Vous êtes sur la page 1sur 223

Rotating BoseEinstein Condensates

Vortex Lattices and Excitations


Andreas Penckwitt
December 2003
A thesis submitted for the degree of
Doctor of Philosophy
Department of Physics
University of Otago
Dunedin
New Zealand
Revision History
Original submitted for examination: December 23, 2003.
First corrected version (post examination): July 8, 2004.
Abstract
The main theme of this thesis is an investigation of rotating BoseEinstein condensates.
This area is of considerable current interest and has been stimulated by several recent
experiments where vortex lattices are created by stirring a BoseEinstein condensate
with an anisotropic trap or by growth from a rotating thermal cloud. These experiments
echo earlier experiments on superuid
4
He in a rotating bucket which rst led to the
discovery of vortex lattices.
One key question, which remained unanswered from the work on
4
He, is the mecha-
nism of vortex lattice formation. Dilute BoseEinstein condensates allow the possibility
of both detailed experimental studies of the dynamics and a priori theoretical treat-
ments. It is well understood that some form of dissipation is necessary to drive a
condensate into a lattice state, however, this makes the standard method of describing
a condensate, the GrossPitaevskii equation, inadequate for this problem.
In this thesis, we provide a simple, unied treatment that describes the process of
vortex lattice formation consistently, from the initiation to the nal lattice stabilization.
Our work is the rst application of a formalism developed by Gardiner et al. [J. Phys.
B, 35, 1555 (2002)] where the dissipation is provided by an exchange of atoms between
the condensate and the thermal cloud. In its simplest form it can be reduced to a
modied GrossPitaevskii equation with growth and loss terms that provide damping
via a bath of thermal atoms.
We model the scenario of vortex lattice formation from a rotating thermal cloud
and show that the basic mechanism of the formation is growth into surface modes of
the condensate. We give an analytic treatment that provides the gain coecients for
this growth in terms of the excitation energies of the modes, and validate it against
our simulations. In our model, the critical angular velocity is given by the condition
for positive gain and coincides with the Landau criterion.
We show that this simple analytical model can be generalized to the case of an
elliptical rotating trap, using the excitation spectrum on top of a vortex free ground
state, which we calculate numerically. We provide possible explanations for the exper-
imentally observed critical frequencies above and below the Landau critical frequency.
In particular, we nd the rst indication for a reason that might explain a lower limit
for vortex nucleation observed by Hodby et al. [Phys. Rev. Lett. 88, 010405 (2002)].
Finally, we simulate non-equilibrium dynamics of rapidly rotating condensates as
well as the excitation of Tkachenko modes. We calculate the excitations of a vortex
lattice and identify the Tkachenko modes in the spectrum.
i
Acknowledgements
A work like this thesis is not possible without the help and support of many people
whom I would like to thank here whole-heartedly. First and foremost, my supervisor
Rob Ballagh, whose enthusiasm for physics is unbroken by the administrative duties
that come with his job. He is one of the friendliest and most supportive supervisors one
could wish for. Many thanks go to Crispin Gardiner, who put us on the right track with
the theoretical side of this work. His insight into the physics behind all the formulae is
almost scary. I would also like to thank my second supervisor Andrew Wilson for the
marvellous job he is doing on the experimental side of BoseEinstein condensation at
the University of Otago.
A surprisingly large number of international visitors nd their way to Dunedin. For
some fruitful discussions and help I would like to thank in particular Sam Morgan,
Matthew Davis, Martin Rusch, Ashton Bradley and Allan Grin, as well as Tapio
Simula who is currently doing a postdoc at Otago.
Past and present members of the theory group at Otago I would like to mention
are Ben Caradoc-Davies, who has the patience to solve any computer problem (as long
as it is Linux related), Blair Blakie, who can explain anything about physics (and is
too modest to talk much about his personal life), Jan Kr uger, formerly known as Max,
who turned into our Linux guru after Ben had left, Adam Norrie, who is just Adam,
Christopher Gies, who introduced us to the beautiful game of carom, Katherine Challis,
who motivated me to run a half-marathon and James Douglas, who is a very worthy
carom opponent.
I also appreciated the exchange of ideas with the experimental BEC group (though
I havent decided yet whether I prefer the door between our two oces open or shut)
and found all sta in the department most helpful at all times.
For keeping me sane I thank all of my friends in Dunedin and around the world,
my atmates of the last two years, who might have wondered at times whether I was
actually still living in the same at, and everyone at Sammys which became kind of a
second home. A very big thank you goes to my family back in Germany, who just had
to get used to the fact that I am as far away as it gets. One day you might even visit
me here.
Last but not least, I thank you, Alexandra, for everything you have done for me.
iii
Contents
Abstract i
Acknowledgements iii
1 Introduction 1
1.1 BoseEinstein Condensation . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Recent Advances in BoseEinstein Condensation . . . . . . . . . . . . . 2
1.2.1 Experimental systems . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Atom lasers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 Non-linear and quantum atom optics . . . . . . . . . . . . . . . 6
1.2.4 Atomic collisions and molecular BoseEinstein condensation . . 6
1.2.5 Superuidity and vortices . . . . . . . . . . . . . . . . . . . . . 7
1.3 This Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Peer-reviewed publications . . . . . . . . . . . . . . . . . . . . . 10
2 Mean Field Theory 11
2.1 Second Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Many-body Hamiltonian . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2 Pseudo-potential approximation . . . . . . . . . . . . . . . . . . 12
2.1.3 Denition of BoseEinstein condensation . . . . . . . . . . . . . 13
2.1.4 Bogoliubov approximation . . . . . . . . . . . . . . . . . . . . . 14
2.1.5 Time-dependent GrossPitaevskii equation . . . . . . . . . . . . 14
2.1.6 Time-independent GrossPitaevskii equation . . . . . . . . . . . 15
2.2 Elementary Excitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Bogoliubov transformation . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Linear response theory . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Computational Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
v
Contents
2.4 ThomasFermi Approximation . . . . . . . . . . . . . . . . . . . . . . . 23
3 Elementary Excitation Families 25
3.1 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Classication of Excitations . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.1 ThomasFermi limit . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 Families in the isotropic case . . . . . . . . . . . . . . . . . . . . 27
3.2.3 Anisotropic case . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Ordering of Quasiparticle Eigenfrequencies . . . . . . . . . . . . . . . . 33
3.3.1 Full solutions of Bogoliubov-de Gennes equations . . . . . . . . 33
3.3.2 Comparison with harmonic oscillator solutions . . . . . . . . . . 35
4 Background on Vortices 39
4.1 Quantization of Circulation . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Characteristics of Vortex . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Healing Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 Energy of Vortex Line . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Critical Frequency of Vortex Nucleation . . . . . . . . . . . . . . . . . 43
4.5.1 Thermodynamic critical frequency . . . . . . . . . . . . . . . . . 43
4.5.2 Landau criterion . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.5.3 Stability and energy barrier . . . . . . . . . . . . . . . . . . . . 45
4.5.4 Anomalous mode and vortex dynamics . . . . . . . . . . . . . . 45
4.5.5 Nucleation of vortices . . . . . . . . . . . . . . . . . . . . . . . . 46
4.6 Vortex Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5 Theory of Growth from Rotating Thermal Cloud 49
5.1 Quantum Kinetic Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.1 System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.1.2 Phase space density . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.3 Master equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1.4 Growth processes . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1.5 Simple growth equation . . . . . . . . . . . . . . . . . . . . . . 53
5.1.6 Evaluation of transition probability W
+
. . . . . . . . . . . . . 55
5.2 Phenomenological Growth Equation . . . . . . . . . . . . . . . . . . . . 56
5.2.1 Growth and loss in GrossPitaevskii equation . . . . . . . . . . 56
5.2.2 Master equation approach . . . . . . . . . . . . . . . . . . . . . 57
5.2.2.1 GrossPitaevskii equation in hydrodynamic form . . . 57
vi
Contents
5.2.2.2 Local energy conservation in hydrodynamic approxima-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.2.3 Application to quantum kinetic theory . . . . . . . . . 58
5.2.2.4 Phenomenological mean value equations . . . . . . . . 60
5.2.3 Rotating thermal cloud . . . . . . . . . . . . . . . . . . . . . . . 61
5.2.4 Stationary solution . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 Comparison with Other Theories . . . . . . . . . . . . . . . . . . . . . 63
6 Numerical Propagation Method 67
6.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.2 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.2.1 General points of consideration for numerical simulations . . . . 69
6.2.2 Choice of grid size and number of points . . . . . . . . . . . . . 70
6.2.3 Temporal step size . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.2.4 Accuracy of angular momentum operator in GrossPitaevskii
equation propagation . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2.5 Accuracy of phenomelogical growth equation . . . . . . . . . . . 80
7 Vortex Lattice Formation in Cylindrically Symmetric Trap 85
7.1 Two Dimensional System . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.2 Simple Growth and Normalization of Wave Function . . . . . . . . . . 87
7.3 Initial State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.4 General Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.5 Initiation Period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.5.1 Gain analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.5.2 Comparison with simulation . . . . . . . . . . . . . . . . . . . . 93
7.5.3 Critical angular velocity . . . . . . . . . . . . . . . . . . . . . . 95
7.5.4 Dominant angular momentum component . . . . . . . . . . . . 97
7.5.5 Role of initial seed . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.6 Equilibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.7 Properties of Equilibrium Lattice . . . . . . . . . . . . . . . . . . . . . 103
8 Vortex Lattice Formation in Elliptical Rotating Trap 109
8.1 No Thermal Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2 With Thermal Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.3 Gain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
vii
Contents
8.3.1 Vortex free solutions of the GrossPitaevskii equation in the ro-
tating frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.3.2 Excited states on vortex free solutions in the rotating frame . . 120
8.3.3 Comparison with simulation . . . . . . . . . . . . . . . . . . . . 122
8.3.4 Critical velocity and importance of quadrupole mode . . . . . . 126
8.4 Experiments on Vortex Nucleation and Lattice Formation . . . . . . . . 130
9 Vortex Lattice Decay 137
9.1 General Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
9.2 Angular Velocity and Radial Velocity . . . . . . . . . . . . . . . . . . . 141
9.3 Single Vortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.4 Experiments on Vortex Lattice Decay . . . . . . . . . . . . . . . . . . . 148
10 Non-equilibrium Dynamics and Lattice Excitations 151
10.1 Simple Non-Equilibrium Lattice . . . . . . . . . . . . . . . . . . . . . . 151
10.2 Deformation of Rapidly Rotating Condensates . . . . . . . . . . . . . . 153
10.2.1 Frequency splitting of m = 2 modes . . . . . . . . . . . . . . . 154
10.2.2 Excitation of m = 2 mode . . . . . . . . . . . . . . . . . . . . 155
10.2.3 Excitation of m = 2 mode . . . . . . . . . . . . . . . . . . . . . 159
10.3 Excitations on Vortex Lattices . . . . . . . . . . . . . . . . . . . . . . . 161
10.3.1 Excitations on single vortex . . . . . . . . . . . . . . . . . . . . 161
10.3.2 Excitations on vortex lattice . . . . . . . . . . . . . . . . . . . . 164
10.3.2.1 Frequency spectrum . . . . . . . . . . . . . . . . . . . 164
10.3.2.2 Tkachenko modes . . . . . . . . . . . . . . . . . . . . . 165
11 Conclusion 169
11.1 This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
11.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
A Optimization Methods 173
A.1 Partial Dierential Equations Solver . . . . . . . . . . . . . . . . . . . . 174
A.1.1 Cylindrically symmetric case in three dimensions . . . . . . . . 174
A.1.2 Two-dimensional system in the rotating frame . . . . . . . . . . 175
A.2 Conjugate-Gradient Method . . . . . . . . . . . . . . . . . . . . . . . . 176
A.2.1 Optimality function . . . . . . . . . . . . . . . . . . . . . . . . . 176
A.2.2 Conjugate-gradient optimization . . . . . . . . . . . . . . . . . . 179
A.3 Basis State Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
viii
Contents
A.3.1 Basis state expansion . . . . . . . . . . . . . . . . . . . . . . . . 180
A.3.2 Construction of harmonic oscillator states . . . . . . . . . . . . 182
A.3.3 Angular momentum operator . . . . . . . . . . . . . . . . . . . 183
A.3.4 Gaussian quadrature . . . . . . . . . . . . . . . . . . . . . . . . 184
A.3.5 Eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
A.3.6 Bogoliubovde Gennes equations . . . . . . . . . . . . . . . . . 186
B Propagation Methods For Dynamical Simulations 189
B.1 GrossPitaevskii Equation Propagation . . . . . . . . . . . . . . . . . . 189
B.1.1 Denition of the problem . . . . . . . . . . . . . . . . . . . . . . 189
B.1.2 Transformation into the interaction picture . . . . . . . . . . . . 190
B.1.3 Fourth-order Runge Kutta . . . . . . . . . . . . . . . . . . . . . 191
B.2 Phenomenological Growth Equation with Rotating Thermal Cloud . . . 192
References 194
ix
Chapter 1
Introduction
1.1 BoseEinstein Condensation
In classical mechanics a system of particles is described in the phase space of the canon-
ical variables of position and momentum. The evolution of the system is completely
determined once the position and momentum of all particles are known at some initial
time. Hence, it is in principle possible to uniquely label identical particles by specifying
their position and momentum. In a quantum mechanical many-body system, however,
the Heisenberg uncertainty principle forbids the simultaneous measurement of the ex-
act position and momentum of any particle. Instead, the evolution of the system is
described by a wavefunction that represents the probability of any of the particles be-
ing at a certain position. Identical particles become fundamentally indistinguishable.
However, certain symmetry conditions have to be imposed on the wavefunction. The
exchange of two identical particles does not change any physical observables. That
means that such a particle exchange can only introduce a phase factor in the wave-
function. Additionally, after a second exchange of the same two particles the original
wavefunction has to be recovered. It follows that the phase change is either -1 or
+1, and the wavefunction is symmetric or antisymmetric under particle exchange, re-
spectively. Hence, all elementary particles are classed into two categories: fermions
that are described by an antisymmetric wavefunction and obey FermiDirac statistics,
and bosons that are described by a symmetric wavefunction and obey BoseEinstein
statistics. Remarkably, in quantum eld theory it has been shown that this statistical
property is linked to an internal property of the particle, namely its intrinsic angular
momentum or spin. While this applies rigorously only to elementary particles, of which
very few are bosons, composite particles can also be regarded as bosons if their total
1
Chapter 1. Introduction
spin is integer as long as their internal structure is not apparent in the collisions, i.e.
their internal energy spacing is much larger than their interaction energy.
While at high temperatures the dierence between fermion and boson gases is slight,
suciently cooled down the dierent statistical behaviour leads to dramatically dierent
eects. Fermions, as a consequence of their antisymmetric wave function, obey the Pauli
exclusion principle which forbids any two fermions to occupy the same single-particle
quantum mechanical state. In contrast, an arbitrarily large number of bosons can
occupy the same state, and indeed, the more bosons occupy the same single-particle
state, the more particles are scattered into it. This Bose-enhanced scattering can lead
to an avalanche eect at very low temperatures and a certain phase space density,
where suddenly the ground state will be macroscopically occupied by atoms. This
eect is called BoseEinstein condensation (BEC). It can be considered as a phase
transition, where the atoms lose their individual identity. The condensate acts as one
entity exhibiting quantum mechanical properties on a macroscopic scale.
BEC was rst predicted by Einstein in 1924 as a consequence of Bose statistics,
which was introduced in an earlier paper by Bose. Bose developed the Bose statistics
for the case of photons, and Einstein generalized the idea to the case of indistinguish-
able particles. In 1961 an important equation was derived for the treatment of weakly
interacting dilute Bose condensed gases the GrossPitaevskii equation, which con-
siders the mean eld of the quantum system. In 1995 the eld of BEC took a huge
step forward when Anderson et al. [1] achieved BEC for the rst time in a dilute
weakly-interacting gas of trapped
87
Rb atoms.
1.2 Recent Advances in BoseEinstein Condensa-
tion
The rst experimental observation of BEC in a dilute alkali gas [1] in 1995 initiated
an enormous renewed interest in the eld of ultra-cold atoms and degenerate gases. It
has brought together researchers from elds as dierent as atomic physics, condensed
matter and quantum optics. The importance of this fast growing eld has been rec-
ognized by two Nobel prizes the rst one awarded in 1997 to Steven Chu, Claude
Cohen-Tannoudji and William D. Phillips for development of methods to cool and trap
atoms with laser light [2]. These developments paved the way for the experimental
realization of BoseEinstein condensates, which was rewarded with a Nobel prize in
2001 to Eric A. Cornell, Wolfgang Ketterle and Carl E. Wieman for the achievement
2
1.2. Recent Advances in BoseEinstein Condensation
of BoseEinstein condensation in dilute gases of alkali atoms, and for early fundamental
studies of the properties of the condensates [3].
BoseEinstein condensates are most commonly created using
87
Rb and
23
Na, but
have also been achieved in all other stable alkali species
7
Li [4],
85
Rb [5],
41
K [6], and
most recently
133
Cs [7] as well as in
1
H [8] and meta-stable
4
He [9, 10]. There is a
huge number of publications in the eld of BEC every year, which makes it impossible
to give a complete summary of all developments. A good summary of the theory of
BEC can be found in [11, 12, 13]. A highly recommended collection of review articles
on ultracold matter has been published in Nature [14, 15, 16, 17, 18, 19]. There is also
a comprehensive list of references in form of a Resource Letter [20]. In this section, we
will give a general overview concentrating mainly on recent experimental advances.
1.2.1 Experimental systems
Conventional traps
To reach the critical temperature of BEC, the atomic cloud is cooled in two steps.
First, laser cooling is applied to the cloud held in a magneto-optical trap (MOT) which
will cool it down to the region of tens of micro-Kelvins. In the next step, the cloud is
typically trapped in a quadrupole eld of a magnetic trap, and the hottest atoms are
removed by evaporative cooling. In this technique, a suitably tuned radio-frequency
(RF) eld ips the spin of the atoms at the edge of the trap to an untrapped m
F
state.
Because the hottest atoms are most likely located at the outer region of the trap,
they will be removed, while the remaining atoms re-thermalize. A slowly decreasing
RF frequency will cut further into the atomic cloud because the resonance condition
is dependent on the Zeeman shift of the atoms, and therefore on the magnetic eld
strength of the trap.
There is one major problem with a magnetic quadrupole trap. The magnetic eld
is zero at the centre of the trap, which allows atoms to escape undergoing Majorana
spin ips to untrapped m
F
states. The two most commonly used trap geometries solve
this problem in dierent ways. The time-orbiting potential (TOP) trap uses a rapidly
rotating bias eld on top of the quadrupole eld so that the point of zero magnetic
eld rotates constantly around the centre of the trap. Because the movement is so fast,
the atoms only experience an averaged harmonic magnetic eld. A TOP trap gives
rise to a pancake-shaped condensate. The second design is a Ioe-Pritchard (IP) type
trap where the point of zero magnetic eld is removed altogether by the application of
a static bias eld using a complex design of cloverleaf coils [21]. Condensates in IP
3
Chapter 1. Introduction
traps are usually long cigar-shaped.
All-optical traps and spinor condensates
Magnetic traps have the disadvantage that only atoms with certain low-eld-seeking
m
F
states can be trapped. Even though binary mixtures of atoms in dierent hyperne
states have been explored in magnetic traps [22, 23, 24, 25], purely optical traps have
the advantage that they can conne all possible hyperne states. This allows the
exploration of so called spinor condensates that consist of atoms in all possible hyperne
states [26, 27]. Typically, the condensate is created in a magnetic trap and subsequently
transfered into the optical trap. However, the BEC transition has also been achieved
in an all optical dipole trap formed by CO
2
laser beams [28].
Micro-traps
In recent years, there has been the trend to miniaturization of magnetic traps on mi-
crochips. The traps are formed by lithographically created arrays of current-carrying
wires. The BEC transition has already been achieved on microchips [29, 30]. The main
advantage of these setups is the easy accessibility to manipulate and study the con-
densates compared to conventional techniques where the optical and mechanical access
is very limited. The JILA group has demonstrated the transport of a condensate in a
magnetic waveguide build on such micro-structures, and are also making advances with
beamsplitters for cold atoms [31] towards beamsplitters for BoseEinstein condensates.
BoseEinstein condensation in lower dimensions
The reduction of a physical system to lower dimensionality can give rise to completely
new physics with formerly unknown phenomena, e.g. the Quantum Hall eect in a two-
dimensional electron gas. It is well known that BEC cannot occur in a uniform one-
dimensional or two-dimensional system
1
. However, a conning trap allows the BEC to
take place even in lower dimensional systems. Such (quasi)low-dimensional condensates
have been realized by conning them in highly elongated optical traps where the energy-
spacing in one or two dimensions exceeds the interaction energy between the atoms [32].
Other possibilities are the formation of quasi-condensates [33] that behave locally like
condensates, but do not show phase coherence accross the whole system due to large
phase uctuations [34].
1
Theoretically, a uniform two-dimensional BEC can exist at T = 0 K.
4
1.2. Recent Advances in BoseEinstein Condensation
BoseEinstein condensation in optical lattices
When condensates are loaded into weak optical lattices, the phase coherence of the con-
densate is maintained accross the lattice sites due to tunneling between them. Equiv-
alent to the Josephson eect, atoms oscillate back and forward between the dierent
lattice sites. This coherence has been demonstrated by the pulsed output of a conden-
sate falling under gravity in an optical lattice [35]. As the lattice barriers are increased,
tunneling is suppressed, and the number of atoms in each well becomes more sharply
dened while the phase coherence between lattice sites is lost. This can also be de-
scribed in terms of number squeezing since condensate phase and number are conjugate
variables [36]. In a three-dimensional lattice, this eect leads to a phase transition from
a superuid to a Mott insulator state [37] where number uctuations are completely
suppressed.
An optical lattice can also impart momentum to a condensate via diraction [38]
or Bragg scattering [39]. Bragg scattering has become an important experimental
tool because it allows the condensate to be put into a superposition of well-dened
momentum states.
1.2.2 Atom lasers
In analogy to an optical laser, an atom laser is a source of coherent matter waves [40].
A trapped BoseEinstein condensate is essentially a single macroscopically occupied
mode of the trap, spatially coherent over the whole condensate region [41, 42], and as
such suitable as a source for an atom laser. In early experimental realizations of atom
lasers [43, 44] the output beam was pulsed, coupling small amounts of condensate by
RF induced spin-ips or Majorana spin ips to an untrapped state. Later, a quasi-
continuous output was achieved [45, 46] with a much better collimated beam.
In all of these experiments, the lasing time of the atom lasers was limited by the
size of the condensate because there was no mechanism to replenish the condensate
while atoms were being output-coupled. Recently, the MIT group has demonstrated
a continuous source of Bose condensed atoms [47]. The condensate held in an optical
trap is reloaded with new condensates delivered using optical tweezers that allow Bose
Einstein condensates to be transported over distances of tens of centimeters [48].
5
Chapter 1. Introduction
1.2.3 Non-linear and quantum atom optics
A number of non-linear eects known from non-linear photon optics has also been
demonstrated with BoseEinstein condensates. In non-linear optics, a medium is nec-
essary to couple light elds non-linearly. With matter waves there is no need for a
medium, but the non-linearity arises from the interactions between atoms in the con-
densate. In four-wave mixing experiments [49] the production of a condensate in three
carefully phase-matched momentum states yields to the population of a fourth momen-
tum state. Similarly, the interaction of light waves with a condensate can lead to eects
like superradiant scattering [50] or wave-matter amplication [51, 52].
A completely dierent phenomena is the propagation of solitons, which are wavepack-
ets that can travel over long distances in non-linear media without spreading. So called
dark solitons have been imprinted onto condensates [53, 54].
Very remarkable experiments have been performed by Hau et al. [55] who were able
to eectively slow down the speed of light travelling through a condensate and even
stop it completely [56] by storing the coherent information contained in the laser eld
in the internal states of the atoms. These experiments made use of a quantum eect
called electromagnetically induced transparency that allows the propagation of light
through an otherwise opaque atomic medium.
1.2.4 Atomic collisions and molecular BoseEinstein conden-
sation
At low temperatures the nature of collisions between atoms is determined by the s-
wave scattering length. If the s-wave scattering length is positive, atoms repell each
other, and the conning force of the potential is balanced by the mean-eld repulsion.
If the scattering length is negative as in
7
Li, however, the attractive forces between the
atoms lead to a collapse of the condensate [57]. Only for a very small number of atoms
a stable condensate is possible if the self-attractive forces are balanced by a repulsion
arising due to a momentum-position uncertainty in the trap [4, 58].
However, some alkali elements show Feshbach resonances where the free state of
the colliding atoms couples resonantly to a quasibound molecular state. This coupling
strongly aects the scattering length in the collision. Because the free state and quasi-
bound state have dierent magnetic moments, a magnetic eld can be used to tune the
scattering length from positive to negative values around a Feshbach resonance [59].
When the scattering length of
85
Rb was tuned from positive to negative, besides the
6
1.2. Recent Advances in BoseEinstein Condensation
expected collapse of the condensate, a blast of hot atoms was observed [60]. Because
this phenomena is similar to the neutrino burst of a collapsing star during a supernova
it was named Bose nova. Using a more controlled collapse the point of instability in
a condensate with attractive interactions was determined [61].
With the control over the scattering length Feshbach resonances provide they are a
very useful experimental tool. They were used to Bose condense
133
Cs [7], which has
an enormous negative scattering length of 2000 Bohr radii for zero magnetic eld, and
also in the formation of bright solitons in
7
Li [62].
One way to produce ultra-cold molecules is to use photo-association starting with
atoms in a BoseEinstein condensate [63, 64] where two colliding atoms collectively
absorb a photon forming a bound, excited-state molecule. Whether this process is
coherent and whether a molecular BoseEinstein condensate is formed is still subject
to further research. Donley et al. [65] have coherently coupled atoms and molecules
in a BoseEinstein condensate using a time-varying magnetic eld near a Feshbach
resonance.
1.2.5 Superuidity and vortices
Already in 1938, London suggested that the cause of the superuid behaviour of
4
He
might be related to BEC although such a system is highly depleted. With the advent of
dilute alkali BoseEinstein condensates, there is the chance to study superuid eects
in almost fully condensed systems.
One criterion of superuidity is that obstacles moving through the condensate ex-
perience zero or much reduced friction as long as their velocity is below some critical
velocity given by the Landau criterion
v
c
= min
_
E
p
p
_
, (1.1)
where E
p
and p are the energy and momentum of an excitation. Below this velocity, no
excitation can be created. This eect has been observed in several experiments using
blue-detuned laser beams [66, 67] and impurity atoms [68] moving through the conden-
sate. Similarly, a condensate can move dissipationless through an optical lattice with
velocities below the critical velocity [69]. Another interesting eect is the occurence of
a certain excitation called scissors mode [70].
A fundamental property of a superuid is the fact that it can only support irrota-
tional ow. For low rotation rates below the critical velocity for vortex formation, this
7
Chapter 1. Introduction
leads to a characteristic ow pattern which has experimentally been observed directly
[71]. For higher rotation rates, angular momentum can only be aquired sustaining
an irrotational ow by the formation of vortices. The rst vortex experimentally ob-
served was created in a two component condensate where the vortex core is lled by
atoms in a dierent spin state [72], a conguration which is sometimes referred to as
skyrmion rather than vortex. The rst observation of a pure vortex in a single compo-
nent condensate [73] has stimulated a large number of experiments. Early experiments
concentrated on states with a single vortex or a small number of vortices [74, 75, 76].
Usually, a pure condensate is set into rotation by means of an anisotropic rotating po-
tential, which is typically ellipsoidal. Potentials with dierent symmetries show distinct
resonance frequencies for vortex formation [77]. Haljan et al. [78] demonstrated the
nucleation of a vortex lattice from a rotating thermal cloud without the need of any
rotating anisotropy.
The rst realization of a large vortex lattice with more than hundred vortices [79]
initiated research into the properties of vortex lattices such as their formation and decay
[80], non-equilibrium deformations under compression [81], excitations of Tkachenko
oscillations [82] and giant vortices [83]. A major part of this thesis is concerned with
vortex lattices. Chapter 4 gives an overview on the theory of vortices and vortex lattices,
and a more detailed account on experiments on vortex lattices will be presented in later
chapters when relevant.
1.3 This Work
1.3.1 Overview
In chapter 2, we review the mean-eld theory of a BoseEinstein condensate. We
present a brief derivation of the time-dependent and time-independent GrossPitaevskii
equation, which provides a very good description of a BoseEinstein condensate at
zero temperature. The linear excitations of a BoseEinstein condensate are described
by the Bogoliubovde Gennes equations. We present two dierent derivations of the
Bogoliubovde Gennes equations that illustrate the equivalence of elementary and col-
lective excitations in a BoseEinstein condensate. And nally, we introduce our choice
of computational units for numerical calculations.
In chapter 3 we consider the excitation spectrum of a BoseEinstein condensate in
a three dimensional cylindrically symmetric trap. This work extends earlier results re-
ported in my Masters thesis, in presenting a systematic classication of the excitations
8
1.3. This Work
that generalizes the concept of families rst identied by Hutchinson and Zaremba [84].
The extension involves a complete revision of the procedure for assigning a family clas-
sication to any mode in the anisotropic case. We have also been able to determine
the energy ordering of the modes, and give a simple model that explains this ordering.
A large part of this thesis is concerned with the nucleation of vortices and the for-
mation of vortex lattices in rotating BoseEinstein condensates. Chapter 4 summarizes
the fundamental properties of superuids and vortices which are well known from the
work on superuid Helium, as well as more recent results on vortices in BoseEinstein
condensates. In particular, we discuss the question of a critical angular velocity for
vortex nucleation in a rotationally stirred BoseEinstein condensate.
In chapter 5, we present the formalism we have used to describe the process of
vortex lattice formation. Dissipation is known to be an essential element in vortex
lattice formation. Therefore, the standard GrossPitaevskii equation cannot describe
this phenomenon, and a new approach is needed. We provide a summary of the theory
developed by Gardiner et al. [85] to treat condensate dynamics in the presence of a
thermal cloud. That formalism is based on quantum kinetic theory, and shows how
a dissipation mechanism is introduced by exchange of atoms between the condensate
and the surrounding thermal cloud of atoms. In its most simplied form, Gardiners
theory reduces to a modied GrossPitaevskii equation that includes growth and loss
terms, and which we call the phenomenological growth equation. This equation forms
the basis of much of our treatment of rotating condensates.
In order to simulate the phenomenological growth equation we needed to develop
a numerical method for propagating the equation, and we describe this method in
chapter 6. We also give a detailed account of the accuracy of the method, including
the optimal choice of grid size, number of points, and temporal step size. The angular
momentum operator is required in this equation to describe a rotating thermal cloud,
and we discuss its eect on the stability and reliability of the method.
In chapter 7, we present results on the formation of vortex lattices from a rotating
thermal cloud in a cylindrically symmetric trap. We analyse the initiation process
in terms of a gain process for surface modes with non-zero angular momentum, and
obtain gain coecients in terms of the excitation energies in good agreement with the
simulations.
In chapter 8, we apply the phenomenological growth equation to the case of vortex
lattice formation in a rotating elliptical trap. Simply stirring a condensate in the
absence of a thermal cloud does not lead to a vortex lattice. However, in the presence
of a thermal cloud, the stirring can seed angular momentum components which may
9
Chapter 1. Introduction
then grow from the thermal cloud by stimulated collisions. To analyse this scenario,
we numerically calculate the stationary vortex free states of a rotating elliptical trap,
and their excitations by solving the GPE and BdG equations in the rotating frame.
We also give a critical evaluation of our model in relation to experimental results for
vortex nucleation through rotational stirring.
The phenomenological growth equation can also be applied to the decay of vortex
lattices in the limiting case of a stationary thermal cloud, which we consider in chapter
9. And in chapter 10, we present simulations on non-equilibrium lattice dynamics,
in particular on the deformations of rapidly-rotating vortex lattices in the presence of
quadrupole excitations as explored in a recent experiment by Engels et al. [81]. Finally,
we calculate the excitations of a vortex lattice and identify the Tkachenko modes in
the excitation spectrum, which have recently been observed experimentally [82].
1.3.2 Peer-reviewed publications
Some of the work presented in this thesis has been published in peer-reviewed jour-
nals. The work on the excitation families of a cylindrically symmetric BoseEinstein
condensate presented in chapter 3 has appeared in Journal of Physics B [86].
Some of the main results from chapters 7 and 8 on the nucleation, formation and
stabilization of vortex lattices have been published in Physical Review Letters [87].
The methods developed in this thesis have also formed the basis for some work on
giant vortices and Tkachenko oscillations [88], which has been published in Physical
Review Letters. A small part of those results are presented in chapter 10.
10
Chapter 2
Mean Field Theory
In this chapter the basic equations describing a single-component BoseEinstein con-
densate will be derived. There are many dierent ways to derive these equations,
e.g. via Greens functions [89] or variational principles [90, 91]. We have chosen to
follow the mean eld approach because it provides a particularly straightforward pro-
cess for obtaining the basic equations. The underlying idea is that in a BoseEinstein
condensate all atoms occupy the same single-particle quantum state and can, therefore,
be described by the same wave function. A single atom is not aware of the individual
behaviour of the others and loses its individuality moving through the condensate. To
a good level of approximation it only sees the mean eld generated by the condensate
as a whole. The condensate can be thought of as acting coherently on a single atom.
In the nal sections of this chapter, we briey show how to transform the Gross
Pitaevskii equation into a form suitable for numerical solution. We introduce our choice
of dimensionless units and also introduce the ThomasFermi approximation, which is
a very good description of a condensate in the hydrodynamic limit.
2.1 Second Quantization
2.1.1 Many-body Hamiltonian
The starting point of a theoretical treatment of a single-component BoseEinstein con-
densate is the exact many-body Hamiltonian for a system of identical, structureless
bosons. If the system is suciently dilute only pairwise interactions have to be taken
into account. This condition is well satised in current experiments on alkali atoms
with lifetimes of the order of seconds where three-body recombinations are the main
loss factor that limits the lifetime. The many-body Hamiltonian can be written in
11
Chapter 2. Mean Field Theory
terms of the boson eld operator

(r, t) in second quantization as

H =
_
dr

(r, t)

H
0

(r, t)
+
1
2
_
dr dr

(r, t)

(r

, t)V (r r

(r

, t)

(r, t), (2.1)


where the single particle Hamiltonian is given by

H
0
=
_

2
2m

2
+V
ext
(r, t)
_
. (2.2)
Here, V
ext
(r, t) represents an external potential, which may consist of a trapping poten-
tial and other parts, e.g. from an external laser, V
ext
(r, t) = V
T
(r) + V
other
(r, t). If the
external potential is set to zero, we retrieve the equation for a homogeneous interacting
Boson gas. The eld operator

(r, t) creates a particle of mass m at position r at time


t. It satises the boson commutation relation
_

(r, t),

(r

, t)
_
= (r r

)
_

(r, t),

(r

, t)
_
= [

(r, t),

(r

, t)] = 0.
(2.3)
The interatomic potential V (r r

) represents the interaction strength between the


atoms in a binary collision. While this Hamiltonian is exact it is intractable to analytical
or numerical solutions except in the case of only a few particles.
2.1.2 Pseudo-potential approximation
At low temperatures s-wave scattering is the dominant collision process in dilute gases.
If the s-wave scattering length a is small compared to the de-Broglie wavelength
dB
=
_
2
2
/mk
B
T, a good approximation for the interatomic potential is the replacement
by a pseudo-potential [92]
V (r r

) = U
0
(r r

), (2.4)
where U
0
represents the eective interaction strength which is related to the s-wave
scattering length by
U
0
=
4
2
a
m
. (2.5)
Eectively, the collisions are now treated as hard sphere collisions. We note that this is
a low momentum approximation. For high momentum collisions the pseudo-potential
approximation gives rise to ultra-violet divergences because it scatters high momentum
12
2.1. Second Quantization
particles just as eectively as low momentum ones. This assumption is unphysical since
in real collisions the energy transfer is less eective for higher momenta. Hence, care
has to be taken if high momentum collisions are included. Morgan [93] has shown that
a high energy renormalization can be achieved in a natural way if the pseudo-potential
approximation is done on the scattering T matrix rather then V (which is simply
the rst term of the T matrix series). In this thesis, we will consider only very low
temperatures so that the pseudo-potential approximation of equation (2.4) can safely
be used.
2.1.3 Denition of BoseEinstein condensation
BoseEinstein condensation (BEC) is the macroscopic occupation of a single quantum
state. Following Legget [12], we can specify this by considering the one-particle reduced
density matrix
(r, r

, t)
_

(r, t)

(r

, t)
_
, (2.6)
where the average indicated by the angled brackets is in general statistical as well as
quantum mechanical. It is always possible to write the boson eld operator in terms
of an orthonormal set of single-particle wave funtions
i
(r, t) [91]

(r, t) =

i
(r, t) a
i
(t), (2.7)
where a
i
are the corresponding boson annihilation operators. These annihilation oper-
ators and their creation counterparts a

i
are dened in Fock space by
a

i
[n
0
, n
1
, . . . , n
i
, . . .) =

n
i
+ 1[n
0
, n
1
, . . . , n
i
+ 1, . . .)
a
i
[n
0
, n
1
, . . . , n
i
, . . .) =

n
i
[n
0
, n
1
, . . . , n
i
1, . . .), (2.8)
where n
k
are the occupation numbers of the single-particle states. The creation and
annihilation operators obey the usual boson commutation rules at the same time
_
a
i
, a

j
_
=
ij
[ a
i
, a
j
] = [ a

i
, a

j
] = 0.
(2.9)
In this notation the reduced density matrix (2.6) takes the form
(r, r

, t) =

i
n
i
(t)

i
(r, t)
i
(r

, t). (2.10)
13
Chapter 2. Mean Field Theory
Here, n
i
(t) a

i
(t) a
i
(t)) are the the expectation values of the number operator. We
shall say that the system shows BEC at any given time t if one or more of the eigenvalues
n
i
(t) are of the order of the total number of particles N, and in particular simple BEC
if only one of the eigenvalues is of the order N, while all others are of order 1. In the
case of simple BEC we will use the index zero to indicate the state of macroscopical
occupation with N
0
n
0
N. We shall call the single-particle state
0
(r, t) the
condensate wave function and N
0
the (mean) number of particles in the condensate.
The simplest and most direct choice of an order parameter for the BEC phase transition
is then (r, t) =
_
N
0
(t)
0
(r, t), which is simply the single-particle wave function into
which condensation occurs scaled by the number of atoms in this state. It may be
stressed that the overall phase of this order parameter has no physical signicance.
2.1.4 Bogoliubov approximation
An approximation commonly used in the BoseEinstein literature is the Bogoliubov
approximation in which the operators a
0
and a

0
are replaced by

N
0
. It is based on
the idea that states with N
0
and N
0
+ 1 N
0
atoms in the condensate correspond
to essentially the same physical conguration. The boson eld operator can then be
separated as

(r, t) = (r, t) +

(r, t) (2.11)
into a condensate part described by the order parameter (r, t) dened in this context
as (r, t) =

(r, t)) and an operator



(r, t) representing uctuations, whose expec-
tation value is zero by denition. However, this approach is not consistent with the
conservation of number of atoms because the boson eld operator has a non-zero expec-
tation value, which is only possible if the condensate wave function is a superposition
of states with dierent numbers of atoms N. The same diculty is often referred to as
the problem of spontaneously broken symmetry which occurs due to the denite phase
associated with the condensate part. Phase and particle numbers are conjugate vari-
ables, i.e. it is not possible to have a denite condensate phase and simultaneously a
denite number of condensate atoms.
2.1.5 Time-dependent GrossPitaevskii equation
Despite the conceptual diculties associated with the Bogoliubov approximation we
will use it here to derive the GrossPitaevskii equation (GPE), the central equation in
the description of a dilute BoseEinstein condensate. Ultimately, its use is justied by
14
2.1. Second Quantization
more complex number-conserving approaches which lead to essentially the same results
[94].
The Heisenberg equation of motion for the boson eld operator is
i


(r, t)
t
=
_

(r, t),

H
_
=
_

H
0
+U
0

(r, t)

(r, t)
_

(r, t). (2.12)
Inserting the decomposition (2.11) into (2.12) and taking the expectation value yields
i

t
=
_

H
0
+U
0
[[
2
_
+U
0
_
2

) +

)
_
. (2.13)
The term

) is identied with the non-condensate density n acting back on the


condensate, and the term m

) is known as the anomalous average which modies


the interaction strength between condensate atoms due to virtual processes. However,
a careful consideration shows that the dominant part of this term has already been
included by the introduction of the contact potential [93]. Finally, the term

)
represents collisions of two thermal atoms in which one of them enters the condensate.
In the limit of zero temperature all these terms can be neglected because the fraction
of thermal atoms is very small. Then, we are left with the famous GPE
i
(r, t)
t
=
_

H
0
+U
0
[(r, t)[
2
_
(r, t). (2.14)
The GPE is strictly valid only when all atoms are in the condensate and can,
therefore, be used only for temperatures near to T = 0 K. Actually, even at T = 0 K
some atoms are depleted from the condensate due to interactions. In liquid
4
He this
depletion is about 90% so that the mean eld approach is not useful to obtain qualitative
results. In the case of condensed alkali gases, however, the depletion is of the order of
0.5% [95] and the GPE enables accurate quantitative predictions.
2.1.6 Time-independent GrossPitaevskii equation
If we use the usual ansatz (r, t) = (r)e
it/
for a stationary solution of the GPE
(2.14) we arrive at the time-independent GPE
_

H
0
+U
0
[(r)[
2
_
(r) = (r). (2.15)
The lowest energy state of this equation is the ground state into which condensation
occurs. At T = 0 K the eigenvalue can be identied as the chemical potential of the
15
Chapter 2. Mean Field Theory
system.
To include thermal eects, terms of higher order in

from equation (2.13) must be
accounted for. In the Popov approximation only the term n =

) is kept

H
0
(r) +U
0
[n
c
(r) + 2 n(r)](r) = (r), (2.16)
where n
c
(r) = [(r)[
2
denotes the condensate density. If the anomalous average m(r) =

) is also kept we obtain the Hartree-Fock-Bogoliubov (HFB) equation

H
0
(r) +U
0
[n
c
(r) + 2 n(r)](r) +U
0
m(r)

(r) = (r). (2.17)


We used the symbol for the eigenvalue of these latter two equations because at nite
temperatures it diers from the thermodynamic quantity of the chemical potential
by = + k
B
T ln(1 + 1/N
0
) [96]. Whereas the HFB-Popov equation (2.16) is
self-consistent, the HFB equation (2.17) yields a gap in the excitation spectrum [97].
Hutchinson et al. [84] have introduced a gapless HFB approximation by modifying the
interaction strength U
0
. This is justied by an approximation for the many-body T
matrix [98].
2.2 Elementary Excitations
In this section we will derive the so called Bogoliubovde Gennes equations (BdG equa-
tions) at zero temperature which describe the excitations of a BoseEinstein condensate.
In an ideal gas the only excitations possible are single-particle excitations, i.e. a single
particle occupies an energy state above the macroscopically occupied ground state. In
the homogeneous case, these excitations are simply plane waves, while in a conned
system they are the eigenstates of the trapping potential. If the gas is interacting, the
nature of these elementary excitations changes because as the excited particle moves
through the sytem it interacts with the neighbouring atoms. However, often it is pos-
sible to describe the combined system of the single particle and the surrounding cloud
of atoms it interacts with in terms of a ctitious quasiparticle, similar to the concept
of a dressed state of an atom in an electromagnetic eld. Quasiparticles represent the
excited energy levels of an interacting many-body system.
Interacting many-body systems also have a completely dierent type of excitations:
collective excitations. They are associated with density uctuations and involve the
collective wave-like motion of all particles. Since interactions are crucial for those kind
16
2.2. Elementary Excitations
of collective excitations, there is no equivalent in a non-interacting gas.
A peculiar property of a BoseEinstein condensate is that its elementary and col-
lective excitations are identical. This can be understood by the fact that all atoms in a
condensate are described by the same single-particle wave function. Thus, any excita-
tion involving one particle (or quasiparticle) automatically involves all others leading
to a collective response. To illustrate this property we will derive the BdG equa-
tions in two distinctively dierent ways. The rst approach uses the grand-canonical
Hamiltonian, which is diagonalized by a Bogoliubov transformation into a collection of
non-interacting quasiparticles [99]. In this pure quantum mechanical approach the exci-
tations are necessarily orthogonal to the condensate since only excited non-condensate
atoms take part in the collective modes [100]. The second derivation uses linear re-
sponse theory around the time-dependent GPE. Thus, the excitations are considered
as collective motions of condensate atoms. These modes are not necessarily orthogonal
to the ground state wave function. We shall outline both derivations for zero temper-
ature following closely Ref. [100] and [101] and make the dierences between the two
approaches clear.
2.2.1 Bogoliubov transformation
The Bogoliubov transformation is a well known method to diagonalize a quadratic
Hamiltonian and gives a transformed set of bosonic operators, which are called quasi-
particle operators due to their particle-like character.
A many atom system can be described by the grand canonical, many-body Hamil-
tonian

K =

H

N, where

H is the many-body Hamiltonian,

N the number operator
and the chemical potential. Within the pseudo-potential approximation described in
section 2.1.2 this is written in terms of eld operators as

K =

H

N =
_
dr

(r)[

H
0
]

(r) +
1
2
U
0
_
dr

(r)

(r)

(r)

(r). (2.18)
Inserting the decomposition of the eld operator (2.11) into (2.18) and neglecting terms
17
Chapter 2. Mean Field Theory
in

(r) higher than quadratic yields

K =
_
dr

[

H
0
+
1
2
U
0
[[
2
]
+
_
dr

[

H
0
+U
0
[[
2
]

+
_
dr

[

H
0
+U
0
[[
2
] (2.19)
+
_
dr

[

H
0
+ 2U
0
[[
2
]

+
1
2
U
0
_
dr

.
The rst term in the above equation is just a c-number. The second and third term
vanish if satises the time-independent GPE (2.29). The remaining Hamiltonian can
be diagonalized by a linear canonical transformation, i.e. a transformation of creation
and annihilation operators that preserves the commutation relations. This is done by
the Bogoliubov transformation

(r) =

i
[u
i
(r)
i
+v

i
(r)

i
]

(r) =

i
[u

i
(r)

i
+v
i
(r)
i
],
(2.20)
which expresses the uctuation operator in terms of the quasiparticle creation and
annihilation operators

i
and
i
that are required to fulll the usual boson commutation
relations
[
i
,

j
] =
ij
[
i
,
j
] = [

i
,

j
] = 0.
(2.21)
The Hamiltonian is diagonalized if the functions u
i
(r) and v
i
(r) are chosen to satisfy
the following equations
_
dr u

i
[Lu
j
+U
0

2
v
j
] +v

i
[Lv
j
+U
0

2
u
j
] =
i

ij
_
dr u
i
[Lv
j
+U
0

2
u
j
] +v
i
[Lu
j
+U
0

2
v
j
] = 0
(2.22)
where L is dened as
L =

H
0
+ 2U
0
[[
2
. (2.23)
From the equations (2.22) it can be shown [102] that the u
i
s and v
i
s obey the
orthogonality relation
_
dr u
i
u

j
v
i
v

j
=
ij
(2.24)
18
2.2. Elementary Excitations
and the symmetry relation
_
dr u
i
v
j
v
i
u
j
= 0. (2.25)
The normalization for i = j in the orthogonality condition is forced to be one. This is
a consequence of the commutation rule of the operator

(r)
[

(r),

(r

)] = (r r

) (r)

(r

), (2.26)
which follows from the decomposition (2.11) and the commutation rules of the boson
eld operator (2.3).
The quasiparticle amplitudes u
i
and v
i
must be orthogonal to the ground state
wave function , which can be understood as follows. In the expansion of the eld
operator (2.7) the single particle wave functions
i
have to be orthonormal to preserve
the commutation relations for the boson eld operators (2.3). This implies in particular
that the coecients
i
(r, t), i = 1, 2, . . ., of the operator

(r, t) =

i=1

i
(r, t) a
i
(t) are
orthogonal to the ground state wave function (r, t) =
_
N
0
(t)
0
(r, t). The Bogoliubov
transformation (2.20) simply casts the operator

into a collection of non-interacting
quasiparticles represented by creation and annihilation operators

i
,
i
, but preserves
this orthogonality to the ground state. Hence, u
i
and v
i
are orthogonal to the ground
state wave function.
2.2.2 Linear response theory
At very low temperatures a Bose condensed gas is described by the time-dependent
GPE
i
(r, t)
t
=
_

H
0
+U
0
[(r, t)[
2
_
(r, t). (2.27)
Assuming that all atoms are in the condensate the wave function is normalized to the
total number of particles in the trap
_
dr [(r, t)[
2
= N. The lowest energy eigen-
state solution of equation (2.27) is of the form (r, t) =
g
(r)e
it/
, where is the
eigenvalue of the system. To nd the linear excitations consider a small harmonic dis-
turbance of frequency . Assuming that the excitations are only weakly occupied so
that they do not aect the condensate ground state and do not couple to each other,
we can look for solutions of the form
(r, t) = e
i

t
_

g
(r) +u(r)e
it
+v

(r)e
it

, (2.28)
19
Chapter 2. Mean Field Theory
where
g
(r) is a solution of the time-independent GPE (2.15). If this is substituted
into equation (2.27) and only terms linear in u(r) and v(r) are retained one obtains the
BdG equations by equating terms of e
it
,
_

H
0
+U
0
[
g
(r)[
2
_

g
(r) =
g
(r), (2.29)
Lu
i
(r) +U
0

2
g
(r)v
i
(r) =
i
u
i
(r),
Lv
i
(r) +U
0

2
g
(r)u
i
(r) =
i
v
i
(r),
(2.30)
where L is dened as before L =

H
0
+ 2U
0
[
g
(r)[
2
.
The rst equation is just the time-independent GPE, which does not contain the
functions u
i
(r) and v
i
(r). Hence, it can be rst solved independently. The following
two equations are a set of coupled equations, which are also dependent on the solution
of the GPE through
g
(r) and . They determine the shape and frequencies of the
linear excitations completely.
The BdG equations (2.30) imply the orthogonality and symmetry relations (2.24)
and (2.25), respectively [99]. In this approach, however, the normalization in the or-
thogonality relation, when i = j, is not forced to unity, but could be chosen arbitrarily.
Only the quantum mechanical approach from the previous section reveals that the
bosonic character of the quasiparticles forces the normalization to unity. A particular
solution of the BdG equations is the Goldstone mode u
0
(r) =
g
(r) and v
0
(r) =

g
(r)
with
0
= 0. Both the orthogonality and symmetry relation for this solution take the
form
_
dr

g
u
i
+
g
v
i
= 0. (2.31)
It can be easily veried that the BdG relations (2.30) satisfy equations (2.22) and
are, therefore, sucient to diagonalize the Hamiltonian. However, they do not nec-
essarily preserve the orthogonality to the ground state whereas the relations (2.22)
allow one to choose the excitations orthogonal to the ground state. This can be done
by solving the BdG equations (2.30) rst and then projecting out the overlap with the
condensate [103, 100]. The corresponding projection operator

P acting on any function
f(r) is dened as

Pf(r) = f(r)
g
(r)
_
dr

g
(r

)f(r

). (2.32)
20
2.3. Computational Units
Thus, the orthogonal excitations become
u
i
(r) =

Pu
i
(r) = u
i
(r) g
i

g
(r) (2.33)
v

i
(r) =

Pv

i
(r) = v

i
(r) +g

g
(r), (2.34)
where g
i

_
dr

g
(r

)u
i
(r

) =
_
dr

g
(r

)v
i
(r

). The latter equality follows from


equation (2.31). The orthogonal excitations u
i
and v
i
satisfy the following modied
BdG relations
L u
i
+U
0

2
g
v
i
=
i
( u
i
+g
i

g
)
L v
i
+U
0

2
g
u
i
=
i
( v
i
g
i

g
)
(2.35)
giving the same eigenvalues
i
, and they still diagonalize the Hamiltonian. They also
fulll the orthogonality and symmetry relations.
Following the nomenclature of reference [100] we will call the excitations satisfy-
ing the ordinary BdG equations (2.30) linear excitations and the ones orthogonal to
the ground state satisfying (2.22) orthogonal excitations. Because the orthogonal ex-
citations can always be obtained from the linear ones by the projection method just
described, we will numerically solve for the linear excitations if not explicitly stated
otherwise.
2.3 Computational Units
We now briey introduce some important quantities for the numerical treatment of the
GrossPitaevskii and Bogoliubovde Gennes equations. In computational physics it is
customary to use dimensionless units for notational simplicity and numerical accuracy.
In most experimental realizations of BEC, the conning trap is cylindrically symmetric
with a harmonic trapping potential, which can be written as
V
T
=
1
2
m
2
r
[x
2
+y
2
+ (z)
2
], (2.36)
where we have chosen the z-axis of the coordinate system parallel to the symmetry
axis of the trap. Here,
r
is the radial trapping frequency. The anisotropy of the trap
is given by =
z
/
r
, the ratio between the trapping frequency in z-direction
z
and

r
. In terms of the radial trapping frequency, we dene the following units for distance
21
Chapter 2. Mean Field Theory
and time
r
0
=
_

2m
r
, (2.37)
t
0
=
1

r
. (2.38)
The unit r
0
is known as the harmonic oscillator length. From those units, we can obtain
derivative units for momentum, angular momentum and energy
p
0
=
_
2m
r
(2.39)
L
0
= , (2.40)
E
0
=
r
, (2.41)
respectively.
As an example, we will transform the time-independent GPE (2.15) to computa-
tional units. We introduce dimensionless quantities, which we will indicate by a tilde,
using
r = rr
0
, =
r
. (2.42)
It is worth noting that the wave function itself has units of [length]
d/2
, where d is the
number of spatial dimensions. If the wave function is expressed as
(r) =

N
r
d
0

(r), (2.43)

(r) is dimensionless and normalized to unity


_
dr[

(r)[ = 1. Substituting the trans-


formations (2.42) and (2.43) in the GPE (2.15) gives the GPE in dimensionless form
_

2
r
+

V
T
(r) +C[

(r)[
2
_

(r) =

(r), (2.44)
where the non-linearity parameter C is dened as
C =
NU
0

r
r
d
0
. (2.45)
The trapping potential (2.36) takes the form

V
T
(r) =
1
4
[ x
2
+ y
2
+ ( z)
2
]. (2.46)
22
2.4. ThomasFermi Approximation
In all chapters of this thesis where numerical results are discussed we will use compu-
tational units exclusively, and for convenience we henceforth omit the tilde denoting
the dimensionless units. It will be clear from the form of the equations and the context
whether SI or computational units are used.
2.4 ThomasFermi Approximation
When there is a large number of atoms in the condensate (large C) the non-linear term
in the GPE (2.44) will dominate the kinetic contribution. In that case, it is a good
approximation to neglect the kinetic energy completely and write the GPE (2.44) in
the ThomasFermi limit as
_
V
T
+C[[
2

= . (2.47)
This has the analytical solution

TF
(r) =
_
_
_
_
V
T
C
for V
T
> 0
0 otherwise
(2.48)
The value of the chemical potential is determined by the normalization condition of
the wave function and depends on the dimensionality of the equation.
Three dimensions
In three dimensions with a trapping potential of the form (2.46), the chemical potential
is given by

3D
=
_
15C
64
_2
5
. (2.49)
Within the ThomasFermi approximation the radial and axial extent of the wave func-
tion are
R
r
= 2
_

3D
, (2.50)
R
z
=
2

3D
, (2.51)
and the peak density at the centre of the condensate is
n
p
=

3D
C
. (2.52)
23
Chapter 2. Mean Field Theory
Two dimensions
In two dimensions, we are mainly interested in elliptical harmonic traps of the form
V
T
=
1
4
[(1 )x
2
+ (1 +)y
2
]. (2.53)
The chemical potential becomes

2D
= [(1 )(1 +)]
1
4
_
C
2
, (2.54)
and the extent of the wave function is
R
x
= 2(1 )

1
4
_

2D
, (2.55)
R
y
= 2(1 +)

1
4
_

2D
. (2.56)
One dimension
Any one dimensional harmonic trap can be cast into the form
V
T
=
1
4
x
2
. (2.57)
Then, the chemical potential is given by

1D
=
_
3C
8
_2
3
. (2.58)
24
Chapter 3
Elementary Excitation Families
In this chapter we consider the excitation spectrum of a BoseEinstein condensate in
a three dimensional cylindrically symmetric trap. We will present a systematic classi-
cation of these excitations that generalizes the concept of families rst identied by
Hutchinson and Zaremba [84]. We will also relate the energy ordering of the modes to
their family classication and provide a simple model that explains this relationship.
The work presented in this chapter extends initial results obtained earlier for my Mas-
ters thesis [102]. The new elements of the work carried out for this PhD thesis are (i)
a complete revision of the procedure to assign a family classication to any mode in
the anisotropic case (section 3.2.3) and (ii) the determination of the energy ordering
of the modes, along with a simple model that give the basis for the explanation of the
ordering (section 3.3). These two signicant extensions of the earlier work have led to
it being published [86].
3.1 Symmetries
In the case of an isotropic harmonic trap, the time-independent GrossPitaevskii equa-
tion (GPE) is completely separable and reduces to a one-dimensional problem in the
radial coordinate r. However, for a cylindrically symmetric harmonic trap of the form
V
T
(x) =
1
2
m
2
r
[x
2
+y
2
+ (z)
2
], (3.1)
where the anisotropy parameter of the trap is dened as the ratio between the axial
and radial trapping frequencies =
z
/
r
, a complete separation of variables is not
possible. Although the ordinary Schrodinger equation is separable in Cartesian or
25
Chapter 3. Elementary Excitation Families
cylindrical coordinates [104], this separation is not possible for the GPE due to the
non-linear term. Nevertheless, solutions of the form
(x) = (, z)e
im
c

(3.2)
can be found, where , and z denote the usual cylindrical coordinates. The magnetic
quantum number m
c
is a good quantum number, and the ground state solution of the
GPE corresponds to m
c
= 0. Since no further separation is possible, the equation has
to be solved in the two variables and z. In non-dimensional units, the GPE takes
then the form
_

_
1

_
+

2
z
2

m
2
c

2
_
+
1
4
_

2
+ (z)
2

+C[(, z)[
2
_
(, z) = (, z).
(3.3)
Correspondingly the normal modes of the Bogoliubovde Gennes (BdG) equations
will also have specic angular momentum compositions if (x) is given by (3.2). If
u
i
(x) is an eigenfunction of

L
z
with eigenvalue m, then v
i
(x) will be an eigenfunction
with eigenvalue (m2m
c
) [105]. Thus, the BdG equations can be written as
_

_
1

m
2

2
+

2
z
2
_
+
1
4
_

2
+ (z)
2

(3.4)
+2C[[
2

_
u
i
(, z) +C
2
v
i
(, z) =
i
u
i
(, z)
_

_
1

(m2m
c
)
2

2
+

2
z
2
_
+
1
4
_

2
+ (z)
2

(3.5)
+2C[[
2

_
v
i
(, z) +C
2
u
i
(, z) =
i
v
i
(, z).
We note that excitations on the ground state (m
c
= 0) with [m[ are degenerate
because m enters the BdG equations quadratically.
The axially symmetric trap potential has also a reection symmetry with respect
to the x-y plane, and thus the solutions to the BdG equations can be chosen to have
a well-dened parity [84]. In the isotropic case, where l and m are good quantum
numbers for the excitations, the parity is simply given by = (1)
lm
.
To solve these equations we employed Matlabs partial dierential equations tool-
box which uses nite element methods based on a triangular segmentation. More details
can be found in appendix A.1 and my Masters thesis [102].
26
3.2. Classication of Excitations
3.2 Classication of Excitations
3.2.1 ThomasFermi limit
In the ThomasFermi limit, Fliesser et al. [106] recognized some underlying symme-
tries of the GPE equation in its hydrodynamic form [107] by identifying three operators
which commute with each other. They introduced three corresponding quantum num-
bers (n, j, m) that classify the solutions completely. An explicit separation of the wave
equation was achieved in cylindrical elliptical coordinates , and , and in terms of
these variables the quantum numbers represent
n: order of polynomial in and
j: index to label dierent eigenvalues for xed n and [m[ ;
j runs from 0 to N = 1 + int
_
n
2

m: z-component of angular momentum.


Although the solutions of the full BdG equations do not strictly conserve these
quantum numbers, we nd that they exhibit in general the same patterns and symme-
tries, and we will show, in the appropriate regime, how the family classication scheme
we develop can be related to n, j, m.
3.2.2 Families in the isotropic case
We rst consider the isotropic case because then the patterns of the dierent mode
families can be easily described in terms of Legendre polynomials. As the trap geom-
etry is changed from spherical to cylindrical symmetry these patterns are continously
modied, being squeezed in the direction of the stronger connement, but the basic
character remains recognizable.
In an isotropic trap, the solutions for the excitations can either be completely sep-
arated in spherical coordinates as u(x) = u
r
(r)Y
lm
(, ), or partially separated in
cylindrical coordinates as u(x) = u(, z)e
im
. In the latter case, u(, z) is essentially
the radial function u
r
(r) modulated by the Legendre polynomial P
lm
(cos ), where
cos = z(
2
+ z
2
)
1/2
. Thus, the general shape of the families is determined by the
symmetries of the Legendre polynomials. We now show that the family classication
suggested by Hutchinson and Zaremba [84] can be generalized, in the isotropic case, as
follows. First we assign a principal family number which is given by
f = l [m[ + 1, (3.6)
27
Chapter 3. Elementary Excitation Families
and then an additional number characterizing the radial function is needed to complete
the classication into families. We shall introduce the nodal family number n
r
for
this purpose, which in the isotropic case is simply the number of nodes in the radial
function. The family is given by the pair (f, n
r
), which together with the magnetic
quantum number m uniquely species any mode. A generalization of this classiction
to the anisotropic case is presented in section 3.2.3.
We illustrate the spatial character of the family assignment by considering rst the
excitation modes with no radial node (n
r
= 0). We begin with the case m ,= 0, which
we illustrate in gure 3.1 with contour plots of full numerical solutions of u(, z) for the
specic case of the degenerate modes of the lowest l = 3 excitation with m = 3, 2, 1.
Their principal family numbers are f = 1, 2, 3 respectively. Since the radial function
4 2 0 2 4
6
4
2
0
2
4
6

Family (1,0)
xposition
z

p
o
s
i
t
i
o
n
(a)
4 2 0 2 4
+ +

Family (2,0)
xposition
(b)
4 2 0 2 4

+ +

Family (3,0)
xposition
(c)
Figure 3.1: General shape of mode families 1 to 3 with no radial node. Contour plots in
the x-z plane of the quasiparticle amplitude u(, z) are given for the degenerate
modes l = 3, (a) m = 3, (b) m = 2, (c) m = 1 modes.
is the same for each of these modes, the relative overall shape is determined by the
Legendre polynomials. The important property of the Legendre polynomials P
lm
(cos )
for our purposes is that they have n

= l [m[ nodes between 0 < < . Thus, the


number of angular nodal surfaces n

beween 0 < < , i.e. surfaces of zero density that


are characterized by a constant value of in the isotropic case, determines the principal
family number f since f is given by equation (3.6) as f = l [m[ + 1 = n

+ 1. We
note also that for m ,= 0 all Legendre polynomials are zero along the z-axis, but this is
not a nodal surface. Because the sign of the wave function changes as it crosses a nodal
surface, family 1 members have even parity, family 2 have odd parity, and in general the
parity of the mode is related to the principal family number by = (1)
f1
= (1)
lm
.
28
3.2. Classication of Excitations
The m = 0 member of each family has a shape that derives from the P
l0
Legendre
polynomial. We call it the anomalous member of the family, since its shape diers from
other members of the family only in that it is non-zero along the symmetry axis, which
does not change the character of the excitation signicantly. We illustrate the shape
of the anomalous modes of the families (2, 0) and (3, 0) in gure 3.2. The anomalous
6 4 2 0 2 4 6
6
4
2
0
2
4
6
+

(a)
Family (2,0)
xposition
z

p
o
s
i
t
i
o
n
6 4 2 0 2 4 6
+
+

(b)
Family (3,0)
xposition
Figure 3.2: Contour plots in the x-z plane of the amplitude u(, z) of the anomalous rst
members of family 2 and 3 (m = 0).
member of family (1, 0) is the ground state, which is a solution of the BdG equations
[100].
The case where the radial function has a non-zero number of nodes (i.e. n
r
,= 0)
can now be easily visualized. The principal family number f determines the number of
angular nodal surfaces (f 1) between 0 < < , while n
r
determines the number of
radial nodal surfaces, which intersect the angular nodal surfaces. In the isotropic case
they are spherical and centered on the origin. In gure 3.3 we illustrate the rst two
modes having one node in the radial function, which both belong to the family (1, 1).
The mode in gure 3.3 (a) is the anomalous rst member of this family (m = 0), while
all other modes of this family have the general shape shown in gure 3.3 (b), which
can be recognized as the same shape as in gure 3.1 (a), but with one radial node.
We stress that all members of the same family, apart from the anomalous one, have
the same general shape, i.e. the same number of peaks (in the contour plot) in similar
spatial distribution. The main qualitative dierence between modes of the same family
is that the peaks move radially outwards and become narrower in both radial and
azimuthal direction with increasing eigenfrequency. This is illustrated in gure 3.4 for
the so-called surface modes (n
r
= 0) of a condensate.
29
Chapter 3. Elementary Excitation Families
5
0
5
5
0
5
0
0.1
0.2
Family (1,1)
xposition
zposition
A
m
p
l
i
t
u
d
e

u
(

,
z
)



(a)
5
0
5
5
0
5
0.05
0
0.05
0.1
0.15
Family (1,1)
xposition
zposition
A
m
p
l
i
t
u
d
e

u
(

,
z
)



(b)
Figure 3.3: Family 1 with one radial node (n
r
= 1). (a) Anomalous rst member (l = 0,
m = 0), (b) general shape (here l = 1, m = 1).
0 2 4 6 8 10
0.05
0
0.05
0.1
0.15
0.2
0.25
position
r
a
d
i
a
l

u
Family 1
Figure 3.4: Quasiparticle amplitude u along the -axis for family 1 modes with n
r
= 0. The
curves from left to right correspond to (l, m) = (1, 1), (2, 2), . . . , (5, 5) respec-
tively. The non-linearity is C = 332.
To illustrate the mode classication by family and m value, we list in table 3.1 the
18 lowest lying modes for the C = 332 case in an isotropic harmonic trap.
3.2.3 Anisotropic case
The main value of the concept of families is in its extension to the anisotropic cylin-
drically symmetric case. Hutchinson and Zaremba identied the rst four families by
the dependence of the eigenvalue on trap anisotropy. Here, we show that the mode
topology determines the family.
In gure 3.5 we have plotted the quasiparticle wave functions for three families: the
30
3.2. Classication of Excitations
Mode Family Mode Family
l m f n
r
l m f n
r
0 0 0.000 1 0 0 0 2.193 1 1
1 0 1.000 2 0 4 0 2.660 5 0
1 1 0 1 4 0
2 0 1.526 3 0 2 3 0
1 2 0 3 2 0
2 1 0 4 1 0
3 0 2.065 4 0 1 0 2.872 2 1
1 3 0 1 1 1
2 2 0
3 1 0
Table 3.1: Lowest quasiparticle modes of a condensate in an isotropic trap for C = 332,
listed by family.
anomalous member of the (3, 0) family, and the (2, 1) and (3, 1) families, for the case
of prolate, spherical and oblate traps. These graphs illustrate the general features that
are found in the anisotropic case of cylindrical symmetry. We see that just as for the
ground state of the GPE the quasiparticles are squeezed in the direction of the stronger
connement of the trap and expand in the other direction. This distortion has no eect
on the zero line along the symmetry axis, but the nodal surfaces of the isotropic case
are distorted. The radial nodal surfaces are no longer spherical, but are changed to a
shape that approximately follows the equipotentials of the trap. In all cases the total
number of nodal crossings along the positive half of the axis of strong connement
remains exactly n
r
. The angular nodal surfaces can no longer be parameterized by
constant , and they do not meet at the origin: in the prolate case they generally
intersect the symmetry axis at various (non-zero) values of z, while in the oblate case
they generally intersect the z = 0 plane in circles of dierent radii. We will call the
former planar nodal surfaces and the latter cylindrical nodal surfaces, which is a good
description in highly anisotropic traps as we will see in section 3.3.1. Exceptions are
rare and occur for the case where the trap is close to spherical. Then, it is possible in
the prolate case that adjacent planar nodal surfaces join just before meeting the z-axis,
while in the oblate case it is possible that adjacent cylindrical nodal surfaces join just
before the z = 0 plane.
The nodal crossings of the isotropic case, which are at the centre of four peaks of
alternating sign, can also change character in the anisotropic case, as illustrated in
the right hand column of gure 3.5 (a). There we see that pairs of peaks of the same
sign can begin fusing; in other words the crossing has become a saddle point, or anti-
31
Chapter 3. Elementary Excitation Families
8
4
0
4
8

+
+
= 0.5
(a)
Family (3,0)
z

p
o
s
i
t
i
o
n

+ +
+ +

= 0.5
Family (2,1)
+ +
+ +


+ +
= 0.5
Family (3,1)
8
4
0
4
8
+
+

= 1
(b)
z

p
o
s
i
t
i
o
n

+ +
+ +

= 1


+ +
+ +
+ +
= 1
8 4 0 4 8
8
4
0
4
8
+
= 8
(c)
xposition
z

p
o
s
i
t
i
o
n
8 4 0 4 8

+ +
+ +

= 8
xposition
8 4 0 4 8



= 8
xposition
Figure 3.5: Eects of trap anisotropy on quasiparticle shapes. Left hand column f = 3,
n
r
= 0, m = 0 mode; middle column f = 2, n
r
= 1, m = 1 mode; right hand
column f = 3, n
r
= 1, m = 1 mode. The anisotropy parameter is (a) = 0.5,
(b) = 1, (c) =

8.
crossing. Despite these distortions, the character of the isotropic excitations remains
clearly evident in the anisotropic case, and thus a family assignment can be made.
The rst step is to determine the number of radial nodes (by inspection along the
positive part of the axis of strong connement), which gives the radial family number
n
r
. Next, if there is no radial node (i.e. n
r
= 0), we count the number n
p
of distinct
peaks in a single quadrant (including peaks that are centered on the axes). If a radial
32
3.3. Ordering of Quasiparticle Eigenfrequencies
node does exist, we identify the rst radial nodal surface, and count the number n
p
of
distinct peaks in a single quadrant inside this radial nodal surface. In either case, the
principal family number is then given by f = 2n
p
1 for even modes and f = 2n
p
for odd modes. The n
r
= 1 nodal surface can be traced out by following a path from
the initial point on the axis of strong connement, through successive crossings or
anti-crossings, until the axis of weak connement is reached. When anti-crossings are
encountered, the path continues over the saddle (along the line of minimum amplitude)
to the nodal line leaving the anti-crossing opposite to the entry. Of course this line
(or radial surface) is no longer strictly nodal, but nevertheless serves to determine a
region which is characteristic of the family. Just as for the isotropic case, a mode
is uniquely identied when the family assignment (f, n
r
) is given together with the
magnetic quantum number m, and the parity is given by = (1)
f1
.
We can relate the quantum numbers n and j as introduced by Fliesser et al. [106]
in the ThomasFermi limit to our family assignment as follows:
f = n 2j + 1, n
r
= j. (3.7)
The similarity of shape of modes of the same family explains the similarity of fre-
quency dependence on anisotropy for modes in a given family found by Hutchinson
and Zaremba [84]; changes in trap geometry aect every mode in the family in much
the same way.
3.3 Ordering of Quasiparticle Eigenfrequencies
In this section we show how the family classication can be related to the energy
ordering of the quasiparticles for given m, and we explain the relationship comparing
the solutions to the harmonic oscillator states.
3.3.1 Full solutions of Bogoliubov-de Gennes equations
In gure 3.6 we illustrate the shape of the rst few family members with n
r
= 0 and 1,
for the case of very prolate and very oblate traps (gures 3.6 (a) and (b) respectively).
We have chosen to represent the m = 0 modes, because the mode shapes are slightly
simpler than the m ,= 0 modes (which dier only in having a zero along the z-axis). In
highly anisotropic traps, the family shapes follow very well dened patterns, as we can
see, and it is convenient to discuss the prolate and oblate case separately.
33
Chapter 3. Elementary Excitation Families
12
0
12
(1,0)
+
Even
(a)
6 0 6
12
0
12
(3,0)
+

z
(1,1)
+
Even
6 0 6
(3,1)
+ +

+
+

(2,0)
+

Odd
6 0 6
(4,0)

+
x
(2,1)

+
+

Odd
6 0 6
(4,1)
+

+
2
0
2
(1,0)
+
Even
(b)
2
0
2
(3,0)
+
2
0
2
(1,1)
+

+
z
4 0 4
2
0
2
(3,1)
+
+
+

+
+
x
(2,0)
+

Odd
(4,0)
+

+
+

(2,1)

+
+

4 0 4
(4,1)

+
+

+
+

Figure 3.6: Highly anisotropic quasiparticle eigenstates in (a) prolate trap with 1/ = 7 and
(b) oblate trap with = 7. The cases shown are the f = 1, 2, 3, 4 anomalous
(m = 0) family members with n
r
= 0 and 1, respectively. The contour plots
show contour lines only at 0.001.
Prolate case:
For the prolate case an increase in the principal family number f simply adds an ad-
ditional planar nodal surface perpendicular to the z-axis while changing n
r
from 0 to
1 adds a radial nodal surface which appears in this projection as two lines symmetric
about, and almost parallel to the z-axis. The energy ordering of these modes is plotted
in gure 3.7 (a) where we see that, for a given anisotropy, the energy ordering initially
follows the family assignment (f, 0) with alternating even and odd modes. This se-
quence is interrupted by the (1, 1) mode, at an energy near 2, where it has become
more favourable to have a single radial nodal surface than many planar ones across
the narrow dimension. We note that the higher the anisotropy (i.e. the larger 1/) the
more (f, 0) modes t in before the (1, 1) mode becomes favourable.
Oblate case:
Figure 3.6 (b) shows that in the oblate case it is natural to separate the sequences of
even and odd modes. Then, successive members of each sequence (for xed n
r
) are
obtained by adding an additional cylindrical nodal surface about the z-axis. Changing
n
r
from 0 to 1 adds again a radial nodal surface.
34
3.3. Ordering of Quasiparticle Eigenfrequencies
2 4 6 8
0
1
2
3
4

(1,1)
(2,1)
(a)
Anisotropy 1/
E
i
g
e
n
f
r
e
q
u
e
n
c
y
C = 332, m
u
= 0
even
odd
2 4 6 8
0
5
10
15
20
25
30
(2,1)
(1,1)
(b)
Anisotropy
E
i
g
e
n
f
r
e
q
u
e
n
c
y
C = 332, m
u
= 0
even
odd

Figure 3.7: Frequency spectrum versus anisotropy of the low-lying n
r
= 0 modes and the
rst two n
r
= 1 modes (m = 0) in (a) prolate, (b) oblate geometry. The even and
odd modes with n
r
= 0 are shown as dotted and dashed lines, respectively, and
the n
r
= 1 modes as dashed-dotted lines. The solid line shows the condensate
chemical potential . The mode eigenfrequencies are measured relative to the
condensate eigenvalue .
In gure 3.7 (b), where the energies of the oblate modes are plotted as a function of
the anisotropy , the distinction between the even and odd modes is clearly revealed.
The odd modes are shifted as a group to higher energies than the even modes (reecting
the energy cost of the nodal surface through z = 0), so that the energy ordering at a
given anisotropy no longer simply alternates between even and odd. We see too that
the even modes have essentially constant spacing except for very low , and the spacing
of the odd modes, while compressed for low excitation numbers, becomes equal to the
spacing of the even modes for higher excitation numbers. Once again, as for the prolate
case, the higher the anisotropy the more (f, 0) modes t in before the rst n
r
= 1 mode.
3.3.2 Comparison with harmonic oscillator solutions
For higher excitations the quasiparticle amplitude v
i
is negligible compared to u
i
and
the BdG equations (2.30) reduce to the eigenvalue problem
[

H
0
+ 2NU
0
[
g
(r)[
2
]u
i
(r) =
i
u
i
(r), (3.8)
which has the form of a single particle equation [108, 109]. The presence of the conden-
sate gives an eective potential that is broadly harmonic, but with a repulsive dimple
35
Chapter 3. Elementary Excitation Families
in the middle, of height approximately . At high excitations, the energy levels for
equation (3.8) will be essentially those of the harmonic oscillator, while for mode ener-
gies comparable or less than , the presence of the dimple will shift the energy levels
upwards, eectively compressing them.
If the condensate density is ignored in equation (3.8) the equation is simply that of
the harmonic oscillator and is separable in cylindrical coordinates. The energy levels
are given in units of
r
by
E = [m[ + 1 + 2n

+ 2(n
z
+
1
4
) (even modes), (3.9)
E = [m[ + 1 + 2n

+ 2(n
z
+
3
4
) (odd modes), (3.10)
where n

and n
z
are the number of nodes in the - and z-wave function respectively
(n

= 0, 1, . . ., n
z
= 0, 1, . . .). Equations (3.9) and (3.10) provide considerable insight
into the behaviour of the solutions we have obtained from the full BdG equations.
Prolate case:
In the prolate case, increasing n
z
corresponds to increasing f in the BdG solutions,
while increasing n

corresponds to increasing n
r
in the BdG solutions. For < 1 it
becomes energetically favourable to increase n
z
rather than n

, and we can t roughly


1/ nodes in the z-direction before it is energetically favourable to put a node in the
-direction. In gure 3.8 we show from calculations of the BdG equations the number
N
(f,0)
of (f, 0) modes of lower energy than the rst n
r
= 1 mode and compare this
qualitatively to the predictions from equations (3.9) and (3.10). We see that while
there is relatively good agreement at C = 332, the full BdG calculation predicts more
(f, 0) modes to t in, and the discrepancy increases at higher values of C. This is due
to the fact that in the full equations the presence of the condensate causes the spacing
of the lower levels (up to an energy of order above the condensate) to be compressed,
and increases with C.
Oblate case:
In the oblate case, the agreement between the harmonic oscillator solutions and the
BdG solutions is much better. As explained in the previous section, it is useful to
separate the even and the odd modes. From gure 3.6 (b) we realize that for even
modes n
z
corresponds to 2n
r
, and n

to (f 1)/2 while for odd modes n


z
corresponds
to 2n
r
+ 1, and n

to (f 2)/2. Since > 1, it is clear from equations (3.9) and


36
3.3. Ordering of Quasiparticle Eigenfrequencies
2 4 6 8
0
5
10
15
20
25
30
Anisotropy 1/
N
(
f
,
0
)
HO
C=332
C=10000
Figure 3.8: Comparison of energy ordering predictions of full BdG and harmonic oscillator
solutions in a prolate trap. Plotted points show N
(f,0)
, the number of n
r
= 0
modes with energy lower than the (1, 1) mode. All modes are m = 0.
(3.10) that increasing n
z
, i.e. increasing n
r
in the BdG solutions, is more energetically
costly than increasing n

, which corresponds to increasing f. Figure 3.9 provides a


quantitative comparison of the predictions of the harmonic oscillator solutions and the
BdG solutions for the number of (f, 0) modes of lower energy than the rst n
r
= 1
mode. We have separated the even and odd modes so that gure 3.9 (a) shows the
number N
(f,0)
even
of even modes with eigenfrequency lower than that of the (1, 1) mode,
while gure 3.9 (b) shows the number N
(f,0)
odd
of odd modes with eigenfrequency lower
2 4 6 8
0
2
4
6
8
10
12
14
(a)
Anisotropy
N
e
v
e
n
(
f
,
0
)
HO
C=332
C=10000
2 4 6 8
0
2
4
6
8
10
12
14
(b)
Anisotropy
N
o
d
d
(
f
,
0
)
HO
C=332
C=10000
Figure 3.9: Comparison of energy ordering predictions of full BdG and harmonic oscillator
solutions in the oblate case for (a) the number N
(f,0)
even
of even n
r
= 0 modes with
energy lower than (1, 1) mode, (b) the number N
(f,0)
odd
of odd modes with energy
lower than the (2,1) mode. All modes are m = 0.
37
Chapter 3. Elementary Excitation Families
than that of the (2, 1) mode. The agreement is excellent for the C = 332 case and
still very good for the C = 10000 case and is much better than for the prolate case of
corresponding asymmetry, because the chemical potential lies lower relative to the
eigenfrequencies in question, as can be seen in gure 3.7.
Finally, we remark that the magnetic quantum number m simply gives an energy
oset in equations (3.9) and (3.10), i.e. modes with a given number of nodes have higher
energies for higher m. It is clear therefore, that the results discussed above for m = 0
will also apply for m ,= 0. The energy oset by m also explains that for a given family
(f, n
r
) (i.e. a given number of nodes) the energy increases with increasing m (see table
3.1).
38
Chapter 4
Background on Vortices
A large part of this thesis is concerned with vortices in rotating BoseEinstein conden-
sates. A vortex is an irrotational ow pattern which is typically a circular ow with a
1/r velocity dependence around a core with zero density. Examples known from classi-
cal physics are whirlpools in water or tornados in air. They are of particular interest in
the study of superuids [110] because superuids do not support rotational (e.g. solid-
body) motion. If one tries to set a superuid into rotation, the superuid will resist
any rotation below a certain critical rotation speed. Above the critical angular velocity,
however, it can take on angular momentum by forming vortices.
It was recognized by London in 1938, that the phenomenon of superuidity in liquid
4
He was probably related to BoseEinstein condensation. However, strong interactions
cause a depletion of as large as 90% in
4
He, which obscures the simple phenomenon
of BoseEinstein condensation predicted on the basis of weakly interacting bosons.
With the advent of almost pure BoseEinstein condensates in dilute alkali gases [1],
the opportunity has arisen to test and extend the theories developed throughout the
research on superuid Helium.
In this chapter we will review well known results from the superuid Helium liter-
ature [110] and more recent results concerning vortices in BoseEinstein condensates
[111]. In the literature on superuid Helium it is commonly assumed that the density
is homogeneous because, even if Helium is trapped in a bucket, the density is constant
throughout the vessel. A BoseEinstein condensate in a dilute gas, on the other hand,
is usually conned in a harmonic trap, and its density falls down with increasing dis-
tance from the trap centre. This leads to some important dierences in the physical
properties of a BoseEinstein condensate compared to superuid Helium.
39
Chapter 4. Background on Vortices
4.1 Quantization of Circulation
In the mean eld description a BoseEinstein condensate is represented by a single
particle wave function (r, t), and this can always be written in terms of its amplitude
and phase as
(r, t) = [(r, t)[e
iS(r,t)
. (4.1)
While the absolute phase has no physical signicance, the relative phase of the con-
densate is important. The wave function has to be single-valued, but the phase is only
dened within multiples of 2. Thus, the phase circulation, which is the change of the
phase over a closed path, is
_
S dl = 2n, (4.2)
where n = 0, 1, . . . is the winding number. It is worth pointing out that a non-zero
phase circulation is only possible if the region is multiply connected. That means the
density has to go to zero somewhere in the enclosed region, so that the phase is not
dened there. In the literature on superuid Helium it is conventional to express this
result in terms of the superuid velocity v
s
. Using the form of (r, t) from equation
(4.1) the current density
j(r, t) =

2mi
[

)] (4.3)
can immediately be written in the hydrodynamic form
j(r, t) = n(r, t)v
s
(r, t), (4.4)
where n(r, t) = [(r, t)[
2
is the condensate density and the superuid velocity is dened
as
v
s
(r, t)

m
S(r, t). (4.5)
The ow pattern can be characterized by the quantity
_
v
s
dl known as circulation.
With the denition (4.5) and the result for the phase circulation (4.2), it follows im-
mediately that the circulation is quantized
_
v
s
dl = n, (4.6)
40
4.2. Characteristics of Vortex
where is the quantum of circulation

h
m
. (4.7)
Relation (4.6) is known as the OnsagerFeynman quantization rule.
4.2 Characteristics of Vortex
An important property of a superuid is that the ow is irrotational. This follows from
the denition of the superuid velocity (4.5), which implies
v
s
=

m
S = 0. (4.8)
This result means in particular that a superuid cannot support a rigid-body rotation.
That does not mean that a superuid cannot aquire angular momentum. Recati et
al. [112] have shown that there are eigenstates of a BoseEinstein condensate in a
rotating elliptical trap that have a small amount of angular momentum even without
the presence of a vortex. In general, however, when more than a certain critical angular
momentum is transferred to a superuid it can only be supported by the formation of
vortices. This is the case, because the vortex states have minimum energy for a given
angular momentum.
A vortex is a topological defect and is characterized by two properties: a line of
zero density and a phase circulation of 2n around this line. While the ow around the
vortex is irrotational as required by equation (4.8), the core of the vortex corresponds
to a singularity that carries the local vorticity v
s
= n.
In practice, the phase circulation of a vortex is always 2 because a vortex of
higher winding number n is energetically unstable and quickly decays into [n[ singly-
quantized vortices of the same orientation. That is the case because the energy of a
vortex goes n
2
[13]. The total circulation along any closed path within a superuid
is simply the sum of the circulations of the enclosed vortices. The circulation within
the enclosed region is conserved unless a vortex passes through the boundary of the
region. This does not mean, however, that the number of vortices in the region is xed.
Two vortices of opposite circulation can be created or annihilated.
41
Chapter 4. Background on Vortices
4.3 Healing Length
Let us consider the circulation around a singly-quantized vortex line. From equation
(4.6) we obtain
_
v
s
dl = 2v
s
(r) = , (4.9)
where r is radial distance form the vortex line. This gives the superuid velocity
v
s
(r) =

2r
. (4.10)
The dependence of the velocity on 1/r predicts a divergence as r 0. This is only
avoided by the fact that the density goes to zero at the core of the vortex. The typical
distance over which the condensate density heals back to its bulk value is called the
healing length . The healing length can be estimated from a simple argument that only
considers the change in density, and not the energy associated with the ow pattern
of a vortex. If the condensate density grows from zero to n
s
within a distance , the
energy that comes from the quantum pressure, i.e. the part of the kinetic energy due
to the change in density, is
2
/(2m
2
), and the interaction energy U
0
n
s
where
U
0
= 4
2
a/m. Equating them gives the following expression for the healing length
=

2mn
s
U
0
= (8n
s
a)

1
2
. (4.11)
In an inhomogeneous superuid, as in the case of a trapped BoseEinstein conden-
sate, it is customary to dene the healing length in terms of the central density n
s
(0)
= (8an
s
(0))
1/2
.
4.4 Energy of Vortex Line
The main contribution to the energy of a vortex line comes from the kinetic energy of
the circulating superuid. For a vortex line of unit length this is
E
v
=
_
R
0

n
s
v
2
s
rdr, (4.12)
where R
0
is some outer radial cut-o. The lower cut-o is taken as the healing length
because the density goes quickly to zero for smaller radii. For a single vortex line
R
0
might be given by the radius of the container in the case of superuid Helium, or
42
4.5. Critical Frequency of Vortex Nucleation
the ThomasFermi radius in the case of a trapped BoseEinstein condensate. If other
vortices are present, R
0
is of the order of half the distance between the vortices. With
expression (4.10) we obtain the energy [110]
E
v
=
n
s

2
4
ln
_
R
0

_
. (4.13)
A numerical analysis based on the GrossPitaevskii equation, that includes the energy
contribution due to the local compression of the superuid around the vortex core yields
[111]
E
v

n
s

2
4
ln
_
1.46
R
0

_
, (4.14)
which is the same as (4.13) apart from an additive numerical constant.
4.5 Critical Frequency of Vortex Nucleation
The critical frequency of vortex nucleation in a BoseEinstein condensate has been a
topic of greate debate. In experiments at ENS in Paris the critical angular velocity was
measured to be 0.7

[73, 74], which is much higher than one would expect from
thermodynamic considerations or the Landau criterion that determines the minimum
velocity to create any excitation in a superuid with a moving perturber. Here, we
present the dierent approaches for calculating the critical velocity that have appeared
in the literature and summarize the consensus which seems to have emerged on this
question, which is supported by our own work presented in chapter 7.
4.5.1 Thermodynamic critical frequency
In this section, we introduce the thermodynamic critical frequency that comes from
energetic considerations of the vortex state compared to the vortex free state of the
superuid. A vortex becomes energetically favourable when its appearance leads to a
decrease of the quantity
F

= F L , (4.15)
where F is the free energy of the rotating uid and L is its total angular momentum
[110].
Let us rst consider the case of superuid Helium in a container of radius R
0
. A
43
Chapter 4. Background on Vortices
vortex is energetically favourable when
E
v
l
v
(R
0
) (4.16)
becomes negative, where l
v
(R
0
) is the angular momentum of a single vortex whose ow
pattern lls the entire container of radius R
0
. The angular momentum is
l
v
=
_
R
0

n
s
v
s
(2r)rdr =
1
2
n
s
R
2
0
, (4.17)
where it is assumed R
0
. Using expression (4.13) and (4.17) in (4.16) the following
estimate of the critical angular frequency for vortex nucleation is obtained

v
=

mR
2
0
ln
_
1.46
R
0

_
. (4.18)
The thermodynamic critical rotation frequency for a trapped BoseEinstein con-
densate is obtained similarly to the homogeneous case. At a certain rotation frequency

v
the vortex state becomes energetically favourable. This can be expressed simply
in terms of the energies of a condensate with and without a vortex evaluated in the
laboratory frame

v
=
E
1
E
0
N
,
where N is the number of atoms in the condensate. Evaluating the energies in the
ThomasFermi limit for an axisymmetric trap leads to the estimate [111]

v

5
2

mR
2

ln
_
0.671
R

_
, (4.19)
where R

is the ThomasFermi radius of the cloud in the x-y plane perpendicular to


the vortex line. Note that this estimate exceeds the estimate of the homogeneous case
(4.18).
4.5.2 Landau criterion
The Landau criterion states that no excitations can be generated in a superuid by a
moving obstacle if its velocity is smaller than the critical velocity [113, 13]
v
c
= min
_
E
p
p
_
, (4.20)
44
4.5. Critical Frequency of Vortex Nucleation
where E
p
and p are the energy and momentum of an excitation [113]. Raman et al. [66]
found evidence for this critical velocity by the observation of strong heating only above
the critical velocity when a condensate was disturbed with a small laser beam that
moved on a line within the condensate in an oscillating manner. This is in agreement
with numerical simulations [114].
As we will discuss in section 4.5.5, the surface modes of a trapped BoseEinstein
condensate play an important part in the nucleation of vortices. In terms of the excita-
tion frequencies of the surface modes
l
and their angular momentum quantum number
l along the z-direction (azimuthal quantum number m = l) the Landau criterion reads
[109]

c
= min
_

l
l
_
. (4.21)
4.5.3 Stability and energy barrier
Before the rst experimental observation of a vortex in a BoseEinstein condensate
[72, 73], the main theoretical question was whether a vortex in a condensate would be
stable at all, and therefore observable. Rokhsar [115] found that a single vortex in a
trapped BoseEinstein condensate is not stable at T = 0 if an imposed rotation ceases,
that is necessary to create the vortex. This in contrast to superuid Helium where
the circulating ow persists in the absence of a rotating drive. Isoshima et al. [116]
argued that at nite temperature the vortex is eectively stabilized by the presence of
the thermal cloud.
With a rotating drive, the vortex state becomes of course stable as discussed in sec-
tion 4.5.1 when the rotation frequency exceeds the thermodynamical critical frequency.
But there are also rotation frequencies for which a vortex state is only locally stable or
metastable, not globally stable [117, 118].
The reason that the critical angular velocity for vortex nucleation is denitely higher
than the thermodynamic critical frequency is that there exists an energy barrier for
vortex penetration at the outer surface of the condensate [118, 119]. This energy
barrier becomes narrower as the rotation frequency is increased, but remains even for
rotation frequencies where the energy of the vortex state is a global energy minimum.
4.5.4 Anomalous mode and vortex dynamics
The dynamics of a single vortex line in a BoseEinstein condensate has been considered
in detail by Svidzinsky and Fetter [120, 111]. Here, we only mention the signicance of
45
Chapter 4. Background on Vortices
the anomalous modes, which have a negative excitation energy in the laboratory frame
and a positive norm. An axisymmetric condensate with a singly quantized vortex has
at least one anomalous mode [105, 115] in a non-rotating frame. The anomalous mode
corresponds to the precession of an o-centre vortex around the origin [121, 118], which
can also be understood in terms of the Magnus force acting on the vortex line that
arises from the mutual friction between the vortex and the thermal cloud [122]. In a
rotating frame, the anomalous frequency increases linearly with the angular frequency
of rotation, and the single vortex becomes locally stable when the frequency of the
anomalous mode exceeds zero.
4.5.5 Nucleation of vortices
To create a vortex, angular momentum has to be transfered to the condensate. The
most common method is to use an ellipsoidal rotating trap to stir a cloud of condensed
atoms [123, 77]. It is also possible to condense into a vortex state from a rotating
thermal cloud without the continued need of an anisotropic stirring potential [78]. In
either case, the vortex has to overcome the energy barrier for vortex penetration because
the circulation in the condensate can only change if a vortex enters the condensate from
the outer regions.
It has been argued [124, 125] that vortices are nucleated at the condensate border
from shape oscillations corresponding to surface modes. This can only happen if the
rotation speed given by the Landau criterion (4.21) is exceeded. A careful analysis of
the surface excitations by Anglin [126, 127] revealed that the energy barrier for vortex
penetration disappears when the Landau critical velocity is exceeded.
In light of these arguments the question remains as to why the critical angular
velocity observed in [73, 74] is much higher than one would expect from the Landau
criterion. One argument was that only for the high angular velocity observed did the
lowest anomalous mode disappear, thereby making the vortex state locally stable [128].
However, by now experiments have shown that the anomalous mode instability does not
inhibit vortex nucleation [76, 79]. Sinha and Castin [129] found dynamical instabilities
for certain stirring procedures with an anisotropic trap that could provide an initiating
mechanism for the production of vortices. Finally, Dalfovo and Stringari [130] argued
that an ellipsoidal rotating trap mainly excites the l = 2 quadrupole mode. Hence, one
should use the excitation frequency of the quadrupole mode in the Landau criterion
(4.21), which gives a universal critical angular velocity of
c,l=2
0.71

.
Our theoretical analysis, as presented in chapters 7 and 8, reveals that an expo-
46
4.6. Vortex Lattice
nential growth in the population of surface modes leads to the formation of vortices.
The vortices appear from the interference of highly occupied surface modes with the
condensate wave function, initially occuring at innity and moving towards the dense
region of the condensate as the population in the surface modes increases. Hence, the
Landau criterion in principle gives the correct critical frequency. However, in practice,
the initiation process might be too slow to lead to a visible vortex in the bulk of the
condensate density [127] unless the condensate is driven on resonance with a certain
surface mode following Dalfovos and Stringaris argument [130]. This may also explain
the fact that in experiments with stirring potentials of dierent multipolarities discrete
resonances for vortex nucleation were found [77].
4.6 Vortex Lattice
In the presence of a uniform array of vortex lines, a superuid can mimic a solid-body
rotation. While the ow pattern is strictly irrotational away from the vortex cores,
on average the rotation is solid-body-like with the vortex array rotating at constant
angular velocity.
A solid-body rotation can be described by the velocity v
sb
= r at position r.
This kind of motion has the constant vorticity v
sb
= 2. Using Stokes theorem,
it follows that the circulation of a solid-body rotation is
=
_
v
s
dl = 2A
v
, (4.22)
where A
v
is the enclosed area. Let us now consider a vortex array with N
v
vortices [131].
The total circulation along a closed contour containing the complete vortex lattice is
=
_
v
s
dl = N
v
. (4.23)
In order to mimic a solid-body rotation the circulation has to be = 2A
v
. Thus, the
areal vortex density is given by
n
v
=
N
v
A
v
=
2

. (4.24)
We note that the vortex density depends only on the angular velocity , not the
superuid density.
The lattice structure of the rotating vortex array can be obtained by minimizing the
47
Chapter 4. Background on Vortices
quantity (4.15). It turns out that it is energetically more favourable to have an array of
lines with small circulation than, say, a single line with large circulation. Tkachenko has
shown that the quantity (4.15) is not only minimized if the vortex lines form a triangular
lattice [132], but also that this conguration is stable against small perturbations [133].
In a triangular lattice the vortex lines are most closely packed, which leads to the typical
hexagonal arrangement of the nearest neighbours of each vortex. In the superuid
literature this lattice structure is also known as Abrikosov lattice after Abrikosov who
predicted such an arrangement of vortices in superconductors of type 2 [134].
There has been a large number of publications that study the stationary lattice
states of rotating BoseEinstein condensates with small and moderate numbers of vor-
tices numerically [135, 136, 137, 138]. The obtained lattice structures are in good
agreement with experiments [73, 74, 76]. The formation of very regular Abrikosov lat-
tices in rotating BoseEinstein condensates has been most impressively demonstrated
by the MIT group [79, 80] with lattices of more than hundred vortices.
48
Chapter 5
Theory of Growth from Rotating
Thermal Cloud
In this chapter, we will review the theory behind the phenomenological equation used in
our dynamical simulations of the growth of a BoseEinstein condensate from a rotating
thermal cloud. The main motivation of our work was to a nd a model that would
describe the formation of a vortex lattice, and a central feature incorporated in the
model is that the trapped BoseEinstein condensate is immersed in a bath of thermal
atoms. The dynamics of the BoseEinstein condensate is described by a modied
GrossPitaevskii equation (GPE) that allows particle exchange between the condensate
and the thermal cloud. The thermal cloud is taken to be an innite bath of atoms with
a constant temperature and chemical potential. While we do not treat the thermal
atoms dynamically, they are considered to be part of a cloud rotating with xed angular
velocity.
5.1 Quantum Kinetic Theory
The challenge in describing a BoseEinstein condensate interacting with a thermal
cloud is the fact that, on one hand, the mutual interaction between the particles in the
condensate and the thermal cloud change the one-particle wave functions and energies
of the gas, and on the other hand, alter the occupation numbers of those states. While
the former processes are coherent, the latter are incoherent. The non-equilibrium dy-
namics of such a system is in principle quite complicated, and its description requires
the combination of kinetic theory and quantum mechanics to incorporate both eects
simultaneously. Such quantum kinetic theories have been developed by several groups
49
Chapter 5. Theory of Growth from Rotating Thermal Cloud
using various techniques. The work in this thesis is based on the quantum kinetic the-
ory developed by Gardiner, Zoller and co-workers who have applied stochastic methods
commonly used in quantum optics. It has been presented in a series of papers on Quan-
tum Kinetic Theory we will refer to as QKI [139], QKII [140], QKIII [141], QKIV [142],
QKV [143], QKVI [144], and QKVII [145]. This theory has been particularly successful
describing the formation and growth of a BoseEinstein condensate in excellent agree-
ment with experimental data [146, 144, 145, 147]. We will briey review their theory
of growth from a bath of thermal atoms with xed temperature and chemical potential
as presented in QKIII [141]. This includes a simple model of condensate growth that
has subsequently been rened in QKVI [144].
Other quantum kinetic theories use various approaches. For example, Stoof has
applied path-integral methods from quantum eld theory [148], Zaremba et al. [149,
150] have developed a two-uid hydrodynamic model, Walser et al. [151] have used
non-equilibrium statistical mechanics to obtain a kinetic approach somehow similar to
the work by Gardiner, Zoller and co-workers [139, 140, 141, 142, 143, 144, 145].
5.1.1 System
One of the signatures of a BoseEinstein condensate formed in a harmonic trap is a
large density peak localized in the centre of the trap. The repulsive mean-eld of the
condensate, which depends on the condensate density, alters the low-lying energy levels
of the trap, but above a certain energy value E
R
the energy spectrum is essentially
unaected. This allows a division of the system into two separate bands [141] as
illustrated in Fig. 5.1. The so called condensate band R
C
contains all energy levels
below E
R
and includes all the levels which are signicantly modied by the presence
of the condensate. The non-condensate band R
NC
consists of all the levels above E
R
,
and these have negligible modications due to the interactions with the condensate.
In QKV it has been shown that a legitimate division into condensate band and
non-condensate band can be made in which one distinguishes between particle-like
excitations, which are occupied with a denite number of single atoms, and phonon-like
excitations, which are collective modes that usually involve a large number of atoms.
In practice, it has been shown [109] that the energy above which all excitations are
essentially particle-like is relatively small. On the other hand, in dynamical simulations
of the condensate growth process [146, 144] it is assumed that the non-condensate band
is time-independent. Thus, all energy levels whose populations are signicantly aected
during the growth process should be included in the condensate band. The precise value
50
5.1. Quantum Kinetic Theory
Figure 5.1: Condensate and non-condensate band in a harmonic trap. Figure is taken from
reference [141].
of E
R
has little eect on numerical results as long as it is reasonably high. For practical
simulations [90, 146] it was found that a value of E
R
between 2 and 3 is a good choice,
where is the nal equilibrium value for the chemical potential of the condensate, since
it is high enough that the energy perturbations of higher energy levels do not have a
large eect on the growth, and it is low enough that the majority of atoms are found
at higher energies, giving an undepleted bath [144].
5.1.2 Phase space density
For a large system the majority of atoms will be in the non-condensate band. Thus, it is
possible to describe R
NC
by a single-particle phase space distribution function f(K, x),
where K is the momentum wave vector and x the position of the particle. In QKIII
this is taken to be fully thermalized with a well dened temperature T and chemical
potential
NC
, so that it can be written as
f(K, x)
1
e
((K,x)
NC
)/k
B
T
1
, (5.1)
where (K, x) =
2
K
2
/(2m) + V
T
(x). The density operator for the non-condensate
band is given by

NC
= exp
_

NC

N
NC


H
NC
k
B
T
_
(5.2)
51
Chapter 5. Theory of Growth from Rotating Thermal Cloud
In later work [143, 144], the dynamics of the thermal cloud is included via a description
by a modied quantum Boltzmann equation. This approach takes also into account
the eective change of the potential for the thermal cloud due to the mean eld of the
condensate, and the corresponding back action of the thermal cloud on the condensate,
but we will not include these eects here.
5.1.3 Master equation
The condensate band is described fully quantum-mechanically by a master equation
for the density matrix. Starting from the standard many-body Hamiltonian (2.1), the
eld operator is split into condensate band and non-condensate band components

(x) =

(x) +

NC
(x). (5.3)
This decomposition is similar to the one used in the Bogoliubov approximation (2.11).
However, the condensate band is still described by an operator

(x) and not approx-
imated by a c-number wave function because it consists not only of the condensate
mode, but all modes whose energy levels are aected by the presence of the conden-
sate, even though they might not be macroscopically occupied. After the substitution
of the decomposition (5.3) into the Hamiltonian, the latter is separated into three parts.
The rst describes only processes within the non-condensate band R
NC
and the second
only processes within the condensate band R
C
, while the third part involves both R
NC
and R
C
and gives rise to particle and/or energy transfer between the two bands.
Assuming that the spectrum of the condensate band does not change signicantly
due to the interaction with the non-condensate band, and making a Markov and random
phase approximation, a master equation for the condensate band density operator

C
= Tr
NC
() is derived, which is given as equations (50a)-(50f) in QKIII.
5.1.4 Growth processes
A stochastic master equation for the occupation probabilities in the condensate band
p(N, n) can be obtained by considering the diagonal matrix elements of the condensate
band density operator
C
p(N, n) = N, n[
C
[N, n), (5.4)
where N is the number of atoms in the condensate band and n n
m
the set of all
quasi-particle occupation numbers. This stochastic master equation is given in QKIII
52
5.1. Quantum Kinetic Theory
as equation (189).
Simple rate equations for the mean number of atoms in the condensate can be
obtained by neglecting all uctuations. These rate equations take the form
n
m
= n
+
m
+ n

m
, (5.5)

N = 2(N + 1)W
+
(N) 2NW

(N) +

m
n
+
m
n

m
, (5.6)
where
n
+
m
2(n
m
+ 1)W
++
(N) 2n
m
W

(N), (5.7)
n

m
2(n
m
+ 1)W
+
(N) 2n
m
W
+
(N), (5.8)
and the Ws are transition probabilities explained as follows. There are six processes
described by these equations:
1. N N + 1 with no change in n in term 2(N + 1)W
+
(N).
2. N N 1 with no change in n in term 2NW

(N).
3. N N + 1 with n
m
n
m
+ 1 in term 2(n
m
+ 1)W
++
(N).
4. N N 1 with n
m
n
m
1 in term 2n
m
W

(N).
5. N N + 1 with n
m
n
m
1 in term 2n
m
W
+
(N).
6. N N 1 with n
m
n
m
+ 1 in term 2(n
m
+ 1)W
+
(N).
The Ws are the transition probabilities for these processes where the rst subscript
indicates a change in the actual particle number in the condensate band by 1, while
the second subscript (if present) indicates a change in the quasi-particle occupations
in the same manner. The exact form of these coecients is given in QKIII, equations
(120)-(128). For our purposes it is sucient to consider only the rst two, and they
are discussed in the next section. Scattering processes that do not change the number
of particles in either of the bands are not discussed here, but have been included in the
more detailed treatments of QKV and QKVI [143, 144].
5.1.5 Simple growth equation
The rates which alter the number of particles in the condensate band without changing
the quasiparticle populations are stimulated by a factor N or N+1, while all the others
53
Chapter 5. Theory of Growth from Rotating Thermal Cloud
are only of the order of the populations n
m
. Because the latter will never be nearly as
large as N, in a rst approximation we can neglect all the terms involving quasiparticles
to nd the following simple equation for the growth of the condensate mode

N = 2(N + 1)W
+
(N) 2NW

(N). (5.9)
The transition probabilities are given by
W
+
(N) =
U
2
0
(2)
5

2
_
d
3
x
_
d(,
(N)

)f
1
f
2
(1 +f
3
)(x, k), (5.10)
W

(N) =
U
2
0
(2)
5

2
_
d
3
x
_
d(,
(N)

)(1 +f
1
)(1 +f
2
)f
3
(x, k), (5.11)
where we use the notation
d d
3
K
1
d
3
K
2
d
3
K
3
d
3
k, (5.12)
(, ) (
123
(x) )(K
1
+K
2
K
3
k), (5.13)

123
(x) (K
1
, x) +(K
2
, x) (K
3
, x). (5.14)
and the Wigner function for the condensate wave function
N
for N atoms is written
as
(x, k) =
1
(2)
3
_
d
3
v

N
(x +v/2)
N
(x v/2) exp(ik v). (5.15)
Generically, the functions f
i
f(K
i
, x) are Wigner functions similar to (5.15), but
for the non-condensate part. If we assume thermodynamic equilibrium for the thermal
cloud, they reduce to the phase space density (5.1). The coecient W
+
represents
collisions of two thermal atoms with momentum K
1
and K
2
, whereby one of them
is scattered into the condensate band with momentum k, while the other remains in
the non-condensate band with momentum K
3
. Equivalently, W

describes the reverse


process. We can see that the transition probabilities W

include only momentum and


energy conserving collision processes due to the delta functions in (5.13). Also, we note
the (1+f) factors associated with the output momentum states of the scattered atoms
arise in a similar fashion to those in the collision integrals in the quantum Boltzmann
equation, indicating a Bose-enhanced process.
54
5.1. Quantum Kinetic Theory
Using the phase space density (5.1) we can evaluate
f
1
1 +f
1
f
2
1 +f
2
1 +f
3
f
3
= e
(
NC

123
(x))/k
B
T
= e
(
NC
(N))/k
B
T
,
and thus obtain the relation
W
+
(N) = e
(
NC
(N)/k
B
T
W

(N) (5.16)
between the forward and backward transition probabilities. The simple growth equation
(5.9) then takes the form

N = 2W
+
(N)
_
(1 e
((N)
NC
)/k
B
T
)N + 1
_
. (5.17)
It comes to equilibrium when the term in the curly brackets is zero, which gives the
equilibrium number of atoms in the ground state
N
eq
=
1
e
((N
eq
)
NC
)/k
B
T
1
. (5.18)
Since it is assumed that N
eq
1, this implies that (N
eq
) =
NC
to the order of 1/N
eq
.
5.1.6 Evaluation of transition probability W
+
The nal task is to evaluate the transition probability W
+
(5.10) occuring in equation
(5.9). In the simplest approximation, we note that the condensate is sharply peaked at
x = 0 compared to the phase space distribution function, and so f(K, x) is replaced
by f(K, 0). Similarly, the range of condensate momenta k is small compared to the
range of K in the phase space density. Thus, the integral over k can be carried out
independently (which gives unity). Finally, the integration of K
i
is done assuming
a classical Maxwell-Boltzmann phase space density. With these approximations W
+
takes the value
W
+
=
4m(ak
B
T)
2

3
e
2
NC
/k
B
T
_
(N)
k
B
T
K
1
_
(N)
k
B
T
__
, (5.19)
where K
1
(x) is a modied Bessel function. A more careful evaluation of W
+
using the
BoseEinstein distribution instead [144] gives a value about three times larger than that
of expression (5.19) depending on T,
NC
and the trap parameters. For our purposes,
55
Chapter 5. Theory of Growth from Rotating Thermal Cloud
we are using the simplied form
W
+
g
4m(ak
B
T)
2

3
(5.20)
where the adjustable parameter g = 3.
5.2 Phenomenological Growth Equation
In this section we will show how a simple growth as described by equation (5.17) can
be incorporated into the GPE. We rst give a simple indicative derivation to make
the physical meaning of the dierent terms clear. This is followed by a more formal
approach based on quantum kinetic theory.
5.2.1 Growth and loss in GrossPitaevskii equation
The simple growth equation (5.17) is a rate equation that describes the change in
condensate atoms over time. By adding loss and growth terms to the GPE we can
implement the same behaviour into a wave equation formulation that would describe
all the basic properties of a condensate in contact with a bath of thermal atoms phe-
nomenologically. If the system is not too far from equilibrium we can expand the
exponential in the simple growth equation (5.17) to rst order to obtain

N = 2
W
+
(N)
k
B
T
(
NC
(N))N 2(
g

l
)N, (5.21)
where we have also omitted the spontaneous +1 term. The linear growth and loss
terms determined by the coecients
g
W
+

NC
/k
B
T and
l
W
+
(N)/k
B
T can be
included in a GrossPitaevskii type of equation employing the mapping

N = 2N

= . The rate constant in the wave function formulation is only half the size
because the total number of atoms is given by N =
_
d
3
x[[
2
, and therefore

N =
_
d
3
x[


] = 2N, provided is real. Using this mapping we get two terms
Linear growth :

N = 2
W
+
k
B
T

NC
N i

= i
NC
, (5.22)
Linear loss :

N = 2
W
+
k
B
T
N i

= i, (5.23)
56
5.2. Phenomenological Growth Equation
where we have dened W
+
/k
B
T. However, there remains one problem. The
condensate chemical potential (N) is dened as the eigenvalue of the stationary GPE
for N atoms in the condensate. Since we want to develop a dynamic model of the
condensate growth, where the number of atoms is not xed, we have to allow the
chemical potential to be time-dependent. Hence, we make the generalization
i

, (5.24)
which may make local (i.e. (x)).
Adding the growth term (5.22) and the loss term (5.23) with the substitution (5.24)
to the GPE (2.14) gives
(i )
(r, t)
t
=
_

H
0
+U
0
[(r, t)[
2
+i
NC
_
(r, t). (5.25)
This equation describes phenomenologically the growth of a BoseEinstein condensate
from a bath of thermal atoms with chemical potential
NC
at temperature T. An
equilibrium state is reached when the chemical potential of the condensate becomes
equal to the chemical potential of the thermal atoms. In that case, the condensate
wave function is given by (r, t) = (r)e
it/
with =
NC
, where (r) is a solution
of the time-independent GPE for xed , and the loss and growth terms in (5.25) cancel
each other.
5.2.2 Master equation approach
In the previous section, we gave an indication of how the simple growth equation (5.17)
derived in the frame work of quantum kinetic theory could be implemented in a mean
eld approach by adding growth and loss terms to the GPE. A more formal derivation of
the phenomenological growth equation (5.25) has been presented by Gardiner et al. [85].
Here, we will only outline the main steps and approximations in their argument and
refer the interested reader to the paper itself for more details.
5.2.2.1 GrossPitaevskii equation in hydrodynamic form
Writing the wave function in the GPE in terms of condensate density n
C
and phase
(x, t) =
_
n
C
(x, t)e
i(x,t)
, (5.26)
57
Chapter 5. Theory of Growth from Rotating Thermal Cloud
the GPE takes the form
n
C
(x, t)
t
= [v
C
(x, t)n
C
(x, t)], (5.27)

(x, t)
t
=
C
(x, t), (5.28)
in which the condensate velocity v
C
and the local energy density
C
are dened by
v
C
(x, t)

m
(x, t), (5.29)

C
(x, t) (x, t) +
1
2
mv
C
(x, t)
2
, (5.30)
where is the the local chemical potential

2
_
n
C
(x, t)
_
n
C
(x, t)
+V
T
(x) +U
0
n
C
(x, t). (5.31)
This description is completely equivalent to the ordinary GPE [107].
5.2.2.2 Local energy conservation in hydrodynamic approximation
In the hydrodynamic approximation it is assumed that the condensate density is slowly
varying in space and time, so that the Laplacian in the local chemical potential (5.31)
can be dropped. Thus, the chemical potential becomes a strictly local quantity and is
a simple function of the density n
C
(x, t). On the other hand, the local energy density
turns out to be explicitly equal to the derivative with respect to n
C
(x, t) of the energy
density E(x, t) evaluated in the hydrodynamic limit
E(x, t) =
1
2
mv
C
(x, t)
2
n
C
(x, t) +V
T
(x)n
C
(x, t) +
1
2
U
0
n
C
(x, t). (5.32)
Hence, it is quite intuitive to view the addition of a particle to the condensate as a
local event at position x that adds the energy
C
(x, t) and the momentum mv
C
(x, t)
[149, 150].
5.2.2.3 Application to quantum kinetic theory
In the derivation of the master equation for the condensate band, terms occur containing

(x)

(x

, ), where

(x, t) is the condensate band operator in the interaction picture

(x, t) = e
i

H
0
t/

(x)e
i

H
0
t/
. (5.33)
58
5.2. Phenomenological Growth Equation
In QKIII the operators

(x) were expanded in terms of eigenoperators of the condensate
band Hamiltonian to account for the time evolution of

(x

, ). The current derivation


applies the idea of local energy conservation instead. Basically, it is assumed that the
condensate band is dominated by a single condensate wave function and that the main
dependence on space and time arises largely from the phase of this wave function. It
is also desirable to express the result in terms of operators at the same location, which
is taken as the midpoint u = (x +x

)/2. This leads to the approximation

(x)

(x

, )

(u)

(u)e
i[(y/2,t)+(y/2,t)]
(5.34)

(u)

(u)e
i

[mv
C
(u,t)y
C
(u,t)]
, (5.35)
where y = xx

. The approximation (5.35) requires that only small y and contribute.


This is the case if the non-condensate band is fully thermalized because the expression

(x)

(x

, ) only occurs in convolution with a similar term containing the non-


condensate operator in third order

NC

NC

NC
.
All irreversible terms can then be expressed in terms of the condensate band oper-
ator

(x), and the master equation takes the form

C
=
i

_
dx
_

(x)
_

2
2m
V
T
(x) 2u
NC
(x, t)
1
2
u

(x)

(x)
_

(x),
C
_
+
_
dx
_
G
(+)
[x,
C
(x, t)](2

(x)
C

(x)
C

(x)

(x)

(x)

(x)
C
)
+G
()
[x,
C
(x, t)](2

(x)
C

(x)
C

(x)

(x)

(x)

(x)
C
)
_
, (5.36)
where

NC
(x)
_
dKF(K, x).
Here, F(K, x) is the phase space density of the non-condensate band. The transition
rates G

are dened as
G
(+)
[x, ] =
U
2
0
(2)
5

2
_
d

f
1
f
2
(1 +f
3
)
= (
123
(x)

)(K
123

mv
C
(x, t)

), (5.37)
G
()
[x, ] =
U
2
0
(2)
5

2
_
d

(1 +f
1
)(1 +f
2
)f
3
= (
123
(x)

)(K
123

mv
C
(x, t)

), (5.38)
59
Chapter 5. Theory of Growth from Rotating Thermal Cloud
where
d

d
3
K
1
d
3
K
2
d
3
K
3
, (5.39)

123

K
1
(x) +
K
2
(x)
K
3
(x), (5.40)
K
123
K
1
+K
2
K
3
. (5.41)
The form of the transition probabilities (5.37) and (5.38) is the same as that in
(5.10) and (5.11), whereby the relevant frequency and wavevector are
C
(x, t)/ and
mv
C
(x, t)/. In the local energy approximation, the transition probabilities represent-
ing the growth processes 3 to 6 from section 5.1.4 are all identically zero.
5.2.2.4 Phenomenological mean value equations
From the master equation (5.36), phenomenological equations are obtained considering
the evolution of the mean values

(x, t) Tr

(x, t)
C
(t), (5.42)
n
C
(x, t) Tr

(x, t)

(x, t)
C
(t), (5.43)

C
(x, t) Tr
_
i
2m
_

(x, t)

(x, t) [

(x, t)]

(x, t)
_

C
(t)
_
. (5.44)
This leads to a modied GPE and a local growth equation

(x, t)
t
=
i

_

2
2m

(x, t) V
T

(x, t) 2U
0

NC

(x, t) U
0

(x)

(x)
2
)
_
+G
+
[x,
C
(x, t)]
_
1 e
(
C

NC
)/k
B
T
_

(x, t), (5.45)


n
C
(x, t)
t
= (x, t) + 2G
+
[x,
C
(x, t)]
__
1 e
(
C

NC
)/k
B
T
_
n
C
(x, t) + 1
_
.(5.46)
Assuming that the temporal derivative of n
C
(x, t) can be neglected, we can approximate
(5.28) using (5.29)

C
(x, t) =
(x, t)
t

i

(x, t)

(x, t)
t
. (5.47)
Finally, we factorize all averages of products of

operators in equation (5.45) and
expand the exponential to rst order to retrieve the phenomenological growth equation
60
5.2. Phenomenological Growth Equation
as presented in equation (5.25)

(x, t)
t
=
i

_

2
2m

2
V
T
2U
0

NC
U
0
[

(x, t)[
2
_

(x, t)
+
W
+
k
B
T
_

NC

(x, t) i

(x, t)
t
_
(5.48)
with the replacement G
+
W
+
.
5.2.3 Rotating thermal cloud
In the derivation so far, only the case of growth of a condensate from a stationary
thermal cloud has been considered. However, our main interest is concerned with the
processes involved in the growth of a vortex lattice. Although in most experimental
realizations of vortex lattices a condensate is stirred by an anisotropic rotating trap
[73, 76, 77, 79, 152], the JILA group has demonstrated [78] that the formation of a
vortex lattice can also be achieved in a cylindrically symmetric harmonic trap if the
thermal cloud has been set into rotation beforehand, and is then evaporatively cooled
below the BoseEinstein condensation transition temperature.
This latter scenario can be considered by assuming that the thermal bath is a cloud
rotating at angular velocity in a cylindrically symmetric trap. It is clear that an
equilibrium state can only be achieved in the frame rotating at angular velocity ,
since only in that frame is the Hamiltonian of the whole system time-independent.
Hence, we make the transformation

R
= e
iLt/
(5.49)
in (5.25) to obtain the phenomenological growth equation in the rotating frame
i

R
t
=
_

2
2m

2
+V
T
+U
0
[
R
[
2
L
_

R
+i
W
+
k
B
T
_

NC
i

t
_

R
. (5.50)
However, the transition probability W
+
has to be evaluated in the rotating frame, too.
The phase space density for a rotating thermal cloud is given by
F(x, p) =
1
e
((x,p)
NC
xp)/k
B
T
1
, (5.51)
61
Chapter 5. Theory of Growth from Rotating Thermal Cloud
where the energy function is
(x, p) =
p
2
2m
+V
T
(x). (5.52)
The parameter determines the mean value of the angular momentum in the same
way that the chemical potential
NC
determines the mean number of atoms. We can
rewrite (5.52) using the substitution
p
2
2m
x p =
(p mx)
2
2m

m(x)
2
2
(5.53)
as
(x

, p

) =
p

2
2m
+V
e
(x

), (5.54)
where p

p m x is the momentum in the rotating frame. Thus, the energy


function appears like an ordinary kinetic energy in an eective potential modied by a
negative centrifugal term V
e
(x) V
T
(x) m( x)
2
/2. In this frame, the relation
between the transition probabilities W
+
and W

is still the same as given in equation


(5.16). Also, the result for W
+
does not change if we make the same approximations as
outlined in section 5.1.6. We therefore take the value (5.19) for W
+
with the correction
factor g = 3.
If the trapping potential is anisotropic and rotating at angular velocity , we can
distinguish two cases. For the case where = , the trapping potential is time-
independent in the rotating frame, and the argument carries through unaltered to
obtain equation (5.50). However, if ,= the thermal cloud cannot truly be in
thermodynamic equilibrium. In the frame rotating at the clouds angular velocity, we
chose the approximate form V
e
(x)

V
T
(x) m(x)
2
/2 for the eective potential,
where

V
T
(x) is the time averaged trap potential over the dierential rotation between
the trap and the cloud. With this approximation, equation (5.50) is still valid in a
frame rotating at the clouds angular velocity , but the trapping potential would be
time-dependent now. More common is a transformation into a frame rotating at the
traps angular velocity , where the trapping potential is time-independent. Then, the
phenomenological growth equation (5.50) takes the form
i

R
t
=
_

2
2m

2
+

V
T
+U
0
[
R
[
2
L
_

R
+i
W
+
k
B
T
_

NC
+ () L i

t
_

R
. (5.55)
62
5.3. Comparison with Other Theories
5.2.4 Stationary solution
As mentioned true equilibrium is only possible if = . In that case equation (5.55) is
identical to (5.50). A vortex lattice is a stationary solution of the GPE in the rotating
frame
i

R
t
=
_

2
2m

2
+V
T
+U
0
[
R
[
2
L
_

R
(5.56)
satisfying
i

R
t
=
R
, (5.57)
where is the chemical potential of the condensate. Substituting (5.57) in equation
(5.50) shows that the phenomenological growth equation is identical with the GPE in
the rotating frame (5.56) if =
NC
. That means that the phenomenological growth
equation drives the condensate state to a stationary state in the rotating frame if
= , which is a lattice state if the rotation rate is above some critical frequency.
5.3 Comparison with Other Theories
One method to nd the solutions of the GPE numerically is to propagate the time-
dependent GPE with an added chemical potential in imaginary time while constantly
renormalizing the wave function. This method nds a state with an energy minimum.
Tsubota et al. [153] proposed a modication of this method as a phenomenological
model to describe the stabilization of vortices into a lattice. They used the equation
(i )

t
=
_

2
2m

2
+V
T
+U
0
[[
2
L
z
_
, (5.58)
which is in appearance very similar to our phenomenological growth equation (5.50) if
we set W
+
/k
B
T. However, there are some signicant dierences [85]. Tsubota
et al. gave no physical justication for this equation. While their results might be
qualitatively attractive, they provide no real physical understanding of the process. In
contrast, our treatment explains the source of the damping parameter from the ex-
change of atoms between the condensate and the thermal cloud and gives an expression
for in terms of the transition probability W
+
from quantum kinetic theory.
Secondly, in equation (5.58) appears the real term on the RHS while the
phenomenological growth equation has an imaginary term i
NC
. Tsubota et al.
adjusted with time to preserve the number of atoms in the condensate. Our
NC
is the
physical chemical potential of the thermal cloud. The driving force for the formation of
63
Chapter 5. Theory of Growth from Rotating Thermal Cloud
a regular lattice arises from the spatially dependent local chemical potential of a random
lattice of vortices. In equilibrium the local chemical potential of the condensate must
balance the chemical potential of the thermal cloud. Thus, the state is driven towards a
state with uniform chemical potential accross the whole condensate, which is a regular
equilibrium lattice. This irreversible process is included in the phenomenological growth
equation.
A formalism conceptually very similar to our formulation has been developed by
Zaremba et al. [150, 154]. Their formulation of a quantum kinetic theory is based on a
two-uid picture of the condensate and the thermal cloud. The condensate is described
by two equations in hydrodynamic form
n
c
t
+ (n
c
v
c
) =
12
[f], (5.59)
m
_

t
+v
c

_
v
c
=
c
, (5.60)
where n
c
, v
c
and
c
are the condensate density, velocity and local chemical potential.
The non-condensate is described by a distribution function f(r, p, t) that is governed by
a quantum Boltzmann equation. The collisional term
12
[f] describes collisions between
one condensate atom and a thermal atom that do not conserve the number of atoms
in the condensate. This corresponds to the same collisions included in the transition
probablity W
+
in our description. The numerical implementation of this description
is quite sophisticated [155] because separate methods are required to propagate the
condensate wave function and the distribution function, and additionally Monte-Carlo
methods to treat the collisional term. The problem of vortex lattice formation has so
far been intractable with these methods because it showed numerical instabilities [156].
Lobo et al. [157] used a classical eld approach to simulate the formation of a vor-
tex lattice using the GPE without any explicit damping term. The requisite damping
is provided solely by the noise arising from the initial conditions. Based on a Wigner
representation it can be shown that the GPE includes the evolution of the thermal
modes, not only the condensate [158, 159, 160]. However, in this approach it is impor-
tant to carefully implement an appropriate energy cut-o for the modes included in the
calculations [161, 162]. In the simulations by Lobo et al. [157] the energy cut-o was
provided by the neness of the numerical grid, which makes the exact interpretation of
the boundary condition dicult. Very recently, Gardiner and Davis [163] have unied
the ideas of quantum kinetic theory [139, 141, 143] and the nite temperature GPE
[164], and some preliminary simulations [165] indicate a promising application of this
64
5.3. Comparison with Other Theories
new treatment for simulations on vortex lattice formation at nite temperatures.
65
Chapter 6
Numerical Propagation Method
In this chapter, we introduce the numerical method to propagate the phenomenological
growth equation, and thoroughly test its accuracy. In particular we determine the
optimal choice of grid size, number of grid points and temporal step size.
In our dimensionless units, the phenomenological growth equation in the frame
rotating with the trap at angular velocity reads
(i )

R
(r, t)
t
=
_

2
+V
T
+C[(r, t)[
2

L
z
_

R
(r, t)
+i
_

NC
+ ( )

L
z
_

R
(r, t), (6.1)
where we have arbitrarily chosen the z-axis as axis of rotation. There are four pa-
rameters: the damping parameter , the nonlinearity C, the chemical potential of
the thermal cloud
NC
and the angular velocity of the thermal cloud . The angular
momentum operator is given by

L
z
= i
_
x

y
y

x
_
. (6.2)
Most of our simulations are carried out in the stationary laboratory frame, where the
equation takes the form
(i )
(r, t)
t
=
_

2
+V
T
+C[(r, t)[
2
+i
_

NC
+

L
z
__
(r, t). (6.3)
67
Chapter 6. Numerical Propagation Method
6.1 Algorithm
To propagate the phenomenological growth equation (6.3) with a rotating thermal
cloud we used an algorithm developed in our Otago group by Ballagh and Fox called
RK4IP (Runge-Kutta Fourth-Order Interaction Picture) [166, 167]. It is similar to the
symmetrized split-step method used in optics to propagate the non-linear Schrodinger
equation. The actual code we used is an extension of Caradoc-Davies work [167] who
implemented a highly ecient version of the algorithm using a combination of Python
and C to minimize memory usage and bandwidth. This enabled the application of
the method to fully three dimensional problems. A detailed review of the algorithm is
presented in Appendix B.
Most of our simulations were performed on a machine with an AMD Athlon 1.8
GHz processor and 512 MB of DDR-RAM. On this machine, it takes about 131 s to
propagate a wave function on a 512 512 point grid for 100 steps. On a 1024 1024
grid, 100 steps take about 564 s. That means the algorithm scales roughly linearly with
number of grid points.
6.2 Accuracy
The numerical accuracy of the standard RK4IP algorithm applied to the pure Gross
Pitaevskii equation (GPE) has been discussed by Caradoc-Davies in-depth [167]. How-
ever, most simulations previously performed in our group ran only for a few trap periods
up to about 16. Our intention was to run comparatively long simulations for up to 100
trap periods to explore vortex lattice formation and also fully three dimensional sim-
ulations that could only be performed with a very small number of points in each
direction due to memory restrictions. In either case, a minimal amount of grid points
is desirable to carry out such simulations eciently. Our main aim was therefore to
nd the minimal requirements on the grid size and the accuracy of the algorithm over
long time periods.
As far as the numerical implementation is concerned, the main dierence of the
phenomenological growth equation (6.3) to the GPE is the inclusion of the angular
momentum operator. Thus, we also need to determine the accuracy of the numerical
representation of the angular momentum operator, and its eect on the dynamical
propagation using the phenomenological growth equation.
68
6.2. Accuracy
6.2.1 General points of consideration for numerical simula-
tions
Here, we summarize the main points of Caradoc-Davies ndings [167] about the accu-
racy of the RK4IP algorithm. There are three important parameters associated with
our time-dependent simulations on a discretized spatial grid: the temporal step size,
the spatial extent of the grid and the number of grid points in each direction.
Since the temporal evolution of an eigenstate with eigenvalue is a pure phase
rotation exp(it), a good guide for the temporal step size t to assure an accurate
representation of the phase evolution is given by the condition
t
1
TF
, (6.4)
where
TF
is the ThomasFermi estimate of the chemical potential as given in section
2.4.
The spatial grid has to be big enough to accomodate the condensate wave function.
Hence, it is required that
R
x,y
2

TF
, (6.5)
where R
x,y
is the radial spatial extent of the grid in x- and y-direction.
On the other hand, the grid spacing has to be small enough to represent sharp
variations in the density eld, in particular vortices. This is equivalent to the require-
ment that the Nyquist frequency of the momentum representation is sucently large
to include all necessary high frequency components. As an estimate of the shortest
wave length necessary to represent the core of a vortex, we can use the healing length
that is approximately given by =
1/2
TF
at the centre of the condensate. Thus, the
requirement on the spatial grid spacing reads
x =
1/2
TF
. (6.6)
Since the GPE conserves the number of atoms, the normalization of the wave func-
tion should remain constant. The main result of Caradoc-Davies exploration is the fact
that the normalization of the wave function is a very good indication of the over-all
accuracy of the simulation as long as the Nyquist limit is not violated. However, a
violation of the Nyquist limit can occur without obvious symptoms. The best way
we have found to detect the occurence of aliasing is to carefully monitor the density
distribution of the wave function at the edges of the grid in position and momentum
69
Chapter 6. Numerical Propagation Method
space. A signicant increase at the edges of the grid in either space is most likely
caused by aliasing since any part of the wave function that reaches the edge of the
grid is reected back into it. In the next section, we will explore how the choice of the
discrete numerical grid aects the accuracy and validity of the simulation.
6.2.2 Choice of grid size and number of points
In this section we present our results from simulations we have carried out to verify the
validity of the RK4IP on dierent spatial grids. For our dynamical simulations, the grid
size is typically the same in x- and y-direction with the same number of equi-distant
points symmetrically spaced around the origin. Choosing l
x
to be the total extent of
the grid in the x-direction, then the grid covers the range [l
x
/2, l
x
/2] and the spatial
grid spacing is x = l
x
/(n
x
1) where n
x
is the number of points
1
.
The wave function is transformed into momentum space via a discrete Fourier trans-
formation giving the Nyquist limit
k
c
= 2
1
2x
, (6.7)
which is the maximum wave vector in the x-direction that is still adequately represented
by the discrete spatial grid. Using this expression we can rewrite condition (6.6) as
k
c
k
heal

. (6.8)
The grid spacing in momentum space in the x-direction is given by
k
x
= 2
1
n
x
x
= 2
n
x
1
n
x
l
x
. (6.9)
It is clear that to have a large spatial grid, but to represent also high momentum
components, requires a large number of grid points.
As a test case, we used the single vortex cycling regime of a condensate stirred by
a small laser beam discussed by Caradoc-Davies in [168]. The stirrer is given by a
Gaussian potential of the form
W(x, t) = W
s
(t) exp
_

_
[x x
s
(t)[
w
s
/2
__
(6.10)
with width w
s
= 1 and amplitude W
s
= 10 rotating around the origin at the xed radius
1
Since the same results are obtained in the y-direction, we discuss here only the x-direction.
70
6.2. Accuracy
r
s
= [x
s
(t)[ = 3 with angular frequency
f
= 0.4. The stirrer is ramped on linearly
over a time period of t = , and the non-linearity coecient is C = 200 which gives
a ThomasFermi radius of R
TF
= 4.75 and chemical potential
TF
= 5.64. According
to condition (6.8), the grid in momentum space should include wave vectors greater
than k
heal
= / =

TF
=7.46. As a benchmark we carried out one simulation on a
large grid with l
x
= 60 and n
x
= 1024 with a time step of 500/. This has a Nyquist
limit of k
c
= 53.56 and saties the conditions (6.4), (6.5) and (6.8) assuring an accurate
simulation. For brevity we will refer to this simulation in the following as the reference
simulation.
While the extent of the grid is large compared to the ThomasFermi radius, in
a logarithmic plot (Fig. 6.1 (a))
2
the wave function appears quite wide approaching
zero exponentially. The same holds for the wave function in momentum space as can
30 20 10 0 10 20 30
10
20
10
15
10
10
10
5
10
0
Position x
|

(
x
,
y
=
0
)
|
(a)
t = 3
t = 50
t = 100
40 20 0 20 40
10
20
10
15
10
10
10
5
10
0
Wave vector k
x
|

(
k
x
,
k
y
=
0
)
|
(b)
t = 3
t = 50
t = 100
Figure 6.1: Cross sections of the wave function obtained in a pure GPE simulation of single
vortex cycling in a C = 200 condensate stirred by a small Gaussian perturber as
described in the text. (a) Position space wave function at y = 0, (b) momentum
space wave function at k
y
= 0 for times t = 3, 50 and 100.
be seen in Fig. 6.1 (b). Over the course of the simulation the wave function spreads
even further to the edges in position space, while the width in momentum space stays
the same after a small initial spread due to the insertion of the stirrer (not shown in
Fig. 6.1). In momentum space the limit of numerical noise is around 10
10
. If we take
this value also as zero point in position space, which seems reasonable considering the
wave function at t = 100, we see from Fig. 6.1 that the minimum radius of the grid in
2
The plot shown is the wave function after propagation for t = 3 because the initial wave function
was obtained using the optimization method on a smaller grid, and then padded with zeros on the
edges. That means the initial state would not show the same kind of exponential decay to zero at the
edges.
71
Chapter 6. Numerical Propagation Method
position space should be around R
min
= 20 and in momentum space around k
min
= 8
to avoid aliasing eects. This corresponds to R
min
= 4.2R
TF
and k
min
= 1.1k
heal
. In
other words the spatial extent has to be much bigger than the ThomasFermi radius to
accomodate the wave function while k
heal
is indeed a very good estimate of the width
of the wave function in momentum space. These observations lead to the conclusion
that a large grid in position space is more important than in momentum space.
Fig. 6.2 shows the angular momentum of this simulation run over a long period
of time of 100 (50 trap periods). The normalization in this simulation decreases by
0 20 40 60 80 100
0
0.2
0.4
0.6
0.8
1
1.2
Time []
A
n
g
u
l
a
r

m
o
m
e
n
t
u
m
Figure 6.2: Angular momentum in single
vortex cycling simulation.
10
4
10
3
10
2
10
1
Position x
P
o
s
i
t
i
o
n

y
5 0 5
5
0
5
Figure 6.3: Density plot of nal state at
t = 100 with vortex trapped
in stirring potential.
1.30 10
6
which is acceptable over a time period of 100. We can see that the single
vortex cycling reported by Caradoc-Davies [168] attenuates gradually after 60 until
the vortex is trapped in the potential of the stirring beam as shown in Fig. 6.3, which
is a density plot of the nal state.
To make accurate comparisons to the reference simulation we made sure that the
spatial points in the smaller, less dense grids matched with points on the reference grid,
so that we could avoid interpolation. This was achieved by setting x = m
s
x
ref
,
where the spacing m
s
is an odd integer to insure a symmetric grid around the origin,
and x
ref
= 60/1023 is the spatial grid spacing of the reference simulation. With the
number of grid points n
x
, the extent of the new grid is given by l
x
= m
s
x
ref
(n
x
1).
To compare the new simulation to the reference solution we calculate the variance in
72
6.2. Accuracy
the wave function weighted by the density
=

ij
[
ij
[
2
[
ij
(
ref
)
ij
[
2
n
2
x
1
. (6.11)
Fig. 6.4 shows that the simulation with the most points on the smaller grid (n
x
= 256
and m
s
= 3) is very accurate compared to the reference simulation to about t = 60.
Even though the accuracy becomes increasingly worse after that point, it is still in
reasonable agreement with the reference simulation, and the nal state shows the same
density deformations as shown in Fig. 6.3. The other two simulations on much smaller
0 20 40 60 80 100
10
10
10
5
10
0
Time []


w
e
i
g
h
t
e
d

b
y

|

|
2
(a)
0 20 40 60 80 100
10
10
10
5
10
0
Time []


w
e
i
g
h
t
e
d

b
y

|

|
2
(b)
Figure 6.4: Comparison of three simulations on dierent sized grids to reference simulation
in (a) position space and (b) momentum space. Solid line n
x
= 256, m
s
= 3
with l
x
= 44.9 and k
c
= 17.8; dashed line n
x
= 128, m
s
= 5 with l
x
= 37.2 and
k
c
= 10.7; dotted line n
x
= 64, m
s
= 7 with l
x
= 25.9 and k
c
= 7.65.
grids with n
x
= 128 and m
s
= 5 or even n
x
= 64 and m
s
= 7 are reasonably good
within the single vortex cycling regime. The former has a spatial extent of l
x
/2 = 18.62
and k
c
= 10.71 and thus roughly fulls the minimum requirement we identied for the
reference simulation of R
min
= 20 and k
min
= 8. The latter, however, with l
x
/2 = 12.93
and k
c
= 7.65, hardly meets this requirement, but still produces qualitatively the same
behaviour in the single vortex cycling regime. While the variance in the wave function
increases with smaller grids, as can be seen from Fig. 6.4, movies of the evolution
of the density show that the main physical eect, i.e. the single vortex cycling, is
well captured with the same shape deformations of the condensate occuring in the
reference simulation. All simulations also show the eventual capture of the vortex in
the trapping potential, but the exact shape deformations of the condensate after the
end of the single vortex cycling show some dierences for the two simulations on the
73
Chapter 6. Numerical Propagation Method
smaller grids compared to the reference simulation. Mainly, the cloud-like contour line
at the edge of the high density region shows dierent proles, and the exact positions
of nearby vortices vary. Apparently, the propagation after the vortex cycling ceases
is computationally less stable due to the almost turbulent density deformations of the
condensate.
Interestingly, we found examples where the simulations were running ne for a small
number of grid points, but crashed when the number of points was increased. This was
the case for l
x
= 40 and n
x
= 128, which was in good agreement with the reference
simulation, and n
x
= 512 with the same spatial extent that failed after around 30. A
simple lter in position space could suppress the failure of the simulation. Following
Kr uger [169], we multiplied the wave function after every time step with a lter that
rises like sin
1/8
, and falls o like cos
1/8
over 15% of the total grid size at all sides of the
grid and is unity everywhere else.
With such a lter in place, the simulation on the 512 point grid runs smoothly.
This indicates that edge eects in position space are responsible for the failure of the
simulation. In Fig. 6.5 (b) we can see that gradually peaks in the momentum space
wave function occur at high momenta. We have checked that they correspond to the
ripples seen at the edge of the position space plot in Fig. 6.5 (a). Presumably, a small
fraction of the condensate wave function reaches the grid border where it is reected
back, and interference leads to the high momentum ripples. As these high momentum
peaks grow rapidly, the wave function in position space grows by orders of magnitude
at the edge of the grid, and the simulation eventually fails. There are also second order
20 10 0 10 20
10
20
10
15
10
10
10
5
10
0
Position x
|

|

a
t

y

=

0
(a)
40 20 0 20 40
10
20
10
15
10
10
10
5
10
0
Wave vector k
x
|

|

a
t

y

=

0
(b)
t = 10
t = 20
t = 30
t = 10
t = 20
t = 30
Figure 6.5: Instability in l
x
= 40, n
x
= 512 simulation. Shown are cross sections of the wave
function at y = 0 in (a) position and (b) momentum space for times t = 10, 20
and 30.
74
6.2. Accuracy
peaks occuring in momentum space at [k[ = 17.6 for t = 30. They actually correspond
to the rst order peaks frequency doubled to [k[ = 62.5. But due to aliasing they are
folded back into momentum space to [k[ = 2k
c
62.7 = 17.6, where k
c
= 40.13.
For the smaller grid, Fig. 6.6 (b) reveals that those high momenta are not resolved
since the Nyquist limit is k
c
= 9.97, and thus do not cause any problems. Even though
20 10 0 10 20
10
20
10
15
10
10
10
5
10
0
Position x
|

|

a
t

y

=

0
(a)
5 0 5
10
20
10
15
10
10
10
5
10
0
Wave vector k
x
|

|

a
t

y

=

0
(b)
t = 10
t = 20
t = 30
t = 10
t = 20
t = 30
Figure 6.6: Cross sections of the wave function at y = 0 in (a) position and (b) momentum
space for l
x
= 40, n
x
= 128 simulation for times t = 10, 20 and 30.
the wave function is much larger at the edge of the grid in momentum space, the
important point is that in position space there is no cumulative build-up at the edge
of the grid, i.e. the wave function does not exceed 10
10
.
We conclude that it is important to nd an appropriate combination of spatial extent
and momentum extent. While both of them have to be reasonably large to include the
whole wave function, too large a grid in momentum space should be avoided. For a
xed spatial extent this is easily achieved by reducing the number of grid points, which
has also the advantage that it is computationally more ecient. In general, it is more
important to have a large extent in position space than in momentum space. As the
rule of thumb, the spatial extent should be at least two or preferably three times as
large as the ThomasFermi radius unless a spatial lter is used. The minimum number
of points can then be chosen in such a way that k
c
is just greater than k
heal
. Even
on small grids that hardly full the conditions (6.5) and (6.8), simulations are possible
that capture the essential physical behaviour although the absolute value of the wave
function might deviate from a more exact solution by the order of 10
4
. It seems best
if the wave function has about the same value at the edge of the grid in either space
to avoid aliasing eects that might cause the simulation to fail completely.
75
Chapter 6. Numerical Propagation Method
6.2.3 Temporal step size
The standard temporal step size we use for our simulations is /500 following Caradoc-
Davies example [167]. One might think that a smaller step size should in general
improve the accuracy of a simulation. But if we run the reference simulation with half
that step size, namely /1000, the result becomes actually worse as shown in Fig. 6.7
(a). It deviates signicantly from the reference simulation, especially for times greater
0 20 40 60 80 100
10
10
10
8
10
6
10
4
10
2
10
0
Time []


w
e
i
g
h
t
e
d

b
y

|

|
2
(a)
0 20 40 60 80 100
10
10
10
8
10
6
10
4
10
2
10
0
Time []


w
e
i
g
h
t
e
d

b
y

|

|
2
(b)
Figure 6.7: Comparison of two simulations with half the step size of t = /1000 to refer-
ence simulation (dotted lines). (a) On the same grid as the reference simulation
with l
x
= 60 and n
x
= 1024. (b) On the grid given by n
x
= 256, m
s
= 3
with l
x
= 44.9 and k
c
= 17.8. The solid line is reproduced from Fig. 6.4 (a)
for the sake of comparison showing the result on the same grid with step size
t = /500.
than t = 60. But how can we be sure that the reference simulation is indeed more
accurate, and not this one with half the step size? There are two reasons. First,
the density plots show increased noise in the low density region coupled with strange
interference eects not seen in the reference simulation. And secondly, we observe the
same kind of build-up in the wave function with the corresponding high momentum
peaks as discussed in connection with Fig. 6.5.
If we halve the step size on the smaller grid with n
x
= 256, m
s
= 3 the deviation from
the reference simulation also increases considerably compared to the same simulation
with /500 as seen in Fig. 6.4 (b). It is even worse than that on the smaller grid with
n
x
= 128, m
s
= 5 and a step size of /500 as can be seen by comparison with Fig. 6.4
(a). On the other hand, doubling the step size results in a complete failure of the
simulation on either grid. After only a few steps the wave function contains NaN (not
a number), and no further propagation is possible.
76
6.2. Accuracy
In summary, too big a step size is easily detected since the simulation fails com-
pletely. However, if the step size is too small it can also cause numerical instabilities
with much more subtle eects, a well known fact from Crank-Nicholson propagation
schemes. Just as there is a relation between grid size and number of points to achieve
reliable results, there is a similar relation between the temporal step size and the num-
ber of points. For our algorithm, Simula [170] found that in general a smaller step size
is required for grids with more points but the same spatial extent. Our results show
that a step size of /500 suits most of our simulations.
6.2.4 Accuracy of angular momentum operator in GrossPita-
evskii equation propagation
To check the eect of the angular momentum operator in a dynamical simulation, we
consider an eigenstate of a cylindrically symmetric trap with at least one vortex. We
propagate the state using the GPE in the rotating frame
i

R
(r, t)
t
=
_

2
+V
T
+C[
R
(r, t)[
2

L
z
_

R
(r, t), (6.12)
where is the angular frequency of the rotation. Because the trap is cylindrically
symmetric and thus time-independent in the rotating frame, the state is also an eigen-
state in the rotating frame. However, because the state carries angular momentum due
to the vortex, we will get some contribution from the

L
z
term in equation (6.12).
Thus, we can see directly what eect the angular momentum operator has on the over-
all accuracy. As a test case, we chose the rst excited harmonic oscillator state with a
phase circulation of 2 given by

n=1,l=1
(, ) =
1
2

2
/4
e
i
. (6.13)
This is an eigenstate of the GPE for C = 0 with the eigenvalue = 2 and angular
momentum l = 1.
If we propagate the state in the laboratory frame with a temporal step size of /500
using the pure GPE on a grid with a spatial extent of l
x
= l
y
= 30 and 128 points
in each direction, the normalization and eigenvalue of the wave function are very well
conserved as shown in Fig. 6.8 (a) and (b). However, if the same state is propagated in
a frame rotating at = 0.8, the normalization and angular momentum show periodic
oscillations with an amplitude of the order of 10
3
as shown in Fig. 6.8 (c) and (d).
77
Chapter 6. Numerical Propagation Method
0 5 10 15 20
1
0
1
x 10
8
Laboratory Frame
Time []
N
o
r
m

1
(a)
0 5 10 15 20
0
2
4
6
8
x 10
10
Time []
(b)
2
<L
z
>1
0 5 10 15 20
0
0.5
1
1.5
2
2.5
3
x 10
3
Rotating Frame
Time []
N
o
r
m

1
(c)
0 5 10 15 20
0
1
2
3
4
5
x 10
4
Time []
(d)
1.2
<L
z
>1
Figure 6.8: Accuracy of conserved quantities in propagation of harmonic oscillator eigen-
state
1,1
(, ) using the pure GPE equation. The left hand column are results
from a laboratory frame, and the right hand column are results from a frame
rotating at = 0.8.
The eigenvalue also displays oscillations with an amplitude of the order of 10
4
and
the same dominant frequency with some additional frequency modulation. A Fourier
transformation of the time evolution of (Fig. 6.9) reveals that the two main frequency
components are
1
= 0.2 and
2
= 1.8, which are
1,2
= 1 . This observation has
been conrmed with the same kind of simulations in the rotating frame for = 0.5 and
= 0.3. In section 10.3.1, we will see that these are exactly the excitation frequencies
of the dipole mode in the rotating frame. Apparently, the propagation method leads to
a considerable excitation of the dipole modes when propagated in the rotating frame.
This must be caused by some intrinsic asymmetry in the method. What comes to mind
is the fact that in position space the grid is exactly centred around the origin. In the
78
6.2. Accuracy
0 0.5 1 1.5 2 2.5
10
12
10
10
10
8
Frequency component
P
o
w
e
r

S
p
e
c
t
r
u
m

o
f

Figure 6.9: Frequency components of the temporal oscillation of eigenvalue in rotating


frame propagation from Fig. 6.8.
momentum space, however, the discrete Fourier transformation leads to an asymmetric
grid because the Nyquist frequency occurs only once. As expected the error in the norm
and the eigenvalue decreases with decreasing , and is about an order of magnitude
smaller for = 0.3 than for = 0.8.
Table 6.1 summarizes several combinations of grid size and step size, we tested.
Using the same number of grid points, but half the step size gave exactly the same
Grid size Step size Result
128 points /500 Oscillation in norm 3 10
3
128 points /1000 Oscillation in norm 3 10
3
Simulation fails after t = 16
512 points /500 Simulation fails completely
512 points /1000 Oscillation in norm 1.5 10
2
Simulation fails after t = 19
1024 points /1000 Simulation fails completely
Table 6.1: Test results for dierent combinations of number of grid points and step size.
error. However, the simulation became numerically unstable after a time period of
about 16. Increasing the number of spatial points to 512 512 and using a step size
of /1000 actually worsened the error in the norm and eigenvalue. For 1024 points
in each direction, the simulation failed even with a step size of /1000 after only a
couple of steps. This conrms the observation from section 6.2.3 that there are certain
79
Chapter 6. Numerical Propagation Method
combinations of grid size and step size which work best while an increase in one or the
other does not automatically give more accurate results.
While this error is still small it is much larger than one would expect from a purely
numerical error in the calculation of the derivatives in the

L
z
term. The regular
periodicity suggests a peculiar systematic problem of the propagation in the rotating
frame, which we do not fully understand. We also found the simulations in the rotating
frame to be much more unstable than in the laboratory frame. This observation was
conrmed by Bradley [165] who notes that at least half the step size is required to
run a simulation in the rotating frame with the same accuracy than in the laboratory
frame. Because of these diculties, all simulations of physical results presented in the
remaining chapters of this thesis are carried out in the laboratory frame.
We also note that the contribution from the angular momentum term in the GPE
propagation of
1,1
is uncharacteristically high compared to the kinetic energy term.
When using the phenomenological growth equation (6.3), the angular momentum term
appears in quadrature to the kinetic energy and is also attenuated by the factor
which we typically choose to be small ( 0.1). Its relative importance is further
diminished due to the non-linear term. As we will show in the next section, the self-
stabilizing eect of the phenomenological growth equation places much less demand
on the accuracy of the angular momentum operator than the test simulations in the
rotating frame presented in this section.
6.2.5 Accuracy of phenomelogical growth equation
In this section we present similar tests, but using the phenomenological growth equation
(6.3) in the laboratory frame. One main dierence to the GPE is the fact that the phe-
nomenological growth equation does not conserve the norm, but allows atom exchange
between the thermal cloud and the condensate. Thus, the normalization of the wave
function cannot be used as a diagnostic tool to determine the numerical accuracy of
the simulation. Nevertheless, because we expect the phenomenological growth equation
to drive any wave function to an eigenstate of the trap with eigenvalue =
NC
(for
= 0), the eigenstate (6.13) should not change if we set
NC
= 2 with the additional
quantities = 0 and = 0.1. The results in Fig. 6.10 (a) and (b) show that the
normalization, eigenvalue and angular momentum are constant to a high degree.
If ,= 0, we have to be a little bit more careful. As we will see in section 7.4,
the phenomenological growth equation drives the state into an eigenstate of the frame
rotating at with
R
=
NC
, where
R
is the eigenvalue of the condensate in the
80
6.2. Accuracy
0 5 10 15 20
4
2
0
2
4
x 10
9
= 0
Time []
N
o
r
m

1
(a)
0 5 10 15 20
0
2
4
6
8
x 10
10
Time []
(b)
2
<L>1
0 5 10 15 20 25
0.005
0
0.005
0.01
0.015
0.02
0.025
0.03
= 0.8
Time []
N
o
r
m

1
(c)
0 5 10 15 20
10
5
0
5
x 10
7
Time []
(d)
1.2
<L>1
Figure 6.10: Accuracy of conserved quantities in propagation of harmonic oscillator eigen-
state
1,1
(, ) using the phenomenological growth equation. For the left hand
column the thermal clouds angular velocity is set to = 0 and its chemical
potential to
NC
= 2 while for the right hand column = 0.8 and
NC
= 1.2.
In both cases the damping parameter is = 0.1.
rotating frame, and
NC
the chemical potential of the thermal cloud. In a cylindrically
symmetric trap, the eigenvalue in the rotating frame is related to the eigenvalue in the
laboratory frame by
R
=

L
z
). Thus, in this case we need to set
NC
= 2
to match the eigenvalue in the rotating frame. In Fig. 6.10 (c) and (d) we can see that
for = 0.8, while still showing some oscillatory behaviour, the eigenvalue and angular
momentum are much better conserved using the phenomenological growth equation
than in the case of the pure GPE in the rotating frame in Fig. 6.8. For C = 0,
the problem is linear which means that the normalization of the wave function is not
determined by the eigenvalue , but can be chosen arbitrarily. Hence, the change in
norm in this case does not indicate a numerical problem.
81
Chapter 6. Numerical Propagation Method
As a next test, we looked at a C = 1000 single vortex state. The state was created
using a one dimensional optimization with 1024 points on a grid of range [0, 30], and
subsequently extrapolated onto the two dimensional grid with spatial extent l
x
= l
y
=
45 and 256 points in each direction for the simulations. Propagation of the state with
step size /500 using the pure GPE, leads to a loss in normalization and eigenvalue of
the order of 10
5
over a time period of 20 as shown in Fig. 6.11 (a), which suggests
that the temporal step size is too small. The same simulation with a step size of /1000
0 5 10 15 20
8
6
4
2
0
2
x 10
5
Time []
GPE propagation
(a)
12.8407
Norm1
0 5 10 15 20
2
1
0
1
2
3
x 10
5
Time []
Phen. growth equation
(b)
12.8407
Norm1
Figure 6.11: Accuracy of conserved quantities in propagation of C = 1000 ground state
using (a) the GPE equation, (b) the phenomenological growth equation with
= 0,
NC
=
init
) and = 0.1.
is about one order of magnitude better in and almost two orders of magnitude in
the normalization. However, an oscillation in the eigenvalue similar to the one in
Fig. 6.11 (a) with an angular frequency of 2 remains indicating that the state is not
a perfect eigenstate. This frequency is universal for the rst breathing mode in a two
dimensional system [171].
If we propagate the state with the phenomenological growth equation setting =
0 and
NC
to the expectation value
init
) of the initial state, we obtain the values
shown in Fig. 6.11 (b). We can see that the normalization and the eigenvalue quickly
settle down to some constant value even though the step size is only /500. The
expectation value ) of the nal state is not exactly
NC
though, but slightly higher.
We understand this, recalling the loss that occurs even if the state is propagated in
the pure GPE. A smaller normalization corresponds to smaller chemical potential.
Apparently, the propagation code systematically underestimates the actual value of the
chemical potential. Thus, if an equilibrium is reached in the dynamical simulation, the
82
6.2. Accuracy
true chemical potential of the nal state is actually slightly higher than the equilibrium
value of =
NC
.
If we take the nal state of this simulation and propagate it in the GPE, the loss in
normalization and eigenvalue is the same as before but the oscillation in is of much
smaller amplitude, which proves that the damped nal state of the phenomenological
growth equation is a better eigenstate than the state we obtained by interpolation from
the one dimensional optimized state.
Finally, we set = 0.4, which is below the critical frequency for vortex nucleation
in the C = 1000 case as we will show in section 7.5.3. As opposed to the pure GPE,
where a transformation into the rotating frame does not change the physics of the
system, using the phenomenological growth equation corresponds to a change from a
stationary thermal cloud to a rotating thermal cloud. Nevertheless, an eigenstate of
the laboratory frame should still be a stationary state in the phenomenological growth
equation if
NC
is set to
NC
=

L
z
). However, Fig. 6.12 (a) shows that the
change in normalization and eigenvalue is now almost two orders of magnitude worse
than for = 0. The eect of the angular momentum operator in this case is much larger
0 5 10 15 20
2
0
2
4
6
8
10
x 10
4
Time []
(a)
12.4407
Norm1
0 5 10 15 20
2
1
0
x 10
8
Time []
<
L
>

1
(b)
Figure 6.12: Accuracy of conserved quantities in propagation of C = 1000 ground state
using the phenomenological growth equation with = 0.4,
NC
=
init
)
and = 0.1.
than the test with the harmonic oscillator state shown in Fig. 6.10 or even the tests
in the rotating frame from section 6.2.4 would indicate. A smaller step size of /1000
does not improve this behaviour either. The angular momentum of the condensate
decreases ever so slightly as shown in Fig. 6.12 (b) while in all other simulations with
C = 1000 mentioned above the angular momentum is constant within the order of
83
Chapter 6. Numerical Propagation Method
10
13
. It seems that the change in normalization and eigenvalue seen in Fig. 6.11 (b)
for = 0 is enhanced due to the loss of angular momentum which occurs when particles
from the thermal cloud are scattered into the condensate. Important to note is in any
case that the phenomenological growth equation is self-stabilizing in the sense that the
wave function reaches a stationary state that does not show continued oscillations in
normalization, eigenvalue or angular momentum.
84
Chapter 7
Vortex Lattice Formation in
Cylindrically Symmetric Trap
We are now in a position to apply the phenomenological growth equation to the problem
of vortex lattice formation. In this chapter, we consider the growth of a vortex lattice
from a rotating thermal cloud in a cylindrically symmetric trap. This is the scenario
of the experiment by Haljan et al. [78], where a thermal cloud of
87
Rb atoms was
set in rotational motion and evaporatively cooled until a condensate was formed in a
cylindrically symmetric trap.
For numerical reasons the simulations are performed in the stationary laboratory
frame. In this frame the phenomenological growth equation takes the form
(i )
(r, t)
t
=
_

2
+V
T
+C[(r, t)[
2
+i
_

NC
+

L
z
__
(r, t) (7.1)
in our dimensionless units. We recall that
NC
is the chemical potential of the thermal
cloud and the clouds angular velocity.
7.1 Two Dimensional System
All of the simulations of the phenomenological growth equation presented in this the-
sis are carried out in two dimensions. While three-dimensional (3D) simulations are
in principle possible with our algorithm, they are computationally very time-intensive
which prohibits in practice a thorough investigation of the parameter space. We have
only performed very few 3D simulations, and they conrm the general behaviour ob-
tained in the two-dimensional (2D) case. In this section, we briey discuss the relevance
85
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
of 2D simulations to experiments.
If we assume that all the interesting dynamics occurs in the x-y plane, while the
shape of the condensate remains essentially unchanged in the z-direction, we can sep-
arate the wave function as
(r, t) = (x, y, t)(z). (7.2)
Because (r, t) is normalized to unity, we require that both (x, y) and (z) are sepa-
rately normalized to unity. To obtain a 2D equation, ansatz (7.2) is inserted into the
time-independent GrossPitaevskii equation (GPE)
_

2
+V
T
(r) +C[(r)[
2

(r) = (r), (7.3)


and (z) is integrated out by multiplying the whole equation by

(z) and integrating


over the z-coordinate. The result is a 2D equation essentially of the same form as (7.3),
but includes a constant arising from the kinetic energy of (z) and C is replaced by
C
2D
= C, where =
_
[(z)[
4
.
To be able to compare 2D simulations with real experiments, we need to make some
assumptions about the form of (z) in order to calculate C
2D
. There are two cases where
a two-dimensional approximation might be valuable [172] if the condensate is very
oblate in a pancake-shaped trap with
z
/
r
1, or to describe the central part
of an oblate condensate in a cigar-shaped trap with 1.
In the rst case, if 1 so that the interaction strength is small compared to
the axial connement, (z) is well approximated by the ground state of a harmonic
oscillator
(z) =
_

2
_1
4
e

z
2
4
. (7.4)
The parameter C
2D
becomes then
C
2D
=

2
C. (7.5)
In the second case of a cigar-shaped condensate, we assume a cylindrical cong-
uration of length 2R
z
, where R
z
is the axial ThomasFermi radius (2.51). We then
obtain
C
2D
=
C
2R
z
. (7.6)
In principle, a similiar procedure can be carried out for a condensate in a very tight
radial connement so that it is essentially 1D to obtain a corresponding value C
1D
, but
86
7.2. Simple Growth and Normalization of Wave Function
we are only concerned with 3D and 2D simulations.
7.2 Simple Growth and Normalization of Wave
Function
First, we would like to remind the reader that the normalization of the wave function
is not xed in the phenomenological growth equation (7.1). If we set = 0, the
equation describes a simple exponential growth in atom numbers from a non-rotating
thermal cloud, until the chemical potential of the condensate reaches equilibrium with
the chemical potential of the thermal cloud
NC
(compare section 6.2.5). The nal wave
function fulls the GPE (7.3) with =
NC
, but the wave function is not necessarily
normalized to unity depending on the choice of
NC
and C. Say the wave function is
normalized to n

, it is easy to show that the new wave function = (n

)
1/2
satises
the modied GPE
_

2
+V
T
(r) +C

[(r)[
2

(r) =
NC
(r), (7.7)
where C

= n

C. This new C

can then be compared to real experiments in the usual


fashion using the denition of the non-linearity parameter (2.45), or the appropriate
relation from section 7.1 in the case of a 2D simulation.
The nal size of the condensate is determined by the choice of C or
NC
in an
equivalent way
1
. The larger C or
NC
, the higher is the nal normalization of the wave
function, i.e. the more atoms are in the condensate.
7.3 Initial State
Now, we consider the case of vortex lattice growth from a rotating thermal cloud in
a cylindrically symmetric trap. If not mentioned explicitly, all simulations commence
with a condensate wave function
0
in the ground state of the trap with unit norm,
C = 1000 and chemical potential = 12.7. As discussed in the previous section, C
determines the size of the condensate. We have also performed simulations with much
smaller and larger values of C that show qualitatively the same behaviour. But for the
sake of presentation, we use here the same C value for all simulations, which makes a
1
But not with the same functional dependence. An estimate of the nal normalization could be
obtained from the ThomasFermi solutions (see section 2.4).
87
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
comparison between the dierent simulations more convenient
2
.
Because our equation does not include non-stimulated processes that can initiate
growth into modes of non-zero angular momentum, we decided to add to the initial
wave function a uniform superposition of angular momentum states with l = 1 to 30
on a Gaussian radial prole. The initial state takes then the form

0
+
A
n
n

l=1
g
l
(, ), (7.8)
where
g
l
(, ) = e
2(R)
2
/W
e
il
(7.9)
in polar coordinates and . Typically, the Gaussian had a total amplitude A =
5 10
6
, centred at the ThomasFermi radius R = 7.1 with width W = 4. This
particular condition approximates a small seed in the surface modes, i.e. modes with
no radial node, of corresponding angular momentum l. We have checked that the exact
prole of the initial seed is not important as long as all relevant angular momentum
components are equally populated. The eect of various dierent seed conditions is
discussed in more detail in section 7.5.5.
7.4 General Behaviour
Fig. 7.1 shows density plots of the process of vortex lattice formation from a thermal
cloud rotating at = 0.65 with xed chemical potential
NC
= 12. The damping
parameter is = 0.1. We emphasize that the simulation covers the whole process
from the initiation of the vortex nucleation up to the stabilization of the lattice. In
the current section, we briey give an overview of the dierent stages in this process,
which are considered separately in more detail in later sections. The simulation starts
at t = 0 with a condensate ground state (Fig. 7.1 (a)). After some initiation time a
ring of 19 vortices approaches the condensate from innity (Fig. 7.1 (b)). When this
ring shrinks further (Fig. 7.1 (c)), several vortices are shed from the condensate, and a
dominant ring of 16 vortices passes through the ThomasFermi radius into the interior
of the condensate
3
. As the vortices penetrate, the condensate expands around them,
2
The choice C = 1000 gives medium-sized condensates with about 18 vortices at a rotation of
= 0.75. This is a convenient size for numerical simulations. Much larger condensates would require
a large grid size and smaller time step.
3
Theoretically, the wave function of a condensate extends of course to innity. Whenever we talk
about the interior or dense region of a condensate, or sometimes in short simply the condensate, we
88
7.4. General Behaviour
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 7.1: Density plots of lattice formation in cylindrically symmetric trap with thermal
cloud rotating at = 0.65 for times (a) t = 0, (b) t = 9.32, (c) t = 12,
(d) t = 16.68, (e) t = 21.84, (f) t = 59.6. Other parameters C = 1000,

NC
= 12, = 0.1.
and its angular momentum increases dramatically. The ring of vortices breaks up, and
the vortices distribute themselves irregularly but quasi-uniformly over the condensate
(Fig. 7.1 (d)). Subsequently, over a long period of time further vortices leave the dense
region (Fig. 7.1 (e)), until nally a regular lattice of 12 vortices is formed (Fig. 7.1 (f)).
The nal lattice rotates exactly at the angular velocity of the thermal cloud.
The overall time scale of the process is illustrated in Fig. 7.2, where the key physical
quantities are plotted versus time: the angular momentum expectation value L
z
), the
norm of the condensate and the expectation value ) and its spatial variance

.
During the formation process of the vortex lattice, the chemical potential is in general
not in equilibrium with the thermal cloud. That means the expectation value
)
_
dr

2
+V
T
+C[(r, t)[
2

L
z
_

_
dr[[
2
(7.10)
mean the region of visible density in our density plots.
89
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
0
4
8
<
L
z
>
0 10 20 30 40 50 60
12
13
Time []
(b) <>
<>+

0
1
2
N
o
r
m
(a)
<L
z
>
Norm
Figure 7.2: Time evolution of condensate quantities for case of Fig. 7.1. (a) Angular momen-
tum L
z
) and normalization of wave function; (b) expectation value of chemical
potential ) and its spatial variance

.
is in general dierent from the local chemical potential dened as

loc

[
2
+V
T
+C[[

L
z
]

. (7.11)
Thus, the spatial variance

, which is the standard deviation of


loc
to ) on the
spatial grid, is a measure of the deviation of the solution from equilibrium.
We can identify two main stages. Firstly, the initiation period during which the
angular momentum builds up until it reaches its maximum value at around t 16.
And secondly, the equilibration process of the lattice during which the vortices rearrange
themselves into a regular lattice where

becomes zero in equilibrium with the thermal


cloud.
While Fig. 7.2 (a) shows a sharp increase in the angular momentum at t 13 when
the 16 vortices enter the dense part of the condensate, a logarithmic plot of the angular
momentum versus time in Fig. 7.3 reveals that the growth of the angular momentum is
essentially a continous exponential growth right up to the end of the initiation period.
At the same time as the vortices penetrate the condensate, the norm of the condensate
increases as the condensate takes in the vortices. During the equilibration period, the
angular momentum of the condensate, which reaches its maximum value just after the
penetration of the vortices, decreases slightly and remains then essentially constant.
90
7.4. General Behaviour
0 10 20 30 40 50 60
10
10
10
5
10
0
Time []
<
L
z
>
Figure 7.3: Angular momentum of Fig. 7.2 (a) on logarithmic scale.
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
8
9
10
11
12
13
14
15
16
Figure 7.4: Local chemical potential of condensate for case of Fig. 7.1. (a) t = 0, (b)
t = 9.32, (c) t = 12, (d) t = 16.68, (e) t = 21.84, (f) t = 59.6. Other
parameters C = 1000,
NC
= 12, = 0.1.
Fig. 7.4 shows the local chemical potential for the same time sequence as Fig 7.1
4
.
To start with, the condensate is in the ground state of the trap with a uniform chemical
potential of = 12.7. In a very short time period it roughly equilibrates with the ther-
mal cloud adjusting
NC
by means of losing atoms, i.e. its normalization decreases
(compare Fig. 7.2 (a)). At t 15 there is a dip in the spatially averaged chemical po-
4
At the borders of the images the density of the condensate is too low to calculate a well-dened
local chemical potential, which is the reason for the white regions.
91
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
tential ), and the number of condensate atoms increases while the vortices penetrate
the condensate. At the positions of the vortices the chemical potential shows a large
gradient (Fig. 7.4 (b) and (c)). The vortices move towards greater chemical potential.
This is the case because the atoms tend towards the smaller chemical potential. Thus,
as the atoms move around the vortex, the vortex is pushed towards the lower density
region the atoms leave behind. During the equilibration time, the chemical potential
becomes gradually uniform across the whole condensate as the vortices rearrange them-
selves into a regular lattice (Fig. 7.1 (d) to Fig. 7.1 (f)). We can see from

in Fig. 7.2
(b) that the condensate is far removed from equilibrium as the vortices enter the dense
region of the condensate. The subsequent slow decay of

indicates the long time


needed for stabilization of the vortex lattice.
7.5 Initiation Period
We now consider the dierent stages in the vortex lattice formation separately. Here,
we start with the initiation period, which we analyse in terms of a growth process into
certain surface modes.
7.5.1 Gain analysis
First, we examine the rst stage of the process in which the angular momentum of the
condensate builds up slowly until its maximum value is reached just after the vortices
have penetrated the condensate. To understand this nucleation and growth process of
the vortices, we consider the linearized Bogoliubov excitations above the initial state
with angular momentum l and energy eigenvalues
n,l
, where n contains any other
quantum numbers that classify the state, measured relative to the eigenvalue of the
condensate
C
. Following the linear response theory of section 2.2.2, we write
(t) = e
i
C
t
_

0
+

n,l
e
il
[b
n,l
(t)u
n,l
e
i
n,l
t
+b

n,l
(t)v

n,l
e
i
n,l
t
]
_
, (7.12)
where
0
is the initial rotationally symmetric condensate, is the azimuthal angle and
P
n,l
[b
n,l
(t)[
2
gives the population in the quasi-particle mode with energy
n,l
. As long
as their is no signicant non-linear mixing between the quasi-particles, we can apply
the linear response theory to equation (7.1). That yields an exponential growth of the
quasi-particle populations according to P
n,l
(t) = P
n,l
(0)e
2Gt
, where the gain coecient
92
7.5. Initiation Period
is given by
G =

2
+ 1
(
C
+l
n,l
) (
C
+l
n,l
), (7.13)
since typically
2
1.
We see that only excitations with
n,l
<
C
+l will experience positive growth;
all other modes will experience loss. Because the condensate mode is the n = l = 0
component with the lowest energy
n,l
= 0, this will cause a rapid initial adjustment of
the condensate number until
C
as we see in Fig. 7.2 (b). Subsequently, the gain
of the other components is
G (l
n,l
). (7.14)
7.5.2 Comparison with simulation
The analytic treatment can be veried by decomposing the condensate into angular
momentum components. Instead of using the exact Bogoliubov modes, we project out
all modes with phase circulation l simultaneously using the projector
A
l
(r) =
_
2
0
de
il
(r, ). (7.15)
The occupation of angular momentum l is then given by
P
l
=
1
2
_
dr r[A
l
(r)[
2
. (7.16)
The occupation values P
l
are plotted against time in Fig. 7.5. We see that the l
components above threshold grow exponentially up to the time t 13 (when the
vortex ring passes through the ThomasFermi radius), and the gain is maximum for
l
v
= 16.
Peculiar is the fact that the dipole mode with l = 1 shows such a high occupation
throughout the whole period. This is probably caused by the same eect we have men-
tioned in section 6.2.4. The propagation method has an intrinsic asymmetry because
the computational grid is symmetric around the origin in position space, but not in
momentum space. The permanent transformation from one into the other represen-
tation seems to slightly excite the dipole mode which corresponds to a little centre of
mass motion.
In Fig. 7.6 we compare the gain prediction (7.14) G = (l
0,l
) obtained by
93
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
0
5
10
15
20
0
10
20
30
10
12
10
7
10
2
Time [] l
P
l
Figure 7.5: Evolution of angular momentum occupation probabilities P
l
for l = 1 to 30 for
the case of Fig. 7.1 during initiation time.
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
Angular momentum l
G
a
i
n
Figure 7.6: Comparison of gain coecients of l components obtained from the simulation in
Fig. 7.1 (solid line) with the predicted gain coecients G = (l
0,l
) (dotted
line).
calculating the surface mode energies
0,l
for a condensate with = 12 against the gain
rates measured from the simulation. Our gain prediction is very accurate. We also nd
that the spatial particle density for each l component (7.15) matches the corresponding
Bogoliubov amplitude as shown in Fig. 7.7 for some selected modes. This indicates
that higher excitations are negligible, and justies the use of the simplied projector
(7.15).
94
7.5. Initiation Period
3 4 5 6 7 8 9 10
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
Radial position
|
u
|
2
l = 5
l = 16
l = 24
Figure 7.7: Comparison of projected l components [A
l
(r)[
2
from simulation in Fig. 7.1 to
theoretical Bogoliubov amplitudes [u[
2
. The l values of the Bogoliubov ampli-
tudes are 5 (dotted line), 16 (solid line) and 24 (dashed line) going from left to
right. The amplitudes obtained from the simulation are are indicated by large
dots and rescaled to t the normalization of the theoretical amplitudes.
7.5.3 Critical angular velocity
Expression (7.14) of the gain coecient gives the critical value of for positive gain

c
= min
_

n,l
l
_
. (7.17)
This is the well known Landau criterion [113, 130] which was originally derived as a
condition to create any excitations in a superuid.
In Fig. 7.8 we consider the change of the gain coecients with for the surface
modes. For small all modes have a negative gain, monotonically increasing in absolute
value with increasing l. As increases, a local maximum forms in the gain coecients.
This maximum value grows with increasing , and the gain becomes positive for some
of the modes. As becomes even larger, more and more modes experience positive
gain, and the maximum gain shifts to higher values of l.
A careful analysis reveals that for the above parameters the critical value is =
0.444 with l
v
= 9. We have checked with dynamical simulations that there are no
vortices produced for = 0.44. On the other hand, the same simulation with = 0.45
and = 1 shows a ring of 10 vortices after an initiation time of t 96 as shown
in Fig. 7.9 (a). The value = 1 was chosen that large to accelerate the initiation
95
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
0 5 10 15 20 25
0.8
0.6
0.4
0.2
0
0.2
Angular momentum l
G
a
i
n
Figure 7.8: Gain coecients of surface modes for = 0.3 (dots), = 0.46 (crosses) and
= 0.58 (triangles) in the case of C = 1000 and
NC
= 12. The damping
parameter is chosen = 0.1.
Position x
P
o
s
i
t
i
o
n

y
(a)
5
0
5
Position x
(b)
10
4
10
3
10
2
10
1
Figure 7.9: Density plots of lattice formation in cylindrically symmetric trap with thermal
cloud rotating at = 0.45, just above the critical rotation frequency. (a)
t = 96.12, (b) t = 200. Other parameters C = 1000,
NC
= 12, = 1.
process, because just above the critical angular frequency the gain is very small. The
nal conguration in this simulation consists of a lattice of three vortices depicted in
Fig. 7.9 (b).
96
7.5. Initiation Period
7.5.4 Dominant angular momentum component
The dominant mode with angular momentum component l
v
is the one that maximizes
the gain, and can be found by calculating

l
(l
0,l
) = 0. (7.18)
For our simulation the dominant value is l
v
= 16, which is exactly the number of
vortices that enter the dense region of the condensate in Fig. 7.1 (c).
To understand the essential behaviour, let us consider only the dominant mode in
the expansion (7.12). Because u(r) v(r), we also neglect v(r) and are left with
only two terms, the condensate wave function
0
and the quasi-particle amplitude
b(t)u
l
v
(r)e
il
v
. While the condensate mode has a at phase, the quasi-particle has an
l
v
phase circulation. The interference between the two terms gives rise to vortices at
points of destructive interference (Fig. 7.10). That means that there is a ring of l
v
0 2 4 6 8 10 12
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Radial position
A
m
p
l
i
t
u
d
e
Ground state
u
0,l=16
Figure 7.10: Radial position of vortex core indicated by dotted line as a result of destructive
interference between ground state wave function and quasi-particle amplitude
u. Shown are the ground state of a C = 1000 condensate with = 12, and
the l = 16 surface mode with arbitrary population.
vortices at the radius where the quasi-particle amplitude is the same as the condensate
wave function. Since the long distance behaviour of u(r) is less rapid than the one of

0
, this ring of vortices, which appears initially at innity, will steadily shrink as the
excitation grows. In a more detailed picture, more values of l should be included, and
the corresponding superposition will be an imperfectly circular ring of vortices.
97
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
From equation (7.12) we see that the phase of the ground state evolves as t while
the phase of the dominant excitation is given by t
0,l
v
t +l
v
. The azimuthal angle
of the position of the vortices is given by the condition for destructive interference

v
=
1
l
v
[
0,l
v
t + (2n 1)], (7.19)
where n is an integer between 1 and l
v
. Hence, the angular velocity of the initial ring
of vortices is

v
=
d
v
dt
=

0,l
v
l
v
. (7.20)
In our case, the energy of the dominant excitation l
v
= 16 is
0,16
= 7.78, which
gives the angular velocity
v
= 0.485. Fig. 7.11 is a plot of the angular velocity of
5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
Time []
A
n
g
u
l
a
r

v
e
l
o
c
i
t
y
Figure 7.11: Angular velocity of vortex during lattice formation from thermal cloud rotating
at = 0.65 with C = 1000,
NC
= 12, = 0.1. The velocity is calculated
from the change in polar angle in movies of the density evolution averaged over
seven frames that are separated by t = /25.
one of the vortices extracted from the simulation. While the vortex is still in the low
density region for t < 10, its angular velocity agrees well with the expected value

v
= 0.485. As the vortex enters the high density region its angular velocity increases.
It appears that the ring of vortices breaks up when it reaches the velocity of the thermal
cloud , and the vortices form the lattice subsequently. The nal lattice rotates at the
angular velocity of the thermal cloud. The plot shows some residual oscillation around
which might be explained by the resolution of our spatial grid. For every vortex,
98
7.5. Initiation Period
our detection algorithm returns the four grid points surrounding the vortex centre as
a possible position of the vortex. We chose arbitrarily the upper right corner as the
position of the vortex. Consider a vortex at half the ThomasFermi radius R
TF
. As the
vortex completes half a rotation the maximum possible error in the polar position is
about d = 4dx/R
TF
, where dx is the grid spacing (which is equal in x- and y-direction
in our simulations). Thus, the error in the angular velocity is
d =
2d
T
=
d

=
4dx
R
TF
, (7.21)
where T is the time period to complete a full rotation. This simulation was carried out
on a grid with 256 points and spatial extent of 45. With dx = 45/255 = 0.176 and
R
TF
= 7.1, this gives d = 0.02 for = 0.65, which is approximately the amplitude of
the oscillation of
v
around in Fig. 7.11.
We can obtain a crude estimate of l
v
in the following way. The harmonic trap
combined with a repulsive dimple in the middle due to the condensate density provide
an eective potential that can be written as V
e
= [r
2
/4[ + in the ThomasFermi
limit. The main contribution to the kinetic energy of a mode with angular momentum
l is simply l
2
. Thus, its total energy in a frame rotating at angular velocity is given
by
E
e
(r, l) = [
1
4
r
2
[ + +l
2
l (7.22)
as a function of r and l. This has a minimum very close to the point
r = R
TF
2

, l =
R
2
TF
2
. (7.23)
The question arises how it is possible to obtain a single vortex state. The lowest-
lying l = 1 mode is the Kohn mode with
0,l
= 1. Thus, its gain coecient is always
negative, as can be seen from expression (7.14) (because < 1). Hence, the initial
number of vortices that penetrate the condensate will always be greater than one. From
the estimate (7.23), we can see that the dominant mode l
v
decreases as C decreases
since R
2
TF

C in 2D (see section 2.4). This still holds even though


c
turns out to
increase with decreasing C. Nevertheless, a simulation with small C = 100 still shows
2 vortices in the nal lattice at = 0.65 just above
c
= 0.645, and
NC
= 4, = 1.
We conclude that there might be another mechanism involved in the production of a
single central vortex state, and we return to this question in section 9.3.
99
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
7.5.5 Role of initial seed
In experiments, the initial occupation of non-zero angular momentum components
comes from non-stimulated processes and quantum noise, which are not accounted
for in the phenomenological growth equation. However, these processes are in nature
just as random as numerical noise. Even if a simulation commences with a pure ground
state of the trap without any seed of specic angular momentum, vortices are nucleated
and form a regular lattice. That suggests that very small numerical noise is sucient
to initiate the process. At the outer regions of the grid the phase of the wave function
is not well dened because the density is too low to be represented accurately. Thus,
the phase is basically random, and vortices of any orientation are present, which makes
it possible for non-zero angular momentum modes to grow.
When the noise level in the low density region was articially increased to a level of
up to 10
5
in the amplitude of the real and imaginary part of the initial wave function,
the initiation process was not accelerated at all. Although, initially, a ring of vortices
in the outer low density region appears, this ring disappears quickly, and the whole
process does not seem to be any further aected by the initial condition. This is an
indication that the growth process is governed by a strong selection process due to the
distinct gain for the various l components.
We chose to seed some angular momentum components in the form given by equa-
tions (7.8) and (7.9) for two reasons. First of all, to guarantee reproducible results we
felt we should not rely on numerical noise of unkown magnitude. The second reason
is rather practical. The initiation process is accelerated by the use of an initial seed,
which in turn reduces the time to run a simulation. If we dene the end of the initiation
period as the time when L
z
) reaches 0.1, the initiation process takes t = 21.1 with-
out any seed for the above parameters as opposed to t = 12.3 with a Gaussian ring
of total amplitude A = 5 10
6
. The amplitude has to be bigger than about 10
8
to
accelerate the initiation process discernibly. From Fig. 7.6, we can see that the range
of l = 1 to 30 essentially covers all angular momentum components with signicant
positive gain.
We have also run simulations where we only seeded some selected angular momen-
tum components. For all of the following examples the total amplitude of the Gaussian
ring was A = 10
5
. From our gain analysis, it is no surprise that a seed of the l = 9
component only produces a ring of 9 vortices at the end of the initiation process shown
in Fig. 7.12 (a), and a seed of the l = 20 components a ring of 20 vortices (not de-
picted). If a range of angular components between 4 and 10 is equally populated at the
100
7.5. Initiation Period
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b)
Position x
P
o
s
i
t
i
o
n

y
(c)
5 0 5
5
0
5
Position x
(d)
5 0 5
10
4
10
3
10
2
10
1
Figure 7.12: Density plots after initiation period of lattice formation from thermal cloud
rotating at = 0.65 for dierent initial seeds. (a) Seed of l = 9 component,
t = 10.76, (b) uniform seed of l = 4 to l = 10 components, t = 12.72, (c)
seed of l = 16 and l = 20 components, t = 9.92, (d) seed of l = 4 component,
t = 10.76. Other parameters C = 1000,
NC
= 12, = 0.1.
beginning of the simulation, the mode with the highest gain dominates the initiation
process, which is the l = 10 mode in this case leading to a ring of 10 vortices. How-
ever, Fig. 7.12 (b) shows that this ring is not as regular as the former because other l
components are signicantly occupied, simultaneously.
A superposition of an l = 16 and l = 20 seed yields the dominant ring of 16 vortices
with a regular ring of four additional vortices at a larger radius shown in Fig. 7.12
(c). A more interesting eect was observed when we seeded the l = 4 component only.
Because the gain of the l = 4 component is very small as seen in Fig. 7.6 (b), even
an initial seed of amplitude A = 10
5
is not enough for the l = 4 mode to catch up
with the dominant mode of l = 16 at the end of the initiation process. However, the
resulting conguration of 16 vortices in Fig. 7.12 (d) shows clearly a four-fold symmetry
forming more of a rectangle as opposed to a circle as before in Fig. 7.1 (c).
101
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
7.6 Equilibration
Because the equilibration period is fairly chaotic, it is hard to analyse quantitatively. As
discussed, a good measure of the equilibration time is the time it takes for

to settle
down to zero after its maximum is reached when the vortices enter the high density
region of the condensate. However, as can be seen from Fig. 7.2 (b), this process is
not a steady monotonic decay, but rather happens in several stages. Depending on
the particular conguration of the transient lattice congurations, it can take a long
time to reach full equilibrium. Occasionally, we found cases in our simulations that did
not settle down to a at chemical potential in true equilibrium with the thermal cloud
within a reasonable time frame, in particular if the damping parameter was chosen too
big. The value of = 0.1 that we typically use in our simulations is chosen to guarantee
relative short simulation times without compromising the end result.
Despite these diculties, the equilibration time is roughly proportional to 1/ as
shown in Fig. 7.13 for two dierent rotation speeds of the cloud. While the initiation
period decreases with increasing , because the gain of the dominant mode is higher,
the equilibration time seems unaected by the rotation speed of the cloud.
0 20 40 60 80 100
0
50
100
150
Inv. damping parameter 1/
E
q
u
i
l
i
b
r
a
t
i
o
n

t
i
m
e
= 0.65
= 0.75
Figure 7.13: Equilibration time vs. inverse damping parameter 1/ for two dierent angular
velocities of the thermal cloud.
102
7.7. Properties of Equilibrium Lattice
7.7 Properties of Equilibrium Lattice
The nal lattice is stationary in the frame rotating at the angular velocity of the thermal
cloud . We can obtain the vortex density n
v
using the quantized circulation rule
_
dl v(x) = N
v
, (7.24)
where N
v
is the number of vortices in the enclosed area and is the quantum of
circulation which is = 4 in our dimensionless units. In the laboratory frame, the
lattice is essentially rotating like a rigid body. Hence, on the boundary of the lattice
the velocity eld is given by v(x) = x. Integrating along a circle of radius R, the
quantization rule yields 2RR = n
v
R
2
. This gives
n
v
=
2

=

2
, (7.25)
which is independent of the density of the condensate.
An estimate of the number of vortices in the nal lattice is obtained using the
ThomasFermi radius in the rotating frame R

TF
as the boundary of the vortex lattice
as follows. We recall from section 5.2.3 that the rotation of the lattice gives rise to an
additional centrifugal potential which can be combined with the trap in an eective
potential
V
e
(x) =
1
4
(1
2
)(x
2
+y
2
). (7.26)
The ThomasFermi wave function then takes the form

TF
=


1
4
(1
2
)(x
2
+y
2
)
C
. (7.27)
Usually, is obtained through the normalization condition. However, in our treatment
the norm of the wave function is a dynamical quantity determined by the non-linearity
C and the chemical potential of the thermal cloud
NC
as discussed in section 7.2.
Integration of
_
dx
2
[
TF
[
2
gives
n

=
2
C(1
2
)
. (7.28)
In Fig. 7.14 we compare this prediction (solid line), assuming that =
NC
for the nal
lattice, with the normalization obtained from simulations with C = 1000 and
NC
= 12
for several (open circles). Though the shape of the curve is well represented, the
103
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
0.4 0.5 0.6 0.7 0.8 0.9
0
1
2
3
4
5
Angular velocity
N
o
r
m
a
l
i
z
a
t
i
o
n
Simulation
Pred. =12
Pred. =10.7
Figure 7.14: Normalization of the nal wave function containing a vortex lattice for dierent
angular velocities of the thermal cloud with chemical potential
NC
= 12. The
non-linearity is C = 1000.
values are systematically too large. A much better t to the simulations is achieved by
taking = 10.7. This is due to the fact that in the ThomasFermi approximation the
kinetic energy is neglected, which also includes the energy associated with the vortex
lines.
The radius of the ThomasFermi solution (7.27) is
R

TF
= 2
_

1
2
. (7.29)
Fig. 7.15 attests that this is a good representation of the size of the condensate con-
taining a lattice. Using the ThomasFermi radius, we can nally obtain the number of
vortices in a lattice as
N
v
= n
v
R

2
TF
=
R

2
TF
2
, (7.30)
where we made use of (7.25) for the vortex density. Inserting (7.29) into (7.30) yields
N
v
= 2

1
2
. (7.31)
In Fig. 7.16 (a) we compare the prediction from this expression for =
NC
(solid
line) with the number of vortices in the lattice obtained from the simulations (open
circles). Again, while the shape is well represented, the actual numbers of vortices in
the simulations are considerably lower. Even taking = 10.7, which represents the
normalization of the wave function much better as discussed, does not improve the
104
7.7. Properties of Equilibrium Lattice
15 10 5 0 5 10 15
0
0.002
0.004
0.006
0.008
0.01
0.012
Position x
D
e
n
s
i
t
y

|

|
2
Figure 7.15: Comparison of radial extent of wave function with lattice in case of C = 1000,

NC
= 12 and = 0.65. The solid line is the cross section from the simulation.
The dotted line is the ThomasFermi wave function given by equation (7.27)
assuming =
NC
= 12. The dashed line is the ThomasFermi wave function
assuming = 10.96, which is obtained from equation (7.28) with n

= 1.31,
the actual normalization of the wave function from the simulation.
0.4 0.5 0.6 0.7 0.8 0.9
0
20
40
60
80
100
120

N
u
m
b
e
r

o
f

v
o
r
t
i
c
e
s
(a)
Simulation
Prediction
0 1 2 3
0
10
20
30
40
50
/ (1
2
)
N
u
m
b
e
r

o
f

v
o
r
t
i
c
e
s
(b)
Figure 7.16: Number of vortices in nal vortex lattice for dierent angular velocities of
the thermal cloud with chemical potential
NC
= 12. The non-linearity is
C = 1000.
prediction signicantly. However, Fig. 7.16 (b) shows clearly that the main functional
dependence is given by /(1
2
) in agreement with expression (7.31).
To resolve this disagreement we rst check the possibility that the actual vortex
density is smaller than the one given by expression (7.25). For a perfect hexagonal
105
Chapter 7. Vortex Lattice Formation in Cylindrically Symmetric Trap
lattice with spacing a, the vortex density is n
v
= 2

3/(3a
2
). From the simulation for
= 0.9 (Fig. 7.17), we extract the mean lattice spacing a = 2.89, which gives n
v
=
0.138, while expression (7.25) gives n
v
= 0.143, which is actually in good agreement.
Using expression (7.30) with the former value for n
v
yields 97 vortices and with the
10
4
10
3
10
2
10
1
Position x
P
o
s
i
t
i
o
n

y
10 5 0 5 10
10
5
0
5
10
Figure 7.17: Vortex lattice from simulation with C = 1000,
NC
= 12, = 0.9 and = 0.1
at time t = 200. The white lines indicate the inter-vortex distances we used
to extract the lattice spacing.
latter 101, but the simulation has only 80 vortices in the nal lattice. Our result
that the number of vortices is always smaller than the number predicted by expression
(7.30) is in agreement with experimental observations [79] and also numerical solutions
by Feder et al. [119].
On one hand, the vortex density in our simulations agrees with expression (7.25),
however, on the other hand, there is a small but signicant area at the edge of the
condensate without any vortices which leads to a smaller number of vortices in the nal
lattice than expected from expression (7.30). For small vortex lattices at moderate
rotation rates this area can be quite large as can be seen in Fig. 7.1 (f), but is less
apparent for higher rotation rates as seen in Fig. 7.17 [165]. In experimental density
pictures, this area is also apparent for small lattices [79], but for large vortex lattices at
higher rotation rates the vortices seem to extend out to the very edge of the condensate.
The numerical solutions by Feder et al. [119] even show vortices at the very edge
of the condensate for small vortex lattices. We should keep in mind, however, that
the experimental results and Feders results show the density of a three dimensional
condensate integrated along the orientation of the vortices on a linear scale, which
might underrate the low density region in comparison to the density in the centre of the
106
7.7. Properties of Equilibrium Lattice
condensate, while our results show logarithmic plots of a two-dimensional condensate.
This makes a direct comparison impossible.
107
Chapter 8
Vortex Lattice Formation in
Elliptical Rotating Trap
In this section we will consider the connection between stirring and the growth process
described in the previous section. In most experiments, the vortex lattice is formed by
stirring a pure condensate with an ellipsoidal harmonic trap. In two dimensions, such
a trap takes the form of an ellipse which can be written as
V
T
(x, y) =
1
4
_
(1 +
x
)x
2
+ (1 +
y
)y
2

. (8.1)
In our simulations we always take
y
>
x
, where
2
x,y
= (1 +
x,y
). Ocassionally, we
use a single ellipticity parameter
=

2
y

2
x

2
x
+
2
y
, (8.2)
which is equivalent to
y
=
x
= .
8.1 No Thermal Cloud
As an example of pure stirring, that means stirring in the absence of a thermal cloud
( = 0), we have used a rotating elliptical trap. A time sequence of this process is shown
in Fig. 8.1 for the case of C = 1000 and = 0.65 with a trap ellipticity of
x
= 0.05
and
y
= 0.15. The trap ellipticity is suddenly switched on at t = 0. The condensate
undergoes a periodic oscillation between its original circular shape and an elliptical
shape. After some initiation time, a similarly deformed ring of 28 vortices appears
109
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
Position x
P
o
s
i
t
i
o
n

y
(a)
6 0 6
6
0
6
Position x
(b)
6 0 6
Position x
(c)
6 0 6
10
4
10
3
10
2
10
1
Figure 8.1: Density plots of pure GPE simulation with elliptical trap (
x
= 0.05,
y
= 0.15)
rotating at = 0.65 for times (a) 9.2, (b) 12.84, (c) 17.84. Non-linearity is
C = 1000. The initial state is the ground state of the undeformed trap ( = 0).
in the region of low density, but never approaches the region of higher density. Only
the even angular momentum components become signicantly occupied and undergo
periodic cycling as shown in Fig. 8.2. This can be well understood in terms of the Rabi
0
10
20
30
0
10
20
30
10
10
10
6
10
2
Time [] l
P
l
Figure 8.2: Evolution of angular momentum occupation probabilities P
l
for case of Fig. 8.1.
cycling model [168], noting that an elliptical potential couples predominantly into the
l = 2 mode, from where subsequent coupling into higher even modes is possible. As
before, resonant mixing into surface modes will occur when =
0,l
/l, which gives the
same condition for criticality as found in references [130, 126].
A simulation with the same parameters as before, but = 0.75 and = 0.05
reveals a dynamical instability around the l = 2 resonance as discussed by Sinha and
110
8.1. No Thermal Cloud
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 8.3: Density plots of pure GPE simulation with elliptical trap ( = 0.05) rotating at
= 0.75 for times (a) t = 0, (b) t = 6, (c) t = 13.32, (d) t = 81.36, (e)
t = 160, (f) t = 800.
Castin [129] (see also [172]). Fig. 8.3 (a) and (b) show that the initial behaviour is the
same as in the = 0.65 case with elliptical distortions of the condensate. However,
when the vortices approach the condensate from the low density region, the condensate
shows an s-shape deformation and streaks of low condensate density form around the
vortices (Fig. 8.3 (c)). This is the beginning of a very turbulent phase, whereby
vortices actually enter the centre regions of the condensate as seen in Fig. 8.3 (d) and
(e). But even after a very long propagation time, the vortices do not form a regular
lattice within the condensate (Fig. 8.3 (f)). Because the GrossPitaevskii equation
(GPE) conserves energy, a true equilibrium lattice, which is an energy minimum in the
rotating frame, can never be achieved. However, Lobo et al. [157] have shown that a
lattice like structure can be obtained using the GPE, if the propagation is done in the
rotating frame, not in the laboratory frame. They state that this is due to the dierence
in the periodic boundary conditions which are implicitly imposed by the propagation
method.
111
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
8.2 With Thermal Cloud
Now, we repeat the simulation for = 0.65, but with a thermal cloud. The thermal
cloud is rotating at the same speed as the trap = , and the damping parameter
is chosen to be = 0.1. From Fig. 8.4 we see that the behaviour is now dramatically
dierent. At the beginning, the same elliptical distortions of the condensate occur
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 8.4: Density plots of lattice formation with elliptical trap (
x
= 0.05,
y
= 0.15)
rotating at = 0.65 for times (a) t = 0, (b) t = 4.64, (c) t = 23.08, (d)
t = 27.72, (e) t = 40.04, (f) t = 100 using the phenomenological growth
equation. Other parameters C = 1000,
NC
= 12 and = 0.1. The long axis of
the trap is indicated by the solid line.
(Fig. 8.4 (b)), but the eect tends to smooth out with time. The initial ring of vortices
appears with about 14 vortices. However, it does not remain in the low density region,
but shrinks into the condensate, initially penetrating the condensate at its narrowest
radius (Fig. 8.4 (c)). As the vortices enter the high density region, the prole of
the condensate returns to a circular shape (Fig. 8.4 (d)) while the vortices rearrange
themselves into a regular lattice (Fig. 8.4 (e) and (f)).
The angular momentum projections in Fig. 8.5 show the dierence very clearly. At
rst, the behaviour is very similar to the one shown in Fig. 8.2 without a thermal cloud
where the stirring with an elliptical trap facilitates mixing into even angular momentum
112
8.3. Gain Analysis
0
10
20
30
0
10
20
30
10
10
10
6
10
2
Time [] l
P
l
Figure 8.5: Evolution of angular momentum occupation probabilities P
l
for case of Fig. 8.4.
states. However, the vortex lattice is a result of the growth process, which is essentially
superposed on the eects of stirring, eventually masking them.
8.3 Gain Analysis
In the previous two sections we have used the same projection method to obtain the
population of angular momentum l for Fig. 8.2 and Fig. 8.5 as in the cylindrically
symmetric case given by equations (7.15) and (7.16). We only considered the total
dependence by projecting onto the functions exp(il). As we have seen in section 7.5.2
this is sucient in the cylindrically symmetric case because all eigenstates of a certain
angular momentum l have the same -dependence exp(il). However, even though this
projection method gives an indication of the angular momentum components involved
in the growth process in the case of an anisotropic trap potential, to perform a careful
gain analysis requires the calculation of the excitations in a frame where the trap
potential is time-independent. That means we have to solve the GPE and Bogoliubov
de Gennes (BdG) equations in the frame rotating at angular frequency . In our
dimensionless units, the GPE equation becomes
_

2
+V
T
+C[
R
(r)[
2

L
z
_

R
(r) =
R
(r) (8.3)
113
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
in the rotating frame, and the BdG equations read
Lu
R
i
(r) +C
2
R
(r)v
R
i
(r) =
R
i
u
R
i
(r),
L

v
R
i
(r) +C
2
R
(r)u
R
i
(r) =
R
i
v
R
i
(r),
(8.4)
where
L =
2
+V
T
+ 2C[
R
(r)[
2

L
z
. (8.5)
It is worth emphasizing that in the second line of (8.4) appears the complex conjugate
operator L

, which must not be confused with L

. While in chapter 3 there was


no need to make this distinction in the BdG equations because L was purely real,
with the inclusion of

L
z
we have to be more careful. Because

L
z
is a Hermitian
operator

L

z
=

L
z
. In contrast, the explicit form of

L

z
in the position representation is

z
= i(x
y
y
x
) =

L
z
, i.e. simply the complex conjugate of

L
z
= i(x
y
y
x
).
A transformation into the rotating frame shifts the eigenfrequencies of the excita-
tions due to the angular momentum term in the operator L. In cylindrical symmetry,
when the excitations are eigenfunctions of the angular momentum operator with eigen-
value l, this shift is given by

R
n,l
=
n,l
l. (8.6)
If the excitations are not eigenfunctions of the angular momentum operator, we can
still assign a denite angular momentum to the excitations relative to the angular
momentum of the condensate wave function
R
using the denition [173]
q =
(A(u) A(
R
))
_
dr[u[
2
+ (A(v) +A()
R
))
_
dr[v[
2
_
dr ([u[
2
+[v[
2
)
, (8.7)
where the angular momentum of any wave function is given by
A() =
_
dr

L
z

_
dr[[
2
. (8.8)
Because of this energy shift, the gain coecient in the rotating frame changes
accordingly and is given by
G
R
=
_
( )q
R
_
. (8.9)
For = 0, we immediately retrieve the expression (7.14) with q = l in the cylindrically
symmetric case. In this section, we consider only the case of the thermal cloud rotating
114
8.3. Gain Analysis
at the same angular velocity as the trap ( = ). Then, the gain (8.9) takes the simple
form
G
R
=
R
. (8.10)
8.3.1 Vortex free solutions of the GrossPitaevskii equation
in the rotating frame
From our gain analysis in the cylindrically symmetric case, we know that during the
initiation period of the vortex lattice formation, the populations of the surface modes
on top of the vortex free ground state grow exponentially. Similarly, we expect a growth
of those surface modes in a rotating elliptical trap. Recati et al. [112] have shown that
in the ThomasFermi approximation there exist vortex free solutions of the GPE in the
rotating frame for a wide range of dierent trap ellipticities and angular velocities
as shown in Fig. 8.6. These solutions are characterized by a velocity eld
v = (xy), (8.11)
where is a solution of the equation
0 0.5 1 1.5
0
0.2
0.4
0.6
0.8
1
1
0
3
2
1
=
x
=
y
Rotation frequency
E
l
l
i
p
t
i
c
i
t
y

Figure 8.6: Phase diagram representing the number of stationary solutions of equation
(8.12). The solid line is given by =
_
4(1 2
2
)
3
/27
2

1/2
. This graph
represents the results of Fig. 1 of reference [112], but in our computational
units.
115
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap

3
+(1 2
2
) + = 0. (8.12)
This parameter turns out to describe the eective elliptical deformation of the conden-
sate in the rotating trap
=
x
2
) y
2
)
x
2
) +y
2
)
. (8.13)
We are only interested in the so called normal branch for small to moderate values of
and where the condensate is elliptically deformed with its long axis aligned with
the direction of the weak connement of the trap. There is also an overcritical branch
where deformation of the condensate is orthogonal to the trap potential.
To solve equation (8.3) numerically turns out to be tremendously dicult even
though we could simplify the problem by searching for solutions with xed eigenvalue
, but arbitrary normalization because we were interested in states with =
NC
. This
is considerably easier than the search for an eigenstate with unkown eigenvalue and
unity norm, because depends in a non-trivial way on all other values of the wave
function. Due to the angular momentum operator we have to allow for a complex wave
function which doubles the number of variables compared to the same problem in
the laboratory frame where the phase can be chosen in such a way that is real (in
the vortex free case). Also, the problem seems to be numerically very sti. Even with
a very good initial guess, the solution does not improve very much with every step of
any kind of iterative optimization method. Often it would terminate at a point which
is not a minimum because no further resonable improvement could be achieved.
We tried three dierent methods to search for the vortex free solutions which are
all described in appendix A in more detail. The rst one uses Matlabs PDE toolbox
based on nite element methods which we had successfully applied to nd the ground
and excited states of a three dimensional cylindrically symmetric condensate as pre-
sented in chapter 3. Unfortunately, this toolbox does not allow much control over the
optimization process. In particular, when the step size becomes too small the program
terminates with an error and does not even return the partially optimized state. Also,
when the method is applied to nd the excitations, it tries to solve the resulting linear
matrix equation iteratively to return all solutions at once. Consequently, it can take a
long time to return any results with no real feedback of how far it has come or whether
it is converging at all.
For the second one, we adapted the optimization routines written by Blakie [174]
based on a conjugate gradient approach. This method has been very successful in
nding two and three dimensional states in cylindrical symmetry. For our problem,
116
8.3. Gain Analysis
however, the convergence to vortex free ground states was very slow. The optimization
for states became increasingly worse with larger and . Another disadvantage is the
fact that the excitations have to be found one by one. This requires a good initial guess
to nd any particular solution.
To nd the excitations more easily we adapted an optimization method based on
a basis state expansion. Although the optimization of the vortex free ground states
did not signicantly improve using this method compared to the solutions found with
the conjugate gradient method, the excitations could be calculated all at once using
Matlabs matrix eigenvalue solver eig. However, memory restrictions only allowed
the use of a relatively small number of basis states.
To obtain any reasonably good vortex free ground state it was absolutely crucial to
start with a very good initial guess. Hence, for any set of parameters we always solved
equation (8.12) rst to obtain the phase prole
=

2
xy (8.14)
that corresponds to the velocity eld (8.11). As an initial guess we tried a Thomas
Fermi wave function for the eective trap parameters that take into account the cen-
trifugal forces due to the rotation, and superimposed the appropriate phase prole
(8.14). However, this did not lead to very satisfactory states. A better approach was to
obtain the states with consecutively increasing and . Starting with a cylindrically
symmetric ground state for = 0, we calculated states with gradually increasing us-
ing as initial guess the amplitude of the state from the previous with the new imposed
phase prole (8.14). Similarly, we extend the solutions to states with ,= 0, using as
initial guess the state with same ellipticity, but the previous value for . Despite this
very systematic approach, the nal optimality of states with higher and became
increasingly worse as shown in Fig. 8.7. Here, the optimality is the sum of the square
residuals scaled by the number of grid points as explained in Appendix A.2.1.
Nevertheless, we believe that the states we obtained are good enough to give us some
indication whether the gain theory that works so well for the cylindrically symmetric
case is still valid for an elliptical rotating trap. Fig. 8.8 shows the vortex free ground
state for = 0.4 and = 0.04. The phase prole has the expected saddle like shape
described by equation (8.14). Also, the parameter = 0.0227 obtained by evaluating
(8.13) for this state compares well to the value = 0.0235 we would expect from
equation (8.12). The angular momentum of this state due to the deformation of the
condensate is

L
z
) = 0.0103. We can compare this to the value obtained by Recati et
117
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
0.4 0.5 0.6 0.7
10
6
10
5
10
4
10
3
Angular frequency
O
p
t
i
m
a
l
i
t
y
= 0
= 0.04
= 0.1
Figure 8.7: Optimality of vortex free solutions of the GPE in the rotating frame obtained
with conjugate-gradient optimization method.
10
4
10
3
10
2
10
1
Position x
P
o
s
i
t
i
o
n

y
(a)
5 0 5
5
0
5
1
0.5
0
0.5
1
Position x
P
o
s
i
t
i
o
n

y
(b)
5 0 5
Figure 8.8: Vortex free ground state in an elliptical trap with ellipticity = 0.04 rotating
at = 0.4. (a) Condensate density, (b) phase prole.
al. [112] in the ThomasFermi limit which is given by

L
z
) = . (8.15)
Here, is the irrotational value of the moment of inertia given by
=
1
2n
(y
2
x
2
))
2
x
2
+y
2
)
. (8.16)
in our dimensionless units, where n is the normalization of the wave function. For our
118
8.3. Gain Analysis
parameters this gives

L
z
) = 0.0104 which is in very good agreement with the value
for the state numerically obtained. For higher and , this agreement becomes worse.
E.g. for the state = 0.65 and = 0.04 we obtain = 0.1258 and

L
z
) = 0.1977 while
equations (8.12) and (8.15) give = 0.1472 and

L
z
) = 0.2011, respectively. In both
cases, the values of the numerically obtained states are smaller than the theoretical
values which indicates that the elliptical deformation is too small. This might well be
the case because of the way we progressively increase when calculating the states.
The optimization method seems to have diculties following the very subtle increase
in the spatial extent of the wave function due to the higher rotation rate of the trap.
A nal attempt was made to improve the optimality by propagation using the phe-
nomenological growth equation with
NC
= 12 and = 0.1. This approach is only
possible as long as is smaller than the critical angular velocity for vortex nucleation,
i.e. if the vortex free state is really the lowest energy state. Otherwise the phenomeno-
logical growth equation will drive the state towards a vortex lattice. For = 0.4 the
optimality improves by an order of magnitude as shown in Fig. 8.9. The oscillation
seen at the end has frequency , which must be due to the way the optimality is cal-
culated using nite dierences. For the nal state = 0.0228 and

L
z
) = 0.0105 are
only slighly dierent and not much closer to the theoretical values calculated in the
ThomasFermi limit.
0 5 10 15 20
10
6
10
5
10
4
Time []
O
p
t
i
m
a
l
i
t
y
Figure 8.9: Improvement of optimality by propagation with phenomenological growth equa-
tion. The initial state was created with the conjugate gradient method for
C = 1000, = 0.04 and = 12. Additional parameters in phenomenological
growth equation
NC
= 12 and = 0.1.
119
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
8.3.2 Excited states on vortex free solutions in the rotating
frame
Despite the diculties with the accuracy of the vortex free ground state solutions we
proceed to calculate the excitations on top of the vortex free solutions by solving the
BdG equations in the rotating frame (8.4). This gives us the relevant excitation modes
for a careful gain analysis in an elliptical rotating trap. Due to the error in the ground
state solution, we expect slight inaccuracies in the excitation spectrum. Nevertheless,
the states will be a much better basis for a gain analysis than the simple exponentials
exp(il).
Fig. 8.10 shows the excitation spectrum on the vortex free solution rotating at
= 0.66 for dierent elliptical deformations of the trap. All modes with negative en-
ergy have a positive gain coecient according to (8.9). We can see that the excitation
energies increase slighly for the majority of the modes as the trap ellipticity increases.
However, as increases, there are also side branches developing with modes accumu-
lating around the angular momentum l
v
16 of the minimum energy excitations in
the main branches. The excitation frequencies of these modes extend to considerably
smaller negative frequencies than the main branches of the excitation modes.
The modes of the lowest side branch have a distinctively dierent shape compared
to the normal surface modes. Fig. 8.11 shows two of the ordinary surface modes with
no radial node for = 0.04, which consist, as in the cylindrically symmetric case, of a
density ring with a certain phase rotation. The ring is more elliptical due to the form
of the trapping potential, and the density is slighly higher further away from the centre
due to centrifugal forces. The phase prole is distorted accordingly, but the overall
phase rotation corresponds roughly to the angular momentum value q of the modes,
which are not integer numbers anymore because of the elliptical deformation.
The modes of the new side branches, however, have distinct density peaks. Fig. 8.12
shows the rst three point-symmetric modes of the lowest side branch.
The rst of those modes has one peak at either side of the narrow waist of the
condensate (Fig. 8.12 (a)), and each subsequent mode has one more peak at either
side as can be seen from Fig. 8.12 (c) and Fig. 8.12 (e). When we compare the phase
diagrams of Fig. 8.12 (d) and (f) to the one of the rst mode in Fig. 8.12 (b), we
can see that the gaps arise due to additional phase singularities (points of vortex like
phase rotations) that are pushed in from the outside. As we can see from the spectrum
in Fig. 8.10 (c) the corresponding odd modes coincide with the even ones. This is
a consequence of their density distribution which is essentially the same as the one
120
8.3. Gain Analysis
4
3
2
1
0
1
2
3
(a) (a)
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
(b) (b)
10 0 10 20 30 40 50
4
3
2
1
0
1
2
3
(c) (c)
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
10 0 10 20 30 40 50
(d) (d)
Angular momentum q
Figure 8.10: Excitation spectrum on vortex free ground state with C = 1000 and = 12
rotating at = 0.66 for trap ellipticities (a) = 0, (b) = 0.02, (c) = 0.04
and (d) = 0.06. Even modes with respect to a point-reection on the origin
are marked by a circle, and odd modes with a cross.
of the even modes in Fig. 8.12. They only dier in the phase prole as can be seen
in Fig. 8.13. The winding number of these modes is odd and one smaller than for
the corresponding modes in Fig. 8.12. However, this dierence occurs in a region of
negligible density so that it does not alter the energy or angular momentum of the
modes considerably. Because the modes in the side branches have the lowest energy,
and therefore the highest gain, they are populated more rapidly than the others. This
is the reason vortices penetrate the condensate at its narrowest radius in the vortex
lattice formation process in a rotating elliptical trap.
121
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(c)
0
0.005
0.01
0.015
0.02
Position x
P
o
s
i
t
i
o
n

y
(b)
5 0 5
5
0
5
Position x
(d)
5 0 5


0

Figure 8.11: Ordinary surface modes in trap with = 0.04 rotating at = 0.66 on vortex
free state with C = 1000 and = 12. (a) Density [u[
2
and (b) phase of q = 5.55
mode with = 0.77. (c) Density [u[
2
and (d) phase of q = 29.82 mode with
= 0.55.
8.3.3 Comparison with simulation
With the appropriate excited states on top of the vortex free irrotational state, we are
now in a position to look at the gain of these states in our simulations of vortex lattice
formation in a rotating elliptical trap. As we have seen in section 8.2, a sudden switch
on of the ellipticity of the trap leads to oscillations of the condensate shape between
a circular and elliptical density distribution. In the case of Fig. 8.4, the simulation
commences with a ground state of an isotropic trap, and at t = 0 the trap is suddenly
set into rotation at = 0.65 with the ellipticities
x
= 0.05 and
y
= 0.15. In Fig. 8.14
(a) the resulting evolution of the deformation parameter (8.13) is shown. The large
oscillations are clearly visible at the beginning of the simulation. After the vortices
enter the condensate at around t = 23 (cp. Fig. 8.4 (c)) the shape of the condensate
returns to a more circular one, and the oscillations subside.
However, this initial oscillation is detrimental to a careful gain analysis because
with the change in the shape of the ground state all the relevant excited modes change
as well. To avoid this diculty we performed a simulation starting with a vortex
122
8.3. Gain Analysis
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(c) (e)
0
0.01
0.02
0.03
0.04
0.05
Position x
P
o
s
i
t
i
o
n

y
(b)
5 0 5
5
0
5
Position x
(d)
5 0 5
Position x
(f)
5 0 5


0

Figure 8.12: Point-symmetric modes of lowest side branch in trap with = 0.04 rotating at
= 0.66 on vortex free state with C = 1000 and = 12. (a) Density [u[
2
and
(b) phase of q = 13.78 mode with = 3.47. (c) Density [u[
2
and (d) phase
of q = 14.41 mode with = 3.81. (e) Density [u[
2
and (f) phase of q = 15.02
mode with = 2.91.
Position x
P
o
s
i
t
i
o
n

y
(a)
5 0 5
5
0
5
Position x
(b)
5 0 5
Position x
(c)
5 0 5


0

Figure 8.13: Phase of the rst modes of the lowest side branch with odd symmetry in the
same case as Fig. 8.12. The angular momentum and energy of these modes
coincide with the ones with even symmetry.
free ground state of the trap with ellipticity = 0.04 rotating at = 0.6 which
was obtained by the optimization method described in section 8.3.1. Experimentally,
123
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
0 10 20 30 40
0
0.05
0.1
0.15
Time []
D
e
f
o
r
m
a
t
i
o
n

P
a
r
a
m
e
t
e
r

(a)
0 10 20 30 40
0
0.05
0.1
0.15
Time []
D
e
f
o
r
m
a
t
i
o
n

P
a
r
a
m
e
t
e
r

(b)
Figure 8.14: Deformation of condensate in rotating elliptical trap. (a) Sudden switch on of
trap ellipticities
x
= 0.05,
y
= 0.15 at t = 0 rotating at = 0.65 correspond-
ing to case of Fig. 8.4. Initial state is the ground state of the non-rotating
isotropic trap. (b) Initial condensate vortex free eigenstate of elliptical trap
with = 0.04 rotating at = 0.65. In both cases, other parameters are
C = 1000,
NC
= 12, = 0.1.
such a state can be achieved in a trap with xed ellipticity by slowly increasing the
rotation frequency of the trap from zero to the desired value [76]. As expected in that
case there is no large scale oscillation of the condensate shape at the beginning of the
simulation as can be seen from Fig. 8.14 (b). However, the elliptical deformation of
the condensate increases at the beginning of the simulation which points to the fact
that the initial optimized state is not a perfect solution of the GPE in the rotating
frame. As mentioned in section 8.3.1, the values we obtain for our optimized states
are always smaller than the theoretical values predicted in the ThomasFermi limit by
equation (8.12). Nevertheless, we analyse the occupation of the excited states during
the initiation period in terms of the excitation modes calculated for the initial state.
The projection
P
q
=

_
dr(

u
q
v
q
)

2
. (8.17)
onto those states gives us the population shown in Fig. 8.15. Since the initial state
is already of elliptical shape, we do not observe the oscillatory behaviour of Fig. 8.5.
Instead, the initial population levels reect the multipolarity of the trapping potential
with the highest population in the q 2 and q 4 mode, and also a fairly high
population in the q = 14.89 mode with a sixteen-fold phase rotation, while there is only
little population in modes with odd angular momentum. The population in the former
stays essentially constant throughout the initiation period up to t 11 when the
124
8.3. Gain Analysis
0
10
20
0
10
20
30
10
10
10
6
10
2
Time [] q
P
q
Figure 8.15: Evolution of angular momentum occupation probabilities P
l
for case of
Fig. 8.14 (b) analysed in terms of Bogoliubov quasiparticle amplitudes.
rst vortices enter the dense region of the condensate. For higher angular momentum
modes we can see an exponential growth in the occupation numbers as expected.
Fig. 8.16 shows the comparison of the gain coecients extracted from the simulation
of Fig. 8.15 to the values obtained from the Bogoliubov excitation energies (8.10). While
0 5 10 15 20 25 30
0.1
0
0.1
0.2
0.3
0.4
0.5
Angular momentum q
G
a
i
n
Figure 8.16: Gain coecients for population growths in Fig. 8.15. The dots represent the
theoretical values obtained from the numerical solution of the BdG equations
in the rotating frame. The circles are obtained from the dynamical simulation
in Fig. 8.15.
125
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
the agreement is not as good as in the cylindrical symmetric case of Fig. 7.6, the general
trend is clearly recognizable with the maximum gain at around q 14, where the modes
of the rst side branch in the excitation spectrum lie. Interestingly, we do not see lower
gain for the second side branch of modes with one radial node that we would expect
from the excitation energies. They must be excited due to the high occupation of the
corresponding modes with no radial node.
Because this particular simulation was only performed on a grid of 256256 points,
we were not able to extract the population of modes with q > 26. Also, some of the
modes with odd angular momentum are missing because their initial population is small
and there is no signicant mixing into them. Thus, their population does not reach a
level high enough to be resolved by our projection method during the short initiation
time.
8.3.4 Critical velocity and importance of quadrupole mode
1
In section 8.3.3, we have shown that our analysis of the initiation period of vortex
nucleation from chapter 7 is also valid in the case of a rotating elliptical trap. Hence,
the critical angular velocity is still given by the Landau criterion. In principle, the
critical angular velocity does not change much using a rotating elliptical trap as we
can deduce from the excitation spectrum in Fig. 8.10. However, as discussed in section
4.5, in practice the critical frequency observed experimentally exceeds the frequency
given by the Landau criterion considerably. Dalfovo and Stringari [130] were the rst to
suggest that the relevant threshold for vortex nucleation depends on the multipolarity of
the stirring potential, and could thus be signicantly larger than the frequency given by
the general Landau criterion. In the case of a rotating elliptical potential, for example,
it is expected that mostly the quadrupole mode l = 2 is excited. This agrees with
our results presented in section 8.2 (Fig. 8.5), where the l = 2 mode has the highest
population (during the initiation stage) and mixes only into higher angular momentum
modes with even multipolarity. Dalfovo and Stringari also showed that the Landau
criterion for the quadrupole mode

c,l=2
=

l=2
2
0.71 (8.18)
1
After the completion of the examination process, I noticed that some of the results presented in this
section might not be accurate because I did not explicitly assure the orthogonality of the excitations to
the vortex free ground state. However, the content of this section has been left unaltered to document
the state of this research at the time of submission. We will address the problem in further research.
126
8.3. Gain Analysis
is essentially independent of the particular anisotropy of the trap =
z
/
r
in three
dimensions. This result still holds in the two dimensional case.
Hodby et al. [152] have systematically performed experiments to study the region of
vortex nucleation for a rotating elliptical trap for a wide range of rotation frequencies
and trap ellipticities. One of their results is shown in Fig. 8.17, where

is the rotation
frequency of the trap in units of

=
_
(
2
x
+
2
y
)/2, identical to our . The trap
Figure 8.17: Critical conditions for vortex nucleation. The data points mark the minimum
trap deformation for nucleation at a particular

. Vortices may be formed in
region 2. The solid line shows the theoretical limit of stability of the quadrupole
mode, which is stable in region 3. This gure is taken from Ref. [152].
ellipticity is adiabatically ramped on from = 0 to its nal value over a time period in
the range of 200 ms to 1 s. They measured the minimum ellipticity required for a given
rotation rate to nucleate vortices. From this data, they identied three distinct regions
of behaviour as labelled in Fig. 8.17 with boundaries delimited by the data points. In
region 1, no vortices are created. In region 2, vortices may be formed. In region 3,
the condensate follows the so-called overcritical branch of reference [130], where the
condensate has an elliptical density distribution orthogonal to the trap potential. The
boundary between region 2 and 3 is well understood in terms of a stability analysis of
this particular overcritical quadrupole mode. The solid line in Fig. 8.17 indicates the
theoretical boundary of where this mode exists, and is given by
=
2

_
2 1
3
_
3/2
, (8.19)
127
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
in good agreement with the experimental data. However, the linear increase in the
ellipticity necessary to create vortices below
c,l=2
, which gives the boundary between
region 1 and 2, cannot be explained by the stability of a quadrupole mode, because
the normal branch is stable in both regions, and the overcritical branch is not stable
in either.
With this experimental result in mind we searched for an anomaly of the lowest
quadrupole mode in the excitation spectra. We identied the quadrupole mode by
examining the quasiparticle amplitudes in detail looking for a 4 phase circulation.
For = 0.02 the quadrupole mode is depicted in Fig. 8.18. It has the typical features
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(c)
0
0.01
0.02
0.03
Position x
P
o
s
i
t
i
o
n

y
(b)
5 0 5
5
0
5
Position x
(d)
5 0 5


0

Figure 8.18: Lowest l = 2 quadrupole mode in trap rotating at = 0.66 on vortex free
state with C = 1000 and = 12. (a) Density [u[
2
and (b) phase of mode with
q = 1.93 and = 0.12 for trap ellipticity = 0.02. (c) Density [u[
2
and (d)
phase of mode with q = 0.60 and = 0.15 for trap ellipticity = 0.06.
of a surface mode with an elliptically deformed, closed density ring and a quadrupole
phase circulation. Its angular momentum is q = 1.93 and its excitation frequency
= 0.12. In the excitation spectrum of Fig. 8.10 (b), this mode is marked by the left
most circle in the lowest main branch. Because the trap ellipticity is small, its angular
momentum is close to the integer value of q = 2 of the quadrupole mode in the isotropic
case shown in Fig. 8.10 (a). Because the excitation frequency of this mode is positive,
128
8.3. Gain Analysis
its gain coecient is negative.
As for the other modes, the excitation frequency of the quadrupole mode increases
slightly with increasing . At the same time, its angular momentum decreases. As
we see in Fig. 8.10 (c), when = 0.04, the angular momentum is now closer to the
q 1 dipole mode than to q = 2. Eventually, for = 0.06 (Fig. 8.10 (d)), the angular
momentum of the quadrupole mode becomes smaller than the the dipole modes angular
momentum. The quadrupole mode in this case is shown in Fig. 8.18 (c) and (d). Even
though the angular momentum is q = 0.60, the phase prole in Fig. 8.18 (d) still
exhibits a quadrupole symmetry. However, the density distribution in Fig. 8.18 (c)
is very dierent from the density distribution for the = 0.02 case of Fig. 8.18 (a).
Instead of a density ring with no discernible density contribution in the centre of the
trap, there is now a large central density peak besides the ring. Also, as opposed to the
usual density distribution of surface modes in a rotating trap that show density peaks
at the end of the largest axis of the elliptical ring, as depicted in Fig. 8.11, the density
is now concentrated along the shortest axis of the condensate.
We have examined this anomaly in the mode shape in detail for in the range of
[0.6, 0.7] and [0, 0.1], and found that it has the consequence that the q value of the
quadrupole mode decreases relative to the dipole mode. While the dipole mode has
always a q value very close to 1, the q value of the quadrupole mode may even go below
1. The results are summarized in Fig. 8.19 where we have indicated with triangles the
smallest ellipticity values for which the quadrupole surface mode has a smaller angular
momentum q than the lowest dipole mode. We were only able to nd this phenomena
for 0.66. In the case of = 0.64, we found for = 0.08 that q = 1.28, and
for = 0.1 that q is of about the same size. While the angular momentum of the
quadrupole mode thus approached 1, it did not become smaller than 1. For = 0.62,
we were only able to identify a surface mode with a quadrupole phase prole in the
spectrum when 0.08, but not for = 0.1. This is marked in Fig. 8.19 by a cross.
Despite these diculties, the results match the slope of the boundary between region
one and two in Fig. 8.17 from Hodbys et al. [152] experimental exploration of vortex
formation reasonably well. This leads us to conclude that the change in nature of the
quadrupole mode for high enough ellipticities plays an important role in the nucleation
of vortices for <
c,l=2
. Because of the quadrupole phase prole, an elliptical rotating
trap would still primarily excite this mode, but because its angular momentum is now
actually around one or even smaller, presumably mixing into modes of any multipolarity
is possible. This initiates a rapid vortex nucleation process.
129
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
0.6 0.65 0.7
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Rotation frequency
E
l
l
i
p
t
i
c
i
t
y

Figure 8.19: Critical conditions for vortex nucleation. The experimental data from Fig. 8.17
[152] are reproduced as lled circles. The triangles represent the minimum trap
deformation for which the lowest quadrupole mode has an angular momentum
smaller than one. For = 0.62 we could not identify a surface mode with
quadrupole phase prole for = 0.1, which is marked with a cross.
8.4 Experiments on Vortex Nucleation and Lattice
Formation
To conclude our discussion on vortex nucleation and vortex lattice formation we give
a critical evaluation of our model in relation to the experiments carried out in this
area. For this purpose, we rst give an overview of the experimental results and then
emphasize any discrepancies with our model. In this section, we concentrate only on the
nucleation and formation of vortex lattices in those experiments that use some form of
rotational stirring as considered in our simulations. In other experiments, vortices have
also been produced using a spatially selective coupling of atoms into a dierent internal
spin state [72, 175], and by a linear motion of a laser beam through the condensate
[176, 79], but these will not be discussed here.
Experiments at JILA
Cornells group at JILA have carried out experiments that have the simplest conceptual
basis, and are most readily related to our theory. In the work reported by Haljan et
al. [78], a thermal cloud of
87
Rb atoms was set in rotational motion in a cylindrically
130
8.4. Experiments on Vortex Nucleation and Lattice Formation
symmetric trap, and evaporatively cooled until a condensate was formed. The cloud of
about 6 10
6
atoms is held in a shallow, oblate trap with

,
z
= 26.8, 13.6 Hz,
and the trap is deformed in the x-y plane to a large ellipticity of 25% and rotated
around the z-axis until the rotation of the thermal cloud reaches a steady-state value
after about 15 s. The large ellipticity is found to be necessary not only to get the
cloud rotating, but also to sustain ongoing rotation, probably (as suggested by the
JILA group) to overcome a small, static asymmetry in the trap that acts to despin the
cloud. After the trap ellipticity is switched o, the trap is adiabatically distorted to
a prolate shape with

,
z
= 28.35, 5.45 Hz. The cloud is then evaporatively
cooled by removing atoms with large axial displacement that carry hardly any angular
momentum. Hence, the remaining atoms need to spin up for the angular momentum
per particle to stay xed. At the end of the evaporation process they end up with
between 3 10
5
and 8 10
5
atoms in a condensate rotating at very high rotation rate
of up to 0.94

.
This experiment clearly demonstrates the importance of the thermal cloud in the
formation of a vortex lattice because there is no other stirring mechanism that could
supply angular momentum to the condensate. The calculations we have presented in
chapter 7 models exactly this scenario, where the lattice formation is driven entirely
by the rotating thermal cloud. Our treatment is simplied, of course, by neglecting
any possible dynamical evolution of the thermal cloud. The major dierence between
the simple predictions of our model and the experiment is that the JILA group report
a threshold for the observation of a single vortex which lies below the angular velocity
given by the Landau criterion, while in our model the critical frequency for vortex
nucleation is given by the Landau criterion. In section 9.3 we discuss this issue in
detail and give a possible explanation for the dierence, based on a slowing down of
the thermal cloud during the course of the experiment.
Experiments at ENS
Dalibards group at ENS in Paris were the rst to create vortices in a pure single-
component condensate [73]. The basic strategy of that experiment was to start with a
thermal cloud of
87
Rb atoms, and while evaporatively cooling it to condensation, to also
stir it continuously with an elliptically shaped trap. The trap potential was formed by
a combination of a prolate Ioe-Pritchard magnetic trap (

,
z
= 2219, 11.7 Hz)
and a laser beam that was rapidly switched between two symmetric positions around
the trap axis and at the same time rotated around this axis at a uniform angular
131
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
velocity. The time average of the combined potentials is a rotating harmonic ellipsoid.
In this rst experiment, the stirring was performed throughout the evaporation process
for about 500 ms to obtain a condensate of about 10
5
atoms with between one and
four vortices. Using the same technique, they have also observed up to 11 vortices in a
slightly shallower trap with

,
z
= 2169, 11.7 Hz [123].
In later experiments [74], the stirring was done after the evaporation process had
been completed, i.e. a non-rotating condensate was created rst, and then stirred with
the ellipsoidal harmonic potential. This has become the standard technique to create
vortex lattices. Both methods lead to the same result, but the latter one gives a better
control over the number of atoms in the condensate. Chevy et al. [74] measured the
angular momentum of the condensate as a function of the rotation frequency of the
trap by exciting the quadrupole modes with m = 2 using the ellipsoidal trap potential
at xed axes. The frequency split between these two modes in the presence of vortices,
which can be extracted from the precession of the eigenaxes of the modes relative to
each other (see also section 10.2.1), is proportional to the angular momentum of the
condensate [177]. By stirring for 900 ms they obtained a sharp critical frequency for
vortex nucleation of
c
= 0.65

which was not very sensitive to the number of atoms


in the condensate. At higher rotation rates, they rst observed multiple vortices with
up to ve vortices, but for even higher angular velocities the lattice structure became
turbulent and the angular momentum decreased.
In another experiment, Madison et al. [76] explored the steady state solutions of an
elliptically deformed condensate in a rotating trap [112] for a variety of trap ellipticities
and rotation rates, a scenario we have discussed in section 8.3.1. In one series of
experiments they xed and increased at a constant rate from 0, or decreased it
from

to follow the normal branch and the overcritical branch, respectively. Vortices
occured only while was increasing, and for > 0.75

. In another series, was


kept constant while was increased and it was found that vortex nucleation was only
possible in a certain window of values. In all of these experiments they discovered
that the vortex nucleation was related to a dynamical instability. This is in contrast to
a third series of experiments, where the trap ellipticity was switched on rapidly. Then,
vortex nucleation was found to be possible only for a range of values around the
quadrupole resonance 1/

with the range of increasing with increasing .


A key question that arose when these experiments were performed was why the
observed critical angular velocity was much higher than expected from the thermo-
dynamic critical velocity or the Landau criterion. We recall that the critical angular
velocity is always higher than the thermodynamic critical velocity because of the en-
132
8.4. Experiments on Vortex Nucleation and Lattice Formation
ergy barrier for vortex penetration [118, 119]. From the results in our model, we would
expect vortex nucleation to always occur for angular velocities above the Landau crit-
ical velocity. Obviously, the growth rate (gain) of any surface mode is too small to
have a signicant eect in these experiments on the time scales involved. This can
be understood in terms of our model because when the condensates are fairly small
(as they are in the experiment) the Landau critical velocity is quite high [130], and
therefore the gain very small for all modes even at high rotation rates. This would also
explain why no vortices are observed in the experiments above a certain rotation rate.
Vortices can only nucleate if one or more surface modes become highly populated by
some other means.
Depending on the particular procedure in the experiment, we believe that there
are two distinct mechanisms involved that lead to a signicant population in certain
surface modes. First of all, in the experiments where the ellipticity or the rotation
frequency are gradually ramped on the condensate follows either the normal branch or
the overcritical branch depending on the starting point of the experiment. In that case,
vortices are only observed when the respective branch becomes dynamically unstable at
which point the vortex free stationary state decays, and many other angular momentum
states become signicantly occupied. In the other series of experiments where the trap
ellipticity is suddenly switched on, the condensate follows neither branch, but shows
some initial shape oscillations. If the ellipsoidal trap is rotating at a frequency close to
the resonance of the quadrupole mode, this mode becomes occupied as suggested by
Dalfovo and Stringari [130], and subsequent mixing into higher modes of even angular
momentum occurs. In any case, as soon as there is a signicant population in non-zero
angular momentum modes the gain process can take over, which leads to the nucleation
of vortices.
Experiments at Clarendon Laboratory, Oxford
A closely related experiment was carried out by Foots group at Oxford [152] where
they started with an oblate
87
Rb condensate of 2 10
4
atoms at a temperature of
0.5T
c
in a TOP trap with trapping frequencies

,
z
= 262, 175 Hz. The trap
ellipticity was then linearly switched on over a time period of 200 ms while rotating at
a xed angular velocity . The condensate was held in the rotating anisotropic trap for
another 800 ms before being released. They systematically mapped out the region of
vortex nucleation for a wide range of and values and found it to be centred around
the quadrupole resonance at 1/

with the frequency range increasing for higher


133
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
ellipticities. They also noted that the temperature had no eect on the nucleation
condition, but did aect the equilibration and decay of the vortex lattices.
While the high limit border of the region of vortex nucleation could be explained
in terms of a dynamical instability, the low limit has not been explained prior to this
thesis. In section 8.3.4 we showed that it is due to a change of nature of the quadrupole
mode which can facilitate mixing into a number of surface modes, and then the gain
process for these modes leads to vortex nucleation.
Experiments at MIT
Ketterles group at MIT has also done a number of stirring experiments, and have pro-
duced some strikingly large lattices with more than 100 vortices [79]. Using a relatively
shallow trap with

,
z
= 284, 20 Hz that suppressed inelastic losses they formed
a large
23
Na condensate with 5 10
7
atoms and a condensate fraction of more than
90%. The condensate was stirred at the condensates diameter by a pattern of two
blue-detuned laser beams, whose beam waist was comparable to the ThomasFermi
radius, rotating symmetrically around the trap axis. This dipole pattern was ramped
up over 20 ms and ramped down over 20 ms after a stirring for a variable length of
time, typically of the order of 500 ms. They observed highly-ordered triangular lattices
in the condensate, so called Abrikosov lattices. As in the experiments described above
the peak number of vortices was observed for a stirring frequency at the quadrupole
resonance. The critical frequency for vortex nucleation, on the other hand, agreed well
with the Landau criterion. They state that the instability predicted by the Landau
criterion occurs on a practical time scale only when the cloud is suciently deformed
to excite high-l modes. The observed number of vortices was always smaller than the
estimate
N
v
= 2R
2
/, (8.20)
an observation that agrees with our simulations. However, in our simulations this was
the case because there was a considerable area at the edge of the condensate without
any vortices (see section 7.7). In the experiment, on the other hand, the vortex lattice
appears to ll the whole condensate with remarkable regularity up to the very edge of
the condensate.
With the same setup, Raman et al. [77] stirred the condensate with anisotropies
of dierent symmetries (l = 2, 3, 4) using laser beams forming 2-, 3-, and 4-point
patterns. They found discrete resonances for vortex nucleation conrming the role
of discrete surface modes in vortex formation. The central observation was that a
134
8.4. Experiments on Vortex Nucleation and Lattice Formation
stirrer with a 2-point symmetry, but much smaller than the condensate size could
generate vortices stirring at a frequency of 7 Hz, well below the frequency of 21 Hz
predicted by the Landau criterion for their condensate. This stirrer created 100 and
more vortices very rapidly in 100 ms of rotation. In that case, there were no discrete
frequencies for vortex nucleation, but the number of vortices increased linearly with
the stirring frequency. The number of vortices also increased with stirring times up to
300 ms, but then showed a saturation and even decreased above 300 ms accompanied
by visible heating. They assume that the small stirrer gives rise to a local nucleation
meachnism not governed by the surface modes. The ow eld around the stirrer can
reduce the Landau critical velocity locally, which might lead to the formation of a
pair of vortices of opposite orientation. While the corotating vortex is favoured by the
rotating system, the counterrotating vortex is expelled from the condensate. They also
conducted experiments to determine the optimal radial position for the stirrer to create
vortices, which turned out to be at intermediate radii (between 0.4R
TF
and 0.7R
TF
).
From that they concluded that the surface excitations are not the dominant mechanism
in their experiment and that the thermal cloud is unlikely to play a crucial role in this
process because it has its maximum density outside the condensate radius.
The discrete resonances in the nucleation of vortices for stirring potentials of dier-
ent multipolarities is an excellent experimental conrmation for the role discrete surface
modes play in the process of vortex formation according to Dalfovo and Stringari [130].
This importance has also been conrmed by our simulations. We believe that the same
process governs also the vortex nucleation when the condensate is stirred by a small
laser beam. We disagree with most of the conclusions Raman et al. come to concern-
ing the nucleation process with the small stirrer. Simulations by Caradoc-Davies et al.
[168, 178] have shown that stirring with a small stirrer draws vortices from the region
of low density into the condensate. While pair production of vortices is possible with a
small stirrer, the angular momentum is gained through vortices that are drawn in from
the outside. So, the nucleation of vortices is not a local process, but still governed by
collective excitations. The stirrer simply facilitates the mixing into those modes. The
excitation spectrum might be signicantly altered by the presence of the stirrer, which
could explain a lower critical rotation frequency. We have also carried out simulations
with a small stirrer using the phenomenological growth equation which are not pre-
sented in this thesis. The thermal cloud plays the same role as before. The exchange
of atoms with the thermal cloud is crucial to provide the dissipation necessary to form
a regular vortex lattice.
Abo-Shaeer et al. [80] explored the formation and decay of vortex lattices at nite
135
Chapter 8. Vortex Lattice Formation in Elliptical Rotating Trap
temperature. Here, we only discuss the formation, and postpone the discussion of their
results on the decay of vortex lattices to the next chapter, section 9.3. The condensate
was stirred with an elliptical trap for 200 ms and then allowed to equilibrate in an
axisymmetric trap. As in [79], the formation of a vortex lattice took typically hundreds
of ms after the rotation was stopped. They presume that the absence of equilibration
during the rotation period is due to heating and excitation of collective lattice modes
by the stirring beams. They studied the formation of vortices for temperatures between
337 nK and 442 nK and found that the lattice formation depends very weakly, if at all,
on the temperature, in contrast to the decay rate, which is highly dependent on the
temperature (see section 9.4).
This observation is in contradiction with our simulations using the phenomenological
growth equation. In our model the initiation, equilibration and decay times are depen-
dent on 1/ T. However, in the experiment it is hard to distinguish between dierent
eects. Abo-Shaeer et al. [80] do actually observe more vortices at lower temperatures
in the lattice, but attribute that to the faster decay rate at higher temperatures. Other
experiments seem to see a temperature dependence for the equilibration of a lattice
[152]. So, we believe this result by Abo-Shaeer et al. is not conclusive.
The other point, that they never observe vortex lattices during the rotation phase
can be explained in terms of our model as follows. They report that the thermal cloud
rotates at a rate 2/3 of that of the condensate at the end of the stirring process.
That means there is a highly non-equilibrium situation. But a regular lattice can only
be formed when an equilibrium with the thermal cloud is reached. This equilibration
happens after the rotating drive is switched o, and the resulting lattice rotates at the
same angular velocity as the thermal cloud.
136
Chapter 9
Vortex Lattice Decay
In this chapter, we study the decay of vortex lattices within the model of our phe-
nomenological growth equation. In a treatment given by Zhuravlev et al. [179], the
decay of a vortex lattice has two limiting cases, one where the thermal cloud is nearly
stationary and the other where the thermal cloud is closely following the rotating con-
densate. The rst limiting case is realized when the trap is suciently anisotropic to
stop the rotation of the thermal cloud quickly. The second case corresponds to the
case of a nearly perfect trap, where the friction between the thermal cloud and the
condensate is stronger than that between the thermal cloud and the trap.
In this chapter we only consider the rst case, which is easily implemented in our
equation by setting the rotation of the thermal cloud to zero, = 0. Guery-Odelin
[180] found that very small trap anisotropies are sucient to dissipate the angular
momentum of a rotating thermal cloud on a short time scale, and suggested that this
sensitivity to a residual static anisotropy might have been of importance in the vortex
decay experiments of Madison et al. [73]. The second limiting case of vortex lattice
decay is beyond the scope of our simple phenomenological growth equation because our
model does not treat the dynamics of the thermal cloud.
9.1 General Behaviour
We consider here the process which is the reverse of vortex lattice formation, namely
lattice decay. We take as the initial state a vortex lattice and observe its propagation
in a cylindrically symmetric trap (
x
=
y
= 0) in the presence of a stationary thermal
cloud ( = 0). Fig. 9.1 and Fig. 9.2 show the time evolution of the density for two
dierent initial lattices
1
for a damping parameter of = 0.1.
137
Chapter 9. Vortex Lattice Decay
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 9.1: Density plots of vortex lattice decaying in the presence of a stationary thermal
cloud ( = 0) at times (a) t = 0, (b) t = 0.68, (c) t = 1.08, (d) t = 1.72, (e)
t = 3.16, (f) t = 7.65 with C = 1000,
NC
= 12 and = 0.1. Initial lattice
was created using
x
= 0.05,
y
= 0.15, = 0.65.
As can be seen in Figs. 9.1 and 9.2 (a)-(c), the outer ring of vortices is lost very
rapidly. The outer vortices escape essentially in the radial direction (in the rotating
frame) along the shortest path to the condensate border. As the vortices leave the
condensate, the condensate shrinks in size. There are two eects responsible for that.
First, with the loss of angular momentum the eective centrifugal potential is reduced
which results in a smaller spatial extent of the condensate (see equation (7.29)). At the
same time, the normalization of the wave function decreases, i.e. there is a loss of atoms
in the condensate. This is the opposite of the eect we saw when the condensate grew
around the vortices as the vortices were taken into the high density region described
in section 7.4 during the vortex lattice formation. During the decay, the condensate
recedes when the vortices are released from the dense part into the low density region.
During the decay, the relative rotational frequencies of the vortices are also aected.
In an equilibrium lattice all vortices rotate at the angular velocity of the thermal cloud,
1
Both lattices were obtained dynamically using the phenomenological growth equation in a slightly
elliptical trap with
x
= 0.05 and
y
= 0.15. But this is of no real importance for our consideration
here. We are only interested in showing the decay process for two qualitatively dierent initial lattice
congurations.
138
9.1. General Behaviour
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 9.2: Density plots of vortex lattice decaying in the presence of a stationary thermal
cloud ( = 0) at times (a) t = 0, (b) t = 0.36, (c) t = 1.12, (d) t = 3.32,
(e) t = 9.64, (f) t = 20 with C = 1000,
NC
= 12 and = 0.1. Initial lattice
was created using
x
= 0.05,
y
= 0.15, = 0.75.
but during the decay the inner vortices rotate slower than the outer ones, as can be
seen by comparison of Fig. 9.1 (b) and (c). Consequently, the inner vortices take a
longer path to the edge of the condensate along a spiral. In Fig. 9.1 (d)-(f), there are
two vortices left, that leave the condensate simultaneously, while in Fig. 9.2 (d)-(f)
there is only one. This qualitative dierence is explained by the vortex constellation
in the initial lattice. In Fig. 9.1 (a), we see that the initial lattice has two vortices
symmetric around the centre of the condenate, while the lattice in Fig. 9.2 (a) shows
clearly a single vortex very close to the centre of the condensate. This central vortex
is especially stable and spirals out very slowly.
The corresponding loss of angular momentum for both cases is shown in Fig. 9.3,
where the second row of graphs plot the same information as the rst row, but in a
logarithmic plot. From Figs. 9.3 (a) and (c) we can see that the fast loss of the outer
vortices takes about the same time for both lattices, namely 2. However, the loss
of the two inner vortices in the lattice decay from Fig. 9.1 is concluded at t 6, while
the loss of the single central vortex in the lattice decay from Fig. 9.2 takes to t 14.
This shows, as mentioned, that the central vortex is indeed much more stable.
139
Chapter 9. Vortex Lattice Decay
0 5 10 15
0
2
4
6
8
10
Time []
A
n
g
u
l
a
r

m
o
m
e
n
t
u
m
(a)
Lattice = 0.65
0 5 10 15
10
4
10
2
10
0
Time []
A
n
g
u
l
a
r

m
o
m
e
n
t
u
m
(b)
0 5 10 15
0
2
4
6
8
10
Time []
A
n
g
u
l
a
r

m
o
m
e
n
t
u
m
(c)
Lattice = 0.75
0 5 10 15
10
4
10
2
10
0
Time []
A
n
g
u
l
a
r

m
o
m
e
n
t
u
m
(d)
Figure 9.3: Loss of angular momentum during vortex lattice decay. (a) and (b) correspond
to the lattice decay shown in Fig. 9.1, (c) and (d) correspond to the lattice decay
shown in Fig. 9.2.
From the logarithmic plots in Fig. 9.3 (b) and (d), it is particularly apparent that
the decay is not a simple exponential loss of angular momentum, but happens in stages.
Basically, one ring of vortices is lost at a time.
In order to quantify the loss rate, we need to assign a decay rate to the process.
However, since the decay time does not have a simple exponential dependence, there
is no single decay constant that would describe the whole process. For the = 0.65
lattice, we have chosen to measure the decay time as the time it takes for the angular
momentum to reach approximately zero in linear graphs, as in Fig. 9.3 (a). In the
case of = 0.75 lattice, however, this decay time would be very long because of the
remaining central vortex, which decays only slowly. Thus, we have opted to measure
the decay time in this case as the time it takes to lose all the vortices, but the last
140
9.2. Angular Velocity and Radial Velocity
0 20 40 60 80 100
0
50
100
150
200
Inv. damping parameter 1/
D
e
c
a
y

t
i
m
e
= 0.65
= 0.75
Figure 9.4: Vortex lattice decay times in dependence on damping parameter .
one. This can be estimated from the angular momentum graphs, such as the one in
Fig. 9.3 (c), as the point were the angular momentum attens out after the initial
fast loss of angular momentum. The results are given in Fig. 9.4, where we plot the
decay time dependence on the damping parameter . Note that because of the diering
denitions of the decay times in the two cases, the two curves in Fig. 9.4 cannot directly
be compared to each other. However, both show roughly a 1/ dependence, and the
decay times are of the same order of magnitude as the initiation times of Fig. 7.13.
9.2 Angular Velocity and Radial Velocity
First, we consider the loss of the single vortex from Fig. 9.2 (d)-(f) more closely. Fig. 9.5
(a) and (b) show the angle and radial position of the single vortex versus time, re-
spectively. Because of the limited spatial resolution of the density images from the
simulation, there is an uncertainty in the exact position of the vortex as discussed in
section 7.5.4. In order to reduce the noise in the calculation of the velocities, we rst
t a polynomial of fth order to the position curves and obtain the angular and radial
velocities from the ts analytically. They are shown in Fig. 9.5 (c) and (d). In Fig. 9.5
(e) and (f) we plot the angular and radial velocities versus the radial position from
Fig. 9.5 (a) and (b). This allows us to compare the angular velocity from Fig. 9.5
(e) to results numerically obtained by Jackson et al. [122] shown in Fig. 9.6 which is
the precession frequency of an o-centred vortex as a function of its radial position
141
Chapter 9. Vortex Lattice Decay
0 5 10 15 20
0
2
4
6
8
10
12
Time []
A
n
g
l
e

(a)
Simulation
Fit
0 5 10 15 20
0.1
0.2
0.3
0.4
0.5
0.6
Time []
A
n
g
u
l
a
r

v
e
l
o
c
i
t
y
(c)
0 5 10 15 20
0
2
4
6
8
10
12
Time []
R
a
d
i
a
l

p
o
s
i
t
i
o
n
(b)
Simulation
Fit
0 5 10 15 20
0
0.2
0.4
0.6
0.8
1
Time []
R
a
d
i
a
l

v
e
l
o
c
i
t
y
(d)
0 2 4 6 8 10 12
0
0.1
0.2
0.3
0.4
0.5
0.6
Radial position
A
n
g
u
l
a
r

v
e
l
o
c
i
t
y
(e)
0 2 4 6 8 10 12
0
0.2
0.4
0.6
0.8
1
Radial position
R
a
d
i
a
l

v
e
l
o
c
i
t
y
(f)
Figure 9.5: Precession of single vortex from simulation in Fig. 9.2. (a) Polar angle and
(b) radial position of vortex versus time. (c) Angular velocity and (d) radial
velocity versus time. (e) Angular velocity and (d) radial velocity versus radial
position.
r
0
in a pure two-dimensional (2D) GrossPitaevskii equation (GPE) simulation with
C = 1000 (i.e. the same value as used in our simulations). In that case there is no dis-
142
9.2. Angular Velocity and Radial Velocity
Figure 9.6: Angular velocity as function of vortex position in pure 2D GPE simulation by
Jackson et al. [122] for C = 1000.
sipation, and the vortex precesses at a constant radius. Qualitatively, our results show
the same behaviour as Jacksons [122]. The angular velocity increases with increasing
radial position. Our results are slightly smaller because the dissipation provided by the
phenomenological growth equation slows the vortex down. Jacksons results approach a
constant value with decreasing radius, which is in good agreement with the expression
[[ =
3
r
4
ln
_
R
TF

_
(9.1)
obtained by Svidzinsky and Fetter [118] for a small o-set of the vortex using a time-
dependent variational analysis. Our result in Fig. 9.5 (e), however, shows a steep
increase for a small radial o-set, and as we now discuss, a possible reason for this
is that the precession frequency at this early stage in the vortex lattice decay is still
inuenced by other vortices.
Let us, for example, consider the angular velocity and radial velocity for one of the
two inner vortices from the decay of the = 0.65 lattice in Fig. 9.1. This is shown
in Fig. 9.7 versus radial position starting from the initial lattice position at t = 0.
We can see that this is very dierent from the single vortex case in Fig. 9.5 (e) and
(f). Initially, both the angular and radial velocities decrease while the radial position
increases. This occurs during the rst stage of the lattice decay when the outer vortices
quickly leave the condensate. The large decrease of angular momentum leads to a slow-
down of the vortex velocity, but once the outer vortices have disappeared, the velocity
143
Chapter 9. Vortex Lattice Decay
0 2 4 6 8 10 12
0
0.1
0.2
0.3
0.4
0.5
0.6
Radial position
A
n
g
u
l
a
r

v
e
l
o
c
i
t
y
(a)
0 2 4 6 8 10 12
0
0.2
0.4
0.6
0.8
1
Radial position
R
a
d
i
a
l

v
e
l
o
c
i
t
y
(b)
Figure 9.7: Precession of double vortex from simulation in Fig. 9.1. (a) Angular velocity
and (b) radial velocity as function of vortex position.
increases again with radial position as in the single vortex case from Fig. 9.5. We note
that the increase in velocity is not linear with radial position as in the single vortex
case. If we compare the angular velocity of Fig. 9.7 (a) with Fig. 9.5 (e) for radial
positions r > 6, we note that the angular velocity in the double vortex case is rst
faster until about r 8, and then slower for r 8 than in the single vortex case. A
comparison of Fig. 9.7 (b) with Fig. 9.5 (f), on the other hand, reveals that the radial
velocity in the double vortex case is consistently greater than in the single vortex case
for r > 6. We conclude that the presence of other vortices can signicantly alter the
angular frequency and radial velocity of a particular vortex. The reason is that the
precession of an o-centre vortex can be understood as a collective excitation of the
condensate and, thus, depends on the state of the system.
Now, we summarize a simple model for the dissipative dynamics of a vortex state
by Fedichev and Shlyapnikov [181] that is based on the two-uid model commonly used
in the superuid Helium literature [110]. A vortex moving in a superuid of density
s
experiences a force given by Magnus law
F =
s
[(v
L
v
S
) ], (9.2)
where v
L
is the velocity of the vortex line, and v
S
is the velocity of the superuid at
the position of the vortex line. The vector is parallel to the vortex line and equal
to the circulation carried by the vortex. The force F originates from mutual friction
between the normal component (i.e. the thermal cloud) and the vortex line. The major
source of this friction is the collision of rotons with the cores of vortex lines [110]. These
144
9.2. Angular Velocity and Radial Velocity
elementary excitations are scattered by the vortex lines. Assuming a small friction in
equation (9.2), it can be linearized as

s
[(v
L
v
S
) ] = Dv
S
D

[v
S
]/ (9.3)
in terms of the dissipation coecients D and D

that represent longitudinal and trans-


verse friction, respectively.
For a single vortex at position r
0
in a condensate of radius R
TF
, where R
TF
is the
ThomasFermi radius, the superuid velocity is given by
v
S
=
r
0
R
2
TF
r
2
0
. (9.4)
The velocity of the vortex line, on the other hand, can be split into a radial component
v
(r)
L
= Dv
S
/
s
(9.5)
and tangential component
v
()
L
= (1 D

)v
S
/
s
. (9.6)
The transverse friction (the Iordanskii force) slows the drift precession frequency of
the vortex down depending on the transverse friction coecient, which is given by the
universal expression D

=
n
. The radial velocity is governed by the longitudinal
friction coecient. For very low temperatures the longitudinal force is extremely weak.
For higher temperatures T , Fedichev and Shlyapnikov [181] nd
D n
0
, (9.7)
where n
0
is the condensate density and a small parameter of nite-temperature per-
turbation theory
= (n
0
a
3
)
1/2
T

. (9.8)
Here, a is the scattering length of the atoms. For a large condensate, the dependence
on the condensate density is well approximated using the maxium condensate density.
Then, the longitudinal friction coecient is simply proportional to the temperature.
Therefore, we expect the typical decay time in our simulations to be proportional to
1/ T, which is conrmed by the result in Fig. 9.4. The same principal mechanism
is discussed by Zhuravlev et al. [179] for the decay of vortex lattices.
145
Chapter 9. Vortex Lattice Decay
9.3 Single Vortex
As discussed in section 7.5.4 the growth process with a xed angular velocity for the
thermal cloud cannot produce a single central vortex. If the angular velocity is above
the critical velocity for vortex formation given by the Landau criterion, we always
obtain a lattice with more than one vortex. If the angular velocity is below the critical
rotation frequency, no vortex can overcome the energy barrier for vortex penetration,
even if the angular velocity is higher than the thermodynamic critical velocity so that
a vortex state would be a global energy minimum.
However, if is reduced after some vortices entered the condensate, it is possible
to create a single central vortex as illustrated in Fig. 9.8. In this case, the lattice is
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 9.8: Density plots of partial decay of vortex lattice to a single vortex for times (a)
t = 0, (b) t = 10.2, (c) t = 19.08, (d) t = 21.32, (e) t = 31.84, (f) t = 100.
The initial lattice was created using C = 1000, = 0,
NC
= 12, = 0.1 with a
thermal cloud rotating at = 0.45. At time t = 0 the rotation of the cloud is
set to = 0.27.
created in a cylindrical trap with the rotation of the cloud at = 0.45, just above
the critical frequency for vortex nucleation for C = 1000,
NC
= 12. The resulting
vortex lattice shown in Fig. 9.8 (a) has three vortices. When the angular velocity of
the thermal cloud is reduced to = 0.27, well below the critical frequency for vortex
nucleation, the lattice expands uniformly to begin with (Fig. 9.8 (b)). But a slight
146
9.3. Single Vortex
assymmetry brings one of the vortices closer to the condensate edge (Fig. 9.8 (c) and
(d)). This vortex is eventually lost in Fig. 9.8 (e) and two vortices remain. The one
closer to the edge spirals out of the condensate while the other one spirals inwards. At
last, we are left with a single central vortex shown in Fig. 9.8 (f).
Repeating the same simulation, but with set to = 0.25, leads to the loss of
all three vortices. This does not necessarily mean that a vortex cannot be supported
at that rotation frequency. For example, starting the simulation with the lattice from
Fig. 9.2 (a) leads to a single central vortex for = 0.25. This is the case even though
the initial lattice has many more vortices, but one of them is close to the centre. As we
know from section 9.1 this central vortex is particularly stable. It might even survive
for angular velocities below the thermodynamic critical velocity as long as a central
vortex is a local energy minimum. This is the case when
m
< <
v
, where
m
is
the angular velocity for a metastable central vortex, and
v
the thermodynamic critical
angular velocity as discussed in section 4.5.
Let us summarize our view on the occurence of a single vortex state in experiments.
To create any vortex in a condensate the rotation of the thermal cloud has to exceed
the critical rotation frequency given by the Landau criterion. In that case, a number
of vortices (more than one) enters the dense region of the condensate. During the
equilibration period, however, the thermal cloud loses some angular momentum due to
small residual trap asymmetries and mutual friction between the thermal cloud and the
vortices, just as in the process of vortex lattice decay. This, in turn, reduces the angular
momentum of the condensate. Thus, some vortices are shed, and an equilibrium state
with a single central vortex might be reached. This is particularly likely if during the
equilibration period one of the vortices has come closer to the centre of the condensate
than the others.
This process would explain the observation of a vortex state below the Landau criti-
cal frequency as reported by Haljan et al. [78]. They state that for their system with 3 to
810
5 87
Rb atoms in a cylindrically symmetric trap with

,
z
= 28.35, 5.45 Hz,
the thermodynamical critical frequency is 0.2 to 0.25

and the Landau criterion 0.4

.
They determine that the threshold for the detection of a vortex lies somewhere between
0.32 and 0.38

. They start the evaporation process with a rotating thermal cloud in


an axially symmetric trap. After the evaporation process, which helps to spin up the
cloud by selectively removing atoms with low angular momentum, they allow the sys-
tem a time of 5 s ( 42 trap periods) to rethermalize before they take a non-destructive
image. From the aspect ratio of the condensate the rotation rate is determined. Eec-
tively, they measure the rotation rate of the nal state, but not the maximum rotation
147
Chapter 9. Vortex Lattice Decay
of the system during the rethermalization. Thus, it is conceivable that the rotation of
the thermal cloud had exceeded the critical frequency given by the Landau criterion,
but that the system slowed down during the rethermalization period.
This would be consistent with our model that actually requires a reduction of the
rotation rate to obtain a single central vortex. Within this model, it is not surprising
to nd a vortex at a rotation rate lower than the Landau critical velocity, as long as
the rotation of the system remains higher than the thermodynamic critical frequency,
as is the case in the experiment.
9.4 Experiments on Vortex Lattice Decay
There are only a few experiments on the decay of vortex lattices. The ENS group
considered the lifetime of a single vortex [73] and the decay of small vortex lattices
[123], and at MIT Abo-Shaeer et al. [80] explored the decay of large vortex lattices at
nite temperature.
For the single vortex state Madison et al. [73] found a characteristic lifetime of
400 to 1000 ms for 2.3 10
5
and 1.2 10
5 87
Rb atoms in the condensate (

,
z
=
2219, 11.7 Hz), respectively. This suggests that the temperature was higher in the
former case, but they could not conrm this assumption because the temperature was
in both cases below their detection limit of 80 nK. By plotting the fraction of images
showing a vortex as a function of time spent by the gas in the axisymmetric trap after
the end of the stirring phase, they found a non-exponential decay curve. In particular,
they note that for long times the vortex was rarely centred. This is in agreement with
our simulations where a single central vortex slowly spirals out of the condensate while
the angular momentum of the condensate is lost to the non-rotating thermal cloud. In
the experiment, they measured a residual trap anisotropy of
x
/
y
= 1.012 0.002,
which is probably enough to quickly stop the rotation of the thermal cloud [180].
In [123] Madison et al. examined the decay of a vortex lattice with ve vortices in
the same manner. By counting the remaining number of vortices for dierent evolution
times in an axisymmetric trap after the end of stirring, they obtained a characteristic
time for the reduction of the vortex number by a factor of two of 750 ms. This is
comparable to the decay time of the single vortex. The decay was non-exponential,
and they found that the vortex pattern decays by losing one vortex at a time. The
remaining vortex pattern readjusts itself shortly after the loss into a new regular lattice.
The observed loss of one vortex at a time is in contrast to our simulations, where all
148
9.4. Experiments on Vortex Lattice Decay
vortices in a ring that have roughly the same distance from the condensate centre
leave the condensate simultaneously. In those simulations, we set the rotation of the
thermal cloud suddenly to zero. However, when we decrease the angular velocity of
the thermal cloud gradually, the loss of vortices happens one-by-one. Essentially, the
lattice adjusts to the new environment by shedding as many vortices as necessary to
t the rotation rate of the thermal cloud, while the remaining vortices still form a
regular lattice. Hence, we conclude that in the experiment the decrease in the rotation
rate of the thermal cloud must happen over an extented period of time which is not
much shorter than the decay time of the lattice. As the thermal cloud slows down,
the vortices are lost from the lattice accordingly. However, in the limiting case of a
co-rotating thermal cloud, the decay is predicted to be exponential [179] whereas the
decay is non-exponential for a xed thermal cloud. Possibly, the decay of such a small
lattice takes place in an intermediate regime between these two limiting cases.
Abo-Shaeer et al. [80] examined the temperature dependence of the decay of large
vortex lattices with about 130 vortices. They use large condensates of
23
Na with up to
7.5 10
7
atoms in a trap with

,
z
= 286.1, 21.1 Hz. They infer the number
of vortices non-destructively from the rotation rate of the condensate and observe an
exponential decay. This is in agreement with the theoretical study by Zhuravlev et al.
[179] in the limiting case of vortex lattice decay where the thermal cloud is co-rotating
with the condensate. The decay constant is strongly temperature dependent, and varies
between 0.3 s
1
for low temperatures of T 0.35T
c
and 3 s
1
for high temperatures
around T 0.58T
c
. As mentioned before, we did not consider vortex lattice decay in
the case of a co-rotating thermal cloud because our phenomenological growth equation
does not include a model for the dynamics of the thermal cloud. Nevertheless, in the
case of a xed thermal cloud we also found a strong dependence of the decay rate on
1/ T with shorter decay times for higher temperatures (Fig. 9.4).
Finally, we mention briey the experiment on vortex nucleation by Hodby et al.
[152] who work with
87
Rb in a trap with

,
z
= 226, 175 Hz. Although they
do not study explicitly the decay of vortex lattices they remark that at high tempera-
tures of 0.8T
c
vortices were only occasionally found in equilibrium congurations and
400 ms after formation had already moved to the edge of the condensate before
disappearing. On the other hand, at lower temperatures around 0.5T
c
the lifetime of
vortices was around 4 s, limited only by the decay of the condensate itself. This is
further evidence that the thermal cloud plays a very crucial role in the decay of vortices.
In our model, this is the case because the condensate loses angular momentum through
the exchange of atoms with a thermal cloud that rotates at a lower angular velocity
149
Chapter 9. Vortex Lattice Decay
than the vortex lattice.
150
Chapter 10
Non-equilibrium Dynamics and
Lattice Excitations
So far, we have only considered examples where the trap was either cylindrically sym-
metric, which means that only the angular velocity of the rotating thermal cloud
is relevant, or examples where the thermal cloud was rotating at the same speed as
the trap = . In this chapter, we consider simulations with ,= . This allows us
to simulate non-equilibrium situtations that show a variety of interesting eects. In
reality, such a situation is of course not maintainable indenitely. Eventually, some
form of equilibrium must be obtained. However, the spin-down time of a thermal cloud
is suprisingly short [180] while a condensate can maintain its angular momentum for a
relatively long period of time [80, 179]. Therefore, it is possible to carry out experiments
in highly non-equilibrium situations.
10.1 Simple Non-Equilibrium Lattice
In Fig. 10.1 we consider a simple example. The trap is elliptical with
x
= 0.05,

y
= 0.15 and rotates at = 0.65, while the thermal cloud is rotating at = 0.5. The
other parameters are the same as before with C = 1000,
NC
= 12 and = 0.1. As
in most of our simulations, the initial state was the ground state of the cylindrically
symmetric trap. In this simulation the initiation time takes around 100, which is very
long because of the small . The rst density plot in Fig. 10.1 (a) is the resulting state
after a propagation time of t = 140, but for the purpose of the discussion here we label
it with t = 0, because we are only interested in the dynamics of the non-equilibrium
lattice. The solid line indicates the long axis of the rotating elliptical potential, and the
151
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
5 0 5
5
0
5
Position x
(e)
5 0 5
Position x
(f)
5 0 5
10
4
10
3
10
2
10
1
Figure 10.1: Density plots of non-equilibrium lattice for times (a) t = 0, (b) t = 0.76,
(c) t = 1.52, (d) t = 2.28, (e) t = 3.04, (f) t = 3.8. The solid line
indicates the long axis of the elliptical trap with
x
= 0.05,
y
= 0.15 rotating
at = 0.65. The dashed line indicates the rotation of the thermal cloud at
= 0.5.
dashed line the rotation of the thermal cloud. Both rotations are counter-clockwise.
In Fig. 10.1 (a) the two lines coincide, and the vortex that is the furthest on the left
hand side is initially very close to the lines. Following this vortex in Fig. 10.1 (b)-(f),
we see that it stays close to the dashed line, not to the solid line. That means the
vortex lattice is rotating with the angular velocity of the thermal cloud, not with the
trap. The overall shape of the condensate, on the other hand, is slightly elliptical and
follows the trap rotation. This constant deformation of the condensate with respect to
the lattice leads to a permanent adjustment of the lattice structure.
We can understand this behaviour in the context of the two-uid model summarized
in section 9.2. Without any thermal cloud, the force on the vortex line in Magnus law
(9.2) is F = 0. In that case, the vortex line moves with the velocity of the superuid.
In the presence of a thermal cloud, on the other hand, the frictional force on the vortex
line depends on the relative velocity of the vortex line to the normal component. This
will cause the vortices to rotate at the angular velocity of the thermal cloud. In the
limit of a small frictional force, Magnus law reduces to equation (9.3), which describes
152
10.2. Deformation of Rapidly Rotating Condensates
a friction-induced relative velocity of the vortex to the superuid. This small force
yields to the constant adjustment of the vortex lattice relative to the shape of the
condensate.
Fig. 10.2 shows the local chemical potential of the initial lattice of Fig. 10.1 (a). We
see that the chemical potential is not uniform across the condensate, but has two ap-
parent features. On the one hand, it has an overall prole reminiscent of the hyperbolic
11.5
12
12.5
Position x
P
o
s
i
t
i
o
n

y
5 0 5
5
0
5
Figure 10.2: Local chemical potential of non-equilibrium lattice shown in Fig. 10.1 (a).
phase prole of a vortex free state in a rotating elliptical trap shown in Fig. 8.8. This
hyperbolic prole is rotating with the trap. Usually, a rotating lattice state mimics
solid-body rotation, but in this case, the lattice rotates at a dierent angular veloc-
ity, and hence, there is an additional irrotational ow pattern necessary to adjust the
condensate shape permanently to the rotating trap. On the other hand, the local chem-
ical potential shows a gradient at the position of the vortices. This indicates that the
vortices are constantly moving to readjust their positions to the condensate density
that has changed relative to their frame of rotation. Quite clearly, the state is not a
stationary state, neither in the frame rotating at , nor in the frame rotating at .
10.2 Deformation of Rapidly Rotating Condensates
Engels et al. [81] have studied the non-equilibrium dynamics of large, rapidly rotating
condensates. With a special evaporation technique they achieve vortex lattices in al-
most pure
87
Rb condensates with about 10
6
atoms that contain 130 or more vortices
153
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
and rotate at = 0.95, close to the centrifugal limit
1
. A typical initial lattice is shown
in Fig. 10.3 in a trap with

,
z
= 28.35, 5.45 Hz. They then distort the lat-
tice by deforming the trapping potential from an almost axially symmetric form to an
elliptical shape in the x-y plane. This ellipse can also be rotated around the z-axis.
Figure 10.3: Typical initial lattice in experiments by Engels et al. [81]. (a) Expansion
picture along the axis of rotation. (b) One pixel wide cross section along the
white line in (a).
10.2.1 Frequency splitting of m = 2 modes
As mentioned several times in this thesis, an ellipsoidal trap deformation excites mainly
the quadrupole mode. For a condensate in the vortex-free ground state the l = 2,
m = 2 modes are degenerate and have the excitation frequency =

2 = 1.414 in
the hydrodynamic limit [107]. Zambelli and Stringari [177] have shown with a sum-rule
approach that in the presence of a quantized vortex, the frequencies of these two modes
split according to

=
2
m
L
z
)
r
2

)
(10.1)
with r
2

= x
2
+ y
2
. This frequency split has been successfully used to measure the
angular momentum of condensates [74].
In the limit of solid-body rotation for rotating condensates with vortex lattices, this
1
As we have seen in section 10.1, the lattice rotates with the angular velocity of the thermal cloud.
Thus, in all the cases of this section, refers to both the angular velocity of the lattice and of the
thermal cloud.
154
10.2. Deformation of Rapidly Rotating Condensates
frequency split can be written as

= 2, (10.2)
where is the angular velocity of the lattice. The m = +2 mode is co-rotating with
the lattice, while the m = 2 mode is counter-rotating. This prediction has been
experimentally conrmed by Haljan et al. [78] in very good agreement with the theory.
In particular, in the limit of 1, the frequency
+
2 and

0. That means
that the shape deformations induced by the m = +2 mode are rotating at
+
/[m[ = 1,
while the shape deformations of the m = 2 mode are stationary in the laboratory
frame.
10.2.2 Excitation of m = 2 mode
The angular velocity = 0.95 is suciently close to the centrifugal limit, so that the
m = 2 mode can conveniently be excited by a stationary elliptical trap deformation.
When this is done in the experiments with an ellipticity of = 3.6% in the x-y plane,
lattice deformations are observed as shown in Fig. 10.4 and most remarkably sheet-like
structures as shown in Fig. 10.5.
Figure 10.4: Change of lattice structure in experiment [81]. (a) Hexagonal structure in an
undisturbed lattice. (b) Near orthorhombic structure seen transiently during
lattice evolution in the presence of an m = 2 quadrupolar surface mode.
We have simulated these experiments in two dimensions. The initial lattice depicted
in Fig. 10.6 (a) was created dynamically in a cylindrically symmetric trap using the
phenomenological growth equation with the parameters C = 1000, = 0.95,
NC
=
6, = 0.1. This lattice with about 80 vortices is a little smaller than the initial
155
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
Figure 10.5: Sheet like structure as seen during lattice evolution in experiment [81] during
presence of an m = 2 quadrupolar surface mode.
P
o
s
i
t
i
o
n

y
(a)
10
5
0
5
10
(b) (c)
Position x
P
o
s
i
t
i
o
n

y
(d)
10 5 0 5 10
10
5
0
5
10
Position x
(e)
10 5 0 5 10
Position x
(f)
10 5 0 5 10
10
4
10
3
10
2
10
1
Figure 10.6: Density plots in two dimensional simulation of lattice structure while the el-
lipticity of the trap is ramped from 0 to = 0.06 over a time period of 6. (a)
t = 0, (b) t = 1.92, (c) t = 2.08, (d) t = 3.12, (e) t = 4.32, (f) t = 7.28.
The damping parameter was = 0.01.
156
10.2. Deformation of Rapidly Rotating Condensates
lattice from the experiment which has about 130 vortices (see Fig. 10.3). To match the
experiment better we could have chosen larger C and
NC
values. However, that would
require a larger numerical grid with more points and possibly a smaller time step. For
convenience, we decided to use the smaller lattice and look only for the qualitative
features of the experiment.
The initial lattice (Fig. 10.6 (a)) is disturbed by a deformation of the trap with
ellipticity = 0.06 that is stationary in the laboratory frame ( = 0) with the long axis
aligned in x-direction. This deformation is switched on over a time period of t = 6,
while the rotation of the thermal cloud is kept at = 0.95. As in the experiment, the
condensate ellipticity increases signicantly (Fig. 10.6 (b)-(d)). The long axis of the
condensate is tilted by about 45

relative to the long axis of the elliptical trap. The


lattice structure must continuously change to accomodate the elliptical shape of the
condensate as seen in Fig. 10.6. In particular, we can identify an almost orthorhombic
structure in Fig. 10.6 (b), very similar to the one observed experimentally (Fig. 10.4
(b)), and also stripe formation in Fig. 10.6 (c), which occurs periodically with a period of
1/6. The stripes become more pronounced as the condensates aspect ratio increases
(Fig. 10.6 (d)). The stripes form whenever a lattice vector lies along the smaller axis
of the condensate. While the elliptical deformation is xed in the laboratory frame,
the lattice continues to rotate at = 0.95. The periodicity of the stripe occurence
can therefore be understood in terms of the six-fold symmetry of the vortex lattice
[81, 182]. Eventually, the elliptically deformed condensate wave function disintegrates
as is apparent in Fig. 10.6 (e), before it recovers a more circular shape in Fig. 10.6 (f)
with a lattice-like structure that rotates at the angular velocity of the thermal cloud.
However, since the discrepancy between the stationary trap and the rotating thermal
cloud remains indenitely in our simulations, this cannot be an equilibrium state.
One signicant dierence to the experiment is, however, that the visibility of the
stripes is much higher in the experiment than in our simulations. In the experiment,
there is hardly any trace of condensate density in the stripes as can be seen from
Fig. 10.3 (a) and (b). In our simulation, on the other hand, the stripes are formed by
neighbouring vortices that moved very close together due to the lattice deformation,
but there is still a perceptible condensate fraction in between the vortices as seen in
Fig. 10.6 (d). Furthermore, a phase plot in Fig. 10.7, corresponding to Fig. 10.6 (d),
clearly reveals that the stripes consist of distinct individual vortices. Contrary to the
interpretation by Haljan et al. [81], the vortices do not merge into one single structure
with a combined phase rotation around, but keep their individuality with each of them
showing a phase circulation of 2.
157
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
Position x
P
o
s
i
t
i
o
n

y
10 5 0 5 10
10
5
0
5
10


0

Figure 10.7: Phase of condensate wave function corresponding to Fig. 10.6 (d).
This is the case even though in our simulation the ellipticity of = 0.06 is almost
twice as large as in the experiment with = 0.036. When we use = 0.03, the stripes are
less pronounced because the aspect ratio of the condensate does not increase as much,
i.e. the condensate is wider across its shortest axis. Therefore, the vortices do not move
that close together. One reason that the contrast in the experimental pictures of the
stripe formation is that much higher than in our simulations might be that they use
time of ight expansion before imaging. The separation between the vortices increases,
and the residual low density might be undetectable. Another possibility is that in their
experiment the vortices move closer together because they have more vortices in the
initial lattice to start with.
The stripe formation is only a transient phenomenon that ceases after abou 2.5 trap
periods in our simulation. For the experiment, it is reported [81] that vortex lattices
can be observed up to 400 ms, which corresponds to about 3.3 trap periods. We have
checked whether a change in the damping parameter would have any eect. It turns
out that the time before the condensate starts to disintegrate is independent of .
However, the stripe formation is slightly less pronounced if the damping is increased to
= 0.1, because the wave function tends more quickly to a state as in Fig. 10.6 (f). We
also performed simulations using the pure GrossPitaevskii equation ( = 0) and still
observed the stripe formation. The main dierence is that the condensate disintegrates
even further than shown in Fig. 10.6 (e), and does not recover at all. In any case, we
conclude that the observation of stripes does not require dissipation.
The duration for which we observe stripe formation is also independent of the time
period over which the ellipticity is switched on. When the ellipticity is applied instantly,
158
10.2. Deformation of Rapidly Rotating Condensates
as done in the experiment [81], the stripes are again less pronounced because the
elliptical deformation of the condensate is smaller. The transition to the lattice like
state of Fig. 10.6 (f) is smoother. For a very long ramp time of the trap ellipticity of
t = 30, the condensate does not return to a circular shape, but keeps expanding
indenitely along the long axis of the trap.
Mueller and Ho [182] have analysed the experiment analytically using a variational
form of the initial vortex structure that is valid in the high angular momentum limit
of the mean-eld quantum Hall regime. Their model recovers all the main features
observed in the experiment, which we also see in our simulations. They come to the
conclusion that the dynamics is solely governed by the kinetic energy whereby the
interactions only perturbatively enter the dynamics. The vortices simply follow the
quadrupolar ow. In particular, the stripe formation is caused by the periodic defor-
mation of the lattice and not by melting of vortices into sheets, which is in agreement
with our own view.
10.2.3 Excitation of m = 2 mode
Engels et al. [81] also carried out experiments where they excited the l = 2, m = 2 mode
by rapidly switching on an elliptical deformation of the magnetic trap
2
and rotating
that ellipse around in the x-y plane. They state that the resonance of that mode lies
only 0.4 Hz above the rotation of the lattice for their rotation rate of = 0.95 and trap
parameters of

,
z
= 28.35, 5.45 Hz. They observe an increase of the ellipticity
of the condensate of 50% when driving the condensate about 1 Hz from resonance,
and approaching 100% when driving on resonance. In the latter case, they observe a
gradual increase of disorder in the lattice, and after 400 ms, which corresponds to about
3.3 trap periods, they nd a complete loss of long-range order in the vortex structure
as shown in Fig. 10.8.
In our simulations, we never saw this kind of complete loss of order in the vortex
lattice structure. Some of our results are shown in Fig. 10.9. The initial vortex lattice
was created with C = 1000 and
NC
= 12 rotating at = 0.9. At t = 0 the elliptical
trap is switched on with ellipticity = 0.06 rotating at = 0.84. For the case in
Fig. 10.9 (a) the ellipticity is linearly ramped on over a time period of t = 1, in
Fig. 10.9 (b) over t = 3 and in Fig. 10.9 (c) over t = 6. The elliptical deformation
of the condensate shape increases as the ramp time increases. For even longer ramp
2
Actually, the paper does not explicitly mention for the m = 2 case whether the trap is switched on
abruptly. We infer this from the statement that the trap is jumped from a round trap to an elliptical
for the excitation of the m = 2 mode.
159
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
Figure 10.8: Irregular assembly of vortex cores as observed in experiment [81] after 500 ms
in the presence of an m = 2 quadrupolar surface mode.
Position x
P
o
s
i
t
i
o
n

y
(a)
12 6 0 6 12
12
6
0
6
12
Position x
P
o
s
i
t
i
o
n

y
(b)
12 6 0 6 12
Position x
P
o
s
i
t
i
o
n

y
(c)
12 6 0 6 12
10
4
10
3
10
2
10
1
Figure 10.9: Lattice structure at t = 5.12 after distortion of the lattice rotating at = 0.9
by linearly ramping on an elliptical trap rotating at = 0.84 over a time
period of (a) t = 1, (b) t = 3, (c) t = 6. The ellipticity of the trap is
= 0.06. Other parameters C = 1000,
NC
= 12 and = 0.01.
times, the condensate keeps expanding indenitely just as in the case of the m = 2
excitation described in section 10.2.2. For all the simulations of Fig. 10.9 the damping
parameter was chosen small with = 0.01. When the damping parameter is increased
the deformation of the condensate is smaller.
In all cases of Fig. 10.9, we can see that the vortices do not show a regular lattice
structure, either hexagonal, or orthorhombic, but they still resemble some kind of
lattice, similar to a lattice with defects. However, we can see that the disorder of the
lattice structure increases going from Fig. 10.9 (a) to Fig. 10.9 (c) as the aspect ratio
of the condensate increases.
For the experiment, very large ellipticities of the condensate are reported, if the
m = 2 mode is excited resonantly. Unfortunately, we were not able to go much beyond
160
10.3. Excitations on Vortex Lattices
the deformation of the condensate shown in Fig. 10.9 (c) for numerical reasons
3
. This
also prohibited us from increasing the rotation speed of the elliptical trap above the
rotation speed of the lattice > , where the resonance of the m = 2 mode should
lie. A simulation with = , which is closer to the resonance than the value = 0.84
used for the results in Fig. 10.9, does not show any more disorder. Presumably, the
lattice is not disturbed as much if the elliptical trap rotates at the rotation rate of the
lattice although the drive is closer to the m = 2 resonance.
In summary, while there is some loss of order in the lattice structure for our simu-
lations, we see nothing as dramatic as the experimental result from Fig. 10.8, possibly
because we do not drive the m = 2 mode on resonance. Also, some other mechanisms
might be involved such as residual oscillations in the rotating trap potential, but we
have not tested this assumption.
10.3 Excitations on Vortex Lattices
The basis state method described in appendix A.3 enables us to calculate excitations on
vortex lattices. We take the lattice states from our dynamical simulations and obtain
the appropriate excitations by solving the Bogoliubovde Gennes (BdG) equations (8.4)
in the frame rotating at the angular velocity of the lattice = .
10.3.1 Excitations on single vortex
As an introductory example, we consider the excitations on a single central vortex.
Because a central vortex is an eigenstate in the stationary laboratory frame and any
rotating frame, this enables us to directly compare the eect of a rotating frame on the
excitation energies.
Fig. 10.10 (a) shows the excitation spectrum on a single central vortex in a non-
rotating frame ( = 0). The excitation energies are plotted against their angular
momentum dened in chapter 8 by equations (8.7) and (8.8) relative to the angular
momentum of the condensate wave function. Because the central vortex is an eigen-
state of the angular momentum operator

L
z
and the trapping potential is cylindrically
symmetric, the problem is in principle separable in polar coordinates. Hence, the quasi-
particle amplitudes are also eigenstates of the angular momentum operator with integer
3
In Fig. 10.9 (c), the condensate seems already to extend beyond the borders of the grid. The
simulation was, however, carried out on a grid with twice the spatial extent in both x- and y-direction.
161
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
20 10 0 10 20 30
0
2
4
6
8
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

(a)
20 10 0 10 20 30
0
2
4
6
8
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
(b)
Figure 10.10: Spectrum of single vortex state (a) in stationary laboratory frame ( = 0)
and (b) in frame rotating at = 0.3. The vortex state is calculated for
C = 1000, and its eigenvalue is = 12.84.
angular momentum. We will use this fact at the end of this section to see how accurate
our solutions are.
The so-called anomalous mode, which is the lowest mode in the spectrum of Fig. 10.10
(a), has the negative energy = 0.134 and q = 1. It is associated with the preces-
sion of an o-centred vortex [183]. The fact that its energy is negative is a sign that the
vortex is not actually a stable state in the non-rotating frame (see also section 4.5.4).
The density prole of the anomalous mode is shown in Fig. 10.11. The quasi-particle
10
0
10
10
0
10
0
0.5
1
1.5
Position x
Position y
(a)
A
m
p
l
i
t
u
d
e

|
u
|
2
10
0
10
10
0
10
0
0.05
0.1
0.15
Position x
Position y
(b)
A
m
p
l
i
t
u
d
e

|
v
|
2
Figure 10.11: Anomalous mode on single vortex state for C = 1000. The plots show the
quasi-particle amplitudes (a) [u[
2
and (b) [v[
2
. The frequency of this mode is
= 0.124.
amplitude [u[
2
in Fig. 10.11 (a) lls the vortex core, while [v[
2
in Fig. 10.11 (b) has a
density dip at the vortex position.
The rst two modes in Fig. 10.10 (a) with a positive energy are the degenerate
162
10.3. Excitations on Vortex Lattices
dipole modes at an energy of = 1 with q = 1. They correspond to a centre of mass
motion of the condensate in the harmonic trap.
The rst breathing mode with q = 0 has the energy = 2, which is a universal
value for a two dimensional system. The excitation is rotationally symmetric as shown
in Fig. 10.12. The condensate density expands and contracts uniformly in the radial
direction.
10
0
10
10
0
10
0
0.02
0.04
0.06
Position x
Position y
(a)
A
m
p
l
i
t
u
d
e

|
u
|
2
10
0
10
10
0
10
0
0.02
0.04
0.06
Position x
Position y
(b)
A
m
p
l
i
t
u
d
e

|
v
|
2
Figure 10.12: First breathing mode on single vortex state for C = 1000. Depicted are the
quasi-particle amplitudes (a) [u[
2
and (b) [v[
2
. The frequency is the universal
frequency of the rst breathing mode = 2 in a two-dimensional system.
When the same equations are solved in a rotating frame, we obtain the spectrum
shown in Fig. 10.10 (b) for = 0.3 . Compared to Fig. 10.10 (a), the spectrum is
screwed to one side because the excitation frequencies are shifted by q, where q is
the angular momentum of the respective modes. Consider for example the two dipole
modes ([q[ = 1). They have their degeneracy lifted with
R
= 1 for q = 1. The
breathing mode at
R
= 2, on the other hand, is unaected by the transformation into
the rotating frame, because it is a rotationally symmetric mode with q = 0. We also
note that the mode which was anomalous in the laboratory frame has now a positive
energy in the rotating frame. This indicates the stabilization of the vortex through the
rotation.
In Fig. 10.13 we plot the spectrum from the laboratory frame and rotating frame in
the same graph, but reverse the energy shift due to the rotation by plotting
R
+ q.
The agreement is very good which shows that our method to solve for the excitations
is robust in the rotating frame. A slight disagreement for the two centre of mass modes
is apparent though. Their energies are the same in either frame, but the calculated
values for q are not exactly q = 1 in the laboratory frame. We are not sure what
causes this problem. Especially puzzling is the fact that it only occurs for these two
163
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
20 10 0 10 20
1
0
1
2
3
4
5
6
7
8
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
+q
Figure 10.13: Spectra of single vortex state from Fig. 10.10 combined in one plot. The
frequencies in the laboratory frame are the same as in Fig. 10.10 (a), but the
frequencies from the rotating frame ( = 0.3) are shifted by +q.
modes, while all the others have integer values for q as expected.
10.3.2 Excitations on vortex lattice
10.3.2.1 Frequency spectrum
We now turn to excitations on vortex lattices. Fig. 10.14 (a) shows the excitation
spectrum on a lattice with three vortices rotating at = 0.35, in the frame rotating
at = . The angular momentum of the vortex lattice wave function is l = 2.137.
Since that is not an integer number, we do not expect the angular momentum of the
excitations to be integer either in contrast to the case of the single central vortex.
Because the lattice has three vortices there are three modes that ll the vortex core
similar to the anomalous mode in the single vortex case. These three modes have the
lowest energies in Fig. 10.14 (a), but their eigenvalues are positive due to the rotation.
The breathing mode is still exactly at = 2. In general, the energies of the excitations
are lower than that of the corresponding modes in the single vortex case, and the
branches in the spectrum are not as well dened.
As we add vortices, the spectrum becomes more and more complicated. Fig. 10.14
(b) is an example for the lattice with 12 vortices shown in Fig. 9.1 (a). While the
energy of the breathing mode is still exactly at = 2, the mode carries now slight
164
10.3. Excitations on Vortex Lattices
10 0 10 20 30 40
0
2
4
6
8
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
(b)
10 0 10 20 30 40
0
2
4
6
8
Angular momentum q
E
x
c
i
t
a
t
i
o
n

f
r
e
q
u
e
n
c
y

R
(b)
Figure 10.14: Spectrum of lattice states with (a) three vortices rotating at = 0.35 and
(b) twelve vortices rotating at = 0.65. For each case the spectrum was
calculated in the frame corotating with the lattice. Both lattices were cre-
ated dynamically for the parameters C = 1000 and
NC
= 12 using the
phenomenological growth equation.
angular momentum. This is the case because the vortex lattice is not centred on the
rotation axis. There are now twelve low-lying modes with lled vortex cores.
10.3.2.2 Tkachenko modes
The recent observation of Tkachenko modes [133, 184] by Coddington et al. [82] has
stimulated considerable interest in these modes of a rotating vortex lattice [185, 186,
187]. Tkachenko modes are transverse elastic modes of the vortex lattice as a whole.
Mizushima et al. [187] have identied Tkachenko modes in the BdG spectrum as
core lled modes, where the phase rotation around the whole condensate relative to
the phase circulation of the lattice wave function is zero. What we mean by this is
the following. Say the condensate has n
v
vortices. That means the phase circulation
around the condensate is 2n
v
. Then, if the quasi-particle amplitude u has a total
phase circulation of 2m
u
= 2(n
v
+ m), the quasi-particle amplitude v has a phase
circulation 2m
v
= 2(n
v
+m). This is a generalization of the condition from section
3.1 for the magnetic quantum numbers of the quasiparticle amplitudes m
v
= m
u
2m
c
,
if the condensate is an eigenstate of

L
z
with the magnetic quantum number m
c
. For
Tkachenko modes m = 0. That means they have no angular momentum relative to the
condensate.
In the experiment [82], the condensed cloud contains 1.5 to 2.9 10
6 87
Rb atoms
in a trap with

,
z
= 28.3, 5.2 Hz. The lattice typically rotates at = 0.84 to
165
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
= 0.975. Fig. 10.15 shows the rst Tkachenko mode in a BoseEinstein condensate
excited via atom removal from the centre of the trap by a blue-detuned laser beam of
16 m width with 10 fW for approximately one lattice rotation period.
Figure 10.15: First Tkachenko mode excited by atom removal (a) taken 500 ms after the
end of the blasting pulse, (b) taken 1650 ms after the end of the blasting
pulse. Condensate rotation is counter-clockwise. Lines are sine ts to the
vortex lattice. Figure is taken from reference [82].
We were able to simulate their results using the pure GrossPitaevskii equation
(GPE) in two dimensions [88]. The initial vortex lattice was created dynamically with
the phenomenological growth equation using = 0.95 and
NC
= 11. We then apply
the potential
V
beam
= 0.02ie
4(x
2
+y
2
)/4
(10.3)
for one trap period. This simulates a weak resonant laser beam that removes atoms
from the centre of the trap. Our results are shown in Fig. 10.16.
Unfortunately, we were not able to calculate the Tkachenko modes for a lattice that
large because it would require many basis states, which is too memory intensive for
our computers. However, we have identied the Tkachenko modes of the lattice with
twelve vortices whose spectrum is shown in Fig. 10.14 (b). Following the condition by
Mizushima et al. [187] described at the beginning of this section, we looked for the
core lled modes that have the same phase circulation as the lattice.
The densities of the quasi-particle amplitudes for the rst Tkachenko mode are
plotted in Fig. 10.17 (a) and (b) for the lattice with twelve vortices. We see that the
quasi-particle amplitudes u and v are very similar. Only the vortices of the outer ring
of the lattice have lled cores. These lled cores are connected by a closed density
ring. Fig. 10.17 (c) and (d) show the second Tkachenko mode. Here, the u amplitude
lls the cores of the outer ring of vortices, and the v amplitude lls the cores of the
166
10.3. Excitations on Vortex Lattices
Position x
(a)
12 6 0 6 12
12
6
0
6
12
Position x
(b)
12 6 0 6 12
Position x
(c)
12 6 0 6 12
10
4
10
3
10
2
10
1
Figure 10.16: First Tkachenko mode in GPE simulation. (a) Initial lattice conguration
rotating at = 0.95 for C = 1000 and
NC
= 11. At time t = 0 atoms are
removed from the trap centre by application of the weak potential (10.3) for
a time period of 2. (b) Lattice conguration at time t = 6 and (c) t = 18.
Straight lines are overlaid to make the recognition of the lattice structure
oscillation easier.
inner ring of vortices. Presumably, that is the reason that the two rings oscillate out
of phase. There are no other Tkachenko modes for this lattice because it consists only
of two rings of vortices. Bigger lattices have more rings of vortices and hence more
Tkachenko modes.
167
Chapter 10. Non-equilibrium Dynamics and Lattice Excitations
P
o
s
i
t
i
o
n

y
(a)
5
0
5
(b)
Position x
P
o
s
i
t
i
o
n

y
(c)
5 0 5
5
0
5
Position x
(d)
5 0 5
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Figure 10.17: Quasi-particle amplitudes of Tkachenko modes for lattice with twelve vortices.
The rst Tkachenko mode is depicted in the rst row with the quasi-particle
amplitudes (a) [u[
2
and (b) [v[
2
. The second row shows the second Tkachenko
mode with the quasi-particle amplitudes (c) [u[
2
and (d) [v[
2
.
168
Chapter 11
Conclusion
11.1 This Thesis
In this thesis, we have presented extensive studies of the excitations and dynamical
behaviour of BoseEinstein condensates. The study has required the development
and implementation of sophisticated numerical techniques, as well as the development
of simple analytic models to aid understanding and interpretation of the phenomena
found.
The rst part of this thesis examined the ground state and the excitations of a
stationary non-rotating cylindrically symmetric condensate, which we treated within
the standard mean-eld formalism of the GrossPitaevskii equation (GPE) and the
Bogoliubovde Gennes equations. We presented solutions of these equations in three
dimensions and showed that the concept of mode families introduced by Hutchinson
and Zaremba [84] could be extended to include all excitation modes. We devised a
procedure to identify the mode family of any mode in an anisotropic trap on the basis
of the characteristic shape of all modes of the same family. We also determined the
energy ordering of the modes and, by relating it to the mode shape, explained the
ordering in terms of a simple model based on harmonic oscillator states.
The main part of this thesis was concerned with rotating BoseEinstein condensates.
Because a BoseEinstein condensate is a superuid, any ow pattern in the condensate
must be irrotational. When a superuid is set into rotation, vortices appear above a
certain critical rotation rate because for a given angular momentum vortices lower the
free energy of the system. Prior to this thesis it was known that a vortex lattice is a
stationary state solution that minimizes the energy of the GPE in the rotating frame.
However, the dynamics that lead to such a stationary state in a rotating BoseEinstein
169
Chapter 11. Conclusion
condensate cannot be adequately described by the GPE because that equation does
not include the dissipation required to achieve a lattice state.
In order to address some of the key questions of vortex lattice formation such as the
nucleation mechanism, the critical frequency of vortex nucleation and the stabilization
process of a vortex lattice, we have implemented for the rst time a formalism developed
by Gardiner et al. [85] to treat the condensate dynamics in the presence of a thermal
cloud. In its most simplied form this reduces to a modied GPE, which includes
growth and loss terms arising from the exchange of atoms between the condensate and
a thermal bath. We call this equation the phenomenological growth equation, and we
have developed some numerical schemes to simulate this equation. We have shown that
it provides a simple, unied treatment from the initiation to the nal stabilization of a
vortex lattice.
We have simulated in two dimensions lattice formation experiments from a rotating
thermal cloud and in a rotating elliptical trap. Our results show that the basic mecha-
nism in the nucleation of vortex lattices is an exponential growth into particular surface
modes of non-zero angular momentum. We have given a simple analytic treatment of
this process, with explicit expressions for the gain coecients that govern the growth
of each mode, and shown that it agrees very well with the full numerical simulations.
Our model is valid for both the lattice formation from a rotating thermal cloud and in
an elliptical rotating trap, and thus gives a unied explanation of all the experiments.
The critical value of the rotation rate is obtained when the mode with maximum gain
rst exceeds the loss rate, and coincides with the Landau criterion. We have discussed,
in relation to our model, why the critical angular velocity has been observed in some
experiments to be raised or lowered.
By calculating the vortex free eigenstates and their excitations in an elliptical ro-
tating trap, we found a possible explanation for the low limit on the critical angular
frequency observed by Hodby et al. [152] when the ellipticity of the trap is increased.
This is due to a peculiar change of nature in the quadrupole mode for suciently large
trap ellipticities.
Finally, we have presented some simulations relevant to recent experiments on the
decay of vortex lattices, the excitations of vortex lattices such as the Tkachenko modes,
and on non-equilibrium lattice dynamics. In particular, we have considered the defor-
mation of large vortex lattices in rapidly rotating condensates, which show sheet like
structures when excited by an m = 2 mode, and lattice melting under the inuence
of an m = +2 excitation [81].
In summary, we have shown that the phenomenological growth equation can be
170
11.2. Future Work
applied to a wide variety of dierent scenarios, and shows a rich spectrum of interesting
physics. In particular, it provides a tractable unied model for vortex lattice formation
from the initiation process of vortex nucleation through to the nal stabilization of the
lattice, and highlights the crucial role of the thermal cloud.
11.2 Future Work
The major shortcoming of the phenomenological growth equation is the fact that the
thermal cloud is treated as an innite bath of thermal atoms with xed temperature,
chemical potential and angular momentum. It does not include a detailed treatment
of the dynamics of the thermal cloud. But as we have shown, the thermal cloud plays
a crucial role in the formation of vortex lattices. Therefore, it would be valuable to
see how the thermal cloud spins up or slows down due to the rotating trap or the
interaction with the vortex lattice. The question of heating caused by the stirring is
also important. In some experiments, a loss of atoms from the condensate by up to
30% was observed [77].
To include all these eects is a very formidable task, but our work should provide a
rm basis of results on which to build. Promising steps have been made towards a more
complete treatment [163] by combining the approach of a stochastic GPE [85] with the
nite temperature GPE [164]. Here, the dynamics of the thermal cloud is included in
the GPE (a good approximation as long as the thermal modes are highly occupied)
and a projector is used to cut o higher energy modes. First simulations based on this
treatment on vortex lattice formation show very similar behaviour as that observed
with our simple treatment [165].
171
Appendix A
Optimization Methods
Because the time-independent GrossPitaevskii equation (GPE) is a non-linear dier-
ential equation, it is in general necessary to use iterative numerical methods to nd its
solution. The Bogoliubovde Gennes (BdG) equations, on the other hand, are linear,
and thus considerably easier to solve. A transformation into a simple linear matrix
equation makes them solvable by standard methods of variable elimination.
In this appendix, we describe the dierent numerical methods we used to solve
the GPE and BdG equations. All of them can loosely be categorized under the term
optimization methods, whereby an optimization problem is dened [188] as a set of
independent variables or parameters, whose acceptable values are often subject to some
conditions or restrictions, and a single measure of goodness, the objective or optimality
function, which depends in some way on the variables. The solution of an optimization
problem is a set of allowed values of the variables for which the optimality function
assumes an optimal value.
To solve a given problem with an optimization method one has to address two very
distinct tasks. First, the problem has to be expressed in a suitable way in terms of
a single optimality function. And secondly, some method has to be applied to nd
the optimal value of this optimality function. This usually involves some form of
extremizing or root-nding. We have used various approaches for both tasks which we
summarize in the following.
173
Appendix A. Optimization Methods
A.1 Partial Dierential Equations Solver
A.1.1 Cylindrically symmetric case in three dimensions
The solutions to the GPE and BdG equations in three dimensions with cylindrical
symmetry (3.3)(3.5) presented in chapter 3 were obtained using Matlabs partial
dierential equations (PDEs) toolbox. This toolbox is designed to solve a set of coupled
dierential equations in two dimensions using nite element methods (FEMs). The two-
dimensional (2D) plane is divided into many interconnected elements and a solution
is sought as a piecewise approximation in terms of interpolation functions within each
element [189].
The basic equation the toolbox can handle is of the form
(cu) +au = f (A.1)
on a bounded domain . The coecients a and f and the unkown u are scalar, complex
valued functions of two variables that describe the plane dened on . The coecient
c may be a 2 2 matrix function on to allow for the two spatial components. All of
these coecients may depend on u to allow for non-linear problems.
The PDE toolbox can also solve the basic linear eigenvalue problem
(cu) +au = qu, (A.2)
where is the eigenvalue and q might be a two-component vector. In this case, the
coecients are not allowed to depend on u. Both solvers can also handle coupled PDEs
using multi-component variables.
Since the GPE is not only non-linear, but also an eigenvalue problem, it cannot
directly be cast into the form of equation (A.1) nor equation (A.2). A further step is
required if we want to solve the GPE using the PDE toolbox. We make use of the
fact that the normalization of the wave function is a continuous monotonic function
of for xed non-linearity C. This allows to dene an optimality function () =
_
dx[()[
2
1, where () is the solution of the GPE for a certain , which is not
174
A.1. Partial Dierential Equations Solver
necessarily normalized to unity. This solution () is found with the PDE solver setting
c =
_
1
1
_
, (A.3)
a =
_
m
2
c

2
+
1
4
_

2
+ (z)
2

+C[u[
2
_
, (A.4)
f = u. (A.5)
A properly normalized solution of the GPE (3.3) is obtained when () = 0, which
we determine with Matlabs simple root solver fzero. This requires, of course, the
solution of equation (A.1) for every succinct value of , which is by itself an iterative
process. Thus, to nd a very accurate solution may be quite time-consuming. For a
discussion of the performance and accuracy of this method the interested reader may
be referred to my Masters thesis [102].
The BdG equations are easily cast into the form (A.2) suitable for a direct solution
by the eigenvalue solver setting
c =
_
_
_
_
_
_
1 0 0
1 0 0
0 0 1
0 0 1
_
_
_
_
_
_
, (A.6)
a =
_
a
11
a
12
a
21
a
22
_
, (A.7)
q =
_
1
1
_
, (A.8)
with
a
11
=
_
m
2

2
+
1
4
_

2
+ (z)
2

+ 2C[u[
2

_
, (A.9)
a
22
=
_
(m2m
c
)
2

2
+
1
4
_

2
+ (z)
2

+ 2C[u[
2

_
, (A.10)
a
12
= a

21
= Cu
2
. (A.11)
A.1.2 Two-dimensional system in the rotating frame
The same method can be used to solve the two-dimensional GPE (8.3) and BdG equa-
tions (8.4) in the rotating frame. As an example, we consider here briey the GPE (8.3)
175
Appendix A. Optimization Methods
in a frame rotating at angular velocity . We assume that the trapping potential is of
the form (8.1) with the trap ellipticity =
x
=
y
and that the chemical potential is
xed at =
NC
. This can be solved with the non-linear PDE solver setting
c =
_
1 i

2
(x
2
+y
2
)
i

2
(x
2
+y
2
) 1
_
, (A.12)
a =
_
1
4
_
(1 )x
2
+ (1 +)z
2

+C[u[
2
_
, (A.13)
f =
NC
u. (A.14)
A.2 Conjugate-Gradient Method
In this section, we present the conjugate gradient method, which we have used to
nd the vortex free solutions of a condensate in a rotating elliptical trap, presented in
section 8.3.1. The optimality function is constructed using nite-dierence operators
to calculate the derivatives occuring in the GPE. The conjugate gradient method is
then used to minimize this optimality function.
A.2.1 Optimality function
We are interested in solving the GrossPitaevskii equation in the rotating frame in
two dimensions for a xed eigenvalue . With the appropriate boundary condition the
problem can be expressed as
_


2
x
2


2
y
2
+V
T
(x, y) +C[(x, y)[
2
+i
_
x

y
y

x
_

_
(x, y) = 0,
(A.15)
(x, y)[
x
2
+y
2
0
= 0.
(A.16)
We discretize the problem on a uniform rectangular grid dened in the x- and y-
directions by the vectors
x = [x
1
, x
2
, . . . , x
m
x
]
T
, (A.17)
y =
_
y
1
, y
2
, . . . , y
m
y

T
(A.18)
176
A.2. Conjugate-Gradient Method
with the grid spacings x = x
i+1
x
i
and y = y
i+1
y
i
. On this grid, the wave
function and trapping potential are represented by matrices

ij
= (x
i
, y
j
), (A.19)
V
ij
= V
T
(x
i
, y
j
). (A.20)
Because of the angular momentum term in (A.15), we have to allow for a complex wave
function
ij
. We split
ij
in a real and complex part

ij
= u
ij
+iv
ij
, (A.21)
where u
ij
= (
ij
) and v
ij
= (
ij
). This leads to two sets of real functions
f
R
ij
=
m
x

k=1
D
(2)
ik
u
kj

m
y

k=1

D
(2)
jk
u
ik

_
x
ij
m
y

k=1

D
(1)
jk
v
ik
y
ij
m
x

k=1
D
(1)
ik
v
kj
_
_
V
ij
+C(u
2
ij
+v
2
ij
)

u
ij
, (A.22)
f
I
ij
=
m
x

k=1
D
(2)
ik
v
kj

m
y

k=1

D
(2)
jk
v
ik
+
_
x
ij
m
y

k=1

D
(1)
jk
u
ik
y
ij
m
x

k=1
D
(1)
ik
u
kj
_
_
V
ij
+C(u
2
ij
+v
2
ij
)

v
ij
, (A.23)
that represent the residual of the GPE (A.15) on the grid point (x
i
, y
j
), where 1 i
m
x
and 1 j m
y
. When all residuals are zero, equation (A.15) is fullled.
Here, D
(2)
ij
and D
(1)
ij
are operators based on nite dierences for which
m
x

k=1
D
(2)
ik
u
kj


2
x
2
u(x
i
, y
j
), (A.24)
m
x

k=1
D
(1)
ik
u
kj


x
u(x
i
, y
j
). (A.25)
The operators

D
(2)
ij
and

D
(1)
ij
are the corresponding operators for the y-derivatives.
In appendix B.2, we briey illustrate how nite dierence derivatives work. The full
matrix form of these operators can be found in [174].
It remains to implement the boundary condition (A.16). Because the grid is of nite
size, it is not possible to enforce the boundary condition at innity. Instead, we require
the wave function to be zero at the edge of the grid. Since the wave function approaches
zero exponentially with increasing radial position, this is a good approximation as long
177
Appendix A. Optimization Methods
as the grid is suciently large. This gives the residuals for the boundary condition
f
R
B1j
= u
1j
, f
R
B2j
= u
m
x
j
, (A.26)
f
R
B3i
= u
i1
, f
R
B4i
= u
im
y
, (A.27)
f
I
B1j
= v
1j
, f
I
B2j
= v
m
x
j
, (A.28)
f
I
B3i
= v
i1
, f
I
B4i
= v
im
y
. (A.29)
As a single valued optimality function, we take the least squares of the residuals
F(u
ij
, v
ij
) =
1
2
_
m
x

i=1
m
y

j=1
_
(f
R
ij
)
2
+ (f
I
ij
)
2

+
m
y

j=1
_
(f
R
B1j
)
2
+ (f
R
B2j
)
2
+ (f
I
B1j
)
2
+ (f
I
B2j
)
2

+
m
x

i=1
_
(f
R
B3i
)
2
+ (f
R
B4i
)
2
+ (f
I
B3i
)
2
+ (f
I
B4i
)
2

_
. (A.30)
The gradient of the optimality function F(u
ij
, v
ij
) can be calculated analytically
and is given by
F
u

=
m
x

i=1
D
(2)
i
f
R
i

m
y

j=1

D
(2)
j
f
R
j
+
_
V

+C(3u
2

+v
2

f
R

+ 2Cu

f
I

+
_
x

m
y

j=1

D
(2)
j
f
I
j
y

m
x

i=1
D
(2)
i
f
I
i
_
+f
R
B1

1
+f
R
B2

m
x

+f
R
B3

1
+f
R
B4

m
y
(A.31)
F
v

=
m
x

i=1
D
(2)
i
f
I
i

m
y

j=1

D
(2)
j
f
I
j
+
_
V

+C(u
2

+ 3v
2

f
I

+ 2Cu

f
R

_
x

m
y

j=1

D
(2)
j
f
R
j
y

m
x

i=1
D
(2)
i
f
R
i
_
+f
I
B1

1
+f
I
B2

m
x

+f
I
B3

1
+f
I
B4

m
y
. (A.32)
178
A.2. Conjugate-Gradient Method
A.2.2 Conjugate-gradient optimization
We now need a method to nd the minimum of the optimality function. Because of
the way the optimality function is constructed as a sum of least squares, the minimum
value is zero. That corresponds to the desired solution of the GPE. The conjugate-
gradient method is a method to nd the minimum of a function F(X) where X =
[X
1
, X
2
, . . . , X
m
]
T
is a vector of m variables. In our case, X is the vector that consists
of all the components u
ij
, v
ij
. Only the gradient, without a need for the Hessian
matrix, is used to calculate the search directions in the minimization process, which
makes the conjugate-gradient method well suited for large-scale problems.
The main idea of the conjugate-gradient method is to minimize the function suc-
cessively in dierent subspaces without spoiling the minimization in the previous sub-
space. This is achieved by choosing successive search directions that are conjugate to
each other with respect to the Hessian matrix. Only for a quadratic optimality function
achieves the algorithm truely conjugate search directions. In that case, the minimum
is reached in m steps, where m is the number of independent variables.
The conjugate gradient minimization can be summarized as follows. Starting with
an initial guess X
(0)
and
0
= 0, and the initial search direction p
0
= g(X
(0)
), where
g is the gradient
g(X) =
_
_
_
_
F
X
1
.
.
.
F
X
m
_
_
_
_

X
, (A.33)
the following steps are iterated:
1. Find the step length
i
that minimizes the objective function along the search
direction p
i
. We approximate this with

i
=
[g(X
(i)
)[
2
D
p
i
(g)[
X
(i)
, (A.34)
where D
p
(g) is the change of g along the direction p given by
D
p
(g(X))[
X
(i) = lim
h0
g(X
(i)
+hp
i
) g(X
(i)
)
h
. (A.35)
The limit to zero is unnecessary for a quadratic optimality function F(X). In our
implementation, we use a small value for h, namely h = 10
7
.
179
Appendix A. Optimization Methods
2. Update the estimate of the minimum
X
(i+1)
= X
(i)
+
i
p
i
. (A.36)
3. Compute the new search direction
p
i+1
= g(X
(i+1)
) +
i
p
i
, (A.37)
where

i
=
[g(X
(i+1)
)[
2
[g(X
(i)
)[
2
. (A.38)
For a non-linear function, the quadratic model the conjugate-gradient algorithm as-
sumes might not be appropriate. Hence, in general it is necessary to check in the
second step whether F(X) really decreases. If F(X) increases, at some step i say, then
the algorithm is reset by taking
i
= 0, p
i
= g(X
(i)
). This is also done routinely
after a xed number of steps, typically of the order of 200 steps.
A.3 Basis State Method
Another method we tried to nd the excitations of a rotating condensate is a basis state
expansion. This method has been extensively used to calculate the ground state and
its excitations in a non-rotating frame and is described in detail in [190]. Our program
is based on code for a cylindrically symmetric three-dimensional case that was kindly
provided by Sam Morgan. We had to adjust it for the use in two dimensions and also
include the angular momentum term which arises from the transformation into the
rotating frame. We will outline the basic method here, and will comment in particular
on the changes that were necessary to accomodate our problem.
A.3.1 Basis state expansion
Because the GPE is a non-linear eigenvalue problem, there is no way around some form
of root-nding or optimization method. For small to medium sized non-linearities,
however, a basis state expansion might reduce the number of variables necessary to
represent a good solution. Instead of optimizing the wave function directly on a dis-
cretized spatial grid, the wave function is described in terms of an expansion in terms
of a basis state set, and the optimization is carried out on a set of equations for the
180
A.3. Basis State Method
expansion coecients. With a suitable choice of basis states, the solution might be well
approximated with only a few basis functions. In particular, if the problem displays
certain symmetries, the number of basis states can be reduced by using only basis states
with the same symmetries.
The GPE in the rotating frame is given by
[

h
0
+C[[
2

L
z
] = 0, (A.39)
where

h
0
=
2
+V
T
. (A.40)
We are intererested in the case of an elliptical trap of the form
V
T
=
1
4
_
(1 )x
2
+ (1 +)y
2

. (A.41)
The basic idea is to expand the wave function in some nite set of 2D basis states

ij
as
=
m
x
,m
y

m,n=0
c
mn

mn
(A.42)
and derive a coupled set of equations for the expansion coecients c
mn
by calculating
the matrix elements
M
ijmn

ij
[

h
0
+C[[
2

L
z
[
mn
). (A.43)
With these matrix elements, we can rewrite (A.39) as a set of coupled equations f
ij

f
ij

m,n
M
ijmn
c
mn
= 0. (A.44)
To solve this set of equations, we used Matlabs optimization solver fsolve. While
fsolve can in principle solve for complex valued variables, its performance is better if
we separate the coecients explicitly into real and complex parts
c
ij
= a
ij
+ib
ij
. (A.45)
In the next sections we will discuss in detail how we evaluate the matrix elements
(A.43). We will also show how to calculate the Jacobian analytically because for a
large number of variables the performance of the optimization routine fsolve is only
181
Appendix A. Optimization Methods
satisfactory if the Jacobian is supplied.
A.3.2 Construction of harmonic oscillator states
As basis states we choose harmonic oscillator states. As is well known [191] the har-
monic oscillator solutions in cylindrical symmetry can either be expressed in Cartesian
coordinates using Hermite polynomials or in polar coordinates using Laguerre polyno-
mials. However, in our case the trap is elliptical, which allows separation of variables
only in Cartesian coordinates, and thus we have to use the Hermite polynomials.
Because the standardnormalization of the Hermite polynomials may cause a com-
putational overow for higher orders [192], we are using the orthonormal set of poly-
nomials, which can be constructed from the recursion relation

H
1
(x) = 0, (A.46)

H
0
(x) = (
1

)
1
4
, (A.47)

H
j+1
(x) = x
_
2
j + 1

H
j
(x)

j
j + 1

H
j1
(x). (A.48)
They obey the derivative formula
d
dx

H
j
(x) =
_
2j

H
j1
(x) (A.49)
and are orthonormal with respect to the weight function W(x) = e
x
2
, i.e.
_
W(x)

H
i

H
j
=
ij
. (A.50)
For our trap geometry (A.41), we obtain the following orthonormal harmonic oscil-
lator states

ij
(x, y) =
_

y
e
(
2
x
x
2
+
2
y
y
2
)/2

H
i
(
x
x)

H
j
(
y
y), (A.51)
where
2
x,y
=

1 /2 respectively. The energies of these states are given by


E
ij
= 2(i +
1
2
)
2
x
+ 2(j +
1
2
)
2
y
. (A.52)
Because these states are eigenstates of the operator

h
0
, we obtain as part of the full
182
A.3. Basis State Method
matrix element (A.43) the diagonal term
M
(1)
ijmn

ij
[

h
0
[
mn
) =
im

jn
E
mn
. (A.53)
For the rst part f
(1)
ij
=

mn
M
(1)
ijmn
c
mn
of the full optimization functions (A.44)
the Jacobian is simply given by
f
(1)
ij
a
mn
= M
(1)
ijmn
, (A.54)
f
(1)
ij
b
mn
= iM
(1)
ijmn
. (A.55)
A.3.3 Angular momentum operator
A big advantage of the basis state method is that the angular momentum operator

L
z
= i(x
y
y
x
) can be implemented analytically within the accuracy of the nite
basis set using the relations (A.48) and (A.49). We will illustrate how to do this by
calculating the term i
ij
[y
x
[
mn
).
Rearranging (A.48) gives

x
x

H
j
(
x
x) =
_
j
2

H
j1
(
x
x) +
_
j + 1
2

H
j+1
(
x
x). (A.56)
Using this relation and (A.49) one can write

mn
=
2
x
x
mn
+
x

2i
m1,n
=
x
_

_
m
2

m1,n

_
m+ 1
2

m+1,n
+

2m
m1,n
_
=

x

2
_

m
m1,n

m+ 1
m+1,n
_
. (A.57)
With the equivalent relation to (A.56) for the y-component, one obtains
y

x

mn
=

x
2
y
_

m
_

n
m1,n1
+

n + 1
m1,n+1
_

m+ 1
_

n
m+1,n1
+

n + 1
m+1,n+1
__
. (A.58)
183
Appendix A. Optimization Methods
Thus, for this part of the angular momentum operator, the matrix element is
i
ij
[y
x
[
mn
) = i

x
2
y
_

m
_

n
i,m1

j,n1
+

n + 1
i,m1

j,n+1
_

m+ 1
_

n
i,m+1

j,n1
+

n + 1
i,m+1

j,n+1
__
. (A.59)
We note that the basis functions
mn
are symmetric with respect to a point reection
at the origin if m + n is even, and anti-symmetric if m + n is odd. We can see from
expression (A.58) that the application of the angular momentum operator preserves
this symmetry since only basis states of the same symmetry are involved.
With M
(2)
ijmn

ij
[

L
z
[
mn
), the Jacobian for f
(2)
ij
=

mn
M
(2)
ijmn
c
mn
is given by
the equivalent equations of (A.54) and (A.55).
A.3.4 Gaussian quadrature
Another problem is the evaluation of the non-linear term. A way to calculate the
integrals exactly is to use Gaussian quadratures [192]. Similar to other numerical
integration techniques, as for example Simpsons rule, an integral is approximated by a
weighted sum of function values at each grid point. However, the idea of the Gaussian
quadrature method is not to use equidistant grid points, but to choose them in such
a way that the integral becomes exact with the appropriate weights for a class of
integrands of polynomials times some known function. Since we do not rely on nite
dierences to calculate the angular momentum term and we can also calculate our basis
set functions on any grid, we can readily use a non-equidistant grid.
For the function W(x) = e
x
2
, we use a standard procedure from Numerical
Recipes [192] to calculate the appropriate weights w

and abscissa points x

that
make the integral approximation
_
W(x)f(x)dx
N

=1
w

f(x

) (A.60)
excact if f(x) is a polynomial. Because we are interested in the matrix elements

ij
[C[[
2
[
mn
) we have to adjust those for the integration of a product of four harmonic
oscillator states (A.51). This is done by the coordinate transformation
x

=
x

2
x
, (A.61)
184
A.3. Basis State Method
and the equivalent for the y-coordinate. We also need to adjust the weights according
to
w

=
w

2
x
e
x
2

, (A.62)
because the exponential function is already contained in the harmonic oscillator states
(A.51).
Before we can calculate the matrix elements, we need to evaluate the basis states on
the x- y grid, which we will denote by

ij

ij
( x

, y

). We also write

ij
c
ij

ij
.
The matrix element becomes then
M
(3)
ijmn

ij
[C[[
2
[
mn
) = C

ij
[

[
2

mn
, (A.63)
where w

is the product of the weights at x

and y

. Note that we do not need to


take the conjugate complex of

ij
in expression (A.63) because our basis states are
real functions.
However, when we calculate the Jacobian for f
(3)
ij
=

mn
M
(3)
ijmn
c
mn
we have to take
care because is a complex function. We obtain
f
(3)
ij
a
mn
= M
(3a)
ijmn
+ 2M
(3)
ijmn
, (A.64)
f
(3)
ij
b
mn
= iM
(3a)
ijmn
+ 2iM
(3)
ijmn
, (A.65)
where
M
(3a)
ijmn

ij
[C
2
[
mn
) (A.66)
is calculated in the same way as (A.63).
A.3.5 Eigenvalue
The eigenvalue is an unknown variable. In order to obtain a self-consistent solution
of the GPE, has to equal the expectation value [

h
0
+C[[
2

L
z
[)/[). If we
wish for a solution normalized to unity we simply set
= [

h
0
+C[[
2

L
z
[) =

klpq
c

kl
c
pq
M
(4a)
klpq
, (A.67)
185
Appendix A. Optimization Methods
where M
(4a)
klpq

kl
[

h
0
+C[[
2

L
z
[
pq
). This can be evaluated using all the techniques
described in the previous sections. With this value of , the matrix element is
M
(4)
ijmn

ij
[[
mn
) =
im

jn
. (A.68)
A self-consistent solution would then automatically be normalized to unity, and obey
the GPE (A.39) with given by (A.67).
On the other hand, if we are looking for a solution with a specic chemical potential
=
NC
, we can use this value for directly without the need to calculate (A.67). We
then obtain a solution with the appropriate normalization, which may not be unity.
If is a constant =
NC
, the Jacobian for the term is trivial. However, if is
calculated using (A.67), we have to take the inner derivative as well. Thus, one obtains
the Jacobian for f
(4)
ij
=

mn
M
(4)
ijmn
c
mn
f
(4)
ij
a
mn
=
im

jn
+

pq
c
pq
M
(4a)
mnpq
+

kl
c

kl
M
(4a)
klmn
, (A.69)
f
(4)
ij
b
mn
= i
_

im

jn

pq
c
pq
M
(4a)
mnpq
+

kl
c

kl
M
(4a)
klmn
_
, (A.70)
These expressions could be slightly simplied if we used an explicit evaluation of M
(4a)
klpq
,
but we refrain from it here because it would not add any more insight.
A.3.6 Bogoliubovde Gennes equations
A convenient way to solve the ordinary BdG equations is to introduce

i
= u
i
v

i
,
which leads to two decoupled equations for

. If the ground state wave function is


real, the resulting equations contain only the operator

h
0
and its square. However,
in our case, where the ground state is a complex function, the resulting equations are
considerably more complicated and cannot easily be separated. So, we decided to solve
the full BdG equations with the functions u
i
and v
i
.
In the same way as the GPE, the BdG equations can be cast into a matrix equation
for the expansion coecients of u
i
and v
i
. Because the BdG equations are linear,
the solution can be found without employing optimization methods. Instead, we used
Matlabs simple eigenvalue solver eig. This gives all eigenvalues and eigenfunctions
(up to a certain energy) at once. This is a big advantage compared to optimization
methods. For any optimization method, a reasonably good guess of the excitation has
186
A.3. Basis State Method
to be supplied as an initial starting point for the optimization. Often it takes quite a
bit of trial and error to nd a particular solution with a certain number of nodes. And
there is never any guarantee that all excitations are found.
187
Appendix B
Propagation Methods For
Dynamical Simulations
B.1 GrossPitaevskii Equation Propagation
B.1.1 Denition of the problem
To understand the basic idea of the RK4IP (Runge-Kutta Fourth-Order Interaction
Picture) algorithm we will rst review the propagation of the ordinary GrossPitaevskii
equation (GPE). The time-dependent GPE can be written in the form

t
=
_

D +

N
_
, (B.1)
with the operators

D = i
2
, (B.2)

N = i[V
T
+C[[
2
]. (B.3)
The operator D is linear, while N includes the non-linear part of the equation. Equation
(B.1) can be cast into an ordinary rst order dierential equation by a transformation
into an interaction picture

I
= e
(tt

)

D
, (B.4)
where t

is an arbitrary time oset. The resulting equation is given by

I
=

N
I

I
, (B.5)
189
Appendix B. Propagation Methods For Dynamical Simulations
where the non-linear operator in the interaction picture is dened as

N
I
= e
(tt

)

D

Ne
(tt

)

D
. (B.6)
In the interaction picture, the evolution equation (B.5) is suitable to be integrated by
the standard fourth-order Runge-Kutta method.
B.1.2 Transformation into the interaction picture
A convenient and ecient way to transform into the interaction picture makes use
of the Fourier transformation. If we dene the Fourier transform of (x) as
(k) =
_
(x)e
ikx
dx, (B.7)
we can write in terms of its Fourier components as
(x) =
_
1
2
_
M
_
(k)e
ikx
dk, (B.8)
where M is the dimensionality of the space over which x is dened. The gradient of
can be obtained by dierentiating (B.8) with respect to the components of x
(x) =
_
1
2
_
M
_
(ik)(k)e
ikx
dk, (B.9)
and further dierentiation gives the Laplacian

2
(x) =
_
1
2
_
M
_
(k
2
)(k)e
ikx
dk. (B.10)
Expanding the exponential of the transformation (B.4) in a power series one can
see that
e
(tt

)

D
=
_
1
2
_
M
_
e
i(tt

)k
2
(k)e
ikx
dk. (B.11)
This transformation is easily carried out numerically using the FFTW routine and
its inverse IFFTW, a very ecient implementation of a Fast Fourier Transformation,
giving
e
(tt

)

D
= IFFTW
_
e
i(tt

)k
2
FFTW()
_
. (B.12)
190
B.1. GrossPitaevskii Equation Propagation
B.1.3 Fourth-order Runge Kutta
Runge-Kutta methods are designed to integrate a system of dierential equations
dy
dt
= f (y, t) (B.13)
stepwise from y
i
at time t
i
to y
i+1
at time t
i+1
= t
i
+ t. The fourth-order algorithm
requires the evaluation of four coecients
k
1
= tf (y
i
, t
i
), (B.14)
k
2
= tf (y
i
+k
1
/2, t
i
+ t/2), (B.15)
k
3
= tf (y
i
+k
2
/2, t
i
+ t/2), (B.16)
k
4
= tf (y
i
+k
3
, t
i
+ t) (B.17)
to construct the solution
y
i+1
= y
i
+ [k
1
+ 2(k
2
+k
3
) +k
4
]/6. (B.18)
We can see that two of the coecients are evaluated at the mid-point t
i
+ t/2.
When we apply the Runge-Kutta method to our problem (B.5) we can greatly reduce
the number of necessary Fourier transformations by moving the arbitrary origin t

of
the interaction picture (B.4) at each step of the Runge-Kutta algorithm to
t

= t
i
+ t/2. (B.19)
This optimizes the speed of the propagation algorithm because the Fourier transforma-
tions are the most time-consuming part.
Putting everything together, the following steps have to be performed to propagate

i
to
i+1

I
= e
(t/2)

D

i
, (B.20)
k
1
= te
(t/2)

D

N
i
, (B.21)
k
2
= t

N(
I
+k
1
/2), (B.22)
k
3
= t

N(
I
+k
2
/2), (B.23)
k
4
= t

Ne
(t/2)

D
(
I
+k
3
), (B.24)

i+1
= e
(t/2)

D

I
+ [k
1
+ 2(k
2
+k
3
)]/6 +k
4
/6. (B.25)
191
Appendix B. Propagation Methods For Dynamical Simulations
B.2 Phenomenological Growth Equation with Ro-
tating Thermal Cloud
The only problem that arises applying the RK4IP algorithm to the phenomenological
growth equation with a rotating thermal cloud
(i )
(r, t)
t
=
_

2
+V
T
+C[(r, t)[
2
+i ( +L
z
)
_
(r, t) (B.26)
is the question how to handle the operator

L
z
= i(x
y
y
x
). (B.27)
One might think at rst that it could be included in the operator

D of equation (B.1).
However, since

L
z
is not diagonal in Fourier space, there would be no straightforward
way to calculate e
(tt

)

D
for the transformation into the interaction picture. Hence,
the angular momentum term has to be included in the non-linear operator

N. With
the denitions

D =
1
i

2
, (B.28)

N =
1
i
[V
T
+C[[
2
+i( +

L
z
)], (B.29)
the RK4IP algorithm can be applied in exactly the same way as described in equations
(B.20)-(B.25).
The chemical potential of the thermal cloud and its rotation angular frequency
are scalars and, therefore, do not pose any problem. The only task that remains is
to evaluate

L
z
whenever

N acts on the wave function . In principle, that could be
done using Fourier methods equivalent to (B.9). However, this would require additional
Fourier transformations essentially nullifying the advantage we won choosing the origin
of the interaction picture at mid-point of every time step. Hence, we decided to use nite
dierences to calculate the derivatives occuring in the angular momentum operator

L
z
.
To illustrate the nite dierence approximation of a derivative, let us consider a
one dimensional discretized function f
i
where f
i
f(x
i
) is the function on the discrete
uniformly spaced grid x
i
with grid spacing x = x
i+1
x
i
. The derivative at point x
i
192
B.2. Phenomenological Growth Equation with Rotating Thermal Cloud
can be approximated by a weighted sum
f

i
=
1
x
r

k=l
c
k
f
i+k
, (B.30)
where l and r are integers with l 0 r. The vector of the weights c is called a nite
dierence stencil. We only used stencils in a central conguration, i.e. with [l[ = [r[ so
that the same number of function values on either side of x
i
are employed to calculate
the derivative at position x
i
. In general, a central stencil gives more accurate results
than a forward or backward conguration with the same number of points in the stencil.
We used derivative operators of up to 13 points whose stencils are given by
c
(3)
=
_

1
2
, 0,
1
2
_
, (B.31)
c
(5)
=
_
1
12
,
2
3
, 0,
2
3
,
1
12
_
, (B.32)
c
(7)
=
_

1
60
,
3
20
,
3
4
, 0,
3
4
,
3
20
,
1
60
_
, (B.33)
c
(9)
=
_
1
280
,
4
105
,
1
5
,
4
5
, 0,
4
5
,
1
5
,
4
105
,
1
280
_
, (B.34)
c
(11)
=
_

1
1260
,
5
504
,
5
84
,
5
21
,
5
6
, 0,
5
6
,
5
21
,
5
84
,
5
504
,
1
1260
_
, (B.35)
c
(13)
=
_
1
5544
,
1
385
,
1
56
,
5
63
,
15
56
,
6
7
, 0,
6
7
,
15
56
,
5
63
,
1
56
,
1
385
,
1
5544
_
,
(B.36)
which we calculated using a method described by Blakie in his PhD thesis [174]. He
has shown that this achieves the same order of accuracy as a Fourier method, but is
appreciably faster.
193
References
[1] M. H. Anderson, J. R. Ensher, M. R. Matthews, C. E. Wieman, and E. A. Cornell,
Observation of BoseEinstein Condensation in a Dilute Atomic Vapor, Science
269, 198 (1995).
[2] 1997 Nobel prize in Physics, http://www.nobel.se/physics/laureates/1997/illpres.
[3] 2001 Nobel prize in Physics, http://www.nobel.se/physics/laureates/2001/illpres.
[4] C. C. Bradley, C. A. Sackett, J. J. Tollett, and R. G. Hulet, Evidence of Bose
Einstein Condensation in an Atomic Gas with Attractive Interactions, Phys.
Rev. Lett. 75, 1687 (1995).
[5] S. L. Cornish, N. R. Claussen, J. L. Roberts, E. A. Cornell, and C. E. Wie-
man, Stable
85
Rb BoseEinstein Condensates with Widely Tunable Interac-
tions, Phys. Rev. Lett. 85, 1795 (2000).
[6] G. Modugno, G. Ferrari, G. Roati, R. J. Brecha, A. Simoni, and M. Inguscio,
BoseEinstein Condensation of Potassium Atoms by Sympathetic Cooling, Sci-
ence 294, 1320 (2001).
[7] T. Weber, J. Herbig, M. Mark, H.-C. Nagerl, and R. Grimm, BoseEinstein
Condensation of Cesium, Science 299, 232 (2003).
[8] D. G. Fried, T. C. Killian, L. Willmann, D. Landhuis, S. C. Moss, D. Kleppner,
and T. J. Greytak, BoseEinstein Condensation of Atomic Hydrogen, Phys.
Rev. Lett. 81, 3811 (1998).
[9] A. Robert, O. Sirjean, A. Browaeys, J. Poupard, S. Nowak, D. Boiron, C. I.
Westbrook, and A. Aspect, A BoseEinstein Condensate of Metastable Atoms,
Science 292, 461 (2001).
195
References
[10] F. P. D. Santos, J. Leonard, J. Wang, C. J. Barrelet, F. Perales, E. Rasel, C. S.
Unnikrishnan, M. Leduc, and C.Cohen-Tannoudji, BoseEinstein Condensation
of Metastable Helium, Phys. Rev. Lett. 86, 3459 (2001).
[11] F. Dalfovo, S. Giorgini, L. P. Pitaevskii, and S. Stringari, Theory of Bose
Einstein condensation in trapped gases, Rev. Mod. Phys. 71, 463 (1999).
[12] A. J. Leggett, BoseEinstein condensation in the alkali gases: Some fundamental
concepts, Rev. Mod. Phys. 73, 307 (2001).
[13] C. J. Pethick and H. Smith, BoseEinstein Condensation in Dilute Gases (Cam-
bridge University Press, 2001).
[14] S. Chu, Cold atoms and quantum control, Nature 416, 206 (2002).
[15] J. R. Anglin and W. Ketterle, BoseEinstein condensation of atomic gases,
Nature 416, 211 (2002).
[16] S. L. Rolston and W. D. Phillips, Nonlinear and quantum atom optics, Nature
416, 219 (2002).
[17] K. Burnett, P. S. Julienne, P. D. Lett, E. Tiesinga, and C. J. Williams, Quantum
encounters of the cold kind, Nature 416, 225 (2002).
[18] T. Udem, R. Holzwarth, and T. W. Hansch, Optical frequency metrology, Na-
ture 416, 233 (2002).
[19] C. Monroe, Quantum information processing with atoms and photons, Nature
416, 238 (2002).
[20] D. S. Hall, Resource Letter: BEC-1: BoseEinstein condensates in trapped dilute
gases, Am. J. Phys. 71, 649 (2003).
[21] M.-O. Mewes, M. R. Andrews, N. J. van Druten, D. M. Kurn, D. S. Durfee, and
W. Ketterle, BoseEinstein Condensation in a Tightly Conning dc Magnetic
Trap, Phys. Rev. Lett. 77, 416 (1996).
[22] D. S. Hall, M. R. Matthews, C. E. Wieman, and E. A. Cornell, Measurements
of Relative Phase in Two-Component BoseEinstein Condensates, Phys. Rev.
Lett. 81, 1543 (1998).
196
References
[23] D. S. Hall, M. R. Matthews, J. R. Ensher, C. E. Wieman, and E. A. Cornell,
Dynamics of Component Separation in a Binary Mixture of BoseEinstein Con-
densates, Phys. Rev. Lett. 81, 1539 (1998).
[24] M. R. Matthews, D. S. Hall, D. S. Jin, J. R. Ensher, C. E. Wieman, E. A. Cornell,
F. Dalfovo, C. Minniti, and S. Stringari, Dynamical Response of a BoseEinstein
Condensate to a Discontinuous Change in Internal State, Phys. Rev. Lett. 81,
243 (1998).
[25] M. R. Matthews, B. P. Anderson, P. C. Haljan, D. S. Hall, M. J. Holland, J. E.
Williams, C. E. Wieman, and E. A. Cornell, Watching a Superuid Untwist
Itself: Recurrence of Rabi Oscillations in a BoseEinstein Condensate, Phys.
Rev. Lett. 83, 3358 (1999).
[26] D. M. Stamper-Kurn, M. R. Andrews, A. P. Chikkatur, S. Inouye, H.-J. Miesner,
J. Stenger, and W. Ketterle, Optical Connement of a BoseEinstein Conden-
sate, Phys. Rev. Lett. 80, 2027 (1998).
[27] J. Stenger, D. M. Stamper-Kurn, M. R. Andrews, A. P. Chikkatur, S. Inouye,
H.-J. Miesner, and W. Ketterle, Optically conned BoseEinstein condensates,
J. Low Temp. Phys. 113, 167 (1998).
[28] M. D. Barrett, J. A. Sauer, and M. S. Chapman, All-optical formation of an
atomic BoseEinstein condensate, Phys. Rev. Lett. 87, 010404 (2001).
[29] H. Ott, J. Fortagh, G. Schlotterbeck, A. Grossmann, and C. Zimmermann, Bose
Einstein condensation in a surface microtrap,Phys. Rev. Lett. 87, 230401 (2001).
[30] W. Hansel, P. Hommelho, T. W. Hansch, and J. Reichel, BoseEinstein con-
densation on a microelectronic chip, Nature 413, 498 (2001).
[31] D. M. M uller, E. A. Cornell, M. Prevedelli, P. D. D. Schwindt, A. Zozulya, and
D. Z. Anderson, A waveguide atom beamsplitter for laser-cooled neutral atoms,
Optics Lett. 25, 1382 (2000).
[32] A. Gorlitz, J. M. Vogels, A. E. Leanhardt, C. Raman, T. Gustavson, J. R. Abo-
Shaeer, A. P. Chikkatur, S. Gupta, S. Inouye, T. Rosenband, and W. Ketterle,
Realization of BoseEinstein condensates in lower dimensions, Phys. Rev. Lett.
87, 130402 (2001).
197
References
[33] A. I. Safonov, S. A. Vasilyev, I. S. Yasnikov, I. I. Lukashevich, and S. Jaakkola,
Observation of Quasicondensate in Two-Dimensional Atomic Hydrogen, Phys.
Rev. Lett. 81, 4545 (1998).
[34] S. Dettmer, D. Hellweg, P. Ryytty, J. J. Arlt, W. Ertmer, and K. Sengstock, Ob-
servation of phase uctuations in elongated BoseEinstein condensates, Phys.
Rev. Lett. 87, 160406 (2001).
[35] B. P. Anderson and M. A. Kasevich, Macroscopic Quantum Interference from
Atomic Tunnel Arrays, Science 282, 1686 (1998).
[36] C. Orzel, A. K. Tuchman, M. L. Fenselau, M. Yasuda, and M. A. Kasevich,
Squeezed States in a BoseEinstein Condensate, Science 291, 2386 (2001).
[37] M. Greiner, O. Mandel, T. Esslinger, T. W. Hansch, and I. Bloch, Quantum
phase transition from a superuid to a Mott insulator in a gas of ultracold atoms,
Nature 415, 39 (2002).
[38] Y. B. Ovchinnikov, J. H. M uller, M. R. Doery, E. J. D. Vredenbregt, K. Helmer-
son, S. L. Rolston, and W. D. Phillips, Diraction of a Released BoseEinstein
Condensate by a Pulsed Standing Light Wave, Phys. Rev. Lett. 83, 284 (1999).
[39] M. Kozuma, L. Deng, E. W. Hagley, J. Wen, R. Lutwak, K. Helmerson, S. L.
Rolston, and W. D. Phillips, Coherent Splitting of BoseEinstein Condensed
Atoms with Optically Induced Bragg Diraction,Phys. Rev. Lett. 82, 871 (1999).
[40] R. J. Ballagh and C. M. Savage, The Theory of Atom Lasers, in BoseEinstein
condensation: From atomic physics to quantum uids, edited by C. M. Savage
and M. P. Das, volume 153 (World Scientic, Singapore, 2001).
[41] M. R. Andrews, C. G. Townsend, H.-J. Miesner, D. S. Durfee, D. M. Kurn, and
W. Ketterle, Observation of Interference Between Two BoseEinstein Conden-
sates, Science 275, 637 (1997).
[42] I. Bloch, T. W. Hansch, and T. Esslinger, Measurement of the spatial coherence
of a trapped Bose gas at the phase transition, Nature 403, 166 (2000).
[43] M.-O. Mewes, M. R. Andrews, D. M. Kurn, D. S. Durfee, C. G. Townsend, and
W. Ketterle, Output Coupler for BoseEinstein Condensed Atoms, Phys. Rev.
Lett. 78, 582 (1997).
198
References
[44] J. L. Martin, C. R. McKenzie, N. R. Thomas, J. C. Sharpe, D. M. Warrington,
P. J. Manson, W. J. Sandle, and A. C. Wilson, Output coupling of a Bose
Einstein condensate formed in a TOP trap, J. Phys. B 32, 3065 (1999).
[45] E. W. Hagley, L. Deng, M. Kozuma, J. Wen, K. Helmerson, S. L. Rolston, and
W. D. Phillips, A Well-Collimated Quasi-Continuous Atom Laser, Science 283,
1706 (1999).
[46] I. Bloch, T. W. Hansch, and T. Esslinger, Atom Laser with a cw Output Cou-
pler, Phys. Rev. Lett. 82, 3008 (1999).
[47] A. P. Chikkatur, Y. Shin, A. E. Leanhardt, D. Kielpinski, E. Tsikata, T. L.
Gustavson, D. E. Pritchard, and W. Ketterle, A Continuous Source of Bose
Einstein Condensed Atoms, Science 296, 2193 (2002).
[48] T. L. Gustavson, A. P. Chikkatur, A. E. Leanhardt, A. Gorlitz, A. Gupta, D. E.
Pritchard, and W. Ketterle, Transport of BoseEinstein condensates with optical
tweezers, Phys. Rev. Lett. 88, 020401 (2001).
[49] L. Deng, E. W. Hagley, J. Wen, M. Trippenbach, Y. Band, P. S. Julienne, J. E.
Simsarian, K. Helmerson, S. L. Rolston, and W. D. Phillips, Four-wave mixing
with matter waves, Nature 398, 218 (1999).
[50] S. Inouye, A. P. Chikkatur, D. M. Stamper-Kurn, J. Stenger, D. E. Pritchard,
and W. Ketterle, Superradiant Rayleigh Scattering from a BoseEinstein Con-
densate, Science 285, 571 (1999).
[51] M. Kozuma, Y. Suzuki, Y. Torii, T. Sugiura, T. Kuga, E. Hagley, and L. Deng,
PhaseCoherent Amplication of Matter Waves, Science 286, 2309 (1999).
[52] S. Inouye, T. Pfau, S. Gupta, A. P. Chikkatur, A. Gorlitz, D. E. Pritchard, and
W. Ketterle, Phase-coherent amplication of atomic matter waves,Nature 402,
641 (1999).
[53] S. Burger, K. Bongs, S. Dettmer, W. Ertmer, K. Sengstock, A. Sanpera, G. V.
Shlyapnikov, and M. Lewenstein, Dark Solitons in BoseEinstein Condensates,
Phys. Rev. Lett. 83, 5198 (1999).
[54] J. Denschlag, J. E. Simsarian, D. L. Feder, C. W. Clark, L. A. Collins, J. Cu-
bizolles, L. Deng, E. W. Hagley, K. Helmerson, W. P. Reinhardt, S. L. Rolston,
199
References
B. I. Schneider, and W. D. Phillips, Generating Solitons by Phase Engineering
of a BoseEinstein Condensate, Science 287, 97 (2000).
[55] L. V. Hau, S. E. Harris, Z. Dutton, and C. H. Behroozi, Light speed reduction
to 17 metres per second in an ultracold atomic gas, Nature 397, 594 (1999).
[56] C. Liu, Z. Dutton, C. H. Behroozi, and L. V. Hau, Observation of coherent op-
tical information storage in an atomic medium using halted light pulses, Nature
409, 490 (2001).
[57] J. M. Gerton, D. Strekalov, I. Prodan, and R. G. Hulet, Direct observation of
growth and collapse of a BoseEinstein condensate with attractive interactions,
Nature 408, 692 (2000).
[58] C. C. Bradley, C. A. Sackett, and R. G. Hulet, BoseEinstein Condensation of
Lithium: Observation of Limited Condensate Number, Phys. Rev. Lett. 78, 985
(1997).
[59] S. Inouye, M. R. Andrews, J. Stenger, H.-J. Miesner, D. M. Stamper-Kurn, and
W. Ketterle, Observation of Feshbach resonances in a BoseEinstein conden-
sate, Nature 392, 151 (1998).
[60] E. A. Donley, N. R. Claussen, S. L. Cornish, J. L. Roberts, E. A. Cornell, and
C. E. Wieman, Dynamics of a collapsing and exploding BoseEinstein conden-
sates, Nature 412, 295 (2001).
[61] J. L. Roberts, N. R. Claussen, S. L. Cornish, E. A. Donley, E. A. Cornell, and
C. E. Wieman, Controlled Collapse of a BoseEinstein Condensate, Phys. Rev.
Lett. 86, 4211 (2001).
[62] L. Khaykovich, F. Schreck, G. Ferrari, T. Bourdel, J. Cubizolles, L. D. Carr,
Y. Castin, and C. Salomon, Formation of a matter-save bright soliton, Science
296, 1290 (2002).
[63] R. Wynar, R. S. Freeland, D. J. Han, C. Ryu, and D. J. Heinzen, Molecules in
a BoseEinstein Condensate, Science 287, 1016 (2000).
[64] C. McKenzie, J. H. Denschlag, H. Haner, A. Browaeys, L. E. E. de Araujo,
F. K. Fatemi, K. M. Jones, J. E. Simsarian, D. Cho, A. Simoni, E. Tiesinga,
P. S. Julienne, K. Helmerson, P. D. Lett, S. L. Rolston, and W. D. Phillips,
200
References
Photoassociation of sodium in a BoseEinstein condensate, Phys. Rev. Lett.
88, 120403 (2002).
[65] E. A. Donley, N. R. Claussen, S. T. Thompson, and C. E. Wieman, Atom-
molecule coherence in a BoseEinstein condensate, Nature 417, 529 (2002).
[66] C. Raman, M. Kohl, R. Onofrio, D. S. Durfee, C. E. Kuklewicz, Z. Hadzibabic,
and W. Ketterle, Evidence for a Critical Velocity in a BoseEinstein Condensed
Gas, Phys. Rev. Lett. 83, 2502 (1999).
[67] R. Onofrio, C. Raman, J. M. Vogels, J. R. Abo-Shaeer, A. P. Chikkatur, and
W. Ketterle, Observation of Superuid Flow in a BoseEinstein Condensed Gas,
Phys. Rev. Lett. 85, 2228 (2000).
[68] A. P. Chikkatur, A. Gorlitz, D. M. Stamper-Kurn, S. Inouye, S. Gupta, and
W. Ketterle, Suppression and Enhancement of Impurity Scattering in a Bose
Einstein Condensate, Phys. Rev. Lett. 85, 483 (2000).
[69] S. Burger, F. S. Cataliotti, C. Fort, F. Minardi, M. Inguscio, M. L. Chiofalo, and
M. P. Tosi, Superuid and Dissipative Dynamics of a BoseEinstein Condensate
in a Periodic Optical Potential, Phys. Rev. Lett. 86, 4447 (2001).
[70] O. M. Marago, S. A. Hopkins, J. Arlt, E. Hodby, G. Heckenblaikner, and C. J.
Foot, Observation of the Scissors Mode and Evidence for Superuidity of a
Trapped BoseEinstein Condensed Gas, Phys. Rev. Lett. 84, 2056 (2000).
[71] G. Hechenblaikner, E. Hodby, S. A. Hopkins, O. M. Marag`o, and C. J. Foot, Di-
rect Observation of Irrotational Flow and Evidence of Superuidity in a Rotating
BoseEinstein Condensate, Phys. Rev. Lett. 88, 070406 (2002).
[72] M. R. Matthews, B. P. Anderson, P. C. Haljan, D. S. Hall, C. E. Wieman, and
E. A. Cornell, Vortices in a BoseEinstein Condensate, Phys. Rev. Lett. 83,
2498 (1999).
[73] K. W. Madison, F. Chevy, W. Wohlleben, and J. Dalibard, Vortex Formation
in a Stirred BoseEinstein Condensate, Phys. Rev. Lett. 84, 806 (2000).
[74] F. Chevy, K. W. Madison, and J. Dalibard, Measurement of the Angular Mo-
mentum of a Rotating BoseEinstein Condensate, Phys. Rev. Lett. 85, 2223
(2000).
201
References
[75] F. Chevy, K. Madison, V. Bretin, and J. Dalibard, Formation of quantized vor-
tices in a gaseous BoseEinstein condensate, e-print, cond-mat/0104218 (2001).
[76] K. W. Madison, F. Chevy, V. Bretin, and J. Dalibard, Stationary States of a
Rotating BoseEinstein Condensate: Routes to Vortex Nucleation, Phys. Rev.
Lett. 86, 4443 (2001).
[77] C. Raman, J. R. Abo-Shaeer, J. M. Vogels, K. Xu, and W. Ketterle, Vortex
Nucleation in a Stirred BoseEinstein Condensate, Phys. Rev. Lett. 87, 210402
(2001).
[78] P. C. Haljan, I. Coddington, P. Engels, and E. A. Cornell, Driving BoseEinstein-
Condensate Vorticity with a Rotating Normal Cloud,Phys. Rev. Lett. 87, 210403
(2001).
[79] J. R. Abo-Shaeer, C. Raman, J. M. Vogels, and W. Ketterle, Observation of
Vortex Lattices in BoseEinstein Condensates, Science 292, 476 (2001).
[80] J. R. Abo-Shaeer, C. Raman, and W. Ketterle, Formation and Decay of Vortex
Lattices in BoseEinstein Condensates at Finite Temperatures, Phys. Rev. Lett.
88, 070409 (2002).
[81] P. Engels, I. Coddington, P. C. Haljan, and E. A. Cornell, Nonequilibrium Ef-
fects of Anisotropic Compression Applied to Vortex Lattices in BoseEinstein
Condensates, Phys. Rev. Lett. 89, 100403 (2002).
[82] I. Coddington, P. Engels, V. Schweikhard, and E. Cornell, Observation of
Tkachenko Oscillations in Rapidly Rotating BoseEinstein Condensates, Phys.
Rev. Lett. 91, 100402 (2003).
[83] P. Engels, I. Coddington, P. C. Haljan, V. Schweikhard, and E. A. Cornell, Ob-
servation of Long-lived Vortex Aggregates in Rapidly Rotating BoseEinstein
Condensates, Phys. Rev. Lett. 90, 170405 (2003).
[84] D. Hutchinson and E. Zaremba, Excitations of a Bose-condensed gas in
anisotropic traps, Phys. Rev. A 57, 1280 (1998).
[85] C. W. Gardiner, J. R. Anglin, and T. I. A. Fudge, The stochastic Gross
Pitaevskii equation, J. Phys. B 35, 1555 (2002).
202
References
[86] A. A. Penckwitt and R. J. Ballagh, Elementary excitation families and their fre-
quency ordering in cylindrically symmetric BoseEinstein condensates, J. Phys.
B: At. Mol. Opt. Phys. 34, 1523 (2001).
[87] A. A. Penckwitt, R. J. Ballagh, and C. W. Gardiner, Nucleation, Growth, and
Stabilization of BoseEinstein Condensate Vortex Lattices, Phys. Rev. Lett. 89,
260402 (2002).
[88] T. P. Simula, A. A. Penckwitt, and R. J. Ballagh, Giant Vortex Lattice Defor-
mations in Rapidly Rotating BoseEinstein Condensates, Phys. Rev. Lett. 92,
060401 (2004).
[89] D. A. Huse and E. D. Siggia, The density distribution of a weakly interacting
Bose gas in an external potential, J. Low Temp. Phys. 46, 137 (1982).
[90] C. W. Gardiner, P. Zoller, R. J. Ballagh, and M. J. Davis, Kinetics of Bose
Einstein Condensation in a Trap, Phys. Rev. Lett. 79, 1793 (1997).
[91] J. P. Blaizot and G. Ripka, Quantum Theory of Finite Systems (MIT Press,
1986).
[92] K. Huang, Statistical Mechanics, second edition (John Wiley & Sons, Inc., 1987).
[93] S. A. Morgan, A gapless theory of BoseEinstein condensation in dilute gases at
nite temperature, J. Phys. B 33, 3847 (2000).
[94] C. W. Gardiner, Particle-number-conserving Bogoliubov method which demon-
strates the validity of the time-dependent GrossPitaevskii equation for a highly
condensed Bose gas, Phys. Rev. A 56, 1414 (1997).
[95] D. Hutchinson, E. Zaremba, and A. Grin, Finite Temperature Excitations of
a Trapped Bose Gas, Phys. Rev. Lett. 78, 1842 (1997).
[96] M. J. Davis, Dynamics of BoseEinstein condensation, Ph.D. thesis, University
of Oxford (2001).
[97] A. Grin, Conserving and gapless approximations for an inhomogeneous Bose
gas at nite temperatures, Phys. Rev. B 53, 9341 (1996).
[98] N. P. Proukakis, S. A. Morgan, S. Choi, and K. Burnett, Comparison of gapless
mean-eld theories for trapped BoseEinstein condensates, Phys. Rev. A 58,
2435 (1998).
203
References
[99] A. L. Fetter, Nonuniform States of an Imperfect Bose Gas, Annal. Phys. 70, 67
(1972).
[100] S. A. Morgan, S. Choi, and K. Burnett, Nonlinear mixing of quasiparticles in an
inhomogeneous Bose condensate, Phys. Rev. A 57, 3818 (1998).
[101] M. Edwards, R. J. Dodd, C. W. Clark, and K. Burnett, Zero-Temperature,
Mean-Field Theory of Atomic BoseEinstein Condensates, J. Res. Natl. Inst.
Stand. Tech. 101, 553 (1996).
[102] A. A. Penckwitt, Computational Investigation of the Elementary Excitations of
BoseEinstein Condensates, Masters thesis, University of Otago (1999).
[103] C. W. Gardiner, private communication.
[104] L. P. Eisenhart, Enumeration of Potentials for Which OneParticle Schrodinger
Equations Are Separable, Phys. Rev. 74, 87 (1948).
[105] R. J. Dodd, K. Burnett, M. Edwards, and C. W. Clark, Excitation spectroscopy
of vortex states in dilute BoseEinstein condensed gases, Phys. Rev. A 56, 587
(1997).
[106] M. Fliesser, A. Csordas, P. Szepfalusy, and R. Graham, Hydrodynamic excita-
tions of Bose condensates in anisotropic traps, Phys. Rev. A 56, R2533 (1997).
[107] S. Stringari, Collective Excitations of a Trapped Bose-Condensed Gas, Phys.
Rev. Lett. 77, 2360 (1996).
[108] L. You, W. Hoston, and M. Lewenstein, Lowenergy excitations of trapped Bose
condensates, Phys. Rev. A 55, R1581 (1997).
[109] F. Dalfovo, S. Giorgini, M. Guilleumas, L. Pitaevskii, and S. Stringari, Collective
and single-particle excitations of a trapped Bose gas, Phys. Rev. A 56, 3840
(1997).
[110] D. R. Tilley and J. Tilley, Superuidity and Superconductivity (Institute of
Physics Publishing, Bristol, 1990).
[111] A. L. Fetter and A. A. Svidzinsky, Vortices in a trapped dilute BoseEinstein
condensate, J. Phys. 13, R135 (2001).
204
References
[112] A. Recati, F. Zambelli, and S. Stringari, Overcritical Rotation of a Trapped
BoseEinstein Condensate, Phys. Rev. Lett. 86, 377 (2001).
[113] E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics, Part 2 (Pergamon Press,
Oxford, 1980).
[114] B. Jackson, J. F. McCann, and C. S. Adams, Dissipation and vortex creation in
BoseEinstein condensed gases, Phys. Rev. A 61, 051603(R) (2000).
[115] D. S. Rokhsar, Vortex Stability and Persistent Currents in Trapped Bose Gases,
Phys. Rev. Lett. 79, 2164 (1997).
[116] T. Isoshima and K. Machida, Vortex stabilization in BoseEinstein condensate
of alkali-metal atom gas, Phys. Rev. A 59, 2203 (1999).
[117] T. Isoshima and K. Machida, Instability of the nonvortex state toward a quan-
tized vortex in a BoseEinstein condensate under external rotation, Phys. Rev.
A 60, 3313 (1999).
[118] A. A. Svidzinsky and A. L. Fetter, Stability of a Vortex in a Trapped Bose
Einstein Condensate, Phys. Rev. Lett. 84, 5919 (2000).
[119] D. L. Feder and C. W. Clark, Superuid-to-Solid Crossover in a Rotating Bose
Einstein Condensate, Phys. Rev. Lett. 87, 190401 (2001).
[120] A. Svidzinsky and A. Fetter, Dynamics of a Vortex in a Trapped Bose-Einstein
Condensate, Phys. Rev. A 62, 063617 (2000).
[121] D. L. Feder, C. W. Clark, and B. I. Schneider, Vortex Stability of Interacting
BoseEinstein Condensates Conned in Anisotropic Harmonic Traps,Phys. Rev.
Lett. 82, 4956 (1999).
[122] B. Jackson, J. F. McCann, and C. S. Adams, Vortex line and ring dynamics in
trapped BoseEinstein condensates, Phys. Rev. A 61, 013604 (1999).
[123] K. W. Madison, F. Chevy, W. Wohlleben, and J. Dalibard, Vortices in a stirred
BoseEinstein condensate, J. Mod. Opt. 47, 2715 (2000).
[124] P. O. Fedichev and G. V. Shlyapnikov, Critical velocity in cylindrical Bose
Einstein condensates, Phys. Rev. A 63, 045601 (2001).
205
References
[125] A. E. Muryshev and P. O. Fedichev, Surface modes and critical velocity in
trapped Bose-condensates, e-print, cond-mat/0106462 (2001).
[126] J. R. Anglin, Local Vortex Generation and the Surface Mode Spectrum of Large
BoseEinstein Condensates, Phys. Rev. Lett. 87, 240401 (2001).
[127] J. R. Anglin, Vortices near surfaces of BoseEinstein condensates, Phys. Rev.
A 65, 063611 (2002).
[128] D. L. Feder, A. A. Svidzinsky, A. L. Fetter, and C. W. Clark, Anomalous Modes
Drive Vortex Dynamics in Conned BoseEinstein Condensates,Phys. Rev. Lett.
86, 564 (2001).
[129] S. Sinha and Y. Castin, Dynamic Instability of a Rotating BoseEinstein Con-
densate, Phys. Rev. Lett. 87, 190402 (2001).
[130] F. Dalfovo and S. Stringari, Shape deformations and angular-momentum transfer
in trapped BoseEinstein condensates, Phys. Rev. A 63, 011601 (R) (2001).
[131] A. L. Fetter and J.-K. Kim, Vortex Precession in a Rotating Nonaxisymmetric
Trapped BoseEinstein Condensate, J. Low Temp. Phys. 125, 239 (2001).
[132] V. K. Tkachenko, On vortex lattices, Sov. Phys. JETP 22, 1282 (1966).
[133] V. K. Tkachenko, Stability of vortex lattices,Sov. Phys. JETP 23, 1049 (1966).
[134] A. A. Abrikosov, On the magnetic properties of superconductors of the second
type, Sov. Phys. JETP 5, 1174 (1957).
[135] D. A. Butts and D. S. Rokhsar, Predicted signatures of rotating BoseEinstein
condensates, Nature 397, 327 (1999).
[136] D. L. Feder, C. W. Clark, and B. I. Schneider, Nucleation of vortex arrays in
rotating anisotropic BoseEinstein condensates, Phys. Rev. A 61, 011601(R)
(1999).
[137] Y. Castin and R. Dum, BoseEinstein condensates with vortices in rotating
traps, Euro. Phys. J. D 7, 399 (1999).
[138] G. M. Kavoulakis, B. Mottelson, and C. J. Pethick, Weakly interacting Bose
Einstein condensates under rotation, Phys. Rev. A 62, 063605 (2000).
206
References
[139] C. W. Gardiner and P. Zoller, Quantum kinetic theory: A quantum kinetic
master equation for condensation of a weakly interacting Bose gas without a
trapping potential, Phys. Rev. A 55, 2902 (1997).
[140] D. Jaksch, C. W. Gardiner, and P. Zoller, Quantum kinetic theory. II. Simulation
of the quantum Boltzmann master equation, Phys. Rev. A 56, 575 (1997).
[141] C. W. Gardiner and P. Zoller, Quantum kinetic theory. III. Quantum kinetic
master equation for strongly condensed trapped systems, Phys. Rev. A 58, 536
(1998).
[142] D. Jaksch, C. W. Gardiner, K. M. Gheri, and P. Zoller, Quantum kinetic theory.
IV. Intensity and amplitude uctuations of a BoseEinstein condensate at nite
temperature including trap loss, Phys. Rev. A 58, 1450 (1998).
[143] C. W. Gardiner and P. Zoller, Quantum kinetic theory. V. Quantum kinetic
master equation for mutual interaction of condensate and noncondensate, Phys.
Rev. A 61, 033601 (2000).
[144] M. D. Lee and C. W. Gardiner, Quantum kinetic theory. VI. The growth of a
BoseEinstein condensate, Phys. Rev. A 62, 033606 (2000).
[145] M. J. Davis, C. W. Gardiner, and R. J. Ballagh, Quantun kinetic theory: VII.
The inuence of vapor dynamics on condensate growth,Phys. Rev. A 62, 063608
(2000).
[146] C. W. Gardiner, M. D. Lee, R. J. Ballagh, M. J. Davis, and P. Zoller, Quantum
Kinetic Theory of Condensate Growth: Comparison of Experiment and Theory,
Phys. Rev. Lett. 81, 5266 (1998).
[147] M. J. Davis and C. W. Gardiner, Growth of a BoseEinstein condensate: a
detailed comparison of theory and experiment, J. Phys. B 35, 733 (2002).
[148] H. T. C. Stoof, Coherent Versus Incoherent Dynamics During BoseEinstein
Condensation in Atomic Gases, J. Low Temp. Phys. 114, 11 (1999).
[149] E. Zaremba, A. Grin, and T. Nikuni, Two-uid hydrodynamics for a trapped
weakly-interacting Bose gas, Phys. Rev. A 57, 4695 (1998).
[150] E. Zaremba, T. Nikuni, and A. Grin, Dynamics of Trapped Bose Gases at
Finite Temperatures, J. Low Temp. Phys. 116, 277 (1999).
207
References
[151] R. Walser, J. Williams, J. Cooper, and M. Holland, Quantum kinetic theory for
a condensed bosonic gas, Phys. Rev. A 59, 3878 (1999).
[152] E. Hodby, G. Hechenblaikner, S. A. Hopkins, O. M. Marag`o, and C. J. Foot,
Vortex nucleation in BoseEinstein condensates in an oblate, purely magnetic
potential, Phys. Rev. Lett. 88, 010405 (2002).
[153] M. Tsubota, K. Kasamatsu, and M. Ueda, Vortex lattice formation in a rotating
BoseEinstein condensate, Phys. Rev. A 65, 023603 (2002).
[154] T. Nikuni, E. Zaremba, and A. Grin, Two-Fluid Dynamics for a BoseEinstein
Condensate out of Local Equilibrium with the Noncondensate, Phys. Rev. Lett.
83, 10 (1999).
[155] B. Jackson and E. Zaremba, Modeling BoseEinstein condensed gases at nite
temperature with N-body simulations, Phys. Rev. A 66, 033606 (2002).
[156] A. Grin, private communication.
[157] C. Lobo, A. Sinatra, and Y. Castin, Vortex lattice formation in BoseEinstein
condensates, Phys. Rev. Lett. 92, 020403 (2004).
[158] A. Sinatra and Y. Castin, A Monte Carlo formulation of the Bogolubov theory,
J. Mod. Optics 47, 2629 (2000).
[159] A. Sinatra, C. Lobo, and Y. Castin, A classical eld method for time dependent
Bose condensed gases, Phys. Rev. Lett. 87, 210404 (2001).
[160] A. Sinatra, C. Lobo, and Y. Castin, The truncated Wigner or classical eld
method for Bose condensed gases: limits of validity and applications, J. Phys.
B 35, 3599 (2002).
[161] M. J. Davis, S. A. Morgan, and K. Burnett, Simulations of Bose elds at nite
temperature, Phys. Rev. Lett. 87, 160402 (2001).
[162] M. J. Davis, S. A. Morgan, and K. Burnett, Simulations of thermal Bose elds
in the classical limit, Phys. Rev. A 66, 053618 (2002).
[163] C. W. Gardiner and M. J. Davis, The stochastic GrossPitaevskii equation: II,
J. Phys. B 36, 4731 (2003).
208
References
[164] M. J. Davis, R. J. Ballagh, and K. Burnett, Dynamics of thermal Bose elds in
the classical limit, J. Phys. B 34, 4487 (2001).
[165] A. S. Bradley, private communication.
[166] R. J. Ballagh, Partial Dierential Equation Algorithm: Conceptual, unpub-
lished personal paper (1995).
[167] B. M. Caradoc-Davies, Vortex Dynamics in BoseEinstein Condensates, Ph.D.
thesis, University of Otago (2000).
[168] B. M. Caradoc-Davies, R. J. Ballagh, and K. Burnett, Coherent Dynamics of
Vortex Formation in Trapped BoseEinstein Condensates, Phys. Rev. Lett. 83,
895 (1999).
[169] J. M. W. Kr uger, Dynamical Behaviour of Radio Frequency Output Couplers for
a BoseEinstein Condensate, Masters thesis, University of Otago (2000).
[170] T. P. Simula, private communication.
[171] L. P. Pitaevskii and A. Rosch, Breathing modes and hidden symmetry of trapped
atoms in two dimensions, Phys. Rev. A 55, R853 (1997).
[172] K. Kasamatsu and M. Tsubota, Nonlinear dynamics of vortex lattice formation
in a rotating BoseEinstein condensate, Phys. Rev. A 67, 033610 (2003).
[173] T. Isoshima, J. Huhtamaki, and M. M. Salomaa, Instabilities of o-centered
vortices in a BoseEinstein condensate, Phys. Rev. A 68, 033611 (2003).
[174] P. B. Blakie, Optical Manipulation of BoseEinstein Condensates, Ph.D. thesis,
University of Otago (2001).
[175] B. P. Anderson, P. C. Haljan, C. E. Wieman, and E. A. Cornell, Vortex Preces-
sion in BoseEinstein Condensates: Observations with Filled and Empty Cores,
Phys. Rev. Lett. 85, 2857 (2000).
[176] S. Inouye, S. Gupta, T. Rosenband, A. P. Chikkatur, A. Gorlitz, T. L. Gustavson,
A. E. Leanhardt, D. E. Pritchard, and W. Ketterle, Observation of Vortex Phase
Singularities in BoseEinstein Condensates, Phys. Rev. Lett. 87, 080402 (2001).
[177] F. Zambelli and S. Stringari, Quantized Vortices and Collective Oscillations of
a Trapped BoseEinstein Condensate, Phys. Rev. Lett. 81, 1754 (1998).
209
References
[178] B. M. Caradoc-Davies, R. J. Ballagh, and P. B. Blakie, Three-dimensional vortex
dynamics in BoseEinstein condensates, Phys. Rev. A 62, 011602(R) (2000).
[179] O. N. Zhuravlev, A. E. Muryshev, and P. O. Fedichev, Dissipative dynamics of
vortex arrays in anisotropic traps, Phys. Rev. A 64, 053601 (2001).
[180] D. Guery-Odelin, Spinning up and down a Boltzmann gas, Phys. Rev. A 62,
033607 (2000).
[181] P. O. Fedichev and G. V. Shlyapnikov, Dissipative dynamics of a vortex state in
a trapped Bose-condensed gas, Phys. Rev. A 60, R1779 (1999).
[182] E. J. Mueller and T.-L. Ho, Stripe formation in BoseEinstein condensates with
large number of vortices, Phys. Rev. A 67, 063602 (2003).
[183] M. Linn and A. L. Fetter, Small-amplitude normal modes of a vortex in a trapped
BoseEinstein condensate, Phys. Rev. A 61, 063603 (2000).
[184] J. R. Anglin and M. Crescimanno, Inhomogeneous vortex matter,e-print, cond-
mat/0210063 (2002).
[185] G. Baym, Tkachenko modes of vortex lattices in rapidly rotating BoseEinstein
condensates, Phys. Rev. Lett. 91, 110402 (2003).
[186] L. O. Baksmaty, S. J. Woo, S. Choi, and N. P. Bigelow, Tkachenko waves in
rapidly rotating BoseEinstein condensates, Phys. Rev. Lett. 92, 160405 (2004).
[187] T. Mizushima, Y. Kawaguchi, K. Machida, T. Ohmi, T. Isoshima, and M. M.
Salomaa, Collective Oscillations of Vortex Lattices in Rotating BoseEinstein
Condensates, Phys. Rev. Lett. 92, 060407 (2004).
[188] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization (Academic
Press, 1981).
[189] K. H. Huebner, The Finite Element Method for Engineers (John Wiley & Sons,
Inc., 1975).
[190] D. Hutchinson, K. Burnett, R. J. Dodd, S. A. Morgan, M. Rusch, E. Zaremba,
N. P. Proukakis, M. Edwards, and C. W. Clark, Gapless mean-eld theory of
BoseEinstein condensates, J. Phys. B 33, 3825 (2000).
210
References
[191] C. Cohen-Tannoudji, B. Diu, and F. Laloe, Quantum mechanics, volume 1 (Wiley,
New York, 1992).
[192] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numeri-
cal Recipes in C, The Art of Scientic Computing, second edition (Cambridge
University Press, 1992).
211

Vous aimerez peut-être aussi