Introduction to Density Functional

Theory of Classical Systems: Theory

and Applications

Lecturenotes

by
Roland Roth
ITAP, Universität Stuttgart
and
Max-Planck-Institut für Metallforschung, Stuttgart
Germany

1
Fukuoka, November 2006
2
Contents

1 Introduction 5

2 Basics of Density Functional Theory 7

2.1 Short history of DFT . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Statistical mechanics in the grand canonical ensemble . . . . . . . . . 8
2.3 Functional of the “grand potential” . . . . . . . . . . . . . . . . . . . 10
2.4 Gibbs inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Hohenberg-Kohn-Mermin variational principle . . . . . . . . . . . . 12
2.6 Classical analog to Kohn-Sham equations . . . . . . . . . . . . . . . 15
2.7 Generalization to Mixtures . . . . . . . . . . . . . . . . . . . . . . . 16
2.8 Excess free energy Fex . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Fundamental measure theory 17

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Exact result in d = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Rosenfeld’s Fundamental Measure Theory (d = 3) . . . . . . . . . . 19
3.4 The White Bear Version of FMT . . . . . . . . . . . . . . . . . . . . 25
3.5 Test for self-consistency . . . . . . . . . . . . . . . . . . . . . . . . 26
3.6 The White Bear Version of FMT Mark II . . . . . . . . . . . . . . . . 27

4 Application 31
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Hard-Sphere Fluid at a Hard Wall . . . . . . . . . . . . . . . . . . . 31
4.2.1 Minimizing Ω[ρ] through a Picard Iteration . . . . . . . . . . 32
4.2.2 Weighted Densities . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.3 One-Body Direct Correlation c(1) (z) . . . . . . . . . . . . . . 34
4.2.4 Hard-Sphere Fluid at a Hard Wall: the density profile . . . . . 35
4.3 Square-Well Fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.1 Bulk Fluid Phase Diagram . . . . . . . . . . . . . . . . . . . 37
4.3.2 Free Interface . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3
4 CONTENTS

4.3.3 Surface Tension of the Free Interface . . . . . . . . . . . . . 40

4.3.4 Square-Well Fluid at a Hard Wall . . . . . . . . . . . . . . . 41
4.3.5 Square-Well Fluid in a Slit . . . . . . . . . . . . . . . . . . . 44
Chapter 1

Introduction

In this lecture I want to cover the basics of density functional theory of classical sys-
tems and want to give a flavor of its possible applications.
Density functional theory started as a theory for electrons. Walter Kohn could
show that instead of solving the N-particle Schrödinger equation, it is possible to
obtain all the information of the ground state (T = 0) of an electron system from its
one-particle density distribution. He went on to show that there exists a functional of
the ground state energy that can be written as a functional of the density distribution.
This functional possesses two important properties: (i) for the ground state one-particle
density distribution this functional recovers the ground state energy of the system, and
(ii) for any other one-particle density distribution the functional takes a values that is
larger than the ground state energy. Density functional theory was born. About the
same time, in the mid 1960s, Mermin showed that these ideas also hold for an electron
system at temperature T > 0. His formulation of the proof of density function theory
was then re-casted for classical systems, i.e. statistical systems that obey the rules of
classical mechanics.
While not the first to apply density functional theory to problems of classical statis-
tics, Bob Evans was one who spread the word by his review paper on the gas-liquid in-
terface [4] in which he introduced the formalism of density functional theory to a broad
audience. This paper was also my first contact to density functional theory. Closely
following Evans’ review, we will introduce the formalism of density functional theory
in Chapter 2.
Beside the formalism there are the applications of density functional theory. For
several systems of great interest there are now reliable and powerful functionals avail-
able. Unfortunately, it is in general not possible to construct a density functional from
the knowledge of the interparticle interactions alone. In order to construct a func-
tional one needs insight and intuition. One elegant and very successful approach in
density functional theory of classical systems is the fundamental measure theory for

5
6 CHAPTER 1. INTRODUCTION

hard-sphere mixtures by Yasha Rosenfeld [6]. We will take a close look at this theory
in Chapter 3.
Finally, we discuss some typical applications of density functional theory in Chap-
ter 4. These application should demonstrate one of the key points of density functional
theory: once a functional for the excess free energy has been found, it is possible to
study a large variety of phenomena simply by changing the external potential acting
on the system under consideration. We show this for a square-well fluid for which we
study the free interface, the fluid at a single planar hard wall, where we observe the
drying transition (wetting by the gas phase), and the fluid in a slit geometry, where we
observe the capillary evaporation transition.
My hope is it that you as participant of this lecture and reader of this lecture notes
get a flavor of what density functional is, how one particular type of functional (the
excess free energy functional of the fundamental measure theory) looks like and what
it can do. Clearly, it is impossible to cover the whole field of density functional is just
a few lectures or on a few pages. The selection of the material reflects my personal
experience and taste. After this lecture, however, it should be possible for you to read
and understand the literature on density functional theory and its applications.
I would like to thank Prof. Ryo Akiyama for inviting me to give these lectures
at the Kyushu University in Fukuoka, Japan, and the Front Researcher Development
Program of the Kyushu University for their support.

Fukuoka, November 2006 Roland Roth

Chapter 2

Basics of Density Functional Theory

2.1 Short history of DFT

Walter Kohn

• 1964: Hohenberg and Kohn (HK) variational principle for the inhomogeneous
electron gas at T = 0 (P. Hohenberg and W. Kohn, Inhomogeneous Electron
Gas, Phys. Rev. 136, B 864 (1964))

7
8 CHAPTER 2. BASICS OF DENSITY FUNCTIONAL THEORY

– electron density n(r) in the ground state |Ψi as basic variable

n(r) = hΨ|Φ∗ (r)Φ(r)|Ψi

– electron density n(r) determines uniquely the external potential Vext (r)
– it exists an unique energy functional Ev [n] with the following properties
that Ev [n0 ] = E0 and Ev [n 6= n0 ] > E0 .

• 1965: Mermin formulates HK for T > 0 (N. D. Mermin, Thermal Properties of

the Inhomogeneous Electron Gas, Phys. Rev. 137, A 1441 (1965))

• 1965: Kohn and Sham equations (Kohn and Sham, Self-Consistent Equations
Including Exchange and Correlation Effects, Phys. Rev. 140, A 1133 (1965)).

• around 1976: application of DFT to classical systems

• 1998: Nobel Prize in Chemistry for W. Kohn for his development of the Density
Functional Theory

2.2 Statistical mechanics in the grand canonical ensem-

ble
We start by considering a classical system of N identical particles. The generalization
to mixtures is straightforward. Each particle has mass m and is located at position ri
and has momentum pi , i = 1, . . . N. The N-particle Hamiltonian is given by

HN = Tkin + U + Vext ,

with the kinetic energy

N
X p2i
Tkin = ,
i=1 2m
the potential energy of interparticle interaction

U = U(r1 , . . . , rN ),

and potential energy due to external potential

N
X
Vext = Vext (ri ).
i=1

Using this Hamiltonian one can calculate the grand canonical partition sum

Zgc ≡ T rcl exp(−β(HN − µN)), (2.1)

2.2. STATISTICAL MECHANICS IN THE GRAND CANONICAL ENSEMBLE 9

with β = 1/kB T , where kB is Boltzmann’s constant and T the absolute temperature.

µ is the chemical potential. Note that in the grand canonical ensemble the system of
interest has a volume V , is coupled to a heat bath at temperature T , and to a particle
reservoir with chemical potential µ. In Eq. (2.1) we have used the classical trace T rcl
as a shorthand notation for the integral over all particle momenta, all particle positions
and the sum over all possible particle numbers N = 0, . . . , ∞. Explicitly, we have
∞
1
Z Z Z Z
d 3 r1 . . . d 3 rN d 3 p1 . . . d 3 pN ,
X
T rcl = 3N N!
N =0 h

where h is Planck’s constant.

The grand canonical partition sum Zgc contains all the information of the system
in thermal equilibrium. From Zgc we can (in principle) calculate everything. If it
would be possible to calculate Zgc exactly, we would not require the formalism of DFT.
However, in general it is not possible to calculate Zgc easily, especially for arbitrary
external fields Vext (r). It turns out that it is simpler to make useful approximations
within the framework of density functional theory than in the evaluation of the partition
sum.
With the help of the partition sum we can define the equilibrium probability density
f0 for N particles at temperature T at particle positions ri with momenta pi , i =
1, . . . , N
1
f0 ≡ exp(−β(HN − µN)). (2.2)
Zgc
The definition of the probability density is such that the classical trace over the proba-
bility distribution gives unity, i.e.
T rcl f0 = 1
Using the equilibrium probability density f0 we can calculate the (ensemble) averages
of operators Ô by
hÔi ≡ T rcl f0 Ô.
One example of such an average, that we will need in the following is the average
equilibrium density distribution ρ0 (r) which can be written as the ensemble average
over the density operator
ρ0 (r) = hρ̂(r)i, (2.3)
with the density operator
N
X
ρ̂(r) = δ(r − ri ).
i=1
Finally, we note that the grand potential Ω of the system and the grand canonical
partition sum are related via
βΩ = − ln Zgc ,
10 CHAPTER 2. BASICS OF DENSITY FUNCTIONAL THEORY

which is simply the definition of the grand potential.

2.3 Functional of the “grand potential”

Following Mermin [3] we consider the following functional of a probability density f .
As mentioned before, we require that the probability density is normalized T rcl f = 1.
The functional [3] is given by

Ω[f ] = T rcl f (HN − µN + β −1 ln f ). (2.4)

It possesses the important feature that for the the equilibrium probability density, given
by Eq. (2.2) the functional reduces to the grand potential of the system Ω, as can be
seen easily from
 
Ω[f0 ] = T rcl f0 HN − µN + β −1 ln f0
 
= T rcl f0 −β −1 ln Zgc
= −β −1 ln Zgc
≡ Ω.

If we now consider a probability distribution different from the equilibrium distribu-

tion, i.e. f 6= f0 , T rcl f = 1, and we evaluate the functional, Eq. (2.4), we find the
inequality
 
Ω[f ] = T rcl f HN − µN + β −1 ln f
 
= T rcl f Ω[f0 ] + β −1 ln f − β −1 ln f0
= Ω[f0 ] + β −1 T rcl f (ln f − ln f0 )
> Ω[f0 ],

which is an important result, as we shall see in the following. The variational principle
of density functional theory is based on this result. In order to show that the inequality
holds we used

HN − µN = −β −1 ln (f0 Zgc )
= −β −1 ln f0 − β −1 ln Zgc
= −β −1 ln f0 + Ω[f0 ],

which follows directly from the definition of f0 in Eq. (2.2), and the Gibbs inequality,
which we shall discuss in the next Section.
2.4. GIBBS INEQUALITY 11

2.4 Gibbs inequality

Here we show that for any two probability densities f1 and f2 , with T rcl fi = 1, i = 1, 2
we have the inequality
T rcl (f1 ln f1 − f1 ln f2 ) ≥ 0, (2.5)
and the equality holds only if f1 = f2 . In order to show the inequality we rewrite
Eq. (2.5) as

T rcl (f1 ln f1 − f1 ln f2 ) = T rcl f1 (ln f1 − ln f2 )

f1
= T rcl f1 ln
f2
!
f1 f1
= T rcl f2 ln .
f2 f2

Next we observe that there is a complicated way of writing a zero by noting that
!
f1
T rcl f2 = T rcl f1 = 1,
f2

so that we find !
f1
T rcl f2 1− ≡0
f2
Using this observation we can conclude that
! !
f1 f1 f1 f1 f1
T rcl f2 ln = T rcl f2 ln + 1 − ≥ 0. (2.6)
f2 f2 f2 f2 f2

In order to see that we obtain this inequality we introduce the variable x = f1 /f2 and
rewrite Eq. (2.6) in terms of x as

hx ln x − (x − 1)i ≥ 0

because
x ln x ≥ x − 1,
as can be seen in Fig. 2.1. Hence we obtain

hx ln xi ≥ hx − 1i

but we have seen before that

hx − 1i = 0.
It is possible to make the statement of the Gibbs inequality even stronger by noting
that x ln x = x − 1 only for x ≡ f1 /f2 = 1. This, however, implies that equality holds
12 CHAPTER 2. BASICS OF DENSITY FUNCTIONAL THEORY

3
x ln(x)
x-1

-1
0 0.5 1 1.5 2 2.5 3

Figure 2.1: x ln x (full line) is greater or equal x − 1 (dotted line), as can be seen in
this plot. The equality x ln x = x − 1 holds only for x = 1, denoted by the circle.

only if f1 = f2 . Thus, if we consider two different probability distributions f1 6= f2 we

can conclude that
T rcl (f1 ln f1 − f1 ln f2 ) > 0.

As a consequence we can deduce that for any a probability distribution different

from the equilibrium distribution, i.e. f 6= f0 , T rcl f = 1 we obtain the result of the
last section
Ω[f ] > Ω[f0 ] ≡ Ω. (2.7)

2.5 Hohenberg-Kohn-Mermin variational principle

Through the Hamiltonian HN , the equilibrium probability distribution f0 , Eq. (2.2),
becomes a functional of the external potential Vext (r). It follows that the equilibrium
density distribution ρ0 (r) also becomes a functional of Vext (r) though Eq. (2.3). The
2.5. HOHENBERG-KOHN-MERMIN VARIATIONAL PRINCIPLE 13

next step is to show that f0 is a functional of ρ0 (r). This can be concluded from the
fact that the external potential Vext (r) is uniquely determined by ρ0 (r).
To show this we assume for a moment that a second external field Vext(r)′ 6=
Vext (r) gives rise to the same equilibrium density profile ρ0 (r) (at the same µ and tem-
perature T ) and construct a contradiction. The external potentials Vext (r) and Vext (r)′
give rise to the Hamiltonians

HN = Tkin + U + Vext , (2.8)

HN′ = Tkin + U + Vext
′
.

We also can rewrite HN′ as

HN′ = HN − Vext + Vext

′
. (2.9)

With the help of HN we can define the equilibrium probability distribution f0 and
with the help HN′ we define f ′ 6= f0 . If we evaluate the functional of the grand
potential, Eq. (2.4), for the probability distribution f ′ we obtain with the help of the
Gibbs inequality

Ω[f ′ ] = T rcl f ′ (HN′ − µN + β −1 ln f ′ )

< T rcl f0 (HN′ − µN + β −1 ln f0 ).

By using Eq. (2.9) the r.h.s. of the inequality can be written first as

T rcl f0 (HN′ − µN + β −1 ln f0 ) = T rcl f0 (HN − µN + β −1 ln f0 − Vext + Vext

′
)

and finally as
Z
Ω[f0 ] + ′
T rcl f0 (Vext − Vext ) = Ω[f0 ] + d3 rρ0 (r) [Vext
′
(r) − Vext (r)] , (2.10)

where we have made use of the definition of the equilibrium density distribution,
Eq. (2.3). Now we can also evaluate the functional for the distribution f0 :

Ω[f0 ] = T rcl f0 (HN − µN + β −1 ln f0 )

< T rcl f ′ (HN − µN + β −1 ln f ′ )

Again, we follow the same steps as before and rewrite the r.h.s. of the inequality. In
addition we make use of our assumption that the distribution f ′ gives rise to the same
equilibrium density distribution ρ0 (r) as f0 . We conclude that the r.h.s. gives
Z
′ ′
Ω[f ] + T rcl f (Vext − ′
Vext ) ′
= Ω[f ] + d3 rρ0 (r) [Vext (r) − Vext
′
(r)] . (2.11)
14 CHAPTER 2. BASICS OF DENSITY FUNCTIONAL THEORY

By adding Eqs. (2.10) and (2.11) we see that

Ω[f0 ] + Ω[f ′ ] < Ω[f0 ] + Ω[f ′ ]

follows, which cannot be true. Therefore the assumption that Vext (r) and Vext (r)′ 6=
Vext (r) gives rise to the same equilibrium density profile ρ0 (r) is wrong. As a conse-
quence we conclude that the equilibrium probability distribution f0 is a functional of
the equilibrium density distribution ρ0 (r)

f0 = f0 [ρ0 (r)],

which further implies that the functional of the grand potential, Eq. (2.4), is also a
functional of ρ0 (r), i.e.
Ω[f0 ] = Ω[ρ0 ].
This implies that the functional of the grand potential can be rewritten in the form
Z
Ω[ρ] = F [ρ] + d3 rρ(r)(Vext (r) − µ)

with the unique functional of the intrinsic Helmholtz free energy

F [ρ] = T rcl f (Tkin + U + β −1 ln f ). (2.12)

Hence, we can express the functional of the grand potential as a functional of the
density distribution ρ(r). This feature is the reason why the theory is called density
functional theory.
Now we also can rewrite the minimum property of the functional, Eq.(2.7), in terms
of density profiles:
Ω[ρ(r) 6= ρ0 (r)] > Ω[ρ0 (r)].
The main result of this chapter can be summarized by the variational principle

δΩ[ρ] 
= 0,
δρ(r)) ρ(r)=ρ0 (r)


which expresses the minimum property in mathematical terms. Furthermore we have

for the equilibrium density profile ρ0 (r)

Ω[ρ0 ] ≡ Ω.

Thus, by minimizing the functional of the grand potential we obtain the thermodynamic
properties of the system via its grand potential Ω and the structure of the system in form
of ρ0 (r).
2.6. CLASSICAL ANALOG TO KOHN-SHAM EQUATIONS 15

2.6 Classical analog to Kohn-Sham equations

It is possible to split the intrinsic free energy functional into two parts

F [ρ] = Fid [ρ] + Fex [ρ],

where Fid [ρ] is the intrinsic free energy of an ideal (non interacting) gas. The second
contribution, Fex [ρ], is the excess (over the ideal gas) free energy functional and con-
tains all the information about the interparticle interaction. The ideal gas contribution
can be calculated exactly to be
Z  
Fid [ρ] = β −1 d3 rρ(r) ln λ3 ρ(r) − 1

with the thermal de Broglie wavelength

s
h2 β
λ= .
2πm
Formally, the functional can be minimized through the variational principle. The result
is
δΩ[ρ] δFex
= 0 = β −1 ln λ3 ρ(r) + + Vext (r) − µ.
δρ(r) δρ
We can split the chemical potential µ = µid + µex into an ideal gas contribution µid =
β −1 ln λ3 ρbulk , where ρbulk is the constant bulk density, and an excess contribution µex .
The variational principle leads to the self-consistent equations for the density profile

ρ(r) = ρbulk exp(−βVext (r) + c(1) (r) + βµex ). (2.13)

In this equation we have introduced the one-body direct correlation function

δFex [ρ]
c(1) (r) = −β .
δρ(r)
Note that for a constant bulk density, the one-body direct correlation function becomes
−βµex . If the external potential is of finite range the argument of the exponential
function in Eq. (2.13) vanishes in the limit r → ∞ when c(1) (r → ∞) → −βµex . In
this limit we obtain ρ(r → ∞) → ρbulk .
This formal solution is not too helpful, because both the left and the right hand
side of Eq. (2.13) depend on ρ(r), because Fex [ρ] and hence c(1) (r) are functionals of
the density profile ρ(r). For an ideal gas, for which Fex [ρ] = 0, we can solve for the
equilibrium density profile and we find the well-known result

ρid
0 (r) = ρbulk exp(−βVext (r)).
16 CHAPTER 2. BASICS OF DENSITY FUNCTIONAL THEORY

2.7 Generalization to Mixtures

For completeness we give the form of the functional of the grand potential for a ν
component mixture without derivation:
ν Z  ν Z
 X  
3
ln λ3i ρi (r) d3 rρi (r) Vext
i
X −1
Ω[{ρi }] = Fex [{ρi }]+ β d rρi (r) −1 + (r) − µi ,
i=1 i=1

with s
h2 β
λi = .
2πmi
The notation {ρi } indicates that there is a set of density profiles for all components i =
1, . . . , ν. The excess free energy functional, which is still unspecified, is a functional
of all density profiles of the mixture. The remaining terms simply turn into sums over
all species. In order to minimize the functional of the grand potential one has to solve
the coupled equations

δΩ[{ρi }] 
= 0, i = 1, . . . , ν.
δρi (r) ρi (r)=ρ0,i (r)


2.8 Excess free energy Fex

So far we have presented the general formalism of density functional theory. In order to
be able to perform any calculation we have to specify the system under consideration,
i.e. the interparticle interaction. Formally we then can employ Eq. (2.12) to obtain
the functional of the free energy. Unfortunately, this is in practice not possible. This
would be equivalent to calculating the partition sum exactly.
Since there is no direct way to derive the functional of the excess free energy
Fex from the Hamiltonian, we used as starting point, one requires some approximate
schemes to construct the functional. Different approaches prove useful depending on
the system under consideration. One system of particular interest is the hard-sphere
mixture, which often is used as a reference system for mixtures of simple fluids or for
colloidal mixtures. In those cases the short-ranged strong repulsion is mapped onto an
hard-sphere diameter and the longer ranged attraction is taken into account through a
perturbation theory treatment.
For hard-sphere mixture we have a rather accurate and successful approach called
fundamental measure theory (FMT) introduced by Yasha Rosenfeld. We shall discuss
this approach in detail in the following chapter.
Chapter 3

Fundamental measure theory

3.1 Introduction

Yasha Rosenfeld
In 1989 Rosenfeld [6] introduced novel ideas for deriving a density functional the-
ory (DFT) for hard-sphere mixtures. His approach, which is distinctly different from
earlier non-local, weighted density approximations [5], is based on the fundamen-
tal geometrical properties of the spheres and is termed fundamental measure theory
(FMT). The original version met with considerable success when applied to a variety
of inhomogeneous situations, including the hard-sphere fluid adsorbed at walls and
confined in model pores [5, 6]. Although the original version could not describe a
stable crystalline phase the FMT was refined [7, 8] in order to incorporate the freez-
ing transition. These refinements and subsequent improvements/modifications of FMT

17
18 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

have all focused on the zero-dimensional (0D) limit, i.e. the limit which pertains to
a narrow cavity that can contain at most one sphere. Requiring the DFT to yield the
exact free energy in the 0D limit provided new insight into the structure of FMT and
suggested new prescriptions for functionals that could describe situations of extreme
confinement [7, 8]. More recently Tarazona and Rosenfeld [9–11] have argued that the
hard-sphere free energy functional can be constructed solely from the requirement that
the functional reproduces the exact 0D limit for cavities of different shapes; the equa-
tion of state and the correlation functions of the homogeneous fluid are then given as
output from, rather than input to, the DFT. This particular strategy is reviewed briefly
in Refs. [10, 12].
One of the main limitations of the original FMT, and indeed of its successors, is
that the underlying bulk fluid equation of state is the Percus-Yevick (PY) compress-
ibility equation, equivalent to scaled particle theory. As is well-known, for the case
of the pure hard-sphere fluid this implies that the pressure p is overestimated for fluid
densities approaching that at bulk freezing [13]. A serious consequence of the inac-
curacy of the underlying PY fluid equation of state is that the FMT, suitably modified
to include a tensor measure, predicts coexisting fluid and solid densities that are rather
low w.r.t. computer simulation results [12].
In this chapter we present a derivation of Rosenfeld’s fundamental measure theory
functional and then show how the derivation can be adjusted in order to enforce an
accurate equation of state, which is done in the White Bear version of FMT [39, 40]
and more recently in the White Bear version Mark II [42].

3.2 Exact result in d = 1

The structure of Rosenfeld’s fundamental measure theory follows the structure of the
exact excess free energy functional for the ν-component hard-rod mixtures in d = 1
[18, 19]. The radius of component i is Ri , so that the length of a rod of component i is
2 Ri . A sketch of the system is shown in Fig. 3.1. The derivation of this functional is
technical quite involved and is specific to the one-dimensional case. Therefore we shall
only quote the functional and its structure, so that we can understand how Rosenfeld
got inspired.
The exact excess free energy functional in d = 1 can be written as
Z
1d
βFex [{ρi }] = dz Φ({nα }),

where Φ is the excess free energy density, which is a function (not a functional) of a
3.3. ROSENFELD’S FUNDAMENTAL MEASURE THEORY (D = 3) 19

set of weighted densities nα . The weighted densities are defined as

ν Z
dz ′ ρi (z ′ ) ωαi (z − z ′ ),
X
nα (z) =
i=1

i.e. they are sums over all components i = 1, . . . , ν of convolutions of weight functions
ωαi , which are specific to the geometry of component i. In d = 1 one has two different
weight functions for each component, namely
1
ω0i (z) = (δ(z − Ri ) + δ(z + Ri )) ,
2
which can be interpreted as a weight function that marks the surface of the rod, which
consist of the two points at z − Ri and z + Ri , and

ω1i (z) = Θ(Ri − |z|),

which can be interpreted as a weight function that marks the volume of the rod. For the
definition of the weight functions we have used the Dirac Delta function δ(x) and the
Heaviside step function Θ(x), which is 1 for x > 0 and 0 otherwise. We can represent
the Mayer-f function between a rod of species i and one of species j, which is defined
by (
−1 |z| < Ri + Rj
fij (z) = exp(−βVij (z)) − 1 =
0 otherwise,
for hard-rod interactions, in terms of weight functions

−fij (z) = ω1i ⊗ ω0j + ω0i ⊗ ω1j ,

where the The symbol ⊗ denotes the convolution of the weight functions
Z
ωiα ⊗ ωjβ (z = zi − zj ) = dz ′ ωiα (z ′ − zi ) ωjβ (z ′ − zj ).

The excess free energy density is given by

Φ({nα }) = −n0 ln(1 − n1 ),

which fully specifies the d = 1 functional.

It is quite remarkable that the excess free energy possesses such a simple structure.

3.3 Rosenfeld’s Fundamental Measure Theory (d = 3)

In order to construct a density functional for a mixture consisting of ν species of hard
spheres, with ν ≥ 1, Rosenfeld used the exact low density result for the excess (over
20 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

2R i 2R j

z
Figure 3.1: Sketch of a one-dimensional hard-rod mixture with radii Ri , i = 1, . . . ν.
The rods can move along a line.

ideal gas) Helmholtz free energy functional, valid in the limit where all the one-body
densities {ρi (r)} → 0,
1X
Z Z
3
βFex[{ρi }] = − dr d3 r ′ ρi (r)ρj (r′ )fij (|r − r′ |) (3.1)
2 i,j

as a starting point. He noted that the Mayer-f function between a sphere of component
i and one of component j, which is defined, analog to the one-dimensional case, by

fij (r) = exp(−βVij (r)) − 1.

Vij (r) is the pair potential between two species i and j. fij (r) has a purely geometrical
interpretation because of the hard-sphere potential
(
∞ r < Ri + Rj
Vij (r) =
0 otherwise,

which gives rise to the Mayer-f function

(
−1 r < Ri + Rj
fij (r) =
0 otherwise.

The Mayer-f function fij (r) of two hard spheres with radii Ri and Rj marks the vol-
ume which is not accessible to the center of one sphere, say of species i, close to the
other of species j. This volume is a sphere of radius Ri + Rj . In general, the volume
of two joined convex bodies Vi+j can be written as

Vi+j = Vi + Si Rj + Ri Sj + Vj ,

where Vi , Si , and Ri are the volume, the surface area and the mean radius of curvature
of the body, respectively. The validity of this relation can be checked easily for two
spheres where Vi+j = 4π/3(Ri + Rj )3 and Vi = 4π/3Ri3 and Si = 4πRi2 .
Analog to the one-dimensional case, the Mayer-f functions of a hard-sphere mix-
ture can be decomposed into the form

−fij (r) = ω3i ⊗ ω0j + ω0i ⊗ ω3j + ω2i ⊗ ω1j + ω1i ⊗ ω2j − ω ~ 1j − ~ω1i ⊗ ~ω2j
~ 2i ⊗ ω (3.2)
3.3. ROSENFELD’S FUNDAMENTAL MEASURE THEORY (D = 3) 21

with the weight functions given by

ω3i (r) = Θ(Ri − r),

ω2i (r) = δ(Ri − r),
ω2i (r)
ω1i (r) =
4πRi
i
ω 2 (r)
ω0i (r) =
4πRi2
r
~ω2i (r) = δ(Ri − r),
r
~ω2i (r)
~ω1i (r) = ,
4πRi

where Θ(r) is again the Heaviside function and δ(r) is the Dirac-delta function. The
symbol ⊗ in Eq. (3.2) denotes the three-dimensional convolution of the weight func-
tions Z
ωiα ⊗ ωjβ (r = ri − rj ) = dr′ ωiα (r′ − ri ) ωjβ (r′ − rj ).

It is important to note that the deconvolution, Eq. (3.2), would appear to be unnecessar-
ily complicated if only a pure hard-sphere fluid were to be considered. For a mixture,
however, this particular structure is suggested by that of the exact one-dimensional
functional of the mixture of hard rods [18, 19] – see Sec. 3.2. It is also interesting to
note that an alternative deconvolution of the Mayer-f function, suggested by Kierlik
and Rosinberg [20], avoids vector-like weight functions but introduces instead weights
containing first and second derivatives of the Dirac-delta function. It was shown later
that both deconvolutions are equivalent [21].
The weight functions give rise to a set of weighted densities {nα (r)} for the ν
component mixture. These are defined, again analog to the one-dimensional case, as
ν Z
d3 r ′ ρi (r′ ) ωαi (r − r′ ),
X
nα (r) = (3.3)
i=1

i.e. the sum of the convolutions of the density profiles of each species with its weight
function. α labels the four scalar and two vector weights. In the bulk, where the
density profiles in the absence of any external field reduce to constant bulk densi-
ties ρibulk , both vector weighted densities ~n1 and ~n2 vanish while the scalar weighted
densities reduce to the so-called scaled particle theory (SPT) [24] variables: n3 →
ξ3 = 4π i ρibulk Ri3 /3, n2 → ξ2 = 4π i ρibulk Ri2 , n1 → ξ1 = i ρibulk Ri and
P P P

n0 → ξ0 = i ρibulk . Note that ξ3 then corresponds to the total packing fraction.

P

As an appropriate ansatz for the excess free-energy functional, Rosenfeld followed

again the structure of the exact one-dimensional functional. He wrote the excess free-
22 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

energy functional in the form

Z
βFex [{ρi }] = d3 r ′ Φ({nα (r′ )}) (3.4)

where Φ, the reduced free energy density, is a function of the weighted densities. As
ansatz for Φ Rosenfeld employed dimensional analysis and used

Φ = f1 (n3 )n0 + f2 (n3 )n1 n2 + f3 (n3 )n~1 · n~2 + f4 (n3 )n32 + f5 (n3 )n2 n~2 · n~2 . (3.5)

Each term in (3.5) has the dimension of a number density, i.e. [length]−3 . In order to
ensure that the ansatz, Eqs. (3.4) and (3.5), recovers the deconvolution of the Mayer-f
function, Eq. (3.2), it is necessary to demand that to lowest order in n3 the unknown
functions f1 , f2 , f3 have expansions of the form f1 = n3 + O(n23 ), f2 = 1 + O(n23 ) and
f3 = −1 + O(n23 ). f4 = 1/24π + O(n23 ), and f5 = −3/24π + O(n23 ).
Although the ansatz in Eq. (3.5) is constructed to reproduce exactly the low density
limit, it is clear that for intermediate and high densities this ansatz introduces the ap-
proximation that the weight functions, and hence the weighted densities, required by
the low density limit are sufficient to approximate the simultaneous interaction of three
or more spheres. This approximation turns into a serious problem in the case of asym-
metric mixtures where radii of different components are significantly different [28].
The functions f1 , . . . , f5 can be determined by demanding that the resulting func-
tional satisfies a thermodynamic condition. In the original derivation Rosenfeld used
the SPT equation [24]
µi
lim ex = p, (3.6)
Ri →∞ Vi

with Vi = 4πRi3 /3, the volume of a spherical particle with radius Ri and µiex the excess
chemical potential of species i. This relation relates the excess chemical potential for
insertion of a big spherical particle with a radius Ri to the leading order term pVi , the
reversible work necessary to create a cavity big enough to hold this particle. The l.h.s.
of Eq. (3.6) can be determined self-consistently in terms of the weighted densities from
Eq. (3.5)
∂Φ X ∂Φ ∂nα
βµiex = = .
∂ρi α ∂nα ∂ρi

Due to the geometrical meaning of the weight functions we find ∂n3 /∂ρi = 4π/3Ri3 ≡
Vi , ∂n2 /∂ρi = 4πRi2 ≡ Si , ∂n1 /∂ρi = Ri , and ∂n0 /∂ρi = 1. In the limit under
consideration all but one term vanish and we obtain
1 i ∂Φ
lim βµex = .
Ri →∞ Vi ∂n3
The equation of state can be obtained from the thermodynamic bulk relation Ωbulk =
−pV . Since the grand potential density in the bulk is Ωbulk /V = Φ + fid − i ρibulk µi
P
3.3. ROSENFELD’S FUNDAMENTAL MEASURE THEORY (D = 3) 23

we obtain
X ∂Φ
βp = −Φ + nα + n0 . (3.7)
α ∂nα
The last term, n0 , results from the ideal gas contribution. We can combine these results
in order to obtain the SPT differential equation, Eq. (3.6),

∂Φ X ∂Φ
= −Φ + nα + n0 .
∂n3 α ∂nα

By collecting all terms proportional to n0 one sees that the differential equation for
f1 (n3 ) takes the form
f1′ (n3 ) n0 (1 − n3 ) = n0 ,
which is solved by
f1 (n3 ) = const1 − ln(1 − n3 ),
with an integration constant const1 that vanishes. It is easy to find the differential
equations for the remaining functions. The integration constants are chosen such that
the correct behavior at low densities is recovered. The solution found by Rosenfeld [6]
and denoted RF, is

f1RF (n3 ) = − ln(1 − n3 )

1
f2RF (n3 ) =
1 − n3
f3 (n3 ) = −f2RF (n3 )
RF
(3.8)
1
f4RF (n3 ) =
24π(1 − n3 )2
f5RF (n3 ) = −3f4RF (n3 ), (3.9)

and it is straightforward to see that these solutions satisfy the aforementioned condi-
tions for the low density limit. It is worthwhile to note that the conditions f3 = −f2
and f5 = −3f4 , that fix the dependence of the functional on the vector weighted den-
sities ~n1 and ~n2 , follow from Eq. (3.6) only if it is assumed that the SPT differential
equation, which is by construction a bulk equation, remains valid for slightly inhomo-
geneous situations. Since the vector weighted densities vanish in the bulk limit it is,
strictly speaking, impossible to determine the functions f3 and f5 from bulk thermo-
dynamics alone. Given the success of the Rosenfeld functional in various applications
we choose to adopt the conditions (3.8) and (3.9) in the subsequent modifications.
The resulting functional, that we refer to as the original Rosenfeld (RF) functional,
is usually written in the form Φ = Φ1 + Φ2 + Φ3 with

ΦRF
1 = −n0 ln(1 − n3 ), (3.10)
24 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

n1 n2 − ~n1 · ~n2
ΦRF
2 = , (3.11)
1 − n3
n3 − 3n2~n2 · ~n2
ΦRF
3 = 2 . (3.12)
24π(1 − n3 )2

Although this functional was found to be very successful and often very accurate
in accounting for various properties of highly inhomogeneous fluid phases, it failed to
predict the fluid to solid phase transition of the pure hard-sphere system. This failing
was first remedied empirically by Rosenfeld et al. [7, 8] who modified the dependence
of Φ3 on the weighted densities n2 and ~n2 , taking into account certain features of
’dimensional crossover’. The modifications were found to perform better than the
original Rosenfeld DFT for densely packed fluids in spherical cavities – a situation of
extreme confinement [25, 26]
Subsequently Tarazona and Rosenfeld derived a FMT especially designed to study
the properties of the one-component hard-sphere solid [9–11]. They began with the so-
called 0D-limit which considers a narrow cavity that can hold at most a single sphere.
Starting with the free energy function for this narrow pore, functionals are derived
for higher embedding dimensions. A three dimensional functional based on this idea
reproduces the original Rosenfeld functional. In Ref. [11] it is pointed out, however,
that there are shapes of 0D cavities which cannot be described by the particular set of
weight functions chosen in FMT. The problem becomes more acute with increasing
embedding dimension. In three dimensions this prevents the Rosenfeld functional
or equivalently a functional based solely on the 0D limit from describing the fluid-
solid phase transition of the pure hard-sphere system. In order to remedy this defect,
Tarazona [11] introduced a new second rank tensor-like weight function ωm2 (r) and
adapted the contribution Φ3 to the functional. In the notation introduced in Ref. [27],
we write the tensor weight function as

ωm2 (r) = ω2 (r)(rr/r 2 − 1̂/3), (3.13)

with 1̂ denoting the unit matrix. This gives rise to a new tensor weighted density nm2 .
The new ΦT3 term of the Tarazona FMT is given by [11, 27]

1 
3

3

ΦT3 = n − 3n ~
n
2 2 · ~
n 2 + 9 ~
n n ~
n
2 m2 2 − Tr(n )/2 , (3.14)
24π(1 − n3 )2 2 m2

and the application of the augmented functional to the hard-sphere solid provided an
excellent account of simulation results for the equation of state and for other proper-
ties of the solid. The extension of this approach to hard-sphere mixtures requires the
introduction of a new third rank tensor-like weight function [28].
3.4. THE WHITE BEAR VERSION OF FMT 25

3.4 The White Bear Version of FMT

Building upon the ideas presented so far, we are now ready to construct a new func-
tional. We retain the same weight functions and the same form (3.4) for the func-
tional but use a different thermodynamic condition in order to specify the coefficients
f1 , . . . , f3 of the ansatz (3.5). In contrast to existing FMT functionals which output the
equation of state [for fluid states this is the Percus-Yevick (PY) compressibility equa-
tion] we use the Mansoori-Carnahan-Starling-Leland (MCSL) equation of state [14],
which is a generalization to the ν-component hard-sphere fluid of the accurate, one-
component Carnahan-Starling equation of state [15], as an input. We prescribe the
functions f1 , . . . , f5 , retaining the two conditions (3.8) and (3.9), such that the equation
of state which underlies the new functional is the MCSL pressure. For this approach
to be feasible it is important that the MCSL equation of state is based on the same SPT
variables which enter the PY compressibility equation of state underlying the original
FMT. The MCSL pressure is given by
n0 n1 n2 n32 n3 n32
βpM CSL = + + − . (3.15)
1 − n3 (1 − n3 )2 12π(1 − n3 )3 36π(1 − n3 )3
The final term in (3.15) is absent in the PY result.
Incorporating the deconvolution of the Mayer-f function and imposing the condi-
tions (3.8) and (3.9), we employ an ansatz for Φ of the form

Φ = f1 (n3 )n0 + f2 (n3 )(n1 n2 − ~n1 · ~n2 ) + f4 (n3 )(n32 − 3n2~n2 · ~n2 ). (3.16)

In order to determine the three unknown functions f1 , f2 , and f4 we employ Eq. (3.7) in
a slightly different way than Rosenfeld did. Instead of the SPT differential equation we
demand that thermodynamic pressure, given by Eq. (3.7) equals the MCSL equation
of state:
3
X ∂Φbulk
−βpM CSL = Φbulk − nα − n0 , (3.17)
α=0 ∂nα
with the sum over the scalar weighted densities only. Substituting (3.15) and (3.16)
into (3.17) we obtain differential equations for f1 , f2 and f4 by collecting all the terms
proportional to n0 , n1 n2 , and n32 , respectively. These differential equations can be
solved easily and we find f1 (n3 ) = f1RF (n3 ), f2 (n3 ) = f2RF and
n3 + (1 − n3 )2 ln(1 − n3 )
f4 (n3 ) = (3.18)
36πn23 (1 − n3 )2
The resulting excess free-energy density is given by
n1 n2 − ~n1 · ~n2 n3 + (1 − n3 )2 ln(1 − n3 )
Φ = −n0 ln(1 − n3 ) + + (n32 − 3n2~n2 · ~n2 )
1 − n3 36πn23 (1 − n3 )2
(3.19)
26 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

which should be compared with the original Rosenfeld form, Eqs. (3.10)–(3.12). Note
that in the low density limit we obtain limn3 →0 f4 (n3 ) = 1/(24π), i.e. the same value
as from the original Rosenfeld functional [see Eq. (3.12)]. Thus we are guaranteed to
recover the exact low density limit.
As the derivation of the new functional has followed that of the original Rosenfeld
FMT very closely, it faces similar problems when it is applied to the freezing transition.
However, the same procedures that remedied the failings for the original FMT can be
used for the new functional. Thus, it is possible to follow the empirical procedure
of Refs. [7, 8] and modify the dependence of Φ3 on the weighted densities n2 and
~n2 in the new functional. This approach would enable the functional to treat a hard-
sphere mixture. Equally well it is possible to follow Tarazona [11] who introduced
a tensor-like weighted density in order to study the properties of the one-component
hard-sphere solid. This is the route we employ here, i.e. in the present calculations for
the solid phase we replace the term (n32 − 3n2~n2 · ~n2 ) in Eq. (3.19) by the numerator
of Tarazona’s expression (3.14) so that the present Φ3 is given by

n3 + (1 − n3 )2 ln(1 − n3 )  3 
3

Φ3 = n2 − 3n2 ~
n 2 · ~
n 2 + 9 ~
n 2 nm ~
n 2 − Tr(nm2 )/2 .
36πn23 (1 − n3 )2 2

(3.20)

3.5 Test for self-consistency

As mentioned earlier, Rosenfeld [6] used the scaled particle equation (3.6) to determine
the functions f1 , . . . , f5 . Here we re-examine this equation in the context of self-
consistency for the functional.
First we note that the excess chemical potential of inserting a single big hard sphere
of species i and radius Ri into a fluid of hard spheres is the reversible work done to
create a cavity that is large enough to hold this inserted hard sphere. In SPT one starts
with a point-like cavity and increases its size until it is sufficiently large. Clearly, when
increasing the cavity size one must work against the pressure of the fluid resulting in a
term pVi , where Vi = 4πRi3 /3. Since the surface area of the sphere is also increased,
work must also be done against the surface tension. This second term is proportional
to the surface area Si = 4πRi2 . Moreover for finite values of Ri the surface tension
will also depend on the radius of curvature so there will be an additional term that is
proportional to Ri . If, however, we divide the excess chemical potential by the volume
Vi it is easy to see that Eq. (3.6) follows and that it is exact in the limit Ri → ∞.
The connection to FMT can be made by noting that the excess chemical potential
3.6. THE WHITE BEAR VERSION OF FMT MARK II 27

takes the form

3
∂Φbulk ∂nα
βµex
X
i = i
(3.21)
α=0 ∂nα ∂ρbulk
∂Φbulk ∂Φbulk ∂Φbulk ∂Φbulk
= Vi + Si + Ri + , (3.22)
∂n3 ∂n2 ∂n1 ∂n0

and we used the definition of the SPT variables n3 , . . . , n0 given earlier. Equation (3.22)
has precisely the same form as the SPT expansion so it is clear that for any FMT func-
tional the coefficient of the leading volume term should be identified with βp, i.e. the
relation
∂Φbulk
= βp (3.23)
∂n3
should be obeyed.
In the derivation of the original Rosenfeld functional Eq. (3.23) is imposed, i.e. the
left hand side of Eq. (3.17) is identified with −∂Φbulk /∂n3 and the resulting (SPT) dif-
ferential equation is solved. The pressure which results is the SPT or, equivalently, the
Percus-Yevick compressibility equation of state. For the present functional, however,
Eq. (3.23) is not imposed and we find from Eq. (3.19) that

∂Φbulk n0 n1 n2 n32 (2 + n3 (n3 − 5)) n32 ln(1 − n3 )

= + − − , (3.24)
∂n3 1 − n3 (1 − n3 )2 36πn23 (1 − n3 )3 18πn33

which evidently is different from the MCSL equation of state (3.15). The difference
arising from this inconsistency was examined within the context of the one-component
fluid where the pressure inputted into the theory is the accurate Carnahan-Starling
equation of state, pCS . We show both the Carnahan-Starling equation of state (solid
line) and the pressure obtained from Eq. (3.24) (dashed line) in Fig. 3.2. The deviation
between these two curves is at most 2%. In contrast, the Percus-Yevick compressibility
equation of state pcP Y , also shown in Fig. 3.2 (dotted line), overestimates the pressure
of a hard-sphere fluid close to freezing by up to 7%.

3.6 The White Bear Version of FMT Mark II

Based on the observation that the MCSL equation of state leads to a excess free energy
density that is slightly inconsistent, recently a new generalization of the Carnahan-
Starling equation of state to mixtures was proposed [41]
   
n0 n1 n2 1 + 31 n23 n32 1 − 32 n3 + 13 n23
βpCSIII = + + .
1 − n3 (1 − n3 )2 12π(1 − n3 )3
28 CHAPTER 3. FUNDAMENTAL MEASURE THEORY

12 Percus-Yevick pcPY
Carnahan-Starling pCS
10 Eq.(2.24)
β p / ρbulk

0 0.1 0.2 0.3 0.4 0.5

η
Figure 3.2: The equation of state of the pure hard-sphere fluid versus packing fraction
η = ρbulk 4π/3R3. For the present DFT the Carnahan-Starling pressure is imposed by
the theory. The pressure given by ∂Φbulk /∂n3 in Eq. (3.24) deviates very slightly from
Carnahan-Starling, attesting to the high degree of self-consistency of the approach.
3.6. THE WHITE BEAR VERSION OF FMT MARK II 29

This equation of state reduces to the Carnahan-Starling equation of state in the one-
component case and represents data for binary and ternary mixtures obtained by com-
puter simulations more accurate than the MCSL result. Based on this new equation of
state we can, following the derivation of Sec. 3.4, derive an excess free energy func-
tional, which improves the level of self consistency. We find [42]
n
1 n2 − ~n1 · ~n2

ΦWBII = −n0 ln(1 − n3 ) + 1 + 91 n23 φ2 (n3 )
1 − n3
  n32 − 3n2~n2 · ~n2
+ 1 − 94 n3 φ3 (n3 ) (3.25)
24π(1 − n3 )2

with  
φ2 (n3 ) = 6n3 − 3n23 + 6(1 − n3 ) ln(1 − n3 ) /n33 ,
and  
φ3 (n3 ) = 6n3 − 9n23 + 6n33 + 6(1 − n3 )2 ln(1 − n3 ) /(4n33 ).
This functional is similar in complexity as the White Bear version, Eq. 3.19, or the
Rosenfeld functional, but is constructed such that for a one-component fluid we find
∂Φbulk
= βpCS ,
∂n3
with Carnahan-Starling equation of state pCS .
30 CHAPTER 3. FUNDAMENTAL MEASURE THEORY
Chapter 4

Application

4.1 Introduction

Bob Evans
Bob Evans wrote the first review on density functional theory for classical systems
in 1979 [4], just a few years after the theory was translated from quantum mechanical
systems to classical ones. He was among the first to realize the power of density
functional theory and pioneered several of its application. With his review article [4]
Bob Evans inspired and influenced many people and made density functional theory
for classical system available to a broad audience. My first contact to density functional
theory was by reading his papers on the subject.

4.2 Hard-Sphere Fluid at a Hard Wall

The first example application we discuss here is one of the simplest inhomogeneous
system one can consider: a hard-sphere fluid at a planar hard wall. Note that in the case
of a fluid the equilibrium density profile ρ0 (r) has the same symmetry as the external

31
32 CHAPTER 4. APPLICATION

potential. For the planar hard wall this means that ρ0 (r) = ρ0 (z). All interactions are
hard-core like which makes temperature a simple scaling parameter. The only param-
eter in the system is the bulk density ρbulk or equivalently the bulk packing fraction
η = ρbulk 4π
3
R3 , where R is the radius of the spheres.
The general outline of the problem of finding the equilibrium density profile ρ0 (z)
is as follows:

• choose liquid density ρbulk or packing fraction η = ρbulk 4π

3
R3

• initialize density profile

ρ(z) = ρbulk exp(−βVext (z))

with (
∞ z<R
βVext(z) =
0 otherwise

• minimize the density functional

δΩ[ρ]
=0
δρ(r)

There are several points to be addressed in order to make clear what we mean by
minimizing the density functional and how we perform this task in practice. Functional
minimization is a standard problem in numerical mathematics and there are several
more or less clever algorithms available. Each algorithm has its advantages as well as
drawbacks. In the present context we wish to keep things as simple as possible and
restrict our consideration to a simple Picard iteration, which is in general robust but
converges slower than a clever minimization.

4.2.1 Minimizing Ω[ρ] through a Picard Iteration

First I give the four simple steps of the Picard iteration and then I discuss their meaning.

1. initialize density profile: ρ(0) (z) = ρbulk exp(−βVext (z))

2. calculate [see Eq. (2.13)] using ρ(i) (z)

ρ̃(i) (z) = ρbulk exp(−βVext (z) + c(1) (z) + βµex )

3. mix solutions with mixing parameter α

ρ(i+1) (z) = (1 − α)ρ(i) (z) + αρ̃(i) (z)

4.2. HARD-SPHERE FLUID AT A HARD WALL 33

4. goto step 2 until solution is converged

The initialization given here is only one possible choice. The closer this initial
guess is to the equilibrium density distribution the faster the minimization will con-
verge. Often, however, it is very difficult to have a good guess. If the external field is
strongly attractive, then the choice given here can turn out to be a bad one. Once an
initial profile is chosen we can start with the iteration. In step 2 we use Eq. (2.13) to
calculate a different guess, which we call ρ̃(i) (z). Clearly, if we input the equilibrium
density profile into the r.h.s. of Eq. (2.13), we again recover the equilibrium density
profile from the l.h.s. of Eq. (2.13). If we input any other density profile we obtain a
different guess for the density profile. In order to keep the iteration from making too
rapid changes, which might result in unphysical density distributions such as negative
densities or local packing fraction larger than 1, it is useful to mix the old and the new
guess, as specified in step 3. The choice of the mixing parameter α is very important.
If we choose it too small, the convergence of the iteration is very slow. If we choose
it too large, we end up with the same problem mentioned above: the changes in the
density profile might be too rapid and one might end up with an unphysical result. Step
4 is to check if the iteration converged already. If the change in the density profile is
smaller than a threshold then we can stop the iteration.
In the course of the minimization either through the described iteration or through
any other algorithm one has to calculate the weighted densities nα (z) from a given
density distribution very often.

4.2.2 Weighted Densities

Here, to keep our considerations simple, we restrict ourself to a planar geometry. For
any (not just hard-wall) external potential that possesses a planar geometry so that
Vext (r) = Vext (z), where z is the distance normal to the wall, a fluid density profile
also possesses the planar geometry: ρ(r) = ρ(z). It is easy to show that the weighted
densities, Eq. (3.3), also take on the planar geometry, i.e. nα (r) = nα (z). By using
the symmetry of the problem, one can perform two of the three integrals in Eq. (3.3)
analytically and thereby reduce the calculation of the weighted densities to a single
integral. In the planar geometry one finds that this integral is of the convolution type,
just as the original three-dimensional integral in Eq. (3.3). Note it is not always the
case that a three-dimensional convolution integral remains a convolution after some
integrals are performed analytically. For example, in the cylindrical geometry, one
looses the convolution property by performing two integrals analytically, in the spher-
ical geometry, however, one finds results similar to the planar geometry.
34 CHAPTER 4. APPLICATION

One finds that

Z
nα (r) = nα (z) = dz ′ ρ(z ′ ) ωα (z − z ′ )

with the one-dimensional weight functions ω3 (z) = π(R2 − z 2 ), ω2 (z) = 2πR, and
~ω2 (z) = 2πz~ez , with the unity vector in the direction normal to the wall ~ez . The
remaining weight functions are related to ω2 (z) and ~ω2 (z) via ω1 (z) = ω2 (z)/(4πR),
ω0 (z) = ω2 (z)/(4πR2 ), and ~ω1 (z) = ~ω2 (z)/(4πR).
Since the integrals are still convolutions one can exploit the convolution theorem
and perform the calculation in Fourier space, where the convolution is a simple multi-
plication. Using the FFT (fast Fourier transform) we get
Z
dz ′ ρ(z ′ ) ωα (z − z ′ ) = F T −1 (F T (ρ) ∗ F T (ωα )) ,

where F T denotes the fast Fourier transform of a function and F T −1 the fast inverse
Fourier transform. The advantage of using FFT is the speed. Convolutions performed
in Fourier space are in general much faster than those performed in real space. If one
wants to implement an integration scheme of higher order, which is straightforward in
the real space, one has to be careful in Fourier space.
Once the weighted densities are evaluated, one is ready to calculate the one-body
direct correlation function c(1) (z)

4.2.3 One-Body Direct Correlation c(1) (z)

From the definition of the one-body direct correlation function and the structure of the
excess free energy functional within fundamental measure theory one finds

δFex [ρ] ∂Φ({nα̃ }) δnα (z ′ )

Z
(1)
X
c (z) = −β =− dz ′ .
δρ(z) α ∂nα δρ(z)

The main problem is to calculate the variation of the weighted densities nα (z ′ ) w.r.t.
the density profile ρ(z). The result (in planar geometry) is quite simple

δnα (z ′ ) δ
Z
= dz ′′ ρ(z ′′ ) ωα (z ′ − z ′′ ) = ωα (z ′ − z).
δρ(z) δρ(z)
However, one has to be careful because of the argument of the weight function. Com-
pared to the argument entering the weight function of the weighted densities, the argu-
ment entering the calculation of c(1) (z) is negative, i.e. z − z ′ becomes z ′ − z. For the
scalar weight functions this is unimportant, since the scalar weight functions are even

ωα (z ′ − z) = ωα (z − z ′ ),
4.2. HARD-SPHERE FLUID AT A HARD WALL 35

but the vector-like weight functions are odd

~ α (z ′ − z) = −~ωα (z − z ′ ).
ω

Taking this sign into account, it is possible to perform the convolutions in Fourier space
using FFT methods:
!
∂Φ({nα̃ })
c(1) (z) = −
X
F T −1 FT ( ) ∗ F T (±ωα ) .
α ∂nα

4.2.4 Hard-Sphere Fluid at a Hard Wall: the density profile

After discussing the practical issues of minimizing a fundamental measure theory den-
sity functional, we can have a look at a typical density profile of a hard-sphere fluid
at a planar hard wall. The packing fraction is choose to be η = ρbulk 4π
3
R3 = 0.4257,
which is quite high. Note that for η > 0.494 hard spheres freeze and form a fcc crystal.
In Fig. 4.1 we show the result. The full line denotes the density profile of the
White Bear version of FMT, the dashed line that of the original Rosenfeld functional
and the symbols denote results obtained from simulation [30]. The results from the
two version of DFT lie almost on top of each other except for values of z very close to
contact (see inset). The reason for the small deviation close to the wall is the contact
theorem, which states that at a planar hard wall the contact density ρ(R+ ) is equal to
the bulk pressure
ρ(R+ ) = βp,

where R+ indicates that the contact value of the density profile is the value at z = R
plus an infinite displacement.
The equation of state underlying the Rosenfeld functional is the Percus-Yevick
compressibility pressure, which is known to overestimate the actual pressure of the
hard-sphere fluid. By construction, the equation of state of the White Bear version of
FMT is the Carnahan-Starling pressure, which is closer to the pressure of the hard-
sphere fluid and agrees well with computer simulations.
There are several other properties of the one-component hard-sphere system in the
fluid and the crystal phase studied in detail [39]. In general the agreement found with
simulations is excellent. Hard-sphere mixtures can also be studied within FMT [34].
If the size ratio is not too asymmetric the agreement with simulations is very good. As
the sizes of the species in the mixture get more asymmetric some problems of FMT
functionals become apparent [28].
36 CHAPTER 4. APPLICATION

6
6

ρ(z) σ3
4

0
ρ(z) σ3

4 0.5 0.55 0.6

z/σ
MC simulation
White Bear functional
2 Rosenfeld functional

0
0.5 1 1.5 2 2.5 3

z/σ
Figure 4.1: Density profile of a one-component hard-sphere fluid at a planar hard wall
for η = 0.4257.

4.3 Square-Well Fluid

The hard-sphere fluid is often employed as a useful reference system for a fluid with a
hard-core repulsion at short separation and an additional attraction. As an example for
such a fluid we consider a fluid with a square-well interparticle interaction given by



 ∞ r < 2R
βVsw (r) = −ε 2R < r < 2Rsw

 0

otherwise.

Rsw denotes the square-well radius. Even for such a simple interparticle interaction
potential it is in general not possible to construct a density functional of the intrinsic
excess free energy Fex analog to the fundamental measure theory for hard-sphere mix-
tures. Very often the additional attraction is taken into account in an perturbative way
4.3. SQUARE-WELL FLUID 37

by splitting the excess free energy into a hard-sphere contribution plus a perturbation
1Z 3 Z
Fex [ρ] = HS
Fex [ρ] + d rρ(r) d3 r ′ ρ(r′ ) φsw (|r − r′ |).
2
The perturbation term underestimates the correlation in the system. To compensate
this effect, one usually introduces a modified square-well potential
(
−ε r < 2Rsw
βφsw (r) =
0 otherwise,

where the square-well is extended into the core, i.e. to r → 0. While this seems phys-
ical meaningless at first, it helps to empirically correct for the error in the correlations.

4.3.1 Bulk Fluid Phase Diagram

In the following we will restrict the considerations to the fluid phase and neglect the
possibility of crystallization. In the fluid phase, if the temperature is sufficiently small,
the square-well fluid can phase separate into a low density gas and a high density
liquid. In order to locate at which temperatures this phase separation can take place
we require the chemical potential and the equation of state, which can be obtained
from the density functional, by inputting a constant density profile ρ(r) = ρbulk . The
resulting chemical potential is
 3
∂f  Rsw

µ(ρbulk ) = = µHS (ρbulk ) − εη + ln λ3 ρbulk ,
∂ρ ρ=ρbulk R


with a hard-sphere contribution µHS (ρbulk ), a square-well contribution, and an ideal

gas term ln(λ3 ρbulk ). At first it seems as if the value of λ and hence of the mass of the
particles plays an role in determining the phase diagram. However, this is not the case
and one can see easily that the value of λ shift the value of µ but does not affect the
phase diagram at all. Therefore, it is possible to replace the ideal gas contribution to the
chemical potential, ln λ3 ρbulk by a simpler term of the form ln η, which is equivalent
to a particular choice of λ.
The equation of state follows from the grand potential of a bulk system to be
3
ε Rsw

p(ρbulk ) = µ(ρbulk )ρbulk − f (ρbulk ) = pHS (ρbulk ) − ρbulk η ,
2 R
with the hard-sphere equation of state of the reference system and a square-well term.
In order to describe a phase equilibrium between two fluid phases I and II with cor-
responding densities ρI and ρII , we demand chemical and mechanical equilibrium

µ(ρI ) = µ(ρII ) and p(ρI ) = p(ρII ).

38 CHAPTER 4. APPLICATION

1.4

1.2
T/(β ε)

0.8

0 0.1 0.2 0.3 0.4

Figure 4.2: The bulk fluid phase diagram of a square-well fluid with Rsw = 3R. The
full circle denotes the critical point. Below the critical temperature the fluid can phase
separate into a low density gas and a high density liquid.

Note that these conditions can be fulfilled only at sufficiently low temperatures, where
the equation of state and the chemical potential display van der Waals loops. The
temperature at which these loops appear is called the critical temperature Tc . To locate
the critical temperature one demands that the first and the second derivative of the
pressure w.r.t. the density vanishes, i.e.
∂ 2 p 
 
∂p 
= 0 and = 0.
∂ρ T =Tc ∂ρ2 T =Tc
 

For T < Tc , the system can separate into a low density gas and a high density liquid
phase.
The bulk fluid phase diagram for a square-well radius of Rsw = 3R is shown in
Fig. 4.2 in the η-T representation. The critical point is denoted by the full circle. Below
the critical temperature a low density gas and a high density liquid can coexist, if their
respective densities are on the binodal (full line). The coexistence is indicated for the
4.3. SQUARE-WELL FLUID 39

temperatures T /(βε) = 1.2, 1.0 and 0.83 by the dotted lines. Outside the binodal
line there are single phase regions. At densities below the coexisting gas density, the
gas phase is the single stable bulk phase and at densities above the coexisting liquid
density, the liquid phase is the single stable bulk phase. Inside the binodal line there is
a region of metastable and unstable states, which shall not be discussed here.

4.3.2 Free Interface

In the case of bulk coexistence between a low density gas and a high density liquid
one finds an inhomogeneous density distribution of the free interface. The interface is
called free, because it can form without the presence of an external field, i.e. in the
bulk.
The calculation of the free interface density profile can be done by the following
steps:

• for a fixed temperature T < Tc choose coexisting densities ρI and ρII so that

µ(ρI ) = µ(ρII ) and p(ρI ) = p(ρII )

• initialize density profile

(
ρI z < 0
ρ(z) =
ρII z > 0

• minimize density functional

δΩ[ρ]
=0
δρ(r)

Note that the minimization has to be performed while enforcing the boundary con-
ditions: for z ≪ 0 the density profile approaches ρI and for z ≫ 0 it approaches
ρII .
For the temperatures T /(βε) = 1.2, 1.0 and 0.83, marked by the dotted lines in
Fig. 4.2, we show the density profiles of the free interface in Fig. 4.3.
For the lowest temperature considered, T /(βε) = 0.83, the difference in the co-
existing densities is considerably large, as can be seen in the phase diagram shown in
Fig. 4.2. The width of the corresponding interface, the region where the density goes
from the a gas-like density to a liquid-like density, is of the order of 4R. Note that on
the liquid side of the interface one can see the onset of an oscillatory structure.
As we increase the temperature to T /(βε) = 1.0 and 1.2, the difference in the
coexisting densities becomes smaller and the interface broader, which is to be expected
40 CHAPTER 4. APPLICATION

0.8

T/(β ε) = 0.83
0.6 T/(β ε) = 1.0
T/(β ε) = 1.2
ρ(z) σ3

0.4

0.2

0
-4 -2 0 2 4

z/σ

Figure 4.3: The free interface of a square-well fluid at three different temperatures –
see Fig. 4.2. σ = 2R is the hard-sphere diameter.

as we approach the critical point. At the critical point, the difference between the gas
and the liquid density vanishes and so does the interface.
The free interface density profiles predicted by density functional theory are smooth
functions which do not display the fluctuations caused by the capillary waves.

4.3.3 Surface Tension of the Free Interface

From the density profiles ρ0 (z) of the free interface we can calculate the energy cost
of the formation of the interface. The grand potential of the system is given by

Ω = Ω[ρ0 (z)].

The energy cost is measured by the liquid (l) and vapor (v) interface tension γlv , which
is defined by
1 1
γlv = (Ω − Ωbulk ) = (Ω[ρ0 (z)] + pV ) ,
A A
4.3. SQUARE-WELL FLUID 41

0.5

0.4

0.3
γlv σ2

0.2

0.1

0
0.8 0.9 1 1.1 1.2 1.3

T/(β ε)

Figure 4.4: The interfacial tension γlv of a square-well fluid as a function of tempera-
ture. σ = 2R is the hard-sphere diameter.

with the area of the interface A and Ωbulk = −pV . As the critical point is approached
(from below, i.e. T < Tc ), the energy cost of the interface formation becomes smaller
and at the critical point, at which the interface vanishes, the liquid-vapor interfacial
tension γlv vanishes.
The result for the interfacial tension γlv is shown in Fig. 4.4. Close to the critical
point, indicated by the full circle, the interfacial tension γlv vanishes with a power law,
according to the theory of critical phenomena. However, the exponent of the power
law is predicted by density functional theory is the (incorrect) mean-field exponent.
To determine the correct power law is quite involved and shall not be discussed here.

4.3.4 Square-Well Fluid at a Hard Wall

The next problem we consider is the inhomogeneous density distribution of a square-
well fluid close to a planar hard-wall. For a particular interesting behavior, we restrict
our considerations here to temperatures T below the critical temperature, i.e. T < Tc ,
42 CHAPTER 4. APPLICATION

1.4

1.2
T/(β ε)

0.8

0 0.1 0.2 0.3 0.4

Figure 4.5: The path through the phase diagram is indicated by the full line. The
dashed line is a guide to the eye.

and to liquid densities ρbulk , away from the binodal line. The path we take in the phase
diagram is shown in Fig. 4.5. We fix the temperature at T /(βε) = 1.0 and consider
the liquid packing fractions η=0.31000, 0.30610, and 0.30571. The coexisting liquid
density at this temperature is at ηco = 0.305700789. The actual path is indicated by the
full line in Fig. 4.5 at the temperature T /(βε) = 1.0. Since all values of η considered
here are rather close to the coexisting density, we also plot, as a guide to the eye, the
dashed line.
The density profile of the square-well fluid close to a planar hard wall can be cal-
culated by the following steps:

• choose liquid density ρbulk or packing fraction η, T < Tc

• initialize density profile

ρ(z) = ρbulk exp(−βVext (z))

4.3. SQUARE-WELL FLUID 43

0.6

0.5

0.4
ρ(z) σ3

0.3

0.2

η = 0.31000
0.1 η = 0.30610
η = 0.30571
0
0 2 4 6 8 10

z/σ

Figure 4.6: The density profiles of a square-well fluid at a planar hard wall along the
path through the phase diagram shown in Fig. 4.5. σ = 2R is the hard-sphere diameter.

with (
∞ z<R
βVext(z) =
0 otherwise

• minimize density functional

δΩ[ρ]
=0
δρ(r)

These steps are the same as for the calculation of the density profiles of the hard-sphere
fluid close to the hard wall. The resulting density profiles, however, are very different.
They are shown in Fig. 4.6.
Through the presence of a van der Waals loop, the pressure close to the binodal
is smaller than the bulk density of the liquid, βp < ρbulk . This implies through the
contact theorem mentioned above that the contact density ρ(R+ ) of the square-well
fluid has to be smaller than the liquid bulk density. What one finds is the phenomenon
called complete drying. The hard wall is hydrophobic and prefers the low density gas
44 CHAPTER 4. APPLICATION

over the high density liquid. If a liquid state point sufficiently close to bulk coexistence
is considered, the square-well fluid develops a gas film close to the wall. In the drying
case, the density profile of the fluid shows gas-like behavior close to the wall and a
liquid behavior far away from the wall. In between one finds a vapor-liquid interface.
The thickness of the gas-like film can be measured by the excess adsorption Γ,
defined by Z
Γ = dz (ρ(z) − ρbulk ).
In the complete drying regime, sufficiently close to coexistence one finds

Γ ∝ ln δµ,

with δµ = µ − µco , the distance in chemical potential from its value at coexistence µco.
The scenario of complete drying is completely confirmed by density functional theory.

4.3.5 Square-Well Fluid in a Slit

As a final application of density functional theory we consider the behavior of a square-
well fluid in a slit geometry. The slit is made of two parallel hard walls, which are
separated by a distance L. Again, to keep the application interesting we restrict the
temperature to T < Tc and choose the density to be in the liquid regime.
The steps to calculate the density profile in this geometry are given by

• choose liquid density ρbulk or packing fraction η, T < Tc

• initialize density profile

ρ(z) = ρbulk exp(−βVext (z))

with (
∞ |z| > L/2 − R
βVext(z) =
0 otherwise

• minimize density functional

δΩ[ρ]
=0
δρ(r)
The state point choose for the present calculation is shown in Fig. 4.7. We choose
T /(βε) = 1 and η = 0.32. Now we keep the density of the square-well fluid fixed and
vary the width of the slit L. The resulting density profiles are shown in Fig. 4.8.
We find that for large values of L a liquid in the slit. If the slit width is sufficiently
small, the liquid, which is the stable bulk phase, becomes metastable compared to
a gas phase. This phenomenon is called capillary evaporation. The reason for this
4.3. SQUARE-WELL FLUID 45

1.4

1.2
T/(β ε)

0.8

0 0.1 0.2 0.3 0.4

Figure 4.7: The state point of a square-well fluid in a slit geometry is indicated by the
full circle.

phase transition is a competition between the volume term and the surface term in the
slit geometry. To highlight this competition we recall that the grand potential Ω has
the following forms:

• bulk system (unconfined):

Ω = −pV

• fluid at a single wall:

Ωw = −pV + γA

• fluid in a slit of width L:

Ωs ≈ −pAL + 2γA

If the wall is hydrophobic, as in the case of a hard wall, then the wall-vapor interface
tension γv is lower than the wall-liquid interface tension γl . This can compete with the
volume term that prefers the stable bulk phase because pl > pv .
46 CHAPTER 4. APPLICATION

ρ(z) σ3 0.6

0.4

L = 10 σ
0.2 L=6σ
L=5σ

0
-4 -2 0 2 4

z/σ

Figure 4.8: Density profiles of a square-well fluid in a slit geometry for three different
values of the slit width L.

Note that in the case of a hydrophilic wall one finds γv > γl and capillary evap-
oration cannot take place. However, a phenomenon called capillary condensation can
be observed if a stable bulk vapor phase is confined in a narrow slit of hydrophilic
(sufficiently attractive) walls. In that case a high density liquid, which is metastable in
the bulk, is stabilized by the walls.
The main purpose of this chapter was to show that the same functional of the intrin-
sic excess free energy Fex can be employed to study quite different physical scenarios
simply by changing the external potential Vext (r). This is part of the power and the
beauty of density functional theory.
Bibliography

[1] P. Hohenberg and W. Kohn, Phys. Rev. 136, B 864 (1964).

[2] W. Kohn and L.J. Sham, Phys. Rev. 140, A 1133 (1965)).

[3] N. D. Mermin, Phys. Rev. 137, A 1441 (1965).

[4] R. Evans, Adv. Phys. 28, 143 (1979).

[5] Evans R 1992 in Fundamentals of Inhomogeneous Fluids (New York: Dekker) p

85

[6] Rosenfeld Y 1989 Phys. Rev. Lett. 63 980; see also Rosenfeld Y, Levesque D and
Weis J-J 1990 J. Chem. Phys. 92 6818

[7] Rosenfeld Y, Schmidt M, Löwen H, and Tarazona P 1996 Phys. Rev. E 55 4245

[8] Rosenfeld Y, Schmidt M, Löwen H, and Tarazona P 1996 J. Phys.: Condens.

Matter 8 L577

[9] Tarazona P and Rosenfeld Y 1997 Phys. Rev. E 55 R4873

[10] Tarazona P and Rosenfeld Y 1999 in New Approaches to Problems in Liquid

State Theory (Dordrecht: Kluwer Academic) p 293

[11] Tarazona P 2000 Phys. Rev. Lett. 84 694

[12] Tarazona P 2002 Physica A 306 243

[13] See, e.g., Hansen J P and McDonald I R 1986 Theory of Simple Liquids (London:
Academic Press)

[14] Mansoori G A, Carnahan N F, Starling K E, and Leland Jr. T W 1971 J. Chem.

Phys. 54 1523

[15] Carnahan N F and Starling K E 1969 J. Chem. Phys. 51 635

47
48 BIBLIOGRAPHY

[16] Tarazona P 1985 Phys. Rev. A 31, 2672

[17] Gonzalez A, White J A, and Evans R 1997 J. Phys.: Condens. Matter 9 2375

[18] Percus J K 1976 J. Stat. Phys. 15 505

[19] Vanderlick T K, Davis H T, and Percus J K 1989 J. Chem. Phys. 91 7136

[20] Kierlik E and Rosinberg M L 1990 Phys Rev A 42 3382

[21] Phan S, Kierlik E, Rosinberg M L, Bildstein B, and Kahl G 1993 Phys Rev E 48
618

[22] Rosenfeld Y 1994 Phys. Rev. E 50 R3318

[23] Rosenfeld Y 1995 Mol. Phys. 86 637

[24] Reiss H, Frisch H L, Helfand E, and Lebowitz J L 1960 J. Chem. Phys. 32 119

[25] González A, White J A, Román F L, Velasco S, and Evans R 1997 Phys. Rev.
Lett. 79 2466

[26] González A, White J A, Román F L, and Evans R 1998 J. Chem. Phys. 109 3637

[27] Schmidt M, Löwen H, Brader J M, and Evans R 2000 Phys. Rev. Lett. 85 1934

[28] Cuesta J A, Martinez-Raton Y, and Tarazona P 2002 J. Phys.: Condes. Matter 14

11965

[29] Groot R D, van der Eerden J P, and Faber N M 1987 J. Chem. Phys. 87 2263

[30] Groot R D, Faber N M, and van der Eerden J P 1987 Mol. Phys. 62 861

[31] van Swol F and Henderson J R 1987 Phys. Rev. A 40 2567

[32] Tarazona P and Evans R 1984 Mol. Phys. 52 847

[33] Noworyta J P, Henderson D, Sokołowski D, and Chan K-Y 1998 Mol. Phys. 95
415

[34] Roth R and Dietrich S 2000 Phys. Rev. E 62 6926

[35] Hoover W G and Ree F H 1968 J. Chem. Phys 49 3609

[36] Young D A and Alder J 1974 J. Chem. Phys 60 1254

[37] Alder B J, Hoover W G, and Young D A 1968 J. Chem. Phys. 49 3688

BIBLIOGRAPHY 49

[38] Bryk P private communication

[39] R. Roth, R. Evans, A. Lang, and G. Kahl, J. Phys.: Condens. Matter 14, 12063
(2002).

[40] Y.-X. Yu and J. Wu, J. Chem. Phys. 117, 10156 (2002).

[41] H. Hansen-Goos and R. Roth, J. Chem. Phys. 124, 154506 (2006).

[42] H. Hansen-Goos and R. Roth, J. Phys.: Condens. Matter 18, 8413 (2006).