

Then, for any f ∈ H there exist a unique f1 ∈ M and a unique f2 ∈ M⊥ such that
f = f1 + f2 .

Let us close this section with the so-called Riesz Lemma. For that purpose, recall
first that the dual H∗ of the Hilbert space H consists in the set of all bounded linear
functionals on H, i.e. H∗ consists in all mappings φ : H → C satisfying for any f, g ∈ H
and α ∈ C

(i) φ(f + αg) = φ(f ) + αφ(g), (linearity)

(ii) |φ(f )| ≤ c∥f ∥, (boundedness)

where c is a constant independent of f . One then sets

|φ(f )|
∥φ∥H∗ := sup .
0̸=f ∈H ∥f ∥

Clearly, if g ∈ H, then g defines an element φg of H∗ by setting φg (f ) := ⟨g, f ⟩.

Indeed φg is linear and one has

1 1
∥φg ∥H∗ := sup |⟨g, f ⟩| ≤ sup ∥g∥ ∥f ∥ = ∥g∥.
0̸=f ∈H ∥f ∥ 0̸=f ∈H ∥f ∥

In fact, note that ∥φg ∥H∗ = ∥g∥ since ∥g∥

1 1
φg (g) = ∥g∥ ∥g∥2 = ∥g∥.
The following statement shows that any element φ ∈ H∗ can be obtained from an
element g ∈ H. It corresponds thus to a converse of the previous construction.

Lemma 1.1.14 (Riesz Lemma). For any φ ∈ H∗ , there exists a unique g ∈ H such
that for any f ∈ H
φ(f ) = ⟨g, f ⟩.
In addition, g satisfies ∥φ∥H∗ = ∥g∥.

Since the proof is quite standard, we only sketch it and leave the details to the
reader, see also [Amr, Prop. 1.8].
1.2. VECTOR-VALUED FUNCTIONS 9

Sketch of the proof. If φ ≡ 0, then one can set g := 0 and observe trivially that φ = φg .
If φ ̸= 0, let us first define M := {f ∈ H | φ(f ) = 0} and observe that M is
a subspace of H. One also observes that M ̸= H since otherwise φ ≡ 0. Thus, let
h ∈ H such that φ(h) ̸= 0 and decompose h = h1 + h2 with h1 ∈ M and h2 ∈ M⊥ by
Proposition 1.1.13. One infers then that φ(h2 ) = φ(h) ̸= 0.
For arbitrary f ∈ H one can consider the element f − φ(h φ(f )
h2 ∈ H and observe that
( ) 2)

φ f − φ(h
φ(f )
2)
h2 = 0. One deduces that f − φ(h φ(f )
2)
h2 belongs to M, and since h2 ∈ M⊥
one infers that
φ(h2 )
φ(f ) = ⟨h2 , f ⟩.
∥h2 ∥2
One can thus set g := φ(h2 )
h
∥h2 ∥2 2
∈ H and easily obtain the remaining parts of the state-
ment.
As a consequence of the previous statement, one often identifies H∗ with H itself.
Exercise 1.1.15. Check that this identification is not linear but anti-linear.

1.2 Vector-valued functions

Let H be a Hilbert space and let Λ be a set. A vector-valued function if a map f : Λ → H,
i.e. for any λ ∈ Λ one has f (λ) ∈ H. In application, we shall mostly consider the special
case Λ = R or Λ = [a, b] with a, b ∈ R and a < b.
The following definitions are mimicked from the special case H = C, but different
topologies on H can be considered:
Definition 1.2.1. Let J := (a, b) with a < b and consider a vector-valued function
f : J → H.
(i) f is strongly continuous on J if for any t ∈ J one has limε→0 ∥f (t+ε)−f (t)∥ = 0,

(ii) f is weakly continuous on J if for any t ∈ J and any g ∈ H one has

⟨ ⟩
lim g, f (t + ε) − f (t) = 0,
ε→0

(iii) f is strongly differentiable on J if there exists another vector-valued function

f ′ : J → H such that for any t ∈ J one has
( )
lim 1ε f (t + ε) − f (t) − f ′ (t) = 0,
ε→0

(iii) f is weakly differentiable on J if there exists another vector-valued function f ′ :

J → H such that for any t ∈ J and g ∈ H one has
⟨ ( ) ⟩
lim g, 1ε f (t + ε) − f (t) − f ′ (t) = 0,
ε→0
10 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

The map f ′ is called the strong derivative, respectively the weak derivative, of f .
Integrals of vector-valued functions can be defined in several senses, but we shall
restrict ourselves to Riemann-type integrals. The construction is then similar to real or
complex-valued functions, by considering finer and finer partitions of a bounded interval
J. Improper Riemann integrals can also be defined in analogy with the scalar case by
a limiting process. Note that these integrals can exist either in the strong sense (strong
topology on H) or in the weak sense (weak topology on H). In the sequel, we consider
only the existence of such integrals in the strong sense.
Let us thus consider J := (a, b] with a < b and let us set Π = {s0 , . . . , sn ; u1 , . . . un }
with a = s0 < u1 ≤ s1 < u2 ≤ s2 < · · · < un ≤ sn = b for a partition of J. One also
sets |Π| := maxk∈{1,...,n} |sk − sk−1 | and the Riemann sum
∑
n
Σ(Π, f ) := (sk − sk−1 )f (uk ).
k=1

If one considers then a sequence {Πi }i∈N of partitions of J with |Πi | → 0 as i → ∞ one
writes ∫ ∫ b
f (t)dt ≡ f (t)dt = s− lim Σ(Πi , f )
J a i→∞

if this limit exists and is independent of the sequence of partitions. In this case, one
says that f is strongly integrable on (a, b]. Clearly, similar definitions hold for J = (a, b)
or J = [a, b]. Infinite intervals can be considered by a limiting process as long as the
corresponding limits exist.
The following statements can then be proved in a way similar to the scalar case.
Proposition 1.2.2. Let (a, b] and (b, c] be finite or infinite intervals and suppose that
all the subsequent integrals exist. Then one has
∫b ∫c ∫c
(i) a f (t)dt + b f (t)dt = a f (t)dt,
∫b( ) ∫b ∫b
(ii) a αf1 (t) + f2 (t) dt = α a f1 (t)dt + a f2 (t)dt,
∫b ∫b
(iii) a
f (t)dt ≤ a
∥f (t)∥dt.
For the existence of these integrals one has:
Proposition 1.2.3. (i) If [a, b] is a finite closed interval and f : [a, b] → H is
∫b
strongly continuous, then a f (t)dt exists,
∫b ∫b
(ii) If a < b are arbitrary and a ∥f (t)∥dt < ∞, then a f (t)dt exists,
(iii) If f is strongly differentiable on (a, b) and its derivative f ′ is strongly continuous
and integrable on [a, b] then
∫ b
f ′ (t)dt = f (b) − f (a).
a
1.3. BOUNDED LINEAR OPERATORS 11

1.3 Bounded linear operators

First of all, let us recall that a linear map B between two complex vector spaces M
and N satisfies B(f + αg) = Bf + α Bg for all f, g ∈ M and α ∈ C.

Definition 1.3.1. A map B : H → H is a bounded linear operator if B : H → H is a

linear map, and if there exists c > 0 such that ∥Bf ∥ ≤ c∥f ∥ for all f ∈ H. The set of
all bounded linear operators on H is denoted by B(H).

For any B ∈ B(H), one sets

∥B∥ := inf{c > 0 | ∥Bf ∥ ≤ c∥f ∥ ∀f ∈ H}

∥Bf ∥
= sup . (1.8)
0̸=f ∈H ∥f ∥

and call it the norm of B. Note that the same notation is used for the norm of an
element of H and for the norm of an element of B(H), but this does not lead to any
confusion. Let us also introduce the range of an operator B ∈ B(H), namely

Ran(B) := BH = {f ∈ H | f = Bg for some g ∈ H}. (1.9)

This notion will be important when the inverse of an operator will be discussed.

Exercise 1.3.2. Let M1 , M2 be two dense linear manifolds of H, and let B ∈ B(H).
Show that
∥B∥ = sup |⟨f, Bg⟩|. (1.10)
f ∈M1 ,g∈M2 with ∥f ∥=∥g∥=1

Exercise 1.3.3. Show that B(H) is a complete normed algebra and that the inequality

∥AB∥ ≤ ∥A∥ ∥B∥ (1.11)

holds for any A, B ∈ B(H).

An additional structure can be added to B(H): an involution. More precisely, for

any B ∈ B(H) and any f, g ∈ H one sets

⟨B ∗ f, g⟩ := ⟨f, Bg⟩. (1.12)

Exercise 1.3.4. For any B ∈ B(H) show that

(i) B ∗ is uniquely defined by (1.12) and satisfies B ∗ ∈ B(H) with ∥B ∗ ∥ = ∥B∥,

(ii) (B ∗ )∗ = B,

(iii) ∥B ∗ B∥ = ∥B∥2 ,

(iv) If A ∈ B(H), then (AB)∗ = B ∗ A∗ .

12 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

The operator B ∗ is called the adjoint of B, and the proof the unicity in (i) involves
the Riesz Lemma. A complete normed algebra endowed with an involution for which
the property (iii) holds is called a C ∗ -algebra. In particular B(H) is a C ∗ -algebra. Such
algebras have a well-developed and deep theory, see for example [Mur]. However, we
shall not go further in this direction in this course.
We have already considered two distinct topologies on H, namely the strong and
the weak topology. On B(H) there exist several topologies, but we shall consider only
three of them.

Definition 1.3.5. A sequence {Bn }n∈N ⊂ B(H) is uniformly convergent to B∞ ∈

B(H) if limn→∞ ∥Bn − B∞ ∥ = 0, is strongly convergent to B∞ ∈ B(H) if for any
f ∈ H one has limn→∞ ∥Bn f − B∞ f ∥ = 0, or is weakly convergent to B∞ ∈ B(H)
if for any f, g ∈ H one has limn→∞ ⟨f, Bn g − B∞ g⟩ = 0. In these cases, one writes
respectively u − limn→∞ Bn = B∞ , s − limn→∞ Bn = B∞ and w − limn→∞ Bn = B∞ .

Clearly, uniform convergence implies strong convergence, and strong convergence

implies weak convergence. The reverse statements are not true. Note that if {Bn }n∈N ⊂
B(H) is weakly convergent, then the sequence {Bn∗ }n∈N of its adjoint operators is also
weakly convergent. However, the same statement does not hold for a strongly convergent
sequence. Finally, we shall not prove but often use that B(H) is also weakly and strongly
closed. In other words, any weakly (or strongly) Cauchy sequence in B(H) converges
in B(H).

Exercise 1.3.6. Let {An }n∈N ⊂ B(H) and {Bn }n∈N ⊂ B(H) be two strongly conver-
gent sequence in B(H), with limits A∞ and B∞ respectively. Show that the sequence
{An Bn }n∈N is strongly convergent to the element A∞ B∞ .

Let us close this section with some information about the inverse of a bounded
operator. Additional information on the inverse in relation with unbounded operators
will be provided in the sequel.

Definition 1.3.7. An operator B ∈ B(H) is invertible if the equation Bf = 0 only

admits the solution f = 0. In such a case, there exists a linear map B −1 : Ran(B) → H
which satisfies B −1 Bf = f for any f ∈ H, and BB −1 g = g for any g ∈ Ran(B). If B is
invertible and Ran(B) = H, then B −1 ∈ B(H) and B is said to be invertible in B(H)
(or boundedly invertible).

Note that the two conditions B invertible and Ran(B) = H imply B −1 ∈ B(H) is
a consequence of the Closed graph Theorem. In the sequel, we shall use the notation
1 ∈ B(H) for the operator defined on any f ∈ H by 1f = f , and 0 ∈ B(H) for the
operator defined by 0f = 0.
The next statement is very useful in applications, and holds in a much more general
context. Its proof is classical and can be found in every textbook.
1.4. SPECIAL CLASSES OF BOUNDED LINEAR OPERATORS 13

Lemma 1.3.8 (Neumann series). If B ∈ B(H) and ∥B∥ < 1, then the operator (1−B)
is invertible in B(H), with
∑
∞
(1 − B)−1 = Bn,
n=0
−1 −1
and with (1 − B) ≤ (1 − ∥B∥) . The series converges in the uniform norm of
B(H).

Note that we have used the identity B 0 = 1.

1.4 Special classes of bounded linear operators

In this section we provide some information on some subsets of B(H). We start with
some operators which will play an important role in the sequel.

Definition 1.4.1. An operator B ∈ B(H) is called self-adjoint if B ∗ = B, or equiva-

lently if for any f, g ∈ H one has

⟨f, Bg⟩ = ⟨Bf, g⟩. (1.13)

For these operators the computation of their norm can be simplified (see also Ex-
ercise 1.3.2) :

Exercise 1.4.2. If B ∈ B(H) is self-adjoint and if M is a dense linear manifold in

H, show that
∥B∥ = sup |⟨f, Bf ⟩|. (1.14)
f ∈M, ∥f ∥=1

A special set of self-adjoint operators is provided by the set of orthogonal projec-

tions:

Definition 1.4.3. An element P ∈ B(H) is an orthogonal projection if P = P 2 = P ∗ .

It not difficult to check that there is a one-to-one correspondence between the set
of subspaces of H and the set of orthogonal projections in B(H). Indeed, any orthog-
onal projection P defines a subspace M := P H. Conversely by taking the projection
Theorem (Proposition 1.1.13) into account one infers that for any subspace M one can
define an orthogonal projection P with P H = M.
In the sequel, we might simply say projection instead of orthogonal projection.
However, let us stress that in other contexts a projection is often an operator P satisfying
only the relation P 2 = P .
We gather in the next exercise some easy relations between orthogonal projections
and the underlying subspaces. For that purpose we use the notation PM , PN for the
orthogonal projections on the subspaces M and N of H.

Exercise 1.4.4. Show the following relations:

14 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

(i) If PM PN = PN PM , then PM PN is a projection and the associated subspace is

M ∩ N,
(ii) If M ⊂ N , then PM PN = PN PM = PM ,
(iii) If M⊥N , then PM PN = PN PM = 0, and PM⊕N = PM + PN ,
(iv) If PM PN = 0, then M⊥N .
Let us now consider unitary operators, and then more general isometries and partial
isometries. For that purpose, we recall that 1 denotes the identify operator in B(H).
Definition 1.4.5. An element U ∈ B(H) is a unitary operator if U U ∗ = 1 and if
U ∗ U = 1.
Note that if U is unitary, then U is invertible in B(H) with U −1 = U ∗ . Indeed,
observe first that U f = 0 implies f = U ∗ (U f ) = U ∗ 0 = 0. Secondly, for any g ∈ H, one
has g = U (U ∗ g), and thus Ran(U ) = H. Finally, the equality U −1 = U ∗ follows from
the unicity of the inverse.
More generally, an element V ∈ B(H) is called an isometry if the equality
V ∗V = 1 (1.15)
holds. Clearly, a unitary operator is an instance of an isometry. For isometries the
following properties can easily be obtained.
Proposition 1.4.6. a) Let V ∈ B(H) be an isometry. Then
(i) V preserves the scalar product, namely ⟨V f, V g⟩ = ⟨f, g⟩ for any f, g ∈ H,
(ii) V preserves the norm, namely ∥V f ∥ = ∥f ∥ for any f ∈ H,
(iii) If H ̸= {0} then ∥V ∥ = 1,
(iv) V V ∗ is the projection on Ran(V ),
(v) V is invertible (in the sense of Definition 1.3.7),
(vi) The adjoint V ∗ satisfies V ∗ f = V −1 f if f ∈ Ran(V ), and V ∗ g = 0 if g⊥ Ran(V ).
b) An element W ∈ B(H) is an isometry if and only if ∥W f ∥ = ∥f ∥ for all f ∈ H.
Exercise 1.4.7. Provide a proof for the previous proposition (as well as the proof of
the next proposition).
More generally one defines a partial isometry as an element W ∈ B(H) such that
W ∗W = P (1.16)
with P an orthogonal projection. Again, unitary operators or isometries are special
examples of partial isometries.
As before the following properties of partial isometries can be easily proved.
1.4. SPECIAL CLASSES OF BOUNDED LINEAR OPERATORS 15

Proposition 1.4.8. Let W ∈ B(H) be a partial isometry as defined in (1.16). Then

(i) one has W P = W and ⟨W f, W g⟩ = ⟨P f, P g⟩ for any f, g ∈ H,

(ii) If P ̸= 0 then ∥W ∥ = 1,

(iii) W W ∗ is the projection on Ran(W ).

For a partial isometry W one usually calls initial set projection the projection
defined by W ∗ W and by final set projection the projection defined by W W ∗ .
Let us now introduce a last subset of bounded operators, namely the ideal of compact
operators. For that purpose, consider first any family {gj , hj }N
j=1 ⊂ H and for any f ∈ H
one sets
∑ N
Af := ⟨gj , f ⟩ hj . (1.17)
j=1

Then A ∈ B(H), and Ran(A) ⊂ Vect(h1 , . . . , hN ). Such an

( operator
) A is called a finite
rank operator. In fact, any operator B ∈ B(H) with dim Ran(B) < ∞ is a finite rank
operator.

Exercise 1.4.9. For the operator A defined in (1.17), give an upper estimate for ∥A∥
and compute A∗ .

Definition 1.4.10. An element B ∈ B(H) is a compact operator if there exists a

family {An }n∈N of finite rank operators such that limn→∞ ∥B − An ∥ = 0. The set of all
compact operators is denoted by K (H).

The following proposition contains sone basic properties of K (H). Its proof can be
obtained by playing with families of finite rank operators.

Proposition 1.4.11. The following properties hold:

(i) B ∈ K (H) ⇐⇒ B ∗ ∈ K (H),

(ii) K (H) is a ∗-algebra, complete for the norm ∥ · ∥,

(iii) If B ∈ K (H) and A ∈ B(H), then AB and BA belong to K (H).

As a consequence, K (H) is a C ∗ -algebra and an ideal of B(H). In fact, compact

operators have the nice property of improving some convergences, as shown in the next
statement.

Proposition 1.4.12. Let K ∈ K (H)).

(i) If {fn }n∈N ⊂ H is a weakly convergent sequence with limit f∞ ∈ H, then the
sequence {Kfn }n∈N strongly converges to Kf∞ ,
16 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

(ii) If the sequence {Bn }n∈N ⊂ B(H) strongly converges to B∞ ∈ B(H), then the
sequences {Bn K}n∈N and {KBn∗ }n∈N converge in norm to B∞ K and KB∞∗
, re-
spectively.

Proof. a) Let us first set φn := fn − f∞ and observe that w− limn→∞ φn = 0. By an

application of the uniform boundedness principle, see Theorem 1.1.8, it follows that
{∥φn ∥}n∈N is bounded, i.e. there exists M > 0 such that ∥φn ∥ ≤ M for any n ∈ N.
Since K is compact, for any ε > 0 there exists a finite rank operator A of the form
given in (1.17) such that ∥K − A∥ ≤ 2M ε
. Then one has

ε ∑
N
∥Kφn ∥ ≤ ∥(K − A)φn ∥ + ∥Aφn ∥ ≤ + |⟨gj , φn ⟩| ∥hj ∥.
2 j=1

Since w− limn→∞ φn = 0 there exists n0 ∈ N such that ⟨gj , φn ⟩| ∥hj ∥ ≤ 2N ε

for any
j ∈ {1, . . . , N } and all n ≥ n0 . As a consequence, one infers that ∥Kφn ∥ ≤ ε for all
n ≥ n0 , or in other words s− limn→∞ Kφn = 0.
b) Let us set Cn := Bn − B∞ such that s− limn→∞ Cn = 0. As before, there exists
M > 0 such that ∥Cn ∥ ≤ M for any n ∈ N. For any ε > 0 consider a finite rank
operator A of the form (1.17) such that ∥K − A∥ ≤ 2M ε
. Then observe that for any
f ∈H

∥Cn Kf ∥ ≤ M ∥(K − A)f ∥ + ∥Cn Af ∥

∑
N
≤ M ∥K − A∥ ∥f ∥ + |⟨gj , f ⟩| ∥Cn hj ∥
j=1
{ ∑
N }
≤ M ∥K − A∥ + ∥gj ∥ ∥Cn hj ∥ ∥f ∥.
j=1

Since Cn strongly converges to 0 one can then choose n0 ∈ N such that ∥gj ∥ ∥Cn hj ∥ ≤ 2N
ε

for any j ∈ {1, . . . N } and all n ≥ n0 . One then infers that ∥Cn K∥ ≤ ε for any n ≥ n0 ,
which means that the sequence {Cn K}n∈N uniformly converges to 0. The statement
about {KBn∗ }n∈N can be proved analogously by taking the equality ∥KBn∗ − KB∞ ∗
∥=
∗ ∗ ∗
∥Bn K − B∞ K ∥ into account and by remembering that K is compact as well.

Exercise 1.4.13. Check that a projection P is a compact operator if and only if P H

is of finite dimension.

Extension 1.4.14. There are various subalgebras of K (H), for example the algebra of
Hilbert-Schmidt operators, the algebra of trace class operators, and more generally the
Schatten classes. Note that these algebras are not closed with respect to the norm ∥ · ∥
but with respect to some stronger norms |||·|||. These algebras are ideals in B(H).
1.5. OPERATOR-VALUED MAPS 17

1.5 Operator-valued maps

In analogy with Section 1.2 it is natural to consider function with values in B(H). More
precisely, let J be an open interval on R, and let us consider a map F : J → B(H).
The notion of continuity can be considered with several topologies on B(H), but as in
Definition 1.3.5 we shall consider only three of them.

Definition 1.5.1. The map F is continuous in norm on J if for all t ∈ J

lim F (t + ε) − F (t) = 0.
ε→0

The map F is strongly continuous on J if for any f ∈ H and all t ∈ J

lim F (t + ε)f − F (t)f = 0.

ε→0

The map F is weakly continuous on J if for any f, g ∈ H and all t ∈ J

⟨ ( ) ⟩
lim g, F (t + ε) − F (t) f = 0.
ε→0

One writes respectively u − limε→0 F (t + ε) = F (t), s − limε→0 F (t + ε) = F (t) and

w − limε→0 F (t + ε) = F (t).

The same type of definition holds for the differentiability:

Definition 1.5.2. The map F is differentiable in norm on J if there exists a map

F ′ : J → B(H) such that

1( )
lim F (t + ε) − F (t) − F ′ (t) = 0.
ε→0 ε
The definitions for strongly differentiable and weakly differentiable are similar.
∫
If J is an open interval of R and if F : J → B(H), one defines J F (t) dt as a
Riemann integral (limit of finite sums over a partition of J) if this limiting procedure
exists and is independent of the partitions of J. Note that these integrals can be defined
in the weak topology, in the strong topology or in the norm topology ∫(and in other
topologies). For example,
∫ if F : J → B(H) is strongly continuous and if J ∥F (t)∥dt <
∞, then the integral J F (t) dt exists in the strong topology.

Proposition
∫ 1.5.3. Let J be an open interval of R and F : J → B(H) such that
J
F (t) dt exists (in an appropriate topology). Then,

(i) For any B ∈ B(H) one has

∫ ∫ (∫ ) ∫
B F (t) dt = BF (t) dt and F (t) dt B = F (t)B dt,
J J J J
18 CHAPTER 1. HILBERT SPACE AND BOUNDED LINEAR OPERATORS

∫ ∫
(ii) One has J
F (t) dt ≤ J
∥F (t)∥ dt,

(iii) If C ⊂ B(H) is a subalgebra of B(H), closed with respect to a norm |||·|||, and
∫ the map F : J → C∫ is continuous with respect to this norm and satisfies
if
J
|||F (t)||| dt < ∞, then J F (t) dt exists, belongs to C and satisfies
∫ ∫
F (t) dt ≤ |||F (t)||| dt.
J J

Note that the last statement is very useful, for example when C = K (H) or for
any Schatten class.
Chapter 2

Unbounded operators

In this chapter we define the notions of unbounded operators, their adjoint, their resol-
vent and their spectrum. Perturbation theory will also be considered. As in the previous
chapter, H denotes an arbitrary Hilbert space.

2.1 Unbounded, closed, and self-adjoint operators

In this section, we define an extension of the notion of bounded linear operators. Obvi-
ously, the following definitions and results are also valid for bounded linear operators.
( )
Definition 2.1.1. A linear operator on H is a pair A, D(A) , where D(A) is a linear
manifold of H and A is a linear
( map)from D(A) to H. D(A) is called the domain of A.
One says that the operator A, D(A) is densely defined if D(A) is dense in H.

Note that one often just says the linear operator A, but that its domain D(A) is
implicitly taken into account. For such an operator, its range Ran(A) is defined by

Ran(A) := A D(A) = {f ∈ H | f = Ag for some g ∈ D(A)}.

In addition, one defines the kernel Ker(A) of A by

Ker(A) := {f ∈ D(A) | Af = 0}.

Let us also stress that the sum A + B for two linear operators is a priori only
defined on the subspace D(A) ∩ D(B), and that the product AB is a priori defined only
on the subspace {f ∈ D(B) | Bf ∈ D(A)}. These two sets can be very small.

Example 2.1.2. Let H := L2 (R) and consider∫the operator X defined by [Xf ](x) =
xf (x) for any x ∈ R. Clearly, D(X) = {f ∈ H | R |xf (x)|2 dx < ∞} ( H. In addition,
by considering the family of functions {fy }y∈R ⊂ D(X) with fy (x) := 1 in x ∈ [y, y + 1]
and fy (x) = 0 if x ̸∈ [y, y +1], one easily observes that ∥fy ∥ = 1 but supy∈R ∥Xfy ∥ = ∞,
which can be compared with (1.8).

19
20 CHAPTER 2. UNBOUNDED OPERATORS

Clearly, a linear operator A can be defined on several domains. For example the
operator X of the previous example is well-defined on the Schwartz space S(R), or on
the set Cc (R) of continuous functions on R with compact support, or on the space D(X)
mentioned in the previous example. More generally, one has:
( ) ( )
Definition 2.1.3. For any pair of linear operators A, D(A) ( and B,)D(B) satisfying
( ⊂ D(B)
D(A) ) and Af = Bf for( all f ∈ )D(A), one says that B,
( D(B) )is an extension
of A, D(A) to D(B), or that A, D(A) is the restriction of B, D(B) to D(A).
( )
Let us now note that if A, D(A) is densely defined and if there exists c ∈ R such
that ∥Af ∥ ≤ c∥f ∥ for all f ∈ D(A), then there exists a natural continuous extension
A of A with D(A) = H. This extension satisfies A ∈ B(H) with ∥A∥ ≤ c, and is called
the closure of the operator A.
Exercise 2.1.4. Work on the details of this extension.
Let us now consider a similar construction but in the absence of a constant c ∈ R
such that ∥Af ∥ ≤ ( c∥f ∥ for
) all f ∈ D(A). More precisely, consider an arbitrary densely
defined operator A, D(A) . Then for any f ∈ H there exists a sequence {fn }n∈N ⊂ D(A)
strongly converging to f . Note that the sequence {Afn }n∈N will not be Cauchy in
general. However, let us assume that this sequence is strongly Cauchy, i.e. for any
ε > 0 there exists N ∈ N such that ∥Afn − Afm ∥ < ε for any n, m ≥ N . Since H is
complete, this Cauchy sequence has a limit, which we denote by h, and it would then
be natural to set Af = h. In short, one would have Af := s− limn→∞ Afn . It is easily
observed that this definition is meaningful if and only if by choosing a different sequence
{fn′ }n∈N ⊂ D(A) strongly convergent to f and also defining a Cauchy sequence {Afn′ }n∈N
then s− limn→∞ Afn′ = s− limn→∞ Afn . If this condition holds, then Af is well-defined.
Observe in addition that the previous equality can by rewritten as s− limn→∞ A(fn −
fn′ ) = 0, which leads naturally to the following definition.
( )
Definition 2.1.5. A linear operator A, D(A) is closable if for any sequence {fn }n∈N ⊂
D(A) satisfying s− limn→∞ fn = 0 and such that {Afn }n∈N is strongly Cauchy, then
s− limn→∞ Afn = 0.
As shown before this definition, in such a case one can define an extension A of
A with D(A) given by the sets of f ∈ H such that there exists {fn }n∈N ⊂ D(A) with
s− limn→∞ fn = f and such that {Afn }n∈N is strongly
( Cauchy.
) For such an element f
one sets Af = s− limn→∞ Afn , and the extension A, D(A) is called the closure of A.
In relation with the previous construction the following definition is now natural:
( )
Definition 2.1.6. An linear operator A, D(A) is closed if for any sequence {fn }n∈N ⊂
D(A) with s− limn→∞ fn = f ∈ H and such that {Afn }n∈N is strongly Cauchy, then one
has f ∈ D(A) and s− limn→∞ Afn = Af .
Exercise 2.1.7. Prove the following assertions:
(i) A bounded linear operator is always closed,
2.1. UNBOUNDED, CLOSED, AND SELF-ADJOINT OPERATORS 21
( ) ( )
(ii) If A, D(A) is closable and B ∈ B(H), then A + B, D(A) is closable,
( ) ( )
(iii) f A, D(A) is closed and B ∈ B(H), then A + B, D(A) is closed,
( ) ( ) ( )
(iv) If A, D(A) admits a closed extension B, D(B) , then A, D(A) is closable and
( ) ( )
its closure satisfies A, D(A) ⊂ B, D(B) .

( Let us) still introduce the notion of the graph of an operator: For any linear operator
A, D(A) one sets
Γ(A) := {(f, Af ) | f ∈ D(A)} ⊂ H ⊕ H (2.1)
and call it the graph of A. Note that the norm of an element (A, Af ) in H ⊕ H is given
by (∥f ∥2 + ∥Af( ∥ ) . It
2 1/2
) is also clear that Γ(A) is a linear manifold in H ⊕ H, and one
observes that A,(D(A) is) closed if and only if the graph Γ(A) is closed in H ⊕ H. On
the other hand, A, D(A) is closable if and only if the closure of its graph does not
contain any element of the form (0, h) with h ̸= 0. Let us finally observe that if one sets
for any f, g ∈ D(A)
⟨f, g⟩A := ⟨f, g⟩ + ⟨Af, Ag⟩, (2.2)
( )
and if A, D(A) is a closed operator, then the linear manifold D(A) equipped with the
scalar product (2.2) is a Hilbert space, with the natural norm deduced from this scalar
product.
Let us now come back to the notion of the adjoint of an operator. This concept is
slightly more subtle for unbounded operators than in the bounded case.
( )
Definition 2.1.8. Let A, D(A) be a densely defined linear operator on H. The adjoint
A∗ of A is the operator defined by
{ }
D(A∗ ) := f ∈ H | ∃f ∗ ∈ H with ⟨f ∗ , g⟩ = ⟨f, Ag⟩ for all g ∈ D(A)

and A∗ f := f ∗ for all f ∈ D(A∗ ).

Let us note that the density of D(A) is necessary to ensure that A∗ is well defined.
Indeed, if f1∗ , f2∗ satisfy for all g ∈ D(A)

⟨f1∗ , g⟩ = ⟨f, Ag⟩ = ⟨f2∗ , g⟩,

then ⟨f1∗ − f2∗ , g⟩ = 0 for all g ∈ D(A), ∗ ∗

) equality implies f1 = f2 only if D(A)
( ∗ and ∗this
is dense in H. Note also that once A , D(A ) is defined, one has

⟨A∗ f, g⟩ = ⟨f, Ag⟩ ∀f ∈ D(A∗ ) and ∀g ∈ D(A).

( )
Exercise 2.1.9. Show that if A, D(A) is closable, then D(A∗ ) is dense in H.

Some relations between A and its adjoint A∗ are gathered in the following lemma.
( )
Lemma 2.1.10. Let A, D(A) be a densely defined linear operator on H. Then
22 CHAPTER 2. UNBOUNDED OPERATORS
( )
(i) A∗ , D(A∗ ) is closed,
(ii) One has Ker(A∗ ) = Ran(A)⊥ ,
( ) ( ) ( ∗ ∗
)
(iii) If
( ∗ B, D(B)) is an extension of A, D(A) , then A , D(A ) is an extension of
∗
B , D(B ) .
Proof. a) Consider {fn }n∈N ⊂ D(A∗ ) such that s− limn→∞ fn = f ∈ H and s− limn→∞ A∗ fn =
h ∈ H. Then for each g ∈ D(A) one has
⟨f, Ag⟩ = lim ⟨fn , Ag⟩ = lim ⟨A∗ fn , g⟩ = ⟨h, g⟩.
n→∞ n→∞

Hence f ∈ D(A ) and A f = h, which proves that A∗ is closed.

∗ ∗

b) Let f ∈ Ker(A∗ ), i.e. f ∈ D(A∗ ) and A∗ f = 0. Then, for all g ∈ D(A), one has
0 = ⟨A∗ f, g⟩ = ⟨f, Ag⟩
meaning that f ∈ Ran(A)⊥ . Conversely, if f ∈ Ran(A)⊥ , then for all g ∈ D(A) one has
⟨f, Ag⟩ = 0 = ⟨0, g⟩
meaning that f ∈ D(A∗ ) and A∗ f = 0, by the definition of the adjoint of A.
c) (Consider) f ∈ D(B ∗ ) and observe
( that) ⟨B ∗ f, g⟩ = ⟨f, Bg⟩ for any g ∈ D(B).
Since B, D(B) is an extension of A, D(A) , one infers that ⟨B ∗ f, g⟩ = ⟨f, Ag⟩ for
any g ∈ D(A). Now, this equality means that f ∈ D(A∗ ) and that A∗ f = B ∗ f , or more
explicitly that A∗ is defined on the domain ∗
( ∗ of B∗ ) and coincide with (this operator
) on
∗ ∗
this domain. This means precisely that A , D(A ) is an extension of B , D(B ) .
Extension 2.1.11. Work on the additional properties of the adjoint operators as pre-
sented in Propositions 2.20 to 2.22 of [Amr].
Let us finally introduce the analogue of the bounded self-adjoint operators but in
the unbounded setting. These operators play a key role in quantum mechanics and their
study is very well developed.
( )
Definition 2.1.12. A densely defined linear operator A, D(A) is self-adjoint if
D(A∗ ) = D(A) and A∗ f = Af for all f ∈ D(A).
Note that as a consequence of Lemma 2.1.10.(i) a self-adjoint operator is always
closed. Recall also that in the bounded case, a self-adjoint operator was characterized
by the equality
⟨Af, g⟩ = ⟨f, Ag⟩ (2.3)
for any f, g ∈ H. In the unbounded case, such an equality still holds if f, g ∈ D(A).
However, let us emphasize that (2.3) does not ( completely
) characterize a self-adjoint
operator. In fact, a densely defined operator A, D(A) satisfying (2.3) is called a sym-
metric operator, and self-adjoint operators are special instances of symmetric operators
(but not all symmetric
( ∗ operators
) are self-adjoint).(In fact, for
) a symmetric operator the
∗
adjoint operator A , D(A )( is an extension
) of A, D(A) , but the equality of these
two operators holds only if A, D(A) is self-adjoint. Note also that for any symmetric
operator the scalar ⟨f, Af ⟩ is real for any f ∈ D(A).
2.2. RESOLVENT AND SPECTRUM 23

Exercise 2.1.13. Show that a symmetric operator is always closable.

Let us add one more definition related to self-adjoint operators.

( )
Definition 2.1.14. A symmetric operator A, D(A) is essentially self-adjoint if its
( )
closure A, D(A) is self-adjoint. In this case D(A) is called a core for A.

A following fundamental criterion for self-adjointness is important in this context,

and its proof can be found in [Amr, Prop. 3.3].
( )
Proposition 2.1.15. Let A, D(A) be a symmetric operator in a Hilbert space H.
Then
( )
(i) A, D(A) is self-adjoint if and only if Ran(A + i) = H and Ran(A − i) = H,
( )
(ii) A, D(A) is essentially self-adjoint if and only if Ran(A + i) and Ran(A − i) are
dense in H.

We still mention that the general theory of extensions of symmetric operators is

rather rich and can be studied on its own.

Extension 2.1.16. Study the theory of extensions of symmetric operators (there exist
several approaches for this study).

2.2 Resolvent and spectrum

We come now to the important notion of the spectrum of an operator. As already
mentioned in the previous section we shall often speak about a linear operator A, its
domain D(A) being implicitly taken into account. Recall also that the notion of a closed
linear operator has been introduced in Definition 2.1.6.
The notion of the inverse of a bounded linear operator has already been intro-
duced in Definition 1.3.7. By analogy we say that any linear operator A is invertible if
Ker(A) = {0}. In this case, the inverse A−1 gives a bijection from Ran(A) onto D(A).
More precisely D(A−1 ) = Ran(A) and Ran(A−1 ) = D(A). It can then be checked that
if A is closed and invertible, then A−1 is also closed. Note also if A is closed and if
Ran(A) = H then A−1 ∈ B(H). In fact, the boundedness of A−1 is a consequence of
the closed graph theorem1 and one says in this case that A is boundedly invertible or
invertible in B(H).

Definition 2.2.1. For a closed linear operator A its resolvent set ρ(A) is defined by
{ }
ρ(A) := z ∈ C | (A − z) is invertible in B(H)
{ }
= z ∈ C | Ker(A − z) = {0} and Ran(A − z) = H .
(
1
Closed graph theorem: If B, H) is a closed operator, then B ∈ B(H), see for example [Kat,
Sec. III.5.4]. This can be studied as an Extension.
24 CHAPTER 2. UNBOUNDED OPERATORS

For z ∈ ρ(A) the operator (A − z)−1 ∈ B(H) is called the resolvent of A at the point z.
The spectrum σ(A) of A is defined as the complement of ρ(A) in C, i.e.

σ(A) := C \ ρ(A). (2.4)

The following statement summarized several properties of the resolvent set and of
the resolvent of a closed linear operator.

Proposition 2.2.2. Let A be a closed linear operator on a Hilbert space H. Then

(i) The resolvent set ρ(A) is an open subset of C,

(ii) If z1 , z2 ∈ ρ(A) then the first resolvent equation holds, namely

(A − z1 )−1 − (A − z2 )−1 = (z1 − z2 )(A − z1 )−1 (A − z2 )−1 (2.5)

(iii) If z1 , z2 ∈ ρ(A) then the operators (A − z1 )−1 and (A − z2 )−1 commute,

(iv) In each connected component of ρ(A) the map z 7→ (A − z)−1 is holomorphic.

Exercise 2.2.3. Provide the proof of the previous proposition.

As a consequence of the previous proposition, the spectrum of a closed linear opera-

tor is always closed. In particular, z ∈ σ(A) if A−z is not invertible or if Ran(A−z) ̸= H.
The first situation corresponds to the definition of an eigenvalue:

Definition 2.2.4. For a closed linear operator A, a value z ∈ C is an eigenvalue of

A if there exists f ∈ D(A), f ̸= 0, such that Af = zf . In such a case, the element f
is called an eigenfunction of A associated with the eigenvalue z. The dimension of the
vector space generated by all eigenfunctions associated with an eigenvalue z is called
the multiplicity of z. The set of all eigenvalues of A is denoted by σp (A).

Let us still provide two properties of the spectrum of an operator in the special
cases of a bounded operator or of a self-adjoint operator.

Exercise 2.2.5. By using the Neumann series, show that for any B ∈ B(H) its spec-
trum is contained in the ball in the complex plane of center 0 and of radius ∥B∥.

Lemma 2.2.6. Let A be a self-adjoint operator in H.

(i) Any eigenvalue of A is real,

(ii) More generally, the spectrum of A is real, i.e. σ(A) ⊂ R,

(iii) Eigenvectors associated with different eigenvalues are orthogonal to one another.
2.3. PERTURBATION THEORY FOR SELF-ADJOINT OPERATORS 25

Proof. a) Assume that there exists z ∈ C and f ∈ D(A), f ̸= 0 such that Af = zf .

Then one has

z∥f ∥2 = ⟨f, zf ⟩ = ⟨f, Af ⟩ = ⟨Af, f ⟩ = ⟨zf, f ⟩ = z∥f ∥2 .

Since ∥f ∥ ̸= 0, one deduces that z ∈ R.

b) Let us consider z = λ + iε with λ, ε ∈ R and ε ̸= 0, and show that z ∈ ρ(A).
Indeed, for any f ∈ D(A) one has

∥(A − z)f ∥2 = ∥(A − λ)f − iεf ∥2

⟨ ⟩
= (A − λ)f − iεf, (A − λ)f − iεf
= ∥(A − λ)f ∥2 + ε2 ∥f ∥2 .

It follows that ∥(A − z)f ∥ ≥ |ε|∥f ∥, and thus A − z is invertible.

Now, for any for any g ∈ Ran(A − z) let us observe that

∥g∥ = (A − z)(A − z)−1 g ≥ |ε| (A − z)−1 g .

Equivalently, it means for all g ∈ Ran(A − z), one has

1
(A − z)−1 g ≤ ∥g∥. (2.6)
|ε|

Let us finally observe that Ran(A − z) is dense in H. Indeed, by Lemma 2.1.10 one
has
( )
Ran(A − z)⊥ = Ker (A − z)∗ = Ker(A∗ − z) = Ker(A − z) = {0}
since all eigenvalues of A are real. Thus, the operator (A − z)−1 is defined on the
dense domain Ran(A − z) and satisfies the estimate (2.6). As explained just before the
Exercise 2.1.4, it means that (A − z)−1 continuously extends to an element of B(H),
and therefore z ∈ ρ(A).
c) Assume that Af = λf and that Ag = µg with λ, µ ∈ R and λ ̸= µ, and
f, g ∈ D(A), with f ̸= 0 and g ̸= 0. Then

λ⟨f, g⟩ = ⟨Af, g⟩ = ⟨f, Ag⟩ = µ⟨f, g⟩,

which implies that ⟨f, g⟩ = 0, or in other words that f and g are orthogonal.

2.3 Perturbation theory for self-adjoint operators

( )
Let A, D(A) be a self-adjoint operator in a Hilbert space H. The invariance of the
self-adjoint property under the addition of a linear operator B is an important issue,
especially in relation with quantum mechanics.
26 CHAPTER 2. UNBOUNDED OPERATORS
( ) ( )
First of all, observe that if A, D(A) and B, D(B) are symmetric operator in H
and if D(A) ∩ D(B) is dense in H, then A + B defined on this intersection is still a
symmetric operator. Indeed, one has for any f, g ∈ D(A) ∩ D(B)

⟨(A + B)f, g⟩ = ⟨Af, g⟩ + ⟨Bf, g⟩ = ⟨f, Ag⟩ + ⟨f, Bg⟩ = ⟨f, (A + B)g⟩.

However, even if both operators are self-adjoint, their sum is not self-adjoint in general.
In the sequel we present some situations where the self-adjointness property
( is )preserved.
In the simplest case, if B ∈ B(H) and( B is self-adjoint,
) and if A, D(A) is a self-
adjoint operator one easily observes that A + B, D(A) ( is still a)self-adjoint operator.
∗
Indeed, this will automatically ) one shows∗ that D (A+B) ⊂ D(A+B) = D(A).
( follow∗if
For that purpose, let f ∈ D (A + B) and let f ∈ H such that ⟨f, (A + B)g⟩ = ⟨f ∗ , g⟩
for any g ∈ D(A). Then one has

⟨f ∗ , g⟩ = ⟨f, (A + B)g⟩ = ⟨f, Ag⟩ + ⟨f, Bg⟩ = ⟨f, Ag⟩ + ⟨Bf, g⟩.

One thus infers that ⟨f, Ag⟩ = ⟨f ∗ − Bf, g⟩ for any g ∈ D(A). Since f ∗ − Bf ∈ H this
means that f ∈ D(A∗ ) ≡ D(A).
Now, if B is not bounded, the question is more subtle. First of all, a notion of
smallness of B with respect to A has to be introduced.
( ) ( )
Definition 2.3.1. Let A, D(A) and B, D(B) be two linear operators in H. The
operator B is said to be A-bounded or relatively bounded with respect to A if D(A) ⊂
D(B) and if there exists α, β ≥ 0 such that for any f ∈ D(A)

∥Bf ∥ ≤ α∥Af ∥ + β∥f ∥. (2.7)

The infimum of all α satisfying this inequality is called the A-bound of B with respect
to A, or the relative bound of B with respect to A.
Clearly, if B ∈ B(H), then B is A-bounded with relative bound 0. In fact, such a
A-bound can be 0 even if B is not bounded. More precisely, the following situation can
take place: For any ε > 0 there exists β = β(ε) such that

∥Bf ∥ ≤ ε∥Af ∥ + β∥f ∥.

In such a case, the A-bound of B with respect to A is 0, but obviously one must have
β(ε) → ∞ as ε → 0.
( ) ( )
Exercise 2.3.2. Consider two linear operators A, D(A) and B, D(B) with D(A) ⊂
D(B), and assume that A is self-adjoint.
(i) Let also a, b ≥ 0. Show that

∥Bf ∥2 ≤ a2 ∥Af ∥2 + b2 ∥f ∥2 ∀f ∈ D(A)

is equivalent to ( )−1
B A ± i ab ≤ a,
2.3. PERTURBATION THEORY FOR SELF-ADJOINT OPERATORS 27

(ii) Show that if B(A − z)−1 ∈ B(H) for some z ∈ ρ(A), then B is A-bounded,
(iii) Show that if B is A-bounded, then B(A − z)−1 ∈ B(H) for any z ∈ ρ(A). In
addition, if α denotes the A-bound of B, then
α = inf ∥B(A − z)−1 ∥ = inf ∥B(A − iκ)−1 ∥ = inf ∥B(A + iκ)−1 ∥. (2.8)
z∈ρ(A) κ>0 κ>0

Let us still mention that a self-adjoint operator A is said to be lower semibounded

or bounded from below if there exists λ ∈ R such that (−∞, λ) ⊂ ρ(A), or in other
words if A has no spectrum below the value λ. Similarly, A is upper semibounded if
there exists λ′ ∈ R such that A has no spectrum above λ′ .
( )
Theorem
( 2.3.3
) (Rellich-Kato theorem). Let A, D(A) be a self-adjoint operator and
let B, D(B) be a A-bounded symmetric operator with A-bound α < 1. Then
(i) The operator A + B is self-adjoint on D(A),
(ii) B is also (A + B)-bounded,
(iii) If A is semibounded, then A + B is also semibounded.
The proof of this theorem is left as an exercise since some preliminary work is
necessary, see [Amr, Prop. 2.44] for the proof. Also one notion from the functional
calculus for self-adjoint operator is used in the proof, and this concept will only be
studied later on.
The following statement is often called the second resolvent equation.
( ) ( )
Proposition 2.3.4. Let A, D(A) be a self-adjoint operator and let B, D(B) be a
symmetric operator which is A-bounded with A-bound α < 1. Then the following equality
holds for any z ∈ ρ(A) ∩ ρ(A + B):
(A − z)−1 − (A + B − z)−1 = (A − z)−1 B(A + B − z)−1 (2.9)
= (A + B − z)−1 B(A − z)−1 . (2.10)
Proof. By assumption, both operators (A − z)−1 and (A + B − z)−1 belong to B(H).
In addition, B(A + B − z)−1 and B(A − z)−1 also belong to B(H), as a consequence
of the previous theorem and of Exercise 2.3.2. Since (A + B − z)−1 maps H onto D(A)
one has
B(A + B − z)−1 = (A + B − z)(A + B − z)−1 − (A − z)(A + B − z)−1
= 1 − (A − z)(A + B − z)−1 .
By multiplying this equality on the left by (A − z)−1 one infers the first equality of the
statement.
For the second equality, one starts with the equality
B(A − z)−1 = (A + B − z)(A − z)−1 − (A − z)(A − z)−1
= (A + B − z)(A − z)−1 − 1
and multiply it on the left by (A + B − z)−1 .
28 CHAPTER 2. UNBOUNDED OPERATORS

( Let us) close this section with the notion of A-compact

( ) operator. More precisely, let
A, D(A) be a closed linear operator, and let B, D(B) be a second operator. Then
B is A-compact if D(A) ⊂ D(B) and if there exists z ∈ ρ(A) such that B(A − z)−1
belongs to K (H). In such a case, the operator B is A-bounded with A-bound equal to
0, as shown in the next statement.
( )
Proposition 2.3.5. Let A, D(A) be self-adjoint, and let B be a symmetric operator
which is A-compact. Then

(i) B(A − z)−1 ∈ K (H) for any z ∈ ρ(A),

(ii) B is A-bounded with A-bound equal to 0,

( )
(iii) if C, D(C) is symmetric and A-bounded with A-bound α < 1, then B is also
(A + C)-compact.

Proof. a) By assumption, there exists z0 ∈ ρ(A) with B(A − z0 )−1 ∈ K (H). By the
first resolvent equation (2.5) one then infers that
( )
B(A − z)−1 = B(A − z0 )−1 1 + (z − z0 )(A − z)−1 .
( )
Since B(A − z0 )−1 is compact and 1 + (z − z0 )(A − z)−1 is bounded, one deduces from
Proposition 1.4.11.(iii) that B(A − z)−1 is compact as well.
b) By (2.8) it is sufficient to show that for any ε > 0 there exists z ∈ ρ(A) such
that ∥B(A − z)−1 ∥ ≤ ε. For that purpose, we shall consider z = iµ and show that
∥B(A − iµ)−1 ∥ → 0 as µ → ∞. Indeed observe first that
( )
B(A − iµ)−1 = B(A − i)−1 (A − i)(A − iµ)−1
( )
= B(A − i)−1 1 + i(µ − 1)(A − iµ)−1 .

As a consequence of (2.6) one then deduces that (A − i)(A − iµ)−1 ∈ B(H) with

µ−1
∥(A − i)(A − iµ)−1 ∥ ≤ 1 + ≤2 for any µ ≥ 1. (2.11)
µ
( )∗
If one shows that (A − i)(A − iµ)−1 converges strongly to 0 as µ → ∞ then one gets
from Proposition 1.4.12.(ii) and from the compactness of B(A − i)−1 (as a consequence
of the point (i) of this statement) that limµ→∞ ∥B(A − iµ)−1 ∥ = 0.
One easily observes that for any f ∈ D(A) one has
( )∗
(A − i)(A − iµ)−1 f = (A + iµ)−1 (A + i)f

and that
( 1
∥ A + iµ)−1 (A + i)f ∥ ≤ ∥(A + iµ)−1 ∥ ∥(A + i)f ∥ ≤ ∥(A + i)f ∥ (2.12)
µ
2.3. PERTURBATION THEORY FOR SELF-ADJOINT OPERATORS 29

where (2.6) has been used once again. Clearly, the r.h.s. of (2.12) converges to 0 as
µ → ∞, and by the upper bound obtained in (2.11) one deduces by density that
( )∗
s− lim (A − i)(A − iµ)−1 = 0,
µ→0

as expected.
c) By the second resolvent equation (2.9) one gets

B(A + C − i)−1 = B(A − i)−1 − B(A − i)−1 C(A + C − i)−1

( )
= B(A − i)−1 1 − C(A + C − i)−1 .

Then B(A−i)−1 is compact, by the point (i), and C(A+C −i)−1 is bounded by Theorem
2.3.3.(ii) and by Exercise (2.3.2).(iii). As in (a) one deduces that B(A + C − i)−1 is a
compact operator.
30 CHAPTER 2. UNBOUNDED OPERATORS
Chapter 3

Examples

In this chapter we present some examples of operators which often appear in the liter-
ature. Most of them are self-adjoint and were first introduced in relation with quantum
mechanics. Indeed, any physical system is described with such an operator. Self-adjoint
operators are the natural generalization of Hermitian matrices. Obviously, the following
list of examples is only very partial, and many other operators should be considered as
well.

3.1 Multiplication and convolution operators

In this section, we introduce two natural classes of operators on Rd . This material is
standard and can be found for example in the books [Amr] and [Tes]. We start by
considering multiplication operators on the Hilbert space L2 (Rd ).
For any measurable complex function φ on Rd let us define the multiplication op-
erator φ(X) on H(Rd ) := L2 (Rd ) by

[φ(X)f ](x) = φ(x)f (x) ∀x ∈ Rd

for any ∫
( ) { }
f ∈ D φ(X) := g ∈ H(Rd ) | |φ(x)|2 |g(x)|2 dx < ∞ .
Rd

Clearly, the properties of this operator depend on the function φ. More precisely:

Lemma
( 3.1.1.
) Let φ(X) be the multiplication operator on H(Rd ). Then φ(X) belongs
to B H(Rd ) if and only if φ ∈ L∞ (Rd ), and in this case ∥φ(X)∥ = ∥φ∥∞ .

Proof. One has

∫ ∫
∥φ(X)f ∥ =
2
|φ(x)| |f (x)| dx ≤
2 2
∥φ∥2∞ |f (x)|2 dx = ∥φ∥2∞ ∥f ∥2 .
Rd Rd
( )
Thus, if φ ∈ L∞ (Rd ), then D φ(X) = H(Rd ) and ∥φ(X)∥ ≤ ∥φ∥∞ .

31
32 CHAPTER 3. EXAMPLES

Now, assume that φ ̸∈ L∞ (Rd ). It means that for any n ∈ N there exists a measur-
able set Wn ⊂ Rd with 0 < |Wn | < ∞ such that |φ(x)| > n for any x ∈ Wn . We then
set fn = χWn and observe that fn ∈ H(Rd ) with ∥fn ∥2 = |Wn | and that
∫ ∫
∥φ(X)fn ∥ =
2
|φ(x)| |fn (x)| dx =
2 2
|φ(x)|2 dx > n2 ∥fn ∥2 ,
Rd Wn

from which one infers that ∥φ(X)fn ∥/∥fn ∥ > n. Since n is arbitrary, the operator φ(X)
can not be bounded.
Let us finally show that if φ ∈ L∞ (Rd ), then ∥φ(X)∥ ≥ ∥φ∥∞ . Indeed, for any ε > 0,
there exists a measurable set Wε ⊂ Rd with 0 < |Wε | < ∞ such that |φ(x)| > ∥φ∥∞ − ε
for any x ∈ Wε . Again by setting fε = χWε one gets that ∥φ(X)fε ∥/∥fε ∥ > ∥φ∥∞ − ε,
from which one deduces the required inequality.
If φ ∈ L∞ (Rd ), one easily observes that φ(X)∗ = φ(X), and thus φ(X) is self-
adjoint if and only if φ is a real function. The operator φ(X) is a projection if and only
if φ(x) ∈ {0, 1} for almost every x ∈ Rd . Similarly, the operator φ(X) is unitary if and
only if |φ(x)| = 1 for almost every x ∈ Rd . Observe also that φ(X) is a partial isometry
if and only if |φ(x)| ∈ {0, 1} for almost every x ∈ Rd . However, since φ(X) and φ(X)
commute, it is impossible to obtain φ(X)∗ φ(X) = 1 without getting automatically that
φ(X) is a unitary operator. In other words, there does not exist any isometry φ(X)
which is not unitary. ( ( ))
If φ is real but does not belong to L∞ (Rd ), one can show that φ(X), D φ(X)
defines a self-adjoint operator in H(Rd ), see also [Ped, Example 5.1.15]. In particular,
if φ ∈ C(Rd ) or if |φ| is polynomially bounded, then the mentioned operator is self-
adjoint, see [Amr, Prop. 2.29]. For example, for any j ∈ {1, . . . , d} the operator Xj
defined by [Xj f ](x) = xj f (x) is a self-adjoint operator with domain D(Xj ). Note that
the d-tuple (X1 , . . . , Xd ) is often referred to as the position operator in H(Rd ). More
generally, for any α ∈ Nd one also sets
X α = X1α1 . . . Xdαd
and this expression defines a self-adjoint operator on its natural domain. Other useful
multiplication operators are defined for any s > 0 by the functions
( ∑
d )s/2
R ∋ x 7→ ⟨x⟩ := 1 +
d s
x2j ∈ R.
j=1
( )
The corresponding operators ⟨X⟩s , Hs (Rd ) , with
{ } { ∫ }
Hs (R ) := f ∈ H(R ) | ⟨X⟩ f ∈ H(R ) = f ∈ H(R ) |
d d s d d
⟨x⟩2s |f (x)|2 dx < ∞ ,
Rd

are again self-adjoint operators on H(Rd ). Note that one usually calls Hs (Rd ) the
weighted Hilbert space
∫ with weight 2ss since it is naturally a Hilbert space with the scalar
product ⟨f, g⟩s := Rd f (x)g(x)⟨x⟩ dx.
3.1. MULTIPLICATION AND CONVOLUTION OPERATORS 33

Exercise 3.1.2. For any real φ ∈ C(Rd ) or φ ∈ L∞ (Rd ), show that the spectrum of
the self-adjoint multiplication operator φ(X) coincides with the closure of φ(Rd ) in R.
We shall now introduce a new type of operators on H(Rd ), but for that purpose
we need to recall a few results about the usual Fourier transform on Rd . The Fourier
transform F is defined on any f ∈ Cc (Rd ) by the formula1
∫
1
ˆ
[Ff ](ξ) ≡ f (ξ) := d/2
e−iξ·x f (x) dx. (3.1)
(2π) Rd

This linear transform maps the Schwartz space S(Rd∫) onto itself, and its inverse is
provided by the formula [F −1 f ](x) ≡ fˇ(x) := (2π)1d/2 Rd eiξ·x f (ξ) dξ. In addition, by
∫ ∫
taking Parseval’s identity Rd |f (x)|2 dx = Rd |fˆ(ξ)|2 dξ into account, one obtains that
the Fourier transform extends continuously to a unitary map on H(Rd ). We shall keep
the same notation F for this continuous extension, but one must be aware that (3.1) is
valid only on a restricted set of functions.
Let us use again the multi-index notation and set for any α ∈ Nd

(−i∂)α = (−i∂1 )α1 . . . (−i∂d )αd = (−i)|α| ∂1α1 . . . ∂dαd

with |α| = α1 + · · · + αd . With this notation at hand, the following relations hold for
any f ∈ S(Rd ) and any α ∈ Nd :

F(−i∂)α f = X α Ff,

or equivalently (−i∂)α f = F ∗ X α Ff . Keeping these relations in mind, one defines for

any j ∈ {1, . . . , d} the self-adjoint operator Dj := F ∗ Xj F with domain F ∗ D(Xj ).
Similarly, for any s > 0, one also defines the operator ⟨D⟩s := F ∗ ⟨X⟩s F with domain
{ } { }
Hs (Rd ) := f ∈ H(Rd ) | ⟨X⟩s Ff ∈ H(Rd ) ≡ f ∈ H(Rd ) | ⟨X⟩s fˆ ∈ H(Rd ) . (3.2)

Note that the space Hs (Rd ) is called the Sobolev space of order s, and (D1 , . . . , Dd ) is
usually called the momentum operator 2 .
We can now introduce the usual Laplace operator −∆ acting on any f ∈ S(Rd ) as

∑
d ∑
d ∑
d
−∆f = − ∂j2 f = (−i∂j )2 f = Dj2 f. (3.3)
j=1 j=1 j=1
( )
This operator admits a self-adjoint extension with domain H2 (Rd ), i.e. − ∆, H2 (Rd )
is a self-adjoint operator in H(Rd ). However, let us stress that the expression (3.3) is not
valid (pointwise) on all the elements of the domain H2 (Rd ). On the other hand, one has
1
Even if the group Rd is identified with its dual group, we will keep the notation ξ for points of its
dual group.
2
In physics textbooks, the position operator is often denoted by (Q1 , . . . , Qd ) while (P1 , . . . , Pd ) is
used for the momentum operator.
34 CHAPTER 3. EXAMPLES
∑
−∆ = F ∗ X 2 F, with X 2 = dj=1 Xj2 , from which one easily infers that σ(−∆) = [0, ∞).
Indeed, this follows from the content of Exercise 3.1.2 together with the invariance of
the spectrum through the conjugation by a unitary operator.
Before going on with other operators of the form φ(D), let us provide some addi-
tional information on the space H2 (Rd ) for d ∈ {1, 2, 3}.
Lemma 3.1.3. Let d ≤ 3 and assume that f ∈ H2 (Rd ). Then f ∈ C0 (Rd ), and for any
α > 0 there exists β > 0 such that
∥f ∥∞ ≤ α∥∆f ∥ + β∥f ∥, f ∈ H2 (Rd ).
In the following proof we shall denote by ∥g∥1 the L1 -norm of g ∈ L1 (Rd ), i.e.
∫
∥g∥1 = |g(x)|dx. (3.4)
Rd

Proof. For any γ > 0 let us set gγ : R → R defined by gγ (ξ) := (ξ 2 + γ 2 )−1 . The key
3

observation is that gγ belongs to L2 (Rd ) if d ≤ 3. Then, for f ∈ H2 (Rd ) one has

[ ]
(X 2 + γ 2 )fˆ = (X 2 + γ 2 )⟨X⟩−2 ⟨X⟩2 fˆ ∈ L2 (Rd )

by (3.2), and one infers by the Cauchy-Schwarz inequality that

∫
∥fˆ∥1 = (ξ 2 + γ 2 )−1 (ξ 2 + γ 2 )fˆ(ξ) dξ ≤ ∥gγ ∥ ∥(X 2 + γ 2 )fˆ∥ < ∞
Rd

which implies that fˆ belongs to L1 (Rd ). By the Riemann-Lebesgue lemma (as presented
for example in [Tes, Lem. 7.6]) one deduces that f ∈ C0 (Rd ), and more precisely that
∥f ∥∞ ≤ (2π)−d/2 ∥gγ ∥ ∥(X 2 + γ 2 )fˆ∥
( )
≤ (2π)−d/2 γ −2+d/2 ∥g1 ∥ ∥∆f ∥ + γ 2 ∥f ∥
( )
= (2π)−d/2 ∥g1 ∥ γ −2+d/2 ∥∆f ∥ + γ d/2 ∥f ∥ .

For any
( measurable
) { function φ on R(d let )} us now set φ(D) := F ∗ φ(X)F, with
ˆ
domain D φ(D) = f ∈ H(R ) | f ∈ D φ(X) , and as before this operator is self-
d

adjoint in H(Rd ), as for example for a continuous function φ or for a polynomially

bounded function φ. Then, if one defines the convolution of two (suitable) functions on
Rd by the formula ∫
1
[k ∗ f ](x) = k(y)f (x − y) dy
(2π)d/2 Rd
and if one takes the equality ǧ ∗ f = F ∗ (g fˆ) into account, one infers that the operator
φ(D) corresponds to a convolution operator, namely
φ(D)f = φ̌ ∗ f. (3.5)
Obviously, the meaning of such an equality depends on the class of functions f and g
considered.
3.2. SCHRÖDINGER OPERATORS 35

Exercise 3.1.4. Show that the following relations hold on the Schwartz space S(Rd ):
[iXj , Xk ] = 0 = [Dj , Dk ] for any j, k ∈ {1, . . . , d} while [iDj , Xk ] = 1δjk .

3.1.1 The harmonic oscillator

An example of an operator which can be expressed easily in terms of the families of
operators {Xj } and {Dj } is the harmonic oscillator, namely the operator

∑
d ∑
d ∑
d
( )
H = −∆ + ω X =2 2
Dj2 +ω 2
Xj2 = Dj2 + ω 2 Xj2 ,
j=1 j=1 j=1

where ω is a strictly positive constant. This operator can be defined on several domain,
as for example on Cc∞ (Rd ) or on the Schwartz space S(Rd ). An other domain which is
quite convenient is the following:
( )
D(H) := Vect Rd ∋ x 7→ xα e−x /2 ∈ R | α ∈ Nd ⊂ L2 (Rd ).
2

It is easily observed that the operator H is symmetric on D(H), and it can be

shown that this operator is in fact essentially self-adjoint on the domain D(H), see
Definition 2.1.14 for the notion of essential self-adjointness. In addition, H can be
completely studied by some algebraic methods, by considering the so-called creation
and annihilation operators. Let us simply mention that

σ(H) = {(2n + d)ω | n ∈ N}

and that the corresponding eigenfunctions can be expressed in terms of the Hermite
polynomials.
Extension 3.1.5. Work on the details of the algebraic methods for the harmonic os-
cillator. In particular, describe the eigenvalues, the corresponding eigenfunctions and
determine the multiplicity of each eigenvalue.

3.2 Schrödinger operators

In this section, we introduce some well-studied operators
First of all, let h : Rd → R be a continuous real function which diverges at infinity.
Equivalently, we assume that h satisfies (h − z)−1 ∈ C0 (Rd ) for some z ∈ C \ R.
The corresponding convolution ( operator
) h(D), defined by F ∗ h(X)F, is a self-adjoint
operator with domain F ∗ D h(X) . Clearly, the spectrum of such an operator is equal
to the closure of h(Rd ) in R.
Some examples of such a function h which are often considered√ in the literature
are the functions defined by h(ξ) = ξ , h(ξ) = |ξ| or h(ξ) = 1 + ξ 2 − 1. In these
2

cases, the operator h(D) = −∆ corresponds to the free Laplace operator, the operator
h(D) = |D| is the relativistic Schrödinger operator without mass, while the operator
36 CHAPTER 3. EXAMPLES
√
h(D) = 1 − ∆ − 1 corresponds ( to
) the relativistic Schrödinger
( ) operator with mass. In
these three cases, one has σ h(D) = [0, ∞) while σp h(D) = ∅.
Let us now perturb this operator h(D) with a multiplication operator V (X). If
the measurable function V : Rd → R is not essentially bounded, then the operator
h(D) + V (X) can only be defined on the intersection of the two domains, and checking
that there exists a self-adjoint extension of this operator is not always an easy task. On
the other hand, if one assumes that V ∈ L∞ (Rd ), then we can define the operator
( )
H := h(D) + V (X) with D(H) = D h(D) (3.1)

and this operator is self-adjoint. A lot of investigations have been performed on such
an operator H when V vanishes at infinity, in a suitable sense. On the other hand,
much less is known on this operator when the multiplication operator V (X), also called
the potential, has an anisotropic behavior. Let us just mention that a C ∗ -algebraic
framework has been developed for dealing with this anisotropic behavior and that it
involves crossed product C ∗ -algebras.

3.2.1 The hydrogen atom

Let us briefly introduce the operator used for describing a simple model of an atom
with a single electron in R3 . It is assumed that the nucleus of the atom is fixed at the
origin, and the electron moves in the external potential generated by the nucleus. If the
electrostatic force is taken into account, the resulting operator has the form
γ
H = −∆ − (3.2)
|X|

where −∆ is the Laplace operator introduced in (3.3), γ > 0 is called the coupling
1
constant, and |X| is the operator of multiplication by the function R3 ∋ x 7→ |x| 1
∈ R.
A priori, the operator exhibited in (3.2) is well-defined only on D(−∆) ∩ D(|X|−1 ).
However, it follows from Lemma 3.1.3 that D(−∆) ⊂ D(|X|−1 ). Indeed, since any
f ∈ D(−∆) also belongs to C0 (R3 ) one has (B1 (0) denotes the open ball centered at 0
and of radius 1)
∫ ∫ ∫
1 2 1 2 1 2
f (x) dx = f (x) dx + f (x) dx
R3 |x| B1 (0) |x| R3 \B1 (0) |x|
∫ ∫
1
≤ ∥f ∥2∞ dx + |f (x)|2 dx
B1 (0) |x|2
R \B1 (0)
3

< ∞.

As a consequence, D(−∆) ∩ D(|X|−1 ) = D(−∆).

In order to check that the operator H is self-adjoint on D(−∆) = H2 (R3 ), let us
decompose the function x 7→ −γ |x| 1
into V1 + V2 with V1 (x) = −γ |x|
1
if |x| < 1 and 0
otherwise, and V2 (x) = −γ |x| if |x| ≥ 1 and 0 otherwise. Clearly, V2 defines a bounded
1
3.3. THE WEYL CALCULUS 37

self-adjoint operator, and adding it to −∆ will not cause any problem. For V1 , one can
observe that for any c > 0 the function

R3 × R3 ∋ (x, ξ) 7→ V1 (x)(ξ 2 + c)−1 ∈ R

belongs to L2 (R3 × R3 ). Since an operator with L2 -kernel corresponds to a Hilbert-

Schmidt operator, the operator V1 (X)(−∆ + c)−1 is Hilbert-Schmidt (see Extension
1.4.14) and thus is compact. As a consequence of Proposition 2.3.5 one deduces that
V1 (X) is a multiplication operator which is −∆-bounded with relative bound equal to
0. By the Rellich-Kato theorem (see Theorem 2.3.3) one infers that −∆ + V1 (X) is
self-adjoint on H2 (R3 ), and then that −∆ + V1 (X) + V2 (X) = −∆ − γ |X| 1
is self-adjoint
on H (R ).
2 3

Let us add a few information about the operator (3.2) and refer to [Tes, Chap. 10] for
more information. One has [0, ∞) ⊂ σ(H). In fact, [0, ∞) corresponds to the essential
spectrum of H, as defined in the next chapter. In addition, the operator H possesses an
infinite number of eigenvalues, which can be computed explicitly. More precisely, by a
decomposition of this operator into the spherical harmonics {Ylm } where l ∈ N, m ∈ Z
with |m| ≤ l, and by studying the resulting operator for each index l one can get that
{ ( )2 }
σp (H) = − γ
2(n+1)
|n∈N .

Each eigenvalue has a multiplicity (n + 1)2 , which means that there are (n + 1)2 linearly
( γ )2
independent functions in H2 (R3 ) satisfying Hf = − 2(n+1) f . These functions can be
expressed in terms of the Laguerre polynomials.

Extension 3.2.1. Work on the details of the Hydrogen atom, and in particular study
its eigenvalues and the corresponding eigenfunctions.

3.3 The Weyl calculus

In section 3.1 we have seen how to define multiplication operators φ(X) and convolution
operators φ(D) on the Hilbert space H := L2 (Rd ). A natural question is how to define
a more general operator f (X, D) on L2 (Rd ) for a function f : Rd × Rd → C.
This can be seen as the problem of constructing a functional calculus f 7→ f (X, D)
for the family X1 , . . . , Xd , D1 , . . . , Dd of 2d self-adjoint, non-commuting operators. One
also would like to define a multiplication (f, g) 7→ f ◦ g satisfying (f ◦ g)(X, D) =
f (X, D)g(X, D) as well as an involution f → f ◦ leading to f ◦ (X, D) = f (X, D)∗ . The
deviation of ◦ from pointwise multiplication is imputable to the fact that X and D do
not commute.
The solution of these problems is called the Weyl calculus, or simply the pseudod-
ifferential calculus. In order to define it, let us set Ξ := Rd × R̂d ∼ = Rd × Rd , which
corresponds to the direct product of a locally compact Abelian group G and of its dual
38 CHAPTER 3. EXAMPLES

group Ĝ. Elements of Ξ will be denoted by x = (x, ξ), y = (y, η) and z = (z, ζ). We also
set ( )
σ(x, y) := σ (x, ξ), (y, η) = y · ξ − x · η
for the standard symplectic form on Ξ. The prescription for f (X, D) ≡ Op(f ) with
f : Ξ → C is then defined for u ∈ H and x ∈ Rd by
∫ ∫ ( )
1 i(x−y)·η x+y
[Op(f )u](x) := e f , η u(y) dy dη, (3.1)
(2π)d Rd R̂d 2

the involution is f ◦ (x) := f (x) and the multiplication (called the Moyal product) is
∫ ∫
4d
(f ◦ g)(x) := e−2iσ(x−y,x−z) f (y) g(z) dy dz. (3.2)
(2π)2d Ξ Ξ
Obviously, these formulas must be taken with some care: for many symbols f and g
they need a suitable reinterpretation. Also, the normalization factors should always be
checked once again, since they mainly depend on the conventions of each author.
Exercise 3.3.1. Check that if f (x, ξ) = f (ξ) (f is independent of x), then Op(f ) =
f (D), while if f (x, ξ) = f (x) (f is independent of ξ), then Op(f ) = f (X).
Beside the encouraging results contained in the previous exercise, let us try to show
where all the above formulas come from. We consider the strongly continuous unitary
maps Rd ∋ x 7→ Ux ∈ U (H) and R̂d ∋ ξ 7→ Vξ := e−iX·ξ ∈ U (H), acting on H as

[Ux u](y) = u(y + x) and [Vξ u](y) = e−iy·ξ u(y), u ∈ H, y ∈ Rd .

These operators satisfy the Weyl form of the canonical commutation relations

Ux Vξ = e−ix·ξ Vξ Ux , x ∈ Rd , ξ ∈ R̂d , (3.3)

as well as the identities Ux Ux′ = Ux′ Ux and Vξ Vξ′ = Vξ′ Vξ for x, x′ ∈ Rd and ξ, ξ ′ ∈ R̂d .
These can be considered as a reformulation of the content of Exercise 3.1.4 in terms of
bounded operators.
A convenient way to condense the maps U and V in a single one is to define the
Schrödinger Weyl system {W (x, ξ) | x ∈ Rd , ξ ∈ R̂d } by

W (x) ≡ W (x, ξ) := e 2 x·ξ Ux Vξ = e− 2 x·ξ Vξ Ux ,

i i
(3.4)
i
which satisfies the relation W (x)W (y) = e 2 σ(x,y) W (x+y) for any x, y ∈ Ξ. This equality
encodes all the commutation relations between the basic operators X and D. Explicitly,
the action of W on u ∈ H is given by

[W (x, ξ)u](y) = e−i( 2 x+y)·ξ u(y + x),

1
x, y ∈ Rd , ξ ∈ R̂d . (3.5)

Now, recall that for a family of m commuting self-adjoint operators

∫ S1 , . . . , −it·S
Sm one
1 ˇ
usually defines a functional calculus by the formula f (S) := (2π)m/2 Rm f (t) e dt,
 1
3.4. SCHRÖDINGER OPERATORS WITH X2
-POTENTIAL 39

where t · S = t1 S1 + . . . + tm Sm and fˇ is the inverse Fourier transform of f , see

the next chapter for a simplified version of this formula. The prescription (3.1) can
be obtained by a similar computation. For that purpose, let us define the symplectic
Fourier transformation FΞ : S ′ (Ξ) → S ′ (Ξ) by
∫
1
(FΞ f )(x) := eiσ(x,y) f (y) dy.
(2π)d Ξ
Now, for any function f : Ξ → C belonging to the Schwartz space S(Ξ), we set
∫
1
Op(f ) := (F −1 f )(x) W (x) dx. (3.6)
(2π)d Ξ Ξ
By using (3.5), one gets formula (3.1). Then it is easy to verify that the relation
Op(f )Op(g) = Op(f ◦g) holds for f, g ∈ S(Ξ) if one uses the Moyal product introduced
in (3.2).
Exercise 3.3.2. Check that the above statements are correct, and in particular that the
normalization factors are suitably chosen.

1
3.4 Schrödinger operators with x2
-potential
In this section we consider various realizations of an operator on H := L2 (R+ ), and
study some properties of the resulting operators. Our aim is to emphasize the role played
by the realization, or in other words by the choice of the domain for this operator. As
mentioned at the end of the section, depending on the realization, this operator can
have zero eigenvalue, a finite number of eigenvalues, or even an infinite number of
eigenvalues. On the other hand, the rest of the spectrum is stable and corresponds to
the half-line [0, ∞).
For any α ∈ C we consider the differential expression
( 1)
Lα := −∂x2 + α − x−2
4
acting on distributions on R+ , and denote by Lmin
α and Lmax
α the corresponding minimal
and maximal operators associated with it in H, see [BDG, Sec. 4 & App. A] for details.
We simply recall from this reference that

D(Lαmax ) = {f ∈ H | Lα f ∈ H}
∞
and that D(Lmin
α ) is the closure of the restriction of Lα to Cc (R+ ). In fact, it can be
shown that the following relation holds:
( min )∗
Lα = Lmax
ᾱ .

Let us recall some additional results which have been obtained in [BDG, Sec. 4].
For that purpose, we say that f ∈ D(Lmin α ) around 0, (or, by an abuse of notation,
40 CHAPTER 3. EXAMPLES

∞
( )
f (x) ∈ D(Lmin
α ) around 0) if there exists ζ ∈ Cc [0, ∞[ with ζ = 1 around 0 such that
f ζ ∈ D(Lmin
α ). In addition, it turns out that it is useful to introduce a parameter m ∈ C
such that α = m2 , even though there are two m corresponding to a single α ̸= 0. In
other words, we shall consider from now on the operator
( 1)
Lm2 := −∂x2 + m2 − x−2 .
4
With this notation, if ℜ(m) ≥ 1 (we use the notation ℜ(m) for the real part of the
complex number m) then Lmin m2 = Lm2 , while if |ℜ(m)| < 1 then Lm2 ( Lm2 and
max min max

D(Lm2 ) is a closed subspace of codimension 2 of D(Lm2 ). More precisely, if |ℜ(m)| < 1

min max

m2 ) then there exist a, b ∈ C such that:

and if f ∈ D(Lmax

f (x) − ax1/2−m − bx1/2+m ∈ D(Lmin

m2 ) around 0 if m ̸= 0,
f (x) − ax1/2 ln(x) − bx1/2 ∈ D(Lmin
0 ) around 0.

In addition, the behavior of any function g ∈ D(Lminm2 ) is known, namely g ∈ H0 (R+ )

1

(the completion of Cc1 (R+ ) with the H1 -norm) and as x → 0:

( ) ( )
g(x) = o x3/2 and g ′ (x) = o x1/2 if m ̸= 0,
( 3/2 ) ′
( 1/2 )
g(x) = o x ln(x) and g (x) = o x ln(x) if m = 0.

3.4.1 Two families of Schrödinger operators

Let us first recall from [BDG, Def. 4.1] that for any m ∈ C with ℜ(m) > −1 the
operator Hm has been defined as the restriction of Lmax
m2 to the domain
{
D(Hm ) = f ∈ D(Lmaxm2 ) | for some c ∈ C,
}
f (x) − cx1/2+m ∈ D(Lm min
2 ) around 0 .

It is then proved in this reference that {Hm }ℜ(m)>−1 is a holomorphic family of closed
operators in H. In addition, if ℜ(m) ≥ 1, then

Hm = Lmin max
m2 = Lm2 .

For this reason, we shall concentrate on the case −1 < ℜ(m) < 1, considering a larger
family of operators.
For |ℜ(m)| < 1 and for any κ ∈ C ∪ {∞} we define a family of operators Hm,κ :
{
D(Hm,κ ) = f ∈ D(Lmax m2 ) | for some c ∈ C, (3.7)
( 1/2−m ) }
f (x) − c κx +x 1/2+m
∈ D(Lm2 ) around 0 ,
min
κ ̸= ∞;
{
D(Hm,∞ ) = f ∈ D(Lmax m2 ) | for some c ∈ C, (3.8)
}
f (x) − cx1/2−m ∈ D(Lmin
m2 ) around 0 .
 1
3.4. SCHRÖDINGER OPERATORS WITH X2
-POTENTIAL 41

For m = 0, we introduce an additional family of operators H0ν with ν ∈ C ∪ {∞} :

{
D(H0ν ) = f ∈ D(Lmax0 ) | for some c ∈ C, (3.9)
( 1/2 ) }
f (x) − c x ln(x) + νx1/2 ∈ D(Lmin 0 ) around 0 , ν≠ ∞;
{
D(H0∞ ) = f ∈ D(Lmax 0 ) | for some c ∈ C, (3.10)
}
f (x) − cx ∈ D(L0 ) around 0 .
1/2 min

The following properties of these families of operators are immediate:

Lemma 3.4.1. (i) For any |ℜ(m)| < 1 and any κ ∈ C ∪ {∞},

Hm,κ = H−m,κ−1 . (3.11)

(ii) The operator H0,κ does not depend on κ, and all these operators coincide with
H0∞ .

As a consequence of (ii), all the results about the case m = 0 will be formulated in
terms of the family H0ν .
Let us now derive an additional result for this family of operators. For its proof, we
recall that the Wronskian W (f, g) of two continuously differentiable functions f, g on
R+ is given by the expression

Wx (f, g) ≡ W (f, g)(x) := f (x)g ′ (x) − f ′ (x)g(x). (3.12)

Proposition 3.4.2. For any m ∈ C with |ℜ(m)| < 1 and for any κ, ν ∈ C ∪ {∞}, one
has
(Hm,κ )∗ = Hm̄,κ̄ and (H0ν )∗ = H0ν̄ (3.13)
with the convention that ∞
¯ = ∞.

Proof. Recall from [BDG, App. A] that for any f ∈ D(Lmax m2 ) and g ∈ D(Lm̄2 ), the
max

functions f, f ′ , g, g ′ are continuous on R+ , and that the equality

⟨Lmax ¯
m2 f, g⟩ − ⟨f, Lm̄2 g⟩ = −W0 (f , g)
max

holds with W0 (f¯, g) = lim Wx (f¯, g) and Wx defined in (3.12). In particular, if f ∈

x→0
D(Hm,κ ) one infers that

⟨Hm,κ f, g⟩ = ⟨f, Lmax ¯

m̄2 g⟩ − W0 (f , g).
( )
Thus, g ∈ D (Hm,κ )∗ if and only if W0 (f¯, g) = 0, and then (Hm,κ )∗ g = Lmax
m̄2 g. Then, by
taking into account the explicit description of D(Hm,κ ), straightforward computations
show that W0 (f¯, g) = 0 if and only if g ∈ D(Hm̄,κ̄ ). One then deduces that (Hm,κ )∗ =
Hm̄,κ̄ .
A similar computation leads to the equality (H0ν )∗ = H0ν̄ .
42 CHAPTER 3. EXAMPLES

Corollary 3.4.3. (i) The operator Hm,κ is self-adjoint for m ∈] − 1, 1[ and κ ∈

R ∪ {∞}, and for m ∈ iR and |κ| = 1.

(ii) The operator H0ν is self-adjoint for ν ∈ R ∪ {∞}.

Proof. For the operators Hm,κ one simply has to take formula (3.13) into account for
the first case, and the same formula together with (3.11) in the second case. Finally for
the operators H0ν , taking formula (3.13) into account leads directly to the result.
Let us finally state the a result about the point spectrum for the self-adjoint oper-
ators. In this statement, Γ denotes the Γ-function.

Theorem 3.4.4. (i) If m ∈] − 1, 1[\{0}, then Hm,κ is self-adjoint if and only if

κ ∈ R ∪ {∞}, and then
{( Γ(−m) )−1/m }
σp (Hm,κ ) = −4 κ for κ ∈] − ∞, 0[,
Γ(m)
σp (Hm,κ ) = ∅ for κ ∈ [0, ∞],

(ii) If m = imi ∈ iR \ {0}, then Himi ,κ is self-adjoint if and only if |κ| = 1, and then
{ ( ( ) ) }
arg κ Γ(−im i)
Γ(imi )
+ 2πj
σp (Himi ,κ ) = − 4 exp − |j∈Z
mi

(iii) H0ν is self-adjoint if and only if ν ∈ R ∪ {∞}, and then

{ }
σp (H0ν ) = − 4 e2(ν−γ) for ν ∈ R,
σp (H0∞ ) = ∅.

The proof of this theorem, as well as much more information about these families
of operators, can be found in the preprint [DR].
Chapter 4

Spectral theory for self-adjoint

operators

In this chapter we develop the spectral theory for self-adjoint operators. As already seen
in Lemma 2.2.6, these operators have real spectrum, however much more can be said
about them, and in particular the spectrum can be divided into several parts having
distinct properties. Note that this chapter is mainly inspired from Chapter 4 of [Amr]
to which we refer for additional information.

4.1 Stieltjes measures

We start by introducing Stieltjes measures, since they will be the key ingredient for the
spectral theorem. For that purpose, let us consider a function F : R → R satisfying the
following properties:

(i) F is monotone non-decreasing, i.e. λ ≥ µ =⇒ F (λ) ≥ F (µ),

(ii) F is right continuous, i.e. F (λ) = F (λ + 0) := limε↘0 F (λ + ε) for all λ ∈ R,

(iii) F (−∞) := limλ→−∞ F (λ) = 0 and F (+∞) := limλ→+∞ F (λ) < ∞.

Note that F (λ + 0) := limε↘0 F (λ + ε) and F (λ − 0) := limε↘0 F (λ − ε) exist since F

is a monotone and bounded function.

Exercise 4.1.1. Show that such a function has at most a countable set of points of
discontinuity. For that purpose you can consider for fixed n ∈ N the set of points λ ∈ R
for which F (λ) − F (λ − 0) > 1/n.

With a function F having these properties, one can associate a bounded Borel
measure mF on R, called Stieltjes measure, starting with
( )
mF (a, b] := F (b) − F (a), a, b ∈ R (4.1)

43
44 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

and extending then this definition to all Borel sets of R (we denote by AB the set of all
Borel sets on R). More precisely, for any set V ∈ AB one sets
∑
mF (V ) = inf mF (Jk )
k

with {Jk } any sequence of half-open intervals covering the set V , and the infimum is
taken over all such covering of V . With this definition, note that mF (R) = F (+∞) and
that ( ) ( )
mF (a, b) = F (b − 0) − F (a), mF [a, b] = F (b) − F (a − 0)
( )
and therefore mF {a} = F (a) − F (a − 0) is different from 0 if F is not continuous at
the point a.
) starting with a bounded Borel measure m on R and setting F (λ) :=
( Note that
m (−∞, λ] , then F satisfies the conditions (i)-(iii) and the associated Stieltjes measure
mF verifies mF = m. Observe also that if the measure is not bounded one can not have
F (+∞) < ∞. Less restrictively, if the measure is not bounded on any bounded Borel
set, then the function F can not even be defined.
Exercise 4.1.2. Work on the examples of functions F introduced in Examples 4.1 to
4.5 of [Amr], and describe the corresponding Stieltjes measures.
Let us now recall the three types of measures on Rthat are going to play an impor-
tant role in the decomposition of any self-adjoint operator. First of all, a Borel measure
is called pure point or atomic if the measure is supported by points only. More precisely,
a Borel measure m is of this type if for any Borel set V there exists a collection of points
{xj } ⊂ V such that ∑
m(V ) = m({xj }).
j

Note that for Stieltjes measure, this set of points is at most countable. Secondly, a Borel
measure m is absolutely continuous with respect to the Lebesgue measure if there exists
a non-negative measurable function f such that for any Borel set V one has
∫
m(V ) = f (x)dx
V

where dx denotes the Lebesgue measure on R. Thirdly, a Borel measure m is singular

continuous with respect to the Lebesgue measure if m({x}) = 0 for any x ∈ R and
if there exists a Borel set V of Lebesgue measure 0 such that the support of m is
concentrated on V . Note that examples of such singular continuous measure can be
constructed with Cantor functions, see for example [Amr, Ex. 4.5].
The following theorem is based on the Lebesgue decomposition theorem for mea-
sures.
Theorem 4.1.3. Any Stieltjes measure m admits a unique decomposition

m = mp + mac + msc
4.2. SPECTRAL MEASURES 45

where mp is a pure point measure (with at most a countable support), mac is an ab-
solutely continuous measure with respect to the Lebesgue measure on R, and msc is a
singular continuous measure with respect to the Lebesgue measure R.

4.2 Spectral measures

We shall now define a spectral measure, by analogy with the Stieltjes measure intro-
duced in the previous section. Indeed, instead of considering non-decreasing functions F
defined on R and taking values in R, we shall consider non-decreasing functions defined
on R but taking values in the set P(H) of orthogonal projections on a Hilbert space H.
Definition 4.2.1. A spectral family, or a resolution of the identity, is a family {Eλ }λ∈R
of orthogonal projections in H satisfying:
(i) The family is non-decreasing, i.e. Eλ Eµ = Emin{λ,µ} ,

(ii) The family is strongly right continuous, i.e. Eλ = Eλ+0 = s − limε↘0 Eλ+ε ,

(iii) s − limλ→−∞ Eλ = 0 and s − limλ→∞ Eλ = 1,

It is important to observe that the condition (i) implies that the elements of the
families are commuting, i.e. Eλ Eµ = Eµ Eλ . We also define the support of the spectral
family as the following subset of R:
{ }
supp{Eλ } = µ ∈ R | Eµ+ε − Eµ−ε ̸= 0, ∀ε > 0 .

Given such a family and in analogy with (4.1), one first defines
( )
E (a, b] := Eb − Ea , a, b ∈ R, (4.2)

and extends this definition to all sets V ∈ AB . As a consequence of the construction,

note that (∪ ) ∑
E Vk = E(Vk ) (4.3)
k k

whenever {Vk } is a countable family of disjoint elements of AB . Thus, one ends up

with a projection-valued map E : AB → P(H) which satisfies E(∅) = 0, E(R) = 1,
E(V1 )E(V2 ) = E(V1 ∩ V2 ) for any Borel sets V1 , V2 . In addition,
( ) ( )
E (a, b) = Eb−0 − Ea , E [a, b] = Eb − Ea−0
( )
and therefore E {a} = Ea − Ea−0 .
Definition 4.2.2. The map E : AB → P(H) defined by (4.2) is called the spectral
measure associated with the family {Eλ }λ∈R . This spectral measure is bounded from
below if there exists λ− ∈ R such that Eλ = 0 for all λ < λ− . Similarly, this spectral
measure is bounded from above if there exists λ+ ∈ R such that Eλ = 1 for all λ > λ+ .
46 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

Let us note that for any spectral family {Eλ }λ∈R and any f ∈ H one can set

Ff (λ) := ∥Eλ f ∥2 = ⟨Eλ f, f ⟩.

Then, one easily checks that the function Ff satisfies the conditions (i)-(iii) of the
beginning of Section 4.1. Thus, one can associate with each element f ∈ H a Stieltjes
measure mf on R which satisfies

mf (V ) = ∥E(V )f ∥2 = ⟨E(V )f, f ⟩ (4.4)

for any V ∈ AB . Note in particular that mf (R) = ∥f ∥2 . Later on, this Stieltjes measure
will be decomposed according the content of Theorem 4.1.3.
Our next aim is to define integrals of the form
∫ b
φ(λ) E(dλ) (4.5)
a

for a continuous function φ : [a, b] → C and for any spectral family {Eλ }λ∈R . Such
integrals can be defined in the sense of Riemann-Stieltjes by first considering a partition
a = x0 < x1 < ... < xn = b of [a, b] and a collection {yj } with yj ∈ (xj−1 , xj ) and by
defining the operator
∑n
( )
φ(yj )E (xj−1 , xj ] . (4.6)
j=1

It turns out that by considering finer and finer partitions of [a, b], the corresponding
expression (4.6) strongly converges to an element of B(H) which is independent of the
successive choice of partitions. The resulting operator is denoted by (4.5).
The following statement contains usual results which can be obtained in this con-
text. The proof is not difficult, but one has to deal with several partitions of intervals.
We refer to [Amr, Prop. 4.10] for a detailed proof.

Proposition 4.2.3 (Spectral integrals). Let {Eλ }λ∈R be a spectral family, let −∞ <
a < b < ∞ and let φ : [a, b] → C be continuous. Then one has
∫b
(i) a
φ(λ) E(dλ) = supµ∈[a,b]∩ supp{Eλ } |φ(µ)|,
(∫ )∗ ∫
b b
(ii) a
φ(λ) E(dλ) = a φ(λ) E(dλ),

∫b 2 ∫b
(iii) For any f ∈ H, a
φ(λ) E(dλ)f = a
|φ(λ)|2 mf (dλ),

(iv) If ψ : [a, b] → C is continuous, then

∫ b ∫ b ∫ b
φ(λ) E(dλ) · ψ(λ) E(dλ) = φ(λ) ψ(λ) E(dλ).
a a a
4.2. SPECTRAL MEASURES 47

Let us now observe that if the support supp{Eλ } is bounded, then one can consider

∫ ∞ ∫ M
φ(λ) E(dλ) = s − lim φ(λ) E(dλ). (4.7)
−∞ M →∞ −M

Similarly, by taking property (iii) of the previous proposition into account, one observes
that this limit can also be taken if φ ∈ Cb (R). On the other hand, if φ is not bounded
on R, the r.h.s. of (4.7) is not necessarily well defined. In fact, if φ is not bounded
on R and if supp{Eλ } is not bounded either, then the r.h.s. of (4.7) is an unbounded
operator and can only be defined on a dense domain of H.

Lemma 4.2.4. Let φ : R → C be continuous, and let us set

∫
{ ∞ }
Dφ := f ∈ H | |φ(λ)|2 mf (dλ) < ∞ .
−∞

(∫ )
∞
Then the pair −∞
φ(λ) E(dλ), D φ defines a closed linear operator on H. This oper-
ator is self-adjoint if and only if φ is a real function.

∫∞
Proof. Observe first that Dφ = Dφ , and set A := −∞ φ(λ) E(dλ). A is densely defined
because its domain contains all elements with( compact ) support with respect to {Eλ },
i.e. it contains all g ∈ H satisfying g = E (−N, N ] g for some N < ∞. Thus, for
f, g ∈ Dφ one has by the point (ii) of Proposition 4.2.3

⟨ ∫ M ⟩
⟨f, Ag⟩ = lim f, φ(λ) E(dλ)g
M →∞ −M
⟨∫ M ⟩ ⟨∫ ∞ ⟩
= lim φ(λ) E(dλ)f, g = φ(λ) E(dλ)f, g .
M →∞ −M −∞

∫∞
It thus follows that Dφ ⊂ D(A∗ ), and that A∗ f = −∞ φ(λ) E(dλ)f for any f ∈ Dφ . As
∫∞
a consequence A∗ is an extension of −∞ φ(λ) E(dλ), and in order to show that these
two operators are equal it is sufficient to show that D(A∗ ) ⊂ Dφ .
For that purpose, recall that if f ∈ D(A∗ ) there exists f ∗ ∈ H such that for any
g ∈ Dφ
⟨f, Ag⟩ = ⟨f ∗ , g⟩.

In particular this equality holds if g has compact support with respect to {Eλ }. One
48 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

then gets for any M ∈ (0, ∞)

( ) ⟨ ( ) ⟩
E (−M, M ] f ∗ = sup E (−M, M ] f ∗ , g
g∈H,∥g∥=1
⟨ ( ) ⟩
= sup f ∗ , E (−M, M ] g
g∈H,∥g∥=1
⟨ ∫ M ⟩
= sup f, φ(λ) E(dλ)g
g∈H,∥g∥=1 −M
⟨∫ M ⟩
= sup φ(λ) E(dλ)f, g
g∈H,∥g∥=1 −M
∫ M (∫ M )1/2
= φ(λ) E(dλ)f = |φ(λ)| mf (dλ)
2
.
−M −M

As a consequence, one has

∫ M
sup |φ(λ)|2 mf (dλ) ≤ ∥f ∗ ∥2 < ∞
M >0 −M
∫∞
from which one infers that −∞ |φ(λ)|2 mf (dλ) < ∞. This shows that if f ∈ D(A∗ ) then
f ∈ Dφ . ∫∞
Since A∗ is always closed by (i) of Lemma 2.1.10, one infers that −∞ φ(λ) E(dλ)
on Dφ is a closed operator. So the same holds for A on Dφ . Finally, since Dφ = Dφ , the
second statement is a direct consequence of the first one.
A function φ of special interest is the function defined by the identity function id,
namely id(λ) = λ.
(∫ )
∞
Definition 4.2.5. For any spectral family {Eλ }λ∈R , the operator −∞
λ E(dλ), D id
with ∫ ∞
{ }
Did := f ∈ H | λ2 mf (dλ) < ∞
−∞

is called the self-adjoint operator associated with {Eλ }.

By this procedure, any spectral family defines a self-adjoint operator on H. The

spectral Theorem corresponds to the converse statement:

Theorem 4.2.6 (Spectral Theorem). With any self-adjoint operator (A, D(A)) on a
Hilbert space H one can associate a unique∫ spectral family {Eλ }, called the spectral
∞
family of A, such that D(A) = Did and A = −∞ λ E(dλ).

In summary, there is a bijective correspondence between self-adjoint operators and

spectral families. This theorem extends the fact that any n × n hermitian matrix is
diagonalizable. The proof of this theorem is not trivial and is rather lengthy. In the
sequel, we shall assume it, and state various consequences of this theorem.
4.2. SPECTRAL MEASURES 49

Extension 4.2.7. Study the proof the Spectral Theorem, starting with the version for
bounded self-adjoint operators.
Based on this one-to-one correspondence it is now natural to set the following
definition:
Definition 4.2.8. Let A be a self-adjoint operator in H and {Eλ } be the corresponding
spectral family. For any bounded and continuous function φ : R → C one sets φ(A) ∈
B(H) for the operator defined by
∫ ∞
φ(A) := φ(λ) E(dλ). (4.8)
−∞

Exercise 4.2.9. For any self-adjoint operator A, prove the following equality:

supp{Eλ } = σ(A). (4.9)

Note that part of the proof consists in showing that if φz (λ) = (λ − z)−1 for some
z ∈ ρ(A), then φz (A) = (A − z)−1 , where the r.h.s. has been defined in Section 2.2. Let
us also mention a useful equality which can be proved in this exercise: for any z ∈ ρ(A)
one has [ ( )]−1
(A − z)−1 = dist z, σ(A) . (4.10)
For the next statement, we set Cb (R) for the set of all continuous and bounded
complex functions on R.
Proposition 4.2.10. a) For any φ ∈ Cb (R) one has
(i) φ(A) ∈ B(H) and ∥φ(A)∥ = supλ∈σ(A) |φ(λ)|,

(ii) φ(A)∗ = φ(A), and φ(A) is self-adjoint if and only if φ is real,

(iii) φ(A) is unitary if and only if |φ(λ)| = 1.

b) The map Cb (R) ∋ φ 7→ φ(A) ∈ B(H) is a ∗-homomorphism.
c) If φ ∈ C(R), then (4.8) defines a closed operator φ(A) with domain
∫ ∞
( ) { }
D φ(A) = f ∈ H | |φ(λ)|2 mf (dλ) < ∞ . (4.11)
−∞

In the point (iii) above, one can consider the function φt ∈ Cb (R) defined by φt (λ) :=
e−iλt for any fixed t ∈ R. Then, if one sets Ut := φt (A) one first observes that Ut Us =
Ut+s . Indeed, one has
∫ ∞ ∫ ∞ ∫ ∞
−iλs
Ut Us = iλt
e E(dλ) e E(dλ) = e−iλt e−iλs E(dλ)
∫−∞
∞
−∞ −∞

= e−iλ(t+s) E(dλ) = Ut+s .

−∞
50 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

In addition, by an application of the dominated convergence theorem of Lebesgue,

one infers that the map R ∋ t 7→ Ut ∈ B(H) is strongly continuous. Indeed, since
2
e−iλ(t+ε) − e−iλt ≤ 4 one has
∫ ∞
2
∥Ut+ε f − Ut f ∥ =
2
e−iλ(t+ε) − e−iλt mf (dλ) → 0 as ε → 0.
−∞

As a consequence, such a family {Ut }t∈R is called a strongly continuous unitary group.
∑∞ (−iλt)k
Note that since e−iλt = k=0 k!
one also infers that whenever A is a bounded
operator
∑ ∞
(−itA)k
Ut = (4.12)
k=0
k!
with a norm converging series. On the other hand, if A is not bounded, then this series
converges on elements f ∈ ∩∞ k
k=0 D(A ). In particular, it converges strongly on elements
of H which have compact support with respect to the corresponding spectral measure.
Let us now mention that the above construction is only one part of a one-to-one
relation between strongly continuous unitary groups and self-adjoint operators. The
proof of the following statement can be found for example in [Amr, Prop. 5.1].
Theorem 4.2.11 (Stone’s Theorem). There exists a bijective correspondence between
self-adjoint operators on H and strongly continuous unitary groups on H. More pre-
cisely, if A is a self-adjoint operator on H, then {e−itA }t∈R is a strongly continuous
unitary group, while if {Ut }t∈R is a strongly continuous unitary group, one sets
{ 1 }
D(A) := f ∈ H | ∃ s − lim [Ut − 1]f
t→0 t
( )
and for f ∈ D(A) one sets Af = s − limt→0 ti [Ut − 1]f , and then A, D(A) is a self-
adjoint operator.
Exercise 4.2.12. Provide a precise proof of Stone’s theorem.
Let us close with section with two important observations. First of all, the map
φ 7→ φ(A) can be extended from continuous and bounded φ to bounded and measur-
able functions φ. This extension can be realized by considering the Lebesgue-Stieltjes
integrals in the weak form. In particular, this extension is necessary for defining φ(A)
whenever φ is the characteristic function on some Borel set V .
The second observation is going to provide an alternative formula for φ(A) in terms
of the unitary group {e−itA }t∈R . Indeed, assume that the inverse Fourier transform φ̌
of φ belongs to L1 (R), then the following equality holds
∫ ∞
1
φ(A) = √ φ̌(t) e−itA dt. (4.13)
2π −∞
Indeed, observe that
∫ ∫ ∫
1
⟨f, φ(A)f ⟩ = φ(λ)mf (dλ) = mf (dλ) √ e−iλt φ̌(t)dt.
R R 2π R
4.3. SPECTRAL PARTS OF A SELF-ADJOINT OPERATOR 51

By application of Fubini’s theorem one can interchange the order of integrations and
obtain
∫ ∫
1
⟨f, φ(A)f ⟩ = √ dtφ̌(t) e−iλt mf (dλ)
2π ∫R R
∫
1 ⟨ 1 ⟩
−itA
=√ dt φ̌(t)⟨f, e f ⟩ = f, √ dt φ̌(t) e−itA f ,
2π R 2π R
and one gets (4.13) by applying the polarisation identity (1.1).

4.3 Spectral parts of a self-adjoint operator

In this section, we consider a fixed self-adjoint operator A (and its associated spectral
family {Eλ }), and show that there exists a natural decomposition of the Hilbert space
H with respect to this operator. First of all, recall from Lemma 2.2.6 that the spectrum
of any self-adjoint operator is real. In addition, let us recall that for any µ ∈ R, one has
( )
Ran E({µ}) = {f ∈ H | E({µ})f = f }.

Then, one observes that the following equivalence holds:

( )
f ∈ Ran E({µ}) ⇐⇒ f ∈ D(A) with Af = µf.

Indeed, this can be inferred from the equality

∫ ∞
∥Af − µf ∥ =
2
|λ − µ|2 mf (dλ)
−∞

which itself can be deduced from the point (iii) of Proposition 4.2.3. In fact, since the
integrand is strictly positive for each λ ̸= µ, one has ∥Af − µf ∥ = 0 if and only if
mf (V ) = 0 for any Borel set V on R with µ ̸∈ V . In other words, the measure mf is
supported only on {µ}.
( )
Definition 4.3.1. The set of all µ ∈ R such that Ran E({µ}) ̸= 0 is called the point
spectrum of A or the set of eigenvalues of A. One then sets
⊕ ( )
Hp (A) := Ran E({µ})

where the sum extends over all eigenvalues of A.

In accordance with what has been presented in Theorem 4.1.3, we define two addi-
tional subspaces of H.
Definition 4.3.2.
{ }
Hac (A) := f ∈ H | mf is an absolutely continuous measure
{ }
= f ∈ H | the function λ 7→ ∥Eλ f ∥2 is absolutely continuous ,
52 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS
{ }
Hsc (A) := f ∈ H | mf is a singular continuous measure
{ }
= f ∈ H | the function λ 7→ ∥Eλ f ∥2 is singular continuous ,

for which the comparison measure is always the Lebesgue measure on R.

Note that one also uses the notation Hc (A) for the set of f ∈ H such that mf is
continuous, i.e. mf ({x}) = 0 for any x ∈ R. One also speaks then about the continuous
subspace of H with respect to A.
The following statement provides the decomposition of any self-adjoint operator into
three distinct parts. Note that depending on the operators, some parts of the following
decomposition can be trivial. The proof of the statement is not difficult and consists in
some routine computations.

Theorem 4.3.3. Let A be a self-adjoint operator in a Hilbert space H.

a) This Hilbert space can be decomposed as follows

H = Hp (A) ⊕ Hac (A) ⊕ Hsc (A),

and the restriction of the operator A to each of these subspaces defines a self-adjoint
operator denoted respectively by Ap , Aac and Asc .
b) For any φ ∈ Cb (R), one has the decomposition

φ(A) = φ(Ap ) ⊕ φ(Aac ) ⊕ φ(Asc ).

Moreover, the following equality holds

σ(A) = σ(Ap ) ∪ σ(Aac ) ∪ σ(Asc ).

Exercise 4.3.4. Provide a full proof of the above statement.

Note that one often writes Ep (A), Eac (A) and Esc (A) for the orthogonal projection
on Hp (A), Hac (A) and Hsc (A), respectively, and with these notations one has Ap =
AEp (A), Aac = AEac (A) and Asc = AEsc (A). In addition, note that the relation between
the set of eigenvalues σp (A) introduced in Definition 2.2.4 and the set σ(Ap ) is

σ(Ap ) = σp (A).

Two additional sets are often introduced in relation with the spectrum of A, namely
σd (A) and σess (A).

Definition 4.3.5.
( An eigenvalue
) λ belongs to the discrete spectrum σd (A) of A if
and only if Ran E({λ}) is of finite dimension, and λ is isolated from the rest of the
spectrum of A. The essential spectrum σess (A) of A is the complementary set of σd (A)
in σ(A), or more precisely
σess (A) = σ(A) \ σd (A).
4.3. SPECTRAL PARTS OF A SELF-ADJOINT OPERATOR 53

Since we have now all type of spectra at our disposal, let us come back to the
examples of Chapter 3. As already mentioned in Exercise 3.1.2 the spectrum of any
2 d d
self-adjoint multiplication operator φ(X) ( in L) (R ) is given by the closure of φ(R ).
Now, it is easily observed that λ ∈ σp φ(X) if and only if there exists a Borel set
V with strictly positive Lebesgue measure such that φ(x) = λ for any x ∈ V . In this
case, the multiplicity of the eigenvalue is infinite, since an infinite family of orthogonal
eigenfunctions corresponding to the eigenvalue λ can (easily )be constructed. Obviously,
the previous requirement is not necessary for λ ∈ σess φ(X) . Let us also mention that
if the function φ is continuously differentiable and if ∇φ(x) ̸= 0 for any x ∈ Rd , then
the operator φ(X) has only absolutely continuous spectrum. Such a statement will be a
consequence of the conjugate operator method introduced in subsequent chapters. Note
also that for a convolution operator φ(D) the situation is rather similar, and this can
be deduced easily from Remark 4.3.6.
For the harmonic oscillator of Section 3.1.1, the corresponding operator has only
discrete eigenvalues and no continuous spectrum, i.e. Hc (A) = {0}. For Schrödinger
operators, one has typically a mixture of continuous spectrum and of point spectrum.
Note that the eigenvalues can be embedded in the continuous spectrum, but that the
situation of eigenvalues below the continuous spectrum often appears for such operators.
For the hydrogen atom of Section 3.2.1 this situation takes place, since the continuous
spectrum corresponds to [0, ∞) while the point spectrum consists in the eigenvalues
which are all located below 0 and are converging to 0. For more general Schrödinger
operators, it is also often expected that Hsc (A) = {0}, but proving such a statement
can be a difficult task. We shall come back to this question in the following chapters.
Let us still consider any compact self-adjoint operator
( A in) a Hilbert
( space ) H. It
is easily observed that for any ε > 0 the subspace E (−∞, −ε] H ∪ E [ε, ∞) H is of
finite dimension, where E(·) corresponds to the spectral measure associated with A. In
other words, away from 0 the spectrum of A consists of eigenvalues of finite multiplicity,
and these eigenvalues can only converge to 0. On the other hand, 0 can be either an
eigenvalue with finite or infinite multiplicity, or a point of accumulation of the spectrum
without being itself a eigenvalue.

Remark 4.3.6. If A is a self-adjoint operator in a Hilbert space H and if U is a unitary

operator in H, conjugating A by U does not change its spectral properties. More pre-
cisely, in one considers AU := U AU ∗ with domain D(AU ) = U D(A), then this operator
is self-adjoint and the following equalities hold: σ(AU ) = σ(A), σp (AU ) = σp (A), ...
These facts are a consequence of the following observations: {U Eλ U ∗ } corresponds to
the spectral family for the operator AU , and then for any Borel set V
⟨( ) ⟩
mU f (V ) = U E(V )U ∗ U f, U f = ⟨E(V )f, f ⟩ = mf (V ).

We end this section with a few results which are related to the essential spectrum of
a self-adjoint operator. The first one provides another characterization of the spectrum
or of the essential spectrum of a self-adjoint operator A.
54 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

Proposition 4.3.7 (Weyl’s criterion). Let A be a self-adjoint operator in a Hilbert

space H.
a) A real number λ belongs to σ(A) if and only if there exists a sequence {fn }n∈N ⊂
D(A) such that ∥fn ∥ = 1 and s − limn→∞ (A − λ)fn = 0.
b) A real number λ belongs to σess (A) if and only if there exists a sequence {fn }n∈N ⊂
D(A) such that ∥fn ∥ = 1, w − limn→∞ fn = 0 and s − limn→∞ (A − λ)fn = 0.
Exercise 4.3.8. Provide a proof of the above statement. For convenience, you can first
provide such a proof in the special case of a multiplication operator in the Hilbert space
L2 (R).
The second result deals with the conservation of the essential spectrum under a rela-
tively compact perturbation. Before its statement, recall that the addition of a relatively
compact perturbation does not change the self-adjointness property, see Proposition
2.3.5 in conjunction with Rellich-Kato theorem.
Proposition 4.3.9. Let A be a self-adjoint operator in a Hilbert space H, and let B be
a symmetric operator in H which is A-compact. Then the following equality holds:

σess (A + B) = σess (A). (4.14)

Proof. Let us consider λ ∈ σess (A), and choose a sequence {fn }n∈N ⊂ D(A) such that
∥fn ∥ = 1, w − limn→∞ fn = 0 and s − limn→∞ (A − λ)fn = 0. Note that the existence of
such a sequence is provided by Proposition 4.3.7. By the same proposition, one would
get λ ∈ σess (A + B) if one shows that s − limn→∞ (A + B − λ)fn = 0, which is itself
implied by s − limn→∞ Bfn = 0.
For that purpose, let us fix z ∈ ρ(A) such that B(A − z)−1 ∈ K (H) and write

Bfn = [B(A − z)−1 ](A − z)fn

= [B(A − z)−1 ](A − λ)fn + (λ − z)[B(A − z)−1 ]fn .

Observe now that both terms converge to 0 as n → ∞. For the first term, this follows
directly from the assumptions. For the second one, recall that w − limn→∞ fn = 0 and
that a compact operator transform a weak convergence into a strong convergence, see
Proposition 1.4.12, which means that s − limn→∞ [B(A − z)−1 ]fn = 0. As a consequence
one infers that σess (A) ⊂ σess (A + B).
The converse statement σess (A + B) ⊂ σess (A) can be obtained similarly by consid-
ering first A + B and by perturbing this operator with the relatively compact operator
−B. Note that the relative compactness of −B with respect to A + B is a direct con-
sequence of the point (iii) of Proposition 2.3.5.
In the previous statement we have obtained the stability of the essential spectrum
under relatively compact perturbation. However, this stability does not imply anything
about the conservation of the nature of the spectrum. In that respect the following
statement shows that the nature of the spectrum can drastically change even under a
small perturbation.
4.4. THE RESOLVENT NEAR THE SPECTRUM 55

Proposition 4.3.10 (Weyl-von Neumann). Let A be an arbitrary self-adjoint operator

in a Hilbert space. Then, for any ε > 0 there exists a self-adjoint Hilbert-Schmidt
operator B with its Hilbert-Schmidt norm ∥B∥HS satisfying ∥B∥HS < ε such that A + B
has only pure point spectrum.

Exercise 4.3.11. Study the proof of this proposition, as presented for example in [Kat,
Sec. X.2].

4.4 The resolvent near the spectrum

In this section we study the resolvent (A − z)−1 of any self-adjoint operator A when
z approaches a value in σ(A). Such investigations lead quite naturally to the spectral
theorem, but also allow us to deduce useful information on the spectrum of the operator
A. The spectral family {Eλ } associated with the operator A can also be deduced from
such investigations.
As a motivation, consider the following function defined for any λ ∈ R and any
ε > 0:
1 1 2iε
R ∋ x 7→ − = ∈ C.
x − λ − iε x − λ + iε (x − λ)2 + ε2
It is known that this function converges as ε → 0 and in the sense of distributions to
2πiδ0 (x − λ). Thus, if we replace x by the self-adjoint operator A one formally infers
that
∫ b ∫ b
1 [ −1 −1
]
(A − λ − iε) − (A − λ + iε) dλ → δ0 (A − λ)dλ = χ(a,b) (A)
2πi a a
( )
= E (a, b) .

The next statement shows that this argument is almost correct, once the behavior
of the operator A at the endpoints a and b is taken into account.

Proposition 4.4.1 (Stone’s formula). Let A be a self-adjoint operator with associated

spectral family {Eλ }. Then for any −∞ < a < b < ∞ the following formulas hold:
∫
1 b [ ]
s − lim (A − λ − iε)−1 − (A − λ + iε)−1 dλ
2πi ε↘0 a
( ) 1 1
= E (a, b) + E({a}) + E({b}) (4.15)
2 2
and
∫
( ) 1 b+δ [ ]
E (a, b] = s − lim s − lim (A − λ − iε)−1 − (A − λ + iε)−1 dλ . (4.16)
2πi δ↘0 ε↘0 a+δ

Note that in the last formula the order of the two limits is important.
56 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

Proof. i) For any ε > 0 and λ ∈ R let us set

∫ b( )
1 1 1
Ψε (λ) = − dx
2πi a x − λ − iε x − λ + iε
∫ b
1 2iε
= dx
2πi a (x − λ)2 + ε2
1[ (b − λ) ( a − λ )]
= arctan − arctan .
π ε ε

) π in λ and satisfies |Ψε (λ)| ≤( x−λ

Clearly, Ψε is (continuous ) any πλ ∈ R. In addition, since
1 for
limε↘0 arctan x−λ ε
= 2
if λ < x and lim ε↘0 arctan ε
= − 2 if λ > x one infers that

1( )
Ψ0 (λ) := lim Ψε (λ) = χ(a,b) (λ) + χ[a,b] (λ) . (4.17)
ε↘0 2
∫
ii) For any ε > 0 the operator Ψε (A) := R Ψε (λ)E(dλ) is well-defined and belongs
to B(H). Our aim is to show that this operator is strongly continuous in ε and strongly
convergent for ε ↘ 0 to a limit corresponding to what is suggested by (4.17). For any
( )⊥
f ∈ H one can write f = E({a})f + E({b})f + f0 with f0 ∈ E({a})H + E({b})H .
One then has
1 (b − a)
Ψε (A)E({a})f = Ψε (a)E({a})f = arctan E({a})f,
π ε
which converges strongly to 21 E({a})f as ε ↘ 0. Similarly,

1
s − lim Ψε (A)E({b})f = E({b})f.
ε↘0 2
( ) ( ) ∫b
On the other hand, since E (a, b) f0 = E [a, b] f0 = a E(dλ)f0 one has
( )
E (a, b) f0 − Ψε (A)f0
∫ b ∫ a ∫ ∞
( )
= 1 − Ψε (λ) E(dλ)f0 − Ψε (λ)E(dλ)f0 − Ψε (λ)E(dλ)f0 .
a −∞ b

By the dominated convergence theorem, one finds that each term on the right-hand
side converges strongly to zero. Thus we have obtained that
( ) 1 1
Ψ0 (A) := s − lim Ψε (A) = E (a, b) + E({a}) + E({b}).
ε↘0 2 2

iii) To obtain the validity of (4.15) it is now sufficient to verify that for ε > 0 one
has ∫
1 b [ ]
Ψε (A) = (A − λ − iε)−1 − (A − λ + iε)−1 dλ.
2πi a
4.4. THE RESOLVENT NEAR THE SPECTRUM 57

By the polarization identity, such an equality holds if one knows that for any f ∈ H
∫ b
1 [ ]
⟨f, Ψε (A)f ⟩ = ⟨f, (A − λ − iε)−1 − (A − λ + iε)−1 f ⟩dλ.
2πi a
To prove this equality, one uses the first resolvent equation for the equality

(A − λ − iε)−1 − (A − λ + iε)−1 = 2iε(A − λ − iε)−1 (A − λ + iε)−1

∫
1
= 2iε E(dµ).
R (µ − λ) + ε
2 2

It then follows that

∫ b ∫ ∫
1 [ −1 −1
] ε b 1
(A − λ − iε) − (A − λ + iε) dλ = dλ E(dµ)
R (µ − λ) + ε
2πi a π a 2 2

and as a consequence
∫ b ∫ ∫
1 [ −1 −1
] ε b mf (dµ)
⟨f, (A − λ − iε) − (A − λ + iε) f ⟩dλ = dλ .
R (µ − λ) + ε
2πi a π a 2 2

By an application of Fubini’s theorem one deduces that

∫ b
1 [ ]
⟨f, (A − λ − iε)−1 − (A − λ + iε)−1 f ⟩dλ
2πi a
∫ ∫
ε b 1
= mf (dµ) dλ
R π a (µ − λ)2 + ε2
= ⟨f, Ψε (A)f ⟩,

as expected.
iv) Let us finally deduce (4.16) from (4.15). As a consequence of the latter formula
one has
∫ b+δ
1 [ ]
s − lim (A − λ − iε)−1 − (A − λ + iε)−1 dλ
2πi δ↘0 a+δ
( ) 1 1
= E (a + δ, b + δ) + E({a + δ}) + E({b + δ})
2 2
1 1
= Eb+δ − Ea+δ + E({a + δ}) − E({b + δ}). (4.18)
2 2
Observe also that if µ ∈ R and δ > 0, then by the right continuity of the spectral family
one has
( ) 2
∥E({µ + δ})f ∥2 ≤ E (µ, µ + δ] f = ∥(Eµ+δ − Eµ )f ∥2 → 0 as δ → 0,

or in other words s − limδ↘0 E({µ + δ}) = 0. Thus, by taking the right continuity
of {Eλ } again
( into) account, one infers that the strong limit as δ ↘ 0 of (4.18) is
Eb − Ea ≡ E (a, b] , which proves (4.16).
58 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

Let us now comment on the relation between the previous proposition and the
proof of the spectral theorem, as stated in Theorem 4.2.6. First of all, observe that the
uniqueness of the spectral family {Eλ } is a consequence∫of Stone’s formula.
∫ Indeed, if
′ ′
{Eλ } and {Eλ } are two spectral families satisfying
( A)= R λE(dλ)
( ) = R λE (dλ), then
it would follow from the equality (4.16) that E (a, b] = E ′ (a, b] (the l.h.s. of (4.16)
depends only A). Hence,
( ) ( )
Eλ = s − lim E (a, λ] = s − lim E ′ (a, λ] = Eλ′ .
a→−∞ a→−∞

For the existence of the spectral family, the proof is much longer, and can be found
in several textbooks. However, let us just sketch it. The starting point for such a proof
is (almost) always the r.h.s. of (4.16), and one has to prove the existence of its r.h.s.
at least in the weak sense, and to show that the corresponding operators have the
properties of a spectral measure. For the existence of the limit, one has to consider
1 ( ∫ b+δ ∫ b+δ )
−1
lim lim ⟨f, (A − λ − iε) f ⟩dλ − ⟨f, (A − λ + iε)−1 f ⟩dλ
2πi δ↘0 ε↘0 a+δ a+δ
∫ b+δ
1
= lim lim ℑ⟨f, (A − λ − iε)−1 f ⟩dλ.
δ↘0 ε↘0 π a+δ

Thus, let us set Φ(z) := ⟨f, (A − z)−1 f ⟩ and observe that this C-valued function is
holomorphic in the upper half complex plane. Observe also that this function has a
non-negative imaginary part and satisfies the estimate |Φ(z)| ≤ c/ℑ(z) for some finite
constant c (which means that Φ(z) is a Nevanlinna function). Then one can use a
theorem on analytic functions saying
∫ that in such a case there exists a finite Stieltjes
−1
measure m on R satisfying Φ(z) = R (λ − z) m(dλ). In addition, this measure satisfies
∫
( ) 1 b+δ
m (a, b] = lim lim ℑΦ(λ + iε)dλ.
δ↘0 ε↘0 π a+δ

As a consequence, the application of this theorem provides us for any f ∈ H the measure
m ≡ mf . The rest of the proof consists in routine computations.
Let us add another comment in relation with the previous proposition. As already
mentioned in Exercise 4.2.9, and more precisely in (4.10), for any λ ∈ σ(A) one has
1
∥(A − λ ∓ iε)−1 ∥ = .
|ε|
However, for some particular f ∈ H the expressions (A−λ∓iε)−1 f could be convergent
as ε ↘ 0. In particular, this is the case if the associated measure mf is supported away
from the value λ, or equivalently if there exists κ > 0 such that E([λ − κ, λ + κ])f = 0.
Indeed, in this situation one has
∫ λ−κ ∫ ∞
−1 −2
∥(A − λ ∓ iε) f ∥ = 2
|µ − λ ∓ iε| mf (dµ) + |µ − λ ∓ iε|−2 mf (dµ)
−∞
∫ λ+κ
1 1
≤ 2 mf (dµ) = 2 ∥f ∥2 .
κ R κ
4.4. THE RESOLVENT NEAR THE SPECTRUM 59

Since this estimate holds for any ε > 0, and even for ε = 0, one easily obtains from the
dominated convergence theorem that limε↘0 ∥(A − λ ∓ iε)−1 f − (A − λ)−1 f ∥ = 0.
Let us also observe that if λ ̸∈ σp (A), then the previous argument holds for a dense
set of elements of H. Indeed, in such a case one has s − limκ↘0 E([λ − κ, λ + κ]) = 0.
However, if E({λ}) ̸= 0 then the above set of vectors can not be dense since it is
orthogonal to E({λ})H.
Let us still assume that λ ̸∈ σp (A). In the previous paragraphs, it was shown that
the set of vectors such that the limit s − limε↘0 (A − λ ∓ iε)−1 f exists is dense in
H, but the choice of this dense set was depending on λ. A more interesting situation
would be when this set can be chosen independently of λ, or at least for any λ in some
interval (a, b). In the next statement, we show that if this situation takes place, then the
spectrum of A in (a, b) is absolutely continuous. In order to get a better understanding
of the subsequent result, let us recall that if ϕ : R → C is a smooth function, then
∫ ∫
ϕ(µ) ϕ(µ)
lim dµ = iπϕ(λ) + Pv dµ, (4.19)
ε↘0 R µ − λ − iε R µ−λ

where Pv denotes the principal value integral. In particular, let us now choose f ∈
Hac (A), which implies
∫ that there exists a non-negative measurable function θf : R → R+
such that mf (V ) = V θf (µ)dµ. If we assume in addition some regularity on θf , as for
example θf ∈ C 1 , then one infers from (4.19) that
∫ ∫
−1 θf (µ) θf (µ)
⟨f, (A − λ − iε) f ⟩ = dµ → iπθf (λ) + Pv dµ as ε ↘ 0.
R µ − λ − iε R µ−λ

The next statement clarifies the link between the existence of a limit for (A−λ−iε)−1
as ε ↘ 0 and the existence of absolutely continuous spectrum for A.

Proposition 4.4.2. Let A be a self-adjoint operator and let J := (α, β) ⊂ R.

(i) Let f ∈ H such that for each λ ∈ J the expression ℑ⟨f, (A − λ − iε)−1 f ⟩ admits
a limit as ε ↘ 0 and that this convergence holds uniformly in λ on any compact
subset of J. Then E(J)f ∈ Hac (A),

(ii) Assume that there exists a dense set D ⊂ H such that the assumptions of (i) hold
for any f ∈ D, then E(J)H ⊂ Hac (A). In particular, it follows that

σp (A) ∩ J = ∅ = σsc (A) ∩ J.

Proof. i) First of all, for z in the upper half complex plane let us set

ϕ(z) := ℑ⟨f, (A − z)−1 f ⟩.

By assumption ϕ(λ) := limε↘0 ϕ(λ + iε) exists for any λ ∈ J and defines a bounded
and uniformly continuous function on any compact subset of J.
60 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS

Let us also consider λ ∈ [a, b] ⊂ J and deduce from Stone’s formula that
∫ λ+δ
⟨ ( ) ⟩ 1 ⟨ ⟩
f, E (a, λ] f = lim lim ℑ f, (A − µ − iε)−1 f dµ
π δ↘0 ε↘0 a+δ
where we can choose δ small enough such that [a+δ, λ+δ] ⊂ J. Then, by the observation
made in the previous paragraph and by an application of the dominated convergence
theorem one infers that
∫ λ+δ ∫
⟨ ( ) ⟩ 1 1 λ
f, E (a, λ] f = lim ϕ(µ)dµ = ϕ(µ)dµ.
π δ↘0 a+δ π a
( )
( ϕ )is continuous on [a, b] if follows that ϕ ∈ ⟨L ([a,
1
Since ( b]). Finally,
) ⟩ since E (a, λ] f =
E (a, λ] E(J)f one infers that the map λ 7→ f, E (a, λ] f defines an absolutely
continuous measure on (a, b].
ii) One has E(J)f ∈ Hac (A) for each f ∈ D. Since D is assumed to be dense in H, it
easily follows that {E(J)f | f ∈ D} is dense in E(J)H, and hence E(J)H ⊂ Hac (A).
The following result will be at the root of the commutator methods introduced later
on.
Theorem 4.4.3 (Putnam’s theorem). Let H and A be bounded self-adjoint operators
satisfying [iH, A] ≥ CC ∗ for some C ∈ B(H). Then for all λ ∈ R, any ε > 0 and each
f ∈ H one has
ℑ⟨Cf, (H − λ − iε)−1 Cf ⟩ ≤ 4∥A∥ ∥f ∥2 , (4.20)
and Ran(C) ⊂ Hac (H). In particular, if Ker(C ∗ ) = {0}, then the spectrum of H is
purely absolutely continuous.
Proof. In this proof, we use the notation R(z) for (H − z)−1 when z ∈ ρ(H).
i) For any f ∈ H one has
1⟨ ⟩
ℑ⟨Cf, R(λ + iε)Cf ⟩ = Cf, [R(λ + iε) − R(λ − iε)]Cf
2i⟨ ⟩
= ε Cf, R(λ + iε)R(λ − iε)Cf
= ε∥R(λ − iε)Cf ∥2
≤ ε∥R(λ − iε)C∥ ∥f ∥2 . (4.21)
Since for any bounded operator T one has ∥T ∥2 = ∥T T ∗ ∥ one infers that
∥R(λ − iε)C∥2 = ∥R(λ − iε)CC ∗ R(λ + iε)∥
⟨ ⟩
= sup R(λ + iε)f, CC ∗ R(λ + iε)f
f ∈H,∥f ∥=1
⟨ ⟩
≤ sup R(λ + iε)f, [iH, A]R(λ + iε)f
f ∈H,∥f ∥=1

= R(λ − iε)[iH, A]R(λ + iε)

= R(λ − iε)[i(H − λ + iε), A]R(λ + iε)
≤ ∥AR(λ + iε)∥ + ∥R(λ − iε)A∥ + ∥2iεR(λ − iε)AR(λ + iε)∥.
4.4. THE RESOLVENT NEAR THE SPECTRUM 61

Finally, since ∥R(λ ± iε)∥ ≤ ε−1 one gets ∥R(λ − iε)C∥2 ≤ 4ε−1 ∥A∥. By inserting this
estimate into (4.21) one directly deduces the inequality (4.20).
ii) Let {Eλ } be the spectral family of H. For any J = (a, b] one has by Stone’s
formula
∫ b+δ
⟨ ( ) ⟩ 1 ⟨ ⟩
Cf, E (a, b] Cf = lim lim ℑ Cf, (H − λ − iε)−1 Cf dλ.
π δ↘0 ε↘0 a+δ

Now since one has

∫ b+δ
⟨ ⟩
ℑ Cf, (H − λ − iε)−1 Cf dλ ≤ 4∥A∥ ∥f ∥2 (b − a)
a+δ

one infers that

⟨Cf, E(J)Cf ⟩ ≡ mCf (J) ≤ 4∥A∥ ∥f ∥2 |J|,
where |J| means the Lebesgue measure of J. Such an inequality implies that mCf (V ) ≤
4∥A∥ ∥f ∥2 |V | for any Borel set V of R, and consequently that the measure mCf
is absolutely continuous with respect to the Lebesgue measure. It thus follows that
Cf ∈ Hac (H). Finally, if Ker(C ∗ ) = {0}, then Ran(C) is dense in H, as proved in (ii)
of Lemma 2.1.10. As a consequence, one obtains that Hac (H) = H.
62 CHAPTER 4. SPECTRAL THEORY FOR SELF-ADJOINT OPERATORS
Chapter 5

Scattering theory

In this chapter we introduce the basic tools of scattering theory. This theory deals
with the unitary group generated by any self-adjoint operator and corresponds to a
comparison theory. More precisely, if A and B are self-adjoint operators in a Hilbert
space H, and if the corresponding unitary groups are denoted by {Ut }t∈R and {Vt }t∈R ,
then one typically considers the product operator Vt∗ Ut and study its behavior for
large |t|. Understanding the limit limt→±∞ Vt∗ Ut in a suitable sense, provides many
information on the relation between the operator A and B.
Scattering theory was first developed in close relation with physics. However, it is
now a mathematical subject on its own, and new developments are currently taking
place in a more interdisciplinary framework.

5.1 Evolution groups

Let A be a self-adjoint operator in a Hilbert space H. The elements of the correspond-
ing strongly continuous unitary group {Ut }t∈R provided by Stone’s theorem in Theorem
4.2.11 are often denoted by Ut = e−itA . This group is called the evolution group associ-
ated with A. Let us also recall that if {Ut }t∈R is a strongly( continuous
) unitary group,
then its generator corresponds to the self-adjoint operator A, D(A) defined by
{ 1 }
D(A) := f ∈ H | ∃ s − lim [Ut − 1]f
t→0 t

and for f ∈ D(A) by Af = s − limt→0 ti [Ut − 1]f . Obviously, the relation Ut = e−itA
then holds, and the domain D(A) is left invariant by the action of Ut for any t ∈ R.
Given a strongly continuous unitary group {Ut }t∈R it is often not so simple to
compute explicitly D(A). However, in applications one can often guess a smaller domain
D ⊂ H on which the computation of s − limt→0 1t [Ut − 1]f is well-defined. The following
statement provides a criterion for checking if the domain D is large enough for defining
entirely the operator A.

63
64 CHAPTER 5. SCATTERING THEORY

Proposition 5.1.1 (Nelson’s criterion). Let {Ut }t∈R be a strongly continuous unitary
group, and let A denotes its self-adjoint generator. Let D be a dense linear submanifold
of H such that D is invariant under the action of Ut for any t ∈ R and such that
s − limt→0 1t [Ut − 1]f has a limit for any f ∈ D. Then A is essentially self-adjoint on
D.
Proof. Let us denote by A0 the restriction of A to D. Since A is self-adjoint, A0 is clearly
symmetric. In order to show that A0 is essentially self-adjoint, we shall use the criterion
(ii) of Proposition 2.1.15. More precisely, A0 is essentially self-adjoint if Ran(A0 ± i) are
dense in H, or equivalently if Ker(A∗0 ∓ i) = {0}. Note that we have also used Lemma
2.1.10 for the previous equivalence.
Let us show that Ker(A∗0 − i) = {0}. For that purpose, assume that h ∈ Ker(A∗0 − i),
i.e. h ∈ D(A∗0 ) and A∗0 h = ih. For any f ∈ D one has
d
⟨Ut f, h⟩ = ⟨−iAUt f, h⟩ = i⟨A0 Ut f, h⟩
dt
= i⟨Ut f, A∗0 h⟩ = i⟨Ut f, ih⟩ = −⟨Ut f, h⟩

where the invariance of D under Ut has been used for the second equality. Thus if one
sets ϕ(t) := ⟨Ut f, h⟩ one gets the differential equation ϕ′ (t) = −ϕ(t), whose solution is
ϕ(t) = ϕ(0) e−t . If ϕ(0) ̸= 0 it follows that |ϕ(t)| → ∞ as t → −∞ which is impossible
since |ϕ(t)| ≤ ∥f ∥ ∥h∥. One deduces that ϕ(0) = 0 which means that ⟨f, h⟩ = 0. It
follows that h is perpendicular to D, but by density of D in H one concludes that
h = 0.
Remark 5.1.2. Dealing with the group {e−itA }t∈R let us provide two formulas which
could also have been introduced in the previous chapter, namely
∫ ∞
−1
(A − z) = i eizt e−itA dt, for ℑ(z) > 0 , (5.1)
0
∫ 0
−1
(A − z) = −i eizt e−itA dt, for ℑ(z) < 0 . (5.2)
−∞

Since the map t 7→ eizt e−itA is strongly continuous and integrable in norm, the above
integrals exist in the strong sense by Proposition 1.5.3. Their equality with the resolvent
of A at z can be checked directly, as shown for example in the proof of [Amr, Prop. 5.1].
Let us now present a few examples of evolution groups and their corresponding
self-adjoint generators.
Example 5.1.3. In the Hilbert space H := L2 (R) we consider the translation group,
namely [Ut f ](x) := f (x − t) for any f ∈ H and x ∈ R. It is easily checked that {Ut }t∈R
defines a strongly continuous unitary group in H. In addition, its self-adjoint generator
can by computed on Cc∞ (R), according to Proposition 5.1.1. One then finds that the
generator is −i dx
d
, or in other words the operator D already considered in Chapter 3.
Note that the operator D is indeed essentially self-adjoint on Cc∞ (R).
5.1. EVOLUTION GROUPS 65

Example 5.1.4. In H := L2 (Rd ) we consider the dilation group acting on any f ∈ H

by [Ut f ](x) = edt/2 f (et x) for any x ∈ Rd . It is also easily checked that {Ut }t∈R defines
a strongly continuous unitary group in H. The self-adjoint generator of this group can
by computed on the Schwartz space S(Rd ), according to Proposition 5.1.1. A direct
computations shows that this generator A is given on S(Rd ) by the expression

1∑
d
1
A=− (Xj Dj + Dj Xj ) ≡ − (X · D + D · X).
2 j=1 2

Example 5.1.5. Consider the Hilbert space H := L2 (Rd ) and the Laplace operator
−∆ = D2 , as already introduced in equation (3.3). The unitary group generated by
this operator has obviously a[very simple ] expression once a Fourier transformation is
−itD −itξ
performed, or more precisely F e [Ff ](ξ) for any f ∈ H and ξ ∈ Rd .
2 2
f (ξ) = e
Without this Fourier transformation, this operator corresponds to the following integral
operator: ∫
[ −itD2 ] 1 i|x−y|2
e f (x) = e 4t f (y)dy
(4πit)d/2 Rd
with the square root given by
( 4πit )−n/2 { e−inπ/4 if t > 0,
=
|4πit| e+inπ/4 if t < 0.

Exercise 5.1.6. Work on the details of the results presented in the previous three
examples.

Later on, we shall often have to compute the derivative with respect to t of the
product of two unitary groups. Since the generators of these groups are often unbounded
operators, some care is necessary. In the next Lemma, we provide some conditions in
order to deal with the Leibnitz rule in this setting. The proof of this Lemma is provided
for example in [Amr, Prop. 5.5]. Note that the first statement can even be used in the
special case B = 1.
( ) ( )
Lemma 5.1.7. Let A, D(A) and B, D(B) be self-adjoint operators in a Hilbert space
H.

(i) Let C ∈ B(H) be such that CD(B) ⊂ D(A). Then for any f ∈ D(B) the map
t 7→ eitA C e−itB f is strongly differentiable and
d itA (
e C e−itB f = i eitA AC − CB) e−itB f . (5.3)
dt

(ii) Let C be B-bounded and such that CD(B 2 ) ⊂ D(A). Then for any f ∈ D(B 2 ) the
map t 7→ eitA C e−itB f is strongly differentiable and its derivative is again given
by (5.3).
66 CHAPTER 5. SCATTERING THEORY

In the next statements, we study the asymptotic behavior of different parts of

the Hilbert space under the evolution group. First of all, we consider the absolutely
continuous subspace.
Proposition 5.1.8. Let A be a self-adjoint operator and let {Ut }t∈R be the correspond-
ing unitary group. Let also f ∈ Hac (A). Then,
(i) Ut f converges weakly to 0 as t → ±∞,
(ii) If B ∈ B(H) is A-compact, then ∥BUt f ∥ → 0 as t → ±∞.
∫
Proof. i) For any h ∈ Hac (A) and since mh (R) = R mh (dλ) = ∥h∥∫2 , one infers that
there exists a non-negative function θ ∈ L1 (R) such that mh (V ) = V θ(λ)dλ for any
V ∈ AB . It thus follows that
∫ ∫
−itλ
φ(t) := ⟨h, Ut h⟩ = e mh (dλ) = e−itλ θ(λ)dλ.
R R

Thus, φ is the Fourier transform of the function θ ∈ L1 (R), and consequently belongs
to C0 (R) by the Riemann-Lebesgue lemma. It follows that limt→±∞ ⟨h, Ut h⟩ = 0.
We now show that limt→±∞ ⟨g, Ut f ⟩ = 0 for any g ∈ H and f ∈ Hac (A). Since
Ut f ∈ Hac (A) for any t ∈ R it follows that ⟨g, Ut f ⟩ = 0 if g ∈ Hs (A) := Hac (A)⊥ . Thus,
one can assume that g ∈ Hac (A). By the polarization identity, one then obtains that
⟨g, Ut f ⟩ is the sum of four terms of the form ⟨g + αf, Ut (g + αf )⟩ for some α ∈ C, and
since g + αf belongs to Hac (A) one infers from the previous paragraph that these four
contributions converge to 0 as t → ±∞, and this proves the statement (i).
ii) Observe first that it is sufficient to prove the statement (ii) for a dense set of
elements of Hac (A). Let us take for this dense set the linear manifold Hac (A) ∩ D(A),
and for any f in this set we define g := (A + i)f . Clearly g ∈ Hac (A) and one has
f = (A + i)−1 g. It follows that
BUt f = BUt (A + i)−1 g = B(A + i)−1 Ut g. (5.4)
Since Ut g converges weakly to 0 by the statement (i) and since B(A + i)−1 belongs to
K (H), one deduces from Proposition 1.4.12 that B(A + i)−1 Ut g converges strongly to 0
as t → ±∞. By (5.4), it means that BUt f converges strongly to 0, as stated in (ii).
For f ∈ Hsc (A), the previous result does not hold in general. However, once a certain
average is taken, similar results can be deduced. Since t is often interpreted as the time,
one speaks about a temporal mean. We state the result for arbitrary f ∈ Hc (A), and
refer to [Amr, Prop. 5.9] for its proof.
Proposition 5.1.9. Let A be a self-adjoint operator and let {Ut }t∈R be the correspond-
ing unitary group. Let also f ∈ Hc (A). Then,
(i) For any h ∈ H one has
∫ ∫
1 T 1 0
lim |⟨h, Ut f ⟩|2 dt = 0 and lim |⟨h, Ut f ⟩|2 dt = 0.
T →∞ T 0 T →∞ T −T
5.2. WAVE OPERATORS 67

(ii) If B ∈ B(H) is A-compact, then one has

∫ T ∫ 0
1 1
lim ∥BUt f ∥2 dt = 0 and lim ∥BUt f ∥2 dt = 0.
T →∞ 0 T T →∞ T −T

The first result can in fact be deduced from a stronger statement, usually called
RAGE theorem in honor of its authors Ruelle, Amrein, Georgescu and Enss.

Theorem 5.1.10 (RAGE Theorem). Let A be a self-adjoint operator and let {Ut }t∈R
be the corresponding unitary group. Let also E({0}) denote the spectral projection onto
Ker(A). Then for any f ∈ H one has
∫ T
1
s − lim Ut f dt = E({0})f.
T →±∞ T 0

In particular, if f ⊥ Ker(A), then the previous limit is 0.

Exercise 5.1.11. Provide a proof of RAGE theorem, see for Example [RS3, Thm
XI.115].

Let us close this section with one more result about Ut f for any f ∈ Hc (A). Its
proof can be found in [RS3, Corol. p. 343]. We emphasize that in the statement, the
family {tk } can be chosen independently of the element f ∈ Hc (A).

Corollary 5.1.12. Let A be a self-adjoint operator and let {Ut }t∈R be the corresponding
unitary group. Then there exists a sequence {tk }k∈N ⊂ R with tk → ∞ as k → ∞ such
that w − limk→∞ Utk f = 0 for any f ∈ Hc (A). In addition, if B ∈ B(H) is A-compact,
then limk→∞ ∥BUtk f ∥ = 0 for any f ∈ Hc (A).

5.2 Wave operators

Scattering theory is mainly a comparison theory. Namely, given a self-adjoint operator H
on a Hilbert space H one wonders if the evolution group {e−itH }t∈R can be approximated
by a simpler evolution group {e−itH0 }t∈R as t → ±∞. More precisely, let f ∈ H and
consider the family of elements e−itH f ∈ H. The previous question reduces to looking
for a “simpler” operator H0 and for two elements f± ∈ H such that

lim e−itH f − e−itH0 f± = 0. (5.5)

t→±∞

Obviously, one has to be more precise in what “simpler” means, and about the set of f
which admit such an approximation.
Observe first that there is not a single procedure which leads to a natural candidate
for H0 . Such a choice depends on the framework and on the problem. However, the
68 CHAPTER 5. SCATTERING THEORY

initial question can be rephrased very precisely. By using the unitarity of eitH , observe
that (5.5) is equivalent to

lim f − eitH e−itH0 f± = 0. (5.6)

t→±∞

For that reason, a natural object to consider is s − limt→±∞ eitH e−itH0 . However, this
limit has a better chance to exist if considered only on a subspace of the Hilbert space.
So let E be an orthogonal projection and assume that E e−itH0 = e−itH0 E for any t ∈ R.
Equivalently, this means that the subspace EH is left invariant by the unitary group
{e−itH0 }t∈R . We say in that case that E commutes with the evolution group {e−itH0 }t∈R .
Note that we have chosen the notation E for this projection because in most of the
applications E is related to the spectral family of H0 . However, other choices can also
appear.

Definition 5.2.1. Let H, H0 be two self-adjoint operators in a Hilbert space H, and let
E be an orthogonal projection which commutes with {e−itH0 }t∈R . The wave operators
are defined by
W± (H, H0 , E) := s − lim eitH e−itH0 E (5.7)
t→±∞

whenever these limits exit. If E = 1 then these operators are denoted by W± (H, H0 ).

Note that we could have chosen two different projections E± for t → ±∞. Since
the general theory is not more difficult in this case, we do not mention it in the sequel.
However, in applications this slight extension is often useful. We now provide some
information about these operators. Recall that some properties of isometries and partial
isometries have be introduced in Propositions 1.4.6 and 1.4.8.

Proposition 5.2.2. Let W := W± (H, H0 , E) be one of the wave operators. Then

(i) W is a partial isometry, with initial set EH. In particular, W is an isometry if

E = 1,

(ii) W intertwines the two operators H0 and H, or more precisely e−itH W = W e−itH0
for any t ∈ R. More generally, E H (V )W = W E H0 (V ) for any Borel set V , and
φ(H)W = W φ(H0 ) for any φ ∈ Cb (R). The following equality also holds:

HW f = W H0 f ∀f ∈ D(H0 ). (5.8)

Proof. In this proof we consider only W := W+ (H, H0 , E), the case W = W− (H, H0 , E)
being similar.
i) If f ⊥EH then clearly W f = 0. On the other hand, if f ∈ EH then ∥W f ∥ =
limt→∞ ∥ eitH e−itH0 f ∥ = limt→∞ ∥f ∥ = ∥f ∥, where Lemma 1.1.5 has been used for the
first equality. It follows that W is a partial isometry, or an isometry if E = 1.
5.2. WAVE OPERATORS 69

ii) Observe that

e−itH W = e−itH s − lim eisH e−isH0 E

s→∞

= s − lim e i(s−t)H
e−isH0 E
s→∞
′ ′
= s − lim
′
e−is H e−i(s +t)H0 E
s →∞
= W e−itH0

where we have used that E and e−itH0 commute. This proves the first part of the
statement (ii). Then, by multiplying this equality with ±i eizt and by integrating with
respect to t on [0, ∞) for ℑ(z) > 0 or on (−∞, 0] if ℑ(z) < 0, one gets as in Remark
5.1.2 the equality

(H − z)−1 W = W (H0 − z)−1 for any z ∈ C \ R.

One then also deduces for any α, β ∈ R that

∫ β ( )
(H − λ − iε)−1 − (H − λ + iε)−1 W dλ
α
∫ β
( )
=W (H0 − λ − iε)−1 − (H0 − λ + iε)−1 dλ
α

Thus, by considering α = a + δ, β = b + δ and ( by) taking consecutively

( ) the two limits
H H0
limε↘0 and then limδ↘0 one infers that E (a, b] W = W E (a, b] for any a < b.
Considering the limit a → −∞ one finds EλH W = W EλH0 for any λ ∈ R. The equality
mentioned in the statement for any Borel set follows then from the equality for any
elements of the spectral family. The equality φ(H)W = W φ(H0 ) follows also from the
previous equality and from the definition of the function of an operator.
For (5.8) observe that if f ∈ D(H0 ), then W e−itH0 f is strongly differentiable at
t = 0, with derivative −iW H0 f . However, since W e−itH0 f = e−itH W f , this function
is also strongly differentiable at t = 0. It then follows from Stone’s theorem that W f ∈
D(H) and that the derivative at t = 0 is given by −iHW f . The equality (5.8) follows
then directly.

Note that the properties mentioned in the point (ii) are usually referred to as
the intertwining properties of the wave operators. Note also that the different steps
presented in the proof, namely how to go from an intertwining relation for the unitary
group to an intertwining relation for arbitrary continuous and bounded functions, is a
quite common procedure. One can prove similarly that if B ∈ B(H) satisfies CUt = Ut C
for an arbitrary unitary group, then Cφ(A) = φ(A)C for any bounded and continuous
function of its generator A.
Let us now state some additional properties of the wave operators.
70 CHAPTER 5. SCATTERING THEORY

Proposition 5.2.3. Let W± := W± (H, H0 , E) be the wave operators for the pair
(H, H0 ) and the initial set projection E. Let F± be the final range projection, i.e. the
orthogonal projection on Ran(W± ) which is given by F± := W± W±∗ . Then F± commute
with the elements of the unitary group {e−itH }t∈R and the limits

W± (H0 , H, F± ) := s − lim eitH0 e−itH F±

t→±∞

exist and satisfy W± (H0 , H, F± ) = W± (H, H0 , E)∗ . In addition, the following implica-
tions hold:

EH ⊂ Hac (H0 ) =⇒ F± H ⊂ Hac (H),

EH ⊂ Hc (H0 ) =⇒ F± H ⊂ Hc (H).

Proof. In this proof we consider only W := W+ (H, H0 , E), the case W = W− (H, H0 , E)
being similar. Accordingly, we simply write F for F+ .
i) Recall first that e−itH W = W e−itH0 for any t ∈ R. By taking the adjoint on both
sides, and by switching t to −t one infers that W ∗ e−itH = e−itH0 W ∗ for any t ∈ R. It
then follows that

e−itH F = e−itH W W ∗ = W e−itH0 W ∗ = W W ∗ e−itH = F e−itH ,

which corresponds to the expected commutation relation.

ii) Let us now consider g ∈ Ran(W ). There exists thus f ∈ EH such that g = W f .
Then we have
( )
∥ eitH0 e−itH g − f ∥ = eitH e−itH0 eitH0 e−itH g − f
= ∥g − eitH e−itH0 f ∥ = ∥W f − eitH e−itH0 f ∥,

which converges to 0 as t → +∞. It thus follows that the operator W+ (H0 , H, F ) exists.
One also deduces from the previous equalities that s − limt→+∞ eitH0 e−itH W = E, and
as a consequence

W+ (H0 , H, F ) = s − lim eitH0 e−itH F

t→+∞

= s − lim eitH0 e−itH W W ∗ = EW ∗ = (W E)∗ = W ∗ ,

t→+∞

as mentioned in the statement.

iii) Let again g ∈ Ran(W ) and let f ∈ EH such that W f = g. One then infers from
Proposition 5.2.2 that for any V ∈ AB one has

⟨g, E H (V )g⟩ = ⟨W f, E H (V )W f ⟩ = ⟨W ∗ W f, E H0 (V )f ⟩ = ⟨f, E H0 (V )f ⟩.

Thus, if the measure mH f is absolutely continuous, then the same property holds for the
0

measure mH g . Similarly, if f belongs to Hc (H0 ), then g = W f belongs to Hc (H).

5.2. WAVE OPERATORS 71

Up to now, we have studied some properties of the wave operators by assuming

their existence. In the next statement, we give a criterion which ensures their existence.
Its use is often quite easy, especially if the evolution group generated by H0 is simple
enough.

Proposition 5.2.4 (Cook criterion). Let H0 , H be two self-adjoint operators in a

Hilbert space H, let M be a subspace of H invariant under the group {e−itH0 }t∈R , and
let D be a linear subset of M satisfying

(i) The linear combinations of elements of D span a dense set in M,

(ii) e−itH0 f ∈ D(H) ∩ D(H0 ) for any f ∈ D and t ∈ R,

∫ ±∞
(iii) ±1 ∥(H − H0 ) e−iτ H0 f ∥dτ < ∞ for any f ∈ D.

Then W± (H, H0 , E) exists, with E the orthogonal projection on M.

Proof. As in the previous proofs, we consider only W+ (H, H0 , E), the proof for the
other wave operator being similar.
For any f ∈ D and by the assumption (ii) one infers that

d ( itH −itH0 ) ( d itH ) −itH0 ( d −itH0 )

e e f = e e f + eitH e f
dt dt dt
= i eitH (H − H0 ) e−itH0 f.

By the result of Proposition 1.2.3.(iii) and for any t > s > 1 one gets that
∫
−itH0 −isH0
t
d ( iτ H −iτ H0 )
e itH
e f −e isH
e f= e e f dτ
dτ
∫
s
t
=i eiτ H (H − H0 ) e−iτ H0 f dτ
s

from which one infers that

∫
−itH0 −isH0
t
d ( iτ H −iτ H0 )
eitH
e f −e isH
e f = e e f dτ
dτ
∫ t
s

≤ eiτ H (H − H0 ) e−iτ H0 f dτ
∫s t
= (H − H0 ) e−iτ H0 f dτ .
s

Since the latter expression is arbitrarily small for s and t large enough, one deduces
that the map t 7→ eitH e−itH0 f is strongly Cauchy for any f ∈ D, and thus strongly
convergent for any f ∈ D. The strong convergence on M directly follows by a simple
density argument.
72 CHAPTER 5. SCATTERING THEORY

Let us still mention a rather famous result about trace-class perturbations, see also
Extension 1.4.14. For its proof we refer for example to [Kat, Thm. X.4.4].

Theorem 5.2.5 (Kato-Rosenblum theorem). Let H and H0 be two self-adjoint oper-

ators in a Hilbert space such that H − H0 is a trace
( class operator) (or in particular a
finite rank operator). Then the wave operators W± H, H0 , Eac (H0 ) exist.

Extension 5.2.6. Work on this theorem and on its proof.

Let us now provide a few examples for which the existence of the wave operators
has been shown. Note that most the time, the existence is proved by using Proposition
5.2.4 or a slightly improved version of it. The first example corresponds to a Schrödigner
operator with a short-range potential.

Example 5.2.7. In the Hilbert space H := L2 (Rd ), let H0 be the Laplace operator −∆
and let H := H0 +V (X) with V (X) a multiplication operator by a real valued measurable
function which satisfies
1
|V (x)| ≤ c
(1 + |x|)1+ε
for some constant c > 0 and some ε > 0. Then the projection E can be chosen equal to
1 and the wave operators W± (H, H0 ) exist. Note that such a result is part of the folklore
of scattering theory for Schrödinger operators and that the proof of such a statement
can be found in many textbooks.

The second example is a very simple system on which all the computations can be
performed explicitly, see [Yaf, Sec. 2.4].

Example 5.2.8. Let H := L2 (R) and consider the operator H0 defined by the operator
D. The corresponding unitary group acts as [e−itH0 f ](x) = f (x − t) as mentioned in
Example 5.1.3. Let also q∫ : R → R belong to L1 (R) and consider the unitary operator V
x
defined by [V f ](x) = e−i 0 q(y)dy f (x) for any f ∈ H and x ∈ R. By setting H := V H0 V ∗
one checks that H is the operator defined on
{ }
D(H) := f ∈ H | f is absolutely continuous and − if ′ + qf ∈ L2 (R)

with Hf = −if ′ + qf for any f ∈ D(H). The unitary group generated by H can then
be computed explicitly and one gets
∫ x+t
[eitH e−itH0 f ](x) = [V eitH0 V ∗ e−itH0 f ](x) = ei x q(y)dy
f (x).

We can then conclude that the wave operators W± (H, H0 ) exist and are given by
∫ ±∞
[W± (H, H0 )f ](x) = ei x q(y)dy
f (x).

We add one more example for which all computations can be done explicitly and
refer to [Ric, Sec. 2] for the details.
5.3. SCATTERING OPERATOR AND COMPLETENESS 73

Example 5.2.9. Let H := L2 (R+ ) and consider the Dirichlet Laplacian HD on R+ .

d2
More precisely, we set HD = − dx 2 with the domain D(HD ) = {f ∈ H (R+ ) | f (0) = 0}.
2

Here H2 (R+ ) means the usual Sobolev space on R+ of order 2. For any α ∈ R, let
d2
us also consider the operator H α defined by H α = − dx 2 with D(H ) = {f ∈ H (R+ ) |
α 2
′ α
f (0) = αf (0)}. It can easily be checked that if α < 0 the operator H possesses only one
eigenvalue, namely −α2 , and the corresponding eigenspace is generated by the function
x 7→ eαx . On the other hand, for α ≥ 0 the operators H α have no eigenvalue, and so
does HD .
Let us also recall the action of the dilation group in H, as already introduced in
Example 5.1.4. This unitary group {Ut }t∈R acts on f ∈ H as
( )
[Ut f ](x) = et/2 f et x , ∀x ∈ R+ .

Its self-adjoint generator is denoted by A. For this model, the following equality can be
proved
√
( −1
)[ α + i HD ]
W− (H , HD ) = 1 + 2 1 + tanh(πA) − i cosh(πA)
α 1
√ −1 . (5.9)
α − i HD

and a similar formula holds for W+ (H α , HD ). Let us still mention that the function of
A in the above formula is linked to the Hilbert transform.

5.3 Scattering operator and completeness

In this section we consider again two self-adjoint operators H and H0 in a Hilbert space
H, and assume that the wave operators W± (H, H0 ) exist. Note that for simplicity we
have set E = 1 but the general theory can be considered without much additional
efforts.

Definition 5.3.1. In the framework mentioned above, the operator

( )∗
S ≡ S(H, H0 ) := W+ (H, H0 ) W− (H, H0 )

is called the scattering operator for the pair (H, H0 ).

We immediately state and prove some properties of this operator. For simplicity,
we shall simply write W± for W± (H, H0 ).

Proposition 5.3.2. (i) The scattering operator commutes with H0 , or more precisely

[S, e−itH0 ] = 0 ∀t ∈ R, (5.10)

and for any f ∈ D(H0 ) one has Sf ∈ D(H0 ) and SH0 f = H0 Sf .

(ii) S is an isometric operator if and only if Ran(W− ) ⊂ Ran(W+ ),

74 CHAPTER 5. SCATTERING THEORY

(iii) S is a unitary operator if and only if Ran(W− ) = Ran(W+ ).

Proof. i) The first statement directly follows from the intertwining relations as presented
in Proposition 5.2.2 and in its proof for the adjoint operators. Indeed one has

S e−itH0 = W+∗ W− e−itH0 = W+∗ e−itH W− = e−itH0 W+∗ W− = e−itH0 S.

Then, for any f ∈ D(H0 ) observe that ti S(e−itH0 −1)f = ti (e−itH0 −1)Sf , and since the
l.h.s. converges to SH0 f , the r.h.s. must also converge and it converges then to H0 Sf ,
which proves the second part of the statement (i).
ii) Let us set F± for the final range projection, i.e. F± := W± W±∗ . The assumption
Ran(W− ) ⊂ Ran(W+ ) means F+ W− = W− . Under this hypothesis one has

S ∗ S = W−∗ W+ W+∗ W− = W−∗ F+ W− = W−∗ W− = 1,

which means that S is isometric. On the other hand if Ran(W− ) ⊂ Ran(W+ ) is not
satisfied, then there exists g ∈ Ran(W− ) with g ̸∈ Ran(W+ ) ≡ Ran(F+ ). By setting
f = W−∗ g (so that g = W− f ) one infers that ∥F+ g∥ < ∥g∥ = ∥W− f ∥ = ∥f ∥, since W−
is an isometry. It follows that

∥Sf ∥ = ∥W+∗ W− f ∥ = ∥W+∗ g∥ = ∥W+ W+∗ g∥ = ∥F+ g∥ < ∥f ∥,

which means that S can not be isometric.

iii) S is unitary if and only if S and S ∗ are isometric. By (ii) S is isometric if and
only if Ran(W− ) ⊂ Ran(W+ ). Since S ∗ = W−∗ W+ one infers by exchanging the role of the
two operators that S ∗ is isometric if and only if Ran(W+ ) ⊂ Ran(W− ). This naturally
leads to the statement (iii).
In relation with the previous statement, let us assume that H0 is purely absolutely
continuous. In that case, one often says that the scattering system for the pair (H, H0 )
is complete 1 if Ran(W− ) = Hac (H) or if Ran(W+ ) = Hac (H). We also say that the
asymptotic completeness holds if Ran(W− ) = Ran(W+ ) = Hp (H)⊥ . Note that this
latter requirement is a very strong condition. In particular it implies that H has not
singular continuous spectrum, and that for any f ∈ Hac (H) there exists f± such that

lim e−itH f − e−itH0 f± = 0. (5.11)

t→±∞

In other words, the evolution of any element of Hac (H) can be described asymptotically
by the simpler evolution e−itH0 on a vector f± .
In the Examples 5.2.7, 5.2.8 and 5.2.9, the corresponding scattering systems are
asymptotically complete. Note also that the Kato-Rosenblum theorem leads to the
existence and to the completeness of the wave operators. On the other hand, Cook
criterion, as presented in Proposition 5.2.4, does not provide any information about the
completeness or the asymptotic completeness of the wave operators.
1
Be aware that this terminology is not completely fixed and can still depend on the authors.
5.3. SCATTERING OPERATOR AND COMPLETENESS 75

Let us close this section with a variant of the spectral theorem. The following
formulation is a little bit imprecise because a fully rigorous version needs some more
information on the structure of direct integrals of Hilbert spaces and on the notion of
the multiplicity theory. We refer to [BM, Chap. 2] for the notion of multiplicity, and to
the same book but Chapter 4 and 5 for more information on direct integral of Hilbert
spaces and of operators. We also refer to [Yaf, Sec. 1.5] for a very short presentation of
the same material.
For a σ-finite measure m on (R, AB ) we define
∫ ⊕
H := H(λ) m(dλ) (5.12)
R

as the Hilbert space of equivalence class of vector-valued functions R ∋ λ 7→ f(λ) ∈ H(λ)

taking values in the Hilbert space H(λ) and which are measurable and square integrable
with respect to the measure m. The scalar product in H is given by
∫
⟨f, g⟩H := ⟨f(λ), g(λ)⟩λ m(dλ)
R

with ⟨·, ·⟩λ the scalar product in H(λ). If the fiber H(λ) is a constant Hilbert space H
independent of λ, then the above construction corresponds to L2 (R, m; H) ∼= L2 (R, m)⊗
H.
In this context, one of the formulation of the spectral theorem can be expressed
as a decomposition of any self-adjoint operator into a direct integral of operator. More
precisely, for any self-adjoint operator H in a Hilbert space H there exists a measure
σ-finite measure m on R and a unitary transformation F : H → H such that
∫
⟨E(V )f, g⟩H = ⟨[F f ](λ), [F g](λ)⟩λ m(dλ),
V

where E(·) is the spectral measure associated with H and V is any Borel set on R. A
different way of writing the same information is by saying that H is a diagonal operator
in the direct integral representation provided by H . In other words, the following
equality holds: ∫ ⊕
∗
F HF = λ m(dλ).
R

Note that such a decomposition is called the direct integral representation of H. This
representation is often highly non-unique, but in applications some natural choices often
appear. In addition, if H is purely absolutely continuous, then the measure m can be
chosen as the Lebesgue measure, as mentioned in [BM, Sec. 5.2.4]. In any case, the
support of the measure m coincides with the spectrum of H, and thus we can restrict
the above construction to the spectrum σ(H) of H.
Once the notion of a direct integral Hilbert space H is introduced, as in (5.12), di-
rect integral operators operators acting on this Hilbert space can naturally be studied.
76 CHAPTER 5. SCATTERING THEORY

We refer again to [BM] for more information, or to [RS4, XIII.16] for a short intro-
duction to this theory and a few important results. Our only aim in this direction is
the statement of the following result. Note that its proof is based on the commutation
relation provided in (5.10) and that such an argument is quite standard.

Proposition 5.3.3. Let H0 be an absolutely continuous self-adjoint operator in a

Hilbert space H and let F0 and H0 be a direct integral representation of H0 , i.e. H0
is a direct integral Hilbert space as constructed in (5.12) with ∫m the Lebesgue measure,
⊕
and F0 : H → H0 is a unitary map satisfying F0 H0 F0∗ = σ(H0 ) λ dλ. Let H be an-
other self-adjoint operator in H such that the wave operators W± (H, H0 ) exist and are
asymptotically complete. Then there exists a family {S(λ)}λ∈σ(H0 ) of unitary operator
in H(λ) for almost every λ such that
∫ ⊕
∗
F SF = S(λ) dλ.
σ(H0 )

The operator S(λ) is called the scattering matrix at energy λ even if S(λ) is usually
not a matrix but a unitary operator in H(λ). Note that there also exist expressions for
the operators S and the operator S(λ) in terms of the difference of the resolvent of H
and the resolvent of H0 on the real axis. Such expressions are usually referred to as the
stationary approach of scattering theory. This approach will not be developed here, but
the reference [Yaf] is one of the classical book on the subject.

Extension 5.3.4. Work on the notion of multiplicity and on the theory of direct integral
of Hilbert spaces and direct integral of operators.
Chapter 6

Commutator methods

Let us consider a self-adjoint operator H in a Hilbert space H. As seen in Section 4.4

the resolvent (H − λ ∓ iε)−1 does not possess a limit in B(H) as ε ↘ 0 if λ ∈ σ(H).
However, the expression ⟨f, (H − λ ∓ iε)−1 f ⟩ may have a limit as ε ↘ 0 for suitable
f . In addition, if this limit exists for sufficiently many f , then H is likely to have only
absolutely continuous spectrum around λ, see Proposition 4.4.2 for a precise statement.
Our aim in this chapter is to present a method which allows us to determine if
the spectrum of H is purely absolutely continuous in some intervals. This method is
an extension of Theorem 4.4.3 of Putnam which is valid if both operators H and A
are unbounded. Again, the method relies on the positivity of the commutator [iH, A],
once this operator is well-defined and localized in the spectrum of H. In fact, it was E.
Mourre who understood how the method of Putnam can be sufficiently generalized.
Since the proof of the main result is rather long and technical, we shall first state
the main result in a quite general setting. Then, the various tools necessary for under-
standing this result will be presented, as well as some of its corollaries. Only at the end
of the chapter, a proof will be sketched, or presented in a restricted setting.

6.1 Main result

As mentioned above, we will state the main results of the chapter even if it is not
fully understandable yet. Additional explanations will be provided in the subsequent
sections. Let us however introduce very few information. A self-adjoint operator H in
a Hilbert space H has a gap if σ(H) ̸= R. In the sequel, we shall give a meaning to
the requirement “H is of class C 1 (A) or of class C 1,1 (A)”, but let us mention that
the condition H being of class C 1 (A), and a fortiori of class C 1,1 (A), ensures that the
commutator [iH, A], between two unbounded self-adjoint operators, is well-defined in a
sense explained later on.
The following statement corresponds to [ABG, Thm. 7.4.2].

Theorem 6.1.1. Let H be a self-adjoint operator in H, of class C 1,1 (A) and having a
spectral gap. Let J ⊂ R be open and bounded and assume that there exist a > 0 and

77
78 CHAPTER 6. COMMUTATOR METHODS

K ∈ K (H) such that

E H (J)[iH, A]E H (J) ≥ aE H (J) + K.

Then H has at most a finite number of eigenvalues in J, multiplicity counted, and has
no singular continuous spectrum in J.
In fact, this statement is already a corollary of a more general result that we provide
below. For its statement, let us still introduce some information.
( If) A is a second self-
adjoint operator in H, with domain D(A), let us set G := D(A), H 1/2,1 for the Banach
space obtained by interpolation between D(A) and H (explained later on). Its dual space
is denoted by G ∗ and one has G ⊂ H ⊂ G ∗ with dense and continuous embeddings, as
well as B(H) ⊂ B(G, G ∗ ). If H is of class C 1 (A) we also define the subset µA (H) by
{ }
µA (H) := λ ∈ R | ∃ε > 0, a > 0 s.t. E(λ; ε)[iH, A]E(λ; ε) ≥ aE(λ; ε) , (6.1)
( )
where E(λ; ε) := E H (λ − ε, λ + ε) .
The following statement is a slight reformulation of [ABG, Thm 7.4.1].
Theorem 6.1.2. Let H be a self-adjoint operator in H and assume that H has a gap
and is of class C 1,1 (A). Then, for each λ ∈ µA (H) the limits limε↘0 ⟨f, (H − λ ∓ iε)−1 f ⟩
exist for any f ∈ G and uniformly on each compact subset of µA (H). In particular, if
T is a linear operator from H to an auxiliary Hilbert space, and if T is continuous when
H is equipped with the topology induced by G ∗ , then T is locally H-smooth on the open
set µA (H).
Note that the notion of H-smooth operator will be introduced later on, but that
these operators play an important role for proving the existence and the completeness
of some wave operators.
Remark 6.1.3. In the above two statements, it is assumed that H has a spectral gap,
which is a restricting assumption since there also exist operators H with σ(H) = R.
For example, the operator X of multiplication by the variable in L2 (R) has spectrum
equal to R. However, there also exists a version of the Theorems 6.1.1 and 6.1.2 which
do not require the existence of a gap. The main interest in the gap assumption is that
there exists λ0 ∈ R such that (H − λ0 )−1 is bounded and self-adjoint. This operator can
then be used in the proofs and this fact is quite convenient. If such a λ0 does not exist
proofs are a little bit more involved.
Our main task now is to introduce all the notions such that the above statements
become fully understandable.

6.2 Regularity classes

Most of the material presented in this section is borrowed from Chapter 5 of [ABG] to
which we refer for more information and for a presentation in a more general setting.
6.2. REGULARITY CLASSES 79

Let us consider a self-adjoint operator A in H which generates the strongly contin-

uous unitary group {e−itA }t∈R . We also consider a bounded operator S in H. In this
setting, the map

R ∋ t 7→ Ut [S] ≡ S(t) := eitA S e−itA ∈ B(H) (6.2)

is well-defined and its regularity can be studied. In fact, Ut : B(H) → B(H) defines a
weakly continuous representation of R in B(H).
We start by some conditions of regularity indexed by a positive integer k.

Definition 6.2.1. Let k ∈ N.

(i) C k (A) denotes the Banach space of all S ∈ B(H) such that the map (6.2) is
k-times strongly continuously differentiable, and endowed with the norm

(∑
k )1/2
∥S∥C k := ∥S (0)∥
(j) 2
, (6.3)
j=0

where S (j) (0) denotes the j th derivative of S(t) evaluated at t = 0.

(ii) Cuk (A) denotes the Banach space of all S ∈ B(H) such that the map (6.2) is
k-times continuously differentiable in norm, and endowed with the norm defined
by (6.3).

It can be shown that these spaces are indeed complete and that Cuk (A) ⊂ C k (A). An
equivalent description of the elements of C k (A) or Cuk (A) is provided in the following
statement, see [ABG, Thm. 5.1.3] for its proof.

Proposition 6.2.2. Let k ∈ N∗ and S ∈ B(H). Then S belongs to C k (A) or to Cuk (A)
if and only if limt↘0 t−k (Ut − 1)k [S] exists in the strong or in the norm topology of
B(H).

Note now that a formal computation of dt d

S(t)|t=0 gives S ′ (0) = [iA, S]. However,
this formula has to be taken with some care since it involves the operator A which
is often unbounded. In fact, a more precise and alternative description of the C 1 (A)
condition is often very useful, see [ABG, Lem. 6.2.9] for its proof.

Lemma 6.2.3. The bounded operator S belongs to C 1 (A) if and only if there exists a
constant c < ∞ such that

⟨Af, iSf ⟩ − ⟨S ∗ f, iAf ⟩ ≤ c∥f ∥2 , ∀f ∈ D(A). (6.4)

In fact, the expression ⟨Af, iSf ⟩ − ⟨iS ∗ f, Af ⟩ defines a quadratic form with domain
D(A), and the condition (6.4) means precisely that this form is bounded. Since D(A)
is dense in H, this form extends to a bounded form on H, and there exists a unique
80 CHAPTER 6. COMMUTATOR METHODS

operator in B(H) which corresponds to this form. For an obvious reason we denote this
bounded operator by [iA, S] and the following equality holds for any f ∈ D(A)

⟨Af, iSf ⟩ − ⟨S ∗ f, iAf ⟩ = ⟨f, [iA, S]f ⟩.

However, let us stress that the bounded operator [iA, S] has a priori no explicit expres-
sion on all f ∈ H.
Let us add some rather simple properties of the class C k (A) and Cuk (A).
Proposition 6.2.4. (i) If S belongs to C k (A) then S ∈ Cuk−1 (A),

(ii) If S, T ∈ C k (A) then ST ∈ C k (A), and if S, T ∈ Cuk (A) then ST ∈ Cuk (A),

(iii) If S is boundedly invertible, then S ∈ C k (A) ⇐⇒ S −1 ∈ C k (A), and S ∈

Cuk (A) ⇐⇒ S −1 ∈ Cuk (A),

(iv) S ∈ C k (A) ⇐⇒ S ∗ ∈ C k (A), and S ∈ Cuk (A) ⇐⇒ S ∗ ∈ Cuk (A).

Exercise 6.2.5. Provide a proof of the above statements. Note that compared with the
general theory presented in [ABG, Sec. 5.1] we are dealing only with a one-parameter
unitary group, and only in the Hilbert space H. Multiparameter C0 -group acting on
arbitrary Banach spaces are avoided in these notes.
Let us now present some regularity classes of fractional order. Consider s ≥ 0,
p ∈ [0, ∞] and let ℓ ∈ N with ℓ > s. We can then define
(∫ )
p dt 1/p
∥S∥s,p := ∥S∥ +
(ℓ)
|t|−s (Ut − 1)ℓ [S] ,
|t|≤1 |t|

with the convention that the integral is replaced by a sup when p = ∞. It can then be
(ℓ′ )
shown that if ∥S∥s,p < ∞ then ∥S∥s,p < ∞ for any integer ℓ′ > s. For that reason, the
(ℓ)

following definition is meaningful:

Definition 6.2.6. For any s ≥ 0 and p ∈ [0, ∞] we set C s,p (A) for the set of S ∈ B(H)
(ℓ)
such that ∥S∥s,p < ∞ for some (and them for all) ℓ ∈ N with ℓ > s. For two different
(ℓ′ )
integers ℓ, ℓ′ > s, the maps S 7→ ∥S∥s,p and S 7→ ∥S∥s,p define equivalent norms on
(ℓ)

C s,p (A). Endowed with any of these norms, C s,p (A) is a Banach space.
Let us state some relations between these spaces and the spaces C k (A) and Cuk (A)
introduced above. All these relations are proved in a larger setting in [ABG, Sec. 5.2].
Proposition 6.2.7. Let k ∈ N, s, t ≥ 0 and p, q ∈ [1, ∞].
(i) C s,p (A) ⊂ C t,q (A) if s > t and for p, q arbitrary,

(ii) C t,p (A) ⊂ C t,q (A) if q > p and in particular for any p ∈ (1, ∞)

C s,1 (A) ⊂ C s,p (A) ⊂ C s,∞ (A), (6.5)

6.2. REGULARITY CLASSES 81

(iii) If s = k is an integer one has

C k,1 (A) ⊂ Cuk (A) ⊂ C k (A) ⊂ C k,∞ (A).

Note that a very precise formulation of the differences between C k,1 (A), Cuk (A),
C (A) and C k,∞ (A) is presented in [ABG, Thm. 5.2.6]. Another relation between some
k

of these spaces is also quite convenient:

Proposition 6.2.8. Let s ∈ (0, ∞) and p ∈ [1, ∞], and write s = k + σ with k ∈ N
and 0 < σ ≤ 1.
(i) If S ∈ C s,p (A) and j ≤ k, then S (j) (0) ∈ C s−j,p (A),

(ii) If ∥ · ∥C σ,p is one norm on C σ,p (A), then

∥S∥C k + ∥S (k) (0)∥C σ,p

defines a norm on C s,p (A). In particular S ∈ C s,p (A) if and only if S ∈ C k (A)
and S (k) (0) belongs to C σ,p (A).
Relations similar to the one presented in Proposition 6.2.4 also hold in the present
context:
Proposition 6.2.9. (i) If S, T ∈ C s,p (A), then ST ∈ C s,p (A),

(ii) If S is boundedly invertible, then S ∈ C s,p (A) ⇐⇒ S −1 ∈ C s,p (A),

(iii) S ∈ C s,p (A) ⇐⇒ S ∗ ∈ C s,p (A).

Let us still mention one more regularity class with respect to A which is quite con-
venient in applications. The additional continuity condition is related to Dini continuity
in classical analysis. For any integer k ≥ 1 we set C k+0 (A) for the set of S ∈ C k (A) and
such that S (k) (0) satisfies
∫
dt
(Ut − 1)[S (k) (0)] < ∞.
|t|≤1 |t|

Once endowed with the norm

∫
dt
∥S∥C k+0 := ∥S∥ + (Ut − 1)[S (k) (0)]
|t|≤1 |t|

the set C k+0 (A) becomes a Banach space and the following relations hold for any k ∈ N

C k+0 (A) ⊂ C k,1 (A) ⊂ Cuk (A) ⊂ C k (A) ⊂ C k,∞ (A) (6.6)

with C 0+0 (A) := C 0,1 (A).

Up to now, we have considered only bounded elements S. In the next section we
show how these notions can be useful for unbounded operators as well.
82 CHAPTER 6. COMMUTATOR METHODS

6.3 Affiliation
In this section we consider two self-adjoint operators A and H in a Hilbert space H.
The various regularity classes introduced before are defined in term of the unitary group
generated by A, and the next definition gives a meaning to the regularity of the operator
H with respect to A, even if H is unbounded. Before this definition, let us state a simple
lemma whose proof depends only on the first resolvent equation and on some analytic
continuation argument, see [ABG, Lem. 6.2.1] for the details.

Lemma 6.3.1. Let k ∈ N, s ≥ 0 and p ∈ [0, ∞], and let H, A be self-adjoint operators
in H. Assume that there exists z0 ∈ C such that (H − z0 )−1 belongs to C k (A), Cuk (A),
or to C s,p (A). Then (H − z)−1 belongs to the same regularity class for any z ∈ ρ(H).
In addition, if H is bounded, then H itself belongs to the same regularity class.

The following definition becomes then meaningful.

Definition 6.3.2. Let k ∈ N, s ≥ 0 and p ∈ [0, ∞], and let H, A be self-adjoint

operators in H. We say that H is of class C k (A), Cuk (A), or C s,p (A) if (H − z)−1 belongs
to such a regularity class for some z ∈ ρ(H), and thus for all z ∈ ρ(H).

Clearly, if the resolvent of H belongs to one of these regularity classes, then the
same holds for linear combinations of the resolvent for different values of z ∈ ρ(H). In
fact, by functional calculus one can show that η(H) belongs to the same regularity class
for suitable function η : R → R. We state such a result for a rather restricted class of
functions η and refer to Theorem 6.2.5 and Corollary 6.2.6 of [ABG] for a more general
statement. Note that the proof is rather technical and depends on an explicit formula
for the operator η(H) in terms of the resolvent of H. For completeness we provide such a
formula but omit all the details and the explanations. For suitable functions η : R → R
the following formula holds:

∑ ∫
r−1
1 ( )
η(H) = η (k) (λ)ℑ ik (H − λ − i)−1 dλ
k=0
πk! R
∫ 1 ∫
1 ( )
+ r−1
µ dµ η (r) (λ)ℑ ir (H − λ − iµ)−1 dλ. (6.7)
π(r − 1)! 0 R

Proposition 6.3.3. Assume that H is of class C k (A), Cuk (A), or C s,p (A), and let η be
a real function belonging to Cc∞ (R). Then η(H) belongs to the same regularity class.

Let us now mention how the condition H is of class C 1 (A) can be checked. For
a bounded operator H, this has already been mentioned in Lemma 6.2.3. For an un-
bounded operator H the question is more delicate. We state below a quite technical
result. Note that the invariance of the domain of H with respect to the unitary group
generated by A is often assumed, and this assumption simplifies quite a lot the argu-
mentation. However, in the following statement such an assumption is not made.
6.3. AFFILIATION 83

Theorem 6.3.4 (Thm. 6.2.10 of [ABG]). Let H and A be self-adjoint operators in a

Hilbert space H.
a) The operator H is of class C 1 (A) if and only if the following two conditions hold:
(i) There exists c < ∞ such that for all f ∈ D(A) ∩ D(H)
( )
⟨Af, Hf ⟩ − ⟨Hf, Af ⟩ ≤ c ∥Hf ∥2 + ∥f ∥2 ,

(ii) For some z ∈ C \ σ(H) the set

{ }
f ∈ D(A) | (H − z)−1 f ∈ D(A) and (H − z̄)−1 f ∈ D(A)

is a core for A.
b) If H is of class C 1 (A), then the D(A) ∩ D(H) is a core for H and the form [A, H]
has a unique extension to a continuous sesquilinear form on D(H) endowed with the
graph topology1 . If this extension is still denoted by [A, H], then the following identify
holds on H:
[A, (H − z)−1 ] = −(H − z)−1 [A, H](H − z)−1 . (6.8)
Let us still mention how the equality (6.8) can be understood. If we equip D(H) with
the graph topology (it is then a Banach space), and denote by D(H)∗ its dual space,
then one has the following dense inclusions D(H) ⊂ H ⊂ D(H)∗ , and the operator
R(z) is bounded from H to D(H) and extends to a bounded operator from D(H)∗ to
H. Then, the fact that [A, H] has a unique extension to a sesquilinear form on D(H)
means that its continuous extension [A, H] is a bounded operator from D(H) to D(H)∗ .
Thus, the r.h.s. of (6.8) corresponds to the product of three bounded operators

[A, (H − z)−1 ] = − (H − z)−1 [A, H] (H − z)−1 (6.9)

| {z } | {z } | {z }
D(H)∗ →H D(H)→D(H)∗ H→D(H)

which corresponds to a bounded operator in H. By setting R(z) := (H − z)−1 , formula

(6.9) can also re rewritten as

[H, A] = (H − z)[A, R(z)](H − z) (6.10)

where the r.h.s. is the product of three bounded operators, namely (H −z) : D(H) → H,
[A, R(z)] : H → H and (H − z) : H → D(H)∗ .
Another formula will also be useful later on. For τ ̸= 0 let us set Aτ := iτ1 (eiτ A −1),
and observe that if H is of class C 1 (A) one has for any z ∈ ρ(H)
1 ( iτ A )
[A, R(z)] = s − lim e R(z) e−iτ A −R(z)
τ →0 iτ
1 [ iτ A ]
= s − lim e , R(z) e−iτ A = s − lim [Aτ , R(z)] . (6.11)
τ →0 iτ τ →0
(
1
The graph topology on D(H) corresponds to the topology obtained by the norm ∥f ∥D(H) = ∥f ∥2 +
)1/2
∥Hf ∥2 for any f ∈ D(H).
84 CHAPTER 6. COMMUTATOR METHODS

In relation with formula (6.9) let us observe that if J ⊂ R is a bounded

( Borel)
set, then E (J) is obviously an element of B(H) but it also belongs to B H, D(H) .
H

Indeed, this fact follows from the boundedness of the operator HE H (J). By duality, it
∗
also follows that the operator
( ∗
) extends to a bounded operator from D(H) to H,
E H (J)
or in short E (J) ∈ B D(H) , H . This fact is of crucial importance. (Indeed, if H is)
H

of class C 1 (A), then as shown above the operator [iH, A] belongs to B D(H), D(H)∗
and therefore the product
E H (J)[iH, A]E H (J)
belongs to B(H). Such a product was already mentioned in Section 6.1 for the special
choice J = (λ − ε, λ + ε) for some fixed λ ∈ R and ε > 0.
We now introduce an easy result which is often called the Virial theorem. Note that
this result is often stated without the appropriate assumption.

Proposition 6.3.5. Let H and A be self-adjoint operator in H such that H is of class

C 1 (A). Then E H ({λ})[A, H]E H ({λ}) = 0 for any λ ∈ R. In particular, if f is an
eigenvector of H then ⟨f, [A, H]f ⟩ = 0.

Proof. We must show that if λ ∈ R and f1 , f2 ∈ D(H) satisfy Hfk = λfk for k = 1, 2,
then ⟨f1 , [A, H]f2 ⟩ = 0. Since f1 = (λ − i)(H − i)−1 f1 and f2 = (λ + i)(H + i)−1 f2 we
get by (6.10) and (6.11) that

⟨f1 , [A, H]f2 ⟩ = −(λ + i)2 ⟨f1 , [A, (H + i)−1 ]f2 ⟩

{ }
−1 −1
= −(λ + i) lim ⟨f1 , Aτ (H + i) f2 ⟩ − ⟨(H − i) f1 , Aτ f2 ⟩ .
2
τ →0

We finally observe that for τ ̸= 0, the term into curly brackets is always equal to 0.

In relation with the definition of µA (H) mentioned in (6.1) let us still introduce
two functions which play an important role in Mourre theory. These functions are well
defined if the operator H is of class C 1 (A). They provide what is called the best Mourre
estimate, and the first one is defined for any λ ∈ R by
{ }
ϱAH (λ) := sup a ∈ R | ∃ε > 0 s.t. E(λ; ε)[iH, A]E(λ; ε) ≥ aE(λ; ε) .

The second function looks similar to the previous one, but the inequality holds modulo
a compact operator, namely

ϱ̃A
H (λ)
{ }
:= sup a ∈ R | ∃ε > 0 and K ∈ K (H) s.t. E(λ; ε)[iH, A]E(λ; ε) ≥ aE(λ; ε) + K .

The relation between ϱA A

H and µ (H) is rather clear. By looking back to the definition
of (6.1) one gets
µA (H) = {λ ∈ R | ϱA
H (λ) > 0}.

Many properties of these functions have been deduced in [ABG, Sec. 7.2].
6.3. AFFILIATION 85

Proposition 6.3.6. (i) The function ϱA H : R → (−∞, ∞] is lower semicontinuous

and ϱH (λ) < ∞ if and only if λ ∈ σ(H),
A

H : R → (−∞, ∞] is lower semicontinuous and satisfies ϱ̃H ≥ ϱH .

(ii) The function ϱ̃A A A

H (λ) < ∞ if and only if λ ∈ σess (H).

Furthermore ϱ̃A

Let us now state and prove a corollary of the Virial theorem showing that when ϱ̃A
H
is strictly positive, then only a finite number of eigenvalues can appear.

Corollary 6.3.7. Let H and A be self-adjoint operator in H such that H is of class

H (λ) > 0 for some λ ∈ R then λ has a neighbourhood in which there is at
C 1 (A). If ϱ̃A
most a finite number of eigenvalues of H, each of finite multiplicity.

Proof. Let ε > 0, a > 0 and K ∈ K (H) such that

E(λ; ε)[iH, A]E(λ; ε) ≥ aE(λ; ε) + K. (6.12)

If g is an eigenvector of H associated with an eigenvalue in (λ − ε, λ + ε) and if ∥g∥ =

1, then (6.12) and the Virial theorem imply that ⟨g, Kg⟩ < −a. By contraposition,
assume that the statement of the lemma is false. Then there exists an infinite orthogonal
sequence {gj } of eigenvectors of H in E(λ; ε)H. In particular, w − limj→∞ gj = 0, as
a consequence of the orthogonality of the sequence. However, since K is compact, Kgj
goes strongly to 0 as j → ∞, and then one has ⟨gj , Kgj ⟩ → 0 as j → ∞. This contradicts
the inequality ⟨gj , Kgj ⟩ ≤ −a < 0.

The previous result can then be used for showing that the two functions ϱA A
H and ϱ̃H
are in fact very similar. The proof of the following statement can be found in [ABG,
Thm. 7.2.13].

Theorem 6.3.8. Let H and A be self-adjoint operator in H such that H is of class

C 1 (A), and let λ ∈ R. If λ is an eigenvalue of H and ϱ̃A A
H (λ) > 0, then ϱH (λ) = 0.
A A
Otherwise ϱH (λ) = ϱ̃H (λ).

As a conclusion of this section let us mention a result about the stability of the
function ϱ̃. Once this stability is proved, the applicability of the previous corollary is
quite enlarged.

Theorem 6.3.9. Let H, H0 and A be self-adjoint operators in a Hilbert space H and

such that H0 and H are of class Cu1 (A). If the difference (H + i)−1 − (H0 + i)−1 is
compact, then ϱ̃A A
H (λ) = ϱ̃H0 (λ).

Note that the content of Proposition 6.3.3 for η(H) is necessary for the proof of
this statement, and that the Cu1 (A)-condition can not be weakened. We refer to [ABG,
Thm. 7.2.9] for the details.
86 CHAPTER 6. COMMUTATOR METHODS

6.4 Locally smooth operators

An important ingredient for showing the absence of singular continuous spectrum and
for proving the existence and the completeness of some wave operators is the notion
of locally smooth operators. Such operators were already mentioned in the statement
of Theorem 6.1.2 and we shall now provide more information on them. Note that an
operator T is always smooth with respect to another operator, it is thus a relative notion.
( )
Definition 6.4.1. Let J ⊂ R be an open set, and let H, D(H) be a self-adjoint
operator in H. A linear continuous operator T : D(H) → H is locally H-smooth on J
if for each compact subset K ⊂ J there exists a constant CK < ∞ such that
∫ ∞
∥T e−itH f ∥2 dt ≤ CK ∥f ∥2 , ∀f ∈ E H (K)H. (6.13)
−∞

In the next statement, we shall show that this notion can be recast in a time-
independent framework. However some preliminary observations are necessary. As al-
ready mentioned in the previous section, by endowing D(H) with its graph norm, one
gets the continuous and dense embeddings

D(H) ⊂ H ⊂ D(H)∗ ,

−1
( )
and
( a continuous
) extension of R(z) ≡ (H − z) ∈ B H, D(H) to an element R(z) ∈
B D(H)∗ , H , for any z ∈ ρ(H). By choosing z = λ + iµ with µ > 0 one observes that

1
δ(µ) (H − λ) := ℑR(λ + iµ)
π
1 ( ) µ
= R(λ + iµ) − R(λ − iµ) = R(λ ∓ iµ)R(λ ± iµ)
2πi π
( ) ( )
and infers that δ(µ) (H − λ) ∈ B D(H)∗ , D(H) . Therefore, if T ∈ B D(H), H then
( )∗
T R(z) ∈ B(H) and R(z)T ∗ = T R(z̄) ∈ B(H), and consequently T δ(µ) (H − λ)T ∗ ∈
B(H). Finally, by the C ∗ -property ∥S ∗ S∥ = ∥S ∗ ∥2 = ∥S∥2 one deduces that
µ µ
∥T δ(µ) (H − λ)T ∗ ∥ = ∥T R(λ ∓ iµ)∥2 = ∥R(λ ± iµ)T ∗ ∥2 .
π π

We are now ready to state:

Proposition 6.4.2. A linear continuous operator T : D(H) → H is locally H-smooth

on an open set J if and only if for each compact subset K ⊂ J there exists a constant
CK < ∞ such that

T ℑR(z)T ∗ ≤ CK if ℜ(z) ∈ K and 0 < ℑ(z) < 1. (6.14)

6.4. LOCALLY SMOOTH OPERATORS 87

The proof of this statement is not difficult but a little bit too long. The main
ingredients are the equalities
∫ ±∞
R(λ ± iµ) = i eitλ e−itH−µ|t| dt
0

and ∫ ∞
1
δ(µ) (H − λ) = eitλ e−itH−µ|t| dt
2π −∞
which were already mentioned in Remark 5.1.2.
Exercise 6.4.3. Provide the proof of the above statement, which can be borrowed from
[ABG, Prop. 7.1.1].
We shall immediately provide a statement which shows the importance of this notion
of local smoothness for the existence and the completeness of the wave operators.
Theorem 6.4.4. Let H1 , H2 be two self-adjoint operators in a Hilbert space H, with
spectral measure
( ) respectively by E1 and E2 . Assume that for j ∈ {1, 2} there
denoted
exist Tj ∈ B D(Hj ), H which satisfy

⟨H1 f1 , f2 ⟩ − ⟨f1 , H2 f2 ⟩ = ⟨T1 f1 , T2 f2 ⟩ ∀fj ∈ D(Hj ).

If in addition there exists an open set J ⊂ R such that Tj are locally Hj -smooth on J,
then ( )
W± H1 , H2 , E2 (J) := s − lim eitH1 e−itH2 E2 (J) (6.15)
t→±∞

exist and are bijective isometries of E2 (J)H onto E1 (J)H.

Proof. The existence of the limits (6.15) is a simple consequence of the following asser-
tion: for any f2 ∈ H such that E2 (K2 )f2 = f2 for some compact set K2 ⊂ J, and for
any θ1 ∈ Cc∞ (J) with θ1 (x) = 1 for any x in a neighbourhood of K2 the limit

s − lim θ1 (H1 ) eitH1 e−itH2 f2 (6.16)

t→±∞

exists, and [ ]
s − lim 1 − θ1 (H1 ) eitH1 e−itH2 f2 = 0. (6.17)
t→±∞

Let us now prove (6.16). For that purpose we set W (t) := θ1 (H1 ) eitH1 e−itH2 and
observe that for any f1 ∈ H and s < t:
∫ t
|⟨f1 , [W (t) − W (s)]f2 ⟩| = ⟨T1 e−iτ H1 θ1 (H1 )f1 , T2 e−iτ H2 f2 ⟩dτ
[∫ t ]1/2 [ ∫ t
s
]1/2
−iτ H1
≤ ∥T1 e θ1 (H1 )f1 ∥ dτ
2
∥T2 e−iτ H2 f2 ∥2 dτ
[∫ t
s s
]1/2
≤ CK1 ∥f ∥ ∥T2 e−iτ H2 f2 ∥2 dτ
s
88 CHAPTER 6. COMMUTATOR METHODS

with K1 = supp θ1 . We thus obtain that ∥[W (t) − W (s)]f2 ∥ → 0 as s → ∞ or t → −∞,

which proves (6.16).
For the proof of (6.17) let θ2 ∈ Cc∞ (J) with θ2 (x) = 1 if x ∈ K2 and such that
θ1 θ2 = θ2 . Then f2 = θ2 (H2 )f2 and [1 − θ1 (H1 )]θ2 (H2 ) = [1 − θ1 (H1 )][θ2 (H2 ) − θ2 (H1 )].
Hence (6.17) follows from

s − lim [θ2 (H2 ) − θ2 (H1 )] e−itH2 f2 = 0. (6.18)

|t|→∞

In fact, we shall prove this estimate for any θ2 ∈ C0 (R). Let us set rz (x) := (x − z)−1 for
any x ∈ R and z ∈ C \ R. Since the vector space generated by the family of functions
{rz }z∈C\R is a dense subset of C0 (R) it is enough to show (6.18) with θ2 replaced by rz
for any z ∈ C \ R. Set Rj = (Hj − z)−1 for some fixed z ∈ C \ R and observe that for
any gj ∈ H

|⟨g1 , (R1 − R2 )g2 ⟩| = |⟨R1∗ g1 , H2 R2 g2 ⟩ − ⟨H1 R1∗ g1 , R2 g2 ⟩|

= |⟨T1 R1∗ g1 , T2 R2 g2 ⟩|
≤ ∥T1 R1∗ ∥ ∥g1 ∥ ∥T2 R2 g2 ∥.

Taking g2 = e−itH2 f2 we see that it is enough to prove that ∥T2 R2 e−itH2 f2 ∥ → 0 as

|t| → ∞. But this is an easy consequence of the fact that both the function F (t) :=
T2 R2 e−itH2 f2 and its derivative are square integrable on R.
As a consequence of the previous arguments, we have thus obtained
( that (6.16)
)
exists. Clearly the same arguments apply for the existence of W± H2 , H1 , E1 (J) . It
( ) ( )∗
then follows that W± H2 , H1 , E1 (J) = W± H1 , H2 , E2 (J) from which one deduces
the final statement, see also Proposition 5.2.3.

6.5 Limiting absorption principle

Since the utility of locally smooth operators has been illustrated in the previous theo-
rem, it remains to show how such operators can be exhibited. In this section, we provide
this kind of information, and start with the so-called limiting absorption principle.
By looking back to the equation (6.14) and by assuming that T ∈ B(H), one
observes that the main point is to obtain an inequality of the form
⟨ ⟩
f, ℑR(λ + iµ)f ≤ CK ∥f ∥2G (6.19)

for some compact set K ⊂ R, all λ ∈ K and µ > 0. Note that we have used the notation
G := T ∗ H endowed with the norm ∥f ∥G := inf{∥g∥ | T ∗ g = f } for any f ∈ G.
Let us be a little bit more general. Consider any Banach space G such that G ⊂
D(H)∗ continuously and densely. By duality it implies the existence of a continuous
embedding D(H) ⊂ G ∗ , but this embedding is not dense in general. The closure of
D(H) inside G ∗ is thus denoted by G ∗◦ , and is equipped with the Banach space structure
6.5. LIMITING ABSORPTION PRINCIPLE 89
( )
inherited from G ∗ . It then follows that B D(H)∗ , D(H) ⊂ B(G, G ∗◦ ) ⊂ B(G, G ∗ ) and
in particular
( )
ℑR(z) ∈ B D(H)∗ , D(H) ⊂ B(G, G ∗◦ ) ⊂ B(G, G ∗ ), ∀z ∈ ρ(H).
Definition 6.5.1. Let J be an open set, H a self-adjoint operator and G ⊂ D(H)∗ .
i) The generalized limiting absorption holds for H in G and locally on J if for each
compact subset K ⊂ J there exists CK < ∞ such that (6.19) holds for all f ∈ G, any
λ ∈ K and µ > 0, or equivalently if
sup ∥ℑR(λ + iµ)∥B(G,G ∗ ) < ∞
λ∈K,µ>0

for any compact subset K ⊂ J.

ii) The strong generalized limiting absorption holds for H in G and locally on J if
lim ⟨f, ℑR(λ + iµ)f ⟩ =: ⟨f, ℑR(λ + i0)f ⟩
µ↘0

exists for any λ ∈ J and f ∈ G, uniformly in λ on any compact subset of J.

Note that by an application of the uniform boundedness principle the generalized
limiting absorption (GLAP) holds if the strong GLAP is satisfied. In the next statement
we shall make the link between GLAP and locally smooth operators. For that purpose,
we first recall a consequence of Stone’s formula, see Proposition 4.4.1:
∫ b
( ) 1 1
E (a, b) + E({a}) + E({b}) = w − lim δ(µ) (H − λ)dλ. (6.20)
2 2 µ↘0 a

Namely, for any a < b and f ∈ H one has

1 ( ) 1 ( ) 2
mf (a, b) = E (a, b) f
b−a b−a
1
≤ sup ⟨f, δ(µ) (H − λ)f ⟩ = sup ⟨f, ℑR(λ + iµ)f ⟩. (6.21)
a<λ<b, µ>0 a<λ<b, µ>0 π

Thus, if J ⊂ R is open and | π1 ⟨f, ℑR(λ + iµ)f ⟩| ≤ C(f ) < ∞ for all λ ∈ J and
µ > 0, then
(
d
mf is) absolutely continuous on J and dλ ∥Eλ f ∥2 ≤ C(f ) on J (recall that
Eλ = E (−∞, λ] ) If this holds for each f in a dense subset of H then the spectrum of
H in J is purely absolutely continuous.
Proposition 6.5.2. Let G be a Banach space with G ⊂ D(H ∗ ) continuously and densely,
and let J ⊂ R be open.
i) If the GLAP holds for H locally on J, then H has purely absolutely continuous
spectrum in J. If the strong GLAP holds for H in G and locally on J, then for each
λ0 ∈ R the function λ 7→ Eλ − Eλ0 ∈ B(G, G ∗ ) is weak*-continuously differentiable on
J, and its derivative is equal to
d 1
Eλ = ℑR(λ + i0). (6.22)
dλ π
90 CHAPTER 6. COMMUTATOR METHODS

ii) Assume that (G ∗◦ )∗ = G and that the GLAP holds in G locally on J. Let T :
D(H) → H be a linear operator which is continuous when D(H) is equipped with the
topology induced by G ∗ , or in other terms let T ∈ B(G ∗◦ , H), then T is locally H-smooth
on J.
Proof. i) Let us first note that for any f ∈ D(H)∗ , the expression ∥E(·)f ∥2 is a well-
defined
( positive )Radon measure on R, since for any bounded J ∈ AB one has E(J) ∈
B D(H)∗ , D(H) . Note that this measure is usually unbounded if f ∈ D(H)∗ \ H. As
a consequence, (6.20) will hold in D(H)∗ , and by assumption (6.21) will hold for any
f ∈ G. It thus follows that mf is absolutely continuous on J for any f ∈ G. If the strong
GLAP holds, then (6.22) is a direct consequence of (6.20).
ii) By assumption one has T ∗ ∈ B(H, G). Since ℑR(z) maps G into D(H) ⊂ G ∗◦ we
get
∥T ℑR(z)T ∗ ∥ ≤ ∥T ∥G ∗ →H ∥ℑR(z)∥G→G ∗ ∥T ∗ ∥H→G
which means that T is locally H-smooth on J, by Proposition 6.4.2.
Let us still mention that quite often, the limiting absorption principle is formally
obtained by (replacing ℑR(λ ) + iµ) by R(λ + iµ). However, since R(λ + iµ) does not
belong to B D(H)∗ , D(H) , the expression ⟨f, R(z)f ⟩ is not a priori well-defined for f
in the space
( 1/2G used
) before. One natural way to overcome this difficulty is to consider the
space D |H| , which is called the form domain of H. Then the following embeddings
are continuous and dense
( ) ( )∗
D(H) ⊂ D |H|1/2 ⊂ H ⊂ D |H|1/2 ⊂ D(H)∗ .
( ( ) ( 1/2 ))
1/2 ∗
The main point in this construction is that R(z) ∈ B D |H| , D |H| for any
( 1/2 )∗
z ∈ ρ(H). So we can mimic the previous construction and consider G ⊂ D |H|
continuously
( 1/2 )∗ and densely. Note that in the ( applications
) one often considers G ⊂ H ⊂
∗
D |H| . Then, it follows that that D |H| 1/2
⊂ G continuously, and consequently
( ( )∗ ( ))
B D |H|1/2 , D |H|1/2 ⊂ B(G, G ∗◦ ) ⊂ B(G, G ∗ ).

In this context, the limiting absorption principle (LAP) or the strong LAP corresponds
to the content of (6.5.1) with ℑR(λ + iµ) replaced by R(λ + iµ). Note that the LAP is
usually a much stronger requirement than the GLAP since the the real part of R(λ+iµ)
is a much more singular object (in the limit µ ↘ 0) than its imaginary part.
Let us still reformulate the strong limiting absorption principle in different terms:
Consider C± := {z ∈ C | ±ℑ(z) > 0} and observe that C± ∋ z 7→ R(z) ∈ B(G, G ∗ ) is a
holomorphic function. The strong LAP is equivalent to the fact that this function has
a weak*-continuous extension to the set C± ∪ J. The boundary values R(λ ± i0) of the
resolvent on the real axis allow us to expression the derivative of the spectral measure
on J as
d 1 [ ]
Eλ = R(λ + i0) − R(λ − i0) . (6.23)
dλ 2πi
6.6. THE METHOD OF DIFFERENTIAL INEQUALITIES 91

6.6 The method of differential inequalities

In this section we provide the proofs of Theorems 6.1.1 and 6.1.2. In fact, we shall mainly
deal with a bounded operator S and show at the end how an unbounded operator H
with a spectral gap can be treated in this setting. For an unbounded operator H without
a spectral gap, we refer to [ABG, Sec. 7.5].
Our framework is the following: Let A be a self-adjoint (usually unbounded) op-
erator in a Hilbert space H and let S be a bounded and self-adjoint operator in H.
We assume that S belongs to C 1,1 (A), as introduced in Section 6.2. Observe that this
regularity condition means
∫
dt
(Ut − 1)2 [S] 2 < ∞
|t|≤1 t
∫
dt
⇐⇒ e2itA S e−2itA −2 eitA S e−itA +S 2 < ∞
|t|≤1 t
∫
dt
⇐⇒ eitA S e−itA + e−itA S eitA −2S 2 < ∞
|t|≤1 t
∫ 1
dt
⇐⇒ eitA S e−itA + e−itA S eitA −2S 2 < ∞.
0 t

Our first( aim is to) provide the proof the following theorem. In its statement, the
space G := D(A), H 1/2,1 appears, and we will explain its definition when necessary.
The important information is that G ⊂ H continuously and densely. We also recall that
µA (S) = {λ ∈ R | ϱA S (λ) > 0}.

Theorem 6.6.1. Let S be a bounded and self-adjoint operator which belongs to C 1,1 (A).
Then the holomorphic function C± ∋ z 7→ (S − z)−1 ∈ B(G, G ∗ ) extends to a weak*-
continuous function on C± ∪ µA (S).

As explained in the previous sections, such a statement means that a strong limiting
absorption principle holds for S in G and locally on µA (S). Some consequences of this
statement is that S has purely absolutely continuous spectrum on µA (S), that the
derivative of its spectral measure can be expressed by the imaginary part of its resolvent
on the real axis, as mentioned in (6.23), and that some locally S-smooth operators on
µA (S) are automatically available.
The proof of Theorem 6.6.1 is divided into several lemmas. Note that since S ∈
C 1,1 (A) ⊂ C 1 (A), then the commutator B := [iS, A] is well-defined and belongs to
B(H). Before starting with these lemmas, let us provide a heuristic explanation about
the approach. Since the aim is to consider ⟨f, (S − λ ∓ iµ)−1 f ⟩ and its limits when
µ ↘ 0, we shall consider a regularized version of such an expression, with an additional
parameter ε, namely ⟨ ( )−1 ⟩
fε , Sε − λ ∓ i(εBε + µ) fε . (6.24)
Then, it is mainly a matter of playing with the ε-dependence of all these terms...
92 CHAPTER 6. COMMUTATOR METHODS

The following lemma is really technical, but it is of crucial importance since it is

there that the assumption ϱAS (λ0 ) > 0 will play an essential role. For completeness we
provide its proof. Note that the reason for getting an estimate of the form (6.25) will
become clear only in the subsequent lemma.

Lemma 6.6.2. Let {Sε }0<ε<1 and {Bε }0<ε<1 be two families of bounded self-adjoint
operators satisfying ∥Sε − S∥ + ε∥Bε ∥ ≤ cε for some constant c as well as the condition
limε→0 ∥Bε − B∥ = 0. Let λ0 ∈ R and a ∈ R such that ϱA S (λ0 ) > a > 0. Then there exist
some strictly positive numbers δ, ε0 and b such that for |λ − λ0 | ≤ δ, for 0 < ε ≤ ε0
and for any µ ≥ 0 the following estimate holds for all g ∈ H:

b [ ] 2
a∥g∥2 ≤ ⟨g, Bε g⟩ + S ε − λ ∓ i(εBε + µ) g . (6.25)
µ2 + δ 2

Proof. Let us
( choose numbers) a < a0 < a1 < ϱAS (λ0 ) and δ > 0 such that a1 E ≤ EBE
for E = E (λ0 − 2δ, λ0 + 2δ) . Let us also choose ε1 > 0 such that ∥Bε − B∥ ≤ a1 − a0
S

for any 0 < ε ≤ ε1 . This implies that EBε E ≥ EBE − (a1 − a0 )E ≥ a0 E. Let us set
E ⊥ = 1 − E and consider from now on λ and µ real with |λ − λ0 | ≤ δ and µ ≥ 0.
Observe then that ∥(S − λ ∓ iµ)−1 E ⊥ ∥ ≤ (µ2 + δ 2 )−1/2 , and hence
[ ] 2
∥E ⊥ g∥2 = (S − λ ∓ iµ)−1 E ⊥ Sε − λ ∓ i(εBε + µ) + S − Sε ± iεBε g
2 [ ] 2 2c2 ε2
≤ 2 2
S ε − λ ∓ i(εBε + µ) g + 2 2
∥g∥2 .
µ +δ µ +δ

We then get for ε ≤ ε1 and for any ν > 0:

a0 ∥g∥2 = a0 ⟨g, Eg⟩ + a0 ∥E ⊥ g∥2

≤ ⟨g, EBε Eg⟩ + a0 ∥E ⊥ g∥2
= ⟨g, Bε g⟩ − 2ℜ⟨Eg, Bε E ⊥ g⟩ − ⟨E ⊥ g, Bε E ⊥ g⟩ + a0 ∥E ⊥ g∥2
≤ ⟨g, Bε g⟩ + ν∥g∥2 + ν −1 ∥Bε ∥2 ∥E ⊥ g∥2 + ∥Bε ∥ ∥E ⊥ g∥2 + a0 ∥E ⊥ g∥2
[ ] 2c2 ε2
≤ ⟨g, Bε g⟩ + ν∥g∥2 + ν −1 ∥Bε ∥2 + ∥Bε ∥ + a0 2 ∥g∥2
µ + δ2
[ ] 2 [ ] 2
+ ν −1 ∥Bε ∥2 + ∥Bε ∥ + a0 2 S ε − λ ∓ i(εB ε + µ) g .
µ + δ2

Since we have ∥Bε ∥ ≤ c and since we may assume c ≥ 1 and ν ≤ 1, it follows that
[ 2c2 ε2 (a0 + 2c2 ) ] 2a0 + 4c2 [ ] 2
a0 − ν − ∥g∥ 2
≤ ⟨g, B ε g⟩ + S ε − λ ∓ i(εBε + µ) g .
νδ 2 ν(µ2 + δ 2 )

We can finally chose ν > 0 and ε0 ∈ (0, ε1 ) such that the term in the square brackets
in the l.h.s. is bigger of equal to a for any 0 < ε ≤ ε0 . We then get (6.25) with
b = ν −1 (2a0 + 4c2 ).
6.6. THE METHOD OF DIFFERENTIAL INEQUALITIES 93

Lemma 6.6.3. Under the same assumptions as in the previous lemma, the operators
Sε − λ ∓ i(εBε + µ) are invertible in B(H) whenever |λ − λ0 | ≤ δ, 0 < ε ≤ ε0 and
µ ≥ 0. For any fixed λ and µ satisfying these conditions we set
[ ]−1
G±ε := Sε − λ ∓ i(εBε + µ)

Then one has (G± ∗ ∓

ε ) = Gε and

1 [ cε + ∥S∥ + |λ + iµ| ]
∥G±
ε∥ ≤ 1 + bε . (6.26)
aε + µ µ2 + δ 2
Moreover, for any h ∈ H
1 1 ( b )1/2
∥G±
ε h∥ ≤ √ |ℑ⟨h, G+
ε h⟩|1/2
+ ∥h∥. (6.27)
aε δ a
Proof. Let us set
Tε± := Sε − λ ∓ i(εBε + µ)
and deduce from (6.25)

bε
(aε + µ)∥g∥2 ≤ ⟨g, (εBε + µ)g⟩ + ∥T ± g∥2
µ2 + δ2 ε
bε
= ∓ℑ⟨g, Tε± g⟩ + 2 ∥T ± g∥2
µ + δ2 ε
{ bε }
± ±
≤ ∥g∥ ∥Tε g∥ 1 + 2 ∥T ∥
µ + δ2 ε
{ ∥S∥ + cε + |λ + iµ| }
±
≤ ∥g∥ ∥Tε g∥ 1 + bε .
µ2 + δ 2

It follows from this equality and from the boundedness (and thus closeness) of Tε± that
these operators are injective and with
( ±closed
) range, see also [Amr, Lem. 3.1]. In addition,
± ∗ ∓
since (Tε ) = Tε and since Ker (Tε ) = Ran(Tε ) , one infers that Ran(Tε± ) = H.
∗ ± ⊥

One then easily deduces all assertions of the lemma, except the estimate (6.27).
To prove (6.27), let us set g = G±
ε h in (6.25) and observe that

bε
aε∥G± ± ±
ε h∥ ≤ ⟨Gε h, (εBε + µ)Gε h⟩ +
2
∥h∥2 . (6.28)
µ2 + δ2
By taking the following identities into account,

⟨G± ± −1 ∓ ∓ ± ±
ε h, (εBε + µ)Gε h⟩ = ±(2i) ⟨h, Gε (Tε − Tε )Gε h⟩
= ±(2i)−1 ⟨h, (G± ∓
ε − Gε )h⟩
= ℑ⟨h, G+
ε h⟩,

one directly obtains (6.27) from (6.28).

94 CHAPTER 6. COMMUTATOR METHODS

We keep the assumptions and the notations of the previous two lemmas, and set
simply Gε and Tε for G+ +
ε and Tε . The derivative with respect to the variable ε will be
denoted by ′ , i.e. G′ε = dε
d
Gε .

Lemma 6.6.4. Assume that the maps ε 7→ Sε and ε 7→ Bε are C 1 in norm, and that
for any fixed ε the operators Bε and Sε belong to C 1 (A). Then the map ε 7→ Gε is C 1
in norm, and also Gε ∈ C 1 (A) for any fixed ε.

Proof. The differentiability of Gε with respect to ε follows easily from the equality
( )
Gε′ − Gε = Gε′ (Tε − Tε′ )Gε = Gε′ Sε − Sε′ − i(εBε − ε′ Bε′ ) Gε .

We thus get that G′ε = −Gε Tε′ Gε . For the regularity of Gε with respect to A, this
follows from the statement (iii) of Proposition 6.2.4. In addition, one can also infer that
[A, Gε ] = −Gε [A, Tε ]Gε , see [ABG, Prop. 6.1.6] for the details. Note that this equality
can also be obtained with the operator Aτ , as already used in (6.11).

Note that from the previous statement and from its proof, one deduces the following
equality
{ ( )}
G′ε + [A, Gε ] = iGε Bε − [iSε , A] + ε iε−1 Sε′ + Bε′ + [A, Bε ] Gε . (6.29)

In the next lemma we provide a precise formulation of what had been mentioned
in (6.24).

Lemma 6.6.5. Let us keep the assumptions of the previous three lemmas, and let
{fε }0<ε<1 be a bounded family of elements of H such that ε 7→ fε is strongly C 1 and
such that fε ∈ D(A) for any fixed ε. Set

Fε := ⟨fε , Gε fε ⟩.

Then the map ε 7→ Fε is of class C 1 and its derivatives satisfies

Fε′ = ⟨fε , (G′ε + [A, Gε ])fε ⟩ + ⟨G∗ε fε , fε′ + Afε ⟩ + ⟨fε′ − Afε , Gε fε ⟩. (6.30)

In addition, if we set

ℓ(ε) := ∥fε′ ∥ + ∥Afε ∥, q(ε) := ∥ε−1 (Bε − [iSε , A]) + iε−1 Sε′ + Bε′ + [A, Bε ]∥, (6.31)

then Fε satisfies the differential inequality

1 ′ [ ] ℓ(ε) q(ε)
|Fε | ≤ ω∥fε ∥ ℓ(ε) + ωεq(ε)∥fε ∥ + √ |Fε |1/2 + |Fε |, (6.32)
2 aε a

with ω = a−1/2 b1/2 δ −1 and 0 < ε ≤ ε0 .

6.6. THE METHOD OF DIFFERENTIAL INEQUALITIES 95

Proof. The equality (6.30) is obvious, since the commutator can be opened on D(A).
By using (6.29) it can then be rewritten as
⟨ { } ⟩
Fε′ =i G∗ε fε , Bε − [iSε , A] + ε(iε−1 Sε′ + Bε′ + [A, Bε ]) Gε fε
+ ⟨G∗ε fε , fε′ + Afε ⟩ + ⟨fε′ − Afε , Gε fε ⟩. (6.33)

As a consequence one infers that

( )
|Fε′ | ≤ εq(ε)∥Gε fε ∥ ∥G∗ε fε ∥ + ℓ(ε) ∥Gε fε ∥ + ∥G∗ε fε ∥ . (6.34)

In addition it follows from (6.27) that

1
∥G±
ε fε ∥ ≤ √ |Fε |1/2 + ω∥fε ∥.
aε

By inserting these inequalities in (6.34) and by using the inequality (p + q)2 ≤ 2p2 + 2q 2
for any p, q ≥ 0 one directly obtains (6.32).

The differential inequality (6.32), from which the method takes its name, is quite
remarkable in that the spectral variable λ + iµ does not appear explicitly in the coeffi-
cients. In fact, the only conditions on these parameters are |λ − λ0 | ≤ δ and µ ≥ 0.
Let us still rewrite (6.32) in the simple form

|Fε′ | ≤ η(ε) + φ(ε)|Fε |1/2 + ψ(ε)|Fε |.

∫ε
By using the trivial inequality |Fs | ≤ |Fε0 |+ s 0 |Fτ′ |dτ , we then obtain for 0 < ε < s < ε0
∫ ∫
ε0 ε0 [ ]
|Fs | ≤ |Fε0 | + η(τ )dτ + φ(τ )|Fτ |1/2 + ψ(τ )|Fτ | dτ. (6.35)
s s

We shall now apply an extended version of the Gronwall lemma to this differential
inequality. More precisely, let us first state such a result, and refer to [ABG, Lem. 7.A.1]
for its proof.

Lemma 6.6.6 (Gronwall lemma). Let (a, b) ⊂ R( and let ) f, φ, ψ be non-negative real
functions on (a, b) with f bounded, and φ, ψ ∈ L1 (a, b) . Assume that, for some con-
stants ω ≥ 0 and θ ∈ [0, 1) and for all λ ∈ (a, b) one has
∫ b [ ]
f (λ) ≤ ω + φ(τ )f (τ )θ + ψ(τ )f (τ ) dτ.
λ

Then one has for each λ ∈ (a, b)

[ ∫ b ∫b ]1/(1−θ) ∫ b
f (λ) ≤ ω 1−θ
+ (1 − θ) φ(τ ) e(θ−1) τ ψ(s)ds dτ e λ ψ(τ )dτ .
λ
96 CHAPTER 6. COMMUTATOR METHODS

Thus, by applying this result for θ = 1/2 to (6.35) one gets that
[( ∫ ε0 )1/2 1 ∫ ε0 ∫ ]2 ∫ ε
1 ε0
|Fε | ≤ |Fε0 | + φ(τ ) e− 2 τ ψ(s)ds dτ e ε ψ(τ )dτ
0
η(τ )dτ +
ε 2 ε

for all 0 < ε < ε0 . We can then deduce the simpler inequality
[ ∫ ε0 ( ∫ ε0 )2 ] ∫ ε
|Fε | ≤ 2 |Fε0 | +
0
η(τ )dτ + φ(τ )dτ e ε ψ(τ )dτ .
ε ε

Our final purpose is to get a bound on |Fε | < const. < ∞ independent of z = λ + iµ
as ε → 0. From the above estimate we see that this is satisfied if
∫ ε0
[η(τ ) + φ(τ ) + ψ(τ )]dτ < ∞.
0

By coming back to the explicit formula for these functions, it corresponds to the con-
dition ∫ ε0 [ ]
ℓ(ε)
ℓ(ε)∥fε ∥ + εq(ε)∥fε ∥2 + √ + q(ε) dε < ∞.
0 ε
In fact, it is easily observed (see also page 304 of [ABG]) that this condition is satisfied
if the following assumption holds:
∫ 1
[ε−1/2 ℓ(ε) + q(ε)]dε < ∞.
0

By looking back at the definitions of ℓ and q in (6.31) we observe that the condition
on ℓ corresponds to a condition on the family of elements {fε } while the condition on
q corresponds to conditions on the families {Sε } and {Bε }. For the condition on q let
us just mention that a suitable choice for Sε is given by
∫ ∞
dτ
e−iετ A S eiετ A e−τ /4
2
Sε := .
−∞ (4π)1/2

Then by setting Bε := [iSε , A] and by assuming that S ∈ C 1,1 (A) it is shown in [ABG,
Lem. 7.3.6] that all assumptions on the families {Sε } and {Bε } are satisfied. Note that
the proof of this statement is rather technical and that we shall not comment on it.
For the condition involving ℓ, let us consider f ∈ H and set for any ε > 0

fε := (1 + iεA)−1 f.

Then one has fε ∈ D(A), fε′ = −i(1 + iεA) −1

Afε , ∥fε ∥ ≤ ∥f ∥, ∥fε − f ∥ → 0 as ε → 0,
∫ 1 −1/2
and ℓ(ε) ≤ 2∥Afε ∥. Then the condition 0 ε ℓ(ε)dε < ∞ holds if
∫ 1
dε
ε1/2 A(1 + iεA)−1 f < ∞. (6.36)
0 ε
6.6. THE METHOD OF DIFFERENTIAL INEQUALITIES 97

Such a condition corresponds to a regularity condition of f with respect to A. In fact,

many Banach spaces of elements of H having a certain regularity with respect A can
be defined, and Chapter 2 of [ABG] is entirely devoted to that question. Here, the
(elements )of H satisfying condition (6.36) are precisely those belonging to the space
D(A), H 1/2,1 , as shown in [ABG, Prop. 2.7.2]. Note that this space is an interpolation
space between D(A) and H and contains ( the space
) D(⟨A⟩
1/2+ϵ
)∫ for any ϵ > 0.
1
By summing up, for any f ∈ G := D(A), H 1/2,1 one has 0 ε−1/2 ℓ(ε)dε ≤ c1 ∥f ∥G
for some c1 < ∞ independent of f ∈ G. By looking at the explicit ∫ 1 form of the functions
η and φ one also obtains that there exists c2 < ∞ such that 0 η(τ )dτ ≤ c2 ∥f ∥2G and
∫1
0
φ(τ )dτ ≤ c2 ∥f ∥G . One then infers that |Fε | ≤ c∥f ∥2G for any 0 < ε ≤ ε0 , |λ − λ0 | ≤ δ
and µ ≥ 0 with a constant c independent of f ∈ G, ε, λ and µ. The proof of Theorem
6.6.1 can now be provided:
Proof of Theorem 6.6.1. By all the previous arguments, there exists an integrable func-
tion κ : (0, ε0 ) → R such that |Fε′ | ≤ κ(ε) for all ε, λ, µ as above. Now, fix µ > 0. Since
Sε − λ − i(εBε + µ) converges to S − λ − iµ ≡ S − z in norm as ε → 0 we shall have
Gε → (S − z)−1 in norm too, and
∫ ε0
−1
⟨f, (S − z) f ⟩ = lim⟨fε , Gε fε ⟩ = ⟨fε0 , Gε0 (z)fε0 ⟩ − Fε′ (z)dε. (6.37)
ε→0 0

Note that we have explicitly indicated the dependence on z = λ + iµ of Gε0 and Fε′ .
Let us set Ω := {λ + iµ | |λ − λ0 | < δ, µ ≥ 0}. It follows from (6.26) that ∥Gε0 (z)∥ ≤
const. < ∞ independently of z ∈ Ω. For each ε > 0 the continuity of z ∈ Ω of
Fε′ (z) follows from (6.33). By the dominated convergence theorem, with the fact that
|Fε′ | ≤ κ(ε), the equation (6.37) gives the existence of a continuous extension of the
function ⟨f, (S − z)−1 f ⟩ from the domain {z ∈ Ω | ℑ(z) = µ > 0} to all Ω. The
polarization principle shows that this holds for ⟨f, (S − z)−1 g⟩ for any f, g ∈ G.
Let us finally show how the two theorems stated at the beginning of the chapter
follow from the various results obtained subsequently. First of all we provide a proof of
Theorem 6.1.2.
Proof of Theorem 6.1.2. Let λ0 ∈ R \ σ(H) and set S := (H − λ0 )−1 . Then S is a
bounded self-adjoint operator, and the resolvents of S and H are related by the identity

(H − z)−1 = (λ0 − z)−1 [S − (λ0 − z)−1 ]−1 S, ℑ(z) ̸= 0. (6.38)

Let J ⊂ µA (H) be a compact set with λ0 ̸∈ J and set J˜ := {(λ0 − λ)−1 | λ ∈ J}. Note
that there is no restriction on the generality if we assume that λ does not belong to a
neighbourhood of λ0 , since (H − z)−1 is holomorphic in such a neighbourhood. Then
S ∈ C 1,1 (A) and J˜ is a compact subset of µA (S), see Proposition 7.2.5 of [ABG]. In
addition, Theorem 6.6.1 says that the map ζ 7→ (S − ζ)−1 ∈ B(G, G ∗ ) extends to a
weak*-continuous function on C± ∪ J. ˜ Since z 7→ ζ = (λ0 − z)−1 is a homeomorphism
of C± ∪ J onto C± sup J, ˜ we see that z 7→ [S − (λ0 − z)−1 ]−1 ∈ B(G, G ∗ ) extends to a
98 CHAPTER 6. COMMUTATOR METHODS

weak*-continuous function on C± ∪ J. The result of the theorem now follows from the
identity (6.38) and the fact that SG ⊂ G, as a consequence of [ABG, Thm. 5.3.3].
The second part of the statement is a direct consequence of what has been presented
in Section 6.4. Note in particular that in the Definition 6.4.1 of a locally H-smooth
operator, one could have considered T : D(H) → K with K an arbitrary Hilbert space.
It is in this generality that the statement of Theorem 6.1.2 is provided, and this slight
extension can easily be taken into account.
In the same vein one has:
Proof of Theorem 6.1.1. The first assertion about the finiteness of the set of eigenval-
ues is a direct consequence of Corollary 6.3.7. For the second statement, observe that
Theorem 6.3.8 implies the inclusion J \ σp (H) ⊂ µA (H), and then use the consequences
of the limiting absorption principle, as presented in Section 6.5.
Bibliography

[Amr] W.O. Amrein, Hilbert space methods in quantum mechanics, Fundamental Sci-
ences. EPFL Press, Lausanne; distributed by CRC Press, Boca Raton, FL, 2009.

[ABG] W.O. Amrein, A. Boutet de Monvel, V. Georgescu, C0 -groups, commutator

methods and spectral theory of N-body Hamiltonians, Birkhäuser, Basel, 1996.

[BM] H. Baumgärtel, M. Wollenberg, Mathematical scattering theory, Birkhäuser ver-

lag, Basel, 1983.

[BDG] L. Bruneau, J. Dereziński, V. Georgescu, Homogeneous Schrödinger operators

on half-line, Ann. Henri Poincaré 12 no. 3 (2011), 547–590.

[DR] J. Dereziński, S. Richard, On almost homogeneous Schrödinger operators, Preprint

arXiv:1604.03340.

[Kat] T. Kato, Perturbation theory for linear operators, Classics in mathematics,

Springer, 1995.

[Mur] G.J. Murphy, C ∗ -algebras and operator theory, Academic Press, Inc., Boston,
MA, 1990.

[Ped] G. Pedersen, Analysis now, Graduate texts in mathematics 118, Springer, 1989.

[RS3] M. Reed, B. Simon, Methods of modern mathematical physics III: scattering the-
ory, Academic Press, Inc., 1979.

[RS4] M. Reed, B. Simon, Methods of modern mathematical physics IV: analysis of

operators, Academic Press, Inc., 1978.

[Ric] S. Richard, Levinson’s theorem: an index theorem in scattering theory, in Proceed-

ings of the Conference Spectral Theory and Mathematical Physics, Santiago 2014,
Operator Theory Advances and Applications 254, 149 203, Birkhuser, 2016.

[Tes] G. Teschl, Mathematical methods in quantum mechanics, with applications to

Schrödinger operators, Graduate Studies in Mathematics 99, American Mathe-
matical Society, Providence, RI, 2009.

99
100 BIBLIOGRAPHY

[Yaf] D.R. Yafaev, Mathematical scattering theory. General theory, Translations of

Mathematical Monographs 105, American Mathematical Society, Providence, RI,
1992.