Lattice Kinetic Methods & Thermal Rectification Phenomena
-
Upload
khangminh22 -
Category
Documents
-
view
1 -
download
0
Transcript of Lattice Kinetic Methods & Thermal Rectification Phenomena
ETH Library
Transport: Lattice KineticMethods & Thermal RectificationPhenomena
Doctoral Thesis
Author(s):Solórzano, Sergio
Publication date:2017
Permanent link:https://doi.org/10.3929/ethz-b-000228259
Rights / license:In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection.For more information, please consult the Terms of use.
Diss. ETH No. 24682
Transport: Lattice KineticMethods & Thermal
Rectification Phenomena
A thesis submitted to attain the degree of
Doctor of Sciences of ETH Zurich
(Dr. sc. ETH Zurich)
presented by
Sergio Daniel Solorzano Rocha
MSc. Physics, ETH Zurich
born 11.10.1989
citizen of Colombia
accepted on the recommendation of
Prof. Dr. Hans J. Herrmann, examiner
Prof. Dr. Sauro Succi, co-examiner
2017
Acknowledgments
Unas gracias infinitas a mi mama, Tomasa, mi papa, la pilla, y el nuevo
miembro de la familia Cutu la pantera negra, por el apoyo incondicional du-
rante este proyecto 24/7/365 en cualquier continente y huso horario. Merci
beaucoup ma petite amie Rocıo pour couter et me montrer que tout est juste
de petits morceaux de la mosaıque de la vie. To Prof. Dr. Hans Herrmann
deep gratitude for giving me the opportunity to work on very interesting
and challenging problems in a very rich environment. Under his supervision
I learnt not only about science, but also and perhaps more important how
to do science. To Dr. Miller Mendoza Deep and Wide gratitude for his vital
guidance and support throughout these years, his ideas and insights always
enriched my work. He patiently helped me through every project, always
having plenty of time for good news and just a little for bad news, but al-
ways having time to guide, discuss and teach. I am also grateful with Miller
for pushing me the right amount and teaching me that things can be done.
To Prof. Dr. Sauro Succi, for the interesting ideas and input regarding the
Wigner function as well as for his patience and thoughtful comments in, the
perhaps, most complex project of my Ph.D. To Prof. Dr. Nuno Araujo for
his patient guidance in the, rich and full of surprises, ratchet project. A
los amigos de la U: J, Henry-Kathe-Chiara, Dani, Alfred, Checho, Laura,
Dra. L. Martin, Nico, El paisa, Juan Pablo del Risco, Monica, Nataly, Juan
David, Rafa, por todas las pendejadas alegres, paseos a Mariquita, Holanda,
Alemania, tareas innecesarias y cosas serias. A mis amigos Luis German y
Angelica por nunca perder el contacto y siempre tener cosas para hablar.
To the set of Zurich friends: Sidharta-Sidharta, Ekin, Ryan, Deep, Nastya,
i
Tom, Hassam, Leila, Tim, Timo, Mathia, Jan, Raphaela, Gabriel, Nishant
Johu. Too many movies, dinners, conversations to remember. Thanks for
the great time together. To the group friends: Robin, Jens-Daniel, Ilario,
Farhang, Lucas, Ryuta, Oliver, Dominik, Trivik, Gautam, Kyriakos, Lau-
rens, Pavel, Kornel, Mirko, Jan, Rodrigo, Juliana. I learn from them, and
they made the time at the group a very enjoyable experience. Thanks for
the time together.
Contents
Zusammenfassung i
Abstract iii
Related publications v
1 Introduction 1
1.1 Overview and motivation . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Part I . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Part II . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Structure and organization . . . . . . . . . . . . . . . . . . . 7
2 General formulation of lattice kinetic methods 9
2.1 Lattice kinetic methods . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 From Boltzmann to Lattice Boltzmann . . . . . . . . 9
2.1.2 Velocity space discretization . . . . . . . . . . . . . . 11
2.1.3 Equilibrium and source term distributions . . . . . . 14
2.1.4 Algorithmic details . . . . . . . . . . . . . . . . . . . 14
3 Lattice kinetic approach to Density Functional Theory 17
3.1 Density functional theory and imaginary time propagation . 17
3.1.1 Density functional theory . . . . . . . . . . . . . . . 17
3.1.2 Imaginary time propagation . . . . . . . . . . . . . . 20
3.1.3 Kinetic approach to DFT . . . . . . . . . . . . . . . 21
3.2 Lattice Kinetic model for DFT . . . . . . . . . . . . . . . . . 26
3.2.1 Formal correction of discrete lattice effects . . . . . . 26
3.2.2 Equilibrium and source distributions construction . . 29
3.2.3 Semi implicit correction . . . . . . . . . . . . . . . . 30
3.3 Application of the improved lattice kinetic approach . . . . . 31
3.3.1 Model comparison . . . . . . . . . . . . . . . . . . . 31
3.3.2 Ethane molecule test . . . . . . . . . . . . . . . . . . 35
3.3.3 Pseudopotentials . . . . . . . . . . . . . . . . . . . . 35
3.3.4 Lattice performance . . . . . . . . . . . . . . . . . . 38
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4 Lattice Wigner Model 41
4.1 Wigner formalism . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Lattice Wigner model . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.1 Harmonic potential . . . . . . . . . . . . . . . . . . . 48
4.3.2 Anharmonic potential . . . . . . . . . . . . . . . . . 52
4.4 Computational cost . . . . . . . . . . . . . . . . . . . . . . . 55
4.5 Lattice Wigner application . . . . . . . . . . . . . . . . . . . 56
4.5.1 1D system . . . . . . . . . . . . . . . . . . . . . . . . 56
4.5.2 2D system . . . . . . . . . . . . . . . . . . . . . . . . 62
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5 Thermal gas rectification using a sawtooth channel 67
5.1 Ratchet systems . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Canonical molecular dynamics . . . . . . . . . . . . . . . . . 69
5.2.1 Molecular dynamics . . . . . . . . . . . . . . . . . . . 69
5.2.2 Predictor corrector integration . . . . . . . . . . . . . 69
5.2.3 Canonical molecular dynamics . . . . . . . . . . . . . 71
5.3 Model and methods . . . . . . . . . . . . . . . . . . . . . . . 73
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6 Conclusion and outlook 83
Appendix 85
A.1 Wigner function derivation . . . . . . . . . . . . . . . . . . . 85
A.2 Generalized Hermite polynomials . . . . . . . . . . . . . . . 89
A.3 Wigner Forcing term calculation . . . . . . . . . . . . . . . . 89
A.4 General Wigner function calculation . . . . . . . . . . . . . . 91
A.5 Multiple barriers effective transmission coefficient . . . . . . 93
A.6 Lattice specification . . . . . . . . . . . . . . . . . . . . . . . 93
References 97
Zusammenfassung
Diese Arbeit besteht aus zwei Teilen. Im Ersten verfolgen wir die Idee
die sogenannten Lattice-Kinetic Methoden, d.h. Erweiterungen der Lattice-
Boltzmann Methode, welche normalerweise fr Fluiddynamik verwendet wer-
den, auf andere Gebiete anzuwenden. Konkret wird eine neue halb-implizite
Korrekturmethode zweiter Ordnung fur das kinetische Kohn-Sham Modell
[1] entwickelt. Mit der neuen Methode werden die Austauschkorrelations-
energien von Atomen und Dimeren fur die ersten beiden Reihen des Peri-
odensystems berechnet und mit Erfahrungswerten verglichen. Dabei wur-
de eine ausgezeichnete Ubereinstimmung gefunden. Des Weiteren wird das
Ethanmolekul simuliert, wobei die Bindungslangen bestimmt und mit Stan-
dardmethoden vergleichen werden. Zuletzt wird gepruft inwiefern Pseudopo-
tentiale im Lattice-Kinetic Kohn-Sham Ansatz Anwendung finden konnen.
Wir verfolgen die Idee Lattice-Kinetic Methoden jenseits der Hydrodynamik
zu untersuchen weiter und entwickeln eine neue Methode um die Wigner-
Gleichung zu losen. Wir validieren den Ansatz fur die Falle eines harmo-
nischen und unharmonischen Potentials. Zudem untersuchen wir ein- und
zweidimensionale offene Quantensysteme mit Potentialbarrieren. Wir zeigen
auf, dass die Methode auch auf dreidimensionale Systeme erweitert werden
kann.
Im zweiten Teil untersuchen wir die Gleichrichtung eines zweidimensionalen
thermischen Gases in einem Kanal mit dissipativen Wanden. Mithilfe von
Molecular-Dynamics Simulationen zeigen wir, dass ein Gas aus Lennard-
Jones Partikeln gleichgerichtet werden kann, sofern es eine raumliche Asy-
metrie und Dissipation im System gibt. Ausserdem demonstrieren wir, dass
i
es eine nichtmonotone Abhangigkeit des Partikelflusses von Systemparame-
tern wie z.B. der Thermostattemperatur, der Kanalasymmetrie und Parti-
keldichte gibt, und finden drei Bereiche mit einer jeweils charakteristischen
Dynamik.
ii
Abstract
This work consists of two parts, in the first one the idea of using lattice
kinetic methods i.e. extensions of the Lattice Boltzmann method to areas
different than fluid dynamics is explored. Particularly, a new semi-implicit
second order correction scheme to the kinetic Kohn-Sham lattice model in-
troduced in Ref.[1] is developed. The new approach is validated by perform-
ing realistic exchange-correlation energy calculations of atoms and dimers
of the first two rows of the periodic table finding good agreement with the
expected values. Additionally, we simulate the ethane molecule where we
recover the bond lengths and compare the results with standard methods.
Finally the current applicability of pseudopotentials within the lattice kinetic
Kohn-Sham approach is also discussed. Continuing with the idea of study-
ing lattice kinetic methods beyond hydrodynamics, a lattice kinetic scheme
to solve the Wigner equation was developed. The approach was validated
for the case of quantum harmonic and anharmonic potentials, showing good
agreement with theoretical results. It was also applied to the study of the
transport properties of one and two dimensional open quantum systems with
potential barriers. The computational viability of the scheme for the case of
three-dimensional systems is also illustrated.
In the second part, the subject of the rectification of a two-dimensional
thermal gas in a channel of asymmetric dissipative walls is studied. We
use molecular dynamics simulations to show that a gas of Lennard-Jones
particles can be rectified provided there is spatial asymmetry and dissipation
in the systems. Further, we find that there is a non-monotonic dependence
of the particle flux on systems parameters such as thermostat temperature,
iii
Related publications
This thesis contains content of the following published or for peer-review
submitted articles:
• S. Solorzano, M. Mendoza, and H. J. Herrmann, “Second order kinetic
Kohn-Sham lattice model,” Phys. Rev. A., vol. 93, p. 062504, 2016.
(arXiv:1605.05032v1)
• S. Solorzano, M. Mendoza, S. Succi and H. J. Herrmann, “Lattice
Wigner equation,” submitted to Phys. Rev. E.
• S. Solorzano, N. A. M. Araujo, and H. J. Herrmann, “Thermal gas
rectification using a sawtooth channel,” Phys. Rev. E., vol. 96, p.
032901, 2017 (arXiv:1706.04429)
For each of these three articles, the author contributed the most effort,
including: design, implementation, and execution of all simulations; analysis
and interpretation of all data generated from simulations; all conceptual
and mathematical derivations; execution of all numerical analyses; literature
research; writing of the manuscript; and design and creation of all figures.
v
Chapter 1
Introduction
1.1 Overview and motivation
This work is composed of two parts. The first one is of quantum nature and
explores the relations between Lattice Kinetic methods, Density Functional
Theory and Wigner Function formalism. The second part is classical and
explores the possibility of using thermal fluctuations, spatial asymmetry and
dissipation to rectify the motion of a gas of particles.
1.1.1 Part I
According to some accounts, Quantum Mechanics is no more than linear
algebra with some physics into it. This is clearly a crass oversimplification
however, there is a bit of truth to it. This impression is well deserved if
the only thing one learns or aims to do is just to diagonalize matrices as
it is usually done at introductory levels. Physical understanding requires
two ingredients to progress, deeper ideas and appropriate tools to handle
these ideas. In the following, we overview two very interesting examples of
the first ingredient, the Density Functional Theory (DFT) and the Wigner
Formulation of quantum mechanics. Later, we comment on how they can
both be handled with the highly versatile tool of Lattice Kinetic Models.
The starting point for the description of microscopic systems is usually
Schrodinger’s equation. It can describe for instance, the dynamics of the
1
CHAPTER 1. INTRODUCTION
protons and electrons that made up matter, ultimately defining electrical,
optical, magnetic and thermal characteristics of materials. In general the
set up of the Schrodinger equation is relatively simple, its solution however
is not. In order to provide a perspective on the DFT approach to the many
electron problem, we provide a concise historical account of the methods
that have been developed to solve the many particle Schrodinger equation.
One of the first systematic approaches to solve Schrodinger’s equation for
a system of many electrons dates back to D.R.Hartree [2] 1928 method. It
was later improved by V. A. Fock [3](1930) by the use of trial wave func-
tions that respected the Fermi exclusion principle and is currently known
as the Hartree-Fock method. In essence, it is a variational approach that
enforces the physically correct symmetry of the wave function. Around the
same time Hylleraas [4, 5] studied the Helium atom problem obtaining good
agreement with experiments. Instead of a single wave function as in the
Hartree-Fock method, he used a superposition of them [6] effectively per-
forming a Configuration Interaction calculation. Also in 1930, Fermi devel-
oped the idea of effective potentials to study scattering problems, providing
the foundation to the modern concept of pseudopotentials [7]. Pseudopoten-
tials are a theoretical as well as a computational tool that allows to reduce
the number of studied electrons by including additional potential terms in
the Schrodinger equation while preserving most of the physics of the system.
The approximation is based on the observation that most of the chemical
and macroscopic properties of materials are determined by the outer most
electrons [8]. Various types have been developed over the years. Some of the
most popular ones, due to their high accuracy and transferability properties,
are Vanderbilt ultrasoft pseudopotentials [9], the family of pseudopotentials
designed by J. Hutter et. al. [10, 11] that include relativistic effects useful
for calculations involving transition elements; and also pseudopotentials for
high-throughput calculations [12].
Fast forward to the second half of the 20th century and a wealth of tech-
niques like DMRG techniques for 1D spin chains [13], Quantum Montecarlo
(QM) methods [14], Many Bodies Perturbation Theory (MBPT) [15], Dy-
2
1.1. OVERVIEW AND MOTIVATION
namical Mean Field Theory (DMFT) [16] and Density Functional Theory
(DFT) [17, 18] where developed to study many electron systems. From a
fundamental perspective, the most important difference between these tech-
niques is the way they incorporate the many body effects. QM, MBPT and
DMFT directly include many body effects whereas DFT only does so in
the independent particle approximation. This may seem as flaw for DFT
methods however, the method is exact for ground state properties, is com-
putationally tractable up to millions of atoms [19] and its output is used as
an input for more accurate methods [20]. This shows that DFT methods
are still a key piece for understanding many electron systems.
Quantum mechanics is without doubt one of the most successful theories
in physics. It is interesting however, that it has at least nine equivalent
formulations [21]. The most common ones are the Schrodinger, Heisenberg
and Feynmann formulations that place the emphasis on wave functions, op-
erators and classical trajectories respectively. In addition to these, there
is also the Wigner formulation or more generally the quantum mechanics
in phase space approach. In broad terms, phase space quantum mechanics
puts the emphasis on complex valued functions of the classical position an
momentum variables. Historically, the Wigner formulation was first intro-
duced in a paper [22] by E. Wigner in 1932 where he was exploring quantum
corrections to thermodynamic equilibrium. Nonetheless, by 1927 H. Weyl
had already given a prescription on how to associate Hilbert space opera-
tors with complex valued functions i.e a quantization rule. This prescription
was used and partially extended by von Neuman in 1932 [23], being in it-
self a precursor of the star product (?) developed by Groenewald [24] in
1946. The star product allows to connect the classical Poisson bracket to
a similar phase space structure known as Moyal bracket that provides the
time evolution of any phase space distribution [25]. Nowadays the idea of
quantum mechanics in phase space has matured [26, 27]. It is known that
there is not a unique way of building a quantum mechanics phase space rep-
resentation, and thus there are many different and equivalent formulations
that differ from each other in the way they associate Hilbert space opera-
3
CHAPTER 1. INTRODUCTION
tors to phase space functions [28]. That is, for example, Wigner function
(symmetric ordering) [22], Mehta function (standard ordering) [29], Kirwood
function (anti standard ordering) [30, 31], Glauber-Sudarshan P and Q func-
tions (normal/antinormal order) [32, 33] and Hussimi function (generalized
order) [34]. Beyond the historical perspective and the formal elegance of
quantum mechanics in phase space in general and of the Wigner function
in particular, the approach has been successfully used in the study of many
areas in physics e.g. Nonlinear dynamics [35, 36], chaotic systems [37, 38],
quantum optics [39] and even ultra cold atoms [40]. More recently, the
Wigner formalism has been useful in the study of noncommutative gauge
theories [41] and it is expected that it can contribute to the growing field
of quantum computing and communication as it is suited to study quantum
decoherence and environmental interactions [23].
Lattice Boltzmann (LB) methods are among the most flexible and efficient
methods for fluid dynamics calculations. The origin of the method can be
traced back to the cellular gas models. These models aim to describe the
complex macroscopic behaviour of fluids using simple microscopic models,
following the idea of Kadanoff [42] that the macroscopic behaviour is mostly
insensitive to the microscopic details. Some of the seminal works are those
of Wolfram [43], d’Humires [44] who developed a 3D model using a 4D hy-
perlattice and Frisch et. al. [45] who discovered the importance of lattice
symmetry in order to recover the Navier-Stokes equations. Even though
these first attempts of discrete fluid dynamics simulations where promising,
their short comings: lack of Galilean invariance, statistical noise, exponen-
tial complexity and spurious invariants where soon evident [46]. The first
answer to these problems particularly to the statistical noise problem was
given by McNamara and Zanetti [47] who proposed to switch from “hard”
occupation numbers to ensemble average populations. Shortly after, a fur-
ther improvement was made by Higuera and Jimenez [48] who proposed to
linearise the kinetic collision term leading to the LBGK model after the
Bhatnagar-Gross-Krook collision operator [49]. Similar ideas where inde-
pendently proposed by Qian [50] and Chen [51]. A few years afterwards, it
4
1.1. OVERVIEW AND MOTIVATION
was shown that the LB models could also be derived from the continuous
Boltzmann equation in the BGK approximation [52, 53] and recently that it
is an exact consequence of the Boltzmann equation [54]. Over the years, it
has become clear that LB methods for fluid dynamics are higly flexible for
introducing new physics in the models as well as for handling curved bound-
aries [55] and complex domains such as blood vessels [56, 57]. Furthermore,
the methods are highly amenable for parallel computing as their parallel
content in the sense of Amdahl law is 90% and its ratio of communica-
tion to computation is relatively low [46]. Demonstrations of these benefits
have been presented as early as 1997 for simulations of turbulent flows [58]
to recent implementations using nearly 300.000 processors [59]. Perhaps as
impressive as the previous accomplishments of LB methods in its original do-
main of application, are its extensions to curved manifolds [60, 61], graphene
physics [62, 63, 64], electrodynamics [65], quantum mechanics [1, 66] and
even relativistic physics [67, 68].
It is clear that Lattice Boltzmann methods are flexible enough to go beyond
fluid dynamics. How flexible are they? would then be the next question
to ask. As it was expressed in Ref. [69], Lattice Kinetic methods could
be generators of broad families of partial differential equations (PDEs) by
“uplifting” the existing theory from its original space, to a phase space
in which the kinetic moments are used to recover the original macroscopic
equations. In this work a few steps into that direction are taken and the idea
of using Lattice Kinetic Methods for Density Functional Theory and Wigner
Equation calculations is explored. This is interesting not only because it will
help to elucidate how flexible Lattice Kinetic Methods really are, but also
because it opens the door to the thought provoking idea of whether or not
certain theories can be considered as kinetic theories with adequate quasi
particles and collisions rules.
1.1.2 Part II
History has it that Brownian motion was discovered by the botanist Robert
Brown while observing pollen grains on water under the microscope. This
5
CHAPTER 1. INTRODUCTION
apparently simple phenomena was soon explained by Einstein [70] and von
Smoluchowski [71]. Shortly after, the idea of harnessing random fluctuations
that is, Brownian motion, was introduced by von Smoluchowski in the form
of a trap door model that aims to emulate a Maxwell demon [72]. The idea
was taken to its modern form by Feynmann, who framed it in the form of a
ratchet and pawl mechanism [73] and explained that it could not work due to
thermal fluctuations. Just like Maxwell demons, ratchet mechanism can not
violate the second law of thermodynamics in equilibrium conditions. How-
ever, transforming random fluctuations in directed motion i.e motion recti-
fication, is possible provided the system is not in equilibrium and that there
is spatial symmetry breaking [74, 75]. Now a days, there are many exam-
ples of systems that function by exploiting Brownian motion e.g. biological
processes [76, 77], particle segregation systems [78, 79, 80, 81] and trans-
port [82, 83, 84]. Also possible is particle motion rectification on sawtooth
channels in the presence of pulsating potential both at a macroscopic [85]
and microscopic levels [79]. Asymmetric objects immersed in granular gases
also display rectification phenomena [86, 87, 88, 89, 90, 91]. As well as active
matter systems from microscopic bacteria to humans [92, 93, 94, 95, 96].
Although these systems rectify the motion of the particles or objects im-
mersed within the particle bath, they still require either a pulsating potential
or active particles. Examples, that in some sense relax these requirements,
include rectification using differentiated noise sources [97, 98] or asymmetric
piston models [99, 100, 101] that show rectification effects, even when work-
ing at a single temperature, provided there is friction and the particle-piston
collisions are different on both sides of the piston. So far there are only few
examples [102, 103] in which the motion of a single particle in a single di-
mension is rectified without external driving forces. In the present work, a
novel example of collective particle motion rectification in two dimensions is
given. We show that the motion of a gas of Lennard-Jones particles can be
rectified without external pulsating potentials only by means of dissipation
and broken spatial asymmetry.
6
1.2. STRUCTURE AND ORGANIZATION
1.2 Structure and organization
This work is organized as follows:
• Part I: Lattice Kinetic Methods for DFT and Wigner function formal-
ism
– Chapter 2 presents the general formulation of the lattice kinetic
method that will be used in chapters 3 and 4. The discussion is of
a general character as the particular specializations for DFT and
Wigner function formalism will be addressed in the corresponding
chapters.
– Chapter 3 starts by introducing the necessary background on den-
sity functional theory (DFT), imaginary time propagation (ITP)
and the kinetic approach to DFT. Afterwards, the second order
correction method is presented and validated with realistic calcu-
lations of atoms and molecules. Finally, the use of pseudopoten-
tials is also addressed.
– Chapter 4 first presents the relevant background on the Wigner
formalism. Afterwards, the Lattice Wigner Method is explained
in detail and validated using quantum harmonic and anharmonic
potentials. Finally, the method is applied to simple one and two
dimensional open driven systems and the viability in three di-
mensions is explored.
• Part II: Thermal gas rectification
– In chapter 5 it is shown that the motion of a thermal gas of
Lennard-Jones particles can be rectified using broken spatial sym-
metry and dissipation. To this end, first the Smoluchowski-Feynmman
ratchet is discussed to show that dissipation and broken spatial
symmetry are required to observe rectification phenomena. Af-
terwards the background on canonical molecular dynamics, which
is the tool used to perform the simulations, is given. Next, the
7
CHAPTER 1. INTRODUCTION
ratchet model is introduced and the dependency of the particle
flux on various system parameters is discussed.
8
Chapter 2
General formulation of lattice
kinetic methods
In this section the general approach to Lattice Kinetic methods is presented.
The discussion is kept general in order to highlight that this family of meth-
ods can go, in principle, beyond the fluid dynamics applications. First it is
shown how the standard Lattice Boltzmann equation is derived, afterwards
the discretization of the velocity space via quadratures will be explained and
then the role of the equilibrium function and source terms is discussed.
2.1 Lattice kinetic methods
2.1.1 From Boltzmann to Lattice Boltzmann
The standard Lattice Boltzmann equation (LBE) is derived from the Boltz-
mann equation (BE) following the formalism presented in [104]. Consider
the Boltzmann equation in the relaxation time approximation
∂f
∂t+ v · ∇f = − 1
τk(f − f eq) + S, (2.1)
where f(x,v, t) is a distribution function in phase space, f eq(x,v, t) is the
equilibrium distribution of the considered system, S(x,v, t) is a general
source term and τk is the kinetic relaxation time.
9
CHAPTER 2. GENERAL FORMULATION OF LATTICE KINETICMETHODS
Observe that Eq. (2.1) can be formally written as an ordinary differential
equationdf
dt+f
τk=
g
τk, (2.2)
where g = f eq + τkS and ddt
= ∂∂t
+ v · ∇ is the time derivative along the
characteristic line v. Eq. (2.2) can be formally integrated in time over the
interval [0, δt] leading to the result
f(x+vδt,v, t+δt) =1
tke−δt/tk
∫ δt
0
et′/tkg(x+vt′,v, t+t′)dt′+e−δt/tkf(x,v, t)
(2.3)
if δt is small enough, g can be linearly approximated in the interval [0, δt] as
g(x + vt′,v, t+ t′) =
(1− t′
δt
)g(x,v, t) +
t′
δtg(x + vδt,v, t+ δt) +O(δt2).
(2.4)
If Eq. (2.4) is substituted in Eq. (2.3) and the integral is solved, the result
is
f(x + vδt,v, t+ δt)− f(x,v, t) = (e−δt/tk − 1)(f(x,v, t)− g(x,v, t))
+
(1 +
tkδt
(eδt/tk − 1)
)(g(x + vδt,v, t+ δt)− g(x,v, t)). (2.5)
Finally if eδt/tk is Taylor expanded and terms of order O(δt2) or higher are
neglected the result is the simultaneous space and time discretization of
Eq. (2.2) namely,
f(x + vδt,v, t+ δt)− f(x,v, t) = −δtτk
(f(x,v, t)− f eq(x,v, t)) + δtS. (2.6)
In the processes of deriving the Lattice Boltzmann equation (Eq. (2.6))
from the Boltzmann equation (Eq. (2.2)) a number of approximations where
made. It could be argued that more sophisticated integration schemes or
softer approximations could lead to additional or different results. This
question was addressed in Ref. [54] and the conclusion is that Eq. (2.6) is an
exact consequence of Eq. (2.1). In different terms, better integration schemes
in Eq. (2.3) and more terms in the Taylor expansion of eδt/tk eventually lead
to the LB equation.
10
2.1. LATTICE KINETIC METHODS
2.1.2 Velocity space discretization
In order to explain the velocity space discretization, it is convenient to first
introduce the velocity moments Π(g)nα1,α1,...,αnof an arbitrary phase space
function, g(x,v, t), as
Π(g)nα1,α2,...,αn=
∫dvvα1vα2 · · · vαng(x,v, t), (2.7)
where n indicates the order of the moment and vα denotes the α component
of the velocity variable. The relevance of the moments steams, as it will be
seen in subsequent chapters, from the fact that these are the quantities that
contain the physical information (particles, energy or momentum densities)
of the problem under consideration.
The velocity space is discretized using quadratures [105] instead of a regular
grid. This approach avoids first, the necessity of fixing a cutoff in veloc-
ity space, which also results in an inaccurate computation of the moments
Eq. (2.7) and second, it provides a better discretization of the ∇ opera-
tor [106], that due to the structure of Eq. (2.1) appears repeatedly in the
concerning problems.
Formally, discretization by quadratures requires that the moments Eq. (2.7)
in the velocity space of the functions f , f eq and S can be exactly calculated
up to a fixed order NΠ. This is achieved with the aid of a quadrature defined
as a finite set vi, ωiNqi=1 of Nq vectors, vi, and corresponding weights, ωi,
that satisfy the relations
Π(n)α1α2...αn
(f) =
∫dvvα1vα2 ...vαnf(x,v, t) (2.8a)
=
Nq∑i
vi,α1vi,α2 ...vi,αnωif(x,vi, t) (2.8b)
=
Nq∑i
vi,α1vi,α2 ...vi,αnfi(x, t) (2.8c)
for n ≤ NΠ. vi,αn denotes the αn component of the i-th velocity vector and
fi(x, t) ≡ ωif(x,vi, t). Similar expressions must also hold for the moments
of f eq and S using the same quadrature vi, ωiNqi=1. The quantities fi, feqi
11
CHAPTER 2. GENERAL FORMULATION OF LATTICE KINETICMETHODS
and Si are respectively known as (discrete) “distributions”, “equilibrium
distributions” and “source distributions”.
The problem of finding quadratures that satisfy Eq. (2.8) can be solved
by observing that they are finite and therefore can not reproduce infinitely
many moments i.e., for a fixed quadrature, Eq. (2.8) can not be satisfied
for an arbitrary large value of n. This imply that only a limited number
of moments of f will be reliably accessible, and that it is not necessary to
work with the full analytical forms of f eq and S. Instead, an expansion in
orthonormal polynomials will be assumed for f eq, S and implicitly for f .
More specifically, given a family of polynomials Pn(v) orthonormal under
the weight function ω(v)∫dvω(v)Pn(v)Pm(v) = δm,n,
f can be represented approximately as
f(x,v, t) ≈ ω(v)
Np∑n
an(x, t)Pn(v), (2.9)
whereNp is the maximum order of the polynomials used in the representation
and the expansion coefficients are given by
an(x, t) =
∫d3vf(x,v, t)Pn(v). (2.10)
The representation Eq. (2.9) simplifies the problem of finding the quadra-
tures, notice that any combination of the form viα1viα1 · · · viαn can be exactly
represented as a linear combination of the Pn,
viα1viα1 · · · viαn =∑m
βmPm(v). (2.11)
If Eq. (2.9) and Eq. (2.11) are substituted in Eq. (2.8a) and Eq. (2.8b), it
can be concluded that the requirement of Eq. (2.8) is then equivalent to
solving the set of constrained algebraic equations
Nq∑i=0
ωiPn(vi)Pm(vi) = δn,m ∀n,m ≤ Np (2.12)
vi ∈ Zd ∀i
wi ≥ 0 ∀i,
12
2.1. LATTICE KINETIC METHODS
for vi, ωi. Technically the constrain vi ∈ Zd is not necessary, however, not
imposing it requires the use of interpolation schemes whenever x + viδt is
not a node of the spatial lattice implied by the space-time discretization of
the Boltzmann equation (Eq. (2.6)). The lattice-Boltzmann equation that
defines the time evolution of the distribution functions fi(x, t) is thus given
by
fi(x + viδt, t+ δt)− fi(x, t) = −δtτk
(fi(x, t)− f eqi (x, t)) + δtSi, (2.13)
where by construction the f eqi and Si are given by
f eqi (x, t) = ωi
Np∑n
aeqn (x, t)Pn(vi) (2.14)
Si(x, t) = ωi
Np∑n
sn(x, t)Pn(vi). (2.15)
A simple, yet effective, approach to construct lattices in any number of
dimension is to find quadrature rules in one dimension and then use ten-
sor products to extend the lattices in two or three dimensions. A family
of orthonormal polynomials that lends itself to such a program is that of
the Hermite polynomials. This is because the Hermite weight function in d
dimensions is the product of d one dimension weight functions. From the
theory of numerical integration [107] it is known that in one dimension a
possible set of quadrature points (vectors) vi is that of the roots of the
polynomial of highest degree for which the quadrature rule is required. Up
to 4th degree Hermite polynomials, this way of proceeding leads to sets of
vectors vi that satisfy the constrain vi ∈ Zd upon a suitable renormaliza-
tion. For polynomials of higher degree the fact that their roots can not, in
general, be expressed as radicals, precludes this way of proceeding. The gen-
eral approach to solve Eq. (2.12) is then heuristic, first the desired discrete
lattice is fixed, and then Eq. (2.12) is solved for the values of the weights.
Typically the weight function and family of polynomials have extra param-
eters, when that is the case Eq. (2.12) becomes a nonlinear problem. The
specification of the generated lattices using the approach just described is
13
CHAPTER 2. GENERAL FORMULATION OF LATTICE KINETICMETHODS
given in Appendix.A.6. It is important to notice that for the particular case
of fluid dynamics, there are a number of ways [108, 109, 110] to system-
atically generate lattices that satisfy Eq. (2.8) and that take into account
additional constrains. Finally the lattice Boltzmann convention of desig-
nating the quadratures or velocity lattices by the scheme DnQm where n
denotes the dimensionality and m the number of velocity vectors or quadra-
ture points will be used throughout.
2.1.3 Equilibrium and source term distributions
It is evident that Eq. (2.13) dictates the time evolution of the distributions
fi, but it is not clear what physical system is actually being modelled. This
information is encoded in the equilibrium distribution, source terms and
also possibly on the boundary conditions. For example, in the traditional
application of Lattice Kinetic methods to fluid dynamics, f eq is a Gaussian
(Maxwellian) distribution such that the moments of fi follow the Navier-
Stokes equations. In general f eqi and S can be physically motivated, but it
is not a requirement that they correspond to a physical equilibrium function
or external source. Instead, they are engineered such that the moments
of fi follow a prescribed dynamic thus defining the system being modelled.
It is this freedom that allows the lattice kinetic methods to be extended
to a variety of situations ranging from Wave equations [111] to Maxwell
equations [65], relativistic hydrodynamics [112] and even fluid dynamics in
curved manifolds [113].
2.1.4 Algorithmic details
Given the explicit expressions for Si and f eqi that define what system is being
modelled and a quadrature vi, ωiNqi=1 that is able to satisfy the necessary
moments (Eq. (2.8)), the way in which a Lattice Kinetic scheme proceeds
is rather simple. The time propagation of the fi is accomplished using
Eq. (2.13) which is evaluated in two stages known as collision and streaming.
14
2.1. LATTICE KINETIC METHODS
In the former one the quantity
f ∗i (x, t) = fi(x, t)−δt
τk(fi(x, t)− f eqi (x, t)) + δtSi(x, t) (2.16)
is evaluated, and in the later it is streamed
fi(x + viδt, t+ δt) = f ∗i (x, t), (2.17)
observe that the collision stage is purely local in space, whereas the stream-
ing stage only involves information flow. In order to repeat the streaming-
collision cycle Si and f eqi need to be updated, in most cases these two
functions are functionals of fi i.e Si = Si[Π0(f),Π1(f), ...Πk(f)] and f eqi =
f eqi [Π0(f),Π1(f), ...Πk(f)] where the different moments are calculated ac-
cording to Eq. (2.8c). The effect of the boundaries can be introduced at
any convenient point, depending on whether they affect the distribution
functions or the moments. The process is illustrated in Fig.2.1
Figure 2.1: Illustration of the Lattice Kinetic algorithm. At every lattice
site there are three different distributions, the distribution function itself,
the equilibrium distribution function and the source term distribution. In
the first step, these distributions are mixed according to Eq. (2.16) leading
to f ∗i . In the second step f ∗i is streamed according to Eq. (2.17). In the
last step Π0(f),Π1(f), ...Πk(f) are calculated and used to update f eqi and Si
before the cycle starts over again.
15
Chapter 3
Lattice kinetic approach to
Density Functional Theory
3.1 Density functional theory and imaginary
time propagation
3.1.1 Density functional theory
Density functional theory can be thought simultaneously as a theory and
method to study the properties of many interacting quantum particles. It
is mostly applied to ground state properties of atomic, molecular and solid
state systems although there are also examples in fields such as cold atomic
gases [114] where the effective particle-particle interaction is short ranged.
From a formal point of view, the problem of finding the properties of a
physical system like a molecule or solid is related to the problem of solving
the stationary Schrodinger equation
Hψi = εiψi, (3.1)
where ψi is the many particle wave function with associated eigenenergy εi.
In the Born approximation the Hamiltonian, H, of a molecular or atomic
17
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
system is generally given by
H = T + Vex + Vee (3.2)
=∑i
− ~2
2m∇2 −
∑i,n
e2zn|Rn − ri|
+1
2
∑i 6=j
e2
|ri − rj|. (3.3)
where m denotes the electron’s mass, ri are the “coordinates” of the i-th
electron, e is the value of the fundamental electric charge and Rn is the
location of the n-th atomic nucleus with z number zn. Observe that the first
term in the r.h.s of Eq. (3.2) is the kinetic energy of the system of electrons,
the second term is related to the external potential due to the atomic nuclei
and the last term is the electron-electron Coulomb interaction.
The difficulty in solving Eq. (3.1) with the Hamiltonian given by Eq. (3.2)
steams from the electron-electron interaction term and from the requirement
that ψ must be antisymmetric on the exchange of any two electrons i.e. the
Pauli exclusion principle.
The DFT approach consist of replacing the original many interacting parti-
cles problem defined by H, with an auxiliary problem HKS of independent
particles that has the same ground state properties as H [115, 116, 117]. The
problem reformulation is carried out using the Hoenberg-Kohn theorems [17]
and the Kohn-Sham ansatz [18].
Let the particle (electron) density associated to an arbitrary state φ be given
by n(r) = 〈φ|n|φ〉 where n =∑
i δ(r − ri) is the particle density operator.
The Hoenberg-Kohn theorems assert that given a physical system described
by a Hamiltonian of the form Eq. (3.2) i) there exists a one to one relation
between the external potential Vex and the ground state particle density n0
that is independent of the electron-electron interaction. And ii) that there
exist an energy functional E[n] that depends on the particle density and
is such that its minimum is the actual ground state energy of the system.
Furthermore, it also holds that the density that minimizes E[n] corresponds
to the ground state density of the system. The Kohn-Sham ansatz starts
from an independent particle system
HKS = T + V KSex , (3.4)
18
3.1. DENSITY FUNCTIONAL THEORY AND IMAGINARY TIMEPROPAGATION
that is assumed to have the same ground state particle density as the original
problem. Under that assumption, the Hoenberg-Kohn theorems show that
the form of V KSex is given by
V KSex = Vex + VH + Vxc, (3.5)
where Vex is the external potential of the original problem and VH is the
Hartree potential. Vxc = δExc[n]δn
is known as the exchange correlation poten-
tial, and is defined via the functional derivative of the exchange correlation
energy functional Exc[n] with respect to the particle density. The important
aspect of Exc[n] is its universality i.e. it is independent of the external po-
tential and therefore the same for all possible problems that have the same
electron-electron interaction.
The crucial consequence of the Kohn-Sham approach is that it is possible to
find the ground state particle density of a many interacting particles problem
by solving the much simpler problem of non interacting particles Eq. (3.4).
The downside however, is that in spite of its universal nature and the fact
that all many particle effects are accounted by Exc[n], only approximations
to it are known. Currently there are many different exchange correlation
potentials. Depending on their degree of sophistication and accuracy, they
are catalogued at different levels on the Jacob Ladder [118] of exchange
correlation potentials. For the present work the BLYP exchange correlation
potential from Becke [119] and Lee et. al [120] will be used throughout.
For practical purposes, once a particular form of Exc[n] has been chosen the
problem is to solve the Kohn-Sham equations[−1
2∇2 + Vex + VH + Vxc
]ψi = εiψi. (3.6)
Since VH and Vxc depend directly on the particle density n, Eq. (3.6) must
be solved in a self consistent way that is, an initial guess n(0) is made,
from there V(0)H and V
(0)xc are calculated and used to solve the Kohn-Sham
equations leading to ψ(0)i . Since the system is non interacting, the ground
state wave function is a single Slater determinant and the updated particle
density can be calculated as n(1) =∑
i |ψ(0)i |2 which can then be used to
iterate the process until |n(i)−n(i+1)| is smaller than a prescribed tolerance.
19
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
3.1.2 Imaginary time propagation
The imaginary time propagation is a technique to find the lowest lying eigen-
functions and eigenenergies of a quantum mechanical problem described by
Eq. (3.1). The method is based on the fact that the time evolution of any
state ϕ, is given by the time dependent Schrodinger equation
i~∂ϕ
∂t= Hϕ, (3.7)
as
ϕ(t) = e−iHt/~ϕ(0) =∑i
aie−iεit/~ψi (3.8)
where
ϕ(0) =∑i
aiψi (3.9)
is the spectral decomposition of ϕ in the eigen basis of H. Observe that if
the time variable is changed as t → −iτ , the imaginary time evolution of
ϕ(0) is given by
ϕ(τ) = eHτ/~ϕ(0) =∑i
aie−εiτ/~ψi. (3.10)
Since the evolution is no longer unitary, the norm of ϕ(0) is not preserved.
In fact, depending on the nature of the spectrum of H, |ϕ(τ)| will either
shrink to zero, or grow without bound. The former case is associated to free
states εi > 0 whereas the later to bound states εi < 0. It is precisely this
behaviour that allows to extract both the eigen functions ψi and the cor-
responding eigen energies εi. To see this, assume without loss of generality
that the system is bounded, for example a molecule, and that the eigenen-
ergies are sorted by increasing value that is, ε0 ≤ ε1 ≤ · · · ≤ 0. In this case
Eq. (3.10) shows that the term a0e−ε0τ/~ψ0 in the spectral expansion of ϕ(τ)
will be the one growing the fastest due to the factor e−ε0τ/~. If after a fixed
time of evolution ϕ(τ) is renormalized, ϕ(τ) → ϕ(τ)/|ϕ(τ)|, then it holds
that ϕ(τ) ≈ ψ0. In the case of an unbounded system a similar argument
holds, but the largest contributing term to ϕ(τ) is the one that decreases
20
3.1. DENSITY FUNCTIONAL THEORY AND IMAGINARY TIMEPROPAGATION
the slowest. Assuming that ϕ(τ) ≈ ψ0, ε0 can be found as follows
ε0 = − 1
δτln
(|ϕ(τ + δτ)||ϕ(τ)|
). (3.11)
Once ϕ0 is found, the exited states and energies can be found by repeating
the process, but starting from the state ϕ(1)(0) = ϕ(0)− 〈ϕ(0)|ψ0〉.
3.1.3 Kinetic approach to DFT
The kinetic approach to density functional theory is based on the observation
that the time dependent Kohn-Sham equations in imaginary time can be
recovered as a special macroscopic limit of the Boltzmann equation in the
BGK approximation [49]. The connection is established in two steps, the
first one is to observe that upon a Wick rotation the time dependent Kohn-
Sham equation is formally equivalent to a diffusion equation. The second
step is to show that an equilibrium distribution function, f eq, and source
term S can be tailored such that the moments of the distribution function
follow a diffusive dynamic consistent with that of the Wick rotated Kohn-
Sham equations.
The first step readily leads to
∂ϕ
∂τ=
~2m∇2ϕ− 1
~V KSϕ, (3.12)
where the T + V form of the Hamiltonian was used. Notice that formally,
Eq. (3.12) is a diffusion equation for the field ϕ, with a diffusion constant~
2mand a source/sink term given by 1
~VKSϕ.
For clarity of explanation, the result of the second step will be first presented
and afterwards it will be shown how it yields the desired diffusive behaviour.
To fix the notation let the α1α2...αn component of the n-th moment of the
distribution function, equilibrium distribution function, and source term be
21
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
respectively defined as
Π(n)α1α2...αn
=
∫dvvα1vα2 ...vαnf(x,v, t), (3.13a)
Π(n)eqα1α2...αn
=
∫dvvα1vα2 ...vαnf
eq(x,v, t), (3.13b)
Σ(n)α1α2...αn
=
∫dvvα1vα2 ...vαnS(x,v, t), (3.13c)
where vα is the α component of the phase space velocity vector v.
If the moments of the equilibrium distribution function and source term are
chosen as
Π(0)eq = Π(0), (3.14a)
Π(1)eq = 0, (3.14b)
Π(2)eqij = C2
sΠ(0)δij, (3.14c)
Π(n)eq = 0 for n > 2, (3.14d)
Σ0 = S, (3.14e)
Σ(n)eq = 0 for n > 0, (3.14f)
where C2s is a characteristic speed of the system and S is a function of space
and time, then the 0-th moment of the distribution function evolves in time
according to the diffusion equation
∂Π0
∂t= τkC
2s∇2Π0 + S. (3.15)
To see how the moments definition Eq. (3.14) lead to Eq. (3.15), the tool
of choice is the Chapman-Enskog expansion [121]. Let the the distribution
function and time derivative be expanded as
f = f (0) + εf (1) + ε2f (2) + · · · , (3.16)
∂
∂t= ε
∂
∂t1+ ε2
∂
∂t2+ · · · , (3.17)
and let the spatial derivative and source term be rescaled as ∇ = ε∇1 and
S = εS(1), where ε is regarded as a small quantity. Substituting these
22
3.1. DENSITY FUNCTIONAL THEORY AND IMAGINARY TIMEPROPAGATION
relations in Eq. (2.1) and equating terms of equal order in ε the following
set of equations are found
f (0) = f eq, (3.18)
∂t1f(0) + v · ∇1f
(0) = − 1
τkf (1) + S(1), (3.19)
∂t1f(1) + ∂t2f
(0) + v · ∇1f(1) = − 1
τkf (2). (3.20)
Taking the 0-th moment of Eq. (3.19) and (3.20) we found
∂t1Π(0,0) +∇1 · Π(1,0) = − 1
τkΠ(0,1) + Σ(0,1), (3.21)
∂t1Π(0,1) + ∂t2Π(0,0) +∇1 · Π(1,1) =1
τkΠ(0,2). (3.22)
Where Π(i,j) is the i-th moment of f (j) and Σ(0,1) is the 0-th moment of the
rescaled source term. Eqs. (3.16), (3.18) and the constraint Eq. (3.14a)
implies that the previous equations simplify as
∂t1Π(0) +∇1 · Π(1,0) = Σ(0,1), (3.23a)
∂t2Π(0) +∇1 · Π(1,1) = 0. (3.23b)
Eq. (3.18) and constraint Eq. (3.14b) implies that Π(1,0) = 0. Thus Eq. (3.23)
can be written as
∂t1Π(0) = Σ(0,1), (3.24a)
∂t2Π(0) +∇1 · Π(1,1) = 0. (3.24b)
Π(1,1) can be calculated by taking the first moment of Eq. (3.19), due to the
constraints Eq. (3.14b) and (3.14f) only the term v · ∇1f(0) will contribute,
explicitly
Π(1,1)k = −τk
∫vkv · ∇1f
(0)d3v
= −τk∂i∫vkvif
(0)d3v
= −τk∂iΠ(0)i,j . (3.25)
23
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
∇1 · Π(1,1) is then calculated as
∇1 · Π(1,1) = ∂iΠ(1,1)i (3.26)
= −τk∂i∂kΠ(0)i,k (3.27)
= −τkC2s∇1 · ∇1Π(0) (3.28)
= −τkC2s∇2
1Π(0), (3.29)
where the constraint Eq. (3.14c) was used. Eq. (3.24b) then reads
∂t2Π(0) − τkC2s∇2
1Π(0) = 0. (3.30)
Finally, Eq. (3.15) is obtained by multiplying Eq. (3.24a) by ε and Eq. (3.30)
by ε2 and adding them taking into account the scaling of the spatial deriva-
tive and source term, as well as the expansion of the temporal derivative.
If the identifications Π0 ≡ ϕ, τkC2s ≡ ~
2mand S ≡ −V KSex
~ ϕ are made, then
Eq. (3.15) can be rewritten as
∂ϕ
∂t=
~2m∇2ϕ− V KS
ex
~ϕ, (3.31)
which is the Wick rotated time dependent Kohn-Sham equation Eq. (3.12).
At this point is important to notice that in the kinetic approach to DFT the
relation between the quantum mechanical wave functions and the kinetic
distribution functions is not direct, instead the quantum mechanical wave
functions emerge as the zeroth moment of the distribution function.
From the previous discussion on imaginary time propagation, it follows that
for any initial condition ϕ(0), that has a non vanishing projection on the
ground state ψ0, the imaginary time evolution guarantees that as the time
increases and the wave function is renormalized, only the ψ0 contribution is
obtained. It is clear then, that by sequentially removing lower lying eigen-
states from the initial condition ϕ(0) all possible states will be eventually
found. Therefore, the Lattice Kinetic approach contains the same physical
information as the Schrodinger equation.
Even though all the physical information regarding a given quantum system
is contained in Eq. (2.1) and Eq. (3.14), for computational purposes it will
24
3.1. DENSITY FUNCTIONAL THEORY AND IMAGINARY TIMEPROPAGATION
prove convenient to make it explicit. This is done by introducing an ex-
tended kinetic model in which each Kohn-Sham orbital ψl l = 1, 2, . . . , N is
associated to its own distribution function fl, and the different distributions
interact with each other via a dynamical orthonormalization potential Wl.
That is, each of the fl evolves according to
∂fl∂t
+ v · ∇fl = − 1
τk(fl − f eql ) + S +
1
τkWl, (3.32)
where the moments of Wl are
Ω(0) = −∑i<l
〈ψl|ψi〉〈ψi|ψi〉
ψi, (3.33a)
Ω(1) = 0, (3.33b)
Ω(2) = −τkC2s
∑i<l
〈ψl|ψi〉〈ψi|ψi〉
ψi, (3.33c)
Ω(n) = 0 for n > 2. (3.33d)
Since the structure of the moments of Wl is the same as that of f eql , each
Π(0)l will be given by
Π(0)l = ψl −
∑i<l
〈ψl|ψi〉〈ψi|ψi〉
ψi. (3.34)
Thus, Π(0)k has no contributions from ψl with l < k. Since ϕl → ψl before
ϕk → ψk for l < k it follows that effectively ϕk has no components along
any of the ψl eigenstates and thus the next lowest available eigenstate is the
one that is going to be selected by the imaginary time evolution. Notice
also that as the different orbitals start converging, the effect of Wl becomes
weaker due to the orthonormality of the wave functions and once the different
orbitals have converged it plays no further role. In other terms, Wl is only
used to drive the different fl in such a way that they converge to different
Kohn-Sham orbitals. Finally, the relation between the kinetic approach to
DFT and the Kohn-Sham equations is summarized in Fig.3.1
25
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
Figure 3.1: Diagram of the relation between the time dependent Kohn-Sham
equations and the Kinetic approach to DFT via the diffusion equation. From
the time dependent Kohn-Sham equation the diffusion equation is reached
by means of the Wick rotation. Starting from the Boltzmann equation, it
can be reached by the right definition of the moments of the equilibrium
function, Πeq, and source term Σ.
3.2 Lattice Kinetic model for DFT
Mendoza et.al proposed a Lattice Kinetic model that evolves according to
Eq. (2.13) and satisfies both the explicit constraints Eq. (3.14) on the equi-
librium and source term distribution functions and the requirement that
the different Kohn-Sham orbitals need to be orthonormal. In the discrete
setting of that model, it can be show that the 0-th order moment (Kohn-
Sham orbital), of the lattice distribution function fil evolves according to
the diffusion equation
∂Π(0)
∂t= δt
(τk −
1
2
)C2s∇2Π(0) + S +O(δt), (3.35)
which is equivalent to Eq. (3.31), up to terms of order O(δt), if the identi-
fications Π0 ≡ ψ, δt(τk − 1
2
)C2s ≡ ~
2mand S ≡ −V
~ ψ are made. The extra
“−1/2” term in the identification of ~/2m compared to that of Eq. (3.15) is
due to the spacial discretization and is well known in the Lattice Boltzmann
community.
3.2.1 Formal correction of discrete lattice effects
In the following it is shown how to improve the proposal of Mendoza et.al
using a semi implicit correction term. First the Lattice Kinetic model is
26
3.2. LATTICE KINETIC MODEL FOR DFT
formally improved by the introduction of a correction term. Afterwards,
the equilibrium and source term distributions for the model are explicitly
constructed and finally the semi implicit implementation of the correction
term is explained.
The second order Lattice Kinetic model is defined as
fi(x+viδt, t+δt)−fi(x, t) = − 1
τk(fi(x, t)−f eqi (x, t))+δtSi+
δt2
2DiSi, (3.36)
where Di = ∂∂t
+vi ·∇ and δt2
2DiSi is the forcing correction term that cancels
spurious discretization effects. In simple terms, the addition of δt2
2DiSi works
because it exactly compensates a similar term that manifests itself at the
second scale level in the Chapman-Enskog expansion of the standard Lattice
Boltzmann equation with source term Eq. (2.13). To see this in more detail
first consider the Taylor expansion of the l.h.s of Eq. (3.36)
δtDifi +δt2
2D2i fi = − 1
τk(fi − f eqi ) + δtSi +
δt2
2DiSi, (3.37)
observe that the convective terms of the original Boltzmann equation are
already present in the first order terms of Eq. (3.37), while the second order
ones appear due to the fact that the Lattice-Boltzmann equation is a discrete
approximation. The diffusion equation is recovered by following a similar
path as that used in the continuous Boltzmann equation case. That is, a
multi-scale expansion of the distribution functions and time derivatives in
a small parameter ε (that in fluids dynamics plays the role of a Knudsen
number) is assumed
fi = f(0)i + εf
(1)i + ε2f
(2)i + · · · , (3.38)
∂
∂t= ε
∂
∂t1+ ε2
∂
∂t2+ · · · , (3.39)
whereas space derivatives and source term are rescaled as ∇ = ε∇1 and
Si = εS(1)i respectively.
Replacing Eq. (3.38) and Eq. (3.39) into Eq. (3.37) and collecting terms of
27
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
equal order in ε, yields the result
O(ε0) : f(0)i = f eqi , (3.40)
O(ε1) : D1ifeqi = − 1
τkδtf
(1)i + S
(1)i , (3.41)
O(ε2) :∂f eqi∂t2
+D1if(1)i +
δt
2D2
1ifeqi = − 1
τkδtf (2) +
δt
2D1iS
(1)i . (3.42)
Eq. (3.41) is further substituted in the l.h.s of Eq. (3.42) and the result is
summed over all discrete velocities leading to the relation
∂Π0
∂t2+
(1− 1
2τk
)∇1 ·
∑vif
(1)i = 0, (3.43)
where the moments definition of f eq and S Eq. (3.14) were used. The first
moment of the first correction to the distribution function,∑
vif(1)i , is eval-
uated by taking the product of Eq. (3.41) with vi and summing over all
velocities. Substituting this result back in Eq. (3.43) yields
∂Π(0)
∂t2= δt
(τk −
1
2
)C2s∇2
1Π(0). (3.44)
In a similar manner if Eq. (3.41) is summed over all velocities the result is
∂Π(0)
∂t1= S(1). (3.45)
Finally when Eq. (3.44) is multiplied by ε2 and added to Eq. (3.45) multiplied
by ε it is found that Π(0) evolves according to
∂Π(0)
∂t= δt
(τk −
1
2
)C2s∇2Π(0) + S +O(δt2). (3.46)
The result is that the modified model (Eq. (3.36)) together with the same
moments definition (Eq. (3.14)) as in the continuous case lead to a second
order improvement in the recovered equation for the zeroth order moment of
the distribution function, that is for the Kohn-Sham orbitals. The inclusion
of the term δt2
2DiSi in Eq. (3.36) is not intuitive, however it can be seen that
if it had not been added then the additional term δt2∂S(1)
∂t1would be present
in Eq. (3.43) and then propagated to Eq. (3.46). It is important to notice
28
3.2. LATTICE KINETIC MODEL FOR DFT
that the inclusion of the correction term is only possible because, as shown
before, it does not change the macroscopic limit of the Lattice-Boltzmann
equation i.e. the zeroth moment of the distribution function still follows a
diffusive dynamics.
The extension to a system of coupled equations for many orbitals is given
by
fil(x + viδt, t+ δt)− fil(x, t) =
− 1
τk(fil(x, t)− f eqik (x, t)) +
δt2
2DiSil + δtSil +
1
τkWil, (3.47)
where the index l is associated to the l-th orbital.
3.2.2 Equilibrium and source distributions construc-
tion
Up to this point, the equilibrium and source term distribution functions are
only implicitly known in terms of their moments structure Eq. (3.14). How-
ever, to have a practical Lattice Kinetic scheme their explicit form is needed.
Following the discussion of Sec.2.1.2, the form Eq. (2.9) is assumed for the
equilibrium distribution function. Then, using the quadrature definition of
the moments Eq. (2.8c) and the explicit moment definition Eq. (3.14) the
set of linear equations
Π(0) =∑i
∑n
ωiaeqn Pn(vi)
Π(1)αi
= 0 =∑i
vαi∑n
ωiaeqn Pn(vi) (3.48)
Π(2)αiαj
= C2sΠ(0)δij =
∑i
vαivαj∑n
ωiaeqn Pn(vi)
Π(k)αij ···αk = 0 =
∑i
vαi · · ·vαk∑n
ωiaeqn Pn(vi) ∀k > 2
for the unknown coefficients an is defined. Taking the Pn as Hermite poly-
nomials and using the D3Q19 and D3Q111 lattices, the explicit forms of f eqi
29
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
and Si are
f(D3Q19)i (x, t) = Π0(x, t)wi
(1 +
1
2C4s
(C2s −D)(3C2
s − vi · vi)), (3.49a)
S(D3Q19)i (x, t) = S(x, t)wi
(1 +
1
2C4s
(C2s )(3C2
s − vi · vi)), (3.49b)
f(D3Q111)i (x, t) = Π0(x, t)wi
(1 +
1
2C4s
(C2s −D)(3C2
s − vi · vi)+ (3.49c)
1
8C2s
(C2s − 2D)(15C4
s − 10C2svi · vi + (vi · vi)2)
)(3.49d)
S(D3Q111)i (x, t) = Swi
(1 +
1
2C4s
(C2s )(3C2
s − vi · vi)+ (3.49e)
1
8C2s
(C2s − 2D)(15C4
s − 10C2svi · vi + (vi · vi)2)
), (3.49f)
where the diffusion constant is given by D = (τk− 12)C2
s = ~2m
and the value
of C2s depends on the lattice.
3.2.3 Semi implicit correction
From a computational perspective the implementation of Eq. (2.13) (model
1) requires no special discussion as it conforms to standard Lattice Boltz-
mann schemes. However, in our approach, (model 2) Eq. (3.36), there are
various ways to implement the correction term DiSi. These can be explicit
DiSi =1
δt(Si(x, t)− Si(x− viδt, t− δt)) , (3.50a)
DiSi =1
δt(Si(x + viδt, t)− Si(x, t− δt)) , (3.50b)
or implicit
DiSi =1
δt(Si(x + viδt, t+ δt)− Si(x, t)) . (3.51)
Given that explicit implementations do not require solving a system of
equations at every iteration of the algorithm, we performed tests using
Eqs. (3.50a) and (3.50b), in both cases we found that the procedure was
numerically unstable leading to wild oscillations of the measured quantities.
30
3.3. APPLICATION OF THE IMPROVED LATTICE KINETICAPPROACH
To use the implicit form of DiSi and avoid the necessity of solving systems
of equations, Eq. (3.51) was approximated using the formally known imag-
inary time evolution of the different orbitals. That is, for the n-th orbital
DiSi,n = DiV~ ϕn, and its discretized version is given by
DiSi,n =1
δt
(V
~ϕn(x + viδt, t+ δt)− V
~ϕn(x, t)
). (3.52)
An approximation of Eq. (3.52) can be obtained if ϕn(x, t + δt) can be
estimated. From the imaginary time evolution it is known that
ϕn(x, t) =∑j≥n
cjψje−εjt
~ , (3.53)
where ψj and εj are the eigenfunctions and eigenenergies of the Kohn-Sham
Hamiltonian and cj are the projection coefficients of ϕ in the basis ψj.Therefore, we consider the approximation
ϕn(x, t+ δt) ≈ ϕn(x, t)e−εnδt~ . (3.54)
Notice that this approximation improves after every iteration and is exact
once the steady state has been reached, this follows from the time projection
technique that progressively drives all the cj → 0 for j 6= n.
Finally it is worth noticing that the actual implementation of the correction
term does not add any extra complexity to the scheme. It corresponds to an
extra scalar-matrix-vector multiplication of the same kind used to calculate
the original source term.
3.3 Application of the improved lattice ki-
netic approach
3.3.1 Model comparison
The improved scheme was used to calculate the exchange and correlation
energies of H, He, Be and Ne atoms as well as the bond lengths of H2 and
LiH dimers. We compared it with the model 1 using a D3Q19 lattice and
31
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
the BLYP exchange correlation potential [119, 120] . The physical length
of the simulation box is given by Lp = L∆x where L is the number of grid
points in one direction, ∆x is the distance between two successive sites and
the resolution of the system is defined as ∆x−1.
Results from the He atom and H2 molecule are shown in Fig.3.2 a,c). Both
models show that for a fixed resolution ∆x−1, as the number of lattice sites
i.e the physical size of the simulation box increases, the value of the measured
quantities tend to converge to a limiting value, and as the resolution level is
improved, the limiting value approaches the expected BLYP values Ref. [20,
122]. These two behaviours are consistent with the fact that as the physical
system size and resolution increase, the boundary effects are reduced and the
system better approximates an atom or molecule in free space. Furthermore,
the limiting values of model 2 are closer to the expected BLYP values than
those of model 1 for a fixed ∆x, and as the resolution improves both models
tend to agree.
The degree to which model 2 is more accurate than model 1 with respect
to the expected BLYP values depends on the considered atom or molecule
and measured quantity (Ex, Ec, bond length, etc). For example, Fig.3.2
b,d) show respectively the relative error, ∆Ex and ∆Ec, of the exchange
and correlation energies as a function of ∆x−1 when the system size is fixed
at L = 150. In the case of Ex, ∆Ex ∝ ∆x1.5 for the first model and
∆Ex ∝ ∆x1.6 for the second. In contrast the behaviour of ∆Ec is non
monotonic. It is worth noting that ∆Ex ranges from 4% to less than 1%
whereas ∆Ec is always smaller than 0.3%.At this point the difference between
the two models seems small, however, this is due to the fact that Ex and
Ec are the integrals of non trivial functions of the density and the density
gradient, where the later has to be numerically calculated. To better observe
the difference between both models, the ground state energy of the H atom,
which only requires the norm of the wave function at two consecutive time
steps, was calculated. Its relative error ∆EH as a function of the resolution
is shown in Fig.3.3, where it can be observed that ∆EH in the second model
(m2) is one order of magnitude smaller than in the first model (m1). That is
32
3.3. APPLICATION OF THE IMPROVED LATTICE KINETICAPPROACH
Figure 3.2: Model comparison: The top and medium left panels are respec-
tively the calculated exchange and correlation energies of the He atom as
a function of system size for three different resolutions using models 1 and
2. Top and medium right panels show the relative error as a function of
∆x−1 for a system size of L = 150. The bottom panel shows the equilibrium
length of the H2 molecule. The solid blue line is the DFT result using the
BLYP functional reported in Ref. [20, 122]
33
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
0.5 1.0 1.5 2.0 2.5 3.0Log(∆x−1)
−10
−9
−8
−7
−6
−5
−4
Log(|E
−Eex|)
m1m2
Figure 3.3: Relative error of the ground state energy of the H atom as
a function of the resolution for m1 and m2. The circles correspond to the
simulation data and the dashed lines are the linear fits. In both cases ∆EH ∝∆x1.6
Atom −Ex −Ex BLYP −Ec −Ec BLYP
H -0.301 -0.301
He -1.0197 -1.0255 -0.0437 -0.0438
Be -2.6741 -2.6578 -0.0965 -0.0945
Ne -12.0532 -12.1378 -0.3827 -0.3835
Molecule Bond length Bond length (BLYP)
H2 1.3867 1.4000
LiH 3.005 3.016
Table 3.1: Exchange and correlation energies of different atoms, and bond
lengths of different molecules calculated using model 2 with D3Q19 lattice
compared to the known BLYP values.
an indication that m2 indeed solves the kinetic Kohn-Sham equation more
accurately than m1. Finally the results for simulations of the other systems
are summarized in Table.3.1.
34
3.3. APPLICATION OF THE IMPROVED LATTICE KINETICAPPROACH
3.3.2 Ethane molecule test
As a test of the proposed model, the C2H6 (ethane) molecule was simulated.
The carbon atoms were initially located such that their center of mass was
in the center of the simulation box and they were aligned along the z axis.
The H atoms were randomly located, three of them closer to the upper
carbon atom, and the remaining ones closer to lower carbon atom (Fig 3.4
left). This set up mimics the common scenario in which there is only partial
information available. The final configuration, obtained after 3.6 days of
run time on a single core, is shown in Fig 3.4 where the qualitatively cor-
rect shape of the ethane molecule and electronic density distribution can be
observed, compared to the initial configuration. The relative errors of the
bond lengths and angles with respect to the expected ones [122] are 1.3%
for the C − C bond length, a mean relative error of 2.1% for the H − C
bond length and a 7% for the H − C − H angles. Except for the angles,
the accuracy is comparable to that of a Carr-Parinello Molecular dynam-
ics (CPMD) simulation performed with identical initial conditions using a
wavefunction cutoff of 100Ry. It achieves a 2.7% C − C bond length error,
2.0% mean H − C bond length error and a 0.05% mean H − C −H angle
error. The CPMD simulation took about three hours, which is a small frac-
tion of the computational time spent by our model. However, CPMD uses
pseudopotentials while our model considers the bare Coulomb potential.
3.3.3 Pseudopotentials
Pseudopotentials are a way to reduce the computational cost of atomistic
simulations that works under the approximation that core electrons are
mostly inert [7] and play a minimal role in most of the chemistry. Although
pseudopotentials are designed to be highly accurate and transferable, it is
not always clear, a priori, how do they couple to different simulation meth-
ods. For instance pseudopotentials are known to be problematic or not
directly applicable in diffusion Monte Carlo and Green functions approaches
[123, 124, 125].
In order to assess how pseudopotentials couple to our method, tests were
35
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
Figure 3.4: (Color online). On the left side, the initial configuration of the
atoms that conform the ethane molecule is depicted. On the right side, the
final configuration is shown along with the electron density, the regions of
high(low) electron density are indicated by red(blue) surfaces
performed using the dual-space Gaussian pseudopotentials (DSGPP) intro-
duced in Ref. [10]. The DSGPP were chosen because they are optimized
for the BLYP exchange correlation potential used in this work and because
their real space representation is compatible with the real space nature of
our method.
The first non trivial example that includes both local and non local contri-
butions of the DSGPP is the BH molecule where the two inner electrons of
the boron atom are neglected. In this case the B atom is described by the
pseudopotential Vpp = Vloc +Hnonloc where
Vloc(r) =−Zione|r−RB|
erf
(|r−RB|√
2rloc
)+ exp
(|r−RB|2
2r2loc
)(3.55)
×[C1 + C2
|r−RB|2
r2loc
+ C3|r−RB|4
r4loc
+ C4|r−RB|6
r6loc
],
and
Hnonloc(r, r′) =
2∑i=1
Y0,0(r)p0i (r)h0
i p0i (r′)Y ∗0,0(r′) (3.56)
+∑m
Y1,m(r)p11(r)h1
1p11(r′)Y ∗1,m(r′).
36
3.3. APPLICATION OF THE IMPROVED LATTICE KINETICAPPROACH
The values of the constants Ci i = 1, ..4, rloc and hi as well as the functional
form of the projectors p0i (r),p1
1(r) can be found in Ref. [10]. The DSGPP
for boron was implemented and used to calculate the bond length of the
BH molecule. It was found that when pseudopotentials are used within our
approach the scheme becomes unstable (black dashed line Fig.3.5). The
instabilities were partially controlled by artificially resetting the wave func-
tions and electron density to their initial values after a fixed number of it-
erations while keeping the current position of the ions (blue dashed-dot line
Fig.3.5), but eventually instabilities arise. Different initial conditions and
resolution levels also suffer from instabilities (green doted line Fig.3.5). The
reason why our approach becomes unstable may be related to the overall
nonlinear nature of the system and the nonlocal component of the pseu-
dopotential that requires the evaluation of projection integrals of the form∫p0i (r′)Y ∗l,m(r′)ψ(r′), that may not be sufficiently resolved due to the fact
that ψ(r′) is only known at a limited number of lattice points.
10000 20000 30000 40000 50000Iterations
2.0
2.5
3.0
3.5
Bond
leng
tha0
BH Lx=10∆x = 0.1BH Lx=10∆x = 0.1BH Lx=10∆x = 0.06BLYP value
Figure 3.5: (Color online). BH bond length calculated using pseudopoten-
tials. The different lines show that the use of pseudopotentials within our
approach leads to numerical instabilities.
37
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
3.3.4 Lattice performance
The performance and accuracy of model 2 using either the D3Q19 or D3Q111
lattice was investigated by simulating the He atom as well as H2 and LiH
dimers. The generally observed trend is that for small resolutions the re-
sults obtained using both lattices differ, but as the resolution increases the
difference is reduced and the results converge.
Although both lattices lead to practically the same results, the D3Q111
lattice allows to chose a higher value for the diffusivity ~/m, that can be
used to control the convergence rate of the procedure. As an example, the H2
molecule was simulated using the D3Q111 lattice for three different values
of ~/m Fig.3.6(inset). It can be seen that in all cases the system converges
to the same value of the bond length, but for ~/m = 1 the approach is faster
than for ~/m = 1/3 or ~/m = 0.69. The D3Q111 lattice has almost six
times more velocity vectors than D3Q19, and the time of a single iteration
using the D3Q111 lattice was measured to be about 1.5 times longer than
that of the D3Q19 lattice. After accounting for this, the comparison between
the speed of convergence using both lattices for two different system sizes is
presented in Fig.3.6. It can be observed that the use of the D3Q111 lattice
allows for a faster convergence. However, for large resolutions, since both
lattices lead to the same accuracy, using D3Q111 presents no advantage in
terms of computational time. For instance for the Be atom we found that
the D3Q111 lattice with ~/m = 1 converges equally fast as D3Q19.
3.4 Summary
In this section a new and more accurate Lattice Boltzmann scheme to solve
the kinetic Kohn-Sham equations has been introduced and validated. The
scheme uses a novel way of implementing a semi implicit second order cor-
rection to the forcing term, that makes use of the known asymptotic be-
haviour of the simulated orbitals. This approach avoids the instabilities of
the explicit implementations and the computational load of solving implicit
systems of equations.
38
3.4. SUMMARY
0 200000 400000 600000Iterations
1.0
1.5
2.0
2.5
3.0
3.5
Bond
Leng
th(a
.u.)
Functionalh/m = 1h/m = cs2 = 0.69h/m = 1/3
0 1000 2000 3000 4000 5000 6000 7000Time/t0
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
Bond
Leng
th(a
.u.)
FunctionalD3Q19 L=80D3Q19 L=100D3Q111 L=80 h/m = 1D3Q111 L=100 h/m = 1
Figure 3.6: Effect of the numerical value of the diffusivity ~/m on the conver-
gence speed using the D3Q111 lattice in the simulation of the H2 molecule.
The use of pseudopotentials within our approach requires further work to
eliminate the associated instabilities and computational demands, not only
for the tested case, but also for general pseudopotentials. Possible ap-
proaches in that direction include subgrid refinements.
The results of the ethane molecule simulation show that our method can
39
CHAPTER 3. LATTICE KINETIC APPROACH TO DENSITYFUNCTIONAL THEORY
reproduce the bond lengths of complex molecules, but that further work
is required to achieve an overall performance similar to that of established
methods such as CPMD, including the full integration with pseudopoten-
tials. It was also confirmed that the D3Q19 and D3Q111 lattices lead to the
same results for high enough resolution, giving an advantage to the D3Q19
lattice in terms of computational resources.
40
Chapter 4
Lattice Wigner Model
4.1 Wigner formalism
The Wigner formalism is a kinetic formulation of quantum mechanics phys-
ically equivalent to the Schrodinger representation [21]. It is however, very
different, as it treats both position and momenta as independent real vari-
ables instead of operators, similar to classical Hamiltonian dynamics and
kinetic theory.
The Wigner function is defined as
W (q,p, t) =1
(2π~)dW(ρ)
=1
(2π~)d
∫ ∞∞
dyρ(q− y/2,q + y/2)eip·y/~, (4.1)
where ρ(x,x′) is the real space representation of the density matrix of the
quantum system under consideration, d is the dimensionality of the sys-
tem and the Weyl transform, W(·), of a quantum mechanical operator O is
defined as:
O(q,p) =W(O) =
∫eip·y/~〈q− y/2|O|q + y/2〉dy. (4.2)
In general, W (q,p) is real and normalised in phase-space, i.e∫dp dqW (p,q, t) = 1. (4.3)
41
CHAPTER 4. LATTICE WIGNER MODEL
However, due to quantum interference effects, it is not positive semi definite,
and consequently, it cannot be regarded as a proper distribution function,
but rather as a quasi-distribution.
Expectation values of a physical observable O are obtained through the
prescription:
tr(ρO) =
∫dpdqO(q,p)W (p,q, t). (4.4)
The moments of the Wigner function with respect to the momentum variable
are defined as
Π(W )nα1,...,αn=
∫dppα1 ...pαnW (q,p), (4.5)
where n indicates the order of the moment and pαi denotes the αi component
of the momentum variable. The first two moments Π(W )0 and Π(W )1αi
can
be identified with the particle density ρ(x,x) and momentum density respec-
tively, whereas the sum of the diagonal terms of Π(W )2αiαj
is proportional
to the kinetic energy density.
The time evolution of the Wigner function can be obtained as the Weyl
transform of the Liouville-von Neumann equation, namely:
∂ρ
∂t=
1
i~[H, ρ],
where H = p2
2m+ V (x) is the Hamiltonian of the system.
The result (see Appendix A.1) is known as the Wigner equation and it reads
as follows:∂W
∂t+
p
m· ∇W + Θ[V ]W = 0, (4.6)
where Θ[V ]W can be written as
Θ[V ]W =
∫ ∞−∞
δ[V ](q,p− p′)W (q,p′)dp′, (4.7)
δ[V ](q, p) =
i
2π~2
∫ ∞−∞
(V (q− y/2)− V (q + y/2))eiy·p/~dy, (4.8)
or alternatively
Θ[V ]W = −∑|s|∈Nodd
(~2i
)|s|−11
s!
∂sV
∂qs∂sW
∂ps, (4.9)
42
4.2. LATTICE WIGNER MODEL
where s is a vector of non negative integers, |s| =∑d
i=1 si,∂s
∂as≡ Πd
i=1∂si∂asi
for a = q, p. Finally, it is important to notice that the different terms of
the Wigner equation Eq. (4.6) can be linked to the different terms of the
Liouville-von Neumann equation.
The convective term arises solely from the kinetic energy term in the Hamil-
tonian, whereas the force term Θ[V ]W originates from the potential energy
contribution. To be noted that spatial derivatives of the potential at various
orders couple to corresponding derivatives in momentum space, multiplied
by the corresponding power of the Planck’s constant ~. Such higher-order
terms are responsible for the “quantumness” of the Wigner representation
and the occurrence of negative values due to quantum interference effects.
4.2 Lattice Wigner model
In this section we introduce the lattice Wigner scheme. For this, it is con-
venient to work in the dimensionless form of Eq. (4.6). Upon the change
of variables q → l0x, p → m(l0/t0)v, t → t0τ where x, v, τ are the new
dimensionless variables and l0, t0 are characteristic length and time scales,
respectively, Eq. (4.6) and Eq. (4.9) can be written as:
∂W
∂τ+ v · ∇xW + Θ[V ]W = 0, (4.10)
and
Θ[V ]W = −∑|s|∈Nodd
(H
2i
)|s|−11
s!
∂sV
∂xs∂sW
∂vs, (4.11)
where H = ~t0ml20
, V = Vm(l0/t0)2 are the dimensionless reduced Planck constant
and potential terms, respectively.
Using the formalism introduced in Ch.2 the lattice kinetic discretization of
Eq. (4.10) is given by
Wi(x + viδt, t+ δt)− Wi(x, t) = −δt(Θ[V ]W )i. (4.12)
Following the lattice Boltzmann nomenclature, the Wi and (Θ[V ]W )i are
termed respectively “distributions” and “source distributions”. Observe that
43
CHAPTER 4. LATTICE WIGNER MODEL
the time evolution of the Wi distributions is given by Eq. (4.12) and that at
every spatial lattice point x there are Nq distributions (see sec. 2.1.2). From
these distributions, the moments such as density Π(W )0 = ρ or momentum
density Π(W )1α = ρuα, can be calculated at every time step using Eq. (2.8).
It is important to notice that, although a discretization by quadratures re-
quires no cutoff in velocity space, it does nonetheless involve a ceiling on the
highest moment for which Eq. (2.8) holds. In other words, it is a truncation
in discrete momentum space.
It is in principle possible to use Eq. (4.12) to track the time evolution of the
moments of the Wigner function under the action of a specified potential.
However, it was shown in Ref. [126] that the resulting structure of the forcing
term leads to numerical instabilities. To address this problem, the lattice
Wigner model is introduced as
Wi(x + viδt, t+ δt)− Wi(x, t) = δtΩi + δtSi +δt2
2DiSi,
Ωi = − 1
τw(Wi(x, t)− W eq
i (x, t)), (4.13)
where Si = −(Θ[V ]W )i, Di ≡ ∂∂t
+vi·∇ and W eq is an artificial “equilibrium”
distribution such that Π(W eq)nα1,α1,...,αn= Π(W )nα1,α1,...,αn
for n ≤ NΠ. τw > 0
and NΠ ∈ N are model parameters.
Compared to Eq. (4.12), Eq. (4.13) exhibits two additional terms. As it will
be shown, the term δt2
2DiSi [127, 61] also eliminates first-order discretization
artefacts in the present case. The term Ωi, is a regularizing artificial collision
term. Since Ω is a relaxation-type collision term, its use is allowed because
it preserves the positive semi definite character of the density matrix that
underlies the Wigner function [128, 129]. Its role is to improve the stability
of the numerical scheme by inducing selective numerical dissipation without
directly affecting the dynamics of the first n ≤ NΠ moments of the Wigner
equation. This can be seen as follows, let us consider the Taylor expansion
up to second order of the l.h.s of Eq. (4.13), namely
DiWi +δt
2D2i Wi = Ω + Si +
δt
2DiSi, (4.14)
44
4.2. LATTICE WIGNER MODEL
by solving for DiWi and recursively substituting back in the second term of
the l.h.s of Eq. (4.14), it is found that
DiWi+
δt
2Di
(−δt
2D2i Wi + Ω + Si +
δt
2DiSi
)= (4.15)
Ω + Si +δt
2DiSi.
From Eq. (4.15), it can be seen that had the term δt2
2DiSi not been intro-
duced in the definition of the model, there would be an uncompensated term
of order δt. Further, if the velocity moments of Eq. (4.15) are calculated,
it can be seen that all the contributions involving Ωi vanish, provided that
the order of the moment is not larger than NΠ. Thus, up to terms of order
O(δt2) and n ≤ NΠ, the resulting set of equations
∂
∂tΠ(W )nα1,α1,...,αn
+∇ · Π(W )n+1α1,α1,...,αn+1
= Π(S)nα1,α1,...,αn+O(δt2), (4.16)
is consistent, with the moments of Eq. (4.10). In summary, Eq. (4.13) ap-
proximately solves the Wigner Equation by solving the corresponding trun-
cated hierarchy of equations Eq. (4.16).
So far, the series representation of the source term Eq. (4.9) was assumed. It
is clear that its equivalence to the natural integral representation Eq. (4.8)
depends on the potential smoothness [26]. Furthermore, in passing from
Eq. (4.8) to Eq. (4.9) it is also assumed that the potential can be Taylor
expanded and that radius of convergence is the whole domain of integration.
This extra condition may not be fulfilled by some smooth functions. This
rises the question on whether the Lattice Wigner model depends on the
chosen representation of the source term. To address this question, first
notice that only the moments of the source term are relevant to the dynamics
of the moments of the Wigner function Eq. (4.16), thus it is sufficient to
show that Eq. (4.9) and Eq. (4.8) have the same moments. First consider
45
CHAPTER 4. LATTICE WIGNER MODEL
the moments of the source term using the series representation,
Π(S)n =
∫pn
∑|s|∈Nodd
(H
2i
)|s|−11
s!
∂sV
∂xs∂sW
∂vs
dp, (4.17)
since only the Wigner function depends on the momentum variable the pre-
vious expression can be simplified as
Π(S)n =∑|s|∈Nodd
(H
2i
)|s|−11
s!
∂sV
∂xs
∫pn∂sW
∂vsdp, (4.18)
using iterated integration by parts it is found that∫pn∂sW
∂vsdp = (−1)λ
n!
(n− s)!Π(W )(n−s), for n ≥ s (4.19)
thus the final result is
Π(S)n =n∑
s∈Nodd
(−1)s(~2i
)s−1(n
s
)∂sV
∂xsΠ(W )(n−s). (4.20)
For the Integral representation of the source term, the starting points are
Eq. (4.7) and Eq. (4.8) and the moment definition that leads to
Π(S)n =
∫ ∫ ∫W (q, p)
(i
π~2
)[V (q + y)− V (q − y)
]pne−i2(j−p)y/~dpdydj.
(4.21)
The previous expression can be reorganized as
Π(S)n =
∫ ∫W (q, p)
(i
π~2
)[V (q + y)− V (q − y)
]e−2jy/~
∫pnei2py/~dpdydj,
(4.22)
using the Fourier property∫xne−iνxdx = 2πinδ(n)(ν) it follows that
Π(S)n =
∫ ∫W (q, p)
(i
~
)[V (q + y)− V (q − y)
]e−2jy/~
(~2i
)nδ(n)(y)dydj.
(4.23)
The terms involving the y variable are integrated as follows∫ [V (q + y)− V (q − y)
]e−2jy/~δ(n)(y)dy =
(−1)n(dn
dyn[V (q + y)− V (q − y)]e−2jy/~
)|y=0 =
(−1)nn∑s=0
(n
s
)(1− (−1)k)
∂sV
∂xs
(−2ij
~
)n−s, (4.24)
46
4.2. LATTICE WIGNER MODEL
finally replacing Eq. (4.24) back in Eq. (4.23) leads to Eq. (4.20). It can
be seen that the general moments of the source term are independent of
the chosen representation. The only depend on the local existence of the
derivatives, not on the more restrictive requirement that the function has a
Taylor expansion with infinite radius of convergence.
Since the Wigner function is bounded over the phase space [130] and only a
limited number of moments are required, due to the truncation in Eq. (4.16),
an expansion in orthonormal polynomials can be assumed for W , S and W eq,
from which the expressions of the corresponding distributions can be derived.
From sec. 2.1.2 the formal result is
W (x,v, t) ≈ ω(v)
Np∑n
an(x, t)Pn(v), (4.25)
where Pn(v) is a family of polynomials orthonormal under the weight
function ω(v) and Np is the maximum order of the polynomials used in the
representation. The expansion coefficients are given by
an(x, t) =
∫d3vW (x,v, t)Pn(v). (4.26)
It is interesting to note that, since the expansion coefficients are linear com-
binations of the moments of the distribution, this procedure is similar to
Grad’s method [131], although not restricted to Hermite polynomials.
In practice, Hermite polynomials are a convenient choice, as they permit the
systematic generation of lattices in any number of dimensions [108, 109, 110].
For example, in one dimension and using Hermite polynomials, Hn(v; cs)
with weight function ω(v; cs) = 1√2πc2s
e− v2
2c2s and parameter cs > 0 [132], the
expressions for Wiand W eqi are, by construction, given by
Wi = ωi
Np∑n
an(x, t)Hn(vi; cs), (4.27)
W eqi = ωi
NΠ∑n
an(x, t)Hn(vi; cs), (4.28)
47
CHAPTER 4. LATTICE WIGNER MODEL
whereas Si can be explicitly calculated (see Appendix. A.3) leading to
Si = −ωi∑n,s
an(x, t)
√(n+ s)!
n!
(−H/i)s−1
csss!
∂sV
∂xsHn+s(vi; cs), (4.29)
where H is the dimensionless reduced Planck constant. It should be noted
that, in general, the condition NΠ < Np must hold, for otherwise Ωi becomes
trivially zero.
4.3 Validation
We validate our model, first for the harmonic oscillator and then for the case
of the anharmonic potentials with up to sixth order.
4.3.1 Harmonic potential
As a first example to illustrate the lattice Wigner method described in the
previous sections, we consider the quantum harmonic oscillator described by
the following Hamiltonian Eq. (4.30):
H =p2
2+
1
2x2. (4.30)
We track the time propagation of the Wi distributions from the initial condi-
tions, for different choices of the spatial resolution and number of moments
NΠ.
The initial condition consists of an equally weighted superposition of the first
two eigenstates of the quantum harmonic oscillator, |φ〉 = 1√2(|ψ0〉 + |ψ1〉).
Since this state is non stationary, it shows time oscillations all along the
evolution.
The Wigner function corresponding to |φ〉 can be calculated from the defi-
nition Eq. (4.1), the result being:
W|φ〉(x, v) =e−
v2+x2
H
(√2√Hx+ v2 + x2
)πH2
. (4.31)
48
4.3. VALIDATION
Observe that if Hermite polynomials are used, W|φ〉(x, v) is already of the
form Eq. (4.25). It follows that the distributions, Wi, are given by
Wi = ωie− x2
2c2s√2πc2
s
(H2(vi; cs)
2+c2s + 2csx+ x2
√2c2s
H1(vi; cs)
)(4.32)
where the specific values of wi, vi and cs for different lattices are given in
the Appendix A.6 and H was taken to be numerically equal to 2c2s.
The results of our simulation using the D1Q3 lattice with a lattice spacing
δx = 0.06 and a equilibrium function with NΠ = 3, are shown in Fig. 4.1.
On the upper panel, it can be seen that both the zeroth and first order
moments (ρ, ρu) are correctly propagated and agree with the theoretical
values at different times.
From Eq. (4.25), it is clear that, given the expansion coefficients an(x, t), it
is possible to reconstruct an approximation of the Wigner function. These
coefficients can be obtained from Eq. (4.27) as linear combinations of the
moments of the Wigner function, which, in turn, can be calculated by means
of quadratures.
The results for the quantum harmonic oscillator are presented on the lower
panel of Fig.4.1, which shows the phase-space representation of the Wigner
function. From this figure, a prototypical shape is clearly recognized, in-
cluding the expected non classical regions of negative values.
To quantitatively characterize the present method, we have studied the ef-
fects of the spatial resolution, lattice configuration and number of preserved
moments (NΠ)
To this end, the root mean square error between the theoretical density and
the simulated one after a full oscillator period, T0,
∆ =
√1
Nx
∑x
(ρtheory(x)− ρsim(x))2, (4.33)
was evaluated for different conditions.
In Fig.4.2 a), the effect of using different lattices and resolution levels is
49
CHAPTER 4. LATTICE WIGNER MODEL
Figure 4.1: (Color on line) a) First moment of the Wigner function (den-
sity) for different times being fractions of the oscillation period T0. The
inset shows the corresponding time evolution for the second moment (ve-
locity density). The symbols denote the simulation and the solid lines the
analytical solution. b) Phase space reconstruction of the quantum harmonic
oscillator Wigner function at τ = T0. The dashed contour line shows where
the Wigner function vanishes.
50
4.3. VALIDATION
1.5 1.6 1.7 1.8 1.9 2.0 2.1log(1/δx)
−5.0
−4.5
−4.0
−3.5
log(∆)
a) D1Q3D1Q8D1Q16
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15NΠ
10-6
10-5
10-4
10-3
10-2
10-1
100
∆
b)D1Q8D1Q16
Figure 4.2: (Color online) a) Root mean square error of the density, ∆, for
different velocity lattices and spatial resolution using an equilibrium function
that preserves the first three moments. b) Effect on ∆ of using different
values of NΠ for two different lattices with δx = 0.008, Np = 8 and Np = 16
for D1Q8 and D1Q16 respectively.
shown; two features are apparent, namely that the error ∆ decreases quadrat-
ically as a function of 1/δx and that, at a given value of the resolution δx,
schemes with higher number of preserved moment provide better results.
The effect of changing the value of NΠ is presented in Fig.4.2 b). From this
figure, an ideal range for NΠ can be identified. If NΠ is low, the order of
truncation of Eq. (4.16) leads to a crude approximation which in turn yields
large values of ∆. On the other end, if NΠ is equal to Np, the model becomes
unstable (this is why we have chosen NΠ < Np for both lattices), because
then the collision term Ωi, Eq. (4.13), vanishes, which implies no artificial
dissipation, hence the onset of stability issues discussed in Ref. [126]. The
anomalous point NΠ = 13, in the D1Q16 case on Fig.4.2 b), may be due to
51
CHAPTER 4. LATTICE WIGNER MODEL
compensated high order modes that reduce the artificial dissipation leading
to a larger than expected error. That was only observed for the particular
case of the harmonic oscillator (and won’t be the case for the anharmonic
potential).
4.3.2 Anharmonic potential
As a second example, we simulate the anharmonic quantum oscillator de-
scribed by the Hamiltonian
H =p2
2+
1
2x2 + αx4 + βx6 , (4.34)
where the parameters α and β determine the strength of the anharmonic
terms.
As discussed earlier on, anharmonic terms involve genuinely quantum effects
in the forcing expansion described in Eq. (4.9).
Similar to the previous example, the initial condition is taken to be the equal
superposition of the first two eigenstates of the Hamiltonian Eq. (4.34) |φ〉 =1√2
(|ψ0〉+ |ψ1〉). Here, |ψ0〉, |ψ1〉 are obtained by direct diagonalization of
Eq. (4.34), using a truncated basis set of 50 eigenvectors, ϕn, from the
quantum harmonic oscillator.
The density matrix for this system is given by
ρ(x, x′) =
∑m,n
cnc∗mϕm(x)ϕn(x′), (4.35)
where the coefficients cn are easily obtained from the diagonalization proce-
dure. Given ρ, the corresponding Wigner function W|φ〉(x, v) was calculated
(see Appendix A.4) with the help of the results in Ref. [24], leading to the
following expression:
W|φ〉(x, v) =1
2πH
∑n≤m
2
1 + δm,n<(cnc
∗mkn,m), (4.36)
52
4.3. VALIDATION
where <(·) denotes the real part, the coefficients km,n are given by
km,n =2(−1)min (m,n)
√min (m,n)!
max (m,n)!e−
x2+v2
H
(2
H(x2 + v2)
) |m−n|2
L|m−n|min (m,n)
(2
H(x2 + v2)
)e(i(m−n) arctan (v/x)) (4.37)
and Lmn is the m order n degree associated Laguerre polynomial.
In order to find the corresponding Wi, it is used that each term of Eq. (4.36)
can be written as the product of a polynomial and a Gaussian function in the
velocity space. Once the Gaussian is factored out, the result in Eq. (4.38)
is readily cast into the form of Eq. (4.25), namely
W (x, v) =e−
v2
H
√2πH
∑n≤m
2
1 + δm,n(crnmk
rnm − cinmkinm), (4.38)
In the above, cnm = (crncrm + cinc
im) + i(cinc
rm − cimcrn) and kmn = ev
2/H
2√πHkmn,
where the superscripts r and i denote real and imaginary parts, respectively.
Since Hermite quadratures are in use, the Wi follow directly.
The results for the anharmonic oscillator Eq. (4.34) with parameters α = 0.1,
β = 0.05 are summarized in Fig.4.3 a), from which it is apparent that for
mild anharmonicities, the method is able to properly evolve the given initial
condition. In Fig.4.3 b) the error as a function of the used lattice and
resolution is reported; the general trend is an error decrease at increasing
resolution; it decreases as NΠ increases and ∆ tends to saturate relatively
fast. In Fig.4.3 c), the behaviour of ∆ as function of NΠ is shown. Similarly
to the harmonic case, as NΠ increases, ∆ decreases, until it reaches and
optimal value (NΠ = 8) and then saturates.
In order to study stronger anharmonic cases, not only larger resolutions,
but also more terms in the representation Eq. (4.27) of the Wigner function
are required because as the strength of the anharmonicity increases, so does
the number of terms in Eq. (4.38). In order to account for them, both the
number of polynomials in Eq. (4.27) and the size of the velocity lattice needs
to be increased.
53
CHAPTER 4. LATTICE WIGNER MODEL
−10 −5 0 5 10x
0.000.050.100.150.200.250.300.35
ρ
a)τ= T0/3
τ=3T0/4
τ= T0
1.6 1.8 2.0 2.2 2.4log(1/δx)
−4−3−2−1012
log(∆)
b)
D1Q16 NΠ = 1D1Q16 NΠ = 15D1Q20 NΠ = 1D1Q20 NΠ = 15
1 3 5 7 9 11 13 15 17 19NΠ
10-4
10-3
∆
c) D1Q16D1Q20
Figure 4.3: (Color online) Results for the anharmonic oscillator with pa-
rameters α = 0.1, β = 0.05. a) Comparison between the density obtained
using the Lattice Wigner method and the one obtained directly from the
Schrodinger equation. (b) Error as a function of resolution, used lattice and
number of projections NΠ. It can be seen that as the resolution increases the
error saturates and that the error decreases upon increasing the number of
projections. (c) Effect on ∆ of using different values of NΠ for two different
lattices with δx = 0.008.
The effect of the relaxation time τw was also studied. By definition, this
parameter controls dissipative effects and consequently, it is not expected to
affect the results. However, numerically it was found that this is the case
only in the range 0.56 ≤ τw ≤ 5, which is similar to the allowed range of τw
in the closely related lattice Boltzmann schemes. This is possibly due to a
marginal coupling between high order moments and the ones relevant to the
Wigner dynamics.
54
4.4. COMPUTATIONAL COST
4.4 Computational cost
For an arbitrary problem, it is a priori not known how many polynomi-
als are required to give an accurate representation of the Wigner function,
Eq. (4.25). The number of polynomials, Np, determines the smallest lattice
that is able to support the orthogonality constraints, Eq. (2.12), and also
the computational cost of solving the respective problem. The scaling of the
cost can be estimated by observing that a single update of the complete set
of lattice points involves four basic steps: 1) the calculation of the expansion
coefficients an in Eq. (4.27), 2) the update of the source term distributions
in Eq. (4.29), 3) the update of W eq in Eq. (4.28), and 4) the update of Wi
according to Eq. (4.13).
The number of floating point operations (+,−,×, /) required at each step
scales respectively as O(NpNq), O(NsNpNq), O(NpNq) and O(1), where Ns
is the number of terms in Eq. (4.29) that are consistent with a cutoff at s
in H.
Under a worst-case scenario, i.e. the largest possible NΠ, NΠ ∼ Np and
Nq ∼ Np, the total cost of updating a single site scales as:
O(NsN2p +N2
p ), (4.39)
In 1D, Ns is effectively O(1) and therefore the cost per site update scales as
O(N2p ). This bound was tested and the corresponding results are reported in
Fig.4.4, from which it is seen that the cost of updating a single site scales like
N2p . The difference with respect to the theoretical value can be accounted
for by the time to access data, which becomes dominant as the size of the
problem is increased.
In 2D, Ns scales O(s2), and since the number of polynomials and lattice
vectors also scale quadratically, the update cost per site is expected to grow
as O(s2N4p ).
55
CHAPTER 4. LATTICE WIGNER MODEL
101
Np
10-6
10-5
t(s)
Figure 4.4: (Color online) The symbols show the time it takes to update
a single site as a function of the number of polynomials. The dashed line
shows the scaling t ∼ N2p . The simulations were performed for the quantum
harmonic oscillator, in every case NΠ was set to the highest value compatible
with numerical stability.
4.5 Lattice Wigner application
4.5.1 1D system
As an application of the proposed model, next we study the dynamics of
the zeroth and first moment of the Wigner function for a system subject to
the combined action of an external drive and potential barriers. As a model
of an homogeneous system, we assume that the initial state is given by the
following thermal density matrix:
ρ =∑|p〉〈p|e−βp2/2m, (4.40)
where |p〉 are plane waves, m is the mass of the particle and β is the inverse
temperature.
The system is taken to be of finite length L, which implies quantization of
the allowed momenta. However, L is assumed sufficiently large to justify a
continuum limit.
56
4.5. LATTICE WIGNER APPLICATION
The potential barriers extend throughout the domain according to:
V (x) =v0
2(erf ((x+ δ/2)ξ)− erf ((x− δ/2)ξ)) , (4.41)
where v0, δ and ξ define the height, width and stiffness of the barrier,
respectively. The barriers were symmetrically distributed at the points
xi = ±Di, i = 0, 1, . . . , Nb where D is the inter barrier distance.
The system is driven by the potential Vd(x) = −ax, where a determines
the strength of the forcing, and is assumed to be open, i.e. each end of the
domain is connected to a fixed reservoir, also described by Eq. (4.40).
Similar to the previous examples, a lattice Wigner representation of the form
Eq. (2.14) is required for the initial condition. In this case, the Wigner
transform of Eq. (4.40) is given by
W (x, v) =1
2πHe−
v2β2 , (4.42)
where β = βm(l0/t0)2.
Comparing Eq. (4.42) with the form of the Hermite polynomials weight
function, ω(v; cs) = 1√2πc2s
e− v2
2c2s , and using Eq.(4.25,2.14) it follows that if
β and H are fixed respectively to 1/c2s and cs then only the a0 expansion
coefficient, that corresponds to the constant Hermite polynomial, is required.
That is, the representation of the initial condition is optimal and the the
distributions Wi are proportional to the weights of the lattice configuration
Wi = ωi1√2π. (4.43)
Finally, it is important to observe that the barrier potential Eq. (4.41)
has infinitely many non-zero derivatives, as opposed to the harmonic and
anharmonic potentials. This implies that a cutoff in Eq. (4.29) needs to
be chosen. For the present simulations, the parameters characterising the
barriers were fixed as v0 = 0.4β−1, δ = 2 and ξ = 1. In this case, the
cutoff is taken at s = 9, since the next contribution, s = 11, is six orders of
magnitude smaller than the first order contribution.
The first two moments of the Wigner function were studied for different
values of the driving force Vd, number and location of the barriers. Fig. 4.5 a)
57
CHAPTER 4. LATTICE WIGNER MODEL
Figure 4.5: (Color online) a): Reconstructed steady state Wigner function.
b): The second moment of the Wigner function as function of Vd in steady
state, for different number of barriers Nb. c) σ as a function of the number of
barriers for different inter barrier distances. The system size is set to 400 (in
dimensionless units) and all simulations were performed with a resolution
δx = 0.004, with NΠ = 14, using a D1Q16 lattice.
shows the reconstructed steady state Wigner function, W (x, v), of a system
with Vd = 10−4 and four barriers randomly located across the domain.
Similar results were obtained for different configurations of barriers and driv-
ing force. The first visible feature is that W (x, v) shows a number of “cuts”
along the v axis at given values of x. These cuts are located at the potential
barriers. Along the barriers, the Wigner function attains lower values as
compared to the nearby regions. This implies that the density in the cuts is
smaller compared to the surroundings.
A second feature is that the Wigner function is nearly translationally invari-
58
4.5. LATTICE WIGNER APPLICATION
ant in the interstitial region between two subsequent cuts, as long as the
cuts are sufficiently far apart, which implies that the density ρ is uniform
between cuts.
Further, from Fig. 4.5 a) it seems that the Wigner distribution is symmetric
along the v = 0 axis, although this is not the case. The driving potential
slightly shifts the distribution, leading to a finite and spatially uniform first
moment (ρu), which is consistent with the continuity equation ∂ρ∂t
+∇ρu = 0,
at steady state. Finally, it can be seen that the Wigner function is nowhere
negative, first, because the reservoir naturally tends to wash out quantum
coherence and second, because the ratio between the height of the barriers
and the thermal energy is about 0.4, whereas in applications such as resonant
tunnelling diodes, such ratio is about ten [133]. Similarly to the case of strong
anharmonicities, to treat systems with higher energy barriers, more terms
i.e. polynomials in the representation of the Wigner function (Eq. (4.25))
are needed, along with the corresponding increase in the velocity lattice size.
From Fig. 4.5 b) it can be seen that the relation between the velocity density
ρu and the forcing potential Vd is linear for a fixed number of barriers, uni-
formly and symmetrically distributed across the domain. Further, Fig. 4.5
b) also implies that, as the number of barriers increases, the electric conduc-
tivity, σ, decreases.
In other words, the capacity of the system to transport momentum from one
end to the other, declines with number of barriers. To quantify this relation,
simulations with a fixed number of barriers, Nb, but different interbarrier
distances, D, were performed. The results, reported in Fig.4.5 c), show
that the overall tendency is a decreasing σ at increasing Nb. However, this
decrease shows a dependence on the interbarrier separation D. For D = 2
and D = 2.5, σ is nearly constant, whereas for D ≥ 3 it decreases rapidly
with Nb. Furthermore, σ saturates above D ≥ 5.
The above picture can be understood as follows: once the barriers are suffi-
ciently close together, they overlap and the resulting potential is no longer
a set of disjoint barriers, but rather a single larger barrier Fig.4.6 a).
In this case, it is known that all incoming plane waves with energy below
59
CHAPTER 4. LATTICE WIGNER MODEL
D a b c
5 53.3± 0.2 38.2± 0.2 0.898± 0.002
5.5 53.7± 0.2 37.6± 0.2 0.896± 0.002
6 53.6± 0.2 37.7± 0.2 0.898± 0.003
6.5 53.4± 0.2 37.8± 0.2 0.899± 0.003
8 53.3± 0.2 37.9± 0.2 0.898± 0.003
Table 4.1: Individual fitting parameters of Eq. (4.44) for different interbar-
rier distances
the barrier are exponentially attenuated as a function of the barrier length,
whereas those with energy above the barrier manage to penetrate, if only
with a non-zero reflection probability. It follows then that the number of
states that can cross the barrier diminishes as the length of the barrier
increases thereby limiting the amount of momentum transported across the
system, thus leading to an overall decrease of σ.
When the separation between the barriers is sufficiently large, the system
can be approximated as a sequence of disjoint barriers. If the system was
closed, this would imply that, T being the transmission coefficient for a
single incoming plane-wave on a single barrier, the transmission coefficient
for n barriers, would be TN = TN(1−T )+T
(See Appendix A.5), without any
dependence on the interbarrier separation. Since this holds for every plane
wave contributing to the thermal density matrix, the system as a whole is
expected to follow a similar trend.
From the previous picture, it can be inferred that the σ − n relation must
have a similar form for the D ≥ 5 settings. The semi empirical formula
σ = a+b cn(1−c)+c , where a, b, c are parameters depending on the interbarrier
separation, offers a good fit to the casesD = 5, 5.5, 6, 6.5, 8. From Table4.5.1,
it is apparent that the parameters a, b and c are constant within error bars.
Therefore, for D ≥ 5, the relation between σ and n and D, is effectively
independent of D and given by
σ = a+ bc
n(1− c) + c, (4.44)
60
4.5. LATTICE WIGNER APPLICATION
Figure 4.6: (Color online) a) The total potential as a function of the inter-
barrier separation. For D = 2 the barriers are close enough such that the
total resulting potential acts as a single barrier. As the interbarrier separa-
tion increases, the resulting potential exhibits the structure shown for the
D = 2.5, 3 cases. b) The symbols show the behaviour of ρu, averaged
over 50 random samples, as a function of the driving potential. The dashed
lines show the behaviour of ρu in the uniformly distributed case with an
interbarrier distance D = 8.
with a = 53.4± 0.2, b = 37.8± 0.2, c = 0.899± 0.002.
The intermediate case 2.5 < D < 5, when the barriers do not form a single
monolithic barrier and the system can no longer be regarded as a superposi-
tion of disjoint subsystems, requires a deeper analysis which is left for future
work.
We have also studied the momentum transport in the presence of a random
distribution of barriers.
Simulations were performed for a fixed number of barriers Nb, randomly
located across the domain. The minimum distance between any two barriers
was constrained to be larger than 2 lattice sites, in order to avoid excessive
overlap, leading to an effective single larger barrier instead of two distinct
ones. The results are presented in Fig. 4.6 b), where for every instance 50
random realizations were considered.
The main observation is that the relationship between the current ρu and
Vd is, on average, the same as with uniformly distributed barriers, with an
interbarrier distance D > 5. This result can be understood as follows; since
61
CHAPTER 4. LATTICE WIGNER MODEL
the barriers are constrained to be far apart, most configurations behave as a
collection of subsystems. This, in turn, implies that σ only depends on the
number of barriers Eq. (4.44) and, as a consequence, the average relation
between ρu and Vd does not depart significantly from the case of a regular
distribution of barriers.
4.5.2 2D system
The transport properties of a square shaped two-dimensional system of side
length L, were also studied. Open boundary conditions were used at the
x = 0 and x = L ends, while periodic boundary conditions are used at the
y = 0 and y = L ends. The system is driven by an external potential of the
form Vd(x) = −ax, where a controls the strength of the external driving.
The barriers are described by the potential
V (x) = v0e− |x|
2
2ξ2 , (4.45)
where v0 determines the height of the barrier and ξ its stiffness.
The initial state is also given by Eq. (4.40), where |p〉 is assumed to be
two dimensional. Following calculations similar to the 1D case, the initial
condition for the lattice Wigner model is given by
Wi = ωi1
2π. (4.46)
The cutoff of Eq. (4.11) was set to s = 9 and the simulations where carried
out on a 256× 256 grid, using the D2Q16 lattice (see Appendix for details).
Similarly to the 1D case, regular and a random settings for the location of
the potential barriers were considered. Fig.4.7 a) shows a sample result for
a simulation with 16 randomly placed barriers. The location of the poten-
tial barriers can be easily identified through the blue color spots, denoting
density depletion. Further, it can be seen that the streamlines bend around
the potential barriers, similarly to the way fluid streamlines turn around
obstacles in porous or campylotic media [134, 113].
The relation between the flux Φ (2D analogue of ρu in 1D) and the driving
potential is presented in Fig. 4.7 b). From this figure, it is seen that the
62
4.5. LATTICE WIGNER APPLICATION
Figure 4.7: (Color online) a) Density map for a 2D system with 16 randomly
located barriers. The effect of the barriers can be observed on the regions
that get depleted (blue color) and on the streamlines that bend around them.
b) Behaviour of Φ as a function of the driving strength. The red dashed line,
blue dot-dashed line and green dotted line correspond to systems where the
barriers are arranged in regular grids of 2× 2, 3× 3 and 4× 4 barriers with
an interbarrier distance of D = 9. The circle, square and triangle symbols
represent respectively the mean flux of 50 random samples of 2 × 2, 3 × 3
and 4× 4 randomly located barriers. The solid lines are a guide to the eye
showing the trend of Φ as a function of Vd for the case of random barriers.
relation σ versus Φ and Vd is linear when the barriers are regularly organised
on a square grid, and that σ decreases at increasing number of barriers.
Furthermore, when the barriers are randomly placed, the average behaviour
63
CHAPTER 4. LATTICE WIGNER MODEL
of Φ is close to the regular case, as it was also observed in 1D. However, as the
number of barriers increases, specific realizations can deviate significantly
from the regular grid behaviour, this can be seen from the error bars of the
red triangles in Fig. 4.7 b).
Finally, for the purpose of showing the viability of the present method also
in three spatial dimensions, we have simulated a three-dimensional open
quantum system. The simulation was performed on a 20 × 20 × 20 lattice,
with a D3Q125 velocity set, which was chosen because it includes terms of
order H2 in the force expansion Eq. (4.29). The boundary conditions are
open (thermal density matrix) at the planes normal to F (See Fig.4.8) and
periodic on the remaining boundaries.
Figure 4.8: (Color online) The figure shows the density ρ and streamlines of
ρu of an open driven system in 3D. The drive is given by a constant force,
F , along the x direction.
In addition to the driving potential generating a force in the x direction,
a random potential is included. It is modelled as a smooth Gaussian with
varying amplitude at different locations in the domain. From Fig.4.8, it
is seen that the streamlines tend to circumvent the regions of low density,
64
4.6. SUMMARY
where the potential is high, and concentrate in the regions of high density,
thus effectively avoiding “impurities”. A systematic analysis of the transport
properties of this three-dimensional open quantum system is left for future
work.
4.6 Summary
In this work, a new numerical method to track the time evolution of the
Wigner function has been introduced. The stability problem previously de-
scribed in Ref. [126], is handled through the inclusion of an artificial collision
term, designed in such a way as to preserve the dynamics of the relevant mo-
ments of the Wigner function. Reducing the momentum space to a compar-
atively small set of representative momentum vectors, opens up interesting
prospects for the simulation of one, two and also three dimensional quan-
tum systems. Preliminary results for 1D systems with regular and random
potentials provide evidence of linear transport laws which are independent
of the barrier configuration for dilute systems. In the 2D case, we find the
same transport laws at low barrier density, while for higher concentrations,
deviations from the linear behaviour are observed (as shown in Fig.4.7 b)
when the barriers are randomly located. Finally, we also presented a pre-
liminary simulation of a 3D open quantum system, to illustrate the ability
of the model to handle the three-dimensional Wigner equation.
The computational cost of the method scales polynomially with the number
of basis functions. However, the simulations show that just a few equilibrium
moments and comparatively small lattices, are often sufficient to obtain
reasonably accurate results.
The present work opens up a number of research directions for the future.
Technically, the performance can be improved by choosing alternative fam-
ilies of lattice configurations and orthonormal polynomials, or by directly
designing orthonormal polynomials that fit the specific problems under in-
vestigation. Since our model is computationally viable also in 3D, problems
like the heat transport properties of three-dimensional semiconductor struc-
65
CHAPTER 4. LATTICE WIGNER MODEL
tures, which are highly relevant to the next generation electronics [135],
could be studied. In addition, the method could also be used as a practical
tool to explore fundamental issues, such as the relation between quantum
entanglement and the Wigner function in diverse systems [136, 137].
66
Chapter 5
Thermal gas rectification using
a sawtooth channel
5.1 Ratchet systems
The subject of ratchet systems has been studied in a number of contexts,
from classical and quantum physical systems to biological examples. In the
following we illustrate the main physical ideas of such systems following the
analysis of the Smoluchowski ratchet done by Feynmman.
The set up of the Smoluchowski-Feynmman ratchet is shown in Fig. 5.1.
It consists of two chambers filled with a gas temperature T , and an axle.
On the right chamber the axle is connected to a set of paddles, on the left
chamber it is connected to a ratchet pawl mechanism and in the middle to a
load. The intended working mechanism of the contraption is as follows: On
the right chamber, the gas particles will collide with the paddles and given
the symmetry of the system, it is expected that the particle-paddle collisions
will not generate a rotation of the axle in any preferred direction. However,
it is also expected that the fluctuations will be such that in some occasions
there is an excess of collision that lead to a small motion in a preferred
direction. If the direction of rotation is that of the “forward” or “easy”
direction of the ratchet in the left chamber, then the axle will rotate and in
the process it will perform work on the load rising it. From the second law of
thermodynamics such a mechanism can not function for it could extract work
67
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
from a single heat reservoir. From a mechanical point of view the system
would not be able to function because if it assumed that the ratchet-pawl
interaction is conservative, on average, the pawl would not be able to lock
in on the ratchet tooth [73]. If the interaction is made dissipative then the
system will heat up and thermal fluctuations will preclude its functioning
unless the excess heat is taken away by a reservoir at lower temperature,
case in which no violation of the second law occurs.
In summary, the previous example shows that the two necessary elements
for a system to exhibit a ratchet or rectifying behaviour are broken spatial
symmetry and dissipation (which leads to non equilibrium conditions) [74,
75]. Notice, that the example also shows that spatial asymmetry alone is not
enough to observe a ratchet effect, this is due to the intrinsic time symmetry
of the mechanical equations of motion [73].
Figure 5.1: Ratchet and pawl mechanism: the mechanism has two parts,
on the right chamber there are a set of paddles attached to an axle that
is connected to the ratchet and pawl system on the left chamber. (Picture
taken from [73] )
68
5.2. CANONICAL MOLECULAR DYNAMICS
5.2 Canonical molecular dynamics
5.2.1 Molecular dynamics
Given the practical impossibility of analytically solving the Newton equa-
tions of motion that describe systems of many interacting particles for all but
a small number of examples, the use of alternative methods is necessary. Be-
sides the tools from statistical physics, one of the most successful approaches
to understand systems of many classical particles is that of molecular dy-
namics simulations. The method was introduced in the 1950 and it has been
applied to a number of systems in fields ranging from physics [138, 139, 140]
to biology [141, 142]
The general idea of molecular dynamics is to integrate in time the Newton
equations of motion of a system of particles. More specifically, it is assumed
that the system is composed by N particles that interact with each other via
a potential of the form V (x1,x2, . . . ,xN). This leads to the set of equations
xi = − 1
mi
∇xiV (x1,x2, . . . ,xN) i = 1, 2, . . . , N (5.1)
that are integrated in time using any one of a number of available methods
such as Verlet [143], Leap frog [144] or predictor-corrector algorithms [145].
The choice of method depends on the particular system under considera-
tion, for instance Verlet methods are suitable for conservative systems be-
cause they are known to be simplectic integrators i.e. they conserve phase
space volume and are therefore guaranteed to be energy conserving [146].
Similarly, predictor-corrector algorithms are better suited for modified first
and second order equations of the general form x = g(x, x) [147]. Since
the system of interest in this work is not conservative and includes terms
proportional to x, a predictor-corrector algorithm of fifth order is used.
5.2.2 Predictor corrector integration
Predictor-corrector algorithms for molecular dynamics are based on the idea
that dynamical quantities such as position, velocity and acceleration at a
future time t+δt can be predicted using a forward in time Taylor expansion,
69
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
and then corrected using the information from the equation of motion. More
specifically, let the scaled time derivatives of the position x be defined as
r0 = x, r1 = δt ddt
x, r2 = δt2
2d2
dt2x, r3 = δt3
6d3
dt3x, r4 = δt4
24d4
dt4x then the
predicted values rpi , i = 0, 1, 2, 3, 4 at a future time t + δt that would be
obtained from a Taylor expansion are given byrp0(t+ δt)
rp1(t+ δt)
rp2(t+ δt)
rp3(t+ δt)
rp4(t+ δt)
=
1 1 1 1 1
0 1 2 3 4
0 0 1 3 6
0 0 0 1 4
0 0 0 0 1
r0(t)
r1(t)
r2(t)
r3(t)
r4(t)
(5.2)
It is clear that Eq.(5.2) alone can not lead to the correct dynamics of the
system because it is missing the information from the forces. This is ac-
counted by the correction step. Given the predicted location and possibly
velocity of the particles, the force at time t + δt is calculated and from it
the corrected acceleration, rc2, is found. Using rc2 the correction term ∆r2 is
defined as
∆r2 = rc2 − rp2, (5.3)
and the full correction step is given byrc0(t+ δt)
rc1(t+ δt)
rc2(t+ δt)
rc3(t+ δt)
rc4(t+ δt)
=
rp0(t+ δt)
rp1(t+ δt)
rp2(t+ δt)
rp3(t+ δt)
rp4(t+ δt)
+
c0
c1
c2
c3
c4
∆r2 (5.4)
where the coefficients ci where derived by Gear [145] and for the specific case
of a fifth order predictor corrector are given by
c0 c1 c2 c3 c4
19/90 3/4 1 1/2 1/12
Table 5.1: Fifth order predictor corrector Gear coefficients
70
5.2. CANONICAL MOLECULAR DYNAMICS
5.2.3 Canonical molecular dynamics
Assuming there are no dissipative forces, the molecular dynamics approach
as previously described is only capable of performing microcanonical sim-
ulations. Observe that in principle the system of particles remains al-
ways in a manifold of constant energy. To perform a simulation at some
fixed temperature T there exist a number of techniques such as stochastic
methods [148, 149], constraint methods [150], and extended system meth-
ods [151, 152]
Here we use the so called Nose Hoover thermostat [152]. This approach is
based on the idea of coupling the actual system of particles to an additional
degree of freedom which has the role of a reservoir. The dynamics of the
combined systems undergoes a microcanonical evolution, but the dynamics
of the system of particles is canonical. Notice that this approach is analo-
gous to the way the canonical ensamble is typically introduced in statistical
physics.
In more detail, the reservoir degree of freedom is denoted by s, its contribu-
tion to the total energy has both a “kinetic” and a potential energy term.
The former is given by 12Qs2 where Q is known as a thermal inertia, and the
later is described by the potential
V (s) = (3N + 1)KbT ln (s), (5.5)
where N is the number of particles in the system and Kb is the Boltzmann
constant. The coupling between the system of particles and the reservoir is
given by the relations
xi = p/mi, (5.6)
pi = fi − ξpi, (5.7)
ξ =s
s, (5.8)
ξ =1
Q
(N∑i=1
p2i
mi
− (3N + 1)KbT
). (5.9)
Notice that the net effect of the reservoir degree of freedom is to act as
a dynamic “friction coefficient” that can slow down or speed up particles
71
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
in order to reach the appropriate mean kinetic energy. An example of the
velocity distributions obtained with this method is show in Fig. 5.2.
0 5 10 15 20 25 30 35 40 45v
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
P(v
)
T=2.5
T=10.5
T=100.5
Figure 5.2: Results of a test simulation to verify the proper working of the
Nose-Hoover thermostat. The obtained velocity distributions were checked
against the expected Maxwell Boltzmann distribution for different tempera-
tures. In all cases the difference between the mean and standard deviation of
the velocity distribution of the simulations was within 1% of the theoretical
ones
72
5.3. MODEL AND METHODS
5.3 Model and methods
We consider a two-dimensional sawtooth channel of linear length L consist-
ing of a sequence of N equal cells, as represented in Fig. 5.3, with periodic
boundary conditions along the horizontal axis. The geometry of the chan-
nel is characterized by four lengths: the length of each cell l = L/N , the
aperture size h, and the horizontal position a and height d of the edge. To
systematically study the dependence on the asymmetry of the channel, we
fix h and define the adimensional asymmetry coefficient α as,
α = 1− 2a
l, (5.10)
where α ∈ [−1, 1]. For α = 0 the channel is symmetric with respect to the
vertical axis, while for α = ±1 the cells look triangular. We classify channels
of negative and positive α as left and right asymmetric, respectively.
We consider a gas of particles interacting pairwise, where the force of particle
j on particle i is conservative and given by Fij = −∇iULJ. ULJ is the 12-6-
d
Figure 5.3: Schematic representation of the channel of size L, with N = 4
cells. The shape of each cell is characterized by four lengths: the linear
length l, the aperture size h, and the horizontal position a and height d of
the peak. The depicted channel is classified as right asymmetric (see text)
and the arrows indicate the corresponding direction of particle flow. The
dashed lines delimit the region where particles are initially released.
73
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
Lennard-Jones potential,
ULJ (rij) = 4ε
[(σ
rij
)12
−(σ
rij
)6]
, (5.11)
where rij = |rj − ri|, and ri and rj are the positions of particles i and j,
respectively. ε corresponds to depth of the potential well which is located
at rm = 21/6σ. The force Fiw of the wall on a particle i is described as
the superposition of two contributions: a conservative force, Fciw, and a
dissipative one, Fdiw. The conservative force is described as a Lennard-Jones
interaction with the closest point on the wall, with the same ε and σ of the
particle/particle interaction. The dissipative force is given by,
Fdiw = −γ (ri · niw) niw , (5.12)
where niw = ri−rwriw
is the unit vector pointing from the closest point on the
wall rw to the particle i and γ ≥ 0 is a friction constant. The particle/wall
interaction is conservative for γ = 0 and dissipative otherwise. Interactions
with the wall are truncated at a cutoff distance dc = 2.5σ and if the par-
ticle is within the cutoff distance of multiple points the contributions are
superimposed. The particle/wall interaction model was chosen to study the
effect of the wall geometry on the particles dynamics. It is assumed that the
particles locally bounce off the wall, thus the use of a cutoff distance and
the nearest point prescription. dc = 2.5σ has been found to be a reasonable
cut-off Ref. [147]. The analysis of more complex particle/wall interaction
models where the walls are directly modelled as a fixed set of particles is left
for future work.
We performed canonical molecular dynamics simulations, using the Nose-
Hoover thermostat [147, 140, 152]. Accordingly, the equation of motion of
particle i is,
ri =1
mi
(∑j 6=i
Fij(rij) + Fiw
)+ FNH , (5.13)
FNH = −ξx ,
74
5.4. RESULTS
where rij = rj − ri, mi is the particle mass and FNH is the force per unit
of mass, resulting from the coupling with the thermostat [152] and ξ is the
variable that describes the thermostat, its dynamics is given by Eq.(5.9) and
the thermal inertia was set to Q = 0.05 throughout the work.
For simplicity, we set mi ≡ m and consider reduced units, such that: mass is
in units of m, distance in units of σ, and energy in units of ε. The equations
of motion are integrated using a fifth-order predictor-corrector algorithm,
with a time step dt = 10−5 and we run the simulation up to t = 165.
To generate the initial configurations, all particles were released within the
region delimited by the dashed lines in Fig. 5.3, with an initial velocity
drawn from a uniform distribution of zero mean. Particles are thermalized at
the thermostat temperature within the dashed region, considering periodic
boundary conditions along the horizontal direction and reflective top and
bottom boundaries, without interacting with the channel walls. At t = 15,
the constraint imposed by the dashed lines is removed and particles move
inside the channel, following the dynamics described by Eq. ( 5.13).
5.4 Results
To characterize the effect of the asymmetry of the channel walls on the
overall flux, we fixed N = 24, l = 15, d = 3, and h = 6, and performed
simulations for different values of the asymmetry coefficient α ∈ [−1, 1]. To
reduce statistical noise, instead of directly measuring the outlet flux φ(t), we
introduce an integrated quantity B(t), which we call balance. B(t) is defined
as the difference between the cumulative number of particles crossing the
rightmost boundary from the left to the right and the ones crossing it in the
opposite direction, up to time t. In the continuum limit,
φ(t) =1
hB(t) . (5.14)
Asymptotically, we expect that the balance scales linearly in time, and so
we estimated the flux from a linear regression fit of the curve B(t) in the
linear regime.
75
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
Figure 5.4: (Color online) Snapshots of the time evolution of a system with
700 particles at T = 5 with γ = 1 and α = 1. The top row shows the t = 15
configuration of the system. The successive rows show the configurations at
t = 20, t = 30, t = 150 respectively
The time evolution of a particular realization of the studied system can be
observed in Fig. 5.4. On the top row the initial condition can be observed.
The particles velocities along the horizontal axis are color coded, blue(red)
correspond to right(left) velocity. It can be seen that the distribution is
random and that there are approximately as many particles with positive
velocity as there are with negative. In the subsequent rows, it can be seen
how as time progresses the number of particles travelling to the left dimin-
ishes, and in the last row it can be observed that the particles going to the
right are the majority.
On a quantitative basis, figure 5.5(a) shows the balance as a function of time
for different values of α. Clearly, spontaneous flow emerges as a result of the
asymmetry of the channel walls. For right symmetric channels (α > 0), the
flow is from the left to the right, while for left symmetric channels (α < 0),
the flow is in the opposite direction. Figure 5.5(a) also shows examples
for α = 0, 0.8, 1, for a particle-wall cutoff distance of 5σ. These examples
show that the observed rectification of the particle motion is still observed for
a larger cutoff distance, but the quantitative values of the flux are obviously
different. To analyse the transition at α = 0, we define Π as the fraction of
samples where B(t) > 0 for large values of t (t = 165). The dependence of
76
5.4. RESULTS
0 20 40 60 80 100 120 140t
-200
-100
0
100
200B(t)
a)
−1.0 −0.5 0.0 0.5 1.0α
0.0
0.2
0.4
0.6
0.8
1.0
Π(α
)
b)
Figure 5.5: (Color online) a) Time evolution of the balance B(t) for differ-
ent values of α = −1.0,−0.8,−0.6,−0.4,−0.2, 0, 0.2, 0.4, 0.6, 0.8, 1.0 (from
bottom to top). Results are averages over 500 samples of systems of 100 par-
ticles at a thermostat temperature T = 2.5 and γ = 1. The black dashed
lines correspond to simulations with α = 0, 0.8, 1 and dc = 5σ which show
that the rectification phenomena is not unique to the choice dc = 2.5σ b)
Fraction of samples Π for which B(t) > 0, for T = 2.5,ρ = 0.143, and
N = 250 (blue circles), N = 500 (green squares), N = 1000 (red triangles).
Π on α is in Fig. 5.5(b), for three different sizes of the channel. One sees
that Π is 0.5 for α = 0 and it seems to converge to a step function as the
system size increases.
The dependence of the flux on the density is shown in Fig. 5.6(a), for α = 1
and T = 2.5. One clearly observes an optimal density (ρopt ≈ 0.45) at
which the flux is maximized. The data for ρ < ρopt suggests two different
regimes (see inset of Fig. 5.6(a)): a low-density regime, for ρ < 0.1, and
an intermediate-density regime, for 0.1 < ρ < ρopt. It is expected that the
flux of particles is a monotonic increasing function of the density up to the
point where it either saturates or starts to decrease. For systems where the
particle/particle collisions are more frequent than the particle/wall collisions,
kinetic theory [153] suggests that the flux is linear in the density. However,
for low densities, most of the particle collisions are with the wall; This implies
that the flux is mostly determined by the chance of a particle to bounce off
the wall and eventually cross the boundary at either end of the channel thus
77
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
0.00 0.25 0.50 0.75 1.00ρ
0.0
0.5
1.0
1.5
2.0φ
a)
10-2 10-1 10010-2
10-1
1001.0
1.4
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8ρ
0.00
0.25
0.50
0.75
1.00
1.25
1.50
1.75
v x
b)
Figure 5.6: (Color online) a) Flux φ as a function of the density ρ, for
T = 2.5 and α = 1. The two initial regimes are shown in the inset in
a double logarithmic plot, where the dashed lines represent the power-law
fits, φ ∼ ρβ, with β = 1.47 ± 0.07 and β = 1.01 ± 0.04 for the low and
intermediate density regimes, respectively. b) Average horizontal component
of the particle velocity (vx) as a function of the density, for the same set of
parameters.
increasing or decreasing the net flux. In this case the monotonic increase of
the flux is not necessarily linear. The numerical data suggests a power law
scaling: φ ∼ ρβ. Assuming a power-law scaling, we estimate β = 1.47±0.07,
for the low-density regime, and β = 1.01±0.04, for the intermediate-density
one.
Figure 5.6(b) shows the average horizontal component of the velocity (vx) as
a function of the density. For the first regime, vx increases with the density
thus, the enhancement of the flux with the density stems from an increase in
the number particles per cell and possible collective effects affecting the par-
ticle velocity. Similar flux enhancement was reported in the context of comb
systems, where the comb tooth would take the role of the sawtooth [154].
However, in Ref. [154], the flux enhancement is observed for high densities
and is related to the saturation of the traps. Here, instead, we observe a
flux increase for much lower densities suggesting a different mechanism.
By contrast, for the intermediate-density regime, vx does not significantly
change with the density and so the flux only increases due to an increase in
78
5.4. RESULTS
0.0 0.2 0.4 0.6 0.8 1.0
α
0
5
10
15
20
25
30
φ/ρβ
ρ=0. 013
ρ=0. 018
ρ=0. 024
ρ=0. 030
ρ=0. 068
ρ=0. 091
ρ=0. 106
ρ=0. 152
ρ=0. 213
ρ=0. 244
ρ=0. 275
ρ=0. 382
Figure 5.7: (Color online) a) Flux φ rescaled by ρβ as a function of α for
T = 2.5 and different values of the density. The values reported on the
left(right) legend were rescaled using β = 1.47(β = 1.01).
the number of particles per cell, yielding a linear scaling. For ρ > ρopt, a
third regime is observed, for which the flux simply decreases with the density
due to crowding effects. That is, as the density increases, the available space
for a particle to move diminishes. Thus, the motion of a single particle is
strongly constrained by the presence of others. This in turn implies that
for a particle to move over an extended region, it requires a concerted re-
arrangement of several other particles. As the density increases, the chances
that this concerted motion leads to a majority of particles moving in a
preferred direction diminishes. Instead, it is more likely that the particles
re-arrange by moving with no preferred direction leading to a decrease in
the flux. Notice that this is consistent with the fact that vx approaches zero
(Fig. 5.6(b)) which implies that roughly the same number of particles travel
in each direction.
79
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
−10 −5 0 5 10 15 20 25
T− Tmax
0
2
4
6
8
10
12
φ/ρ
γ=1. 0
γ=0. 9
γ=0. 8
γ=0. 7
γ=0. 6
γ=0. 5
0.5 0.6 0.7 0.8 0.9 1.0
γ
2
4
6
8
10
12
Tmax
Figure 5.8: (Color online) Scaled flux, φ/ρ, as a function of the thermostat
temperature, for α = 1, ρ = 0.0917, and different values of the friction
constant, γ. The thermostat temperature was shifted by Tmax, defined as
the optimal temperature at which a maximum is observed in the flux. The
dependence of Tmax on γ is shown in the inset.
To study the dependence on α, we plot in Fig. 5.7, the flux rescaled by
ρβ, using the estimated values of β for the corresponding regime. We ob-
serve a data collapse for each regime, suggesting that the power-law scal-
ing is resilient over the entire range of α values. The low density regime
shows an optimal value of α whereas the intermediate regime shows in-
stead a nearly constant behaviour. We think that, for high-enough density,
the rate of particle-particle collisions is significantly higher than the one of
particle-wall collisions and thus the geometry of the walls does not play a
significant role on the overall dynamics. By contrast, for low density, the
rates of particle-particle and particle-wall collisions are comparable and a
competition between the two is observed, leading to the maximum in the
flux. Although there is some dispersion on the data collapse, it is clear
80
5.5. SUMMARY
that there are two distinct regimes. The origin of the dispersion can be due
to finite-size effects or scaling corrections. The study of the nature of the
transition and crossover between these two regimes requires further study,
that is beyond the scope of this work. It is interesting to notice that the
results for low densities in Fig. 5.7 are similar to those obtained by A. Sar-
racino [102]. Namely, the quantity that indicates the presence of motion
rectification (φ in our case and 〈V 〉 in Ref. [102]) shows a qualitatively sim-
ilar non-monotonic behaviour as a function of the asymmetry parameter. In
both cases, it vanishes for the symmetric case and initially grows with the
asymmetry, having a maximum for intermediate values, and then decreasing
towards a saturation value for large values of the asymmetry parameter.
The dependence of the flux on the thermostat temperature is shown in
Fig. 5.8, for different values of γ. For the entire range of values of γ, a
maximum is observed at an optimal temperature, Tmax, that increases with
γ (see inset). Also, the optimal flux grows with dissipation (increasing γ).
Note that, for γ = 0 the overall flux vanishes and thus a dissipative inter-
action with the walls is necessary to rectify the thermal motion of the gas.
This is consistent with the work of Prost et al. that suggests that time-
reversal symmetry of trajectories needs to be broken to obtain rectification
from asymmetric walls [155].
Finally, to quantify the flux for a specific system, let us consider a channel of
total length L = 216µm, single cell length l = 9µm, α = 0, p = 1.8µm and
h = 3.7µm at room temperature, with colloidal particles of σ = 6× 10−7 m,
m = 2.49 × 10−16 kg, and γ = 2 × 10−14 kg/s [156, 157]. If we assume
ε/kb = 0.01414 K (where kb is the Boltzmann constant), we obtain a flow
velocity φ/ρ ≈ 42µm/s.
5.5 Summary
In this work, we systematically study the dependence of the rectification of
the motion of a thermal gas on a channel of asymmetric dissipative walls.
We found that the overall flux enhances with the friction constant of the
81
CHAPTER 5. THERMAL GAS RECTIFICATION USING ASAWTOOTH CHANNEL
particle/wall interaction and that it shows a nonmonotonic dependence on
three other model parameters, namely, the thermostat temperature, chan-
nel asymmetry, and particle density. For the dependence of the flux on the
density of particles, we found three different regimes. For low density, the
flux scales superlinearly with the density, as collective effects lead also to
an increase in the horizontal component of particle velocity. For interme-
diate density, the horizontal component of particle velocity saturates at a
constant value and the overall flux scales linearly with the density. Finally,
above an optimal value of the density, the flux monotonically decreases due
to crowding effects. Future work might consider different geometries and a
generalization to the three-dimensional case. The effect of different dissipa-
tion mechanisms as well as particle shapes are still open questions.
82
Chapter 6
Conclusion and outlook
In this work we have studied transport ideas from very different yet interest-
ing perspectives. We have shown that Lattice Kinetic methods (LKM) are
flexible enough to handle diverse scenarios. More specifically, we improved
the original application of LKM to the field of density functional theory.
This was achieved by a mixed use of lattice kinetic and physical ideas i.e.
the introduction of correction terms to the original Lattice Kinetic scheme
was possible by using an approximation to the imaginary time evolution
of wave functions hence, the semi-implicit nature of the method. The new
approach was tested on a number of realistic calculations and it showed
good agreement with the known results. For a balanced assessment of the
approach, it is also important to say that the incorporation of pseudopoten-
tials was problematic and that further research on that direction is required.
A very interesting path for future research is to use the diffusive character
of the Schrodinger equation not in imaginary time, but in the real time. It
is easy to see that by considering the real and imaginary parts of the wave
function, the Schrodinger equation naturally splits into two coupled diffusion
equations that are amenable to study using lattice kinetic methods.
As a further case study for the applicability of LKM, its use for the Wigner
function formalism was studied. In this case, a new approach to perform
the time evolution of the moments of the Wigner equation was developed.
The proposed approach improved the previously known stability problems
of similar attempts. This was achieved by introducing a collision term that
83
CHAPTER 6. CONCLUSION AND OUTLOOK
although, is not present in the original problem formulation, does not affect
the dynamics of the moments of the Wigner function and also keeps the pos-
itive semidefinite character of the density matrix that underlies the Wigner
function. The idea was tested for the quantum harmonic and anharmonic
problems where good agreement with the theory was found. The approach
was also used to study transport on simple open driven one dimensional
systems with potential barriers, there a linear relation between the particle
current and the driven potential was found. Also studied was the relation
between the conductivity and the number of barriers in the system. Here it
was found that if the barrier’s separation was large enough, then the conduc-
tivity was independent of the separation both in regular or random barriers
arrangements. The two dimensional analogue of the system just described
was also considered, in this case it was found that for random arrangements
of barriers there could be deviations from the linear behaviour between par-
ticle current and driving potential. Directions for future work include the
extension of the method to handle magnetic fields, as well as interacting
particles.
In the second part of this work, a novel example of Brownian motion rectifi-
cation was introduced. It was shown that it is possible to rectify the motion
of a gas of Lenard-Jones particles by using dissipation and broken spatial
symmetry. The dependence of the flux of particles on various system pa-
rameters was studied and two distinct regimes where found for the quantity
φ/ρβ, one at low and the other at intermediate densities. We expect these
results to be useful in the future in microfluidic applications or lab on a chip
settings.
84
Appendix
A.1 Wigner function derivation
The time evolution of the Wigner function can be obtained by transforming
the Liouville-von Neumann equation
∂ρ
∂t= − i
~[H, ρ], (A.1)
using the prescription of Eq.(4.1), where
H =p2
2m+ V (x), (A.2)
is the system Hamiltonian. For simplicity the derivation is made using a
pure state (ρ = |ψ〉〈ψ|) in a single dimension. The general result for mixed
states follows from the linearity of the theory and the extension to more
dimensions is straightforward.
Equation Eq.(A.1) can explicitly be written as
∂ρ
∂t= − i
~[p2
2m, ρ]− i
~[V , ρ]. (A.3)
The first commutator on the right hand side of Eq.(A.3) reads:
〈q−y2|[ p
2
2m, ρ]|q+y
2〉 =
1
2m
(〈q − y
2|p2|ψ〉〈ψ|q +
y
2〉 − 〈q − y
2|ψ〉〈ψ|p2|q +
y
2〉),
(A.4)
and in real space this is
〈q−y2|[ p
2
2m, ρ]|q+y
2〉 = −~22
m
(∂2ψ(q − y
2)
∂y2ψ∗(q +
y
2)− ψ(q − y
2)∂2ψ∗(q + y
2)
∂y2
).
(A.5)
85
APPENDIX
The Wigner transformation of the first commutator in the r.h.s of Eq.(A.3)
is then given by
− ~πm
∫ ∞−∞
(∂2ψ(q − y
2)
∂y2ψ∗(q +
y
2)− ψ(q − y
2)∂2ψ∗(q + y
2)
∂y2
)eipy/~dy.
(A.6)
Integrating by parts the preceding equation yields
~πm
∫ ∞−∞
(∂ψ(q − y
2)
∂y
∂(ψ∗(q + y2)eipy/~)
∂y−∂ψ∗(q + y
2)
∂y
∂(ψ(q − y2)eipy/~)
∂y
)dy
=ip
πm
∫ ∞−∞
(∂ψ(q − y
2)
∂yψ∗(q +
y
2)−
∂ψ∗(q + y2)
∂yψ(q − y
2)
)eipy/~dy. (A.7)
Changing the derivatives in the y variable for derivatives in the q variable
gives:
− ip
2πm
∂
∂q
∫ ∞−∞
ψ(q − y
2)ψ∗(q +
y
2)eipy/~dy = −ip~
m
∂W
∂q(A.8)
The second commutator of the r.h.s of Eq.(A.3) is evaluated as follows
〈q− y
2|[V , ρ]|q+
y
2〉 = 〈q− y
2|V |ψ〉〈ψ|q+
y
2〉 − 〈q− y
2|ψ〉〈ψ|V |q+
y
2〉 (A.9)
which in real space representation reads
〈q − y
2|[V , ρ]|q +
y
2〉 = V (q − y
2)ψ(q − y
2)ψ∗(q +
y
2)− ψ(q − y
2)V (q +
y
2)ψ∗(q +
y
2)
(A.10)
=(V (q − y
2)− V (q +
y
2))ψ∗(q +
y
2)ψ(q − y
2).
(A.11)
The Wigner transformation of Eq.(A.11) is:
1
2π~
∫ ∞−∞
(V (q − y
2)− V (q +
y
2))ψ∗(q +
y
2)ψ(q − y
2)eipy/~dy. (A.12)
If the potential is smooth enough such that it can be expanded as
V (q + y) =s=∞∑s=0
(y/2)s
s!
∂sV
∂qs, (A.13)
86
A.1. WIGNER FUNCTION DERIVATION
then Eq.(A.12) is transformed into
− 1
π~
∫ ∞−∞
∑s∈Nodd
(y/2)s
s!
∂sV
∂qsψ∗(q +
y
2)ψ(q − y
2)eipy/~dy. (A.14)
Using the fact that ys = (~i)s ∂
seipy/~
∂psEq.(A.14) can be simplified as
− 1
π~∑s∈Nodd
(~2i
)s1
s!
∂sV
∂qs
∫ ∞−∞
ψ∗(q +y
2)ψ(q − y
2)eipy/~dy =
− 2∑s∈Nodd
(~2i
)s1
s!
∂sV
∂qs∂s
∂ps1
2π~
∫ ∞−∞
ψ∗(q +y
2)ψ(q − y
2)eipy/~dy =
− 2∑s∈Nodd
(~2i
)s1
s!
∂sV
∂qs∂s
∂psW. (A.15)
Finally combining Eq.(A.8) and Eq.(A.15) with the corresponding prefactors
in Eq.(A.3) results in
∂W
∂t+p
m∂qW (q, p)−
∑s∈Nodd
(~2i
)s−11
s!
∂sV
∂qs∂sW
∂ps= 0,
∂W
∂t+p
m∂qW (q, p) + Θ[V ]W = 0,
(A.16)
where
Θ[V ]W = −∑s∈Nodd
(~2i
)s−11
s!
∂sV
∂qs∂sW
∂ps. (A.17)
Alternatively, the Wigner transform of the potential term in (A.3)
V [p, q] =1
2π~i
~
∫ ∞−∞〈q − y/2[V , ρ]q + y/2〉eipy/~dy, (A.18)
can be evaluated as follows, using the result from (A.11)
V [p, q] =1
2π~i
~
∫ ∞−∞
(V (q − y
2)− V (q +
y
2))ψ∗(q +
y
2)ψ(q − y
2)eipy/~dy,
(A.19)
or more generally
87
APPENDIX
V [p, q] =1
2π~i
~
∫ ∞−∞
(V (q − y
2)− V (q +
y
2))ρ(q−y
2, q+
y
2)eipy/~dy. (A.20)
Using the inverse representation of the Wigner function∫ ∞−∞
dpe−ipy/~W (q, p)dp = ρ(q − y/2, q + y/2), (A.21)
V [p, q] takes the form
V [p, q] =1
2π~i
~
∫ ∞−∞
(V (q − y
2)− V (q +
y
2))W (q, p′)eiy(p−p′)/~dydp′.
(A.22)
The final result is then:
∂W
∂t+p
m∂qW (q, p)
+1
2π~i
~
∫ ∞−∞
(V (q − y
2)− V (q +
y
2))W (q, p′)eiy(p−p′)/~dydp′ = 0.
(A.23)
or if δV is defined as
δ[V ](q, p) =i
2π~2
∫ ∞−∞
(V (q − y/2)− V (q + y/2))eiyp/~dy, (A.24)
then
∂W
∂t+p
m∂qW (q, p) +
∫ ∞−∞
δ[V ](q, p− p′)W (q, p′)dp′ = 0, (A.25)
where the identification
Θ[V ]W =
∫ ∞−∞
δ[V ](q, p− p′)W (q, p′)dp′ (A.26)
can be made.
88
A.2. GENERALIZED HERMITE POLYNOMIALS
A.2 Generalized Hermite polynomials
The physicist Hermite polynomials Hn are defined as
Hn = (−1)nex2 dn
dxne−x
2
(A.27)
and satisfy the orthogonality relation∫RHn(x)Hm(x)e−x
2
dx =√π2nn!δmn, (A.28)
with weight function ω(x) = e−x2.
Upon the change of variables x→ v√2cs
the following relation holds∫RHn
(v√2cs
)Hm
(v√2cs
)e− v2
2c2sdx√2cs
=√π2nn!δmn. (A.29)
Using the definition of the general weight function ω(v; cs) = 1√2πc2s
e− v2
2c2s ,
taking m = n and dividing by the normalization constant it is found that∫R
[1√n!2n
Hn
(v√2cs
)][1√n!2n
Hn
(v√2cs
)]ω(v; cs)dx = 1. (A.30)
The generalized Hermite polynomials Hn(x; cs) orthonormal with respect to
the weight function ω(v; cs) are then defined as
Hn(x; cs) =1√n!2n
Hn
(v√2cs
). (A.31)
From Eq.(A.31) the equivalent expression to Eq.(A.27) is given by
ω(v; cs)Hn(x; cs) = (−1)n(2cs)
n/2
√n!2n
dn
dxnω(v; cs). (A.32)
A.3 Wigner Forcing term calculation
In this section the calculation of the the Wigner forcing term in the Lattice
Kinetic approach is presented for the case of Hermite polynomials. The
derivation is carried out in one dimension, and the general result is provided
at the end.
89
APPENDIX
In dimensionless variables, the Wigner forcing term is given by
Θ[V ]W = −∑s∈Nodd
(H
2i
)s−11
s!
∂sV
∂xs∂sW
∂vs, (A.33)
and the Wigner function is assumed to be represented as
W (x, v, t) = ω(v; cs)
Np∑n
an(x, t)Hn(v; cs), (A.34)
where Hn(v;Cs) are the Hermite polynomials with weight function
ω(v; cs) =1√2πc2
s
e− v2
2c2s . (A.35)
If W is replaced in Eq.(A.33), it is clear that the terms that need to be
evaluated are of the form
∂s
∂vs(ω(v; cs)Hn) . (A.36)
From the generalized Hermite Polynomials definition Eq.(A.32), it follows
that∂s
∂vs(ω(v; cs)Hn) = (−1)n
(2cs)n/2
√n!2n
dn+s
dxn+sω(v; cs). (A.37)
Using Eq.(A.32) again, the result reads
∂s
∂vs(ω(v; cs)Hn) = (−1)s
√(n+ s)!
n!(c2s)
ω(v; cs)Hn+s. (A.38)
Using this result, the Wigner forcing term finally reads
Θ[V ]W = ω(v; cs)∑n,s
an(x, t)
√(n+ s)!
n!
(H/i)s−1
Csss!
∂sV
∂xsHn+s(vi; cs). (A.39)
Given that Eq.(A.39) is already expressed as the weight function ω(v; cs)
times Hermite polynomials, the source distribution follows directly from the
prescription ω(vi; cs)→ ωi and Si = Θ[V ]W (vi).
For dimension D > 1, the procedure that was used in the one dimensional
case can also be used to find Θ[V ]W in terms of the expansion coefficients
90
A.4. GENERAL WIGNER FUNCTION CALCULATION
of the Wigner function. That is, if the Wigner function is assumed to be
represented as
Θ[V ]W = ω(v; cs)∑n
anHn(v) (A.40)
where n = (n1, n2, . . . , nD) is a D dimensional set of non-negative integer
indices ni and Hn is a D dimensional tensor Hermite polynomial defined as
Hn = Hn1 ⊗Hn1 · · · ⊗ HnD , then Θ[V ]W can be written as
Θ[V ]W = ω(v; cs)∑
|s|∈Nodd,n
anK(v,n, s)∂s
∂xsV, (A.41)
where the kernel K is given by
K(v,n, s) =
(H
2i
)|s|−1 [1
(2c2s)|s|/2
∏l sl!
]∏l
Hnl+sl(v). (A.42)
In this case the source term distribution also follows from the prescription
ω(vi; cs)→ ωi and Si = Θ[V ]W (vi).
A.4 General Wigner function calculation
The calculation of the Wigner function can be conveniently performed if the
underlying density matrix is written in terms of Hermite functions (eigen-
functions of the quantum harmonic oscillator) that are defined as
ϕn =1√2nn!
(1
πH
)1/4
e−x2
2HHn
(x√H
). (A.43)
Let the density matrix of a given physical system be given by
ρ(x, x′) =∑n,m
cnc∗mϕn(x)ϕm(x). (A.44)
By definition the Wigner function is given by
W =1
2πH
∑n,m
cnc∗m
∫ϕn
(x− y
2
)ϕm
(x+
y
2
)eivyH dy, (A.45)
=1
2πH
∑n,m
cnc∗mKnm(x, v), (A.46)
=1
2πH
∑n≤m
2
1 + δm,n<(cnc
∗mkn,m), (A.47)
91
APPENDIX
where the Knm are found using Groenewold’s formula [24] and are given by
km,n =2(−1)min (m,n)
√min (m,n)!
max (m,n)!e−
x2+v2
H
(2
H(x2 + v2)
) |m−n|2
L|m−n|min (m,n)
(2
H(x2 + v2)
)e(i(m−n) arctan (v/x)) (A.48)
where Lmn is the m order n degree associated Laguerre polynomial.
At this point it is important to mention that there is a mistake, possibly
due by a typo, in Ref [24] where it is stated that Lmn is a Legendre polyno-
mial instead of a Laguerre one. In the following it is shown that the right
polynomial is the later one.
The starting point is the second line of Eq (5.16) on page 457 of Ref [24],
from which the relevant part is
min (a,b)∑k=0
(−1)k
(m− k)!(n− k)!k!
(~2
(p2 + q2)
)min(m,n)−k
, (A.49)
for simplicity let z =(~
2(p2 + q2)
)and let S denote the sum on Eq.(A.49).
Without loss of generality let m > n thus
S =n∑k=0
(−1)k
(m− k)!(n− k)!k!zn−k, (A.50)
using the change of variables u = n− k, S can be written as
S = (−1)nn∑u=0
(−1)u
(m− n+ u)!(n− u)!(u)!zu, (A.51)
with the help of the identity S = Sm!m!
, S is transformed into
1
m!
[n∑u
(−1)u(
m
n− u
)zu
u!
]=
(−1)n
m!Lmn . (A.52)
Note that the expression in square brackets corresponds to the definition of
Laguerre (see Ref.[158]) and not Legendre polynomials.
92
A.5. MULTIPLE BARRIERS EFFECTIVE TRANSMISSIONCOEFFICIENT
A.5 Multiple barriers effective transmission
coefficient
Let T1, T2 be the transmission coefficients of two barriers for a given incoming
state ψ. The effective transmission coefficient T12 is given by the probability
that the incoming state passes through the barriers without reflection, with
two reflections, with four reflections..., since each of these events are exclusive
the total probability is given by the sum of the individual events probabilities
i.e.
T12 = T1T2 + T1T2R1R2 + T1T2(R1R2)2 + · · · =∞∑i=0
T1T2(R1R2)i, (A.53)
where Ri = 1− Ti. Since |R1R2| < 1 the series can be summed leading to
T12 =T1T2
1−R1R2
, (A.54)
taking the reciprocal of the last equation and using again that Ri = 1− Ti,it follows:
1
T12
=1
T1
+1
T2
− 1, (A.55)
that can be casted in the form
1− T12
T12
=1− T1
T1
+1− T2
T2
, (A.56)
Thus for N identical barriers the result is
TN =T
N(1− T ) + T. (A.57)
A.6 Lattice specification
Here the full specification of the lattices used through this work is given. 1D
lattices where generated as described in the main text. 2D lattices and in
general n dimensional Hermite based lattices can be constructed by taking
n times the tensor product of the set of vectors and weights of a fixed
93
APPENDIX
1D Lattice. For example the D2Q4 lattice is given by Table.A.6. It is
important to notice that this way of building higher dimensional lattices
does not exhaust all possible lattices.
Table A.1: D1Q4 Lattice with Cs = 0.60625445810016454
vi wi
0 0.63664690312607816284434609283846
-1,1 0.18141458774368577505004149208377
-3,3 0.00026196069327514352778546149699
Table A.2: D1Q8 Lattice with Cs = 1.0658132602705641
vi wi
0 0.37428019874212190129215011724318
-1,1 0.24105344284458452784844296921093
-2,2 0.06434304152476086575379872184362
-3,3 0.00713156628791277339406557854605
-4,4 0.00032523057375714836476726255033
-5,5 6.6163470389851878681133911638949×10−6
-7,7 3.0508847488049822958363118638543×10−9
Table A.3: D1Q10 Lattice with Cs = 1.229594448425497
vi wi
0 0.32444899174631946866086595194671
-1,1 0.23309081165504033632566413700874
-2,2 0.08642582836940192624063184539752
-3,3 0.01653989847863324979993319254793
-4,4 0.00163342485156222352004541584861
-5,5 0.00008333063878279730921268566542
-6,6 2.1783167706100240902344965225688×10−6
-7,7 3.1805869765623071575130276965860×10−8
-9,9 1.0779356826917937931616055896767×10−11
94
A.6. LATTICE SPECIFICATION
Table A.4: D1Q16 Lattice with Cs = 1.6215048099592275
vi wi
0 0.24603212869787232483785340883852
-1,1 0.20342468717937742901117526034797
-2,2 0.11498446042457243913866706495342
-3,3 0.04443225067964028999644337006636
-4,4 0.01173764938741580915572505702913
-5,5 0.00211976456798849884644315007219
-6,6 0.00026170845228301249011385925086
-7,7 0.00002208877826469659955769726449
-8,8 1.2745253026359480126112714680367×10−6
-9,9 5.0275261810959383411581297192576×10−8
-10,10 1.3556297819769484757032262002820×10−9
-11,11 2.5012031341031852003279373539252×10−11
-12,12 3.1243604817078012360750317883072×10−13
-13,13 2.9655118189640940365948400709026×10−15
-15,15 5.5758174181938354491800200913102×10−19
95
APPENDIX
Table A.5: D1Q20 Lattice with Cs = 1.8357424381402594
vi wi
0 0.21731931022112109059537537887018
-1,1 0.18735357499686018912399983787195
-2,2 0.12004746243830897823022161375249
-3,3 0.05717041140835294313179190148076
-4,4 0.02023564183037203154174508370450
-5,5 0.00532341082536521716813053993040
-6,6 0.00104085519989277817787032717819
-7,7 0.00015125787069729717011289372472
-8,8 0.00001633702528266419558030012576
-9,9 1.3114608069909806258412013955825×10−6
-10,10 7.8246508661616857867473666191193×10−8
-11,11 3.4697808952346470123102609636720×10−9
-12,12 1.1435779630395075964965610648417×10−10
-13,13 2.8013182217362421082623834683491×10−12
-14,14 5.0995884226301388644438757982605×10−14
-15,15 6.9079520892785667788901676695952×10−16
-16,16 6.8680470174442627832690379600090×10−18
-17,17 5.7551467186859264824045886746476×10−20
-19,19 8.3761764243303081227469285304101×10−24
Table A.6: D2Q4 Lattice with Cs = 0.60625445810016454
vi wi
(0,0) w20
(0,± 1),(± 1,0) w0w1
(± 1,± 1) w1w1
(0,± 3)(± 3,0) w0w3
(± 3,± 1),(± 1,± 3) w1w3
(± 3,±3) w23
96
References
[1] M. Mendoza, S. Succi, and H. J. Herrmann, “Kinetic formulation of the
kohn-sham equations for ab initio electronic structure calculations,”
Phys. Rev. Lett., vol. 113, p. 096402, Aug 2014.
[2] D. R. Hartree, “The wave mechanics of an atom with a non-coulomb
central field. part i. theory and methods,” Mathematical Proceedings
of the Cambridge Philosophical Society, vol. 24, no. 1, p. 89110, 1928.
[3] V. Fock, “Naherungsmethode zur losung des quantenmechanischen
mehrkorperproblems,” Zeitschrift fur Physik, vol. 61, no. 1-2, pp. 126–
148, 1930.
[4] E. A. Hylleraas, “Neue berechnung der energie des heliums im
grundzustande, sowie des tiefsten terms von ortho-helium,” Zeitschrift
fur Physik, vol. 54, pp. 347–366, May 1929.
[5] E. A. Hylleraas, “Uber den grundterm der zweielektronenprobleme von
h, he, li+, be++ usw.,” Zeitschrift fur Physik, vol. 65, pp. 209–225,
Mar 1930.
[6] E. W. Schmid, G. Spitz, and W. Losch, The Ground State of the He-
lium Atom by the Hylleraas Method, pp. 165–180. Berlin, Heidelberg:
Springer Berlin Heidelberg, 1990.
[7] P. Schwerdtfeger, “The pseudopotential approximation in electronic
structure theory,” ChemPhysChem, vol. 12, no. 17, pp. 3143–3155,
2011.
97
REFERENCES
[8] F. Nogueira, A. Castro, and M. A. L. Marques, A Tutorial on Density
Functional Theory, pp. 218–256. Berlin, Heidelberg: Springer Berlin
Heidelberg, 2003.
[9] D. Vanderbilt, “Soft self-consistent pseudopotentials in a generalized
eigenvalue formalism,” Phys. Rev. B, vol. 41, pp. 7892–7895, Apr 1990.
[10] S. Goedecker, M. Teter, and J. Hutter, “Separable dual-space gaussian
pseudopotentials,” Phys. Rev. B, vol. 54, pp. 1703–1710, Jul 1996.
[11] C. Hartwigsen, S. Goedecker, and J. Hutter, “Relativistic separable
dual-space gaussian pseudopotentials from h to rn,” Phys. Rev. B,
vol. 58, pp. 3641–3662, Aug 1998.
[12] K. F. Garrity, J. W. Bennett, K. M. Rabe, and D. Vanderbilt,
“Pseudopotentials for high-throughput dft calculations,” Computa-
tional Materials Science, vol. 81, pp. 446 – 452, 2014.
[13] S. R. White, “Density matrix formulation for quantum renormalization
groups,” Phys. Rev. Lett., vol. 69, pp. 2863–2866, Nov 1992.
[14] W. M. C. Foulkes, L. Mitas, R. J. Needs, and G. Rajagopal, “Quantum
monte carlo simulations of solids,” Rev. Mod. Phys., vol. 73, pp. 33–83,
Jan 2001.
[15] W. G. Aulbur, L. Jnsson, and J. W. Wilkins, “Quasiparticle calcula-
tions in solids,” Solid State Physics, vol. 54, pp. 1 – 218, 2000.
[16] A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg, “Dynamical
mean-field theory of strongly correlated fermion systems and the limit
of infinite dimensions,” Rev. Mod. Phys., vol. 68, pp. 13–125, Jan 1996.
[17] P. Hohenberg and W. Kohn, “Inhomogeneous electron gas,” Phys.
Rev., vol. 136, pp. B864–B871, Nov 1964.
[18] W. Kohn and L. J. Sham, “Self-consistent equations including ex-
change and correlation effects,” Nov 1965.
98
REFERENCES
[19] J. VandeVondele, U. Bortnik, and J. Hutter, “Linear scaling self-
consistent field calculations with millions of atoms in the condensed
phase,” Journal of Chemical Theory and Computation, vol. 8, no. 10,
pp. 3565–3573, 2012. PMID: 26593003.
[20] R. M. Martin, Electronic Structure Basic Theory and Practical Meth-
ods. United Kingdom: Cambridge University Press, 2004.
[21] D. F. Styer, M. S. Balkin, K. M. Becker, M. R. Burns, C. E. Dudley,
S. T. Forth, J. S. Gaumer, M. A. Kramer, D. C. Oertel, L. H. Park,
M. T. Rinkoski, C. T. Smith, and T. D. Wotherspoon, “Nine formu-
lations of quantum mechanics,” American Journal of Physics, vol. 70,
no. 3, pp. 288–297, 2002.
[22] E. Wigner, “On the quantum correction for thermodynamic equilib-
rium,” Phys. Rev., vol. 40, pp. 749–759, Jun 1932.
[23] T. L. Curtright, D. B. Fairlie, and C. K. Zachos, A concise treatise
on quantum mechanics in phase space. World Scientific Publishing Co
Inc, 2013.
[24] H. Groenewold, “On the principles of elementary quantum mechanics,”
Physica, vol. 12, no. 7, pp. 405 – 460, 1946.
[25] J. E. Moyal, “Quantum mechanics as a statistical theory,” in Math-
ematical Proceedings of the Cambridge Philosophical Society, vol. 45,
pp. 99–124, Cambridge University Press, 1949.
[26] M. Hillery, R. O’Connell, M. Scully, and E. Wigner, “Distribution
functions in physics: Fundamentals,” Physics Reports, vol. 106, no. 3,
pp. 121 – 167, 1984.
[27] H.-W. Lee, “Theory and application of the quantum phase-space dis-
tribution functions,” Physics Reports, vol. 259, no. 3, pp. 147–211,
1995.
99
REFERENCES
[28] L. Cohen, “Generalized phase-space distribution functions,” Journal
of Mathematical Physics, vol. 7, no. 5, pp. 781–786, 1966.
[29] C. Mehta, “Phase-space formulation of the dynamics of canonical vari-
ables,” Journal of Mathematical Physics, vol. 5, no. 5, pp. 677–686,
1964.
[30] J. G. Kirkwood, “Quantum statistics of almost classical assemblies,”
Physical Review, vol. 44, no. 1, p. 31, 1933.
[31] A. Rihaczek, “Signal energy distribution in time and frequency,” IEEE
Transactions on Information Theory, vol. 14, no. 3, pp. 369–374, 1968.
[32] R. J. Glauber, “Coherent and incoherent states of the radiation field,”
Physical Review, vol. 131, no. 6, p. 2766, 1963.
[33] E. Sudarshan, “Equivalence of semiclassical and quantum mechani-
cal descriptions of statistical light beams,” Physical Review Letters,
vol. 10, no. 7, p. 277, 1963.
[34] K. Husimi, “Some formal properties of the density matrix,” Proceed-
ings of the Physico-Mathematical Society of Japan. 3rd Series, vol. 22,
no. 4, pp. 264–314, 1940.
[35] M. V. Berry, “Semi-classical mechanics in phase space: a study of
wigner’s function,” Philosophical Transactions of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences, vol. 287,
no. 1343, pp. 237–271, 1977.
[36] K. Takahashi, “Distribution functions in classical and quantum me-
chanics,” Progress of Theoretical Physics Supplement, vol. 98, pp. 109–
156, 1989.
[37] S. A. Gardiner, D. Jaksch, R. Dum, J. I. Cirac, and P. Zoller, “Non-
linear matter wave dynamics with a chaotic potential,” Phys. Rev. A,
vol. 62, p. 023612, Jul 2000.
100
REFERENCES
[38] W. H. Zurek, “Sub-planck structure in phase space and its relevance
for quantum decoherence,” Nature, vol. 412, pp. 712–717, 08 2001.
[39] M. A. Alonso, “Wigner functions in optics: describing beams as ray
bundles and pulses as particle ensembles,” Adv. Opt. Photon., vol. 3,
pp. 272–365, Dec 2011.
[40] D. Leibfried, T. Pfau, and C. Monroe, “Shadows and Mirrors: Recon-
structing Quantum States of Atom Motion,” Physics Today, vol. 51,
pp. 22–28, Apr. 1998.
[41] A. Stern, “Noncommutative point sources,” Phys. Rev. Lett., vol. 100,
p. 061601, Feb 2008.
[42] L. P. Kadanoff, “Computational physics: Pluses and minuses,” Physics
Today, vol. 39, no. 7, pp. 7–9, 1986.
[43] S. Wolfram, “Cellular automaton fluids 1: Basic theory,” Journal of
Statistical Physics, vol. 45, no. 3, pp. 471–526, 1986.
[44] D. d’Humires, P. Lallemand, and U. Frisch, “Lattice gas models for
3d hydrodynamics,” EPL (Europhysics Letters), vol. 2, no. 4, p. 291,
1986.
[45] U. Frisch, B. Hasslacher, and Y. Pomeau, “Lattice-gas automata for
the navier-stokes equation,” Phys. Rev. Lett., vol. 56, pp. 1505–1508,
Apr 1986.
[46] S. Succi, The lattice Boltzmann equation: for fluid dynamics and be-
yond. Oxford university press, 2001.
[47] G. R. McNamara and G. Zanetti, “Use of the boltzmann equation to
simulate lattice-gas automata,” Phys. Rev. Lett., vol. 61, pp. 2332–
2335, Nov 1988.
[48] F. J. Higuera and J. Jimnez, “Boltzmann approach to lattice gas sim-
ulations,” EPL (Europhysics Letters), vol. 9, no. 7, p. 663, 1989.
101
REFERENCES
[49] P. L. Bhatnagar, E. P. Gross, and M. Krook, “A model for collision
processes in gases. i. small amplitude processes in charged and neutral
one-component systems,” Phys. Rev., vol. 94, pp. 511–525, May 1954.
[50] Y. H. Qian, “Lattice gas and lattice kinetic theory applied to the
navier-stokes equations,” Doktorarbeit, Universite Pierre et Marie
Curie, Paris, 1990.
[51] S. Chen, H. Chen, D. Martnez, and W. Matthaeus, “Lattice boltzmann
model for simulation of magnetohydrodynamics,” Phys. Rev. Lett.,
vol. 67, pp. 3776–3779, Dec 1991.
[52] X. He and L.-S. Luo, “A priori derivation of the lattice boltzmann
equation,” Phys. Rev. E, vol. 55, pp. R6333–R6336, Jun 1997.
[53] T. Abe, “Derivation of the lattice boltzmann method by means of
the discrete ordinate method for the boltzmann equation,” Journal of
Computational Physics, vol. 131, no. 1, pp. 241 – 246, 1997.
[54] F. Bosch and I. V. Karlin, “Exact lattice boltzmann equation,” Phys.
Rev. Lett., vol. 111, p. 090601, Aug 2013.
[55] R. Mei, W. Shyy, D. Yu, and L.-S. Luo, “Lattice boltzmann method for
3-d flows with curved boundary,” Journal of Computational Physics,
vol. 161, no. 2, pp. 680 – 699, 2000.
[56] S. Melchionna, M. Bernaschi, S. Succi, E. Kaxiras, F. J. Rybicki,
D. Mitsouras, A. U. Coskun, and C. L. Feldman, “Hydrokinetic ap-
proach to large-scale cardiovascular blood flow,” Computer Physics
Communications, vol. 181, no. 3, pp. 462 – 472, 2010.
[57] R. W. Nash, H. B. Carver, M. O. Bernabeu, J. Hetherington, D. Groen,
T. Kruger, and P. V. Coveney, “Choice of boundary condition for
lattice-boltzmann simulation of moderate-reynolds-number flow in
complex domains,” Phys. Rev. E, vol. 89, p. 023303, Feb 2014.
102
REFERENCES
[58] G. Amati, S. Succi, and R. Piva, “Massively parallel lattice-boltzmann
simulation of turbulent channel flow,” International Journal of Modern
Physics C, vol. 8, no. 04, pp. 869–877, 1997.
[59] J. Gotz, K. Iglberger, M. Sturmer, and U. Rude, “Direct numerical
simulation of particulate flows on 294912 processor cores,” in High Per-
formance Computing, Networking, Storage and Analysis (SC), 2010
International Conference for, pp. 1–11, IEEE, 2010.
[60] M. Mendoza, S. Succi, and H. Herrmann, “Flow through randomly
curved manifolds,” Scientific reports, vol. 3, p. 3106, 2013.
[61] J.-D. Debus, M. Mendoza, S. Succi, and H. J. Herrmann, “Poiseuille
flow in curved spaces,” Phys. Rev. E, vol. 93, p. 043316, Apr 2016.
[62] M. Mendoza, H. J. Herrmann, and S. Succi, “Preturbulent regimes in
graphene flow,” Phys. Rev. Lett., vol. 106, p. 156601, Apr 2011.
[63] I. Giordanelli, M. Mendoza, and H. Herrmann, “Modelling electron-
phonon interactions in graphene with curved space hydrodynamics,”
arXiv preprint arXiv:1702.04156, 2017.
[64] O. Furtmaier, M. Mendoza, I. Karlin, S. Succi, and H. J. Herrmann,
“Rayleigh-benard instability in graphene,” Physical Review B, vol. 91,
no. 8, p. 085401, 2015.
[65] M. Mendoza and J. D. Munoz, “Three-dimensional lattice boltzmann
model for electrodynamics,” Phys. Rev. E, vol. 82, p. 056708, Nov
2010.
[66] S. Succi and R. Benzi, “Lattice boltzmann equation for quantum me-
chanics,” Physica D: Nonlinear Phenomena, vol. 69, no. 3, pp. 327 –
332, 1993.
[67] M. Mendoza, B. Boghosian, H. Herrmann, and S. Succi, “Derivation of
the lattice boltzmann model for relativistic hydrodynamics,” Physical
Review D, vol. 82, no. 10, p. 105008, 2010.
103
REFERENCES
[68] E. Ilseven and M. Mendoza, “Lattice boltzmann model for numerical
relativity,” Phys. Rev. E, vol. 93, p. 023303, Feb 2016.
[69] S. Succi, “Lattice boltzmann 2038,” EPL (Europhysics Letters),
vol. 109, no. 5, p. 50001, 2015.
[70] A. Einstein, “ber die von der molekularkinetischen theorie der
wrme geforderte bewegung von in ruhenden flssigkeiten suspendierten
teilchen,” Ann. Phys(Leipzig), vol. 322, no. 8, pp. 549–560, 1905.
[71] M. von Smoluchowski, “ber die von der molekularkinetischen theo-
rie der wrme geforderte bewegung von in ruhenden flssigkeiten sus-
pendierten teilchen,” Phys. Z, vol. XIII, no. 13, p. 1069, 1912.
[72] K. Maruyama, F. Nori, and V. Vedral, “Colloquium,” Rev. Mod. Phys.,
vol. 81, pp. 1–23, Jan 2009.
[73] R. Feynman, R. B. Leighton, and M. L. Sands, The Feynman Lectures
on Physics. Addison-Wesley, 1963. 3 volumes.
[74] P. Reimann, “Brownian motors: noisy transport far from equilibrium,”
Phys. Rep., vol. 361, no. 24, pp. 57 – 265, 2002.
[75] P. Hanggi, F. Marchesoni, and F. Nori, “Brownian motors,”
Ann.Phys(Leipzig), vol. 14, no. 1-3, pp. 51–70, 2005.
[76] J. Howard, “Molecular motors: structural adaptations to cellular func-
tions,” Nature, vol. 389, no. 6651, p. 561, 1997.
[77] K. Svoboda, C. F. Schmidt, B. J. Schnapp, and S. M. Block, “Direct
observation of kinesin stepping by optical trapping interferometry,”
Nature, vol. 365, no. 6448, pp. 721–727, 1993.
[78] I. Derenyi and R. D. Astumian, “ac separation of particles by biased
brownian motion in a two-dimensional sieve,” Phys. Rev. E, vol. 58,
pp. 7781–7784, Dec 1998.
104
REFERENCES
[79] C. Kettner, P. Reimann, P. Hanggi, and F. Muller, “Drift ratchet,”
Phys. Rev. E, vol. 61, pp. 312–323, Jan 2000.
[80] S. Matthias and F. Muller, “Asymmetric pores in a silicon mem-
brane acting as massively parallel brownian ratchets,” Nature, vol. 424,
no. 6944, pp. 53–57, 2003.
[81] J. Rousselet, L. Salome, A. Ajdari, and J. Prostt, “Directional motion
of brownian particles induced by a periodic asymmetric potential,”
Nature, vol. 370, pp. 446–447, Aug. 1994.
[82] C. Marquet, A. Buguin, L. Talini, and P. Silberzan, “Rectified motion
of colloids in asymmetrically structured channels,” Phys. Rev. Lett.,
vol. 88, p. 168301, Apr 2002.
[83] L. P. Faucheux, L. S. Bourdieu, P. D. Kaplan, and A. J. Libchaber,
“Optical thermal ratchet,” Phys. Rev. Lett., vol. 74, pp. 1504–1507,
Feb 1995.
[84] J. Prost, J.-F. Chauwin, L. Peliti, and A. Ajdari, “Asymmetric pump-
ing of particles,” Phys. Rev. Lett., vol. 72, pp. 2652–2655, Apr 1994.
[85] S. Mobarakabadi, E. N. Oskoee, M. Schroter, and M. Habibi, “Granu-
lar transport in a horizontally vibrated sawtooth channel,” Phys. Rev.
E, vol. 88, p. 042201, Oct 2013.
[86] G. Costantini, U. Marini Bettolo Marconi, and A. Puglisi, “Granular
brownian ratchet model,” Phys. Rev. E, vol. 75, p. 061124, Jun 2007.
[87] B. Cleuren and C. V. den Broeck, “Granular brownian motor,” EPL,
vol. 77, no. 5, p. 50003, 2007.
[88] C. Van den Broeck, R. Kawai, and P. Meurs, “Microscopic analysis of
a thermal brownian motor,” Phys. Rev. Lett., vol. 93, p. 090601, Aug
2004.
105
REFERENCES
[89] P. Eshuis, K. van der Weele, D. Lohse, and D. van der Meer, “Ex-
perimental realization of a rotational ratchet in a granular gas,” Phys.
Rev. Lett., vol. 104, p. 248001, Jun 2010.
[90] A. Gnoli, A. Petri, F. Dalton, G. Pontuale, G. Gradenigo, A. Sarracino,
and A. Puglisi, “Brownian ratchet in a thermal bath driven by coulomb
friction,” Phys. Rev. Lett., vol. 110, p. 120601, Mar 2013.
[91] R. Balzan, F. Dalton, V. Loreto, A. Petri, and G. Pontuale, “Brownian
motor in a granular medium,” Phys. Rev. E, vol. 83, p. 031310, Mar
2011.
[92] L. Angelani, R. Di Leonardo, and G. Ruocco, “Self-starting micromo-
tors in a bacterial bath,” Phys. Rev. Lett., vol. 102, p. 048104, Jan
2009.
[93] R. Di Leonardo, L. Angelani, D. DellArciprete, G. Ruocco, V. Iebba,
S. Schippa, M. P. Conte, F. Mecarini, F. De Angelis, and E. Di Fab-
rizio, “Bacterial ratchet motors,” Proceedings of the National Academy
of Sciences, vol. 107, no. 21, pp. 9541–9545, 2010.
[94] A. Sokolov, M. M. Apodaca, B. A. Grzybowski, and I. S. Aranson,
“Swimming bacteria power microscopic gears,” Proceedings of the Na-
tional Academy of Sciences, vol. 107, no. 3, pp. 969–974, 2010.
[95] H. Li and H. P. Zhang, “Asymmetric gear rectifies random robot mo-
tion,” EPL, vol. 102, no. 5, p. 50007, 2013.
[96] C. L. N. Oliveira, A. P. Vieira, D. Helbing, J. S. Andrade, and H. J.
Herrmann, “Keep-left behavior induced by asymmetrically profiled
walls,” Phys. Rev. X, vol. 6, p. 011003, Jan 2016.
[97] V. Holubec, A. Ryabov, M. H. Yaghoubi, M. Varga, A. Khodaee, M. E.
Foulaadvand, and P. Chvosta, “Thermal ratchet effect in confining
geometries,” Entropy, vol. 19, no. 4, p. 119, 2017.
106
REFERENCES
[98] A. Ryabov, V. Holubec, M. H. Yaghoubi, M. Varga, M. E. Foulaad-
vand, and P. Chvosta, “Transport coefficients for a confined brownian
ratchet operating between two heat reservoirs,” Journal of Statistical
Mechanics: Theory and Experiment, vol. 2016, no. 9, p. 093202, 2016.
[99] A. Sarracino, A. Gnoli, and A. Puglisi, “Ratchet effect driven by
coulomb friction: The asymmetric rayleigh piston,” Phys. Rev. E,
vol. 87, p. 040101, Apr 2013.
[100] G. Costantini, U. M. B. Marconi, and A. Puglisi, “Noise rectification
and fluctuations of an asymmetric inelastic piston,” EPL (Europhysics
Letters), vol. 82, no. 5, p. 50008, 2008.
[101] J. Talbot and P. Viot, “Effect of dynamic and static friction on an
asymmetric granular piston,” Phys. Rev. E, vol. 85, p. 021310, Feb
2012.
[102] A. Sarracino, “Time asymmetry of the kramers equation with nonlin-
ear friction: Fluctuation-dissipation relation and ratchet effect,” Phys.
Rev. E, vol. 88, p. 052124, Nov 2013.
[103] A. Baule and P. Sollich, “Singular features in noise-induced transport
with dry friction,” EPL (Europhysics Letters), vol. 97, no. 2, p. 20001,
2012.
[104] X. He and L.-S. Luo, “Theory of the lattice boltzmann method: From
the boltzmann equation to the lattice boltzmann equation,” Phys. Rev.
E, vol. 56, pp. 6811–6817, Dec 1997.
[105] F. B. Hildebrand, Introduction to Numerical Analysis. Unites States
of America: McGraw-Hill, 1974.
[106] S. P. Thampi, S. Ansumali, R. Adhikari, and S. Succi, “Isotropic
discrete laplacian operators from lattice hydrodynamics,” Journal of
Computational Physics, vol. 234, pp. 1 – 7, 2013.
107
REFERENCES
[107] R. Bulirsch and J. Stoer, Introduction to Numerical Analysis. Berlin-
Heidelberg: Springer, 2011.
[108] S. S. Chikatamarla and I. V. Karlin, “Entropy and galilean invariance
of lattice boltzmann theories,” Phys. Rev. Lett., vol. 97, p. 190601,
Nov 2006.
[109] S. S. Chikatamarla and I. V. Karlin, “Lattices for the lattice boltzmann
method,” Phys. Rev. E, vol. 79, p. 046701, Apr 2009.
[110] P. C. Philippi, L. A. Hegele, L. O. E. dos Santos, and R. Surmas, “From
the continuous to the lattice boltzmann equation: The discretization
problem and thermal models,” Phys. Rev. E, vol. 73, p. 056702, May
2006.
[111] J. Zhang, G. Yan, and X. Shi, “Lattice boltzmann model for wave
propagation,” Phys. Rev. E, vol. 80, p. 026706, Aug 2009.
[112] M. Mendoza, I. Karlin, S. Succi, and H. J. Herrmann, “Relativistic
lattice boltzmann model with improved dissipation,” Phys. Rev. D,
vol. 87, p. 065027, Mar 2013.
[113] J. D. Debus, M. Mendoza, S. Succi, and H. J. Herrmann, “Energy
dissipation in flows through curved spaces,” Scientific Reports, vol. 7,
pp. 42350 EP –, 02 2017.
[114] P. N. Ma, S. Pilati, M. Troyer, and X. Dai, “Density functional theory
for atomic fermi gases,” Nat Phys, vol. 8, pp. 601–605, 08 2012.
[115] E. Engel and R. M. Dreizler, Density Functional Theory an advanced
course. Berlin-Heidelberg: Springer, 2011.
[116] R. G. Parr and Y. Weitao, Density-Functional Theory of Atoms and
Molecules. Oxford University Press, 1989.
[117] J. Kohanoff, Electronic Structure Calculations for Solids and
Moleculer. United Kingdom: Cambridge University Press, 2006.
108
REFERENCES
[118] J. P. Perdew and K. Schmidt, “Jacobs ladder of density functional
approximations for the exchange-correlation energy,” AIP Conference
Proceedings, vol. 577, no. 1, pp. 1–20, 2001.
[119] A. D. Becke, “Density-functional physrev.35.210.2exchange-energy ap-
proximation with correct asymptotic behavior,” Phys. Rev. A, vol. 38,
pp. 3098–3100, Sep 1988.
[120] C. Lee, W. Yang, and R. G. Parr, “Development of the colle-salvetti
correlation-energy formula into a functional of the electron density,”
Phys. Rev. B, vol. 37, pp. 785–789, Jan 1988.
[121] S. Chapman and T. Cowling, The Mathematical Theory of Non-
Uniform Gases: An Account of the Kinetic Theory of Viscosity, Ther-
mal Conduction and Diffusion in Gases. Cambridge, England: Cam-
bridge University Press, 1970.
[122] NIST, “Nist computational chemistry comparison and benchmark
database standard reference database number 101,” 2015. [On-
line,Release 17b, September 2015, Editor: Russell D. Johnson III ;
accessed 10-November-2015].
[123] M. Casula, S. Moroni, S. Sorella, and C. Filippi, “Size-consistent vari-
ational approaches to nonlocal pseudopotentials: Standard and lattice
regularized diffusion monte carlo methods revisited,” The Journal of
Chemical Physics, vol. 132, no. 15, 2010.
[124] L. Mit, E. L. Shirley, and D. M. Ceperley, “Nonlocal pseudopotentials
and diffusion monte carlo,” The Journal of Chemical Physics, vol. 95,
no. 5, 1991.
[125] S. Fahy, X. W. Wang, and S. G. Louie, “Variational quantum monte
carlo nonlocal pseudopotential approach to solids: Cohesive and struc-
tural properties of diamond,” Phys. Rev. Lett., vol. 61, pp. 1631–1634,
Oct 1988.
109
REFERENCES
[126] O. Furtmaier, S. Succi, and M. Mendoza, “Semi-spectral method for
the wigner equation,” Journal of Computational Physics, vol. 305,
pp. 1015 – 1036, 2016.
[127] B. Shi, B. Deng, R. Du, and X. Chen, “A new scheme for source
term in LBGK model for convectiondiffusion equation,” Computers
& Mathematics with Applications, vol. 55, no. 7, pp. 1568 – 1575, 2008.
Mesoscopic Methods in Engineering and Science.
[128] A. Jungel, Transport Equations for Semiconductors. Berlin-
Heidelberg: Springer, 2009.
[129] R. Rosati, F. Dolcini, R. C. Iotti, and F. Rossi, “Wigner-function
formalism applied to semiconductor quantum devices: Failure of the
conventional boundary condition scheme,” Phys. Rev. B, vol. 88,
p. 035401, Jul 2013.
[130] W. B. Case, “Wigner functions and weyl transforms for pedestrians,”
American Journal of Physics, vol. 76, no. 10, pp. 937–946, 2008.
[131] H. Grad, “On the kinetic theory of rarefied gases,” Communications
on Pure and Applied Mathematics, vol. 2, no. 4, pp. 331–407, 1949.
[132] M. M. Mizrahi, “Generalized hermite polynomials,” Journal of Com-
putational and Applied Mathematics, vol. 1, no. 4, pp. 273 – 277, 1975.
[133] A. Dorda and F. Schrrer, “A weno-solver combined with adaptive
momentum discretization for the wigner transport equation and its
application to resonant tunneling diodes,” Journal of Computational
Physics, vol. 284, pp. 95 – 116, 2015.
[134] M. Mendoza, S. Succi, and H. J. Herrmann, “Flow through randomly
curved manifolds,” Scientific Reports, vol. 3, pp. 3106 EP –, 10 2013.
[135] W. Lu and C. M. Lieber, “Nanoelectronics from the bottom up,” Nat
Mater, vol. 6, pp. 841–850, 11 2007.
110
REFERENCES
[136] R. McConnell, H. Zhang, J. Hu, S. Cuk, and V. Vuletic, “Entangle-
ment with negative wigner function of almost 3,000 atoms heralded by
one photon,” Nature, vol. 519, pp. 439–442, 03 2015.
[137] N. Gonzalez, G. Molina-Terriza, and J. P. Torres, “Properties of the
spatial wigner function of entangled photon pairs,” Phys. Rev. A,
vol. 80, p. 043804, Oct 2009.
[138] A. Laio, S. Bernard, G. L. Chiarotti, S. Scandolo, and E. Tosatti,
“Physics of iron at earth’s core conditions,” Science, vol. 287, no. 5455,
pp. 1027–1030, 2000.
[139] C. Pinilla, M. Acua-Rojas, N. Seriani, and S. Scandolo, “An atomistic
model of mgsio3 perovskite and post-perovskite phases,” Computa-
tional Materials Science, vol. 126, pp. 351 – 359, 2017.
[140] T.Poschel and T. Schwager, Computational Granular Dynamics.
Springer Berlin Heidelberg, 2005.
[141] M. Karplus and J. A. McCammon, “Molecular dynamics simulations
of biomolecules,” Nat Struct Mol Biol, vol. 9, pp. 646–652, 09 2002.
[142] M. Karplus and G. A. Petsko, “Molecular dynamics simulations in
biology,” Nature, vol. 347, pp. 631–639, 10 1990.
[143] L. Verlet, “Computer ”experiments” on classical fluids. i. thermody-
namical properties of lennard-jones molecules,” Phys. Rev., vol. 159,
pp. 98–103, Jul 1967.
[144] R. W. Hockney, “The Potential Calculation and Some Applications,”
Methods Comput. Phys., vol. 9, pp. 136–210, 1970.
[145] C. W. Gear, Numerical Initial Value Problems in Ordinary Differential
Equations. Upper Saddle River, NJ, USA: Prentice Hall PTR, 1971.
[146] E. Hairer, C. Lubich, and G. Wanner, Geometric numerical integra-
tion: structure-preserving algorithms for ordinary differential equa-
tions, vol. 31. Springer Science & Business Media, 2006.
111
REFERENCES
[147] M.P.Allen and D.J.Tildesley, Computer simulation of liquids. Claren-
don press, 1989.
[148] H. C. Andersen, “Molecular dynamics simulations at constant pressure
and/or temperature,” The Journal of Chemical Physics, vol. 72, no. 4,
pp. 2384–2393, 1980.
[149] T. A. Andrea, W. C. Swope, and H. C. Andersen, “The role of long
ranged forces in determining the structure and properties of liquid
water,” The Journal of Chemical Physics, vol. 79, no. 9, pp. 4576–
4584, 1983.
[150] W. G. Hoover, A. J. C. Ladd, and B. Moran, “High-strain-rate plastic
flow studied via nonequilibrium molecular dynamics,” Phys. Rev. Lett.,
vol. 48, pp. 1818–1820, Jun 1982.
[151] S. Nos, “A molecular dynamics method for simulations in the canonical
ensemble,” Molecular Physics, vol. 52, no. 2, pp. 255–268, 1984.
[152] W. G. Hoover, “Canonical dynamics: Equilibrium phase-space distri-
butions,” Phys. Rev. A, vol. 31, pp. 1695–1697, Mar 1985.
[153] C. Cercignani, “The boltzmann equation,” in The Boltzmann Equation
and Its Applications, pp. 40–103, Springer, 1988.
[154] O. Benichou, P. Illien, G. Oshanin, A. Sarracino, and R. Voituriez,
“Diffusion and subdiffusion of interacting particles on comblike struc-
tures,” Phys. Rev. Lett., vol. 115, p. 220601, Nov 2015.
[155] J. Prost, J. F. Chauwin, L. Peliti, and A. Ajdari, “Asymmetric pump-
ing of particles,” Phys. Rev. Lett., vol. 72, p. 2652, 1994.
[156] A. Tomilov, A. Videcoq, M. Cerbelaud, M. A. Piechowiak, T. Chartier,
T. Ala-Nissila, D. Bochicchio, and R. Ferrando, “Aggregation in col-
loidal suspensions: Evaluation of the role of hydrodynamic interactions
by means of numerical simulations,” The Journal of Physical Chem-
istry B, vol. 117, no. 46, pp. 14509–14517, 2013.
112
REFERENCES
[157] A. Tomilov, A. Videcoq, T. Chartier, T. Ala-Nissil, and I. Vattulainen,
“Tracer diffusion in colloidal suspensions under dilute and crowded
conditions with hydrodynamic interactions,” The Journal of Chemical
Physics, vol. 137, no. 1, pp. –, 2012.
[158] G. B. Arfken, H. J. Weber, and F. E. Harris, “Chapter 18 - more
special functions,” in Mathematical Methods for Physicists (Seventh
Edition) (G. B. Arfken, H. J. Weber, and F. E. Harris, eds.), pp. 871
– 933, Boston: Academic Press, seventh edition ed., 2013.
113