Diss Phaase
Diss Phaase
Diss Phaase
Correlated Systems
Dissertation
zur Erlangung des mathematisch-naturwissenschaftlichen Doktorgrades
”Doctor rerum naturalium”
der Georg-August-Universität Göttingen
im Promotionsprogramm ProPhys
der Georg-August University School of Science (GAUSS)
vorgelegt von
Patrick Haase
aus Soltau
Göttingen, 2015
Betreuungsausschuss
Prof. Dr. Thomas Pruschke,
Institut für Theoretische Physik,
Georg-August-Universität Göttingen
1 Introduction 1
2 Models 5
2.1 Interacting Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 The Hubbard Model . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 The Falicov-Kimball Model . . . . . . . . . . . . . . . . . . . 8
2.2 Disordered Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 The Anderson Disorder Models . . . . . . . . . . . . . . . . . 11
2.2.2 Model for the Disorder Potential . . . . . . . . . . . . . . . . 12
2.2.3 Modeling Disorder in Macroscopic Systems . . . . . . . . . . . 13
2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
v
Contents
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Appendices 115
Bibliography 133
vi
1 Introduction
Solid state physics is situated at the boundary between the microscopic and macro-
scopic world. The classical sample of interest is a cube of dimension 1cm3 containing
6·1023 particles. This imaginary object is of tangible size, still many of its properties
are governed by microscopic laws of physics, namely quantum mechanics. Electrons
and atoms determine the properties of the sample, Coulomb interaction and Pauli
principle are the driving forces. Gravitation, paramount in the macroscopic world,
and weak and strong interaction, the forces of the subatomic world of the nuclei of
the atoms, are mostly irrelevant for the physics of the solid. Thus, the stage is set.
Quantum mechanics provides a clear description for every system composed of
electrons and atoms, in principle allowing for an exact solution of the problem. The
large number of particles, however, renders an exact solution impossible in almost
all cases. This is true for classical systems and even more so for quantum mechanical
systems, due to the exponential growth of the Hilbert space.
Two approaches exist to overcome this obstacle. The first is the simplification
of the Schrödinger equation. Instead of solving the Hamiltonian involving all de-
grees of freedom of the system, simplifications are introduced to reduce the number
of degrees of freedom. To this end, degrees of freedom and interactions that are
deemed irrelevant for the problem at hand are removed and, if necessary, are re-
placed by effective interactions. An example is the neglect of atomic motion, known
as the Born-Oppenheimer approximation [1], and the coupling of the electrons to
the atomic degrees of freedom. If the coupling of the electrons to the atoms cannot
be neglected, the concept of elementary lattice vibrations, so-called phonons, can be
introduced along with an effective interaction between electrons and phonons.
The second approach consists in the development of new methods that are partic-
ularly well suited for a given problem, exploiting symmetries and special properties
that are unique for a given situation. Symmetries can include translational and
rotational invariance. Special properties can be a small entanglement entropy in
one dimension, that is exploited for example in the density matrix renormalization
group [2]. Another example is nearly-free electron behavior in many metals, where
Fermi liquid theory [3, 4] can be applied successfully. Other approaches rely on
an expansion around a known solution in terms of a small parameter. These ap-
proaches are known as perturbation theory [5]. An approach can be analytical, i. e.
a pen-and-paper solution, purely numerical, e. g. a straight forward diagonalization
of the Hamiltonian by means of numerical algorithms, or a combination of both.
In general, it is not enough to use one of the two approaches above. A successful
1
1 Introduction
solution for a given problem often requires to find an effective, minimal model that
covers all relevant aspects of the problem and then find a method that is suitable for
solving the model. Usually, this requires some analytical work. This can lead to a
satisfactory solution of the problem or be the starting point for numerical methods.
In this thesis we are interested in the situation of strongly correlated electron
systems with disorder, where disorder refers to defects in the crystal and strongly
correlated means that electrons strongly repel each other via the Coulomb interac-
tion. Materials with strongly correlated electrons include transition-metal oxids [6]
and heavy fermions [7, 8, 9, 10, 11]. They are interesting because of their rich phase
diagram [12, 13] and are promising candidates for applications [14], e. g. smart
windows using VO2 [15].
Disorder is often neglected in effective models, but to a certain degree, disorder in
the form of impurities, vacancies, dislocations, etc. is present in all systems at finite
temperature. Hence, it is necessary, to study the effects of disorder and see how it
affects the physical properties. Strong disorder can even be the dominant factor,
e. g. in the event of a metal-insulator transition known as Anderson localization
[16]. Close to the Anderson transition from the metal to the Anderson localized
state the wave function takes on a multifractal nature [17, 18], highlighting the
complex behavior associated with disorder. Even in strongly correlated electron
systems, the impact of disorder on transport and thermodynamic properties is often
not negligible [19, 20].
An important model for disordered interacting systems is the disordered Hubbard
model [21, 22, 23], also known as the Anderson-Hubbard model [24]. It is a lattice
model for electrons that can hop from one site to another. The electrons can have
up or down spin and interact via a repulsive Coulomb interaction if two electrons
occupy the same site. Disorder is taken into account by a local random potential
that equally affects up and down electrons. In the model, electrons obey the Pauli
principle, i. e. only electrons of different spins can occupy the same site. There is
only one orbital in the simplest form of the model. This is an oversimplification that
is only justified, if one is interested in general properties of the Hamiltonian. For
applications to real materials the single orbital model is not suited.
The Anderson-Hubbard model is a seemingly simple model, but even the Hubbard
part alone, i. e. hopping and interaction, is hard to solve. Generally, quantum Monte
Carlo [25] can be used, but it only allows to treat small systems, far away from the
thermodynamic limit. Also, away from half-filling it suffers from the infamous sign
problem [25], which means that the computational cost scales exponentially with
system size and inverse temperature. This is believed to be a manifestation of the
exponential growth of the Hilbert space.
In 1989 Metzner and Vollhardt introduced dynamical mean-field theory (DMFT)
[26, 27, 28, 29] that allows to treat interacting systems in the thermodynamic limit
and is exact in infinite dimensions. It was extended in [30, 31] to treat disordered
2
interacting systems. DMFT treats local correlations exactly by mapping the lattice
problem to an impurity problem that is solved self-consistently, but DMFT ignores
non-local correlations altogether when employed for finite-dimensional systems. This
is an approximation away from the limit of infinite dimensions and it fails when
non-local correlations become important. Unlike quantum Monte Carlo for lattice
models, it does not suffer from the sign problem for a single orbital model.
Cluster extensions of DMFT are the dynamical cluster approximation (DCA)
[32, 33] and cellular DMFT [34]. For the Hubbard model both methods often use
quantum Monte Carlo to solve the cluster problem and thus indirectly suffer from
the sign problem.
The methods mentioned so far are either not working in the thermodynamic limit,
neglect non-local correlations, or suffer from the sign problem. A possible way out
are diagrammatic extensions of DMFT [35] that rely on an impurity problem to
capture the local physics and treat non-local correlations using perturbation theory.
Two such methods for purely interacting systems are the dual fermion approach
[36, 37] and the dynamical vertex approximation [38, 39]. Prior to this work, the
dual fermion approach has also been applied to purely disordered systems [40] as
well.
In this thesis, we want to combine the two dual fermion approaches for disordered
and interacting systems to a method that treats both disorder and interaction on
equal footing. We develop the formalism for the Anderson-Hubbard model and
Anderson-Falicov-Kimball model [41], which is a simplified version of the Anderson-
Hubbard model. Both models are introduced and discussed in chapter 2.
We are interested in the case of macroscopic systems. In this context a macro-
scopic system is a system that is much larger than the coherence length of the
electrons. In general, such systems are too large to be simulated directly. Thence,
we exploit the self-averaging property of electrons and use disorder-averaging. This
allows to model a macroscopic system by a set of smaller subsystems that can be sim-
ulated, and physical properties of the original system are calculated as the average of
the corresponding properties of the subsystems. A drawback of the disorder-average
is that it is not suited for an accurate treatment of Anderson localization, meaning
results around the critical disorder strength have to be taken with care.
In chapter 3 we introduce DMFT and DCA for disordered interacting systems.
DMFT is the method we want to improve and DCA, an established cluster method,
acts as a benchmark for our approach.
Chapter 4 is the main result of this thesis. We present a detailed discussion of the
dual fermion approach for disordered interacting systems. This includes a discussion
of the perturbative treatment of the dual single-particle Green function and the self-
consistency loop. The chapter ends with the calculation of vertex corrections for the
conductivity, which is used as an example to show how the dual fermion approach
can be used to calculate two-particle quantities.
3
1 Introduction
DMFT, DCA and the dual fermion approach require the solution of an impurity
or cluster problem. In all three cases quantum Monte Carlo can be used. The
interaction expansion algorithm [42, 43, 44] is discussed in chapter 5, along with
modifications needed for the inclusion of disorder.
Chapters 6 and 7 present results for the Anderson-Falicov-Kimball model and the
Anderson-Hubbard model. The Anderson-Falicov-Kimball model serves as a test
case, therefore we limit our studies to the 1d and 2d system. We compare the dual
fermion approach to DCA in 1d and obtain a phase diagram in 2d as a function of
interaction and disorder strength.
The discussion of the Anderson-Hubbard model starts with a comparison between
dual fermions and DCA in 1d, as well. After that, we turn to the 3d case and inves-
tigate phase transitions, namely the antiferromagnetic transition, Mott transition
and Anderson transition. Prominent 3d materials with interactions and disorder are
alloys like brass, bronze or cast iron.
Throughout this thesis we set ~ = kB = 1, where ~ is the reduced Planck constant
and kB is the Boltzmann constant. A quantity is regarded as finite if it is less than
infinity and non-zero.
4
2 Models
The purpose of this chapter is to introduce two models for disordered interacting
systems. It is organized as follows.
First, in section 2.1 some phenomena in interacting systems will be presented and
the Hamiltonian for the Hubbard model [21, 22, 23] will be introduced, as well as
the Hamiltonian for the Falicov-Kimball model [45]. The Hubbard model is a lattice
model that describes two types of electrons that repel each other if they occupy the
same site. The Falicov-Kimball model is similar, except only one electron species is
mobile, the other is localized.
Second, in section 2.2 concepts for disordered systems will be introduced along
with some phenomena caused by disorder. The Hamiltonians of the Anderson-
Hubbard model and Anderson-Falicov-Kimball model for disordered-interacting sys-
tems will be introduced along with a discussion of the disorder potential in sec-
tion 2.2.2. Section 2.2.3 concludes this chapter with a discussion of the self-averaging
property of electrons and we supplement the Hamiltonians given in section 2.2 with
the disorder-average as the final ingredient for the disorder models for macroscopic
systems.
5
2 Models
U/t Pressure
500
critical point
Temperature (K) 400
insulator
300 metal
200
100 antiferromagnetic
insulator
0
0.04 0.02 0 0.02 0.04 0.06
+Cr (V1-xMx)2O3 +Ti
Figure 2.1: The schematic phase diagram of V2 O3 based on the data of [13]. The phase
diagram shows three different phases that can be explained by the Hubbard
model. The metal for high temperatures and weak interactions is dominated
by the kinetic energy term in the Hamiltonian. Upon increasing the pressure
via doping, V2 O3 becomes a Mott insulator, dominated by the local Coulomb
repulsion. For low temperatures and sufficient pressure, the material enters an
antiferromagnetic state, caused by the interplay between both the kinetic and
potential energy terms in the Hamiltonian. The Coulomb repulsion prevents
double occupancies, but the antiferromagnetic order allows for virtual hopping,
which lowers the energy.
6
2.1 Interacting Systems
(†)
Here, i and j denote site indices, σ = ↑, ↓ is the spin index, cσi annihilates (creates)
an electron of spin σ at site i, nσi = c†σi cσi is the particle number operator for
electrons of spin σ at site i. U parametrizes the local Coulomb repulsion, µef f =
µ − U2 is the effective chemical potential and µ is the chemical potential. For half-
filling µef f is zero. tij,σ is the hopping amplitude for an electron with spin σ and
only depends on the difference |i − j| between sites i and j because a translationally
invariant lattice is assumed. In the Hubbard model the spin dependence of the
hopping amplitude is usually dropped. The Hamiltonian in eq. (2.1) neglects non-
local Coulomb interactions. This is a simplification that is justified if the screening of
the Coulomb interaction in the material is strong. In this simple form, the Hubbard
model has only one orbital.
Overall, the Hubbard model is an oversimplification if quantitative predictions for
7
2 Models
experiments are required. Although this Hamiltonian is very simple, too simple to
model real materials, no one has succeeded in finding a general analytic solution in
more than one [56] and less than ∞ [26] dimensions. Despite these two problems,
the Hubbard model is very popular. It was successfully applied to study the Mott
metal-insulator transition as well as itinerant magnetism, where it provides a simple
environment [57]. In these studies the simplicity of the model can be an advantage
as it allows to determine more easily the physical origin, hopping or local Coulomb
repulsion, of observed features. Compared to models with more parameters and
thus more possible origins this is a clear benefit.
Although no general analytic solution exists, a lot of progress in understanding
this model has been made [58]. Besides analytic treatment of the large and small
U limit, numerical methods shed some light on the underlying physics [59]. Such
methods include exact diagonalization (ED) [60], quantum Monte Carlo (QMC) [25],
dynamical mean-field theory [26, 27, 28, 29] and non-local extensions thereof. These
include the dynamical cluster approximation (DCA) [32, 33, 61], cellular DMFT
(CDMFT) [34] or the dual fermion approach [42, 43]. The dual fermion approach is
used and developed further in this thesis.
8
2.2 Disordered Systems
In this thesis we opted to use the Falicov-Kimball model as a test case. The
impurity problem that one has to solve in dynamical mean-field theory and the dual
fermion approach is particularly simple for the Falicov-Kimball model and thus
testing numerical algorithms for this model is fast.
k = k̃ + G, (2.3)
where the crystal momentum k̃ is a vector in the first Brillouin zone and here G is a
reciprocal lattice vector. This facilitates analytic as well as numerical calculations.
However, it completely ignores that no crystal is ever perfect, which follows from
entropy based arguments.
In this thesis, we want to go beyond the perfect crystal. Instead, we will take
into account that there will always exist a number of defects in the group of atoms
forming the lattice or crystal. A few possible deviations from the perfect crystal are
illustrated in fig. 2.2.
Whatever it is, the potential and hopping parameters around those defects will be
changed and depending on the nature of those defects and their concentration the
behavior of a sample may change as well. A famous example is the Kondo effect [53],
where a low concentration of magnetic impurities in a metal may cause an increase
in resistivity at low temperatures. Another effect of introducing defects in a crystal
is visible in electron diffraction. Here, the clear peaks in the diffraction pattern of
the clean sample begin to smear out when disorder starts to break the translational
invariance and long-range order of the crystal [68, 69]. Arguable the most prominent
influence of disorder is its diminishing effect on the conductivity which can localize
electrons entirely. This effect is known as Anderson localization [16] and it is caused
by coherent backscattering of electrons from impurities. If the backscattering does
not lead to localization but only to a reduction of the conductivity it is called weak
localization [70].
Before we move on, let us mention a few effects where disorder and interaction
act together. As was presented above, both interaction and disorder can lead to
localization and insulating behavior [16, 49]. For a long time it was believed that
the ground state of a 2d system is always insulating if one has disorder or interaction.
This does not hold if both are present [71, 72, 73]. In this case, a metallic ground
state is possible.
9
2 Models
Impurity atom
Interstitial impurity
Self-interstitial atom
Vacancy
Figure 2.2: Possible deviations from the perfect crystal. An atom can be replaced by an
impurity atom or be removed from its position and leave a vacancy. It is
also possible to have atoms that do not occupy a lattice site. This can be
an interstitial impurity as well as a self-interstitial atom, which is a displaced
atom from the regular lattice.
10
2.2 Disordered Systems
In the disordered system the hopping tij,σ depends on the two individual sites in-
volved and viσ is a random site-dependent potential, also called the disorder poten-
tial. It may be spin-dependent. Finally, the Hubbard interaction U could become a
site-dependent random variable as well.
These modifications make for a quite complex Hamiltonian. The explicit site-
dependence of the hopping is not very well suited for an approach in momentum
space. In this thesis it is assumed that the hopping and disorder potential do not
depend on spin, i. e. the energy of the impurities does not depend on whether an
electron with up or down spin is added. We will restrict ourselves to the disordered
potential and neglect the influence of the potential on the hopping. This is called
diagonal disorder. We obtain the Anderson-Hubbard Hamiltonian
1 1
(tij + µδij )(c†σi cσj + h.c.) −
X X X
HAH = − vi niσ + U (ni↑ − )(ni↓ − ) (2.5)
ij,σ i,σ i 2 2
11
2 Models
12
2.2 Disordered Systems
13
2 Models
{Z m } − 1
{ln Z} = lim (2.11)
m→0 m
allows to integrate out theR disorder analytically and thus restore pseudo-momentum
conservation. {. . . } = dvp(v) . . . denotes the disorder-average according to a
probability density p for the disorder. Equation (2.11) follows from
Z x
1
ln x = dt
1 t
Z x 1
= lim dt
m→0 1 t1−m
m
x −1
= lim . (2.12)
m→0 m
m is taken to be an integer and it is assumed that in the end the limit m → 0
can be performed. Within the replica trick one has to deal with m replicas of Z
instead of {ln Z}. This facilitates analytic calculations and is a standard trick used
for disordered systems as well as spin-glass systems [83].
14
2.2 Disordered Systems
where
XZ β XZ β
SRv c̄αωkσ (ω k )cαωkσ dτ vi nαi (τ ) dτ U nα↑i (τ )nα↓i (τ )
XX
=− +µ− − +
α ωkσ i,α 0 i,α 0
(2.14)
is the action for a given disorder configuration and Z v is the corresponding partition
function. α is a replica index. In the above, the replica limit introduced limm→0 −1 m
.
This terms is a constant with respect to derivatives of the partition function and thus
it can be ignored. As such derivatives are performed before the replica limit, there
will be no problem with 01 . Now, the disorder can be integrated out analytically
[84]: Z R P
dτ nα
dvi p(vi )evi α i (τ ) = eW (ñi ) (2.15)
with ∞
X 1 l l
W (ñi ) = hv ic ñ (2.16)
l=2 l! i i
and XZ
ñi = dτ nαi (τ ). (2.17)
α
hvil ic denotes the lth cumulant of the disorder distribution. This can be seen by
means of the cumulant-generating function which is defined as the logarithm of the
left-hand side of eq. (2.15). It follows that
∞
1 l l
Z
W 0 (ñ) = ln dvp(v)evñ =
X
hv ic ñ . (2.18)
l=1 l!
Finally, it is assumed that hv 1 ic = 0 such that the disorder does not change the filling.
If hv 1 ic 6= 0, it can be absorbed in the chemical potential. The effective interaction
W that originates from the disorder-average is local in space and translationally
invariant. This implies that any single-particle observable Okk0 ∝ δkk0 is diagonal in
momentum. In this sense, momentum conservation is restored.
15
2 Models
2.3 Summary
In this chapter the Hubbard model and Falicov-Kimball model were introduced.
Both models describe two types of interacting electrons on a lattice, the difference
between the two models is that both types of electrons are mobile in the Hubbard
model, but only one type is for the Falicov-Kimball model.
In section 2.2 disorder was incorporated in both models with a disorder potential
that acts on the mobile electrons. The resulting models are the Anderson-Hubbard
model and the Anderson-Falicov-Kimball model. Section 2.2.2 dwelled on the physi-
cal origin of the disorder potential and the section concluded with the two probability
distributions that are used in chapters 6 and 7 to model disorder.
We convinced ourselves that for microscopic systems the solution of the model
depends strongly on the disorder realization, therefore the solution of a microscopic
system is hardly suitable as an approximation for macroscopic systems. Hence, it
was necessary to add disorder-averaging in section 2.2.3 as an additional step in
modeling macroscopic disordered systems.
Using the replica trick, it was possible to perform the disorder-average by inte-
grating out the disorder term analytically in favor of an effective translationally
invariant interaction between different replicas. The disorder-average restored a
pseudo-momentum conservation such that Green functions and other quantities are
diagonal in momentum space. This simplifies the methods for disordered interacting
lattice problems that are introduced in the following two chapters.
16
3 Conventional Mean-Field Methods
for Correlated Lattice Models
The purpose of this thesis is to develop a method that treats disorder and interac-
tions on equal footing. To this end, we will first introduce the dynamical mean-field
theory (DMFT) [26, 27, 28, 29] to approximately handle systems with local inter-
actions. DMFT includes local quantum fluctuations exactly but ignores non-local
correlations altogether. This can lead to wrong result when non-local correlations
become important, e. g. DMFT predicts the wrong line shape for the metal-insulator
transition on the U T -plane in 2d [85]. Therefore, an important aspect of this thesis
is to also introduce and further develop a method that allows to incorporate these
non-local correlations.
In spite of its deficiencies DMFT is a useful method. An advantage is that it allows
for calculations at very low temperatures even if quantum Monte Carlo (QMC) is
used to solve the impurity problem. It does so by mapping the lattice problem to
an impurity problem which is numerically relatively cheap to solve. Thus, DMFT
allows for quick parameter scans to get a general idea of what to expect for a given
system. After such a parameter scan more elaborate methods can be used to study
the physics of the system in more detail, taking into account non-local correlations.
In section 3.1 the basic ideas of DMFT are introduced, i. e. the mapping of the
lattice problem to an impurity problem and the self-consistency loop. The discussion
of the self-consistency loop starts with the original loop of the clean system and is
then extended to include disorder [30, 31]. The section on DMFT ends with a brief
discussion of the so-called typical-medium-theory (TMT) [86]. TMT is designed to
distinguish localized states from extended states which enables it to detect Anderson
localization. In this thesis, the critical disorder strength from TMT is compared with
the dual fermion result to estimate how good the dual fermion approach is in the
strong disorder regime.
In section 3.2 the dynamical cluster approximation (DCA) [32, 33] is introduced.
It has been used for disordered, interacting and disordered interacting systems. DCA
includes short-range correlations by replacing the impurity of DMFT by a cluster
embedded in a momentum-dependent dynamical mean-field. Here, it will be shown
how to modify the corresponding self-consistency loop to include both disorder and
interactions [87, 88].
DCA is numerically much costlier than DMFT. If QMC is used to solve the
17
3 Conventional Mean-Field Methods for Correlated Lattice Models
Figure 3.1: In dynamical mean-field theory the lattice of interacting sites is mapped to
a single impurity in a non-interacting bath. Instead of solving the lattice
problem, an impurity problem is solved self-consistently.
cluster problem, DCA will also suffer from the infamous sign problem away from
half-filling. The sign problem is discussed in appendix A. As QMC is the standard
cluster solver for interacting quantum systems in DCA, the range of applicability
of DCA is limited. Nevertheless, DCA is a well established method that adds back
non-local correlations. Comparing the dual fermion approach with DCA will show
how good the dual fermion approach is compared to established cluster methods.
Thus, it will allow us to estimate how reliable the results will be when new regions
of the parameter space are explored.
18
3.1 Dynamical Mean-Field Theory
In the Hubbard model the hopping term is the only non-local term. DMFT treats
the hopping processes in terms of a mean-field function G0 (τ ) that depends on
imaginary time τ . The action describing the impurity problem is a functional of G0 .
It reads
ZZ
dτ dτ 0 c†oσ (τ )G0 (τ − τ 0 )−1 coσ (τ 0 )
X
Sef f = −
σ
Z β (3.1)
1 1
+U dτ no↑ (τ ) − no↓ (τ ) − .
0 2 2
Here, o labels the impurity site. In frequency space G0 can be calculated from the
impurity self-energy Σ = Σ[Sef f ] and the local lattice Green function Gloc = Gloc [Σ]
according to
G0 (iω)−1 = Gloc (iω)−1 + Σ(iω). (3.2)
The following definition of the single-particle Green function is used:
1 P 1
1. Calculate Gloc (iω) = N k iω+µ−k −Σ(iω) .
2. Calculate G−1 −1
0 (iω) = Gloc (iω) + Σ(iω).
3. Fourier transform G0 (iω) to obtain G0 (τ ). This step is only needed for quan-
tum Monte Carlo methods.
19
3 Conventional Mean-Field Methods for Correlated Lattice Models
1 P 1
Gloc (iωn ) = N k iωn +µ−ǫk −Σ(iωn )
Figure 3.2: DMFT loop for the self-consistent solution of the lattice problem.
4. Solve the impurity problem defined by the medium G0 and obtain Σnew (iω).
in real space for disorder realization n. G0 is the bath Green function for the
translationally invariant system. The self-consistency loop of DMFT is altered to
the following:
20
3.1 Dynamical Mean-Field Theory
1 P 1
1. Calculate Gloc (iω) = N k iω+µ−k −Σ(iω) .
2. Calculate G−1 −1
0 (iω) = Gloc (iω) + Σ(iω).
b) Solve the impurity problem defined by the medium Gv0n and obtain g vn (iω),
which is the impurity Green function for the disorder potential vn . Do
this for all disorder realizations.
1 PN
c) Calculate g = N n=1 g vn , where N is the number of disorder realizations.
21
3 Conventional Mean-Field Methods for Correlated Lattice Models
typically zero. This means that the average value of the LDOS for a given energy
does not reflect what one expects to measure. These considerations have been used
to develop the typical-medium-theory (TMT) [86]. It can be described as a DMFT
where the arithmetic disorder-average is replaced by a geometric disorder-average
for the LDOS ρ in the self-consistency according to
Z
ρtyp (ω) = exp dvn p(vn ) ln ρ(ω, vn ) . (3.5)
ρtyp is used in the Hilbert transform for the typical local lattice Green function
Z
ρtyp (ω 0 )
Gtyp (ω) = dω 0 , (3.6)
ω − ω0
which corresponds to step 1 in the DMFT self-consistency loop.
TMT can distinguish extended states from localized states, therefore it is capable
of detecting Anderson localization. The typical LDOS is used as an order parameter.
If the typical LDOS vanishes, the system enters the Anderson insulator. TMT is
surprisingly good at detecting Anderson localization but the estimate for the critical
interaction strength VA can be improved. This is done by extending the single-site
TMT approach to a cluster approach, namely cluster-TMT [91] or typical-medium
dynamical cluster approximation (TMDCA) [92].
Although TMT is very good at detecting Anderson localization, it is not the
method of choice in all cases where disorder is involved. In fact, it has a number
of severe drawbacks, i. e. it does not produce correct two-particle quantities and
without cluster-extensions it cannot capture true non-local physics. Additionally, it
is not capable of producing correct critical exponents [86]. The cluster extensions
require the solution of an interacting cluster, thus from a numerical point of view
they share the general problems of the dynamical cluster approximation that will
be introduced in the next section.
22
3.2 Dynamical Cluster Approximation for Disordered Interacting Systems
The dynamical cluster approximation (DCA) [32, 33] is an approximation that goes
beyond DMFT by including short range correlations within a finite cluster that
is used as a reference system. Similar to DMFT the cluster is embedded in a
momentum dependent mean-field, which is illustrated in fig. 3.3. The original lattice
coordinate x = x̃+X is written in terms of the cluster position x̃ and the position X
inside the cluster. The lattice momentum k = K + k̃ is written in terms of the cluster
momentum K and the momentum k̃ of the superlattice. This is illustrated in fig. 3.4.
Real space and momentum space are related by discrete Fourier transformations
Nc X ik̃x̃
f (X, x̃) = e f (X, k̃) (3.7)
N k̃
e−ik̃x̃ f (X, x̃)
X
f (X, k̃) = (3.8)
x̃
1 X i(K+k̃)X
f (X, k̃) = e f (K, k̃) (3.9)
Nc K
e−i(K+k̃)X f (X, k̃)
X
f (X, k̃) = (3.10)
X
for some function f . Nc is the number of cluster sites and N the number of lattice
sites. The self-energy Σ(iω, k) is replaced by the cluster self-energy Σc (iω, K) which
only depends on the cluster momentum K. For k̃ 6= 0 we set
such that |k − K| is minimal. The cluster self-energy is obtained from the Green
function Gc of the isolated cluster via
h i−1
Gc (iω) = (iω + µ)1 − tc − Σc (iω) . (3.12)
Bold symbols denote matrices in K-space. tc is the hopping matrix inside the cluster.
For the models considered, all quantities are diagonal in spin space.
In a diagrammatic treatment of the full lattice problem, momentum sums have
to be performed and momentum conservation is retained at each vertex. This is
enforced by the Laue function
0 0
eix(k1 +k2 +···−k1 −k2 −... ) = N δk1 +k2 +...;k10 +k20 +... .
X
∆= (3.13)
x
23
3 Conventional Mean-Field Methods for Correlated Lattice Models
Figure 3.3: The lattice problem is mapped to a cluster embedded in a mean field or bath.
In DCA momentum conservation is retained only up to the cluster momenta
K.
x2 k2
π
k k̃
... ... K
X −π π k1
x x̃
−π
1st Brillouin Zone
x1
Figure 3.4: The lattice is tiled and a superlattice is introduced. The coordinates of the
superlattice are denoted by x̃. The position inside the tiles or clusters are
denoted X. This leads to a tiling of the Brillouin zone. The coordinates of
the superlattice in k-space are K and the coordinates within the tiles are k̃.
Patches around equivalent K-points have the same color.
24
3.2 Dynamical Cluster Approximation for Disordered Interacting Systems
This means that the cluster approximation neglects the full momentum conservation.
For Nc → ∞ it is systematically restored and for Nc = 1 one obtains DMFT.
3.2. Calculate
Gc = {(G−1 −1
0 − V) } (3.18)
where {. . . } indicates the disorder average and V is the diagonal
disorder potential matrix in real space.
25
3 Conventional Mean-Field Methods for Correlated Lattice Models
Σnew
c = G−1 −1
0 − Gc . (3.19)
Now, we have to merge these two algorithms that differ in step 3. To this end, let
us understand what the main difference is. For the purely interacting problem all
quantities are diagonal in momentum space. For the disordered system, however, V
breaks the translational invariance of the cluster and therefore the Green function
Gconf
c for a single disorder configuration is no longer diagonal in momentum space.
It will only be diagonal after the disorder average {. . . } has been performed. The
step we have to modify is 3.2. for the disordered system, where the term in {. . . }
has to be replaced by the solution of the interacting disordered cluster. The effect
of disorder is to change G0 for the individual disorder configurations:
−1
(Gconf
0 ) = G−1
0 − V. (3.20)
We are now dealing with a Gconf 0 that is neither diagonal in real space nor in mo-
mentum space.
QMC (c. f. chapter 5) is used in this work to solve the disordered interacting
cluster problem. In the following we present in detail, how Gconf 0 (τ, i, j) is obtained.
0
Additionally, Gconf
0 (iω, K, K ) is needed to calculate Gconf
c (iω, K, K 0 ) as we will see
later.
Starting from G0 (iω, K) we do the following steps:
26
3.2 Dynamical Cluster Approximation for Disordered Interacting Systems
−1 −1
2. Then Gconf
0 (iω, i, j) = (G0 − V) (iω, i, j) is evaluated.
3. Next, a Fourier transform leads back to the reciprocal space Green function
0 1 X conf
Gconf
0 (iω, K, K ) = G (iω, i, j) exp[−i(Kxi − K 0 xj )]. (3.22)
Nc ij 0
1
The factor Nc
is needed to adjust the dimensionality in terms of Nc .
0
4. Then Fourier transforming Gconf
0 (iω, K, K ) to the imaginary time domain leads
0
to Gconf
0 (τ, K, K ).
With Gconf
0 (τ, i, j) one can calculate Σc using QMC to solve the cluster problem. At
first glance it may seem we perform some unnecessary Fourier transforms. However,
for the imaginary time Fourier transform we need high-frequency conditioning. High-
frequency conditioning means that we treat the high-frequency behavior of the Green
function analytically. This is particularly simple in (iω, K)-space where the Green
function decays as 1/(iω) [93, 94].
27
3 Conventional Mean-Field Methods for Correlated Lattice Models
At the end, Gconf will be disorder averaged and only the terms diagonal in K will
survive.
Alternatively, one can measure the cluster Green function in r and τ space. One
has to measure only the diagonal elements in r space, because all other elements will
not survive the disorder-averaging. This is because the Green function is diagonal
in k-space after the disorder-average as shown in section 2.2.3.
3.3 Summary
In this chapter, we introduced two mean-field methods for the solution of interacting
electrons on a disordered lattice. These methods have been developed prior to the
dual fermion approach. Both rely on the mapping to an impurity or cluster problem
that has to be solved self-consistently.
First, in section 3.1 DMFT was introduced. It is exact in infinite dimensions, but
ignores non-local correlations altogether if applied to finite dimensional systems.
This neglect of non-local correlations is what we want to overcome. The section
about DMFT was closed with an introduction of TMT, which is a modification of
DMFT that is able to detect Anderson localization.
Second, in section 3.2 DCA was discussed, a method developed to improve DMFT
by including short-range correlations. DCA is a powerful method for interacting
systems, but it is limited to relatively small system sizes, especially in 3d. Generally
Quantum Monte Carlo is used to solve the cluster problem, which means that DCA
indirectly suffers from the sign problem away from half-filling. We use DCA as a
benchmark for the dual fermion approach presented in the next chapter.
28
4 The Dual Fermion Approach for
Disordered Interacting Systems
The dynamical mean-field theory (DMFT) presented in section 3.1 neglects non-
local correlations. Thus, depending on the system under investigation, DMFT can
be quite a crude approximation giving qualitatively wrong results, for example in
the vicinity of antiferromagnetism or charge-density waves.
One way to add back non-local correlations is by means of cluster methods, e. g.
the dynamical cluster approximation (DCA) [32, 33], cellular DMFT (cDMFT) [34]
or large scale quantum Monte Carlo (QMC) [42, 43] simulations. If DCA and
cDMFT use QMC to solve the cluster problem, all three methods will in general
suffer from the sign problem. Therefore, these three methods are often limited to
rather short-ranged correlations as the cluster sizes are usually quite small because
an increasing cluster size worsens the sign problem. This is especially problematic
in more than one dimension.
In one dimension it is possible to overcome the sign problem [44]. For translation-
ally invariant single-orbital Hubbard-like systems the sign problem can be overcome
at half-filling as well. However, in the case of disorder there will generally be disor-
der realizations that are away from half-filling. Thus, in more than one dimension
the sign problem is typically present.
Other cluster methods in the context of disorder are the traveling cluster approxi-
mation [95, 96], the molecular coherent potential approximation [97, 98, 99] and the
cluster coherent potential approximation [100, 101, 102, 103, 104].
Another way to add back non-local correlations is to use a perturbative approach.
The calculations involved are usually less costly and they do not suffer from the
sign problem if a corresponding reference system can be chosen. However, ordinary
perturbation theory suffers from the problem that one has to find a small expansion
parameter, which is not an easy task for the Hubbard model.
A way out is to introduce new degrees of freedom that allow to do a perturbative
expansion in the non-local correlations only. Around the DMFT result for small
values of U non-local correlations will be small. On the other hand, for strong inter-
actions, i. e. around the atomic limit, non-local correlations will not be significant
either. Therefore, an expansion in terms of the non-local contributions is desirable.
29
4 The Dual Fermion Approach for Disordered Interacting Systems
1
The transformation used is actually more general than a Hubbard-Stratonovich transformation,
but the name has established itself in the field of dual fermions and thus it is used in this thesis.
2
An introduction to path integrals can be found in [107].
30
4.1 Derivation of the Formalism
where η is a source field and Z is the partition function. The partition function is
defined as Z Y
Z= dc̄ωk dcωk e−S (4.2)
ωk
in the path integral formalism. c and c̄ are Grassmann numbers. To shorten the
notation, ω ≡ iωn is used to label Matsubara frequencies and k is a momentum
label.
The action for a given disorder configuration is
∆ω is a hybridization function that is not yet specified. In section 4.4 we will show
how it can be determined self-consistently.
As it is inconvenient to deal with {ln Z}, we use the replica trick [81, 82] to
integrate out the disorder as shown in section 2.2.3. In this approach, we use the
replica trick to derive the formalism and to obtain diagrams. We will not use it to
do calculations and extrapolate m → 0 for our data. This extrapolation is done only
for the diagrams. Therefore, we are not concerned about replica symmetry breaking
[83].
31
4 The Dual Fermion Approach for Disordered Interacting Systems
where
XZ β XZ β
SRv = − c̄αωkσ (ω + µ − k )cαωkσ − dτ vi nαi (τ ) + dτ U nα↑i (τ )nα↓i (τ )
XX
with
v,imp
XZ β XZ β
c̄αωiσ (ω + µ − ∆ω )cαωiσ − dτ vi nαi (τ ) + dτ U nα↑i (τ )nα↓i (τ )
XX
Si,R =−
α ωσ α 0 α 0
(4.9)
Q α α
and D[c̄i , ci ] = ωα dc̄ωi dcωi .
Greek letters α, β . . . that appear as indices here
and in the following are replica labels. Be aware not to confuse the replica label β
and with the inverse temperature β and the dispersion relation . In the above,
the replica limit introduced limm→0 −1 m
. This terms is a constant with respect to
derivatives of the partition function and because such derivatives are performed
before the replica limit there will be no problem with 10 .
It is now possible to integrate out the disorder as was done in section 2.2.3. This
leads to the action
i
c̄αωkσ (∆ω − k )cαωkσ
X XX
SR = Simp − (4.10)
i α ωkσ
with
XZ β
i
c̄αωiσ (ω + µ − ∆ω )cαωiσ − W (ñi ) + dτ U nα↑i (τ )nα↓i (τ ).
XX
Simp =− (4.11)
α ωσ α 0
32
4.1 Derivation of the Formalism
and XZ
ñi = dτ nαi (τ ). (4.13)
α
hvil ic is the lth cumulant of the disorder distribution. This was derived in sec-
tion 2.2.3.
In the following, we will omit the spin index for better readability. It is easy to
reintroduce the spin index if needed. All configurations which conserve the spin are
possible, e. g., hc↑ c↓ c̄↑ c̄↓ i is possible, hc↑ c↓ c̄↑ c̄↑ i is not.
The next step is to perform a Hubbard-Stratonovich transformation
λ2
c̄α 2 A2ωk Z ¯
−λω (c̄α α ¯α α ω ¯α α
ωk fωk +fωk cωk )− A2 fωk fωk
e ωk Aωk cωk = 2 D[f , f ]e ωk , (4.14)
λω
A2ωk = ∆ω − k , (4.15)
with
i i
λω (c̄αωi fωi
α
+ f¯ωi
α α
X
Ssite = Simp + cωi ). (4.17)
ωα
Ssite does not contain any non-local terms. In fact, all non-local correlations have
been removed from the c-degrees of freedom and have been transfered to the f -
degrees of freedom by means of the Hubbard-Stratonovich transformation. There-
fore, the c-degrees of freedom can be integrated out.
In the following, the dual potential will be defined and an expression for it will
be derived. It will turn out that the dual potential is related to two-particle vertex
3
To shorten the notation and increase the readability, we sometimes refer to the c/f Grassmann
numbers as c/f operators.
33
4 The Dual Fermion Approach for Disordered Interacting Systems
Only even orders of n survive the integration, because expectation values like hci,
hc̄cci, . . . with an odd number of operators vanish. Also, for even orders, only contri-
butions with the same number of c̄ and c have to be considered because expectation
values like hcci vanish for the models in this thesis.
In lowest order the expansion yields
Z
n=0: D[c̄, c]e−Simp = Zimp . (4.21)
34
4.1 Derivation of the Formalism
Equation (4.22c) follows from eq. (4.22b) by noting that there is no travel of electrons
between different replicas and time translational invariance, i. e. hcαω c̄βω0 i ∝ δαβ δωω0 .
Finally, the definition of the Green function eq. (3.3) leads to eq. (4.22f).
In fourth order the expansion yields
1 Z
D[c̄, c]e−Simp λω1 λω2 λω3 λω4 (c̄αω1 fωα1 + f¯ωα1 cαω1 )
X
n=4:
24 αβγδ
ω1 ω2 ω3 ω4
× (c̄βω2 fωβ2 + f¯ωβ2 cβω2 )(c̄γω3 fωγ3 + f¯ωγ3 cγω3 )(c̄δω4 fωδ4 + f¯ωδ4 cδω4 ) (4.23a)
1 Z
D[c̄, c]e−Simp
X
= λω1 λω2 λω3 λω4
24 αβγδ
ω1 ω2 ω3 ω4
× (c̄ω1 cω2 c̄ω3 cω4 fω1 fω2 fω3 fω4 + c̄αω1 c̄βω2 cγω3 cδω4 fωα1 fωβ2 f¯ωγ3 f¯ωδ4
α β γ δ α ¯β γ ¯δ
+ c̄αω1 cβω2 cγω3 c̄δω4 fωα1 f¯ωβ2 f¯ωγ3 fωδ4 + cαω1 c̄βω2 c̄γω3 cδω4 f¯ωα1 fωβ2 fωγ3 f¯ωδ4
+ cαω1 c̄βω2 cγω3 c̄δω4 f¯ωα1 fωβ2 f¯ωγ3 fωδ4 + cαω1 cβω2 c̄γω3 c̄δω4 f¯ωα1 f¯ωβ2 fωγ3 fωδ4 ) (4.23b)
1 Z
D[c̄, c]e−Simp λω1 λω2 λω3 λω4 cαω1 cβω2 c̄γω3 c̄δω4 f¯ωα1 f¯ωβ2 fωγ3 fωδ4
X
= (4.23c)
4 αβγδ
ω1 ω2 ω3 ω4
Zimp X
= λω+ν λ−ω λ−ω0 λω0 +ν hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp f¯ω+ν
α
f¯−ω
α α
f−ω α
0 fω 0 +ν
4 α
ωω 0 ν
Zimp X = ¯α
+ λω+ν λ−ω λ−ω0 λω0 +ν hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iVimp fω+ν f¯−ω
β β
f−ω α
0 fω 0 +ν
4 αβ
ωω 0 ν
Zimp X × ¯α
+ λω+ν λ−ω λ−ω0 λω0 +ν hcαω+ν cβ−ω c̄α−ω0 c̄βω0 +ν iVimp fω+ν f¯−ω
β α
f−ω β
0 fω 0 +ν
4 αβ
ωω 0 ν
(4.23d)
Zimp X
= λω+ν λ−ω λ−ω0 λω0 +ν hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp f¯ω+ν
α
f¯−ω
α α
f−ω α
0 fω 0 +ν
4 α
ωω 0 ν
Zimp X = ¯α
+ λω+ν λ−ω λ−ω0 λω0 +ν hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iVimp fω+ν f¯−ω
β β
f−ω α
0 fω 0 +ν .
2 αβ
ωω 0 ν
(4.23e)
V =(×)
Here, h. . . iimp denotes an average in the horizontal (crossed) channel where no
Hubbard interaction lines connect Green function lines on the two-particle level,
i. e. all the two-particle diagrams one obtains for U = 0 and V > 0. Corresponding
diagrams are shown in a) of fig. 4.1. The diagrams in the top row of a), b) and c)
correspond to the horizontal channel, the diagrams in the lower rows belong to the
crossed channel.
35
4 The Dual Fermion Approach for Disordered Interacting Systems
β β β β β β
a)
...
α α α α α α
β α β α
...
α β α β
α α α α α α
b)
α
...
α α α α α
α α α α
...
α α α α
c)
α α α α α α
...
α α α α α α
α α α α
...
α α α α
Figure 4.1: Different contributions to the vertex function. Depending on the spin config-
uration not all shown diagrams may be possible. The contributions in the top
row of a) are part of hccc̄c̄iV = and the contributions in the bottom row of a)
are part of hccc̄c̄iV × . These two expectation values also contain the uncon-
nected two-particle Green functions, which are not depicted in the above. The
contributions in b) and c) are part of hccc̄c̄iU . Note that the Green function
lines denote dressed single-particle Green functions. (Based on figure by S.-X.
Yang)
36
4.1 Derivation of the Formalism
h. . . iUimp denotes an average, where the two Green functions involved are connected
by at least one Hubbard interaction line that fixes replica indices. Alternatively,
one can describe h. . . iUimp as containing all the additional diagrams that are created
on the two-particle level when the Hubbard interaction U is turned on. Example
diagrams are shown in b) and c) of fig. 4.1.
The final expression in eq. (4.23) is obtained as follows. From eq. (4.23a) to
eq. (4.23b) it was used that only six from the possible sixteen terms of the product
can survive the integration. Then the c- and f -operators are commuted and relabeled
such that only one term is left. Equation (4.23d) follows by integrating out the c-
electrons.
Evaluating the replica indices requires some care as contributions with Hubbard
interactions are different from contributions without. Hubbard interaction lines fix
all replica indices as the Hubbard interaction does not act between different replicas.
This leads to the first line of eq. (4.23d) and explains the need to define h. . . iUimp .
V =(×)
h. . . iimp contains terms without Hubbard interactions, thus replica indices are
only fixed by electron travel. Two connections are possible: the one in the second
line of eq. (4.23d) corresponds to the horizontal channel and the one in the third
line corresponds to the crossed channel.
For the final expression, the c- and f - numbers are commuted and the frequencies
relabeled such that only one term is left for the disorder contribution. We opted for
the horizontal channel because in the following it will be related to the disorder chan-
nel of the two-particle vertex, which has no crossed contribution. Time-translational
invariance was used for the frequency indices. The separation of the two-particle
Green function in a purely disordered part and the rest is necessary for the deriva-
tion of the formalism. Later on, it can be convenient to combine both contributions
to the dual potential for calculations. Reintroducing the crossed channel for the
disorder part can facilitate calculations as well.
We continue with expanding the right hand side of eq. (4.18). To this end, the
ansatz
Vd = a2 f¯α f β + a4 f¯α f¯β f γ f δ + . . . (4.24)
is used. In principle, one has to go up to infinite order in f¯ and f and then the
mapping to the dual degrees of freedom is exact. However, for practical reasons, we
will limit ourselves to second order.
The expansion reads
λ2ω g(ω)f¯ωα fωα −Vd [f¯α ,f β ]
P
Zimp e− λ2ω g(ω)f¯ωα fωα
X
ωα = Zimp 1 −
αω
1 X 2 2
+ λω λω0 g(ω)g(ω 0 )f¯ωα fωα f¯ωβ0 fωβ0 − Vd ± . . . .
2 αβωω0
(4.25)
37
4 The Dual Fermion Approach for Disordered Interacting Systems
The ansatz for Vd is inserted and the coefficients for the dual potential are read off
by comparing order by order. One finds that
a2 =0 (4.26)
1
a4 = − λω1 λω2 λω3 λω4 hcα cα c̄α c̄α iUimp
4
1
− λω1 λω2 λω3 λω4 δαβ hcα cβ c̄β c̄α iVimp
=
(4.27)
2
1 α α
+ λ2ω1 λ2ω2 gimp (ω1 )gimp (ω2 )δω1 ω3 δω2 ω4 δαβ .
2
Next, hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp and hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iVimp
=
are examined in more de-
tail such that we know how to calculate them. Both terms contribute to the two-
particle Green function of the replicated system. This is clear from eq. (4.23) which
implies
hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iimp = 2hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iVimp
=
+ hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp δαβ , (4.28)
Taking into account the third contribution in eq. (4.27), λ = g −1 seems a natural
choice as the dual potential then becomes a two-particle vertex function.
hcαω+ν cβ−ω c̄β−ω0 c̄αω0 +ν iVimp is particularly easy, as this is the disordered particle-particle
Green function. Combining it with the third contribution in eq. (4.27) yields
as one coefficient for the dual potential. hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp is obtained from
the disorder-averaged particle-particle Green function by subtracting the disorder-
averaged unconnected particle-particle Green functions
hcαω+ν cα−ω c̄α−ω0 c̄αω0 +ν iUimp δαβ = hcαω+ν cβ−ω c̄γ−ω0 c̄δω0 +ν iimp δαδ δβγ − δαγ δβδ
(4.31)
×
− hcα cβ c̄β c̄α iVimp
=
+ hcα cβ c̄α c̄β iVimp
38
4.1 Derivation of the Formalism
tracted from the particle-particle Green function, which follows from eq. (4.23d).
The final expression for the dual potential reads
1 1
Vd = − γ U p f¯α f¯α f α f α − γ vp f¯α f¯β f β f α . (4.33)
4 2
Now the dual action can be written as
with
h i−1
G0d (ω, k) = −gω gω + (∆ω − k )−1 gω
(4.35)
= G(ω, k) − gω
after the replica limit. Note that this term can only survive the replica limit as part
of a diagram because there are two replica indices left before limm→0 is evaluated.
It is sometimes convenient to have the crossed component
as well. For γ vp and γ vp× all possible spin configurations have the same numerical
value as we look at systems with spin symmetry. We introduce the symbols
vp
Vvp ≡ γσσ 0 (4.38)
× vp×
Vvp ≡ γσσ 0 (4.39)
39
4 The Dual Fermion Approach for Disordered Interacting Systems
for the disorder part of the dual potential. Correspondingly, one obtains
for the terms with Hubbard interactions. Note that the crossing-symmetric uncon-
nected two-particle Green function is subtracted before the disorder-average. Only
a limited number of spin configurations is needed, e. g. only the pairing channel for
the dual potential in the particle-particle channel
Up
Vp (ν)ωω0 = γ↑↓;↑↓ (ν)ωω0 . (4.41)
Crossing symmetry was used in the above. The other two components of the triplet
Up Up
channel are γ↑↑;↑↑ and γ↓↓;↓↓ . All three contributions are equal due to SU(2) symme-
try.
For higher-order approximations the particle-particle channel and particle-hole
channel are not equivalent. Therefore the dual potential for the particle-hole channel
is needed. It is given by
for the part containing Hubbard interactions. Here, it is convenient to introduce the
vertical disorder potential
40
4.2 Perturbation Theory for the Self-Energy
v
where Vv ≡ γσσ 0 . For the interacting part we use
U U
Vd/m (ν)ωω0 = γ↑↑;↑↑ (ν)ωω0 ± γ↑↑;↓↓ (ν)ωω0 . (4.46)
These are the density and magnetic channels. A more detailed discussion of two-
particle vertices is presented in appendix D.
This concludes the introduction of the dual degrees of freedom. The bare dual
Green function and the dual potential define the dual lattice problem.
41
4 The Dual Fermion Approach for Disordered Interacting Systems
a) β β b) ǫ β β ǫ
β β ǫ ǫ
α α α α
ǫ ǫ
→ c) d)
α α
β β ǫ ǫ β β
ǫ ǫ
α α α α
Figure 4.2: The box represents the dual potential, the green lines with arrows represent
f (f¯) operators. In first order, there are four possible ways to connect the in-
and outgoing f (f¯) operators to the dual potential. For the crossing symmetric
component of the dual potential all replica indices are fixed to , because the
Hubbard interaction is local within a replica and electrons travel only within
one replica. Therefore only one free replica index is left and the replica sum is
of order m. The diagrams will survive the replica limit. Because of the crossing
symmetry of the interaction, all four contributions a)-d) give the same result.
This cancels the factor 41 in the dual potential. For the crossing asymmetric
component, diagrams a) and b) are not possible, because top and bottom are
of the diagram are only connected by the effective disorder interaction which
connects different replica. The replica sum is of order m2 and does not survive
the replica limit. For diagrams c) and d), the internal Green function fixes
the replica indices and the replica sum will be of order m. Thus, the diagram
survives the replica limit. The two diagrams are equivalent and cancel the
factor 12 in the dual potential.
a) b)
β β δ δ β β δ δ β β δ δ
ǫ
ǫ ǫ
→
α α γ γ ǫ α α γ γ ǫ ǫ α α γ γ
Figure 4.3: In second order up to sixteen connections are possible, depending on whether
all internal Green function lines are distinguishable. Two of these diagrams are
shown in this figure. All possibilities are equivalent to one of the two chosen
above. If one of the vertices is crossing-symmetric it is enough to pick one of
the two as they are related by crossing symmetry. In this case, only one free
replica index remains as the crossing-symmetric vertices have only one free
replica index. If both vertices are not crossing-symmetric both diagrams have
to be taken into account as they are not equivalent. In this case, only diagram
b) survives the replica limit. Diagram a) has two free replica indices remaining.
Both times all prefactors cancel up to a factor 21 for two indistinguishable
1
Green function lines. This generalizes to k! for k indistinguishable lines in
higher orders.
42
4.2 Perturbation Theory for the Self-Energy
Figure 4.4: Two diagrams with closed Fermi loops that have a free replica index. The red
line shows the closed Fermi loop. The example on the right uses the crossed
disorder vertex. It is not part of the dual potential as it has been derived, but
at times it can be convenient to add the crossed contribution back to the dual
potential.
energy:
1. Draw n boxes and connect them in all topologically different ways via directed
lines. Exactly two corners must have no connections.
3. The boxes are associate with a factor γ U + γ V . Draw diagrams for every com-
bination of γ U and γ V .
6. Remove all diagrams that have more than one free replica index. These dia-
grams contain closed Fermi loops that are connected to the rest of the diagram
only through disorder vertices. Figure 4.4 shows two examples of closed Fermi
loops. Now, the replica limit has been performed and the replica labels can
be removed from the diagrams.
7. Attach frequency, spin and momentum labels to each Green function line.
8. Sum over all internal frequencies, spins and momenta, obeying energy, spin
and momentum conservation.
43
4 The Dual Fermion Approach for Disordered Interacting Systems
Figure 4.5: Contracting the box that represents the vertex to a single line determines
whether one has a closed loop or not. In the upper example one has a closed
loop, but not in the lower one.
9. Every independent frequency summation gives a factor 1/β and every inde-
pendent momentum summation a factor 1/N . β is the inverse temperature
and N the number of dual lattice sites.
10. Multiply by 1/k! for every set of k indistinguishable lines, e. g. the two internal
Green function lines in the two diagrams in fig. 4.4 are indistinguishable for
equal spin.
11. For every diagram determine the number of closed loops and multiply with
a factor of (−1)nl where nl is the number of closed loops. Closed loops are
identified as follows. For interaction vertices contract the boxes to lines. Do
this consistently for the whole diagram. If a Green function line ends at its
starting point, i. e. it is local, it counts as a closed loop. This is illustrated
in fig. 4.5. These loops are not to be confused with the closed Fermi loops
illustrated in fig. 4.4 For disorder vertices the corresponding diagrams are
removed by the replica limit in step 6.
These rules are an adapted form of the rules given in [105] to accommodate the
changes due to disorder.
A natural question to ask is why dual diagrams with closed Fermi loops are
removed. If one constructs real fermion diagrams with the same topology, one finds
that it is not possible to construct such diagrams with the Feynman rules of the
disorder-averaged real fermion system. This is illustrated in fig. 4.6.
A good starting point for approximations of the self-energy is the Schwinger-
Dyson equation. It relates the self-energy to the full vertex function. It is shown for
the dual degrees of freedom in figure 4.7. For the particle-hole channel the equation
44
4.2 Perturbation Theory for the Self-Energy
Figure 4.6: Two examples of diagrams that are removed by the replica limit. The topology
of the dual fermion diagrams (left) is mapped to corresponding real fermion
diagrams (right). These diagrams are not possible for real fermions. The
conclusion is that the replica limit removes diagrams with a topology that is
unphysical for real fermions. This fact underlines that there is a close relation
between dual and real fermion excitations. Section 4.3 investigates this in more
detail.
a) p+q p+q
p′ + q
Σ p
= Ṽ + Ṽ F̃
p p p p′ p
b) −p + q −p + q
−p′ + q
Σ p
= Ṽ p + Ṽ p F̃ p
p p p p′ p
Figure 4.7: Illustration of the Schwinger-Dyson equation for the particle-hole and particle-
particle channel. (Illustration by S.-X. Yang [106])
45
4 The Dual Fermion Approach for Disordered Interacting Systems
ph pp
Figure 4.8: Spin configurations in first order for the particle-hole (ph) and particle-particle
(pp) channel.
reads
T X
Σph
σ (ω, k) = − Ṽ (ν)ω;ω Gdσ0 (ω + ν, k + q)
N qσ0
T
Ṽ (ν)ω,ω0 Gdσ0 (ω 0 + ν, k 0 + q)Gdσ00 (ω 0 , k 0 )×
X
− (4.47)
2N ω 0 ,k0 ,ν,q
σ 0 ,σ 00 ,σ 000
where Ṽ is the dual potential, i. e. the impurity vertex, and F̃ is the full dual vertex.
Spin indices on Ṽ and F̃ have been suppressed for readability. Note that for U = 0
the bare vertex does not exist for the real fermion system. Therefore, the Schwinger-
Dyson equation does not exist. However, for the dual degrees of freedom the bare
vertex can be defined and thus the Schwinger-Dyson equation can be derived for the
dual system.
The first order contribution has two possible spin configurations. These are shown in
fig. 4.8. In first order, contributions from the particle-particle channel and particle-
hole channel are topologically equivalent, therefore only one contribution has to be
taken into account. For the particle-particle channel it follows that
T X ×
Σ(1)
pp (ω, k) = (Vp (ν, q)−ω;−ω + Vt (ν, q)−ω;−ω + Vvp (ν = 0, q)−ω;−ω )×
N ν,q
× Gd (ω + ν, k + q)
(4.48)
T X
= (2Vp (ν, q)−ω;−ω + Vp0 (ν, q)−ω;−ω + Vvp
×
(ν = 0, q)−ω;−ω )×
N ν,q
× Gd (ω + ν, k + q).
46
4.2 Perturbation Theory for the Self-Energy
Figure 4.9: The three different spin configurations for the particle-particle channel. The
first contribution has a symmetry factor of 2!1 due to the two indistinguishable
Green function lines. This is an example of Feynman rule 10. The second and
third contributions are topologically equivalent, therefore there is a symmetry
factor 12 .
in second order. The name vertex ladder comes from the included diagrams which
look like ladders with the vertices as rungs. Here, Ṽp is defined as
χpp d d
0 (ν, q)ωk = −G (ω + ν, k + q)G (−ω, −k). (4.53)
The second term in equation 4.50 removes the unphysical disorder contribution.
47
4 The Dual Fermion Approach for Disordered Interacting Systems
This is due to the replica limit. For the triplet channel, i. e. the left diagram in
fig. 4.9, one obtains
(2)
Ṽt (ν)ωω00 χ̄pp Vvp (ν)ω χ̄pp
X X
Φt (ν, q) = 0 (ν, q)ω 00 Ṽt (ν)ω 00 ω 0 − 0 (ν, q)ω 00 Vvp (ν)ω 0 δω;ω 0 ,
ω 00 ω 00
(4.54)
where
Ṽt = Ṽp + Ṽp0 (4.55)
and
X 0 (ν)ω;ω0 := −X(ν)ω;−ω0 −ν (4.56)
(2)
defines a rotation for some vertex function X. With this, Φt becomes
(2)
Ṽp (ν)ωω00 χ̄pp
X
Φt (ν, q)ωk;ω0 k0 = 0 (ν, q)ω 00 Ṽp (ν)ω 00 ω 0
ω 00
Ṽp (ν)ωω00 χ̄pp
X
+ 0 (ν, q)ω 00 Ṽp0 (ν)ω 00 ω 0
ω 00 (4.57)
Vvp (ν)ω δω;ω00 χ̄pp
X
− 0 (ν, q)ω 00 Vvp (ν)ω 0 δω 00 ;ω 0
ω 00
=Φ(2)
p (ν, q)ωk;ω 0 k0 + Φ(2)0 (2)
p (ν, q)ωk;ω 0 k0 − Φvp (ν, q)ωk;ω 0 k0 .
Here, we defined
Therefore, the total vertex ladder for second order in the particle-particle channel
reads
Φ(2) (2) (2) (2)0
pp = 2(Φp − Φvp ) + Φp . (4.61)
The self-energy is calculated as
T X (2)
Σ(2)
pp (ω, k) = Φ (ν, q)−ω;−ω Gd (ω + ν, k + q). (4.62)
N ν,q pp
Alternatively, it is possible to calculate the second order self-energy from the particle-
hole channel. This is equivalent to the particle-particle channel because the diagrams
are topologically equivalent.
Three different spin configurations are possible for the particle-hole channel. These
are shown in fig. 4.10. The vertices can be expressed in terms of the magnetic and
48
4.2 Perturbation Theory for the Self-Energy
Figure 4.10: The three different spin configurations for the particle-hole channel. The
second and third contributions are topologically the same, therefore there is
a symmetry factor 12 . The first contribution has the same symmetry factor
due to the two indistinguishable Green function lines.
For the particle-hole channel the k-averaged unconnected two-particle Green func-
tion is
T X ph
χ̄ph
0 (ν, q)ω = χ (ν, q)ωk (4.65)
N k 0
with the unconnected two-particle Green function
χph d d
0 (ν, q)ωk = G (ω + ν, k + q)G (ω, k). (4.66)
(4.70)
49
4 The Dual Fermion Approach for Disordered Interacting Systems
Figure 4.11: Example of an (unphysical) ring type diagram. The red line marks the way
of the electron through the diagram.
Here, Γ̃ is the irreducible dual fermion vertex. For the FLEX the irreducible dual
vertex is approximated by the dual potential Γ̃ ≈ Ṽ . Then the equation for F̃ can
be solved as
1
F̃ (ν, q)ωk;ω0 k0 = × Ṽ (ν)
1 − Ṽ (ν)χ̄0 (ν, q) ωk;ω 0 k0
(4.72)
1
= Ṽ (ν) × .
1 − χ̄0 (ν, q)Ṽ (ν) ωk;ω0 k0
For the calculation of the self-energy, Φ(2) has to be subtracted from Φ because
the second-order contribution is special and has to be treated separately. First,
additional symmetry factors do exist in second order. These symmetry factors do
not appear in higher order diagrams and therefore the prefactor is not suitable for
the ladder summation. Second, this avoids double counting because in second order
50
4.2 Perturbation Theory for the Self-Energy
Figure 4.12: Particle-hole diagrams for the third order contributions to the FLEX.
From this the FLEX contribution to the self-energy can be calculated for the particle-
particle channel as
T Xh
ΣFppLEX (ω, k) = 2(Φp (ν, q)−ω,−ω − Φp(2) − Φvp (ν, q)−ω,−ω )
N ν,q (4.76)
p(2)0
i
p0 d
+(Φ (ν, q)−ω,−ω − Φ (ν, q)−ω,−ω ) G (ω + ν, k + q)
where
1
Φvp (ν, q)ωk;ω0 k0 = × Vvp (ν) − Vvp (ν)ω;ω0 χ̄pp
0 (ν, q)Vvp (ν)ω;ω 0
1 − Vvp (ν)χ̄pp
0 (ν, q) ωk;ω 0 k0
− Vvp (ν)ω;ω0
(4.77)
51
4 The Dual Fermion Approach for Disordered Interacting Systems
1 Ṽd
ΦFphLEX = − Ṽd χ̄ph
0 Ṽd − Ṽd
2 1 − Ṽd χ̄ph 0
1 1
+ ph Vv
||
ph − Ṽd χ̄ph || || ph
0 Vv − Vv χ̄0 Ṽd − Vv
||
1 − Ṽd χ̄0 1 − χ̄0 Ṽd
(4.78)
3 Ṽm ph
+ − Ṽ m χ̄ 0 Ṽm − Ṽm
2 1 − Ṽm χ̄ph 0
Vv= = ph = =
−2 − V v χ̄ 0 V v − V v
1 − Vv= χ̄ph0
This concludes the calculation of the self-energy. The next section discusses the
connection between dual and real quantities. This will enable us to formulate a
self-consistency loop in section 4.4.
52
4.3 Relations between Dual and Real Quantities
real vertex function. The relations obtained here are very similar to the relations
derived in [105]. Similar relations do exist for higher Green functions, as well as for
the corresponding cumulants and ultimately the vertex functions.
In section 4.5 vertex corrections for the conductivity are calculated using the real
fermion vertex. Non-local corrections to the vertex are calculated at the level of
the dual degrees of freedom and a simple relation for the vertex functions allows to
translate the dual vertex to the real vertex.
The starting point for both examples is the functional
Z
F [K̄, K] = ln Zf D[c̄, c]D[f¯, f ] exp(−S[c̄, c; f¯, f ] + K̄1 f1 + f¯2 K2 ) (4.81)
spin, frequency and momentum. Integrating out the c-degrees of freedom yields
Z
F [K̄, K] = ln Z̃f D[f¯, f ] exp(−Sd [f¯, f ] + K̄1 f1 + f¯2 K2 ) (4.82)
Z
with Z̃f = Zd
and
Z
¯
Zd = D[f¯, f ]e−Sd [f ,f ] . (4.83)
Now functional derivatives like
δ 2 F [K̄, K] δ 1 Z
− = − D[f¯, f ]f1 exp(−Sd [f¯, f ] + K̄1 f1 + f¯2 K2 )
δK2 δ K̄1 K̄1 =K2 =0 δK Zd
1 Z
= − D[f¯, f ]f1 f¯2 exp(−Sd [f¯, f ] + K̄1 f1 + f¯2 K2 )
Zd
1 Z
+ 2 D[f¯, f ]f1 exp(−Sd [f¯, f ] + K̄1 f1 + f¯2 K2 )×
Zd
Z
× D[f¯, f ]f¯2 exp(−Sd [f¯, f ] + K̄1 f1 + f¯2 K2 )
1 Z
= − ¯ ¯ ¯ ¯
D[f , f ]f1 f2 exp(−Sd [f , f ] + K̄1 f1 + f2 K2 )
Zd
=Gd
(4.84)
53
4 The Dual Fermion Approach for Disordered Interacting Systems
δ 4 F [K̄, K] 1 δ4
= lim Z̃fm ×
δK4 δK3 δ K̄2 δ K̄1 K̄1 =K̄2 =K3=K4=0 m→0 m δK1 δK2 δ K̄3 δ K̄4
Z
× D[f¯, f ] exp(−Sd [f¯, f ] + K̄µ fµ + f¯ν Kν )
1 Z
D[f¯, f ] f1α f2β f¯3γ f¯4δ exp(−Sd [f¯, f ] + K̄µ fµ + fν Kν )
X
= lim
m→0 m
αβ
γδ
1 Z X
= lim D[f¯, f ] δαβ f1α f2α f¯3α f¯4α + f1α f2β f¯3β f¯4α + f1α f2β f¯3α f¯4β ×
m→0 m
αβ
is obtained. Next, the f -degrees of freedom are integrated out from eq. (4.81). One
obtains
Z
F [K̄, K] = K̄1 [g(∆ω − k )g]K1 + ln D[c̄, c] exp(−S[c̄, c]
(4.86)
+K̄1 [g(∆ω − k )]c1 + c̄1 [(∆ω − k )g]K1 ) .
λ2 ¯
− f f − λc̄f − λf¯c + K̄f + f¯K
A2
λ A λ A A2 A2 A2
=− f + Ac̄ − K f + Ac̄ − K + A2 c̄c + 2 K̄K − K̄c − c̄K.
A λ A λ λ λ λ
(4.87)
The same derivatives as in eqs. (4.84) and (4.85) are performed to obtain Gd and
χd1234 + Gd14 Gd23 − Gd13 Gd24 in terms of the c-degrees of freedom. For the dual Green
function the result is
δ 2 F [K̄, K]
− = −gω (∆ω − k )gω
δK2 δ K̄1 K̄1 =K2 =0
1 Z
− D[c̄, c] exp(−S[c̄, c])× (4.88)
exp(F ) K̄1 =K2 =0
× [gω (∆ω − k )c2 c̄1 (∆ω − k )gω ]
54
4.3 Relations between Dual and Real Quantities
2 Z
1
+ D[c̄, c] exp(−S[c̄, c])× (4.89)
exp(F ) K̄1 =K2 =0
Z
× c2 D[c̄, c] exp(−S[c̄, c]c̄1 )
= − gω (∆ω − k )gω + gω (∆ω − k ){Gv }(∆ω − k )gω = Gd .
(4.90)
δ4F δ4 1 Z
= D[c̄µ , cµ ] exp(−S[c̄µ , cµ ]+
δK4 δK3 δ̄K2 K̄1 K=0 δK4 δK3 δ̄K2 K̄1 K=0 m
for the single particle real lattice Green function. Combining eqs. (4.85) and (4.91)
yields
Gdα Gdβ
χα1234 δαβ + Gα14 Gβ23 − Gα13 Gβ24 = 11
α
22
×
(g1 (∆1 − 1 )G11 ) (g1 (∆1 − 1 )Gβ )
dβ dβ
× χdα
1234 δαβ + Gdα
14 G23 − Gdα
13 G24 × (4.93)
Gdβ
33 Gdα
44
× β
(g3 (∆3 − 3 )G33 ) 4 4 4 )Gα44 )
(g (∆ −
for the two-particle vertex. In eq. (4.93) replica indices are kept because the vertex
55
4 The Dual Fermion Approach for Disordered Interacting Systems
is used to construct diagrams. If the replica limit is performed at the level of the
vertex all terms that contain only disorder are removed. With that vertex one would
miss diagrams that are important for the vertex corrections of the conductivity in
section 4.5.
Dual and real fermion quantities are related by linear relations. These relations
show that excitations of the real and dual system are the same [105], on the single-
particle level as well as for many-particle excitations.
where ξ ∈ [0, 1] controls the mixing between the old and new hybridization function.
The solution is self-consistent if Gd,loc ≡ 0. This means that all diagrams containing
a closed dual loop are summed at the level of the impurity problem, independently
of their topology.
Calculating local quantities is equivalent to coarse-graining. Thus, it is at this
point where a certain mean-field character is introduced in the dual fermion ap-
proach. Let us stress that this is only because the dual problem is not solved
exactly. Figure 4.13 shows the self-consistency loop, which goes as follows:
1. Solve the impurity problem and calculate the impurity self-energy and impu-
rity vertex function.
2. Parametrize the dual fermion calculation with the impurity self-energy and
vertex function.
56
4.4 Self-Consistency in the Dual Fermion Approach
Figure 4.13: Self-consistency loop for the dual fermions. First, the dual fermion self-
consistency loop is initialized, e. g. with the DMFT solution. Second, the
dual fermion problem is solved and defines a new impurity problem. Third,
the impurity problem is solved and the output is used as input for the dual
fermion problem. This is iterated until self-consistency is reached. (Based on
illustration by S.-X. Yang [106])
57
4 The Dual Fermion Approach for Disordered Interacting Systems
It should be remarked that the self-consistency loop presented here does not guar-
antee that one finds the optimal solution. The self-consistency condition given in
eq. (4.94) is derived heuristically. In practice it turns out to give good results but a
priori it is not clear that it sums relevant diagrams or is biased in some way.
This concludes the derivation of the self-consistent dual fermion approach. In the
last section of this chapter we will show how the dual degrees of freedom can be
used to construct an approximation for the vertex correction to the conductivity.
58
4.5 Observable of Interest: Conductivity
Vertex corrections to the conductivity vanish in DMFT [109] because the vertex lad-
der has no momentum dependence. Dual fermions retrieve momentum dependence
and thus vertex corrections remain finite and can be taken into account by means
of the parquet equations [110]. According to [106, 111] the dc conductivity can be
approximated as
β2 β
σdc = χxx (q = 0, τ = ). (4.95)
π 2
χxx = hjx (q, τ )jx (−q, 0)i is the current-current correlation function. In the deriva-
tion it was assumed that Imχxx ≈ ωσdc holds for an energy range Ω much larger than
the temperature T [111]. This assumption does not hold for Fermi liquids as Ω ∝ T 2
at low temperatures. Hence, for small values of U and V in the Anderson-Hubbard
model, eq. (4.95) is not a good description of the conductivity. Fortunately, we are
interested in the conductivity around metal-insulator transitions, i. e. not in the
Fermi liquid regime.
χxx is related to its Fourier transform in frequency space via
β β
e−iνm 2 χxx (q = 0, iνm ).
X
χxx (q = 0, τ = )=T (4.96)
2 iνm
T X †
J(q = 0) = vk ckσ ckσ , (4.97)
N kσ
d
where vk = dk
(k). χxx (q = 0, iνm ) can be calculated as
T2 X
χxx (q = 0, iνm ) = χxx (q = 0, iνm )ω,ω0 ;k,k0 (4.98)
N2 ω,ω 0 ;k,k0
T X 2
=− v χ0 (q = 0, iνm )ω;k
N ω;k k
T2 X
− vk χ0 (q = 0, iνm )ω;k ×
N2 ω,ω 0 ;k,k0
Here, F rf is the real fermion vertex which can be calculated from the dual fermion
vertex F d via
F rf (q)pp0 = T (p + q)T (p)F d T (p0 + q)T (p0 ) (4.100)
59
4 The Dual Fermion Approach for Disordered Interacting Systems
with
1
T (p) = − . (4.101)
1 + g(p)Σd (p)
The transformation formula follows from eq. (4.93) and
The last equality is shown in [105]. F d can be approximated by the vertex ladder.
From the calculation of the self-energy the construction for the second order and
FLEX approximations are known, but several changes have to be made compared
to the self-energy. The external Green functions impose different constraints for the
conductivity than for the self-energy. For example,
c†σ cσ
X
J∝ (4.103)
σ
allows only for contributions where the in- and out-going Green functions on each
side of the vertex ladder carry the same spin index. Also, a chain only consisting
of disorder vertices is physical here, since the replica indices are now all fixed due
to the in and out-going Green functions. The corresponding diagrams are created
when the bubble diagram is disorder-averaged. The bubble diagram is obtained by
contracting the ends of the unconnected horizontal two-particle Green function, and
it is given by the contribution in the first line of eq. (4.99).
60
4.5 Observable of Interest: Conductivity
1 3
F F F F
2 4
F F F F
Figure 4.14: Possible spin configurations for the particle-hole horizontal channel.
parts which represent expectation values of the current operator. Without external
fields these expectation values are zero and thus the corresponding diagrams cannot
be present after the disorder-average. All other diagrams are removed because the
spin configurations are not possible for the given combination of vertices.
The diagrams shown in figs. 4.15 to 4.18 are diagrams for the conductivity. Dia-
grams for the vertex ladder are obtained by removing the outer four Green functions.
First, diagram fig. 4.15 translates to:
1 ph 1 1
Φph
1 = (Vd + Vm )χ̄0 (Vd + Vm ) + (Vd + Vm )χ̄ph 0 V0
2 2 2 (4.105)
1
+V0 χ̄ph
0 (Vd + Vm ) + V0 χ̄ph
0 V0 .
2
For fig. 4.16 one gets
1 ph 1
Φph
2 = (Vd − Vm )χ̄0 (Vd − Vm ). (4.106)
2 2
For fig. 4.17 one obtains
1 ph 1 1
Φph
3 = (Vd + Vm )χ̄0 (Vd − Vm ) + V0 χ̄ph
0 (Vd − Vm ). (4.107)
2 2 2
Finally, for fig. 4.18 one gets
1 ph 1 1
Φph
4 = (Vd − Vm )χ̄0 (Vd + Vm ) + (Vd − Vm )χ̄ph
0 V0 . (4.108)
2 2 2
Defining
Ṽd = Vd + V0 (4.109)
the particle-hole vertex ladder can be written as
Φph= ph ph ph ph ph
2nd = Φ1 + Φ2 + Φ3 + Φ4 = Ṽd χ̄0 Ṽd . (4.110)
61
4 The Dual Fermion Approach for Disordered Interacting Systems
F F =
+ +
+ +
+ +
+ +
Figure 4.15: There are nine possible diagrams for the second order contribution to the
conductivity. Above, the (↑↑) → (↑↑) → (↑↑) contribution is analyzed. For
all possible spin configuration, the bottom right diagram is removed by the
replica limit. Additionally, all diagrams with the vertical disorder vertex
vanish, because they are proportional to h~ji = 0 before the disorder average.
62
4.5 Observable of Interest: Conductivity
F F =
+ +
+ +
+ +
+ +
Figure 4.16: For the (↑↑) → (↓↓) → (↑↑) contribution only the upper right diagram is
allowed. This is because all interactions conserve spin.
This can be continued for higher order terms and one can derive
Ṽd
Φph=
F LEX = (4.111)
1 − Ṽd χ̄ph
0
for the FLEX approximation. The horizontal channel has no momentum depen-
dence, therefore it does not contribute to the vertex corrections. It has been pre-
sented in this much detail for instructional purposes.
63
4 The Dual Fermion Approach for Disordered Interacting Systems
F F =
+ +
+ +
+ +
+ +
Figure 4.17: For the (↑↑) → (↑↑) → (↓↓) contribution two diagrams are possible. The
difference between the two diagrams on the second line is that on the left the
disorder vertex connects different spins, which is not allowed, whereas on the
right it connects the same spin.
64
4.5 Observable of Interest: Conductivity
F F =
+ +
+ +
+ +
+ +
Figure 4.18: For the (↑↑) → (↓↓) → (↓↓) contribution two diagrams are possible. The first
diagram on the second line is possible because the disorder vertex connects
the same spin. In the second diagram the disorder vertex connects different
spins and thus the diagram is not possible.
65
4 The Dual Fermion Approach for Disordered Interacting Systems
and diagram 3
ph|| 1 1 1 || || ph 1
Φ3 = (Vd − Vm )χ̄ph
0 (Vd − Vm ) + (Vd − Vm )χ̄ph
0 V0 + V0 χ̄0 (Vd − Vm ). (4.115)
2 2 2 2
Overall, the vertex ladder reads
ph|| 1 || ph || 3
Φ2nd = Vd∗ χ̄ph ∗ ph ph
0 Vd − 2V0 χ̄0 V0 + Ṽm χ̄0 Ṽm − 2V0 χ̄0 V0 (4.116)
2 2
3 1
F 2↔3 2 1
F F
4 3
4 F 2
Figure 4.19: To calculate the contributions from the particle-hole vertical channel, the
vertex is rotated to the particle-hole horizontal channel by exchanging 2 ↔ 3.
Note that this rotation comes with a minus sign.
66
4.5 Observable of Interest: Conductivity
1 2 3
F F F F F F
Figure 4.20: The possible spin configurations for the second order contributions.
with
||
Vd∗ = Vd + V0 + 2V0 Ṽm = Vm + V0 . (4.117)
The pure disorder term is removed because left and right of the diagram are only
connected via disorder and two replica indices remain. Including higher orders, the
FLEX approximation is obtained as
Ṽp = Vp + V0 (4.120)
67
4 The Dual Fermion Approach for Disordered Interacting Systems
F F F
Figure 4.22: By exchanging the bottom left and bottom right corner of the vertex ladder,
the maximally crossed diagram is transformed to a diagram that is internally
a particle-particle diagram.
1 2 3
F F F F F F
Figure 4.23: Spin configurations for the particle-particle channel. The first contribution
carries a factor 21 because the two internal lines are indistinguishable. The
second and third diagrams are topologically equivalent.
Using Ṽp it is easy to include the disorder contributions. Ṽt is crossing symmetric,
therefore one has to take into account only the first diagram in fig. 4.23.
For the first diagram one gets
1 0 V0×
ΦtF LEX = (Ṽp + Ṽ ) − . (4.123)
1 − Ṽp χ̄pp
0
p
1 − V0 χ̄pp
0
F d =2(Φph=
c (q = 0; ν)ω,ω0 ;k,k0 − Φph=
c (k 0 − k; ω 0 − ω)ω,ω+ν;k,k
0 0 (4.124)
− Φppc (k + k ; ω + ω + ν)−ω 0 ,−ω;−k0 ,−k ).
68
4.6 Summary
The factor of 2 comes from the fact that we ignored half of the spin configurations
which are connected to the contributions above by spin symmetry.
The next step is to transform the dual vertex to the real fermion space. This is
easily done using the transformation matrices for vertices ([105]A.136-138)
where
1
T =− . (4.126)
1 + gΣd
This concludes the calculation of the conductivity. It is possible to apply the same
reasoning for other quantities like the susceptibility. This is shown for the clean
system in [112].
4.6 Summary
In this chapter we presented the dual fermion approach for disordered interacting
systems, the main result of this thesis. We started with the derivation of the for-
malism, where we used the replica trick to replace {ln Z} by powers of the disorder-
averaged partition function {Z}. After the replica trick, it was possible to integrate
out the disorder in favor of an effective interaction between replicas. The dual de-
grees of freedom were introduced and the dual potential has been calculated. With
the dual potential at our disposal, we were able to obtain a perturbative solution
of the dual fermion lattice problem in terms of a self-consistent second order and
FLEX approximation.
Unlike for the clean system [105], the replica limit imposes topological restrictions
on the dual fermion diagrams. These restrictions remove dual diagrams that have
a topology that is impossible for real fermion diagrams. This was already a hint at
the strong relation between dual and real excitations.
Next, dual and real fermion quantities were related by exact linear relations, thus
giving a prescription to obtain physical quantities from the dual degrees of freedom.
The fact that these relations are linear, is interpreted as dual and real fermions
having the same excitations.
To conclude the formalism, we wrote down the self-consistency loop for the dual
fermion approach, which is analogous to the DMFT self-consistency loop. We use
the self-consistency condition that the local dual Green function vanishes. Hence,
all dual fermion diagrams with a local dual Green function are implicitly summed.
At the end, the exact relation for the vertex function was used to calculated vertex
corrections for the conductivity. The conductivity will be used later on to determine
the metal-insulator transition for strong disorder.
So far, we did not specify how to solve the impurity and cluster problems in sec-
69
4 The Dual Fermion Approach for Disordered Interacting Systems
tions 3.1 and 3.2 and chapter 4. In the next chapter, the continuous-time Quantum
Monte Carlo method is introduced to tackle this task.
70
5 Quantum Monte Carlo for
Impurity Problems
One common aspect of all the methods presented in sections 3.1 and 3.2 and chap-
ter 4 is the underlying impurity problem. Several methods exist for the solution
of impurity problems. These methods are referred to as impurity solvers or just
solvers. Which solver is used depends on the specifics of the problem at hand, i. e.
calculations at T = 0 or T > 0, are two-particle quantities needed or how many
orbitals or sites comprise the impurity?
The solver of choice for dynamical mean-field theory (DMFT) involving only one-
particle quantities is arguably the numerical renormalization group (NRG) [113,
114]. It provides real frequency spectra and can be applied over a wide range of
parameters. However, practical calculations are mostly limited to one orbital or
one site, respectively. This eliminates NRG from the list of possible solvers for the
dynamical cluster approximation (DCA). The access to two-particle quantities is
very expensive which makes it a bad choice for the dual fermion approach, where
two-particle quantities are needed with their full dependence on three independent
frequencies.
An alternative method for impurity problems is exact Diagonalization (ED). ED
in the context of DMFT and cluster extensions thereof is reviewed in [115]. As the
name suggests, the Hamiltonian is diagonalized by means of numerical algorithms.
The solution is exact at most up to machine precision, depending on the algorithm
used. Popular examples are the Lanczos algorithm [116] and the Jacobi-Davidson
algorithm [117]. Which ED algorithm is used depends on the problem, e. g. are
there degeneracies that one wishes to resolve, is one only interested in the ground
state or are excited states needed?
Despite its flexibility the applicability of ED is limited to small systems. The
exponential growth of the Hilbert space restricts ED to a relatively small number of
sites, e. g. ≈ 20 for the square lattice Hubbard model around half-filling [118, 119]. A
further drawback is the limitation to a discretized bath as the continuum of states of
the infinite lattice is mapped to a finite cluster with discrete energy levels. This can
lead to systematic errors. Therefore, one has to check that the self-energy converges
as the number of bath sites is increased. If this convergence can be achieved, ED
is a good choice for DMFT. Its applicability to dual fermions is less ideal, though.
The calculation of vertex functions with the full dependence on three frequencies is
71
5 Quantum Monte Carlo for Impurity Problems
quite expensive.
In [115] the number of possible cluster sites for cluster-extensions of DMFT is
given as six. This small number comes from the fact that bath sites are needed.
Even in 1d, this does not give a converged result in the number of cluster sites and
hence ED is not suited for DCA in the context of this work.
Finally, a very popular method for 1d systems is the density matrix renormal-
ization group (DMRG) [2]. There are attempts to utilize it as an impurity solver
[120, 121, 122] but it has not yet established itself as such. It is a promising can-
didate for DMFT, but its applicability to DCA, which is mainly used in 2d and 3d
systems, has yet to be explored. DMRG is not suited for dual fermions because
access to the two-particle Green function is numerically too costly.
Quantum Monte Carlo (QMC) can be readily applied for 1d, 2d and 3d systems as
well as for a single impurity. Two particle quantities can be expensive to calculate,
but still come at reasonable computational cost. An advantage of QMC is that
it can be easily parallelized such that it can exploit high performance computing
resources. Until the beginning of the 21st century, the standard method was Hirsch-
Fye quantum Monte Carlo [123]. It suffers from a systematic error which is due to the
discretized imaginary time variable, but this error can be removed by a systematic
extrapolation ∆τ → 0. Nowadays, new algorithms [42, 43, 124, 125, 25] allow to
avoid the discretization of the time variable altogether and thereby avoid systematic
errors. These methods are referred to as continuous-time quantum Monte Carlo
(ctQMC). One major drawback of QMC is that efficient equilibrium algorithms
work in imaginary time, whereas one desires dynamical quantities in real-time. This
requires analytic continuation [126] if one wishes to look at quantities like the density
of states. Real-time algorithms do exist [127, 128] but their usage is not advisable
for equilibrium systems.
Its versatility and the high potential for parallelization make ctQMC the method
of choice for this thesis. As the same ctQMC solver can be applied to the impurity
problem in section 3.1 and chapter 4 and to the cluster problem in section 3.2, it
is better to adopt this single method and use a highly optimized code than to try
different impurity solvers and find the best solver for each individual problem. Also,
method development is the main scope of this thesis, such that a solver that can be
easily adapted to handle more complex impurities and interactions is desirable and
at present only ctQMC fulfills this requirement. In the remainder of this chapter
the basic ideas of the interaction expansion ctQMC [42, 43, 44] are introduced.
Z = Tre−βH . (5.1)
72
5.1 Basic Ideas of the Monte Carlo Method
In general, the trace cannot be evaluated exactly. The associated sums or integrals
are usually high-dimensional. In our case we have a bath in the thermodynamic limit
which means we have infinite dimensions. Integration1 routines sum contributions
to an integral by employing a systematic grid to discretize a continuous system
and approximating integrals by sums. This can become very inefficient for such
c
a high-dimensional system. The scaling of the error is generally of the form N − d
[129] where N is the number of sample points, c is a constant that depends on the
algorithm and d is the dimension of the integral.
A way out are Monte Carlo simulations [130]. In a Monte Carlo simulation sample
points are chosen at random to approximate the integral. Using this method, a
1
scaling of the error as N − 2 can be achieved. This scaling is independent of the
dimension of the integral [131].
Monte Carlo methods require that a weight or probability is attached to every
sample point. To this end, we rewrite the partition function as a sum
X
Z= wC , (5.2)
C
Here, sgn(wC ) is the sign of wC . Now |wC | ≥ 0 can be used as a weight for a
configuration. Expectation values can be calculated according to
PN PN
i=1 OCi wCi i=1 OCi sgn(wCi )|wCi |
hOi ≈ hOiM C = PN = PN , (5.4)
i=1 wCi i=1 sgn(wCi )|wCi |
1
Here and in the following “integration“ includes summation as well.
73
5 Quantum Monte Carlo for Impurity Problems
if the right scheme for creating configurations is employed. This scheme will be
introduced in this section. The variance Var(O) is estimated according to
Equation (5.6) is only valid if the Oi are uncorrelated [133], otherwise eq. (5.6) will
underestimate the variance.
Importance sampling generally improves the efficiency of Monte Carlo algorithms.
The gain in efficiency depends on the problem at hand. For a constant function,
every sample point has the same weight and is as important as all others. There
would be no gain in efficiency if importance sampling is used. Trying to obtain
uncorrelated data generally means one has to create additional configurations and
as a result Monte Carlo procedures with importance sampling would even be less
efficient than the algorithm without. If the integrand has non-zero contributions
in only one percent of the sample space, only one percent of truly random sample
points will contribute. Only those relevant sample points will be created if impor-
tance sampling is used. That way, the ratio of sample points to important sample
points would be 100 times higher. If ten Monte Carlo steps are needed to obtain
uncorrelated data the total gain in efficiency will be a factor of ten.
A random walk in configuration space is used to implement importance sampling.
Starting from a configuration C0 , a chain of configurations
C0 → C1 → C2 → · · · → Cn (5.7)
74
5.2 The Interaction Expansion
p(Cn+1 → Cn ) · |wCn+1 |
a(Cn → Cn+1 ) = min 1, . (5.10)
p(Cn → Cn+1 ) · |wCn |
This means that if the probability of accepting a move from Cn to Cn+1 is greater
than the probability of accepting a move from Cn+1 to Cn we set it equal to one and
accept the move. Using eq. (5.10) to determine the acceptance probability ensures
a random walk with importance sampling.
75
5 Quantum Monte Carlo for Impurity Problems
where δ is a parameter that can be adjusted to reduce the infamous sign problem
(cf. appendix A). This modification does not change the Hamiltonian except for a
constant, so the physics is not affected. In the next step the partition function is
expanded as
∞ n Z β
Z −U Z τn−1
X X
= dτ1 ... dτn ×
Z0 n=0 2 0 i1 ,s1 0
X Y (5.13)
× h[ni1 ,σ (τ1 ) − ασ (s1 )] . . . [nin ,σ (τn ) − ασ (sn )]i0
in ,sn σ
2
It is possible to calculate determinants using fast matrix multiplication and thereby achieving a
complexity of n2.373 [134]
76
5.2 The Interaction Expansion
where Mσ is the matrix on the right-hand side of eq. (5.15). The weight of a
configuration for the random walk is
n Y
U
|wCn | = − det Mσ (Cn )dτ n , (5.20)
2 σ
where the infinitesimals from the integrals are included in the weight. For simplicity,
addition and removal of a vertex are suggested with equal probability. This gives a
factor of 21 for the addition and removal probability and thus cancels out in eq. (5.10).
Every vertex has three components that have to be chosen at random. Without prior
knowledge about the importance of configurations, it is advisable to attach equal
probability to the possible choices for spin up and spin down, the choice of cluster
site and the imaginary time τ . Choosing a spin thus results in a factor 12 , a factor of
1
Nc
comes from the choice of one out of Nc equally likely cluster sites, and a factor dτ
β
from picking a random time τ in the interval [0, β). The probability for proposing
the addition of a certain vertex is therefore
dτ
P (n → n + 1) = . (5.21)
2βNc
The proposal probability of removing a given vertex is just inversely proportional
77
5 Quantum Monte Carlo for Impurity Problems
This formula follows from eq. (5.24) and the fast matrix update formulas in ap-
pendix B. Equation (5.25) is the building block for all quantities that one wishes to
calculate, because Wick’s theorem holds in all configurations and hence all expecta-
tion values of creation and annihilation operators can be decomposed and rewritten
in terms of the single-particle Green function.
78
5.2 The Interaction Expansion
Here, the dependence on spin indices has been ignored. Note that within a single
Monte Carlo configuration, g c is the value of the impurity Green function and it
depends on two frequencies. Only after the QMC calculation becomes the impurity
single-particle Green function g diagonal in frequency space, i. e.
Note that the disconnected part has to be subtract in the vertical channel. This
follows from the definition of the particle-hole Green function and in section 4.1 we
established that the vertex obtained from the disorder averaged particle-hole Green
function is part of the dual fermion potential. For the second case the result is
1
χph
↑↑;↓↓ (ν)ω,ω 0 = hg↑c (ω + ν, ω)g↓c (ω 0 , ω 0 + ν) + g↓c (ω + ν, ω)g↑c (ω 0 , ω 0 + ν)iQMC
2β
1
− hg↑c (ω + ν, ω) + g↓c (ω + ν, ω)iQMC ×
2
× hg↑c (ω 0 , ω 0 + ν) + g↓c (ω 0 , ω 0 + ν)iQMC .
(5.29)
For the particle-particle channel the interaction of an up spin with a down spin is
79
5 Quantum Monte Carlo for Impurity Problems
χp (ν)ω,ω0 +ν ≡ χpp
↑↓;↑↓ (ν)ω,ω 0 . (5.31)
χpp p
↑↓;↓↑ (ν)ω,ω 0 = −χ (ν)ω,−ω 0 −ν . (5.32)
Using eq. (D.15) and SU(2) symmetry, it is now possible to obtain the two-particle
Green function and vertex function for all possible spin configurations that we may
need.
1 NX Dis
1 NX Dis
for the particle-hole channel. g i denotes the impurity Green function obtained by
the solver for the disorder configuration i. This quantity depends on only two
frequencies and does not require additional measurements in the Monte Carlo run,
80
5.3 Summary
therefore it is much cheaper than the interacting two-particle Green functions. The
measurement is illustrated in fig. 5.3.
5.3 Summary
In this chapter we have motivated the use of continuous-time quantum Monte Carlo
and introduced the basic ideas of Monte Carlo methods. The interaction expansion
algorithm (CT-INT) was discussed in more detail as this is the algorithm that was
used in this thesis. CT-INT relies on an expansion in terms of the interaction
part of the Hamiltonian and is in principle exact. However, it is best for small to
intermediate interaction strength as the size of the configuration matrix growth as
U 3 [44].
In section 5.2.3 the measurement of two-particle quantities was explained as these
quantities, namely the two-particle Green functions, are key to the dual fermion
approach. In the end, we combined the interacting impurity problem with the
disorder-average that is needed for DMFT, DCA and dual fermions. The disorder-
average is performed over multiple realization of the disorder for the impurity or
cluster problem.
This concludes the discussion of numerical methods and models. The following
two chapters present results for the Anderson-Falicov-Kimball model and Anderson-
Hubbard model that were obtained by means of the dual fermion approach for
disordered interacting systems.
81
5 Quantum Monte Carlo for Impurity Problems
w′ + ν w+ν
χ
w′ w
= − { + − δν,0}
{ }
= −
{ δ ′−
} w,w F
w′ + ν w+ν
χ
w′ w
= { − δν,0}
Figure 5.1: Definition of the particle-hole Green functions in terms of the Monte-Carlo
measurement. h. . . i denotes a Monte Carlo average for a single configuration,
{. . . } denotes the disorder average (Based on illustration by S.-X. Yang [106]).
82
5.3 Summary
w′ + ν w+ν
χp
−w′ −w
= { }
Figure 5.2: Definition of the two-particle pairing Green function in terms of the Monte-
Carlo measurement. h. . . i denotes a Monte Carlo average for a single configu-
ration, {. . . } denotes the disorder average (Based on illustration by S.-X. Yang
[106]).
w w
χ′
w′ w′
= { }
Figure 5.3: Definition of the particle-hole disorder Green function in terms of the Monte
Carlo measurement. For this illustration spins have been omitted as the result
does not depend on the spin configuration (Illustration by S.-X. Yang [106]).
83
6 Test Case for the Dual Fermion
Algorithm: The
Anderson-Falicov-Kimball Model
Material published in S.-X. Yang, P. Haase, H. Terletska, Z. Y. Meng, T. Pruschke,
J. Moreno, and M. Jarrell. Dual-fermion approach to interacting disordered fermion
systems. Phys. Rev. B, 89:195116, May 2014
A first test for our algorithm is the Anderson-Falicov-Kimball model, which was
introduced in section 2.2.1 and has the Hamiltonian
It is a good test case, because the solution of the impurity problem is trivial and thus
the computational cost is low. The particle-particle interaction is static and leads
to a vertex function that depends only on two frequencies. Dynamical mean-field
theory (DMFT) gives a temperature-independent solution. This is because DMFT
for the Anderson-Falicov-Kimball model at half-filling is equivalent to DMFT for the
Anderson model, which has no dynamics and is thus temperature-independent. This
is not in accordance with results from the dynamical cluster approximation (DCA).
Thus, it is clear that dual fermions have to recover a temperature-dependent solution
to be considered an improvement over DMFT.
First, the 1d system is addressed. Comparisons with DCA calculations are easily
done because DCA is relatively cheap in 1d. We examine the local Green function
to observe how well the dual fermion approach performs compared to DCA and to
get a first idea of the influence of disorder. Next, we present results for the relative
corrections compared to DMFT results for the local Green function. The corrections
allow us to identify in which parameter region non-local correlations are important.
Second, we move on to the 2d system and calculate a phase diagram on the U V
plane. To this end, we calculate the difference
The sign of ImδG indicates a qualitative change in the local Green function from
metallic behavior to more insulator-like behavior. This quantity is not suited to
85
6 Test Case for the Dual Fermion Algorithm: The Anderson-Falicov-Kimball Model
0.0
-0.4
-0.6
ImGloc
Figure 6.1: The local Green function obtained from DMFT and dual fermions for U =
0.8W and various temperatures. Results for the clean system are shown on
the left, and results for disorder with V = 1.0W on the right. For the lowest
temperature T = 0.01W we added DCA results for comparison. (Data from
[106])
study the Anderson transition as it does not distinguish between localized and ex-
tended states. Box disorder cannot open a gap in the spectrum and thus ImδG
cannot become negative. We supplement the Green function data by the conduc-
tivity and find that it is strongly reduced for large values of the disorder strength
V.
In the following, all results were obtained for half-filling of the system with c and
f electrons. The disorder is distributed according to a box probability distribution
pBox (vi ) = V1 Θ( V2 − |vi |), where V denotes the strength of the disorder. In plots and
equations, DF stands for the dual fermion approach, DF-2nd represents the second
order approximation and DF-FLEX the dual fermion FLEX, which is sometimes
reduced to FLEX in the text. The dispersion relation reads k = −2t di=1 cos( 2π
P
k ),
Li i
where t = 0.25 is the hopping constant, d the dimensionality, Li the linear dimension
of the lattice in direction i and ki the ith component of the momentum vector.
In [135] the dual fermion approach has been applied to the clean Falicov-Kimball
model.
86
6.1 Dual Fermion Corrections in 1d
1.0
T=0.01 V=0.0W, DCA T=0.02 V=0.0W, DF-2nd
0.8 V=1.0W, DCA V=1.0W, DF-2nd
V=0.0W, DF-2nd V=2.0W, DF-2nd
V=1.0W, DF-2nd
0.6
σ(Gloc)
0.4
0.2
0.0
-0.2
0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
U/W U/W
Figure 6.2: Relative corrections to the DMFT result for the local Green function at
ω = iπT . The results on the left are for T = 0.01W and are compared to
DCA results. On the right, the results are for T = 0.02W and show how the
corrections from dual fermions change with increasing disorder strength. (Dual
fermion data from [106].)
be inferred from ImGloc (iπT ) going to zero for decreasing temperature. The DMFT
solutions are essentially on one curve, indicating that the result is temperature-
independent. This is in contrast with the dual fermion results which show a clear
temperature-dependence.
To obtain an estimate for the quality of the results we add a DCA solution for
Nc = 8 for the lowest temperature T = 0.01W . The dual fermion and DCA results
are essentially the same. These results show that the temperature dependence in the
Falicov-Kimball model at half-filling is strongly linked with non-local correlations
[106].
For the interacting system with disorder, ImGloc is shown in the right panel of
fig. 6.1. The DMFT result is again essentially temperature-independent and shows
metal-like behavior. For the dual fermions, the low-temperature result shows insu-
lating behavior and for increasing temperatures we observe a transition to a metallic
solution.
The DCA solution for the disordered interacting system shows the same trend as
the dual fermion result, namely an increase in the absolute value of ImGloc (iπT )
compared to the clean system, but the agreement with the dual fermion result is
not as good as for the clean system. However, the results underline the importance
of the non-local correlations for the physics of the Anderson-Falicov-Kimball model,
which appear even stronger in case of a disordered interacting system.
To get a better understanding of how and where non-local correlations change the
87
6 Test Case for the Dual Fermion Algorithm: The Anderson-Falicov-Kimball Model
88
6.2 2d Phase Diagram
1.4 0.72
1.2 0.60
0.48
1.0
0.36
0.8
V/W
Metal
0.24
0.6 0.12
00
0.
0.4 Bad metal 0.00
0.2 −0.12
0.0 −0.24
0.0 0.2 0.4 0.6 0.8 1.0
U/W
(a) DMFT
V/W
Metal Metal
0.24 0.24
0.6 0 0.6 0
0.0 0.0
0.12 0.12
0.4 Bad metal 0.00 0.4 Bad metal 0.00
0.2 −0.12 0.2 −0.12
Figure 6.3: Phase diagram on the U V -plane from ImδG from (a) DMFT, (b) 2nd order
dual fermions and (c) dual fermion FLEX for T = 0.025W . (Data from [106])
89
6 Test Case for the Dual Fermion Algorithm: The Anderson-Falicov-Kimball Model
1.4 1.08
1.2 0.96
0.84
1.0
0.72
0.8
V/W
0.60
80
0. 90 0.48
0
0.6
0.
0.5
0.36
0.4
0.24
0.2 0.12
0.0 0.00
0.0 0.2 0.4 0.6 0.8 1.0
U/W
(a) DMFT
0.96 0.9
1.4 1.4
0.8
1.2 0.84 1.2
0.7
0.72
1.0 1.0 0.6
0.60
0.8 0.8 0.5
V/W
V/W
0.48
0.6 0.4
0.6 0
80 0.36
0.8
0. 90 0.3
0.4 0. 0.24 0.4
0
0.2
0
0.5
0.5
Figure 6.4: Leading eigenvalue of the charge-density channel on the U V -plane from (a)
DMFT, (b) 2nd order dual fermions and (c) dual fermion FLEX for T =
0.025W . (data from [106])
90
6.3 Summary
for the charge density channel, where Γ is the irreducible vertex and χ0 the bare
lattice susceptibility. We find that the LEV is close to one around the metal-bad-
metal crossover. This shows that the bad metal in the vicinity of the clean system
is susceptible to the formation of a charge-density-wave.
ImδG is only a rough estimate, therefore we supplement it by calculating the dc
conductivity as well. We include vertex corrections, thereby taking into account
corrections to the bubble diagram, which contains the lowest order contributions.
The conductivity is shown in fig. 6.5 which reveals that the conductivity decreases
for large values of U and V , but always remains finite. This can be due to the
approximation, as well as the fact that at the temperature under consideration the
system is indeed merely a bad metal and not yet an insulator. We fix the critical
value of the conductivity by determining Uc for V = 0. The corresponding value of
the conductivity is σdc = 0.04. We find a smooth connection on the U V -plane from
the U -axis to the V -axis, thereby determining an estimate for the phase boundary in
the whole plane. For V W2 the phase diagram from σdc and ImδG are very similar
and show that small amounts of disorder help to mobilize the charge carriers. For
strong disorder and weak interaction the slopes of the lines of equal conductivity
are negative, indicating that interactions help to localize the c-electrons [106]. This
behavior is in contrast with the result from [41, 136]. Those results were obtained
within DMFT on the infinite dimensional Bethe lattice for the ground-state phase
diagram, whereas we work in 2d and for T > 0. More importantly, however, we
take into account non-local correlations through our dual fermion formalism. As we
have seen already in the previous sections, especially the latter can become quite
important for stronger disorder, thus explaining the difference. Let us stress that it is
possible that we nevertheless still miss the correct behavior as Anderson localization
is not a self-averaging effect.
6.3 Summary
The Anderson-Falicov-Kimball model served as a first test for the new formalism.
We started by looking at the local single-particle Green function and how it com-
pares to DMFT and DCA. We found that dual fermions significantly improve the
DMFT results and compare quite well with the DCA results. The best agreement
is achieved for the clean or weakly disordered system. For strong disorder dual
fermions and DCA still show qualitative agreement but quantitatively we observe
noticeable deviations, not just in value but also in the qualitative line shape of the
corrections.
For the Anderson-Falicov-Kimball model in 2d the dual fermion approach gives
good results for weak disorder. However, in the region of strong disorder it possibly
fails to produce the correct phase diagram. This is not completely unexpected as the
arithmetic averaging is not able to detect Anderson localization. Also, for strong
91
6 Test Case for the Dual Fermion Algorithm: The Anderson-Falicov-Kimball Model
0.64
1.4
0.56
1.2
0.48
1.0
0.40
0.8
V/W
Metal 0.32
04
0.
0.6 0.24
Bad metal
0.4 0.16
0.2 0.08
0.0 0.00
0.0 0.2 0.4 0.6 0.8 1.0
U/W
(a) DMFT
0.64 0.64
1.4 1.4
0.56 0.56
1.2 1.2
0.48 0.48
1.0 1.0
0.40 0.40
0.8 0.8
V/W
V/W
Metal Metal
04
0.32 0.32
4
0.0
0.
Figure 6.5: Phase diagram from the conductivity from (a) DMFT, (b) 2nd order dual
fermions and (c) dual fermion FLEX for T = 0.025W . (data from [106])
92
6.3 Summary
disorder the quantitative agreement between DCA and dual fermions is reduced,
which might indicate that dual fermions miss some important contributions.
For the Anderson-Hubbard model in the next section, the shape of the metallic
region for strong disorder is qualitatively different and does indeed meet the expecta-
tions based on the previous work in reference [137]. We suspect that the particularly
bad behavior of the DMFT, i. e. temperature independent Green functions, provides
a bad starting point for the dual fermions in case of the Anderson-Falicov-Kimball
model at half-filling. A bad starting point means that the small parameter for the
perturbative expansion is actually not as small as one would like. As the solution of
the impurity problem and the disorder-averaging look the same, it is possible that
the disorder vertex and the interaction vertex act together in a way such that the
effective expansion parameter is not small. However, we did not find a way to verify
that this conjecture is true.
93
7 Results for the Anderson-Hubbard
Model
In this chapter, we present results for the Anderson-Hubbard model that were ob-
tained using the dual fermion approach. The Anderson-Hubbard model was intro-
duced in section 2.2.1 and has the Hamiltonian (c. f. eq. (2.5))
1 1
(tij + µδij )(c†σi cσj + h.c.) −
X X X
HAH = − vi niσ + U (ni↑ − )(ni↓ − ). (7.1)
ij,σ i,σ i 2 2
In section 7.1 we compare the dual fermion results with dynamical cluster ap-
proximation (DCA) results in 1d to see how well dual fermions perform compared
to established cluster methods. First, we look at the imaginary part of the local
Green function ImGloc to get a general idea of the quality of the dual fermion ap-
proach. Second, we look at the relative corrections to the dynamical mean-field
theory (DMFT) result for ImGloc (iπT ). This quantity allows to show results for a
wide range of parameters in compact form.
In section 7.2 we continue with results for the 3d system. We opt for the 3d
system because it displays an Anderson transition, a finite temperature transition
from the paramagnet to the antiferromagnet and a Mott metal-insulator transition
for the paramagnetic Hubbard model. All three transitions are investigated within
the dual fermion approach. We find quantitative corrections as well as qualitative
ones compared to the DMFT results, e. g. for the temperature dependence of the
Mott metal-insulator transition in section 7.2.2.
All results were obtained for the half-filled system. If not mentioned otherwise, we
use binary disorder and second order dual fermions. The dispersion relation reads
k = −2t di=1 cos( 2π
P
k ), where t = 0.25 is the hopping constant, d the dimension-
Li i
ality, Li the linear dimension of the lattice in direction i and ki the ith component
of the momentum vector. In plots and equations, DF stands for the dual fermion
approach, DF-2nd represents the second order approximation and DF-FLEX the
dual fermion FLEX, which is sometimes reduced to FLEX in the text.
95
7 Results for the Anderson-Hubbard Model
-0.2
-0.5
ImGloc
-0.6
-0.7
Figure 7.1: Imaginary part of the dual Green function for βW = 20 and V = 1.0. On the
left we have U = 2W and on the right U = 0. For U = 0 the dual fermion
FLEX calculation did not converge.
7.1 1d Results
The one-dimensional case is the worst-case scenario for methods that retain a mean-
field character as non-local correlations become particularly important in 1d. Nev-
ertheless, it is a good test ground for the dual fermion approach, as it allows for a
comparison with DCA. This is because DCA calculations in 1d are relatively cheap
for reasonably large cluster sizes. The performance of DCA in 1d [138] is well known,
so we can estimate the quality of the dual fermion results. But what we are really
interested in is the comparison of the dual fermion approach with established cluster
approximations, as these are the competitors of our approach.
The goal is to establish a reliable alternative technique for simulations off half-
filling when quantum Monte Carlo based methods are needed, which is often the
case in more than one dimension. The expectation is that if we are able to capture
the DCA results with some accuracy in 1d, we should expect the same in higher
dimensions. Also, the overall quality of the approximation should improve in higher
dimensions as the importance of non-local correlations is reduced, and the results
become more mean-field like.
First, we take a look at the imaginary part of the local Green function ImGloc .
Figure 7.1 shows examples of ImGloc from various approximations for U = 2.0W
and V = 1.0W . The bandwidth is W = 1.0 in our units. Both dual fermion results
and the DCA result show insulator-like behavior in the left panel of fig. 7.1, i. e.
ImGloc has a minimum. Dynamical mean-field theory (DMFT) gives a metallic
solution. The plot on the right of fig. 7.1 shows results for the non-interacting
system. DMFT overestimates the effect of the disorder, showing an insulator-like
96
7.1 1d Results
0.8 0.7
DF-2nd DF-2nd
0.7 DF-FLEX
0.6 DF-FLEX
0.6 DCA,Nc = 12 0.5 DCA,Nc = 12
0.5 0.4
σ(Gloc)
σ(Gloc)
0.4 0.3
0.3 0.2
0.2 0.1
V = 0.5W , βW = 20
0.1 0.0
V = 0.0, βW = 20
0.0 −0.1
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
U/W U/W
(a) (b)
0.6 0.25
DF-2nd DF-2nd
DF-FLEX DF-FLEX
0.4 DCA,Nc = 12 0.20 DCA,Nc = 12
0.2 0.15
−σ(Gloc)
σ(Gloc)
−0.2 0.05
U = 0.0, βW = 20
−0.4 0.00
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
U/W V/W
(c) (d)
Figure 7.2: Relative correction as a function of U for multiple values of V for binary
disorder at βW = 20.
solution. Both second order dual fermions and DCA show a solution on the edge
between metallic and insulator-like behavior. There is no FLEX result, as the FLEX
calculation did not converge.
For a more systematic study of the effect of non-local correlations we compare
dual fermions and DCA by looking at the relative corrections compared to the
DMFT result according to eq. (6.3). The relative corrections are a measure for the
importance of non-local correlations. Results are shown in figs. 7.2 and 7.3.
In all plots results for 2nd order dual fermions, dual fermion FLEX and DCA with
Nc = 12 are shown. With increasing interaction strength (figs. 7.2a to 7.2c) the
correction increase until a maximum is reached. For the clean system (fig. 7.2a) and
the weakly disordered system (fig. 7.2b) the strongest correction are around 1.5W .
This is the region of the metal-insulator transition when non-local correlations are
included, whereas the transition happens in DMFT for larger values of U . With
97
7 Results for the Anderson-Hubbard Model
0.16 0.40
DF-2nd DF-2nd
0.14 0.35
DF-FLEX DF-FLEX
0.12 DCA,Nc = 12 0.30 DCA,Nc = 12
0.10 0.25
σ(Gloc)
σ(Gloc)
0.08 0.20
0.06 0.15
0.04 βW = 20, binary disorder 0.10
0.02 0.05 βW = 20, box disorder
0.00 0.00
−0.02 −0.05
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0
U/W = V /W U/W = V /W
(a) (b)
Figure 7.3: Relative correction of ImGloc (iπT ) with respect to the DMFT as a function of
U = V for (a) binary and (b) box disorder for βW = 20.
increasing disorder strength (from fig. 7.2a to fig. 7.2c) the maximal corrections are
reduced from about 80 percent to about 60 percent for the FLEX and shifted to
larger values of U from 1.4W to 2W .
Figure 7.2d shows corrections for the non-interacting system. The picture is the
same as for the interacting system but this time the maximal corrections occur
slightly below the bandwidth W . In figs. 7.2c and 7.2d the FLEX calculations did
not converge around V = W and weak interactions. This is the region where binary
disorder opens a gap. We did not observe this problem for the Mott gap and a possi-
ble explanation is that disorder opens the gap quicker than the Hubbard interaction
does, leading to stronger fluctuations that interfere with the FLEX calculation.
In fig. 7.3 results are shown for the special case U = V . For binary disorder the
maximal corrections we found were small, i. e. below 15 percent. At U = V =
4.5W the maximum has not yet been reached. For box disorder the corrections are
significantly larger, the maximal corrections are reached around U = V . 2W and
are about 35 percent.
The qualitative agreement between all three approaches is good, except for the
non-convergent FLEX results. In fig. 7.2 the FLEX results show good quantitative
agreement with the DCA results below the interaction strength UM of the maxi-
mal corrections. Above UM the 2nd order results show better agreement, but are
not as good as the FLEX for U < UM . Exceptionally good agreement is found
between FLEX and DCA for U = V in fig. 7.3. The results are nearly on top of
each other, whereas the 2nd order results show qualitative agreement but otherwise
underestimate the corrections, especially pronounced for box disorder.
We use the results in fig. 7.2 to compare the numerical efficiency. For the case of
binary disorder for U = 4W the computational effort to obtain the DCA result is
98
7.1 1d Results
roughly ten times higher than for the dual fermion FLEX. The reason is that for
binary disorder there are only two disorder configurations for dual fermions but 4096
configurations for DCA with Nc = 12. We sampled 200 random configurations for
DCA which led to a converged solution with respect to the number of configurations.
Still, the DCA data is much noisier than the dual fermion data. In case of box dis-
order, the computational effort becomes comparable. The dual fermion calculation
requires a larger number of configurations whereas DCA needs less configurations
compared to binary disorder as the different random configurations are more similar
to each other. In that case the efficiency depends on the specific parameters and the
number of iterations needed. It can be expected that dual fermions will gain an edge
over DCA as the temperature is lowered and the sign problem becomes more severe.
We had to abort the DCA calculations at smaller values of U = V as we were not
able to obtain converged results. This is probably caused by the sign problem.
In figs. 7.2c and 7.2d the weakness of the dual fermion FLEX approximation be-
comes apparent. In the case of binary disorder around V = W the FLEX calculation
does not converge. This is around the transition from a metallic local Green func-
tion to an insulating one. The failure is not apparent from the leading eigenvalue of
the Bethe-Salpeter equation. This underlines the fact that the FLEX approximation
has to be used with care. As we can see from the small U limit in fig. 7.2c the FLEX
result becomes much worse than the 2nd order approximation but still converges. It
is therefore advisable to always check the FLEX results against the 2nd order results
for consistency and look at surrounding data points to see whether any divergences
are close by.
The physical interpretation of the results is that disorder reduces the effect of non-
local correlations from the Hubbard term [24]. This is apparent from the reduced
maximal correction. Also, stronger interactions are necessary for maximal non-local
correlations. This result confirms our expectations. If the underlying lattice does
not show long-range order, it is natural to expect that the electrons on the lattice
do not show long-range order either.
We conclude that the dual fermion approach is able to recover non-local correla-
tions on a level that is comparable to DCA. Thus, we can expect good qualitative
and quantitative results in higher dimensions. If applicable, the FLEX approxima-
tion is the method of choice for small to intermediate values of U and weak disorder.
For larger values of U the 2nd order approximation gives results closer to DCA. For
a final evaluation comparisons with other quantum cluster methods are desirable,
but go beyond the scope of this thesis.
For the case of binary disorder the FLEX has to be handled with care as it
can show non-convergent behavior and give nonsensical results. We point out that
the dual fermion approach gives better results in the presence of disorder for the
Anderson-Hubbard model than for the Anderson-Falicov-Kimball model. The reason
is possibly that for the Hubbard model the DMFT solution is already a better
99
7 Results for the Anderson-Hubbard Model
7.2 3d Results
Our main interest for the Anderson-Hubbard model lies in the 3d case. Here, we
want to take a closer look at phase transitions. These are the antiferromagnetic
transition, Mott transition and Anderson localization.
First, for the antiferromagnetic transition we look at the leading eigenvalue of
M = Ṽm χ̄0 , which determines where the sum over the ladder diagrams for the
antiferromagnetic susceptibility diverge. We compare results obtained from DMFT,
2nd order dual fermions and dual fermions with FLEX.
Second, we address the Mott metal-insulator transition. At high temperatures
one observes a crossover, whereas for lower temperatures the transition is of first
order. This leads to a hysteresis for the double occupancy D as a function of U . We
investigate the effect of disorder and temperature on the hysteresis.
Third, for Anderson localization, we have two quantities we look at. These are
ImδG and the conductivity. ImδG is only applicable for binary disorder, because
box disorder will not open a gap in the spectrum. As we cannot detect localized
states from the arithmetically averaged density of states, we have to go beyond
single-particle quantities for box disorder. We will use the conductivity, which is a
defining quantity for an insulator. Therefore, it is supposed to work for both binary
and box disorder.
In this chapter vertex corrections for the conductivity have been obtained only
for the disorder component. In the region of the metal-bad-metal transition due to
interactions the vertex corrections become very noisy and the noise translates to the
conductivity, thus making calculations too costly for our purpose. Additionally, we
suspect that the vertex correction do not fulfill the Ward identities [139], which relate
the irreducible vertex with the self-energy. Instead, we relate the self-energy to the
full vertex via the Schwinger-Dyson equation. The reason that we keep the vertex
correction from disorder is that eq. (4.95) was originally introduced for disordered
systems and it worked well in [40].
100
7.2 3d Results
1.4
U=2.0
1.3 U=2.5
U=3.0
1.2
U=3.5
U=4.0
1.1
LEV
1.0
0.9
0.8
0.7
0.070 0.075 0.080 0.085 0.090 0.095 0.100
T
Figure 7.4: The LEV for the magnetic channel as a function of temperature for V /W = 1/3
and various values of U . To extract the transition temperature we make a linear
extrapolation to the data around LEV= 1.
101
7 Results for the Anderson-Hubbard Model
0.05 0.05
V/W=0.0 V/W=0.0
0.04 DMFT V/W=1/3 0.04 DF-2nd V/W=1/3
V/W=2/3 V/W=2/3
0.03 0.03
TN /W
TN /W
0.02 0.02
0.01 0.01
0.00 0.00
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
U/W U/W
0.02
0.01
0.00
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
U/W
(c) DF-FLEX
Figure 7.5: The phase diagram on the U T plane for various values of V for binary disorder.
In (a) the result for the DMFT calculation is shown, (b) shows the 2nd order
dual fermion result and (c) shows the dual fermion FLEX result. All three
methods agree qualitatively. Disorder suppresses antiferromagnetism for small
values of U and enhances it for large values of U . The effect of non-local
correlations is to reduce the transition temperature for intermediate and large
values of U . This effect is stronger for the FLEX calculation than for the 2nd
order dual fermion results.
102
7.2 3d Results
where λ is a parameter that depends on the disorder distribution. This shows that
weak disorder increases TN for large U and U V and it qualitatively explains the
phase diagram that we observe. The DMFT result does indeed show an increase
from 2.7% to 10% for the relative increase of TN when V is doubled. We do not
observe this for the dual fermion results. The results show some noise for large values
of U and the relative change is very small such that the quantitative agreement is
not very good. Also, U V is not fulfilled, which may explain why we cannot see
this behavior.
103
7 Results for the Anderson-Hubbard Model
0.045
0.09 V/W=0.0 V/W=0.0
0.08
V/W=1/3 V/W=1/3
0.040
V/W=2/3 V/W=2/3
0.07
0.035
D
D
0.06
0.05 0.030
0.04
DF-2nd, βW = 120 0.025 DMFT, βW = 120
0.03
0.7 0.8 0.9 1.0 1.1 1.2 0.9 1.0 1.1 1.2 1.3 1.4
U/W U/W
Figure 7.6: Hysteresis at βW = 120 for binary disorder and various values of V . On the
left, the 2nd order dual fermion result is shown, the right figure shows the
DMFT result. In both cases disorder shifts the hysteresis to larger values of
U and reduces the area of the hysteresis. The effect of non-local correlations
is to move the hysteresis to smaller values of U and increase the area of the
hysteresis. Note the different scaling of the D axis in both figures to see the
increase in the area for the green and red curves. (DMFT data by S.-X. Yang)
104
7.2 3d Results
0.12 0.055
V=0 βW = 120 V=0,DMFT βW = 120
0.10 0.050
βW = 180 βW = 180
0.045
0.08
0.040
0.06
D
D
0.035
0.04
0.030
0.02 0.025
0.00 0.020
0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76 0.78 0.90 0.92 0.94 0.96 0.98 1.00 1.02 1.04
U/W U/W
Figure 7.7: The dual fermion result on the left for the hysteresis shows a decreasing Uc1
with decreasing temperature. This is in contrast with the DMFT result on the
right, where Uc1 increases with decreasing temperature. (DMFT data by S.-X.
Yang)
0.12 0.12
V/W=1/6 βW = 120 V/W=1/3 βW = 120
0.10 0.10
βW = 180 βW = 180
0.08 0.08
0.06 0.06
D
0.04 0.04
0.02 0.02
0.00 0.00
0.65 0.70 0.75 0.80 0.85 0.80 0.82 0.84 0.86 0.88 0.90 0.92
U/W U/W
Figure 7.8: Double occupancy for V /W = 1/6 and V /W = 1/3. The qualitative picture
from fig. 7.7 does not change with disorder.
105
7 Results for the Anderson-Hubbard Model
106
7.2 3d Results
V/W
0.024 0.30
0.4 0.4
Metal Metal 0.24
0.3 0.000 0.3
0.18
4
00
0.0
0.2 Bad 0.2 Bad
0.
−0.024 0.12
metal metal
0.1 0.1 0.06
−0.048
0.0 0.0 0.00
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
U/W U/W
(a) (b)
Figure 7.9: (a) ImδG and (b) the conductivity per spin as function of U and V for βW =
60. For both quantities we find the same general shape of the phase diagram.
For such a high temperature, the elevated region can be interpreted as a metal
whereas the lower region as a bad metal.
V/W
0.032 0.30
−0.016 0.12
0.
Bad Bad
0.2 0.2
metal −0.032 metal 0.06
0.0 −0.048 0.0 0.00
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
U/W U/W
(a) (b)
Figure 7.10: (a) ImδG and (b) the conductivity per spin as function of U and V for βW =
60. In this case, ImδG and the conductivity show similar behavior close to the
clean system. For stronger disorder the difference is quite noticeable because
for box disorder ImδG does not show any signs of Anderson localization.
107
7 Results for the Anderson-Hubbard Model
in qualitative agreement with the phase diagram in [137, 145]. However, the critical
disorder strength VA does not match the TMT and TMDCA results. For binary
disorder, the estimate is relatively close to the numerical solution of [144]. Using
the conductivity worsens the result in this case.
For box disorder, VA cannot be estimated from ImδG. Our estimate from the
conductivity is too small, but improved results for the conductivity on the U axis
may help to improve the estimate for VA and for the overall phase diagram.
If precise values for the critical disorder strength are needed, typical medium based
methods may provide better results that can be used to obtain better estimates for
the defining conductivity for the metal-insulator or metal-bad-metal transition.
7.3 Summary
In this chapter we applied the dual fermion formalism to the Anderson-Hubbard
model. We started with a 1d system in section 7.1, where the dual fermion approach
was compared to DCA. Both methods gave qualitatively comparable results for
the non-local corrections to DMFT. The quantitative agreement depends on the
approximation used to solve the dual fermion problem. The FLEX approximation
tends to provide better results if it is applicable.
We continued with the 3d system in section 7.2 and looked at various phase transi-
tions. In section 7.2.1 we investigated the effect on the antiferromagnetic transition.
From [105] it is known, that dual fermions are capable of giving good approximate
results for the transition temperature. The addition of weak disorder suppresses an-
tiferromagnetism for small values of U and enhances it for large values. This result
was know from DMFT [24], dual fermions provide a quantitative improvement and
possibly a qualitative improvement for the large U limit. Unfortunately we were not
able to study this limit as the computational cost becomes prohibitive.
Next, we investigated the effect of disorder on the Mott metal-insulator transition.
To this end, we calculated the hysteresis of the double occupancy. The results show
that disorder moves the critical values Uc1 and Uc2 to larger values and reduces the
area of the hysteresis, indicating that for strong disorder the nature of the phase
transition changes, possibly to a crossover. The effect of dual fermions is two-fold.
For fixed temperature both Uc1 and Uc2 are decreased and the area of the hysteresis
increases, but otherwise the qualitative picture stays the same.
A qualitative change appeared when the temperature dependence was investi-
gated. DMFT predicts that Uc1 and Uc2 increase with decreasing temperature. Our
dual fermion results show that Uc1 decreases with decreasing temperature, which is
qualitatively different from the DMFT result. At least in the case of disorder Uc2
increases with decreasing temperature, which indicates that the upper critical value
of U is governed by mean-field behavior.
Finally, the Anderson transition was addressed in section 7.2.3. Although the dual
108
7.3 Summary
fermions do not distinguish between localized and extended states, we were able to
determine values for VA for binary and box disorder. In the case of box disorder, the
results for VA do not agree very well with the results for VA from typical medium
based methods, which have proved to be good approximations to the exact solution
at T = 0. This indicates that dual fermions have trouble dealing with non-self-
averaging effects. Additionally, we calculated the phase diagram on the U V plane,
which agrees qualitatively with the results of [143, 145].
We conclude that dual fermions add non-trivial modifications to the DMFT re-
sults. Quite often these are of a quantitative nature, but when non-local correlations
are particularly important the changes can be qualitative as well, as we saw for the
temperature dependence of the hysteresis.
109
8 Summary and Perspective
The main goal of this thesis was the development of a numerical method that treats
disorder and interaction on equal footing and goes beyond the dynamical mean-field
approximation. To this end we started from the dual fermion approach which had
been developed for interacting systems and which had been applied to disordered
systems. We combined both approaches to treat disordered interacting systems.
Disorder was modeled with a local random potential that did not affect the hop-
ping matrix elements. To obtain answers for macroscopic systems, i. e. systems
which are self-averaging, we employed disorder-averaging.
In the derivation of the dual fermion formalism, the replica trick was used to
perform the disorder-average on powers of the partition function Z instead of the
free energy F = −T ln Z. As a result, the disorder-average became part of the
impurity problem for the dual fermions and no disorder-average had to be done for
the dual degrees of freedom. The replica trick introduced restrictions for the dual
diagrams, thereby eliminating excitations for the dual degrees of freedom that are
unphysical for the real degrees of freedom.
First, the dual fermion formalism was applied to the Anderson-Falicov-Kimball
model. It is a good test ground, because the impurity problem for the dual fermion
approach is trivial, but the model itself has non-trivial features like long-range order.
The simplicity of the impurity problem manifests itself in the two-particle vertex
function that only depends on two frequencies instead of three, as is the case for the
Anderson-Hubbard model. Also, the DMFT solution for the single-particle Green
function is essentially temperature independent. As our DCA calculations showed
in 1d, this is not the correct answer but an artifact from ignoring the non-local
correlations. We found that the dual fermion approach is able to reintroduce the
correct temperature dependence.
For the 2d Anderson-Falicov-Kimball model we calculated the phase diagram on
the U V plane at T = 0.025W , where U parametrizes the local Coulomb repulsion
and V the disorder strength. We found a metal-bad-metal transition for large values
of U and for large values of V , albeit the shape of the metallic region does not
coincide with results obtained by other groups [41, 136]. We include non-local
correlations whereas [41, 136] use geometric averaging, which is better suited to
observe Anderson localization. It is thus hard to tell which result is correct. Possibly
TMDCA [144] could provide an answer. The case of strong disorder and weak
interaction poses the main challenge for the dual fermion approach as the physics
becomes non-self-averaging.
111
8 Summary and Perspective
The driving force of the metal-bad-metal transition for weak disorder and strong
Coulomb interactions is the formation of a charge density wave, which we learned
from the leading eigenvalue of the Bethe-Salpeter equation. This is in accordance
with reference [136, 146, 62]. On a bipartite lattice for interaction and weak disorder
the ground state of the Anderson-Falicov-Kimball model is a checkerboard ordering
of the c and f electrons. For the Anderson-Falicov-Kimball model results were shown
exclusively for box disorder.
Second, the dual fermion approach was applied to the Anderson-Hubbard model,
which is our model of interest. It has true quantum correlations, i. e. the two-particle
vertex has the full dependence on three frequencies. Unless otherwise stated, all
results were obtained for binary disorder. Again, we started with a comparison with
DCA in 1d. The agreement is better than for the Anderson-Falicov-Kimball model,
which might be due to the fact, that DMFT, which acts as a starting point for the
dual fermion approach, provides a better, temperature dependent solution for the
lattice problem.
The comparison with DCA is followed by results for the 3d system, which shows
a variety of phase transitions. Emphasis was put on those phase transitions. We
started with the phase diagram on the U T plane and distinguished two phases,
namely paramagnetic metal and antiferromagnetic insulator. Results were shown
for different values of the disorder strength V . DMFT, 2nd order dual fermions
and dual fermion FLEX were used to obtain the phase diagram. All three methods
gave the same shape of the phases and the effects of disorder were the same as
well, specifically the reduction of antiferromagnetism for weak interaction and the
enhancement of antiferromagnetism for strong interactions. In comparison with
DMFT, 2nd order dual fermions reduced the critical temperature and dual fermion
FLEX even more so. This was known for the clean system [105], and is assumed to
be the correct behavior, based on DCA and quantum Monte Carlo studies [147, 140].
We continued with the Mott transition. To this end, we calculated the hysteresis
of the double occupancy, which was approximated by the double occupancy of the
impurity problem. As the Mott transition is concealed by the antiferromagnetic
phase in the three dimensional unfrustrated system, the leading eigenvalue of the
Bethe-Salpeter equation is greater than one, thus dual fermion FLEX could not be
used. Therefore, we only obtained results for 2nd order dual fermions and DMFT.
At equal temperature, DMFT and dual fermions showed the same qualitative be-
havior. Disorder increases the critical interaction strength and reduces the area of
the hysteresis. Dual fermions lower the critical interaction strength and increase the
area of the hysteresis compared to DMFT. This is due to short-range antiferromag-
netic correlations that are not present in DMFT. It would be interesting to extend
the calculation to the strongly frustrated system and see the effect of FLEX on the
Mott transition.
A qualitative difference between dual fermions and DMFT was found when we
112
investigated the temperature dependence of the hysteresis. In DMFT the upper
and lower critical value of the interaction strength are increased with decreasing
temperature. With dual fermions we found that at least the lower critical interaction
strength decreases. Our data was too noisy to extract unambiguous results for the
upper critical value for the clean system. Upon the introduction of disorder the
upper critical value is clearly increased with decreasing temperature, thus showing
mean-field behavior. We expect this for the lower critical value as well when disorder
is strong enough, but we were not able to obtain data at sufficiently strong disorder
as increasing disorder meant calculations at large values of U , unaccessible to us.
Lastly, Anderson localization was investigated. The main focus was placed on
the question of how to detect it and how it compares with established methods like
typical-medium-theory or the newer and very promising TMDCA [148]. For detec-
tion, we used the imaginary part of the local Green function, namely the difference
between the lowest two frequencies, which determines whether the imaginary part
has a minimum or not. This criterion only worked for binary disorder. Therefore, we
calculated the conductivity as well, which allowed to determine a phase boundary.
Our results for the shape of the metallic region agree with results obtained previously
[143]. Unfortunately, our estimate for the critical disorder strength for box disorder
does not match results obtained within typical medium theory and TMDCA. This
is not completely unexpected, as Anderson localization is a non-self-averaging effect
and thus the disorder-average we used is not suited to detect it.
Our results from both the Anderson-Falicov-Kimball model and the Anderson-
Hubbard model suggest that dual fermions add non-trivial corrections to the DMFT
result and do so on a level that is qualitatively and oftentimes quantitatively com-
parable with DCA. Hence, we believe that the formalism developed in this thesis
can help further grow the understanding of disordered interacting systems. For ex-
ample, it is an interesting topic, how disorder influences superconductivity. Is it
possible, that disorder enhances superconductivity in certain cases as it does with
antiferromagnetism and can this be quantified?
A particularly interesting case consists in the application to real materials. For this
purpose, the dual fermion formalism has to be combined with first-principle meth-
ods like density functional theory (DFT)[149, 150]. In order to include quenched
disorder, it has to be worked out, whether it is possible to start from a DFT solution
for a clean system and add disorder only at the impurity and dual fermion level or
whether disorder has to be treated already on the DFT level. In the latter case, the
interfacing between the DFT part and the dual fermion part has to be addressed.
Finally, we want to propose a further extension of the dual fermion formalism.
In [151] experiments on layered f -electron systems are presented. These layered
systems have been investigated by means of inhomogeneous or real-space DMFT
(rDMFT) [152]. To this end, the dual fermion formalism for the periodic Anderson
model could be used to solve individual layers. The coupling between different layers
113
8 Summary and Perspective
114
Appendices
115
A The Sign Problem in Quantum
Monte Carlo Simulations
The following discussion is based on the explanation of the sign problem in [25].
Quantum-Monte Carlo methods are very powerful tools if applicable. Although
many algorithm scale at least with (U Nc β)3 , calculations at lower temperatures and
larger cluster sizes will be achievable as long as computers become faster. However,
if it is impossible to retain positive weights for all Monte Carlo configurations for a
given system, Quantum-Monte Carlo methods hit a hard wall. In the following we
will discuss why that is.
We start from the partition function
Z = Tre−βH =
X
wC . (A.1)
C
P
C denotes a sum over configurations as discussed is chapter 5 and wC are the
corresponding contributions to the partition function. The form of wC depends on
the algorithm that is used.
Now we want to use Monte Carlo methods to sample the configurations C. This
requires that wC ≥ 0. We will assume that this is not the case for our problem and
that there is no known trick to achieve this. In that case, we can write
X X
Z= wC = sgn(wC )|wC |, (A.2)
C C
and we can use |wC | as weights for the Monte Carlo simulation. For the expectation
value of an observable O we have
1X X hsgnOi|w|
hOi = OC wC = sgn(wC )OC |wC | = . (A.3)
Z C C hsgni|w|
117
A The Sign Problem in Quantum Monte Carlo Simulations
sgn(wC )|wC |
P
C
hsgni|w| = . (A.4)
C |wC |
P
The nominator is just the partition function of our system, the denominator is a par-
tition function for a system with only positive weights. According to thermodynam-
ics, the partition function can be written in terms of the free energy F = −kB T ln Z
where 1/β = kB T . kB is the Boltzmann constant and T temperature. It follows
that
hsgni|w| = e−β∆F . (A.5)
We know that hsgni|w| ≤ 1, therefore ∆F ≥ 0. This means the average sign becomes
exponentially smaller for increasing β and for increasing ∆F . When we approach
the thermodynamic limit, i. e. all relevant correlations are included in the cluster
for the Monte Carlo simulation, we will have ∆F ∝ Nc . This means the sign scales
exponentially in β and NC . This is still no unsolvable problem, all we need is the
corresponding numerical accuracy. This costs time but in principle it should be
possible.
Next, we look at the variance of the sign:
after N Monte Carlo measurements. This means that in order to obtain the sign with
reasonable accuracy the number of Monte Carlo measurements has to be increased
exponentially. This is the reason the sign is a problem. The time needed to solve a
given problem with the needed accuracy scales exponentially.
118
B Fast Matrix Updates for Quantum
Monte Carlo
For the ctQMC random walk it is necessary to calculate the ratio of two determi-
nants. For arbitrary matrices the calculation of the determinant is of order N 3 and
such would be the time to calculate the ratio. However, the smaller of our two
matrices appears in the larger one. This can be expressed as [25]
!
(n+1) N (n) Q
N = , (B.1)
R S
where N (n) is a matrix of rank n. For the cases of importance in this work, S is a
number, Q a column vector and R a row vector. The inverse of N (n+1) is
!
(n+1) −1 P̃ Q̃
(N ) = . (B.2)
R̃ S̃
which is verified by direct calculation. If one of the off-diagonal blocks is zero, the
determinant is !
A B
det = det A · det D. (B.4)
0 D
The matrices S̃, Q̃, R̃ and P̃ are given by [131]
119
B Fast Matrix Updates for Quantum Monte Carlo
det N (n+1) 1
(n)
= det(S − R(N (n) )−1 Q) = (B.9)
det N det S̃
for the matrix ratios. In the ctQMC algorithm N −1 is stored, therefore eq. (B.9)
is particularly cheap for the removal of a vertex. It is only necessary to look up
S̃, which is O(1). For the addition of a vertex, matrix-matrix multiplication is
necessary, which is O(n2 ). The same holds true for the update of the matrix.
120
C Dual Fermion Mapping for the
Anderson-Falicov-Kimball Model
The derivation of the dual fermion approach for the Anderson-Falicov-Kimball model
is similar to the Anderson-Hubbard model. Instead of spin up and down we have the
c- and f -degrees of freedom. Here, we closely follow the derivation that we published
in [106]. In the following, the f -degrees of freedom will be denoted as cf so they are
not confused with the dual degrees of freedom f . We start the derivation from the
disorder-averaged Green function
δ
{Gk } = − {ln Z v [ηk ]}|ηk =0 , (C.1)
δηk
with Z β Z β
Siv [c̄i , ci ; c̄fi , cfi ] = dτ vi ni (τ ) + dτ U nci (τ )nfi (τ ). (C.4)
0 0
We rewrite the action in terms of an impurity action
Siv,imp [c̄i , ci ; c̄fi , cfi ] = − c̄ωi [(ω + µ) − ∆ω ]cωi + Siv [c̄i , ci , c̄fi , cfi ].
X
(C.5)
ω
121
C Dual Fermion Mapping for the Anderson-Falicov-Kimball Model
in terms of the impurity action Siv,imp . Note that the hybridization function drops
out if we insert the impurity action.
Next, we apply the replica trick [81, 82]
hZ m idis − 1
hln Zidis = lim (C.7)
m→0 m
and the Green function now reads
δ Z Y
hGk idis = − dvi p(vi ) ln Z iv
δηk i
(C.8)
1 δ Z Y Z Z Y
v
= − lim dvi p(vi ) D[cα , c̄α ]D[cfα , c̄fα ]eSα .
m→0 m δηk
i α
The f -numbers are Grassmann numbers, which follow fermionic commutation rela-
tions. For A we obtain
A2ωk = ∆ω − k − ηωk , (C.13)
and λω remains arbitrary as in the derivation for the Anderson-Hubbard model.
After the Hubbard-Stratonovich transformation the Green function reads
1 δ Y ∆ω − k − ηωk m
Gωk = − lim ×
m→0 m δηωk
ωk λ2ω
Z Z (C.14)
λ2 f¯α (∆ω −k −ηωk )−1 fωk
P α
P
Si
× Df¯Df e − ωkα ω ωk Dc̄Dce − i site .
ηωk =0
122
The c- and cf -degrees of freedom both appear only locally and can be integrated
out. Now, the local part of the action reads
i i
λω (c̄αωi fωi
α
+ f¯ωi
α α
X
Ssite = Simp + cωi ) (C.15)
ωα
with
XZ β
i
c̄αωi [(ω ∆ω ]cαωi dτ U nc,α f,α
XX
Simp =− + µ) − − W (ñi ) + i (τ )ni (τ ). (C.16)
α ω α 0
Note that only the c-degrees of freedom couple to the f -degrees of freedom. This is
because the cf -degrees of freedom only appear locally in the Hamiltonian eq. (2.6)
and are thus part of the impurity problem.
We can now write down the defining equation for the dual potential which reads
Z Y
λ2ω gimp (ω)f¯ωα fωα −Vd [f¯α ,f β ]
P
dc̄α dcα dc̄f,α dcf,α e−Ssite = Zimp e− ωα (C.17)
αω
(∆ − k )−1 + gω α
Sd0 = f¯ωk
α
fωk . (C.22)
gω2
123
C Dual Fermion Mapping for the Anderson-Falicov-Kimball Model
For non-interacting dual fermions this gives the DMFT result, which therefore ap-
pears as a zeroth order approximation in this formalism.
124
D Two-Particle Green Functions
The following definitions are taken from [153, 105]. The two-particle Green functions
are defined via functional derivatives of the free energy as
δ 4 F [K̄, K]
χph = (D.1)
δK4 δ K̄3 δK2 δ K̄1 K̄1 =K2 =K̄3=K4=0
δ 4 F [K̄, K]
χpp = (D.3)
δK4 δK3 δ K̄2 δ K̄1 K̄1 =K̄2 =K3=K4=0
For the particle-hole Green function the vacuum term is subtracted which follows
from the chain rule. For the particle-particle channel the chain rule leads to hccihc† c† i
which is zero for all models in this thesis.
We obtain the Fourier transformed two-particle quantities for the particle-hole
Green function via
Z β Z β Z β Z β
ph
χ (ω1 , ω2 ; ω3 , ω4 ) = dτ1 dτ2 dτ3 dτ4 χ(τ1 , τ2 ; τ3 , τ4 )ei(ω1 τ1 −ω2 τ2 +ω3 τ3 −ω4 τ4 ) .
0 0 0 0
(D.5)
Time-translational invariance requires ω1 +ω3 = ω2 +ω4 . This means χ only depends
on three frequencies, two fermionic ones and one bosonic one, which is the difference
ν := ω1 − ω2 = ω4 − ω3 of two fermionic frequencies. ν denotes the total energy of
the particle-hole pair, 1(4) denotes the initial (final) electron, 2(3) the initial (final)
hole. We use time-translational invariance to introduce the three frequency notation
[105]
χph (ν)ω,ω0 := χph (ω + ν, ω; ω 0 , ω 0 + ν), (D.6)
which is shown in fig. D.1. The particle-particle Green function is Fourier trans-
125
D Two-Particle Green Functions
ω′ + ν ω+ν
χph
↑↑;↑↑ (ν)ωω ′ =
ω′ ω
ω′ + ν ω+ν
χph
↑↑;↓↓ (ν)ωω ′ =
ω′ ω
Figure D.1: The two particle-hole Green function shown here are all Green functions
needed for an SU(2) symmetric system.
ω′ + ν ω+ν
χpp
↑↓;↑↓ (ν)ωω ′ =
−ω ′ −ω
Figure D.2: In the particle-particle channel we need only the pairing channel Green func-
tion. We can construct all other particle-particle Green functions from it.
formed via
Z β Z β Z β Z β
pp
χ (ω1 , ω2 ; ω3 , ω4 ) = dτ1 dτ2 dτ3 dτ4 χ(τ1 , τ2 ; τ3 , τ4 )ei(ω1 τ1 +ω2 τ2 −ω3 τ3 −ω4 τ4 ) .
0 0 0 0
(D.7)
In this case the bosonic frequency is defined as ν := ω1 + ω2 = ω4 + ω3 . 1 and 2 are
the initial electrons, 3 and 4 the final electrons. The three frequency notation is
which is illustrated in fig. D.2. The difference in sign for the frequencies in the
particle-hole and particle-particle channel comes from the convention that a particle
with frequency label ω carries the energy ω whereas a hole with the frequency label
ω carries the energy −ω. The total energy is the sum of the top and the bottom
where a hole introduces an additional minus sign. If ν is to denote the total energy
of the particle-hole (particle-particle) pair we obtain the notation as as shown above.
126
D.1 Channels
D.1 Channels
For the two-particle Green function we have not only the particle-hole and particle-
particle channel, but we have six different possible spin configurations as well. For
the particle-hole channel these are (↑↑; ↑↑), (↓↓; ↓↓), (↑↑; ↓↓), (↓↓; ↑↑), (↑↓; ↓↑) and
(↓↑; ↑↓). The first four have sz = 0 for the initial and final state, whereas the last
two have sz = 1 and sz = −1. A hole of spin up is effectively a down spin because
the lack of an up spin leaves an uncompensated down spin.
We have to measure these channels using quantum Monte Carlo. It is possible
to exploit symmetries to reduce the number of channels we have to measure down
to two for the particle-hole channel. This reduces the memory needed to store the
vertex function and the time needed to measure the particle-hole Green functions
by a factor of three.
First we note that the last two configurations have a total spin S = 1. There is
one state missing for a triplet. We can create this state via
1
χm0 = √ (χph ph ph ph
↑↑;↑↑ + χ↓↓;↓↓ + χ↑↑;↓↓ + χ↓↓;↑↑ ) (D.9)
4
which defines the magnetic channel m with sz = 0. For the magnetic channel the
initial and final states have a total spin S = 1. we can see this by symbolically
writing
1 1 1
√ (↑↑; ↑↑) + (↓↓; ↓↓) + (↑↑; ↓↓) + (↓↓; ↑↑) = √ (↑↑) + (↓↓) √ (↑↑) + (↓↓) .
4 2 2
(D.10)
The initial and final states are the third triplet state.
We can create a singlet state with S = 0 according to
1
χd = √ (χph ph ph ph
↑↑;↑↑ + χ↓↓;↓↓ − χ↑↑;↓↓ − χ↓↓;↑↑ ). (D.11)
4
We call this channel the density channel d. In the symbolic notation from above we
obtain
1 1 1
√ (↑↑; ↑↑) + (↓↓; ↓↓) − (↑↑; ↓↓) − (↓↓; ↑↑) = √ (↑↑) − (↓↓) √ (↑↑) − (↓↓) .
4 2 2
(D.12)
The initial state and final state have indeed a total spin S = 0.
We can now exploit the symmetry between spin up and down. Two channels that
are related by flipping all spins are equal because of the spin symmetry. For the
density channel d and the magnetic channel m0 we obtain
127
D Two-Particle Green Functions
Furthermore, SU(2) symmetry demands that χm1 = χm0 = χm−1 . This reduces the
number of independent particle-hole channels to two. In this thesis we work with
χd/m0 and measure the two particle-hole Green functions on the right hand side of
eq. (D.13).
For the particle-particle channel we have six possible spin configurations as well.
These are (↑↑; ↑↑), (↓↓; ↓↓), (↑↓; ↑↓), (↓↑; ↓↑), (↑↓; ↓↑) and (↓↑; ↑↓). The first two
have a total spin S = 1 for the inital and final state, the last four have sz = 0 but
they are no eigenstates of the square of the total spin operator.
Again, we can exploit symmetries to reduce the number of channels we have to
measure. We start from the pairing channel Green function
χp (Q)ωω0 = χpp
↑↓;↑↓ (Q)ωω 0 . (D.14)
The corresponding diagram is shown in fig. D.2. The pairing channel is related to
the singlet/triplet channel via
From the spin symmetry we get the (↓↑; ↓↑) channel. The crossing symmetry relates
the pairing channel with χpp↑↓;↓↑ . We proceeded the same way as we did for the
particle-hole channel to obtain eq. (D.15) and find the singlet and triplet channels.
In this case, we only have to measure the particle-particle Green function for the
pairing channel.
The disorder two-particle Green function is defined via
χv (ν)ωω0 = χ=
v (ν)ωω 0 = {gω+ν gω } (D.16)
which connects the horizontal channel with the vertical channel. The vertical chan-
128
D.2 Vertices
nel is unphysical, but can be convenient to use in some calculations. We find this
ω′ + ν ω+ν ω ω+ν
→ −
ω′ ω ω′ ω′ + ν
Figure D.3: Relation of the horizontal channel and the vertical channel for the disorder
vertex.
relation from the crossing symmetry by switching the upper right corner of the two-
particle Green function with the lower left one as is shown in fig. D.3. For the
particle-particle channel we relate the crossed channel to the horizontal one via
χ× =
vp (ν)ω,ω 0 = −χvp (ν)ω,−ω 0 −ν δω+ω+ν,0 . (D.19)
a) b)
D.2 Vertices
χph
d/m = - - Fd/m
The full vertex is obtained from the two-particle Green function via the Bethe-
Salpeter equation
129
D Two-Particle Green Functions
χp = - Fp
for the particle-particle channel. It follows that we have to remove the disconnected
part first, and then remove the four Green function legs by division.
For the particle-hole channel we obtain
χ=
v (ν)ωω − gω gω+ν
Fv= (ν)ωω0 = −β δωω0 . (D.24)
gω+ν gω gω gω+ν
As a summary, in fig. D.7 all the vertices we need in this thesis are presented with
their diagrammatic representation. Here, Fd/m/p have been replaced in the diagrams
by graphic symbols.
Fp = =
Fvp = ×
Fvp =
Figure D.7: Definition of the different vertices in terms of diagrammatic building blocks.
The solid green lines for the disorder vertices indicate for which connection
from corner to corner the energy is conserved.
130
E Comment on the Use of the
Replica Trick
In several discussions, the question was raised whether the replica trick is necessary
to derive the dual fermion approach for disordered or disordered interacting systems
or it was just stated that it should be. We use the replica trick to easily integrate
out the disorder and for book-keeping purposes, that is to account for the correct
diagrams and remove unphysical ones.
Alternatively, one could proceed to solve the lattice problem for a single disor-
der configuration and perform the disorder-average for the dual quantities. Before
the disorder-average the vertex has only one contribution, namely from the interac-
tion. For the purely disordered system the dual potential would be zero before the
disorder-average.
Unlinked diagrams for the dual Green function would not appear, the same way
as they do not appear for the real fermion Green function. Thus, the diagrams
that contain what we call Fermi loops cannot be created. The disorder-average for
the dual diagrams might be possible the same way that it can be done for the real
fermion system [70] and it is thus possible to guess the correct equations.
Nevertheless, without an in-depth analysis the details of the disorder-average for
the dual fermion system remain unclear. Practically, this is more involved as it may
seem at a first glance. In the end, we found using the replica trick easier.
In that sense, it is pretty straight forward to see what the replica limit does for
the diagrams and where it comes from. It is not as straight forward to actually do
the derivation.
131
Bibliography
[1] M. Born and R. Oppenheimer. Zur Quantentheorie der Molekeln. Annalen
der Physik, 389(20):457–484, 1927.
[3] Gordon Baym and Christopher Pethick. Landau Fermi-Liquid Theory and
Low Temperature Properties of Normal Liquid 3He, pages 1–121. Wiley-VCH
Verlag GmbH, 2007.
[6] Sang-Wook Cheong. Transition metal oxides: The exciting world of orbitals.
Nat Mater, 6:927928, 12 2007.
[10] Qimiao Si and Frank Steglich. Heavy Fermions and Quantum Phase Transi-
tions. Science, 329(5996):1161–1166, 2010.
133
Bibliography
[15] Jan M. Tomczak and Silke Biermann. Optical properties of correlated mate-
rials - Or why intelligent windows may look dirty. physica status solidi (b),
246(9):1996–2005, 2009.
[18] Michael Schreiber and Heiko Grussbach. Multifractal wave functions at the
Anderson transition. Phys. Rev. Lett., 67:607–610, Jul 1991.
134
Bibliography
[29] Antoine Georges, Gabriel Kotliar, Werner Krauth, and Marcelo J. Rozenberg.
Dynamical mean-field theory of strongly correlated fermion systems and the
limit of infinite dimensions. Rev. Mod. Phys., 68(1):13, Jan 1996.
[34] Gabriel Kotliar, Sergej Y. Savrasov, Gunnar Pálsson, and Giulio Biroli. Cel-
lular Dynamical Mean Field Approach to Strongly Correlated Systems. Phys.
Rev. Lett., 87:186401, Oct 2001.
[35] V. Janiš. Parquet approach to nonlocal vertex functions and electrical con-
ductivity of disordered electrons. Phys. Rev. B, 64:115115, Aug 2001.
135
Bibliography
[40] H. Terletska, S.-X. Yang, Z. Y. Meng, J. Moreno, and M. Jarrell. Dual fermion
method for disordered electronic systems. Phys. Rev. B, 87:134208, Apr 2013.
[46] D.R. Hartree. The wave mechanics of an atom with a non-Coulomb central
field. Proc. Cambridge Phil. Soc., 24(1-2):89–132, 1928.
[48] N.W. Ashcroft and N.D. Mermin. Solid State Physics. Saunders College
Publishing, 1976.
[49] N F Mott. The Basis of the Electron Theory of Metals, with Special Refer-
ence to the Transition Metals. Proceedings of the Physical Society. Section A,
62(7):416, 1949.
136
Bibliography
137
Bibliography
138
Bibliography
[82] Mark Kac. On certain Toeplitz-like matrices and their relation to the problem
of lattice vibrations. Arkiv for det Fysiske seminar i Trondheim, 1978.
[83] Tommaso Castellani and Andrea Cavagna. Spin-glass theory for pedestrians.
Journal of Statistical Mechanics: Theory and Experiment, 2005(05):P05012,
2005.
[90] Antoine Georges and Werner Krauth. Numerical solution of the d =∞ Hub-
bard model: Evidence for a Mott transition. Phys. Rev. Lett., 69:1240–1243,
Aug 1992.
139
Bibliography
[91] Chinedu E Ekuma, Hanna Terletska, Zi Yang Meng, Juana Moreno, Mark
Jarrell, Samiyeh Mahmoudian, and Vladimir Dobrosavljević. Effective cluster
typical medium theory for the diagonal Anderson disorder model in one- and
two-dimensions. Journal of Physics: Condensed Matter, 26(27):274209, 2014.
[92] C. E. Ekuma, H. Terletska, K.-M. Tam, Z.-Y. Meng, J. Moreno, and M. Jarrell.
Typical medium dynamical cluster approximation for the study of Anderson
localization in three dimensions. Phys. Rev. B, 89:081107, Feb 2014.
[94] Armin Comanac. Dynamical mean field theory of correlated electron systems:
New algorithms and application to local observables. PhD thesis, Columbia
University, New York, 2007.
[97] M. Tsukada. A New Method for the Electronic Structure of Random Lattice
-the Coexistence of the Local and the Band Character-. J. Phys. Soc. Jpn.,
26:684–696, 1969.
[98] M. Tsukada. Contribution to the Many Site Theory of the Disordered System.
J. Phys. Soc. Jpn., 32:1475, 1972.
140
Bibliography
[113] Kenneth G. Wilson. The renormalization group: Critical phenomena and the
Kondo problem. Rev. Mod. Phys., 47:773–840, Oct 1975.
[114] Ralf Bulla, Theo A. Costi, and Thomas Pruschke. Numerical renormalization
group method for quantum impurity systems. Rev. Mod. Phys., 80:395–450,
Apr 2008.
[115] Ansgar Liebsch and Hiroshi Ishida. Temperature and bath size in exact di-
agonalization dynamical mean field theory. Journal of Physics: Condensed
Matter, 24(5):053201, 2012.
[116] C. Lanczos. An Iteration Method for the Solution of the Eigenvalue Problem of
Linear Differential and Intergral Operators. J. Res. Nat. Bur. Stand., 45:255–
82, 1950.
141
Bibliography
[117] Gerard L. G. Sleijpen and Henk A. van der Vorst. A Jacobi-Davidson Iter-
ation Method for Linear Eigenvalue Problems. SIAM J. Matrix Anal. Appl.,
17(2):401–425, April 1996.
[120] Martin Ganahl, Patrik Thunström, Frank Verstraete, Karsten Held, and
Hans Gerd Evertz. Chebyshev expansion for impurity models using matrix
product states. Phys. Rev. B, 90:045144, Jul 2014.
[121] F. Alexander Wolf, Ian P. McCulloch, and Ulrich Schollwöck. Solving nonequi-
librium dynamical mean-field theory using matrix product states. Phys. Rev.
B, 90:235131, Dec 2014.
[123] J. E. Hirsch and R. M. Fye. Monte Carlo Method for Magnetic Impurities in
Metals. Phys. Rev. Lett., 56:2521–2524, Jun 1986.
[124] Philipp Werner, Armin Comanac, Luca de’ Medici, Matthias Troyer, and An-
drew J. Millis. Continuous-Time Solver for Quantum Impurity Models. Phys.
Rev. Lett., 97:076405, Aug 2006.
[126] Carey Huscroft, Richard Gass, and Mark Jarrell. Maximum entropy method of
obtaining thermodynamic properties from quantum Monte Carlo simulations.
Phys. Rev. B, 61(14):9300–9306, Apr 2000.
[127] Philipp Werner, Takashi Oka, and Andrew J. Millis. Diagrammatic Monte
Carlo simulation of nonequilibrium systems. Phys. Rev. B, 79:035320, Jan
2009.
[128] Emanuel Gull, David R. Reichman, and Andrew J. Millis. Numerically exact
long-time behavior of nonequilibrium quantum impurity models. Phys. Rev.
B, 84:085134, Aug 2011.
142
Bibliography
[129] Aicke Hinrichs, Erich Novak, Mario Ullrich, and Henryk Woźniakowski. The
curse of dimensionality for numerical integration of smooth functions II. Jour-
nal of Complexity, 30(2):117 – 143, 2014. Dagstuhl 2012.
[130] N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller, and E. Teller.
Equation of State Calculations by Fast Computing Machines. The Journal of
Chemical Physcis, 21:1087–1092, 1953.
[132] W. K. Hastings. Monte Carlo sampling methods using Markov chains and
their applications. Biometrika, 57(1):97–109, 1970.
[135] Andrey E. Antipov, Emanuel Gull, and Stefan Kirchner. Critical Exponents
of Strongly Correlated Fermion Systems from Diagrammatic Multiscale Meth-
ods. Phys. Rev. Lett., 112:226401, Jun 2014.
[139] G. D. Mahan. Many Particle Physics, Third Edition. Plenum, New York,
2000.
143
Bibliography
144
Acknowledgements
Im Laufe meiner Promotion habe ich von vielen verschiedenen Personen immer
wieder Unterstützung erhalten. Zuerst möchte ich mich dafür bei meinem Dok-
torvater Thomas Pruschke bedanken. Er hat mir während meiner Promotion die
Freiheit gelassen, eigene Ideen und Lösungsansätze zu verfolgen, aber mir auch im-
mer dann geholfen und mich in die richtige Richtung gelenkt, wenn ich nicht weiter
wusste. Ebenso möchte ich meinem zweiten Betreuer Reiner Kree danken, sowohl
für inhaltliche Diskussionen und Feedback zu meiner Arbeit, wie auch zu formalen
Fragen im Zusammenhang mit meiner Promotion.
Weiterhin möchte ich den Mitgliedern der Festkörpertheorie für die angenehme At-
mosphäre in dieser Gruppe danken, insbesondere denjenigen, die meine Arbeit gele-
sen und mir mit ihren Rückmeldungen geholfen haben, meine Ergebnisse möglichst
verständlich und klar zu präsentieren. Namentlich möchte ich insbesondere Rainer
Härtle, Florian Sohn und Benjamin Lenz erwähnen.
Ebenso danke ich Silke Teschke dafür, dass sie meine Arbeit im Hinblick auf
sprachliche Richtigkeit korrekturgelesen hat.
Special thanks goes to Shuxiang Yang from the Louisiana State University (LSU).
Shuxiang supervised me on the whole project on disordered interacting systems. He
helped me learn the necessary tools to address this topic and supported me along
the way in our weekly discussions. I also thank Mark Jarrell and Juana Moreno
from LSU for useful discussions and for having me on the dual fermion project.
Meinen Eltern danke ich für ihre fortwährende Unterstützung, sowie dafür, dass
sie mir den Wert von Bildung vermittelt haben. Zum Schluss danke ich Merle für
ihre Unterstützung und aufmunternde Worte, wenn es mal nicht so gut lief.
I acknowledge financial support from the SFB 602, FOR 1807 as well as funding by
the National Science Foundation under the NSF EPSCoR Cooperative Agreement
No. EPS-1003897 with additional support from the Louisiana Board of Regents.
145
Curriculum Vitae
Persönliche Daten
Name: Patrick Haase
Geburtsdatum: 05. April 1985
Geburtsort: Soltau, Deutschland
Staatsangehörigkeit: deutsch
Bildung
Jan. 2014-März 2014: Aufenthalt als “Short Term Scholar” an der Louisiana State Uni-
versity, Baton Rouge, USA in der Arbeitsgruppe von Prof. Mark
Jarrell, PhD.
Seit Apr. 2011: Promotionsstudium an der Georg-August-Universität Göttingen,
Betreuer: Prof. Dr. Thomas Pruschke und Prof. Dr. Reiner
Kree
März 2011: Abschluss mit dem Titel ”Diplom-Physiker” (Note: sehr gut), Ti-
tel der Diplomarbeit: The Hubbard-Holstein Model with Disper-
sive Phonons and Graphene with an Impurity, Betreuer: Prof. Dr.
Thomas Pruschke
Sept. 2010-Nov. 2010: Arbeit in der Gruppe von Dr. Francisco Guinea am ICMM-CSIC
in Madrid
Febr. 2009-Juni 2009: Studienaufenthalt an der Universidad Autónoma de Madrid
2005-2011: Studium an der Georg-August-Universität Göttingen, Studiengang:
Physik
Juni 2004: Gymnasium Walsrode, Allgemeine Hochschulreife (Note: 2,0)
Publikation im Zusammenhang mit dieser Arbeit
S.-X. Yang, P. Haase, H. Terletska, Z. Y. Meng, T. Pruschke, J. Moreno, and M. Jar-
rell. Dual-fermion approach to interacting disordered fermion systems. Phys. Rev.
B, 89:195116, May 2014
Sonstige Publikationen
P. Haase, S. Fuchs, T. Pruschke, H. Ochoa, and F. Guinea. Magnetic moments and
Kondo effect near vacancies and resonant scatterers in graphene. Phys. Rev. B,
83:241408, Jun 2011