Mathematical Quantum Theory

Marcello Porta

Eberhard Karls Universität Tübingen, Department of Mathematics, Auf der

Morgenstelle 10, 72076 Tübingen, Germany

July 23, 2019

Abstract
These are the lecture notes for the courses Mathematical Quantum Theory and Advanced
Topics in Quantum Mechanics, both given at the University of Tübingen in the academic
year 2018/2019.

Contents
1 Introduction 3
1.1 The Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Function spaces 4
2.1 C k spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 The free Schödinger equation 7

3.1 The Fourier transform on L1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Solution of the free Schrödinger equation . . . . . . . . . . . . . . . . . . . . . 11
3.2.1 Comparison between Schrödinger, heat and wave equations . . . . . . 14
3.3 Tempered distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Long time asymptotics of the momentum operator . . . . . . . . . . . . . . . 18
3.5 Properties of Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.6 The Fourier transform in L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.7 Unitary groups and their generators . . . . . . . . . . . . . . . . . . . . . . . 26

4 Selfadjoint operators 29
4.1 The Hilbert space adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Criteria for symmetry, selfadjointness and essential selfadjointness . . . . . . 35
4.3 Selfadjoint extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 From quadratic forms to operators . . . . . . . . . . . . . . . . . . . . . . . . 41

5 The spectral theorem 42

5.1 The spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 Postulates of quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2.1 Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2.2 Time evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.3 Projection valued measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.4 Functional calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.5 Construction of projection valued measures . . . . . . . . . . . . . . . . . . . 54
5.6 Unitary equivalence of self-adjoint operators with multiplication operators . . 61
5.7 Decomposition of the spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 63

1
6 Quantum dynamics 65
6.1 Existence and uniqueness of the solution of the Schrödinger equation . . . . . 65
6.2 Stone’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.3 The RAGE theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

7 General Schrödinger operators 72

7.1 Kato-Rellich theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.2 Relatively compact perturbations and Weyl’s theorem . . . . . . . . . . . . . 76
7.3 Two examples of Schrödinger operators . . . . . . . . . . . . . . . . . . . . . 78
7.3.1 The harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7.3.2 Finite well potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.4 General Schrödinger operators: existence of stationary states . . . . . . . . . 83
7.4.1 Energy functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.4.2 Weak continuity of the potential energy . . . . . . . . . . . . . . . . . 84
7.4.3 Existence of minimizers . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.4.4 Excited states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.4.5 Min-max principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.4.6 Generalized min-max principle . . . . . . . . . . . . . . . . . . . . . . 90

8 Semiclassical approximations 91
8.1 Dirichlet Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8.2 Lower bound on the sum of Dirichlet eigenvalues . . . . . . . . . . . . . . . . 91
8.3 Asymptotic behavior of eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . 94
8.4 Upper bound on the sum of Dirichlet eigenvalues . . . . . . . . . . . . . . . . 95
8.4.1 Coherent states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.4.2 Proof of Theorem 8.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.5 General Schrödinger operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

9 Many-body quantum mechanics 98

9.1 Bosons and fermions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
9.2 Reduced density matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9.3 Atoms and molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.4 Thomas-Fermi theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
9.4.1 The free Fermi gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
9.4.2 The Thomas-Fermi energy functional . . . . . . . . . . . . . . . . . . . 104
9.4.3 Existence and uniqueness of the minimizer in DN . . . . . . . . . . . . 106
9.4.4 Ionization in TF theory . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.4.5 Scaling properties of the TF energy . . . . . . . . . . . . . . . . . . . 117
9.5 Stability of matter of the second kind via TF theory . . . . . . . . . . . . . . 117
9.5.1 The no-binding theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 117
9.5.2 Proof of stability of matter . . . . . . . . . . . . . . . . . . . . . . . . 120
9.6 TF theory as the N Ñ 8 limit of quantum mechanics . . . . . . . . . . . . . 122
9.6.1 Upper bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9.6.2 Lower bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

A Properties of Sobolev spaces 127

A.1 Sobolev inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A.2 Weak to strong convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

B Bathtub principle 131

References 133

2
1 Introduction
1.1 The Schrödinger equation
Let us consider the evolution of one particle in Rd , with d “ 1, 2, 3 the physically relevant
choices of the dimension d. We will assume the particle to be pointlike. We suppose that
the particle is exposed to the action of an external potential V : Rd Ñ R.
In quantum mechanics, the state of the system is described by the wave function ψpt, xq,
ψ : R ˆ Rd ÞÑ C, square integrable:
ż
}ψpt, ¨q}22 :“ |ψpt, xq|2 dx “ 1 . (1.1)
Rd

The physical interpretation of |ψpt, xq|2 is that of probability distribution for finding the
particle at px, tq. That is, the probability for finding the particle at the time t in the region
A Ă Rd is: ż
Pψt pAq “ |ψpt, xq|2 dx . (1.2)
A
The evolution of the particle is defined by the time-dependent Schrödinger equation:
B ~2
i~ ψpt, xq “ ´ ∆x ψpt, xq ` V pxqψpt, xq “: Hψpt, xq , (1.3)
Bt 2m
where ~ is called the (reduced) Planck constant, and it has the dimensions of an action,
r~s “ renergys ˆ rtimes. The Laplace operator is defined as:
d
ÿ B2
∆x “ . (1.4)
j“1
Bx2j

The differential operator H is called the Hamiltonian of the system. The Schrödinger equa-
tion is an example of partial differential equation, and the discussion of existence and unique-
ness of solutions will be part of the present course.
Given a Hamiltonian H, the corresponding time-independent Schrödinger equation is:

Hψ “ Eψ , (1.5)

where the (real) number E has the interpretation of energy of the system. A square inte-
grable solution of the time-independent Schrödinger equation is called an eigenstate of the
Hamiltonian H. Notice that if ψ is an eigenstate of H, then ψptq “ e´iEt{~ ψ is a solution of
the time-dependent Schrödinger equation.

Comparison with classical mechanics. Recall the motion of particle in classical

mechanics. The trajectory qptq P Rd of a classical particle is determined by Newton’s equa-
tion:
m:q ptq “ F pqptqq “ ´∇V pqptqq , pqp0q, qp0qq
9 “ pq0 , q90 q . (1.6)
This second order ordinary differential equation can be rewritten as a first order differential
equation for the pair ppptq, qptqq, with pptq “ mqptq
9 the momentum of the particle. The
Hamilton’s equation of motion for the particle is:
ˆ ˙ ˆ ˙ ˆ ˙
d pptq ´∇V pqptqq ´∇q Hpq, pq
“ 1 ” , (1.7)
dt qptq m pptq ∇p Hpp, qq
2
with Hpp, qq “ |p|m ` V pqq the Hamiltonian of the particle. The Hamiltonian appearing
in the Schrödinger equation is called the canonical quantization of the classical Hamilto-
nian, obtained by replacing the position variable q by a multiplication operator x, and the
momentum variable p by the differential operator ´i~∇x .
Quantum mechanics is a more fundamental theory of nature than classical mechanics. A
natural question is to understand how classical mechanics emerges from quantum mechanics.
This question will be discussed later in the course, while introducing semiclassical analysis.

3
 The main goal of this course is to develop the mathematical theory of the Schrödinger
equation, for one particle and for many particle systems. Notice that the Schrödinger equa-
tion is a linear evolution equation, in contrast to Hamilton’s equation of motion; this seems
to suggest that its mathematical study should be “easy”. This is not true, due to the fact
that the solution of the equation lives in an infinite dimensional space, and that the operator
H is unbounded.

2 Function spaces
In this section we shall introduce function spaces that will play an important role in the
mathematical formulation of quantum mechanics. We shall only review some basic results,
and we will refer the reader to [3, 5] for more details.

2.1 C k spaces
Definition 2.1. A multiindex α P Nd0 is a d-tuple α “ pα1 , . . . , αd q, with αj P N0 , and
řd
|α| “ j“1 αj . For x P Rd we define:

αd B |α|
xα “ xα1 α2
1 x2 ¨ ¨ ¨ xd and Bxα :“ . (2.1)
Bxα
1
1
¨ ¨ ¨ Bxα
d
d

Definition 2.2. Let A Ď Rd , k P N0 . We define:

! )
C k pAq “ f | f : A Ñ C, Bxα f is continuous for all α such that |α| ď k . (2.2)

Also, we denote by Cbk pAq the restriction of C k pAq to functions with bounded derivatives:
! )
Cbk pAq “ f | f P C k pAq and there exists cα ą 0 such that @|α| ď k sup |Bxα f pxq| ď cα .
xPA
(2.3)
Remark 2.3. It turns out that the space Cbk pAq is a Banach space, if endowed with the
following norm:
k
ÿ ÿ
}f }Cbk pAq “ sup |Bxα f pxq| . (2.4)
n“0 α:|α|“n xPA

We also define the space of C k functions with compact support.

Definition 2.4. Let:
supppf q “ tx P Dompf q | f pxq ‰ 0u (2.5)
d
be the support of the function f . Let A Ď R , k P N0 . We define:
! )
Cck pAq “ f | f P C k pAq s.t. supppf q X A is compact. (2.6)

Remark 2.5. Cck pAq Ď Cbk pAq Ď C k pAq.

Example 2.6. (i) Let A “ R, and f pxq “ x. We have f P C 8 pRq. However, f R Cb8 pRq,
since f is unbounded. Also, f R Cc8 pRq, since supppf q “ R.
(ii) Consider the “bump function”:
" ` ˘
exp ´ 1{p1 ´ x2 q x P p´1, 1q
f pxq “ (2.7)
0 otherwise.

It is easy to see that all derivatives of f are continuous in x P R, and are compactly
supported in p´1, 1q. Thus, f P Cc8 pRq.

4
2.2 Lp spaces
Definition 2.7. Let A Ď Rd , measurable. Let p P R, 1 ď p ă 8. We define:
! ż )
p
L pAq :“ f | f : A Ñ C, f measurable, dx |f pxq|p ă 8 . (2.8)
A
ş
Remark 2.8. The integral A dx ¨ ¨ ¨ has to be understood as a Lebesgue integral. If the
function f is Riemann integrable, then it coincides with the standard Riemann integral.
More generally, one could replace dx by a Lebesgue measure µpdxq. In that case, we shall
denote the corresponding Lp space by Lp pA, dµq. One can check that Lp is a vector space.
Besides being vector spaces, Lp spaces are also Banach spaces, if endowed with the
following norm.
Definition 2.9. Let f P Lp pAq, 1 ď p ă 8. We define:
´ż ¯1{p
}f }Lp pAq :“ dx |f pxq|p . (2.9)
A

One can check that the map } ¨ }Lp pAq has the following properties.
(i) }λf }Lp pAq “ |λ|}f }Lp pAq , λ P C.
(ii) }f }Lp pAq “ 0 ô f pxq “ 0 a.e.
(iii) }f ` g}Lp pAq ď }f }Lp pAq ` }g}Lp pAq (Minkowki inequality).
These properties imply that } ¨ }Lp pAq is a semi-norm. The reason why it is not a norm is
that it is easy to imagine functions such that }f }Lp pAq “ 0 and f pxq ‰ 0 (take f to be zero
everywhere except at a point). To ensure that }¨}Lp pAq defines a norm, one has to redefine Lp
by identifying functions that differ on a zero measure set (e.g., on a countable set of points).
Given f P Lp , we define an equivalent class of functions as
f˜ “ f 1 P Lp | f ´ f 1 “ 0 a.e.
(
(2.10)

We redefine Lp as the set of the equivalence classes of functions f˜.

The L8 space is defined as follows.
Definition 2.10.
L8 pAq :“ tf | f : A Ñ C , f measurable , DK ą 0 s.t. |f pxq| ď K a.e. u . (2.11)
A norm on L8 is defined by taking the essential supremum of f :
}f }L8 pAq :“ inf tK | |f pxq| ď K a.e. in Au . (2.12)
Here we shall list some important facts about Lp spaces, without proof. We refer the
reader to [3] for details. Whenever it does not generate ambiguity, we might replace } ¨ }Lp pAq
by } ¨ }p .
Theorem 2.11 (Completeness). Let 1 ď p ď 8, and let f i , i “ 1, 2, 3, ¨ ¨ ¨ be a Cauchy
sequence in Lp pAq:
lim }f i ´ f j }p “ 0. (2.13)
i,jÑ8

Then, there exists f˚ P Lp pAq such that

lim }fi ´ f˚ }p “ 0 . (2.14)
iÑ8

Remark 2.12. We use the notation fi Ñ f˚ and we say that f i converges strongly to f˚ in
Lp .
Another important property of Lp spaces, for p ă 8, is that their elements can be
approximated arbitrarily well by smooth, compactly supported functions. In other words,
Cc8 pAq is dense in Lp pAq.
Theorem 2.13 (Approximation by Cc8 functions.). Let f P Lp pRd q, 1 ď p ă 8. Then,
there exists a sequence of functions tf i uiPN , f i P Cc8 pRd q such that f i Ñ f in Lp .

5
2.3 Hilbert spaces
Let H be a vector space over C. A map x¨, ¨y : H ˆ H Ñ C is called a scalar product (or a
inner product) over H if:
(i) it is linear in its second variable, that is:
xψ, αϕ1 ` βϕ2 y “ αxψ, ϕ1 y ` βxψ, ϕ2 y (2.15)

(ii) it is antisymmetric, that is:

xψ, ϕy “ xϕ, ψy (2.16)
(iii) it is positive definite, that is:
xψ, ψy ě 0 (2.17)
for all ψ P H, with xψ, ψy “ 0 if and only if ψ “ 0.
Every scalar product induces a norm on H, defined through:
a
}ψ} “ xψ, ψy . (2.18)
The triangle inequality for } ¨ } follows from the Cauchy-Schwartz inequality
|xψ, ϕy| ď }ψ}}ϕ} . (2.19)
In fact:
a
}ψ ` ϕ} “ xψ ` ϕ, ψ ` ϕy
a
“ }ψ}2 ` }ϕ}2 ` 2Rexψ, ϕy
a
ď }ψ}2 ` }ϕ}2 ` 2}ψ}}ϕ}
“ }ψ} ` }ϕ} . (2.20)
If the vector space H equipped with the scalar product x¨, ¨y is complete, the pair pH, x¨, ¨yq
is called a Hilbert space.
Example 2.14. (a) The space Cn equipped with the scalar product:
n
ÿ
xx, yyCn “ x̄j yj (2.21)
j“1

is a Hilbert space.
(b) The space `2 of the square summable sequences pxj qjPN , equipped with the scalar product:
8
ÿ
xx, yy`2 “ x̄j yj (2.22)
j“1

is a Hilbert space.
Example 2.15 (L2 space.). In quantum mechanics, a special role is played by the space of
square integrable functions, L2 pAq. This space turns out to be a Hilbert space, if equipped
with the following scalar product:
ż
xf, gy “ dx f pxqgpxq . (2.23)

It is easy to see that the scalar product xf, gy is well defined, for all f, g P L2 pAq:
ż
|xf, gy| ď dx |f pxq||gpxq|
ż ż
1 2 1
ď dx |f pxq| ` dx |gpxq|2
2 2
1 1
” }f }L2 pAq ` }g}L2 pAq ă 8 . (2.24)
2 2
Also, it is easy to see that Eq. (2.23) fulfills the properties (i)–(iii) spelled above.

6
3 The free Schödinger equation
To start our mathematical study of the Schrödinger equation we shall consider the simplest
possible situation, corresponding to a free particle in Rd . We look for a solution ψ : RˆRd Ñ
C of the equation:
1
iBt ψpt, xq “ ´ ∆x ψpt, xq , (3.1)
2
where we set ~ “ 1 and m “ 1. A special solution can be found by separation of variables.
Consider first the time-independent Schrödinger equation:
1
´ ∆x φpxq “ λφpxq . (3.2)
2
Then, a solution of Eq. (3.1) is obtained by setting ψpt, xq “ e´iλt φpxq. We are left with
finding a solution of the time-independent equation (3.2). A family of solutions for such
equation is given by the plane waves on Rd :

φk pxq “ eik¨x “ eipk1 x1 `...`kd xd q for k P Rd . (3.3)

In fact:
1 2 |k|2
´ ∆x φk pxq “ pk1 ` . . . ` kd2 qeik¨x ” φk pxq . (3.4)
2 2
Thus, we found a first solution of the free Schrödinger equation, Eq. (3.1):
k2
t ik¨x
ψk px, tq “ e´i 2 e . (3.5)

However, the above solution does not make sense in quantum mechanics, since ψpt, ¨q R
L2 pRd q for all t: ż
dx |ψk pt, xq|2 “ `8 . (3.6)

Nevertheless, we can use the above unphysical solutions to construct physical solutions of the
Schrödinger equation, by using the fact that the Schrödinger equation is a linear equation:
a linear combination of solutions of Eq. (3.1) is a solution of Eq. (3.1). More precisely, we
shall consider solutions of the form:
ż ż
k2
ψpx, tq “ ρpkqψk px, tqdk ” ρpkqe´ip 2 t´k¨xq dk . (3.7)
Rd Rd

Formally, ψpx, tq is a solution of Eq. (3.1), with initial datum at t “ 0:

ż
ψpx, 0q ” ψ0 pxq “ ρpkqeik¨x dk . (3.8)
Rd

The questions we will address here are: for which class of ρpkq does the function ψpt, xq
makes sense from a quantum mechanical viewpoint, namely ψpt, ¨q P L2 pRd q?

3.1 The Fourier transform on L1

We are now ready to introduce the Fourier transform for L1 functions.
Definition 3.1. Let f P L1 pRd q. We define the Fourier transform fˆ ” Ff as
ż
1
pFf qpkq ” fˆpkq “ d dx e´ik¨x f pxq, k P Rd . (3.9)
p2πq 2

We define the inverse Fourier transform fˇ ” F ´1 f as:

ż
1
pF ´1 f qpkq ” fˇpkq “ d dx eik¨x f pxq . (3.10)
p2πq 2 Rd

7
Remark 3.2. Since |e´ik¨x | “ 1 and f P L1 pRd q, fˆ and fˇ are well defined:
ż
ˆ 1 1
|f pkq| ď d dx |f pxq| “ d }f }1 . (3.11)
p2πq 2 p2πq 2

The next lemma will be useful to study the regularity properties of the Fourier transform.
d 1 d
Lemma 3.3. Let Γ Ă R be ş an open interval, and f : R ˆ Γ Ñ C such that f px, γq P L pRx q
for all γ P Γ. Let Ipγq “ Rd f px, γqdx. Then, the following is true.
(a) If the map γ ÞÑ f px, γq is continuous for almost all x P Rd , and if there exists a function
g P L1 pRd q such that supγPΓ |f px, γq| ď gpxq for almost all x P Rd , then Ipγq is also
continuous.
(b) If the map γ ÞÑ f px, γq is continuously differentiable for almost all x P Rd , and if there
exists a function g P L1 pRd q such that supγPΓ |Bγ f px, γq| ď gpxq for almost all x P Rd ,
then Ipγq is also continuously differentiable. Moreover:
ż ż
dI d B
pγq “ f px, γqdx “ f px, γqdx . (3.12)
dγ dγ Rd R d Bγ

Proof. The proof immediately follows from the dominated convergence theorem, see [3].

Lemma 3.3 has important consequences on the behavior of the Fourier transform.
Theorem 3.4 (Riemann-Lebesgue.). Let f P L1 pRd q. Then:
! )
fˆ P C8 pRd q :“ f P CpRd q | lim sup |f pxq| “ 0 . (3.13)
RÑ8 |x|ąR

Proof. The continuity immediately follows from Lemma 3.3. The falloff at infinity will follow
from a result we will discuss later on.

Next, we will focus on the properties of the “nicest possible” functions, namely the
Schwartz functions. Later, we will come back on a more general class of functions, by using
approximation arguments.
Definition 3.5 (Schwartz functions.). The Schwartz space SpRd q is the set of functions
f P C 8 pRd q such that:
}f }α,β :“ }xα Bxβ f }8 ă 8 , (3.14)
for all multiindices α, β.
That is, the functions in SpRd q decay faster than any inverse polynomial in x, and the
same is true for all their partial derivatives. Obviously, if f P S then xα Bβ f P S for all
multiindices α and β. Also, SpRd q Ă Lp pRd q. Finally, the maps } ¨ }α,β : S Ñ r0, 8q are
norms.
Remark 3.6. Notice that Cc8 pRd q Ă SpRd q, which means that SpRd q is dense in Lp pRd q,
1 ď p ă 8.
Definition 3.7. We say that fn Ñ f in S if limnÑ8 }f ´ fn }α,β Ñ 0 for all α, β P Nd0 .
Proposition 3.8 (S is a metric space.). Convergence in S is equivalent to convergence with
respect to the metric:
8
ÿ }f ´ g}α,β
dS pf, gq “ 2´n sup . (3.15)
n“0 |α|`|β|“n 1 ` }f ´ g}α,β

Remark 3.9. Notice that dS pf, gq ď 2.

8
Proof. Let us first check that dS pf, gq is a metric. Positivity is trivial, and also symmetry:
dS pf, gq “ dS pg, f q. From the definition, we see that dS pf, gq “ 0 implies }f ´ g}0,0 “ }f ´
g}8 “ 0, that is f “ g. Also, the triangle inequality holds true: dS pf, gq ď dS pf, hq`dS ph, gq,
since } ¨ }α,β satisfies the triangle inequality and the function hpxq “ x{p1 ` xq is monotone
increasing and satisfies hpx ` yq ď hpxq ` hpyq. This shows that dS is a metric. Convergence
in S immediately implies convergence with respect to dS pf, gq. On the other hand, suppose
that dS pfn , f q Ñ 0. To prove convergence in S we use that, for all α, β there exists a constant
Cα,β ą 0 such that:
}fn ´ f }α,β ď Cα,β dS pfn , f q . (3.16)
Therefore, convergence with respect to dS implies convergence in S.

Theorem 3.10. The Schwartz space is complete.

Proof. Let pfm q be a Cauchy sequence in S. Then, pfm q is a Cauchy sequence with respect
to the (semi-)norms } ¨ }α,β . Also, convergence in S implies that xα Bxβ fm Ñ gα,β pxq in L8
norm, with gα,β P Cb pRd q, the space of continuous, bounded functions. This last fact is
implied by the completeness of Cb pRd q with respect to the } ¨ }8 norm, recall Remark 2.3.
We are left with showing that g :“ g0,0 P C 8 pRd q, and that xα Bxβ g “ gα,β . If so, g P S
and dS pfm , gq Ñ 0. For simplicity, let us consider the case d “ 1. We would like to show that
g P C 1 pRq and that Bx g “ g0,1 . Higher derivatives and higher dimensions can be studied in
the same way. For fm P S, we write:
żx
1
fm pxq “ fm p0q ` fm pyq dy . (3.17)
0

1
We know that fm Ñ g and fm Ñ g0,1 uniformly. Therefore, the m Ñ 8 limit of Eq. (3.17)
is: żx
gpxq “ gp0q ` g0,1 pyq dy . (3.18)
0
1 1
This proves that g P C pRq with g “ g0,1 .

Lemma 3.11 (Properties of F on S.). The maps F and F ´1 are continuous, linear maps
from S into itself. Moreover, for all α, β it holds:
´ ¯ ´ ¯
pikqα Bkβ Ff pkq “ FBxα p´ixqβ f pkq . (3.19)

Remark 3.12. In particular,

zqpkq “ ip∇k fˆqpkq
pxf and p∇
{ ˆ
x f qpkq “ ik f pkq . (3.20)

Proof. Let f P S. Recall:

ż
1
fˆpkq “ f pxqe´ik¨x dx . (3.21)
p2πqd{2 Rd

Then:
´ ¯ ż
p2πqd{2 pikqα Bkβ Ff pkq “ pikqα Bkβ e´ik¨x f pxq dx
Rd
ż
“ pikqα p´ixqβ e´ik¨x f pxq dx
Rd
ż
“ p´1q|α| pBxα e´ik¨x qp´ixqβ f pxq . (3.22)
Rd

Integrating by parts:
´ ¯ ż
p2πqd{2 pikqα Bkβ F pkq
` ˘
“ e´ik¨x Bxα p´ixqβ f pxq dx
Rd
´ ¯
” p2πqd{2 FBxα p´ixqβ f pkq . (3.23)

9
This shows that, in particular, Ff P C 8 . Moreover:

p1 ` |x|2 qd
ż
1
}fˆ}α,β “ ›k α Bkβ fˆ›8 ď
› › α β
|B x f pxq| dx
p2πqd{2 Rd x p1 ` |x|2 qd
ż
1 ˇ
2 d α β
ˇ 1
sup x f dx
ˇ ˇ
ď ` |x| q B x pxq
d{2 |x|2 qd
ˇp1 ˇ
p2πq xPRd Rd p1 `
m
ÿ
ď C sup }f }α̃,β̃ , (3.24)
j“0 |α̃|`|β̃|“j

with m “ maxt|α|, |β|u ` 2d, and for C ą 0 independent of f . Therefore, Ff P S. Eq. (3.24)
also shows that fn Ñ f in S implies fˆn Ñ fˆ in S. In particular, Eq. (3.24) can be used to
show that F : S Ñ S is continuous, with respect to the topology induced by dS p¨, ¨q. In fact,
suppose that fn Ñ f with respect to dS . Then, by Eq. (3.24), there exists Cα.β ą 0 such
that:
}fˆn ´ fˆ}α,β ď Cα,β dS pfn , f q . (3.25)

Theorem 3.13. The map F : S Ñ S is a continuous bijection, with inverse F ´1 .

Proof. We will show that F ´1 ˝ F “ 1S (the same proof gives F ˝ F ´1 “ 1S ). Since F ´1 ˝ F

and 1S are both continuous in S, it is sufficient to prove their equality on a dense subset of
S.
Lemma 3.14. Cc8 pRd q is dense in SpRd q.

Proof. (of Lemma 3.14.) Let:

" ` ˘
exp ´ 1{p1 ´ |x|2 q ` 1 for |x| ă 1
Gpxq “ (3.26)
0 otherwise.

Let f P SpRd q, and let fn pxq “ f pxqGpx{nq. Clearly, fn P Cc8 pRd q. Moreover, limnÑ8 }fn ´
f }α,β “ 0 for all α, β.

Let us now come back to the proof of Theorem 3.13. By Lemma 3.14, it is sufficient to
prove the claim of Theorem 3.13 on Cc8 pRd q. Let f P Cc8 pRd q. Let us denote by Wm Ă Rd
a cube in Rd , centered in the origin, with side 2m. Let us choose m large enough so that
supppf q Ă Wm . Let Km “ π{mZd . We can express the function f on Wm as the uniformly
convergent Fourier series: ÿ
f pxq “ fk eik¨x , (3.27)
kPKm

with Fourier coefficients:

p2πqd{2
ż ż
1 1
fk “ f pxqe´ik¨x dx “ f pxqe´ik¨x dx “ pFf qpkq . (3.28)
VolpWm q Wm VolpWm q Rd p2mqd

Therefore we have:
ÿ pFf qpkqeik¨x ´ π ¯d
f pxq “ . (3.29)
kPK
p2πqd{2 m
m

The observation is that the right-hand side of Eq. (3.29) is a Riemann sum, over cubes of
volume pπ{mqd and with k the center of the cube. Therefore, we have:
ÿ pFf qpkqeik¨x ´ π ¯d 1
ż
f pxq “ lim d{2
“ d{2
pFf qpkqeik¨x dk “ pF ´1 ˝ Ff qpxq .
mÑ8
kPKm
p2πq m p2πq Rd

(3.30)
This proves that F ´1 ˝ F “ 1Cc8 pRd q .

10
Proposition 3.15. Let f, g P SpRd q. Then:
ż ż
fˆpxqgpxqdx “ f pxqĝpxqdx . (3.31)
Rd Rd

Moreover,
}f }2 “ }fˆ}2 . (3.32)

Proof. By Fubini’s theorem,

ż ´ż ¯ ż ´ż ¯
´ik¨x
e f pkqdk gpxqdx “ e´ik¨x gpxqdx f pkqdk . (3.33)
Rd Rd Rd Rd

Therefore, p2πqd{2 dx fˆpxqgpxq “ p2πqd{2 dk ĝpkqf pkq. This proves Eq. (3.33). To prove
ş ş

Eq. (3.32), we use that Ff pxq “ F ´1 f pxq, which can be easily checked. Thus, Eq. (3.32)
follows as a special case of Eq. (3.33), choosing gpxq “ Ff pxq.

`
Example 3.16 (The Fourier transform of a Gaussian.). Let λ ą 0, and let gλ pxq “ exp ´
2˘
λ |x|2 be the Gaussian function. Then, we claim that:

|k|2
ˆ ˙
´d
ĝλ pkq “ λ exp ´
2 . (3.34)
2λ
To prove Eq. (3.34), we proceed as follows.
´ 2 ¯ By scaling, it is enough to consider the case
śd x
λ “ 1. Also, since g1 pxq “ i“1 exp ´ 2i , it is enough to consider the case n “ 1. We
have:
ż ż
1 ´ik¨x ´ x2
2 1 px`ikq2 k2
ĝ1 pkq “ 1 dxe e “ 1 dxe´ 2 ´ 2 ” g1 pkqf pkq, (3.35)
p2πq 2 p2πq 2
ş px`ikq2
1
where we defined f pkq “ 1 dxe´ 2 . By dominated convergence, we can differentiate
p2πq 2
under the integral sign:
ż ż
d dx ´
px`ikq2 dx d ´ px`ikq2
f pkq “ 1 p´px ` ikqqie 2 “ 1 i e 2 “ 0. (3.36)
dk R p2πq 2 R p2πq 2 dx

This means that f pkq is a constant and, in particular, f pkq “ f p0q “ 1. This proves Eq.
(3.34).

3.2 Solution of the free Schrödinger equation

Let us now come back to the Schrödinger equation for one free particle in Rd :
1
iBt ψpt, xq “ ´ ∆x ψpt, xq . (3.37)
2
Let us take the Fourier transform in both sides. Proceeding formally, we get:
1 2
iBt ψ̂pt, kq “ |k| ψ̂pt, kq . (3.38)
2
The advantage of taking the Fourier transform is that now we are left with an ordinary
differential equation of the first order. The solution is:
|k|2
t
ψ̂pt, kq “ e´i 2 ψ̂p0, kq . (3.39)
To get a solution of the original equation (3.37), we have to take the inverse Fourier transform.
We get:
|k|2
t
ψpt, xq “ pF ´1 e´i 2 Fψ0 qpxq , (3.40)
with initial datum ψp0, xq “ ψ0 pxq. The next theorem shows that the above formal manip-
ulation can be made rigorous for a suitable class of regular initial data.

11
Theorem 3.17 (Existence of a unique global solution for the free Schrödinger equation.). Let
ψ0 P SpRd q. Then, there exists a global solution ψ P C 8 pRt , SpRd qq of the free Schrödinger
equation with ψp0, xq “ ψ0 pxq for t ‰ 0, given by the expression:
ż
1 |x´y|2
ψpt, xq “ d{2
ei 2t ψ0 pyqdy . (3.41)
p2πitq Rd

Moreover, }ψpt, ¨q}L2 pRd q “ }ψ0 }L2 pRd q .

Proof. To begin, notice first that, for ψ0 P S, the expression (3.40) is well defined. Hence,
Eq. (3.40) is a solution of the free Schrödinger equation (3.37). Next, we shall show that
ψ P C 8 pRt , SpRd qq. Let us start by showing that t ÞÑ ψptq is differentiable. Let: ψpt,
9 xq :“
2 |k|2
´1 |k| ´i 2 t d
9 ¨q P SpR q. Furthermore, we claim that:
´ipF 2 e Fψ0 qpxq. Then, ψpt,
› ψpt ` hq ´ ψptq ›
lim ›
› 9 ››
´ ψptq “0 (3.42)
hÑ0 h α,β

with respect to any } ¨ }α,β . By continuity of F and of F ´1 , this is equivalent to:

› ψ̂pt ` hq ´ ψ̂ptq ›
lim ›
› 9̂ ››
´ ψptq “0, (3.43)
hÑ0 h α,β

|k|2
t
for all α, β. This follows from the smoothness of e´i 2 and from the decay of ψ̂0 pkq:
|k|2 |k|2
ˇ α β e´i 2 pt`hq ´ e´i 2 t |k|2 ´i |k|2 t ¯
› ψ̂pt ` hq ´ ψ̂ptq › ˇ ´ ˇ
›
› 9̂ ››
´ ψptq “ sup ˇk Bk `i e 2 pFψ0 qpkqˇ
ˇ
h α,β kPRd h 2
Ñ 0 as h Ñ 0. (3.44)

In the same way, one can prove that ψpt, xq P C k pRt , SpRd qq for any k ě 1, and hence
that ψpt, xq P C 8 pRt , SpRd qq. The uniqueness of the solution for ψ0 P S follows from the
uniqueness of the solution of (3.38). The formula (3.41) follows from an explicit computation,
using that:
żR c
´αx2 π
lim e dx “ , (3.45)
RÑ8 ´R α
for all α P C such that Re α “ 0. Finally, the isometry in L2 follows from the isometry
2
property of the maps F and F ´1 , proven in Eq. (3.32), and from the fact that |e´i|k| t{2 | “ 1.

Remark 3.18 (Decay of the solutions of the Schrödinger equation.). The formula (3.41)
immediately implies that:
}ψ0 }L1
sup |ψpt, xq| ď Ñ0 as t Ñ 8. (3.46)
xPRd p2πtqd{2

However, as we just proved, the L2 norm stays constant. This means that the solution of the
Schrödinger equation spreads in space. One speaks about the “spreading of the wave packet”.
Definition 3.19 (Polynomially bounded functions.). Let Cpol 8
pRd q be the space of the poly-
8
nomially bounded smooth functions: g P Cpol pRd q if g P C 8 pRd q and if:
npαq
|B α gpxq| ď Cα xxynpαq :“ Cα p1 ` |x|2 q 2 , (3.47)

for all α.
Motivated by Lemma 3.11, we introduce the notion of pseudodifferential operator.
8
Definition 3.20 (Pseudodifferential operator.). Let f P Cpol pRd q. Let Mf : S Ñ S be
the multiplication operator ψpxq Ñ f pxqψpxq. We define the pseudodifferential operator
f p´i∇x q : S Ñ S as:

pf p´i∇x qψqpxq :“ pF ´1 Mf Fψqpxq “ pF ´1 f pkqFψqpxq . (3.48)

12
Remark 3.21. Notice that the mapping Mf : S Ñ S is continuous. The continuity of Mf
and of F implies the continuity of f p´i∇x q. For f pkq “ k α , one naturally has f p´i∇q “
p´iq|α| Bxα . For polynomial functions f , the corresponding pseudodifferential operators are
differential operators.
Example 3.22 (Translations and the free propagator.). Let a P Rd and Ta “ e´ia¨k . One
8
has Ta P Cpol and for ψ P SpRd q one has:
ż ż
1 ´ik¨a ik¨x 1
pTa p´i∇qψqpxq “ dk e e ψ̂pkqdk “ eik¨px´aq ψ̂pkqdk “ ψpx ´ aq .
p2πqd{2 p2πqd{2
(3.49)
The operator Ta p´i∇q is called the translation operator. Another example is Pf pt, kq “
|k|2
t
e ´i 2 8
. One has Pf pt, ¨q P Cpol pRd q and hence:
ψpt, xq “ pPf pt, ´i∇x qψ0 qpxq . (3.50)
This operator is also called the free propagator, and one also writes:
i
ψptq “ e 2 ∆x t ψ0 . (3.51)
Example 3.23 (The heat equation and diffusion.). We can apply the previous theory to
solve the heat equation:
1
Bt f pt, xq “ ∆x f pt, xq , (3.52)
2
for f p0, ¨q “ f0 P SpRd q. Let t ą 0. The solution of Eq. (A.38) reads:
1
f ptq “ e 2 ∆x t f p0q “ W pt, ´i∇x qf0 , (3.53)
2
t ´ k2 8
with W pt, kq “ e . Notice that W ptq P Cpol only for t ě 0. In general, one cannot
establish existence of solutions of the heat equation for t ă 0. However, if fˆ0 has compact
support, the corresponding solution of the heat equation exists for all times.
Definition 3.24 (Convolutions.). Let f, g P S. We define the convolution f ˚ g as:
ż
pf ˚ gqpxq :“ f px ´ yqgpyqdy . (3.54)
Rd
Here we list some properties of the convolution operation.
Theorem 3.25. Let f, g, h P S. The following is true.
(i) pf ˚ gq ˚ h “ f ˚ pg ˚ hq and f ˚ g “ g ˚ f .
(ii) The map g ÞÑ f ˚ g from S to S is continuous.
(iii) It follows that:
fz˚ g “ p2πqd{2 fˆ ¨ ĝ , (3.55)
and also fxg “ p2πq´d{2 fˆ ˚ ĝ. Moreover, one has:
gp´i∇qf “ F ´1 pg fˆq “ p2πq´d{2 ǧ ˚ f . (3.56)
Proof. The properties piq and piiiq easily follows from the definition. Concerning piiq, conti-
nuity follows from:
f ˚ g “ p2πqd{2 F ´1 fˆFg ; (3.57)
that is, the convolution with f corresponds to the combination of Fourier transform, mul-
tiplication by f , and inverse Fourier transform. All these maps are continuous, and their
composition preserves continuity. Thus piiq holds true.

Example 3.26 (The heat equation.). Consider:

1 |x|2
Gpt, xq :“ p2πq´d{2 pF ´1 W qpt, xq “ d{2
e´ 2t . (3.58)
p2πtq
The function Gpt, xq is called the fundamental solution of the heat equation, and can be used
to construct more general solutions. In fact:
ż
1 |x´y|2
´ 2t
f pt, xq “ pW pt, ´i∇x qf0 qpxq “ pGptq ˚ f0 qpxq “ e f0 pyqdy . (3.59)
p2πtqd{2 Rd

13
3.2.1 Comparison between Schrödinger, heat and wave equations
To conclude this section, let us compare the free Schrödinger equation to the heat equation
and the wave equation. For simplicity, we shall consider the case d “ 1.

The wave equation. The wave equation can be used to describe the motion of an
oscillating string of length L. Let f px, tq be the wave deflection. The equation reads:

B2 B2
f pt, xq “ f pt, xq , (3.60)
Bt2 Bx2
with boundary conditions:
f pt, 0q “ f pt, Lq “ 0 . (3.61)
The acceleration of the string at the point is x is proportional to the curvature at the same
point, and this explains why the string oscillates.

The heat equation. The temperature profile for the temperature f px, tq in a rod of
length L, which temperature is kept to zero at both ends, satisfies the heat equation:

B B2
f pt, xq “ 2 f pt, xq , (3.62)
Bt Bx
with boundary condition:
f pt, 0q “ f pt, Lq “ 0 . (3.63)
The rate at which the temperature changes at the position x is proportional to the curvature
at that point. Therefore, the temperature converges to the constant value f pxq “ 0.

The Schrödinger equation. The motion of one free quantum particle in one dimension
is described by the Schrödinger equation:

B B2
ψpt, xq “ i 2 ψpt, xq , (3.64)
Bt Bx
with boundary condition:
ψpt, 0q “ ψpt, Lq “ 0 . (3.65)
As for the heat equation, it depends on the first time derivative. However, due to the presence
of the factor i, it gives rise to an oscillatory behavior of the solution. In fact, the function
ψpt, xq is now complex values, which we can picture as a time-dependent vector field in R2 .
Even though the rate of change of the wave function is proportional to the curvature at the
point x, because of the i factor it is described by an orthogonal vector to ψpxq. Therefore,
in general both the argument and the modulus of ψpt, xq change in time.

3.3 Tempered distribution

The goal of this section is to extend the notion of partial differential equation to functions
that are not smooth, in fact not even differentiable in the standard sense. In particular, we
shall be interested in formulating the Schrödinger equation for initial data which are only in
L2 pRd q.
Definition 3.27. The elements of the dual space S 1 pRd q of SpRd q are called tempered dis-
tribution.
Remark 3.28. The dual space V 1 of a topological vector space V is the space of continuous
linear maps from V to C. For f P V and T P V 1 , one defines the pairing of f and T as:

pf, T qV,V 1 :“ T pf q . (3.66)

Example 3.29. Let us discuss some examples of tempered distributions.

14
 (a) Let g : Rd ÞÑ C such that p1 ` |x|2 q´m gpxq P L1 pRd q for m P N. Then, the mapping
ż
Tg : S ÞÑ C , f ÞÑ gpxqf pxq dx (3.67)
Rd

is linear and continuous, hence Tg P S 1 .

Proof. Let fn Ñ f in S. Then,

ż
lim |Tg pfn ´ f q| ď lim |gpxq||fn pxq ´ f pxq| dx
nÑ8 nÑ8 Rd

ď }p1 ` |x|2 q´m g}1 lim }p1 ` |x|2 qm |fn ´ f |}8 “ 0 . (3.68)
nÑ8

(b) The delta-distribution is defined as:

δ:SÑC, f ÞÑ δpf q :“ f p0q . (3.69)

Therefore, δ P S 1 . One also writes:

ż
δpf q “ δpxqf pxq dx (3.70)
Rd

and: ż
δpx ´ aqf pxq dx “ f paq . (3.71)
Rd

The expression Eq. (A.37) is formal: there exists no function δ : Rd ÞÑ C that

gives (A.37). Nevertheless, one can approximate δ P S 1 by functions, more and more
“peaked”
ş at a, such that in the limit Eq. (A.37) holds true. For example, let g P L1 pRq
with dx gpxq “ 1. Let:
gn pxq :“ nd gpnxq . (3.72)
Then, by dominated convergence, for any continuous bounded function f , and in par-
ticular for all f P S, one has:
ż ´ż ż ¯
lim Tgn pf q “ lim gn pxqf pxq dx “ lim gn pxqf p0q dx ` gn pxqpf pxq ´ f p0qq dx
nÑ8 nÑ8 R nÑ8 R R
ż
“ f p0q ` lim gpyqpf py{nq ´ f p0qq dy “ f p0q ” δpf q . (3.73)
nÑ8 R

In the last step we used that the argument of the integral converges to zero pointwise in
x, as n Ñ 8, and dominated convergence theorem to bring the limit inside the integral.
Next, we shall introduce the notions of weak and weak˚ convergence.
Definition 3.30. Let V be a topological vector space and V 1 its dual.
(i) A sequence pmn q in V converges weakly to m P V if:

lim T pmn q “ T pmq , for all T P V 1 . (3.74)

nÑ8

One also writes w ´ limnÑ8 mn “ m or mn á m.

(ii) A sequence pTn q in V 1 converges in the weak˚ topology to T P V 1 if:

lim Tn pmq “ T pmq , for all m P V . (3.75)

nÑ8

˚
One also writes w˚ ´ limnÑ8 Tn “ T or Tn á T .
Theorem 3.31 (The adjoint map.). Let A : S Ñ S be a linear and continuous map. Then,
the map
A1 : S 1 Ñ S 1 , pA1 T qpf q :“ T pAf q for all f P S (3.76)
is weak˚ continuous. The map A1 is called the adjoint of A.

15
Proof. One has A1 T P S 1 , where A1 T ” T ˝ A is a continuous map on S. To prove the weak˚
˚
continuity of A1 : S 1 Ñ S 1 , we proceed as follows. Let Tn á T . Then, for each f P S:
lim pA1 Tn qpf q “ lim Tn pAf q “ T pAf q “ pA1 T qpf q , (3.77)
nÑ8 nÑ8

˚
that is A1 Tn á A1 T .

Remark 3.32. Strictly speaking, the above proof only shows sequential continuity in S 1 .
This does not immediately imply continuity in S 1 , since the topology of S 1 is not defined
through a metric. Nevertheless, the above argument can be repeated for a net on S 1 , and net
continuity would imply continuity.
Next, we define the Fourier transform on S 1
Definition 3.33. For T P S 1 , the Fourier transform Tp P S 1 is defined as:
Tppf q :“ T pfˆq for all f P S. (3.78)
Remark 3.34. In other words, FS 1 :“ FS1 . That is, the Fourier transform on S 1 is defined
as the adjoint of the Fourier transform on S.
Lemma 3.35. The Fourier transform F : S 1 Ñ S 1 is a weak˚ continuous bijection. More-
over, for f P S, Tpf “ Tfˆ.

Proof. Since F : S Ñ S is continuous, it follows from Theorem 3.31 that F : S 1 Ñ S 1 is

weak˚ continuous. Also, since pF ´1 FT qpf q “ T pFF ´1 f q “ T pf q, the Fourier transform on
S 1 is also bijective, with inverse F ´1 . Finally, let f P L1 . Then:
ż ż
Tpf pgq ” Tf pĝq “ f pxqĝpxq dx “ fˆpxqgpxq dx “ Tfˆpgq , (3.79)

where the second equality follows from Proposition 3.15.

Example 3.36 (The Fourier transform of the δ-distribution.). Let δpf q be the delta distri-
bution, δpf q “ f p0q. Then:
ż ż
ˆ ˆ 1 1
δ̂pf q “ δpf q “ f p0q “ f pxq dx ” f pxq dx “ Tg pf q , (3.80)
p2πqd{2 p2πqd{2

with g “ p2πq´d{2 the constant function. That is, the Fourier transform of the delta distri-
bution is the constant function g.
Let us now introduce the notion of derivative on the space of distributions S 1 .
Definition 3.37 (The distributional derivative.). For T P S 1 , we define its distributional
derivative Bxα T P S 1 as:
pBxα T qpf q :“ T pp´1q|α| Bxα f q . (3.81)
Lemma 3.38. The distributional derivative Bxα : S 1 Ñ S 1 is weak˚ continuous and extends
the notion of derivative on S; that is, for g P S we have:
Bxα Tg “ TBxα g . (3.82)
Proof. As an adjoint map, the derivative Bxα is continuous thanks to Theorem 3.31. The
property Eq. (3.82) follows from the integration by parts formula:
ż ż
pBxα Tg qpf q “ Tg pp´1q|α| Bxα f q “ gpxqp´1q|α| Bxα f pxq dx “ f pxqBxα gpxq dx “ TBxα g pf q .
(3.83)

Example 3.39 (The derivative of the delta distribution.). It follows that:

pBxα δqpf q “ δpp´1q|α| Bxα f q “ p´1q|α| Bxα f p0q . (3.84)
d
For the Heaviside function θpxq “ 1r0,8q pxq on R one has: dx θ “ δ.

16
 8
Lemma 3.40. Let g P Cpol . Then, pgT qpf q “ T pgf q defines a weak˚ continuous map from
1 1
S to S . In general, one cannot define the product of two distributions, but one can define
8
the product of a distribution and of a function in Cpol .

Proof. Exercise.

Lemma 3.41. Let g P S and g̃pxq “ gp´xq. Then pg ˚ T qpf q :“ T pg̃ ˚ f q defines a weak˚
continuous map from S 1 to S 1 , which extends the convolution on S: g ˚ Th “ Tg˚h for h P S.

Proof. Exercise.

This result allows to prove the following theorem.

Theorem 3.42. S is dense in S 1 in the weak˚ topology.

Proof. Let us give a sketch of the proof. We want to show that for all T P S 1 there exists
pϕn q Ă S such that:
˚
Tϕn á T . (3.85)
We proceed as follows.
ş Let pgn q Ă S such that pgn ˚ f q Ñ f in S (e.g., gn pxq “ nd gpnxq,
with g P S and g “ 1.) Then, we write:

pgn ˚ T qpf q “ T pg̃n ˚ f q

´ż ¯
“ T dy g̃n p¨ ´ yqf pyq
ż
“ dy T pg̃n,y qf pyq , (3.86)

with g̃n,y p¨q “ g̃n p¨ ´ yq. Thus, we would be tempted to say that pgn ˚ T q “ Tξn , with
8
ξn pyq “ T pg̃n,y q. To prove this, we simply notice that ξn P Cpol pRd q (exercise), which
implies that ξn f P S, and hence that it is an integrable function. Thus, by the weak˚
continuity of the convolution, Lemma 3.41, we just proved that for each T P S 1 there exists
8
ξn P Cpol such that:
˚
Tξn á T . (3.87)
To conclude, we would like to show that the sequence pξn q can be replaced by a sequence
pϕn q in S. We proceed as follows. Let Gpxq as in Eq. (3.26). Let: ϕn pxq “ ξn pxqGpx{nq.
Then, being Gpx{nq compactly supported, ϕn P S. Notice that Tϕn pf q “ Tξn pGp¨{nqf q. Fix
ε ą 0. By what we just proved, for n large enough:
ˇ ˇ
ˇTξn pGp¨{nqf q ´ T pGp¨{nqf qˇ ď ε{3 . (3.88)
ˇ ˇ

(Notice that the argument of the distributions is n-dependent. Nevertheless, this is not a
problem, since the } ¨ }α,β norms of Gp¨{nqf are all bounded uniformly in n.) Also, by the
continuity of T : ˇ ˇ
ˇT pGp¨{nqf q ´ T pf qˇ ď ε{3 , (3.89)
ˇ ˇ

where we used that Gp¨{nqf ´ f Ñ 0 in S, as n Ñ 8. Finally, again by Eq. (3.87):

ˇ ˇ
ˇT pf q ´ Tξn pf qˇ ď ε{3 . (3.90)
ˇ ˇ

All together, for any f P S and for any ε ą 0 there exists n0 P N such that for n ě n0 :
ˇ ˇ
ˇTξn pf q ´ Tϕn pf qˇ ď ε . (3.91)
ˇ ˇ

This, together with Eq. (3.87), implies that:

˚
Tϕn á T . (3.92)

17
 Next, we discuss the solution of the free Schrödinger equation in the sense of distributions.
We say that ψptq P C 8 pRt , S 1 pRd qq is a distributional solution of the Schrödinger equation
if:
d 1
i pf, ψptqqS,S 1 “ pf, ´ ∆ψptqqS,S 1 , (3.93)
dt 2
for all functions f P SpRd q.
Proposition 3.43. Let ψ0 P S 1 . Then, there exists a unique, global solution ψptq P
C 8 pRt , S 1 pRd qq of the Schrödinger equation, given by
|k|2
t
ψptq “ F ´1 e´i 2 Fψ0 . (3.94)

Proof. By Lemma 3.35 and by the fact that F and F ´1 are maps from S 1 to S 1 , we know
that ψptq P S 1 pRd q. To conclude, we show that ψptq is a solution of the Schrödinger equation
in the sense of distributions. Let f P S be a test function. Then:
d d |k|2
i pf, ψptqqS,S 1 “ i pFe´i 2 t F ´1 f, ψ0 qS,S 1
dt dt
|k|2 |k|
2
“ pFe´i 2 t F ´1 f, ψ0 qS,S 1
2
|k|2 1
“ p´Fe´i 2 t F ´1 ∆f, ψ0 qS,S 1
2
1 |k|2
“ p´ ∆f, F ´1 e´i 2 t Fψ0 qS,S 1
2
1
“ pf, ´ ∆ψptqqS,S 1 . (3.95)
2
The regularity in time of the mapping ψptq : S Ñ C can be easily checked.

3.4 Long time asymptotics of the momentum operator

We have proven that, for ψ0 P S, the solution of the free Schrödinger equation is given by:
ż
1 |x´y|2
ψpt, xq “ d{2
dy ei 2t ψ0 pyq . (3.96)
p2πitq

The probability for finding the quantum particle in the region A Ă Rd is given by:
ż
P pXptq P Aq “ |ψpt, xq|2 dx . (3.97)
A

Next, we want to determine the “velocity distribution” of the quantum particle. Since the
velocity at a fixed time is not defined in standard quantum mechanics, we shall consider the
asymptotic speed for large times, which we define as:
´ Xptq ¯ ż
lim P P A :“ lim P pXptq P tAq “ lim |ψpt, xq|2 dx . (3.98)
tÑ8 t tÑ8 tÑ8 tA

Notice that choice of the origin of the reference frame does not play any role. To get an
expression for the above limit, we shall use the next lemma.
Lemma 3.44. Let ψptq be the solution of the free Schrödinger equation, with ψp0q “ ψ0 P S.
Then:
x2
ei 2t
ψpt, xq “ ψ̂0 px{tq ` rpt, xq , (3.99)
pitqd{2
with limtÑ8 }rptq}L2 “ 0.

18
Proof. We have, by Eq. (3.96):
x2
ei 2t
ż
1 ´i xt y
´ y2
i 2t
¯
ψpt, xq “ e e ` 1 ´ 1 ψ0 pyq dy
pitqd{2 p2πqd{2
x2
ei 2t ´
ż
1 ´i xt y
´ y2
i 2t
¯ ¯
“ ψ̂0 px{tq ` e e ´ 1 ψ0 pyq dy
pitqd{2 p2πqd{2
x 2
ei 2t ´ ¯
“ ψ̂0 px{tq ` ĥpt, x{tq , (3.100)
pitqd{2
and hence:
x2
ei 2t
rpt, xq “ ĥpt, x{tq . (3.101)
pitqd{2
To prove the claim on the L2 norm, we proceed as follows:
ż ż ż ż
1
2
}rpt, ¨q}L2 “ |rpt, xq| dx “ d |ĥpt, x{tq| dx “ |ĥpt, yq| dy “ |hpt, yq|2 dy . (3.102)
2 2 2
t
Now, notice that hpt, xq Ñ 0 pointwise as t Ñ 8. Also, |hpt, xq|2 ď 4|ψ0 pxq|2 . Therefore, by
dominated convergence theorem:
ż
lim |hpt, xq|2 dx “ 0 . (3.103)
tÑ8

This concludes the proof.

Theorem 3.45. Let ψpt, xq be a solution of the free Schrödinger equation and let A Ă Rd
measurable. Then:
´ Xptq ¯ ż
lim P P A “: lim Pψt ptΛq “ |ψ̂0 ppq|2 dp . (3.104)
tÑ8 t tÑ8 A

Proof. By Lemma 3.44, we have:

ż ż ż
1
|ψpt, xq|2 dx “ d |ψ̂0 px{tq|2 dx ` Rptq “ |ψ̂0 ppq|2 dp ` Rptq , (3.105)
tA t tA A

where, following the proof of the Lemma:

ż ´1 ż ¯
lim Rptq “ lim |rpt, xq|2 dx ` lim 2Re d ψ̂0 px{tqĥpt, x{tqdx
tÑ8 tÑ8 tA tÑ8 t tA
´ż ¯
“ lim 2Re ψ̂0 ppqĥpt, pq . (3.106)
tÑ8 A

By the Cauchy-Schwarz inequality we have:

ˇż ˇ
lim ˇ ψ̂0 ppqĥpt, pqdpˇ ď lim }ψ̂0 }L2 }ĥptq}L2 “ 0 . (3.107)
ˇ ˇ
tÑ8 tA tÑ8

Remark 3.46. • If we would not have set the mass m to 1, the probability in the left-
hand side of Eq. (3.104) should have been replaced by P pmXptq{t P Λq. Therefore,
the above result allows to control the asymptotic distribution of the momentum of the
quantum particle.
• The operator P :“ ´i∇x is called the momentum operator. The expectation value of
the momentum operator is given by:
ż ż ż
Eψt pP q :“ xψt , P ψt y :“ ψpt, xqpP ψqpt, xqdx “ ψ̂pt, pqpψ̂pt, pqdp “ p|ψ̂p0, pq|2 dp ,
Rd Rd Rd
(3.108)
where we used that |ψ̂pt, pq| “ |ψ̂p0, pq|. Thus, the quantum mechanical expectation
value of the momentum operator is equal to its expectation value with respect to the
asymptotic momentum distribution.

19
 3.5 Properties of Hilbert spaces
Recall the definition of Hilbert space, given in Section 2.3. In this section we shall spell out
some important properties of Hilbert spaces, that will play a role in the following discussion.
Definition 3.47. Let H be a Hilbert space. A sequence pϕn q in H is called an orthonormal
sequence if xϕn , ϕm y “ δn,m .
The next proposition is an immediate consequence of notion of orthogonality.
Proposition 3.48. Let pϕj qjPN be a orthonormal sequences in H. For any ψ P H, let us
rewrite:
n
ÿ ´ n
ÿ ¯
ψ “ xϕj , ψyϕj ` ψ ´ xϕj , ψyϕj
j“1 j“1

“: ψn ` ψnK . (3.109)

Then, xψn , ψ K y “ 0 and:

xψ, ψy “ xψn , ψn y ` xψnK , ψnK y . (3.110)

Proof. Exercise.

Proposition 3.48 implies the validity of two important inequalities, the Cauchy-Schwarz
inequality and the Bessel inequality.
Corollary 3.49. (a) Let pϕj qjPN be an orthonormal sequences in H. Let ψ P H and n P N.
Then:
n
ÿ
}ψ}2 ě |xϕj , ψy|2 (Bessel inequality). (3.111)
j“1

(b) Let ϕ, ψ P H. Then:

|xϕ, ψy| ď }ϕ}}ψ} , (Cauchy-Schwarz inequality). (3.112)

Proof. Eq. (3.111) immediately follows from Proposition 3.48. Eq. (3.112) follows from Eq.
(3.111), after choosing ϕ1 “ ϕ{}ϕ} and n “ 1.

Proposition 3.50 (Polarization identity.). Let H be a Hilbert space. Let ψ, ϕ P H. Then:

1
xϕ, ψy “ p}ϕ ` ψ}2 ´ }ϕ ´ ψ}2 ´ i}ϕ ` iψ}2 ` i}ϕ ´ iψ}2 q . (3.113)
4
Proof. Eq. (3.113) follows from the following identity, valid for any sesquilinear form1 B :
X ˆ X Ñ C, with X a complex vector space:
1
Bpx, yq “ pBpx`y, x`yq´Bpx´y, x´yq´iBpx`iy, x`iyq`iBpx´iy, x´iyqq . (3.114)
4

Definition 3.51. An orthonormal sequence pϕj qjPN in H is called an orthonormal basis if

for all ψ P H:
ÿ8
ψ“ xϕj , ψyϕj . (3.115)
j“1

Remark 3.52. Notice that the series converges in H. In fact, by Bessel’s inequality,
n
ÿ
|xϕj , ψy|2 ď }ψ}2 .
j“1

1
A map B : X ˆ X Ñ C is called a sesquilinear form if it is linear in the second variable and antilinear in the
first variable.

20
 řn ´ř ¯
2 n
Thus, limnÑ8 j“1 |xϕj , ψy| exists. Consider the sequence of partial sums j“1 xϕj , ψyϕj .
1
Let n ą n. We have:
›ÿn n1
ÿ ›2 n1
ÿ
, ψyϕ , ψyϕ |xϕj , ψy|2 , (3.116)
› ›
› xϕj j ´ xϕj j› “
j“1 j“1 j“n
´ř ¯
n
which vanishes as n Ñ 8. Hence, j“1 xϕj , ψyϕj is a Cauchy sequence in H. Being H
ř8
complete, j“1 xϕj , ψyϕj P H.
Definition 3.53. A topological vector space is called separable if it contains a countable,
dense subset.
Proposition 3.54. A Hilbert space is separable if and only if it contains an orthonormal
basis.

Proof. Let pϕj q be a ONB. Then, the following set is a dense and countable subset of H:
N
!ÿ )
spanQ`iQ tϕj | j P Nu :“ paj ` ibj qϕj | N P N , aj , bj P Q . (3.117)
j“1

Let us now prove the converse statement. Suppose that pϕj qjPN is a dense and countable
subset of H. Let pϕj qjPJ Ď pϕj qjPN be a subset of linearly independent vectors in pϕj qjPN ,
dense in H. This subset can be used to define a ONB, via the Gram-Schmidt method.

Proposition 3.55. Let pϕj q be an orthonormal basis for H. Then, the following inequality
holds true:
8
ÿ
}ψ}2 “ |xϕj , ψy|2 (Parseval equality.) (3.118)
j“1

Proof. Eq. (3.118) immediately follows from the definition and the continuity of the scalar
product:
A N
ÿ M
ÿ E
}ψ}2 “ lim xϕj , ψyϕj , lim xϕi , ψyϕi
N Ñ8 M Ñ8
j“1 i“1
A N
ÿ M
ÿ E N
ÿ
“ lim lim xϕj , ψyϕj , xϕi , ψyϕi “ lim |xϕj , ψy|2 . (3.119)
N Ñ8 M Ñ8 N Ñ8
j“1 i“1 j“1

Remark 3.56 (`2 as a coordinate space for a separable Hilbert space.). Let pϕj q Ă H be a
ONB. Then, the Parseval equality implies that the following mapping is an isometry:
U : H Ñ `2 , ϕ ÞÑ pxϕj , ψyqjPN . (3.120)
2
ř8
In particular, for each sequence c P ` we can associate a series j“1 cj ϕj , which converges
in norm:
8
› ÿ ›2 8
ÿ
cj ϕj › “ |cj |2 Ñ 0 as N Ñ 8; (3.121)
› ›
›
j“N j“N

this means that U is also surjective, i.e. it is an isometric isomorphism. Therefore, each
separable Hilbert space is isometrically isomorphic to `2 and each ONB generates an isometric
isomorphism. Thus, we can identify `2 as the coordinate space for separable Hilbert spaces
of infinite dimension.
Example 3.57. Consider L2 pr0, 2πsq. It is a separable Hilbert space, and a ONB is provided
by ϕk pxq “ ?12π eikx , k P N. Let ψ P L2 pr0, 2πsq, and consider its Fourier series:
8
ÿ
ψ“ xϕk , ψyϕk . (3.122)
k“´8

The Fourier series provides an isometric isomorphism between `2 and L2 .

21
Proposition 3.58 (Characterization of an orthonormal basis.). An orthonormal sequence
pϕj qjPI in H is an orthonormal basis of H if and only if:

xϕj , ψy “ 0 for all j P I ñψ“0. (3.123)

Proof. Let pϕj qjPI be a ONB of H. Suppose that xϕj , ψy “ 0 for all j P I. Then, by
definition of ONB, Eq. (3.115), ψ “ 0. Let us now prove the converse implication. Let pϕj q
be an orthonormal sequence in H, and let φ P H. By Bessel’s inequality, we have, for all
n P N:
ÿn
|xϕj , φy|2 ď }φ}2 . (3.124)
j“1
řn
Being the sequence n ÞÑ j“1 |xϕj , φy|2 nondecreasing and bounded, the n Ñ 8 limit
řn
exists: limnÑ8 j“1 |xϕj , φy|2 “ jPI |xϕj , φy|2 . In particular, this implies that the series
ř
ř
jPI xϕj , φyϕj is convergent in H. Consider the vector:
ÿ
ψ “φ´ xφ, ϕj yϕj . (3.125)
jPI

By construction, xψ, ϕj y “ 0 for all j P I. By assumption, this implies that ψ “ 0, hence:

ÿ
φ “ xφ, ϕj yϕj , for all φ P H. (3.126)
jPI

Therefore, tϕj ujPI is an ONB of H. This concludes the proof.

Definition 3.59. Let M Ă H. We define its orthogonal complement as:

! )
M K :“ ψ P H | xϕ, ψy “ 0 for all ϕ P M . (3.127)

Remark 3.60. It follows that M X M K “ t0u. Also, being xϕ, ¨y linear and continuous, M K
is a closed subspace of M .
Theorem 3.61. Let M Ă H be a closed subspace of H. Then:

H “ M ‘ MK . (3.128)

That is, every element ψ P H can be rewritten in a unique way as ψ “ ϕ ` ϕK with ϕ P M

and ϕK P M K .

Proof. Let ψ P H. If ψ P M , or ψ P M K , there is nothing to prove. Suppose that ψ R M ,

ψ R M K . Let pvk q be a minimizing sequence:

lim }ψ ´ vk }2 “ inf }ψ ´ v}2 . (3.129)

kÑ8 vPM

By using that } ¨ }2 “ x¨, ¨y, we have:

}ψ ´ v}2 “ F pvq ` }ψ}2 , F pvq :“ }v}2 ´ 2Re xψ, vy . (3.130)

Therefore, limkÑ8 F pvk q “ inf vPM F pvq “: α. Our preliminary goal is to show that vk Ñ v
in M . To prove this, we write:

F pvk q ` F pvl q “ }vk }2 ´ 2Re xψ, vk y ` }vl }2 ´ 2Re xψ, vl y

1´ ¯
“ }vk ` vl }2 ` }vk ´ vl }2 ´ 2Re xψ, vk ` vl y
2
› v ` v ›2 A v `v E 1
› k l› k l
“ 2› › ´ 4Re ψ, ` }vk ´ vl }2
2 2 2
´v ` v ¯ 1 1
k l
“ 2F ` }vk ´ vl } ě 2α ` }vk ´ vl }2 .
2
(3.131)
2 2 2

22
Since F pvk q, F pvl q Ñ α as k, l Ñ 8, we get that }vk ´vl } Ñ 0. Being pvk q a Cauchy sequence,
and since H is complete, vk Ñ v in H. Also, since M is closed, v P M . By continuity of the
scalar product, α “ F pvq. Our next goal is to show that ψ ´ v P M K . If so, this provides
one decomposition ψ “ v ` v K , with v P M and v K P M K .
Let ṽ P M and let f ptq :“ F pv ` tṽq. Then, by definition of v:

f ptq ě F pvq ” f p0q , for all t P R. (3.132)

Thus, t “ 0 is a minimum of f ptq. In particular, f 1 p0q “ 0. Let us compute the derivative.

A simple computation shows that:

0 “ f 1 p0q “ 2Re xψ ´ v, ṽy . (3.133)

Replacing ṽ with iṽ, we get the same identity but with Re replaced by Ran. Hence:

0 “ xψ ´ v, ṽy “ 0 , for all ṽ P M . (3.134)

In conclusion, ψ ´ v P M K , as claimed; thus, ψ “ v` K v. Let us now prove uniqueness of

the splitting. Suppose there exists v1 , v2 P M and v1K , v2K such that:

ψ “ v1 ` v1K “ v2 ` v2K . (3.135)

Then, v1 ´v2 “ v2K ´v1K , which means that v1 ´v2 “ 0 and v1K ´v2K “ 0, since M XM K “ t0u.

3.6 The Fourier transform in L2

Definition 3.62. Let X and Y be two normed spaces. An operator L : X Ñ Y between X
and Y is called bounded if there exists C ă 8 such that:

}Lx}Y ď C}x}X , for all x P X. (3.136)

Proposition 3.63. Let X and Y be two normed spaces. Let LpX, Y q be the set of the
bounded linear operators from X to Y . Let:

}L}LpX,Y q :“ sup}x}X “1 }Lx}Y . (3.137)

Then, } ¨ }LpX,Y q defines a norm on LpX, Y q. Moreover, if Y is complete then LpX, Y q is

complete as well, that is it is a Banach space.

Proof. It is easy to check that } ¨ }LpX,Y q defines a norm on LpX, Y q. Let now prove that if
Y is complete then LpX, Y q is complete as well. Let pLn q be a Cauchy sequence in LpX, Y q:

}Ln ´ Lm }LpX,Y q Ñ 0 as n, m Ñ 8. (3.138)

Then, pLn xq is Cauchy sequence in Y , since

}Ln x ´ Lm x}Y ď }Ln ´ Lm }LpX,Y q }x}Y . (3.139)

Being Y complete, Ln x Ñ y P Y , as n Ñ 8. We define Lx :“ y. It is easy to show that L is

a linear operator. Let us prove that L is a bounded operator. By the Cauchy property, we
have, for all ε ą 0, for n, m large enough:

sup }Ln x ´ Lm x}Y ď ε . (3.140)

}x}X “1

Therefore, dropping the sup and taking the m Ñ 8 limit:

}Ln x ´ Lx}Y ď ε ñ }Lx}Y ď C , (3.141)

uniformly in x, for all x such that }x}X “ 1. This proves that L P LpX, Y q. Due to the
arbitrariness of ε, Eq. (3.141) also proves that Ln Ñ L in LpX, Y q. This concludes the
proof.

23
Theorem 3.64. Let X and Y be two normed spaces. Let L : X Ñ Y be a linear operator.
Them, the following statements are equivalent:
(i) L is continuous at 0.
(ii) L is continuous.
(iii) L is bounded.

Proof. piiiq ñ piq. In fact, let }xn } Ñ 0. Then, }Lxn } ď }L}}xn } Ñ 0.

Let us now show that piq ñ piiq. Let }xn ´ x} Ñ 0 and let L be continuous at 0. Then,
}Lxn ´ Lx} “ }Lpxn ´ xq} Ñ 0.
Finally, let us prove that piiq ñ piiiq. Suppose that L is continuous but not bounded: that
xn
is, there exists a sequence pxn q with }xn } “ 1 such that }Lxn } ě n. Then, let zn :“ }Lx n}
.
1
It follows that }zn } ď n , but }Lzn } “ 1, which contradicts continuity.

Example 3.65 (Unbounded linear operators.). Let `0 “ tpxn q P `1 | DN ř8P N : xn “ 0 @n ě

N u be the space of finite sequences, equipped with the norm }x}`1 :“ n“1 |xn |. Then, the
operator T : `0 Ñ `0 such that x ÞÑ T x “ px1 , 2x2 , 3x3 , . . .q is unbounded, since }T en } “ n
but }en } “ 1.
Theorem 3.66 (Extension of densely defined linear bounded operators.). Let Z Ă X be
a dense subspace of a normed space X and let Y be a Banach space. Let L : Z Ñ Y be
linear and bounded. Then, L admits a unique linear and bounded extension L̃ P LpX, Y q
with L̃ æZ “ L and
}L̃}LpX,Y q “ }L}LpZ,Y q . (3.142)

Proof. Let x P X. Then, there exists a sequence pzn q Ă Z such that }zn ´ x}X Ñ 0.
Being pzn q convergent, the sequence pzn q is also a Cauchy sequence. Thus, }Lzn ´ Lzm }Y “
}Lpzn ´ zm q}Y ď }L}}zn ´ zm }X , which means that pLzn q is also a Cauchy sequence in
Y . Since Y is complete, Lzn Ñ y P Y . Let us now prove that the limit y does not
depend on the choice of the sequence pzn q (provided it converges to x). Let pzn1 q be another
sequence in Z, such that }zn1 ´ x}X Ñ 0. Consider the new sequence z1 , z11 , z2 , z21 , . . .. By
assumption, also this new sequence converges to x, and by following the previous argument,
Lz1 , Lz11 , Lz2 , Lz21 . . . converges to ỹ P Y . But since every subsequence of a convergent
sequence converges to the same limit, we have y “ lim Lzn “ lim Lzn1 “ ỹ. Therefore, we
can define L̃x :“ y. The linearity of L follows immediately from the previous construction.
The boundedness follows from:

}L̃x}Y “ lim }Lzn }Y ď lim }L}}zn }X “ }L}}x}X . (3.143)

nÑ8 nÑ8

Therefore, L̃ is bounded, and also continuous, by Theorem 3.64. Finally, the extension L̃ of
L is unique: this follows from the fact that two continuous maps which coincide on a dense
subset are equal.

Next, we shall extend the Fourier transform on L2 .

Theorem 3.67 (The Fourier transform on L2 .). The Fourier transform F : pSpRd q, }¨}L2 q Ñ
L2 pRd q can be uniquely extended to a bounded linear operator on L2 pRd q. Moreover, for all
f P L2 :
}Ff }L2 “ }f }L2 (3.144)
and FF ´1 “ F ´1 F “ 1L2 .
Remark 3.68. Eq. (3.144) takes the name of Plancherel’s theorem.

Proof. By Theorem 2.13, the space S is dense in L2 . The extension of F to a bounded linear
operator on L2 follows from Theorem 3.66. Moreover, as proven in Theorem 3.13,

F ´1 F æS “ FF ´1 æS “ 1S . (3.145)

Being F, F ´1 , 1 continuous, and being S dense in L2 , Eq. (3.145) holds as an identity on

L2 .

24
Definition 3.69 (Unitary operator.). A bounded linear operator U P LpH1 , H2 q is called
unitary if it is surjective and isometric, that is }U ψ}H2 “ }ψ}H1 for all ψ P H1 .
Remark 3.70. By the polarisation identity, it immediately follows that U “preserves angles”,
that is:
xU ψ, U ϕyH2 “ xψ, ϕyH1 for all ϕ, ψ P H1 . (3.146)
Remark 3.71. The Fourier transform F : L2 Ñ L2 is unitary.
As an application of the Fourier transform in L2 , consider the propagator of the free
Schrödinger equation, defined in Eq. (3.50). By extending the Fourier transform to L2 , the
free propagator can also be extended to an operator on L2 :
k2
Pf ptq : L2 pRd q Ñ L2 pRd q , Pf ptq “ F ´1 e´i 2 t
F. (3.147)

It follows that Pf ptq is a unitary operator, for all t P R. Moreover, it satisfies the following
composition property:
k2 k2 k2
s t
Pf psqPf ptq “ F ´1 e´i 2 FF ´1 e´i 2 F “ F ´1 e´i 2 ps`tq
F “ Pf ps ` tq . (3.148)

Therefore, one says that Pf : R Ñ LpL2 q is a unitary group. In the next section we will show
that the function:
ψptq :“ Pf ptqψ0 , ψ0 P L2 pRd q (3.149)
solves the Schrödinger equation in the L2 sense. Before doing that, let us first check that

ψ : R Ñ L2 pRd q , ψ0 ÞÑ ψptq “ Pf ptqψ0 (3.150)

is continuous. By dominated convergence:

ż ˇ k2 ˇ2
k2
ψpt0 q}2L2 Pf pt0 qqψ0 }2L2
ˇ ´i 2 t
´ e´i 2 t0 ˇ |ψ̂0 pkq|2 dk Ñ 0 (3.151)
ˇ
}ψptq ´ “ }pPf ptq ´ “ ˇe
Rd

as t Ñ t0 . This proves the continuity of ψptq. Let us now check differentiability. Again by
dominated convergence, we see that ψ : R Ñ L2 pRd q is differentiable if and only if:

|k|4 |ψ̂0 pkq|2 (3.152)

is integrable, that is when |k|2 ψ̂0 pkq P L2 pRd q. To conclude, let us discuss the continuity
properties of the unitary group Pf . In particular, let us consider }Pf ptq ´ Pf pt0 q}LpL2 q , with
} ¨ }LpL2 q defined in Proposition 3.63. We have:
› k2 k2
› ˇ k2 k2
ˇ
}Pf ptq ´ Pf pt0 q}LpL2 q “ ›e´i 2 t ´ e´i 2 t0 › “ sup ˇe´i 2 t ´ e´i 2 t0 ˇ “ 2 , (3.153)
› › ˇ ˇ
LpL2 q kPRd

where we used that F is unitary, and that it leaves L2 invariant. Therefore, the unitary
group Pf is not continuous with respect to the topology of the bounded operators. However,
one might have continuity with respect to different topologies.
Definition 3.72. Let pAn q be a sequences in LpHq and A P LpHq.
(a) We say that An converges to A in norm if:

lim }An ´ A}LpHq “ 0 . (3.154)

nÑ8

One writes also limnÑ8 An “ A or An Ñ A.

(b) We say that An converges strongly (or pointwise) to A if:

lim }An ψ ´ Aψ}H “ 0 for all ψ P H. (3.155)

nÑ8

s
One writes also s ´ limnÑ8 An “ A or An Ñ A.

25
 (c) We say that An converges weakly to A if:

lim |xϕ, pAn ´ Aqψy| “ 0 for all ϕ, ψ P H. (3.156)

nÑ8

w
One writes also w ´ limnÑ8 An “ A or An Ñ A.
Remark 3.73. These notions of convergence verify the following chain of implications:

norm convergence ñ strong convergence ñ weak convergence. (3.157)

The reverse implications are in general not correct.

3.7 Unitary groups and their generators

In this section we shall discuss in which sense ψptq “ Pf ptqψ0 with ψ0 P L2 solves the free
Schrödinger equation:
d 1
i ψptq “ ´ ∆ψptq . (3.158)
dt 2
As we have seen in the previous section, ψptq is differentiable in the strong sense if |k|2 ψ̂ptq P
L2 . Moreover, the distributional derivative:

1 |k|2
´ ∆ψptq “ F ´1 ψ̂ptq (3.159)
2 2

is in L2 if and only if |k|2 ψ̂ptq P L2 . Also,

k2
|k|2 ψ̂ptq “ |k|2 e´i 2 t
ψ̂0 P L2 (3.160)

if and only if |k|2 ψ̂0 P L2 . Therefore, if the initial datum satisfies |k|2 ψ̂0 P L2 , then |k|2 ψ̂ptq P
L2 for all times, and ψptq solves the Schrödinger equation in the L2 sense: Eq. (3.158) holds
as an identity between L2 functions.
Definition 3.74 (Sobolev spaces.). Let m P Z. The m-th Sobolev space H m pRd q Ă S 1 pRd q
is the set of distributions f P S 1 pRd q such that fˆ is a measurable function and:
m
p1 ` |k|2 q 2 fˆ P L2 pRd q . (3.161)

For m ě 0, it follows that H m Ă L2 .

Remark 3.75. Let us consider again the propagator of the free Schrödinger equation:
k2
Pf : R Ñ LpL2 q , t ÞÑ Pf ptq “ F ´1 e´i 2 t
F. (3.162)

It satisfies the following properties:

(a) Pf ptq is unitary for all t P R.
(b) Pf is strongly continuous: t ÞÑ Pf ptqψ is continuous for all ψ P L2 .
(c) Pf has the group property: Pf psqPf ptq “ Pf pt ` sq for all s, t P R.
Moreover,
(d) For all ψ0 P L2 , ψptq “ Pf ψ0 is a solution in the sense of distributions.
(e) For all ψ0 P H 2 Ă L2 , ψptq “ Pf ptqψ0 is a solution in the L2 sense: the map R Q t ÞÑ
ψptq P L2 is differentiable and the derivative satisfies:

d 1
i ψptq “ ´ ∆ψptq (3.163)
dt 2
where ´ 12 ∆ψptq P L2 .
The items paq ´ pcq motivate the following definition.

26
Definition 3.76 (Strongly continuous one-parameter group.). A family U ptq, t P R, of
unitary operators U ptq P LpHq is called a strongly continuous one-parameter group if:
(i) U : R Ñ LpHq, t ÞÑ U ptq is strongly continuous.
(ii) U pt ` sq “ U ptqU psq for all t, s and moreover U p0q “ 1H .
The items pdq ´ peq motivate the following definition.
Definition 3.77 (Generator of a unitary group.). A densely defined linear operator H with
domain DpHq Ď H is called a generator of a strongly continuous unitary group if:
(i) DpHq “ tψ P H | t ÞÑ U ptqψ is differentiableu.
d
(ii) For all ψ P DpHq it follows that i dt U ptqψ “ U ptqHψ.
Example 3.78 (The free Hamilton operator.). Consider the free Hamilton operator:
1
H0 “ ´ ∆ with DpH0 q “ H 2 pRd q (3.164)
2
is the generator of the unitary group Pf ptq. This can easily be checked from the definition
(3.162), and from the fact that FF ´1 “ F ´1 F “ 1.
Proposition 3.79 (Properties of the generators.). Let H be a generator for U ptq. Then:
(i) DpHq is invariant under U ptq, that is U ptqDpHq “ DpHq for all t P R.
(ii) H commutes with U ptq, that is:

rH, U ptqsψ :“ HU ptqψ ´ U ptqHψ “ 0 for all ψ P DpHq. (3.165)

(iii) H is symmetric, that is:

xHψ, ϕy “ xψ, Hϕy for all ϕ, ψ P DpHq. (3.166)

(iv) U is uniquely determined by H.

(v) H is uniquely determined by U .

Proof. (i) We notice that the map s ÞÑ U psqU ptqψ “ U ps ` tqψ is differentiable if and
only if the map s ÞÑ U psqψ “ U p´tqU ps ` tqψ is differentiable. The derivative of the
first map at s “ 0 is: p´iqU ptqHψ. The derivative of the second map at s “ 0 is:
p´iqU p´tqU ptqHψ. Thus, ψ P DpHq if and only if ψ P U ptqDpHq.
(ii) Let ψ P DpHq. Then:

d d d
U ptqHψ “ U ptqi U psqψ |s“0 “ i U ptqU psqψ |s“0 “ i U psqU ptqψ |s“0 “ HU ptqψ .
ds ds ds
(3.167)
To get the third equality we used that U ptqU psq “ U pt`sq “ U psqU ptq, and that U ptqψ
is in DpHq, by what we proved before.
(iii) By unitarity, xψ, ϕy “ xU ptqψ, U ptqϕy for all ψ, ϕ P H. Therefore,

d d
0 “ xψ, ϕy “ xU ptqψ, U ptqϕy “ x´iHU ptqψ, U ptqϕy ` xU ptqψ, ´iHU ptqϕy
dt dt
“ ixU ptqHψ, U ptqϕy ´ ixU ptqψ, U ptqHϕy “ ixHψ, ϕy ´ ixψ, Hϕy . (3.168)

(iv) Suppose that U

r ptq is generated by H. Then, by symmetry of H:

d ›› r ptqqψ ›2
› d´ ¯
pU ptq ´ U “ 2 }ψ}2 ´ Re xU ptqψ, U
r ptqψy
dt dt ´ ¯
“ ´2Re x´iHU ptqψ, U r ptqψy ` xU ptqψ, ´iH Ur ptqψy
´ ¯
“ ´2Re ixHU ptqψ, Ur ptqψy ´ ixU ptqψ, H U
r ptqψy

“ 0, (3.169)

27
 for all ψ P DpHq (for the second term, we actually use that U r ptqDpHq “ DpHq). Eq.
(3.169) together with U p0q “ Ur p0q “ 1, implies that U ptq æDpHq “ U
r ptq æDpHq for all
t P R. Moreover, from DpHq “ H (recall that, by definition, the generator H is densely
defined in H), we conclude that U “ U r on H.
(v) This is an immediate consequence of the definition of H.

Example 3.80 (Translations as unitary groups on L2 ). (a) Let T ptq : L2 pRq Ñ L2 pRq
with ψ ÞÑ pT ptqψqpxq :“ ψpx ´ tq be the group of translations. It follows that T ptq is
d
a strongly continuous unitary group, generated by D0 “ ´i dx , with domain DpD0 q “
1
H pRq.
(b) The definition of the translations on L2 pr0, 1sq is a bit more delicate. Let 0 ď t ă 1
and θ P r0, 2πq. We define:
"
ψpx ´ tq if x ´ t P r0, 1s
pTθ ptqψqpxq :“ (3.170)
eiθ ψpx ´ t ` 1q if x ´ t ă 0.

This definition allows to define the translation to the right for all t ě 0. Intuitively,
whatever “exits the interval r0, 1s from the right”, “comes back from the left” with a
phase factor eiθ . One can easily check that Tθ ptq is unitary, and that it satisfies the
group composition property. However notice that for θ ‰ θ1 one has Tθ ptq ‰ Tθ1 ptq for
t ‰ 0: different phase factors produce different translation groups. Thus, according to
Proposition 3.79, these groups must have different generators.
However, for t small enough the function pTθ ptqψqpxq does not depend on θ: how can
this be, if the generators of Tθ , Tθ1 ptq differ for different θ, θ1 ? The difference lies in
d
the domains of Dθ , which differ for different values of θ. One has Dθ “ ´i dx , with
domain:
DpDθ q “ tψ P H 1 pr0, 1sq | eiθ ψp1q “ ψp0qu . (3.171)
One can check that DpDθ q is invariant under Tθ ptq, and that Dθ is the generator of
Tθ . Here, H 1 pr0, 1sq is the local Sobolev space, defined as follows:

H 1 pr0, 1sq :“ tψ P L2 pr0, 1sq | such that there exists ϕ P H 1 pRq with ϕ ær0,1s “ ψu .
(3.172)
As we will prove later H 1 pRq Ă CpRq, which means that the pointwise constraint in the
definition of DpDθ q makes sense.
d
Remark 3.81. The operator ´i dx equipped with the maximal definition domain Dmax “
H pr0, 1sq does not generate any unitary group, since H 1 is not invariant under Tθ . The
1

same is true if one chooses a too small domain, for instance Dmin “ tψ P H 1 pr0, 1sq |
ψp0q “ ψp1q “ 0u.
Remark 3.82. For ψ, ϕ P H 1 pr0, 1sq it follows that:
ż1 ż1
d d d
xψ, ´i ϕy “ dx ψpxqp´i ϕpxqq “ ´ipψp1qϕp1q ´ ψp0qϕp0qq ` dx p´i ψpxqqϕpxq
dx 0 dx 0 dx
d
“ ´ipψp1qϕp1q ´ ψp0qϕp0qq ` x´i ψ, ϕy . (3.173)
dx
d
That is, the operator ´i dx on Dmax is not symmetric. As we shall see later, this implies
d d
that ´i dx is not a generator. Instead, ´i dx on Dθ and on Dmin is a symmetric operator,
d
since the boundary term in Eq. (3.173) vanishes. However, ´i dx is a generator only if
defined on Dθ . The symmetry of the operator is a necessary but not sufficient condition to
define the generator of a unitary group.
Before discussing further how to characterize the generator of a unitary group, we con-
clude this section by discussing a regularity result for functions in Sobolev spaces.
Lemma 3.83 (Sobolev.). Let ` P N0 and f P H m pRd q with m ą ` ` d2 . Then, f P C ` pRd q
and B α f P C8 pRd q for all |α| ď `.

28
Proof. We will prove that k α fˆpkq P L1 pRd q for all α P Nd0 with |α| ď `. Then, B α f P C8 pRd q
follows thanks to the Riemann-Lebesgue lemma, Theorem 3.4.
From the definition of H m one has p1`|k|2 qm{2 fˆpkq P L2 pRd q, and therefore for all α P Nd0
with |α| ď `:
ż ż
αˆ
|k f pkq| dk ď p1 ` |k|2 q`{2 |fˆpkq| dk
Rd Rd
ż
`´m
“ p1 ` |k|2 qm{2 |fˆpkq|p1 ` |k|2 q 2 dk
R d
´ż 1 ¯1{2
ď }p1 ` |k|2 qm{2 fˆpkq}L2 pRd q 2 m´`
dk , (3.174)
Rd p1 ` |k| q

where in the last step we used the Cauchy-Schwarz inequality. The last integral is finite if
and only if 2pm ´ `q ą d.

4 Selfadjoint operators
4.1 The Hilbert space adjoint
Let V and W be normed spaces and A P LpV, W q. Then, the dual spaces V 1 and W 1 are
Banach spaces and one can define the adjoint operators A1 : W 1 Ñ V 1 for w1 P W 1 :

pA1 w1 qpvq :“ w1 pAvq for all v P V . (4.1)

Therefore, A1 P LpW 1 , V 1 q and from the Hahn-Banach theorem one also has }A1 } “ }A}. For
Hilbert spaces, it follows that H1 – H, which means that if A P LpHq then A1 P LpH1 q can
also be seen as an operator in LpHq. We shall clarify these points in the following.
Theorem 4.1 (Riesz). Let H be a Hilbert space and T P H1 . Then, there exists a unique
ψT P H such that:
T pϕq “ xψT , ϕyH , for all ϕ P H. (4.2)

Proof. Let T P H1 . We would like to prove that T can be understood as a “projection” over
a vector ψT P H. If so, we can think M :“ KerpT q as being the orthogonal complement of
ψT . Since T is continuous, M is closed. If M “ H then T “ 0 and ψT “ 0 provides the
required vector.
Suppose that M ‰ H. Then, we claim that M K is one dimensional. Let ψ0 , ψ1 P M K zt0u.
Let α :“ TT pψ0q
pψ1 q . We have:

T pψ0 ´ αψ1 q “ T pψ0 q ´ αT pψ1 q “ 0 . (4.3)

That is, ψ0 ´ αψ1 P M X M K “ t0u, which proves that ψ0 “ αψ1 , and hence that M K is
one-dimensional. Now, by Theorem 3.61, for any ϕ P H there is a unique splitting:

xψ0 , ϕy
ϕ “ ϕM ` ϕM K “ ϕM ` ψ0 , (4.4)
}ψ0 }2

T pψ0 q
where the last step follows from the fact that dimpM K q “ 1. Now, let ψT :“ }ψ0 }2 ψ0 . We
have:
xψ0 , ϕy T pψ0 q
T pϕq “ T pϕM ` 2
ψ0 q “ xψ0 , ϕy ” xψT , ϕy , (4.5)
}ψ0 } }ψ0 }2
where the second equality follows from the linearity of T , and from the fact that ϕM P KerpT q.
This proves the claim (4.2). The uniqueness follows from the definition of scalar product.

Riesz Theorem, together with the next proposition, shows that H and H1 are isometrically
isomorphic. In other words, H is selfdual.

29
Proposition 4.2 (Selfduality of Hilbert spaces). Consider the map:

J : H Ñ H1 , ϕ ÞÑ Jϕ :“ xϕ, ¨y . (4.6)

J is a linear map. Moreover, J is an isometry:

}Jϕ}H1 “ }ϕ}H . (4.7)

Remark 4.3. Theorem 4.1 proves that H and H1 are isomorphic. Proposition 4.2 proves
that the isomorphism that associates to an element of H an element of H1 is an isometry.

Proof. The linearity of J immediately follows from its definition. Let us now prove Eq. (4.7).
We have:
|Jϕpψq|
}Jϕ}H1 “ sup
ψPH }ψ}H
|xϕ, ψy|
“ sup
ψPH }ψ}H
“ }ϕ}H , (4.8)

since |xϕ, ψy| ď }ϕ}}ψ} by Cauchy-Schwarz inequality and xϕ, ϕy “ }ϕ}2 .

Definition 4.4 (Hilbert space adjoint). Let A P LpHq. The Hilbert space adjoint operator
A˚ : H Ñ H is defined as:
A˚ “ J ´1 A1 J . (4.9)
Proposition 4.5. For A P LpHq it follows:

xψ, Aϕy “ xA˚ ψ, ϕy for all ψ, ϕ P H. (4.10)

This relation defines A˚ uniquely.

Proof. By the definition of A˚ it follows that:

xψ, Aϕy “ JψpAϕq “ A1 Jψpϕq “ JJ ´1 A1 Jψpϕq “ JA˚ ψpϕq “ xA˚ ψ, ϕy . (4.11)

Also, the map ϕ ÞÑ xψ, Aϕy is continuous and linear. Therefore, by Theorem 4.1 there exists
a unique vector η P H with xψ, Aϕy “ xη, ϕy for all ϕ P H. This proves uniqueness of A˚ .

Theorem 4.6 (Properties of the Hilbert space adjoint). Let A, B P LpHq and λ P C. Then:
(i) pA ` Bq˚ “ A˚ ` B ˚ and pλAq˚ “ λA˚ .
(ii) pABq˚ “ B ˚ A˚ .
(iii) }A˚ } “ }A}.
(iv) A˚˚ “ A.
(v) }AA˚ } “ }A˚ A} “ }A}2 .
(vi) Ker A “ pRan A˚ qK and Ker A˚ “ pRan AqK .

Proof. piq ´ piiiq follows immediately from the definition of Hilbert space adjoint. The
property pivq follows from:

xψ, Aϕy “ xA˚ ψ, ϕy “ xϕ, A˚ ψy “ xA˚˚ ϕ, ψy “ xψ, A˚˚ ϕy for all ψ, ϕ P H. (4.12)

The property pvq follows from:

}Aϕ}2 “ xAϕ, Aϕy “ xϕ, A˚ Aϕy ď }ϕ}2 }A˚ A} , (4.13)

therefore:
}A}2 “ sup }Aϕ}2 ď }A˚ A} ď }A˚ }}A} “ }A}2 . (4.14)
}ϕ}“1

30
To conclude, the property pviq follows from:

ϕ P Ker A ðñ Aϕ “ 0
ðñ xψ, Aϕy “ 0 for all ψ P H (4.15)
˚
ðñ xA ψ, ϕy “ 0 for all ψ P H (4.16)
˚ K
ðñ ϕ P pRan A q . (4.17)

Example 4.7. Let T : `2 Ñ `2 be the right shift, px1 , x2 , . . .q ÞÑ p0, x1 , x2 , . . .q. We have:
8
ÿ 8
ÿ
xx, T yy “ xj yj´1 “ xj`1 yj “: xT ˚ x, yy , (4.18)
j“2 j“1

with T ˚ the left shift operator, px1 , x2 , . . .q ÞÑ px2 , x3 , . . .q. Notice that the rightshift is
isometric, but not surjective and hence not unitary. It follows that T ˚ T “ 1, but T T ˚ ‰ 1.
Proposition 4.8. U P LpHq is unitary if and only if U ˚ “ U ´1 .

Proof. Suppose that U is unitary. Then:

pU ˚ U ψ ´ ψ, ϕq “ pU ψ, U ϕq ´ xψ, ϕy “ 0 for all ψ, ϕ P H. (4.19)

Therefore, U ˚ U “ 1. Since U is surjective, for any ϕ P H there exists ψ P H such that

U ψ “ ϕ. Also, U U ˚ ϕ “ U U ˚ U ψ “ U ψ “ ϕ. This implies that U U ˚ “ 1. That is,
U ˚ “ U ´1 .
Suppose now that U ˚ “ U ´1 . Then, U is surjective, and moreover:

xU ϕ, U ψy “ xU ˚ U ϕ, ψy “ xU ´1 U ϕ, ψy “ xϕ, ψy . (4.20)

This proves that U is unitary.

Definition 4.9 (Bounded selfadjoint operator). A P LpHq is called selfadjoint if A “ A˚ .

Proposition 4.10. Let A P LpHq. Then:

A is selfadjoint ðñ A is symmetric. (4.21)

Proof. The proof immediately follows from Proposition 4.5.

Remark 4.11. In general, for unbounded operators the implication ð does not hold true.
Theorem 4.12 (Bounded generator.). Let H P LpHq with H ˚ “ H. Then, the operator
8
ÿ p´iHtqn
e´iHt “ (4.22)
n“0
n!

defines a unitary group with generator H, with DpHq “ H. Moreover, the map R Ñ LpHq :
t ÞÑ e´iHt is differentiable.

Proof. Exercise.

Definition 4.13 (Unbounded operators.). (a) An unbounded operator is a pair pT, DpT qq
of a subspace DpT q Ă H together with a linear operator T : DpT q Ñ H. If DpT q “ H,
we say that T is densely defined.
(b) An operator pS, DpSqq is called an extension of pT, DpT qq if DpSq Ą DpT q and S æDpT q “
T . We say that T Ă S.
(c) An operator pT, DpT qq is called symmetric if for all ϕ, ψ P DpT q it follows that:

xϕ, T ψyH “ xT ϕ, ψyH . (4.23)

31
Example 4.14. The free Hamilton operator H0 “ ´ 12 ∆ on DpH0 q “ H 2 pRd q is a symmetric
unbounded operator, densely defined.
As we have seen in Example 3.80, the solution of the Schrödinger equation generated
by a symmetric operator H might leave DpHq, if DpHq is chosen too small. We would
like to understand what is exactly missing to imply that a given symmetric operator is the
generator of a unitary group. Let pH0 , DpH0 qq be a symmetric operator, and let pH1 , DpH1 qq
be a symmetric extension. Suppose that the equation:
d
i ψptq “ H1 ψptq , (4.24)
dt
with initial datum ψp0q P DpH0 q has, at least for small times, a solution ψptq that belongs
at least to DpH1 q but not to DpH0 q. The question we ask is where does ψptq go after leaving
DpH0 q. For ϕ P DpH0 q Ă DpH1 q it follows that:

xH1 ψptq, ϕy “ xψptq, H1 ϕy “ xψptq, H0 ϕy . (4.25)

Therefore, if ψptq does not belong to DpH0 q, then it is at least in the domain of the adjoint
operator H0˚ , defined as follows.
Definition 4.15 (The adjoint operator). Let T be a densely defined linear operator on a
Hilbert space H. Then, the domain DpT ˚ q of the adjoint operator T ˚ is defined as:

DpT ˚ q :“ tψ P H | Dη P H s.t. xψ, T ϕy “ xη, ϕy @ϕ P DpT qu . (4.26)

Since DpT q is densely defined, η is uniquely defined and we define, for all ψ P DpT ˚ q:

T ˚ : DpT ˚ q Ñ H , ψ ÞÑ T ˚ ψ :“ η . (4.27)

Remark 4.16. By Theorem 4.1, Definition 4.15 is equivalent to:

DpT ˚ q :“ tψ P H | ϕ ÞÑ xψ, T ϕy is continuous on DpT q.u (4.28)

Proposition 4.17. pT ˚ , DpT ˚ qq is a linear (not necessarily densely defined) operator and:

xψ, T ϕy “ xT ˚ ψ, ϕy for all ψ P DpT ˚ q and ϕ P DpT q. (4.29)

Proof. It immediately follows from Definition 4.15.

Definition 4.18 (Self-adjoint operator). Let pT, DpT qq be a densely defined linear operator.
If DpT ˚ q “ DpT q and T “ T ˚ holds true on DpT q, then we say that pT, DpT qq is a selfadjoint
operator.
Example 4.19. In order to clarify the above definition, let us come back to Example 3.80.
d
(a) Let us consider first Dmin “ ´i dx with:

DpDmin q “ tϕ P H 1 pr0, 1sq | ϕp0q “ ϕp1q “ 0u . (4.30)

For ϕ P DpDmin q we have:

ż1 ¯ ż1 ´
´ d d ¯ d
xψ, Dmin ϕy “ dx ψpxq ´ i ϕpxq “ dx ´ i ψpxq ϕpxq “ x´i ψ, ϕy
0 dx 0 dx dx
“: xη, ϕy (, 4.31)
d
provided dx ψ P L2 pr0, 1sq, which is implied by ψ P H 1 pr0, 1sq. Therefore, one has
˚
DpDmin q “ H 1 pr0, 1sq Ľ DpDmin q which implies that Dmin is not selfadjoint.
d
(b) Let Dθ “ ´i dx with:

DpDθ q “ tϕ P H 1 pr0, 1sq | eiθ ϕp1q “ ϕp0qu . (4.32)

32
 One has, for ϕ P DpDθ q:
ż1 ´ d ¯
xψ, Dθ ϕy “ dx ψpxq ´ i ϕpxq
0 dx
ż1 ´ d ¯ d
“ ipψp0qϕp0q ´ ψp1qϕp1qq ` dx ´ i ψpxq ϕpxq “ x´i ψ, ϕy
0 dx dx
” xη, ϕy , (4.33)

provided that ψ P H 1 pr0, 1sq and that:

ψp0q ϕp1q
ψp0qϕp0q ´ ψp1qϕp1q “ 0 ðñ “ “ e´iθ . (4.34)
ψp1q ϕp0q
d
It follows than that DpDθ˚ q “ DpDθ q and that Dθ˚ “ ´i dx “ Dθ . That is, Dθ is
selfadjoint.
Theorem 4.20 (Generator of a unitary group). A densely defined operator pH, DpHqq is a
generator of a unitary group U ptq “ e´iHt if and only if H is selfadjoint.
Remark 4.21. The Spectral Theorem, to be stated later, will imply that every selfadjoint
operator generates a unitary group. The converse implication, that is that every unitary
group is generated by a selfadjoint operator, is called the Stone Theorem. Both will be proven
later; Theorem 4.20 will then follow as an immediate corollary.
Definition 4.22 (Direct sum of Hilbert spaces). Let H1 and H2 be two Hilbert spaces. Then,
their direct sum is defined as:
H1 ‘ H2 :“ H1 ˆ H2 , (4.35)
equipped with the scalar product

xϕ, ψyH1 ‘H2 :“ xϕ1 , ψ1 yH1 ` xϕ2 , ψ2 yH2 . (4.36)

Remark 4.23. pH1 ‘ H2 , x¨, ¨yH1 ‘H2 q is a Hilbert space.

Definition 4.24 (Graph of an operator, closed operator, closure). (a) The graph of a lin-
ear operator T : DpT q Ñ H is the space:

ΓpT q “ tpϕ, T ϕq P H ‘ H | ϕ P DpT qu Ă H ‘ H . (4.37)

(b) An operator T is called closed if ΓpT q is a closed subspace of H ‘ H.

(c) An operator T is called closable if it admits a closed extension. In this case, the smallest
closed extension T is called the closure of T .
Remark 4.25. It is easy to see that:

ΓpT q “ ΓpT q . (4.38)

Remark 4.26. Therefore, an operator T is closed if for every sequence pϕn q Ă DpT q such
that ϕn Ñ ϕ and T ϕn Ñ η in H, then ϕ P DpT q and T ϕ “ η.
Theorem 4.27 (The adjoint of an operator is always closed.). Let pT, DpT qq be densely
defined. Then, T ˚ is closed.

Proof. We shall show that ΓpT ˚ q is a closed subspace of H ‘ H. To do this, let us first notice
that:

pψ, ηq P ΓpT ˚ q ðñ xψ, T ϕy “ xη, ϕy for all ϕ P DpT q

ðñ xψ, T ϕy ´ xη, ϕy “ 0 for all ϕ P DpT q (4.39)
ðñ xpψ, ηq, p´T ϕ, ϕqyH‘H “ 0 for all ϕ P DpT q. (4.40)

Let us introduce the unitary map:

W : H ‘ H Ñ H ‘ H : pϕ1 , ϕ2 q ÞÑ p´ϕ2 , ϕ1 q . (4.41)

33
Therefore, we rewrite Eq. (4.39) as:

pψ, ηq P ΓpT ˚ q ðñ xpψ, ηq, φyH‘H “ 0 for all φ P W pΓpT qq. (4.42)

That is, ΓpT ˚ q “ pW pΓpT qqqK . Being the orthogonal complement a closed set, it follows
that ΓpT ˚ q is closed and hence that T ˚ is a closed operator.

Proposition 4.28 (Extension of symmetric operators via their adjoint). A densely defined
operator T is symmetric if and only if T Ă T ˚ .

Proof. If T is symmetric, it follows that DpT q Ă DpT ˚ q, because for every ψ P DpT q one
can set η “ T ψ “: T ˚ ψ. Conversely, if T Ă T ˚ , then for every ψ P DpT q Ă DpT ˚ q we have
xψ, T ϕy “ xT ˚ ψ, ϕy “ xT ψ, ϕy for all ϕ P DpT q.

Remark 4.29 (Symmetric operators are closable). Since for symmetric operators one has
T Ă T ˚ and T ˚ is closed, then the symmetric operators are always closable.
Remark 4.30. For general symmetric operators T , the identity T “ T ˚ does not have to be
true. In fact, it is not difficult to see that T is symmetric, while T ˚ may not be.
Proposition 4.31. Let T be densely defined and T Ă S. Then, S ˚ Ă T ˚ .

Proof. With the notation of the proof of Theorem 4.27, one has ΓpS ˚ q “ pW ΓpSqqK . Since
T Ă S, one has ΓpT q Ă ΓpSq, and also W ΓpT q Ă W ΓpSq. Hence:

ΓpS ˚ q “ pW ΓpSqqK Ă pW ΓpT qqK “ ΓpT ˚ q . (4.43)

Proposition 4.32. Let T be densely defined and closable. Then, T ˚ is also densely defined.

Proof. We shall prove that DpT ˚ q is dense in H by showing that DpT ˚ qK “ 0. Let η P
DpT ˚ qK . Then (recall that the orthogonal complement is a closed set):

pη, 0q P ΓpT ˚ qK “ pW ΓpT qqKK “ W ΓpT q . (4.44)

Since W ΓpT q “ tp´T ϕ, ϕq | ϕ P DpT qu, there exists a sequence pϕn q in DpT q with ϕn Ñ 0,
such that ´T ϕn Ñ η. Being T closable, we have that T 0 “ η “ 0.

Proposition 4.33. Let T densely defined and closable. Then:

(a) T ˚˚ “ T .
(b) pT q˚ “ T ˚ “ T ˚˚˚ .

Proof. Being W unitary, it follows that for every subspace M Ă H ‘ H then W pM K q “

pW pM qqK .
(a) We already know that ΓpT ˚ q “ pW ΓpT qqK . Replacing T with T ˚ we have:

ΓpT ˚˚ q “ pW ΓpT ˚ qqK “ pW ppW ΓpT qqK qqK “ W ˝W pΓpT qKK q “ ´ΓpT q “ ´ΓpT q “ ΓpT q .
(4.45)
˚
(b) Thanks to the previous equality it turns out that T “ T ˚˚˚ . Moreover,
˚ K
ΓpT q “ pW ΓpT qqK “ W ΓpT q “ pW ΓpT qqK “ ΓpT ˚ q . (4.46)

34
4.2 Criteria for symmetry, selfadjointness and essential selfadjoint-
ness
Selfadjoint operators play an important role in quantum mechanics, since they are the only
operators that can generate time evolution. Nevertheless, we would like to have criteria that
allows to check whether a given operator is selfadjoint. Before doing this, let us discuss a
simple criterion to determine whether an operator is symmetric.
Lemma 4.34 (Criterium for symmetry). Let T be a linear operator on a complex Hilbert
space H. Then:

xϕ, T ϕy P R for all ϕ P DpT q ðñ T is symmetric. (4.47)

Proof. The fact that T is symmetric immediately implies that xϕ, T ϕy P R, since xϕ, T ϕy “
xT ϕ, ϕy. Let us now prove the converse implication. Suppose that xϕ, T ϕy P R for all
ϕ P DpT q. We would like to show that

xϕ, T ψy “ xT ϕ, ψy for all ψ, ϕ P DpT q. (4.48)

Consider the identity:

xϕ, T ψy “ (4.49)
1
pxϕ ` ψ, T pϕ ` ψqy ´ xϕ ´ ψ, T pϕ ´ ψqy ´ ixϕ ` iψ, T pϕ ` iψqy ` ixϕ ´ iψ, T pϕ ´ iψqyq
4
Let us take the complex conjugate of both sides, recalling that, by assumption, xϕ, T ϕy P R
for all ϕ P DpT q. We have:

xϕ, T ψy “ xT ψ, ϕy “ (4.50)
1
pxϕ ` ψ, T pϕ ` ψqy ´ xϕ ´ ψ, T pϕ ´ ψqy ` ixϕ ` iψ, T pϕ ` iψqy ´ ixϕ ´ iψ, T pϕ ´ iψqyq .
4
Therefore, interchaging ψ with ϕ:

xT ϕ, ψy “ (4.51)
1
pxϕ ` ψ, T pϕ ` ψqy ´ xϕ ´ ψ, T pϕ ´ ψqy ` ixψ ` iϕ, T pψ ` iϕqy ´ ixψ ´ iϕ, T pψ ´ iϕqyq
4
1
“ pxϕ ` ψ, T pϕ ` ψqy ´ xϕ ´ ψ, T pϕ ´ ψqy ` ixiψ ´ ϕ, T piψ ´ ϕqy ´ ixiψ ` ϕ, T piψ ` ϕqyq
4
” xϕ, T ψy

where the last step follows by comparison with Eq. (4.49).

Example 4.35. (i) Let f : R Ñ C measurable. Consider the multiplication operator

pAf ψqpxq “ f pxqψpxq, for all ψ P DpAf q “ tψ P L2 pRq | f ψ P L2 pRqu. We then have
that Af is a symmetric operator if and only if f pxq is real valued.
Let us compute the adjoint of A˚f . To begin, notice that DpAf q is dense in L2 pRq. This
follows from Cc8 pRq Ă DpAf q Ă L2 pRq. The adjoint operator on DpAf q is given by:

pA˚f ψqpxq “ f pxqψpxq . (4.52)

Thus, A˚f “ Af if and only if f is real valued.

(ii) Consider the distributional Laplacian ´∆ on H 2 pRd q. For all ψ P H 2 pRd q Ă L2 pRd q:
ż
xψ, ´∆ψy “ xFψ, F ´ ∆F ´1 Fψy “ dk |ψ̂pkq|2 k 2 P R . (4.53)

Hence, ´∆ is a symmetric operator.

Sometimes, one has to deal with non-closed symmetric operators. Of course, these oper-
ators cannot be self-adjoint (self-adjoint operators are always closed). The relevant question
here is whether the closure of a symmetric operator is selfadjoint.

35
Definition 4.36 (Essentially selfadjoint operator). A symmetric, densely defined operator
is called essentially selfadjoint if its closure is selfadjoint.
Corollary 4.37. A symmetric, densely defined operator T is essentially selfadjoint if and
only if T ˚ is symmetric. In this case T “ T ˚ and T is the unique selfadjoint extension of T .

Proof. Suppose that T ˚ is symmetric. We would like to show that pT q˚ “ T , that is T is

essentially selfadjoint. By Proposition 4.33 pbq, pT q˚ “ T ˚ , hence it is enough to check that
T ˚ “ T . By Theorem 4.27, T ˚ is closed. Moreover, being T symmetric, by Proposition 4.28
T Ă T ˚ . Thus, T Ă T ˚ . To conclude, we would like to show that T ˚ Ă T . We claim that Tutto sbagliato.
T ˚˚˚ Ă T ˚˚ . If so, by Proposition 4.33, we have: T ˚ “ T ˚˚˚ Ă T ˚˚ “ T , which proves Si conclude
that T ˚ Ă T and hence that T ˚ “ T . The claim T ˚˚˚ Ă T ˚˚ follows from the fact that, for semplicemente
notando che
T symmetric, T ˚˚ Ă T ˚ . In fact: by Proposition 4.33 pbq, we have T ˚ “ pT q˚ ; since T is essendo T*
symmetric and densely defined, pT q˚ Ą T , by Proposition 4.28; finally, Proposition 4.33 paq simmetrico, risulta
implies that T “ T ˚˚ . che T**=barT
Now, suppose that T is essentially selfadjoint. Then, T is selfadjoint, and in particular contiene T*.
symmetric. Moreover, T ˚ is symmetric as well, since, by Proposition 4.33, T ˚ “ pT q˚ “ T ,
where the last equality follows from the definition of essential selfadjointness.
To conclude, we have to show that T is the unique selfadjoint extension of T . Suppose
that S is another selfadjoint extension of T . Then, T Ă S implies that T Ă S “ S (since,
by Theorem 4.27, selfadjoint operators are closed). The reverse implication follows from
Proposition 4.31: S “ S ˚ Ă T ˚ “ T , i.e. S “ T .

Definition 4.38. Let pT, DpT qq be a selfadjoint operator. A subspace D0 Ă DpT q, dense in
H, is called core of T if pT, D0 q is essentially selfadjoint, that is if:

T æD0 “ T . (4.54)

Remark 4.39. Equivalently, D0 is a core for pT, DpT qq if and only if D0 is dense in DpT q
with respect to the graph norm:

}ϕ}2ΓpT q :“ }T ϕ}2H ` }ϕ}2H . (4.55)

d
Example 4.40. (a) As we have seen in Example 4.19, the operator p´i dx , Dmin q is sym-
metric but not selfadjoint. Let us check whether it is essentially selfadjoint. To do so,
let us compute the closure of the operator, and check whether the closure is selfadjoint.
d
Being T “ ´i dx symmetric on its domain, we know that T “ T ˚˚ Ă T ˚ . Therefore,
d
for all ψ P DpT ˚ q “ H 1 pr0, 1sq and all ϕ P DpT q, recalling that T Ă T ˚ “ ´i dx :

0 “ xψ, T ϕy ´ xT ˚ ψ, ϕy
d d
“ xψ, ´i y ´ x´i ψ, ϕy “ irϕp0qψp0q ´ ϕp1qψp1qs , (4.56)
dx dx
which implies that ϕp0q “ ϕp1q “ 0 (because ψ P DpT ˚ q “ H 1 pr0, 1sq does not need
to satisfy any boundary condition). We conclude that DpT q Ă tψ P DpT ˚ q | ψp0q “
ψp1q “ 0u ” Dmin . On the other hand, it is easy to check that every ψ P H 1 pr0, 1sq
with ψp0q “ ψp1q “ 0 is also in DpT ˚˚ q “ DpT q. In fact, for any ψ P Dmin and any
ϕ P DpT ˚ q “ H 1 pr0, 1sq, integrating by parts:

d d
xψ, T ˚ ϕy “ xψ, ´i ϕy “ x´i ψ, ϕy “: xη, ϕy , (4.57)
dx dx
d d
with η P L2 pRq given by ´i dx ψ. Therefore, DpT q “ Dmin , and T ψ “ ´i dx ψ for all
ψ P DpT q. Hence, T is a symmetric operator on Dmin , but not selfadjoint; that is
pT, Dmin q is not essentially selfadjoint.
d
(b) We already know that p´i dx , Dθ q is selfadjoint. Hence, it is in particular essentially
selfadjoint.

36
 The distinction between closed symmetric operators and self-adjoint operators may seem
just a technicality, but it is actually very important. The spectral theorem, which plays
a very important role in quantum mechanics, only holds for selfadjoint operators, not for
general closed symmetric operators. Similarly, only selfadjoint operators, and not general
closed symmetric operators, generate a unitary evolution. Unfortunately, while it is easy to
check whether an operator is symmetric, it is much more difficult to decide whether it is
selfadjoint; we need criteria to prove selfadjointness. The basic criterium is stated in the
following theorem.
Theorem 4.41 (Criteria for seldadjointness). Let pH, DpHqq be densely defined and sym-
metric. Then, the following statements are equivalent:
(i) H is selfadjoint.
(ii) H is closed and KerpH ˚ ˘ iq “ t0u.
(iii) RanpH ˘ iq “ H.

Proof. piq ñ piiq. Let H be selfadjoint. Then, H is closed (since H ˚ is closed, Theorem
4.27). Let ϕ˘ P KerpH ˚ ˘ iq. Then, Hϕ˘ “ ¯iϕ˘ . Since the eigenvalues of a symmetric
operators are always real, it follows that ϕ˘ “ 0.
piiq ñ piiiq. This implication will be postponed to the next lemma.
piiiq ñ piq. Being H symmetric, it follows that H Ă H ˚ , by Proposition 4.28. We are
left with showing that H ˚ Ă H. To this end, let ψ P DpH ˚ q. Then, by the assumption
RanpH ˘ iq “ H, there exists ϕ P DpHq such that
pH ˚ ´ iqψ “ pH ´ iqϕ . (4.58)
By H Ă H ˚ , it also follows that:
pH ˚ ´ iqψ “ pH ˚ ´ iqϕ , (4.59)
that is ϕ ´ ψ P Ker pH ˚ ´ iq. As the next lemma will show, this implies that ϕ ´ ψ “ 0,
that is ψ “ ϕ P DpHq, which shows that DpH ˚ q Ă DpHq. Also, by Eq. (4.58), H “ H ˚ on
DpHq, which concludes the proof.

Lemma 4.42. Let pT, DpT qq be densely defined. Then:

(a) For all z P C it follows that Ker pT ˚ ˘ zq “ Ran pT ˘ zqK . In particular:
Ker pT ˚ ˘ zq “ t0u ðñ Ran pT ˘ zq “ H . (4.60)

(b) If T is closed and symmetric, then the sets Ran pT ˘ iq are closed.
Remark 4.43. Let us check how this lemma allows to conclude the proof of Theorem 4.41.
Let us check that piiq ñ piiiq. Eq. (4.60) implies that: Ker pH ˚ ˘ iq “ t0u ñ Ran pH ˘ iq “
H. Finally, being H closed and symmetric, item pbq above implies that Ran H is closed. This
proves the implication piiq ñ piiiq.
To conclude the proof of the implication piiiq ñ piq above, we have to show that piiiq
implies that Ker pH ˚ ´ iq “ t0u. Since Ran pH ˘ iq Ă Ran pH ˘ iq, and Ran pH ˘ iq “ H
by assumption, Eq. (4.60) implies that Ker pH ˚ ´ iq “ t0u, which concludes the proof of
Theorem 4.41.

Proof. (of Lemma 4.42.) To prove paq, notice first that pT ` zq˚ “ T ˚ ` z. Then:
ψ P Ran pT ˘ zqK ðñ xψ, pT ˘ zqϕy “ 0 for all ϕ P DpT q
ðñ ψ P DpT q and pT ˚ ˘ zqψ “ 0
˚

ðñ ψ P Ker pT ˚ ˘ zq . (4.61)
This proves paq. Let us now prove pbq; we start by choosing `i. The proof for ´i is exactly the
same. For symmetric T , it follows that xψ, T ψy “ xT ψ, ψy “ xψ, T ψy, that is xψ, T ψy P R.
Therefore, for any ψ P DpT q:
}pT ` iqψ}2 “ xpT ` iqψ, pT ` iqψy “ }T ψ}2 ` }ψ}2 ´ 2Re ixψ, T ψy
“ }T ψ}2 ` }ψ}2 ě }ψ}2 . (4.62)

37
Therefore, T ` i is injective and pT ` iq´1 : RanpT ` iq Ñ DpT q exists and it is bounded.
Let pψn q be a sequence in Ran pT ` iq such that ψn Ñ ψ. Let ϕn :“ pT ` iq´1 ψn . The
boundedness of pT ` iq´1 implies that ψn is a Cauchy sequence, which therefore converges
to ϕ P H. Being T closed, ΓpT q is a closed set; therefore, the sequence pϕn , ψn q P ΓpT ` iq
converges to pϕ, ψq “ pϕ, pT ` iqϕq P ΓpT ` iq, which shows that ψ P Ran pT ` iq.

Remark 4.44. Suppose that H is nonnegative, that is xψ, Hψy ě 0 for all ψ P DpHq. Then,
it is not difficult to see that the condition for selfadjointness Ran pH ˘ iq “ H in Theorem
4.41 can be replaced by Ran pH ` 1q “ H.
From Theorem 4.41, we also obtain criteria for essential selfadjointness.
Corollary 4.45 (Criteria for essential selfadjointness). Let H be densely defined and sym-
metric. Then, the following statements are equivalent:
(i) H is essentially selfadjoint.
(ii) Ker pH ˚ ˘ iq “ t0u.
(iii) Ran pH ˘ iq “ H.

Proof. Exercise.
d
Example 4.46. (a) Let us give a simple proof of the fact that the operator H “ ´i dx on
1
Dmin “ tψ P H pr0, 1sq | ψp1q “ ψp0q “ 0u is not essentially selfadjoint, based on
Corollary 4.45. The equation:
d
H ˚ ϕ˘ “ ´i ϕ˘ “ ¯iϕ˘ (4.63)
dx
is solved by ϕ˘ “ e˘x , which lies in DpH ˚ q “ H 1 pr0, 1sq. Therefore, KerpH ˚ ˘iq ‰ t0u,
which disproves essential selfadjointess.
(b) For H0 “ ´∆ on Cc8 pRd q it follows that DpH0˚ q “ H 2 pRd q and the equation

H0˚ ϕ˘ “ ´∆ϕ˘ “ ¯iϕ˘ (4.64)

has no solution in H 2 , since ´∆ is a symmetric operator. Therefore, Ker pH0˚ ˘iq “ t0u
and H0 is essentially selfadjoint on Cc8 pRd q.
To conclude this section, let us prove that p´∆, H 2 pRd qq is a selfadjoint operator. We
could use Theorem 4.41, by checking that Γp´∆q is closed. An easier proof will be provided
by the following lemma.
Lemma 4.47. Let U : H1 Ñ H2 be a unitary operator, and pH, DpHqq be a selfadjoint
operator on H1 . Then, pU HU ˚ , U DpHqq is selfadjoint on H2 .

Proof. Exercise.

Let H1 “ H2 “ L2 pRd q, H “ ´∆ and Dp´∆q “ H 2 pRd q. Choose U “ F, the Fourier

transform on L2 pRd q. Then, U HU ˚ “ F ´ ∆F ´1 ” Af with f “ k 2 (multiplication
operator). Being f measurable and real valued, selfadjointness immediately follows from
Example 4.35.

4.3 Selfadjoint extensions

If a symmetric operator is nonnegative, there is a simple way of constructing a selfadjoint
extension via the Friedrichs extension.
Definition 4.48. A densely defined linear operator pT, DpT qq on a Hilbert space H is called
nonnegative, T ě 0, if:

qT pψq :“ xψ, T ψy ě 0 for all ψ P DpT q. (4.65)

It is called positive, T ą 0, if qT pψq ą 0 for all ψ P DpT q.

38
Remark 4.49. The functional qT p¨q is called the quadratic form associated to T .
Remark 4.50. Lemma 4.34 implies that every nonnegative operator is symmetric.
Proposition 4.51. Let pT, DpT qq be a densely defined, linear, nonnegative operator. Given
ψ, ϕ P DpT q, let xϕ, ψyT :“ xϕ, T ψy ` xϕ, ψy. Then, x¨, ¨yT defines a scalar product on DpT q.

Proof. Exercise.
a
Remark 4.52. Therefore, } ¨ }T :“ x¨, ¨yT defines a norm on DpT q. Being T nonnegative,
we have }ψ}2T “ xψ, T ψy ě xψ, ψy “ }ψ}2 .
Definition 4.53. The completion HT of DpT q is the set of equivalence classes of sequences
in DpT q which are Cauchy with respect to the } ¨ }T norm. Two sequences pψn q, pϕn q belong
to the same equivalence class in HT if }ψn ´ ϕn }T Ñ 0.
Remark 4.54. If a sequence is Cauchy with respect to the } ¨ }T norm, it is also Cauchy
with respect to the } ¨ } norm (recall Remark 4.52).
Proposition 4.55. Let rpϕn qnPN s P HT , such that ϕn Ñ ϕ P H. The map rpϕn qnPN s ÞÑ ϕ is
well defined and injective.

Proof. Let us start by proving that the map is well defined. Let pϕn q, pψn q be two sequences
in HT , with }ϕn ´ ψn }T Ñ 0. That is, the two sequences belong to the same equivalence
class, and have the same limit ϕ in H since, by Remark 4.52, }ϕn ´ ψn } Ñ 0. Thus, the map
rpϕn qnPN s ÞÑ ϕ is well defined.
Let us now prove that the map is injective. Suppose that pψn q, pϕn q are two sequences
in HT . Suppose that they converge to the same limit, }ϕn ´ ψn } Ñ 0. Then, we claim that
}ϕn ´ ψn }T Ñ 0, that is they belong to the same equivalence class. This follows from:

}ψn ´ ϕn }2T “ xψn ´ ϕn , ψn ´ ϕn ´ pψm ´ ϕm qyT ` xψn ´ ϕn , ψm ´ ϕm yT (4.66)

ď }ψn ´ ϕn }T }ψn ´ ϕn ´ pψm ´ ϕm q}T ` }pT ` 1qpψn ´ ϕn q}}ψm ´ ϕm }
ď C}ψn ´ ϕn ´ pψm ´ ϕm q}T ` }pT ` 1qpψn ´ ϕn q}}ψm ´ ϕm } ,

where we used that every Cauchy sequence is bounded and that T is a symmetric operator.
For any ε ą 0, by choosing n, m large enough, C}ψn ´ ϕn ´ pψm ´ ϕm q}T ď ε{2. Also, for
any n we can choose m large enough so that }pT ` 1qpψn ´ ϕn q}}ψm ´ ϕm } ď ε{2. Therefore,
}ψn ´ ϕn }2T ď ε, that is }ϕn ´ ψn }T Ñ 0.

Remark 4.56. (i) This proposition is useful because it allows to identify HT with a sub-
space QpT q Ă H, by associating to each equivalence class rpϕn qn s its limit ϕ P H.
Obviously, DpT q Ă QpT q Ă H (every element of DpT q is the limit of a sequence in HT :
just take the constant sequence).
(ii) The scalar product x¨, ¨yT , originally defined on DpT q, can be naturally extended to
QpT q. This is done by using the continuity of the scalar product on H, and the fact
that every element of QpT q is the limit of a sequence in DpT q. (Exercise).
Definition 4.57. The subspace QpT q is called the form domain T . The extension of the
quadratic form qT to QpT q is defined as:

qT pψq :“ xψ, ψyT ´ }ψ}2 for all ψ P QpT q, (4.67)

where x¨, ¨yT is the extension of the scalar product induced by T to QpT q ˆ QpT q.
Remark 4.58. If ψ P DpT q, then qT pψq “ xψ, T ψy.
Theorem 4.59 (Friedrichs extension). Let pT, DpT qq be a linear, symmetric, densely defined
operator, bounded from below by γ: xψ, T ψy ě γ for all ψ P DpT q. Let:

DpTrq :“ DpT ˚ q X QpT ´ γq

Trψ :“ T ˚ψ for all ψ P DpTrq. (4.68)

Then:

39
 (i) Tr is an extension of T , and Tr ě γ.
(ii) Tr is selfadjoint.
(iii) Tr is the only selfadjoint extension of T with DpTrq Ă QpT ´ γq.

Proof. For simplicity, we shall set γ “ 0. If not, replace T by T ´ γ in what follows.

(i) We claim that T Ă Tr. By Proposition 4.28, we have that T Ă T ˚ . Since DpT ˚ q Ą DpT q
and QpT q Ą DpT q, then DpT q Ă DpTrq. Moreover, T “ Tr on DpT q, since T “ T ˚ on
DpT q. This proves that T Ă Tr. Let us now prove that Tr ě 0. Let ψ P DpTrq, and
pψn q Ă DpT q such that ψn Ñ ψ and pψn q is Cauchy in } ¨ }T . Then:

xψ, Trψy “ lim xψn , Trψy . (4.69)

nÑ8

We further write:

xψn , Trψy “ xψn , T ˚ ψy (4.70)

“ xT ψn , ψy
“ xT ψn , ψm y ` xT ψn , ψ ´ ψm y
“ xT ψm , ψm y ` xT pψn ´ ψm q, ψm y ` xT ψn , ψ ´ ψm y “: I ` II ` III .

Clearly, I ě 0. Pick ε ą 0. Consider II. We have, for n, m large enough:

ε
|II| ď }ψn ´ ψm }T }ψm }T ď , (4.71)
2
where we used that pψn q is Cauchy in }¨}T and that every Cauchy sequence is bounded.
Consider now III. We have, for m large enough:
ε
|III| ď }ψn ´ ψm }T }ψ ´ ψm } ď . (4.72)
2

Therefore, xψ, Trψy ě 0.

(ii) Let us now show that Tr is selfadjoint. We shall use Theorem 4.41 piiq. Being Tr ě 0, Tr
is symmetric. Our goal is to show that Ran pT ` 1q “ H (recall Remark 4.44). Recall:

DpTrq :“ tψ P QpT q | Dη P H s.t. xψ, T ϕy “ xη, ϕy for all ϕ P DpT qu , (4.73)

where the vector η is unique (by density of DpT q is H). From the definition x¨, ¨yT , this
is equivalent to:

DpTrq “ tψ P QpT q | Dη P H s.t. xψ, ϕyT “ xη, ϕy for all ϕ P DpT qu . (4.74)

Also, being DpT q dense in QpT q:

DpTrq “ tψ P QpT q | Dη P H s.t. xψ, ϕyT “ xη, ϕy for all ϕ P QpT qu , (4.75)

where now x¨, ¨y is the extension of x¨, ¨yT to QpT q ˆ QpT q (see Remark 4.56). By
definition, Trψ “ T ˚ ψ “ η ´ ψ for all ψ P DpTrq, that is:

pTr ` 1qψ “ η . (4.76)

We will show that for every η P H there exists ψ such that Eq. (4.76) holds true, i.e.
that Ran pTr ` 1q “ H, as claimed. For any η P H, the map QpT q Q ϕ ÞÑ xη, ϕy is a
bounded linear functional on QpT q, with respect to } ¨ } and hence to } ¨ }T . Thus, by
Riesz theorem (Theorem 4.1), there exists ξ P QpT q such that xη, ϕy “ xξ, ϕyT for all
ϕ P QpT q. Comparing this equation with Eq. (4.75), we find that ξ P DpTrq. Also,
by Eq. (4.76), we have pTr ` 1qξ “ η, which shows that Ran pTr ` 1q “ H; therefore,
Theorem 4.41 and Remark 4.44 imply that Tr is selfadjoint.

40
(iii) To conclude, let us prove uniqueness of the selfadjoint extension. Suppose that Tp is
another selfadjoint extension of T with DpTpq Ă QpT q. Let ψ P DpTpq and ϕ P DpT q Ă
DpTpq. Then:

xϕ, pTp ` 1qψy “ xpTp ` 1qϕ, ψy “ xpT ` 1qϕ, ψy “ xψ, pT ` 1qϕy “ xψ, ϕyT “ xϕ, ψyT .
(4.77)
By density of DpT q in QpT q and continuity of the scalar product, taking the complex
conjugate:
xpTp ` 1qψ, ϕy “ xψ, ϕyT for all ψ, ϕ P DpTpq. (4.78)
This implies that ψ P DpTrq, since ψ P QpT q and xψ, ϕyT “ xη, ϕy holds for all ϕ P
DpT q Ă DpTpq, with η “ pTp ` 1qψ. Thus, DpTpq Ă DpTrq. Moreover, by Eq. (4.76),
pTr ` 1qψ “ η: therefore, Trψ “ Tpψ for all ψ P DpTpq. In other words, Tp Ă Tr. By taking
the adjoint, and recalling Proposition 4.31, we also have Tr˚ Ă Tp˚ , but then Tr “ Tp,
since Tr˚ “ Tr and Tp “ Tp˚ .

4.4 From quadratic forms to operators

Theorem 4.59 shows how to construct a selfadjoint extension of a nonnegative operator using
the quadratic form associated with the operator. Later, we will be interested in defining a
selfadjoint operator given a certain quadratic form.
Proposition 4.60. Let Q Ă H, let spϕ, ψq be a sesquilinear form on Q ˆ Q, with quadratic
form qpψq “ spψ, ψq. Suppose that q is real valued and that q is semibounded: there exists
γ P R such that qpψq ě γ}ψ}2 . Let:

xψ, ϕyq :“ spψ, ϕq ` p1 ´ γqxψ, ϕy . (4.79)

Then, x¨, ¨yq is a scalar product on Q.

Proof. Exercise.

Remark 4.61. Recall that a map sp¨, ¨q : Q ˆ Q Ñ C is called a sesquilinear form if it is

linear in the second variable and antilinear in the first variable.
We would like to know whether x¨, ¨yq can be thought as the scalar product generated
by an operator T with quadratic form qT “ q and form domain Q “ QpT q. This is true,
provided we make some assumptions on q.
Definition 4.62. A real valued quadratic form q is called closable if for any sequence pψn q Ă
Q such that }ψn } Ñ 0 and which is Cauchy with respect to } ¨ }q then }ψn }q Ñ 0.
Remark 4.63. This is the analog of the property that allowed us to identify HT with QpT q Ă
H, recall Eq.(4.66).
Let Hq be the completion of Q with respect to } ¨ }q . For closable q, this space can be
identified with a subspace of H, that we shall denote by Qq .
Definition 4.64. The extension of q to Qq is called the closure of q. The quadratic form is
called closed if Qq “ Q.
Theorem 4.65. For every densely defined, closed, semibounded form q : Q Ñ R there is a
unique selfadjoint operator T such that Q “ QpT q and q “ qT . If s is the sesquilinear form
associated with q, then:

DpT q “ tψ P Q | Dη P H s.t. spψ, ϕq “ xη, ϕy for all ϕ P Qu (4.80)

and T ψ “ η.

41
Proof. For simplicity, we assume that q ě 0 (that is, γ “ 0). Since Q is dense, T is well
defined (there cannot be two different η1 , η2 with spψ, ϕq “ xη1 , ϕy “ xη2 , ϕy for all ϕ P Q).
By construction, we have qT pψq “ qpψq for all ψ P DpT q. It follows that T is symmetric and
nonnegative. Proceeding as in the proof of Theorem 4.59, we find that Ran pT ` 1q “ H and
hence T is selfadjoint. Uniqueness is proven again as in the proof of Theorem 4.59.

Definition 4.66. A quadratic form is called bounded if |qpψq| ď C}ψ}2 . The norm of q is
given by:
}q} “ sup |qpψq| . (4.81)
}ψ}“1

Remark 4.67. For bounded quadratic forms, the norm induced by x¨, ¨yq is equivalent to
the standard norm. In this case, we obtain Hq “ H and the operator T associated with
q is bounded, by the Hellinger-Toeplitz theorem (every symmetric operator defined on the
full Hilbert space H is bounded). Together with the polarization identity, it is not difficult to
check that a closed semibounded form q is bounded if and only if the corresponding selfadjoint
operator T is bounded. In this case, }T } “ }q}. In particular, it follows that:

}A} “ sup |xψ, Aψy| (4.82)

}ψ}“1

for all symmetric operators.

5 The spectral theorem

5.1 The spectrum
Definition 5.1 (Resolvent, resolvent set and spectrum). Let pT, DpT qq be a linear operator
on H. We define the resolvent set of T as:

ρpT q :“ tz P C | pT ´ zq : DpT q Ñ H is a bijection with continuous inverse.u (5.1)

For z P ρpT q we define the resolvent of T at z as:

Rz pT q :“ pT ´ zq´1 P LpHq . (5.2)

The spectrum of T is defined as the complement of the resolvent set:

σpT q :“ CzρpT q . (5.3)

Remark 5.2. For closed operators, the continuity requirement in Eq. (5.1) can be dropped.
This is a consequence of the closed graph theorem, stating that a linear map T : X Ñ Y
between two Banach spaces X, Y is continuous if and only if T is closed.
Proposition 5.3. If T is not closed, then ρpT q “ H.

Proof. Suppose that pT ´ zq : DpT q Ñ H is a bijection. Then, pT ´ zq is invertible, and it is

not difficult to see that ΓpT q “ ΓpT ´ zq “ ΓppT ´ zq´1 q (modulo switching the order of the
pairs in the definition of graph). Thus, if ΓpT q is not closed, ΓppT ´ zq´1 q is not closed as
well. This means that there exists pϕn q Ă H such that ϕn Ñ 0 but limnÑ8 pT ´ zq´1 ϕn ‰ 0.
Therefore, pT ´ zq´1 is not continuous. Hence, ρpT q “ H.

Definition 5.4. Let pT, DpT qq be a closed, linear operator. Then, its spectrum σpT q is
partitioned according to the following criteria:
(a) σp pT q :“ tz P C | T ´ z is not injectiveu
is called the point spectrum, and it coincides with the set of eigenvalues of the operator.
(b) σc pT q :“ tz P C | T ´ z is injective, not surjective, with dense rangeu
is called the continuous spectrum.
(c) σr pT q :“ tz P C | T ´ z is injective, not surjective, with no dense rangeu
is called the residual spectrum.

42
Remark 5.5. In conclusion, for closed operators:
σpT q “ σp pT q Y σc pT q Y σr pT q , (5.4)
and if dim H ă 8 then σpT q “ σp pT q is the set of eigenvalues.
Example 5.6. (i) Consider the position operator x̂, with domain:
D̂pxq “ tψ P L2 pRq | xψpxq P L2 pRqu (5.5)
defined via x̂ : ψ ÞÑ xψ. It follows that px̂ ´ zq´1 is the multiplication by the function
px ´ zq´1 , which is bounded for all z P CzR. Therefore, σpx̂q “ R.
The map px̂ ´ λq has a dense range for all λ P R. To see this, for all ψ P L2 we define:
ψ
ϕn :“ χRzrλ´ n1 ,λ` n1 s . (5.6)
x´λ
Then, px ´ λqϕn Ñ ψ in L2 , and hence the range of x ´ λ is dense. Therefore,
σpx̂q “ σc px̂q “ R.
(ii) Let U P LpHq unitary. Then, σpT q “ σpU T U ´1 q. This follows from the fact that T ´ z
is bijective if and only if U pT ´ zqU ´1 “ U T U ´1 ´ z is bijective.
d
Therefore, the momentum operator p̂ “ ´i dx on L2 pRq has real continuous spectrum,
´1
σpp̂q “ σc pp̂q “ R, since p̂ “ F x̂F and the Fourier transform is unitary.
Theorem 5.7 (Properties of the resolvent and of the spectrum). Let pT, DpT qq be a densely
defined operator on a Hilbert space H. Then:
(a) ρpT q is open, that is the spectrum σpT q is closed.
(b) The resolvent map:
ρpT q Ñ LpHq , z ÞÑ Rz pT q :“ pT ´ zq´1 (5.7)
is analytic, that is Rz pT q can be written locally as a pointwise convergent series with
coefficients in LpHq.
(c) If T P LpHq, then |z| ď }T } for all z P σpT q. In particular, the spectrum is compact.
(d) For z, w P ρpT q the first resolvent identity holds true:
Rz pT q ´ Rw pT q “ pz ´ wqRw pT qRz pT q . (5.8)
In particular, the resolvents commute:
Rw pT qRz pT q “ Rz pT qRw pT q . (5.9)

The proof of this theorem is based on the following proposition.

Proposition 5.8 (Neumann series). Let X be a Banach space and T P LpXq with }T } ă 1.
Then, 1 ´ T is continuously invertible and:
8
ÿ
p1 ´ T q´1 “ Tn , (5.10)
n“0

and:
}p1 ´ T q´1 } ď p1 ´ }T }q´1 . (5.11)

Proof. Exercise.

Proof. (of Theorem 5.7.)

(a) Let z0 P ρpT q and |z ´ z0 | ă }Rz0 }´1 . Then,
T ´ z “ T ´ z0 ´ pz ´ z0 q “ pT ´ z0 qp1 ´ pz ´ z0 qRz0 pT qq . (5.12)
Then, the next proposition implies that }pz ´ z0 qRz0 } ă 1, which means that 1 ´
pz ´ z0 qRz0 is continuously invertible, and hence pT ´ zq is continuously invertible.
Therefore, z P ρpT q.

43
 (b) Thanks to the Neumann series :
8
ÿ
Rz “ p1 ´ pz ´ z0 qRz0 q´1 Rz0 “ pz ´ z0 qn Rzn`1
0
, (5.13)
n“0

where the coefficients Rzn`1

0
belong to LpHq.
T
(c) Let |z| ą }T }. Then, 1 ´ z is invertible, and T ´ z as well. Therefore, z P ρpT q.
(d) We have:

Rz pT q ´ Rw pT q “ Rz pT qpT ´ wqRw pT q ´ Rz pT qpT ´ zqRw pT q “ pz ´ wqRz pT qRw pT q .

(5.14)

Theorem 5.9 (Spectrum of a selfadjoint operator). Let pH, DpHqq be a selfadjoint operator.
Then, σpHq Ă R and for all z P CzR:
1
}pH ´ zq´1 } ď . (5.15)
|Impzq|

Proof. Let z “ λ ` iµ, with λ, µ P R and µ ‰ 0. Then, pH ´ λq{µ is selfadjoint on DpHq

and, by Theorem 4.41:
´H ´ λ ¯
Ker ´ i “ Ker pH ´ λ ´ iµq “ t0u (5.16)
µ
and: ´H ´ λ ¯
Ran ´ i “ Ran pH ´ λ ´ iµq “ H . (5.17)
µ
Eq. (5.16) implies that H ´ z : DpHq Ñ H is injective, while Eq. (5.17) implies that it is
surjective, Therefore, H ´ z : DpHq Ñ H is a bijection. Moreover, the inverse is bounded,
since:
}pH ´ λ ´ iµqψ}2 “ }pH ´ λqψ}2 ` }µψ}2 ě µ2 }ψ}2 , (5.18)
which implies that }pH ´ zq´1 } ď 1{|µ|. Therefore, z P ρpHq.

Lemma 5.10. Let T : DpT q Ñ H be a symmetric operator, and suppose that σpT q Ă R.
Then, T is selfadjoint.

Proof. If σpT q Ă R, then T ´ z : DpT q Ñ H is a bijection for all z P CzR. In particular,

RanpT ´ zq “ H; being T symmetric, Theorem 4.41 implies that it is selfadjoint.

Remark 5.11. Therefore, Theorem 5.9 and Lemma 5.10 imply that a symmetric operator
T is selfadjoint if and only if σpT q Ă R.
Lemma 5.12. Let T : DpT q Ñ H be a closed, densely defined operator. Then,

}Rz0 pT q} ě distpz0 , σpT qq´1 (5.19)

for all z0 P C.
Remark 5.13. If T is bounded, we have tz P C | |z| ą }T }u Ă ρpT q.

Proof. The radius of convergence of the Neumann series (5.13) is }Rz0 pT q}´1 . Also, the
series cannot converge if z P σpT q; therefore, }Rz0 pT q}´1 ď distpz0 , σpT qq.

Remark 5.14. For selfadjoint operator, one actually has:

1
}pH ´ zq´1 } “ . (5.20)
distpz, σpHqq

The next theorem provides a useful criterion to decide whether z P σpAq.

44
Theorem 5.15 (Weyl criterion.). Let T : DpT q Ñ H be a closed densely defined operator.
Suppose that there exists a sequence ψn P DpT q with }ψn } “ 1 for all n P N and such
that }pT ´ zqψn } Ñ 0 (such a sequence is known as a Weyl sequence at z). Then, z P σpT q.
Conversely, if z P BρpT q Ă σpT q (recall that σpT q is closed), then there exists a Weyl sequence
at z.

Proof. Let ψn be a Weyl sequence at z. If z P ρpT q, we would have

}ψn } “ }Rz pT qpT ´ zqψn } ď }Rz pT q}}pT ´ zqψn } ď C}pT ´ zqψn } Ñ 0 , (5.21)

thus giving a contradiction. Hence, z P σpT q. On the other hand, suppose that z P BσpT q.
Then, there exists a sequence zn P ρpT q with zn Ñ z. From Theorem 5.12, we have
}Rzn pT q} Ñ 8. Hence, there exists pϕn q Ă H such that }Rzn pT qϕn }{}ϕn } Ñ 8. Let
ψn “ Rzn pT qϕn {}Rzn pT qϕn }. Then, }ψn } “ 1 for all n and:
}ϕn }
}pT ´ zqψn } ď }pT ´ zn qψn } ` |z ´ zn |}ψn } “ ` |z ´ zn | Ñ 0 . (5.22)
}Rzn pT qϕn }
Hence ψn is a Weyl sequence.

Another useful result is the following lemma, that establishes a relation between the
spectrum of T and the one of its inverse T ´1 (which is a densely defined operator on H, if
T is injective and RanT is dense).
Lemma 5.16. Let T be injective and Ran T be dense. Then, T ´1 : Ran T Ñ H is such that:

σpT ´1 qzt0u “ pσpT qzt0uq´1 . (5.23)

Furthermore, T ψ “ λψ if and only if T ´1 ψ “ λ´1 ψ.

Proof. Let z P ρpT qzt0u. Since, for every ϕ P H:

pT ´1 ´ z ´1 qp´zqT Rz pT qϕ “ pT ´ zqRz pT qϕ “ ϕ (5.24)

and for all ψ P DpT ´1 q “ Ran pT q we can write ψ “ T ϕ, we have:

p´zqT Rz pT qpT ´1 ´ z ´1 qψ “ p´zqT Rz pT qpT ´1 ´ z ´1 qT ϕ

“ T Rz pT qpT ´ zqϕ “ T ϕ “ ψ . (5.25)

This shows that T ´1 ´ z ´1 : DpT ´1 q Ñ H is a bijection, with inverse given by p´zqT Rz pT q.

Therefore, z ´1 P ρpT ´1 q and:

Rz´1 pT ´1 q “ ´zT Rz pT q “ ´z ´ z 2 Rz pT q . (5.26)

Inverting the roles of T and T ´1 we have that z ´1 P ρpT ´1 qzt0u implies z P ρpT q. Thus,
recalling that σpT q “ CzρpT q, we have that z P σpT qzt0u if and only if z ´1 P σpT ´1 qzt0u.
To prove the relation between point spectra, notice that if T ψ “ λψ holds, then λψ is in
the range of T , and hence ψ is in the range of T . Therefore, we can apply T ´1 to both sides
of the equation and obtain ψ “ λA´1 ψ, that is λ´1 ψ “ A´1 ψ.

5.2 Postulates of quantum mechanics

5.2.1 Observables
As discussed already in Section 1, quantum mechanical systems are described by vector
in Hilbert spaces. Physically measurable quantities, called observables, correspond to self-
adjoint operators on H. The expected value associated with the self-adjoint operator T in
the state ψ is given by xψ, T ψy.
The vector ψ does not only determine the expectation of T , but also the distribution of
its possible values. Let us consider the simple case in which A has the decomposition:
ÿ
T “ λj Pϕj , (5.27)
j

45
with λj P R the eigenvalues of T , and Pϕj the orthogonal projection onto the normalized
eigenvector ϕj . That is:
Pϕ ψ “ xϕ, ψyϕ . (5.28)
One also uses the notation Pϕ “ |ϕyxϕ|. Then, we have:
ÿ
xψ, T ψy “ λj |xψ, ϕj y|2 . (5.29)
j

Eq. (5.27) is called the spectral representation of the operator T . The spectral theorem
for unbounded operators, that will be discussed later on, implies that the vectors ϕj form
an ONB forřH (this is clear if dim H ă 8, from the spectral theorem for matrices). In
particular, j |xψ, ϕj y|2 “ 1. So far, we are assuming that the spectrum of the observable
T coincides with its point spectrum. As we shall see, the spectral theorem will allow to
generalize the expression (5.27) to cases in which σp pT q ‰ σpT q, introducing the concept of
projection-valued measure.
The interpretation of the identity (5.29) is the following: the eigenvalues λj are the
possible values of the observable T and |xψ, ϕj y|2 is the probability that, if the system is
in the state ψ, a measurement of T gives the value λj . If for example ψ “ ϕj , then a
measurement of T will produce the value λj with probability 1. In general, however, ψ will
be a linear combination of different ϕj ’s. Hence, a measurement of T will give different
values with different probabilities. It makes sense, therefore, to define the variance of T in
the state ψ by setting:

∆Tψ “ xψ, pT ´ xψ, T ψyq2 ψy “ xψ, T 2 ψy ´ xψ, T ψy2 . (5.30)

ř
If, as before, T “ j λj Pϕj , a simple computation shows that:
ÿ
∆Tψ “ pλj ´ xψ, T ψyq2 |xψ, ϕj y|2 . (5.31)
j

An important property of quantum systems is that noncommuting observables cannot be

measured simultaneously with arbitrary precision.
Theorem 5.17 (Heisenberg’s uncertainty principle.). Let A, B be two self-adjoint operators
acting on H. Then, we have:
1
∆Aψ ∆Bψ ě |xψ, rA, Bsψy|2 . (5.32)
4
Proof. For simplicitly, suppose that xψ, Aψy “ xψ, Bψy “ 0 (if not, redefine A, B by sub-
tracting their average values on ψ). Then,

xψ, rA, Bsψy “ xψ, ABψy ´ xψ, BAψy “ 2iImxψ, ABψy . (5.33)

Therefore,
1 1
|xψ, rA, Bsψy| ď 2|xψ, ABψy| ď 2|xAψ, Bψy| ď 2}Aψ}}Bψ} “ 2p∆Aψ q 2 p∆Bψ q 2 . (5.34)

That is:
1
∆Aψ ∆Bψ ě |xψ, rA, Bsψy|2 . (5.35)
4

In particular, choosing A “ x̂i (position operator) and B “ p̂j ” ´i∇j (momentum

operator), assuming that }ψ}2 “ 1, we obtain the relation:

δij
∆xi,ψ ∆pj,ψ ě . (5.36)
4

46
5.2.2 Time evolution
In every quantum system there is an observable that plays a particularly important role, the
Hamiltonian. It generates time evolution via the Schrödinger equation:

iBt ψptq “ Hψptq . (5.37)

If H is a bounded operator, the unique solution of the Schrödinger equation can be written
as
ψptq “ e´iHt ψp0q , (5.38)
where the exponential of H is defined via its Taylor expansion, which converges forřall times
for bounded operators. More generally, if H has the spectral decomposition H “ j λj Pϕj ,
the exponential map is defined as:
ÿ
e´iHt “ e´iλj t Pϕj . (5.39)
j

In particular, the solution of the Schrödinger equation associated to the initial datum ψp0q “
ϕj is simply given by:
ψptq “ e´iλi t ϕi . (5.40)
In this case, the expectation of an arbitrary self-adjoint operator T is given by:

xψptq, T ψptqy “ xϕi , T ϕi y , (5.41)

and does not depend on t. Physically, the vectors ψptq “ e´iλj t ϕj describe the same state
for all times.
The spectral theorem will allow to introduce a spectral decomposition for any self-adjoint
operators, even unbounded ones, and will allow to make sense of the exponential of the
Hamilton operator. This in particular proves existence and uniqueness of the solution of the
Schrödinger equation for general Hamiltonians.

5.3 Projection valued measures

As explained in Section 5.2.1, the spectral representation of a self-adjoint operator T is often
useful in quantum mechanics. It tells us what are the possible outcomes of a measurement of
the observable associated to T , and the probability with which possible values are assumed.
Moreover, as we shall see later, it allows to define a functional calculus, that is to make sense
of functions of operators. An important example is the unitary evolution e´iHT associated
to the Hamiltonian H.
In this section we will discuss how to define functions of self-adjoint operators, satisfying
the properties:

pf ` gqpT q “ f pT q ` gpT q , pf gqpT q “ f pT qgpT q , f pT q “ f pT q˚ . (5.42)

The question is, for which class of functions f do we want to define f pT q. As long as f
is a polynomial, we can define f pT q by simply takinng powers of T . However, for several
purposes, including solving the Schrödinger equation, taking powers of T is not enough. The
next guess would be to consider functions that can be approximated by polynomials, like
analytic functions. This works for bounded operators, but does not work well for unbounded
operators: taking high powers of an unbounded operator typically makes the domain smaller
and smaller.
A better approach consists in defining χΩ pT q for all characteristic functions of Borel sets
Ω Ă R, and then in using the bounded operators χΩ pT q to construct measurable functions of
A. The main advantage of this approach is that, since χ2Ω “ χΩ “ χΩ , the operator χΩ pT q is
an orthogonal projection, for all Borel sets Ω Ă R. On the other hand, we have to show how
to use the orthogonal projections χΩ pT q to define f pT q for a general measurable function f ,
We start by discussing the second step, and we postpone the first.

47
Definition 5.18 (Projection-valued measure). Let H be a Hilbert space. Let BpRq be the
Borel σ-algebra over R. We say that a map P : BpRq Ñ LpHq is a projection valued measure
if:
(i) P pΩq2 “ P pΩq “ P pΩq˚ , for all Ω P BpRq.
(ii) P pRq “ 1H .
Ť
(iii) (Strong σ-additivity) If Ω “ nPN Ωn with Ωn X Ωm “ H for all n ‰ m, then:

ÿ N
ÿ
P pΩn qψ “ lim P pΩn qψ “ P pΩqψ , (5.43)
N Ñ8
nPN n“0

for all ψ P H.
Example 5.19. (a) Let H “ Cd and T P LpCd q be a symmetric d ˆ d matrix. Let
λ1 ă λ2 ă . . . ă λd be the eigenvalues of T , and P1 , . . . , Pd be the corresponding
eigenprojectors (for simplicity, we assume the eigenvalues to be simple). Then, we can
define: ÿ
P pΩq “ Pj . (5.44)
j:λj PΩ

It is easy to check that P : BpRq Ñ LpCd q is a projection-valued measure.

(b) Let H “ L2 pRq and set P pΩq “ χΩ pxq, with χΩ the characteristic function of the set
Ω. Also in this case, P defines a projection valued measure on H.
Remark 5.20. In the definition of projection valued measure we request σ-additivity to hold
in a strong sense (that is, after application to a fixed ψ P H), and not in norm (that is,
taking the supremum over all ψ). This is an important point. Already in the simple example
discussed above, where P pΩq “ χΩ pxq is a multiplication operator over L2 pRq, we do not
have σ-additivity in norm, because the operator norm of multiplication operators is the L8
norm and thus:
"
0 if µpΩ∆Ω1 q = 0
}P pΩq ´ P pΩ1 q} “ }χΩ∆Ω1 }8 “ (5.45)
1 if µpΩ∆Ω1 q ą 0

where Ω∆Ω1 “ pΩzΩ1 q Y pΩ1 zΩq is the symmetric difference of the two sets and µp¨q denotes
the Lebesgue measure on R. Eq. (5.45) implies that σ-additivity does not hold in norm.
Remark 5.21. In Definition 5.18, strong σ-additivity is actually equivalent to weak σ-
additivity. In other words, Eq. (5.43) is equivalent to the condition:
ÿ
xψ, P pΩn qϕy “ xψ, P pΩqϕy , for all ψ, ϕ P H. (5.46)
nPN

This follows from the fact that, if Pn is a sequennce of orthogonal projections and P is an
orthogonal projection with w ´ limnÑ8 Pn “ P then, for any ψ P H:

}Pn ψ}2 “ xPn ψ, Pn ψy “ xψ, Pn ψy Ñ xψ, P ψy “ }P ψ}2 . (5.47)

The weak convergence Pn Ñ P together with }Pn ψ} Ñ }P ψ} implies that Pn ψ Ñ P ψ. Hence,

Pn Ñ P strongly.
Next, we discuss some important properties of projection-valued measures.
Proposition 5.22. The following properties are true.
(i) P pHq “ 0 and P pΩc q “ 1 ´ P pΩq
(ii) P pΩ1 Y Ω2 q “ P pΩ1 q ` P pΩ2 q ´ P pΩ1 X Ω2 q.
(iii) P pΩ1 X Ω2 q “ P pΩ1 qP pΩ2 q
(iv) P pΩ1 q ď P pΩ2 q if Ω1 Ă Ω2 .

Proof. Exercise.

48
Definition 5.23 (Resolution of the identity). For every projection-valued measure P we
define the resolution of the identity p : R Ñ LpHq via ppλq :“ P pp´8; λsq.
Remark 5.24. Then, ppλq is clearly an orthogonal projection for all λ P R. Monotonicity
of P implies that ppλ1 q ď ppλ2 q if λ1 ď λ2 . Also, strong σ-additivity implies that for every
ψ P H and every sequence λn such that λn ď λ for all n P N and such that λn Ñ λ as
n Ñ 8,
lim ppλn qψ “ ppλqψ . (5.48)
nÑ8

That is, s ´ limnÑ´8 ppλn q “ ppλq. Another consequence of strong σ-additivity is that:

s ´ lim ppλq “ 0 , s ´ lim ppλq “ 1 . (5.49)

λÑ´8 λÑ8

As above, strong convergence of an orthogonal projection towards an orthogonal projection is

equivalent to weak convergence.
Definition 5.25 (Measure and distribution associated to a projection-valued measure). For
any fixed ψ P H, we define the finite measure µψ : BpRq Ñ r0; 8q via µψ pΩq “ xψ, P pΩqψy
for all Ω P BpRq. The corresponding distribution function dψ : R Ñ r0; 8q is given by
dψ pλq “ µψ pp´8, λsq.
Remark 5.26. Notice that µψ pΩq ď }ψ}2 . Therefore, dψ pλq ď }ψ}2 . Also, dψ pλq “
}P pp´8; λsqψ}2 “ xψ, ppλqψy.
More generally, starting from the projection valued measure we can also introduce, for
every ψ, ϕ P H, the complex measures µψ,ϕ pΩq “ xψ, P pΩqϕy. They are related to the
positive measures µψ via the polarization identity:
1
µψ,ϕ pΩq “ rµψ`ϕ pΩq ´ µψ´ϕ pΩq ` iµψ´iϕ pΩq ´ iµψ`iϕ pΩqs . (5.50)
4
Also, they satisfy |µψ,ϕ pΩq| ď }P pΩqψ}}P pΩqϕ} ď }ψ}}ϕ}.
Remark 5.27. Every distribution function is associated with a unique measure on the Borel
σ-algebra BpRq. One can also show that every resolution of the identity p : R Ñ LpHq with
the properties listed above is associated with a unique projection valued measure. This follows
from the fact that the resolution of the identity allows us to define distribution functions dψ ,
which in turn can be used to reconstruct the measure µψ . Then, it is easy to check that for
all Ω P BpRq there is a unique orthogonal projection P pΩq such that µψ pΩq “ xψ, P pΩqψy.
This follows from the fact that a linear operator can be reconstructed from the corresponding
quadratic form, via the polarization identity.

5.4 Functional calculus

We shall now use the projection valued measure P : BpRq Ñ LpHq to define a functional
calculus, that is a map from a class of functions to operators. We start with the set of
measurable simple functions.
Definition 5.28 (Simple function.). We say that the function f is a simple measurable
function on R if
n
ÿ
f“ αj χΩj , nPN, αj P C , Ωj P BpRq , (5.51)
j“1

with Ωj X Ω` “ H for all j ‰ `. We denote by SpRq the space of simple measurable functions
on R (or simple functions, for short).
ř
Definition 5.29 (Functional calculus for simple functions.). Let f P S, f “ j αj χΩj . Let
P : BpRq Ñ LpHq be a PVM. We define the functional calculus Φ : S Ñ LpHq as:
n
ÿ
Φpf q :“ αj P pΩj q . (5.52)
j“1

49
Remark 5.30. We shall also define:
ż n
ÿ
f pλqdppλq :“ αj P pΩj q . (5.53)
j“1

Remark 5.31. Notice that for arbitrary ϕ, ψ P H we have:

n
ÿ n
ÿ ż
xϕ, Φpf qψy “ αj xϕ, P pΩj qψy “ αj µϕ,ψ pΩj q “: f pλqdµϕ,ψ pλq . (5.54)
j“1 j“1

The right-hand side is the Lebesgue integral with respect to the complex measure µϕ,ψ (which
is just a linear combination of real measures, according to the polarization identity (5.50)).
Proposition 5.32. The functional calculus Φ : pS, } ¨ }8 q Ñ LpHq is a bounded linear map,
with }Φ} ď 1.

Proof. Linearity immediately follows from the definition. Let us prove boundedness. For
ψ P H, we have:
›ÿn ›
›Φpf qψ ›2
› ›
αj P pΩj qψ ›
› ›
“ ›
j“1
n
ÿ
“ |αj |2 }P pΩj qψ}2
j“1
ÿn
“ |αj |2 µψ pΩj q
j“1
ż
“ |f pλq|2 dµψ pλq . (5.55)

In particular,
}Φpf qψ} ď }f }8 }ψ} , (5.56)
2
where we used that µψ pΩj q ď }ψ} . Therefore:
}Φpf q}
}Φ} :“ ď1. (5.57)
}f }8

Recall the notion of Borel measurable function on R. We say that a function f : R Ñ C

is called Borel measurable if for any Borel set Ω Ă BpCq one has f ´1 pΩq Ă BpRq. We denote
by Mb the space of bounded Borel functions.
Proposition 5.33. The functional calculus Φ : pS, } ¨ }8 q Ñ LpHq extends uniquely to a
bounded linear map Φ : pMb , } ¨ }8 q Ñ LpHq.

Proof. The proof is an application of Theorem 3.66. To begin, recall that any bounded
measurable function can be approximated in L8 norm by simple function. Therefore, S is
dense is Mb with respect to the } ¨ }8 norm. By Theorem 3.66, there is a unique extension
of Φ to a bounded linear map Φ : Mb Ñ LpHq, with norm }Φ} ď 1. This defines Φ for all
f P Mb .

The Lebesgue integral of functions in Mb is defined as the limit of the Lebesgue integral
of simple functions. We have, for any f P Mb :
ż
xψ, Φpf qϕy “ f pλqdµϕ,ψ pλq . (5.58)

We shall also generalize the definition (5.53) by setting:

ż
f pλqdppλq “ Φpf q . (5.59)

50
Theorem 5.34. Let P : BpRq Ñ LpHq be a projection-valued measure. Then, Φ : Mb Ñ
LpHq is a C ˚ -algebra homomorphism with norm one. Moreover, for every sequence fn P Mb
and f P Mb such that fn Ñ f pointwise and with }fn }8 bounded, we have Φpfn q Ñ Φpf q
strongly.
Remark 5.35. The fact that Φ is a C ˚ -algebra homomorphism means that Φ is linear, that
Φp1q “ 1, that Φpf gq “ Φpf qΦpgq for all f, g P Mb and that Φpf q “ Φpf q˚ .

Proof. For simple measurable functions, It is easy to check that Φ is linear, that it satisfies
Φpf gq “ Φpf qΦpgq and that Φpf q “ Φpf q˚ . For general bounded measurable f , these
properties follow by approximation.
If fn Ñ f pointwise and }f }8 ď K, then, by dominated convergence theorem:
ż ż
xϕ, Φpfn qψy “ fn pλqdµϕ,ψ pλq Ñ f pλqdµϕ,ψ pλq “ xϕ, Φpf qψy . (5.60)

This shows that Φpfn qψ Ñ Φpf qψ weakly, as n Ñ 8. Moreover, again by dominated

convergence theorem:
ż ż
}Φpfn qψ} “ |fn pλq| dµψ pλq Ñ |f pλq|2 dµψ pλq “ }Φpf qψ}2 .
2 2
(5.61)

This implies that Φpfn qψ Ñ Φpf qψ, which means that Φpfn q Ñ Φpf q strongly.

Remark 5.36. Since Φ : Mb Ñ LpHq is a C ˚ -homomorphism, we find that:

xΦpgqϕ, Φpf qψy “ xϕ, Φpgq˚ Φpf qψy

ż ż
“ xϕ, Φpgf qψy “ pgf qpλqdµϕ,ψ pλq “ gpλqf pλqdµϕ,ψ , (5.62)

for all f, g P Mb and for all ϕ, ψ P H. Hence, we have:

ż
µΦpgqϕ,Φpf qψ pΩq “ xΦpgqϕ, ΦpχΩ qΦpf qψy “ χΩ pλqgpλqf pλqdµϕ,ψ pλq , (5.63)

which implies that

dµΦpgqϕ,Φpf qψ “ gf dµϕ,ψ . (5.64)
Example 5.37. Let H “ Cd . Let T P Cdˆd matrix. Let λ1 ă λ2 . . . ă λd be the eigen-
values of T , that we assume to be disjoint. Let P1 , . . . , Pd be the corresponding (rank 1)
eigenprojectors. We already defined the projection valued measure associated to T as:
ÿ
PT pΩq “ Pj . (5.65)
j:λj PΩ

Let Mb be the space of bounded measurable functions on σpT q. The functional calculus
associated to this space of functions is the map ΦT : Mb Ñ LpCd q:
d
ÿ
ΦT pf q “ f pλj qPj . (5.66)
j“1

We have, for any ψ P Cd : ÿ

µψ pp´8, λsq “ }Pj ψ}2 (5.67)
j:λj ďλ

or equivalently:
ż d
ÿ
xψ, ΦT pf qψy “ f pλqdµψ pλq “ f pλj q}Pj ψ}2 . (5.68)
j“1

51
 The above discussion allows to define a functional calculus for bounded functions. Next,
we shall introduce a functional calculus for unbounded functions; this is relevant for un-
bounded self-adjoint operators (like the Laplacian).
For f unbounded, we expect Φpf q to be an unbounded operator. Hence, we first have to
define its domain. Recall that. for every bounded measurable function f , we have:
ż
}Φpf qψ}2 “ |f pλq|2 dµψ pλq . (5.69)

Hence, we expect that even for unbounded f , the operator Φpf q can be applied on it, if
f P L2 pR, dµψ q.
Definition 5.38. Given f : R Ñ C, we define the domain of the functional calculus associ-
ated to f as:
Df :“ tψ P H | f P L2 pR, dµψ qu . (5.70)
Proposition 5.39. Df is a linear subspace, dense in H.

Proof. For every Borel set Ω Ă R, we have µαψ pΩq “ |α|2 µψ pΩq and:
µψ`ϕ pΩq ď 2µψ pΩq ` 2µϕ pΩq . (5.71)
This bound implies that f P L2 pR, dµαψ`ϕ q if f P L2 pR, dµψ qXL2 pR, dµϕ q and α P C. Hence
αψ ` ϕ P Df if ψ, ϕ P Df and α P C.
To prove that Df is dense in H we proceed as follows. Let Ωn “ tλ P R | |f pλq| ď nu.
Then, for any ψ P H, we define ψn “ P pΩn qψ. Since dµψn “ χΩn dµψ , we have ψn P Df for
any n. Moreover, since χΩn Ñ 1 pointwise, it follows that ψn Ñ ψ strongly. This proves
that Df is dense.

Proposition 5.40. Let f be a Borel measurable function on R. Let ψ P Df . Let pfn q Ă Mb ,

such that fn Ñ f pointwise and such that }fn }L2 pR,dµψ q is bounded uniformly in n. Then,
the limit limnÑ8 Φpfn qψ “: Φpf qψ exists in H and does not depend on the sequence pfn q. It
defines a linear map Φpf q on Df , such that for all ψ, ϕ P Df :
ż ż
2 2
}Φpf qψ} “ |f pλq| dµψ pλq , xψ, Φpf qϕy “ f pλqdµψ,ϕ pλq . (5.72)

Remark 5.41. The first integral makes sense by definition of Df . The second integral also
makes sense, since by Cauchy-Schwarz L2 pR, dµψ q Ă L1 pR, dµψ q (recall that dµψ is a finite
measure, that is it has finite mass).

Proof. By dominated convergence, we have fn Ñ f in L2 pR, dµψ q. Therefore,

ż
}Φpfn qψ ´ Φpfm qψ} “ }Φpfn ´ fm qψ}2 “ |fn pλq ´ fm pλq|2 dµψ pλq (5.73)

which implies that Φpfn qψ is a Cauchy sequence in H. Therefore, the limit exists and we
set:
Φpf qψ :“ lim Φpfn qψ . (5.74)
nÑ8
It is easy to see that the limit does not depend on the sequence. Therefore, it defines a linear
map Φpf q on Df , and moreover:
ż
}Φpf qψ}2 “ |f pλq|2 dµψ pλq (5.75)

for all ψ P Df . Since µψ is a finite measure, we have that L2 pR, dµψ q Ă L1 pR, dµψ q and
therefore: ż
xψ, Φpf qψy “ f pλqdµψ pλq , (5.76)

or more generally: ż
xψ, Φpf qϕy “ f pλqdµψ,ϕ pλq . (5.77)

52
Remark 5.42. We shall set: ż
Φpf q “: f pλqdppλq . (5.78)

Theorem 5.43. For every Borel measurable function f : R Ñ C, the operator Φpf q : Df Ñ
H is a normal operator (meaning that DpΦpf qq “ DpΦpf q˚ q) and }Φpf qψ} “ }Φpf q˚ ψ} for
all ψ P Df . Moreover, for f, g Borel measurable and α, β P C, we have Φpf q˚ “ Φpf¯q,

αΦpf q ` βΦpgq Ă Φpαf ` βgq , (5.79)

with DpαΦpf q ` βΦpgqq “ D|f |`|g| and:

Φpf qΦpgq Ă Φpf gq (5.80)

where DpΦpf qΦpgqq “ Dg X Df g .

Proof. Fix a Borel measurable function f : R Ñ C. For n P N, let Ωn “ tλ P R | |f pλq| ă nu

and let fn “ f χΩn . Then, fn P Mb and thus Φpfn q˚ “ Φpfn q by Theorem 5.34. For any
ϕ, ψ P Df “ Df “ D|f | , we have:

xϕ, Φpf qψy “ lim xϕ, Φpfn qψy “ lim xΦpfn qϕ, ψy “ xΦpf qϕ, ψy . (5.81)
nÑ8 nÑ8

This implies that DpΦpf q˚ q Ą DpΦpf qq “ DpΦpf qq “ Df , and that, for all ϕ P Df , one has
Φpf q˚ ϕ “ Φpf qϕ. To conclude that Φpf q˚ “ Φpf q we still have to show that DpΦpf q˚ q Ă Df .
To this end, let us fix ϕ P DpΦpf q˚ q. Then, there exists ϕ
r P H such that xϕ, Φpf qψy “ xϕ,
r ψy
for all ψ P DpΦpf qq. By definition of Φpf q we find, for every ξ P H:

Φpf qΦpχΩn qξ “ lim Φpfm qΦpχΩn qξ “ lim Φpf χΩm χΩn qξ “ Φpfn qξ , (5.82)
mÑ8 mÑ8

since χΩm χΩn “ χΩn for all m ě n. Hence, we find:

xΦpf n qϕ, ξy “ xϕ, Φpfn qξy “ xϕ, Φpf qΦpχΩn qξy “ xϕ,
r ΦpχΩn qξy “ xΦpχΩn qϕ,
r ξy (5.83)

for all ξ P H. This implies that Φpfn qϕ “ ΦpχΩn qϕ r and therefore that:
ż
|fn pλq|2 dµϕ pλq “ }Φpfn qϕ}2 “ }ΦpχΩn qϕ}
r 2 Ñ }ϕ}
r2, as n Ñ 8. (5.84)

Since f is the pointwise limit of fn , the monotone convergence theorem implies that f P
L2 pR, dµψ q, with: ż
|f pλq|2 dµϕ pλq “ }ϕ}
r2. (5.85)

Hence ϕ P Df . We obtain Φpf q˚ “ Φpf q, for all Borel measurable functions f over R. This
also implies that:
ż
}Φpf qψ}2 “ |f pλq|2 dµψ pλq “ }Φpf qψ}2 “ }Φpf q˚ ψ}2 (5.86)

for all ψ P Df “ Df . Hence, Φpf q is a normal operator.

Next, we observe that for two Borel measurable functions f, g : R Ñ C and for α, β P C,
we have DpαΦpf q ` βΦpgqq “ DpΦpf qq X DpΦpgqq “ Df X Dg “ D|f |`|g| , because |f | ` |g| P
L2 pR, dµψ q if and only if f P L2 pR, dµψ q and g P L2 pR, dµψ q. Since |αf ` βg| ď Cp|f | ` |g|q,
it is easy to check that D|f |`|g| Ă Dαf `βg . It remains to show that αΦpf qψ ` βΦpgqψ “
Φpαf ` βgqψ for all ψ P D|f |`|g| . To this end, for n P N, set:

Ωn “ tλ P R | |f pλq| ` |gpλq| ď nu , fn “ f χΩn , gn “ gχΩn . (5.87)

For ψ P D|f |`|g| , we have Φpfn qψ Ñ Φpf qψ, Φpgn qψ Ñ Φpgqψ, αΦpfn qψ ` βΦpgn qψ “
Φpαfn ` βgn qψ “ Φppαf ` βgqχΩn qψ Ñ Φpαf ` βgqψ.

53
 Finally, we prove Eq. (5.80). To this end, assume first that g is bounded. Then:

DpΦpf qΦpgqq “ tψ P H | Φpgqψ P Df u “ tψ P H | f P L2 pR, dµΦpgqψ qu

“ tψ P H | f P L2 pR, |g|2 dµψ qu
“ tψ P H | f g P L2 pR, dµψ qu “ DpΦpf gqq ” Df g . (5.88)

Thus, for all ψ P DpΦpf gqq, we have Φpgqψ P DpΦpf qq and (recalling that fn “ χΩn f , with
Ωn “ tλ P R | |f pλq| ď nu):

Φpf qΦpgqψ “ lim Φpfn qΦpgqψ “ lim Φpfn gqψ “ Φpf gqψ . (5.89)
nÑ8 nÑ8

This shows that, if g is bounded, Φpf gq “ Φpf qΦpgq. If now g is not necessarily bounded,
we define Ωn “ tλ P R | |gpλq| ď nu, gn “ gχΩn . Suppose that ψ P Dg X Df g . Then, we
have Φpgn qψ Ñ Φpgqψ. Moreover, ψ P Df gn “ DpΦpf gn qq “ DpΦpf qΦpgn qq implies (from
the case considered above) that Φpf qΦpgn qψ “ Φpf gn qψ Ñ Φpf gqψ. Since Φpf q is closed
(which follows from Φpf q “ Φpf q˚˚ “ Φpf¯q “ Φpf q), this shows that Φpgqψ P Df and that
Φpf qΦpgqψ “ Φpf gqψ.

5.5 Construction of projection valued measures

The discussion of the previous section allowed us to define the functional calculus, given
a family of projection valued measures.
ş In particular, given P : BpRq Ñ LpHq, we can
associate a self-adjoint operator T “ λdppλq with domain:
ż
DpT q “ tψ P H | λ2 dµψ pλq ă 8u . (5.90)

The question we shall consider is this section is: given a self-adjoint operatorş T , is it possible
to find a projection valued measure P such that T can be expressed as T “ λdppλq? If yes,
this provides a spectral representation for the operator T . We shall first answer this question
for the resolvent of T , Rz pT q, and later for T .
Definition 5.44. Let µp¨q : BpRq Ñ R be a Borel measure. For all z P Czsuppµ, we define
the Borel transform F of µ as:
ż
1
F pzq “ dµpλq . (5.91)
λ´z
Remark 5.45. The support of the measure is defined as:

suppµ “ tλ P R | µpOq ą 0 for all open neighbourhoods O of λu . (5.92)

Remark 5.46. Since ż

1
ImF pzq “ Imz dµpλq , (5.93)
|λ ´ z|2
we conclude that z ÞÑ F pzq is a holomorphic function mapping the upper half complex plane
tz P C | Imz ą 0u into itself. Such functions are called Herglotz or Nevanlinna functions.
Theorem 5.47. Every Herglotz function F has the form:
ż ”
1 λ ı
F pzq “ bz ` a ` ´ dµpλq , (5.94)
R λ´z 1 ` λ2
with b ě 0, a P R and µ a Borel measure on R with:
ż
1
dµpλq ă 8 . (5.95)
1 ` λ2
Conversely, for every b ě 0, a P R and for every Borel measure µ satisfying Eq. (5.95), the
function (5.94) is holomorphic on Czsuppµ. It is such that F pzq “ F pzq and:
ż
” 1 ı
ImF pzq “ Imz b ` dµpλq (5.96)
|λ ´ z|2

54
for all z P Czsuppµ. Moreover, if F is a Herglotz function, the triple pa, b, µq satisfying
(5.94) is uniquely determined by
ż
1
a “ ReF piq , b “ ImF piq ´ dµpλq (5.97)
λ2 ` 1
and by the Stieltjes inversion formula:
ż λ2
1“ ‰ 1
µppλ1 , λ2 qq ` µprλ1 , λ2 sq “ lim ImF pλ ` iεqdλ . (5.98)
2 εÑ0 π
`
λ1

Remark 5.48. That is, this theorem allows us to construct a measure starting from a
Herglotz function. Later, we shall take as Herglotz function the quadratic form associated to
Rz pT q, and use this theorem to construct the projection valued measure.

Proof. Let f pzq “ ipi´zq{pi`zq. It is easy to see that f is holomorphic in D “ tz P C | |z| ă

1u and that it takes values in C` “ tz P C | Imz ą 0u. More precisely, f maps the lower
disk D´ “ tz P D | Imz ă 0u into C` zD and it maps the upper disk D` “ tz P D | Imz ą 0u
into itself. Also, the map is invertible, and f ´1 : C` Ñ D is simply f ´1 pzq “ f pzq. Let:

Cpzq :“ ´iF pf pzqq (5.99)

One easily sees that if the map F is Herglotz then C is a Caratheodory function, that is an
holomorphic function on D with Re Cpzq ě 0 for all z P D. Also, we can invert Eq. (5.99)
and obtain:
F pzq “ iCpf pzqq , (5.100)
which shows that if C is a Caratheodory function then F is a Herglotz function. Thus, F is
Herglotz if and only if C is Caratheodory.
We claim now that every Caratheodory function C : D Ñ C has the form:
ż π iϕ
e `z
Cpzq “ ic ` iϕ ´ z
dνpϕq (5.101)
´π e

for c “ Im Cp0q P R and for a finite measure ν, with:

żπ
dνpϕq “ Re Cp0q . (5.102)
´π

To prove this claim, let C : D Ñ C be a Caratheodory function and fix 0 ă r ă 1. Fix z P D

with |z| ă r. By Cauchy theorem, we have the identity:

r2 {ξ ` z ı
ż
1 ”ξ ` z dξ
Cpzq “ ` 2 Cpξq
4πi |ξ|“r ξ ´ z r {ξ ´ z ξ
ż
1 ´ ξ`z ¯ dξ
“ Re Cpξq
2πi |ξ|“r ξ´z ξ
żπ iϕ
1 ´ re ` z ¯
“ Re Cpreiϕ qdϕ . (5.103)
2π ´π reiϕ ´ z

We take the real part:

żπ
Re Cpzq “ P|z|{r pargpzq ´ ϕqdνr pϕq , (5.104)
´π

where we set:
1 ` reiϕ dϕ
Pr pϕq “ Re , dνr pϕq “ Re Cpreiϕ q . (5.105)
1 ´ reiϕ 2π
Notice that dνr is a Borel measure, thanks to Re C ě 0. Setting z “ 0, we obtain:
żπ
dνr pϕq “ Re Cp0q ă 8 , (5.106)
´π

55
uniformly in r ă 1. This implies that there exists a sequence rn Ñ 1 and finite Borel measure
ν on r´π; πs such that, as n Ñ 8:
ż ż
f pϕqdνrn pϕq Ñ f pϕqdνpϕq (5.107)
r´π;πs r´π;πs

for all f P Cpr´π; πsq. In fact, uniform boundedness implies the existence of a subsequence
of measures converging vaguely, that is after testing with compactly supported continuous
functions; this can be proven approximating compactly supported continuous functions with
simple functions, and from the convergence of νrn prλ1 ; λ2 sq on subsequences, for any interval
rλ1 ; λ2 s.
For |z| ă 1, we also have P|z|{r pargz ´ ϕq Ñ P|z| pargz ´ ϕq as r Ñ 1, uniformly in ϕ. We
conclude that:
żπ
Re Cpzq “ lim P|z|{rn pargpzq ´ ϕqdνrn pϕq
nÑ8 ´π
żπ
“ lim P|z| pargpzq ´ ϕqdνrn pϕq
nÑ8 ´π
żπ
“ P|z| pargpzq ´ ϕqdνpϕq
´π
żπ ” eiϕ ` z ı
“ Re dνpϕq . (5.108)
´π eiϕ ´ z

The claim (5.101) now follows because every holomorphic function is determined by its real
part, up to an imaginary constant. In fact, let f pzq be a holomorphic function, such that
Ref “ 0. Then, the Cauchy-Riemann equation implies that Imf “ constant. Therefore,
f pzq “ ic. This proves Eq. (5.101).
Let now F be an arbitrary Herglotz function and C the corresponding Caratheodory
function, defined as in (5.99). Then we can write F pzq “ iCpipi ´ zq{pi ` zqq, or F pzq “
iCppi
r ´ zq{pi ` zqq for the function Cpzq
r “ Cpizq, which is also a Caratheodory function and
therefore admits a representation as in (5.101). Hence:

F pzq “ iCppi
r ´ zq{pi ` zqq
eiϕ ` i´z
ż
i`z
“ ´c ` i iϕ ´ i´z
dνpϕq
r´π;πs e i`z
ipeiϕ ` 1q ` zpeiϕ ´ 1q
ż
“ ´c ` i iϕ ´ 1q ` zpeiϕ ` 1q
dνpϕq
r´π;πs ipe
iϕ
i ` z eeiϕ `1
´1
ż
“ ´c ` i peiϕ ´1q
dνpϕq
r´π;πs i eiϕ `1 ` z
ż8
1 ` λz
“ ´c ` νpt´π, πuqz ` dr
µpλq , (5.109)
´8 λ ´ z

where we changed variables, setting λ “ f pϕq with the function f : p´π; πq Ñ R defined
through f pϕq “ ip1 ´ eiϕ q{p1 ` eiϕ q, we introduced the Borel measure µ r over R such that
rpAq “ νpf ´1 pAqq, and we took into account the weight of ν at ˘π. Setting a “ ´c,
µ
b “ νpt˘πuq and dµpλq “ p1 ` λ2 qdr µpλq, we obtain the desired representation of F .

56
 Suppose now that a Herglotz function F has the form (5.94). Then, we find
ż λ2
1
lim Im F pλ ` iεqdλ
εÑ0` π λ1

1 λ2
ż ż
ε
“ lim dµpxqdλ (5.110)
εÑ0 π λ
1
px ´ λq2 ` ε2
ż ż λ2
1 ε
“ lim dλdµpxq
ε π λ1 px ´ λq2 ` ε2
ż
1
“ lim rarctgppλ2 ´ xq{εq ´ arctgppλ1 ´ xq{εqsdµpxq
εÑ0 π
ż
1
“ rχ pxq ` χpλ1 ,λ2 q pxqsdµpxq
2 rλ1 ,λ2 s
1
“ pµprλ1 ; λ2 sq ` µppλ1 ; λ2 qqq (5.111)
2
where we used the dominated convergence theorem to take the limit ε Ñ 0, since
1 1
rarctgppλ2 ´ xq{εq ´ arctgppλ1 ´ xq{εqs Ñ rχrλ1 ,λ2 s pxq ` χpλ1 ,λ2 q pxqs (5.112)
π 2
pointwise, and
1 C
rarctgppλ2 ´ xq{εq ´ arctgppλ1 ´ xq{εqs ď (5.113)
π 1 ` x2
for an appropriate constant C depending on λ1 , λ2 . The formula for a, b follows evaluating
(5.94) at z “ i.

The next proposition allows to establish a link with the resolvent of selfadjoint operators.
Proposition 5.49. Let pT, DpT qq be a selfadjoint operator. Let FψT pzq be the quadratic form
associated to Rz pT q:
FψT pzq “ xψ, Rz pT qψy . (5.114)
Then, FψT pzq is a Herglotz function, and it can be written as:
ż
1
FψT pzq “ dµpλq , (5.115)
R λ´z

for a unique finite Borel measure µ.

Proof. By the analyticity of z ÞÑ Rz pT q, recall Theorem 5.7, we see that FψT pzq is analytic
in ρpT q, and in particular in C` . Also, FψT pzq maps C` into itself, since:

1
ImFψT pzq “ rxψ, Rz pT qψy ´ xψ, Rz pT qψys
2i
1
“ xψ, pRz pT q ´ Rz pT q˚ qψy
2i
1
“ xψ, pRz pT q ´ Rz pT qqψy
2i
z´z
“ xψ, Rz pT qRz pT qψy (5.116)
2i
where in the last step we used Eq. (5.8). Therefore, ImFψT pzq “ Imz}Rz pT qψ}2 ě 0 for
z P C` . Hence, FψT pzq is a Herglotz function, which means that it can be rewritten as in Eq.
(5.94), for some pa, b, µq. We clam that a “ b “ 0, and that µ is a finite Borel measure. In
fact, by Eq. (5.15) one has }Rz pT q} ď 1{|Imz|, which implies that

|yFψT piyq| ď }ψ}2 , @y P R . (5.117)

57
This implies that FψT pzq has the form:
ż
1
FψT pzq “ dµpλq . (5.118)
R λ´z

The fact that the measure is finite, µpRq ă 8, follows from

y2
ż
yImF piyq “ dµpλq ď }ψ}2 , (5.119)
λ2 ` y 2
and from dominated convergence.

Remark 5.50. Moreover, theorem 5.47 tells us that we can reconstruct the Borel measure
associated to FψT by the inverse Stieltjes transform. In particular, the distribution function
dTψ pλq “ µpp´8; λsq is:
ż λ`δ
1
dTψ pλq “ lim lim Im FψT pt ` iεqdt . (5.120)
δÑ0` εÑ0` π ´8

Since this is a distribution function, it can be used to reconstruct the corresponding Borel
measure µTψ : BpRq Ñ r0; 8q (write the measure of any Borel set via the complement, count-
able union or intersection of sets p´8, λs, λ P R).
We are now left with constructing the projection valued measure. For every Ω P BpRq,
we define the quadratic form:
ż
qΩ pψq “ µψ pΩq “ χΩ pλqdµTψ pλq .
T T
(5.121)

T
Through the polarization identity, we also find a sesquilinear form sTΩ pϕ, ψq such that qΩ pψq “
T
sΩ pψ, ψq. Clearly,
sTΩ pψ, ϕq “ µTψ,ϕ pΩq , (5.122)
with µTψ,ϕ defined from µTψ via the polarization identity. Since 0 ď qT pψq ď }ψ}2 , we have,
by the Cauchy-Schwarz inequality for sesquilinear forms:
1 1
|sTΩ pψ, ϕq| ď qΩ
T T
pψq 2 qΩ pϕq 2 ď }ψ}}ϕ} . (5.123)

By Riesz’ representation theorem, we can write the map ϕ ÞÑ sTΩ pψ, ϕq as sTΩ pψ, ϕq “ xη, ϕy,
for a unique η P H. By the antilinearity of the sesquilinear form, it is not difficult to see that
η “ QT pΩq˚ ψ, for a bounded linear operator QT pΩq with }QT pΩq} ď 1. We then have:

sTΩ pψ, ϕq “ µTψ,ϕ pΩq “ xψ, QT pΩqϕy , T

qΩ pψq “ µTψ pΩq “ xψ, QT pΩqψy . (5.124)

Lemma 5.51. The map QT : BpRq Ñ LpHq is a projection valued measure.

Proof. That is, we have to prove that:

(i) QT pΩq2 “ QT pΩq “ QT pΩq˚ .
(i) QT pRq “ 1H .
(iii) Strong σ-additivity.
We prove first that QT pΩ1 qQT pΩ2 q “ QT pΩ1 X Ω2 q for all Ω1 , Ω2 P BpRq. This implies, in
particular, that for Ω1 “ Ω2 :
QT pΩq2 “ QT pΩq . (5.125)
To this end, we observe that, for all z, z̃ P CzR, by definition of dµTRz̃ pT qϕ,ψ pλq:
ż
1
dµT pλq “ xRz̄ pT qϕ, RzrpT qψy “ xϕ, Rz pT qRzrpT qψy
λ ´ z̃ Rz̄ pT qϕ,ψ
1
“ rxϕ, Rz pT qψy ´ xϕ, RzrpT qψys , (5.126)
z ´ z̃

58
where we used the resolvent identity:
Rz pT q ´ RzrpT q “ pz ´ zrqRz pT qRzrpT q . (5.127)
We conclude that:
ż ż ”
1 1 1 1 ı T
dµT pλq “ ´ dµψ,ϕ pλq
λ ´ zr Rzr pT qϕ,ψ z ´ zr λ´z λ ´ zr
ż
1 1
“ dµT pλq . (5.128)
λ ´ zr λ ´ z ψ,ϕ
Since this identity holds for all zr P CzR, we must have:
1
dµTRz̄ pT qϕ,ψ pλq “ dµT pλq . (5.129)
λ ´ z ψ,ϕ
Therefore,
ż ż
1
dµT “ dµTRz̄ pT qϕ,QpΩqψ pλq
λ ´ z ϕ,QpΩqψ
“ xϕ, Rz pT qQT pΩqψy
ż
“ χΩ pλqdµRz̄ pT qϕ,ψ pλq
ż
1
“ χΩ pλqdµϕ,ψ pλq , (5.130)
λ´z
which means that:
dµϕ,QT pΩqψ pλq “ χΩ pλqdµϕ,ψ . (5.131)
Hence:
ż
T T
xψ, Q pΩ1 qQ pΩ2 qϕy “ dµϕ,ψ pλqχΩ1 pλqχΩ2 pλq
ż
“ χΩ1 XΩ2 pλqdµϕ,ψ pλq

“ xϕ, QT pΩ1 X Ω2 qψy , (5.132)

which means that QT pΩ1 X Ω2 q “ QT pΩ1 qQT pΩ2 q. Also, we claim that QT pΩq˚ “ QT pΩq.
This easily follows from QT pΩq ě 0. Therefore, QT is an orthogonal projection.
Let us now prove that QT pRq “ 1H . Suppose it is false, QT pRqψ ‰ ψ. Then, we write:
ψ “ QT pRqψ ` ϕ (5.133)
with ϕ P Ker QT pRq. Then we have, for any ξ P H:
0 “ dµξ,QT pRqϕ “ χR pλqdµξ,ϕ pλq (5.134)
which implies xξ, Rz pT qϕy “ 0 for all ξ P H and for all z P CzR. Since CzR Ă ρpT q, Rz pT q
is invertible: for any η P H there exists ξ such that Rz̄ pT qξ “ η. Therefore, ϕ “ 0, thus
implying a contradiction: QT pRqψ “ ψ.
Finally, we have to prove the strong σ-additivity. For orthogonal projection, the strong
σ-additivity is equivalent to the weak σ-additivity, since }Qψ} “ xψ, Qψy (hence }Qn ψ} Ñ
}Qψ} is implied by weak convergence). Let pΩn q Ă BpRq, such that Ωn XΩm “ H for n ‰ m.
Let Ω “ Yn Ωn . Therefore, for all ψ P H, for N Ñ 8:
N
ÿ N
ÿ
xψ, QT pΩn qψy “ µψ pΩn q Ñ µψ pΩq “ xψ, QT pΩqψy , (5.135)
n“1 n“1

where the convergence follows from the strong σ-additivity of the measure µψ . By polariza-
tion,
N
ÿ
xψ, QT pΩn qϕy Ñ xψ, QT pΩqϕy (5.136)
n“1
for all ψ, ϕ, which implies strong σ-additivity.

59
 In conclusion, starting from a self-adjoint operator T p, DpT qq we constructed a PVM
P : BpRq Ñ LpHq such that, for all z P CzR:
ż
1
Rz pT q “ dppλq . (5.137)
λ´z
This easily implies the spectral theorem for unbounded self-adjoint operators.
Theorem 5.52. For any self-adjoint operator pT, DpT qq there exists a unique PVM P T such
that: ż
DpT q “ tψ P H | λ2 dµψ pλq ă 8u , (5.138)

and: ż
T “ λdppλq . (5.139)
ş
Proof. Given the PVM constructed before, we know that A “ λdppλq defines an unbounded
self-adjoint operator, with domain DpAq “ Dλ . We claim that A “ T . By construction:
ż
1
Rz pT q “ pT ´ zq´1 “ dppλq , for z P CzR, (5.140)
λ´z

with Rz pT q : H Ñ DpT q. We claim that DpT q Ă Dλ . This follows from the fact that for
any ϕ P DpT q there exists ψ P H such that: Φpλ ´ zqϕ “ Φpλ ´ zqΦp1{pλ ´ zqqψ “ ψ. Also,
Φpλ ´ zq Ą T ´ z, since, for any ψ P Dλ ,

Φpλ ´ zqΦp1{pλ ´ zqqψ “ Φp1{pλ ´ zqqΦpλ ´ zqψ “ ψ . (5.141)

This shows that Φpλ ´ zq “ T ´ z on DpT q, hence Φpλq Ą T . Using that both operators are
self-adjoint, we get Φpλq “ T . To prove uniqueness, notice that the measure µψ is uniquely
determined by Rz pT q via the Stieltjes inversion formula. Uniqueness of P T follows from the
fact that it is uniquely determined by µψ .

Finally, as one could expect, the projection valued measure associated with T is supported
on the spectrum of T .
Theorem 5.53. Let T : DpT q Ñ H be a self-adjoint operator, with projection-valued measure
P T : BpRq Ñ LpHq. Then:

σpT q “ tλ P R | P T ppλ ´ ε, λ ` εqq ‰ 0 , @ε ą 0u . (5.142)

Also,
P T pσpT qq “ 1H , P T pRzσpT qq “ P T pR X ρpT qq “ 0 . (5.143)
Remark 5.54. The condition P T pΩq ‰ 0 has to be understood as there exists ψ P H such
that P T pΩqψ ‰ 0.

Proof. Let λ0 P R, Ωn “ tλ0 ´ 1{n, λ0 ` 1{nu. Suppose that P T pΩn q ‰ 0 for all n P N.
Then, for all n P N we can find ψn P RanP T pΩn q with }ψn } “ 1. We have:

}pT ´ λ0 qψn }2 “ }pT ´ λ0 qP T pΩn qψn }2

ż
1
“ |λ ´ λ0 |2 χΩn pλqdµψn pλq ď 2 . (5.144)
n

Therefore, from the Weyl criterium, λ0 P σpT q. This proves that tλ P R | P T ppλ´ε, λ`εqq ‰
0 , @ε ą 0u Ă σpT q. On the other hand, suppose that there exists ε ą 0 such that
P T ppλ0 ´ ε, λ0 ` εqq “ 0. Define:
1
fε pλq “ χ pλq . (5.145)
λ ´ λ0 Rztλ0 ´ε,λ0 `εu

60
By the properties of the functional calculus,

pT ´ λ0 qΦT pfε q “ ΦT ppλ ´ λ0 qfε q

“ P T pRzpλ0 ´ ε, λ0 ` εqq
“ 1H ´ P T ppλ0 ´ ε, λ0 ` εqq
“ 1H . (5.146)

Analogously, ΦT pfε qpT ´ λ0 qψ “ ψ for all ψ P DpT q. Therefore pT ´ λ0 q is invertible, and

λ0 P ρpT q. This proves Eq. (5.142).
Let us now prove that P T pR X ρpT qq “ 0. For all λ P R X ρpT q, let Iλ Q λ be an open
neighbourhood of λ and P T pIλ q “ 0 (otherwise λ P σpT q, as we just proved). Let us cover
RXρpT q with intervals Iλ , and let tJn unPN be a countable subcovering. Let Ωn “ Jn zYn´1
i“1 Ji .
so that tΩn u is a disjoint covering. By σ-additivity of the projection valued measure,
N
ÿ
P T pR X ρpT qq “ lim P T pΩn q “ 0 . (5.147)
N Ñ8
n“0

Remark 5.55. Therefore, ΦT pf q “ P pσpT qqΦT pf q “ ΦT pχσpT q f q. That is, changing f on

RzσpT q does not change ΦT pf q.

5.6 Unitary equivalence of self-adjoint operators with multiplica-

tion operators
In this section we shall show that self-adjoint operators are unitarily equivalent to multipli-
cation operators. We say that two operators T on H and Tr on H r are unitarily equivalent if
there exists a unitary operator U : H Ñ H such that U T “ T U , with U DpT q “ DpTrq.
r r
Let ψ P H. Let P be a projection valued measure, generating a functional calculus Φ,
and a Borel measure µψ “ xψ, P pΩqψy. Let

Hψ “ tΦpgqψ | g P L2 pR, dµψ qu Ă H . (5.148)

It is not difficult to see that Hψ is closed. Therefore, by Theorem 3.61, we can split the
K
original Hilbert space as H “ Hψ ‘Hψ . In what follows, we shall denote by Pψ the projection
onto Hψ .
Lemma 5.56. The subspace Hψ reduces Φpf q:

Pψ Φpf q Ă Φpf qPψ . (5.149)

Remark 5.57. That is, if ϕ P Df then Pψ ϕ P Df , i.e. Pψ Df Ă Df . Also, for all ϕ P Df ,

Pψ Φpf q “ Φpf qPψ ϕ. We shall also say that Hψ is invariant under Φpf q.

Proof. (Sketch). Suppose f is bounded. Any ϕ P H can be written as ϕ “ Pψ ϕ ` ϕK , with

Pψ ϕ “ Φpgqψ for some g P L2 pR, dµψ q. We claim that Φpf qϕK P Hψ
K
. In fact:

xΦpf qϕK , Φphqψy “ xϕK , Φpf¯hqψy “ 0 , (5.150)

because f¯h P L2 pR, dµψ q since f is bounded. It follows that:

Pψ Φpf qϕ “ Pψ Φpf qΦpgqψ “ Pψ Φpf gqψ

“ Φpf gqψ “ Φpf qΦpgqψ “ Φpf qPψ ϕ . (5.151)

This proves the claim for bounded f . The case of unbounded f follows by an approximation
argument, we omit the details.

61
 Therefore, we can decompose Φpf q “ Φpf q|Hψ ‘ Φpf q|HKψ ; this means that if ϕ “ ϕ1 ` ϕ2
K
with ϕ1 P Hψ and ϕ2 P Hψ , then Φpf qϕ “ Φpf qϕ1 ` Φpf qϕ2 with Φpf qϕ1 P Hψ and
K
Φpf qϕ2 P Hψ .
The domain of Φpf q|Hψ is defined as:

Pψ Df “ Df X Hψ “ tΦpgqψ | g, f g P L2 pR, dµψ qu . (5.152)

On Pψ Df the action of Φpf q is then given by:

Φpf qΦpgqψ “ Φpf gqψ (5.153)

This implies that the operator Φpf q can be interpreted, when considering its action of Hψ ,
as a multiplication operator by f . To be more precise, we can define the map:

Uψ : Hψ Ñ L2 pR, dµψ q , (5.154)

by setting Uψ Φpf qψ “ f . Since }Φpf qψ} “ }f }2 , the map Uψ is unitary. Furthermore, it

follows that:

Uψ DpΦpf q|Hψ q “ Uψ Pψ Df “ Uψ pDf X Hψ q “ tg P L2 pR, dµψ q | f g P L2 pR, dµψ qu (5.155)

and:
Uψ Φpf q|Hψ “ f Uψ , (5.156)
where f also denotes the multiplication operator, pf gqpλq “ f pλqgpλq, with domain Dpf q “
Uψ DpΦpf q|Hψ q.
We say that the vector ψ is cyclic if Hψ “ H. In this case the picture is complete: the
operator Φpf q is unitarily equivalent to the multiplication operator f , acting on its domain
Dpf q “ Uψ Df . In general however Hψ ‰ H, and Eq. (5.156) only shows that the restriction
of Φpf q on the space Hψ (more precisely, on the dense domain Hψ XDf ) is unitarily equivalent
to multiplication with f .
K
What can we say about the restriction of Φpf q on the orthogonal complement Hψ ? Also
K 1
on Hψ we can choose a vector ψ ; the corresponding space Hψ1 will again be invariant with
respect to the action of Φpf q. We can iterate the procedure; tψj ujPJ is called a family of
spectral vectors, if Hψi K ÀHψj for all i ‰ j. We say that a family of spectral vectors if a
spectral basis of H if H “ jPJ Hψj . Such family always exists.
Lemma 5.58. Let H be a separable Hilbert space, and P and projection À valued measure.
Then there exists a, at most countable, spectral basis tψj ujPJ with H “ jPJ Hψj . We can
define a unitary map U “ jPJ Uψj : H Ñ jPJ L2 pR, dµψj q, where Uψj is defined as in Eq.
À À
(5.154), through the identity Uψj Φpf qψj “ f . Then, for any Borel measurable f : R Ñ C:
à
U Df “ Dpf q “ tg P L2 pR, dµψj q | f g P L2 pR, dµψj qu , (5.157)
jPJ

L2 pR, dµψj q.
À
and U Φpf q “ f U , where f acts as a multiplication on each component of jPJ

This last lemma shows, in particular, that any selfadjoint operator is unitarily equivalent
to the multiplication operator λ̂: pλ̂gqpλq “ λgpλq.
Remark 5.59. Notice that the spectral basis is not unique, and its cardinality is not well
defined: there might exists different spectral bases with different cardinality. However, since
we are only considering separable Hilbert spaces, the cardinality of every spectral basis is at
most countable. The minimal cardinality of a spectral basis for a given self-adjoint operator
T , or more generally for a given projection valued measure P , is called the spectral multiplicity
of T (or of P ). We shall say that the spectrum of T is simple if the spectral multiplicity of
T is one (this means that there exists a cyclic vector).

62
5.7 Decomposition of the spectrum
Let us start by reminding some well-known facts about Borel measures. For any Borel
measure µ there exists a decomposition µ “ µac `µs , where µac is absolutely continuous with
respect to the Lebesgue measure (meaning that µac pΩq “ 0 for all Ω P BpRq with Lebesgue
measure |Ω| “ 0) while µs is singular with respect to the Lebesgue measure (meaning that
there exists a set Ω with |Ω| “ 0 and µs pRzΩq “ 0).
The singular measure µs can be further decomposed as µs “ µpp ` µsc , where µpp is
pure point (meaning that the distribution function dpp pλq is a step function on R) and µsc
is singular continuous (meaning that the distribution function is continuous on R).
The measures µac , µsc , µpp are mutually singular: there exist disjoint sets Mac , Mpp , Msc Ă
R such that µac is supported on Mac , µpp is supported on Mpp and µsc is supported on Msc .
Observe that the choice of the sets Mac , Msc , Mpp is not unique: one can always add sets
with zero µ measure. We will choose Mpp as the set of all jump points of the distribution
function µpλq and Msc with Lebesgue measure equal to zero.
At first, suppose that the spectrum of T is simple, and that ψ is a cyclic vector. Let P ”
P T be the projection-valued measure associated to T , and let µ ” µTψ be the corresponding
spectral measure. We then introduce the orthogonal projections:
Pac “ ΦpχMac q , Psc “ ΦpχMsc q , Ppp “ ΦpχMpp q , (5.158)
such that Pac ` Psc ` Ppp “ 1H . By the orthogonality of the projections, we write:
H “ Hac ‘ Hsc ‘ Hpp , (5.159)
with H7 “ P7 H. Recall that the Hilbert space H ” Hψ is unitarily equivalent to L2 pR, dµq,
Uψ Hψ “ L2 pR, dµψ q. Writing Uψ Hψ “ Uψ pPac ` Psc ` Ppp qHψ and using that Uψ P7 Uψ˚ “
χM7 , we get the following orthogonal splitting:

L2 pR, dµq “ L2 pR, dµac q ‘ L2 pR, dµsc q ‘ L2 pR, dµpp q . (5.160)

This means that every function g P L2 pR, dµq can be written as g “ gac ` gsc ` gpp , with
g7 “ g|M7 . Being the sets M7 disjoint, the functions appearing in the splitting are orthogonal.
Notice that, by construction, if ϕ P M7 , then µϕ ” µϕ,7 , with 7 “ ac, sc, pp. In fact, being ψ
cyclic, ϕ “ Φpg7 qψ for some g7 P L2 pR, dµ7 q, and dµψ pλq “ |g7 pλq|2 dµψ pλq, with g7 supported
in M7 .
Also,
T “ pT Pac q ‘ pT Psc q ‘ pT Ppp q . (5.161)
We define the absolutely continuous, singular continuous and pure point spectrum of T as:
σac pT q :“ σpT Pac q , σsc pT q :“ σpT Psc q , σpp pT q :“ σpT Ppp q . (5.162)
Being the subspaces H7 invariant under T , we have P7 T P7 “ T P7 . Hence, T P7 are selfadjoint,
and σ7 pT q are closed subsets of R.
Remark 5.60. One has σp pT q Ă σpp pT q, with σp pT q the set of eigenvalues of T . This also
implies σp pT q Ă σpp pT q. It is possible to prove that σpp pT q “ σp pT q. See next example.
Example 5.61. Let H “ `2 pNq, let T δn “ n1 δn with δn the sequence equal to 1 at the
n-th place and zero otherwise. That is T is a diagonal matrix with elements 1{n. Then,
σp pT q “ t1{n | n P Nu. We claim that

σpT q “ σpp pT q “ σp pT q Y t0u “ σp pT q . (5.163)

We claim that t0u belongs to σpT q. To see this, notice that T is injective, but not surjective:
not every vector in `2 pNq can be written as T ϕ for some ϕ P `2 pNq. Finally, notice that all
points z P C which are not in t n1 | n P NuYt0u are in ρpT q. This simply follows by computing
the resolvent:
n
pT ´ zq´1 δn “ δn , (5.164)
1 ´ zn
and observing that pT ´ zq´1 is bounded for all z P t n1 | n P Nu Y t0u. Therefore, σpT q “
σp pT q Y t0u. At the same time, we know that σpT q “ σpp pT q Y σac pT q Y σsc pT q. Being
σp pT q Ă σpp pT q with σp pT q open and σpp pT q closed, it follows that σp pT q Y t0u “ σpp pT q.

63
Example 5.62. An example of purely absolutely continuous spectrum is obtained taking µ
to be the Lebesgue measure. An example with purely singular continuous spectrum is given
by taking µ to be the Cantor measure.
To conclude, we are left with discussing the case in which the spectrum of T is not
simple. In this case there is no cyclic vector, and we need to introduce a spectral basis. After
introducing such basis, the operator T is unitarily equivalent to a multiplication operator,
after conjugating with the unitary map: U H Ñ ‘j L2 pR, dµψj q. In general, however, it is
difficult to exclude that the splitting (5.159) depend on the choice of the spectral basis. For
this reason, we introduce the following definition of spectral subspaces of H:

Hac :“ tψ P H | µψ is absolutely continuousu

Hsc :“ tψ P H | µψ is singular continuousu
Hpp :“ tψ P H | µψ is pure pointu . (5.165)

Lemma 5.63. We have:

H “ Hac ‘ Hsc ‘ Hpp . (5.166)
As for the simple case, the absolutely continuous, singular continuous and pure point
spectrum are defined as:
σ7 pT q “ σpT |H7 q “ σpT P7 q , (5.167)
where P7 is the projector over H7 .
To conclude, we discuss a simple consequence of the fact that σpp pT q “ σp pT q.
Proposition 5.64. Let pT, DpT qq be a selfadjoint operator. Suppose that ψ P Hpp . Let
pϕj qjPN be the eigenvectors of T , T ϕj “ λj ϕj . Then, there exists pαj q Ă C such that
› N
ÿ ›
lim ›ψ ´ αj ϕj › “ 0 . (5.168)
› ›
N Ñ8
j“1

Proof. The proof immediately follows from the fact that Mpp “ Mp , with Mp “ tλ P R |
λ is an eigenvalue of T u. Therefore, Hp “ φpχMp qH ” Pp H is dense in Hpp .

Remark 5.65. Recall that λj ‰ λk implies that xϕj , ϕk y “ 0. This follows from, for ε ą 0:
1
xϕj , ϕk y “ xϕj , pH ` iε1H qϕk y
λk ` iε
1
“ xpH ´ iε1H qϕj , ϕk y
λk ` iε
λj ´ iε
“ xϕj , ϕk y (5.169)
λk ` iε
which implies that xϕj , ϕk y “ 0 (since λj , λk P R, the ratio in the r.h.s. is ‰ 1).
To conclude, let us discuss a simple example of self-adjoint operator with purely absolutely
continuous spectrum.
Example 5.66. The Laplacian p´∆, H 2 pRd qq is a selfadjoint operator, with:

σp´∆q “ σac p´∆q “ r0, 8q . (5.170)

The selfadjointness of the Laplacian has been proved in Section 4.2, using that it is unitarily
equivalent to multiplication by |k|2 (real-valued measurable function), recall Lemma 4.47.
Also, σp´∆q “ r0, 8q, since σp´∆q “ σpFp´∆qF ´1 q “ σpAk2 q, and σpAk2 q “ r0, 8q, since
k ÞÑ pk 2 ´ zq´1 is a bounded function for all z R Rzr0, 8q.
Let us now prove that σp´∆q “ σac p´∆q. To do this, it is enough to show that the
spectral measure µψ is absolutely continuous with respect to the Lebesgue measure, for all
ψ P H 2 pRd q. Observe that, for all ψ P L2 pRd q, z P ρp´∆q:

|ψ̂pkq|2
ż ż
1
xψ, Rz p´∆qψy “ xψ̂, Rz pAk2 qψ̂y “ 2
dk “ 2
dµ̃ψ prq (5.171)
Rd k ´ z R r ´z

64
where ´ż ¯
d´1
µψ prq “ χr0,8q prqr
dr |ψ̂prωq|2 dω dr . (5.172)
S d´1
After a simple change of variables, we have:
ż
1
xψ, Rz p´∆qψy “ dµψ pλq , (5.173)
R λ´z

with µψ pλq given by:

1 d
´ ż ? ¯
dµψ pλq “ χr0,8q pλqλ 2 ´1 |ψ̂p λωq|2 dω dλ . (5.174)
2 S d´1

This measure is absolutely continuous, since it is of the form dµψ pλq “ f pλqdλ, with f P
L1 pRd , dλq given by:
1 d
´ż ? ¯
fψ pλq “ χr0,8q pλqλ 2 ´1
|ψ̂p λωq|2 dω n´1 . (5.175)
2 S d´1

Absolute continuity of the measure follows from the fact that the integral of an Lp function
over a set with zero Lebesgue measure is zero.

6 Quantum dynamics
In this section we shall apply the spectral theorem to study solutions of the Schrödinger
equation:
iBt ψptq “ Hψptq , (6.1)
where H is a selfadjoint operator, the Hamiltonian, defined on a domain DpHq Ă H.

6.1 Existence and uniqueness of the solution of the Schrödinger

equation
In the next theorem we shall prove that the solution to this equation is given by ψptq “
U ptqψp0q, with U ptq “ expp´iHtq, define via functional calculus:
ż
´iHt
e “ e´iλt dppλq , (6.2)

with P the projection-valued measure associated to pH, DpHqq.

Theorem 6.1. Let pH, DpHqq be a selfadjoint operator and let U ptq “ e´iHt . Then:
(i) U ptq is a strongly continuous one-parameter unitary group.
(ii) The limit:
1
lim rU ptq ´ 1sψ (6.3)
tÑ0 t

exists if and only if ψ P DpHq. In this case:

1
lim rU ptq ´ 1sψ “ ´iHψ . (6.4)
tÑ0t

(iii) We have U ptqDpHq “ DpHq and, on DpHq, rU ptq, Hs “ 0 for all t P R.

Remark 6.2. That is, H is the generator of U ptq, recall Definition 3.77.

Proof. Let us prove piq. The spectral representation of U ptq, Eq. (6.2), together with the
rules of functional calculus, implies that U ptq´1 “ U ptq˚ , and that U pt ` sq “ U ptqU psq for
all t, s P R. To prove that U ptq is strongly continuous, fix ψ P H and consider:
ż
lim }e´iHt ψ ´ e´iHt0 ψ}2 “ lim |e´iλt ´ e´iλ0 t |2 dµψ pλq “ 0 (6.5)
tÑt0 tÑt0

65
 by dominated convergence. This proves piq. Let us now consider piiq. Suppose that ψ P
DpHq. Then, we have:
›1 ›2 ż ˇ ˇ2
´iHt ˇ1
lim › pe ´ 1qψ ` iHψ › “ lim ˇ pe´iλt ´ 1q ` iλtˇ dµψ pλq “ 0 , (6.6)
› › ˇ
tÑ0 t tÑ0 t
again by dominated convergence. Here we used the bound |e´iλt ´ 1| ď |tλ| and the fact
that, since ψ P DpHq: ż
λ2 dµψ pλq ă 8 . (6.7)

One the other hand, define the operator H r Ñ H by:

r : DpHq

r “ tψ : lim i rU ptqψ ´ ψs
DpHq existsu (6.8)
tÑ0 t

and by:
r “ lim i rU ptqψ ´ ψs
Hψ (6.9)
tÑ0 t

for all ψ P DpHq.

r The operator H r is the generator of the one-parameter group U ptq. It
follows from Eq. (6.6) that H Ă H. Moreover, for all ϕ, ψ P DpHq
r r we have:
i p´iq
xϕ, Hψy
r “ lim xϕ, rU ptqψ ´ ψsy “ lim x rU p´tqϕ ´ ϕs, ψy “ xHϕ,
r ψy . (6.10)
tÑ0 t tÑ0 t
We conclude that H r is a symmetric extension of H. However, self-adjoint operators are
maximal: they do not have symmetric extensions2 , which means that H r “ H. This proves
piiq. The proof of piiiq follows from Proposition 3.79 piiq.

Therefore, it follows from Eq. (6.4) that, for ψ0 P DpHq, the vector ψptq P U ptqψ0 with
U ptq “ e´iHt is a solution of the Schrödinger equation with initial datum ψp0q “ ψ0 . In fact:
i i
iBt U ptqψ0 “ lim rU pt ` hq ´ U ptqsψ0 “ lim rU phq ´ 1sU ptqψ0 “ HU ptqψ0 (6.11)
h
hÑ0 hÑ0 h

because U ptqψ0 P DpHq if ψ0 P DpHq. It turns out that U ptqψ0 is the unique solution of the
Schrödinger equation.
Lemma 6.3. Let ψ0 P DpHq and let ψptq be a solution of the Schrödinger equation with
initial datum ψp0q “ ψ0 . Then ψptq “ U ptqψ0 .

Proof. Let ψptq be a solution of the Schrödinger equation. In particular, ψptq is differentiable
and ψptq P DpHq for all t P R (or for all t in the time-interval on which ψptq is a solution).
Let ϕptq “ U p´tqψptq. Then:
i
iBt ϕptq “ lim rU p´t ´ εqψpt ` εq ´ U p´tqψptqs
ε
εÑ0
” ψpt ` εq ´ ψptq U p´εq ´ 1 ı
“ lim iU p´t ´ εq `i U p´tqψptq . (6.12)
εÑ0 ε ε
Since ψ is differentiable and U is strongly continuous, we have, as ε Ñ 0:
ψpt ` εq ´ ψptq
iU p´t ´ εq Ñ iU p´tqψ 1 ptq “ U p´tqHψptq “ HU p´tqψptq . (6.13)
ε
On the other hand, ψptq P DpHq implies that U p´tqψptq P DpHq and therefore that:
U p´εq ´ 1
i U p´tqψptq Ñ ´HU p´tqψptq . (6.14)
ε
We conclude that ϕ1 ptq “ 0 for all t and therefore that ϕptq “ ϕp0q “ ψp0q “ ψ0 . Hence,
ψptq “ U ptqψ0 .

Remark 6.4. Since DpU ptqq “ H, the dynamics can be extended to all initial data ψ0 P H.
However, notice that U ptqψ0 is a solution of the Schrödinger equation if and only if ψ0 P
DpHq.
2 r ˚ Ă H. Also, being H
Suppose that H Ă H.
r Then, by Proposition 4.31 H r symmetric, by Proposition 4.28
˚
H Ă H . That is, H Ă H, hence H “ H.
r r r r

66
6.2 Stone’s theorem
In the previous section we proved that any self-adjoint operator generates a unitary evolution.
Conversely, Stone’s theorem shows that any strongly continuous one-parameter unitary group
U ptq is generated by a selfadjoint operator such that U ptq “ e´iHt .
Theorem 6.5. Let U ptq be a weakly continuous one-parameter unitary group. Let H :
DpHq Ñ H be the generator of U ptq, defined by:
1
DpHq “ tψ P H | lim rU ptqψ ´ ψs existu (6.15)
tÑ0 t

and by:
i
Hψ “ lim rU ptqψ ´ ψs for all ψ P DpHq. (6.16)
tÑ0 t

Then, H is selfadjoint and U ptq “ e´iHt .

Proof. First of all, we notice that the weak continuity of U ptq also implies strong continuity,
since, for any ψ P H and for t Ñ t0 :

}U ptqψ ´ U pt0 qψ}2 “ 2}ψ}2 ´ 2Re xψ, U pt0 ´ tqψy Ñ 0 (6.17)

if U pt0 ´ tq Ñ 1 weakly. Next, we claim that DpHq is dense in H. For any ψ P H and τ ą 0,
we set: żτ
ψτ :“ U ptqψdt . (6.18)
0
´1
This implies that τ ψτ Ñ ψ as τ Ñ 0. In fact, given ε ą 0, by the strong continuity of U ptq
we can find t0 ą 0 such that }U ptqψ ´ ψ} ď ε for all 0 ă t ă t0 . Then, for all 0 ă τ ă t0 we
have:
1 τ
ż
}τ ´1 ψτ ´ ψ} ď }U ptqψ ´ ψ}dt ď ε . (6.19)
τ 0
Since ε ą 0 is arbitrary, this shows that τ ´1 ψτ Ñ ψ. Moreover, we claim that ψτ P DpHq.
In fact, for any τ ą 0, we have:

1 ” t`τ
ż żτ
1 ı
pU ptqψτ ´ ψτ q “ U psqψds ´ U psqψds
t t t 0
1 ” τ `t
ż żt ı
“ U psqψds ´ U psqψds
t τ 0
1 t
ż
“ pU pτ q ´ 1q U psqψds Ñ rU pτ q ´ 1sψ , as t Ñ 0. (6.20)
t 0

This implies that ψτ P DpHq. Hence, for arbitrary ψ P H, we found a sequence τ ´1 ψτ P

DpHq with τ ´1 ψτ Ñ ψ. This proves that DpHq is dense. Next, we show that H is essentially
self-adjoint. From Corollary 4.45, it is enough to check that Ker pH ˚ ˘ iq “ t0u. To this
end, suppose that H ˚ ϕ “ ¯iϕ. Then, proceeding as in the proof of Theorem 6.1 piiiq, for
any ψ P DpHq, we have U ptqψ P DpHq for all t P R and therefore:

d
xϕ, U ptqψy “ xϕ, ´iHU ptqψy “ ´ixH ˚ ϕ, U ptqψy “ ˘xϕ, U ptqψy . (6.21)
dt
Hence,
xϕ, U ptqψy “ e˘t xϕ, ψy . (6.22)
Since the left-hand side is bounded, uniformly in t P R, we must have xϕ, ψy “ 0. Since
ψ P DpHq is arbitrary and DpHq is dense, we conclude that ϕ “ 0. Therefore, H is essentially
selfadjoint, and its closure H is selfadjoint. We can therefore define the one-parameter group
V ptq “ e´iHt . We claim now that V ptq “ U ptq. This would also imply, by Theorem 6.1, that
H “ H (because it would imply that DpHq “ DpHq) and therefore it would conclude the
proof of the theorem.

67
 To show that indeed V ptq “ U ptq, we pick ψ P DpHq and we set ψptq “ U ptqψ ´ V ptqψ.
Then, we compute:
ψpt ` sq ´ ψptq pU psq ´ 1q V psq ´ 1
lim “ lim U ptqψ ´ lim V ptqψ
sÑ0 s sÑ0 s sÑ0 s
“ iHU ptqψ ´ iHV ptqψ “ iHψptq , (6.23)
where we used that U ptqψ P DpHq if ψ P DpHq, V ptqψ P DpHq if ψ P DpHq Ă DpHq, and
that HU ptqψ “ HU ptqψ for ψ P DpHq (because H is an extension of H). We obtain:
d d
}ψptq}2 “ xψptq, ψptqy “ 2Re xψptq, iHψptqy “ 0 (6.24)
dt dt
since xψptq, Hψptqy P R (which follows from the fact that H is selfadjoint). With ψp0q “ 0,
it follows that ψptq “ 0 for all t and therefore that U ptqψ “ V ptqψ for all ψ P DpHq. Since
DpHq is dense in H and U ptq, V ptq are unitary (in particular, bounded), this also implies
that U ptq “ V ptq on H.

6.3 The RAGE theorem

There is an interesting relation between the spectrum of a selfadjoint operator H and the
properties of the quantum dynamics U ptq “ e´iHt . This relation is summarized in a theorem
due to Ruelle-Amrein-Georgescu-Enss. The goal here is to understand, based on the spectral
properties of H, whether a quantum system whose evolution is generated by H remains
confined in a bounded region for all times or whether instead it moves to infinity as t Ñ 8.
A first simple observation is as follows. Let H be a selfadjoint operator, and Hac , Hsc ,
Hpp the corresponding spectral subspaces, so that H “ Hac ‘ Hsc ‘ Hpp .
If ψ P Hac , then the spectral measure µψ is absolutely continuous with respect to the
Lebesgue measure. This also implies that µϕ,ψ is absolutely continuous with respect to
Lebesgue, for all ϕ P H, since
|µϕ,ψ pΩq| “ |xϕ, P pΩqψy| ď }xϕ, P pΩqy}1{2 |xψ, P pΩqψy|1{2 “ µϕ pΩq1{2 µψ pΩq1{2 . (6.25)
Therefore, setting U ptq “ e´iHt we find:
ż
xϕ, U ptqψy “ e´iλt dµϕ,ψ pλq Ñ 0 as t Ñ 8, (6.26)

by the Riemann-Lebesgue lemma. This is because any Borel measure µ which is absolutely
continuous with respect to Lebesgue can be written as dµpλq “ f pλqdλ, with f P L1 pR, dλq
and dλ the volume measure. In fact, by Theorem 3.4:
ż
xϕ, U ptqψy “ e´iλt fϕ,ψ pλqdλ ” fˆϕ,ψ ptq Ñ 0 as t Ñ 8. (6.27)

This means that, if ψ P Hac , the time evolved state U ptqψ becomes orthogonal to any fixed
ϕ P H, as t Ñ 8. This of course cannot be true for all ψ P H. In particular, if ψ is an
eigenvector of H, that is if Hψ “ Eψ, one has:
|xϕ, U ptqψy| “ |xϕ, ψy| , for all t P R. (6.28)
A more exhaustive understanding of the asymptotic behavior of xϕ, U ptqψy in the limit of
large t is provided by the following theorem.
Theorem 6.6. [Wiener] Let µ a finite complex Borel measure on R and:
ż
µ̂ptq :“ e´itλ dµpλq . (6.29)
R

Then,
żT
1 ÿ
lim |µ̂ptq|2 dt “ |µptλuq|2 , (6.30)
T Ñ8 T 0 λPR

where the sum on the r.h.s. is finite (because µ is a finite measure).

68
Remark 6.7. Recall that any Borel measure has can be written as µ “ µac `µ ř sc `µpp . Also,
2
since µac , µsc have continuous distribution, µptλuq “ µpp ptλuq. Therefore, λPR |µptλuq| “
2
ř
λPR |µpp ptλuq| . The sum is over the support of Mpp of theŤ pure point measure µpp , which
is a countable set. This follows from the fact that Mpp “ nPN Mn with Mn “ tλ P R |
µptλuq ą 1{nu ” tλ P R | µpp ptλuq ą 1{nu. Each set Mn is countable and finite: otherwise,
µpMn q “ 8, which is impossible since µ is finite. Therefore, Mpp is the countable union of
finite sets, and hence it is countable.

Proof. We apply Fubini’s theorem to write:

1 T 1 T
ż ż ż ż
µptq|2 dt “
|p e´ipx´yqt dµpxqdµpyqdt
T 0 T 0 R R
ż ż ” żT
1 ı
“ e´ipx´yqt dt dµpxqdµpyq . (6.31)
R R T 0

Since
ˇ1 żT ˇ
e´ipx´yq dtˇ ď 1 (6.32)
ˇ ˇ
ˇ
T 0
and, as T Ñ 8:
żT "
1 0 if x ‰ y
e´ipx´yqt dt Ñ (6.33)
T 0 1 if x “ y.
Therefore, by dominated convergence:

1 T
ż ż ż ż ÿ
µptq|2 dt Ñ
|p χt0u px ´ yqdµpxqdµpyq “ µptyuqdµpyq “ |µptyuq|2 . (6.34)
T 0 R R R yPR

Let us now apply this theorem to study the quantity |xϕ, U ptqψy|, describing the proba-
bility of finding the evolved state in the state ϕ at time t. If ψ P Hac ‘ Hsc and ϕ P H is
arbitrary, the measure µϕ,ψ has not atoms, i.e. it is such that µϕ,ψ ptλuq “ 0, for all λ P R.
Therefore, by Theorem 6.6:

1 T
ż
lim |xϕ, e´iHt ψy|2 dt “ 0 . (6.35)
T Ñ8 T 0

Hence the probability of finding the evolved state in ϕ tends to zero, but only in an averaged
sense.
Notice that |xϕ, U ptqψy|2 “ }Pϕ U ptqψ}2 , with Pϕ the orthogonal projection onto ϕ. We
can extend Eq. (6.35) to a more general class of operators, called compact operators. Com-
pact operators are the natural generalization of finite-rank operators, that is operators that
can be written as finite linear combination of orthogonal projectors. In the following, we
shall denote by B1 p0q the unit ball in H, that is:

B1 p0q “ tψ P H | }ψ} ď 1u . (6.36)

Definition 6.8. An operator K P LpHq is called compact if KB1 p0q Ă H is pre-compact in

H, that is if KB1 p0q is compact.
Remark 6.9. (i) Equivalently, an operator K P LpHq is compact if and only if for any
bounded sequence ψn P H, Kψn has a convergent subsequence.
(ii) The space of all compact operator KpHq is a closed linear subspace of LpHq. Also, K ˚
is compact if K is compact, and KA, AK are compact if K P KpHq and A P LpHq.
Furthermore, compact operators can be approximated in norm by sequences of finite
rank operators.
Definition 6.10. An operator K : DpKq Ñ H is called relatively compact with respect to
the self-adjoint operator H if there exists z P ρpHq such that KRz pHq “ Kpz ´ Hq´1 is
compact.

69
Remark 6.11. (i) Using the first resolvent identity, Rz pHq´Rz0 pHq “ pz´z0 qRz pHqRz0 pHq,
one can check that if KRz pHq is compact for one z P ρpHq, then it is compact for all
z P ρpHq.
(ii) If K is relatively compact with respect to H, then DpHq Ă DpKq, because every ψ P
DpHq can be written as ψ “ RA pzqϕ for a ϕ P H.
The results (6.27), (6.35) can now be extended as follows.
Theorem 6.12. Let H be a selfadjoint operator. Let K be relatively compact with respect
to H. Then, for all ψ P DpHq:

1 T
ż
lim }Ke´iHt Pc pHqψ}2 dt “ 0 , (6.37)
T Ñ8 T 0

where Pc pHq “ Pac pHq ` Psc pHq is the orthogonal projection onto Hac ‘ Hsc . Also, for all
ψ P DpHq:
lim }Ke´itH Pac pHqψ}2 “ 0 . (6.38)
tÑ8

If we also assume that K is bounded, then Eqs. (6.37), (6.38) hold true for any ψ P H.

Proof. To prove Eqs. (6.37), (6.38), we can assume that ψ P Hc and, respectively, that
ψ P Hac , and drop the orthogonal projections. If K is a rank-one projector,
řn the claims follow
from Eqs. (6.27), (6.35). If K is a finite-rank operator, Kψ “ j“1 αj xψj , ψyϕj for two
orthonormal families tϕ1 , . . . , ϕn u, tψ1 , . . . , ψn u then:
n
ÿ
}Ke´iHt ψ}2 “ |xψj , e´iHt ψy|2 , (6.39)
j“1

and the problem reduces to the rank-1 case. If K is compact, we can find a sequence of
finite-rank operators Kn with }K ´ Kn } ď 1{n. Then:

}Ke´iHt ψ}2 ď 2}Kn e´iHt ψ}2 ` 2n´2 }ψ}2 , (6.40)

and the problem reduces to the finite-rank case (by choosing first n large enough, and then
T or t large enough). Finally, it K is relatively compact with respect to H and ψ P DpHq,
we write ψ “ pH ´ zq´1 ξ for a ξ P H (notice that, if ψ P Hc or ψ P Hac , then also ξ P Hc or,
respectively, ξ P Hac ). Thus, it is enough to apply the result for compact operators to the
operator KpH ´ zq´1 , because the operator pH ´ zq´1 commutes with e´iHt .

Example 6.13. A simple application of these results is obtained by taking H “ ´∆ and K

the multiplication operator χBR p0q pxq. It turns out that the operator K is relatively compact
with respect to H. More generally, one can prove that all operators of the form f pi∇qgpx̂q,
or gpx̂qf p´i∇q, for f, g P C8 pRn q and gp´i∇q “ F ´1 gpkqF are compact. In our case,
gpxq “ χBR p0q pxq and f pkq “ pk 2 ` zq´1 , with z P CzR.
Since H has purely absolutely continuous spectrum, we conclude that:

}χBR p0q eit∆ ψ} Ñ 0 as t Ñ 8, (6.41)

for every ψ P H and for every R ą 0. In other words, if the evolution is generated by the
Laplace operator, the probability that the system is found in a ball of radius R around the
origin decays to zero as t Ñ 8, for all R ą 0 and for all initial data ψ P H: the system
moves to infinity.
As we will see later, more realistic Hamilton operators have the form H “ ´∆ ` V ,
for a potential V . Depending on the form of V , the spectrum of H may contain absolutely
continuous, singular continuous and pure point parts. Taking again K “ χBR p0q pxq (which is
still relatively compact with respect to H, at least for reasonable choices of V ), we conclude
that
}χBR p0q pxqe´iHt ψ} Ñ 0 as t Ñ 8, (6.42)
if ψ P Hac , that:
żT
1
}χBR p0q pxqe´iHt ψ}2 dt Ñ 0 as t Ñ 8, (6.43)
T 0

70
if ψ P Hac ‘ Hsc , and that:
}χBR p0q pxqe´iHt ψ} “ }χBR p0q ψ} Ñ }ψ} (6.44)
as R Ñ 8, if ψ is an eigenvector of H. In other words, if the initial data ψ is an eigenvector
(hence, it belongs to Hpp ), its evolution remains localized within a ball of radius R, if R is
large enough.
If ψ is contained in the spectral subspace Has of H, the its evolution moves to infinity,
while if it is contained in the spectral subspace Hc , with possibly a component in Hsc , the
probability for finding the state within a ball of radius R still goes to zero, but only in an
average sense.
It turns out that the behavior of }Ke´iHt ψ} can be used to dynamically characterize the
spectral subspaces Hc and Hpp associated with H.
Theorem 6.14 (RAGE theorem). Let H be a selfadjoint operator and suppose that Kn
is a sequence of relatively compact operators with respect to H, converging strongly to the
identity. Then:

1 T
! ż )
Hc “ ψ P H | lim lim }Kn e´iHt ψ} “ 0
nÑ8 T Ñ8 T 0
! )
Hpp “ ψ P H | lim sup }p1 ´ Kn qe´iHt ψ} “ 0 . (6.45)
nÑ8 tě0

Proof. Pick first ψ P Hc . By Cauchy-Schwarz and by Theorem 6.12, we find:

1 T ”1 żT
ż ı1{2
}Kn e´iHt ψ}dt ď }Kn e´iHt ψ}2 dt Ñ0 (6.46)
T 0 T 0
as T Ñ 8. Hence:
1 T
! ! ż )
Hc Ă ψ P H | ψ P H | lim lim }Kn e´iHt ψ} “ 0 . (6.47)
nÑ8 T Ñ8 T 0

On the other hand, suppose that ψ R Hc . We want to show that:

1 T
ż
}Kn e´iHt ψ}2 dt (6.48)
T 0
does not converge to zero, if we let first T Ñ 8 and then n Ñ 8. Since ψ R Hc , we have
ψ “ ψc ` ψpp , for a ψc P Hc and for ψpp P Hpp , with ψpp ‰ 0. Since }Kn e´iHt ψ} ě
}Kn e´iHt ψpp } ´ }Kn e´iHt ψc } and since we know that:

1 T
ż
lim }Kn e´iHt ψc }dt “ 0 , (6.49)
T Ñ8 T 0

it is enough to show that

1 T
ż
}Kn e´iHt ψpp }dt (6.50)
T 0
does not converge to zero, it T Ñ 8 and then n Ñ 8. To prove this, we shall show that:
sup }Kn e´iHt ψpp ´ e´iHt ψpp } Ñ 0 (6.51)
tě0

as n Ñ 8. If this is true, we obtain that:

1 T
ż
}Kn e´iHt ψpp }dt ě }ψpp } ´ sup }Kn e´iHt ψpp ´ e´iHt ψpp } Ñ }ψpp } ą 0 (6.52)
T 0 tě0

as n Ñ 8, which implies the claim. To show (6.51), we use that ψpp can be approximated
by a sequence ψN , having the form:
N
ÿ
ψN “ αj ϕj (6.53)
j“1

71
where pϕj qjPN are orthonormal eigenfunctions of H, associated with eigenvalues λj , recall
Proposition 5.64. This implies that:
N
ÿ
e´iHt ψN “ αj e´iλj t ϕj . (6.54)
j“1

Hence, for every fixed N , as n Ñ 8:

N
ÿ
sup }Kn e´iHt ψN ´ e´iHt ψN } ď |αj |}Kn ϕj ´ ϕj } Ñ 0 , (6.55)
tPR j“1

because Kn Ñ 1H strongly. Since, on the other hand, }e´iHt ψpp ´e´iHt ψN } “ }ψpp ´ψN } Ñ
0 and also:
}Kn e´iHt ψpp ´ Kn e´iHt ψN } ď }Kn }}ψpp ´ ψN } ď C}ψpp ´ ψN } Ñ 0 (6.56)
as N Ñ 8, uniformly in t and in n, we obtain Eq. (6.51). (We used that strong convergence
of Kn to 1H implies that pKn q is a bounded sequence, whose proof is left as an exercise).
This proves the first identity in Eq. (6.45). Let us now prove the second identity. The
inclusion: ! )
Hpp Ă ψ P H | lim sup }p1H ´ Kn qe´iHt ψ} “ 0 (6.57)
nÑ8 tě0

follows from Eq. (6.51). Conversely, it ψ R Hpp , then ψ “ ψc ` ψpp for ψc P Hc , with ψc ‰ 0.
Applying again Eq. (6.51), it is enough to show that
sup }p1H ´ Kn qe´iHt ψc } does not converge to zero as n Ñ 8. (6.58)
tě0

To this end, let us proceed by contradiction and assume that suptě0 }p1H ´Kn qe´iHt ψc } Ñ 0
as n Ñ 8. Then, we would conclude:
1 T
ż
0 “ lim lim }p1H ´ Kn qe´iHt ψc }dt
nÑ8 T Ñ8 T 0

1 T
ż
ě }ψc } ´ lim lim }Kn e´iHt ψc }dt “ }ψc } ą 0 (6.59)
nÑ8 T Ñ8 T 0

which is a contradiction.

7 General Schrödinger operators

7.1 Kato-Rellich theorem
Often in quantum mechanics one has to deal with perturbations H of simple reference oper-
ators H0 . As an example, one might consider Hamiltonians of the form H “ H0 ` V , with
H0 “ ´∆ and V ” V px̂q a multiplication operator, describing an external potential.
Perturbation theory aims at establishing properties of H, starting from the properties
of H0 , assumed to be well-known. Of course, to reach this goal, we will also need some
information about H ´ H0 . For example, it is easy to check that if H ´ H0 is bounded and
seldadjoint, then H is again selfadjoint (provided H0 is selfadjoint). More generally, in this
section we will show that relatively bounded perturbations of self-adjoint operators remain
self adjoint (if the relative bound is less than one).
Definition 7.1. Let A : DpAq Ñ H, B : DpBq Ñ H be two densely defined linear operators.
We say that B is relatively bounded with respect to A (or A-bounded) if DpAq Ă DpBq and
if there are constants a, b ą 0 such that:
}Bψ} ď a}Aψ} ` b}ψ} (7.1)
for all ψ P DpAq. If B is relatively bounded with respect to A, then the infimum over all
a ą 0 such that Eq. (7.1) holds true is called the relative bound of B with respect to A (or the
A-bound of B). If the A-bound of B is zero, then we say that B is infinitesimally A-bounded.

72
 The next theorem is the main result of this section.
Theorem 7.2 (Kato-Rellich). Let A be self-adjoint and B a symmetric operator, bounded
with respect to A and with A-bound less than one. Then, A`B defined on DpA`Bq “ DpAq is
selfadjoint. The statement remains true if we replace everywhere selfadjoint with essentially
selfadjoint. In this case, we have DpAq Ă DpBq and A ` B “ A ` B.

Proof. We shall only consider the case in which A is selfadjoint. We shall prove that Ran pA`
B ˘ iλ0 q “ H for a suitable λ0 ą 0. This implies that pA ` Bq{λ0 is selfadjoint, hence that
A ` B is selfadjoint.
Let ϕ P DpAq. We have, for every λ ą 0:

}pA ` iλqϕ}2 “ }Aϕ}2 ` λ2 }ϕ}2 . (7.2)

Being A selfadjoint, pA ˘ iλq´1 is bounded. Setting ϕ “ pA ` iλq´1 ψ, we have, for all ψ P H:

}ψ}2 ě }ApA ` iλq´1 ψ}2 and }ψ}2 ě λ2 }pA ` iλq´1 ψ}2 . (7.3)

Therefore, }ApA ` iλq´1 } ď 1 and }pA ` iλq´1 } ď λ´1 . From the relative boundedness, it
follows that, for ϕ “ pA ` iλq´1 ψ:
´ b¯
}BpA ` iλq´1 ψ} ď a}ApA ` iλq´1 ψ} ` b}pA ` iλq´1 ψ} ď a ` }ψ} . (7.4)
λ
Choosing λ0 ą b{p1´aq ą 0 (recall that a ă 1 by assumption), it follows that }BpA`iλq´1 } ă
1. Therefore, by the Neumann series

1H ` BpA ` iλ0 q´1 “ 1H ´ p´BpA ` iλ0 q´1 q (7.5)

is continuously invertible, and hence Ran p1H ` BpA ` iλ0 q´1 q “ H. Using that, for all
ϕ P DpAq:
p1H ` BpA ` iλ0 q´1 qpA ` iλ0 qϕ “ pA ` B ` iλ0 qϕ (7.6)
and that Ran pA ` iλ0 q “ H (recall that A is selfadjoint), we find Ran pA ` B ` iλ0 q “ H.
The same argument applies for ´iλ0 ; this proves that A ` B is selfadjoint.

Let us now discuss applications of the above theorem. We will be interested in operators
of the form H “ ´∆ ` V px̂q. We will use the Kato-Rellich theorem to establish under which
conditions on V the operator H is self-adjoint.
Theorem 7.3. (´∆-bounded potentials on R3 .) Let V : R3 Ñ R, with V P L2 pR3 q `
L8 pR3 q, that is one can write V “ V1 ` V2 with V P L2 and V2 P L8 . Then, V is
infinitesimally H0 -bounded, with H0 “ ´∆ on DpH0 q “ H 2 pR3 q. In particular, the operator
H “ H0 ` V is selfadjoint on DpH0 q.

Proof. Let DpV q “ tψ P L2 | V ψ P L2 u. DpV q contains Cc8 pRd q, and it is therefore dense
in L2 . Let V “ V1 ` V2 with V1 P L2 and V2 P L8 . Then, by the Sobolev lemma 3.83, any
function ϕ P H 2 pR3 q is continuous and bounded. Therefore:

}V ϕ}L2 pR3 q ď }ϕ}8 }V1 }L2 pR3 q ` }V2 }L8 pR3 q }ϕ}L2 pR3 q , (7.7)

that is, H 2 pR3 q Ă DpV q. The next lemma will allow us to complete the proof of infinitesimal
boundedness of V with respect to ´∆.

Lemma 7.4. For every a ą 0 there exists b ą 0 such that for all ϕ P H 2 pR3 q:

}ϕ}8 ď a}∆ϕ}L2 ` b}ϕ}L2 . (7.8)

Remark 7.5. Eq. (7.8) together with Eq. (7.7) concludes the proof of infinitesimal bound-
edness of V with respect to ´∆.

73
Proof. By Cauchy-Schwarz inequality:
}ϕ}8 ď p L1 “ }p1 ` k 2 qp1 ` k 2 q´1 ϕ}
}ϕ} p L1
ď }p1 ` k 2 q´1 }L2 }p1 ` k 2 qϕ}
p L2
2
ď Cp}k ϕ}p L2 ` }ϕ} p L2 q . (7.9)
pr pkq “ r3 ϕprkq,
Setting ϕ p one has:
}ϕ
pr }L1 pR3 q “ }ϕ}
p L1 pR3 q for all r ‰ 0. (7.10)
At the same time, we also have:
3
}ϕ
pr }L2 pR3 q “ r 2 }ϕ}
p L2 pR3 q (7.11)
and:
1
}k 2 ϕ
pr }L2 pR3 q “ r´ 2 }k 2 ϕ}
p L2 pR3 q . (7.12)
All together, we have:
}ϕ}8 ď pr }L1 ď Cp}k 2 ϕ
p L1 “ }ϕ
}ϕ} pr }L2 ` }ϕ
pr }L2 q
1 3
“ Cr´ 2 }k 2 ϕ}
p L2 ` Cr 2 }ϕ}
p L2
1 3
“ Cr´ 2 }∆ϕ}L2 ` Cr 2 }ϕ}L2 . (7.13)
Being r a free parameter, the claim follows.
e
Example 7.6 (The Coulomb potential). Let V pxq “ ´ |x| be the Coulomb potential (and ´e
the electric charge). We write:
e e e
V pxq “ ´ “ ´χ|x|ďR ´ χ|x|ąR
|x| |x| |x|
” V1 ` V2 , (7.14)
e
where V1 P L2 pR3 q and V2 P L8 . Therefore, the previous results imply that H “ ´∆ ´ |x| is
2
selfadjoint on H pRq. Analogously, it is possible to check that the N -body Hamiltonian:
N
ÿ ÿ ejk
H“ ´∆j ´ (7.15)
j“1 jăk
|xj ´ xk |

is a selfadjoint operator on H 2 pR3N q.

If the operator A is bounded below, under the same assumptions of Kato-Rellich theorem
one can also prove that A ` B is bounded below. We will not discuss the proof of this fact.
Instead, we shall focus on a special important case, the one of the hydrogenic atom:
Z
H “ ´∆ ´ , (7.16)
|x|
on DpHq “ H 2 pRd q. As we proved above, this operator is selfadjoint on H 2 pRd q. The
parameter Z ą 0 plays the role of nuclear charge (here we set e “ 1). We will prove that
this model is stable, in the sense that the Hamiltonian is bounded below by a constant. We
shall prove an optimal lower bound which matches the ground state energy of the model,
xψ, Hψy
EGS “ inf . (7.17)
ψPH 2 pRd q xψ, ψy
Notice that this is very much in contrast with what happens in classical mechanics. Clas-
sically, the Hamiltonian Hpp, qq “ p2 ´ Z{|q| is not bounded from below: one can lower
the energy by taking the electron closer and closer to the nucleus (that is, sending |q| to
zero, and choosing p “ 0). In quantum mechanics, we know from the uncertainty principle,
Eq. (5.36), that particles cannot be simultaneously localized both in space and in velocity:
this ultimately means that a particle that is very close to the nucleus should have a large
kinetic energy. The compensation between these two energies is ultimately responsible for
the stability of the hydrogenic atom, and more generally for the stability of matter. This
heuristic principle is captured by the following inequality.

74
Lemma 7.7 (Coulomb uncertainty principle.). Let H P H 1 pR3 q. Then:
ż
1
dx |ψpxq|2 ď }∇ψ}L2 pR3 q }ψ}L2 pR3 q . (7.18)
|x|

Before discussing the proof, let us use this lemma to prove the stability of the hydrogenic
atom.
Proposition 7.8. Let ψ P H 1 pRd q, Eψ “ xψ, Hψy. Then, the following inequality holds
true:
Z2
Eψ ě ´ }ψ}22 . (7.19)
4
Equality is reached for ψ “ Ke´pZ{4q|x| .
2
In particular, this proposition proves that EGS “ ´ Z4 (recall that H 2 pRd q Ă H 1 pRd q,
which follows from the definition of Sobolev space, Definition 3.74, together with |k| ď
p1{2qp1 ` |k|2 q). This inequality proves the stability of the hydrogenic atom.

Proof. (of Proposition 7.8.) Suppose that }ψ}2 “ 1. By Lemma 7.7, we have:

Z2
Eψ ě }∇ψ}22 ´ Z}∇ψ}2 ě ´ , (7.20)
4
as it follows from x2 ´ Zx “ px ´ Z{2q2 ´ Z 2 {4. Equality for ψ “ Ke´pZ{4q|x| is left as an
exercise.

To conclude, let us prove Lemma 7.7.

Proof. (of Lemma 7.7.) The starting point is the following identity:
1 ÿ xj
2xψ, ψy “ xψ, rBxj , sψy , (7.21)
|x| j“1,2,3
|x|

where we used that:

” xj ı 1 x2j
Bxj , “ ´ 3 . (7.22)
|x| |x| |x|
Therefore, integrating by parts:
1 ÿ ´ xj xj ¯
2xψ, ψy “ ´ xBxj ψ, ψy ` x ψ, Bxj ψy
|x| j“1,2,3
|x| |x|
ÿ xj
“ ´2Re xBxj ψ, ψy
j“1,2,3
|x|
ÿ xj
ď 2 |xBxj ψ, ψy| .
j
|x|

By Cauchy-Schwarz inequality:
1 ÿ ›x ›
› j ›
2xψ, ψy ď 2 }Bxj ψ}L2 › ψ › 2
|x| j
|x| L
´ÿ ¯1{2 ´ ÿ › x ›2 ¯1{2
› j ›
ď 2 }Bxj ψ}2L2 › ψ› 2
j j
|x| L
ď 2}∇ψ}L2 }ψ}L2 . (7.23)

This concludes the proof.

75
 7.2 Relatively compact perturbations and Weyl’s theorem
Kato-Rellich theorem allowed us to prove that selfadjointness survives perturbations, if they
are small enough. It is also natural to ask whether perturbations preserve other properties of
self-adjoint operators. For example, how does the spectrum of an operators looks like after
perturbation?
Let T be a selfadjoint operator, and let λ be an eigenvalue of T . Let ϕ be the eigenvector
of T with eigenvalue λ, and consider the perturbation T ` εPϕ , where Pϕ is the projector
onto ϕ. Then, ϕ is still an eigenvector of T `εPϕ , with new eigenvalue λ`ε. This shows that
the eigenvalues of a selfadjoint operator are, in general, not invariant under perturbations.
The question we would like to address here is whether there exists subsets of the spectrum
that are invariant under a class of perturbations.
Given a selfadjoint operator T with projection-valued measure PT , let us define the
discrete spectrum:

σd pT q “ tλ P R | rank PT ppλ ´ ε; λ ` εqq ă 8 for some ε ą 0u (7.24)

and the essential spectrum:

σess pT q “ tλ P R | rank PT ppλ ´ ε; λ ` εqq “ 8 for all ε ą 0u . (7.25)

Obviously, σpT q “ σd pT q Y σess pT q, and σd pT q X σess pT q “ H. The essential spectrum con-

tains the absolutely continuous spectrum, the singular continuous spectrum, accumulation
points of eigenvalues and isolated eigenvalues of infinite multiplicity. Instead, the discrete
spectrum σd pT q contains isolated eigenvalues of finite multiplicity. From what we discussed
above, we know that the discrete spectrum is not invariant under finite rank perturbations.
Instead, as we shall show, the essential spectrum is invariant under finite rank and, more
generally, compact perturbations.
Lemma 7.9 (Weyl criterion for the essential spectrum). Let T be a self-adjoint operator.
Then, λ P σess pT q if and only if there exists a sequence ψn P DpT q such that }ψn } “ 1 for
all n P N, ψn converges weakly to 0 as n Ñ 8, }pT ´ λqψn } Ñ 0. Moreover, if λ P σess pT q,
the sequence ψn can be chosen to be orthonormal. Such a sequence is called a singular Weyl
sequence at λ.
Remark 7.10. With respect to the Weyl criterion we discussed with Theorem 5.15, here
ψn Ñ 0 weakly.

Proof. Let ψn be a Weyl sequence at λ. Then, by Theorem 5.15, λ P σpT q. It is therefore

enough to show that λ R σd pT q. We proceed by contradiction. Suppose that λ P σd pT q.
Then, there is ε ą 0 such that the spectral projection Pε :“ PT ppλ ´ ε, λ ` εqq is of finite
rank. Let ϕn :“ Pε ψn . Since, by assumption, ψn Ñ 0 weakly and Pε is finite rank, we have
ϕn “ Pε ψn Ñ 0 strongly.3 On the other hand, by the spectral theorem:

}ψn ´ ϕn }2 “ xψn , PT ppλ ´ ε; λ ` εqc qψn y

ż
“ χppλ ´ ε; λ ` εqc qpxqdµψn pxq
ż
1 1
ď px ´ λq2 dµψn pxq “ 2 }pT ´ λqψn }2 Ñ 0 . (7.28)
ε2 ε
3 řM
Since Pε is finite rank, it can be written as Pε “ `“1 α` Pφ` , where tφ` u is an orthonormal family, Pφ` is
the projector over φ` and M “ rank of Pε . Therefore,
M
ÿ
P ε ψn “ α` φ` xφ` , ψn y , (7.26)
`“1

and the norm }Pε ψn } is:

M
ÿ
}Pε ψn }2 “ |α` |2 |xφ` , ψn y|2 . (7.27)
`“1

By weak convergence, xφ` , ψn y Ñ 0. Hence, }Pε ψn } Ñ 0.

76
Since }ψn } “ 1 by assumption, and }ψn } ´ }ψn ´ ϕn } ď }ϕn } ď }ψn } ` }ψn ´ ϕn }, we
conclude that }ϕn } Ñ 1. This gives rise to a contradition: Pε cannot be of finite rank, hence
λ R σd pT q.
Conversely, suppose that λ P σess pT q. We claim that there exists a singular Weyl sequence
at λ. There are two possibilities: either λ is isolated, or it is not. Suppose that λ is isolated.
Then, λ has to be an eigenvalue of infinite multiplicity. We can choose an orthonormal
sequence ψn in the eigenspace of T associated to λ. It is clear that ψn Ñ 0 weakly. Thus,
tψn u is a singular Weyl sequence.
Suppose that λ is not isolated. In this case, consider the sequence of orthogonal projec-
tions:
Pn “ PT prλ ´ 1{n; λ ´ 1{pn ` 1qq Y pλ ` 1{pn ` 1q; λ ` 1{nsq . (7.29)
Since λ is not isolated, there must be an infinite subsequence nj , such that rank Pnj ą 0 for
all j. Hence, we construct a singular Weyl sequence by choosing a normalized ψj P ran Pnj ,
for all j P N.

We are now ready to prove stability of the essential spectrum with respect to compact
perturbations.
Corollary 7.11. Let T be a selfadjoint operator and K selfadjoint and compact. Then,
σess pT ` Kq “ σess pT q.
Remark 7.12. In particular, if T is a compact selfadjoint operator, this theorem recovers
the well-known result σess pT q “ t0u.

Proof. Let λ P σess pT q and let ψn be a singular Weyl sequence at λ. Then, we have:

}pT ` K ´ λqψn } ď }pT ´ λqψn } ` }Kψn } Ñ 0 (7.30)

because ψn Ñ 0 weakly, which implies that Kψn Ñ 0 strongly. Therefore, ψn is also a

singular Weyl sequence at λ for the operator T `K, and therefore λ P σess pT `Kq. Reversing
the roles of T and T ` K, we can also show that λ P σess pT ` Kq implies λ P σess pT q.

Since, as observed at the beginning of the section, any point in the discrete spectrum can
be moved away by a finite rank perturbation, we obtain the following characterization of the
essential spectrum, whose proof will be omitted.
Theorem 7.13. Let T be a selfadjoint operator. Then,
č
σess pT q “ σpT ` Kq . (7.31)
K compact
self-adjoint

Before discussing applications, let us mention that the essential spectrum is not only
preserved by compact operators, but even by relatively compact operators. Recall that for
a selfadjoint operator T , we say that K is relatively compact with respect to T if KRT pzq is
compact for a z P ρpT q.
Theorem 7.14 (Weyl.). Let A, B be selfadjoint operators such that RA pzq ´ RB pzq is
compact, for a z P ρpAq X ρpBq. Then, σess pAq “ σess pBq.

Proof. Fix z P ρpAq X ρpBq. Let λ P σess pAq and ψn be a singular Weyl sequence for A at λ.
Then: ” 1 ı RA pzq
RA pzq ´ ψn “ ´ pA ´ λqψn . (7.32)
λ´z λ´z
Since RA pzq is bounded, we obtain that ψn is also a singular Weyl sequence for RA pzq at
the point pλ ´ zq´1 . We claim that this proves that pλ ´ zq´1 P σess pRA pzqq. Notice that
this does not directly follow from Lemma 7.9, since the operator RA pzq is not selfadjoint.
Nevertheless, the proof of Lemma 7.9 directly applies to this case as well, since the spectral
projector of A is, by construction, equal to the spectral projection of RA pzq. Also, the proof
of Corollary 7.11, together with the assumption that RA pzq ´ RB pzq is compact, implies that
pλ ´ zq´1 P σess pRB pzqq.

77
 We are left with showing that λ P σess pBq. Setting ϕn “ RB pzqψn , we find that:
›´ 1 ¯ ››
}pB ´ λqϕn } “ |z ´ λ|› RB pzq ´ ψn › Ñ 0 as n Ñ 8. (7.33)
›
λ´z
Moreover, since

lim }ϕn } “ lim }pλ ´ zq´1 ψn ` pRB pzq ´ pλ ´ zq´1 qψn } “ |λ ´ z|´1 ‰ 0 (7.34)
nÑ8 nÑ8

it follows that ϕ
rn “ ϕn {}ϕn } is a singular Weyl sequence for B at λ and that λ P σess pBq.
Reverting the roles of A and B, we conclude that σess pAq “ σess pBq.

The invariance of the essential spectrum with respect to relatively bounded perturbations
is now a simple corollary of the last theorem.
Corollary 7.15. Let T be a selfadjoint operator and let K be selfadjoint and relatively
compact with respect to T . Then, σess pT q “ σess pT ` Kq.

Proof. To begin, notice that T ` K is a selfadjoint operator. In fact:

KRT piλq “ pKRT piqqpT ´ iqRT piλq (7.35)

from which we get }KRT piλq} Ñ 0 as λ Ñ 8, since K is relatively compact with respect
to T . This implies that K is relatively bounded with respect to T , with relative bound 0:
hence, T ` K is selfadjoint, and RT `K pzq is bounded for all z P CzR.
To prove the corollary, it is enough to observe that

RT `K pzq ´ RT pzq “ RT `K pzqKRT pzq (7.36)

is the product of a bounded operator RT `K pzq and a compact operator KRT pzq, and it is
therefore compact. The claim then follows from Theorem 7.14.

7.3 Two examples of Schrödinger operators

In this section we shall discuss applications of Kato-Rellich and Weyl’s theorems. We shall
consider operators of the form H “ ´∆ ` V pxq, for suitable, explicit choices of the external
potential V pxq. As we shall see, the spectrum of H will depend dramatically on the behavior
of the function V .
Operators of this form are called Schrödinger operators. They play an important role
in quantum mechanics. The nature of the spectrum of H will allow us to understand the
dynamics generated by H, via the Schrödinger equation.

7.3.1 The harmonic oscillator

Consider the operator Hharm “ ´∆ ` ω 2 x2 , depending on a fixed parameter ω P R. For
simplicity, suppose first that the system is one-dimensional: x P R and ∆ “ d2 {dx2 . Observe
that the perturbation V pxq “ ω 2 x2 is not relatively bounded with respect to ´∆. Never-
theless, using the positivity of V pxq, we can construct a selfadjoint extension of Hharm by
means of the Friedrichs extension.
The spectrum. Next, remark that pHharm ` 1q´1 is compact. Therefore, the spectrum of
pHharm ` 1q´1 is discrete, and can only accumulate at zero (recall that the essential spectrum
is given by t0u). This implies that the spectrum of Hharm consists of isolated eigenvalues,
diverging at infinity.
To determine the eigenvalues of Hharm , we define the operators:
1 ”? 1 d ı
A˘ “ ? ωx ¯ ? . (7.37)
2 ω dx
Note that A` “ A˚´ . A simple computation shows that rA´ , A` s “ 1 and that H “
ωp2N ` 1q where N “ A` A´ . Next, we observe that:

rN , A˘ s “ ˘A˘ . (7.38)

78
Hence, if N ψ “ nψ for ψ ‰ 0, then N A˘ ψ “ pn ˘ 1qA˘ ψ. Moreover, we have }A` ψ}2 “
xψ, A´ A` ψy “ pn ` 1q}ψ}2 and }A´ ψ}2 “ xψ, A` A´ ψy “ n}ψ}2 . This implies that n ě
0, and therefore that σpN q Ă N, because if n R N was an eigenvalue, then applying A´
sufficiently many times we would find a negative eigenvalue of N .
if N ψ0 “ 0, we must have A´ ψ0 “ 0: if A´ ψ0 ‰ 0, it would be an eigenvector of N with
eigenvalue ´1. The condition A´ ψ0 “ 0 implies that:
? 1
ωxψ0 pxq “ ? ψ01 pxq (7.39)
ω

which has a unique normalized solution (up to an irrelevant phase):

2
ψ0 pxq “ pω{πq1{4 e´ωx {2
. (7.40)

Starting from ψ0 , we can construct eigenvectors ψn associated with the eigenvalue n P N of

N for all n P N, setting:
1
ψn pxq “ ? An` ψ0 pxq . (7.41)
n
We find that:
1 ? 2
ψn pxq “ ? pω{πq1{4 Hn p ωxqe´ωx {2 , (7.42)
n
2 n!
where Hn is the Hermite polynomial of degree n, given by:
n
2
” d ın ´x2 {2 2 d 2
Hn pxq “ ex {2
x´ e “ p´1qn ex e´x . (7.43)
dx dxn
It turns out that the eigenvectors tψn u form a basis of the Hilbert space L2 pRq. This might be
checked from the properties of the Hermite polynomials, or from the spectral theorem: since
Hharm is selfadjoint and has discrete spectrum, the set of eigenvectors must form a complete
basis of the Hilbert space. Summarizing, the Hamiltonian of the harmonic oscillator has the
spectrum:
σpHharm q “ σpp pHharm q “ tωp2n ` 1q : n P Nu . (7.44)
Each eigenvalue λn “ ωp2n`1q is simple, and it is associated with the normalized eigenvector
ψn . Notice that the difference λn`1 ´ λn is independent of n. In other words, the energy is
quantized: each quantum carries the energy 2ω. Applying the operator A` , we generate an
additional energy quantum, applying the operator A´ we annihilate a quantum of energy.
The operator A` is therefore called a creation operator, while A´ is called an annihilation
operator.
Properties of eigenvectors.
? In terms of the creation operator A` , the eigenvectors can be
written as ψn “ p n!q´1 An` ψ0 . A simple computation shows that the expectation values of
the position and of the momentum operator on the state ψn vanish. In fact:
1
xψn , x̂ψn y “ ? xψn , pA` ` A´ qψn y
2ω
1
“ ? xAn` ψ0 , pA` ` A´ qAn` ψ0 y
n! 2ω
2
“ ? Re xAn` ψ0 , An`1
` ψ0 y “ 0 (7.45)
n! 2ω
and similarly, with the momentum operator p̂ “ id{dx:
a
xψn , p̂ψn y “ i ω{2xψn , pA´ ´ A` qψn y
?
ω
“ i ? xAn` ψ0 , pA´ ´ A` qAn` ψ0 y (7.46)
n! 2
?
2ω
“ Im xAn` ψ0 , An`1
` ψ0 y “ 0 . (7.47)
n!

79
To have an idea of the distribution of position and momentum in the state ψn , we have to
consider the variance of these quantities. We find:
∆xψn “ xψn , x̂2 ψn y
1
“ xAn ψ0 , pA` ` A´ q2 An` ψ0 y
2ωn! `
1
“ xAn ψ0 , pA` A´ ` A´ A` qAn` ψ0 y
2ωn! `
1
“ xAn ψ0 , p2A´ A` ´ 1qAn` ψ0 y
2ωn! `
1 1 1
“ }An`1
` ψ0 } ´
2
}An` ψ0 }2 “ pn ` 1{2q . (7.48)
ωn! 2ωn! ω
Similarly,
∆pψn “ xψn , p̂2 ψn y
ω
“ ´ xAn ψ0 , pA´ ´ A` q2 An` ψ0 y
2n! `
ω
“ xAn ψ0 , pA` A´ ` A´ A` qAn` ψ0 y “ ωpn ` 1{2q . (7.49)
2n! `
We conclude that:
∆xψn ∆pψn “ pn ` 1{2q2 . (7.50)
Observe that for n “ 0, corresponding to the state ψ0 with smallest energy (the vacuum state,
with no energy quanta), the product of the variance is minimal, according to Heisenberg
uncertainty principle. For larger n, on the other hand, the uncertainty in the state ψn grows.

7.3.2 Finite well potential

The harmonic oscillator is a special example, is the sense that the Hamiltonian operator H
has a purely discrete spectrum. Here we shall consider a simple one-dimensional system,
where the spectrum of the Hamilton operator has a discrete and continuous component. We
consider a Schrödinger operator H “ ´∆ ` V , with V : R Ñ R by setting:
"
´b if |x| ă a
V pxq “ (7.51)
0 if |x| ě a

for some a, b ą 0. It is easy to check that V pxq is relatively compact with respect the Laplace
operator ´∆ “ ´d2 {dx2 . Therefore, it follows from Weyl’s theorem that the Hamilton
operator H is such that:
σess pHq “ σess p´∆q “ r0; 8q . (7.52)
We can ask whether H has additional eigenvalues. To answer this question, we shall solve the
eigenvalue problem (also known as the time-independent Schrödinger equation) Hψ “ Eψ,
i.e.: ” d2 ı
´ 2 ` V pxq ψpxq “ Eψpxq . (7.53)
dx
We find:
´ ψ 2 pxq “ Eψpxq (7.54)
for |x| ě a and:
´ ψ 2 pxq “ pE ` bqψpxq (7.55)
? ?
i Ex ´i Ex
for |x| ă a. ?It follows that,
?
if E ě 0, ψpxq “ Ae ` Ae
r if x ą a and, similarly,
i Ex ´i Ex 2
ψpxq “ Be ` Ber if x ă ´a. But then, ψ R L pRq. Hence, H has no positive
eigenvalues.
Negative eigenvalues. Let us assume now E ă 0. In this case, explucing exponentially
increasing solutions, we obtain that:
# ?
Ae´? |E|px´aq if x ě a
ψpxq “ (7.56)
Be |E|px`aq if x ď ´a

80
for some constants A, B. For |x| ă a, on the other hand, we find:

ψpxq “ C cospωxq ` C
r sinpωxq (7.57)
a
where we set ω “ b ´ |E| (the case E ă ´b can be easily excluded, since H ě ´b cannot
have eigenvalues below ´b). Next, we have to make sure that ψ and ψ 1 are continuous at
x “ ˘a (otherwise ψ is not a solution of Hψ “ Eψ on R). We obtain the conditions:

B “ C cospωaq ´ C
r sinpωaq
a
|E|B “ ωC sinpωaq ` ωC cospωaq (7.58)

at x “ ´a and

A “ C cospωaq ` C
r sinpωaq
a
´ |E|A “ ´ωC sinpωaq ` ω Cr cospωaq (7.59)

at x “ a. Thus:
ω ω r
C cospωaq ´ C
r sinpωaq “ a C sinpωaq ` a C cospωaq
|E| |E|
ω ω r
C cospωaq ` C
r sinpωaq “ a C sinpωaq ´ a C cospωaq (7.60)
|E| |E|

or equivalently
ω ω r
C ´C
r tanpωaq “ a C tanpωaq ` a C
|E| |E|
ω ω
C `C r tanpωaq “ a C tanpωaq ´ a C r (7.61)
|E| |E|
a
To
a solve these equations, we must either
a have
? C
r “ 0 and |E| “ ω tanpωaq or C “ 0 and
?
|E| tanpωaq 2 we can find solutions ω P r0; bq of the
a “ ´ω. Noticing that |E| “ b ´ ω ?
equation |E| “ ω tanpωaq intersecting the graphs b ´ ω 2 and of ω tanpωaq. Depending
on the value of b, we find finitely many solutions ω1 , . . . , ωn . It is interesting to notice that,
no matter how a small b ą 0 is, we can always find a solutions ω1 ą 0. Similarly, we can find
solutions of? |E| tanpωaq “ ´ω, by looking at the intersections of the graphs of tanpωaq
and of ´ω{ b ´ ω 2 . Also in this case, depending on the value of b, we obtain finitely many
solutions ωr1 , . . . , ω
rn2 (in this case, for b small enough, there is no solutions). For each value
of ω P tω1 , . . . , ωn1 , ω
r1 , . . . , ω
rn2 u, we can find the corresponding eigenvalue E1 , . . . , En1 `n2
and a corresponding eigenvector ψ1 , . . . , ψn1 `n2 . Let us stress, once again, that the number
of eigenvalues depend on the parameter a, b and that, no matter how small a, b ą 0 are, there
is always at least one negative eigenvalue.
Generalized eigenvectors for positive energies. We can ask whether we can find solutions of
the equation Hψ “ Eψ for E ą 0, that are associated with the continuous spectrum of H. As
noticed above, for E ą 0, that are associated with the continuous spectrum of H. As noticed
above, for E ą 0 solutions of Hψ “ Eψ are not in L2 pRq, they cannot be normalized. Still,
we can look for so-called generalized eigenfunctions, oscillating at infinity, playing the same
role as plane waves eikx play for the Laplace operator (notice that ´d2 {dx2 eikx “ k 2 eikx ;
hence eikx is a solution of the eigenvalue equation ´∆f “ Ef , with E “ k 2 ě 0).
For E ą 0, we find that solutions of Hψ “ Eψ must have the form
$
& e1 eikx ` a1 e´ikx for x ă ´a
ψpxq “ c1 eiωx ` c2 e´iωx for |x| ď a (7.62)
e2 e´ikx ` a2 eikx for x ą a
%

? ?
for appropriate coefficients e1 , e2 , a1 , a2 , c1 , c2 and where k “ E and ω “ E ` b. The
coefficients e1 and e2 are known as the incoming coefficients since they are associated to
waves eikx for x ă ´a and e´ikx for x ą a that are moving towards the obstacle (described

81
by the potential). The coefficients a1 , a2 are known as outgoing coefficients, since they are
associated to waves moving away from the obstacles, towards infinity.
The continuity of ψ, ψ 1 at x “ ˘a gives four conditions relating the six coefficients
e1 , e2 , a1 , a2 , c1 , c2 . It follows that, for every E ą 0, we can find two linearly independent
solutions of the equation Hψ “ Eψ. We can, for example, use the continuity relations
to express c1 , c2 , a1 , a2 as linear combinations of e1 , e2 (of course, the coefficients of these
combinations will depend on E and on the parameters a, b in the Hamilton operator). The
2 ˆ 2 matrix S “ SpEq giving the outgoing coefficients as a functions of the incoming
coefficients, i.e. such that pa1 , a2 q “ Spe1 , e2 q, is known as the scattering matrix of the
system. It can be shown to be a unitary matrix, describing the scattering of waves at the
obstacle.
We can build two linearly independent solutions by fixing once e1 “ 1 and e2 “ 0 (this
solution describes a wave incoming from the left), and then e1 “ 0 and e2 “ 1 (describing
a solution incoming from the right). Alternatively, we can classify solutions according to
their parity. In other words, we can find a solutions ψE,` taking e1 “ e2 “ 1 (this solutions
has positive parity, i.e. ψE,` pxq “ ψE,` p´xq) and another solution ψE,´ taking e1 “ 1
and e2 “ ´1 (this solution has negative parity, ψE,´ p´xq “ ´ψE,´ pxq). Comparing with
the case H “ ´∆, the solution of Hψ “ Eψ incoming from the left is just eikx while the
solution incoming from the right is e´ikx . The solution with positive parity is cospkxq and
the solution with negative parity is just sinpkxq. For the Laplace operator, the scattering
matrix is just S “ 1.
Completeness relation. One can prove that the states ψE,˘ (or also the two states with
energy E ą 0 associated with e1 “ 1, e2 “ 0 and with e1 “ 0, e2 “ 1) build, together with
the finitely many-eigenfunctions of H associated with negative energies, a complete set of
functions, meaning that
n ż8
dk r|ψEpkq,` yxψEpkq,` | ` |ψEpkq,´ yxψEpkq,´ |s “ 1L2 pRq
ÿ
|ψj yxψj | ` (7.63)
j“1 0

with Epkq “ k 2 . Futhermore, they satisfy the orthogonality relations:

ż ż
1
dx ψEpkq,˘ pxqψEpk1 q,˘ pxq “ δpk´k q while dx ψ Epkq,˘ pxqψEpk1 q,¯ pxq “ 0 (7.64)

(of course, also the eigenfunctions ψ1 , . . . , ψn are orthonormal). Although the generalized
eigenfunctions ψE,˘ (or also the two states with energy E ą 0 associated with e1 “ 0, e2 “ 0
and with e1 “ 0, e2 “ 1) are not in L2 pRq, they can nevertheless be used to construct
singular Weyl sequences for H at every energy E0 ą 0. To this end, it is enough to consider
linear combinations of the form
ż8
dk αpkqψEpkq,˘ pxq (7.65)
0
2
for a sequence
? of α P L pRq with }α} “ 1 and concentrating closer and closer to the fixed
value k0 “ E0 . Hence, the existence of generalized eigenfunctions for all E ą 0 is related
to the fact that σess pHq “ r0; 8q (and the fact that we can find two linearly independent
solutions ψE,˘ , for all E ą 0, is related with the multiplicity of the essential spectrum).
Time-evolution of arbitrary initial data. Because of the completeness and of the orthogonality
relations, we can also use the true eigenfunctions and the generalized eigenfunctions ψE,˘ to
compute the time-evolution of arbitrary initial data (similarly as we used Fourier transform
to describe the free evolution generated by the Laplace operator). A given ψ P L2 pRq can
be written, according to Eq. (7.63), as
ÿn ÿ ż8
ψpxq “ xψj , ψyψj pxq ` dk xψEpkq,α , ψyψEpkq,α pxq (7.66)
j“1 α“˘ 0

with ż
xψEpkq,α , ψy “ dx ψEpkq pxqψpxq . (7.67)

82
Hence,
n ÿ ż8
ÿ 2
e´iHt ψpxq “ e´iEj t xψj , ψyψj pxq ` dk eik t xψEpkq,α , ψyψEpkq,α pxq (7.68)
j“1 α“˘ 0

in close analogy with the evolution generated by the Laplace operator, computed by means
of Fourier transform.

7.4 General Schrödinger operators: existence of stationary states

So far, we considered two simple examples of Hamilton operators, whose eigenvalues and
eigenvectors (and generalized eigenvectors) could be computed explicitly. For a general
choice of the potential V pxq, we know, if V pxq is relatively compact with respect to ´∆,
that σess pHq “ r0; 8q, but it is impossible to determine explicitly the eigenvalues of H “
´∆ ` V pxq. Still, it is often possible to show the existence of negative eigenvalues through
the method of calculus of variations. This is the goal of the present section.

7.4.1 Energy functional

We consider a quantum system in d dimensions, described on L2 pRd q by the Hamilton
operator H “ ´∆ ` V pxq, assuming for now only that V P Lsloc pRd q, for a 1 ď s ď 8
(stronger conditions will come later). We consider the quadratic form associated with H,
defining the energy functional
ż ż
εpψq “ xψ, Hψy “ dx |∇ψpxq|2 ` dx V pxq|ψpxq|2 . (7.69)

We are going to establish conditions that guarantee that the functional ε attains a minimum
on the unit sphere tψ P L2 pRd q | }ψ}2 “ 1u. We will show then that the minimizer ψ0 of ε
on the unit sphere is an eigenvector of H with eigenvalue E0 “ εpψ0 q. E0 is going to be the
ground state of H, i.e. the smallest eigenvalue of H. Later, we will show how to construct
excited eigenvalues (if they exists) by similar minimization problems.
Boundedness from below. The first question we have to consider, to show the existence of a
minimizer for Eq. (7.69), is whether ε is bounded below. Consider, for example, for d “ 3,
the potential V pxq “ ´|x|´5{2 . Then, V P Lsloc pR3 q, for all s ă 6{5. For every ψ P C08 pRd q
with }ψ}2 “ 1 and for λ ą 0 we set:
ψλ pxq “ λ´3{2 ψpx{λq . (7.70)
Then, }ψλ }2 “ 1 for all λ ą 0 and
ż ż
εpψλ q “ |∇ψλ pxq| dx ´ |x|´5{2 |ψλ pxq|2
2

ż ż
“ λ´2 |∇ψpxq|2 dx ´ λ´5{2 dx |x|´5{2 |ψpxq|2 . (7.71)

For λ Ñ 0 we notice that the second term dominates and that the energy takes arbitrarily
large negative values. In this case, εpψq is not bounded below and the minimum cannot be
attained. The following theorem provides sufficient conditions to make sure that the energy
is bounded below. We use the notation
ż
T pψq “ dx |∇ψpxq|2 (7.72)

for the kinetic energy of the particle.

Theorem 7.16. Assume that V P L8 pRd q`Ld{2 pRd q, if d ě 3, that V P L8 pRd q`L1`ε pRd q,
if d “ 2, for an arbitrary ε ą 0, and that V P L8 pRd q ` L1 pRd q, if d “ 1. Then, there exists
constants C, D ą 0 with
εpψq ě CT pψq ´ D}ψ}2 . (7.73)
In particular,
E0 :“ inftεpψq | }ψ}2 “ 1u ą ´8 . (7.74)

83
Remark 7.17. Here V P Lp1 ` Lp2 means that there are V1 P Lp1 and V2 P Lp2 such that
V “ V1 ` V2 .

Proof. We consider only the case d ě 3 (the other cases can be handled analogously). By
assumption, we have V1 P L8 , V2 P Ld{2 with V “ V1 ` V2 . We claim that, for arbitrary
δ ą 0, there exists W1 P L8 , W2 P Ld{2 with V “ W1 ` W2 and }W2 }d{2 ď δ. In fact, since
|V2 pxq|d{2 χp|V2 pxq| ě µq ď |V2 pxq|d{2 for all x P Rd and since |V2 pxq|d{2 χp|V2 pxq| ě µq Ñ 0
for almost all x P Rd , as µ Ñ 8, it follows from dominated convergence that
ż
|V2 pxq|d{2 χp|V2 pxq| ě µq Ñ 0 (7.75)

as µ Ñ 8. Hence, there exists µ0 ą 0 large enough with

ż
|V2 pxq|d{2 χp|V2 pxq| ě µq ď δ d{2 . (7.76)

Then, W2 pxq “ V2 pxqχp|V2 pxq| ě µ0 q and W1 pxq “ V1 pxq ` V2 pxqχp|V2 pxq| ď µ0 q have the
desired properties. Thus
ż ż
εpψq “ |∇ψ| dx ` V |ψ|2
2

ż ż
“ }∇ψ}22 ` W1 pxq|ψpxq|2 ` W2 pxq|ψpxq|2

ě }∇ψ}22 ´ }W1 }8 }ψ}22 ´ }W2 }d{2 }ψ}22d{pd´2q

ě p1 ´ Cδq}∇ψ}22 ´ }W1 }8 }ψ}22 (7.77)

where in the last bound we used the Sobolev inequality. The theorem follows by choosing δ
small enough.

For example, for d “ 3, the last theorem can be applied to the hydrogen atom, where
V pxq “ ´1{|x| “ ´χp|x| ď 1q{|x|´χp|x| ě 1q{|x| P Lp pR3 q`L8 pR3 q, for all p ă 3. Theorem
7.16 implies that the spectrum of the hydrogen atom is bounded below, something we already
knew from Proposition 7.8. We stress again that the stability of the hydrogen atom and of
other quantum systems with attractive potentials (that is, the fact that the spectrum is
bounded below) was a crucial success of quantum mechanics. In the classical counterpart
of such systems, the energy can take arbitrarily negative values. In quantum mechanics,
stability follows thanks to the fact that the negative potential energy is compensated by the
positive kinetic energy, so that the total energy is always bounded below. In order to localize
the electron close to the singularity of the potential, we pay a price in terms of kinetic energy
(this is a formulation of Heisenberg’s uncertainty principle).

7.4.2 Weak continuity of the potential energy

Next, we look for conditions that guarantee the existence of a minimum of the energy (bound-
edness from below is a necessary but not sufficient condition). We will make use of the
following result.
Theorem 7.18. Let V : Rd Ñ R with V P L8 pRd q ` Ld{2 pRd q, if d ě 3, V P L8 pRd q `
L1`ε pRd q, if d “ 2, and V P L8 pRd q ` L1 pRd q, if d “ 1. We assume moreover that
V P L8 pRd zBR p0qq for sufficiently large R ą 0, with }V }L8 pRd zBR p0qq Ñ 0 as R Ñ 8. The
potential energy ż
P pψq “ dx V pxq|ψpxq|2 (7.78)

is then weakly continuous in H 1 pRd q. In other words, if ψj Ñ ψ weakly in H 1 pRd q, then

P pψj q Ñ P pψq as j Ñ 8.

84
Proof. We consider the case n ě 3, the other cases can be handled analogously. Let ψj be a
sequence in H 1 pRd q with ψj Ñ ψ weakly in H 1 pRd q. Then, the sequence ψj is bounded in
H 1 pRd q, i.e. }ψj }H 1 ď C for all j P N. Since
ż
V pxq|ψj pxq|2 ď }V }L8 pBRc p0qq }ψj }22 ď C}V }L8 pBRc p0qq Ñ 0 (7.79)
|x|ěR

uniformly in j, it is enough to show that:

ż ż
χBR p0q pxqV pxq|ψj pxq| Ñ χBR p0q pxqV pxq|ψpxq|2
2
(7.80)

as j Ñ 8, for an arbitrary, but fixed R ą 0. We write now V pxq “ V1 pxq ` V2 pxq, with
V1 P Ld{2 pRd q and V2 P L8 pRd q. For δ ą 0 we set
"
V1 pxq if |V1 pxq| ď 1{δ
V1,δ pxq “ (7.81)
0 otherwise

and Vδ “ V1,δ ` V2 . Then, |V1,δ pxq| ď |V1 pxq| for all δ ą 0, and V1,δ pxq Ñ V1 pxq almost
everywhere. Dominated convergence implies that:
ż ż
dx |V pxq ´ Vδ pxq| “ |V1 pxq ´ V1,δ pxq|d{2 dx Ñ 0
d{2
as δ Ñ 0. (7.82)

Therefore,
ˇż ˇ ż
ˇ χBR p0q pxqpVδ pxq ´ V pxqq|ψj pxq|2 ˇ ď |Vδ pxq ´ V pxq||ψj pxq|2 dx
ˇ ˇ
ż
ď }ψj }2d{pd´2q |Vδ pxq ´ V pxq|d{2 dx
2

ż
ď }ψj }2H 1 |Vδ pxq ´ V pxq|d{2 dx Ñ 0 (7.83)

where in the last step we used Sobolev inequality. This means that it is enough to show that:
ż ż
χBR p0q Vδ |ψj |2 Ñ χBR p0qVδ |ψ|2 (7.84)

as j Ñ 8, for all fixed δ, R ą 0. To this end, notice ψj Ñ ψ weakly in H 1 pRd q implies that
ψj Ñ ψ strongly in Lq pBR p0qq for all 1 ď q ă 2n{pn ´ 2q; see Theorem A.5. In particular,
|ψj |2 Ñ |ψ|2 strongly in Lq{2 pBR p0qq. Hence,
ˇż ˇ
ˇ χBR p0q Vδ p|ψj |2 ´ |ψ|2 qˇ ď }Vδ }8 }|ψj |2 ´ |ψ|2 }L1 pBR p0qq Ñ 0 (7.85)
ˇ ˇ

as j Ñ 8.

7.4.3 Existence of minimizers

We are now ready to show the existence of a minimum of the energy functional.
Theorem 7.19. Let V : Rd Ñ R with V P L8 pRd q ` Ld{2 pRd q, if d ě 3, V P L8 pRd q `
L1`ε pRd q, if d “ 2, and V P L8 pRd q ` L1 pRd q, if d “ 1. Moreover, let V P L8 pRd zBR p0qq
for R large enough, with }V }L8 pBRc p0qq Ñ 0 as R Ñ 8. We assume that:

E0 “ inftεpψq | ψ P H 1 pRd q, }ψ}2 “ 1u ă 0 . (7.86)

Then, there exists ψ0 P H 1 pRd q, with }ψ0 }2 “ 1 and εpψ0 q “ E0 . Moreover, the function ψ0
satisfies the Schrödinger equation in the sense of distributions:

p´∆ ` V qψ0 “ E0 ψ0 . (7.87)

85
Proof. Let ψj a sequence in H 1 pRd q with }ψj }2 “ 1 and εpψj q Ñ E0 as j Ñ 8. Theorem
7.16 implies that
1
εpψj q ě }∇ψj }22 ´ C , (7.88)
2
which implies that }∇ψj } is bounded. Hence, }ψj }H 1 ď C for all j. By the Banach-
Alaoglu theorem, this implies that there exists a subsequence ψnj and ψ0 P H 1 pRd q such that
ψnj Ñ ψ0 weakly in H 1 pRd q (in other words, ψnj Ñ ψ0 weakly in L2 pRn q and ∇ψnj Ñ ∇ψ0
weakly in L2 pRd q). Since in the weak limit the norm can only get smaller, we obtain:

}ψ0 }2 ď 1 , }∇ψ0 }2 ď lim inf }∇ψnj }2 . (7.89)

jÑ8

From Theorem 7.18 we have that P pψ0 q “ limjÑ8 P pψnj q. This implies that

E0 }ψ0 }22 ď εpψ0 q “ }∇ψ0 }22 ` P pψ0 q ď lim inf p}∇ψnj }22 ` P pψnj qq “ lim inf εpψnj q “ E0 .
jÑ8 jÑ8
(7.90)
Since E0 ă 0, we find }ψ0 }2 ě 1. This means that }ψ0 }2 “ 1 and εpψ0 q “ E0 . To show
that ψ0 satisfies the Schrödinger equation, we consider the variation of ψ0 . For δ P R and
f P C08 pRd q, let ψδ “ ψ0 ` δf and Rpδq “ εpψδ q{}ψδ }22 . Then Rpδq has a minimum in δ “ 0.
Hence, since R is differentiable in δ “ 0,

dRpδq dεpψδ q d}ψδ }22

0“ |δ“0 “ |δ“0 ´E0 |δ“0 . (7.91)
dδ dδ dδ
A simple computation shows that
ż
dεpψδ q
|δ“0 “ 2Re dx p∇f ¨ ∇ψ0 ` V f ψ0 q (7.92)
dδ
and that
d}ψδ }22
ż
|δ“0 “ 2Re dx f ψ0 . (7.93)
dδ
Thus, ż
Re rp´∆ ` V ´ E0 qf sψ0 “ 0 (7.94)

for all f P C08 pRd q. If we replace f by if we conclude that

ż
rp´∆ ` V ´ E0 qf sψ0 “ 0 , (7.95)

for all f P C08 pRn q. This shows that ψ0 solves the Schrödinger equation in the sense of
distributions.

Remark 7.20. Since inf σpHq “ inf ψPDpHq,}ψ}2 “1 εpψq, and since DpHq is dense in H 1
(DpHq is dense in H 2 , the domain of the Laplacian, which is dense in H 1 ), we conclude
that E0 “ inf σpHq.

7.4.4 Excited states

Theorem 7.19 gives a variational characterization of the smallest eigenvalue of H. It is also
possible to give a variational characterization of higher eigenvalues and eigenfunctions. Let
us assume that
E0 “ inftεpψq | ψ P H 1 pRd q, }ψ}2 “ 1u ă 0 . (7.96)
Then Theorem 7.19 implies that the energy functional εpϕq has a minimizer ψ0 on the unit
sphere of L2 pRd q which is an eigenvector of H with eigenvalue E0 . We can then define:

E1 “ inftεpψq | ψ P H 1 pRd q, }ψ}2 “ 1 and xψ, ψ0 y “ 0u (7.97)

that is we look for the infimum of the energy functional among all normalized vectors,
orthogonal to the eigenvector ψ0 . If this minimum is attained, we denote the minimizing

86
vector by ψ1 . We can proceed recursively. Given that we already constructed the normalized
vectors ψ0 , ψ1 , . . . , ψk´1 , we define:

Ek “ inftεpψq | ψ P H 1 pRd q, }ψ}2 “ 1 and xψ, ψj y “ 0, for all j “ 0, 1, . . . , k ´ 1u . (7.98)

In the next theorem, we show that if Ek ă 0 then Ek is an eigenvalue of H and the minimizer
ψk is the corresponding eigenvector.
Theorem 7.21. Let V be as in Theorem 7.19. Assume Ek ă 0. Then, the infimum in Eq.
(7.98) is attained and the minimizer ψk is such that Hψk “ Ek ψk .

Proof. The proof of the existence of a minimizer follows the same ideas as the proof of
Theorem 7.19. From a minimizing sequence ψkj , we extract a weak limit ψk . As in the proof
of Theorem 7.19, one can show that εpψk q “ Ek and that }ψk } “ 1. The only additional
observation here is that xψk , ψ` y “ 0, for all ` “ 0, 1, . . . , k ´ 1. This follows from ψkj Ñ ψk
weakly, since xψkj , ψ` y “ 0 for all ` “ 0, . . . , k ´ 1 and for all j.
To show that ψk solves the eigenvalue equation Hψk “ Ek ψk , we first show, proceeding
as in the proof of Theorem 7.19, that xf, pH ´Ek qψk y “ 0 for all f P C08 pRd q with xf, ψ` y “ 0
for all ` “ 0, 1, . . . , k ´ 1. This implies that:
k´1
ÿ
pH ´ Ek qψk “ α` ψ` , (7.99)
`“1

for appropriate coefficients α` P C. Multiplying the equation with ψi and using the orthog-
onality xψi , ψk y “ 0 for i “ 0, . . . , k ´ 1 we conclude that αi “ 0 for all i “ 0, . . . , k ´ 1 and
therefore that
Hψk “ Ek ψk . (7.100)

It follows from the recursion sketched above to define E0 , E1 , . . . only stops when it reaches
Em “ 0. Also, it is not difficult to see that the eigenvalues have finite multiplicity. Let us
sketch the proof. Suppose that Ek has infinite multiplicity, and let pψk,j q be a orthonormal
basis for the spectral subspace of Ek . Then, being orthonormal, the sequence ψk,j converges
to zero weakly in L2 . Weak convergence to zero in H 1 can be proven via an approximation
argument, using that every element in the dual of H 1 can be approximated with a Schwartz
function, and that }ψk,j }2 “ 1. By the continuity of the potential energy, we then have:

P pψk,j q Ñ 0 as j Ñ 8, (7.101)

which implies that Ek “ limjÑ8 εpψk,j q “ limjÑ8 T pψk,j q ě 0, which contradicts Ek ă 0.

Therefore, Ek is in the discrete spectrum of H. Furthermore, it is not difficult to check that
there cannot be any additional eigenvalues in p´8; 0q. In fact, if E ă 0 and if ψ P L2 pRd q are
such that Hψ “ Eψ and }ψ} “ 1, and if j ą 0 is such that E0 ď E1 ď ¨ ¨ ¨ ď Ej ă E ă Ej`1 ,
then xψ, ψ` y “ 0 for all ` “ 0, 1, . . . , j and:

xψ, Hψy “ εpψq “ mintεpϕq | }ϕ} “ 1 and xϕ, ψ` y “ 0 for all ` “ 0, 1, . . . , ju (7.102)

which by definitions implies that E “ Ej`1

Let us now comment on the essential spectrum of H. If V is a relatively compact
perturbation of the Laplacian, we know from Weyl’s theorem that σess pHq “ σess p´∆q “
r0; 8q. The next theorem gives more general conditions under which this is true, that allow
in particular to include the Coulomb potential.
Theorem 7.22. Suppose that V is a Kato class potential, that is V can be written as V1 `V2
with V1 P L8 pRd q and V2 P Lp pRd q, with }V1 }8 ă ε. Here, p “ 2 for d ď 3 and p ą d{2 for
d ě 4. Then, V is relatively compact with respect to ´∆.
Thus, under the slightly more restrictive condition that the potential is Kato class,
σess pHq “ r0; 8q.
To conclude the section, let us state, without proof, some important properties of eigen-
vectors of Schrödinger operators, that hold true in great generality.

87
 1) Uniqueness of the ground state. Under the same assumption of Theorem 7.19,
the normalized minimizer ψ0 of the energy functional ε. (which by Theorem 7.19 is
an eigenvector of H with eigenvalues E0 “ mintεpψq | }ψ} “ 1u) can be chosen (by
appropriate choice of the overall phase) to be a strictly positive function. Moreover,
up to a constant phase, ψ0 is the unique normalized minimizer. This implies that the
ground state energy E0 of H, that is the smallest eigenvalue of H, is nondegenerate.
2) Elliptic regularity. Let B1 Ă Rd be an open ball and let ψ and V be functions on B
with p´∆`V qψ “ 0 in the sense of distributions. Then, for any ball B Ă Rd concentric
with B1 and with strictly smaller radius, we have:
(i)If d “ 1, ψ is continuously differentiable on B.
(ii)If d “ 2, ψ P Lq pBq for all q ă 8.
(iii)If d “ 3, ψ P Lq pBq for all q ă d{pd ´ 2q.
(iv) If d ě 2 and V P Lp pB1 q for a d{2 ă p ď d, then ψ is Hölder continuous with
exponent α in B, for all α ă 2 ´ d{p.
(v) If d ě 1 and V P Lp pB1 q for a p ą d, ψ is continuously differentiable and the
derivative is Hölder continuous with exponent α in B, for all α ă 1 ´ d{p.
(vi) If d ě 1 and V P C k,α pB1 q (this is the subspace of C k pB1 q of functions whose k-th
derivative is Hölder continuous with exponent α) for some k ě 0 and 0 ă α ă 1,
then ψ P C k`2,α pBq.
In other words, there is a gain in regularity of two derivatives between the potential and
eigenvectors of Schrödinger operators (which by definition are only in L2 ). Note that this
regularity results hold locally. For example, this result implies that the eigenvectors of the
hydrogen atom, with V pxq “ ´Z{|x|, are C 8 in any ball away from the original. In a ball
containing the origin, on the other hand, V P Lp for all p ă 3; hence, the result above
implies that eigenvectors of the hydrogen atom are Hölder continuous with exponent α, for
any α ă 1.

7.4.5 Min-max principles

Theorem 7.22 gives a variational characterization of all negative eigenvalues of H. However,
it is usually difficult to use in practice, since it defines Ek by using all the eigenvectors ψj
associated with the eigenvalues Ej ă Ek . To compute the eigenvalues of H, the following
min-max principles are much more practical.
We denote as above by E0 ď E1 ď E2 ď ¨ ¨ ¨ ď EN ď ¨ ¨ ¨ ă 0 the eigenvalues of the
Schrödinger operator H “ ´∆ ` V . If H has a finite number J of eigenvalues, we set
EN “ 0 for all N ě J.
Theorem 7.23 (Min-max principles.). Let V as in Theorem 7.21.
Version 1. Choose φ0 , . . . , φN P H 1 pRd q such that V |φi |2 P L1 pRd q for all i and such that
xφi , φj y “ δij . We define the pN ` 1q ˆ pN ` 1q self-adjoint matrix h “ phij q0ďi,jďN by
setting hij “ xφi , Hφj y. Then, the eigenvalue problem hv “ λv has pN ` 1q eigenvalues
λ0 ď λ1 ď ¨ ¨ ¨ ď λN such that λi ě Ei for all i “ 0, 1, . . . , N .
Version 2. If N ă J,

EN “ max mintεpφN q | }φN } “ 1 and xφN , φj y “ 0, for all j “ 0, . . . , N ´ 1u (7.103)

φ0 ,...,φN ´1

where the maximum is taken over all orthonormal families φ0 , . . . , φN ´1 .

Version 3. If N ă J,

EN “ min maxtεpφq | }φ} “ 1 and φ P spanpφ0 , . . . , φN qu (7.104)

φ0 ,...,φN

where the minimum is taken over all orthonormal families φ0 , . . . , φN .

If N ě J, Version 2 and Version 3 hold true with max-min and min-max replaced by
max-inf and inf-max.

88
Remark 7.24. As it will be clear from the proof, the max and the min in the above expres-
sions are attained.

Proof. Let us assume N ă J, which is the most interesting case. We begin with Version
1. Let v0 , . . . , vN be the orthonormalřN eigenvectors of the matrix h. We use these eigenvec-
tors to define functions ξi pxq “ j“0 vi pjqφj pxq for i “ 0, 1, . . . , N . These functions are
orthonormal, since
ÿ ÿ
xξi , ξk y “ v i pjqvk p`qxφj , φ` y “ v i pjqvk pjq “ δik (7.105)
j,` j

and moreover
N
ÿ
xξi , Hξj y “ v i p`qvj pmqxφ` , Hφm y “ xvi , hvj y “ δij λi . (7.106)
`,m“0

We clearly have:
E0 ď εpξ0 q “ xξ0 , Hξ0 y “ λ0 . (7.107)
Let us now assume that Ei ď λi for all i ď k ´ 1. We prove that Ek ď λk . To this end,
we observe that dim spanpξ0 , . . . , ξk q “ k ` 1 and therefore that it must contain a function
řk
ξ “ j“0 cj ξj with }ξ} “ 1 and such that xξ, ψi y “ 0 for all i “ 0, 1, . . . , k ´ 1. By Theorem
7.22, we find:
ÿk k
ÿ
Ek ď εpξq “ ci cj xξi , Hξj y “ |cj |2 λj ď λk . (7.108)
i,j“1 j“0

This completes the proof of Version 1. To show Version 2, we set:

γN “ max mintεpφN q | xφN , φj y “ 0, for all j “ 0, . . . , N ´ 1u . (7.109)

φ0 ,...,φN ´1

Clearly, by Theorem 7.22, we have:

γN ě mintεpφN q | }φN } “ 1 and xφN , ψj y “ 0, for j “ 0, 1, . . . , N ´ 1u “ EN . (7.110)

On the other hand, for an arbitrary choice of orthonormal φ0 , . . . , φN ´1 we can find a linear
řN
combination f “ j“0 cj ψj such that f is normalized and orthogonal to all φj (because
dim spanpψ0 , . . . , ψN q “ N ` 1). Then, we have:
N
ÿ
εpf q “ xf, Hf y “ |cj |2 Ej ď EN . (7.111)
j“0

Hence, γN ď EN . To prove Version 3, we define:

γ
rN “ min maxtεpφq | }φ} “ 1 and φ P spanpφ0 , . . . , φN qu (7.112)
φ0 ,...,φN

řN řN
Choosing φ0 , . . . , φN to be ψ0 , . . . , ψN and noticing that for φ “ j“0 cj ψj with j“0 |cj |2 “
1 we have
ÿN
εpφq “ xφ, Hφy “ |cj |2 Ej ď EN (7.113)
j“0

we conclude that:

γ
rN ď maxtεpφq | }φ} “ 1 and φ P spanpψ0 , . . . , ψN qu “ EN . (7.114)

On the other hand, for arbitrary φ0 , . . . , φN , we can find f P spanpφ0 , . . . , φN q with }f } “ 1

such that xf, ψj y “ 0 for all j “ 0, 1, . . . , N ´ 1. This implies, from Theorem 7.21, that:

EN “ inftεpφq | }φ} “ 1 and xφ, ψj y “ 0 for all j “ 0, 1, . . . , N ´ 1u ď εpf q . (7.115)

Therefore, γ
rN ě EN .

89
7.4.6 Generalized min-max principle
Let us mention a simple extension of the min-max principle, which is very useful to get
bounds on sums of eigenvalues. From Version 1 of Theorem 7.23, we find in particular that:
N
ÿ N
ÿ N
ÿ N
ÿ
Ej ď λj “ Tr h “ hjj “ εpφj q . (7.116)
j“0 j“0 j“0 j“0

for any orthonormal family φ0 , . . . , φN . We can generalize this statement to the case where
the functions φj are not orthonormal. Let φ0 , . . . , φL be the pL ` 1q functions in H 1 pRd q
such that θij “ xφi , φj y defines a pL ` 1q ˆ pL ` 1q matrix θ with 0 ď θ ď 1. Suppose that
řL
Tr θ “ j“0 θjj “ N ` 1 ` δ, for a δ P p0; 1q. Then, we have:
L
ÿ N
ÿ
εpφj q ě Ej ` δEN `1 . (7.117)
j“0 j“0

To prove Eq. (7.117), consider first the case in which the functions are orthogonal (but not
necessarily normalized). Then, Tj “ θjj “ }φj }2 ď 1 (from the assumption θ ď 1). Let us
reorder the indices 0, . . . , L such that
0 ă TL ď TL´1 ď ¨ ¨ ¨ ď T0 ď 1 . (7.118)
a
Let ψj “ φj { Tj (then ψj in an orthonormal family). Then, by a telescopic rearrangement
of sums:
L
ÿ L
ÿ
εpφj q “ Tj εpψj q (7.119)
j“0 j“0
L
ÿ L´1
ÿ
“ TL εpψj q ` pTL´1 ´ TL q εpψj q `
j“0 j“0
1
ÿ
¨ ¨ ¨ ` pT1 ´ T2 q εpψj q ` pT0 ´ T1 qεpψ0 q
j“0
L
ÿ L´1
ÿ
ě TL Ej ` pTL´1 ´ TL q Ej ` . . . ` pT0 ´ T1 qE0
j“0 j“0
L
ÿ
“ Tj Ej
j“0
L
!ÿ L
ÿ ) ÿN
ě min Tj Ej | 0 ď Tj ď 1 , Tj “ N ` 1 ` δ “ Ej ` δEN `1 .
j“0 j“0 j“0

Now, let us consider the general case. Define µα and gα to be the eigenvalues and the
corresponding eigenvectors of the matrix θ. We denote by G the pL ` 1q ˆ pL ` 1q matrix
řL
with the eigenvectors gα as columns. We set Φα “ j“0 gα pjqφj . Then, we have:
ÿ
xΦα , Φβ y “ g α piqgβ pjqxφi , φj y
i,j
ÿ
“ g α piqgβ pjqθij
i,j

“ pG˚ θGqα,β “ δα,β µα . (7.120)

řL
Since α“0 Trθ “ N ` 1 ` δ, we apply the result for the case of orthogonal (but not normal-
ized) functions established above. We find that:
N
ÿ L
ÿ L
ÿ L
ÿ L
ÿ
Ej ` δEN `1 ď εpΦα q “ g α piqgα pjqxφi , hφj y “ hjj “ εpφj q , (7.121)
j“0 α“0 α,i,j j“0 j“0

which proves Eq. (7.117).

90
8 Semiclassical approximations
8.1 Dirichlet Laplacian
In the last section we gave a variational characterization for the eigenvalues of Schrödinger
operators of the form H “ ´∆`V . The question we want to address in this section is whether
it is possible to obtain information on the eigenvalues Ej by looking at the corresponding
classical system, at least in some particular regime. To simplify the analysis, we will focus
here on a special class of potentials V , For an open bounded subset Ω Ă Rd , we will consider
the potential: "
0 if x P Ω
VΩ pxq “ (8.1)
`8 if x R Ω
This (mathematically not very precise) choice means that we look at the Laplace operator
on Ω, imposing Dirichlet boundary conditions at the boundary of Ω. In other words, for
a bounded open subset Ω Ă Rd , we will consider the operator HΩ “ ´∆, defined on the
Hilbert space H02 pΩq, the closure of C08 pΩq with respect to the H 2 -norm. The eigenvalues
have a variational characterization, similarly as the Schrödinger operators discussed in the
previous section. Defining:
!ż )
E0 “ inf |∇ϕpxq|2 dx | ϕ P H02 pΩq, }ϕ}2 “ 1 (8.2)
Ω

we can show (as in the proof of Theorem 7.19, using also the fact that ψj Ñ ψ weakly in
H 1 pΩq for a bounded set Ω implies also that ψj Ñ ψ strongly in L2 pΩq) that E0 is attained
by a minimizer ψ0 with }ψ0 } “ 1, which is then a solution of HΩ ψ0 “ E0 ψ0 . Recursively,
after constructing the eigenvectors ψ0 , . . . , ψk´1 , we find that:
!ż )
Ek “ inf |∇ϕpxq|2 dx | ϕ P H02 pΩq, }ϕ}2 “ 1, xϕ, ψ` y “ 0, ` “ 0, 1, . . . , k ´ 1 (8.3)
Ω
is attained by a normalized minimizer ψk such that HΩ ψk “ Ek ψk . In this case, the recursion
never stops, HΩ has infinitely many eigenvalues (tending to infinity) and eigenvectors. Simi-
larly as discussed in the previous section, alls eigenvalues and eigenvectors of HΩ are obtained
by this recursion. Finally, it is not difficult to see that the spectrum of HΩ is purely dis-
crete: σess pHΩ q “ H. To prove this, recall Weyl’s characterization of the essential spectrum,
Lemma 7.9. A number E P R belongs to the essential spectrum of HΩ if and only if there
exists a singular Weyl sequence pψn q at E, that is a sequence such that ψn Ñ 0 weakly in L2 ,
}ψn }2 “ 1 and }pH ´ Eqψn }2 Ñ 0. This last condition, together with }ψn }2 “ 1, implies that
}ψn }H 1 ď C uniformly in n. Therefore, we can extract a weakly convergent subsequence in
H 1 , ψnj Ñ ψ. Suppose that ψ ‰ 0. Then, }ψ}22 “ limjÑ8 xψ, ψnj y “ limnÑ8 xψ, ψn y “ 0,
which gives a contradiction. Therefore, ψ “ 0.
By Theorem A.5, weak convergence in H 1 implies strong convergence in L2 on bounded
sets. Therefore, }ψnj }2 Ñ 0 as n Ñ 8, which contradicts }ψnj }2 “ 1. Thus, pψn q is not a
singular Weyl sequence. This shows that the spectrum is purely discrete, i.e. it is given by
eigenvalues of finite multiplicity.

8.2 Lower bound on the sum of Dirichlet eigenvalues

Our goal in this section is to extract information about the eigenvalues Ej . It turns out that
it is quite difficult to approximate the single eigenvalues Ej . Instead, it is easier to obtain
information about sums of eigenvalues. Besides the mathematical interest for the question
of approximation sums of eigenvalues, this is also a relevant question in physics, since, as
we shall see later, this allows to estimate the energy of a system of many non-interacting
fermions. The first result we want to discuss is a lower bound for the sum of the first N
eigenvalues of HΩ , a result due to Li-Yau and Berezin.
Theorem 8.1. Let Ω Ă Rd be open and bounded, φ0 , . . . , φN ´1 P H01 pΩq an orthonormal
family in L2 pΩq. Then:
Nÿ
´1
d ´ d ¯2{d 1`2{d ´2{d
}∇φj }22 ě p2πq2 N |Ω| , (8.4)
j“1
d ` 2 |Sd´1 |

91
where |Sd´1 | is the area of the pd ´ 1q-dimensional unit sphere. In particular:
Nÿ
´1
d ´ d ¯2{d 1`2{d 2{d
Ej ě p2πq2 N |Ω| . (8.5)
j“0
d ` 2 |Sn´1 |

Proof. Since H01 pΩq is the closure of C08 pΩq with respect to the H 1 norm, it is enough to
show the statement for orthonormal families φ0 , . . . , φN ´1 P C08 pΩq, compactly supported
away form the boundary of Ω. We extend φ0 , . . . , φN ´1 to functions in C08 pRd q, by setting
them equal to zero outside of their support. Now, we can express the H 1 norm of φj by
means of its Fourier transform. We find:
ż
}∇φj }2 “ k 2 |φpj pkq|2 dk .
2
(8.6)

Hence,
Nÿ
´1 ż
}∇φj }22 “ k 2 ρpkqdk , (8.7)
j“0

where we set
Nÿ
´1
ρpkq “ |φpj pkq|2 . (8.8)
j“0

Notice that: ż
ρpkqdk “ N . (8.9)

Moreover, with the definition ek pxq “ e´ikx χΩ pxq{p2πqd{2 , we find:

φpj pkq “ xφj , ek y . (8.10)

2 d
Extending φ0 , . . . , φN ´1 to an orthonormal basis tφj u8
j“0 of L pR q, we find:

φ̂j pkq “ xφj , ek y . (8.11)

We conclude that, by the bathtub principle, see Appendix B:

Nÿ
´1 ż
}∇φj }22 “ k 2 ρpkqdk
j“0
!ż ż )
ě inf k 2 ρpkq | ρ P L1 pRd q, ρpkqdk “ N and 0 ď ρpkq ď |Ω|{p2πqd
ż
|Ω|
“ k 2 χp|k| ď M qdk , (8.12)
p2πqd
where M ą 0 is chosen so that
ż
|Ω|
χp|k| ď M q dk “ N (8.13)
p2πqd
which implies that
´ d ¯1{d
M “ p2πq N 1{d |Ω|´1{d . (8.14)
|Sd´1 |
Therefore, Eq. (8.12) yields:
Nÿ
´1 ż ¯1{d
|Ω| ´ ´ d ¯
}∇φj }22 ě k 2 χ |k| ď p2πq N 1{d |Ω|´1{d
j“0
p2πqd |Sd´1 |
d ´ d ¯2{d 1`2{d ´2{d
“ p2πq2 N |Ω| . (8.15)
d ` 2 |Sd´1 |

92
 řN ´1
As an example, let us consider the sum SpN q “ j“0 Ej of the eigenvalues of the
operator HΩ , for the simple case in which Ω “ r0; Lsd of a cube with side length L. In this
case, eigenvectors of HΩ are products of eigenvectors of the one-dimensional Laplace operator
on the interval r0; Ls, with Dirichlet boundary conditions. Hence, we look for solutions of:

´ ψ 2 pxq “ Eψpxq ,with ψp0q “ ψpLq “ 0. (8.16)

?
The condition ψp0q “ 0 implies that ψpxq “ A sinpkxq with k “ E. The condition ψpLq “ 0
implies that k “ mπ{L, for an m P N. This gives the eigenvalues

pmπq2
Em “ , (8.17)
L2
and the eigenvector ψm pxq “ A sinpmπ{Lq (for an appropriate normalization constant A),
śd
for m P N. The energy of the product wave function ψpm1 ,...,md q px1 , . . . , xd q “ j“1 ψmj pxj q
is then given by:
d
π2 ÿ 2
Epmq “ 2 m , (8.18)
L j“1 j

for any m “ pm1 , . . . , md q P Nd . Let us now fix κ ą 0 such that the set

Kκ “ tx “ px1 , . . . , xd q P Rd | }x} ď κ and xj ě 0 for all j “ 1, . . . , du (8.19)

has volume N . In other words, we require that:

1 κd
d
|Sd´1 | “N , (8.20)
2 d
or equivalently, we fix:
´ d ¯1{d
κ “ 2N 1{d . (8.21)
|Sd´1 |
Then, the set Kκ certainly contains less than N points pm1 , . . . , md q P Nd , because every such
point can be associated uniquely with a square with unit volume (the square tpx1 , . . . , xn q |
mj ´1 ď xj ď mj u) contained in Kκ (the case with exactly N points can be excluded because
one cannot cover a ball with finitely many nonoverlapping unit cubes). Hence (remember
that we use the notation EN ´1 for the N -th eigenvalue of HΩ )

π2 2 2
´ d ¯2{d
EN ´1 ě κ “ p2πq N 2{d |Ω|´2{d (8.22)
L2 |Sd´1 |

and
Nÿ
´1 ¯2{d N
´ d ÿ
SpN q “ Ej ě p2πq2 j 2{d
|Ω|´2{d
j“0
|S d´1 | j“1
d ´ d ¯2{d
ě p2πq2 |Ω|´2{d N 1`2{d , (8.23)
d ` 2 |Sd´1 |

in agreement with the result of Theorem 8.1. A famous conjecture in mathematics, due to
George Polya, states that the bound (8.22) for the N -th eigenvalue of HΩ holds not only if
Ω is a cube but for arbitrary open bounded Ω Ă Rd . From the lower bound in Theorem 8.1
we obtain the bound:
N ´1
1 ÿ d ´ d ¯2{d ´2{d 2{d
EN ´1 ě Ej ě p2πq2 |Ω| N , (8.24)
N j“0 d ` 2 |Sd´1 |

which however misses Polya’s conjecture because of the factor d{pd ` 2q ă 1. Although
Polya’s conjecture is known to hold true for special classes of domains Ω, it remains open in
its full generality.

93
8.3 Asymptotic behavior of eigenvalues
For the cube Ω “ r0; Lsd , the right hand side of Eq. (8.23) is not only a lower bound for the
sum SpN q. Instead, it really capture the leading behavior of SpN q, in the limit of large N .
With κ as defined in Eq. (8.21), we have:
ÿ π2
SpN q » |m|2 (8.25)
L2
mPNd :|m|ďκ

řd
where |m|2 “ j“1 m2j , for m P Nd . Defining k “ m{N 1{d and λ “ κ{N 1{d “ 2pd{|Sd´1 |q1{d ,
we find
π 2 1`2{d ÿ 1 2
SpN q » N k . (8.26)
L2 N
kPNd {N 1{d :|k|ďλ

The sum on the right-hand side is a Riemann sum; as N Ñ 8, it approaches

π 2 1`2{d
ż
d ´ d ¯2{d ´2{d 1`2{d
SpN q » 2 N k 2 dk “ p2πq2 |Ω| N (8.27)
L |k|ďλ d ` 2 |Sd´1 |

up to errors of lower order in N . It turns out that the same asymptotics behavior of the sum
SpN q holds for a more general class of domains. This important result is known as Weyl’s
law. In order to state Weyl’s law, we need to introduce first the notion of boundary area.
Let Ω Ă Rd be a bounded set, BΩ its boundary. We define the boundary area ApΩq of BΩ by
1
ApΩq “ lim sup r|tx P Ωc | distpx, Ωq ă ru| ` |tx P Ω | distpx, Ωc q ă ru|s . (8.28)
rÑ0` 2r

Theorem 8.2. Let Ω Ă Rd open, bounded and with finite boundary area ApΩq. Then:
Nÿ
´1
d ´ d ¯2{d ´2{d 1`2{d
SpN q “ Ej “ p2πq2 |Ω| N ` opN 1`2{d q (8.29)
j“0
d ` 2 |Sd´1 |

in the limit N Ñ 8.
In fact, the error opN 1`2{d q in Eq. (8.29) can be estimated more precisely by:

0 ď opN 1`2{d q ď CN pApΩq{|Ω|q2{3 pd{|Sd´1 |q4{3d pN {|Ω|q4{3d (8.30)

for a universal constant C ą 0.

The result can be interpreted as a semiclassical estimate. The postulate of semiclassical
analysis is that every quantum state occupies p2πqd in the classical phase space. We are
interested in the total energy of the N states with the smallest possible energies. The
classical counterpart of the Laplace operator with Dirichlet boundary boundary conditions
is the classical Hamiltonian Hpp, xq “ p2 χpx P Ωq. To minimize the total energy, we fill the
phase space with x P Ω and |p| ď κ, where κ ą 0 is chosen, so that

κd
|Sd´1 ||Ω| “ p2πqd N , (8.31)
d
i.e. so that there is enough space for N quantum states (according to the postulate that
every quantum state occupy the volume p2πqd in phase space). We find
´ d ¯1{d
κ “ p2πq N 1{d |Ω|1{d . (8.32)
|Sd´1 |

Hence, semiclassical analysis suggests that the total energy of the N states with smallest
energy is given by:

d ´ d ¯2{d N 1`2{d
ż
1
p2 dpdx “ p2πq2 (8.33)
p2πqd Ωˆt|p|ďκu d ` 2 |Sd´1 | |Ω|2{d

94
which is exactly the statement of Theorem 8.2.
The goal of the rest of this section consists in proving Theorem 8.2. Since a lower bound
for SpN q has already been established in Theorem 8.1 (in fact, the lower bound holds for all
N , not only in the limit N Ñ 8), we need only to prove an upper bound for SpN q, coinciding
to leading order with the right-hand side of Eq. (8.29). To find such an upper bound, we
will use coherent states; this is not surprising, since we pointed out above that Weyl’s law is
a semiclassical estimate, and coherent states are as close as possible to classical states.

8.4 Upper bound on the sum of Dirichlet eigenvalues

8.4.1 Coherent states
In the present setting, coherent states are wave functions of the form

Fk,y pxq “ eik¨x Gpx ´ yq , (8.34)

where G is centered Gaussian function, y, k P Rd . Since |Fk,y pxq| “ Gpx ´ yq and |Fpk,y ppq| “
Gpp
p ´ kq, the coherent state Fk,y is localized around y in position space and it is localized
around k in momentum space. In the sequel, we will not need to assume that G is a Gaussian.
We will only assume that G P L2 pRd q with Gp´xq “ Gpxq and }G}2 “ 1 (so that }Fk,y }2 “ 1
for all k, y P Rd ).
For an arbitrary ψ P L2 pRd q, we define the coherent state transform
ż ż
ψpk, yq “ xFk,y , ψy “ Fk,y pxqψpxqdx “
r e´ik¨x Gpx ´ yqψpxqdx . (8.35)
Rd

Since by Cauchy-Schwarz ż
|Gpx ´ yq||ψpxq|dx ď }G}}ψ} (8.36)

r yq is the Fourier transform of an L1 -function; hence ψpk,

the transform ψpk, r yq is bounded.
We denote by πk,y the orthogonal projection onto Fk,y , so that

pπk,y ψqpxq “ Fk,y pxqxFk,y , ψy “ Fk,y pxqψpk,

r yq . (8.37)

The integral kernel of πk,y is given by πk,y px; zq “ Fk,y pxqFk,y pzq.
Lemma 8.3. Let G P L2 pRd q with Gp´xq “ Gpxq and }G}2 “ 1. Let ψ P L2 pRd q. Then
ż
1 r yq|2 dk “ p|ψ|2 ˚ |G|2 qpyq
|ψpk,
p2πqd
ż
1 r yq|2 dy “ p|ψ|p 2 ˚ |G|
p 2 qpkq
|ψpk,
p2πqd
ż
1 r yq|2 dydk “ }ψ}2 “ }ψ} p2.
|ψpk, 2 2 (8.38)
p2πqd
Moreover, ż
r yq “ e´iky iqy p
ψpk, ψpqqe
p Gpq ´ kqdq . (8.39)

Proof. Set Hpx, yq “ |ψpxq|2 |Gpx ´ yq|2 . By Fubini:

ż ”ż ı ż ”ż ı
Hpx, yqdx dy “ Hpx, yqdy dx “ }ψ}22 ă 8 . (8.40)
ş
Hence, the function y Ñ Hpx, yqdx “ p|ψ|2 ˚ |G|2 qpyq is in L1 pRd q and thus it is finite for
a.e. y P Rd . This means that the function x Ñ ψpxqGpx ´ yq is in L2 pRd q for a.e. y P Rd .
By Cauchy-Schwarz, this function is also in L1 pRd q, for all y P Rd , ψpk,
r yq{p2πqd{2 is the
Fourier transform of this function. Hence, by Plancherel,
ż ż
1
|ψpk, yq| dk “ |ψpxq|2 |Gpx ´ yq|2 dx “ p|G|2 ˚ |ψ|2 qpyq
r 2
(8.41)
p2πqd

95
and thus, by Fubini,
ż ”ż ı ż ż ż
1
dy dk |ψpk, yq| “ p|ψ| ˚ |G| qpyqdy “ dx dy |ψpxq|2 |Gpx ´ yq|2 “ }ψ}22 .
r 2 2 2
p2πqd
(8.42)
The second formula can be proven similarly, Finally, we show (8.39). By Plancherel,

ψpk,
r yq “ xFk,y , ψy “ xFpk,y , ψy
p . (8.43)

Since Fpk,y pqq “ e´iy¨k Gpq

p ´ kq, this proves (8.39).

Remark 8.4. The relation:

ż ż
1 r yq|2 dkdy “ 1
}ψ}2 “ |ψpk, xψ, Fk,y yxFk,y , ψydkdy , (8.44)
p2πqd p2πqd

expresses the completeness of the coherent states Fy,k , i.e.

ż
1
|Fk,y yxFk,y | dkdy “ 1L2 pRd q . (8.45)
p2πqd

8.4.2 Proof of Theorem 8.2

We are now ready to prove the main result of this section, Theorem 8.2.

Proof. (of Theorem 8.2.) For R ą 0, we consider the domain ΩpRqr “ tx P Ω | distpx, Ωc q ą
Ru. By definition of the boundary area ApΩq, we have |ΩpRq| ě |Ω| ´ 4RApΩq, for R ą 0
r
small enough.
Let now M pk, yq be a function on phase space, with 0 ď M pk, yq ď 1 for all k, y, with
supp M pk, ¨q Ă ΩpRq
r for all k P Rd , and with
ż
1
dkdy M pk, yq “ N ` ε (8.46)
p2πqd

for an arbitrary ε ą 0. We construct the operator K on L2 pRd q by defining its integral kernel
ż ż
1 1
Kpx, zq “ M pk, yqπk,y px, zqdkdy “ M pk, yqFk,y pxqFk,y pzqdkdy . (8.47)
p2πqd p2πqd

Here Fk,y are the coherent states defined by

Fk,y pxq “ eik¨x GR px ´ yq , (8.48)

where GR pxq “ R´d{2 Gpx{Rq and G P L2 pRd q is a non-negative smooth function with
Gpxq “ Gp´xq, }G}2 “ 1 and with supp G Ă B1 p0q. This guarantees that GR is non-
negative, smooth, GR p´xq “ GR pxq, }GR }2 “ 1 for all R ą 0 and that supp GR Ă BR p0q.
For any f P L2 pRd q,
ż ż
1 1
0 ď xf, Kf y ď M pk, yqxf, πk,y f y dkdy “ M pk, yq|xf, Fk,y y|2 dkdy
p2πqd p2πqd
ż
1
ď |frpk, yq|2 dkdy “ }f }22 (8.49)
p2πqd

where we used Lemma 8.3. Hence, 0 ď K ď 1. Furthermore, we find

ż
dx Kpx, xq “ N ` ε . (8.50)

In particular, this implies that K has discrete spectrum. We denote by:

λ1 ě λ2 ě λ3 ě . . . (8.51)

96
the eigenvalues of K, and by f1 , f2 , . . . the corresponding normalized eigenvectors. Note that
by
ř the restriction on the support of M and of G, supp fj Ă Ω, for all j. Note moreover, that
j λj “ N ` ε. Hence, we can find an integer L large enough with
L
ÿ
λj ą N . (8.52)
j“1
řL
We set KL “ j“1 λj |fj yxfj |. Then K ´ KL ě 0. Now we apply the generalized min-max
1{2
principle, Eq. (7.117), with φj “ λj fj , j “ 1, . . . , L. We are allowed to do so, because
řL
θij “ xφi , φj y “ λi δij and therefore 0 ď θ ď 1 and Tr θ “ j“1 λj ą N . We conclude that:
Nÿ
´1 L
ÿ L
ÿ ż
Ej ď εpφj q “ λj |∇fj pxq|2 dx
j“0 j“1 j“1
8
ÿ ż ż
ď λj |∇fj pxq|2 dx “ ∇x ∇z Kpx, zq |z“x dx
j“1
ż ż
1
“ dkdy M pk, yq dx |∇Fk,y pxq|2 . (8.53)
p2πqd
With:
∇Fk,y pxq “ ikeik¨x GR px ´ yq ` eik¨x ∇GR px ´ yq , (8.54)
and noticing that
ż ż
dx rGR px ´ yq∇GR px ´ yq ` GR px ´ yq∇GR px ´ yqs “ dx ∇|GR px ´ yq|2 “ 0 , (8.55)

we obtain:
Nÿ
´1 ż
1
Ej ď k 2 M pk, yq dkdy ` pN ` εq}∇GR }2
j“0
p2πqd
ż
1
ď k 2 M pk, yq dkdy ` CR´2 pN ` εq , (8.56)
p2πqd
because, by definition of GR , }∇GR }22 “ R´2 }∇G}22 “ CR´2 .
This bounds hold for all choices of M with 0 ď M pk, yq ď 1 for all k, y with supp M pk, ¨q Ă
ΩpRq
r for all k P R3 , and with
ż
1
dkdy M pk, yq “ N ` ε (8.57)
p2πqd
for an arbitrary ε ą 0. To minimize the average of k 2 , we choose
M pk, yq “ χpy P ΩpRqqχp|k|
r ď κq (8.58)
where we fix κ ą 0 such that
ż
1 1 r |Sd´1 | d
d
dkdy M pk, yq “ d
|ΩpRq| κ “N `ε. (8.59)
p2πq p2πq d
Hence, κ “ p2πqpN ` εq1{d pd{|Sd´1 |q1{d |ΩpRq|
r ´1{d
. With this choice of M , we compute
κd`2
ż
1 2 1 r
k M pk, yq dkdy “ | ΩpRq||S d´1 |
p2πqd p2πqd d`2
´2{d d
´ |S | ¯´2{d
d´1
“ p2πq2 pN ` εq1`2{d |ΩpRq|r . (8.60)
d`2 d
Since |ΩpRq|
r ě |Ω|´4RApΩq, we can choose R “ N ´α , for a sufficiently small α ą 0. Letting
ε Ñ 0, we conclude that
Nÿ
´1
d ´ |Sd´1 | ¯´2{d
Ej ď p2πq2 N 1`2{d |Ω|´2{d ` opN 1`2{d q (8.61)
j“0
d`2 d

which implies the theorem.

97
8.5 General Schrödinger operators
To conclude, let us briefly discuss the extension of the previous result for Schrödinger opera-
tors of the form H “ ´∆ ` V on L2 pRd q. Semiclassical analysis also give predictions for the
sum of negative eigenvalues of such Hamiltonians, for potentials V decaying at infinity, corre-
sponding to relatively compact perturbations of the Laplacian. By Weyl’s theorem, the essen-
tial spectrum of the Hamiltonian is not affected by the potential: σess pHq “ σp´∆q “ r0, 8q.
However, the negative part of the potential V´ pxq “ ´mintV pxq, 0u might generate negative
eigenvalues.
Arguing semiclassically, that is associating a volume p2πqd in phase space for every quan-
tum state, we can predict that the sum of all negative eigenvalues of H can be approximated
by:
ż
ÿ 1
Ej » d
pp2 ´ V´ pxqqχp|p2 ´ V´ pxq| ă 0qdxdp
j
p2πq
ż ż ż ż
1 2 1
“ dx p dp ´ dx V´ pxq dp
p2πqd 1{2
|p|ďV´ pxq p2πqd 1{2
|p|ďV´ pxq
ż
1 ” 1 1ı
“ |S d´1 | ´ dx V´ pxq1`d{2
p2πqd d`2 d
ż
1 2|Sd´1 |
“ ´ dx V´ pxq1`d{2 . (8.62)
p2πqd dpd ` 2q

One can prove that this prediction is indeed correct in the semiclassical limit. In fact, in
analogy with the Dirichlet Laplacian, we expect the prediction of semiclassical analysis to
become more accurate after summming a large number of eigenvalues. Here, the number of
negative eigenvalues is fixed by the choice of the potential V . In order to increase the number
of negative eigenvalues, we perform the semiclassical limit: that is, instead of considering
the Hamiltonian H, we consider:

H~ “ ´~2 ∆ ` V , ~ą0. (8.63)

The parameter ~ plays the role of Planck constant in Physics. We shall be interested in the
limit ~ Ñ 0` ; in this limit, the number of negative eigenvalues of H~ diverges. This is clear
after rewriting H~ “ ~2 p´∆ ` ~´2 V pxqq, since the negative part of ~´2 V becomes deeper in
the semiclassical limit ~ Ñ 0` . Semiclassical analysis allows to prove that, as ~ Ñ 0` :
ż
ÿ 1 2|Sd´1 | ´d
Ej “ ´ ~ dx V´ pxq1`d{2 ` op~´d q . (8.64)
j
p2πqd dpd ` 2q

Also, in analogy with the Li-Yau inequality, Theorem 8.1, one can prove that the semiclas-
sical prediction gives a lower bound to the sum of the negative eigenvalues, for the initial
Schrödinger operator H. This is encoded by the Lieb-Thirring inequality:
ÿ ż
Ej ě CLT dx V´ pxq1`d{2 , (8.65)
j

for a suitable constant CLT ă Csc , where Csc is the constant predicted by the semiclassical
1 2|Sd´1 |
approximation, Csc “ ´ p2πq d dpd`2q . Proving that the inequality Eq. (8.65) holds with CLT

replaced by Csc is a longstanding open problem in mathematical physics.

9 Many-body quantum mechanics

9.1 Bosons and fermions
In this Section we will consider quantum mechanical models for many particle systems. The
wave function for a system of N quantum particles in Rd is described by a wave function
ψN px1 , . . . , xN q P L2 pRdN q, where xi corresponds to the location of the i-th particle. More

98
generally, one might want to include the presence of extra degrees of freedom for each par-
ticle, labelled by σi “ 1, . . . , M ; in that case, the wave function of the system is denote by
ψN pz1 , . . . zM q P L2 pRdN ; CM N q. For instance, σi might denote the spin of the particle: in
that case, M “ 2. The scalar product in the presence of this extra degree of freedom is
defined as:
ÿ ż
xψN , φN y “ dx1 . . . dxN ψN pz1 , . . . , zN qφN pz1 , . . . , zN q
σ1 ,...,σN
ż
” dz1 . . . dzN ψN pz1 , . . . , zN qφN pz1 , . . . , zN q . (9.1)

We shall consider identical particles. These correspond to wave functions satisfying the
property:
|ψN p¨ ¨ ¨ zi ¨ ¨ ¨ zj ¨ ¨ ¨ q| “ |ψN p¨ ¨ ¨ zj ¨ ¨ ¨ zi ¨ ¨ ¨ q| . (9.2)
That is, the probability density for finding the particles in a given configuration does not
change is one exchanges two particles. It turns out that in Nature there exists only two type
of particles: bosons and fermions. Bosonic wave functions are symmetric with respect to
exchange of particles:
ψN p¨ ¨ ¨ zi ¨ ¨ ¨ zj ¨ ¨ ¨ q “ ψN p¨ ¨ ¨ zj ¨ ¨ ¨ zi ¨ ¨ ¨ q . (9.3)
We shall denote by L2sym pRdN ; CM N q the restriction of L2 pRdN ; CM N q to functions such that
Eq. (9.3) holds true. Example of bosonic particles are photons, the elementary constituents
of light. Instead, fermions correspond to wave functions that are antisymmetric with respect
to exchange of particles:
ψN p¨ ¨ ¨ zi ¨ ¨ ¨ zj ¨ ¨ ¨ q “ ´ψN p¨ ¨ ¨ zj ¨ ¨ ¨ zi ¨ ¨ ¨ q . (9.4)
We shall denote by L2anti pRdN ; CM N q the restriction of L2 pRdN ; CM N q to functions such that
Eq. (9.4) holds. Example of fermionic particles are electrons, neutrons and protons, which
form all elements in Nature. The antisymmetry of the wave function immediately implies
Pauli exclusion principle: a fermionic wave function is vanishing whenever xi “ xj , for any
i “ j. The probability density for finding two fermionic particles at the same location is
zero.
As a matter of fact, there is a deep connection between the possible values of the spin of
the particle and its bosonic or fermionic type: the spin-statistics theorem states that particles
with an even number of spin states are fermions, while particles with an odd number of spin
states are bosons. In the following, we shall neglect this fact, and keep the number of spin
states arbitrary for both bosons and fermions. Also, for simplicity we shall often set M “ 1.
A simple example of bosonic wave function ψN P L2sym is given by:
ψN pz1 , . . . zN q “ f pz1 q ¨ ¨ ¨ f pzN q , (9.5)
for some f P L2 . Instead, the simplest example of fermionic wave function is provided by a
Slater determinant, defined as follows. Let fi pzi q, i “ 1, . . . , N be N orthonormal functions
in L2 pRd ; CM q. The N -particle wave function
1
ψN pz1 , . . . , zN q “ ? detpfi pzj qqN
i,j“1 (9.6)
N!
is antisymmetric and normalized. It is called the Slater determinant associated to f1 , . . . , fN .
By Leibnitz formula, Eq. (9.6) can be rewritten as:
1 ÿ
ψN pz1 , . . . , zN q “ ? sgnpπqfπp1q px1 q ¨ ¨ ¨ fπpN q pxN q , (9.7)
N ! πPSN

where SN is the set of all permutations π of t1, . . . , N u, with sign sgnpπq “ ˘1. Notice
that the Slater determinant vanishes if fi “ fj for some i ‰ j, which is another instance
2
of Pauli principle. If pfi q8
i“1 form a basis of L , it is not difficult to see that a basis for
2 dN NM
Lanti pR ; C q is given by the set of all Slater determinants that can be constructed
choosing N functions among pfi q8 i“1 .

99
9.2 Reduced density matrices
Given the wave function ψN of N identical particles, the k-particle reduced density matrix
pkq
γψN is an operator on L2 pRdk q with integral kernel:
pkq
γψN py1 , . . . , yk ; x1 , . . . , xk q (9.8)
ˆ ˙ż
N
:“ dxk`1 . . . dxN ψN px1 , . . . , xk , xk`1 , . . . , xN qψN py1 , . . . , yk , xk`1 , . . . , xN q .
k
Equivalently, one writes:
ˆ ˙
pkq N
γψN “ Trk`1,...,N |ψN yxψN | . (9.9)
k
pkq
Notice that TrL2 pRdk q γψN “ N
` ˘
k . Density matrices are interesting because they allow to
compute averages of k-particle observables. For instance, consider:
Oi “ 1bpi´1q b O b 1bpN ´iq ,
ÿ
ON “ Opiq , (9.10)
i

with O acting on L2 pRd q. Then:

ÿ
xψN , ON ψN y “ xψN , Oi ψN y
i
“ N xψN , O1 ψN y
p1q
“ TrL2 pRd q OγψN . (9.11)
In general,
ř the k-particle density matrix allows to compute the average of observables of the
type ti1 ,...,ik u Opi1 ,...,ik q . In particular, let us consider the many-body Hamiltonian,
ÿ ÿ
HN “ hi ` Vij , (9.12)
i iăj

with Vij “ V pxi ´ xj q. One has:

ˆ ˙
p1q N
xψN , HN ψN y “ Tr hγψN ` xψN , V12 ψN y
2
p1q p2q
“ Tr hγψN ` Tr V12 γψN . (9.13)

Therefore, the many-body ground state energy is completely specified by γ p1q and γ p2q . It
is therefore important to know the mathematical properties of the density matrices. Being
pkq
partial traces of a nonnegative operator, γψN ě 0. The next lemma will provide an important
upper bound for the reduced one-particle density matrix of identical fermions.
Lemma 9.1. Let ψN P L2anti pRdN q. Then:

0 ď γψN ď 1L2 pRd q .

p1q
(9.14)
pkq
Remark 9.2. Being a trace class operator, γψN can be approximated by finite rank operators.
pkq ř8
That is, γψN “ j“1 λj |fj yxfj | with tfj u a ONB of L2 pRd q. The bounds in Eq. (9.14) imply
that 0 ď λj ď 1.

Proof. We shall use a Fock space formalism. We define the fermionic Fock space as:
à 2
F “C‘ Lanti pRdn q . (9.15)
n

That is, an element of F has the form ψ “ pψ p0q , ψ p1q , . . . , ψ pnq , . . .q with ψ pnq P L2anti pRdn q.
The space F becomes a Hilbert space if endowed with the standard scalar product
ÿ
xψ, ϕyF “ xψ pnq , ϕpnq yL2 pRdn q . (9.16)
ně0

100
Given f P L2 pRd q, we define the creation and annihilation operators a˚ pf q and apf q as:
a ż
pnq
papf qψq px1 , . . . , xn q “ pn ` 1q dx f pxqψ pn`1q px, x1 , . . . , xn q (9.17)
n
1 ÿ
pa˚ pf qψqpnq px1 , . . . , xn q “ ? p´1qj f pxj qψ pn´1q px1 , . . . , xj´1 , xj`1 , . . . , xn q .
n j“1

It is not difficult to see that a˚ pf q “ apf q˚ . Physically, the operator apf q destroys a fermion
with wave function f , while the operator a˚ pf q creates a fermion with wave function f . Let
tA, Bu be the anticommutator of the operators A, B: tA, Bu “ AB ` BA. It is a simple
algebraic exercise to check that:

tapf q, a˚ pgqu “ xf, gyL2 pRd q 1F , tapf q, apgqu “ ta˚ pf q, a˚ pgqu “ 0 . (9.18)

The above relations are called the canonical anticommutation relations (CAR). An important
consequence of the CAR is the boundedness of the fermionic operators:

xψ, a˚ pf qapf qψy “ }f }22 }ψ}2F ´ xψ, apf qa˚ pf qψy

ď }f }22 }ψ}2F . (9.19)

We used that xψ, apf qa˚ pf qψy “ }a˚ pf qψ}2 ě 0. As a consequence,

}apf qψ}
}apf q} “ sup ď }f }2 . (9.20)
ψPF }ψ}

This bound easily implies the desired statement for the one-particle density matrix. Let
ψ P F be an N -particle vector in the Fock space: ψ “ p0, 0, . . . , 0, ψ pN q , 0, . . . , 0, . . .q, with
ψ pN q “ ψN a normalized fermionic wave function. A simple computation shows that:
p1q
xψ, a˚ pf qapgqψy “ xpapf qψqpN ´1q , papgqψqpN ´1q y “ xg, γψ f yL2 pRd q . (9.21)

Therefore:
p1q
xf, γψ f y “ xψ, a˚ pf qapf qψy ď }f }22 , (9.22)

which implies that γψ ď 1.

p1q

Remark 9.3. The above upper bound is not true for bosons: there, γψ ď N 1. This suggests
p1q

that bosonic one-particle density matrices might have large eigenvalues. One can check that
for factorized states the reduced one-particle density matrix has one eigenvalue equal to N .
To conclude, as an example let us compute the reduced one-particle density matrix of the
simplest fermionic wave functions for N fermions, namely Slater determinants. Consider:
1 ÿ
ψN “ ? sgnpπqfπp1q px1 q ¨ ¨ ¨ fπpN q pxN q , (9.23)
N! π

with tfi u orthonormal. A simple computation shows:

N
ÿ
p1q
γψ “ |fi yxfi | . (9.24)
i“1

p1q p1q p1q˚ p1q2 p1q

That is, γψ is a rank´N orthogonal projector: γψ “ γψ “ γψ , Trγψ “ N . In this
case, the eigenvalues of the density matrix are either 0 or 1.

9.3 Atoms and molecules

In the following, we shall focus on a specific model in quantum mechanics, of great relevance
for physics and chemistry. The model describes a system of N fermions (electrons) interacting

101
with K fixed nuclei. For K “ 1, this model describes an atom with N electrons, for K ą 1
it describes a molecule. The Hamiltonian is:
N N ÿ
K N K
ÿ ÿ Zi ÿ 1 ÿ Zi Zj
HN,K pZ, Rq “ ´∆i ´ ` ` , (9.25)
i“1 i“1 j“1
|xi ´ Rj | iăj“1 |xi ´ xj | iăj“1 |Ri ´ Rj |

on L2 pR3N ; CM N q. Let us discuss the various terms. The first term describes the kinetic
energy of the N particles; ∆i is the Laplacian acting on the i-th particle,

p∆i ψN qpz1 , . . . , zN q “ ∆xi ψN pz1 , . . . , zN q , i “ 1, . . . , N . (9.26)

The second term takes into account the interaction between the electrons, with positions xi ,
and the nuclei, located at Rj . Units are chosen so that the charge of the electron is ´1, and
the charge of the nuclei is Zj P N. The sign of the Coulomb potential shows that the energy
decreases when the particles and the nuclei are close: the interaction is attractive. The third
term describes the electrostatic interaction among the electrons: the sign of the Coulomb
potential shows that the energy increases when two electrons are close: the interaction is
repulsive. Finally, the last term takes into account the Coulomb repulsion of the nuclei.
Notice that xi is a multiplication operator, while Rj is a fixed vector in R3 : that is, the
nuclei are treated as fixed in space. This is motivated by the fact that that, physically, the
masses of the nuclei are much larger than the masses of the electrons (chosen to be equal
to 1{2 in our units). Later, we shall minimize over the positions of the nuclei, to find the
optimal energy of the system.
At zero temperature, the state of the system coincides with the ground state of the
Hamiltonian HN,K pZ, Rq. Since we are interested in describing a system of N electrons, and
since electrons are fermions, we shall consider the fermionic ground state energy:

EN,K pZ, Rq “ inf xψN , HN,K pZ, RqψN y (9.27)

ψN PL2anti pR3N ;CN M q
}ψN }2 “1

These are the typical questions we shall study in the following.

1. Is the system stable? That is, EN,K pZ, Rq ą ´8? If so, this is called stability of matter
of the first kind. Stability is a purely quantum mechanical phenomenon: it is false in classical
mechanics. We have seen that stability holds for N “ 1, K “ 1.
2. From experience, we know that the energy of a physical system scales linearly with the
number of constituents. If not, this would imply a huge release or absorption of energy as
two systems are merged together. Suppose for instance that EN,K pZ, Rq „ ´CpN ` Kq2 .
Then,
E2N,2K “ ´4CpN ` Kq2 ! 2EN,K . (9.28)
This means that it would be energetically much more convenient to merge together two sys-
tems composed by N particles and K nuclei, which is not what we observe. This gain is
incompatible with the observation that matter is extensive (doubling the number of parti-
cles of a physical system corresponds to a macroscopic variation of the volume the system
occupies). We say that stability of matter of the second kind occurs if:

EN,K pZ, Rq ě ´CpZ, RqpN ` Kq . (9.29)

Of course, stability of matter of the second kind implies stability of matter of the first kind.
Does stability of matter of the second kind occurs for the model in Eq. (9.25)?
3. In order for an atom or a molecule to be stable, the ionization energy to remove an electron
must be positive. That is, if EN `1,M ă EN,M : it is energetically more convenient for the
system to attract one more electron. Under which conditions the ionization energy is posi-
tive? We know from experience that there are no atoms with N ą Z ` 2. This is intuitively
clear: the ionization energy will be zero, when the total charge of the electrons compensates
the total charge of the nucleus, so that the atoms looks neutral at large distances. Can one
prove this mathematically?

102
4. As N increases, the model becomes quickly intractable from an analytic point of view.
Can we say anything quantitative about, e.g., the ground state energy of the system for N
large?
In order to understand these questions, we shall first consider them in a simplified theory,
the Thomas-Fermi model. Later, we shall discuss the rigorous connection between Thomas-
Fermi theory and the original many-body problem.

9.4 Thomas-Fermi theory

Thomas-Fermi (TF) theory is an effective theory for many-body quantum mechanics, which
takes as only input the density of the quantum system, defined as:
ż
ρpzq “ N dz2 . . . dzN |ψN pz, z2 , . . . , zN q|2 , (9.30)
ş
where ψN is the many-body wave function of the system. Clearly, ρ ě 0, dz ρpzq “ N . The
quantity ρpzq{N describes the probability density for finding a particle at z “ px, σq. Any
wave function ψN determines uniquely a density ρ; clearly, the converse does not hold.
TF theory is much easier to study than the full many-body problem, due to the fact
that it depends on much less degrees of freedom (the density is a function on R3 , while the
wave function is a function on R3N ). Later, we will discuss the rigorous validity of this
approximation, in the limit in which N, Z Ñ 8.
The main approximation introduced in TF theory is the replacement of the kinetic energy
of the system with a functional of ρ. In order to understand this approximation, let us first
discuss a simple example.

9.4.1 The free Fermi gas

Consider a system of N spinless, noninteracting particles confined in a cubic box Λ, of side
1, with periodic boundary conditions: Λ “ T3 , with T3 the unit torus in three dimensions.
Let us first start with N “ 1. The Hamiltonian is H “ ´∆, on L2 pT3 q. The eigenfunctions
of the Hamiltonian are given by plane waves:
Hfp “ |p|2 fp , fp pxq “ e´ip¨x , p P p2πqZ3 . (9.31)
The vector p is called the momentum of the particle; the constraint p P p2πqZ3 is due
to the requirement of periodic boundary conditions. The energy of the quantum particle
with momentum p is |p|2 . Consider now a system of N ą 1 particles. The Hamiltonian
řN
is HN “ j“1 ´∆j . Clearly, if pfpi q are eigenstates of the Laplacian, then their product
řN
fp1 ¨ ¨ ¨ fpN is an eigenstate of HN with energy i“1 |pi |2 . Since we are interested in fermionic
eigenstates, we shall consider antisymmetric combinations of products of plane waves, that
is Slater determinants:
1 ÿ
? sgnpπqfpπp1q px1 q ¨ ¨ ¨ fpπpN q pxN q . (9.32)
N! π
Notice that as soon as pi “ pj for i ‰ j, the Slater determinant vanishes (Pauli principle).
The fermionic ground state of HN is given by the Slater ř
determinant with the smallest energy.
N
To find such state, we have to minimize the quantity i“1 |pi |2 under the constraints that
3
pi ‰ pj for i ‰ j, and pi P p2πqZ . The solution to this problem is provided by “filling the
Fermi ball”: one considers the N momenta pi with smallest modulus. In general, one has
that |p| ď pF , where the Fermi momentum pF scales as pF „ cN 1{3 for some constant c ą 0.
Notice that in general not all states with momenta such that |p| ď pF will be occupied: the
Fermi ball might be only partially filled, and the ground state might be degenerate.
Suppose, for the sake of simplicity, that the number of particles N is chosen so that the
Fermi ball is completely filled. The ground state energy of the system is:
ÿ ÿ 1 2
EN “ |p|2 “ N |p| . (9.33)
3 3
N
pPp2πqZ pPp2πqZ
|p|ďcN 1{3 |p|ďcN 1{3

103
Changing variable, one has:
2
ÿ 1 2
EN “ N 1` 3 |p| . (9.34)
N
pP 2π
Z3
N 1{3
|p|ďc

As N Ñ 8, the sum converges to an integral. One has:

ż
5 5 5 5
EN “ N 3 dp |p|2 ` opN 3 q “ CN 3 ` opN 3 q . (9.35)
|p|ďc

Thus, the ground state energy of the system, which is purely kinetic, scales as N 5{3 . More
2
generally, in d-dimensions one would find N 1` d . This asymptotic behavior is in agreement
with the Weyl law for the sum of the first N eigenvalues of the Dirichlet Laplacian, recall
Theorem 8.2. In the present case, however, the domain Ω has no boundary, hence Theorem
8.2 does not apply directly. One can actually show that the constant C is equal to the
constant appearing in the Weyl asymptotics. In the present example, the density ρpxq
associated to the ground state is constant: ρpxq “ ρ “ N . Thus, the kinetic energy of the
confined system scales as ρ5{3 . This connection between kinetic energy and density turns out
to be much more general, and it plays a crucial role in defining the Thomas-Fermi energy
functional.

9.4.2 The Thomas-Fermi energy functional

In TF theory, the energy of the system is determined by the electron density via the following
functional (we omit the spin of the system for simplicity):
ż ż ż
5{3 1 ρpxqρpyq
ETF pρq “ cTF dx ρpxq ´ dx ρpxqV pxq ` dxdy `U , (9.36)
2 |x ´ y|
where V pxq is the electrostatic potential generated by the K fixed nuclei, and U is the
electrostatic repulsion of the nuclei:
K K
ÿ Zj ÿ Zi Zj
V pxq “ , U“ . (9.37)
j“1
|x ´ Rj | iăj“1
|R i ´ Rj |

The constant cTF is positive, and later it will be suitably chosen, in order to connect with
the original many-body problem. The following discussion will only use that cTF ą 0.
The domain of the TF functional is given by the set of allowed densities:

FN “ tρ : R3 Ñ R | ρpxq ě 0 , }ρ}1 “ N , ρ P L5{3 pR3 qu . (9.38)

As we shall prove later, the TF functional is well-defined on this domain. The TF ground
state energy is:
TF
EN “ inf ETF pρq . (9.39)
ρPFN

Before discussing the mathematical properties of the functional, let us discuss its physical
origin. The first term in Eq. (9.36) takes into account the kinetic energy of the system. As
we have seem for a homogeneous electron gas, Section 9.4.1, the kinetic energy of the system
grows as ρ5{3 . For a general system, one cannot expect the density ρpxq associated to the
ground state to be constant. Nevertheless, in general it will vary on a scale that is much
smaller that the mean interparticle distance; to approximate the ground state, one ş fills a
“local” Fermi ball, with radius ρpxq1{3 , and integrates over space. This yields the ρpxq5{3
term in the TF energy functional. This approximation of the kinetic energy turns out to be
rigorously justified, as we shall discuss later with the Lieb-Thirring kinetic energy inequality.
The second term describes the electrostatic interaction between the electrons and the
nuclei. In the full many-body problem, this is given by:
N ÿ
K
ÿ Zj
xψN , ψN y . (9.40)
i“1 j“1
|x i ´ Rj |

104
We have:
N ÿ
K N ÿ
K ż
ÿ Zj ÿ 1
xψN , ψN y “ Zj dx1 . . . dxN |ψN px1 , . . . , xN q|2
i“1 j“1
|xi ´ Rj | i“1 j“1
|xi ´ Rj |
K ż
ÿ 1
“ N Zj dx1 . . . dxN |ψN px1 , . . . , xN q|2
j“1
|x1 ´ Rj |
ż
” dx ρpxqV pxq , (9.41)

where in the second step we used the (anti)symmetry of the wave function. The right-hand
side reproduces exactly the second term in the TF energy functional: hence, no approxima-
tion is made here. Consider now the third term. This describes the electrostatic repulsion of
the electrons: it appears as a classical electrostatic energy, generated by the charge density
ρpxq. In the full many-body problem, this terms corresponds to:
N
ÿ 1
xψN , ψN y . (9.42)
iăj“1
|xi ´ xj |

Consider the electrons as classical point particles, with positions xi ; treat them as indepen-
dent, identically distributed random variables, with probability distributions ρpxq{N . The
law of the large numbers implies:
ż
1 ÿ 1 ρpxq 1
» dx . (9.43)
N j:j‰j |xi ´ xj | N |xi ´ x|

Under this approximation, we replace Eq. (9.42) by:

N N
ÿ 1 1 ÿ 1
xψN , ψN y “ xψN , ψN y
iăj“1
|x i ´ xj | 2 i‰j
|x i ´ xj |
N
1 ÿ
» xψN , wpxi qψN y , (9.44)
2 i“1

with wpxq “ pρ ˚ | ¨ |´1 qpxq. The big conceptual simplification here is that we replaced a sum
of two-body operators by a sum of one-body operators, exploiting an averaging principle.
Then, we can repeat the computation in Eq. (9.41). We have:
N ż
1 ÿ 1 1
xψN , wpxi qψN y “ dxdy ρpxqρpyq , (9.45)
2 i“1
2 |x ´ y|

which is precisely the third term appearing in the TF energy functional. Finally, the fourth
term appearing in the TF functional is equal to the corresponding term appearing in the full
many-body problem, hence no further approximation is introduced at this point.
The mathematical foundations of TF theory have been developed by Lieb and Simon
in the seventies, see [2] for a review, fifty years after the introduction of the functional by
Thomas and Fermi. It is a milestone in mathematical physics; its development played a
crucial role in understanding the problem of stability of matter for large quantum systems.
Here we shall discuss the mathematics of the TF energy functional, and in particular how
to solve the problems 1.-4. spelled out in Section 9.3 within the framework of TF theory.
Later, we will show how the TF approximation can be rigorously justified starting from the
original many-body problem.
Let us now prove that the TF energy functional is well-defined on its domain FN . The
finiteness of the first term in Eq. (9.36) follows from ρ P L5{3 . Consider the second term.
We rewrite it as:
ż ż ż
dx ρpxqV pxq “ dx ρpxqVă pxq ` dx ρpxqVą pxq , (9.46)

105
where:
K K
ÿ Zj ÿ Zj
Vă pxq “ χp|x ´ Rk | ď 1q , Vą pxq “ χp|x ´ Rj | ą 1q . (9.47)
j“1
|x ´ Rj | j“1
|x ´ Rj |

Consider the first term. By Hölder inequality, we have:

ż
χp|x ´ Rj | ď 1q › χp| ¨ | ď 1q ›
dx ρpxq (9.48)
› ›
ď }ρ}5{3 › ›
|x ´ Rj | |¨| 5{2

which is finite, thanks to the fact that ρ P L5{3 . The second term can be estimated immedi-
ately, using that: ż
χp|x ´ Rj | ą 1q
dx ρpxq ď }ρ}1 “ N . (9.49)
|x ´ Rj |
All together: ż ÿ
dx ρpxqV pxq ď C Zj p}ρ}1 ` }ρ}5{3 q . (9.50)
j“1

Finally, consider the third term in Eq. (9.36). This will be estimated using the Hardy-
Littlewood-Sobolev inequality. Let f P Lp pRd q and h P Lr pRd q. Then, for p1 ` 1r ` λd “ 2:
ˇż 1 ˇ
ˇ dxdy f pxqhpyq ˇ ď Cpλ, d, pq}f }p }h}r . (9.51)
ˇ ˇ
|x ´ y|λ
See [3] for a proof. To apply this inequality to the TF functional, we choose λ “ 1, d “ 3,
and f “ h “ ρ. Choosing p “ r, one has p “ 6{5; hence:
ż
1 1
Dpρ, ρq “ dxdy ρpxqρpyq ď C}ρ}26{5 . (9.52)
2 |x ´ y|
The right-hand side is finite, since by interpolation:

}ρ}6{5 ď }ρ}λ1 }ρ}1´λ

5{3 , (9.53)

with λ “ 7{12. This shows that ETF is well defined on FN .

9.4.3 Existence and uniqueness of the minimizer in DN

In this section we shall start the study of the variational problem associated to the TF
functional. Before starting, let us comment about the fact that in general one does not
expect the minimizer to exists for all values of N . In fact, one expects the system tořbe able
K
to bind a finite number of electrons, dependent on the total nuclear charge Ztot “ j“1 Zj .
This is due to the fact that, for N “ Ztot , the total charge of the electrons is equal to the
total charge of the nuclei. Hence, at large distances, the system will look charge neutral, and
will not be able to attract any further electron. This is confirmed by the fact that in nature
one does not observe stable atoms with N ą Ztot ` 2.
Mathematically, one does not expect the minimizer to exists in FN , for any N . Calling
ρ˚ the minimizer, it might happen that:
ż
dx ρ˚ pxq ă N , (9.54)

which means that ρ˚ R FN . Therefore, in order to avoid this problem for the moment, we
will consider the functional on a larger domain,
! ż )
1 5{3
DN “ ρ P L X L | ρ ě 0, dx ρpxq ď N . (9.55)

This new space allows to take into account the “loss” of electrons at infinity. We will first
prove the existence and uniqueness of the minimizer in this domain, and then later we will
prove that, for suitable values of N , the minimizer actually belongs to FN (particles are not
lost at infinity).

106
Theorem 9.4 (Existence of minimizers in DN .). There exists ρ˚ P DN such that the follow-
ing is true:
inf ETF pρq “ ETF pρ˚ q . (9.56)
ρPDN

The proof will be based on the following auxiliary result.

Lemma 9.5. Let ρ1 , ρ2 in DN , and ρj Ñ ρ2 weakly in Lp for all p P p1, 5{3s. Then:

lim Dpρ1 , ρj q “ Dpρ1 , ρ2 q , Dpρ1 , ρ2 q ď Dpρ1 , ρ1 q1{2 Dpρ2 , ρ2 q1{2 . (9.57)

jÑ8

Proof. (of Lemma 9.5) Let us prove the first property. To this end, we rewrite:
ż ż
1 1
Dpρ1 , ρj q “ dxdy ρ1 pyqρj pxq ” dx ρj pxqf pxq , (9.58)
2 |x ´ y|

where f pxq “ p1{2qpρ1 ˚ | ¨ |´1 qpxq. We decompose the function f as f “ fă ` fą , where:

ż ż
1 χp|x ´ y| ď 1q 1 χp|x ´ y| ą 1q
fă pxq “ dy ρ1 pyq , fą pxq “ dy ρ1 pyq (9.59)
2 |x ´ y| 2 |x ´ y|

Consider fă . By Hölder inequality,

› χp| ¨ | ď 1q ›
(9.60)
› ›
}fă }8 ď }ρ1 }5{3 › › ă8
|¨| 5{2

and: ż
χp|x ´ y| ď 1q
}fă }1 “ dxdy ρ1 pxq “ C}ρ1 }1 ă 8 . (9.61)
|x ´ y|
Therefore, by interpolation fă P Lp pR3 q for all p P r1, 8q. Hence, by weak convergence:
ż ż
lim dx ρj pxqfă pxq “ dx ρ˚ pxqfă pxq . (9.62)
jÑ8

Consider now fą pxq. By Young’s inequality for convolutions,

› χp| ¨ | ą 1q ›
(9.63)
› ›
}fą }p ď }ρ1 }q › ›
|¨| r

1
with p1 “ 1q ` 1r ´ 1 ď 1r ă 13 . Therefore, p ą 3. Using that Lp is equal to the dual of Lp
with p1 P p1, 3{2q, by weak convergence:
ż ż
lim dx ρj pxqfą pxq “ dx ρ˚ pxqfą pxq . (9.64)
jÑ8

This together with Eq. (9.62) proves the first claim in Eq. (9.57). Consider now the second
claim. To prove it, we proceed as follows. Let hpxq “ Ce´c|x| , and let Kpxq “ ph ˚ hqpxq.
Notice that K is a radial function: Kpxq “ Kp|x|q. Let us choose the constant C such that:
ż8
1
dt Kptq “ . (9.65)
0 2

In particular, by a change of variables:

ż8
1
“ dt Kpt|x ´ y|q . (9.66)
2|x ´ y| 0

We can further rewrite this as:

ż8 ż
1
“ dt t3 dz ht px ´ zqht py ´ zq , ht px ´ zq “ hptpx ´ zqq . (9.67)
2|x ´ y| 0

107
Therefore, using this decomposition of the Coulomb potential:
ż ż8 ż
3
Dpρ1 , ρ2 q “ dxdy ρ1 pxqρ2 pyq dt t dz ht px ´ zqht py ´ zq
0
ż8 ż
“ dt t3 dz pρ1 ˚ ht qpzqpρ2 ˚ ht qpzq (9.68)
0

where the exchange of integrations is allowed by Fubini’s theorem. By Cauchy-Schwarz

inequality:
´ż 8 ż ¯1{2 ´ ż 8 ż ¯1{2
Dpρ1 , ρ2 q ď dt t3 dz pρ1 ˚ ht q2 pzq dt t3 dz pρ2 ˚ ht q2 pzq
0 0
“ Dpρ1 , ρ1 q1{2 Dpρ2 , ρ2 q1{2 . (9.69)

This concludes the proof of the second of Eq. (9.57), and of the Lemma.

We are now ready to prove Theorem 9.4.

Proof. Let ρj be a minimizing sequence in DN . The bounds used to prove the wellposedness
of the TF functional on FN easily imply that:
5{3
ET F pρj q ě a}ρj }5{3 ´ bN . (9.70)

Therefore, using that |ET F pρj q| ď C (which follows from the finiteness of the j Ñ 8 limit),
the above estimate allows to prove an a priori bound on }ρj }5{3 :

}ρj }5{3 ď K , (9.71)

for some constant K independent of j. Since }ρj }1 ď N for all j, by interpolation we get:

}ρj }p ď C , @p P r1, 5{3s . (9.72)

By Banach-Alaoglu theorem, we know that, up to the extraction of a subsequence, ρj Ñ ρ˚

weakly in Lp , for p P p1, 5{3s (one can actually prove that the limit is independent of p).
Also, one can easily check that ρ˚ P DN . Let us first prove that ρ˚ ě 0. Suppose it is false.
Then, there exists a bounded set A Ă R3 such that:
ż
dx ρ˚ pxqχA pxq ă 0 . (9.73)
ş ş
However, since by weak convergence dx ρ˚ pxqχA pxq “ limjÑ8 dx ρj pxqχA pxq, and ρj ě 0,
Eq. (9.73) would imply a contradiction. Thus, ρ˚ ě 0. In a similar way, one can prove that
}ρ˚ }1 ď N . Suppose it is false. Then, there exists a bounded set A such that:
ż
dx ρ˚ pxqχA pxq ą N . (9.74)

Repeating the same argument as before, this implies a contradiction. Thus, ρ˚ P DN . To

prove the claim (9.56), we shall show that:

ETF pρ˚ q ď ETF . (9.75)

Consider the kinetic energy contribution. By the lower semicontinuity of norms, one gets:

lim inf }ρj } ě }ρ˚ }5{3 . (9.76)

j

Consider now the electrons-nuclei interaction. We claim that:

ż ż
lim dx ρj pxqV pxq “ dx ρ˚ pxqV pxq . (9.77)
j

108
To prove this, we write:
V pxq “ Vă pxq ` Vą pxq , (9.78)
with:
K K
ÿ Zj ÿ Zj
Vă pxq “ χp|x ´ Rj | ď 1q , Vą pxq “ χp|x ´ Rj | ą 1q . (9.79)
j“1
|x ´ Rj | j“1
|x ´ Rj |

1
Consider first Vą . This function belongs in Lp for p ą 3, which is the dual of Lp , for
p1 P p1, 3{2q. Since ρj Ñ ρ˚ in Lp with p1, 5{3s and p1, 3{2q Ă p1, 5{3s, we have:
ż ż
lim dx ρj pxqVą pxq “ dx ρ˚ pxqVą pxq . (9.80)
jÑ8

Consider now Vă . This function belongs to L5{2 , which is the dual of L5{3 . Thus, by weak
convergence: ż ż
lim dx ρj pxqVă pxq “ dx ρ˚ pxqVă pxq . (9.81)
jÑ8

Eqs. (9.80), (9.81) imply Eq. (9.77). Finally, we claim that:

lim inf Dpρj , ρj q ě Dpρ˚ , ρ˚ q . (9.82)

j

The proof of this inequality follows from Lemma 9.5. From

Dpρ˚ , ρ˚ q “ lim Dpρ˚ , ρj q (9.83)

jÑ8

and:
Dpρ˚ , ρj q ď Dpρ˚ , ρ˚ q1{2 Dpρj , ρj q1{2 , (9.84)
we get:
lim inf Dpρj , ρj q1{2 ě Dpρ˚ , ρ˚ q1{2 (9.85)
j

which proves Eq. (9.83). All in all,

ż
5{3
ETF “ lim ETF pρj q ě lim inf cTF }ρj }5{3 ´ lim dx V pxqρj pxq ` lim inf Dpρj , ρj q
j j j j
ě ETF pρ˚ q , (9.86)

which concludes the proof of the theorem.

To conclude, we will prove convexity of the TF energy functional, that will be important
in establishing the uniqueness of the minimizer, and to understand the behavior in N of the
TF ground state energy.
Lemma 9.6 (Convexity of the TF functional.). The domain DN is convex. Moreover, the
TF functional is strictly convex: for any ρ1 , ρ2 P DN , ρ1 ‰ ρ2 and λ P p0; 1q:

ETF pλρ1 ` p1 ´ λqρ2 q ă λETF pρ1 q ` p1 ´ λqETF pρ2 q . (9.87)

Proof. The convexity of DN is a simple exercise (if ρ1 and ρ2 belong to DN then it is easy to
check that the convex combination ρλ “ λρ1 ` p1 ´ λqρ2 belongs to DN ). Next, let us prove
the convexity of the TF functional. We shall
ş study the different contributions separately.
Consider the kinetic energy term cTF dx ρpxq5{3 . This term is strictly convex, thanks to
the strict convexity of the function s ÞÑ s5{3 , şfor s ě 0.
Consider the electron-nuclei interaction, dx ρpxqV pxq. Being linear in ρ, this term is
trivially convext.
Finally, consider the electron-electron interaction, Dpρ, ρq. We have:

Dpλρ1 ` p1 ´ λqρ2 , λρ1 ` p1 ´ λqρ2 q “ λ2 Dpρ1 , ρ1 q ` p1 ´ λq2 Dpρ2 , ρ2 q ` 2λp1 ´ λqDpρ1 , ρ2 q

(9.88)

109
By Lemma 9.5, Dpρ1 , ρ2 q ď Dpρ1 , ρ1 q1{2 Dpρ2 , ρ2 q1{2 ď p1{2qpDpρ1 , ρ1 q ` Dpρ2 , ρ2 qq. Hence,

Dpλρ1 ` p1 ´ λqρ2 , λρ1 ` p1 ´ λqρ2 q ď λ2 Dpρ1 , ρ1 q ` p1 ´ λq2 Dpρ2 , ρ2 q

´ ¯
`λp1 ´ λq Dpρ1 , ρ1 q ` Dpρ2 , ρ2 q
ď λDpρ1 , ρ1 q ` p1 ´ λqDpρ2 , ρ2 q . (9.89)

This proves convexity of the electron-electron interaction, and concludes the proof of con-
vexity of ETF pρq.

Uniqueness of the minimizer is an immediate consequence of strict convexity.

Corollary 9.7 (Uniqueness of the minimizer.). Let ρ1 , ρ2 be two minimizers of ETF pρq in
DN . Then, ρ1 “ ρ2 .

Proof. Suppose ρ1 ‰ ρ2 . By convexity of DN , ρλ “ λρ1 ` p1 ´ λqρ2 P DN , for λ P p0; 1q. By

convexity of the TF functional:
TF TF TF
ETF pρλ q ă λETF pρ1 q ` p1 ´ λqETF pρ2 q “ λEN ` p1 ´ λqEN ” EN . (9.90)
TF
But this is absurd, since EN is the smallest energy that can be reached in DN . Hence
ρ1 “ ρ2 .

9.4.4 Ionization in TF theory

In this section we shall investigate the behavior of the TF energy as a function of the
number of particles N . In particular, we would like to understand under which conditions
TF TF
the ionization energy is positive: EN `1 ă EN . As we shall see, the validity of this inequality
is related to whether the minimizer in DN is actually in FN . We will start from the following
lower bound on the TF energy, that improves on (9.70). The bound shows that the energy
cannot be arbitrarily negative as N increases.
Theorem 9.8. There exists a universal constant C ą 0 such that, for all ρ P DN :
K
´ÿ ¯1{3
5{3
ETF pρq ě ´CZtot Zj2 . (9.91)
j“1

The proof of this theorem is based on the following important result, see [3] for a proof.
Theorem 9.9 (Newton’s theorem). Let µ be a rotation invariant measure on R3 . Then:
ż ż ż
1 1 1
φpxq :“ µpdxq “ µpdxq ` µpdxq . (9.92)
R3 |x ´ y| |x| |y|ă|x| |y|ą|x| |y|

If one thinks of µ as describing a charge distribution, the function φpxq has the inter-
pretation of electric potential generated by µ. As a consequence, this theorem shows that
spherically symmetric charged objects are equivalent to pointlike charges. Another impor-
tant consequence of this result is that the electric potential generated by a uniformly charged
sphere is constant inside the sphere.

Proof. (of Theorem 9.8.) The N -dependence of the nonoptimal lower bound (9.70) came from
a naive control of the tail of the Coulomb attraction between the nuclei and the electrons.
Here, we will control the growth in N of this energetic contribution with the positive mutual
Coulomb repulsion of the electrons.
To begin, we write:
V pxq “ Vă pxq ` Vą pxq , (9.93)
where, for R ą 0 to be chosen later:
K
ÿ ! 1 1)
Vą pxq “ Zj min , . (9.94)
j“1
|x ´ Rj | R

110
The function Vą pxq captures the long range contribution to the electron-nuclei electro-
static interaction,
! )while Vă pxq takes into account the singularity. By Newton’s theorem,
1
Zj min |x´R , 1 is the electrostatic potential generated by a uniformly charged sphere,
j| R
centered in Rj , with radius R:
ż
! 1 1) 1 Zj
Zj min , “ µj pdxq , µj pxq “ δp|x ´ Rj | ´ Rq . (9.95)
|x ´ Rj | R |x ´ y| 4πR2

Therefore,
ż K
1 ÿ
Vą pxq “ µpdxq , µpxq “ µj pxq . (9.96)
|x ´ y| j“1

Hence: ż ż
1
dx Vą pxqρpxq “ dxdy µpdyqρpxq ” 2Dpµ, ρq , (9.97)
|x ´ y|
where, for two measures µ1 , µ2 , not necessarily absolutely continuous:
ż
1 1
Dpµ1 , µ2 q “ µ1 pdxqµ2 pdyq . (9.98)
2 |x ´ y|

We then rewrite the TF energy functional as:

ż
5{3
ETF pρq “ cTF }ρ}5{3 ´ dx Vă pxqρpxq ´ 2Dpµ, ρq ` Dpρ, ρq ` U
ż
5{3
“ cTF }ρ}5{3 ´ dx Vă pxqρpxq ` Dpρ ´ µ, ρ ´ µq ´ Dpµ, µq ` U . (9.99)

The next crucial remark is that Dpρ ´ µ, ρ ´ µq, the electrostatic interaction of the net charge
distribution ρ ´ µ, is positive: Dpρ ´ µ, ρ ´ µq ě 0. The proof of this fact follows again from
the representation of the Coulomb interaction as in Eq. (9.67). In fact, setting ν “ ρ ´ µ:
ż ż ż8 ż
1 1 3
Dpν, νq “ νpdxqνpdyq “ νpdxqνpdyq dt t dz ht px ´ zqht py ´ zq
2 |x ´ y| 0
ż8 ż
“ dt t3 dz pν ˚ ht qpzq2 ě 0 (9.100)
0

where in the last step we exchanged integrations thanks to Fubini’s theorem. Using this fact,
we can bound from below the TF energy as:
ż
5{3
ETF pρq ě cTF }ρ}5{3 ´ dx Vă pxqρpxq ´ Dpµ, µq ` U . (9.101)

Next, let us estimate the energetic contribution due to Vă . We have:

K
ÿ ´ 1 ! 1 1 )¯
Vă pxq “ V pxq ´ Vą pxq “ Zj ´ min ,
j“1
|x ´ Rj | |x ´ Rj | R
K
ÿ ´ 1 1¯
“ Zj ´ χp|x ´ Rj | ď Rq . (9.102)
j“1
|x ´ Rj | R

Therefore, by Hölder inequality:

ż K ż
ÿ 1
dx ρpxqVă pxq ď Zj dx ρpxqχp|x ´ Rj | ď Rq
j“1
|x ´ Rj |
ÿ › χp| ¨ | ď 1q › ÿ
Zj }ρ}5{3 › › ď CR1{5 Zj }ρ}5{3 . (9.103)
› ›
ď
j
|¨| 5{2
j

111
Finally, let us consider the Dpµ, µq term. We have:
ż
1 1
Dpµ, µq “ µpdxqµpdyq
2 |x ´ y|
ż
1
“ µpdxqVą pxq
2
ż K K
1 ÿ Zj ÿ ! 1 1)
“ dx δp|x ´ Rj | ´ Rq Z i min ,
2 j“1
4πR2 i“1
|x ´ Ri | R
ż
1 ÿ Zi Zj ! 1 1)
“ dx δp|x| ´ Rq min , . (9.104)
2 i,j 4πR2 |x ´ Ri ` Rj | R

Separating the i “ j terms from the i ‰ j terms:

ż
1ÿ 1 Zi Zj
Dpµ, µq “ dx δp|x| ´ Rq
2 i‰j |x ` Ri ´ Rj | 4πR2
1 Zi2
ż
1ÿ
` dx δp|x| ´ Rq
2 i |x| 4πR2
ÿ Zi Zj 1 ÿ Zi2 1 ÿ Zi2
ď ` ”U` . (9.105)
iăj
|Ri ´ Rj | 2 i R 2 i R

Eqs. (9.101), (9.103), (9.105) imply:

K
5{3 1 ÿ 2
ETF pρq ě cTF }ρ}5{3 ´ CZtot }ρ}5{3 R1{5 ´ Z . (9.106)
2R i“1 i

The final statement, Eq. (9.91), follows optimizing over R (that is, choosing the R ą 0 that
maximizes the right-hand side).

The next lemma is an immediate consequence of convexity and of the uniform lower
bound.
TF
Lemma 9.10. The TF ground state energy EN is convex, nonincreasing and bounded below.

Proof. Boundedness follows from Theorem 9.8. Let us prove convexity. Let ρ1 be the
minimizer in DN1 and ρ2 be the minimizer in DN2 . We have:
TF
EλN1 `p1´λqN2
ď ETF pλρ1 ` p1 ´ λqρ2 q (9.107)
TF TF
ď λETF pρ1 q ` p1 ´ λqETF pρ2 q “ λEN1
` p1 ´ λqEN2
,

which proves convexity. To prove that the energy is nonincreasing in N , we simply notice
that the set DN grows with N , hence DN can only decrease.

The previous result implies that the limit:

TF
lim EN “ E8 (9.108)
N Ñ8

exists. We define the critical number of particles Nc as:

! )
TF
Nc “ inf N | EN “ E8 . (9.109)

Notice that we do not know yet whether Nc ă 8. The next theorem characterizes the shape
TF
of EN as a function of N .
Theorem 9.11. For N ď Nc , there exists a unique minimizer on ETF in FN . The function
TF
EN is strictly convex and decreasing in r0, Nc s. If Nc ă 8 and N ą Nc , there is no
TF
minimizer in FN . The function ρNc is the unique minimizer in DN . Moreover, EN is
constant in rNc , 8q.

112
Proof. Let N ď Nc and let ρ˚ be the minimizer of ETF in DN . We claim that:
ż
dx ρ˚ pxq “ N . (9.110)
ş TF TF
Suppose that ρ˚ “ N 1 ă N . Then, ρ˚ P DN 1 , which implies that EN 1 “ EN . Since
TF TF 1 TF
EN is nonincreasing, E¨ is constant in rN , N s. Also, by convexity E¨ is constant for
all N 2 ě N 1 , which implies that N 1 ě Nc . This however contradicts N 1 ă N ď Nc . Hence,
N 1 “ N . The above argument also proves strict convexity.
Suppose now that Nc ă 8 and that N ą Nc . Suppose that there is a minimizer in FN .
Then, consider the trial state:
1´ ¯
ρNc ` ρN (9.111)
2
which has pN ` Nc q{2 particles. We have:
´1 ¯ ´1 ¯ 1´ ¯
TF
EN “ E pNc ` N q ď ETF pρN c
` ρ N q ă E TF pρ Nc
q ` E TF pρ N q
c
2 2 2
1 TF TF
“ pE ` EN q
2 Nc
TF
“ EN c
. (9.112)
TF
The first step follows from the (assumed) constant profile of EN for N ě Nc ; the second
from the variational principle; the third from strict convexity (since Nc ă N , ρNc ‰ ρN !);
TF
the fourth from the fact that ρN and ρNc are minimizers; and the last from the fact that EN
is constant for N ą Nc . This gives a contradiction, and shows that there is no minimizer in
FN for N ą Nc .

All we are left to do is to determine the value of Nc . To do this, we shall use that the
TF minimizer satisfies a self-consistent equation, called the TF equation.
Theorem 9.12. Let N ď Nc . Then, there exists µ ě 0 such that the unique minimizer
ρN P FN satisfies the equation:
2{3
´ 1 ¯ 5
γρN pxq “ V pxq ´ ˚ ρN pxq ´ µ , pγ “ cTF q . (9.113)
|¨| ` 3
For N “ Nc , µ “ 0.
Remark 9.13. Eq. (9.113) is called the Thomas-Fermi equation. One can actually prove
that solutions of the TF equation in L1 X L5{3 are minimizers of the TF functional. The
number of particles is determined by the chemical potential µ.

Proof. Let ρN be the minimizer in FN . For any δ ą 0, for any bounded function f such
that: ż
dx χpρN pxq ě δqf pxq “ 0 , (9.114)

define:
ρε “ ρN ` εf pxqχpρN pxq ě δq . (9.115)
For ε small enough (dependent of δ), ρε ě 0. Also, the assumption (9.114) implies that
}ρε }1 “ }ρN }1 “ N . Finally, since f is bounded and χpρN pxq ě δq is supported on a
bounded set, ρε P L5{3 . Hence, ρε P FN , and so:

ETF pρε q ě ETF pρN q . (9.116)

Together with differentiability in a neighbourhood of ε “ 0 (which can be easily proven),

this implies:
d
0“ ETF pρε q |ε“0 . (9.117)
dε
Writing explicitly the right-hand side, one has:
ż ´
5{3 1 ¯
0“ dx γρN pxq ` pρN ˚ qpxq ´ V pxq f pxq (9.118)
ρN ěδ |¨|

113
 ş
for all bounded f such that ρN pxqěδ
dx f pxq “ 0. The arbitrariness of f in this class of
functions implies that:

2{3 1
γρN pxq ` ˚ ρN pxq ´ V pxq “ ´µ , (9.119)
|¨|

for some constant µ, and for all x such that ρN pxq ě δ. Being ρN nonnegative, this implies:

2{3
´ 1 ¯
γρN pxq “ V pxq ´ ˚ ρN pxq ´ µ (9.120)
|¨| `

for all x such that ρN pxq ą δ. Taking the δ Ñ 0 limit, we found that ρN pxq satisfies the TF
equation for all x such that ρN pxq ą 0. Let us now explore the region tx | ρN pxq “ 0u. To
this end, consider the function:

ρε pxq “ ρN pxq ` εf pxq , (9.121)

ş
for ε ě 0, f P L5{3 X L1 , f ě 0 on tx | ρN pxq “ 0u and dx f pxq “ 0. The function ρε
belongs to FN . Hence, ETF pρε q ě ETF pρN q for ε ě 0; taking the right derivative, one gets:
ż ´
1 ¯
0ď γρN pxq2{3 ´ V pxq ` ˚ ρN pxq f pxq dx . (9.122)
|¨|
Next, we split the integral in the right-hand side as:
ż ż ż
dx p. . .q “ dx p. . .q ` dx p. . .q ; (9.123)
ρN pxq“0 ρN pxqą0

using the previous result for the region ρN pxq ą 0, we have:

ż ż
´ 1 ¯
0 ď dx γρN pxq2{3 ´ V pxq ` ˚ ρN pxq f pxq ´ µ dx f pxq
ρN pxq“0 |¨| ρN pxqą0
ż ´ 1 ¯
“ dx γρN pxq2{3 ´ V pxq ` ˚ ρN pxq ` µ f pxq (9.124)
ρN pxq“0 |¨|
ş ş
where in the last step we used that ρN pxq“0 dx f pxq “ ´ ρN pxqą0 dx f pxq. The arbitrariness
of f implies that:
1
´ µ ` V pxq ´ ρN pxq ď 0 on x s.t. ρN pxq “ 0. (9.125)
|¨|
Taking the positive part:
´ 1 ¯
0 “ ´ µ ` V pxq ´ ρN pxq on x s.t. ρN pxq “ 0. (9.126)
|¨| `

Equivalently,

2{3
´ 1 ¯
γρN pxq “ ´ µ ` V pxq ´ ρN pxq on x s.t. ρN pxq “ 0. (9.127)
|¨| `

Eqs. (9.127), (9.120) give the TF equation for all values of x. To conclude, let us comment
1
on the chemical potential µ. Since ρN pxq, V pxq and p |¨| ˚ ρN qpxq decay at infinity, the TF
equation implies that µ ě 0 (otherwise the right-hand side would be nonzero for |x| Ñ 8,
2{3
which would contradict decay for ρN ). Let us prove that for N “ Nc one as µ “ 0.
We repeat the trial state argument, with ρε “ ρNc ` εf . We only assume that ε ě 0,
f ě 0 and that f P L1 X L5{3 . In this way, the number of particles is not Nc ; this is
however not important, since ρNc is the density with the smallest energy. Hence, one has
ETF pρε q ě ETF pρNc q, for ε ě 0. Taking the right derivative and proceeding as above:
´ ¯
2{3
γρNc pxq ě V pxq ´ ρ ˚ V pxq ; (9.128)
`

114
by the TF equation:
´ ¯ ´ ¯
V pxq ´ ρ ˚ V pxq ´ µ ě V pxq ´ ρ ˚ V pxq . (9.129)
` `

Notice that the function V pxq´|¨|´1 ˚ρNc pxq has to be positive for some x, otherwise the TF
equation would prove that ρNc “ 0. The assumption ρNc P L1 X L5{3 implies that | ¨ | ˚ ρNc pxq
is bounded, and V pxq Ñ `8 in proximity of the nuclei; hence V pxq ´ | ¨ |´1 ˚ ρNc pxq is
positive in a neighbourhood of the nuclei. For these values of x, Eq. (9.129) implies that
µ “ 0. This concludes the proof of Theorem 9.12.

The function:
1
φpxq “ V pxq ´ ˚ ρN pxq , (9.130)
|¨|
is called the Thomas-Fermi potential. It describes the net electrostatic potential generated
by the nuclei plus the electrons. In terms of this function the TF equation reads:
´ ¯
2{3
γρN pxq “ φpxq ´ µ . (9.131)
`

Therefore, the TF minimizer is supported for the values of x such that φpxq ě µ. This is
certainly true if x is close enough to one of the nuclei, since V pxq Ñ 8 there; hence, the
TF equation is telling us that the electrons are localized close to the nuclei, as expected.
The next proposition is an important property of the TF functional, that will be crucial to
compute the critical number of particles Nc .
Proposition 9.14. Let N ď Nc . Then:
φpxq ě 0 . (9.132)
Proof. Let ∆ be the distributional Laplacian. Then:
1
∆ “ ´δp¨q , (9.133)
|¨|
with δp¨q the Dirac’s delta. Hence, away from the nuclei:
∆φpxq “ ρN pxq . (9.134)
Now, using the TF equation we can rewrite the density as a function of φ as:
´ ´ 1 ¯ ¯3{2
3{2
ρN “ γ ´1 V pxq ´ ˚ ρN pxq ´ µ ” γ ´3{2 pφpxq ´ µq` . (9.135)
|¨| `

Therefore, the TF equation is equivalent to the following PDE:

3{2
∆φpxq “ γ ´3{2 pφpxq ´ µq` . (9.136)
Let us consider the set:
A “ tx P R3 | φpxq ă 0u . (9.137)
Notice that the nuclei do not belong to such set, since V pxq Ñ 8 there and ρN ˚ | ¨ |´1 is
bounded. Away from the nuclei, the function φpxq is continuous, and hence A is open. By
continuity the function φpxq vanishes on the boundary of A. Notice that the set A need not be
compact; nevertheless, the function φpxq is also vanishing at infinity, since | ¨ |´1 ˚ ρN pxq Ñ 0
as |x| Ñ 8, see Chapter 2 of [3]. Hence, φpxq vanishes on the boundary of A, and at infinity.
In A, the function ´φ is harmonic:
∆p´φqpxq “ 0 for all x P A. (9.138)
By the maximum principle for harmonic functions [3, 1], the function ´φ reaches its max-
imum on the boundary of A (or at infinity, if A is unbouded). (Strictly speaking, the
maximum principle holds for open and connected domains; in case A is not connected, we
split A into connected components, and we repeat the argument in each component.)
Hence ´φpxq ď 0 for all x P A, which implies that φpxq “ 0 in A. That is, A is the empty
set. This proves that φpxq ě 0 for all x.

115
 To conclude the section, we compute the critical number of particles Nc .
Theorem 9.15. Nc “ Ztot .

Proof. The starting point is Newton’s theorem, for a uniformly distributed charge on a sphere
of radius r: ż
1 1 !1 1 )
2
dω “ min , . (9.139)
4πr |ω|“r |ω ´ y| r |y|
We then compute:
ż K ż ż
1 ÿ 1 1 1 ´ 1 ¯
dω φpωq “ Zj dω
´ dω ˚ ρ N pωq
4πr2 |ω|“r j“1
4πr2
|ω|“r |ω ´ Rj | 4πr2 |ω|“r |¨|
K !1 1 ) ż
ÿ 1 ! 1 1)
“ Zj min , ´ dy ρN pyq min ,
j“1
r |Rj | 4πr2 |y| r
ě 0. (9.140)

where the last inequality follows from Proposition 9.14. Taking r ě |Rj | for all j, we have:
ż
Ztot !1 1 )
´ dy min , ρN pyq (9.141)
r |y|ďr r |y|

which implies: ż
Ztot ě dy ρN pyq . (9.142)
|y|ďr

Taking the r Ñ 8 limit:

Ztot ě N ñ Ztot ě Nc . (9.143)
To conclude the proof, suppose now that Ztot ą Nc . Let ρNc be the minimizer with Nc
particles. By the TF equation:

2{3
´ 1 ¯
γρNc pxq “ V pxq ´ ˚ ρNc pxq
|¨| `
1
“ V pxq ´ ˚ ρNc pxq (9.144)
|¨|

where in the last step we used again the positivity of the TF potential. Averaging over a
sphere of radius r we get:
ż ż
γ 2{3 1
dω ρ Nc pωq “ dω φpωq . (9.145)
4πr2 |ω|“r 4πr2 |ω|“r

On one hand, by concavity of s ÞÑ s2{3 :

ż ´ 1 ż ¯2{3
1 2{3
2
dω ρNc pωq ď 2
dω ρNc pωq ; (9.146)
4πr |ω|“r 4πr |ω|“r

on the other hand, by Newton’s theorem, taking r ě |Rj | for all j:

ż ż ż
1 Ztot !1 1 ) Ztot 1 Ztot ´ Nc
dω φpωq “ ´ dy min , ρNc pyq ě ´ dy ρNc pyq “ .
4πr2 r r |y| r r r
(9.147)
All together, for r ě |Rj |:
´ 1 ż ¯2{3 Ztot ´ Nc C
2
dω ρ N c pωq ě ě (9.148)
4πr |ω|“r γr r

116
for some C ą 0, since by assumption Ztot ą Nc . Now, let rewrite the number of particles
Nc in spherical coordinates:
ż ż8 ż
2
Nc “ dx ρNc pxq “ dr r dω ρNc prωq
0 |ω|“1
ż8 ż
1
“ dr r2 2 dω ρNc pωq (9.149)
0 r |ω|“r

where in the last step we performed a change of variables. By the lower bound (9.148):
ż8
C
Nc ě dr r2 3{2 “ `8 , (9.150)
maxt|Rj |u r

which contradicts Nc ă Ztot . Hence, Nc “ Ztot .

9.4.5 Scaling properties of the TF energy

One of the advantages of TF theory is that it allows to obtain a very simple prediction for
the energy of neutral atoms (K “ 1, Z “ N ). As we shall see, this prediction becomes exact
as N Ñ 8.
Z2 1{3
Proposition 9.16. Let K “ 1. We have ρZ pxq “ γ 3 ρpZ x{γq, with ρ the minimizer of:
ż ż
3 5{3 1
eTF pρq “ dx ρpxq ´ dx ρpxq ` Dpρ, ρq (9.151)
5 |x|

on F1 . Oe has:
EZTF “ pe0 {γqZ 7{3 , (9.152)
with e0 the ground state energy of eTF pρq on F1 . Numerically, e0 » ´3.678.

Proof. For a given ` ą 0, let us define the corresponding function ρ as:

ρZ pxq “ Z`3 ρpZ 1{3 `xq . (9.153)

Notice that ρ P L1 X L5{3 , and }ρ}1 “ 1. To find the correct value of `, let us rewrite the
energy of ρZ as:
ż ż
3 1
EZTF “ γZ 5{3 `2 dx ρpxq5{3 ´ Z 2 ` dx ρpxq ` Z 2 `Dpρ, ρq . (9.154)
5 |x|

The final claim follows setting γZ 5{3 `2 “ Z 2 `, that is ` “ Z 1{3 {γ.

Remark 9.17. The above proposition shows that the TF density profile has amplitude OpZ 2 q,
and that it varies on scale Z ´1{3 . In other words, TF theory shows that the TF minimizer
is concentrated in a region of diameter OpZ ´1{3 q around the position of the nucleus. The
energy of a neutral atom takes a particularly simple form, Eq. (9.152); remarkably, in the
N » Z Ñ 8 limit, this prediction becomes quantitatively correct.

9.5 Stability of matter of the second kind via TF theory

9.5.1 The no-binding theorem
One of the limitations of TF theory is that it does not predict the existence of molecules:
this is the content of the no-binding theorem, due to Teller in ’62 and proven by Lieb and
Simon in ’77.
Suppose we are given a system of nuclei. Let us partition them in two sets, A and B.
We define the electrostatic potential of the nuclei in the set 7 “ A, B as:
ÿ Zj ż
1
V7 pxq “ “ dµ7 pyq . (9.155)
jP7
|x ´ R j | |x ´ y|

117
The TF energy of the system corresponding to the nuclei in 7 is:
ż ż
TF 5{3
ÿ Zk Zj
E7 pρq “ cTF dx ρpxq ´ dx V7 pxqρpxq ` Dpρ, ρq ` . (9.156)
kăj
|Ri ´ Rj |
k,jP7

TF TF TF
We denote by E7,N the corresponding ground state energy: E7,N “ inf ρPDN E7,N pρq.
řK
Theorem 9.18 (No binding theorem.). Suppose N ď Ztot “ j“1 Zj . Then:

TF TF TF
mintEA,N 1
` EB,N2
| N1 ` N2 “ N u ď EN . (9.157)
TF TF
We can think of EA,N 1
` EB,N 2
as the energy of the system after the sets A and B have
been pushed infinitely far away from each other, so that their mutual interaction is negligible.
Consider the configuration of N1 and N2 particles in the sets A and B such that the sum
of the two energies is minimal. The theorem is telling us that there is no energetic gain in
bringing the two systems close together. Of course, the argument can be iterated for A and
B separately, and so on.
The conclusion is that the energetic gain due to the formation of a molecule (a stable
system composed by more than one nucleus) is missed by TF theory. As we shall see later,
this limitation of TF theory will be used in a positive way, to give a very simple proof of
stability of matter of the second kind. But first, let us prove the no binding theorem.
A crucial role in the proof we shall present is due to the following lemma, due to Baxter.
Lemma 9.19. Let ρ ě 0, ρ P L1 pR3 q X Lp pR3 q, with p ą 3{2. There exists g such that
0 ď g ď ρ such that:
1
p ˚ gqpxq “ VA pxq if ρpxq ą gpxq ě 0. (9.158)
|¨|

Moreover,
1
p ˚ gqpxq ď VA pxq if ρpxq “ gpxq. (9.159)
|¨|
Proof. The proof is based on calculus of variations. Let us define the functional:
ż
µA pyq
Ipgq “ Dpg, gq ´ dxdy gpxq . (9.160)
|x ´ y|

The functional is well defined on Dρ “ tg | 0 ď g ď ρu, and it is bounded below. Let tgj u be
a minimizing sequence in Dρ . Then, }gj }p ď }ρ}p ď C for p ą 3{2. This means that there
exists a weakly convergent subsequence g P Dρ in Lp , that we shall still denote by gj with a
slight abuse of notation. We claim that:

lim inf Ipgj q ě Ipgq . (9.161)

j

This shows that g is a minimizer of I. The proof of (9.161) follows from lim inf j Dpgj , gj q ě
Dpg, gq, Eq. (9.83), and from limj Dpgj , µA q “ Dpg, µA q, Eq. (9.77).
To prove Eqs. (9.158), (9.159) we shall use a trial state argument. Let us first explore
the region x : 0 ă gpxq ă ρpxq. Consider:

gε pxq “ gpxq ` εf pxqχpδ ď gpxq ď ρpxq ´ δq , (9.162)

with f bounded and δ ą 0. Clearly, gε P Dρ for |ε| small enough. Hence,

Ipgε q ě Ipgq . (9.163)

Taking the derivative with respect to ε (it can be proven that the function is differentiable):
ż
d ´ 1 ¯
0“ Ipgε q |ε“0 ñ 0 “ dx f pxq p ˚ gqpxq ´ VA pxq . (9.164)
dε δďgpxqďρpxq´δ |¨|

118
By arbitrariness of f , taking the δ Ñ 0` limit:
1
p ˚ gqpxq “ VA pxq if 0 ă gpxq ă ρpxq. (9.165)
|¨|

Let us now explore the region gpxq “ ρpxq. Consider the trial state:

gε pxq “ gpxq ` εf pxqχpmaxtρpxq ´ δ, δu ď gpxq ď ρpxqq , (9.166)

for ε ď 0, f ě 0. Taking the left derivative:

ż ´ 1 ¯
0ě dx f pxq p ˚ gqpxq ´ VA pxq . (9.167)
maxtρpxq´δ,δuďgpxqďρpxq |¨|

Again by arbitrariness of f , for δ Ñ 0` :

1
p ˚ gqpxq ď VA pxq if gpxq “ ρpxq. (9.168)
|¨|

Finally, let us consider the region gpxq “ 0. Let us introduce the trial state:

gε pxq “ gpxq ` εf pxqχp0 ď gpxq ď mintδ, ρpxq ´ δuq , (9.169)

for ε ě 0, f ě 0. Taking the right derivative:

ż ´ 1 ¯
0ď dx f pxq p ˚ gqpxq ´ VA pxq (9.170)
0ďgpxqďmintδ,ρpxq´δu |¨|

which implies, as δ Ñ 0` :
1
p ˚ gqpxq ě VA pxq if gpxq “ 0. (9.171)
|¨|
1
We are left with excluding the case p |¨| ˚ gqpxq ą VA pxq. To this end, consider the set:
! 1 )
P “ x|p ˚ gqpxq ě VA pxq . (9.172)
|¨|

Clearly, P Ă tx | gpxq “ 0u. Notice that the points Rj , the center of the nuclei, do not
1
belong to P : this is due to the fact that VA pxq “ `8 there, and |¨| ˚ gqpxq is bounded. Away
from these points, the function
1
p ˚ gqpxq ´ VA pxq (9.173)
|¨|
1
is continuous. Hence, the set P is open, and p |¨| ˚ gqpxq ´ VA pxq “ 0 on BP . Moreover,
´ 1 ¯
∆x p ˚ gqpxq ´ VA pxq “ ´gpxq ` µA pxq “ 0 @x P P , (9.174)
|¨|
1
since x ‰ Rj and P Ă tx | gpxq “ 0u. Therefore, function p |¨| ˚ gqpxq ´ VA pxq is harmonic in
1
P . By the maximum principle, p |¨| ˚ gqpxq ´ VA pxq “ 0, which proves that P is empty.

We are now ready to prove the no-binding theorem, Theorem 9.18.

Proof. (of Theorem 9.18.) Let ρ ” ρN be the minimizer of E TF in FN , for N ď Ztot . To

prove the theorem, it is enough to show that there exists g, h such that 0 ď g, h ď ρ such
that g ` h “ ρ and:
TF TF
EA pgq ` EB phq ď E TF pρq . (9.175)
Consider the kinetic energies. Since pa ` bq5{3 ě a5{3 ` b5{3 for all a, b ě 0, we immediately
get: ż ż ż
dx gpxq5{3 ` dx hpxq5{3 ď dx pgpxq ` hpxqq5{3 . (9.176)

119
Let us now study the interaction. In view of Eq. (9.176), to conclude the proof it is enough
to show that:
ÿ Zj Zk
´2Dpg, µA q ` Dpg, gq `
kăj
|R j ´ Rk |
k,jPA
ÿ Zj Zk
´2Dph, µB q ` Dph, hq `
kăj
|Rj ´ Rk |
k,jPB
ÿ Zk Zj
ď ´2Dpg ` h, µA ` µB q ` Dpg ` h, g ` hq ` . (9.177)
kăj
|Rj ´ Rk |

The inequality (9.177) can be rewritten as:

ÿ Zk Zj
0 ď ´2Dpg, µB q ´ 2Dph, µA q ` 2Dpg, hq `
kăj
|Rk ´ Rj |
kPA,jPB
“ ´2Dpg, µB q ´ 2Dph, µA q ` 2Dpg, hq ` 2DpµA , µB q
“ 2Dpg ´ µA , h ´ µB q . (9.178)

Let now now choose g, ρ to be as in Lemma 9.19. We rewrite Dpg ´ µA , h ´ µB q as:

ż ´ 1 ¯
Dpg ´ µA , h ´ µB q “ dx phpxq ´ µB pxqq ˚ p ˚ gqpxq ´ VA pxq
x:gpxq“ρpxq |¨|
ż ´ 1 ¯
` dx phpxq ´ µB pxqq ˚ p ˚ gqpxq ´ VA pxq
x:gpxqăρpxq |¨|
ż ´ 1 ¯
” ´ dx µB pxq ˚ p ˚ gqpxq ´ VA pxq (9.179)
x:gpxq“ρpxq |¨|
1
where we used that hpxq “ 0 if gpxq “ ρpxq and that p |¨| ˚ gqpxq ´ VA pxq “ 0 if gpxq ă ρpxq,
1
by Lemma 9.19. Also, the same lemma implies that p |¨| ˚ gqpxq ´ VA pxq ď 0 for gpxq “ ρpxq,
hence Dpg ´µA , h´µB q ě 0, which proves Eq. (9.177). This concludes the proof of Theorem
9.18.

The next result is a simple corollary of the no-binding theorem, that will play a crucial
role in the proof of stability of matter of the second kind for the many-body problem.
Corollary 9.20. For any ρ P L1 X L5{3 , ρ ě 0, for any γ ą 0:
K
3.678 ÿ 7{3
ETF pρq ě ´ Z . (9.180)
γ j“1 j

Proof. Consider a collection of K nuclei, and separate one, say the one in R1 with charge
Z1 , from the rest: A “ tR1 u and B “ tRj uK j“2 . By the no-binding theorem, using that
TF TF
E7,N ě E7,Ztot
:
TF
ETF pρq ě EA,Z 1
` inftEZTF
2 ,...,Zk
pρq | ρ P FZ2 `...ZK u
3.678 7{3
“ ´ Z1 ` inftEZTF 2 ,...,Zk
pρq | ρ P FZ2 `...ZK u , (9.181)
γ
where in the last step we used Proposition 9.16. Iterating the argument K ´ 1 times, the
claim follows.

9.5.2 Proof of stability of matter

As we shall see in this section, the no binding theorem of Thomas-Fermi theory can be used
to proved stability of matter of the second kind for the full many-body probblem, described
by the Hamiltonian HN,K pZ, Rq on L2anti pR3N q. We shall prove the following theorem.

120
Theorem 9.21 (Stability of matter of the second kind.). There exists a constant CpZq ą 0
1
such that, for all ψN P Hanti pR3N q, }ψN }2 “ 1:

xψN , HN,K pZ, RqψN y ě ´CpZqpN ` Kq . (9.182)

This lower bound is compatible with the fact that the ground state energy of the initial
many-body problem grows linearly with the number of particles. If not, matter could not
be extensive (recall the discussion in Section 9.3): splitting the system into subsystems
could produce an enormous increase/decrease of the energy. The first proof of stability of
matter was given by Dyson and Lenard in ’67. Here we shall discuss the proof of Lieb and
Thirring ’77, much simpler than the original one, based on Thomas-Fermi theory and on the
Lieb-Thirring kinetic energy inequality.
Theorem 9.22 (LT kinetic energy inequality.). There exists K ą 0 such that for any
ψN P L2anti pR3N q:
N
ÿ ż
5
xψN , ´∆j ψN y ě K dx ρψ pxq 3 , (9.183)
j“1
ş
where ψψ pxq “ N dx2 . . . dxN |ψN px, x2 , . . . , xN q|2 is the density associated to ψN .
This inequality is a consequence of another important result in quantum mechanics, the
Lieb-Thirring inequality for sums of negative eigenvalues. Let H “ ´∆ ` V on L2 pRd q be
a self-adjoint Schrödinger operator, with V P L1`d{2 pRd q, and V satisfying the assumptions
in Section 7.4 needed in order to define the eigenvalues with the min-max principles. Let Ej
be the eigenvalues of H (which can be defined as in Section 7.4). Then, the Lieb-Thirring
inequality states that: ż
ÿ
|Ej | ď Ld dx V pxq1`d{2 , (9.184)
j:Ej ď0

for some explicit Ld ą 0 (see [3] for generalizations). This inequality is compatible with
the semiclassical approximation for the sum of negative eigenvalues, recall the discussion of
Section 8.5. Let us show how Eq. (9.184) implies the kinetic energy inequality (9.183).
p1q
Proof. (of Theorem 9.22) Let γψ be the reduced one-particle density matrix of ψN , and
consider a class
ř of Schrödinger operators H “ ´∆ ` V with V such that (9.184) holds true.
Let HN “ i H piq . By the definition of density matrix:
p1q
xψN , HN ψN y “ Tr Hγψ . (9.185)

p1q p1q ř
Being γψ a nonnegative, trace-class operator, it can be written as γψ “ λj |fj yxfj |, for
j
2 d
some orthonormal fj P L pR q and 0 ď λj ď 1, due to the fact that 0 ď
p1q
1
γψ ď , recall
Section 9.2. Therefore:
ÿ 8
ÿ
xψN , HN ψN y “ λj Tr HPj “ λj xfj , Hfj y . (9.186)
j j“1

Clearly, xf0 , Hf0 y ě E0 , the ground state of H. Hence:

8
ÿ
xψN , HN ψN y ě E0 ` λj xfj , Hfj y . (9.187)
j“2

Being f1 orthogonal to f0 , xf1 , Hf1 y ě E1 , the first excited state of H. The argument can
be iterated for all negative eigenvalues; we have:

ÿ ˚
ÿ
xψN , HN ψN y ě Ej ` λj xfj , Hfj y (9.188)
j:Ej ď0 j

121
where the asterisk denotes that fj are orthogonal to all eigenfunctions φi of the negative
eigenvalues. Therefore, by the variational characterization of eigenvalues, xfj , Hfj y ě 0,
which gives: ż
ÿ d
xψN , HN ψN y ě Ej ě ´C dx V´ pxq1` 2 , (9.189)
j:Ej ď0

where the last step follows from Eq. (9.184). Now, let us choose:
2
V pxq “ ´cρψ pxq d . (9.190)

Since ρψ P L1 , V pxq P Ld{2 . For this choice of V , Eq. (9.189) implies:

N
ÿ ż N
ÿ
xψN , ´∆j ψN y ě ´Cc1`2{d dx ρψ pxq1`2{d ´ xψN , V pxj qψN y
j“1 j“1
ż ż
“ ´Cc1`2{d dx ρψ pxq1`2{d ` c dx ρψ pxq1`2{d . (9.191)

The claim follows after optimizing over c ą 0.

We are now ready to prove Theorem 9.21.

Proof. By the Lieb-Thirring kinetic energy inequality:

ż K ż
ÿ 1 ÿ 1
xψN , HN,K ψN y ě K dx ψψ pxq5{3 ´ Zj dx ρψ pxq ` xψN , ψN y ` U .
j“1
|x ´ Rj | iăj
|xi ´ xj |
(9.192)
Let us now find a useful lower bound for the many-body interaction, in terms of the TF
interaction. Let us apply Corollary 9.20 for Zj “ 1, Rj “ xj and K “ N . In this setting,
1
ř
U “ iăj |xi ´x j|
. The bound (9.180) implies the following lower bound:

N ż N ż
ÿ 1 3.678 3 ÿ 1
ě´ N´ γ dx ρpxq5{3 ` dx ρpxq ´ Dpρ, ρq (9.193)
iăj
|xi ´ xj | γ 5 j“1
|x ´ xj |

for any γ ą 0 and for any ρ P L1 X L5{3 , ρ ě 0. Choose ρ ” ρψN , Then, plugging the bound
(9.193) in (9.192) we get:
ż K ż
3 ÿ 1
xψN , HN,K ψN y ě pK ´ γq dx ρψ pxq5{3 ´ Zj dx ρψ pxq
5 j“1
|x ´ Rj |
N ż
ÿ 1 3.678
` dx ρpxqxψN , ψN y ´ Dpρψ , ρψ q ` U ´ N
j“1
|x ´ xj | γ
ż K ż
3 5{3
ÿ 1
” pK ´ γq dx ρψ pxq ´ Zj dx ρψ pxq ` Dpρψ , ρψ q ` U
5 j“1
|x ´ Rj |
3.678
´ N. (9.194)
γ

The first line reconstructs the TF energy functional, with a new constant cTF (positive,
choosing the old γ small enough). Therefore, the final claim follows from Corollary 9.20.

9.6 TF theory as the N Ñ 8 limit of quantum mechanics

In this section we shall give a rigorous derivation of TF theory starting from many-body
quantum mechanics. As we shall see, TF theory becomes exact once the number of particles
goes to infinity. We shall prove a theorem that provides a rigorous bound for the energy
difference of large quantum systems and the TF approximation.

122
 The result will hold in a suitable scaling regime, that we shall describe here. Let N 0 P N,
and let QN “ N {N 0 . Also, let Z 0 P NK , and let Z “ QN Z 0 . Finally, let R0 P R3K , and let
´1{3
R “ QN R0 . It is not difficult to see that:
TF TF 7{3
0 0
EN pZ, Rq “ QN EN 0 pZ , R q . (9.195)
7{3
That is, as N Ñ 8, every contribution to the TF ground state energy scales as QN . We
shall suppose that N 0 ď Ztot
0
, to make sure that the TF minimizer has N0 particles. We will
prove the following theorem.
Theorem 9.23. Let γ “ p6π 2 q2{3 . Then, there exists δ ą 0 independent of N such that:
Q 7{3
|EN TF
pZ, Rq ´ QN EN0
pZ 0 , R0 q| ď CN 7{3´δ . (9.196)

Therefore, the theorem proves that Thomas-Fermi theory becomes exact in the N Ñ 8,
for the ground state energy:
Q
EN pZ, Rq
lim 7{3
“1. (9.197)
N Ñ8 TF pZ 0 , R0 q
QN EN0

The proof of the theorem will be based on matching upper and lower bounds for the ground
state energy.

9.6.1 Upper bound

Let us start by proving an upper bound for the ground state energy, that gives the TF energy
at leading order in N . To do this, we shall use Hartree-Fock theory.
Consider a Slater determinant:
1 ÿ
ψN “ ? sgnpπqfπp1q px1 q ¨ ¨ ¨ fπpN q pxN q , (9.198)
N! π

with reduced one-particle density matrix given by:

N
ÿ
ω“ |fj yxfj | . (9.199)
j“1

As we have seen in Section 9.2, the energy of a fermionic wave function is expressed in terms
of the one- and two-particle density matrices. For the special case of Slater determinants, it
turns out that the energy is a functional of the one-particle density matrix only. Consider a
many-body Hamiltonian of the form:
ÿ ÿ
HN “ hi ` V p xi ´ xj q . (9.200)
i iăj

Then:
xψN , HN ψN y “ EHF pωq , (9.201)
where EHF is the Hartree-Fock energy functional:
ż
1
EHF pωq “ Trhω ` dxdy V px ´ yqpωpx; xqωpy; yq ´ |ωpx; yq|2 q . (9.202)
2
The last term describes the many-body interaction in HF theory. It is given by a sum of two
terms: the first is called the direct term, the second is called the exchange term. Notice that
for a positive potential, the exchange term is negative.
We define the HF ground state energy as the smallest energy of a Slater determinant.
Due to the one-to-one correspondence between Slater determinants and rank´N orthogonal
projections, the HF ground state energy is:
HF
EN “ inf EHF pωq , (9.203)
ωPPN

123
where PN :“ tω : L2 pR3 q Ñ L2 pR3 q | ω 2 “ ω ˚ “ ω, Trω “ N u. Since Slater determiants
form a subset of L2a pR3N q, we trivially have:

Q xψN , HN ψy HF
EN “ inf ď EN . (9.204)
ψN PL2a pR3N q xψN , ψN y

The idea will be to come up with a good trial state for the HF energy functional, that
TF
reproduces EN at leading order in N . To do that, we shall rely on the next Theorem.
Theorem 9.24. Suppose that V ě 0. Let K : L2 pR3 q Ñ L2 pR3 q be an admissible one-
particle density matrix: 0 ď K ď 1, Tr K “ N . Then, there exists a Slater determinant ψN
such that:
xψN , HN ψy ď EHF pKq . (9.205)
This theorem immediately implies the Lieb’s variational principle:
TF
EN “ inf EHF pKq . (9.206)
K admissible

Hence, it gives us the freedom to look for a larger class of trial states:
Q HF
EN ď EN ď EHF pKq for any admissible K. (9.207)

Proof. (of Theorem 9.24.) We shall prove the theorem in the case K is finite rank: K “
řM
i“1 λi |fi yxfi |, with 0 ď λi ď 1. The general case follows from an approximation argument,
that we leave as an exercise. Define:

hpkq “ xfk , hfk y , V pk`q “ xfk ^ f` , V fk ^ f` y , (9.208)

where fk ^ f` “ ?1 pfk b f` ´ f` b fk q. Then, a simple computation gives:

2

ÿ 1ÿ
EHF pKq “ λk hpkq ` λk λ` V pk`q . (9.209)
k
2 k,`

Suppose M ą N . If not, M “ N and there is nothing to prove. Then, there exists at least
two eigenvalues λp and λq such that 0 ă λp , λq ă 1. Without loss of generaily, we assume
that: ÿ ÿ
hpqq ` λk V pkqq ď hppq ` λk V pkpq . (9.210)
k k

Let δ “ mintλp , 1 ´ λ1 u. Clearly, δ ą 0. Define:

ÿ
K “ λk |fk yxfk | ` pλp ´ δq|fp yxfp | ` pλq ` δq|fp yxfp |
kRtp,qu
ÿ
” λk |fk yxfk | . (9.211)
k

Obviously, K is admissible. Notice that if δ “ λp then λp “ 0, and if δ “ 1 ´ λq then λq “ 1.

Hence, the number of eigenvalues of K that are neither 0 or 1 is strictly smaller than the
same quantity for K. After iterating the above procedure at most M times, we will be left
with a density matrix with eigenvalues equal to either 0 or 1, i.e. the reduced one-particle
density matrix of a Slater determinant. To conclude the proof, we have to make sure that
the energy does not increase in the process:

EHF pKq ď EHF pKq . (9.212)

To do this, we compute, recalling Eq. (9.209):

´ ÿ ÿ ¯
EHF pKq ´ EHF pKq “ δ hppq ` V pp`q λ` ´ hpqq ´ V pq`q λ` ´ δ 2 Vqp , (9.213)
` `

where the term in parenthesis is ě 0, by Eq. (9.210), and Vqp ě 0, as a consequence of

V ě 0. This implies Eq. (9.212), and concludes the proof.

124
 Legitimated by Lieb’s variational principle, we will choose a trial state K for the HF
functional that is not a projetcion. Since TF theory is a semiclassical approximation of
quantum mechanics, it makes sense to use coherent states. We define:
ż
1
K“ dpdq M pp, qqπpq , πpq “ |Fpq yxFpq | , Fpq pxq “ eipx Gpx ´ qq . (9.214)
p2πq3
1
ş
We shall assume that 0 ď M pp, qq ď 1, and that p2πq 3 dpdq M pp, qq “ N . This implies that
K is admissible. The idea is to try to choose M pp, qq so to make the energy as small as
possible. A good ansatz is:

M pp, qq “ χpp2 ´ φTF pqq ď ´µq , (9.215)

with φTF the Thomas-Fermi potential and µ ě 0 chosen so that K is admissible. Using the
TF equation, Eq. (9.215) can be rewritten as:

M pp, qq “ χpp2 ď pφTF pqq ´ µq` q “ χpp2 ď γρpqq2{3 q , (9.216)

with ρ the N -particle TF minimizer. Notice that:

4π γ 3{2
ż ż
1
TrK “ dpdq M pp, qq “ dq ρpqq “ N , (9.217)
p2πq3 p2πq3 3

provided γ “ p6π 2 q2{3 . Also:

ż
1
Tr ´ ∆K “ dqdp M pp, qqpp2 ` }∇G}22 q
p2πq3
ż
3
“ γ dq ρpqq5{3 ` N }∇G}22 . (9.218)
5
Hence, we get, using that the exchange term is nonpositive:
ż ż ż
3 1 1
EHF pKq ď γ dx ρpxq5{3 ´ dx V pxqKpx, xq ` dxdy Kpx; xqKpy; yq
5 2 |x ´ y|
`U ` CN }∇G}22 . (9.219)

The first line looks very much like the TF functional, except that Kpx, xq appears instead
of ρpxq. We compute:
ż ż
1
Kpx; xq “ dpdq M pp, qq|Gpx ´ qq| “ dq ρpqq|Gpx ´ qq|2 ” pρ ˚ |G|2 qpxq . (9.220)
2
p2πq3

Consider the density-density interaction. We rewrite it as:

ż
1 1
Dpρ ˚ |G|2 , ρ ˚ |G|2 q “ dx pρ ˚ |G|2 qpxqpρ ˚ |G|2 ˚ qpxq , (9.221)
2 |¨|

that is as a L2 scalar product. Rewriting it in Fourier space, and using that the Fourier
transform of the convolution is the product of the Fourier transforms:
ż
1 y2 ppq|2 1 ď Dpρ, ρq ,
Dpρ ˚ |G|2 , ρ ˚ |G|2 q “ dp |ρ̂ppq|2 ||G| (9.222)
2 p2
where we used that:
||G| y2 |}8 ď }|G|2 }1 “ 1 .
y2 ppq| ď }||G| (9.223)
Thus, the bound (9.222) reproduces the TF density-density interaction. To conclude, con-
sider the electron-nuclei interaction. We rewrite it as:
ż ż ż
´ dx V pxqKpx; xq “ ´ dx V pxqρpxq ` dx V pxqpρpxq ´ Kpx; xqq . (9.224)

125
The last term is an error term, that we have to estimate. We rewrite it as:
ż K ż
ÿ ´ 1 1 ¯
dx V pxqpρpxq ´ Kpx; xqq “ Zj dx ρpxq ´ |G|2 ˚ pxq . (9.225)
j“1
|x ´ Rj | | ¨ ´Rj |

Let us suppose that:

Gpxq “ R´3{2 G0 px{Rq , G0 pxq ” G0 p|x|q , (9.226)

with G0 compactly supported for |x| ď 1 and }G0 }2 “ 1, }∇G0 }2 ď C. With this choice,

}∇G}22 “ R´2 }∇G0 }22 . (9.227)

Thus, |Gpxq|2 is a spherically symmetric charge distribution, with total charge 1: by Newton’s
theorem,
1 1
|G|2 ˚ pxq “ , for |x ´ Rj | ą R. (9.228)
| ¨ ´Rj | |x ´ Rj |
Hence, we can rewrite Eq. (9.225) as:
ż K ż
ÿ ´ 1 1 ¯
dx V pxqpρpxq ´ Kpx; xqq “ Zj dx ρpxq ´ |G|2 ˚ pxq
j“1 |x´Rj |ďR |x ´ Rj | | ¨ ´Rj |
K ż
ÿ 1
ď Zj dx ρpxq . (9.229)
j“1 |x´Rj |ďR |x ´ Rj |

Now, from the TF equation for the minimizer:

K
´ÿ Zj 1 ¯
γρpxq2{3 “ ´ρ˚ pxq ´ µ
j“1
|x ´ Rj | |¨| `

K
ÿ Zj
ď . (9.230)
j“1
|x ´ Rj |

Assuming that R ! |Ri ´ Rj | for i ‰ j, we get:

3{2
Zmax
ρpxq ď C . (9.231)
|x ´ Rj |3{2

Hence, plugging this bound in Eq. (9.229) we get:

ż ż
5{2 1
dx V pxqpρpxq ´ Kpx; xqq ď CZmax dx ď CN 5{2 R1{2 . (9.232)
|x|ďR |x|5{2

All in all, plugging the bounds (9.222), (9.232) in Eq. (9.219) we find:
TF
EHF pKq ď EN ` CN R´2 ` CN 5{2 R1{2 . (9.233)

The optimal value of R is R “ CN ´3{5 , which is indeed such that R ! |Ri ´ Rj | for i ‰ j
(recall that, by assumption, the internuclear distance is order N ´1{3 ). With this choice:
Q TF
EN ď EHF pKq ď EN ` CN 11{5 . (9.234)
TF
Since EN „ N 7{3 , the error term is subleading as N Ñ 8. This concludes the proof of the
upper bound.

9.6.2 Lower bound

To conclude the proof of Theorem 9.23, we need a lower bound on the ground state energy
that agrees with TF at leading order.

126
A Properties of Sobolev spaces
Here we shall collect some basic properties of Sobolev spaces. We refer the reader to [3, 1]
for more details. Given an open set U Ď Rd , recall the definition of Sobolev space W k,p pU q:

W k,p pU q “ tu : U Ñ C | B α u P Lp pU q for all α s.t. |α| ď ku . (A.1)

The norm } ¨ }W k,p pU q is defined as:

ÿ
}u}pW k,p pU q “ }B α u}pLp pU q . (A.2)
α:|α|ďk

A special role in quantum mechanics is played by H 1 pU q :“ W 1,2 pU q. More generally,

we define H k pU q “ W k,2 pU q. We shall also denote by W0k,p pU q the space of functions in
W k,p pU q which are vanishing on the boundary of U .

A.1 Sobolev inequality

The Sobolev inequality allows to bound from below Lp norms of Du with Lq norms of u.
As we shall see, this cannot be true for all p, q. Let u P Cc8 pRd q, u ‰ 0, uλ pxq :“ upλxq.
Suppose that there exists C ą 0, independent of λ, such that:

}uλ }Lq pRd q ď C}∇uλ }Lp pRd q . (A.3)

By a change of variables,
ˆż ˙ q1 ˙1 ˆ
q 1 q
}uλ }Lq pRd q “ dx|uλ pxq| “ }u}Lq pRd q ,
λd
ˆż ˙ p1 ˆ ˙ (A.4)
p λ
}∇uλ }Lp pRd q “ dx|∇uλ pxq| “ d }∇u}Lp pRd q .
λp
Therefore, Eq. (A.3) implies:
d d
}u}Lq pRd q ď Cλ1´ p ` q }∇u}Lp pRd q . (A.5)

Thus, if 1´ dp ` dq ‰ 0, by taking either λ Ñ 0 or λ Ñ 8, Eq. (A.5) would imply }u}Lp pRd q ď 0,

that is u “ 0, which is a contradiction.
Therefore, we might only hope to prove Eq. (A.3) for:
d d 1 1 1
1` ´ “0ñ “ ´ . (A.6)
q p q p d
Let 1 ď p ă d. We define the Sobolev conjugate of p as the number q ” p˚ for which Eq.
(A.6) holds true:
dp
p˚ :“ . (A.7)
d´p
Notice that p˚ ą p.
Theorem A.1 (Gagliardo-Nirenberg-Sobolev inequality). Let 1 ď p ă d. There exists a
constant C ” Cpd, pq such that

}u}Lp˚ pRd q ď C}∇u}Lp pRd q , @u P Cc1 pRd q. (A.8)

The proof will be based on the generalized Hölder inequality:

ż ˇźm m m
ˇ ź ÿ 1
ui ˇdx ď }ui }Lpi pU q , “ 1. (A.9)
ˇ ˇ
ˇ
U i“1 i“1
p
i“1 i

Remark A.2. The proof crucially relies on the fact that u is compactly supported: the
inequality is trivially false if u “ 1. However, the constant C does not depend on the support
of u.

127
Proof. Let us start with the case p “ 1. Using the compact support of u,
ż xi
upxq “ dyi uxi px1 , ¨ ¨ ¨ , xi´1 , yi , xi`1 , ¨ ¨ ¨ , xd q (A.10)
´8

thus ż8
|upxq| ď dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ , xd q|. (A.11)
´8
d
For p “ 1 the Sobolev conjugate of p is p˚ “ d´1 . Therefore, it is natural to consider:
1
d ˆż 8 ˙ d´1
d ź
|upxq| d´1 ď dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ , xd q| . (A.12)
i“1 ´8

We have:
1
ż8 ż8 d ˆż 8 ˙ d´1
d ź
dx1 |upxq| d´1 ď dx1 dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q| “ (A.13)
´8 ´8 i“1 ´8
1 1
ż8 ˆż 8 ˙ d´1 n ˆż 8
ź ˙ d´1
“ dx1 dy1 |∇upy1 , ¨ ¨ ¨ q| dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q|
´8 ´8 i“2 ´8
1 1
ˆż 8 ˙ d´1 ż8 d ˆż 8
ź ˙ d´1
“ dy1 |Dupy1 , ¨ ¨ ¨ q| dx1 dyi |Dupx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q| .
´8 ´8 i“2 ´8

Let us now apply the generalized Hölder inequality, with pi “ d ´ 1. We have

1 1
ż8 ˆż 8 ˙ d´1 n ˆż 8 ˙ d´1
d ź
dx1 |upxq| d´1 ď dy1 |∇upy1 , ¨ ¨ ¨ q| dx1 dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q| .
´8 ´8 i“2 ´8
(A.14)

Next, let us integrate over x2 . We get:

ż8
d
dx1 dx2 |upxq| d´1
´8
1 1
ż ˆż 8 ˙ d´1 d ˆż 8
ź ˙ d´1
ď dx2 dy1 |∇upy1 , ¨ ¨ ¨ q| dx1 dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q|
´8 i“2 ´8
ˆż 8 1
˙ d´1 ż ˆż 8 1
˙ d´1 (A.15)
“ dx1 dy2 |∇upx1 , y2 ¨ ¨ ¨ q| dx2 dy1 |∇upy1 , x2 , ¨ ¨ ¨ q| ¨
´8 ´8
1
d ˆż 8
ź ˙ d´1
¨ dx1 dyi |∇upx1 , x2 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ q| .
i“3 ´8

Using again the generalized Hölder inequality for the x2 integration, choosing pi “ d ´ 1, we
have
ˆż 8 2
˙ d´1
(A.15) ď dx1 dx2 |∇upx1 , x2 ¨ ¨ ¨ q| ¨
´8
n ˆż 8 ż8 1
˙ d´1 (A.16)
ź
¨ dx2 dx1 dyi |∇upx1 , ¨ ¨ ¨ , yi , ¨ ¨ ¨ , xd q| .
i“3 ´8 ´8

Iterating the same procedure n times (i.e. integrating again over dx3 , ¨ ¨ ¨ , dxd ) we finally
get
ż ˆż d
˙ d´1
d
dx1 ¨ ¨ ¨ dxd |upxq| d´1 ď dx1 ¨ ¨ ¨ dxd |∇upx1 , ¨ ¨ ¨ , xd q| , (A.17)

which proves the inequality for p “ 1.

128
 Let us now consider 1 ă p ă d. Let v :“ |u|γ , γ ą 1 to be chosen later. By Eq. (A.17),
we have
ˆż ˙ d´1
d
ż ż ˆż ˙ p´1
p
ˆż ˙ p1
γd p
|u| d´1 dx ď |∇|u|γ |dx “ γ |u|γ´1 |∇u|dx ď γ |u|pγ´1q p´1 dx |∇u|p ,
(A.18)
where in the last step we used the Hölder with q “ p{pp ´ 1q. Now let us choose γ such that
γd p
d´1 “ pγ ´ 1q p´1 . That is,
ˆ ˙ ˆ ˙
d p p p´d p
γ ´ “´ ñγ “´ , (A.19)
d´1 p´1 p´1 pd ´ 1qpp ´ 1q p´1
i.e. γ “ ppd ´ 1q{pd ´ pq ą 1. Plugging this choice into Eq. (A.18) we get:
ˆż ˙ d´1
d
ˆż ˙´ p´1
p
ˆż ˙ p1
γd γd
p
|u| d´1 |u| d´1 ďγ |Du| , (A.20)

with
d´1 p´1 ppd ´ 1q ´ dpp ´ 1q d´p 1
´ “ “ ” ˚, (A.21)
d p dp dp p
and hence
γd pd
“ “ p˚ . (A.22)
d´1 d´p
We conclude that:
ˆż ˙ 1 ˆż ˙ p1
p˚
p˚ p
dx|u| ďγ dx|∇u| , 1ăpăd, (A.23)

which is what we wanted to prove.

This inequality can be used to prove that, in some cases, Sobolev spaces are embedded in
Lq spaces.
Theorem A.3 (Sobolev embedding). Let U Ă Rd open and bounded. Let u P W01,p pU q,
1 ď p ă d. Then
}u}Lq pU q ď C}∇u}Lp pU q , @q P r1, p˚ s, (A.24)
with C ” Cpp, q, U q.
Remark A.4. i) In particular, q “ p is allowed, since p˚ ą p. We have:
}u}Lp pU q ď C}∇u}Lp pU q , (A.25)
which takes the name of Poincaré inequality.
ii) The Poincaré inequality allows us to prove that on H0p pU q, the norms }∇u}Lp pU q and
}u}H p pU q are equivalent. In fact, one trivially has:
}∇u}Lp pU q ď }u}H p pU q (A.26)
and, by Poincaré inequality:
´ ¯ p1
}u}H p pU q ď }u}pLp pU q ` }∇u}pLp pU q ď C}∇u}Lp pU q . (A.27)

iii) Theorem A.3 is telling us that

u P H0p pU q ñ u P Lq pU q, @q P r1, p˚ s. (A.28)
˚
We stress that the smallest such Lq pU q space is Lp pU q. Indeed, by Hölder:
ˆż ˙ q1 ˆż 1 ˆż
˙ qp ˙ 11
p
q qp
}u}q “ dx|upxq| ď dx|upxq| dx ď C}u}p˚ , (A.29)
U U U

p˚ ˚
where 1 1
p1 ` p “ 1 and p “ q ą 1. We say that the space H0p pU q is embedded in Lp pU q,
dp
p˚ “ d´p , 1 ď p ă n.

129
 iv) Finally, the Sobolev inequality (A.24) can be extended to functions in H p pU q, under
the assumption that the boundary of U is of class C 1 .

Proof. Let u P W01,p pU q. Then, there exists tum umPN , um P Cc8 pU q such that um Ñ u in
W 1,p pU q. Let us extend um to Rd , setting um “ 0 on Rd zU . By the GNS inequality,

}um ´ ul }p˚ ď C}Dpum ´ ul q}p Ñ 0 as m, l Ñ 8. (A.30)

˚ ˚
Thus, tum u is a Cauchy sequence in Lp pU q, and hence um Ñ ũ in Lp pU q. Being U
bounded, ũ P Lq pU q, @q : 1 ď q ď p˚ . In particular, ũ P Lp pU q, which shows that u “ ũ, and
therefore that u P Lq for all q P r1, p˚ s.
By the GNS inequality:

}um }Lp˚ pU q ď C}∇um }Lp pU q . (A.31)

Then, by convergence in W 1,p pU q:

}∇um }Lp pU q “ }∇pum ´ u ` uq}Lp pU q Ñ }∇u}Lp pU q as m Ñ 8. (A.32)

Also,
}um }Lp˚ pU q ě C}um }Lq pU q , @1 ď q ď p˚ , (A.33)
and
}um }Lq pU q “ }um ´ u ` u}Lq pU q Ñ }u}Lq pU q as m Ñ 8. (A.34)
˚
by convergence in Lp pU q. All in all:

}u}Lq pU q ď C}∇u}Lp pU q , (A.35)

for some C ” CpU, d, pq.

A.2 Weak to strong convergence

Here we shall use the result of the previous section to prove the following result.
Theorem A.5 (Weak to strong convergence). Suppose that ψj Ñ ψ weakly in H 1 pRd q. Let
A Ă Rd be any set of finite measure and let χA be its characteristic function. Then:

χA ψj Ñ χA ψ strongly in Lp pRd q (A.36)

for every 1 ď p ă 2d{pd ´ 2q when d ě 3, every p ă 8 when d “ 2 and every p ď 8 when

d “ 1. In fact, for d “ 1 the convergence is pointwise and uniform.

Proof. The proof is based on an approximation argument via the heat kernel e∆t . We claim
that for any ψ P H 1 pRd q: ?
}ψ ´ e∆t ψ}2 ď }∇ψ}2 t , (A.37)
where: ż
∆t 1
pe ψqpxq “ expt´|x ´ y|2 {4tuψpyqdy . (A.38)
p4πtqd{2 Rd

The estimate (A.37) follows from Plancherel theorem:

ż
∆t 2
}ψ ´ e ψ}2 “ |ψ̂pkq|2 p1 ´ expt´|k|2 tuq2 dk
Rd
ż
ď |ψ̂pkq|2 |k|2 t ” }∇ψ}22 t , (A.39)
Rd

where we used that p1 ´ expt´|k|2 tu ď mint1, |k|2 tu. By the uniform boundedness principle,
see Theorem 2.12 of [3], the weak convergence ψj Ñ ψ in H 1 pRd q implies that }∇ψj }2 ď C
uniformly in j. Therefore, ?
}ψj ´ e∆t ψj }2 ď tC . (A.40)

130
Now, let φj “ e∆t ψj . Assuming for the moment that φj converges strongly in L2 pRd q to
φ “ e∆t ψ, we shall prove that χA ψ j converges strongly to χA ψ. Simply note that:

}χA pψj ´ ψq}2 ď }χA pψj ´ φj q}2 ` }χA pφj ´ φq}2 ` }χA pφ ´ ψq}2 . (A.41)

The first and the last term can be bounded using that:
?
}χA ψj ´ φj }2 ď }ψj ´ φj }2 ď t}∇ψj }2
?
}χA pφ ´ ψq}2 ď }φ ´ ψ}2 ď t}∇ψ}2 (A.42)

Again by the uniform boundedness principle, }∇ψj }2 ď C. Also, by the lower semicontinuity
of norms, Theorem 2.11 of [3],

}∇ψ}2 ď lim inf }∇ψj } ď C (A.43)

jÑ8
?
we see that the sum of the first and of the last term in Eq. (A.41) is estimated by 2C t.
Therefore, under the assumption that φj converges strongly to φ, for any ε ą 0 we can find
t and j such that: ?
}χA pψj ´ ψq}2 ď 2C t ` }χA pφj ´ φq}2 ď ε . (A.44)
This proves the claim Eq. (A.36) for p “ 2. It remains to prove that χA φj Ñ χA φ strongly
in L2 pRd q. To see this, note that by (A.38) and Hölder inequality:
´ż ¯1{2
χA pxq|φj pxq| ď p4πtq´d{2
expt´2x2 {4tudx }ψj }2 χA pxq . (A.45)

By the uniform boundedness principle, }ψj }2 ď }ψj }H1 ď C. Therefore, φj is bounded

uniformly in j. On the other hand, φj pxq Ñ φpxq pointwise, since ψj Ñ ψ weakly in H 1
and for every fixed x the function y ÞÑ expt´|x ´ y|2 {4tu is in the dual of H 1 . Therefore,
pointwise convergence follows from dominated convergence. This concludes the proof of
(A.36) for p “ 2.
Let us now prove Eq. (A.36) for all p such that 1 ď p ă 2d{pd ´ 2q. Consider first
1 ď p ď 2. By Hölder inequality:

}χA pψ ´ ψj q}p ď }χA }r }χA pψ ´ ψj q}2 , (A.46)

with 1{p “ 1{r ` 1{2. Using that χA P Lr for all r, this proves the theorem for 1 ď p ă 2.
Finally, consider p ą 2. Again by Hölder:

}χA pψ ´ ψj q}p ď }χA pψ ´ ψj q}α 1´α

2 }χA pψ ´ ψj q}q (A.47)

with α “ p1{p ´ 1{qqp1{2 ´ 1{qq, which is strictly positive if p ă q. Then, by the Sobolev
inequality:

}χA pψ ´ ψj q}Lq pRd q “ }ψ ´ ψj }Lq pAq ď Cp}∇ψ}L2 pRd q ` }∇ψ}L2 pAq q ď C 1 , (A.48)

where we used again the uniform boundedness principle and the lower semicontinuity of the
norm.

B Bathtub principle
In this appendix we shall briefly discuss the bathtub principle, used for instance in Section
8.2. We refer the reader to [3] for a more extensive discussion. Let µ be a Borel measure,
and let f be a real valued function, such that µpx : f pxq ă tq is bounded for all t. Consider
the functional ż
Ipgq “ dµpxq f pxqgpxq , (B.1)
ş
defined on gpxq such that 0 ď gpxq ď 1 and dµpxq gpxq “ G. We are interested in minimizing
Ipgq over all such functions. We claim that:

inf Ipgq “ Ipg˚ q , g˚ pxq “ χpf pxq ă sq ` cχpf pxq “ sq , (B.2)

131
where the numbers s and c are defined as:

s “ suptt | µpx : f pxq ă tq ď Gu , cµpx : f pxq “ sq “ G ´ µpx : f pxq ă sq . (B.3)

ş
Notice that dµpxqg˚ pxq “ G. First of all, let us prove that c ď 1. Suppose that µpf pxq “
sq ‰ 0; otherwise, c is undetermined. Suppose that c ą 1. Then:
ż ż
dµpxq χpf pxq ď sq ă dµpxq g˚ pxq “ G . (B.4)

Moreover,
ż
µpf pxq ď sq “ dµpxq χpf pxq ď sq
ż
“ dµpxq lim χpf pxq ă s ` 1{nq
nÑ8

“ lim µpf pxq ă s ` 1{nq , (B.5)

nÑ8

where the last step follows from monotone convergence theorem: the function χpf pxq ă
s ` 1{nq is nonincreasing in n, and its integral is bounded uniformly in n, by the assumptions
on f . Therefore:
lim µpf pxq ă s ` 1{nq ă G (B.6)
nÑ8

which means that there exists N large enough such that:

µpf pxq ă s ` 1{N q ă G . (B.7)

This however contradicts the definition of s in Eq. (B.3), hence c ď 1.

To prove that
ş g˚ is a minimizer, it is enough to show that given any h such that 0 ď
hpxq ď 1 and dµpxq hpxq “ G, one has Ipg˚ ´ hq ď 0. Let us check this. We write:
ż ż
Ipg˚ ´ hq “ dµpxqf pxqpg˚ pxq ´ hpxqq ` dµpxqf pxqpg˚ pxq ´ hpxqq
f ăs f ąs
ż
` dµpxqf pxqpg˚ pxq ´ hpxqq
f “s
ż ż
ď s dµpxqpg˚ pxq ´ hpxqq ` dµpxqf pxqp´hpxqq
f ăs f ąs
ż
`s dµpxqpg˚ pxq ´ hpxqq
f “s
ż ż
ď s dµpxqpg˚ pxq ´ hpxqq ` s dµpxqp´hpxqq
f ăs f ąs
ż
`s dµpxqpg˚ pxq ´ hpxqq
f “s
´ ż ż
“ s dµpxqpg˚ pxq ´ hpxqq ` dµpxqpg˚ pxq ´ hpxqq
f ăs f ąs
ż ¯
` dµpxqpg˚ pxq ´ hpxqq (B.8)
f “s

where in the first inequality we used that g˚ pxq “ 1 is f pxq ă s, and in the second inequality
that g˚ pxq “ 0 if f pxq ą s. Therefore,
ż
Ipg˚ ´ hq ď s dµpxqpg˚ pxq ´ hpxqq “ spG ´ Gq “ 0 , (B.9)

which concludes the proof of the claim.

132
References
[1] L. C. Evans. Partial Differential Equations. AMS, (2010).
[2] E. H. Lieb. Thomas-Fermi and related theories of atoms and molecules. Rev. Mod.
Phys. 53, 603-641 (1981).
[3] E. H. Lieb, M. Loss. Analysis. 2nd Edition. Graduate Studies in Mathematics, vol.
14. AMS (2001).
[4] E. H. Lieb and R. Seiringer. Stability of Matter. Cambridge University Press, (2010).
[5] M. Reed, B. Simon. Methods of Modern Mathematical Physics I: Functional Anal-
ysis. Revised and Enlarged Edition. Academic Press (1980).
[6] G. Teschl. Mathematical Methods in Quantum Mechanics. 2nd Edition. Graduate
Studies in Mathematics, vol. 157. AMS (2014).

133