Computational Approach to Bohm's Quantum Mechanics

Computational Approach to Bohm's Quantum

Mechanics

by

PAULO ALEXANDRE MACHADO

B.Sc. (Oporto), M.Sc. (McMaster)

A Thesis

Submitted to the School of Graduate Studies

in Partial Fulfilment of the Requirements

for the Degree

Doctor of Philosophy

McMaster University

@Copyright by Paulo Alexandre Machado, 2007.

DOCTOR OF PHILOSOPHY (2007)

McMaster University

(Department of Physics and Astronomy)

Hamilton, Ontario

TITLE: Computational Approach to Bohm's Quantum Mechanics

AUTHOR: Paulo Alexandre Machado
SUPERVISOR: Prof. D. W. L. Sprung

NUMBER OF PAGES: xiv, 146

11

Abstract

Bohmian mechanics is an alternative formulation of quantum mechanics that in

corporates the familiar and intuitive picture of particles moving along trajectories and
yet predicts the same results as the more widely accepted Copenhagen interpretation.
In recent years there has been renewed interest in this Bohmian view, in part for
the novel approach that it suggests to certain problems, such as decay processes, both
from a theoretical and computational stand point. In this thesis we focus on using the
concepts introduced by the Bohmian framework as a practical computational tool.
I evaluate a number of implementations of the Bohmian method, get a sense of
their strengths and weaknesses and attempt to overcome some stability issues that
arise. For problems in one-dimension (lD), accurate solutions of the time-dependent
Schrodinger equation produce a wave function from which Bohmian trajectories can
be computed by integrating along flux lines. For direct integration of the quantum
Hamilton-Jacobi equations, the main problems that arise are related to evauating the
quantum potential (QP), especially in regions of low probability density. Sufficient
accuracy is required to avoid unphysical trajectory crossings. A number of interpola
tion schemes were investigated, and smoothed splines with special treatment of edge
effects gave the best results.
For problems in 2D the alternating direction implicit (ADI) method was employed
to produce the wave function. Ways of dealing with unphysical reflections from the
boundaries of a finite size domain were studied.
iii

The use of cellular automata, especially the lattice-Boltzmann method (LBM)

were also considered. Here Bohm trajectories would be propagated by following a
small set of rules. The main problem identified is that, unless a scheme can be found
in which the quantum potential is self-generating from an equation of continuity, the
overhead of computing the QP at each time step, is prohibitive.

IV

Acknowledgements
I would like to express my gratitude to my supervisor D. W. L. Sprung for the
thoughtful guidance, kindness, patience and constructive criticism that I enjoyed over
the years under his supervision.
I also thank Y. Nogami and D. Pelinovsky for their helpful suggestions and interest
in the work as well as the encouragement given by R. Dumont.
This work would not be possible without the positive learning environment and
financial support provided by McMaster University.
I would also like to thank Hannah Sprung, Rajat and Manju Bhaduri as well as
Natalia and Tanya Wolf, Walter Phillips and Errol Rennie for keeping me occupied
and making sure I didn't starve throughout the year.
Finally I would like to thank Jen and my Mom who helped me through the more
challenging moments.

Contents

List of Figures

List of Tables

xiv

1 Introduction

1.1

Motivation and Objectives

1.2

Quantum Mechanics and its Interpretations

1.2.1

Schrodinger Picture: Wave Mechanics .

1.2.2

Bohmian Formulation

1.3

Historical Perspective . . . . .

1.4

Objections to Bohmian Mechanics .

2 Bohmian Mechanics

11

2.1

Hamilton-Jacobi Theory

11

2.2

Derivation from the Schrodinger Equation

13

Eulerian/Lagrangian Frames

14

2.2.1
2.3

The Quantum Potential

15

2.4

Bohmian Trajectories .

16

2.4.1

Definition . . .

16

2.4.2

Important Properties

16

vi

2.5

Compatibility with the Copenhagen Interpretation .

18

2.6

Practical Advantages of the Trajectory Formulation

19

2.7 Calculating Trajectories

2.8
3

21

2.7.1

From Wavefunction

21

2.7.2

Direct Trajectory Integration

22

Conservation of Probability

23

Calculations

3.1

Prelude: Units

25

25

3.2 Relevance of Time-Dependent Methods

26

3.3 Wavefunction Based Calculations

26

3.3.1

Time Propagation

27

3.3.2

Boundary conditions

27

3.3.3

Calculation of Trajectories

29

3.3.4

Web Interface .

31

3.4 Bohmian Calculations

33

3.4.1

Particle Approaches .

33

3.4.2

Algorithms- Newtonian Equations of Motion

34

3.4.3

Algorithms- QHJ Equations of Motion

35

3.4.4

Construction of the Wavefunction

36

3.4.5

Choice of Integrators . . . . .

36

3.5 Evaluation of the Quantum Potential

37

3.6 Interpolation Schemes

39

3.6.1

Quantum Potential by Polynomial Interpolation

39

3.6.2

Least Squares Approximation

41

3.6.3

Other Approaches

42

vii

Numerical Breakdown

45

4.1

Anatomy of a Trajectory Crossing .

46

4.2

Remedies

48

4.2.1

Artificial Friction Term .

48

4.2.2

Fourier Residues

48

4.2.3

Spline Interpolation .

52

4.2.4

Smoothed Splines .

54

4.2.5

Train of Gaussians

55

4.3

Interference

59

4.3.1

Dynamic Resampling

62

4.3.2

Dynamic Timescale .

65

5 Validation, Test Cases

5.1

70

Comparisons of Accuracy and Performance with Standard Techniques

70

5.1.1

Table Description . . . . . . .

71

5.1.2

Free Wavepacket Propagation

73

5.1.3

Harmonic Oscillator

.....

74

5.1.4

Interference of Two Gaussians

76

5.2

Numerical Instabilities and an Alternate Criterion for !lt

78

5.3

Conclusion on Interpolation Methods Tested . . . . . . .

81

5.4

Other Existing Methods - Covering function, W splitting, etc ..

82

6 2D and Higher Dimensions

6.1

83

Wave Based Calculations .

83

6.1.1

ADI Method . . . .

84

6.1.2

Boundary Conditions .

84

Vlll

6.2

Trajectories and Phase Unwrapping in 2D

87

6.3

Bohmian Methods in 2D . . . . . . . . . .

89

6.3.1

89

Instance of Generalizing from 1D to 2D: Least Squares Method

7 Cellular Automata and Lattice Boltzmann Methods

7.1

7.2

91

Cellular Automata . . . . . . .

91

7.1.1

Conway's Game of Life .

92

Cellular Automata and Fluid Mechanics

93

7.2.1

Lattice Gas Cellular Automata (LGCA)

93

7.3

Lattice Boltzmann Methods . . . . . . . . .

96

7.4

Cellular Automata and Quantum Mechanics

96

7.4.1

Dirac Equation . . . . . .

97

7.4.2

Majorana Representation .

97

7.4.3

Formal Connection to LBE .

98

Numerical Experiments .

98

7.5.1

Algorithm . . . .

98

7.5.2

Quantum Systems studied

99

7.6

Critique of the Majorama LBE Method .

103

7.7

A Look Back at the Bohmian Picture .

103

7.7.1

Lattice Boltzmann Revisited ..

104

7.7.2

Lattice Boltzmann Method and Bohmian Mechanics.

105

7.5

7.8

109

Comments

8 Conclusion

111

Bibliography

113

IX

List of Figures

2.1

Node avoidance property . . . . . . . . . . . . . . . . . .

17

2.2

Evolution of Bohmian Trajectories in an escape problem

21

2.3

Bohmian trajectories x(t) for a two-gaussian initial state.

23

3.1 Time evolution of a free wavepacket in 1D

28

3.2 Schematic of a double barrier potential ..

29

3.3 Time evolution of the wavepacket incident on a double barrier poten

tial, from the left.

30

3.4 Example of phase wrapping around

?T and its solution, in 1D

30

3.5 Bohmian trajectories corresponding to the wave function of Fig. 3.3

31

3.6 Typical run of the code via web interface

32

3. 7 Induced oscillations in a high degree polynomial fit

40

3.8 Least squares approximation to the same points as in Fig. 3.7

42

3.9 Example of free packet evolution using least squares interpolation

43

4.1 Trajectory crossings caused by numerical errors

46

4.2 Rotated view of \7 S

47

4.3 The smoothing property of Least Squares

47

4.4 Wavepacket evolution with a large friction coefficient

49

4.5 Resulting linear time-evolution of the system of Fig. 4.4

49

4.6 Example of interference . . . . . . . . . . . . . . . .

50

4. 7

The residue of a particular wavefunction and the result of low pass

filtering

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

4.8 Comparison of two interpolation methods with same arbitrarily dis

tributed points

. . . . . . . . . .

53

4.9 Interpolation methods compared .

54

4.10 Trajectories x(t) where edge instabilities develop .

55

4.11 p representation by a train of Gaussians

56

. . . . .

4.12 Trajectories x(t) for the harmonic oscillator in the classical limit (t vs x) 58

4.13 Detail of effect of "turning on" the quantum potential, over a quarter

period of the oscillation (t vs x) . . . . . . . . . . . . . . . . . . . . .

59

4.14 Avoided crossings between two Gaussians using sharply defined inter
60

polation kernels . . . . . . . .
4.15 Trajectory detail of Fig. 4.14

60

4.16 Wavefunction of two compact Gaussian packets interfering

61

4.17 Trajectories corresponding to Fig. 4.16 . . . . . . . . . . .

61

4.18 Trajectory detail of interference between two extended Gaussians

62

4.19 Evolution of the log-density (upper graphs) and quantum potential Q

(lower graphs) at selected times.

63

4.20 Evolution of the velocity profiles for two Gaussian packets with differ
ent separations (x vs v) . . . . . . . . . . . . . . . . . . . . . . . . . .

64

4.21 Trajectory compression and expansion in a barrier scattering problem

65

4.22 Dynamic timescale, abrupt declines represent time step adjustments

caused by error conditions . .

67

xi

4.23 Linear time scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

4.24 Perturbation in secondary fields not yet apparent in the log probability 68

4.25 High frequency oscillations in V 4 S for the system of Fig. 4.24

69

5.1

Sample run for the free wavepacket . . . . . . .

73

5.2

Sample run in an harmonic oscillator potential

75

5.3

Trajectories x(t) for smoothed spline calculation, 10,000 timesteps

77

5.4

Detail of Fig. 5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

6.1

Initial wavepacket (t=O)

85

6.2

Hard Wall boundary conditions in 2D, at t=800

86

6.3 Periodic boundary conditions in 2D, and the resulting wavefunction

86

6.4 Profile of the absorbing imaginary potential . . . . . . . . . . . . .

87

6.5 Pseudo-Transparent boundary conditions (with an imaginary potential

near the walls) and resulting wavefunction

88

6.6 A simple phase angle problem in 2d

88

6. 7 A practical case of the phase problem and its solution

89

6.8 2D Bohmian evolution of an inhomogeneous packet as viewed from the

z-ax1s . . . . . . . . . . . . . . . . . . . . .

90

7.1 Lattice and collision rules for HPP model .

94

7.2 Lattice and collision rules for FHP model .

95

7.3 Example of 3D lattice for a LGCA model .

95

7.4 Free packet dispersion (2 different momenta)

100

7.5 Mean position (left) and standard deviation (right) of the wavepacket

101

vs time . . . . . . . . . . . . .
xu

7.6 Wavepacket in an quadratic potential (left) and its average position

(right) . . . . . . . . . . . . .

101

7. 7 Wavepacket hitting a barrier .

102

7.8 Interference and exponential decay inside the potential barrier

103

xiii

List of Tables
3.1

Atomic units ([23]) . . . . . . . . . . . . . . . . . . . . . . . .

25

3.2

Coefficient values for method S4a at each intermediate step k .

37

5.1

Free Gaussian packet evolution

.......... .

74

5.2

Gaussian packet evolution in an harmonic potential

75

5.3

Two Gaussian packets interfering . . . . . . . . . .

76

XIV

Chapter 1
Introduction
"Things on a very small scale behave like nothing that you have any
direct experience about. They do not behave like waves, they do not
behave like particles, they do not behave like clouds, or billiard balls, or
weights on springs, or like anything that you have ever seen."
Richard Feynman

Almost since the inception of quantum mechanics, alternative interpretations have

existed. One of these, developed by de Broglie and later systematized by Bohm,
makes use of the familiar and intuitive picture of particles moving along trajectories
to describe the unfamiliar realm of quantum mechanics.
Paradoxically Bohmian mechanics (as it came to be known), was met with little
enthusiasm, even outright hostility in the early days of quantum mechanics and as a
result proceeded to lay dormant for a few decades.
Recently there has been renewed interest in this interpretation, in part for the
novel approach that it suggests to certain problems, such as decay processes, both
from a theoretical and computational stand point.
1

CHAPTER 1. INTRODUCTION

1.1

Motivation and Objectives

In this thesis I will further explore the Bohmian perspective as an alternative

framework for numerical calculations on a given quantum system.
The motivation for this study is the promise of better CPU scaling with the
dimensionality of the problem at hand. Common grid based methods often become
impractical in high dimensionality problems because of their exponential scaling.
It is said that there is no such thing as a free lunch on Wall Street and the same can

be said of Computing Street. The performance increase offered by Bohmian methods,

due to the sparse grids used, normally has a big trade off in the form of inconsistent
results wherever nodes or interference phenomena develop.
Since the long time evolution of a quantum system gives rise precisely to such
interference phenomena, Bohmian methods are usually riddled with stability issues.
I will investigate the causes of these stability issues, and how to minimize them. I
will study how different approaches and implementations of the particle method fare,
in terms of speed and long time stability, the latter usually enhanced at the cost of
additional computational complexity. The aim is to improve where possible on these
methods and recognize their strengths and weaknesses.

1.2

Quantum Mechanics and its Interpretations

That quantum theory "works" has long been settled by an impressive succession
of successful experimental predictions, and explanations, so its validity is no longer in
question. However the thorny issue of how the mathematical formulation of quantum
theory is to be interpreted is quite another matter. ..
As a clear sign of its inherent richness, the quantum theory accommodates very

1.2. QUANTUM MECHANICS AND ITS INTERPRETATIONS

many, sometimes disconcerting, interpretations. To paraphrase Richard Feynman in

one of his series of Auckland lectures on QED:
"If you think you understand quantum mechanics, you don't understand
quantum mechanics!"
We will not exhaustively describe all proposed formulations/interpretations of quan
tum mechanics, which at last count numbered at least nine, and rapidly increasing.
Instead we will refer you to the excellent introductory paper "Nine formulations of
quantum mechanics" [1] for details .
The list includes some of the more debatable and trendy ones such as "Many
Worlds" and "Transactional" interpretations of quantum mechanics. For curiosity's
sake though, I will enumerate from the aforementioned paper [1] the nine formulations
and two extra interpretations of quantum mechanics:
Matrix Formulation (Heisenberg)
Wavefunction Formulation (Schrodinger)
Path Integral Formulation (Feynman)
Phase Space Formulation (Wigner)
Density Matrix Formulation
Second Quantization Formulation
Hamilton-Jacobi Formulation
Variational Formulation
Many-Worlds interpretation (Everett)

CHAPTER 1. INTRODUCTION

4
Transactional interpretation (Cramer)

Last but not least: The Pilot Wave Formulation (de Broglie-Bohm)

To some, the Pilot Wave or Bohmian theory is one of those disconcerting formulations,
regressing as it were to classical concepts such as the moving point particle with a
definite momentum and position. One must not be prejudiced towards the theory
because of this, but use it as a tool, and like any tool use it where it works best.
The two formulations with which we will be concerned are the usual Schrodinger
representation or wavefunction formulation as it is the basis for common grid based
computational methods; and the Bohmian formulation as an alternative computa
tional route.

1.2.1

Schrodinger Picture: Wave Mechanics

Erwin Schrodinger had hoped that this formulation would cast quantum mechanics
in a "congenial" and "intuitive" form but was distressed when he found that his wave
functions were set in configuration space and not in ordinary three-dimensional space

([1]).
In the Schrodinger representation of quantum mechanics we deal with a complex
wavefunction w(x, t) =< xl'll(t) > where the ket is in Hilbert space representing the
state of the system.
The time evolution of the wavefunction is obtained by solving the Time-Dependent
Schrodinger Equation (TDSE):

."'&w

'tn-

&t

(1.1)

1.2. QUANTUM MECHANICS AND ITS INTERPRETATIONS

Observations of physical quantities (observables) like position, momentum, an

gular momentum, etc are associated with the "collapse" of the wavefunction onto
eigenstates of the corresponding observable, i.e. a quantum measurement acts as a
filter. Those systems that pass through the filter corresponding to some value of an
observable emerge in the associated eigenstate. The apparatus is described by a sum
of projectors onto the eigenstates of the operator associated with the particular phys
ical quantity. The probability amplitude for a particular result is the scalar product
of the initial wave function with the particular eigenket. In the Schrodinger repre
sentation, the scalar product is carried out by integration over all space coordinates,
and summation over spin indices.

1.2.2

Bohmian Formulation

The Bohmian formulation of quantum mechanics is also known as the Pilot-Wave

theory or causal interpretation of quantum mechanics.
In the practical implementation of this method use is made of an ensemble of
sampling points representative of the state of the system at time zero. Usually these
points are chosen to be more dense in regions where the particle has a high occupation
probability in the initial state.
In Bohmian mechanics the analogy to classical mechanics is so compelling that
we will find ourselves sometimes referring to these sampling points very loosely as
"particles" whose "trajectories" develop in time. It should be understood that we are
not dealing with particles in the usual sense of the word , but instead when we refer
to particles in a Bohmian context we are actually referring to the sampling points
carrying with them some probability volume, which taken as a whole may represent
the evolution of some "real" particle (such as an electron).

CHAPTER 1. INTRODUCTION

It is only in the initial state when we assign these sampling points or "particles"

that probabilistic arguments are invoked and the rest of the theory is completely
deterministic thus earning the denomination "causal interpretation".
Following the initial setup, these point particles evolve in time according to forces
derived from the sum of the classical and a newly introduced quantum potential.
The quantum potential in turn depends on the particular distribution of the point
particles at a given time.
It is this circular self consistency of the theory that is responsible for all the

complex quantum behaviour it is able to describe (and also for some computing
headaches we may add).

1.3

Historical Perspective

The historical foundations of Bohmian mechanics can be traced back to the hydro
dynamical formulation of quantum mechanics due to de Broglie[2] [3] and Madelung
[4] in the early days of quantum theory (i.e. late 1920's to early 1930's).
Some basic aspects of pilot-wave theory were already anticipated in de Broglie's
thesis of 1924. His talk at the 5'th Solvay conference in 1927 included an almost
complete exposition of the theory (see [5]).
Probably because of immediate criticism of the theory by Pauli and others (as early
as the 1927 Solvay conference [5]) the pilot-wave theory of quantum mechanics was
promptly cast into oblivion until David Bohm independently developed and presented
it in the early 50's (see [6]).
David Bohm (from whom Bohmian mechanics takes the name) was incidentally a
man of many and varied contributions and not only to physics. We pay tribute to his
memory by quoting Diirr, Goldstein and Zanghi [7] celebrating the life of the then

1.3. HISTORICAL PERSPECTIVE

recently deceased David Bohm:

"... We focus here on Bohm's contributions to physics. However, he also
made profound contributions to many other disciplines: to the philosophy
of science and the philosophy of mind, to ethics and moral philosophy.
And Bohm labored long for peace and disarmament, for dialog and mutual
understanding."
"Richard Feynman once declared Bohm the smartest man he had ever
met. He was certainly a man of extraordinary commitment to princi
ple, both scientific - as witnessed by his often lonely pursuit of scientific
truth, without regard for prevailing fashion - and moral - as witnessed
by his refusal in 1951 to testify against colleagues before the House of
Un-American Activities Committee. This act led to his indictment for
contempt of Congress and his banishment from Princeton and, indeed
from all of American academia. "
Bohm's seminal work was followed by another period of relative disinterest in the
theory until the 1970's, when John Bell became one of the most celebrated physicists
to back the Bohmian view. In fact the theory is even closely related to one of his most
celebrated discoveries "Bell's Inequality". According to Bell it was the non-locality
present in the Bohmian theory that inspired him to develop that result.
Included in the same paper by Diirr, Goldstein and Zanghi [7] is a private com
munication by Bell to the Philosopher Renee Webber where the ironic connection
between David Bohm's Theory and Bell's theorem (which is widely cited in support
of quantum orthodoxy) is exposed in the words of J. Bell:
"When I first realized that [Bohm's theory is nonlocal], I asked: "Is that
inevitable or could somebody smarter than Bohm have done it differently

CHAPTER 1. INTRODUCTION

and avoided this nonlocality?" That is the problem that [Bell's) Theorem
is addressed to. The theorem says: "No! Even if you are smarter than
Bohm, you will not get rid of nonlocality", that any sharp mathematical
formulation of what is going on will have that nonlocality... "

"In my opinion the picture which Bohm proposed then completely disposes
of all the arguments that you will find among the great founding fathers of
the subject- that in some way, quantum mechanics was a new departure of
human thought which necessitated the introduction of the observer, which
necessitated speculations about the role of consciousness, and so on."

"All those are simply refuted by Bohm's 1952 theory... So I think that it
is somewhat scandalous that this theory is so largely ignored in textbooks
and is simply ignored by most physicists. They don't know about it."

The historical record on the computational front, where Bohmian mechanics is used
in a more practical way as a basis for direct calculation of trajectories, is more recent.
Early works can be traced to the 1970's in the works of Weiner et al. [8). Then
after another long hiatus, the computational side of Bohmian mechanics was picked
up again in the nineties in works by Hua Wu and Sprung[9), Wyatt and Lopreore
[10], and by Sales Mayor, Askar and Rabitz [11) which set in motion another wave of
interest in the theory this time from a numerical perspective. A series of works then
followed over the following years up to the present day by many authors, including
those aforementioned and, amongst others, E. Bittner , B. Poirier , S. Garashchuk,
V. Rassolov, S. Goldstein and R.E. Wyatt.

1.4. OBJECTIONS TO BOHMIAN MECHANICS

1.4

Objections to Bohmian Mechanics

Bohmian mechanics has never been a short of critics. At the 1927 Solvay confer
ence Pauli was the first of a long list of eminent physicists to criticize de Broglie's first
proposal of a Pilot-Wave theory. In 1952, Pauli again leveled criticism on the now
more developed de Broglie-Bohm theory in his contribution to de Broglie's 60th birth
day volume deeming it "artificial metaphysics" [12, 13] (Pauli's objections revolved
around the break in the correspondence between classical and Bohmian mechanics
pertaining to the symmetric treatment of canonically conjugate variables such as po
sition and momentum. For other historical criticism a good source is Myrvold's paper
[14]). Concerning other less damaging accusations, Peter Holland's book ([15]) does
a good job of summarizing and then shortly addressing them:

Cannot prove trajectories are real - Cannot prove empirically the completeness
postulate either.
Predicts nothing new - Does permit more detailed predictions pertaining to
individual processes.
Regression to classical physics - It does depend on the "state" of the whole
system, represented by the guiding wave (a very non-classical concept)
Non-locality- Non-locality seems to be a small price to pay if the alternative is
to forego any account of objective processes.
More complicated than quantum mechanics - It is just a reformulation of quan
tum mechanics
Counter intuitive- Quantum phenomena require quantum intuition

10

CHAPTER 1. INTRODUCTION

No reciprocal action of the particle on the wave - in fact the wave depends on
the positions of the representative particle's trajectories.

Chapter 2
Bohmian Mechanics
For a more exhaustive and complete theoretical treatment of Bohmian mechanics
the definitive reference is Peter Holland's book " The Quantum Theory of motion"
[15]. In this chapter we will limit ourselves to aspects of the theory that are relevant
to calculations in this thesis.

2.1

Hamilton-Jacobi Theory

We take the classical Hamilton-Jacobi equation as our starting point since Hamil
ton Jacobi theory in the shape of the Quantum Hamilton-Jacobi equation (QHJE)
figures so prominently in Bohmian mechanics.

as
at

(VS)2- V
2m

(2.1)

In fact it can be argued (see [16]) that the seed of wave mechanics and Bohmian
mechanics is contained in the relationship between classical Hamilton-Jacobi theory
and Geometrical Optics.
11

12

CHAPTER 2. BOHMIAN MECHANICS

Goldstein argued along the following line:

If we consider that the surfaces of constant S (Hamilton's principal function)

correspond to wavefronts propagating in configuration space we can assign a wave

velocity to them u

= ~.

Taking the scalar wave equation of optics ,

(2.2)

which for slowly changing index of refraction n, has a solution of the form

A(r) exp(ik0 (L(r)- ct)) where A is the amplitude of the wavefunction and Lis called
the optical path length or eikonal of the wave. Substituting into eq. 2.2 results ,
in the short wavelength approximation, in the eikonal equation of geometrical optics

(\7 L) 2 = n 2 . The surfaces of constant L are surfaces of constant optical phase that
define wave fronts just as in the case of S from the classical Hamilton Jacobi equation.
Classical mechanics corresponds then to the geometrical optics limit of wave mo
tion. One might ask then what would happen to classical mechanics if we do not take
the short wavelength approximation. The answer would be the same as in the case of
wave motion: you would get the full wave equation. In the context of mechanics the
analog of the full wave equation is of course the celebrated the Schrodinger equation.
So could Hamilton or his peers have stumbled upon wave mechanics nearly 100
years before its actual discovery, with all the associated tantalizing scientific and
technological advances that such an early leap in human knowledge would carry?
Probably not... there was no reason to suspect that h in .\

= ~

was anything other

than zero, and classical mechanics was deemed to be rigorously true. Experimental
evidence suggesting otherwise would not be available until the beginning of the XXth
century in experiments such as those of Davisson and Germer, and others.

2.2. DERIVATION FROM THE SCHRODINGER EQUATION

2.2

13

Derivation from the Schrodinger Equation

The central element in the Bohmian picture is the concept of particle with an
associated (definite) position and momentum. One can obtain the Bohmian equations
of motion by writing the wavefunction in polar form: 'll(x, t)

R(x, t)e~s(x,t), where

Rand S are respectively the amplitude and phase.

Inserting the polar form of 'll(x, t) into the Schrodinger equation:

After some manipulations leads to two equations arising from the real and imagi
nary parts respectively ([15])

&R2

8t

as=
8t

+ 'V(R2\1S)
m

_(VS)
2m

= 0

+.!f._ v 2 R

(2.3)

2mR

The first equation in 2.3 can be interpreted as a continuity equation relative to

the probability flow R 2 ";;

The second equation is an analog of the Hamilton-Jacobi equation with an ex

tra term Q

;~ v~R, the quantum potential. This extra term depends on the

whole of the wavefunction, thus the name pilot wave theory, as Bohmian mechanics
is sometimes referred to (the wave guides the particle).
This is the usual way of interpreting that extra term as an non-classical potential.
Alternatively (and equivalently) we can consider the - 2n~ v~R as being another con
tribution to the kinetic energy (~~t, a form of a shape induced internal energy akin
to the role that the internal stress tensor plays in a classical fluid dynamics context.
Quickly glancing at the quantum term, already we can make a few preliminary

CHAPTER 2. BOHMIAN MECHANICS

14
observations. We note that as

n tends to zero the quantum potential disappears and

we should expect to get classical behaviour from the system. Because R appears both
in the denominator and in the numerator (as its gradient) Q will not depend on the
magnitude of R but on its curvature. When evaluating it numerically, however, we
can anticipate some difficulties in places where R is zero or very small.

2.2.1

Eulerian/Lagrangian Frames

To further clarify on the equations of motion we can, as is done in Hamilton Jacobi

theory of classical mechanics, associate a velocity field given by v = ; : , parallel to
the gradient of S which therefore will be perpendicular to a wavefront (a level surface)
of S.
At this point we need to discuss the concepts of Lagrangian and Eulerian frames of
reference. In classical fluid dynamics if we take a "moving with the fluid" perspective,
then the change of a physical quantity at each fluid "particle" is given by its material
derivative (also called substantive derivative), its intrinsic change plus its variation in
space: 1 = %t

+ 'Vv

. This is the so called Lagrangian frame where the grid evolves

in time with the moving fluid particles as opposed to the Eulerian view where the
grid is fixed in space.
Taking the gradient of the Quantum Hamilton Jacobi equation (QHJE) eq 2.3,
with p = R 2 and working in the Lagrangian frame of reference, we obtain a more
familiar set of equations([15]):

-ftp + p'Vv = 0
~~

-~'V(V

Q=

+ Q)

_!f._ V' 2R
2m

(2.4)

15

2.3. THE QUANTUM POTENTIAL

Here we can still recognize the first as the continuity equation, but now the second
equation reveals itself as just a statement of Newton's second law F

= ma

with the

now familiar extra quantum force term.

2.3

The Quantum Potential

That last term in eq.

mechanics.

2.4 is the quantum potential, the essence of Bohmian

It is Q which is solely responsible for all the quantum behaviour of

the system and it incorporates both properties of the particle such as momentum
and position, and properties of the whole system such as potential and observing
apparatus.

n,2 \72R
Q=---2m R

(2.5)

It is quite peculiar in that it can vary rapidly in places where V, the classical
potential, is almost constant. Even more strange is that since we have a normalizing
factor in the denominator, it does not depend on the magnitude of the density p itself,
but rather on its curvature (or more appropriately on the curvature of its square root
R), so we could have a very large quantum potential in an area where you have a
negligible probability density (as is the case for instance of interference between two
Gaussians as we shall see in later chapters). This can be seen by scaling R by an
2

arbitrary factor .X: Q =_..!E._

\7 >.R
2m >.R

= _..!E._ \7 R
2m

Numerically speaking, given that in general p can vary exponentially, and thus
has an enormous dynamic range, computationally it will be beneficial to represent it
instead by its logarithm C

log p. Not only are we compressing the range of p but

also for instance in the case of Gaussian wavepackets we are going from

ea 2 x +a 1 x+ao

16

CHAPTER 2. BOHMIAN MECHANICS

to a2x 2 + a1 x
So with p

+ ao

obviously a form more amenable to polynomial interpolation.

= ecand R = e~c, inserted into equation 2.5 the quantum potential

then takes the simplified form:

(2.6)
We can see how, as a consequence of the presence of the differential operators \7 and
\7 2 , the quantum potential encodes at each point information about its neighbours

and ultimately about the state of the whole system.

This is essentially the reason for the non-locality observed in quantum phenomena
such as entangled states where two potentially very distant parts of the system main
tain high correlation among themselves, as for example in Einstein's EPR paradox
(see [17]).

2.4
2.4.1

Bohmian Trajectories
Definition

The Bohmian trajectories are determined, just as in classical mechanics, from the
velocity field generated from S according to the definition v

= ;: .

At each point the particle moves perpendicular (parallel to the gradient) to the
wavefront of S.

2.4.2

Important Properties

In Bohmian mechanics trajectories may not cross or even touch each other. If they
were to cross, at the crossing point two "particles" with distinct momenta would share

2.4. BOHMIAN TRAJECTORIES

17

Figure 2.1: Node avoidance property

( From Hua Wu and D.W.L. Sprung [9])

the same space-time point which would imply that the underlying wavefunction should
have two distinct values at the same point in space. Since the wavefunction is single
valued, trajectory crossings are forbidden (see [15) and [9]). So if we see trajectories
crossing we immediately know that something is wrong in our calculations.
Furthermore at nodes of the pilot wavefunction W, the phase S is undefined and
there will be no particle paths going through those points , instead we may have
avoided crossings and or vortices (see [9]) develop around these points as illustrated
in Fig. 2.1. Fig. 2.1 represents the steady state of a 2D L-shaped quantum wire at
a particular energy (in this case JtJ ~ 0.25). Solid lines enclose regions containing
2

vortices and the inset shows behaviour near a stationary point of the velocity, where
the flow divides. At nodes S is no longer single valued and may undergo discrete
jumps: Sn = S + 21rnn where n is an integer. The contour integral of v around a node
will be

f v.dl = f;;: dl = f dS =

~1in and will define a vortex when n =I= 0 (see [9])

These sudden jumps in S will give rise to big gradients and can cause numerical
instability in algorithms, commonly referred to the node problem.
To illustrate the above, we use a compact Gaussian as the envelope of our ini

18

CHAPTER 2. BOHMIAN MECHANICS

tial wavepacket. It is positive definite everywhere thereby making the node problem
aforementioned less likely to occur. The reason for that resides with the continu
ity equation which determines evolution of p(x, t) = p(x 0 , 0) exp(-

J; \7v(x(t) , t)dt)

therefore if p(x 0 , 0) > 0 then p(x , t) > 0 for all t ime.

We can also make use of Ehrenfest's theorem which gives the evolution of the
expectation value of an Hermitian operator to calculate the average values for the
position and momentum of the system. For an arbitrary operator A,
d

dt < A>=<

8A

8t > +h < [H, A]

>

(2.7)

here the expectation value < A > stands for the average value of measurements of
the observable A in a state 17/J > . By taking A to be the position operator x or
momentum operator p, we obtain Ehrenfest 's theorem, the equations of motion for
their mean values. (eq. 2.8)

Ji
dt

<

>=

<p>
m

(2.8)

ft < p >=- < \7V(x ) >

We can then use eqs. 2.8 which agree with the classical equations of motion, to
monitor the quality of our numerical solution by monitoring the average position and
momentum along the trajectories.

2.5 Compatibility with the Copenhagen Interpreta

tion
The ultimate test for any physical theory is agreement with experiment so we duly
note that both standard quantum theory and Bohmian mechanics predict the same

2.6. PRACTICAL ADVANTAGES OF THE TRAJECTORY FORMULATION 19

experimental results, at least for single-particle problems.
One can always recover the wavefunction having solved the Bohmian trajectories
by integrating along them the density and phase given by respectively the continuity
equation and the Quantum Hamilton Jacobi equation.
But how does a totally deterministic theory with point particles and trajectories
accommodate key quantum concepts such as uncertainty in results of measurement?
The reason we "recover" these uncertainties and thus converge to the same results
as conventional quantum mechanics, is that the initial conditions are not perfectly
known (by design?) . We are given a statistical ensemble for the initial conditions
corresponding to the wavefunction at that time, say p(r0 , t 0 ) =

lw(r0 , t 0 )l 2

(where t 0

is the initial time and r 0 is some surface on which the boundary conditions are ap
plied). From the solution of the TDSE, everything else will follow and the trajectories
will retain their statistical weight through time, yielding measurements that are also
probabilistic in nature.

2.6 Practical Advantages of the Trajectory Formu

lation
Not counting the supposed practical computing advantages, sometimes it is even
advantageous to recalculate Bohmian trajectories from pre-existing time-dependent
wavefunction solutions, since the former will give us a clearer understanding of some
underlying physical concept(such as arrival time).
In some problems , when we are trying to answer kinds of questions that are bet
ter posed in a classical domain, the trajectory formulation makes it possible to give
intuitive yet quantum mechanically correct answers. Such is the case for instance

CHAPTER 2. BOHMIAN MECHANICS

20

when calculating lifetimes of particles when we want to know the probability a par
ticular wavefunction will be contained by some potential. In the trajectory view one
simply counts the number of particles/ trajectories that have crossed a predetermined
boundary.

An example of this is illustrated in figure 2.2, the deterministic individual trajec

tories are perfectly suited to explore questions of time of flight , lifetime etc. which
are present in time-dependent quantum problems such as a decay. We can see that
in the second case this probability is actually an oscillatory function of time [18], as
we have trajectories that for a while tunnel in and out of the boundary .

Note that when using Bohmian trajectories we can determine with no ambiguity
which part of the wavepacket passes and which part remains within the barrier. For
instance we can also see in Fig. 2.2 that most of the transmittance is due to the front
of the wavepacket and not the rear. This is something unthinkable in conventional
quantum theory where the wavefunction is taken as a whole, and such information
would be unavailable.

Besides these practical advantages of Bohmian mechanics, there is an important

conceptual continuity between classical and quantum physics afforded by Bohmian
theory[12] (at the cost of non-locality).

The emergence of classical phenomena out of a quantum world continues to be a

major unsolved problem in Physics . The fact that Bohmian mechanics preserves the
concepts of trajectories and particles from classical physics is a hint that Bohmian
mechanics may have a role to play and be a more appropriate tool to approach this
problem.

2. 7. CALCULATING TRAJECTORIES

21

r(t)

~----~----~----~--~

10

11

12

13

14

Figure 2.2: Evolution of Bohmian Trajectories in an escape problem

(From Nogami et al.[18])

2. 7
2.7.1

Calculating Trajectories
From Wavefunction

Calculating trajectories from the wavefunction w(x, t) is really a two step process:
1. Solve the time dependent Schrodinger equation (TDSE).

2. Integrate the velocity field determined by the phase of the newly calculated W.
We can solve the TDSE analytically for a relatively small number of simple systems
( see [15, 19]).
For the vast majority of cases that cannot be solved analytically, with the advent
of cheap and powerful computing, a suitable time propagation method can be used
to numerically solve the TDSE in low dimensional systems.
There are many methods that can be used to solve the TDSE. We can for instance
project the initial wavefunction onto stationary states of the system's Hamiltonian,
and then evolve these according to their eigen-energies (spectral methods).

CHAPTER 2. BOHMIAN MECHANICS

22

We can discretize the TDSE on a fixed space-time grid. This can be done using
FFT methods, Feynman Path integrals, Monte Carlo techniques, or operator splitting
methods, etc. We use the latter in a Crank-Nicholson scheme as represented in eq.
2.10 in conjunction with transparent boundary conditions as described in Moyer's
paper ([20]) and in appendix A to setup a "golden" standard to which we can compare
the various particle method implementations.

w(r, t + bt)
-zH8t

e-n-

exp( -iHM/n) x 'll(r, t)

1-iHk
21l
1 + iH8t
21l

(2.9)
(2.10)

In fact, we should note that this simple Crank-Nicholson method for 1D problems,
has recently been generalized to higher orders described recently by W. van Dijk and
F. M. Toyama (see [21]). They have obtained a dramatic improvement in attainable
precision, for a given amount of computation.

2.7.2

Direct Trajectory Integration

In this method we directly use the Bohmian framework to make numerical cal
culations. The trajectories come out naturally as a constituent of the calculating
procedure.
The direct Bohmian method is a more efficient way of calculating trajectories and
because we use grids that are much sparser than those used in wave based methods,
it usually is orders of magnitude faster. However it is vulnerable to numerical insta
bilities when interference effects are dominant. We will give more details on this in
the next chapter.

23

2.8. CONSERVATION OF PROBABILITY

Figure 2.3: Bohmian trajectories x(t) for a two-gaussian initial state.

2.8

Conservation of Probability

After we discretize the initial wavepacket into a representative statistical ensem

ble of particles, each of them can be thought of as carrying an associated probability
volume. This probability volume will evolve in time according to (d~:)
which using the continuity equation can be put in the form pl. \lv

p~~

+ J~,

+ J( -p\lv)

= 0

which is always zero. So it will remain constant in time, that is to say that each
particle will have the volume "assigned to it" expand and compress as the probability
density decreases or increases respectively, but it always represents the same prob
ability volume. This is illustrated on Fig. 2.3, the inital state has two Gaussians
placed symmetrically about the origin. As the wave packets spread they interfere
but the trajectories do not cross. The shaded volume represents the same probability
volume 1/N at all times and initially experiences expansion as the probability density
decreases while the Gaussian packet is expanding, but then experiences compression
when the Gaussian packet "meets" trajectories from the neighbouring packet, giving
rise to an increased probability density.
This probability volume can be used to effectively label each particle (as is done

24

CHAPTER 2. BOHMIAN MECHANICS

in [22]) since each one can easily (at least in lD) be assigned a cumulative probability
that will remain constant throughout time: P(xi, t)

= f~oo

p(x', t)dx'

In fact, in the Newtonian code of Sec. 3.4.2, which is how we will select our particle
ensemble from w(x, 0): we will place a particle whenever the cumulative probability
has increased by 1/N where N is the chosen number of particles. Thus between two
trajectories there will always be the same amount of probability
exception of the two particles at the extreme boundaries.

-ft

with the possible

Chapter 3
Calculations

3.1

Prelude: Units

Quantum mechanical phenomena are mainly observed in the atomic domain, al

though increasingly they are relevant in meso and nanoscopic systems. When studying
such systems, physicists and chemists find it convenient to work in the atomic sys
tem of units where

n= m

= 1. The resulting Schrodinger equation for an electron

simplifies to: i~~ = -~\7 2 \lf

+ Vw

On Table 3.1 for convenience, we include the conversion factors between the atomic
system of units and the SI system.

I Physical quantity I Atomic Units I

Action
length
mass
Energy

n=1
a0 = 1
me= 1
Eh = 1

Description
i; Planck's constant
Bohr Radius
electron's mass
Hartree energy

Table 3.1: Atomic units ([23])

25

SI units
1.054571 * 10 34 J s
5.2917 * w- 11 m
9.109382 * w- 31 kg
4.35974 * 10 lS J

II

CHAPTER 3. CALCULATIONS

26

3.2

Relevance of Time-Dependent Methods

Why use a time dependent method? Since inception of quantum theory, time
independent methods have always been preferred as being more tractable and efficient,
and in most cases synthesizing the relevant physical processes taking place in a system.
Obviously if we are dealing with a time dependent Hamiltonian we have no choice
but to use a time dependent method. However even in the case of a time independent
Hamiltonian it can sometimes be advantageous to use a time dependent method.
To illustrate this let us consider spectral methods. The standard way these work
is by projecting the initial wavefunction onto the eigenvector basis (stationary states)
of the system, and then evolve in time a linear combination of these projections.

(3.1)
with ak =< klw(O) >and Hk = Ekk
This is often the preferred route since initially we deal with space only and not
space and time. However in some cases (i.e. scattering) the eigenstates become con
tinuum functions making the discrete sum in equation 3.1 an integral over continuum
states and if the calculation is required at very large times, it may be difficult to
perform the Fourier integrals with sufficient accuracy. In such cases a time dependent
method may be advantageous, an early example of which is in Heller's work ([24]).

3.3

Wavefunction Based Calculations

As we saw at the end of the previous chapter, Bohmian trajectories can be calcu
lated a posteriori, after the full time evolution of the wavefunction 'll(x, t) is known.
Since, apart from a handful of cases, analytical solutions are generally not available we

3.3. WAVEFUNCTION BASED CALCULATIONS

27

must turn to calculating the time evolution of the system via some standard method
that is known to work. One such method, augmented by transparent boundary con
ditions, is described in the next sections.

3.3.1

Time Propagation

If we wish to compare the efficiency and accuracy of a Bohmian method, we need

some standard to which it can be compared. Towards that objective we implemented
a fixed grid split operator method that will serve as a "gold standard" in later sections.
The main reason we choose a particular method (see [20]) over other standard
methods is that for this one, transparent boundary conditions have been devised
and it is quite efficient. As for the method specifically, it uses the Crank-Nicholson
scheme to propagate the system in time by symmetrizing the propagator :
1 + iH <5t- ~ H 2 <5t 2 into the Cayley form:

e-iHlit =

~~:Z~~j~

+ 0( <5t 3 )

e-tH 8t ~

which is second

order accurate, and more importantly unitary. The Numerov algorithm is used to
extend the accuracy in the spatial domain to fifth order.

3.3.2

Boundary conditions

Transparent Boundary Conditions (TBC) are very convenient in that they allow
us to concentrate on a small volume with great detail where the "interesting" physics
may be happening, without worrying about unwanted wave reflections from artificial
boundaries. Complex potentials are sometimes used as an absorbing layer on the
boundaries to avoid having to enlarge the computing domain, but they never work
perfectly.
In our case TBCs are also essential for easy comparisons between wave and particle
methods, as TBCs are inherent to the Bohmian method (that is one of its advantages).

28

CHAPTER 3. CALCULATIONS

200
400
600
~

800

1000

::..
o>

o>

~ 1200
1400
1600
1800
2000

50

100

150

200
250
300
350
X From -40 to 40 (nm)

400

450

500

Figure 3.1: Time evolution of a free wavepacket in 1D

(no reflection is apparent at the domain boundaries, TBCs are working)

TBCs for different kinds of differential equations are discussed by Matthias Ehrhardt
m [25]. In particular discrete TBCs in one dimension for the Schrodinger equation
were derived in [26] and an adaptation to the Numerov method is presented by Moyer

([20]).
We can observe the time evolution of a free Gaussian with some initial momentum
to the right , in Fig. 3.1; the packet spreads in time but no reflection occurs when it
hits the domain walls because TBCs are enabled.
We should note that implementation of TBCs comes at a cost of the minor incon
venience that we need to record the wavefunction at the boundaries at each time step,
and store this history. This has an increasing performance cost as time progresses.
The reason these values are needed , is that at the system boundaries, there are fluxes
of probability in both directions, and it is not valid to suppose purely outgoing flux,
even when the net flux is to the right.

3.3. WAVEFUNCTION BASED CALCULATIONS

29

Figure 3.2: Schematic of a double barrier potential

A more detailed description of the TBC used is presented in appendix A.

3.3.3

Calculation of Trajectories

Let us take a more interesting example of a Gaussian packet hitting a double

barrier, as illustrated by the potential in Fig. 3.2. We can observe the resulting
time evolution in picture 3.3. Some reflection to the left from the first barrier, some
transmission to the right past the second barrier and indeed some metastable states
that become temporarily trapped in between the barriers (and eventually leak out)
are seen.
We were faced with an interesting problem when trying to calculate trajectories
from the wavefunction results. Remembering from Sec. 2.4 that the trajectories are
determined by the particle velocities v = V S/ m, we then need a way to extract the
phase information from the wavefunction, which is represented by a set of complex
numbers. Unfortunately the standard way of simply taking the arctan(Imag/Real)
will only give results in the

1r domain, so we need to monitor the function for points

where sudden jumps of 27r are required for continuity. These are indicate a crossing
of the ordinate axis as illustrated by Fig. 3.4 . If we were to take just the gradient
of the phase provided by the computer code, we would be presented with unphysical
velocity jumps at those points, resulting in erroneous trajectories.

30

CHAPTER 3. CALCULATIONS

Figure 3.3: Time evolution of the wavepacket incident on a double barrier potential,
from the left.
( otice the transmitted and metastable states to the right and in the middle
respectively)

60,----,---~-~-~-~----,

so
40

30

20
-1

_, .

10

-3

_, OL_------,1-',0
0--2,-',00------,'-300,------~
40-0---,50~0_

_JSOO

(a) computed phase showing jumps of 1r

-lo o;------;,~oo--,:;;;oo;-------;;:
3o:-o-~.o;:-o---;so=o-----:!.
soo

(b) same, but with discontinuities resolved.

Figure 3.4: Example of phase wrapping around

1r and its solution, in lD

3.3. WAVEFUNCTION BASED CALCULATIONS

31

50
40

_./,
/

30

/"

,'

"

/,..
.

/.
/~

20

./

10

'E

.s

-10
-20
-30
-40
-50

100

200

300

400
500
600
T1mestep ( 1 Is)

700

800

900

1000

Figure 3.5: Bohmian trajectories corresponding to the wave function of Fig. 3.3
( viewing angle is rotated by 90 degrees, time is now x axis)
Once we have the correctly calculated velocities we can proceed to integrate the
position x

J vdt

from v

~~ to get the updated positions of the particles and

thereby define the trajectories.

An example of trajectories calculated from the wavefunction can be seen in Fig.
3.5. The metastable orbits inside the double barrier can more easily be seen than
in Fig. 3.3 (as can their eventual tunneling to escape to the right or left side of the
barrier).

3.3.4

Web Interface

We should add that, for convenience, at the time we developed a website that took
as inputs the various parameters of a calculation such as dimensions, time steps, po
tential , initial wavepacket, etc and then fed these to a background running program.
After a calculation related delay, a webpage with both the wavefunction solution and

32

CHAPTER 3. CALCULATIONS

Location

~dlt

~ew

,G.o

S,ook:marl<s

Iools

.S.ettlngs ,Window

Q.QQO 0 , ...,,K '"""~

;:J

t:felp

-If..

G L9_cation [f-"'httpiil27o~O~ld/web;e~dy~~P..(tm\_-

____ ~-

__

d _-,!j\\GL

- --tfme Dependent-SimulatioriofQuantum systems

10 code
Paulo Machado 2005

I
' GMSMASSES (relative to effective mass In GaAs).ll
SIZE (In# cells):

TIMESTEPS:

1512_____

riQOQ- --- - - ---

DELTA: (fsec)
r-"'""

XL: (left boundary nm) :-40

XR: (light boundary nm) 1+40--

POTENTIAL: (function of X 1n eV)

iirx;-io)&CX<-s)+rx>Sll

PSIO ( probability distrlb 1n terms of X) :exp(-1/2/sqrt(lO)A2'{X

i BOHM: (# Bohm1an trajectones each representing 1/# probability) '300

i:

Submit job j

Feel free to experiment with alternate values for

variables/potentials/etc

Figure 3.6: Typical run of the code via web interface

the corresponding calculated trajectories is output by the program.

A screenshot of a typical screen is shown in figure 3.6, the resulting page will
present similar contents to figs 3.3 and 3.5.
We should also mention that recently, an improvement to the Crank-Nicholson
scheme was introduced by W. van Dijk and F. M. Toyama (see [21]) that can be ap
plied to further increase the accuracy of these wavefunction based calculations. Their
method uses high order formulae in both space and time domains, typically of order
h 20 , where his the step length. The extra labour per step is greatly overcompensated
by using large steps in both space and time, reducing the overall computing time by
several orders of magnitude. However, they have not yet incorporated TBC into their

3.4. BOHMIAN CALCULATIONS

33

methods.

3.4

Bohmian Calculations

First a word about boundary conditions. In trajectory based calculations the

system is generally unbounded, as we simply keep track of the numerical value of
the current position of each particle. In fact we can easily incorporate any kind of
boundary conditions in the potential calculation.
This is in a marked contrast to grid based methods as illustrated in appendix A
where we note that to enforce transparent boundary conditions (zero outside poten
tial) an elaborate procedure is necessary that becomes more and more computation
ally intensive as time progresses (we need to keep a history of boundary values of the
wavefunction).

3.4.1

Particle Approaches

As we noted in Sec. 2.4, the cumulative probability can be used as a convenient

label for each particle, i.e. each particle carries with it a certain volume of probability
volume through time that can be used to calculate the probability density. We can
then choose one of two courses:
We can work with probability densities which we will have to update at each step
via ~ = p0 e-"il.v and then evaluate p' and p" to compute the quantum potential.
Or we can label each particle with its cumulative probability P(x) at each spatial
point which will then be constant in time at the cost of when p is required we use ~~
(from P(x) =

J:.oo p(y)dy) , thus increasing by one order the derivatives required for

the Q calculation. We will implement both methods of keeping track of the probability

CHAPTER 3. CALCULATIONS

34

density and compare them in terms of stability.

3.4.2

Algorithms- Newtonian Equations of Motion

Let us recall the Bohmian Equations of motion of Sec. 2.2.1:

ftp+ p\lv = 0
dv

dt

= _..!.\J(V
m

Q=

+ Q)

(3.2)

_.!E._ V'zR

2m

Using this set of equations the procedure to follow is this:

Start with a statistical ensemble representative of the initial state of the system
{Xi, Vi,

Pi} (we shall use a linked list for storage).

If using P, from {xi,~} calculate p.

Calculate Q from {xi,~} or {xi, Pi} depending on the labeling method used
(more on this later).
Update vi via the gradient of V
Update xt from vt (x

+Q

(~~ = -~\l(V

+ Q)).

~~).

If we are not using P we need to update Pi according to

1f =

p 0 e-'V.v

At this point we have advanced by <5t in time and the whole process is repeated.
If we want to avoid dealing with P (and its 3rd order derivatives for Q) we can use
p instead, provided we update it at each time step using the continuity equation as

illustrated in the following section.

3.4. BOHMIAN CALCULATIONS

3.4.3

35

Algorithms- QHJ Equations of Motion

Also we can minimize the numerically troublesome task of taking gradients by

using the Quantum Hamilton-Jacobi version of the equations of motion from 2.2, this
time taking care to update the velocity of each particle as v

~ + \7 (p '::)

as=
at

_C'i1S)2 _
2m

= ';;:

(V + Q)

(3.3)

=_.ft._ V'2R
2m R

Using this set of equations the procedure to follow is this:

Start with a statistical ensemble representative of the initial state of the system

Calculate Q from

{Xi,

Pi}

Update S via the Quantum Hamilton-Jacobi Equation ~~

= -

<~~

(V

+ Q)

( i.e. in an explicit central differences implementation, time evolution will take

the form: S(t + 8t) = S(t- 8t) - 28t [ (V'~~))
Calculate vi from v =

Update

Xi

from

+ (V + Q) J )

Y'S

vi

Update Pi according to !f{f

p0 e-V'.v or equivalently !f{f = p0 e-(V'

At this point we have advanced by

repeated.

ot in time and

S)jm

the whole process can be

CHAPTER 3. CALCULATIONS

36

3.4.4

Construction of the Wavefunction

At any time we can recover the standard wavefunction view of the system ([8]) by
integrating both phase and density along a trajectory .
The evolved density is determined by integrating the equation of continuity:

p(x, t)

p(x, O)e- f~C'v.v(x,t))dt

Evolution of the phase is obtained by integrating the quantum Hamilton-Jacobi

Equation:

S(x, t) = S(x, 0) +

J(dt
dS

)dt = S(x, 0)

+ J(

(\7S)
m - (V
2

+ Q))dt

Or combining both terms and referring to the original wavefunction we get:

w(x, t)

3.4.5

w(x, O)ef~('v.v)dte* J~((~~t -(V+Q))dt

(3.4)

Choice of Integrators

Given the equations of motion the question arises, what kind of integrator to use
to evolve these in time. Since the time integration will be at the core of our procedure,
special care in terms of accuracy and efficiency is required in picking the numerical
integrator to use.
One other consideration is that since, as we can see from Sec.

3.4.2, we are

integrating an Hamiltonian system, we would like certain constants of motion such as

energy for a classical system and normalization, in a quantum system, to be conserved.
The symplectic class of integrators, widely used in the fields of celestial mechanics
and molecular dynamics, are designed with those properties in mind, so its remains

3.5. EVALUATION OF THE QUANTUM POTENTIAL

I Coefficient I I
ak
bk

k=l

k=2

k=3

2+2!+T!

1-2-3-2--3

1-2-3-2--3

~
2-23

~
1-23

37

k=4
2+2!+T-3
6

~
2-23

Table 3.2: Coefficient values for method S4a at each intermediate step k
(reproduced from (28)[29])
to choose one from that class.
To help us do that we refer to the paper by Gray (1994) (28] where he compares
a range of symplectic integrators as well as more standard ones such as fourth order
Runge-Kutta, in terms of their speed and conservation properties in a molecular
dynamics context.
We selected the fourth order method S4a (which was originally introduced in (29])
because from all the methods examined it provided the best balance between speed
and accuracy. The coefficients used, ak and bk, are taken from (28] and listed in table
3.2. They are used to integrate the equations of motion as directed by eq. 3.5.

&H
.
p=-aq
{

&H

(3.5)

q = &p

where F = - ~~ and G = ~~ = ~ = v and you take four steps k = 1..4 to update

the values of p and q.

3.5

Evaluation of the Quantum Potential

We have seen in Sec. 2.4 that each particle carries with it a predetermined amount
of probability. We usually pick ~ for each trajectory when using the Newtonian form
of the Bohmian equations. In the QHJE case we often use an initially uniform grid

CHAPTER 3. CALCULATIONS

38

in which each probability volume will be set but different for each trajectory. In
either case the cumulative probability that it carries (at least in lD) unequivocally
identifies the particle and can be used to label it. When required we can calculate the
probability density asp= ~~ which follows from the definition of P(x) = f~oo p(y)dy
Depending then on which variables we choose to work with (they are going to
vary as with different methods of interpolation), Q the quantum potential and FQ
the quantum force will be given by:

or, with R =

Vf5,
(3.6)

(3.7)
In terms of P,

(3.8)

(3.9)
Or instead, using p and determining its logarithm we obtain with C = log R =
~ logp:

(3.10)

3.6. INTERPOLATION SCHEMES

39

(3.11)
This logarithmic form has a number of advantages:
It more naturally represents an exponentially decaying probability.
It avoids the problem of dividing by a possibly very small number .

Last but not least, C lends itself more naturally to a polynomial interpolation
than p (i.e. for a Gaussian C is a quadratic polynomial).
The non-locality of Q is patent in the V' and \7 2 terms and presents a significant
numerical challenge. In the next few sections we will present some ways of extracting
their values from our discrete set of trajectories.

3.6

Interpolation Schemes

3.6.1

Quantum Potential by Polynomial Interpolation

We are given a set of discrete points and are asked to calculate gradients, the
simplest and more intuitive choice is to approximate these points by continuous poly
nomial functions which can easily be differentiated.
This interpolation scheme yields the most local approximations to Q and in fact
for the case of 1 or 2 neighbouring points it will reduce to the common two and three
point formulas for the first derivative, respectively.
Given a particle and its N-1 nearest neighbours at distances

~Xn

with values

we want to determine the coefficients of the interpolation polynomial y

C3X

+ ...

where

Cn

~! ~ lx=xo

c1 + c2 x

Yn

40

CHAPTER 3. CALCULATIONS

Figure 3. 7: Induced oscillations in a high degree polynomial fit

( the underlying is a simple parabola with only one offset point near x=5)

They are determined by solution of the linear system:

L}.xl

(L}.x 1)

L}.x2

(L}.x2)

2
2

(L}.x 1)
(L}.x 2)

3
3

(.6.x3) 2 (.6.x 3)3

3
2
1 Llx4 (.6.x 4) (.6.x 4)
1

Y1

(3.12)

L}.x3

Y4

As the book Numerical Recipes [31] reminds us, this is a Vandermonde Matrix
which can be quite ill-conditioned, so we can expect the coefficients thus obtained to
be undesirably sensitive to small variations in the data points.
Instability is especially obvious when a large number of points is involved in the
polynomial interpolation. This is illustrated in Fig. 3. 7 where we start with a perfect
parabolic distribution of points and slightly shift only one point (near 5). We can
observe how a high degree polynomial fit tends to oscillate wildly around the interpo
lated points as it tries to pass exactly through each and every one of them (Runge's
Phenomenon).
Lagrange interpolation thus has some desirable characteristics such as simplicity

3.6. INTERPOLATION SCHEMES

41

and the fact that it hits each point exactly but as one makes it less local (by including
more neighbouring points in the interpolation) it becomes less and less stable.
One should note as well that given the form of eq. 3.11 and in particular its

\1 3 C term that the polynomial used should be of at least fourth degree in order to
contribute to that term in the quantum force. This observation holds for other types
of interpolation that depend on polynomials such as the one discussed in the next
section.

3.6.2

Least Squares Approximation

We noted before that Q is non-local, so one can attempt to capture more of this
non-locality by including a larger number of nearest neighbours in the polynomial
interpolation. Unfortunately as we saw in Sec. 3.6.1 and illustrated by Fig. 3.7 the
polynomial interpolation does not do well with many points, becoming hypersensitive
to noise.
If we increase the number of points in the interpolation without increasing the

degree of the polynomial we get an overdetermined system of equations that we can

solve optimally using the least squares method.
We wish to fit a polynomial function of the type:

Again we get a linear system similar to eq. 3.12 but now for M points, we have
a MxN matrix with M > N which we can solve using for instance the SVD method
(Singular Value Decomposition, see [56) ). This is now a semi-local evaluation of the
derivatives since we are taking information from more and more distant neighbouring
particles when evaluating gradients.

CHAPTER 3. CALCULATIONS

42

-35,'---'----'--"'--___.,.._....___.__

__..__---:------',0

Figure 3.8: Least squares approximation to the same points as in Fig. 3. 7

The difference is dramatic in Fig. 3.8, specially when we compare it with Fig.
3. 7 of the polynomial interpolation. The interpolating points are the same for both
cases, yet we can see how the least squares approximation reproduces almost exactly
the original generating quadratic form.
Overall the least squares method results in a smoother approximation to the un
derlying distribution and therefore should lead to more physically valid results when
evaluating its gradients (especially if we have some previous knowledge of the un
derlying distribution). We note however that because of this weighted approach the
interpolating function will not, in general, pass through each point. In fact in Fig.
3.8 we can see that it basically ignores the sixth point from the left. This averaging
property can lead to trajectory crossings down the line (more on these in the next
chapter).

3.6.3

Other Approaches

The interpolation methods described in this chapter tend to be unsatisfactory in

all but the simplest situations. In the next chapter we will explore alternative ways
of evaluating Q with better stability properties.

3.6. INTERPOLATION SCHEMES

43

i -05~:t

-1~--~----~--~----~--~----~--~--~

-2000

-1500

-1000

-500

1000

500

1500

2000

space

a_:-:f--,-:: _,
1

200

400

600

800

i:r : , , :
0

200

400

600

800

1000
t1me

1200

1400

:;?J

1000
step

1200

1400

1600

1800

2000

Figure 3.9: Example of free packet evolution using least squares interpolation

Top: Phase diagram

Middle: Space-time diagram

Bottom: Dynamical timescale (here it is always increasing)

44

CHAPTER 3. CALCULATIONS
As a testimony to the computational versatility of the Bohmian interpretation,

there are a plethora of numerical implementations of the basic Bohmian equations.

Some of the methods we didn't mention here include the derivative propagation
method, covering function , Wigner function approach, etc. (see [8] for an overview).

Chapter 4

Numerical Breakdown

The particle method works quite well and is known to be stable for Gaussian
packets in free space, linear and quadratic potentials, where the time evolution of
the wavefunction is shape preserving (these are incidentally, cases where analytical
solutions also exist).
However, if we introduce a minute cubic or quartic term in the potential or a
barrier, the system will, given enough time steps, become unstable, eventually leading
to nonsensical conditions such as trajectory crossings, the telltale sign that something
has gone wrong.
In this chapter we will explore the causes and conditions that affect the stability
of the particle method. We will be very interested in what role numerical instabili
ties caused by interference play in this, and finally we will introduce some alternate
ways of evaluating the quantum potential that may have better stability and fidelity
properties.
45

46

CHAPTER 4. NUMERICAL BREAKDOWN

Figure 4.1: Trajectory crossings caused by numerical errors

4.1

Anatomy of a Trajectory Crossing

Referring back to Sec. 2.4.2, we recall that one of the important properties men
tioned was that particle trajectories should not cross; this is a consequence of the
single-valuedness of the underlying wavefunction. If we do observe crossings, we can
be assured that something has gone wrong with the calculations that led to those
crossings (see Fig. 4.1).
When two trajectories approach each other in a sudden manner, it becomes diffi
cult to reliably estimate p from ~~. Furthermore because of the denominator term in

Q, when

p tends to 0 , we can expect to see big swings in the value of

Q ( and high

gradients) which in turn lead to high particle velocities which may then catapult a
particle towards another. In fact because p

= ~~

when they cross, dx tends to zero

and thus p becomes infinite, or in numerical terms "diverges", that can be seen by the
sudden scattering of particles in Fig. 4.1.
In the real world there has to be a force that prevents the trajectories from getting
too close. This is the quantum force term (Sec. 3.5), and a trajectory crossing is an
indication of too big a time step or a badly calculated quantum force (i.e. insufficient
magnitude or pointing in the wrong direction).
To illustrate how this could happen, let us take the example of the least squares

4.1. ANATOMY OF A TRAJECTORY CROSSING

47

Figure 4.2: Rotated view of \7 S

(points near 0.2 are about to cause a trajectory crossing)

Figure 4.3: The smoothing property of Least Squares

(it largely ignores the particle shift near 5)

method of Sec. 3.6.2.

We start with an uniform sampling of points taken from a simple quadratic form.
Now we artificially move particle 6 to the left (see Fig. 4.3). We can see that the
characteristic smoothing responsible for the stabilization of the derivatives here works
against us and smoothes out the crucial area between the now almost overlapping
points 5 and 6. This smoothing prevents the particles from "seeing" each other when
calculating Q with the least squares interpolation and thus we don't have a suitable
quantum force to "repel" the particles and thereby prevent a crossover.
There are then two issues that should be addressed in order to prevent particles

CHAPTER 4. NUMERICAL BREAKDOWN

48

from getting excessively close to each other: The proper interpolation of Q and a
time step small enough to enable the particles to respond to Q and decelerate. The
time step issue will be addressed in the next Sec. 4.3.2.

4.2
4.2.1

Remedies
Artificial Friction Term

By introducing a term of the form Vj

= -

fc ~~ in the Hamiltonian ([44][45]) where

fc is the friction coefficient, we artificially introduce dissipation into the system. We

will obviously destroy any energy conservation properties so it is certainly not a way
to get the proper time evolution of the system. However a slight amount of artificial
friction can be beneficial to the stability of the simulation and may be introduced so
as to dampen spurious numerical oscillations that typically occur in the evaluation of
gradients. If the friction is low enough, we can do this without completely changing
the character of the system.
If one takes this to an extreme, and increases the friction too much, the system

will evolve into its ground state. The ground state in a Bohmian context is peculiar in
that all movement stops as the particles remain static at the bottom of the potential.
In Fig. 4.4 we can see the system relaxing towards the ground state of a potential
well.

4.2.2

Fourier Residues

This is an alternate interpolation scheme that combines a polynomial least squares

interpolation with a Fourier expansion to approximate the log density.
The rough rationale for this combination is that the polynomial form arises nat

4.2. REMEDIES

49

space

<:E---------~

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

3000

3500

4000

4500

5000

tune

1500

2000

2500
slep

Figure 4.4: Wavepacket evolution with a large friction coefficient

From top to bottom: Phase space, time-evolution and dynamic time scale

Figure 4.5: Resulting linear time-evolution of the system of Fig. 4.4

(note the convergence to the ground state, where the Bohmian particles remain

motionless)

CHAPTER 4. NUMERICAL BREAKDOWN

50

"

"'~

"

''

,-~~~-

Figure 4.6: Example of interference

Top: log density profile

Bottom: Corresponding -Q

(Sampling particles spaced by 1% probability )

urally from the use of Gaussian wavepackets (where the log density is a quadratic
form) and the fact that when we have potential reflection we should expect interfer
ence terms to start developing in the density function (see Fig. 4.6).
These interference terms are characteristically oscillatory in nature and thus fa
vorably inclined to be represented by a Fourier expansion.
Mathematically the approximation to the density with an oscillatory residue is:

The polynomial (of order m) part of C, the log density, can be determined for

51

4.2. REMEDIES

x 1o

6
4

0
-2
-4

50

100

150

x10

Figure 4.7: The residue of a particular wavefunction and the result oflow pass filtering
instance in a least squares way.

Once that is determined, the residue Res(x)

C(x)- Polym(x) is calculated and can be expanded in a (discrete) sine series (DST)
as in equation 4.1.

y(k) -

~:;'= 1 x(n) sin(n Nk: 1 ), k =

1, ... , N

(4.1)

Recalling equation 3.11, in order to calculate the quantum force we will require
C',C"and C 111 :

The calculation of derivatives for the polynomial part of C is straightforward. To

determine the derivatives for the residue part we should differentiate in Fourier space.

52

CHAPTER 4. NUMERICAL BREAKDOWN

Since we are now working in Fourier space, at this point if desired, we can lowpass filter the residue in momentum space (Fig. 4. 7) by adjusting the coefficients
of Res(k), and then make use of the differentiating property of Fourier transforms

F('/!J

ikF(J) to calculate the needed derivatives of the residue as shown on the

right side of eqs. 4.2

C' -

Poly'+ F- 1 (ik.Res(k))

C"
C 111

(4.2)

Poly"'+ F- 1 ( -ik 3 .Res(k))

Here F- 1 is the inverse discrete sine transform (IDST) that takes us back to position
space and is the reverse of eq. 4.1:

x(n)

L.k= 1y(k)

sm(1r N

kn

+ ), n
1

1, ... , K

(4.3)

This is the ultimate non-local approximation to Q since we take information from

every other point to evaluate the gradients C(n) at each location. Practically, although
the estimates for

C(o)

correlate very well with C , the estimated gradients tend to

oscillate for points near the edge, giving rise to problems in future time steps.

4.2.3

Spline Interpolation

We have seen in Sec. 3.6.2 that the least squares method, although less local,
smoother and more reasonable looking than the polynomial interpolation, will not
prevent trajectory crossings because it doesn't pass through every point.

4.2. REMEDIES

(a) Polynomial interpolation

53

(b) Spline interpolation

Figure 4.8: Comparison of two interpolation methods with same arbitrarily dis
tributed points
(The spline method is distinctly more physical)

Ideally we would like to have an interpolation of C or p, that although smooth and

taking information from as many points as possible, still manages to pass through
each point exactly so that trajectory crossings may be avoided.
This should be in the spirit of the polynomial interpolation but without the sta
bility issues that riddle the former, and somewhat in the spirit of the least squares
approximation which takes information from as many points as possible, but being
certain to pass through each fitted point, exactly. A method that satisfies both re
quirements is the spline approximation, where you approximate piecewise with poly
nomials.
In Fig. 4.8 illustrate the difference between the polynomial and the spline methods,
and in particular the fact that the splines still pass through every point, while the
overall function looks much more smooth( and physical) than the polynomial.
In Fig. 4.9 we see the effect of shifting that sixth point towards the left in the
least squares and splines method, the latter passes through each point and doesn't

54

CHAPTER 4. NUMERICAL BREAKDOWN

(a) Least Squares interpolation (b) Least Squares interpolation

(c) Spline interpolation

Figure 4.9: Interpolation methods compared

seem to suffer from big oscillations as does Lagrange interpolation.
To calculate the quantum force, we will use C instead of p, as its predominantly
polynomial shape is better suited to being interpolated by splines. Because we need
terms of order \7 3 C in eq. 3.11, if we are using P , the cumulative probability, the
degree of splines we should use is four to make the quantum force term continuous .
Therefore we want to use at a minimum quartic splines instead of the more commonly
used cubic splines (see [31]) to interpolate C (a good guide on the method of splines
is Carl De Boor's book[38] ).

4.2.4

Smoothed Splines

Unfortunately the spline interpolation is liable to develop oscillations in Q at the

edge of the interpolating domain. These propagate in time and eventually compromise
the whole simulation. An example of this phenomenon is presented in Fig. 4.10.
As we saw in the previous section the spline method has so many other desirable
properties, that we should try to fix this edge effect. One way of doing that is to use a
smoothing spline with a site dependent parameter p that will make the interpolation
smoother at the edges, and less sensitive to numerical oscillations.

55

4.2. REMEDIES

Figure 4.10: 'Ifajectories x(t) where edge instabilities develop

The smoothing spline SS will minimize:

w(j)(C- SS(x)) 2

+ (1- p)

)..(x)('\7 2 SS) 2 dx

j=l

The parameter p will control how smooth the spline will be by emphasizing the
point distance (p- 1) or the low gradients (p

---t

0). Cis the usual log-density known

at sites x 1 and w and ).. are site dependent weights. We can, by adjusting this local
parameter p, generate a smoother interpolation of C near the edges while remaining
close to the original splines in the middle points.

4.2.5

Train of Gaussians

This alternate interpolation scheme takes its inspiration from the field of digital
signal processing, in particular in the Shannon sampling of a discrete signal, and
specifically of an irregularly sampled discrete signal as our Bohmian points really are.
Because the density is at all times positive definite we should enforce this by
substituting the sine function from Shannon sampling which allows for negative values

CHAPTER 4. NUMERICAL BREAKDOWN

56

Figure 4.11: p representation by a train of Gaussians

(the solid line is the resulting p and arrows represent sampling points )

with a Gaussian shape that is convolved at each sample point represented by a delta
function. This procedure is illustrated graphically in Fig. 4.11.
Normally each point represents the same probability volume, thus the Gaussians
will have the same volume, resulting in sharper distributions when points are closer
together and more diffuse ones when points are far apart. Mathematically what we are
doing is convolving a series of delta functions (the particle locations) with Gaussians
of varying standard deviation but the same area (thus the different heights in Fig.
4.11).

p(x) -

1 N
N'L.n=
1 6(x- Xn)

p(x)

~L.;[=lGn(X- Xn, ern)

* Gn ( x, ern)

where Gn(x, ern) represents a Gaussian centered at 0 with a standard deviation of ern.

4.2. REMEDIES

57

(4.4)

Recalling the form of the quantum force when using p as in eq. 3. 7 we are going
to need terms of the form

v;p, ("7) and 'il~~ P. Since we can analytically differentiate

2

the Gaussians, it is just a (laborious) matter of calculating Gn,G~,G~ and G~' and
combining these according to eq. 3.7 to get an analytic form of Q for a given particle
distribution:

G' (x, a)
G" (x, a)

G111 (x, a)

Having the different derivatives of G we can now express the needed terms of p (x)
as:

1 N
N~n=lGn (x- Xn,an)

p (x)
Vp(x)
p (x)

a2 N p (x)

~~= 1 ( (x- Xn) - a 2 ) Gn (x- Xn, an)

a4 N p (x)
2
~~= 1 ( (x- Xn) (3a 2 - (x- Xn) )) Gn (x- Xn, an)
a6 Np(x)
2

V 2 p (x)
p (x)

\13p (x)
p (x)

~~=l- (x- Xn) Gn (x- Xn, an)

Finally we can plug these values in the quantum potential or the quantum force

CHAPTER 4. NUMERICAL BREAKDOWN

58

Figure 4.12: Trajectories x(t) for the harmonic oscillator in the classical limit (t vs

x)
as required:

Classical limit from vanishing Q The classical limit is approached whenever the
system's action is large compared to nor alternatively Q

---+

0.

By turning off the quantum potential we can appreciate the differences between
the classical and quantum worlds.
In Figure 4.12, which represents an harmonic oscillator, we can see that with Q = 0
we have trajectory crossings and every particle follows a sinusoidal path independent
of the other particles.
Fig.

4.13 represents the same potential.

Here we can see that the quantum

trajectories are no longer independent, and they do not cross.

Finally in Fig. 4.14 we have a hint of how classical behaviour may emerge from

4.3. INTERFERENCE

59

16

02

1B

(a) Classical

04

06

08

12

14

ta

(b) Quantum case, note the trajectories do not

colapse

Figure 4.13: Detail of effect of "turning on" the quantum potential, over a quarter
period of the oscillation (t vs x)
Bohmian trajectories. In this case we use the train of Gaussians interpolation method,
but with somewhat sharper density distributions than usual.
Looking at Fig. 4.15 it is apparent that we don't actually have trajectory crossings
but if we add a property of indistinguishability to the neighbouring particles those
trajectories would appear to cross when we zoom out back to Fig. 4.14 . This is
speculative but it would seem that Bohmian mechanics would, in this way, enable a
natural emergence of classical behaviour.

4.3

Interference
"The double slit experiment has in it the heart of quantum mechanics. In
reality it contains the only mystery.."
Richard Feynman

How does interference arise in our bohmian context? To study that, we are going
to use two Gaussians freely evolving in space and eventually interfering together. To

CHAPTER 4. NUMERICAL BREAKDOWN

60

lmeartme

Figure 4.14: Avoided crossings between two Gaussians using sharply defined interpo
lation kernels

-2

-6

-8

1DO

200

300

400

800

Figure 4.15: Trajectory detail of Fig. 4.14

61

4.3. INTERFERENCE

50

100

150

200

250

Figure 4.16: Wavefunction of two compact Gaussian packets interfering

1~

1=

Timestep (.1 fs}

1 1 1

Figure 4.17: Trajectories corresponding to Fig. 4.16

(In this figure the space and time axes have been interchanged as compared to Fig.

4.16)

obtain the proper trajectories, we first calculate the wavefunction using the techniques
of Sec. 3.3. In Fig. 4.16 we can see regions of compression (in red) and expansion (in
blue) typical of interference patterns emerging as time passes.
In the frames of Fig. 4.19 we can readily see one reason why it may be difficult to
capture this behaviour using the traditional particle treatment: insufficient sampling.

62

CHAPTER 4. NUMERICAL BREAKDOWN

E'

.s
><

20

40

60

80

100

120

140

160

T1mestep ( 1 Is)

Figure 4.18: Trajectory detail of interference between two extended Gaussians

The interference emerges from a zone of negligible density where there are very few
particles encoding for the log-density field. In a density weighted particle distribution
for this case, the majority of the particles will initially be far away from the zone where
interference develops and so will not be able to account for it, some resampling of that
area will be in order if we want to capture that information earlier in the simulation.

4.3.1

Dynamic Resampling

It is that little blip near x

0 in the quantum potential of Fig. 4.19 and the

associated discontinuity in the velocity fields of Fig. 4.20 that are going to self feed
in time and grow into the interference patterns of the density distribution.
There is no mystery as to how a trajectory on the far right "knows" that there is
a trajectory on the far left, that information is carried by the quantum potential and
it has its origin in that small bump in Q between the two Gaussians.

63

4.3. INTERFERENCE

lt:S?:1l\CY1
.
- -. - . . . . . - -. - . .

. .

i~:rrIIioc
- - -

(~) t~3oo

- - (d)

t~2odb

"

Figure 4.19: Evolution of the log-density (upper graphs) and quantum potential Q
(lower graphs) at selected times.
On each panel, representing a particular timestep, the log density and corresponding
Q are drawn versus space

64

CHAPTER 4. NUMERICAL BREAKDOWN

.....

Figure 4.20: Evolution of the velocity profiles for two Gaussian packets with different
separations (x vs v)
(Left panels: the Gaussians are initially centered at 7, and right panels, 15 units
of distance)

4.3. INTERFERENCE

65

(a) linear

Figure 4.21: Thajectory compression and expansion in a barrier scattering problem

How can we hope to capture this behaviour? Besides the obvious importance of
having a physically acceptable interpolation method, one possibility is that we must
change the premise that we need only sample regions of space where the probability
density is significant, as we have already seen that the quantum potential does not
depend on absolute densities.
Instead we should resample regions where the inter-particle distance is contracting
and where that distance is above a certain threshold (to avoid oversampling) in order
to properly capture the highly variable gradients in Q and S or v that will develop in
the region. The reason for the distance criterion is related to the Shannon sampling
theorem in that, if we have interparticle spacing of
momentum information of order :S

ix.

we can only hope to obtain

Since we will have high gradients in the

interference area, we should decrease the average

4.3.2

~x,

~x

in this region.

Dynamic Timescale

In order to estimate the error in our calculation at the end of each time step, we
compare the results (as measured by the particle positions or more accurately by the

CHAPTER 4. NUMERICAL BREAKDOWN

66

gradients of a field like S or C) with the errors obtained by taking two half steps. This
way we can monitor the error and maintain it below some predetermined threshold
by making dynamical adjustments to the time step. At the end of the process we get
to keep the more accurate two half step result so the extra calculations involved are
not completely wasted.
In our particular calculations we are cutting the time step in half in case of too
large an error and increasing it by 1% if the error is acceptable, this enables the time
step to "recover" from very low values in regions where such is possible, as illustrated
in Fig. 4.22.
In the case of a Gaussian packet in a parabolic potential, we get a decreasing time
step as the packet approaches the bottom of the potential where it has maximum
momentum and highest density of trajectories (and therefore where we need a smaller
time step).
Following this point of maximum trajectory density, circa step 750 in Fig. 4.22,
the time step gradually recovers to larger values accelerating the simulation.
Furthermore if a more serious error condition is detected such as a trajectory
crossing (as opposed to the error simply exceeding some threshold), the time step
is again cut in half, and in addition we restore the system to the state of some
predetermined number of time steps in the past, to enable it to avoid an evolution to
the serious error condition.
In Fig. 4.24 we can see the reason why we have to go back a couple of time steps
in order to correct a trajectory crossing. It is that usually, by the time it becomes
apparent that the positions of the particles are erroneous, the background fields (such
as Q and S ) have been compromised for quite a while. Indeed some interpolation
methods (i.e. splines) are prone to develop nasty oscillatory perturbations by the

4.3. INTERFERENCE

67

-5
-10
-15
-20

-25 o'-----=-'2oo.,---4:7:oo,----=GOO-'-::----:'so':-o-1c:':ooo.,---::':-:---:-:'::--:':::---c:-:':-:-~
bme

llC::: .. d
0

200

400

600

BOO

1000

1200

1400

1600

1800

2000

step

Figure 4.22: Dynamic timescale, abrupt declines represent time step adjustments
caused by error conditions
Top:Space-time graph
Bottom:Time step versus time

l1neartJme

Figure 4.23: Linear time scale

(Using the recorded time steps from Fig. 4.22 we can convert the space-time graph

back to a linear time scale where the Harmonic oscillator trajectories are more easily

recognized)

CHAPTER 4. NUMERICAL BREAKDOWN

68

10

Figure 4.24: Perturbation in secondary fields not yet apparent in the log probability
Clockwise from top left: log probability C , Q , \7 2 8 and \7 8

edges that will self-feed after many time steps and give rise to trajectory crossings.
Because of this lag in cause and effect, we must use different monitoring functions for
different methods.
Areas that are troublesome when evaluating derivatives with splines, are the com
pression areas, where the trajectories are coming closer together and thus the density
is increasing leading to higher numerical error in the differentiation. These areas can
be identified by \7 2 8 < 0 or equivalently \7v < 0, that area is readily visible in the
right bottom panel of Fig. 4.24.
A useful monitoring function for problematic zones that does a good job of catching
unstable conditions early on, is the time evolution of the quantum potential,

69

4.3. INTERFERENCE
025
0~

015
01
005

-005
-01
-015
-02
-025

_,;

_,;

-2

Figure 4.25: High frequency oscillations in \1 4 S for the system of Fig. 4.24
which we can for error monitoring purposes approximate by its more problematic first
term \1 4 S that is represented in Fig. 4.25. By comparing all fields represented in that
figure and the previous one, we can see an illustration of the lag in the propagation of
instabilities as these are quite evident in \14 S as high frequency spacial oscillations and
are already somewhat developed in Q of Fig. 4.24 but they are not yet recognizable
in C at that particular time step.

Quality of Fit

An added bonus of using a dynamical timescale is that we are given

a measure of the quality of the simulation at each point in time. The dynamical time
step tends to be very small in regions where the system is encountering errors and
when the time evolution is relatively error free it will increase as in Fig. 4.22.
Furthermore, we can make use of the fact that the Ehrenfest theorem should
hold, as this is a quantum system, so we can take the average of observables such
as position and momentum and compare their values to those expected in a purely
classical system, according to eq. 2.8. At each time step we compute the average
position < x > of the trajectories we are following, and compare them to those of the
classical trajectories arising from the mean initial values.

Chapter 5
Validation, Test Cases
To check the validity of the trajectories and to test the performance of the various
interpolation methods we start by choosing a couple of (the very few) systems for
which an analytical solution is known. We then apply the same tests to a propagating
problem with interference similar to that of Sec. 4.3, that is known to have long time
stability issues in Bohmian calculations.

5.1 Comparisons of Accuracy and Performance with

Standard Techniques
The following tests were done on an Intel Dual Core Duo 3GHz machine with
4Gb RAM under MATLAB in Linux, we used single threaded code for our tests.
The tests consisted of measuring the running time and flaggin g error conditions (if
any) for wave packet propagation in certain standard quantum potentials using our
Bohmian code.
It is our experience that using the quantum Hamilton-Jacobi equations (QHJE) of

70

5.1. COMPARISONS OF ACCURACY AND PERFORMANCE WITH STANDARD TECHNIQ

motion of Sec. 3.4.3 as opposed to the Newtonian code of 3.4.2, results in both faster
execution and more stable numerics. The reason for that is probably linked with the
higher order of numerical differentiations that the usage of P and the algorithms of
Sec. 3.4.2 carry and the fact that \7 Q is not calculated directly but instead Q is
combined with the kinetic energy and the classical potential in ~~.
Therefore in the following calculations, intended to compare interpolation rou
tines, we will be using the QHJE code, modified in the following way:
So that comparisons are easier to make we use a fixed time step (dt

5 x 10- 3 )

for all trials. This fixed time step, which we get by turning off the dynamic time
step machinery, will cause trajectory crossings to be more frequent than would nor
mally happen, but the trajectory crossings will still depend on the quality of the
interpolation methods used.
When the run is successful we compare the end wavepacket with the one obtained
using the split operator method (see sec. 3.3). If on the other hand error conditions
do occur a note is taken of the time step when that happened, giving us a rough
measure of which interpolation used is more susceptible to numerical instabilities.

5.1.1

Table Description

A brief explanation of the parameters seen in the following tables 5.1, 5.2 and 5.3
follow.
The column entitled CPU(s) is self explanatory and it is simply the time in seconds
that it takes to run the code. Note that unlike the Moyer code, no optimization was
attempted on the Bohmian code and visualization and debugging code is run inside
the main calculating loop. For our purposes of comparing interpolation methods this
is not an issue though since they all run the same unoptimized code.

CHAPTER 5. VALIDATION, TEST CASES

72

Tstop

is the time step at which execution is interrupted, either because we have

reached our predetermined maximum number of time steps or because an error con
clition such as trajectory crossing is encountered, that information is included in the
next column "Traj".
In the next two columns estimates of the error of the final wavefunction are given
as compared to the analytic solution of the same problem or to the Moyer solution
(it was not necessary to use a numerical solution for the first two tables as analytic
solutions exist for those potentials but we always compute the Moyer solution in all
cases).
The error is the difference between our solution and the wavebased solution in the
spatial domain common to both. In< (P-Pm) >we have the error in the density while
in < (C - Cm) >the error in the log density ( x ~) is represented, these are presented
as an average value the standard deviation

CJP

J(P-;t_"i_)

or

CJc

J(C;~rr_;_)
2

Finally the last column gives us an estimate of the relative speed of all the methods
in time steps per second of CPU time.
The Moyer grid method with a parameter 2048 x 2000 means that it was run on
a 2048 point space grid for 2000 time steps .
P is the number of trajectories or particles used and in the polynomial method D
stands for the polynomial degree used. At each point we find the D nearest neighbours
and fit a polynomial to them to get estimates of the derivatives (for instance if D=2
we get the usual central difference formula for the curvature).
Part of the least square method is an extra parameter N, the number of nearest
neighbours considered at each point. In the case N>D+l, the higher theN the more
averaging occurs in the least squares estimate.
In the sines method, the LPFilter parameter represents the percentage of low pass

5.1. COMPARISONS OF ACCURACY AND PERFORMANCE WITH STANDARD TECHNIQ

J.. ;.. :. . .. ..: <

-10

-8

-6

-4

-2

J~ . . :~-10

-8

-6

-4

-2

..... : . :-J

10

:::l

10

Figure 5.1: Sample run for the free wavepacket

Top to bottom panels are: C, Q, '\! S, '\! 4 S

filtering done in momentum space as discussed in Sec. 4.2.2.

In the smoothed splines method B represents the number of points considered at
left or right "edges", and smoothed according to the procedure of Sec. 4.2.4.

5.1.2

Free Wavepacket Propagation

This is the simplest case where there is no external potential and only one initial
Gaussian packet exists. No interference effects are expected.
The free Gaussian wavepacket evolves ([15]) in time according to :

'lf(x, t)

(21fSt)-t exp(- (~:::r) with

St

ao(1

+ i 2n!aJ

All interpolation methods deal quite well with this very simplest of cases of a single
decaying Gaussian. The one exception is when we increase the number of points in
the polynomial interpolation to 32; there the simulation doesn't quite make it to 500
timesteps and there is a trajectory crossing at 469. This is probably due to the high
order of polynomial used causing oscillations as pictured in the Runge phenomenon
in Fig. 3.7.

CHAPTER 5. VALIDATION, TEST CASES

74
Description/Parameters
Moyer grid 2048x2000
Poly P=50 D=2
Poly P=50 D=5
Poly P=50 D=6
Poly P=50 D=8
Poly P=50 D=16
Poly P=50 D=32
LeastSq P=50 D=2 N=6
LeastSq P=50 D=6 N=8
Smes P=50 LPFllter=10%
Gaussians P=50
Splines P=50
I
I SmoothSplines P=50 B=12

I
I

I CPU(s) I Tstop I Traj I <p-Pm > 10

I 49 I 2000 I NI A I
0
83
137
141

142
145
161
85
139
16.4
23.3
16.8
21

500
500
500
500
500
469
500
500
500
500
500
500

Stable
Stable
Stable
Stable
Stable
Cross
Stable
Stable

I Stable I
I Stable I
I Stable I
I Stable I

-1.54 2.3
-1.54 2.3
-1.54 2.3
-1.54 2.3
-1.54 2.3
N/A
-1.54 2.3
-1.54 2.3
-1.54 2.3
-1.54 2.3
-1.54 2.3
-1.54 2.3

I steps/s I
I 40.8 I
6.02
3.65
3.55
3.52
3.45
3.11
5.88
3.60
30.5
21.5
30.5
23.8

Table 5.1: Free Gaussian packet evolution

One can also observe where the different methods stand in terms of performance
relative to each other.

We see that in general the polynomial and least squares

methods are much slower than the rest. In the polynomial or least squares methods,
the routine has to consider each point, determine a local neighbourhood of size N, and
then construct an approximation in that neighbourhood. In other methods (such as
splines) we first get a global approximation to the log density and then we get updated
values to every trajectory at the same time. In other words we do operations to the
whole vector of positions instead of each point separately, with performance gains.
This is particularly true as the number of trajectories increases.

5.1.3

Harmonic Oscillator

Another analytically solvable system, the harmonic oscillator's time evolution al

ternates between periods of compression and expansion as illustrated in Fig. 5.2. Here
the potential used is a weak quadratic centered around x = 0, V(x) = 5 x

w- 2 (x)2.

5.1. COMPARISONS OF ACCURACY AND PERFORMANCE WITH STANDARD TECHNIQ

12

Figure 5.2: Sample run in an harmonic oscillator potential

Fig. 5.2 looks somewhat different from Fig. 4.23 because we are using an uniform
distribution of initial points as opposed to distributed as

Description/Parameters I CPU(s)
Moyer grid 2048x2000
I 49
Poly P=50 D=2
173
170
Poly P=50 D=3
Poly P=50 D=6
295
Poly P=50 D=10
295
164
LeastSq P=50 D=2 N=6

277
LeastSq P=50 D=6 N=8

Smes P=50 LPFilter=10%

31
Gaussians P=50
45
Splines P=50
33
40.8
SmoothSplines P=50 B=12
361
SmoothSplines P=50 B=12

I Tstap
I 2ooo
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
10000

Traj
NI A
Stable
Stable
Stable
Stable
Stable
Stable
I Stable
I Stable
I Stable
Stable
Stable

1 in Fig. 4.23.

I < p-Pm > 10-5 I steps/s I

I

I
I
I

0
-9.73 539
-9.73 539
-9.73 539
-9.73 539
-9.73 539
290 5300
5400 15200
-13.7 41
-1990 4610
-9.73 539
-182 820

Table 5.2: Gaussian packet evolution in an harmonic potential

I 40.8 I
5.78
5.88
3.39
3.39

6.10

3.61
32.2
22.2
30.5
24.5
27.7

CHAPTER 5. VALIDATION, TEST CASES

76

5.1.4

Interference of Two Gaussians

Here we apply the same tests to a system that will develop interference patterns
similar to those that we discussed in Sec. 4.3. We have two Gaussians that are placed
side by side and let them interfere together with no external potential applied. The
evolution of the system is represented in Fig. 5.3 which was taken from a run of the
smoothed splines method (last item on table 5.3).
Description/Parameters
Moyer grid=2048x2000
Poly P=50 D=2
Poly P=50 D=3
Poly P=50 D=4
Poly P=50 D=5
Poly P=50 D=6
Poly P=50 D=10
Least Sq P=50 D=2 N=4
Least Sq P=50 D=2 N=5
Least Sq P=50 D=2 N=6
/ Sines P=50 LPFilter=10%
Gaussians P=50
Splines P=50
SmoothedSplines B=12
SmoothedSplines B=12

I CPU(s) I Tstop
I

49
132
135
212
211

NjA
NjA

70
108
110
17.5

NjA
NjA
75
392

2ooo
862
840
814
737
736
652
416
640
686
472

332
629
2000
10000

Traj
N/A
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Cross
Stable
Stable

I < p-Pm > 10 3 I stepsjs \

I
0
I 40.8 I
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
-0.15 1.5
-La 2.8

6.5
6.2
3.8
3.5
N/A
N/A
5.9
5.9
6.2
27
N/A
N/A
26.7
25.5

Table 5.3: Two Gaussian packets interfering

We can see from table 5.3 that most methods lead to trajectory crossings. We
remind the reader that we have turned off the dynamic time scale control mechanism.
The objective here is not to avoid trajectory crossings, but to evaluate different in
terpolation methods, and so by looking at table 5.3, and specifically at which time
step trajectory crossings were observed, we can say that a certain method is more
stable than another. Of course, the proper stable choice here should be the smoothed

5.1. COMPARISONS OF ACCURACY AND PERFORMANCE WITH STANDARD TECHNIQ

Figure 5.3: Trajectories x(t) for smoothed spline calculation, 10,000 timesteps

Figure 5.4: Detail of Fig. 5.3

splines method as it is the only one that doesn't result in crossed trajectories.

A surprising fact is that D

2 poly, which is really nothing more than fitting

parabolas at each point, (resulting in the familiar central differences formula) doesn't
do too badly compared to other methods and outperforms the least square method
and even the basic spline method (crosses at 862 vs 629).

CHAPTER 5. VALIDATION, TEST CASES

78

5.2

Numerical Instabilities and an Alternate Crite

rion for

~t

As we saw in Sec. 4.3.2 and Figs. 4.24,4.25, by the time we notice that there is
something wrong with the C or the S fields, usually the root cause of that problem
already occured a few time steps previous, and a typical signature is high frequency
oscillations in '\1 4 8.
Oscillations suggest that we are integrating some trajectories too far in the direc
tion of other particles, resulting in an excessive repulsive force and oscillations down
the line. Why this tends to happen in problem areas of compression is linked to the
fact that the absolute error of a derivative goes approximately as

fx

times the error

of the underlying field, so as particles are packed closer together and t5x becomes
smaller, so grows the error in the gradients estimated in that area. In fact for '\1 2 the
error will be of order 4/ (8x )2; therefore it is imperative, when using non-averaging
(such as least square) interpolations, that we reduce the time intervals in lockstep
with a minimum inter-particle distance to minimize integration errors .
Let's take a look at the QHJE equations of motion:

VS

1
2
--(VS)
+-21("\! 2 C + (VC) 2 ) - V
2

_!v 2 s
2

where X is the position, S the phase and C the log-density. For a moment let's ignore
any errors in the X field and concentrate in the fields S s and C c with absolute

5.2. NUMERICAL INSTABILITIES AND AN ALTERNATE CRITERION FOR l::iT79

errors s and c respectively.
In the case of simple numerical differentiation, every time we are evaluate \7 S we
2

get an error of order 2sj8x, for \7 2 8 the error is 4s/(8x)2 and for (V'S) we get

(4s/(8x) 2 ) \i'S. Similarly for C.

Let's see what happens in just one Euler integration over dt, of the equations of
motion:

If now we consider S 0

= S0 s 0 and C0 = 0 0 c0 where the bar denotes the

noiseless fields, using the expressions for the errors in the gradients we have the error
estimates:

s1

so + dt x

~ ( ( 8:\7 So) .so 8~ 2 co \7 Co! co)

c1

co dt x

~ ( 8~ 2 so)

For small 8x, the 8x 2 term will dominate in the first equation and we can further
simplify:

sl

,....,_
,....,_

2dt
so (8x) 2co

cl

,....,_
,....,_

2dt
co (8x) 2so

80

CHAPTER 5. VALIDATION, TEST CASES

If we extend to further time steps, these equations define a recursive relation

between two variables s and c for time step n:

OrWl"th

an=

2dtn-1

Sn

Sn-1

(~

Cn

Cn-1

(~

VXn-1

)2 Cn-1

2dtn-1

UXn-1

)2 Sn-1

2dtn-1
~,
oXn-1

To get a general expression for the errors we sum both equations,

So in general

Now an

2oXn
..d t2n , so we want an as small as possible which implies an alternative

criterion for t:lt :

5.3. CONCLUSION ON INTERPOLATION METHODS TESTED

81

Where t5xmin is the minimum inter-particle distance. This will mean that some
times the whole simulation will have to slow down by virtue of a small dt in order
to integrate accurately zones of compression. If we are using a local interpolation
method, this could probably be relaxed to integrating only the neighbourhood of
small t5x with a reduced timescale, in a sort of adiabatic approximation where the
rest of the field would be "frozen", but we have not yet explored that possibility.

5.3

Conclusion on Interpolation Methods Tested

There is a delicate balance between the demands of having a stable estimation

and having an accurate estimation of the gradients in our Bohmian calculations. As
we saw in chapter 4 the accurate estimation is crucial, to make sure that particles
avoid each other. However with this added sensitivity to local conditions also comes
sensitivity to numerical noise and the possibility of inducing numerical oscillations in
the solution. On the other hand, the low noise and stable estimation of derivatives
afforded by an averaging method come at the cost of not being accurate enough when
genuine local fluctuations do exist.
A compromise between these is present in the smoothing splines method where
we can specify (sec. 4.2.4) zones of higher or less sharpness in the interpolation. (A
more computationally costly but valid possibility is to have a least squares method
with a locally adjustable number of fitting neighbours)
In any case, from the data in the previous tables, it seems reasonable to pick the
smoothing splines method of Sec. 4.2.4 as it is the only method in our sample to deal

CHAPTER 5. VALIDATION, TEST CASES

82

successfully with the interference of Fig. 5.3. Also, because the smoothed splines are
calculated as a global approximation to all points, it is relatively fast compared to
other interpolation methods, that require a local calculation in the neighbourhood of
each point.

5.4 Other Existing Methods - Covering function, W

splitting, etc.
For completeness sake we should mention that besides variations on resampling
and or interpolation methods, there are some very different numerical takes on the
Bohmian interpretation that we did not explore. One such case is the derivative
propagation method (see (32]). Another one is based on the Wigner function formu
lation (see [57]) . The arbitrary Lagrangian-Eulerian method of Wyatt and coworkers
([8]) is a way to deal with insufficient data points in zones of interest for the fields,
conveniently letting us specify the grid that we wish to work with at each time step.
Finally as a way to deal with the problematic nodal zones, we mention the covering
function method of Babyuk and Wyatt ([33]), where the wavefunction is split into
two nodeless components (to be independently evolved), and in the work of Poirier
and Trahan ([34, 35, 36]) where the wavefunction is represented by two counter
propagating traveling waves.

Chapter 6
2D and Higher Dimensions
Generalization of Bohmian methods to two and higher dimensions should be rel
atively straightforward. Unfortunately the same cannot be said of the wave based
methods. When it comes to boundary conditions it is not so straightforward and even
when it comes to integrating techniques they can look quite different in lD, 2D and
higher dimensions. In fact we do not know of an equivalent transparent boundary
condition scheme in 2D to the one we used in Sec. 3.3.2 for lD. Instead we revert
to the imperfect technique of complex absorbing potentials as an approximation to
TBCs.

6.1

Wave Based Calculations

As we did in Sec. 3, we begin with the wavefunction based methods to serve as a

baseline for comparison to the trajectory methods.

83

84

CHAPTER 6. 2D AND HIGHER DIMENSIONS

6.1.1

ADI Method

To get the time evolution of the wavefunction we will use the ADI method
(Alternating-Direction Implicit see [56]). This method is second order accurate both
in space and time and it is an efficient way of solving a parabolic diffusion equation
such as the Schrodinger equation.
For a general parabolic partial differential equation (assuming D to be constant):

the 2D ADI method in discretized form reads:

t+~

'1/Jx,y

/,t+1 _

'f/x,y -

8t

t+~

t+~

+ D 28x2 ( 'l/Jx+1,y + 'l/Jx-1,y /,t+2

D 8t (n/,t+2
n/,t+2
'f/X,y +
28x2 'f/x+1,y + 'f/x-1,y '1/Jx,y
1

t+~

+ '1/Jx,y+l + '1/Jx,y-1 2n/,t+2

/,t+1
/,t+1
'f/X,y + 'f/x,y+1 + 'f/x,y-1 2'1/Jx,y

t
'1/Jx,y)

(fi.l)

/,t+1)

'f/x,y

The name ADI comes from the fact that we split the time evolution into two half
steps. In the first we mix a forward and a backwards Euler scheme for the x and y
directions and in the second half step those roles are reversed, thus earning the name
alternating direction.

6.1.2

Boundary Conditions

As we mentioned in Sec. 6, in 2D and higher dimensions we are faced with the

need to minimize unphysical reflections of the wavefunction from artificial domain
walls. To achieve this we tried a couple of different approaches, with varying degrees
of success.
We present in the next few sections the propagation of a free Gaussian in a box (as

6.1. WAVE BASED CALCULATIONS

85

0035

003
0025
002

0015
001
0005
0
100
100

Figure 6.1: Initial wavepacket (t=O)

shown in Fig.6.1) with different boundary conditions: Hard wall (Fig. 6.2), periodic
(Fig. 6.3) and complex absorbing potential (Figs. 6.5 6.4). The wavepacket is given
some initial momentum in the y direction so that after a few time steps, interaction
with the boundary can be more easily be seen.
Obviously the ideal result would be the analytical one where the wavepacket re
tains its Gaussian shape for all time and merely becomes more diffuse (larger (j)
with time. While that does not happen for any of the pseudo transparent boundary
condition methods tried, some do better than others in minimizing reflections and
retaining the Gaussian shape, as we shall see.

Hard Wall Boundary Conditions

This is just the default type of boundary condition and we include it for comparison
with the others. In this case we assume that the wavefunction suddenly encounters an
infinite potential at the domain walls and thus necessarily vanishes on the boundaries.
Unsurprisingly, as we can see in Fig. 6.2, reflections from the boundaries are quite
obvious at t=800.

86

CHAPTER 6. 2D AND HIGHER DIMENSIONS

100

Figure 6.2: Hard Wall boundary conditions in 2D, at t = 800

_j___.. L
t- 1
100

(a) Schematic of the bound

ary conditions

(b) t = 800

Figure 6.3: Periodic boundary conditions in 2D , and the resulting wavefunction

Periodic Boundary Conditions

In this case the domain walls are assumed to be transparent and periodic. The
wavefunction exits from one side, only to reappear on the oppos ite side domain wall.
As we can see by comparing Fig. 6.3 to Fig. 6.2 at the same time step, the periodic
system seems to have fewer interference fringes than the hard wall case, so it is
marginally better suited to our purposes.

6.2. TRAJECTORIES AND PHASE UNWRAPPING IN 2D

87

-1

-2

.
3

5
100
100

Figure 6.4: Profile of the absorbing imaginary potential

Complex Absorbing Potentials

In this section, we combine the marginal benefits of the periodic boundary con
ditions with a more effective complex absorbing potential "coating" on the domain
walls, as used by Muga et al. [37]. The complex part of the potential is pictured in
Fig. 6.4. The top of the potential surface represents zero potential, while the sides of
that surface are where absorption occurs.
This combination seems to be a more sensible approximation to 2D transparent
boundary conditions as judged by the relative integrity of the Gaussian shape shown
in Fig. 6.5 at the same time step, as compared to the previous figures. For longer
simulation times we still get reflection and interference phenomena that eventually
overcome the characteristic Gaussian shape of the original packet, because we are
dealing with only approximations to 2D TBCs.

6.2

Trajectories and Phase Unwrapping in 2D

Recall that to calculate the trajectories in Sec. 3.3.3 we had to unwrap the lD
phase of the wavefunction. In 2D we have the same problem, except that now we
have to monitor for jumps of 21r in any direction.
In Fig. 6.6 we display a simple case where the phase is increasing mainly in one

88

CHAPTER 6. 2D AND HIGHER DIMENSIONS

100

(a) schematic of boundary

conditions

(b) t = 800

Figure 6.5: Pseudo-Transparent boundary conditions (with an imaginary potential

near the walls) and resulting wavefunction

100

80

60

40

zo

Figure 6.6 : A simple phase angle problem in 2d

89

6.3. BOHMIAN METHODS IN 2D

.
150

.. ..

.......... .

100

..

511

100

100

100

100

(b) unwrapped phase

(a) original

Figure 6.7: A practical case of the phase problem and its solution
direction a nd in Fig. 6. 7 a more realistic case is shown, as well as its resolution.

6.3 Bohmian Methods in 2D

The basic algorithms (Sec.

3.4.2 and 3.4.3) that we used to implement the

Bohmian equations of motion remain the same. The only adjustments that we have
to make are to account for the fact that we are now dealing with position and mo
mentum vectors with N= 2 coordinates instead of just scalars. Most interpolation
techniques translate to 2-D and higher dimensions relatively straightforwardly.

6.3.1 Instance of Generalizing from ID to 2D: Least Squares

Met hod
As a simple example of how these methods transpose to higher dimensions, we
show the time evolution of an initially resting Gaussian wavepacket in Fig. 6.8.
Here we use the least squares method to calculate Q from the density of the 25
particle cloud. Note that given this particular visualization method of projecting

90

CHAPTER 6. 2D AND HIGHER DIMENSIONS

-6 ' - - - - - - - - ' - - - - - ' - - - ' - - - - - - - - - ' - - - - ' - - - - _ _ _ J

-24
-22
-20
-18
-26
30
28

Figure 6.8: 2D Bohmian evolution of an inhomogeneous packet as viewed from the

z-axis
a cross-section of what really is a 3D surface (2 space+ l time) into a 2D picture,
the trajectories may appear to cross. In the actual 3D space the non-crossing of
trajectories is of course still observed.
As one increases the number of particles the computational effort should scale
linearly, independent of the dimensionality of the space in which we are working.
This is in marked contrast to traditional wave based methods, where in going from
lD to 2D we square the number of grid points, and in general that number scales
exponentially as ND, leaving us with no choice but to use coarser grids when working
in higher dimensions. Here lies one of the big advantages of Bohmian methods where
the grid stays with the wavefunction.

Chapter 7
Cellular Automata and Lattice
Boltzmann Methods
In this chapter we explore a possible connection between simulation of Bohmian
mechanics and the world of cellular automata, of which lattice Boltzmann methods
are a subset.

7.1

Cellular Automata

Since they were invented by von Neumann is the 1940s, cellular automata have
been successfully used to simulate an enormous range of systems, including bacterial
growth, forest wildfires, rush hour traffic in a busy city and classical fluid dynamics
[39, 40). These cellular automata operate according to a tantalizingly simple set of
rules (that is supposed to capture the microphysics of the process), from which the
macroscopically complex behaviour of the system gradually emerges. It is conceivable
that by using a proper set of rules we may be able to reproduce the behaviour of a
quantum fluid in the Bohmian sense, thus approaching the problem from a completely
91

92CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

different viewpoint .
Recently re-popularized by Steven Wolfram [40], Cellular Automata or CA have
been with us for decades. As we mentioned they are particularly well suited to model
complex and chaotic systems as well as more "classical" problems. This versatility
should not surprise us since they have been demonstrated [42] to be formally equiv
alent to a Thring machine, so theoretically they are capable of carrying out any well
posed calculation. To emphasize this point we should note that the calculations of
previous chapters were done on Thring machines (standard computers).
The question remains, are there are advantages to using cellular automata in the
context of time evolution of a quantum system, as opposed to more conventional
techniques , and in particular can we marry the framework of Bohmian mechanics to
the gear of cellular automata? That is what we explore in this section.

7.1.1

Conway's Game of Life

There are many and interesting cellular automata implementations that have been
studied ([40]). The reader is probably already familiar with at least one of these. In
Conway's "Game of life", a simple set of rules gives rise, totally det erministically, to
quite complex behaviour of its constituents, mimicking organisms that live, die and
even self replicate on the computer screen. Incidentally a Thring machine has been
constructed with Conway's game of life "parts"; it is described in [43].

7.2. CELLULAR AUTOMATA AND FLUID MECHANICS

7.2

93

Cellular Automata and Fluid Mechanics

7.2.1

Lattice Gas Cellular Automata (LGCA)

Historically, these models are the original catalysts for the usage of cellular au
tomata in fluid simulation. Originally introduced in the 1970's by Hardy, Pomeaux
and de Pazzis, these simulations consist of an underlying lattice on which two basic
steps are performed to advance in time :

A free streaming step, where particles move with whatever momentum has been
assigned to them.
A collision step, where particles collide with each other in momentum and par
ticle number conserving exchanges.

Finally if at some point in time we wish to recover the physically relevant macroscopic
fields, an averaging over many lattice sites is done.
Amazingly, together with a judicious choice of underlying lattice, this is all that
it takes to make a respectable simulation of a simple fluid. Following are some his
torically famous choices for the lattice and the corresponding collision rules.

"HPP" LCGA
This is the original model by Hardy, Pomeaux and de Pazzis (HPP). Here, the
chosen lattice is square, and the simple collision rules are illustrated in Fig. 7.1. This
is mostly a toy model and the simple choice of lattice ends up having consequences in
terms of the resultant flow not being isotropic. In fact this lattice gives rise to fluid
vortices which are square instead of circular.

94CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

Incoroming

- H - =>

Outgoing

1
T
Figure 7.1 : Lattice and collision rules for HPP model

"FHP" LGCA

Later introduced by Frish, Hasslacher and Poumeaux, this CA does not suffer from
the anisotropic anomaly of the HPP model. This is thanks to using an hexagonal
lattice and a correspondingly larger set of collision rules (see Fig. 7.2).

95

7.2. CELLULAR AUTOMATA AND FLUID MECHANICS

a) _____.

t--

I
\

b) _____.

=>

=>

d)

-----.0

=>

e)

_____.0t--

=>

!)

_____..-

g)

=>

OR\

=>

c)

G------.

<
I \

e6

OR

A r
OR

I =>~OR~
1\
conserving
0

(a) Momentum
collision rules

es

e1
ez

e!\e,

(b) underlying lattice

Figure 7.2: Lattice and collision rules for FHP model

Other Lattices
As you move to higher dimensions, the complexity of the lattices necessary to
preserve isotropy quickly escalates and the number of collision rules that we need to
track, rises exponentially.

Figure 7.3: Example of 3D lattice for a LGCA model

The main advantage of LGCA methods for 2D is the small storage requirement
and fast collision calculations (dealing with logical operations) given the Boolean
nature of the collision operators. In the 1980's and 1990's a couple of machines were

96CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

specifically built to efficiently implement these methods in hardware. On the other

hand modern computers are more and more engineered towards efficient floating point
calculations.

7.3

Lattice Boltzmann Methods

One of the disadvantages of the previously mentioned LGCA methods is their high
statistical noise, a consequence of their discrete nature whereby a lattice direction
either is or is not occupied. A natural evolution from these ideas is presented in
the Lattice Boltzmann method where one replaces the Boolean particle number by
its average value resulting in a continuous density distribution function f(f, iJ, t),
defined in phase space . The evolution off is governed by the Boltzmann equation

(7.1)
Here

Vx

the velocity and

Fx

the force in the x direction. The collision operator D can

be a very complex function that needs to be approximated. Approximation of D by

Aj(fj - Jr) constitutes the quasi-linear LBE method. Going one step further we
get the fully linear approximation, the BGK (Bhatnagar,Gross, Krook [58]) collision
operator:D = ~(fo- f).

7.4

Cellular Automata and Quantum Mechanics

Now that we have established a link between cellular automata, lattice Boltz
mann and fluid mechanics, we use Succi ([50]) as a guide to illustrate the link in one
dimension between the Dirac equation and lattice Boltzmann methods.

97

7.4. CELLULAR AUTOMATA AND QUANTUM MECHANICS

7.4.1

Dirac Equation

In 3D the usual way of writting the Dirac equation is,

with

CJx,y,z

ax,y,z =
[

] ' (3 = [ 0 I ] and where a x,y,z are the usual Pauli

-CJx,y,z
I 0
0

matrices:

CJx,y,z =

7.4.2

0 1) ' (
1 0
'/,

-i ) ' ( 1

0 )
0 -1

Majorana Representation

In the Majorana representation, where w1 = (a~k

+ if31k)wk , the Dirac equation

is represented by

(7.2)

here J.L

0, 1, 2, 3 and Wlk

r51k ,W1

= ax,

matrices). The modified mass term M1k

W2

= -imaY

(3 , W 3 =

+ qV61k

-az

(all real 4 by 4

contains the rest mass

and an external potential term. This form is reminiscent of a 4-phase fluid (the two
Dirac spinors each with their 2 components) interacting via a scattering matrix

Mjk

, where u and d represent the

Restricting ourselves to 1-D and taking W =

dr

d!
top and bottom Dirac spinors with components u 1 = ur, u 2

u!, d 1

dr, d2 = d!,

98CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

eq. 7.2 becomes:

8tu1,2- azu1 ,2 = md2,1 + igu2,1

(7.3)

{ 8td1,2 + 8zd1,2 = -mu2,1 + igd1 ,2

7.4.3

Formal Connection to LBE

Looking at eq. 7.3 and eq. 7.1 , if we make the substitutions

f 1e = ~
h
9

, f ?. =

/i

d, u , vi = =F1 ,

-~!I , we can make a formal mapping of the Dirac equation onto

9

two Lattice Boltzmann equations, one for fi

= d and another for f i = u,

That is to say that the quantum system is formally equivalent to a 1-D Lattice
Boltzmann system with four channels (up and down spinors) that will intermix be
cause

7.5
7.5.1

!I

depends on

and vice versa .

Numerical Experiments
Algorithm

We use natural units as in Sec. 3, setting

n = c = 1.

If we discretize the eqs. 7.3

in a Crank-Nicolson scheme we obtain:

ui,2- u1,2 = W'(d;,2d2,1) +

{ d;,2 - d1 ,2 =

Where the

hat ~

i~(u2,1 + u2,1)

W' (u2,~1 +u2,1) + iHd;,2 + d1,2 )

(7.4)

on a variable means that time has advanced by one step. This is an

implicit scheme, as can be seen by the presence of advanced terms on both sides of

99

7.5. NUMERICAL EXPERIMENTS

the equation. To make it explicit and refer only to known values, we solve the linear
system for

u and d, giving us the following equations:

Where the coefficients a and bare given by:

a=
{ b-

~1-0/4
1+0/4-ig

m
1+0/4-ig

This convenient linear system will be the basis of our lattice propagation and
collision calculations. Note that if desired we can easily include non-linear and/or
self-consistent terms by suitably changing the potential term g (i.e. if we want it to
be a function of the density we can make it a function of (u 2 + d2 )). In that case the
coefficients a and b will no longer be constant and will have to be recomputed at each
time step.

7.5.2

Quantum Systems studied

To test the algorithm we are going to use similar systems to those of Sec. 5.1,
whose analytical solutions are well known.

Free Packet
We start with a Gaussian packet centered in the middle of our array and given a
certain momentum p towards the right. The fact that the Gaussian disperses in time

lOOCHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

WavefuncUon Evolut1on
~~~~--~

~~--~~~

800

1000

1200

1400

200

400

600

800

1000

1200

1400

1600

1800

2000

Spe.te

Figure 7.4: Free packet dispersion (2 different momenta)

is clearly visible in Fig. 7.4. How that dispersion relates to the expected analytical
value is represented in Fig. 7.5. Also represented in the same figure is the average
value for the position of the packet, which follows a path consistent with Ehrenfest's
theorem. The expressions for both expection values are:

Po

<X>= Xo + -t
m

and

a(t)

Looking carefully at the graphs of Fig. 7.5 we note that as the wavepacket ap
proaches the boundary our values deviate further from the theoretical values, as we
start to observe some numerical reflection from the walls. This is particularly visible
in the right panel of Fig. 7.4.

101

7.5. NUMERICAL EXPERIMENTS

Packet position <X:o

1\40,r---.--.--..----r----'~~-~-~~~

1120

180

1\00

160

1060

100

80
1040

1020'-----'----'--'---___.,._
0
200
400
600
800

__,____,,--_.___
1000
1200
1400
Time

_.___,___J
1600

1800

200

2000

400

600

800

1000
Time

1200

1400

1600

1800

2000

Figure 7.5: Mean position (left) and standard deviation (right) of the wavepacket vs
time

Packet positiOn <X:>

Wa.vefuncllon EvolutiOn

1300

500

1200

1000

1\00

~ 1500

1000

2000

900

2500

600

3000

200

400

600

800

1000
Spooe

1400

1600

1600

2000

7000

500

1000

1500
Time

2000

2500

Figure 7.6: Wavepacket in an quadratic potential (left) and its average position (right)

102CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

Figure 7.7: Wavepacket hitting a barrier

Harmonic Oscillator
In the next case we begin with a Gaussian packet with zero initial momentum
placed in an harmonic oscillator potential. In Fig. 7.6 we see when plotting the mean
position of the packet that there is some numerical dissipation present as the position
is slightly smaller after one period.

Barrier Penetration
Finally in the last case, instead of having zero potential everywhere, we place a

20 unit long barrier at position 1300. In Fig. 7.8 intense interference can be seen at
the edge of the "wall" and a weak evanescent wave continues along the interrupted
path of the Gaussian packet, across the barrier to the left side of Fig. 7.7. Here we
don't face the same problems with nodes or interference as in the Bohmian methods
we considered before. The reason for this is that we are dealing more with a wave
based method closer in its internals to our reference Moyer method of Sec. 3.3.

7.6. CRITIQUE OF THE MAJORAMA LBE METHOD

(a) t=800

103

(b) t=llOO

Figure 7.8: Interference and exponential decay inside the potential barrier

7.6

Critique of the Majorama LBE Method

As we have seen the Majorama Lattice Boltzmann method seems to work, however
as we just saw it really should belong to the class of wavefunction based methods
such as the split operator method we used previously in 3.3. We still have hard
domain walls and a fixed space-time grid, with all the associated limitations. What
this method does do however, is give us a glimpse of a connection between cellular
automata and quantum simulation, and in doing so it gives us a hint that an equivalent
Bohmian based cellular automaton may also be devisable.

7.7

A Look Back at the Bohmian Picture

Looking back at the Bohmian equations of motion, with Q = - ;~

'iljf and where

gt represents the material derivative:

{

~~ + \7 (pv) =

0 {::}

~~

-~\l(V + Q)

+ (v\l)v =

~ + p \7 v =
=

~~

And then looking at the equation for a classical fluid (Euler equation):

(7.5)

104CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

avi

at

+ (vV)vt

1
m

= --811;t

1
mp

-86
J t]P

(7.6)

We see that the first equation 7.5 expresses the conservation of probability, and the
second maps term by term onto the classical fluid equation 7.6 except for the pressure
term that must map to Q for the analogy to hold. In fact ([15]), if we substitute the
pressure tensor [{1 by O"ij

= -

';~Paij log p plus any tensor whose divergence is zero,

we get two formally equivalent equations.

So a Bohmian system is, at least formally, equivalent to an inviscid (i.e. time
reversible and non-dissipative) irrotational and compressible fluid. When trying to
extrapolate this fact to the lattice Boltzmann framework we will have two macroscopic
real fields to work with, v the velocity and p the probability density, an analog of the
fluid density.

7. 7.1

Lattice Boltzmann Revisited

We reintroduce the Boltzmann formalism in a lattice space with N discretized

velocity vectors Vi. In the i direction, the Boltzmann equation in the BGK([58])
approximation reads:

At each lattice site, these N discrete directions combine in the following way to
reconstitute the real macroscopic fields:

105

7. 7. A LOOK BACK AT THE BOHMIAN PICTURE

(7.7)

higher order terms

where the lattice specific weights

mi

are determined in order to preserve isotropy. For

instance, in the D2Q9 lattice, a two dimensional lattice with 9 velocity vectors, the
diagonal terms will have to be weighted differently since they have different lengths.
Now fo can be quite complicated, but we are going to consider a Taylor expan
sion in the macroscopic fields v and p, thus establishing a connection between the
macroscopic equilibrium and the collisional microphysics.

0
fi

= ap

fii.il
+ bp+ CPViaVtf3Ua Uf3 + ....
v2

Here a, b and c are coefficients to be determined; p and v are the macroscopic fields,
and ui are the lattice vectors.
We can now insert this Taylor expanded form of the Boltzmann equilibrium func
tion in the equations of eq. 7. 7. The first equation implies that in a "collision" or
"step"

L mdP =

p, which will give a lattice-dependent value for

a = ~). Similarly the remaining coefficients of

fP

a (in the D2Q9 case

can be determined ([52]), given

enough conservation relations as those of eq. 7. 7.

7.7.2

Lattice Boltzmann Method and Bohmian Mechanics

In general, the Lattice Boltzmann method works for problems that can be set
up in the form equations of continuity which are the result of one or more conserva
tion laws([53]). For instance for some hypothetical macroscopic fields Y and Z, the
conservation relations for the zeroth and first moments,

106CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

(7.8)

together with the space-time evolution given by the Boltzmann equation f(r+f:lte"a,, t+
f:lt)

JP + (1 -

~)

JP,

will generate t he macroscopic differential relation (of conti

nuity):

a~~) + \l (Z) = 0 + O(c)

So for instance if we want

fJf -

\7 (pu)

(7.9)

= 0 to apply, we can identify Y =

p and

Z = pu, which are generated if we set X= 1, giving:

these may now used to determine coefficients in our particular form for t he function
f~, as shown in Sec. 7.7.1.

Bohmian Case
We need to work with the Bohmian equations of motion, in an Eulerian frame
since the LBM lattice is fixed in space:

fff+ \lpu=O
~~

+ u\lu = - ~ \l(V + Q)
Q=

(7.10)

_.!f_ \lyP
2m yP

We already saw in the previous section how we can generate t he first of these
equations, the probability conservation relation by setting X

1, Y

= p

and Z

= pu.

7. 7. A LOOK BACK AT THE BOHMIAN PICTURE

Now we wish to generate something of the form, with Q

8u

-8t + u\lu =

La f~ =

We certainly cannot use

We can however set X=

e~

107
= -

2~ "j/

--\l(V + Q)
M

u for this one since we already set it to p.

in eq. 7.8, generating:

(7.11)

We now see that Y

pu this time and Z should be chosen so that the final

equation coincides with the Bohmian momentum equation . This is a little bit tricky
as we shall see. Looking at the equations 7.11 we have set Y = pu, so we want a
term in ~ =

P&u
&t

!!1!:'
!:.
&t

u!fjf + p~~.

+ u\lpu =

!!1!:'
!:.
&t

We can use the equation of continuity here, and obtain

+ \lupu- pu\lu

Substituting back in eqs. 7.10 we get:

!f/i+Vpu=O
~

+ Vupu- pu\lu + pu\lu = -if\l(V + Q) {::}

Q

8pu

ot

+ \lupu = - M \l(V + Q)

= - !_ "v'P
2m

,;p

which is equivalent to an Euler Fluid with a pressure term -iJV(V

+ Q)

(or an

external force term).

We can further pull the potential term inside the differential operator, thus ex
pressing all the Bohmian equations in terms of continuity relations

108CHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

~+\lpu=O

+ \1 (upu + if(V + Q)) = 0

Q=

!i... '\l.,ft
2m ..;P

Thus our Z should be equal to upu+if(V +Q) , giving the complete set of conservation
relations for our model:

~ccfg = P
~a fg eAa = pu

(7.12)

~a fgeai"e~j = upu + if(V + Q)

ow we define

(7.13)

which , together with the 2DQ9 lattice properties:

will set some constraints on the coefficients in eq. 7.13 via the same procedure of Sec.
7.7.1. The resultant behaviour of this lattice Boltzmann system should be a close
analog of the Bohmian equations:

7.8. COMMENTS

109

Wf + V pu =
~

0 + 0 (E2 )

+ V (upu + fi(V + Q))

Q

0 + ER2

+ 0(E 2 )

= _.!f._ '\ly!p
2m yip

with an extra term ([59]) R2, a viscous term that can be tuned via
(T-

8 n;~)
2 ) ( 8t
8x
0
1
1

88t P,~
2 ) 'th
8xk Wl
2

0 _

7Tij -

"'

LJa

T,

since R2 =

f.OaeaieaJand pO~jk _- "'

jO
LJa aeaieaJeak

We pause here because we realize that by setting Z

= upu

+ fi(V + Q) in eqs.

7.12, we are going to end up with Q intact in the Boltzmann distribution function

f.

That is to say that, unlike in the standard lattice Boltzmann method, where we have
constant coefficients in

f, here at each different lattice point, we are going to have to

compute V 2 p and V p, which is an enormous extra overhead for the method.

7.8

Comments

First it should be noted that the classic Lattice Boltzmann method already has a
significant overhead on its own (since it deals with N velocity vectors at each site),
this hindrance is usually compensated by the ability to tune the microphysics to deal
with specialized problems, complex boundaries, phase mixtures, etc.
Secondly, our Bohmian lattice Boltzmann method is set in an Euler frame (see
eq. 7.10), so it does not benefit from the speed increase yielded by sparse grids of the
Lagrangian implementations of previous chapters.
Finally, to my knowledge the only other way to include the quantum term in the
Lattice Boltzmann theory is in the form of an external applied force. This means that
the quantum force term still has to be calculated at each lattice point and does not
"emerge" as some sort of collective behaviour (as say viscosity does in classic LBM).
Given the non-local form of Q, this is a very expensive computational operation to

llOCHAPTER 7. CELLULAR AUTOMATA AND LATTICE BOLTZMANN METHODS

perform at each point.
Having concluded that the lattice Boltzmann method is probably not the optimal
way to implement the Bohmian equations, we should take care not to discount cellular
automata methods in general as a means of a quantum simulation. We are aware of
the amazing things that different cellular automata can simulate ([43]) and other CA
methods certainly will warrant further investigations.

Chapter 8

Conclusion
"I think I can safely say that nobody understands quantum mechanics"
Richard Feynman [55]

On a conceptual level it is somewhat comforting that Bohmian mechanics provides

us with an ontological continuity between the worlds of the microscopic and macro
scopic, quantum and classical, as we make use of the same concepts of trajectories,
velocities, etc. which are familiar in the macroscopic world. This is what drew us
to explore some aspects of the numerical side of Bohmian mechanics. To this effect
we implemented fixed grid methods in lD and 2D from which Bohmian trajectories
were calculated. After a partial survey of direct numerical Bohmian methods, we
considered different interpolation methods as ways of improving the stability of the
calculations. In particular, we concluded that the smoothed splines method was more
robust when simulating a model system for interference. Finally we investigated a
possible link between Lattice Boltzmann Methods of theCA world and our Bohmian
equations of motion.
We have concluded that as a numerical tool, direct Bohmian methods have some

111

112

CHAPTER 8. CONCLUSION

stability shortcomings in specific situations (node problem, interference) , and some

performance drawbacks (Lattice Boltzmann). One has to realize though , we can't get
something for nothing and this comes as a trade off from the extreme sparseness of
the grids it uses which makes this method a viable alternative in situations where the
dimensionality of the problem makes traditional grid methods with high resolution
impracticable. Furthermore we saw problems that afflict Bohmian methods can be
minimized by judicious use of point resampling, appropriate interpolation techniques
and dynamical time integration scales.
Despite these remedies, one should not indiscriminately use Bohmian methods to
solve every numerical problem, just as it would be unwise to use fixed grid methods
to approach high dimensional problems, or situations where specific boundary condi
tions are required. (An exception is for lD where TBCs are practical for grid based
methods).
These two aspects, sparse grids and ability to impose arbitrary boundary condi
tions are then the major strengths of the Bohmian methods. These, incidentally are
the shortcomings of fixed grid methods, just as the ability to deal well with nodes
and interference is a source of trouble in the Bohmian methods , in a complementary
fashion.
We should take away the idea that Bohmian methods and fixed grid methods
complement each other nicely, and that , armed with both views we are much bet
ter equipped to face such numerical problems. This duality between these disparate
approaches to the same numerical problems would almost be disconcerting were we
not talking about quantum mechanics, the mother of innumerable dualities and para
doxes ....

Bibliography
[1] Daniel Styer et al. "Nine Formulations of Quantum Mechanics", Am. J. Phys. 70
(2002) 288-97.
[2] L. de Broglie, "Recherches sur la theorie des quanta" Doctoral thesis, Universite
de Paris (1925)
[3] L. de Broglie, Ann. de Phys., Paris 3 (1925) 22.
[4] E. Madelung, "Quantentheorie in hydrodynamischer Form", Zeit. Phys. 40 (1926)
322-6.
[5] Bacciagaluppi, Valentini: "Quantum Theory at the Crossroads: Reconsidering the
1927 Solvay Conference", quant-ph/0609184, (2006)
[6] David Bohm, "A suggested interpretation of the quantum theory in terms of "hid
den" variables", Phys. Rev. 85 (1952) 1:166-79; II: 180-93
[7] D. Diirr, S. Goldstein and N. Zanghi, "David Joseph Bohm: 1917-1992", Founda
tions of Physics Letters, 6 (1993) 551-4; alsoP. Holland and J-P Vigier, Founda
tions of Physics 23 (1993) 5-6.
[8] R. E. Wyatt, "Quantum Dynamics with Trajectories", Springer Science, (2006)
[9] Hua Wu and D.W.L. Sprung, Phys. Lett. A 183 (1993) 413-7.
113

114

BIBLIOGRAPHY

[10] C.L. Lopreore and R.E. Wyatt, "Quantum Wave packet dynamics with trajecto
ries" , Phys. Rev. Lett. 82 (1999) 5190-4.
[11] F. Sales Mayor, A. Askar, and H.A Rabitz , "Quantum fluid dynamics in the La
grangian representation and applications to photodissociation problems", J. Chern.
Phys. 111 (1999) 2423-35
[12] A. Matzkin, " Are Bohmian trajectories real? On the dynamical mismatch be
tween de Broglie-Bohm and classical dynamics in semiclassical systems", quant
ph/ 0609172 ' (2006)
[13] W. Pauli, "Remarques sur le probleme des parametres caches dans la mecanique
quantique et sur la theorie de l'onde pilote" Andre George Ed. , "Louis de Broglie
Physicien et Penseur", Paris: Albin Michel
[14] W.

Myrvold,

national

Studies

"On some early objections to

in

the

Philosophy of Science,

Bohm's Theory",
17

(2003)

Inter

7-24,

see

http: / / publish.uwo.ca/ -wmyrvold/ pub.htm

[15] P. R. Holland, "The Quantum Theory of Motion", Cambridge University Press,
(1993)
[16] H. Goldstein, "Classical Mechanics", Addison-Wesley (1950)
Note that subsequent editions omit the discussion on the relationship between
classical Hamilton-Jacobi theory, Geometrical Optics and Wave-Mechanics
[17] A. Einstein, B. Podolsky, N. Rosen, "Can Quantum-Mechanical Description of
Physical Reality Be Considered Complete? ", Phys. Rev. 47 (1935) 777- 780.
[18] Y. Nogami, F. Toyama and W. van Dijk, Phys. Lett. A 270 (2000) 279-287.

BIBLIOGRAPHY

115

[19] C. Phillipidis, C. Dewdney and B.J. Hiley, Nuovo Cimento 52 B (1979) 15-28.
"Quantum interference and the quantum potential"
[20] C. Moyer, "Numerov extension of transparent boundary conditions for the
Schrodinger equation in one dimension.", Am. J. Phys 72 (2004) 351-360.
[21] W. van Dijk, F. M. Toyama, "Accurate numerical solutions of the time-dependent
Schrodinger equation" , physics/0701150v1 or Phys. Rev. E 75 (2007) 036707
[22] S. Garashuk, V. Rassolov, "Semiclassical Dynamics based on quantum trajecto
ries, Chern. Phys. Lett. 364 (2002) 562-7.
[23] http:/jphysics.nist.gov/cuu/Constants/
[24] E.J. Heller , J. Chern. Phys. 62 (1975) 1544-55.
[25] M. Ehrhardt, "Finite Difference Schemes on unbounded Domains" , World Sci
entific, 2005, pp. 343-384
[26] A. Arnold, M. Ehrhardt, I. Sofronov,"Discrete transparent boundary conditions
for the Schrodinger equation: Fast calculation, approximation and stability",
Comm. Math. Sci. , (2003)
[27] Zisowsky A, Arnold A, Ehrhardt M and T. Koprucki, "Discrete transparent
boundary conditions", Zeithschrift fiir angewandte Mathematik und Mechanik 58
(2005) 793-805.
[28] Gray S.K., Noid D.W. and Sumpter B.G., J. Chern. Phys. 101 (1994) 4062-72.
[29] J. Candy W. Rozmus, J. Comput. Phys. 92 (1991) 230-56.
[30] H. Yoshida, Celest. Mech. Dynam. Astron. 56 (1993) 27

BIBLIOGRAPHY

116

[31] W. Press, S. Teukosky, W. Vetterling, B. Flannery , "Numerical Recipes: The

Art of Scientific Computing", Cambridge University Press, (2002)
[32] C. 'frahan, K. Hughes and R. Wyatt, "A new method for wave packet propa
gation: Derivative propagation along quantum trajectories", J. Chern. Phys. 118
(2003) 9911
[33] D. Babyuk and R. Wyatt, "Coping with the node problem in the hydrodynamic
representation of quantum mechanics: The covering function method", J. Chern.
Phys. 121 (2004) 9230
[34] B. Poirier, "Reconciling semiclassical and Bohmian mechanics:

I stationary

states", J. Chern. Phys. 121 (2004) 4501-4515

[35] C. Trahan and B. Poirier, "Reconciling semiclassical and Bohmian mechanics: II
Scattering states for discontinuous potentials", J. Chern. Phys. 124 (2006)
[36] C. Trahan and B. Poirier, "Reconciling semiclassical and Bohmian mechanics:
III Scattering states for continuous potentials", J. Chern. Phys. 124 (2006)
[37] J.G. Muga, J.P. Palao, B.

avarro, I.L. Egusquiza, "Complex absorbing Poten

tials", Phys. Rep. 395 (2004) 357-426

[38] C. de Boor, "A Practical Guide to Splines", Springer-Verlag, (1978)
[39] S. Wolfram, "Theory and Applications of Cellular Automata", World Scientific
(1986)
[40] S. Wolfram, "A

ew Kind of Science", Wolfram publishing (2002)

[41] S. Wolfram, Cellular automaton fluids 1: Basic theory. Journal of Statistical

Physics, 45 (1986) 471-529.

BIBLIOGRAPHY

117

[42] K. Zuse, "Rechnender Raum", Friedrich Vieweg & Sohn, Braunschweig, (1969)
[43] A. Adamatzky et al, "Collision-Based Computing", Springer, (2002) , 491
[44] J Maddox and E. Bittner, J. Chern. Phys. 115 (2001), 6309-16
[45] B. Kendrick, J. Chern Phys. 119 (2003) 5805-17
[46] T. Usuki, M. Saito, M. Takatsu, R.A. Kiehl and N. Yokoyama, Phys. Rev. B 52
(1995) 8244.
[47] R. Akis and D.K. Ferry and J.P. Bird, "Numerical simulation of quantum dots",

Super. and Microstruc., 20 (1996) 323-9.

[48] L. Shifren, R. Akis, D.K. Ferry, "Correspondence between quantum and classical

motion" Phys. Lett. A 274 (2000) 75-83.

[49] Hua Wu and D.W.L. Sprung, Phys. Lett. A 196 (1994) 229-236.
[50] S. Succi ,"The Lattice Boltzmann Equation for Fluid Dynamics and Beyond" ,

Oxford University Press (2001)

[51] S. Succi, "Lattice Boltzmann Equation: Failure or Success?", Physica A 240
(1997) 221
[52] J. G. Zhou, "Lattice Boltzmann Methods for Shallow Water Flows", Springer
(2004)
[53] D. Wolf-Gladrow, "Lattice-Gas Cellular Automata and Lattice Boltzmann Mod

els: An Introduction", Springer (2000)

[54] G. McNamara and G. Zanetti, Phys. Rev. Lett. 56 (1988), 1065-70

BIBLIOGRAPHY

118

[55] R.P. Feynman "The Character of a Physical Law" , Cambridge Mass. (1965)
[56]

. Gershenfeld," The Nature of Mathematical Modeling", Cambridge University

Press, (1999)
[57] A. Donoso, C.C. Martens," Quantum tunneling using entangled classical trajec
tories", Phys. Rev. Lett. 87 (2001) 223202
[58] P. Bhatnagar, E. Gross and M. Krook, "A model for collison processes in gases",

Phys. Rev . 94 (1954) 511-525

[59] Y. Guangwu , C. Yaosong and H. Shouxin, "Simple lattice Boltzmann model for

simulating flows with shock wave ", Phys. Rev. E 59 , 454- 459 (1999)

Appendix A- Transparent Boundary

Conditions
This is a summary of the algorithm we used, based on Moyer's([20]) implementa
tion of transparent boundary conditions (TBC).

It is a fixed grid method whose core is the Cayley approximation to the system
propagator:

ll!(x, t

+ !:!.)

1- iH .6./2
~ 1 + iH .6./2 ll!(x, t)

This is rewritten with y(x, t) = ll!(x, t+l:!.) + ll!(x, t), a function of two consecutive
time steps as:

11
.2) y='t
.4
y
- ( V-'t-

.6.

or y" = gy + f with g = V- 2i and

,6.

i-;!: , a form suitable to be discretized on a

uniform grid with spacing h. The Numerov scheme is used resulting in:

(8.1)
where d = 1- h 2 f2 and w = dy- h 2 --f-2 . We see that Eq. 8.1 is a recursion relation
119

BIBLIOGRAPHY

120
that depends on three consecutive points j -1, j, j

+ 1.

It can be put in terms of two

recursion relations that involve two new variables e and q defined by wi+ 1 = e1w1 +q1 :

To enforce the TBCs we define at the left boundary (j = 0), w1+1 = aw0

eowo

+ qo

+ {3

, with a and {3 determined according to

eo = a = ao
{ qo

J a5 -

d0(a 0- eo)Wo + do(ao- ao) 2:~= 1 ln-k+lw~

subject to Je 0 J > 1 , with a0 = 1 + h 2 g0 /2d 0 and d0 = 1 - h 2 g0 /12 (the letter super

scripts denote time steps and subscripts the position). The last term, a convolution,
implies that a history of W must be kept at the boundary, constituting an increasing
performance penalty as the number of time steps increases. The convolved value
has the form

l = exp(-in ) (P, ( ) _ P, _ ( ))
n

2n _ 1

n 1-L

n 2 1-L

where Pn is the Legendre polynomial of order n, and

..\- 2h2
t:,.

c = 1- i~
6d

a= 1 + h 2 ..9...
2d

ln

121

BIBLIOGRAPHY
The same procedure applied to the other border on the right yields,

which, using the definition of w, gives '11 advanced in time:

w~+l =
J

h2 wn
w'r!
1 + _J
-wn + i3L).
-d.
d.
J

Appendix B- MATLAB Code

Listings

Split Operator code / Fixed Grid Code

%% SYSTEM TIME AND SPACE

SIZE=512
%cells

XL=-40 ;XR=40;
%nanometers

Tl M ESTEPS=2000

eV=1.6022*1e-19 ;

hbar=1.0546*1e-34;

c_ light=2 .9979*1e8 ;

Melectron=0.51099*1e6;

Meffective=0 .067; % GaAs

MASS=Melectron*Meffective; %eV

MASSE=MASS ;

MASS _ KG=MASS*eV/c_light A2;

m=MASS*eVjc_ light A2;

MASS=l/2;
Lenght_scale=l ;
Time_scale=le3*hbar/(2*m );
Energy_ scale=eV*2*m*le-18/hbarA 2;
Velocity_scale=le-3*2*m/hbar;
M=m ;
fsec= Time _ scale;
%Check
cprime=c _ lightjle6*Velocity_ scale ;
%timestep in femtoseconds
DELTA=.! *fsec;
J=SIZE ;
H=(XR-XL)/(SIZE-1) ;
%grid size

122

BIBLIOGRAPHY
X=XL:H:XR;
temp=X*O;
%% SYSTEM MASS AND POTENTIAL
Vj=temp; % Potential in eV
%Vj=Vj+1e3*exp(-112/.0125A 2*{X-. 75). A2);
%X=X*Xmsca le;
Vj=Vj+(X>-10)&{X<-5)+{X>5)&(X<10);
Vj=Vj*.25;
Vj=Vj*Energy _scale;
%Initial wavefunction
PsiO=temp;
wave_ sigma=sqrt{10);
X0=-25;
K=0.31;
PsiO=exp( -112lwave _sigma A2*{X-XO). A2). *exp(i*K*X);
%Normalize
%Psi0=Psi0*1lsqrt( norma liz(abs(PsiO. *PsiO),X) );
%SYSTEM DEFINITION ENDS
PSI=zeros{TIMESTEPS,SIZE);
PSI{1,:)=Psi0;
ALPHAj=temp;
%Fixed
Gj=Vj-2*iiDELTA;
Gj=Gj*{2*MASS);
Dj=1-H A2I12*Gj;
Aj=1+H A2*Gj.I{2*Dj);
j=J;
niuJ={1-abs(Aj(j) A2)) I abs{1-Aj(j) A2);
j=1;
niu1={1-abs(Aj(j) A2)) I abs{1-Aj(j) A2);
% If V1=V J then niu1=niuJ symmetrical
%Calculate all LEGENDRE coeficients (#timesteps) ahead of time index
% 1->n=O
%REPEAT 1 For X=1
j=1;
lambda=2*HA2IDELTA; %Fixed
c=1-i*lambdai{6*Dj(j)); %Fixed
phi=angle{{Aj(j). A2-1).1c); %Fixed
niu={1-abs(Aj{j) A2)) I abs{1-Aj(j) A2);
% PLEASE NOTE PN{n) is actually Legrendre polynomial of order n-1
% So use to refer to PN{m) m=n+1
PN=temp;LN=temp;
PN{1)=1;PN{2)=niu; % These are PNO and PN1
for m=2:{TIMESTEPS-1)

123

124

BIBLIOGRAPHY
%generate Legendre polynomials;
n=m-1;
PN ( m+ 1)=2*niu*PN ( m)-PN( m-1 )-( niu*PN( m)-PN ( m-1)) I(n+ 1);
end ;
LN(1)=niu*exp(-i*phi); % LN(O) would be -1; LNn is LN(n)
for n=2:(TIMESTEPS-1)
m=n+1 ;
LN( n)=exp( -i*n*phi)l(2*n-1 )*(PN( m)-PN( m-2) );
%LN ( n)=exp( -i*n*phi)*( niu*PN (m-1 )-PN ( m-2) )*(2*n+ 1) I (2*n-1) I(n+ 1);
end;
LN1=LN ;
%REPEAT 2 for X=J
j=J;
lambda=2*H A2IDELTA; %Fixed
c=1-i*lambdai(6*Dj(j)) ; %Fixed
phi=angle((Aj(j) . A2-1).1c); %Fixed
niu=(1-abs( Aj(j) A2)) I abs(1-Aj(j) A2);
% PLEASE NOTE PN(n) is actually Lagrage polynomial of order n-1
%So use to refer to PN(m) m=n+1
PN=temp;LN=temp;
PN(1)=1 ;PN(2)=niu ; % These are PNO and PN1
for m=2:(TIMESTEPS-1)
%generate Legendre polynomials;
n=m-1;
PN ( m+ 1)=2*niu*PN ( m)-PN( m-1 )-( niu*PN ( m)-PN ( m-1)) I(n+ 1);
end ;
LN(1)=niu*exp(-i*phi); % LN(O) would be -1 ; LNn is LN(n)
for n=2:(TIMESTEPS-1)
m=n+1 ;
LN( n)=exp( -i*n*phi)l(2*n-1)*(PN( m)-PN( m-2) );
%LN ( n)=exp( -i* n*phi)*( niu*PN (m-1 )-PN( m-2) )*(2*n+ 1) I (2*n-1) I (n+ 1);
end;
LNJ=LN ;

%###################################
%###################################
%

LOOP

%###################################

%###################################

for N_ INDEX=l:(TIMESTEPS-1)

N INDEX

%###################################

%STEP 1

%###################################

Fj=4*i*PSI(N _ INDEX, :)IDELTA;

BIBLIOGRAPHY
Fj=Fj*2*MASS;

%###################################
%STEP 2

%###################################

%From the left

% Calc Ej Fixed

t1=Aj(1)+sqrt(Aj(1) A2-1);

t2=Aj( 1)-sqrt(Aj(1) A2-1 );

if(abs(t1)>1) ALPHAj(1)=tl; else ALPHAj(1)=t2; end;

Ej(1)=ALPHAj(1);

%Recurrence

for j=2:SIZE

Ej(j)=2* Aj(j)-1/Ej(j-1 );
end
% Calc Qj Variable
j=1;
%convolution
SUM=O;
for k=1:N INDEX
n=N_INDEX-k+1;
SUM=SUM+PSI(k,j)*LN1(n);
end;
j=1;
Qj(j)=conj(Dj(j) )*( conj(Aj(j) )-ALPHAj(j) )*PSI(N _INDEX,j)+Dj(j)*(Aj(j)
ALPHAj(j) )*SUM;
%Recurrence
for j=2:SIZE
Qj(j)=Qj(j-1) /Ej(j-1 )+H A2*Fj(j) /Dj(j);
end

%###################################
%STEP 3

%###################################

%From the right

tl=Aj(J)+sqrt(Aj(J) A2-1 );

t2=Aj( J)-sqrt(Aj( J) A2-1 );

if(abs(t1)>1) ALPHAj{J)=tl; else ALPHAj{J)=t2; end;

j=J;

%convolution

SUM=O;

for k=1:N INDEX

n=N_INDEX-k+1;

SUM=SUM+PSI(k,j)*LNJ(n);

end;

j=J;

125

126

BIBLIOGRAPHY
BETAj(j)=conj(Dj(j) )*( conj( Aj(j) )-ALPHAj(j) )*PSI(N _IN DEX,j)+Dj(j)*( Aj(j)

ALPHAj(j) )*SUM;

Wj( J)=( Qj(J-1 )+BETAj( J)*Ej( J-1)) I (1-ALPHAj( J)*Ej( J-1));

%Recurrence for Wj

for j=(J):-1 :2

Wj(j-1 )=(Wj(j)-Qj(j-1)) /Ej(j-1);

end

%##############################

%STEP 4

%##############################

% Calculate Psi from Wj

PSI(N _INDEX+1, :)=PSI(N _INDEX, :).*(i*H 2./(3*DELTA *Dj)-1 )+Wj./Dj;

end;

figure

image( abs(PS1)*100);

ylabei('Timestep (.1 fs)');xlabei('X From -40 to 40 (nm)');

Bohmian Code
Newtonian
%SIZE=512
%cells
XL=-40;XR=40;
nump=40;
porder=2;
maxgauss=1;
TOLERANCE=1e-5*nump;
SIZE=10000; %for psi,etc
TIMESTEPS=2000 %
DELTA=1;
DELTA=DELTA*fsec;
%timestep in femtoseconds
POTENTIAL='0*0.0061.*X.-4+50*cos(X)*0';
J=SIZE;
H=(XR-XL)/(SIZE-1);
%grid size
X=XL :H:XR;
temp=X*O;
%% SYSTEM MASS AND POTENTIAL
Vj=temp; % Potential in eV
%Vj=Vj+1e3*exp(-1/2/ . 0125-2*(X- . 75).-2);
%X=X*Xmscale;
%Vj=Vj+(X>-10)&(X<-5)+(X>5)&(X<10);

BIBLIOGRAPHY

127

%Vj=Vj*.25;
Vj=Vj+eval(POTENTIAL);
Vj=Vj*Energy_scale;
VV=Vj;
%Initial wavefunction
PsiO=temp;
wave_sigma=sqrt(10);
X0=20;
K=0.31;
%PsiO=exp(-1/2/wave_sigma-2*(X-X0).-2).*exp(i*K*X);
Psi0=exp(-1/2/wave_sigma-2*(X-X0).-2).*exp(i*K*X);
Psi1=exp(-1/2/wave_sigma-2*(X+X0).-2).*exp(i*K*X);
PsiO=Psi0+1/2*Psi1;
%PsiO=eval(PSIO);
%Normalize
%PsiO=Psi0*1/sqrt(normaliz(abs(PsiO.*PsiO),X));
% SYSTEM DEFINITION ENDS
PSI=zeros(TIMESTEPS,SIZE);
PSI(1, :)=PsiO;
rho=abs(Psi0).-2;
S=angle(PsiO);
%S=unwrap(S,[] ,2);
S=unwrap(S, [],1);
vel=diff(S);
%pO=log(abs(PsiO));
pO=abs(Psi0).-2;
cpO=cumtrapz(pO);
cpO=cpO/cpO(SIZE);
clear XX;
XX=ones(nump,TIMESTEPS)*NaN;
QQ=ones(nump,TIMESTEPS)*NaN;
PP=XX;
for i=1:nump
i;

%XX(i,1)=(X(siz)-X(1))/(nump+1)*i; %trial points uniformly distributed in spa

% can also be density dependent on PsiO
% XX(i,1)=-35+(i-1)*30/nump;
% tmpr=O;
% while(!)
% tmpi=ceil(rand(1)*siz);
%tmpr=rand(1);
%if(tmpr<pO(tmpi))XX(i,1)=X(tmpi);break;end;
%end;
prob1=1/(nump+1)*i;

BIBLIOGRAPHY

128

XX(i,1)=interpolate(X,cp0,prob1);
tmpi=getindex(cp0,1/(nump+1)*i);
XX(i,1)=X(tmpi); %dont use X
%
% use this instead

PD(i,1)=1/(nump+1)*i;

end;

QQ=XX;

PP=O*XX;

FF=XX*O;

%QF=XX*O;

%%%%%% LOOP

dt=DELTA;

VG=diff(VV);tmp=VG;VG=[tmp VG(SIZE-1)];

go_back1step=O;

tstep=2;

global FF;

while(tstep<=TIMESTEPS)

% this routine will try to get an analytical expression for rho;

[coefs resiO resi1 resi2 resi3]=analyse_rho21(QQ,PD,tstep-1,porder,maxgauss);
%
resi3=0*resi3;
%
resi2=0*resi2;
%
resi1=0*resi1;
myresidues=[resi0;resi1;resi2;resi3];
if(go_back1step)if(tstep>2)tstep=tstep-1;go__ back1step=O;end;end; % not run h
tstep
dt
notvalid=1;

error=O;

while notvalid
%notvalid
%
go_back1step
%
tstep
if(go_back1step)if(tstep>2)tstep=tstep-1;go_back1step=O;end;end;
%
tstep
%error
%dt+9
error=O;
notvalid=O;
for i=1:nump
qO=QQ(i,tstep-1);
pO=PP(i,tstep-1);
[qf, pf, ff] = symplectic21(qO,pO,dt,X,QQ,PP,VG,i,nump,tstep,PD,coefs, ...
porder,myresidues);

BIBLIOGRAPHY

129

%[qf_2, pf_2, ff_2] = symplectic(qO,p0,2*dt,X,QQ,PP,VG,i,nump,tstep);

[qh1, phi, fh1] = symplectic21(qO,pO,dt/2,X,QQ,PP,VG,i,nump,tstep,PD, ...

coefs,porder,myresidues);
[qh2, ph2, fh2] = symplectic21(qh1,ph1,dt/2,X,QQ,PP,VG,i,nump,tstep,PD, ...
coefs,porder,myresidues);
if(-isfinite(qh2)) fsdadsaerrrrrrr;end;

%
error=abs(ph2-pf); %momentum

error=abs(qh2-qf); %space

%error_lagre=abs(qf_2-qf); %space

%MATLAB BUG ;break; end;

if(error>TOLERANCE)

not_valid=1;dt=dt/2;

go_back1step=1;

end;

% Check for trajectory crossing

if((i>1)&(i<nump))

if(qh2<QQ(i-1,tstep-1))1(qh2>QQ(i+1,tstep-1))

%% bad HACK HACK

qh2=1/2*(QQ(i-1,tstep-1)+QQ(i+1,tstep-1));

ph2=1/2*(PP(i-1,tstep-1)+PP(i+1,tstep-1));

%
cxzcxz

if(tstep>3)tstep=tstep-3;end

not_valid=1;dt=dt/2;

go_back1step=1;

break cant use break because of a matlab bug

%
%go back 1 step;

end;end;

%if(error_large>TOLERANCE) not_valid=1;dt=dt*1.5;break; end;

QQ(i,tstep)=qh2;
FF(i,tstep)=fh2;
PP(i,tstep)=ph2;
TT(i,tstep)=dt;
end;
%Test for crossings.if crossed try again with timestep=1/2 current
%Test for accuracy.try with 2x timestep .. if accuracy below threshold timestep=
if(error<TOLERANCE)&(-go_back1step)not_valid=O;end;

end;

%we got valid

%endkiio

dt=dt* 1 . 002;

tstep=tstep+1;

130

BIBLIOGRAPHY
end;

subplot(3,1,1);

pp=QQ+j*PP;

%plot(pp(:, :) ', 'x-')

plot(QQ',PP');xlabel('space');ylabel('momentum');

subplot(3,1,2);

% plot(PP','x-')
plot(QQ');xlabel('time');ylabel('space');
subplot(3,1,3);
plot(TT(1, : ));xlabel('step');ylabel('time delta');
z=cumsum(TT');
figure
plot(z,QQ');;ylabel('space');xlabel('linear time ');
grid
dt*TIMESTEPS
( sum(QQ(: ,TIMESTEPS)-=sort(QQ(:,TIMESTEPS))))/nump
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function
[coefs resiO resi1 resi2 resi3]=analyse_rho22(QQ,PD,time,porder, ...
maxgauss,fourier_residue)
% this routine will try to get an analytical expression for rho;
%Get quadratic coeficients ...
zx=QQ(:,time);zy=PD;
rho=gradient(zy,zx);
CC=1/2*log(rho);
[siza sizb]=size(rho);

siz=max(siza,sizb);

%[maxi,mini,num_max,num_min]=findmaxmin(CC,siz,10,9);

maxi=siz;mini(1)=1;mini(2)=siz;num_max=1;num_min=2;

ngauss=num_min-1;

if(ngauss>maxgauss)ngauss=maxgauss;end;

%%% ONLY WORKS FOR 1 GAUSSIAN NOW FIXME

coef=zeros(ngauss,porder+1);
%generate coefs
for n=1 : (num_min-1)
istart=mini(n);

iend=mini(n+1);

%%BUG remove this hack

istart=istart+1;

iend=iend-1;

%%BUG remove this hack

BIBLIOGRAPHY

131

tempy=CC(istart:iend);
tempx=QQ(istart:iend,time);
coef(n,:)=polyfit(tempx,tempy,porder);
%SANITY CHECK
if(coef(n,1)>0)coef(n,:)=coef(n,:)*O;end;
residy=tempy-polyval(coef(n,:),tempx);
fresidy=dst(residy);
end;
specsize=max(max(size(residy)));
%resO=idst(residy); will be a check later
resO=zeros(1,specsize);
res1=zeros(1,specsize);
res2=zeros(1,specsize);
res3=zeros(1,specsize);
%should calculate this with idst ... but like this it is a good check
N=specsize;
for n=1:N
for k=1:N

%% IF WE WANT TO LOWPASS THIS IS WHERE TO DO IT

if(k>N/10)fresidy(k)=O;end; % 20%

ss=sin(pi*k*n/(N+1));

cc=cos(pi*k*n/(N+1));

resO(n)=resO(n)+fresidy(k)*ss;
res1(n)=res1(n)+fresidy(k)*Cc*(pi*k/(N+1));
res2(n)=res2(n)+fresidy(k)*(-ss)*(pi*k/(N+1))*(pi*k/(N+1));;
res3(n)=res3(n)+fresidy(k)*(-cc)*(pi*k/(N+1))*(pi*k/(N+1))*(pi*k/(N+1));;
end;

end;

resO=res0*2/(N+1);

res1=res1*2/(N+1);

res2=res2*2/(N+1);

res3=res3*2/(N+1);

temperr=resO-residy';

err_in_idst=max(max(temperr));

%coef=polyfit(QQ(:,time),CC,porder);

%###############

%DEBUG GOODNESS OF FIT

figure(1);

hold off

subplot(2,1,1);

plot(QQ(: ,time) ,CC, 'x')

hold on

for n=1:ngauss

132

BIBLIOGRAPHY
istart=mini(n);

iend=mini(n+1);

%%BUG remove this hack

istart=istart+1;

iend=iend - 1;

xstart=QQ(istart,time);
xend=QQ(iend,time);
vect=xstart:(xend-xstart)/200:xend;
drawnow
subplot(2,1,1);
plot(vect,polyval(coef(n,:),vect));
hold off
subplot(2,1,2);
plot(tempx,resO,'x-')
drawnow;
end;
hold off
%###############

figure(2)
plot(QQ(: , 1 :time)');
coef

%z
coefs=coef;
resiO=resO;
resi1=res1;resi2=res2;resi3=res3;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [q2, p2, f1 ] = symplectic21(qO,pO,dt,X,QQ,PP,VG,i,nump,tstep,PD,coefs, . . .
porder,myresidues)
%
%

qO=QQ(i,tstep-1);
pO=PP(i,tstep-1);
%get force term from potential
%
curi=getindex(X,qO);
%if(curi>O)
fO=-VG(curi);end;
%f0= - 2*q0;
%porder
fO=extforces21(qO,pO,QQ,PP,i,nump,tstep,PD,coefs,po r der,myresidues);
fO=fO+O;

p1=p0+dt/2*f0;

q1=qO+dt/2*p1;

%get force term from potential

%curi=getindex(X,q1) ;
f1=-VG(curi);end;
%if(curi>O)

BIBLIOGRAPHY

133

%f1=-2*q1;
f1=extforces21(q1,p1,QQ,PP,i,nump,tstep,PD,coefs,porder,myresidues);
f1=f1+0; %other forces
p2=p1+dt/2*f1;
q2=q1+dt/2*p2;

[q2, p2, f1];

return

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function [retemp QF] =extforces21(qO,pO,QQ,PP,i,nump,tstep,PD,coefs, ...

porder,myresidues)
temp=O;

resi0=myresidues(1,:);

resi1=myresidues(2,:);

resi2=myresidues(3,:);

resi3=myresidues(4,:);

%for n=1:nump

% if n==i continue;end;

% d=((qO-QQ(n,tstep-1)));

% u=sign(d);

%temp=temp+100*1/d~2*(u);

%end;
%DEPRECATED ################################### DERIVATIVES
%
%
%
%

QVector=QQ(:,tstep-1);

list=get_nearest_neigh_wself(QVector,i,5);

PD_DIFFS= maple_deriv(QQ(list),PD(list)');

PO=PD_DIFFS(1);P1=PD_DIFFS(2); P2=PD_DIFFS(3);P3=PD_DIFFS(4);P4=PD_DIFFS(5)

% This is the Q term already diff to give force not potential

%
QF=-1/2*(P4*P1~2+P2~3-2*P2*P3*P1)/P1~3; %

%
temp=QF*0.001;

%DEPRECATED ###################################

% #### C=CO+C1*X +C2*X~2

+C3 x~3 ....

% CD1=C1+2*C2*X+3*C3*X~2+ ....

% CD2=2*C2+3*2*C3*X+ ....

%coefs

%
CO=coefs(3);
coef(49) (porder 48)

coef(48)

%
C1=coefs(2);
%
C2=coefs(1);
coef(47)

% CALC DERIVATIVES ANALYTICALLY

coef=O;

BIBLIOGRAPHY

134
X=qO;
CD1=0;
for n=1:(porder+1-1);

coef=coefs(n);

m=porder-n+1;

CD1=CD1+coef*(m)*x-cm-1);

end;
CD2=0;
for n=1 : (porder+1-2);

coef=coefs(n);

m=porder-n+1;

CD2=CD2+coef*(m)*(m-1)*x-cm-2);

end ;
CD3=0;
if(porder>2)
for n=1:(porder+1-3);

coef=coefs(n);

m=porder-n+1;

CD3=CD3+coef*(m)*(m-1)*(m-2)*X-(m-3);

end;
end;
% %%%%%%%%%% QF=-(2*cd1*cd2+cd3_
QF=(2*CD1*CD2+CD3);
%
temp=0.1*QF;

% ###############################################
[ngauss temp]=size(coefs);
rhoO=O;rho1=0;rho2=0;rho3=0;
% FOURIER RESIDUES

%APPROX TO LAST PARTICLE ... NOT PRECISELY CORRECT

resO=O;

res1=0;

res2=0;

res3=0;

if(i>1)&&(i<nump-1)

resO=resiO(i);

res1=resi1(i);

res2=resi2(i) ;

res3=resi3(i);

end;

% ONLY WORK WITH 1 GAUSSIAN

%for n=1 :ngauss

n=1;

%% RESOLVE ILL CONDITIONING

a2=coefs(n,1);

BIBLIOGRAPHY

135

a1=coefs(n,2);
aO=coefs(n,3);

%n

expo=exp(a2*q0-2+a1*qO+a0); %%BUG

rhoO=rhoO+expo*exp(resO);
rho1=rho1+expo*(2*a2*qO+a1+res1);
rho2=rho2+expo*(2*a2+res2+(2*a2*q0+a1+res1)-2);
rho3=rho3+expo*(res3+3*(2*a2+res2)*(2*a2*qO+a1+res1)+(2*a2*qO+a1+res1)-3);
end;

%rho0

%rhoO=rhoO+expo*resO;

%rho1=rho1+expo*res1;

%rho2=rho2+expo*res2;

%rho3=rho2+expo*res3;

temp=1/2*(rho1-3-2*rho1*rho2*rhoO+rho3*rho0-2)/rho0-3;

%checkc=(temp/2/alphan-2+miu)/q0;

% Harmonic osc

%get force term from potential

%curi=getindex(X,q1);

%f1=0;

%if(curi>O)
f1=-VG(curi);end;

%f1=-2*q1;

%
temp=0.1*QF;

%% THE -0.001*p0 friction term will stabilize the density function

%tempi=-.1*q0-0.005*p0;%-0.003*q0-3;;

xx=q0-5;

%testing diff force terms

%

% tempi=+0.001*sech(xx)*tanh(xx)*1-0*qO-O.OOOO*q0-4;

tempi=-sech(xx)*tanh(xx)*i-O*qO-O.OOOO*q0-4;

ternpi=-sech(xx)*tanh(xx)*O-i*qO-O.OOOO*q0-4;

retemp=temp+tempi;

QHJE
tic

POTENTIAL='O*t+l*x.-2'

BIBLIOGRAPHY

136
ONEHALF=1/2*(1);

nump=50;porder=6; %linear

TDLERANCE=1e-5*nump;

TIMESTEPS=10000 %

dt=10--2*1;

clear order points

global order points

order=2

points=order+1

%points=

% this last part is assuming X spacing of 1

clear XX;

XX=ones(nump,TIMESTEPS)*NaN;

QQ=ones(nump,TIMESTEPS)*NaN;

RH=XX;PP=XX;

VP=XX;

QQ=XX;
%PP=O*XX;
FF=XX*O;
%Q=XX*O;

DDSS=PP;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

X=-10*1:20*1/(nump-1) : 10*1;

% for i=1 :max(size(X)) y(i)=QPotential(1,X(i));end; plot(X,y,'x-')

clear XX;

XX=ones(nump,TIMESTEPS)*NaN;

QQ=ones(nump,TIMESTEPS)*NaN ;

RH=XX;PP=XX;

SS=XX;

VP=XX;

VP=VP*O;

for i=1:nump
QQ(i,1)=X(i);
RH(i,1)=log(rho(O,X(i)))/2;
SS(i,1)=Sphase(O,X(i));
PP(i,1)=vel(O,X(i));
VP(i,1)=Potential(O,X(i));
end
for i=1:TIMESTEPS
VP( : ,i)=VP(:,1); %time indep for now
end

BIBLIOGRAPHY
%initial phase fix
temp=SS(:,1);
temp=unwrap(temp,2);
SS(:,1)=temp;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%STARTHERE
tstep=1;
ctime=-dt; %not zero
while(tstep<=TIMESTEPS),
dt
notvalid=1;go_back1step=O;
while not valid,
if(go_back1step)
tstep=tstep-go_back1step;
tstep=max(tstep,1);
go_back1step=O;
end;
notvalid==O;
%{
%this is slow somehow
dt2==dt/2;
t==ctime;
for i==1:nump
x==QQ(i,tstep);
VP(i,tstep)=O;%eval(POTENTIAL); %time indep for now
end
plot(VP, '>-');
%%%%%%%%%%% 2 half steps

%}
%faster here
x=QQ(:, tstep);
%VP(:,tstep)=10*10~-3*(x-2).~2;
VP(:,tstep)==0*1*10~-4*(x-2).~2.*(x-0).~2;

%% POTENTIAL HERE
%VP(:,tstep)==10~1*sech(4*(x-3));
VP(:,tstep)=0*8*10~-2*(x).~2;

qqOO==QQ(:,tstep);

%{

if(tstep>1)

qp11=calcQP3(qqOO,RH(:,tstep),RH(:,tstep-1),QP(:,tstep-1));

else

qp11=calcQP(qqOO,RH(:,tstep));

end

[pp11 ddss11]==calcPPDDSS(qqOO,SS(:,tstep));%

137

138

BIBLIOGRAPHY
qq12=QQ( : ,tstep)+dt2*pp11;% only 1st order
%VP(:,tstep)=10*pp11.~2;%0*1*10~-4*(x-2).~2.*(x-0) . ~2;

%% POTENTIAL HERE
ss12=SS(:,tstep)+dt2*(0NEHALF*pp11.~2-(VP(:,tstep)+qp11));

[ptemp ddss12]=calcPPDDSS(qq12,ss12);

rh12=RH(: ,tstep)+dt2*(-1/2*ddss12);

%rh12=RH(: ,tstep)+dt2*(-1/2*ddss11);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

if (tstep>1)

qp12=calcQP3(qq12,rh12,RH(:,tstep),qp11);

else

qp12=calcQP(qq12,rh12);

end

[pp12 ddss12]=calcPPDDSS(qq12,ss12);%

qq22=qq12+dt2*pp12; % only 1st order

ss22=ss12+dt2*(0NEHALF*pp12.~2-(VP(:,tstep)+qp12));

[pp22 ddss22]=calcPPDDSS(qq22,ss22);
rh22=rh12+dt2*(-1/2*ddss12);
%rh22=rh12+dt2*(-1/2*ddss11);

%}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%{
if(tstep>1)
qp11=calcQP3(qqOO,RH(:,tstep),RH(:,tstep-1),QP(: ,tstep-1));
else
qp11=calcQP(qqOO,RH(:,tstep),qqOO,qqOO);
end
% qp11=threshold_smooth(qqOO,qp11,1);
%qp11=fixborders2(qqOO,qp11,3,0,1);
[pp11 ddss11]=calcPPDDSS(qqOO,SS( : ,tstep)); %% v=grad(S)
%pp11=fixborders2(qqOO,pp11,3,0,1);
%ddss11=fixborders2(qqOO,ddss11,3,0,1);
% ddss11=threshold_smooth(qqOO,ddss11,10);
%no need for this block use from prev one
qq1f=QQ( : ,tstep)+dt*pp11;% only 1st order
ss1f=SS(:, t step)-dt*(DNEHALF*pp11.~2+(VP(: ,tstep)+qp11));
[ptemp ddss1f]=calcPPDDSS(qq1f,ss1f);
%
ddss1 f =fixborders2(qqOO,ddss1f,11,0,1);
% ddss1f=threshold_smooth(qq1f,ddss1f,10);
rh1f=RH(:,tstep)+dt*(-1/2*ddss1f);
%rh1f=RH( : , tstep)+dt*(-1/2*ddss11);
%}
ITER_HIGHTHRESH=16;
ITER_LOWTHRESH=9;

BIBLIOGRAPHY
THRESHOLD=10~-13;

XO=QQ(: ,tstep);SO=SS(:,tstep);CO=RH(:,tstep);

X1=XO;S1=SO;C1=CO;

% [DSO DDSO]=calcPPDDSS(X1,S1);

% [DCO DDCO]=calcPPDDSS(X1,C1);

[DSO DDSO DCO DDCO]=calcPPDDSS_alternate_zones(X1,S1,C1);

% SO=SO+dt/2*(-(DS0.~2)+DDCO+DC0.~2);

% DDCO=DDCO*O;DCO=DCO*O;

% S1=SO+dt/2*(-(DS0.~2)+DDCO+DC0.~2); %jumpstart S1 for zones

% figure(3);plot(S1);

notvalid=1;

while notvalid,

iter=O;DDS=XO;DDC=XO;err=100;

while abs(err)>THRESHOLD,

iter=iter+1;

LASTDDS=DDS;

% [DS DDS]=calcPPDDSS_alternate(X1,S1);

% [DC DDC]=calcPPDDSS_alternate(X1,C1);

[OS DDS DC DDC]=calcPPDDSS_alternate_zones_varpoly(X1,S1,C1);

thres=10~-1o;

%
%
%
%

DDC=threshold_smooth2(X1,DDC,thres);
DDS=threshold_smooth2(X1,DDS,thres);
DC=threshold_smooth2(X1,DC,thres);
DS=threshold_smooth2(X1,DS,thres);
% DDC=DDC*O;DC=DC*O;
%
DDS=fixborders2(X1,DDS,2,0,1);
DS=fixborders2(X1,DS,2,0,1);
%
%
DDC=fixborders2(X1,DDC,2,0,1);
%
DC=fixborders2(X1,DC,2,0,1);
err=max(max(LASTDDS-DDS));
X1=XO+dt*DS;
S1=SO+dt/2*(-(DS.~2)+DDC+DC.~2);

C1=CO-dt/2*DDS;

if(iter>50) errorinfiniteloop; end;

% plot(XO,DDS);grid;drawnow

end

iter

if(iter>ITER_HIGHTHRESH)

dt=dt/2; notvalid=1;
else
notvalid=O;
end;
end;
%trapezoid semiimpl

139

140

BIBLIOGRAPHY
X1=XO+dt/2*(DS+DSO);

S1=SO+dt/4*(-(DS . -2)-(DS0.-2)+DDC+DDCO+DC . -2+DC0.-2);

C1=CO-dt/4*(DDS+DDSO);

% Xl=XO+dt*DS;

% S1=SO+dt/2*(-(DS.-2)+DDC+DC.-2);

% C1=CO-dt/2*DDS;

%%%%%%%%%%%%%%%%%

%compare qlf to qq22

%maxerror=max(abs(qq22-qq1f));

%if(maxerror>TOLERANCE)

% not_valid=l;dt=dt/2;

% go_back1step=1;

%end

qq1f=X1 ;

if(min(diff(qqlf))<=O) %there was a crossing

toe

fefrefre ;

end

end;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% use the 2 half step results anyways

% QP(:,tstep)=qp12;

% PP(:,tstep)=pp12;

% DDSS( : ,tstep)=ddss12;

% QQ(:,tstep+1)=qq22;

% SS( : ,tstep+1)=ss22;

% DDSS(: ,tstep+1)=ddss22;

% RH(: ,tstep+1)=rh22;

% Or force usage of single big step

QP( : ,tstep)=-1/2*(DDC+DC.-2);
PP (: , tstep) =DS;
DDSS( : ,tstep)=DDS;
QQ( : ,tstep+1)=X1;
SS(:,tstep+1)=S1;
DDSS(:,tstep+l)=DDS;
RH( :, tstep+1)=C1;
%pause
%order bug tstep+l
%plot(QQ(: , tstep),(RH(: ,tstep)),'x-');drawnow;
%plot(QQ( : ,tstep),exp(RH( : ,tstep)),'x-');drawnow;
%plot(QQ( : ,tstep),SS(: ,tstep),'x-');drawnow ;
%plot(QQ(: ,tstep),QP(:,tstep),'x-');drawnow;
figure (1);

BIBLIOGRAPHY

141

subplot(4,1,1);
plot(QQ(:,tstep),(RH(:,tstep)),'x-');
subplot(4,1,2);
plot(QQ(:,tstep),QP(:,tstep),'x-');
subplot(4,1,3);
plot(QQ(:,tstep),(SS(:,tstep)),'x-');
%p12
%plot(QQ(:,tstep),-1/2*2*ddss22,'x-');

[tempi temp2]=calcPPDDSS(QQ(:,tstep),QP(:,tstep));

plot(QQ(:,tstep),DS,'x-')

subplot(4,1,4);

plot(QQ(:,tstep),DDS,'x-')

%plot(QQ(:,tstep),temp2,'x-')

%if(tstep>2)plot(QQ(:,tstep),(QP(:,tstep)-QP(:,tstep-1))./dt,'x-');end

drawnow;

if pauses pause; end;

%plot(QQ(:,tstep),gradient(SS(:,tstep)),'x-');drawnow;

%pause(2)

deltat(tstep)=dt;

ctime=ctime+dt;

time(tstep)=ctime;

dt=dt;

ctime

tstep=tstep+l

if(iter<ITER_LOWTHRESH) dt=dt*1.01; end;

end

toe

%}%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function [OS DDS DC DDC]=calcPPDDSS_alternate_zones(X,S,C)

% needs accurate initial S to work!

siz=max(size(X));

my_S=spapi(4,X,S);

spl=fnder(my_S,1);t1=fnval(sp1,X);

sp2=fnder(my_S,2);t2=fnval(sp2,X);

sp3=fnder(my_S,3);t3=fnval(sp3,X);% may have to bump up order of splines

DSO=tl; DDSO=t2; DDDSO=t3;

my_C=spapi(4,X,C);

sp1=fnder(my_C,1);t1=fnval(sp1,X);

sp2=fnder(my_C,2);t2=fnval(sp2,X);

sp3=fnder(my_C,3);t3=fnval(sp3,X);% may have to bump up order

zones=zeros(l,siz);

zones=(DDC0>10--8);

%enlarge by 1

142

BIBLIOGRAPHY
enl=4;
zones=smooth(zones,enl*2+1);%zones=zones>O;
[numz breaks]=getnumtypezones(zones);
% figure(2);plot(zones,'-x');drawnow;
starts=!;
for i=1:numz
% 0 for normal transition zones
% -1 for always splines
% 3 for always least sq
if(zones(starts)>O)
%splines or local
ends=breaks(i);
DC(starts:ends)=DCO(starts:ends);
DDC(starts:ends)=DDCO(starts:ends);
DS(starts:ends)=DSO(starts :ends);
DDS(starts:ends)=DDSO(starts : ends);
starts=ends+1;
else
ends=breaks(i);
%least sq or global
gx=X(starts:ends);gc=C(starts:ends);gs=S(starts:ends);
my_C=spap2(1,3,gx,gc);my_S=spap2(1,3,gx,gs);
sp1=fnder(my_C,1);t1=fnval(sp1,gx);
sp2=fnder(my_C,2);t2=fnval(sp2,gx);
DC(starts : ends)=t1;
DDC(starts:ends)=t2;
sp1=fnder(my_S,1);t1=fnval(sp1,gx);
sp2=fnder(my_S,2);t2=fnval(sp2,gx);
DS(starts:ends)=t1 ;
DDS(starts:ends)=t2;
starts=ends+1;
end ;
end;
DC=DC';DDC=DDC';DDS=DDS';DS=DS';
return;
% sp=spapi(4,x,yerr);
% ch = spapi(augknt(x,4,2), [x x], [yerr dy]);
% ls=spap2(18,3,x,yerr); %1 quadratic pieces=x-k+1;
% my_sp=spapi(4,X,S);% 3rd order spline (cubic)
my_S=spap2(1,3,X,S);
my_C=spap2(1,3,X,C);
%my_sp=csapi(X,S) ; %reg spline
% my_sp=csaps(X,S); %smoothed
sp1=fnder(my_S,1);t1=fnval(sp1,X);

BIBLIOGRAPHY
sp2=fnder(my_S,2);t2=fnval(sp2,X);
DS=t1;
DDS=t2;
sp1=fnder(my_C,1);t1=fnval(sp1,X);
sp2=fnder(my_C,2);t2=fnval(sp2,X);
DC=t1;
DDC=t2;
% t1=fixborders(X,t1,2,2); %linear
% t2=fixborders(X,t2,2,1);% quadratic
return

% %
% %
siz=max(size(X));
order=2;
points=order+1;
interpx=zeros(1,points);interpy=interpx;
t1=interpx;t2=interpx;
for i=1:siz
originx=X(i);
indexes=get_n_neighbours(points,X,originx);
interpx=X(indexes);interpy=S(indexes);
sp=polyfit(interpx,interpy,order);
sp1=polyder(sp);sp2=polyder(sp1);
tt1=polyval(sp1,originx);tt2=polyval(sp2,originx);
% t1=tt1;t2=tt2;
%
PP(i)=t1;
%
DDSS(i)=t2;
t1(i)=tt1;

t2(i)=tt2;

end
t1=fixborders(X,t1,1,2)'; %
t2=fixborders(X,t2,1,2)';
PP=t1';
DDSS=t2';
return;
%my_sp_extended=my_sp;
%fnxtr(my_sp,2); %linear outside domain
%CC(1)=fnval(my_sp_extended,zx(1));
%CC(nump)=fnval(my_sp_extended,zx(nump));
%siz=max(size(S));
%pvalues=0.99; %optimal value
%weights=[! 1 ones(l,siz-4) 1 1];
%my_sp=csaps(X,S,pvalues,[],weights);
i. smoothing aps
%my_sp=spapi(4,X,S); % 3rd order spline (cubic)

143

144

BIBLIOGRAPHY
siz=max(size(S));
pvalues=0 . 99; %optimal value
weights=[1 1 1*ones(1,siz-4) 1 1];
%my_sp=csaps(X,S,pvalues,[],weights);
my_sp=csapi(X,S);% no blending
sp1=fnder(my_sp,1);
sp2=fnder(my_sp,2);
cub=csapi(X,S);
cub1=fnder(cub,1);
cub2=fnder(cub,2);
dasdsadsa
blend=[O 0 0 0 0 ones(1,siz-10) 0 0 0 0 0]';
%must be between 0-->cubic and 1--->smoothed
PP=fnval(sp1,X).*(blend)+fnval(cub1,X) . *(1-blend);
DDSS=fnval(sp2,X).*(blend)+fnval(cub2,X).*(1-blend);
%must blend in csplines with smoothed ones at borders
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%{
siz=max(size(X));
interpx=zeros(1,points);interpy=interpx;
for i=1:siz
originx=X(i);
indexes=get_n_neighbours(points,X,originx);
interpx=X(indexes);interpy=S(indexes);
sp=polyfit(interpx,interpy,order);
sp1=polyder(sp);sp2=polyder(sp1);
t1=polyval(sp1,originx);t2=polyval(sp2,originx);
PP(i)=t1;
DDSS(i)=t2;
end

%}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function Q=calcQP2_splines(X,RH,prevRH,prevQ)
delt=RH-prevRH;
%rescale
pvalues=0.99-(delt>0)*0.5;
pvalues=0.99;
%siz=max(size(X));
Q=O;
%my_sp=spap2(1,3,X,RH);% least sq quadratic
%my_sp=spapi(4,X,RH);% 3rd order spline (cubic) %continuous 2nd
%my_sp=csapi(X,RH);% cubic spline
%my_sp=spaps(X,RH,O);% smoothing aps

BIBLIOGRAPHY

145

siz=max(size(RH));

%pvalues=0.99; %optimal value

pvalues=0.99; %optimal value

weights=[! 1 ones(1,siz-4) 1 1];

my_sp=csaps(X,RH,pvalues',[],weights);
%smoothing aps
%my_sp=spap2(1,3,X,RH); %leastsq quadratic->correct behaviour for single gaussians
my_sp=csapi(X,RH);% cubic spline
%my_sp=spaps(X,RH,10--3);
my_sp_extended=my_sp;

%fnxtr(my_sp,3); %quadratic outside domain

%CC(1)=fnval(my_sp_extended,zx(1));
%CC(nump)=fnval(my_sp_extended,zx(nump));

sp1=fnder(my_sp_extended,1);

sp2=fnder(my_sp_extended,2);

t1=fnval(sp1,X);t2=fnval(sp2,X);

Q=-1/2*(t2+t1.-2);

%Q=Q;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%{

siz=max(size(X));

Q=O;

interpx=zeros(1,points);interpy=interpx;

for i=1:siz

%order=6;

%if(i<floor(points/2))l(i>(siz-floor(points/2)))

%order=2;

%end;

originx=X(i);

originy=RH(i);

% get n closest points

indexes=get_n_neighbours(points,X,originx);

interpx=X(indexes);interpy=RH(indexes);

sp=polyfit(interpx,interpy,order);

sp1=polyder(sp);sp2=polyder(sp1);

t1=polyval(sp1,originx);t2=polyval(sp2,originx);

tempQ=-1/2*(t2+t1.-2);

Q(i)=tempQ;

Q=Q';

end

%}

function indexes=get_n_neighbours(n,X,originx)

% gets indexes to nearest n neighbours .. n should be less than sizeX

[temps tempi]=sort(abs(X-originx));

146

BIBLIOGRAPHY
result=tempi(1:n);

indexes=sort(result)';

% return

siz=max(size(X));

me=tempi (1);

idx=1:n;

idx=me+idx-ceil(n/2);

adjl=min(idx(1)-1,0);

adjr=max(idx(n)-siz,O);

index=idx-adjl-adjr;

%siz=max(size(X));

%if(indexes(1)>1)&(indexes(n)<siz)

%indexes=(result(1):result(1)+n-1)-floor(n/2);

%end;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

4146

32