Density functional

approach to the
many-body problem: key
concepts and exact
functionals
Robert van Leeuwen
Theoretical Chemistry, Materials Science Centre,
Nijenborgh 4, 9747 AG, Rijksuniversiteit Groningen
Groningen, The Netherlands
Abstract
We give an overview of the fundamental concepts of density functional
theory. We give a careful discussion of the several density functionals and
their dierentiability properties. We show that for nondegenerate ground
states we can calculate the necessary functional derivatives by means of
linear response theory, but that there are some dierentiability problems
for degenerate ground states. These problems can be overcome by extending
the domains of the functionals. We further show that for every interacting
v-representable density we can nd a noninteracting v-representable density
arbitrarily close and show that this is sucient to set up a Kohn-Sham
scheme. We nally describe two systematic approaches for the construction
of density functionals.
1. Introduction
2. Denition of the problem
3. Conditions on the electron density and external potentials
4. The Hohenberg-Kohn theorem
5. Kohn-Sham theory by Legendre transforms
6. Denition of the functional derivative
7. Static linear response of the Schr odinger equation
8. Invertability of the density response function
9. Functional derivatives and v-representability
10. The Hohenberg-Kohn theorem for degenerate ground states
11. Linear response for degenerate ground states and functional
dierentiability
12. The Levy and Lieb functionals F
LL
and F
L
13. Dierentiability of F
L
14. Ensemble v-representability
15. The Kohn-Sham approach and noninteracting v-representability
16. The gradient expansion
17. The optimized potential method and the e
2
-expansion
18. Outlook and conclusions
19. Acknowledgments
References
1. Introduction
In this paper we give an overview of the foundations of density functional
theory for stationary systems. In the discussion we try to be as precise
as possible and we therefore carefully discuss several exact density func-
tionals and the domains of densities on which they are dened. At the
heart of almost any application of density functional theory lie the Kohn-
Sham equations. These equations describe a noninteracting system that
is required to produce the ground state density of an interacting system.
Therefore any discussion of the validity of these equations has to focus
on how well we can approximate a density of an interacting system with
that of a noninteracting system. Closely related to this question is whether
or not the exact functionals that we dened have functional derivatives.
This is because the various potentials, such as the Kohn-Sham potential,
in density functional theory appear as functional derivatives of the energy
with respect to the density. The investigation of the existence of functional
derivatives will form a central theme of this review.
A large part of this work will follow the proofs of the fundamental papers
by Lieb [1] and Englisch and Englisch [2, 3]. In this paper we try to make
these two important works more accessible by providing some background
on the mathematics involved. We further show, from a more physical view-
point, how to calculate functional derivatives by response theory and show
that for nondegenerate ground states the static density response function is
invertable. We also show that for every interacting v-representable density
there is a noninteracting v-representable density arbitrarily close to it and
that this is sucient to set up a Kohn-Sham scheme. We nally discuss
two systematic approaches for the construction of the exchange-correlation
functional and provide an outlook and conclusions.
2. Denition of the problem
Density-functional theory in its earliest formulation by Hohenberg, Kohn
and Sham [5, 8] aims at a description of the description of ground state
properties of many-electron systems in terms of the electron density. One
may wonder why this is possible. Let us therefore investigate this point
more closely. Consider a Hamiltonian of a stationary many-body system

H
v
=

T +

V +

W (1)
where

T is the kinetic energy of the electrons,

V the external potential,
and

W the two-particle interaction. We denote the Hamilton operator

H
v
with a subindex v to indicate that we will consider the Hamiltonian as a
functional of the external potential v. The constituent terms are explicitly
given as

T =
N

i=1

1
2

2
i
(2)

V =
N

i=1
v(r
i
) (3)

W =
N

i>j
w([r
i
r
j
[) (4)
where w([r[) = 1/[r[ will in our discussion always be the Coulomb potential.
We are interested in electronic systems, i.e. molecules and solids. For all
these systems the kinetic energy operator

T and two-particle interaction

W are identical. They only dier in the form of the external potential v(r)
and the number of electrons N. The properties of all these systems can
therefore be regarded as a functional of the external potential v. This is in
particular the case for the ground state wavefunction [[v]) and the ground
state energy E[v], which are of course related by the Schrodinger equation
(

T +

V +

W)[[v]) = E[v][[v]) (5)
At rst sight we have not gained much by this viewpoint, the problem looks
as dicult as before. However, the problem will have a dierent appearance
once we eliminate the potential in favor of the electron density. One may
wonder what is so particular about the relation between the density and
the potential. Let us therefore look specically at the term which describes
the external potential. It can be written as

V =
_
d
3
rv(r) n(r) (6)
where we dened the density operator by
n(r) =
N

i=1
(r r
i
) (7)
The expectation value of

V is given by
[

V [) =
_
d
3
rn(r)v(r) (8)
where n(r) the electron density . The electron density is obtained from
the many-body wavefunction, which in our case will usually be the ground
state wavefunction of Hamiltonian

H
v
, by
n(r
1
) = [ n(r
1
)[) = N

1...N
_
d
3
r
2
. . . d
3
r
N
[(r
1

1
, . . . , r
N

N
)[
2
(9)
In this expression
i
is the spin variable for electron i. The physical inter-
pretation of the density is that n(r)V is proportional to the probability
of nding an electron in an innitesimal volume V around r. Note that
in the denition Eq.(9) we sum over the spin coordinate
1
so we do not
distinguish between the probabilities of nding up or down spin electrons
at r. We see from Eq.(8) that the external potential v(r) and the electron
density n(r) are conjugate variables. With this we mean that they occur
as a simple product in the contribution of the external potential to the
energy. It is exactly this property that allows us to prove the Hohenberg-
Kohn theorem which establishes a 1-1-correspondence between the density
and the external potential. We can therefore go from a functional E[v] of
the external potential to a functional E[n] of the density. We can ask our-
selves whether there is any reason that E[n] would be easier to calculate
than E[v]. A priori there is no reason to expect this. However, we will
see that the determination of E[n] is equivalent to the solution of a set of
one-particle equations, known as the Kohn-Sham equations, with a poten-
tial v
s
[n] that is also a functional of the density. By now we know that
we can nd practical and useful approximations for this potential v
s
[n] and
that the Kohn-Sham equations have been succesfully applied to the calcu-
lation of properties of many electronic systems. However, as in any physical
theory, there are a number of assumptions made in the transformation to
the Kohn-Sham one-particle equations. The aim of this paper is to discuss
these assumptions and thereby the validity of the Kohn-Sham equations.
3. Conditions on the electron density and the external
potentials
We begin by giving a discussion of the properties of the two key objects in
density functional theory, the density n and the external potential v. The
density has the obvious properties
n(r) 0 ,
_
d
3
rn(r) = N (10)
These properties follow directly from the denition of the density and the
usual normalization condition on the wavefunction. If we take into account
that the density is obtained as the density of a bound eigenstate of Hamil-
tonian (1) we can derive further conditions. For this we put the physical
constraint on the many-body system that it has a nite expectation value
of the kinetic energy, i.e.
T[] =
1
2
N

i=1

1...N
_
d
3
r
1
. . . d
3
r
N
[
i
[
2
< (11)
At this point it is useful to introduce a new space of functions. We say that
a function f is in H
1
(
n
) ( denotes the real numbers) if
|f| =
__
d
n
r([f(r)[
2
+[f(r)[
2
_
1/2
< (12)
The space of functions H
1
(
n
) is called a Sobolev space. The supindex 1
refers to the fact that the denition of the norm contains only rst order
derivatives. We therefore see that niteness of the kinetic energy implies
that is an element of the function space H
1
(
3N
). Dierentiation of
Eq.(9) and use of the Cauchy-Schwarz inequality then leads to [1]
(
1
n(r
1
))
2
4Nn(r
1
)

1...N
_
d
3
r
2
. . . d
3
r
N
[
1
[
2
(13)
and hence
1
2
_
d
3
r(
_
n(r))
2
T[] < (14)
We see that the niteness of the kinetic energy puts a constraint on the
density. From Eq.(12) we see that

n belongs to H
1
(
3
). We also see that
if we consider systems with a nite kinetic energy, then we only need to
consider the following set of densities
o = n [ n(r) 0,

n H
1
(
3
),
_
d
3
rn(r) = N (15)
This set of densities has a property which will be of importance later,
namely o is convex. With this we mean that if n
1
and n
2
are elements of
o, then also n
1
+ (1 )n
2
is an element of o where 0 1. This
property is easily proven using the Cauchy-Schwarz inequality. Now we will
derive constraints on the allowed set of external potentials. In order to do
this we introduce some other function spaces. We say that a given function
f belongs to the space L
p
if
|f|
p
=
__
d
3
r[f(r)[
p
_
1/p
< (16)
Note that here we only consider functions on the usual threedimensional
coordinate space
3
. The letter L refers to Lebesque integration, a feature
that assures that the function spaces are complete (complete normed spaces
are also called Banach spaces). We will, however, not go into the detailed
mathematics and refer the interested reader to the literature [4]. We just
note that for continuous functions the integral is equivalent to the usual
(Riemann) integral. Equation (16) denes a norm on the space L
p
and
we see from Eq.(10) that the density belongs to L
1
. From the condition
of nite kinetic energy and the use of a Sobolev inequality one can show
that [1]
_
d
3
rn
3
(r) C
_
d
3
r(
_
n(r))
2
< (17)
where C = 3(/2)
4/3
. In other words, the niteness of the kinetic energy
implies that the density is also in the space L
3
. Since we already know
that n L
1
we nd that the density is element of the intersection of both
spaces, i.e. n L
1
L
3
. We also see from the inequality (17) that o is a
subset of L
1
L
3
. Let us now see what this implies for the allowed set of
external potentials when we require the expectation value of the external
potential to be nite, i.e. we require

_
d
3
rn(r)v(r)

|nv|
1
< (18)
If the potential is bounded, i.e. [v(r)[ < M for some nite number M then

_
d
3
rn(r)v(r)

sup[v(r)[N < (19)

The space of bounded functions is called L

and has the norm

|f|

= sup[f(r)[ (20)
The supremum is dened to be the smallest number M such that [f(r)[ M
almost everywhere. The term almost everywhere has a precise mathemat-
ical meaning for which we refer to the literature [4]. We almost never use
it in the remainder of this paper. We therefore conclude that if v L

then the expectation value of the external potential is nite. To show this
we used that n L
1
. But we also know that n L
3
and if we make use of
the Holder inequality
|fg|
1
|f|
p
|g|
q
(21)
with 1/p + 1/q = 1 we obtain
|nv|
1
|n|
3
|v|3
2
(22)
which is nite if v L
3
2
. Therefore the most general set of potentials for
which the expectation value [

V [) is nite, is the set

L
3
2
+L

= v[v = u +w, u L
3
2
, w L

(23)
i.e. the set of potentials that can be written as a sum of a function from L
3
2
and a function from L

. This is also a normed function space with norm

|v| = inf|u|3
2
+|w|

[v = u +w (24)
In the remainder of this paper we will always consider the densities to be in
the space L
1
L
3
and the potentials in the space L
3
2
+L

. It is important
to note that the Coulomb potential is in the latter set since we can write
1
[r[
=
(1 [r[)
[r[
+
([r[ 1)
[r[
(25)
where is the Heaviside function, (x) = 0 if x 0 and (x) = 1 if
x > 0. One can readily check that the rst and the second term on the
right hand side are in L
3
2
and L

respectively. One can easily extend this

result to a nite sum of Coulomb potentials and therefore every molecule
can be described with the potentials in the space L
3
2
+L

. One may nally

wonder what the condition of nite Coulombic electron-electron repulsion,
i.e. [

W[) < , would imply for the density. However, one can show
that if the kinetic energy [

T[) is nite then also [

W[) is nite [1],
so this does not yield any new constraints on the density.
4. The Hohenberg-Kohn theorem
The basis of density-functional theory is provided by the Hohenberg-Kohn
theorem [5]. We will provide a proof of this theorem for nondegenerate
ground states. The case of degenerate ground states will be discussed later.
The Hohenberg-Kohn theorem states that the density n(r) of a nondegen-
erate ground state uniquely determines the external potential v(r) up to an
arbitrary constant. This means that the external potential is a well-dened
functional v[n](r) of the density.
In the proof essential use is made of the fact that the density and the po-
tential are conjugate variables. For the same reason we can, for instance,
prove that the 2-particle interaction is a unique functional of the diago-
nal 2-particle density matrix. The general mapping between N-particle
density matrices and N-body potentials is discussed by De Dominicis and
Martin [6].
Let us consider the subset 1 L
3
2
+L

of potentials that yield a normaliz-

able nondegenerate ground state. The solution of the Schrodinger equation
provides us with a mapping from the external potential to the ground state
wavefunction, v(r) [[v]). Since we assume that we are dealing with
nondegenerate ground states [[v]) is uniquely determined apart from a
trivial phase factor. We have therefore established a map C : 1 ,
where is the set of ground states.
We will rst prove that the map C is invertible. Suppose that [
1
) and
[
2
) H
1
(
3N
) correspond to external potentials v
1
and v
2
L

+ L
3
2
where v
1
,= v
2
+C. We have to show that [
1
) ,= [
2
). If we assume that
[
1
) = [
2
) = [) then by subtraction of the Hamiltonian for [
1
) and
[
2
) we nd that
(

V
1

V
2
)[) = (E
1
E
2
)[) (26)
If v
1
v
2
is not constant in some region then must vanish in this region
for the above equation to be true. However if v
1
, v
2
L

+ L
3
2
then [)
cannot vanish on an open set (a set with nonzero measure) by the unique
continuation theorem [1]. So we obtain a contradiction and hence we must
have made a wrong assumption. Therefore [
1
) ,= [
2
) and we obtain the
result that dierent potentials ( diering more than a constant) give dier-
ent wavefunctions. Consequently we nd that the map C is invertible.
We now dene the set / as the set of densities which come from a nonde-
generate ground state, where we only consider ground state densities from
potentials in the set L

+L
3
2
. The set / is obviously a subset of the previ-
ously dened set o. From a given wavefunction in the set of ground states
we can calculate the density according to Eq.(9). This provides us with
a second map D : / from ground state wavefunctions to ground state
densities. Also this map is invertible. To show this we calculate
E[v
1
] = [v
1
][

T +

V
1
+

W[[v
1
]) < [v
2
][

T +

V
1
+

W[[v
2
])
= [v
2
][

T +

V
2
+

W[[v
2
]) +
_
n
2
(r)(v
1
(r) v
2
(r))dr
= E[v
2
] +
_
n
2
(r)(v
1
(r) v
2
(r))dr (27)
Likewise we nd
E[v
2
] < E[v
1
] +
_
n
1
(r)(v
2
(r) v
1
(r))dr (28)
Adding both inequalities then yields the new inequality
_
d
3
r(n
2
(r) n
1
(r))(v
2
(r) v
1
(r)) < 0 (29)
If we assume that n
1
= n
2
then we obtain the contradiction 0 < 0 and we
conclude that dierent ground states must yield dierent densities. There-
fore the map D is also invertible. Consequently the map DC : 1 / is
also invertible and the density therefore uniquely determines the external
potential. This proves the Hohenberg-Kohn theorem.
Let us now pick an arbitrary density out of the set / of densities of non-
degenerate ground states. The Hohenberg-Kohn theorem then tells us that
there is a unique external potential v (to within a constant) and a unique
ground state wavefunction [[n]) (to within a phase factor) corresponding
to this density. This also means that the ground state expectation value of
any observable, represented by an operator

O, can be regarded as a density
functional
O[n] = [n][

O[[n]) (30)
In particular we can thus dene the Hohenberg-Kohn functional F
HK
on
the set / as
F
HK
[n] = [n][

T +

W[[n]) (31)
With this functional we can dene the energy functional E
v
as
E
v
[n] =
_
n(r)v(r)dr +F
HK
[n] (32)
If n
0
is a ground state density corresponding to external potential v
0
and
n an arbitrary other ground state density then
E
v0
[n] =
_
n(r)v
0
(r)dr +F
HK
[n] = [n][

T +

V
0
+

W[[n])
[n
0
][

T +

V
0
+

W[[n
0
]) = E
v0
[n
0
] (33)
Therefore
E[v] = inf
nA
__
n(r)v(r)dr +F
HK
[n]
_
(34)
and we see that the ground state energy of a many-body system can be
obtained by minimization of a density functional. For application of this
formula we have to know F
HK
on the set /. In practice it is, of course,
impossible to calculate F
HK
exactly on this set of densities. Instead one
would prefer to make an explicit approximation for F
HK
as is usually done
within the Kohn-Sham scheme. Before we go into that let us rst discuss
some properties of F
HK
. The functional F
HK
is a convex functional, i.e. if
n
1
, n
2
/ and if
1
n
1
+
2
n
2
/ with 0
1
,
2
1 and
1
+
2
= 1
then
F
HK
[
1
n
1
+
2
n
2
]
1
F
HK
[n
1
] +
2
F
HK
[n
2
] (35)
This is readily proven. Suppose that the ground state densities n
1
, n
2
,
1
n
1
+

2
n
2
/ correspond to the external potentials v
1
, v
2
and v. Then
F
HK
[n] = [n][

T +

V +

W[[n])
_
d
3
rn(r)v(r)
=
1
[n][

H[[n]) +
2
[n][

H[[n])
_
d
3
rn(r)v(r)

1
[n
1
][

T +

W[[n
1
]) +
2
[n
2
][

T +

W[[n
2
])
+
_
(
1
n
1
(r) +
2
n
2
(r))v(r)d
3
r
_
d
3
rn(r)v(r)
=
1
F
HK
[n
1
] +
2
F
HK
[n
2
] (36)
and we obtain the convexity of F
HK
. Note, however, that the domain /
of F
HK
does not need to be convex, i.e. if n
1
, n
2
/ then not necessarily

1
n
1
+
2
n
2
/ with 0
1
,
2
1 and
1
+
2
= 1. We will come back
to this point later when we consider the dierentiability of F
HK
. Let us
rst collect our results in the form a theorem
Theorem 1 (Hohenberg-Kohn) The density n corresponding to a non-
degenerate ground state species the external potential v up to a constant
and the ground state wavefunction [[n]) up to a phase factor. Moreover,
1. Any ground state expectation value corresponding to an observable

O
is a functional of the density according to
O[n] = [n][

O[[n])
2. The ground state energy of a system with a nondegenerate ground
state and an external potential v can be obtained from
E[v] = inf
nA
__
n(r)v(r)dr +F
HK
[n]
_
where F
HK
[n] = [n][

T +

W[[n]).
3. F
HK
is convex
We nally discuss an interesting consequence of the Hohenberg-Kohn the-
orem. Suppose that in Eq.(29) we take v
2
= v
1
+ v where v is not a
constant. We therefore obtain potential v
2
by a small perturbation from
potential v
1
. By means of perturbation theory we can then calculate the
change in the density which gives
n
2
(r) = n
1
(r) +n(r) +O(
2
) (37)
where n can be calculated from the static density response function
n(r) =
_
d
3
r

(r, r

)v(r

) (38)
The properties and explicit form of are described in more detail in a later
section. Therefore
(v
2
(r) v
1
(r))(n
2
(r) n
1
(r)) =
2
v(r)n(r) +O(
3
) (39)
If we insert this expression in Eq.(29) and divide by
2
> 0 then we obtain:
_
d
3
rv(r)n(r) +O() < 0 (40)
Now taking the limit 0 and expressing n in terms of the response
function we obtain
_
d
3
rd
3
r

v(r)(r, r

)v(r

) < 0 (41)
This is true for an arbitrary nonconstant potential variation. We therefore
see that the eigenvalues of , when we regard as an integral operator,
are negative. Moreover we see that the only potential variation that yields
a zero density variation is given by v = C where C is a constant. This
implies that is invertable. We will go more closely into this matter in a
later section where we will prove the same using the explicit form of the
density response function.
5. Kohn-Sham theory by Legendre transforms
The method described in this section goes back to the work of De Do-
minicis and Martin [6]. This work discusses the relations between N-body
potentials and N-particle density matrices, of which the density-potential
relation to be discussed here is a special case. The ground state energy E[v]
and wave function [[v]) are considered to be functionals of the external
potential through solving the time-independent Schrodinger equation
(

T +

V +

W)[[v]) = E[v][[v]) (42)
where the two-particle interaction

W is kept xed. From this equation we
see that the ground state energy as a functional of the external potential v
can also be written as
E[v] = [v][

H
v
[[v]) (43)
Our goal is now to go from the potential as our basic variable, to a new
variable, which will be the electron density. The deeper reason that this is
possible is that the density and the potential are conjugate variables. With
this we mean that the contribution of the external potential to the total
energy is simply an integral of the potential times the density. We make use
of this relation if we take the functional derivative of the energy functional
E[v] with respect to the potential v:
E
v(r)
=

v(r)
[

H
v
[) +[

H
v
[

v(r)
) +[


H
v
v(r)
[)
= E[v]

v(r)
[) +[ n(r)[)
= [ n(r)[) = n(r) (44)
where we used the Schrodinger equation

H
v
[) = E[v][) and the nor-
malization condition [) = 1. Note that the equation above is nothing
but a functional generalization of the well-known Hellmann-Feynman the-
orem [7]. Now we can go to the density as our basic variable by dening a
Legendre transform
F[n] = E[v]
_
d
3
rn(r)v(r) = [v][

T +

W[[v]) (45)
where v must now be regarded as a functional of n. The uniqueness of such
a mapping is garanteed by the Hohenberg-Kohn theorem [5]. The set of
densities for which the functional F[n] is dened is the set of so-called v-
representable densities. These are ground state densities for a Hamiltonian
with external potential v. The question which constraints one has to put
on a density to make sure that it is v-representable is known as the v-
representability problem. We postpone a discussion of these matters to
later sections. From E/v = n it follows immediately that
F
n(r)
= v(r) (46)
This is our rst basic relation. In order to derive the Kohn-Sham equations
we dene the following energy functional for a system of noninteracting
particles with external potential v
s
and with ground state wave function
[[v
s
]):
E
s
[v
s
] = [v
s
][

T +

V
s
[[v
s
]) (47)
with Legendre transform
F
s
[n] = E[v
s
]
_
d
3
rn(r)v
s
(r) = [v
s
][

T[[v
s
]) (48)
and derivatives
E
s
v
s
(r)
= n(r) (49)
F
s
n(r)
= v
s
(r) (50)
We see that F
s
[n] in Eq.(48) is the kinetic energy of a noninteracting system
with potential v
s
and density n. For this reason the functional F
s
is usually
denoted by T
s
. In the following we will adopt this notation. Finally we
dene the exchange-correlation functional E
xc
[n] by the equation
F[n] = T
s
[n] +
1
2
_
d
3
rd
3
r

n(r)n(r

)w([r r

[) +E
xc
[n] (51)
This equation assumes that the functionals F[n] and T
s
[n] are dened on
the same domain of densities. We thus assume that for a given ground
state density of an interacting system there is a noninteracting system with
the same density. In other words, we assume that the interacting density
is noninteracting-v-representable. If we dierentiate Eq. (51) with respect
to the density n we obtain
v
s
(r) = v(r) +
_
d
3
r

n(r

)w([r r

[) +v
xc
(r) (52)
where
v
xc
(r) =
E
xc
n(r)
(53)
denes the exchange-correlation potential. Now the state [[v
s
]) is a ground
state for a system of noninteracting particles, and can therefore be written
as an antisymmetrized product of single-particle orbitals
i
(r). If we now
collect our results we see that we have converted the ground state problem
into the following set of equations
E[v] =
N

i=1

1
2
_
d
3
r

i
(r)
2

i
(r) +
_
d
3
rn(r)v(r)
+
1
2
_
d
3
rd
3
r

n(r)n(r

)w([r r

[) +E
xc
[n] (54)
(
1
2

2
+v(r) +
_
d
3
r

n(r

)w([r r

[) +v
xc
(r))
i
(r) =
i

i
(r) (55)
n(r) =
N

i=1
[
i
(r)[
2
(56)
The above equations constitute the ground state Kohn-Sham equations [8].
These equations turn out to be of great practical use. If we can nd a
good approximation for the exchange-correlation energy, we can calculate
the exchange-correlation potential v
xc
and solve the orbital equations self-
consistently. The density we nd in this way can then be used to calculate
the ground state energy of the system.
The exchange-correlation functional is often split up into an exchange func-
tional E
x
and a correlation functional E
c
as
E
xc
[n] = E
x
[n] +E
c
[n] (57)
in which the exchange and correlation functionals are dened by
E
x
[n] =
s
[n][

W[
s
[n])
1
2
_
d
3
rd
3
r

n(r)n(r

)w([r r

[) (58)
E
c
[n] = [n][

H
v
[[n])
s
[n][

H
v
[
s
[n]) (59)
In this equation [
s
[n]) is the Kohn-Sham wavefunction and [[n]) the true
ground state wavefunction of the interacting system with density n. Since
the Kohn-Sham wavefunction is not a ground state wavefunction of the true
system we see immediately from the variational principle that E
c
< 0. We
see also that the form of the exchange functional in terms of the Kohn-Sham
orbitals is identical to that of the exchange energy within the well-known
Hartree-Fock approximation. However, since the Kohn-Sham and Hartree-
Fock orbitals dier, the value of E
x
[n] is not equal to the Hartree-Fock
exchange. We nally remark that splitting up E
xc
into an exchange and
a correlation part has several disadvantages. First of all, this splitting has
only meaning if the ground state of the true or Kohn-Sham system is non-
degenerate. We will later see that E
xc
[n] is a well-dened functional even
for degenerate ground states, but that exchange and correlation separately
are ill-dened in that case. Secondly, there are many cases, notably molec-
ular dissociation cases, where the exchange-only theory is a bad starting
point for the treatment of correlation eects and for which it is much easier
to nd good approximations for the combined exchange-correlation func-
tional.
In this section we gave a derivation of the Kohn-Sham equations. However,
in this derivation we made a number of assumptions. We assumed that
for every ground state density of an interacting system there is a noninter-
acting system which has the same density in its ground state. Secondly,
we assumed that the density functionals were dierentiable. This assumes
that the values of the functionals change smoothly with changes in the
density. In the following sections we will investigate to which extent these
assumptions are justied.
6. Denition of the functional derivative
Let us start by dening what we mean with a functional derivative. The
derivative we will talk about is what in the mathematical literature [9, 10,
11, 12, 13, 14] is refered to as a G ateaux derivative. Let G : B be
a functional from a normed function space B to the real numbers . If
for every h B there exist a continuous linear functional G/f : B
dened by:
G
f
[h] = lim
0
G[f +h] G[f]

(60)
then G/f is called the G ateaux derivative in f B. If the limit exists
but the resulting functional of h is not linear or continuous then this limit
is called a G ateaux variation. Note that the denition of the G ateaux
derivative is very similar to the denition of the directional derivative in
vector calculus where the linear functional corresponds to the inner product
of the gradient vector with the vector which species the direction of the
dierentiation. Often the linear functional G/f can be written in the
form
G
f
[h] =
_
d
3
rg(r)h(r) (61)
If this is the case we write
g(r) =
G
f(r)
(62)
which we will call the functional derivative of G. Note further that although
G/f species a linear functional when acting on h, the function g(r)
depends in general on f in a nonlinear way. We also see that the set of
functional derivatives on a function space B is equal to the set of continuous
linear functionals on that space. This set is called the dual space of B and
denoted B

. For instance, the dual space of B = L

1
L
3
is known to be the
space B

= L
3
2
+ L

. From this we see that if the derivative of a density

functional dened on the set of densities L
1
L
3
exists then its derivative
is in the set L
3
2
+L

.
Let us now mention a straightforward consequence of the denition of the
G ateaux derivative. Suppose that the functional G which we assume to
be G ateaux dierentiable, has a minimum at f
0
, i.e. G[f] G[f
0
] for all
f B. Then the function
g() = G[f
0
+h] G[f
0
] (63)
has a minimum in = 0 and therefore the derivative of g() in = 0
vanishes. Thus
0 =
dg
d
(0) = lim
0
g() g(0)

=
G
f
(f
0
)[h] (64)
Therefore a necessary condition for a dierentiable functional for having a
minimum at f
0
is that its G ateaux derivative vanishes at f
0
. We further
prove one other fact that we will use later. If G is a convex functional, i.e.
G[f
0
+ (1 )f
1
] G[f
0
] + (1 )G[f
1
] (65)
for 0 1 then the function
g() = G[f
0
+(f
1
f
0
)] (66)
is a continuous function on the interval [0, 1]. To show this we will show
that g() is convex. Take
1
,
2
and from the interval [0, 1]. Then
g(
1
+ (1 )
2
) = G[(f
0
+
1
(f
1
f
0
)) + (1 )(f
0
+
2
(f
1
f
0
))]
g(
1
) + (1 )g(
2
) (67)
Therefore g() is a convex real function on the interval [0, 1] and hence
continuous. The fact that convexity implies continuity is true for functions
on the real axis, but this does not extend to innite-dimensional spaces.
Let us nally make some remarks on higher order derivatives. If g() =
G[f
0
+ h] denes a n-fold dierentiable function of then we dene the
n-th G ateaux variation of G as

n
G
f
n
[h, . . . , h] =
d
n
g
d
n
(0) (68)
where
n
G/f
n
has now n arguments h. If this expression denes a multi-
linear continuous functional then we call this the nth-order G ateaux deriva-
tive of G. For more details on this point we refer to reference [9]. Again, if
this expression can be written in the form

n
G
f
n
[h, . . . , h] =
_
d
3
r
1
. . . d
3
r
n
g(r
1
. . . r
n
)h(r
1
) . . . h(r
n
) (69)
then we call

n
G
f(r
1
) . . . f(r
n
)
= g(r
1
. . . r
n
) (70)
the nth-order functional derivative of G. Sofar our discussion has been
rather abstract. Let us therefore apply the denition and calculate some
functional derivatives.
7. Static linear response of the Schr odinger equation
We now consider the eect of small changes in the external eld on the
expectation values of physical observables. This is exactly what is studied
in most experimental situations where one switches on and o an external
eld and studies how the system reacts to this. We will here study a more
specic case in which we look at static changes in the external potential
and their accompanying changes in the ground state expectation values. By
investigating this problem we will learn how to take functional derivatives
and how the existence of these derivatives is related to the existence of the
linear response function.
Suppose that we have solved the following ground state problem:

H[
0
) = (

T +

V +

W)[
0
) = E
0
[
0
) (71)
where [
0
) is a nondegenerate ground state [[v]) of Hamiltonian

H with
external potential

V . From the wavefunction we can then calculate the
expectation value of any operator

O which is a well-dened functional of
the external potential
O[v] = [v][

O[[v]) (72)
Let us now calculate the functional derivative O/v at a given potential
v. According to our denition in the previous section we have to calculate
the quantity
O
v
[v] = lim
0
O[v +v] O[v]

(73)
To evaluate this limit we have to calculate O[v+v] which we will do using
static perturbation theory. We therefore make a slight change

V =
_
d
3
r n(r)v(r) (74)
in the external potential of Hamiltonian

H, i.e. we change the potential
to

V +

V . The new ground state wavefunction which we will denote by

[()) satises
(

H +V )[()) = E()[()) (75)
We will solve this equation to rst order in with the condition [(0)) =
[
0
). We note that the solution of Eq.(75) is not unique, because if [())
is a solution then also [()) = e
i()
[()) is a solution, where () is an
arbitrary function of . If we choose (0) = 0 then [()) also satises the
condition [(0)) = [
0
). The arbitrariness of the phase factor obviously
does not aect the value of any expectation value, i.e.
O() = ()[

O[()) = ()[

O[()) (76)
which is a unique function of . However, it aects the appearance of our
rst order expansion, which for both functions looks like
[()) = [
0
) +[

(0)) +O(
2
) (77)
[()) = [
0
) +([

(0)) +i

(0)[
0
)) +O(
2
) (78)
where [

(0)) and

(0) are the rst order derivatives of [()) and () in

= 0. We see that [()) and [()) dier in rst order by the amount
i

(0)[
0
), i.e. by an imaginary number times the unperturbed ground
state. This is exactly the freedom we will nd in our expansion when we
try to obtain the rst order change in the wavefunction.
Let us expand the wavefunction and energy in Eq.(75) to rst order in .
We then obtain
(

H E
0
)[

(0)) = (E

(0)

V )[
0
) (79)
where E

(0) is the rst order derivative of E() in = 0. To solve this

equation we expand [

(0)) in an orthonormal set of eigenstates of the

unperturbed Hamiltonian

H as
[

(0)) =

i=0
c
i
[
i
) (80)
If the unperturbed Hamiltonian has a continuous spectrum then for the
corresponding energy eigenstates the summation in this equation has to be
replaced by an integration. If we insert the expansion (80) into Eq.(79) for
[

(0)) we nd the equation

i=0
c
i
(E
i
E
0
)[
i
) = (E

(0)

V )[
0
) (81)
where the energies E
i
for i > 0 are the eigenenergies of the excited states
of the unperturbed Hamiltonian. If we multiply this equation from the left
with
0
[ we obtain for the change in energy
E

(0) =
0
[

V [
0
) =
_
d
3
rn
0
(r)v(r) (82)
where n
0
is the density of the unperturbed system. We have therefore
shown that
E

(0) = lim
0
E[v +v] E[v]

=
_
d
3
rn
0
(r)v(r) (83)
As has been derived before we nd E/v(r) = n
0
(r). If we now multiply
Eq.(81) from the left with
k
[ for k > 0 we nd
c
k
=

k
[

V [
0
)
E
k
E
0
(for k > 0) (84)
Note that these coecients are well-dened because E
k
> E
0
, since we are
dealing with an isolated nondegenerate ground state. We therefore nd for
the rst order change in the wavefunction
[

(0)) = c
0
[
0
)

k=1
[
k
)
k
[

V [
0
)
E
k
E
0
(85)
The coecient c
0
remains undetermined by these equations. However, from
the requirement that the perturbed system remains normalized we nd
0 =
d[)
d
(0) =

(0)[
0
) +
0
[

(0)) = c

0
+c
0
(86)
and hence c
0
must be purely imaginary. This is exactly the kind of freedom
in the rst order wavefunction that we noted before. However, as we have
seen this is related to choice of phase and does not aect the expectation
values. We now want to calculate the rst order change in an arbitrary
expectation value due to a change v of the external potential. This change
is given by
O

(0) = lim
0
O[v +v] O[v]

(0)[

O[
0
) +
0
[

O[

(0)) (87)
where O

(0) is the derivative of O() in = 0 and where we assumed that

the operator

O does not depend on

V (as for instance the Hamiltonian
does). From Eq.(85) we obtain
O

(0) = (c

0
+c
0
)
0
[

O[
0
)

k=1

0
[

V [
k
)
k
[

O[
0
) +
0
[

O[
k
)
k
[

V [
0
)
E
k
E
0
(88)
From the normalization condition Eq.(86) we see that the rst term on the
right hand side vanishes and we can rewrite the equation as
O

(0) =
O
v
[v] =
_
d
3
r
O
v(r)
v(r) (89)
where
O
v(r)
=

k=1

0
[

O[
k
)
k
[ n(r)[
0
)
E
k
E
0
+c.c. (90)
For a system with a nondegenerate ground state we therefore have obtained
an explicit expression for the functional derivative O/v of the expectation
value O[v], evaluated at potential v(r). We will use this formula later when
we are studying functional derivatives with respect to the density.
8. Invertability of the density response function
In the previous section we obtained an expression for the functional deriva-
tive of an arbitrary expectation value. We now make a special choice for
the operator

O and we choose

O = n(r

). In that case we obtain from

Eq.(90):
(r

, r) =
n(r

)
v(r)
=

k=1

0
[ n(r

)[
k
)
k
[ n(r)[
0
)
E
k
E
0
+c.c. (91)
where (r

, r) is the static density response function. From this expression

we see immediately that is real and symmetric, i.e. (r, r

) = (r

, r).
The density response function relates rst order changes in the potential to
rst order changes in the density according to
n(r) =
_
d
3
r

(r, r

)v(r

) (92)
If we change the potential by a constant then the density does not change.
We therefore must have
_
d
3
r

(r, r

) = 0 (93)
which can also be checked immediately from Eq.(91). Because of this prop-
erty we also see that the induced density n(r) automatically integrates
to zero. The density response function has therefore indeed the physical
properties that we expect. On the basis of the Hohenberg-Kohn theorem
we may further expect that the only potential variation that yields a zero
density variation is the constant potential, i.e. if
0 =
_
d
3
r

(r, r

)v(r

) (94)
then v(r) = C. This readily proven from the properties of the response
function. Suppose Eq.(94) is true for some v(r

). Then we obtain by
integration with v(r):
0 =
_
d
3
rd
3
r

v(r)(r, r

)v(r

) = 2

k=1
[a
k
[
2
E
k
E
0
(95)
where we dened a
k
by
a
k
=
_
d
3
r
k
[ n(r)[
0
)v(r) (96)
Since E
k
E
0
> 0 equation (95) can only be true if a
k
= 0 for all k 1.
But this implies that
0 =

k=1
a
k
[
k
) =

k=1
_
d
3
r[
k
)
k
[ n(r)[
0
)v(r)
=
_
d
3
rv(r) (1 [
0
)
0
[) n(r)[
0
)
=
_
d
3
rv(r) ( n(r) n
0
(r)) [
0
) (97)
where n
0
(r) =
0
[ n(r)[
0
) is the ground state density. Written in rst
quantization this implies that
N

i=1
v(r
i
)[
0
) = 0 (98)
where N is the number of electrons and v the potential
v(r) = v(r)
1
N
_
d
3
rn
0
(r)v(r) (99)
Now Eq.(98) implies that v(r) = 0 which together with Eq.(99) implies
that v(r) = C. We have therefore shown that only constant potentials
yield a zero-density variation, and therefore the density response function
is invertible up to a constant. One should, however, be careful with what
one means with the inverse response function. The response function de-
nes a mapping : 1 / from the set of potential variations from a
nondegenerate ground state, which we call 1 and is a subset of L
3
2
+L

,
to the set of rst order densities variations, which we call /, that are
produced by it. We have just shown that the inverse
1
: / 1 is
well-dened modulo a constant function. However, there are density vari-
ations that can never be produced by a potential variation and which are
therefore not in the set /. An example of such a density variation is one
which is identically zero on some nite volume.
From our analysis we can further derive another property of the static re-
sponse function. Since (r, r

) is a real Hermitian integral kernel it has an

orthonormal set of eigenfunctions which can be chosen to be real. Let f(r)
be such an eigenfunction, i.e.
_
d
3
r

(r, r

)f(r

) = f(r) (100)
Multiplication from the right with f(r) and subsequent integration yields

_
d
3
rf
2
(r) =
_
d
3
rd
3
r

f(r)(r, r

)f(r

) = 2

k=1
[b
k
[
2
E
k
E
0
0 (101)
where
b
k
=
_
d
3
r
k
[ n(r)[
0
)f(r) (102)
We therefore nd that 0. However, we already know that = 0 is only
possible if f(r) is constant. We have therefore obtained the result that if
a non-zero density variation is proportional to the potential that generates
it, i.e. n = v, then the constant of proportionality is negative. This is
exactly what one would expect on the basis of physical considerations. In
actual calculations one indeed nds that the eigenvalues of are negative.
Moreover it is found that there is a innite number of negative eigenvalues
arbitrarily close to zero, which causes considerable numerical diculties
when one tries to obtain the potential variation that is responsible for a
given density variation. We nally note the invertability proof for the static
response function can be extended to the time-dependent case. For a recent
review we refer to [15].
9. Functional derivatives and v-representability
We are now ready to tackle the question how to calculate functional deriva-
tives with respect to the density. By the Hohenberg-Kohn theorem every
density associated with a nondegenerate ground state uniquely determines
that ground state and the external potential that produced it. Because
the density n determines the external potential v and vice versa we can
parametrize the ground state either by the density or the external poten-
tial. The same is, of course, true for the expectation value of any operator

O that we calculate from the ground state. We will therefore write

O[n] = [n][

O[[n]) = [v][

O[[v]) = O[v] (103)

Since we know how to calculate the functional derivative with respect to
v the functional derivative with respect to n seems rather straightforward.
By the chain rule for dierentiation we obtain
O
n(r)
=
_
d
3
r

O
v(r

)
v(r

)
n(r)
=
_
d
3
r

O
v(r

1
(r

, r) (104)
where we used that the density response function has an inverse (modulo a
constant). However, great care must be taken when deriving this equation.
One needs to consider under which conditions the use of this chain rule
is applicable. Let us therefore start with the denition of the functional
derivative and try to calculate
O
n
[n] = lim
0
O[n
0
+n] O[n
0
]

(105)
where n
0
is the density of a given nondegenerate ground state with external
potential v
0
. Now we run into the following problem. In order to make the
value of O[n
0
+ n] well-dened we have to make sure that n
0
+ n
/, i.e. we have to make sure that this perturbed density belongs to a
nondegenerate ground state. Whether or not this is true obviously depends
on the choice we make for n. A natural choice would be to take n /,
i.e. we take a n from the set of rst order density variations produced
by some potential variations v 1. In this way we know that for every
such n there is a potential v
0
+v that generates, to rst order in , the
required density. However, this still does not imply that n
0
+ n /.
This will only be so if we can nd an additional potential that would make
the higher order terms in disappear and it is not clear how to prove the
existence of such a potential.
Nevertheless, this idea is sucient to make a useful statement about the
functional derivative. Let us therefore take some n /. Then, by the
invertability of the density response function, there is a unique v modulo
a constant, such that
n(r) =
_
d
3
r

0
(r, r

)v(r

) (106)
where
0
is the static density response function belonging to the nondegen-
erate ground state [
0
) with density n
0
. Consider now the set of external
potentials v
0
+ v. For each we can solve the Schrodinger equation and
obtain a ground state density n

(r), which for small enough will corre-

spond to a nondegenerate ground state, i.e. n

(r) /. By construction
n

n
0
will, to rst order in , be equal to the chosen n we started with,
i.e. n

can be written as
n

(r) = n
0
(r) +n(r) +m

(r) / (107)
where the term m

(r) satises
lim
0
m

(r)

= 0 (108)
Since n

is the density of a nondegenerate ground state the expectation

value O[n

] is now well-dened and we can calculate

lim
0
O[n

] O[n
0
]

= lim
0
O[v
0
+v] O[v
0
]

=
_
d
3
r
O
v(r)
v(r) (109)
We have therefore found a parametrized set of densities n

/ for which
the limit above is well-dened. Moreover, in this equation v is uniquely
dened by the n / we started with by v =
1
0
n, so that we can
write
lim
0
O[n
0
+n +m

] O[n
0
]

=
_
d
3
r
O
n(r)
n(r) (110)
where the functional derivative O/n(r) is given (modulo a constant) by
O
n(r)
=
_
d
3
r

k=1

0
[

O[
k
)
k
[ n(r

)[
0
)
E
k
E
0
+c.c.
_

1
0
(r

, r) (111)
Note that this derivative is dened up to a constant, since n(r) integrates
to zero. The functional derivative O/n[n] in Eq.(110), regarded as a
linear functional acting on n, is of course independent of this constant.
We have therefore a result that is in accordance with our naive expectation
of Eq.(104). The functional derivative in Eq.(111) is well-dened in terms
of the properties of the unperturbed system and independent of the choice
for n, and therefore independent of the parametrized path n

/ that
we used to approach n
0
. Moreover, n

approaches the straight path n

0
+
n arbitrarily closely for 0 and this is exactly the limit that we are
interested in. We will therefore call the O/n of Eq.(111) the functional
derivative of O[n]. Let us check that Eq.(111) gives us back some known
results. Let us take

O =

T +

W so that O[n] = F
HK
[n]. If we insert this
operator in Eq.(111) we see that we have to calculate

0
[

T +

W[
k
) =
0
[

H
0

V
0
[
k
)
= E
k

k0

_
d
3
rv
0
(r)
0
[ n(r)[
k
) (112)
where

H
0
=

T +

V
0
+

W. If we insert this matrix element into Eq.(111) we
obtain:
F
HK
[n]
n(r)
=
_
d
3
r

d
3
r

v
0
(r

)
0
(r

, r

)
1
0
(r

, r) = v
0
(r) (113)
This is exactly the result that we derived earlier using Legendre transforms.
We can, of course, also do this derivation for a noninteracting system where

W = 0 and therefore

O =

T, in which case we obtain
T
s
n(r)
= v
s
[n
0
](r) (114)
where v
s
[n
0
] is the potential that in a noninteracting system generates
density n
0
, i.e. the Kohn-Sham potential corresponding to density n
0
.
10. The Hohenberg-Kohn theorem for degenerate ground
states
Until now we considered only nondegenerate ground states. These ground
states have the simplifying feature that they are determined uniquely (up to
a phase factor) by the external potential and therefore any expectation value
calculated from the ground state wavefunction is a well-dened functional
of the external potential v. However, degenerate ground states do occur
(for instance in open shell atoms) and in that case the external potential
does not generate a unique ground state but a linearly independent set of
q dierent ground states. The expectation value of any operator (except
the total energy) will then depend on which ground state out of the ground
state manifold we choose to compute the expectation value from. This is
in particular true for the density operator, and therefore dierent ground
states [
i
) out of the ground state multiplet will yield dierent densities
n
i
. If we want to consider degenerate ground states, then the density is
no longer a unique functional of the external potential. However, we will
show that the inverse mapping n v is well-dened, i.e. every ground
state density determines uniquely the external potential that generated it.
We will rst generalize this statement somewhat. Instead of pure-state
densities, which come from an eigenstate of the Hamiltonian

H, we will
consider ensemble densities. To dene this concept we rst introduce for a
q-fold degenerate ground state [
i
), i = 1 . . . q the density matrix

D =
q

i=1

i
[
i
)
i
[
q

i
= 1 (0
i
1) (115)
where the ground state wavefunctions [
i
) are chosen to be orthonormal.
We then dene the ground state expectation value of operator

O as
O) = Tr

D

O (116)
where the trace operation for an arbitrary operator is dened as
Tr

A =

i=1

i
[

A[
i
) (117)
where [
i
) is an arbitrary complete set of states. The trace is independent
of the complete set that we choose, as follows simply by insertion of a
dierent complete set [
i
)
Tr

A =

i=1

i
[

A[
i
) =

i,j=1

i
[

A[
j
)
j
[
i
)
=

i,j=1

j
[
i
)
i
[

A[
j
) =

j=1

j
[

A[
j
) (118)
If we choose the complete set to be the set of eigenstates of Hamiltonian

H
then we nd
Tr

D

O =

i=1

i
[

D

O[
i
) =
q

i=1

i
[

O[
i
) (119)
This denes the expectation value of an observable

O in an ensemble de-
scribed by density matrix

D. For the particular case of the density operator
we have
n(r) = Tr

D n(r) =
q

i=1

i
[ n(r)[
i
) =
q

i=1

i
n
i
(r) (120)
We will denote densities n(r) of this type, which are obtained from an
orthonormal set of ground states [
i
), i = 1 . . . q corresponding to a po-
tential v, as ensemble v-representable densities, or for short E-V-densities.
We further denote the set of all E-V-densities generated by potentials in
L
3
2
+ L

as B. A density will be called a pure state v-representable den-

sity or for short PS-V-density if it can be written as n(r) = [ n(r)[) ,
where [) is a ground state. Obviously PS-V-densities are special cases of
E-V-densities and the set of PS-V-densities is therefore a subset of the set
of E-V-densities.
We are now ready to formulate the Hohenberg-Kohn theorem for degen-
erate ground states: Every E-V-density determines the external potential
that generated it up to an arbitrary constant. Let us make this statement a
bit more specic. Suppose that

D
1
and

D
2
are density matrices belonging
to ground state ensembles for potentials v
1
and v
2
respectively, with corre-
sponding ensemble densities n
1
and n
2
. If v
1
,= v
2
+C with C a constant,
then n
1
,= n
2
. The proof is analogous to the proof of the nondegenerate
case.
Suppose v
1
generates the ground state multiplet A
1
= [
i
), i = 1 . . . q
1

and v
2
generates the ground state multiplet A
2
= [
i
), i = 1 . . . q
2
. All
the wavefunctions within these multiplets may, without loss of generality,
be chosen orthonormal. Then none of the wavefunctions in the sets A
1
and
A
2
are equal. This follows from the same argument as used in the proof of
the Hohenberg-Kohn theorem for the nondegenerate case. In particular, as
the sets A
1
and A
2
are only dened to within a unitary transformation no
[
i
) in A
2
is a linear combination of the [
i
) in A
1
. This then implies that
two ground state ensemble density matrices constructed from the ground
states in A
1
and A
2
are dierent

D
1
=
q1

i=1

i
[
i
)
i
[ ,=
q2

i=1

i
[
i
)
i
[ =

D
2
(121)
where

i
=

i
= 1. This follows, for instance, by taking the inner
product on both sides with [
m
) as the [
i
) are not linear combinations
of the [
i
). We have thus established that the sets of ground state density
matrices for the two dierent potentials v
1
and v
2
are disjoint. We now have
to prove that the density matrices in these sets yield dierent densities.
If

H
1
=

T +

V
1
+

W and

H
2
=

T +

V
2
+

W then
Tr

D
1

H
2
> Tr

D
2

H
2
(122)
This follows directly from
Tr

D
1

H
2
=
q1

i=1

i
[

H
2
[
i
) >
q1

i=1

i
[

H
2
[
i
) =
=
q1

i=1

i
E[v
2
] = E[v
2
] =
q2

i=1

i
[

H
2
[
i
) = Tr

D
2

H
2
(123)
We can now show that

D
1
and

D
2
yield dierent densities. We proceed
again by reductio ad absurdum. We have
E[v
1
] = Tr

D
1

H
1
= Tr

D
1
(

H
2
+

V
1

V
2
)
= Tr

D
1

H
2
+
_
n
1
(r)(v
1
(r) v
2
(r))dr
> Tr

D
2

H
2
+
_
n
1
(r)(v
1
(r) v
2
(r))dr
= E[v
2
] +
_
n
1
(r)(v
1
(r) v
2
(r))dr (124)
Likewise we have
E[v
2
] > E[v
1
] +
_
n
2
(r)(v
2
(r) v
1
(r))dr (125)
which added to the last inequality leads to
_
d
3
r(n
2
(r) n
1
(r))(v
2
(r) v
1
(r)) < 0 (126)
This leads again to the contradiction 0 < 0 if we assume that n
1
= n
2
.
Therefore

D
1
and

D
2
must give dierent densities, which proves the theo-
rem.
Within the set of ensemble ground state density matrices corresponding
to the same potential however, two dierent density matrices can yield the
same density. The simplest example is the hydrogen atom with the degener-
ate ground state wavefunctions for an electron with spin up and spin down,
i.e.

= (1s) and

= (1s) where and are the spin wavefunc-

tions. Both wavefunctions obviously have the same total density. Another
example is provided by the the two degenerate ground state wavefunctions
(1s)
2
2p
+
and (1s)
2
2p

of the noninteracting lithium atom, where 2p

are
p-orbitals with angular momentum quantum numbers l = 1. Both wave-
functions lead to the same density
n(r) = 2[
1s
(r)[
2
+[
2p
+(r)[
2
= 2[
1s
(r)[
2
+[
2p
(r)[
2
(127)
We have therefore constructed two degenerate ground states with the same
density. As a consequence the ground state expectation value of a given
operator may no longer be considered as a functional of the density (take,
for instance, the expectation values of the spin and angular momentum in
our examples). However, if two dierent ground state density matrices have
the same density then also the energy Tr

D

H for those dierent density ma-
trices is the same. For every E-V-density n we can therefore unambiguously
dene the ensemble version of the F
HK
functional as [16]
F
EHK
[n] = Tr

D[n](

T +

W) (128)
where

D[n] is any of the ground state ensemble density matrices corre-
sponding to n. We can now dene an extension of the energy functional
E
v
to the set of E-V-densities
E
v
[n] =
_
n(r)v(r)dr +F
EHK
[n] = Tr

D[n]

H (129)
Similarly as for F
HK
we easily can prove
E[v] = inf
nB
__
n(r)v(r)dr +F
EHK
[n]
_
(130)
The functional F
EHK
is an extension of F
HK
since
F
EHK
[n] = F
HK
[n] if n / (131)
This follows directly from the fact that for a non-degenerate ground state
[[n]) corresponding to n we have

D[n] = [[n])[n][, so
F
EHK
[n] = Tr

D[n](

T +

W) = [n][

T +

W[[n]) = F
HK
[n] (132)
We can furthermore prove that F
EHK
is convex by the same proof as for
F
HK
. Nothing is however known about the convexity of the set of E-V-
densities B itself which constitute the domain of F
EHK
. Let us collect our
results in the form of a theorem
Theorem 2 (Hohenberg-Kohn) The E-V-density n species the exter-
nal potential up to a constant. Moreover,
1. The ground state energy of a system with external potential v can be
obtained from
E[v] = inf
nB
__
n(r)v(r)dr +F
EHK
[n]
_
where F
EHK
[n] = Tr

D[n](

T +

W) and

D[n] is a ground state density
matrix with Tr

D[n] n(r) = n(r).
2. F
EHK
is convex
As we will now demonstrate the subset of PS-V-densities of B is not convex.
More precisely, we will now show that there are E-V-densities which are
not PS-V-densities [1, 17]. As any E-V-density is a convex combination
of PS-V densities this then demonstrates the non-convexity of the set of
PS-V-densities.
Consider an atom with total angular momentum quantum number L > 0
which has a 2L + 1-degenerate ground state. The external potential v is
the spherically symmetric Coulomb potential of the atomic nucleus and
the degeneracy is a result of rotational invariance of the Hamiltonian of
the system. The ground state wavefunctions then transform among one
another according to a 2L + 1-dimensional unitary representation of the
rotation group. We assume that there is no accidental degeneracy. If we
denote the ground state wave functions by [[n
i
]) = [
i
), i = 1 . . . 2L+1
and the corresponding electron densities by n
i
then the following convex
combination
n =
1
2L + 1
2L+1

i=1
n
i
(133)
is invariant under all rotations and therefore spherically symmetric. How-
ever the n
i
are not spherical. In fact, there is no linear combination of the
ground states [
i
) that leads to a spherically symmetric density. As the n
j
is obtained from [
j
), which by a unitary transformation can be obtained
from any other [
i
), and the external potential is invariant under rotations
we nd that
_
n
i
(r)v(r)dr =
_
n
j
(r)v(r)dr =
_
n(r)v(r)dr (134)
for all 0 i, j 2L + 1. Let us now suppose that n is a ground state
density obtained from a pure state wavefunction [[ n]). This wavefunction
is not a linear combination of the [
i
) otherwise n would not be spherically
symmetric. We then nd
F
EHK
[ n] = [ n][

H
v
[[ n])
_
n(r)v(r)dr
>
2L+1

i=1
1
2L + 1
[n
i
][

H[[n
i
])
_
n(r)v(r)dr
=
2L+1

i=1
1
2L + 1
F
EHK
[n
i
] (135)
This then gives
F
EHK
[ n] >
2L+1

i=1
1
2L + 1
F
EHK
[n
i
] (136)
But we also know that F
EHK
is convex on the set of E-V-densities. This
leads to a contradiction and hence we must conclude that n can not be a
pure state density of any potential. The density n is, however, a convex
combination of ground state densities corresponding to the same external
potential and therefore, by denition, an ensemble v-representable density.
We therefore have constructed an E-V-density which is not a PS-V-density.
For an explicit numerical example of such a density we refer to the work of
Aryasetiawan and Stott [18].
11. Linear response for degenerate ground states and
the dierentiability of F
EHK
In the previous sections we learned to take functional derivatives on the
basis of linear response theory. We concentrated on nondegenerate ground
states for which the response functions are well-dened. We will now use
response theory for degenerate ground states to study the dierentiability
of the functional F
EHK
.
It is clear that in the case of degeneracy the expectation value of all ob-
servables, except the energy, depends on which particular ground state we
choose to calculate the expectation value from. This poses a clear diculty
for the denition of general density functionals and their functional deriva-
tives. One may, however, argue that if a potential v
0
leads to a degenerate
ground state then we can always nd an arbitrarily small perturbation of
the potential v that lifts the degeneracy and therefore the expectation
value
O[v
0
+v] = [v
0
+v][

O[[v
0
+v]) (137)
of an observable described by an operator

O does exist for any > 0, where
[[v
0
+ v]) is the ground state wavefunction of the perturbed system.
Therefore also the limit

O[v
0
, v] = lim
0
O[v
0
+v] (138)
is dened, although it will in general depend on the potential variation v.
One can therefore dene the functional derivative as follows
O

[v] = lim
0
O[v
0
+v]

O[v
0
, v]

(139)
However, we will see that these derivatives depend in a nonlinear way on v
and are therefore only dened as G ateaux variations rather than G ateaux
derivatives. We shall especially be interested in the case where

O is the
density operator. For the energy the limit in Eq.(138) is independent of v
and we may wonder if a G ateaux derivative still exists. Let us investigate
this in more detail. Suppose we have an Hamiltonian

H
0
with external
potential v
0
which has a q-fold degenerate ground state. Let us now apply
a perturbation v which lifts the degeneracy. Then there are q eigenstates
[
k
()), k = 1 . . . q with energies E
k
() of the perturbed Hamiltonian
that are continuously connected to degenerate ground states [
k
(0)) of the
unperturbed Hamiltonian

H
0
, i.e.
lim
0
[
k
()) = [
k
(0)) (140)
with
E
k
(0) =
k
(0)[

H
0
[
k
(0)) = E
0
(141)
Note that since the states [
k
()) are orthonormal, also the limiting states
states [
k
(0)) in the degenerate ground state multiplet are orthonormal.
Which particular ground states of the ground state manifold are reached
by the 0 limit depends obviously on the form of the perturbation v.
In which way it depends on v we will investigate now. The ground state
wavefunctions [
k
()) satisfy the equation
(

H
0
+

V +

2
2

U)[
k
()) = E
k
()[
k
()) (142)
where, for the later discussion, we also added a one-body potential

U
which is of second order in . It turns out that this term inuences the
rst order density response. We note, like in the nondegenerate case, that
if [
k
()) is a solution to this equation then also [
k
()) = e
i
k
()
[
k
()) is
a solution, where
k
() is an arbitrary function of . This freedom will, as
in the nondegenerate case, not aect any of the expectation values. If we
expand the Schrodinger equation (142) to rst order in we obtain
(

H
0
E
0
)[

k
(0)) = (E

k
(0)

V )[
k
(0)) (143)
where E

k
(0) and [

k
(0)) are the rst order derivatives of E
k
() and [
k
())
with respect to in = 0. In order to solve this equation for [

k
(0)) we
expand this quantity in the q -connected ground states and all the other
eigenstates of

H
0
, i.e.
[

k
(0)) =

i=1
c
k
i
[
i
) (144)
where [
i
) = [
i
(0)) for i = 1 . . . q and the states [
i
), i > q are eigen-
states of

H
0
with eigenenergies E
i
> E
0
. If we insert this expansion into
the rst order Schrodinger equation (143) we obtain the equation

i=1
c
k
i
(E
i
E
0
)[
i
) = (E

k
(0)

V )[
k
(0)) (145)
If we multiply this equation from the left with
j
[ where 1 j q we
obtain
0 = E

k
(0)
jk

j
(0)[

V [
k
(0)) (146)
This is the equation tells us exactly how v picks out the ground states in
the degenerate multiplet. They are exactly the ones that diagonalize

V
within the q-dimensional space of degenerate ground state functions. This
equation also tells us that
E

k
(0) =
k
(0)[

V [
k
(0)) =
_
d
3
rn
k
(r)v(r) (147)
If the state perturbed state with the lowest energy has label k = 1 then we
see that the functional derivative of the ground state energy is given as
E
v(r)
= n
1
(r) (148)
However, since n
1
implicitely depends on the perturbation v which picked
out a particular ground state, this is not a proper G ateaux derivative but
only a G ateaux variation. So even the energy functional has no functional
derivative in the proper sense. Let us investigate the consequences for
dierent expectation values. We therefore rst have to determine the rst
order wavefunction. If we multiply Eq.(145) from the right with
i
[ where
i > q we obtain the coecient c
k
i
for i > q and the following expression for
[

k
(0)):
[

k
(0)) =
q

i=1
c
k
i
[
i
)

i=q+1
[
i
)
i
[

V [
k
(0))
E
i
E
0
(149)
However, the coecients c
k
i
for 1 i q are not determined by Eq.(143).
We know from the orthonormality of the states [
k
()) that
0 =
d
i
[
k
)
d
(0) =

i
(0)[
k
(0)) +
i
(0)[

k
(0)) = c
i
k
+c
k
i
(150)
i.e. the coecients c
k
i
form an anti-Hermitian matrix. In particular we nd
that the coecients c
k
k
are purely imaginary. The value of these diagonal
elements is undetermined by the Schrodinger equation as they are related
to the arbitrary phase
k
() which we can choose for each [
k
()). As in the
nondegenerate case the value of any expectation value does not depend on
the coecients c
k
k
. This leaves us with the determination of the o-diagonal
terms c
k
i
for i ,= k. These coecients can be found by expanding Eq.(142)
to second order in
2
. This yields
(

H
0
E
0
)[

k
(0)) = 2(E

k
(0)

V )[

k
(0)) + (E

k
(0)

U)[
k
(0)) (151)
where [

k
(0)) and E

k
(0) are the second order derivatives of [
k
()) and
E
k
() in = 0. If we multiply this equation from the left with
l
(0)[ where
1 l q and use Eq.(146) and Eq.(149) then we obtain
0 = 2(E

k
(0) E

l
(0))c
k
l
+E

k
(0)
lk

l
(0)[

U[
k
(0))
+ 2

i=q+1

l
(0)[

V [
i
)
i
[

V [
k
(0))
E
i
E
0
(152)
If we in this equation take k = l we obtain an expression for the energy to
second order in :
E

k
(0) = 2
_
d
3
rd
3
r

v(r)
k
(r, r

)v(r

) +
_
d
3
rn
k
(r)u(r) (153)
where n
k
is the density corresponding to [
k
(0)) and
k
the following
function

k
(r, r

) =

i=q+1

k
(0)[ n(r)[
i
)
i
[ n(r

)[
k
(0))
E
i
E
0
+c.c. (154)
If, on the other hand, we take l ,= k we obtain our desired equation for the
coecients c
k
i
:
c
k
l
=
1
2

l
(0)[

U[
k
(0))
E

k
(0) E

l
(0)
+
1
E

k
(0) E

l
(0)

i=q+1

l
(0)[

V [
i
)
i
[

V [
k
(0))
E
i
E
0
(155)
We see that indeed c
k
l
= c
l
k
as expected. We have therefore completely
determined the rst order change in the wavefunction. Let us see what this
implies for the observables. The rst order change in the expectation value
of an observable

O is given by
O

(0) =
k
(0)[

O[

k
(0)) +c.c. (156)
If we insert our expression for the rst order change in the wavefunction
we obtain
O

(0) =
_
d
3
r
k
(r)v(r) +
1
2
_
d
3
r
k
(r)u(r)
+
_
d
3
rd
3
r

k
(r, r

)v(r)v(r

) (157)
where we dened the following functions

k
(r) =

i=q+1

k
(0)[

O[
i
)
i
[ n(r)[
k
(0))
E
i
E
0
+c.c. (158)

k
(r) =
q

l=1(l=k)

k
(0)[

O[
l
)
l
[ n(r)[
k
(0))
E

k
(0) E

l
(0)
+c.c. (159)

k
(r, r

) =
q

l=1

i=q+1

k
(0)[

O[
l
)
l
[ n(r)[
i
)
i
[ n(r

)[
k
(0))
(E

k
(0) E

l
(0))(E
i
E
0
)
(160)
If we choose

O = n(r

) then we obtain an expression for the rst order

change in the density
n
k
(r

) =
_
d
3
r
k
(r

, r)v(r) +
1
2
_
d
3
r
k
(r

, r)u(r)
+
_
d
3
rd
3
r

k
(r

, r, r

)v(r)v(r

) (161)
where the functions
k
,
k
and
k
are obtained by inserting n(r

) for

O in
Eqns.(158), (159) and (160). If we denote the perturbed state with label
k = 1 as the one with the lowest energy then n
1
represents the change
in ground state density. We see that this density change does not depend
linearly on the rst order change v in the potential and even the second
order change u of the potential contributes to its value. Let us now take
u = 0 and write
n
1
(r

) =
_
d
3
r
1
(r

, r)v(r)
+
_
d
3
rd
3
r

1
(r

, r, r

)v(r)v(r

) (162)
We can ask ourselves the question whether a given rst order variation n
1
uniquely determines the rst order density change v. One can show from
the Hohenberg-Kohn theorem for degenerate states that this is indeed the
case. If we in Eq.(126) take v
2
= v
1
+ v(r) where v is not a constant
function we obtain

2
_
d
3
rn
1
(r)v(r) +O(
3
) < 0 (163)
Dividing by
2
and taking the limit 0 then yields
_
d
3
rn
1
(r)v(r) < 0 (164)
We see that the rst order density change n
1
can not be zero if v is not
a constant. In contrast to the nondegenerate case Eq.(164) does not imply
that the function
1
is invertable, only the relation in Eq.(162) is invertable
where the inverse only exists on the set of v-representable density variations
within the set B.
Let us now see what the linear response theory can tell us about the dier-
entiability of F
EHK
. We consider again a system with a degenerate ground
state and external potential v
0
. Let us take out of the set of ground state
ensemble densities for this system a particular density n
0
. We may then
wonder whether the following limit
F
EHK
n
[n] = lim
0
F
EHK
[n
0
+n] F
EHK
[n
0
]

(165)
exists. For this limit to be well-dened we have to make sure that n
0
+
n B and we run into the same problem as for the nondegenerate case.
However, in the nondegenerate case we could solve this diculty using
response theory. We will do the same for the degenerate case, but we will
see that some diculties remain. Suppose we look at the perturbed system
with potential v
0
+ v in which the degeneracy is lifted. For this system
there is a well-dened ground state density n

= n[v
0
+v]. From response
theory we known that we can write n

as
n

(r) = n
1
(r) +n
1
(r) +m

(r) (166)
where
n
1
(r) = lim
0
n[v
0
+v] (167)
0 = lim
0
m

(r)

(168)
and where n
1
is explicitly given in Eq.(162). Now by construction the
density n

is in B for all values of . We can now consider the limit

F
EHK
n
[n
1
] = lim
0
F
EHK
[n
1
+n
1
+m

] F
EHK
[n
1
]

= lim
0
F
EHK
[v
0
+v]

F
EHK
[v
0
, v]

(169)
where

F
EHK
[v
0
, v] is dened as in Eq.(138). The latter limit is readily
calculated by inserting

T +

W =

H
0

V
0
for the operator

O in Eqns.(158)
and (160). We obtain for the rst order change of F
EHK
F
EHK
=
_
d
3
rd
3
r

v
0
(r

)
1
(r

, r)v(r)

_
d
3
rd
3
r

d
3
r

v
0
(r

)
1
(r

, r, r

)v(r)v(r

)
=
_
d
3
rv
0
(r)n
1
(r) (170)
and we thus obtain
lim
0
F
EHK
[n
1
+n
1
+m

] F
EHK
[n
1
]

=
_
d
3
rv
0
(r)n
1
(r) (171)
We see that this limit is linear in n
1
and that furthermore v
0
is independent
of n
1
. Therefore we can write for the functional derivative of F
EHK
in n
1
:
F
EHK
n(r)
(n
1
) = v
0
(r) (172)
up to an arbitrary constant. However, we see that this derivative only exists
for some special set of ground state densities corresponding to potential v
0
.
The derivative exists for those ground state densities that correspond to
pure states that can be obtained in the 0 limit for a perturbed system
with potential v
0
+v. It is not clear how to take the functional derivative
at an arbitrary ensemble density n
0
for potential v
0
. We saw that there
exist E-V-densities that are not PS-V-densities. If we consider an ensemble
corresponding to such a density and change the external potential to v
0
+v
where v lifts the degeneracy then for > 0 the ensemble will change into a
pure state and the density will change abruptly. We must therefore conclude
that for general E-V-densities the functional derivative of F
EHK
does not
exist. This poses not only a theoretical problem but also a practical one. As
we will discuss later, it is known from numerical investigations that there
are ground state densities of interacting systems that are not pure state
densities for a noninteracting system and therefore there is a clear need for
establishing a Kohn-Sham scheme for arbitrary E-V-densities. Fortunately
it turns out that one can dene an extension of the functional F
EHK
to a
larger domain of densities which can be shown to be dierentiable at the
set of all E-V-densities. This functional is the Lieb functional F
L
and will
be studied in the next section.
12. The Levy and Lieb functionals F
LL
[n] and F
L
[n]
The functionals F
HK
and F
EHK
have the unfortunate mathematical dif-
culty that their domains of denition / and B, although they are well-
dened, are dicult to characterize, i.e. it is dicult to know if a given
density n belongs to / or B. It is therefore desirable to extend the domains
of denition of F
HK
and F
EHK
to an easily characterizable (preferably
convex) set of densities. This can be achieved using the constrained search
procedure introduced by Levy [19]. We dene the Levy-Lieb functional F
LL
as
F
LL
[n] = inf
n
[

T +

W[) (173)
where the inmum is searched over all normalized anti-symmetric N-particle
wave functions in H
1
(
3N
) yielding density n. As shown earlier such a den-
sity is always in the convex set o which is again a subspace of L
1
L
3
. One
can furthermore show, as has been done by Simon [1], that the inmum is
always a minimum, i.e. there is always a minimizing wave function. As this
is the rst important result that we obtain for F
LL
we put it in the form
of a theorem
Theorem 3 For any n o there is a [[n]) H
1
(
3N
) such that
F
LL
[n] = [n][

T +

W[[n])
Let us discuss some properties of F
LL
. The functional F
LL
is an extension
of the Hohenberg-Kohn functional F
HK
, which is dened on /, to the larger
set o, i.e
F
LL
[n] = F
HK
[n] if n / (174)
This is readily derived. Suppose n is some ground state density correspond-
ing to some external potential v and ground state [[n]) then
_
n(r)v(r)dr +F
HK
[n] = [n][

H[[n]) = inf
n
[

H[) =
_
n(r)v(r)dr + inf
n
[

T +

W[) =
_
n(r)v(r)dr +F
LL
[n] (175)
We dene a corresponding energy functional
E
v
[n] =
_
n(r)v(r)dr +F
LL
[n] (176)
If n
0
is the ground state density for potential v with corresponding ground
state wave function [n
0
] then
E
v
[n] = inf
n
[

H[) [n
0
][

H[[n
0
]) = E
v
[n
0
] (177)
Minimizing E
v
over the set o therefore yields the ground state density n
0
corresponding to external potential v. The functional F
LL
has however the
inconvenient property that it is not convex.
Theorem 4 The functional F
LL
is not convex
To show this we take the example of a previous section where we presented
a density n which did not correspond to a ground state wavefunction. It
was a convex combination of 2L + 1 degenerate ground state densities n
i
with corresponding ground states [[n
i
]) for an external potential v. Then
we nd
_
n(r)v(r)dr +F
LL
[ n] = inf
n
[

H[)
>
1
2L + 1
2L+1

i=1
[n
i
][

H[[n
i
]) =
1
2L + 1
2L+1

i=1
F
LL
[n
i
] +
_
n(r)v(r)dr
(178)
and we nd
F
LL
[ n] >
1
2L + 1
N

i=1
F
LL
[n
i
] (179)
which proves the non-convexity of F
LL
. This is somewhat unfortunate as
convexity is an important property which can be used to derive dieren-
tiability of functionals. We will therefore now dene a dierent but related
convex functional with the same domain o. This is the Lieb functional F
L
dened as
F
L
[n] = inf

Dn
Tr

D(

T +

W) (180)
where the inmum is searched over all N-particle density matrices

D =

i=1

i
[
i
)
i
[

i=1

i
= 1 [
i
) H
1
(
3N
) (181)
which yield the given density n(r) = Tr

D n(r) where [
i
) is an orthonor-
mal set. Also for this case one can prove the inmum to be a minimum,
i.e. there is a minimizing density matrix. We put the result again in the
form of a theorem
Theorem 5 For every n o there is a density matrix

D[n] such that
F
L
[n] = Tr

D[n](

T +

W)
For the proof of this theorem we again refer to Lieb [1]. The functional F
L
is an extension of F
EHK
to the larger set o, that is
F
L
[n] = F
EHK
[n] if n B (182)
This follows directly from the fact that if n B then there is a potential v
which generates a ground state ensemble density matrix

D[n] which yields
n. So
_
n(r)v(r)dr +F
EHK
[n] =
_
n(r)v(r)dr + Tr

D[n](

T +

W)
= Tr

D[n]

H = inf

Dn
Tr

D

H
=
_
n(r)v(r)dr +F
L
[n] (183)
We can again dene an energy functional
E
v
[n] =
_
n(r)v(r)dr +F
L
[n] (184)
which by a similar proof as for F
LL
assumes its minimum at the ground
state density corresponding to potential v. We further have the following
relations
F
L
[n] = F
LL
[n] if n / (185)
and
F
L
[n] < F
LL
[n] if n B and n , / (186)
The rst relation follows from the fact that if the density n is a pure state
v-representable density then the minimizing density matrix for F
L
is a pure
state density matrix. The second relation also easily follows. We take n to
be an E-V-density which is not a PS-V density. There is a ground state
ensemble density matrix

D[n] for which we have
_
n(r)v(r)dr+F
L
[n] =
_
n(r)v(r)dr+Tr

D[n](

T+

W) =
i
[

H[
i
) (187)
where [
i
) is any of the ground states in the degenerate ground state multi-
plet. Any wave function yielding density n can not be a linear combination
of these ground state wave-functions otherwise n would be pure state v-
representable. Therefore its expectation value with the Hamiltonian must
be larger, i.e

i
[

H[
i
) < inf
n
[

H[) =
_
n(r)v(r)dr +F
LL
[n] (188)
which proves our statement.
We will now demonstrate another important property of F
L
, which is its
convexity.
Theorem 6 The functional F
L
is convex
This is easily shown. If n =
1
n
1
+
2
n
2
with
1
+
2
= 1 and 0
1
,
2
1
then we have

1
F
L
[n
1
] +
2
F
L
[n
2
] =
1
inf

D1n1
Tr

D
1
(

T +

W)
+
2
inf

D2n2
Tr

D
2
(

T +

W)
= inf

D1,

D2n1,n2
Tr(
1

D
1
+
2

D
2
)(

T +

W)
inf

Dn
Tr

D(

T +

W) = F
L
[n] (189)
We therefore now have established that F
L
is a convex functional on a
convex space. This is important information which enables us to derive
the G ateaux dierentiability of the functional F
L
at the set B of ensemble
v-representable densities. We will discuss this feature of F
L
in the next
section.
Having obtained some desirable convexity properties of F
L
we try to obtain
some analytic properties of this functional. An obvious question to ask is
whether this functional is continuous. To be more precise, suppose that
a series of densities n
k
approaches a given density n in some sense, for
instance |n n
k
|
1
0 and |n n
k
|
3
0 for k in L
1
L
3
. Does
this imply that [F
L
[n
k
] F
L
[n][ 0. ? It turns out that this question is
not easily answered. However, one can prove a weaker statement. Suppose
n is an E-V-density corresponding to potential v of Hamiltonian

H
v
. Then
F
L
[n] +
_
d
3
rn(r)v(r) = Tr

D[n]

H
v
Tr

D[n
k
]

H
v
= F
L
[n
k
] +
_
d
3
rn
k
(r)v(r) (190)
and therefore
F
L
[n] F
L
[n
k
] +
_
d
3
rv(r)(n
k
(r) n(r)) (191)
Further, since n
k
n in the norms on L
1
L
3
for each > 0 there is an
integer M such that |n n
k
|
1
and |n n
k
|
3
for k > M. Now we
can split up v L
3
2
+L

as v = u +w where u L
3
2
and w L

and we
have
[
_
d
3
rv(r)(n
k
(r) n(r))[ |u|3
2
|n n
k
|
3
+|w|

|n n
k
|
1
(192)
So if |n n
k
]|
1
/C and |n n
k
]|
3
/C, where C = |u|3
2
+ |w|

then
F
L
[n] F
L
[n
k
] + (193)
for k suciently large. We therefore see that if we take 0 from above
then F
L
[n
k
] approaches F
L
[n] from above and n
k
n in the norms on
L
1
L
3
. This, of course, does not imply that F
L
[n] is continuous in n. For
that we would have to prove that [F
L
[n] F
L
[n
k
][ rather than F
L
[n]
F
L
[n
k
] . What we have proven is a weaker form of continuity, known
as semicontinuity. Because the limit point is a lower bound, the functional
with the property in Eq.(193) is called lower semicontinuous. This can
also be characterized dierently. If we dene inf F[n
m
] by inf F
L
[n
m
] =
infF
L
[n
k
][k m then lower semicontinuity implies
F
L
[n] lim
m
inf F
L
[n
m
] (194)
Since we required that n is an E-V-density we have proven that F
L
is lower
semicontinuous on the set of E-V-densities. It turns out that one can prove
that F
L
is lower semicontinuous on all densities in o (see theorem 4.4 of
Lieb [1]). However, the proof of this is not simple and we will therefore not
try to reproduce it here. Since the result is important we present it here in
the form of a theorem
Theorem 7 Suppose n
k
and n o and n
k
n for k in the norms
on L
1
L
3
. Then
F
L
[n] lim
k
inf F
L
[n
k
]
In other words F
L
is lower semicontinuous on the set o.
One can prove an even stronger theorem in which we only need weak con-
vergence of the series n
k
. We will, however, not need that property in
the remainder of this review. The notion of lower semicontinuity is an
important property for convex functionals which will allow us to make sev-
eral other useful statements about other properties of F
L
. One particular
consequence of lower semicontinuity that we will use, is that for a lower
semicontinuous convex functional G : B the following set of points
epi(G) = (n, r) B [G[n] r (195)
is closed and convex [11]. This set is called the epigraph of G and is simply
the set of points that lie above the graph of G. The closedness property
means that if a series of points (n
k
, r
k
) in the epigraph of G converges to a
point (n, r) then this point also lies in the epigraph. This also implies that
the set of interior points of epi(G), i.e. the set of points that lie strictly
above G, is an open convex set (this means that for every point in this set
there is a neighborhood that contains it). We will need this property in the
next section.
13. Dierentiability of F
L
In this section we will prove that the Lieb functional is dierentiable on the
set of E-V-densities and nowhere else. The functional derivative at a given
E-V-density is equal to v where v is the external potential that generates
the E-V-density at which we take the derivative. To prove existence of the
derivative we use the geometric idea that if a derivative of a functional G[n]
in a point n
0
exists, then there is a unique tangent line that touches the
graph of G in a point (n
0
, G[n
0
]). To discuss this in more detail we have
to dene what we mean with a tangent. The discussion is simplied by
the fact that we are dealing with convex functionals. If G : B is a
dierentiable and convex functional from a normed linear space B to the
real numbers then from the convexity property it follows that for n
0
, n
1
B
and 0 1 that
G[n
0
+(n
1
n
0
)] G[n
0
] (G[n
1
] G[n
0
]) (196)
From the fact that G is dierentiable we then nd
G[n
1
] G[n
0
] lim
0
G[n
0
+(n
1
n
0
)] G[n
0
]

=
G
n
[n
1
n
0
] (197)
Therefore
G[n
1
] G[n
0
] +
G
n
[n
1
n
0
] (198)
Now G/n is a continuous linear functional which will be identied with
a tangent. This is the basis of the following denition. If for a convex
functional G there is a continuous linear functional L : B such that
G[n
1
] G[n
0
] +L[n
1
n
0
] (199)
then L is called a subgradient or tangent functional in n
0
. The geometric
idea behind this denition is illustrated in Figure 1.
F[n]
n
n
0
F[n
0
]+L[n-n
0
]
Figure 1: The convex functional F has a unique tangent
functional L in the point n
0
For the Lieb functional F
L
we will now prove the following statement:
Theorem 8 The functional F
L
has a unique tangent functional for every
E-V-density and nowhere else. Moreover the tangent functional at an E-V
density n can be identied with v where v is the potential that generates
this density.
In order to show this we rst dene the energy functional
E[v] = inf

D
Tr

D

H
v
(200)
where the inmum is searched over all density matrices of the form given
in Eq.(181). This is an extension of the previously dened functional in
Eq.(130) to all potentials in L
3
2
+ L

. One can further show that if an

inmum exists the the minimizing density matrix satises the Schrodinger
equation [1]. Let us rst suppose that n
0
is an E-V-density corresponding
to potential v of Hamiltonian

H
v
. Then
F
L
[n
0
] +
_
d
3
rn
0
(r)v(r) = Tr

D[n
0
]

H
v
Tr

D[n]

H
v
= F
L
[n] +
_
d
3
rn(r)v(r) (201)
and therefore
F
L
[n] F
L
[n
0
]
_
d
3
rv(r)(n(r) n
0
(r)) (202)
Therefore the functional L
L[n] =
_
d
3
rv(r)n(r) (203)
denes a linear and continuous functional (continuity follows from Eq.(192)
) and is therefore a tangent functional. Now we have to show its uniqueness.
Suppose L

is a dierent tangent functional. Since F is dened on B =

L
1
L
3
the tangent functional must represent an element in the dual space
B

= L
3
2
+L

and can be written as

L

[n] =
_
d
3
r v(r)n(r) (204)
with v L
3
2
+L

and v ,= v +C. Now since we assumed that L

was also
tangent functional we have
F
L
[n] F
L
[n
0
]
_
d
3
r v(r)(n(r) n
0
(r)) (205)
and hence
F
L
[n] +
_
d
3
rn(r) v(r) F
L
[n
0
] +
_
d
3
rn
0
(r) v(r) (206)
We take the inmum over all densities in o on the left hand side. Since
we know that there is ground state density matrix corresponding to n
0
we
obtain from Eq.(206):
E[ v] = inf
nS
F
L
[n] +
_
d
3
rn(r) v(r) Tr

D[n
0
]

H
v
(207)
Now there are two cases, either the Hamiltonian with potential v is able to
support a bound ground state or it is not. If it does support a ground state
then
Tr

D[n
0
]

H
v
> E[ v] (208)
since

D[n
0
] is not a ground state density matrix for v. Together with
Eq.(207) this immediately leads to the contradiction E[ v] > E[ v]. If the
Hamiltonian with potential v does not support a bound ground state then
there is no normalized density matrix for which the inmum of E[ v] as in
Eq.(200) is attained. We then again obtain Eq.(208) and the same contra-
diction E[ v] > E[ v]. We therefore conclude that the E-V-density n
0
has a
unique tangent functional equal to v where v is the potential that yields
ground state density n
0
.
Let us now suppose that n
0
is not an E-V-density. Let us further suppose
that there is a tangent functional at n
0
, i.e. that for some v Eq. (205)
and therefore also Eq.(206) are satised. The constrained search for F
L
[n
0
]
always has a minimizing density matrix, which we call

D[n
0
]. With this
density matrix then also Eq.(207) is true. Since n
0
is not an E-V-density

D[n
0
] can not be a ground state density matrix for the Hamiltonian with
potential v. With the same arguments as before we nd that Eq.(208) must
be true which again leads to the contradiction E[ v] > E[ v]. Therefore there
is no tangent functional at n
0
if n
0
is not an E-V-density. We have there-
fore proven our statement, there is a unique tangent functional at every
E-V-density and nowhere else.
As a next step we will show that this implies that the functional F
L
G ateaux
dierentiable at every E-V-density and nowhere else. For the application
of the next theorem it is desirable to extend the domain of F
L
to all of
L
1
L
3
. We follow Lieb [1] and dene
F
L
[n] =
_
inf

Dn
Tr

D(

T +

W) if n o
+ if n L
1
L
3
and n , o
(209)
The reader may seem surprised with the appearance of + in this def-
inition. However, innite values are well-dened in the theory of convex
functionals [11] and they are usually introduced to deal in a simple way
with domain questions. With this denition the functional F
L
is a convex
lower semicontinuous functional of the whole space L
1
L
3
. We are now
ready to introduce the following key theorem which we will use to prove
dierentiability of F
L
at the set of E-V-densities:
Theorem 9 Suppose F : B is a lower semicontinous convex func-
tional which is nite at a convex subset V B of a normed linear space
B. If F has a unique continuous tangent functional L : B at n
0
V
then F is G ateaux dierentiable at n
0
and L = F/n.
The fact that uniqueness of a tangent functional implies dierentiability is
clear from a geometric point of view. In Figure 2 we display an example of
a functional (in this picture the functional is simply a function) which is not
dierentiable at n
0
. Consequently, there is no unique tangent functional
at that point. In the example there are in fact innitely many tangent
functionals at n
0
of which there are two drawn in the plot.
F[n]
n n
0
L
1
L
2
Figure 2: The convex functional F has no unique tangent
functional in the point n
0
. Both L
1
and L
2
are tangent
functionals
In the case of innitely many dimensions one should, however, be careful
in drawing conclusions from simple plots and we have to resort to formal
proofs. Since the proof of the previous theorem is important for our discus-
sion and not dicult to follow we will present it here. Since V is a convex
set we have that if n
0
V and n
1
V then n
0
+ (n
1
n
0
) V for
0 1 and since F is convex we obtain
F[n
0
+(n
1
n
0
)] F[n
0
] +(F[n
1
] F[n
0
]) (210)
Furthermore since F has a unique continuous tangent functional at n
0
we
have
F[n
0
+(n
1
n
0
)] F[n
0
] +L[n
1
n
0
] (211)
Note that n
1
n
0
need not be in V but is always in B since that is a linear
space. If we combine the two inequalities we nd that
F[n
1
] F[n
0
]
F[n
0
+(n
1
n
0
)] F[n
0
]

L[n
1
n
0
] (212)
Now since F is convex the function g() = F[n
0
+(n
1
n
0
)] F[n
0
] is a
convex real function on the interval [0, 1] and hence continuous. Moreover,
as the equation above shows, the function g()/ has a nite upper and
lower bound and therefore the limit 0 of g()/ exists. We therefore
nd
F[n
1
] F[n
0
] F

[n
0
, n
1
] L[n
1
n
0
] (213)
where
F

[n
0
, n
1
] = lim
0
F[n
0
+(n
1
n
0
)] F[n
0
]

(214)
It remains to show that F

[n
0
, n
1
] is continuous and linear and equal to L.
Now F

has the property that for n

2
V
F

[n
0
, n
0
+(n
2
n
0
)] = lim
0
F[n
0
+(n
2
n
0
)] F[n
0
]

= F

[n
0
, n
2
]
(215)
Therefore from Eq.( 213) we nd
F[n
0
+(n
1
n
0
)] F[n
0
] +F

[n
0
, n
1
] (216)
This means that the straight line
L = (n
0
+(n
1
n
0
), F[n
0
] +F

[n
0
, n
1
]) B [ [0, 1] (217)
lies below the graph of functional F. Now since F is convex and lower
semicontinuous the set of points above the graph of F (the interior points
of epi(F)) form an open convex set which is nonempty since F is nite
on V [11]. Now a famous theorem of functional analysis, the Hahn-Banach
theorem, tells us that if we have an open convex set A (the interior of epi(F)
in our case) and an ane subspace L (which is the line L in our case) that
does not intersect A then there is a hyperplane H that contains L in which
the hyperplane H has the form
H = (n, r) B [L

[n] +r = (218)
where L

: B is a continuous linear functional and and are real

numbers (We can take ,= 0 since F

[n
0
, n
1
] is nite and we will have
no vertical line L or vertical hyperplane H). We will not prove the Hahn-
Banach theorem as it is geometrically intuitive and can be found in most
textbooks on functional analysis [20, 21]. We will just use its consequences.
Now we know that our line L is contained in H. The coecient is then
determined by the fact that (n
0
, F[n
0
]) Hand we nd = L

[n
0
]+F[n
0
]
and therefore
H = (n, r) B [L

[n n
0
] +(r F[n
0
]) = 0 (219)
Now from the fact that (n
0
+ (n
1
n
0
), F[n
0
] + F

[n
0
, n
1
]) H we nd
that
L

[n
1
n
0
] +F

[n
0
, n
1
] = 0 (220)
From this equation and Eq.(213) we then see immediately that L

/ is a
continuous tangent functional at n
0
which then must be equal to L since
that was the only continuous tangent functional at n
0
. Consequently
F

[n
0
, n
1
] = L[n
1
n
0
] (221)
and thus
F
n
[n
1
n
0
] = lim
0
F[n
0
+(n
1
n
0
)] F[n
0
]

= L[n
1
n
0
] (222)
We therefore obtained what we wanted to prove, F is G ateaux dieren-
tiable at n
0
and F/n = L. Now we can apply this theorem to the Lieb
functional. If we take F = F
L
, B = L
1
L
3
, V = o and use that F
L
has a
unique tangent functional L = v at every E-V-density and nowhere else,
we obtain:
Theorem 10 The functional F
L
is G ateaux dierentiable for every E-V-
density in the set o and nowhere else. Moreover the functional derivative
at an E-V-density is equal to v where v is the potential that generates this
density.
This leaves us with the question which densities in set o actually are
E-V-densities. A useful result on this point is obtained in the next section.
14. Ensemble v-representability
We have seen that the Lieb functional F
L
is dierentiable at the set of E-
V-densities in o and nowhere else. For this reason it is desirable to know a
bit more about these densities. The question therefore is: which densities
are ensemble v-representable? In this section we will prove a useful result
which will enable us to put the Kohn-Sham approach on a rigorous basis.
Theorem 11 The set of E-V-densities is dense in the set o with respect
to the norm on L
1
L
3
.
This statement means the following. Suppose we take an arbitrary
density n
0
from the set o, then for every > 0 we can nd an E-V-density
n such that |n
0
n|
1
and |n
0
n|
3
. In other words, for every
density in the set o there is an E-V-density arbitrarily close to it. This
can also be phrased dierently. For every density n
0
in the set o there is a
series of E-V-densities n
k
such that
lim
k
|n
0
n
k
|
p
= 0 (223)
where the subindex for the norm takes the values p = 1 and p = 3. In order
to establish this result we need to use a theorem due to Bishop and Phelps.
For the clarity of the discussion we split the theorem into two parts which
both yield interesting results for our functional F
L
. The rst part gives
some insight in the set of potentials 1 that generate a bound ground state.
From the relation
E[v] = inf
nS
F
L
[n] +
_
d
3
rn(r)v(r) (224)
we obtain immediately that for any n o and any v L
3
2
+L

F
L
[n]
_
d
3
rn(r)v(r) +E[v] (225)
Now the functional on the right hand side of the inequality sign is, for
a given v, a linear functional of n. The inequality sign tells us that this
functional lies below the graph of F
L
[n]. A linear functional with this
property is called F
L
-bounded. Let us give a general denition of these
linear functionals. Let F be a functional F : B from a normed
function space (a Banach space) B to the real numbers. Let B

be the
dual space of B, i.e. the set of continuous linear functionals on B. Then
L B

is said to be F-bounded if there is a constant C such that for all

n B
F[n] L[n] +C (226)
where the constant C may depend on L but not on n. We therefore see
from Eq.(225) that every v L
3
2
+ L

denes an F
L
-bounded functional
where L is dened as in Eq.(203). The geometric picture of an F-bounded
functional is displayed in Figure 3.
F[n]
n
L
1
L
2
Figure 3: Both L
1
and L
2
are F-bounded by the convex
functional F
Note that the tangent functionals of F
L
are special cases of F
L
-bounded
functionals, they are the F
L
bounded functionals that also touch the graph
of F
L
. Therefore the set of tangent functionals is a subset of the set of F
L
-
bounded functionals. This can be illustrated with an example. Consider
the function f(x) = exp (x) on the real axis. This function is convex and
all tangent functions are of the form g

(x) = x + where > 0 and

() 1. The constant function g
0
(x) = with 0 is an f-bounded
function, but not a tangent. However, we can always nd a tangent with a
slope arbitrary close to zero, i.e. arbitrary close to the slope of the constant
f-bounded function g
0
. The following theorem (due to Bishop and Phelps)
ensures that a similar situation occurs for the case of general Banach spaces:
Theorem 12 (Bishop-Phelps I) Let F : B be a lower semicontin-
uous convex functional on a real Banach space B. The functional F can take
the value + but not everywhere. Then the continuous tangent functionals
to F are B

-norm dense in the set of F-bounded functionals in B

.
This means that if L
0
is some F-bounded functional then we can nd a
set of tangent functionals L
k
such that
lim
k
|L
0
L
k
|
B
= 0 (227)
where the limit is taken in the norm on B

. We will not prove this theorem

here. The proof is clearly described with a geometric interpretation in
reference [22]. Let us apply this theorem to our functional F
L
. We know
that every tangent functional of F
L
can be identied with v, where v is
a potential that yields a bound ground state, i.e. v 1. Now we know
that every element v
0
for an arbitrary v
0
L
3
2
+ L

corresponds to an
F
L
-bounded functional. Therefore for such a v
0
there is a series of v
k
1
such that
lim
k
|v
0
v
k
| = 0 (228)
where the norm is the L
3
2
+L

-norm described in Eq.(24). Therefore every

potential v
0
L
3
2
+ L

can be approximated to arbitrary accuracy by a

potential that yields a bound ground state. This may seem counterintuitive
at rst sight as one can imagine v
0
to be a repulsive potential, until one
realizes that one may put the system in big box with a wall of nite height.
This system will have a bound state if the size of the box is chosen to be
big enough. The particles will then spread out over the box in order to
minimize the repulsion between them. If we choose v
k
= v
0
+w
k
where w
k
describes a series of boxes of increasing size but with decreasing height of
the potential wall then we see that we have created a series of potentials
in 1 that approaches v
0
to an arbitrary accuracy. Let us now discuss the
second part of the Bishop-Phelps theorem
Theorem 13 (Bishop-Phelps II) Let F : B be a lower semicon-
tinuous convex functional on a real Banach space B. The functional F can
take the value + but not everywhere. Suppose n
0
B with F[n
0
] <
and let L
0
B

be an F-bounded functional. Then for every 0 there

exists n

B and a functional L

such that
1. |L

L
0
|
B

2. F[n] F[n

] +L

[n n

] for all n.
3. |n

n
0
|
B
F[n
0
] L
0
[n
0
] inf
nB
F[n] L
0
[n]
For the details of the proof we again refer to [22]. The rst two points of
this theorem are equivalent to the previous theorem. They say that any
F-bounded functional L
0
can be approximated to arbitrary accuracy by a
tangent functional L

. The inequality in the third point of this theorem

allows us to make statements about distances between elements in B, which
in our case will be densities. Note that the right hand side of this inequality
has the geometric meaning of being the dierence of the vertical distance of
the functionals L
0
and F in n
0
and the shortest possible distance between
L
0
and F. Let us now apply this theorem to the functional F
L
and show
that that every density in o is arbitrarily close to an E-V-density. We rst
need some preliminaries. From Eq.(224) we see that we can write
F
L
[n] = sup
vL
3
2 +L

E[v]
_
d
3
rn(r)v(r) (229)
If the supremum is attained for some v then n is an E-V-density. This
follows because then there is a density matrix

D[n] that yields density n
(see Theorem 5) such that
Tr

D[n]

H
v
= F
L
[n] +
_
d
3
rn(r)v(r) = E[v] = inf

D
Tr

D

H
v
(230)
The density matrix

D[n] must therefore be a ground state density matrix.
If n is not an E-V-density then the supremum is not attained for any v. In
any case, for every integer k and any density n
0
o we can always nd
some v
k
such that
E[v
k
]
_
d
3
rn
0
(r)v
k
(r) sup
vL
3
2 +L

E[v]
_
d
3
rn
0
(r)v(r)
1
k
= F
L
[n
0
]
1
k
(231)
where we note that the series v
k
does not converge to any v if n
0
is not an
E-V-density. Furthermore, for any n we have
E[v
k
] = inf
nS
F
L
[n] +
_
d
3
rn(r)v
k
(r) F
L
[n] +
_
d
3
rn(r)v
k
(r) (232)
which in combination with the previous inequality yields
F
L
[n] +
_
d
3
rn(r)v
k
(r) E[v
k
] F
L
[n
0
] +
_
d
3
rn
0
(r)v
k
(r)
1
k
(233)
We are now ready to apply the Bishop-Phelps theorem to the Lieb func-
tional F
L
. We take = 1 , n
0
o and let v
k
correspond to the F-bounded
functional L
0
of the theorem. According to the theorem we can then nd
a tangent functional w
k
(the L

of the theorem) such that

|v
k
w
k
| 1 (234)
in the norm on L
3
2
+L

. From our previous investigations we know that

the tangent w
k
touches the graph of F
L
in an E-V-density n
k
, i.e.
F
L
[n] F
L
[n
k
]
_
d
3
rw
k
(r)(n(r) n
k
(r)) (235)
where w
k
is the potential that generates density n
k
. The Bishop-Phelps
theorem then tells us that
|n
k
n
0
|
p
F
L
[n
0
] +
_
d
3
rn
0
(r)v
k
(r)
inf
nS
F
L
[n] +
_
d
3
rn(r)v
k
(r) (236)
where the normindex has the values p = 1 and p = 3. Note that the
inmum in this equation can be taken over o rather than L
1
L
3
since F
L
is dened to be + outside o. Then from Eq.(233) we see immediately
that
inf
nS
F
L
[n] +
_
d
3
rn(r)v
k
(r) F
L
[n
0
] +
_
d
3
rn
0
(r)v
k
(r)
1
k
(237)
and therefore
|n
k
n
0
|
p

1
k
(238)
for p = 1 and p = 3. Since this equation is true for any k we see that any
density n
0
o can be approximated to any accuracy by the E-V-density
n
k
in the norms on L
1
L
3
. This proves the theorem in the beginning of
this section.
15. The Kohn-Sham approach and noninteracting v-
representability
We have now come to the discussion of the central equations which form
the basis of almost any practical application of density functional theory:
the Kohn-Sham equations. Kohn and Sham [8] introduced an auxiliary
noninteracting system of particles with the property that it yields the same
ground state density as the real interacting system. In order to put the
Kohn-Sham procedure on a rigorous basis we introduce the functional
T
L
[n] = inf

Dn
Tr

D

T (239)
We see that this is simply the Lieb functional with the two-particle inter-
action omitted. All the properties of the functional F
L
carry directly over
to T
L
. The reason is that all these properties were derived on the basis
of the variational principle in which we only required that

T +

W is an
operator that is bounded from below. This is, however, still true if we omit
the Coulomb repulsion

W. We therefore conclude that T
L
is a convex lower
semicontinuous functional which is dierentiable for any density n that is
ensemble v-representable for the noninteracting system and nowhere else.
We refer to such densities as noninteracting E-V-densities and denote the
set of all noninteracting E-V-densities by B
0
. Let us collect all the results
for T
L
in a single theorem:
Theorem 14 T
L
is a convex lower semicontinuous functional with the fol-
lowing properties:
1. For any n o there is a minimizing density matrix

D[n] with the
property T
L
[n] = Tr

D[n]

T.
2. T
L
is G ateaux dierentiable at the set of noninteracting E-V-densities
and nowhere else.
3. The functional derivative at a noninteracting E-V-density n is given
by:
T
L
n(r)
= v
s
[n](r)
where the potential v
s
generates the density n in a noninteracting
system.
From the last point in this theorem we see that if we want to know if
a given density n from the set o can be obtained as an E-V-density of
a noninteracting system we may try to calculate the derivative of T
L
for
this density. If it exists then the derivative yields the potential that we
were looking for. Now Kohn and Sham [8] assumed that for any density of
an interacting system there is a noninteracting system that has the same
density as its ground state. We can now ask the question whether the sets
of interacting and noninteracting E-V-densities are equal. This is currently
not known. However, we can make a number of useful conclusions. First of
all, if we apply the Bishop-Phelps theorem to the functional T
L
we obtain
the following result:
Theorem 15 The set of noninteracting E-V-densities is dense in the set
o with respect to the norm on L
1
L
3
.
This means that for any density n o there is an noninteracting E-V-
density arbitrarily close. We can in particular choose n to be an interacting
E-V-density, i.e. n B and nd a noninteracting E-V-density arbitrarily
close. We also know from the Bishop-Phelps theorem applied to F
L
that the
set of interacting E-V-densities is dense in o and therefore for any n o, in
particular n B
0
, we can nd an interacting E-V-density arbitrarily close.
We therefore conclude
Theorem 16 The set B
0
of noninteracting E-V-densities is dense in the
set B of interacting E-V-densities, and vice versa.
If we combine this result with previous theorems we obtain the following
important consequence for the Kohn-Sham scheme:
Theorem 17 Suppose n B is an interacting E-V-density. Then for every
> 0 there is a noninteracting E-V-density n

B
0
such that
1. |n n

|
p
for p = 1 and p = 3.
2. The density n

is a ground state ensemble density of a noninteracting

system with potential
v
s,
(r) =
T
L
n(r)
(n

)
i.e. we can always set up a Kohn-Sham scheme that produces a given in-
teracting E-V-density to arbitrary accuracy.
This theorem tells us that in practice we can always set up a Kohn-Sham
scheme. We also see that if we want to prove that the sets B and B
0
of in-
teracting and noninteracting E-V-densities are equal, then we have to show
that the potentials v
s,
in this theorem approach some potential v
s,0
for
0 in some smooth way. This, however, has not been proven until now.
In numerical calculations (see the discussion at the end of this section) one
has indeed always succeeded in obtaining a Kohn-Sham potential for given
interacting E-V-densities obtained from accurate Conguration Interaction
(CI) calculations. In these calculations convergence to a given Kohn-Sham
potential is sometimes dicult to obtain, but seems to happen in a rather
smooth way. One might therefore expect that the sets of interacting and
noninteracting E-V-densities are in fact equal. In order to encourage further
work in this eld we put it here as a conjecture
Conjecture 1 The sets B and B
0
of interacting and noninteracting E-V-
densities are equal, i.e. B = B
0
.
Let us describe a couple of cases for which we know this conjecture to be
true. Consider a system of two particles. Let the interacting system have
density n. Then we can construct a noninteracting Kohn-Sham system with
ground state wavefunction

KS
(r
1

1
, r
2

2
) =
1

2
(r
1
)(r
2
)((
1
)(
2
) (
1
)(
2
)) (240)
where and are the usual spin functions with (
1
2
) = (
1
2
) = 1 and
(
1
2
) = (
1
2
) = 0. This wavefunction has a density n(r) = 2[(r)[
2
and
the Kohn-Sham orbital satises
_

1
2

2
+v
s
[n](r)
_
(r) = (r) (241)
If we now choose (r) =
_
(r)/2 then for an interacting E-V-density n
the Kohn-Sham Hamiltonian with potential
v
s
[n](r) =
1
2

n
+ (242)
has orbital as an eigenfunction. Because the density is positive the orbital
=
_
n/2 has no nodes and must be a ground state orbital. We therefore
have constructed a noninteracting system with ground state n, which is
even a pure state density.
Another case where the conjecture is true is for lattice systems. The lattice
system is obtained by discretizing the many-body Hamiltonian on a grid,
i.e. we replace the dierential equation by a dierence equation as one
might do to solve the Schrodinger equation numerically. Then any inter-
acting E-V-density on the grid is also a noninteracting E-V-density. This
was proven by Chayes, Chayes and Ruskai [23]. This work may be used
to prove the general conjecture above if one can prove that the continuum
limit can be taken in some smooth way.
Let us nally discuss the exchange-correlation functional which is the cen-
tral object in any application of density functional theory. We dene the
exchange-correlation functional E
xc
by
F
L
[n] = T
L
[n] +
1
2
_
d
3
r
1
d
3
r
2
n(r
1
)n(r
2
)
[r
1
r
2
[
+E
xc
[n] (243)
Since both T
L
and F
L
are dened on the set o the exchange-correlation
functional E
xc
is also dened on that set. Since F
L
and T
L
are dierentiable
respectively on the sets B and B
0
and nowhere else, the functional E
xc
is
dierentiable on B B
0
and nowhere else. The derivative of Eq.(243) on
that set is given by
v(r) = v
s
(r) +
_
d
3
r

n(r

)
[r r

[
+
E
xc
n(r)
(244)
If we dene the exchange-correlation potential by
v
xc
(r) =
E
xc
n(r)
(245)
then we see that the Kohn-Sham potential can be written as
v
s
(r) = v(r) +
_
d
3
r

n(r

)
[r r

[
+v
xc
(r) (246)
We see that on the set B B
0
we have obtained the usual Kohn-Sham
equations
_

T +

V +

V
H
[n] +

V
xc
[n]
_
[
i
) = E
0
[
i
) (247)
where i = 1 . . . q runs over the q degenerate ground states [
i
) of the Kohn-
Sham system. The density must now be calculated from a ground state
ensemble

D
s
[n] =
q

i=1

i
[
i
)
i
[ (248)
of the Kohn-Sham system and is explicitly given as
n(r) = Tr

D
s
[n] n(r) =
q

i=1

i
[ n(r)[
i
) (249)
These equations can be written more explicitely in terms of the Kohn-Sham
orbitals dened to be the eigenstates of the single particle equation
_

1
2

2
+v
s
(r)
_

i
(r) =
i

i
(r) (250)
It is then easily seen that any Slater determinant
[) = [
i1
. . .
iN
) (251)
built out of Kohn-Sham orbitals is an eigenstate of the Kohn-Sham system
with eigenvalue E
0
=
i1
+ . . . +
iN
. However, not every eigenstate is
necessarily a Slater determinant. For instance, if two Slater determinants
have the same energy eigenvalue, then a linear combination of them also has
the same eigenvalue. This leads to a subtle point for Kohn-Sham theory of
degenerate states. In Eq.(249) we require that the density is representable
by a ground state ensemble of a noninteracting system. However, we did
not require that it must be representable by an ensemble of ground state
Slater determinants. That one should be careful at this point, has been
shown by Englisch and Englisch [3] and Lieb [1]. They constructed an
explicit example of a pure state density of a noninteracting system that
can not be obtained from a single Slater determinant. We will therefore in
general have
[
i
) =
q

j=1

ij
[D
j
) (252)
were [D
j
) is a Slater determinant. We further know that every [
i
) must
be a ground state for potential v
s
. This means that all orbitals below the
highest occupied level must be occupied. If this were not the case then we
could lower the energy by transferring an electron from the highest level to
a lower level which would lower the energy and therefore [
i
) would not be
a ground state for potential v
s
. This also means that if two determinants
[D
i
) and [D
j
) dier, then this dierence must be due to dierent orbitals
in the highest level. Because both determinants are eigenfunctions of the
Kohn-Sham Hamiltonian it is easily seen that these dierent orbitals must
have the same Kohn-Sham eigenvalue. This eigenvalue is the Kohn-Sham
orbital energy of the highest occupied state which we will denote by .
From these considerations we nd that the density is of the general form
n(r) =

i<
[
i
(r)[
2
+

k
=
l
=

kl

k
(r)

l
(r) (253)
where
kl
is an Hermitian matrix with eigenvalues between 0 and 1. This
means that we can diagonalize this matrix with a unitary matrix U
ij
. We
then introduce new orbitals for the highest occupied level:

i
(r) =

j
U
ij

j
(r) (254)
Since these orbitals are linear combinations of degenerate orbitals, they are
again eigenfunctions of the Kohn-Sham single-particle Hamiltonian for the
highest occupied level. We can now write the density as
n(r) =

i<
[
i
(r)[
2
+

k
=
n
k
[
k
(r)[
2
(255)
with occupation numbers n
k
between 0 and 1. We see that for the case of a
degenerate ground state we have to solve Eq.(250) together with Eq.(255)
in which also the coecients n
k
must be determined. When do we need
fractional occupation numbers in practice? In practice one sometimes nds
that when solving the Kohn-Sham equations with occupations equal to
one that the converged solution has a hole below the highest level. This
happens for instance for the iron atom within the local density approxima-
tion [24]. This is an indication that the true Kohn-Sham density could be
an E-V-density. In practice the lowest energy state is then obtained by a
procedure called evaporation of the hole. In this method one increases
the occupation of the hole while descreasing the occupation of the highest
level until both levels are degenerate [24, 25]. The matter has been in-
vestigated in detail by Schipper et al. [26, 27]. These authors calculated
by CI-methods accurate charge densities for a couple of molecules. Sub-
sequently they constructed the Kohn-Sham potentials that generate these
densities. A particularly interesting case is the C
2
molecule [26]. When
one tries to construct a Kohn-Sham potential that produces the ground
state density of the C
2
molecule using occupation numbers equal to one,
i.e. for a pure state, one nds that the solution corresponds to a state with
a hole below the highest occupied level. This is then not a proper Kohn-
Sham state since it is not the state of lowest energy for the corresponding
potential, but an excited state. If one then tries to construct an ensem-
ble solution by the technique of evaporation of the hole, one does nd a
proper ground state ensemble for the Kohn-Sham system. From this we can
conclude that the ground state density of the C
2
molecule is a noninteract-
ing E-V-density but not a noninteracting PS-V-density. This means that
the extension of density functionals to E-V-densities is not only of theoret-
ical, but also practical interest. As an illustration we display in Figure 4,
the exchange-correlation potentials of the C
2
molecule for four bonding dis-
tances. The potentials are plotted along the CC bond axis as functions of
the distance z from the bond midpoint. The corresponding bond distances
R(CC) are displayed in the inset. The potentials v
PShole
xc
correspond to
the exchange-correlation potentials of the improper Kohn-Sham solutions
with a hole. The corresponding state is a pure state consisting of a sin-
gle Slater determinant. The potentials v
E
xc
are the Kohn-Sham potentials
corresponding to a ground state ensemble. Both v
PShole
xc
and v
E
xc
give
an accurate representation of the true density of the molecule, although
the accuracy attained with v
E
xc
is a bit higher. The potential v
E
xc
has the
features found in many diatomic molecules, i.e. the usual well around the
atom, the small intershell peaks and a plateau around the bond midpoint.
The potential v
PShole
xc
on the other hand, is heavily distorted as compared
to v
E
xc
. This may be explained by the fact that the wavefunction of the C
2
molecule has a strong multideterminantial character. With this we mean
that the simple Hartree-Fock approximation has a relatively small coe-
cient in the CI-expansion of the wave function. The distortions in v
PShole
xc
therefore appear to be a price we have to pay for producing the density of
such a strongly multideterminantial state with a single Slater determinant.
The use of an ensemble formulation of Kohn-Sham theory seems therefore
especially relevant in cases of strong electron correlations. Other examples
are provided by near-degeneracy situations such as in avoided crossings [25]
and the H
2
+ H
2
reaction [27]. A recent discussion on Kohn-Sham theory
for ensembles can be found in Ref. [28] which also gives the beryllium series
as an explicit example of systems which are ensemble v-representable but
not pure state v-representable.
distances to be E-V
s
representable, (or class b, as
described in Sect. 1).
Applying the results of Levy [1] and Lieb [2] to
noninteracting systems, we know that there are many
densities that are not PS-V
s
representable but that are
E-V
s
representable. It is thus certainly possible that a
given interacting ground state density does not belong to
the set of PS noninteracting ground state densities, but
does belong to the set of ensemble-representable non-
interacting densities. Our results prove this not to be
only an academic possibility, but E-V
s
representability is
in fact called for to handle cases with strong electron
correlation, i.e., essentially multideterminantal character
of the interacting ground-state wavefunction.
The KS solution for the lowest singlet states of CH
2
and C
2
has been constructed from the ab initio CI q. To
obtain the ensemble solution in the cases when an at-
tempt to construct the PS leads to a non-Aufbau solu-
tion, the procedure of ``evaporation of the hole below
the Fermi level'' has been employed [15, 17, 37]. Already
for CH
2
, for which the wave function has a relatively
weak multideterminantal character, calculation in the
TZ basis yields the ensemble solution and only with the
extended QZ basis the pure-state one-determinantal KS
solution has been obtained.
With the example of C
2
, the possibility has been
demonstrated of an essentially accurate ensemble KS
solution with accidental degeneracy (E-V
s
represent-
ability) for a q that is PS-V representable for the inter-
acting system. A variety of KS solutions has been
obtained depending on the R(CC) distance, ranging
from a single-determinantal pure-state at R(CC) 1:8
a.u. through an ensemble with a weak accidental de-
generacy at R
e
to an ensemble with a strong degeneracy
Fig. 3ad The exchange-corre-
lation potential m
xc
along the
bond axis of C
2
. m
E
xc
(solid lines)
is the potential corresponding
to the ensemble solution;
m
PShole
xc
(dotted lines) corre-
sponds to the single-determi-
nant solutions with a hole
below the Fermi level. a R
e
, b
2.8 a.u., c 3.0 a.u., d 4.0 a.u.
341
Figure 4: Exchange-correlation potentials producing the accu-
rate interacting ground state density of the C
2
molecule at various
bonding distances. The potential v
E
xc
is the exchange-correlation
potential for a ground state ensemble. The potential v
PShole
xc
is
an exchange-correlation potential that reproduces the same den-
sity for a single Slater determinant with a hole.
16. The gradient expansion
Until now we only considered the formal framework of density functional
theory. However, the theory would be of little use if we would not be able to
construct good approximate functionals for the exchange-correlation energy
and exchange-correlation functional. Historically the rst approximation
for the exchange-correlation functional to be used was the local density
approximation. In this approximation the exchange-correlation functional
is taken to be
E
LDA
xc
[n] =
_
d
3
r
xc
(n(r)) (256)
where
xc
(n) is the exchange-correlation energy per volume unit for a homo-
geneous electron gas with density n. We therefore treat the inhomogeneous
system locally as an electron gas. This simple and crude approximation
turns out to be surprisingly succesful for realistic and very inhomogeneous
systems although one would expect that the LDA would only be accurate
for systems with slowly varying densities. The LDA therefore seems a suit-
able starting point for more accurate approximations. In this section we
will show that the LDA is in fact the rst term in a systematic expan-
sion of the exchange-correlation functional in terms of spatial derivatives
with respect to the density. This expansion is commonly referred to as the
gradient expansion [5, 29, 30] and has the following form
E
xc
[n] = E
LDA
xc
[n] +
_
d
3
rg
1
(n(r))(n(r))
2
+
_
d
3
rg
2
(n(r))(
2
n(r))
2
+. . . (257)
where the functions g
i
(n) are uniquely determined by the density response
functions (of arbitrary order) of the homogeneous electron gas. The gradi-
ent expansion presents an, in principle, exact way to construct the exchange-
correlation functional for solids, provided the series converges. We will come
back to the question of convergence. To derive the gradient expansion we
start out by expanding the exact exchange-correlation energy functional
around its homogeneous electron gas value n
0
as follows
E
xc
[n] = E
xc
[n
0
]+

m=1
1
m!
_
d
3m
rK
(m)
xc
(n
0
; r
1
. . . r
m
)n(r
1
) . . . n(r
m
) (258)
where d
3
r = d
3
r
1
. . . d
3
r
m
and n(r) = n
0
+n(r). We further dened
K
(m)
xc
(n
0
; r
1
. . . r
m
) =

m
E
xc
n(r
1
) . . . n(r
m
)
[
n=n0
(259)
The rst term E
xc
[n
0
] in Eq.(258) is the exchange-correlation energy of a
homogeneous system with constant density n
0
and is therefore a function of
n
0
rather than a functional. We will therefore write this term as E
xc
(n
0
).
This function is by now well-known from extensive investigations of the
homogeneous electron gas [31]. Since the electron gas has translational,
rotational and inversion symmetry the functions functions K
(m)
xc
satisfy
K
(m)
xc
(n
0
; r
1
. . . r
m
) = K
(m)
xc
(n
0
; r
1
+a. . . r
m
+a) (260)
K
(m)
xc
(n
0
; r
1
. . . r
m
) = K
(m)
xc
(n
0
; Rr
1
. . . Rr
m
) (261)
K
(m)
xc
(n
0
; r
1
. . . r
m
) = K
(m)
xc
(n
0
; r
1
. . . r
m
) (262)
where a is an arbitrary translation vector and where R is an arbitrary
rotation matrix. Furthermore, the function K
(m)
xc
has full permutational
symmetry, i.e.
K
(m)
xc
(n
0
; r
1
. . . r
m
) = K
(m)
xc
(n
0
; r
P(1)
. . . r
P(m)
) (263)
for an arbitrary permutation P of the indices 1 . . . m. If we choose a = r
1
in Eq. (260) we see that
K
(m)
xc
(n
0
; r
1
. . . r
m
) = K
(m)
xc
(n
0
; 0, r
2
r
1
. . . r
m
r
1
)
= L
(m)
(n
0
; r
2
r
1
. . . r
m
r
1
) (264)
where the latter equation denes the function L
(m)
as a function of m1
variables. Let us give some specic examples. The function K
(1)
xc
is given
by
K
(1)
xc
(n
0
; r
1
) =
E
xc
n(r
1
)
[
n=n0
= v
xc
(n
0
) (265)
and we see that this function is simply the exchange-correlation potential
of the electron gas which, due to translational invariance, does not depend
on r
1
. Since n(r) integrates to zero the term with K
(1)
xc
in the expansion
Eq.(258) does not contribute to E
xc
[n]. The rst nontrivial term is therefore
K
(2)
xc
(n
0
; r
1
, r
2
) =

2
E
xc
n(r
1
)n(r
2
)
[
n=n0
= L
(2)
(n
0
; r
2
r
1
) (266)
which leads to an expansion of the form
E
xc
[n] = E
xc
(n
0
) +
1
2
_
d
3
r
1
d
3
r
2
L
(2)
(n
0
; r
2
r
1
)n(r
1
)n(r
2
) +. . . (267)
The function L
(2)
(often denoted by K
xc
in the literature) has been the
subject of many investigations [32, 33]. Let us now go back to the more
general case. If we introduce the Fourier transform of a function f as follows

f(q
1
. . . q
m
) =
_
d
3m
rf(r
1
. . . r
m
)e
iq1r1...iqmrm
(268)
then

K
(m)
xc
(n
0
; q
1
. . . q
m
) = (2)
3
(q
1
+. . . +q
m
)

L
(m)
(n
0
; q
2
. . . q
m
) (269)
The expansion for E
xc
then becomes
E
xc
[n] = E
xc
(n
0
)+

m=1
1
m!
_
d
3m
q
(2)
3m

K
(m)
xc
(n
0
; q
1
. . . q
m
) n(q
1
) . . . n(q
m
) (270)
where d
3m
q = d
3
q
1
. . . d
3
q
m
. This can also be written in terms of the
functions L
(m)
as follows
E
xc
[n] = E
xc
(n
0
)+

m=1
1
m!
_
d
3
q
2
(2)
3
. . .
d
3
q
m
(2)
3

L
(m)
(n
0
; q
2
. . . q
m
)
n(q
2
. . . q
m
) n(q
2
) . . . n(q
m
) (271)
The gradient expansion is then obtained by expanding L
(m)
in powers of
q
i
around q
i
= 0 and subsequently transforming back to real space. If we
do this our nal expansion will then still depend on our reference density
n
0
which is undesirable for application of the functional to general systems.
However, we will show that we can get rid of this dependence on n
0
by
doing innite resummations. The key formula that will enable us to do
these resummations is the following rst order change of K
(m)
xc
:
K
(m)
xc
(r
1
. . . r
m
) =
_
d
3
rK
(m+1)
xc
(n
0
; r, r
1
. . . r
m
)n(r) (272)
If we take n(r) = n
0
to be a constant change we obtain
K
(m)
xc
n
0
(n
0
; r
1
. . . r
m
) =
_
d
3
rK
(m+1)
xc
(n
0
; r, r
1
. . . r
m
) (273)
Note that by taking n(r) = n
0
we violate the condition that n(r) inte-
grates to zero. Nevertheless Eq.(273) can alternatively be derived directly
from the properties of the response functions [30] and is related to a gener-
alized form of the well-known compressibility sumrule of the electron gas.
An important special case is obtained for m = 1:
v
xc
n
0
=
K
(1)
xc
n
0
(n
0
) =
_
d
3
r
2
K
(2)
xc
(n
0
; r
1
, r
2
)
=
_
d
3
r
2
L
(2)
(n
0
; r
2
r
1
) =

L
(2)
(q = 0) (274)
We see that the q = 0 value of

L
(2)
can be directly calculated from the
knowledge of v
xc
(n
0
). For m 2 relation (273) directly implies the follow-
ing relation between L
(m)
and L
(m+1)
L
(m)
n
0
(n
0
; q
2
. . . q
m
) = L
(m+1)
(n
0
; q
2
. . . q
m
, q
2
. . . q
m
) (275)
The importance of this formula will become clear if we work out, as an
explicit example, the lowest order in the gradient expansion. From the
symmetry properties of K
(m)
xc
one nds that its Fourier transform has the
following expansion in powers of q
i
:

K
(m)
xc
(n
0
; q
1
. . . q
n
) = (2)
3
(q
1
+. . . +q
m
)
(K
(m)
0
+K
(m)
1
P
(m)
1
(q
1
. . . q
m
) +K
(m)
2
P
(m)
2
(q
1
. . . q
m
)+. . .)(276)
in which the K
(m)
i
are coecients and P
(m)
i
are symmetric polynomials.
This is a direct consequence of the permutational symmetry of K
xc
. The
rst polynomial P
(m)
1
(taking into account the -function) has the explicit
form
P
(m)
1
(q
1
. . . q
m
) = q
2
1
+. . . +q
2
m
(277)
and is invariant under permutation of the indices, i.e. it transforms accord-
ing to a one-dimensional representation of the permutation group. More
details on the group theoretical treatment of these reponse functions can be
found in the work of Svendsen and von Barth [34]. From the properties of

K
(m)
xc
we nd that for m 2 the function L
(m)
has the following expansion
L
(m)
(n
0
; q
2
. . . q
m
) = L
(m)
0
(n
0
) +L
(m)
1
(n
0
)p
(m)
(q
2
. . . q
m
) +. . . (278)
where the polynomial p
(m)
is given by
p
(m)
(q
2
. . . q
m
) = (q
2
. . . q
m
)
2
+q
2
2
+. . . +q
2
m
= 2
m

i=2
q
2
i
2
m

i>j2
q
i
q
j
(279)
It is easily seen that this polynomial has the property
p
(m+1)
(q
2
. . . q
m
, q
2
. . . q
m
) = p
(m)
(q
2
. . . q
m
) (280)
If we use this property together with Eq.(278) and Eq.(275) we obtain for
the coecients in the expansion of L
(m)
the following equations
L
(m)
0
n
0
= L
(m+1)
0
(n
0
) (281)
L
(m)
1
n
0
= L
(m+1)
1
(n
0
) (282)
where m 2. Together with

L
(2)
(q = 0) =
v
xc
n
0
(n
0
) =

2

xc
n
2
0
(n
0
) (283)
where
xc
(n
0
) is the exchange energy per volume unit of the electron gas,
this yields
L
(m)
0
(n
0
) =

m

xc
n
m
0
(n
0
) (284)
L
(m)
1
(n
0
) =

m2
L
(2)
1
n
m2
0
(n
0
) (285)
It is these two equations that will allow us to eliminate the dependence on
n
0
in the lowest order gradient expansion. If we insert the explicit form of
L
(m)
of Eq.(278) into Eq.(271) and Fourier transform back to real space we
obtain
E
xc
[n] = E
xc
(n
0
)+

m=1
1
m!
L
(m)
0
(n
0
)
_
d
3
r(n(r))
m
+

m=2
1
m!
L
(m)
1
(n
0
)m(m1)
_
d
3
r(n(r))
2
(n(r))
m2
+. . .(286)
Since n(r) = n(r) this can be rewritten with help of Eq.(284) and
Eq.(285) as
E
xc
[n] = E
xc
(n
0
)+

m=1
1
m!

xc
(n
0
)
n
m
0
_
d
3
r(n(r))
m
+

m=2
1
(m2)!

m2
L
(2)
1
(n
0
)
n
m2
0
_
d
3
r(n(r))
2
(n(r))
m2
+. . .(287)
We see that the rst two terms simply give the expansion of the LDA
functional around the constant density n
0
,i.e.
E
LDA
xc
[n] =
_
d
3
r
xc
(n
0
+n(r)) (288)
whereas the third term involves an expansion of the coecient L
(2)
1
(n
0
+
n(r)) around n
0
. We obtain
E
xc
[n] = E
LDA
xc
[n] +
_
d
3
rL
(2)
1
(n(r))(n(r))
2
+. . . (289)
and we see that we succeeded in eliminating the dependence on n
0
in the
gradient expansion. Morever we see that the local density approximation
naturally appears as the rst term in the gradient expansion. Furthermore
we nd that the coecient g
1
(n) in our rst equation (257) is completely
determined by the q-expansion of the function

L
(2)
(n
0
; q) to order q
2
. We
also see that the replacement n
0
n(r) requires a resummation over re-
sponse functions of arbitrary order. The dependence on n
0
can also be
removed to higher order in powers of q. For an explicit example up to
order q
4
we refer again to Svendsen and von Barth [34].
Let us now address the question of convergence of the gradient expan-
sion. This question has been investigated for the gradient expansion of the
exchange-energy functional for which a comparison with exact exchange
energies is possible. For this case the analytic form of

L
(2)
(n
0
; q) is known
from the impressive work of Engel and Vosko [35, 36] and some higher order
gradient coecients have been determined by Svendsen, Springer and von
Barth [34, 37]. One nds [37] that the gradient expansion performs very
well for metallic systems, but that the quality deteriorates as soon as the
system acquires an energy gap. This may not be so surprising if one real-
izes that the appearance of a gap drastically changes the low q-behavior of
the response functions K
(m)
xc
which determine the gradient coecients [38].
This means that the standard gradient expansion can not deal with insu-
lators or nite systems (which may be modelled as insulators by repeating
them periodically with a large lattice constant). Further progress along the
lines of the gradient expansion may be obtained by study of the so-called
gapped electron gas [39]. Currently the most fruitful approach to the con-
struction of simple and accurate functionals is the so-called Generalized
Gradient Approximation (GGA) [40, 41]. In this approach one species a
form of the pair-correlation function of the many-electron system and de-
termines the parameters in this function by sumrules and information from
the straightforward gradient expansion. These functionals have lead to
large improvements in molecular binding energies as compared to the local
density approximation [42]. However, since the approach is not systematic
it is dicult to improve the quality of the current GGA functionals.
17. The optimized potential method and the e
2
-expansion
In this section we will describe a second systematic method to construct
density functionals, namely perturbation theory starting from the Kohn-
Sham Hamiltonian [43, 44]. The method is based on traditional pertur-
bation theory and is comparable in computational cost. For this reason
the method is less suited to the calculation of properties of large systems.
However, the method has the theoretical advantage that it can be used as a
benchmark to test the quality of dierent approximate density functionals.
We consider the Hamiltonian

=

H
s
+(

H

H
s
) =

H
s
+(

W

V
Hxc
) (290)
where

W =
N

i>j
e
2
[r
i
r
j
[
(291)

V
Hxc
=
N

i=1
v
H
(r
i
) +v
xc
(r
i
) (292)
where e
2
is the square of the electronic charge and v
H
is the usual Hartree
potential
v
H
(r) = e
2
_
d
3
r

n(r

)
[r r

[
(293)
and v
xc
the exchange-correlation potential. The perturbation term is simply
the dierence between the true and the Kohn-Sham Hamiltonian and we
are interested in the case = 1 although this makes expansion in powers
of rather doubtful. We will come back to the point of convergence later.
We can now do standard perturbation theory and obtain the ground state
energy
E() = E
s
+E
(1)
+
2
E
(2)
+. . . (294)
where E
s
is ground state energy of the auxiliary Kohn-Sham system, i.e.
simply a sum over orbital energies and the terms E
(i)
is the energy to order
i in powers of . The rst two terms are explicitely given by [43, 44]
E
(1)
=
s
[

W

V
Hxc
[
s
) (295)
E
(2)
=

i=1
[
s
[

W

V
Hxc
[
s,i
)[
2
E
s
E
s,i
(296)
where E
s,i
and
s,i
are the excited state energies and wave functions of
the Kohn-Sham system. We now note that the energies E
(i)
are implicit
functionals of the density through their dependence on the Kohn-Sham
potential, orbitals, and energies, i.e.
E
(i)
[n] = E
(i)
[
k
[n],
k
[n], v
xc
[n]] (297)
This follows directly from the Hohenberg-Kohn theorem applied to a non-
interacting system. The density n uniquely determines the Kohn-Sham
potential v
s
(up to a constant) and therefore the also the orbitals (up to
a phase factor) and eigenvalues (up to constant). The arbitrariness with
respect to a constant shift and with respect to the phase factor cancels out
in the energy expression and therefore the i-th order energy becomes a pure
density functional. We therefore have the following series of implications
n(r) v
s
(r)
k
(r),
k
E
(i)
(298)
Note that the perturbation theory that we constructed is not yet in a form
that we can use in practical calculations. This is because the perturbing
Hamiltonian contains the exchange-correlation potential which is unknown
from the start and has to be determined self-consistently from the pertur-
bation series. However, the equations can be simplied if we expand the
energies and potentials in powers of the interaction strength e
2
and take
= 1. This leads to the following set of equations [44]
E[n] = T
s
[n] +
_
d
3
rn(r)v(r)
+
e
2
2
_
d
3
rd
3
r

n(r)n(r

)
[r r

[
+

i=1
e
2i
E
(i)
xc
(299)
v
Hxc
(r) = v
H
(r) +

i=1
e
2i
v
(i)
xc
(r) (300)
v
(i)
xc
(r) =
E
(i)
xc
n(r)
(301)
The e
2
and e
4
terms in the expansion of the exchange-correlation energy
have the explicit form
e
2
E
(1)
xc
[n] =
s
[

W[
s
)
e
2
2
_
d
3
rd
3
r

n(r)n(r

)
[r r

[
(302)
and
e
4
E
(2)
xc
[n] =

i=1
[
s
[

W

V
H
e
2

V
(1)
xc
[
s,i
)[
2
E
s
E
s,i
(303)
The term of order e
2
can be written explicitly in terms of the Kohn-Sham
orbitals as follows:
E
(1)
xc
[n] =
1
2
N

k,l=1
_
d
3
rd
3
r

k
(r)
l
(r)

l
(r

)
k
(r

)
[r r

[
(304)
We see that this expression has exactly the same form as the usual expres-
sion of the exchange energy within the Hartree-Fock approximation. The
dierence, however, is that the orbitals in this expression are Kohn-Sham
orbitals which, in contrast to the Hartree-Fock orbitals, are eigenfunctions
of a single-particle Hamiltonian with a local potential. The numerical value
of E
(1)
xc
[n] and the Hartree-Fock exchange will therefore in general dier
from each other. However, because of the similarity to the Hartree-Fock
denition of exchange we will dene the exchange functional within density
functional theory to be E
x
[n] = E
(1)
xc
[n]. Corresponding to the exchange-
functional there is a local exchange potential dened as
v
x
(r) =
E
x
n(r)
=
E
(1)
xc
n(r)
= v
(1)
xc
(r) (305)
We see from Eq.(303) that we need to know this potential in order to
calculate the e
4
contribution to the exchange-correlation energy. This is
a general feature of the present perturbation theory. In order to calculate
E
(i)
xc
we need to calculate v
(i1)
xc
rst. Let us therefore start by calculating
v
x
. For this we have to calculate the rst order change E
x
in the exchange
functional due to a change n in the density. Since densities and potentials
are in 1-1-correspondence this task amounts to calculating the change in
E
x
due to a change v
s
in the Kohn-Sham potential. This is readily done
by perturbation theory. The change
k
and
k
of the orbitals and orbital
energies of the Kohn-Sham system due to a small change v
s
in the potential
is given by solution of the equation
_

1
2

2
+v
s
(r)
k
_

k
(r) = (
k
v
s
(r))
k
(r) (306)
This gives

k
=
_
d
3
r

k
(r)
k
(r)v
s
(r) (307)

k
(r) =
_
d
3
r

G
k
(r, r

)
k
(r

)v
s
(r

) (308)
G
k
(r, r

) =

l=k

l
(r)

l
(r

l

k
(309)
From these equations we obtain the following functional derivatives

k
v
s
(r)
=

k
(r)
k
(r) (310)

k
(r)
v
s
(r

)
= G
k
(r, r

)
k
(r

) (311)
The density change n is given by
n(r) =
N

k=1

k
(r)
k
(r) +c.c. =
_
d
3
r

s
(r, r

)v
s
(r

) (312)
where we dened the static density response function
s
of the Kohn-Sham
system by

s
(r, r

) =
n(r)
v
s
(r

)
=
N

k=1

k
(r)G
k
(r, r

)
k
(r

) +c.c. (313)
We can now readily derive the following integral equation for v
x
:

x
(r) =
E
x
v
s
(r)
=
_
d
3
r

E
x
n(r

)
n(r

)
v
s
(r)
=
_
d
3
r

s
(r, r

)v
x
(r

) (314)
where the inhomogeneity
x
is given by

x
(r) =
N

k=1
_
d
3
r

E
x

k
(r

)
G
k
(r

, r)
k
(r) +c.c. (315)
Since both
x
and
s
are given as explicit functionals of the Kohn-Sham
orbitals and orbital energies, the exchange potential can be found from
simultaneous solution of the equations

k
(r) =
_

1
2

2
+v(r) +v
H
(r) +e
2
v
x
(r)
_

k
(r) (316)

x
(r) =
_
d
3
r

s
(r, r

)v
x
(r

) (317)
Once we have solved these equations and determined v
x
we can go on
and calculate E
(2)
xc
from Eq.(303). To calculate the e
6
contribution to the
exchange-correlation energy we rst have to evaluate v
(2)
xc
. This potential
is the solution of the integral equation

(2)
xc
(r) =
E
(2)
xc
v
s
(r)
=
_
d
3
r

s
(r, r

)v
(2)
xc
(r

) (318)
Since E
(2)
xc
is an explicit functional of the orbitals, orbital energies and v
x
the inhomogeneity
(2)
xc
can be calculated from

(2)
xc
(r) =
_

k=1
_
d
3
r

E
(2)
xc

k
(r

k
(r

)
v
s
(r)
+c.c.
_
+

k=1
_
d
3
r

E
(2)
xc

k
v
s
(r)
+
_
d
3
r

E
(2)
xc
v
x
(r

)
v
x
(r

)
v
s
(r)
(319)
All terms in this equation are explicitly known, except for the term v
x
/v
s
.
However, for this term we can nd an integral equation by dierentiation
of Eq.(317) with respect to v
s
:

x
(r
1
)
v
s
(r
2
)
=
_
d
3
r
3

s
(r
1
, r
3
)
v
s
(r
2
)
v
x
(r
3
) +
_
d
3
r
3

s
(r
1
, r
3
)
v
x
(r
3
)
v
s
(r
2
)
(320)
Since both
x
and
s
are explicitly known in terms of orbitals and orbital
energies their derivatives with respect to v
s
are also explicitly known in
terms of these quantities (see reference [44] for more details) and therefore
Eq.(320) determines v
x
/v
s
uniquely. So we see that in order to deter-
mine v
(2)
xc
we have to solve two integral equations. For realistic systems
these equations have only be solved approximately (for an explicit solution
of Eq.(320) see reference [34])). From v
(2)
xc
one could go on along the same
lines to determine E
(3)
xc
and subsequently v
(3)
xc
. The determination of these
higher order energies and potentials becomes more involved. Nevertheless,
the perturbation series represents an explicit construction procedure for the
exact exchange-correlation energy and potential, provided that the series
converges. Before we go on to discuss the convergence properties of this
series, let us briey review the rst order equations from a dierent view-
point. It is readily seen that one can write the total energy to order e
2
,
which we denote by E
1
[n], as
E
1
[n] = T
s
[n] +
_
d
3
rn(r)v(r) +
e
2
2
_
d
3
rd
3
r

n(r)n(r

)
[r r

[
+e
2
E
x
[n]
=
s
[

H[
s
) (321)
This is just the expectation value of the true Hamiltonian

H of the system
with the Kohn-Sham wave function [
s
). Because of the 1-1-correspondence
between the density n and the potential v
s
we can also regard E
1
as a
functional of the potential v
s
, i.e.
E
1
[v
s
] =
s
[v
s
][

H[
s
[v
s
]) (322)
We may now try to nd an approximation to the true total energy of the
system by choosing a local potential v
s
that minimizes the energy expression
E
1
[v
s
]. This means that we have to solve the variational equation
0 =
E
1
v
s
(r)
=
N

k=1
_
d
3
r

E
1

k
(r

k
(r

)
v
s
(r)
+c.c. (323)
If one works out this expression one obtains equations that are identical
to Eqns.(316) and (317). These equations were rst derived by Talman
and Shadwick [45]. Since in our procedure we optimized the energy of
a Slater determinant wavefunction under the constraint that the orbitals
in the Slater determinant come from a local potential, the method is also
known as the Optimized Potential Method (OPM). We have therefore ob-
tained the result that the OPM and and the expansion to order e
2
are
equivalent procedures. The OPM has many similarities to the Hartree-
Fock approach. Within the Hartree-Fock approximation one minimizes the
energy of a Slater determinant wavefunction under the constraint that the
orbitals are orthonormal. One then obtains one-particle equations for the
orbitals that contain a nonlocal potential. Within the OPM, on the other
hand, one adds the additional requirement that the orbitals must satisfy
single-particle equations with a local potential. Due to this constraint the
OPM total energy E
1
will in general be higher than the Hartree-Fock en-
ergy E
HF
, i.e. E
1
E
HF
. We refer to [46, 47] for an application of the
OPM method for molecules.
We nally make some comments on the calculation of functional derivatives
in this section. We stress this point since careless use of the chain rule for
dierentiation has led to wrong results in the literature [48]. As an example
we consider the exchange functional E
x
. When we regard this functional
as an explicit functional of the orbitals then the functional derivative with
respect to
k
is given by
_
E
x

k
(r)
_
e
=
N

i=1
_
d
3
r

i
(r)
i
(r

)
[r r

k
(r

) (324)
where we used the subindex e to indicate that we regard the functional
as an explicit orbital functional. This functional derivative represents the
change in E
x
due to a change
k
in orbital
k
, while keeping all other
orbitals xed. Moreover, we regard
k
and

k
as independent variables. If
we regard E
x
as a density functional we must require that all orbitals are
eigenfunctions of a noninteracting Hamiltonian with a local potential v
s
. It
is clear that we can we can never nd a change v
s
in a local potential that
induces a change in only one orbital while keeping the other orbitals xed.
If a potential changes one orbital
k
to
k
+
k
then all other orbitals will
change too. The change in E
x
regarded as a functional of v
s
is then given
by the functional derivative
E
x
v
s
(r)
=
N

k=1
_
d
3
r

_
E
x

k
(r

)
_
e

k
(r

)
v
s
(r)
+
N

k=1
_
d
3
r

_
E
x

k
(r

)
_
e

k
(r

)
v
s
(r)
(325)
We see that in this case the explicit orbital derivatives only occur in the
sum in which all the orbital changes must be taken into account. We may
also introduce an implicit derivative (E
x
/
k
)
i
, which can be given the
meaning of giving the change in E
x
if we know that there is a potential
change v
s
that changes
k
to
k
+
k
. This implicit orbital derivative,
which we denote by subidex i is expressed in terms of the explicit orbital
derivatives as follows
_
E
x

k
(r)
_
i
=
N

i=1
_
d
3
r

_
E
x

i
(r

)
_
e

i
(r

k
(r)
+
N

i=1
_
d
3
r

_
E
x

i
(r

)
_
e

i
(r

k
(r)
(326)
The derivatives
i
/
k
appear now to take into account that, if orbital
k
changes to
k
+
k
by some potential change v
s
, then the other orbitals

i
change to
i
+
i
. By using the implicit orbital derivative we regard
E
x
as a functional of the potential, or equivalently of the density, and we
can therefore use the chain rule
_
E
x

k
(r)
_
i
=
_
d
3
r

E
x
n(r

)
_
n(r

k
(r)
_
i
=
_
d
3
r

v
x
(r

)
_
n(r

k
(r)
_
i
(327)
In this equation (n/
k
)
i
is also an implicit derivative given by:
_
n(r

k
(r)
_
i
=
N

i=1
_
d
3
r

i
(r

i
(r

k
(r)
+
i
(r

i
(r

k
(r)
_
(328)
which should be compared to the explicit derivative
_
n(r

k
(r)
_
e
=

k
(r) (329)
for k = 1 . . . N. We stress that Eq.(327) is only true when using the im-
plicit derivatives. If in Eq.(327) we replace the implicit derivatives by the
explicit ones of Eq.(324) and Eq.(329) we obtain a result that is wrong. As
pointed out this is due to the fact that a change in just one orbital can
not be induced by a change in a local potential which in density functional
theory is in 1-1-correspondence with the density. Most of the confusion can
be avoided by regarding all functionals as functionals of the potential v
s
and by calculating the change in the functionals by means of perturbation
theory. In this way one can avoid use of the implicit orbital derivatives as
in Eq.(326).
Let us nally come back to the question of convergence of the perturbation
series. The perturbation series presented in this section is very similar to
Mller-Plesset perturbation theory starting from the Hartree-Fock approx-
imation. For the Mller-Plesset perturbation theory it is known that it is
in general divergent [49, 50, 51]. However, it is well-known that when car-
ried out to low orders this perturbation theory gives reasonable answers.
The Mller-Plesset perturbation series has therefore all the features of an
asymptotic series [52]. Since the perturbation series in this section is very
similar to the Mller-Plesset series it will in general also diverge. This
has indeed been found in a perturbation theory on the basis of some ap-
proximate Kohn-Sham Hamiltonians [53]. Nevertheless, it has been found
for the method presented in this section that low orders in perturbation
theory give good results [54] and we conclude that our series give at least
an asymptotic expansion for the exchange-correlation energy and potential.
18. Outlook and conclusions
In this work we have given on overview of the mathematical foundations
of stationary density functional theory. We discussed in great detail the
question of dierentiability of the functionals and showed that the Kohn-
Sham theory can be put on a solid basis for all practical purposes, since
the set of noninteracting E-V-densities is dense in the set of interacting
E-V-densities. The question whether these two sets are in fact identical is
still an open question. We further discussed two systematic approaches for
the construction of the exchange-correlation functional and potential.
What can we say about future developments within density functional the-
ory? There have been many extensions of density functional theory in-
volving spins, relativistic eects, temperature, superconductivity and time-
dependent phenomena. The last few years we have, for instance, seen many
applications of response properties, rather than ground state properties, us-
ing time-dependent density functional theory. For these extended density
functional theories it is, of course, more dicult to provide a rigorous the-
oretical basis. This is, however, not particular to density functional theory,
but applies to any method that deals with many-body systems, especially if
one is interested in phenomena like superconductivity or interactions with
laser elds. Nevertheless, also for these complicated cases simple density
functionals have provided encouraging results although there is still a clear
need for more accurate density functionals. In view of the success of den-
sity functional theory for ground state calculations it seems worthwhile to
explore these new areas.
19. Acknowledgments
I would like to thank Evert Jan Baerends, Pieter Schipper and Oleg Grit-
senko for providing me with Fig.4.
References
[1] E.H.Lieb Int.J.Quant.Chem. 24, 243 (1983)
[2] H.Englisch and R.Englisch Phys.Stat.Solidi B123, 711 (1984)
[3] H.Englisch and R.Englisch Phys.Stat.Solidi B124, 373 (1984)
[4] G.de Barra Introduction to Measure Theory Van Nostrand Reinhold,
London (1974)
[5] P.Hohenberg and W.Kohn Phys.Rev., 136B, 864 (1964)
[6] C.De Dominicis and P.C.Martin J.Math.Phys., 5, 14 (1964)
[7] H.Hellmann, Einf uhrung in die Quantenchemie, Leipzig (1937) and
R.P.Feynman Phys.Rev 56, 340 (1939)
[8] W.Kohn and L.J.Sham Phys.Rev. 140, A1133 (1965)
[9] H.Jeggle Nichtlineare Funktionalanalysis Teubner, Stuttgart (1979).
[10] L.B.Rall Nonlinear functional analysis and applications Academic
Press, New York, (1971)
[11] I.Ekeland and R.Temam Convex analysis and variational problems
North-Holland Publishing, Amsterdam (1976)
[12] M.M.Vainberg Variational method and method of monotone operators
in the theory of nonlinear equations John Wiley, New York (1973)
[13] M.C.Joshi and R.K.Bose Some topics in nonlinear functional analysis
Wiley Eastern, New Delhi (1985).
[14] A.H.Siddiqi Functional analysis with applications McGraw-Hill, New
Delhi (1986)
[15] R.van Leeuwen Int.J.Mod.Phys. B15,1969 (2001)
[16] H.Englisch and R.Englisch Physica 121A, 253 (1983)
[17] M.Levy Phys.Rev. A26, 1200 (1982)
[18] F.Aryasetiawan and M.J.Stott Phys.Rev. B38, 2974 (1988)
[19] M.Levy Proc.Nat.Acad.Sci.USA 76, 6062 (1979)
[20] D.Werner Funktionalanalysis, Springer, Berlin (1995)
[21] W.Rudin Functional Analysis, McGraw-Hill, New Delhi (1973)
[22] R.B.Israel Convexity in the Theory of Lattice Gases Princeton Univer-
sity Press, Princeton, New Jersey (1979)
[23] J.T.Chayes, L.Chayes and M.B.Ruskai J.Stat.Phys. 38, 497 (1985)
[24] F.W.Averill and G.S.Painter Phys.Rev. B15, 2498 (1992)
[25] S.G.Wang and W.H.E.Schwarz J.Chem.Phys. 105, 4641 (1996)
[26] P.R.T.Schipper, O.V.Gritsenko and E.J.Baerends Theor.Chem.Acc.
99, 329 (1998)
[27] P.R.T.Schipper, O.V.Gritsenko and E.J.Baerends J.Chem.Phys. 111,
4056 (1999)
[28] C.A.Ullrich and W.Kohn Phys.Rev.Lett. 87, 093001 (2001)
[29] D.J.W.Geldart in
Topics in Current Chemistry 180, 31, ed.R.F.Nalewajski, Springer,
Berlin (1996)
[30] Y.Osaka J.Phys.Soc.Japan 36,376 (1974)
[31] R.M.Dreizler and E.K.U.Gross Density Functional Theory: An Ap-
proach to the Many-Body Problem Springer, Berlin (1990)
[32] D.C.Langreth and S.H.Vosko Adv.Quant.Chem. 21, 175 (1990)
[33] E.Engel Phys.Rev. A51, 1159 (1995)
[34] P.S.Svendsen and U.von Barth Phys.Rev. B54, 17402 (1996)
[35] E.Engel and S.H.Vosko Phys.Rev. B42, 4940 (1990)
[36] M.L.Glasser Phys.Rev. B51, 7283 (1995)
[37] M.Springer, P.S.Svendsen and U.von Barth Phys.Rev. B54, 17392
(1996)
[38] M.Rasolt Phys.Rev. B36, 5041 (1987)
[39] J.Rey and A.Savin Int.J.Quant.Chem. 69, 581 (1998)
[40] K.Burke, J.P.Perdew and Y.Wang in Electronic Density Func-
tional Theory: Recent Progress and New Directions eds.J.F.Dobson,
G.Vignale and M.P.Das, Plenum, New York (1997)
[41] J.P.Perdew, K.Burke and Y.Wang Phys.Rev. B54, 16533 (1996)
[42] H.L.Schmider and A.D.Becke J.Chem.Phys. 109, 8188 (1998)
[43] A.Gorling and M.Levy Phys.Rev. A50, 196 (1994)
[44] E.Engel and R.M.Dreizler J.Comp.Chem. 20, 31 (1999)
[45] J.D.Talman and W.F.Shadwick Phys.Rev. A14, 36 (1976)
[46] S.Ivanov, S.Hirata and R.J.Bartlett Phys.Rev.Lett. 83, 5455 (1999)
[47] A.Gorling Phys.Rev.Lett. 83, 5459 (1999)
[48] R.K.Nesbet and R.Colle Phys.Rev. A61, 012503 (1999)
[49] O.Christiansen, J.Olsen, P.Jrgensen, H.Koch and P.-

A. Malmqvist
Chem.Phys.Lett. 261, 369 (1996)
[50] J.Olsen, P.Jrgensen, T.Helgaker and O.Christiansen J.Chem.Phys.
112, 9736 (2000)
[51] F.H.Stillinger J.Chem.Phys. 112, 9711 (2000)
[52] C.M.Bender and S.A.Orszag Advanced Mathematical Methods for Sci-
entists and Engineers, McGraw-Hill, Singapore (1978)
[53] M.Warken Chem.Phys.Lett. 237, 256 (1995)
[54] A.FaccoBonetti, E.Engel, R.N.Schmid and R.M.Dreizler
Phys.Rev.Lett. 86, 2241 (2001)