Inequalities for Schrödinger Operators

and Applications to the Stability of

Matter Problem∗
Robert Seiringer
Department of Physics
Princeton University

May 21, 2009

Abstract
Notes of lectures given at “Entropy and the Quantum”, a school on
analytic and functional inequalities with applications, Tucson, Arizona,
March 16–20, 2009.

Contents
1 Uncertainty Principles in Quantum Mechanics 2
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Hardy’s Inequality . . . . . . . . . . . . . . . . . . . . . 3
1.3 Sobolev Inequalities . . . . . . . . . . . . . . . . . . . . 7
1.4 Consequences for Schrödinger Operators . . . . . . . . . 9

2 Lieb-Thirring Inequalities 10
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 The Semiclassical Approximation . . . . . . . . . . . . . 11
2.3 The Sharp Constants . . . . . . . . . . . . . . . . . . . . 11
2.4 The Birman-Schwinger Principle . . . . . . . . . . . . . 12
2.5 Possible Extensions . . . . . . . . . . . . . . . . . . . . . 14
2.6 Kinetic Energy Inequalities . . . . . . . . . . . . . . . . 15
∗
c 2009 by the author. This paper may be reproduced, in its entirety, for non-

commercial purposes.

1
3 Application: The Stability of Matter 16
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Baxter’s electrostatic inequality . . . . . . . . . . . . . . 17
3.3 Proof of Stability of Matter . . . . . . . . . . . . . . . . 20
3.4 Further Results on Stability of Matter . . . . . . . . . . 21

4 Hardy-Lieb-Thirring Inequalities 21
4.1 Hardy Inequalities with Remainder . . . . . . . . . . . . 22
4.2 Consequences: Hardy-Sobolev and Hardy-Lieb-Thirring
Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Application: Stability of Relativistic Matter . . . . . . . 24

References 25

Acknowledgments. Many thanks to Rupert Frank for his help in prepar-

ing both the lectures and these notes. This work was partially sup-
ported by U.S. National Science Foundation grant PHY 06052356.

1 Uncertainty Principles in Quantum Me-

chanics
1.1 Introduction
One of the most important differences between quantum and classical
mechanics is the uncertainty principle. Among many other things,
it implies that position and momentum of a particle can not simulta-
neously take on definite values. To make this more quantitative, recall
that the state of a quantum system is described by a wave func-
tion ψ, which is a function in L2 (Rd ), the space of square integrable
complex-valued function on Rd . For a particle of mass m, the kinetic
energy equals
Z
~2 ~2
(ψ, −∆ ψ) = |∇ψ(x)|2 dx ,
2m 2m Rd
where ~ denotes Planck’s constant. For convenience, we will set m =
1/2 and ~ = 1 in the following. We will also often write −∆ = p2 , with
p = −i∇.
The potential energy in a potential V (x) is
Z
(ψ, V ψ) = V (x)|ψ(x)|2 dx .
Rd

Historically, the most famous uncertainty principle is Heisenberg’s:

For ψ ∈ L2 (Rd ) with (ψ, ψ) = 1,

d2
−1
ψ, p2 ψ ≥ ψ, x2 ψ , (1)
4

2
with equality if and only if ψ is a centered Gaussian. This inequality
says that ψ can not be localized too close around the origin without
the kinetic energy being big. The proof of (1) can be found in any
standard textbook on quantum mechanics, and we shall not repeat it
here. It uses the fact that [p · a, x · b] = −ia · b for vectors a, b ∈ Cd .
Heisenberg’s uncertainty principle is not very useful in practice,
however. Specifically, while a small value of (ψ, x2 ψ) means that ψ is
localized close to the origin, a large value (ψ, x2 ψ) does not mean it is
spread out. In fact, (ψ, x2 ψ) could be huge even if most of the mass
of ψ is localized around the origin, if only the remaining small part of
the mass is far away.
A more useful way to quantify the localization properties of ψ
around some point (the origin, say) is via Hardy’s and Sobolev’s in-
equality, which we shall discuss next.

1.2 Hardy’s Inequality

Hardy’s inequality looks very similar to Heisenberg’s given in Eq. (1)
above, with one important difference: On the right side, on has the
expectation value of the inverse of x2 instead of the inverse of the
expectation value of x2 . More precisely, the following holds:
Theorem 1 (Hardy’s inequality). For d ≥ 3,
 
2

(d − 2)2 1
ψ, p ψ ≥ ψ, 2 ψ . (2)
4 x

The theorem also holds for d = 1 if ψ is required to vanish suitably

at the origin. With this assumption, it trivially also holds for d = 2,
of course. Note that (ψ, x−2 ψ) ≥ (ψ, x2 ψ)−1 if (ψ, ψ) = 1, by an
application of Jensen’s inequality and convexity of x 7→ x−2 .
The constant (d − 2)2 /4 in (2) is sharp in the sense that the in-
equality with a larger constant is false for some ψ.

Proof. We ignore some technical details and assume ψ to be sufficiently

smooth. We can then write ψ(x) = |x|1−d/2 g(x) with g having the
property that g(0) = 0, for otherwise both sides of the inequality will
be infinite. After taking the derivative of ψ and squaring the resulting
expression, we obtain

(d − 2)2 |ψ(x)|2
|∇ψ(x)|2 = + |x|2−d |∇g(x)|2
4 x2
∂
+ (2 − d)ℜ |x|1−d g(x) g(x) . (3)
∂|x|

3
The last term vanishes after integration over x, since
∂ 1 ∂
ℜ g(x) g(x) = |g(x)|2
∂|x| 2 ∂|x|

and hence the radial integral over |x| vanishes at every fixed angle.
The second term on the right side of (3) is strictly positive, and hence
we arrive at the desired result.

What the proof really showed is that

(d − 2)2
−∆ − = |x|d/2−1 ∇|x|2−d ∇|x|d/2−1 . (4)
4x2
The term on the right side is positive, which gives Hardy’s inequality.
It is in fact strictly positive, in the sense that it does not have a zero
eigenvalue. It annihilates the function |x|1−d/2 , but this is not a square-
integrable function. In particular, (2) is strict for any ψ that is not
identically zero.
The fact that the constant (d − 2)2 /4 is sharp can also be easily
seen from (4). For a (otherwise smooth) function that diverges as
|x|1−d/2 at the origin the expectation value of |x|−2 is infinite while
the expectation value of the right side of (4) is finite.

In a way similar to Theorem 1, one can in fact prove that

Z  p Z
d−p |ψ(x)|p
|∇ψ(x)|p dx ≥ dx (5)
Rd p Rd |x|p

for any 1 ≤ p < d. One writes ψ(x) = |x|(1−d/p) g(x) and uses the
convexity inequality

|a + b|p ≥ |a|p + p|a|p−2 ℜ a · b

for vectors a, b ∈ Cd .

In order to take effects of special relativity

p into account, it is use-
ful to consider the kinetic energy to be p2 + m2 instead of p2 /(2m).
By definition,
 p  Z p
ψ, −∆ + m2 ψ = b
|ψ(k)|2
|2πk|2 + m2 dk
Rd

where ψb is the Fourier transform of ψ, i.e.,

Z
b
ψ(k) = ψ(x) e−2πik·x dx .
Rd

4
The operator p = −i∇ thus acts as multiplication by 2πk in momentum
space. p
Note
p that p2 + m2 ≈ m + p2 /(2m) for small |p| (or large m),
2 2
while p + m ≈ |p| for large |p| (or small m). In any case,
p
|p| ≤ p2 + m2 ≤ |p| + m ,

and for the questions of stability discussed in the following sections

one might as well set m = 0. This is sometimes referred to as the
‘ultra-relativistic’ limit.
In the case m = 0, the relativistic kinetic energy has the following
nice double integral representation:
Lemma 1 (Integral Representation of Relativistic Kinetic En-
ergy).
 √  Γ((d + 1)/2) Z |ψ(x) − ψ(y)|2
ψ, −∆ ψ = dxdy (6)
2π (d+1)/2 R2d |x − y|d+1

Proof. Write |p| = limt→0 t−1 (1 − e−t|p| ) and use the fact that e−t|p|
has the integral kernel

Γ((d + 1)/2) t
e−t|p| (x, y) = (d+1)/2
.
π (t + |x − y|2 )(d+1)/2
2

This leads immediately to (6).

The analogue of Hardy’s inequality (2) in the relativistic case is

the following. It is also sometimes referred to as Kato’s inequality [16,
Eq. (V.5.33)].
Theorem 2 (Relativistic Hardy Inequality). For d ≥ 2
 √   
Γ((d + 1)/4)2 1
ψ, −∆ ψ ≥ 2 ψ, ψ (7)
Γ((d − 1)/4)2 |x|

As in the non-relativistic case in Theorem 1, the constant in (7) is

sharp. Note that for d = 3 it equals 2/π.

Proof. For simplicity, we restrict our attention to the case d = 3, which

is the most relevant case in view of applications in physics. For general
d, the proof works the same way, but the integrals involved are slightly
more complicated. See [14, 30, 11].
According to Lemma 1,
Z
1 |ψ(x) − ψ(y)|2
(ψ, |p| ψ) = 2 dx dy ,
2π R6 |x − y|4

5
which we can write as
Z
1 |ψ(x) − ψ(y)|2
(ψ, |p| ψ) = lim 2 dx dy . (8)
ǫ→0 2π R6 (|x − y|2 + ǫ)2
The purpose of ǫ is to avoid the singularity at x = y. The limit ǫ → 0
will be taken at the end of the calculation.
We can write

||x|ψ(x) − |y|ψ(y)|2
|ψ(x) − ψ(y)|2 =
|x| |y|
   
2 |x| 2 |y|
+ |ψ(x)| 1 − + |ψ(y)| 1 − .
|y| |x|
Hence the integral in (8) can be written as
Z
||x|ψ(x) − |y|ψ(y)|2 dx dy
R6 (|x − y|2 + ǫ)2 |x| |y|
Z  
|ψ(x)|2 |x|
+2 1− dx dy .
R6 (|x − y|2 + ǫ)2 |y|
After performing the y angular integration and denoting t = |y|, the
second integral equals
Z Z ∞
t(t − |x|)
8π dx |ψ(x)|2 dt .
R3 0 (|x|2 − t2 )2 + 2ǫ(|x|2 + t2 ) + ǫ2
The t integral is non-negative and converges, as ǫ → 0, to the principle-
value integral Z ∞
t(t − |x|) 1
p.v. 2 2 2
dt = .
0 (|x| − t ) 2|x|
It is tedious but elementary to justify exchanging the x integration and
the ǫ → 0 limit.
We conclude that
Z
2 1
(ψ, |p| ψ) − |ψ(x)|2 dx
π R3 |x|
Z
1 ||x|ψ(x) − |y|ψ(y)|2 dx dy
= lim 2 .
ǫ→0 2π R6 (|x − y|2 + ǫ)2 |x| |y|
Since the right side is positive, this proves (7).

The proofs of Theorems 2 and 1 are very similar. The main idea
is to write ψ(x) = f (x)g(x), with f (x) the formal solution of the
corresponding variational equation, which is f (x) = |x|1−d/2 in the
non-relativistic case and f (x) = |x|(1−d)/2 in the relativistic case. This
procedure is sometimes referred to as ground state substitution.

6
 The Hardy inequalities discussed above can be generalized in vari-
ous ways. One is to kinetic energies of the form (−∆)s for s > 0, and
this generalization is straightforward. More involved is the Lp general-
ization (which is (5) for s = 1), in which case one considers expressions
of the form Z
|ψ(x) − ψ(y)|p
dx dy
R2d |x − y|d+ps
for general 0 < s < 1 and 1 ≤ p < d/s. The sharp constants for the
corresponding Hardy inequalities have only been found very recently
in [12].

1.3 Sobolev Inequalities

An alternative way to quantify the uncertainty principle in quantum
mechanics is Sobolev’s inequality. Recall the definition of the Lp (Rd )
norms Z  1/p
kψkp = |ψ(x)|p dx
Rd
for 1 ≤ p < ∞, and

kψk∞ = ess sup |ψ(x)| .

Theorem 3 (Sobolev’s Inequality). For d ≥ 3

(ψ, −∆ ψ) ≥ Sd kψk22d/(d−2) (9)

with Sd = d(d − 2)|Sd |2/d /4. For d = 2 one has, for some S2,p > 0,
−4/(p−2)
(ψ, −∆ ψ) ≥ S2,p kψk2 kψkp2p/(p−2) for all 2 < p < ∞

while for d = 1
(ψ, −∆ ψ) ≥ kψk−2 4
2 kψk∞ .

For d = 1 and d ≥ 3 the constants are sharp, while for d = 2

the value of the optimal constant S2,p is unknown. We shall skip
the proof of this theorem here, and refer the interested reader to [19].
Appropriate Sobolev inequalities hold also for fractional powers of −∆,
of course. In the relativistic case, the following holds.
Theorem 4 (Relativistic Sobolev Inequality). For d ≥ 2,
 √ 
ψ, −∆ ψ ≥ Sed kψk22d/(d−1) (10)

with Sed = (d − 1)|Sd |1/d /2. For d = 1 one has, for some Se1,p > 0,
 √ 
−2/(p−2)
ψ, −∆ ψ ≥ Se1,p kψk2 kψkp2p/(p−2) for all 2 < p < ∞.

7
 Using the powerful tool of symmetric-decreasing rearrange-
ment (see, e.g., [19]), one can actually derive Sobolev’s inequalities
from Hardy’s, except for the value of the optimal constants. In this
sense, Hardy’s inequality is stronger than Sobolev’s. The argument
goes as follows [12].
For a radial, decreasing function ψ,
Z
ψ(x)p dx = kψkpp ≥ ψ(y)p |y|d |Bd |
Rd

for any y ∈ Rd . Now take this to the power 1 − 2/p, multiply by

|ψ(y)|2 |y|d(2/p−1) and integrate over y. This gives
Z
|ψ(y)|2
d(1−2/p)
dx ≥ |Bd |1−2/p kψk2p .
Rd |y|

Hence Sobolev’s inequality follows from Hardy’s for radial, decreasing

functions.
To extend this result to arbitrary functions, the crucial observation
is that (ψ, −∆ψ) goes down under symmetric-decreasing rearrange-
ment of ψ, while kψkp stays the same. Therefore it suffices to prove
Sobolev’s inequality for symmetric-decreasing functions.
In order to define symmetric-decreasing rearrangement, note first
that for any measurable function ψ,
Z ∞
|ψ(x)| = χ{|ψ|>t} (x)dt
0

where χ{|ψ|>t} denotes the characteristic function of the set where

|ψ| > t, i.e., (
1 if |ψ(x)| > t
χ{|ψ|>t} (x) =
0 if |ψ(x)| ≤ t .
For a general set A ⊂ Rd , we denote by χ∗A the characteristic function of
a ball of volume |A| centered at the origin. The symmetric-decreasing
rearrangement of ψ, denoted by ψ ∗ , is defined as
Z ∞
∗
ψ (x) = χ∗{|ψ|>t} (x)dt .
0

Note that ψ ∗ is clearly symmetric-decreasing. From the definition, it

is also obvious that kψ ∗ kp = kψkp for any p since the rearrangement
does not change the values of ψ, only the places where these values are
taken. What is less obvious is that



ψ ∗ , p2 ψ ∗ ≤ ψ, p2 ψ (11)
and we refer to [19] for its proof. Inequality (11) also holds with p2
replaced by |p|, and hence the argument just given also applies to the
relativistic case.

8
1.4 Consequences for Schrödinger Operators
If V is a (real-valued) potential that goes to zero at infinity, the spec-
trum of −∆ + V consists of discrete points in (−∞, 0] and the contin-
uum [0, ∞). The infimum of the spectrum is called the ground state
energy, and it is determined by the variational principle

E0 = inf (ψ, (−∆ + V ) ψ) .

kψk2 =1

We shall investigate the question for what potentials V the ground

state energy is finite.
Recall Hölder’s inequality, which states that
Z
f (x)g(x)dx ≤ kf kp kgkq for 1 ≤ p ≤ ∞, 1/p + 1/q = 1.
Rd

By combining this with Sobolev’s inequality (9) we see that, for d ≥ 3

Sd
(ψ, −∆ ψ) ≥ (ψ, |V | ψ) .
kV kd/2

In particular, −∆ + V ≥ 0 if kV kd/2 ≤ Sd . That is, E0 = 0 in this

case.
More generally, if V ∈ Ld/2 + L∞ then V can be written as V (x) =
w(x) + u(x) with kwkd/2 ≤ Sd and u bounded, and hence E0 > −∞.
We leave the demonstration of this fact as an exercise. We can proceed
in a similar way for d ≤ 2 and conclude that E0 is finite if

 d/2 ∞
L + L if d ≥ 3
1+ε
V ∈ L + L∞ if d = 2 (12)

 1 ∞
L +L if d = 1.
√
Similarly, in the relativistic case, the ground state energy of −∆ + V
is finite if (
Ld + L∞ if d ≥ 2
V ∈ (13)
L1+ε + L∞ if d = 1.

As an example, consider the Coulomb potential V (x) = −|x|−1 in

d = 3. Clearly V ∈ L3/2 + L∞ , which explains the stability of the
hydrogen atom with non-relativistic kinematics. The relativistic case
3
is borderline, however, since V just fails + L∞ . We have in
√to be in L −1
fact seen in Subsection 1.2 above that −∆ − λ|x| is bounded from
below (in fact, positive) only if λ ≤ 2/π.
As this example shows, one can deduce from Hardy’s inequality
that the singularities in V can actually be slightly stronger than in
(12) and (13) above for E0 to be finite: In the non-relativistic case

9
E0 > −∞ if V (x)P≥ −(d − 2)2 /(4|x|2 ) − C or, more generally, if
V (x) ≥ − 14 (d − 2)2 i |x − Ri |−2 − C for finitely many distinct points
Ri 6= Rj . We leave the proof of this last statement as an exercise.
After having found conditions on V that guarantee the finiteness
of the ground state energy, we will study sums of powers of all the
negative eigenvalues of Schrödinger operators in the next section.

2 Lieb-Thirring Inequalities
2.1 Introduction
Let E0 ≤ E1 ≤ E2 . . . be the negative eigenvalues of the Schrödinger
operator −∆ + V on L2 (Rd ), with V satisfying the condition (12).
Lieb-Thirring inequalities concern bounds on the moments
X
|Ei |γ
i≥0

for some γ ≥ 0. The case γ = 0 corresponds to the number of

negative eigenvalues.
Theorem 5 (Lieb-Thirring Inequalities). The negative eigenval-
ues Ei of −∆ + V satisfy the bounds
X Z
γ+d/2
|Ei |γ ≤ Lγ,d V (x)− dx (14)
i≥0 Rd

where V (x)− = max{−V (x), 0} denotes the negative part of V . The

(sharp) values of γ ≥ 0 for which (14) holds with Lγ,d < ∞ (indepen-
dent of V ) are
• for d = 1, γ ≥ 1/2 [22, 29]
• for d = 2, γ > 0 [22]
• for d ≥ 3, γ ≥ 0 [4, 18, 25, 22]
The fact that γ > 0 is necessary for d = 2 follows from the fact
that −∆ − λV has a negative eigenvalue for arbitrarily small λ if V is
negative. For d = 1 this eigenvalue is of the order λ2 as λ → 0, from
which one easily deduces that (14) can hold only for γ ≥ 1/2 for d = 1.
The special (and most difficult) case γ = 0 for d ≥ 3 is also known
as the Cwikel-Lieb-Rosenblum [4, 18, 25] bound. Below we will
prove (14) only for γ > 0 for d ≥ 2 and γ > 1/2 for d = 1. Our proof
follows the original work by Lieb and Thirring [22].

10
2.2 The Semiclassical Approximation
Another way to write sum of the negative eigenvalues to the power γ
is X
γ
|Ei |γ = Tr − ∆ + V (x) − , (15)
i≥0

where Tr denotes the trace. A semiclassical approximation of the

trace leads to the phase space integral
ZZ Z

γ γ+d/2
(2π)−d p2 + V (x) − dp dx = Lscl
γ,d V (x)− dx (16)
Rd
Rd ×Rd

where
Z
Γ(γ + 1)
Lscl
γ,d = (2π)
−d
(1 − p2 )γ dp = .
|p|≤1 (4π)d/2 Γ(γ + 1 + d/2)

Note that the factor (2π)−d in front of the integral in (16) is really
Planck’s constant to the power −d; in our units Planck’s constant
equals 2π.
The reason for calling (16) the semiclassical approximation to (15)
is that
Z

γ γ+d/2
lim hd Tr − h2 ∆ + V (x) − = Lsclγ,d V (x)− dx (17)
h→0 Rd

under certain assumptions on the potential V . See [19, Sect. 12.12] or

[24, Sect. XIII.15]. Note that because of (17) it is necessarily true that
Lγ,d ≥ Lscl
γ,d . But when is Lγ,d < ∞, and when does it equal Lγ,d ?
scl

2.3 The Sharp Constants

As mentioned above, Lγ,d < ∞ if and only if γ ≥ 1/2 for d = 1, γ > 0
for d = 2 and γ ≥ 0 for d ≥ 3. Some facts about the sharp values for
Lγ,d are known:
• Lγ,d = Lscl
γ,d for all γ ≥ 3/2 and d ≥ 1 [22, 1, 17]

• L1/2,1 = 1/2 while Lscl

1/2,1 = 1/4 [15]

• Lγ,d > Lscl

γ,d if γ < 1 [13]
The optimal constant in the physically most interesting case, γ = 1
and d = 3, remains an open problem. It is conjectured to be L1,3 =
Lscl
1,3 . The best current bound was obtained in [6] as

π
L1,3 ≤ √ Lscl
1,3 . (18)
3
We will use this bound in Section 3 in our proof of Stability of Matter.

11
2.4 The Birman-Schwinger Principle
We shall now explain the proof of the LT inequalities (14). For sim-
plicity we restrict our attention to the non-critical cases, i.e., to γ > 0
for d ≥ 2 and γ > 1/2 for d = 1.
Without loss of generality, we assume that V (x) ≤ 0 from now on.
The eigenvalue equation

−∆ψ(x) + V (x)ψ(x) = −eψ(x)

is equivalent to Z
φ(x) = Ke (x, y)φ(y)dy
Rd
p
where φ(x) = −V (x)ψ(x) and Ke is the Birman-Schwinger ker-
nel p p
Ke (x, y) = −V (x)Ge (x − y) −V (y)
with Z
1
Ge (x − y) = e2πik·(x−y) dk .
Rd (2πk)2 + e
√ √
In one dimension, the√ function Ge equals Ge (x) = e− e|x| /(2 e),
while it is Ge (x) = e− e|x| /(4π|x|) for d = 3.
In other words, −∆ + V having an eigenvalue −e < 0 is equivalent
to Ke having an eigenvalue 1; also the multiplicities coincide. This
fact is known as the Birman-Schwinger principle. The study of
the negative eigenvalues of −∆ + V thus reduces to a study of the
spectrum of the family Ke of compact and positive operators.
From the definition it is obvious that Ke is monotone decreasing
in e. Hence all the eigenvalues λi of Ke are monotone decreasing in e.
These are sketched in Figure 1.
From the figure we easily deduce the fact that the number of
eigenvalues of Ke that are ≥ 1 equals the number of eigenval-
ues of −∆ + V that are ≤ −e ! We shall call this number Ne . For
any m > 0 we thus have the simply but important inequality

Ne ≤ Tr (Ke )m . (19)

Since V is assumed to be non-positive, we can write Ke as the product

Ke = |V |1/2 Ge |V |1/2 , with |V |1/2 a multiplication operator and Ge
a convolution operator. It is a fact that for any two non-negative
operators A and B and any m ≥ 1,

Tr(B 1/2 AB 1/2 )m ≤ Tr B m/2 Am B m/2 . (20)

For the proof for general m ≥ 1 we refer the reader to [22, 26, 2], [20] or
to the lecture notes by E. Carlen in this volume. For m = 2, however,

12
 λ

λ0
1
λ1
λ2

λ3
λ4
e

e2 e1 e0

Figure 1: Sketch of the eigenvalues of the Birman-Schwinger op-

erator Ke as a function of e. An eigenvalue −e of the Schrödinger
operator −∆ + V is equivalent to Ke having an eigenvalue 1.

the proof is very simple. Since AB − BA is anti-hermitian, its square

is non-positive, i.e., (AB − BA)2 ≤ 0. Using this and the cyclicity of
the trace one concludes that

Tr ABAB ≤ Tr A2 B 2 .

The case m = 2 is actually the one needed to prove the LT inequality

in the physically relevant case γ = 1 and d = 3, as we shall see below.
If one is only interested in this special case, one does not need the
general bound (20).
From (19) and (20) we conclude that, for m ≥ 1,

Ne ≤ Tr |V |m/2 (Ge )m |V |m/2

Z Z
1
= 2 m
dk |V (x)|m dx .
Rd ((2πk) + e) Rd

The k integral is finite if m > d/2. In fact, it is given by

Z
1 Γ(m − d/2)
2 m
dk = e−m+d/2 =: e−m+d/2 Cd,m .
Rd ((2πk) + e) (4π)−d/2 Γ(m)

13
Hence we have the upper bound
Z
Ne ≤ Cd,m e−m+d/2 |V (x)|m dx
Rd

for any m satisfying the conditions m ≥ 1 and m > d/2.

To obtain information on the sum of powers of negative eigenvalues
of −∆ + V , note that
X Z ∞
|Ei |γ = γ eγ−1 Ne de . (21)
i≥0 0

Using the above bound on Ne , the e integral diverges, however, either

at 0 or at ∞. As a way out, consider

We (x) = max{−V (x) − e/2, 0} .

Then
Ne (V ) = Ne/2 (V + e/2) ≤ Ne/2 (−We )
and hence
X Z Z −2V (x)
|Ei |γ ≤ γCd,m eγ−1−m+d/2 (−V (x) − e/2)m de dx
i≥0 Rd 0
Z
′
= Cγ,d,m |V (x)|γ+d/2 dx .
Rd

′
For Cγ,d,m to be finite we need d/2 < m < γ + d/2, i.e., γ > 0 and
γ > 1/2 for d = 1. A possible choice is m = (γ + d)/2. This completes
the proof.

2.5 Possible Extensions

LT inequalities are known for a more general class of operators. Ex-
tensions that are important for applications in physics are:
• Magnetic fields: One can replace −∆ by −(∇ − iA(x))2 for a
real-valued vector-potential A (whose curl is the magnetic field).
Recall the diamagnetic inequality, which states that


ψ, −(∇ − iA(x))2 ψ ≥ (|ψ|, −∆ |ψ|) .

Hence the lowest eigenvalue E0 always goes up when a vector field

is introduced, but not necessarily the sum of powers of the eigen-
values. Hence the extension of the LT inequalities to magnetic
fields is non-trivial.

14
 √
• Fractional Schrödinger operators: One can use −∆ instead
of −∆ for the kinetic energy [5]. The appropriate Lp -norm of the
potential is determined by semiclassics as
√ γ Z
Tr −∆ + V ≤ Kγ,d V (x)γ+d
− dx .
−

A similar result holds for (−∆)s for arbitrary s > 0 and appro-
priate γ.

2.6 Kinetic Energy Inequalities

For N particles satisfying Fermi-Dirac statistics (e.g., electrons),
the wave functions have to be antisymmetric, i.e.,

ψ(x1 , . . . , xi , . . . , xj , . . . , xN ) = −ψ(x1 , . . . , xj , . . . , xi , . . . , xN )

for every 1 ≤ i 6= j ≤ N . (We ignore spin for simplicity.) We leave it

as an exercise to show that for any such ψ with (ψ, ψ) = 1,
 X  NX
−1
N
ψ, [−∆j + V (xj )] ψ ≥ Ei (22)
j=1
i=0

with Ej the negative eigenvalues of −∆ + V . The LT inequality for

γ = 1 implies that the latter sum is bounded from below by
Z
1+d/2
−L1,d V (x)− dx . (23)
Rd

For a given (antisymmetric) ψ, let ̺ψ denote its one-particle den-

sity, i.e.,
Z
̺ψ (x) = N |ψ(x, x2 , . . . , xN )|2 dx2 · · · dxN .
Rd(N −1)

If we choose
V (x) = −c ̺ψ (x)2/d
for some c > 0, we conclude from (22)–(23) that
 XN   Z
ψ, − ∆j ψ ≥ c − L1,d c1+d/2 ̺ψ (x)1+2/d dx .
j=1 Rd

To make the right side as large as possible, the optimal choice of c is

c = [L1,d (1 + d/2)]−2/d . This yields
 XN   2/d Z
d 2
ψ, − ∆j ψ ≥ ̺ψ (x)1+2/d dx .
j=1 d+2 L1,d (d + 2) Rd
(24)

15
 Inequality (24) can be viewed as an uncertainty principle for
many-particle systems. We emphasize that the antisymmetry of ψ is
essential for (24) to hold with an N -independent constant on the right
side. For general ψ, (24) holds only if the right side is multiplied by
N −2/d .
Inequality (24) is equivalent to the LT inequality (14) for γ = 1
in the sense that validity of (24) for all antisymmetric ψ implies (14)
with the corresponding constant L1,d . The demonstration of this fact
is left as an exercise.
The right side of (24), with L1,d replaced by Lscl
1,d , is just the semi-
classical approximation to the kinetic energy of a many-body sys-
tem. To see this, let us calculate the sum of the lowest N eigenvalues of
the Laplacian on a cube of side length ℓ. For large N , boundary condi-
tions are irrelevant, and hence we can use periodic boundary conditions
in which case the eigenvalues of −∆ are just (2πk)2 with k ∈ (Z/ℓ)d .
Replacing sums by integrals the sum of the lowest N eigenvalues is
Z
(2π)2 ℓd |k|2 dk (25)
|k|≤µ

with µ determined by
Z
ℓd dk = N .
|k|≤µ

A simple computation thus shows that for this value of µ (25) equals
!2/d  
1+2/d
d 2 d N
ℓ . (26)
d + 2 Lscl1,d (d + 2) ℓd
In a semiclassical approximation, one can estimate the lowest energy
of a system with particle density ̺(x) by (26) replacing N/ℓd by ̺(x)
and integrating over x instead of multiplying by ℓd . One indeed arrives
at the right side of (24) this way, except for the prefactor.

3 Application: The Stability of Matter

3.1 Introduction
Ordinary matter composed of electrons and nuclei is described by the
Hamiltonian
N
X N X
X M
Z
H =− ∆i −
i=1 i=1 j=1
|xi − Rj |
X 1 X Z2
+ + . (27)
|xi − xj | |Rk − Rl |
1≤i<j≤N 1≤k<l≤M

16
The nuclei have charge Z and are located at positions Rj ∈ R3 , j =
1, . . . , M , which are treated as parameters. The electron coordinates
are xi ∈ R3 , i = 1, . . . , N , and ∆i denotes the Laplacian with respect
to xi . Since elections are fermions, the wave functions ψ(x1 , . . . , xN )
are antisymmetric functions of xi ∈ R3 . (For simplicity, we ignore
spin here.)
Let E0 (N, M ) denote the ground state energy of H, i.e.,

E0 (N, M ) = inf inf (ψ, H ψ) .

{Rj } (ψ,ψ)=1

Note that besides minimizing over electron wavefunctions ψ, we also

minimize over all nuclear coordinates Rj , j = 1, . . . , M . Everyday
experience tells us that

E0 (N, M ) ≥ −c(N + M )

for some c > 0. If not, you would not be reading this notes! If the
energy would decrease faster than linearly with the particle number,
bulk matter would not be stable but rather implode; adding two half-
filled glasses of water would release a huge amount of energy.
Why is the energy bounded from below by N + M ? After all, there
are (N + M )2 terms in the Hamiltonian, N M of which are negative. In
other words, why is matter stable? The antisymmetry of the allowed
wave functions ψ turns out to be crucial. Without it, E0 (N, M ) would
decease as −C min{N, M }5/3 for large particle number.
Stability of Matter was first proved by Dyson and Lenard in 1967
[7]. In 1975, Lieb and Thirring [21] gave a very elegant proof, using the
kinetic energy inequalities derived in Section 2.6 and the “no-binding
theorem”in Thomas-Fermi theory.
We shall follow a different route to prove Stability of Matter (due
to Solovej [27]) which is probably the shortest. One uses
• The Lieb-Thirring inequality for d = 3 and γ = 1.
• Baxter’s electrostatic inequality
In the former, the antisymmetry of the wave functions enters. The
latter is a pointwise bound on the total Coulomb potential and has
nothing to do with quantum mechanics. We shall explain it next.

3.2 Baxter’s electrostatic inequality

Baxter proved the following in 1980 [3].
Theorem 6 (Baxter’s Electrostatic Inequality). For any xi , Rj ∈

17
R3 , i = 1, . . . , N , j = 1, . . . M , and any Z > 0,
N X
X M X X
Z 1 Z2
− + +
i=1 j=1
|xi − Rj | |xi − xj | |Rk − Rl |
1≤i<j≤N 1≤k<l≤M
N
X 2Z + 1
≥− (28)
i=1
D(xi )

where D(x) = minj |x−Rj | denotes the distance to the nearest nucleus.
This inequality quantifies electrostatic screening. Effectively, an
electron sees only the nearest nucleus. The coefficient 2Z + 1 on the
right side is not sharp. Inequality (6) was later improved by Lieb and
Yau [23] to yield the sharp value Z for x close to one of the nuclei.
The proof of Theorem 6 given here does not follow the original
one in [3]. We shall instead follow a suggestion by Solovej (private
communication) and deduce it from a Lemma by Lieb and Yau [23].
The argument uses essentially only two things:
• The Coulomb potential 1/|x| has positive Fourier transform
• Newton’s theorem, which states that
Z Z Z
f (x) 1 f (x)
dx = f (x)dx + dx
R 3 |x − y| |y| |x|≤|y| |x|≥|y| |x|

for any radial function f .

Proof of Theorem 6. For a measure µ(dx), let D(µ) denote the elec-
trostatic energy
Z
1 1
D(µ) = µ(dx)µ(dy) .
2 R6 |x − y|

Note that D(µ) ≥ 0 even if µ is not positive; this follows from the fact
the |x|−1 has a positive Fourier transform. Let

X M
Z Z
φ(x) = − + (29)
D(x) j=1 |x − Rj |

denote the Coulomb attraction to all but the nearest nuclei. We shall
first show that
Z X Z2
D(µ) − φ(x)µ(dx) + ≥0 (30)
|Rk − Rl |
1≤k<l≤M

for any measure µ.

18
 The function φ is harmonic, i.e., ∆φ = 0, except on the surfaces
{x : |x − Rk | = |x − Rj | for some k 6= j}. An explicit computation
shows that φ is superharmonic on all of R3 , i.e.,

−∆φ(x) = 4π ν(dx) (31)

for some non-negative measure ν which is supported on these surfaces.

Using D(µ − ν) ≥ 0 we have
Z
D(µ) − φ(x)µ(dx) = D(µ − ν) − D(ν) ≥ −D(ν) .

It remains to calculate D(ν). Using (29) and (31) it is not difficult to

see that
Z
1
D(ν) = φ(x)ν(dx)
2
X Z
Z2 1 Z
= − ν(dx) .
|Rk − Rl | 2 R3 D(x)
1≤k<l≤M

Since ν ≥ 0 the last expression is negative and can be dropped for an

upper bound. This proves (30).
For simplicity, denote di = D(xi ), and let µi (dx) be the normalized
uniform measure supported on a sphere of radius di /2 centered at xi ,
i.e., µi (dx) = (d2i π)−1 δ(|x − xi | − di /2)dx. We shall use (30) with µ =
PN
i=1 µi . If we replace the electron point charges by the smeared out
spherical charges µi , the electrostatic interaction among the electrons
is reduced because the interaction energy between two spheres is less
than or equal to that between two points. This follows from Newton’s
Theorem, which also implies that the interaction between the smeared
electrons and the nuclei is not changed, since the radius di /2 is less
than the distance of xi to any of the nuclei. Hence

X XX N M
1 Z
−
|xi − xj | i=1 j=1 |xi − Rj |
1≤i<j≤N

X Z XM Z
1 Z
≥ µi (dx)µj (dy) − µ(dx)
R6 |x − y| j=1 R
3 |x − Rj |
1≤i<j≤N
M Z
X XN
Z 1
= D(µ) − µ(dx) − .
j=1 R3 |x − Rj | d
i=1 i

We have used that D(µi ) = 1/di . An application of (30) yields the

19
lower bound
N X
X M X X
Z 1 Z2
− + +
i=1 j=1
|xi − Rj | |xi − xj | |Rk − Rl |
1≤i<j≤N 1≤k<l≤M
N Z
X 
Z 1
≥− µi (dx) + .
i=1 R3 D(x) di

To finish the proof it suffices to show that for x in the support of µi ,

we have D(x) = minj |x − Rj | ≥ di /2. This follows from the triangle
inequality, which implies that for any k

|x − Rk | ≥ |xi − Rk | − |xi − x| ≥ di − di /2 = di /2 .

In the last step, we used that |xi −Rk | ≥ di by definition, and |xi −x| =
di /2 for x on the sphere centered at xi .

3.3 Proof of Stability of Matter

The proof of Stability of Matter for the Hamiltonian (27) is an easy
consequence of (22), Theorem 5 for d = 3 and γ = 1, as well as
Theorem 6.
Using Baxter’s electrostatic inequality (28), the Hamiltonian (27)
is bounded from below by
N 
X 
2Z + 1
H≥ −∆i − .
i=1
D(xi )

The right side is just a sum of independent one-particle terms! For an-
tisymmetric functions, it is bounded below by the sum of the lowest
N eigenvalues of −∆ − (2Z + 1)/D(x). We can pick a µ > 0 and use
the LT inequality (14) for d = 3 and γ = 1 to obtain
Z  5/2
2Z + 1
H ≥ −µN − L1,3 µ− dx .
R3 D(x) −

The integral is bounded by

Z  5/2 Z  5/2
2Z + 1 2Z + 1
µ− dx ≤ M µ− dx
R3 D(x) − R3 |x| −
2
5π (2Z + 1)3
=M √ .
4 µ

We can now optimize over µ. If we use, in addition, the bound (18) on

the LT constant L1,3 , we have completed the proof of

20
Theorem 7 (Stability of Matter). On the space of antisymmetric
functions, the Hamiltonian (27) is bounded from below by

π 2/3
H≥− (2Z + 1)2 M 2/3 N 1/3 .
4
Since M 2/3 N 1/3 ≤ N/3 + 2M/3, this yields the desired linear lower
bound on the ground state energy.

3.4 Further Results on Stability of Matter

Since the work of Lieb and Thirring, stability of matter has been proved
for many more models, including
√
• Relativistic kinematics, where −∆ replaces −∆ in the ki-
netic energy
• Magnetic fields and their interaction with the electron spin
• models with kinetic energy described by the Dirac operator
• certain approximations to quantum electrodynamics, where
the electromagnetic field is quantized.
For details and a pedagogical presentation of this subject, we refer the
reader to [20].

4 Hardy-Lieb-Thirring Inequalities
Recall Hardy’s inequalities (2) and (7). For general fractional pow-
ers of the Laplacian, they read
 

ψ, (−∆)s − Cs,d |x|−2s ψ ≥ 0 (32)

for 0 < s < d/2 and ψ ∈ L2 (Rd ). The sharp constant Cs,d is known to
be [14]
Γ((d + 2s)/4)2
Cs,d = 22s
Γ((d − 2s)/4)2
but there is no optimizer for this inequality. The variational equation
is satisfied for ψ(x) = |x|s−d/2 .

Questions:
• Can (32) be improved by adding a Sobolev term kψk2q on the
right side?
• Do Lieb-Thirring inequalities hold for (−∆)s − Cs,d |x|−2s − V (in
terms of kV kq alone)?

21
Although the second question seems stronger than the first, these two
are actually intimately related. Both answers are Yes, as was proved
in [11]. A simpler proof was obtained by Frank in [9], using an elegant
Lemma of Solovej, Sørensen, and Spitzer [28]. We will follow the latter
route here.

4.1 Hardy Inequalities with Remainder

The following theorem is the key to understanding the answers to the
questions just raised.
Theorem 8 (Hardy with Remainder). For 0 < t < s < d/2, there
exists a κd,s,t > 0 such that
 

s/t 2(1−s/t)
ψ, (−∆)s − Cs,d |x|−2s ψ ≥ κd,s,t ψ, (−∆)t ψ kψk2 (33)

for all ψ ∈ L2 (Rd ).

The proof of this theorem can be found in [9]. In the special case
d = 3 and s = 1/2 it was obtained earlier in [28]. We only give a sketch
of the proof here, and refer to [9, 28] for details.

Sketch of Proof. In terms of the Fourier transform, the Hardy poten-

tial can be written as
Z Z b ψ(q)
b
Γ(d/2 − s) ψ(k)
|ψ(x)|2 |x|−2s dx = π 2s−d/2 dk dq .
Rd Γ(s) R2d |k − q|d−2s

A simple Schwarz inequality shows that, for any positive function h on

Rd ,
Z b ψ(q)
b Z
ψ(k) b 2
d−2s
dk dq ≤ th (k)|ψ(k)| dk ,
R 2d |k − q| R d

where Z
1 h(q)
th (k) = dq .
h(k) Rd |k − q|d−2s
Hardy’s inequalities (32) follow from this by choosing h(k) = |k|−s−d/2 .
The improvement (33) can be obtained by choosing instead h(k) =
|k|−s−d/2 (1 + λ|k|2(t−s) )−1 for λ > 0 and optimizing over the value of
λ.

4.2 Consequences: Hardy-Sobolev and Hardy-Lieb-

Thirring Inequalities
We can now combine Theorem 8 with the Sobolev inequality


ψ, (−∆)t ψ ≥ St,d kψk2q , q = 2d/(d − 2t)

22
which is the generalization of (9)–(10) to general fractional Laplacians.
This yields
Corollary 1 (Hardy-Sobolev Inequalities). For ψ ∈ L2 (Rd ), 0 <
s < d/2 and 2 < q < 2∗ := 2d/(d − 2s),
 

ψ, (−∆)s − Cs,d |x|−2s ψ ≥ Sd,s,q kψk2(1+σ)
q kψk2−2σ

with σ = 2(1 − q/2∗ )/(q − 2) > 0.

The value of σ is determined by scale invariance. Note that q is
assumed to be strictly smaller than the critical Sobolev exponent
2∗ = 2d/(d − 2s). One can, in fact, show that Sd,s,q → 0 as q → 2∗ .

Given the fact that Lieb-Thirring inequalities hold for (−∆)t + V ,

the improved Hardy inequality of Theorem 8 implies
Corollary 2 (Hardy-Lieb-Thirring Inequalities). For some con-
stants 0 < Mγ,d,s < ∞,
Z
s −2s

γ γ+d/2s
Tr (−∆) − Cs,d |x| + V (x) − ≤ Mγ,d,s V (x)− dx (34)
Rd

for γ > 0 and 0 < s < d/2.

A sketch of the proof of Corollary 2 is given below. The special
case s = 1 for d ≥ 3 was proved earlier in [8].
It is not difficult to see that γ > 0 is necessary. An arbitrarily
small non-positive potential V will always create a bound state of the
operator (−∆)s − Cs,d |x|−2s + V (x) !
In order to prove (34), one uses again (21) to reduce the question
to bounding the number of eigenvalues less or equal to e. One then
uses Theorem 8 together with the simple convexity bound

s/t 2(1−s/t)
ψ, (−∆)t ψ kψk2
 1−t/s
s e
e
≥ ψ, (−∆)t ψ − kψk22
t 4(s/t − 1) 4

and proceeds as in the proof of Theorem 5. We refer to [9] for the

details.
For s ≤ 1, the inequalities (34) also hold in the magnetic case, i.e.,
with (−∆)s replaced by |∇−iA(x)|2s , with constants independent of A
[11, 9]. This is relevant for applications to (pseudo-)relativistic models
in physics, as we shall discuss below.

23
4.3 Application: Stability of Relativistic Matter
As in Section 3, we now consider a quantum-mechanical model of N
electrons and M fixed nuclei. Besides the electrostatic Coulomb in-
teraction, the electrons interact with an arbitrary external magnetic
field B = curl A, in which case p has to be replaced by p − A in the
kinetic energy. To take effects of special relativity into account,
we shall also replace the kinetic energy operator −(∇ − iA)2 by the
corresponding (pseudo-)relativistic expression
q
−(∇j − iA)2 + m2 − m .

Note that, for large m, this reduces to −(∇ − iA)2 /(2m) to leading
order.
The resulting many-particle Hamiltonian is then
N q
X 
HN,M = −(∇j − iA(xj ))2 + m2 − m + αVN,M ,
j=1

with VN,M the total electrostatic potential,

VN,M (x1 , . . . , xN ; R1 , . . . RM )
X XX N M X
1 Z Z2
= − + .
|xi − xj | j=1 |xj − Rk | |Rk − Rl |
1≤i<j≤N k=1 1≤k<l≤M

The constant α > 0 in front of VN,M is the fine structure constant,

which equals α = 1/137 in nature. The reason it did not appear in the
non-relativistic model (27) is that in the latter case it can simply be
removed by scaling. Not so in the relativistic case; in fact, from the
relativistic Hardy inequality (7) we see that necessarily Zα ≤ 2/π for
HN,M to be bounded from below, even for N = M = 1.
The Pauli exclusion principle dictates that HN,M acts on antisym-
metric functions ψ(x1 , . . . , xN ). Stability of matter means that on
this space
HN,M ≥ −const. (N + M )
with a constant that is independent of the positions Rk of the nuclei
and of the magnetic vector potential A. In the relativistic case, one
observes that, by scaling, either inf Rk ,A (inf spec HN,M ) ≥ −mN or
inf Rk ,A (inf spec HN,M ) = −∞.
The following theorem was proved in [10].
Theorem 9 (Stability of Relativistic Matter with Magnetic
Fields). Assume that Zα ≤ 2/π and α ≤ 1/133. Then (on the space
of antisymmetric wave functions)
HN,M ≥ −mN

24
for all values of the nuclear coordinates Rj and the magnetic vector
potential A.
Theorem 9 was proved by Lieb and Yau [23] in the non-magnetic
case A = 0, but the more general case of non-zero A was open. One
of the key ingredients in the proof in [10] is the Hardy-Lieb-Thirring
inequality (34) (more precisely, its generalization to include arbitrary
magnetic vector potentials) in the case γ = 1, d = 3 and s = 1/2. In
order to get the explicit bound on the allowed values of α, which barely
covers the physical case α = 1/137, one needs a good bound on the
relevant constant M1,3,1/2 . The ones currently available are not good
enough to cover the physical case. In [10] only the special case of V
being constant inside a ball centered at the origin and +∞ outside the
ball was needed, however. In this special case, a good enough bound
on the corresponding constant was obtained.
Lieb and Yau also showed that inf spec HN,M = −∞ if α > αc for
some some αc > 0 independent of Z. Hence a bound on α is indeed
needed!

References
[1] M. Aizenman and E.H. Lieb, On Semi-Classical Bounds for
Eigenvalues of Schrödinger Operators, Phys. Lett. A 66, 427–429
(1978).
[2] H. Araki, On an inequality of Lieb and Thirring, Lett. Math.
Phys. 19, 167–170 (1990).
[3] R.J. Baxter, Inequalities for potentials of particle systems, Ill. J.
Math. 24, 645–652 (1980).
[4] M. Cwikel, Weak type estimates for singular values and the num-
ber of bound states of Schrödinger operators, Ann. of Math. 106,
93–100 (1977).
[5] I. Daubechies, An uncertainty principle for fermions with gener-
alized kinetic energy, Commun. Math. Phys. 90, 511–520 (1983).
[6] J. Dolbeault, A. Laptev, and M. Loss, Lieb-Thirring inequalities
with improved constants, J. Eur. Math. Soc. 10, 1121–1126 (2008).
[7] F.J. Dyson and A. Lenard, Stability of matter I, J. Math. Phys.
8, 423–434 (1967); II, ibid. 9, 1538–1545 (1968).
[8] T. Ekholm and R.L. Frank, On Lieb-Thirring Inequalities for
Schrödinger Operators with Virtual Level, Commun. Math. Phys.
264, 725–740 (2006)].
[9] R.L. Frank, A simple proof of Hardy-Lieb-Thirring inequalities,
arXiv:0809.3797, Commun. Math. Phys. (in press).

25
[10] R.L. Frank, E.H. Lieb, and R. Seiringer, Stability of Relativistic
Matter with Magnetic Fields for Nuclear Charges up to the Critical
Value, Commun. Math. Phys. 275, 479–489 (2007).
[11] R.L. Frank, E.H. Lieb, and R. Seiringer, Hardy-Lieb-Thirring in-
equalities for fractional Schrödinger operators, J. Amer. Math.
Soc. 21, 925–950 (2008).
[12] R.L. Frank and R. Seiringer, Non-linear ground state represen-
tations and sharp Hardy inequalities, J. Funct. Anal. 255, 3407–
3430 (2008).
[13] B. Helffer and D. Robert, Riesz means of bounded states and
semi-classical limit connected with a Lieb-Thirring conjecture I,
J. Asymp. Anal. 3, 91–103 (1990); II, Ann. Inst. H. Poincare 53,
139–147 (1990).
[14] I.W. Herbst, Spectral theory of the operator (p2 + m2 )1/2 − Ze2 /r,
Comm. Math. Phys. 53, 285–294 (1977).
[15] D. Hundertmark, E.H. Lieb, and L.E. Thomas, A sharp bound for
an eigenvalue moment of the one-dimensional Schrödinger oper-
ator, Adv. Theor. Math. Phys. 2, 719–731 (1998).
[16] T. Kato, Perturbation theory for linear operators, Springer (1976).
[17] A. Laptev and T. Weidl, Sharp Lieb-Thirring inequalities in high
dimensions, Acta Math. 194, 87–111 (2000); See also: Recent
results on Lieb-Thirring inequalities, Journées ”Équations aux
Dérivées Partielles” (La Chapelle sur Erdre, 2000), Exp. No. XX,
Univ. Nantes, Nantes (2000).
[18] E.H. Lieb, The Number of Bound States of One-Body Schrödinger
Operators and the Weyl Problem, Proc. Amer. Math. Soc. Sym-
posia in Pure Math. 36, 241–252 (1980).
[19] E.H. Lieb and M. Loss, Analysis, Second Edition, American Math-
ematical Society, Providence, Rhode Island (2001).
[20] E.H. Lieb and R. Seiringer, The Stability of Matter in Quantum
Mechanics, Cambridge University Press, to appear.
[21] E.H. Lieb and W. Thirring, Bound for the Kinetic Energy of
Fermions which Proves the Stability of Matter, Phys. Rev. Lett.
35, 687–689 (1975). Errata ibid., 1116 (1975).
[22] E.H. Lieb and W. Thirring, Inequalities for the Moments of the
Eigenvalues of the Schrödinger Hamiltonian and Their Relation
to Sobolev Inequalities, in Studies in Mathematical Physics, E.
Lieb, B. Simon, A. Wightman, eds., Princeton University Press,
269–303 (1976).
[23] E.H. Lieb and H.-T. Yau, The Stability and Instability of Rela-
tivistic Matter, Commun. Math. Phys. 118, 177–213 (1988).

26
[24] M. Reed and B. Simon, Methods of Modern Mathematical Physics
IV. Analysis of Operators, Academic Press (1978).
[25] G.V. Rosenbljum, Distribution of the discrete spectrum of singular
differential operators, Izv. Vyss. Ucebn. Zaved. Matematika 164,
75–86 (1976); English transl. Soviet Math. (Iz. VUZ) 20, 63–71
(1976).
[26] E. Seiler and B. Simon, Bounds in the Yukawa2 quantum field the-
ory: upper bound on the pressure, Hamiltonian bound and linear
lower bound, Commun. Math. Phys. 45, 99–114 (1975).
[27] J.P. Solovej, Stability of Matter, Encyclopedia of Mathematical
Physics, eds. J.-P. Francoise, G.L. Naber and S.T. Tsou, vol. 5,
8–14, Elsevier (2006).
[28] J. P. Solovej, T.Ø. Sørensen, and W.L. Spitzer, The relativistic
Scott correction for atoms and molecules, arXiv:0808.2163.
[29] T. Weidl, On the Lieb-Thirring constants Lγ,1 for γ ≥ 1/2, Com-
mun. Math. Phys. 178, 135–146 (1996).
[30] D. Yafaev, Sharp constants in the Hardy-Rellich inequalities, J.
Funct. Anal. 168, 121–144 (1999).

27