Mathematical Theory of Many-Body Quantum Systems

Benjamin Schlein
January 14, 2025

These notes cover the class “Mathematical Theory of Many-Body Quantum Systems”,
taking place at the University of Zurich in the Spring Term of 2025. They are based on:

• E. H. Lieb, M. Loss. Analysis 2. Second Edition. Graduate Studies in Mathematics.

Volume 14, American Mathematical Society.

• E. H. Lieb, R. Seiringer. The Stability of Matter in Quantum Mechanics. Cambridge

University Press, 2010.

• E. H. Lieb. Thomas-Fermi and related theories of atoms and molecules. Rev. Mod.
Phys. 53, no. 4, 1981.

• V. Bach. Error bound for the Hartree-Fock energy of atoms and molecules. Comm.
Math. Phys. 147 (1992), no. 3, 527-548.

• G.M. Graf, J.P. Solovej. A correlation estimate with applications to quantum systems
with Coulomb interactions. Rev. Mod. Phys. 6 (1994), 977-997.

Contents
1 Mathematical Background 2
1.1 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Distributions and Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Sobolev Inequalities and Sobolev Embeddings . . . . . . . . . . . . . . . . . . 12

2 Introduction to Quantum Mechanics 15

2.1 Classical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Postulates of Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 The Harmonic Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 The Hydrogen Atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1
 2.5 Schrödinger Operators: Existence of Stationary States . . . . . . . . . . . . . . 36
2.6 Many-Body Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 43

3 Stability of Matter 53
3.1 Atoms and Molecules: Hamilton Operators and their Properties . . . . . . . . 53
3.2 Lieb-Thirring Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 Electrostatic Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 Indirect Part of the Coulomb Energy . . . . . . . . . . . . . . . . . . . . . . . 73
3.5 Proof of Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4 Thomas-Fermi Theory 83
4.1 Definition of Thomas-Fermi Functional and Existence of Minimizer . . . . . . 83
4.2 The No-Binding Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.3 Stability of Matter via Thomas-Fermi Theory . . . . . . . . . . . . . . . . . . 101
4.4 Thomas-Fermi Theory as Limit of Quantum Mechanics . . . . . . . . . . . . . 103

5 Hartree-Fock Theory 118

5.1 The Hartree-Fock Energy Functional . . . . . . . . . . . . . . . . . . . . . . . 118
5.2 An Abstract Correlation Inequality . . . . . . . . . . . . . . . . . . . . . . . . 125
5.3 Estimate for the Many-Body Interaction Energy . . . . . . . . . . . . . . . . . 130
5.4 Semiclassical Approximation through Coherent States . . . . . . . . . . . . . . 138
5.5 Proof of Theorem 5.1.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

1 Mathematical Background
1.1 Lp spaces
Definition and basic properties. Let (Ω, A, µ) be a measure space (ie. Ω is a set, A a
σ-algebra on Ω and µ a measure on A). For 1 ≤ p < ∞, we define
 Z 
p p
L (Ω, A, µ) = f : Ω → C measurable s.t. |f | dµ < ∞

For f ∈ Lp (Ω, A, µ), let

Z 1/p
p
∥f ∥p = |f | dµ

For p = ∞, we define

L∞ (Ω, A, µ) = {f : Ω → C measurable s.t. ∃ C > 0 with |f (x)| ≤ C almost everywhere}

2
For f ∈ Lp (Ω, A, µ), let
∥f ∥∞ = inf{C > 0 : |f (x)| ≤ C almost everywhere}
By definition, we have ∥.∥p : Lp (Ω, A, µ) → [0; ∞), for any 1 ≤ p ≤ ∞. One can also check
that
∥λf ∥p = |λ|∥f ∥p
for all λ ∈ C and all f ∈ Lp (Ω, A, µ) and that
∥f + g∥p ≤ ∥f ∥p + ∥g∥p (1.1)
for all f, g ∈ Lp (Ω, A, µ). Nevertheless, ∥.∥p does not define a norm on the vector space
Lp (Ω, A, µ), because ∥f ∥p = 0 does not imply that f = 0 (it only implies that f (x) = 0
almost everywhere).
To make ∥.∥p into a norm, we define, on Lp (Ω, A, µ), an equivalence relation. For f, g ∈
Lp (Ω, A, µ) we write f ∼ g if and only if f (x) = g(x) µ-almost everywhere. It is easy to check
that ∼ defines an equivalence relation. We can therefore define, for any 1 ≤ p ≤ ∞,
Lp (Ω, A, µ) = Lp (Ω, A, µ)/ ∼ = {[f ] : f ∈ Lp (Ω, A, µ)}
where [f ] = {g ∈ Lp (Ω, A, µ) : g ∼ f } is the equivalence class associated to f , which
contains all functions that coincide with f almost everywhere. Switching from Lp (Ω, A, µ) to
Lp (Ω, A, µ), we make the space smaller, because we identify functions that coincide almost
everywhere. By definition, elements of Lp (Ω, A, µ) are not functions, they are equivalence
classes of functions.
On Lp (Ω, A, µ) we can naturally define the structure of a vector space, by setting
[f ] + [g] = [f + g], λ[f ] = [λf ]
We can also define ∥.∥p : Lp (Ω, A, µ) → [0; ∞) through ∥[f ]∥p := ∥f ∥p (∥.∥p is well defined,
because ∥f ∥p = ∥g∥p if f ∼ g). When defined on Lp (Ω, A, µ), ∥.∥p is indeed a norm, because
∥[f ]∥p = 0 implies that f (x) = 0 almost everywhere, and therefore [f ] = [0]. We conclude
that, for every 1 ≤ p ≤ ∞, (Lp (Ω, A, µ), ∥.∥p ) is a normed vector space. As every norm, ∥.∥p
induces a metric on Lp (Ω, A, µ) (the metric is defined by d(f, g) = ∥f − g∥p ), and thus a
notion of convergence for sequences in Lp (Ω, A, µ).
Theorem 1.1.1. Let (Ω, A, µ) be a measure space, 1 ≤ p ≤ ∞. Then Lp (Ω, A, µ), equipped
with the norm ∥.∥p . is a Banach space, ie. it is a normed vector space, complete with respect to
the metric induced by ∥.∥p . In other words, every Cauchy sequence on Lp (Ω, A, µ) converges.
For p = 2, we can also define an inner product (a scalar product) on L2 (Ω, A, µ). For
f, g ∈ L2 (Ω, A, µ), we set Z
⟨f ; g⟩ = f g dµ (1.2)

3
It is easy to check that ⟨.; .⟩ is linear in its second argument, and antisymmetric (ie. ⟨g; f ⟩ =
⟨f ; g⟩). Moreover,
⟨f ; f ⟩ = ∥f ∥22 ≥ 0
Thus ⟨.; .⟩ is an inner product on L2 (Ω, A, µ) and it induces the norm ∥.∥2 . Hence, L2 (Ω, A, µ)
is a Hilbert space. In particular, we have the Cauchy-Schwarz inequality
Z
f g dµ = |⟨f ; g⟩| ≤ ∥f ∥2 ∥g∥2

for all f, g ∈ L2 (Ω, A, dµ).

An important extension of the Cauchy-Schwarz inequality to Lp -spaces, for p ̸= 2, is
Hölder’s inequality.
Theorem 1.1.2 (Hölder’s inequality). Let 1 ≤ p, p′ ≤ ∞, with 1/p + 1/p′ = 1. For f ∈
′
Lp (Ω, A, µ), g ∈ Lp (Ω, A, µ), we have
Z Z
f g dµ ≤ |f ||g|dµ ≤ ∥f ∥p ∥g∥p′ (1.3)

The dual of Lp (Ω, A, µ). For an arbitrary normed vector space (X, ∥.∥) over C, a linear
functional on X is a linear map L : X → C. A linear functional L is continuous if and only if
it is bounded, ie. if there exists a constant C > 0 such that |L(f )| ≤ C∥f ∥ for all f ∈ X. We
define the dual space of X as
X ∗ = {L : X → C : L is a continuous linear functional on X}
On X ∗ we can naturally introduce a sum and a multiplication with complex numbers. Hence
X ∗ is a vector space. Since continuous functionals are bounded, we can also introduce a norm
on X ∗ , by setting, for L ∈ X ∗ ,
|L(f )|
∥L∥ = sup |L(f )| = sup (1.4)
f ∈X:∥f ∥≤1 f ∈X:f ̸=0 ∥f ∥

Equipped with this norm, X ∗ is a normed vector space. In fact, X ∗ is always complete (w.r.t.
the norm (1.4), of course), and thus a Banach space.
In the following, we would like to characterize the dual space of Lp (Ω, A, µ). To this end,
′
we observe that, choosing 1 ≤ p′ ≤ ∞ so that 1/p + 1/p′ = 1, for every g ∈ Lp (Ω, A, µ), we
can define a linear functional Lg : Lp (Ω, A, µ) → C by setting
Z
Lg (f ) = f gdµ

Linearity of Lg is clear, continuity follows from Hölder’s inequality. For 1 ≤ p < ∞, it turns
out that all linear functionals on Lp (Ω, A, µ) have this form.

4
Theorem 1.1.3. Let (Ω, A, µ) be a measure space, 1 ≤ p < ∞ and 1 < p′ ≤ ∞ with
1/p+1/p′ = 1. If p = 1 and p′ = ∞, we make the additional assumption that the measure space
′
is sigma-finite. The map ϕ : Lp (Ω, A, µ) → Lp (Ω, A, µ)∗ defined by ϕ(g) = ϕg : Lp (Ω, A, µ) →
C, with Z
ϕg (f ) = f gdµ

is an isometric isomorphism, meaning that it is linear, it preserves the norm, in the sense
that ∥ϕg ∥Lp (Ω,A,µ)∗ = ∥g∥Lp′ , and it is a bijection.
Remark: If p = ∞ and p′ = 1, the map ϕ : L1 (Ω, A, µ) → L∞ (Ω, A, µ)∗ is still well defined,
linear, isometric (and therefore injective), but it is (in general) not surjective.
Approximation with smooth functions. We will mostly deal with functions defined on
subsets Ω ⊂ Rn , equipped with the Lebesgue σ-algebra M and with Lebesgue measure λ. In
this case, we will use the notation Lp (Ω) = Lp (Ω, M, λn ).
An important property of functions in Lp (Ω), for 1 ≤ p < ∞, is that they can be ap-
proximated by smooth functions, defined through convolutions. To introduce the notion of
convolution, we use Young’s inequality.
Theorem 1.1.4. Let 1 ≤ p, q, r ≤ ∞, with 1/p + 1/q + 1/r = 2. Let f ∈ Lp (Rn ), g ∈ Lq (Rn ),
h ∈ Lr (Rn ). Then
Z
f (x)g(x − y)h(y) dλn (x)dλn (y) ≤ ∥f ∥p ∥g∥q ∥h∥r (1.5)
Rn ×Rn

Let us also mention here a useful extension of Young’s inequality, known as the Hardy-
Littlewood-Sobolev inequality, which allows us to take g(x) = |x|−κ , 0 < κ < n, in (1.5).
Theorem 1.1.5. Let p, r > 1 and 0 < κ < n with 1/p + κ/n + 1/r = 2. Then there exists a
constant C (depending on n, κ, p) such that
Z
f (x)|x − y|−κ h(y)dλn (x)dλn (y) ≤ C∥f ∥p ∥h∥r (1.6)
Rn ×Rn

for all f ∈ Lp (Rn ), h ∈ Lr (Rn ).

We can use Young’s inequality to define the convolution of two functions. For 1 ≤ p, q ≤ ∞,
f ∈ Lp (Rn ) and g ∈ Lq (Rn ), let
Z
h(x) := |f (x − y)||g(y)|dλ(y) (1.7)

Since the function we are integrating is positive, h is certainly well-defined, although it could be
infinite. Using Young’s inequality, we find (by duality) that h ∈ Lr (Rn ), if 1+1/r = 1/p+1/q,

5
with ∥h∥r ≤ ∥f ∥p ∥g∥q . In particular, h(x) < ∞ almost everywhere. Thus, the function
y → f (x − y)g(y) is integrable in y ∈ Rn , for almost all x ∈ Rn . Therefore, we can define
 R
f (x − y)g(y)dλn (y) if f (x − .)g(.) is integrable
(f ∗ g)(x) = (1.8)
0 otherwise

We obtain
∥f ∗ g∥r ≤ ∥f ∥p ∥g∥q (1.9)
if 1 + 1/r = 1/p + 1/q, ie. f ∗ g ∈ Lr (Rn ).
Using convolutions, we can approximate functions in Lp (Rn ) (and more generally in Lp (Ω),
for Ω ⊂ Rn measurable) by sequences of smooth functions (note the the following theorem
does not hold if p = ∞).

jdλ = 1. For ε > 0, let jε (x) = ε−n j(x/ε), so that

1 n
R
Theorem
R 1.1.6. Let j ∈ L (R ) with
jε dλ = 1 and ∥jε ∥1 = ∥j∥1 for all ε > 0. Let f ∈ Lp (Rn ) for 1 ≤ p < ∞ and let fε = jε ∗ f ,
defined as in (1.8). We have

i) fε ∈ Lp (Rn ), with ∥fε ∥p ≤ ∥j∥1 ∥f ∥p .

ii) fε → f in Lp (Rn ), as ε → 0.

iii) If we additionally assume that j ∈ Cc∞ (Rn ), then fε ∈ C ∞ (Rn ) for all ε > 0 and
Dα fε = Dα jε ∗ f . Here and in the following we use the notation

Cc∞ (Rn ) = {f ∈ C ∞ (Rn ) : f has a compact support} .

Recall that a normed vector space is called separable if there exists a countable subset
B ⊂ X that is dense in X. From the last theorem, we obtain separability of Lp (Ω), for all
1 ≤ p < ∞.

Theorem 1.1.7. For every measurable Ω ⊂ Rn and 1 ≤ p < ∞, the Banach space Lp (Ω) is
separable.

Weak convergence. On an arbitrary normed vector space (X, ∥.∥), we say a sequence
{fj }j∈N in X converges strongly to f ∈ X, if ∥fj − f ∥ → 0, as j → ∞.

Definition 1.1.8. Let (X, ∥.∥) be a normed vector space, and X ∗ its dual space (namely the
space of all continuous linear functionals on X). Let {fj }j∈N be a sequence in X and f ∈ X.
We say that fj converges weakly to f , and we write fj ⇀ f if, for every L ∈ X ∗ , L(fj ) → L(f ).

6
 It is easy to check that strong convergence implies weak convergence, ie. if {fj }j∈N is a
sequence on a normed space X, with fj → f , for some f ∈ X, then we also have fj ⇀ f .
In the following, we focus on weak convergence on Lp (Ω, A, µ). For 1 ≤ p < ∞ (assuming
additionally that (Ω, A, µ) is sigma-finite, if p = 1), it follows from Theorem 1.1.3 that a
sequence fj in Lp (Ω, A, µ) converges weakly to f ∈ Lp (Ω, A, µ) if and only if
Z Z
fj gdµ → f gdµ

′
as j → ∞, for all g ∈ Lp (Ω, A, µ), with 1 < p′ ≤ ∞ so that 1/p + 1/p′ = 1. Here are some
important properties of weak convergence on Lp spaces. 1

• Uniqueness of weak limit. Let (Ω, A, µ) be a measure space (if p = ∞, we assume

additionally that the measure space if sigma-finite). Let fj be a sequence in Lp (Ω, A, µ),
with fj ⇀ f and fj ⇀ g as j → ∞, for some f, g ∈ Lp (Ω, A, µ). Then f = g.

• Lower semicontinuity of the norm. Let (Ω, A, µ) be a measure space (if p = ∞, we

assume additionally that the measure space if sigma-finite). Suppose that {fj }j∈N is a
sequence on Lp (Ω, A, µ) and f ∈ Lp (Ω, A, µ), with fj ⇀ f . Then

∥f ∥p ≤ lim inf ∥fj ∥p (1.10)

j→∞

Moreover, if 1 < p < ∞ and limj→∞ ∥fj ∥p = ∥f ∥p , then fj → f .

• Uniform boundedness principle. Let (Ω, A, µ) be a measure space. Suppose that {fj }j∈N
is a sequence on Lp (Ω, A, µ) and f ∈ Lp (Ω, A, µ), with fj ⇀ f . Then there exists C > 0
such that ∥fj ∥p ≤ C, for all j ∈ N.

Finally, a very important and useful properties of weak convergence on Lp -spaces is the
Banach-Alaoglu Theorem. We state it here for Lp (Ω, M, λ) where Ω ⊂ Rn is a Lebesgue
measurable subset of Rn , M is the Lebesgue σ-algebra on Ω and λ is Lebesgue measure on M
(the theorem can be extended to general reflexive Banach spaces, and even to non-reflexive
separable Banach spaces, using the notion of weak-* convergence).

Theorem 1.1.9. Let Ω ⊂ Rn be a measurable set and let {fj }j∈N be a bounded sequence in
Lp (Ω), for a 1 < p < ∞. Then there exists a subsequence {fnj }j∈N and f ∈ Lp (Ω) such that
fnj ⇀ f weakly, as j → ∞.
1
These properties hold for general normed spaces, but in general they rely on the axiom of choice; for Lp
spaces, they can be proved without further assumptions

7
Fourier transform. For f ∈ L1 (Rn ), we define the Fourier transform fˆ : Rn → C of f ,
setting Z
ˆ 1
f (k) = f (x)e−ik·x dλn (x) (1.11)
(2π)n/2
The Fourier transform has the following important properties.
• Boundedness and continuity. Let f ∈ L1 (Rn ). Then fˆ ∈ L∞ (Rn ) with

∥fˆ∥∞ ≤ (2π)−n/2 ∥f ∥1

Moreover, fˆ is continuous.

• Riemann-Lebesgue Lemma. Let f ∈ L1 (Rn ). Then fˆ(k) → 0, as |k| → ∞.

• Fourier transform of convolutions. For f, g ∈ L1 (Rn ) we have (from Young’s inequality)

f ∗ g ∈ L1 (Rn ). Then
f[∗ g(k) = (2π)n/2 fˆ(k)ĝ(k)

• Fourier transform of a Gaussian. Let fα (x) = e−αx

2 /2
. Then
2
fˆα (k) = α−n/2 e−k /2α

• Derivatives and Fourier transform. Let f ∈ L1 (Rn ), and g : Rn → Cn be defined

through g(x) = −ixf (x). If g is integrable, then fˆ ∈ C 1 (Rn ) with ∇fˆ = ĝ.

• Plancherel’s identity. Let f ∈ L1 (Rn ) ∩ L2 (Rn ). Then fˆ ∈ L2 (Rn ) and ∥fˆ∥2 = ∥f ∥2 .

Plancherel’s identity can be used to extend the Fourier transform to L2 . For an arbitrary
f ∈ L2 (Rn ), we can find a sequence fj ∈ L1 (Rn ) ∩ L2 (Rn ) with fj → f in L2 . Since
fj ∈ L1 (Rn ), we can define the Fourier transform fˆj . From Plancherel’s identity, it follows
that fˆj ∈ L2 (Rn ), with ∥fˆj ∥2 = ∥fj ∥2 . Moreover

∥fˆj − fˆm ∥2 = ∥f\

j − fm ∥2 = ∥fj − fm ∥2 → 0

as j, m → ∞. We conclude that fˆj is a Cauchy sequence in L2 (Rn ) and thus it must converge.
We can therefore define
fˆ = lim fˆj . (1.12)
j→∞

It is important to observe that fˆ does not depend on the choice of the sequence fj ∈ L1 (Rn ) ∩
L2 (Rn ).
The map ˆ· : L2 (Rn ) → L2 (Rn ) is clearly linear and isometric, ie. ∥fˆ∥2 = ∥f ∥2 . It also
preserves the inner product, ie. ⟨fˆ; ĝ⟩ = ⟨f ; g⟩ for all f, g ∈ L2 (Rn ). In fact, ˆ· : L2 (Rn ) →

8
L2 (Rn ) is also invertible (the inverse is the map ˇ· : L2 (Rn ) → L2 (Rn ), defined by fˇ(k) := fˆ(−k)
for all f ∈ L2 (Rn )). Thus, the Fourier transform ˆ· : L2 (Rn ) → L2 (Rn ) is a unitary map.
So far, we defined the Fourier transform as a linear map from L1 (Rn ) to L∞ (Rn ) and as a
linear map from L2 (Rn ) to L2 (Rn ). By interpolation, it can also be extended as a linear map
′
from Lp (Rn ) to Lp (Rn ), for all 1 < p < 2, choosing 2 < p′ < ∞ so that 1/p + 1/p′ = 1. The
extension is based on the Hausdorff-Young inequality (which replace the Plancherel’s identity,
for p < 2).
Theorem 1.1.10. Let 1 ≤ p ≤ 2 and 2 ≤ p′ ≤ ∞ such that 1/p + 1/p′ = 1. Then there exists
′
a constant C > 0 (depending on the dimension n and on p) such that fˆ ∈ Lp (Rn ) and
∥fˆ∥p′ ≤ C∥f ∥p (1.13)
for any f ∈ L1 (Rn ) ∩ Lp (Rn ).

1.2 Distributions and Sobolev Spaces

Definition and basic properties. We start with the definition of the space of test functions.
Definition 1.2.1. Let n ∈ N\{0} and Ω ⊂ Rn be an open non-empty set (Ω = Rn is allowed).
We denote the space of test functions on Ω by D(Ω). It consists of all functions in Cc∞ (Ω) (the
space of infinitely differentiable functions that are compactly supported inside Ω), equipped with
the following notion of convergence. A sequence ϕm ∈ Cc∞ (Ω) converges in D(Ω) to ϕ ∈ Cc∞ (Ω)
if and only if there exists a fixed compact set K ⊂ Ω such that supp (ϕm − ϕ) ⊂ K for all
m ∈ N and
Dα ϕm → Dα ϕ
uniformly on K, for all multi-index α ∈ Nn . In other words,
∂ |α| ϕm (x) ∂ |α| ϕ(x)
sup − →0
x∈K ∂xα1 1 . . . ∂xαnn ∂xα1 1 . . . ∂xαnn
as m → ∞, for all α = (α1 , . . . , αn ) ∈ Nn (with |α| = α1 + · · · + αn ).
Distributions are then continuous linear functionals on D(Ω) (since the notion of conver-
gence on D(Ω) is very strong, it is very easy for a linear functional on D(Ω) to be continuous).
Definition 1.2.2. A distribution T is a continuous linear functional on the space of test
functions D(Ω). In other words, it is a map T : D(Ω) → C, with
T (ϕ1 + λϕ2 ) = T (ϕ1 ) + λT (ϕ2 )
for all λ ∈ C, ϕ1 , ϕ2 ∈ D(Ω), such that T (ϕm ) → T (ϕ) as m → ∞, whenever ϕm → ϕ
in D(Ω). We denote by D′ (Ω) the space of distributions on Ω. Observe that D′ (Ω) has the
structure of a vector space (distributions can be summed and multiplied with scalars λ ∈ C).

9
 An important observation is that every Lp -function defines a distribution. Since the be-
havior at infinity is not important, we define local Lp -spaces.
Definition 1.2.3. For 1 ≤ p ≤ ∞, we define the space of locally p-th power integrable
functions
Lploc (Ω) = {f : Ω → C measurable such that ∥f ∥Lp (K) < ∞ for all K ⊂ Ω compact}
Notice that Lploc (Ω) is a vector space. In contrast with Lp (Ω), however, there is no natural
norm on Lploc (Ω). Notice also that Lqloc (Ω) ⊂ Lploc (Ω) if 1 ≤ p ≤ q ≤ ∞. In particular,
L1loc (Ω) ⊃ Lploc (Ω) ⊃ Lp (Ω) for all 1 ≤ p ≤ ∞.
Every locally integrable function defines a distribution.
Lemma 1.2.4. Let Ω ⊂ Rn open, non-empty, f ∈ L1loc (Ω). We define Tf : D(Ω) → C setting
Z
Tf (ϕ) = f ϕ dλn (x) (1.14)
Ω

for all ϕ ∈ D(Ω). Then Tf ∈ D′ (Ω) is a distribution (it is the distribution associated with f ).
While every locally integrable function f ∈ L1loc (Ω) defines a distribution Tf ∈ D′ (Ω), not
every distribution in D′ (Ω) has the form Tf for an f ∈ L1loc (Ω). An example of a distribution
which is not associated with a locally integrable function is the Dirac delta-distribution
δx (ϕ) = ϕ(x) (1.15)
for all ϕ ∈ D(Ω) and for a x ∈ Ω.
Distributional derivatives and Sobolev spaces. The derivative of a distribution is defined
as follows.
Definition 1.2.5. Let Ω ⊂ Rn be open and non-empty, T ∈ D′ (Ω) a distribution. For a
multi-index α = (α1 , . . . , αn ) ∈ Nn , we define Dα T ∈ D′ (Ω) through
Dα T (ϕ) = (−1)|α| T (Dα ϕ) (1.16)
with |α| = α1 + · · · + αn .
Remark: Let α = (α1 , . . . , αn ) ∈ Nn be a multi-index and f ∈ C |α| (Ω). In this case,
we have Dα Tf = TDα f . In other words, if we identify functions with the corresponding
distributions, classical derivatives coincide, when they exist, with distributional derivatives.
But, of course, distributional derivatives are much more general, since they can be applied
to arbitrary locally integrable functions (for which classical derivatives do not need to exist)
and, even more generally, to arbitrary distributions.
We can now define the Sobolev space W m,p (Ω) as the space of all functions f ∈ Lp (Ω) whose
distributional derivatives of order up to m (defined through the distribution Tf associated with
f ) are again functions in Lp (Ω).

10
Definition 1.2.6. For Ω ⊂ Rn open, non-empty, m ∈ N and 1 ≤ p ≤ ∞, we define

W m,p (Ω) = {f ∈ Lp (Ω) : ∀ α ∈ Nn with |α| ≤ m there is gα ∈ Lp (Ω) with Dα Tf = Tgα }

For f ∈ W m,p (Ω) and α ∈ Nn with |α| ≤ m, we define Dα f = gα ∈ Lp (Ω) (the functions Dα f
are known as the distributional or weak derivatives of f ). For 1 ≤ p < ∞, we define
 1/p
X
∥f ∥W m,p =  ∥Dα f ∥pp 
α∈Nn :|α|≤m

For p = ∞, we set X
∥f ∥W m,∞ = ∥Dα f ∥∞ .
|α|≤m

Then ∥.∥W m,p is a norm on W m,p (Ω) and (W m,p (Ω), ∥.∥W m,p ) is a complete Banach space.
An important property of functions in the Sobolev space W m,p (Ω), for 1 ≤ p < ∞, is the
fact that they can be approximated by sequences of smooth functions. This is known as the
Meyers-Serrin theorem.
Theorem 1.2.7. Let m ∈ N, 1 ≤ p < ∞, Ω ⊂ Rn open, non-empty. Let f ∈ W m,p (Ω). Then
there exists a sequence fj ∈ C ∞ (Ω) ∩ W m,p (Ω) such that

∥f − fj ∥W m,p (Ω) → 0

as j → ∞. If Ω = Rn , the sequence fj can be chosen in Cc∞ (Rn ) (ie. to have compact support).
Remark. If Ω ⊂ Rn has a boundary and m ∈ N\{0}, Cc∞ (Ω) is not dense in W m,p (Ω). The
completion of Cc∞ (Ω) with respect to the W m,p (Ω) norm is a subspace of W m,p (Ω), typically
denoted by W0m,p (Ω), which is used to solve differential equations with Dirichlet boundary
conditions.
The Hilbert spaces H m (Ω). Let Ω ⊂ Rn open and non-empty, m ∈ N. We will use the
notation H m (Ω) for the space W m,2 (Ω). For f, g ∈ H m (Ω), we define the inner product
X
⟨f ; g⟩H m = ⟨Dα f ; Dα g⟩2
α∈Nn :|α|≤m

where ⟨.; .⟩2 denotes the usual inner product on L2 (Ω). Clearly,

∥f ∥2W m,2 ≡ ∥f ∥2H m = ⟨f ; f ⟩H m .

Thus, (H m (Ω), ⟨.; .⟩H m ) is a Hilbert space.

For Ω = Rn , there is a useful characterisaztion of H m (Ω) in terms of Fourier transform.

11
Theorem 1.2.8. Let f ∈ L2 (Rn ), and fˆ ∈ L2 (Rn ) be its Fourier transform. Then f ∈
H m (Rn ) if and only if the function k → |k|m fˆ(k) is in L2 (Rn ). In this case, the distributional
derivatives of f satisfy D dα f (k) = i|α| k α fˆ(k) for all α ∈ Nn with |α| ≤ m and therefore

X Z
2
∥f ∥H m = |k1 |2α1 |k2 |2α2 . . . |kn |2αn |fˆ(k)|2 dk
α∈Nn :|α|≤m Rn

Green’s functions of the Laplacian. Let G2 (x) = (2π)−1 log |x − y| and, for n ∈ N, n ≥ 3,
1 1
Gn (x) =
(n − 2)|S | |x|n−2
n−1

where |S n−1 | = 2π n/2 /Γ(n/2) is the measure of the (n − 1)-dimensional sphere with radius
one. Then, we have, in the sense of distributions,

−∆Gn = δ (1.17)

where δ ∈ D′ (Rn ) is defined through δ(ϕ) = ϕ(0). Thus, if f ∈ L1loc (Rn ) such that y →
Gn (x − y)f (y) is integrable, for almost every x ∈ Rn , then the function u = Gn ∗ f ∈ L1loc (Rn )
solves the Poisson equation −∆u = f , in the sense of distributions.

1.3 Sobolev Inequalities and Sobolev Embeddings

Sobolev inequalities for gradients. We begin with the standard Sobolev inequality for
functions in H 1 (Rn ), n ≥ 3.
Theorem 1.3.1. Let n ∈ N, n ≥ 3. Let f ∈ H 1 (Rn ). Then f ∈ Lq (Rn ), for q = 2n/(n − 2).
Moreover, there exists a constant Cn (depending only on the dimension n) such that

∥f ∥q ≤ Cn ∥∇f ∥2 (1.18)

for all f ∈ H 1 (Rn ).

In dimensions n = 1, 2, we cannot bound high Lq norms just with the L2 norm of ∇f , we
also need the L2 norm of f .
Theorem 1.3.2. One-dimension: There exists a constant C > 0 such that

∥f ∥q ≤ C∥f ∥H 1

for all f ∈ H 1 (R) and all 2 ≤ q ≤ ∞. Moreover, every f ∈ H 1 (R) has a representative
satisfying
|f (x) − f (y)| ≤ ∥∂f ∥2 |x − y|1/2

12
for all x, y ∈ R. In other words, H 1 -functions in one-dimension are automatically Hölder
continuous with exponent α = 1/2.
Two-dimensions: For every 2 ≤ q < ∞ there exists C > 0 such that

∥f ∥q ≤ C∥f ∥H 1

for all f ∈ H 1 (R2 ).

Remark: in two dimensions, the L∞ -norm cannot be bounded by the H 1 -norm.

Compact Sobolev embedding. The Sobolev inequality implies that the unit ball in H 1 (Rn )
is a subset of Lq (Rn ), if q ≤ 2n/(n − 2) (or if q < ∞ in two dimensions, q ≤ ∞ in one
dimension). When restricting on sets with finite measure, it turns out that the unit ball of
H 1 is sequentially compact subset of Lq . This follows from the following theorem.

Theorem 1.3.3. Let {fj }j∈N be a weakly convergent sequence in H 1 (Rn ), let f denote its weak
limit. Let moreover A ⊂ Rn be a set of finite measure and χA the corresponding characteristic
function. Then χA fj → χA f strongly in Lq (Rn ), for all 1 ≤ q < 2n/(n − 2), if n ≥ 3, and
for all 1 ≤ q < ∞, if n = 1, 2. For n = 1, the convergence is pointwise and uniform on every
compact subset of R.

General Sobolev inequalities and embedding theorems. Sobolev inequalities can be

extended to domains with the cone property, defined as follows.

Definition 1.3.4. For r > 0 and θ ∈ [0; 2π], let

Kr,θ := {x ∈ Rn : x ̸= 0, 0 < xn < |x| cos θ} ∩ Br (0)

denote a finite cone of angle θ and length r. We say that an open domain Ω ⊂ Rn has the
cone property if there exists r, θ such that, for all x ∈ Ω there is a finite cone Kx congruent
to Kr,θ which is contained in Ω and has vertex x.

We state now the general form of Sobolev inequality.

Theorem 1.3.5. Let Ω ⊂ Rn be an open set, with the cone property for some r, θ. Let
1 ≤ p ≤ q ≤ ∞, m, k ∈ N, with m ≥ 1 and k ≤ m. Then there exists a constant C > 0
(depending on m, k, q, p, θ, r) such that, for all f ∈ W m,p (Ω), we have

i) If kp < n,
∥f ∥W m−k,q (Ω) ≤ C∥f ∥W m,p (Ω)
for all p ≤ q ≤ np/(n − kp).

13
 ii) If kp = n,
∥f ∥W m−k,q (Ω) ≤ C∥f ∥W m,p (Ω)
for all p ≤ q < ∞.
iii) If kp > n,
∥f ∥W m−k,q (Ω) ≤ C∥f ∥W m,p (Ω)
for all p ≤ q ≤ ∞. In particular,
max sup |Dα f (x)| ≤ C∥f ∥W m,p (Ω)
0≤|α|≤m−k x∈Ω

The next theorem, known as the Rellich-Kondrachov theorem, provides the general state-
ment for compact Sobolev embeddings.
Theorem 1.3.6. Let Ω ⊂ Rn be an open set, with the cone property for some r, θ. Let
1 ≤ p ≤ q ≤ ∞, m, k ∈ N, with m ≥ 1 and k ≤ m. Let f ∈ W m,p (Ω) and {fj }j∈N be a
sequence in W m,p (Ω) with fj ⇀ f weakly in W m,p (Ω) (meaning that Dα fj ⇀ Dα f weakly in
Lp (Ω), for all multiindex α ∈ Nn with 0 ≤ |α| ≤ m). Let ω ⊂ Ω be open and bounded (in
particular, we can take ω = Ω, if Ω ⊂ R3 is bounded). Then, we have
i) If kp < n, 1 ≤ q < np/(n − kp), then ∥fj − f ∥W m−k,q (ω) → 0, as j → ∞.
ii) If kp = n, 1 ≤ q < ∞, then ∥fj − f ∥W m−k,q (ω) → 0, as j → ∞.
iii) If kp > n, 1 ≤ q ≤ ∞, then ∥fj − f ∥W m−k,q (ω) → 0, as j → ∞. In particular, taking
q = ∞, we conclude that
max sup |Dα fj (x) − Dα f (x)| → 0
0≤|α|≤m−k x∈ω

as j → ∞.

Poincaré inequality. Finally, let us also mention a useful variant of the Sobolev inequality,
known as the Poincaré inequality, which allows us to bound fluctuations of functions in Sobolev
spaces.
Theorem 1.3.7. Let Ω ⊂ Rn be a bounded, connected, open set with the cone property for
p′
R
some r, θ. Let 1 ≤ p < ∞, and let g ∈ L (Ω) with gdλn = 1. Then there exists a constant
C = C(Ω, g, p, q) such that
Z
f − f g dλn ≤ C∥∇f ∥Lp (Ω)
Ω Lq (Ω)

for all f ∈ W 1,p (Ω) and all 1 ≤ q < pn/(n − p) if p < n, all 1 ≤ q < ∞ if p = n, and all
1 ≤ q ≤ ∞ if p > n.

14
2 Introduction to Quantum Mechanics
2.1 Classical Mechanics
Consider a particle moving on Rn . Classically, such a particle is described by a vector (x, p) ∈
R2n , where x denotes the position of the particle and p its momentum (p = mv, with m the
mass and v = ẋ the velocity of the particle). The space R2n of all vectors (x, p) is known as
the phase space of the system.
To determine the evolution of the system, we need to find (x(t), v(t)) for all times t ∈ R.
This can be achieved solving Newton’s equation mẍ(t) = F (x(t)), where F (x) is the force at
point x. If the force field F is a gradient field, induced by a potential V with F = −∇V ,
Newton’s equation takes the form

ẋ(t) = p(t)/m
ṗ(t) = −∇V (x(t))

Classical mechanics is deterministic: if we know (x, p) at time t = t0 , we can determine

(x(t), p(t)) for all t ∈ R, solving Newton’s equation.
An important quantity is the Hamiltonian
p2
H(x, p) = + V (x)
2m
The numerical value of the function H gives the energy of the system. Newton’s equation can
be rewritten as
ẋ(t) = ∂H

∂p
(x(t), p(t))
∂H
ṗ(t) = − ∂x (x(t), p(t))
Thus, we obtain
d ∂H ∂H ∂H ∂H ∂H ∂H
H(x(t), p(t)) = · ẋ(t) + · ṗ(t) = · (x(t), p(t)) − · (x(t), p(t)) = 0
dt ∂x ∂p ∂x ∂p ∂p ∂x
In other words, the Hamiltonian is preserved by the time-evolution.
More generally, observables of the system are functions A(x, p) on the phase space. The
time evolution of an observable is determined by
d ∂A ∂A
A(x(t), p(t)) = · ẋ(t) + · ṗ(t)
dt ∂x ∂p
(2.19)
∂A ∂H ∂A ∂H
= · (x(t), p(t)) − · (x(t), p(t)) = {A, H} (x(t), p(t))
∂x ∂p ∂p ∂x
where
∂A ∂H ∂A ∂H
{A, H} = · − ·
∂x ∂p ∂p ∂x

15
is known as the Poisson bracket of the observables A, H (in other words, the evolution of
the observable A is determined by the Hamiltonian H, taking the Poisson bracket). Every
observable A with the property {A, H} = 0 is therefore preserved along the time-evolution.
Conservation of energy is a consequence of {H, H} = 0. If H(x, p) = p2 /2m (a free particle,
with no force), we find {p, H} = 0, which means that momentum is preserved.
Conservation laws are very important, they often allow us to solve Newton’s equation.
Consider a particle moving in the three-dimensional space, under the action of a radial po-
tential V (|x|). Then, we claim that the angular momentum L = x ∧ p is preserved by the
time-evolution (each one of the three components of L are preserved). Using the notation εijk
for the fully antisymmetric tensor (with ε123 = ε231 = ε312 = 1, ε321 = ε213 = ε132 = −1, and
εijk = 0 if at least two indices coincide), we find
X
{(x ∧ p)i , H} = εijk {xj pk , H}
j,k
X ∂(xj pk ) ∂H ∂(xj pk ) ∂H
= εijk −
j,k
∂xℓ ∂pℓ ∂pℓ ∂xℓ
X
= εijk [δjℓ pk pℓ /m − xj δkℓ ∂ℓ V ]
j,k
1
= (p ∧ p)i − (x ∧ ∇V )i = 0
m
because ∇V (x) is parallel to x, for radial V .
Conservation of the angular momentum L implies, first of all, that the particle moves on a
fixed plane, going through the origin and orthogonal to L. Introducing polar coordinates (r, φ)
on this fixed plane, we can parametrize the position of the particle by the coordinates x =
(r cos φ, r sin φ). This implies ẋ = ṙ(cos φ, sin φ) + rφ̇(− sin φ, cos φ). Thus ∥ẋ∥2 = ṙ2 + r2 φ̇2
and ∥x∧ ẋ∥ = r2 φ̇. From conservation of angular momentum, we obtain the area law (Kepler’s
second law), stating that the line joining the particle with the origin sweeps out equal areas
during equal intervals of time. Moreover, expressing the energy of the particles through the
coordinates (r, φ), we find
p2 1 2 1 2 1 2 2 1 2 L2
E= + V (r) = mẋ + V (r) = mṙ + mr φ̇ + V (r) = mṙ + + V (r)
2m 2 2 2 2 2mr2
which implies
2
ṙ2 =
(E − U (r))
m
with the effective potential U (r) = L2 /(2mr2 ) + V (r). Therefore
r
dr 2
= (E − U (r))
dt m

16
and Z r(t)
dr
q =t (2.20)
2
r(0)
m
(E − U (r))

Solving the integral, we can find r(t). Inserting into ∥L∥ = mr(t)2 φ̇(t), we can find φ̇(t)
and then, again through integration, φ(t). Alternatively, considering t = t(r) (inverting the
relation from (2.20)) and φ = φ(t(r)) as a function of r, we can observe that

dφ φ̇(t(r)) ∥L∥ ∥L∥

= = 2
= q
dr ṙ(t(r)) mr ṙ(t(r)) mr2 m2 (E − U (r))

which implies
r
∥L∥
Z
φ(r) − φ0 = p dx (2.21)
r0 x2 2m(E − U (x))
Solving this equation determines the trajectory of the particle starting from the initial position
(r0 , φ0 ). In other words, the solution of Newton’s equation is reduced, using conservation of
energy and angular momentum, to the computation of integrals (such a system is called
integrable).
Kepler applied this approach to solve Newton’s equation already at the beginning of the
XVII century, with V (r) = −GM m/|x|, to describe the motion of planets around the sun (here
m is the mass of the planet, M is the mass of the sun and G is the gravitational constant). In
this case, we find U (r) = L2 /(2mr2 )−GM m/r and, solving (2.21), we obtain that trajectories
are ellipses, with one focus in the origin (the sun) if E < 0, parabolas if E = 0 and hyperbolas
if E > 0 (Kepler’s first law).

2.2 Postulates of Quantum Mechanics

States and observables. Quantum mechanical systems are described on Hilbert spaces.
States of the system are associated with vectors ψ ∈ H, normalized so that ∥ψ∥ = 1. Ob-
servables are associated with self-adjoint operators on H. The expectation of an observable
described by the operator A in the state ψ is given by the inner product ⟨ψ, Aψ⟩.
Let us consider, for example, a system consisting of a particle moving in Rn , which can be
described on the Hilbert space H = L2 (Rn ; C) of complex-valued, square integrable functions
on Rn . Physically, |ψ(x)|2 is interpreted as the probability density for finding the particle
at position x. Accordingly, the probability that the particle is located in a region B ⊂ Rn
(assume B to be measurable) is given by
Z Z
|ψ(x)| dx = χB (x)|ψ(x)|2 dx = ⟨ψ, χB (x)ψ⟩
2
B

17
where χB is the characteristic function of B. Similarly, the expectation value of the position
of the particle is determined by
Z
x|ψ(x)|2 dx = ⟨ψ, xψ⟩

The wave function does not only determine the distribution of the position of the particle.
Instead, every observable physical quantity is associated with a self-adjoint operator A acting
on the Hilbert space L2 (Rn ) so that the expectation value of the observable is given by the
inner product ⟨ψ, Aψ⟩ and the probability that a measurement of A gives a result in a set
B ⊂ R is determined by ⟨ψ, χB (A)ψ⟩, where χB (A) is defined through functional calculus.
As we saw above, the self-adjoint operator associated with the observable “position of the
particle” is the multiplication operator mapping ψ(x) to xψ(x) (the position is a vector-valued
observable, it would be better to consider separately each component of x). Notice that the
position operator, defined as multiplication through x, is an unbounded self-adjoint operator
on L2 (Rn ). For this reason, it is not defined on the full Hilbert space L2 (Rn ) but only on
the (dense) subspace consisting of functions ψ ∈ L2 (Rn ) such that xψ(x) is again square
integrable.
The momentum of the particle is associated with the self-adjoint operator p = −iℏ∇ (also p
is a vector-valued observable; each component p = −iℏ∂xj , with j = 1, . . . , n is an unbounded
self-adjoint operator). Here, ℏ is Planck’s constant, which we will henceforth assume to be
ℏ = 1. The expected value of the momentum is therefore given by the expectation
Z
⟨ψ, −i∇ψ⟩ = −i ψ(x)∇ψ(x)dx

The probability that the momentum is measured in a set B ⊂ Rn is given by ⟨ψ, χB (−i∇)ψ⟩.
Since the Fourier transform diagonalizes the derivative, we can write
Z Z
2
⟨ψ, χB (−i∇)ψ⟩ = χB (p)|ψ̂(p)| dp = |ψ̂(p)|2 dp
Z B

⟨ψ, −i∇ψ⟩ = p|ψ̂(p)|2 dp

In other words, |ψ̂(p)|2 is the probability density for finding momentum close to p. This
observation explains why it is important for ψ to be complex valued (while |ψ(x)|2 determines
the distribution of the position observable, it is the phase of ψ, the oscillations of ψ, which
contributes to |ψ̂(p)|2 and determines the distribution of the momentum observable).
The position operator is a multiplication operator (and thus already diagonal); its distri-
bution is determined directly by the probability density |ψ(x)|2 . The momentum operator is
a differential operator; it is diagonal in Fourier space; its distribution is determined by the
probability density |ψ̂(p)|2 . For a general self-adjoint operator A on L2 (Rn ), its distribution is

18
 P
determined by the spectral decomposition of A. Suppose namely that A = j λj Pψj , where
λj ∈ R are the eigenvalues of A and Pψj is the orthogonal projection onto the eigenvector ψj
of A. Then we can write, for a general wave function ψ ∈ L2 (Rn ) and a set B ⊂ R,
X
⟨ψ, Aψ⟩ = λj |⟨ψ, ψj ⟩|2
j
X X
⟨ψ, χB (A)ψ⟩ = χB (λj )|⟨ψ, ψj ⟩|2 = |⟨ψ, ψj ⟩|2
j j:λj ∈B

Hence, the observable A can take the values λj ; A takes the value λj with probability |⟨ψ, ψj ⟩|2 .
In general, the spectrum of A will not be purely discrete, it will also have a continuous
component. Denoting by (Eλ )λ∈R the projection-valued measure associated with A, we can
write Z
A = λdEλ

We find
Z
⟨ψ, Aψ⟩ = λd⟨ψ, Eλ ψ⟩
Z Z
⟨ψ, χB (A)ψ⟩ = χB (λ)d⟨ψ, Eλ ψ⟩ = d⟨ψ, Eλ ψ⟩
B

Heisenberg’s uncertainty principle. Consider a quantum system described on a Hilbert

space H. States of the systems are associated to vectors ψ ∈ H, normalized so that ∥ψ∥ = 1.
Since observable quantities are given by expectations of the form ⟨ψ, Aψ⟩, for self-adjoint
operators A, the vector ψ is only defined up to a phase (ie. ψ and eiθ ψ describe the same
state, if θ ∈ R is a constant).
Taking into account that ψ determines simultaneously the distribution of all observables
of the system, it follows that these distributions cannot be independent of each other. An
important consequence of this observation is Heisenberg’s uncertainty principle.
Theorem 2.2.1. Let A, B be two self-adjoint operators acting on a Hilbert space H and let
ψ ∈ H be a normalized (so that ∥ψ∥ = 1). We define the variance of A in the state ψ by
∆Aψ = ⟨ψ, (A − ⟨ψ, Aψ⟩)2 ψ⟩ = ⟨ψ, A2 ψ⟩ − ⟨ψ, Aψ⟩2
Then, we have
1
∆Aψ ∆Bψ ≥ |⟨ψ, [A, B]ψ⟩|2
4
e = A − ⟨ψ, Aψ⟩ and B
Proof. Replacing A and B by A e = B − ⟨ψ, Bψ⟩, we can assume that
⟨ψ, Aψ⟩ = ⟨ψ, Bψ⟩ = 0. Then
⟨ψ, [A, B]ψ⟩ = ⟨ψ, ABψ⟩ − ⟨ψ, BAψ⟩ = 2iIm ⟨ψ, ABψ⟩

19
This implies that
1/2 1/2
|⟨ψ, [A, B]ψ⟩| ≤ 2|⟨ψ, ABψ⟩| = 2|⟨Aψ, Bψ⟩| ≤ 2∥Aψ∥∥Bψ∥ = 2∆Aψ ∆Bψ

and therefore that

1
∆Aψ ∆Bψ ≥ |⟨ψ, [A, B]ψ⟩|2
4

Example. Consider a particle moving on R, take A = x and B = −i∂x . Then [A, B] = i.

It follows that
1
∆xψ ∆pψ ≥
4
In other words, x and p cannot be measured simultaneously with arbitrary precision. If ψ
concentrates very much around a position x0 (meaning that the variance ∆xψ is very small),
then ψ̂ is flat, meaning that the variance ∆pψ is large. Only commuting observables can be
measured simultaneously.
The Hamilton operator and the time evolution. In every quantum system there is an
observable that plays a particularly important role. The Hamilton operator, or the Hamilto-
nian, is a self-adjoint operator H on the Hilbert space H, corresponding to the energy of the
system. Through the Schrödinger equation

i∂t ψ(t) = Hψ(t) (2.22)

for ψ(t) ∈ H, the Hamilton operator generates the time evolution of the quantum system.
Notice that
d d
∥ψ(t)∥2 = ⟨ψ(t), ψ(t)⟩
dt dt
= i [⟨i∂(t)ψ(t), ψ(t)⟩ − ⟨ψ(t), i∂t ψ(t)⟩]
= i [⟨Hψ(t), ψ(t)⟩ − ⟨ψ(t), Hψ(t)⟩]
=0

since H is self-adjoint. Hence, the Schrödinger equation defines a unitary evolution on H,

preserving the norm. Denoting by e−iHt the unitary group generated by H, we can write the
solution of the Schrödinger equation (2.22) with initial data ψ(0) as ψ(t) = e−iHt ψ(0).
Suppose that φ ∈ H is a normalized eigenvector of H, ie Hφ = Eφ. Then, the solution of
the Schrödinger equation with initial data ψ(0) = φ is simply given by ψ(t) = e−iEt φ. This
implies that ⟨ψ(t), Aψ(t)⟩ = ⟨φ, Aφ⟩, independently of t. For this reason, eigenvectors of H
are referred to as stationary states.
Consider now a quantum particle in Rn , described on the Hilbert space L2 (Rn ). As ex-
plained above, the observables depending on the position are associated with multiplication

20
operators, while the momentum is associated with the differential operator p = −i∇. Recall-
ing the classical energy H = p2 /(2m) + V (x), we can guess that the Hamilton operator for
this system takes the form
∆
H=− + V (x) (2.23)
2m
where ∆ = ∇ · ∇ is the Laplace operator on L2 (Rn ). Operators of the form (2.23) are known
as Schrödinger operators. Understanding their spectral properties is crucial to understand
properties of the time-evolution they generate.
Conservation laws. Let A be a self-adjoint operator on the Hilbert space H. Let ψ(t)
denote the solution of the Schrödinger equation

i∂t ψ(t) = Hψ(t)

The variation of the expectation of the observable A has the form

d
i ⟨ψ(t), Aψ(t)⟩ = ⟨ψ(t), [A, H]ψ(t)⟩
dt
where [A, H] = AH − HA is the commutator between the operators A and H. This should
be compared with (2.19) in the classical setting; the commutator replaces here the Poisson
brackets.
In particular, for an observable A satisfying [A, H] = 0, the expectation ⟨ψ(t), Aψ(t)⟩
remains constant over time.
For example, if H = −∆/2m on L2 (Rn ) (this is the Hamilton operator of a free particle
moving in Rn ), then [H, −i∇] = 0, ie. momentum is conserved. Similarly, the Hamilton
operator H = −∆/2m + V (|x|) acting on L2 (R3 ), with a radial potential (ie. V only depends
on the ∥x∥), then [H, x ∧ −i∇] = 0, ie. (each component of) the angular momentum operator
L = −ix ∧ ∇ is conserved. We will use this conservation law below, to compute the spectrum
of the Hamilton operator describing the hydrogen atom.
Measurements in quantum mechanics. What happens to the state ψ ∈ H of a quantum
system when we perform a measurement of an observable A (a self-adjoint operator on A)?
Measurements in quantum mechanics correspond to interactions with classical devices. For
example, to measure the position of an electron we let it interact with a detector. We do not
want to describe the detector as part of the quantum system; we want to consider it instead
as a classical device, obeying the rules of classical mechanics. As a consequence, we cannot
describe the measurement process through the Schrödinger equation. We need to establish
new rules to describe what happens to the quantum system when it undergoes a measurement.
This is still a debated subject in quantum mechanics (the problem of foundations of quantum
mechanics). The most common interpretation of the measurement process (the Copenhagen
interpretation) is as follows. Suppose that a quantum system is described by a normalized

21
vector ψ in a Hilbert space H. We perform a measurement of an observable associated with the
self-adjoint
P operator A. For simplicity, let us assume that A has the spectral representation
A = j λj Pφj , where λj are the eigenvalues and φj the corresponding normalized eigenvectors
of A. As explained above a measurement of A produces the outcome λj with probability
|⟨ψ, φj ⟩|2 . If the measurement of A produces the outcome λj , then, after the measurement,
the system is described by the eigenvector φj . More precisely, as a result of the measurement,
the vector ψ collapses into the vector Pj ψ/∥Pj ψ∥, where Pj denotes the orthogonal projection
onto the eigenspace associated with the eigenvalue λj (we have Pj ψ/∥Pj ψ∥ = φj , up to an
irrelevant phase, if the eigenvalue λj is not degenerate).
Example: the Stern-Gerlach experiment. In quantum mechanics, elementary particles
have an internal degree of freedom, known as spin. The spin is an angular momentum; it
has three components, σx , σy , σz . Experimentally, the existence of a spin can be verified
with a Stern-Gerlach experiment, using the fact that spins couple to magnetic fields. In the
Stern-Gerlach experiment, neutral particles (in the original experiment in 1922, silver atoms)
are sent through a non-constant magnetic field, which deflects them according to their spin.
Particles are then detected on a screen. The outcome of the experiment was, at the beginning,
surprising (it led to the Nobel Prize for Stern in 1943); the screen does not show a continuous
distribution but, instead, it reveals discrete accumulation points, showing that the spin (more
precisely, the component of the spin in the direction parallel to the applied magnetic field) is
quantized. Particles are associated with a spin number n ∈ N/2; each spin component has
then 2n + 1 possible values. In particular, for spin 1/2 particles (like electrons, protons), each
spin component has two possible values (spin up or spin down).
If we forget about all other degrees of freedom, we can describe the setting of the Stern-
Gerlach experiment, restricting for simplicity our attention to spin 1/2 particles, as a quantum
system on C2 . If the magnetic field in the Stern-Gerlach apparatus points in the z-direction,
it is convenient to choose the vectors (1, 0), (0, 1) ∈ C2 to describe particles with the z-
component of the spin pointing up and, respectively, down. In this representation of the
system, the z-component of the spin is associated with the self-adjoint matrix
 
1 0
σz = (2.24)
0 −1

The x- and the y-component of the spin are associated instead with the matrices
   
0 1 0 −i
σx = , σy = (2.25)
1 0 i 0

The matrices σx , σy , σz are known as Pauli matrices. They satisfy the commutation relation

[σx , σy ] = 2iσz (2.26)

22
and its cyclic permutations. In particular, notice that spins in different directions do not
commute. Notice that we did not try to justify the choice of the matrices (2.24), (2.25). In fact,
as we will discuss later, relations between operators associated to different spin components
(in particular the commutator relations (2.26)), are determined by the properties of the group
of rotations SO(3).
The Stern Gerlach experiment is a measurement of the z-component of the spin. If the
initial state of the system is described by the vector ψ = (α, β) ∈ C2 , normalized so that
|α|2 +|β|2 = 1, the measurement of σz will give the value +1 with probability |α|2 and the value
−1 with probability |β|2 . In other words, after going through the Stern-Gerlach appartus, the
particle will move up and hit the upper accumulation point on the screen with probability |α|2
and it will move down and hit the lower accumulation point on the screen with probability
|β|2 . If the measurement of σz gives the value +1, after going through the apparatus the
particle is described by the vector (1, 0). In this case, if we let the particle go through a
second Stern-Gerlach experiment, again with magnetic field pointing in the z direction, the
particle will show spin up with probability one. On the other hand, if the measurement of σz
gives the value −1, after going through the apparatus the particle is described by the vector
(0, 1). In this case, a second measurement of σz will give again the value −1 with probability
one. It is also possible to insert, after the first Stern-Gerlach apparatus measuring the z-
component of the spin, a Stern-Gerlach apparatus measuring, say, the x-component of the
spin. Let us assume, for example, that the first measurement of σz returns the value +1 and
therefore that after the first measurement the system is described by the vector (1, 0). The
expectation value of σx is given by
   
0 1
(1, 0), (1, 0) = 0
1 0

Since σx2 = 1, the variance of σx is given by ∆σx = 1. A different approach to reach the same
conclusion consists in noticing that
√ the√matrix σx has √ eigenvalues ±1 associated with
√ the two
the normalized eigenvectors (1/ 2, 1/ 2) and (1/ 2, −1/ 2). Since
1 √ √ 1 √ √
(1, 0) = √ (1/ 2, 1/ 2) + √ (1/ 2, −1/ 2)
2 2
we conclude that the measurement of σx gives the value +1 and the value −1 both with
probability 1/2 (this explains also why the expectation of σx vanishes and its variance equals
one). This result shows a well-known phenomenon in quantum systems. Non-commuting
observables cannot be measured simultaneously. It is impossible to know the precise value
of σz and of σx at the same time (mathematically, this means that two non-commuting self-
adjoint operators cannot be diagonalized simultaneously).
Density matrices and mixed states. As explained above a quantum mechanical system is
described on a Hilbert space H. States of the system are associated with normalized vectors

23
ψ ∈ H. Observables are associated with self-adjoint operators A acting on H. The expectation
value for the measurement of the observable A in the state ψ is then given by the inner product
⟨ψ, Aψ⟩. The same state can be described through the orthogonal projection onto ψ, denoted
by
γ = |ψ⟩⟨ψ|
The expectation of the observable A is then given by

⟨ψ, Aψ⟩ = trAγ

In this case, describing the system through γ is fully equivalent to the description given by ψ
(the description through γ shows that eiθ ψ and ψ describe the same state). The description
through γ has the advantage that it can be easily extended to mixed states.
Definition 2.2.2. A density matrix over the Hilbert space H is a non-negative (in particular
self-adjoint) trace class operator γ on H, with trγ = 1. A trace class operator is a com-
pact operator, with the property that the sequence {λj }j of its eigenvalues
P is summable, ie.
2
P
j |λj | < ∞ . The trace of a trace class operator γ is defined by trγ = j λj

Given a density matrix γ over H, we can always decompose it as

X
γ= λj |ψj ⟩⟨ψj |
j
P
where {ψj }j is an orthonormal basis in H and where λj ≥ 0 for all j, with j λj = 1. Such a
density matrix describes the system, if it is in the state ψj with probability λj . If the density
matrix γ is not an orthogonal projection, it is said to describe a mixed state (on the other
hand, if γ = |ψ⟩⟨ψ| is an orthogonal projection, it is said to describe the pure state, which
can be equivalently be described by the vector ψ ∈ H).
It is important to observe the difference between pure states, given by linear combinations
of vectors in H and taking instead mixed states, in which case the density matrix is given by
a (convex) linear combination of orthogonal projections. Suppose for example ψ1 , ψ2 are two
orthogonal normalized vectors in H. Then
1
ψ = √ (ψ1 + ψ2 ) (2.27)
2
is again a normalized vector in H. The expectation of an observable A in the pure state ψ is
given by
1
⟨ψ, Aψ⟩ = [⟨ψ1 , Aψ1 ⟩ + ⟨ψ2 , Aψ2 ⟩ + ⟨ψ1 , Aψ2 ⟩ + ⟨ψ2 , Aψ1 ⟩] (2.28)
2
2
Recall that the spectrum of compact operators consists of discrete eigenvalues with finite multiplicity
which can only accumulate at 0.

24
On the other hand, consider the mixed state, described by the density matrix
1
γ= [|ψ1 ⟩⟨ψ1 | + |ψ2 ⟩⟨ψ2 |] (2.29)
2
The expectation of A in the mixed state with density matrix γ is given by
1
trAγ = [⟨ψ1 , Aψ1 ⟩ + ⟨ψ2 , Aψ2 ⟩]
2
In contrast to (2.28), there are here no “interference” terms of the form ⟨ψ1 , Aψ2 ⟩, ⟨ψ2 , Aψ1 ⟩.
We call (2.27) a coherent superposition of the quantum states ψ1 , ψ2 . On the other hand,
(2.29) is called an incoherent superposition of quantum states.
Mixed states are needed, when there is no complete information about the system. For
example, equilibrium states at positive temperature are mixed.

2.3 The Harmonic Oscillator

An harmonic oscillator with frequency ω > 0 can be described by the self-adjoint operator
Hharm = −d2 /dx2 + ω 2 x2 acting on L2 (R).
The spectrum of Hharm . Since (H + 1)−1 is a compact operator, the spectrum of H is
discrete and can only accumulate at infinity. To find the eigenvalues of H, we define the
operators
1 √
 
1 d
a± = √ ω x ∓ i√
2 ω dx
Note that a+ = a∗− . A simple computation shows that [a− , a+ ] = 1 and that H = ω(2N + 1)
where we set N = a+ a− . Next , we observe that
[N, a± ] = ±a±
Hence, if N ψ = nψ for a ψ ̸= 0, then N a± ψ = (n ± 1)a± ψ. Moreover, we have ∥a+ ψ∥2 =
⟨ψ, a− a+ ψ⟩ = (n + 1)∥ψ∥2 and ∥a− ψ∥2 = ⟨ψa+ a− ψ⟩ = n∥ψ∥2 . This implies that n ≥ 0 and
therefore that σ(N ) ⊂ N (because if n ̸∈ N was an eigenvalue, then applying a− sufficiently
many times we would find a negative eigenvalue of N ). To prove that σ(N ) = N, we need
to find ψ0 ̸= 0 with N ψ0 = 0 (because then we can construct eigenvectors for all n ∈ N by
applying a+ sufficiently many times). If N ψ0 = 0, we must have a− ψ0 = 0 (if a− ψ0 ̸= 0, it
would be an eigenvector of N with eigenvalue −1). a− ψ0 = 0 implies that
√ 1
ωxψ0 (x) = √ ψ0′ (x)
ω
which has the unique (up to an irrelevant phase) normalized solution
2 /2
ψ0 (x) = (ω/π)1/4 e−ωx . (2.30)

25
Starting from ψ0 , we can construct eigenvectors ψn associated with the eigenvalue n ∈ N of
N , for all n ∈ N, setting
1
ψn (x) = √ an+ ψ0 (x)
n!
We find that
1 √ 2
ψn (x) = √ (ω/π)1/4 Hn ( ωx)e−ωx /2 (2.31)
n
2 n!
where Hn is the so called Hermite polynomial of degree n, given by
 n n
x2 /2 d 2 2 d −x2
Hn (x) = e x− e−x /2 = (−1)n ex e
dx dxn
It is easy to check that the eigenvectors {ψn }n∈N build an orthonormal basis of the Hilbert
space L2 (R) (since Hharm is self-adjoint and has simple pure-point spectrum, the set of eigen-
vectors must build an orthonormal basis). Summarizing, the Hamilton operator Hharm has
the spectrum
σ(Hharm ) = σpp (Hharm ) = {ω(2n + 1) : n ∈ N}
Each eigenvalue λn = ω(2n + 1) is simple and it is associated with the normalized eigenvec-
tor ψn in (2.31). It is worth remarking that the difference λn+1 − λn between consecutive
eigenvalues of H 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. a+ is therefore called a creation
operator, its adjoint a− an annihilation operator.
Properties
√ −1 n of eigenvectors. Let us have a closer look at the eigenvectors ψn =
( n!) a+ ψ0 . A simple computation shows that the expectation of position and momentum
in the state ψn vanish. In fact, denoting by x̂ the multiplication operator with x,
1
⟨ψn , x̂ψn ⟩ = √ ⟨ψn , (a+ + a− )ψn ⟩
2ω
1
= √ ⟨an+ ψ0 , (a+ + a− )an+ ψ0 ⟩
n! 2ω
2
= √ Re ⟨an+ ψ0 , an+1
+ ψ0 ⟩ = 0
n! 2ω
and similarly, with the momentum operator p̂ = id/dx,
p
⟨ψn , p̂ψn ⟩ = i ω/2⟨ψn , (a− − a+ )ψn ⟩
√
ω
= i √ ⟨an+ ψ0 , (a− − a+ )an+ ψ0 ⟩
n! 2
√
2ω
= Im ⟨an+ ψ0 , an+1
+ ψ0 ⟩ = 0
n!
26
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 = ⟨ψn , x̂2 ψn ⟩

1
= ⟨an+ ψ0 , (a+ + a− )2 an+ ψ0 ⟩
2ωn!
1
= ⟨an+ ψ0 , (a+ a− + a− a+ )an+ ψ0 ⟩
2ωn!
1 1 1 1
= ⟨an+ ψ0 , (2a− a+ − 1)an+ ψ0 ⟩ ∥an+1 2
+ ψ0 ∥ − ∥an+ ψ0 ∥2 = (n + 1/2)
2ωn! ωn! 2ωn! ω
Similarly, we obtain

∆pψn = ⟨ψn , p̂2 ψn ⟩

ω n
=− ⟨a ψ0 , (a− − a+ )2 an+ ψ0 ⟩
2n! +
ω n
= ⟨a ψ0 , (a+ a− + a− a+ )an+ ψ0 ⟩ = ω(n + 1/2)
2n! +
We conclude that
∆xψn ∆pψn = (n + 1/2)2
Observe that for n = 0, i.e. in the state ψ0 with the smallest possible energy (the vacuum
state, with no energy quanta), the product of the variances is minimal, according to Heisenberg
uncertainty principle. For larger n, on the other hand, the uncertainty in the state ψn grows.
Coherent states. Instead of considering the eigenvectors ψn , there is another family of
states in the Hilbert space L2 (R3 ) which minimize the uncertainty and are therefore closer to
classical states. Such states are known as coherent states. For α ∈ C, we define the coherent
state
ψ(α) = exp(αa+ − ᾱa− )ψ0
with ψ0 as defined in (2.30). Since exp(αa+ − ᾱa− ) is a unitary operator, it follows that
∥ψ(α)∥ = 1 for all α ∈ C. Since moreover [a− , a+ ] = 1 commutes with a− , a+ , we find that
2 /2
ψ(α) = e−|α| eαa+ e−ᾱa− ψ0
2 /2 2 /2
X 1 2
X 1 (2.32)
= e−|α| eαa+ ψ0 = e−|α| αn an+ ψ0 = e−|α| /2 √ α n ψn
n≥0
n! n≥0
n!

Hence, ψ(α) is a linear combination of all eigenvectors ψn ; with different probabilities, the
coherent state ψα contains different numbers of energy quanta. In (2.32), we used the formula

eA+B = eA eB e−1/2[A,B] (2.33)

27
valid under the assumption that the commutator [A, B] commutes with A and with B. To
prove this formula, we notice first that, under the assumption that [A, [A, B]] = [B, [A, B]] = 0,
we have, for any t > 0,
Z t Z 1
−tA d sA −sA
tA
e Be =B+ ds e Be =B+ ds esA [A, B]e−sA = B + t[A, B]
0 ds 0

Hence, we can define F (t) = et(A+B) e−tA e−tB and compute the derivative

F ′ (t) = et(A+B) Be−tA e−tB − et(A+B) e−tA Be−tB = −t[A, B]F (t)
2
With the initial data F (0) = 1, we find the solution F (t) = e−t [A,B]/2 and thus F (1) =
e−[A,B]/2 , which proves (2.33)
An important property of coherent states ψ(α) is the fact that they are eigenvectors of the
annihilation operator A− . This follows from the observation that

e−(αa+ −ᾱa− ) a− eαa+ −ᾱa− = a− + α

e−(αa+ −ᾱa− ) a+ eαa+ −ᾱa− = a+ + ᾱ
Hence
a− ψ(α) = a− eαa+ −ᾱa− ψ0 = eαa+ −ᾱa− (a− + α)ψ0 = αψ(α)
for all α ∈ C. With this remark, we can compute the expectations
r
1 2
x̄ψ(α) = √ ⟨ψ(α), (a+ + a− )ψ(α)⟩ = Re α
2ω ω

r
p ω
p̄ψ(α) = i ω/2⟨, ψ(α), (a− − a+ )ψ(α)⟩ = − Im α
2
The variances are instead given by
1
∆xψ(α) = ⟨ψ(α), x̂2 ψ(α)⟩ − ⟨ψ(α), x̂ψ(α)⟩2 =
2ω
2 ω
∆pψ(α) = ⟨ψ(α), p̂ ψ(α)⟩ =
2
which implies that ∆xψ(α) ∆pψ(α) = 1/4, independently of α. This shows that, in coherent
states, the products of the variances is minimal; this makes them similar to classical states.
This last fact is also confirmed by looking at the time-evolution of a coherent state. To
compute the time-evolution of a coherent state, we observe that
d ita+ a−
e a− e−ita+ a− = ieita+ a− [a+ a− , a− ]e−ita+ a− = −ieita+ a− a− e−ita+ a−
dt
28
Hence
eita+ a− a− e−ita+ a− = e−it a−
Since Hharm = ω(2a+ a− + 1), this implies that

eitHharm a− e−itHharm = e−2itω a−

and similarly that

eitHharm a+ e−itHharm = e2itω a+
Hence
2itω a −ᾱe−2itω a
e−iHharm t ψ(α) = e−iHharm t eαa+ −ᾱa− ψ0 = eαe + −
e−iHharm t ψ0 = e−iωt ψ(e2itω α)

which means that, up to the irrelevant phase e−iωt , the time evolution of a coherent state is
again coherent. The evolution of the expectation of position and momentum in e−iHharm t ψ(α)
are therefore given by
r r
2 2
x̄e−iHharm t ψ(α) = Re e2itω α = [cos(2tω)Re α − sin(2tω)Im α]
ω ω
and by
r r
ω ω
p̄e−iHharm t ψ(α) =− Im e2itω α = − [cos(2tω)Im α + sin(2tω)Re α]
2 2
Hence, the expectation of position and momentum follow the evolution of a classical harmonic
oscillator. Coherent states describe therefore quantum states with minimal variances moving
along classical trajectories.
Let us conclude this section with the remark that coherent states are complete, ie.
Z
1
dαdᾱ |ψ(α)⟩⟨ψ(α)| = 1L2 (R) (2.34)
π

where we used the notation |ψ(α)⟩⟨ψ(α)| to denote the orthogonal projection onto the state
ψ(α) but they are not orthonormal, since
2
|⟨ψ(α), ψ(β)⟩|2 = e−|α−β| (2.35)

To prove (2.34), we let Z

1
A= dαdᾱ |ψ(α)⟩⟨ψ(α)|
π

29
and we compute, using (2.32),
Z
1
⟨ψn , Aψm ⟩ = dαdᾱ⟨ψn , ψ(α)⟩⟨ψ(α), ψm ⟩
π
Z
1 2
= √ √ dαdᾱe−|α| αn ᾱm
π n! m!
This implies that ⟨ψn , Aψm ⟩ = 0 if n ̸= m and that
2 ∞ 2n+1 −r2 1 ∞ n −x
Z Z Z
1 2n −|α|2
⟨ψn , Aψn ⟩ = dαdᾱ|α| e = r e dr = x e dx = 1
πn! n! 0 n! 0
Hence ⟨ψn , Aψm ⟩ = δnm for all n, m ∈ N, which already shows that A = 1. To prove (2.35),
we use again (2.32) to compute
1
X ᾱn β m
2 +|β|2 )
⟨ψ(α), ψ(β)⟩ = e− 2 (|α| √ √ δn,m
n,m≥0
n! m!
1 2 2
X (ᾱβ)n 1 2 2
= e− 2 (|α| +|β| ) = e− 2 (|α| +|β| −2ᾱβ)
n≥0
n!

which implies (2.35).

2.4 The Hydrogen Atom

Consider a nucleus with electric charge +1, fixed at the origin (this is a good approximation,
because the nucleus is much heavier than the electron). The electron has charge −1 and moves
in R3 in the electrostatic field generated by the nucleus, with the potential V (x) = −1/|x|.
In appropriate units, the Hamilton operator describing the electron has the form
1
H = −∆ −
|x|
and acts on L2 (R3 ). Since −1/|x| is relatively compact w.r.t. −∆, the operator H has
the essential spectrum σess (H) = σess (−∆) = [0; ∞). The attractive potential, however,
generates additional negative eigenvalues, which only accumulate at 0. In this section, we aim
at computing these eigenvalues. To reach this goal, we observe that, since the potential is
radial, the Hamilton operator H commutes with the angular momentum operator L = −ix∧∇.
As remarked above, this implies that the expectation of L is preserved by the time evolution.
The fact that [H, L] = 0 can also be used to diagonalize H. The three components of the
angular momentum operator L = −ix ∧ ∇ do not commute among each others. In fact, we
find the relations
[Lk , Lj ] = iεkjℓ Lℓ (2.36)

30
for all k, j, ℓ ∈ {1, 2, 3}. Each component, though, commutes with L2 , since
3
X 3
X X
2 2
[Li , L ] = [Li , Lj ] = [Li , Lj ]Lj + Lj [Li , Lj ] = i εijk (Lk Lj + Lj Lk ) = 0
j=1 j=1 j

because of the antisymmetry of the tensor ε. This means that we can diagonalize H, L2
and one of the component of L, for example L3 , simultaneously. In other words, we can first
diagonalize L3 , L2 and then look for eigenvectors of H separately in the joint eigenspaces of
L2 and L3 (which are invariant spaces for H, since [H, L] = 0). We start by determining the
spectrum and the eigenspaces of L2 , L3 .
Spectrum of angular momentum operators. Observe that angular momentum opera-
tors, which generate rotations around axis going through the origin, do not act in the radial
direction. To extract spectral information about the operators L2 and L3 , it is therefore useful
to think of these operators as operators acting on the Hilbert space L2 (S 2 ), where S 2 is the
two-dimensional sphere, embedded in R3 . On this Hilbert space, L2 and L3 have discrete
spectrum.
Since L2 ≥ 0, eigenvalues of L2 must be positive. We look therefore for joint eigenvectors
ϕℓ,m of L2 and L3 , with L2 ϕℓ,m = ℓ(ℓ + 1)ϕℓ,m and L3 ϕℓ,m = mϕℓ,m , for some m ∈ R and ℓ ≥ 0.
From the operator inequality L2 ≥ L23 , we immediately conclude that m2 ≤ ℓ(ℓ + 1). Consider
now the operator L± = L1 ± iL2 , so that L∗+ = L− . From the commutation relations (2.36),
we obtain that
[L3 , L± ] = ±L± , [L2 , L± ] = 0
and therefore, that

L3 L± ϕℓ,m = L± L3 ϕℓ,m + [L3 , L± ]ϕℓ,m = (m ± 1)L± ϕℓ,m

and
L2 L± ϕℓ,m = L± L2 ϕℓ,m = ℓ(ℓ + 1)L± ϕℓ,m
In other words, applying the operator L± to the eigenvector ϕℓ,m we either get L± ϕℓ,m = 0 or
we obtain a joint eigenvector of L3 and L2 associated with the eigenvalue m±1 and ℓ(ℓ+1). If
we start with any ϕℓ,m and we continue to apply L+ , in order for m2 ≤ ℓ(ℓ + 1) to be satisfied,
there must be a maximal m̄ such that L+ ϕℓ,m̄ = 0. Since

L− L+ = (L1 − iL2 )(L1 + iL2 ) = L21 + L22 + i(L1 L2 − L2 L1 ) = L2 − L23 − L3

L+ ϕℓ,m̄ = 0 implies that ℓ(ℓ + 1) − m̄2 − m̄ = 0 and therefore that m̄ = ℓ. Similarly, if we

start with ϕℓ,m and we apply L− , there must be a minimal m such that L− ϕℓ,m = 0. From
L+ L− = L2 − L23 + L3 , the condition L− ϕℓ,m = 0 implies that m = −ℓ. Hence, if there exists
a joint eigenvector ϕℓ,m0 of L2 and L3 for a ℓ ≥ 0 and for an m0 ∈ R with m20 ≤ ℓ(ℓ + 1), then
we find 2ℓ + 1 eigenvectors ϕℓ,m , with m = −ℓ, −ℓ + 1, . . . , ℓ. This is of course only possible if

31
2ℓ + 1 ∈ N. Hence, only based on the commutation relations (2.36), we reach the conclusion
that the eigenvalues of L2 can only have the form ℓ(ℓ + 1), for an ℓ ∈ {0, 1/2, 1, 3/2, . . . }. If
ℓ(ℓ+1) is an eigenvalue of L2 , we can construct 2ℓ+1 joint eigenvectors ϕℓ,m of the operators L2
and L3 , such that L2 ϕℓ,m = ℓ(ℓ + 1)ϕℓ,m and L3 ϕℓ,m = mϕℓ,m , with m = −ℓ, −ℓ + 1, . . . , ℓ − 1, ℓ.
Spherical Harmonics. To obtain more precise information on the joint eigenvectors of
L2 and L3 , it is useful to parametrize the sphere S 2 with spherical coordinates (θ, φ) ∈
(0; π) × (0; 2π). Expressed in these coordinates, the operators L3 , L± and L2 take the form

d
L3 = −i
dφ
 
±iφ d d
L± = e ± + i cot θ (2.37)
dθ dφ
1 d d 1 d2
L2 = − sin θ −
sin θ dθ dθ sin2 θ dφ2

Notice here that L2 is the Laplace-Beltrami operator −∆S 2 on the sphere S 2 .

Using these expressions, we look for solutions ϕℓ,m of the eigenvalue equations

L2 ϕℓ,m = ℓ(ℓ + 1)ϕℓ,m , and L3 ϕℓ,m = mϕℓ,m

for ℓ ∈ N/2 and m ∈ {−ℓ, −ℓ + 1, . . . , ℓ}. To solve these equations, we first look for
the eigenfunction ϕℓ,ℓ , with the maximal value of m = ℓ (from ϕℓ,ℓ , we can construct all
other eigenvectors applying the operator L− ). The equation L3 ϕℓ,ℓ = ℓϕℓ,ℓ has the solution
ϕℓ,ℓ (θ, ϕ) = eiℓφ f (θ) for a function f depending only on the angle θ. This already implies that
ℓ ∈ N. Half-integer values of ℓ (which are not excluded by the commutation relations (2.36),
lead to functions that are not 2π-periodic in φ, i.e. functions that are not defined on the
sphere S 2 ). From now on, we assume that ℓ ∈ N. To determine the θ-dependence of ϕℓ,ℓ , we
solve the equation L+ ϕℓ,ℓ = 0, which gives f ′ (θ) − ℓ cot θf (θ) = 0 and therefore

ϕℓ,ℓ (θ, φ) = const · eiℓφ (sin θ)ℓ

Applying recursively the operator L− , we find

ϕℓ,m (θ, φ) = const · Lℓ−m

− ϕℓ,ℓ (θ, φ)

for m = −ℓ, −ℓ + 1, . . . , ℓ. Applying this procedure to all integer values of ℓ ∈ N, we obtain

the spherical harmonics {Yℓ,m : ℓ ∈ N, m = −ℓ, −ℓ + 1, . . . , ℓ}. These functions are given by
s
2ℓ + 1 (ℓ − m)!
Yℓ,m (θ, φ) = Pℓ,m (cos θ)eimφ
4π (ℓ + m)!

32
where Pℓ,m (z) is the associated Legendre polynomial defined by

(−1)m 2 m/2 d
ℓ+m
Pℓ,m (z) = (1 − z ) (z 2 − 1)ℓ
2ℓ ℓ! dz ℓ+m
satisfying the recursion formula

(ℓ − m + 1)Pℓ+1,m (z) = (2ℓ + 1)zPℓ,m (z) − (ℓ + m)Pℓ−1,m (z)

The first spherical harmonics are given by

1
Y00 (θ, φ) = √
4π
r r r
3 3 3
Y1,−1 (θ, φ) = sin θe−iφ , Y1,0 (θ, φ) = cos θ, Y1,1 (θ, φ) = − sin θeiφ
8π 4π 8π
r r
15 15
Y2,±2 (θ, φ) = sin2 θe±2iφ , Y2,±1 (θ, φ) = ∓ sin θ cos θe±iφ ,
32π 8π
r
5
Y2,0 (θ, φ) = (3 cos2 θ − 1)
16π
They are normalized so that
Z
Y ℓ,m (θ, φ)Yℓ′ ,m′ (θ, φ) sin θdθdφ = δℓ,ℓ′ δm,m′

for all ℓ ∈ N and all m = −ℓ, . . . , ℓ. Spherical harmonics also satisfy the completeness relation
ℓ
X X
Y ℓ,m (θ, φ)Yℓ,m (θ′ , φ′ ) = δ(φ − φ′ )δ(cos θ − cos θ′ ) (2.38)
ℓ∈N m=−ℓ

and build therefore an orthonormal basis for L2 (S 2 ), consisting of joint eigenvectors of L2 and
L3 .
Remark. Although half-integers values of ℓ do not appear in the spectrum of the total
angular momentum on L2 (S 2 ), these values are not excluded by the commutation relations
(2.36). For this reason, they can and they do occur in the spectrum of generators of rotations
that do not act on the position space R3 . As first proposed by Pauli, quantum particles carry
an intrinsic angular momentum, known as spin (classically, the spin, or “Eigendrehimpuls”
in German, generates rotation around the center of mass of a composite particle; the Stern-
Gerlach experiment discussed at the beginning of this class implies that quantum particles
carry a spin, even though they appear as point-particles). If S denotes the spin-operator,
S 2 can now have the eigenvalues s(s + 1) for any s ∈ {0, 1/2, 1, 3/2, . . . }; here, half-integer

33
values are also allowed (but since S 2 commutes with every other observable, a so-called super
selection rule holds and for each particle there can be only one value of s; for example, electrons
are spin 1/2 particles, meaning that s = 1/2).
Energy levels of the hydrogen atom. With (2.37), we write the Laplace operator in
spherical coordinates as
1 d2 1 1 d2 1 2
−∆ = − r − ∆S 2 = − r + L
r dr2 r2 r dr2 r2
where L2 = 3j=1 L2j is the total angular momentum operator discussed in the previous section.
P
We can therefore rewrite the eigenvalue equation Hψ = Eψ as
1 d2
 
1 2 1
− r + 2L − ψ = Eψ
r dr2 r r
To solve this equation, we separate variables by setting
g(r)
ψ(r, θ, φ) = Yℓ,m (θ, φ).
r
We obtain that
ℓ(ℓ + 1) 1
g ′′ (r) − 2
g(r) + g(r) = −Eg(r) (2.39)
r r
Separating variables is possible because, by symmetry, H is invariant with respect to rotations
and therefore it commutes with all components of the angular momentum operator L. This
implies that H can be diagonalized separately within each joint eigenspace of L2 and L3 .
From (2.39) we observe that, for large r ≫ 1, g ′′ (r) ≃ −Eg(r). If E > 0, all solutions
of this equation oscillate, and we will not be able to find eigenvectors in L2 (R3 ). Hence,
eigenvalues of H must be negative. For E < 0, excluding exponentially increasing factors, we
must have √
g(r) ≃ e− |E|r ,
for r ≫ 1. For r → 0, on the other hand, (2.39) can be approximated by the simplified
equation g ′′ (r) ≃ ℓ(ℓ + 1)g(r)/r2 . This equation has the two solutions rℓ+1 and r−ℓ . For
ℓ ̸= 0, the choice g(r) ≃ r−ℓ for r close to zero would lead to ψ(r, θ, φ) ≃ r−ℓ−1 Yℓ,m (θ, φ),
which is not square integrable, in three dimensions. Also for ℓ = 0, the choice g(r) ≃ 1 is not
relevant, since it would imply that ψ(r, θ, φ) ≃ r−1 Yℓ,m (θ, φ) and therefore that ∆ψ ≃ δ(r),
which cannot be compensated by the 1/r potential. We conclude that relevant solutions to
(2.39) behave like g(r) ≃ rℓ+1 close to zero. The study of the behavior of g at zero and at
infinity motivates the following power series ansatz for g:
√ X
g(r) = e− |E|r rℓ+1 cj r j
j≥0

34
Inserting in (2.39) we find the recursive relation
p
cj+1 2 |E|(ℓ + j + 1) − 1
=
cj (j + 1)(j + 2ℓ + 2)
If the power series is infinite, we would find the asymptotic behavior
p
cj+1 |E|
≃2
cj j+1
for large j. This would correspond to an asymptotic behavior
X
p
X (2 |E|r)j √
j
cj r ≃ = e2 |E|r
j j
j!
√
Even after multiplying with e− |E| , the function g would be exponentially increasing at in-
finity, hence the corresponding ψ(r, θ, φ) = g(r)r−1 Yℓ,m (θ, φ) would not be in L2 (R3 ). We
conclude
p that the power series must be finite. In other words, there must be nr ∈ N such that
2 |E|(ℓ + nr + 1) = 1 which leads to the eigenvalue
1
E=−
4(ℓ + nr + 1)2
and to the corresponding eigenvector
g(r)
ψ(r, θ, φ) = Yℓ,m (θ, φ)
r
with
√ nr
X
− |E|r ℓ+1
g(r) = e r cn r n
n=0
It is useful to introduce the quantum number n = ℓ + nr + 1. We have shown that, for every
n ∈ N\{0}, every ℓ = 0, 1, . . . , n − 1, every m = −ℓ, . . . , ℓ, we find a eigenvector
ψn,ℓ,m (x) = fn,ℓ (r)Yℓ,m (θ, φ)
where
gn,ℓ (r) √ n−ℓ−1
X
fn,ℓ (r) = = e− |E|r rℓ cj r j
r j=0

such that Hψn,ℓ,m = En ψn,ℓ,m , with the eigenvalue En = −1/(4n2 ). Hence, σ(H) = {En }n∈N ∪
[0; ∞). Note that the eigenvalue En has the degeneracy
n−1
X
(2ℓ + 1) = n2
ℓ=0

35
Part of the degeneracy was expected; since [Lj , H] = 0 for j = 1, 2, 3 we also have [L± , H] = 0
and therefore the energy cannot depend on m (because the value fo m can be changed applying
L± ). The additional degeneracy (namely the fact that, if n is fixed, energies do not depend on
ℓ) is known as accidental degeneracy (although also this degeneracy can be explained in terms
of symmetries; it is related with the conservation of the so-called Lenz vector). The smallest
eigenvalue of the hydrogen atom, also known as the ground state energy of the hydrogen atom,
is given by E1 = −1/4 (in physical units, we find E1 = −13.605eV = −2.18 · 10−18 Joule).
This is the energy needed to ionize an hydrogen atom at rest.

2.5 Schrödinger Operators: Existence of Stationary States

So far we considered a couple of examples of Hamilton operators (the harmonic oscillator, the
hydrogen atom in three dimensions) whose eigenvalues and eigenvectors could be computed
more or less explicitly. For a general choice of the potential V (x), we know (if V (x) is relatively
compact w.r.t. −∆) that σess (H) = [0; ∞), but it is impossible to determine explicitly the
eigenvalues of H = −∆ + V (x). 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.
We consider a quantum system in n dimensions, described on L2 (Rn ) by the Hamilton
operator H = −∆ + V (x), assuming for now only that V ∈ Lsloc (Rn ), for a 1 ≤ s ≤ ∞
(stronger conditions will come later). We consider the quadratic form associated with H,
defining the energy functional
Z Z
ε(ψ) = ⟨ψ, Hψ⟩ = dx |∇ψ(x)| + dx V (x)|ψ(x)|2
2
(2.40)

We are going to establish conditions that guarantee that the functional ε attains a minimum
on the unit sphere {ψ ∈ L2 (Rn ) : ∥ψ∥2 = 1}. We will show then that the minimizer ψ0 of ε
on the unit sphere is an eigenvector of H with eigenvalue E0 = ε(ψ0 ). 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 (2.40), is whether ε is bounded below. Consider, for example, for n = 3, the
potential V (x) = −|x|−5/2 . Then V ∈ Lsloc (R3 ), for all s < 6/5. For every ψ ∈ C0∞ (Rn ) with
∥ψ∥2 = 1 and for λ > 0 we set
ψλ (x) = λ−3/2 ψ(x/λ)

36
Then ∥ψλ ∥2 = 1 for all λ > 0 and
Z Z
ε(ψλ ) = |∇ψλ (x)| dx − |x|−5/2 |ψλ (x)|2
2

Z Z
−2 −5/2
=λ 2
|∇ψ(x)| dx − λ dx |x|−5/2 |ψ(x)|2

For λ → 0 we notice that the second term dominates and that the energy takes arbitrarily
large negative values. In this case, ε(ψ) 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
Z
T (ψ) = dx |∇ψ(x)|2

for the kinetic energy of the particle.

Theorem 2.5.1. Assume that V ∈ L∞ (Rn )+Ln/2 (Rn ), if n ≥ 3, that V ∈ L∞ (Rn )+L1+ε (Rn ),
if n = 2, for an arbitrary ε > 0, and that V ∈ L1 (Rn ) + L∞ (Rn ) if n = 1. Then there exist
constants C, D > 0 with
ε(ψ) ≥ CT (ψ) − D∥ψ∥2
In particular
E0 := inf {ε(ψ) : ∥ψ∥2 = 1} > −∞

Remark: Here V ∈ Lp1 + Lp2 means that there are V1 ∈ Lp1 and V2 ∈ Lp2 such that
V = V1 + V2 .
Proof. We consider only the case n ≥ 3 (the other cases can be handled analogously). By
assumption, we have V1 ∈ L∞ , V2 ∈ Ln/2 with V = V1 + V2 . We claim that, for arbitrary
δ > 0 there exists W1 ∈ L∞ , W2 ∈ Ln/2 with V = W1 + W2 and ∥W2 ∥n/2 ≤ δ. In fact, since
|V2 (x)|n/2 χ(|V2 (x)| ≥ µ) ≤ |V2 (x)|n/2 for all x ∈ Rn and since |V2 (x)|n/2 χ(|V2 (x)| ≥ µ) → 0 for
almost all x ∈ Rn , as µ → ∞, it follows from dominated convergence that
Z
|V2 (x)|n/2 χ(|V2 (x)| ≥ µ) → 0

as µ → ∞. Hence, there exists µ0 > 0 large enough with

Z
|V2 (x)|n/2 χ(|V2 (x)| ≥ µ) ≤ δ n/2

37
Then W2 (x) = V2 (x)χ(|V2 (x)| ≥ µ0 ) and W1 (x) = V1 (x) + V2 (x)χ(|V2 (x)| ≤ µ0 ) have the
desired properties. Thus
Z Z
ε(ψ) = |∇ψ| dx + V |ψ|2
2

Z Z
= ∥∇ψ∥2 + W1 (x)|ψ(x)| + W2 (x)|ψ(x)|2
2 2

≥ ∥∇ψ∥22 − ∥W1 ∥∞ ∥ψ∥22 − ∥W2 ∥n/2 ∥ψ∥22n/(n−2)

≥ (1 − 2δ)∥∇ψ∥22 − ∥W1 ∥∞ ∥ψ∥22

where in the last inequality we used Sobolev inequality. The theorem follows by choosing
δ < 1/2.
For example, for n = 3, the last theorem can be applied to the hydrogen atom, where
V (x) = −1/|x| = −χ(|x| ≤ 1)/|x| − χ(|x| ≥ 1)/|x| ∈ Lp (R3 ) + L∞ (R3 ), for all p < 3. Theorem
2.5.1 implies that the spectrum of the hydrogen atom is bounded below, something we already
knew from its explicit computation in Section 2.4. Notice that the stability of the hydrogen
atom and of other quantum system with attractive potentials (ie. the fact that the spectrum
is bounded below), was a crucial success for quantum mechanics. In a classical system with
potential V (x) = −1/|x| (or with any potential that is not bounded from below), the energy
can take arbitrarily large negative values. As a consequence, the system will constantly try to
decrease its energy, it would not be stable (in a classical atom, the electron would collapse onto
the nucleus). In quantum mechanics, instead, the negative potential energy is compensated
by the positive kinetic energy, so that the total energy is always bounded below. In other
words, to localize the electron close to the nucleus (i.e. to make the potential energy very
negative), we pay a price in terms of kinetic energy, compensating for the negative potential
(this is a formulation of Heisenberg’s uncertainty principle). This explains the stability of
the hydrogen atom and of other system with potential that are not bounded from below
in quantum mechanics (the conditions on V guarantee exactly that the localization price
dominates the expectation of the potential).
Weakly continuity of potential energy. Next, we look for conditions that guarantee the
existence of a minimum of the energy (boundedness from below is a necessary condition, not
sufficient). We will make use to the following theorem.

Theorem 2.5.2. Let V : Rn → R, with V ∈ L∞ (Rn ) + Ln/2 (Rn ), if n ≥ 3, V ∈ L∞ (Rn ) +

L1+ε (Rn ), if n = 2, and V ∈ L∞ (Rn ) + L1 (Rn ), if n = 1. We assume moreover that V ∈
L∞ (Rn \BR (0)) for sufficiently large R > 0, with ∥V ∥L∞ (Rn \BR (0)) → 0 as R → ∞. The
potential energy Z
P (ψ) = dxV (x)|ψ(x)|2

38
is then weakly continuous in H 1 (Rn ). In other words, if ψj → ψ weakly in H 1 (Rn ), then
P (ψj ) → P (ψ) as j → ∞ (Notice that weak continuity is stronger than stron continuity!)

Proof. We consider the case n ≥ 3, the other cases can be handled analogously. Let ψj be a
sequence in H 1,2 (Rn ) with ψj → ψ weakly in H 1,2 (Rn ). Then the sequence ψj is bounded in
H 1,2 (Rn ), i.e. ∥ψj ∥H 1,2 ≤ C für alle j ∈ N. Since
Z
V (x)|ψj (x)|2 ≤ ∥V ∥L∞ (BRc (0)) ∥ψj ∥22 ≤ C∥V ∥L∞ (BcR (0)) → 0
|x|≥R

uniformly in j, it is enough to show that

Z Z
χBR (0) (x)V (x)|ψj (x)| → χBR (0) (x)V (x)|ψ(x)|2
2

as j → ∞, for an arbitrary, but fixed, R > 0. We write now V (x) = V1 (x) + V2 (x), with
V1 ∈ Ln/2 (Rn ) and V2 ∈ L∞ (Rn ). For δ > 0 we set

V1 (x) falls |V1 (x)| ≤ 1/δ
V1,δ (x) =
0 sonst

and Vδ = V1,δ + V2 . Then |V1,δ (x)| ≤ |V1 (x)| for all δ > 0, and V1,δ (x) → V1 (x) almost
everywhere. Dominated convergence implies that
Z Z
|V (x) − Vδ (x)| dx = |V1 (x) − V1,δ (x)|n/2 dx → 0
n/2
als δ → 0

and therefore
Z Z
χBR (0) (x) (Vδ (x) − V (x)) |ψj (x)| ≤ |Vδ (x) − V (x)| |ψj (x)|2 dx
2

Z
2
≤ ∥ψj ∥2n/n−2 |Vδ (x) − V (x)|n/2 dx
Z
2
≤ ∥ψj ∥H 1,2 |Vδ (x) − V (x)|n/2 dx → 0

This means that it is enough to show that

Z Z
2
χBR (0) Vδ |ψj | → χBR (0) Vδ |ψ|2

as j → ∞, for all fixed δ, R > 0. To this end, notice that ψj → ψ weakly in H 1,2 (BR (0)).
This implies, by the Sobolev Embedding Theorem, that ψj → ψ strongly in Lq (BR (0)), for all

39
1 ≤ q < 2n/(n − 2). Hence, |ψj |2 → |ψ|2 strongly in Lq/2 (BR (0)) for all 1 ≤ q < 2n/(n − 2).
In particular |ψj |2 → |ψ|2 strongly in L1 (BR (0)). Hence,
Z
χBR (0) Vδ |ψj |2 − |ψ|2 ≤ ∥Vδ ∥∞ |ψj |2 − |ψ|2 L1 (B (0)) → 0


R

as j → ∞.

Existence of minimizers. Now, we can show the existence of a minimum.

Theorem 2.5.3. Let V : Rn → R with V ∈ L∞ (Rn ) + Ln/2 (Rn ), if n ≥ 3, V ∈ L∞ (Rn ) +
L1+ε (Rn ), if n = 2, and V ∈ L∞ (Rn ) + L1 (Rn ), if n = 1. Moreover, let V ∈ L∞ (Rn \BR (0))
for R large enough, with ∥V ∥L∞ (BRc (0)) → 0 as R → ∞. We assume that

E0 = inf ε(ψ) : ψ ∈ H 1,2 (Rn ), ∥ψ∥2 = 1 < 0 .



Then there exists ψ0 ∈ H 1,2 (Rn ), with ∥ψ0 ∥2 = 1 and ε(ψ0 ) = E0 . ψ0 satisfies (in a weak
sense, after integration against a smooth test function) the Schrödinger equation

(−∆ + V )ψ0 = E0 ψ0

Proof. Let ψj be a sequence in H 1,2 (Rn ) with ∥ψj ∥2 = 1 and ε(ψj ) → E0 as j → ∞. Since
1
ε(ψj ) ≥ ∥∇ψj ∥22 − C
2
it follows from Theorem 2.5.1, that ∥∇ψj ∥2 is bounded. Hence, ∥ψj ∥H 1,2 ≤ C for all j. This
implies that there exists a subsequence nj and ψ0 ∈ H 1,2 (Rn ) with ψnj → ψ0 weakly in
H 1,2 (Rn ) (in other words ψnj → ψ0 weakly in L2 (Rn ) and ∇ψnj → ∇ψ0 weakly in L2 (Rn )).
Since in the weak limit the norm can only get smaller, we obtain

∥ψ0 ∥2 ≤ 1, and ∥∇ψ0 ∥2 ≤ lim inf ∥∇ψnj ∥2

j→∞

From Theorem 2.5.2, we have P (ψ0 ) = limj→∞ P (ψnj ). This implies that

E0 ∥ψ0 ∥22 ≤ ε(ψ0 ) = ∥∇ψ0 ∥22 + P (ψ0 ) ≤ lim inf ∥∇ψnj ∥22 + P (ψnj ) = lim inf ε(ψnj ) = E0


j→∞ j→∞

Since E0 < 0, we find ∥ψ0 ∥2 ≥ 1. This means that ∥ψ0 ∥2 = 1 and ε(ψ0 ) = E0 . To show
that ψ0 satisfies the Schrödinger equation, we consider the variation of ψ0 . For δ ∈ R and
f ∈ C0∞ (Rn ), let ψδ = ψ0 + δf and R(δ) = ε(ψδ )/∥ψδ ∥22 . Then R(δ) has a minimum in δ = 0.
Hence, since R is differentiable in δ = 0, we must assume that
dR(δ) dε(ψδ ) d∥ψδ ∥22
0= |δ=0 = |δ=0 − E0
dδ dδ dδ
40
A simple computation shows that
Z
dε(ψδ )
dx ∇f¯ · ∇ψ0 + V f¯ψ0


|δ=0 = 2Re
dδ
and that
d∥ψδ ∥22
Z
|δ=0 = 2Re dxf¯ψ0
dδ
Thus Z
(−∆ + V − E0 ) f¯ ψ0 = 0
 
Re

for all f ∈ C0∞ (Rn ). If we replace f by if we conclude that

Z
(−∆ + V − E0 ) f¯ ψ0 = 0
 

for all f ∈ C0∞ (Rn ). This shows that ψ0 solves the Schrödinger equation (in the weak sense).

Excited eigenvalues and their eigenfunctions. Theorem 2.5.3 gives a variational char-
acterization of the smallest eigenvalue of H. It is also possible to give a variational character-
ization of higher eigenvalues and eigenfunctions. Let us assume that

E0 = inf ε(ψ) : ψ ∈ H 1,2 (Rn ), ∥ψ∥2 = 1 < 0



Then Theorem 2.5.3 implies that the energy functional ε(φ) has a minimizer ψ0 on the unit
sphere of L2 (Rn ) which is an eigenvector of H with eigenvalue E0 . We can then define

E1 = inf ε(ψ) : ψ ∈ H 1,2 (Rn ), ∥ψ∥2 = 1 and ⟨ψ, ψ0 ⟩ = 0



i.e. 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 vector by
ψ1 . We can proceed recursively. Given that we already constructed the normalized vector
ψ0 , ψ1 , . . . , ψk−1 , we define

Ek = inf ε(ψ) : ψ ∈ H 1,2 (Rn ), ∥ψ∥2 = 1 and ⟨ψ, ψj ⟩ = 0, for all j = 0, 1, . . . , k − 1 (2.41)


In the next theorem, we show that, if Ek < 0, then Ek is an eigenvalue of H and the minimizer
ψk is the corresponding eigenvector.

Theorem 2.5.4. Let V be as in Theorem 2.5.3. Assume Ek < 0. Then the infimum in (2.41)
is attained and the minimizer ψk is such that Hψk = Ek ψk .

41
Proof. The proof of the existence of a minimizer follows the same ideas as the proof of Theorem
2.5.3. From a minimizing sequence ψkj we extract a weak limit ψk . As in the proof of Theorem
2.5.3 one can show that ε(ψk ) = Ek and that ∥ψk ∥ = 1. The only additional observation here
is the fact that ⟨ψk , ψℓ ⟩ = 0, for all ℓ = 0, . . . , k − 1. This follows from ψkj → ψk weakly, since
⟨ψkj , ψℓ ⟩ = 0 for all ℓ = 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 2.5.3, that ⟨f, (H −Ek )ψk ⟩ = 0 for all f ∈ C0∞ (Rn ) with ⟨f, ψℓ ⟩ = 0 for
Pk−1
all ℓ = 0, 1, . . . , k − 1. This implies that (H − Ek )ψk = ℓ=1 αℓ ψℓ for appropriate coefficients
αℓ ∈ C. Multiplying the equation with ψi and using the orthogonality ⟨ψi , ψk ⟩ = 0, we
conclude that αi = 0 for all i = 0, . . . , k − 1 and therefore that Hψk = Ek ψk .
Remarks. To conclude this section, let us state two more important properties of eigen-
vectors of Schrödinger operators that hold true in great generality.

1) Under the same assumptions as in Theorem 2.5.3, the normalized minimizer ψ0 of the
energy functional ε (which by Theorem 2.5.3 is an eigenvector of H with eigenvalue
E0 = min{ε(ψ) : ∥ψ∥ = 1}) 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, ie. the
smallest eigenvalue of H, is simple, non-degenerate.

2) Elliptic regularity: Let B1 ⊂ Rn be an open ball and let ψ and V be functions on B with
(−∆ + V )ψ = 0, in the sense of distributions. Then, for any ball B ⊂ Rn concentric
with B1 and with strictly smaller radius, we have

i) If n = 1, ψ is continuously differentiable on B.
ii) If n = 2, ψ ∈ Lq (B) for all q < ∞.
iii) If n = 3, ψ ∈ Lq (B) for all q < n/(n − 2).
iv) If n ≥ 2 and V ∈ Lp (B1 ) for a n/2 < p ≤ n, then ψ is Hölder continuous with
exponent α in B, for all α < 2 − n/p.
v) If n ≥ 1 and V ∈ Lp (B1 ) for a p > n, ψ is continuously differentiable and the
derivative is Hölder continuous with exponent α in B, for all α < 1 − n/p.
vi) If n ≥ 1 and V ∈ C k,α (B1 ) (this is the subspace of C k (B1 ) of functions whose k-th
derivative is Hölder continuous with exponent α) for some k ≥ 0 and 0 < α < 1,
then ψ ∈ C k+2,α (B).

In other words, there is a gain of regularity (of two derivatives) between the potential and
eigenvectors of Schrödinger operators (which by definition are only in L2 (Rn )). Note that
this regularity result holds locally. For example, this result implies that the eigenvectors
of the hydrogen atom, with V (x) = −1/|x|, are C ∞ in any ball away from the origin.

42
 In a ball containing the origin, on the other hand, V ∈ Lp for all p < 3; hence the
result above implies that eigenvectors of the hydrogen atom are Hölder continuous with
exponent α, for any α < 1, which is consistent with the explicit formula proportional
to e−|x| that we found above (in fact, the function e−|x| behaves a little bit better, it is
Hölder continuous with exponent α = 1, i.e. it is Lipschitz continuous).

2.6 Many-Body Quantum Mechanics

Bosonic and fermionic statistics. A system of N particles moving in the 3-dimensional
space is described on the Hilbert space

HN = L2 (RdN , dx1 . . . dxN ) = L2 (R3 )⊗N

given by the tensor product of N copies of the one-particle space L2 (R3 ). Wave functions
ψN ∈ HN are normalized so that
Z
∥ψN ∥ = |ψN (x1 , . . . , xN )|2 dx1 . . . dxN = 1
2

Accordingly, |ψN (x1 , . . . , xN )|2 is the probability density for finding particle 1 close to x1 ,
particle 2 close to x2 , and so on.
When considering systems of N indistinguishable particles, we need to restrict the Hilbert
space to subspaces with appropriate statistics. Indistinguishability plays an important role in
quantum mechanics, much more important than in classical mechanics. Classical particles fol-
low trajectories and can be labeled. Quantum particles do not. Indistinguishability translates,
in quantum system, into a restriction of the set of physical observables. Only observables that
are invariant with respect to permutations of the N particles can be measured.
Let us denote by SN the group of permutations π of the set {1, 2, . . . , N }. We define a
unitary representation of SN on the N -particle Hilbert space L2 (R3N ) setting

(Uπ ψ)(x1 , . . . , xN ) = ψ(xπ1 , . . . , xπN )

Physically meaningful observables of the N -particle system are self-adjoint operators A on

L2 (R3N ) with the property Uπ∗ AUπ = A (or, equivalently, [A, Uπ ] = 0) for all π ∈ SN (this
is the meaning of invariance w.r.t. permutations). An example of such an observable is the
kinetic energy
XN
−∆xj
j=1

of the N -particles (the operator −∆x1 , measuring the kinetic energy of particle 1, is, on the
other hand, not invariant w.r.t. permutations).

43
 Let now ψ ∈ L2 (R3N ) be a normalized wave function, describing the state of the N -particle
system. Then, we have
∥Uπ ψ∥ = ∥ψ∥
and
⟨Uπ ψ, AUπ ψ⟩ = ⟨ψ, Uπ∗ AUπ ψ⟩ = ⟨ψ, Aψ⟩
for every permutation π ∈ SN and for every allowed (ie. invariant w.r.t. permutations)
observable A. This means that the wave function Uπ ψ describes exactly the same state as
the wave function ψ, for every π ∈ SN . In other words, there is no one-to-one correspondence
between physical states and wave functions in L2 (R3N ) (beyond the standard non-uniqueness
due to multiplication with phases). To avoid this redundancy, we postulate that wave functions
of many-body quantum system must be chosen so that Uπ ψ is proportional to ψ, for all
π ∈ SN . It turns out that there are only two possible behaviors (because there are exactly
2 one-dimensional irreducible representation of the group SN ). The wave function ψ can be
completely symmetric w.r.t. permutation, ie Uπ ψ = ψ, meaning that

ψN (xπ1 , xπ2 , . . . , xπN ) = ψN (x1 , . . . , xN )

for all π ∈ SN . Alternatively, the wave function ψ can be completely antisymmetric w.r.t.
permutation, meaning that Uπ ψ = σπ ψ, meaning that

ψN (xπ1 , xπ2 , . . . , xπN ) = σπ ψN (x1 , . . . , xN )

for all π ∈ SN . Here, σπ ∈ {−1, 1} is the sign of the permutation π (ie. σπ = 1, if π consists
of an even number of transpositions, σπ = −1 if π consists of an odd number of transposi-
tions). For each quantum particle, either completely symmetric or completely antisymmetric
wave functions can be considered, not both at the same time. Particles described by wave
functions that are symmetric w.r.t. permutations are called bosons. Particles described by
antisymmetric wave function are called fermions. Systems of N bosons moving in the three
dimensional space are described on the Hilbert space

L2s (R3N ) = L2 (R3 )⊗s N = {ψ ∈ L2 (R3N ) : ψ(xπ1 , . . . , xπN ) = ψ(x1 , . . . , xN )}

Systems of N fermions on R3 are described, on the other hand, on the Hilbert space

L2a (R3N ) = L2 (R3 )∧N = {ψ ∈ L2 (R3N ) : ψ(xπ1 , . . . , xπN ) = σπ ψ(x1 , . . . , xN )}

The choice between bosonic and fermionic statistics is related with the internal degree of
freedom of quantum particles, known as spin (already discussed in the Stern-Gerlach experi-
ment). Particles with integer spin are bosons (they are described by wave functions that are
symmetric w.r.t. permutations). Particles with half-integer spin are fermions (they must be
described by wave functions that are antisymmetric w.r.t. permutations).

44
 Examples of bosonic wave function: The simplest example of a bosonic wave function is
the product of N copies of the same one-particle normalized wave function φ ∈ L2 (R3 ), ie
N
Y
ψN (x1 , . . . , xN ) = φ(xj )
j=1

which is clearly normalized and symmetric w.r.t. permutations.

More generally, if {φj }j∈N is an orthonormal basis in L2 (R3 ), we can describe the N -
particle state with one particle in the state φi1 , one particle in the state φi2 , and so on, the
symmetrized wave function
r
1 X
ψB (x1 , . . . , xN ) = φi1 (xπ1 )φi2 (xπ2 ) . . . φiN (xπN )
ℓ1 !ℓ2 ! . . . ℓk !N ! π∈S
N

if the set {i1 , . . . , iN } contains k different indices, appearing respectively ℓ1 , . . . ℓk times, with
ℓ1 + ℓ2 + · · · + ℓk = N .
Similarly, we can define construct fermionic wave functions. In this case, however, each
one-particle state can be “occupied” by at most one-particle, otherwise the antisymmetrization
is going to vanish (this is known as the Pauli principle). The fermionic state with one particle
in the state φi1 , one particle in the state φi2 , and so on (with all different indices i1 , . . . , iN ) ,
can be described by the wave function
1 X
ψF (x1 , . . . , xN ) = √ σπ φi1 (xπ1 ) . . . φiN (xπN )
N ! π∈SN

The wave function ψF can be written as

1
ψF (x1 , . . . , xN ) = √ det (φij (xk ))1≤j,k≤N
N!
Such wave functions are known as Slater determinants.
Reduced density matrices. The density matrix of a normalized N -particle wave function
ψN ∈ L2 (R3N ), with either bosonic or fermionic statistics, is defined as γN = |ψN ⟩⟨ψN |, ie. γN
is the orthogonal projection onto ψN . More generally, mixed states of an N -particle system
are described by non-negative trace class operators γN with trγN = 1.
For a density matrix γN , we denote by γN (x1 , . . . , xN ; y1 , . . . , yN ) the integral kernel of γN .
This is defined so that
Z
(γN ξN )(x1 , . . . , xN ) = γN (x1 , . . . , xN ; y1 , . . . , yN )ξN (y1 , . . . , yN )

45
Every trace class operator can be associated with an integral kernel (since every trace class
operator is Hilbert Schmidt, the integral kernel is in fact a L2 (R3N × R3N )-function).
Using the integral kernel of γN , we can define its reduced density matrices. For 1 ≤ k ≤ N ,
we define the k-particle reduced density matrix associated with γN (and therefore with ψN )
by
N!
γ (k) = trk+1,k+2,...,N γN
(N − k)!
where trk+1,...,N γN denotes the partial trace of γN over the last (N − k) particles. Taking the
partial trace means integrating out the degrees of freedom associated with the last (N − k)
(k)
particles. In other words, γN is defined through its integral kernel
(k)
γN (x1 , . . . , xk ; y1 , . . . , yk )
Z
N!
= γN (x1 , . . . xk , xk+1 , . . . , xN ; y1 , . . . yk , xk+1 , . . . , xN )dxk+1 . . . dxN
(N − k)!
(k)
By definition, it is easy to check that γN is a non-negative operator, for all 1 ≤ k ≤ N .
In fact, for an arbitrary φ ∈ L2 (R3k ) we have
Z
(k) (k)
⟨φ, γ φ⟩ = dx1 . . . dxk dy1 . . . dyk φ̄(x1 , . . . , xk )γN (x1 , . . . , xk ; y1 , . . . , yk )φ(y1 , . . . , yk )
Z Z 2
= dxk+1 . . . dxN dx1 . . . dxk φ̄(x1 , . . . , xk )ψN (x1 , . . . , xk , xk+1 , . . . , xN ) ≥ 0

Moreover,
Z
(k) N! (k) N!
trγN = dx1 . . . dxk γN (x1 , . . . , xk ; x1 , . . . , xk ) =
(N − k)! (N − k)!
Reduced density matrices are useful, because they allow us to compute the expectation of
observables that depend only on a fixed number k of particles. Let O(k) be an operator on
the k-particle space L2 (R3k ). Then O(k) ⊗ 1(N −k) is an operator on L2 (R3N ), acting as O(k) on
the first k particles and as the identity on the other (N − k) particles. A simple computation
shows that
(N − k)! (k)
⟨ψN , (O(k) ⊗ 1(N −k) )ψN ⟩ = tr O(k) γN
N!
2 3k
where, on the r.h.s., tr denotes the trace over L (R ). Interesting k-particle observables in
N -body systems must be invariant w.r.t. permutations of the N particles and are therefore
given by sums over all possible choices of the k indices, over which O(k) must act non-trivially.
Using the permutation invariance of ψN , their expectation is given by
X (k)
X (N − k)! (k) 1 (k)
⟨ψN , Oi1 ,...,ik ψN ⟩ = tr O(k) γN = tr O(k) γN
1≤i <i <···<i ≤N 1≤i <i <···<i ≤N
N ! k!
1 2 k 1 2 k

46
For example, the kinetic energy of an N -body system is the one-particle operator N
P
j=1 −∆xj .
Its expectation in the state ψN can be computed through the one-particle reduced density
(1)
γN :
XN
(1)
⟨ψN , −∆xj ψN ⟩ = tr(−∆)γN
j=1
PN
A two-body interaction has the form i<j V (xi − xj ). To compute its expectation we need
its 2-particle reduced density:
N
X 1 (2)
⟨ψN , V (xi − xj )ψN ⟩ = trV (x1 − x2 )γN
i<j
2

Let us compute a couple of reduced density matrices, for simple example of N -particle
states. In the bosonic setting, the simplest QN class of N -particle states are factorized wave
functions of the form ψN (x1 , . . . , xN ) = j=1 φ(xj ). In this case, it is very easy to compute
reduced density matrices. We find

(k) N!
γN = |φ⟩⟨φ|⊗k
(N − k)!
where |φ⟩⟨φ| is the orthogonal projection onto φ.
In the fermionic setting, the simplest class of N -particle states are Slater determinant,
having the form
1
ψN (x1 , . . . , xN ) = √ det (φi (xj ))1≤i,j≤N
N!
where {φj }N 2 3
j=1 is an orthonormal system on L (R ). The one-particle reduced density matrix
associated with ψN has the integral kernel
Z N
(1) N X Y
γN (x; x′ ) = ′
σπ σπ′ φπ1 (x)φπ′ 1 (x ) dx2 . . . dxN φπj (xj )φπ′ j (xj )
N ! π,π′ ∈S j=2
N

N
1 X X
= φπ1 (x)φπ1 (x′ ) = φj (x)φj (x′ )
(N − 1)! π j=1

(1) PN
Hence γN = j=1 |φj ⟩⟨φj | is the orthogonal projection onto the N -dimensional subspace
2 3
of L (R ), spanned by φ1 , . . . , φN . Similarly, one can compute higher order reduced density
matrices of the Slater determinant ψN . It turns out, they satisfy Wick’s theorem, stating that

X k
Y
(k) (1)
γN (x1 , . . . xk ; x′1 , . . . , x′k ) = σπ γN (xj ; x′πj )
π∈Sk j=1

47
In particular, the 2-particle reduced density is given by
(2) (1) (1) (1) (1)
γN (x1 , x2 ; x′1 , x′2 ) = γN (x1 ; x′1 )γN (x2 ; x′2 ) − γN (x1 ; x′2 )γN (x2 ; x′1 )

Typical N -particle Hamilton operators have the form

N
X   X
HN = − ∆xj + Vext (xj ) + V (xi − xj )
j=1 i<j

The first term is a one-particle operator, it describes the kinetic energy of the particles and
their interaction with external fields. The second term is a two-body operator, describing the
interaction among the particles. The energy of the N -particle system in a state ψN can be
(2)
expressed in terms of the reduced density matrix γN associated with ψN . We find
  (1) 1 (2)
⟨ψN , HN ψN ⟩ = tr − ∆ + Vext γN + tr V (x1 − x2 )γN (2.42)
2
This identity suggests that a possible approach to compute the ground state energy of an
N -particle system, ie. the infimum of ⟨ψN , HN ψN ⟩ over all normalized wave functions ψN ∈
L2 (R3N ) with the appropriate statistics, consists in minimizing the r.h.s. of (2.42) over all
(2) (1) (2)
possible choices of γN (the one-particle density matrix γN can be computed from γN , taking
the partial trace over one of the two particles). At first sight, this seems much easier than
(2)
minimizing the l.h.s. of (2.42) over all N -particle wave function ψN , because γN is an operator
on the 2-particle space L2 (R6 ), which is of course much smaller than L2 (R3N ). Unfortunately,
this approach does not work, in general, because it is very difficult to characterize the space of
all two-particle reduced densities (ie. it is very difficult to decide whether a non-negative trace
class operator γ (2) over L2 (R6 ) is the two-particle reduced density of an N -particle states or
not). It is possible, however, to characterize admissible one-particle reduced densities (hence,
the suggested approach works well for Hamilton operators without two-body interactions).
This is the content of the next two theorems, where we consider separately systems with
bosonic and fermionic statistics.

Theorem 2.6.1. Let γ (1) be a self-adjoint, non-negative operator on L2 (R3 ), with trγ (1) =
N , for some N ∈ N. Then there exists a bosonic density matrix γ on L2 (R3N ) such that
γ (1) = N tr2,...,N γ. If N ≥ 2, γ can be chosen to be a pure state, ie. there exists ψ ∈ L2 (R3N ),
symmetric w.r.t. permutations, such that γ (1) = N tr2,...,N |ψ⟩⟨ψ|.

Remark: a bosonic density matrix γ on L2 (R3N ) is a non-negative trace-class operator on

L2 (R3N ) with γ(xπ1 , . . . xπN ; x′π′ 1 , . . . , x′π′ N ) = γ(x1 , . . . , xN ; x′1 , . . . , x′N ) for all permutations
π, π ′ ∈ SN ).

48
Proof. If N = 1, we can just take γ = γ (1) . Let us consider the case N ≥ 2. Suppose that
X
γ (1) = λj |φj ⟩⟨φj |
j

with orthonormal eigenfunctions φj ∈ L2 (R3 ), 0 < λj ≤ 1 for all j and

P
j λj = N . Consider

N
1 X 1/2 Y
ψ(x1 , . . . , xN ) = √ λj φj (xi )
N j i=1

Then ψ(xπ1 , . . . , xπN ) = ψ(x1 , . . . , xN ) and

1 X
∥ψ∥22 = λj = 1
N j

Moreover,
Z X
N dx2 . . .dxN ψ̄(x, x2 , . . . , xN )ψ(x′ , x2 , . . . , xN ) = λj φ̄j (x)φj (x′ ) = γ (1) (x; x′ )
j

While in the bosonic case every non-negative trace class operator γ (1) over L2 (R3 ) with
trγ (1) = N is admissible, in the fermionic case admissibility requires the additional condition
γ (1) ≤ 1, expressing the Pauli principle.

Theorem 2.6.2. Let γ (1) be a self-adjoint, non-negative, trace-class operator on L2 (R3 ), with
trγ (1) = N , for some N ∈ N. Then there exists a fermionic density matrix γ over L2 (R3N )
with γ (1) = N trγ if and only if γ (1) ≤ 1.

Remarks: 1) a density matrix γ over L2 (R3N ) is called fermionic if

γ(xπ1 , . . . xπN ; x′π′ 1 , . . . , x′π′ N ) = σπ σπ′ γ(x1 , . . . , xN ; x′1 , . . . , x′N )

for all permutations π, π ′ ∈ SN ; 2) in general γ cannot be chosen as a pure state.

Proof. First, we assume 0 ≤ γ (1) ≤ 1 and we show the existence of γ on L2 (R3N ) with the
desired properties. Suppose that
X
γ (1) = λj |φj ⟩⟨φj |
j

49
where φj is an orthonormal system in L2 (R3 ). We assume here that the eigenvalues λj are
ordered, ie.
1 ≥ λ1 ≥ λ2 ≥ . . .
P
If λN +1 = 0, the condition j λj = N implies that λj = 1 for all j = 1, . . . , N , and therefore
that γ (1) = N
P
j=1 |φj ⟩⟨φj |. In this case, the statement follows taking

1 X
ψ(x1 , . . . , xN ) = √ σπ φπ1 (x1 ) . . . φπN (xN )
N ! π∈SN

as a Slater determinant in L2 (R3N ) and setting γ = |ψ⟩⟨ψ|.

Hence, we can assume that λN +1 > 0. Because of the ordering of the eigenvalues, we have
N
X +1
N≥ λj ≥ (N + 1)λN +1 (2.43)
j=1

which implies that λN +1 ≤ N/(N + 1) < 1. Hence, defining

ε = min{λN , 1 − λN +1 }
we certainly have 0 < ε < 1. Using the notation

0 if j > N
χ[1,N ] (j) =
1 if 1 ≤ j ≤ N
we write
λj = εχ[1,N ] (j) + (1 − ε)fj
with
1  
fj = λj − εχ[1,N ] (j)
1−ε
We observe that, for 1 ≤ j ≤ N , we have
λj − ε
fj =
1−ε
and therefore 0 ≤ fj ≤ 1 (because λj ≥ λN ≥ ε and λj ≤ 1, for all 1 ≤ j ≤ N ). On the other
hand, for j > N , we find
λj
fj =
1−ε
and therefore again 0 ≤ fj ≤ 1, because λj ≤ λN +1 ≤ 1 − ε (from the choice ε < 1 − λN +1 ).
Hence, we find 0 ≤ fj ≤ 1 for all j and
X 1
fj = (N − εN ) = 1
j
1 − ε

50
In other words, the coefficients fj satisfy the same conditions as the original coefficients λj .
After reshuffling the terms so that the coefficients fj are decreasing, we can repeat the
procedure. After M iteration, we arrive at
M
X
λj = ck χk (j) + RjM (2.44)
k=1

where
k−1
Y
ck = ε k (1 − εj ),
j=1

0 < εj < 1 is the ε-coefficient used in the j-th iteration, χk is the characteristic function of a
subset of N containing exactly N points and
(M )
RjM = (1 − ε1 )(1 − ε2 ) . . . (1 − εM )fj
(M ) P (M )
with 0 ≤ fj ≤ 1 for all j and j fj = N . Remark that

M
Y M
X
(1 − εk ) = 1 − ck (2.45)
k=1 k=1

and therefore
M
X
(M )
RjM = fj (1 − ck ) (2.46)
k=1
PM
The identity (2.45) implies, in particular, that k=1 ck ≤ 1, uniformly in M .
Now, we claim that, as M → ∞,
X M
ck → 1. (2.47)
k=1
(M )
By (2.46) (recalling that 0 ≤ fj ≤ 1), this would also imply that RjM → 0 and therefore
∞
X
λj = ck χk (j)
k=1

which means that

X ∞
X hX i
(1)
γ = λj |φj ⟩⟨φj | = ck χk (j)|φj ⟩⟨φj |
j k=1 j

51
Hence, γ (1) would be a convex (since 0 ≤ ck ≤ 1 for all k ∈ N and ∞
P
k=1 ck = 1) combination of
projections onto N -dimensional subspaces of L2 (R3 ). Introducing the notation ψk ∈ L2a (R3N )
for the Slater determinant defined by the one-particle wave functions φj , for all j with χk (j) =
1 (for every k ∈ N, there are exactly N such labels j), we could conclude that
X
N tr2,...,N |ψk ⟩⟨ψk | = χk (j)|φj ⟩⟨φj |
j

Defining the density matrix

∞
X
γ= ck |ψk ⟩⟨ψk |
k=1

on L2a (R3N ), we would obtain therefore that N tr2,...,N γ = γ (1) , as desired.

PM
It remains to show the claim (2.47). Since
PM k=1 ck ≤ 1, uniformly in M , the limit
limM →∞ k=1 ck = d exists and satisfies d ≤ 1. We proceed by contradiction and as-
sume d < 1. From (2.44), also limM →∞ RjM = R Pj exists, for all j. P By (2.46), it satisfies
0 ≤ Rj ≤ (1 − d) and, by monotone convergence, j Rj = limM →∞ j RjM = N (1 − d), ie.
Rj X Rj
0≤ ≤ 1 for all j and =N (2.48)
1−d j
1 − d
(M ) (M )
= RjM /(1 − M
P
Let us write fj k=1 ck ). Then fj → Rj /(1 − d), as M → ∞. We
(M )
rearrange the coefficients fj so that the are monotonically decreasing. We call the resulting
(M ) (M )
coefficients fej . Arguing as in (2.43), we find 1 − feN +1 ≥ 1/(N + 1) > 0, uniformly in M .
(M )
On the other hand, feN converges, as M → ∞, to the N -th largest value in the sequence
Rj /(1 − d). From (2.48), we know that this limit is strictly positive (because otherwise the
(M ) (M )
sum of these coefficients could not reach N ). Since εM +1 = min(feN , 1 − feN +1 ), it follows
PM
that εM does not converge to 0. From (2.45), we conclude that k=1 ck → 1, in contradiction
to the assumption d < 1 (the limit of M
Q
k=1 (1 − εk ) is 0, unless the sequence εk is summable).
Next, suppose that γ (1) = N tr2,...,N γ for a fermionic density matrix γ on L2 (R3N ). We
claim that γ (1) ≤ 1. It is enough to show the statement for pure states, ie if γ = |ψ⟩⟨ψ|
for a normalized ψ ∈ L2a (R3N ). The general case follows by linearity. To this end, it is
convenient to introduce creation and annihilation operators. For ϕ ∈ L2 (R3 ), we define
aN,ϕ : L2a (R3N ) → L2a (R3(N −1) ) setting
√ Z
(aN,ϕ ψ)(x1 , . . . , xN −1 ) = N dxϕ̄(x)ψ(x, x1 , . . . , xN −1 )

We also define the creation operator a∗N,ϕ : L2a (R3(N −1) → L2a (R3N ), setting
N
1 X
(a∗N,ϕ ψ)(x1 , . . . , xN ) =√ (−1)N −j ϕ(xj )ψ(x1 , . . . , xj−1 , xj+1 , . . . , xN )
N j=1

52
It is easy to check (as suggested by the notation, that a∗N,ϕ is the adjoint of aN,ϕ , ie. that

⟨ψ1 , aN,ϕ ψ2 ⟩ = ⟨a∗N,ϕ ψ1 , ψ2 ⟩

Another important property of creation and annihilation operators is the identity:

aN +1,ϕ a∗N +1,ϕ + a∗N,ϕ aN,ϕ = ∥ϕ∥2

Therefore, we obtain

⟨ϕ, γ (1) ϕ⟩ = ⟨aN,ϕ ψ, aN,ϕ ψ⟩ = ⟨ψ, a∗N,ϕ aN,ϕ ψ⟩

= −⟨ψ, aN +1,ϕ a∗N +1,ϕ ψ⟩ + ∥φ∥2 = −∥a∗N +1,ϕ ψ∥2 + ∥φ∥2 ∥ψ∥2 ≤ ∥φ∥2

which proves that γ (1) ≤ 1.

3 Stability of Matter
3.1 Atoms and Molecules: Hamilton Operators and their Proper-
ties
The goal of this section is to introduce many-body Hamilton operator describing matter and
to study their basic properties.
The simplest class of many-body Hamilton operators consists of sums of one-body opera-
tors, having the form
X N
free
HN = 1 ⊗ ··· ⊗ h ⊗ ··· ⊗ 1 (3.49)
j=1

for a self-adjoint h on the one-particle L2 (R3 ). For example, the one-particle Hamilton op-
erator may have the form h = −∆ + Vext . The time-evolution generated by (3.49) factorize,
ie.
free
e−iHN t = e−iht ⊗ e−iht ⊗ · · · ⊗ e−iht
In other words, the evolution of factorized initial data remains factorized and is determined by
the one-particle Hamilton operator h. Similarly, eigenvectors of the non-interacting Hamilton
operator HNfree are given by products of the eigenvectors of h. More precisely, for bosonic
systems, HNfree acts on permutation symmetric functions, and its eigenvectors are symmetrized
products of eigenvectors of h. For fermionic systems, on the other hand, eigenvectors of HNfree
are given by anti-symmetrized products of eigenvectors of h. Eigenvalues of HNfree are therefore
given by sums of eigenvalues of h. For fermionic systems, because of Pauli’s principle, only
sums of eigenvalues that are associated with different eigenvectors are allowed. In particular,
if e1 < e2 ≤ e3 ≤ . . . are the eigenvalues of h (assuming the lowest eigenvalue to be simple),

53
associated with the eigenvectors φ1 , φ2 , φ3 , . . . , the ground state energy of HNfree is given by
N e1 in the bosonic case (corresponding to the eigenvector φ⊗N 1 ) and by e1 +e2 +· · ·+eN in the
fermionic case (corresponding with the Slater determinant det(φi (xj ))1≤i,j≤N ). In particular,
this implies that fermionic systems typically have a larger energy than bosonic systems, in the
limit of large N .
As an example, let us consider the Hamilton operator of an atom with N electrons and a
nucleus with charge Z in three dimensions (Z = N for electrically neutral atoms), assuming
the nucleus to be fixed (at the origin) and the electron to not interact. The Hamilton operator
in this case is given by
N  
free
X Z
HN = −∆xj −
j=1
|xj |

The one-particle Hamilton operator h = −∆ − Z/|x| has the eigenvalues en = −Z 2 /n2 , with
degeneracy n2 (the dependence on Z follows by simple scaling). A basis of the eigenspace
associated with the eigenvalue en consists of the eigenfunctions ψn,ℓ,m , with ℓ = 0, 1, . . . , n − 1
and m = −ℓ, −ℓ + 1, . . . , ℓ − 1, ℓ. Taking into account that electrons have spin 1/2, each
eigenstate of h can be occupied by two electrons. The ground state of HNfree can therefore
by constructed by filling the eigenstates of h, starting with those with smaller energy and
by occupying each eigenstate with at most two electrons. Eigenstates with ℓ = 0 are called
s-states, eigenstates with ℓ = 1 are called p-states, eigenstates with ℓ = 2 are called d-states
and eigenstates with ℓ = 3 are called f -states. The silicium atom Si14 has the electronic
configuration 1s2 2s2 2p6 3s2 3p2 (two electrons in the state with n = 1, ℓ = 0, two electrons in
the state with n = 2, ℓ = 0, six electrons in states with n = 2, ℓ = 1, two electrons in the
state with n = 3, ℓ = 0, two electrons in states with n = 3, ℓ = 1). This simplified approach,
which completely neglects the interaction among the electrons, predicts the correct electronic
configuration for atoms with only few electrons. To understand heavier atoms (and to get
more precise information about light atoms), however, it is crucial to take into account the
interaction among electrons.
We are going to consider system of N electrons and M static nuclei, fixed at positions
R1 , . . . , RM and with charges Z1 , . . . , ZM (M = 1 corresponds to an atom, M > 1 to a
molecule). For N, M ∈ N, R = (R1 , . . . , RM ) ∈ R3M and Z = (Z1 , . . . , ZM ) ∈ RM + (physically,
nuclei have integer charge, but the systems we are going to define make sense, mathematically,
for arbitrary positive charges Zj ), we define the Hamilton operator
N
" M
# N M
X X Zi X 1 X Zi Zj
HN,M (R, Z) = −∆xj − + + (3.50)
j=1 i=1
|Ri − x j | i<j
|x i − x j | i<j
|Ri − R j |

54
 PN
We write HN,M (R, Z) = − j=1 ∆xj + V (R, Z, x), with the total potential energy operator

N M N M
X 1 X Zi Zj XX Zi
V (R, Z, x) = + −
i<j
|xi − xj | i<j |Ri − Rj | j=1 i=1 |Ri − xj |

Observing that
N X
M
X Zi
V (R, Z, x) ≥ − (3.51)
j=1 i=1
|Ri − xj |

and therefore that

N X
M 
−∆xj

X Zi
HN,M (R, Z) ≥ −
j=1 i=1
M |xj − Ri |

we conclude, using the stability of the hydrogen atom, that

M
X
inf ⟨ψN , HN,M (R, Z)ψ⟩ ≥ −N M Zi2 > −∞ (3.52)
ψN ∈L2 (R3 )⊗N :∥ψN ∥=1
i=1

for every choice of (R, Z). Eq. (3.52) establishes stability of the first kind, which is essentially
equivalent to the stability of the one-particle system (of the hydrogen atom, in this case).
From the point of view of physics, however, stability of the first kind is not quite satis-
factory. Suppose in fact that the ground state energy had the form −C(N + M )2 , growing
(in absolute value) quadratically in the total number of particles N + M . This would imply
that the ground state energy of a system with 2N electrons and 2M nuclei would be given
by −4C(N + M )2 . This would imply that we could generate the energy 2C(N + M )2 just
by merging two initially separated systems, each having N electrons and M nuclei. Matter
would collapse and there would be no stability, in the sense we are used to. This observation
leads to a new definition of stability, known as stability of the second kind, stating that the
ground state energy of a system with N electrons and M nuclei grows (in absolute value) at
most linearly in the total number of particles (N + M ).
Compared with stability of the first kind, stability of the second kind is more subtle. It
was only proved for the first time in 1967 by Dyson and Lenard, about 40 years after the birth
of quantum mechanics. In fact, stability of the second kind relies on Pauli’s principle, ie on
the fact that electrons are fermions and must therefore be described by antisymmetric wave
functions (stability of the first kind, on the other hand, holds independently of the statistics;
in (3.52) we can take the infimum over arbitrary normalized ψ ∈ L2 (R3N )).
The next theorem establishes stability of the second kind for systems of electrons and
nuclei described by the Hamilton operator (3.50). The goal of the rest of this section consists
in proving this theorem.

55
Theorem 3.1.1 (Stability of the second kind). Let

EN,M (R, Z) = inf ⟨ψN , HN,M (R, Z)ψN ⟩ . (3.53)

ψ∈L2a (R3N ),∥ψ∥2 =1

denote the ground state energy of the Hamilton operator (3.50). Then we have

EN,M (Z) = inf{EN,M (R, Z) : R = (R1 , . . . , RM ) ∈ R3M } ≥ −C(Z)(N + M ) (3.54)

where Z = max(Z1 , . . . , ZM ).

Remark: in the definition (3.50) of the Hamilton operator HN,M (R, Z) we included the
term M
P
i<j Zi Zj /|Ri − Rj |, describing the interaction among nuclei. Since nuclei are kept fixed
in this model, this term is a constant. For stability of the first kind, it does not play any
role (in fact, we estimated it by zero, in (3.51). For stability of the second kind, on the other
hand, it plays a crucial role. Removing it, the infimum of the energy would be attained taking
R1 = · · · = RM = 0, which leads to the operator
N
" # N
( M
P
X
i=1 i Z ) X 1
HN,M (R = (0, . . . , 0), Z) = −∆xj − + (3.55)
j=1
|x j | i<j
|x i − x j |

Ignoring
P the 2interaction 2among the electrons, the ground state of (3.55) is of the order
−N ( M i=1 Zi ) ≃ −N M . Since the effect of the interaction among the electrons can be
bounded unformly in M , this would contradict stability of the second kind.
Remark: it is also possible to consider systems of N electrons and M “dynamical” nuclei,
of charge Z1 , . . . , ZM . In this case, the Hamilton operator would act on the larger Hilbert
space L2a (R3N ) ⊗ L2 (R3M ) (assuming the nuclei are disinguishable particles, otherwise one has
to restrict the space for the nuclei to the P right subspace of L2 (R3M )) and, compared with
(3.50) would contain the additional term M i=1 −∆Rj /2mj , describing the kinetic energy of
the nuclei (here, mj is the mass of the j-th nucleus). The results that we will prove for the
model with static nuclei can be extended, sometimes with some additional work, to the case
of dynamical nuclei. In the following we focus on the model with static nuclei to simplify a
bit the analysis.
An important property of the ground state energy (3.53) is its monotonicity in the nuclear
charges, to be understood in the sense of following proposition.

Proposition 3.1.2. Assume Zj ≤ Z for all j = 1, . . . , M (in other words, Z is the maximal
charge of the M nuclei). Then

EN,M (R, Z) ≥ min

′
min EN,M ′ ((Ri1 , . . . , RiM ′ ), (Z, . . . , Z))
M ≤M i1 <i2 <···<iM ′

56
 Remark: In other words, a lower bound for EN,M (R, Z) can be obtained by replacing
each nuclear charge Zk either by zero (which is equivalent to removing the nucleus) or by
the common upper bound Z (the maximal charge), and by taking the minimum over all such
nuclear configurations.
Proof. For any fixed ψ ∈ L2a (R3N ), the expectation ⟨ψ, HN,M (R, Z)ψ⟩ is an affine function
of Zi , for every i = 1, . . . , M (keeping here all other charges Zj , j ̸= i, fixed). This im-
plies that the infimum of ⟨ψ, HN,M (R, Z)ψ⟩ is a concave function of Zi . In fact, writing
⟨ψ, HN,M (R, Z)ψ⟩ = a(ψ) + Zi b(ψ) and defining

f (Zi ) = inf ⟨ψ, HN,M (R, Z)ψ⟩ = inf a(ψ) + Zi b(ψ)

ψ∈L2a (R3N ):∥ψ∥2 =1 ψ∈L2a (R3N ):∥ψ∥2 =1

we conclude that, for any α ∈ [0; 1],

h i
(1) (2) (1) (2)
f (αZi + (1 − α)Zi ) = inf a(ψ) + (αZi + (1 − α)Zi )b(ψ)
ψ
n h i h io
(1) (2)
= inf α a(ψ) + Zi b(ψ) + (1 − α) a(ψ) + Zi b(ψ)
ψ
h i h i
(1) (2)
≥ α inf a(ψ) + Zi b(ψ) + (1 − α) inf a(ψ) + Zi b(ψ)
ψ ψ
(1) (2)
= αf (Zi ) + (1 − α)f (Zi )

A concave function defined on the interval [0; Z] can only take its minimal value at the edge,
ie. either for Zi = 0 or for Zi = Z. Repeating this argument M times, we obtain the claim of
the proposition.

Another interesting observation concerns the role played by the restriction to permuta-
tion symmetric states. There is an important difference here between fermionic and bosonic
systems. It turns out that the restriction on bosonic states has no effect in the search for
minimizers (we will see later that the restriction on fermionic state, on the other hand, has
important consequences for the minimizers; in fact stability of the second kind, as stated in
(3.54), only holds true because of this restriction).

Proposition 3.1.3. Let E(ψ) a quadratic form on L2 (R3N ), with

i) E(ψ) ≥ C∥ψ∥2 for a constant C > −∞.

ii) E(|ψ|) ≤ E(ψ) for all ψ ∈ L2 (R3N ).

iii) E(ψπ ) = E(ψ) for all permutation π ∈ SN . Here we used the notation ψπ (x1 , . . . , xN ) =
ψ(xπ1 , . . . , xπN ).

57
Then we have

E0 = inf{E(ψ) : ∥ψ∥2 = 1} = inf{E(ψ) : ∥ψ∥2 = 1 and ψπ = ψ for all π ∈ SN }

Proof. From a) we have E0 > −∞. From b), we have

inf{E(ψ) : ∥ψ∥2 = 1} = inf{E(ψ) : ∥ψ∥2 = 1and ψ = |ψ|}

Since |ψ| is permutation symmetric if ψ is, we also obtain

inf{E(ψ) : ∥ψ∥2 = 1 and ψπ = ψ for all π ∈ SN }

= inf{E(ψ) : ∥ψ∥2 = 1, ψ = |ψ| and ψπ = ψ for all π ∈ SN }

Suppose now that ψ ≥ 0 is an unrestricted minimizer in L2 (R3N ), ie. E(ψ) = E0 and ∥ψ∥2 = 1.
Then we can decompose ψ = ψs + ψr , with
1 X
ψs = ψπ
N ! π∈S
N

the symmetrization of ψ (ψr is defined as the reminder, ie. ψr = ψ − ψs ). Observe that (since
⟨ψπ , ψσ ⟩ ≥ 0 for all π, σ ∈ SN )
1 X 1 X 1
∥ψs ∥2 = ⟨ψs , ψs ⟩ = 2
⟨ψπ , ψσ ⟩ ≥ 2
⟨ψπ , ψπ ⟩ = ∥ψ∥2 > 0
N ! π,σ N! π N!

which implies that ψs ̸= 0. Moreover, we have

E(ψ) = E(ψs ) + E(ψr ) (3.56)

To show (3.56), we define the sesquilinear form Q : L2 (R3N ) × L2 (R3N ) → C associated with
E, defined by
1 1
Q(φ, ψ) = [E(φ + iψ) − E(φ − iψ)] + [E(φ + ψ) − E(φ − ψ)]
4i 4
so that E(ψ) = Q(ψ, ψ) for all ψ ∈ L2 (R3N ). By defintion, we have Q(φπ , ψπ ) = Q(φ, ψ), for
all π ∈ SN . Therefore, we find
1 X
Q(ψs , ψ − ψs ) = Q(ψπ , ψ − ψσ )
N !2 π,σ
1 X
= Q(ψ, ψπ−1 − ψπ−1 σ )
N !2 π,σ
= Q(ψ, ψs − ψs ) = 0

58
Writing E(ψ) = Q(ψ, ψ) and decomposing ψ = ψs + ψr , we obtain (3.56). Similarly, we also
find
1 = ∥ψ∥2 = ∥ψs ∥2 + ∥ψr ∥2
Thus, we conclude

E0 = E(ψ) = E(ψs ) + E(ψr ) ≥ E0 ∥ψs ∥22 + E0 ∥ψr ∥2 = E0

and thus that E(ψs ) = E0 ∥ψs ∥2 . Since ∥ψs ∥2 ̸= 0, it follows that E(ψs /∥ψs ∥2 ) = E0 . If there
is no minimizer, we can proceed similarly, using a minimizing sequence.

Corollary 3.1.4. Consider the Hamilton operator HN,M (R, Z) defined in (3.50), for a
system with N electrons and M static nuclei. Then unrestricted minimizers of E(ψ) =
⟨ψ, HN,M (R, Z)ψ⟩ are bosonic. In other words,

inf{⟨ψ, HN,M (R, Z)ψ⟩ :ψ ∈ L2 (R3N ), ∥ψ∥2 = 1}

= inf{⟨ψ, HN,M (R, Z)ψ⟩ : ψ ∈ L2s (R3N ), ∥ψ∥2 = 1}

Proof. It is enough to show that the quadratic form E(ψ) = ⟨ψ, HN,M (R, Z)ψ⟩ satisfies the
assumptions of Proposition 3.1.3. Assumption i) is a consequence of stability of the first kind.
Assumption iii) is easy to verify. As for assumption ii), it follows from the remark that

|∇ψ(x)|2 ≥ |∇|ψ(x)||2

for almost all x ∈ Rd and any ψ ∈ H 1 (Rd ) (proof: exercise).

3.2 Lieb-Thirring Inequalities

d
Consider
PN a system of N non-interacting particles in R , described by the Hamilton operator
H = j=1 hj , with hj denoting the one-particle operator

h = −∆ + V (x) (3.57)

acting only on the variable xj . Let us assume that h has eigenvalues e1 < e2 ≤ e3 ≤ e4 ≤ . . . ,
associated with orthonormal eigenvectors φ1 , φ2 , . . . . For bosons, the operator H acts on the
symmetric subspace L2s (RdN ) of L2 (RdN ) and its ground state has the form ψN = φ⊗N 1 . For
fermions, on the other hand, we cannot have two particles in the same state. For this reason,
the ground state of H is the Slater determinant ψN (x1 , . . . , xN ) = N !−1/2 det(φi (xj ))1≤i,j≤N
and it is associated with the energy
XN
EN = ej
j=1

59
This simple computation shows why it is important to develop tools to estimate sums of
eigenvalues of one-particle operators of the form (3.57). In particular, since we are going to
consider one-particle Hamilton operator with a potential V that vanishes at infinity, we will
be interest in estimates for the sum of the negative eigenvalues of h (the continuous spectrum
starts at 0). This is the goal of the Lieb-Thirring inequalities that we are going to discuss in
this section.
To understand the origin of the Lieb-Thirring inequalities, we follow the heuristics of semi-
classical analysis, which postulates that every quantum state is associated with a volume (2π)d
in the classical phase space. According to this principle, the sum of all negative eigenvalues
of h = −∆ + V (x) can be approximated by the integral
Z
X 1
|ej | ≃ d
(p2 + V (x))χ(p2 + V (x) < 0)dxdp
j≥1
(2π) d
R ×Rd

over the classical phase-space.

p Decomposing V = V+ − V− , with V− (x) = − min(0, V (x)), and
integrating over |p| ≤ V− (x), we find
Z
X 1 2(d + 1)
|ej | ≃ |S | V− (x)(d+2)/2 dx
d d(d + 2) d−1
(3.58)
j≥1
(2π)

where |Sd−1 | denotes the area of the (d − 1)-dimensional sphere. In the appropriate (semi-
classical) limit, one can show that (3.58) holds true, in an approximate sense; this result is
known as the Weyl law. The Lieb-Thirring inequality shows that, in general (ie. not only in
the semiclassical limit), up to a multiplicative constant, the right hand side of (3.58) provides
an upper bound for the sum of the absolute value of the negative eigenvalues of h. More
generally, it provides bounds for the sum of powers of the eigenvalues of h.

Theorem 3.2.1. Fix γ ≥ 0 and assume V− = max(0, −V ) ∈ Lγ+d/2 (Rd ). Let e0 ≤ e1 ≤ e2 ≤

. . . be the non-positive eigenvalues of h = −∆ + V . If d = 1 and γ ≥ 1/2, if d = 2 and γ > 0,
or if d ≥ 3 and γ ≥ 0, there exists a constant C = C(γ, d) > 0, independent of V , such that
X Z
γ
|ej | ≤ C V− (x)γ+d/2 dx (3.59)
j≥0 Rd

Remark: For γ = 0, (3.59) gives an upper bound on the number of negative eigenvalues of
the operator h = −∆ + V . Notice that, for dimensions d = 1, 2, (3.59) does not necessarily
hold true for γ = 0. This is a consequence of the fact that, in one or two dimensions, an
arbitrarily small negative potential can already generate a negative eigenvalue (this is not
true in dimensions three or higher).

60
Proof. We show the theorem for the most important case γ = 1. We follow an argument due
to M. Rumin3 . In the exercises, we will discuss the case γ = 0.
It is enough to show that
N
X Z
⟨ϕj , hϕj ⟩ ≥ −C V− (x)1+d/2 dx (3.60)
j=1

for all finite orthonormal families {ϕj }N 2 d

j=1 in L (R ) (choose φj the eigenvectors associated
with the negative eigenvalues of h).
For any ϕ ∈ L2 (Rd ) and e > 0, we can decompose now ϕ = ϕe,+ + ϕe,− with ϕe,± defined
through Fourier transform by
e,+ (p) = ϕ(p)χ(|p|
√ e,− (p) = ϕ(p)χ(|p|
√
ϕd b > e), ϕd b ≤ e)

Using the unitarity of the Fourier transform in L2 (Rd ), we obtain

Z ∞ Z  Z ∞ Z 
e,+ 2 e,+ 2
|ϕ (x)| dx de = |ϕ (p)| dp de
d
0 0
Z Z ∞  Z "Z p2 #
= e,+ (p)|2 de dp =
|ϕd b 2 de dp
|ϕ(p)| (3.61)
0 0
Z Z
2 b 2 dp =
= |p| |ϕ(p)| |∇ϕ(x)|2 dx

Since, for any j = 1, . . . , N , we have

|ϕj (x)|2 = |ϕe,+ e,− 2 e,+ 2 e,− 2 e,+ e,−

j (x) + ϕj (x)| ≤ |ϕj (x)| + |ϕj (x)| + 2|ϕj (x)||ϕj (x)|,

we find
N
X N
X N
X N
X
|ϕj (x)|2 ≤ |ϕe,+ 2
j (x)| + |ϕe,− 2
j (x)| + 2 |ϕe,+ e,−
j (x)||ϕj (x)|
j=1 j=1 j=1 j=1

N N N
!1/2 N
!1/2
X X X X
≤ |ϕe,+ 2
j (x)| + |ϕe,− 2
j (x)| + 2 |ϕe,+
j (x)|
2
|ϕe,−
j (x)|
2

j=1 j=1 j=1 j=1


N
!1/2 N
!1/2 2
X X
= |ϕe,+
j (x)|
2
+ |ϕe,−
j (x)|
2 
j=1 j=1

3
M. Rumin. Balanced distribution-energy inequalities and related entropy bounds. Duke Math. J., 160
(2011).

61
by Cauchy-Schwarz. Thus

N
!1/2 N
!1/2 N
!1/2
X X X
|ϕj (x)|2 ≤ |ϕe,+
j (x)|
2
+ |ϕe,−
j (x)|
2

j=1 j=1 j=1

which leads us to
!1/2  !1/2 !1/2 
N
X N
X N
X
|ϕe,+
j (x)|
2
≥ |ϕj (x)|2 − |ϕe,−
j (x)|
2  (3.62)
j=1 j=1 j=1
+

where we used the notation s+ = max(0, s). Next we estimate, using the unitarity of the
Fourier transform and Bessel’s inequality,
N N Z 2
X X 1
|ϕe,−
j (x)|
2
= eip·x ϕbj (p)χ(|p| ≤e 1/2
)dp
j=1 j=1
(2π)d/2
N
1 X i.·x 2
= ⟨e χ(|.| ≤ e1/2 ), ϕbj ⟩
(2π)d j=1
(3.63)
N
1 X
\
2
= ⟨ei.·x χ(|.| ≤ e1/2 ), ϕj ⟩
(2π)d j=1
1 |Sd−1 | d/2
≤ d
∥ei.·x χ(|.| ≤ e1/2 )∥22 = e
(2π) (2π)d

Combining (3.61), (3.62), (3.63), we arrive at

 !1/2 2
N Z ∞ N 1/2
|Sd−1 |
X Z Z X
|∇ϕj (x)|2 dx ≥  |ϕj (x)|2 − ed/4  dedx
j=1 0 j=1
(2π)d/2
+

With e0 = (2π)2 |Sd−1 |−2/d ( N 2 2/d

P
j=1 |ϕj (x)| ) , we obtain
 !1/2 2
N Z e0 N 1/2
|Sd−1 |
X Z Z X
|∇ϕj (x)|2 dx ≥  |ϕj (x)|2 − ed/4  dedx
j=1 0 j=1
(2π)d/2
!1+2/d (3.64)
2 2 Z N
(2π) d X
= 2/d
|ϕj (x)|2 dx
|Sd−1 | (d + 2)(d + 4) j=1

62
 With Hölder’s inequality, we find
N N N Z
2 2
X X X
2
|∇ϕj (x)|2 − V− (x)|ϕj (x)|2 dx
   
⟨ϕj , Hϕj ⟩ = |∇ϕj (x)| + V (x)|ϕj (x)| dx ≥
j=1 j=1 j=1

N
!(d+2)/d
(2π)2 d2
Z X
≥ |ϕj (x)|2 dx
|Sd−1 |2/d (d + 2)(d + 4) j=1
 !(d+2)/d d/(d+2)
Z 2/(d+2) Z N
X
− V− (x)(d+2)/2 dx  |ϕj (x)|2 dx
j=1

Finally, we use the fact that, for any parameter a, b > 0 and any t ≥ 0,
 d/2
2 d
d/(d+2)
at − bt ≥− a−d/2 b(d+2)/2
d+2 d+2

(the minimum is attained at t = (d/(d + 2))(d+2)/2 (b/a)(d+2)/2 ) to conclude that

N  d/2
|Sd−1 |
Z
X 2 d+4
⟨ϕj , Hϕj ⟩ ≥ − d
V− (x)(d+2)/2 dx
j=1
d+2 d (2π)

This shows (3.60) and concludes the proof of the theorem (for the case j = 1).
The following bounds for the expectation of the kinetic energy in a fermionic state are a
simple consequence of the Lieb-Thirring inequality.

Theorem 3.2.2. There exists a constant C > 0 such that, for every fermionic wave function
ψ ∈ L2a (RdN ) we have
D X N E Z
(d+2)/d
ψ, −∆xj ψ ≥ C ρψ (x)dx (3.65)
j=1

where Z
ρψ (x) = N |ψN (x, x2 , . . . , xN )|2 dx2 . . . dxN
(1)
is the density of particles close to x ∈ Rd , in the state ψ. In other words, ρψ (x) = γψ (x; x)
is the diagonal of the one-particle reduced density associated with ψ.

Remark: in fact, the proof of the Lieb-Thirring inequality given above essentially already
contains the kinetic energy inequality (3.65). Decomposing the one-particle reduced density
(1) P
of ψ in its spectral representation as γψ = j λj |ϕj ⟩⟨ϕj |, with weights 0 ≤ λj ≤ 1 satisfying

63
 λj = 1 and with orthonormal wave functions ϕj ∈ L2 (Rd ), it is easy to check that (3.65) is
P
equivalent to
!1+2/d
X Z Z X
λj |∇ϕj (x)|2 dx ≥ C λj |ϕj (x)|2 dx
j j

which is a simple extension of (3.64). Below, we give a different proof, which uses the statement
of the Lieb-Thirring inequality, but not directly its proof, as given above (the original proof
of the Lieb-Thirring inequality followed a different strategy).
Proof. For an arbitrary V ∈ L(d+2)/2 (Rd ), we define the one-particle
PN operator h = −∆+V and
the many-body non-interacting Hamilton operator H = j=1 hj where, as usual, hj denotes
h acting only on particle j. We have
N
X N
X N
X Z
⟨ψ, Hψ⟩ = ⟨ψ, −∆xj ψ⟩ + ⟨ψ, V (xj )ψ⟩ = ⟨ψ, −∆xj ψ⟩ + V (x)ρψ (x)dx (3.66)
j=1 j=1 j=1

On the other hand

(1)
⟨ψ, Hψ⟩ = tr hγψ (3.67)
(1) (1)
where we recall that tr γψ = N and 0 ≤ γψ ≤ 1 because of the fermionic symmetry. This
(1)
implies that (3.67) is minimized by choosing γψ as the projection onto the N eigenvectors of
h associated with its N smallest eigenvalues (counted with multiplicity, of course). It follows
(1)
that tr hγψ is certainly larger than if we sum all negative eigenvalues of h (this even holds if
there are less than N eigenvectors below the essential spectrum of h). With the Lieb-Thirring
inequality, we obtain Z
⟨ψ, Hψ⟩ ≥ −C V− (x)1+d/2 dx (3.68)

With (3.66), it follows that

N
X Z Z
(d+2)/2
⟨ψ, −∆xj ψ⟩ ≥ −C V− (x) dx − V (x)ρψ (x)dx
j=1

2/d
for any V ∈ L(d+2)/2 (Rd ). Choosing V (x) = −Kρψ (x), for a constant K > 0 to be specified
later on, we conclude that
N
X Z
(d+2)/2 (d+2)/d
⟨ψ, −∆xj ψ⟩ ≥ (K − CK ) ρψ (x)dx
j=1

Choosing K > 0 small enough, we obtain (3.65).

64
 Remark: The fermionic symmetry of ψ plays a crucial role in (3.65). For bosonic wave
functions, we only have 0 ≤ γψ ≤ N , and thus (3.68) has to be replaced by the weaker estimate
Z
⟨ψ, Hψ⟩ ≥ −CN V− (x)1+d/2 dx

leading to the bound

N Z
D X E C (d+2)/d
ψ, −∆xj ψ ≥ 2/d ρψ (x)dx
j=1
N

3.3 Electrostatic Inequalities

In this section, we restrict our attention to three-dimensional systems. The potential associ-
ated with a possibly signed Borel measure µ on B(R3 ) is given by
Z
1
ϕ(x) = dµ(y) (3.69)
|x − y|
If µ is a positive measure, ϕ is always well-defined, although, of course, it could be infinite
(take for example µ = δ, then ϕ(x) = |x|−1 is infinity at x = 0). If µ = σ − τ is a signed
measure, given by the difference of two positive Borel measures σ, ν, the potential is not
necessarily well-defined because there could be cancellations between the contributions of σ
and ν.
In the following, we will consider potentials associated with Borel measures of the form
µ = σ − ν, with σ, ν positive Borel measures satisfying
Z Z
1 1
dσ(x), dν(x) < ∞ (3.70)
1 + |x| 1 + |x|
This assumption guarantees that ϕ ∈ L1loc (R3 ) (proof: exercise).
The Coulomb energy of a positive measure ν is defined as
Z
1 1
D(ν, ν) = dν(x)dν(y)
2 |x − y|
A charge distribution is a Borel measure µ = σ − ν, where σ, ν are positive Borel measure
satisfying (3.70) and having finite Coulomb energy, ie. D(σ, σ), D(ν, ν) < ∞. We can then
define the Coulomb energy of the charge distribution µ through
Z
1 1
D(µ, µ) = dµ(x)dµ(y)
2 |x − y|
More generally, given two charge distributions µ, τ , we can define their interaction energy
Z
1 1
D(µ, τ ) = dµ(x)dτ (y)
2 |x − y|

65
Proposition 3.3.1. Let µ be a charge distribution. Then 0 ≤ D(µ, µ) < ∞. Moreover, for
any two charge distributions µ, τ , we have

D(µ, τ )2 ≤ D(µ, µ)D(τ, τ )

Proof. Observe that

Z Z
1 1
dz = dz
|x − z|2 |y − z|2 |z|2 |(x − y) − z|2

is only a function of |x − y|. Hence, with e3 = (0, 0, 1),

Z Z
1 1 1
2 2
dz = 2
dz 2
|x − z| |y − z| |x − y| z
|z|2 |x−y| − e3
Z
1 1 C
= dz 2 =
|x − y| |z| |z − e3 |2 |x − y|

for an appropriate constant C > 0. Thus, we obtain

Z Z Z 2
1 1 1
D(µ, µ) = dµ(x)dµ(y) = dz dµ(x) ≥0
|x − y| C |x − z|2

Moreover,
Z Z Z  Z 
1 1 1 1
D(µ, τ ) = dµ(x)dτ (y) = dz dµ(x) dτ (y)
|x − y| C |x − z|2 |y − z|2
( Z Z 2 )1/2 ( Z Z 2 )1/2
1 1 1 1
≤ dz dµ(x) dz dτ (x)
C |x − z|2 C |y − z|2
= D(µ, µ)1/2 D(τ, τ )1/2

An important property of Coulomb potentials of the form (3.69) is that, for spherically
symmetric charges, they can be often controlled by the potential generated by a pointwise
charge. More precisely, we have the following Newton’s theorem.

Theorem 3.3.2. Let µ be a charge distribution, invariant w.r.t. rotations (ie. µ(R(A)) =
µ(A) for all A ∈ B(R3 ) and all R ∈ SO(3)). Then
Z Z Z
1 1 1
ϕ(x) = dµ(y) = dµ(y) + dµ(y) (3.71)
|x − y| |x| |y|<|x| |y|>|x| |y|

66
Thus, if supp µ ⊂ BR (0) (meaning that µ(A) = 0 for all A ⊂ BR (0)c ), we find

µ(R3 )
ϕ(x) =
|x|

for all |x| > R. If instead µ({y ∈ R3 : |y| < R}) = 0, then
Z
1
ϕ(x) = dµ(y)
|y|>R |y|

for all |x| < R (hence ϕ is constant on |x| < R).

Remark: if µ is a positive measure, (3.71) also implies that

µ(R3 )
ϕ(x) ≤ (3.72)
|x|

for all x ∈ R3 .
Proof. Since ϕ(Rx) = ϕ(x) for all R ∈ SO(3), we can write
Z
1
ϕ(x) = dω ϕ(|x|ω)
4π S 2
Z Z Z Z
1 1 1 1
= dµ(y) dω = dµ(y) dω
4π S2 |y − |x|ω| 4π S2 ||y|e3 − |x|ω|

We compute
Z Z π Z 2π
1 1
dω = dθ sin θ dφ p
S2 ||y|e3 − |x|ω| 0 0 |x|2 sin2 θ + (|x| cos θ − |y|)2
Z π
sin θ
= 2π p dθ
0 |x| + |y|2 − 2|x||y| cos θ
2

2π 2π 2|x| if |x| < |y|
= (||x| + |y|| − ||x| − |y||) =
|x||y| |x||y| 2|y| if |y| ≤ |x|

Thus Z Z
1 1
ϕ(x) = dµ(y) + dµ(y)
|x| |y|≤|x| |x|<|y| |y|

67
 We will need the notion of Voronoi cell with respect to a collection of nuclei of equal charge
Z. Let R1 , . . . , RM be M distinct points in R3 . We define the cell Γj ⊂ R3 by

Γj = {x ∈ R3 : |x − Rj | < |x − Ri | for all i ̸= j}

In other words, Γj is the set of points in R3 that are closer to Rj than to any other Ri , i ̸= j.
By definition Γj is open and convex. Its boundary ∂Γj consists of a finite collection of plane
segments. For j = 1, . . . , M , we set
1
Dj = min |Ri − Rj |
2 i̸=j
Then Dj is the distance from Rj to ∂Γj .
The potential produced by the nuclear configuration at x is given by
M
X 1
W (x) = Z
k=1
|x − Rk |

Note that W is harmonic on R3 \{R1 , . . . , RM }. Let

D(x) = min{|x − Ri | : 1 ≤ i ≤ M }

be the distance from x to the closest nucleus. By definition, we have D(x) = |x − Rj | for all
x ∈ Γj . We set
Z
ϕ(x) = W (x) −
D(x)
for all x ∈ R3 . For x ∈ Γj , ϕ(x) is the potential generated at x ∈ Γj by all nuclei outside Γj .
Note that ϕ is a continuous functions on R3 . ϕ is harmonic in each Voronoi cell. However, ϕ
is not harmonic on R3 , because ϕ is not differentiable on ∂Γj .
The next inequality, known as the basic electrostatic inequality, will be used to estimate
the energy of an electron due to the interaction with all nuclei, excluding the closest one.

Theorem 3.3.3. For any charge distribution µ and for every set of points R1 , . . . , RM ∈ R3 ,
we have
M M
Z2 Z2 X 1
Z X
D(µ, µ) − ϕ(x)dµ(x) + ≥
k<ℓ
|Rk − Rℓ | 8 j=1 Dj

Proof. First of all, we would like to find a charge distribution ν generating the potential ϕ, in
the sense that Z
1
ϕ(x) = dν(y)
|x − y|

68
which implies −∆ϕ = 4πν. Therefore, for f ∈ Cc∞ (R3 ), we compute
Z Z M Z
X
−4π f (x)dν(x) = ϕ(x)∆f (x)dx = ϕ(x)∆f (x)dx
j=1 Γj

M Z
X M Z
X
= ∇ · (ϕ(x)∇f (x))dx − ∇ϕ(x) · ∇f (x)dx (3.73)
j=1 Γj j=1 Γj

M Z
X M Z
X
= ϕ(x)∇f (x) · nj dσ − ∇ϕ(x) · ∇f (x)dx
j=1 ∂Γj j=1 Γj

Notice that every segment in ∂Γj appears twice, with opposite sign. Since ∇f and ϕ are
continuous, the first term on the r.h.s of (3.73) vanishes. Let us focus on the second term on
the r.h.s of (3.73). We have
M Z
X M Z
X M Z
X
− ∇f (x) · ∇ϕ(x)dx = − ∇ · (f (x)∇ϕ(x)) dx + f (x)∆ϕ(x)dx
j=1 Γj j=1 Γj j=1 Γj

Since ∆ϕ = 0 inside every cell Γj , we arrive at

M Z
X M Z
X
− ∇f (x) · ∇ϕ(x)dx = − f (x)∇ϕ(x) · nj dσ
j=1 Γj j=1 ∂Γj

We conclude that
Z M Z
X
−4π f (x)dν(x) = − f (x)∇ϕ(x) · nj dσ (3.74)
j=1 ∂Γj

Although every segment ∂Γj appears twice, in opposite directions, the integral is not zero,
because ∇ϕ jumps through ∂Γj . Recall that
Z
ϕ(x) = W (x) −
D(x)
Since W is differentiable on ∂Γj (W is differentiable everywhere away from the nuclei), its
contribution to the r.h.s. of (3.74) vanishes. On the segment ∂Γj ∩ ∂Γℓ , we have
1 1
nj · ∇ = nℓ · ∇
|x − Rj | |x − Rℓ |
and thus Z Z
1
−4π f (x)dν(x) = 2Z f (x) nj · ∇ dσ
S
j ∂Γj |x − Rj |

69
We conclude that ν is a surface charge density, concentrated on the boundaries ∪j ∂Γj . On
the segment ∂Γj ∩ ∂Γℓ , we find that ν has the magnitude

Z 1 Z 1
− nj · ∇ = − nℓ · ∇
2π |x − Rj | 2π |x − Rℓ |

Notice that
1 (x − Rj )
−nj · ∇ = nj · ≥0
|x − Rj | |x − Rj |3
by the convexity of the Voronoi cells. Thus, ν is a positive measure. With it, we can write
Z Z
1
ϕ(x)dµ(x) = dν(y)dµ(x) = 2D(µ, ν) = 2D(ν, µ)
|x − y|

Hence
M M
Z2 Z2
Z X X
D(µ, µ) − ϕ(x)dµ(x) + = D(µ − ν, µ − ν) − D(ν, ν) +
k<ℓ
|Rk − Rℓ | k<ℓ
|Rk − Rℓ |
X Z2
≥ −D(ν, ν) +
|Rk − Rℓ |
(3.75)

because D(µ − ν, µ − ν) ≥ 0 by positivity of the Coulomb energy. To conclude the proof, we

compute D(ν, ν). To this end, we notice that
M N Z
X 1 X 1
W (x) = Z =Z δ(y − Rj )dy
j=1
|x − Rj | j=1 R3 |y − x|

Thus
Z M Z Z M M
X 1 X X 1
W (x)ν(dx) = Z dy δ(y − Rj ) dν(x) =Z ϕ(Rj ) = 2Z 2
j=1 R3 |x − y| j=1 k<ℓ
|Rk − Rℓ |

Hence
Z Z Z Z
1 1 1 1 Z 1
D(ν, ν) = dν(x)dν(y) = ϕ(x)dν(x) = W (x)dν(x) − dν(x)
2 |x − y| 2 2 2 D(x)
M Z
2
X 1 Z 1
=Z − dν(x)
k<ℓ
|R k − Rℓ | 2 D(x)

70
From (3.75), we conclude that
M
Z2
Z X
D(µ, µ) − ϕ(x)dµ(x) +
k<ℓ
|Rk − Rℓ |
M Z
Z2 X
Z
Z 1 1 1
≥ dν(x) = − nj · ∇ dσ
2 D(x) 8π j=1 ∂Γj D(x) |x − Rj |
M Z
Z2 X 1 1
=− nj · ∇ dσ
8π j=1 ∂Γj |x − Rj | |x − Rj |
M Z M Z
Z2 X 1 Z2 X 1
=− nj · ∇ dσ = ∆ dx
16π j=1 ∂Γj |x − Rj |2 16π j=1 Γcj |x − Rj |2
M Z
Z2 X 1
= dx
8π j=1 Γcj |x − Rj |4

Since the integrand is positive, the integral over Γcj is larger than if we integrate over the whole
half-plane defined by the prolongation of the segment of ∂Γj at distance Dj from Rj . Hence
M
Z2
Z X
D(µ, µ) − ϕ(x)dµ(x) +
k<ℓ
|Rk − Rℓ |
M Z Z ∞
Z2 X
Z
1
≥ dy dz 2 2 2 2
dx
8π j=1 R R Dj (x + y + z )

M Z M Z M
Z2 X ∞
Z ∞
1 Z2 X ∞ 1 Z2 X 1
= rdr dx 2 = dx =
4 j=1 0 Dj (r + x2 )2 8 j=1 Dj x2 8 j=1 Dj

As an application of the basic electrostatic inequality (Theorem 3.3.3), we can show a

pointwise bound on the potential energy of N electrons interacting with M nuclei, all having
charge Z, known as Baxter’s inequality.

Theorem 3.3.4. Let

N N M M
X 1 XX Z X Z2
V (x, R) = − +
i<j
|xi − xj | i=1 j=1 |xi − Rj | k<ℓ |Rk − Rℓ |

71
As above, we denote by Γj ⊂ R3 the Voronoi cell with Rj , D(x) = min{|x − Rj |, j = 1, . . . , M }
and Dj = min{|Ri − Rj |/2, i ̸= j}. Then
M M
X 1 Z2 X 1
V (x, R) ≥ −(2Z + 1) +
j=1
D(xj ) 8 j=1 Dj

Proof. We fix x = (x1 , . . . , xN ). Consider the measure on R3 defined by

1
µi (dx) = δ(|x − xi | − di /2)dx
d2i π
where we set di = D(xi ). Note that (setting y = x − xi )
Z Z
1 1
µi (dx) = 2 δ(|y| − di /2)dy = 4π(di /2)2 2 = 1
R3 di π di π
Thus µi is a positive probability measure (with total mass one). Next, we observe that, by
Newton’s theorem (more precisely, from the bound (3.72)),
Z Z
1 1 1 1 1 1
D(µi , µj ) = dµi (x)dµj (y) ≤ µj (dy) ≤
2 |x − y| 2 |xi − y| 2 |xi − xj |
Moreover, Z
Z Z
dµi (y) =
|y − Rj | |xi − Rj |
Therefore
N N X
M Z M
X X 1 X Z2
V (x, R) ≥ 2D(µi , µj ) − Z dµi (x) +
i<j i=1 j=1
|x − Rj | ℓ<k
|Rℓ − Rk |
PN
Take µ = i=1 µi . Then
N N Z
X X 1 X Z2
V (x, R) ≥ D(µ, µ) − D(µi , µi ) − Z dµ(x) +
i=1 j=1
|x − Rj | ℓ<k
|Rℓ − Rk |

where
Z
1 1 1
D(µi , µi ) = 2 2
dxdy δ(|xi − x| − di /2)δ(|y − xi | − di /2)
2 (di π) |x − y|
δ(|z| − di /2)δ(|w| − di /2)
Z Z
1 1 1 1
= 2 2
dzdw = 2
dω1 dω2
2 (di π) |z − w| 16π di S2 ×S2 |ω1 − ω2 |
Z π Z 2π
1 1 1
= dθ sin θ dφ =
4πdi 0 0 |(0, 0, 1) − (sin θ cos φ, sin θ sin φ, cos θ)| di

72
Thus
N Z M N
X 1 X Z2 X 1
V (x, R) = D(µ, µ) − Z dµ(x) + −
j=1
|x − Rj | k<ℓ
|Rk − Rℓ | i=1 di
With
PM the basic electrostatic inequality in Theorem 3.3.3 (and recalling that ϕ(x) =
Z k=1 1/|x − Rk | − Z/D(x)), we obtain
N M
Z2 X 1
Z
1 X 1
V (x, R) ≥ −Z µ(dx) − +
D(x) d
i=1 i
8 j=1 Dj
N Z M
Z2 X 1

X Z 1
≥− dµi (x) + +
i=1
D(x) di 8 j=1 Dj

Now note that, for x ∈ supp µi , we have |x − Rℓ | ≥ |xi − Rℓ | − |xi − x| ≥ di /2 for all ℓ. Thus
D(x) ≥ di /2 on supp µi . This implies that
N M
X 1 Z2 X 1
V (x, R) ≥ −(2Z + 1) +
d
j=1 i
8 j=1 Dj

3.4 Indirect Part of the Coulomb Energy

For any antisymmetrix wave function ψN ∈ L2a (R3N ), we define
N N Z
D X 1 E X 1
Iψ = ψ, ψ = dx1 . . . dxN |ψ(x1 , . . . , xN )|2
i<j
|xi − xj | i<j
|xi − xj |

Recall moreover the definition

Z
ρψ (x) = N dx2 . . . dxN |ψ(x, x2 , . . . , xN )|2

of the space density associated with ψ, normalized so that

Z
ρψ (x)dx = N

The space density is defined so that ρψ (x) = γ (1) (x; x), where γ (1) is the one-particle reduced
density matrix associated with ψ. The electrostatic energy associated with the electron density
ρψ is then given by Z
1 ρψ (x)ρψ (y)
D(ρψ , ρψ ) = dxdy
2 |x − y|

73
 Note that if ψ ∈ L2 (R3 ) was a product, we would clearly have Iψ = D(ρψ , ρψ ), because the
two-particle reduced density that is needed to compute Iψ would just be given by the product
of two copies of the one-particle reduced density (like in probability, product wave functions
mean that the particles are independent, so that the probability to find a particle at xi and
a particle at xj would just be the product of the two probabilities). Thus, the difference
between Iψ and D(ρψ , ρψ ) comes from correlation among particles (which can arise because of
the interaction or because of the antisymmetry). In the next theorem, we give a lower bound
for the difference Iψ − D(ρψ , ρψ ), which is known as the indirect part of the Coulomb energy
(while D(ρψ , ρψ ) is known as the direct part of the Coulomb energy).
Theorem 3.4.1. For every ψ ∈ L2a (R3N ), we have
Z
4/3
Iψ − D(ρψ , ρψ ) ≥ −C ρψ (x)dx

Proof. The proof is divided into several intermediate steps.

Step 1. For any z ∈ R3 , let µz be a non-negative, spherical symmetric w.r.t. z, bounded
function with Z
µz (y)dy = 1

Then for every ρ ∈ L1 (R3 ) non-negative, we have

N N N
X 1 X X
≥ −D(ρ, ρ) + 2 D(ρ, µxi ) − D(µxi , µxi ) (3.76)
i<j
|xi − xj | i=1 i=1

To show (3.76), we start by observing that

N N
! N N
X X X X
0≤D ρ− µ xi , ρ − µxi = D(ρ, ρ) − 2 D(ρ, µxi ) + D(µxi , µxj )
i=1 i=1 i=1 i,j

Thus
X N
X N
X
D(µxi , µxj ) ≥ −D(ρ, ρ) + 2 D(ρ, µxi ) − D(µxi , µxi )
i̸=j i=1 i=1

Using the spherical symmetry of µxi , we have, by Newton’s theorem,

Z Z
1 µxi (x)µxi (y) 1 1 1 1
D(µxi , µxj ) = dxdy ≤ dxµxi (x) ≤
2 |x − y| 2 |x − xj | 2 |xi − xj |
Therefore
N N N N
X 1 X X X
≥ D(µxi , µxj ) ≥ −D(ρ, ρ) + 2 D(ρ, µxj ) − D(µxj , µxj )
i<j
|xi − xj | i̸=j j=1 j=1

74
 Step 2. We have
Iψ − D(ρψ , ρψ ) ≥ −F1 − F2 (3.77)
with
ρψ (x)(δz (y) − µz (y))
Z
F1 = ρψ (z)dxdydz
|x − y|
and Z
1 µz (x)µz (y)
F2 = ρψ (z)dxdydz
2 |x − y|
for any µ as in Step 1.
To show (3.77), we choose ρ = ρψ and we apply Step 1. We find
N N Z
D X 1 E X
Iψ = ψ, ψ ≥ −D(ρψ , ρψ ) + 2 dx1 . . . dxN |ψ(x1 , . . . , xN )|2 D(ρψ , µxj )
i<j
|x − y| j=1
N
X Z
− dx1 . . . dxN |ψ(x1 , . . . , xN )|2 D(µxj , µxj )
j=1
Z Z
= −D(ρψ , ρψ ) + 2 dzρψ (z)D(ρψ , µz ) − dzρψ (z)D(µz , µz )

Hence
Z Z
Iψ − D(ρψ , ρψ ) ≥ −2D(ρψ , ρψ ) + 2 dzρψ (z)D(ρψ , µz ) − dzρψ (z)D(µz , µz )

Note that
Z Z Z
1 ρψ (x)δz (y)
dzρψ (z)D(ρψ , δz ) = dzρψ (z) dxdy
2 |x − y|
Z
1 ρψ (x)ρψ (y)
= dxdy = D(ρψ , ρψ )
2 |x − y|

It follows that
Z Z
Iψ − D(ρψ , ρψ ) ≥ −2 dz ρψ (z)D(ρψ , δz − µz ) − dz ρψ (z)D(µz , µz ) = −F1 − F2

R
Step 3. Suppose now µ e ≥ 0 is bounded and such that µ e(x)dx = 1 and µ e(x) = 0 if
−1
|x| > 1. For example, we can take µ
e(x) = (4π/3) if |x| ≤ 1 and µ
e(x) = 0 if |x| > 1. Set
1/3
µ(ρψ (z)(x − z))
µz (x) = ρψ (z)e (3.78)

75
Then we claim that Z
F2 = C ρψ (y)4/3 dy . (3.79)

To show (3.79) we observe that

1/3 1/3
e(ρψ (z)(y − z))e
µ µ(ρψ (z)(x − z))
Z Z
1 µz (x)µz (y)
F2 = dxdydz ρψ (z) = dxdydzρ3ψ (z)
2 |x − y| |x − y|
1/3 1/3
We introduce new variables w = ρψ (z)(y − z), t = ρψ (z)(x − z). We find
Z Z Z
1 4/3 µ
e(w)e
µ(t) 4/3 4/3
F2 = dzdwdt ρ (z) = D(e µ, µ
e) ρψ (z)dz = C ρψ (z)dz
2 |w − t|

Step 4. With the same choice of µz as in Step 3, we have

Z
4/3
F1 ≤ C ρψ (x)dx (3.80)

To show (3.80), we compute

Z
ρψ (x) h 1/3
i
F1 = dxdydz ρψ (z) δ(z − y) − ρψ (z)e
µ(ρψ (z)(z − y))
|x − y|
1/3
µe(ρψ (z)(z − y))
Z Z Z
ρψ (x)ρψ (y)
= dxdy − dxdzρψ (x)ρψ (z) dyρψ (z)
|x − y| |x − y|
We observe that
1/3
e(ρψ (z)(z − y))
µ
Z Z
dxdzρψ (x)ρψ (z) dyρψ (z)
|x − y|
Z Z
4/3 µ
e(y)
= dxdz ρψ (x)ρψ (z) de y 1/3
|ρψ (z)(x − z) − ye|
Z
4/3
= dxdzρψ (x)ρψ (z)ϕ(ρ1/3 (z)|x − z|)

where we introduced the notation

Z
µ
e(y)
ϕ(t) = dt
|y − t|
for the potential generated by µ
e. We obtain
Z Z Z
ρψ (x)ρψ (y) 4/3 1/3
F1 = dxdy − dxdyρψ (x)ρψ (y)ϕ(ρ (z)|x − z|) = dxρψ (x)R(ρψ (y), |x − y|)
|x − y|

76
with
a
− a4/3 ϕ(a1/3 r)
R(a, r) =
r
e is spherically symmetric, Newton’s theorem implies that ϕ(r) ≤ 1/r. Thus R(a, r) ≥ 0.
Since µ
Moreover, R(a, r) = 0 if a1/3 r ≥ 1. We conclude that

a/r if a1/3 r < 1
0 ≤ R(a, r) ≤
0 if a1/3 r ≥ 1
which implies that Z
ρψ (x)ρψ (y)
F1 ≤ dxdy
ρ1/3 (y)|x−y|≤1 |x − y|
We use the representation Z ∞
ρψ (x) = dα χ(ρψ (x) ≥ α)
0
With χα (x) = χ(ρψ (x) ≥ α), we find
Z ∞ Z ∞ Z
χα (x)χβ (y)
F1 ≤ dα dβ dxdy
0 0 β 1/3 |x−y|≤1 |x − y|
Z ∞ Z α Z Z ∞ Z ∞ Z
χα (x) χβ (y)
≤ dα dβ dxdy + dα dβ dxdy
0 0 β 1/3 |x−y|≤1 |x − y| 0 α β 1/3 |x−y|≤1 |x − y|
=: I + II
We estimate
Z ∞ Z α Z   Z
dy
I= dα dβ χα (x)dx
0 0 |y|≤β −1/3 |y|
Z ∞ Z α Z Z ∞ Z
−2/3 1/3
≤ 2π dα dββ χα (x)dx = 6π dαα χα (x)dx
0 0 0

and
Z ∞ Z ∞ Z Z ∞ Z Z ∞
−2/3
II = 2π dα dββ
dyχβ (y) = 2π dα dy dββ −2/3 χ(ρψ (y) ≥ β)
0 α 0 α
Z ∞ Z Z ρψ (y)
= 2π dα dy χ(ρψ (y) ≥ α) dβ β −2/3
0 α
Z ∞ Z
1/3
= 6π dα dyχα (y)(ρψ (y) − α1/3 )
0

From the last two bounds, we conclude that

Z ∞ Z Z
1/3 4/3
F1 ≤ I + II = 6π dα dyρψ (y)χα (y) = C dyρψ (y)
0

77
3.5 Proof of Stability
We are going to give two proofs of (3.54). First of all, we recall notations that will be used.
Let N, M ∈ N, R = (R1 , . . . , RM ) ∈ R3M , Z = (Z1 , . . . , ZM ) ∈ RM
+ . Let x ∈ R
3N
. Then we
define
N M N XM
X 1 X Zi Zj X Zi
V (R, Z, x) = + −
i<j
|xi − xj | i<j |Ri − Rj | i=1 j=1 |xi − Rj |
and the Hamilton operator
N
X
HN,M (R, Z) = −∆xj + V (R, Z, x) (3.81)
j=1

acting as a self-adjoint operator on the space L2a (R3N ).

Theorem 3.5.1. There exists C > 0 such that

inf inf ⟨ψ, HN,M (R, Z)ψ⟩ ≥ −C(1 + 2Z)2 (N + M )

R∈R3M ψ∈L2a (R3N ):∥ψ∥2 =1

where Z = max(Z1 , . . . , ZM ).
Proof. It follows from Prop. 3.1.2 that it is enough to consider the case Z1 = · · · = ZM = Z.
Under this assumption, we can use Baxter’s inequality, Theorem 3.3.4, to bound
N M N
X 1 Z2 X 1 X 1
V (R, Z, x) ≥ −(2Z + 1) + ≥ −(2Z + 1)
i=1
D(xi ) 8 i=1 Dj i=1
D(xi )

where D(xi ) = minj=1,...,M |xi − Rj | and Dj = (1/2) mini̸=j |Ri − Rj |. Hence, we have
N  
X (2Z + 1)
⟨ψ, HN,M (R, Z)ψ⟩ ≥ ⟨ψ, −∆xj − ψ⟩
j=1
D(xj )
N   
X 1
= ⟨ψ, −∆xj − (2Z + 1) − b ψ⟩ − (2Z + 1)b∥ψ∥22 N
j=1
D(x j )

for an arbitrary b > 0 (we introduced the parameter b > 0 to avoid divergencies arising from
the slow decay of the Coulomb potential at infinity).
By the Lieb-Thirring inequality, we find
N    Z  5/2
X 1 5/2 1
⟨ψ, −∆xj − (2Z + 1) −b ψ⟩ ≥ −C(2Z + 1) −b dx
j=1
D(xj ) D(x) +

78
We have
 5/2  5/2 M  5/2
1 1 X 1
−b = max −b ≤ −b
D(xj ) +
j=1,...,M |x − Rj | + j=1
|x − Rj | +

Thus
Z  5/2 M Z  5/2
1 X 1
−b dx ≤ −b dx
D(x) + j=1
|x − R j | +
M Z  5/2
X 1
≤ −b dx
j=1 |x−Rj |<1/b
|x − Rj |
(3.82)
Z Z  5/2
−1 5/2 5/2 1
=M (|x| − b) dx = M b −1 dx
|x|<1/b b|x|≤1 b|x|
Z  5/2
1
= M b−1/2 −1 = CM b−1/2
|y|≤1 |y|
Thus, we conclude that
inf inf ⟨ψ, HN,M (R, Z)ψ⟩ ≥ −C(2Z + 1)5/2 M b−1/2 − (2Z + 1)bN
R∈R3M ψ∈L2a (R3N ):∥ψ∥2 =1

Optimizing over b > 0 (choosing b = C(2Z + 1)(M/N )2/3 ), we obtain

inf inf ⟨ψ, HN,M (R, Z)ψ⟩
R∈R3M ψ∈L2a (R3N ):∥ψ∥2 =1

≥ −C(2Z + 1)2 N 1/3 M 2/3 ≥ −C(2Z + 1)2 (N + M )

We can also give a second proof of stability of the second kind.

Theorem 3.5.2. There exists C > 0 with
inf inf ⟨ψ, HN,M (R, Z)ψ⟩ ≥ −CN (1 + Z 2 (N/M )2/3 )
R ψ∈L2a (R3N ):∥ψ∥2 =1

Proof. Again, by Prop. 3.1.2, it is enough to consider the case Z1 = · · · = ZM = Z. From the
kinetic energy inequality, Theorem 3.2.2, and the estimate for the indirect Coulomb energy,
Theorem 3.4.1, we find
Z Z
5/3 4/3
⟨ψ, HN,M (R, Z)ψ⟩ ≥ C ρψ (x) dx + D(ρψ , ρψ ) − C ρψ (x)dx
M Z M
X Z X Z2
− ρψ (x)dx +
j=1
|x − Ri | i<j
|Ri − Rj |

79
With the basic electrostatic inequality, Theorem 3.3.3, we obtain
Z Z Z
5/3 4/3 Z
⟨ψ, HN,M (R, Z)ψ⟩ ≥ C ρψ (x) dx − C ρψ (x)dx − C ρψ (x)dx
D(x)

We bound now
4/3 1/3 5/3
ρψ (x) = ρψ (x)ρψ (x) ≤ aρψ (x) + a−1 ρψ (x)
Thus, choosing a > 0 small enough,
Z Z  
5/3 1
⟨ψ, HN,M (R, Z)ψ⟩ ≥ C ρψ (x) dx − CZ − b ρψ (x)dx − CN − CZbN
D(x) +

With Hölder, we conclude that

Z
5/3
⟨ψ, HN,M (R, Z)ψ⟩ ≥ − CN (1 + bZ) + C ρψ (x)dx
!2/5 Z
Z   5/2 3/5
1 5/3
− CZ −b dx ρψ (x)dx
D(x) +

Minimizing over all possible ρψ , using namely that

inf t − at3/5 = −Ca5/2

t>0

and proceeding as in (3.82), we obtain

Z  5/2
5/2 1
⟨ψ, HN,M (R, Z)ψ⟩ ≥ −CN (1 + bZ) − CZ −b dx
D(x) +
= −CN (1 + bZ) − CZ 5/2 M b−1/2

Minimizing over b > 0, we conclude that

⟨ψ, HN,M (R, Z)ψ⟩ ≥ −CN (1 + Z 2 (M/N )2/3 )

Both proofs of stability used the restriction to antisymmetric wave functions (the Lieb-
thirring inequality and the kinetic energy inequality both rely on the fermionic statistics). It
turns out that stability does not hold true for bosons (particles described by wave functions
that are symmetric w.r.t. permutations).

80
Theorem 3.5.3. Consider the Hamilton operator (3.81), but now acting on the bosonic
subspace L2s (R3N ) of permutation symmetric wave functions. Assume that Zj = Z for all
j = 1, . . . , M , and N = ZM (neutrality; these assumptions are not important, they only
simplify slightly the computations). Then, we have
inf inf ⟨ψ, HN,M (R, Z)ψ⟩ ≤ −CZ 4/3 N 5/3
R ψ∈L2s (R3N ):∥ψ∥2 =1

Proof. For a fixed g ∈ L2 (R3 ), we set ϕλ (x) = λ3/2 g(λx). We consider the bosonic wave
function
N
Y
ψ(x1 , . . . , xN ) = ϕλ (xj )
j=1
Then we find Z
2
⟨ψ, HN,M (R, Z)ψ⟩ = N λ |∇g(x)|2 dx + λW (N, R)
with
M M
|g(x)|2 |g(y)|2 |g(x)|2 Z2
Z Z
1 X X
W (N, R) = N (N − 1) dxdy − ZN dx +
2 R3 ×R3 |x − y| j=1
|x − Rj | i<j
|Ri − Rj |
We choose g with compact support. We divide supp g into M cells Γj , j = 1, . . . , M , with
Z
|g(x)|2 dx = 1/M (3.83)
Γj

for all j = 1, . . . , M . We arrange one nucleus in each cell Γj (since we are trying to show a
bound for the infimum over all possible locations of the M nuclei, we can choose where to
put them). We average over the position of the nucleus in the cell Γj , according to the weight
M |g(x)|2 (because of (3.83), M |g|2 is a probability density over Γj ). The average value of
W (N, R) is given by
|g(x)|2 |g(y)|2
Z
1
ER W (N, R) = N (N − 1) dxdy
2 R3 ×R3 |x − y|
M Z
X |g(x)|2 M |g(y)|2
− ZN dxdy
3
j=1 R ×Γj
|x − y|
M Z
2
X M |g(x)|2 M |g(y)|2
+Z dxdy
i<j Γ j ×Γj
|x − y|
|g(x)|2 |g(y)|2
 Z
1 1 2 2
= N (N − 1) − ZN M + Z M dxdy
2 2 R3 ×R3 |x − y|
M Z
1 2 2X |g(x)|2 |g(y)|2
− Z M dxdy
2 j=1 Γj ×Γ j
|x − y|

81
If ZM = N , we find
1 1 N
N (N − 1) − ZN M + Z 2 M 2 = − < 0
2 2 2
Hence, there exists R0 , with R0,j ∈ Γj , such that
M
|g(x)|2 |g(y)|2
Z
1 X
W (N, R0 ) ≤ − N 2 dxdy
2 j=1 Γj ×Γj |x − y|

Let now rj > 0 be the radius of the smallest ball containing Γj . Then

|g(x)|2 |g(y)|2
Z
1
dxdy ≥ inf D(ρ, ρ)
2 Γj ×Γj |x − y|
R
where the infimum is taken over all ρ ≥ 0, with supp ρ ⊂ Brk and ρ(x)dx = 1/M . The infi-
mum is attained distributing the charge uniformly on the sphere of radius rk (proof: exercise).
Hence:
|g(x)|2 |g(y)|2
Z Z
1 1 1 1
dxdy ≥ 2
dσdω =
2 Γj ×Γj |x − y| 2(4πM ) S 2 ×S 2 |rk σ − rk ω| 2rk M 2

which implies that

M
1 2 1 X 1 1 1 1
W (N, R0 ) ≤ − Z M ≤ − Z 2M 1
PM ≤ − Z 2 M 4/3
4 M k=1 rk 4 M k=1 rk
4

where we used the convexity of x → 1/x on R+ in the second inequality and the fact that,
with an appropriate choice of g and the cells Γj , we can make sure that the average radius
M −1 M −1/3
(a possible choice of g, {Γj }M
P
j=1 rj ≤ CM j=1 is discussed on the Lieb-Seiringer
book). We conclude that

⟨ψ, HN,M (R0 , Z)ψ⟩ ≤ CN λ2 − CλZ 2 M 4/3

Minimizing over λ (which leads to the choice λ = CN 1/3 Z 2/3 ), we arrive at

⟨ψ, HN,M (R0 , Z)ψ⟩ ≤ −CZ 4/3 N 5/3

82
4 Thomas-Fermi Theory
The energy of a system of N electrons and M nuclei of charges Z1 , . . . , ZM located at
R1 , . . . , RM ∈ R3 is obtained by minimizing the expectation of the Hamilton operator
N
X
HN,M (R, Z) = −∆xj + αVC (x, R)
j=1

with
N M N M
X 1 X Zi Zj XX Zj
VC (x, R) = + −
i<j
|xi − xj | i<j |Ri − Rj | i=1 j=1 |Rj − xi |

For large values of N + M , it is practically impossible to compute the infimum, neither

analytically nor numerically. For this reason, effective theories have been developed which, on
the one hand, are mathematically simpler to approach and, on the other hand, approximate the
full many body quantum theory in interesting regimes. The simplest such theory is Thomas-
Fermi theory. The goal of this section is to define Thomas-Fermi theory, to study its basic
mathematical properties and to discuss in which sense and in which regime Thomas-Fermi
theory provides a good approximation to the quantum theory.

4.1 Definition of Thomas-Fermi Functional and Existence of Min-

imizer
The basic idea of Thomas-Fermi theory is to describe the N electrons in terms of an electron
density ρ ∈ L1 (R3 ) defined so that ρ(x)dx is the average number of electrons close to x. The
density is therefore normalized so that
Z
dx ρ(x) = N ∈ N .

Starting form an N electron wave function ψ ∈ L2 (R3 × {1, . . . , q})⊗a N we can compute the
electron density
X Z
ρψ (x) := N dx2 . . . dxN |ψ(x, σ1 , . . . , xN , σN )|2
σ1 ,..,σN

In TF, the idea is to forget about the complicated N -particle wave function ψ and, instead,
focus only on the simpler electron density ρ. Of course, the problem is that the energy of the

83
N -particle state with wave function ψ cannot be expressed using only the density ρ. In fact
N Z
X
⟨ψ, HN,M ψ⟩ = dz1 . . . dzN |∇xj ψ(z1 , . . . , zN )|2
j=1
N Z
X 1
+ dz1 . . . dzN |ψ(z1 , . . . , zN )|2 (4.84)
i<j
|x i − x j |
M N X M Z
X Zi Zj X 1
+ − Zj dz1 . . . dzN |ψ(z1 , . . . , zN )|2
i<j
|Ri − Rj | i=1 j=1 |xi − Rj |

where we used the notation z = (x, σ) ∈ R3 × {1, . . . , q}, and, to simplify the expression, we
assumed that the coupling constant α = 1. The terms in the last line can indeed be expressed
in terms of the density ρ, since
N XZ Z
X 1 1
dx1 . . . dxN |ψ(x1 , σ1 , . . . , xN , σN )|2 = dx ρψ (x)
i=1 σ
|xi − Rj | |x − Rj |

The first two terms on the r.h.s. of (4.84), however, cannot be expressed through ρψ . Nev-
ertheless, we can try to approximate these terms with expressions depending only on ρψ . To
approximate the kinetic energy, we use the kinetic energy inequality, which states that
N
X Z
5/3
⟨ψ, −∆xj ψ⟩ ≥ C ρψ (x)dx
j=1 R3

Although this is only a lower bound, one can hope (and we will show) that in certain regimes
the r.h.s. represents a good approximation for the kinetic energy. To approximate the interac-
tion among the electrons, on the other hand, we will neglect the indirect part of the Coulomb
energy, and we will use the direct Coulomb energy
Z
1 1
D(ρψ , ρψ ) = dxdyρψ (x)ρψ (y)
2 |x − y|
Recall that approximating the Coulomb interaction by its direct part means that we are
neglecting the correlations among the electrons; we replace essentially the expectation of the
products of the number of electrons at x and at y by the product of the expectations. As
we will see later on, it turns out that this approximation (like the replacement of the kinetic
energy) is indeed justifiable in the limit of large Z (in the case of neutral systems, N = Z).
We arrive therefore to the Thomas-Fermi energy functional
Z Z
3 5/3
εT F (ρ) = γ dx ρ (x) − dx ρ(x)V (x) + D(ρ, ρ) + U (4.85)
5

84
for some γ > 0 and where
M M
X Zj X Zi Zj
V (x) = and U= .
j=1
|x − Rj | i<j
|Ri − Rj |

Note that although the Thomas-Fermi functional is well defined for any value of γ > 0, we
can only expect it to approximate the full quantum theory for a precise value of γ (as we will
see, in units with m = 1/2 and ℏ = 1, the physically relevant value is γ = (6π 2 /q)2/3 , where
q is the dimension of the spin space).
5/3 3
In order for the first term to be finite, we can restrict our attention to ρ ∈ L R (R ). Since
1 3
ρ is interpreted as a density, we also require that ρ ≥ 0, ρ ∈ L (R ) with ρ(x)dx = N .
Under these conditions, it is simple R to check (by Hölder and the Hardy-Littlewood-Sobolev
inequality) that also D(ρ, ρ) and ρ(x)V (x) are finite.
We obtain the main postulate of Thomas-Fermi theory; the energy of a system of N
electrons and M nuclei of charge Z1 , . . . , ZM located at R1 , . . . , RM is given, in Thomas-Fermi
theory, by the minimization of the functional εT F (ρ) over all ρ in the set
Z
5/3 3 1 3
FN := {ρ ∈ L (R ) ∩ L (R ), ρ ≥ 0, ρ(x)dx = N } .

Note that while the parameters M and Z1 , . . . , ZM enter the functional R εT F directly, the
number of electrons N only enters through the normalization condition dxρ(x) = N . It
makes sense, therefore, to think of M and of Z1 , . . . , ZM as fixed, and, on the other hand, to
allow N to vary.
For fixed M and Z1 , . . . , ZM , we should not expect that a minimizer for εT F exists for
arbitrary N . If N is too large, we know from the discussion about binding that a minimizer
for the full Schrödinger energy does not exist. Hence, if N is too large we do not expect a
minimizer for εT F over the set FN to exist. Instead, if N is very large, in order to approximate
the infimum of εT F over FN , we will have to move some of the N electrons towards spatial
infinity.
By the same arguments, we always expect a minimizer of εT F to exist over the larger set
Z
5/3 3 1 3
DN := {ρ ∈ L (R ) ∩ L (R ), ρ ≥ 0, ρ(x)dx ≤ N }

where we only prescribe an upper bound for the number of electrons. In this case, instead of
moving the electrons to infinity, to approximate the infimum we just decrease their number.
The plan for this section is as follows. First, we show that a minimizer for εT F on the set
DN always exist. Then we will study the properties of this minimizer P and we will conclude
that a minimizer for εT F on the set FN exists if an only if N ≤ Ztot = M j=1 Zj . This means
that, within Thomas-Fermi theory, binding occurs for N ≤ Ztot (there is no stable negative
ion in Thomas-Fermi theory).

85
 The first step to show the existence of a minimizer on DN consists in proving a lower
bound to the Thomas-Fermi energy.
Theorem 4.1.1. There exists a universal constant C > 0 such that
M
!1/3
5/3
X
εT F (ρ) ≥ −CZtot Zj2
j=1

for all ρ ∈ DN .
Remarks. The bound is uniform in N . We are able to find a bound which is uniform in the
number of electrons because M and Z1 , . . . , ZM are kept fixed, and therefore (as we will see
shortly) increasing N above a critical level does not decrease the energy any more. Following
the ideas used in the proof of stability of matter (here, one only needs the monotonicity in the
nuclear charge and the basic electrostatic inequality) one can of course also derive a bound of
the form εT F (ρ) ≥ −C(Z)(N + M ) where C(Z) only depends on the maximal nuclear charge
Z = maxj=1,...,M Zj . For the case of one nucleus, M = 1 of charge Z, the resulting bound
εT F (ρ) ≥ −CZ 7/3 is, as we will see, optimal.
Proof. We write V (x) = M −1
P
j=1 Zj |x − Rj | as V (x) = V< (x) + V> (x), with
M  
X 1 1
V> (x) = Zj min ,
j=1
|x − Rj | R

for a fixed R > 0. Observe that min(|x − Rj |−1 , R−1 ) is the potential generated by a uniform
charge distribution on the sphere |x − Rj | = R. In other words,
Z
1
V> (x) = dµ(y)
|x − y|
where µ is the Borel measure
M
X Zk
µ(x) = 2
δ(|x − Rk | − R)
k=1
4πR

This implies that Z

V> (x)ρ(x)dx = 2D(ρ, µ)

Therefore
Z Z
3 5/3
εT F (ρ) = γ ρ (x)dx − V< (x)ρ(x)dx − 2D(ρ, µ) + D(ρ, ρ) + U
5
Z Z (4.86)
3 5/3
≥ γ ρ (x)dx − V< (x)ρ(x)dx − D(µ, µ) + U
5

86
because, by the positivity of the Coulomb potential D(ρ − µ, ρ − µ) ≥ 0. Next observe that
Z M Z  
X 1 1
V< (x)ρ(x)dx = Zj χ(|x − Rj | < R) − ρ(x)dx
j=1
|x − Rj | R
M (4.87)
χ(|x − Rj | < R)
X Z
≤ Zj ρ(x)dx
j=1
|x − Rj |
≤ Ztot R1/5 ∥ρ∥5/3
Hence
Z  
3 5/3 3 5/3 5/2
γ∥ρ∥5/3 − V< (x)ρ(x)dx ≥ inf γx − Ztot R1/5 x = −CZtot R1/2 (4.88)
5 x≥0 5
for some C > 0 depending only on γ > 0. On the other hand, a simple computation shows
that
Z Z
1 dµ(x)dµ(y) 1
D(µ, µ) = = dµ(x)V> (x)
2 |x − y| 2
Z  
1X 1 1 1
= Zk Zℓ dxδ(|x − Rk | − R) min ,
2 k,ℓ 4πR2 |x − Rℓ | R
1 X Zk2
Z
1X 1 1
≤ Zk Zℓ dxδ(|x| − R) +
2 k̸=ℓ 4πR2 |x + Rk − Rℓ | 2 k R
X Zk Zℓ 1 X 2
≤ + Z
k<ℓ
|Rk − Rℓ | 2R k k

which means that

M
1 X 2
−D(µ, µ) + U ≥ − Z (4.89)
2R k=1 k
Combining the bounds (4.86), (4.88) and (4.89), we find
M
5/2 CX 2
εT F (ρ) ≥ −CZtot R1/2 − Z
R j=1 j

Minimizing over R > 0, we conclude that

M
!1/3
5/3
X
εT F (ρ) ≥ −CZtot Zj2
j=1

87
 The second important observation to conclude the existence of a (unique) minimizer is the
convexity of εT F .

Lemma 4.1.2. The domain DN is convex and εT F (ρ) is a strictly convex functional on DN ,
that is εT F (λρ1 + (1 − λ)ρ2 ) < λεT F (ρ1 ) + (1 − λ)εT F (ρ2 ), for all λ ∈ (0, 1), ρ1 ̸= ρ2 .

Proof. For ρ1 , ρ2 ∈ DN , and λ ∈ [0, 1], we clearly have λρ1 + (1 − λ)ρ2 ∈ L1 ∩ L5/3 and
λρ1 + (1 − λ)ρ2 ≥ 0. Moreover, by the linearity of the integral,
Z
dx (λρ1 (x) + (1 − λ)ρ2 (x)) = N

Hence λρ1 + (1 − λ)ρ2 ∈ DN , and thus DN is convex. To prove the convexity

R of εT F (ρ) observe
that theR map s → s5/3 is convex, and therefore the kinetic energy ρ(x)5/3 dx is convex. The
term − V (x)ρ(x)dx is linear and hence convex. The self-energy D(ρ, ρ) is convex, because

D(λρ1 + (1 − λ)ρ2 , λρ1 + (1 − λ)ρ2 ) = λ2 D(ρ1 , ρ1 ) + (1 − λ)2 D(ρ2 , ρ2 ) + 2λ(1 − λ)D(ρ1 , ρ2 )

≤ λD(ρ1 , ρ1 ) + (1 − λ)D(ρ2 , ρ2 )

where we used the estimate

1
|D(ρ1 , ρ2 )| ≤ D(ρ1 , ρ1 )1/2 D(ρ2 , ρ2 )1/2 ≤
(D(ρ1 , ρ1 ) + D(ρ2 , ρ2 ))
2
R
Strict convexity follows from the strict convexity of ρ → ρ5/3 (x)dx.
Combining the last two results, we are now ready to show the existence of a unique
minimizer for εT F (ρ) over the domain DN . For an arbitrary N > 0, we define

E(N ) := inf{εT F (ρ) : ρ ∈ DN }

(physically, only N ∈ N are interesting, mathematically, the restriction to integer values of N

is not necessary).

Theorem 4.1.3. For any N > 0 there exists a unique ρ0 ∈ DN such that

εT F (ρ0 ) = E(N ) .

Proof. From (4.86), (4.87), (4.89), we conclude that

Z M
! Z
3 1 X
εT F (ρ) ≥ γ ρ5/3 (x)dx − Ztot R1/5 ∥ρ∥5/3 − Zk2 ≥ c1 ρ5/3 (x)dx − C2
5 2R j=1

88
for constants C1 > 0 (depending only on γ) and C2 > 0 (depending on the nuclear charges).
In particular, Z
ρ5/3 (x)dx ≤ C
e1 εT F (ρ) + C
e2 (4.90)

for appropriate constants C e2 . Suppose now that ρj is a minimizing sequence in DN , i.e.

e1 , C
ρj ∈ DN for all j ∈ N, with εT F (ρj ) → E(N ) as j → ∞. From (4.90), we find
Z
5/3
ρj (x)dx ≤ C for all j ∈ N

for a constant C depending on γ and on the nuclear charge, but independent of N and of j.
Hence ρj is a bounded sequence in L5/3 (R3 ); therefore there exists ρ0 ∈ L5/3 (R3 ) such that
(after passing to a subsequence, still denoted ρj ) ρj ⇀ ρ0 weakly in L5/3 (R3 ). By the weakly
lower semicontinuity of the norm,
Z Z
5/3 5/3
lim inf ρj (x)dx ≥ ρ0 (x)dx (4.91)
j→∞

Since Z Z
dxρ0 (x)f (x) = lim dxρj (x)f (x) ≥ 0
j→∞

5/2 3
R
R any positive f ∈ L (R ), we conclude that
for ρ0 ≥ 0. Moreover, dxρ0 (x) ≤ N . In fact, if
dxρ0 (x) > N , there would be a set A ∈ R3 with finite measure such that
Z
dxρ0 (x)χA (x) > N

This would contradict the fact that

Z Z
dxρ0 (x)χA (x) = lim dxρj (x)χA (x) ≤ N
j→∞

Summarizing ρ0 ∈ DN .
Next, we want to show that εT F (ρ0 ) = limj→∞ εT F (ρj ) = E(N ), which will imply that ρ0
is indeed a minimizer. Note that it is enough to show that

εT F (ρ0 ) ≤ lim εT F (ρj ). (4.92)

j→∞

To prove (4.92), we observe that

Z
1 ρj (x)ρj (y) 7/5 3/5
D(ρj , ρj ) = dxdy ≤ C∥ρj ∥26/5 ≤ C∥ρj ∥1 ∥ρj ∥5/3
2 |x − y|

89
where we used the Hardy-Littlewood-Sobolev inequality and simple interpolation. Hence,
D(ρj , ρj ) is bounded uniformly in j (by an N dependent constant). Moreover (after passing to
a subsequence, still denoted ρj ), we have D(ρj , ρ) → D(ρ0 , ρ) as j → ∞, for any ρ ∈ L5/3 ∩ L1 .
In fact, Z
D(ρj , ρ) = f (x)ρj (x)

with
Z Z Z
ρ(y) ρ(y) ρ(y)
f (x) = dy = dy + dy =: f1 (x) + f2 (x)
|x − y| |x−y|≤1 |x − y| |x−y|>1 |x − y|

Note that ∥f1 ∥∞ ≤ C∥ρ∥5/3 and ∥f1 ∥1 ≤ C∥ρ∥1 . Hence f1 ∈ L1 ∩ L∞ . By interpolation, we

conclude that f1 ∈ Lq , for all q ∈ [1, ∞]. On the other hand, Young’s inequality implies that
f2 ∈ Lq for all q > 3. Hence f ∈ Lq for all q > 3. Fix q ′ ∈ (1, 3/2). Since ρj ∈ L1 ∩ L5/3
′
with uniformly bounded norms, it follows that ρj is uniformly bounded in Lq . Hence, ρj
′
has a weakly convergent subsequence in Lq ; the weak limit must be again ρ0 . Denoting the
′
subsequence again with ρj , we conclude that ρj ⇀ ρ0 in Lq . Since f ∈ Lq , where q > 3 is
such that 1/q + 1/q ′ = 1, this implies that
Z Z
D(ρj , ρ) = f (x)ρj (x) → f (x)ρ0 (x) = D(ρ0 , ρ)

as j → ∞. Since ρ0 ∈ L1 ∩ L5/3 , we find

D(ρ0 , ρ0 ) = lim D(ρj , ρ0 ) ≤ lim inf D(ρj , ρj )1/2 D(ρ0 , ρ0 )1/2

j→∞ j→∞

and hence
D(ρ0 , ρ0 ) ≤ lim inf D(ρj , ρj ) (4.93)
j→∞

We also claim that (along the same subsequence as above)

Z Z
V (x)ρ0 (x)dx = lim V (x)ρj (x) (4.94)
j→∞

In fact, writing
M M
X χ(|x − Rj | ≥ 1) X χ(|x − Rj | < 1)
V (x) = Zj + Zj =: V> (x) + V< (x)
j=1
|x − Rj | j=1
|x − Rj |

we note that V> ∈ Lq for all q > 3 and V< ∈ L5/2 . Hence
Z Z
V< (x)ρj (x)dx → V< (x)ρ0 (x)dx

90
because ρj ⇀ ρ0 in L5/3 , while
Z Z
V> (x)ρj (x)dx → V> (x)ρ0 (x)dx

′
because ρj ⇀ ρ0 in Lq for a 1 < q ′ < 3/2. Eqs. (4.91), (4.93), (4.94) imply (4.92). Hence ρ0
is a minimizer. The uniqueness of the minimizer follows from the strict convexity. Suppose
in fact that ρ1 , ρ2 ∈ DN are two minimizers, with εT F (ρ1 ) = εT F (ρ2 ) = E(N ). Then, for
λ ∈ (0, 1), Lemma 4.1.2 implies that λρ1 + (1 − λ)ρ2 ∈ DN and

εT F (λρ1 + (1 − λ)ρ2 ) < λεT F (ρ1 ) + (1 − λ)εT F (ρ2 ) = E(N )

in contradiction to the minimality of E(N ).

Next, we would like to understand under which conditions a minimizer for εT F exists in
the (physically most interesting) set FN . To reach this goal, the first step is to study the
properties of the infimum E(N ) = inf ρ∈DN εT F (ρ), defined for all N > 0.
Lemma 4.1.4. The function E(N ) is convex, non-decreasing and bounded below.
Proof. Convexity follows from the convexity of the functional εT F . In fact, for fixed N1 , N2 > 0
and λ ∈ [0, 1], let ρ1 , ρ2 be the minimizers of εT F on DN1 and, respectively, DN2 . Then, by
convexity of εT F ,

E(λN1 +(1−λ)N2 ) ≤ εT F (λρ1 +(1−λ)ρ2 ) ≤ λεT F (ρ1 )+(1−λ)εT F (ρ2 ) = λE(N1 )+(1−λ)E(N2 ) .

The fact that E(N ) is bounded below follows from Theorem 4.1.1. The fact that E(N ) is
non-increasing follows because the set DN , over which we minimize, becomes larger as N
increases.
From the last lemma, it follows that E∞ = limN →∞ E(N ) exists. We can also define the
critical number of particles Nc by

Nc := inf{N : E(N ) = E∞ }

By monotonicity of E(N ), we have E(N ) = E∞ for all N ≥ Nc if Nc < ∞. At this point it

is still possible that Nc = ∞ (we will see shortly that this is not the case). It turns out that
a minimizer for εT F over FN exists if and only if N ≤ Nc . This is the content of the next
theorem.
Theorem 4.1.5. For N ≤ Nc , there R exists a unique minimizer ρN of εT F (ρ) on the set
1 5/3 3
FN = {ρ ∈ L ∩ L (R ) : ρ ≥ 0, ρ = N }. Moreover, on [0, Nc ], the function E(N ) is
strictly convex and strictly decreasing. If Nc < ∞, and N > Nc , there is no minimizer of
εT F on FN . In this case, ρNc is the unique minimizer of εT F on DN . Moreover, the function
E(N ) is constant on [Nc , ∞).

91
 R
Proof. PickRN ≤ Nc , and let ρ0 be the minimizer of εT F on DN . We claim that ρ0 (x)dx = N .
In fact, if ρ0 (x)dx = N ′ < N , then we would conclude that ρ0 is also the minimizer
of εT F over DN ′ , which would imply that E(N ′ ) = E(N ). Since E is non-increasing and
convex, it must be constant in the interval [N ′ , N ]. Again by convexity, this is only possible
if E(N ′ ) = E(N ) = E∞ and thus N ′ ≥ Nc , which contradicts the assumption N ′ < N . The
same argument shows the strict convexity of E on [0, Nc ]. Suppose now that Nc < ∞ and
N > Nc , thenR we claim that there is no minimizer of εT F on FN . In fact, if a minimizer ρN
existed, with ρN = N , we could use (ρNc + ρN )/2 as a trial function, to conclude, using the
strict convexity of εT F , that
   
Nc + N ρNc + ρN
E(Nc ) = E ≤ εT F
2 2
εT F (ρNc ) + εT F (ρN ) E(Nc ) + E(N )
< = = E(Nc ) (4.95)
2 2
R
which gives a contradiction. Here we noted that, since ρN = N > Nc , ρN ̸= ρNc and we used
the fact that E(N ) = E∞ for all N ≥ Nc (as a convex function E(N ) is certainly continuous).
Hence there is no minimizer of εT F on FN , for N > Nc , if Nc < ∞.
Next, we want to establish the value of Nc . To this end, we characterize minimizers as
solutions of the Thomas-Fermi differential equation.
Theorem 4.1.6. Assume that N ≤ Nc . Then there exists a constant µ = µ(N ) ≥ 0 so that
the unique minimizer ρN satisfies the equation
 
2/3 1
γρN (x) = V (x) − ∗ ρN − µ .
|.| +

For N = Nc , we have µ(Nc ) = 0.

Proof. Let ρN be the minimizer of εT F (ρ) over FN (since N ≤ Nc , ρN is also the minimizer
of |epsT F (ρ) over DN ). For any fixed δ > 0, for any bounded function f with
Z
χ(ρN (x) ≥ δ)f dx = 0 (4.96)

and for sufficiently small ε (ε can be positive or negative, it is small in absolute value), consider

ρε = ρN + εf χ(ρN ≥ δ)

Since ρε ∈ FN , we have εT F (ρε ) ≥ εT F (ρN ) and therefore

d
εT F (ρε )|ε=0 = 0
dε
92
which implies that
Z    
2/3 1
γρ (x) − V (x) + ∗ ρN (x) f (x)dx = 0
ρN (x)≥δ |.|
for all f satisfying (4.96). We conclude that there exists µ ∈ R with
 
2/3 1
γρ (x) − V (x) + ∗ ρN (x) = −µ (4.97)
|.|
for all x ∈ R3 with ρN (x) ≥ δ. Letting δ → 0, we conclude that (4.97) must hold for all x ∈ R3
with ρN (x) ̸= 0. To explore the region where
R ρN (x) = 0, we consider ρε (x) = ρN (x) + εf (x)
for ε > 0, f ≥ 0 on {x : ρN (x) = 0} and f = 0. Variation w.r.t. ε > 0 gives
Z    
2/3 1
γρN (x) − V (x) + ∗ ρN (x) f (x)dx ≥ 0
|.|
and thus Z Z    
1
−µ f (x) + −V (x) + ∗ ρN (x) f (x) ≥ 0 .
ρN (x)>0 ρN (x)=0 |.|
R
Since f (x)dx = 0, we arrive at
Z    
1
−V (x) + ∗ ρN (x) + µ f (x) ≥ 0
ρN (x)=0 |.|
and therefore, we conclude that
 
1
V (x) − ∗ ρN (x) − µ ≤ 0 on {x : ρN (x) = 0} . (4.98)
|.|
Eqs. (4.97) and (4.98) can be combined into
 
2/3 1
γρN (x) = V (x) − ∗ ρN − µ .
|.| +
R
In the case N = Nc , we do not need the constraint ρN (x)≥δ f (x) = 0 because, in that case,
R
ρN is a minimizer among all ρ with ρ ≥ Nc . Hence µ = 0 in this case. In general, in order
to make sure that ρN (x) → 0 as |x| → ∞ (since V (x) → 0 as |x| → ∞), one needs µ ≥ 0.
We obtain the following corollary.
Corollary 4.1.7. Fix N < ∞, with N ≤ Nc . Then the potential
1
ϕ(x) = V (x) − ∗ ρN (x)
|.|
is non-negative.

93
Proof. By definition, we have ∆ϕ = ρN away from the nuclei (because ∆|x−Rj |−1 = δ(x−Rj )).
From Theorem 4.1.6, we conclude that
3/2
∆ϕ = γ −3/2 [ϕ(x) − µ]+ (4.99)

away from the nuclei. It is simple to check that ϕ is continuous, away from the nuclei (V (x)
is clearly continuous, away from the nuclei; the continuity of |.|−1 ∗ ρN follows because ρN ∈
L5/3 ∩ L1 ).
Let A := {x : ϕ(x) < 0}. Note that a neighborhood of the nuclei is contained in the
complement of A. By (4.99), ϕ is harmonic on A (recall that µ ≥ 0). Moreover, since
ϕ(x) → 0 as |x| → ∞, we have that ϕ = 0 on the boundary of A. Harmonicity implies that
ϕ = 0 on A, and therefore that A is empty.
Using the last corollary, we can show now that Nc = Ztot .
Theorem 4.1.8. We have Nc = Ztot = M
P
j=1 Zj .

Proof. The main ingredient of this proof is the observation that

Z  
1 1 1 1
dω = min , . (4.100)
4π |x|=r |rω − y| r |y|

This follows easily using spherical coordinates.

Assume now Nc > Z, and choose Z < N < Nc . Recall
M  
X 1 1
ϕ(x) = Zj − ∗ ρN (x) ≥ 0
j=1
|x − Rj | |.|

Averaging over x in a sphere of radius r, we find from (4.100), choosing r sufficiently large
(r > |Rj |, for every j), Z  
Z 1 1
− dy min , ρN (y) ≥ 0
r r |y|
Hence Z
Z 1
− dy ρN (y) ≥ 0
r r |y|<r
R
and Z ≥ |y|<r ρN (y). Since this is true for every r large enough, we conclude that Z ≥
R
dyρN (y) = N (since N < Nc ). This contradicts the assumption. R
Assume, on the other hand, that Nc < Z. Let ρc be the minimizer of εT F with ρc = Nc .
Recall that ρc satisfies the equation
 
2/3 1
γρc (x) = V (x) − ∗ ρc (x)
|.|

94
where the r.h.s. is exactly the potential ϕ(x). Averaging over the sphere |x| = r, we find
Z Z
γ 2/3 1
dωρc (rω) = dω ϕ(rω) .
4π S 2 4π S 2
Since, on the one hand,
Z  Z 2/3
γ 1
dωρ2/3
c (rω) ≤γ ρc (rω)
4π S2 4π S2

and, on the other hand,

  R
Z − ρc (y)dy
Z Z
1 Z 1 1
dω ϕ(rω) = − dy min , ρc (y) ≥
4π S 2 r r |y| r

we conclude that  2/3

Z − Nc
Z
1
ρc (rω) ≥ .
4π S2 rγ
Therefore, we find Z
1
ρc (rω) ≥ Cr−3/2
4π S2
and Z Z ∞ Z
2
Nc = ρc (x)dx = drr dω ρc (rω) = ∞
0 S2
which is a contradiction.
Summarizing: E(N ) = inf{εT F (ρ) : ρ ∈ DN } is monotone decreasing and convex. It is
strictly decreasing for N ≤ Z. It is constant for N ≥ Z. ARunique minimizer for εT F exists on
DN , for everyR N . For N ≤ Z, the minimizer is such that ρ = N , for N > Z, the minimizer
is such that ρ = Z. For N ≤ Z, the minimizer is therefore also a minimizer of εT F on FN .
For N ≥ Z, on the other hand, there is no minimizer of εT F on FN . In Thomas-Fermi Theory,
a configuration of nuclei with total charge Z binds exactly Z electrons.
Virial identities give more precise information about E(N ) and the minimizers ρN .

Theorem 4.1.9. Assume N ≤ Z, and let ρN be the unique minimizer of εT F on DN (or FN ).

Then Z Z Z
5/3
γ ρN (x)dx − V (x)ρN (x)dx + D(ρN , ρN ) = −µ ρN (x)dx

where µ = µ(N ) is defined in Theorem 4.1.6. For a single atom (M = 1), we find also
Z Z
6 5/3
γ ρN (x)dx − V (x)ρN (x)dx + D(ρN , ρN ) = 0
5

95
In the neutral case N = Z (again assuming M = 1), we find
Z Z
3 5/3 3
E(Z) = − γ ρZ (x)dx = − V (x)ρZ (x)dx = −3D(ρZ , ρZ )
5 7
Proof. From the Thomas-Fermi equation
   
2/3 1
γρN = V (x) − ∗ ρN (x) − µ
|.| +

we conclude that    
5/3 1
γρN (x) = ρN (x) V (x) − ∗ ρN (x) − µ
|.|
Therefore Z Z Z
5/3
γ ρN (x)dx − ρN (x)V (x) + 2D(ρN , ρN ) = −µ ρN (x)dx
R
To show the second identity, let fs (x) = s3 ρN (sx). Then fs (x)dx = N for all s > 0 and, for
M = 1,
Z Z
3 5/3
εT F (fs ) = γ fs (x)dx − fs (x)V (x) + D(fs , fs )
5
3s2
Z Z
5/3
= γ ρN (x)dx − s ρN (x)V (x)dx + sD(ρN , ρN )
5
Therefore
Z Z
d 6 5/3
0 = εT F (fs )|s=1 = γ ρN (x)dx − ρN (x)V (x)dx + D(ρN , ρN )
ds 5

The third identity follows combining the first two (using the fact that, if N = Z, µ = 0).
For one neutral atom (M = 1, N = Z) it is also possible to obtain the following charac-
terization of the energy E(Z) and of the corresponding minimizer.

Theorem 4.1.10. If M = 1, we have ρZ (x) = (Z 2 /γ 3 )ρ̄(Z 1/3 x/γ), where ρ̄ is the minimizer
of the problem
 Z Z Z 
3 5/3 1
e = inf ρ (x)dx − ρ(x)dx + D(ρ, ρ) : ρ(x)dx = 1 .
5 |x|

Moreover E(Z) = (e/γ)Z 7/3 . Numerically, e ≃ −3.678

96
Proof. Defining ρ̄ by the equation ρZ (x) = aℓ3 ρ̄(ℓx) for some ℓ, a > 0, we find
Z Z
5/3 2 3 5/3 1
εT F (ρZ ) = a ℓ γ ρ̄ (x)dx − Zaℓ ρ̄(x)dx + a2 ℓD(ρ̄, ρ̄)
5 |x|

Setting a5/3 ℓ2 γ = Zaℓ = a2 ℓ gives a = Z, ℓ = Z 1/3 /γ. With this choice, we find

Z 7/3 3
 Z Z 
5/3 ρ̄(x)
εT F (ρZ ) = ρ̄ (x)dx − dx + D(ρ̄, ρ̄)
γ 5 |x|

and thus we conclude that ρ̄ must minimize the functional on the r.h.s. of the last equation.

4.2 The No-Binding Theorem

An important fact about Thomas-Fermi theory, which is also one of its most important limi-
tations, is the complete absence of binding; to minimize the energy of a system with 2 nuclei,
they have to be shifted infinitely apart from each others. According to Thomas-Fermi theory
there would be no stable molecule in nature.
The statement of the no-binding theorem is even stronger. If we consider a system of
M nuclei and we divide them into two groups A, B, the sum of the energies of A and B
(minimized over all possible allocations of the electrons into the two groups) is always smaller
or equal to the energy of the full system. This statement is true for any fixed location of the
nuclei.
Next, we state the no-binding theorem more precisely. Consider a system with M nuclei
of charges Z1 , . . . , ZM located at R1 , . . . , RM ∈ R3 . The Thomas-Fermi energy of this system
is given by Z Z
3 5/3
ε(ρ) = γ ρ (x)dx − V (x)ρ(x) + D(ρ, ρ) + U
5
where
M
X Zk
V (x) =
k=1
|x − Rk |
and X Zk Zℓ
U= .
k<ℓ
|Rk − Rℓ |

As before, let E(N ) denote the

R ground state energy, obtained by minimizing ε(ρ) over all
1 5/3
ρ ∈ L ∩ L with ρ ≥ 0 and ρ(x)dx = N .
We divide the nuclei into two groups A and B. Correspondingly, we define the measures
X X
µA (x) = Zk δ(x − Rk ), µB (x) = Zk δ(x − Rk )
k∈A k∈B

97
Hence, the potential generated by the nuclei in group A is given by
X Zk Z
dµA (y)
VA (x) = =
k∈A
|x − Rk | |y − x|

and, similarly, the potential generated by the nuclei in group B is

X Zk Z
dµB (y)
VB (x) = = .
k∈B
|x − Rk| |y − x|

The Thomas-Fermi energies of the subsystems A and B are given by

Z Z
3 5/3
X Zk Zℓ
εA (ρ) = γ ρ (x)dx − VA (x)ρ(x) + D(ρ, ρ) +
5 k<ℓ: k,ℓ∈A
|Rk − Rℓ |

and, respectively,
Z Z
3 5/3
X Zk Zℓ
εB (ρ) = γ ρ (x)dx − VB (x)ρ(x) + D(ρ, ρ) +
5 k<ℓ: k,ℓ∈B
|Rk − Rℓ |

We define the ground state energy

Z
5/3 1
EA (N ) := inf {εA (ρ) : ρ ∈ L ∩ L , ρ ≥ 0, ρ(x)dx = N }

and similarly EB (N ).
PM
Theorem 4.2.1. Assume that N ≤ Ztot = j=1 Zj . Then

inf{EA (N1 ) + EB (N2 ) : N1 + N2 = N } ≤ E(N ) .

The no-binding theorem was first predicted by Teller in 1962; the first mathematically
rigorous proof was obtained by Lieb-Simon in 1977. The crucial ingredient in the proof of the
no-binding theorem is the following lemma, due to Baxter.
Lemma 4.2.2. Assume that ρ ∈ L1 ∩ Lp , for some p > 3/2, and ρ ≥ 0. Then there exists g,
with 0 ≤ g ≤ ρ so that Z Z
g(y) 1
dy ≤ dµA (y)
|x − y| |x − y|
for every x ∈ R3 . Moreover,
Z Z
g(y) 1
dy = dµA (y)
|x − y| |x − y|
on {x ∈ R3 : g(x) < ρ(x)}.

98
Proof of Lemma. Consider the problem of minimizing the functional
Z
g(x)µA (y)
I(g) := D(g, g) − dxdy
|x − y|

under the condition 0 ≤ g ≤ ρ. Note that

Z Z
g(x)µA (y) X g(x)
dxdy = Zk dx
|x − y| k∈A
|x − Rk |
!
X Z ρ(x) X

≤ Zk dx ≤ Zk ∥ρ∥1 + C∥ρ∥5/3 < ∞
k∈A
|x − R k|
k∈A

Hence, the functional I(g) is bounded below. We claim now that a minimizer for I(g) exists.
In fact, let gj be a minimizing sequence. Since 0 ≤ gj ≤ ρ, gj is uniformly bounded in Lp , for
some p > 3/2. Thus, there exists a subsequence (again denoted by gj ) which converges weakly
to g ∈ Lp . Let now ξ be a smooth function of compact support, equal one in a neighborhood of
′
the origin. The function ξ(x−y)/|x−y| is then in Lp , where p′ < 3 is such that 1/p+1/p′ = 1,
and therefore
ξ(x − y) (gj (x) − g(x))
Z
dx →0
|x − y|
for every fixed y. On the other hand, (1 − ξ(x))/|x| is in Lr , for every r > 3. Let r′ < 3/2 be
′
such that 1/r + 1/r′ = 1. Since ρ ∈ L1 ∩ Lp , gj is uniformly bounded in Lr . Hence, passing
′
again to a subsequence, gj → g weakly in Lr and

(1 − ξ(x − y)) (gj (x) − g(x))

Z
dx →0
|x − y|

for every fixed y ∈ R3 . We conclude that

(gj (x) − g(x))

Z
dx →0
|x − y|

for every fixed y ∈ R3 , and therefore that

Z Z
gj (x)µA (y) g(x)µA (y)
dxdy → dxdy
|x − y| |x − y|

as j → ∞. Similarly as in Theorem 4.1.3 (see, in particular, (4.93)), one can also show that

lim inf D(gj , gj ) ≥ D(g, g)

j→∞

99
Hence I(g) ≤ lim inf j→∞ I(gj ) = inf{I(f ) : 0 ≤ f ≤ ρ}. From the weak convergence of gj to
g, we also conclude that 0 ≤ g ≤ ρ. Therefore, a minimizer exists. Simple variations shows
that
1 1
∗g = ∗ µA on {x : 0 < g(x) < ρ(x)}
|.| |.|
that
1 1
∗g ≤ ∗ µA on {x : g(x) = ρ(x)}
|.| |.|
and that
1 1
∗g ≥ ∗ µA on {x : g(x) = 0} .
|.| |.|
Let ( )
1 1 X Zk
P := x: ∗ g(x) > ∗ µA (x) = .
|.| |.| k∈A
|Rk − x|

P is an open set, because |.|−1 ∗ g is continuous. The function |.|−1 ∗ (g − µA ) vanishes on the
boundary of P . Since, Rk ̸∈ P for all k ∈ A, and since P is a subset of {x : g(x) = 0}, we
conclude that  
1
∆ ∗ (g − µA ) = −(g − µA )
|.|
vanishes on P , and therefore that the function |.|−1 ∗ (g − µA ) is harmonic on P . Since it
vanishes on the boundary of P , we conclude that it vanishes everywhere in P . Hence P is
empty.
ProofR of no-binding Theorem. Since N ≤ Z, there is a minimizer ρ ∈ L1 ∩ L5/3 , with ρ ≥ 0,
and ρ(x)dx = N . The theorem follows if we can find non-negative functions g, h with
g + h = ρ and with
εA (g) + εB (h) ≤ ε(ρ) = E(N )
Since a5/3 + b5/3 ≤ (a + b)5/3 for any a, b ≥ 0, we conclude that, if g + h = ρ, then
Z Z Z
g (x)dx + h (x)dx ≤ ρ5/3 (x)dx
5/3 5/3

Hence, to prove the theorem, it is enough to show the existence of non-negative g, h, with
g + h = ρ such that
X Zk Zℓ X Zk Zℓ
−2D(µA , g) + D(g, g)+ − 2D(µB , h) + D(h, h) +
k≤ℓ:k,ℓ∈A
|Rk − Rℓ | k≤ℓ:k,ℓ∈B
|Rk − Rℓ |
X Zk Zℓ
≤ −2D(µA + µB , g + h) + D(g + h, g + h) +
k<ℓ
|Rk − Rℓ |

100
This is equivalent to
X Zk Zℓ
0 ≤ −2D(µA , h) − 2D(µB , g) + 2D(g, h) + (4.101)
k∈A,ℓ∈B
|Rk − Rℓ |

Since X Zk Zℓ
= 2D(µA , µB )
k∈A,ℓ∈B
|Rk − Rℓ |

it is enough to find g, h with g + h = ρ such that

D(g − µA , h − µB ) ≥ 0 .

Using the g constructed in Lemma 4.2.2, and taking h = ρ − g, we find

Z Z 
dy
2D(g − µA , h − µB ) = dx (g(y) − µA (y)) (h(x) − µB (x))
|x − y|
Z Z 
dy
= dx (g(y) − µA (y)) (h(x) − µB (x))
g(x)=ρ(x) |x − y|
Z Z 
dy
=− dx (g(y) − µA (y)) µB (x) ≥ 0
h(x)=0 |x − y|

which completes the proof of the theorem.

4.3 Stability of Matter via Thomas-Fermi Theory

As an immediate consequence of the no-binding theorem, we obtain the following corollary.

Corollary 4.3.1. For any non-negative function ρ ∈ L5/3 (R3 ), for any positions R1 , . . . , RM ∈
R3 , and for any charges Z1 , . . . , ZM ≥ 0, we find
Z M Z M
3 5/3
X ρ(x) X Zk Zℓ 3.678 X 7/3
γ ρ (x)dx− Zk dx+D(ρ, ρ)+ ≥− Z . (4.102)
5 k=1
|x − Rk | k<ℓ
|Rk − Rℓ | γ k=1 k

Proof. Denote by I the l.h.s. of (4.102). Then, by the results of Sect. 4.1,
( Z M
)
X
5/3 1
I ≥ inf εZ1 ,...,ZM (ρ) : ρ ∈ L ∩ L , ρ ≥ 0 and ρ(x)dx = Zj
j=1

where εZ1 ,...,ZM denotes the Thomas-Fermi functional with M nuclei of charges Z1 , . . . , ZM > 0
located at R1 , . . . , RM . Separating the nuclei into two groups, A and B, where A consists only

101
of the nucleus of charge Z1 located at R1 , and B is the rest, we conclude from the no-binding
theorem, and from Theorem 4.1.10 that
( M
)
−3.678 7/3
Z X
I≥ Z1 + inf εZ2 ,...,ZM (ρ) : ρ ∈ L5/3 ∩ L1 , ρ ≥ 0 and ρ= Zj
γ j=2

Iterating M times, we obtain the claim.

Using this Corollary, it is possible to give an alternative proof of stability of matter.
The only other ingredient which is needed is the kinetic energy estimate discussed in the
previous semester. Recall that the kinetic energy estimate states that, for any ψ ∈ L2 (R3 ×
{1, . . . , q})⊗N ,
N Z
X C 5/3
⟨ψ, −∆xj ψ⟩ ≥ (1) 2/3 ρψ (x)dx
j=1 ∥e
γψ ∥∞
(1)
where γ
eψ is the spin-summed one-particle reduced density matrix associated with the wave
(1)
function ψ, and ρψ (x) = γ
eψ (x; x) is the corresponding density. In particular, for fermionic
(1)
ψ, we have ∥e γψ ∥∞ ≤ q (q being the number of different spin-states, for example q = 2 for
electrons), and therefore
N Z
X C
⟨ψ, −∆xj ψ⟩ ≥ 2/3 ρ5/3 (x)dx (4.103)
j=1
q
Consider now the Hamilton operator
N M N M
X X 1 X Zi Zj XX Zj
HN,M (Z, R) = −∆xj + + −
j=1 i<j
|xi − xj | i<j |Ri − Rj | i=1 j=1 |xi − Rj |

describing N electrons and M nuclei fixed at positions R1 , . . . , RM with charges Z1 , . . . , ZM .

The expectation ⟨ψ, HN,M (Z, R)ψ⟩ for a fermionic wave function ψ ∈ L2 (R3 × {1, . . . , q})⊗a N
can be bounded below using (4.103) by
Z * +
C 5/3
X 1
⟨ψ, HN,M (Z, R)ψ⟩ ≥ 2/3 ρψ (x)dx + ψ, ψ
q i<j
|xi − xj |
M M Z (4.104)
X Zi Zj X ρψ (x)
+ − Zj dx
i<j
|Ri − R j | j=1
|x − Rj |

Applying the Corollary 4.3.1 (but replacing the variables Rj by the positions of the N electrons,
and the charges Zj by 1) we conclude that, for any ρ ∈ L5/3 ,
N Z N Z
X 1 3 5/3
X ρ(x) 3.678
≥ − ε ρ (x)dx + dx − D(ρ, ρ) − N.
i<j
|x i − x j | 5 k=1
|x − x k | ε

102
In particular, taking ρ = ρψ , we find
* N
+ Z
X 1 3 5/3 3.678
ψ, ψ ≥ − ε ρψ (x)dx + D(ρψ , ρψ ) − N.
i<j
|xi − xj | 5 ε

Plugging into (4.104), we obtain

 Z
C 3 5/3
⟨ψ, HN,M (Z, R)ψ⟩ ≥ 2/3
− ε ρψ (x)dx + D(ρψ , ρψ )
q 5
M
X Zi Zj M Z
X ρψ (x) 3.678
+ − Zj dx − N
i<j
|Ri − Rj | j=1
|x − R j | ε

Using again Corollary 4.3.1, we find

 
M 7/3
3.678 X 7/3 3.678 Zmax M N
⟨ψ, HN,M (Z, R)ψ⟩ ≥ −   Zj − N ≥ −3.678   +
5C
−ε j=1
ε 5C
−ε ε
3q 2/3 3q 2/3

7/3
Optimizing over ε > 0, we conclude that ⟨ψ, HN,M (Z, R)ψ⟩ ≥ −CZmax (M +N ), for a universal
constant C > 0, and thus we obtain stability of the second kind.

4.4 Thomas-Fermi Theory as Limit of Quantum Mechanics

In the section, we are interested in the limit of the energy of quantum mechanical models of
matter as the total nuclear charge, Z, diverges to infinity, and in particular in its relation
with the Thomas Fermi energy. We consider a system with a fixed number, M , of nuclei.
We assume that the total nuclear charge Z and the number of electrons N both tend to
infinity, so that their ratio stays constant. Since the system contains M nuclei, there are
several possibilities to let Z → ∞; we choose here to keep the ratio between the charges of
the different nuclei constant. To this end, we fix nuclear charges Z0 = (Z10 , . . . , ZM
0
), with
0
Zj > 0 for all j = 1, . . . , M , and the number of electrons N0 ∈ N. For every N ∈ N,
we define aN = N/N0 and Z = (Z1 , . . . , ZM ), with Zj = aN Zj0 , for j = 1, . . . , M . The
Thomas-Fermi Theory tells us that also the positions of the M nuclei must be rescaled. For
R = (R1 , . . . , RM ) ∈ R3M , let
TF
EN (Z, R) = inf ET F (ρ)
ρ∈FN
R
where FN = {ρ ∈ L5/3 ∩ L1 (R3 ) : ρ ≥ 0, ρ(x)dx = N } and
Z M Z Z M
3 5/3
X ρ(x) 1 ρ(x)ρ(y) X Zi Zj
ET F (ρ) = γ ρ (x)dx − Zj dx + dxdy +
5 j=1
|x − Rj | 2 |x − y| i<j
|Ri − Rj |

103
with γ = (6π 2 )2/3 (in the previous sections, the value of γ was not fixed; however, to get an
approximation to quantum mechanics, γ = (6π 2 )2/3 must be fixed). Choosing Zj = aN Zj0 and
ρ(x) = Kℓ3 ρ̄(ℓx), we observe that
Z M Z
3 5/3 2 5/3
X
0 ρ̄(x)
ET F (ρ) = K ℓ γ ρ̄ (x)dx − aN Kℓ Zj dx
5 j=1
|x − ℓR j |
M
K 2ℓ Zi0 Zj0
Z
ρ(x)ρ(y) 2
X
+ dxdy + aN ℓ
2 |x − y| i<j
|ℓRi − ℓRj |

1/3
Hence, setting K = aN , ℓ = aN , we have K 5/3 ℓ2 = aN Kℓ = K 2 ℓ = a2N ℓ, and
( Z M Z
7/3 3 5/3
X
0 ρ̄(x)
ET F (ρ) = aN γ ρ̄ (x)dx − Zj 1/3
dx
5 j=1 |x − aN Rj |
M
) (4.105)
Zi0 Zj0
Z
1 ρ(x)ρ(y) X
+ dxdy + 1/3 1/3
2 |x − y| i<j |aN Ri − aN Rj |
R
which has to be minimized over all ρ̄ with the property that ρ̄(x)dx = N0 . We conclude
that in the limit N → ∞, which is equivalent to aN → ∞, the interaction among the nuclei
−1/3
remains of the same order as the rest of the energy if and only if Rj = aN Rj0 for fixed
Rj0 ∈ R3 , j = 1, . . . , M . In the following, we will assume this scaling for the positions of the
M nuclei. It is easy to understand what happens when different scalings are considered; if
the positions of the nuclei are chosen independently of N , in the limit it is as if they were
infinitely apart (of course, for M = 1, the position of the single nucleus does not need to be
rescaled).
Summarizing, for fixed M ∈ N, Z0 = (Z10 , . . . , ZM 0
) and R0 = (R10 , . . . , RM0
), we con-
sider a system of N = aN N0 electrons and M nuclei with charges Z = (Z1 , . . . , ZM ) =
−1/3 −1/3 0
(aN Z10 , . . . , aN ZM
0
) located at R = (R1 , . . . , RM ) = (aN R10 , . . . , aN RM ). According to
(4.105), the Thomas-Fermi energy of this system is given by
TF 7/3
TF
EN (Z, R) = aN EN 0
(Z0 , R0 )
( Z M Z
7/3 3 5/3
X
0 ρ(x)
= aN R inf γ dxρ (x) − Zj dx
ρ=N0 5 j=1
|x − Rj0 |
M
)
ρ(x)ρ(y) X Zj0 Zi0
Z
1
+ dxdy +
2 |x − y| i<j
|Ri0 − Rj0 |

104
The quantum model is described, on the other hand, by the N -electron Hamilton operator
−1/3
HN,M (Z, R) = HaN N0 ,M (aN Z0 , aN R0 )
N N X
M N M
X X Zi X 1 X Zi Zj
=− ∆xj − + +
j=1 j=1 i=1
|xj − Ri | i<j |xi − xj | i<j |Ri − Rj |

The quantum mechanical energy is then given by

Q
EN (Z, R) = inf ⟨ψ, HN,M (Z, R)ψ⟩
∥ψ∥=1,ψ∈L2a (R3N )

The goal of this section is to prove the following theorem, which establishes the relationship
between the quantum mechanical and the Thomas-Fermi energy in the limit of large N (or,
equivalently, of large Z).

Theorem 4.4.1. We have

Q 0 −1/3
EN =aN N0 (aN Z , aN R0 ) TF
lim 7/3
= EN0
(Z0 , R0 )
N →∞ aN

To prove this theorem, we will first show an upper bound for the quantum mechanical
energy, and later we will prove a matching lower bound.
Upper Bound on the Quantum Energy. To produce an upper bound on the quantum
energy, we will use Hartree-Fock (HF) Theory. In HF Theory, one restrict the attention to
wave functions given by Slater determinants. For an orthonormal system {φ1 , . . . , φN } in
L2 (R3 ), we define the Slater determinant
1 X
ψN (x1 , . . . , xN ) = Slater(φ1 , . . . , φN )(x1 , . . . , xN ) = √ σπ φ1 (xπ1 )φ(xπ2 ) . . . φ(xπN )
N ! π∈SN

Note that ψN has automatically the correct symmetry. It is simple to compute the one-particle
reduced density γ (1) associated with Slater(φ1 , . . . , φN ). In fact, its kernel is given by
Z
γ (x; x ) = N dx2 . . . dxN ψN (x, x2 , . . . , xN )ψ N (x′ , x2 , . . . , xN )
(1) ′

N
1 X
′
Y
= σπ σπ φπ1 (x)φπ′ 1 (x )
′ δπj,π′ j
(N − 1)! π,π′ ∈S j=2
N

N
1 X
′
X
= φπ (x)φ(x ) = φj (x)φj (x′ )
(N − 1)! π∈S j=1
N

105
Hence γ (1) = N
P
j=1 |φj ⟩⟨φj | is the orthogonal projection onto the space spanned by the N
orbitals φ1 , . . . , φN . We can also compute the two-particle reduced density
N
1 X Y
γ (2) (x1 , x2 ; x′1 , x′2 ) = σπ σπ′ φπ1 (x1 )φπ2 (x2 )φπ′ 1 (x′1 )φπ′ 2 (x′2 ) δπj,π′ j
2(N − 2)! π,π′ j=3
1 (1)
γ (x1 ; x′1 )γ (1) (x2 ; x′2 ) − γ (1) (x1 ; x′2 )γ (1) (x2 ; x′ 1)


=
2
Observe here that, for Slater determinants, the two-particle (and in fact every other k-particle)
reduced density can be expressed as a function of γ (1) . Using these formulas for γ (1) and γ (2) ,
we can now compute the energy of a Slater determinant. We consider in the following a
general Hamiltonian with two-body interaction, having the form
N
X N
X
H= hj + v(xi − xj )
j=1 i<j

where hj = 1⊗· · ·⊗h⊗. . . 1 acts only on the j-th particle. In our application (up to a constant),
h = −∆ − M −1
and v(x) = |x|−1 . The expectation of the Hamiltonian H in the
P
j=1 Zj |x − Rj |
Slater determinant state ψN is then given by
Z
(1) 1
dxdyv(x − y) γ (1) (x; x)γ (1) (y, y) − |γ (1) (x; y)|2 =: F (γ (1) )


⟨ψN , HψN ⟩ = tr hγ +
2
We define the Hartree-Fock energy as the infimum of ⟨ψN , HψN ⟩ over all Slater determinants
ψN . Since every rank N orthogonal projection γ (1) over L2 (R3 ) is the one-particle density of
a Slater determinant, we have

EHF (N ) = inf ⟨ψN , HψN ⟩ = inf F (γ (1) )

Slater dets ψN rank N orthog. proj. γ (1)

Let E(N ) denote the ground state energy of H, that is,

E(N ) = inf ⟨ψN , HψN ⟩

ψ∈L2a (R3N ), ∥ψN ∥=1

By definition, we clearly have E(N ) ≤ EHF (N ). So, for every rank N orthogonal projec-
tion γ (1) over L2 (R3 ), F (γ (1) ) gives an upper bound to the quantum energy E(N ) (and to
the Hartree-Fock energy EHF (N )). It turns out that, as an upper bound for EHF (N ) (and
therefore also for E(N )), we can even take F (K), for certain K not being a rank N projection
over L2 (R3 ). We say that a non-negative, trace class operator K over L2 (R3 ) is admissible if
0 ≤ K ≤ 1 and tr K = N . We know from last semester that for every admissible K there
exists an N particle state whose one-particle marginal equals K. In general, however, this

106
N -particle state will not be a Slater determinant, and therefore its energy will not be given by
F (K). Nevertheless, under the assumption that v ≥ 0, it turns out that we can always find
a Slater determinant whose energy is smaller than F (K). This is the statement of the next
theorem. It implies that, for every admissible K, F (K) gives an upper bound to EHF (N ) and
therefore also to E(N ).

Theorem 4.4.2. Assume that v ≥ 0. Then, for every admissible non-negative trace class
operator K, there exists a slater determinant ψN such that
Z
1
dxdy v(x − y) K(x; x)K(y; y) − |K(x; y)|2


⟨ψN , HψN ⟩ ≤ F (K) = tr hK +
2

Hence E(N ) ≤ EHF (N ) ≤ F (K) for every admissible K.

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

Lemma 4.4.3. Let c1 ≥ c2 ≥ . . . be an infinite sequence with 0 ≤ cj ≤ 1 and ∞

P
cj = N ∈
1 N 2
PN j=1j 2
N. Then there exist N orthonormal vectors V , . . . , V ∈ ℓ (C) such that j=1 |Vi | = ci for
all i ∈ N.
1/2
Proof of the Lemma. We use induction over N . If N = 1, we simply take Vi = ci . Let
us now assume the lemma holds for N = n − 1. We show it for N = n. First, we assume
cPj = 0 for all j ≥ 2n. Let dj = 1 − cj for 1 ≤ j < 2n and dj = 0 for j ≥ 2n. Then
∞
i=1 dj = (2n − 1) − n = n − 1 and 0 ≤ dj ≤ 1 for all j.PHence, from the induction
n−1
assumption, we find orthonormal W 1 , . . . , W n−1 ∈ ℓ2 (C) with j=1 |Wij |2 = di for all i. In
particular, this means that Wij = 0 for all i ≥ 2n. Hence W 1 , . . . , W n−1 can be thought of
as an orthonormal set in C2n−1 . We can find therefore orthonormal V 1 , . . . , V n ∈ C2n−1 so
that {W 1 , . . . , W n−1 , V 1 , . . . , V n } is an orthonormal basis of C2n−1 . From the completeness
of these vectors, we conclude that
n−1
X n
X
|Wij |2 + |Vij |2 = 1
j=1 j=1

for every i = 1, . . . , 2n − 1. This implies that nj=1 |Vij |2 = 1 − di = ci for all i = 1, . . . , 2n − 1.

P

Defining Vij = 0 for all i ≥ 2n, and for all j = 1, . . . , n, we find the desired vectors V 1 , . . . , V n ∈
ℓ2 (C). Next, we show the statement of the lemma for N = n, assuming that ci = 0 for all
j ≥ J, for an arbitrary J ≥ 0. We proceed by induction, starting from J = 2n. We assume
therefore that the statement of the lemma (with N = n) holds true if cj = 0 for all j ≥ J,
for some J ≥ 2n, and we prove it is true if we only assume that cj = 0 for all j ≥ J + 1.
Let bj = cj for all j < J − 1, bJ−1 = cJ−1 + cJ , bj = 0, for all j ≥ J. Note that, since
the ci ’s are ordered, cJ−1 + cJ ≤ 1 for all J ≥ 2n. The numbers bj may not be ordered,

107
but this
Pn will not play an important role. Let W 1 , . . . , W n be orthonormal vectors in ℓ2 (C),
with j=1 |Wi | = bi for all i. Then, we define vectors V 1 , . . . , V n by setting Vij = Wij for all
j 2

j j
i < J −1, VJ−1 = (cJ−1 /bJ−1 )1/2 WJ−1 , VJj = (cJ /bJ−1 )1/2 WJ−1
j
, and Vij = 0 for all i ≥ J +1. It
is then simple to check that the vectors V 1 , . . . , V n are orthonormal and that nj=1 |Vij |2 = ci
P
for all i. Finally, we have to consider the case cj > 0 for all j ∈ N. To this end, we choose
L ∈ N such that
X∞
bL := cj ≤ 1
j=L

Applying the statement of the lemma to the sequence P (c1 , . . . , cL−1 , bL , 0, 0, . . . ), we find or-
thonormal vectors W 1 , . . . , W n ∈ ℓ2 (C) such that nj=1 |Wij |2 = bi for all i ∈ N. Set now
Vij = Wij , for all i < L and Vij = (ci /bL )1/2 WLjPfor i ≥ L. It is then simple to check that the
vectors V 1 , . . . , V n are orthonormal, and that nj=1 |Vij |2 = ci for all i ∈ N.
We are now ready to prove Theorem 4.4.2.
Proof of TheoremP4.4.2. Let K be an admissible trace class operator with kernel K(x; y). We
write K(x; x′ ) = j cj fj (x)fj (x′ ) for an orthonormal basis {fj } of L2 (R3 ). Since 0 ≤ K ≤ 1,
tr K = N , we conclude that 1 ≤ cj ≤ 1 and ∞ 1 N
∈ ℓ2 (C) be such
P
j=1 cj = N . Let V , . . . , V
PN
that j=1 |Vij |2 = ci for all i ∈ N. Let θ = {θ1 , θ2 , . . . } be an infinite sequence of real numbers.
Then, for k = 1, . . . , N , we define
∞
X
Fkθ (x) = eiθj Vjk fj (x)
j=1

Observe that X k
⟨Fkθ , Fℓθ ⟩ = ei(θi −θj ) Viℓ V j ⟨fj , fℓ ⟩ = δk,ℓ
i,j

The functions F1θ , . . . , FNθ are therefore orthonormal for every choice of θ. Let
θ
= Slater F1θ , . . . , FNθ


ψN
θ θ θ θ θ
and γN = |ψN ⟩⟨ψN |. We consider then the mixed state γN := ⟨γN ⟩θ obtained by averaging γN
over θ. This definition is a bit formal, since it involves averaging over infinitely many variables,
but it can be made precise by appropriate approximation arguments. The one-particle density
associated with γN is given by
* N + N X
(1)
X θ X j
γN (x; x′ ) = Fjθ (x)F j (x′ ) = Vij V ℓ fi (x)f ℓ (x′ )⟨ei(θi −θℓ ) ⟩θ
j=1 θ j=1 i,ℓ

X N
X
= fi (x)f i (x′ ) |Vij |2 = K(x; x′ )
i j=1

108
where we used that ⟨ei(θi −θℓ ) ⟩θ = δiℓ . The two-point reduced density associated with γN is
(1) θ ′
similarly given by (setting γθ (x; x′ ) = N θ
P
j=1 Fj (x)F j (x ))

(2) (1) (1) (1) (1)

2γN (x, y; x′ , y ′ ) = ⟨γθ (x; x′ )γθ (y; y ′ ) − γθ (x; y ′ )γθ (y; x′ )⟩θ
N
X X j ℓ
= Vij1 V i2 Viℓ3 V i4 ⟨ei(θi1 −θi2 +θi3 −θi4 ) ⟩θ
j,ℓ=1 i1 ,i2 ,i3 ,i4

× fi1 (x)f i2 (x′ )fi3 (y)f i4 (y ′ ) − fi1 (x)f i2 (y ′ )fi3 (y)f i4 (x′ )

= K(x; x′ )K(y; y ′ ) − K(x; y ′ )K(y; x′ )

N 2
X X j
Vij1 V i2 fi1 (x)f i1 (x′ )fi2 (y)f i2 (y ′ ) − fi1 (x)f i1 (y ′ )fi2 (y)f i2 (x′ )


−
i1 ,i2 j=1

where we used the fact that ⟨ei(θi1 −θi2 +θi3 −θi4 ) ⟩θ = δi1 ,i2 δi3 ,i4 + δi1 ,i4 δi2 ,i3 − δi1 ,i2 ,i3 ,i4 . Hence the
energy of the mixed state γN is given by
θ θ
tr γN H = ⟨ψN , HψN ⟩ θ
= F (K)
Z N 2
X X j
Vij1 V i2 |fi1 (x)|2 |fi2 (y)|2 − fi1 (x)f i1 (y)fi2 (y)f i2 (x)


− dxdy v(x − y)
i1 ,i2 j=1

≤ F (K)

using the assumption that v ≥ 0. In particular, there exists at least one sequence θ0 =
{θ10 , θ20 , . . . } such that
θ0 θ0
⟨ψN , HψN ⟩ ≤ F (K)
θ 0
Since ψN is a Slater determinant, this concludes the proof of the theorem.
We are now ready to prove an upper bound for the quantum mechanical energy EQM (N ) :=
Q −1/3
EN (Z, R)= EaQN N0 (aN Z, aN R). For simplicity, we ignore here the spin of the electrons.
From Theorem 4.4.2, we find
M
! M
X Zj X Zi Zj
EQM (N ) ≤ F (K) ≤ tr −∆ − K+
|x − Rj | |Ri − Rj |
Z
j=1 i<j (4.106)
1 1
+ dxdy K(x, x)K(y, y)
2 |x − y|
for every non-negative trace class operator K (with kernel K(x; y)) with 0 ≤ K ≤ 1 and
tr K = N . Here we neglected the negative part of the interaction energy (this is allowed, since

109
we are looking for an upper bound). We choose now
Z
′ 1
K(x; x ) = dpdr M (p, r) fpr (x)f pr (x′ )
(2π)3
where 0 ≤ M (p, r) ≤ 1 is such that
Z
1
dpdr M (p, r) = N
(2π)3
R
and fpr (x) = g(x − r)eipx for some g with |g|2 = 1. K is admissible because
Z Z
1
tr K = dx K(x; x) = dpdrM (p, r) = N
(2π)3
and because, for arbitrary φ ∈ L2 (R3 ) with ∥φ∥ = 1,
Z
⟨φ, Kφ⟩ = dxdx′ φ(x′ )K(x; x′ )φ(x′ )
Z Z 2
1
= dpdr M (p, r) dxg(x − r)eipx φ(x)
(2π)3
Z Z 2
1 ipx
≤ dpdr dxg(x − r)e φ(x)
(2π)3
Z
= drdx|g(x − r)|2 |φ(x)|2 = 1
2/3 2
Next, we compute the energy F (K). We fix, from now on,R M (p, r) = θ(γρ(r) − p ), 2where
ρ is the minimizer of the Thomas-Fermi functional, with ρ(x) = N (and with γ = (6π )2/3 ).
The kernel of ∆K is given by
∆K(x; x′ ) = ∆x K(x; x′ )
Z
1 ′
= 3
dpdr M (p, r)(∆g)(x − r)g(x′ − r)eip(x−x )
(2π)
Z
2i ′
+ 3
dpdr p M (p, r)(∇g)(x − r)g(x′ − r)eip(x−x )
(2π)
Z
1 ′
− 3
dpdr p2 M (p, r) g(x − r)g(x′ − r)eip(x−x )
(2π)
Using the fact that M (r, p) depends only on p2 , we find (computing the integrals over p)
Z
1
tr (−∆)K = dxdpdr θ(γρ(r)2/3 − p2 )|∇g(x − r)|2
(2π)3
Z
1
+ dxdpdr p2 θ(γρ(r)2/3 − p2 )|g(x − r)|2
(2π)3
Z Z
3
=N |∇g(x)| dx + γ dx ρ(x)5/3
2
5

110
Moreover, since (after integration over p), K(x; x) = (ρ ∗ |g|2 )(x), we find that
M Z M
X Zj 2
X Zj
tr K= dx (ρ ∗ |g| )(x)
j=1
|x − Rj | j=1
|x − Rj |
Z Z
2
= dxdy ρ(y) |g| (x − y)V (y) = dy Vg (y)ρ(y)

where we put V (x) = M −1

and Vg = V ∗ |g|2 (we will consider the limit where
P
j=1 Zj |x − Rj |
|g|2 tends to a delta-function, and therefore Vg ≃ V ). In the last two equations, we computed
tr (−∆ − M −1
P
j=1 Zj |x − Rj | ) K. Next, we estimate the last term on the r.h.s. of (4.106).
Again, since K(x; x) = (ρ ∗ |g|2 )(x), we find
Z
1
dxdy K(x; x) K(y; y)
|x − y|
Z
1
= dxdydzdw |g|2 (z − x) ρ(z) |g|2 (y − w)ρ(w)
|x − y|
Z Z
1
= dwdzρ(w)ρ(z) dxdydpdp1 dp2 2 e−ip(x−y) |g| d2 (p1 )eip1 (x−w) |g|
d2 (p2 )eip2 (y−w)
|p|
Z Z
1 d2 d2 (p2 )δ(p − p1 )δ(p − p2 )eip(w−z)
= dwdzρ(w)ρ(z) dpdp1 dp2 2 |g| (p1 )|g|
|p|
Z
1 d2 2
= dp 2 |g| (p) |b ρ(p)|2
|p|
Z
2 1
2
≤ ∥|g| ∥∞
d ρ(p)|2
|b
|p|2
Z
2 2 1
≤ ∥|g| ∥1 dxdy ρ(x)ρ(y) = 2D(ρ, ρ)
|x − y|

Inserting these estimates into (4.106), we conclude that

Z Z Z M
3 2 5/3
X Zi Zj
EQM (N ) ≤ N |∇g(x)| dx + γ ρ (x)dx − Vg (x)ρ(x)dx + + D(ρ, ρ)
5 i<j
|Ri − R j |
Z Z
= ET F (N ) + N |∇g(x)|2 dx + dx (Vg (x) − V (x)) ρ(x)
(4.107)

To obtain the desired upper bound, we have to choose g to be close enough to a delta-function
2
R the2 last term sufficiently small. When |g| approaches a δ-function, however, the
to make
term |∇g| diverges to infinity. So we have to find a good compromise for g. To this end we

111
set, for some R > 0 to be fixed later on,
(
sin(π|x|/R)
√1 for |x| < R
g(x) := 2πR |x|
,0 for |x| > R
R
It is then simple to check that |g|2 = 1 for arbitrary R > 0. Moreover, by simple scaling,
we find Z
|∇g(x)|2 = CR−2 (4.108)

for a constant C. On the other hand, we have

Z M Z  
X 1 1 2
(V − Vg ) (x)ρ(x) = Zj dxρ(x) − ∗ |g| (x − Rj )
j=1
|x − Rj | |.|
M Z
X ρ(x)
≤ Zj dx
j=1 |x−Rj |≤R |x − Rj |

where we used that, by Newton’s Theorem (since g is radial), |.|−1 ∗|g|2 (x) = |x|−1 for |x| > R.
To conclude the proof, we just have to make sure that ρ(x) is not too large when |x − Rj | < R.
To this end, we use the Thomas-Fermi equation for the minimizer ρ:
 
2/3 1
γρ (x) = V (x) − ∗ ρ(x) − µ
|.| +

for some µ ≥ 0. This equation implies that

M
X Zj
γρ2/3 (x) ≤ V (x) =
j=1
|x − Rj |

−1/3
Choosing R > 0 so small that |Ri − Rj | ≫ R for all i ̸= j (recall that |Ri − Rj | ≃ aN ≃
N −1/3 ), we find that
Zj
ρ2/3 (x) ≤ C
|x − Rj |
if |x − Rj | ≤ R, for an appropriate constant C > 0. Hence
Z M Z
X 5/2 dx
(V (x) − Vg (x)) ρ(x) ≤ C Zj 5/2
≤ CR1/2 N 5/2
j=1 |x−Rj |≤R |x − Rj |

Inserting the last bound, together with (4.108) into (4.107), we conclude that

EQM (N ) ≤ ET F (N ) + CN R−2 + CR1/2 N 5/2

112
for all R ≪ N −1/3 . Taking R = N −3/5 ≪ N −1/3 , we find

EQM (N ) ≤ ET F (N ) + CN 11/5 (4.109)

Since N 11/5 ≪ N 7/3 ≃ ET F (N ), this provides the desired upper bound to the quantum
mechanical energy.
Lower Bound on the Quantum Energy. We need a lower bound for the energy
* N
+ * N
+
X
X 1
⟨ψ, HN,M (Z, R)ψ⟩ = ψ, −∆xj − V (xj ) ψ + U + ψ, ψ
j=1 i<j
|xi − xj |

where
M M
X Zj X Zi Zj
V (x) = , U=
j=1
|x − Rj | i<j
|Ri − Rj |

and ψ ∈ L2a (R3N ) is such that ∥ψ∥ = 1. The electron interaction can be bounded from below
using the bound on the indirect part of the Coulomb energy that we proved last semester:
* N
+ Z
X 1 4/3
ψ, ψ ≥ D(ρψ , ρψ ) − C ρψ (x)dx
i<j
|x i − x j |
R
where ρψ (x) = N dx2 . . . dxN |ψ(x, x2 , . . . , xN )|2 is the electron density associated with ψ.
Hence
* N
+ Z
X
4/3
⟨ψ, HN,M (Z, R)ψ⟩ ≥ ψ, −∆xj − V (xj ) ψ + U + D(ρψ , ρψ ) − C ρψ (x)dx
j=1

Next, we observe that for an arbitrary density ρe,

D(ρψ , ρψ ) = D(ρψ − ρe, ρψ − ρe) + 2D(ρψ , ρe) − D(e

ρ, ρe) ≥ 2D(ρψ , ρe) − D(e
ρ, ρe)

and thus
⟨ψ, HN,M (Z, R)ψ⟩
* N     + Z
X 1 4/3
≥ ψ, −∆xj − V − ∗ ρe (xj ) ψ + U − D(e ρ, ρe) − C ρψ (x)dx
j=1
|.| (4.110)
XN Z N
X
4/3
≥ ⟨ψ, hj ψ⟩ + U − D(e
ρ, ρe) − C ρψ (x)dx + ε ⟨ψ, −∆xj ψ⟩
j=1 j=1

113
where we defined hj = −(1 − ε)∆xj − ϕ(x e j ), for some 0 < ε < 1 to be chosen later on, and
ϕe = V − |.|−1 ∗ ρe (we will make use of the ε part of the kinetic energy to bound the negative
contribution given by the third term on Rthe r.h.s. of (4.110)). We fix ρe to be the minimizer,
among all ρ ∈ L5/3 ∩ L1 with ρ ≥ 0 and ρ = N , of the Thomas-Fermi energy functional
Z Z
3 5/3
εeT F (ρ) = (1 − ε)γ dx ρ (x) − dx V (x)ρ(x) + D(ρ, ρ) + U
5

where we replaced the constant γ = (6π 2 )2/3 by (1 − ε)γ. By Theorem 4.1.6, ρe satisfies the
Thomas-Fermi equation h i
2/3
(1 − ε)γ ρe (x) = ϕ(x) − µ
e e (4.111)
+

e ≥ 0 determined by the condition

for some µ
Z "e #3/2
ϕ(x) − µ
Z
e
ρe(x)dx = dx = N . (4.112)
(1 − ε)γ
+

We begin by estimating the infimum of the first term on the r.h.s. of (4.110). Since N
P
j=1 hj
is a sum of one-particle Hamiltonians, its minimum is achieved on a Slater determinant. Hence
N
X N
X
⟨ψ, hj ψ⟩ ≥ inf tr h |fj ⟩⟨fj |
{f1 ,...,fN }
j=1 j=1

where the infimum is taken over all orthonormal sets {f1 , . . . , fN } in L2 (R3 ). To compute this
infimum, we use again the coherent states introduced when we were looking for the upper
bound to the energy. We set fpr (x) = g(x − r)eipx , with g(x) = (2πR)−1/2 sin(π|x|/R)/|x| for
|x| < R and g(x) = 0 for |x| ≥ R. We denote moreover by πpr = |fpr ⟩⟨fpr | the orthogonal
projection onto fpr , with kernel πpr (x; x′ ) = fpr (x)f pr (x′ ). We have the completeness relation
Z
1
dpdr πpr (x; x′ ) = δ(x − x′ )
(2π)3

which implies that, for any m ∈ L2 (R3 ),

Z
1
(m, m) = dpdr (m, πpr m) (4.113)
(2π)3

Moreover, we can compute

Z Z Z Z
2 1
|∇m(x)| dx = dpdr p (m, πpr m) − |m(x)| dx |∇g(x)|2 dx
2 2
(4.114)
(2π)3

114
In fact, using that g is real, we obtain, integrating by parts,
Z Z
1 1
3
2
dpdr p (m, πpr m) = 3
dpdrdxdx′ p2 m̄(x)fpr (x)f¯pr (x′ )m(x′ )
(2π) (2π)
Z
1 ′ 2 ′ ′ ipx −ipx′
= dpdrdxdx p m̄(x) m(x ) g(x − r) g(x − r) e e
(2π)3
Z
1 ′
= 3
dpdrdxdx′ m̄(x) m(x′ ) g(x − r) g(x′ − r)∇x eipx ∇x′ e−ipx
(2π)
Z
1 ′
= 3
dpdrdxdx′ ∇m(x) ∇m(x′ ) g(x − r) g(x′ − r)eip(x−x )
(2π)
Z
1 ′
+ 3
dpdrdxdx′ m(x) m(x′ ) ∇g(x − r) ∇g(x′ − r)eip(x−x )
(2π)
Z Z
= drdx |∇m(x)| |g(x − r)| + drdx |m(x)|2 |∇g(x − r)|2
2 2

Z Z Z
2 2
= dx |∇m(x)| + dx |m(x)| dx|∇g(x)|2

We also observe that

Z Z Z
1
3
dpdr ϕ(r)⟨m, πpr m⟩ = dxdr ϕ(r)|m(x)| |g(x − r)| = dx|m(x)|2 ϕeg (x) (4.115)
e e 2 2
(2π)

where we introduced the notation ϕeg (x) = ϕe ∗ |g|2 (x). From (4.110), we find that
N
X Z  
⟨ψ, HN,M (Z, R)ψ⟩ ≥ inf tr hg |fj ⟩⟨fj | + ϕeg (x) − ϕ(x)
e ρψ (x)
f1 ,...,fN
j=1
Z N
X
4/3
+ U − D(e
ρ, ρe) − C ρψ (x)dx +ε ⟨ψ, −∆xj ψ⟩
j=1

where we defined hg = −(1 − ε)∆ − ϕeg (x) and where the infimum in the first term is taken
over all orthonormal sets {f1 , . . . , fN } in L2 (R3 ). For any suchPset of N orthonormal functions
N
{f1 , . . . , fN }, and for any p, r ∈ R3 , we define M (p, r) = j=1 ⟨fj , πpr fj ⟩. Then we have
3
0 ≤ M (p, r) ≤ 1 for every p, r ∈ R (since M (p, r) = tr πpr K, where K is the orthogonal
projection onto the space spanned by f1 , . . . , fN ) and (using (4.113)),
Z
1
dpdr M (p, r) = N .
(2π)3

115
From (4.114), (4.115), we conclude that
Z Z
1 h i
⟨ψ, HN,M (Z, R)ψ⟩ ≥ inf dpdr (1 − ε)p − ϕ(r) M (p, r) − N |∇g(x)|2 dx
2 e
M (2π)3
Z  
+ ϕeg (x) − ϕ(x)
e ρψ (x)
Z N
X
4/3
+ U − D(e
ρ, ρe) − C ρψ (x)dx + ε ⟨ψ, −∆xj ψ⟩
j=1
(4.116)
where the infimum is taken over all M (p, r) with 0 ≤ M (p, r) ≤ 1 and (2π)−3 M (p, r)dpdr =
R

N . It is simple to check that the infimum in the first term on the r.h.s. of the last equation
e − (1 − ε)p2 − µ) for some µ ≥ 0 (here θ(s) = 1 if s > 0 and
is achieved at M0 (p, r) = θ(ϕ(r)
θ(s) = 0 if s ≤ 0). The value of µ is determined by the condition (using that γ = (6π 2 )2/3 )
" #3/2
e − p2 − µ) = dr ϕ − µ
Z Z
1 e
N= 3
dpdrθ(ϕ(r)
(2π) γ(1 − ε)
+

We conclude from (4.112) that µ = µ e in the Thomas-Fermi equation (4.111). We compute

the energy associated with the minimizer M0 (p, r):
Z
1 h
2
i
dpdr (1 − ε)p − ϕ(r)
e M0 (p, r)
(2π)3
!
1
Z h i e −µ
ϕ(r)
dpdr (1 − ε)p2 − ϕ(r) χ p2 ≤
e
= 3
e
(2π) (1 − ε)
" # 5/2 " #3/2
4π(1 − ε) e −µ e −µ
Z Z
ϕ(r) 4π ϕ(r)
−
e e
= dr dr ϕ(r)
e
5(2π)3 (1 − ε) 3(2π)3 (1 − ε)
Z Z + +
3
= (1 − ε)γ dr ρe5/3 (r) − ρe(r)ϕ(r) e
5
Z Z
3 5/3
= (1 − ε)γ dr ρe (r) − V (r)ϕ(r) e + 2D(e ρ, ρe)
5
where we used the Thomas-Fermi equation (4.111), and the fact that ϕe = V − |.|−1 ∗ ρe. From
(4.116), we find
Z  
−2
⟨ψ, HN,M (Z, R)ψ⟩ ≥ ET F (N ) − CN R +
e ϕeg (x) − ϕ(x)
e ρψ (x)
Z N
X (4.117)
4/3
−C ρψ (x)dx +ε ⟨ψ, −∆xj ψ⟩
j=1

116
where we used the choice of ρe as Rminimizer of the TF-functional εeT F (with γ replaced by
(1 − ε)γ). Here, we also estimated |∇g(x)|2 dx ≤ CR−2 , for a universal constant C > 0.
Next, we bound the third term on the r.h.s. of (4.117). We have
Z   Z
V ∗ |g|2 (x) − V (x) ρψ (x)dx


ϕeg (x) − ϕ(x)
e ρψ (x) =
Z  
1 1 2
+ ∗ ρe − ∗ ρe ∗ |g| (x)ρψ (x)dx
|.| |.|
Observe that  
1
∆ ∗ ρe = −e
ρ≤0
|.|
Hence |.|−1 ∗ ρe is superharmonic, and |.|−1 ∗ ρe ≥ |.|−1 ∗ ρe ∗ |g|2 pointwise. Therefore
Z   Z
V ∗ |g|2 (x) − V (x) ρψ (x)dx


ϕg (x) − ϕ(x) ρψ (x) ≥
e e

M Z
X ρψ (x)
≥ − Zj dx
j=1 |x−Rj |≤R |x − Rj |
≥ − CZR1/5 ∥ρψ ∥5/3

where we used the fact that, by Newton’s Theorem, |. − Rj |−1 ∗ |g|2 (x) = |x − Rj |−1 if
PM
|x − Rj | ≥ R, and Hölder’s inequality (and we set Z = j=1 Zj to be the total nuclear
charge). From (4.117), we obtain (since Z ≃ N )
eT F (N ) − CN R−2 − CN R1/5 ∥ρψ ∥5/3
⟨ψ, HN,M (Z, R)ψ⟩ ≥ E
N
(4.118)
Z X
4/3
− C ρψ (x)dx + ε ⟨ψ, −∆xj ψ⟩
j=1

To bound the last two terms, we use that, by the kinetic energy inequality,
N
X Z
5/3
ε ⟨ψ, −∆xj ψ⟩ ≥ C0 ε ρψ (x)dx
j=1

for an appropriate constant C0 > 0 and that, by simple interpolation

Z Z Z Z
4/3 5/3 −1 −1 5/3
C ρψ (x)dx ≤ Cαε ρψ (x)dx + Cα ε ρψ (x)dx = Cαε ρψ (x)dx + Cα−1 ε−1 N

From (4.118), we find (choosing α of order one, but sufficiently small)

eT F (N ) − CN R−2 − CN ε−1 − CN R1/5 ∥ρψ ∥5/3 + Cε∥ρψ ∥ 5/3
⟨ψ, HN,M (Z, R)ψ⟩ ≥ E 5/3 (4.119)

117
We still have to replace, in the r.h.s., the Thomas-Fermi energy E eT F (N ) (associated with the
modified functional εeT F having γ replaced by (1 − ε)γ) by the “true” Thomas-Fermi energy
ET F (N ). Using ρe as a trial function for the “true” Thomas-Fermi functional, we deduce that
Z
3 eT F (N ) + CεN 7/3
ET F (N ) ≤ E eT F (N ) + γε ρe(x)dx ≤ E
5
(by the scaling properties of the Thomas-Fermi functional, it is simple to check that
R 5/3 7/3
ρe (x)dx is of the order aN ≃ N 7/3 ). Finally, to control the last two terms on the r.h.s. of
(4.119), we remark that
5/3
−CN R1/2 ∥ρψ ∥5/3 + Cε∥ρψ ∥5/3 ≥ inf −CN R1/2 X + CεX 5/3 = −CN 5/2 R1/2 ε−3/2
X≥0

Summarizing, we find the bound

⟨ψ, HN,M (Z, R)ψ⟩ ≥ ET F (N ) − CεN 7/3 − CN R−2 − CN 5/2 R1/2 ε−3/2 − CN ε−1 (4.120)

valid for all ψ ∈ L2a (R3N ) with ∥ψ∥ = 1, for all 0 < ε < 1 and all R > 0. Choosing, for
example, R = N −1/2 and ε = N −1/30 , we find

⟨ψ, HN,M (Z, R)ψ⟩ ≥ ET F (N ) − CN 7/3−δ

for some δ > 0. This is the desired lower bound. Together with the upper bound (4.109), this
completes the proof of Theorem 4.4.1.

5 Hartree-Fock Theory
5.1 The Hartree-Fock Energy Functional
In this section, we are going to study the relation between many body quantum mechanics
and the Hartree-Fock theory. Our goal is to prove that Hatree-Fock theory provides a very
good approximation for the ground state energy of quantum systems of atoms and molecules
in the Born-Oppenheimer approximation.
We consider the N -electron Hamiltonian
N
" k
# N
X X zi X 1
HN = −∆xj − +
j=1 i=1
|xj − Ri | i<j
|xi − xj |

acting on the antisymmetric tensor product L2 (R3 ; C2 )∧N ≃ L2a (R3N ; C2N ). We are interested
in estimating the ground state energy
Q
({zi , Ri }) = inf ⟨ψN , HN ψN ⟩ : ψN ∈ L2a (R3N ; C2N ) and ∥ψN ∥2 = 1

EN

118
 For a one-particle observable A (a self-adjoint operator on L2 (R3 ; C2 )), we define Ai =
1 ⊗ · · · ⊗ A ⊗ · · · ⊗ 1 as the operator acting as A on the i-th particle and acting trivially on
the other (N − 1) particles. For any many body wave function ψN ∈ L2a (R3N ; C2N ), we define
the one-particle density matrix γ so that
* N
+
X
ψN , Ai ψN = tr Aγ
i=1

In particular, this implies that tr γ = N . An important consequence of the Pauli principle

(the antisymmetry of ψ), is the operator inequality 0 ≤ γ ≤ 1. In fact, every operator γ on
L2 (R3 ; C2 ) with 0 ≤ γ ≤ 1 and such that tr γ = N is the one-particle density matrix of a (in
general mixed) N -particle state.
An important class of wave functions ψN ∈ L2a (R3N ; C2N ) are Slater determinants, having
the form
ψN (x1 , . . . , xN ) = (N !)−1/2 det (φi (xj )) (5.121)
where {φ1 , . . . , φN } is an orthonormal system in L2 (R3 ; C2 ). The reduced one-particle density
of the Slater determinant (5.121) is the orthogonal projection
N
X
ω= |φj ⟩⟨φj |
j=1

onto the N -dimensional subspace spanned by φ1 , . . . , φN . It is easy to see that every projection
with rank N on L2 (R3N ; C2N ) corresponds to a Slater determinant.
The expectation of the energy of a Slater determinant with one-particle density ω is given
by
Z
1 1 
ω(x, x)ω(y, y) − |ω(x, y)|2

EHF (ω) = tr (−∆ + V ) ω + dxdy (5.122)
2 |x − y|
where we introduced the notation
k
X zi
V (x) = −
i=1
|x − Ri |

Eq. (5.122) defines the Hartree-Fock energy functional. The Hartree-Fock ground state energy
is then given by
HF
({zi , Ri }) = inf EHF (ω) : ω is orthogonal projection on L2 (R3 ; C2 ) with trω = N

EN
Q HF
We are interested in the relation between EN and its approximation EN . A first simple
observation in this direction is given by the following theorem.

119
Theorem 5.1.1. Let Ri , zi , i = 1, . . . , k depend on N such that zi /N → λi , R1 = 0 and
Ri N −1/3 → ri or to infinity, for all i = 1, . . . , k. Then
HF
EN ({zi , Ri })
lim Q
=1
N →∞ E ({z , R })
N i i

HF Q
Proof. Since clearly EN ({zi , Ri }) ≥ EN ({zi , Ri }), we immediately find that
HF
EN ({zi , Ri })
lim inf Q
≥1
N →∞ EN ({zi , Ri })
On the other hand, in the first part of the class, when we proved that Thomas-Fermi theory
approximates many body quantum mechanics, we found a sequence of Slater determinants
ψN such that ∥ψN ∥ = 1 and
⟨ψN , HN ψN ⟩
lim T F =1
N ({zi , Ri })
N →∞ E

HF
Since EN ({zi , Ri }) ≤ ⟨ψN , HN ψN ⟩ and since we showed that
Q TF
EN ({zi , Ri })/EN ({zi , Ri }) → 1,

as N → ∞, we conclude that
HF
EN ({zi , Ri })
lim sup Q
≤1
N →∞ EN ({zi , Ri })

This theorem proves that Hartree-Fock theory, like Thomas-Fermi theory captures the
leading order term in the ground state energy. In fact, we will show that Hartree-Fock theory
Q HF TF
does much better than that. While EN ({zi , Ri }), EN ({zi , Ri }) and EN ({zi , Ri }) are all of
7/3
order N is the limit N → ∞, we will prove that
Q HF
|EN ({zi , Ri }) − EN ({zi , Ri })| ≤ CN 5/3−ε (5.123)

for a sufficiently ε > 0. In particular, this implies that Hartree-Fock theory captures the
correct contribution of order N 2 (the Scott corrections) and of the order N 5/3 (the Dirac-
Schwinger corrections). Notice, on the other hand, that Thomas-Fermi theory only predicts
correctly the leading contribution to the ground state energy. The first proof of (5.123) was
found by Bach in 1992. Two years later, Graf and Solovej extended and simplified the result
of Bach. In the following, we are going to present the proof of Graf-Solovej.
An important observation, which goes back to Lieb, concerning the Hartree-Fock ground
state energy, is that minimizing the Hartree-Fock functional EHF (ω) over all projections ω

120
with rank N gives the same result as minimizing EHF (ω) over all ω with 0 ≤ ω ≤ 1 with
trω = N . Here we are using the r.h.s. of (5.122) to extend the Hartree-Fock functional to
arbitrary one-particle density matrices ω; notice, however, that if ω is not a rank N projection,
EHF (ω) is not the energy of the N -particle state ψN with reduced density ω (because ψN is
not a Slater determinant). We prove this fact following an idea of Bach.

Lemma 5.1.2. Let 0 ≤ ω ≤ 1, trω = N be a one-particle density matrix with finite rank. Then
there exists an orthogonal projection ω
e=ω e ∗ with tr ω
e2 = ω e = N such that EHF (e
ω ) ≤ EHF (ω).

Proof. We may assume EHF (ω) < ∞ since otherwise there is nothing to prove. We write
M
X
ω= λk |φk ⟩⟨φk |
k=1

for a finite M < ∞, for appropriate 0 < λk ≤ 1 with M

P
k=1 λk = N , and for orthonormal
φ1 , . . . , φM . We define
hk = ⟨φk , (−∆ + V ) φk ⟩
and
1
Vkℓ = ⟨φk ∧ φℓ , φk ∧ φℓ ⟩
|x − y|
Z
1 
|φk (x)|2 |φℓ (y)|2 − φk (x)φk (y)φℓ (x)φℓ (y)

= dxdy
|x − y|

Notice that Vkk = 0 and Vkℓ ≥ 0 for all k, ℓ. Moreover

M M
X 1 X
EHF (ω) = λk hk + λk λℓ Vkℓ
k=1
2 k,ℓ=1

We can assume M > N , since otherwise, there is nothing to prove. Then there exist at least
two indices i, j ∈ {1, . . . , M } with 0 < λi , λj < 1. Without loss of generality, we can assume
that
XM XM
hj + λk Vkj ≤ hi + λk Vki
k=1 k=1

We set δ = min(λi , 1 − λj ) and we define

M
X
ω
e= λk |φk ⟩⟨φk | + (λi − δ)|φi ⟩⟨φi | + (λj + δ)|φj ⟩⟨φj |
k̸=i,j

121
Notice, first of all, that 0 ≤ ω
e ≤ 1 and tre
ω = N . Since

e = ω − δ|φi ⟩⟨φi | + δ|φj ⟩⟨φj |

ω

we find
Z
1 1 
ω (x, y)|2

EHF (e
ω ) = tr(−∆ + V )eω+ dxdy ω ω (y, y) − |e
e (x, x)e
2 |x − y|
= EHF (ω) − δhi + δhj
Z
1
ω(x, x) |φj (y)|2 − |φi (y)|2
 
+ δ dxdy
|x − y|
Z
1  
− δ dxdy ω(x, y) φj (y)φj (x) − φi (y)φi (x)
|x − y|
2 Z
δ 1 
|φj (x)|2 − |φi (x)|2 |φj (y)|2 − |φi (y)|2
 
+ dxdy
2 |x − y|
2 Z
δ 1   
− dxdy φj (x)φj (y) − φi (x)φi (y) φj (y)φj (x) − φi (y)φi (x)
2 |x − y|

This implies
" M
! M
!#
X X
EHF (e
ω ) = EHF (ω) − δ hi + λk Vki − hj + λk Vkj − δ 2 Vij
k=1 k=1

and therefore EHF (e

ω ) ≤ EHF (ω). Moreover, setting

n(ω) = |{λk : 0 < λk < 1}|

we have n(eω ) = n(ω) − 1. Repeating this construction M − N times, we find an orthogonal

projection ω b = N and EHF (b
b with tr ω ω ) ≤ EHF (ω).

Corollary 5.1.3. We have

HF
EN ({zi , Ri }) = inf{EHF (ω) : 0 ≤ ω ≤ 1, tr ω = N } .

Proof. We only have to prove that

HF
EN ({zi , Ri }) ≤ inf{EHF (ω) : 0 ≤ ω ≤ 1, tr ω = N }

since the other inequality is clear. Assume instead that

HF
EN ({zi , Ri }) > inf{EHF (ω) : 0 ≤ ω ≤ 1, tr ω = N }

122
Then there exists 0 ≤ ω ≤ 1 with trω = N and
HF
EHF (ω) < EN ({zi , Ri })

Writing
∞
X
ω= λk |φk ⟩⟨φk |
k=1
P∞
with 0 < λj ≤ 1, j=1 λj = N , we find
∞
X ∞
X
EHF (ω) = λj ⟨φj , (−∆ + V )φj ⟩ + λi λj Vij
j=1 i,j=1

We also find C > 0 such that −∆ + V + C ≥ 0. Then we can write

∞
X ∞
X
EHF (ω) + CN = λj ⟨φj , (−∆ + V + C)φj ⟩ + λi λj Vij
j=1 i,j=1

For any fixed ε > 0, we find M ∈ N such that

M
X
λj ≥ N − ε and
j=1
M
X M
X
HF
λj ⟨φj , (−∆ + V + C)φj ⟩ + λi λj Vij ≤ EN ({zi , Ri }) − ε
j=1 i,j=1

Then we also have

M
X N
λj PM ⟨φj , (−∆ + V + C)φj ⟩
j=1 j=1 λj
M
!2
X N
+ λi λj PM Vij − CN
i,j=1 j=1 λj
!2
N HF

HF
≤ PM EN ({zi , Ri }) − ε + CN − CN < EN ({zi , Ri })
j=1 λj

for ε > 0 small enough.Applying the lemma above to the one-particle reduced density
M
X N
λ j PM |φj ⟩⟨φj |
j=1 j=1 λj

123
we deduce that there exists an orthogonal projection ω
e with tre
ω = N and
HF
EHF (e
ω ) < EN ({zi , Ri })
HF
This contradicts the definition of EN ({zi , Ri }).
In order to compare the many body and the Hartree-Fock ground state energy, it is useful
to introduce another theory which lays halfway between Thomas-Fermi and Hartree-Fock. We
recall the Thomas-Fermi energy functional
Z Z
3 5/3
ET F (ρ) = cT F dxρ (x) − dxV (x)ρ(x) + D(ρ, ρ)
5
where cT F = (3π 2 )2/3 , in appropriate units and where we introduced the notation
Z
1 ρ1 (x)ρ2 (y)
D(ρ1 , ρ2 ) = dxdy
2 |x − y|
It is known P functionalRhas a unique minimizer ρT F , among all ρ ≥ 0
that the Thomas-Fermi
with ρ(x)dx ≤ N . If N ≤ Z = ki=1 zi , then ρT F (x)dx = N . On the other hand, if N > Z,
R

∥ρT F ∥1 = Z < N . The Thomas-Fermi density ρT F defines the Thomas-Fermi potential

1
ϕT F = V + ρT F ∗
|.|
This is the electrostatic potential generated by the nuclei (the V part) and by the configuration
of the electrons, distributed according to the Thomas-Fermi density ρT F .
We use the Thomas Fermi density ρT F and the Thomas-Fermi potential ϕT F to define the
Dirac-Schwinger Hamiltonian
XN Z
4/3
HDS = hi − D(ρT F , ρT F ) − cD dxρT F (x)
i=1

acting on L2a (R3N ; C2N ). Here we introduced the one-particle hamiltonian h = −∆ + ϕT F (x).
Notice that the Dirac-Schwinger Hamiltonian is a non-interacting Hamiltonian; the interaction
among the electrons is described as interaction with the mean field potential ϕT F . Since this
potential overcounts the Coulomb repulsion among the electrons, we have to subtract the term
D(ρT F , ρT F ). We define also the Dirac-Schwinger ground state energy by
DS
({zi , Ri }) = inf ⟨ψN , HDS ψN ⟩ : ψN ∈ L2a (R3N ; C2N ), ∥ψN ∥ = 1

EN
DS
Similarly as in Theorem 5.1.1, it is possible to show that EN ({zi , Ri }) captures
the correct leading order contribution to the many body ground state energy. Hence,
Q TF HF DS
EN ({zi , Ri }), EN ({zi , Ri }), EN ({zi , Ri }), EN ({zi , Ri }) all scale as N 7/3 in the limit of large
N . Our goal is to show the following theorem.

124
Theorem 5.1.4. For sufficiently small ε > 0, there exists a constant C > 0 such that
HF Q
EN ({zi , Ri }) − CZ 5/3−ε ≤ EN HF
({zi , Ri }) ≤ EN ({zi , Ri })
and
Q DS
EN ({zi , Ri }) ≥ EN ({zi , Ri }) − CZ 5/3−ε
Pk
for all N ≤ Z = i=1 zi .

5.2 An Abstract Correlation Inequality

The main obstacle to show Theorem 5.1.4 consists in estimating the interaction energy and,
more precisely, the difference between the total many body interaction energy and the mean
field approximation given by the direct and the exchange term. A useful tool to bound the
interaction energy is given by the Fefferman-de la Llave representation
Z Z
1 1 dr
= dz χ(|x − z| ≤ r)χ(|y − z| ≤ r)
|x − y| π r5
This implies that
* N
+ Z Z ∞
X 1 1 dr X
ψ, ψ = dz ⟨ψ, χ(|xi − z| ≤ r)χ(|xj − z| ≤ r)ψ⟩
i<j
|xi − xj | π 0 r5 i<j

With this formula in mind, it is enough to study expectations of two-particle observables

having the form
XN
X i Xj
i<j

where Xi , Xj are two copies of a single one-particle operator X acting on particles i and,
respectively, on particle j. For this type of operators, we have the following algebraic identity,
which hold in a very abstract setting.
Theorem 5.2.1. Let X, P be bounded operators on a Hilbert space H with X ≥ 0, 0 ≤ P ≤ 1.
For all α ≥ 0 we have then the operator inequality
N N  N
α2 X

X X 1
Xi Xj ≥ αXi − Xi Pi Xi − − Bi
i<j i=1
2 2 i=1

on the antisymmetric tensor product H∧N . The error term can be taken to be
   
1 1 1
B = α+ 5 + 3λ + tr P X (1 − P )X(1 − P ) + (X 2 + P X 2 P )
2 λ 2λ
for any λ ≥ 0.

125
 Remark that the inequality only holds on the antisymmetric subspace H∧N of the tensor
product space H⊗N . The inequality is useful because it estimates a two-particle operator in
terms of a one-particle operator.
Proof. Let Q = 1 − P . We decompose
N
X N
X
Xi Xj = (Qi + Pi )(Qj + Pj )Xi Xj (Qj + Pj )(Qi + Pi )
i<j i<j
X
= (Qi Qj + Qi Pj + Pi Qj )Xi Xj (Qi Qj + Qi Pj + Qj Pi )
i<j
X
+ (Pi Pj Xi Xj Pi Pj + 2Pi Pj Xi Xj Pi Qj + 2Pi Qj Xi Xj Pi Pj )
i<j
X
+ (Pi Pj Xi Xj Qi Qj + Qi Qj Xi Xj Pi Pj )
i<j

= A1 + A2 + A3

with
X
A1 = (Qi Qj + Qi Pj + Pi Qj )Xi Xj (Qi Qj + Qi Pj + Qj Pi )
i<j
1X
A2 = (Pi Pj Xi Xj Pi Pj + 2Pi Pj Xi Xj Pi Qj + 2Pi Qj Xi Xj Pi Pj )
2 i̸=j
1X
A3 = (Pi Pj Xi Xj Qi Qj + Qi Qj Xi Xj Pi Pj )
2 i̸=j

We observe that A1 ≥ 0 (each summands in A1 is non-negative, since X ≥ 0). So, we focus

126
on A2 , A3 . We notice that, for any α ≥ 0,
X 2
α− Pi Xi (1 + Qi )
i
X X X
2
=α −α Pi Xi (1 + Qi ) − α (1 + Qi )Xi Pi + (1 + Qi )Xi Pi Pj Xj (1 + Qj )
i i i,j
X X
2
=α −α (1 − Qi )Xi (1 + Qi ) − α (1 + Qi )Xi (1 − Qi )
i i
X X
+ (1 + Qi )Xi Pi2 Xi (1 + Qi ) + (Pi + 2Qi )Pj Xi Xj Pi (Pj + 2Qj )
i i̸=j
X X X
= α2 − 2α Xi + 2α Qi Xi Qi + (1 + Qi )Xi Pi2 Xi (1 + Qi )
i i i
X
+ (Pi Pj Xi Xj Pi Pj + 2Qi Pj Xi Xj Pi Pj + 2Pi Pj Xi Xj Pi Qj )
i̸=j
X
+4 Qi Pj Xi Xj Pi Qj
i̸=j

Hence, we conclude that

1 X 2 α2
A2 = α− Pi Xi (1 + Qi ) −
2 i
2
 
X 1 2
+ αXi − αQi Xi Qi − (1 + Qi )Xi Pi Xi (1 + Qi )
i
2
X
−2 Qi Xi Pi Pj Xj Qj
i̸=j

127
Similarly, we rewrite also the term A3 . We compute
1 X √ X 2
√ Q i Xi P i + λ Pi Xi Qi
λ i i
1X X
= Pi Xi Qi Qj Xj Pj + λ Qi Xi Pi Pj Xj Qj
λ i,j i,j
X X
+ Pi Xi Qi Pj Xj Qj + Qi Xi Pi Qj Xj Pj
i,j i,j
X
= (Pi Pj Xi Xj Qi Qj + Qi Qj Xi Xj Pi Pj )
i̸=j
 X
1
+ +λ Qi Xi Pi Pj Xj Qj
λ i̸=j
X 1

2 2
+ Qi Xi Pi Qi Xi Pi + Pi Xi Qi Pi Xi Qi + λQi Xi Pi Xi Qi + Pi Xi Qi Xi Pi
i
λ

Therefore, we find
1 1 X √ X 2
A3 = √ Qi Xi Pi + λ Pi Xi Qi
2 λ i i
 X
1 1
− λ+ Qi Xi Pi Pj Xj Qj
2 λ i̸=j
 
1X 2 1 2
− Qi Xi Pi Qi Xi Pi + Pi Xi Qi Pi Xi Qi + λQi Xi Pi Xi Qi + Pi Xi Qi Xi Pi
2 i λ

We note that A2 and A3 can bePwritten as sums of positive terms, one-particle observable,
and the two particle observable i̸=j Qi Xi Pi Pj Xj Qj . To deal with this term, we observe that
X
(Qj Pi − Qi Pj )Xi Xj (Qj Pi − Qi Pj )
i̸=j
X
= (Qj Pi Xi Xj Qj Pi − Qj Pi Xi Xj Qi Pj − Qi Pj Xi Xj Qj Pi + Qi Pj Xi Xj Qi Pj )
i̸=j
X
=2 (Pi Xi Pi Qj Xj Qj − Qi Xi Pi Pj Xj Qj )
i̸=j

128
Hence
X 1X X
Qi Xi Pi Pj Xj Qj = − (Qj Pi − Qi Pj )Xi Xj (Qj Pi − Qi Pj ) + Pi Xi Pi Qj Xj Qj
i̸=j
2 i̸=j i̸=j
X
1/2 1/2
≤ (Qj Xj Qj ) (Pi Xi Pi )(Qj Xj Qj )
i̸=j
X
≤ tr(P XP ) Qj Xj Qj
j
X
≤ trP X Qj Xj Qj
j
(5.124)
Here we used the fact that, for any fixed j = 1, . . . , N , the form
X
⟨ψ, Pi Xi Pi ψ⟩
i̸=j

is maximized on H∧(N −1) by taking ψ = Slater(φ1 , . . . , φN −1 ), where φ1 , . . . , φN −1 are the

eigenvectors of P XP associated with the (N − 1) largest eigenvalues λ1 , . . . , λN −1 of P XP .
This gives
X N
X −1
⟨ψ, Pi Xi Pi ψ⟩ ≤ λj ≤ trP XP ≤ trP X
i̸=j j=1

since 0 ≤ P ≤ 1. Inserting the bound (5.124) in the expressions for A2 and A3 , we conclude
that
N
α2 X
 
X 1 2
Xi Xj ≥ − + αXi − αQi XQi − (1 + Qi )Xi Pi Xi (1 + Qi )
i<j
2 i
2
 
1X 2 1 2
− Qi Xi Pi Qi Xi Pi + Pi Xi Qi Pi Xi Qi + λQi Xi Pi Xi Qi + Pi Xi Qi Xi Pi
2 i λ
  
1 1 X
− 2+ λ+ tr (P X) Qi Xi Qi
2 λ i

The right hand side is now a one-particle operator. We can simplify it a bit by noticing that
(1 + Q)XP 2 X(1 + Q) = XP (1 − Q)X + QXP 2 X + XP 2 XQ + QXP 2 XQ
≤ XP X − XP QX + (1 + λ)QXP 2 XQ + λ−1 XP 2 X
Since moreover
XP QX = (P + Q)XP QX(P + Q) ≥ P XP QXQ + QXP QXP

129
we find
X X 1 α2
Xi Xj ≥ (αXi − Xi Pi Xi ) −
i<j i
2 2
X 1 1

2 2 2
− αQi Xi Qi + (1 + 2λ)Qi Xi Pi Xi Qi + (Xi Pi Xi + Pi Xi Qi Xi Pi )
i
2 2λ
  
1 1 X
− 2+ λ+ trP X Qi Xi Qi
2 λ i

Finally, we use
QXP 2 XQ ≤ ∥X 1/2 P 2 X 1/2 ∥QXQ ≤ tr(X 1/2 P 2 X 1/2 ) QXQ ≤ (tr XP )QXQ
and
XP 2 X + P XQ2 XP ≤ X 2 + P X 2 P
This gives
X X 1 α2 X
Xi Xj ≥ (αXi − Xi Pi Xi ) − − Bi
i<j i
2 2 i

with    
1 1 1
B = α+ 5 + 3λ + tr(XP ) QXQ + (X 2 + P X 2 P )
2 λ 2λ

5.3 Estimate for the Many-Body Interaction Energy

We apply now Theorem 5.2.1, specializing on X being a projection. Later, we will choose
X as the multiplication operator with the characteristic function χ(|x − z| ≤ r), to derive a
bound for the interaction energy.
Corollary 5.3.1. If X 2 = X = X ∗ , we have, for any ψ ∈ H∧N antisymmetric,
* N
+ r !
X 1 1 77 p
(trXγ)2 − trXγXγ − tr(Xγ) min 1,

ψ, Xi Xj ψ ≥ tr X(γ − γ 2 )
i<j
2 2 2
(5.125)

P Remark. Notice that the first term on the r.h.s. is exactly the expectation of the operator
i<j Xi Xj in a Slater determinant. The second term on the r.h.s. is an error; the reason
this bound is very useful is that the error term vanishes on Slater determinants (since then
γ 2 = γ), and it is small for states close to Slater determinants (the distance from a Slater
determinant is measured in terms of the difference γ − γ 2 ).

130
Proof. Taking α = tr Xγ, P = 1, λ = ∞ in Theorem 5.2.1, we find
* N
+
X 1 1 1  1
ψ, Xi Xj ψ ≥ (trXγ)2 − trXγ ≥ (trXγ)2 − trXγXγ − trXγ
i<j
2 2 2 2

which proves the bound when the minimum is one. Let’s now consider the case
r
77 p
trX(γ − γ 2 ) < 1
2
In this case, we pick α = tr(Xγ), P = γ and
r
1 1 77 p
= trX(γ − γ 2 )
λ 4 2
so that λ ≥ 4. The theorem gives
* +
X 1
(trXγ)2 − trXγXγ

ψ, Xi Xj ψ ≥
i<j
2
   
1 1
− tr(1 − γ)X(1 − γ)γ · trXγ + 5 + 3λ + trXγ
2 λ
1
trXγ + trγXγ 2


−
2λ
Observe that
tr(1 − γ)X(1 − γ)γ = trXγ − 2trγXγ + trγXγ 2 ≤ trXγ − trγXγ = trX(γ − γ 2 )
while
trγXγ 2 ≤ trXγ
Hence, we find
* +
X 1
(trXγ)2 − trXγXγ

ψ, X i Xj ψ ≥
i<j
2
 
2 7 3 1
− tr(Xγ)trX(γ − γ ) + λ+
2 2 2λ
1
− tr(Xγ)
λ
Since
r r
7 3 1 29 3 77 77 2 1 1 77 1
+ λ+ ≤ + λ≤ λ= p = p
2 2 2λ 8 2 32 8 77 trX(γ − γ )
2 4 2 trX(γ − γ 2 )

131
we conclude that
* + r
X 1 2
 1 77 p
ψ, X i Xj ψ ≥ (trXγ) − trXγXγ − (trXγ) trX(γ − γ 2 )
i<j
2 2 2

Let us stress once again that the bound (5.125) is useful because the error becomes small
when wep consider states close to Slater determinants. This follows from the presence of the
factor trX(γ − γ 2 ). It is natural to ask whether the exponent 1/2 for this factor is optimal
or, instead, whether one can increase it to get a stronger estimate. The following example
shows that the exponent is indeed optimal.
Consider an orthonormal system φ1 , . . . , φN +2 . We build the two N -particle Slater deter-
minants Ψ1 = (N !)−1/2 φ1 ∧ · · · ∧ φN and Ψ2 = (N !)−1/2 φ3 ∧ · · · ∧ φN +2 . For ε > 0, we define
Ψ = (1 − ε)1/2 Ψ1 + ε1/2 Ψ2 . Clearly Ψ1 is orthogonal to Ψ2 and therefore ∥Ψ∥ = 1. For ε = 0,
Ψ is a Slater determinant; for 0 < ε < 1, Ψ is not a Slater determinant. We denote by
N
X
P = |φj ⟩⟨φj |
j=1

and by
N
X +2
′
P = |φj ⟩⟨φj |
j=3

the one-particle density of Ψ1 and, respectively, of Ψ2 . Then, the one-particle density of Ψ is

given by
γ = (1 − ε)P + εP ′
This follows from the observation that ⟨Ψ1 , (A ⊗ 1)Ψ2 ⟩ = 0 for all one-particle operators A
(because there are two particles in Ψ1 that are orthogonal to all particles in Ψ2 , and A only
acts on one particle). We conclude that
γ 2 − γ = ε(1 − ε)P ′′
with
P ′′ = |φ1 ⟩⟨φ1 | + |φ2 ⟩⟨φ2 | + |φN +1 ⟩⟨φN +1 | + |φN +2 ⟩⟨φN +2 |
We define now X to be the projection onto span(φ1 + φN +1 , φ2 + φN +2 ), i.e.
1
X= (|φ1 + φN +1 ⟩⟨φ1 + φN +1 | + |φ2 + φN +2 ⟩⟨φ2 + φN +2 |)
2
1
= (|φ1 ⟩⟨φ1 | + |φ2 ⟩⟨φ2 | + |φN +1 ⟩⟨φN +1 | + |φN +2 ⟩⟨φN +2 |
2
+|φ1 ⟩⟨φN +1 | + |φN +1 ⟩⟨φ1 | + |φ2 ⟩⟨φN +2 | + |φN +2 ⟩⟨φ2 |)

132
Then we have
tr Xγ = 1, and tr X(γ − γ 2 ) = 2ε(1 − ε)
Since Ψ1 , Ψ2 are Slater determinants, we have
X 1
(tr(XP ))2 − trXP XP


⟨Ψ1 , Xi Xj Ψ1 ⟩ =
i<j
2
X 1
(tr(XP ′ ))2 − trXP ′ XP ′


⟨Ψ2 , Xi Xj Ψ2 ⟩ =
i<j
2

Moreover, we find
X 1
⟨Ψ1 , Xi Xj Ψ2 ⟩ =
i<j
4

Hence
X (1 − ε)
ε
(tr(XP ))2 − trXP XP + (tr(XP ′ ))2 − trXP ′ XP ′


⟨Ψ, Xi Xj Ψ⟩ =
i<j
2 2
(1 − ε)1/2 ε1/2
−
4
1 2

ε1/2
= (tr(XP )) − trXP XP − + O(ε)
2 4
1
ε1/2
= (tr(Xγ))2 − trXγXγ − + O(ε)
2 4
This example shows that the exponent 1/2 on the r.h.s. of (5.125) is optimal.

We will make use of another bound for the expectation of the electron interaction, which
also follows from Theorem 5.2.1.

Corollary 5.3.2. Let X = X 2 = X ∗ be an orthogonal projection, 0 ≤ P ≤ 1 and α ≥ 0.

Then
* N
+
X α2 1
ψ, Xi Xj ψ ≥ αtrXγ − − trP XP X
i<j
2 2
 p 
− const · (α + trXP + trXγ) min 1, trX(1 − P )γ(1 − P )

Remark: notice that the bound of Corollary 5.3.2 is an extension of the estimate in Corol-
lary 5.3.1, which can be obtained putting α = trXγ and P = γ.

133
Proof. Let us first assume that
trX(1 − P )γ(1 − P ) ≥ 1
The theorem implies that
* N
+
X 1 α2
ψ, Xi Xj ψ ≥ αtrγX − trXP Xγ −
i<j
2 2
   
1 1
− α+ 5 + 3λ + tr(P X) tr(1 − P )X(1 − P )γ
2 λ
1
− (trXγ + trP XP γ)
2λ
α2
≥ αtrγX − − const · [α + trP X + trγX]
2
where we used that
trP XP γ ≤ trP X and trXP Xγ ≤ trP X
Next, we assume that
trX(1 − P )γ(1 − P ) ≤ 1
From the theorem, we deduce that
N X α2 X

X 1
Xi Xj ≥ αXi − Xi Pi Xi − − Bi
i<j i
2 2 i

with
   
1 1 1
B = α+ 5 + 3λ + trP X (1 − P )X(1 − P ) + (X + P XP )
2 λ 2λ
We notice that
X X
X i Pi X i = (Pi + Qi )Xi Pi Xi (Pi + Qi )
i
i 
1 X X
≤ 1+ Pi Xi Pi Xi Pi + (1 + λ) Qi Xi Pi Xi Qi
λ i i
1 X
≤ trP XP XP + ∥X 1/2 P X 1/2 Pi X i Pi
λ i
X
+ (1 + λ)∥X 1/2 P X 1/2 ∥ Qi Xi Qi
i
1X X
≤ trP XP X + Pi Xi Pi + (1 + λ)tr(P X) Qi Xi Qi
λ i i

134
Hence, we obtain
X X 1 α2
Xi Xj ≥ α Xi − trP XP X −
i<j i
2 2
   X
1 1
− α+ 5 + 3λ + tr(P X) Q i Xi Q i
2 λ i
1 X
− (Xi + Pi Xi Pi )
2λ i

and therefore
* +
X 1 α2
ψ, Xi Xj ψ ≥ αtrγX − trP XP X −
i<j
2 2
   
1 1
− α+ 5 + 3λ + trP X trQXQγ
2 λ
1
− (trXγ + trP XP γ)
2λ
We choose
1 p
= trXQγQ
λ
and note that λ ≥ 1.This gives
* +
X 1 α2
ψ, Xi Xj ψ ≥ αtrγX − trP XP X −
i<j
2 2
 
p 5 3 1
− α trQXQγ − + λ+ tr(P X)trQXQγ
2 2 2λ
1p
− trQXQγ (trXγ + trXP )
2
1 α2
≥ αtrγX − trP XP X −
2  2 
p 5 3 1
− α trQXQγ − + λ+ tr(P X)trQXQγ
2 2 2λ
1p
− trQXQγ (trXγ + trXP )
2
1 α2
≥ αtrγX − trXP XP −
2 2 p
− const [α + tr(P X) + tr(Xγ)] trQXQγ

135
 Next, we apply Corollary 5.3.2 to bound the interaction energy of the electrons.
Lemma 5.3.3. Fix 0 < ε < 1/6. For any density matrix P and any density ρ0 (x) ≥ 0, we
have
* N
+
|P (x, y)|2
Z
X 1 1
ψ, ψ ≥ 2D(ρ0 , ργ ) − D(ρ0 , ρ0 ) − dxdy
i<j
|x − y| 2 |x − y|
5/6 1/6+ε
− const · ∥ρ∥5/3 ∥ρ∥1 δ(γ, P )1/3−ε

where ρ = ρ0 + ργ + ρP and we recall that δ(γ, P ) = tr γ(1 − P ).

Proof. We use the Fefferman-de la Llave representation
Z Z ∞
1 1 dr
= dz χ(|x − z| ≤ r)χ(|y − z| ≤ r)
|x − y| π 0 r5
to estimate
D X 1 E
ψ, ψ
i<j
|xi − xj |
Z Z * +
1 dr X
= dz ψ, (Xr,z )i (Xr,z )j ψ
π r5 i<j
2
αr,z
Z Z  
1 dr 1
≥ dz αr,z trXr,z γ − − trP Xr,z P Xr,z
π r5 2 2
Z Z  q 
dr
− const · dz (αr,z + trXr,z P + trXr,z γ) · min 1, trXr,z (1 − P )γ(1 − P )
r5
We observe that
Z
trXr,z γ = dxργ (x)
Br (z)
Z
trXr,z P = dxρP (x)
Br (z)
Z
trXr,z P Xr,z P = dxdy|P (x, y)|2 χ(|x − z| ≤ r)χ(|y − z| ≤ r)
Z
trXr,z (1 − P )γ(1 − P ) = dxρQγQ (x)
Br (z)

We set Z
αr,z = dxρ0 (x)
Br (z)

136
Then we have Z
αr,z + trXr,z γ + trXr,z P = dxρ(x)
Br (z)

Hence, we find
* N
+
|P (x, y)|2
Z
X 1 1
ψ, ψ ≥ 2D(ρ0 , ργ ) − D(ρ0 , ρ0 ) − dxdy
i<j
|xi − xj | 2 |x − y|
Z Z ∞
dr
− const · dz E(r, z)
0 r5
with the error Z   Z 
E(r, z) = ρ(x)dx · min 1, ρQγQ (x)dx
Br (z) Br (z)

For a function f ∈ L1 (R3 ), we define the Hardy-Littlewood maximal function f ∗ by

Z
∗ 1
f (x) = sup dyf (y)
r>0 |Br (x)| Br (x)

The Hardy-Littlewood maximal inequality states that, for all 1 < p ≤ ∞, there exists a
constant C > 0 such that ∥f ∗ ∥p ≤ C∥f ∥p . With this definition, we have
r !
4 3 ∗ 4 3 ∗
E(r, z) ≤ πr ρ (z) · min 1, πr ρQγQ (z)
3 3

This implies that

Z Z ∞ Z Z R Z ∞ 
dr ∗ dr q ∗ dr
dz E(r, z) ≤ const dzρ (z) · ρQγQ (z) +
0 r5 0 r
1/2
R r
2
Z h q i
≤ const dzρ∗ (z) 2R1/2 ρ∗QγQ (z) + R−1

Optimizing the choice of R (taking R = (ρ∗QγQ (z))−1/3 ), we find

Z Z ∞ Z
dr
dz 5
E(r, z) ≤ const dρ∗ (z)(ρ∗QγQ (z))1/3
0 r
1/3
≤ const∥ρ∗ ∥3/(2+ε) ∥ρ∗QγQ ∥1/(1−ε)
1/3
≤ const∥ρ∥3/(2+ε) ∥ρQγQ ∥1/(1−ε)
(5−5ε)/6 (1+5ε)/6 5ε/6 (2−5ε)/6
≤ const∥ρ∥5/3 ∥ρ∥1 ∥ρQγQ ∥5/3 ∥ρQγQ ∥1

137
where we used the Hardy-Littlewood maximal inequality and simple interpolation in the last
two inequalities. Since

QγQ = (1 − P )γ(1 − P ) ≤ (1 − P )γ(1 − P ) + (1 + P )γ(1 + P ) = 2γ + 2P γP ≤ 2(γ + P )

we have ρQγQ (x) ≤ 2ρ(x) pointwise and hence ∥ρQγQ ∥5/3 ≤ 2∥ρ∥5/3 . On the other hand

∥ρQγQ ∥1 = trQγQ ≤ trγ(1 − P ) = δ(γ, P )

This concludes the proof of the lemma (after appropriately redefining ε).

5.4 Semiclassical Approximation through Coherent States

In order to apply Lemma 5.3.3 to estimate the difference between the quantum mechanical
ground state energy and its Hartree-Fock approximation, we need an approximation for the
one-particle density matrix of low energy N -particle wave functions. To this end, we go back
to Thomas-Fermi theory.
In Thomas-Fermi theory, the distribution of the electrons in space is described
R by the
density ρT F which minimizes the Thomas-Fermi functional under the constraint ρ(x)dx ≤ N .
We will assume, from now on, that N ≤ Z, so that ∥ρT F ∥1 = N . The Thomas-Fermi density
satisfy the equation
2/3
cT F ρT F = (µ − ϕT F )+
with cT F = (3π 2 )2/3 and with the Thomas-Fermi potential
1
ϕT F = V + ∗ ρT F
|.|

The chemical potential µ < 0 has to be chosen so, that ∥ρT F ∥1 = N .

Let us define the classical Hamiltonian

h(p, q) = p2 + ϕT F (q)

The Thomas-Fermi density emerges as the marginal of the phase space density M (p, q) =
2χ(h(p, q) ≤ µ), in the sense that
Z
1
dpM (p, q) = ρT F (q)
(2π)3
The classical energy of this electron distribution is given by
Z
1
EC = dpdqh(p, q)M (p, q) = ET F (ρT F ) + D(ρT F , ρT F )
(2π)3

138
The factor D(ρT F , ρT F ) is due to the overcounting of the interaction energy. This remark ex-
plains that, classically, electrons are distributed on phase-space to minimize the self-consistent
Hamiltonian h(p, q). Now, to implement this semiclassical approximation to the computation
of the quantum mechanical ground state energy, we are going to use coherent states. Fix
g ∈ L2 (R3 ) even with ∥g∥2 = 1. For any q, p ∈ R3 , set

fpq (x) = g(x − q)eip·x

Clearly, fpq ∈ L2 (R3 ), with ∥fpq ∥2 = 1. For σ = ±1, we define fpqσ ∈ L2 (R3 ; C2 ) so that it
takes the value fpq in the σ-component and it vanishes in the other component. We define the
corresponding orthogonal projection Πp,q,σ = |fpqσ ⟩⟨fpqσ |. Following the idea that each quan-
tum state occupy the volume (2π)3 on phase-space, we define the semiclassical approximation
for the one-particle density of the quantum ground state
Z
1 X
PSC = dpdqΠp,q,σ
(2π)3 σ h(p,q)≤µ

For any φ ∈ L2 (R3 ; C2 ) we have

Z
1 X
⟨φ, PSC φ⟩ = dpdq |⟨fpqσ , φ⟩|2
(2π)3 σ h(p,q)≤µ
Z
1 X
≤ dpdqdxdyg(x − q)g(y − q)eip(x−y) φσ (x)φσ (y)
(2π)3 σ
XZ
= dxdq|φσ (x)|2 |g(x − q)|2 = ∥φ∥2
σ

Hence 0 ≤ PSC ≤ 1. We also find

Z
2
trPSC = dpdq = N
(2π)3 h(p,q)≤µ

which implies that PSC is a well defined fermionic one-particle density matrix. The integral
kernel of PSC is given by
Z
δσ,σ′
(PSC )σ,σ′ (x, y) = dpdqfpq (x)f pq (y)
(2π)3 h(p,q)≤µ
Z
δσ,σ′
= dpdqg(x − q)g(y − q)eip·(x−y)
(2π)3 h(p,q)≤µ
Z
= δσ,σ′ dqG(q, x − y)g(x − q)g(y − q)

139
where we defined
Z
1
G(q, x) = dpχ(h(p, q) ≤ µ)eip·x
(2π)3
Z
1 
1/2

= dpχ |p| ≤ (µ − ϕTF (q))+ eip·x
(2π)3
Z
1
= dpχ (|p| ≤ pT F (q)) eip·x
(2π)3
1/2 1/3
where we defined the Thomas-Fermi momentum pT F (q) = cT F ρT F (q) and we used the Thomas-
Fermi equation. We conclude that G(q, x) = p3T F (q)G0 (pT F (q)x), with
Z
1 ip·x 1
G0 (x) = dpe = [sin |x| − |x| cos |x|]
(2π)3 |p|≤1 2π 2 |x|3
In particular, the density of PSC is given by
X Z Z
ρPSC (x) = (PSC )σ,σ (x, x) = 2 dqG(q, 0)|g(x−q)| = ρT F (q)|g(x−q)|2 dq = (ρT F ∗|g|2 )(x)
2

σ
We also note that the exchange term associated with PSC takes the form
2
XZ |(PSC )σ,σ′ (x, y)|2
Z Z
1
dxdy = 2 dxdy dq G(q, x − y)g(x − q)g(y − q)
σ,σ ′
|x − y| |x − y|
Z Z
1
= 2 drds dqG(q, r)g(s − q + r/2)g(s − q − r/2)
|r|
Z Z
dr
=2 ds |(G ∗1 f )(s, r)|2
|r|
where f (r, s) = g(s + r/2)g(s − r/2) and G ∗1 f denotes the convolution on the first variable
only. Young’s inequality implies that
Z Z  Z 2 Z
2 2
ds |(G ∗1 f )(s, r)| ≤ ds|G(s, r)| ds|f (s, r)| ≤ ds|G(s, r)|2

since ∥g∥2 = 1. We find

XZ |(PSC )σ,σ′ (x, y)|2
Z
dr
dxdy =2 ds|G(s, r)|2
σ,σ ′
|x − y| |r|
Z
1 1
= 4 drds 7 (sin(pT F (s)|r|) − pT F (s)|r| cos(pT F (s)|r|))2
4π |r|
∞ 2 Z
c2T F
 Z
1 2
= 3 dr 5 (sin r − r cos r) dsρT F (s)4/3
π r
Z 0
4/3
= cD ρT F (x)

140
with the Dirac constant cD = 2c2T F /(2π)3 .
It will also be useful to introduce the operator
Z
1 X
hSC = dpdqh(p, q)Πp,q,σ
(2π)3 σ

A simple computation shows that

Z XZ
1 X 2 2
3
dpdqp |⟨φ, fpqσ ⟩| = dx|∇φσ (x)|2 + ∥∇g∥2 ∥φ∥2
(2π) σ σ

and that
Z XZ
1 X 2
dpdqϕT F (q) |⟨φ, fpqσ ⟩| = dx(ϕT F ∗ |g|2 )(x)|φσ (x)|2
(2π)3 σ σ

This implies that

hSC = −∆ + ϕT F ∗ |g|2 + ∥∇g∥2 = h − ϕT F − ϕT F ∗ |g|2 + ∥∇g∥2

 
(5.126)

with the Hamilton operator

h = −∆ + ϕT F (x)
appearing in Dirac-Schwinger theory. For future reference, we note that
2/5
∥ϕT F − ϕT F ∗ |g|2 ∥5/2 ≤ constZ∥|x|1/2 |g|2 ∥1 (5.127)

To prove this inequality, we observe that

Z
|x| − |.| ∗ |g| ≤ dy|g(y)|2 |x|−1 − |x − y|−1
−1 −1 2

Z 2/5 Z 3/5
−1 −1 5/2 2 2
≤ dy |x| − |x − y| |g(y)| dy|g(y)|

Hence
Z Z
−1 −1 5/2 5/2
∥|x| − |.| ∗ |g|2 ∥5/2 ≤ dx dy |x|−1 − |x − y|−1 |g(y)|2
Z Z 
5/2
= dy dx |x|−1 − |x − y|−1 |g(y)|2

= const · ∥|.|1/2 |g|2 ∥1

because, by scaling, Z
5/2
dx |x|−1 − |x − y|−1 = const · |y|1/2

141
We conclude that
2/5
∥ϕT F − ϕT F ∗ |g|2 ∥5/2 ≤ const · (Z + ∥ρT F ∥1 )∥|.|1/2 |g|2 ∥1

as claimed.
Finally, we will need some bound for the L5/3 -norm of the densities ρT F , ρSC and for the
density ργ of many-body states with low energy. We begin with an estimate for ∥ρT F ∥5/3 . To
this end, we notice that
Z Z Z
1 1 1
ρ(x)dx = ρ(x)dx + ρ(x)dx
|x − R| |x−R|>λ |x − R| |x−R|≤λ |x − R|
Z 2/5
−1 −5/2
≤ λ ∥ρ∥1 + |x| dx ∥ρ∥5/3
|x|≤λ
−1
≤ λ ∥ρ∥1 + Cλ1/5 ∥ρ∥5/3

Optimizing over λ, we find

Z
1 1/6 5/6
ρ(x)dx ≤ C∥ρ∥1 ∥ρ∥5/3
|x − R|

Hence, any ρ with ET F (ρ) ≤ 0 satisfies

h i
5/3 1/6 5/6 5/6 5/6
0 ≥ ET F (ρ) ≥ C∥ρ∥5/3 − CZ∥ρ∥1 ∥ρ∥5/3 ≥ C∥ρ∥5/3 ∥ρ∥5/3 − Z 7/6

5/3 5/3
and therefore ∥ρ∥5/3 ≤ CZ 7/3 . In particular, this implies that ∥ρT F ∥5/3 ≤ CZ 7/3 . By Young’s
inequality it follows also that
5/3 5/3 5/3
∥ρSC ∥5/3 = ∥ρT F ∗ |g|2 ∥5/3 ≤ C∥ρT F ∥5/3 ≤ CZ 7/3

Finally, consider a normalized N -particle wave function in L2a (R3N ; C2N ) such that ⟨ψ, HN ψ⟩ ≤
0. Then we have, using the Lieb-Thirring inequality,
* N
+ k Z
X X 1
0 ≥ ⟨ψ, HN ψ⟩ ≥ ψ, −∆xj ψ − zi ργ (x)dx
j=1 i=1
|x − Ri |
5/3 5/6
≥ C∥ργ ∥5/3 − CZN 1/6 ∥ργ ∥5/3

5/3
Again, it follows that ∥ργ ∥5/3 ≤ CZ 7/3 for all N -particle states with negative energy expecta-
tion (since we consider N ≤ Z).

142
5.5 Proof of Theorem 5.1.4
We proceed now with the proof of Theorem 5.1.4. We show first the lower bound on the
quantum energy in terms of the Dirac Schwinger model. Applying Lemma 5.3.3 with ρ0 = ρT F
and P = PSC , we find (with ρ = ργ + ρT F + ρSC ) that
* N
+
X
⟨ψ, HN ψ⟩ ≥ ψ, (−∆xj + V (xj ))ψ + 2D(ρT F , ργ ) − D(ρT F , ρT F )
j=1

1
Z
|P (x, y)|2 5/6 1/6+ε
(5.128)
− dxdy − C∥ρ∥5/3 ∥ρ∥1 δ(γ, PSC )1/3−ε
2 |x − y|
≥ ⟨ψ, HDS ψ⟩ − CZ 7/6 Z 1/6+ε δ(γ, PSC )1/3−ε

for any ψ ∈ L2a (R3N , C2N ) with ⟨ψ, HN ψ⟩ ≤ 0 (to use the bounds for the L5/3 norm of ργ ).
We note that with the trivial bound

δ(γ, PSC ) = trγ(1 − PSC ) ≤ tr γ = Z

we already obtain that

⟨ψ, HN ψ⟩ ≥ ⟨ψ, HDS ψ⟩ − CZ 5/3 (5.129)
Any improvement on the trivial bound for δ(γ, PSC ) will imply the desired estimate. The
reason we insist on doing better than Z 5/3 is that we want to show that the Dirac-Schwinger
and later the Hartree-Fock theory correctly describe the exchange energy, which is given by
Z
4/3 1/2 5/6
ρSC (x)dx ≤ C∥ρSC ∥1 ∥ρSC ∥5/3 ≤ CZ 1/2 Z 7/6 = CZ 5/3

and is exactly of the size Z 5/3 .

Starting from (5.129) and recalling the definition of HDS , we obtain

⟨ψ, HN ψ⟩ ≥ tr hγ − D(ρT F , ρT F ) − const · Z 5/3

 
2/5
≥ tr hSC γ − D(ρT F , ρT F ) − C Z 5/3 + Z 12/5 ∥|x|1/2 |g|2 ∥1 + Z∥∇g∥22

where we used (5.126) and (5.127) (and the fact that |tr(ϕT F − ϕT F ∗ |g|2 )γ| ≤ ∥ϕT F − ϕT F ∗
|g|2 ∥5/2 ∥ργ ∥5/3 ). Now, we scale the function g. For λ > 0, we set g(x) = λ−3/2 g0 (x/λ), for a
smooth and even g0 with ∥g0 ∥2 = 1. Then we have
2/5 2/5
∥|.|1/2 |g|2 ∥1 = λ1/5 ∥|.|1/2 |g0 |2 ∥1

and
∥∇g∥22 = λ−2 ∥∇g0 ∥22

143
We find
⟨ψ, HN ψ⟩ ≥ tr hSC γ − D(ρT F , ρT F ) − C Z 5/3 + λ1/5 Z 12/5 + Zλ−2

To minimize, we choose λ = Z −7/11 and we find

⟨ψ, HN ψ⟩ ≥ tr hSC γ − D(ρT F , ρT F ) − CZ 7/3−2/33
(5.130)
≥ tr hSC γ − EC − CZ 7/3−2/33 + ET F
recalling that EC = ET F + D(ρT F , ρT F ). For states with energy close to the ground state
energy, we have ⟨ψ, HN ψ⟩ ≃ ET F . Hence, we may expect that trhSC γ − EC is small for these
states; this should imply that γ ≃ PSC and therefore that δ(PSC , γ) is small, exactly what
we want to show. To make his argument more precise, we use Feynman-Hellmann technique,
considering
tr hSC γ − EC − αδ(γ, PSC )
= tr(hSC − α(1 − PSC ))γ − EC
Z
1 X
= dpdq [h(p, q) + αχ(h(p, q) ≤ µ)] trΠp,q,σ γ
(2π)3 σ
Z
2
− dpdq [h(p, q) + α] χ(h(p, q) ≤ µ)
(2π)3
Z
1 X
= dpdq [h(p, q) − µ + αχ(h(p, q) ≤ µ)] trΠp,q,σ γ
(2π)3 σ
Z
2
− dpdq [h(p, q) − µ + α] χ(h(p, q) ≤ µ)
(2π)3
Next, we observe that trΠp,q,σ γ = ⟨fpqσ , γfpqσ ⟩ ≤ 1. We conclude that
Z
1 X
dpdq [h(p, q) − µ + αχ(h(p, q) ≤ µ)] trΠp,q,σ γ
(2π)3 σ
 Z 
1
≥ inf dpdq [h(p, q) − µ + αχ(h(p, q) ≤ µ)] M (p, q) : 0 ≤ M (p, q) ≤ 2
(2π)3
Z
2
= dpdq [h(p, q) − µ + αχ(h(p, q) ≤ µ)] χ (h(p, q) − µ + αχ(h(p, q) ≤ µ) ≤ 0)
(2π)3
Z
2
= dpdq [h(p, q) − µ + αχ(h(p, q) ≤ µ)] χ(h(p, q) ≤ µ − α)
(2π)3
for any α > 0 (a moment thought shows that, for χ(h(p, q) − µ + αχ(h(p, q) ≤ µ) ≤ 0) =
χ(h(p, q) − µ + α+ ≤ 0)). Therefore
trhSC γ − EC − αδ(γ, PSC )
Z
2
≥− dpdq [h(p, q) − µ + α] χ(−α ≤ h(p, q) − µ ≤ 0)
(2π)3

144
For −α ≤ h(p, q) − µ ≤ 0, we find
Z α
h(p, q) − µ + α = χ(h(p, q) − µ + α′ ≥ 0) dα′
0

Hence
trhSC γ − EC − αδ(γ, PSC )
Z Z α
2
≥− dpdqχ(−α ≤ h(p, q) − µ ≤ 0) dα′ χ(h(p, q) − µ + α′ ≥ 0)
(2π)3 0
Z α Z
2 ′
=− 3
dα dpdqχ(−α′ ≤ h(p, q) − µ ≤ 0)
(2π) 0
Z α Z
1 ′
h
3/2 ′ 3/2
i
=− 2 dα dq (µ − ϕT F (q))+ − (µ − α − ϕT F (q))+
3π 0
Z α Z α′ Z
1 ′ ′′ 1/2
=− 2 dα dα dq(µ − ϕT F (q) − α′′ )+
2π 0 0

We claim that Z
1/2
dq (µ − ϕT F (q) − α′′ )+ ≤ C(α′′ )−1/4 (5.131)

To prove (5.131) we need an upper bound for (µ − α′′ − ϕT F (q))+ . To this end, we define

ϕj (x) = c3T F (3/π)2 |x − Rj |−4

Pk
and ϕ(x) = j=1 ϕj (x). For x ̸= Rj , we notice that

1 −3/2 3/2
∆ϕj = cT F ϕj
4π
This easily implies that
1 −3/2
∆ϕ ≤ cT F ϕ3/2
4π
Let ψ = ϕT F + µ − ϕ and B = {x ∈ R3 : ψ(x) > 0}. Since ϕT F (x) → −∞ for x → Rj while
ϕ(x) → +∞ for x → Rj , we have Rj ̸∈ B for all j = 1, . . . , k. Since ψ is continuous away
from the nuclei positions Rj , B must be open. For x ∈ B, we have

1 1 −3/2 3/2
− ∆ψ = ∆ϕ − ρT F ≤ cT F (ϕ3/2 − (ϕT F + µ)+ ) ≤ 0
4π 4π
Hence ψ is subharmonic on B. Since moreover ψ|∂B = 0 and lim|x|→∞ ψ(x) = µ ≤ 0, it follows
that ψ ≤ 0 on B, and therefore that B = ∅. This argument proves that ϕT F (x) + µ ≤ ϕ(x)

145
for all x ∈ R3 . Therefore
Z Z
1/2 1/2
dx (ϕT F + µ − α)+ ≤ dx (ϕ − α)+
Z k
!1/2
X
= dx ϕj (x) − k −1 α
j=1 +
Z k
!1/2
X
≤ dx (ϕj (x) − k −1 α)+
j=1
k Z
X
1/2
≤ dx ϕj (x) − k −1 α +
j=1
Z

1/2
=k dx c3T F (3/π)2 |x|−4 − k −1 α +
= Cα−1/4

by a simple scaling argument. This shows (5.131).

Integrating (5.131) twice over α, we conclude that

trhSC γ − EC − αδ(γ, PSC ) ≥ Cα7/4

for all α > 0. Optimizing over α, we find

trhSC γ − EC ≥ Cδ(γ, PSC )7/3

Inserting in (5.130), we obtain

⟨ψ, HN ψ⟩ ≥ ET F + Cδ(γ, PSC )7/3 − CZ 7/3−2/33

It follows that every many body wave function with ⟨ψ, HN ψ⟩ ≤ ET F + CZ 7/3−2/33 has a
one-particle density matrix γ satisfying

δ(γ, PSC ) ≤ CZ 1−2/77

Note that the existence of many-body wave functions with ⟨ψ, HN ψ⟩ ≤ ET F + CZ 7/3−2/33
Q TF
follows because |EN − EN | ≤ CZ 7/3−2/33 , as proven in the first part of this class (where it
was proven that Thomas-Fermi theory approximate many body quantum mechanics).
Inserting in (5.128), we conclude that any ψ with ⟨ψ, HN ψ⟩ ≤ ET F + CZ 7/3−2/33 satisfies

⟨ψ, HN ψ⟩ ≥ ⟨ψ, HDS ψ⟩ − CZ 5/3−ε ≥ EN

DS
− CZ 5/3−ε
Q DS
for a sufficiently small ε > 0. Hence EN ≥ EN − CZ 5/3−ε , as claimed.

146
 We prove now the first bound in Theorem 5.1.4, giving a lower bound for the quantum
mechanical energy in terms of the Hartree-Fock energy. Using Lemma 5.3.3 with ρ0 = ργ and
P = γ we conclude that

⟨ψ, HN ψ⟩ ≥ EHF (γ) − CZ 4/3+ε δ(γ, γ)1/3−ε

for any N -particle wave function ψ ∈ L2a (R3N ; C2N ) satisfying ⟨ψ, HN ψ⟩ ≤ 0. The factor
δ(γ, γ) can be bounded by

δ(γ, γ) = trγ(1 − γ)
1/2 1/2
= trPSC γ(1 − γ)PSC + tr(1 − PSC )1/2 γ(1 − γ)(1 − PSC )1/2
1/2 1/2
≤ trPSC (1 − γ)PSC + tr(1 − PSC )1/2 γ(1 − PSC )1/2
= δ(γ, PSC ) + δ(PSC , γ)

But
tr(1 − γ)PSC = trPSC − trγPSC = trγ − trγPSC = trγ(1 − PSC )
implies that δ(γ, PSC ) = δ(PSC , γ) ≤ CZ 1−2/77 , for all ψ ∈ L2a (R3N ; C2N ) with ⟨ψ, HN ψ⟩ ≤
ET F + CZ 7/3−2/33 . This shows that

⟨ψ, HN ψ⟩ ≥ EHF (γ) − CZ 5/3−ε ≥ EN

HF
− CZ 5/3−ε (5.132)

after appropriate redefinition of ε. Here we used Lieb’s variational principle (Corollary 5.1.3),
stating that the Hartree-Fock ground state energy is given by the infimum of EHF (ω) over all
0 ≤ ω ≤ 1 with trω = N . Since (5.132) holds for all ψ ∈ L2a (R3N ; C2N ) with ⟨ψ, HN ψ⟩ ≤
ET F + CZ 7/3−2/33 , we conclude that
Q HF
EN ≥ EN − CZ 5/3−ε

as claimed. This concludes the proof of Theorem 5.1.4.

147