[D2] LDA + DMFT Investigation of NiO

LDA + DMFT Investigation of NiO
Von der Mathematisch-Naturwissenschaftlichen Fakultät

der Universität Augsburg
zur Erlangung eines Doktorgrades der Naturwissenschaften
genehmigte Dissertation
von
Xinguo Ren
aus
China
November 2005
Vorsitzender: Priz-Doz Dr. Alexander Krimmel
Erstgutachter: Prof. Dr. Dieter Vollhardt
Zweitgutachter: Priz-Doz Dr. Volker Eyert
Prüfer: Prof. Dr. Ulrich Eckern, Prof. Dr. Siegfried Horn
Tag der mündlichen Prüfung: 16.02.2006

Contents i
CONTENTS
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1. Electronic-Structure Calculations with Density Functional Theory 5

1.1 Density Functional Theory . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Single-Particle Description and Local Density Approximation . . . 8
1.2.1 Kohn-Sham Approach . . . . . . . . . . . . . . . . . . . . 8
1.2.2 Local Density Approximation . . . . . . . . . . . . . . . . 9
1.3 The Linear Muffin-Tin Orbital Method . . . . . . . . . . . . . . . 11
1.3.1 Energy Band Methods . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Linear Muffin-Tin Orbitals . . . . . . . . . . . . . . . . . . 13
1.3.3 The LMTO Band Calculation . . . . . . . . . . . . . . . . 16
1.4 Wannier Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2. The LDA+DMFT Approach . . . . . . . . . . . . . . . . . . . . . . 23

2.1 The Hubbard Model . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Dynamical Mean-Field Theory . . . . . . . . . . . . . . . . . . . . 27
2.2.1 The Infinite Dimension Limit . . . . . . . . . . . . . . . . 27
2.2.2 DMFT Equations — Exact Solution of the D = ∞ Hub-
bard Model . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.3 Quantum Monte-Carlo Method . . . . . . . . . . . . . . . 31
2.2.4 Case Studies with DMFT . . . . . . . . . . . . . . . . . . 35
2.3 The LDA+DMFT Formulation . . . . . . . . . . . . . . . . . . . 38
ii Contents
3. LDA+DMFT Investigation of NiO . . . . . . . . . . . . . . . . . . . 43

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.1 Crystal Structure . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.2 Electronic and Magnetic Properties . . . . . . . . . . . . . 44
3.1.3 Previous Studies . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Method and Results . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2.1 Wannier Function Construction . . . . . . . . . . . . . . . 51
3.2.2 LDA Results . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.3 LDA+DMFT Results and Comparison with Experiment . 56
3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4. Towards A Self-Consistent LDA+DMFT Scheme . . . . . . . . . . 63

4.1 General Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Flow Diagram and Formulation . . . . . . . . . . . . . . . . . . . 64
4.3 Difficulties and Challenges . . . . . . . . . . . . . . . . . . . . . . 68
5. Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Appendix 73
A. Proof That The External Potential Is A Unique Functional of The

Ground-State Density . . . . . . . . . . . . . . . . . . . . . . . . . . 75
B. Estimation of The Local Magnetic Moment and The Coulomb In-

teraction Parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
C. Energy Moment Calculation from Matsubara Green function . . 79
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Curriculum Vitae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Introduction 1
INTRODUCTION
From a microscopic ab-initio point of view, a solid material is an interacting

many-particle system involving both ions and electrons. However, according to
the Born-Oppenheimer approximation (Born and Oppenheimer, 1927), the ions
and electrons can be treated separately. Indeed, many properties of solids can be
well described by the electronic degree of freedom, while the ions only contribute
through a static potential. The resultant electronic Hamiltonian is still far too
complicated to be fully solved and immediately requires further approximations,
among which the one-electron approximation plays an important role. Within
this approximation, the electron-electron interaction is taken into account at a
mean-field level, behaving like an effective potential, and therefore the many-
electron problem reduces to a single-electron one described by a single-electron
Schrödinger equation. Solving this Schrödinger equation leads to the energy
band theory of solids. The effective potential can be determined in different
ways, governed by different approximations, among which the Hartree-Fock (HF)
approximation (Hartree, 1928; Fock, 1930) is the most famous example. At
present the most satisfactory picture for single-electron theory is based on the
density functional theory (DFT) (Hohenberg and Kohn, 1964) through Kohn-
Sham’s approach (Kohn and Sham, 1965). The single-particle approximation is
very successfull for explaining the properties of weakly correlated systems, e.g.,
simple metals, ordinary insulators and some semiconductors, but generally fails
for systems with strongly correlated electrons such as Mott insulators, cuprates,
manganites and rare earth systems. A satisfactory description of these systems
requires an explicit treatment of the interactions between electrons.
The problem of understanding the properties of strongly correlated materials is
one of the main challenges for modern condensed matter physics. In this case one
has to go beyond the one-electron approximation and employ more sophisticated
treatment of electron-electron interaction. For that purpose, practically one has
to restrict oneself to the most important orbitals so that the many-electron in-
teractions can be explicitly treated. For instance, the valence d electrons are the
most relevant ones responsible for the properties of transition metal compounds,
and a model Hamiltonian can be formulated involving only these electrons. The
simplest model appropriate for the strongly correlated electrons is one-band Hub-
bard model (Gutzwiller, 1963; Hubbard, 1963), in which only the on-site Coulomb
2 Introduction
interaction is considered, whereas the long-range ones are neglected. However,

it turned out even for such a highly simplified model it is still very difficult to
solve, and the exact solution only exists for one dimension. Therefore various
approximations were developed to gain insights into the model and to arrive at
some understandings of the experimental behaviors for real materials in the end.
Dynamical mean-field theory (DMFT) is such an approximation which maps the
the lattice model onto a quantum impurity model subject to a self-consistent con-
dition (Metzner and Vollhardt, 1989b; Georges and Kotliar, 1992; Jarrell, 1992).
The particular advantages of DMFT lie in two aspects: first, it is a “controlled"
approximation meaning that it has a well-defined limit— the infinite coordination
number (or the infinite dimensions) where the theory becomes exact (Metzner and
Vollhardt, 1989b); second, the practical solution of DMFT consists in solving an
effective Anderson impurity model (Anderson, 1961) iteratively and several ana-
lytical and numerical techniques have existed to deal with it. The application of
DMFT to Hubbard model has produced fruitful results and substantial progress
has been made in understanding the nature of Mott metal-insulator transition
(Georges et al., 1996; Rozenberg et al., 1999; Blümer, 2002).
The model Hamiltonian approach is helpful in understanding some qualitative
features or identifying the underlying physical mechanism of the strongly corre-
lated systems, but it can’t explain the detailed features of real materials. This
is not surprising since the material-specific information can’t be contained in a
highly-simplified, Hubbard-like model. One the other hand, the efforts of de-
scribing the real materials at a quantitative level persist, and for that purpose
the model Hamiltonian approach has to be used with the help of the “ab-initio"
approach for incorporating the material-specific information. This is actually the
basic idea of the LDA+DMFT approach formulated by Anisimov et al (1997b)
(see also Lichtenstein and Katsnelson (1998)) in which the band-structure cal-
culation based on DFT within its local density approximation (LDA) and the
DMFT treatment of the localized orbitals are combined. The strategy here is
based on the observation that although LDA often leads to qualitatively wrong
results for the strongly correlated materials, it can usually provides quite reli-
able parameters for these systems. These parameters can be in turn used to
construct a many-body Hamiltonian which is specific for the particular material
under study. In most of the practical applications of LDA+DMFT, one first
performs a LDA band calculation to drive a material-specific generalized model
Hamiltonian, and solve this Hamiltonian by DMFT.
In the past few years LDA+DMFT approach has been successfully applied to
transition metals, e.g. nickel (Lichtenstein et al., 2001), transition metal com-
pounds, e.g. La1−x Srx TiO3 (Anisimov et al., 1997b; Nekrasov et al., 2000),
LiV2 O4 (Nekrasov et al., 2003), Ca(Sr)VO3 (Nekrasov et al., 2005), V2 O3 (Held
et al., 2001a; Keller, 2005), and f -electron systems such as plutonium (Savrasov
et al., 2001) and cerium (Held et al., 2001b). In this thesis, we will use the
Introduction 3
LDA+DMFT approach to investigate the electronic structure of NiO. NiO is a

classical Mott insulator which has been under intensive studies for many years.
The recent theoretical investigations mainly fall into two categories, i.e., the cal-
culations from first principles and that based on the localized cluster model.
However, a satisfactory description of its electronic spectrum is still not avail-
able, and this is due to the reason that the first principles studies usually can not
treat the strong local Coulomb interaction adequately whereas the local cluster
approach completely ignores the band effect which also plays an important role
in this system. In this connection it is very interesting to see if the LDA+DMFT
approach works better for NiO, considering its previous successes for strongly
correlated materials. It turns out that within the LDA+DMFT approach the
calculated energy gap and local magnetic moment are in good agreement with
experiment, and the obtained electronic energy spectrum shows impressive quan-
titatively improvement over previous results.
The plan of this thesis is as follows. In chapter 1 we give an account of the

density-functional based band structure calculations which is the starting point
for performing a LDA+DMFT calculation. In particular, Emphasis will be given
to LDA which is the most commonly used approximation for carrying out the
self-consistent band structure calculations. Then we will introduce one of the
most favorable methods for calculating the band structures of transition metal
compounds, namely linear muffin-tin orbital (LMTO) method (Andersen, 1975),
within which the LMTO basis is used for solving the one-electron Schrödinger
equation. Finally we discuss the concept of Wannier functions (WFs) and its
historical development, and point out its usefulness in realistic modelling of ma-
terials with localized orbitals.
An introduction of the LDA+DMFT approach in general is then presented in

Chapter 2. In this chapter we first give an elementary review of the stereotyped
strongly-correlated fermionic lattice model, namely the Hubbard model. This
is followed by a presentation of the DMFT equations, illustrating how a lattice
model, in the limit of infinite dimension, can be mapped to a single-site quantum
impurity embedded in an average medium. As a powerful, numerically exact
solver of the quantum impurity problem, the Hirsch-Fye quantum Monte-Carlo
(QMC) method (Hirsch and Fye, 1986) is then described. Finally the general
motivation of combining the many-body technique-DMFT, and the state-of-the-
art band structure method-DFT(LDA) is discussed. This naturally leads to the
LDA+DMFT approach. In particular we show how the LDA band-structure
is incorporated into the DMFT equations, giving rise to the formalism of the
practical LDA+DMFT scheme.
In chapter 3, we apply the LDA+DMFT approach to the prototypical Mott

insulator-NiO. First in the introduction the main properties and the previous
studies of NiO are reviewed. In addition we point out why it is worthwhile to
4 Introduction
perform a LDA+DMFT study of NiO. We then describe the new procedure for
implementing the LDA+DMFT scheme, in which a set of WFs are constructed
and used as the basis for the DMFT calculation. The LDA+DMFT scheme is
applied to calculate the electronic properties of NiO, and the obtained results are
presented and compared with experiment. This chapter is closed with comments
on the successfull aspects and limitations of the present study, and the possible
directions of improvement.
The possible extensions of the present LDA+DMFT scheme are discussed in
Chapter 4. This consists of two respects: firstly, not only the transition metal d
but also ligand p orbitals should be included in the DMFT calculation when there
is strong hybridization between them, and secondly the LDA part and DMFT part
should be merged self-consistently rather than performed in a subsequent order,
as done in the present implementation. Here we follow a fully self-consistent
scheme recently proposed by Anisimov et al. (2005), and its implementation is a
still ongoing work.
Chapter 5 concludes this thesis with a summary and outlook.
5
1. ELECTRONIC-STRUCTURE
CALCULATIONS WITH DENSITY
FUNCTIONAL THEORY
Density-functional theory (DFT) is nowadays a popular and successful approach

to study the ground-state properties of an interacting many-particle system, in-
cluding atoms, molecules, and crystalline solids. This approach concentrates on
the electron density n(r) instead of the much more complicated many-body wave
function Ψ(r1 , r2 , ..., rN ), the solution of the latter is an impossible task for an
interacting system with more than a few electrons. The idea of using electron
density n(r) instead of the wave function Ψ(r1 , r2 , ..., rN ) as the basic variable to
study many-body systems dates back to the Thomas-Fermi (TF) model proposed
by Thomas (1927) and independently by Fermi (1928) in late 1920s. However, the
framework of DFT was put on a firm rooting only after the work of Hohenberg
and Kohn (1964), known as Hohenberg-Kohn (HK) theorems.
1.1 Density Functional Theory
To get an idea of what the HK theorems are, let us start by considering a system
with N interacting electrons moving in an external static potential v(r). For this
system the many-electron Hamiltonian reads
X
Ĥ = T̂ + Û + v(ri ), (1.1)
i
where
X ~ 2 ∇2
T̂ = , (1.2)
i
2m
X e2
Û = (1.3)
i<j
|ri − rj |
are the kinetic and electron-electron interaction operator respectively, and m

in (1.2) is the electron mass. We note that under the Born-Oppenheimer ap-
proximation, a N-electron Coulomb system is specified solely by the form of the
6 1. Electronic-Structure Calculations with Density Functional Theory
external potential v(r), since both T̂ and Û are universal. HK showed that for
a given ground-state density n(r), the external potential v(r) can be uniquely
determined up to an unimportant constant (for a proof, see Appendix A). Since
v(r) in turn fixes the full N-electron Hamiltonian, it is clear that the ground state
wave function Ψ0 , and in particular the kinetic energy hT̂ i0 = hΨ0 |T̂ |Ψ0 i and the
interaction energy hÛi0 = hΨ0 |Û |Ψ0 i, are all functionals of n(r). Therefore one
can define a universal functional including only the kinetic and integration energy
as
F [n] = hΨ0 [n]|T̂ + Û |Ψ0 [n]i = T [n] + U[n] (1.4)
which does not refer to any external potential v(r). And Ψ0 [n] here is the ground
state wave function associated with some particular density n(r).
Now suppose we have some arbitrary external potential v(r), and associated with
it can we define the following energy functional,
Z
Ev [n] = v(r)n(r)dr + F [n]. (1.5)
Note that in Eq. (1.5) n(r), as the basic variable of the functional, is not neces-
sary to be the ground-state density associated with v(r) here.1 However, it can be
easily shown that Ev [n] assumes its minimum at the ground-state density n0 (r)
of the present system, i.e., associated with v(r). Thus we have briefly demon-
strated the essential ideas of the HK theorems which state that for an interacting
electronic system there exist an energy functional of the electron density, and this
functional is minimized by the ground-state density. Combining Eqs. (1.4) and
(1.5) Ev [n] can be written as
Z
Ev [n] = v(r)n(r)dr + T [n] + U[n]. (1.6)
The original proof of the HK theorems is given in the space of V-representable

electron densities. Levy (1979) and independently Lieb (1983) generalized the
proof to the N-representable2 electron-density space through the approach of
“constrained search”. Now it has been known that all the non-negative function
of electron density is N-representable.
The energy-functional Ev [n] is easy to write down, but its explicit form is not
known. Thus one has to make approximations to T[n] and U[n] before doing
any practical calculations based on the variational principle. The TF model
was actually one particular approximation to Ev [n] in which the electrons are
1
But it should be the ground-state density corresponding to some other v 0 (r) in this context.
This is the so-called “V-representability”.
2
N-representability means that the electronR density
R can be realized for some antisymmetric
N-electron wave function, i.e., n(r) = N ... Ψ∗ (r, r2 , ..., rN )Ψ(r, r2 , ..., rN )dr2 ...drN .
1.1. Density Functional Theory 7
treated as independent particles and the interaction energy is approximated by

the electrostatic energy. This model was frequently use in the past, but there
are serious deficiencies within it, e.g., for atoms the electron density decays too
slow far away from the nucleus3 , and for molecules and solids the chemical bonds
calculated with this model are not stable, and so on. These deficiencies can be
by large ascribed to the local approximation to the kinetic energy4 ,
Z
T [n] ≈ drt0 [n(r)]. (1.7)
Here t0 [n] = (3~2 /10m)(3π 2 )2/3 n5/3 is the kinetic energy density of a noninter-
acting homogeneous electron gas with a constant density n. Actually two kinds
of approximations are involved in (1.7), the first is the local approximation which
assumes that the kinetic energy density at some particular spatial point only de-
pends on the density precisely at that point, and the second is to use the kinetic
energy density of the noninteracing system to replace that of the interacting one
since the latter is not known.
The drawbacks arising the local approximation to T [n] was removed through
Kohn-Sham (KS) approach (Kohn and Sham, 1965), which maps a system with
interacting electrons to one with noninteracing electrons moving in an effective
potential. This mapping is achieved by introducing auxiliary single-particle or-
bitals, by means of which the noninteracing part of the kinetic energy can be
treated exactly. This represents a substantial improvement over TF model, and
many pathologies are thus cured. Furthermore, a single-particle picture arises
with a set of self-consistent equations which are analogous to the Hartree-Fock
(HF) equations. The resultant effective potential includes the external static po-
tential, the Hartree or electrostatic potential, and the remaining part known as
exchange-correlation potential. KS equations play a fundamental role in DFT.
Nowadays electronic-structure calculations based on DFT through KS approach
are routinely performed for atoms, molecules and solids, and the application of
DFT to organic materials has just appeared. A large number of review arti-
cles and books exist, and here we only list a few of them, e.g., Lundqvist and
March (1983); Parr and Yang (1989); Jones and Gunnarsson (1989); Dreizler and
Gross (1990). An excellent elementary introduction into DFT was given recently
by Capelle (2003). In this thesis we are only concerned with the application of
DFT to crystalline solid, where the single-particle picture arising from the KS
approach leads to a band theory due to the periodicity of the effective potential.
3
The electron density for a single atom calculated with TF model decays as power law (1/r6 )
away from the nucleus, whereas the physically correct behavior should be an exponential
decay
4
This is effectively an approximation that the motion of the electrons are treated as inde-
pendent particles, and should be distinguished from the local density approximation to the
exchange-correlation functional.
1.2. Single-Particle Description and Local Density Approximation 9
where Z
n(r0 )
φ(r) = dr0 . (1.11)
|r − r0 |
is the Hartree potential. The original problem given by Eq. (1.10) is mathemat-
ically identical to the one of a system of noninteracting electrons moving in an
effective potential
δExc [n]
vef f (r) = v(r) + φ(r) + . (1.12)
δ(r)
The latter problem can be solved by the single-particle Schrödinger equation,
2
~ 2
∇ − vef f (r) ψi (r) = i ψi (r) (1.13)
2m
which is required to yield the same electron density as the interacting electron
system5 ,
N
X
n(r) = |ψi (r)|2. (1.14)
i=1
Equations (1.11) to (1.14) are the famous KS equations which have to be solved
self-consistently.
For crystalline solids, the periodicity can be fully retained in the effective po-
tential vef f (r), and the effective single-particle problem (1.13) naturally leads to
the Bloch’s energy band theory. In this connection an approximation has been
implicitly invoked to interpret the auxiliary single-particle eigenvalues i in (1.13)
as the physical excitation energies. In practice such an interpretation is found to
be a good approximation for weakly correlated systems.
1.2.2 Local Density Approximation
As has been shown, KS theorems guarantee that the ground-state energy of a

quantum many-electron system can be obtained by minimizing an energy func-
tional Ev [n] with respect to the electron density, and KS approach maps the
problem of minimizing Ev [n] to a set of self-consistent equations for a single
electron. Thus KS theorems and KS mapping together provide a single-particle
description of interacting many-particle systems. So far these two steps are both
exact in principle. However, as mentioned before, for any practical implemen-
tation of DFT, one has to introduce approximation to the exchange-correlation
energy functional Exc [n] defined in expression (1.9).
5
In case that the electron density P
is not representable by a single Slater determinant, one
N
can replace Eq. (1.14) by n(r) = i=1 fi |ψi (r)|2 , in which the states above the Fermi level
can be occupied and holes can be left below the Fermi level.
The most popular approximation that has been used for decades is the local
density approximation (Kohn and Sham, 1965), which is usually expressed as,
Z
Exc [n] = drn(r)xc (n(r)). (1.15)
Here xc (n(r)) is the exchange-correlation energy per electron for a homogeneous
gas of interacting electrons with constant density n. The basic idea behind it is
to separate the whole inhomogeneous electron system into infinitely small pieces
and treat every piece as if its neighbors do not have influences on it. This kind of
approximation has appeared for treating the kinetic energy in TF theory where
it is quite problematic. However, the LDA treatment of Exc [n] proved to be very
successful and this is due to the reason that the nonlocal correction to Exc [n] is
relatively small in cases that the variation of n(r) is not too rapid.
Now let’s have a closer look at xc (n). xc (n) consists of two components: the
exchange energy x (n) and correlation energy c (n) per electron. x (n) is known
exactly6
3e2 3n 1/3
x (n) = − ( ) , (1.16)
4 π
whereas the precise form of c (n) is not known. The study of c (n) by itself is a
very difficult problem in many-body theory, and the best description of c (n) so
far is given numerically by Quantum Monte Carlo method (Ceperley and Alder,
1980). The practical expression of c (n) in the modern calculations is based on
the parameterization of these numerical data.
LDA has been successfully used in the band-structure calculations of quite a
large number of solid state systems, but it failed for one particular group of ma-
terials, namely those with strongly correlated electrons. Most of the transition
metals and their compounds, as well as rare-earth materials, belong to this cat-
egory. Another example is chemistry where LDA is usually not accurate enough
to describe quantitatively the chemical bonding in molecules. These problems
call for better approximations beyond LDA, and among many of them we here
only mention a particular one which is commonly used in chemistry, known as
generalized-gradient approximation (GGA). Instead of considering xc (n(r)) is
a local function of n(r), GGA takes it as a function of n(r) and its gradient
∇n(r)(Perdew and Wang, 1986),
xc (n(r)) = f (n(r), ∇n(r)). (1.17)
GGA enjoys a great success in chemistry by giving reliable results of the chemical
bonding, and often improves over the LDA results for the strong-correlated solid
6
See, e.g. Mahan (1990, p. 385), x (n) = −(3/4π)kF e2 where kF = (3π 2 n)1/3 is the Fermi
wave vector.
1.3. The Linear Muffin-Tin Orbital Method 11
materials. Different choices of the form of f (n, ∇n) represents different kinds of
GGA, and one has the freedom to choose a best one appropriate for the particular
type of system under investigation.
1.3 The Linear Muffin-Tin Orbital Method
The linear muffin-tin orbital (LMTO) method is one particular technique for
solving the one-electron problem in crystalline solids. Among the many methods
of solving band-structure problems, the LMTO method is often more favorable
because it is relatively easy to implement and computationally cheap, and it has
the accuracy required in most cases. In this section we first briefly review the
energy band methods in general, and then discuss the LMTO method specifically.
1.3.1 Energy Band Methods
DFT through KS Ansatz offers a self-consistent way of calculating the band

structures of crystalline solids. The energy bands of electrons arise from the
translational symmetry of crystals and determine many physical properties of
the system. To calculate the band structures accurately and efficiently is one of
the basic tasks in solid state physics. Lot’s of experience had been gained much
earlier before DFT was widely accepted as an efficient tool for band-structure
calculations.
Indeed, even without considering any self-consistency, solving the Schrödinger
equation of a single electron moving in a given, periodic potential is a highly
nontrivial problem. To be specific, we consider the following problem
2
~ 2
∇ − v(r) ψjk (r) = Ejk ψjk (r) (1.18)
2m
where the potential v(r) is translationally invariant,
v(r + R) = v(r), (1.19)
with R being a lattice constant. The eigenfunctions ψjk (r) in Eq. (1.18), known
as Bloch functions, have been chosen to be simultaneously the eigenfunctions of
both the Hamiltonian operator Ĥ and the translation operator T̂ , and hence are
labelled by both the band index j and Bloch vector k. The justification for doing
so is provided by the Bloch theorem,
T̂R ψjk (r) ≡ ψjk (r + R) = eik·R ψjk (r). (1.20)
The Bloch vector k is a vector in the reciprocal space, and is usually chosen to
be restricted inside the first Brillouin zone.
The different energy-band methods differ from one another by the set of func-
tions chosen as the basis to expand the unknown eigenfunctions ψjk (r). Histori-
cally, these methods are divided into two classes: one works with fixed, energy-
independent basis functions, like plane waves, atomic orbitals, or orthogonalized
plane waves (OPW) (Herring, 1940), and the other uses energy-dependent basis,
in particular the partial waves. Examples for the latter are the cellular (Wigner
and Seitz, 1933), augmented plane wave (APW) (Slater, 1937), and Korringa-
Kohn-Rostoker (KKR) (Korringa, 1947; Kohn and Rostoker, 1954) methods.
Both of these methods have advantages and drawbacks. The fixed basis method,
say LCAO (linear combination of atomic orbitals), has the advantage that the
variational procedure for one-electron Hamiltonian leads to an algebraic eigen-
value problem so that all the eigenvalues and eigenvectors at a given k point
can be obtained by a single diagonalization. However, this method requires a
large number of atomic orbitals to form a complete basis set, and the Hamil-
tonian matrix elements involve a lot of two- and there-center integrals which
are very difficult to calculate. On the other hand, the methods employing the
energy-dependent partial waves as basis functions can have good accuracies with
a smaller basis set, but the resultant secular matrix has a nonlinear energy-
dependence so that the eigenvalues can only be found one by one, thus requiring
much more computation time than the linear problem.
Under this background a linear procedure was proposed by Andersen (1975) in

order to combine the positive features of both kinds of energy band methods
but avoid their difficulties. The idea is to linearize (Taylor expanded up to first
order) the energy-dependence of the partial waves around some arbitrary but
fixed energy points. The linear energy-independence vanishes by a proper linear
combination of the partial wave functional and their energy derivatives at these
energy energy points (one energy point for each partial wave), leading to energy-
independent orbitals. With these energy-independent basis functions, the secular
equations of the eigenvalue problem become linear in energy. The linear method
was first applied to the muffin-tin orbitals (MTOs) (Andersen and Wooley, 1973)
and then to augmented plane waves, leading to linear muffin-tin orbitals (LM-
TOs) and linear augmented plane waves (LAPWs). The LMTOs and LAPWs,
as their conventional counterparts are defined with respect to the muffin-tin po-
tential which is a reasonable approximation to the real crystal potential. In this
approximation, a so-called muffin-tin sphere is inscribed inside each atomic poly-
hedron: inside the sphere the potential is assumed to be spherically symmetric,
and out of the spheres it is flat. A schematic behavior of the MT potential is
shown in Fig. 1.1. In the next section we will give an illustration of how the
LMTOs are constructed.
Figure 1.1: A schematic picture of the MT potential. The potential well should
go to infinity at the center of the atom.
1.3.2 Linear Muffin-Tin Orbitals
A detailed description of the LMTO method was given by Skriver (1984), which
we are following here. To arrive at a final definition of LMTO, several steps are
needed to take. First, instead of treating a full MT potential, we only consider
a single MT well embedded in a constant potential environment. Namely we are
dealing with a single-electron problem with the following potential,

V (r) r ≤ SMT ,
v(r) = (1.21)
VMTZ r ≥ SMT ,
where V (r) is the spherically symmetric potential inside the MT sphere, and
VMTZ is constant potential outside the sphere, with SMT the radius of the MT
sphere. For convenience we define

V (r) − VMTZ r ≤ SMT ,
VMT (r) = v(r) − VMTZ = (1.22)
0 r ≥ SMT ,
and
κ2 = E − VMTZ . (1.23)
Therefore the Schrödinger equation of a single electron moving in the potential
v(r) with behaving like (1.21) reads

~2
− 2
+ VMT (r) − κ ψL (E, r) = 0. (1.24)
2m
Due to the spherical symmetry of the total potential v(r) under consideration, the
eigenfunction ψL (E, r) can be classified by the combined angular and magnetic
quantum number L = lm is , and and can be written as a product of a radial
part and an angular part, i.e.,
ψL (E, r) = il Ylm (r̂)ψl (E, r). (1.25)

The radial part of the Schrödinger equation satisfied by ψl (E, r) inside the MT
sphere and in the constant potential region respectively look like,
2 2
~ d l(l + 1)
+ + VMT (r) − κ rψl (E, r) = 0, for r ≤ SMT , (1.26a)
2
2m dr 2 r2
2 2
~ d l(l + 1)
+ − κ rψl (E, r) = 0, for r ≥ SMT . (1.26b)
2
2m dr 2 r2
Leaving Eq. (1.26a) aside for a while, let us concentrate on the Helmholtz equation
(1.26b) which has two linearly independent solutions. For a positive κ2 , these
are the spherical Bessel function jl (κr) and Neumann function nl (κr), and for
κ2 < 0, i.e., the kinetic energy is negative in the constant potential region, the
Neumann function nl (κr) should be replaced by the first kind Hankel function
(1)
−ihl = nl − ijl . Here we only present the formulations of the positive κ2 case,
and those for κ2 < 0 can be obtained by a simple replacement.
Summarizing the above analysis, we can have the partial waves solving the
Schrödinger equation of (1.24) ,

 ψl (E, r) r ≤ SMT ,
l m
ψL (E, κ, r) = i Yl (r̂) (1.27)

κ[nl (κr) − cl (E, κ)jl (κr)] r ≥ SMT .
Here the coefficient cl (E, κ), usually expressed as cot(ηl (E, κ)), is determined so
that ψL (E, κ, r) is continuous and differentiable across the boundary of the MT
sphere. This requires
nl (κSMT ) Dl (E) − D{nl }
cl (E, κ) = cot(ηl (E, κ)) = · , (1.28)
jl (κSMT ) Dl (E) − D{jl }
where
SMT dψl (E, r)
Dl (E) = ,
ψl (E, SMT ) dr r=SMT
SMT dn(κr)
D{nl } = , (1.29)
nl (κSMT ) dr r=SMT
SMT dj(κr)
D{jl } = ,
jl (κSMT ) dr r=SMT
are the logarithmic derivative of ψl (E, r), n(κr), and j(κr) at the sphere boundary
respectively. The ηl (κ, E) defined in (1.28) can be view as the phase shift of the
free spherical wave for r → ∞ due to the scattering of the MT potential.
The partial waves (1.27) are not suitable for serving basis functions. This is
particularly because the presence of the term −κcl (E, κ)jl (κr) in the constant
potential region make them not normalizable for negative κ2 . The trick that can
be employed here to cure this problem is to subtract this term from ψl (E, r) in
both regions (inside and outside the MT sphere) while maintaining the continuity
and differentiability, ending up with

 ψl (E, r) + κcl (E, κ)jl (κr) r ≤ SMT ,
l m
χL (E, κ, r) = i Yl (r̂) (1.30)

κnl (κr) r ≥ SMT .
These orbitals χL (E, κ, r) are actually the energy dependent MTOs. Although
they are not the solutions of the problem (1.24), the Bloch sum of χL (E, κ, r) and
ψL (E, κ, r) give the identical results except for the k points satisfying |k + G|2 =
k 2 with G being the reciprocal lattice vector. In addition, they are reasonably
localized, and regular over the whole space.
In (1.27) and (1.30), the parameter E and κ are related through Eq. (1.23).
However, the continuity and differentiability of ψL (E, κ, r) and χL (E, κ, r) are
guaranteed by the chosen value (1.28) of cl (E, κ) irrespective of their possible
relation between E and κ. In this connection we can disregard (1.23) and treat κ
as an independent parameter. By doing so the tails of ψL (E, κ, r) and χL (E, κ, r)
are no longer the exact solution of the Schrödinger equation (1.24) in the region
of the consant potential any more, but they have the advantage of being energy
independent. Moreover, the head of χL (E, κ, r) (i.e., the part inside the MT
sphere) can also be made energy independent around a fixed energy Eν up to the
first order by replacing (augmenting) jl (κr) and nl (κr) inside the MT sphere by
more appropriate functions which are attached to the original functions at the
sphere boundary in a continuous and differentiable fashion. For this purpose, we
define the augmented Bessel function Jl (κr) as

 −ψ̇l (Eν , r)/(κċl (Eν , κ)) r ≤ SMT
Jl (κr) = (1.31)

jl (κr) r ≥ SMT
where ψ̇l and c˙l are the energy derivative of ψl and cl respectively. It is easy to
verify that Jl (κr) defined in (1.31) is everywhere continuous and differentiable.
A proper definition of the augmented Neumann function Nl (κr) is more delicate.
Before giving an explicit form of Nl (κr), it is illustrating to present the following
expansion theorem of nL (κ, r) = nl (κr)il Ylm (r̂) and jL (κ, r) = jl (κr)il Ylm (r̂),
namely, XX
nL (k, r) = 4π CLL0 L00 jL0 (κ, r − R)n∗L00 (κ, −R) (1.32)
L0 L00
which is valid inside the sphere |r| < R. Here the Gaunt coefficients CLL0 L00 are
defined as Z
0 m00
CLL0 L00 = Ylm (r̂)Ylm
0 (r̂)Yl00 (r̂)dr̂. (1.33)
The band theory of solid crystals, initiated by Bloch, Brillouin and Wilson, has
been tremendously advanced since the emergence of DFT.
Practically, it is inevitable to introduce approximations to the exchange-correlation
potential. The local density approximation (LDA), which we will discuss below,
is the most frequently used one.
1.2 Single-Particle Description and Local Density

Approximation
DFT is turned into a tractable framework through Kohn-Sham approach, or

Kohn-Sham ansatz (Kohn and Sham, 1965), which assumes that a system of in-
teracting particles can be represented by one of noninteracting particles moving in
an effective potential. This potential contains an unknown exchange-correlation
term, and approximations have to be employed to deal with this term. In this
section which first discuss the Kohn-Sham approach, and then introduce the local
density approximation (LDA).
1.2.1 Kohn-Sham Approach
Among the different parts of contributions to the electronic energy, the exter-
nal potential energy and the classic electron-electron interacton energy can be
expressed explicitly in terms of electron density n(r), and all the remaining con-
tributions, denoted as G[n], are not known explicitly as a functional of n(r). Thus
we can rewrite the energy functional Ev [n] Eq. (1.6).
Z Z Z
1 n(r)n(r0 )
Ev [n] = drv(r)n(r) + dr dr0 + G[n]. (1.8)
2 |r − r0 |
KS further separate G[n] into Ts [n] and Exc [n],
G[n] = Ts [n] + Exc [n], (1.9)
where the Ts [n] is the kinetic energy for a system of noninteracing electrons
with density n(r), and Exc [n] is the remaining parts, defined as the exchange-
correlation energy.
Combining Eqs. (1.8) and (1.9), and applying the variational principle, one
arrives at the following Euler equation,
δTs [n] δExc [n]

+ v(r) + φ(r) + = 0, (1.10)
δn(r) δn(r)
The augmented spherical Neumann and Bessel functions are also required to
satisfy the above expansion theorem, and this lead to the following definition of
Nl (κr), including the angular part,
 P P

 4π L0 L00 CLL0 L00 jL0 (κ, r − R)n∗L00 (κ, −R)

|r − R| ≤ SMT , ∀R 6= 0
NL (κ, r) = (1.34)



nL (κ, r) otherwise.
To have a clear understanding of (1.34), one may think of the full MT potential
composed of nonoverlapping array of MT wells: inside every MT sphere except
the one where the present Neumann function is centered, NL (κ, r) is defined as
the linear expansion of the augmented Bessel functions centered at that particular
MT sphere. In other regions, both the MT sphere at the origin and the interstitial
region, the augmented Neumann function is simply defined as the normal one.
With JL (κ, r) and NL (κ, r) defined, we finally end up with the following definition
of the augmented MTO

 ψl (E, r) + κcl (E, κ)Jl (κr) r ≤ SMT ,
l m
χL (E, κ, r) = i Yl (r̂) (1.35)

κNl (κr) r ≥ SMT .
The augmented MTO defined in (1.35) is energy independent up to the first order
in (E −Eν ). It is everywhere continuous and differentiable, and it is orthogonal to
the core states. By neglecting the high-order energy dependence of JL (κ, r), i.e.,
fixing E = Eν , we are led to the linear (energy independent) MTOs (LMTOs)
χL (κ, r).
One disadvantage of the MTOs defined above is their infinite range which makes
the practical calculations cumbersome. It has nevertheless been shown (Ander-
sen and Jepsen, 1984; Andersen et al., 1986) that these conventional MTOs can
be exactly transformed into a set tight-binding (TB) orbitals. These TB-MTOs,
basically formed by a linear combination of the conventional ones, are rather
localized and particularly suitable for first-principles electronic structure calcula-
tions.
1.3.3 The LMTO Band Calculation

Now we can consider solving the band structure problem with single-electron
crystal potential modelled by MT approximation, within which the potential is
formed by a array of MT wells centered at sites R of a three-dimensional periodic
lattice. In the spirit of the LCAO method, the Bloch function can be represented
as X
ψ k (E, r) = αLk (E)χkL (κ, r) (1.36)
L
where the coefficients αLk (E) are to be determined and χkL (κ, r) is the Bloch sum
of the energy independent MTOs
1 X ik·R
χkL (κ, r) = √ e χL (κ, r − R)
L R
!
1 X
= √ χL (κ, r) + eik·R χL (κ, r − R) . (1.37)
L R6=0
Here L is the number of the lattice sites or unitary cells. The last term in (1.37)
consists of the contributions from all the MT spheres except the one at the origin.
In the region that is inside the sphere centered at the origin and passing through
the nearest-neighbor sites but outside the neighboring MT spheres, this term can
be written as a one-center expansion,
X X
eik·R χL (κ, r − R) = eik·R κNl (κ, r − R) (1.38a)
R6=0 R6=0
X
= JL0 (κ, r)BLk0 L (κ) (1.38b)
L0
where the KKR structure constants BLk0 L (κ), according to the expansion theorem
(1.32), should be defined as
X X
BLk0 L (κ) = 4π CLL0 L00 eik·R κn∗L00 (κ, R). (1.39)
L00 R6=0
The above stated region of convergence is the intersection of the two regions
where (1.38a) and (1.38b) are valid respectively. Therefore, inside this region,
the Bloch sum of MTOs can be expressed in terms of a one-center expansion,
!
1 X
χkL (κ, r) = √ χL (κ, r) + JL0 (κ, r)BLk0 L (κ) . (1.40)
L L0
With the set of Bloch summed MTOs χkL (κ, r), by applying standard variational
techniques, the band structure problem is reduced to a set of linear equations at
each k point, X
hχkL |H − E|χkL0 i αLk 0 (E) = 0, (1.41)
L0
which has solutions in case that
det {hχkL|H − E|χkL0 i} = 0. (1.42)
Thus we need to evaluate the secular matrix element hχkL (κ, r)|H − E|χkL0 (κ, r)i.
Due to the translational properties of χkL0 (κ, r) and H, one can verify that
hχkL |H − E|χkL0 i = NhχkL |H − E|χkL0 i0 (1.43)
where hi0 means the integral over the atomic polyhedron at the origin. Within
the atomic polyhedron, χkL (κ, r) can be expanded as (1.40), and therefore
NhχkL |H − E|χkL0 i0 = hχL |H − E|χL0 i0

X
+ hχL |H − E|JL00 i0 BLk00 L0 + BLL
k
00 hJL00 |H − E|χL0 i0
L00
XX
+ k
BLL k
00 hJL00 |H − E|JL000 i0 BL000 L0 . (1.44)
L00 L000
For the spherically symmetric potential, the angular part of the wave functions
can be first integrated out, and we are finally left with
hχkL |H − E|χkL0 i = hχl |H − E|χl i0 δLL0

k
+ [hχl |H − E|Jl i0 + hχl0 |H − E|Jl0 i0 ] BLL0
X
+ k k
BLL00 hJl00 |H − E|Jl00 i0 BL00 L0 (1.45)
L00
The simplification from (1.44) to (1.45) arises from the fact the secular matrix
element between two χL (κ, r) or JL (κ, r) with two different L indices vanishes.
The matrix elements appearing on the righthand side of (1.45), is defined as
integrals over radial variable r, e.g.,
Z
d2 l(l + 1) 2
hχl |H − E|χl i0 ≡ dr rχl (κ, r) − 2 + + vMT (r) − κ rχl (κ, r).
0 dr r2
(1.46)
Within the LMTO method, the integral terms on the righthand side of χL (κ, r)
can be parameterized and evaluated at different orders of approximations. The
detailed way of representing these integrals by a set of parameters can be found
in the book of Skriver (Skriver, 1984). Concerning the approximations made to
accomplish this, a simple and popularly used one is the so-called atomic sphere
approximation (ASA), in which the κ2 is set to 0 and the atomic polyhedra are
replaced by the atomic spheres.
The procedure of constructing LMTOs described above is for the simple case
when there is only one atom in a unit cell, but it could be easily extended to
the multiatomic case. In that case, the LMTOs should carry one more index r
distinguishing the different atoms in the unit cell, namely,
χkL → χkrL = χkrlm . (1.47)
Thus in general, the LMTOs have three indices r, l, m to label the atom, the
angular and magnetic quantum number.
1.4. Wannier Functions 19
1.4 Wannier Functions
The electronic states in the system with a periodic potential are naturally repre-
sented by Bloch functions, labelled by the band index n and reciprocal vector k.
An equivalent representation is provided by Wannier functions (WFs) (Wannier,
1937), defined as a Fourier transformation of Bloch functions. These WFs are
hence are labelled by the spatial lattice R and band index n. The existence of a
localized set of WFs, and their general properties have been discussed by various
authors over the years, e.g., Koster (1953), Parzen (1953), Kohn (1959; 1973),
Des Cloizeaux (1963; 1964; 1964), and Blount (1962). Although the concept of
WFs has been employed in the construction of model Hamiltonians and in many
theoretical discussions, the quantitative calculations based on WFs did not ap-
pear until recently (Marzari and Vanderbilt, 1997; Ku et al., 2002; Pavarini et al.,
2004; Anisimov et al., 2005). This state of affairs is partly due to the nonunique
nature of WFs so that there is no general and reliable method to calculate them,
and partly due to the reason that the Bloch description of the electronic states is
usually quite satisfactory. However, in narrow-band systems where the electrons
have strongly atomic natures, there is a great need for a suitable set of localized
orbitals to describe the system properly. In this case, we consider that the WFs
are not just a mathematically unitary transformation of Bloch functions, but
rather represent the real electronic structure of the system.
The different ways to calculate WFs can be roughly classified into two categories.
The first approach, assumed by Koster (1953), Parzen (1953), and Kohn (1973),
attempts to produce the WFs directly through a variational procedure, without
knowing the Bloch functions. The other approach, which are is often used and
will be used here, is to calculate the WFs from a set of Bloch states that has been
already obtained from a energy band calculation. In a single-band case, the WFs
are defined as the Fourier transformation of the Bloch functions,
1 X ik·R k
W (r − R) = √ e ψ (r) (1.48)
L k∈BZ
where the summation is over the first Brillouin zone (BZ) and N is the number
of discrete k points inside this zone. However, even in the simple transformation
above, ambiguity concerning the definition of WFs arises from the indeterminacy
of the phase factor eiφ(k) associated with the Bloch function ψ k (r). On the other
hand, the freedom of choosing φ(k) can be utilized to obtain a set of well-behaved
WFs. For instance, for a one-dimensional lattice with reflection symmetry, one
can obtain a set of real, symmetric and exponentially localized WFs by choosing
ψ k (0) to be real (Kohn, 1959).
The practically more interesting case is that a composite group of bands are in-
terconnected among themselves by degeneracies. (See Fig. 1.2). In this case, one
composite bands
isolated band
Γ X
Figure 1.2: A schematic picture of isolated band, and a composite group of bands.
usually does not perform the transformation (1.48) individually for every branch
of the bands, but rather construct WFs for the composite bands simultaneously
by introducing an additional unitary transformation Umm k
0 among the different
branches at each k point, namely,
1 X ik·R X k
Wm (r − R) = √ e k
Umm0 ψm 0 (r). (1.49)
L k∈BZ m0
Two considerations are involved here: firstly it may be possible that no exponen-
tially localized WF can be obtained at all by including only a single branch, and
secondly WFs with better localization and higher symmetry can be constructed
by treating the composite bands all together. Thus the task of constructing WFs
consist in the determination of Umm
k
0 according to some criterions chosen a priori.
Different methods have been developed in the past. In particular, Marzari and
Vanderbilt (1997) devised a procedure to obtain the maximally P localized WFs
by minimizing a functional that representing the total spread m hr 2 im − hri2m
of these WFs. This minimization procedure starts with some initial guess of the
WFs obtained by projecting the trial localized orbitals onto the chosen set of
composite Bloch bands and it was found that this initial guess is usually quite
good. Ku et al. (2002) then discarded the minimization procedure and took
only the first step of Marzari and Vanderbilt’s method to construct their WFs,
by projecting the Gaussian orbitals onto the DFT all-electron eigenstates. Even
with this simplified procedure they found remarkably good results concerning the
relevant material-specific parameters. Pavarini et al. (2004) built up WFs for t2g
orbitals in some typical 3d1 perovskites by symmetrically orthonormalizing the N-
th order muffin-tin orbitals (NMTOs) (Andersen and Saha-Dasgupta, 2000), and
employed them in the LDA+DMFT investigation for the Mott transition and the
suppression of orbital fluctuations in these systems. Most recently, Anisimov et
al. (2005) proposed a scheme to calculate WFs by projecting the LMTOs onto the
1.4. Wannier Functions 21
chosen set of Bloch bands. This scheme is particularly suitable for LDA+DMFT
calculations, and will be discussed in detail in chapter 3.
23
2. THE LDA+DMFT APPROACH
In the first chapter we discussed the density functional theory and its local den-
sity approximation (DFT-LDA) that have been extensively used in the modern
band-structure calculation for solid crystalline. We also presented the main idea
and formulations of the LMTO method, which is one of the most popular method
for performing the DFT-LDA band calculation. However, the DFT-LDA, and in
general the single-particle description of solid systems generally fail for the sys-
tems with narrow bands and strong electron interactions. For these systems,
one has to invoke another approach, namely the model Hamiltonian approach,
to explicitly take into account the electron-electron interactions. In this context
an appropriate model should on the one hand be able to capture the strongly
correlated nature of these systems, but on the other hand be simple enough to
allow for an analytical or numerical solution, possibly under some reliable ap-
proximation. The one-band Hubbard model is such a “minimal” model aiming at
describing in a simplified way the correlated d electrons in transition metals and
their compounds. The model Hamiltonians, in spite of their apparent relevance
to some basic features of the strongly correlated materials and usefulness in re-
vealing the underlying physical mechanisms, are restricted in their ability to make
quantitative predictions. The LDA+DMFT approach is the first attempt to give
a quantitative description of the materials with strongly correlated electrons by
combining the DFT-LDA band structure calculation and the dynamical mean-
field theory (DMFT) for solving the many-body Hamiltonian. This combination
is based on the fact that the DFT-LDA calculation is a first-principles method
and usually gets the material-specific information quite correctly, and DMFT is
most powerful many-body technique for treating the strongly correlated lattice
fermionic model. In this chapter we first introduce the Hubbard model, then give
a brief review of the dynamical mean-field theroy and the quantum Monte-Carlo
(QMC) method for solving the DMFT equations, and finally close the chapter
with the LDA+DMFT formulations.
2.1 The Hubbard Model
The model referred to as Hubbard model became a standard framework for

studying the Mott transition and ferromagnetic metal since it was introduced
24 2. The LDA+DMFT Approach
(Gutzwiller, 1963; Hubbard, 1963; Hubbard, 1964). The model Hamiltonian

consists of two terms
X X
H= tij (d†iσ djσ + h.c.) + U ni↑ ni↓ , (2.1)
ij,σ i
with the first term describing the electron hopping from site i to site j, and the
second an intraatomic Coulomb repulsion. Here the operator diσ (d†iσ ) annihilate
(creates) an electron at site i with spin σ, niσ = d†iσ diσ , and U is the strength
of the Coulomb interaction. Usually the hopping of the electrons is restricted to
the neighboring site, and is translationally invariant, namely,

−t, (t > 0) i, j are neighboring sites,
tij = (2.2)
0 otherwise.
Apart from the physical parameters t and U, a few other parameters can affect
the the feature of the model, and these are the dimension D of the lattice on
which the model is defined, the band filling δ = Ne /(2L) (N being the total
electron number, and L the number of the lattice site ), and the temperature.
However, it is extremely difficult to get any exact result of (2.1), except in the case
of D = 1 where exact solution was worked out by Lieb and Wu (1968) using the
Bethe ansatz, and these authors showed that there is no metal-insulator transition
for any U > 0 and δ. The fact that exact solutions are not possible in D = 2
and D = 3 cases makes the employment of various approximations unavoidable.
To be confident that some particular approximation at hand is a meaningful one,
it is important to explore the different limiting situations in which the original
model reduces to a simpler one and some reliable results can be obtained. In some
particular limiting regime, a good approximation should become exact or at least
capture some physical ingredients. The ration U/t is a natural quantity according
to which one can have two opposite limiting regimes: the strong coupling regime
(defined as U/t 0) and weak coupling regime ( U/t 0). Another important
quantity is the dimension D. D = 1 is the case where exact solution is available
and thus provides an important test for the validity of new approximations one
wants to employ. Amazingly, in the opposite limit D = ∞ drastic simplification
of the Hubbard model (2.1) (and in general fermionic lattice model) occurs under
a proper scaling of the hopping term (Metzner and Vollhardt, 1989b), and a
nontrivial D = ∞ Hubbard model is formed. It is on this particular limit that
the dynamical mean-field theory (DMFT) (Georges and Kotliar, 1992; Jarrell,
1992) is based.
While leaving the discusson of N = ∞ limit to the next section, let’s first have
a look at the strong and weak coupling regimes. First of all, in the limit of
U/t 0 and at half-filling (η = 1/2), the Hubbard model (2.1) is reduced to an
antiferromagnetic Heisenberg model up to the second order in t/U (Anderson,
2.1. The Hubbard Model 25
1959), X
H=J Si · Sj , (2.3)
i,j
where the exchange constant J = t2 /U, and the the spin Si at site i is defined
through
Siz = ni↑ − ni↓ ,

Si↑ = d†i↑ di↓ , (2.4)
Si↓ = d†i↓ di↑ .
This implies that the ground state is insulating. However, the magnetic property
of the ground state of (2.3) is not exactly known except D = 1. While the half-
filled Hubbard model is reduced to the antiferromagnetic Heisenberg model for
very large U, how about the case that the band filling is slightly less than one
half? In that case the additional holes can hop between the lattice sites without
costing extra energy, and this requires adding a hopping term to the Heisenberg
Hamiltonian (2.3), leading to the so-called t-J model. Nagaoka (1966) rigorously
proved that the ground state of the system with a single whole and infinite U
(i.e., U Ne t) is ferromagnetic. This sheds some lights on the understanding
the itinerant ferromagnetism within the Hubbard model, but the condition for
Nagaoka’s theorem is rather unphysical, and the efforts of extending it to more
realistic cases failed.
Apart from the above simplifications, two analytical approaches have proved to
be useful for the strong-coupling regime. One is the Gutzwiller variational ap-
proach (Gutzwiller, 1963), through which Brinkman and Rice (1970) was able to
determine a criterion for the metal-insulator transition for the half-filling case,
and in particular the metallic phase was found to be a Fermi liquid. In addition,
by using Gutzwiller approximation, the Gutzwiller variational wave function pro-
vides a framework for interpolating the strong and weak coupling regimes (for a
review, see Vollhardt (1984) and Vollhardt et al. (1987)). The other approach
is known as slave boson mean-field theory (Barnes, 1976; Barnes, 1977), which
in many aspects lead to the same results as the Gutzwiller variational approach,
but can be used in a more general context (Kotliar and Ruckenstein, 1986).
The weak coupling regime was also addressed by several authors. However, even
in this limit, a proper treatment of the Hubbard model turned out to be highly
nontrivial, especially for the bipartite lattice where the antiferromagnetic correla-
tion always sets in at half-filling due to the “perfect nesting” of the Fermi surface.
Nevertheless, Metzner and Vollhardt (1989a) showed that the exact second-order
contribution to the ground-state energy can be obtained by standard perturba-
tion theory. A general approach to strongly correlated electrons by employing
“conserving approximation” (i.e., consistent with microscopic conservation laws)
beyond the mean-field level was developed by Bickers et al (1989) and applied to
2D Hubbard model. But in general, these approaches require a large amount of
computation efforts and are cumbersome to be performed. Again, a preferable
route one may take here is to start from the D = ∞ limit where the computation
effort is tremendously reduced, and reach the finite dimension by perturbative
techniques.
Besides the analytical studies of the Hubbard model within various approxima-
tions, numerical investigations have been performed on finite systems. Among
these the exact diagonalization (ED) and quantum Monte Carlo (QMC) are the
two major tools. ED provides exact results and is only doable for rather small
systems due to the exponential growth of the configuration space with the num-
ber of sites. QMC, on the other hand, can be used to study relatively larger
systems, but suffers from the sign problem for large values of U and numerical
instability at low temperatures. QMC technique can be either used directly as
an ab initio approach or within a variational framework (Yokoyama and Shiba,
1987). ED and QMC can often be employed in a complementary way in which
the result of ED offers a check for the efficiency of QMC.
The one-band Hubbard model is the minimal model one can figure out to describe
correlated d-electron systems. But one can ask the question: how well does
this model represent the physics of the correlated electrons? Obviously the one-
band model has neglected the multi-orbital effects and the possible hybridizations
between the d orbitals and s, p orbitals. Indeed important effects may be lost
through such a simplification. For a material-oriented study, one often needs to
generalize the model (2.1) so that the orbital degree of freedom can be taken into
account, giving a multi-orbital Hubbard-like model,
X 0 †
H= tmm
ij dimσ djm0 σ
i,j,m,m0 ,σ
X σσ0
Umm X Jmm0 †
dimσ d†im0 σ̄ dim0 σ dimσ̄ . (2.5)
0 0 0
+ nimσ nim0 σ0 −
2 2
i,m,m0 ,σ,σ0 0
i,m,m ,σ
mm0
Here the hopping parameters tij become a matrix between m orbital on site i
σσ0
and m0 orbital on site j. Umm 0 gives the strength of direct Coulomb interaction
between the spin-orbital channels {m, σ} and {m0 , σ 0 }, and the prime on the
summation excludes the self-interaction with m = m0 and σ = σ 0 . The Jmm0
term describes the “spin-flip” effect between two channels and the prime here
means m 6= m0 . In practice, the following relationships among the parameters
σσ0
Umm 0 and Jmm0 are assumed,
0
σσ
Umm 0 = U − 2J(1 − δmm0 ) − Jδσσ 0 and Jmm0 = J (2.6)
which hold exactly for cubic systems. It should be noted that, in a rigorous
derivation of (2.5), there should exist other terms reflecting the physical pre-
2.2. Dynamical Mean-Field Theory 27
cesses like pair hopping and density-dependent hopping, which are nevertheless
neglected here for simplicity. The above model (2.5) is frequently used in mate-
rial investigations in conjugation with the LDA band-structure calculations which
provide the model parameters. The one-band Hubbard model is already difficult
enough to solve, the task of solving the multi-band model (2.5) is much more
challenging. At present the most powerful tool for dealing with this model seems
to be DMFT, which we will discuss in the next section.
2.2 Dynamical Mean-Field Theory
The essential idea of DMFT is to replace the fermionic lattice model by an quan-
tum impurity model embedded in an effective medium which needs to be de-
termined self-consistently. By doing so the local dynamics is contained in the
impurity problem and the lattice effect is taken care of by the self-consistent
condition, in a rather similar philosophy of the Weiss mean-field theory for the
classical systems. The DMFT is a dynamical theory, however, in the sense that
the local quantum fluctuation is fully accounted for by the impurity model and
only the spatial correlations are treated in a mean-field way. Thus it is standing at
a higher level than static mean-field theores such as the Hartree-Fock approxima-
tion in which both the spatial and local quantum fluctuations are frozen. Similar
to the Weiss mean-field theory, DMFT becomes exact in the limit of infinite
dimension (or the infinite coordination number). Before presenting the DMFT
equations, it is appropriate to have a look at the D = ∞ limit for correlated
lattice fermions.
2.2.1 The Infinite Dimension Limit
For a number of classical problems (e.g., the Ising model and the spin glasses),
many insights into the system can be gained by taking the limit D → ∞. For
the classical spin models, the D → ∞ leads to the Weiss molecular field theory.
In order to keep the total energy finite, the spin coupling constant J (see the
Heisenberg model (2.3) for example) has to be rescaled as J = J ∗ /Z where J ∗ is
a constant and is the coordination number (for a hypercubic lattice, Z = 2D). It
is interesting to see that this limit is also useful for strongly correlated fermionic
lattice model. In their original paper, Metzner and Vollhardt (1989b) pointed
out, in the D → ∞ limit, the diagrammatic treatment of the Hubbard model
simplifies substantially while the many-body nature is still preserved. Again, to
have a nontrivial model where both the kinetic and interaction energy are finite,
the hopping amplitude t (see (2.1) and (2.2)) has to be rescaled, but in a different
way comparing with the spin models, namely,

t∗
t = √ , t∗ = const. (2.7)
2D
The reason for such a choice of the scaling can be easily seen from the noninter-
acting density of states (DOS) of the Hubbard model (2.1) with nearest-neighbor
hopping on a supercubic lattice (Metzner and Vollhardt, 1989b),
" 2 #
1 ε
ND (ε) = √ exp − √ , D→∞ (2.8)
2t πD 2t D
Thus the scaling (2.7) immediately leads to the Gaussian behavior of the nonin-
teracting DOS at infinite dimension,

1 1 ε 2
N∞ (ε) = √ exp − . (2.9)
2π t∗ 2 t∗
Another important example is the Bethe lattice with nearest neighbor hopping.
For this lattice the scaling of (2.7) leads to the the semicircular noninteracting
DOS for D = ∞,
1 p ∗2
N∞ (ε) = ∗ 2 4t − ε2 , |ε| < 2t∗ . (2.10)
2πt
Moreover, due to (2.7), it is easy to see the non-interacting single-particle prop-
agator √
G0ijσ ∼ O(1/ D) (2.11)
for neighboring i, j, and for general i, j sites one obtains (van Dongen et al.,
1989; Metzner, 1989),
G0ijσ ∼ O 1/D||Ri−Rj ||/2 (2.12)
where ||Ri − Rj || is the distance between i and j under the so-called “New York
metric”. As a consequence of the property (2.12), it was shown that the off-site
contribution of the irreducible self-energy, i.e., Σij with i 6= j, is infinitely smaller
than its on-site counterpart Σii for D → ∞ (Metzner and Vollhardt, 1989b;
Müller-Hartmann, 1989b), and thus the full self-energy becomes a purely local
quantity,
Σij (ω) = Σii (ω)δij , for D → ∞. (2.13)
It follows that its Fourier transformation Σ(k, ω) becomes momentum-independent,
Σ(k, ω) = Σ(ω). (2.14)
Remarkable simplifications of the treatment of the Hubbard-like models arise from
(2.14) and this immediately stimulated a number of subsequent works mainly
focusing on the Gutzwiller variational wave function, and the weak-coupling ex-
pansion of the Hubbard model and related models (for a review, see Müller-
Hartmann (1989a), Vollhardt (1991; 1993)).
2.2.2 DMFT Equations — Exact Solution of the D = ∞

Hubbard Model
As observed by Müller-Hartmann (1989b), the irreducible self-energy Σii (ω),

which is purely local, depends only on the site-diagonal full Green’s function
Gii (ω) on the the same site. By using this fact, Brandt and Mielsch (1989; 1990;
1991) obtained the exact solution of the infinite dimensional Falikov-Kimball
model, which can be thought of as a simplified version of the Hubbard model by
permitting only one of the two spin species to hop. Brandt and Mielsch’s work
provided an illustrating guideline that the lattice problem can be understood by
just looking at a single site. Unfortunately, the D → ∞ Hubbard model does not
allow for an analytically exact solution. However, Georges and Kotliar (1992)
showed that its dynamics can be described by a single impurity with the effective
single-site action (see also Jarrell (1992)),
Z β Z β Z β X
Seff = U dτ n↑ (τ )n↓ (τ ) − dτ dτ 0 d†σ (τ )G0−1 (τ − τ 0 )dσ (τ 0 ). (2.15)
0 0 0 σ
The G0 (τ − τ 0 ) here, serving as the “bare” Green’s function for this local site,
describes the influences of the environment on the present site. It essentially
plays the same role as the Weiss field in the classic models, but now it is not just
a constant number but rather (imaginary) time-dependent, so as to account for
the quantum fluctuations on this local site. The reason that one is allowed to
reduce the original lattice problem to a single-site problem (2.15) is due to the
fact that the spatial fluctuations is completely suppressed at D = ∞.
The full Green’s function and the self-energy (represented in the domain of Mat-
subara frequency) of the impurity problem can be calculated from (2.15),
imp (iωn ) (2.16)

Gimp(iωn ) = hd† (iωn )d(iωn )iSeff , and Σimp (iωn ) = G0−1 (iωn ) − G−1
Of course the effective field G0 (τ ) is not known a priori, but it has to be such
that the interacting Green’s function of impurity problem and the site-diagonal
Green’s function of the original lattice problem are identical, and so are the self-
energies, namely,
Gimp (iωn ) = Gii (iωn ), Σimp (iωn ) = Σ(iωn ). (2.17)
On the other hand, the on-site lattice Green’s function is

Z ∞
1X 1 N∞ (ε)
Gii (iωn ) = = (2.18)
L k iωn + µ − εk − Σ(iωn ) −∞ iωn + µ − ε − Σ(iωn )
P
where ε(k) = √1L j tij eik·(Ri −Rj ) is the non-interacting single-particle energy.
Eq. (2.18) is known as the k-integrated Dyson equation, and here the lattice
property enters the equation only through the noninteracting density of states
N(ε)1 . The close set of equations from (2.15) to (2.18) provide a self-consistent
way to determine the local Green’s function G(iωn ) = Gii (iωn ) as well as the self-
energy Σ(iωn ). These equations are known as DMFT equations and consist in a
mean-field theory of the Hubbard that becomes exact for D → ∞. The single-site
representation of the lattice problem can be derived in a mathematically rigorous
manner. The derivation can be carried out in several different ways, among
which there are the “cavity” method, the expansion around the atomic limit, and
effective medium interpretation (For a comprehensive review, see Georges et al
(1996)).
Now we can focus on the single-impurity problem (2.15) which is of course still a
highly nontrivial problem. Practically, one can consider the action (2.15) as aris-
ing from a single impurity coupled to a bath of “conduction electrons”, described
by the following Hamiltonian,
X X X
Himp = ˜k a†k,σ ak,σ + εd d†σ dσ + Und↑ nd↓ + (Vk a†k,σ dσ + h.c.) (2.19)
k,σ σ k
where ndσ = d†σ dσ and ˜k is the single-particle energy of the auxiliary bath electrons
(represented by operators a† , a) and should not be confused with the noninter-
acting single-particle energy ε(k) of the original lattice problem in (2.18). The
Hamiltonian (2.19) is known as single impurity Anderson model (SIAM) (Ander-
son, 1961), on which substantial experience has been gained during thirty years
of studies. After integrating out the degree of freedom of the bath electrons, one
ends up with the action (2.15) for the impurity electron with the effective field,
Z +∞
∆(ε)
−1
G (iωn ) = iωn − d − dε , (2.20)
−∞ iωn − ε
P
in which ∆(ε) = k |Vk |2 δ(ε − ˜k ) is the so-called hybridization function and the
representation(2.20) is general enough to produce any G −1 . It is worthwhile to
point out the interpretation of the single-site action (2.15) in terms of the SIAM
is not the unique way, and alternative interpretation in terms of the Wolff model
(Wolff, 1961) also exists (Georges et al., 1992; Georges et al., 1996).
Thus, as one can see, the problem of solving the D = ∞ Hubbard model is re-
duced to solving the SIAM iteratively. However, an exact solution of SIAM only
exists for a constant hybridization function ∆(ε) = ∆ using the Bethe ansatz.
Therefore, for a general ∆(ε) appearing in a self-consistent procedure, one has
to invoke proper approximations or numerical techniques to get a solution of
(2.19). The different approaches to treat the single impurity problem correspond
to the different methods of solving the DMFT equations. Of these there are
1
This is not true, however, for the nondegenerate multi-orbital cases, see Section (2.3).
numerical techniques such as QMC, ED, and Wilson numerical renormalization

group (NRG), and analytical approaches such as iterative perturbation theory
(IPT) and noncrossing approximation (NCA). QMC studies are based on the
Hirsch-Fye algorithm (Hirsch and Fye, 1986) and were applied in the DMFT
problems independently by Jarrell (1992), Rozenberg, Zhang, and Kotliar (1992)
and Georges and Krauth (1992). A detailed discussion of QMC will be given
in the next section. ED investigations were carried out by Caffarel and Krauth
(1994) and Si et al (1994). The application of NRG in the present context were
first performed by Sakai and Kuramoto (Sakai and Kuramoto, 1994), and later
by Bulla, Hewson and Pruschke (1998) and Bulla (1999). The IPT scheme was
first used in the original work of Georges and Kotliar (1992), and then gener-
alized by Kajueter and Kotliar (1996). NCA was first employed by Jarrell and
Pruschke (1993a; 1993b), and by Pruschke, Cox, and Jarrell (1993a; 1993b).
It should be pointed out that the above equations (2.15) to (2.18) are valid
specially for paramagnetic phase of the Hubbard model. The scheme can be
easily extended to phases with long-range magnetic orders and to other strongly
correlated fermionic lattice models like the periodic Anderson model and the
Kondo lattice model (Georges et al., 1992).
In the next section, we will present the main formalism for the Hirsch-Fye QMC
algorithm which is employed in this thesis for solving the impurity problem.
2.2.3 Quantum Monte-Carlo Method
As mentioned above, the Hirsch-Fye algorithm originally devised for the SIAM
can be straightforwardly employed as the impurity solver for the DMFT prob-
lem. In one-band case this algorithm has empirically proven to be absent of the
sign problem and and not suffering the numerical instability at low temperatures.
More importantly, for the studies of multi-band models and the material-specific
calculations, the QMC is practically the only numerical tool so far to deal with
the DMFT equations. Therefore QMC plays an indispensable role among the
different impurity solvers. Within the QMC method, what one can obtain is the
imaginary-time Green’s function G(τ ). To get any physically interested quanti-
ties, one has to first continue G(τ ) to get the real-time Green’s function, which
is usually accomplished by the maximum entropy method (MEM) (Jarrell and
Gubernatis, 1996).
In the following we will give a discussion of the Hirsch-Fye algorithm, following

the review of Georges et al (1996). To begin with, we rewrite the SIAM (2.19)
as,
Himp = H0 + H1 , (2.21a)
X U X X
H0 = ˜k a†kσ akσ + (εd + ) d†σ dσ + + (Vk a†kσ dσ + h.c.), (2.21b)
k,σ
2 σ k
U d
H1 = Und↑ nd↓ − (n + nd↓ ). (2.21c)
2 ↑
Now we consider the the partition function of the SIAM,
Λ
Y Λ
Y
Z = TreβHimp = Tr e∆τ (H0 +H1 ) ≈ Tr e∆τ H0 e∆τ H1 . (2.22)
i=1 i=1
Here the imaginary time interval [0, β] has been equivalently discretized into Λ
slices and ∆τ = β/Λ. The Trotter breakup is used for the last step and the error
involved in this breakup is ∼ O(∆τ 2 ). The interaction part of the Hamiltonian
can be decoupled via a Hubbard-Stratonovich transformation using the auxiliary
Ising variables (Hirsch, 1983),
1 X λs(nd↑ −nd ↓)
e∆τ H1 = e , (2.23)
2 s=±1
where λ = cosh−1 (exp(∆τ U/2)). From (2.22) and (2.23) we have
1 X ∆τ
Z ≈ Z{s} , (2.24a)
2Λ
{s}=±1
Y σ σ σ
∆τ
Z{s} = Tre∆τ K eV (s1 ) × e∆τ K eV (s2 ) · · · e∆τ K eV (sΛ ) . (2.24b)
σ=±1
Here, {s} denotes the set of Ising variables (s1 , s2 , · · · , sΛ ), one variable corre-
sponding to one time slice. In addition,
 
d + U/2 Vk∗1 Vk∗2 ···
 V k1 ˜k1 0 ··· 
e∆τ K = 
 (2.24c)
V k2 0 ˜k2 ··· 
··· ··· ··· ···
stems from the noninteracting part of the Hamiltonian H0 in (2.21b), and

 σs λ 
e l 0 0 ···
 0 1 0 ··· 
eV (sl ) =  
σ
 0 (2.24d)
0 1 ··· 
··· ··· ··· ···
is associated with the decoupled on-site interaction (2.23). At this point one
should distinguish between the physical spin σ and the auxiliary Ising spins which
can be 1 or −1 at every time slice.
A crucial observation is that the Ising-spin dependent partition function Z{s}
∆τ
can
be expressed as
↑ ↓
∆τ
Z{s} = detO{s} · detO{s} , (2.25)
where
 σ (s 
1 0 ··· 0 e∆τ K eV Λ)
 −e∆τ K V σ (s1 )
e 1 ··· ··· 0 
 
σ
O{s} =
 0 −e∆τ K V σ (s2 )
e 1 ··· ··· .

 ··· ··· ··· 1 0 
∆τ K V σ (sΛ−1 )
0 0 · · · −e e 1
(2.26)
(For a proof, see Hirsch (1985) or Blankenbecler, Scalapino and Sugar (1981)).
The element of the matrix O{s}
σ
is labelled by the combined index {l, p} in which l
corresponds to the time slice and p to electron orbital respectively. The Ising-spin
dependent Green’s function g{s}σ
, defined as
1 σ σ
σ
g{s} = Tre∆τ K eV (s1 ) · · · e∆τ K eV (sl−1 ) apσ (τl ) · · ·
lp;l0 p0 detO{s}
σ
σ (s σ (s
e∆τ K eV l0 −1 ) a†p0 σ (τl0 ) · · · e∆τ K eV Λ)
, (2.27)
is related to the matrix O{s}
σ
by
−1
σ
g{s} σ
= O{s} . (2.28)
In (2.27) we have assumed the correspondence that that a1σ = dσ , a2σ = ak1 σ ,
a3σ = ak2 σ , . . . , and so on. The essential ingredient of the Hirsch-Fye algorithm
is based on the fact, as first observed by Hirsch and Fye (1986), that Green’s
functions for two different Ising-spin configurations {s} and {s0 } are connected
by a Dyson equation
0
g 0 = g + (g − 1)(eV −V − 1)g 0. (2.29)
In (2.29) we have abbreviated g ≡ g{s} σ
, g 0 ≡ g{s
σ
0 } , and the e
V
here should be
V σ (sl ) 0 σ
understood as a diagonal matrix with e ll0 = e
V
δll (the form of eV (sl ) is
0
given in (2.24d)). It is not difficult to see the matrix eV −V − 1 in (2.29) has
the following behavior,
h 0 i 0
e(V −V ) − 1 = eλσ(sl −sl ) δll0 δp1 δp0 1 , (2.30)
lp;l0 p0
i.e., it is nonzero only at the impurity site. Therefore, the Dyson

equation
(2.29)
also holds for the Green’s function of the impurity site G{s} ≡ g{s}
σ σ
,
11
0 −V
G0 = G + (G − 1)(eV − 1)G0 , (2.31)
0
but now eV −V −1 should be understood as an Λ×Λ diagonal matrix with element
0
eλσ(sl −sl ) − 1. Rearranging Eq. (2.31) one can get
h i−1
0 V 0 −V
G = 1 + (1 − G)(e − 1) G (2.32)
Eq. (2.32) provides a way to generate the Green’s function G{s0 } for some Ising
spin configuration {s0 } = (s01 , s02 , · · · , s0Λ ) from the known Green’s function G{s}
for another spin configuration {s} = (s1 , s2 , · · · , sΛ ). For two general {s} and
{s0 }, this involves an inversion of a Λ × Λ matrix. However, in the special case
that only a single spin, say sl , is flipped, one can verify that (2.32) is reduced to
−1
0 λσ(s0l −sl ) λσ(s0l −sl )
Gl1 l2 = Gl1 ,l2 + (Gl1 ,l − δl1 l ) e − 1 1 + (1 − Gll )(e − 1) Gll2 ,
(2.33)
where l denotes some arbitrary but fixed time slice. Moreover, in this case, it can
be shown that the following relation holds,
detO{s}
σ
detg{s
σ
0} detGσ{s0 } 0
= = = 1 + (1 − Gσ{s} ll )(eλσ(sl −sl ) − 1). (2.34)
detO{s
σ
0} detg{s}
σ
detGσ{s}
Now we can consider the physical Green’s function for the d site which is given
by h i
P ↑ ↓
{s} detO {s} detO {s} G σ
{s}
l1 l2
Gσl1 l2 = P ↑ ↓
. (2.35)
{s} detO {s} detO {s}
Here the exact calculation of the physical Gσ involves a summation over 2Λ dif-
ferent spin configurations. This is computationally impossible for a large λ. The
QMC method, however, provides an efficient way to evaluate (2.35) by impor-
tance sampling according to the probability distribution
↑ ↓
detO{s} detO{s}
P ({s}) = P ↑ ↓
. (2.36)
{s0 } detO{s 0 } detO{s0 }
The sampling over the configuration space can be performed according to different
rules, and the Metropolis sampling is the most frequently used one. In which
P ({s} → {s0 }), the probability of accepting the trial move for {s} to {s0 }, is
given by
(
1 if P ({s0}) > P ({s}),
P ({s} → {s0 }) = (2.37)
P ({s0})/P ({s}) otherwise.
The Hirsch-Fye algorithm can be straightforwardly used to solve the effective

impurity model appearing in the DMFT context. For the effective impurity
model, what is known is the Weiss field G σ (τ ) which is determined in the previous
iteration. G σ (τ ), on the other hand, can be regarded as the spin-dependent
Green’s function Gσ{s} with {s} = 0. The initial Ising-spin dependent Green’s
function Gσ{s} for an arbitrary Ising-spin configuration (s1 = ±1, s2 = ±1 . . . SΛ =
±1) can be obtained from Gσ{s=0} = G σ by explicitly using (2.32) (which is valid
for arbitrary {s0 } and {s}). From then on, the Ising-spin configuration space is
visited by once flipping only one spin, and whether the flip is accepted or not
is determined by the Metropolis algorithm (2.37). Once the flip is accepted, the
Green’s function is updated through (2.33). The physical Green’s function Gσ
is calculated by averaging the Ising-spin dependent Green’s functions Gσ{s} . The
process is continued until a desired accuracy is reached.
2.2.4 Case Studies with DMFT

To have an idea that what one can get within DMFT, we present two examples
of the DMFT solutions of the Hubbard model.
(a) One-band Hubbard model away from half-filling
In the first example, we consider the single-band Hubbard model (2.1) defined on
a Bethe lattice with infinite connectivity. For U = 0, the DOS has a semicircular
behavior (2.10), and the integrated Dyson equation (2.18) reduces to
p
G(iωn ) = (ζ − sgn [Im(ζ)]) ζ 2 − 4t∗ 2 /2t∗2 , (2.38)
with ζ = iωn + µ − Σ(iωn ). In this calculation we fix U = 10, and t∗ = 0.5 giving
half-width of the semicircular DOS W = 1. The effective impurity problem
is solved by the QMC method at a temperature T = 1160K corresponding to
β = 10. The QMC result is treated by MEM to get the spectral functions. In
Fig. 2.1 the spectral functions as a function of the electron density are shown.
At half-filling n = 1, it is quite clear that a Mott insulator is formed with the
spectrum equally split into lower Hubbard band (LHB) and upper Hubbard band
(UHB), and the Fermi energy is sitting in the middle of the gap. When the
electron density is away from the half-filling, a quasiparticle resonance takes place
at the Fermi level. There is still a Mott gap formed separating the LHB and UHB,
but not at the Fermi level. The UHB is shifted to a higher energy and becoming
more dispersive, and its spectrum weight is becoming smaller as n decrease. The
studies of the one-band Hubbard model away from half-filling within DMFT were
carried out by Pruschke, Cox, and Jarrell (1993b; 1993a) using the QMC method,
and Kajueter and Kotliar (1996) using the ITP method respectively.
(b) Two-band Hubbard model with hybridization
In the second example we consider a two-band Hubbard model, with one inter-
acting band and one noninteracting band, and the two bands hybridizing with
0.8 n=0.5
0.4
0
0.8
0.4 n=0.7
0
0.8 n=0.9
0.4
0
0.8 n=0.95
0.4
0
DOS
0.8 n=1.0
0.4
0
-8 -4 0 4 8 12 16 20
Energy (eV)
Figure 2.1: The result of the one-band Hubbard model solved by DMFT. The
on-site Coulomb interaction U = 10 and half of the band width W = 1. Fermi
energy is set to zero. From top to bottom, electron numberper site are 0.5, 0.7, 0.9,
0.95, and 1.0 respectively. The spectrum weights of the UHB are approximately
measured as 0.438, 0.658, 0.878, 0.943, and 1.0 respectively
each other. The Hamiltonian reads
X X X X
H = − tp (p†iσ pi0 σ + h.c.) + p npiσ − td (d†jσ dj 0 σ + h.c.) + d ndjσ
hi,i0 i,σ i,σ hj,j 0 i,σ j,σ
X X
+ U nj↑ nj↓ + V (p†iσ djσ + h.c.). (2.39)
i hi,ji,σ
in which the electrons are classified into a p-type which is free and a d-type which
is subject to a on-site Coulomb interaction. These two types of electrons hybridize
with each other with a strength V . The model (2.39) is sometimes also addressed
as a generalized periodic Anderson model by allowing the originally localized
electrons to hop. This model at half-filling case (n = 2) has been studied within
DMFT by Shimizu, Sakai and Hewson (2000) using NRG method and recently
by de’ Medici et al (2005) using ED and Gutzwiller approximation. For the case
of n = 1, Ōno, Bulla and Hewson (2001) has obtained a phase diagram within
the approximation of linearized DMFT.
In this study we put the p- and d-orbitals on different sublattices of the bipartite
Bethe lattice. When the connectivity of the Bethe lattice q → ∞, the self-
consistency equations for the model (2.39) are given by (see e.g., Georges, Kotliar,
and Krauth (1993), or Ōno, Bulla, Hewson (2001)),

Gp (iωn )−1 = iωn + µ − p − t2p Gp (iωn ) − V 2 Gd (iωn ), (2.40a)
Gd (iωn ) −1 2
= iωn + µ − d − V Gp (iωn ) − t2d Gd (iωn ). (2.40b)
Within DMFT, these equations should be solved together with the effective im-
purity problem for the d electrons. Again, the impurity problem is solved by the
Hirsch-Fye QMC method at β = 10 and the obtained imaginary Green’s function
is treated by MEM to get the energy spectrum. At this point we want to point
out the model under investigation and the arrangement of the p, d orbitals on
the lattice is intended to mimic materials with strong correlations among the
transition metal valence d electrons as well as strong hybridizations between the
d electrons and the ligand p electrons. The obtained spectra for a fixed set of td ,
tp , d , and p parameters and different U, V parameters are shown in Fig. 2.2.
0.8 0.8
0.6 U=0, V=0 0.6 U=8, V=0
0.4 0.4
0.2 0.2
0 0
-10 -8 -6 -4 -2 0 2 4 -10 -8 -6 -4 -2 0 2 4
0.8 0.8
0.6 U=0, V=1 0.6 U=8, V=1
0.4 0.4
0.2 0.2
0 0
-10 -8 -6 -4 -2 0 2 4 -10 -8 -6 -4 -2 0 2 4
0.8 0.8
0.6 U=5, V=0 0.6 U=10, V=0
0.4 0.4
0.2 0.2
0 0
-10 -8 -6 -4 -2 0 2 4 -10 -8 -6 -4 -2 0 2 4
0.8 0.8
0.6 U=5, V=1 0.6 U=10, V=1
DOS
DOS
0.4 0.4
0.2 0.2
0 0
-10 -8 -6 -4 -2 0 2 4 -10 -8 -6 -4 -2 0 2 4
Energy (eV) Energy (eV)
Figure 2.2: The energy spectra of d band (solid line) and p band (dashed line)
obtained from the DMFT solution of the two-band Hubbard model. The electron
occupation n = 3, p − d = 3, and the Fermi energy is fixed at 0. The hopping
parameters are set as tp = td = 0.5 so that the half-width of the noninteracting
DOS of both bands is 1. The values of U and V are varied as indicated in the
figure.
From Fig. 2.2, one can clearly see the effect of the on-site interaction U of the
d electrons and the hybridization between d and p electrons on the electronic
properties of the system. For V = 0 and a large U, the d band is split into LHB
and UHB by the on-site interaction U, with the fully occupied noninteracting
p states sitting in between. This is actually the classic picture for the idea of
“charge-transfer” insulator. When V becomes finite, a resonance which has a
mixed character of d, p states is formed on the top of the valence bands. These
results shed illustrating lights on the understanding of the Mott transition in the
materials with strong p-d hybridization.
2.3 The LDA+DMFT Formulation
In the chapter 1 we discussed DFT and its LDA that provide a successful first-
principles description of weakly correlated materials. However, usually the LDA
fails to describe the strongly correlated materials which are characterized by
open d or f shells. These systems exhibit essentially many-body features in their
ground-state and excitation properties which can not be captured by the one-
electron band theory. The many-body effects such as the Mott metal-insulator
transition, Kondo effect and heavy fermion behavior, are usually understood in
the framework of simplified model Hamiltonians, among which the Hubbard
model and Anderson model are famous examples. DMFT, as a powerful ap-
proach to solved these strongly correlated lattice models, has been discussed in
the previous section.
Is it possible to develop a first-principles theory for the strongly correlated ma-
terials as the DFT-LDA for the weakly correlated one? For this purpose one
may first think of a many-body extention of the LDA by supplementing the
one-electron LDA Hamiltonian by a Hubbard-like interaction term among the
correlated orbitals (Anisimov et al., 1997b; Held et al., 2003). By doing so we are
actually separating the electronic states into the weakly correlated ones for which
the LDA has already given a sufficient description, and the strongly correlated
ones which require a further treatment by solving a many-body problem. Based
on these considerations the extended full Hamiltonian reads
H = HLDA + Hint − H dc , (2.41a)

X
HLDA = δirlm,jr0 l0 m0 εrlm nσirlm + tirlm,jr0l0 m0 cσ† σ
irlm cjr 0 l0 m0 , (2.41b)
irlm,jr 0 l0 m0 ,σ
" #
X X σσ0
Umm 0 0
X Jmm0 †σ †σ̄
0
Hint = nσirlm nσirlm0 − 0
cirlm cirlm0 cσirlm0 cσ̄irlm .
2 2
i,r=rd ,l=ld m,m0 ,σ,σ0 0m,m ,σ
(2.41c)
The one-electron LDA Hamiltonian HLDA can be cast into a tight-binding (TB)
form as expressed in (2.41b), by choosing the TB-LMTOs (Andersen and Jepsen (1984),
2.3. The LDA+DMFT Formulation 39
Andersen, Pawlowska, and Jepsen (1986)) as the basis set. cσjr0l0 m0 denotes an an-
nihilation operator for an electron in r-thatom of the j-th elementary unit cell
with an angular momentum number l and magnetic quantum number m. On top
of HLDA , one adds explicitly a many-body interaction term Hint for the strongly
correlated electrons, as expressed in (2.41c), in exact analogy to the multi-band
Hubbard model (2.5) (here we used rd and ld to denote the atom and angular
momentum channel that these electrons are associated with, usually they are
the valence d states of the transition metals). However, the Coulomb interaction
among these correlated electrons has been taken into account in an averaged way
in the LDA through the one-electron effective potential. Therefore a term H dc ,
is subtracted to avoid double-counting. Since there is no direct microscopic link
between the Hubbard model and LDA, the exact form of H dc in terms of U,
J parameters is not known. However, a commonly employed approximation for
H dc assumes that the LDA energy E dc for this term is given by (Anisimov et al.,
1997b; Held et al., 2003)
1 1 X
E dc = Ū nd (nd − 1) − J ndσ (nσσ̄ − 1). (2.42)
2 2 σ
with
U + (M − 1)(U − 2J) + (M − 1)(U − 3J)
Ū = . (2.43)
2M − 1
Here M is the number of correlated orbitals, and U, J parameters PM are related
σσ0
to Umm 0 and Jmm0 in (2.41c) through (2.6). Moreover ndσ = m=1 hnird ld mσ i
is the P
number of electrons occupying these correlated orbitals with spin σ, and
nd = σ ndσ the total number of electrons on these orbitals. The correction
(2.42) for the LDA energy of the interacting orbitals leads to a correction of
one-electron energy εirlm in (2.41b), namely,
∂E dc
ε0rlm = εrlm − = εrlm − δrl,rd ld Ū (nd − 1/2) + J/2(nd − 1) , (2.44)
∂nrlm
P
with nrlm = σ hnσirlm i. Thus we can rewrite Eqs. (2.41a) and (2.41b) in the
following way,
0
H = HLDA + Hint , (2.45a)
X
0
HLDA = δrlm,r0 l0 m0 ε0rlm nσirlm + tirlm,jr0l0 m0 cσ†
irlm cjr 0 l0 m0 . (2.45b)
σ
irlm,jr 0 l0 m0 ,σ
To have a concrete Hamiltonian, the U, J parameters in the interaction term

(2.41c) have to be specified. This can be accomplished by a procedure known as
“constrained LDA” (Gunnarsson et al., 1989). Eqs. (2.45a), (2.45b) together with
(2.41c) represent the ab-initio Hamiltonian for a particular material under inves-
tigation. For later use, it is more convenient to transform the TB Hamiltonian
(2.45b) to the reciprocal space,

1 X ik·(Ri −Rj )
0
HLDA (k) rlm,r 0 l0 m0
= e 0
HLDA irlm,jr 0 l0 m0
. (2.46)
L j
in which L is the number of the elementary unit cell in the periodic system (or
equivalently, the number of k points in the first Brillouin zone).
Such a many-body extention of LDA first appeared in the LDA+U approach
proposed by Anisimov et al (1991) where the interaction term (2.41c) is treated
within the Hartree-Fock approximation. Many interesting phenomena, such as
orbital and charge ordering in transition metal compounds, can be reproduced
by LDA+U. However, for the strongly correlated paramagnetic states, the under-
standing of which relying on the energy-dependence of the self-energy, LDA+U is
too simple to give an adequate description, and more sophisticated methods are
needed. The LDA+DMFT approach, first implemented by Anisimov et al (1997b)
[see also Lichtenstein and Katsnelson (1998)], is such a successful method in which
the Hubbard interaction term is treated by DMFT to preserve the many-body
nature of the system. The essential ingredient of the LDA+DMFT approach lies
in a generalization of the k-integrated Dyson equation (2.18) to incorporate the
material-specific information,

1X 1
G(iωn )rlm,r0l0 m0 = 0
. (2.47)
L k iωn + µ − HLDA (k) − Σ(iωn ) rlm,r0 l0 m0
Here in general an inversion of a matrix with a dimension as large as that of

the LDA Hamiltonian matrix is required at each k point. The self-energy ma-
trix Σ(iωn ) has non-zero elements only among the interacting orbitals, namely,
Σ(iωn )rlm,r0 l0 m0 = δrl,rd ld δr0 l0 ,rd ld Σmm0 (iωn ). The Weiss Green’s function for the
interacting orbitals is given by
−1
Gmm −1
0 (iωn ) = G(iωn )r l m,r l m0 + Σmm0 (iωn ).
d d d d
(2.48)
Simplification is possible when the energy bands of the interacting orbitals are well
separated from other bands, or in another word, the elements in the hybridization
block between the interacting orbitals and other ones in the Hamiltonian matrix
0
HLDA (k) are negligibly small. In this case Eq. (2.47) can be reduced to an
equation for M × M matrice,

1X 1
G(iωn )mm0 = 0
, (2.49)
L k iωn + µ − Hdd (k) − Σ(iωn ) mm0
in which Hdd
0
(k) denotes the interacting-orbital subblock of the full LDA Hamilto-
nian HLDA
0
(k). Further simplification occurs in the case that these interacting or-
2.3. The LDA+DMFT Formulation 41
bitals under consideration are degenerate2 . In this case G(iωn )mm0 = G(iωn )δmm0 ,
Σ(iωn )mm0 = Σ(iωn )δmm0 , and (2.49) furthur reduces to a Hilbert transformation
of the noninteracting LDA DOS N 0 (ε) of the orbitals under investigation,
Z
N 0 (ε)
G(iωn ) = dε . (2.50)
iωn − Σ(iωn ) − ε
In the context of LDA+DMFT, we need to consider a multi-orbital effective

impurity problem. The effective action for this problem is given by
Z β Z β X
−1
Sef f = − dτ dτ 0 d†mσ (τ )Gmm 0 0
0 (τ − τ )dm0 σ (τ )
0 0 mm0 σ
Z β X σσ0
0 Umm 0
+ dτ nmσ (τ )nm0 σ0 (τ )
0 mm0 σσ0
2
Z β X Jmm0
− dτ 0
d†mσ (τ )d†m0 σ̄ (τ )dm0 σ (τ )dmσ̄ (τ ). (2.51)
0 0
2
m,m ,σ
as the multi-orbital generalization of (2.15). Here dmσ = cσo rd ld m and nmσ =

nσo rd ld m with o representing the impurity site, and Gmm −1
0 (τ − τ ) is given through
0
the Fourier transform of (2.48). A QMC solution of the impurity problem (2.51)
can be obtained by straightforwardly generalizing the Hirsch-Fye algorithm to the
multi-orbital case (for details, see, e.g., (Held et al., 2003)). However, in this case
it turned out the spin-flip term in (2.51) is very difficult to deal with due to the
sign problem arising from the discrete Hubbard-Stratonovish decoupling (2.23).
Therefore for simplicity this term is usually ignored in practical LDA+DMFT
calculations. Recently it was reported the sign problem associated with the spin-
flip term can be cured to some extent by introducing a different decoupling scheme
for this term (Sakai et al., 2004). Unfortunately, this improvement has only been
implemented for the model studies (Arita and Held, 2005), but not yet for realistic
LDA+DMFT calculations.
The LDA+DMFT scheme presented above has been successfully applied to many
materials with strongly correlated electrons, e.g., La1−x Srx O3 (Anisimov et al.,
1997a; Nekrasov et al., 2000), V2 O3 (Held et al., 2001a; Keller et al., 2004; Keller,
2005), Ce (Held et al., 2001b; McMahan et al., 2003), and Ca(Sr)VO3 (Sekiyama
et al., 2004). A comprehensive review of the LDA+DMFT approach and its
application to various materials is given by Held et al (2003). However, for the
case that the hybridization between the interacting orbitals and other ones is
2
For materials with cubic symmetry, the crystal-field of the ligand atoms causes the tran-
sition metal d orbitals to split into 3-fold degenerate t2g orbitals and 2-fold degenerate eg
orbitals. In cubic perovskites, it is often a good approximation to take only the three t2g
orbitals around the Fermi level into account.
strong, the separation of the interacting-orbital subblock Hdd

0
(k) from the rest
part of HLDA (k) is not a proper treatment, and thus Eq. (2.49) does not hold to
0
a good approximation. This difficulty arises from the fact that we are actually
using LMTOs as the basis for the DMFT calculation, and can be avoided by
choosing the Wannier functions instead of the LMTOs as the basis. We will
discuss this point in detail when we deal with NiO in the next chapter.
43
3. LDA+DMFT INVESTIGATION OF
NIO
3.1 Introduction
NiO is a prototype Mott insulator (Mott, 1949; Brandow, 1977) which has been
under intensive experimental and theoretical investigation for several decades.
Historically the theoretical studies of NiO roughly fall into two categories: those
performed using first-principles approaches such as the DFT and its various im-
provements, and those using the localized approach such as the ligand field the-
ory and the configuration-interaction treatment of a cluster model. These studies
have led to a good understanding of the underlying physical ingredients in NiO,
but a clear, complete picture of NiO is not yet available. This is perhaps due to
the fact that both of the abovementioned approaches have some drawbacks and
hence can not provide a complete treatment of NiO. While the first-principles
studies incorporate the material-specific information without adjustable parame-
ters, they are not able to describe the many-particle features of NiO adequately.
On the other hand, the studies within the localized approaches correctly capture
the atomic nature of the 3d electrons of Ni, but adjustable parameters are in-
volved in these studies to fit the experiment, and more seriously, the band effects
are neglected completely which is known to play an important role in NiO. The
LDA+DMFT approach, cures the drawbacks of the above methods to some ex-
tend, and has proved to be a powerful method for treating strongly correlated
materials. In this chapter we will apply this approach to NiO, and before pre-
senting the computation details and results, we first give a short introduction of
the physical properties of NiO and a brief review of previous studies.
3.1.1 Crystal Structure

NiO has a rocksalt crystal structure structure (space group F m3̄m) in its para-
magnetic phase (Roth, 1958a; Roth, 1958b). Below it’s Néel temperature TN =
523K (Föex, 1948; Tomlinson et al., 1955), NiO is a type-II fcc antiferromagnetic
compound (Skull et al., 1951; Roth, 1958a; Roth, 1958b). For this type of order-
ing, the magnetic moments of Ni ion align ferromagnetically on every (111) plane,
44 3. LDA+DMFT Investigation of NiO
Ni spin down Oxygen Ni spin up

0.417 nm

Figure 3.1: Crystal and magnetic structure of NiO in its antiferromagnetic phase;
red circle-Ni ion with up spin, blue circle-Ni ion with down spin, blue circle-
Oxygen ion. The small rhombohedral distortion is ignored here.
and antiferromagnetically for adjacent planes, as schematically shown in Fig. 3.1.

Moreover, the transition from the paramagnetic state to the antiferromagnetic
state is accompanied by a tiny rhombohedral distortion (space group R3̄m) due
to the exchange-striction effect (Bartel and Morosin, 1971). The lattice constant
has a small increase from about 4.17 Å to 4.20 Å as the temperature increases
from 7 K to 700 K (Bartel and Morosin, 1971). The NaCl crystal structure
of NiO in its paramagnetic phase is the simplest possible structure one can ex-
pect in transition metal compounds1 , and this simplifies the analysis of its band
structure, as can be seen later.
3.1.2 Electronic and Magnetic Properties
The basic experimental facts concerning the electronic and magnetic properties of
NiO are as follows: (a) it is a good insulator with an energy gap about 4.0 eV, and
this gap remains essentially unchanged above the Néel temperature up to over
1000 K. 2 (b) The local magnetic moment of Ni ion is about 1.7 −1.9µB , and this
1
On the other hand, this structure is typical for the 3d transition metal monoxides. These
compounds, including TiO, VO, MnO, FeO, CoO and NiO, all crystallize in NaCl structure.
CuO is the only exception which has a monoclinic structure (Dominguez Rodriguez et al.,
1984).
2
The melting temperature of NiO is about 2260 K, but the experiment on pure NiO is
limited to 1000 K due to the severe stoichiometry problem.
3.1. Introduction 45
value also persists unchanged above TN . The magnetic behavior of NiO, e.g., the
spin-wave spectrum and the value of the Néel temperature can be well accounted
for by a Heisenberg spin Hamiltonian. The susceptibility of the paramagnetic
phase roughly follows the Curie-Weiss law with the Weiss constant θ ≈ −2000
K. (c) Inside the gap of the optical absorption spectrum, there are structures
of Frenkel excitons showing up,3 which can be explained with the ligand field
theory (Sugano et al., 1970; Griffith, 1961). These properties reveal the localized
nature of the Ni 3d electrons, and this further implies the Coulomb repulsion
parameter U is quite large.
An efficient technique for investigating the electronic structure is the photoemis-
sion spectroscopy (PES), which provide a direct measurement of the electronic
energy spectrum of the system if the “matrix element effect” can be neglected for
the energy range concerned. Fig. 3.2 shows the X-ray PES (XPS) spectrum of
NiO from unpublished data of Kowalczyk et al, first appearing in the review paper
of Brandow (1977). Three peaks can be clearly distinguished in this spectrum: a
very pronounced main peak on the top of the valence bands, a shoulder near it
about 1.6 eV below, and a broad satellite at 7.0 eV below. One issue at the heart
of the electronic structure of NiO is how to assign these peaks to the proper final
state after the emission of 3d electron from the ground state, and this problem is
directly related to the nature of the 4 eV insulating gap. In a localized picture, the
electronic states are labelled by the shell configuration of the valence electrons,
e.g., the ground state of NiO is 3d8 . Conventionally, the main peak in Fig. 3.2
was ascribed to the d7 final state, and the satellite is considered to be a d8 L state, Main peak: d7
where L denotes a ligand hole, resulting from a charge-transfer from the Oxygen
2p state to the Ni 3d state accompanying the 3d emission. Thus the insulating
gap was assumed to be formed due to the excitation process d8 + d8 → d7 + d9 .
However, this picture was revised by Fujimori and Minami (1984) based on a
calculation within the configuration-interaction cluster approach. These authors
concluded that the main peak is primarily due to d8 L final state and the satellite Main peak: d8 L
due to d7 final state. This gives rise to the picture that the gap arises from the
process d8 + d8 → d8 L + d9 . The latter picture is supported by the subsequent
experiments (Sawatzky and Allen, 1984; Hüfner et al., 1984) and has been widely
accepted since then. According to a classification scheme by Zaanen, Sawatzky
and Allen (1985), an insulating material is called Mott-Hubbard insulator in the
former case, and charge transfer insulator in the latter case. In a simple picture
of charge transfer insulator, the Oxygen 2p bands are located in between the
lower and upper Hubbard bands arising from the Ni 3d states, and the chemical
potential is sitting inside the gap formed between the Oxygen bands and upper
Hubbard bands. The hybridization effect will blur this picture to some extent
and lead to a resonant state on the top of the valence bands with predominantly
3
Such structures are not observed in the photoemission experiment, however (Sawatzky and
Allen, 1984).
d7
Conventional
d8 L
Figure 3.2: X-ray PES data for NiO by S. Kowalczyk, L. Ley, R. Pollack, and
Shirley, obtained from a single crystal cleaved in dry nitrogen. The zero energy
point is arbitrarily set at the highest point of the main peak. After Brandow
(1977).
Oxygen character.
3.1.3 Previous Studies
As the classic example of Mott insulator, NiO is perhaps the best known and
thoroughly studied material on which various theoretical approaches have been
tried. The paramagnetic band calculations predict NiO to be a metal (Mattheiss,
1972; Shen et al., 1991)4, in contrast with the experimental fact. The spin-
polarized band calculation based on the local spin density approximation (LSDA)
correctly gives a antiferromagnetic ground state (Terakura et al., 1984a; Ter-
akura et al., 1984b), but the obtained insulating gap and the local magnetic
moment are considerably smaller than the experimental values. These facts
are often taken as the evidence that the LDA (LSDA) are not able to describe
the strongly correlated materials. By contrast, the cluster model treated by
configuration-interaction method (Fujimori et al., 1984; Fujimori and Minami,
1984) was quite successful in explaining most of the known experimental facts.
The success of the cluster model was soon reinforced by the combined x-ray-
4
Mattheiss’s band calculation is conventional one based on the APW method in combination
with the LCAO method, i.e., it is not DFT-based and without selfconsistency. A first DFT-
based nonmagnetic band calculation seems to appear in the paper of Shen et al (1991) in
the context of comparison between LDA calculations and their experimental results. The
two results differ mainly in the distance between the Ni 3d bands and O 2p bands, with
the former being much larger than the latter. But both give a metallic solution.
3.1. Introduction 47
photoemission (XPS) and bremsstrahlung-isochromat-spectroscopy (BIS) mea-

surements (Sawatzky and Allen, 1984), and finally led to the recognition that
NiO is not a Mott insulator in the normal sense, but rather a kind of charge
transfer insulator (Zaanen et al., 1985).
For some time it was thought that the purely localized cluster approach is ade-
quate in describing the electronic properties of NiO , but then the angle-resolved
photoemission (ARPES) experiments (Shen et al., 1991; Kuhlenbeck et al., 1991)
revealed that both the localized and band effects play important roles in this sys-
tem. Especially the dispersion of O 2p states can be well accounted for by the
LDA band structure. In the meantime, various attempts to calculate the elec-
tronic structure of NiO from first-principles beyond L(S)DA persisted, among
which the most prominent were the self-interaction-corrected density functional
theory (SIC-DFT) (Svane and Gunnarsson, 1991), the LDA+U method (Anisi-
mov et al., 1991), and the GW approximation (Aryasetiawan and Gunnarsson,
1995; Massidda et al., 1997). These methods represent corrections of the single-
particle potential or the self-energy in one way or another, and lead to substantial
improvements over the L(S)DA results concerning the values of the energy gap
and local moment. Since then, many new works have been performed along these
lines (Anisimov et al., 1993; Anisimov et al., 1994; Hugel and Kamal, 1997; Shick
et al., 1999; Bengone et al., 2000; Faleev et al., 2004; Li et al., 2005), differ-
ing in the basis used and/or the detailed ways of doing approximations. Within
SIC-DFT and LDA+U method, the occupied and unoccupied states are splitted
by the Coulomb interaction U, whereas within LSDA, this splitting is caused by
the Stoner parameter I, which is typically one order of magnitude smaller than
U. Compared with LSDA, SIC-DFT and LDA+U capture the correct physics for
TM oxides, and improve the energy gap and local moment significantly. However,
both of SIC-DFT and LDA+U, in which the self-energy is energy-independent,
fail to take into account the correlation effects properly and thus can’t give a
sensible description of the electronic energy spectrum. The GW method goes
one step further by calculating the self-energy to the lowest order of the screened
interaction W . But its applications to NiO show that the GW approximation is
usually not adequate to describe the strongly correlated systems, and different
implements can give rise to quite different results (Aryasetiawan and Gunnarsson,
1995; Massidda et al., 1997; Li et al., 2005) concerning the relative positions of
the bands and the magnitude of the energy gap. Most recently, an investigation
based on the cluster perturbation theory (Eder et al., 2005) which can be viewed
as an extension of the local cluster approach was reported.
Although lots of progress has been made in the understandings of NiO, there
are still open questions. First of all, although good accordance with experi-
ment has so far been achieved concerning the values of the energy gap and the
local moment, the agreement is far from satisfactory with respect to quasipar-
ticle energy spectra. This is not surprising, because the self-energies employed
in previous approaches are either energy-independent, or suffering from uncon-

trolled approximations. Secondly, almost all the first-principles calculations so
far were focused on the antiferromagnetic ground state with only few exceptions
(Manghi et al., 1994). However, as is well known, both the band gap and the
local magnetic moment persist essentially unchanged above its Néel temperature
TN ∼ 523 K, and recent experiments showed that the long-range magnetic order
has no significant influences on the valence band photoemission spectra (Tjern-
berg et al., 1996) as well as the electron density distribution (Jauch and Reehuis,
2004). These facts reflect the strongly localized nature of the electronic states
in NiO, and can’t be understood through Slater’s antiferromagnetism approach
(Slater, 1951). Therefore, how to produce the paramagnetic insulating state of
NiO at high temperatures, without embarking on the long-range magnetic or-
der, is still a very important problem to deal with from the physical point of
view. Finally, although the charge transfer nature of the insulating gap of NiO is
widely accepted, different voices do persist from both the theoretical side (Hugel
and Kamal, 1997; Bengone et al., 2000) and experimental side (Hüfner et al.,
1992; Schuler et al., 2005). These authors suggested that NiO be either a Mott-
Hubbard insulator, or a mixture of charge-transfer and Mott-Hubbard characters.
Indeed, the conclusions concerning the nature of NiO drawn by experimentalists
also largely depend on the analysis based on some particular theoretical models.
Therefore, it is still highly desirable to have more first-principles investigations of
the electronic structure of NiO within different approaches to make the situation
more clear.
In this work, we perform a LDA+DMFT calculation of NiO for the paramagnetic
insulating phase, based on the procedure proposed by Anisimov et al (2005),
where the basis of WFs are used in the DMFT calculation. The energy gap,
local magnetic moment, and the electronic spectrum are obtained. Since these
quantities have no significant dependence on the temperature, we are allowed
to compare our results with experimental data at low temperatures and other
theoretical results for the ground state. The comparison shows that not only the
energy gap and the local moment, but also the electronic spectrum are in good
agreement with experiment.
3.2 Method and Results
The LDA+DMFT scheme, in most of its applications so far, is close in spirit to

the model approach, namely, the LDA calculation serves to provide the necessary
parameters of a many-body model Hamiltonian, and this Hamiltonian is in turn
solved by means of DMFT. Since the LDA calculation usually involves a large
number of valence s, p, d orbitals originating from all the atoms in the unit cell,
it is mandatory to project all the orbitals except for a few relevant ones to be
3.2. Method and Results 49
included in the model Hamiltonian so that DMFT can handle. For transition-
metal (TM) compounds, the most relevant orbitals responsible for the physical
properties are the TM valence d orbitals. As a matter of fact, in most of the
previous studies one just takes the LDA density of states (DOS) of the TM d or-
bitals as the input for the DMFT calculation, as indicated by the equation (2.50)
and this means that one is using the atomic-like LMTO basis for constructing
the many-body Hamiltonian (suppose the LDA problem is solved by the LMTO
method). However, as already mentioned in the previous chapter, for some ma-
terials, e.g., the late TM oxides where there is a strong covalency effect between
the TM d orbitals and Oxygen 2p orbitals, the atomic-like d orbitals are not the
proper basis for the Hubbard-like many-body Hamiltonian. Actually in this case
the proper choice of the basis is the d-like Wannier functions (WFs) (Wannier,
1937) which on the one hand have the same symmetry as the atomic-like d or-
bitals, and on the other hand implicitly take into account of the admixture of
Oxygen 2p orbitals resulting from the hybridization effect. This is the strategy
adopted in a recent work of Anisimov et al (2005), where WFs of the d-like bands
were constructed from the solution of the LDA band structure, and used as the
basis for the DMFT calculation. In the abovementioned work, calculations were
carried out for SrVO3 and V2 O3 as examples, and in the present work we apply
the same procedure to NiO.
To begin with, we perform a standard LDA band calculation for the paramagnetic
phase of NiO using the LMTO-ASA (Andersen, 1975) method with the combined
correction term included5 . As has been discussed above, the lattice constant
increases from about 4.17 Å to within a temperature range from 7 K to 700
K (Bartel and Morosin, 1971). Such an increase only causes a small deviation
of the bands far away from the Fermi energy, but no noticeable change of the
Ni 3d-derived bands on which we are focusing. In addition, since in future we
will compare our result with the low-temperature experimental data anyway, we
choose the lattice constant a = 4.17 Å throughout this work. The calculated
nonmagnetic band structure of NiO along the selected high symmetry lines is
shown in Fig. 3.3 (the first Brillouin zone and the denotation of its high symmetry
point is shown in Fig. 3.4 for illustrating purpose.) and it is in agreement with
those published in literatures (Shen et al., 1991; Eder et al., 2005). Based on
the LDA band structure, we can start to construct a set of WFs, which will be
used as the basis for the DMFT calculation, following the procedure proposed by
Anisimov et al (2005). Let’s discuss this procedure in some detail below.
5
Here we use the Stuttgart TB-LMTO code (Vers. 4.7) for the LDA band calculation.
Thanks to Dr. Igor Nekrasov for making the code available.
Nonmagnetic band structure of NiO

10
5
Energy (eV)
-5
L Γ X W L K Γ
Figure 3.3: Nonmagnetic band structure of NiO obtained by the LMTO method;
the Fermi energy is set to zero. The five d-like Bloch bands used for constructing
Wannier functions are highlighted by green color
L
Γ
X
W K
Figure 3.4: The first Brillouin zone of the fcc lattice

3.2.1 Wannier Function Construction

The solution of the LDA band problem is a set of Bloch states |ψik i, given by
one-electron Schrödinger equation,
ĤLDA |ψik i = εi(k)|ψik i. (3.1)
For a given set of Bloch functions, a set of Wannier functions (WFs) can be defined
as a Fourier transformation of these Bloch functions. However, as has been
discussed in Section 1.4, in this way the WFs can not be uniquely defined. This
is because the Bloch functions themselves are determined only up to an arbitrary
phase factor, and for the multi-band case, an additional unitary transformation
connecting the different Bloch functions at each k point is involved, as indicated
in Eq. (1.49). In another word, it means there exist some degrees of freedom of
constructing WFs corresponding to freedom of choice of a unitary transformation
(k)
matrix Uji for the Bloch functions, namely
1 X ik·R ek
|WiR i = √ e |ψi i. (3.2)
L k
with X (k)
|ψeik i = Uji |ψjk i. (3.3)
j
(resulting Bloch function |ψeik i will not be in general case an eigenfunction of the
Hamiltonian but has a meaning of Bloch sum of WFs (see below |W fnk i in Eq.
(k)
(3.4) ) ). There is no rigorous way to determine Uji , and on the other hand
(k)
the freedom of choosing Uji can be utilized to obtain WFs with desired prop-
erties by exerting additional restrictions. Among others Marzari and Vanderbilt
(1997) proposed the condition of maximum localization for WFs, resulting in a
(k)
variational procedure to calculate Uji . To get a good initial guess Marzari and
Vanderbildt (1997) proposed choosing a set of localized trial orbitals |φn i and
projecting them onto the Bloch functions |ψik i. It was found that this starting
guess is usually quite good. This fact later led to the simplified calculating scheme
proposed by Ku et al (2002) where the variational procedure was abandoned and
the result of the projection was considered as the final step.
For constructing a set of WFs with some particular symmetry, one can select
either a set of Bloch bands (N1 , . . . , N2 ), or choose the energy interval (E1 , E2 )
in which the bands are located. We first define the nonorthogonalized WFs in
reciprocal space |Wf k i as the projection of the set of site-centered atomic-like trial
n
orbitals |φn i onto the Bloch functions |ψik i of the chosen bands,
N2
X X
fki ≡
|W |ψik ihψik |φn i = |ψik ihψik |φn i. (3.4)
n
i=N1 i(E1 ≤εi (k)≤E2 )
f R i are given by
Then the real space WFs |Wn
X
f R i = √1
|W f k i.
e−ik·R |W (3.5)
n n
L k
In the present work the energy bands which are included in the construction (3.4)
are the five d-like bands sitting around the Fermi energy, as highlighted in green
in Fig. 3.3. These five Bloch bands are dominated by Ni 3d states, but also
have considerable contributions from O 2p states due to the hybridization effect.
Concerning the trial orbitals |φn i, the simplest choice would be the Bloch sum of
the LMTOs which are also used as the basis for solving the LDA band problem,
1 X ik·R
|χkµ (r)i = √ e |χµ (r − R)i. (3.6)
L R
Here the index µ is a combination of rlm where r denote the different atom in
the unit cell, and l, m are angular momentum and magnetic quantum numbers
respectively (see Eq. (1.47)). In this work n in |φn i (Eq. (3.4)) is chosen to
enumerate the five Ni 3d LMTOs, namely, those |χkrlm i with r corresponding to
Ni ion, l = d, and m = 1, 2, . . . , 5. Note that a WF in reciprocal space |W fnk i
defined in (3.4) does not coincide with the Bloch function |ψn i in the multi-band
k
case due to the summation over band index i in (3.4), but rather with a linear
combination of them, i.e., |ψenk i in Eq. (3.3) after orthonormalization. Actually
one can consider them as Bloch sums of WFs analogous to the Bloch sum of the
basis functions χkµ (r) (Eq. (3.6)). The coefficients hψik |φn i in (3.4) define (after
(k)
orthonormalization) the unitary transformation matrix Uji in Eq. (3.3).
By expressing the LDA Hamiltonian in Eq. (3.1) within the basis of TB-LMTOs
|χkµ i = |χkrlm i, the band problem (3.1) reduces to a linear eigenvalue problem.
Solving (3.1), one can obtain eigenvalues εi (k) and eigenvectors cµi (k), and as
well as the Bloch function |ψik i which are given by
X
|ψik i = cµi (k)|χkµ i. (3.7)
µ
For an orthogonal set of LMTO basis |χkµ i6 , Eq. (3.7) means the coefficients
cµi (k) = hχkµ |ψik i, c∗µi (k) = hψik |χkµ i. (3.8)
By using Eq. (3.4), (3.7) and (3.8), one can arrive at

N2
X N2 X
X X
fki =
|W |ψik ic?ni (k) = cµi (k)c∗ni (k)|χkµ i = b̃kµn |χkµ i, (3.9)
n
i=N1 i=N1 µ µ
6
For a set of nonorthogonal set of LMTO basis |χ̃kµ i, one can first orthogonalize them,
|χ̃kµ i → |χkµ i, and then take the orthogonalized functions |χkµ i as the trial function |φn i in
(3.4).
with
N2
X
b̃kµn ≡ cµi (k)c∗ni (k). (3.10)
i=N1
Eq. (3.9) and (3.10) provide a well-defined way to construct a subset of nonorthog-
onal WFs from a (large) set of LMTOs.
In order to orthonormalize the WFs (3.9) one needs to calculate the overlapping
matrix Onn0 (k)
N2
X
f k |W
Onn0 (k) ≡ hW f k0 i = cni (k)c?n0 i (k), (3.11)
n n
i=N1
and its inverse square root Snn0 (k) is defined as

−1/2
Snn0 (k) ≡ Onn0 (k). (3.12)
(orthogonality of Bloch states hψnk |ψnk0 i = δnn0 was used in the derivation of
(3.11).)
From (3.9) and (3.12), the orthonormalized WFs in k-space |Wnk i can be obtained
as
X N2
X X
|Wnk i = f k0 i =
Snn0 (k)|W |ψik ic̄∗ni (k) = bkµn |χkµ i, (3.13)
n
n0 i=N1 µ
with
X
c̄∗ni (k) ≡ hψik |Wnk i = Snn0 (k)c∗n0 i (k), (3.14)
n0
N2
X
bkµn ≡ hχkµ |Wnk i = cµi (k)c̄∗ni (k). (3.15)
i=N1
The real space site-centered WFs |WnR i are given by a simple Fourier transforma-
tion of |Wnk i. But here all the calculations can be done in the momentum space,
and an explicit calculation of the real space WFs are not needed.
With the set of orthogonal WFs |Wnk i constructed, one can easily obtain the
LDA Hamiltonian matrix H WF (k) within this basis set. By using Eqn. (3.1) and
(3.13), one can get
N2
X
WF
Hnn 0 (k) = hWnk |ĤLDA (k)|Wnk0 i = c̄ni εi (k)c̄∗n0 i
i=N1
X
= bk∗ k
µn HLDA (k)µν bνn0 , (3.16)
µ,ν
10
t2g states in LMTO basis
8 eg states in LMTO basis
DOS ( states/(eV.cell) )
t2g states in WF basis
eg states in WF basis
6
0
-6 -5 -4 -3 -2 -1 0 1 2
Energy (eV)
Figure 3.5: LDA partial DOSs in the basis of Ni 3d LMTOs (t2g -black dashed
curve, eg -blue dashed cure) and d-like WFs (t2g -red solid curve, eg -green solid
curve).
where HLDA (k)µν = hχkµ |ĤLDA (k)|χkν i.

Thus, we have described a simple and practical procedure to construct a set
of WFs which span the same Hilbert space as a specified set of Bloch bands do.
The Hamiltonian matrix within this basis set, as indicated in (3.16), can be easily
obtained directly from (a subset set of) the eigenvalues and eigenvectors of the
LDA band problem. In the case of NiO, at each k point this leads to a set of five
WFs with d symmetry, and Hnn WF
0 (k) is a 5 × 5 Hamiltonian matrix.
3.2.2 LDA Results
In the previous section, we have demonstrated the procedure for calculating the
WFs for the correlated d-like orbitals, and show that the low-energy Hamiltonian
matrix within this basis can be constructed straightforwardly by the solution of
the LDA band problem. The more detailed formulation and derivations can be
found in the original paper of Anisimov et al (2005). In this work the procedure is
applied to NiO, and the WFs are constructed so as to span the Hilbert subspace
containing the five green Bloch bands across the Fermi level (see Fig. 3.3). Here
we first present the density of states (DOS) of these WFs in Fig. 3.5, and as
a comparison Ni 3d LMTO DOSs are also shown. Since the WFs constructed
in the above procedure have d symmetry, they can also be classified into 3-fold
degenerate t2g states and 2-fold degenerate eg states for the cubic system, like
NiO. However, one should keep in mind that these WFs are not pure d states,
but rather the anti-bonding-like states resulting from the hybridization between
Ni 3d and O 2p states. Actually this is self-evident in Fig. 3.5, where in the Ni
10
LMTO LDA DOS (states/( eV.atom )

8
Ni 3d states in LMTO basis
O 2p states in LMTO basis
6
0
-8 -6 -4 -2 0 2
Energy ( eV )
Figure 3.6: LDA partial DOS of Ni 3d states (red curve) and O 2p states (green
curve) in the basis of LMTOs. The blue dashed box indicates the energy window
in which the Bloch bands are included in the WF construction.
3d-dominant region (from −3 to 1.4 eV), WFs have more spectral weight than
LMTOs, with the extra spectral weight mainly coming from the contributions of
Oxygen 2p states. This point is further illustrated in Fig. 3.6, in which the LMTO
DOSs of Ni 3d states and O 2p states are both shown. By comparing Fig. 3.5
and Fig. 3.6, one can see that the difference between the spectral weights of the
WFs and LMTOs in the Ni 3d-dominant region (indicated by the dashed box in
Fig. 3.6) is well accounted for by the O 2p states. On the other hand, these WFs,
by its construction, omit the region of higher binding energies which primarily
has O character but strongly hybridized with Ni d states. Therefore, we do not
expect the present calculation would reproduce the well-known satellite structure
appearing at higher binding energies, but aim at providing a quantitatively good
description of the electronic spectrum around the gap region.
So far we already have the low-energy tight-binding Hamiltonian H WF (k) ex-

pressed in the basis of WFs, and to have a many-body theory of NiO, one needs
to supplement H WF (k) with the multi-orbital Coulomb interaction and Hund-
rule’s coupling exchange terms. On the other hand, these interaction terms have
been taken into account in H WF (k) in an average way, and therefore in princi-
ple a term H dc corresponding to these contributions should be subtracted from
H WF (k) to avoid double-counting. However, in the present case when only the
the interacting orbitals are included in the effective Hamiltonian, H dc is just an
irrelevant constant and does not really need to be considered.
Based on the above consideratons, the full Hamiltonian reads

X
H = H WF (k)nn0 d†knσ dkn0σ
k,n,n0 ,σ
X σσ0
Unn X Jnn0 † †
d†inσ dinσ d†in0 σ0 din0 σ0 −
0 0 0
+ d d 0 din0 σ dinσ̄ (3.17)
2 2 inσ in σ̄
i,n,n0 ,σ,σ0 i,n,n ,σ0
in which d†inσ (dinσ ) creates (destroys) an electron with spin σ in the n-th Wannier
orbital (here n ranging from 1 to 5) centering at lattice site i, and d†k,n,σ (dk,n,σ ) is
σσ0
its Fourier transform. Furthermore Unn 0 denotes the interaction strength between
one electron with spin σ in Wannier orbital n and that with spin σ 0 in orbital
σσ0
n0 , and satisfies the relation that Unn 0 = U − 2J(1 − δnn0 ) − Jδσσ 0 for the cubic
system. In principle the values of the U, J parameters here are accessible from
the constained LDA calculation. However, since now we are using the WFs as
the basis, the suitable U values should be smaller than those obtained by the
constained LDA using LMTO basis. This is due to the fact that WFs are more
extended than LMTOs. Actually concerning the value of the averaged Coulomb
interaction Ū (Eq. (2.43)), we can roughly estimate this value within the WF
basis Ū WF from that within LMTO basis U LMTO by Ū WF = Ū LMTO (1 − x)2
where x is the admixture of the Oxygen states into the WFs (see Appendix B
for details). For the d-band WFs constructed in this work, x ≈ 0.15, and the
constrained LDA calculation within LMTO basis gives Ū LMTO = 8 eV (Anisimov
et al., 1991). This leads to an estimated Ū WF = 5.78 eV. Since the results usually
don’t show significant dependence on the J value, we assume the J value does
not change much from LMTO basis to WF basis, and choose J = 1 eV which is
typical for such systems. From Eq. (2.43), one can obtain U value from Ū and
J by U = Ū + (20/9)J in which the number of orbitals M = 5 is used. All the
above analysis finally give rise to U WF ≈ 8 eV.
3.2.3 LDA+DMFT Results and Comparison with

Experiment
We solve the above material-specific model (3.17) by means of DMFT (Kotliar

and Vollhardt, 2004; Georges et al., 1996; Vollhardt et al., 2005) that represents
the original lattice model by a single-impurity model subject to a self-consistent
condition. The impurity problem is in turn solved by the quantum Monte-Carlo
(QMC) technique (Hirsch and Fye, 1986), and the imaginary Green’s function
data obtained by QMC are treated by maximum entropy method (MEM) (Jarrell
and Gubernatis, 1996) to get the physical spectral function. The QMC simulation
can be performed at different temperatures, but the computation effort scales
with 1/T 3 and therefore the calculations are restricted within relatively high
2.5
t2g states
spectrum of NiO (states/eV.atom)

2 eg states
total states
1.5
0.5
0
-8 -6 -4 -2 0 2 4 6
Energy (eV)
Figure 3.7: Theoretical energy spectrum of NiO obtained by the LDA+DMFT

calculation for T = 1160 K, and U = 8 eV, J = 1 eV. t2g states (blue curve) and
eg states (green curve), and total d states (red dotted curve) are shown.
temperatures. The electronic energy spectrum obtained with QMC simulation at

T = 1160 K (corresponding to β = 10), and the interaction parameters U = 8
eV and J = 1 eV are shown in Fig. 3.7. From which one can see that the t2g
bands are fully filled and form the main peak below the Fermi level. The eg
bands, on the other hand, are splitted into lower and upper Hubbard bands, and
an insulating gap is formed between the occupied t2g bands and unoccupied eg
bands. The occupied eg bands, together with the lower edge of the t2g bands
forms the shoulder below the main peak.
To analyze the effect of the QMC simulation temperature on the spectrum of
NiO, we also performed the calculation at T = 725 K (corresponding to β = 16).
The result is shown is in Fig. 3.8 in comparison with that obtained at T = 1160 K.
From the comparison one can see the results at two different temperatures do not
show any significant difference, and the overall behaviors of the spectra almost fall
on top of each other, in agreement with the conclusion from the photoemission
experiment (Tjernberg et al., 1996). Since there are always some numerical errors
from QMC and some extent of indeterminacy of the MEM spectrum, here we
would not associate the small detailed differences between the results at two
different temperatures with any physical meanings.
Now it is worthwhile to compare our LDA+DMFT results with experiment.
For this purpose we choose the combined XPS+BIS experiment carried out by
Sawatzky and Allen (1984) which can be used directly for comparison with theo-
retical energy spectrum. This experiment was done at room temperature, i.e., in
the antiferromagnetic phase. Although the theoretical calculation and the exper-
iment are performed at different magnetic phases, the comparison is justified by
2.5
T=1160 K
Spectrum of NiO (states/atom.eV)

T=725 K
2
1.5
0.5
0
-8 -6 -4 -2 0 2 4 6
Energy (eV)
Figure 3.8: Theoretical spectrum of NiO obtained by the LDA+DMFT approach

for U = 8 eV, J = 1 eV at T = 1160 K (red curve) and T = 725 K (green curve)
respectively.
the observation that the electronic structure of NiO is not significantly affected
by the magnetic phase transition (Tjernberg et al., 1996; Jauch and Reehuis,
2004), as mentioned earlier in the introduction. Due to the ambiguity involved in
the determination of the position of the chemical potential inside the insulating
gap7 , here we shift the position of the chemical potential coming from the theo-
retical calculation to be in accord with the experimental one. From Fig. 3.9, one
can see that concerning the conduction bands, the theoretical spectrum agrees
with the experimental BIS data almost perfectly, and moreover the main peak
below the Fermi energy is also well reproduced. The second peak appearing in
the theoretical spectrum, on the other hand, is relatively far off the experimental
shoulder structure. On the other hand, the determination of the exact behavior
of the energy spectrum around this region is quite tricky because there is some
degree of ambiguity involved in the MEM for energy ranges relatively far away
from the Fermi level. Therefore, we leave the proper explanation of the shoulder
structure within the LDA+DMFT approach to future studies. A full theory of
NiO should allow for the hybridization between Ni 3d states and O 2p states
evolving under the influence of the interactions among d electrons. This requires
an explicit inclusion of Oxygen 2p states, which is however not performed in this
calculation. Therefore we take the present result as the first approximation of
the full solution of the NiO problem within the LDA+DMFT approach.
Although the interaction parameter U = 8 eV estimated from the constrained
LDA calculation gives rise to a theoretical spectrum in excellent agreement with
7
In principle, for finite temperatures, the chemical potential position in the gap can be
determined exactly, but this requires a rather high accuracy of the calculation and hence
is not feasible in the present case.
Experimental data
Spectrum of NiO (arbi. units)

LDA+DMFT spectrum
-8 -6 -4 -2 0 2 4 6
Energy (eV)
Figure 3.9: Theoretical spectrum obtained by the LDA+DMFT calculation (red

curve) for U = 8 eV and J = 1 eV at T = 1160 K, compared with experimental
XPS+BIS data (black dots) [after Sawatzky and Allen (1984)]. The zero energy
point of the theoretical curve is shifted to fit the energy scale of the experimental
data.
experiment, it is interesting to see how the spectrum will change by varying the
U value. Therefore we also performed the calculations for U = 8.5 eV and 7.5
eV with fixed J = 1 eV at T = 1160 K respectively. The results are shown in
Fig. 3.10, in comparison with that for U = 8 eV and the experimental data. To
illustrate more clearly the change of the insulating gaps for different U values,
the positions of the conduction-band peaks are put together by purpose. From
Fig. 3.10, we see that indeed for U = 8 eV we get the best agreement with
experiment. The Mott insulating gap increases roughly linearly with the U value.
After discussing the electronic energy spectrum, we will then turn to the en-
ergy gap and local magnetic moment. In table 3.1 the values of the energy gap
and local magnetic moment obtained in this work are presented, in comparison
with the experimental results, as well as those obtained within other theoretical
approaches. The energy gap 4.4 eV here is measured by the distance between
the half-maximum points of the first valence peak and the conduction peak, in
the same way as done in the paper of Sawatzky and Allen (1984). Concerning
the magnetic moment, it is worth mentioning that the value of 1.70 µB is ob-
tained in an indirect way here. This is due to the fact that the basis employed
in the DMFT calculation is WFs, and hence the calculated value is not what is
measured in experiment, namely, the local moment in the immediate vicinity of
Ni ion. Specifically, in this calculation the material-specific model Hamiltonian
involves five d-like WFs occupied by eight electrons, and hence gives rise to a
magnetic moment M = 2 µB for these WFs (practically we get M ≈ 1.99 from
U=7.5 eV, J=1 eV

U= 8 eV, J=1 eV
Spectra of NiO (arbi. unit)

U=8.5 eV, J=1 eV
Experimental data
-8 -6 -4 -2 0 2 4 6
Energy (eV)
Figure 3.10: Theoretical spectra obtained by the LDA+DMFT calculations for

U = 7.5, 8, and 8.5 eV (blue, red, and green curves) respectively with fixed J = 1
eV and temperature T = 1160 K. Experimental XPS+BIS data [after Sawatzky
and Allen (1984)] are also shown for comparison.
our QMC calculation, and the small derivation from 2 µB may be due to the
numerical error or the minor minus-sign problem present in the present calcula-
tion). However, we can roughly estimate the magnetic moment associated with
Ni ion by M(1 − x) where M is the moment on Wannier orbitals and x is the
contribution of O 2p states to WFs8 . From the LDA band calculations and the
construction of WFs we know x ≈ 0.15, and this gives M(1 − x) ≈ 1.70 which is
comparable to the experimental value and those obtained within other theoretical
approaches.
3.3 Conclusions
In this work we applied the LDA+DMFT approach, in its recently developed

form, to NiO. Specifically, we constructed a subset of WFs corresponding to
the five Bloch bands across the Fermi level from the LDA band structure, and
used them as the basis to build up a material-specific multi-band Hubbard-like
Hamiltonian. This Hamiltonian is further solved by DMFT. The electronic energy
spectrum obtained in this way is in excellent agreement with the experimental
XPS+BIS results around the gap region. At the same time, we obtained the
energy gap and local magnetic moment which also show good agreement with
experimental results and those obtained by other theoretical approaches.
8
This relationship is derived under the same argument as that for the interaction parameter
U for WFs. See Appendix B
3.3. Conclusions 61
Tab. 3.1: The LSDA, GW, LDA+U, LDA+DMFT (present work) and experi-
mental energy gaps and magnetic moments for NiO
LSDA GWa LDA+Ub LDA+DMFT Expt.
Energy gaps (eV) 0.3 3.7 3.7 4.4 4.3c ,4.0d

Moments (µB ) 1.09 1.83 1.70 1.70 1.64e ,1.77f ,1.90g
a
S. Massidda et al (1997) b
V. I. Anisimov et al (1993)
c
G. A. Sawatzky and J. W. Allen (1984) d
S. Hüfner et al (1984)
e
H. A. Alperin (1962) f
B. E. F. Fender et al (1968)
g
A. K. Cheetham and D. A. O. Hope (1983)
However, our present treatment is not a perfect one. In the construction of

the material-specific Hamiltonian (3.17), we only include the five “anti-bonding”
bands (which have mainly Ni 3d characters in the LDA calculation), but the
three “bonding” bands (which are the mixture of the Ni 3d and O 2p states but
have more O contributions) below them are completely neglected. Because of
this, in the present work, the ratio of the contributions to the WFs from the Ni
3d and O 2p states depends on the LDA results. This means the valence bands
close to the Fermi level have mainly Ni 3d characters, and thus the insulating
gap is of the Mott-Hubbard type. This seems to be in contradiction with the
widely accepted view that NiO is a charge-transfer insulator (Fujimori et al.,
1984; Sawatzky and Allen, 1984; Zaanen et al., 1985). However, since different
voices concerning this problem have persisted (Hugel and Kamal, 1997; Bengone
et al., 2000; Schuler et al., 2005), either saying it is a Mott-Hubbard insulator or
a mixture of both types, we can leave this question open for the moment, and
hope to clarify this problem in future studies. Another drawback of the present
framework is that, by neglecting the “bonding band”, it is not possible for us to
reproduce the satellite structure appearing at high binding energies. A proper
explanation of this structure from first-principles is another important issue of
NiO.
In spite of the limitations of the current scheme, the LDA+DMFT approach is
apparently a very promising method to deal with the late transition metal monox-
ides like NiO, by incorporating the first-principles information and strongly-
correlated physics in one theoretical framework. A more complete treatment
of NiO within the LDA+DMFT approach would require an explicit inclusion of
the Oxygen 2p bands in the DMFT calculation, which is an on-going project. As
a first approximation, the present results are quite encouraging, and we hope this
work will stimulate more theoretical investigations along this direction.
63
4. TOWARDS A SELF-CONSISTENT
LDA+DMFT SCHEME
4.1 General Motivation
In the previous two chapters we have discussed the LDA+DMFT approach, and
applied it to NiO. We also pointed out that in most of its practical applications,
the LDA+DMFT approach is implemented in a way that is close to the model
Hamiltonian approach. Namely, one first performs a LDA band-structure cal-
culation and derives a Hubbard-like many-body Hamiltonian for the correlated
“heavy” orbitals, and then solves this Hamiltonian by DMFT. Such a procedure
is usually quite successful, but it has apparent limitations for dealing with the
materials for which it is not sufficient considering just the correlated orbitals.
This is because some other orbitals may also play important roles in the system
through their hybridizations with these correlated ones. In the case of NiO, these
“other” orbitals are the Oxygen 2p ones, and many controversies about NiO arise
from their interplay with the Ni 3d orbitals. This requires a more sophisticated
description of the system beyond the Hubbard-like model, by treating the both
kinds of orbitals on the same footing.
Moreover, for such kind of systems, it is expected that the DMFT treatment of
the physically relevant orbitals will modify the electron density resulting from
the DFT-LDA calculation. Since the LDA band structure uniquely depends on
the electron density, the change of the electron density will inversely leads to a
change of the LDA band structure, and thus the starting point of the DMFT
calculation. Therefore, in principle one needs a global self-consistent loop which
allows for a feedback from DMFT to LDA. Such a global loop is similar to the
DFT one, but has one more ingredient, namely, a chosen subset of orbitals is
treated by DMFT. In another word, the full orbital space is separated into two
subspaces, one for the correlated orbitals which are treated by DMFT, and the
other for all the rest orbitals which are described by usual DFT-LDA calculation.
Implemented in this way, the LDA+DMFT approach is now close in spirit to the
usual first-principle electronic structure method.
The above picture of the global self-consistent LDA+DMFT scheme, motivated
64 4. Towards A Self-Consistent LDA+DMFT Scheme
by physical considerations, is largely conceptual. A more rigorous formulation of

this scheme can be worked out by applying variational principle to a constructed
energy functional, similar to the approach used in deriving the Kohn-Sham equa-
tions. The new LDA+DMFT energy functional, which can be viewed as an ex-
tension of the LDA electronic energy functional, has an additional basic variable,
i.e. the local Green’s function Gloc , besides the electron density ρ(r) (Savrasov
and Kotliar, 2001; Lichtenstein et al., 2003). The extremization of this functional
can lead to a set of LDA+DMFT equations representing a global self-consistent
loop. A recently proposed GW+DMFT scheme (Biermann et al., 2003) is based
on the similar idea, but the band structure is now solved by the GW method
instead of LDA.
Depending on the basis used to represent the correlated orbitals, the method em-
ployed to solve the DMFT equations, and the interface between the band struc-
ture calculations and the DMFT calculations, the self-consistent LDA+DMFT
scheme can be implemented in different ways. Here we will present a full-orbital
calculation scheme proposed recently by Anisimov et al (2005) in which the band
structure calculation and the DMFT treatment of the strongly correlated orbitals
can be coupled in self-consistent way. In this scheme, the LDA band problem is
solved within a set of LMTO basis as usual, but the correlated orbitals which will
be further treated DMFT are represented by a set of Wannier functions (WFs).
These WFs (labelled by index n) span a subspace of the whole Hilbert space
defined by the LMTOs (labelled by index µ) used in the LDA calculation. The
self-energy obtained within the DMFT calculation, can be transformed from the
WF subspace back into the LMTO full-orbital space. With the self-energy in the
full-orbital space, one can calculate the full interacting Green’s function and of
course the total electron density, thus closing the self-consistent loop. In the next
section we will give more detailed formulations.
4.2 Flow Diagram and Formulation
The flow diagram of the fully self-consistent LDA+DMFT scheme is shown is

Fig. 4.1. Here the basic ingredients are:
1. Solve the LDA band structure within the basis of LMTOs χkµ (r).
2. From the LDA band structure, the WFs Wn (r − R) = WnR (r) for a subset
of correlated orbitals (specified by a set of Bloch bands from N1 to N2 across
the Fermi level), and their Bloch sum Wnk (r) can be calculated, following
the procedure described in Sec. 3.2.1. At each k point, one can derive a low-
energy tight-binding Hamiltonian Hnn WF
0 (k) matrix within the basis Wn (r).
k
The transformation from LMTOs to WFs is defined by a set of coefficients

4.2. Flow Diagram and Formulation 65

DMFT(QMC)

k start
HLDA , bkµnstart - WF
Hnn - ΣW F WF
6
0 (k) nn0 (iω), Gnn0 (iω)

transformation back
to full orbital space

?
Σµν (k, iω)

feedback k- integration

(self-consistent
?
LDA+DMFT)
Gµν (iω)
R
bkµnnew Mµν
(m)

constraint on WFs ?
(m)
? ρ[Mµν ]
WF (0)
∂Hnn
Un = ∂QWF
nn (0)
?
V[ρ]
Figure 4.1: Flow diagram of the fully self-consistent LDA+DMFT scheme. The
red lines indicate the steps which has not been realized so far (after Anisimov et
al. (2005)).
bkµn = hχkµ (r)|Wnk (r)i (see Eq. (3.15) for its definition). The transformation
of the full LDA Hamiltonian matrix HLDA (k) within the LMTO basis to
that within subset WF basis H WF (k) is provided by Eq. (3.16).
3. Calculate the Coulomb interaction parameter U, J within “constrained”

LDA, and form a material-specific generalized multi-orbital Hubbard-like
Hamiltonian. Here the U, J values are defined for the WFs. The procedure
goes similar to the conventional “constrained” LDA for the LMTO basis.
For instance, the U value for orbital n is defined as,
0
∂Enn
Un = , (4.1)
∂Q0nn
where Enn0
and Q0nn are the elements of the energy and occupancy matrix
for the set of real-space WFs centering at the origin,
N2 X
!
1 X
0
Enn 0 = hWn0 | |ψik iεi (k)θ(EF − εi (k))hψik | |Wn00 i
L i=N k
1
N2
X X
1
= c̄ni (k)εi (k)θ(EF − εi(k))c̄∗n0 i (k), (4.2)
L i=N
1 k
N2 X
!
1 X
Q0nn0 = hWn0 | |ψik iθ(EF − εi(k))hψik | |Wn00 i
L i=N k
1
N2
X X
1
= c̄ni (k)θ(EF − εi (k))c̄∗n0 i (k). (4.3)
L i=N
1 k
In above equations εi (k) and |ψik i are the eigenvalues and eigenstates (Bloch
states) of the LDA Hamiltonian, and c̄ni (k), c̄∗n0 i (k) are defined through Eq.
(3.14). In addition EF is the Fermi energy and θ(x) is the step function.
4. The many-body Hamiltonian is solved within DMFT , and in particular,

the impurity problem is solved by Hirsch-Fye quantum Monte Carlo (QMC)
method. From the solution of DMFT, one can obtain the local self-energy
nn0 (iωn ) and Green’s functions Gnn0 (iωn )
ΣWF WF
5. Utilizing the transformation coefficients bkµn , the local self-energy ΣWF

nn0 (iωn )
can be converted back (upfolded) into the full Hilbert space
X
Σµν (k, iωn ) = bkµn (ΣWF dc k∗
nn0 (iωn ) − E )bνn0 . (4.4)
nn0
Here E dc is the double-counting correction for the Coulomb interaction en-

ergy which has already been taken into account by LDA. Now the resultant
4.2. Flow Diagram and Formulation 67
full self-energy Σµν (k, iωn ) is free of double-counting. An approximate form

for E dc is given by Eq. (2.42) which reduces to
1
E dc = Ū nd (nd − 1) (4.5)
2
in the case that the Hund’s rule coupling term is neglected. Here nd should
be understood as the total occupancy of the WFs. In the literature other
form of E dc exists, e.g.,
M
1 X
E dc
= ReΣnn (0) (4.6)
M n=1
is used by Lichtenstein et al (2003) where M is the number of the correlated

orbitals.
6. With Σµν (k, iωn ), one can calculate the full interacting Green’s function
1X
Gµν (iωn ) = [iωn + µ − HLDA (k) − Σ(k, iωn )]−1
µν . (4.7)
L k
Here the matrix inversion is with respect to the LMTO basis.

7. Energy moments can be calculated from the (Matsubara) Green’s function
∞
X
1
(m)
Mµν = (iωn )m Gµν (iωn ). (4.8)
β iωn =−∞
Since from QMC calculation, we only have the Green’s function value at a
finite number (in order of 10) of Matsubara frequencies, the calculation of
(m)
Mµν with 4.8 is a highly nontrivial problem. The detailed procedure to
treat the problem is given is Appendix C.
8. For a given set of LMTOs, the charge density ρ(r) can be obtained as
a functional of energy moment, and new effective LDA potential can be
constructed V (r) = V (ρ(r)). This can be schematically illustrated for an
orthogonal basis set in the following. In the case only the diagonal energy
(n) (n)
moments mµ = Mµµ are nonzero and contribute. The spherically averaged
charge density can be evaluated as
Z Xocc X Z EF
4πρ(r) = dr̂ k 2
|ψj (r)| = Nµ (ε)χ2µ (ε, r)dε, (4.9)
j,k µ
in which χµ (ε, r) is the radial part of the energy-dependent partial wave,

and Nµ (ε) is its partial density of states. The linear approximation, which
leads to the LMTOs, means that
χµ (ε, r) ≈ χµ (εν , r) + (ε − εµ )χ̇µ (εν , r). (4.10)
The energy moments are defined as

Z EF
(n)
mµ = Nµ (ε)(ε − εν )n dε, (4.11)
and therefore the charge density can be express as a function of energy

moments
Xh i
4πρ(r) = m(0)
µ φ(ε µ , r) 2
+ 2m(1)
µ φ(ε µ , r) φ̇(ε µ , r) + m(2)
µ φ̇(ε µ , r) 2
(4.12)
µ
9. Go back to step 1 and repeat the iteration until self-consistency is reached.
4.3 Difficulties and Challenges
So far the self-consistent scheme discussed above has not been fully realized. The
first attempt to implement the scheme described above encounters the following
difficulties.
• Approximations have to be made to calculate the energy moments from

Matsubara Green’s function, since only information at a limited number of
Matsubara frequencies is known from QMC. A high accuracy is thus not
easy to obtain. We have worked out a procedure to overcome this problem,
but it is not fully clear if the accuracy reached so far is enough or not.
• The double counting correction discussed above (Sec. 4.2, point 5.) has
to be made in this context, but its exact form is not known. Different
approximations used to treat this problem can lead to quite different results.
One has to arrive at a better understanding of this problem before one can
perform reliable self-consistent calculations.
• To have a good convergence of the full self-consistent scheme, for some tran-
sition metal oxides, it seems necessary to include in the DMFT calculation
not only the transition metal d orbitals, but also the ligand p orbitals which
strongly hybridize with the former. This by itself is a quite challenging prob-
lem. We have attempted to do this for NiO, but the present calculations
have not produced the right physical scenario, namely, we couldn’t find a
Mott insulator solution when we treat the Ni 3d orbitals and O 2p orbitals
on the same footing within DMFT. Under the present theoretical frame-
work, the system is modelled as a composition of interacting electrons (Ni
3d) and noninteracting electrons (O 2p). Although the total electron num-
ber is an integer, the individual (both the interacting and noninteracting)
number is not an integer due the the hybridization between them. For the
moment it is not fully clear if the failure of getting an insulating solution is
due to the present framework, or the way of solving it.
4.3. Difficulties and Challenges 69
• To get a full energy spectrum, including both Ni 3d and O 2p states, by

maximum entropy method is quite problematic, due to the reason that a
large energy range is involved in this case. This makes the problem more
difficult, and it is highly desirable to find out a way to improve the MEM
performance in future.
Due to the above difficulties and problems, a full implementation of the con-
ceptually appealing scheme is quite challenging, and the studies are still going
on.
71
5. SUMMARY AND OUTLOOK
In this thesis the LDA+DMFT approach, formulated in recent years as an ab

initio tool to study the strongly correlated materials, was reviewed. The two
basic ingredients of this approach, namely the density-functional based band-
structure calculation, and the dynamical mean-field theory for solving the many-
body Hamiltonian were discussed in chapter 1 and chapter 2 respectively. In
chapter 1 we also gave an account of the LMTO method particularly suitable for
solving the energy band problem of transition metal compounds. And in chapter 2
a description of the QMC method as the solver of the quantum impurity problem
was presented. These two chapters consist in the methodological part of this
thesis.
In chapter 3 we applied the LDA+DMFT approach to NiO which is a clas-
sic system studied for many years. In this work a new implementation of the
LDA+DMFT scheme is employed in which a set of Wannier functions (WFs)
are calculated from the band structure and used as the basis of the material-
specific many-body Hamiltonian. The results obtained within this approach are Concorda com
in good agreement with experiment concerning the insulating gap, local magnetic resultados
moment and energy spectrum. The most successful aspect of the LDA+DMFT experimentais
approach in this context is that it can reproduce the photoemission (and inverse
photoemission) spectrum of NiO at a quantitative level, which is something other
theoretical approaches are not able to do. On the other hand, as already pointed
out in the conclusion of chapter 3, the present treatment of NiO is not perfect,
in the sense that the Oxygen 2p-dominated bands are not included in the DMFT
calculation, but it is known that the O 2p bands play an important role in this
system, through its hybridization with Ni 3d bands. Therefore, a complete treat-
ment of NiO should allow the hybridization between the Ni 3d and O 2p electrons
and the strong Coulomb interaction among Ni 3d electron interplaying with each
other. With an extension of its present scheme by including the O 2p bands in the
DMFT treatment, the LDA+DMFT is a promising method to give a full solution
of the NiO problem, considering its present success in producing the experimental
photoemission spectrum.
A self-consistent LDA+DMFT scheme was presented in chapter 4. A new feature
of this scheme is that the interface between the LDA band-structure calculation
and the DMFT calculation is provided by a set of WFs. This set of WFs can
be constructed via a transformation from the LMTO basis set. By utilizing the
72 5. Summary and Outlook
transformation coefficients relating the LMTOs to the WFs, the feedback from
DMFT to LDA can be performed in a well-defined way So far this proposed
scheme has not been fully implemented, and there are some remaining difficulties
to overcome before one can really close the self-consistent loop.
Finally we would like to give an outlook for future studies. In the first place,
we hope to extend the conventional LDA+DMFT scheme so that the transition
metal valence d orbitals and the ligand p orbitals can be treated on the same
footing within DMFT. The solution of this problem will also help to close the
self-consistent LDA+DMFT loop, the realization of which will be the next aim.
In addition, it is also interesting to apply the present scheme to the neighbors
of NiO, such as MnO, FeO to see if it is possible to understand the similari-
ties and differences among these isostructured materials within the LDA+DMFT
approach.
APPENDIX
75
A. PROOF THAT THE EXTERNAL

POTENTIAL IS A UNIQUE
FUNCTIONAL OF THE
GROUND-STATE DENSITY
The first Hohenberg-Kohn theorem states that there is a one-to-one mapping be-
tween a many-body interacting system and its ground-state density, in another
word, the external potential v(r) (which specifies the system) can be uniquely de-
termined (up to an unimportant constant) by the ground-state density ρ(r). The
following proof, first given by Hohenberg and Kohn (1964), is valid for systems
with nondegenerate ground states. It proceeds by reductio ad absurdum (proof
by contradiction).
Suppose two different external potential v(r) and v 0 (r), with ground-state wave-
functions ψ and ψ 0 respectively, lead to the same ground state density ρ(r). Now,
it is clear that unless v(r) − v 0 (r) =constant, ψ 6= ψ 0 since they satisfy different
Schrödinger equations. Let’s denote the Hamiltonians and ground-state energies
corresponding to ψ and ψ 0 by H, H 0 and E, E 0 respectively. By using the minimal
property of the ground-state energy, we have,
E = hψ|H|ψi < hψ 0 |H|ψ 0i = hψ 0 |H 0 + V − V 0 |ψ 0 i,
P P
where V = i v(ri ) and V 0 = i v 0 (ri ). Since hψ 0 |H 0 |ψ 0 i = E 0 , this leads to
Z
E < E + ρ(r)(v(r) − v 0 (r))dr.
0
(A.1)
By interchanging the primed and unprimed quantities, we then find in the same
way that Z
E < E + ρ(r)(v 0 (r) − v(r))dr.
0
(A.2)
Adding (A.1) and (A.2) gives rise to

E0 + E < E + E0 (A.3)
which is of course incorrect. Thus the two different external potentials can not
generate the same ground state density, and therefore, up to a constant, the
external potential is a unique functional of ρ(r).
76 A. Proof That The External Potential Is A Unique Functional of The Ground-State Density
77
B. ESTIMATION OF THE LOCAL

MAGNETIC MOMENT AND THE
COULOMB INTERACTION
PARAMETER
As discussed in the context of the LDA+DMFT study of NiO in Sec. 3.2.2, the
average Coulomb interaction parameter within the WF basis Ū WF can be esti-
mated from that within the LMTO basis Ū LMTO by Ū WF = Ū LMTO (1−x)2 , where
x is the admixture of the Oxygen 2p states into the the WFs. Similarly in Sec.
3.2.3, the local Magnetic moment mloc can be estimated from that associated
with a set of WFs mWF (which should be an integer) by mloc = mWF (1 − x).
These two problems are related, and both can be understood in the following
way1 .
For simplicity we consider a single Wannier function which is composed of a pure
transition metal d atomic-like function (e.g. LMTO) located at the origin and a
pure ligand p function centered around some fixed position d,
ψ WF (r) = aφd (r) + bφp (r − d), (B.1)
where |a| ≈ 1 and |b| 1. The normalization condition of ψ WF (r) requires that
Z
ψ WF∗ (r)ψ WF (r)dr = |a|2 + a∗ bS + ab∗ S ∗ + |b|2 = 1, (B.2)
R
where S = φ∗d (r)φp (r − d)dr is the overlap between the d, p orbitals and the
normalization of φd (r) and φp (r) are implicitly assumed.
The electron density associated with ψ WF (r) is
ρWF (r) = ψ WF∗ (r)ψ WF (r)

= |a|2 ρd (r) + a∗ bφ∗d (r)φp (r − d) + ab∗ φd (r)φ∗p (r − d) + |b|2 ρp (r − d)
(B.3)
1
I would like to thank Professor Anisimov for making this point clear through private
communication
78 B. Estimation of The Local Magnetic Moment and The Coulomb Interaction Parameter
In (B.3) the first term ρd (r) gives a density confined to the atomic sphere of
the transition metal ion, the overlap (second and third) terms are also confined
within the immediate vicinity of the transition metal atomic sphere, but the last
term φp (r − d) represents the p electron density confined to the ligand sphere
and does not contribute to the local moment mloc associated with the magnetic
ion. Based on these considerations, we have
Z
2
mloc = |a| ρd (r) + a∗ bφ∗d (r)φp (r − d) + ab∗ φd (r)φ∗p (r − d) dr
Ω
≈ |a|2 + a∗ bS + ab∗ S ∗ = 1 − |b|2 , (B.4)
where Ω represents the atomic sphere of the transition metal magnetic ion.
Concerning the Coulomb interaction parameter U, it is defined with respect to
some particular basis. For the WF basis,
Z WF
ρ (r)ρWF (r0 )
U WF
= 0
drdr0 (B.5)
|r − r |
and for the atomic-like basis,

Z
ρd (r)ρd (r0 )
Udd = 0
drdr0 (B.6)
|r − r |
Using Eqn. (B.3), (B.5) and (B.6), and assuming only the d − d electron inter-
action is important, we have,
U WF ≈ |a|4 Udd = (1 − a∗ bS − ab∗ S ∗ − |b|2 )2 )Udd (B.7)
For the orthogonal (or nearly orthogonal) LMTO basis set, S → 0, therefore

U WF ≈ (1 − |b|2 )2 U LMTO (B.8)
Eqn. (B.4) and (B.8) can be generalized to a set of WFs with magnetic moment
mWF , and the Oxygen contribution x = |b|2 , and we can get
mloc ≈ (1 − x)mWF , (B.9)

ŪWF ≈ (1 − x)2 Ū LMTO . (B.10)
These are the relationship used in this work for studying NiO (here We have used
the average Coulomb interaction parameter Ū for multiorbital case).
79
C. ENERGY MOMENT
CALCULATION FROM
MATSUBARA GREEN FUNCTION
Matsubara Green function G(iωn + µ) can be expanded in terms of a Laurant

series:
Z ∞
Aα,β ()d
Gα,β (iωn + µ) = (C.1)
−∞ iωn + µ −
(0) (1) (2) (3)
Mα,β Mα,β Mα,β Mα,β
= + + + (C.2)
iωn (iωn )2 (iωn )3 (iωn )4
where Z ∞
(n)
Mα,β = Aα,β ()( − µ)n d (C.3)
−∞
with n = 0, 1, 2, 3, ...
The energy moments m(n) are calculated by:
Z
(0) 1X 1 ∞ X 1
mα,β = Gα,β (iωn ) = Aα,β ()d
β iω β −∞ iωn
iωn + µ −
n
Z ∞
= Aα,β ()f ( − µ)d (C.4)
−∞
Z
(1) 1X 1 ∞ X iωn
mα,β = (iωn )Gα,β (iωn ) = Aα,β ()d
iωn + µ −
n
Z ∞
= Aα,β ()( − µ)f ( − µ)d (C.5)
−∞
Z
(2) 1X 2 1 ∞ X (iωn )2
mα,β = (iωn ) Gα,β (iωn ) = Aα,β ()d
iωn + µ −
n
Z ∞
= Aα,β ()( − µ)2 f ( − µ)d (C.6)
−∞
80 C. Energy Moment Calculation from Matsubara Green function
Here f ( − µ) is just the Fermi function:
1X 1 1
f ( − µ) = = β(−µ) (C.7)
β iω iωn + µ − e +1
n
In practice, we only have a finite number of Matsubara frequencies, it means that

on summing over iωn in (4), (5) and (6), one has to do some truncation at a
particular point n = nmax . For instance, the zeroth moment m(0) is calculated in
the following way:
(0) (0)
(0) 1 X Mα,β Mα,β (0)
mα,β = (Gα,β (iωn ) − )+ + ∆mα,β (C.8)
β iωn 2
|ωn |<=ωnmax
with
(1) (2) (3)
(0)
X Mα,β Mα,β Mα,β
∆mα,β = 2
+ 3
+ + ...,
(iωn ) (iωn ) (iωn )4
|ωn |>ωnmax
(1)
X 1 (3)
X 1
= −2Mα,β + 2Mα,β + ... (C.9)
ωn >ωnmax
(ωn )2 ω >ω
(ωn )4
n nmax
The minimal nmax should be chosen so that the series in (9) converges sufficiently
(0) (n)
fast, and hence the evaluation of ∆mα,β in terms of the coefficients Mα,β rather
(n)
than Gα,β (iωn ) themselves is doable. The estimation of Mα,β is relatively easy
because for that purpose one only needs a few Gα,β (iωn ) at high frequencies.
Similarly:
(1) (1)
(1) 1 X (0) Mα,β Mα,β (1)
mα,β = [(iωn )Gα,β (iωn ) − Mα,β − ]+ + ∆mα,β (C.10)
β iωn 2
|ωn |<=ωnmax
with
(2) (3) (4)
(1)
∆mα,β = + + + ...,
(iωn )2 (iωn )3 (iωn )4
|ωn |>ωnmax
(2)
X 1 (4)
X 1
= −2Mα,β 2
+ 2Mα,β + ... (C.11)
ω >ω
(ωn ) ω >ω
(ωn )4
n nmax n nmax
(2) (2)
(2) 1 X (0) (1) Mα,β Mα,β (2)
2
mα,β = [(iωn ) Gα,β (iωn )−(iωn )Mα,β −Mα,β − ]+ +∆mα,β
β iωn 2
|ωn |<=ωnmax
(C.12)
81
with
(3) (4) (5)
(2)
∆mα,β = 2
+ 3
+ + ...,
(iωn ) (iωn ) (iωn )4
|ωn |>ωnmax
(3)
X 1 (5)
X 1
= −2Mα,β 2
+ 2Mα,β 4
+ ... (C.13)
ω >ω
(ω n ) ω >ω
(ω n )
n nmax n nmax
By this procedure, the energy moments can be calculated with the minimal set
of Matsubara Green functions, but with arbitrarily good accuracy, only if the
M (n) ’s are evaluated exactly.
For SrVO3, nmax ≈ 1000, and for NiO, nmax ≈ 500
It is possible to extrapolate the number of Matsubara frequencies to this order

from QMC data, but the accuracy is still to be checked.
82 C. Energy Moment Calculation from Matsubara Green function
BIBLIOGRAPHY 83
BIBLIOGRAPHY
Alperin, H. A.: 1962, J. Phys. Soc. Jpn. Suppl. B 17, 12

Andersen, O. K.: 1975, Phys. Rev. B 12, 3060
Andersen, O. K. and Jepsen, O.: 1984, Phys. Rev. Lett. 53, 2571
Andersen, O. K., Pawlowska, Z., and Jepsen, O.: 1986, Phys. Rev. B 34, 5253
Andersen, O. K. and Saha-Dasgupta, T.: 2000, Phys. Rev. B 62, R16 219
Andersen, O. K. and Wooley, R. G.: 1973, Mol. Phys. 26, 905
Anderson, P. W.: 1959, Phys. Rev. 115, 2
Anderson, P. W.: 1961, Phys. Rev. 124, 41
Anisimov, V. I., Aryasetiawan, F., and Lichtenstein, A. I.: 1997a, J. Phys. Cond.
Matter 9, 767
Anisimov, V. I., Kondakov, D. E., Kozhevnikov, A. V., Nekrasov, I. A., Pchelkina,
Z. V., Allen, J. W., Mo, S.-K., Kim, H.-D., Metcalf, P., Suga, S., Sekiyama,
A., Keller, G., Leonov, I., Ren, X., and Vollhardt, D.: 2005, Phys. Rev. B 71,
125119
Anisimov, V. I., Kuiper, P., and Nordgren, J.: 1994, Phys. Rev. B 50, 8257
Anisimov, V. I., Poteryaev, A. I., Korotin, M. A., Anokhin, A. O., and Kotliar,
G.: 1997b, J. Phys. Cond. Matter 9, 7359
Anisimov, V. I., Solovyev, I. V., Korotin, M. A., Czyżyk, M. T., and Sawatzky,
G. A.: 1993, Phys. Rev. B 48, 16929
Anisimov, V. I., Zaanen, J., and Andersen, O. K.: 1991, Phys. Rev. B 44, 943
Arita, R. and Held, K.: 2005, Orbital-selective Mott-Hubbard transition in the
two-band Hubbard model, cond-mat/0504040
Aryasetiawan, F. and Gunnarsson, O.: 1995, Phys. Rev. Lett. 65, 3221
Barnes, S. E.: 1976, J. Phys. F 6, 1375
Barnes, S. E.: 1977, J. Phys. F 7, 2637
Ōno, Bulla, R., and Hewson: 2001, Eur. Phys. J. B 19, 375
Bartel, L. C. and Morosin, B.: 1971, Phys. Rev. B 3, 1039
Bengone, O., Alouani, M., Blöchl, P., and Hugel, J.: 2000, Phys. Rev. B 62, 16
392
Bickers, N. E., Scalapino, D. J., and White, S. R.: 1989, Phys. Rev. Lett. 62, 961
Biermann, S., Aryasetiawan, F., and Georges, A.: 2003, Phys. Rev. Lett. 90,
086402
Blankenbecler, R., Scalapino, D. J., and Sugar, R. L.: 1981, Phys. Rev. D 24,
2278
84 BIBLIOGRAPHY
Blount, R. I.: 1962, Solid State Physics, Vol. 13, Academic, New York
Blümer, N.: 2002, Ph.D. thesis, Universität Augsburg
Born, M. and Oppenheimer, J. R.: 1927, Ann. Physik 84, 457
Brandow, B.: 1977, Adv. Phys. 26, 651
Brandt, U. and Mielsch, C.: 1989, Z. Phys. 75, 265
Brinkman, W. F. and Rice, T. M.: 1970, Phys. Rev. B 2, 4302
Bulla, R.: 1999, Phys. Rev. Lett. 83, 136
Bulla, R., Hewson, A. C., and Pruschke, T.: 1998, J. Phys. Cond. Matter 10,
8365
Caffarel, M. and Krauth, W.: 1994, Phys. Rev. Lett. 72, 1545
Capelle, K.: 2003, A bird’s-eye view of density-functional theory, VIII’th Brazil-
ian Summer School on Electronic Structure, p. 1, cond-mat/0211443
Ceperley, D. M. and Alder, B. J.: 1980, Phys. Rev. Lett. 45, 566
Cheetham, A. K. and Hope, D. A. O.: 1983, Phys. Rev. B 27, 6964
Cloizeaux, J. D.: 1964, Phys. Rev. 135, A698
de’ Medici, L., Georges, A., Kotliar, G., and Biermann, S.: 2005, Phys. Rev.
Lett. 95, 066402
Des Cloizeaux, J.: 1963, Phys. Rev. 129, 554
Des Cloizeaux, J.: 1964, Phys. Rev. 135, A685
Dominguez Rodriguez, A., Castaing, J., and Marquez, R.: 1984, Basic Properties
of Binary Oxides, Univ. of Seville Press
Dreizler, R. M. and Gross, E. K. U.: 1990, Density Functional Theory, Springer,
Berlin
Eder, R., Dormeich, A., and Winter, H.: 2005, Phys. Rev. B 71, 045105
Faleev, S. V., van Schilfgaarde, M., and Kotani, T.: 2004, Phys. Rev. Lett. 93,
126406
Fender, B. E. F., Jacobson, A. J., and Wegwood, F. A.: 1968, J. Chem. Phys
48, 990
Fermi, E.: 1928, Z. Phys. 48, 73
Fock, V.: 1930, Z. Phys. 61, 126
Föex, M.: 1948, Compt. Rend. Acad. Sci. 227, 193
Fujimori, A. and Minami, F.: 1984, Phys. Rev. B 30, 957
Fujimori, A., Minami, F., and Sugano, S.: 1984, Phys. Rev. B 29, 5225
Georges, A. and Kotliar, G.: 1992, Phys. Rev. B 45, 6479
Georges, A., Kotliar, G., and Krauth, W.: 1993, Z. Phys. B 92, 313
Georges, A., Kotliar, G., Krauth, W., and Rozenberg, M. J.: 1996, Rev. Mod.
Phys. 68, 13
Georges, A., Kotliar, G., and Si, Q.: 1992, Int. J. Mod. Phys. B 6, 705
Georges, A. and Krauth, W.: 1992, Phys. Rev. Lett. 69, 1240
Griffith, J. S.: 1961, The Theory of Thansition-Metal Ions, Cambridge University,
Cambridge
BIBLIOGRAPHY 85
Gunnarsson, O., Andersen, O. K., Jepsen, O., and Zaanen, J.: 1989, Phys. Rev.
B 39, 1708
Gutzwiller, M. C.: 1963, Phys. Rev. Lett. 10, 159
Hartree, D. R.: 1928, Proc. Cambridge Philos. Soc. 24, 89
Held, K., Keller, G., Eyert, V., Anisimov, V. I., and Vollhardt, D.: 2001a, Phys.
Rev. Lett. 86, 5345
Held, K., McMahan, A. K., and Scalettar, R. T.: 2001b, Phys. Rev. Lett. 87,
276404
Held, K., Nekrasov, I. A., Keller, G., Eyert, V., Blümer, N., McMa-
han, A. K., Scalettar, R. T., Pruschke, T., Anisimov, V. I., and Voll-
hardt, D.: 2003, Realistic investigations of correlated electron systems
with LDA+DMFT, in Psi-k Newsletter, Vol. #56, p. 65, http://psi-
k.dl.ac.uk/newsletters/News_56/Highlight_56.pdf
Herring, C.: 1940, Phys. Rev. 57, 1169
Hirsch, J. E.: 1983, Phys. Rev. B 28, 4059
Hirsch, J. E.: 1985, Phys. Rev. B 31, 4403
Hirsch, J. E. and Fye, R. M.: 1986, Phys. Rev. Lett. 56, 2521
Hohenberg, P. and Kohn, W.: 1964, Phys. Rev. 136, B864
Hubbard, J.: 1963, Proc. Roy. Soc. London A 276, 238
Hubbard, J.: 1964, Proc. Roy. Soc. London A 281, 401
Hüfner, S., Osterwalder, J., Riesterer, T., and Hulliger, F.: 1984, Solid State
Commun. 52, 793
Hüfner, S., Reinert, P. S. I. S. F., and Schmitt, H.: 1992, Z. Phys. B 86, 207
Hugel, J. and Kamal, M.: 1997, J. Phys. Cond. Matter 9, 647
Jarrell, M.: 1992, Phys. Rev. Lett. 69, 168
Jarrell, M.: 1993a, Z. Phys. B 90, 187
Jarrell, M.: 1993b, Phys. Rev. B 90, 187
Jarrell, M. and Gubernatis, J. E.: 1996, Physics Reports 269, 133
Jauch, W. and Reehuis, M.: 2004, Phys. Rev. B 70, 195121
Jones, R. O. and Gunnarsson, O.: 1989, Rev. Mod. Phys. 61, 689
Kajueter, H. and Kotliar, G.: 1996, Phys. Rev. Lett. 77, 131
Keller, G.: 2005, Ph.D. thesis, Universität Augsburg
Keller, G., Held, K., Eyert, V., Anisimov, V. I., and Vollhardt, D.: 2004, Phys.
Rev. B 70, 205116
Kohn, W.: 1959, Phys. Rev. 115, 809
Kohn, W.: 1973, Phys. Rev. B 7, 4388
Kohn, W. and Rostoker, N.: 1954, Phys. Rev. 94, 1111
Kohn, W. and Sham, L. J.: 1965, Phys. Rev. 140, A1133
Korringa, J.: 1947, Physica 13, 392
Koster, G. F.: 1953, Phys. Rev. 89, 67
Kotliar, G. and Ruckenstein, A. E.: 1986, Phys. Rev. Lett. 57, 1362
Kotliar, G. and Vollhardt, D.: 2004, Phys. Today 57(3), 53
Ku, W., Rosner, H., Pickett, W. E., and Scalettar, R. T.: 2002, Phys. Rev. Lett.
86 BIBLIOGRAPHY
89, 167204
Kuhlenbeck, H., Odörfer, G., Jaeger, R., Illing, G., M. Menges, T. M., Freund, H.-
J., Pöhlchen, M., Staemmler, V., Witzel, S., Scharfschwerdt, C., Wennemann,
K., Liedtke, T., and Neumann, M.: 19911991, Phys. Rev. B 43, 1969
Levy, M.: 1979, Proc. Natl. Acad. Sci. (USA) 76, 6062
Li, J.-L., Rignanese, G.-M., and Louie, S. G.: 2005, Phys. Rev. B 71, 193102
Lichtenstein, A. I. and Katsnelson, M. I.: 1998, Phys. Rev. B 57, 6884
Lichtenstein, A. I., Katsnelson, M. I., and Kotliar, G.: 2001, Phys. Rev. Lett. 87,
067205
Lichtenstein, A. I., Katsnelson, M. I., and Kotliar, G.: 2003, Spectral density
functional approach to electronic correlations and magnetism in crystals, in A.
Gonis, N. Kioussis, and M. Ciftan (eds.), Electron Correlations and Materials
Properties 2, Chapt. Phenomological Studies of Correlation Effects, Kluwer,
New York, cond-mat/0211076
Lieb, E. H.: 1983, Int. J. Quantum. Chem. 24, 243
Lieb, E. H. and Wu, F. Y.: 1968, Phys. Rev. Lett. 20, 1445
Lundqvist, S. and March, N. H.: 1983, Theory of the Inhomogeneous Electron
Gas, Plenum, New York
Mahan, G. D.: 1990, Many-Partilce Physcis, Plenum Press, New York and
London
Manghi, F., Calandra, C., and Ossichi, S.: 1994, Phys. Rev. Lett. 73, 3129
Marzari, N. and Vanderbilt, D.: 1997, Phys. Rev. B 56, 12 847
Massidda, S., Continenza, A., Posternak, M., and Baldereschi, A.: 1997, Phys.
Rev. B 55, 13494
Mattheiss, L. F.: 1972, Phys. Rev. B 5, 290
McMahan, A. K., Held, K., and Scalettar, R. T.: 2003, Phys. Rev. B 67, 75108
Metzner, W.: 1989, Z. Phys. B77, 253
Metzner, W. and Vollhardt, D.: 1989a, Phys. Rev. B 39, 4462
Metzner, W. and Vollhardt, D.: 1989b, Phys. Rev. Lett. 62, 324
Mott, N. F.: 1949, Phys. Soc. London Sect. A 62, 416
Müller-Hartmann, E.: 1989a, Int. J. Mod. Phys. B 3, 2169
Müller-Hartmann, E.: 1989b, Z. Phys. B 74, 507
Nagaoka, Y.: 1966, Phys. Rev. 147, 392
Nekrasov, I. A., Held, K., Blümer, N., Poteryaev, A. I., Anisimov, V. I., and
Vollhardt, D.: 2000, Eur. Phys. J. B 18, 55
Nekrasov, I. A., Keller, G., Kondakov, D. E., Kozhevnikov, A. V., Pruschke,
T., Held, K., Vollhardt, D., and Anisimov, V. I.: 2005, Comparative study
of correlation effects in CaVO3 and SrVO3 , cond-mat/0501240, accepted for
Phys. Rev. B
Nekrasov, I. A., Pchelkina, Z. V., Keller, G., Pruschke, T., Held, K., Krimmel,
A., Vollhardt, D., and Anisimov, V. I.: 2003, Phys. Rev. B 67, 085111
Parr, R. G. and Yang, W.: 1989, Density-Functional Theory of Atoms and
Molecules, Oxford University Press, Oxford
BIBLIOGRAPHY 87
Parzen, G.: 1953, Phys. Rev. 89, 237

Pavarini, E., Biermann, S., Poteryaev, A., Lichtenstein, A. I., Georges, A., and
Andersen, O. K.: 2004, Phys. Rev. Lett. 92, 176403
Perdew, J. P. . and Wang, Y.: 1986, Phys. Rev. B 33, 8800
Pruschke, T., Cox, D. L., and Jarrell, M.: 1993a, Europhys. Lett 21, 593
Pruschke, T., Cox, D. L., and Jarrell, M.: 1993b, Phys. Rev. B 47, 3553
Roth, W. L.: 1958a, Phys. Rev. 110, 1333
Roth, W. L.: 1958b, Phys. Rev. 111, 772
Rozenberg, M. J., Chitra, R., and Kotliar, G.: 1999, Phys. Rev. Lett. 83, 3498
Rozenberg, M. J., Zhang, X. Y., and Kotliar, G.: 1992, Phys. Rev. Lett. 69, 1236
Sakai, O. and Kuramoto, Y.: 1994, Solid State Commun. 89, 307
Sakai, S., Arita, R., and Aoki, H.: 2004, Phys. Rev. B 70, 172504
Savrasov, S. Y. and Kotliar, G.: 2001, Spectral Density Functionals for Electronic
Structure Calculations, cond-mat/0106308
Savrasov, S. Y., Kotliar, G., and Abrahams, E.: 2001, Nature 410, 793
Sawatzky, G. A. and Allen, J. W.: 1984, Phys. Rev. Lett. 53, 2339
Schuler, T. M., Ederer, D. L., Itza-Ortiz, S., Woods, G. T., Callcott, T. A., and
Woicik, J. C.: 2005, Phys. Rev. B 71, 115113
Sekiyama, A., Fujiwara, H., Imada, S., Suga, S., Eisaki, H., Uchida, S. I., Takega-
hara, K., Harima, H., Saitoh, Y., Nekrasov, I. A., Keller, G., Kondakov, D. E.,
Kozhevnikov, A. V., Pruschke, T., Held, K., Vollhardt, D., and Anisimov,
V. I.: 2004, Phys. Rev. Lett. 93, 156402
Shen, Z.-X., List, R. S., Dessau, D. S., Wells, B. O., Jepsen, O., Arko, A. J.,
Barttlet, R., Shih, C. K., Parmigiani, F., Huang, J. C., and Lindberg, P.
A. P.: 1991, Phys. Rev. B 44, 3604
Shick, A. B., Liechtenstein, A. I., and Pickett, W. E.: 1999, Phys. Rev. B 60, 10
763
Shimizu, Y., Sakai, O., and Hewson, A. C.: 2000, J. Phys. Soc. Jap. 69, 1777
Si, Q., Rozenberg, M. J., Kotliar, G., and Ruckenstein, A. E.: 1994, Phys. Rev.
Lett. 72, 2761
Skriver, H. L.: 1984, The LMTO Method, Springer-Verlag, Berlin Heidelberg
New York Tokyo
Skull, C. G., Strauser, W. A., and Wollan, E. O.: 1951, Phys. Rev. 83, 333
Slater, J. C.: 1937, Phys. Rev. 51, 846
Slater, J. C.: 1951, Phys. Rev. 82, 538
Sugano, S., Tanabe, Y., and Kamimura, H.: 1970, Multiplets of Transition-Metal
Ions in Crystals, Academic, New York
Svane, A. and Gunnarsson, O.: 1991, Phys. Rev. Lett. 65, 1148
Terakura, K., Oguchi, T., Williams, A. R., and Kübler, J.: 1984a, Phys. Rev. B
30, 4734
Terakura, K., Williams, A. R., Oguchi, T., and Kübler, J.: 1984b, Phys. Rev.
Lett. 52, 1830
Thomas, L. H.: 1927, Proc. Cambridge Philos. Soc. 23, 542
88 BIBLIOGRAPHY
Tjernberg, O., Söderholm, S., Chiaia, G., Girard, R., Karlsson, U. O., Nylén, H.,
and Lindau, I.: 1996, Phys. Rev. B 45, 10 245
Tomlinson, J. R., Domash, L., Hay, R. G., and Montgomery, C. W.: 1955, J.
Am. Chem. Soc. 77, 909
van Dongen, P. G. J., Gebhard, F., and Vollhardt, D.: 1989, Z. Phys. 76, 199
Vollhardt, D.: 1984, Rev. Mod. Phys. 56, 99
Vollhardt, D.: 1991, Physica B 169, 277
Vollhardt, D.: 1993, in V. J. Emery (ed.), Correlated Electrons Systems, World
Scientific, Singapore
Vollhardt, D., Held, K., Keller, G., Bulla, R., Pruschke, T., Nekrasov, I. A., and
Anisimov, V. I.: 2005, J. Phys. Soc. Jap. 74, 136
Vollhardt, D., Wölfle, P. ., and Anderson, P. W.: 1987, Phys. Rev. B 36, 006703
Wannier, G.: 1937, Phys. Rev. 52, 191
Wigner, E. P. and Seitz, F.: 1933, Phys. Rev. 43, 804
Wolff, P. A.: 1961, Phys. Rev. B 124, 1030
Yokoyama, H. and Shiba, H.: 1987, J. Phys. Soc. Jap. 56, 1490
Zaanen, J., Sawatzky, G. A., and Allen, J. W.: 1985, Phys. Rev. Lett. 55, 418
BIBLIOGRAPHY 89
CURRICULUM VITAE
Personal Data
Name Xinguo Ren

Date of Birth September 25, 1976
Place of Birth Henan, China
Nationality Chinese
Education and Studies

Ph.D in Physics Expected in Dec. 2005, University of Augsburg, Germany.
Advisor: Professor Dieter Vollhardt
Dissertation Topic: “LDA+DMFT Investigation of NiO"
M.S. in Physics 2001, Chinese Academy of Sciences, China

Advisor: Professor Yue Yu
Thesis: “The Methods of Hartree-Fock Approximation and Exact
Diagonalization in Two-Dimensional Electron System"
B.S. in Physics 1998, Nankai Univerisity, Tianjin, China

Advisor: Professor Xue-Qian Li
Thesis: “The Possible Source of Heavy Neutral Particles
in Cosmic Rays"
Research Interests
• ab initio electronic-structure calculations with density functional theory.

• Realistic modelling of strongly correlated systems with the LDA+DMFT
approach.
• Mott-Hubbard metal-insulator transition in electronic systems.
• Two-dimensional electronic systems.
90 BIBLIOGRAPHY
BIBLIOGRAPHY 91
ACKNOWLEDGEMENTS
In the first place I want to express my sincere gratitude to Prof. Dr. Dieter
Vollhardt for offering me an opportunity to work on challenging topics under his
supervision. It is his guidance that led me into a new research area of condensed
matter physics, namely, investigating the strongly correlated materials from first
principles. From numerous discussions with him, what I have learned is not only
the scientific knowledge, but also the way of thinking about physics, which is
invaluable for my future academic career.
My special thanks go to Dr. Georg Keller and Dr. Igor Nekrasov for their
generous helps when I started to learn the LDA+DMFT method. I also appreciate
that they have been always available for discussions and answering questions
during all these years.
I also want to thank Prof. Dr. Vladimir Anisimov, Prof. Dr. Thomas Pruschke,
Ivan Leonov, Dr. Karsten Held, Priv.-Doz. Dr Volker Eyert, and Prov.-Doz Ralf
Bulla for helpful discussions. They have helped to broaden my knowledge of the
LDA+DMFT approach and strongly correlated materials in one way or another.
Additionally, I would like to thank all the members of the groups of Prof. Dr.
Dieter Vollhardt, and Prof. Dr. Arno Kampf for the friendly and pleasant en-
vironment they have created. Among them are Priv.-Doz. Stefan Kehrein, Dr.
Marcus Kollar, Dr. Sabine Tornow, Dr. Sanjeev Kumar, Markus Schmid, Michael
Sekania, Hyun Jung Lee, Dmitry Lobaskin, Martin Eckstein, and Andreas Hackl.
Finally I am deeply indebted to my parents to whom I owe everything.

[D2] LDA + DMFT Investigation of NiO

Uploaded by

Copyright:

Available Formats

[D2] LDA + DMFT Investigation of NiO

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

[D2] LDA + DMFT Investigation of NiO

Uploaded by

Copyright:

Available Formats

LDA + DMFT Investigation of NiO

Von der Mathematisch-Naturwissenschaftlichen Fakultät

Tag der mündlichen Prüfung: 16.02.2006

1. Electronic-Structure Calculations with Density Functional Theory 5

2. The LDA+DMFT Approach . . . . . . . . . . . . . . . . . . . . . . 23

3. LDA+DMFT Investigation of NiO . . . . . . . . . . . . . . . . . . . 43

4. Towards A Self-Consistent LDA+DMFT Scheme . . . . . . . . . . 63

5. Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . 71

A. Proof That The External Potential Is A Unique Functional of The

B. Estimation of The Local Magnetic Moment and The Coulomb In-

C. Energy Moment Calculation from Matsubara Green function . . 79

From a microscopic ab-initio point of view, a solid material is an interacting

interaction is considered, whereas the long-range ones are neglected. However,

LDA+DMFT approach to investigate the electronic structure of NiO. NiO is a

The plan of this thesis is as follows. In chapter 1 we give an account of the

An introduction of the LDA+DMFT approach in general is then presented in

In chapter 3, we apply the LDA+DMFT approach to the prototypical Mott

Density-functional theory (DFT) is nowadays a popular and successful approach

1.1 Density Functional Theory

are the kinetic and electron-electron interaction operator respectively, and m

The original proof of the HK theorems is given in the space of V-representable

treated as independent particles and the interaction energy is approximated by

1.2.2 Local Density Approximation

As has been shown, KS theorems guarantee that the ground-state energy of a

xc (n(r)) = f (n(r), ∇n(r)). (1.17)

1.3 The Linear Muffin-Tin Orbital Method

1.3.1 Energy Band Methods

DFT through KS Ansatz offers a self-consistent way of calculating the band

Under this background a linear procedure was proposed by Andersen (1975) in

1.3.2 Linear Muffin-Tin Orbitals

ψL (E, r) = il Ylm (r̂)ψl (E, r). (1.25)

1.2 Single-Particle Description and Local Density

DFT is turned into a tractable framework through Kohn-Sham approach, or

1.2.1 Kohn-Sham Approach

KS further separate G[n] into Ts [n] and Exc [n],

G[n] = Ts [n] + Exc [n], (1.9)

δTs [n] δExc [n]

1.3.3 The LMTO Band Calculation

NhχkL |H − E|χkL0 i0 = hχL |H − E|χL0 i0

hχkL |H − E|χkL0 i = hχl |H − E|χl i0 δLL0

χkL → χkrL = χkrlm . (1.47)

1.4 Wannier Functions

branches at each k point, namely,

2. THE LDA+DMFT APPROACH

2.1 The Hubbard Model

The model referred to as Hubbard model became a standard framework for

(Gutzwiller, 1963; Hubbard, 1963; Hubbard, 1964). The model Hamiltonian

Siz = ni↑ − ni↓ ,

2.2 Dynamical Mean-Field Theory

2.2.1 The Infinite Dimension Limit

way comparing with the spin models, namely,

2.2.2 DMFT Equations — Exact Solution of the D = ∞

As observed by Müller-Hartmann (1989b), the irreducible self-energy Σii (ω),

imp (iωn ) (2.16)

Gimp (iωn ) = Gii (iωn ), Σimp (iωn ) = Σ(iωn ). (2.17)

On the other hand, the on-site lattice Green’s function is

numerical techniques such as QMC, ED, and Wilson numerical renormalization

2.2.3 Quantum Monte-Carlo Method

In the following we will give a discussion of the Hirsch-Fye algorithm, following

where λ = cosh−1 (exp(∆τ U/2)). From (2.22) and (2.23) we have

xc (n(r)) = f (n(r), ∇n(r)). (1.17)