The Density Matrix

Renormalization Group
Applied to mesoscopic structures

M.Sc. Thesis
Dan Bohr
April 2004

MIC – Department of Micro and Nanotechnology

Technical University of Denmark
Lyngby, Denmark
The Density Matrix Renormalization Group
Applied to mesoscopic structures

c Dan Bohr, 2004.

°

M.Sc. Thesis, Technical University of Denmark.

MIC – Department of Micro and Nanotechnology

Ørsteds Plads
DTU – Building 345east
DK-2800 Kgs. Lyngby
Denmark
E-mail: [email protected]
Web: http://www.mic.dtu.dk

Typeset in LATEX 2ε
Abstract

In recent years much attention has been focused on understanding various

nanotechnological devices. Doing exact calculations using quantum mechan-
ical models limits the manageable system size heavily due to the exponen-
tially large Hilbert space, making full quantum calculations unfeasible in
the vast majority of cases. Non-perturbative methods and methods beyond
mean field theories, such as the Density Matrix Renormalization Group
(DMRG), are of great interest when studying systems with strong correla-
tions, such as quantum wires and dots.
We have made our own implementation of DMRG and studied two model
systems: (i) the one-dimensional Hubbard model at half filling and (ii) a
two level quantum dot coupled to semi-infinite leads. For the Hubbard
model we studied the effect of on-site repulsion, while for the quantum dot
with leads we computed the occupation as a function of chemical potential
varying the strength of the dot-lead coupling. In the non-interacting limit we
have compared DMRG and analytic Green’s function results and found good
agreement when keeping sufficiently many states in the DMRG calculation.
We used MatLab to implement DMRG which turned out to make parts
of the implementation cumbersome due to the matrix layout of MatLab.
Based on our experience we believe that low level object oriented languages
would be more suitable for DMRG.
Recently we have studied the possibility of calculating dynamical prop-
erties using a simple Lanczos method combined with DMRG. We have not
yet been successful and probably more sophisticated methods, as proposed
by both Jeckelmann and White, are required to successfully perform these
calculations.
To approach transport we have calculated the occupation of a biased
dot versus gate voltage using DMRG and these results have been compared
to Green’s function results. For small gate voltages the agreement is fine,
while for larger gate voltages there is a qualitative disagreement, and a more
careful analysis is needed to remedy this.

i
ii
Preface

The present thesis is submitted in candidacy for the Master of Science degree
within the Applied Physics program at the Technical University of Denmark
(DTU). It is the result of part of my work at the Department of Micro
and Nanotechnology (MIC) in the period May 2003 to April 2004 under
supervision of Professor Antti-Pekka Jauho.
The thesis concerns the powerful numerical method known as the Density
Matrix Renormalization Group (DMRG). On the theoretical level DMRG is
fairly simple to understand; an elegant derivation of the DMRG states can
be given and superficially the DMRG algorithms are easy to understand.
Actually implementing it turned out to be much more difficult, and I spent
long time setting up matrix representations of operators, expanding the
Hilbert space as the system grows and other vital technicalities.
Having taken care of these issues I found that doing precise DMRG
calculations indeed is very hardware demanding. Virtually any DMRG cal-
culation presented in this thesis took days rather than hours – some even
nearly a week – which obviously limited the number of calculations I was
able to do.

I would like to thank my supervisor for introducing me to the field of

DMRG through this project. During the project many people have helped
me in various ways, and apart from my supervisor I would like to emphasize
Professor Stellan Östlund, who kindly invited me to Chalmers and helped me
with the DMRG implementation on the Hubbard model, Professor Richard
Berkovits, who accepted to answer questions about the quantum dot with
leads setup, substantially improving the quality of my calculations, and
Thomas Frederiksen who’s useful comments to the manuscript I appreciate
very much.

ooOoo

Lyngby
April 30, 2004 Dan Bohr

iii
iv
Contents

1 Introduction 1
1.1 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . 3

2 The Density Matrix Renormalization Group 5

2.1 Retained states . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Infinite system algorithm . . . . . . . . . . . . . . . . 10
2.2.2 Finite system algorithm . . . . . . . . . . . . . . . . . 12
2.2.3 Diagonalization routine . . . . . . . . . . . . . . . . . 13
2.3 Symmetries and commutators . . . . . . . . . . . . . . . . . . 14
2.4 Speeding up the calculation . . . . . . . . . . . . . . . . . . . 14
2.4.1 Quantum numbers . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Wavefunction transformations . . . . . . . . . . . . . . 15
2.4.3 Varying number of retained states . . . . . . . . . . . 15
2.5 Observables and measurements . . . . . . . . . . . . . . . . . 16
2.5.1 Static quantities . . . . . . . . . . . . . . . . . . . . . 16
2.5.2 Dynamic quantities . . . . . . . . . . . . . . . . . . . . 17
2.6 General comments . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Hubbard model 25
3.1 Hamiltonian and matrices . . . . . . . . . . . . . . . . . . . . 25
3.2 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Infinite system sweep . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Finite system sweeps . . . . . . . . . . . . . . . . . . . . . . . 31
3.4.1 Varying number of retained states . . . . . . . . . . . 32
3.5 Computing the ground state . . . . . . . . . . . . . . . . . . . 32
3.6 Magnetic field . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.7 Expectation values . . . . . . . . . . . . . . . . . . . . . . . . 34
3.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.8.1 Fixed U . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.8.2 Fixed L . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.8.3 Eigenspectrum of ρ . . . . . . . . . . . . . . . . . . . . 36

v
vi
CONTENTS

3.9 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.10 Analytic result . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.11 Limitations and future work . . . . . . . . . . . . . . . . . . . 41
3.12 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Quantum dot with single infinite lead 45

4.1 Hamiltonian and matrices . . . . . . . . . . . . . . . . . . . . 45
4.2 Infinite system sweep . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Computing the ground state . . . . . . . . . . . . . . . . . . . 49
4.4 Expectation values . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.5.1 Convergence . . . . . . . . . . . . . . . . . . . . . . . 53
4.5.2 Spectral function . . . . . . . . . . . . . . . . . . . . . 55
4.6 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.7 Analytic result . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.7.1 Occupation . . . . . . . . . . . . . . . . . . . . . . . . 59
4.7.2 Comparison . . . . . . . . . . . . . . . . . . . . . . . . 61
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5 Quantum dot with two infinite leads 65

5.1 Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Superblock setup . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Infinite system sweep . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4.1 Convergence . . . . . . . . . . . . . . . . . . . . . . . 68
5.5 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.6 Analytic result . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.6.1 Occupation . . . . . . . . . . . . . . . . . . . . . . . . 74
5.6.2 Comparison . . . . . . . . . . . . . . . . . . . . . . . . 74
5.7 Towards transport . . . . . . . . . . . . . . . . . . . . . . . . 75
5.7.1 Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . 77
5.7.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.7.3 Analytic result . . . . . . . . . . . . . . . . . . . . . . 79
5.7.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . 81
5.7.5 Attempted current calculation . . . . . . . . . . . . . 81
5.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6 Single particle quantum mechanics 85

6.1 Hamiltonian and matrices . . . . . . . . . . . . . . . . . . . . 85
6.2 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2.1 RDM for single particle problems . . . . . . . . . . . . 86
6.2.2 Particle in a box . . . . . . . . . . . . . . . . . . . . . 87
6.2.3 Infinite system sweep . . . . . . . . . . . . . . . . . . . 88
6.2.4 Finite system sweeps . . . . . . . . . . . . . . . . . . . 89
 vii
CONTENTS

6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3.1 Comparison . . . . . . . . . . . . . . . . . . . . . . . . 91
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

7 Future work 93
7.1 Improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.1.1 Quantum numbers . . . . . . . . . . . . . . . . . . . . 93
7.1.2 Wave function transformations . . . . . . . . . . . . . 94
7.1.3 Programming language . . . . . . . . . . . . . . . . . 94
7.1.4 Diagonalization routine . . . . . . . . . . . . . . . . . 95
7.1.5 Further benchmarks . . . . . . . . . . . . . . . . . . . 95
7.2 Extending the capability of DMRG . . . . . . . . . . . . . . . 96
7.2.1 Transport . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2.2 Phonons . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2.3 Dynamic properties . . . . . . . . . . . . . . . . . . . 96

8 Summary and outlook 97

8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
8.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

A Observables 101

B Properties of the Lanczos basis 105

B.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
B.2 Orthogonality of basis . . . . . . . . . . . . . . . . . . . . . . 105

C Anticommutation and matrices 109

C.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
C.2 Anticommutators . . . . . . . . . . . . . . . . . . . . . . . . . 110
C.3 Commutators . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

D Truncation of a continued fraction 113

E Equation Of Motion Technique 115

E.1 Quantum dot with single infinite lead . . . . . . . . . . . . . 115
E.2 Quantum dot with two infinite leads . . . . . . . . . . . . . . 121

F Evaluation of ni for non-interacting biased quantum dot 123

F.1 Analytic expressions . . . . . . . . . . . . . . . . . . . . . . . 123
F.2 Numerics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
F.2.1 GA (ω, Vg ) . . . . . . . . . . . . . . . . . . . . . . . . . 124
F.2.2 G< ii (ω, Vg ) . . . . . . . . . . . . . . . . . . . . . . . . . 129
F.2.3 ni (Vg ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Bibliography 134
viii
Chapter 1

Introduction

The complexity of models used in solid state physics has increased dra-
matically since the first models were proposed by Drude and Sommerfeld.
Today quantum mechanical models are very powerful tools in the description
of nature, but present one formidable challenge: The Hilbert space needed
to describe a given system grows exponentially with the system size, and
thus even the fastest of computers rapidly reach a maximum system size
that can be handled. No matter how fast computers are available this will
always impose an upper limit to how large systems we can treat using exact
calculations.
Therefore a systematic way of sorting the information according to vi-
tality to the description is useful. When describing certain limits, e.g. the
low energy limit, not all information is needed and we may disregard some
parts without changing the description significantly, and hence increase the
manageable system size. One type is the Renormalization Groups (RG’s)
which are widely used in modern science.
The goal of any renormalization group is to find the best states to retain
when describing a system of l states keeping only m < l states. K. G. Wilson
was among the first to successfully perform numerical renormalization [1]. In
his approach the line of argument is that the most important states to retain
when describing low energy properties are the lowest eigenvalue eigenstates
of the Hamiltonian.
Wilson’s procedure had many successes but for interacting quantum lat-
tice systems it fails, mainly due to problems with the boundary conditions.
Some can be remedied by choosing other boundary conditions or by putting
the system of interest inside a larger system – a setup known as the super-
block setup.1 This failure can be illustrated by the tight binding chain
(TBC). In the continuum limit the TBC describes a particle in a box with
infinitely high potentials at the boundaries. Consider the joining of the
ground state wavefunctions of the two blocks to try to form a ground state
1
Elaboration on Wilson’s renormalization can be found in the original work by Wilson
and Kogut [1], and the Springer Lecture Notes in Physics edited by Peschel et al. [2].

1
2
Introduction

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1

Figure 1.1: Attempt to approximate the superblock ground state wavefunction

by the ground state wavefunctions of two subblocks. The wavefunctions are not
normalized and the units are arbitrary.

wavefunction of the superblock as seen in Fig. 1.1. The true ground state
wavefunction of the superblock deviates qualitatively from the one predicted
by the joint wavefunctions of the separate blocks, due to the boundary con-
dition imposed on each of the blocks.
Faced with such problems White formulated the Density Matrix Renor-
malization Group (DMRG) in 1992 [3, 4]. DMRG is able to handle strongly
interacting quantum systems, including quantum lattice systems, and gen-
erally gives very accurate results, making it a natural and obvious candidate
for many numerical calculations. Since then DMRG has been established as
a very powerful tool used by many groups around the world.

In recent years much attention has been focused on transport phenom-

ena, also in strongly correlated systems where existing methods are limited.
Analytic studies have only been carried out in certain limits, and usually
under quite severe restrictions. Mean field theories neglect higher order
correlations ab initio thereby restricting the kind of systems that can be
described. Contrary DMRG is able to handle strongly interacting quantum
systems in equilibrium, and has also turned out to be a good starting point
for treating transport in strongly interacting systems [5, 6].
Since White introduced DMRG, the basic algorithm has been extended
and modified to treat many different systems, ranging from ab initio quan-
tum chemistry calculations, over applications in high energy physics to cal-
culating dynamical properties using DMRG. In this context it is interesting
to note that several schemes to calculate dynamical properties and transport
have been proposed in the literature [2, 5, 6, 7].

In this thesis we aim at a basic understanding of the theory behind

DMRG and at implementing DMRG on a couple of models. In general we
 3
1.1 Outline of the thesis

attempt to include sufficient technical details such that beginners in the field
can write their own DMRG implementations. Some experience in numeri-
cal physics and mathematics as well as with a mathematical programming
language would however be a great advantage.

1.1 Outline of the thesis

In Chap. 2 we present in general terms the theoretical framework of DMRG
and the important DMRG algorithms are formulated. The general idea
behind DMRG can be understood by a fairly simple derivation of the DMRG
states – and a loose criterion for successful use of DMRG is formulated.
Methods for calculating observables are presented, and a few expansions of
the basic DMRG are discussed.
In Chap. 3 we show how DMRG is implemented on the one-dimensional
Hubbard chain. We use our implementation to calculate ground state prop-
erties at half filling, both with and without a magnetic field, and varying
the on-site repulsion and the chainlength.
In Chap. 4 and 5 we implement DMRG on a quantum dot coupled to
one and two semi-infinite one-dimensional leads. We use DMRG to calculate
the occupation of the dot as a function of chemical potential in the system.
In the two lead case we additionally calculate the occupation when the dot
is biased using different chemical potentials for the left and right lead, and
attempt a very simple calculation of the static current.
To increase the transparency of the text parts of Chap. 2 are repeated in
the specific contexts in Chap. 3, 4 and 5, and therefore some sections may
resemble each other.
In Chap. 6 we present a simpler implementation of DMRG on single
particle quantum mechanics. We show that the truncation procedure of
DMRG reduces to a projection of the wavefunction, making single particle
quantum mechanics a good starting point for DMRG calculations.
Several improvements of the implementation performed in this work
that would render the calculation faster and more reliable are proposed
in Chap. 7. A number of extensions of DMRG are mentioned, enabling
calculations beyond reach of ordinary DMRG.
In Chap. 8 we briefly summarize this work and discuss the current sta-
tus of DMRG in the literature. Perspectives of DMRG and its extensions
are discussed, and the natural continuation of this work on DMRG within
condensed matter physics and nanotechnology is outlined.
4
Chapter 2

The Density Matrix

Renormalization Group

Realizing the boundary condition problem of Wilson’s Numerical Renormal-

ization mentioned in the introduction White formulated the Density Matrix
Renormalization Group (DMRG) algorithms in the early 90s [3, 4]. DMRG
uses a superblock setup consisting of a system block and an environment
block, and the line of argument as to which states to retain is very different
from other renormalization groups.

In this chapter we review DMRG and formulate in general terms the two
important algorithms used in this work. We give a simple derivation of the
DMRG states, which shows there is no perturbative series involved, so that
the systems may be arbitrary strongly correlated.

Based on this derivation we formulate a loose criterion for successful

application of DMRG. Finally we show how expectation values are handled
in the DMRG framework and suggest several ways to speed up calculations.
This chapter is mainly based on the Springer Lecture Notes in Physics edited
by Peschel et al. [2], and the original articles by White [3, 4].

Superblock
z }| {
|ii |ji
| {z }| {z }
System Environment

Figure 2.1: Most general superblock setup. States |ii are system states and states
|ji are environment states.

5
6
The Density Matrix Renormalization Group

2.1 Retained states

The basic idea of DMRG is to optimize the basis in which the system is
described. This is done using a superblock setup where the system is joined
with an environment block to eliminate boundary condition problems. To
see how this works we consider an arbitrary superblock state |ψi. Using the
notation in Fig. 2.1, denoting system states by |ii and environment states
by |ji, this state can be written in product form
I X
X J
|ψi = ψij |ii |ji , (2.1)
i=1 j=1
P
where Ii=1 denotes summation over the complete basis of the system block,
P
and Jj=1 denotes summation over the complete basis of the environment
block. The reduced density matrix (RDM) for the system block is obtained
by tracing out the environment degrees of freedom1
X
ρii0 ≡ ψi∗0 j ψij
j
X †
= ψij ψji 0, (2.2)
j

or in matrix notation

ρ = ψψ † . (2.3)

For any system block operator A we then find the quantum mechanical
average as
­ ¯ ¯ ® X ­ ¯­ ¯ ¯ ®¯ ® †
ψ ¯A¯ψ = j ¯ i¯A¯i0 ¯j 0 ψi0 j 0 ψji
iji0 j 0
X X †
= Aii0 ψi0 j ψji
ii0 j
X
= Aii0 ρi0 i
ii0
= TrρA. (2.4)

Normalizing ρ, Trρ = 1, and α

Pdenoting its eigenstates by |u i and the
eigenvalues by wα ≥ 0 we have α wα = 1. For any system operator A we
may write the trace in this eigenbasis
­ ¯ ¯ ® X
ψ ¯A¯ψ = wα huα | A |uα i . (2.5)
α
1
This isP
the definition used by Feynman [8]. White and many others in the community
∗
use ρii0 = j ψij ψi0 j . This discrepancy seems to be quite persistent in the literature.
 7
2.1 Retained states

¯ ® ¯ ®
The goal is to construct the best possible approximate state, ¯ψ̄ ≈ ¯ψ ,
for the superblock using a given number of states. Eq. (2.5) yields a hint: If
for one particular α we had wα = 0 we would make no error by disregarding
the corresponding state |uα i when calculating quantum mechanical averages.
Mathematically the goal is to minimize
¯ ¯ ®¯2
S = ¯|ψi − ¯ψ̄ ¯ , (2.6a)
¯ ® X
¯ψ̄ = aα,j |uα i |ji , α = 1, . . . , m < I. (2.6b)
α,j

by varying over all coefficients

­ α ¯ α0 ® aα,j and choosing the best possible orthonor-
mal basis states u , u ¯u
α = δαα0 .
In order to map the problem of finding the best states and coefficients
onto a known form we use a change of basis
¯ ® X X
¯ψ̄ = |uα i aα,j |ji
α j
X X
≡ aα |uα i |v α i , aα |v α i = aα,j |ji . (2.7)
α j

Normalization of states |v α i yields

hv α | v α i ≡ 1
X
= hv α | ji hj | v α i
j
X ¯ ¯2
= ¯vjα ¯ , (2.8)
j
P
and aα is chosen so that this is fulfilled, |aα |2 = j |aα,j |2 .
Switching
¯ ® to matrix notation and using m basis states in the approximate
state ¯ψ̄ , the error measure is
X
S = Sij
ij
X³ m
X ´2
= ψij − aα uαi vjα , (2.9)
ij α=1

a form known from linear algebra.2 The state vector of the superblock |ψi
is represented by the rectangular (possibly complex) matrix ψij defined in
Eq. (2.1), and we may write it in singular value decomposition,3
ψ = U DV † , (2.10)
2
S in Eq. (2.9) is the square of the Frobenius norm of the matrix (ψ − ψ̄), see
e.g. http://www.uwlax.edu/faculty/will/svd/norm/.
3
Eric W. Weisstein. ”Singular Value Decomposition.” From MathWorld–A Wolfram
Web Resource. http://mathworld.wolfram.com/SingularValueDecomposition.html.
8
The Density Matrix Renormalization Group

where U and D are I × I matrices and V is J × I, where I and J are the

number of states in the system and environment blocks respectively. U and
V are unitary and D is diagonal with positive elements. The connection
between U , V , and D can be demonstrated as follows:

ψψ † = U DV † V DU †
= U D2 U † ⇒
ψψ † U = U D2
= U diag(σ12 , . . . , σI2 ), (2.11a)
† † †
ψ ψ = V DU U DV
= V D2 V † ⇒
ψ † ψV = V D2
= V diag(σ12 , . . . , σI2 ). (2.11b)

Hence the elements of D can be found as the square root of the eigenvalues
of ψψ † or ψ † ψ. The columns of U are the corresponding eigenvectors of ψψ †
while the columns of V are the corresponding eigenvectors of ψ † ψ.4
Now we see how to minimize S in Eq. (2.9): for aα choose the m largest
values of σi , for uα choose the corresponding columns of U , and for v α
choose the corresponding columns of V . For a given m < I this will be the
optimal choice. Note that the singular value decomposition is used only in
the derivation of the DMRG states and not in numerical calculations.
From Eq. (2.11a) it is clear that the singular value decomposition is
equivalent to diagonalization of the reduced density matrix

ρ = ψψ † ,
ψψ † U = U diag(σ12 , . . . , σI2 ) ⇒
U † ρU = diag(σ12 , . . . , σI2 ), (2.12)

i.e. that U is the matrix of eigenvectors of ρ. Hence the best possible states
to retain are the m largest eigenvalue eigenstates of ρ, or equivalently the
m most probable eigenstates of ρ.
If the system is in a mixed state with probabilities Wk ,5
¯ ® X ¯ ®
¯ψ = Wk ¯ψ k , (2.13)
k
¯ ®
each ¯ψ k can be written as in Eq. (2.1), and the relevant error measure
accordingly
X X³ m
X ´2
S = Wk k
ψij − akα uαi vjk,α . (2.14)
k ij α=1

4
Note that ψ † ψ is the RDM for the environment block.
5
To use finite temperatures Wk is normalized Boltzmann weights [2], p. 39.
 9
2.1 Retained states

We seek a single set of optimal states, hence not allowing uαi to depend on
k, while the environment states are allowed to change with k.
The reduced density matrix is
X X
ρii0 = Wk k
ψij (ψ k )†ji0 , (2.15)
k j

and a similar optimization procedure gives

U † ρU = diag(σ12 , · · · , σI2 ), (2.16)

where
X
σi2 = Wk (aki )2 , (2.17)
k

so that the optimal states again are the largest eigenvalue eigenstates of the
reduced density matrix, defined in Eq. (2.15).
In conclusion we have shown that the m < I optimal DMRG states
are the m largest eigenvalue eigenstates of the reduced density matrix as
properly defined in Eq. (2.2) and (2.15).

2.1.1 Efficiency
It is clear from the above derivation of the DMRG states that success of
DMRG is intimately connected with the spectrum of the reduced density
matrix. For fixed number of DMRG states m, the error of truncation is
larger if many eigenvalues of ρ are approximately equal, as compared to when
a few eigenvalues dominate. In fact one can use the sum of the truncated
eigenvalues as a guideline to the magnitude of the truncation error at each
step, although it does not tell anything about the actual error in observables.
Thus DMRG performs best if the spectrum of eigenvalues of the RDM
decays rapidly. When the decay becomes slower a larger number of states
must be kept to maintain precision and when this number becomes too large
precision can no longer be maintained and eventually DMRG fails.
For certain exactly solvable models it is possible to show analytically that
the spectrum of the RDM decays exponentially explaining the good perfor-
mance of DMRG in these cases. Also it should be noted that the boundary
conditions influence the spectrum of the RDM and generally DMRG per-
forms substantially worse for systems with periodic rather than open bound-
ary conditions [2], p. 41. Hence most often better precision is obtained using
larger systems with open BCs as compared to a small system with periodic
BCs.
10
The Density Matrix Renormalization Group

| {z } | {z } | {z }
Left Central Right

| {z } | {z }
System Environment

Figure 2.2: Generic superblock configuration and the nomenclature used in this
work. We use both notations, and use m to denote the number of states in the
truncated basis for the system block throughout this thesis.

2.2 Algorithms
To implement the idea of optimizing the basis we need algorithms for build-
ing up a system from a sufficiently small size and for improving the descrip-
tion of a given finite system. In this section we describe how the superblock
procedure is combined with the DMRG projection, leading to the two im-
portant DMRG algorithms.
In order not to truncate too heavily at each step the usual way of adding
degrees of freedom is to add one site at a time, hence building up the system
from a very small size in a controlled manner. A site means either a site in a
quantum lattice system or any other generic and sufficiently small building
block of the Hilbert space. The key idea is to divide the system into 4 blocks,
a left/system block, two central blocks, typically being single sites, and a
right block. The central and right blocks constitute the environment block,
as illustrated in Fig. 2.2.
There are two generically different algorithms, the infinite and the fi-
nite system algorithms differing by the choice of environment block. In the
following only ground state properties are targeted, although it is possible
to target also excited states or multiple states at the same time, as shown
in the previous section. We use the terminology of a one-dimensional sys-
tem, although this is not strictly necessary.6 Any block, superblock, system
block etc., is defined entirely by matrix representations of the Hamiltonian
and other relevant operators in the current basis. Hence in the following
the phrase “...form the superblock...” means construct the Hamiltonian and
other operators for the superblock in the current basis.

2.2.1 Infinite system algorithm

In the infinite system algorithm the system is enlarged one site at a time.
As environment block the simplest and most common choice is to use two
6
DMRG has been successfully applied to two-dimensional systems, see e.g. [9] and
references therein.
 11
2.2 Algorithms

(a) Initial 4 site superblock. (b) Intermediate 4 site su-

perblock with enlarged left
and right blocks.

(c) 6 site superblock.

Figure 2.3: Superblock configurations in the first DMRG step.

single sites joined with a spatial reflection of the system block, as seen in
Fig. 2.3. The superblock size thus grows by two sites at each step – giving
rise to the name ‘infinite system algorithm’. The algorithm is:
1. Form a superblock consisting of L sites, where L is small enough that
the ground state of the superblock Hamiltonian can be computed ex-
actly. In Fig. 2.3(a) the case L = 4 is shown.
2. Compute the ground state of the superblock Hamiltonian HLsuper nu-
merically.
3. From the ground state, in general the target state(s), construct the
reduced density matrix ρ, defined in Eq. (2.2), for the new system
block, i.e. the system block obtained when the leftmost central site
has been added, as seen in Fig. 2.3(b).
4. Diagonalize the reduced density matrix ρ obtaining the m most prob-
able eigenstates. The value of m should be chosen sufficiently large
to make the calculation accurate and sufficiently small to make su-
perblock matrices manageable on the computers available. Collect the
m column vectors uα in the truncation matrix O = (u1 , u2 , · · · , um ).
Note that usually a full diagonalization of ρ is fastest since many eigen-
states are needed.
5. Denoting by index l the last site in the old system block, l = 1 in
Fig. 2.3(b), construct the Hamiltonian Hl+1 and other operators Al+1
in the new system block, and transform them to the eigenbasis of the
reduced density matrix by H̄l+1 = OL† Hl+1 OL , and Āl+1 = OL
†
Al+1 OL .
R ,
6. Form a superblock of size L + 2 using H̄l+1 , two single sites, and H̄l+1
Fig. 2.3(c). Note that the right block is the spatial reflection of the
left block.
7. Repeat this procedure using this new superblock, defined by the Hamil-
super
tonian HL+2 , as input to step 2.
12
The Density Matrix Renormalization Group

In the first few steps it is possible to have m larger than the total number
of states in the system. In such cases one should keep all eigenstates of ρ,
and the truncation reduces to a shift of basis.
Using the infinite system algorithm the superblock can be built two sites
at the time from a small size in a controlled manner. These iterations can
either be continued to approximate the infinite system limit, or stopped
when the desired superblock size has been reached. Note that in step 2
it is essential to compute only the target state(s) and not perform a full
diagonalization of the superblock Hamiltonian.

(a) l = 3.

(b) l = 4.

(c) l = 5.

(d) l = 4.

(e) l = 3.

(f) l = 2.

(g) l = 1.

Figure 2.4: Superblock configurations for fixed superblock length, L = 8, in the

initial left to right and right to left phases.

2.2.2 Finite system algorithm

Once a superblock of the desired size has been built one can choose the
environment block in a different way. Keeping the superblock size fixed,
the point of division between the two blocks can be moved back and forth
 13
2.2 Algorithms

through the superblock improving the accuracy of the description at each

such sweep. The algorithm is:
1. Use the infinite system algorithm to build up a superblock of the de-
sired size L storing at each step the left and right block Hamiltonians
and operators relevant to connect these blocks. Also matrices for ob-
servables A should be stored at each step.

2. Carry out steps 3-5 in the infinite system algorithm, computing H̄l+1
and Āl+1 , l being the point of division as illustrated in Fig. 2.4. Store
H̄l+1 and Āl+1 .

3. Form a superblock of unchanged size L using H̄l+1 , two single sites

R . Note that H̄ R was found and stored during the build up
and H̄l−1 l−1
of the initial superblock using the infinite system algorithm.

4. Repeat steps 2-3 until l = L − 3. The above constitutes the left to

right phase in which the left block is enlarged, illustrated in Fig. 2.4(a)-
2.4(c).

5. Reverse the roles of left and right and move the point of division all the
way through the superblock until l = 1. This is the right to left phase,
in which the right block is enlarged, illustrated in Fig. 2.4(c)-2.4(g).

6. Reverse the roles of left and right again, and repeat this ‘zipping’ back
and forth until convergence is reached.
The left to right phase improves the description of the left block, while the
right to left phase improves the description of the right block. When the
calculation has converged the iterations are usually stopped when the left
and right blocks are of equal size. Specific examples of these two algorithms
are given in Chap. 3, 4, and 5.

2.2.3 Diagonalization routine

The largest matrices encountered in DMRG are the superblock matrices. A
major part of the calculational time is spent computing the ground state of
the superblock, or generally the target states. Hence it is essential to use an
efficient algorithm. In addition, since only the ground state – or at least a
very limited number of low lying eigenstates – of the Hamiltonian are needed
it is essential not to perform a full diagonalization since this is unnecessary
and time consuming. Iterative diagonalization schemes like the Lanczos and
Davidson algorithms are well suited for these purposes and implementations
exist for many different programming languages.
It should be noted that it is possible to eliminate the need of constructing
the full superblock Hamiltonian by incorporating a special diagonalization
routine [2], p. 47.
14
The Density Matrix Renormalization Group

2.3 Symmetries and commutators

Preserving symmetries of the system is essential for a successful truncated
description. The DMRG optimization process consists of multiple trunca-
tions and basis changes, and hence it is relevant to check if any symmetries
are unintentionally broken during these operations. Having a symmetry op-
erator S operating only on the system part of the superblock, the relevant
check is

[ρsystem , S] = 0, (2.18)

where ‘0’ is a zero-matrix of the same size as ρsystem and S [10]. Also it
can be useful to perform other checks such as (anti)commutators between
bosonic (fermionic) operators.
These checks should be performed at each step of the DMRG algorithm.
If they are not sufficiently fulfilled it is necessary to keep more states, i.e. use
larger m. Hence these checks aid in determining how many states it is neces-
sary to keep in order to describe the system accurately. The degree to which
these checks should be satisfied varies between models and implementations,
and some checks are more vital than others depending on what observables
we wish to describe accurately. These points should be investigated using
benchmark results to verify the precision of the DMRG within the model
and parameter values chosen.

2.4 Speeding up the calculation

A number of schemes to speed up the basic DMRG algorithms described
above have been proposed, and in this section we consider a few.

2.4.1 Quantum numbers

Identification of good quantum numbers can be used to speed up the DMRG
calculation substantially. In one-dimensional systems with short range inter-
actions the matrix representations of operators contain mostly zeros. The
basis states of each block (system and environment) may be labelled and
sorted according to good quantum numbers, which leaves non-zero entries
of most matrices in rectangular submatrices. Depending on the matrix in
question these submatrices may either connect states with the same set of
quantum numbers (i.e. square submatrices on the diagonal) or connect states
with different sets of quantum numbers (rectangular off-diagonal submatri-
ces). Hence, instead of working with large sparse matrices it is possible to
work with a set of smaller dense matrices. This is the optimal way to store
the data, since only non-zero elements are stored, reducing memory require-
ments. Also finding the ground state of the superblock becomes computa-
 15
2.4 Speeding up the calculation

tionally easier, since we may compute the ground state of a smaller dense
Hamiltonian matrix rather than the large sparse Hamiltonian matrix.
Sorting the superblock states according to quantum numbers addition-
ally allows one to seek the ground state within a given set of quantum
numbers, e.g. particle number or total spin, while working with the entire
superblock Hilbert space would always compute the total ground state.

2.4.2 Wavefunction transformations

A large part of the total calculational time in the DMRG scheme is used
computing the ground state of the superblock, and therefore it is very im-
portant to use an efficient algorithm.
In DMRG convergence to the actual ground state of the current Hamil-
tonian is vital since the basis is optimized for the target state. Hence con-
vergence to a different state would optimize the basis to describe this state
rather than the ground state. Most algorithms for finding eigenvectors and
eigenvalues use a random vector as starting point for an iterative diagonal-
ization scheme. Thus to ensure convergence to the actual ground state of
the Hamiltonian, and not a different low lying eigenstate, a large number of
iterative steps are needed, making the computation lengthy.
One solution to this problem is to have a good initial guess to the actual
ground state. This ensures that not very many iterations are necessary as
well as ensuring convergence to the actual ground state. And such a guess
can indeed be found in the finite system algorithm – using the ground state
from the previous step. To incorporate this initial guess in the DMRG
scheme the ground state wavefunction of the previous step has to be shifted
to the current basis using the transformation matrices from the previous
DMRG step. Hence the transformation matrices should be stored at each
step increasing storage requirements somewhat.

2.4.3 Varying number of retained states

Another way to speed up the calculation is to use a varying number of
retained states throughout the calculation. The idea is to build up the
system and let the finite system algorithm converge using m0 states. Upon
convergence the number of states kept is increased to m1 > m0 and the
finite system algorithm is iterated until convergence is achieved again. One
can then increase the number of states even further if sufficient precision
has not yet been obtained.
By varying m in this way the time spent on the infinite system algorithm
becomes less and we spend most time on the finite system algorithm. When
the finite system algorithm is used the precision of the infinite system sweep
is less critical and hence the calculation is not too sensitive to the number of
states kept in the build up of the system. Hence the optimal strategy is to
16
The Density Matrix Renormalization Group

use a relatively small number of states in the build up, and then increase m
rather fast once convergence has been reached in the finite system algorithm.

2.5 Observables and measurements

DMRG is an indirect method in which operators are known only through
their matrix representations. The basic DMRG is only suitable for static
quantities but several schemes to obtain dynamic quantities using DMRG
have been proposed in the literature [2, 5, 6, 7]. In the following subsections
we show how to obtain expectation values of static operators and a simple
Lanczos method of obtaining expectation values of dynamic correlators.

2.5.1 Static quantities

Expectation values of any static operator A that can be expressed in the
original basis are fairly easy to calculate. One has to construct and update
the matrix representing the operator at each DMRG step and these matrices
should be stored if the finite system algorithm is used.
After the final DMRG step one should have matrix representations of
A and the target state |ψi in the current optimized DMRG basis. The
expectation value can then be found as in Eq. (2.5),

hψ| A |ψi = Tr[ρA]

£ ¤
= Tr ψ † Aψ
X †
= ψji Aii0 ψi0 j , (2.19)
i,i0 ,j

where ψ is the target state matrix defined in Eq. (2.1). This product gives
one scalar, the expectation value, for each target state and the evaluation
gives the numerically exact expectation value for the approximate eigenstate.
Alternatively one can obtain expectation values directly from the vector
representation of |ψi,
­ ¯ ¯ ® X
ψ ¯A¯ψ = ψk∗ Akl ψl , (2.20)
kl

where now ψ is the column ground state vector, and A is any operator in
the basis of the full superblock. In App. A we show that the two methods
in Eq. (2.19) and (2.20) yield identical results for any system operator, and
thus the difference amounts to different organizations of the same data.
The evaluation of correlators often involve operators applied to different
parts of the system. Thus we may have to evaluate the expectation value
of an operator A = B · C, where B operates on one part of the superblock
and C operates on another. Let |ii be states in the system block and |ji be
 17
2.5 Observables and measurements

states in the environment block. The state of the system may be written in
product form
X ¯ ®¯ ®
|ψi = ψi0 j 0 ¯i0 ¯j 0 , (2.21a)
i0 j 0
X †
hψ| = hj| hi| ψji , (2.21b)
ij

so that if B and C operate on the system and environment blocks respec-

tively the expectation value of A is
X ­ ¯ ¯ ®­ ¯ ¯ ® †
hψ| A |ψi = i¯ B ¯i0 j ¯ C ¯j 0 ψji ψi0 j 0
i,i0 ,j,j 0
X †
= ψji Bii0 Cjj 0 ψi0 j 0 . (2.22)
i,i0 ,j,j 0

If, however, B and C operate on the same block one should notice that
X †
ψji Bii0 Ci0 i00 ψi00 j 6= hψ| A |ψi . (2.23)
i,i0 ,i00 ,j

The sum over i0 runs over the truncated basis set and should really run
over a complete set. Instead one should treat the combined operator as
one operator and construct and update matrices for [B · C] throughout the
calculation and then evaluate
X †
hψ| A |ψi = ψji [B · C]ii0 ψi0 j , (2.24)
i,i0 ,j

to obtain the correct expectation value. Notice that Eq. (2.24) is identical
to Eq. (2.19) for the combined operator A = B · C.
Using these relations it is fairly easy to obtain averages of operators if
they can be constructed in the initial basis. One simply updates the matrix
representations exactly the same way as the Hamiltonian and other operators
are updated, and when DMRG has converged use Eq. (2.19), (2.20), or (2.22)
to find expectation values. Which of the two methods is most conveniently
used depends on how data is represented in the implementation. Specific
examples of the calculation of observables using both methods are given in
Chap. 3, 4, and 5.

2.5.2 Dynamic quantities

One scheme to calculate expectation values of dynamic operators was pro-
posed by Wang, Hallberg, and Naef [2]. Following their ideas we show how
to obtain dynamic quantities in the zero temperature limit.
18
The Density Matrix Renormalization Group

For any approximate ground state |ψ0 i of the Hamiltonian H we may

construct an orthogonal basis in a Lanczos procedure. We specialize to
calculating fermionic correlators

C(t) = hψ0 | {A† (t), A(0)} |ψ0 i . (2.25)

We usually seek the Fourier transform

Z ∞
C(ω) = dt eiωt hψ0 | {A† (t), A(0)} |ψ0 i , (2.26)
−∞

and inserting the Heisenberg time evolution we find

Z ∞
C(ω) = dt eiωt hψ0 | {eiHt A† e−iHt , A} |ψ0 i
−∞
Z ∞ Z ∞
† i(ω+E0 −H)t
= hψ0 | A dt e A+A dt ei(ω−E0 +H)t A† |ψ0 i
−∞ −∞
= 2π hψ0 | A δ(ω + E0 − H)A + Aδ(ω − E0 + H)A† |ψ0 i .
†
(2.27)

Additionally we use the identity

1 1
= − iπδ(x), η → 0+ , (2.28)
x + iη x
to rewrite the δ-functions
1 1
δ(ω ± (E0 − H)) = − lim Im , (2.29)
π η→0 + ω ± (E0 − H) + iη

so that

C(ω) = (2.30)
1 1
−2 lim Im hψ0 | A† A+A A† |ψ0 i ,
η→0+ ω + E0 − H + iη ω − E0 + H + iη

or finally using the notation

G+ (z) = hψ0 | A(z + H)−1 A† |ψ0 i , (2.31a)

− † −1
G (z) = hψ0 | A (z − H) A |ψ0 i , (2.31b)

we find
¡ ¢
C(ω) = −2 lim Im G− (ω + E0 + iη) + G+ (ω − E0 + iη) . (2.32)
η→0+

We start the Lanczos procedure by defining

½
A |ψ0 i ,
|f0 i = (2.33)
A† |ψ0 i ,
 19
2.5 Observables and measurements

and iterate to find further Lanczos vectors,

|fn+1 i = H |fn i − an |fn i − b2n |fn−1 i , (2.34)

­ ¯ ®
where n ≥ 0, b0 = 0, and fn ¯fm = 0 for n 6= m as shown explicitly in
App. B. The coefficients are

hfn | H |fn i
an = ­ ¯ ® , (2.35a)
fn ¯fn
­ ¯ ®
fn ¯fn
b2n = ­ ¯ ®. (2.35b)
fn−1 ¯fn−1

In order to reexpress the Green’s functions Eq. (2.31) we first define an

orthonormal basis
|fn i
|ni = p . (2.36)
hfn |fn i

Using Eq. (2.34) the Hamiltonian applied to the basis state |ni is

H |ni = (2.37)
s s
hfn+1 |fn+1 i hfn−1 |fn−1 i
|n + 1i + an |ni + b2n |n − 1i ,
hfn |fn i hfn |fn i

and using the notation

s
hfm |fm i
= nm,n , (2.38)
hfn |fn i

we find the matrix elements

hm| H |ni = nn+1,n δm,n+1 + an δm,n + b2n nn−1,n δm,n−1


 nn+1,n , m = n + 1,
= an , m = n, (2.39)
 2
bn nn−1,n , m = n − 1.

Thus by construction the Hamiltonian is tri-diagonal in the Lanczos basis

 
z ± a0 ±b21 n0,1 0 ··· 0
 .. 
 ±n1,0 z ± a1 ±b22 n1,2 . 
 
(z ± H)n =   0 ±n2,1 z ± a2 0  . (2.40)

 .. . . 
 . . ±bn nn−1,n 
2

0 ··· 0 ±nn,n−1 z ± an
20
The Density Matrix Renormalization Group

First we rewrite Eq. (2.31) as

­ ¯ ¯ ®
G± (z) = f0 ¯(z ± H)−1 ¯f0
X ­ ¯ ®­ ¯ ¯ ®­ ¯ ®
= f0 ¯m m¯(z ± H)−1 ¯m0 m0 ¯f0
mm0
­ ¯ ® ­ ¯ ®
X f0 ¯fm ­ ¯¯ ¯ ® fm0 ¯f0
−1 ¯ 0
= q­ ¯ ® m (z ± H) m q­ ¯ ®
mm 0 fm ¯fm fm0 ¯fm0
­ ¯ ®­ ¯ ¯ ®
= f0 ¯f0 0¯(z ± H)−1 ¯0 . (2.41)
Hence we need to find the first element of the inverse of the matrix (z ± H)
in the basis of Eq. (2.36).7 In order to invert the matrix in Eq. (2.40) we
first consider the solution to 2nd order in the coefficients ai and b2i
µ ¶
z ± a0 ±b21 n0,1
(z ± H)1 = ⇒
±n1,0 z ± a1
­ ¯ ¯ ® [C1 ](1,1)
0¯[(z ± H)1 ]−1 ¯0 = , (2.42)
det(z ± H)1
where [C1 ](1,1) is the complement of [(z ± H)1 ](1,1) , given by the subdeter-
minant, in this case just a scalar,
[C1 ](1,1) = z ± a1 , (2.43)
so that to this order
­ ¯ ¯ ® z ± a1
0¯[(z ± H)1 ]−1 ¯0 =
(z ± a0 )(z ± a1 ) − b21 n0,1 n1,0
1
= b21
, (2.44)
z ± a0 − z±a 1

where we used that ni,i+1 ni+1,i = 1.

Next we consider the solution to (n + 1)’st order in the coefficients ai
and b2i , given by a very similar expression,
­ ¯ ¯ ® [Cn ](1,1)
0¯[(z ± H)n ]−1 ¯0 = , (2.45)
det(z ± H)n
with the complement given by the subdeterminant
¯ ¯
¯ z ± a1 ±b2 n1,2 0 ··· 0 ¯
¯ 2 ¯
¯ .. ¯
¯ ±n2,1 z ± a2 ±b23 n2,3 . ¯
¯ ¯
[Cn ](1,1) = ¯¯ 0 0 ¯
¯
¯ .. .. ¯
¯ . . ±b2n nn−1,n ¯
¯ ¯
¯ 0 ··· 0 ±nn,n−1 z ± an ¯
≡ det(2 : n), (2.46)
7
See e.g. [11] for details on inverting matrices.
 21
2.5 Observables and measurements

which defines the notation det(2 : n) for the subdeterminant of (z ± H)n in

Eq. (2.46). Extending in the obvious way the usage of this notation we have
­ ¯ ¯ ® det(2 : n)
0¯[(z ± H)n ]−1 ¯0 = . (2.47)
det(1 : n)
The full determinant can be rewritten
det(1 : n) = (z ± a0 ) det(2 : n)
¯ ¯
¯ ±n1,0 ±b2 n1,2 0 ··· 0 ¯
¯ 2 ¯
¯ .. ¯
¯ 0 z ± a2 ±b23 n2,3 . ¯
¯ ¯
∓b21 n0,1 ¯¯ 0 0 ¯
¯
¯ .. .. ¯
¯ . . ±b2n nn−1,n ¯
¯ ¯
¯ 0 ··· 0 ±nn,n−1 z ± an ¯
= (z ± a0 ) det(2 : n) − b21 n0,1 n1,0 det(3 : n)
= (z ± a0 ) det(2 : n) − b21 det(3 : n), (2.48)
and hence we find
­ ¯ ¯ ® det(2 : n)
0¯[(z ± H)n ]−1 ¯0 =
(z ± a0 ) det(2 : n) − b21 det(3 : n)
1
= . (2.49)
z ± a0 − b21 det(3:n)
det(2:n)

Due to the tri-diagonal structure of the Hamiltonian matrix in Eq. (2.40)

we can rewrite any fraction det(m + 1 : n)/ det(m : n) in a completely similar
way,
det(m + 1 : n) det(m + 1 : n)
=
det(m : n) (z ± am−1 ) det(m + 1 : n) − b2m det(m + 2 : n)
1
= det(m+2:n)
. (2.50)
z ± am−1 − b2m det(m+1:n)
Hence by successive use of Eq. (2.50) and using Eq. (2.41) we find the final
solution for arbitrary n, in particular n → ∞,
hψ0 | AA† |ψ0 i
G+ (z) = b21
, (2.51a)
z + a0 − b2
2
z+a1 − z+···

hψ0 | A† A |ψ0 i
G− (z) = b21
. (2.51b)
z − a0 − b2
2
z−a1 − z−···

The condition for this method to be feasible is that the continued frac-
tions in Eq. (2.51) converge using a finite and manageable number of Lanc-
zos coefficients ai and b2i . If very many coefficients are needed it becomes
difficult to maintain the accuracy.
22
The Density Matrix Renormalization Group

In principle this procedure is independent of the method used to find

the ground state [12]. In a DMRG implementation of the method above
there are however technical details that need attention. The DMRG basis is
usually heavily truncated compared to the complete basis. The truncation is
performed to optimally describe the target state, usually the ground state.
The basis may therefore not be suitable for describing excitations in the
system, making it impossible to accurately describe dynamical quantities in
this basis.
This apparent incompatibility can be remedied by including exited states
as target states. As was pointed out by Hallberg the first few Lanczos
vectors are candidates for such additional target states [13]. Hence the new
normalized target states become
¯ ®
 ¯ψ0 , l = 1,


¯ l®
¯ψ = ¯ ® (2.52)
 ¯fl−2

­ ¯ ® , l = 2, 3, · · · , M ,
¯
fl−2 fl−2

and the reduced density matrix accordingly

M
X
ρii0 = l
pl ψij (ψ l )†ji0 , (2.53)
l=1
¯ ® P
where pl is the weight for target state ¯ψ l and pl = 1 to maintain the
normalization.
With these additional target states excitations can be described using
a standard DMRG implementation. The additional calculations amount to
computing and including as target states at each step M normalized Lanczos
vectors, given the ground state |ψ0 i, the Hamiltonian H, and the operator
A or A† . Since the precision of DMRG generally decreases when more target
states are included the optimal value of M has to be determined from case
to case. Including more states improves the description of excitations but
lowers the general precision of DMRG.

2.6 General comments

DMRG is an approximate method, which restricts the complete Hilbert
space of the particular model considered to make computations less demand-
ing. From the derivation of the DMRG states it is clear that the method is
variational of nature and that no perturbative series is used. Thus arbitrary
strongly correlated systems may be treated using DMRG.
However one cannot always expect completely monotonous convergence
in energy due to the fact that the variational procedure optimizes the basis
of the target state. Thus the description of the target state improves at each
 23
2.6 General comments

step while the energy is not necessarily lowered [14]. Also it should be noted
that the ground state energy of the superblock depends on the configuration
of the superblock, i.e. the relative size of the left and right blocks.
In DMRG the strength of the truncation is given entirely by m, and
in general increasing m increases the accuracy of the description. Hence
there is full control over the truncation and the severity can be checked by
increasing m. A monotonous increase in the precision is therefore naively
expected for increasing m, and when m equals the total number of states in
the system no truncation is performed. In the limit of keeping all system
states DMRG is conceptually meaningless since we might as well use exact
diagonalization techniques, and the above statements are only included to
emphasize the correspondence between m and the degree of truncation.
There is one exception however, when degeneracies of the spectrum of
the reduced density matrix exist. Such degeneracies correspond to symme-
tries of the Hamiltonian and hence all the degenerate states should be kept.
As an example Caprara and Rosengren argues that spin triplets manifest
themselves in this way and that all degenerate states of the reduced density
matrix therefore must be kept [10].
Finally it should be noted that the range of interactions is closely related
to the dimensionality of lattice systems. Any two or tree-dimensional lat-
tice system with short range interactions can be mapped onto a sufficiently
long one-dimensional system with long range interactions. Hence one might
naively expect that such systems could be treated in the DMRG framework.
This is not the case as it turns out that the calculational time grows rapidly
with the width of quasi-one-dimensional systems making such simple ap-
proaches unfeasible. Similarly the calculational time grows rapidly when
interactions become longer ranged. It should however be noted that DMRG
has been successfully applied to two-dimensional systems, using more so-
phisticated approaches [9].
24
Chapter 3

Hubbard model

To understand in more detail how DMRG works we implemented it on

the one-dimensional Hubbard chain, describing fermions confined to a one-
dimensional chain consisting of sites, on which there can be 0, 1, or 2 spinful
electrons.
In this chapter we review our implementation of DMRG and use it to
calculate low energy properties of the Hubbard chain for various lengths
and couplings, both in zero and finite magnetic field. In the non-interacting
limit we compare the DMRG ground state energy to corresponding ana-
lytic results and find good consistency. We briefly comment on the results
found and conclude the chapter with a discussion of the limitations of our
implementation.

3.1 Hamiltonian and matrices

In order to limit our study of the Hubbard model we have focused on calcu-
lations at half filling. The usual Hubbard term U ni↑ ni↓ does not ensure this
since it does not treat empty and doubly occupied sites equally, and thus
the chemical potential depends on the value of U . To ensure half filling we
use the Hamiltonian

L−1
X X ³ † ´ L
X
H = −t ciσ ci+1σ + c†i+1σ ciσ + U (1 − Ni )2 , (3.1)
i=1 σ=↑,↓ i=1

where Ni = ni↑ + ni↓ , and where the term (1 − Ni )2 ensures that states with
Ni = 1 are energetically favorable compared to Ni = 0 and Ni = 2.
The basis used for the sites is |0i, |↑i ≡ c†↑ |0i, |↓i ≡ c†↓ |0i, and |↑↓i ≡
c†↑ c†↓ |0i. In this basis the matrix representation of annihilation operators
ciσ , the number operator Ni , the total z-spin operator Siz , and the on-site

25
26
Hubbard model

repulsion Hamiltonian HU for the single sites are

 
0 1 0 0
0 0 0 0
ci↑ = 
0 0 0 1 ,
 (3.2a)
0 0 0 0
 
0 0 1 0
0 0 0 −1
ci↓ = 
0 0 0 0  ,
 (3.2b)
0 0 0 0
 
0 0 0 0
0 1 0 0 † †
Ni =  
0 0 1 0 = ci↑ ci↑ + ci↓ ci↓ , (3.2c)
0 0 0 2
 
0 0 0 0
0 1 0 0
Siz =  2
0 0 − 1 0 ,
 (3.2d)
2
0 0 0 0
 
1 0 0 0
0 0 0 0
HU = U  0 0 0 0 .
 (3.2e)
0 0 0 1
These operators have to be expressed in the basis of several sites. Elec-
trons anticommute and hence we must keep track of the number of electrons
‘to the left of’ any given anticommuting operator, provided a numbering of
the basis states. We use the obvious numbering given by the site numbers.
This bookkeeping is done using a matrix counting the number of particles
on each site, and returning a sign in accordance. For the single site this
matrix is
 
1 0 0 0
0 −1 0 0
Pi = 0 0 −1 0 .
 (3.3)
0 0 0 1
In App. C we demonstrate in further detail the properties of P in relation
to anticommutation of fermionic operators.
Denoting fermionic operators by f and bosonic operators by b we enlarge
the basis by performing Kronecker tensor products,
f˜1 = f1 ⊗ δ2 , (3.4a)
f˜2 = P1 ⊗ f2 , (3.4b)
b̃1 = b1 ⊗ δ2 , (3.4c)
b̃2 = δ1 ⊗ b2 , (3.4d)
 27
3.2 Symmetries

where e.g. b̃1 means operator b1 in the enlarged basis of sites 1 and 2. This
illustrates the use of matrix P .

Hermiticity of H
The Hamiltonian operator is hermitian, and thus also the matrix represen-
tation should be. Throughout DMRG calculations we find that the Hamil-
tonian matrix indeed is hermitian – apart from numerical imprecisions on
the scale of the machine working precision (∼ 10−16 ). Some sparse matrix
diagonalization routines require complete hermiticity of the matrices. Hav-
ing ensured it was only numerical imprecisions preventing the Hamiltonian
from being hermitian, i.e. checking that

max(H − H † ) ∼ 10−16 , (3.5)

we use the substitution,

H + H†
→ H. (3.6)
2
to ensure hermiticity. This forces any matrix to be hermitian, and there-
fore it is absolutely crucial to ensure that only numerical imprecisions are
eliminated this way. If the Hamiltonian turns out to be far from hermi-
tian something else is wrong, and using this trick prevents such errors from
displaying themselves directly and leads to wrong or imprecise results.

3.2 Symmetries
Preserving symmetries of the Hamiltonian is essential. We have incorporated
two symmetry checks at each step of the algorithm,
£ ¤
ρ, N L = 0, (3.7a)
£ ¤
ρ, SzL = 0, (3.7b)

as well as various anticommutation checks, e.g.

© ª
cl,σ , c†l,σ = 1, (3.8a)
© ª
cl,σ , c†l+1,σ = 0, (3.8b)

where l denotes the rightmost site of the left block, and ‘0’ and ‘1’ are zero
and identity matrices respectively.
These checks are performed after each truncation in the DMRG calcula-
tion. This is not completely trivial since these checks are never completely
fulfilled. In this implementation the degree to which these checks are satis-
fied depends crucially on the parameter values chosen. Usually the checks are
28
Hubbard model

fulfilled within 10−8 − 10−12 , remarkably being worst in the non-interacting

limit where we often find values as high as 10−6 , which is probably close
to the lower acceptable limit of precision in order for the DMRG calcula-
tions to be precise. Keeping more states would improve this but also make
the calculational time longer. It is worth noting that much DMRG work
reported use significantly higher number of states kept than we have been
able to.

3.3 Infinite system sweep

The DMRG procedure is started by constructing the Hamiltonian and other
operators for a four site superblock. As illustrated in Eq. (3.4) the basis
is enlarged by performing Kronecker tensor products. The Hamiltonian for
sites 1 and 2 is constructed first,

H12 = HU ⊗ δ2 + δ1 ⊗ HU
X h¡ † ¢¡ ¢ ¡ ¢¡ ¢i
−t c1σ ⊗ δ2 P1 ⊗ c2σ + P1 ⊗ c†2σ c1σ ⊗ δ2 , (3.9)
σ

where the first line represents the on-site repulsions and the second line
represents the hopping between sites 1 and 2. Note the usage of matrix P1
to maintain anticommutation of fermionic operators.
Other relevant operators are
z
S12 = S1z ⊗ δ2 + δ1 ⊗ S2z , (3.10a)
P12 = P1 ⊗ P2 , (3.10b)
cL
σ = P1 ⊗ c2σ , (3.10c)
N12 = N1 ⊗ δ2 + δ1 ⊗ N2 , (3.10d)

representing the total z-component of the spin, the fermionic sign operator,
the annihilation operator for the rightmost site, and the number operator
for two sites respectively. Note that cL
σ is needed to connect the left block
to the leftmost central site. All matrices in the combined Hilbert space of
sites 1 and 2 have dimensions 16 × 16.
Completely analogous expressions exist for the right block. Hence we
may construct the Hamiltonian and other operators for the four site su-
perblock, shown in Fig. 3.1(a),

H1234 = H12 ⊗ δ34 + δ12 ⊗ H34 (3.11)

X³ ´
−t ((cL † R R † L
σ ) ⊗ δ34 )(P12 ⊗ cσ ) + (P12 ⊗ (cσ ) )(cσ ⊗ δ34 ) ,
σ

where cRσ is the annihilation operator for the leftmost site in the right block,
and where the second line in Eq. (3.11) is the hopping between the left and
right blocks.
 29
3.3 Infinite system sweep

(a) Initial 4 site superblock. (b) Intermediate 4 site su-

perblock with enlarged left
and right blocks.

(c) 6 site superblock.

Figure 3.1: Superblock configurations in the first infinite system DMRG step.

Usually there is no truncation at this point since the Hilbert space is

sufficiently small to be handled exactly. However to illustrate the DMRG
scheme we do so anyway. The first step is to compute the ground state
eigenvector of the Hamiltonian, denoted H1234 above, currently a 256 × 256
sparse matrix. We use Lanczos solver laneig1 or Davidson solver jdqr2 for
MatLab.
Denoting the generic element of the ground state vector vα ,
|ψi = (v1 , · · · , v256 ) , (3.12)
the ground state matrix for the new left system, consisting of sites 1 and 2
as in Fig. 3.1(b), is constructed as
|0, 0i |0, ↑i · · · |↑↓, ↑↓i
 
ψ = |0, 0i v1 v2 ··· v16 (3.13)
|0, ↑i  v17 v18 v32 
 
..  .. .. .. ,
.  . . . 
|↑↓, ↑↓i v240 v241 ··· v256
where the rows correspond to the complete basis for sites 1 and 2, and the
columns correspond to the complete basis for sites 3 and 4. Note that e.g. v16
corresponds to the basis vector |0, 0, ↑↓, ↑↓i ≡ |0, 0 i| ↑↓, ↑↓i.
The reduced density matrix (RDM) for the left subsystem is given by
ρL = ψψ † . (3.14)
DMRG instructs us to retain the most probable eigenstates of the RDM. In
this model however N L and Sz L are symmetries, which can be demonstrated
as follows: Let |ψN i be the state vector of the superblock for given fixed
particle number N
X
|ψN i = ψαβ |nα i |nβ i , (3.15)
α,β
1
http://soi.stanford.edu/∼rmunk/PROPACK/.
2
http://www.math.uu.nl/people/sleijpen/index.html.
30
Hubbard model

where |nα i and |nβ i are states of the system and environment block re-
spectively, and labeled by their particle numbers. Obviously we have N =
nα + nβ , and thus given N and nβ also nα is fixed. The reduced density
matrix is
X †
ραα0 = ψαβ ψβα 0 (3.16)
β

which implies
£ that¤ |nα i and |nα0 i are states with the same particle number,
and hence ρ, N L = 0. Similar arguments apply for SzL .
Thus we may choose simultaneous eigenstates of ρL , N L , and Sz L , by
computing eigenvectors of

ρ̄L = ρL + r1 N L + r2 Sz L , ri ∈]0, 1]. (3.17)

The eigenvalues of ρL are then found by U † ρL U , where U is the matrix of

eigenvectors found above. With this procedure we ensure that the particle
number and spin remains good quantum numbers for the DMRG basis we
construct, thus keeping matrices N L and SzL diagonal.
Diagonalizing ρ̄L and collecting the m retained column eigenvectors, de-
noted by uα , in a matrix,

O = (u1 , u2 , · · · , um ) , (3.18)

we can truncate and shift to the retained eigenbasis of ρL . The block Hamil-
tonian and all operators needed to reconstruct the superblock Hamiltonian
are shifted to this basis,

H̃1 = O† H12 O, (3.19a)

c̃L
σ = O† cL
σ O, (3.19b)
†
P̃L = O PL O, (3.19c)
†
à = O AO, (3.19d)

where A is the matrix representation of any operator working on the left

block. Truncating in a similar fashion the right block, retaining eigenstates
of ρR , we obtain H̃4 and c̃R L L
σ . Renaming H̃1 → H1 , c̃σ → cσ , P̃L → PL ,
R R
H̃4 → H4 , and c̃σ → cσ we have a truncated description of the enlarged left
and right blocks in Fig. 3.1(b).
Truncated left block matrices H1 , cLσ , and PL , and the corresponding
truncated right block matrices, should be stored for use in the finite system
algorithm. We store additionally matrices for N L and SzL in order to calcu-
late these quantum numbers for the final state. If other observables are to
be found matrices representing those have to be stored as well.
Using the truncated block matrices the Hamiltonian of a superblock
consisting of 6 sites can be constructed. The left and right blocks of this
 31
3.4 Finite system sweeps

superblock contains two sites each and are represented in a truncated basis
by H1 and H4 , and the two central sites are represented in their full real
space basis. The superblock is shown in Fig. 3.1(c).
This setup is formally identical to the initial having matrices for the left
and right block and for the two central sites, and hence we may repeat all
the above steps. We iterate in this way until the chain reaches the desired
length, increasing the superblock chain length by two sites at each iteration.

3.4 Finite system sweeps

Once the superblock reaches the desired length L all matrices for the right
and left block being up to half the total length should be stored on disk or in
memory. At the onset of the finite system algorithm the superblock consists
of a left and right block, each consisting of L/2 − 1 sites and represented in
a truncated basis, and two central sites each represented in their full real
space basis.
Once the superblock Hamiltonian has been constructed the basic step is
similar to the infinite system method. Taking the left to right sweep as an
example the basic DMRG step is:

1. Compute the ground state of the superblock Hamiltonian.

2. Construct the RDM ρL for the new left block, consisting of the old
left block, initially of length L/2 − 1, and the leftmost central site.

3. Diagonalize ρL .

4. Find the retained eigenbasis by keeping the m most probable eigen-

states of ρL .

5. Transform all left matrices to the eigenbasis of ρL and store them.

6. Construct a new superblock Hamiltonian, consisting of the new trun-

cated left block, two central sites, and a new right block containing
one site less. At the first finite system step the length of the new left
block is L/2, and the length of the new right block is L/2−2. Matrices
defining the new right block were found and stored during the infinite
system sweep.

This prescription is repeated using the new superblock as input until the
point of division has been moved all the way through the chain. The basic
right to left step is identical to the above reversing the roles of left and
right. Each finite system sweep improves the description of the block that
is enlarged during the sweep.
Having obtained convergence by sweeping back and forth through the
chain the algorithm is stopped when the left and right blocks are of equal
32
Hubbard model

size. It is not always obvious when convergence has been reached – but one
usable measure is the change in ground state energy from sweep to sweep.
Note however that the energy is not necessarily monotonously decreasing,
as discussed in Sect. 2.6. Typically we iterate 4-6 times and iterating further
does not improve the description.

3.4.1 Varying number of retained states

Our calculations on the Hubbard model uses the infinite as well as the finite
system algorithms. We have experimented with retaining different number
of states in the calculation, as described in Sect. 2.4.3. In the infinite system
sweep (the build up of the system) we retain m = 10 states, which makes
the build up very fast for the system sizes we have used. Once the system
has been built we iterate the finite system algorithm 5 times using m = 10,
ensuring convergence with this number of states kept. Then we increase m
to the final value, typically m = 24, and iterate until convergence has been
reached, usually around 5 times.
In our implementation we use multidimensional arrays to store matrices
for the left and right blocks, and in MatLab these are non-sparse and
declared with a given fixed size. Thus to retain different number of states
in the calculation we need at least two storage arrays increasing memory
requirements.
In general we find lower ground state energies when we use m = 24 in
the whole calculation. Also we find that the precision of the infinite system
algorithm does indeed depend on the number of states kept, hence we cannot
really justify using m = 10 when building up the system and have therefore
used constant m in all calculations presented in this thesis.
It should be noted that we use a significantly smaller number of states
than in most DMRG results reported in the literature. The above conclu-
sions may depend heavily on m and should therefore be viewed as conclusions
for our implementation only; they are not general statements.

3.5 Computing the ground state

The superblock contains basis states corresponding to many different quan-
tum numbers, e.g. particle number and spin. Often a target state of a given
particle number is desired, and hence such a subspace of the total Hilbert
space has to be picked out. The most time consuming part of the algorithm
is computing the ground state, and since such subspaces are often substan-
tially smaller than the full Hilbert space the overall effect is a speed up of
the total calculation.
We have specialized to calculating at half filling using a Hamiltonian
operator that ensures this, cf. Eq. (3.1). This does not imply that the basis
used only contains states at exactly half filling. Rather it means that the
 33
3.6 Magnetic field

ground state is a half filled state. In our calculations it would therefore suffice
to simply compute the ground state of the entire Hamiltonian matrix. But
taking into account the speed up of the ground state calculation explained
above we still use a subspace calculation.
Instead of picking out the subspace completely and working with a
smaller dense matrix we have chosen to (temporarily) set entries of the
superblock Hamiltonian to zero before the diagonalization, hence creating a
much sparser Hamiltonian. In doing so the total DMRG calculation becomes
up to 10 times faster in our implementation.

3.6 Magnetic field

In order to split the spin degeneracy present in the Hubbard model we have
included a small magnetic field in the z-direction. The extra term in the
site Hamiltonian is

L
X
Hh = −h (Sz )i , (3.20)
i=1

with the corresponding matrix representation

 
0 0 0 0

h 0 1 0 0
Hh = −  . (3.21)
20 0 −1 0
0 0 0 0

As h is turned up the ground state energy of the system is changed but for
very small h, of the order t/100, the energy is essentially unchanged.
Due to the spin-splitting for non-zero h symmetries and anticommutation
relations are fulfilled to much greater extent, which can be understood as
follows: The risk of breaking the spin-flip symmetry by the truncation of
the Hilbert space is eliminated once the symmetry itself is eliminated. Thus
there is one less constraint on the number of states kept, reflecting itself in
the symmetry and anticommutator checks.
Also the calculation of the total spin on the chain changes as a conse-
quence of non-zero h. Without a magnetic field the spin states were energy
degenerate. Thus the state of the superblock can be an arbitrary superpo-
sition of a ‘spin up’ and a ‘spin down’ state, giving an arbitrary total Sz
depending on the weight between these states. Due to the spin-splitting
effect of non-zero h the total spin on the chain becomes half-integer, as is
clearly observed in our DMRG calculations.
34
Hubbard model

3.7 Expectation values

In this section we explain how expectation values of various operators are
obtained in this implementation of DMRG. Having constructed and updated
matrices for the total particle number and spin throughout the DMRG cal-
culation it is fairly easy to obtain expectation values of these operators.
At the end of the DMRG routine we have matrix representations of the
operators we want to find expectation values for, in the current DMRG
basis, as well as a ground state vector found from the current Hamiltonian.
The left and right blocks are described by m DMRG states, and hence the
superblock basis consists of 16m2 states. Taking the total particle number
of the superblock as an example we find expectation values in either of the
two ways considered in Sect. 2.5.
Denoting the matrix representation of the particle number operator of
the superblock N

N = Nsystem ⊗ δenvironment + δsystem ⊗ Nenvironment , (3.22)

we find the total particle number by the vector-matrix-vector multiplication

­ ¯ ¯ ®
ψ0 ¯N ¯ψ0 = ψ0† N ψ0 , (3.23)

where ψ0 is the 16m2 component column ground state vector and N is the
16m2 × 16m2 matrix in Eq. (3.22).
Obtaining the particle number of the system block itself can be done in
two ways. Either you expand the representation of the number operator for
the system to the basis of the superblock by,

Ñsystem = Nsystem ⊗ δenvironment , (3.24)

and then proceed as above,

­ ¯ ¯ ®
ψ0 ¯Nsystem ¯ψ0 = ψ0† Ñsystem ψ0 , (3.25)

where Ñsystem is the 16m2 × 16m2 matrix in Eq. (3.24) and ψ0 is the 16m2
component ground state vector.
Alternatively we may use
h i
hψ0 | Nsystem |ψ0 i = Tr ψ0 † Nsystem ψ0 , (3.26)

where now ψ0 is the m × m matrix given in Eq. (3.13) and Nsystem is the
m × m matrix representation of the number operator for the system.
 35
3.8 Results

−19.656 −12.9609

−19.6565

−19.657

−19.6575 −12.9609

−19.658
Energy

Energy
−19.6585

−19.659 −12.9609

−19.6595

−19.66

−19.6605 −12.9609
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Finite system sweep Finite system sweep

(a) L = 16, U = 0. (b) L = 16, U = 1.

−8.8167 −8.8167

−8.8167 −8.8167

−8.8167 −8.8167

−8.8167 −8.8167

−8.8167
Energy

−8.8167
Energy

−8.8167 −8.8167

−8.8168 −8.8168

−8.8168 −8.8168

−8.8168 −8.8168

−8.8168 −8.8168
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 1 2 3 4 5 6 7 8 9 10
Finite system sweep Finite system sweep

(c) L = 16, U = 2. (d) L = 16, U = 2.

Figure 3.2: Convergence of the ground state energy as a function of the finite
system sweep number. Notice the non-monotonous convergence in energy when
U = 0. DMRG calculations performed using m = 24 states.

3.8 Results
Using the implementation described above we have obtained results for the
ground state energy, particle number, and spin of the complete superblock
as well as for the system block. We have targeted only the ground state
but for the final superblock we have calculated properties of exited states
as well. The description of exited states is therefore not expected to be as
accurate as the ground state.
There are odd-even oscillations in the properties of the system, i.e. they
depend on wether there is an odd or even number of sites in the system.
The superblock always consists of an even number of sites and hence does
not exhibit this kind of behavior. Plots of the ground state energy of the
superblock as function of the finite system iteration number are shown in
Fig. 3.2. As can be seen the effect of the finite system sweeps is limited,
which is surprising since Noack and White comments that most often this is
not the case [2], p. 44. Notice also the non-monotonous decrease in energy
36
Hubbard model

in Fig. 3.2(a).
We see two possible explanations for the small improvement using the
finite system algorithm: (i) We retain a quite small number of states. Most
work reported use much larger m, and Noack and White argues that the
infinite system algorithm is relatively insensitive to the number of states kept
[2]. We expect this statement to be valid only for sufficiently large values
of m since very small values would truncate too heavily in the build up of
the system. (ii) We calculate at half filling, which is a special particle-hole
symmetric filling. This introduces symmetries in the Hamiltonian, possibly
influencing the DMRG calculation.

3.8.1 Fixed U
Keeping the on-site repulsion U fixed and varying the length of the chain
L we have calculated the energy, total particle number, and total spin for
the lowest five eigenstates of the superblock. In Fig. 3.3 and 3.4 we have
plotted Ei (L)/L, Ni (L), and Szi (L) for U = 0.5 and U = 2 both with and
without a magnetic field, h = 0 and h = 1/100, and using m = 24 states in
DMRG calculations. These figures show that the behavior of the Hubbard
chain does not change qualitatively when the chain length is changed.

3.8.2 Fixed L
For fixed chain length L and varying on-site repulsion U we have performed
DMRG calculations of the energy, total spin, and total particle number for
the ground state and four lowest exited states of the superblock. In Fig. 3.5
we have plotted Ei (U ), Ni (U ), and Szi (U ) using chainlengths L = 20 and
L = 40, h = 1/100, and m = 24 states in DMRG calculations. These figures
reveal the effect of increasing on-site interaction and as expected exited
states becomes number degenerate with the ground state for sufficiently
large U .

3.8.3 Eigenspectrum of ρ
We have computed the eigenspectrum of the enlarged system block at the
last step in the DMRG calculation for L = 12, L = 14, L = 16, and L = 20
for different couplings both with and without a magnetic field. From the
ground state vector the reduced density matrix for the enlarged left system
consisting of the left block and the leftmost central site has been calculated
and the complete set of eigenvalues found. The result is seen in Fig. 3.6,
and indeed the eigenspectrum decays rather fast. Note that in the generic
DMRG step we truncate to m = 24 states.
 37

−0.49 −0.95
E0 E0
−0.5 E1 E1
−0.96
E E
2 2
−0.51 E E
3 −0.97 3
E E
4 4
−0.52
−0.98
−0.53
E /L

E /L
−0.99
i

i
−0.54
−1
−0.55
−1.01
−0.56

−0.57 −1.02

−0.58 −1.03
10 15 20 25 30 35 40 10 15 20 25 30 35 40
L L

(a) E(L)/L for U = 2. (b) E(L)/L for U = 0.5.

45 45
N0 N0
N1 N1
40 40
N N
2 2
N N
3 3
35 N 35 N
4 4

30 30
N

25 25

20 20

15 15

10 10
10 15 20 25 30 35 40 10 15 20 25 30 35 40
L L

(c) N (L) for U = 2. (d) N (L) for U = 0.5.

1 0.5
Sz,0 Sz,0
0.8 Sz,1 0.4 Sz,1
S S
0.6 z,2 0.3 z,2
S S
z,3 z,3
0.4 S 0.2 S
z,4 z,4

0.2 0.1
z

0 0
S

−0.2 −0.1

−0.4 −0.2

−0.6 −0.3

−0.8 −0.4

−1 −0.5
10 15 20 25 30 35 40 10 15 20 25 30 35 40
L L

(e) Sz (L) for U = 2. (f) Sz (L) for U = 0.5.

Figure 3.3: Lowest 5 states, Ei (L)/L, Ni (L), and Szi (L) for h = 0. Qualitatively
the properties of the Hubbard chain is independent of the chainlength. DMRG
calculations performed using m = 24 states.
38
Hubbard model

−0.95 45
E0 N0
E1 N1
−0.96 40
E N
2 2
E N
−0.97 3 3
E 35 N
4 4

−0.98
30
E /L

N
−0.99
i

25
−1
20
−1.01

−1.02 15

−1.03 10
10 15 20 25 30 35 40 10 15 20 25 30 35 40
L L

(a) E(L)/L. (b) N (L).

0.6
Sz,0
Sz,1
0.4
S
z,2
S
z,3
0.2 S
z,4

0
z
S

−0.2

−0.4

−0.6

−0.8
10 15 20 25 30 35 40
L

(c) Sz(L).

Figure 3.4: Lowest 5 states, Ei (L)/L, Ni (L), and Szi (L) for U = 0.5 and h = 1/100.
Notice the strictly half-integer spin. DMRG calculations preformed using m = 24
states.

3.9 Interpretation
The phase diagram of the Hubbard model is in general rich and contains
many different regions. It is not the purpose of this thesis to analyze in
detail the Hubbard model. We will limit ourselves to a few comments on
the data calculated using our DMRG implementation.
We have limited our implementation to a narrow band around half fill-
ing. More specifically we ensured a half filled ground state by the choice of
Hamiltonian but allowed for exited states to be away from half filling. At
half filling and zero magnetic field the ground state is a spinles state, Sz = 0.
Considering exited states we are able to see the ‘competition’ between double
occupancy, costing energy U , and flipping spins, which eliminates hopping
possibilities and thus costs energy t. We outline a few features.

Large U limit: In the large U limit the lowest exited states are number
 39

−6 −10

−8 −15

−10
−20
−12
−25
−14
−30
i

i
−16
E

E
−35
−18
−40
−20 E E
0 0
E E
1 −45 1
−22 E2 E2
E3 −50 E3
−24
E E
4 4
−26 −55
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
U U

(a) E(U ) for L = 20. (b) E(U ) for L = 40.

21.5 41.5

21 41

20.5 40.5
N

40
N

20

19.5 39.5
N N
0 0
N N
1 1
19 N2 39 N2
N3 N3
N N
4 4
18.5 38.5
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
U U

(c) N (U ) for L = 20. (d) N (U ) for L = 40.

1.5 1.5
S0 S0
z z
S1 S1
z z
1 S2 1 S2
z z
S3 S3
z z
S4 S4
0.5 z 0.5 z
z

0 0
S

−0.5 −0.5

−1 −1

−1.5 −1.5
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
U U

(e) Sz (U ) for L = 20. (f) Sz (U ) for L = 40.

Figure 3.5: Lowest 5 states, Ei (U ), Ni (U ), and Szi (U ) for h = 1/100. Sufficiently

large U makes exited states number degenerate with the ground state. DMRG
calculations performed using m = 24 states.
40
Hubbard model

U=0, h=0 0.45

0.8 U=1, h=0.01
U=0, h=0.01
0.4 U=2, h=0
U=1, h=0.01
0.7 U=2, h=0
0.35
0.6
0.3
Eigenvalue

0.5

Eigenvalue
0.25
0.4
0.2
0.3
0.15

0.2 0.1

0.1 0.05

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Nr Nr

(a) L = 12. (b) L = 14.

0.8
U=0, h=0.01 U=0, h=0
U=0, h=0 U=2, h=0
0.6 0.7

0.6
0.5

0.5
Eigenvalue

Eigenvalue

0.4
0.4
0.3
0.3

0.2
0.2

0.1 0.1

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Nr Nr

(c) L = 16. (d) L = 20.

Figure 3.6: Eigenspectrum for various lengths and on-site interactions. Notice that
a very limited number of eigenvalues dominate and very many are essentially zero.
DMRG calculations performed using m = 24 states such that we usually truncate
to 24 states.

degenerate with the ground state due to the large energy cost of leav-
ing half filling, giving always doubly occupied or empty sites. For
chainlengths L = 20 and L = 40 this kind of behavior is clearly ob-
served, as seen in Fig. 3.5(c) and 3.5(d), and this behavior seems to
be independent of the chainlength, Fig. 3.3(c). In this limit the lowest
exited states differ from the ground state by having one or more spins
flipped.
Length dependence: Qualitatively the properties of the Hubbard chain
are independent of the length of the chain, which is clearly seen in
Fig. 3.3 and 3.4. The very similar properties of chainlengths L = 20
and L = 40 in Fig. 3.5 further demonstrate this.
Magnetic field: Including a magnetic field, h 6= 0, breaks the spin symme-
try, revealed in several ways. Firstly the total spin on the superblock
 41
3.10 Analytic result

becomes half-integer, as seen in Fig. 3.4(c), 3.5(e), and 3.5(f). With no

magnetic field there are two degenerate ground states and the eigen-
routine finds and arbitrary superposition of the two, possibly giving
non-half-integer spins.
Secondly for non-zero h the lowest exited states tend to have posi-
tive superblock spin since this is the preferred direction given by the
magnetic field.

3.10 Analytic result

As benchmark for the performance of DMRG we have calculated numeri-
cally exact results in the non-interacting limit U = 0. In this limit spin
up and down electrons are treated separately. Hence we may construct a
Hamiltonian matrix using a single particle spinles basis,
 
0 1 ··· 0
 .. 
1 0 .

H0 = −t  . , (3.27)
. . . 
. . 1
0 ··· 1 0

where the dimensions of H0 is L × L. For the chain lengths we consider full

diagonalization of H0 is possible since the matrix size scales linearly with
the system size.
The ground state energy of H0 at half filling is obtained by diagonalizing
and summing up the lowest L/2 eigenvalues. Remembering the spinfactor 2
and denoting the eigenvalues ei we thus find the ground state energy of the
spinful non-interacting Hubbard model at half filling
L/2
X
E0 = 2 ei . (3.28)
i=1

A few comparisons of the numerically exact results with the corresponding

DMRG results are shown in Tab. 3.1, enabling us to benchmark the perfor-
mance of DMRG in the non-interacting limit. The agreement is very fine,
considering that e.g. the 40 site chain is described using approximately 104
states, while the entire Hilbert space contains approximately 1024 states. So
despite the heavily truncated Hilbert space produced by DMRG the ground
state properties are still described quite well.

3.11 Limitations and future work

The code produced for the Hubbard model is specialized to calculations near
half filling. But this is not the only limitation of our DMRG implementation.
42
Hubbard model

L Exact DMRG ∆E/E

12 −14.5925 −14.5895 2 · 10−4
14 −17.1335 −17.0927 2 · 10−3
20 −24.7630 −24.7183 2 · 10−3
26 −32.3969 −32.3675 9 · 10−4
30 −37.4873 −37.4472 1 · 10−3
40 −50.2156 −50.1468 1 · 10−3

Table 3.1: Comparison of exact and DMRG results for U = 0 and h = 0. Our
implementation yields quite accurate estimates of the ground state energy. DMRG
calculations performed using m = 24 states.

In this section we consider a few other limitations and discuss possible ways
to eliminate these.

Target states: The current code is limited to a narrow band around half
filling, always targeting a half filled ground state. We are not able to
handle arbitrary fillings or e.g. restraining the spin using our imple-
mentation.
Expanding our implementation to treat arbitrary fillings probably im-
plies reorganizing the code to make full use of quantum numbers, which
also speeds up the calculation. We expect object oriented languages
like C++ or Python to be better suited for this layout than Mat-
Lab.
Calculational time: The system size currently accessible is heavily limited
by the long calculational time. Many further optimizations of the code
can be made, which could improve on this. It would be interesting to
rewrite the code in either of the programming languages mentioned
above, rethinking the structure of the program. Another potential
speedup of the code is the wavefunction transformations described in
Sect. 2.4.2, which has not yet been implemented.
Precision: We have not been able to improve the description by increasing
the number of states kept in the DMRG. This would most likely enable
us to e.g. maintain anticommutation relations to larger extent. It
should be noted that most work presented in the literature use larger
m than we do – some even use m ∼ 500.
Likewise we have not been able to really test what values of m gave
best results. There may be significant improvements of the precision
by increasing or decreasing m moderately, as discussed in Sect. 2.6. We
have mainly used m = 24 since this is the choice used by M. Andersen
et al. on a similar setup [15]. Whether this is actually the optimal
choice should be checked thoroughly.
 43
3.12 Conclusions

Benchmark for U 6= 0: So far we have only tested the performance of

DMRG in the non-interacting limit. To test DMRG further it would
be interesting to compare results also for U 6= 0. Simple approxima-
tions and perturbative results could be compared where applicable to
give a better understanding of DMRG in those limits.
Also we have only benchmarked the ground state energy and verified
that DMRG finds a zero spin ground state. It would be interesting
to also benchmark properties of exited states to see to what extend
these results are reliable. One possible selfconsistency check could be
to include a few exited states as target states and see how the results
changed. Probably more that m = 24 states are needed to main-
tain sufficient precision in such calculations as the general precision of
DMRG decreases when more states are targeted.

3.12 Conclusions
In this chapter we have implemented DMRG on the one dimensional Hub-
bard model at half filling, giving technical details of the implementation as
well as pointing out some difficulties in the implementation. For various
chain lengths we have successfully calculated properties of the ground state
as well as exited states, more specifically the energy, particle number, and to-
tal spin of the states. In the non-interacting limit we have compared DMRG
and analytic results and found the DMRG ground state energies very accu-
rate, with relative energy deviations of the order 10−3 −10−4 . To expand the
model we included a small magnetic field and made additional calculations.
Expected features of the spin splitting are clearly observed in the results.
We briefly commented on the results found using our DMRG implementa-
tion, and finally we discussed the limitations and future improvements of
the implementation.
44
Chapter 4

Quantum dot with single

infinite lead

In recent years quantum dots have received much attention experimentally

as well as theoretically, also within DMRG. One example is Berkovits who
studies a quantum dot coupled to a single one-dimensional semi-infinite
lead using DMRG calculations [16]. In this chapter we follow the work of
Berkovits and implement DMRG on this setup, calculating the occupation
of the dot versus chemical potential. In the non-interacting limit we work
out the occupations analytically using Green’s functions and compare to the
DMRG results. To extend the capability of our DMRG implementation we
attempt to calculate the spectral function of the dot using a simple Lanczos
scheme. Finally we comment on results found with our implementation.

4.1 Hamiltonian and matrices

The model considered here is a quantum dot with NDot levels coupled to
a single one-dimensional semi-infinite lead. On the dot there is coulomb
interaction – modelled by a repulsion term for double occupancy, the elec-
trons are treated as spinles, and electrons in the leads are assumed to be
non-interacting,

N
XDot N
XDot

HDot = (²i − Vg )a†i ai +U a†i ai a†j aj , (4.1a)

i=1 i>j
∞
X
HLead = −t (c†j cj+1 + c†j+1 cj ), (4.1b)
j=1
N
XDot

HDL = Vi (a†i c1 + c†1 ai ), (4.1c)

i=1

45
46
Quantum dot with single infinite lead

and including a chemical potential µ, the complete Hamiltonian is

µ NX
Dot ∞
X ¶
† †
H = HDot + HLead + HDL − µ ai ai + cj cj . (4.2)
i=1 j=1

where ai represents states in the dot and ci represents states in the lead.
Electrons are treated as spinles so the real space basis used for the single
sites is {|0i, |1i}. In this basis the basic operators have the following matrix
representations
µ ¶
0 1
ai = , (4.3a)
0 0
µ ¶
0 1
ci = , (4.3b)
0 0
µ ¶
1 0
Pi = , (4.3c)
0 −1
µ ¶
0 0
Ni = = a†i ai = c†i ci . (4.3d)
0 1

Matrices Pi and Ni are the fermionic sign operator and the number operator
for the single sites respectively.
The basic idea of enlarging the basis is the same as for the Hubbard
model considered in Chap. 3. However, here we study the semi-infinite lead
limit, eliminating the use of the finite system algorithm, and therefore fewer
matrices have to be stored. We specialize our calculations to two dot levels,
i.e. NDot = 2.

Hermiticity of H
Just as with the Hubbard model there are numerical imprecisions preventing
the Hamiltonian from being completely hermitian. We eliminate these by
the replacement

H + H†
→ H. (4.4)
2
This enforces hermiticity onto any matrix and it should therefore be verified
that max(H − H † ) ∼ 10−16 before using this replacement, as explained in
Sect. 3.1.

4.2 Infinite system sweep

The superblock setup for this system is two quantum dots connected by a
one-dimensional lead, i.e. the right block is a spatial reflection of the left,
 47
4.2 Infinite system sweep

(a) First superblock used in the in-

finite system sweep.

(b) Intermediate superblock with

left and right blocks enlarged.

(c) Second superblock used in the infinite sys-

tem sweep.

Figure 4.1: Superblocks used in the first step of the infinite system algorithm.

as seen in Fig. 4.1. The infinite system sweep is initiated by constructing a

left block consisting of the dot and one lead site as seen in Fig. 4.1(a),
µX
2 ¶
† †
HL = HDot + HDL − µ ai ai + c1 c1 , (4.5a)
i=1
cL = PDot ⊗ c1 , (4.5b)
PL = PDot ⊗ P1 , (4.5c)

where it is understood that matrix representations of operators in Eq. (4.5a)

are in the enlarged basis of the dot and one lead site, i.e. 8 × 8 matrices.
Note that cL is the matrix representation of the annihilation operator on
the rightmost site in the left block, needed to connect the left block to the
first central site.
Additionally matrices for the number operator of the two dot levels are
constructed,

N1Dot = N1 ⊗ δ2Dot ⊗ δ1 , (4.6a)

N2Dot = δ1Dot ⊗ N2 ⊗ δ1 , (4.6b)

where δiDot denotes a unit matrix in the Hilbert space of dot site i, and δj
denotes a unit matrix in the Hilbert space of lead site j.
48
Quantum dot with single infinite lead

The basic DMRG step is to include an additional site in the lead and
truncate to m DMRG states. The enlargement is

H̃L = HL ⊗ δ2 − µδL ⊗ N2
³¡ ¢¡ ¢ ¡ ¢¡ ¢´
−t c†L ⊗ δ2 PL ⊗ c2 + PL ⊗ c†2 cL ⊗ δ2 , (4.7a)
c̃L = PL ⊗ c2 , (4.7b)
P̃L = PL ⊗ P2 , (4.7c)
Dot
Ñi = NiDot ⊗ δ2 , i = 1, 2, (4.7d)

where
¡¡ −µδL¢¡⊗ N2 is the
¢ ¡chemical¢¡potential¢¢term for site 2 in the lead and
−t c†L ⊗ δ2 PL ⊗ c2 + PL ⊗ c†2 cL ⊗ δ2 is the hopping between sites 1
and 2 in the lead. Similarly H̃R and c̃R are found and we thus have complete
descriptions of the enlarged left and right blocks in Fig. 4.1(b). Note that
this step has doubled the size of the Hilbert space for the left and right
blocks respectively.
Having computed the ground state of the 4 site superblock in Fig. 4.1(a),
currently a 256 component vector,

|ψi = (v1 , v2 , · · · , v256 ) , (4.8)

we construct the RDM for the enlarged left block in Fig. 4.1(b)

|0, 0, 0, 0i |0, 0, 0, 1i · · · |1, 1, 1, 1i

 
ψ = |0, 0, 0, 0i v1 v2 ··· v16 (4.9a)
|0, 0, 0, 1i  v17 v18 v32 
 
..  .. .. .. ,
.  . . . 
|1, 1, 1, 1i v240 v241 ... v256

ρL = ψψ † . (4.9b)

Note that e.g. the state |1, 1, 1, 1i for the left block means all dot and lead
sites in the left block occupied. Diagonalizing ρL and keeping m column
eigenstates uα , conveniently collected in a matrix,

O = (u1 , u2 , · · · , um ) , (4.10)

we find the new left block matrices in a truncated basis by

O† H̃L O → HL , (4.11a)
†
O c̃L O → cL , (4.11b)
†
O P̃L O → PL , (4.11c)
†
O ÑiDot O → NiDot , i = 1, 2, (4.11d)

with similar expressions for the right block. Note that if m is larger than the
total number of states in the enlarged left block no truncation is performed,
 49
4.3 Computing the ground state

and in that case all eigenstates of ρL should be kept. The truncation reduces
to a basis shift in those instances.
With these truncated representations of the left and right blocks we may
construct the superblock Hamiltonian consisting of the truncated left and
right blocks, and two central sites described in their full basis, as shown
in Fig. 4.1(c). The truncated left block is represented by the matrices in
Eq. (4.11) each being m × m, and hence superblock matrices are in general
4m2 × 4m2 .
This completes the basic step, and we may repeat the above procedure
until convergence to the infinite lead limit is reached. It may be difficult to
determine when convergence is reached. We have chosen to iterate a fixed
and sufficiently large number of times, see also Sect. 4.5.1.

4.3 Computing the ground state

In this setup it is not possible to use the same trick as for the Hubbard model,
finding and calculating in the subspace of a given number of particles. Due
to the open setup the number of particles in the system is determined by
the chemical potential, and thus we cannot restrict the particle number in
this setup. Therefore we have to compute the ground state of the entire
superblock Hamiltonian rather than in a subspace.

4.4 Expectation values

Expectation values of operators are obtained in either of two ways as ex-
plained in Sect. 2.5. At the end of the DMRG calculation the left and right
blocks are described by m DMRG states, and hence the superblock basis
contains 4m2 states. Using the number operator for the dot as an example
we find the expectation value

ÑDot = NDot ⊗ δenvironment , (4.12a)

­ ¯ ¯ ®
ψ0 NDot ¯ψ0 = ψ0† ÑDot ψ0 ,
¯ (4.12b)

where ÑDot is a 4m2 × 4m2 matrix and ψ0 is the 4m2 component column
ground state vector. Alternatively we may evaluate instead
h i
hψ0 | NDot |ψ0 i = Tr ψ0† NDot ψ0 , (4.13)

where now ψ0 is the m × m matrix given in Eq. (4.9a) and NDot is an

m × m matrix. With either of these two methods we can thus calculate the
occupation of the dot using DMRG for arbitrary but fixed parameter values
U , Vi , and µ.
50
Quantum dot with single infinite lead

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 0 0.5 −0.5 0 0.5
µ µ

(a) m = 8. (b) m = 16.

1

0.9

0.8

0.7

0.6
Occupation

0.5

0.4

0.3

0.2
n1
0.1
n
2
0
−0.5 0 0.5
µ

(c) m = 32.

Figure 4.2: Occupations for various values of m. The occupations change qualita-
tively when increasing m to 32, indicating that 8 and 16 states are insufficient for
this calculation. Parameter values used are U = 0.2 and Vi = 0.25. U , Vi , ², and µ
are given in units of t.

4.5 Results
In this section we present results obtained with our implementation of
DMRG as described above. We have mainly been interested in reproduc-
ing data already presented by Berkovits. All calculations presented in this
section were performed with Vg = 0. Initially we used m = 8 which quickly
turned out to be insufficient to obtain accurate results. Also m = 16 turned
out to be insufficient and after consulting with Professor Berkovits we ended
up using m = 32. This turned out to be nearly sufficient, but also close to
the capability of the available hardware. In Fig. 4.2 we show a typical im-
provement when increasing m, where it is clearly seen that the numerical
irregularities are reduced when increasing m.
With m = 32 we have calculated the occupation of the levels on the
dot as a function of the chemical potential. For each set of values U , Vi ,
and µ we have performed a full DMRG calculation of the occupation of the
 51
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) U = 0, Vi = 1. (b) U = 0.1, Vi = 1.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) U = 0.2, Vi = 1. (d) U = 0.4, Vi = 1.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(e) U = 0, Vi = 0.75. (f) U = 0.1, Vi = 0.75.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(g) U = 0.2, Vi = 0.75. (h) U = 0.4, Vi = 0.75.

Figure 4.3: Occupations for ² = [−0.1, 0.1], m = 32 states, and 125 sites in the
lead. The strong coupling Vi makes the filling of the levels occur at nearly the same
energy, while the repulsion U pushes the filling of the levels upwards in energy. U ,
Vi , ², and µ are given in units of t.
52
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) U = 0, Vi = 0.5. (b) U = 0.1, Vi = 0.5.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) U = 0.2, Vi = 0.5. (d) U = 0.4, Vi = 0.5.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(e) U = 0, Vi = 0.25. (f) U = 0.1, Vi = 0.25.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(g) U = 0.2, Vi = 0.25. (h) U = 0.4, Vi = 0.25.

Figure 4.4: Occupations for ² = [−0.1, 0.1], m = 32 states, and 125 sites in the
lead. The coupling Vi renormalizes the levels, while the repulsion U suppresses the
filling of the second level. U , Vi , ², and µ are given in units of t.
 53
4.5 Results

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) U = 0, Vi = 0.05. (b) U = 0.1, Vi = 0.05.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) U = 0.2, Vi = 0.05. (d) U = 0.4, Vi = 0.05.

Figure 4.5: Occupations for ² = [−0.1, 0.1], m = 32 states, and 125 sites in the lead.
The levels are essentially unchanged, but the repulsion U suppresses the filling of
the second level to ²̃2 = ²2 + U . U , Vi , ², and µ are given in units of t.

dot, obtaining occupations n1 (µi ) and n2 (µi ), and therefore the calculation
of such graphs can be quite lengthy. Figures of the occupations of the
single levels have been made for strong (U = 0.4), medium (U = 0.2), weak
(U = 0.1), and non-interacting (U = 0) cases. We have collected figures
with the same coupling to the leads, revealing the effect of increasing on-site
interaction. The results are shown in Fig. 4.3, 4.4, and 4.5. Additionally in
Fig. 4.6 we show the total occupation on the dot n = n1 + n2 for various
interactions and couplings.

4.5.1 Convergence

To test the convergence of the algorithm to the infinite lead limit we have
calculated occupations using lead lengths 75 and 125 for various couplings
and on-site interactions, and computed the difference between these two sets
54
Quantum dot with single infinite lead

2 2

1.8 1.8

1.6 1.6

1.4 1.4
Total occupation

Total occupation
1.2 1.2

1 1

0.8 0.8

0.6 V =0.05 0.6 V =0.05

i i
V =0.25 V =0.25
i i
0.4 Vi=0.5 0.4 Vi=0.5
Vi=0.75 Vi=0.75
0.2 0.2
V =1 V =1
i i
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) U = 0. (b) U = 0.1.

2 2

1.8 1.8

1.6 1.6

1.4 1.4
Total occupation

Total occupation

1.2 1.2

1 1

0.8 0.8

0.6 V =0.05 0.6 V =0.05

i i
V =0.25 V =0.25
i i
0.4 Vi=0.5 0.4 Vi=0.5
Vi=0.75 Vi=0.75
0.2 0.2
V =1 V =1
i i
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) U = 0.2. (d) U = 0.4.

Figure 4.6: Total ocupations for ² = [−0.1, 0.1], m = 32 states, and 125 sites in the
lead. For small U notice the transition from two dot levels when Vi = 0.05 to a
single localized state when Vi = 1. U , Vi , ², and µ are given in units of t.

of results.1
The majority of these results vary on the scale 10−3 so in these cases
the algorithm has converged. There are however also cases where the scale
is 10−2 making it more questionable if convergence really has been reached.
Examining more carefully in which cases these larger deviations occur we
find two cases: (i) When there are plateaus in the occupation, and (ii) when
there is a µ-point in the step in occupation. The first case indicates that
some of the numerical problems present with 75 sites in the lead are due
to lacking convergence of the algorithm. The second case is quite under-
standable; if the position of the step in occupation moves even slightly when
iterating further, a point located in the steep part of the step would move
substantially. Either of the two however indicates lacking convergence to
some degree.
1
The length of the lead corresponds to the number of iterations in the DMRG algorithm,
increasing the size of the lead in the left block by one site at each iteration.
 55
4.5 Results

An example is shown in Fig. 4.7 where the plateaus present with 75

sites in the lead are partly eliminated when increasing the lead length to
125 sites. Increasing the lead length to 200 sites gives practically no further
improvements. In Fig. 4.7(d) we have plotted the difference in occupation for
a calculation with 200 sites and a calculation with 125 sites in the lead. All
differences are on the scale 10−3 and from this we conclude that convergence
is achieved with 125 sites in the lead.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 0 0.5 −0.5 0 0.5
µ µ

(a) 75 sites in the lead. (b) 125 sites in the lead.

−3
x 10
1 2

0.9
1
0.8

0.7
0
0.6
Occupation

Occupation

0.5 −1

0.4
−2
0.3

0.2
−3
n1 n1
0.1
n n
2 2
0 −4
−0.5 0 0.5 −0.5 0 0.5
µ µ

(c) 200 sites in the lead. (d) Difference in occupations when using
200 and 125 sites in the lead.

Figure 4.7: Occupations for ² = [−0.1, 0.1], Vi = 0.25, m = 32 states, and 125 sites
in the lead. The results are improved when using 125 rather than 75 sites in the
lead. No further improvement is achieved by using 200 sites in the lead. U , Vi , ²,
and µ are given in units of t.

4.5.2 Spectral function

To calculate the spectral function of the dot levels using DMRG methods
we use the Lanczos method reviewed in Sec. 2.5.2, in particular a continued
56
Quantum dot with single infinite lead

fraction representation of the Green’s functions

­ ¯ ¯ ®
Ga† a (z) = ψ0 ¯a†i (z − H)−1 ai ¯ψ0 , (4.14a)
i i
­ ¯ ¯ ®
Ga a† (z) = ψ0 ¯ai (z + H)−1 a†i ¯ψ0 , (4.14b)
i i

in a Lanczos basis where

­ ¯ † ¯ ®
ψ0 ¯ai ai ¯ψ0
Ga† a (z) = b21
, (4.15a)
i i
z − a0 − b2
2
z−a1 − z−···
­ ¯ †¯ ®
ψ0 ¯ai ai ¯ψ0
Ga † (z) = , (4.15b)
i ai b21
z + a0 − b2
2
z+a1 − z+···

We find the spectral function by

A(ω) = (4.16)
³ ´
−2 lim Im Ga† a (ω + E0 + iη) + Ga a † (ω − E0 + iη) .
η→0+ i i i i

In the previous calculations the DMRG basis was optimized only for the
ground state of the system. To optimize the basis for properties of exited
states as well we include at each step the first Lanczos vector as target state.
When convergence to the infinite lead limit has been achieved we use 200
Lanczos states in the calculation of Ga† a (ω) and Ga a† (ω) in Eq. (4.15).
i i i i
We have
­ ¯ † ¯ ® limited our calculations
­ ¯ †¯ ® to i = 1. Note that the evaluation
¯ ¯ ¯ ¯
of ψ0 a1 a1 ψ0 and ψ0 a1 a1 ψ0 requires matrix representations of the
combined operators a†1 a1 and a1 a†1 constructed and updated throughout the
calculation, as explained in Sect. 2.5.1. We use different DMRG calculations
to find Ga† a and Ga a† , since the DMRG basis has to be optimized to
1 1 1 1 ¯ ®
different additional target states in the two cases, namely the states a1 ¯ψ0
¯ ®
and a†1 ¯ψ0 respectively. In App. D we show what kind of convergence we
expect based on a truncated representation of a continued fraction formally
similar to the ones in Eq. (4.15). In Fig. 4.8 plots of the two terms in the
spectral function are shown, A1 = Ga† a + Ga a† for 3 values of η and µ = 2.
1 1 1 1
The plots are similar to the truncated continued fraction in App. D, but the
known structure of the spectral function is not reproduced.
Based on work reported in the literature we expect that keeping more
states would improve these calculations. We have not been able to target
additional Lanczos states and maintain precision, which we expect is the
major reason for the failure of these calculations. Wang et al. [2] use m
from 100 and upwards, whereas we use m ≤ 32. They consider a different
model but we still expect the small m to be the major difficulty in these
calculations.
 57
4.6 Interpretation

30 90
Gcd c(ω)) Gcd c(ω))
Gccd(ω)) 80 Gccd(ω))
25
70

20 60

50
15
40

10 30

20
5
10

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) η = 10−2 . (b) η = 10−3 .

700
Gcd c(ω))
Gccd(ω))
600

500

400

300

200

100

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω

(c) η = 10−4 .

Figure 4.8: DMRG calculation of the spectral function, A1 = Ga† a1 + Ga1 a† , using
1 1
m = 32 and 125 sites in the leads, U = 0, µ = 2, and Vi = 1. The analytically
known structure of A1 (ω) is not reproduced. U , Vi , ², µ, and ω are given in units
of t.

4.6 Interpretation
In this section we briefly comment on the results found using our DMRG
implementation. We distinguish two different parameters, the coupling to
the leads Vi and the on-site repulsion U .
For very weak coupling to the leads the levels of the dot are essentially
unchanged, only the second level is renormalized. The filling of the first
level occurs when the chemical potential crosses the energy of the level, and
the filling of the second level occurs when the chemical potential reaches the
renormalized energy ²̃2 = ²2 + U , as seen in Fig. 4.6. When the coupling is
turned up there is a crossover to a single localized state on the dot, revealed
by a sharp jump of unit magnitude in the total occupation, located between
²1 and ²2 , as seen in Fig. 4.6(a).
When interactions and couplings both are turned up some features of
58
Quantum dot with single infinite lead

the non-interacting limit persist but due to the interactions the position of
the jump in occupation of the second level and the localized state is shifted,
as seen in Fig. 4.6(b)-4.6(d). All jumps are smeared by the interactions and
for stronger interactions simple arguments are insufficient.

4.7 Analytic result

In the non-interacting limit it is possible to find analytic results for the occu-
pation using Green’s functions theory, [17, 18]. In this limit the Hamiltonian
is

H = HD + HL + HDL , (4.17a)
b
X b
X
HD = (²i − Vg ) a†i ai ≡ εi a†i ai , (4.17b)
i=a i=a
∞ ³
X ´ ∞
X
HL = −t c†j cj+1 + c†j+1 cj + ε c†j cj , (4.17c)
j=1 j=1
b
X ³ ´
HDL = Vi a†i c1 + c†1 ai . (4.17d)
i=a

In order not to clutter up the notation we use indices a and b to denote

the dot levels and 1, 2, · · · to denote lead sites. Also in retarded Green’s
functions it is understood that a convergence should be included ω → ω +iη.
We consider the full Green’s functions on the dot generally defines as
­© † 0 ª®
GR 0 0
ij (t − t ) = −iθ(t − t ) ai (t), aj (t ) . (4.18)

In App. E we show how the equations of motion (EOMs) for the relevant
Green’s functions are computed. In particular for GR
aa we find the series of
EOMs

(i∂t − εa )GR 0 0 R 0
aa (t − t ) = δ(t − t ) + Va G1a (t − t ), (4.19a)
(i∂t − ε)GR
1a (t −t) = 0
Va GR 0 R 0
aa (t − t ) + Vb Gba (t − t ) (4.19b)
−tGR 0
2a (t − t ),
(i∂t − ε)GR 0
2a (t − t ) = −tGR 0 R 0
1a (t − t ) − tG3a (t − t ), (4.19c)
..
.
(i∂t − ε)GR 0 R 0
ja (t − t ) = −tGj−1a (t − t ) (4.19d)
−tGR
j+1a (t
0
− t ), j ≥ 2,

where the hybrid Green’s function is

­© ª®
GR 0 0 † 0
ja (t − t ) = −iθ(t − t ) cj (t), aa (t ) , j ≥ 2. (4.20)
 59
4.7 Analytic result

These EOMs can be understood by a relatively simple argument, taking

GR1a as an example: From the first site in the lead there are three possible
hoppings, to site a on the dot via hopping element Va , to site b on the dot
via hopping element Vb , and finally to site 2 in the lead via hopping element
−t, and of course it is possible to stay on site 1, with energy ε.
Fourier transforming we thus find the EOMs

(ω − εa )GR R
aa (ω) = 1 + Va G1a (ω). (4.21a)
(ω − ε)GR
1a (ω) = Va GR
aa (ω) + Vb GR
ba (ω) − tGR
2a (ω). (4.21b)
..
.
(ω − ε)GR R R
ja (ω) = −tGj−1a (ω) − tGj+1a (ω), j ≥ 2. (4.21c)

Eq. (4.21a)-(4.21c) form a semi-infinite set of equations. In App. E we show

that the final solutions for GR R
aa (ω) and Gbb (ω) are

1
GR
aa (ω) = , (4.22a)
ΣR R
ab (ω)Σba (ω)
ω − εa − ΣR
aa (ω) − ω−εb −ΣRbb (ω)
1
GR
bb (ω) = . (4.22b)
ΣR R
ba (ω)Σab (ω)
ω − εb − ΣR
bb (ω) − ω−εa −ΣRaa (ω)

4.7.1 Occupation
Having found GR R
aa (ω) and Gbb (ω) it is possible to find the spectral function
and hence the occupations of the levels. The relevant selfenergies are

ΣR 2 R
aa (ω) = Va G̃11 (ω), (4.23a)
ΣR
bb (ω) = Vb2 G̃R
11 (ω), (4.23b)
R
Σab (ω) = Va Vb G̃R
11 (ω), (4.23c)
ΣR
ba (ω) = Vb Va G̃R
11 (ω), (4.23d)

so that G̃R11 (ω) essentially determines the complex structure of the self-
energies. In App. E we show that G̃R 11 can be expressed as
 √

 ω/2t 1−(ω/2t)2
 t − i t , |ω| < 2t,
R
G̃11 (ω) = √ (4.24)


 ω/2t−sign(ω) (ω/2t)2 −1 , |ω| ≥ 2t,
t

and in Fig. 4.9(d) the real and imaginary part of G̃R

11 are plotted.
In equilibrium the lesser Green’s function generally defined as
­ † ®
G<
ij (t) ≡ i aj ai (t) , (4.25)
60
Quantum dot with single infinite lead

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 n , V =0.05 0.6 n , V =0.05

Occupation

Occupation
1 1 1 2
n , V =0.05 n , V =0.05
2 1 2 2
0.5 n , V =0.25 0.5 n , V =0.25
1 1 1 2
n , V =0.25 n , V =0.25
0.4 2 1 0.4 2 2
n , V =0.5 n , V =0.5
1 1 1 2
0.3 n , V =0.5 0.3 n , V =0.5
2 1 2 2
n , V =0.75 n , V =0.75
1 1 1 2
0.2 n , V =0.75 0.2 n , V =0.75
2 1 2 2
n1, V1=1 n1, V2=1
0.1 0.1
n2, V1=1 n2, V2=1
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
µ µ

(a) V1 = V2 . (b) V1 = 2V2 .

1 1

0.9
0.5
0.8

0.7
0
0.6 n , V =0.05
Occupation

1 1
n , V =0.05
2 1
11

0.5 n , V =0.25 −0.5

G

1 1
n , V =0.25
0.4 2 1
n , V =0.5
1 1 −1
0.3 n , V =0.5
2 1
n , V =0.75
1 1
0.2 n , V =0.75
2 1 −1.5
n1, V1=1 real(G (ω))
0.1 11
n2, V1=1 imag(G (ω))
11
0 −2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −4 −3 −2 −1 0 1 2 3 4
µ ω

(c) V2 = 2V1 . (d) Real and imaginary part of G̃11 .

Figure 4.9: Occupations in the non-interacting case for ² = [−0.1, 0.1], computed
using Green’s function theory. U , Vi , ², µ, and ω are given in units of t.

may be reexpressed using the Fluctuation Dissipation Theorem [18], p. 171.

Particularly we find
G<
ii (ω) = inF (ω)Aii (ω), (4.26)
where nF = (e−β(µ−ω) +1)−1 is the Fermi distribution and Aii is the spectral
function for level i. Hence we can express the occupation of the dot level i
by,
­ ®
ni ≡ a†i ai
= −iG< (t = 0)
Z ii∞
dω <
= −i Gii (ω)
−∞ 2π
Z ∞
dω
= nF (ω)Aii (ω). (4.27)
−∞ 2π

Using also the definition of the spectral function, Aii (ω) = −2ImGR ii (ω),
and the fact that in the zero temperature limit nF is a step function at the
 61
4.8 Conclusions

chemical potential, we find the final solution for the occupation,

Z µ
dω
ni (µ) = − ImGR ii (ω), µ ≤ 2t. (4.28)
−∞ π

Note that with the parameter values we use, the renormalized levels remain
within the band, but for generic parameter values this is not necessarily the
case.
Expressing the occupations analytically involves finding the imaginary
part of the Green’s function, which yields very large expressions that do not
simplify. Therefore we have performed these calculations numerically. Some
results obtained are shown in Fig. 4.9 for various couplings to the lead.

4.7.2 Comparison
Having performed both analytic and DMRG calculations of the occupation
of the dot as a function of chemical potential in the non-interacting limit
we are able to benchmark DMRG. In Fig. 4.10 and 4.11 both DMRG and
analytic results are plotted simultaneously, in rather consistency.
Calculations have been performed for m = 8, m = 16, and m = 32,
which enables us to check if the deviations of the DMRG from the analytic
result is caused by a too large truncation, or wether other problems are
significant as well. The calculation with m = 32 still shows irregularities
for certain values of µ, but the curves are generally very much better. The
numerical problems were not completely solved by increasing the number of
states kept to m = 32 but we expect that increasing m even further would
solve the remaining.
However, based on the data presented here we conclude that indeed
DMRG reproduces the analytic data quite well, with the largest deviations
in the vicinity of the step in occupation. Even with m = 8 the resemblance
of DMRG and analytic data is very clear, although also unphysical features
caused by numerical problems are present. Considering that in this case the
dot and (in principle) infinite lead is described using a total of 8 states the
performance of DMRG is quite impressing.

4.8 Conclusions
We successfully implemented DMRG on a quantum dot coupled to a single
lead in the semi-infinite lead limit. Using a setup identical to Berkovits’s
we performed DMRG calculations of the occupation of the dot for various
couplings to the leads, both with and without on-site interaction.
Calculations similar to Berkovits’s were performed using our implemen-
tation of DMRG. We explored the precision of DMRG versus the number
of states kept and found that the precision of calculations depends heavily
62

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) m = 8, Vi = 0.75. (b) m = 16, Vi = 0.75.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) m = 8, Vi = 0.5. (d) m = 16, Vi = 0.5.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(e) m = 8, Vi = 0.25. (f) m = 16, Vi = 0.25.

Figure 4.10: Comparison between analytic and DMRG results for ² = [−0.1, 0.1]
and U = 0. Calculated with m = 8 and m = 16 states, and 250 and 175 sites in
the lead respectively. Notice the improvement in the DMRG results when using
m = 16 rather than m = 8. U , Vi , ², and µ are given in units of t.
 63

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3
n DMRG n DMRG
1 1
0.2 n2 DMRG 0.2 n2 DMRG
n1 Analytic n1 Analytic
0.1 0.1
n Analytic n Analytic
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(a) Vi = 1. (b) Vi = 0.75.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3
n DMRG n DMRG
1 1
0.2 n2 DMRG 0.2 n2 DMRG
n1 Analytic n1 Analytic
0.1 0.1
n Analytic n Analytic
2 2
0 0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5 −0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ µ

(c) Vi = 0.5. (d) Vi = 0.25.

0.9

0.8

0.7

0.6
Occupation

0.5

0.4

0.3
n DMRG
1
0.2 n2 DMRG
n1 Analytic
0.1
n Analytic
2
0
−0.5 −0.25 −0.1 0 0.1 0.25 0.5
µ

(e) Vi = 0.05.

Figure 4.11: Comparison between analytic and DMRG results for ² = [−0.1, 0.1]
and U = 0. Calculated with m = 32 states, and 125 sites in the lead. The DMRG
and analytic results agree very well with the largest deviations in the vicinity of
the step. U , Vi , ², and µ are given in units of t.
64
Quantum dot with single infinite lead

on m. In order to produce quality data we needed to use m = 32, which

eliminated the majority of numerical problems. In the non-interacting limit
we compared DMRG and analytic calculations and found the DMRG results
quite accurate. This comparison was made for various couplings to the lead.
Using even a very limited number of states the resemblance between DMRG
and analytic results is very clear. Calculations for finite on-site repulsion
were made for various couplings to the lead and we briefly commented on
these results. We attempted a combined Lanczos and DMRG calculation
of the spectral function of the dot, but were not successful, and discussed
probable explanations for this.
Chapter 5

Quantum dot with two

infinite leads

The natural setup when considering transport is a system connected to two

leads. One such system studied by many people is a quantum dot coupled to
one-dimensional leads in various configurations and for various parameters.
Berkovits et al. have performed DMRG calculations on a setup in which a
quantum dot is coupled to multiple leads calculating the charging for varying
number of open channels [19].
In this section we use a simpler setup, performing DMRG calculations
on a quantum dot coupled to two one-dimensional semi-infinite leads. Using
as starting point the existing code for a quantum dot with a single lead, we
have performed the most straightforward generalization to two leads. Using
this we calculate the occupation of the dot versus chemical potential in the
system. To approach transport we calculate the occupation of the biased
dot varying the gate voltage. We use different chemical potentials for the
two leads to simulate finite bias. In the non-interacting limit we calculate
analytic Green’s functions results to benchmark both sets of DMRG cal-
culations. Finally we comment on the results calculated using our DMRG
implementation.

5.1 Hamiltonian
Considering a quantum dot with NDot levels and treating the electrons as
spinles the Hamiltonians are essentially the same as for the single lead case,
only here there is an additional lead and couplings of this lead to the quan-
tum dot

N
XDot N
XDot

HDot = (²i − Vg )a†i ai +U a†i ai a†j aj , (5.1)

i=1 i>j

65
66
Quantum dot with two infinite leads

∞
X
HRight Lead = −t (c†j cj+1 + c†j+1 cj ), (5.2a)
j=1
−∞
X
HLef t Lead = −t (c†j cj−1 + c†j−1 cj ), (5.2b)
j=−1
N
XDot ³ ´
HDL = ViR (a†i c1 + c†1 ai ) + ViL (a†i c−1 + c†−1 ai ) , (5.2c)
i=1

and including a chemical potential µ, the complete Hamiltonian is

³ NX
Dot ∞
X ¡ † ¢´
H = HDot + HLead + HDL − µ a†i ai + cj cj + c†−j c−j . (5.3)
i=1 j=1

All site matrices are identical to the single lead case, and we refer to Sect. 4.1
for details. Again we specialize to two dot levels, i.e. NDot = 2.

Hermiticity of H
Also here we find numerical imprecisions preventing the Hamiltonian from
being completely hermitian. We eliminate these by replacing

H + H†
→ H, (5.4)
2
which enforces hermiticity on any matrix and hence should be used with
caution, as explained in Sect. 3.1.

4 6 8 10
2

1
3 5 7 9
(a) Superblock setup for the quantum dot with (b) Numbering of the left
2 leads. block basis states.

Figure 5.1: Superblock configuration and numbering of the basis states for the left
block.

5.2 Superblock setup

We use a superblock setup, in which two leads connect two quantum dots as
seen in Fig. 5.1(a). Note that this is not a real space picture but a convenient
way to picturize the superblock setup. The numbering of the states we use,
 67
5.3 Infinite system sweep

(a) (i), initial superblock. (b) (ii), initial superblock.

(c) (i), superblock with the (d) (ii), superblock with

‘right’ lead enlarged. both leads enlarged.

(e) (i), truncated super- (f) (ii), truncated super-

block. block.

(g) (i), superblock with the (h) (ii), superblock with

‘left’ lead enlarged. both leads enlarged.

Figure 5.2: Two different methods of enlarging the leads.

illustrated in Fig. 5.1(b), is chosen in order to have an environment block

that resembles the system block itself. The dot level i couples to the left
and right leads via couplings ViL and ViR respectively, so there are a total
of four different coupling parameters.

5.3 Infinite system sweep

The basic infinite system step is similar to the single lead setup. The differ-
ence is that here we have two leads and both should be enlarged. This can
be done in several different ways, and we have considered two: (i) Including
a site in one of the leads at each step and then alternate the lead that is
enlarged, and (ii) including a site in both the leads at each step. The two
methods are illustrated in Fig. 5.2.
Enlarging only one lead in the generic DMRG step the ground state of the
superblock has to be computed twice as many times when building up a given
lead length. A number of matrices are updated at each step, making the
total computational time for diagonalization and updating matrices roughly
68
Quantum dot with two infinite leads

twice as long for a given lead length, when only enlarging one lead. Enlarging
both leads at each step turned out to be approximately four times faster.
Therefore we chose to use method (ii) and enlarge both leads at each DMRG
step. The drawback of this choice is that the truncation performed at each
step is larger, possibly reducing the overall precision. Unfortunately we did
not have the time to test this. A careful comparison of the two possibilities
should be performed before conclusions regarding the relative precision can
be made.

5.4 Results
In this setup as in the previous two expectation values of operators can be
obtained in two apparently different ways, as explained in Sect. 2.5 and 4.4.
We have performed calculations for various on-site interactions U and
L/R
coupling to the leads Vi with zero gate voltage Vg = 0. In this setup
the superblock size – for fixed left and right block size m – is four times
bigger than in the single lead case due to the two extra central sites in the
superblock setup. Using the databars at DTU and equipped with generous
amounts of patience we managed to use m = 32 in these calculations as
well.1 The occupations calculated are shown in Fig. 5.3 and 5.4. To expose
the effect of on-site interaction we have collected figures with fixed coupling
to the leads and varying U . Additionally in Fig. 5.6 we have plotted the
total occupation of the dot n(µ) = n1 (µ) + n2 (µ).

5.4.1 Convergence
In order to test the convergence of the algorithm we have calculated the
occupation of the dot using 125 and 200 sites in each lead. All these results
differ on the scale of 10−3 , as seen in the example in Fig. 5.5, and hence
convergence is reached using 125 sites in each lead.

5.5 Interpretation
In this section we briefly comment on the results presented in the previous
section. We distinguish two different sets of parameters, the coupling to the
leads ViL,R and the on-site repulsion U . Generally we use weaker couplings
to the leads in these calculations than in the single lead setup to keep the
renormalized levels inside the band.
For very weak coupling to the leads the levels on the dot are essentially
unchanged, only the second level is renormalized. The filling of the first
level occurs when the chemical potential crosses the energy of the level, and
1
Approximately 800 MB of memory is needed for these calculations. The total number
of superblock states is 24 × m2 = 16384.
 69
5.5 Interpretation

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(a) U = 0, Vi = 0.4. (b) U = 0.1, Vi = 0.4.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(c) U = 0.2, Vi = 0.4. (d) U = 0.4, Vi = 0.4.

Figure 5.3: Occupations for ² = [−1.3, −1.1], m = 32 states, and 125 sites in the
L/R
lead. The strong coupling Vi makes the filling of the levels occur at nearly the
same energy, while the repulsion U pushes the filling of the levels upwards in energy.
L/R
U , Vi , ², and µ are given in units of t.

the filling of the second level occurs when the chemical potential reaches the
renormalized energy ²̃2 = ²2 + U , as seen in Fig. 5.4. When the coupling is
turned up there is a crossover to a single localized state on the dot, revealed
by a sharp jump of unit magnitude in the total occupation, located between
²1 and ²2 , as seen in Fig. 5.6(a).

When interactions and couplings both are turned up some features of

the non-interacting limit persist, as seen in Fig. 5.6(b)-5.6(d), but due to
the interactions the position of the jump in occupation of the second level
and the localized state is shifted. All jumps are smeared by the interactions
and for stronger interactions simple arguments are insufficient.
70
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(a) U = 0, Vi = 0.25. (b) U = 0.1, Vi = 0.25.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(c) U = 0.2, Vi = 0.25. (d) U = 0.4, Vi = 0.25.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(e) U = 0, Vi = 0.05. (f) U = 0.1, Vi = 0.05.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(g) U = 0.2, Vi = 0.05. (h) U = 0.4, Vi = 0.05.

Figure 5.4: Occupations for ² = [−1.3, −1.1], m = 32 states, and 125 sites in the
L/R
lead. The repulsion U suppresses the filling of the second level. U , Vi , ², and µ
are given in units of t.
 71
5.6 Analytic result

−3
x 10
3

−1

−2

−3 n1
n
2
−4
−1.5 −1 −0.5

L/R
Figure 5.5: Difference in occupations for ² = [−1.3, −1.1], Vi = 0.25, m = 32
states, and 125 and 200 sites in each lead respectively. Notice the scale on which
the two sets of occupations differ from each other.

5.6 Analytic result

In the non-interacting limit it is possible to calculate the occupations of the

dot levels analytically using Green’s functions theory [17, 18]. The Hamil-
tonian is

H = HDot + HRight Lead + HLef t Lead + HDL, , (5.5a)

b
X X
HDot = (²i − Vg )a†i ai ≡ εa a†i ai , (5.5b)
i=a i
∞
X ∞
X
† †
HRight Lead = −t (cj cj+1 + cj+1 cj ) + ε c†j cj , (5.5c)
j=1 j=1
−∞
X −∞
X
HLef t Lead = −t (c†j cj−1 + c†j−1 cj ) +ε c†j cj , (5.5d)
j=−1 j=−1
b ³
X ´
HDL = ViR (a†i c1 + c†1 ai ) + ViL (a†i c−1 + c†−1 ai ) . (5.5e)
i=a

In order not to clutter up the notation we denote by a and b the dot levels and
by 1, 2, · · · the lead sites, and in retarded Green’s functions the convergence
factor ω → ω + iη is not written explicitly.
We seek the full Green’s function for the dot levels generally defined as

­ † 0 ®
GR 0 0
ij (t − t ) = −iθ(t − t ) {ai (t), aj (t )} . (5.6)
72
Quantum dot with two infinite leads

2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2
Occupation

Occupation
1 1

0.8 0.8

0.6 0.6

0.4 Vi=0.05 0.4 Vi=0.05

Vi=0.25 Vi=0.25
0.2 0.2
V =0.4 V =0.4
i i
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

(a) U = 0. (b) U = 0.1.

2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2
Occupation

Occupation

1 1

0.8 0.8

0.6 0.6

0.4 Vi=0.05 0.4 Vi=0.05

Vi=0.25 Vi=0.25
0.2 0.2
V =0.4 V =0.4
i i
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

(c) U = 0.2. (d) U = 0.4.

Figure 5.6: Total occupations for ² = [−1.3, −1.1], m = 32 states, and 125 sites in
L/R
the lead. Notice the crossover to a single localized state when the coupling Vi
L/R
increases. U , Vi , ², and µ are given in units of t.

In App. E we show how the EOM for GR

aa are obtained,

(ω − εa )GR R R L R
aa (ω) = 1 + Va G1a (ω) + Va G−1a (ω), (5.7a)
(ω − ε)GR1a (ω) = VaR GR R R R
aa (ω) + Vb Gba (ω) − tG2a (ω), (5.7b)
R
(ω − ε)G−1a (ω) = VaL GR L R R
aa (ω) + Vb Gba (ω) − tG−2a (ω), (5.7c)
(ω − ε)GR2a (ω) = −tGR R
1a (ω) − tG3a (ω), (5.7d)
R
(ω − ε)G−2a (ω) = −tGR R
−1a (ω) − tG−3a (ω), (5.7e)
..
.
(ω − ε)GR R R
ja (ω) = −tGj−1a (ω) − tGj+1a (ω), (5.7f)
(ω − ε)GR
−ja (ω) = −tGR
−j+1a (ω) − tGR
−j−1a (ω), (5.7g)
and dropping the arguments the solution to this semi-infinite set of equations
is
£ ¤ R
ω − εa − ΣR R R
aa Gaa = 1 + Σab Gba . (5.8)
 73
5.6 Analytic result

0.5

11
−0.5

G
−1

−1.5
real(G (ω))
11
imag(G (ω))
11
−2
−4 −3 −2 −1 0 1 2 3 4
ω

Figure 5.7: Plot of real and imaginary part of G̃R

11 (ω).

The selfenergy ΣR
ij was found in App. E,

¡ R R ¢
ΣR
ij = Vi Vj + ViL VJL G̃R
11 , (5.9a)
p
ω/2t − sign(ω) (ω/2t)2 − 1
G̃R
11 = , (5.9b)
t

where G̃R
11 is the full Green’s function for the first site in the lead in absence
of the quantum dot.
Similarly we find for GR
ba

£ ¤ R
ω − εb − ΣR R R
bb Gba = Σba Gaa , (5.10)

and hence the full solution for GR

aa ,

1
GR
aa = , (5.11)
ΣR R
ab Σba
ω − ε a − ΣR
aa − ω−εb −ΣRbb

and similarly

1
GR
bb = . (5.12)
ΣR R
ba Σab
ω − εb − ΣR
bb − ω−εa −ΣRaa

It should be mentioned that the single lead model cannot be mapped onto
the two lead model by simply redefining the coupling constants, and hence
the two models are indeed different.
74
Quantum dot with two infinite leads

5.6.1 Occupation
Having found G̃R11 (ω) it is possible to find the spectral function and hence
the occupations of the levels. The relevant selfenergies are
¡ R 2 ¢ R
ΣR L 2
aa (ω) = (Va ) + (Va ) G̃11 (ω), (5.13a)
¡ ¢
ΣR R 2 L 2 R
bb (ω) = (Vb ) + (Vb ) G̃11 (ω), (5.13b)
R
¡ R R L L
¢ R
Σab (ω) = Va Vb + Va Vb G̃11 (ω), (5.13c)
¡ ¢
ΣR R R L L
ba (ω) = Va Vb + Va Vb G̃11 (ω),
R
(5.13d)

so that G̃R
11 (ω) essentially determines the complex structure of the selfener-
gies. In App. E we show that G̃R 11 can be expressed as
 √

 ω/2t 1−(ω/2t)2
 t −i t , |ω| < 2t,
R
G̃11 (ω) = √ (5.14)


 ω/2t−sign(ω) (ω/2t)2 −1 , |ω| ≥ 2t.
t

The occupation of the dot is found in exactly the same way as in the
single lead setup in Sect. 4.7.1, using the Fluctuation Dissipation Theorem
to rewrite the lesser Green’s function
Z µ
dω
ni (µ) = Ai (ω), (5.15a)
−∞ 2π
Ai (ω) ≡ −2ImGR
ii (ω). (5.15b)

Note that for the parameter values we use the renormalized levels remain
within the band, but for other parameter values this might not be the case.
Expresing the occupations analytically involves finding the imaginary part
of the Green’s function, which yields very large expressions that do not
simplify, and we have therefore performed these calculations numerically for
various combinations of couplings to the two leads, shown in Fig. 5.8.

5.6.2 Comparison
Having performed both exact Green’s function and approximate DMRG
calculations we are able to test the performance of DMRG in the non-
interacting limit. In Fig. 5.9 both exact and DMRG results are plotted
in rather consistency.
The agreement between the two is obvious, but there appear to be some
numeric irregularities in the DMRG result, especially near the step in occu-
pation. We expect that the majority of these problems would be eliminated
by increasing the number of states kept at each DMRG step. There are two
differences compared to the single lead setup, explaining the lesser precision:
(i) For fixed number of states kept m the relative truncation in this setup is
 75
5.7 Towards transport

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5 n , V =0.05
1 L
n , V =0.05
0.4 0.4 2 L
n , V =0.05 n , V =0.2
1 L 1 L
0.3 n , V =0.05 0.3 n , V =0.2
2 L 2 L
n , V =0.2 n , V =0.35
1 L 1 L
0.2 n2, VL=0.2 0.2 n2, VL=0.35
n1, VL=0.35 n1, VL=0.5
0.1 0.1
n , V =0.35 n , V =0.5
2 L 2 L
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
µ µ

(a) ViL = ViR . (b) ViL = 2ViR .

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 n , V =0.05 0.5 n , V =0.05

1 1 1 2
n , V =0.05 n , V =0.05
0.4 2 1 0.4 2 2
n , V =0.2 n , V =0.2
1 1 1 2
0.3 n , V =0.2 0.3 n , V =0.2
2 1 2 2
n , V =0.35 n , V =0.35
1 1 1 2
0.2 n2, V1=0.35 0.2 n2, V2=0.35
n1, V1=0.5 n1, V2=0.5
0.1 0.1
n , V =0.5 n , V =0.5
2 1 2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
µ µ

L/R L/R L/R L/R

(c) V1 = 2V2 . (d) 2V1 = V2 .

Figure 5.8: Occupations in the non-interacting case for ² = [−1.3, −1.1], computed
L/R
using Green’s function theory. U , Vi , ² and µ are given in units of t.

bigger than in the single lead case. (ii) With the numbering of basis states
used, the two leads in this setup are equivalent to a single lead with only
next-nearest neighbor interactions. Thus this setup corresponds to a single
lead setup with longer range interactions, which is computationally heavier,
as argued in Sect. 2.6. Note, however, that enlarging only one lead at a
time in the infinite system step might increase the precision, as discussed in
Sect. 5.3.

5.7 Towards transport

To do transport using DMRG turned out to be unrealistic within the time

limits of this project. However, to approach the subject we have calculated
the occupation of the dot as a function of gate voltage when the chemical
potentials of the two leads are different. We use the setup shown in Fig. 5.10.
76

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3
n DMRG
1
0.2 n2 DMRG 0.2
n1 Analytic n1
0.1 0.1
n Analytic n
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(a) Vi = 0.05. (b) Vi = 0.1.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3
n DMRG n DMRG
1 1
0.2 n2 DMRG 0.2 n2 DMRG
n1 Analytic n1 Analytic
0.1 0.1
n Analytic n Analytic
2 2
0 0
−1.5 −1.3 −1.1 −1 −0.75 −0.5 −1.5 −1.3 −1.1 −1 −0.75 −0.5
µ µ

L/R L/R
(c) Vi = 0.25. (d) Vi = 0.3.

0.9

0.8

0.7

0.6
Occupation

0.5

0.4

0.3
n DMRG
1
0.2 n2 DMRG
n1 Analytic
0.1
n Analytic
2
0
−1.5 −1.3 −1.1 −1 −0.75 −0.5
µ

L/R
(e) Vi = 0.4.

Figure 5.9: Comparison between DMRG and analytic results for ² = [−1.3, −1.1],
m = 32 states, and 125 sites in the lead. The agreement is fine but there are more
L/R
numeric irregularities than in the single lead setup. U , Vi , ², and µ are given in
units of t.
 77
5.7 Towards transport

2t

µL
²b
µR

Vg ²a

−2t

Figure 5.10: Biased quantum dot. Notice that we use indices a and b for the dot
levels and 1, 2, · · · for the lead sites and the different chemical potentials in the left
and right leads.

5.7.1 Hamiltonian
The Hamiltonians are essentially unchanged, but the leads are assumed in
equilibrium with different chemical potentials, µL and µR respectively, giving
the total Hamiltonian

H= (5.16)
∞
X ∞
X
HDot + HRight Lead + HLef t Lead + HDL − µR c†j cj − µL c†−j c−j ,
j=1 j=1

where HDot , HRight Lead , HLef t Lead , and HDL are given in Eq. (5.1) and
(5.2a)-(5.2c).

5.7.2 Results
The DMRG implementation of the biased dot is essentially identical to the
unbiased dot explained in Sect. 5.2 and 5.3 only here we distinguish different
chemical potentials. We calculate the occupation of the dot versus gate
voltage Vg for various couplings and on-site interactions. The parameter
L/R
values used are µL = −µR = 0.1 and three different couplings Vi between
L/R
0.5 and 0.05. For each set of parameters U , Vi , and Vg we do a full
DMRG calculation of the occupation of the dot. In Fig. 5.11 and 5.12 we
show plots of the occupation of the dot versus gate voltage calculated using
DMRG. It should be noted that this calculation does not incorporate the
finite bandwidth of this model, and hence allows filling of the dot even
when the renormalized levels are outside the band. This contradicts energy
conservation since there is no dissipation in this model.
78
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(a) U = 0, Vi = 0.5. (b) U = 0.1, Vi = 0.5.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(c) U = 0.2, Vi = 0.5. (d) U = 0.4, Vi = 0.5.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(e) U = 0, Vi = 0.25. (f) U = 0.1, Vi = 0.25.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(g) U = 0.2, Vi = 0.25. (h) U = 0.4, Vi = 0.25.

Figure 5.11: Occupations for ² = [−1.3, −1.1], µL = −µR = 0.1, m = 32 states,

L/R
and 125 sites in the lead. U , Vi , ², and Vg are given in units of t.
 79
5.7 Towards transport

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

(a) U = 0. (b) U = 0.1.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
n1 n1
0.1 0.1
n n
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

(c) U = 0.2. (d) U = 0.4.

Figure 5.12: Occupations for ² = [−1.3, −1.1], µL = −µR = 0.1, m = 32 states,

L/R
and 125 sites in the lead. U , Vi , ², and Vg are given in units of t.

5.7.3 Analytic result

In the non-interacting limit we can calculate the occupation of the biased

dot analytically, using again Green’s function theory. The basic analysis is
very similar to the analysis already presented, and again the convergence
factor ω → ω + iη is not written explicitly. The EOM found in App. E can
be written more intuitively in terms of the free Green’s functions, (giR )−1 ≡
ω − (²i − Vg ), and the selfenergies, ΣR L L R R R
ij ≡ (Vi Vj + Vi Vj )G̃11 ,

£ R −1 ¤ R
(ga ) − ΣR R R
aa ) Gaa = 1 + Σab Gba ⇒
£ ¤ R
1 − gaR ΣR R R R R
aa Gaa = ga + ga Σab Gba ⇒
GR R R R R R R R
aa = ga + ga Σab Gba + ga Σaa Gaa . (5.17)
80
Quantum dot with two infinite leads

Similarly we find

GR R R R R R R R
bb = gb + gb Σba Gab + gb Σbb Gbb , (5.18a)
GR
ab = gaR ΣR R R R R
aa Gab + ga Σab Gbb , (5.18b)
GR
ba = gbR ΣR R R R R
bb Gba + gb Σba Gaa . (5.18c)

These equations are collected in matrix form,

GR = gR + gR ΣR GR , (5.19)

where
µ ¶
GR R
aa Gab
GR = , (5.20a)
GR R
ba Gbb
µ R ¶
ga 0
gR = , (5.20b)
0 gbR
µ R ¶
Σaa ΣR
ΣR = ab . (5.20c)
ΣR R
ba Σbb

Having verified that we can indeed write the Dyson equation for GR
Eq. (5.19) we can use a result from non-equilibrium Green’s function theory,2

G< = GR Σ< GA , (5.21)

where
¡ R ¢∗
GA
ij (ω, Vg ) = Gij (ω, Vg ) , (5.22a)
¡ R ¢¡ R R R ¢
Σ<
ij
L L L
= −2iIm G̃11 Vi Vj nF (ω) + Vi Vj nF (ω) . (5.22b)

L/R
The notation nF (ω) indicates that the chemical potential of the left and
right leads are different,

L/R 1
nF (ω) = −β(µL/R −ω)
, (5.23)
e +1

and G̃R
11 is again given by
p
ω/2t − θ(|ω|/2t − 1)sign(ω) (ω/2t)2 − 1
ReGR
11 = , (5.24a)
p t
−θ(1 − |ω|/2t) 1 − (ω/2t)2
ImGR
11 = . (5.24b)
t
2
[18], p. 165-166. The derivation of this statement relies on the Keldysh formalism,
defining generalized Green’s functions on the Keldysh contour and extracting various real
time Green’s functions from this.
 81
5.7 Towards transport

By matrix multiplication we find for the two levels on the dot,

G< R < A R < A R < A R < A

aa = Gaa Σaa Gaa + Gab Σba Gaa + Gaa Σab Gba + Gab Σbb Gba , (5.25a)
G<
bb = GR < A
bb Σbb Gbb + GR < A
ba Σab Gbb + GR < A
bb Σba Gab + GR < A
ba Σaa Gab , (5.25b)

where the retarded Green’s functions were found in Sect. 5.6. Hence we may
calculate the occupation of the dot in the standard way,
­ † ®
ni (Vg ) ≡ ai ai
= −iG< (t = 0, Vg )
Z ∞ ii
dω <
= Gii (ω, Vg ). (5.26)
−∞ 2πi

Plots of the occupation calculated using Eq. (5.26) as well as DMRG results
are shown in Fig. 5.13. As discussed in App. F the spectral function, and
hence G< , may have δ-like peaks outside the band. Such peaks are very
difficult to integrate numerically, and hence numerical imprecisions might
be present in the numerical calculations of ni for some values of Vg .

5.7.4 Comparison
Having calculated the occupation using both DMRG and analytic methods
we are able to compare the two. In Fig. 5.13 we show both sets of results.
For small gate voltages these results agree quite well. There is however
one feature of the analytic result that the DMRG results do not exhibit,
namely the decrease of occupation when the gate voltage becomes suffi-
ciently large to push the renormalized levels outside the band. The hopping
matrix elements used in the DMRG calculation allows filling of the dot even
when this is prohibited by energy conservation. Our DMRG implementation
does not incorporate the finite bandwidth, explaining the deviation of the
DMRG and analytic results. Probably a modification of the hopping matrix
elements used in DMRG can remedy this.
Also in future calculations it would be interesting to use a finer grid of
gate voltages near the unrenormalized level energies, and see if the DMRG
results reproduce the detailed structure of the step-up in occupation.

5.7.5 Attempted current calculation

¡­ ®
Naively one would expect that we might directly calculate ViL a†i c1 −
­ † ®¢
c1 ai using DMRG, which apart from some factors is the static current
through the dot,

ie X L ¡­ † ® ­ † ®¢
JL = Vi ai c1 − c1 ai . (5.27)
~
i
82
Quantum dot with two infinite leads

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3
n1, DMRG n1, DMRG
0.2 0.2
n2, DMRG n2, DMRG
0.1 n , Analytic 0.1 n , Analytic
1 1
n , Analytic n , Analytic
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(a) Vi = 0.05. (b) Vi = 0.25.
1

0.9

0.8

0.7

0.6
Occupation

0.5

0.4

0.3
n1, DMRG
0.2
n2, DMRG
0.1 n , Analytic
1
n , Analytic
2
0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v
g

L/R
(c) Vi = 0.5.

Figure 5.13: Comparison between DMRG and analytic occupations for ² =

[−1.3, −1.1], µL = −µR = 0.1, m = 32 states, and 125 sites in the lead. The
large gate occupations are not reproduced by DMRG since we have not incorpo-
L/R
rated the finite band width in our implementation. U , Vi , ², and Vg are given
in units of t.

Very recently we experimented with this approach, but found it impossible

to calculate the current, at least with this extremely simple approach. The
DMRG results turned out to cancel identically, and the single contributions
turned out to be real. In our DMRG approach there is no time dependence,
which could explain this failure. Apparently the time dependence present in
the Green’s function result has been collapsed too soon. We consider current
calculations as future work, and expect that more sophisticated methods are
needed, [5, 6, 7].

5.8 Conclusions
Expanding on the implementation of Chap. 4 we have successfully imple-
mented DMRG on a quantum dot coupled to two one-dimensional leads, in
 83
5.8 Conclusions

the semi-infinite lead limit. We have performed virtually the same calcula-
tions as for the single lead setup, calculating the occupation of the dot as a
function of chemical potential, for various couplings and on-site interactions
and briefly commented on the results found. We compared DMRG to an-
alytic data in the non-interacting limit, and found consistency, but to less
extent than in the single lead case. We discussed possible explanations for
this.
To take the first steps towards transport and to test the capability of
DMRG we forced the dot out of equilibrium by using different chemical
potentials in the left and right leads, calculating the occupation versus gate
voltage on the dot. In the non-interacting limit we compared DMRG and
analytic non-equilibrium Green’s function results. For a limited range of
gate voltages the two results are consistent, but the DMRG calculation does
not reproduce the large gate behavior, when the renormalized dot levels are
pushed outside the band. We explained the origin of this discrepancy and
mentioned a possible solution.
84
Chapter 6

Single particle quantum

mechanics

One very simple application of DMRG is on single particle quantum mechan-

ics. Dealing with single particles does not really unlock the full potential of
DMRG but the implementation is much simpler, making it a good starting
point. This is due to the fact that the truncation usually performed via the
reduced density matrix reduces to a simple projection of the wavefunction
in this case.
In this section we follow Martı́n-Delgado et al. and calculate ground
state energy and gap for a single particle on a tight binding chain in two
different potentials using a simple DMRG implementation [20]. The article
is pedagogical and well written and we will be brief in the explanation of
the algorithm. For further details we refer to the article.1

6.1 Hamiltonian and matrices

The model considered is a single particle in a box experiencing a potential
and represented classically by

H = p2 + V (x), (6.1)

where V (x) is any potential. First we discretize the Hamiltonian by restrict-

ing x, xn = h(n − N 2+1 ), where n = 1, 2, · · · , N . The parameter h gives the
size of the box in which the particle is confined, ∆x = xN − x1 = h(N − 1).
With these definitions the Hamiltonian becomes
 2
 h2 + V (xn ), n = m,
H = −1, n = m ± 1, (6.2)
 h2
0, otherwise.
1
For the historically interested, single particle quantum mechanics using DMRG is
interesting since it was the failure of Wilson’s numerical renormalization group on this
type of problems that lead White to formulate DMRG.

85
86
Single particle quantum mechanics

Using an ordinary superblock setup we find the generic superblock Hamil-

tonian,
 
HL −vL 0 0
 −v † hCL −1/h2 0 
HSB =  L
 0 −1/h2 h

† , (6.3)
CR −vR
0 0 −vR HR

where HL and vL represent the (possibly truncated) Hamiltonian of the left

block and interaction of the left block with the first central site, and similar
for the right block.
Due to the single particle nature of the problem each state is described
by a single degree of freedom – and hence HSB has dimensions (2NE + 2) ×
(2NE + 2), where NE is the total number of target states for each block.

6.2 Algorithm
In this section we show that the eigenstates of the reduced density matrix
(RDM) in the single particle problem reduces to a projection of the wave-
function, and review the DMRG algorithms.

6.2.1 RDM for single particle problems

In single particle problems the complete basis consists of single particle
states, located at the different sites in the chain. To express the state of
the superblock in terms of product states we need an enlarged basis of
the superblock, containing both zero and two particle states. We use the
notation of Noack and White and denote the states of the enlarged basis for
the left block as
¯ ®
¯0 , (6.4a)
L
¯ ® ¯ ®
¯1
L
= c†1 ¯0 L , (6.4b)
..
.
¯® ¯ ®
¯l
L
= c†l ¯0 L , (6.4c)

with similar definitions for the right block [2]. Index l denotes the rightmost
site of the left block, so that the first site in the environment block is l + 1.
Denoting additionally the ground state wavefunction
¯ ®
¯ψ = (ψ1 , ψ2 , · · · , ψl , ψl+1 , · · · , ψL ), (6.5)
 87
6.2 Algorithm

where L is the total chain length we find the ground state matrix in the
usual way,
 
0 ψl+1 · · · ψL
 ψ1 0 0 
 
ψ =  . . ..  , (6.6)
 .. . . . 
ψl 0 ··· 0

where it is clearly seen that only one particle states enter. By matrix mul-
tiplication the reduced density matrix becomes
 PL 2

j=l+1 ψj 0 0 ··· 0
 0 2
ψ1 ψ1 ψ2 · · · ψ1 ψl 
 
 ψ2 ψ1 ψ22 · · · ψ2 ψl 
ρ = ψψ † =  0 . (6.7)
 .. .. .. .. .. 
 . . . . . 
0 ψl ψ1 ψl ψ2 · · · ψl2

The spectrum of this matrix is very simple, containing only two non-zero
eigenvalues, with the corresponding normalized eigenvectors

v1 = (1, 0, · · · , 0) , (6.8a)
1
v2 = qP (0, ψ1 , ψ2 , · · · , ψl ) . (6.8b)
l 2
i=1 ψi

The first eigenvector does not need to be considered since the only nonzero
entry corresponds to a zero particle state. Hence the best state to retain for
the left block is (0, ψ1 , ψ2 , · · · , ψl ), given the ground state vector Eq. (6.5),
or removing the zero particle state,

v = (ψ1 , ψ2 , · · · , ψl ) (6.9)

As in Sect. 2.1 this derivation generalizes when targeting NE states,

i.e. targeting additionally NE − 1 exited states, where you get NE optimal
states, conveniently collected in a matrix
 
ψ1,1 · · · ψ1,NE
 ψ2,1 · · · ψ2,N 
 E 
O =  . ..  . (6.10)
 .. . 
ψl,1 · · · ψl,NE

6.2.2 Particle in a box

Targeting NE states in the left and right blocks the superblock Hamiltonian
can be written as in Eq. (6.3), where HL and HR are NE × NE matrices and
88
Single particle quantum mechanics

vL and vR are NE component column vectors. Each eigenstate of the su-

perblock is a 2NE +2 component vector. Denoting the NE lowest eigenstates
of the superblock Hamiltonian by

(ψ̄L,i , ψCL,i , ψCR,i , ψ̄R,i ), i = 1, 2, · · · , NE , (6.11)

where ψ̄L/R,i are vectors, we find the projection of each target state onto
the enlarged left block by the set of vectors (ψ̄L,i , ψCL,i ), as shown above.
These states are orthonormalized explicitly.
Hence including the central state CL into the left block we have the
truncation onto a NE component DMRG basis given by the matrix
µ ¶
ψ̄L,1 · · · ψ̄L,NE
O = , (6.12)
ψCL,1 · · · ψCl,NE

or explicitly the new left block Hamiltonian

µ ¶
0 † HL −vL
HL = O O, (6.13)
−vL† hCL

and the new interaction with the central site,

0
vL,i = ψCL,i , i = 1, 2, · · · , NE . (6.14)

This interaction is not found using DMRG arguments but by noting that
ψCL,i is the probability of finding the particle on the old central site for the
state i.
We complete the generic DMRG step by renaming
0
vL,i → vL,i , (6.15a)
HL0 → HL . (6.15b)

Similar relations exist for the right block.

6.2.3 Infinite system sweep

We scale all arguments by h2 to get consistency with [20]. We initialize the
infinite system procedure by defining
µ ¶
2 + h2 V (x1 ) −1
HL = , (6.16a)
−1 2 + h2 V (x2 )
µ ¶
0
vL = , (6.16b)
1

and

hCL/CR = 2 + h2 V (x3/4 ), (6.17)

 89
6.3 Results

and similar expressions for the right block. Hence the initial superblock
Hamiltonian is
 
HL −vL 0 0
 −v † 2 + h2 V (x3 ) −1 0 
HSB =   0
L 
† , (6.18)
−1 2 + h2 V (x4 ) −vR
0 0 −vR HR
where the number of sites is N = 6.
At each step the left and right blocks are enlarged by one state and
truncated as explained in the previous subsection. This is repeated until
the total number of sites in the system reaches the desired value N . At each
step all matrices for the left and right block should be stored. Note that the
size of the box in which the particle is confined ∆x is kept constant, only
the number of allowed positions N is increased.

6.2.4 Finite system sweeps

In this part the number of sites in the system is kept constant, but the sizes
of the left and right blocks are varied. The basic left to right step is enlarging
the left block by one site and truncating as explained above. Then form a
superblock using this enlarged left block, two central sites and a right block
with one state less. This right block has been found and stored during the
infinite system sweep.
This procedure is repeated until the right block consists of NE states,
i.e. until the description of the right block is exact. Then the roles of right
and left are reversed and we iterate until the left block contains only NE
states. We repeat this zipping back and forth until convergence is reached,
typically around 10 times.

6.3 Results
Using the algorithm described above and targeting only one exited state
(NE = 2) we are able to calculate the ground state energy and energy
gap for various potentials. We have considered two different potentials, the
harmonic potential and the double well potential,
Vharmonic (x) = x2 , (6.19a)
2 4
VDW (x) = −x + Cx . (6.19b)
A few results calculated using this DMRG scheme are shown in Tab. 6.1
for the harmonic potential and Tab. 6.3 for the double well potential. The
corresponding data presented by Martı́n-Delgado et al. are shown in Tab. 6.2
and 6.4. It should be mentioned that these results require only modest
computer power compared to the results presented in chapters 3, 4, and 5,
as the superblock matrices in this setup are only 6 × 6.
90

(a) V (x) = x2 . (b) V (x) = −x2 + Cx4 .

Figure 6.1: Harmonic and double well potentials in arbitrary units.

N h 1 − E0 3 − E1
1000 0.01 6.24989 × 10−6 3.1243338 × 10−6
5000 0.002 2.498 × 10−7 1.2427 × 10−6

Table 6.1: DMRG results for the harmonic potential.

N h 1 − E0 3 − E1
1000 0.01 6.24989(322) × 10−6 3.1243338(564) × 10−6
5000 0.002 2.498(36) × 10−7 1.2427(28) × 10−6

Table 6.2: DMRG results from [20] for the harmonic potential.

C N h E0 ∆
1 1000 0.01 0.65764425361 2.176825710
0.6 20000 0.0005 0.39195261 1.6332847
0.06 60000 0.0005 -2.82363949 0.0072997661

Table 6.3: DMRG results for the double well potential.

C N h E0 ∆
1 1000 0.01 0.65764425361(29) 2.176825710(298)
0.6 20000 0.0005 0.39195261(873) 1.6332847(928)
0.06 60000 0.0005 -2.82363949(203) 0.0072997661(673)

Table 6.4: DMRG results from [20] for the double well potential.
 91
6.4 Conclusions

6.3.1 Comparison
Our DMRG result agrees completely with the data presented by Martı́n-
Delgado et al. and one could proceed and explore e.g. the continuum limit
of single particle quantum mechanics. It is not the purpose of the present
thesis to do that. Further details of the physics of these models can be found
in [20].

6.4 Conclusions
In this section we used a simpler formulation of DMRG to calculate prop-
erties of a particle in a box, experiencing two different potentials. Results
obtained with our implementation were compared to those presented by
Martı́n-Delgado et al., showing complete consistency.
We showed explicitly that in the single particle case the projection of
DMRG, usually done via the reduced density matrix, reduces to a projection
of the target state wavefunction, making the DMRG implementation much
simpler. All the calculations presented here could have been performed
using the full machinery of DMRG, defining the reduced density matrix,
diagonalizing it, computing the retained eigenstates etc. The simplicity
of the single particle DMRG makes it a good starting point for DMRG
calculations.
92
Chapter 7

Future work

Implementing DMRG turned out to be much more difficult than understand-

ing the theoretical framework. We spent long time learning the basic con-
cepts of DMRG, setting up matrix representations of operators etc. Many
extensions of basic DMRG have been proposed, either speeding up calcula-
tions or expanding the capability of DMRG. We have considered only a few
such additional schemes.
In this chapter we discuss improvements of our implementation of DMRG
as well as several extensions that enable calculations beyond basic DMRG.
Some extensions are highly non-trivial and require thorough studies of the
literature, while others are more straight forward.

7.1 Improvements
7.1.1 Quantum numbers
So far our implementation has not been focused on the use of quantum
numbers to label the basis states, since this implies programmatic difficulties
in MatLab. However, in all models we have considered there are a number
of good quantum numbers that can be used to label the states, such as
particle number and spin.
As explained in Sect. 2.4.1 the use of quantum numbers potentially
speeds up the DMRG calculation substantially, due to the decomposition
of the full Hilbert space into subspaces. Since storing small dense matrices
is more efficient than storing large sparse matrices, also the memory required
is reduced, making the use of quantum numbers inevitable if larger systems
are to be considered, or more states are to be kept.
Efficient programming in MatLab implies using the internal matrix
structure, organizing data in matrices and vectors. This has some obvi-
ous advantages in DMRG but there are also drawbacks; reordering of basis
states in the superblock may sometimes be desirable. In MatLab such
reorderings of superblock states are easily performed, according to various

93
94
Future work

criterions. A rearrangement of the superblock basis corresponds to a rear-

rangement of the basis of the individual blocks, and it is non-trivial to link
these basis rearrangements to each other. Likewise it is difficult to organize
the basis states of the individual blocks such that the superblock automati-
cally is in quantum number block structure. Hence it is not trivial to identify
and calculate in a given subspace of the superblock when using MatLab.
We expect that object oriented programming would facilitate the solu-
tion of these issues. A block object could contain relevant subspaces and a
loop over quantum numbers could then construct a superblock Hamiltonian
in a basis ordered according to quantum numbers. Of course this is also
possible in MatLab but it is less obvious how to construct a fast routine
incorporating these features.

7.1.2 Wave function transformations

The most time consuming part of the DMRG algorithm is computing the
ground state of the superblock Hamiltonian. Usually an iterative scheme like
the Lanczos or Davidson algorithm is used. Most implementations of such
diagonalization schemes use an arbitrary starting vector for the iterative
procedure, making many iterative steps necessary to converge to the ground
state. If a good guess for the ground state can be provided a significantly
smaller number of iterations are needed, speeding up the calculation.
In the DMRG scheme a good guess to the current ground state is the
ground state vector from the previous iteration, since the two superblocks
have almost identical configurations. To use this initial guess the basis of
the ground state vector from the previous DMRG iteration must be shifted
into the basis of the current iteration, as discussed in Sect. 2.4.2.

7.1.3 Programming language

We chose to do all programming in MatLab due to prior knowledge of
that program. However it appears that many mathematical algorithms and
packages exist for languages like Python and C++, facilitating implemen-
tations in these languages as well. Since most algorithms within these lan-
guages are based on fortran or C codes better performance could probably
be achieved in either of these languages. Due to the limited time available
we chose to continue using MatLab, but long term DMRG work is likely
to benefit from using Python or C++. The block structure of DMRG
appeals to object oriented programming, where a nice organization can be
made, and we expect that organizing the code using quantum numbers is
easier using objects, as mentioned in Sect. 7.1.1.
Storing matrices in MatLab is done in non-sparse multidimensional
arrays that are initialized with a given fixed size. This is an inefficient and
not very flexible way of storing data, since such storage arrays can only
 95
7.1 Improvements

hold matrices of a given size. In object oriented languages we expect that

both efficient and flexible storage could be implemented, reducing memory
requirements and increasing the flexibility of the program.

7.1.4 Diagonalization routine

As mentioned in Sect. 2.2.3 Noack and White [2] device a ground state
computation scheme in which the construction of the full superblock Hamil-
tonian matrix can be avoided. Incorporating this would reduce memory
requirements and speed up the ground state computation itself. Since the
most time consuming part of the algorithm is the computation of the ground
state of the superblock Hamiltonian a total speed up is expected.
It is our belief that this improvement is highly non-trivial and requires
thorough studies of the Lanczos or Davidson methods, as well as construc-
tion of specialized diagonalization routines.

7.1.5 Further benchmarks

So far we have tested the performance of DMRG in the non-interacting limit

for the various models. For the Hubbard model we have only benchmarked
the ground state energy and the total spin of the ground state in the non-
interacting limit. To test further the capability and precision of our DMRG
implementations it would be interesting to compare DMRG calculations to
analytics in the interacting cases. No complete solutions exist for arbitrary
U in the various models but in certain limits perturbative results or simple
approximations may apply, making it possible to explore the precision of
DMRG in these limits.
A ‘selfconsistent’ test of DMRG, by comparing results calculated using
different number of states m, would also be interesting. This requires more
efficient implementations of DMRG, since we are not able to increase m very
much further than the values used in this work. In the Hubbard model,
where also exited states of the final superblock are considered it would be
interesting to include these as target states in the DMRG calculation, and
see if the description of the exited states of the final superblock is improved.
Probably more than m = 24 states are needed since the general precision
of DMRG decreases when more states are targeted. We expect future more
efficient implementations of DMRG on the models considered in this work
would enable us to perform such selfconsistent tests.
96
Future work

7.2 Extending the capability of DMRG

7.2.1 Transport
One interesting application would be to calculate transport properties of one-
dimensional systems using DMRG. In this work we have only approached
transport slightly in Sect. 5.7, where we used different chemical potentials
in the left and right lead. We attempted to calculate the current using a
very simple approach, which turned out to be insufficient.
Recently two articles on transport using DMRG appeared on the preprint
archive, by White and Feiguin [6] and by Schmitteckert [5]. Both articles
expand DMRG to include time evolution, and are hence able to calculate
properties of strongly correlated system out of equilibrium. Schmitteckert
argues that his method uses no further approximations than the DMRG
truncation, making it very interesting. Both articles could provide a DMRG
method for calculating transport, and would be interesting to study in fur-
ther detail.

7.2.2 Phonons
One of the hot subjects these days is the coupling of electrons to phonons
in various systems [21, 22, 23]. Many of these systems would be interesting
to study using non-perturbative methods, but also very challenging to im-
plement in DMRG. This is due to the fact that the Hilbert space of bosons
is large (in principle infinite), and that DMRG performs poorly if a large
Hilbert space is added at every step. Hence a mapping from the single site
large bosonic Hilbert space onto multiple pseudo sites with small Hilbert
spaces is needed in order for the DMRG to perform effectively, making
bosonic Hilbert spaces more challenging to treat in DMRG. More informa-
tion on treating bosonic Hilbert spaces in DMRG can be found in the work
of Jeckelmann and White on the Holstein model [24].

7.2.3 Dynamic properties

Several schemes to calculate dynamical properties using DMRG have been
proposed. We have used a very basic method in our attempt to calculate
a spectral function using DMRG. The goal was to see if we could get some
dynamical results using DMRG, rather than to get accurate results.
There are several candidates in the literature for efficient and precise
calculations of dynamics using DMRG, among those Jeckelmann’s Dynami-
cal DMRG [7], Kühner and White’s DMRG scheme in [25], Schmitteckert’s
non-equilibrium transport using DMRG [5] and White and Feiguin’s real
time evolution using DMRG in [6]. Any of those deserves attention on their
own, and probably a thorough study of these methods would make accurate
calculations of dynamics possible within the framework of DMRG.
Chapter 8

Summary and outlook

In the past decade DMRG has been established as a very powerful numerical
tool, and today it is used by many groups around the world. In this section
we summarize our work, give a short presentation of the current status of
DMRG and discuss the perspectives of DMRG.

8.1 Summary
With this work we have taken the initial steps towards establishing DMRG
as a tool within the Theoretical Nanotechnology group at MIC. We have
successfully implemented ordinary DMRG on the one-dimensional Hubbard
chain with zero and finite magnetic field, and on a two level quantum dot
coupled to one and two semi-infinite leads.
For the Hubbard chain we used our implementation to calculate the
energy, particle number and spin of the lowest eigenstates varying both
length and on-site repulsion. We included a small magnetic field in the
DMRG calculation, and studied the effect on the low energy properties of
the model. In the non-interacting limit and with vanishing magnetic field we
compared the DMRG ground state energies to analytical results and found
very good agreement.
We calculated the occupation of the quantum dot as a function of chem-
ical potential varying the coulomb interaction on the dot and couplings to
the leads. With two leads we additionally made calculations on a biased
dot, calculating the occupation versus gate voltage for different couplings
and coulomb interactions. In the non-interacting limit analytic results were
computed in both the single and double lead case, and we compared these
to DMRG results and found good agreement when sufficiently many states
were retained in the DMRG truncation. For the biased dot we found good
agreement for small gate voltages, where the renormalized dot levels re-
main inside the band. For larger gate voltages analytic and DMRG results
deviate qualitatively from each other since our DMRG algorithm does not
incorporate the finite bandwidth in the model. Finally we attempted a

97
98
Summary and outlook

very simple calculation of the static current through the dot. However the
method turned out to be insufficient for calculating the current.
A simpler version of DMRG on single particle quantum mechanics was
reviewed, and we showed explicitly how the DMRG truncation reduces to a
projection of the wavefunction. We implemented this on a particle in a box
experiencing a potential, and calculated the ground state energy and energy
gap. This implementation is simpler than the full DMRG implementation,
making it a suitable starting point for DMRG calculations.
We reviewed the limitations of our implementations and considered pos-
sible solutions to these issues. Several improvements, rendering the code
faster or more reliable were discussed, and a number of schemes extending
the capability of DMRG was mentioned. Of particular interest to this work
is extensions to calculate dynamical properties and transport.

8.2 Outlook
When studying DMRG a one year master project is very short time and this
project leaves many open ends. Having no prior knowledge of DMRG we had
to focus our attention on ordinary DMRG calculations before considering
various extensions. We managed to get ordinary DMRG under control as
well as consider a few extensions.
In the literature however the usage of DMRG has been extended to very
different systems, ranging from ab initio quantum chemistry calculations
over applications in high-energy physics to transport in nanostructures. The
wide range of problems to which DMRG has already been applied indicates
the versatility of the general idea; while White formulated DMRG for quan-
tum lattice systems the general concept of optimizing some functional using
DMRG ideas appears to apply in very many other cases. Obviously DMRG
and its extensions constitute a very powerful and relevant numerical tool,
undergoing continuous development.
Nanotechnology and properties of nanoscale physical systems is receiv-
ing much attention these years, theoretically as well as experimentally. In
general nanoscale systems are strongly correlated and hence difficult to treat
using standard methods. DMRG could turn out to be particularly interest-
ing when treating these systems theoretically and might be predictive once
parameters have been estimated from experiments.
When doing transport the usual setup is ideal non-interacting leads con-
necting the system of interest to large particle reservoirs, hence keeping the
leads in equilibrium. However this distinction looses the physical rationale
for some setups, particularly when the leads are one-dimensional such as
carbon nanotubes and atomic wires, where the underlying Fermi liquid the-
ory breaks down. Traditional methods are thus not able to properly handle
such situations. Contrary DMRG is able to handle interactions also in the
 99
8.2 Outlook

leads, eliminating the distinction between non-interacting leads and an in-

teraction region, and may therefore be able to give results for such system.
It should be noted that setups using carbon nanotubes as leads are being
used experimentally at MIC.

Recent developments of DMRG have been in the direction of transport

and the calculation of dynamical properties. A number of methods have
been proposed, particularly Schmitteckert calculates the transport through
a quantum dot, formulated in terms of wavepackets evolving in time in a
properly truncated DMRG Hilbert space [5].
Much work still remains before we are able to perform efficient transport
calculations. We hope to continue this work in the direction of transport
within the Ph.D. programme at DTU. The ultimate goal is to formulate
and implement a transport theory using DMRG, and while ordinary DMRG
is now mainly under control it is clear that this is a very challenging task,
theoretically as well as numerically. Despite the many proposed schemes in
the literature transport using DMRG is still an unsettled and active research
area.
100
Appendix A

Observables

In this appendix we show explicitly the equivalence between the two methods
of calculating observables within the DMRG. We denote by A any system
operator.

Method 1
The first method uses
hAi = Tr[ρA]
Xm X n
†
= ψji Aii0 ψi0 j , (A.1)
i,i0 =1 j=1

where A is described in the basis of the system block, m denotes the number
of system states and n denotes the number of environment states.

Method 2
The second method is
mn
X
hAi = ψk† Ãkl ψl , (A.2)
k,l=1

where à = A ⊗ δn , i.e. the matrix representation of A is in the basis of the

superblock, and ψ is the column superblock ground state vector. m denotes
the number of system states and n denotes the number of environment
states.

Equivalence
To show that the two apparently different methods are in fact identical we
consider Eq. (A.1),
m X
X n
†
hAi = ψji Aii0 ψi0 j , (A.3)
i,i0 =1 j=1

101
102
Observables

where m denotes the number of system states and n denotes the number of
environment states.
Using the construction of matrix ψ,
 
ψ1 ψ2 · · · ψn
 ψn+1 ψn+2 · · · ψ2n 
 
ψ =  .. ..  , (A.4)
 . . 
ψ(m−1)n+1 ψ(m−1)n+2 · · · ψmn
we rewrite this in terms of the vector ψ,
ψi0 j = ψ(i0 −1)n+j , (A.5a)
† ∗
ψji = ψj+(i−1)n , (A.5b)
where m denotes the total number of states in the system block and n
denotes the total number of states in the environment block.
Writing out in detail Eq. (A.2) gives
hAi = ψ † (A ⊗ δn )ψ
 
A11 δn A12 δn · · · A1m δn  
ψ1
¡ ¢
 A21 δn A22 δn · · · A2m δn

  .. 
= ψ1∗ · · · ψmn
∗
 .. .. ..   .  , (A.6)
 . . . 
ψmn
Am1 δn Am2 δn · · · Amm δn
where δn is an n × n unit matrix in the Hilbert space of the environment.
Multiplying ψ † and A ⊗ δn gives an m · n component row vector. A careful
index analysis reveals that the vector can be written
m
X
£ † ¤ ∗
ψ (A ⊗ δn ) j+(i0 −1)n = ψj+(i−1)n Aii0 , (A.7)
i=1

with j = 1, 2, · · · , n and i0 = 1, 2, · · · , m. Hence the final multiplication in

Eq. (A.6),
X£ ¤
hAi = ψ † (A ⊗ δn ) k ψk , (A.8)
k

where k = j + (i0 − 1)n, yields

m X
X n
∗
hAi = ψj+(i−1)n Aii0 ψ(i0 −1)n+j , (A.9)
i,i0 =1 j=1

or using Eq. (A.5),

X †
hAi = ψji Aii0 ψi0 j , (A.10)
i,i0 ,j

and hence the two methods differ only by arrangement of the data.
 103

Applications
Which of the two methods is most conveniently used depends on the repre-
sentation of the data. The eigenstate routines used in this work returns
eigenstates in column vector form and hence in some cases the second
method is most favorable. Also when calculating properties of the su-
perblock only the second method is applicable.
The first method is most favorable when calculating properties of the
system throughout the DMRG calculation since the state matrix ψ and
operator matrix A is constructed at each step anyway, and hence the expec-
tation value can be found very fast by matrix multiplication.
104
Appendix B

Properties of the Lanczos

basis

In this appendix we demonstrate explicitly the orthogonality of the Lanczos

basis.

B.1 Definition
The Lanczos basis is defined by
¯ ® ¯ ® ¯ ® ¯ ®
¯fn+1 = H ¯fn − an ¯fn − b2n ¯fn−1 , (B.1)

where
­ ¯ ¯ ®
fn ¯H ¯fn
an = ­ ¯¯ ® , (B.2a)
fn fn
­ ¯ ®
fn ¯fn
b2n = ­ ¯ ®, (B.2b)
fn−1 ¯fn−1
¯ ® ¯ ®
¯f0 = A¯ψ0 , (B.2c)
b20 = 0, (B.2d)

and where A is the operator for which we want to construct the basis.

B.2 Orthogonality of basis

First we demonstrate orthogonality for the first Lanczos basis states.
¯ ®
¯f1 :
¯ ® ¯ ® ¯ ®
¯f1 = H ¯f0 − a0 ¯f0 , (B.3a)
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ®
f0 ¯f1 = f0 ¯H ¯f0 − a0 f0 ¯f0 = 0, (B.3b)
­ ¯ ¯ ® ­ ¯ ®
using a0 = f0 ¯H ¯f0 / f0 ¯f0 .

105
106
Properties of the Lanczos basis

¯ ®
¯f2 :
¯ ® ¯ ® ¯ ® ¯ ®
¯f2 = H ¯f1 − a1 ¯f1 − b21 ¯f0 , (B.4a)
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
f1 ¯f2 = f1 ¯H ¯f1 − a1 f1 ¯f1 − b1 f1 ¯f0 = 0, (B.4b)
2

­ ¯ ® ­ ¯ ® ­ ¯ ¯ ® ­ ¯ ®
using f1 ¯f0 = ( f0 ¯f1 )† = 0 and a1 = f1 ¯H ¯f1 / f1 ¯f1 . Also
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
f0 ¯f2 = f0 ¯H ¯f1 − a1 f0 ¯f1 − b21 f0 ¯f0 . (B.5)
­ ¯ ­ ¯ ­ ¯
Using f0 ¯H = f1 ¯ + a0 f0 ¯ gives
­ ¯ ® ­ ¯ ® ­ ¯ ® ­ ¯ ®
f0 ¯f2 = f1 ¯f1 + a0 f0 ¯f1 − b21 f0 ¯f0 = 0 (B.6)
­ ¯ ® ­ ¯ ®
using b21 = f1 ¯f1 / f0 ¯f0 .
¯ ®
¯f3 :
¯ ® ¯ ® ¯ ® ¯ ®
¯f3 = H ¯f2 − a2 ¯f2 − b22 ¯f1 , (B.7a)
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
¯ ¯ ¯ ¯ 2 ¯
f2 f3 = f2 H f2 − a2 f2 f2 − b2 f2 f1 = 0, (B.7b)
¯ ® ¯ ®
using the orthogonality of ¯f2 and ¯f1 .
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
f1 ¯f3 = f1 ¯H ¯f2 − a2 f1 ¯f2 − b22 f1 ¯f1
­ ¯ ® ­ ¯ ® ­ ¯ ® ­ ¯ ®
= f2 ¯f2 + a1 f1 ¯f2 + b21 f0 ¯f2 − b22 f1 ¯f1
= 0, (B.8)
­ ¯ ® ­ ¯ ®
using previous results and b22 = f2 ¯f2 / f1 ¯f1 .
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
f0 ¯f3 = f0 ¯H ¯f2 − a2 f0 ¯f2 − b22 f0 ¯f1
­ ¯ ® ­ ¯ ®
= f1 ¯f2 + a0 f0 ¯f2 = 0, (B.9)
­ ¯ ­ ¯ ­ ¯
using f0 ¯H = f1 ¯ + a0 f0 ¯ and previous results.
Thus the first few Lanczos vectors are orthogonal. Next we consider orthog-
onality of Lanczos state n + 1 for n ≥ 3,
¯ ® ¯ ® ¯ ® ¯ ®
¯fn+1 = H ¯fn − an ¯fn − b2n ¯fn−1 , (B.10a)
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
¯ ¯ ¯ ¯ 2
fn fn+1 = fn H fn − an fn fn − bn fn fn−1 ¯
¡­ ¯ ®¢†
= −b2n fn−1 ¯fn . (B.10b)

Hence orthogonality of state n+1 with state n depends on the orthogonality

of state n with state n − 1. Since n is arbitrarily chosen orthogonality of
state n with state n − 1 depends of orthogonality of state n − 1 with state
n − 2 and so on.
Thus since the first few Lanczos basis states are orthogonal we have
shown that the full Lanczos basis satisfies,
­ ¯ ®
fm ¯fn = 0, m = n ± 1. (B.11)
 107
B.2 Orthogonality of basis

Similarly we consider
­ ¯ ® ­ ¯ ¯ ® ­ ¯ ® ­ ¯ ®
fn−1 ¯fn+1 = fn−1 ¯H ¯fn − an fn−1 ¯fn − b2n fn−1 ¯fn−1
­ ¯ ® ­ ¯ ® ­ ¯ ® ­ ¯ ®
= fn ¯fn + an−1 fn−1 ¯fn + b2n−1 fn−2 ¯fn − fn ¯fn
­ ¯ ®
= b2n−1 fn−2 ¯fn , (B.12)
­ ¯ ®
since above we showed that fn−1 ¯fn = 0. Hence orthogonality of state
n − 1 with state­ n + ¯1 depends
® on orthogonality of state
­ n¯− 2 with
® state
¯ ¯
n. Considering fn−2 fn we find that it depends on fn−3 fn−1 etc. and
hence since the first Lanczos vectors satisfy this relation we have shown that
­ ¯ ®
fm ¯fn = 0, m = n ± 2. (B.13)

Continuing this scheme we may show that

­ ¯ ®
fm ¯fn = 0, m = n ± 3, (B.14a)
­ ¯ ®
fm ¯fn = 0, m = n ± 4, (B.14b)
..
.

and hence all in all,

­ ¯ ® ­ ¯ ®
fm ¯fn = fn ¯fn δm,n . (B.15)
108
Appendix C

Anticommutation and
matrices

In this appendix we demonstrate in detail the properties of the fermionic

sign operator P and (anti)commutation rules for fermionic operators and
bosonic operators in the context of matrix representations. For further de-
tails regarding matrix P we refer to [26]. To be specific we use the Hubbard
chain as example.

C.1 Matrices

We use two fermionic and a bosonic operator as examples and use a site
basis defined by |0i, |↑i ≡ c†↑ |0i, |↓i ≡ c†↓ |0i and |↑↓i ≡ c†↑ c†↓ |0i. In this
basis we find the matrix representations

 
0 1 0 0
0 0 0 0
c↑ = 
0
, (C.1a)
0 0 1
0 0 0 0
 
0 0 1 0
0 0 0 −1
c↓ = 
0
, (C.1b)
0 0 0
0 0 0 0
 
0 0 0 0
0 1 0 0
N = 0
, (C.1c)
0 1 0
0 0 0 2

109
110
Anticommutation and matrices

while the matrix representation of the fermionic sign operator is

 
1 0 0 0
0 −1 0 0
P = 
0
. (C.2)
0 −1 0
0 0 0 1

C.2 Anticommutators

First it should be noted that the site operators themselves, (C.1a) and
(C.1b), satisfy the correct anticommutation relations, cσ c†σ̃ + c†σ̃ cσ = δσ,σ̃ .
Here we show how to maintain these relations when expanding the basis to
several sites.
First we expand the basis of operators c1σ and c2σ to the basis of both
sites,

c̃1σ = c1σ ⊗ δ2 , (C.3a)

c̃2σ = P1 ⊗ c2σ , (C.3b)

or more specifically

c̃1↑ = c1↑ ⊗ δ2
 
0 δ2 0 0
0 0 0 0
= 
0 0 0
, (C.4)
δ2 
0 0 0 0

and

c̃2↑ = P1 ⊗ c2↑
 
c2↑ 0 0 0
 0 −c2↑ 0 0
= 
0
. (C.5)
0 −c2↑ 0
0 0 0 c2↑

We use a compact notation in which entries of matrices can themselves be

matrices, and thus Eq. (C.4) and (C.5) are 16 × 16 matrices. Ordinary
 111
C.2 Anticommutators

matrix multiplication yields

  c† 0 0 0

0 δ2 0 0 2↑

0 0 0 0   0 −c2↑ 0 †
0 
c̃1↑ c̃†2↑ + c̃†2↑ c̃1↑ = 
0 0 0 δ2  
 
0 0 −c†2↑ 0
0 0 0 0 0 0 0 c†2↑
 †  
c2↑ 0 0 0 0 δ2 0 0
 0 −c† 
 2↑ 0 0  0 0 0 0 
+ 
0 0 −c†2↑ 0  0 0 0 δ2 
0 0 0 c†2↑ 0 0 0 0
 
0 [c†2,↑ , δ2 ] 0 0
 
0 0 0 0 
=  † , (C.6)
0 0 0 −[c2,↑ , δ2 ]
0 0 0 0
where
     
0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0
 1 0 0 0 0 1 0 0 0 1 0 0 1 0 0 0
[c†2,↑ , δ2 ] = 
0
 −  
0 0 0 0 0 1 0 0 0 1 0 0 0 0 0
0 0 1 0 0 0 0 1 0 0 0 1 0 0 1 0
 
0 0 0 0
0 0 0 0
= 
0
. (C.7)
0 0 0
0 0 0 0
Thus indeed we find c̃1↑ c̃†2↑ + c̃†2↑ c̃1↑ = 0. Had we used δ1 instead of P1 this
would not have been the case.
Similarly we may consider
  c† 0 0 0

c2↑ 0 0 0 2↑
 0 −c2↑ 0  0 −c† 0
† †  0   2↑ 0 
c̃2↑ c̃2↑ + c̃2↑ c̃2↑ =   
0 0 −c2↑ 0   0 †
0 −c2↑ 0 
0 0 0 c2↑ 0 0 0 c†2↑
 † 
c2↑ 0 0 0 c2↑ 0 0 0

 0 −c† 0
 0  0 −c2↑ 0 0
+ 2↑
 

0 †
0 −c2↑ 0  0 0 −c2↑ 0 
0 0 0 c†2↑ 0 0 0 c2↑
 † 
{c2↑ , c2↑ } 0 0 0
 0 {c†2↑ , c2↑ } 0 0 
 
=   , (C.8)
 0 0 {c†2↑ , c2↑ } 0 
0 0 0 {c†2↑ , c2↑ }
112
Anticommutation and matrices

where
     
0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0
 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0
{c†2↑ , c2↑ } = 
0
 +  
0 0 1 0 0 0 0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0
 
1 0 0 0
0 1 0 0
= 
0
, (C.9)
0 1 0
0 0 0 1

and hence indeed {c̃2↑ , c̃†2↑ } = 1. Perhaps it should be pointed out that this
property would be maintained even if one had used δ1 instead of the correct
P1 .

It should be noted that the fermionic sign matrix for sites 1 and 2 is

P12 = P1 ⊗ P2
 
P2 0 0 0
 0 −P2 0 0
= 
0
, (C.10)
0 −P2 0 
0 0 0 P2

where it is clear that an odd total number of particles on sites 1 and 2 gives
a sign, while an even total number of particles does not.

C.3 Commutators
The number operator (C.1c) is diagonal – and remains diagonal after en-
largement of the basis. Diagonal matrices commute – and hence so does the
number operators before and after enlarging the basis. Hence any commu-
tation relations of diagonal matrices are preserved in the proces of enlarging
the basis. Note that we use δ and not P when enlarging the Hilbert space
for bosonic operators.
Appendix D

Truncation of a continued
fraction

To explore what kind of convergence we can expect for the DMRG calcula-
tion of the spectral function in Sect. 4.5.2 we consider the selfenergy used
in Sect. 4.7. For the selfenergy we have both an exact result and a contin-
ued fraction we can truncate and hence we can compare the two. Writing
explicitly ω + iη we have

ΣR (ω) ≡ V 2 GR (ω),
1
GR (ω) = t2
ω + iη − ² − ω+iη−²−···
p
(ω + iη − ²) − sign(ω − ²) (ω + iη − ²)2 − 4t2
= . (D.1)
2t2
Truncating the continued fraction and comparing to the exact result leads
to Fig. D.1. The value of η is chosen fairly large, η = 0.01 to broaden the
δ-peaks.
The structure of the Lanczos method is similar to the continued
­ fraction
®
and hence we expect this kind of convergence when calculating ci c†i and
­ † ®
ci ci for increased number of Lanczos coefficients ai and b2i .

113
114

35 14
R R
−Im G (ω) to 5‘th order −Im G (ω) to 15‘th order
Exact result Exact result
30 12

25 10

20 8

15 6

10 4

5 2

0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(a) 5th order. (b) 15th order.

2.5 1
R R
−Im G (ω) to 100‘th order −Im G (ω) to 1000‘th order
Exact result 0.9 Exact result

2 0.8

0.7

1.5 0.6

0.5

1 0.4

0.3

0.5 0.2

0.1

0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(c) 100th order. (d) 1000th order.

Figure D.1: Imaginary part of −GR (ω) for ² = 0 and η = 0.01.

Appendix E

Equation Of Motion
Technique

Here we derive the Equations Of Motion (EOM) for the dot and lead Green’s
functions in the two cases relevant for this thesis. In this appendix a conver-
gence factor ω → ω + iη is understood, but not explicitly written, n retarded
Green’s functions .

E.1 Quantum dot with single infinite lead

The Hamiltonian for this setup is

H = HD + HL + HDL , (E.1a)
b
X b
X
HD = (²i − Vg ) a†i ai ≡ εi a†i ai , (E.1b)
i=a i=a
∞ ³
X ´ ∞
X
† †
HL = −t cj cj+1 + cj+1 cj + ε c†j cj , (E.1c)
j=1 j=1
b
X ³ ´
HDL = Vi a†i c1 + c†1 ai . (E.1d)
i=a

In order not to clutter up the notation we denote by a and b the dot levels
and by 1, 2, · · · the lead sites.
The Green’s function for the dot levels is generally defined as
­© † 0 ª®
GR 0 0
ij (t − t ) = −iθ(t − t ) ai (t), aj (t ) . (E.2)

We seek the EOM for GR 0

ij (t − t ), and consider
­© † ª® ­© ª®
i∂t GR 0 0
ij (t − t ) = δ(t − t ) ai (t), aj (t) + θ(t − t0 ) ∂t ai (t), a†j (t0 )
­© ª®
= δ(t − t0 )δij + θ(t − t0 ) ∂t ai (t), a†j (t0 ) . (E.3)

115
116
Equation Of Motion Technique

The time derivative of an operator in the Schrödinger picture is

¡ ¢
∂t ai (t) = ∂t eiHt ai e−iHt
= eiHt (iHai − iai H + ∂t ai ) e−iHt
= i [H, ai ] (t), (E.4)

where we used that ai has no explicit time dependence, and where H is the
full Hamiltonian of the system. Hence we find

[H, ai ] = [HD , ai ] + [HDL , ai ] ,

X £ † ¤
[HD , ai ] = εj aj aj , ai
j
X ³ ´
= εj a†j {aj , ai } − {a†j , ai }aj
j
= −εi ai , (E.5a)
X £ † ¤
[HDL , ai ] = Vj aj c1 + c†1 aj , ai
j
X ³ ´
= Vj a†j {c1 , ai } − {a†j , ai }c1 + c†1 {aj , ai } − {c†1 , ai }aj
j
= −Vi c1 . (E.5b)

Thus the EOM for GR 0

ij (t − t ) becomes
­© † 0 ª®
i∂t GR 0 0 0
ij (t − t ) = δ(t − t )δij − εi iθ(t − t ) ai (t), aj (t )
­© ª®
−Vi iθ(t − t0 ) c1 (t), a†j (t0 )
≡ δ(t − t0 )δij + εi GR 0 R 0
ij (t − t ) + Vi G1j (t − t ), (E.6)

where the hybrid Green’s function GR 1j was defined.

Fourier transformation thus yields,
Z ∞
dω R 0
(i∂t − εi ) Gij (ω)e−iω(t−t ) =
−∞ 2π
Z ∞
0 dω R 0
δ(t − t )δij + Vi Gij (ω)e−iω(t−t ) ⇔
−∞ 2π
Z ∞
dω −iω(t−t0 )
(ω − εi )GR ij (ω)e =
−∞ 2π
Z ∞
dω −iω(t−t0 )
(δij + V GR ij (ω))e , (E.7)
−∞ 2π

so that in fourier space we find the EOM

(ω − εi )GR R
ij (ω) = δij + Vi G1j (ω). (E.8)
 117
E.1 Quantum dot with single infinite lead

We also need the EOM for GR

1j , found similarly,

­© † ª® ­© ª®
i∂t GR 0 0
1j (t − t ) = δ(t − t ) c1 (t), aj (t) + θ(t − t0 ) ∂t c1 (t), a†j (t0 )
­© ª®
= θ(t − t0 ) ∂t c1 (t), a†j (t0 ) . (E.9)

The time derivative of c1 (t) is,

[H, c1 ] = [HL , c1 ] + [HDL , c1 ] ,

Xh † i Xh † i
[HL , c1 ] = ε cj cj , c1 − t cj cj+1 + c†j+1 cj , c1
j j
X³ † ´
= ε cj {cj , c1 } − {c†j , c1 }cj
j
X³ ´
−t c†j {cj+1 , c1 } − {c†j , c1 }cj+1 + c†j+1 {cj , c1 } − {c†j+1 , c1 }cj
= −εc1 + tc2 , (E.10a)
X h † i
[HDL , c1 ] = Vj aj c1 + c†1 aj , c1
j
X ³ ´
= Vj a†j {c1 , c1 } − {a†j , c1 }c1 + c†1 {aj , c1 } − {c†1 , c1 }aj
j
X
= − Vj aj , (E.10b)
j

and hence
­© † 0 ª®
i∂t GR 0 0
1j (t − t ) = −εiθ(t − t ) c1 (t), aj (t )
³ ­© ª®´
−t −iθ(t − t0 ) c2 (t), a†j (t0 )
X ­© ª®
− Vi iθ(t − t0 ) ai (t), a†j (t0 ) . (E.11)
i

Fourier transforming we find

X
(ω − ε)GR R
1j (ω) = −tG2j (ω) + Vi GR
ij (ω). (E.12)
i

Yet another hybrid Green’s function has been introduced, and trying to find
the EOM for GR 2j introduces yet another hybrid Green’s function and the
equations never close.
All these EOMs can be understood by the following simple argument,
taking e.g. GR1j as an example: From the first site in the lead there are
limited number of possibilities of propagating. You can either stay, giving a
term ε, or jump to the dot, via couplings Vi , or jump to the second site in
the lead via coupling −t, hence the EOM in Eq. (E.12).
118
Equation Of Motion Technique

Solution
We specialize to the case ij = aa, and consider solutions to increasing order
of hybrid Green’s functions.

Finite order
0th order: With GR
1a = 0 we find from Eq. (E.8)

1
GR
aa = = gaR , (E.13)
ω − εa

which is the well known free result.

1st order: Here GR

2a = 0 and hence we find from Eq. (E.8) and (E.12)

Va Vb
GR
1a = GR
aa + GR ⇒
ω−ε ω − ε ba
Va Vb
1 ω−ε
GR
aa = Va2
+ Va2
GR
ba . (E.14)
ω − εa − ω−ε ω − εa − ω−ε

To this order we may define selfenergies

Va2
ΣR
1,aa = , (E.15a)
ω−ε
Va Vb
ΣR
1,ab = . (E.15b)
ω−ε

2nd order: Here GR

3a = 0,

−t R
GR
2a = G ⇒
ω − ε 1a
Va Vb
GR
1a = t2
GR
aa + t2
GR ba ⇒
ω−ε− ω−ε ω−ε − ω−ε
Va Vb
t2
1 ω−ε− ω−ε
GR
aa = Va2
+ Va2
GR
ba , (E.16)
ω − εa − t2
ω − εa − t2
ω−ε− ω−ε ω−ε− ω−ε

and again we may define selfenergies,

Va2
ΣR
2,aa = t 2 , (E.17a)
ω − ε − ω−ε
Va Vb
ΣR
2,ab = t2
. (E.17b)
ω−ε− ω−ε
 119
E.1 Quantum dot with single infinite lead

Infinite order

It is easy to see that to infinite order we find

Va Vb
t2
1 ω−ε− ω−ε−···
GR
aa = Va2
+ Va2
GR
ba , (E.18)
ω − εa − t2
ω − εa − t2
ω−ε− ω−ε−··· ω−ε− ω−ε−···

and hence selfenergies

Va2
ΣR
∞,aa = t 2 , (E.19a)
ω − ε − ω−ε−···
Va Vb
ΣR
∞,ab = t2
. (E.19b)
ω−ε− ω−ε−···

Defining the selfenergy and the Green’s function for the first site in the lead
in absence of the quantum dot,

ΣR ≡ V 2 G̃R
11 , (E.20a)
1
G̃R
11 = t2
, (E.20b)
ω − ε − ω−ε−···

we can restate Eq. (E.18) like

1 ΣR
GR
aa = + ab
GR
ba . (E.21)
ω − ε a − ΣR
aa ω − εa − ΣR
aa

In order to solve for GR R

aa we need to find Gba . This is done similarly to
GR
aa ,

ΣR
GR
ba =
ba
GR
aa , (E.22)
ω − εb − ΣR
bb

and hence the final solution is

1
GR
aa (ω) = . (E.23)
ΣR R
ab (ω)Σba (ω)
ω − εa − ΣR
aa (ω) − ω−εb −ΣRbb (ω)

The solution for GR

bb is found analogously,

1
GR
bb (ω) = . (E.24)
ΣR R
ba (ω)Σab (ω)
ω − εb − ΣR
bb (ω) − ω−εa −ΣRaa (ω)
120
Equation Of Motion Technique

Finally using the selfsimilarity of G̃R

11 we find the closed formal solution,

1
G̃R
11 =
ω − ε − t2 G̃R11
1
= ⇒
(g R )−1 − t2 G̃R
11
t2 g R (G̃R 2 R
11 ) − G̃11 + g
R
= 0⇒
p
1± 1 − 4t2 (g R )2
G̃R
11 =
2t2 g R
p
(g R )−1 ± (g R )−1 1 − 4t2 (g R )2
= , (E.25)
2t2

where the free Green’s function g R = 1/(ω − ²) was introduced. In the limit
of vanishing hopping element, t → 0, we expect to find the free result, and
hence using L’Hospitals rule we find

p
1 ± (g R )−1 1 − 4t2 (g R )2 R −1 (g R )2
lim ∼ lim ∓(g ) p
t→0 2t2 t→0 1 − 4t2 (g R )2
= ∓g R , (E.26)

In our case lead energies vanish, ² = 0, and hence (g R )−1 = ω. Thus we find
the solution,

p
ω/2t − ω/2t 1 − 4t2 /ω 2
G̃R
11 =
t p
ω/2t − sign(ω) (ω/2t)2 − 1
= . (E.27)
t

In all the above equations we have ignored the infinitesimal imaginary con-
vergence factor iη. Putting in by hand ω → ω + iη the plot of the real and
imaginary part of the Green’s function G̃R11 is shown in Fig. E.1.
It should be noted that the real and imaginary parts of G̃R
11 (ω) are given
by

p
ω/2t − θ(|ω|/2t − 1)sign(ω) (ω/2t)2 − 1
ReG̃R
11 = , (E.28a)
p t
−θ(1 − |ω|/2t) 1 − (ω/2t)2
ImG̃R
11 = , (E.28b)
t

where there is no hidden iη.

 121
E.2 Quantum dot with two infinite leads

0.5

0
11

−0.5
G

−1

−1.5
real(G (ω))
11
imag(G (ω))
11
−2
−4 −3 −2 −1 0 1 2 3 4
ω

Figure E.1: Plot of real and imaginary part of G̃R

11 (ω).

E.2 Quantum dot with two infinite leads

Here the Hamiltonian is

H = HDot + HRight Lead + HLef t Lead + HDL, (E.29a)

b
X X
HDot = (²i − Vg )a†i ai ≡ εa a†i ai , (E.29b)
i=a i
∞
X ∞
X
† †
HRight Lead = −t (cj cj+1 + cj+1 cj ) + ε c†j cj , (E.29c)
j=1 j=1
−∞
X −∞
X
HLef t Lead = −t (c†j cj−1 + c†j−1 cj ) +ε c†j cj , (E.29d)
j=−1 j=−1
b ³
X ´
HDL = ViR (a†i c1 + c†1 ai ) + ViL (a†i c−1 + c†−1 ai ) , (E.29e)
i=a

and the Green’s functions are defined similarly,

­© † 0 ª®
GR 0 0
ij (t − t ) = −iθ(t − t ) ai (t), aj (t ) . (E.30)
122
Equation Of Motion Technique

The analysis is essentially the same, only there are more couplings available.
The final result resembles the single lead case,
1
GR
aa = , (E.31a)
ΣR R
ab Σba
ω − ε a − ΣR
aa − ω−εb −ΣRbb
1
GR
bb = , (E.31b)
ΣR R
ba Σab
ω − εb − ΣR
bb − ω−εa −ΣRaa

where the selfenergies are

¡ R R ¢
ΣR
ij = Vi Vj + ViL VjL G̃R
11 , (E.32)

and where G̃R 11 is given in Eq. (E.27).

We see that the overall structures of the two results are similar. However
it is not possible to map e.g. the single lead Green’s function onto the double
lead Green’s function by simply redefining coupling constants. This means
that the two systems indeed are different, and hence we cannot model the
two lead system by modifying the couplings in the single lead system.
Appendix F

Evaluation of ni for
non-interacting biased
quantum dot

In this appendix we consider more explicitly the integration of the Green’s

function G<ij in Chap. 5.

F.1 Analytic expressions

The retarded Green’s functions can be written in matrix form,

GR = gR + gR ΣR GR , (F.1)

where GR is the full retarded Green’s function, gR is the corresponding free

Green’s function, and ΣR is the retarded selfenergy. This result enables us
to use a result from non-equilibrium Green’s function theory,1

G< = GR Σ< GA , (F.2)

or more explicitly we use

G< R < A R < A R < A R < A

aa = Gaa Σaa Gaa + Gab Σba Gaa + Gaa Σab Gba + Gab Σbb Gba , (F.3a)
G<
bb = GR < A
bb Σbb Gbb + GR < A
ba Σab Gbb + GR < A
bb Σba Gab + GR < A
ba Σaa Gab , (F.3b)

where

Σ< R R < L L <

ij = Vi Vj G̃11,µR + Vi Vj G̃11,µL , (F.4a)
R/L
G̃<
11,µR/L = −2inF (ω)ImG̃R
11 , (F.4b)
R/L 1
nF (ω) = . (F.4c)
e−β(µL/R −ω) + 1
1
[18], p. 165-166.

123
124
Evaluation of ni for non-interacting biased quantum dot

Using this result we may compute the occupation on the dot in the
standard way

ni (Vg ) = −iG< (t = 0, Vg )
Z ∞ ii
dω <
= Gii (ω, Vg ). (F.5)
−∞ 2πi

F.2 Numerics
In this section we evaluate the different parts entering the expression for the
occupation and evaluate the occupation of the dot.

F.2.1 GA (ω, Vg )
The advanced Green’s function is the complex conjugate of the retarded
Green’s function,
¡ R ¢∗
GA (ω, Vg ) = G (ω, Vg ) , (F.6)

and the retarded Green’s functions were found explicitly using EOM tech-
niques in App. E. Plots of ImGA (ω) for various couplings and gate voltages
are shown in Fig. F.1, F.2, and F.3.
Any properly defined spectral function obey the sum rule
­ ®
A(t) ≡ {c(t), c† (0)} , (F.7a)
Z ∞ Z ∞ Z
dω dω ∞
A(ω) = dteiωt A(t)
−∞ 2π −∞ 2π −∞
Z ∞
= dtδ(t)A(t)
−∞
= A(0) = 1. (F.7b)

When considering the figures of ImGA there are two things to notice: (i) The
scales are very different, and (ii) the integrated weights of these functions
are not equal.
This does not contradict the sum rule in Eq. (F.7), since although the
imaginary part of the selfenergy vanishes outside the band, there is still the
infinitesimal imaginary part ω → ω − iη in the advanced Green’s function.
η is infinitesimal, and the imaginary part of the advanced Green’s function
is proportional to η outside the band, making ImGA mostly zero. But if the
denominator of GA vanishes for particular values of ω then ImGA has δ-like
peaks for those values of ω and hence also integrated weight.
 125

0.09 60
A A
Im Gaa(ω) Im Gaa(ω)
0.08 A A
Im Gbb(ω) Im Gbb(ω)
50
0.07

0.06 40

0.05
30
0.04

0.03 20

0.02
10
0.01

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

60 80
A A
Im Gaa(ω) Im Gaa(ω)
A A
Im Gbb(ω) 70 Im Gbb(ω)
50
60
40
50

30 40

30
20
20
10
10

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

0.2 0.06
A A
Im Gaa(ω) Im Gaa(ω)
0.18 A
Im Gbb(ω)
A
Im Gbb(ω)
0.05
0.16

0.14
0.04
0.12

0.1 0.03

0.08
0.02
0.06

0.04
0.01
0.02

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.1: ImGA (ω) for Vi = 0.5. Note the different scales.
126

0.05 25
A A
Im Gaa(ω) Im Gaa(ω)
0.045 A
Im Gbb(ω)
A
Im Gbb(ω)
0.04 20

0.035

0.03 15

0.025

0.02 10

0.015

0.01 5

0.005

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

18 30
A A
Im Gaa(ω) Im Gaa(ω)
16 A A
Im Gbb(ω) Im Gbb(ω)
25
14

12 20

10
15
8

6 10

4
5
2

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

0.35 0.03
A A
Im Gaa(ω) Im Gaa(ω)
A A
Im Gbb(ω) Im Gbb(ω)
0.3
0.025

0.25
0.02

0.2
0.015
0.15

0.01
0.1

0.005
0.05

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.2: ImGA (ω) for Vi = 0.25. Note the different scales.
 127

−3
x 10 250
3 A
Im Gaa(ω)
A
Im Gaa(ω) A
A Im Gbb(ω)
Im Gbb(ω)
2.5 200

2
150

1.5
100
1

50
0.5

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

200 250
A A
Im Gaa(ω) Im Gaa(ω)
180 A
Im Gbb(ω)
A
Im Gbb(ω)
160 200

140

120 150

100

80 100

60

40 50

20

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

−3
0.06 x 10
Im Gaa(ω)
A 1.4
A
A Im Gaa(ω)
Im Gbb(ω) A
0.05 Im Gbb(ω)
1.2

0.04 1

0.8
0.03

0.6
0.02
0.4

0.01
0.2

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.3: ImGA (ω) for Vi = 0.05. Note the different scales.
128
Evaluation of ni for non-interacting biased quantum dot

8 8
V =−2 V =−2
g g
V =−2 V =−2
6 g 6 g
V =−1 V =−1
g g
V =−1 V =−1
g g
4 V =0 4 V =0
g g
V =0 V =0
g g
2 V =1 2 V =1
g g
V =1 V =1
g g
0 Vg=2 0 Vg=2
Vg=2 Vg=2
−2 −2

−4 −4

−6 −6

−8 −8
−4 −2 0 2 4 −4 −2 0 2 4
ω ω

L/R L/R
(a) Vi = 0.5. (b) Vi = 0.25.

8
V =−2
g
V =−2
6 g
V =−1
g
V =−1
g
4 V =0
g
V =0
g
2 V =1
g
V =1
g
0 Vg=2
Vg=2
−2

−4

−6

−8
−4 −2 0 2 4
ω

L/R
(c) Vi = 0.05.

Figure F.4: Graphical location of singular points for various couplings. Outside the
band a vanishing denominator corresponds to a singular point. Full lines are for
GR R
aa and dotted lines are for Gbb .

Singularities

By considering the denominator of GR or GA it is possible to find graphically

the position of the δ-like peaks outside the band. In general we have

1 −ImX
Im = , (F.8)
ReX + iImX (ReX)2 + (ImX)2

so that if ImX = 0+ in some region the singular points are determined by

ReX = 0.
In this context this corresponds to
µ ¶
ΣR R
ab Σba
0 = Re ω − (εa − Vg ) − ΣR
aa − , (F.9)
ω − (εb − Vg ) − ΣR
bb
 129
F.2 Numerics

where

ΣR L L R R R
ij = (Vi Vj + Vi Vj )G̃11 , (F.10a)
p
ω/2t − θ(|ω|/2t − 1)sign(ω) (ω/2t)2 − 1
ReG̃R
11 = , (F.10b)
p t
−θ(1 − |ω|/2t) 1 − (ω/2t)2
ImG̃R
11 = . (F.10c)
t
Eq. F.9 can easily be solved graphically, to locate the singular points. In
Fig. F.4 we have plotted the real part of the denominator of GR aa and Gbb
R

for a range of differen gate voltages. It is clearly seen that there are indeed
singular points outside the band and hence the spectral function has δ-like
peaks at these points.
A comparison with Fig. F.1-F.3 shows that the values of Vg where sin-
gular points exist outside the band correspond spectral functions without
pronounced δ-like peaks, and where integrated weight is missing.

F.2.2 G<
ii (ω, Vg )

The lesser Green’s functions are given explicitly in Eq. (F.3). In Fig. F.5, F.6
and F.7 we show plots of the imaginary part of the lesser Green’s functions
inside the band for various gate voltages and couplings to the leads.

F.2.3 ni (Vg )
Integrating G< ii (ω, Vg )/2πi over ω ∈ [−∞, ∞], and using a fine grid of gate
voltages gives ni (Vg ). In practice we use integration over a finite interval,
where it is essential to include the singular points discussed above. Plots of
analytical as well as DMRG occupations are shown in Fig. F.8.
It should be noted that numeric errors might be present in the occupa-
tions due to the δ-like peaks in the spectral function discussed above. Such
peaks are very difficult to integrate numerically, which possibly introduces
numerical errors in the occupations presented here. A more careful analytic
study should be performed to eliminate these errors.
130

0.12 0.8
< <
Im Gaa(ω) Im Gaa(ω)
< <
Im Gbb(ω) 0.7 Im Gbb(ω)
0.1
0.6
0.08
0.5

0.06 0.4

0.3
0.04
0.2
0.02
0.1

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

120 140
< <
Im Gaa(ω) Im Gaa(ω)
< <
Im Gbb(ω) Im Gbb(ω)
120
100

100
80

80
60
60

40
40

20
20

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

0.4 0.12
< <
Im Gaa(ω) Im Gaa(ω)
< <
0.35 Im Gbb(ω) Im Gbb(ω)
0.1
0.3
0.08
0.25

0.2 0.06

0.15
0.04
0.1
0.02
0.05

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.5: ImG< (ω) for Vi = 0.5. Notice the cut-off due to the finite band-
width.
 131

0.035 0.4
< <
Im Gaa(ω) Im Gaa(ω)
< <
Im Gbb(ω) 0.35 Im Gbb(ω)
0.03

0.3
0.025

0.25
0.02
0.2
0.015
0.15

0.01
0.1

0.005 0.05

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

35 60
< <
Im Gaa(ω) Im Gaa(ω)
< <
Im Gbb(ω) Im Gbb(ω)
30
50

25
40

20
30
15

20
10

10
5

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

0.7 0.06
< <
Im Gaa(ω) Im Gaa(ω)
< <
Im Gbb(ω) Im Gbb(ω)
0.6
0.05

0.5
0.04

0.4
0.03
0.3

0.02
0.2

0.01
0.1

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.6: ImG< (ω) for Vi = 0.25. Notice the cut-off due to the finite band-
width.
132

−3
x 10 0.016
1.4 <
Im Gaa(ω)
<
Im Gaa(ω) <
< 0.014 Im Gbb(ω)
Im Gbb(ω)
1.2
0.012
1
0.01
0.8
0.008

0.6
0.006

0.4
0.004

0.2 0.002

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(a) Vg = −4. (b) Vg = −2.

400 500
< <
Im Gaa(ω) Im Gaa(ω)
<
Im Gbb(ω) 450 <
Im Gbb(ω)
350
400
300
350
250
300

200 250

200
150
150
100
100
50
50

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(c) Vg = −1. (d) Vg = 0.

−3
0.1 x 10
<
Im Gaa(ω) 3
<
0.09 < Im Gaa(ω)
Im Gbb(ω) <
Im Gbb(ω)
0.08 2.5

0.07
2
0.06

0.05 1.5
0.04

0.03 1

0.02
0.5
0.01

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
ω ω

(e) Vg = 1. (f) Vg = 2.

L/R
Figure F.7: ImG< (ω) for Vi = 0.05. Notice the cut-off due to the finite band-
width.
 133

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Occupation

Occupation
0.5 0.5

0.4 0.4

0.3 0.3
n1, DMRG n1, DMRG
0.2 0.2
n2, DMRG n2, DMRG
0.1 n , Analytic 0.1 n , Analytic
1 1
n , Analytic n , Analytic
2 2
0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v v
g g

L/R L/R
(a) Vi = 0.5. (b) Vi = 0.25.

0.9

0.8

0.7

0.6
Occupation

0.5

0.4

0.3
n1, DMRG
0.2
n2, DMRG
0.1 n , Analytic
1
n , Analytic
2
0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
v
g

L/R
(c) Vi = 0.05.

Figure F.8: Comparison between DMRG and analytic occupations for ² =

[−1.3, −1.1], µL = −µR = 0.1, m = 32 states, and 125 sites in the lead. The
large gate occupations are not reproduced by DMRG since we have not incorpo-
L/R
rated the finite band width in our implementation. U , Vi , ², and Vg are given
in units of t.
134
Bibliography

[1] K. G. Wilson and J. Kogut, The Renormalization Group and the ²

Expansion, Physics Reports 12(2), 75–200 (1973).
[2] I. Peschel, X. Wang, M. Kaulke, and K. Hallberg, Density-Matrix
Renormalization, volume 528 of Lecture Notes in Physics, Springer-
Verlag, 1999.
[3] S. R. White, Density matrix formulation for quantum renormalization
groups, Phys.Rev.Lett. 69(19), 2863–2866 (1992).
[4] S. R. White, Density-matrix algorithms for quantum renormalization
groups, Phys.Rev.B. 48(14), 10345–10356 (1993).
[5] P. Schmitteckert, Nonequilibrium electron transport using the density
matrix renormalization group, cond-mat/0403759 (2004).
[6] S. R. White and A. E. Feiguin, Real time evolution using the density
matrix renormalization group, cond-mat/0403310 (2004).
[7] E. Jeckelmann, Dynamical density-matrix renormalization-group
method, Phys.Rev.B. 66, 045114 (2002).
[8] R. P. Feynman, Statistical Mechanics – A set of lectures, W. A. Ben-
jamin, INC., 1972.
[9] N. Shibata, Application of the density matrix renormalization group
method to finite temperatures and two-dimensional systems, cond-
mat/0310028 (2003).
[10] S. Caprara and A. Rosengren, Density-matrix renormalization group
for fermions: Convergence to the infinite-size limit, Nuclear Physics B.
(493), 640–650 (1997).
[11] J. Eising, Lineær Algebra, Institut for Matematik – Danmarks Tekniske
Universitet, 2nd edition, 1997.
[12] E. R. Gagliano and C. A. Balseiro, Dynamical Properties of Quan-
tum Many-Body Systems at Zero Temperature, Phys.Rev.Lett. 59(26),
2999–3002 (1987).

135
136
BIBLIOGRAPHY

[13] K. Hallberg, Density-matrix algorithm for the calculation of dynamical

properties of low-dimensional systems, Phys.Rev.B. 52(14), 9827–9830
(1995).

[14] J. R. Laguna, Real Space Renormalization Group Techniques and Ap-

plications, PhD thesis, Instituto de matemáticas y Fı́sica Fundamental
(CSIC), 2001.

[15] M. Anderson, Numerical Renormalization, Doped Antiferromagnets,

and Statistics – Topics in Low Dimensional Physics, PhD thesis, De-
partement of Theoretical Physics, Chalmers University of Technology,
Göteborg University, 2001.

[16] R. Berkovits, Density matrix renormalization group study of the

charging of a quantum dot strongly coupled to a single lead, cond-
mat/0306284 (2003).

[17] H. Bruus and K. Flensberg, Quantum field theory in condensed matter

physics, Ørsted Laboratory, Niels Bohr Institute, 2001.

[18] H. Haug and A.-P. Jauho, Quantum Kinetics in Transport and Optics
of Semiconductors, Springer Series in Solid-State Sciences, Springer-
Verlag, 1998.

[19] R. Berkovits, F. v. Oppen, and J. W. Kantelhardt, Discrete charging of

a quantum dot strongly coupled to external leads, cond-mat/0307730
(2003).

[20] M. A. Martı́n-Delgado, G. Sierra, and R. M. Noack, The density matrix

renormalization group applied to single-particle quantum mechanics,
J.Phys.A:Math.Gen. 32(33), 6079–6090 (1999).

[21] I. P. Bindloss, Phases of the interacting one-dimensional electron gas

coupled to phonons, cond-mat/0404154 (2004).

[22] E. Papa, Nonuniversal properties of the single-particle density of states

of 1D system with electron-phonon interactions, cond-mat/0403287
(2004).

[23] T. Frederiksen, Inelastic electron transport in nanosystems, Master’s

thesis, MIC/DTU, 2004.

[24] E. Jeckelmann and S. R. White, Density-matrix renormalization-group

study of the polaron problem in the Holstein model, Phys.Rev.B.
57(11), 6376–6385 (1998).

[25] T. D. Kühner and S. R. White, Dynamical Correlation Functions using

the Density Matrix Renormalization Group, cond-mat/9812372 (2004).
 137
BIBLIOGRAPHY

[26] A. Juozapavičius, Density-Matrix Renormalization-Group Analysis of

Kondo and XY models, PhD thesis, Royal Institute of Technology,
Sweden, 2001.

[27] J. König, Y. Gefen, and G. Schön, Level Statistics of Quantum Dots

Coupled to Reservoirs, Phys.Rev.Lett. 81(20), 4468–4471 (1998).

[28] L. E. Ballentine, Quantum Mechanics – A Modern Development, World

Scientific Publishing Co. Pte. Ltd., 1999.

[29] B. H. Bransden and C. J. Joachain, Quantum Mechanics, Prentice

Hall, 2nd edition, 2000.

[30] C. Kittel, Introduction to Solid State Physics, John Wiley & Sons, 7th
edition, 1996.

[31] C. Zhang, E. Jeckelmann, and S. R. White, Density matrix Approach

to Local Hilbert Space Reduction, Phys.Rev.Lett. 80(12), 2661–2664
(1998).

[32] G. Fano, F. Ortolani, and L. Ziosi, The density matrix renormalization

group method. Application to the PPP model of a cyclic polyene chain.,
cond-mat/980307 (1998).

[33] K. Hallberg, Density Matrix Renormalization: A Review of the Method

and its Applications, cond-mat/0303557 (2003).

[34] M. A. Cazalilla and J. B. Marston, Time-Dependent Density-Matrix

Renormalization Group: A Systematic Method for the Study of Quan-
tum Many-Body Out-of-Equilibrium Systems, Phys.Rev.Lett. 88,
256403 (2002).

[35] G. D. Mahan, Many-Particle Physics, Kluwer Academic/Plenum Pub-

lishers, 2nd edition, 1999.

[36] S. Daul, I. Ciofini, C. Daul, and S. R. White, Full-CI Quantum Chem-

istry Using the Density Matrix Renormalization Group, International
Journal of Quantum Chemistry 79(6), 331–342 (2000).

[37] A. O. Mitrushenkov, G. Fano, F. Ortolani, R. Linguerri, and

P. Palmieri, Quantum chemistry using the density matrix renormal-
ization group, Journal of Chemical Physics 115(15), 6815–6821 (2001).

[38] Y. Park and S. Liang, Charge and spin dynamics of the Hubbard chains,
Physica C 328, 200–206 (1999).

[39] S. Elliot, The Physics and Chemistry of Solids, John Wiley & Sons,
2000.
138
BIBLIOGRAPHY

[40] G. D. Mahan, Applied Mathematics, Kluwer Academic/Plenum Pub-

lishers, 2002.

[41] Y. Nishiyama, Folding of the triangular lattice in a discrete three-

dimensional space: Density-matrix-renormalization-group study, cond-
mat/0403747 (2004).

[42] H. Benthien and F. Gebhard, Spectral function of the one-dimensional

Hubbard model away from half filling, cond-mat/0402664 (2004).

[43] H. Matsueda, T. Tohyama, and S. Maekawa, Dynamical DMRG

study of photoexcited states in one-dimensional Mott insulators, cond-
mat/0402298 (2004).

[44] R. A. Molina, D. Weinmann, and J.-L. Pichard, Length-dependent os-

cillations of the conductance through atomic chains: The importance of
electronic correlations, Europhysics Letters, cond-mat/0402307 (2004).

[45] A. J. Daley, C. Kollath, U. Schollwöck, and G. Vidal, Time-dependent

density-matrix renormalization-group using adaptive effective Hilbert
spaces, cond-mat/0403313 (2004).

[46] D. Garcı́a, K. Hallberg, and M. J. Rozenberg, Dynamical Mean

Field Theory with the Density Matrix Renormalization Group, cond-
mat/0403169 (2004).

[47] S. Nishimoto and E. Jeckelmann, Density-matrix renormalization group

approach to quantum impurity problems, cond-mat/0311291 (2003).

[48] D. J. J. Farnell, Density Matrix Renormalization Group Calculations

for Two-Dimensional Lattices: An Application to the Spin-Half and
Spin-One Square Lattice Heisenberg Models, cond-mat/0309057 (2003).

[49] L. Capriotti, D. J. Scalapino, and S. R. White, Spin-liquid versus

dimerized ground states in a frustrated Heisenberg antiferromagnet,
cond-mat/0404085 (2004).

[50] G. K.-L. Chan and M. Head-Gordon, Highly correlated calculations

with a polynomial cost algorithm: A study of the density matrix renor-
malization group, Journal of Chemical Physics 116(11), 4462–4476
(2002).

[51] I. P. McCulloch and M. Gulácsi, Density Matrix Renormalization

Group Method and Symmetries of the Hamiltonian, Aust.J.Phys. 53,
597–612 (2000).

[52] M. Andersson, M. Boman, and S. Östlund, Density Matrix Renomal-

ization Group of Gapless Systems, cond-mat/9810093 (1998).
 139
BIBLIOGRAPHY

[53] E. Capelluti, B. Cerutti, and L. Pietronero, Charge fluctuations and

electron-phonon interaction in the finite-U Hubbard model, cond-
mat/0312654 (2003).

[54] J. Dukelsky and S. Pittel, The Density Matrix Renormalization Group

for finite Fermi systems, cond-mat/0404212 (2004).

[55] C. A. Büsser, G. B. Martins, K. A. Al-Hassanieh, A. Moreo, and

E. Dagotto, Interference Effects in the Conductance of Multi-Level
Quantum Dots, cond-mat/0404426 (2004).