THE THEORY OF OPEN QUANTUM SYSTEMS

THE THEORY OF OPEN

QUANTUM SYSTEMS

Heinz-Peter Breuer and Francesco Petruccione

., Albert-Ludwigs-Universitdt Freiburg, Fakultdt fiir Physik
and
Istituto Italiano per gli Studi Filosofici

OXFORD
UNIVERSITY PRESS
 OXFORD
UNIVERSITY PRESS
Great Clarendon Street, Oxford 0 x2 6Dp
Oxford University Press is a department of the University of Oxford.
It furthers the University's objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Bangkok Buenos Aires Cape Town Chennai
Dar es Salaam Delhi Hong Kong Istanbul Karachi Kolkata
Kuala Lumpur Madrid Melbourne Mexico City Mumbai Nairobi
Sao Paulo Shanghai Taipei Tokyo Toronto
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
0 Oxford University Press 2002
The moral rights of the authors have been asserted
Database right Oxford University Press (maker)
First published 2002
Reprinted 2002,2003
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose this same condition on any acquirer
A catalogue record for this book is available from the British Library
Library of Congress Cataloging in Publication Data
(Data available)
ISBN 0 19 852063 8
3 5 7 9 10 8 6 4
Printed in Great Britain
on acid-free paper by
T.J. International Ltd., Padstow
 A Gerardo Marotta
Presidente dell' Istituto Italiano per gli Studi Filosofici
per avere promosso e sostenuto questa ricerca

Fiir Heike und Valeria

fiir unendliche Geduld, Verstdndnis und Hilfe
 PREFACE

Quantum mechanics is at the core of our present understanding of the laws of

physics. It is the most fundamental physical theory, and it is inherently prob-
abilistic. This means that all predictions derived from quantum mechanics are
of a probabilistic character and that there is, as far as we know, no underlying
deterministic theory from which the quantum probabilities could be deduced.
The statistical interpretation of quantum mechanics ultimately implies that pre-
dictions are being made about the behaviour of ensembles, i.e. about collections
of a large number of independent, individual systems, and that the statements
of quantum theory are tested by carrying out measurements on large samples of
such systems.
Quantum mechanical-systems must be regarded as open systems. On the one
hand, this is due to the fact that, like in classical physics, any realistic system
is subjected to a coupling to an uncontrollable environment which influences
it in a non-negligible way. The theory of open quantum systems thus plays a
major rôle in many applications of quantum physics since perfect isolation of
quantum systems is not possible and since a complete microscopic description or
control of the environmental degrees of freedom is not feasible or only partially so.
Most interesting systems are much too complicated to be describable in practice
by the underlying microscopic laws of physics. One can say even more: Not
only is such a microscopic approach impossible in practice, it also does not
provide what one really wants to know about the problem. Even if a solution of
the microscopic evolution equations were possible, it would give an intractable
amount of information, the overwhelming part of which is useless for a reasonable
description.
Practical considerations therefore force one to seek for a simpler, effectively
probabilistic description in terms of an open system's dynamics. The use of
probability theory allows the treatment of complex systems which involve a huge
or even an infinite number of degrees of freedom. This is achieved by restricting
the mathematical formulation to an appropriate set of a small number of relevant
variables. Experience shows that under quite general physical conditions the time
evolution of the relevant variables is governed by simple dynamical laws which
can be formulated in terms of a set of effective equations of motion. The latter
take into account the coupling to the remaining, irrelevant degrees of freedom in
an approximate way through the inclusion of dissipative and stochastic terms.
There is another reason for invoking the notion of an open system in quan-
tum theory which is of more fundamental origin. Quantum theory introduces
a deterministic law, the Schrödinger equation, which governs the dynamics of
the probability distributions. This equation describes the evolution of chance,
viii PREFACE

that is the dynamics of ensembles of isolated systems. However, as a probabilis-

tic theory quantum mechanics must also encompass the random occurrence of
definite events which are the realizations of the underlying probability distri-
butions. In order to effect the occurrence of chance events a quantum system
must be subjected to interactions with its surroundings. Any empirical test of
the statistical predictions on a quantum system requires one to couple it to a
measuring apparatus which generally leads to non-negligible influences on the
quantum object being measured. Thus, quantum mechanics in itself involves an
intimate relationship to the notion of an open system through the action of the
measurement process.
This book treats the central physical concepts and mathematical techniques
used to study the dynamics of open quantum systems. The general approach
followed in the book is to derive the open system's dynamics either from an
underlying microscopic theory by the elimination of the environmental degrees
of freedom, or else through the formulation of specific measurement schemes
in terms of a quantum operation. There is a close physical and mathematical
connection between the evolution of an open system, the state changes induced
by quantum measurements, and the classical notion of a stochastic process. The
book provides a detailed account of these interrelations and discusses a series of
physical examples to illustrate the mathematical structure of the theory.
To provide a self-contained presentation Part I contains a survey of the classi-
cal theory of probability and stochastic processes (Chapter 1), and an introduc-
tion to the foundations of quantum mechanics (Chapter 2). In addition to the
standard concepts, such as probability space, random variables and the definition
of stochastic processes, Chapter 1 treats two topics, which are important for the
further development of the theory. These are piecewise deterministic processes
and Levy processes. In Chapter 2 the emphasis lies on the statistical interpre-
tation of quantum mechanics and its relationship to classical probability theory.
As a preparation for later chapters, we also discuss composite quantum systems,
the notion of entangled states, and quantum entropies. A detailed account of the
quantum theory of measurement within the framework of quantum operations
and effects is also included.
The fundamentals of the description of the quantum dynamics of open sys-
tems in terms of quantum master equations are introduced in Part II, together
with its most important applications. In Chapter 3, special emphasis is laid
on the theory of quantum dynamical semigroups which leads to the concept of
a quantum Markov process. The relaxation to equilibrium and the multi-time
structure of quantum Markov processes are discussed, as well as their irreversible
nature which is characterized with the help of an appropriate entropy functional.
Microscopic derivations for various quantum master equations are presented,
such as the quantum optical master equation and the master equation for quan-
tum Brownian motion. The influence functional technique is investigated in the
context of the Caldeira—Leggett model. As a further application, we derive the
master equation which describes the continuous monitoring of a quantum ob-
 PREFACE ix

ject and study its relation to the quantum Zeno effect. Chapter 3 also contains
a treatment of non-linear, mean field quantum master equations together with
some applications to laser theory and super-radiance.
In Chapter 4 we study the important field of environment-induced decoher-
ence and the transition to the classical behaviour of open quantum systems. A
number of techniques for the determination of decoherence times is developed.
As specific examples we discuss experiments on the decoherence of Schrödinger
cat-type states of the electromagnetic field, the destruction of quantum coher-
ence in the Caldeira—Leggett model, and the environment-induced selection of a
pointer basis in the quantum theory of measurement.
While Parts I and II mainly deal with the standard aspects of the theory,
Parts III—V provide an overview of more advanced techniques and of new devel-
opments in the field of open quantum systems. Part III introduces the notion
of an ensemble of ensembles and the concept of stochastic wave functions and
stochastic density matrices. The underlying mathematical structure of probabil-
ity distributions on Hilbert or Liouville space and of the corresponding random
state vectors is introduced in Chapter 5. These concepts are used to describe the
dynamics of continuous measurements performed on an open system in Chap-
ter 6. It is shown that the time -evolution of the state vector, conditioned on
the measurement record, is given by a piecewise deterministic process involv-
ing continuous evolution periods broken by randomly occurring, sudden quan-
tum jumps. This so-called unravelling of the quantum master equation in the
form of a stochastic process is based on a close relation between quantum dy-
namical semigroups and piecewise deterministic processes. The general theory
is illustrated by means of a number of examples, such as direct, homodyne and
heterodyne photodetection.
The general formulation in terms of a quantum operation is given in Chapter
8, where we also investigate further examples from atomic physics and quantum
optics, e.g. dark state resonances and laser cooling of atoms. In particular, the
example of the sub-recoil cooling dynamics of atoms nicely illustrates the inter-
play between incoherent processes and quantum interference effects which leads
to the emergence of long-tail Lévy distributions for the atomic waiting time.
The numerical simulation of stochastic processes on high-performance com-
puters provides an efficient tool for the predictions of the dynamical behaviour
in physical processes. The formulation of the open system's dynamics in terms of
piecewise deterministic processes or stochastic differential equations in Hilbert
space leads to efficient numerical simulation techniques which are introduced and
examined in detail in Chapter 7.
Part IV is devoted to the basic features of the more involved non-Markovian
quantum behaviour of open systems. Chapter 9 gives a general survey of the
Nakajima—Zwanzig projection operator methods with the help of which one de-
rives so-called generalized master equations for the reduced system dynamics. In
the non-Markovian regime, these master equations involve a retarded memory
kernel, that is a time-convolution integral taken over the history of the reduced
 PREFACE

o ee,
.e.,
0.4 cs
0. 4-.'•
,,..-
a.qtS`.# \-1".4*. 4s.• ,bc.s
..- 0,sç'. sç'iz*, sk‘ c., es...,
Z,'' '\'6 -. 1'".cs,<,-'sç.. , -,),q> ef,
eP 0-4, 0- oc' co•tN-
(s..\‘ 'sfs• e" ....o• -gç
. Chapter 1 •
tct
P. Chapter 2 • • Ilio
3
Chapter • • • • •
C:3• Chapter 4 • • • e
Chapter 5 • •
Chapter 6 • • •
Chapter 7 •
Chapter 8 • • •
> Chapter 9 • • •
t
4,1 Chapter 10 • •
> Chapter 11 • • •
P. Chapter 12 • • •
FIG. 0.1. Possible pathways through the book for a first reading. The heavy dots
indicate those chapters in which the basic information on central concepts of
the book my be found.

system. In general, one is thus confronted with the difficult task of treating an
integro-differential equation for the open system's density matrix. We therefore
develop in Chapter 9 a method of particular relevance based on an equation of
motion which is local in time and which is known as the time-convolutionless pro-
jection operator method. This method serves as a starting point for a systematic
perturbation expansion around the Markovian limit and for a numerical treat-
ment. A number of applications to non-Markovian dynamics in physical systems
is investigated in Chapter 10, such as the Jaynes—Cummings model, quantum
Brownian motion and the spin-boson model.
The final Part V is concerned with the relativistic formulation of the dy-
namics of open quantum systems and of quantum measurement theory. Chapter
11 deals with the relativistic formulation of the state reduction postulate in
quantum measurement theory. This postulate is used to study a number of ap-
plications to local and non-local measurements and to the restrictions on the
measurability of non-local quantities imposed by the causality principle. The
relativistic formulation allows us to discuss several important experiments from
 PREFACE xi

a unified perspective, such as EPR-type experiments, measurements of Bell state

operators, exchange measurements, and quantum teleportation. The relativistic
density matrix theory of quantum electrodynamics is developed in Chapter 12
using functional methods from field theory, path integral methods, and the in-
fluence functional formulation. As an important example a detailed theory of
decoherence in quantum electrodynamics is presented.
The cross-disciplinary nature of the field of open quantum systems necessarily
requires the treatment of various different aspects of quantum theory and of
diverse applications in many fields of physics. We sketch in Fig. 0.1 various
possible pathways through the book which may be followed in a first reading.
The book addresses undergraduate and graduate students in physics, theo-
retical physics and applied mathematics. As prior knowledges only a basic under-
standing of quantum mechanics and of the underlying mathematics, as well as
an elementary knowledge of probability theory is assumed. The chapters of the
book are largely written as self-contained texts and can be employed by lecturers
as independent material for lectures and special courses. Each chapter ends with
a short bibliography. Since most chapters deal with a rapidly evolving field it
was sometimes impossible to give a complete account of the literature. Instead
we have tried to give some important examples of original papers and introduc-
tory review articles that cover the subjects treated; personal preferences strongly
influenced the lists of references and we apologize to those authors whose work
is not cited properly. A number of excellent general textbooks and monographs
from which we learned a great deal is listed in the following bibliography.

Bibliography
Alicki, R. and Lendi, K. (1987). Quantum Dynamical Semigroups and Applica-
tions, Volume 286 of Lecture Notes in Physics. Springer-Verlag, Berlin.
Braginsky, V. B. and Khali li, F. Ya. (1992). Quantum Measurement. Cambridge
University Press, Cambridge.
Carmichael, H. (1993). An Open Systems Approach to Quantum Optics, Volume
m18 of Lecture Notes in Physics. Springer-Verlag, Berlin.
Cohen-Tannoudji, C., Dupont-Roc, J. and Grynberg, G. (1998). Atom Photon -

Interactions. John Wiley, New York.

Davies, E. B. (1976). Quantum Theory of Open Systems. Academic Press,
London.
Gardiner, C. W. and Zoller, P. (2000). Quantum Noise (second edition).
Springer-Verlag, Berlin.
Giulini, D., Joos, E., Kiefer, C., Kupsch, J., Stamatescu, I.-0. and Zeh, H. D.
(1996). Decoherence and the Appearence of a Classical World in Quantum
Theory. Springer Verlag, Berlin.
-

Louise 11, W. (1990). Quantum Statistical Properties of Radiation. John Wiley,

New York.
Mandel, L. and Wolf, E. (1995). Optical Coherence and Quantum Optics. Cam-
bridge University Press, Cambridge.
xii PREFACE

Nielsen, M. A. and Chuang, L L. (2000). Quantum Computation and Quantum

Information. Cambridge University Press, Cambridge.
Scully, M. O. and Zubairy, M. S. (1997). Quantum Optics. Cambridge Univer-
sity Press, Cambridge.
Walls, D. F. and Milburn, G. J. (1994). Quantum Optics. Springer-Verlag,
Berlin.
Weiss, U. (1999). Quantum Dissipative Systems, Volume 2 of Series in Modern
Condensed Matter Physics. World Scientific, Singapore.
 ACKNOWLEDGEMENTS

It is a pleasure to thank the Istituto Italiano per gli Studi Filosofici for promoting
the research programme which led to this book. The Istituto supported a series
of fruitful workshops on the theory of open quantum systems. In particular we
owe thanks to its President Avv. Gerardo Marotta, who gave continual help and
encouragement throughout the research and the preparation of the manuscript.
We have received a great deal of help from friends and colleagues. Our thanks
to Robert Alicki, Francois Bardou, Thomas Filk, Domenico Giulini, Gerard Mil-
burn and Ludger Riischendorf for critically reading parts of the manuscript and
for making several suggestions for improvements. We have profited much from
lively discussions with them.
We are also indebted to our students Peter Biechele, Kim Bostriim, Uwe
Dorner, Jens Eisert, Daniel Faller, Wolfgang Huber, Bernd Kappler, Andrea
Ma, Wolfgang Pfersich and Frithjof Weber. They developed with us parts of the
material presented in the book and contributed with several ideas.
The excellent cooperation with the staff of Oxford University Press deserves
special thanks. In particular, we express our gratitude to Siinke Adlung for the
professional guidance through all stages of the preparation of the work.
 CONTENTS

I PROBABILITY IN CLASSICAL AND QUANTUM PHYSICS

1 Classical probability theory and stochastic processes 3

1.1 The probability space 3
1.1.1 The a-algebra of events 3
1.1.2 Probability measures and Kolmogorov axioms 4
1.1.3 Conditional probabilities and independence 5
1.2 Random variables 5
1.2.1 Definition of random variables 6
1.2.2 Transformation of random variables 8
1.2.3 Expectation values and characteristic function 9
1.3 Stochastic processes 11
1.3.1 Formal definition of a stochastic process 11
1.3.2 The hierarchy of joint probability distributions 12
1.4 Markov processes 14
1.4.1 The Chapman—Kolmogorov equation 14
1.4.2 Differential Chapman—Kolmogorov equation 17
1.4.3 Deterministic processes and Liouville equation 19
1.4.4 Jump processes and the master equation 21
1.4.5 Diffusion processes and Fokker—Planck equation 28
1.5 Piecewise deterministic processes 32
1.5.1 The Liouville master equation 33
1.5.2 Waiting time distribution and sample paths 33
1.5.3 Path integral representation of PDPs 37
1.5.4 Stochastic calculus for PDPs 39
1.6 Lévy processes 45
1.6.1 Translation invariant processes 46
1.6.2 The Lévy—Khintchine formula 47
1.6.3 Stable Lévy processes 51
References 57

2 Quantum probability 59
2.1 The statistical interpretation of quantum mechanics 59
2.1.1 Self-adjoint operators and the spectral theorem 59
2.1.2 Observables and random variables 63
2.1.3 Pure states and statistical mixtures 65
2.1.4 Joint probabilities in quantum mechanics 70
2.2 Composite quantum systems 74
xvi CONTENTS

2.2.1 Tensor product 75

2.2.2 Schmidt decomposition and entanglement 77
2.3 Quantum entropies 79
2.3.1 Von Neumann entropy 79
2.3.2 Relative entropy 81
2.3.3 Linear entropy 82
2.4 The theory of quantum measurement 83
2.4.1 Ideal quantum measurements 83
2.4.2 Operations and effects 85
2.4.3 Representation theorem for quantum operations 87
2.4.4 Quantum measurement and entropy 92
2.4.5 Approximate measurements 93
2.4.6 Indirect quantum measurements 96
2.4.7 Quantum non-demolition measurements 102
References 104

II DENSITY MATRIX THEORY

3 Quantum master equations 109
3.1 Closed and open quantum systems 110
3.1.1 The Liouville—von Neumann equation 110
3.1.2 Heisenberg and interaction picture 112
3.1.3 Dynamics of open systems 115
3.2 Quantum Markov processes 117
3.2.1 Quantum dynamical semigroups 117
3.2.2 The Markovian quantum master equation 119
3.2.3 The adjoint quantum master equation 124
3.2.4 Multi-time correlation functions 125
3.2.5 Irreversibility and entropy production 128
3.3 Microscopic derivations 130
3.3.1 Weak-coupling Limit 130
3.3.2 Relaxation to equilibrium 137
3.3.3 Singular-coupling limit 138
3.3.4 Low-density limit 139
3.4 The quantum optical master equation 141
3.4.1 Matter in quantized radiation fields 141
3.4.2 Decay of a two-level system 146
3.4.3 Decay into a squeezed field vacuum 149
3.4.4 More general reservoirs 152
3.4.5 Resonance fluorescence 154
3.4.6 The damped harmonic oscillator 160
3.5 Non-selective, continuous measurements 166
3.5.1 The quantum Zeno effect 166
3.5.2 Density matrix equation 167
 CONTENTS xvii

3.6 Quantum Brownian motion 172

3.6.1 The Caldeira-Leggett model 172
3.6.2 High-temperature master equation 173
3.6.3 The exact Heisenberg equations of motion 182
3.6.4 The influence functional 192
3.7 Non-linear quantum master equations 201
3.7.1 Quantum Boltzmann equation 201
3.7.2 Mean field master equations 203
3.7.3 Mean field laser equations 205
3.7.4 Non-linear Schrödinger equation 208
3.7.5 Super-radiance 210
References 216
4 Decoherence 219
4.1 The decoherence function 220
4.2 An exactly solvable model 225
4.2.1 Time evolution of the total system 225
4.2.2 Decay of coherences and the decoherence factor 227
4.2.3 Coherent subspaces and system-size dependence 231
4.3 Markovian mechanisms of decoherence 232
4.3.1 The decoherence rate 233
4.3.2 Quantum Brownian motion 234
4.3.3 Internal degrees of freedom 235
4.3.4 Scattering of particles 237
4.4 The damped harmonic oscillator 242
4.4.1 Vacuum decoherence 242
4.4.2 Thermal noise 246
4.5 Electromagnetic field states 251
4.5.1 Atoms interacting with a cavity field mode 251
4.5.2 Schrödinger cat states 257
4.6 Caldeira-Leggett model 262
4.6.1 General decoherence formula 263
4.6.2 Ohmic environments 265
4.7 Decoherence and quantum measurement 269
4.7.1 Dynamical selection of a pointer basis 270
4.7.2 Dynamical model for a quantum measurement 275
References 278

III STOCHASTIC PROCESSES IN HILBERT SPACE

5 Probability distributions on Hilbert space 283
5.1 The state vector as a random variable in Hilbert space 283
5.1.1 A new type of quantum mechanical ensemble 283
5.1.2 Stern-Gerlach experiment 288
5.2 Probability density functionals on Hilbert space 291
xviii CONTENTS

5.2.1 Probability measures on Hilbert space 291

5.2.2 Distributions on projective Hilbert space 294
5.2.3 Expectation values 297
5.3 Ensembles of mixtures 299
5.3.1 Probability density functionals on state space 299
5.3.2 Description of selective quantum measurements 301
References 302

6 Stochastic dynamics in Hilbert space 303

6.1 Dynamical semigroups and PDPs in Hilbert space 304
6.1.1 Reduced system dynamics as a PDP 304
6.1.2 The Hilbert space path integral 311
6.1.3 Diffusion approximation 314
6.1.4 Multi-time correlation functions 316
6.2 Stochastic representation of continuous measurements 320
6.2.1 Stochastic time evolution of gp-ensembles 321
6.2.2 Short-time behaviour of the propagator 322
6.3 Direct photodetection 324
6.3.1 Derivation of the PDP 324
6.3.2 Path integral solution 330
6.4 Homodyne photodetection 335
6.4.1 Derivation of the PDP for homodyne detection 336
6.4.2 Stochastic Schrödinger equation 340
6.5 Heterodyne photodetection 342
6.5.1 Stochastic Schrödinger equation 342
6.5.2 Stochastic collapse models 345
6.6 Stochastic density matrix equations 348
6.7 Photodetection on a field mode 350
6.7.1 The photocounting formula 350
6.7.2 QND measurement of a field mode 354
References 358
7 The stochastic simulation method 361
7.1 Numerical simulation algorithms for PDPs 362
7.1.1 Estimation of expectation values 362
7.1.2 Generation of realizations of the process 363
7.1.3 Determination of the waiting time 364
7.1.4 Selection of the jumps 367
7.2 Algorithms for stochastic Schrödinger equations 367
7.2.1 General remarks on convergence 368
7.2.2 The Euler scheme 370
7.2.3 The Heun scheme 370
7.2.4 The fourth-order Runge-Kutta scheme 370
7.2.5 A second-order weak scheme 372
7.3 Examples 373
 CONTENTS xix

7.3.1 The damped harmonic oscillator 373

7.3.2 The driven two-level system 377
7.4 A case study on numerical performance 380
7.4.1 Numerical efficiency and scaling laws 381
7.4.2 The damped driven Morse oscillator 383
References 389
8 Applications to quantum optical systems 390
8.1 Continuous measurements in QED 391
8.1.1 Constructing the microscopic Hamiltonian 391
8.1.2 Determination of the QED operation 393
8.1.3 Stochastic dynamics of multipole radiation 396
8.1.4 Representation of incomplete measurements 398
8.2 Dark state resonances 401
8.2.1 Waiting time distribution and trapping state 401
8.2.2 Measurement schemes and stochastic evolution 405
8.3 Laser cooling and Lévy processes 409
8.3.1 Dynamics of the atomic wave function 410
8.3.2 Coherent population trapping 416
8.3.3 Waiting times and momentum distributions 421
8.4 Strong field interaction and the Floquet picture 428
8.4.1 Floquet theory 429
8.4.2 Stochastic dynamics in the Floquet picture 431
8.4.3 Spectral detection and the dressed atom 434
References 437

IV NON-MARKOVIAN QUANTUM PROCESSES\

9 Projection operator techniques 441
9.1 The Nakajima-Zwanzig projection operator technique 442
9.1.1 Projection operators 442
9.1.2 The Nakajima-Zwanzig equation 443
9.2 The time-convolutionless projection operator method 445
9.2.1 The time-local master equation 446
9.2.2 Perturbation expansion of the TCL generator 447
9.2.3 The cumulant expansion 451
9.2.4 Perturbation expansion of the inhomogeneity 452
9.2.5 Error analysis 455
9.3 Stochastic unravelling in the doubled Hilbert space 456
References 458
10 Non-Markovian dynamics in physical systems 460
10.1 Spontaneous decay of a two-level system 461
10.1.1 Exact master equation and TCL generator 461
10.1.2 Jaynes-Cummings model on resonance 466
XX CONTENTS

10.1.3 Jaynes-Cummings model with detuning 471

10.1.4 Spontaneous decay into a photonic band gap 474
10.2 The damped harmonic oscillator 474
10.2.1 The model and frequency renormalization 475
10.2.2 Factorizing initial conditions 477
10.2.3 The stationary state 481
10.2.4 Non-factorizing initial conditions 483
10.2.5 Disregarding the inhomogeneity 488
10.3 The spin-boson system 490
10.3.1 Microscopic model 490
10.3.2 Relaxation of an initially factorizing state 491
10.3.3 Equilibrium correlation functions 495
10.3.4 Transition from coherent to incoherent motion 496
References 497

V RELATIVISTIC QUANTUM PROCESSES

11 Measurements in relativistic quantum mechanics 501
11.1 The Schwinger-Tomonaga equation 502
11.1.1 States as functionals of spacelike hypersurfaces 502
11.1.2 Foliations of space-time 506
11.2 The measurement of local observables 507
11.2.1 The operation for a local measurement 508
11.2.2 Relativistic state reduction 511
11.2.3 Multivalued space-time amplitudes 514
11.2.4 The consistent hierarchy of joint probabilities 517
11.2.5 EPR correlations 521
11.2.6 Continuous measurements 523
11.3 Non-local measurements and causality 526
11.3.1 Entangled quantum probes 527
11.3.2 Non-local measurement by EPR probes 530
11.3.3 Quantum state verification 536
11.3.4 Non-local operations and the causality principle 538
11.3.5 Restrictions on the measurability of operators 544
11.3.6 QND verification of non-local states 550
11.3.7 Preparation of non-local states 554
11.3.8 Exchange measurements 555
11.4 Quantum teleportation 557
11.4.1 Coherent transfer of quantum states 557
11.4.2 Teleportation and Bell-state measurement 560
11.4.3 Experimental realization 562
References 565
12 Open quantum electrodynamics 568
12.1 Density matrix theory for QED 569
 CONTENTS xxi

12.1.1 Field equations and correlation functions 569

12.1.2 The reduced density matrix 576
12.2 The influence functional of QED 577
12.2.1 Elimination of the radiation degrees of freedom 577
12.2.2 Vacuum-to-vacuum amplitude 583
12.2.3 Second-order equation of motion 585
12.3 Decoherence by emission of bremsstrahlung 588
12.3.1 Introducing the decoherence functional 589
12.3.2 Physical interpretation 593
12.3.3 Evaluation of the decoherence functional 596
12.3.4 Path integral approach 607
12.4 Decoherence of many-particle states 614
References 617
Index 619
 Part I
Probability in classical and
quantum physics
 1

CLASSICAL PROBABILITY THEORY AND STOCHASTIC

PROCESSES

This chapter contains a brief survey of classical probability theory and stochastic
processes. Our aim is to provide a self-contained and concise presentation of the
theory. We concentrate on those subjects which will be important for the devel-
opments of the following chapters. More details and many interesting examples
and applications of classical probability theory may be found, e.g. in the excellent
textbooks by Feller (1968, 1971) and Doob (1953) for the more mathematically
oriented readers, and by Gardiner (1985), van Kampen (1992) and Reichl (1998)
for readers who are more interested in physical applications.

1.1 The probability space

The fundamental concept of probability theory is the probability space. It con-
sists of three basic ingredients, namely a sample space of elementary events,
a a-algebra of events, and a probability measure on the a-algebra. These no-
tions will be introduced and explained below. We shall follow here the axiomatic
approach to probability which is mainly due to Kolmogorov (1956).

1.1.1 The a algebra of events

-

The formal objects to which we want to attribute probabilities are called events.
Mathematically, these events are subsets of some basic set St, the sample space,
or space of events. The subsets of SI containing just one element w E SI are
referred to as elementary events.
Given some sample space S2 one is usually not interested in all possible subsets
of SI (this may happen, for example, if SI is infinite and non-countable), tht is,
we need to specify which kind of subsets A C It we would like to include in our
theory. An important requirement is that the events form a so-called a algebra,
-

which is a system A of subsets of SI with the following three properties.

1. The sample space itself and the empty set belong to the system of events,
that is 1E A and 0E A.
2. If A 1 E A and A2 C A, then also the union A 1 U A2, the intersection
A 1 n A2, and the difference A 1 \ A2 belong to the system A.
3. If we have a countable collection of events A 1 , A2, ... , An , ... E A, then
also their union UncxL I A, belongs to A.
We shall always write A E A to express that the subset A C fl is an event of
our theory. The above requirements ensure that the total sample space SI and the
4 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

empty set O are events, and that all events of A can be subjected to the logical
operations 'AND', 'OR' and 'NOT' without leaving the system of events. This
is why A is called an algebra. The third condition is what makes A a o--algebra.
It tells us that any countable union of events is again an event.

1.1.2 Probability measures and Kolmogorov axioms

The construction of the probability space is completed by introducing a probabil-
ity measure on the a-algebra. A probability measure is simply a map p : A —4 IR
which assigns to each event A of the o--algebra a real number p(A),

A 1-4 p(A) E (1.1)

The number p(A) is interpreted as the probability of the event A. The probability
measure p is thus required to satisfy the following Kolmogorov axioms:
1. For all events A E A we have

0 < p,(A) <1. (1.2)

2. Probability is normalized as

p(Q ) = 1. (1.3)

3. If we have a countable collection of disjoint events

, d42, An , E A, with Ai n Ai = 0 for i j, (1.4)

then the probability of their union is equal to the sum of their probabilities,
00
p (U,T=1 An ) = E p(An ). (1.5)
n=1

On the basis of these axioms one can build up a consistent probability theory.
In particular, the Kolmogorov axioms enable one to determine the probabilities
for all events which arise from logical operations on other events. For example,
one finds

P(Ai u A2) = it(Ai) + p(A2) - it(Ai n A2). (1.6)

Summarizing, a probability space consists of a sample space S-2, a a-algebra

A of events, and a probability measure p on A. This concept of a probability
space constitutes the axiomatic basis of classical probability theory. Of course,
from a physical viewpoint one has to relate these abstract notions to experience
and to specific theoretical models of reality, which may be a non-trivial task.
 RANDOM VARIABLES 5

1.1.3 Conditional probabilities and independence

An important concept of probability theory is the notion of statistical indepen-
dence. This concept is often formulated by introducing the conditional probability
p(A i 1A2 ) of an event A1 under the condition that an event A2 occurred,

ft(Ai n A2)
POI IA2) = (1.7)
p(A2)

Of course, both events A1 and A2 are taken from the a-algebra and it is assumed
that p(A2) > 0. These events are said to be statistically independent if

P(Ai1A2) = P (A I), (1.8)

or, equivalently, if

p,(Ai n A2 ) = ti(Ai) • p(A2)• (1.9)

This means that the probability of the mutual occurrence of the events A1 and
A2 is just equal to the product of the probabilities of A 1 and A2.
If we have several events A1 , A2, . . . A, the condition of statistical inde-
pendence is the following: For any subset (i 1 , j 2 ,... , ik) of the set of indices
(1, 2, ... , n) we must have

(Air n Ai2 n n Aik) = p(Ati )p4'122) ...p(241k), (1.1 0 )

which means that the joint occurrence of any subset of the events A i factorizes.
As simple examples show (Gardiner, 1985), it is not sufficient to check statistical
independence by just considering all possible pairs Ai , Ai of events.
An immediate consequence of definition (1.7) is the relation

P OI 1 A2) = P(A2 ) P(A1)

P(A2)

which is known as Bayes's theorem.

1.2 Random variables

The elements w of the sample space SI can be rather abstract objects. In practice
one often wishes to deal with simple numbers (integer, real or complex numbers)
instead of these abstract objects. For example, one would like to add and multiply
these numbers, and also to consider arbitrary functions of them. The aim is thus
to associate numbers with the elements of the sample space. This idea leads to
the concept of a random variable.
6 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

I X ?t ) I
B = X (A)I
Px (13) = it(X -1 (B))

FIG. 1.1. Illustration of the definition of a random variable. A random variable

X is a map from the sample space to the set of real numbers. The probability
that the random number falls into some Borel set B is equal to the probability
measure p(A) of the event A = X -1 (B) given by the pre-image of B.

1.2.1 Definition of random variables

A random variable X is defined to be a map

X : Il 1-÷ R, (1.12)

which assigns to each elementary event w E 0 a real number X(w). Given some
w the value

x = X(w) (1.13)

is called a realization of X. In the following we use the usual convention to denote

random numbers by capital letters, whereas their realizations are denoted by the
corresponding lower case.
Our definition of a random variable X is not yet complete. We have to impose
a certain condition on the function X. To formulate this condition we introduce
the a-algebra of Borel sets' of R which will be denoted by B. The condition on
the function X is then that it must be a measurable function, which means that
for any Borel set B E B the pre-image A = X -1 (B) belongs to the a-algebra A
of events. This condition ensures that the probability of X -1 (B) is well defined
and that we can define the probability distribution of X by means of the formula

Px (B) = p (X -1 (B)) . (1.14)

A random variable X thus gives rise to a probability distribution Px (B) on the

Borel sets B of the real axis (see Fig. 1.1).

1 The a-algebra of Borel sets of R is the smallest a-algebra which contains all subsets of the
form (—oc, x), x E R. In particular, it contains all open and closed intervals of the real axis.
 RANDOM VARIABLES 7

Particular Borel sets are the sets (—oc, x] with x E IR. Consider the pre-images
of these set, that is the sets

A x E {r.,) E 1X(w) _< x} . (1.15)

By the condition on X these sets are measurable for any x E Et which enables
one to introduce the function

Fx (x) E p(A x ) = p ({w E (-V (w ) < x}) . (1.16)

For a given x this function yields the probability that the random number X
takes on a value in the interval (—co, xj. Fx (x) is referred to as the cumulative
distribution function of X. One often employs the following shorthand notation,

Fx ( x ) E p ( x < z ). (1.17)

As is easily demonstrated, the cumulative distribution function has the following

properties:
1. Fx (x) increases monotonically,

Fx (x i ) < Fx (x2), for z 1 < X2. (1.18)

2. Fx (x) is continuous from the right,

lim Fx (x + E) = Fx (x). (1.19)

E-H-0

3. Fx (x) has the following limits,

lim Fx (x) = 0, lim Fx (x) = 1. (1.20)

X-Y-00 x—H-0.0

The random variable X is said to have a probability density Px (x) if the

cumulative distribution function can be represented as
x
Fx (x) = f dx p x (x). (1.21)
_co

If Fx (x) is absolutely continuous we get the formula

dFy (x)
(1.22)
Pv(x) = dx

In the following we often represent distribution functions by their densities px (x)

in this way, as is common in the physics literature. This is permissible if we
allow Px (x) to involve a sum of ES-functions and if we exclude certain singular
8 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

distribution functions (Feller, 1971). In terms of the density of X we may also

write eqn (1.14) as

Px (B) = f dx px (x), (1.23)

B

where the integral is extended over the Borel set B.

We have just considered a single random variable. One can, of course, study
also an arbitrary collection

X = (Xi • X2, • • • , Xci) (1.24)

of random variables which are defined on the same probability space. The vector-
valued function X : S../ H÷ Rd is called a multivariate random variable or a random
vector, which simply means that each component Xi is a real-valued random
variable. For a given ci.) E 52 the quantity z = X (w) = (X1 GO , ... , X d(w)) is
a realization of the multivariate random variable. The joint probability density
of a multivariate random variable is denoted by px (x). The probability for the
variable to fall into a Borel set B C Rd is then given by

.13x (B) = ft (X -1 (B)) = f dd x Ay (x). (1.25)

B

In accordance with our former definition, two random variables X 1 and X2

on the same probability space are said to be statistically independent if

xi , X2 < x2) = P(X 1 < xi) • P(X2 5 z2) (1.26)

for all x l , x2 . Here, the left-hand side is the shorthand notation for the probability

p(Xi < x i , X2 < z2) E p ({ w E 121X1(w) < xi and X2 (w) < x2 }) . (1.27)

The joint statistical independence of several random variables is defined analo-

gously to definition (1.10).

1.2.2 Transformation of random variables

Given a d-dimensional random variable X we can generate new random variables
by using appropriate transformations. To introduce these we consider a Borel-
measurable function

g : Rd —4 le . (1.28)

Such a function is defined to have the property that the pre-image g -1 (B) of
any Borel set B C Rf is again a Borel set in Rd . Thus, the equation

Y = g (X) (1.29)
 RANDOM VARIABLES 9

defines a new f-dimensional random variable Y. If Px is the probability distri-

bution of X, then the probability distribution of Y is obtained by means of the
formula

Py (B) = Px (g -1 (B)). (1.30)

The corresponding probability densities are connected by the relation

py (y) = f ddx (5 ( f ) (y - g(x))p x (x), (1.31)

where 6(f) denotes the f-dimensional 6-function. This formula enables the de-
termination of the density of Y = g(X). For example, the sum Y = X1 + X2 of
two random variables is found by taking g(x i , x 2 ) = xl + x2 . If X1 and X2 are
independent we get the formula

py (y) = f dxi px i (xi )px2 (y - xi), (1.32)

which shows that the density of Y is the convolution of the densities of X1 and
of X2.

1.2.3 Expectation values and characteristic function

An important way of characterizing probability distributions is to investigate
their expectation values. The expectation value or mean of a real-valued random
variable X is defined as 2
+00 +Do
E(X) _= f xdFx (x) = f dx xp x (x). (1.33)

Here, the quantity dFx (x) is defined as

dFx (x) E Fx (x + dx) - Fx (x) = ti(x <X <z + dx). (1.34)

Correspondingly, the integrals in (1.33) are regarded as Lebesgue-Stieltjes inte-

grals (Feller, 1971). More generally, the expectation value of some measurable
function g(X) of X is defined to be

+Da +Da
E(g(X)) = f g(x)dFx (x) = f dxg(x)px (x). (1.35)
-... -0.0

Particularly important expectation values are the moments of order m:

2 We follow the usual convention of mathematical probability theory and denote the expec-
tation value of a classical random variable by the symbol E, in order to distinguish it from the
quantum mechanical expectation value which will be denoted by angular brackets.
10 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

E(Xm) = f xmc/Fx (x) = f dx xmpx(x). (1.36)

The variance of a random variable X is defined by

— E(X)r) = E(x2) E(X) 2 .Var(X)E (1.37)

The significance of the variance stems from its property to be a measure for
the fluctuations of the random variable X, that is, for the extent of deviations
of the realizations of X from the mean value E(X). This fact is expressed, for
example, by the Chebyshev inequality which states that the variance controls
the probability for such deviations, namely for all E > 0 we have

PDX — E(X)1 > e) < -j

1 Var(X). (1.38)

In particular, if the variance vanishes then the random number X is, in fact,
deterministic, i.e. it takes on the single value x = E(X) with probability 1. The
variance plays an important rôle in the statistical analysis of experimental data
(Honerkamp, 1998), where it is used, for example, to estimate the standard error
of the mean for a sample of realizations obtained in an experiment.
For a multivariate random variable X = (X1 7 X2,... Xd) one defines the
matrix elements of the covariance matrix by

Cov(Xi, Xi ) E ([Xi E(X)] [Xi — E(Xi )]). (1.39)

The d x d matrix with these coefficients is symmetric and positive semidefinite.
As is well known, the statistical independence of two random variables X1 , X2
implies that the off-diagonal element Cov(XI, X2) vanishes, but the converse is
not true. However, the off-diagonal elements are a measure of the linear depen-
dence of X1 and X2. To see this we consider, for any two random variables with
non-vanishing variances, the correlation coefficient

COV(X1 7 X2)
Cor(X1 7 X2) (1.40)
N/Var(X1)Var(X2)
which satisfies 1Cor(Xi , X2 )1 < 1. If the absolute value of the correlation coeffi-
cient is equal to one, 1Cor(X1, X2)1 = 1, then there are constants a and b such
that X2 = aXi + b with unit probability, i.e. X2 depends linearly on X1 .
Let us finally introduce a further important expectation value which may
serve to characterize completely a random variable. This is the characteristic
function which is defined as the Fourier transform of the probability density,

G(k) = E (exp [ikX]) = f dx px(x)exp (ikx) . (1.41)

 STOCHASTIC PROCESSES 11

It can be shown that the characteristic function G(k) uniquely determines the
corresponding probability distribution of X. Under the condition that the mo-
ments of X exist the derivatives of G(k) evaluated at k = 0 yield the moments
of X,

1 dm
E(Xm) = G(k). (1.42)
im dkm k=0

For this reason G(k) is also called the generating function. For a multivariate
random variable the above expression for the characteristic function is readily
generalized as follows,

[ d
G(k1,k2, . . . , kd) =E ( exp i E kj xj (1.43)

An important property of the characteristic function is the following. As we

have already remarked, if X and Y are two independent random variables the
probability density of their sum Z = X + Y is the convolution of the densities
of X and Y. Consequently, the characteristic function of Z is the product of the
characteristic functions of X and Y.

1.3 Stochastic processes

Up to now we have dealt with random variables on a probability space without
explicit time dependence of their statistical properties. In order to describe the
dynamics of a physical process one needs the concept of a stochastic process
which is, essentially, a random variable whose statistical properties change in
time. The notion of a stochastic process generalizes the idea of deterministic
time evolution. The latter can be given, for example, in terms of a differential
equation which describes the deterministic change in time of some variable. In
a stochastic process, however, such a deterministic evolution is replaced by a
probabilistic law for the time development of the variable.
This section gives a brief introduction to the theory of stochastic processes.
After a formal definition of a stochastic process we introduce the family of joint
probability distributions which characterizes a stochastic process in a way that
is fully sufficient for practical purposes. In a certain sense, the differential equa-
tion of a deterministic theory is replaced by such a family of joint probability
distributions when one deals with stochastic processes. This is made clear by a
theorem of Kolmogorov which is also explained in this section.

1.3.1 Formal definition of a stochastic process

In mathematical terms, a stochastic process is a family of random variables X (t)
on a common probability space depending on a parameter t E T. In most physical
applications the parameter t plays the rôle of the time variable. The parameter
space T is therefore usually an interval of the real time axis.
12 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

Corresponding to this definition, for each fixed t the quantity X(t) is a map
from the sample space St into R. A stochastic process can therefore be regarded
as a map
X:ftxT—+R, (1.44)
which associates with each co E Il and with each tETa real number X(co, t).
Keeping co fixed, we call the mapping
t F--+ X (c.o , t), t E T , (1.45)
a realization, trajectory, or sample path of the stochastic process.
In the above definition the map (1.44) can be quite general. Because of this
the notion of a stochastic process is a very general concept. We need, however,
one condition to be satisfied in order for X (t) to represent a random variable
for each fixed t. Namely, for each fixed t the function X(t) which maps 1 2 into R
must be measurable in the sense that the pre-images of any Borel set in 111 must
belong to the algebra of events of our probability space.
A multivariate stochastic process X(t) is defined similarly: It is a vector-
valued stochastic process X(t) = (Xi (t), X2(t), ... , Xd(t)), each component
Xi(t), i = 1,2, ... , d, being a real-valued stochastic process. Thus, formally
a multivariate stochastic process can be regarded as a map
X:0><T-4 Rd . (1.46)
1.3.2 The hierarchy of joint probability distributions
What characterizes a stochastic process is the way in which the random variables
X (t) for different times t are related to each other, that is, the degree of statistical
dependencies between the random variables of the family.
According to the definition given above, a stochastic process is, formally
speaking, nothing but a time-dependent random variable. It can therefore be
characterized uniquely if one constructs a probability space and a map of the
form (1.46) on it. This is however not the way one characterizes a stochastic
process in practice. In most applications one tries to construct an appropriate
process by the observation of the statistical correlations between the random
variables X(t) at a finite set of discrete times tv . This is done on the basis of
experimental data, employing some phenomenological model, or with the help
of some underlying microscopic physical theory.
Thus, the physical theory usually provides a so-called family of finite joint
probability distributions which is defined as follows (see Fig. 1.2). We take a set
t1 , t 2 , ... , tni of discrete times and Borel sets B1 , B2, . .. ,13„ in Rd , and consider
for a multivariate stochastic process X(t) the quantity
P(Bi ,ti; B2, t2; ... ; B, , t ni ) E p (X(ti) E B1, X(t2) E B2, ... , X(t m ) E B in ) •
(1.47)
This quantity is a joint probability distribution of order m. It gives the probabil-
ity that the process X(t) takes on some value in B1 at time t 1 , some value in B2
 STOCHASTIC PROCESSES 13

trn
trn_i

•
•

t2 1
B2
tl I [
B1

FIG. 1.2. A sample path X(t, co) of a stochastic process which passes at
times t i , t 2 , ... , t m _ i , tm the sets B 1 , B2, ... , Bni _i, B m , respectively.
The probability for such a path to occur is given by the joint probability
P(Bi, t1; ... ; B, t 7 ).

at time t 2 , ..., and some value in B m at time t m . The set of all joint probability
distributions for all m = 1, 2, ... , all discrete times t u , and all Borel sets B y is
called the family of finite joint probability distributions of the stochastic process.
Each stochastic process gives rise to such a family of joint probabilities. It
follows immediately from the above definition that the joint probabilities satisfy
the Kolmogorov consistency conditions:
P(Rd , t) = 1, (1.48)
P(Bi ,t i ;... ;B, t) > 0, (1.49)
P(Bi, t1; ;Bm_1,tm_i; Rd , tm) = P(Bi, t1; ... ; B 7 _ 1 , t7 _ 1 ), (1.50)
.N.B7r (i), t 7r (i); . .. ; Ar ern y, t7r ( m )) — P(B1,t 1 ; . . . ; Bm , tm ). (1.51)
The first two conditions state that the distributions must be non-negative and
that the probability for the sure event X(t) E Rd is normalized to 1. The third
condition asserts that for I/ > 1 the sure event X(t) E Rd can always be omitted
from the set of arguments, whereas the fourth condition means that the joint
probability distribution is invariant under all permutations 7 of its arguments.
Of course, for a given stochastic process X(t) the consistency conditions
(1.48)-(1.51) are a trivial consequence of the definition (1.47). The important
point to note is that the following non-trivial theorem is also true: Suppose
that a family of functions is given which satisfy conditions (1.48)-(1.51). Then
there exists a probability space and a stochastic process X(t) on this space such
that the family of joint probabilities pertaining to X(t) coincides with the given
family.
14 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

This is the theorem of Kolmogorov. It ensures that for any consistent family
of joint probabilities there exists a stochastic process X(t) on some probability
space. It should be noted, however, that the process X(t) is not unique; for a
given family of joint probability distributions there may exist different stochastic
processes, where the term different means that these processes may be different
on events with non-zero measure.
In practice, the non-uniqueness of X(t) usually does not cause any difficulty
since the family of finite-dimensional joint probabilities uniquely determines the
probabilities of all events which can be characterized by any finite number of
random variables. Thus, if we assess, for example, some stochastic model by
comparison with a set of experimental data, which is always finite, no problem
due to the non-uniqueness of the process will ever be encountered.

1.4 Markov processes

Markov processes play an important rôle in physics and the natural sciences.
One reason for this fact is that many important processes, as for example the
processes arising in equilibrium statistical mechanics, are Markovian provided
one chooses an appropriate set of variables. Another reason is that many types of
stochastic processes become Markovian by an appropriate extension of the state
space. Finally, Markov processes are relatively easy to describe mathematically.
In this section we define and classify the most important Markov processes and
briefly review their properties.

1.4.1 The Chapman Kolmogorov equation

—

Essentially, a Markov process is a stochastic process X(t) with a short mem-

ory, that is a process which rapidly forgets its past history. This property is
what makes a Markov process so easy to deal with, since it ensures that the
whole hierarchy of joint probabilities introduced in the preceding section can be
reconstructed from just two distribution functions.
The condition for the rapid decrease of memory effects can be formulated in
terms of the conditional joint probabilities as follows,

ti(X(t) E BA(t 7 ) = x,„... ,X(ti) = x1 ) = p(X(t) E B1X(t 7i ) = x ni ).

(1.52)

This is the Markov condition. It is assumed to hold for all m = 1, 2, 3, ... , for
all ordered sets of times,

t1 < t 2 < ... < tin < t, (1.53)

for all Borel sets B and all x 1 , x2 , ... , x,,,,, E Rd . The Markov condition states
that the probability for the event X(t) E B, conditioned on m previous events
X(4) = x i , ... , X(t ni ) = x m , only depends on the latest event X(t 1 ) = xni.
 MARKOV PROCESSES 15

In the following we shall discuss the consequences of the Markov condition

by introducing the joint probability densities

P(B,„ tffi ;... ; Bi,ti) = f dx 7r, ... f dx 1 pm (x,,,t ra ; ... ;xi,ti) (1.54)
B n, Bi

and the corresponding conditional probability densities

pk±i(xk+1,44-1; ••• ;xl • t1)

Plik(Xk+1,4+1; • • • ;Xk4-1•4+11X1c,tki • • • ;Xi,t1) =
Pk(Xkjk; • . . ;x i ,ti )
(1.55)
in terms of which the Markov condition takes the form

Piini(x, tIx Tri, trn ; .. . ; x l, ti) = 23 111 (z, tlxm, tm ). (1.56)

This equation demonstrates that the quantity pi l l (x, tlx', t') plays a crucial rôle
in the theory of Markov processes.
For any stochastic process (not necessarily Markovian) p i l l (x, t z', t') is equal
to the probability density that the process takes on the value x at time t under
the condition that the process took the value x' at some prior time t'. This condi-
tional probability is therefore referred to as the conditional transition probability
or simply as the propagator. We introduce the notation

T(x, t x i , t i ) E Pili(x,tlx i ,e) (1.57)

for the propagator. As follows directly from the definition, the propagator fulfils
the relations,

f dx T(z,t x' ,t') = 1, (1.58)

lim T(x,t x' ,t') = S(x — x'). (1.59)

t—q ,
The first equation expresses the fact that the probability for the process to take
any value at any fixed time is equal to 1, and the second equation states that
with probability 1 the process does not change for vanishing time increment.
The probability density p i (x, t), which is simply the density for the uncondi-
tioned probability that the process takes on the value x at time t, will be denoted
by

p(x , t) -_E pi (x , t). (1.60)

The density p(x,t) is connected to the initial density at some time t o by the
obvious relation

P(x,t) = f dx' T (x, tlx' , to)P(x' , to) • (1.61)

16 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

A stochastic process is called stationary if all joint probability densities are

invariant under time translations, that is if for all 7-

Pm(Xm,tm ± r; — ;z1,t1+ 7) = Prri(Xmltm;... ;Xl,t1). (1.62)

In particular, stationarity implies that the probability density p is independent

of time, p(x , t) = p(x), and that the propagator T (x , tlx' , t') depends only on the
difference t t' of its time arguments. With the help of stationary processes one
-

describes, for example, equilibrium fluctuations in statistical mechanics.

A process is called homogeneous in time if the propagator depends only on the
difference of its time arguments. Thus, a stationary process is homogeneous in
time, but there are homogeneous processes which are not stationary. An example
is provided by the Wiener process (see below).
The great simplification achieved by invoking the Markov condition derives
from the fact that the total hierarchy of the joint probabilities can be recon-
structed from an initial density p(x , to) and an appropriate propagator. Accord-
ing to eqn (1.61) the density p(x , t) for later times t > to is obtained from the
initial density and from the propagator. Thus, also the joint probability distri-
bution p2 (x, t; z', t') is known, of course. By virtue of the Markov condition all
higher-order distribution functions can then be constructed, provided the prop-
agator fulfils a certain integral equation which will now be derived.
To this end, we consider three times t i < t2 < t3 and the third-order distribu-
tion p3 and invoke the definition of the conditional probability and the Markov
condition to obtain

P3(x3,t3;x2,t2;xi,ti) =P112(x3,t31x2,t2;xi,t0p2(x2,t2;xi,ti) (1.63)

=Pi1i(x3, t31x2,t2)Pi1l(x2,t2lxi, ti)pi(xi,t1)•

We integrate this equation over x2 ,

P2(x3, t3; xi, ti.) = pi (x i , t i ),

f dx 2pi i i (x 3 , t 3 42 , t2 )pi i i (x 2 t2 lxi ,t i ), (1.64)

and divide by p i (xi , ti),

P111 (X3 7 6411 tl) = f dX2P1 li (X3, t3 X2,t2)P111(X2,t21X11t1)• (1.65)

On using our notation (1.57) for the propagator we thus have

T(X3, 641, ti) – f dX2 T(X3, t3 4 2, t2)T(X2, t21X1, t1), (1.66)

which is the Chapman Kolmogorov equation.

-

The Chapman-Kolmogorov equation admits a simple intuitive interpretation,

which is illustrated in Fig. 1.3: Starting at the point x i at time t i , the process
reaches the point x3 at a later time t3 . At some fixed intermediate time t2
 MARKOV PROCESSES 17

time

t3

t1
Xi

space

FIG. 1.3. Illustration of the Chapman-Kolmogorov equation (1.66).

the process takes on some value x2. The probability for the transition from
(xi, t i ) to (x3 , t3 ) is obtained by multiplying the probabilities for the transitions
(x i , t i ) —+ (x2, t2) and (x 2 , t2) —+ (x 3 , t3 ) and by summing over all possible
intermediate positions x2.
Once we have a propagator T (x,t x' ,t') and some initial density p(x,t o ) we
can construct the whole hierarchy of joint probability distributions. As we have
already seen, the propagator and the initial density yield the time-dependent
density p(x,t). It is easy to verify that with these quantities all m-th order joint
probability densities are determined through the relation
rrt-1

Prn (Xrn 7 trn; Xrn-1 7 trn-1 ; . • . ; X1 , tl) = H T(x„,,,tv±ilxv,t0p(xi,ti), (1.67)

where to < t i < t2 < ...

In summary, to define a stochastic Markov process we need to specify a
propagator T (x , tx' , t') which obeys the Chapman-Kolmogorov equation (1.66),
and an initial density p(x,t0). The classification of Markov processes therefore
amounts, essentially, to classifying the solutions of the Chapman-Kolmogorov
equation.

1.4.2 Differential Chapman Kolmogorov equation

-

The Chapman-Kolmogorov equation (1.66) is an integral equation for the con-

ditional transition probability. In order to find its solutions it is often useful
to consider the corresponding differential form of this equation, the differential
Chapman Kolmogorov equation.
-

We suppose that the propagator T(x,tlx' ,e) is differentiable with respect

to time. Differentiating eqn (1.66) we get the differential Chapman-Kolmogorov
18 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

equation,

a , t'). (1.68)
, t') = A(t)T (x ,

Here, A is a linear operator which generates infinitesimal time translations. It is

defined by the action on some density p(x) ,

1
A(t) p(x) f dx' [T (x , t + xi, t) 6(x — xi)] p(x' )
. 1.!121KI Kt.
= 211,0 [f dx' T (x t + tV t) p(x' ) — (x)]. (1.69)

In the general case the operator A may depend on t. However, for a homogeneous
Markov process the propagator T (x , t + Atlx' ,t) for the time interval from t to
t + At does not depend on t and, thus, the generator is time independent in this
case.
For a homogeneous Markov process we can write the propagator as Ty (xIx' )
where T = t - t' > 0 denotes the difference of its time arguments. The Chapman-
Kolmogorov equation can then be rewritten as

T 7- +7 , (xIx' ) = fdx" T r (xlx" )T, 7- i ( x" Ix' ) ,

- T , _ 0.
TI > (1.70)

Once the generator A is known, the solution of the Chapman-Kolmogorov equa-

tion for a homogeneous Markov process can be written formally as

TT (XIX ' ) = exp (TA) (5(x - r> 0. (1.71)

These equations express the fact that the one-parameter family { TT 7- > 0} of
conditional transition probabilities represents a dynamical semigroup. The term
semi-group serves to indicate that the family {TT IT > 0} is, in general, not a full
group since the parameter T is restricted to non-negative values.
From a physical viewpoint the semigroup property derives from the irre-
versible nature of stochastic processes: Suppose that at time to an initial density
P(x, to) is given. The above family of conditional transition probabilities then
allows us to propagate uniquely this initial density to times t = to + r > to.
With the help of eqns (1.61) and (1.71) we get

P(x, t) = exp (TA)P(x, to). (1.72)

However, the resulting process is, in general, not invariant with respect to time
inversions, which means that it is not possible to find for each p(x , t o ) a prob-
ability density p(x , t) at some earlier time t < to which evolves into p(x , t o ).
Mathematically, this means that the range of the operator exp (TA) is contract-
ing for increasing 7- , that is, this operator is not invertible in the total space of
 MARKOV PROCESSES 19

t t+T t + 27-

FIG.1.4. Schematic picture of the irreversible nature of a dynamical semigroup.

The figure indicates the contraction of the range of the propagators TT for
increasing 'T.

all probability distributions (Fig. 1.4). This is why irreversible processes enable
one to distinguish the future from the past.
The above situation occurs, for example, if the process relaxes to a unique
stationary state p, (z) in the limit of long times,

lim p(x,t) = p.(x). (1.73)

t—H-00
Such processes arise in statistical mechanics, for example, when one studies the
relaxation of closed physical systems to an equilibrium state. Then, clearly, the
stationary density p.(x) must be a zero mode of the generator,

Ap(x) = 0. (1.74)
Consequently, also backward propagation leaves p(z) invariant, and there is no
way of obtaining any specific distribution other than p,, (z) at any prior time.
In the following sections we shall introduce three basic types of Markov pro-
cesses which can be distinguished by the form of their generator or, equivalently,
by the short-time behaviour of their propagator.

1.4.3 Deterministic processes and Lionville equation

The simplest example of a Markov process is provided by a deterministic process.
It is defined by some initial density p(x,t 0 ) and by a propagator which describes
a deterministic time evolution corresponding to a system of ordinary differential
equations
d
— x(t) = g(x(t)), x(t) E Rd . (1.75)
dt
20 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

Here, g(x) denotes a d-dimensional vector field. For simplicity we assume this
system to be autonomous, i.e. that the vector field g(x) does not depend explic-
itly on time, such that the resulting process becomes homogeneous (see below).
The most prominent examples of this type of process are the processes that
arise in equilibrium statistical mechanics, in which case eqn (1.75) represents the
Hamiltonian equations of motion in phase space.
To such an ordinary differential equation there corresponds the phase flow
which is denoted by (Dt (x). This means that for fixed x the phase curve

t cD t (x) (1.76)
represents the solution of eqn (1.75) corresponding to the initial value 4) 0 (x) = z.
The sample paths of the deterministic process are given by the phase curves
(1.76). Thus, the propagator for such a process is simply

T (x, , t') = (x 4)t-t , ( 4) • (1.77)

This equation tells us that the probability density for the process to reach the
point x at time t, under the condition that it was in x' at time t', is different
from zero if and only if the phase flow carries x' to x in the time interval from
t' to t, that is, if and only if x = t _ t , (x`).
As is easily verified the propagator (1.77) fulfils the relations (1.58) and (1.59).
On using the group property of the phase flow, which may be expressed by

= (1t+8(x), (1.78)

one can also show that (1.77) satisfies the Chapman-Kolmogorov equation. Thus
we have constructed a solution of the Chapman-Kolmogorov equation and de-
fined a simple Markov process.
In order to find the infinitesimal generator A for a deterministic process we
insert expression (1.77) into definition (1.69):

Ap(x) = lim
At-kJ At
f dx' [6(x - (DAt (X I )) - (5 (z - x')] p(x 1 )
d
— f dx' 6(x - (D t (x'))p(x') = f dx' g t (x')[ 3 5(z - x')] p(x 1 ),

dt t=o axi
a [gi (x)p(x)] . (1.79)
axi
Here, the gt (x) denote the components of the vector field g(x), and summation
over the index i is understood. Thus, the generator of the deterministic process
reads
D
=- gi (x), (1.80)
axi
and the differential Chapman-Kolmogorov equation takes the form
 MARKOV PROCESSES 21

—
a T(x,tlx` ,e) = - —
,,a [gi (x)T(x,tlx' ,e)]. (1.81)
at axi

This is the Liouville equation for a deterministic Markov process corresponding

to the differential equation (1.75). Of course, the density p(x , t) fulfils an equation
which is formally identical to eqn (1.81).

1.4.4 Jump processes and the master equation

The deterministic process of the preceding subsection is rather simple in that it
represents a process whose sample paths are solutions of a deterministic equation
of motion; only the initial conditions have been taken to be random. Now we con-
sider processes with discontinuous sample paths which follow true probabilistic
dynamics.

1.4.4.1 Differential Chapman-Kolmogorov equation We require that the sam-

ple paths of X(t), instead of being smooth solutions of a differential equation,
perform instantaneous jumps. To formulate a differential Chapman-Kolmogorov
equation for such a jump process we have to construct an appropriate short-time
behaviour for its propagator.
To this end we introduce the transition rates W(xix', t) for the jumps which
are defined as follows. The quantity W(xix', t)At is equal to the probability
density for an instantaneous jump from the state x' into the state x within the
infinitesimal time interval [t, t + st], under the condition that the process is in x'
at time t. Given that X(t) = x', the total rate for a jump at time t is therefore

r(x i , t) = f dxW(xlx`, t), (1.82)

which means that F(x', t)At is the conditional probability that the process leaves
the state x' at time t by a jump to some other state.
An appropriate short-time behaviour for the propagator can now be formu-
lated as

T (x,t + Atlx' ,t) = W (xix' ,t)At + (1 - F(x' ,t)At) 6(x - x') + 0(At 2 ). (1.83)

The first term on the right-hand side gives the probability for a jump from x' to
x during the time interval from t to t+ At. The prefactor of the delta function of
the second term is just the probability that no jump occurs and that, therefore,
the process is still in x' at time t+ At. As it should do, for At .- 0 the propagator
approaches the delta function 6(x - x') (see eqn (1.59)). In view of (1.82) the
propagator also satisfies the normalization condition (1.58).
It is now an easy task to derive the differential Chapman-Kolmogorov (1.68)
equation for the jump process. Inserting (1.83) into (1.69) we find for the gener-
ator of a jump process
22 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

1
A(t)p(x) = urn f dx' [T (x ,t + AO' , t) - ( 5 (x - 41 p(x` )
At->0 At
=u a1---t-f dx' [W (xlx` , t) At - F (x` , t) AtS (x - x')] p(x' )

= f dxi W (xix' ,t)p(xl) - F (x ,t)p(x)

= f dxi [W (xix' , t) p (x i ) - W (x i ix , t)P(x)] • (1.84)

In the last step we have used definition (1.82) for the total transition rate. This
immediately leads to the equation of motion for the propagator,
a tlx' , t') = A(t)T (x , tlx' , t') (1.85)

= f dx" [W (xix i ' 1 0 7 7 (x" , tix' , t') - W (x" lx , t)T (x , tlx' , t')] .

This is the differential Chapman-Kolmogorov equation for the jump process. It

is also called the master equation. The same equation holds, of course, for the
density p(x , t) ,
a
— p(x , t) = A(t)p(x , t)
at
---- f dx' [W (xl x i , t)p(x l , t) - W (x l Ix , t)p(x , t)] . (1.86)

This equation is also referred to as the master equation. It must be kept in mind,
however, that the master equation is really an equation for the conditional transi-
tion probability of the process. This is an important point since a time-evolution
equation for the first-order density p(x , t) is not sufficient for the definition of a
stochastic Markov process.
The master equation (1.86) has an intuitive physical interpretation as a bal-
ance equation for the rate of change of the probability density at x. The first
term on its right-hand side describes the rate of increase of the probability den-
sity at x which is due to jumps from other states x' into x. The second term is
the rate for the loss of probability due to jumps out of the state x.
Note that in the above derivation we have not assumed that the process is
homogeneous in time. In the homogeneous case the transition rates must be time
independent, W (x Ix' , t) = W(xix'). Then also the total transition rate F(x1 ) is
time independent, of course.
For the case of an integer-valued process, which may be univariate or multi-
variate, we write X(t) = N(t). The first-order probability distribution for such
a discrete process is defined by

P (n , t) = ti (N (t) = n) , (1.87)

whereas the propagator takes the form

 MARKOV PROCESSES 23

T(n,tin',e) = pt(N(t)= niN(e) = n'). (1.88)

The corresponding master equation for the probability distribution then takes
the form

a
— P(n t) =
+co
E [W(nini, t)P(ni , t) — W(ni In, t)P(n,t)1, (1.89)
at '
n'= —co

with a similar equation for the propagator.

1.4.4.2 The homogeneous and the non - homogeneous Poisson process Let us
discuss two simple examples for an integer-valued jump process, namely the
homogeneous and the non-homogeneous Poisson process. These examples will
be important later on.
As a physical example we consider a classical charged matter current de-
scribed by some current density j(i, t) of the form
-0
3 ( x , t) = (I)e —iw ° t + RI)* eiw° t (1.90)

The current density is assumed to be transverse, that is V-4 • = 0. As is known

from quantum electrodynamics (Bjorken and Drell, 1965) such a current cre-
ates a radiation field, emitting photons of frequency wo at a certain rate ry. An
explicit expression for this rate can be derived with the help of the interaction
Hamiltonian

II (t) = — f d3 x t) • ,44(Y) (1.91)

which governs the coupling between the classical matter current j(i, t) and the
quantized radiation field A(x). Treating HI(t) as a time-dependent interaction
and applying Fermi's golden rule one obtains an explicit expression for the rate
of photon emissions

== e2w°3
27rhe
E 14k,A). j(k)1 2 (1.92)
A=1,2

where j(k4) denotes the Fourier transform of the current density,

= f d3 x34(i)e -ii;i. (1.93)

Obviously, the relation (1.92) involves an integral over the solid angle at into the
direction k of the emitted photons and a sum over the two transverse polarization
vectors (k, À).
The rate ry describes the probability per unit of time for a single photon
emission. We are interested in determining the number N(t) of photon emissions
24 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

over a finite time interval from 0 to t. To this end we assume that the current,
acting as a fixed classical source of the radiation field, is not changed during the
emission processes. We then have a constant rate ry for all photon emissions and
each emission process occurs under identical conditions and independently from
all the previous ones. The number N(t) then becomes a stochastic Markov pro-
cess, known as a (homogeneous) Poisson process. It is governed by the following
master equation for the propagator,
a tin', t') = ryT(n — 1, tin', t') — ryT (n, t' ) . (1.94)

In the following we study the deterministic initial condition N(0) = 0, that is

P (n , 0) = 15n,O• (1.95)
The Poisson process provides an example for a one-step process for which,
given that N (t) = n, only jumps to neighbouring states n ± 1 are possible. The
Poisson process is even simpler for only jumps from n to n + 1 are possible and
the jump rate is a constant Fy which does not depend on n. It should thus be no
surprise that the master equation for the Poisson process can easily be solved
exactly. This may be done by investigating the characteristic function
+00
G (k , t) = E eik e" ) T (n , tin', t') , (1.96)
n=re

with n' and t' held fixed. We have extended here the summation from n = n' to
infinity to take into account that
T (n , tin', t') = 0 for n < n'. (1.97)
This is a direct consequence of the fact that the jumps only increase the number
N(t). Inserting the characteristic function (1.96) into the master equation (1.94)
we obtain

Na G (k , t) = [e 1] G (k t), (1.98)
which is immediately solved to yield
G (k , t) = exp [-y(t — t') (eik 1 )] , (1.99)
where we have used the initial condition G (k t = t') = 1 which, in turn, follows
from T (n, in', t') = 8nni . On comparing the Taylor expansion of the character-
istic function,
+00
G (k , t) = E e an [-y(t
n!
e - (t - t'
n=0
+.0
E e ik ( T, [y(t — ti )]71 - 11 e t) (1.100)
n=n'
(n n')!
 MARKOV PROCESSES 25

5 10 15 20 25 30
t

FIG. 1.5. Sample path of the Poisson process which was obtained from a nu-
merical simulation of the process.

with definition (1.96) we obtain for the propagator of the Poisson process

[7(t — t1)]n—ri
T(n,tin',e) = e --y(t—e) , n > n' (1.101)
(n — n')! —'

As should have been expected the process is homogeneous in time. It is also

homogeneous in space in the sense that T (n, tin', t') only depends on the differ-
ence n — n'. The corresponding distribution P(n,t) becomes

P(n,t) = (21) n e -11t , n _

> 0, (1.102)
n!
which represents a Poisson distribution with mean and variance given by

E(N(t)) = Var(N(t)) = ryt. (1.103)

Figure 1.5 shows a realization of the Poisson process.

The homogeneous Poisson process describes the number N(t) of independent
events in the time interval from 0 to t, where each event occurs with a constant
rate 'y. Let us assume that for some reason this rate may change in time, that
is -y = -y(t). The resulting process is called a non homogeneous Poisson process,
-

since it is no longer homogeneous in time. As one can check by insertion into the
master equation,

,
a T(n,tin' ,e) = -y(t)T (n — 1,tin' ,e) — -y(t)T (n,tin' ,e), (1.104)

the propagator of the non-homogeneous Poisson process is given by

26 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

[12(t ti
T (n, tin', t') = e - ti(t ' t1) n>
_ 7 (1.105)
(n - n')!

and by T (n, tin' , ti) = 0 for n < n!. Here, we have set

p(t , t') (1.106)

The probability P (n, t) is accordingly found to be

[p(t, 0)]n _
P (n, t) = 711 e 1-4 ") , n 0. (1.107)

By means of the above explicit expressions we immediately get the general-

ization of the relations (1.103) for the non-homogeneous Poisson process,

E(N(t)) = Var(N(t)) = p(t , 0). (1.108)

Let us also determine the two-time correlation function which is defined by

E(N (ON (s)) = E(N (s)N (t))

= E n2 n i T(n 2 ,tin i ,$)T(n i ,s10,0), (1.109)
ni ,n2

where in the second line we have assumed, without restriction, that t > s. Dif-
ferentiating with respect to t and invoking the master equation we find

a E(N (t)N (s))

= 'y (t) E [n2T(n2 — 1, tini,$) - n2T(n2, 4,4) s )] 11 1 T (n1, 8 1 0

ni ,n2

ry(t) E n i T(n2 ,tin i ,$)T(n i ,s10,0)

ni ,n2

'y(t) E n177 (74,810,0) = ry(t)E(N(s))

= ''y(t)pc(s, 0) .

In the second step we have performed the shift n2 n2 + 1 of the summation

variable n2 in the gain term of the master equation. Hence, on using the initial
condition E(N (s)N (s)) = p(s , 0) 2 + p(s , 0) we find the following expression for
the two-time correlation function of the non-homogeneous Poisson process,

E(N (ON (s)) = p(t, p(s 0) + p(s , 0). (1.111)

 MARKOV PROCESSES 27

1.4.4.3 Increments of the Poisson process It will be of interest for later appli-
cations to derive the properties of the increment dN(t) of the non-homogeneous
Poisson process. For any time increment dt > 0 the corresponding increment of
the Poisson process is defined by

dN(t) N(t + dt) — N(t), dt > O. (1.112)

Note that we do not assume at this point that dt is infinitesimally small. The
probability distribution of the increment dN(t) may be found from the relation
CO

p(dN(t) = dn) ET(rt + drt,t + dtin,t)P(n,t). (1.113)

n=0

Since the process is homogeneous in space the sum over n can immediately be
carried out and using eqn (1.105) we find

[p(t + dt,t)rin -ii(t±dt,t)

p(dN(t) dn) = T (dn,t + dtiO,t) e (1.114)
dn!
where dn = 0,1, 2, .... The first and second moments of the Poisson increment
are therefore given by

E(dN(t)) = tt(t + dt,t), (1.115)

E(dN(t) 2 ) = ti(t + dt,t) + ti(t + dt,t)2 . (1.116)

If we now take dt to be infinitesimally small we see that, apart from terms of

order 0(dt2 ),

E(dN(t) 2 ) =E(dN(t)) = -y(t)dt. (1.117)

In addition, we also observe that this last relation becomes a deterministic re-
lation in the limit dt 0. Namely, according to eqn (1.114) the probability for
the event dN(t) > 2 is a term of order dt2 , and, consequently

p [dN(t) = dN(0] — 1 + 0(dt2 ). (1.118)

Thus, apart from terms of order dt2 , the equation

dN(t) 2 = dN(t) (1.119)

holds with probability 1 and can be regarded as a deterministic relation in that

limit
These results are easily generalized to the case of a multivariate Poisson pro-
cess N(t) = (N1 (t), N2 (t),... ,Nd (t)). We take the components Ni (t) to be sta-
tistically independent, non-homogeneous Poisson processes with corresponding
rates ryi (t). It follows that the relations
28 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

E(dNi (t)) = -yi(t)dt,

dNi(t)dNi(t) =Si i dNi(t),
hold in the limit dt 0, where terms of order 0(dt2 ) have been neglected.
These important results will be useful later on when we construct the stochastic
differential equation governing a piecewise deterministic process.

1.4.5 Diffusion processes and Fokker-Planck equation

Up to now we have considered two types of stochastic Markov processes whose
sample paths are either smooth solutions of a differential equation or else discon-
tinuous paths which are broken by instantaneous jumps. It can be shown that
the realizations of a Markov process are continuous with probability one if
1
lim f dx 7 -1 (x,t + Atlx i ,t) = 0 (1.122)
At—>t3 At

for any E > 0. This equation implies that the probability for a transition during
At whose size is larger than E decreases more rapidly than At as At goes to zero.
Of course, deterministic processes must fulfil this continuity condition. In
fact, we have for a deterministic process (0 denotes the unit step function)
1 f
hm — dxT(x,t + Atlx` ,t)
At—A) At

= llin
At—r0 At
f dx6(x - it'At (X 1 ))

11111 --9(1X 1 — 4) pt (X 1 )1 — E)
At—H) At
= 0. (1.123)

In the case of a jump process the continuity condition is clearly violated if the
jump rate W(xlxi, t) allows jumps of size larger than some E > 0,
1
dxT(x,t + Atlx` ,t) =
B.111 —
At—A) At
jx- x' >6
f
Ix — x'1> e
W (XIX / ,t) >0. (1.124)

There exists, however, a further class of stochastic Markov processes, the

diffusion processes, which are not deterministic processes but also satisfy the
continuity condition. We derive the differential Chapman-Kolmgorov equation
for a multivariate diffusion process X (t) = (Xi (t),X2(t),... ,Xd(t)) by investi-
gating a certain limit of a jump process which is described by the master equation
(1.86). To this end, we write the transition rate as

W (xix' , t) = f (x` ,y,t), (1.125)

 MARKOV PROCESSES 29

where y = x - , that is as a function f of the starting point x', of the jump

increment y, and of time t. Inserting (1.125) into the master equation (1.89) we
find
a
— p(x , t) = f dy f (x - y, y, t)p(x - y t) - p(x, t) f dy f (x, y, t). (1.126)
at

The fundamental assumption is now that f (xi , y, t) varies smoothly with x', but
that it is a function of y which is sharply peaked around y O. In addition it is
assumed that p(x ,t) varies only slowly with x on scales of the order of the width
of f. This enables one to expand the gain term f (x - y, y, t)p(x - y, t) to second
order in y,

a
— p(x,t) =f dy f (x, y, t)p(x, t) - f dy y a [f (x, y, t)p(x t)] (1.127)
axi
a2
f dy 21- yiy3• OXiaXi [f (x, y, t)p(x, t)] - p(x , t) f dy f (x, y, t)

The indices i, j label the different components of the multivariate process and
a summation over repeated indices is understood. Taking fully into account the
strong dependence of f (x' , y, t) on y, we have not expanded with respect to the
y-dependence of the second argument of f (x - y, y, t). We see that the first term
on the right-hand side cancels the loss term. Thus, introducing the first and the
second moment of the jump distribution as

gi (x , t) f dy yif (x , y t), D ii (x, t) f dy yiyi f (x , y, t), (1.128)

we finally arrive at the differential Chapman-Kolmgorov equation for a diffusion

process,

a
— p(x ,t) = -
a [g(x
i , t)p(x , +
a2
[Dii (x,t)p(x,t)j. (1.129)
at oxi 2 OXiaXi
This is the famous Fokker-Planck equation for a diffusion process. A formally
identical equation holds for the propagator of the process.
Again, this equation admits an obvious interpretation as a continuity equa-
tion for the probability density if we rewrite it as

(1.130)
at42 p(x, t) +ax,
— Ji (x,t) o,

where we have introduced the probability current density

1
Ji(X,t) gi (x , t)p(x, -- [Dii (x , t)p(x ,t)] . ( 1. 131)
2 ax3
30 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

FIG. 1.6. Sample path of the Wiener process.

As we already know, the first term on the right-hand side of the Fokker-Planck
equation describes a deterministic drift which corresponds to the differential
equation with vector field g,

dt x (t) =--- g(x(t),t). (1.132)

The second term on the right-hand side of the Fokker-Planck equation describes
the diffusion of the stochastic variable X(t). According to its definition the matrix
D(x, t), known as the diffusion matrix, is symmetric and positive semidefinite.
The most prominent example of a diffusion process is obtained by considering
a one-dimensional diffusion process, setting the drift equal to zero, g(x, t) E. 0,
and by taking D(x,t) =7, 1. This leads to the Gaussian propagator

T(x,tlx` ,t I ) =
1
exp
(x - 42 ) (1.133)
V27r(t -

showing that the process is both homogeneous in time and in space. This process
is often referred to as a Brownian motion process. If we further take the initial
density

p(x,t = 0) = 8(x) (1.134)

the process X(t) becomes the famous Wiener process W (t).

Another important example of a diffusion process is obtained by adding to
a Brownian motion process with constant diffusion coefficient D a deterministic
drift corresponding to the differential equation dxidt = -kx with a constant
coefficient k> O. The corresponding propagator then takes the form
 MARKOV PROCESSES 31

2 4 5

FIG. 1.7. Sample path of the Ornstein—Uhlenbeck process.

k(x — 42)
T(x,tlx' ,t') D [1 _ e-2k(t—t,)] exp (1.135)
D [1 _ e -2k(t—til

and the stationary first-order distribution is given by

kx2
p(x) = exp ( --) . (1.136)
7r
D D
This defines the Ornstein Uhlenbeck process. Its physical significance is due to
—

the fact that, up to trivial linear transformations, it is essentially the only process
which is stationary, Gaussian and Markovian. This assertion is known as Doob's
theorem. A Gaussian process is defined to be a process whose joint probability
distributions are Gaussian functions. The only other process with the above three
properties is the so-called completely random process, for which all m-th order
joint probability distributions are products of first-order Gaussian distributions.
We show in Figs. 1.6 and 1.7 realizations of the Wiener and of the Ornstein—
Uhlenbeck process.
With the help of the explicit expression (1.133) for the conditional transition
probability of the Wiener process we can calculate all relevant quantities of
interest. For example we find for the mean and the variance of the Wiener process

E(W (t)) = f dw wT(w,t10,0) = 0, (1.137)

E(W (0 2 ) =f dw w 2 T(w,t10,0) t. (1.138)

The second relation shows the well-known linear increase of the variance with
time t. The two-time correlation function of the Wiener process will also be of
interest. It is defined by
32 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

E(W(t)W(s) ) = E(W (s)W(t)) (1.139)

f dw2 f dwi w2wiT(w2, tlwi,$)T(wi,s10, 0),

where the second equation presupposed that t > s. The integrals can easily be
evaluated with the help of (1.133). The result is

E(W (t)W (s)) = min(t, s), (1.140)

which holds for t> s and for t < s.

On using these results we can also study the properties of the increment
dW(t) of the Wiener process which is defined for any dt > 0 by
dW (t) W (t + dt) — W (t). (1.141)

From eqns (1.137) and (1.140) we get

E [dW (0] = 0, (1.142)

E [dW(t) 2 ] = dt. (1.143)

For a multivariate Wiener process with statistically independent components

W1 (t) we have the important formulae:
E [dWi (t)] --- 0, (1.144)
E [dWi(t)dWi (t)] = 6 1 dt, (1.145)
E [(dW 1.(t)) 2/1 = 0(dtk ), (1.146)
E RdW1.(t)) 2k +1 1 = 0. (1.147)

The last two relations hold for all positive integers k, and may be shown directly
using the Gaussian propagator for the Wiener process.

1.5 Piecewise deterministic processes

We are now in a position to introduce a certain type of stochastic process which
plays an important rôle in this book, namely the so-called piecewise determin-
istic processes (PDPs). Essentially, such a process is obtained by combining a
deterministic time-evolution with a jump process.
Piecewise deterministic processes have a large variety of applications in the
natural sciences and technology. A number of examples and a mathematically
rigorous treatment of PDPs may be found in the book by Davis (1993). In the
theory of open quantum systems PDPs provide the appropriate mathematical
concept to describe the evolution of the state of a continuously monitored system
(see Chapter 6). We shall present in this section three different mathematical
formulations of a piecewise deterministic process, namely the Liouville master
equation for its propagator, the path integral representation, and the stochastic
differential equation for the random variable.
 PIECEWISE DETERMINISTIC PROCESSES 33

1.5.1. The Liouville master equation

The sample paths of a PDP consist of smooth deterministic pieces which are
given by the solution of some differential equation and which are interrupted by
instantaneous jumps. On the basis of our previous results it is now easy to write
down an appropriate expression for the short-time behaviour of the propagator
of a PDP,

T (x,t + Atlx' ,t) = (1 -1 1 (x' ) At) 6 (x - x' - g(x')At)

+W(xix')At + 0(At 2 ). (1.148)

The 6-function represents the short-time behaviour which is characteristic of a

deterministic process and which is obtained by expanding the flow of the deter-
ministic differential equation (1.75),

4'At(x 1 ) = x' + g(x l )At + 0(At 2 ). (1.149)

f
The prefactor of the 6-function involves the total rate F(x1 ) = dxW(x1x 1 ) for
jumps out of the state x'. The factor (1 — F(x')At) is therefore the probability
that no jump occurs in the interval At.
The second term on the right-hand side of eqn (1.148) is the probability for
a jump from x' to x within time At. Note that the conditions (1.58) and (1.59)
are satisfied. Note further that we consider here, for simplicity, only processes
which are homogeneous in time since the generalization to the inhomogeneous
case is obvious.
On account of the above short-time behaviour we can immediately write down
a differential Chapman-Kolmogorov equation for the propagator,
0 0
— T(x, tlx`, t i ) = - — [gi(x)T (x , tix 1 , t')] (1.150)
Ot axi
+ f dx" [W(xlx")T(x",tlx' ,e) - W(x 11 1x)T(x,tlx` ,e)].

The first term on the right describes the deterministic drift, whereas the sec-
ond part leads to the jumps x' .- x with rate W(x1x 1 ). Accordingly, we call
eqn (1.150) the Liouville master equation.

1.5.2 Waiting time distribution and sample paths

Let us determine a central quantity for the description of PDPs, i.e. the waiting
time distribution F(T Ix', t'). This quantity is defined to be the probability for
the next jump to occur during the time interval [t', t` + T under the condition
]

that we are in the state x' at time t'. To find this quantity we observe that

Fer ± &Ix', t') - F (rix' , t')=

_-.-- dF erix' , t') (1.151)

is just the probability for the next jump to occur in the time interval [t` + r, tl ±
T ± dri. If we divide this probability by the probability 1 — F(7-lx', t') that no
34 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

jump occurred in the previous interval from t' to t' + T, we get the conditional
probability for a jump over the interval dr. The latter quantity must be equal
to dr times the total rate F(C-(x')) for a jump out of the state we are presently
in, namely 4.7 (x') (since, if the jump occurs at time t1 + T, the deterministic
evolution has carried us from the state x' to C-(x')). Hence we have

dF(rixi, t')
( = F(IbT (x 1 ))dr. (1.152)
1.152)
1— F(rixi, V)

This equation shows that F(rix', t') does not, in fact, depend on the time t' and
we shall omit this argument in the following (note however that for a PDP which
is not homogeneous in time, e.g. for a non-autonomous differential equation, it
does depend on t').
The above relation leads to the differential equation

d
F(rlx`)] --,-- —1-1 ( 41-r(x 1 )) (1.153)
di-
ar

which is easily solved to yield

F(rIx') = 1 — exp [ — j. ds r(bs (x')) , (1.154)

where we have used the initial condition F(r ---- Olxi ) = O. Given that we are in
the state x', expression (1.154) is the cumulative distribution function for the
random waiting time T of the PDP.
For a pure jump process (g = 0) we simply have

F(rix') = 1 — exp [—F(x l )r] , (1.155)

demonstrating that for F(x') > 0 pure jump processes have an exponentially
distributed waiting time. For F(x') = 0 we have F:--_-_--. 0 which means that the
waiting time is infinite: The process never leaves the state x' which is called a
trapping state of the process.
For a true PDP (g 0) the waiting time distribution is, in general, not an
exponential function. As eqn (1.154) shows this is a direct consequence of the
deterministic time evolution between the jumps as a result of which the total
transition rate can change in a non-trivial way. Since F(x') is non-negative the
function
T

I(TIX I ) =--- f ds FeI.,(4) (1.156)

o
increases monotonically with T. Thus, we can basically distinguish two cases:
 PIECEWISE DETERMINISTIC PROCESSES 35

1. If the function /(-7- Ix') is not bounded it follows that

lirn /(7- 1x1 ) = +co, (1.157)

7--)00

and, therefore, we have

lirn F(7- Ix i ) = 1. (1.158)

T ->00

This implies that the process jumps with certainty after some time. One
can obtain a realization of the waiting time T for example by making use
of the inversion method. To this end, one first draws a random number r/
which is uniformly distributed over the interval [0, 1]. The random waiting
time may then be determined by solving the implicit equation

F(7-1x 1 ) = n (1.159)

for T (see Fig. 1.8).

2. If, on the other hand, the function /(7- 1x1 ) is bounded it follows that the
limit

liln /(TIX 1 ) = /(001X 1 ) < OC (1.160)

T -}00

exists. In this case we find

lim F(rixi ) = 1 — exp [—/(Doix i )] 2 1 — q < 1. (1.161)

7 - 00

This means that with a finite probability q, where 0 < q < 1, the process no
longer jumps. The quantity q is called the defect. Formally, this situation
can be dealt with by defining the state space of the random waiting time
to be 14 U {oc}. Then q is the probability for the event T -77-- co. Again a
realization of T can be obtained with the help of the inversion method (see
Fig. 1.8). Having drawn a uniformly distributed random number 17 E [0,1]
one first decides whether 7) > 1 — q or 7) < 1 — q. In the former case one
sets T z--- co, in the latter case T is to be determined by eqn (1.159).
We have just described how to determine a realization of the random time
intervals T between the jumps of a PDP. The above procedure allows us to design
a simple algorithm for the generation of a sample path x(t) of the PDP:
1. Assume that at some time t o the realization has reached the state xo
through the preceding jump. In the case that t o is the initial time, xo
must be drawn from the initial probability density p(x o , to ). In any case
we have

x(to) = zo. (1.162)

36 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

= + CO

/ T2 7
.

FIG. 1.8. The figure shows a typical waiting-time distribution F(Tlx`) with de-
fect q and illustrates how a realization of 7 is obtained by the inversion
method. ri i and ri 2 are realizations of a uniform random number 77 E [0, 1].
For ni the corresponding realization of the waiting time is infinite, T1 = +co,
while for 772 the realization of the waiting time is equal to T2.

2. Draw a random waiting time T from the distribution function F(rIxo ) by

employing the procedure described above. In the time interval from t o to
to + T the realization x(t) then follows the solution of the deterministic
equation, that is,

x (to + s) = 4, 3 (xo ) (1.163)

for 0 < s < T .
3. At time to + T the sample path performs an instantaneous jump into some
state z which is a random variable to be drawn from the probability density

W(z1x (to + T))

Q(z1x(to ± T)) = F(x(t o + (1.164)
T)) •

Note that Q(z1x) is just the conditional probability density for a jump into
the state z, given that we already know that a jump takes place out of the
state x. Note further that Q(z1x) is correctly normalized as

f dzQ(z 1x) = 1. (1.165)

Finally, we set
 PIECEWISE DETERMINISTIC PROCESSES 37

X(to ± T) = Z (1.166)

and proceed with the first step above.

This algorithm can easily be translated into a numerical computer program.
Using such a program we can generate a sample of realizations of the process
and then estimate all quantities of interest as sample averages. This is the essence
of the stochastic simulation technique, which will be illustrated with the help of
a number of examples in Chapters 6, 7 and 8.

1.5.3 Path integral representation ofPDPs

The above algorithm for the generation of the sample paths of a PDP yields
the structure of the possible paths of the process and the probability of their
occurrence. In mathematical terms this means that the propagator T(x, tlx', t')
of the PDP may be represented in terms of those sample paths that start in x'
at time t' and end in x at time t. Each of these paths contributes with a certain
probability, and summing over all paths with an appropriate weighting we get
the propagator. This path integral representation of PDPs will be derived in the
present section.
Up to now we have characterized a PDP by its differential Chapman—Kolmo-
gorov equation (1.150), that is by its short-time asymptotics. There is, however,
an equivalent integral representation for the propagtor of a PDP which is given
by

T( x, tlx ', t') =--- T(°) (x, tlx' , t') (1.167)

t
+ f ds f dy f dz TO ) (x, tly, , s)W (yiz)T (z , six' ,e).
t,
Here, we have introduced the quantity

T ° (x, tix', t') = [1 — F(t — ti lx 1 )] 8 (x — (Dt—ti (x')) (1.168)

which is the conditional probability density for the process to reach the state
x at time t without any jump, given that it was in x' at time t'. As is easily
checked, the solutions of (1.167) fulfil eqns (1.58) and (1.59). We may also verify
by an explicit calculation that the solutions of (1.167) satisfy the differential
Chapman—Kolmogorov equation (1.150).
We will not go into the details of these calculations since (1.167) allows a
simple interpretation (see Fig. 1.9). The propagator is written as a sum of two
terms, corresponding to two possibilities for the process to proceed from (z', t')
to (z, t): The first possibility is that no jump occurs in the time interval from t'
to t. The corresponding conditional probability density is given by the quantity
T(°)(x, tlx`, t') introduced above. The second possibility is that at least one jump
occurs. The corresponding contribution is the second term on the right-hand
38 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

time time

+
.

space

FIG. 1.9. Graphical representation of the Kolmogorov forward equation (1.167).

side of eqn (1.167). The variable s denotes the instant of the last jump. 3 Then
T(z, six` ,e) is the probability to be in some state z immediately before the jump,
and W (yiz)ds is the probability for a jump from z to y during ds. Multiplying
this by 70) (x,tly,$) we get the probability that the process reaches the state z
at time s, that it then jumps into the state y, and that it reaches the state x in
the remaining time interval without any further jumps. The total contribution
to T(x, tlxi,e) is found by integrating over all possible jump times s and over
all intermediate states z and y.
Equation (1.167) is the Kolmogorov forward equation. It is an integral rep-
resentation for the propagator of a PDP which directly leads to the path in-
tegral representation. Our above interpretation makes it natural to decompose
the propagator of the process into separate terms T (N) (x, tlxi, t'), each of which
represents the contribution from those paths with exactly N jumps:

00

T(x,tlx' ,e) = E ENT(N) ( x,tixi,e). (1.169)

N=0

The first term in this sum has already been defined above. We have introduced
a formal expansion parameter E which will be set equal to 1 afterwards. We
introduce this parameter also into the integral equation by regarding the rates
W(Y1z) formally as quantities of order E. In this way the decomposition (1.169)
becomes a perturbative expansion in powers of the transition rates.
If we insert eqn (1.169) into eqn (1.167) and collect terms of the same order
in E we find the following recursion relation which is valid for N > 1,

3 We exclude the possibility that an infinite number of jumps accumulates at a point of the
time interval.
 PIECEWISE DETERMINISTIC PROCESSES 39

t
T (N) (x , tlx' , t') = f ds f dy f dz TO ) (x , tly, , s)W (y1z)T (N —1) (z , six' ,t`).
t,
(1.170)

On iterating this equation one obtains T( N) (x, ti xi, t') for all N. Substituting the
result into (1.169) and setting E = 1 we finally get the path integral representation
of the propagator,

T (x ,tlx i ,e) = T(°) (x ,tix t ,t1 ) (1.171)

CO

N=1 t,
t

+ E f ds, f ds,_, ... S

SN

ti
i
ti
de5 1 f dyNdzN f dyN—idzN—i• • f dyidzi

TM (x, tlyN,sN)W(YNIzN)T M (zN,sNlYN-1,sN-1)

xW(yN_1izN_..1)7 0) (zN_1,sN—ilYN-2,sN_2)... W(yi Izi)T ( Nzi,si Ix', t').

We see that the sum over paths involves a sum over the number N of jumps,
a multiple integral over all intermediate jump times 81 < s2 < ... < sN, and
a multiple integral over the intermediate states yi , zi , y2, z2, ... , yN, zN. Here,
zu is the state just before the v-th jump, whereas y u is the state immediately
after the v-th jump. Finally, the integrand in eqn (1.171) is the corresponding
probability for a specific path given by the number N, the 8, and the y u , zu .

1.5.4 Stochastic calculus for PDPs

In the preceding sections we have formulated the dynamics of a stochastic Markov
process X(t) with the help of appropriate time evolution equations for the condi-
tional propagator of the process. Alternatively, one can also describe the process
in terms of an evolution equation for the random variables itself. Of course, such
an equation is, in general, not a deterministic differential equation, but must be
an equation of motion involving random coefficients, that is, a stochastic differ-
ential equation. This way of describing a stochastic process is well known for the
case of diffusion processes. We first review briefly this case as a motivation for
our discussion of piecewise deterministic processes.

1.5.4A Itei calculus for diffusion processes Consider a multivariate diffusion

process X(t) = (Xl (t) , X2 (t) , . . . , X d (t)) which is governed by the Fokker—
Planck equation (1.129). Taking an arbitrary differentiable scalar function f
one finds for the expectation value of f(X(t)) by employing the Fokker—Planck
equation

0
2̀ E [f(X(t))] = f dx f (x) — p(x , t)
at at
0f 1 a2 f
=E[ gi (X (t)) ± D j3 (X (t))1 . (1.172)
OXi 2 OX iDX i
40 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

By use of the increment dX(t) = X (t + dt) X(t) of the process we may rewrite
this as follows,

E [f (X (t) + dX (t)) — f ((X (t))] (1.173)

af 82 f
E[ g (X (t)) + 13,3 (X (t))] dt + 0 (dt2 ).
axi 2 aXiaXi

Now, let us take f to be one of the coordinate functions, f(x) = x i . This

yields

E [dXi (t)] E [gi (X (0)] dt + 0 (dt2 ) (1.174)

To leading order the expectation value of the increment dXi(t) is proportional to

dt, where the coefficient is given by the expectation value of gi (X (t)) describing
the deterministic drift. Next, we take f (x) to be the function f (x) = x,xj for a
fixed pair of indices (i, j). The result is then

E [dXi (t)dX i (0] = E [Dii (X (t))] dt + 0 (dt2 ). (1.175)

For small dt the covariance matrix of the increments dX i(t) is given by dt times
the expectation of the diffusion matrix.
The important point to notice is that the behaviour of the process expressed
by eqns (1.174) and (1.175) can be reproduced by means of an appropriate linear
combination of Wiener increments dWi (t), j = 1, 2, , d. These increments are
supposed to be mutually independent and independent of X (t) . If we set

dX- i (t) 7,- gi (X (t))dt + (X (t))dW i (t), (1.176)

with an appropriate matrix B ii (x), we find with the help of eqns (1.144) and
(1.145) that

E [dX- JO] = E [gi (X (0)] dt,

and

E [dk- (t)d±i (t)] = E [Do (X (t) )] dt + 0 (dt2 ).

The last equation holds provided the matrix B ii (x) is related to the diffusion
matrix D 1 (x) by (remember that the summation convention is in force)

Rik (X)Bik (X) (x). (1.179)

This shows that the mean and the covariance of the increments dX i (t) and
dX i (t) coincide to order dt. Therefore we see that on small time scales the in-
crements dX i (t) of the diffusion process behave as the increments dX- i (t) which
are obtained by adding an appropriate linear combination of Wiener increments
 PIECEWISE DETERMINISTIC PROCESSES 41

to the deterministic drift. Thus, one expects that the increments dX i (t) obey a
stochastic differential equation of the form

dX i(t) = gi (X (t))dt + Bi i (X (t))dWi (t). (1.180)

It can be shown that the diffusion process X(t), which was defined originally
in terms of the Fokker—Planck equation for its conditional propagator, is indeed
equivalent to the stochastic differential equation (1.180). A rigorous mathemati-
cal treatment of stochastic differential equations requires the introduction of the
concept of stochastic integration which enables the development of a stochastic
calculus for diffusion processes, known as Itô calculus. Equation (1.180) is then
to be interpreted as a stochastic differential equation in Itô form. We do not en-
ter here into a detailed discussion of the stochastic calculus, which is the subject
of several excellent textbooks (Gardiner, 1985; Doob, 1953; Kloeden and Platen,
1992; Arnold, 1974).
Let us, however, demonstrate that the increments dX- i(t) defined by eqn
(1.176) reproduce correctly, within the desired order, all expectation values of
the form (1.173) as they are determined through the Fokker—Planck equation.
To show this we have to determine the expectation value

E [f(X(t) + (t)) — f (X (0)1

= E [f (Xi(t) + gi (X (t))dt + Bii (X (t))dW (t)) — f (X (0)] . (1.181)

On using eqns (1.144)—(1.147) we get by expanding f around X(t) and disre-

garding terms of order dt2 ,

E [f(X(t) + (t)) — f (X (0)]

[ (9 2 f
E (gidt + Bii dWi) + B ik B idWkdWi + 0 (dt2 )
2 Ox i Oxi
[ f 1 02 f
=E (X (t)) + D z3 (X (0) dt + 0 (dt2 ). (1.182)
2 aXiaXi

This is seen to coincide within order dt with the expectation value (1.173) de-
termined by the Fokker—Planck equation, which proves our statement.
A similar calculation leads to the following formula for the increment of an
arbitrary function 0(X(t)) of the stochastic variable, known as the Itô formula,

00 1 02 0 00
d0(X (t)) = g i(X (t)) + D, (X (t))1i dt + B • (X (t))dW i (t).
°xi 2 aXiaXi aXi 3
(1.183)

We observe that this expression for the increment of 0 may be obtained directly
by expanding the difference d0(X(t)) = «X(t) + dX (t)) «X (t)) in powers of
—

dX(t) and by employing the following rules of the Itô calculus:

42 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

dWi(t)dWi (t) = ciiidt, (1.184)

dWi (t)dt = 0, (1.185)
[dWi(t)] k = 0, for k > 2. (1.186)

In summary, these rules can be stated as follows: (i) Pick up all terms of order
dt 1 /2 , and dt, whereby dWi(t) is treated as a quantity of order dt 1 / 2 , and (ii)
discard all terms of order dt312 and higher.

1.5.4.2 Itô calculus for PDPs Let us now develop an analogous calculus for
piecewise deterministic processes. To this end, we first consider a real-valued
PDP X(t) and again define the increments by dX(t) E X (t + dt) X (t), dt > 0.
—

Our aim is to construct, in analogy to the procedure of the foregoing subsection,

a stochastic differential equation for the PDP which is of the following form,

dX (t) = g(X (t))dt + dJ(X (t)). (1.187)

Obviously, the first term describes the deterministic drift of the PDP. The second
term dJ(X(t)) represents the jump part of the process. It is intuitively clear that
this term is not small in the usual sense as dt goes to zero, since X(t) undergoes
instantaneous jumps of finite size. This fact makes the stochastic calculus for
PDPs very different from ordinary calculus as we shall see below.
Our task is to find the statistical properties of the jump part dJ(X (t)). Let us
assume that the state X(t) is given. According to the Liouville master equation
(1.150) the system may perform an instantaneous jump from this given state to
some other state z with a rate W(z A(t)). If no jump occurs in the time interval
dt we must have dJ(x) = 0, of course, which means that only the deterministic
drift is present.
Consider now the quantity dNz (t)dz which is defined to be the number of
jumps from the given state X(t) into the volume element dz around z, taking
place during the interval dt. For small dt the average of the random number
dNz (t)dz must be equal to dt times the rate W(zIX(t))dz for jumping into the
element dz, that is,

E [dNz (t)] = W (z1X(t))dt. (1.188)

This equation shows that the expectation value on the left-hand side is a condi-
tional expectation, namely it is conditioned on the given state X(t).
If the random number dNz (t)dz takes on the value 1 we get the new state
X(t 4- dt) = z, that is the increment dX(t) is equal to z — X(t). Thus we are led
to the following ansatz for the stochastic jump term,

dJ(X (t)) = f (z — X (t))dN z (t)dz, (1.189)

such that the stochastic differential equation for the PDP can be written as
 PIECEWISE DETERMINISTIC PROCESSES 43

dX (t) = g(X (t))dt + f (z — X (t))dN z (t)dz. (1.190)

We have not yet fully specified the statistics of the random quantities dNz (t),
only their expectation values have been defined in (1.188). Since the PDP involves
a jump rate which increases linearly with dt one expects that for small dt at most
one jump can occur, that is, we expect that the relation

dNz , (t) • dNz (t) = .5(z' — z)dN z (t) (1.191)

holds.
According to eqn (1.191) the random numbers dNz (t) provide a field of inde-
pendent Poisson increments as may be seen as follows. We suppose that, starting
from the given state X(t), the process may perform a discrete set of transitions

X(t) -- z a = z a (X(t)) (1.192)

labelled by an index a. The rate for a particular transition a will be denoted

by '-ya (X (t)). This rate as well as the final state z a (X(t)) of the transition may
depend on the initial state X(t), of course. Hence we have

W(z1X(t)) = El ic, (X (t))6 [z — z a (X (0)] . (1.193)

a

We further take a disjoint partition of the state space into small elements Az a in
such a way that each z a is contained in precisely on Aza . The random quantities

dNa (t)E_-- f dN z (t)dz (1.194)

Az OE

therefore denote the number of jumps into the element Az a during the time
interval dt. By virtue of eqns (1.188) and (1.193) their expectation values are
given by

E(dNa (t)) = ,ya (X (t))dt. (1.195)

Using eqn (1.191) we also find

dNa (t)dNo(t) = CodNa (t). ( 1. 196)

This shows that the dNa (t) behave for small dt as independent Poisson incre-
ments (compare with eqns (1.120) and (1.121)). Finally, with these definitions
the stochastic differential equation (1.190) reduces to the particularly transpar-
ent form

dX (t) = g(X (t))dt + E {,(x(0) — X (t)} dN a (t). (1.197)

a

This equation has an obvious interpretation: In view of (1.196) either all dNa (t)
are zero, or dNa (t) = 1 for precisely one index a. In the former case X (t) follows
44 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

the deterministic drift, in the latter case X(t) performs a particular jump given
by the index Q.
Equations (1.190) or (1.197) provide the desired formulation of a PDP in
terms of a stochastic differential equation. Let us check that they lead to the
correct dynamical behaviour as dictated by the Liouville master equation (1.150).
From the latter we find for the expectation value of the increment of any function
f

,E rofx
E [f (X (t) + dX (t)) — f (X (t))]

g (X (t))
(1.198)

+ f If (z) — f (x (t))1W (z1X (t))dz] dt + 0 (dt2 ) .

Our aim is to show that the same equation is obtained if one substitutes the
right-hand side of eqn (1.190) for dX (t) and uses the properties (1.188) and
(1.191). To this end we first determine the powers of the jump part dJ(t) in the
stochastic differential equation. Invoking eqn (1.191) it is easy to verify that for
all k =-- 1 5 2, 3, . . . we have

[dJ (X (t))] k =f z { — X (t)} k d1V,(t)dz. (1.199)

This means that, after averaging, all powers of dJ (X (t)) are of the same order
dt. Consider now the expectation value

E , E [f (X (t) + g (X (t))dt + dJ (X (t)))

— — f (X (0)] (1.200)

which must be demonstrated to be equal to the right-hand side of eqn (1.198)

up to terms of order dt2 . This may be achieved by first expanding with respect
to the drift term, keeping the jump term to all orders, which leads to

0 f

E=E[ - (X + dJ)gdt + f (X + dJ) — f (X)1+ 0(dt 2 ). (1.201)

Ox

In the first term the derivative of f is evaluated at the point X + dJ. However,
since this term is already of order dt we can evaluate it at the point X, thereby
making only an error of order dt2 , due to the above property of dJ. Hence we
write

Of

E --,--- E [ gdt + f (X + dJ) — f (X)1+ 0(dt 2 ). (1.202)

Ox

The second term must be treated, however, in an entirely different way, since, on
averaging, all powers of dJ are of the same order dt, that is, we must evaluate the
difference f (X + dJ) f (X) exactly. This is obviously due to the fact that the
—

process performs instantaneous jumps of finite size, such that any approximation
 LEVY PROCESSES 45

of this difference by a finite Taylor expansion fails. Hence we write, denoting by

f( k) the k-th derivative of f,

9 i -i f (k) (X)
E [f (X + c1,1) – f (X)] .,--_. E rÉ [del
k=1 .

Er É) il-
k=1
f (k) (X) f {z — X} k dl Vz (t)dd

E [f If (z) – f (X)} dINT,(t)dzi

E [f Or (z) - f(x)}w(z1x)dddt. (1.203)

Thus we have

Of

E=E[ — g+ f If (z) – f (X)} W (z1X)dzi dt + 0(dt 2 ), (1.204)

Ox

which coincides with the right-hand side of (1.198) and proves that the stochastic
differential equation (1.190) correctly reproduces the expectation values of the
increments of all functions f.
Finally we note that stochastic differential equations similar to (1.190) or
(1.197) also hold in the case of a multivariate process. Explicitly, we may write

dXi(t) = gi (X (t))dt + f { zi - Xi (t)} dN z (t)dz (1.205)

for the stochastic differential equation of a multivariate PDP with components

X i (t).

1.6 Lévy processes

In the preceding sections our investigation of Markov processes was based mainly
on the differential Chapman—Kolmogorov equation. An alternative approach is
to search directly for solutions of the corresponding integral equation. Such a
strategy is indeed possible if certain symmetry and invariance properties are
imposed. An example is the requirement that the process is both homogeneous
in space and time which leads to the important class of Lévy processes (Itô,
1993; Bertoin, 1996). As will be discussed in this section it is possible to give
a complete characterization of this class of stochastic processes. Lévy processes
have a large number of interesting applications (Klafter, Shlesinger and Zumofen,
1996). An example will be given in Chapter 8, where it is shown that certain
stable Lévy distributions arise in the stochastic representation of the dynamics
of open quantum systems.
46 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

1.6.1 Translation invariant processes

A Lévy process is defined to be a stochastic process X(t), t > 0, which is both
homogeneous in space and in time. For simplicity we assume that the process is
real valued (the generalization to d dimensions is straightforward), and that the
process starts at time to = 0 at the point X(0) = 0. The propagator of such a
process depends only on the differences of its time and space arguments,

T(x,tlx i ,t1 ) = Tt _ t ,(x — x`), (1.206)

which means that the propagator is invariant with respect to space-time trans-
lations.
Alternatively, translational invariance may be formulated as follows. We take
any ordered set of times,

to E0< t1 <t2 < . . . < trn . (1.207)

Spatial homogeneity then implies that the m random variables

X(t) — X(t u _ i ), I. = 1, 2,... , m, (1.208)

are mutually independent. Such a process is also called an additive process, or
a process with independent increments. Homogeneity in time means that the
random variable X(t + s) — X (ti + s), with t > t', follows the same distribution
as the random variable X(t) — X(e) for all s > 0. Such processes are called
processes with stationary increments. A Lévy process is thus a stochastic process
with independent and stationary increments.
Employing translational invariance we write the Chapman—Kolmogorov equa-
tion as follows,

T+(y) = f 41 Tt (y — Y i )Tti(Y 1 ), (1.209)

which means that the propagator over the time t +t' is equal to the convolution
of the propagators over the times t and t'. It is thus advantageous to consider
the characteristic function

G(k,t) E E ikx(t)] = f dy e iky Tt(y), (1.210)

[ e

such that the Chapman—Kolmogorov equation leads to

G (k , t 1 + t2) = G (k ,ti)G(k , t2) . (1.211)

This equation shows that the logarithm ln G (k , t) of the characteristic function
(the cumulant generating function) is linear in t and that we can write the
characteristic function as

G(k,t) = exp [011(k)] . (1.212)

The quantity 41(k) is called the characteristic exponent of the process.
 LEVY PROCESSES 47

The characteristic exponent 41 ( k ) is directly related to the generator A of the

Lévy process. We write the propagator as

dk
Tt(y) = f— e -ik vG(k,t)
27r
dk
f 27r e
—
(1+ t4j(k)) 0(t 2 )

dk .k
(5(y) + t f — e' "11(k) 0(t2 ). (1.213)
27r

Hence we get by the definition of the generator

1
Ap(x) = f dy [Tt (y) — S(y)] p(x — y)

f dy[f Trdk e'. k Y41(k)] p(x — y)

(1.214)
f dyA(y)p(x — y).

Thus we observe that the integral kernel A(y) of the generator is equal to the
Fourier transform of the characteristic exponent 41(k).

1.6.2 The Lévy Khintchine formula

-

According to eqn (1.211) the characteristic function G (k , t) is the m-th power

of G(k, t/m). This implies that the propagator over time t is equal to the m-th
convolution power of the propagator over time t/m, that is,

Tt = Tthn * Ton * • • • * Tt/m) (1.215)

where * denotes the convolution and the right-hand side contains m factors. A
distribution with this property is called infinitely divisible. One can see this more
directly as follows. We set t„ = vtlm for v = 0, 1, 2, ... , m. By means of the
identity

X(t) = E (x(t) _ x(t,)) , (1.216)

u=1

the random variable X(t) is represented as a sum of m independent and identi-

cally distributed random numbers. This is precisely the definition for a random
number X(t) (and its corresponding distribution Te) to be infinitely divisible.
It is possible to give a complete characterization of all possible Lévy processes
by making use of the property that their propagators are infinitely divisible. This
is achieved by a fundamental theorem which is due to Lévy and Khintchine.
48 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

Applied to the present case this theorem asserts that the most general form of
the characteristic exponent of a Lévy process is given by the following equation:

4.1 ( k ) = igk - —
D k2 + f [e "i 1 - ikY 0 (h ly1)]W (04. (1.217)
2
Here, g and D are real constants with D > 0. 0(x) is the unit step function such
that
1\
0(h -
f 1, II < (1.218)
y > h.

This function cuts off the last term in the integral over y at some length scale
h, the significance of which will be discussed below. Finally, W(y) is some non-
negative measure, the Lévy measure, defined on R {0 } with the properties

if y 2 W(y)dy < oc,

Ivl<h (ody < 00.

fw
IvI>h

According to the Lévy-Khintchine theorem any translational invariant process

X(t) has a characteristic exponent 111(k) of the form (1.217). For a given length
scale h, the constants g and D, and the measure W(y) in this representation are
uniquely determined by the process.
To get a feeling for the Lévy-Khintchine representation let us consider some
simple examples. First we take W(y) 0. Then we have

D
41 ( k ) = igk - — k 2 , (1.221)
2
and, consequently,

D
G(k, t) = exp [(igk - k 2) t]. (1.222)

On Fourier transforming this characteristic function we immediately get

1
(1.223)
Tt(Y) 'N/2 7Dt exP 2Dt f

which is recognized as the propagator of a Gaussian diffusion process with con-

stant drift coefficient g (corresponding to a linear drift) and constant diffusion
coefficient D.
Of course, g and D must be constant since the process is assumed
to be homogeneous.
 LEVY PROCESSES 49

Now, consider the case g , D --,--- 0 and take the Lévy measure

W(y) ---- 'O W — yo), (1.224)

with some positive constant 1/ and yo > h. The characteristic exponent is then
found to be

41(k) = 1/ (eik Y° — 1) , (1.225)

such that

G(k,t) = exp ["yt (e ik Y° — 1)] . (1.226)

Thus, X(t) is a Poisson process with rate 'y and step size yo (compare with
eqn (1.99)) or, equivalently, a Poisson process N(t) with unit step size times MI>
i.e. X(t) -L--- yo N(t).
This example is readily generalized to the case that the Lévy measure is a
discrete sum of (5-functions,

W(y ) ---- E -ya(5(y — ya), (1.227)

where '-y, > 0 and y, > h. The characteristic function of the corresponding
process is in this case given by

G(k,t) = exp [E 'Tat (e — 1)1 = 1-1 exp ['Tat ( — 1)] . (1.228)

a a

The result is that the Lévy process X(t) is now a discrete sum of independent
Poisson processes with rates -y, and step sizes y,. Introducing the corresponding
independent Poisson processes Na (t) with unit step size we thus have

X(t) -_, E ya N,(t). (1.229)

a

Hence,

dx(t) , E yAlV a (t), (1.230)

a

and

E [dNa (t)] = "ya dt, dNa (t)dNi3(t) = 60dNa (t). (1.231)

This corresponds exactly to our former results (1.195) and (1.196) for the special
case of a PDP without drift and with Poisson increments dNc,(t) the statistics
of which does not depend on X(t), which is due to the translational invariance
50 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

of the process. Accordingly, the Lévy measure W(y) is equal to W(x + ylx) =
W(xlx y), which is just a discrete distribution of jump rates.
In the general case we expect the Lévy measure W(y) to represent a contin-
uous sum of independent Poisson processes indexed by their jump sizes y. This
can be seen directly from the general form of the generator A as follows with the
help of expression (1.214) and the Lévy-Khintchine formula:

Ap(x) = f dy f dk [i k_ k 2 ] e -iky P(.4 y) (1.232)

- 2 ' -

+ f dy f -
2
dk7r f dz [e ik z - 1 - ikz0(h - lz ) ] e-ik YW (z)p(x - y).

The first term is easily seen to be the generator of a linear diffusion process,
whereas the second term is determined as

J dy f dz [ 6 (z y) 6 (y) + z 0 (h izi) 6 (Y) —

aax 1W(z)P(x y)

f dz W(z)[p(x – z) – p(x) + zO(h – 1z1) :1 3 (x)] • ( 1 . 233)

Thus we find for the most general form of the generator of a Lévy process:

0 D 1
Ap(x) = [ p(x) (1.234)
–g Ox + 28x2

+ f dy W (y)[p(x - y) - p(x) + yO(h -

IY1) oax P(x)] •
Apart from the additional term under the integral which is proportional to 0(h -
II) this is just the generator of a process obtained by adding a jump process to
a Gaussian diffusion process. Of course, due to the requirement of translational
invariance of the process, drift and diffusion coefficient must be constant, and
the jump rate W(y) in the master equation part in eqn (1.234) only depends on
the difference y of the positions of the initial and final states.
What is the significance of the additional term y9(h II y) 8p/0x which obvi-
-

ously depends on the arbitrary length scale h? To answer this question we take a
closer look at the conditions (1.219) and (1.220) imposed on the Lévy measure.
These conditions control the behaviour of W(y) at zero and infinity. Up to now
in this section, and also in the preceding ones, we have always tacitly assumed
that the total transition rate

r= f dyw(y) (1.235)

is finite. In this case both conditions (1.219), (1.220) are satisfied for any h > 0.
But also the quantity flyi<h dy yW (y) is then finite and we may absorb the last
term in (1.234) into the drift coefficient, which yields
 LEVY PROCESSES 51

0 D 02 1
Ap(x) = [
g' Ox + 2 ax2J P(x) + f dYw(Y) [P(x — Y) (1.236)

where we have introduced a new drift coefficient by

g' = g — f dy yW(y). (1.237)

IYI<h \

Thus we see that for a finite total transition rate F the scale h is completely
arbitrary and the generator of the Lévy process takes on precisely the form
encountered before.
The important point to note is, however, that the conditions (1.219) and
(1.220) do not exclude the possibility that the total transition rate diverges as a
result of a singularity of W(y) at y = 0. The reason for this divergence is that
a singularity of the Lévy measure at y = 0 can lead to an accumulation of an
infinite number of jumps in a finite time interval. As an example take W(y) to
be proportional to I y. Such a measure decreases sufficiently rapidly for the
integral (1.220) to converge, whereas the singularity at y = 0 is weak enough to
satisfy also (1.219). But the total transition rate given by the integral (1.235)
clearly diverges. In spite of this, the generator (1.234) and the characteristic
exponent T(k) are well defined and finite.
Let us rewrite the general form (1.234) of the generator as follows,

0 D 02 1
A.p(x) = [— g
Ox + 2 Ox 2 ] P(x) + f clY ") [P(x — y) — P(x)]
IYI?h
+ f dyW (y) [p(x — y) — p(x) 4- y -(x)]. (1.238)
Iv1<t,
The second term on the right-hand side yields the big jumps of size ly1 > h,
while the third term involves the small jumps. If we apply the generator to a
smooth probability density p(x) which varies significantly on a typical length
scale L» h, we may expand the integrand in the third term around x to obtain
the expression
1f 2 a2
ox2 P(x).
j dY Y W(Y) --- (1.239)
IYI<h
By condition (1.219) this expression is, for any fixed h > 0, well defined and
finite. We thus see that the contribution (1.239) from the small jumps leads to
diffusion-type behaviour of the process on small length scales.

1.6.3 Stable Lévy processes

In the preceding subsection we have discussed the most general form of the gen-
erator of a Lévy process. A question that arises from the foregoing discussion is,
52 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

are there interesting Lévy measures for which the total transition rate diverges?
As we shall see below there are such measures the existence of which is closely
related to the notions of scale invariance and stability of the corresponding pro-
cesses.

1.6.3.1 Stable distributions and scaling relations A Lévy process X(t) is called
stable4 if it has the following scaling property. For any u > 0 there is a correspond-
ing A(u) > 0 such that the random number X(ut) follows the same distribution
as A(u)X(t). This means that a change of the time scale by a factor of u yields
the same distribution as a change of the length scale by an appropriate factor
Mu) (here and in the following we shall assume, to be definite, that the variable
X has the dimension of length). The stability of the process thus expresses a
certain scale invariance, which is often referred to as self-similarity. We write the
stability condition symbolically as

X (ut) --, A(u)X(t), (1.240)

whereby the symbol — serves to indicate that the left- and right-hand sides of
the expression follow the same distribution.
One can show that a Lévy process is stable if and only if the corresponding
propagator represents a stable distribution. Let us first define stable distribu-
tions. Consider some random number Z with distribution p(z), and mutually
independent copies Z 1 , Z2 , . . . Zi, , . . . of Z all with the same distribution p(z).
Then p(z) is defined to be a stable distribution if for each positive integer m
there exists a factor A(m) such that the sum an Zr, has the same distribution
,

as the random number A(m)Z,

M.

E z, , A(m)Z. (1.241)
v.i
Thus, a distribution p is stable if all m-th convolution powers of p are equal to
p up to an appropriate change of the scale. A prominent example for a stable
distribution is the normal distribution which obviously has this property with
the scaling factor A(m) = -On.
It is easy to show that the propagator Tt (y) of a stable Lévy process X(t)
must be a stable distribution: For any m we can write
771

x(nt) , E (x(vt) - x([v - lit)). (1.242)

,=1

The left-hand side in this equation has the same distribution as

whereas the right-hand side is the sum of m independent, identically distributed
random numbers X(ut) — X([v — 1]t) which follow the same distribution as

4 More precisely, our definition is that of stability in the strict sense.

 LEVY PROCESSES 53

X(t). This shows that the distribution of X(t), namely Tt (y) must be stable.
The converse is also true: If we take some stable distribution, then it yields the
propagator of a stable Lévy process.
Next we investigate what stability means for the characteristic exponent
kli(k). The relation X(ut) - A(u)X(t) leads to the scaling relation

Tut (Y) = (1.243)

A (1u) Tt ( A (Yu) ) \
for the propagator. The corresponding relation for the characteristic functions
becomes
1
‘11(k) =
41 (A(u)k).- (1.244)
u
If we look at the Lévy-Khintchine formula (1.217) we get the following obvious
solutions of the scaling condition (1.244): For D = W = 0 we have A(u) =
u which is the scaling for a linear drift, corresponding to the singular stable
distribution Tt (y) = ä(y - gt). For g = W = 0 one finds A(u) = \Fa which is
the scaling for the Brownian motion case, corresponding to the stable Gaussian
propagator
1 21
Tt(Y) ----' exp (1.245)
V27rDt [ ;Dd .
The question is now, are there non-trivial Lévy measures that lead to stable
Lévy processes? To answer this question we put D ,- - 0, since we already know
that D > 0 for Brownian motion. We keep however a non-vanishing g for reasons
that will become clear later on. On using the Lévy-Khintchine formula the scaling
condition (1.244) then takes the form

igk + f [e ikY - 1 — iky0(h - IYI)] 147 (MY (1.246)

= i (-A ) gk + f [eik9 - 1 1 fy
■
u
- ikyO(Ah - 10] uÀ W A clY.
This equation suggests that we can satisfy the scaling relation by taking Lévy
measures with the property
1 iy\
W(y) (1.247)
OW U) '
provided the scale dependence through the term 9(Alz - y) can be removed in
some way (it can!). Ignoring this term for a moment we are thus led to (1.247),
which, of course, could have been obtained directly by dimensional analysis, since
W(y) has the dimension (length-time) -1 . Thus we find the stable Lévy measures
C+y 1 c (1.248)
Wa(Y) = { ' '

C—IY1-1 al
54 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

with some non-negative constants C. These measures lead to the scaling relation

A(u) = u l /', (1.249)

that is we have X(ut) — u l /aX(t).

The possible values for the exponent a, the scaling exponent, are restricted by
the conditions (1.219) and (1.220) which must be satisfied by any Lévy measure.
Condition (1.219) leads to a < 2, while condition (1.220) yields a > 0, and
hence,

0 <ci <2. (1.250)

Recall that the case a -= 2, corresponding to the scaling relation A(u) = \ru, is
obtained for a non-vanishing diffusion coefficient D and by setting g = W = 0.
The above scaling relation has interesting consequences for a sequence of in-
dependent random numbers X, X 1 , X2,... ,X,..., following the corresponding
stable distribution. Consider the case 0 <ci < 1. Then we have
M

E xv,mvax,
v=1
(1.251)

with 1/a > 1. This shows that the sum grows faster than the number in of
terms in it. In other words, the maximal term of the sum is likely to become
very large and to dominate the value of the sum. It is clear that the central limit
theorem of probability theory is therefore violated which is related to the fact
that the stable distributions do not have a finite variance for a < 2. Indeed, it
can be shown that a stable distribution with finite variance must necessarily be
Gaussian (or concentrated at a single point). This is seen from the fact that the
density p(y) of a stable distribution with exponent a 2 behaves for large y as
-1 —ct
iYi
1.6.3.2 Characteristic exponents of stable Lévy processes Our analysis is not
yet complete since we have neglected above the scale-dependent term in eqn
(1.246) and since we do not yet know the characteristic exponents T a (k) of the
stable processes.
As we have seen, if one changes the length scale h to some other value h', the
corresponding change of the characteristic exponent *a
(k) can be compensated
by an appropriate shift of the drift coefficient which is given by

g -- g' T---- g + f dy yWc,(y)[0(1il

-
— II) — 0 (h — MA . (1.252)

Henceforth, this new coefficient g' will be called the renormalized drift coefficient.
Thus we see that the choice of the length scale is to a large extent arbitrary and
we can employ this freedom to find the characteristic exponents of the stable
distributions. We distinguish three cases.
 LEVY PROCESSES 55

1. Consider first the case 0 <Q < 1. In this case the integral

IYI<h
f cly ywa(y) (1.253)

converges. Thus, we can let le go to zero in eqn (1.252). We further put the
renormalized drift coefficient equal to zero, since a non-zero drift term would
violate the scaling relation (see eqn (1.246)). The characteristic exponent then
reduces to
+00
0 (k) = f (e lk Y — 1) 14 c,(y)dy

o
= C+ (e iky 1) dY
y i+c,
c_ f 1)
dy
(1.254)
iYi l+
- 00

The integrals can be determined explicitly and yield the final result for the
characteristic exponent of the a-stable distributions

(k) = ( Co + i — — C (1.255)
lkI

The new constants Co and C1 are related to the constants C± of the Lévy
measure through

Co = ( C+ + C_M—a) cos (72°)

— (1.256)

— (C+ — C_)F(—a) sin C+:E . (1.257)

Here F denotes the gamma function. As it should be, we have Co > 0 since
r(—a) < O. Co is the scale parameter and C1 the symmetry parameter, for
C1 = 0 the corresponding stable distribution is obviously symmetric.
2. Consider now the case 1 <Q <2. We cannot follow the above procedure
in the present case for the integral (1.253) now diverges. However, for 1 < a < 2
the integral

Y
f
I> h
cly ywa(y) (1.258)

is finite. This allows us to take the limit le oc in eqn (1.252). Putting again
the renormalized drift coefficient equal to zero we now have
56 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

0.5

0.4

0.3

0.2

0.1 \

••■• ■
\
\

-3 -2 -1 0 1 3 4
Y

FIG. 1.10. Plots of the propagators T1 (y) for three a-stable Lévy processes: the
Gaussian propagator (1.245) with a = 2 and D = 1 (dashed-dotted line), the
Cauchy propagator (1.263) with a = 1 and Co = 1 (dashed line), and the
propagator (1.264) corresponding to a = 1/2 (solid line).

+00

Ta (k)
f
- 00
( e iky _ 1 _ i k y) Wa( Y )dY (1.259)

+00
dy dy
C+ f (e 1 - iky) i+a + C_ f (ei " 1 iky)
1Y11+a
O - 00

Again the integrals can be evaluated explicitly and the result is precisely the same
as above, that is, we again get eqn (1.255) with the coefficients given by (1.256)
and (1.257). Note that again Co > 0 since now F(-û) > 0 and cos(7r0/2) <0.
3. The case a = 1 requires special treatment which involves a non-zero
drift coefficient. Note that the drift term satisfies the scaling condition (see eqn
(1.246)) since a = 1 gives A(u) = u. It can be shown that for the process to be
strictly stable the Lévy measure W1 (y) must be taken to be symmetric, Thus in
the special case a = 1 we write
Co 1
— , Co > O.
(Y) = (1.260)
IY12
The characteristic exponent is then given by
+00
iky dy
41 1(k) = iCi k + f (eik Y - 1 (1.261)
7r 1 + (y I h) 2 ) y2
- 00

Note that we have included here a finite drift term with coefficient C1 and that
we have changed the cutoff function from 0 (h - ly1) to the function 1/(1+ (y I h)2)
 REFERENCES 57

This is permissible since this change also amounts to a finite renormalization of

the drift coefficient. The integral in eqn (1.261) can be evaluated explicitly using
the method of residues which yields the h-independent value —7r1kl. As a result
one finds the same formula as for the other cases, namely

k
4, i (k) = iCik — Co* = ( — Co + i l C].) ikl. (1.262)

On Fourier transforming this result for C1 ,--- 0 we immediately get the propagator

Tt(y) _ f dk e —c o lkt—ik y _ 1 Co t
(1.263)
27r 7 (C0 t)2 4_ y 2'

representing the symmetric Cauchy process (Fig. 1.10).

We have obtained above the most general form for the characteristic expo-
nents of stable Lévy processes with scaling exponent a. In addition we have
explicit expressions for the corresponding Lévy measures. However, the stable
distribution functions obtained by Fourier transforming the characteristic func-
tions are known explicitly only in very special cases. For a = 1 we have already
seen that one gets the Cauchy distribution. Another case is provided by the
propagator given by

t t2
Tt (y) = r exp { — 2—y } , y> 0, (1.264)
-V27ry 3

and by Tt (y) --= 0 for y < 0 (see Fig. 1.10). This case corresponds to a stable
distribution with scaling exponent a = 1/2, scale parameter Co = 1, and sym-
metry parameter Ci = 1. Hence we have A(u) = u 2 . We shall discuss in Chapter
8 a physical example where this stable distribution emerges in the distribution
of the waiting times of a PDP.

References
Arnold, L. (1974). Stochastic Differential Equations. John Wiley, New York.
Bertoin, J. (1996). Lévy Processes. Cambridge University Press, Cambridge.
Bjorken, J. D. and Drell, S. D. (1965). Relativistic Quantum Fields. McGraw-
Hill, New York.
Davis, M. H. A. (1993). Markov Models and Optimization. Chapman Sz Hall,
London.
Doob, J. L. (1953). Stochastic Processes. John Wiley, New York.
Feller, W. (1968). An Introduction to Probability Theory and Its Applications,
Volume I. (third edition). John Wiley, New York.
Feller, W. (1971). An Introduction to Probability Theory and Its Applications,
Volume II. (second edition). John Wiley, New York.
Gardiner, C. W. (1985). Handbook of Stochastic Methods for Physics, Chemistry
and the Natural Sciences (second edition). Springer-Verlag, Berlin.
58 CLASSICAL PROBABILITY THEORY AND STOCHASTIC PROCESSES

Honerkamp, J. (1998). Statistical Physics. Springer-Verlag, Berlin.

Itô, Kiyosi (ed.) (1993). Encyclopedic Dictionary of Mathematics (second edi-
tion). The MIT Press, Cambridge, Massachusetts.
Klafter, J., Shlesinger, M. F. and Zumofen, G. (1996, February). Beyond Brow-
nian motion. Phys. Today, 49, 33-39.
Kloeden, P. E. and Platen, E. (1992). Numerical Solutions of Stochastic Dif-
ferential Equations. Springer-Verlag, Berlin.
Kolmogorov, A. N. (1956). Foundations of the Theory of Probability. Chelsea
Publishing Company, New York.
Reichl, L. E. (1998). A Modern Course in Statistical Physics (second edition).
John Wiley, New York.
van Kampen, N. G. (1992). Stochastic Processes in Physics and Chemistry
(second edition). North-Holland, Amsterdam,
 2
QUANTUM PROBABILITY

Having outlined the classical theory of probability and stochastic processes we

now turn to the notion of probability in quantum mechanics. The emphasis in
this chapter lies on the standard formulation of quantum mechanics as may be
found in much more detail in many excellent textbooks (von Neumann, 1955;
Landau and Lifshitz, 1958; Cohen-Tannoudji, Diu and Laloë, 1977; Bohm, 1993).
In the first section we concentrate on those aspects which clarify the relation
between quantum mechanics and classical probability theory. Most importantly,
we discuss that, although quantum mechanics is an intrinsically probabilistic
theory, the application of probabilistic concepts to quantum mechanics is quite
different from that of the classical theory. It turns out that the notion of a prob-
ability space together with the corresponding space of random variables is not
applicable to quantum mechanics. On the contrary, this concept is incompatible
with the basic structures of quantum mechanics, namely a Hilbert space of state
vectors and the corresponding algebra of observables.
We further introduce in this chapter the concept of a composite quantum
system which will be important for the study of open systems. A section on
quantum entropies is also included. The description of measurements performed
on a quantum system provides an important part of the statistical interpreta-
tion of quantum mechanics. We therefore present in this chapter the foundations
of the generalized theory of quantum measurement which is based on the no-
tions of operations and effects. These concepts will play a significant rôle in the
development of a stochastic theory of open quantum systems.

2.1 The statistical interpretation of quantum mechanics

In this section we outline the basic mathematical notions used to formulate
quantum mechanics and its statistical interpretation. We start with some formal
concepts from functional analysis and then present in the following subsections
the fundamentals of their physical and statistical significance. Further mathe-
matical details on the probabilistic interpretation of quantum mechanics may be
found in Holevo (1982).

2.1.1 Self adjoint operators and the spectral theorem

-

In quantum mechanics the state of a closed physical system is described by a

state vector 5 , or wave function V) which is an element of some Hilbert space 9-t.
5 Weshall use both the notation 1/) and the `ket' 10 for an element of the Hilbert space "?-1.
The symbol (V)1 denotes an element of the dual space ) -1* 7-1, i.e. a 'bra' vector.
60 QUANTUM PROBABILITY

The scalar product in the Hilbert space will be denoted by angular brackets,
that is we write MO for the scalar product of two state vectors V), 0 E H.
Accordingly, the norm of I,b is given by

NH E \/(010). (2.1)

The Hilbert space is assumed to be separable, which means that there exists a
finite or countable dense orthonormal basis { v a } of state vectors satisfying

(ValVO) = 60 , (2.2)

such that each state vector 10) has a unique decomposition of the form

lo) = E kooxvam. (2.3)

Û

The measurable quantities, or observables, of the closed physical system are

represented by linear, self-adjoint operators in the Hilbert space. A linear oper-
ator _h,

(2.4)

is self-adjoint if its domain D(k coincides with the domain D(k) of the adjoint
operator M and if RV) = MO on the common domain.
A theorem of fundamental importance in connection with self-adjoint oper-
ators is the spectral theorem (Akhiezer and Glazman, 1981). It states that for
any self-adjoint operator _h there exists a unique spectral family Er such that
+0,0
_h = f r dEr . (2.5)
- 00

The spectral family Er , where r E R, is a one-parameter family of commuting

orthogonal projection operators with the following properties:
1. The family of projections is monotonically increasing,

Er, _
> Er for r' > r. (2.6)

2. The family is continuous from the right,

lim Er+, = Er . (2.7)

5-->--i-0

3. It has the following limits

lim Er = 0, lim Er = I, (2.8)

where / denotes the unit operator.

 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 61

Equation (2.5) represents the spectral decomposition of self-adjoint operators on

which the statistical interpretation of quantum mechanics is based.
The spectral theorem asserts that all projectors Er of the spectral family
commute with .I' and that the domain of 1-=1 can be characterized as follows
+00
V() = { 0 G R r2 401E, ki)) < oo } . (2.9)
LL
00

Note that the integral on the right-hand side is an ordinary (Riemann—Stieltjes)

integral which may be written in the following alternative way
+00 +00
f r2 d(PIErttP) = f r2 cilWr011 2 = 1 1?- 011 2 . (2.10)

More generally, for any continuous function f (r) one can define the operator
function f (i?' ) in terms of the spectral decomposition
+Do
f(1) f (r)dEr . (2.11)
=f
—00

Furthermore, the spectral family gives rise to an orthogonal decomposition

of the unit operator. To this end, consider any disjoint partition of the real axis
into intervals Ar a = (ra _ i ,ra ], and define the corresponding projectors

AE, = Er„ — Er„,. (2.12)

With the help of the properties (2.6)—(2.8) of the spectral family we find the
completeness relation

=I (2.13)
a

and the orthogonality relation

AE, LE 3 = Sa0AE0. (2.14)

The spectrum spec(R) of a self-adjoint operator R can be expressed in terms

of the properties of its spectral family E. To this end, a point r E R is said to
be stationary if for some E > 0

Er+E Er _E- - O. (2.15)

This means that the spectral family is constant in an s-neighbourhood of r. The

spectrum spec(R) is then defined to be the set of all non-stationary points.
62 QUANTUM PROBABILITY

The spectrum spec(Ê) may be decomposed into the discrete spectrum and
the continuous spectrum. The discrete spectrum consists of all so-called jump
points. A point r E R is called a jump point if for any E > 0 one has
Er — Er, 0, (2.16)
which means that Er is not continuous from the left in r. If, for example, the
dimension of Er increases by one in r, we have for some normalized vector ça
Er — Er_E = (2.17)

follows that ço is an eigenvector of ft with eigenvalue r, i.e. ko = rcp. On

It follows
the other hand, a point r E spec(it) belongs to the continuous spectrum if the
spectral family is continuous in r.
Example 2.1 The simplest example of a self-adjoint operator .1' is that of an
operator with a purely discrete spectrum which is obtained from the eigenvalue
equation

-hOn,k — rnOn,k• (2.18)

The index n labels the different eigenvalues r n , whereas k distinguishes the eigen-
vectors belonging to the same eigenvalue. Thus, the projector onto the eigenspace
with eigenvalue r n may be written as

ll, = E 1 On,k) (On,k1 • (2.19)

k
The spectral family takes the form

Er = E rin (2.20)
rn <r

and the spectral decomposition of it is given by

+00
1- = f r dEr = E NIL. (2.21)
—co n

Example 2.2 In the Schrödinger representation the position operator 0 of a

one-dimensional system with Hilbert space 9-1 = L2 (a), the space of square-
integrable functions on R, is defined by

(00)(x) = x0(x). (2.22)

Its spectral family may be defined as follows,
{ 0(x), for x < q1
(Eq 0)(x) =
0, for x > q f = 9(q — x)0(x). (2.23)

Obviously, e-2 has only a continuous spectrum which covers the whole real axis.
 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 63

Example 2.3 The momentum operator P canonically conjugated to e) is de-

fined by (setting Planck's constant h equal to 1)

(N)(x) = - i -(x) . (2.24)

The domain of P consists of all functions 0(x) which are absolutely continuous
and whose derivative belongs to L2 (R). The spectral family of the momentum
operator is most easily defined with the help of the Fourier transform
+00

f dx e —i ldx 0(x) (2.25)

of 0(x) in terms of which we get

P
(Eplp)(x) =
f dk eikx1 , (k).
(2.26)
j 27r
—00
It can easily be verified that the momentum operator also has a purely continuous
spectrum which covers the whole real axis and that (2.26) provides, in fact, the
spectral decomposition of the momentum operator. Namely, if 0(x) is absolutely
continuous we find with the help of (2.26)
+00 +00

f
—00
p dEp O(x) = f
—00
p e'Pxl,b(p)
-d7.3
r

+00
f dp L i d eipx ) 17)(3) _i d 0(x)
f 27r dx ) dx

= (P0)(x). (2.27)

2.1.2 Observables and random variables

The statistical interpretation of quantum mechanics is closely connected to the
spectral decomposition (2.5) of self-adjoint operators .h introduced in the pre-
ceding subsection. It is based on the following postulates.
We consider a statistical ensemble E consisting of a large number of identically
prepared quantum systems S (1) , S(2) , ... , s(N) ,

E = { so) , s(2 ) , . . . , s( N)} . (2.28)

As in the classical theory, the construction of such an ensemble requires the

specification of a certain set of experimental conditions which may be realized,
at least in principle, an infinite number of times. Each realization of this iden-
tical set of conditions leads to the preparation of a single quantum mechanical
64 QUANTUM PROBABILITY

system S (i) which belongs to the ensemble E. The first postulate is that, under
certain conditions (more on this point later), a complete characterization of such
a statistical ensemble is provided by a normalized state vector 10) in the Hilbert
space 1 1 pertaining to the quantum mechanical system.
-

The second postulate is that the measurable quantities of the statistical en-
semble E are represented by self-adjoint operators in the Hilbert space 1 1. The
-

outcomes of the measurements of an observable it, performed on the ensemble

described by 10), represent a real-valued random variable R with a cumulative
distribution function FR(r) which is given by

FR(r) = (2.29)

where Er is the spectral family of ft.

Equation (2.29) represents the basis of the statistical interpretation (von Neu-
mann, 1955). It is easy to see that FR (r) has, in fact, the properties (1.18)—(1.20)
of a cumulative distribution function of a real random variable R. Note that this
is a direct consequence of the properties (2.6)—(2.8) of the spectral family Er .
In functional analysis FR (r) is referred to as the spectral measure. In view of
eqn (2.29) it is also obvious that the possible outcomes of the measurement rep-
resented by it are given by the values of r which belong to the spectrum of R.
Namely, if r is not contained in spec(ii) then the spectral family is constant in a
neighbourhood of r and, thus, also FR(r) is constant there. The probability for
the random variable R to fall into this neighbourhood is therefore equal to zero.
On the basis of eqn (2.29) one can also define the probabilities for more
general events as follows. If B denotes some Borel set of 111 we can introduce a
corresponding projection operator by means of

E(B) = f dE r, (2.30)

such that the probability for the measurement outcome to fall into the set B is
given by

PR(B = ( 0 1E ( B) 10 ).
)
(2.31)

For example, the probability for r to fall into the interval (a, b] is found to
be equal to (//dEb — Ea ) = FR (b) — FR (a). In view of eqn (2.30) we have a
correspondence between the Borel sets B and the projection operators E(B)
defined in terms of the spectral family. For any sequence of disjoint Borel sets
Bi we have

E (U i B i ) (2.32)

and a formally identical relation holds for the corresponding probabilities, namely
 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 65

PR (uiBi) = E PR(Bi)• (2.33)

i
In mathematical terms, the map B 1--> E(B) with the property (2.32) is called a
projection-valued measure.
Thus we see that an observable .1 leads, via its spectral family, to a real
random variable R which describes the probabilities for all possible measurement
outcomes. The sample space Q is the real axis and the algebra of events B is given
by the Borel sets of R. In particular, on using the spectral decomposition one
finds for the mean value of R the well-known expression

+00 +00
E(R) = f r dFR(r) = f r d(P1-Er*) = (2.34)
—cc —00

whereas the variance of R reads

Var(R) = E(R 2 ) — [E(R)] 2 = (0A 2 10) — (70 4 2. (2.35)

2.1.3 Pure states and statistical mixtures

Our characterization of quantum statistical ensembles is not yet the most general
one encountered in the applications of quantum theory. An obvious way to obtain
more general ensembles is the following. Consider a number M of ensembles
6'1 , 6.2, ... ' gm of the type introduced in the preceding subsection. Each of these
ensembles is described by a normalized state vector ,0 a , a = 1, 2, ... , M, in the
underlying Hilbert space R. It is then natural to study the statistics of the total
ensemble g which is obtained by mixing all the Ea with respective weights w a
satisfying

m
w, > 0, E w a = 1. (2.36)
ct=i

The mixing is achieved by taking a large number Na of systems from each Ea .

The total number N = Ea Na of systems then constitutes the new ensemble g
and the weights wa are given by w a = Na/N.

2.1.3.1 The statistical formulae of quantum mechanics According to the rules

of classical probability theory any self-adjoint operator .h now yields a random
variable R with cumulative distribution function

FR(r) = E wa opa 1E, *06). (2.37)

a

This equation generalizes eqn (2.29) to the present case. Accordingly, the mean
value of R is now given by
66 QUANTUM PROBABILITY

E(R)=Ewc,(7palikpa). (2.38)
a

These formulae can be cast into a compact form by introducing the density
matrix, or statistical operator,

P = Ewal0a)(0al, (2.39)

which enables one to write the distribution function of the random variable R as

FR (r) = tr {-ErP}, (2.40)

where tr denotes the trace of an operator. The trace of an operator A is defined

by

trA = E(cpi A kp i ) , (2.41)

where {(pi} is an orthonormal basis of the Hilbert space 7-1. Provided the trace
exists, it is easy to show that it does not depend on the particular choice of the
basis {cpi }.
By the same argument, the mean and the variance of R can now be written
in the following form,

E(R) = tr{kp}, (2.42)

Var(R) = tr{R 2 p} — [tr{Ao}] 2 . (2.43)

Following the usual notation we also write the expectation value of R as

(R) = tr{R- p}, (2.44)

and its variance as

Var(ft) = (f1 2 ) — (2.45)

Thus we have the result that the total ensemble E can, with regard to its
statistical properties, be completely characterized by a density matrix p. With
the help of eqn (2.39) one easily verifies that p is self-adjoint, positive 6 and has
trace one,

pt == p > 0, trp 1. (2.46)

The above reasoning can be confirmed further by an axiomatic approach (von

Neumann, 1955). Namely, consider the set of bounded self-adjoint operators on
6 An operator A is said to be positive if for all E V(A) the inequality 0/00 > 0 holds,
which is often simply written as A > O. Note that we include the equality sign in the definition
of positivity.
 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 67

the Hilbert space R. We denote this set by R. We seek a mathematical expression

for the expectation value, which is regarded as a real functional E(R) on R, that
is, which assigns a real number to every if? E R. A natural set of conditions any
such functional must fulfil is the following:

É(I) = 1, (2.47)
> a, (2.48)

E (Eat) =Ec1t(11,), ci E R. (2.49)

The first condition means that the unit operator, corresponding to a determin-
istic (i.e. dispersion-free) variable R = 1, has expectation 1. The second condi-
tion expresses the natural requirement that the expectation of any projection
operator H must be positive. This requirement results from our former observa-
tion that the events of the theory correspond to the projection operators in the
Hilbert space such that E(H) is the probability for the event represented by H
(see eqn (2.30) and (2.31)). Finally, the third condition states that E should be
a linear functional on R,. A fundamental theorem (Langerholc, 1965) then states
that any functional E(R) with these properties must necessarily be of the form
(2.42), that is there exists a unique operator p with the properties (2.46) such
that E(ft) = E(R) = tr(R p).
In our definition (2.39) of the density matrix we did not assume that the
l'006) are orthogonal. However, given p we can, of course, diagonalize it. Since p
is positive, its eigenvalues are greater than or equal to zero. The spectral theory of
density operators asserts that p has only a countable number of strictly positive
eigenvalues pi > 0. The point 0 is the only possible accumulation point of the
spectrum. Furthermore, the strictly positive eigenvalues are finitely degenerate,
and 0 is the only possible infinitely degenerate eigenvalue. Hence, the spectral
decomposition of p can be written as

P= ( 2.50)

where the sum extends over a complete set of eigenstates ko i ) with the eigenvalues
pi . The normalization condition therefore takes the form

trp = = 1. (2.51)

2.1.3.2 Properties of the density matrix We briefly discuss the most important
general properties of density matrices. The first one is the inequality

trp2 < trp = 1. (2.52)

It may be shown that the equality sign holds if and only if p has the form
p = WO for some unit state vector 0). This is the case considered in the
68 QUANTUM PROBABILITY

previous subsection 2.1.2. An ensemble with this property is said to be in a

pure state. It is clear that one can replace the above kp) by exp(iX)10), with
an arbitrary phase x E [0, 27), without changing the density matrix and the
statistical formulae. Thus, a pure state is uniquely described by a ray in the
underlying Hilbert space. In mathematical terms, one can therefore say that the
set of pure states is isomorphic to the set of rays in 7-1, that is, to the projective
Hilbert space.
If the left-hand side in (2.52) is strictly less than 1, the ensemble is called a
statistical mixture. Alternatively, the difference between pure states and mixtures
may be characterized as follows. The set of all density matrices will be denoted
by S(7-1). This set is convex which means that for any two density matrices pi.
and p2 the convex linear combination

p = Api + (1— A)P2, À E [0, 1- ] , (2.53)

is also a density matrix. Physically, this convex linear combination describes an

ensemble which is obtained from the mixture of the ensembles Ei (corresponding
to pl ) and 82 (corresponding to p2 ) with weights A and 1 — A, respectively. The
pure states lie on the boundary of the set S(7-1) and are distinguished by the
fact that they cannot be represented as a non-trivial convex linear combination
of two different density matrices. More precisely, if p is a pure state and if (2.53)
holds for some A G (0, 1), it follows that pi. = p2 = p. The physical implication is
that any decomposition of a pure statistical ensemble E into two sub-ensembles
El and 82 does not change in any way the statistical properties, that is, Ei and
82 have the same statistics as E.
In eqn (2.39) the density matrix has been introduced as a convex linear com-
bination of normalized, not necessarily orthogonal states kp„) with weights w,
satisfying (2.36). Given some density matrix p this decomposition into a convex
linear combination of pure states is, in general, not unique (Hughston, Jozsa and
Wootters, 1993; Nielsen and Chuang, 2000). In order to characterize the possible
convex linear combinations which give rise to one and the same density matrix it
is convenient to work with non-normalized states = -Vw,10„) and to write
the density matrix as

P = (2.54)

Such a set of states kpo } is said to generate the given density matrix p. Consider
{

now two sets of states {kJ)} and { i„)}. These two sets generate the same
density matrix p, that is

P = 11a)(1 = (2.55)

if and only if there exists a unitary matrix u = (u,8) such that

 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 69

k-ba)=Euaork3). (2.56)

Here, one appends zero vectors to the set with the smaller number of states in
such a way that the number of states in both sets becomes equal.
This statement will be very useful later on, for example in the character-
ization of the freedom one has in the representation of generalized quantum
measurements. To prove the statement we first assume that (2.56) holds. By
direct substitution it is then easily seen that (2.55) holds, that is both sets gen-
erate the same density matrix. Conversely, suppose that (2.55) holds. We write
the spectral decomposition (2.50) of p as p = E i 1;5 i)(k, where kp i ) =
Consider any state 1 ,0) which is orthogonal to the space spanned by the states
{Içoi)}. Then we have

oplpko =0= E (ok-ba)12, (2.57)

a

and we conclude that (010a ) = 0 for all a and for all 10) of the above form. This
means that 1/-- ,:t ) can be expressed as a linear combination of the Iço" i ),

k-ba) = Ecctirpi). (2.58)

This gives

P > iXi =E
a
* i c ai
a
k-lo i) (cAi I,
)
(2.59)

from which we deduce that

Ea
c t ca,
Ja (2.60)

This means that the matrix c = (c, i ) can be supplemented to form a square and
unitary matrix such that

ka) =EcaoPo), (2.61)

where, if appropriate, zero vectors must be appended to the set of the states
4i ). In the same manner we get a relation of the form

via) = Eda,40), (2.62)

0
with a unitary matrix d. The last two equations yield a relation of the desired
form (2.56), where the matrix u is defined by u = cdt which is unitary since c
and d are unitary.
70 QUANTUM PROBABILITY

2.1.3.3 The non-existence of dispersion-free ensembles It is a well-known fact

that quantum mechanics cannot be formulated as a statistical theory on a clas-
sical probability space. There exist many mathematical theorems which express
this fact in various ways (see, e.g. Gudder, 1979). Here and in the following
subsection we shall elucidate this point from a probabilistic viewpoint.
Consider an arbitrary Borel algebra B of events on a sample space St (see
Section 1.1). One can then always introduce a dispersion-free measure p, on B in
the following way: Take some fixed point wo E Q and define for A E B,

for wo E A,
p(A) E { 01 : for wo çt A.
(2.63)

The probability space thus defined has the property that all random variables
X on it are dispersion-free. In fact, the cumulative distribution of X is given by
(see eqn (1.17))

{ 0, for x < xo ,
Fx (x) = 1, for x > xo .
(2.64)

This shows that X is a sharp, deterministic variable which takes on the single
value xo = X(wo ) with unit probability.
By a well-known theorem of von Neumann (von Neumann, 1955) in quan-
tum mechanics the situation is markedly different, namely there do not exist
dispersion-free ensembles E of whatsoever type. This is easily demonstrated with
the help of the statistical formulae derived above. Namely, if E were such an en-
semble and p its statistical operator we must have Var(R) = 0 for all observables
1-=/, that is
e , 2
tr {.fi2 p} = [tr ppl] . (2.65)

Taking for R the one-dimensional projections It = IOW, 11011 = 1, one finds

(01P10) = (01)010) 2 , (2.66)

and, therefore, (00 ) must be identically 0 or 1 for all unit vectors in R.

It follows that p = 0 or p = I, which is a contradiction for both cases do
not represent a statistical operator (they violate the normalization condition
tr(p) = 1; if one admits non-normalizable density matrices the case p = 0 is
clearly excluded, but also the case p = I is excluded since it does not represent
a dispersion-free ensemble as is easily demonstrated directly).

2.1.4 Joint probabilities in quantum mechanics

As we have seen, a self-adjoint operator it leads via its spectral family Er to a
random variable R with cumulative distribution function FR (r) given by (2.29)
(for simplicity we consider here only ensembles in pure states). Consider now
 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 71

two self-adjoint operators it ]. and R2 and the corresponding random variables

R1 , R2 with distribution functions
Fi (r) = (7PLE:,/p), i = 1,2, (2.67)
where 41 and Er2 denote the spectral families of Pi and R2 respectively. An
important question is then, of what nature are the statistical correlations between
the random variables R1 and R2 and can one describe these correlations in terms
of a classical joint probability?
This question is answered by a theorem of Nelson (1967) which asserts, es-
sentially, that two observables .h 1 and R2 can be represented in all states as
random variables on a common, classical probability space if and only if they
commute, that is, if and only if [k , /-42] = 0. Thus, the correlations between
non-commuting observables do not in general admit the characterization with
the help of a classical joint probability distribution.
Let us formulate this theorem more precisely. To this end, consider two ran-
dom numbers R 1 , R2 on a common probability space with probability measure
p. We denote by
F(ri, r2) = P (Ri < r1, R2 < r2) (2.68)
their cumulative joint probability distribution. If R 1 and R2 are statistically inde-
pendent (see eqn (1.26)) the joint probability distribution factorizes, F(ri , r 2 ) =
F1 (r i ) .F2 (r2 ). In that case, F(ri , r2) can be reconstructed, of course, from knowl-
edge of the marginal distributions F1 (ri ) and F2 (7-2 ) . This is not possible if Ri
and R2 are statistically dependent. However, if the distribution function of the
random variable
k • R k i R i + k2R2, (2.69)
namely the function

Fk.R(r) = (kiRI + k2R2 < r) (2.70)

is known for all real kl, k 2 , the joint distribution F(ri , r2 ) is uniquely determined.
This is a theorem of classical probability theory and can be shown as follows.
Recall that F(ri, r2) is uniquely determined by its characteristic function

G(k 1 ,k 2 ) = f exp + k2 r2)] dF(ri , r2 ). (2.71)

R2
This equation can be written as
+00
G(k1,k2) rclFk.R(r), (2.72)
=f
—00
which shows explicitly that the characteristic function of F(ri , r2) and, therefore,
F(r i , r2 ) itself can be obtained from knowledge of Fk.R(r).
72 QUANTUM PROBABILITY

Turn now to quantum mechanics and consider the above self-adjoint operators
-k1, R2. We assume that also the linear combination
k • ft ki Pi + k2î2 (2.73)

is self-adjoint for all real k 1 , k2 and denote by Erk 'k the spectral family of the
operator k • P. One should expect then that the self-adjoint operator (2.73) cor-
responds to the random variable (2.69) which means, according to the statistical
formulae of quantum mechanics, that the distribution function of (2.69) is given
by

Fk.R(r) = (0044 10. (2.74)

If this is true the observables ill and R2 have a joint probability distribution
F(r i , r2 ) which can be determined from Fk . R (r) with the help of eqn (2.72).
Thus we are led to the following definition: Two observables R1 and R2 are said
to have a joint probability distribution in the state 0 if there exist a classical
probability space and two random numbers R I , R2 on it such that for all real
k 1 , k2 eqn (2.74) holds.
Inserting eqn (2.74) into eqn (2.72) we get
+co +00

G( ki , k2) =
f e"dFk.R(r) = (11) f e ir dErk.1:1 , )

= (0 lexp (ik • .fi)10), (2.75)

where we have applied eqn (2.11) to the function f(r) = exp(ir) and to the
operator k• ft. Thus we see that, if I'L and R2 have a joint probability distribution
in the state 0, then (2.75) is necessarily its characteristic function.
Let us investigate the expectation values generated by G(k i , k2 ). The first
terms of its Taylor expansion around k = 0 are given by

G(k1,k2) = 1 + iki(01f1110) + ik2(01-h210) - k4 Win10) - --

k2 (141410)

- -1 kik2(01- 1?‘ ift2 + fl2f1111,b) + 0(k 3 ). (2.76)

2
This shows that G(k i , k2 ) generates the quantum expectation value of the sym-
metrically ordered product (i/11/2 + -1?- 2-ii1)/ 2 , but not that of the operator
i(R 1 R 2 — iat1 )/2. In other words, G does not contain any information on the
difference between the expectations (0 -1-=‘11 1/21'0) and (01-k ii1 kb ) •
With the above definition of joint probability we can now give the precise
formulation of Nelson's theorem: The observables R1 and R2 have a joint proba-
bility distribution in all states 0 if and only if they commute, that is if and only
if [Ri , R2] = O.
 THE STATISTICAL INTERPRETATION OF QUANTUM MECHANICS 73

It is clear that R i and 1?2 have a joint probability distribution if they com-
mute, namely

F(ri,r2) == (01-E 1E 2 ). (2.77)

As is easily checked using the properties of the spectral families Erl , and 422 , this
is a real, non-negative function with the properties of a true joint distribution
function. In addition, it leads to the characteristic function (2.75) and the relation
(2.74) is satisfied.
The general proof of the only if part of the theorem is more difficult. We
refrain from giving the details of this proof, but merely discuss the important
case Ri = 0, the position operator, and f212 = P, the momentum operator,
satisfying the Heisenberg commutation relation (h = 1),

(2.78)

It is then easily seen that the function (2.75) is not the characteristic function
of a classical joint probability distribution. In fact, on Fourier transforming we
find that the density of F (q, p) is given by

32F d2k7ri 2k7r2

f d_e—i(kiq+k2P) e i( k 1 0+k2 15)
f (q, (2.79)
Oqap
d2k7ri f d2k: f dxe—i(k,q+k2p)+ikik212+ikix (s)0(x + k2).

Here we have introduced the position representation of the state vector, 0(x) =
(x10). Further we made use of
ei(k1ed+k2i3) = eik10 e ik2P e ik1k212
(2.80)

and of the fact that exp(ik 2 /5 ) induces the translation x X + k2. Carrying out
the integrations over k i and x and substituting s = -k2 we finally arrive at

+00
ds
f (q, p) — e 8* (q + s12)0(q — s/2). (2.81)
=f
- 00

This is the famous Wigner distribution (see, e.g. Louisell, 1990). It is interesting
to see that, in the present context, it arises as the only possible candidate for a
joint probability density of position and momentum. However, as is well known
it does not represent a true probability density since it is not positive for general
0. An example is shown in Fig. 2.1.
In spite of the fact that the Wigner distribution is not a true probability
distribution it is a useful quantity. Its Fourier transform is known as the Wigner
74 QUANTUM PROBABILITY

-4

FIG. 2.1. The Wigner distribution (2.81) corresponding to the third excited
eigenstate of the harmonic oscillator Hamiltonian H = 21- (p2 02).

characteristic function. For a system described by an arbitrary density matrix p

it is given by

G(ki, k2) = tr exp (i(ki + k2 /5)) P} (2.82)

Introducing the creation and destruction operators bt and b through the relations
= (b+bt)INa P = — i(b — b)/v', and using the complex variable a = ik i —k2
one often writes the Wigner characteristic function as

x(a, a*) = tr {exp (abt — ce*b) p}. (2.83)

Similarly to eqn (2.76), it may be shown by an expansion of the exponential that

x(a, cy*) generates the symmetrically ordered products of b and bt through the
formula
an+rn
tr { (bnbil s pl = Dam x(a, a*) (2.84)
Œ=Œ*=0

Here, the symmetrically ordered product Onbt m s is defined to be the average

over all possibilities of ordering the operators in the product bn bt m

2.2 Composite quantum systems

In the theory of open quantum systems the notion of composite quantum systems
is fundamental. The Hilbert space of a composite quantum system is the tensor
 COMPOSITE QUANTUM SYSTEMS 75

product space of the Hilbert spaces describing its subsystems. The present section
serves to introduce some basic features of tensor products of Hilbert spaces and
of the statistical properties of composite systems and their subsystems.

2.2.1 Tensor product

We consider two quantum systems S(1) and S (2 ) with respective Hilbert spaces
70) and 'H (2 ). The two systems may represent two (distinguishable) particles,
two different composite objects (e.g. two atoms or molecules), or two different
degrees of freedom of the same object (e.g. rotational and translational degrees
of freedom of a molecule). In general, SO-) and S( 2) interact with each other.
According to the postulates of quantum mechanics the state space 7/ of the
combined system S = S (1) + S(2) is given by the tensor product of the Hilbert
spaces pertaining to the subsystems S (1) and S( 2),
71 R (1) 0 74 (2) .
(2.85)

If we take fixed orthonormal bases kc), 1) )1 and v i( 2) ) } in 71 (1) and 71( 2) , re-
{

spectively, a general state in the tensor product space I/ may be written as

141) =E kP,i( 1) ) IV?). (2.86)

ii

This means that the elements k0,1)) 0(2) 110i ) form a basis of the tensor product
space and that the dimension of 71 is equal to the product of the dimensions of
R (1) and 71 (2) .
If A(') is an operator acting in 70 ) and A( 2) is an operator acting in 7t( 2) ,
one defines their tensor product A( 1 ) 0 A (2) by
0 (A (2) bp 3(2) ),
(A M o A (2) )(101) ) blo(i2) )) E (A(')1(p.(1))) (2.87)
and by a linear extension of this formula for arbitrary states (2.86). Any operator
A acting on 71 can be represented as a linear combination of tensor products,

A= E A(2) 0 A(2). (2.88)

a

Specifically, the observables of system S( 1 ) take the form AO ) 0 1( 2), while ob-
servables of system S (2) are given by the expression /( 1 ) 0 A (2) . The identity
operators in 70 ) and 71( 2) are denoted, respectively, by PO and 1 (2) .
The density matrix of the composite system S is an operator in the state space
S(7-1). If the subsystems S (1 ) and S( 2 ) are uncorrelated the total density matrix
takes the form of a tensor product of the density matrices of the subsystems,

(2.89)
This implies that the expectation value of any tensor product of operators per-
taining to the subsystems factorizes,
76 QUANTUM PROBABILITY

(A(1) 0 A(2) ) E tr (A (1) 0 A(2) ) pl = tr (1) p(1) } fr(2) {A(2)p(2)}

= (A ( ' ) ) • (A (2) ). (2.90)

Here, tr (1) and tr (2) denote the partial traces over the Hilbert spaces R (1) and
71( 2), respectively.
If one is only interested in observables of subsystem S (1) , that is only in
operators of the form

A = A' o 1(2) , (2.91)

it is convenient to introduce the reduced density matrix pertaining to the sub-

system by means of

p (l) —= t r (2) p. (2.92)

In this equation tr( 2) denotes the partial trace taken over the second Hilbert space
R( 2) . A density matrix which is obtained in this way through a partial trace of
a density matrix in some larger space is sometimes called an improper mixture.
It completely describes the statistical properties of all observables belonging to
the subsystem under consideration since the expectation value of any observable
of the form (2.91) can be determined with the help of the formula

(A) = tr (1) 0(1) p(1) }. (2.93)

We have seen that the reduced density matrix of subsystem S( 1) is obtained

by taking the partial trace over subsystem S( 2 ). It might be interesting to note
that this way of describing subsystem S( 1) is the only possible. More precisely,
if we seek a function f : 8(R) S(9 1( 1) ) which maps the state space of the
-

composite system to the state space of the subsystem, then the function given
by f (p) = tr (2) p is the only one with the property

tr (1) {A(1) f(p)} tr (A (1) 1 (2) ) pl . (2.94)

Thus, the definition of the reduced density matrix given above is the only possible
one which is compatible with the statistical formulae.
To prove this statement it is convenient to introduce the Liouville space.
Given some Hilbert space R the Liouville space is the space of Hilbert—Schmidt
operators, that is the space of operators A in R for which trAtA is finite.
Equipped with the scalar product

(A, B) trfAtB1 (2.95)

the space of Hilbert—Schmidt operators becomes a Hilbert space. One can there-
fore introduce an orthonormal basis {Bi in this space satisfying the orthogo-
}

nality and the completeness conditions,

 COMPOSITE QUANTUM SYSTEMS 77

(Bi , Bi ) = 6, (2.96)
A = E Bi (Bi , A). (2.97)

Let us consider now the Liouville space associated with the state space 71(1)
of subsystem S (1) . We take an orthonormal basis {Bi } of Hermitian operators
Bi on R (1) • Using the completeness condition (2.97) and the requirement (2.94)
we then get

f (p) = E Bi (Bi, f(P))

E Bi tr (1) {Bi f (p)}

EBitr{(Bi / (2) )p}

= EBitr(1){Bip(1)}

=
p( l) = tr (2) p, (2.98)

which proves that f (p) is necessarily given by the partial trace of p taken over
subsystem S( 2) .

2.2.2 Schmidt decomposition and entanglement

An important characterization of the states of a composite quantum system is
obtained with the help of the Schmidt decomposition theorem. This theorem
asserts that for any given state IT) E R (1) 0 11 (2) there exist orthonormal bases,
the Schmidt bases { lxn} and {lx i( 2)5 } in 71 (1) and 9( ( 2) , respectively, such that

= Eailx1)) ® le). (2.99)

The ai are complex numbers called Schmidt coefficients. For a normalized state
we have obviously,

(TIT) =E 2=1 (2.100)

It must be noted that the Schmidt bases which allow a representation of the
form (2.99) depend, in general, on the given state IT).
The Schmidt decomposition theorem can be proven as follows. First, we may
suppose without restriction that R (1 ) and R( 2 ) have the same dimension. The
matrix a = (aii) of coefficients in the decomposition (2.86) with respect to the
78 QUANTUM PROBABILITY

fixed basis vectors vi(1) ) and Ivi(2) ) is then a square matrix. By use of the singular
value decomposition this matrix can always be written as a = udv, where u and
v are unitary matrices and d is a diagonal matrix with non-negative diagonal
elements ai > 0. Thus, the decomposition (2.86) takes the form

IT) = EuijaiVik kli i) ) 0 1V (k2) ) (2.101)

iik
By virtue of the unitarity of u and v the vectors defined by

Euiibp,(i 1 )), (2.102)

= ( 2) ,
l x)
i( 2) vikivk /, (2. 1 03 )

form orthonormal bases in 11 (1) and 9( (2) , respectively. Using these expressions
in eqn (2.101) we immediately obtain the Schmidt decomposition (2.99). Note
that the proof also demonstrates that the Schmidt bases can always be chosen
such that the Schmidt coefficients are real and non-negative.
The number of non-zero Schmidt coefficients ai is called the Schmidt number.
This number is invariant with respect to unitary transformations u( 1 ) and U(2)
which act only in the respective spaces 7/( 1 ) and 71 (2 ). For the same reason the
Schmidt number does not depend on the particular Schmidt bases chosen and is
thus uniquely defined for a given state IT).
A state IT) E 71 (1) 011 (2) is said to be entangled if it cannot be written as a
tensor product ko(1) ) 0 I v (2) ) of states of the subsystems. If IT) can be written
as a tensor product we call it a product state. It follows from the Schmidt
decomposition theorem that IT) is an entangled state if and only if the Schmidt
number is larger than 1. Equivalently, IT) is a product state if and only if its
Schmidt number is equal to 1. Another notion will be important later on: If the
absolute values of all non-vanishing Schmidt coefficients for a given state Alf) are
equal to each other the state is said to be maximally entangled.
The Schmidt decomposition can be used to prove several interesting state-
ments about the states of a composite system and its subsystems. For example,
if the combined system S is in a pure state p = 1 41 )011 1 the reduced density
matrices p(l) = tr(2) p and p(2) = tr( 1) p have the same eigenvalues. In fact, the
Schmidt decomposition of IT) immediately yields

p(1) tr (2) {1T)(TII = E lad21x(i 1 ))(x(i 1 ) (2.104)

p(2) = tr (1) {0)(TII = E clei121x(i 2 ))(x(i, 2)k (2.105)

from which the statement follows. We note that the subsystems are, in general,
described by mixed states. In particular, if IT) is maximally entangled the re-
duced densities are proportional to the identities in the subspaces spanned by
 QUANTUM ENTROPIES 79

the Schmidt basis vectors belonging to the non-vanishing Schmidt coefficients.

Moreover, these representations show that Ili) is a product state if and only if
the reduced density matrix p(l ) and, therefore, also p( 2) describe pure states.

2.3 Quantum entropies

Quantum entropies play a crucial rôle in quantum statistical mechanics and
quantum information theory. In this section we briefly review the basic defini-
tions and some important properties of quantum entropies, which will be used
in later sections to characterize, for example, the information gained in a quan-
tum measurement and the irreversible nature of the quantum dynamics of open
systems. More details about quantum entropies may be found in a review article
by Werl (1978) and in the book by Nielsen and Chuang (2000).

2.3.1 Von Neumann entropy

The von Neumann entropy provides an important entropy functional used in
quantum statistical mechanics and thermodynamics. Given the state of a quan-
tum statistical ensemble in terms of a density matrix p it is defined by

S(p) —tr {pinp} (2.106)

In the present section we set the Boltzmann constant kB equal to 1, that is we

measure temperatures in units of energy. Using the spectral decomposition of
the density matrix,

P =EPibPi)(Pil, 0, E pi = 1, (2.107)

we can write

S(P) = Pi 1 nPi 1/({Pi}). (2.108)

In these expressions it is understood that 0 • ln O 0. For a mathematically

precise definition one defines S(p) by means of equation (2.106) if the trace is
finite, and sets S(p) = d oo otherwise.
-

Equation (2.108) shows that the von Neumann entropy is equal to the Shan-
non information entropy H({p i }) of the distribution i pi , that is of the random
number / with the distribution pi = = i) given by the spectral decomposi-
tion of the density matrix p. A statistical mixture which is described by p can be
obtained by mixing pure ensembles described by states ko i ) with corresponding
weights pi . Then S(p) expresses our uncertainty, or lack of knowledge about the
realization of a particular state bpi) in the mixture.
We list without proof some important properties of the von Neumann en-
tropy which make evident its importance for quantum statistical mechanics and
quantum information theory.
80 QUANTUM PROBABILITY

1. For all density matrices one has

8(p) > 0, (2.109)

where the equality sign holds if and only if p is a pure state.
2. If the dimension of the Hilbert space is finite, dim 7-1 = D < oo, the von
Neumann entropy is bounded from above S`(p) < ln D, where the equality
sign holds if and only if p is the completely mixed or infinite temperature
state p = I ID.
3. The von Neumann entropy is invariant with respect to unitary transfor-
mations U of the Hilbert space, that is S(UpUt)= S(p).
4. The von Neumann entropy is a concave functional p 1-4 S(p) on the space
of density matrices. This means that for any collection of densities pi and
numbers A i > 0 satisfying Ei Ai = 1 one has the inequality

> (2.110)

The equality sign in this relation holds if and only if all pi with non-
vanishing A i are equal to each other. This property is called strict concavity
of the entropy functional. In physical terms it means that our uncertainty
about the state p = Ei Aipi is greater than or equal to the average uncer-
tainty of the states pi that constitute the total mixture.
5. Consider a composite system with Hilbert space 7-1 = 71 (1) 0 71 (2) and
denote by p the density matrix of the total system and by p( 1 ) = tr( 2)p
and p( 2) = tr(l) p the densities of the subsystems. Then the von Neumann
entropy obeys the so-called subadditivity condition,

S(p) < S(p 1 ) + S(p (2) ), (2.111)

where the equality sign holds if and only if the total density matrix de-
scribes an uncorrelated state, p = p(1) p(). Thus, our uncertainty about
the product state p( 1 ) p( 2) is, in general, greater than our uncertainty
about the state p of the composite system. In other words, by tracing over
the subsystems we lose information on correlations between the subsystems
and, consequently, increase the entropy. If P = 111')(411 is a pure state we
have S(p) = 0. It follows from the Schmidt decomposition that p(1) and
p(2) have the same eigenvalues (see eqns (2.104) and (2.105)) and, thus the
von Neumann entropies of the subsystems are equal to each other. This
yields

S(po) = S(p2) =-E >0, (2.112)

and the left-hand side of the inequality is strictly greater than zero if and
only if IT) is an entangled state.
 QUANTUM ENTROPIES 81

2.3.2 Relative entropy

For a given pair of density matrices p and a the relative entropy is defined by
S(pa) tr {p ln pl tr {pin o- } . (2.113)
To give a rigorous definition one introduces the kernel of a density matrix as the
space spanned by the eigenstates with zero eigenvalue, and the image of a density
matrix as the space spanned by the eigenstates belonging to the non-vanishing
eigenvalues. The relative entropy is then defined to be equal to +co if the kernel
of a has a non-trivial intersection with the image of p.
A physical interpretation for the relative entropy can be given if we consider
again a composite quantum system as in point 5 of Subsection 2.3.1. Then we
have

s(pl Ip(1) ® 10(2) ) = s(P(1) ) + s (P(2) ) s(p). (2.114)

The entropy of the density matrix p of a combined system relative to the corre-
sponding uncorrelated state p(1) p(2) is thus a measure of the change of the von
Neumann entropy resulting from the tracing over the subsystems, and provides
a measure for the corresponding information loss.
Let us summarize some important properties of the relative entropy func-
tional.
1. The relative entropy fulfils the inequality

S(P10- ) >0 (2.115)

for all p and all a, which is known as the Klein inequality. The equality
sign holds if and only if p = a. With the help of eqn (2.114) the Klein
inequality leads to the subadditivity property (2.111) of the von Neumann
entropy.
2. Similarly to the von Neumann entropy, the relative entropy is invariant
with respect to unitary transformations U,

S(UpUtlUo-Ut) = s(p 1 10.)- (2.116)

3. The relative entropy is jointly convex in its arguments. This means that
for 0 < A < 1 the inequality

S(pa) AS(Pdlai) + — A)S(P211 0-2) (2.117)

is satisfied, where p = Api + (1 — A)p2 and a --= Au1 + (1 — A)o- 2 -

4. Setting p( 1 ) = tr (2) p and a(1) = tr( 2 )o- we have

s(po ) dam) < s(pl (2.118)

Thus, the tracing over a subsystem in both input arguments reduces the
relative entropy. In particular, if p is an uncorrelated state we find

s(0( 1) 1 1 .7(1) ) = S(P (1) P(2) 11.7(1) g P(2) ). (2.119)

82 QUANTUM PROBABILITY

We finally comment briefly on a theorem which is of fundamental importance

in many proofs of entropy inequalities. This is Lieb's theorem which states that
the functional

ft (A, B) = —tr {XtA t XB 1-t } (2.120)

is jointly convex in its arguments A and B. Here, A and B are positive operators,
while X is an arbitrary fixed operator and t is a fixed number in the interval [0, 1].
As an example, we note that Lieb's theorem plays a crucial rôle in the proof of
the strong subadditivity of the von Neumann entropy (Lieb, 1973) which asserts
that for any system which is composed of three subsystems S (1) , 8( 2 ) and S(3 )
the inequality
s (p( 1,2,3)) + S(p2 ) < s(p( 1,2 )) + s (p( 2,3)) (2.121)

holds, where p(123) is the density matrix of the total system and
p ,, ,
= tl' (3)(123) (2.122)
= tr (l) p (1,2,3), (2.123)
p(2) -= tr (1 '3) p(1 ' 2 ' 3) , (2.124)

are reduced density matrices pertaining to the subsystems SO) + 8(2) , 8 (2) + S( 3)
and 8 (2), respectively. For classical systems the proof of strong subadditivity is
relatively simple. By contrast, no easy proof of strong subadditivity is known in
the quantum case (for details, see Werl, 1978; Nielsen and Chuang, 2000).
We shall use Lieb's theorem in Chapter 3 to prove that the entropy production
rate for a quantum dynamical semigroup is a convex functional on the state
space of an open quantum system. As a preparatory step we take A = B = p in
eqn (2.120) to get the convex functional

ft (p) - —tr {Xtp t Xp l-t } . (2.125)

Since fo (p) -= —tr{XtXp} is a linear functional we conclude that the derivative

of ft (p) with respect to t taken at t = 0 is also a convex functional. Thus, by
Lieb's theorem the functional

pi-4 —tr { (XpXt — XtXp) lnp} (2.126)

is convex. To prove convexity of the entropy production rate we have to con-

nect the expression under the trace to the generator of an irreversible quantum
dynamical semigroup which will be derived in Chapter 3.

2.3.3 Linear entropy

As another measure for the purity of states it is sometimes useful to use, instead
of the von Neumann entropy, the so-called linear entropy Si (p) which is defined
by
 THE THEORY OF QUANTUM MEASUREMENT 83

St (p) --= tr p - p2 = 1 - trp2 . (2.127)

We can immediately give an upper and a lower bound for this functional,
0 < Si(p) < 1. (2.128)
The first inequality is obvious from the inequality (2.52), which also shows that
Si(p) is equal to zero if and only if p is a pure state. The second inequality follows
from the fact that trp2 is a positive operator.
For a D-dimensional space, dirn1-1 --= D, we have the upper bound
1
Si(p) <1— — (2.129)
D.
The functional Si(p) attains its maximal value for the infinite-temperature state
p , I1D.

2.4 The theory of quantum measurement

In this section we discuss the quantum theory of measurement which is funda-
mental to the statistical interpretation of quantum mechanics. A general intro-
duction may be found in the book by Gottfried (1974).
The measurement process in quantum mechanics plays a dual rôle. On the one
hand, it describes the way in which the state of a quantum system changes if a
measurement is performed on it, thereby influencing the predictions on the future
behaviour of the system. On the other hand, it gives a unique prescription for
the preparation of a quantum system in a definite state. Thus, the measurement
process tells us how to choose the set of experimental conditions we have been
talking about at the beginning of Section 2.1.2 and which led to the realization
of quantum statistical ensembles.
2.4.1 Ideal quantum measurements
The fundamental measurement postulate of quantum mechanics may be formu-
lated as follows. Suppose that some property B with corresponding projection
operator E(B) is measured on a quantum statistical ensemble E described by
the density matrix p. Then, after the measurement, the density matrix
E(B)pE(B)
P = tr {E(B)pE(B)} (2.130)

describes the sub-ensemble E' consisting of those systems for which property B
has been found to be true. This is the well-known von Neumann—Liiders projec-
tion postulate (von Neumann, 1955; Eiders, 1951). Note that the denominator in
(2.130) ensures that p' is normalized, tip' = 1, and that it is just the probability
for the measured property to occur, namely
P(B) --= tr {E(B)pE(B)} = tr {E(B)p} . (2.131)
If we take some self-adjoint operator 1 , then B can be, for example, a Borel
set of 111, the corresponding event being that the random variable R takes a value
84 QUANTUM PROBABILITY

in B. The projection operator E(B) is defined in terms of the spectral family of

.1 by eqn (2.30). For example, consider the orthogonal decomposition of unity
which is given by the projections AE„, defined in eqn (2.12). The probability for
the event R E Ar, then takes the form

= tr {AE,p} , (2.132)

whereas the sub-ensemble corresponding to this particular event is described by

the density matrix

1
(2.133)

The probabilities P(r) are normalized to 1 in view of the completeness relation

(2.13),

= 1. (2.134)
a

The density matrices p% of the sub-ensembles are orthogonal in the sense that
P'aP"3 := 60,010-
The measurement of an orthogonal decomposition AE,„ of unity thus leads
to a decomposition of the original ensemble g into the various sub-ensembles
labelled by the index a. Such a splitting of the original ensemble into various sub-
ensembles, each of which being conditioned on a specific measurement outcome,
is called a selective measurement.
One could also imagine an experimental situation in which the various sub-
ensembles are again mixed with the probabilities P(Ara ) of their occurrence.
The resulting ensemble is then described by the density matrix

P(Ar,)p'o, --= E AE,pAE,. (2.135)

a a

This remixing of the sub-ensembles after the measurement is referred to as non-

selective measurement.
As another example, take a self-adjoint operator .fi with a discrete, possibly
degenerate spectrum (see Example 2.1) and the initial density matrix p --=
The probability for the measurement of the discrete eigenvalue rn , that is, for
the event R = r72 corresponding to the projection operator II„ , is given by

P(r) = tr {ILO — (01 11n10). (2.136)

The corresponding sub-ensemble is described by the density matrix

rin10)(//) Inn — 10 (0 / (2.137)

P'n P(rn)
 THE THEORY OF QUANTUM MEASUREMENT 85

which implies that the measured sub-ensemble may be represented by the nor-
malized state vector
o= (2.138)

Thus we see that the sub-ensemble is again a pure state. This is not true, in
general, if the initial density matrix p is a true mixture. If however the eigenvalue
rn is non-degenerate, the projection being Il n = 10n) (Onl , then its measurement
does lead to a sub-ensemble describable by a pure state 0, even if p is a true
mixture, namely 0 = On with probability P(r) = (On 1P1 On)•
Finally, we study the position operator Q (see Example 2.2). Suppose that
an ideal measurement of the event Q G A EE (a, b], that is, of the event that the
coordinate lies in the interval (a, b] is performed on an ensemble E in the pure
state p ,10)(01. The probability for this event is

J
P(A) = tr {E(A)p} -= f dx (x)2,
a
(2.139)

and the corresponding sub-ensemble is to be described by the pure state p' =

10)(01, where
{p(A)--i / 2 0 , x .,), a<x<_b
0(x) --= (2.140)
0, x < a, x > b

represents the new wave function after the measurement.

2.4.2 Operations and effects

It is important to emphasize that the last example of the preceding subsection
describes the ideal quantum measurement of the projection E(A) which is de-
rived from the continuous spectral family Eq of the position operator Q. One
must expect, however, that in practice it is not the exact spectral family Eq
that is measured, but rather some kind of approximation which involves the fi-
nite resolution of the detector. Moreover, as we shall see, there are many other
important variants of quantum measurement schemes that are encountered in
practice. It turns out that in this context the generalized measurement theory
based on the notions of operations and effects is extremely useful.
We consider some measurement scheme which yields a set M of possible
outcomes Tn G M. One may consider M as a classical sample space, and the
possible outcomes Tn as its elementary events. For simplicity we generally assume
the set M to be discrete. The measurement is performed on some quantum
statistical ensemble which is described by a density matrix p.
The generalized theory of quantum measurements (Kraus, 1983; Davies, 1976;
Braginsky and Khalili, 1992) is based on the following concepts which are illus-
trated in Fig. 2.2 and which may be viewed as natural generalizations of the von
Neumann—Liiders projection postulate (compare with eqns (2.130) and (2.131)):
86 QUANTUM PROBABILITY

1. The measurement outcome m represents a classical random number with

probability distribution

P ( rri) = tr {Fm p} , (2.141)

where F, is a positive operator, called the effect, which satisfies the nor-
malization condition

E F, = I, (2.142)
mEM

such that P(m) is also normalized, that is,

E P(m) = 1. (2.143)
mEM

2. For the case that the measurement is a selective one, the sub-ensemble of
those systems for which the outcome in has been found is to be described
by the density matrix

P,, — Pern) -l cprn(p), (2.144)

where cb, = (1. m (p) is a positive super-operator, called an operation, which

maps positive operators to positive operators. The operation 4., is further
assumed to obey the consistency condition

tr (1)„(p) = tr {F,p} , (2.145)

which yields, together with eqn (2.141), the normalization of the density
matrix p, namely

tr pirn --= P (m) - 1 tr cb m (p) = 1. (2.146)

3. For the corresponding non-selective measurement we get the density matrix

P1 -= E
mEM
P(rn)p/m =
mEM
(2.147)

which is normalized according to eqns (2.145) and (2.142),

trpf --= tr c1( p ) = E tr {Fm p} = trp = 1. (2.148)

mem mem
The von Neumann-Liiders scheme for an ideal quantum measurement is ob-
viously a special case of the above general setting. Namely, taking the orthogonal
decomposition of unity AE, and considering M to be the set of intervals Ar,
of that decomposition we find for the measurement outcome in = Ar„ the effect
 THE THEORY OF QUANTUM MEASUREMENT 87

,=
Pm i+,,)(13 m(P) P' = Ern (Drn(P)
}

P(m) = tr{ F„p}

readout m

FIG. 2.2. Illustration of the operation 4), and the effect Fm, for a generalized
measurement scheme.

Fin = AE, and the operation t(p) = AE,pAE,. Furthermore, the complete-
ness relation (2.13) ensures that eqn (2.142) holds, whereas (2.145) follows from
the fact that the AE, are projection operators.
As a natural generalization of this example one may consider an operation
given by

QmPQ rt n,

and the corresponding effect

Fin = 1-2 17n 1-2„,

where the 52„ are linear operators in the underlying Hilbert space satisfying the
normalization condition

E Fin = E 1-217,12, = I. (2.151)

mEM mEM

As we have seen, an ideal quantum measurement gives rise to a projection-

valued measure which associates with each measurement outcome m = Ar, a
projection operator F, = AE,. The above concept of generalized measurement
theory leads, in a natural way, to the more general idea of a positive operator-
valued measure which associates with each measurement outcome m a positive
operator Fin =1-2tin f2„. A specific example of the above general concepts will be
discussed in Subsections 2.4.5 and 2.4.6.

2.4.3 Representation theorem for quantum operations

For practical applications it is often very helpful to work with an explicit rep-
resentation for the operations cl.„. Such a representation is provided by an im-
portant theorem which is due to Kraus (1983). We first formulate the three
conditions underlying this theorem and then briefly sketch the proof.
We consider a collection of positive maps (1),n (p) labelled by the possible
outcomes m of a generalized measurement. Each of these maps is required to
fulfil the following three conditions.
88 QUANTUM PROBABILITY

1. Since we intend to interpret P(rn) = tr I(p) as the probability of the

outcome m we first demand that

0< tr (1)(p) < 1. (2.152)

2. The map al,n, is further required to be convex linear, that is for any collec-
tion of density matrices pi and non-negative numbers pi > 0 with Ei pi = 1
we require

cb, (Pipi) -= EPicl'rn(Pi)• (2.153)

This condition may be motivated with the help of Bayes's theorem (1.11).
To this end, we suppose that the state p = Ei
pi pi describes an ensemble
which is obtained by preparing the states pi with probabilities pi. We then
expect that the following relation holds,

cDm(P) 4(p)
(2.154)
tr 4,m(p) = EP(4m) tr,Drn (A).

The left-hand side represents the state of the total mixture under the con-
dition that the outcome m occurred. The quantity 4.,,(pi)/tr cl.,,(pi) gives
the state after the outcome in occurred under the condition that the state
pi has been prepared. This state is multiplied with the probability p(ilm)
which is the probability that the state pi has been prepared under the
condition that the outcome in occurred, that is under the condition that
the operation (1),, acts on the state. Applying Bayes's theorem (1.11) we
obtain

Pz
Ailm) P(Tnli) tr in (p) tr cDrn (PO Pz (2.155)

where p(mi) = tr cb m (p,) is the probability of the occurrence of the out-

come in under the condition that pi has been prepared. Note also that pi
is the unconditioned probability of the preparation of pi , while tr cl, ,,(p) is
the unconditioned probability of the occurrence of the outcome in. Sub-
stituting (2.155) into (2.154) leads to the convex linearity (2.153) of the
operation.
3. Finally, the map 4, is required to be completely positive. This condition is
much stronger than positivity (for an example, see Gorini, Kossakowski and
Sudarshan, 1976) and can be formulated as follows. Positivity of (1), means
that (1),(A) is positive for any positive operator A acting on the Hilbert
space 9-1. Consider now a second Hilbert space 9-1 of arbitrary dimension
 THE THEORY OF QUANTUM MEASUREMENT 89

D. We can then define a combined operation rn ®/ which acts on the

tensor product 74 0 74 as follows,

( m ®I) (E /l a Ba ) = 4,m(Aa) Ba . (2.156)

a a

Complete positivity means that not only is 4), positive, but also the com-
bined operation (1),, 01 for all dimensions D, that is (I., maps positive
operators of the composite system to positive operators. Physically, this is
a reasonable condition since the combined operation cbm 0I may be viewed
as an operation which operates locally on the first of two widely separated
systems without influencing the second system.
The representation theorem for quantum operations now states that an oper-
ation 4), satisfies the above three conditions if and only if there exist a countable
set of operators Q m k such that the operation can be written as

(krn(P) = E QmkPQ rnt (2.157)

while the effects satisfy

= E Qrnt < I. (2.158)

These are the most general forms for operations and effects.
We briefly sketch the proof of the representation theorem for finite-dimen-
sional spaces. Suppose first that we have an operation (1. m which is given by
operators Q m k as in eqns (2.157) and (2.158). Then (D in obviously satisfies the
three conditions given above. For example, complete positivity of (I),T, follows
immediately since

(Qrnk I) ilAnt -010 (VmklAkarnk) > 0 (2.159)

for positive operators A acting on the product 94 where we have defined
kornk) (Qtrnk Pp).
Conversely, suppose we have an operation (D rn satisfying the above three
conditions. Let us give an explicit construction for the operators Q m k. To this
end, we take an auxiliary Hilbert space 74 of the same dimension as the original
space 9-1, and choose fixed orthonormal bases {Xi)} and {IV} in 9-1 and 94,
respectively. Next we introduce the maximally entangled state

(2.160)

and define the operator

am (cD, ® (1(p)(vI) • (2.161)
We note that o-n, is an operator in the tensor product space pertaining to the
combined system. By virtue of the complete positivity of (1),, this operator is
 90 QUANTUM PROBABILITY

positive since l(p)((pl is positive. Therefore, am has a spectral decomposition of

the form

am E Ivrfloopmk • (2.162)

The states 11 (pink) are states of the composite system and are, in general, not
normalized. Finally, for any state

(2.163)

in 9-1 we introduce a corresponding state in 94 by means of

Iro) (2.164)

For each index k the equation

1 rnk1 11) ) ( 17*Prnk) = EaAikprno (2.165)

then defines a linear operator on 9-1 . We now prove that the collection of operators
O ink represents the operation (I. rn in accordance with eqn (2.157). To this end,
we first consider the pure case. With the help of our definitions we obtain

E ornkloomwrnk Eo-plvmk)(0mk = (010-rnlq)

no(p)(pDlo)
(m. 0
= ENomoxixxi 1) O Ui)U./16)

cDrn(lXi)(XiDa;ai. (2.166)

Hence, we have

cDrn(10)(01) = EQrnk*)(019trnle (2.167)

This proves the representation (2.157) for pure states. The general case easily
follows with the help of the convex linearity of (1. m . Finally, the inequality in
(2.158) is a direct consequence of the condition (2.152).
The representation theorem shows that a completely positive quantum oper-
ation cb m can be represented in terms of a set of operators Qmk • This gives rise
to the interesting question concerning the freedom in the choice of the operators
 THE THEORY OF QUANTUM MEASUREMENT 91

S-2 in k . Consider two sets Ift ink l and {0,k } of operators which represent opera-
tions (b m and ci, ,, respectively, in accordance with the representation theorem,

cDrn(P)--= EQmkpf2t,,, (2.168)

Ln(P) = Ef2mkpf2tink . (2.169)

Ic

Adding zero operators to the smaller set we can always ensure that both sets
contain the same number of elements. Then both operations are equal, that is
cb ni --= $ n„ if and only if there exists a unitary matrix u --= (uki) such that

Qmk --Euk/omi. (2.170)

Thus, two quantum operations are equal to each other if and only if their cor-
responding Il-operators are related through a linear combination involving a
unitary coefficient matrix.
To prove the above statement let us first suppose that (2.170) is satisfied. By
use of the representations (2.168), (2.169) it follows immediately that cD m --=
Conversely, assuming (b m --= 0i, we have

E , mkpf,trnk E fimkpo tink (2.171)

for all densities p. Using the same construction and the same notation as in the
proof of the representation theorem we define

l'Prnk) = (Qrnk ® -T)R°)) rnk) = (Ornk ® (2.172)

which yields

l'Prnk)(Pmk1 = Pmk)(137nkl• (2.173)

This shows that the positive operator am is generated by the two sets of states
l'Prnk)} and {Wo rnk)}. These states are therefore related through

koink) =Euk/ int) (2.174)

with some unitary matrix u (see eqn (2.56)). For arbitrary 110 we have

Qink1 11) ) --= ( 1'; bPrnk) --= Euk1(0- 1(-pm,/)--= Euk1f4n/p), (2.175)

which gives eqn (2.170), as was to be shown.

An important implication of the above statement is that any operation cb m
acting on the state space of a system with a D-dimensional Hilbert space can be
represented with the help of at most D 2 operators 12mk.
92 QUANTUM PROBABILITY

2.4.4 Quantum measurement and entropy

In general, a quantum measurement induces a change of the von Neumann en-
tropy. We define the entropy change in a non-selective quantum measurement
described by operations €1., as

AS = S(d) — S(p) = S ( m (p)) — S(p). (2.176)

771

Consider first an ideal quantum measurement in which case (1.,(p) =

with orthogonal projections Hm , satisfying Ern Hm = I. From the Klein inequal-
ity (2.115) we get

0 < S(P11P') = — S(P) tr{P In P'}. (2.177)

With the help of the completeness relation and of the fact that p' = Ern lim pH,
commutes with the projections Hm , the second term on the right-hand side of
the equality sign is found to be

—tr{pinp'} = —tr { E li m p On p') II,

m

= —tr {E 11,p11,2 ln p'}

= —tr{p'lnp'}
= S (P'), (2.178)
and, hence,

AS > 0. (2.179)

We conclude that non-selective, ideal quantum measurements never decrease the

von Neumann entropy. We note also that AS = 0 if and only if p' --= p, that is if
and only if the density matrix is left unchanged by the measurement.
The above conclusion is, in general, not true for a generalized measurement
as can be seen from the following example. We consider a two-level system the
Hilbert space of which is spanned by two states le) and 1g). We define the two
operations (I),(p) Q m pfftni , m = 1, 2, where 1 2 1 = 1g)(g1 and St 2 = g (el. Thus
we have the effects 12112 1 = 1g)(g1 and 12 . S1 2 = le) (el, and their sum adds to
the identity I. However, for a normalized initial density matrix p we find the
following state after the non-selective measurement,

10 ' = PQt + PW2 = 19)(91. (2.180)

The final state is therefore always a pure state and S(p') = 0. Thus, the entropy
decreases to zero if we start from an initial density with positive entropy. This
 THE THEORY OF QUANTUM MEASUREMENT 93

example shows that a generalized measurement can decrease the von Neumann
entropy.
Another interesting question concerns the change of entropy involved in the
transition from the selective to the non-selective level of a measurement. We
define the quantity

SS = S (EP(m)6,) — (2.181)
rn in

which may be referred to as the mixing entropy. It is the entropy of the final
density matrix 11 = Ern P(17)pInt on the non-selective level minus the average
of the entropies of the sub-ensembles described by the states p'in . This mixing
entropy satisfies the inequalities

0 < OS < 11 ({P(m)}). (2.182)

Thus, SS is always non-negative. This statement follows immediately from the

concavity of the von Neumann entropy (see eqn (2.110)). The second inequal-
ity shows that OS is never larger than the Shannon entropy associated with the
random variable m following the distribution m i-- P(m). The proof of this state-
ment may be found, for example, in Nielsen and Chuang (2000). We can interpret
the Shannon entropy H({P(m)}) as the information lost in the transition from
the selective to the non-selective level, since during the mixing process we for-
get about the measurement results. In other words, the Shannon entropy of the
distribution P(m) is a measure of the information gained in the measurement
by obtaining a specific realization of the random variable m. Thus we see that
the mixing entropy is never larger than the information gained. It can be shown
further that the mixing entropy is equal to the Shannon entropy if and only if
the densities p' orthogonal images. This occurs for an ideal measurement,
in which case the mixing entropy becomes equal to the Shannon entropy.

2.4.5 Approximate measurements

The generalized framework for the description of quantum measurements intro-
duced above is capable of describing, for example, approximate measurements,
that is, measurement devices which measure the spectrum of some observable _h
only with finite resolution (Braginsky and Khalili, 1992).
We consider some physical situation in which an observable R discrete,
non-degenerate spectrum,

(2.183)

is measured. If the eigenvalues r, are not too closely spaced one can measure
such an observable exactly, provided the resolution of the experimental device is
sufficiently large. A measurement which is only approximate arises, however, if
94 QUANTUM PROBABILITY

the resolution of the apparatus is not high enough to distinguish, for example,
neighbouring eigenvalues. For such a case we introduce a conditional probabil-
ity distribution W(mlin') which represents the probability for the measurement
to yield the outcome m provided the measured system is known to be in the
state Ow with eigenvalue In f . This conditional probability distribution serves to
describe the finite resolution of the apparatus, possible disturbances by the en-
vironment, as well as possible uncertainties in the precise state of the apparatus
before the measurement.
Assuming that the apparatus always yields a definite outcome we have

Ew(rnlini) = 1. (2.184)
171

The probability distribution for the measurement outcomes r m then takes the
form

P(m) := E woontxorn , iolorni), (2.185)

which is obviously normalized to 1,

E P(m) --= 1. (2.186)

771

Defining the effect operator

Frn _= E lonoworont)(orn, (2.187)

rn'

we can rewrite the probability distribution of the outcomes as

P(m) --= tr {Fin p} . (2.188)

As is easily verified, (2.187) is a positive operator and satisfies the normalization

condition (2.142).
To find an appropriate operation we invoke the ansatz (2.149) and conclude
from the consistency condition (2.145) the relation (2.150). The most general
solution of this relation takes the form

Q in = UrnFrn112 ,

where Urn is an (undetermined) unitary operator and

Fri?" 2 = E kprn, Ww(nont)(orni 1

ne

is the square root of Fm . Thus we have the operation

 THE THEORY OF QUANTUM MEASUREMENT 95

(1)(p) =Um Fm112 pFm112 U;tn , (2.191)

showing that the transition from the initial density matrix p to that describing
the sub-ensemble conditioned on the outcome r, may be viewed formally as a
two-step process:

p P(m) -1 F,,112 pFm1/ 2 pir'n =Um p',A. (2.192)

The second step in (2.192) represents a unitary transformation of the density

matrix. During this step the entropy thus remains constant and no information
on the quantum system is gained. The precise form of the unitary operator Um
is determined by the details of the measuring device. The above analysis shows
that one cannot determine the exact form of the operation from the probability
distribution of the measurement outcomes alone. As we shall see, the unitary op-
erator Urn describes the disturbance of the quantum object by the measurement
apparatus (see Section 2.4.7).
The first step in (2.192) describes the gain of information during the measure-
ment: It is this step which splits, in general, the initial ensemble into the various
sub-ensembles. The effect F, is uniquely determined by the spectral family of
and the conditional probability distribution W(indin') that describes the finite
resolution of the measuring device. Since R and F, have the same spectral fam-
ily they commute, [Frn ,RJ = O. As we shall see this is a sufficient condition for
the first step to be a quantum non-demolition measurement.
In the case of an infinite resolution we get W(mIrre) = and, therefore,
the effect reduces to that of an ideal measurement of R namely F,
,

Accordingly, the operation takes a form which is similar to the one given in the
von Neumann-Li_iders postulate, namely

P'r'n = rnki) rn) ml n (2.193)

where the projection is, in general, accompanied by an unitary transformation.

As another example we consider the approximate measurement of the position
operator Q (Davies, 1976). The probability density p(q) for the outcome q is taken
to be

p(q) = f dq' W(q (2.194)

Here, the distribution W(q - q') yields the conditional probability density for the
apparatus to respond with the value q provided the system is in the (generalized)
eigenstate IC of Q. Note that an exact position measurement would be obtained
if W(q) approaches the Dirac delta function. The corresponding effect reads

Fq _-= f dq' q')14/(q - ql(q`1. (2.195)

96 QUANTUM PROBABILITY

ideal
quantum measurement
quantum classical
of _h
object probe measurement
Po
U(7, 0) PP device
Qmk

readout m

FIG. 2.3. Schematic picture of an indirect measurement.

It is interesting to determine the mean value E(Q, W) and the variance

Var(Q, W) of the measurement outcomes in this approximate measurement. We
find with the help of the distribution p(q),

E(Q, W) = E(Q) + f dq qW(q), (2.196)

2
Var(Q, W) = Var(Q) + f dq q 2 W (q) — [f dq qW (q)] , (2.197)

where, as usual, E(Q) and Var(Q) denote the mean and the variance as they
would be obtained in an ideal measurement of Q. The first equation shows that
a non-vanishing mean value of W(q) leads to a systematic bias of the measure-
ment results. The second equation tells us that a finite variance of the distribu-
tion W(q) yields an additional dispersion which is superimposed on the intrinsic
quantum fluctuations given by Var(Q).

2.4.6 Indirect quantum measurements

An important measurement scheme which can also be treated within the frame-
work of the generalized theory of quantum measurements is provided by the
concept of an indirect measurement (Braginsky and Khalili, 1992).
An indirect measurement can be viewed as consisting of three elements (see
Fig. 2.3). The first element is the quantum system to be measured, that is the
system from which information is to be extracted. This system is therefore called
the quantum object and will be considered to live in the Hilbert space 7 10. The
-

second element is the so-called quantum probe with Hilbert space Np. The quan-
tum probe is again a certain quantum system which interacts with the quantum
object. It is assumed that the probe system has been prepared, prior to this in-
teraction, in a suitable state given by some density matrix pp. As a result of the
subsequent interaction, correlations between object and probe are built up. The
third element of the scheme is a classical apparatus by which a measurement
 THE THEORY OF QUANTUM MEASUREMENT 97

on the quantum probe is performed when the interaction between object and
probe is over. The aim of this scheme is to obtain information on the state of the
object by means of the measurement on the probe. This information is deduced
from the correlations between object and probe built up during their interaction.
Thus we can formulate three basic requirements for an ideal measurement:
1. Before the interaction between object and probe at time t --= 0, the probe
has been prepared in a well-defined quantum state pp. At the same time
the quantum object is in a state po.
2. The interaction between quantum object and probe starts at time t = 0
and is finished at some time t = T > 0 before the measurement by means
of the apparatus sets in.
3. The third element of the scheme represents a classical apparatus and the
measurement on the quantum probe can be described as an ideal measure-
ment according to the von Neumann—Liiders projection postulate.
In the following we shall derive explicit expressions for the effect and the opera-
tion pertaining to such an ideal measurement scheme.
At time t --= 0 the total quantum system consisting of object and probe is
described by the density matrix po 0 pp in the total Hilbert space given by the
tensor product 1-1 - = 1-10 0 Hp. The Hamiltonian of the total system is taken to
be

11(t) = Ho + Hp + HIM, (2.198)

where Ho and Hp describe the free evolution of object and probe, respectively,
and 1//(t) their interaction. This time-dependent interaction is assumed to vanish
outside the time interval [0, T]. The corresponding unitary operator describing
the time evolution over this interval according to the Schrödinger equation will
be denoted by U,

U U(T, 0) --= T exp [--i f dt1/(t)] , (2.199)

o
where T denotes the chronological time ordering and we have set h = 1.
The initial density matrix p(0) , po 0 pp of the total system thus evolves
after time T into

p(T) --- u (po ® pp) (it. (2.200)

We now assume that at time T the classical apparatus measures the probe ob-
servable

-h — E rrn korn)(4,orn1 (2.201)

m
98 QUANTUM PROBABILITY

with a non-degenerate, discrete spectrum. Note that _h is a self-adjoint operator

which acts in the Hilbert space 7-1p of the quantum probe. The probability P(m)
for the outcome Tm of this measurement is found to be

P(m) = tr fb,orn)(PrnIP(7)} = tr { utkorn)((pmlu (Po 0 Pp)} , (2.202)

where the trace is taken over the total Hilbert space 9-1. The partial traces over
'Ho and 1-tp will be denoted by tro and trp, respectively. The expression for
P(m) can then be rewritten as

P(Trt) = tro {Fin po} , (2.203)

where the effect takes the form

Frnpo = trp { Ut kPm)( (Pm r (PO 0 pp)} • (2.204)

As is easily verified Fm is a positive operator which acts in the Hilbert space 9-to
of the quantum object and which satisfies the normalization condition, namely
we have

E Fin po = E trp {U kOrn) (C)in 1U (PO 0 PP)} = trP {po 0 pp} = po,
(2.205)
showing that (2.142) is fulfilled.
The operation for the indirect measurement is obtained by applying the von
Neumann-Li_iders projection postulate to the measurement of the probe which
immediately yields the following state of the quantum object:

P'm = P(111 ) -1 (PrnIU (PO 0 PP) Ut kom). (2.206)

Introducing the spectral decomposition of the density matrix of the probe,

PP = P 10k)(Okl, (2.207)

we can write the operation on the quantum object as

cDm(P0) = EQmkpoRnt k , (2.208)

k

where

Qmk = N/Pk(Prn1 11 10k)• (2.209)

On using these operators the effect is given by

Fm = mk • (2.210)
mk
k
 THE THEORY OF QUANTUM MEASUREMENT 99

Thus we observe that operation and effect take exactly the form given in the
representation theorem, see eqns (2.157) and (2.158). If initially the probe is in
a pure state pp = 10) (01, we get

12 in = (ornIU10) (2.211)
which is proportional to the amplitude for a transition of the probe from the
state 0 to the eigenstate (pin of the measured observable R. Note that Sl, is an
operator in the Hilbert space of the quantum object which describes the change
of the state of the object induced by the outcome T., of the measurement on
the probe. The additional sum over the index k in the general expression (2.208)
results from the additional ignorance in the case that the initial state of the
probe is a statistical mixture.
The above results are readily generalized to the case of an observable 1-
with a continuous spectrum. Let us work out a specific example. We take 1-
to be the momentum operator P in a specific direction, that is, the classical
measurement apparatus measures a certain component of the momentum of some
probe particle. Our aim is to measure by means of an indirect measurement the
non-degenerate, discrete observable

A --= Eavlxv)(x,1 (2.212)

v

of the quantum object. To this end, we couple the position operator Q canonically
conjugate to P,

(2.213)

to the observable A, that is, we consider the interaction Hamiltonian

HT (t) = g(t)Â, (2.214)

where the time-dependent coupling g(t) is assumed to vanish outside the time
interval [0, T]. For simplicity we further assume that the interaction time 7 is
very small and that the free evolution generated by Ho + Hp may be neglected
over this time. The unitary time-evolution operator then takes the simple form

U --= exp [—iGA(2] (2.215)

where
T

G _-- f dt g(t) (2.216)

a
denotes the integrated coupling strength. The quantum probe is assumed to be
in the initial state pp --= 1(0(01 such that we obtain for the operator 11p pertaining
to the measurement outcome p (see eqn (2.211))
100 QUANTUM PROBABILITY

(WO)
= exP [ — iGM2] 10)
E lXv)(Plexp [ — iGav(2] 10)(XvL (2.217)

The matrix element in the last expression is equal to 0(p + Gap), where 0(p) E-
(p10) is the initial wave function of the probe in the momentum representation.
Hence we have

Xv)(/)(P + Gav)(xv1= o( p
+ G A) (2.218)
Ii

It follows that the probability density f (p) for the momentum measurement (to
avoid confusion with the momentum we denote here probability densities by f)
is given by the expression

f (p) = tr O pt QpPol = E 10(p + cav)12(xvIpolxv). (2.219)

Ii

The last two equations can now be interpreted as describing an approximate

measurement of the observable A. To see this we must answer the following
question: What information on the observable A of the quantum object can
be inferred from the measurement of P? Suppose for a moment that 0(p)1 2 is
sharply peaked around its mean value

(p) E (01 16 10) E f dP 13 10(13)1 2 (2.220)

Looking at eqn (2.218) we then see that the measurement of p effectively projects
the state of the quantum object onto a group of eigenstates x v of A whose eigen-
values av satisfy p + Ga,, (p). The precise number of states in the group is
determined by the widths of the probe's wave function in the momentum rep-
resentation. We thus define the a-value which is inferred from the measurement
outcome p by means of the relation p + Ga --= (p), that is,
1
a ((P) P) • (2.221)

We then get for the probability density of a

f (a) = E W(a av)(Xv1P01Xv), (2.222)

z,

where

W(a — av) = PI • OM — G[a — avi)1 2 (2.223)

We thus observe that the direct measurement of the momentum P of the probe
particle yields an approximate measurement of the observable A of the quantum
 THE THEORY OF QUANTUM MEASUREMENT 101

object. Expression (2.223) represents the conditional probability density of this

approximate measurement of A: It yields the density for obtaining the value a
provided the object is in an eigenstate with eigenvalue a,.
Clearly, W(a — a v ) is normalized to 1,

f da W (a — ay ) = 1, (2.224)

and satisfies

f da aW (a — ay ) = ay , (2.225)

f da a2 W (a — ay ) = a2, + —
1 Var(P) ,
(2.226)
G2

where Var(P) denotes the variance of I" in the initial state (/) of the probe. It
follows from eqn (2.225) that the expression (2.221) for the value of a inferred
from the measurement outcome p has a vanishing bias, that is, its expectation
value coincides with the exact mean value of A,

E(a) f da a f (a) = tr {Apo} = E(A). (2.227)

Equation (2.226) on the other hand leads to the relation

2
Var(a) E f da a 2 1 (a) — [f da af (a)]
1
= Var(A) + Var(P). (2.228)

This equation tells us that the variance of the inferred value of a equals the
sum of the exact quantum variance Var(A) of A and the variance of the probe
momentum in the initial state divided by G2 . It follows that the measurement
is the more accurate the lower the momentum uncertainty of the probe and the
stronger the coupling between quantum object and probe.
Having obtained a particular value a, the corresponding operation is given
by the operator

9a lx v ) AAGIO((p) — G[a — ad)(x v i = VIGO ((p) — G(a — Â)). (2.229)

i,

The density matrix of the resulting sub-ensemble then takes the form

/ I» lx,)0((p) — G[a — av])(x,IPolxvi)0*((P) — Gla — av, l)(xvi I

Pa =
E, 10((p) — G[a —
av1)1 2 (Xv1Polxv)
flaP0 9 a
(2.230)
— tr{Qtat/aP0}.
102 QUANTUM PROBABILITY

In the ideal case of an infinitely sharp initial momentum Po of the probe we

have 10(p)1 2 ---> Op — p o ) and, therefore,

.f(a) = Ev Oa — av)(XvP01Xv), (2.231)

which shows that a takes on only a discrete set of values given by the eigenvalues
of the operator A. The corresponding probabilities

P(a) = (xvIPolxv) (2.232)

as well as the new density matrix

P a,= lxv)(x,1 (2.233)

are given by the correct quantum mechanical formulae corresponding to an ideal

quantum measurement of A.

2.4.7 Quantum non-demolition measurements

Suppose an ideal, non-selective measurement of some observable A is performed
on a quantum system in the state p. As is well known if A commutes with p then
the state of the system is not changed by the measurement. In fact, if

A (2.234)
z,

represents the spectral decomposition of A, the a,, denoting the eigenvalues and
Hy the corresponding projections, then we find for the non-selective density
matrix p' after the measurement

II = E H11 " = E 1LP = P, (2.235)

where we have used the completeness relation as well as the fact that p commutes
with the projections H.
Consider now some generalized measurement scheme with operations (1 , 77„
defined in terms of operators Q„,, (see eqn (2.149)). Such a scheme is defined
to be a quantum non-demolition measurement (QND measurement) (Braginsky
and Khalili, 1992) for some observable A if the probability distribution for an
ideal measurement of A remains unchanged during the measurement process,
that is if the distribution of the eigenvalues of A is the same in the initial and
the final density matrix. Here, the generalized measurement is assumed to be a
non-selective one. We remark that QND measurement devices are important for
the design of high-precision quantum measurements (Bock() and Onofrio, 1996).
 THE THEORY OF QUANTUM MEASUREMENT 103

On using expression (2.149) for the operation of a generalized measurement

we can cast the above definition into the following mathematical form,

tr {Â' p} = tr {AY} = E tr {A k Qm pf2tn,} . (2.236)

M,

This means that all moments of A coincide in the initial and the final state and,
therefore, also the distribution of its eigenvalues (assuming for simplicity that the
moments exist). For the above example of an ideal measurement of A the II, are
given by the projections Hy and condition (2.236) is obviously satisfied provided
p commutes with the 11 ,. This is a condition for the state of the quantum system
on which the measurement is performed.
Alternatively, we can read eqn (2.236) as a condition for the operators 11„,
which is required to hold for all states p of the quantum system. Employing the
normalization condition (2.151) we can cast eqn (2.236) into the form

E tr {Wm, [A k , S2 m,] p}= 0. (2.237)

171

Since this condition is assumed to hold for all p we conclude that a sufficient
condition for a QND measurement is given by

[A, O m] = 0. (2.238)

A measurement is therefore a QND measurement for the observable A if the

operators 11 „,, describing the change of the quantum system induced by the mea-
surement commute with A. Condition (2.238) is always fulfilled for an ideal
measurement of the observable A, that is an ideal measurement of A is always
a QND measurement in the sense defined above. We also see immediately from
eqn (2.229) that the example for an indirect measurement scheme discussed in
the preceding subsection fulfils the QND condition (2.238). This is due to the fact
that Sla, is diagonal in the A-representation which, in turn, results from the fact
that the interaction Hamiltonian 1//(t) commutes with the measured quantity.
As a result, A is a constant of motion during the measurement process.
Let us study the general setting for an indirect measurement with a probe in
a pure state O. Substituting (2.211) into (2.238) we find

(ciom I [A, U(7, 0)] 10) = 0. (2.239)

Since this must be true for all in and because the (p, represent a basis in the
Hilbert space li p of the probe, we conclude that

[A, U(r, 0)] 10) =0. (2.240)

This equation can possibly be achieved by preparing the probe in an appropriate
state O. Since this is obviously a difficult task in practice one requires that the
commutator itself vanishes, which leads to the following sufficient QND condition
104 QUANTUM PROBABILITY

[A, U(r, 0)1 = 0. (2.241)

According to this equation the Heisenberg picture operator

AH(t) = Ut(t, 0)AU(t, 0) (2.242)

returns to its original value after time t = T, that is, AH(r) = AH(0). The easiest
way to achieve this is to require that A commutes with the total Hamiltonian
11(t) = Ho ± Hp ± 111(t), that is

[A, Ho + I-1/ (t)] =0. (2.243)

If A is a constant of motion under the free evolution of the quantum object, a

general sufficient condition for a QND measurement is found to be

[A, Hi(t)] = O. (2.244)

This condition is also known as the back-action evasion condition, for it guaran-
tees that the interaction of the quantum object with the probe system does not
disturb the measured quantity.
We remark finally that by the same reasoning one can show that a suffi-
cient condition for the approximate measurement discussed in Subsection 2.4.5
to be a QND measurement is given by [f?, Urn] = 0, which means that the
observable measured approximately by the scheme must commute with the uni-
tary operators Urn that occur in the second step of (2.192). Thus, these unitary
transformations disturb, in general, the probability distribution of the measured
quantity.

References
Akhiezer, N. I. and Glazman, I. M. (1981). Theory of Linear Operators in
Hilbert Space. Pitman, Boston.
Bocko, M. F. and Onofrio, R. (1996). On the measurement of a weak classical
force coupled to a harmonic oscillator: Experimental progress. Rev. Mod.
Phys., 68, 775 - 799.
Bohm, A. (1993). Quantum Mechanics (third edition). Springer-Verlag, New
York.
Braginsky, V. B. and Khalili, F. Ya. (1992). Quantum Measurement. Cambridge
University Press, Cambridge.
Cohen-Tannoudji, C., Diu, B. and Laloë, F. (1977). Quantum Mechanics, Vol-
ume I and II. John Wiley, New York.
Davies, E. B. (1976). Quantum Theory of Open Systems. Academic Press,
London.
Gorini, V., Kossakowski, A. and Sudarshan, E. C. G. (1976). Completely posi-
tive dynamical semigroups of N-level systems. J. Math. Phys., 17, 821-825.
 REFERENCES 105

Gottfried, K. (1974). Quantum Mechanics. W. A. Benjamin, Reading, Mas-

sachusetts.
Gudder, S. P. (1979). Stochastic Methods in Quantum Mechanics. Elsevier,
North Holland, New York.
Holevo, A. S. (1982). Probabilistic and Statistical Aspects of Quantum Theory,
Volume I of North-Holland Series in Statistics and Probability. North-Holland,
Amsterdam.
Hughston, L. P., Jozsa, R. and Wootters, W. K. (1993). A complete classifica-
tion of quantum ensembles having a given density matrix. Phys. Lett., A183,
14-18.
Kraus, K. (1983). States, Effects, and Operations, Volume 190 of Lecture Notes
in Physics. Springer-Verlag, Berlin.
Landau, L. D. and Lifshitz, E. M. (1958). Quantum Mechanics. Pergamon
Press, London.
Langerholc, J. (1965). Trace formalism for quantum mechanical expectation
values. J. Math. Phys., 6, 1210-1218.
Lieb, E. H. (1973). Proof of the strong subadditivity of quantum-mechanical
entropy. J. Math. Phys., 14, 1938-1941.
Louise 11, W. H. (1990). Quantum Statistical Properties of Radiation. John
Wiley, New York.
Liiders, G. (1951). Über die Zustandshnderung durch den Mel3proze13. Ann.
Phys. (Leipzig), 8, 322-328.
Nelson, E. (1967). Dynamical Theories of Brownian Motion. Princeton Uni-
versity Press, Princeton.
Nielsen, M. A. and Chuang, I. L. (2000). Quantum Computation and Quantum
Information. Cambridge University Press, Cambridge.
von Neumann, J. (1955). Mathematical Foundations of Quantum Mechanics.
Princeton University Press, Princeton.
Werl, A. (1978). General properties of entropy. Rev. Mod. Phys., 50, 221-260.
 Part II
Density matrix theory
 3
QUANTUM MASTER EQUATIONS

In contrast to the case of a closed systems, the quantum dynamics of an open

system cannot, in general, be represented in terms of a unitary time evolution.
In many cases it turns out to be useful to formulate, instead, the dynamics of
an open system by means of an appropriate equation of motion for its density
matrix, a quantum master equation. In this chapter we shall give a survey of the
various types of master equations which can be employed to analyse the quantum
dynamics of open systems. The emphasis lies on the Markovian dynamics of open
systems. A systematic treatment of non-Markovian quantum processes will be
presented in Part IV.
To begin with, we give a general characterization of the dynamics of closed
and open quantum systems. Quantum Markov processes are then introduced
which represent the simplest case of the dynamics of open systems. They can
be regarded, essentially, as a direct generalization of the classical probabilistic
concept of a dynamical semigroup to quantum mechanics. In analogy to the dif-
ferential Chapman—Kolmogorov equation of classical probability theory, a quan-
tum dynamical semigroup gives rise to a first-order linear differential equation
for the reduced density matrix, which is known as quantum Markovian master
equation in Lindblad form.
A number of derivations of Markovian density matrix equations will be given
which start from various microscopic models, such as the weak-coupling inter-
action of radiation with matter and the Caldeira—Leggett model. The resultant
density matrix equations enable the investigation of general features like irre-
versibility, entropy production, and relaxation to equilibrium. In standard ap-
plications of the dynamics of open systems two limiting cases are of particular
relevance. These are the quantum optical and the quantum Brownian motion
master equation. We shall give a number of examples and discuss basic properties
of the corresponding master equations. For later purposes the Caldeira—Leggett
model will also be studied from a broader perspective using the exact Heisenberg
equations of motion as well as influence functional and path integral techniques.
In the treatment of open many-body systems non-linear quantum master
equations are also encountered. If one takes into account particle correlations on
the mean-field level the reduced density matrix equation involves a non-linear
time-evolution generator. Under certain conditions the solutions of such non-
linear density matrix equations may be obtained with the help of a non-linear
Schrödinger-type equation. As two prominent examples for mean-field master
equations we treat the laser equations and the phenomenon of super-radiance.
110 QUANTUM MASTER, EQUATIONS

3.1 Closed and open quantum systems

3.1.1 The Liouville - von Neumann equation
According to quantum mechanics the state vector 10(t)) evolves in time according
to the Schrödinger equation,
d
i(t)) = H(t)Mt)), (3.1)

where _H(t) is the Hamiltonian of the system and Planck's constant h has been
set equal to 1. The solution of the Schr6dinger equation may be represented in
terms of the unitary time-evolution operator U(t, to ) which transforms the state
*(to)) at some initial time t o to the state 1/)(t)) at time t,

10(t)) = U(t,to )mt o )). (3.2)

If we substitute expression (3.2) into the Schr6dinger equation (3.1) we get an

operator equation for the time-evolution operator U(t, to ),

U(t, to ) = H(t)U(t, to), (3.3)

which is subjected to the initial condition

U(to , to ) = I. (3.4)

It is easy to show with the help of (3.3) and (3.4) that Ut(t, to)U(t, to) =
U(t, to)Ut(t, to ) I and, hence, that U(t, to) is a unitary operator.
For a closed, isolated physical system the Hamiltonian H is time independent
and eqn (3.3) is readily integrated to yield the well-known expression

U(t, to) = exp [—iH(t — to)l. (3.5)

In physical applications one often encounters the situation that the system under
consideration is driven by external forces, an external electromagnetic field for
example. If in such a case the dynamics of the system can still be formulated in
terms of a possibly time-dependent Hamiltonian generator H(t) the system will
again be said to be closed, while we reserve the term isolated to mean that the
Hamiltonian of the system is time independent. For a time-dependent Hamil-
tonian the solution of eqn (3.3) subjected to the initial condition (3.4) may be
represented as a time-ordered exponential,

U(t, to ) = T, exp [ — i f dsH(s) , (3.6)

to
where T, denotes the chronological time-ordering operator which orders prod-
ucts of time-dependent operators such that their time-arguments increase from
right to left as indicated by the arrow.
 CLOSED AND OPEN QUANTUM SYSTEMS 111

If the system under consideration is in a mixed state the corresponding quan-

tum statistical ensemble may be characterized with the help of the statistical
operator p. It is straightforward to derive an equation of motion for the density
matrix starting from the SchrEdinger equation (3.1). Let us assume that at some
initial time to the state of the system is characterized by the density matrix (see
Section 2.1.3)

P(t o ) — w.10, ( to )) (0a(t 0 )1, (3.7)

a

where the w, are positive weights and the 10 „(to)) are normalized state vectors
which evolve in time according to the Schralinger equation (3.2). The state of
the system at time t will therefore be given by

P(t) = Ewau(t,t0)10a(toopa(tort(t, to), (3.8)

a

which may be written more concisely as

p(t) = U(t,to )p(to )Uf (t, to ). (3.9)
Differentiating this equation with respect to time we immediately get an equation
of motion for the density matrix,
d
— p(t) = —i[H (t), p(t)], (3.10)
dt
which is often referred to as the von Neumann or Liouville—von Neumann equa-
tion.
In order to stress the analogy of the von Neumann equation with the cor-
responding equation of motion for the probability density in classical statistical
mechanics it is often written in a form analogous to the classical Liouville equa-
tion,
d
— p(t) = r(t)p(t). (3.11)
dt
Here, f is the Liouville operator which is defined through the condition that
.C(t)p is equal to —i times the commutator of H(t) with p. More precisely, f
is often called a Liouville super-operator since it acts on an operator to yield
another operator. In close analogy to eqn (3.6) the Liouville equation (3.11)
leads to the formal expression
t
p(t) = T, exp [f ds,C(s)1 p(t o ). (3.12)
to
For the case of a time independent Hamiltonian the Liouville super-operator is
also time independent and we obviously have

P(t) = exp [C(t — to)] P(to). (3.13)

112 QUANTUM MASTER EQUATIONS

3.1.2 Heisenberg and interaction picture

In the Schr6dinger picture the time dependence of the density matrix p(t) is gov-
erned by the Liouville—von Neumann equation (3.11). An equivalent description
of the quantum dynamics is obtained by transferring the time dependence from
the density matrix to the operators in the Hilbert space 7 t. This leads to the
-

Heisenberg picture. We assume that at some fixed initial time to the quantum
states in both pictures coincide, that is p(t o ) = pH (to ). Here and in the following
Heisenberg picture operators are indicated by an index H. Schr6dinger picture
and Heisenberg picture operators are related through the canonical transforma-
tion

A H (t) = Uf(t, t o )A(t)U(t, to ), (3.14)

where we allow the Schriidinger picture operator A(t) to depend explicitly on
time. Obviously, at time to Schralinger and Heisenberg picture operators coin-
cide, AH (t0) = A(to ). In the Heisenberg picture the quantum expectation values
are determined through the fixed density matrix pH(to ). The physical equiva-
lence of the two pictures is made evident by the fact that the expectation value
of an observable A(t) is the same in both pictures,

(A(t)) = tr {A(t)p(t)} = tr OH MPH (to)} -

(3.15)

It is straightforward to derive an equation of motion for a Heisenberg picture

operator A H (t) from the transformation law (3.14). Differentiating both sides of
(3.14) with respect to time we obtain the Heisenberg equation of motion
d DAH(t)
— AH(t) = i[II H (t) ,A H ( 01 ± (9t , (3.16)
dt
where I- I H (t) denotes the Hamiltonian in the Heisenberg picture,

HH(t) = Ut(t, to)H(t)U(t, to ). (3.17)

In eqn (3.16) d/dt denotes the total time derivative, while 0/0t is the partial
derivative with respect to the explicit time dependence of the Schriidinger picture
operator. Explicitly, we have

a AH(t)
ot — ut (t ' to)"ot( t ) u (t, to) - (3.18)

If dAH(t)Idt = 0, then AH is a constant of motion. An important special

case of the Heisenberg equation of motion is obtained by choosing A to be the
Hamiltonian of the system, i.e. A = H. For an isolated system 01110t = O.
and the time-evolution operator has the special form (3.5). Hence, the Hamilto-
nian commutes with the time-evolution operator U(t, to ) and it follows that the
Heisenberg picture Hamiltonian HH(t) is a constant of motion, that is
 CLOSED AND OPEN QUANTUM SYSTEMS 113

d
— HH = 0 . (3.19)
dt
Moreover, if the SchrZidinger operator A has no explicit time dependence and if
the system is isolated the Heisenberg equation of motion reduces to the form

d
— A H (t) = 411, AH (t)]. (3.20)
dt
Let us now look at the equation of motion for the expectation value of an
arbitrary Schriidinger observable A(t) , which in general will depend explicitly on
time. Differentiating eqn (3.15) with respect to time and invoking the Heisenberg
equation of motion (3.16) one finds

0 A;(t) )
dcit AH (t)) = tr { (i[HH (t) , A H (01 ±
dd t (AM) = (—
— P11(t0)} - (3.21)

This equation is known as the Ehrenfest equation. It states that the time vari-
ation of the expectation value of an observable A(t) is equal to the expectation
value of the time derivative of the corresponding Heisenberg observable AH (t) .
The Schn5dinger and the Heisenberg pictures are the limiting cases of a more
general picture, which is called the interaction picture. Let us write the Hamil-
tonian of the system as the sum of two parts

H(t) = Ho + Û1 (t). (3.22)

The precise form of the splitting of the Hamiltonian depends upon the particular
physical situation under study. In general, Ho represents the sum of the energies
of two systems when the interaction between the systems is ignored and we shall
assume that it is independent of time; ii,i (t) is then the Hamiltonian describing
the interaction between the systems. Again, the time evolution operator of the
total system will be denoted by U(t, to ). According to eqn (3.15) the expectation
value of a Schriidinger observable A(t) (which may depend explicitly on time) at
time t is given by

(A(t)) = tr {A(t)U(t,to)P(to)Ut(t, t0)}, (3.23)

where p(to) is the state of the system at time to.

We introduce the unitary time evolution operators

Uct (t, to) E eXP[ — iHO (t — to)] (3.24)

and

U1 (t, to) E Uci; (t, to)U(t, to). (3.25)

Then the expectation value (3.23) may be written as

114 QUANTUM MASTER EQUATIONS

(11(t)) = tr { Uci; (t, to )A(t)U0 (t, to)U/ (t)p(to)U1 (t, to)}

E tr { A i (t) p 1 (t)} , (3.26)

where we have introduced Ai(t) as the interaction picture operator

24/ (t) E Uct) (t, to)A(t)U0(t, to), (3.27)

and pi(t) as the interaction picture density matrix

Mt) E U/ (t, to)P(to)0(t, to). (3.28)

It may be important to note that in contrast to the Heisenberg picture the time
evolution of interaction picture operators is not generated by the full Hamiltonian
H but only by the free part 1/0 . If f-1/(t) = 0 we have Uo (t, to ) = U(t, to ) and
U/(t, to ) = I, such that the interaction picture is identical to the Heisenberg
picture. Conversely, in the other limiting case of a vanishing free Hamiltonian,
Ho = 0, we have f-1/(t) = H (t) such that Uo(t, to) = I and U./(t, to) = U(t, to),
and we regain the Schr6dinger picture.
It is clear that the interaction picture time-evolution operator Ur (t, to) is the
solution of the differential equation

0
i — U1 (t o)
Ot
= HI W U' (t, to ) (3.29)
't
subject to the initial condition U/(to, to) = I. Here, we have denoted the inter-
action Hamiltonian in the interaction picture by

Hj(t) E U(I3, (t, to)ffi MU() (t, to). (3.30)

The corresponding von Neumann equation in the interaction picture thus takes
the form

d
— (3.31)
dt P 1 (t) = — i [II i (t) , P i (0] .

Often, the interaction picture von Neumann equation is written in the equivalent
integral form

t
P 1 (0 = P I (to) — i fds [II i (8) , P 1 (s)1 . (3.32)
to

This form of the von Neumann equation may be used as a starting point of a
perturbative approach to the construction of approximate solutions.
 CLOSED AND OPEN QUANTUM SYSTEMS 115

(S + B, 1-t s 0 71 B, P )

FIG. 3.1. Schematic picture of an open quantum system.

3.1.3 Dynamics of open systems

Having briefly sketched the fundamental equations describing the Hamiltonian
dynamics of a closed quantum system let us now turn to the notion of an open
system. In general terms, an open system is a quantum system S which is cou-
pled to another quantum system B called the environment. It thus represents a
subsystem of the combined total system S +B (see Section 2.2), whereby in most
cases it is assumed that the combined system is closed, following Hamiltonian
dynamics. The state of the subsystem S, however, will change as a consequence
of its internal dynamics and of the interaction with the surroundings. The inter-
action leads to certain system—environment correlations such that the resulting
state changes of S can no longer, in general, be represented in terms of uni-
tary, Hamiltonian dynamics. The dynamics of the subsystem S induced by the
Hamiltonian evolution of the total system is often referred to as reduced system
dynamics, and S is also called the reduced system.
Let us denote by 7-15, the Hilbert space of the system and by 1-63 the Hilbert
space of the environment. The Hilbert space of the total system S + B is then
given by the tensor product space 94 = 7-ts ORB (see Section 2.2.1). The total
Hamiltonian H(t) may be taken to be of the form

H(t) = Hs 0 IB ± IS 0 HB ± fil(t), (3.33)

where Hs is the self-Hamiltonian of the open system S, Hg is the free Hamilto-

nian of the environment B, and Hi(t) is the Hamiltonian describing the interac-
tion between the system and the environment. Of course, one may also include
time-dependent interactions into the model. A schematic picture of the typical
situation under study is shown in Fig. 3.1.
116 QUANTUM MASTER EQUATIONS

If we speak of an open system S we shall use the general term environment for
the system B coupled to it. The term reservoir refers to an environment with an
infinite number of degrees of freedom such that the frequencies of the reservoir
modes form a continuum. As will be seen, it is this property which generally
leads to an irreversible behaviour of the open quantum system. Finally, the term
bath or heat bath will be used for a reservoir which is in a thermal equilibrium
state.
We are going to investigate several physical examples of open quantum sys-
tems in this book. Mainly, the motivation for the study of open systems is that
in many physically important situations a complete mathematical model of the
combined system's dynamics is much too complicated. The environment may
represent, for example, a reservoir or a heat bath consisting of infinitely many
degrees of freedom, in which case an exact treatment requires the solution of an
infinite hierarchy of coupled equations of motion. Even if a solution is known one
is confronted with the task of isolating and determining the interesting physical
quantities through an average over the remaining, irrelevant degrees of freedom.
Moreover, one often encounters the situation that the modes of the environment
are neither known exactly nor controllable. One therefore tries to develop a sim-
pler description in a reduced state space formed by a restricted set of physically
relevant variables which is achieved by employing various analytical methods
and approximation techniques.
We regard an open system S to be singled out by the fact that all observations
of interest refer to this subsystem. According to Section 2.2.1 the observables
referring to S are all of the form A 0 IB, where A is an operator acting on
the Hilbert space 'Hs and IB denotes the identity in the Hilbert space 7-IB. If
the state of the total system is described by some density matrix p, then the
expectation values of all observables acting on the open system's Hilbert space
are determined through the formula

(A) = trs {APs} , (3.34)

where

Ps = trBp (3.35)
is the reduced density matrix of the open quantum system S (see eqn (2.92)).
Here and in the following trs denotes the partial trace over the open system's
Hilbert space, while trB denotes the partial trace over the degrees of freedom of
the environment B. It is clear that the reduced density operator ps will be the
quantity of central interest in the description of open quantum systems.
The reduced density matrix PS (t) at time t is obtained from the density
matrix p(t) of the total system by taking the partial trace over the degrees of
freedom of the environment. Since the total density matrix evolves unitarily we
have

Ps (t) = trB fu(t, to)p(to)u f (t, to)} 1 (3.36)

 QUANTUM MARKOV PROCESSES 117

u nitary evolution
P(0) = P s (0) 0 PB > P(t) = U(t,0)[Ps(0) 0 pB]Uf (t,0)

trg i Itrg
dynamical map
PS(0) ) Ps( t) = V (t )p s (0 )

FIG. 3.2. A commutative diagram showing the action of a dynamical map V(t).

where U(t, to ) is the time-evolution operator of the total system. In an analo-

gous way the equation of motion for the reduced density matrix is obtained by
taking the partial trace over the environment on both sides of the Liouville—von
Neumann equation for the total system,

--1c--ps (t) = —itrB [11(t),p(q. (3.37)

dt
In the following sections we give a survey of the most important equations of
motion for the reduced density matrix used to approximate the above exact
dynamical equations.

3.2 Quantum Markov processes

The most important property of a classical, homogeneous Markov process is
the semigroup property which is usually formulated in terms of the differential
Chapman—Kolmogorov equation (1.68) involving a time-independent generator.
The extension of this property to quantum mechanics leads to the concepts of
a quantum dynamical semigroup and a quantum Markov process (Kraus, 1971;
Davies, 1976; Alicki and Lendi, 1987; Alicki and Fannes, 2001). These concepts
are introduced in this section and the general form of a quantum Markovian
master equation is derived. We also discuss some general features of quantum
Markov processes, such as the adjoint master equation, the quantum regression
theorem, irreversibility and entropy production.

3.2.1 Quantum dynamical semigroups

In general the dynamics of the reduced system defined by the exact equations
(3.36) and (3.37) will be quite involved. However, under the condition of short
environmental correlation times one may neglect memory effects and formulate
the reduced system dynamics in terms of a quantum dynamical semigroup.
We first introduce the concept of a dynamical map. To this end, let us suppose
that we are able to prepare at the initial time t = 0 the state of the total system
S + B as an uncorrelated product state p(0) = ps(0) 0 pB, where ps(0) is the
initial state of the reduced system S and pB represents some reference state of
the environment, a thermal equilibrium state, for example. The transformation
describing the change of the reduced system from the initial time t = 0 to some
time t> 0 may then be written in the form
118 QUANTUM MASTER EQUATIONS

ps (0) 1- ps (t) = V (t)ps (0) E trB { U(t, 0) [ps (0) 0 pBlUt (t, 0)}. (3.38)

If we regard the reference state pB and the final time t to be fixed, this relation
defines a map from the space S(I ( s) of density matrices of the reduced system
into itself,

V (t) : S(I s) —4 S(Hs)-

( (3.39)

This map, describing the state change of the open system over time t, is called
a dynamical map (see Fig. 3.2).
A dynamical map can be characterized completely in terms of operators per-
taining to the open system's Hilbert space 'Hs. To this end, we use the spectral
decomposition of the density matrix pB of the environment,

pB = E Aalvax(pod. (3.40)
c,
Here, the Iva ) form an orthonormal basis in RB and the A, are non-negative
real numbers satisfying EOE A, = 1. Definition (3.38) then immediately yields
the following representation,

v(t)p wao mpswat,(0, (3.41)

c3

where the Wao , being operators in lis, are defined by

W(t) = 00(Va 1U(t , 0 )1'100)- (3.42)

According to eqn (3.41) the dynamical map V(t) is of the form (2.157) of an
operation (D in describing a generalized quantum measurement. Moreover, the
operators Wao (t) satisfy the condition

E wo,t s (owa o (t) = IS, (3.43)

a ,0

from which we deduce that

trs {V(t)ps} = trsps = 1. (3.44)

Thus, we conclude that a dynamical map V (t) represents a convex-linear, com-

pletely positive and trace-preserving quantum operation.
We have introduced above a dynamical map V(t) for a fixed time t > 0. If
we now allow t to vary we get a one-parameter family {V(t)lt > 0} of dynamical
maps, where V(0) is the identity map. Such a family describes the whole future
time evolution of the open system, which, in general, could be very involved.
However, if the characteristic time scales over which the reservoir correlation
 QUANTUM MARKOV PROCESSES 119

functions decay are much smaller than the characteristic time scale of the sys-
tematic system evolution, it is justified to neglect memory effects in the reduced
system dynamics. As in the classical theory one thus expects Markovian-type
behaviour. For the homogeneous case the latter may be formalized with the help
of the semigroup property (compare with eqn (1.70))

V( t ) V (t2 ) = V (t + t2) t i , t2 > 0. (3.45)

The precise physical conditions underlying the Markovian approximation will be

given in the microscopic derivations of Markovian quantum master equations (see
Sections 3.3 and 3.4). A systematic investigation and an expansion around the
Markovian limit is postponed to Part IV. Summarizing, a quantum dynamical
semigroup is a continuous, one-parameter family of dynamical maps satisfying
the semigroup property (3.45).

3.2.2 The Markovian quantum master equation

Given a quantum dynamical semigroup there exists, under certain mathematical
conditions (see below), a linear map r, the generator of the semigroup, which
allows us to represent the semigroup in exponential form,

V(t) = exp(ft). (3.46)

This representation immediately yields a first-order differential equation for the

reduced density matrix of the open system,

d
— P s = p s (t) , (3.47)
dt
which is called the Markovian quantum master equation. The generator r of the
semigroup represents a super-operator. It may be regarded as a generalization of
the Liouville super-operator which has been introduced in eqn (3.11) and which
is given by the commutator with some Hamiltonian.
Let us construct the most general form for the generator f of a quantum
dynamical semigroup. To this end, we consider first the simple case of a finite-
dimensional Hilbert space Rs, dimRs = N. The corresponding Liouville space
is a complex space of dimension N2 and we choose a complete basis (see eqn
(2.97)) of orthonormal operators Fi , i = 1,2, ... , N 2 , in this space such that we
have

(Fi , trs {FP } 8j. (3.48)

For convenience one of the basis operators is chosen to be proportional to the

identity, namely FN 2 = (1/N) 1 / 2 /s, such that the other basis operators are
traceless, that is trsFi = 0 for i = 1, 2, ,N2 — 1. Applying the completeness
relation to each of the operators W(t) defined in eqn (3.42), we have
120 QUANTUM MASTER EQUATIONS

N2
Woo W = (3.49)

With the help of the representation (3.41) we can write the action of the dynam-
ical map V(t) as
N2
V WPS E (3.50)

where

cii (t) (Fi , Wco (t))(Fj , Wao (t))* (3.51)

The coefficient matrix c = (cij ) is easily seen to be Hermitian and positive. In

fact, for any N2 -dimensional complex vector y we have
2
(E vi Fi , Wo (t)) > 0, (3.52)

which proves that e > 0.

The definition (3.46) of the generator f now gives, by virtue of eqn (3.50),
1
rps = hill — { V(E)PS — PS}
e—)-0 E
N 2 _1
{ 1 cN 2 N 2(E) — N 1
= liM Ps + (eiN2 (6)
FiPs + eN2i(E) Fi
e—q) N E fl —
V

(3.53)

Next, we define coefficients aii by

cN 2 N 2(E) — N
aN2 N2 (3.54)
e

a iN2 = lirn ei N2 (E) ,N2 1, (3.55)

E *
c(E)
= e—)-0 E 15 ,N2 1, (3.56)

and introduce the quantities

(3.57)
 QUANTUM MARKOV PROCESSES 121

and
1 1
G = —a N 2 N 2/s + - + F), (3.58)
2N 2
as well as the Hermitian operator
1
H= (Ft - F). (3.59)

We note that the matrix formed by the coefficients a ii, i , j = 1, 2, N 2 - 1,

is Hermitian and positive. With the help of these definitions we can write the
generator as
N 2 -1
rPS [H, psi + {G, + E aii FipsFr (3.60)

Since the semigroup is trace preserving we have for all density matrices ps

N2 -1
0 = trs {ps}
f trs { (2G + E ajjFjFi J ps} (3.61)

from which we deduce that

N 2 -1
1 F
G (3.62)
2

Hence, we get from (3.60) the first standard form of the generator,

N 2 -1
fps [H, psi + E aii (F riPsF - -1 Fi, Ps}) • (3.63)
j=1
3 2 3

Since the coefficient matrix a = (a ii) is positive it may be diagonalized with the
help of an appropriate unitary transformation u,

( -Y1
0'y2
uaut (3.64)
00 •. 0
\0 0 • • • -yN2_ 1

where the eigenvalues -y i are non-negative. Introducing a new set of operators

Ak through

N 2 -1
= A (3.65)
k=1
122 QUANTUM MASTER EQUATIONS

the following diagonal form of the generator is obtained

N2 -1
, CPS PS] +E (AkpsAt k — —21 Atk AkPs — —12 PSA kt Ak) (3.66)

This is the most general form for the generator of a quantum dynamical semi-
group. The first term of the generator represents the unitary part of the dy-
namics generated by the Hamiltonian H. The operators Ak, introduced above as
appropriate linear combinations of the basis operators Fi in Liouville space, are
usually referred to as Lindblad operators and the corresponding density matrix
equation (3.47) is called the Lindblad equation. We note that the non-negative
quantities 'yk have the dimension of an inverse time provided the Ak are taken
to be dimensionless. As will be seen later the 'yk are given in terms of certain
correlation functions of the environment and play the rôle of relaxation rates for
the different decay modes of the open system.
The detailed mathematical proof that (3.63) defines the most general gen-
erator of a quantum dynamical semigroup for the case of a finite-dimensional
Hilbert space has been given by Gorini, Kossakowski, and Sudarshan (1976). At
the same time Lindblad (1976) proved in a theorem that bears his name that
(3.66) provides the most general form for a bounded generator in a separable
Hilbert space if the index k is allowed to run over a countable set. In the physics
literature quantum master equations of the form (3.63) or (3.66) appeared much
earlier, for example in the context of spin relaxation dynamics and laser the-
ory (see, e.g. Bausch, 1966; Haake, 1973; Haken, 1984). We emphasize that the
Lindblad theorem presupposes that the generator f is bounded. In physical ap-
plications this is usually not the case: The self-Hamiltonian Hs of the reduced
system as well as the Lindblad operators are, in general, unbounded. However,
all known examples for generators of quantum dynamical semigroups are either
of Lindblad form, or can be cast into it by slight modifications.
In the above presentation we have left out all mathematical details. Let us
make, however, a few remarks on the mathematical theory (for a review see,
e.g. Gorini, Frigerio, Verri, Kossakowski and Sudarshan, 1978; Spohn, 1980; Al-
icki and Lendi, 1987). First, we mention that the theory of quantum dynami-
cal semigroups is usually formulated in terms of the Heisenberg, instead of the
Schrödinger picture. To this end, one characterizes dynamical maps in the Heisen-
berg picture. The Heisenberg picture dynamical map will be denoted by Vt(t)
and acts on operators A in the Hilbert space 7-ts of the open system such that
we have
tr s {A (V (Op s)} = tr s { (V t (t) A) p s (3.67)

for all ps. Since we know that V(t) maps density matrices to density matrices,
the Heisenberg picture map Vt (t) should also transform positive operators into
positive operators and must preserve the identity operator, that is Vt (t)Is = I.
If the mathematical theory is based on an axiomatic approach to dynamical maps
 QUANTUM MARKOV PROCESSES 123

one has to demand not only positivity but the stronger condition of complete
positivity (see Section 2.4.3). However, from a physical point of view the addi-
tional requirement of complete positivity does not introduce a new condition if
dynamical maps are introduced, as we have done above, through the reduced
dynamics induced by the unitary dynamics in a larger system S + B.
We have also required the one-parameter family of dynamical maps V(t) to
be continuous. It can be shown that it suffices to demand ultraweak continuity,
which means that we must have
lim trs {(Vt (OA – A) ps} = 0 (3.68)

for all Ps and all bounded operators A. It can be shown further that quantum
dynamical semigroups always have the property of being contracting, that is
they fulfil the condition
Ilv t
( )
r A AH1, (3.69)
where A = trs IA is known as the trace norm. It is this property which
allows the application of the theory of contracting semigroups and enables the
introduction of the infinitesimal generator f.
It is sometimes convenient to introduce the dissipator

D(Ps) E 2A
7k (AkpSA tk - -1 kt r„
AL.D‹
„, - -
2 t AL)
1 ,OçA k (3.70)

and to write the quantum master equation (3.47) in the form

d
— (3.71)
dt Ps Ps(t)] + D(Ps(t)).
We remark that, in general, the operator H cannot be identified with the free
Hamiltonian Hs of the reduced system S: The Hamiltonian H may contain
additional terms which are due to the coupling of the system to its environment
(see Section 3.3.1 for an example). We also note that the generator f does not
uniquely fix the form of the Hamiltonian H and the Lindblad operators A. In
fact, the generator is invariant under the following transformations:
1. unitary transformations of the set of Lindblad operators,
^Ai Ri A 'i = Eu , (3.72)

where ui3 is a unitary matrix,

2. inhomogeneous transformations
Ai Ati =--- Ai + ai ,
H H' = H+
1
E (a*.3 - aiAti ) + b, (3.73)

where the ai are complex numbers and b is real.

124 QUANTUM MASTER EQUATIONS

Because of the second invariance property it is always possible to choose traceless

Lindblad operators.
As we already noticed the open system may be subjected to an external time-
dependent field. The description of such an open system may require the help of
time-dependent generators. An obvious time-dependent generalization of (3.47)
may be written in the form
d
Ps(t) = (3.74)

where C(t) is the generator of a quantum dynamical semigroup for each fixed
t > O. Let us introduce the corresponding propagator by means of
t

V(t, to ) = T+_ exp ( , (3.75)

Id)
to
which satisfies
a
i v(t,t o ) = OW (t, to). (3.76)

Instead of the semigroup property (3.45) we now have

V(t, ti)V(ti, to) = V(t, to). (3.77)

We study Lindblad generators with time-dependent external fields in Section 8.4.

3.2.3 The adjoint quantum master equation
In full analogy to closed quantum systems, also for open quantum systems it
is possible to define for each system operator A in the Schrödinger picture a
corresponding operator AH(t) in the Heisenberg picture. This is achieved with
the help of eqn (3.67) which we rewrite as

trs {A V ( t, O) ps } =trs I(Vt(t, 0)A) Ps} =trs {AH(t)Ps}, (3.78)

where we allow the Lindblad generator to be time dependent. In analogy to

eqn (3.75) we have introduced the adjoint propagator
t
vt (t , to = 1-1 _ exp (f dset (s)) . (3.79)
to

The symbol T, denotes the antichronological time-ordering operator, while the

adjoint generator ft is defined by

tr s {A,C(t ps} -- trs { (E t ( t )A) Ps} .

) (3.80)
 QUANTUM MARKOV PROCESSES 125

The adjoint propagator Vt (t, to ) satisfies the differential equation

0
— V (t to ) = Vt(t to )ft(t)
, ,
(3.81)
Ot '
and describes the time evolution of operators in the Heisenberg picture,

AH(t) = Vt(t, 0)A. (3.82)

This yields the following equation of motion for the Heisenberg operator AH (t),

d
— AH(t) = Vt(t,0) {t(t)A} , (3.83)
dt
which is known as the adjoint master equation.
Notice that on the right-hand side of eqn (3.83) the adjoint Lindblad operator
first acts on the operator A and then the result is propagated with the help of
the adjoint propagator. In order to determine the form of the equation of motion
at time t knowledge of the Heisenberg operator AH(t) is not sufficient; one also
needs the adjoint propagator, in general.
An important special case is obtained if the Lindblad generator does not
depend explicitly on time, as is the case in eqn (3.66). In this case the adjoint
Lindblad generator ft commutes with Vt(t, 0) and the adjoint quantum master
equation takes the simpler form

d
— AH(t) (3.84)
dt
, r t AH(t)
1 t
i [H AH (t)] +E -Yk (A tkAH(t)Ak - — AH(t)AtA k –
2 k 2 k
AkAH(t)) .
k
Now, the right-hand side of the adjoint master equation only depends on the
Heisenberg operator AH(t) at time t. An example of the use of the adjoint
master equation will be discussed in Section 3.4.6.

3.2.4 Multi - time correlation functions

The quantum master equation (3.47) and the adjoint master equation (3.83)
describe the time evolution of the density matrix and of the Heisenberg picture
operators. With their help we can determine expectation values and matrix el-
ements of all system observables. However, it is important to be aware of the
fact that for open quantum systems these are not the only observable quanti-
ties. A complete statistical description also requires knowledge of mean values
of operators taken at different times, that is of multi-time correlations functions.
In physical applications these are of great interest since they are often directly
related to measurable quantities. A typical example is provided by a radiation
mode a(t) of the electromagnetic field. In this case one might be interested in
126 QUANTUM MASTER EQUATIONS

the fluctuation spectrum which is defined in terms of the Fourier transform of

the two-time correlation function (at (t)a(0)).
To begin with, we first consider a two-time correlation function of the form
g(t, s) (B(t)C(s)) = trs {B(t)C(s)ps (0 )1 (3.85)
where t > s > 0. Suppressing for simplicity the index H we write B(t) and C(t)
for arbitrary Heisenberg operators, where we take the fixed initial time to = 0.
On using the propagator of the quantum master equation we can write eqn (3.85)
as follows,

g(t, s) = tr s IBV(t,$)C17(s, 0)ps(0)} , (3.86)

where B and C are Schrödinger picture operators. We use here and in the fol-
lowing the convention that the super-operators V(t, s) always act on anything
standing to the right of them. Equation (3.86) thus means that, in order to calcu-
late the two-time correlation function, one propagates the density matrix ps(0)
from time 0 to time s, applies from the left the operator C, propagates the new
'density matrix' Cps(s) from time s to time t, and, finally, applies B from the
left and takes the trace over
This result is easily generalized to the most general form of a measurable
multi-time correlation function which is given by (Gardiner and Zoller, 2000;
Gardiner and Collett, 1985)

g(t 1 ,t27 • . • tn; .9 1 7 .9 2, • • . S m ) (3.87)

(B i (s 1 )B 2 (s 2 )... B 7 (s 7 )C(t) . • • C2(t2)C1(t1))
trs {C(t) ... C2 (t2)Ci (ti)Ps(0)Bi (Si )B2 (82) .

Here, the Bi (s), j = 1, . . . , m, and the Ci (ti ), i = 1, ... n, are arbitrary Heisen-
berg operators and the si and ti are ordered in time such that

tn > tn_i > > ti > 0, sm > sm _1 > > S i > 0. (3.88)
In order to bring the multi-time correlation function into a form analogous to
(3.86) we perform a time ordering of the set {t 1 , t2 , , tn , si, s2, , s m l. Note
that the number q of elements of this set may be smaller than n+m since we allow
the possibility that t i = si for certain indices i, j. We denote the time-ordered
elements by r1 1 = 1, 2, ... , q. Thus we have
,

rq > rq _ i > > ri > O. (3.89)

For each / we define a super-operator T1 through

Tip = Cip if ri = ti si for some i and all j,

Tip = pBi if ri = si ti for some j and all i, (3.90)
Tip = CipBi if r = ti =si for some i and j.
 QUANTUM MARKOV PROCESSES 127

Then we have

,tn; si,s2, • • srn) (3.91)

trs {..rg V(r q ,rq _ 1 ),Fg _1V(r q _1,rq ___2) • • • TiV(ri, 0)ps(0)} •
This equation provides a general expression for all time-ordered multi-time corre-
lation functions in terms of the propagator of the underlying quantum Markovian
master equation.
In practical applications it is often useful to formulate the dynamics of corre-
lation functions in terms of a system of differential equations. Suppose we have
a set of system operators {Bi } such that the Markovian master equation (3.47)
gives rise to a closed linear system of first-order differential equations for their
averages,
d
Gi j(Bi(t)) (3.92)
It (13i(t))

with some coefficient matrix G. The two-point correlation function then satisfies
the same system of differential equations,
d
(B i (t + Bt(t)) = E G ii (Bi (t T )B , (0). (3.93)
3

This statement is known as the quantum regression theorem (Lax, 1963; Gardiner
and Zoller, 2000). To prove it we first write

d
= trs {(Lt B) ps (01 = trs Gii Bi ps (0 . (3.94)

Since this is supposed to hold, in particular, for any initial density ps(0) we
conclude that
= E (3.95)
3

Hence we obtain
d
— (B i (t + r)131(t)) = trs {(ft Bi ) V(t + 77 t)Bips(t)}
dr
=E
Gii (Bi (t 7-)B, (0), (3.96)

which was to be shown. The significance of the quantum regression theorem

derives from the fact that it allows us to determine explicit expressions for the
correlation functions once a solution for the mean values is known. An example
will be discussed in Section 3.4.5.
128 QUANTUM MASTER EQUATIONS

3.2.5 Irreversibility and entropy production

The quantum Markovian master equation (3.47) describes the irreversible evolu-
tion of an open quantum system. In order to discuss in some detail the irreversible
character of quantum dynamical semigroups we shall derive in this section a gen-
eral expression for the entropy production rate a in an open quantum system
which admits a stationary state and show that it is always non-negative.
To begin with, we consider an arbitrary dynamical map V(t) and investi-
gate the change of the relative entropy S(pl1P0) of two states p, po of S which
is induced by V(t) (Lindblad, 1975). Representing V(t) as in eqn (3.38) and
employing eqns (2.116) and (2.119) and the inequality (2.118) we find7

S(V(t)pll V (t)p o ) = S (tr{U (t, 0)p 0 pBUf (t,0)}11tr{U (t , 0) p o ® pBUf (t 7 0)1)

< S (U(t,O)p 0 pBUt (t,0)11U(t, o)po ® pBut(t, 0))
= s(p 0 pB 1 Po 0 PB)
= s(P1 1P0). (3.97)

Let us suppose that po is a stationary state of S, that is V (t)p o = po . We can

then conclude that a dynamical map V (t) decreases the relative entropy with
respect to a stationary state,

s(v( t )pl lpo) s(pl IN). (3.98)

If now V(t) = exp ft is a dynamical semigroup it follows from inequality

(3.98) that the negative time derivative of the relative entropy is positive, that
is we have the inequality

d
a(p(t)) -- S(p(t)11Po) >0, (3.99)
= --
dt

where p(t) = V (t)p(0). Invoking arguments from non-equilibrium thermody-

namics we may interpret the functional a(p) as the entropy production rate (see
below). The definition of the relative entropy functional, eqn (2.113), leads to
the explicit expression

a(P) = — kBtr {f(P) in p} + kBtr ff(p) lnpol > O. (3.100)

Note that we have included here the Boltzmann constant kB into the definition
of the entropy.
Equation (3.100) shows that the entropy production is non-negative and van-
ishes in a stationary state. With the help of Lieb's theorem (see eqn (2.120)) we
can prove even more, namely that the map p 1-4 a(p) is a convex functional on
the state space of the open system (Spohn, 1978). To see this we first observe

7 To simplify the notation we omit the index S on quantities of the reduced system.
 QUANTUM MARKOV PROCESSES 129

that the second term in expression (3.100) is a linear functional. Thus, it suffices
to show that the map

pH-* —kBtr fr(p) ln pl (3.101)

is a convex functional. Employing the general form (3.66) of the generator ,C this
map takes the form

p F4 —k B E Ntr { (A k pAtk — Atk A k p) ln p} . (3.102)

k

The convexity of the entropy production functional a(p) now follows immediately
from the convexity of the map (2.126).
Let us motivate the definition (3.100) for the entropy production rate by using
arguments from non-equilibrium thermodynamics. To this end, we suppose that
the canonical equilibrium distribution (Gibbs state)

Pth = 1 exp (—NI ) (3.103)

Z
is a stationary solution of the master equation. This means that f(pth) --
D(pth) ---- 0, where D denotes the dissipator of the quantum master equation
(see eqns (3.70) and (3.71)). The normalization factor Z = tr exp(— 31-1) is the
partition function and

1
/3 = (3.104)
kBT

is the inverse temperature. The precise physical conditions underlying this as-
sumption will be discussed later.
In non-equilibrium thermodynamics the entropy obeys a balance equation
which can be written in the form
dS
o- ,--- + J. (3.105)
Trt
Here, S is the von Neumann entropy (see Section 2.3.1) of the open system.
The quantity J denotes the entropy flux, that is the amount of entropy which is
exchanged per unit of time between the open system and its environment, where
we use the convention that for J> 0 entropy flows from the open system into
the environment. Consequently, the quantity a is the entropy production rate,
that is the amount of entropy produced per unit of time as a result of irreversible
processes.
The time derivative of the von Neumann entropy is easily evaluated to be

dS
kB tr{,C(p) ln p } . (3.106)
dt
130 QUANTUM MASTER EQUATIONS

On the other hand, the entropy flux J is due to those changes of the internal
energy E = tr{Hp} which result from dissipative effects. Thus we may define
the entropy flux by means of

1d 1 1
J -- — E E -- tr{HD(p)} = -- tr{Hr(p)}. (3.107)
T dt diss T T
Using the explicit expression (3.103) for the thermal distribution we find

1
H = kB ln 9th + kB In Z, (3.108)
T
so that the entropy flow can be written as

J = kB tr {,C(p)ln pth} , (3.109)

where we have made use of the fact that the generator is trace-preserving, i.e.
tr{ f(p)} = 0. Adding eqns (3.106) and (3.109) we see that the thermodynamic
entropy production rate a defined by the balance equation (3.105) coincides with
expression (3.100) for the negative rate of change of the relative entropy with
respect to the thermal equilibrium state. In this context the inequality a(p) > 0
expresses the second law of thermodynamics. Note also that cr(p t h) = 0, that
is the entropy production vanishes in the thermal equilibrium state. Thus we
conclude that the entropy production rate a(p) is a convex functional on the
space of density matrices which vanishes in the thermal equilibrium state.

3.3 Microscopic derivations

From a fundamental viewpoint it is desirable to derive the generator of a quan-
tum dynamical semigroup from the underlying Hamiltonian dynamics of the
total system. The aim of this section is to show under which assumptions such
derivations can be given on the grounds of various approximation schemes.

3.3.1 Weak coupling Limit

-

We begin by considering a quantum mechanical system S weakly coupled to a

reservoir B. The Hamiltonian of the total system is assumed to be of the form

H = Hs+ HB + ( 3.110)

where Hs and HB denote respectively the free Hamiltonian of the system and
of the reservoir and H1 describes the interaction between the system and the
reservoir. The derivation of a quantum Markovian master equation is most easily
performed in the interaction picture. Our starting point is thus the interaction
picture von Neumann equation (see Section 3.1.2)

d
— i[Ili(t),P(t)1
dt P(t) =
—
 MICROSCOPIC DERIVATIONS 131

for the total density matrix p(t) and its integral form

t
p(t) = p(0) – if ds[111 (s), p(s)]. (3.112)
o

Note that we omit here for ease of notation the index / which served to indicate
the interaction picture in Section 3.1.2. Inserting the integral form into (3.111)
and taking the trace over the reservoir we find

d
— p s (t) = – f dstrB [H 1 (t),[11 1 (8), p(s)]]. (3.113)
dt
I
o

Here, we have assumed that

trB [HI (t), p(0)] (3.114)

Equation (3.113) still contains the density matrix of the total system p(t) on
its right-hand side. In order to eliminate p(t) from the equation of motion we
perform a first approximation, known as the Born approximation. This approxi-
mation assumes that the coupling between the system and the reservoir is weak,
such that the influence of the system on the reservoir is small (weak-coupling
approximation). Thus, the density matrix of the reservoir pB is only negligibly
affected by the interaction and the state of the total system at time t may be
approximately characterized by a tensor product

p(t)Pz.,-' p s (0 0 pB. (3.115)

We emphasize that this does not imply that there are no excitations in the
reservoir caused by the reduced system. The Markovian approximation to be
derived below provides a description on a coarse-grained time scale and the
assumption is that environmental excitations decay over times which are not
resolved. Inserting the tensor product into the exact equation of motion (3.113)
we obtain a closed integro-differential equation for the reduced density matrix
Ps (0
,

d
— ) – – f dstrB [II1 (t) , [II1 (s), ps(s) 0 PB]l• (3.116)
dt P (
.

In order to simplify the above equation further we perform the Markov ap-
proximation, in which the integrand ps(s) is first replaced by ps(t). In this way
we obtain an equation of motion for the reduced system's density matrix in which
132 QUANTUM MASTER EQUATIONS

the time development of the state of the system at time t only depends on the
present state Ps (0,

t
d
Ps(t) --r- - f dstrB [II1 (t),[111 (s), Ps (t) 0 PBii. (3.117)
o

This equation is called the Redfield equation (Redfield, 1957; Blum, 1981).
The Redfield equation is local in time, but it is not yet a Markovian master
equation since the time evolution of the reduced density matrix still depends
upon an explicit choice for the initial preparation at time t = O. This implies
that the dynamics of the reduced system is not yet described by a dynamical
semigroup. In order to achieve this we substitute s by t s in the integral
-

in eqn (3.117) and let the upper limit of the integral go to infinity. This is
permissible provided the integrand disappears sufficiently fast for s > TB The .

Markov approximation is therefore justified if the time scale TR over which the
state of the system varies appreciably is large compared to the time scale TB
over which the reservoir correlation functions decay. Thus, we finally obtain the
Markovian quantum master equation

jc ps (t) = - I dstrB [H 1 (t),[11 1 (t - s), PO) 0 Pa• (3.118)

dt
o

It is important to realize that in a description of the reduced system dynamics

on the basis of a Markovian quantum master equation the dynamical behaviour
over times of the order of magnitude of the correlation time TB is not resolved.
As mentioned before, the evolution is described in this sense on a coarse-grained
time axis.
The approximations performed above are usually termed the Born Markov -

approximation. In general they do not guarantee, however, that the resulting

equation (3.118) defines the generator of a dynamical semigroup (Davies, 1974:
Diimcke and Spohn, 1979). One therefore performs a further secular approxima-
tion which involves an averaging over the rapidly oscillating terms in the master
equation and is known as the rotating wave approximation. To explain the pro-
cedure let us write the Schrödinger picture interaction Hamiltonian HI in the
form

HI -= E Aa 0 Ba7 (3.119)
a

where Atc, = A, and Bic; = B a . This is the most general form of the interaction.
The secular approximation is easily carried out if one decomposes the interaction
Hamiltonian H1 into eigenoperators of the system Hamiltonian Hs. Supposing
the spectrum of Hs to be discrete this may be achieved as follows. Let us denote
 MICROSCOPIC DERIVATIONS 133

the eigenvalues of Hs by E and the projection onto the eigenspace belonging to

the eigenvalue E by 11(e). Then we can define the operators

A0 (w) E H(e)A,H(e). (3.120)

Er

The sum in this expression is extended over all energy eigenvalues Ei and E of Hs
with a fixed energy difference of w. An immediate consequence of this definition
is that the following relations are satisfied,

[Hs, A c,(w)] = (3.121)

[Hs, A (w )] = -FwAta (w). (3.122)

The operators A 0 (w) and Atc,(w) are therefore said to be eigenoperators of Hs

belonging to the frequencies +w, respectively. It follows from relations (3.121)
and (3.122) that the corresponding interaction picture operators take the form

e i Hs t A a ( w ) e Hs t
(3.123)
e ifIst iqt p) e —ilist =ei-iwtAlt (w) . (3.124)

Finally, we note that

S, A at (W) 43(W)] = 0,
and

Atc,(w) = A0 (—w).

Summing (3.120) over all energy differences and employing the completeness
relation we get

E A0 (w) =>A(w) = A,. (3.127)

This enables us to cast the interaction Hamiltonian into the following form

= E A 0 (w)®B 0 = (w) Btc,. (3.128)

ct,u) ce,w

This is the desired decomposition of the interaction into eigenoperators of the

system Hamiltonian. Note that the frequency spectrum {w} is, in general, degen-
erate: For a fixed w the index a labels the different operators A 0 (w) belonging
to the same frequency. A specific example will be encountered in Section 3.4,
where a labels the Cartesian components of the dipole operator.
134 QUANTUM MASTER EQUATIONS

The reason for introducing the eigenoperator decomposition (3.128) is that

the interaction picture interaction Hamiltonian can now be written in the par-
ticularly simple form

H1(t) . E e - iw t A a (co) 0 B a (t) = E e+iw t itta (w) ® B(t), (3.129)

where

Ba(t) = e iHB t Bae -iHB t (3.130)

are interaction picture operators of the environment. We also note that condition
(3.114) becomes
(B a (t)) E tr {B a (t)pB} = 0, (3.131)

which states that the reservoir averages of the B c,(t) vanish.

Inserting now the form (3.129) into the master equation (3.118) we obtain
after some algebra
DO
d
— Ps(t)= f d3trB fili(t- 8)ps(t)PB-111(t) - 111(tgli(t - s)Ps(t)PB} + h.c.
dt
o
= E Eeiw — w)trao(w) (Aompswitta(w') — Atc,(wi)itocuops(t))
w,w , ct,s
+h.c. (3.132)
Here h.c. means the Hermitian conjugated expression and we have introduced
the one-sided Fourier transforms
00

110P) E f dse'"(B1(t)Bo(t - s)) (3.133)

o
of the reservoir correlation functions

(Bta (t)B(t - s)) E trB {BI(t)B(t — s)PB 1 . (3.134)

Let us suppose that pB is a stationary state of the reservoir, that is [HB, pB].-- 0.
The reservoir correlation functions are then homogeneous in time which yields

(B(t)B0(t - s)) = (B'cr, (s)Bs (0)), (3.135)

showing that the quantities Fo(w) do not depend on time. We remark that
there are interesting cases in which the reservoir correlation functions do depend
on the time argument t. This happens, for example, if the reservoir represents a
squeezed vacuum state (see Section 3.4.3).
 MICROSCOPIC DERIVATIONS 135

As mentioned before, the basic condition underlying the Markov approxima-

tion is that the reservoir correlation functions (3.135) decay sufficiently fast over
a time TB which is small compared to the relaxation time TR. Typical exam-
ples for the behaviour of these correlation functions will be discussed in Sections
3.6.2.1 and 12.1.1.3. It is important to note that a decay of the correlations can
only be strictly valid for an environment which is infinitely large and involves a
continuum of frequencies. In the typical situation the reservoir is provided by a
collection of harmonic oscillator modes bn with frequencies co n and the B a are
given by linear combinations of the modes bn . If the frequency spectrum {co„}
of the reservoir modes is discrete, it is easy to see that, in general, correlation
functions of the type (3.135) are quasi-periodic functions of s. A rapid decay of
the reservoir correlations therefore requires a continuum of frequencies: For an
infinitely small frequency spacing Poincaré recurrence times become infinite and
irreversible dynamics can emerge.
We denote by Ts the typical time scale of the intrinsic evolution of the system
S. This time scale Ts is defined by a typical value for lw' — w , w' co, that
is by a typical value for the inverse of the frequency differences involved. If Ts
is large compared to the relaxation time TR of the open system the non-secular
terms in (3.132), i.e. the terms for which co' co, may be neglected, since they
oscillate very rapidly during the time TR over which ps varies appreciably. This
condition is typically satisfied for quantum optical systems where it is known as
the rotating wave approximation. Thus we have

cips(t )
dt
. E E ro w (43 (w)p s (t)4(w) - AtΠmAo mps (0) + h.c.
w a 03
(3.136)

It is convenient to decompose the Fourier transforms of the reservoir correlation

functions as follows
1
Pas(w) ---, — -y„s(w) + iS,o (w), (3.137)
2
where for fixed w the coefficients
1
So (w) =---- Ti (Fao (co
. ) —
(w ))
(3.138)

form a Hermitian matrix and the matrix defined by

+00

70(w) = Fa (w) + F a (w) = f dse" (Bta (s)Bs(0)) (3.139)

--co

is positive (see below). With these definitions we finally arrive at the interaction
picture master equation
136 QUANTUM MASTER EQUATIONS

ci Ps(t)
dt Ps(t)] +D(Ps(t)).

The Hermitian operator

Ihs = EESco (w) Ato,(‘.4_,) Ao (w)

c , /3

provides a Hamiltonian contribution to the dynamics. This term is often called

the Lamb shift Hamiltonian since it leads to a Lamb-type renormalization of the
unperturbed energy levels induced by the system-reservoir coupling. Note that
the Lamb shift Hamiltonian commutes with the unperturbed system Hamilto-
nian,

[Hs, Ihs] =0, (3.142)

by virtue of eqn (3.125). Finally, the dissipator of the master equation takes the
form

E'Ycto((-0 )(AsP)PsA tc,(w) --1-2 {Ata(w)A0(w),Ps1). (3.143)

cto3

We note that the master equation (3.140) is of the first standard form (3.63).
It can be brought into Lindblad form (3.66) by diagonalization of the matrices
ey(w) defined in eqn (3.139). In order to prove that these matrices are pos-
itive one uses Bochner's theorem according to which the Fourier transform of
a function f(s) is positive provided f(s) has the property of being of positive
type. The latter property means that for arbitrary t 1 t-2, • • • ,t, and all n the
,

(n x n) matrix aki = f (tk - t1) must be positive. Since all homogeneous cor-
relation functions f (s) = (Bt (s)B (0)) are of positive type the positivity of the
matrices 70(w) follows immediately. Finally, we remark that the Schrödinger
picture master equation is obtained from (3.140) simply by adding the free sys-
tem Hamiltonian Hs to His, as is easily verified with the help of the properties
(3.121), (3.122) and (3.125) of the eigenoperators.
Let us summarize the different approximations used in the above deriva-
tion. The first approximation is a consequence of the weak-coupling assumption
which allows us to expand the exact equation of motion for the density matrix
to second order. Together with the condition p(t) ps(t) 0 pB this leads to the
Born approximation to the master equation. The second approximation is the
Markov approximation in which the quantum master equation is made local in
time by replacing the density matrix ps(s) at the retarded time s with that at
the present time ps(t). Furthermore, the integration limit is pushed to infinity
to get the Born-Markov approximation of the master equation. The relevant
physical condition for the Born-Markov approximation is that the bath corre-
lation time TB is small compared to the relaxation time of the system, that is
TB < TR. Finally, in the rotating wave approximation rapidly oscillating terms
 MICROSCOPIC DERIVATIONS 137

proportional to exp[i(Lo' — co)t] for co' co are neglected, ensuring that the quan-
tum master equation is in Lindblad form. The corresponding condition is that
the inverse frequency differences involved in the problem are small compared to
the relaxation time of the system, that is Ts — — w < TR.
3.3.2 Relaxation to equilibrium
In the previous section we have assumed that the environment is in a stationary
state pB which is invariant with respect to the dynamics of the reservoir. Now
we want to consider a situation in which the environment is a heat bath at
the inverse temperature O. In the absence of external time-dependent fields one
expects the Gibbs state
exp(— /3Hs)
Pth (3.144)
trs exp(-01/s)
to be a stationary solution of the quantum master equation (3.140). It can be
shown then that for any initial state the system returns to equilibrium,

Ps(t) pth, for t +cc, (3.145)

provided the quantum dynamical semigroup has the property of being ergodic.
This means that the relations
[X, Atc,(w)] = [X, A,(co)] = 0 for all a, CO (3.146)
imply that X is proportional to the identity.
In order to show that (3.144) is indeed a stationary solution of the master
equation (3.140) we make use of the KMS condition according to which the bath
correlation functions are related through
(B(t)B(0)) = (B,3(0)Bta (t + 0)). (3.147)
The KMS condition can easily be verified if the reservoir is a heat bath with
canonical equilibrium distribution
exp( — OHB)
PB (3.148)
trB exp(—OHB) -
It can also be shown to hold for thermal equilibrium systems in the thermody-
namic limit. Equation (3.147) leads to the following relations between the Fourier
transforms (3.139) of the bath correlation functions,

70 ( — A)) = exP( - 0w)'yoc, (w). (3.149)

We further have by virtue of eqns (3.121) and (3.122),

PthA cE (L0-7) 7-- e 43w AcEMPth, (3.150)

PthA ta (w) = e —d3w Ata (w)Pth. (3.151)
The proof of the stationarity of p th is now easily carried out with the help of the
relations (3.142), (3.149), (3.150) and (3.151).
138 QUANTUM MASTER EQUATIONS

We mention a further important property of the quantum master equation

(3.140). Namely, if the spectrum of the system Hamiltonian Hs = EnEn In)(n
is non-degenerate it gives rise to a closed equation of motion for the populations
P(n,t) = (rips(t)In) (3.152)
of the eigenstates In). Thus, the equation for the diagonals of the density matrix
in the eigenbasis of Hs decouples from the off-diagonal elements. As is easily
checked using the quantum master equation the populations are governed by the
equation
d
— P(n,t) = E [w(on)P(7n, t) - w(illn)P(n, t)] . (3.153)
dt In

This equation is of the form of the classical discrete master equation (1.89) with
time-independent transition rates given by

W(n n) =E'Yas(Eln — En)(7n1Actln)( 71 1Aolm). (3.154)

oe,3

Equation (3.153) is also known as the Pauli master equation. The rates (3.154)
are real and non-negative as a consequence of the positivity of the matrices
'To (w). They are just those obtained with the help of Fermi's golden rule.
The relations (3.149) give
W(mln)exP(--Oen) = W(nlm) exp(-- /km ) (3.155)
which is known as the condition of detailed balance and which leads to the con-
clusion that the equilibrium populations Ps (n) follow the Boltzmann distribution
P8 (n) = const x exp(-- /3E72 ) (3.156)
over the energy eigenvalues en.

3.3.3 Singular coupling limit

-

In the weak-coupling limit the perturbation caused by the interaction between

the system and the environment is assumed to be small. As a result the degrees
of freedom of the environment are the fast variable and can be effectively elimi-
nated. With an appropriate scaling of the time parameter it is possible to derive
under certain conditions a linear quantum master equation also for the case of
strong coupling. In this so-called singular-coupling limit one considers a total
Hamiltonian of the form
H = Hs +e-2 HB +6-1 HI , (3.157)
where the interaction Hamiltonian is again written as

E Act Ba (3.158)
a

with Ata = Ac, and B = B. The aim is to derive an equation of motion for
the reduced density matrix in the limit c 0. To motivate the form of the
 MICROSCOPIC DERIVATIONS 139

Hamiltonian (3.157) we first observe that the decay time of reservoir correlation
functions of the form (3.134) is decreased by a factor of 6-2 through the scaling
HB E -2 HB. The scaling III 6 -1 HI of the interaction Hamiltonian ensures
that the Fourier transforms of the reservoir correlation functions remain finite
in the limit e -4 0.
The derivation of the quantum master equation from this microscopic model
is similar to the weak-coupling case and will not be presented here in detail. The
only essential difference is that it is not necessary to perform the rotating wave
approximation. The result of the derivation is the following Schrödinger picture
master equation,
d
—p s (t) -_-, -i[Hs + HLS 1 PS(0] ± E-yas
1
(A,ps(to, _ _ {A,A3,ps(t)}) ,
dt
ct43
2
(3.159)
where the Lamb shift Hamiltonian reads

HLS ---= Esa0 A,A0 , (3.160)

cr)3

with Sao defined analogously to (3.138), and

+Do
tyco = f ds(B a (s)Bfi(0)). (3.161)

The matrix -yas is again Hermitian and positive. Hence, the generator of the
quantum master equation (3.159) is of the first standard form (3.63) and may be
written in Lindblad form by diagonalization of the matrix -yas . Master equations
of the general form (3.159) will be encountered, for example, in our study of
the quantum Zeno effect in Section 3.5.2 and in the recoilless limit of quantum
Brownian motion (see Section 3.6.4.5).

3.3.4 Low - density limit

We consider a gas of particles with low density n. The particles of the gas inter-
act through collisions involving the excitation of internal degrees of freedom of
the gas particles. The collisions of the gas particles are considered as statistically
independent processes. The aim is to write down a master equation which effec-
tively describes the dynamics of the internal degrees of freedom. To this end, the
internal variables are taken to be the reduced system S, while the translational
degrees of freedom of the gas particles provide the environment B.
The internal degrees of freedom of the particles may be described by a free
Hamiltonian of the form

Hs = E Ed/0 (k 1' (3.162)

k
140 QUANTUM MASTER EQUATIONS

The translational degrees of freedom of a gas particle are described by the Hamil-
tonian

HB = fd 3 pE(15)1A(151, (3.163)

where IA is a momentum eigenstate and E ( 5) = 2 /2m is the free particle

energy. The one-particle states of the environment are given by the density matrix

pB = —vi f d3PG(15)11 5)051, (3.164)

where GO represents the momentum distribution of the gas particles. The den-
sity matrix pB is normalized in some volume V according to

trpB f
0 d3 13 5-1PB113 ) = f (2d3,703pL-7(P)
„= , ( 3.165)

The interaction of the internal degrees of freedom with the translational de-
grees of freedom is given by scattering reactions in which the incoming momen-
tum changes from 17 to while the internal level changes from 1) to k). The
scattering amplitude for this process can be written in the form

(k, OS1 1 ,13) = (4. 15)(5k1 — 27ri5 (ck + E (q ') — — EV) T (k, 41- 1 ,13), (3.166)

with the S-matrix and the T-matrix known from scattering theory. In the low-
density limit the reduced dynamics of the quantum systems formed by the inter-
nal degrees of freedom can be shown to yield the Schrödinger picture quantum
master equation
d
Ps(t) = —i[Hs HLS PS (t)] +D(Ps(t) (3.167)

The Hamiltonian Ths yields a renormalization of the system Hamiltonian,

HLS = —n2 E f d3
Ek =6 /
P G05) (k ,1511,15) + T* ( 1 ,131k,13)} (3.168)

and the dissipator is given by

D(p s) = 2rn Ef d3 p f d 3 qG(13)(5 (EK) E(13) + w) (3.169)

X { TW(4:APS T1(T7 — PSTC:1r (q1/3) 7142 (4.7)3)

where 71,(4.,17) is defined by

(T, = E (3.170)
Ele—El=W
 THE QUANTUM OPTICAL MASTER EQUATION 141

Thus the Lindblad operator 77, (4', /I) is given by a coherent superposition of the
scattering amplitudes for all processes which contribute to a given excitation
energy w of the internal degrees of freedom,

C4) = Ek — q = E(p) — E(0. (3.171)

This is the low-density limit of the quantum master equation which can be
shown to hold in the thermodynamic limit. A rigorous mathematical treatment
may be found in D Eimcke (1985). Equation (3.167) is obviously in Lindblad form,
where we have, however, a continuous family of Lindblad operators. To derive
the master equation an averaging procedure over times which are large compared
to the mean collision time rc is performed, which is similar to the rotating wave
approximation. Let d be the linear dimension of the interaction volume, y the
mean particle velocity, and 1 the mean free path length. Then the low-density
limit master equation is valid as long as

—1,1 v—ln-1/3
re ess V it < V— t 7 (3.172)

which means that the mean collision time Tc must be small compared to the
time between collisions. Hence, we get the low-density condition n <

3.4 The quantum optical master equation

The interaction of matter with electromagnetic radiation in the quantum optical
limit provides a typical field for the application of quantum dynamical semi-
groups and quantum Markovian master equations (Cohen-Tannoudji, Dupont-
Roc and Grynberg, 1998; Gardiner and Zoller, 2000; Walls and Milburn, 1994).
The reason for this fact is that in a typical quantum optical situation the phys-
ical conditions underlying the Markovian approximation are very well satisfied.
A number of examples will be discussed in this section.

3.4.1 Matter in quantized radiation fields

We consider a bound quantum system, e.g. an atom or a molecule, which interacts
with a quantized radiation field. The radiation field represents a reservoir with
an infinite number of degrees of freedom and the bound system is the reduced
system we are interested in. The uncoupled atom or molecule is described by some
Hamiltonian Hs, while the free quantized radiation field will be represented by
the Hamiltonian (subtracting an infinite c-number for the vacuum energy)

HB = EE riwkbio-z)b),(/-,). (3.173)
X=1,2

In this section we reintroduce all physical constants such as Planck's constant

h and the speed of light c. For simplicity we decompose the radiation field into
Fourier modes in a box of volume V, imposing periodic boundary conditions.
142 QUANTUM MASTER EQUATIONS

These modes are labelled by the wave vector k and two corresponding, transverse
unit polarization vectors 6,(1;), such that

( =0, (3.174)
r,(k) • 6, (k) = S, (3.175)
kz k
E e(k)e(k) = 3 i, = 1,2,3. (3.176)
A=1,2 l iC'1 2 7

The dispersion relation is Wk = ck = c. The field operators b),(1) and OA (1-i)

describe the destruction and creation of photons with wave vector 1 and polar-
ization 6,(k). They obey the commutation relations

(P)] = [b, °Ai (P )] = 0, (3.177)

[b),(), b, (k')] 15- P 6)0,1 • (3.178)

Finally, we assume the interaction Hamiltonian to be given in the dipole

approximation by

H1= -it) • f, (3.179)

where B is the dipole operator of the system under consideration and Ê is the
electric field operator in the Schrödinger picture,

f=iE hwk (1-c.)

27v (b),(k) — btx (k)) . (3.180)

With these definitions the total Hamiltonian governing the coupled system of
the matter degrees of freedom and the radiation degrees of freedom is given by

H = Ils +11B + (3.181)

3.4.1.1 Performing the Born-Markov approximation We proceed as in Section

3.3.1 and decompose the dipole operator D into eigenoperators of H. The latter
take the form (see eqn (3.120))

A(w) E ll(s),6H(61) (3.182)

-Eh)

Note that the index a used in Section 3.3.1 labels here the Cartesian components
Di , i = 1, 2, 3, of the dipole operator. In accordance with eqns (3.121), (3.122)
and (3.126) we now have

[H f(w)] = — hw Â(w) , [H s Ât ( o)] = Fhwift(w), - (3.183)

 THE QUANTUM OPTICAL MASTER EQUATION 143

and
Al (
o ) 4 (— w ). (3.184)

The decomposition of the interaction picture dipole operator bb (t) into eigenop-
erators therefore reads
15 (t) E e _iwt .z( ,) _ E e -Fiwt i v (,), (3.185)

and the interaction Hamiltonian can now be written in the interaction picture
in a form analogous to eqn (3.129),
(t) = e - iw t1(w ) (3.186)

where f(t) denotes the electric field operator in the interaction picture.
Assuming as in eqn (3.131) that

(É(t)) E tr {_g(t)p B } =0, (3.187)

we can immediately write down the equation of motion analogous to eqn (3.132),
d
dt s
EE (w) (A (w) p s Ati (co') — (w i ) Ai (co) p s} + h.c.
w,w'
(3.188)
The correlation functions of the electric field operator are defined through
(Ei (t)Ei (t s)) E tr B {Ei (t)Ei (t — s)pB}
and their one-sided Fourier transforms are given by

rii ( r.)) E f dse i w (Ei(t)E (t — s)) .

o
The matrix Fii (w) will be referred to as the spectral correlation tensor. It de-
pends, in general, on t. In fact, without further assumption on the state pB of
the reservoir we find
2 7rruok,1
2 7rhwie exi (k) eiv (p)
(w) = EE (3.191)
i;,W, )01
00

x f ds [(b ),(k) bt),, (i)) exP [+i(wk , wk)t i(wk , — w).9]

+(btx(k)bv(P))exP [ — z(wk— cok)t + i(cok , + co)s]

— (1)),(k)bv(P))exP + cok)t + i(wk , + w) .5]

(btx (k)b tx , (1?)) exP [ F - — (co — co) .51] .

144 QUANTUM MASTER EQUATIONS

3.4.1.2 Thermal reservoir As our first example let us now take the reservoir
of radiation modes to be in an equilibrium state at temperature T,
1
pB = exp [ ,31-1B] = ri(i_exp[_,@riwk])exP [— firkokb t),(0b,„(k)]
— .

k*,x
(3.192)
As above, expectation values with respect to this equilibrium state are denoted
by angular brackets. Note that (f(t)) = 0 since the field operator is linear
in the creation and annihilation operators. Since the thermal equilibrium state
is stationary with respect to the reservoir dynamics, the reservoir correlation
functions are homogeneous in time and we have

(Ei (t)Ei (t — s)) = (Ei (s)Ei (0)). (3.193)

It follows that the spectral correlation tensor fii(w) does not depend on t. Con-
sequently, if we perform the rotating wave approximation in eqn (3.188) only the
diagonal terms b..) = w' of the double sum over the system frequencies survive
and we are left with
d
dt Ps =
EEi,j
{Ai(w)PsA it (w) AIMAi(w)Ps} + h.c. (3.194)

It remains to determine the spectral correlation tensor. To this end we use the
relations

= (OA (ObtA , (P)) = 0, (3.195)

= (5w,(5),), , (1 + N(wk)), (3.196)
= (3.197)

where
1
(3.198)
N(wk) exp [ruok] —1
denotes the Planck distribution, that is the average number of photons in a mode
with frequency Wk. We further perform the continuum limit,
00
d3k 1
f dwkw 12, f dO. (3.199)
171- (270 3 (27r) 3 c3

The integration over the solid angle dQ of the wave vector k is carried out with
the help of
d9 (6,3 kik i 87r
(3.200)
k2 ) 3 3
 THE QUANTUM OPTICAL MASTER EQUATION 145

Equation (3.191) then yields

CC CO

f dwk4 (1+ N(wk)) f dsexp[-i(wk w)s]

11(w)
'3 7r2h, 3 6 "
3c

On making use of the formula

+N(wk) f ds exp [+i(wk + w)sj
o 1 . (3.201)

CXJ
1
se -iE s = 76(E) - iP - , (3.202)
0

where P denotes the Cauchy principal value, we finally arrive at

-
Su -2 7(w) + i S ( w)) (3.203)

where we have introduced the quantities

42
l'(w) = 3h,c3 ( 1 ± N(w)) (3.204)

and
+cc
2 1+ N(wk) N (wk)
(3.205)
S(w) = P f dwk4
h,
37c3 w w bid

Note that the Planck distribution satisfies N (-w) -(1 + N(w)) such that
-y(w) = 42(1 + N(w))13hc 3 for co > 0 and 7(w) = 41w13N(14/3h,c3 for c,.) G 0.
Summing up our results we obtain a Markovian quantum master equation for
the matter degrees of freedom, often called the quantum optical master equation,
which can be written in Lindblad form,

d
— ps (t) = [His, ps(t)] D(ps (0). (3.206)
dt h,
The Hamiltonian

HLS = E h,S(w).2it (w) • _if(w) (3.207)

leads to a renormalization of the system Hamiltonian Hs which is induced by

the vacuum fluctuations of the radiation field (Lamb shift) and by thermally
induced processes (Stark shift). A more detailed treatment of the Lamb shift is
146 QUANTUM MASTER EQUATIONS

given in Section 12.2.3.2. The dissipator of the quantum master equation takes
the form
42 1r
D(Ps) = 3h3 ( 1 + N (w)) (IP)Ps2ft (w) - -2 {At MA (w ), ps
(w)A(w) s})
w>0

+ E 3hc3 N(w) (Ât(w)psif(w) _2 p(w)24 .t(w),ps}) • (3.208)

w>o
To obtain this form we have transformed the sum over the negative frequencies
into a sum over the positive frequencies and have made use of relation (3.184).
We observe that the dissipator of the master equation describes spontaneous
and thermally induced processes. By virtue of eqn (3.121) the Lindblad operators
Ai (co) lower the atomic energy by the amount ruo: If is an eigenstate of Hs with
energy E, then Ai (w)0 is again an eigenstate of Hs belonging to the eigenvalue
E - hw. Correspondingly, the operators AI(w) raise the atomic energy by h‘.4.;.
Thus, the Lindblad operators Ai (w) describe spontaneous and thermally induced
emission processes which occur with the rate 4w 3 (1 + N(w))13hc 3 , while the
Ati (w) describe thermally induced absorption processes taking place with the
rate 4w3 N(w)/3hc 3 .
Let us finally comment on the validity of the quantum optical master equa-
tion. The relaxation time TR is given by the inverse of a typical relaxation rate
'yo . The latter is defined by a typical value for the transition rates 4w3 lc/-2 /3hc3
of electric dipole transitions, where ci is the corresponding matrix element of
the dipole operator. On the other hand, the vacuum correlation time TB of the
reservoir of radiation degrees of freedom is given by the inverse of a typical tran-
sition frequency wo . This will be seen explicitly in our discussion of a systematic
perturbation expansion around the Markovian limit in Section 9.2.1. It can also
be concluded from the well-known derivation of the golden rule in quantum me-
chanics. The condition TB < TR for the Born-Markov approximation thus yields
< wo, which is obviously a weak-coupling condition. This condition is usually
very well satisfied in the quantum optical regime. For example, typical radiative
inverse atomic lifetimes are of the order 107 s -1 to 109 s -1 or even much smaller,
whereas optical frequencies are of the order 10 15 s -1 . Furthermore, it must be
noted that we need the condition Ts < TR if the master equation involves a num-
ber of different transition frequencies (see Section 3.3.1). This condition enables
one to perform the rotating wave approximation and ensures that transitions
belonging to different frequencies may be described as separate decay channels
involving different Lindblad operators.

3.4.2 Decay of a two-level system

The simplest quantum system is a two-level system whose Hilbert space is
spanned by just two states, an excited state le) and a ground state lg). The
Hilbert space of such a system is equivalent to that of a spin 4- system. In fact.
the Pauli operators
 THE QUANTUM OPTICAL MASTER EQUATION 147

07 1 = ION + 19)(e, u2 = — ile)(91+ ilg)(e , cr = le)(el —1001 (3.209)

satisfy the commutation relations

[0-i , a-i] = 2 iEjikak 7
(3.210)
as well as the anticommutation relations

{cr ,cr} = 26 ii . (3.211)

It will be convenient to define also the operators

1 1
= 1001 = (al + ig2) = LOW = (ai — zu2) • (3.212)

The corresponding matrix representations of these operators in the basis le), 1g)
take the form

and
=
(0 1 )

1 0) 7
2 — ri 0 )7 U3 —
(1 0
— 1) 7

(0 1
— 0) 7 — ( 00
10)

We take the free Hamiltonian Hs of the system to be diagonal in the basis le),
19) With an appropriate choice for the ground state energy we then have
1
Hs = -2 w0u3, (3.215)

where wo > 0 is the transition frequency and we set again h = 1. In the following
we work in the interaction picture.
Physically, such a two-level system arises whenever the dynamics of the sys-
tem under study is effectively confined to a two-dimensional subspace, that is
under the condition that transitions to other levels may be neglected. To be spe-
cific we regard here the model to describe the dynamics of a two-level atom with
a transition frequency coo in the optical range. We note that the Pauli operators
a+ represent eigenoperators of the atomic Hamiltonian,

[Hs, u-] = — wou- u-p] = +wou+ (3.216)

and, hence, a+ changes the atomic energy by the amount ±wo , corresponding to
the absorption and emission processes, respectively. We thus have two Lindblad
operators

24.(wo ) _= Â= Â(—wo ) At = cr*a±, (3.217)

where ci= (g1-61e) is the transition matrix element of the dipole operator (it
is assumed here that the diagonals of the dipole operator vanish). Within the
 148 QUANTUM MASTER EQUATIONS

two-level approximation the atomic dipole operator in the interaction picture

can thus be written as

15(t) = der _ e -iwot + dI, a+ e+iw° t . (3.218)

Neglecting the Lamb and Stark shift contribution we can now write the quan-
tum optical master equation (3.206) in the form (writing p instead of ps for
simplicity)

d 1 1
— p(t) = -yo (N + 1) (o- _ p(t)o- + - -o- ±o- _ p(t) - - p(t)o- ± a_)
dt 2 2
1 1
+fiYoN (a± p(t)o- _ - o- _ a± p(t) - (3.219)

with the spontaneous emission rate

4w3 2
7o = (3.220)
3 1 ,c3 .
The dissipator of the master equation describes spontaneous emission processes
(rate 70 ) as well as thermally induced emission and absorption processes (rates
70 N). The total transition rate will be denoted by

7 = 'Yo (2N + 1) , (3.221)

where N = N (co 0 ) denotes the Planck distribution at the transition frequency.

To solve eqn (3.219) it is convenient to represent the density matrix as

p(t) = (I + (6 (0) • 6) = ( 32. (1(0.

++((atT))) 1. ( 1(°- ((0.90)) ) , (3.222)

where the vector

11(t) (46- (t)) = tr {6.- p(t)} (3.223)

is known as the Bloch vector. It represents a real 3-vector satisfying 177(t)1 < 1.
This condition is equivalent to the requirement that p(t) must be positive. For
1'6(01 < 1 the corresponding density matrix describes a true statistical mixture.
while a Bloch vector satisfying Iff(t)1 = 1 represents a pure state. Thus we see
that the set of density matrices of the two-level system is isomorphic to the unit
sphere, known as the Bloch sphere, the surface of which is equal to the set of
pure states.
The matrix elements p H (t) = 1),(0 and p22 (t) = p9 (t) are the populations of
the excited and ground state levels, respectively. The off-diagonals p12 (t) = p'2K 1 (t)
are the coherences given by the expectation values of the atomic raising and
 THE QUANTUM OPTICAL MASTER EQUATION 149

lowering operators u±. With the help of the algebra of the Pauli matrices one
easily derives the differential equations

d
(ai(t)) = (3.224)
d
(3.225)
lt (a2(t)) =
d
(3.226)
Tt (a3(t)) =
We observe that the 3-component of the Bloch vector decays exponentially with
rate -y, while the coherences (u± (t)) decay with rate 7/2. The stationary solution
is given by

1
(Gri)s = (0-2) 8 = 0, = - ')±:) (3.227)
@Os
7 = 2N + 1'

and the stationary population of the upper level is found to be

1 N
(3.228)
P: = -2 (1 ± (u3)s) = 2N + 1 .

Of course, the stationary solution of the master equation is equal to the thermal
equilibrium state Ps = Pth given in eqn (3.144).
If we choose, for example, the initial state p(0) = 1g)(g we find the corre-
sponding time-dependent solution

(u3(t)) = - e -71 ( 1 + (0 - 3) 5 ) + (708, (3.229)

pe (t) = p: (1- e- 1 , (3.230)

showing explicitly the exponential approach to the thermal equilibrium values.

3.4.3 Decay into a squeezed field vacuum

In the preceding subsection we have considered a reservoir which is station-
ary with respect to the bath dynamics. In order to give an example for a non-
stationary environmental state we consider here a two-level system interacting
with a reservoir in a state of the form pB = 10) (0 , where 10) represents a
squeezed vacuum

10) , 11 5, (ek) 0). (3.231)

Here, we have introduced the unitary squeeze operator (Walls and Milburn, 1994)

exp [-21 eZb),(fc) 2 - -21 k btx (fc) 2 1 , (3.232)

150 QUANTUM MASTER EQUATIONS

with k rkeiek. With the help of the transformation properties

= b), (k) cosh rk (k)e iek sinh rk , (3.233)
S t (4)bt (ii)S-k , X (4) = btA (k) cosh rk -b),(fi)ek sinh rk (3.234)
one obtains the expectation values (compare with eqns (3.195)-(3.197))

= (3.235)
= Nk7 (3.236)
(bA (i)bA'(i')) = h'k1 XV Mk, (3.237)
(btx (k)bl, (rCI )) = (3.238)
where we have defined the quantities
Nk sinh2 rk, Mk — cosh rk sinh rke iek. (3.239)
We note that these relations are compatible with the commutation relations
(3.177), (3.178) of the field operators and that
111,41 2 = Nk (N k + 1). (3.240)
As for a thermal reservoir we have (f(t)) = (01É(00) = O. However, the
correlation functions (3.189) of the electric field operator are no longer homoge-
neous in time since pB is not an invariant state. As a consequence the spectral
correlation tensor now depends explicitly on time t. With the help of eqn (3.191)
the spectral correlation tensor for the squeezed vacuum can be written as follows,

26 i
= 37r h,c3
f dwk b-4 [(1 + Nk) f dse - i" - w)s + Nk f dSe ±i(wk+w)s
0 0
00 00

-Mke -2 " t f dse+ i( wk+w ) s Mk*e+2iwk t f dse -i( wk - ' )8

_
(3.241)
It has been assumed that the squeezing is uniform over the total 47r solid angle.
We observe that the spectral correlation tensor consists of two parts: The first
contribution F(w) is independent of t and is formally identical to the thermal
correlation tensor given in eqn (3.201), while the second contribution 1-' 32. ) (w)
involves the rapidly oscillating exponentials exp(±2iwkt).
Correspondingly, the dissipator of the master equation consists of two parts.
d
(3.242)
dt p(t) = D(1) (P(t)) + 1)(2) (P(t)).
—

The contribution DO) is determined through rT(w). Formally it therefore has

the same structure as the thermal dissipator given by the right-hand side of
 THE QUANTUM OPTICAL MASTER EQUATION 151

eqn (3.219), where N is the value of Nk at the resonance frequency co o = cko . It

is given in terms of the resonant squeeze parameter r as

N = sinh 2 r, N +1 = cosh2 r. (3.243)

If we neglect the Lamb and Stark shift contributions we find for the time-
dependent part of the spectral correlation tensor

2) 22
2w3 (5. m* e 2iwt
(

= for co > 0, (3.244)

3hc3 "
and
21w13 me2iwt
F(w)
id 7 for co G 0, (3.245)
3h,C3
where again M is the value of Mk taken at the resonance frequency, that is

M = — cosh r sinh re i° , (3.246)

with the resonant squeeze parameter r and the resonant phase O. In order to
determine the contribution D (2) of the dissipator of the master equation we sub-
stitute F (w) into the right-hand side of eqn (3.188) and perform the rotating
(2)
wave approximation. Since F oscillates with exp(2icot) the rotating wave
approximation selects exactly two resonant terms from the double frequency
sum, namely the terms corresponding to the case w = —coo , co' = +wo and to the
case co = +coo, w' = —we . The first case is seen to lead to the contribution
j(w ) p Ât pi) _
— A.)- pift ci*2 0.±p0.± (3.247)

whereas the second case leads to the contribution

ÂMPÂt (w I ) ÂPX (3.248)

Absorbing the phase of the dipole matrix element into the squeezing phase 61 we
therefore get

1) (2) (p) = — 7
2 ° [M o-+ po-+ + o- _ po- _] + h.c. (3.249)

Adding the first part D ( ' ) (p) of the dissipator we finally arrive at the density
matrix equation

1 1
p(t) = 70 (N +1) (o-_p(t)o-± - - o- +o- _ p(t) p(t)o-+o- _)
2
1 1
+70 (a+P(t)u- - - 0 +P(t) - P(t)(1 - +)

— 70/0-+P(t)a+ 'YoM*0--P(t)a-. (3.250)

152 QUANTUM MASTER EQUATIONS

This is the master equation describing the decay of a two-level system into a
squeezed vacuum. It is important to realize that it can be written in Lindblad
form. In fact, introducing the Lindblad operator

C a_ cosh r + cy+ sinh re if) (3.251)

we find
d 1 1

— p(t) =ryo (C p(t)Ct - - Ct C p(t) - - p(t)Ct C) . (3.252)

dt 2 2

Invoking the algebra of the Pauli matrices we obtain the equations of motion
d d
(a+ (t)) = — (t)) * (3.253)
dt dt
= (cosh2 r + sinh2 r) (o- ± (t)) + sinh r cosh re(cr (0)7
d
(0-3 (0) = (cosh2 r + sinh 2 r) (o-3 (t)) - 70 . (3.254)
dt
These equations describe a phase-dependent relaxation of the coherences and a
relaxation of the 3-component of the Bloch vector which is enhanced in compar-
ison with the decay into the field vacuum. Setting 0 = 0 the equations of motion
for the Bloch vector take the form
d 70 -2
(3.255)
(al (t)) = r(cri (t)),
(a2 (t)) = ') e-F2r (.72 (t)) (3.256)
d
— To (2 sinh2 r + 1) (0-3 (t)) - 'Yo • (3.257)

The corresponding vacuum Bloch equations (eqns (3.224)-(3.226) for ey = 70 )

are obtained in the limit r O. We observe that for a squeezed vacuum the
components (o- 1 ,2) of the Bloch vector decay with different relaxation rates: In
comparison to the vacuum case the relaxation rate is enhanced in one direction
and diminished in the other. The 3-component of the Bloch vector approaches
the stationary state

1
(u3)s = (3.258)
2 sinh2 r + 1

which has the same form as in eqn (3.227), where N = sinh2 r here plays the
rôle of the Planck distribution.

3.4.4 More general reservoirs

The form of the master equation (3.250) holds for a general environmental state
PB which can be represented by a product of independent modes satisfying
 THE QUANTUM OPTICAL MASTER EQUATION 153

(b),(fc)) = (btx (k)) = 0. For example, if we consider a squeezed thermal state,

that is if we replace the vacuum state 10)(01 used in the preceding subsection by
a thermal equilibrium state and squeeze the latter, we get instead of eqns (3.243)
and (3.246)

N = Nth (cosh2 r + sinh2 r) + sinh 2 r, (3.259)

M = - cosh r sinh rem (2Nth + 1), (3.260)

where we have written Nth = 1/(exp[8w0] - 1) for the Planck distribution. Ob-
viously, the quantities N and M are not independent. One easily verifies the
relation

1m1 2 = N(N + 1) - Nth (Nth + 1). (3.261)

The last equation shows that the inequality

1M1 2 < N(N +1) (3.262)

holds. This inequality is satisfied for general pB of the above form and may be
proven with the help of an appropriate uncertainty relation. To this end we take
a fixed mode bb),(k) and write

(bt b) = N (bbt) = N + 1, (3.263)

(bb) = M = (btbt) = m. (3.264)

This is the most general parametrization which is compatible with the commu-
tation relations [b,b1= 1. Next one defines the operators

e _i o/2 b e+ 2 bt ) (3.265)
X =—
1= (e -i9 / 2 b + e±i62 bt)

which are Hermitian and satisfy (X) = (Y) = 0 and [X, = - i. The un-
certainty relation (X 2 ) • (Y 2 ) > 1/4 then immediately leads to the inequality
(3.262).
The significance of the inequality (3.262) is that it guarantees the positivity of
the generator of the master equation (3.250). In fact, the generator of eqn (3.250)
is in the first standard form (3.63) with F1 = u_, F2 = a+ and the coefficient
matrix

(aii) (N + 1 --M*
(3.266)
-M N

The positivity of this matrix leads to the condition (3.262).

154 QUANTUM MASTER EQUATIONS

3.4.5 Resonance fluorescence

Let us consider the case that the atomic transition le) 1g) is driven by an
external, coherent single-mode field on resonance. Invoking the dipole approxi-
mation we obtain the following interaction picture Hamiltonian describing the
interaction of the system with the driving mode,

HL = - E. L(t) • _b. (t). (3.267)

Here, the quantity

EL(t) = ee - iwot + e* eot (3.268)

is the electric field strength of the driving mode. The product
Q -= 2e • cr* (3.269)
is referred to as the Rabi frequency. In the following we choose the phase of
the external field such that 12 is real and positive. With these definitions we can
write the atom-field interaction in the rotating wave approximation as follows,
S2
HL = - — (u+ + u-) . (3.270)
2
Now we couple the system to a thermal reservoir of radiation modes. Accord-
ing to Section 3.4.2 the quantum master equation becomes
d iS2
p(t) = — [4(7 + + (7- ,P(t)]
—

dt 2
+70(N + 1 ) ((7-P(t)(7+ - u-P(t) - -21 P(t)(3"+u-)
1 1
+70N (a+P(t)(7- - -2 0- -(7+P(t) - -2 P(t)(3"-u+) • (3.271)

3.4.5.1 Equation of motion for the Bloch vector The master equation (3.271)
leads to the following system of differential equations,

= G(.6- (t)) +6, (3.272)

dt
where we have introduced the matrix
(-7/2 0 0 )
G= 0 -7/2 Q (3.273)
0 --S2 7 -

and the 3-vector

(0
g= 0 . (3.274)
70 —

Equation (3.272) is known as the optical Bloch equation.

 THE QUANTUM OPTICAL MASTER EQUATION 155

The stationary solution of the differential equation (3.272) is found to be

(0 _ 3 ) 8 707 , (3.275)
72 + 2S12
ifrYo
—

(3.276)
= = .72 + 2S-12 •

The stationary population of the upper level is therefore

p: 1 7(7 — 'To) + 292 (3.277)

2 72 + 2C22 •

At zero temperature (N = 0, -y = 70 ) the stationary solutions take the form

2
70 (3.278)
(a3)8 = 2
70 ± 2112'
(a+ )8 = (a- ): = 2-4 97° 2 ' (3.279)
70 + 21l '

while the stationary population of the upper level is

112
(3.280)
7o + 292.

In the strong driving limit S-) >> -yo we get the limiting forms pse = 1/2 and
(a+ ), = —i-y0 /2Q.
The time-dependent solution of the Bloch equation (3.272) is most easily
found by introducing the vector

(V1(t))) = (6 (t)) — ( 6)8. (3.281)

Representing the difference to the stationary solution this vector satisfies the
homogeneous equation

d
it (V1 (t))) =
E (3.282)

which can easily be solved by diagonalization of the matrix G. The eigenvalues

of G are given by

x (3.283)
A1 — — -2- , '12,3 - - —
4
± tl-i,

where we have introduced

(3.284)
156 QUANTUM MASTER EQUATIONS

0t

FIG. 3.3. The population of the upper level IA (t) for the driven two-level system
as a function of time. Parameters: -yo /SI = 0.1 (solid line), 70 /Q = 0.5 (dashed
line), and '70 /S2 = 1 (dashed-dotted line).

We note that all eigenvalues have negative real parts. It follows that any initial
density matrix is driven to the unique stationary solution (.6), of the Bloch
equation determined above. The time-dependent solution can be represented by

(0-3 (t)» = e -37t /4 [cos lit - iTt sin [it]

iSi
+ e -371 /4 sin pt Ma- ± (0))) - ((a _(0)))) , (3.285)
tt
and
1
((a+ (0)) = et/2 (((a+(0))) + ((a- (0 ))))
1
+ e -3't/4 [cos pt + 7 sin /it] (((a+(0)))
- — -

2 4p

sin pt((u3( 0 ))). (3.286)

2p

Let us consider here the case that the atom is initially in the ground state,
p(0) = g) (g, which gives

((u± OM = - (0-±)8, (3.287)

( c1 3(0)» = -1- (c13)8. (3.288)
 THE QUANTUM OPTICAL MASTER EQUATION 157

S-1t

FIG. 3.4. The imaginary part of the coherence (o-±(t)) for the driven two-level
system as a function of time. Parameters: -yo /Q = 0.1 (solid line), -yo /Q = 0.5
(dashed line), and -yo /Q = 1 (dashed-dotted line).

At zero temperature (-y = 70 ) we find for arbitrary driving strength

112 [
pe (t) = 1 _ e -3-yo t/4 {COS pt + 34—
'TO
tt si n } (3.289)
7,3 + 2S12
iQ'Yo
[i — e -37°t /4 {COS fit + (31)- - --
-

(a+ (0) = 2 °2 ) sin pit}1(3.290)

7o + 2Q2 4p, 70 p

In the underdamped case, which is defined by Q > -y0 /4, the quantity p is real and
both the occupation of the upper level and the coherences exhibit exponentially
damped oscillations as shown in Figs. 3.3 and 3.4. By contrast, in the overdamped
case, Q G 70 /4, the quantity p becomes purely imaginary and we write

(7o\ 2
f . (3.291)
4 ) _ p
The occupation probabilities and the coherences then show a monotonic ap-
proach to their stationary values.
In the limit of very strong driving, Q >> 70 /4, we obtain the asymptotic
expressions,

, 1 _ e -3-yot/4 cos Sit]

Pe (t) 7:-.„' — [I_
7 (3.292)
2
i
(u(t)) ssss — e -37°1/4 sin Qt. (3.293)
158 QUANTUM MASTER EQUATIONS

The tip of the Bloch vector thus follows an exponential spiral in the (2, 3)-plane
around the origin, the circular frequency being equal to the Rabi frequency Q.

3.4.5.2 Spectrum of resonance fluorescence As an application of the quantum

regression theorem (see Section 3.2.4) we determine the fluorescence spectrum of
the emitted radiation in the stationary state. The radiation spectrum is deter-
mined by a certain two-time correlation function of the system. To see this we
first note that the positive frequency component of the electric dipole field radi-
ated by the source is found from the retarded solution of the Maxwell equations
to be
2
f (±) (t, i) = cW2°r ( [ 11 X d
-- X 71) (7_ (t — r/c). (3.294)

Here, the dipole is located at the origin of the coordinate system, r = A and
71 = Z/r is the unit vector pointing into the direction of Z. This equation relates
the positive frequency component of the radiated field to the atomic lowering
operator a_ at the retarded time t — r/c. On the other hand, the spectral intensity
radiated per unit solid angle by the oscillating dipole is given by
+00
d/(w) cr2 f dr e i„(f(—) (t, z)f(+)(t
+ 1-, (3.295)
c/S2 47r j 27r

Substituting eqn (3.294) one finds

di (C41) w4 2
= 0
(ri x d") x ri S(C4J), (3.296)
c/S) 872 c3
which shows the typical angular distribution of dipole radiation, whereas the
frequency dependence is embodied in the spectral function
+00
S(w) = f dTe'"(o-± (7)o- _(0)) s . (3.297)
—00

Here, we have employed the homogeneity in time of the stationary correlation

function.
Thus, in order to obtain the spectrum we need the atomic correlation function
(o-±(T)o- _(0)) 8 . To this end, we consider the vector

— (5.- ) 8 (u _) s . (3.298)
It follows from the quantum regression theorem that the dynamics of this vector
is determined through the homogeneous part of the optical Bloch equation, i.e.
d
cw (V- (T)o-_ (0))) = G((d- (r)o- _ (0))) . (3.299)
 THE QUANTUM OPTICAL MASTER EQUATION 159

1.2

FIG. 3.5. The spectrum of resonance fluorescence according to (3.305) without

the elastic contribution. The plot shows S(w) for the three values 4Q/-yo = 2
(solid line), 4S1/-yo = 8 (dashed line), and 412/')/0 = 16 (dashed-dotted line).

We conclude that the desired correlation function ((a + (T)o- _ (0))) is given by the
right-hand side of eqn (3.286), where the initial values ((o-±(0))) and ((o-3 (0)))
must be replaced, respectively, by
((a+ (0)a_ (0))) = (a+ a_ ), - (a+ ) 8 (cr- )8 = p - ((3- +) s12 (3.300)
= ( 0. - 0- 49 - (3.301)
(0-3 (0)a_ (0))) (a3 a_ ) (o-3 ),(u - ), = (a_ ), (03),(a-), . (3.302)
Finally, it must be taken into account that the Bloch equations have been
solved in the interaction picture. To return to the Schrödinger picture we replace
g±(t) by exp[±iwo t]a±(t). This yields the two-time correlation function in the
stationary state
2 1
= „ -Yo exp [iwo r] + - exP [ - (7o/ 2 - iwo)r]
yo + 2Q2 yg + 2Q2 2

+A+ exp [- CI)4

-((2 - - Ti
3 +
+A_ exp [- (-P iwo) , (3.303)

where we have defined the amplitudes

1 1
A± = 2 [ 2 -2 2 — .2‘r) (10S12 - -y,3)] . (3.304)
4 -yo + 2S2 2
160 QUANTUM MASTER EQUATIONS

Equation (3.303) gives the two-time correlation for T > 0; for negative times it
is found with the help of the relation (a+ (—Oa_ (0)), = (a+(r)o- _(0)):.
The correlation function is seen to consist of four terms. The first term rep-
resents the contribution from elastic Rayleigh scattering which gives rise to a
6-shape spectral function (see below). The other three terms describe inelastic
scattering processes and yield Lorentzian line shapes. In the overdamped case.
<70/4, the quantity ji is real, such that the spectral function contains a sum of
three Lorentzian peaks centred at w = coo. If one goes over to the underdamped
case, Si > -y0 /4, the quantity ji = — ill becomes imaginary and the single inelastic
peak splits into three peaks centred at co = coo and co = coo ± p. In fact, on taking
the Fourier transform of (3.303) we find the spectral function for the fluorescence
radiation of an atomic dipole in the underdamped case (see Fig. 3.5):

2
2Q2 7rN 1 70/2
2 1l 2 6(co co0 ) + (3.305)
2 (70/2) 2 + (w — wo) 2
A+ (370/4 i [ci.) wo + A_ (3-y0/4 i [b.) wo 1-1])
+R (3 7o/4)2 ± _ R
wo + p) 2 (3-7o/4) 2 + (w — wo — }

In the limit of strong driving, 1 » 70 /4, the amplitudes A± approach 1/4 and the
elastic scattering contribution is negligible. This yields the following spectrum
in the strong driving limit,

1 70/2
= 2 (70/2)2 + (co _ wo)2 (3.306)

1 3 7o/ 4 1 3 7o/ 4
+ 4(37/4)2 + (w — wo _Q)2
4 (3-y0 /4) 2 + (co — coo + Si)2

The spectrum thus consists of three Lorentzian peaks: A central peak at co = wo

with width 70/2, and two sideband peaks at ci.) = co0 ± Si with width 370 /4. The
heights of the peaks are in the ratio 1: 3 : 1, while their integrated intensities are
in the ratio 1 : 2 : 1. The above spectrum was derived theoretically by Mollow
(1969), and verified experimentally by Wu, Grove and Ezekiel (1975) and by
Hartig et al. (1976). Further interesting features of the two-level system, such
as the photon statistics and the phenomenon of photon antibunching, will be
discussed in Section 6.3.2.

3.4.6 The damped harmonic oscillator

As a further example of a dynamical semigroup we study the master equation
for a damped harmonic oscillator. The free evolution is generated by the system
Hamiltonian Hs = woata, describing a harmonic oscillator with frequency wo.
The Schrödinger picture master equation reads
 THE QUANTUM OPTICAL MASTER EQUATION 161

— p(t) = -ico0 [at a, ps(t)]

dt
1 1
+70(N +1) faps(t)a - -at aps(t) - - ps(t)at a}
2 2
1 1
+70 {a t ps(t)a — — aat Ps (t) — — ps(t)aat}
2 2
E rPs (t)- (3.307)

This equation may be used, for example, to describe the damping of an electro-
magnetic field mode inside a cavity, in which case at and a denote the creation
and annihilation operators of the cavity mode. The environment is provided, for
example, by the modes outside the cavity and leads to a damping of the cavity
mode with a rate 70 . The quantity

N = [exp(wo /kB T) -1] -1 (3.308)

is the mean number of quanta in a mode with frequency co o of the thermal

reservoir.

3.4.6.1 Pauli master equation and stationary solution We denote by In), n =

0,1, 2, ... , the n-th oscillator eigenstate, i.e. ataln) = nIn) • The master equation
(3.307) leads to a closed equation for the populations of the n-th level,
P(n,t) = (71 1Ps(t) n), (3.309)
which takes on the form of a classical master equation for a one-step process (see
Section 1.4.4.2), the Pauli master equation
d
— P(n t) = 70 (N +1)[(n +1)P(n + 1,t) - nP(n,t)]
dt '
+'YoN [71P(n - 1, t) - (n + 1 )P( 71 ,t)] .
The stationary solution of this equation is given by

P$ (n) =
1 ( N VI
1+N1+Nj .
Invoking eqn (3.308) this can be rewritten as

Ps (n) = [1 — exp (—w o /k B T)] exp [—nwo /kBT] , (3.312)

which is recognized as the Boltzmann distribution over the eigenstates of the
oscillator. The stationary value for the mean number of quanta in the mode is
thus equal to the thermal average,
CO

(ata), = E nPs (n) = N. (3.313)

n=0
162 QUANTUM MASTER EQUATIONS

3.4.6.2 Adjoint master equation To investigate the dynamics one may deduce
equations of motion for expectation values of system operators directly from the
master equation. Alternatively, one can use the adjoint master equation (3.84)
which becomes
d
— A H (t) = +iwo [ata, AH(t)]
dt
+70 (N + 1) {at A H (t)a - 2-1 ataA H (t) - -21 A H (t)ata}

+70N faA H (t)at - -21 aatA H (t)- -21 A H (t)aat} , (3.314)

in order to determine the dynamics of the corresponding Heisenberg operators

AH(t). Solving this equation with the initial condition that Heisenberg and
Schrödinger operators coincide at t = 0 one finds, for example,

aH (t) = e(-iw 0--012)t a, (3.315)

41 (0 = e (+iwo —ro/ 2 )t at (3.316)
(ata) H (t) = e -7° t ata + N (1 - e --)T' t ) . (3.317)
From these equations one immediately deduces the mean amplitude of the oscil-
lator,

(a(t)) = tr faH(t)ps(0)} = (a(0))e(-jw0 0 I2 (3.318)

and the mean number of quanta,

(ata(t)) = tr { (Oa) H (t) p s (0)1 = (ata(0))e—Y0t N ( 1 _ e —yo t) (3.319)

The mean oscillator amplitude thus follows an exponential spiral in the complex
plane, converging for -yo t >> 1 to the origin. The mean number of quanta is
seen to start from the initial value (ata(0)) and approaches, for times which are
large compared to the inverse damping rate, the thermal average N which is
independent of the initial value.

3.4.6.3 Coherent state representation In quantum optical applications one of-

ten studies the master equation with the help of phase space methods. As an
example we consider here the coherent state or P-representation of the master
equation. More details and various other related representations may be found
in Gardiner and Zoller (2000).
For any complex number a the coherent state la) is defined in terms of the
number states In) of the oscillator by
00
la) = expIctI 2j an In) = exp [aat - a* al 10) . (3.320)
n=0
 THE QUANTUM OPTICAL MASTER EQUATION 163

The coherent states are normalized to 1 but are not orthogonal to each other.
The overlap of two coherent states is given by
1 212
(a1r3 ) = exp --2 lal - -2 IA + a * 0 • (3.321)

The coherent states are eigenstates of the non-Hermitian operator a, that is

ala) = ala)• (3.322)

They form an overcomplete set of states satisfying the completeness relation

-1 =1. (3.323)
7r f d2a1a)(a1
The coherent state representation of the system's density matrix goes back
to Glauber (1963) and Sudarshan (1963) and is defined by

p(t) = f d2 aP(a , a* , t) a) (a (3.324)

Thus, p(t) is represented as a mixture of coherent states la) with a correspond-

ing weight function P (a, a* , t) which is uniquely defined by the given density ma-
trix. It should be noted, however, that P is only some kind of quasi-probability for
it is, in general, not a positive distribution function. It can been shown (Klauder,
McKenna and Currie, 1965) that the distribution P exists for any density matrix
provided one permits P to be a generalized, singular function. The normalization
of the density matrix leads to the normalization condition

trs ps(t) = f d2 aP(a, a* ,t) = 1. (3.325)

A product of (equal-time) creation and annihilation operators at (t), a(t) is said

to be normally ordered if all creation operators stand to the left of all annihilation
operators. For such a normally ordered product we find

((at (t))P a(t)q) = f d2 a(a* )P aq P(a, a* , t). (3.326)

This is an important property of the P-representation which means that the

expectation value of any normally ordered product of creation and annihilation
operators transforms into a corresponding moment of the distribution P (a , a* , t).
To derive an equation for P (a, a* ,t) one substitutes the P-representation
into the quantum master equation (3.307) and uses the properties
ala)(al = ala)(al, (3.327)
1a)(ala t = ala)(a1, (3.328)

at la)(al = (79-
8c-t l+a la)(a1, (3.329)

(a
la)(ala = I acts: +a ) (3.330)
164 QUANTUM MASTER EQUATIONS

which may easily be deduced from the definition for the coherent states. Next.
one performs an integration by parts (assuming zero boundary conditions at
infinity) which introduces an additional minus sign for each differential operator.
The above properties then show that we have the following correspondences,
ap s aP, (3.331)
Psat H a* P, (3.332)

atps H (
a P, (3.333)
aa
ps a
( a
, + a) P. (3.334)
oa*
With the help of these relations on easily derives the following Fokker-Planck
type equation of motion in the P-representation, (Scully and Zubairy, 1997),

a 7° )
a a s] P (a, as: ,t)
a * , t) = - [( - iwo - 21) a a + (+iw
) --
2 aa ° 2 aa*
a2
+-yoN P (a a* t) (3.335)
a aaa* ' ' '
The structure of this equation is similar to that of the classical Fokker-Planck
equation (1.129) for a diffusion process: The first term on the right-hand side
represents the deterministic drift, while the second term has the structure of a
diffusion term.
Let us solve eqn (3.335) for the initial value
P(a, a*, 0) = 62 (a— ao), (3.336)

which means that the initial state is the coherent state a o ). The corresponding
solution can be obtained by substituting the Gaussian ansatz

P(a, a*, t) = 1 exp [

7ro-2 (t)
a - Nt)12
0-2 (t)
1 (3.337)

into (3.335) and deriving differential equations for the time-dependent functions
o-2 (t) and /3(0. However, a simpler way is to observe that NO is equal to the
mean amplitude since

w --t
12) .
0 (t) = f d2 aa P(a , a*, t) = (aolaH(t)lao) = ao0yo (3.338)

Correspondingly, we find that a2 (t) is equal to the variance of at a which can be

obtained using (3.317),
o-2 (t) = Var (at (t)a(t))
= (NI (a f a)H (t)lao) - aa0e —Y0t
= N (1 - e 0). (3.339
 THE QUANTUM OPTICAL MASTER EQUATION 165

al

FIG. 3.6. Coherent state representation of the damped harmonic oscillator. The
figure shows the distribution function P (a, a* , t) according to eqn (3.337)
over the complex phase plane with coordinates a = a l + ia2 at three different
times given by co o t = 370, 77r/ 4, 327r. Parameters: 70/coo = 0.1, ao = 1, and
N = 0.025. The trajectory of the mean amplitude NO is also shown.

The P-representation is therefore given by a Gaussian function whose mean value

spirals around the origin of the phase plane. The width o-2 (t) increases from
zero to the thermal average N when the oscillator comes to equilibrium with
the surrounding reservoir (see Fig. 3.6). We also note that at zero temperature
(N = 0) the initial coherent state remains a coherent state under the time
evolution.
The Fokker—Planck equation (3.335) describes the evolution of a complex
Ornstein—Uhlenbeck process (see Section 1.4.5), the equivalent stochastic Itô
differential equation (see Section 1.5.4.1) being given by

da(t) = ( iwo
— — ) a(t)dt ± V -yo N dW (t). (3.340)

Here, dW (0 is the increment of a complex Wiener process which satisfies

E[dW (0] = E[dW * (t)] ,- 0, (3.341)

dW (t)dW (t) = dW* (t)dW* (t) = 0, (3.342)
dW (t)dW * (t) = dt. (3.343)
166 QUANTUM MASTER EQUATIONS

The complex increment can be represented in terms of two statistically indepen-

dent real Wiener processes as follows

1
dW(t) = — (dW i (t) idW2(t)) , (3.344)

which is easily verified with the help of the relations (1.184)–(1.186).

Thus we observe that for appropriate initial conditions which give rise to
a true probability distribution in the P-representation, the dynamics of the
complex amplitude of the oscillator can be modelled in terms of the classical
stochastic process a(t). It must be noted, however, that this does not imply, as
we already know, that all features of the dynamics can be described with the
help of a classical stochastic process on the phase plane. The quantum optical
master equation (3.307) for the damped harmonic oscillator will be encountered
at several occasions in the following chapters. For example, it will be used for
a study of decoherence phenomena in Chapter 4, in the context of a stochastic
analysis of continuous measurements in Section 6.7, and for a test of numerical
simulation algorithms in Section 7.3.1.

3.5 Non selective, continuous measurements

-

Density matrix equations can be used to describe the system evolution which is
induced by a non-selective, continuous measurement of some observable of the
system. An ideal and continuous measurement gives rise to the so-called quantum
Zeno effect which will be described in the first subsection. We then turn to the
description of a continuous, indirect measurement of a system observable and
derive an equation of motion for the density matrix of the object system.

3.5.1 The quantum Zeno effect

We consider the measurement of an observable A whose spectrum is assumed.
for simplicity, to be purely discrete and non-degenerate,

ani 7MONI. (3.345)

A

A series of ideal and instantaneous measurements of the observable A is carried

out in such a way that two successive measurements are separated by a fixed
time interval 0. In between two successive measurements the state vector evolves
according to the Schr6dinger equation

i -:1 7i)(t)) = Hili5(0) (3.346)

with some Hamiltonian H. In the limit 0 —> 0 one speaks of a continuous

measurement of the observable A.
 NON-SELECTIVE, CONTINUOUS MEASUREMENTS 167

Let us assume that the system is initially in an eigenstate 10n ) belonging to

the eigenvalue an of A,

IV)(0)) = (3.347)
For sufficiently small t we then have according to the Schrödinger equation,

110) = [I — iHt — —1 H2 t2 ± ...]111)n ). (3.348)

2
The first ideal measurement of A is carried out at time t = 0. The probability of
obtaining the eigenvalue a, in this measurement is given by

w(0) = Pni/i)(0))12 _ 1 — (AE)202 +..., (3.349)

where the dots indicate terms of higher order and

(AE ) , = (OniI/2 10n) — (Orilli

on )2 (3.350)
is the energy uncertainty in the state W). The quantity w(0) is just the
probability that the system is still in the initial state PO after time O. After
time T = 109, that is after k measurements this probability is thus

tOnn(T) Pe., [1 _ (AE)2no2]k . (3.351)

For large k and fixed T, that is in the limit 0 = rIk 0, this leads to
-k
[i — (AE) 2 122- z,- exp [—(AE) 2n 7-0] 1. (3.352)
nk
This equation tells us that the system remains with probability 1 in the initial
state 10n ) if a continuous, ideal measurement of the observable A is carried out
on the system. As a result of the continuous state reduction induced by the
measurements the system cannot leave its initial state. Formally, the reason for
this fact is that in the limit of small 0 the probability of leaving the state IOn )
is proportional to the square of 0, that is we have 1 — w(8) oc 02 , while the
number k of measurements increases as 8 -1 . The state reduction induced by the
succession of measurements is thus faster than possible transitions into other
states. In analogy to Zeno's paradox this phenomenon is known as the quantum
Zeno effect.

3.5.2 Density matrix equation

The quantum Zeno effect described above results, or course, from an idealization
which presupposes an ideal, continuous measurement. In order to investigate
what happens for non-ideal measurements we consider the following model for
an indirect continuous monitoring of the system.
168 QUANTUM MASTER EQUATIONS

As in the preceding subsection we first subdivide the time axis into inter-
vals of length 1. Over each interval the object system evolves according to the
Schr6clinger equation (3.346). However, instead of performing ideal measure-
ments at the beginning of each interval we carry out indirect measurements of
A (see Section 2.4.6). At the beginning of each interval a single probe particle
with coordinate Q interacts with the quantum object through an interaction
Hamiltonian of the form 1//(t) = g(t) AQ . The probe particles are independent
and supposed to be in identical initial states 100). After the object-probe inter-
action the momentum P canonically conjugated to Q is measured directly on
each probe particle. This leads to a quantum operation which is described by
the operators (compare eqn (2.217))

(pI exp( - iG AQ)I0 0 ) , (3.353)

where p) is a momentum eigenstate and G = f dt g (t) = 0 is the integrated
coupling strength which is assumed to scale with B.
Including the coherent evolution generated by H we thus have an operation
described by the operators

S21, = exp (-i1/8) (pI exP ( --i AQ 0 ) 100) . (3.354)

The evolution of the object's density matrix over a time interval 0 on the non-
selective level can thus be written as

p(t + 0) = f dpQp p(t)S21) . (3.355)

Substituting expression (3.354) and expanding the coherent part to first and the
incoherent part to second order in 0 one obtains

p(t + 0) = p(t) - i[H, p(t)] 0 - i[A, p(t)] (Q) 0 0

1 1
+ [24. p(t) A - -2 A 2 p(t) - -2 p(t) A21 (c2 2 ) 09 2 , (3.356)

where we have defined

(Q)0 = (001(2100), (C22 )0 = (001(22 00)• (3.357)

We now assume that (00 = O. This means that the drift contribution
-i [A, p(t)] (Q)0 vanishes and, therefore, that the measuring device does not lead
to a systematic back-action on the dynamics of the quantum object. Further-
more, it will be assumed that the limit

al? E lim 0(Q2 ) 9 (3.358)

o-40
exists and is finite. It follows that the stochastic back-action induced by the
measurement device remains finite in the continuum limit 0 O.
 NON-SELECTIVE, CONTINUOUS MEASUREMENTS 169

In the continuum limit we therefore obtain the following equation of motion

for the state of the object system,

d 1 1
— p(t) = —i [H, 1)(0] + a6 [A p(t) A — A 2 p(t) — p(t)A 2 1
dt
1
= —i [H, p(t)] — ygl? [A, [A, p(t)]] . (3.359)

The right-hand side of this equation obviously takes on the form of the genera-
tor of a quantum dynamical semigroup. The generator is of the form obtained
in Section 3.3.3 for the singular-coupling limit. It involves a single Hermitian
Lindblad operator A given by the observable being measured indirectly.
The dissipator of the master equation describes the back-action on the object
system induced by the measurement device. We observe that the strength of this
back-action is related to the accuracy of the A-measurement. Namely, assuming
(P)0 , 0 we have by virtue of the uncertainty relation
ip2 \0 1
\ 0 f e(Q 2 )0 ? —4 . (3.360)

Taking the limit 19 0 one obtains

2 2 -.., 1
(3.361)

where we have defined

= li1 p2)0
aA2 rn ‘ , . (3.362)
0-40 bf

The quantity o - A is a measure for the uncertainty of the A-measurement (see Sec-
tion 2.4.6) and eqn (3.361) tells us that the back-action induced by the measure-
ment is the stronger the more accurate is the A-measurement. Equation (3.361)
thus provides a kind of uncertainty relation for the accuracy of the monitoring
and the resulting fluctuating back-action on the measured system. If we take A
to be dimensionless the quantity o-6 has the dimension of an inverse time,

To E [u]', (3.363)

which may be referred to as the Zeno time. In fact, the time To is proportional
to the accuracy of the A-measurement. For vanishing Zeno time, To 0, that
is for an arbitrarily accurate A-measurement, one is led back to the quantum
Zeno effect described above.
The interplay between the characteristic time scale of the coherent motion of
the object system and the Zeno time To can be nicely illustrated by considering
again a two-level system. The free object Hamiltonian is taken to be H = —
describing coherent Rabi oscillations with frequency 51 . The measured quantity
170 QUANTUM MASTER EQUATIONS

is chosen to be A = o-3 , that is the measurement device is designed to monitor

whether the system is in the upper level le) or in the ground state l g). The
corresponding density matrix equation thus takes the form
d iS2 —1
p(t) = — [0-1, P(t)] [0-3, [0-3, P(t)]] (3.364)
dt 2 c,To

which leads to the following system of differential equations for the components
of the Bloch vector,
d
- (ai (t)) = (3.365)
t To
d 2
— (a2(t)) = — (0-2 (t)) + Q( 0-3 (t)) (3.366)
dt TO
d
it (Gr3(t)) = — Q( 0-2(0)- (3.367)

Let us consider the case that the object system is initially in the upper level.
which means p(0) = le)(el such that (cri (0)) = (cr2(0)) = 0 and (o-3 (0)) = 1. The
corresponding solution of the above differential equations leads to
(o-3 (t)) = 1 (p2e — kti t — Pie' t ) (3.368)
tt2 to_
where the characteristic frequencies are given by

tt1,2 = (3.369)
To To

With the help of these relations the population of the upper level is given by
1
pe (t) = (1 + (cr3 (t))) . (3.370)

Since the real parts of ,tt i ,2 are always positive the occupation probability
1),(0 of the upper level approaches 1/2 in the long time limit. This means that
the back-action of the measurement device drives the quantum object into a
stationary state with equal populations of upper and lower level. In the under-
damped case S-27-0 > 1 the characteristic frequencies are complex. We write them
as /1 1,2 = 1/To ± iv with

v= Çl (3.371)

Then we have
1
(0- 3(t)) = (cos vt + — sin vt) e —t/. (3.372)
VT0

This shows that 1),(0 follows an exponentially damped oscillation with frequency
v. For QT0 » 1, that is for a Zeno time which is much larger than the inverse
 NON-SELECTIVE, CONTINUOUS MEASUREMENTS 171

1
nt o .0 (Zeno limit)

0.8 -

0.6 -

aw

0.4

0.2

a
0 5 10 15 20
01

FIG. 3.7. The occupation probability pe (t) (eqn (3.370)) as a function of time
for three different values of the Zeno time To.

of the Rabi frequency, the object system oscillates many times with a nearly
unperturbed frequency y ',:.- Q. If the Zeno time is decreased the frequency y
becomes smaller and the coherent oscillations die out at QT0 , 1.
For the overdamped case QT0 < 1 one has a monotonic approach to the
stationary state. Thus, for a sufficiently small Zeno time the coherent oscillations
of the unperturbed motion are completely suppressed (see Fig. 3.7). In the case
f27-0 < 1 one finds the approximate solution

(o-3 (t)) (3.373)

where the time constant T is given by

2
T=
Q2 To . (3.374)

The decay time of the upper level thus becomes inversely proportional to the
Zeno time To . In the limit TO 0 we have T oc. This limit corresponds to
infinite accuracy of the a3 -measurement and to the emergence of the quantum
Zeno effect: The lifetime of the upper level becomes infinite, i.e. the object system
remains frozen in its initial state. We remark that the occurrence of the quantum
Zeno effect has been observed experimentally in a similar system (Itano, Heinzen,
Bollinger and Wineland, 1990).
172 QUANTUM MASTER EQUATIONS

3.6 Quantum Brownian motion

In the preceding sections we were mainly concerned with the investigation of
master equations which arise in the quantum optical limit. In this limit the ro-
tating wave approximation enabled the derivation of Markovian density matrix
equations defining a quantum dynamical semigroup. The physical condition un-
derlying the rotating wave approximation is that the systematic evolution of the
reduced system is fast, which means that the coherent dynamics goes through
many cycles during a typical relaxation time. In several physical applications
involving strong system-environment couplings and low temperatures this con-
dition is violated. By contrast to the typical situation in quantum optics it may
even happen that the systematic dynamics of the reduced system is slow com-
pared to the correlation times of the environment. This is the characteristic
situation of quantum Brownian motion. Under such circumstances another type
of master equation arises which requires an approximation scheme that differs
from the quantum optical one.
In this section we are going to introduce and discuss in detail a specific
system-reservoir model which is appropriate for the study of several aspects of
quantum Brownian motion. This is the famous Caldeira-Leggett model (Caldeira
and Leggett, 1983) which is a prototype of a system-reservoir model for the de-
scription of dissipation phenomena in solid state physics (Weiss, 1999). In the
high-temperature limit the model leads to a quantum master equation which is
quite different from the ones obtained in the quantum optical case, but is still
of Markovian nature. However, for low temperatures and/or strong couplings
the reduced system dynamics exhibits a pronounced non-Markovian character
and one must resort to different techniques in order to describe the reduced sys-
tem dynamics. We shall discuss here the use of the exact Heisenberg equations
of motion for the coupled system, the fluctuation-dissipation theorem and the
prominent Feynman-Vernon influence functional technique (Feynman and Ver-
non, 1963). More details on the treatment of non-Markovian processes and a
systematic perturbation expansion are postponed to Part IV.
3.6.1 The Caldeira-Leggett model
The model describes a Brownian particle of mass m with coordinate x which
moves in a potential V(x). The free Hamiltonian Hs of the particle is thus taken
to be
1 2
(3.375)
HS 21nP + 17(X)'
= —
where p is the particle momentum. The particle is assumed to be coupled to a
bath consisting of a large number of harmonic oscillators with masses m i, and
frequencies wi, described by the Hamiltonian

(3.376)
 QUANTUM BROWNIAN MOTION 173

Here, the bn , bfn denote the annihilation and creation operators of the bath
modes, while the x n and pn are the corresponding coordinates and canonically
conjugated momenta.
In the model under consideration the coordinate x of the Brownian particle
is assumed to be coupled linearly to the coordinates xn of the bath oscillators.
The corresponding interaction Hamiltonian HT takes the form

H., = _x E K n X n E —XB, (3.377)

n

where the bath operator

\I
E KnXn = E Kn
h
B=
2711 n Wn
(bn + btn ) (3.378)
n n

is a weighted sum over the coordinates x n of the bath modes involving the
corresponding coupling constants Kn . This type of interaction will be seen later
to yield a renormalization of the potential V(x) of the Brownian particle. To
compensate for this renormalization it is convenient to include a further term
into the interaction Hamiltonian which is of the form
2
H, = x2 E 2mnw,22 .
Kn
(3.379)
n

This term, known as the counter term, acts only in the Hilbert space 7 is of the
- -

Brownian particle. It ensures, as will be shown later, that the potential V (z)
involves the physical frequencies of the motion of the Brownian particle.
The total Hamiltonian of the combined system S + B is thus given by

H = Hs + HB + HI ± II, (3.380)

=
2
1
—

m
■11

p' + V,(x) + E(
1 p2n ± _1 in n b.) 2 x2
277-t n 2 nn
_
X E KnXn •
n n

In the second expression we have included the counter-term H, into the potential
of the particle,
2
Ve (X) = V(X) +z2 E 2m w2 • Kn

nn
(3.381)
n

3.6.2 High temperature master equation

-

We start our discussion with an investigation of the simplest case, namely of

the motion of the Brownian particle in the weak-coupling and high-temperature
limit. This limit allows the derivation of a Markovian master equation for the
reduced density matrix of the particle, which is known as the Caldeira—Leggett
master equation.
174 QUANTUM MASTER EQUATIONS

3.6.2.1 Derivation of the master equation The starting point is the Born-
Markov approximation for the reduced density matrix of the Brownian particle
which can be written in the Schrödinger picture as follows,
d
Ps(t) = —ih [Hs + H, Ps (t)] ± 1CPs(t), (3.382)

where we have introduced the super-operator

CO

1
/Cps (t) = — dr trB [1-//, [11i(--r), ps (t) 0 ABU. (3.383)
o
This expression may be derived easily by transforming eqn (3.118) back to the
Schrödinger picture. Note that the counter-term Hc must be treated as a term
of second order in the coupling, while HI is of first order, such that (3.382)
provides a consistent expansion of the equation of motion to second order in
the coupling. Here and in the following, operators with a time argument such
as 1-11(t) denote interaction picture operators with respect to the unperturbed
Hamiltonian Ho = Hs ± HB. We shall assume factorizing initial conditions and
that the bath is in a thermal equilibrium state pB at temperature T =11kBO.

exp(—,3HB )
(3.384)
PB — trB exp( — OHB) .

Again, averages with respect to pB are denoted by angular brackets.

For the discussion of quantum Brownian motion it is convenient to introduce
the following correlation functions,'

D(T) E i([B, B(-7 ) ]) = i[B, B(-7 ) ], (3.385)

D i (T) E ({B , B (-- TM . (3.386)

For reasons which will become clear below, the functions D(r) and D1 (T) are
often referred to as the dissipation and noise kernel, respectively. We note that
the average over the bath may be omitted in the expression for the dissipation
kernel D(r), since the commutator [B, B(--T)] is a c-number. Making use of the
spectral density defined by

= v-■ 2 6 (c.,) —
Kn wri)
(3.387)
J(w)
n
n wn Z—i2rn

we can express the bath correlation functions by

8 We use a notation which is analogous to the usual notation for the corresponding correlation
functions in QED, see Chapter 12.
 QUANTUM BROWNIAN MOTION 175

00

D(T) = 2h, f dw (co) sin WT, (3.388)

D i (r) = 2h, f dw,/(w) coth(hwl2kBT) cos WT. (3.389)

After a little rearrangement the super-operator 1C can now be written as

00

1
1Cps(t) =h--
12 f dT GD(T)[x, {x( — r), Ps(t)}1 — D1(T)[x, [x( — T)) Ps (00 •

(3.390)
The properties of the generator 1C strongly depend on the behaviour of the
dissipation and the noise kernel which, in turn, is determined by the spectral
density J(w). In order to obtain true irreversible dynamics one introduces a
continuous distribution of bath modes and replaces the spectral density by a
smooth function of the frequency w of the bath modes. In phenomenological
modelling one often introduces a frequency-independent damping constant 7
and takes the spectral density to be proportional to the frequency for small w,
2m7
JP) = W for w 0. (3.391)
'
It will be seen below that this form for the spectral density, which is known as
the Ohmic spectral density, gives rise to frequency-independent damping with
the rate -y. On the other hand, the high-frequency modes of the environment
lead to a renormalization of the physical parameters in the particle potential.
To account for this renormalization of the particle Hamiltonian one introduces
a high-frequency cutoff 1 2 into the spectral density. To give an example we take
an Ohmic spectral density with a Lorentz-Drude cutoff function,
2m-y 2
J(w) = (3.392)
71 W Q2 ±
For this type of spectral density the bath correlations can be determined ana-
lytically as
D(T) = 2m-yh1 2 e -Q ITIsign T (3.393)
+" Qe— Ç2 ITI
Di(T) = 4m7kBTS-2 2 E 112 v72i
(3.394)
n=—co

Here, the noise kernel D I (T) is calculated by using the formula

2kB T ‘2
0

H -.

coth(hw 12kBT) = w2 (3.395)

n=—oc

where the vn = 271nkBT/h, are known as the Matsubara frequencies. The remain-
ing integrals are then determined with the help of the method of residues.
176 QUANTUM MASTER EQUATIONS

Table 3.1 Separation of the time scales in the quan-

tum optical limit and in quantum Brownian motion.
quantum optical limit quantum Brownian motion
TB <<
TR TB <TR
Ts <TR TB < TS

We observe that the correlation functions involve the correlation times S-2 -1
and 1 yid -1 for n 0. The largest correlation time is therefore equal to T B =
Max{12 -1 , h/27kBT} and the condition for the applicability of the Born-Markov
approximation becomes
hîy < Min{hn, 27kBT}. (3.396)
This condition corresponds to the condition TB < TR which has already been
used in the derivation of the weak-coupling and the quantum optical master
equations in Sections 3.3.1 and 3.4. However, in the quantum optical case one
proceeds by performing the rotating wave approximation which requires that
the systematic evolution of the reduced system is fast compared to the typical
relaxation time, that is Ts - P i- col -1 < TR. By contrast, here we investigate the
case that the systematic system's evolution is slow in comparison to the bath
correlation time. If we denote by wo = TV a typical frequency of the system
evolution the latter condition becomes
hwo < Min{ hf2, 2R- kB T}. (3.397)
The fundamental difference between the quantum optical and the quantum Brow-
nian motion case is summarized in Table 3.1.
In order to simplify the expression for the generator IC we exploit condition
(3.397) and approximate x(-y) by the free dynamics,
i

C illsr/h Xe illsT/4 `,::.,

' X — — [Hs 47- = x - T. — (3.398)
h ' Pm
Substituting into eqn (3.390) the generator is seen to consist of four terms:
00 oc
D(T)[x, fx, Ps}] d T TD(T)fX, {p, ps)-]
2h2 f d T 2h,2m f fICPs

1 1
- f c/TD I (T)[x, [x, ps]] + f c/TTD i (7) [x,[p, ps]].(3.399)
2h2 2h,m
2

The first term on the right-hand side of this equation can be determined with
the help of the relation
00
1
f dr sin COT = P. (3.400)
co
13
 QUANTUM BROWNIAN MOTION 177

Hence, we have
00 DO
2
f dTD(r) = 2h f dw j(w)
co
= 2h E Kn
27nnWn
2.
(3.401)
n
o o

Since, in addition, [x, {x, Ps}] = [x2 , Ps] the first term on the right-hand side of
eqn (3.399) can be cast into the form

[s2, psi = i (3.402)

J fi [Hc, Ps].
h Ln-d 2m nw2
n

Evidently, this term compensates the contribution from the counter-term to the
Hamiltonian part of eqn (3.382). Thus we indeed see, as claimed earlier, that the
interaction with the bath induces a renormalization of the free-particle Hamil-
tonian Hs which is exactly cancelled by the counter-term H.
In order to determine the second term on the right-hand side of (3.399) we
use
0. 00

f dry sin LOT =——a f dr COS WT = (3.403)

aw
o o

which yields
00

(3.404)
fo dTTD(r) = hli- Jt (0) = 2m7h,

and allows the second term to be written as

i-y
— -v,[x, {P, Ps}]. (3.405)

Correspondingly, we find
00 DO

fo ch-D i (r) = 2h7i- f dwJ(w) coth(hw/2kBT)6(w)

o
= h71- lim J(w)coth(tw.)/2kBT) = 4m-ykBT. (3.406)
w---yo

The third term of (3.399) may thus be cast into the form
2m-ykBT
[x
h2 /
[x ' ps]]. (3.407)

The fourth term of (3.399), finally, depends on the frequency cutoff SI and will
be determined by use of the explicit expression (3.394) for the Ohmic spectral
178 QUANTUM MASTER EQUATIONS

density with Lorentz-Drude cutoff. In the limit of high temperatures such that
kBT > h,S-/ we get
co

f
o
dTTD i (7- )
4Trryç2kBT .
(3.408)

Hence, the fourth term reads

2-yk B T
h2ls [x'[P' Ps]]. (3.409)
To estimate the importance of this contribution we compare it with the third
term (3.407). Since the momentum is of the order of p = rri, - mw o x, we see
that the term (3.409) differs from the term (3.407) by a factor of co0 /S-2, which,
by assumption, is very small. Thus we may neglect the fourth term in the master
equation.
Collecting our results we thus arrive finally at the Caldeira-Leggett master
equation (Caldeira and Leggett, 1983)
d i i-y 2m-ylcB T
Ps(t) = -- [1115)Ps(t)] — —h {x, fp, ps (0 1] — h2 [x, [x, ps(t)]]. (3.410)
dt h
The first term on the right-hand side of the master equation describes the free
coherent dynamics of the system. The second term, which is proportional to
the relaxation rate 7, is a dissipative term which stems from the contribution
involving the dissipation kernel D ( r). The last term, which is proportional to
the temperature, describes thermal fluctuations and will turn out, as we will see
later on, to be of fundamental importance in the theoretical description of the
phenomenon of decoherence.
One might ask the question of whether the generator of the Brownian motion
master equation (3.410) can be brought into Lindblad form. The answer to this
question is negative. However, it can be written in Lindblad form by just adding
a term which is small in the high-temperature limit. To see this we observe that
(3.410) can be written in the form
d i iry
Ps = -- {Hs, Ps] — — (3.411)
dt
—

h 2h [xP + Px, Ps] +D(Ps).

The commutator [xp+ px, PS] gives a Hamiltonian contribution to the generator.
while V(p) is of the form of the dissipator in eqn (3.63) with F1 = x F2 = P ,

and the coefficient matrix

(az3) 7BT/h 2 -i-y/h
_ (4rrryk
i/h o ).
(3.412)

Thus, the generator of the Caldeira-Leggett master equation is of the first stan-
dard form (3.63) under the condition that the matrix aii is positive. This condi-
tion is obviously violated for det a = - (71 h) 2 < 0 . However, we could introduce
 QUANTUM BROWNIAN MOTION 179

a non-vanishing coefficient a22 which corresponds to an additional term of the

form -a22h -p, psn12 in the generator of the master equation. With such a term
the positivity condition reads

4m7kBT
det a = h2 a22 - (7/h) 2 > 0 . (3.413)

Let us employ a minimally invasive modification and require that det a = 0,

which yields a 22 = 7 I 47-nkBT . This amount to adding the term

7 [7) 4 , PS (3.414)
8mkBT
to the right-hand side of the master equation which makes the generator positive
and allows the master equation to be written in Lindblad form (3.66) with a single
relaxation rate -y and a single Lindblad operator

zlinkBT.\11
A= V
h2 X ±2
4m kBT P.
(3.415)

The term (3.414) is small compared to the term with the double x-commutator
provided h I kBT < Lk) 0-1 . This shows that the term added above is in fact small
under the conditions which were used to derive the master equation.

3.6.2.2 Approximate stationary solution In the position representation the

master equation (3.410) takes the form

a , [ih ( 02
— PsV x't) = 3l2) (V (x) - V (x')) (3.416)
at ' 2rn ,ax2 h
( a 9\ 2m7kBT , 2
h2 (x - x ) 1 ps(X, Xi t).
x') — , ) —

Introducing new variables r and q defined through

x = r + hq, x' = r - hq, (3.417)

we obtain an equation for the function f (r, q, t) = ps(x, x' ,t),

a [ i a2 i a 8-ymkBTq2j
— f =
at araq2m, h
(V (r + h,q) - V (r - hq)) - 27q — -
aq
f.
(3.418)

On approximating

V(r + hq) - V (r - hq) ',:.- 217' (r)h,q (3.419)

the stationary solution can be written

180 QUANTUM MASTER EQUATIONS

V(r)
f (r, q) = N exp [- kBT 2mkBTq21, (3.420)

such that an approximate stationary solution of the master equation (3.410) is

found to be

HV((x + x')12) mkBT(x — x') 2 1

P s (x , x' ) = N exp (3.421)
kBT 2h2 j ,

with the normalization factor

+Do
N -1 = f dxexp [ V(x)1 (3.422)
kBT ] '
—00

We see that the diagonal of the position space density matrix represents an
equilibrium distribution proportional to exp[—V(x)/kBT] in accordance with
the result of statistical mechanics. The off-diagonal elements decay exponentially
with the distance Ix — x'I from the diagonal, where the relevant length scale is
given by the thermal wavelength Ath = hIV2mkBT. Expanding the left-hand side
of (3.419) to second order we thus see that (3.419) provides a good approximation
as long as the thermal wavelength is small compared to V241V/1/1Vm1 1.
We also note that (3.419) becomes an equality for a quadratic potential
V(x) = -12-mw,3x 2 , in which case the stationary density is represented by a Gaus-
sian function

0.2
1 [1 (x + x1)2
Ps(x, x') = exp (3.423)
-V27i-o- 2o-2x 2
]

with corresponding expressions for the variances of position and momentum

coordinates,

kBT
ax2 = (x 2 ) = 2 /
TP2 _ (7)2 ) _ rnkB T. (3.424)
"VO

The product of position and momentum uncertainty is therefore

kBT h
ax • ap = >—. (3.425)
wo 2

Thus, in the present high-temperature approximation the product of the uncer-

tainties is large compared to the minimal possible value allowed by the uncer-
tainty relation.

3.6.2.3 Equations of motion for mean values and variances With the help of
the master equation (3.410) one easily obtains the following equations for the
 QUANTUM BROWNIAN MOTION 181

first and second moments of the coordinate and the momentum of the Brownian
particle,

d
(3.426)
lt (x) =
d
—(171 (x)) — (3.427)
(19) =
d2 1
= — (px + xp) , (3.428)
dt m
d
2 (p2 ) — 2(xV ' (x)) — 2-y(px + xp) ,
271 (3.429)
Tt (13x + xP) =
d2
1 (x) + V' (x)p) — 47(P2 ) + 4m-yk B T. (3.430)
= —(PV

These are the Ehrenfest equations for the first and second moments of coordinate
and momentum of a damped particle. They involve the friction force —2-yp. The
corresponding classical equations of motion are given by the stochastic differen-
tial equations

1
m p(t)dt,
dx(t) = — (3.431)

dP(t) = —171 (x(t))dt — 27p(t)dt + V 4m7kBT dW (t) . (3.432)

Let us solve the equations of motion (3.426)-(3.430) for a free Brownian

particle, that is for V = O. The solutions for the first moments yield

(x(t)) = (x(0)) + 1 (1— e -214 ) (p(0)), (3.433)

2m-y
(p(t)) = e -214 (p(0)). (3.434)

Thus, the initial momentum relaxes exponentially to zero over a time scale 1/2-y,
while the average position is displaced asymptotically by the value ((0))/2-y.
Defining

a! (t) = (x2 (t)) — (x (t)) 2 , (3.435)

4(0 = (P2 (0) — (3.436)
apx (t) = ({p(t) , x (OD — (3.437)

we can represent the solutions for the second moments as follows,

182 QUANTUM MASTER EQUATIONS

) g 2 (0)
( 1 _ e -2-yt 2 1 _ e -2-yt
2 2 Pk
ax (t) = ax (0) + m2 +
27 2m7

+
kB T_ [7t — ( 1 _ e -2-yt) + _
m72
41 (1 _ e -4-yt1 I
j / (3.438)
ap2 (t) = e-4'Y t up2 (0) + mkBT (1— e _ 4-t) , (3.439)
1
apx (t) = e -21't u2px(0) + 1'o- ( un)
(1 — e -214 ) e —2t
m7
kB T
+— (1 — e -21') 2 . (3.440)
7
In the long time limit (7t » 1) we thus get the asymptotic expressions
2 kBT kBT
Grx —+ t, a 2p -+›. mkBT, apx —f . (3.441)
my 7
Obviously, the process does not become stationary in the long time limit. The
momentum uncertainty approaches a value which is equal to that given by the
thermal equilibrium value (see eqn (3.424)). The position uncertainty, however.
increases with the square root of time, exactly as in classical Brownian motion.
For a pure Gaussian (minimal uncertainty) initial state we have o-p2 (0) =
h2 /4o(0) and o-px (0) = 0, which leads to
ax2 (0) + h2 (1 – e -2-Yt ) 2
47/2 (q (0) 2,y 2

+
kB T r
mpy 2 rt — ( 1 — e-2-Y ' 41
t, + _ , i _ e _4.-1 1
h2 t2 4kB T
+ 3.442)
g!( 0) +
3m7 2 (-rt) 3
(

4m2 o- (0) ,

where the second relation holds for 7t < 1. The second term of this second
form describes the spreading of the Gaussian wave packet according to the free
Schrödinger equation. For short times the influence of the environment is seen
to yield an additional contribution to the spreading which is proportional to the
third power of t.
3.6.3 The exact Heisenberg equations of motion
For a number of applications it is useful not to resort to an approximate master
equation for the reduced density matrix, but to work instead directly with the
Heisenberg equations of motion of the system. Since in the Caldeira—Leggett
model the bath is a collection of harmonic oscillators and since the coupling
to the reduced system in linear, one can eliminate the dynamics of the bath
variables completely from the Heisenberg equations of motion for the variables
pertaining to the reduced system. This will be done in the present subsection.
The usage of the resulting Heisenberg equations of motion will be illustrated
with the help of several examples for free Brownian motion and by means of the
fluctuation—dissipation theorem.
 QUANTUM BROWNIAN MOTION 183

3.6.3.1 Derivation of the Heisenberg equations The complete Caldeira—Leggett

model leads to the following exact Heisenberg equations of motion for the Brow-
nian particle and the environmental oscillators,

X(t) = [H, x(t)] = lp, (3.443)

h, 771
= x(t)1 = (3.444)
mn
ii(t) = [H,p(t)} = —VAx(t)) + E Knxn (t), (3.445)

ii(t) = [FT (t)] = —in,4,x(t)+Knx(t). (3.446)

The corresponding equation for the coordinate of the Brownian particles is thus

m(t) + V(x(t)) — E Knxn (t) = 0, (3.447)

rt
while the equations for the coordinates of the bath oscillators take the form

Tertnn (t) + tn ncon2 x, (t) — Kn x(t) = 0. (3.448)

The last equation shows that the n-th bath oscillator is driven by the force
ix(t) which depends linearly on the coordinate of the Brownian particle. In
order to get a closed equation of motion for x(t) one solves eqn (3.448) in terms
of x(t) and of the initial conditions for the bath modes and substitutes the result
into eqn (3.447). To this end, it is convenient to express the coordinates of the
bath oscillators in terms of the creation and annihilation operators bt,, bn as

nintkon
n(0) = 2innwn (bn pn(0) = (bn — b!,). (3.449)
2

The solution of eqn (3.448) is then given by

x n (t)= \I 2Tnnc, jn ( e —iwnt br, e icon t,+ + Im

flj
rnnuln f ds sin[wn (t —
(3.450)

Substituting into eqn (3.447) yields

2
K
771(t) Vc1 (X(t)) — n f dS sin[in (t — s)]x(s) = B(t), (3.451)
nW n
0

where we recall that the operator B(t) which appears here on the right-hand
side is the interaction picture operator
184 QUANTUM MASTER EQUATIONS

Kn h
B(t) = E (e i bn +eiwn t bI) (3.452)

corresponding to the Schrödinger operator B =

Kri xr, (o). With the help of
the dissipation kernel we can write the equation of motion for the coordinate of
the Brownian particle as

1 1 1
(t)+ —V,1(x(t))— f dsD(t — s)x(s) = —B(t). (3.453)
o
In the theory of quantum Brownian motion it is useful to express the dis-
sipation kernel in terms of another quantity which is known as the damping
kernel
CO

ry(t — s) = 271
2 dw j (w)
ul cos[w(t — s)], (3.454)
0

which satisfies
d 1
= -- D(t — s), (3.455)
hm
and

-y(0) = (3.456)
m w
o
With the help of this damping kernel we can write the dissipative term of
eqn (3.453) as follows

1 d
hm
f dsD(t — s)x(s) = f ds— -y(t — s)x(s)
dt
(3.457)
jo
d
= — f ds-y(t — s)x(s) — 7(0)x(t).
dt

In view of eqn (3.456) the last term 7(0)x(t) is seen to cancel the contribution
—

from the counter-term contained in the potential Ve (x). Thus we finally arrive
at the following exact Heisenberg equation of motion,

1 d 1
(t)+ —
in V i (x(t)) + — — B(t). (3.458)
dt f ds -Y(t — s)x(s) = 771
0

Equation (3.458) is the desired equation of motion for the coordinate of

the Brownian particle. It may be viewed as the quantum analogue of a clas-
sical stochastic differential equation, involving a damping kernel -y(t — s) and a
 QUANTUM BROWNIAN MOTION 185

stochastic force B(t) whose statistical properties depend on the initial distribu-
tion at t = O. In the case of an environment with an Ohmic spectral density with
an infinite cutoff, 12 oo, we get the damping kernel (see eqns (3.392) and
(3.454))

7(0 = 4'y(t) (3.459)

such that the Heisenberg equations of motion take the form

±(t) = --p(t), (3.460)

= -Triv/(x(0) - 2m-y&(t) + B(t). (3.461)

As in the classical stochastic differential equations (3.431) and (3.432) the friction
force is equal to —2m,04.
The statistics of the forcing term B(t) in eqn (3.458) is described by the
quantum correlation function ({B(t), B(ti)}). We emphasize that the angular
brackets denote here the average over the initial distribution p(0) of the total
system since we are working in the Heisenberg picture. If we use an uncorrelated
initial state of the form p(0) = ps (0) pB this correlation function is equal to the
noise kernel Di (t — t') introduced in eqn (3.386). For an Ohmic spectral density
it is given by

4m7h,
D i (t — t') = ({B(t), B(e)f) = 7r Id ww coth(hw I 2kBT) cos[w(t — t')].

(3.462)

In the high-temperature limit one may assume that 2kBT >> ruo for all relevant
frequencies u.) which gives
00

1 , 2911h2kBT f
({B(t),B(e)}) Pze cos[w(t — t`)] = 4nrykBT (t — t').
—
2
71"
o
(3.463)

This shows that h drops out of the correlation function and that we obtain
precisely the correlation function of the noise term in the classical stochastic
differential equations (3.431) and (3.432).

3.6.3.2 Quadratic potentials For a quadratic potential

1
V (x) = -inw2 x2 (3.464)
2 °

the Heisenberg equation (3.458) becomes

186 QUANTUM MASTER EQUATIONS

t
d 1
(t) + 4x(t) + — f ds-y(t — s)x(s) = --B(t). (3.465)
dt
o
To solve this equation we introduce the fundamental solutions G i (t) and G2 (t) of
the homogeneous part of eqn (3.465), which is obtained by setting the right-hand
side equal to zero. These solutions are defined through the initial conditions
G 1 (0) = 1, 6 1 (0) = 0, (3.466)
G2(0) = 0, 62(0) = 1. (3.467)
Introducing the Laplace transformation
00
f(z) = f dte — z t f (t), (3.468)
o
one can write the Laplace transforms of the fundamental solutions as follows,
z
61(z) =
z2 + wci + z;y(z) '
1
62(z) = (3.469)
2 +4 + z'-f(z)1Z
where ;y(z) is the Laplace transform of the damping kernel.
In terms of the fundamental solutions one can write the general solution of
the Heisenberg equation (3.465) as
t

X(t) = Gi (t)X(0) -I- G2 (t)± (0) + 71

7:1, f dSG2 (t - s)B (s). (3.470)
0

With the help of this solution all desired mean values, variances and correlation
functions of the Brownian particle may be expressed in terms of averages over
the initial distribution p(0) of the total system.

3.6.3.3 Free Brownian motion As an example we study free Brownian mo-

tion, that is V = 0, for an Ohmic spectral density. For an infinite cutoff the
fundamental solutions then take the form
1
G1 (t) = 1, G2 (t) = — (1 — e -2 ''') , (3.471)

which yields
t
1 _ e -2-yt 1 _ e_ 2 (t - 8 )
x(t) = x (0) + &(0) + f ds B (s), (3.472)
2 -y 2m ry
o
t
p(t) = e -21(tp(0) + f dse -214t-5) B(s). (3.473)
 QUANTUM BROWNIAN MOTION 187

Let us consider a factorizing initial state p(0) = ps(0) pB. The initial coordi-
nates of the Brownian particle and of the bath oscillators are then uncorrelated
and averages of the form (x(0)B(s)) and (p(0)B(s)) vanish. With the help of
(3.473) the mean kinetic energy of the Brownian particle, for example, is found
to be
1
E(t) E _ri(p2(t)) (3.474)
t t
1 (s s /)
1 /7.12(0))e-4-yt f ds f ds'e-2-y(2t –8-8 1 )
2m, 4m,
o o
Q
1 2
Ie t _ e -2-yt 1 2
= fli (.7)- (0))e-4t + f dww coth h42kBT) '
(
it W2 + (2-y)2 ,
o

where we have used expression (3.462) for the noise kernel. If we perform here
the limit -yt oo we find that the mean kinetic energy approaches the value

27 f (N(w) +
E(CO) = — hw = Eth Evac • (3.475)
it + (21) 2
o

This equation provides a spectral representation of the mean kinetic energy of

the Brownian particle: Each frequency component is given by the mean energy of
the environmental oscillators at that frequency times a Lorentzian-type spectral
density.
The thermal contribution Eth to the kinetic energy is given by an integral
which converges in the limit of an infinite cutoff,
00

I ruo/VP)
Eth = — dcu , (3.476)
it + (2 y) 2 '
o

In the high-temperature limit kB T >> hry this expression leads to the well-known
result of classical statistical mechanics, Eth r-Z-J -kBT. On the other hand, the
vacuum contribution Eva, is seen to diverge logarithmically with the cutoff,

2
rtWi 2 'Yh7r ln
Evac = f clw 2 + (2,y)2 = T [1+ (y.j . (3.477)
7r
o

The interaction of the Brownian particle with the vacuum fluctuations of the
high-frequency modes of the environment thus yields a logarithmically divergent
contribution to its mean kinetic energy.
188 QUANTUM MASTER EQUATIONS

As a further example we determine the mean square displacement

d2 (t , t') = ([x(t) — x (t` )] 2 ) (3.478)

of the particle which is an important quantity in the theory of Brownian motion.
On using eqns (3.472) and (3.462) one obtains

d2 (t, ti ) = 1 (p -2-yt (p2(0))

(3.479)
4m272
2
2-yh, f wcoth(hw/2kBT) e iwt _ e iwt' e -2ryt e -2-ye
mir J 14C4j W2 + (27) 2 hl)
o
The second term within the squared modulus leads to an integral which diverges
in the limit of an infinite cutoff. In fact, the cutoff-dependent contribution to the
mean squared displacement is found to be
../2 j.i\i 1 (e-2-yt 2 2Evac
Ct »singular e (3.480)
—

4-y2
This shows that the ultraviolet divergent vacuum part of the mean quadratic
velocity 2Evac /m leads to a corresponding divergent contribution of the mean
square displacement: The Brownian particle can absorb an arbitrary amount of
energy from the high-frequency environmental modes and can travel an arbitrary
distance within a finite time interval.
This singular behaviour, known as initial jolts, is clearly a result of the ar-
tificial assumption of an uncorrelated initial state. Namely, if we let the times t
and t' go to infinity, keeping fixed their difference T t — t', all transient terms
vanish and the mean square displacement becomes a function of T which is given
by the ultraviolet convergent integral
co
f _ c,4. coth(Tuo/2k
2-yh, ) (2 )2 BT) 4 sin2 P7/2)
d2 (r) = dw (3.481)
—

M7T CO2 + w2
0
For any finite temperature we can replace the second factor of the integrand by
27 1T1 6 (w) which leads to the asymptotic expression
kBT I
d2 T ) ( (3.482)
rirt

This formula describes the well-known regime of classical diffusion which is valid
for rYiTi » 1 and TI » h/kBT. The same result is found for a classical Brownian
particle. On the other hand, for zero temperature the term w coth(hw/kB T) in
the integrand must be replaced by w and one obtains the asymptotic expression
00 9

d2 (T) = f C/W chi Sin-

(w 2 (WT/2)
(2, 02) h ln(7171), (3.483)
77171" -
rn7 ry
0 \
 QUANTUM BROWNIAN MOTION 189

which is valid for *1 » 1. We observe a strong subdiffusive behaviour of the

particle in the quantum regime. The mean square displacement increases only
weakly with the logarithm of time, the relevant length scale for the quantum
diffusion being given by -/h/miry.

3.6.3.4 Response function, equilibrium correlation function and the fluctuation-

dissipation theorem In order to determine the equilibrium fluctuations of the
total system it is useful to invoke the fluctuation-dissipation theorem (FDT).
Together with the exact solutions of the Heisenberg equations of motion the FDT
allows us to determine all equilibrium correlations of the Brownian particle. To
formulate the FDT we introduce the symmetrized autocorrelation function of
some Heisenberg picture operator z(t) of the total system,
1 1
S(r) = -2- ({z(t + 7), z(t)}) = (3.484)

For simplicity we assume that (z(0) = 0 and zt(t) = z(t). We emphasize that
the angular brackets denote here the average over the equilibrium distribution
of the total system given by
exp(- )3H)
p= (3.485)
tr exp(-0H) '
where we recall that H is the Hamiltonian of the combined system S + B. Due
to the homogeneity in time the autocorrelation only depends on T and satisfies
S*(r) = 8(7) = S( - 7). The spectrum of the equilibrium fluctuations is given
by the Fourier transform of the autocorrelation function,
+co

S' (w) (3.486)

= f
-00
dr co s c.J T S(T ).

We note that the spectrum has the property S'*(w) = :9- (co) =
The response function pertaining to the variable z(t) is defined as

x(r) = -hi 0(7)([z(7), z(0)]). (3.487)

It describes the linear response of the system to an external force F(t) which
is applied for times t > 0 and is represented by a time-dependent perturbation
of the form V(t) = -zF(t) in the Hamiltonian of the total system. First-order
perturbation theory then yields the linear response of the system

(6z(t)) = c f dsx(t - s)F(s), (3A88)

o
where (5 z (t) E Zi/ (t) - z (t) and z(t) denotes the Heisenberg operator correspond-
ing to the perturbed Hamiltonian H + V(t). The average first-order change of
190 QUANTUM MASTER EQUATIONS

the variable z(t) is thus given by a linear integral transform of the applied force
where the integral kernel is provided by the response function. We note that the
linear response at time t only depends on the values of the force at times prior to
t, showing the causal character of the system's response. The Fourier transform
of the response function may be written as
+00 +00

= f ‘-' XM = f
dT,iwr f \
( )
dreiwT X\T)
- E ,( 1 (w) ± iii(w) 1 (3.489)

where we have decomposed i- (w) into real and imaginary parts.

The fluctuation-dissipation theorem provides a relation between the linear
response of the system to an external force and the fluctuations in equilibrium.
In frequency language this relation is given by (see, e.g. Landau and Lifshitz.
1958)

(w) = h coth(tv„,)/2kBT)i ll (w), (3.490)

which expresses the spectrum of the equilibrium fluctuations in terms of the
imaginary part of the Fourier transform of the response function.
The fluctuation-dissipation theorem holds for a general system in thermal
equilibrium and for any Heisenberg picture observable z(t). Let us now apply the
theorem to the coordinate x(t) of a Brownian particle in a harmonic potential.
Since the commutator [x(r), x(0)] is a c-number we can immediately determine
the response function with the help of the exact solution (3.470) of the Heisenberg
equation of motion (3.465),

X(T) = i1•, 9(7)[x(7), x(0)] = 0 (r)G2( 7)P( 0), x(0)] = OW wli G2(T)• (3.491)

This relation can also be inferred directly from (3.470). Namely, adding the
perturbation -xF(t) to the total Hamiltonian of the system amounts to replacing
B(t) by B(t)+F(t) in the equation of motion (3.465). This replacement yields the
additional term ,g dsG2(t - s)F(s)Im on the right-hand side of (3.470), which
shows that the response function is in fact related to the fundamental solution
G2(7) by eqn (3.491).
Expressing the Fourier transform of the response function in terms of its
Laplace transform we therefore obtain
+00
1 1 - . 1 1
T 1, fG2(r)
5C(w) = 7 drei" = wi G2(-no) = 2 . (3.492)
7/2 Wo — CO 2 — iw;y( - ico)
o
The imaginary part of the Fourier transform of the response function is thus

1 wRVY(+iw)]
(w) = (3.493)
m (cuô - w 2 + wh'( - ic4))1) 2 + w2 (RM-iw)i) 2 '
 QUANTUM BROWNIAN MOTION 191

and the FDT (3.490) immediately leads to the following expression for the au-
tocorrelation function of the coordinate of the Brownian particle,
+00

S(T) = f —C1W hCOth(r/LJ/2kBT)V 1 (L) cos )Y. (3.494)

27r
- 00

With the help of eqns (3.494) and (3.493) one can determine the exact equilib-
rium correlations in terms of the Laplace transform ri, (z) of the damping kernel.
For example, we have by definition of the autocorrelation function

(x 2 ) = (3.495)
(px + xp) = = 0, (3.496)
2) _ m2 (0) ,
9
(3.497)
(

and therefore
+00

s 2 ) = f dW
27r
hCOth(hW/ 2 kBT) II (W), (3.498)

+cc dw
f
(p2 ) = m2 h cot (hoi/2kBT)"(w). (3.499)
- OC

As an example, let us consider the Ohmic spectral density (3.392) with a

Lorentz—Drude cutoff. According to definition (3.454) this leads to the damping
kernel

7( 7 ) = 2-yfte — '1171 , (3.500)

with the Laplace transform

2-yf/
''y(z) = (3.501)

Thus, we have

14-54E — ico)] = 2-y 92 w 2 2 ,y, (3.502)

w
:stryjE — iw)] = 27 1-2 1-12 w2 —* 0, (3.503)

where we have indicated the limits of an infinite cutoff, SI —* co. The Laplace
transform of the damping kernel becomes real for an infinite cutoff and the
spectrum of the fluctuations takes the form
192 QUANTUM MASTER EQUATIONS

2h-y
(w) = m (4) _ (4 2)2 + (2 7(4)2 coth(h42kBT). (3.504)

For frequencies such that kBT » twi the Planck constant drops out and the
spectrum becomes
47kB T 1
= m _ (.4) 2)2 + (27(.0 )2 ' (3.505)

which exactly corresponds to the classical fluctuation spectrum of a damped

harmonic oscillator in thermal equilibrium.
3.6.4 The influence functional
In the case of the Caldeira—Leggett model it is possible to eliminate the environ-
mental variables exactly. This is due to the fact that the environment is supposed
to be in a thermal state whose characteristic function is Gaussian, and because
the reduced system couples linearly to the bath operators B(t) whose commuta-
tor represents a c-number. In the present subsection we are going to construct an
exact super-operator representation for the reduced density matrix and derive
from it the corresponding path integral representation. The path integral involves
a certain functional, known as the Feynman—Vernon influence functional, which
represents exactly the influence of the heat bath on the reduced system.
3.6.4.1 The influence super operator To obtain an exact representation for the
-

reduced density matrix, let us assume factorizing initial conditions and that the
initial environmental state is a Gaussian (thermal) state. We emphasize that the
assumption of an uncorrelated initial state is made here only for simplicity; it is
possible to carry out the derivation also for correlated initial states (for a review.
see Grabert, Schramm and Ingo ld, 1988).
An appropriate starting point of the derivation is the following exact repre-
sentation of the reduced density matrix ps(t f ) at time tf in terms of the total
density p(ti ) at some initial time ti ,

Ps(tf) = trs {U(tf, ti)p(ti)Ut(tf,ti)} . (3.506)

Again, we are working in the interaction picture such that U(t f , t i ) is the inter-
action picture time-evolution operator of the total system. To eliminate the bath
variables from this expression one can employ the following method of deriva-
tion. The interaction picture time-evolution operator involves a time-ordering of
the reduced system as well as of the bath variables. First, one eliminates the
time-ordering of the bath variables, which is easily done, for the commutator of
the bath variables is a c-number. One is then left with a simple average taken
over the state of the bath. In a second step this average is determined with the
help of a cumulant expansion, which terminates at second order since the bath
state is Gaussian. This procedure yields the final result

ps(t f) = T exp (—hi41.[X c , X a ]) ps(t i ). (3.507)

 QUANTUM BROWNIAN MOTION 193

The details of the calculations leading to this result will be given in Section 12.2
(see, in particular, eqn (12.99)) in a more general context and will thus be omitted
here. In eqn (3.507) the symbol T, indicates, as usual, the chronological time
ordering in the interaction picture. The phase of the time-ordered exponential
represents a super-operator which is given by
tf
sal
f dt,C,(t) (3.508)

tf t
1
+ f f {- 0 D(t — t f )x,(t)x 0 (e)D1(t — t i )x,(t)x,(t 1 )} .
2h2
—
t, t,
The first term on the right-hand side stems from the counter-term H, in the
Hamiltonian of the total system and is defined in terms of the commutator
super-operator
i M
,C e (t)ps [1-1,(t), ps] = -- - -y(0)[x 2 (t), ps ] .
E (3.509)
h, h 2
The second term on the right-hand side of (3.508) involves the commutator and
the anticommutator super-operators
x,(t)ps [x(t),ps], x a (t)ps _= {x(t),ps}, (3.510)
where all operators carrying a time argument are interaction picture operators
with respect to the free Hamiltonian 1/0 = Hs + HB•
The phase 4, [xe , Sa] subsumes completely the influence of the environment
on the reduced system's dynamics and may thus be called the influence phase
functional. Formally, clqxe , xa ] is both a super-operator acting on the initial state
Ps (tt) and a bilinear functional which involves the dissipation kernel D(t—e) and
the noise kernel D i (t — e). We note that the second-order generator (3.390) may
be obtained from eqn (3.507) by expanding the exponential to first order in the
influence phase and by taking into account that the time integrals in expression
(3.508) for the influence phase are already time-ordered.
3.6.4.2 Path integral representation Equation (3.507) gives rise to an equiva-
lent path integral representation for the reduced density matrix. It thus provides
an operator formulation of the famous Feynman—Vernon influence functional.
To construct the path integral representation we turn back to the Schrödinger
picture and introduce the so-called propagator function J through the relation

ps(x f,x 1f ,tf) = f dx i f dx:: J(x f, X ipt (3.511)

Thus, the propagator function is simply the Green function for the reduced
density matrix in the position representation. In terms of the influence super-
operator in eqn (3.507) it is given through the relation,
194 QUANTUM MASTER EQUATIONS

J(X f,X if,tf;Xj, s , ti ) = (5 ftfl (T, exp Cx,,x,,l) ixiftf),

(3.512)
where 1st) is the eigenstate of the interaction picture operator x(t) belonging to
the eigenvalue x, that is

1st) = exp (+iHs(t — ti)1h)ex). (3.513)

We note that (3.512) leads to the following initial and normalization conditions.

lim J(xf,x'f., tf; xi, ti) = ( 5 flx j )(x`i lx (f )

tf
= 6(x f X i)(5 (X i f (3.514 )

f dX f J(X f X f ,t f; Xi, s , ti) = = (5(x1 — ( 3.515)

We also observe that for vanishing system—environment coupling (4. = 0) the

propagator function reduces to a product of the Green function G5 for the
Schrödinger equation with Hamiltonian Hs,
J (x f , f ,tf;x i , ti) = (x ft fix f t f)
= G s (x f ,t f ; x i ,ti )G*s (x`f ,t f ; (3.516i

The path integral representation of the influence functional is obtained by

writing Cs as a path integral:

GS f t f; Xi, ti) — ( X f I exP (—iHS(tf ti) I h) kJ) = f D x exp So[x])

(3.5171
where
tf

So [x]=j
S0[x] =f — V (x)) (3.518 ,

is the classical action functional of the free Brownian particle. Here, the path
integral is an integral over all paths x(t) subjected to the boundary conditiom..
x(t i ) = xi , and x(t f ) = xf.
Now, if T,O[x(t)] is any time-ordered product, the corresponding matrix
element can be obtained from the path integral as follows

(x f t f IT ,O[x(t)]Ixiti) = f Dx exp ( hi So[x]) O[x(t)]. (3.519

In other words, the path integral performs the time ordering automatically if
the time-ordered operator product is replaced by the corresponding classical
 QUANTUM BROWNIAN MOTION 195

expression under the path integral. In a similar way the matrix elements of the
commutator or the anticommutator super-operators x e , a (t) are given as double
path integrals as follows,

(5f t f IX a ,c(t) (Ix t1 (X il tiOIX(t f)

= ftf IX MIX it i) f) ± ftf IX it i) (t )IXt f t f)

= f Dx f Dx' exp (So [s] — So EST) (x(t) x(e)) • (3.520)

Thus we see that the transition from the super-operator representation to the
path integral representation may be achieved by performing the replacements

sa (t) x(t) + x i (t), x(t) x(t) — x 1 (t). (3.521)

A moment's thought shows that this prescription is also valid in the general
case of an arbitrary time-ordered product of the super-operators x e , a (t). We are
therefore led to the expression

J(x 1,x, t1; Xi, ti) = f Dx f Dx` exp 'u ] ) . (3.522)

This is the double path integral representation for the propagator function which
is well known from the Feynman—Vernon influence functional technique (Feyn-
man and Vernon, 1963). It involves an integration over all paths x(t) and x 1 (t)
satisfying the boundary conditions

x(t i ) = x i , x (t f = xf,
) x 1 (t1 ) = x , x i (tf) = X. (3.523)

The weight factor of the paths is given by the effective action functional

Aix = So [x] — So [x'] + , (3.524)

where the path integral representation of the influence phase functional 4:, [x,
is obtained from (3.508) with the help of the replacements (3.521),
tf
[x, x'] = dt-
171
2 7(0) [x 2 (t) — x 12 (t)] (3.525)
—

tif
f t
1
• f dt f de —
2h
D(t — t 1 )[x(t) — x 1 (t)][x(e) + (e)]
ti tz
f t
• f dt f 2 Di(t — t i )[x(t) — (t)][x(e) — (ti)]•
--

2h
ti ti
196 QUANTUM MASTER EQUATIONS

For the further discussion it is convenient to introduce new variables through

1
q=x—, r= + s'), (3.526)

and to consider the propagator function J as a function of (rf q1, t f ;ri , qi ,t i ).,

The corresponding path integral representation reads

J(r f,qf,t f;ri ,qi ,t i ) = f Dr f Dqexp Alr, ql) , (3.527)

which is a double path integral over all paths r(t), q(t) satisfying the boundary
conditions

r(t i ) = ri , r(t1) = r f , q(ti ) = qi , q(t f) = q f . (3.528)

The initial and normalization conditions for the propagator function read
follows,

lim J(rf,qf,tf;ri ,qi ,ti ) = f — r i )6(q1 — q i ), (3.529

tf
drf J (rf , qf = 0, tf ; ri , qi ,ti ) = (qi ) . (3.530i
f
The effective action functional (3.524) is now given in the new variables by

A[r , = f dt(mq — V (r + q/2) + V(r — q/2) — m-y(0)rq) (3.531

tf t
1
f dt f {—D(t — t i )q(t)r(ti ) + — D1 (t —
2h
ti
3.6.4.3 Classical equation of motion The action functional (3.531) leads to
classical equations of motion which are found by setting the first variation of .4
equal to zero. The variation with respect to q(t) gives the equation of motion for
r(t),

0 d
r(t) + — (V (r + q 1 2) + V (r — +—f (t — )r (t i )
2m Or dt
ti
tf

f depi(t_ (3.532
= 2mh, t,
whereas the variation of r(t) leads to the equation for q(t),
 QUANTUM BROWNIAN MOTION 197

tf
2 d
4w+ — — + q12) + V (r — q12)) — — f cle-y(t 1 — t)q(e) = 0. (3.533)
m Oq dt

Here we have expressed the equations of motion through the damping kernel
7(t — t') (eqn (3.454)) . and the noise kernel D i (t — t'). We note that for q = 0
the dynamical equation for the path r(t) takes on the form of the homogeneous
part of the Heisenberg equation of motion (3.458).

3.6.4.4 Determining the path integral for harmonic potentials To illustrate the
path integral technique we consider a harmonic potential
1
V (x) = (3.534)
2
which leads to the classical equations of motion
tf
d
i:(t) + u,(2) r(t) + — f
dt
,7(t — t i )r(ti ) = ann f deD i (t — e)q(e), (3.535)
ti
and
tf

4m+ cog q(t) — dt f dtiPY(ti t)q(e) = 0. (3.536)

Note that eqn (3.536) represents the backward equation of the homogeneous part
of eqn (3.535), that is, if r(t) solves the homogeneous part of eqn (3.535), then
q(t) E r(t f ti t) is a solution of eqn (3.536).
Since the action functional is quadratic the double path integral (3.527) can
be determined exactly by evaluating the action along the classical solution and
by taking into account the Gaussian fluctuations around the classical paths. Let
r(t), q(t) be a solution of the classical equations of motion (3.535) and (3.536)
satisfying the boundary conditions (3.528). The value of the action functional
(3.531) along this solution can be written as
A[r, q] = m f q f —
tf
d
dt mq(t) (t) + uigr dt f de-y(t — ti)r(ti )
+—
-ft, t,
tf tf

f dt f D i (t t')q(t)q(e)
tf tf

4i4 f dt f deD i (t — t i)q(t)q(t 1 ),

= m(1 f qf — i- i qi ) — — (3.537)
ti t,
where the equation of motion (3.535) has been used in the second step.
198 QUANTUM MASTER EQUATIONS

Due to the inhomogeneous term in eqn (3.535) r(t) couples to q(t) through
the noise kernel D 1 (t — t'). As a result the solution r(t) is, in general, complex.
We decompose r(t) into real and imaginary parts,
r(t) = r 1 (t) + ir(2) (t). (3.538)

The real part r( 1 ) (t) solves the homogeneous part of eqn (3.535), while the imag-
inary part r( 2 ) (t) is a solution of the inhomogeneous equation
tf
(2) (t) + 47- (2) (t) + dtd f -y(t — t 1 )r(2) (t') =
1
2mh,
f D i (t — t 1 )01 ).
ti ii

(3.539)
We now show the following useful property: In order to find the classical action
it suffices to determine the real solution r(1 )(t) and to insert it into the action
functional, which means that we have the relation

A[r , = A [r. (1) , q] (3.540)

if if
•
.( 1) 2 1)qi \) + 4-zh
=m qf - f dt f D i (t t')q(t)q(e).
t,
To prove this statement one first uses the equations of motion for the imaginary
part r (2) (t) and for q(t) to show that

i f dt f dt' D i (t — t')q(t)q(e) = im (7- f( 2) qf

if if

qi ) . (3.541)
2h,
t, t,
To obtain this equation one performs two integrations by part and uses the
condition that r? ) = r f( 2) = 0, since the boundary conditions for the paths are
real. Combining the last equation with (3.537) immediately leads to eqn (3.540
which proves the statement.
The procedure to determine the propagator function can thus be summarized
as follows. One first solves the homogeneous classical equations of motion,

d
i(t) + 4r(t) + — f Py(t t f )r(ti ) = 0, (3.542i
dt
ti
tf
d
4(0 + c4q (t) — — t)q(e) = 0, (3.543
dt

under the boundary conditions (3.528). Substituting these solutions into the
action functional then yields the propagator function
 QUANTUM BROWNIAN MOTION 199

i
J(rf,qf,tf;ri ,qi,ti )=N(tf,ti)exp (— .A[r,q1) (3.544)

=N(tf ,ti )exp (rf qf —r i qi )+F(qf,tf;qi ,ti ))

where
tf tf

1 f dt f dt i D 1 (t—t 1)q(t)q(e).
F(qf ,t f ;qi ,ti )=-- (3.545)
4h2
ti
G2(ol r ent
In eqn (3.544) N(tf, ti) is a time-dependent ± ,:n1
72/2r:(t)tf
imai ,.)l ization factor. Let us set
t, = 0 for simplicity in the following. With the help of the fundamental solutions
G1 (t) and G2(t) (see eqns (3.466) and (3.467)) we can write the solutions of
(3.542) and (3.543) as
i itf)
G(
r(t) =[Gi(t) (3.546)
G2k.tf) k,, rf l
11_72tf
G2 (t f - t)
q(t) = „ [G (t f t) ,G2(t f — t)] q f . (3.547)
L.T2(.tf)
Invoking the normalization condition (3.530) the normalization factor is found
to be
N(t f 1 0) = (3.548)
271-hG2 (tf )
With the above expressions the problem of determining the exact propagator
function is solved completely. Obviously, the propagator function is a Gaussian
function whose exponent is a quadratic form in the coordinates (rf ,qf , ri , qi ),
which is determined by substituting (3.546) and (3.547) into (3.544). We also
note that the normalization factor is directly related to the response function
through N(t f , , 0) = 1/27rhx(tf).

3.6.4.5 Solution of the Caldeira-Leggett master equation To give an example

we consider free Brownian motion (coo = 0) for an Ohmic environment. In the
limit of an infinite cutoff we have the simple expressions G 1 (t) E 1 and G2 (t)=
[1— exp(-2-yt)]/2-y which yields
1 62(tf)
= „(rf —
ri),
(3.549)
G2(t1 ) f G2 (t f )
showing the translational invariance of the porpagator function. Thus we have
J(rf ,qf ,t f ;ri ,qi ,ti) (3.550)
m im
= exp (rf —ri)[d2(tf)qf — q i] r(qf,tf;qi3 O)) .
27rnAL.72(tf) hG2(tf)
This expression for the propagator function can be used to obtain the general
solution of the Caldeira-Leggett master equation (3.410). To this end, we just
200 QUANTUM MASTER EQUATIONS

have to evaluate the function F(qf,tf; qi , 0) in the high-temperature limit which

is provided by the expression

f
tf
2m-ykB T
F(qf,tf;q i3 O) dtq 2 (t) (3.551)
o
Substituting the solution q(t) (eqn (3.547)) and evaluating the time integral
finally yields
J(rf,qf,tf;ri ,qi , 0) (3.552)
2m-y
27h,(1 - e-2-ytf )
2imy h
X exp
1 _ e - 2-y t f (r f - ri
)(e -2 `T t f qf qi)
e -4ytf)
-
2mkBT f -ytf - (1 - e-2 -ytf (1
(qf _ qi )2
h2 1 (1 e -2-ytf )2
f _ 22; ( 1 _ e -2-yto
-
1 _ e-2ytf 2qf(q f - qi ) + Pyt fq . })

This is an explicit expression for the propagator function of the Caldeira-Leggett

master equation for free Brownian motion. With the help of the relations
1
PYti — -2 ( 1 - e -21/tf ) = (7t1) 2 + O((Ytf) 3 ), (3 . 553 )
7tf _ (1 _ e-2-ytf ) + -4-1 (1 - e -4`Yt f ) = (ryt1) 3 + 0((yt f) 4 ), (3.554)
one easily shows that in the limit -ytf 0 the propagator function approaches
im
J(rf ,qf , tf ; ri , qi , 0) —* exp(r f - ri ) (q f - q i )) , (3.555)
f (It f
which is recognized as the Green function corresponding to the von Neumann
equation for a free Schrödinger particle.
Let us finally consider the so-called recoilless limit of the master equation.
This is the limit of a heavy particle which is studied over times which are short
in comparison to the relaxation time. It corresponds to the Caldeira-Leggett
master equation (3.410) without the friction term -i-y[x, p, p}]/h. Thus we set {

A = 2m-ykB T/h2 and perform the limit -ytf —* 0, keeping A fixed. Making use
of eqns (3.553), (3.554) the propagator function is found to be
rn im
J(xf,x'f ,tf; 0) =
27rhtf exP 2h,tf {(xf s 1)2 (xi/ )2} (3.556
A3tf {(xf x ,f)2 ± (xi i) 2 (xf _ x :f )( xi _

where we have again used the original coordinates.

 NON-LINEAR QUANTUM MASTER EQUATIONS 201

3.7 Non-linear quantum master equations

In classical statistical mechanics the dynamics of a many-particle system may
often be approximated in terms of a non-linear kinetic equation for the time-
dependent distribution function Pp', I,t) on one-particle phase space (Reichl,
1998). The most prominent example for such a kinetic equation is the Boltz-
mann equation. For the corresponding approximation of a quantum mechanical
many-body system one has to replace the classical distribution f(13; I , t) by the
one-particle density matrix 0). The latter is then expected to satisfy a non-
linear master equation which provides a mean-field type description of the dy-
namics of the many-body system. Within this framework it is also possible to
consider many-body systems interacting with external reservoirs. The coupling
to an external reservoir may be performed either on the level of the mean-field
approximation or else on the level of the many-particle system. A number of
examples for the resulting non-linear master equations, such as the quantum
Boltzmann equation, the time-dependent Hartree equation, and the non-linear
Schrödinger equation, will be discussed below (Alicki and Messer, 1983).
The general structure of a non-linear quantum master equation may be pos-
tulated to be of the following form,
d
— p(t) = L[P(t)]P(t), P(0 ) = Po. (3.557)
dt

Here, L[u] is a super-operator which is in Lindblad form for each fixed den-
sity matrix u. Thus, eqn (3.557) is a Lindblad-type quantum master equation
whose generator depends parametrically on the density matrix. Under certain
conditions eqn (3.557) has a unique solution

0) = Vt (p0 ), t > 0, (3.558)

where { 14 1t > 01 is a one-parameter family of non-linear maps satisfying the
semigroup property

Vt(Vs(P)) =--- Vt+s(P), Iii =---

Equation (3.557) is therefore said to be in parametric Lindblad form.

3.7.1 Quantum Boltzmann equation

The discrete quantum Boltzmann equation is an example of a non-linear quan-
tum master equation of the form (3.557). In a phenomenological way this equa-
tion can be obtained by exploiting the formal analogy to the classical Boltzmann
equation for the one-particle density f(13, I ,t),

—
a f (g, Y, t) =-- — {h, f(134,Y, t) + C[f](g, I , t) . (3.560)
at

Here, h = 112 I 2m is the one-particle Hamiltonian of a particle with momentum

/I and mass m, the curly brackets denote the Poisson bracket and C[f] is the
202 QUANTUM MASTER EQUATIONS

Boltzmann collision operator, which is a bilinear functional of the one-particle

density f.
To be specific we consider a system of N identical particles. The internal
degrees of freedom of the particles are described by a one-particle Hamiltonian h
with a discrete spectrum. The particles interact through collisions during which
they may exchange internal energy, whereby the total energy is conserved. An ap-
propriate master equation analogous to the classical Boltzmann equation (3.560)
is given by

ip(t)
c = —i[h, p(t)] +D(p(t)), (3.561)
dt
which may be called the quantum Boltzmann equation. The dissipator D(p)
which replaces the collision operator of the classical Boltzmann equation is pos-
tulated to be given by the bilinear expression

D(p) = tr2 {C(p 0 p)} , (3.562)

where k is a Lindblad generator which acts on the two-particle space Ri 0 R2

as follows,

k(R) , E {T,R71,1 — —21 TctT,R — -1- RT,t,,T,„} . (3.563)

a

We have further used the symbol tr 2 to denote the partial trace taken over the
second Hilbert space R2.
The operators To, introduced in eqn (3.563) act on the two-particle space
R1 0 R2. They describe the exchange of energy between the particles during the
collisions and have to satisfy certain conditions. First, to guarantee the conser-
vation of energy one demands

[T,„ h i + h2] =-- 0. (3.564)

Second, micro-reversibility requires that

TIT, = 7-1,71,t, (3.565)

and, third, the To are supposed to be invariant under permutations 7(1, 2) of

the particles,

[7(1,2),T] = 0. (3.566)

It is easy to show that eqn (3.561) is of parametric Lindblad form (3.557).

We note further that the quantum Boltzmann equation, in close analogy to the
classical Boltzmann equation, has the following properties. It conserves energy

d
— tr(hp(t)) = 0, (3.567)
dt
 NON-LINEAR QUANTUM MASTER EQUATIONS 203

the H-theorem holds which states that the von Neumann entropy S(p) is non-
decreasing,
d
— S(p(t)) > 0, (3.568)
dt
and the canonical distribution pth = exp(-0h)/tr exp(-01/) is a stationary so-
lution, that is D(pth ) = 0. These properties can be verified with the help of the
conditions (3.564)—(3.566).

3.7.2 Mean field master equations

We consider N identical quantum systems associated with certain sites j
1, 2, ... ,N of a lattice, a spin lattice for example. The Hilbert space of the j-th
particle is denoted by Ri and its self-Hamiltonian by h = hi regarded as an
operator acting on 'Hi . The systems at two different sites i and j interact via a
two-particle potential V = which is considered to be an operator on 0
of the total system thus takes the form TheHamilton

1
HN = hi + (3.569)
j .1

and acts on the Hilbert space

(3.570)
of the total system. The interaction is scaled by a factor 1/N to ensure that the
total Hamiltonian scales with the particle number (HN) — N.
Let us denote by tr n the partial trace over the n-th Hilbert space Rn and let
us introduce the shorthand notation

tr[ n ,N] tr ntrn ±i . trN (3.571)

for the partial trace taken over the Hilbert spaces labelled by n, n + 1, , N.
Then the following theorem on the mean -field or Hartree approximation holds
(Spohn, 1980). Suppose that the initial state of the system is taken to be a
product state

PN =--p0•••Op (N factors). (3.572)

Then one has in the limit of an infinite particle number
t
lim tr[n+1,N]
N—oo
PNe tHN t = p(t) 0 • • • 0 p(t) (n factors), (3.573)

where p(t) is the solution of the Hartree equation

d
— p(t) = — i[h,p(t)] — itr2 [Vi2 + V i , P(t) o P(t)] (3.574)
dt
corresponding to the initial condition p(0) =--- p. The relation (3.573) holds for all
n and the convergence is to be understood as a convergence in the trace norm.
204 QUANTUM MASTER EQUATIONS

Essentially, the mean-field approximation consists in performing the approxima-

tion p2 (t) p(t) p(t) in the hierarchy of coupled equations of motion for the
n-particle density matrices

Pr), (t ) = t r [n ± 1 PN (t) (3.575)

where

P N (t) = exP( — iliNt)PN exP(±iHNt) (3.576)

represents the exact dynamics of the total system. The above theorem shows
that this approximation holds in the limit N --+ oc.
We will not present the detailed proof of eqn (3.573). Let us demonstrate.
however, that the Hartree equation (3.574) is obtained as a consequence of
(3.573). To this end, we first observe that for n = 1 eqn (3.573) yields

p(t) = lim tr [2 Ni p N (t), (3.577)

N—oo

from which we get

d „
—Plt =
dt
lim tr[2,N]
N—oo
[h» p A (0] + —
1
E [V, pN(t)] • (3.578)

The first term on the right-hand side can be written as

lim tr[2 N] [hi, P N (t)] E[11 j P N (t)]

N co
j =2

= —i [h, lim tr [2 N] pN(t)] i lim tr [2 ,N ] [hi, pN (0]

N oo N co
:7=2
=--- —i[h, p(t)], (3.579)

where we made use of the fact that the partial trace tr[2 , N][hi , pN (t )] vanishes
for j = 2,3, ... , N. Hence, we can write the equation of motion for the density
matrix p(t) in the form
d
— p(t) = —i[h , p(t)] — i lim —
dt N co N
1
E tr [2 , N] [, p N (3.580)

Due to the trace over R2 0 • • • 0 RN only those terms in the second term on the
right-hand side of (3.580) survive for which either i = 1 or j = 1. The second
term in eqn (3.580) is thus equal to

liM —
N—oo
1
N
E tr[2 ,N][17i i + Vii 5 P N ( t )] • (3.581
j
 NON-LINEAR QUANTUM MASTER EQUATIONS 205

Because of the symmetry of pN and pN(t) all terms of the sum over j are equal
to each other. Hence, expression (3.581) can be written as

N-1
lirfl
N
tr[2 04 2 ± 172 1 pN (0] —itr2 [171 2 + V21 p(t) 0 p(0], (3.582)
N—oo '
where we have used tr[2 ,N] = tr2tr[3 ,N] and eqn (3.573) for n = 2, that is

p(t) 0 p(t) = lim tr [3 mpN(t). (3.583)

N—oo

This shows that the Hartree equation (3.574) follows from the mean-field ap-
proximation (3.573).

3.7.3 Mean field laser equations

As an example of the application of the Hartree equation with linear dissipation
we consider a simple model for the laser, namely the model developed by Haken
(1984), Lax (1966), Louisell (1990), and Gordon (1967). To this end, we first
derive the mean-field equation governing the atom-laser mode dynamics, and
then couple external reservoirs to the system to describe the pumping of the
atoms and the field losses.
The atoms are described as two-level systems with transition frequency co
living in the Hilbert space 'HA C2 . The atoms are coupled to a laser mode
described by creation and annihilation operators at, a acting on the single-mode
Fock space 'HF. Employing the rotating wave approximation and assuming exact
resonance we can write the atom-laser mode Hamiltonian as

(.43
(3.584)
2
3=1 1=1

HA and HF are the free Hamiltonians for the atoms and for the laser mode, re-
spectively, and cq, aï, 0; are the j-th atom's 3-component of the Pauli matrices
and the corresponding raising and lowering operators in the usual notation. The
atom-field interaction is given by HAF with coupling constant
In order to bring the Hamiltonian of the system into the form (3.569) required
to apply the mean-field theory we make use of the following trick. We introduce
N identical copies ai of the field mode, satisfying the commutation relations
[ai , at] = 6ii , and perform the following replacements,

1
a —+ \TN. a ata --+ ata,
3 3' (3.585)

If we also scale the coupling constant as

g= g
(3.586)
VN
206 QUANTUM MASTER EQUATIONS

we can write the atom-laser mode Hamiltonian as

wa 3t.a-
3
+ E2 3
_igN E (ato-7 — a ia
3 3 •
(3.587)
j=1 iO3

This representation allows us to exploit the following identification,

h3 = wata + (3.588)
3 2 3
Vij = ig (alcq — a ioJ) (3.589)

Hence, the Hartree equation (3.574) becomes

d
dt P(t)
— [wata u3 ' PM ]
+g [tr(o-- p(Mat tr(o-± p(t))a, p(t)1
+g [tr(at p(t))o-- tr(ap(t))o-± , p(t)] . (3.590)

This is the mean-field approximation of the atom-laser mode dynamics.

The next step is to consider the losses of the radiation field and the pumping
of the two-level atoms. To this end, the laser mode and the atoms are coupled to
two different reservoirs. These couplings are modelled within the weak-coupling
approximation and lead to two Lindblad generators which must be added to the
right-hand side of the mean-field equation:

d
— p(t) = —i[wat a +
dt
+g [tr(o-- p(t))at — tr(o-+ p(t))a, p(t)1
+g [tr(at p(t))o-- — tr(ap(t))u± , p(t)1
+2K (ap(t)at — at ap(t) — p(t)at

+14'2 1 (o-- p(t)o-± — a + cr- p(t) —

+W12 (o-± p(t)o-- — .7+ p(t) — P(t)a- ci-+) (3.591

The damping of the laser mode through losses and output is described by the
Lindblad operator a and by a damping constant 2K. The reservoir for the atom!.
leads to the Lindblad operators u and a+, where we have written W21 and 11 -1 .)
for the rates of downward and upward transitions, respectively. To account for
the pumping the reservoir of the atoms is taken to be a reservoir with negative
temperature, that is W12 > W21
 NON-LINEAR QUANTUM MASTER EQUATIONS 207

It is now easy to verify that eqn (3.591) leads to the following closed system
of equations for the mean values (a(t)) = tr(ap(t)), (a- (t)) = tr(o-- p(t)) and
(o-3 (t)) = tr(o-3 p(t)),

(-iw - K) (a(t)) + g (o-- (t)) , (3.592)

(-iw - -y)(o -- (t)) + g (a(t))(o-3 (t)) , (3.593)

-2g ((at (t))(o-- (t)) + (a(t))(o- ± (t)))

—2-y ((o-3 (t)) — d) . (3.594)

Here, we have introduced the abbreviations

1,
-Y -2 04/12 ± 14721), (3.595)

and
14712 - W21
d , , . (3.596)
VV 12 ± VV21

Introducing as new variables the negative frequency part of the laser field

A(t) = (a(t)) exp(+iwt) , (3.597)

the negative frequency part of the polarization

S(t) (t)) exp(+icot), (3.598)

and the inversion

D (t) = (o-3 (t)), (3.599)

we may write the system of differential equations (3.592) (3.594) as

-

d
— A(t) = KA(t) + gS (t), (3.600)
dt
d
— S (t) = --y S (t) + g A(t)D(t), (3.601)
dt
d
(A* (OS (t) + A(t)S(t)*) - 27(D (t) - d). (3.602)
(t) = -2g
These are the mean-field laser equations. With the help of an appropriate
transformation of the variables A, S, and D this non-linear system of differential
equations can be shown to be equivalent to the famous Lorentz equations of
fluid dynamics (Haken, 1975). We note that the non-linear terms are the terms
which are proportional to g. These terms stem from the mean-field coupling
208 QUANTUM MASTER EQUATIONS

-<-

stable — unstable

-0. 1
o 1 2
C

FIG. 3.8. Bifurcation diagram of the non-linear equation (3.604) for the laser
field amplitude A.

between the field and the atoms. As is easily seen the system (3.600)-(3.602)
has a non-trivial stationary solution A 0 0 corresponding to the emergence of a
mean coherent field (lasing action) provided the so-called pump parameter

dg2
(3.603)
7K
satisfies C > 1. This is obviously possible only for g 0, i.e. lasing action may
be viewed as a collective mean-field coupling.
The system (3.600)-(3.602) is determined through the three time constants
K -1 , g - ', and -y - '. Since -y -3- is the smallest time in the problem we may adi-
abatically eliminate the fast variables S and D. Setting ,./ = ..6 = 0 we find a
differential equation for the field amplitude,

dA C
KA (1 (3.604)
dt — — 1 +1AP/no) '
where no = 72 /2g2 . From this equation we read off the stationary solutions as
well as their stability properties. The stationary solution A =- 0 is unique and
stable for C G 1. At C = 1 a Hopf bifurcation occurs: For C > 1 the state
A = 0 becomes unstable, while a new one-parameter family of stationary states
emerges which is determined by 1Al 2 = no(C — 1) (see Fig. 3.8).

3.7.4 Non-linear Schrödinger equation

In the previous example of the laser equations we first performed the mean-field
approximation and then added linear Lindblad generators to describe the cou-
pling to the reservoirs. This procedure corresponds to a weak-coupling approx-
imation on the mean-field level. In the following we consider the case in which
 NON-LINEAR QUANTUM MASTER EQUATIONS 209

the weak-coupling approximation is performed first on the level of the N-particle

system, and then the mean-field approximation is made. This approximation pro-
cedure leads to non-linear dissipative terms in the master equation and gives rise
to a non-linear Schrödinger-type equation.
The starting point is again a total Hamiltonian of the form (3.569). The total
system of N particles is coupled to some reservoir. In the weak-coupling approx-
imation the interaction with the reservoir is supposed to lead to a Lindblad-type
generator in the master equation for the total density matrix pN(t),

d 1
(3,605)
P N (t) = 3 N Evil
j , PN(t)

1 1 1
{wapo)wat _ W atW a pN(t) –2 PN(t)W at W'}
2
a

The operators Wa in this equation are supposed to be collective Lindblad oper-

ators of the form
wa vza , (3,606)

where the Via are single-particle operators acting on the Hilbert space of the
i-th particle. An example for such a master equation will be given in the next
subsection.
The mean-field approximation of the above master equation is obtained in a
similar way as in Section 3.7.2. Namely, in the limit of large N one obtains the
mean-field master equation

d
— —i[h, p(t)] — itr2[Vi2 + V21, P(t) g P(t)] (3.607)
dt P(t) =
[VD (P(t)), P(t)]

The first two terms on the right-hand side provide the Hartree approximation,
while the third term describes collective dissipative processes through a non-
linear potential given by

vp(p) E ftr(vatp)Va _ tr(vap)Vat} . (3.608)

2

To demonstrate the emergence of this potential we determine the quantity

1 1 at 1 a .I.
N
hm tr[2,N]
cc v
2 2 {Vja PNVa
3
— V Vi aPN
2 3 2
PN V .
3
} (3.609)
a ij
1 a 1
= lirrl tr[2 ,N] {172c' pNV 1't — - V1 tV2a pN — - pN VtV,a} + (1 2).
2 2
a
210 QUANTUM MASTER EQUATIONS

The symbol (1 2) indicates that the terms with indices 1 and 2 interchanged
must be added. Invoking the mean-field approximation (3.573) for n =-- 2 we now
get

E tr2 {V2ap 0 pVi — - 1 at

ï û
7i
a
V2 P 0 p — ?
1
0 pVrt V? } + (1 2)
a

= E {tr(Vap)pVat —1-tr(Vap)17'tp —1-tr(17'p)pV't

a
1 1
+tr(V'tp)V'p — -tr(V'tp)V'p — - tr (V c't p) pV ' 1
1
ftr(17Qtp)[17', p] — tr (V ' p)[1 7 't , p]}
=2 a
= - i[ vp ( p), P t ]
( ) ,
(3.610)

This shows that the master equation (3.605) gives rise to the non-linear potential
(3.608) on the mean-field level.
The non-linear mean-field master equation (3.607) has the property that it
can be written in the form of a non linear Schrödinger equation. Namely, if the
-

state vector OW satisfies the non-linear Schrödinger-type equation

d
(3.611)
(t) = — i [h + 1711 (°) + vD(0)] OM
then the pure state density matrix

p (t ) = 10 (t) ) K O (t ) I (3.612)

is a solution of eqn (3.607). In eqn (3.611) h is the one-particle Hamiltonian.

VH(7,b)= (011712 + 172110) represents the Hartree potential and VD(0) denotes
the non-linear potential

{ (0 117c! 1 0) va — (0 I va I 0 ) vn . (3.613)
a

Obviously, the non-linear Schrödinger equation (3.611) preserves the norm of the
state vector since trp(t) = hb(01 2 = 1 .

3.7.5 Super radiance

-

As an example of the non-linear dissipative processes and the non-linear Schrödin-

ger equation (3.611) we study here the phenomenon of super-radiance (Gross and
Haroche, 1982). To this end, we consider N identical two-level atoms with tran-
sition frequency co interacting with the vacuum of the electromagnetic field. The
 NON-LINEAR QUANTUM MASTER EQUATIONS 211

atoms are located at fixed positions 7:41) , j = 1, 2 5 ... , N. The Pauli matrices as-
sociated with the j-th atom are again denoted by GI, crat. The Hamiltonian of
the atoms is taken to be

3
HA - (3.614)
j=1

while the interaction of the atoms with the radiation field will be described in
the dipole approximation,

H, (0-; (-14. . (3.615)

j=1

The vector cr(o--jF + UT) represents the dipole operator of the j-th atom, â
being the dipole matrix element of the atomic transition. Finally, the Schrödinger
picture operator for the electric field at position i is (compare eqn (3.180))

27rvwk 6,(k) (b),()exp(ik .71 — bfA Ck) exp( d • t4)) .— (3.616)

fc,),
From the general form (3.143) of the dissipator for the weak-coupling limit
we immediately obtain the following master equation for the N-atom system,

d
— p N (t) = -iw[u , p N (0]
dt i=
1 ±_
± E aii cri
{ pNcr;F — 1 +_
- a t• a-j• PN - - PN O-i Uj . (3.617)
2 2
ij

The coefficients aii are given by

00
dt exp(iwt)(J* , t)cr , 0)) . (3.618)
=f
-00

In eqn (3.617) Lamb shift contributions have been neglected and the temperature
of the radiation field was supposed to be zero. Thus, the correlation functions in
eqn (3.618) are vacuum expectation values.
The Fourier transforms of the reservoir correlation functions can be deter-
mined explicitly. Performing the continuum limit of the radiation modes and
proceeding as in Section 3.4.1.2 one gets

f
• 6r— (k d- (i; di)) exp(ifc(i'i (3.619)
27r k2
212 QUANTUM MASTER EQUATIONS

The angular integration can be carried out to yield

4
a - = -3 w3 1d1 2 fio(x'•) + P2 (COS Oij) (Xij)} (3.620)

where we have introduced the Bessel functions

sin x 3 1) . 3
j0 (x) = i2(x) =-- -3 - sin x - - COS X, (3.621)
x x2

and the Legendre polynomial

P2 (COS 9) = - (3 cos2 0 - 1) , (3.622)

2

and defined

(Tii -77'i)1 2
xii = w14 cos2 0ii = (3.623)
1dP • lei -
The eigenvalues of the N x N matrix aii determine the relaxation rates of the
system. The diagonals of this matrix are equal to the spontaneous emission rate
for a single atom,

4 --0 2

a it. ' = -3 L03 1 6/ (3.624)

Let us suppose that all atoms are initially in the excited state which yields
the initial state

'Po) = e, • • • e) (N factors) (3.625)

for the N-atom system. According to eqns (3.620) the matrix aii is approximately
diagonal provided the mean distance between the atoms is large compared to the
wavelength A 1/w of the radiation. Hence, in this case we have aii -yo bij . The
atoms therefore radiate essentially independently of each other and the radiated
intensity follows an exponential decay law

1 (t) = eyo wNe -Y° I . (3.626)

The radiation emitted by the atoms adds incoherently which leads to a total
intensity proportional to N.
We denote by r the linear dimension of the N-atom system. For r < the
matrix elements are approximately equal to each other, i.e. a ii z•-• yo. One then
finds a qualitatively different behaviour: After a certain delay time TD the emitted
radiation occurs in a short burst whose maximal intensity is proportional to N2 .
while its width scales with N -1 . This phenomenon is called super-radiance (see
Fig. 3.9).
 NON-LINEAR QUANTUM MASTER EQUATIONS 213

30

25

20

10

O
a 0.5 1 1.5
Yo t

FIG. 3.9. Qualitative picture of the radiation intensity 1(t) of a system of N

two-level atoms. Dashed line: Independent spontaneous emission processes
lead to the exponential decay law (3.626). Solid line: The emission of su-
per-radiance occurs in a short burst with a width proportional to N -1 and
a height proportional to N2 . The super-radiance intensity profile was taken
from the mean-field result (3.643) with N = 10 and -yo tp = 0.4.

To understand this behaviour qualitatively we first note that due to the con-
dition r < Yk the coupling (3.615) is approximately symmetric under exchanges
of the atoms. Therefore, we may assume that the wave function 141(t)) of the
total system stays symmetric under the time evolution. Identifying the two-level
atoms with spin -i particles we may introduce the total 'angular momentum'
operator

1
f _6 (3.627)
2 3
.

2=1

Its 3 component is proportional to the atomic Hamiltonian HA,

-

1 3 1
J3 = -HA. (3.628)
w
j=1

The corresponding raising and lowering operators take the form

(3.629)
j=1
214 QUANTUM MASTER EQUATIONS

With the help of these operators the energy eigenstates of HA may be char-
acterized by the simultaneous eigenstates J, M) of f2 and J3 , where J =
0, 1, 2, ... , N/2 and M = J,... ,+J. These states are known as Dicke states
—

(see, e.g. Mandel and Wolf, 1995). As is well known the subspace which is to-
tally symmetric under permutations of the atoms is spanned by the states with
maximal angular momentum J = NI2. We can thus order the totally symmetric
eigenstates of the system with decreasing energy as follows,

1 ,1,M= J) = le,e, • • • ,e)

1J,M= J -1 ) = S19,e, • • • ,e)
I J, M =-- J 2) = Slg,g,e,... ,e)
—

• • • • • • • • •

1 ,1, m - -J) - Ig, ... , g ),

where S is the symmetrization operator. In view of this scheme super-radiance

can be considered as a cascade emission along a ladder of 'angular momentum'
eigenstates. The emission rate W(M) for the transition 1J, M) + 1J, M 1) is — —

given by the square of a certain Clebsch-Gordan coefficient,

(3.630)

Thus we see that the emission rate increases from the initial value W(M = J) =
2-y0 J = -yo N to a maximal value

W(M = 0,1) = -y0J(J + 1) = ±N(N + 2), for N even, (3.631)

W(M = 1/2) = 70 (J + 1/2) 2 = T(N + 1) 2 , for N odd, (3.632)

which scales with the square N2 of the number of atoms. Finally, the rate de-
creases again to W(M = J) = 0.
—

Let us now apply the mean-field theory. For r < Yt we get the master equation

d
— pN(t) = iw[J3 , pN(0]
-

dt
1 1
+-YID (J pN(t)J+ - -J+ J- pN(t) - - pN(t)J + J- ) . (3.633)
2 2
This equation is of the form of the mean-field master equation (3.605) with
the collective Lindblad operator J- and we can immediately write down the
corresponding non-linear Schr6dinger equation (3.611),

d w
— OM = -i-cr3 0(t) — iVD (0(0)0(0, (3.634)
dt 2
where
 NON-LINEAR QUANTUM MASTER EQUATIONS 215

i N;° (3.635)

is the non-linear potential. Introducing the parametrization

1p(t) exp(i0(t)) )
OM =-- ( 01(0 _ ( (3.636)
02(0 ) — V1 — p(t) exp(i0(t)) )

and using the representations (3.213) and (3.214) of the Pauli matrices one finds
the system of differential equations

5 + 2ip8 = —kip — N70p(1 —p),

/ (3.637)
---- —ich)(1 — p) — N-y op(1 — p). (3.638)
Separating real and imaginary parts we obtain 2é , —w and 2q5 =---- +c.o. The
phases 0 and q5 of the mean-field wave function thus evolve according to the free
dynamics.
Since OM represents the mean-field approximation of the N-particle system
the quantity Np(t) = N kb' (t)1 2 must be interpreted as the mean number of
atoms in the excited state. Hence,

/(t) = —Nc.o—
ddt p(t) (3.639)

is the mean intensity of the radiation. From the system of differential equations
(3.637) and (3.638) we obtain a differential equation for p(t) ,
15 = -N70p(1 - p) , (3.640)
which has the solution
p(t) PD
exp[--yo N(t — tD)]. (3.641)
1 — p(t) 1 — pp

We identify tEi with the delay time of the super-radiance and p(t =---- t D) =---- pp is
the mean occupation probability of the excited state le) at time t = t p , i.e. at
the time of maximal intensity. The above interpretation in terms of the cascade
process (see eqns (3.631) and (3.632)) yields pp -,-z:,' 1/2, such that we finally
obtain
1
p(t) = (3.642)
exp[-yo N(t — tD)] + 1 .
The intensity of the radiation is therefore given by
-2
0( 4 [cosh ( NN
1- (t) = OWN2 (3.643)
2 (t — tD))] •

We note that this formula already describes the characteristic features of super-
radiance (Gross and Haroche, 1982) as shown in Fig. 3.9: The intensity is pro-
portional to N 2 , the maximal intensity is reached after a delay time tp, and the
216 QUANTUM MASTER EQUATIONS

width of the radiation pulse scales with N -1 . It must be noted, however, that
the delay time exhibits strong fluctuations which are determined by the quan-
tum fluctuations in the initial phase of the super-radiance process. The simple
mean-field description given above cannot account for the statistics of the delay
time.

References
Alicki, R. and Messer, J. (1983). Nonlinear quantum dynamical semigroups for
many-body open systems. J. Stat. Phys., 32, 299-312.
Alicki, R. and Lendi, K. (1987). Quantum Dynamical Semigroups and Applica-
tions, Volume 286 of Lecture Notes in Physics. Springer-Verlag, Berlin.
Alicki, R. and Fannes, M. (2001), Quantum Dynamical Systems. Oxford Uni-
versity Press, Oxford.
Bausch, R. (1966). Bewegungsgesetze nicht abgeschlossener Quantensysteme.
Z. Phys., 193, 246-265.
Blum, K. (1981), Density Matrix Theory and Applications. Plenum Press, New
York, London,
Caldeira, A. O. and Leggett, A. J. (1983). Path integral approach to quantum
Brownian motion. Physica, 121A, 587-616.
Cohen-Tannoudji, C., Dupont-Roc, J. and Grynberg, G. (1998). Atom-Photon
Interactions. John Wiley, New York.
Davies, E. B. (1974). Markovian master equations. Commun. Math. Phys., 39
91-110.
Davies, E. B. (1976). Quantum Theory of Open Systems, Academic Press,
London.
Diimcke, R. and Spohn, H. (1979), The proper form of the generator in the
weak coupling limit. Z. Phys., B34, 419-422,
Diimcke, R. (1985). The low density limit for an N-level system interacting
with a free Bose or Fermi gas. Commun. Math. Phys., 97, 331-359.
Feynman, R. P. and Vernon, F. L. (1963). The theory of a general quantum
system interacting with a linear dissipative system. Ann. Phys. (N. Y.), 24.
118-173.
Gardiner, C. W. and Collett, M. J. (1985). Input and output in damped quan-
tum systems: Quantum stochastic differential equations and the quantum mas-
ter equation. Phys. Rev., A31, 3761-3774.
Gardiner, C. W. and Zoller, P. (2000). Quantum Noise (second edition).
Springer-Verlag, Berlin.
Glauber, R. J. (1963). Coherent and incoherent states of the radiation field.
Phys. Rev., 131, 2766-2788.
Gordon, J. P. (1967). Quantum theory of a simple maser oscillator. Phys.
Rev., 161, 367-386.
Gorini, V., Kossakowski, A. and Sudarshan, E. C. G. (1976). Completely posi-
tive dynamical semigroups of N-level systems. J. Math. Phys., 17, 821-825.
 REFERENCES 217

Gorini, V., Frigerio, A., Verri, M., Kossakowski, A. and Sudarshan, E. C. G.

(1978). Properties of quantum Markovian master equations. Rep. Math.
Phys., 13, 149-173.
Grabert, H., Schramm, P. and Ingo ld, G.-L. (1988). Quantum Brownian motion:
The functional integral approach. Phys. Rep., 168, 115-207.
Gross, M. and Haroche, S. (1982). Superradiance: An essay on the theory of
collective spontaneous emission. Phys. Rep., 93, 301-396.
Haake, F. (1973). Statistical treatment of open systems by generalized master
equations. Springer Tracts in Modern Physics, 66, 98-168.
Haken, H. (1975). Analogy between higher instabilities in fluids and lasers.
Phys. Lett., A53, 77-78.
Haken, H. (1984). Laser Theory. Springer-Verlag, Berlin.
Hartig, W., Rasmussen, W., Schieder, R. and Walther, H. (1976). Study of
the frequency distribution of the fluorescent light induced by monochromatic
radiation. Z. Phys., A278, 205-210.
Itano, W. M., Heinzen, D. J., Bollinger, J. J. and Wineland, D. J. (1990).
Quantum Zeno effect. Phys. Rev., A41, 2295-2300.
Klauder, J. R., McKenna, J. and Currie, D. G. (1965). On "diagonal" coherent-
state representations for quantum-mechanical density matrices. J. Math.
Phys., 6, 734-739.
Kraus, K. (1971). General state changes in quantum theory. Ann. Phys.
(N. Y.), 64, 311-335.
Landau, L. D. and Lifshitz, E. M. (1958). Statistical Physics. Pergamon Press,
London.
Lax, M. (1963). Formal theory of quantum fluctuations from a driven state.
Phys. Rev., 129, 2342-2348.
Lax, M. (1966). Quantum noise. IV. Quantum theory of noise sources. Phys.
Rev., 145, 110-129.
Lindblad, G. (1975). Completely positive maps and entropy inequalities. Com-
mun. Math. Phys., 40, 147-151.
Lindblad, G. (1976). On the generator of quantum dynamical semigroups.
Commun. Math. Phys., 48, 119-130.
Louise 11, W. (1990). Quantum Statistical Properties of Radiation. John Wiley,
New York.
Mandel, L. and Wolf, E. (1995). Optical Coherence and Quantum Optics. Cam-
bridge University Press, Cambridge.
Mollow, B. R. (1969). Power spectrum of light scattered by two-level systems.
Phys. Rev., 188, 1969-1975.
Redfield, A. G. (1957). On the theory of relaxation processes. IBM J. Res.
Deli., 1, 19-31.
Reichl, L. E. (1998). A Modern Course in Statistical Physics (second edition).
John Wiley, New York.
Scully, M. O. and Zubairy, M. S. (1997). Quantum Optics. Cambridge Univer-
sity Press, Cambridge.
218 QUANTUM MASTER EQUATIONS

Spohn, H. (1978). Entropy production for quantum dynamical semigroups. J.

Math. Phys., 19, 1227-1230.
Spohn, H. (1980). Kinetic equations from Hamiltonian dynamics: Markovian
limits. Rev. Mod. Phys., 53, 569-615.
Sudarshan, E. C. G. (1963). Equivalence of semiclassical and quantum mechan-
ical descriptions of statistical light beams. Phys. Rev. Lett., 10, 277-279.
Walls, D. F. and Milburn, G. J. (1994). Quantum Optics. Springer-Verlag.
Berlin.
Weiss, U. (1999). Quantum Dissipative Systems (second edition). World Sci-
entific, Singapore.
Wu, F. Y., Grove, R. E. and Ezekiel, S. (1975). Investigation of the spectrum of
resonance fluorescence induced by a monochromatic field. Phys. Rev. Lett., 35.
1426-1429.
 4
DECOHERENCE

The interaction of an open quantum system with its surroundings creates correla-
tions between the states of the system and of the environment. The environment
carries information on the open system in the form of these correlations. For
certain system—environment interactions the environment behaves similarly to a
quantum probe performing a kind of indirect measurement on the open system:
After tracing over the environmental degrees of freedom a certain set of states of
the open system's Hilbert space exhibits strong stability properties, while super-
positions of these states are destroyed in the course of time, often very rapidly or
even nearly instantaneously. This environment-induced, dynamical destruction
of quantum coherence is called decoherence. It leads to a dynamical selection
of a distinguished set of pure states of the open system and counteracts the
superposition principle in the Hilbert space of the open system.
The theory of decoherence allows a number of interesting physical applica-
tions, ranging from fundamental questions of quantum mechanics to technolog-
ical applications in quantum information processing. Moreover, modern experi-
mental techniques enable the observation and control of decoherence phenomena
and a quantitative comparison with the theoretical analysis. An important ap-
plication of decoherence is the explanation of the extreme sensitivity of coherent
superpositions of macroscopically distinguishable states to the influence of their
surroundings.
In a variety of theoretical models it turns out that the environmental interac-
tion leads to a decay of the coherences of such superpositions on extremely short
time scales, much shorter than the corresponding relaxation time scales of the
open system. Thus, the environment induces the emergence of effective super-
selection sectors. The latter give rise to a decomposition of the reduced system's
Hilbert space into coherent subspaces in such a way that coherences between dif-
ferent subspaces are no longer observable locally, that is through measurements
on the reduced system. This destruction of quantum coherences is of particular
relevance for the formulation of dynamical models of quantum measurements.
It provides a physical mechanism for the selection of a preferred pointer basis,
that is for a definite set of apparatus states which designate classical alternative
outcomes.
We start in Section 4.1 with a general discussion of the dynamical structure
which leads to what may be called ideal environment-induced decoherence, that
is to the destruction of quantum coherence without damping. The basic concepts
developed, such as the dynamical selection of a preferred basis, the decoherence
220 DECOHERENCE

time, and the emergence of coherent subspaces, are illustrated with the help
of a simple, analytically solvable model in Section 4.2. Section 4.3 is devoted
to a discussion of various fundamental physical decoherence mechanisms that
lead to the space localization of composite quantum objects. These are high-
temperature quantum Brownian motion, decoherence through excitation and de-
excitation of internal degrees of freedom and decoherence through the scattering
of particles. The localization in space will be seen to occur almost instantaneously
for macroscopic objects.
The interplay between decoherence and dissipation is investigated in Section
4.4 by means of the example of the damped harmonic oscillator. The central topic
will be the determination of the time it takes for the destruction of quantum co-
herences through a reservoir in the vacuum state and in the presence of thermal
noise. As an application of the theoretical analysis we study in Section 4.5 an
experiment on the decoherence of electromagnetic field states which was per-
formed by Haroche and coworkers. This experiment enabled them to observe the
progression of the decoherence in a mesoscopic variant of Schr6dinger's famous
gedanken experiment involving the superposition of a dead and alive cat.
An exact treatment of decoherence in the Caldeira-Leggett model, including
the full non-Markovian dynamics, is presented in Section 4.6. Finally, we dis-
cuss in Section 4.7 the rôle of environment-induced decoherence in the quantum
theory of measurement.

4.1 The decoherence function

In its purest form decoherence arises for certain types of system-reservoir inter-
actions. These are measurement-type interactions which are used to describe an
ideal, indirect measurement (see Section 2.4.6) on the open system whereby the
environment plays the rôle of the quantum probe.' The characteristic feature of
this type of interaction is that the reduced system affects the environment in
a way that leads to certain system-reservoir correlations whereby, however, the
back-action of the reservoir on certain system states is negligibly small. As a
result the damping of the populations of the reduced density matrix in a specific
basis representation is small, while the coherences are often found to be strongly
decaying on extremely short time scales.
The starting point of our discussion is thus a microscopic Hamiltonian of the
form

H = HS ± IIB ± kli = Ho ± Hi, (4.1)

where the interaction Hamiltonian is taken to be

I// = E in)(ni 0 BnE-- >An ® B ? . (4.2)

n n

9 1t must be emphasized that we are not talking here about a real quantum measurement
involving a reduction of the state vector, but only about a certain type of system—reservoir
interaction whose form is that of an indirect measurement.
 THE DECOHERENCE FUNCTION 221

This interaction Hamiltonian singles out a specific set of orthonormal basis vec-
tors ?),) of the reduced system, while the B„ = B;, are arbitrary reservoir oper-
ators. We assume further that the system Hamiltonian Hs commutes with the
projections An = 1 n) (n 1 which yields

[H0 + HI , An] = [H0 , An} = 0, (4.3)

such that the system operators A, are conserved quantities. As a consequence

the mean energy is constant in time, that is
d
(4.4)
It (Hs(t)) = 0.
T-
The interaction picture interaction Hamiltonian therefore takes the form
e iHo t Bne —ixot ,
H1(t) = e illot Hie—iHot = E In)(T/ 0 B(t), B(t) =
(4 . 5)

while the interaction picture time-evolution operator for the combined system
can be written as
t
U(t) = T, exp Ei f I X ØB(s) . (4.6)
o n

This expressions shows that, as an immediate consequence of the commutation

relation (4.3), the basis states In) are not affected by the coupled dynamics and
that the initial state
41(0)) = (4.7)
n

where 10 is an arbitrary reservoir state, evolves into

(4.8)
n

where
t
kbn (t)) = Texp
, [ —i f dsB n (s) 10) E V(t)ç». (4.9)
o
The state (4.8) is an entangled system—reservoir state given by a superposition
of the states In) 0 ,IOn (t)). The latter represent perfect correlations between the
various system states In) and corresponding reservoir states lOn (t)). Due to the
measurement-type interaction the reservoir carries information on the system
state. However, OM) is still a superposition involving all system states In)
222 DECOHERENCE

already present in the initial state 1*(0)). As a consequence the coherences are
still present in the reduced system's density matrix which is given by

Ps(t) = trB {1 41(0)( 41(t)} = c c 7,,In gin I ( O m ( 01 O n ( t )) • (4.10)

n ,m

It follows from unitarity that (0„(t)kbn (t)) , 1, and, thus, the diagonal ele-
ments of PS (t) are constant in time. The off-diagonal elements of ps(t), however.
do change with time, in general. The time dependence of the matrix element
(n I Ps (t) I In) is given by the overlap of the corresponding reservoir states On (t))
and 10,,(t)) which will be written as

= exp [Fnm (t)] , F„,,,(t) < 0. (4.11)

For n m the quantity r„,,(t) describes the behaviour of the off-diagonals of
the reduced density matrix and will be called the decoherence function in the
following.
The time dependence of the decoherence function Enn,(t) strongly depends.
in general, on the specific form of the system—reservoir coupling, on the various
parameters of the underlying microscopic model, and also on the properties of
the initial state. For many physical systems it turns out that the irreversible
dynamics induced by the system—reservoir interaction leads to a rapid decrease
of the overlap (0„(t)10,,(t)) when n m. Several examples of this behaviour
will be discussed in the following sections. Let us consider here the extreme case
that for n m the overlap of the states ç(t)) and 10,,(t)) decreases to zero
after times which are large compared to a typical time scale TD, the decoherence
time,

(On(t)10m (t)) (5nm for t >> TD- (4.12)

This leads to the reduced system's density matrix

Ps(t) I 1 2 1n)(nl. (4.13)

n

The coherences of the density matrix in the basis In) have disappeared as a
result of the interaction with the environment: After times t »TD the state
p5(t) of the reduced system behaves as an incoherent mixture of the states In)
in the sense that interference terms of the form (mlAn), n m, no longer
appear in the expectation value of any system observable A. Superpositions of
the states In) are therefore effectively destroyed locally which means that they
are unobservable for all measurements performed solely on the system S.
The dynamical transition expressed by (4.13) is called decoherence. According
to this relation the reduced density matrix becomes diagonal in a particular set of
basis states In) which is sometimes called the preferred basis. It is clear that these
basis states are distinguished by the form of the interaction (4.2) between system
 THE DEC OHERENCE FUNCTION 223

and environment and by the behaviour of the decoherence function embodied in

(4.12). Moreover, by virtue of the condition (4.3) the preferred basis consists of
those states which are not affected during the time evolution. These states are
thus stable with respect to the system—environment interaction.
More generally, one can study an initial state of the combined total system
of the form

P(0) == k5 (0 )) (0(0)1 0 i9B (0), (4.14)

where

10(0)) — C In) (4.15)

n

is the system's initial state and pB(0) may be any reservoir density matrix, a
thermal equilibrium state, for example. The reduced system's density at time t
can then be written as

Ps(t) , cm cIn)(ml trB {V7,7 1 (t)Vn(t)PB(°)} , (4.16)

mm

such that the decoherence function takes the form

rnm (t) = i n 1 07177 ' (t)vn (t) ) I. (4.17)

Here, the angular brackets denote the expectation value taken over the initial
density pB(0) of the reservoir.
As we have seen in eqn (4.4) the mean system energy is constant in time for
the simple class of models studied here. By contrast, the entropy of the reduced
system does depend on time, in general, since the initial pure state is transformed
into a statistical mixture in the course of time. Let us determine the behaviour
of the linear entropy (2.127),

St (OS (t)) = 1— cn 2 cm 2 exP [21-1nrn(01 • (4.18)

nm

This shows that the linear entropy can be expressed in terms of the initial pop-
ulations lcm 1 2 and of the decoherence function F mm (t). Note that Si (ps(0)) = 0
since we started out from a pure initial state. For complete decoherence in the
long time limit we get

st (ps (+00) = 1 — (4.19)

n

For example, if the initial state (4.15) is maximally entangled, that is if all
non-vanishing amplitudes cm are equal in magnitude, 1 cm l = 1/a) for n =
224 DECOHERENCE

1, 2, ... , D, we find that the linear entropy attains its maximal possible value in
a D-dimensional space,

(ps (+00) = 1— (4.20)

The interaction Hamiltonian (4.2) distinguishes a specific set of basis states

n) It is precisely this basis of states which is not influenced by the environ-
mental interaction since the projections An = In) (n1 are conserved quantities.
By contrast, superpositions of these basis states react, in general, extremely
sensitively to this interaction. Let us consider again an interaction of the form
= > A n B n , where now the operators
dn

rt.i)(n.11 (4.21)
i=1

single out an orthogonal decomposition of the system's state spacel-is into linear
subspaces A n l-is of dimensions dn > 1. The solution of the time-dependent
Schr6dinger equation corresponding to the initial condition

10(0)) = 10), (4.22)

then immediately yields the reduced density matrix

ps(t) ni)(mi'(07n(t)kbn(t))- (4.23)

nm jj ,

Under the condition of complete decoherence expressed by (4.12) this becomes

ps(t) EE Inj)(nfl for i i» TD (4.24)

n jj/

Thus we see that the coherences between the states nj) for different j and a
fixed n are still visible in the reduced system's density matrix. In other words.
coherences referring to one and the same subspace A n l-ls are not affected by the
environmental interaction. The latter thus gives rise to an orthogonal decomposi-
tion of the systems Hilbert space into coherent subspaces A nl-t s , or superselection
sectors,

= An1-is, (4.25)

such that coherences are only observable locally within one and the same sub-
space. The measurement of any system observable of the form A = En
that is of any system observable which commutes with the projections An , consti-
tutes a QND measurement on the state (4.24). This is the general framework for
 AN EXACTLY SOLVABLE MODEL 225

an environment-induced superselection rule, which has been called einselection

by Zurek (1998).
We close this section with an investigation of the short-time behaviour of the
decoherence function. To this end, we consider for simplicity the important case
that the average of the reservoir operators Bn taken over the initial state pB(0)
are equal to zero, that is (B(t)) = O. Using then the short-time expansion of
the unitary evolution operator V(t) introduced in eqn (4.9) one easily deduces

t2
Fn,n (t) — —
2 ((Bn — 13,0 2 ) , (4.26)

showing that F(t) is proportional to the square of t for short times. Corre-
spondingly, the short-time behaviour of the linear entropy (4.18) is found to
be
Si(ps(t)) t 2 E len1 2 1 cm,1 2 ((Br, — B,n ) 2 ) , (4.27)
Th ^ 7m

which grows with the square of the time.

4.2 An exactly solvable model

The general discussion of the preceding section can be illustrated with the help
of a specific system—reservoir model (Unruh, 1995; Palma, Suominen and Ekert,
1996). The advantage of this model is that, on the one hand, it shows several
important features of decoherence and that, on the other hand, it is simple
enough to allow for an exact analytic solution. As will be seen the model exhibits
decoherence in its purest form, namely the destruction of quantum coherence
without decay of populations.
4.2.1 Time evolution of the total system
We consider a two-state system which is coupled to a reservoir of harmonic
oscillators. The total Hamiltonian in the Schrödinger picture is taken to be
H= Hs + HI
= 110 ±
WO
= —0-3 ± Wkbtkbk 3 (9kbtk + g k4`bk) , (4.28)
2

where wo is the level spacing of the two-state system and k labels the reservoir
modes with frequencies Wk and bosonic creation and annihilation operators btk
and bk, satisfying

[bk , bte ] = - (4.29)

The gk are coupling constants which describe the coupling of the two-state sys-
tem to the reservoir modes bk through the Pauli matrix u3 . The Hamiltonian
226 DECOHERENCE

(4.28) provides a particularly simple example of a spin-boson model. Originally,

the model was introduced to study the influence of decoherence in quantum com-
puters (DiVincenzo, 1995). In this context the two-state system is referred to as
a qubit , the elementary building block of a quantum computer (Steane, 1998:
Bouwmeester, Ekert and Zeilinger, 2000).
Let us introduce a basis of states vectors of the qubit through o-310) = —10)
and 0- 34) = +11). In the usual representation we thus have

(1 0
C73 = (]1 -i)' 1::1) = ( 131) 1 11) = CO) ' (4.30)

The Hamiltonian (4.28) is of the form (4.1) and (4.2) used in our general discus-
sion. In particular, the Pauli matrix o-3 is a conserved quantity for this model.
since it commutes with the total Hamiltonian, [H, o-3 ] = 0. It follows that the
populations

Pi' = trs+B {1 1 )(11P(t)} = (1Ips(t)1), (4.31)

Poo = trs + B {1 0)(01P(t)} = (01Ps(010), (4.32)
are constant in time, where p(t) is the density matrix of the total system. To
determine the decoherence function for this model we first observe that the
interaction picture interaction Hamiltonian is given by

Hi(t) = e iHot Hire _iRot

= E (73 (gkek eiw kt + gk* bk e) ,
(4.33)
k

and the unitary time-evolution operator in the interaction picture can be written
t
U(t) = T, exp [—i f dslii (8)] . (4.34)
O
Since the commutator of the interaction Hamiltonian at two different times is a
c-number function,

[Hi(t),III (e )] = — g k 2 sin cuk (t — t') EE —2i(p(t — t'), (4.35)

k

we obtain,
t t { t
1
U(t) = exp — — f ds f ds' [Ili- (s), II I (8')} 0 (s — sr ) exp —i f ds.///(s)
2
[ o o o
t t
= exp [ i f ds f ds' ,o(s — 8' )0 (s — s') V (t), (4.36)
0 0
 AN EXACTLY SOLVABLE MODEL 227

where the unitary operator V(t) is defined by,

[1
V(t) = exp — o- akbtk — (4.37)
2 ]

with the amplitudes

1 — e i wk t
ak = 2gk (4.38)
Wk

Thus, we see that apart from an overall, time-dependent phase factor, the time
evolution of the total system is governed by the operator V(t) defined above. The
time development is exactly of the form given in the preceding section. Namely,
we find for an arbitrary reservoir state 0),

V (t) (10) 0 kb)) 13) 011D ( ak 2 ) kb) E 1 0) 0 100(0)1

— (4.39)

v(t) (11) kb)) = HD (+a/c/2) kb) 1 1 ) 0101(0), (4.40)

where

D(ak) = exp [akbtk — a*kbk] (4.41)

is the coherent state generator. Thus, the interaction of the system with its en-
vironment creates correlations between the system states 10) and 11) and certain
reservoir states 00 (t)) and AM), respectively. If the reservoir is in the vacuum
state initially, we find that the reservoir states

100(0) = fl — ak/ 2), 1951(t)) = 11 I + ak/ 2) (4.42)

are products of coherent states with amplitudes +ak/2, where the sign of the
shift generated by D(ak) is triggered by the quantum number of the system
state.

4.2.2 Decay of coherences and the decoherence factor

Let us suppose that the initial state of the total system is given by

p(0) = ps(0) ®pB, pB = (4.43)

Here, the reservoir is in a thermal equilibrium state at temperature T, where

3 = 11kBT and ZB is the reservoir partition function. The matrix elements of
the system's density matrix can be determined from the relation

p(t) = (ilPs(01.i) = (iltrB {VMP(0)1 7-1 (01 (4.44)

228 DECOHERENCE

where i,j = 0,1. As is easily verified, the populations stay constant in time.
Pt r (t) = Pi 1 (0)) Poo (t) = Poo (0), while the coherences behave as
Plo(t) = Pot (t) * = P100(0)e r(t) . (4.43

With the help of eqns (4.37) and (4.44) the decoherence function is found to be

F(t) = ln trB {ex p [E (a kbtk — alb k)] p B} L-- E ln (exp [cekbtk -

k k
(4.46
The angular brackets denote the expectation value with respect to the thermal
distribution pB . The expectation value

X (ak, aZ)E.--- (exp [akbtk — aZbk]) (4.47

is the Wigner characteristic function of the bath mode k. It can easily be de-
termined by noting that it represents a Gaussian function, which immediately
leads to the expression

x (a k , al,) = exp [---21 1ak 1 2 ({b k ,btk })] . (4.48'

Thus, we find

`Ok 1 2 coth (wkl2kBT) (1 — cos w k t) .

lak 2 (t b k, b tk 1 ) = 2
Wk
(4.49
We now perform the continuum limit of the bath modes. Introducing th(
density f(w) of the modes of frequency w and defining the spectral density as
Au)) = 4f(w) I g(w) 1 2 , (4.50

we can write the decoherence function as

.
F(t) =— f
0
dwAw) coth (w/2kBT)
1 — cos wt
C4J
2
(4.51

Thus, we have obtained an explicit expression for the decoherence function for
the present model. Obviously, r(t) depends on the temperature T of the environ-
ment and on the form of the spectral density J(w). To illustrate the dynamical
behaviour of the decoherence function let us take a spectral density of the form
J(w) ,-- Awe -4° . (4.52

We assume a linear increase of J(w) for small frequencies and an exponential

frequency cutoff at the cutoff frequency Q. Such a form for the spectral density
 AN EXACTLY SOLVABLE MODEL 229

is typically obtained in the quantum optical regime, where g(w) — \AT.; and by
assuming a one-dimensional field of bath modes with a constant mode density,
f (w) = constant. Note also that for this case we have a dimensionless coupling
constant, that is, the constant prefactor A of the spectral density is dimensionless
and will be taken to be equal to 1.
In order to determine the decoherence function it is helpful to split it into a
vacuum part F,e (t) and a thermal part F th(t) as follows,

F(t) = r vac (t) +rth(t). (4.53)

The vacuum contribution can be determined explicitly,

co
1
rvac (t) EZ — f dwe-4s21 — cos cut = -- ln (1 ± Q 2 t2 ) . (4.54)
w
a
It is independent of the temperature and describes how the fluctuations of the
field vacuum affect the coherence of the open system. This part depends on the
cutoff frequency Q. The thermal contribution to the decoherence function is given
by
00
1 — cos wt
Fth(t) EE — f &le' I Q [coth (w12kBT) — 1]
0
t
1
-
a f ds f dx e—Tx/O
=— ks
[coth(x/2) — 1] sin(sx/O). (4.55)
'•-- o o
If we assume that kB T < 12 the thermal contribution is found to be
t co

13 f
rth (t) 'r:' — T ds f dx[coth(x / 2) — 1] sin(sx/0)
o o
= _ 1n [sinh (t/TB)1
(4.56)
L OTB ]

Here we have carried out the x-integration with the help of the formula

i
o
dx[coth(x/2) — 1] sin(ax) = 7r coth(rra) — --(: (4.57)

and we have introduced the thermal correlation time

)3 1 19 5
2.43 10 --- (4.58)
7r 7rkB T T [K] •
230 DECOHERENCE

10

0
10-2 10-1 100 101

FIG. 4.1. Semilogarithmic plot of the decoherence function F(t) according to

eqn (4.59) as a function of the dimensionless time variable T = t / TB, where
S2 TB = 20 (solid line). The three approximations given in eqns (4.60), (4.61
and (4.62) are also indicated (dashed lines).

Summarizing these results, we can write the decoherence function as

1
F(t) = -- ln (1 + S22 t2 )
rs inh (t/TB )1 (4.59
2 [ t/TB •

One clearly recognizes three different regimes of time:

1. The short-time regime t < Sr i : In this regime the magnitude of F increases
with the square of t,

F(t) -- Q2 t2 , (4.60
2
which also follows directly from the short-time expansion of the time-
evolution operator (see eqn 4.26). Note that the short-time behaviour in
this regime is fully determined by the vacuum contribution Fvac .
2. The vacuum regime Sr' « t < TB: Here, we may approximate the deco-
herence function as

F(t) — ln Qt. (4.61

In this range of time, decoherence effects are mainly due to the vacuum
fluctuations of the field.
 AN EXACTLY SOLVABLE MODEL 231

3. The thermal regime regime may also be called the Markovian

TB t: This
regime since the magnitude of the decoherence function increases linearly
with time,

(4.62)

which means that the off-diagonals of the reduced density matrix decay
exponentially with a rate given by 77,6 1 •
A plot of the decoherence function F(t) is shown in Fig. 4.1 together with these
approximations.

4.2.3 Coherent subspaces and system-size dependence

The above simple model can also be used to illustrate the emergence of coherent
subspaces and the dependence of the decoherence function on the system size
(Palma, Suominen and Ekert, 1996). To this end, we consider N qubits, labelled
by an index j, which interact with the reservoir through an interaction of the
form used in eqn (4.28). Thus, the total Hamiltonian reads

H =—
2
kb tkb k EE (gkib kt + g;i bk) . (4.63)
k

We assume here that the qubits do not interact directly with each other. The
coupling constant gki describes the coupling of the j-th qubit with mode k. We
further suppose that the qubits take on fixed positions FO) in space and that the
reservoir may be characterized through a certain correlation length r e . We can
then distinguish two extreme situations.
First, we consider the case that the minimal distance between the qubits is
large compared to the correlation length r e of the reservoir. The qubits may
then be supposed to interact with independent reservoirs, one for each qubit.
Denoting a specific qubit configuration by {in(i ) }, where the m(i ) take on the
values 0 or 1, we have for the matrix elements of the N-qubit density matrix

({ 15-1(i) } 1PS f in(i) }) ({ 171(i) } S

{m(i)}) (4.64)

x 11 (exp [(mu) _ mu)) E akek — a*k bk)1) •

The decoherence function can therefore be written as

_ m (i F(t),
)
(4.65)
:1=1

where F(t) is the decoherence function for a single qubit, given by eqn (4.46).
Thus, the decoherence function is an integer multiple of the decoherence function
232 DECOHERENCE

of a single qubit. It vanishes if and only if fri (j) rn(i) for all qubits, that is on
the diagonal of the N-qubit density matrix. If exactly two single-qubit quantum
numbers are different from each other it is given by the expression for a single
qubit, while it is proportional to the total number of qubits if the quantum
numbers of all qubits are different. In the latter case we have

= N • F (t) , (4.66)

corresponding to maximal decoherence.

Let us now consider the case that the linear dimension of the total N-qubit
system is small in comparison to the correlation length r e . The qubits then inter-
act collectively with the reservoir, that is we may assume that for a fixed mode
k all g ki are equal to each other. The system of qubits therefore effectively inter-
acts with the reservoir through the collective operator E .7 o.3
decoherence
function is thus found to be

Feih( 3 ), m (i ) , — F(t), (4.67)

where M = E . mu) and M = E in ( . This shows that the decoherence func-

tion for the NAubit system is proportional to the square of the difference of the
sums M and M of the quantum numbers { m (j) } and Wn (i) }.
The coherent subspaces are defined by the condition M = M, that is a specific
coherent subspace is spanned by the basis states 1{m(-1) 1) with a fixed value for
the sum M = E . mu). For example, if N = 2 the two-qubit states 110) and 101)
form a non-trivial, two-dimensional coherent subspace.
The formula (4.67) demonstrates that collective interactions may lead to a
strong amplification of decoherence. The case of maximal decoherence is obtained
for M = N M = 0, or for M = 0, M= N, which yields

= N2 • F (t) (4.68)

Summarizing, in the case of independent reservoir interactions the maximal de-

coherence increases linearly with the system size N, while it grows with the
square N 2 of the system size in the case of collective interaction.

4.3 Markovian mechanisms of decoherence

A number of important physical processes leading to decoherence can already
be analysed with the help of simple Markovian models. Here we discuss some
basic physical mechanisms which cause the decoherence of the centre of mass
coordinate of a composite quantum object and, thus, lead to a localization of
the object in position space. To make the discussion as simple as possible we
restrict ourselves here to the recoilless limit, that is we neglect damping effects.
The generalizations to include these and, in addition, non-Markovian effects will
be developed in the following sections.
 MARKOVIAN MECHANISMS OF DECOHERENCE 233

4.3.1 The decoherence rate

We consider a master equation of the following general form,

d
Ps(t) = [11s, Ps(t)] — A [Z, [Z, Ps (t)]i (4.69)

where
1 ,2
HS =2m
-P (4.70)

generates the free evolution of the centre-of-mass coordinate Y of the quantum

object with total mass in. As will be shown in the next sections, the decay of
the off-diagonals in the position representation of the reduced density matrix Ps
often occurs on an extremely short time scale, much shorter than those corre-
sponding to the damping of the diagonals and those of the free evolution of the
reduced system. To a first approximation we may thus neglect the free evolution
altogether and solve the master equation by disregarding the Hamiltonian part.
This yields

ps(t,Y, exp [—A (Y Y`) 2 ps(0,Z,P), (4.71)

showing that the off-diagonals are damped by a factor exp [—AAx2 t] , where

Ax (4.72)

measures the distance to the diagonal of the density matrix. The quantity A
is referred to as the decoherence rate with dimension (time) — ' x (length) -2 .
The decoherence function introduced in the preceding section therefore takes
the form °

F(t) = —AAx2 t. (4.73)

This allows us to define a corresponding decoherence time as

1
= AAx 2 (4.74)

In the following we discuss three basic physical mechanisms which may be

described by a Markovian master equation of the form (4.69). These are quan-
tum Brownian motion in the high-temperature limit, decoherence through spon-
taneous and thermally induced transitions of internal degrees of freedom, and
decoherence by scattering of an incoming particle flux. It will be seen that in all
1 °The above argument of neglecting the free evolution will be made more rigorous in the
next sections, see in particular eqn (4.234).
234 DECOHERENCE

cases the decoherence rate typically takes the form of a product of a characteristic
rate and a squared wavenumber,

A — (rate) x (wavenumber) 2 . (4.75)

The rate characterizes the physical process causing the decoherence, i.e. it is equal
to the relaxation, to the transition, or to the scattering rate. The wavenumber
is determined by the thermal wavelength of the decohering object, by the wave-
length of the emitted radiation, or by the de Broglie wavelength of the scattered
particles.

4.3.2 Quantum Brownian motion

As our first case we discuss the emergence of decoherence in those situations
in which the motion of the collective degree of freedom can be described by
the Caldeira—Leggett quantum master equation (3.410). In the recoilless limit
(see the end of Section 3.6.4.5) the master equation for the collective degree of
freedom reduces to the form (4.69), where the decoherence rate is given by
(reintroducing factors of h)
= 2m-ykBT
A (4.76)
h2 •
Introducing the thermal wavelength
h
Ath - (4.77)
V217/kBT

and the corresponding wavenumber kth = 1/Ath we can cast the decoherence
rate into the general form (4.75),

A = -yqh . (4.78)
As expected, the relevant rate here is the relaxation rate -y while the characteristic
wavenumber is given by the thermal wavelength of the object.
Due to the neglect of the free evolution and of damping effects, the diagonal
elements of the density matrix are not affected in the course of the time evolution
whereas the off-diagonal elements are strongly decaying. We generally define the
relaxation time TR to be the decay time of the square of the momentum of the
particle. According to eqn (3.427) the momentum of a free Brownian particle
relaxes as exp(-27t) which gives TR = 1/4-y. From eqns (4.74) and (4.78) we
thus infer that the ratio of decoherence time to relaxation time may be written
as
2
TD = 4 ( Ath) (4.79 )
Ax
Since for macroscopic objects the thermal wavelength is extremely small thi!-)
estimate shows that the decoherence time can differ from the relaxation time by
 MARKOVIAN MECHANISMS OF DECOHERENCE 235

many orders of magnitude (Caldeira and Leggett, 1985; Unruh and Zurek, 1989).
For example, if we consider a particle of mass in = 1 g at room temperature,
T = 300 K, and assume that Ax = 1 cm, we find that the ratio of the two time
scales TD and TR is of the order TD/TR — 10-40 (Zurek, 1991). Thus, even if
we choose TR to be of the order of the age of the universe (TR — 10 17 s) the
decoherence time scale is found to be extremely small, namely TD 10 -23 S.
4.3.3 Internal degrees of freedom
The destruction of coherences of the centre-of-mass density matrix can be caused
by spontaneous or induced transitions involving internal degrees of freedom of the
composite object. Again we denote the centre-of-mass coordinate of the object by
ff and concentrate on two internal energy levels le) (excited state) and g) (ground
state) separated by a transition frequency cuo. The environment is taken to be a
thermal radiation field at temperature T. Thus, we have spontaneous emissions
with a rate -yo and thermally induced emission and absorption processes. The
Markovian master equation for the density matrix p(t) of the object, including
the centre-of-mass as well as the internal degree of freedom, takes the form of a
Lindblad equation
d
— p(t) = [Hs + wo le)(el,P(t)] (4.80)
dt

+7°(N4ir± 1) f d 9 [A(rc)P(t)A t — 2 {A t (rc) P(t)}]

+ d [At AP(t)A(k) — —1 {A(k)A t (k),P(t)}]

47r 2
Here, N = N (w o ) is the Planck distribution and the integration is performed
over the solid angle element c/ 12 into the directions ic7liztl of the photon emission,
which, for simplicity, is taken to be isotropic.
The Lindblad operator A(k) provides both the transition e g of the
internal state and the momentum recoil caused by the emission of a photon with
wave vector rc. The momentum recoil is described by the operator exp(— i)
which changes the momentum of the object by Hence we have
A(k) = At(k) = e i" o-± (4.81)
where u_ = 1g)(el and a+ = le)(gl. We remark that in the master equation (4.80)
the frequency shifts induced by Doppler effect and recoil are neglected, which is
justified in the limit of a large total mass in of the object.
Our aim is now to derive an effective master equation for the reduced density
matrix of the centre-of-mass coordinate,
ps(t) = trint {p(t)} , (4.82)
where the trace is taken over the internal degrees of freedom. On using p
PS 0 pint we immediately find with the help of eqn (4.80),
236 DEC OHERENCE

d
Ps(t) = [II s Ps(t)] (4.83)
+ 1)
pe (t) f d 12[e -4-Î ps(t)e i" — ps(0]
47r
± 70N p f [eik •g ps(t ) e — tk'.d _ ps(t)]
47r g
where pe (t) = (elpint (t)le) and 1)9 (0 = 1 — p(t) are the populations of the
excited and ground state levels, respectively. These populations are to be deter-
mined from the internal density matrix pint (t), which is obtained from p(t) by
taking the trace over the centre-of-mass coordinate. Since the dynamics govern-
ing the internal degrees of freedom decouples from the centre-of-mass motion
one immediately finds

d „
pe (t) = 'Yo(2N 1)p(t) 'y o N. (4.84)

If the object is initially in the excited state, pe (0) = 1, for example, we get the
solution
N+1
pe (t) = exp[ -yo (2N + 1)t] + • (4.85)
2N + 1 2N + 1
Performing the angular integration in (4.83) now yields the following position
space representation for the master equation governing the centre-of-mass density
matrix,
d
— ps(t,Y ±") = — i41[11s,Ps(t)11±") (4.86)
dt '
sin (kV _x 1)1
(N + pe (0) [1 Ps(t,x,x ).
1 ,1

This is the master equation which describes the motion of the centre-of-mass co-
ordinate of the composite object, where k = w0 / c = 1/ is the wavenum-
ber of the emitted radiation.
We can distinguish two important limiting cases. First, we suppose that
kAx >> 1, that is Ax » À. In this limit the second term in (4.86) approaches

— 110 (N + Pe (0)Ps(t, (4.87)

showing that the decay rate of the off-diagonals becomes independent of Ax.
The decoherence thus saturates for distances from the diagonal which are large
compared to the wavelength of the radiation. Moreover, we note that the de-
coherence rate is approximately equal to the sum of the rates for spontaneous
emission (contribution -yope ), for induced emission (contribution -yo Npe ), and for
induced absorption (contribution -yo Np g ). The rate given in eqn (4.87) depends
on time through the internal dynamics given by eqn (4.84). However, for small
 MARKOVIAN MECHANISMS OF DECOHERENCE 237

times, -yo (2N+ 1)t < 1, and with the initial condition pe (0) = 1 the decoherence
time is found to be
1
(4.88)
70 (N +1) .
For example, at zero temperature, N = 0, we simply get TD -= 1/-yo. This is an
obvious result: If the wavelength of the radiation is small compared to the dis-
tance Ax =Ii—'1 between two superimposed localized wave packets describing
the centre-of-mass coordinate, the emission of a single photon already enables
an approximate localization of the object and thus leads to partial destruction
of the coherence of the superposition. At zero temperature the decoherence time
must therefore be equal to the average time for the emission of a photon, that
is to the inverse of the spontaneous emission rate.
Let us now consider the other extreme case, namely kAx <1 which means
Ax < 3t. An expansion of the term 1 — sin(kAx)/kAx in eqn (4.86), yields a
master equation of the general form (4.69) with the following expression for the
decoherence rate,
1
A es,' — 70 (N + pe (t))k 2 . (4.89)
6
Again, the decoherence rate depends on time. However, for high temperatures,
N» 1, we get a time-independent rate,
1 1 kB T 2
A R.,' — -yo Nk2 R', 6 -yo riwo k . (4.90)
6
This decoherence rate is again of the general form (4.75): The characteristic
wavenumber k is that of the transition radiation, while the characteristic rate is
provided by the rate of thermally induced processes.
On the other hand, at low temperatures and for times t satisfying -yo t < 1
the decoherence rate is found to be
1
A R,' — -yo k2 (4.91)
6
if the system is initially in the excited state. To give an example, let us consider
the 2p 1s transition of hydrogen. For a hydrogen atom we have -yo — a3 coo ,
where a = e 2 Ihc is the fine structure constant. This leads to the simple estimate
1 1
c) , io 2o
A — —c(al3 (4.92)
6 cm2 s
for the decoherence rate of the atomic transition.

4.3.4 Scattering of particles

We consider finally the destruction of coherence through the scattering of an
incoming flux of particles off the composite object. In general, many types of
238 DECOHERENCE

scattering reactions contribute to the decoherence. In the case of an incoming

photon flux, for example, we could have Thomson scattering, elastic Rayleigh or
inelastic Raman scattering. To derive a master equation for the reduced density
matrix of the centre-of-mass motion (Gallis and Fleming, 1990; Joos and Zeh.
1985; Joos, 1996) we shall use here the assumptions that the scattering is elastic
and that recoil may again be neglected. This means that only the state of the
scattered particles changes during the scattering reaction, while the state of a
localized target object is left unchanged.
If we take the centre-of-mass coordinate of the object to be a position eigen-
state 4') we may thus write for the scattering reaction

4')Çb) S (0 ) =1410z), (4.93)

where 10) is the incoming wave function and S the scattering matrix. The out-
going wave function is denoted by 10i). It represents the scattered wave for a
scattering centre located at Y. Note that it is assumed here that the scattering
time Tsca,tt is small compared to the typical time scale for the systematic system
evolution. As a result of a single scattering reaction the density matrix of the
centre-of-mass coordinate of the object then undergoes the transition:

Ps (i, Z 1 ) Ps (Y, ±")(Og'10i)• (4.94)

Thus, the matrix element ps (i,Z') is multiplied by the overlap of the waves
scattered by Y' and i'.
To determine the overlap of the scattered waves we invoke the assumption
that the S-matrix commutes with the total momentum, that is with the sum of
the object's momentum ./7 and of the momentum 4.' of the scattered particle,

[S,g+ 41 = 0. (4.95)

On writing the initial state as

14 kb) = exP( — ig' Z)4' = 0)10 = exP [ — i(g+ 0 . 4 4' = 0) exp(q. ) kb)
(4.96)

the conservation of total momentum thus gives

S (0))
= exp [ — i(g+ 0 ' 4 W = 0)S (exip(iT• 414 5))
= 1Y) exP [ — iOE - 4 so (exP(q . z)10))
E lz)100 , (4.97)

where So denotes the S-matrix for the scattering at Z = 0. The overlap of the
scattered waves can therefore be written as

Ksbi , Is/5 0 = KO exp(—q. z')s(c; exp [—i

4'. (i — i')] so exp(iq'. z) 0) . (4.98)
 MARKOVIAN MECHANISMS OF DECOHERENCE 239

Next we suppose that the incoming particle state çl)) represents a momentum
eigenstatelk) which is normalized to 1 in a quantization volume L3 . Introducing
the T-matrix through

So = I + iTo (4.99)

and employing the unitarity of the S-matrix one easily deduces

100 = 1+ E (exp [i (krci)

_ _ it)] - I (re IT01 2 (4.100)

We perform the continuum limit, replacing

27)3 f
(T d3 k t = ( i )3 fdkk12 f dif , (4.101)

and

1 6 (k' k)1 2 LS(k i — k). (4.102)

The scattering amplitude f(i', ict) is then defined in terms of the T-matrix as

(k''Tolk) 27ri k f (k' r (ki k), (4.103)

such that the overlap of the scattered waves can be cast into the form,

= 272 f c/S/' (exp [i(k' — re)(Z — — 01.0, 14 2 . (4.104)

1+ L

With this expression we can write the change Aps of the density matrix during
a time interval At as a result of a single scattering event as follows,

APs (Oirkb i- ) —1 (4.105)

At At
Ps.

Let us now suppose that many scattering reactions take place during the
time interval At, which implies that the interval At must be chosen in such a
way that it is much larger than • scatt and much smaller than the characteristic
time scale of the free evolution of the system. In addition, we also assume that
the incoming state may be described by an incoherent mixture of momentum
eigenstates 11-c) and that the corresponding distribution of incoming momenta
is isotropic. To describe the incoming state we define dk 1(k) to be the flux of
incoming momenta in the interval [k, k + dk], that is the quantity dk 1(k) 1.2 At
is the number of incoming particles during the time interval At and having
momenta lying in the interval [k, k + dk]. The total rate of change of the density
240 DECOHERENCE

matrix is obtained with the help of an integral over all momenta and by an
average over all directions. Hence, we have

dPs f dk f df2 I (k) L 2 At Ps . (4.106)

dt 47r At
Substituting expression (4.105) and adding the free evolution, we thus obtain
the density matrix equation (Gallis and Fleming, 1990)

d
— ps(t, , Y 1 ) =
dt [He ps(t )] ) – F (Y ±")ps (t, (4.107)

where

F — = f dk I (k) f df2df21 (1 exp [i(rc — rci)(Y(I?

— )1f , k)1 2 .
(4.108)

As in the preceding subsection we may distinguish two important limiting

cases. Suppose first that koAx >> 1, where ko is a typical wavenumber of the
scattered particles. The exponential in expression (4.108) then averages to zero
and we get
1
F f dk I (k) f df2df2 1 1 f (le , lc)
— = 271 f dk I (k)u(k) = 27r-y/scatt, (4.109)
2

where a(k) is the total cross-section and 'NJscatt the total scattering rate. Similarly
to the decoherence through transitions between internal levels, the decoherence
induced by scattering thus saturates for large distances Ax and occurs at a
rate which is equal to 271 times the total scattering rate. The saturation of the
decoherence function is easily understood: For ko Ax >> 1 the wavelength of the
scattered particles is small compared to the distance Ax. A single scattering
reaction thus provides sufficient information to localize the object. Increasing
Ax any further cannot enhance this information.
On the other hand, for small distances, ko Ax < 1, we find

F( — R.,' f dk I (k) f df2df2 1 4-1 [( k — – ±")] 2 f k)I2 (4.1101

Introducing spherical coordinates such that cif/ = d cos 84 , (K2' = d cos 01 dcio' we
find

F (Y – f dk I (k)o-eff (k)k 2 (Y Y 1 ) 2 , (4.111

where
f(1
C eff (k) = f dildf 11 1 711 (cos – cos 01)2 itc) 1 2 (4.1121
 MARKOVIAN MECHANISMS OF DECOHERENCE 241

may be viewed as an effective cross-section. For example, if the differential cross-

section is isotropic one obtains
27r
aeff(k) = — 0- (k). (4.113)
3
These expressions reveal that the decoherence rate describing the decoherence
through scattering off the object may be written

A= f dkI(k)Gr eff(k)k 2 7scattk a2v . (4.114)

This is an expression which is again of the general form (4.75). Obviously the
relevant wavenumber is here equal to an appropriate average ka, of the de Broglie
wavenumber of the scattered particles and the characteristic rate is given by the
total scattering rate 'Y,scatt •
As a specific example we consider the decoherence through scattering in a
photon gas at temperature T (Joos and Zeh, 1985). If the thermal wavelength
of the photons in the gas is large compared to the radius a of the object we may
assume that the scattering cross-section is given by the Rayleigh cross-section.
The evaluation of the formula (4.114), employing an average over the Planck
distribution of the photon gas, yields the following estimate for the decoherence
rate,

T) 9 ( a 6 1
A — 102° (--- (4.115)
cm )cm2 s
Note the extremely strong dependence on the size a of the object and on the
temperature of the gas. The decoherence rate increases with the sixth power of
a which is due to the a-dependence of the Rayleigh cross-section in the limit of
large wavelengths. The increase of A with the nineth power of the temperature T
can be understood as follows. First, the photon flux I is proportional to the third
power of T, which is just the Stefan—Boltzmann law of black-body radiation. In
addition, the average of the product a(k)k 2 increases with the sixth power of the
thermal wavenumber kt h = kBT !he since the Rayleigh cross-section increases
with the fourth power of the wavenumber. For example, for an object of size
a = 10 -6 cm corresponding to a large molecule we get A = 10_ 12 cril -2 s—i
at T = 3 K (cosmic background radiation), and A = 106 cm -2 s -1 at T =
300 K (room temperature). For a small dust particle, say a = 10 -5 cm, the
corresponding decoherence rates increase to A = 10-6 cm-2-1 s (T = 3 K) and
A = 10 12 cm -2 s -1 (T = 300 K).
For objects whose size is large compared to the wavelength the cross-section
is approximately equal to the geometric cross-section. The decoherence rate may
then be estimated from
5
(71 ) ( a ) 2 12
A — 5 10" (4.116)
1() cm) cm s
242 DECOHERENCE

In this range A increases with the second power of a and with the fifth power of
T. For an object of size a = 10-1 cm this yields A --, 1025 cm -2 s -1 at T = 300
K. Further examples of the decoherence caused by scattering have been discussed
by Tegmark (1993).

4.4 The damped harmonic oscillator

In the foregoing section we have treated the simplest case, namely the destruc-
tion of quantum coherence without damping. In more realistic physical models
decoherence occurs together with a decay of populations. To study the effects of
the combination of both phenomena we first investigate the damped harmonic
oscillator in the quantum optical limit given by the master equation (3.307).
Let us consider as initial state a superposition of two coherent states,

OM ) = Ar (la) + IS)) .
(4.117)

In reference to Schr6dinger's famous gedanken experiment (Schr6dinger, 1935)

such a superposition is sometimes called a Schr6dinger cat state. The super-
posed states c)z) and 1,(3) are considered to be macroscopically distinct with an
extremely small overlap, representing the dead and the alive cat, respectively.
The normalization factor ./1/- reads

A( = (2 + 2R(aP)) 1/2 (4.118)

'
where the overlap (alfi) is given by eqn (3.321). The corresponding initial density
matrix of the oscillator is thus

PS( 0 ) = Af2 (1a)( a l + 1/3 )(S ± la)(SI + Ma l) • (4.119)

Our aim is to determine the time evolution of this initial state and to define the
decoherence function through the behaviour of the interference term la)(fil +
1S)(al in this equation. In the following Section 4.4.1 we present the solution to
this problem for zero temperature. The general finite temperature case will be
discussed in Section 4.4.2.

4.4.1 Vacuum decalterence

At zero temperature the master equation (3.307) can be written as

d
7 ) ps(t)at a + -yo aps Mat
— ps(t) = (-iwo - 1°-) at aps(t) + (+ 0 - —
dt 2 2
Lps(t). (4.120)
As we know already, a coherent state remains a coherent state under time evolu-
tion, which makes the solution of the problem particularly simple for the vacuum
optical case. We therefore try the ansatz

a (t) E exp[Lt]la)(3 1 = f (t)la(t))(fi(t)I, (4.121)

 THE DAMPED HARMONIC OSCILLATOR 243

where a(0) a and NO) 0 and f (t) is a c-number function with f(0) = 1,
such that o- (0) = la)(OL On using

d (eil ( t f 1 d
Tt ict(t)) = ,cv (t ))a a — Tit ct(t)12) la(t)) (4.122)

as well as the fact that the coherent states are eigenstates of the annihilation
operator, one verifies that o- (t) given by (4.121) indeed solves the master equation
(4.120) provided the differential equations

6(0 _ t (t) . 'Yo (4.123)

a(t) OM zw° 2'

0(01 2 + 1t3 (t) 2 ) + 700 * (t ) a (t) (4.124)

f (t) 2 dt

are satisfied. These equations are easily solved to yield

_ 'Yot)
a(t) = a exp (_i wo , (4.125)
2
-Yot)
/3 ( t) = fi exp (_i wo _ , (4.126)
2
L 1ial2
f (t) = exp [(1 _ e —yot)

= (Sla)[1—exP(--Y0t)]. (4.127)

Summarizing, the solution of the master equation (4.120) corresponding to

the initial state (4.119) takes the following form,

PS(t) = AP (1 a (t))( Ct (t )1 + 0(t))(0(0 + f(t)a(t))(/3(t) 1 + f* (t)Ifi(t))(a(01) •

(4.128)

The decoherence function r(t) can now be defined as the logarithm of the mod-
ulus of the factor f(t) multiplying the off-diagonals in the coherent state repre-
sentation,

1 2 f ,y t\
IT(t) = lnlf(t)1= - -
a— r(3- 0- — e ''' ) • (4.129)

One observes that r(t) is proportional to the square of the distance of the initial
coherent state amplitudes a and 0 in the complex plane. For --yot >> 1 the deco-
herence function approaches the value —la —,31 2 /2, that is exp(r) approaches the
absolute value (43)1 of the overlap of the initial states. For widely separated
coherent states this overlap is extremely small which means that the coherences
practically vanish in the long time limit.
244 DECOHERENCE

FIG. 4.2. Wigner distribution corresponding to the superposition (4.117) of two

coherent states with a = ao exp[i0] and ,3 = ao exp[—i0], where ao = fib
and 0 = 0.7.

On the other hand, for -yo t < 1 the decoherence function is proportional to
time,
1
r(t) ,-- -a — /3 1 2 7ot.

This allows us to define a decoherence time TD through

TD 2
—=,

where we make use of the relaxation time TR = -y(V. . One observes that the ratio
of the decoherence time to the relaxation time is inversely proportional to the
square of the distance of the initial states in the complex plane. Equation (4.131)
is an important relation which is encountered in many microscopic models of de-
coherence. It tells us that for widely separated initial states the decoherence time.
that is the time over which coherences are destroyed through the interaction with
the environment, is much smaller than the relaxation time, which characterizes
the time after which the system loses its energy through dissipative effects.
The difference of relaxation and decoherence time is illustrated in Figs.
4.2 and 4.3, where we have taken the initial superposition (4.117) with a =
 THE DAMPED HARMONIC OSCILLATOR 245

FIG. 4.3. Wigner distribution of the state which evolves from the initial state
(4.117) shown in Fig. 4.2 after time t = 0.15-yo— ' according to the master
equation (4.120).

ao exP{i0i, = ao exp[—i0] and a() real. The initial phase components are thus
separated by an angle 20 in the complex plane, such that la — /312 = 4a8 sin 2
Figure 4.2 shows the interaction picture Wigner distribution (2.81) of this initial
state while Fig. 4.3 shows the corresponding state at time t = 0.15TR . The time
TR characterizes the time scale it takes for the two phase components to merge
at the origin of the phase plane. We observe that the coherences of the initial
state rapidly decay over a time which is much shorter than the relaxation time.
To interpret the result (4.131) physically (Caldeira and Leggett, 1985) we
assume for simplicity that 0 = 0, which means that one of the superposed states
is the ground state of the oscillator. Equation (4.131) then yields
TD = 2
_ (4.132)
TR n
where 1a1 2 = a1a) n » 1. The question is then, why is the decoherence
time obtained above smaller than the relaxation time by the factor n, that is
by the average number of quanta in the initial state? First, we note that at
zero temperature the environment is in the ground state vacuum which will be
denoted by 10)B. Thus we have the following initial state of the combined system
(neglecting the overlap (al0)),
246 DECOHERENCE

1
(la + 1 0 ) 0 10)B.
) )
(4.133)
. ■/-
The rate for emission of a quantum of energy from this initial state is approxi-
mately equal to W = -yon/2. Accordingly, the time it takes for the emission of a
quantum is of the order 1/W = 2TRIn. After this time the initial state evolves
approximately into the state

1
IT .f) -'•:-:, —fr.-, (la') OMB +1 0) 0 1 0 )B) • (4.134)
vz
Here, la') is the state 1a) after the emission of one quantum, while 11) B denotes
a state of the reservoir containing one quantum of energy. The reduced density
matrix of the oscillator is then given by

1
PS = trB{ I llif O f 11-z-',- (1a1 ) (ct 'l + 0) (0l), (4.135)

since the reservoir states 10)B and 11) B are orthogonal. This shows that coher-
ences are already destroyed after the emission of one quantum, that is already
after a time of order 27-R /n, whereas the dissipation of energy takes the emission
of n quanta, that is a time of order TR. This is what is expressed by eqn (4.132).

4.4.2 Thermal noise

It must be expected that thermal noise leads to an enhancement of decoherence.
To determine the decoherence time in the case of finite temperatures we write
the solution of the master equation (3.307) corresponding to the initial state
(4.119) as

PS(t) = Ar2 (Pact(t) + POO(t)+ p(t) + P0a(t)) ) (4.136)

where we have introduced

p(t) = exp [ft]Ict)(al, (4.137)

p(t) = exP [0]10)011 (4.138)
p 0 (t) = pli3 a (t) = exp [Ct] la)(31. (4.139)

Our strategy is to investigate the probability density in position space

P(X, t) (XIPS(t)1X) (4.140)

= Ar2 [pace (X, t) + poo(x,t)+ pc,o(x,t)-F po,(x,t)],
involving the matrix elements

rocta(x)t) = (x1Pcycy(t)lx), POO(x,t) = (x1/9,30(t)lx), (4.141)

po(x,t) = (xlp ao(t)lx) = A(x,t), (4.142)
 THE DAMPED HARMONIC OSCILLATOR 247

FIG. 4.4. The probability density p(x , t) according to eqns (4.143), (4.160) and
(4.161). The x-coordinate increases from left to right and time increases from
back to front. The picture shows the time evolution over 1.5 periods of the
oscillator. The initial distance between the centres of the superposed wave
packets is equal to 12a0 . Parameters: 27r-yo/wo = 0.05 and N = 0.5.

and to determine the decoherence function from the reduction of the interference
terms occurring in the expression for p(x , t). In fact, it will be shown that p(x , t)
can be written as
p(x , t) = (4.143)

.A1‘2 [pa,a (x, t) + A@ 0 (x ,t) + 2 N/p„(x, Opoo(x, t) exp[F(t)] cos (p(x ,t)1 .

The density p(x , t) exhibits the typical structure of an interference pattern shown
in Fig. 4.4. The first two contributions in eqn (4.143) represent an incoherent sum
of the superposed wave packets, while the third term describes the interference
pattern. As will be seen, the coupling to the environment yields a time and
temperature dependent modulation of the pattern given by the phase cp(x, t) , as
well as a reduction of the interference contrast (or the visibility of the pattern)
which is determined by the factor exp F(t) in eqn (4.143). Figure 4.4 shows a plot
of the function p(x I t) , where we use the analytical expressions for r(t) and (p(x ,t)
which will be derived below. The figure shows the decrease of the interference
contrast with increasing time, while damping effects are still negligible over the
time interval shown in the figure.
The task is now to determine the quantities defined in (4.141) and (4.142).
This can be done, of course, by determining the solution of the master equation
248 DECOHERENCE

(3.307) corresponding to the initial condition (4.119). However, we are going to

employ a more direct method which turns out to be useful for the treatment of
other models as well. The key point is that p„(x, 0), poo(x, 0) and po(x, 0)
are Gaussian functions. Since the system considered here is linear this Gaussian
property is preserved under time evolution. We thus conclude that the quantities
defined in (4.141) and (4.142) must take on the form of Gaussian functions of x
and all we have to do is to evaluate their mean and their variance. For example,
the function p„(x,t) may be written

p„(x,t) =
1
exp [
(x — x a (0) 2 ] , (4.144)
V27ro-,i (t) 2o-,i (t)
where

x(t) = f dxxp„(x,t) = trs {xPace(t)} = (alx(t)la),

and

(7 (t) = (alx 2 (t)la) — (04(010) 2 .

Here, x(t) and x 2 (t) are Heisenberg picture operators which have to be deter-
mined from the adjoint master equation (see below). Of course, analogous ex-
pressions hold for poo(x,t). Note that both p„(x,t) and poo(x,t) are correctly
normalized to 1. Finally, the function pao (x, t) must be written as

pco x, t) = (Ma) exp r (x — xao(t)) 2 1 (4.147)

V2 1- 0-20 (t)
(

[ 2o-2c(t)
i 0 , 7
with
1
x(t) = (sloe) f dxxpo(x,t) = (lx(t)a)
(fila) , (4.148)

and

a 2 (t) (S1 X2 (t)l a) (/3X (t) l a) 2 (4.149)

(M a) (fil a) 2 •

Note that (4.147) is a complex-valued Gaussian function and that it is correctly

normalized to the overlap of the initial coherent states,

f dxpo(x,t) = trs {p(t)} = (01a). (4.150)

Our next step is to calculate the mean values and variances introduced in the
above formulae. To this end we solve the adjoint master equation with the initial
 THE DAMPED HARMONIC OSCILLATOR 249

condition that at t = 0 the Heisenberg and Schrödinger operators coincide. This

yields
1
x(t) = ( e —"t a + e iw° tat) e —Y0t/2 (4.151)
V2mwo
1 ( e -2icu o t a2 e 2icu0t a f 2
X2 (t) = 2ata) e --" t
2mwo
1
+ [2N (1 — e --Y° t ) ± 1] • (4.152)
2 mwo

where at, a are Schrödinger operators and N =[exp(c.Jo lkBT) 1 ]'. With the
help of these relations the desired quantities are easily found to be
1 'Yot/2
x(t) = (4.153)
V2mwo (e—i'"ta eiw°ta*) e—
1 '
x o (t) = e iccots*) _---yot/2 (4.154)
V2mwo (e—iw t
x(t) = .v2212
1 coo +e t/*) e —t/ 2 , (4.155)

o-2 (t) (t) = o-20 (t) = o-20 (t) = u [2N(1 e —Y° t ) +1] . (4.156)

We observe that the variances introduced are equal to each other and that o-2 (0)
equals the initial width of the wave packets,

u2(0) = 0.? 1
(4.157)
2mwo •
As our final step we substitute the expressions found for p„(x,t), poo(x,t),
and p,o(x,t) into eqn (4.140). After a little rearrangement one obtains the de-
coherence function

F(t) = R [(x — x, o (t)) 2 — (x x,(t)) 2 — (x — x fi (t)) 2 ]

2o-2(t)
(4.158)

and the phase

{
(x, t) = (x—x, o (t)) 2 —(x (t) ) 2 (x x0(t)) 2 }] •
20.21(t) —

[S*a
(4.159)

If we insert here the relations (4.153)—(4.156) we finally get

Ap(t) • x
(x, t) = (4.160)
2N (1— e —yot) + 1 + Yo(t) ,

and
250 DECOHERENCE

1 2N + 1
F(t) = (1 — e —"fc' t ) . (4.161)
2N (1 — e --rot) + 1
The quantity yo(t) in eqn (4.160) is a space-independent phase given by

1 e —l'ot
Sao (t) = Tm [fi*cv (e —iwt) ta + (4.162)
2 2N (1 — e —Yot) + 1
while

AO) E (CdP(t)la) (01P(t)10) (4.163)

irtwo e —yo t/2 {( e —iwo t a _ e iwotal _ ( e - 0t
e iw0t)(3*)]
2
is the difference of the average momenta of the superposed wave packets.
These equations describe the modulation of the phase and the reduction of
the interference contrast through the interaction with the environment (Savage
and Walls, 1985a). For -yo t < 1 we have from eqns (4.160) and (4.162)

1
cp (x, t) Ap(t) • s — .(1'' [e- 2iwota2 e2iwot 1:3*21j (4.164)
2
This expression describes the usual interference pattern as it occurs for the case
of a vanishing system—environment coupling. Equation (4.161) shows that r(t)
reduces to expression (4.129) in the limit of zero temperature (N = 0). In the
long time limit expression (4.161) approaches the temperature-independent value
given by the overlap of the initial states of the superposition,

F(t) - $12, 'YOt (4.165)

For small times, -yo t < 1, we find

1 1 wo ) 70t.
F(t) R-- --la - 01 2 (2N + 1)-y0 t = --la - 01 2 coth (4.166)
2 2 (2kBT
Thus, the ratio of decoherence time to relaxation time becomes
2
TD
— =- (4.167)
TR — 01 2 (2N + 1) .
Comparing this expression with eqn (4.131) we see that in the presence of thermal
noise the decoherence time TD is reduced by an additional factor of 1/(2N + 1).
Let us assume that a and fi are real. This means that the initial momenta of
the superposed states vanish and that

(a 0)2
(4.168)
4ag
 ELECTROMAGNETIC FIELD STATES 251

where

Ax E X a (0) — x0(0) (4.169)

represents the initial space distance between the centres of the superposed states.
We then find with the help of eqn (4.167) in the high-temperature limit, kB T >>
hwo ,
-, 2
TD 2 hwo Ath )
(4.170)
TR la — M 2 2kB T (

Here, we have included factors of h and in the second equation we have intro-
duced the thermal wavelength Ath = h/V2772,k1371 . The same result has been
obtained previously from the high-temperature Brownian motion master equa-
tion (compare with eqn (4.79)).

4.5 Electromagnetic field states

Having discussed some basic features of the theory, let us take a look at experi-
ments on decoherence. Present day experimental technology is not only capable
of investigating and controlling decoherence phenomena, but also to resolve them
in time as dynamical processes and to verify results of the theory in quantitative
terms. For example, it has been shown to be possible to prepare a superposition
of two motional states of the centre-of-mass coordinate of a Be+ ion stored in
a Paul-trap and to measure its decoherence in controllable environments (My-
att et al., 2000; Turchette et al., 2000). Moreover, experimental evidences for a
coherent superposition of two macroscopically distinct magnetic-flux states in a
superconducting quantum interference device have been reported by Friedman
et al. (2000).
In another famous experiment, which was performed by Brune et al. (1996),
the progression of the decoherence of a mesoscopic superposition of two coherent
field states in a high-Q cavity has been observed for the first time. The decohering
field state observed in the experiment was of the form of the Schrödinger cat
state (4.117). In the present section we wish to discuss this experiment in some
detail. For our discussion the most important feature of the experiment is that
it enabled a direct measurement of the decoherence function F(t) and, thus, a
direct comparison with the theoretical analysis (Maitre et al., 1997).

4.5.1 Atoms interacting with a cavity field mode

The experimental setup is sketched in Fig. 4.5. With a variable time delay T, two
atoms A 1 and A2 traverse the setup which consists of two microwave resonators
R1 and R2 and a superconducting microwave cavity C with cavity frequency v.
The high cavity quality factor of Q = 5 • 107 leads to a cavity relaxation rate of
-yo = (160 us) — '. The atoms are described by two circular Rydberg levels (with
principal quantum numbers n = 51 and n = 50) which will be denoted by le)
and Ig), respectively. The corresponding transition frequency w is detuned from
252 DECOHERENCE

De1 9

IT» TR TRC TC TCR TR TRD

Time

FIG. 4.5. Schematic setup of the experiment performed by Brune et al. The
atoms A i and A2 move along the dashed horizontal line with a time delay T
and cross the resonator R 1 , the cavity C, the resonator R2 and detectors De
and Dg . The times of flight between and within the different components are
also indicated.

the cavity frequency by an amount A. The resonators Ri and R2 are fed by the
same classical microwave source of frequency coR . Finally, the state of the atoms
is analysed in field ionization detectors De and Dg . The atomic level scheme is
depicted in Fig. 4.6.
To begin with let us briefly describe the basic idea underlying the experiment.
The first resonator Ri serves to prepare the atoms in a certain superposition of
the states le) and lg). Initially, the cavity C contains a small coherent field la) as
shown in Fig. 4.7. As will be demonstrated below, the interaction of the atoms
with the field in C effectively induces a phase shift of ±0 on the field state whose
sign is triggered by the atomic state. Thus, the interaction of the first atom A1
with the field in C leads to an entanglement between the two atomic states
and the two phase components la exp(±i0)). The second resonator R2 induces a
further mixing of the states of A i such that the final measurement of the atomic
states in the field emission detector De and Dg does not give any information on
the state in which A i has passed cavity C. As a result, the state measurement on
A i projects the field in C onto a Schr6dinger cat-type superposition of two phase
components which are separated by an angle 20 in the complex plane (see Fig.
4.8). Ignoring for a moment field damping, normalization factors, and further
phase factors this state essentially takes the form (for details see the following
subsection)

+ (4.171)
where xi = 0 if A i was found to be in the state Ig) and x i = 7r if it was found
to be in the state le).
A similar transformation is then induced by the interaction of the second
atom with the cavity field. The final state of the cavity field is thus
220) + ei(xi+x2+20) (4.172)
lac- lae2içb ) + (e ix1 + 64'12 ) ekb
 ELECTROMAGNETIC FIELD STATES 253

••• 4 ..... • 1n,

.•

I e (71 = 51 , / = m = 50)

li WR 427 = 51.099 GHz

1g) (n 50 , / m = 49)

FIG. 4.6. Two-level scheme displaying the atomic transition frequency w, the
frequency v of the field in the cavity C and the frequency c.oft of the fields in
the resonators Ri and R2.

where x2 again takes on the values 0 or 7r, depending on the outcome of the
measurement on A2. Thus, if both atoms are found to be in the same state
(x i = x2 ) the field state (4.172) becomes (see Fig. 4.9)

lac -2i° ) + c2ilac2i°) ± 2c0),

while it takes the form

lac -21° ) - c2i° lac2° )

if the atoms are found to be in different states (xi x2). We see that the
interaction of the field with both atoms leads to some kind of constructive inter-
ference of the phase component la) if both atoms are detected in the same state.
If the atoms are detected in different states the phase component la) interferes
destructively and disappears from the final state.
In the experiment the conditional probabilities W„, are measured. These
are defined to be the probabilities of finding atom A2 in state lE1 ) under the
condition that atom A i has been found to be in state Is), where s = e, g. The
detailed analysis reveals that in the absence of any field damping the probability
of finding both atoms in the same state (i.e. to have constructive interference)
is larger than that of finding them in different states (destructive interference).
If field damping and, therefore, the decoherence of the Schrödinger cat state
(4.171) is taken into account this difference of conditional probabilities decays
to zero for increasing delay time T. Thus, by measuring a certain difference of
conditional probabilities as a function of the time delay T between the two atoms,
it is possible to measure the decoherence of the Schn5dinger cat state.
Before presenting the detailed theoretical analysis of the experiment in the
next subsection, we investigate the dynamics induced by the interaction of the
atoms in the central cavity C. The atom-field interaction in C may be described
by the Jaynes-Cummings Hamiltonian,
254 DECOHERENCE

4 0.65

0.6

2
0.4

10.2

o
-2
-0.2

- 0.4
-4 - 0.45
-4 -2 2 4

FIG. 4.7. Schematic representation and Wigner distribution of the initial coher-
ent field in the central cavity C for a real and a2 -= 6. The quadratures x
and p have been scaled such that s = Re a and p = Tm a.

1
Hjc(t) = — coo-3 + vat + 1-2(t) [au+ + ata_]. (4.175)
2

Note that the Rabi frequency SI(t), providing the coupling of the field mode a to
the atomic raising and lowering operators u+, depends on time. This is due to
the space-dependence of the mode function, given by a Gaussian envelope, that
describes the field mode in C. When traversing the cavity the atoms thus feel
a time-varying coupling to the field mode and the Hamiltonian Hjc (t) depends
parametrically on time.
 ELECTROMAGNETIC FIELD STATES 255

2
0.4

o_ 0

-2

-4
- 4 -2 2 4

FIG. 4.8. Schematic representation and Wigner distribution of the superposition

of coherent fields in the cavity C after the first atom has crossed C and was
detected in the ground state. Parameters: a 2 = 6, 0 = 0.7 and cpo = 0.

For a fixed time, that is for a fixed position of the atom within the cavity,
the dressed energy eigenvalues of the Jaynes—Cummings Hamiltonian are given
in linear approximation by

1) A S1 2 (t)(n, + 1)
= v 2(n + — + —
2± A , (4.176)

with the corresponding dressed states (we consider without restriction the case
of positive detuning, A w — y> 0)
 I 0 45

04

-2

- 4
- 4 - 2 o 2 4

FIG. 4.9. Schematic representation and Wigner distribution of the superposition

of coherent fields in the cavity C after detection of the second atom in the
ground state. Parameters: a' = 6, 0 = 0.7 and po := 0.

+,rt) = 1). (4.177)

Under the given experimental circumstances the off-resonant atom—field inter-

action in C is predominantly adiabatic, which means that real transitions be-
tween the dressed atomic states may be neglected and that only virtual pro-
cesses must be taken into account. Applying the adiabatic theorem of quantum
mechanics we then see that the dressed states of an atom that traverses the
cavity take up a phase factor which is given by the integrated dynamical phase
 ELECTROMAGNETIC FIELD STATES 257

exp[-i f dtE±,n (t)]. Thus, in addition to the contribution from the unperturbed
energies, the interaction of the atom with the field mode a leads to the phase
shifts given by
—> exp(ic,o e )e, n) , g, n) —> exp(icpg )Ig,n), (4.178)
where

S-22 (t)(n + 1)1

exp(ip e ) = exp dt exp [0(n + 1)] ,
A
-00

ç2 (t) m 1
exp(icpg ) = exp ±[-i f dt = exp[-iOn] , (4.179)
A
- 00

and
+00
E f dt-
92(t)
A
(4.180)
-

Accordingly, we may write the transformations of the states e)la) and 1g)la) in
the form
I ea) eXp (ata + 1)] le)Ice) ezçble)laekb), (4.181)
g)ct) exp {-i0ata]Ig)la) = g)Icte — z` ) (4.182)
This result will be used in the next subsection to describe the preparation of the
superposition of field states whose decoherence is studied in the experiment.
4.5.2 Schradinger cat states
We now turn to a detailed analysis of the experiment. Our man intention is to
derive a formula for the difference W6 — Wge of the conditional probabilities
measured in the experiment and to relate it to the decoherence function pertain-
ing to the Schrödinger cat. The experiment may be described as a succession of
seven steps as follows.
4.5.2.1 Initial state of A1 and interaction in R1 First, atom A 1 is prepared in
state le) and undergoes a 7r/2-pulse in resonator R 1 , yielding the following state
of the atom-field system prior to the interaction in C,
1
(le) + 19)) la). (4.183)

4.5.2.2 Atom field interaction in C Atom A 1 enters the central cavity C. The
-

field mode in C induces a phase shift which can be described by the transfor-
mations (4.181) and (4.182) such that the state (4.183) is transformed into the
state
1 . .
(eiçb ncte 4) ) +19)1cve -4b )) • (4.184)
 ▪ ▪
258 DECOHERENCE

4.5.2.3 Interaction in R2 and measurement on A 1 The second 7/2-pulse in

R2 induces the transformations
1 1
le) —> (l e) + e7470 19)) 19) i
( - e -i7° le) + 19)) (4.185)
vz vz
of the states of atom A 1 , where cpo -= (R -4r is the dynamical phase difference
acquired during the time of flight T between R1 and R2. Thus, the atom-field
state takes the following form after atom A 1 has left resonator R2,
1r /
(6 ) + e (0) e)) e 0 e) (la ç() ) ei(çb-H°°) Ietei ))] •
Llg)
(4.186)
The measurement of the state of atom A 1 projects the field state onto the states

0) + e i(x+0 + 4,0 o )1 i0)) (4.187)

1 (lae
Ar(X)
where the measurement of g) implies x = 0, while the measurement of 1e) means
x = 7r. The normalization factor is given by

A((X) = -V2 (1 + exp [-21a1 2 sin2 0] cos [x + + + 1a1 2 sin 20]). (4.188)

4.5.2.4 Field damping in C The state (4.187) corresponds to the density ma-
trix
1
pF(0,x) = {Ict i(6)(ae i(61 Ictei(4)(cteiç61 eilaei(6)( 14'1 h.c.]
A[2 (x) -

(4.189)
with = X + cpo + 0. Here and in the following h.c. means that the Hermitian
conjugate of the last term must be added to the expression. We denote the
time delay between the first and the second atom by T. In the experiment the
temperature of the environment corresponds to a mean number N = 0.05 of
thermal photons in the field mode, such that we may use the vacuum optical
master equation (4.120). During the time interval T the field mode thus evolves
into the density matrix
1
pF(T,x)= [1a(T))(a(T)1 10(71))(0(T)1 f(T)1[3(T))(a(T)1±h.c.],
Al2 (x)
(4.190)

where (compare with eqns (4.125)-(4.127))

a(T) = exP( - NT/ 2 - (4.191)

fi(T) = exP( - NT/ 2 + 45), (4.192)
f (T) = ez exp [-1a1 2 (1 - e 2 ) (1 - e -110T )]
=e k(cee -i(licte i0 ) (1- e - "°T) . (4.193)
 ELECTROMAGNETIC FIELD STATES 259

Note that dynamical phase factors are absent since we are working in the inter-
action picture. We see that the decoherence function is given by

F(T) = 1n1f (T) = -21a1 2 sin2 0 (1 - e - " T ), (4.194)

which leads to a decoherence time of the form

TD = (4.195)
2 1 (1 2 sin2 0 .

4.5.2.5 Initial state of A2 and interaction in R 1 Atom A2 is prepared into the

initial state le) and subjected to a 7r/2-pulse in R1 such that the state of atom
A2 and the field, just before the atom enters the cavity, is given by
1
- [10(6 + 1001+ 10(g1 + 19)(ell PF(T, X). (4.196)
2

4.5.2.6 Atom field interaction in C

- The interaction of atom A2 with the field
in C yields the state
1 e x e l e i0 (at a+1) p F (T1 x ) e —ic/,(at a+1) g ) (g l e— io at a pF x)e ioat a

+ e x g e ic/y (a t a+ 1 ) pF (T, x ) e i0a t a + h.c] . (4.197)

4.5.2.7 Interaction in R2 and measurement on A2 Atom A2 traverses reso-

nantor R2, which transforms the state (4.197) into the state

1 . (T, x)e —io(ata+i)

PAF(T, X) = -4 (le) + eiv019)) ((e1+ e -iv0 (g1) eio(ata+1 ) pF
1 (e)
+ 4— (19) — e—" le)) ((gl — e 0 a pF(T X)e zcl' at a
1
+ -4 (1 e) + e w0 19)) (01 - v° (el) e 45(ata+1) pF(T,x)e t a
1 ioata pF(T, x)e _0( a t a+i ) .
+ -4 (19) e° e)) ((el + e° (g) e -

Finally, the state of atom A2 is measured and the conditional probabilities Wee
and Wge are determined. Here, WE , is defined to be the probability of finding
atom A2 in the state le) under the condition that atom A 1 was detected in the
state 1s), where E = e, g, depending on the outcome of the measurement on atom
Al .
According to the obtained expression for PAF we now have

We = trAF {1e)(e1PAF(T X)} (4.198)

= El — e z( v° ±(6) trF e2zç'at a PF(T,X)}] •

In the first expression the trace is taken over atom A2 and the field mode, while
in the second expression the trace is taken over the field mode only. Furthermore,
260 DECOHERENCE

we have to set x = 0 if E = g and x = 7r if E = e. With the help of the field

density matrix (4.190) we find

1
trF { e2ioata pF (T, x ) = Ar2(x) [(a(T)1a(T) exp 20) + (T)0(T) exp 245)

f* (T)(,3(T) (7') exp 20)

+ei V(T)(a(T)10(T) exp2i0)] (4.199)

which yields

1 1 2B+Ccosx+Dcosi
— [1 (4.200)
2 1 + A cos x 2
where

A = exp [-2 al 2 sin2 (1)] cos Roo + + laI 2 sin 2ç], (4.201)
r
B = exp [ 2.ct12 e — Y oT s•n2 pi cos L(PCI (/) + a 2 e - T sin 20] ,
- (4.202)
0 T) sin 2 0] cos [1 (1 12(i _
C = exp [-2la1 2 (1 — e l'° T ) sin 20] , (4.203)
-

D = exp [-21ct2e--y0T sin2 20 _ 2 k:1 12 (1 _ c --yo T ) sin 2

x cos [2,00 + 20 +1(11 2 e —Y° T sin40+ 2 (1 — e ----Y°T ) sin 20] . (4.204)

The term 1 + A cos x in eqn (4.200) stems from the normalization factor of
the field density matrix, while the quantities B, C, and D arise from the four
scalar products that contribute to the trace in eqn (4.199). Namely, the term B
is due to the product

(a(T) a(T)exp2i0) = (i3(T)10(T)exp2i0) = (o(T)8(T))

= exp {-1a 2 e -11° T (1 — e2 )]. (4.205)

Thus, B is determined by the overlap of the two phase components of the original
superposition created by the first atom. These phase components are separated
by an angle 20 (see Fig. 4.8). The contribution C stems from the product

(i3(T)la(T)exp 2i(/)) = ([3(T)1 /3(T)) = 1, (4.206)

whereas D is determined by the product

(a(T)10(T) exp 2i0) = exp (1— e4i0)] (4.207)

The term D is therefore determined by the overlap of two phase components

which are separated by an angle of 40, while C is obtained from the overlap
of two phase components corresponding to the same angle. The term C thus
describes the contribution from the recombination of the phase components at
the location of the original coherent field state la) (see Fig. 4.9).
 ELECTROMAGNETIC FIELD STATES 261

0.80

0.70

0.60

0.40

0.30

0.20 - 2 exP( - y0 1 ) -
0.10

0.00
-0.10
00 0.5 1.0 1.5 20
-y oT
0.80

0.70

0.60

0.50

0.40

0.30
1
0.20 _ 2 exp(--yon
-

0.10

0.00

-0.10
00 0.5 1.0 1.5 20
y oT

FIG. 4.10. Top: The conditional probabilities Wge and W„, their difference
n and the probability WeD as a function of the delay time T for (po = 0,
,

= 0.7 and a 2 = 6. To make visible the difference between decoherence

and relaxation of the photon number we also show the function 0.5 exp[-NT].
Bottom: The same quantities averaged over the angle (Po.

It may be seen from the relations (4.201)-(4.204) that A, B and D are expo-
nentially small in la12 for NT < 1, provided the angle 0 is not too close to 0,
7r/2, or 7r. Thus we have

1[ 1
TI
ree - 1 - - C] . (4.208)
2 2

In the experiment the following difference of the conditional probabilities is de-

termined,
262 DECOHERENCE

n(71 , WO) = Wee ge 1 (4.209)

or rather the angle-averaged one
ir
1
7r f (490 17( 71 ) WO)*
11(71) = — (4.210)
o
With the help of eqn (4.200) one finds

1C+D-2AB
11( 71,490 = 2 (4.211)
1 — A2
Since C represents the dominant contribution in ri(T, (p 0) and since it is inde-
pendent of (po we find that ri(T, (p 0) P-.,' fi (T)s-.,, C/2. The expression (4.203) for
C shows that it is equal to the real part of the quantity f (T) (see eqn (4.193))
which multiplies the interference term of the Schrödinger cat state (4.190). Hence
we get

71(71 PO
, (

1
Ref (T)
1
= — exp [r(T)] cos
2 [H 2 ( 1 — e —Y° T ) sin 20] . (4.212)

This is the desired expression which relates the experimentally observed quantity
fi(T) to the decoherence function F(T).
We show in Fig. 4.10 the conditional probabilities Wge and Wee and their
difference 17 as a function of the delay time T. For comparison we also depict the
corresponding angle-averaged quantities, the function exp[—NT]/2, and

WeD = —21 (1 — B) , (4.213)

which is defined to be the probability of detecting A2 in the excited state under

the assumption of complete decoherence, that is under the condition that the
interference terms of the field density matrix are put equal to zero.
According to eqn (4.212) the difference ri of the conditional probabilities is
directly related to the relevant off-diagonal element of the field density matrix
and thus also to the decoherence function F(t). As was demonstrated in the
experiment performed by Haroche and coworkers this relation enabled the direct
observation of the decoherence of the Schrödinger cat and a nice quantitative
verification of the theory. A more detailed theoretical analysis of the experiment
was given by Breuer, Dorner and Petruccione (2001).

4.6 Caldeira—Leggett model

Up to now we have discussed decoherence and dissipation for Markovian dynam-
ics defined by the quantum optical or by the quantum Brownian motion master
 CALDEIRA-LEGGETT MODEL 263

equation. The use of these equations presupposes the weak coupling and/or the
high-temperature limit to be valid. In order to investigate effects from non-
Markovian dynamics and from strong system—environment coupling we study
here the destruction of quantum coherence arising in the full Caldeira—Leggett
model for the damped harmonic oscillator discussed in Section 3.6 (Caldeira and
Leggett, 1985; Unruh and Zurek, 1989).

4.6.1 General decoherenee formula

To determine the decoherence function F(t) corresponding to the superposition
of two Gaussian wave packets la) and 1 13) we apply the same technique as in
Section 4.4. A simple consideration reveals that eqn (4.158) can also be used for
an exact treatment of the Caldeira—Leggett model for arbitrary spectral densities,
temperatures and coupling strengths. The reason is that the derivation for this
equation relies solely on the Gaussian property of the initial states and of the
propagator function. Thus, eqn (4.158) holds also in the general case with the
only modification that the quantities x,(t), xo(t), x o (t) and a2 (t) are now
defined in terms of expectation values of the exact Heisenberg picture operator
x(t) through
XŒ(t) = tr fx(t)la)(a1PB} ) (4.214)
x(t) = tr fx(00)(MPB} , (4.215)
tr fx(t)la)(filPB}
x(t) = (4.216)
(0 cE) '
0-2 (t) = tr fx 2 (t)a)(ctIPB } — [tr -{x(t) a)(a PB 1] 2 ' (4.217)
Here the trace is taken over the total system. We also note that o-ci (t) = o-20 (t) =
o-c,2 0 (t) E a 2 (t), as in the case of the quantum optical master equation. We may
therefore determine the variance a 2 (t) with the help of the state a), for example.
It is assumed for simplicity that initially there are no system—bath correlations.
A similar technique may be used, however, if correlations are present in the initial
state.
To determine the decoherence function we first express the quantity a — V
used in eqn (4.158) in terms of the initial separation of the superposed wave
packets in position space,
Ax E X 0 (0) - (4.218)
and in momentum space,
Ap E Tri±0 (0) - 7/10 (0). (4.219)
This yields the identity
2 _ Ax2 + 4442
Vi-13 l (4.220)
4o-g
We know already that the Heisenberg picture position operator x(t) and the
corresponding momentum operator p(t) obey the equations of motion p(t) =
264 DECOHERENCE

m(t) and (3.465). As in Section 3.6.3.2 we introduce the fundamental solutions

G 1 (t) and G2(t) of the homogeneous part of eqn (3.465) satisfying G 1 (0) =
d2(0) = 1 and G 1 (0) = G2 (0) = 0 and write the Heisenberg operator x(t) as

X(t) = G i (t)x(0) + G2 (t)±(0) 7

,1 1n f d8G2 (t — s)B (s)
E x h(t) + 1(t), (4.221)
where xh(t) denotes the solution of the homogeneous equation, while /(t) is a
solution of the inhomogeneous equation satisfying 1 (0) = 1 (0) = O.
With the help of eqn (4.221) it is now an easy task to evaluate the required
quantities. We find that
1
[(x - (t)) 2 - -21 (x - x,(t)) 2 - -2 (x - xo (t)) 2 1

= - /3 1 2 (crôGi(t) + 4Grn%.t)g ) (4.222)

and that the variance is given by

G(t) 2
u2 (t) = c)-(3GT (t) (4.223)
47722 1702 ± (t))'

where
t t
(12 (t)) = 2m
1 2 f ds f ds' G 2(t - s)G 2 (t - ) D (s - s'), (4.224)
o o
00
2
1

We recall that
= f d o (w) coth )( 2 :BT
c
o I dsG 2 (s)e i w

00
CA)
D (S — = 2 f dw J (w) coth ( 2kBT ) cos w(s - s') (4.225)
o
is the noise kernel expressed here in terms of the spectral density J (w) of the
underlying model.
Substituting eqns (4.220), (4.222) and (4.223) into (4.158) we arrive at

r(t)
Ax2 + 40-g Ap2 (12 (t» (4.226)
80-g (12(0) + 0-gGT(t) + G(t) I 4m 2
This equation provides a general expression for the decoherence function F(t). It
may be used for all linear models with coordinate-coordinate coupling, involving
 CALDEIRA-LEGGETT MODEL 265

Gaussian initial states, arbitrary coupling strengths and spectral densities. In eqn
(4.226) the decoherence function has been expressed in terms of the initial width
a-F, of the superposed wave packets. If ug is taken to be related to the oscillator
frequency through eqn (4.157) the superposed wave packets represent coherent
states of the oscillator. However, since ag may be chosen arbitrarily eqn (4.226)
is also valid for squeezed initial states.
Obviously, r(0) = 0 and r(t) tends to the value given by the initial overlap
in the long time limit provided (12 (t)) » o-gG21 (t) + G3(t)/4m 2 o-ci in this limit.
Moreover, we have the following limits
2 ( -12 (t)) 2
0, (4.227)
2G(t)/m2
F(t) (4.228)
2GT (t)
The problem of determining the decoherence function I'm is thus reduced to the
determination of the fundamental solutions G 1 (t) and G2 (t) of the homogeneous
part of the Heisenberg equation of motion, and of the reservoir average (12 (t))
of the square of the inhomogeneous part.
4.6.2 Ohmic environments
Let us discuss in some detail the case of an Ohmic spectral density J(w) =
2m/ywe(1 — w)/7r with some cutoff frequency ft
4.6.2.1 High-temperature limit In the high-temperature limit, that is in the
case 2k13 T » I » wo,'Y, we get (see eqns (3.462) and (3.463))
D i (s — s') 8rrrykBTS(s — s'), (4.229)
which gives, by virtue of eqn (4.224),

(I2(0) 47rn
kB T
f dsG(8). (4.230)

To give an example we study free Brownian motion for which we have the fun-
damental solutions G 1 (t) = 1 and G2 (t) = (1 - exp(-27t))/27. This leads to
kB T 1
(1-2 (t)) Eyt — [1 — exp( —270] + — [1 — exp(-47t)]) . (4.231)
17/7 4
Substitution of these relations into eqn (4.226) yields the decoherence function
for high-temperature free Brownian motion. Let us investigate times such that
2-yt < 1. Introducing A = 2m-ykBT we find
2A +3
AX2 ± 40-g AP2 3Tn 2V

2A t3 u2 t2
(4.232)
8ci-g 3m2 0 4m 2
'

This shows that the behaviour of the decoherence function depends crucially
on the initial width of the superposed wave packets. For example, if the noise
266 DECOHERENCE

contribution to the variance o-2 (t) is small compared to the initial width cig of
the wave packets and if the free spreading t2 /4m2 o-g may be neglected, we get

Ax2 + 4o- p2 2A
F(t) 3m2u2 t3 1 (4.2331
8o-g
showing that the magnitude of r(t) increases as the third power of t.
In the limit of vanishing initial width (see eqn (4.227)) eqn (4.232) leads to

AAx2 t. (4.234)
3

Apart from the factor this relation was already used in the estimation of the
decoherence rate in Section 4.3.1. On the other hand, if we let the initial width
tend to infinity (see eqn (4.228)) we get

1 A
F(t) - - Ap2 t3 . (4.235 ,
3m
Again, the magnitude of the decoherence function grows with the third power
of time. This result corresponds to the case of an interference device involving
plane waves which has been discussed by Savage and Walls (1985 b).

4.6.2.2 Harmonic oscillator The homogeneous part of the Heisenberg equa-

tion of motion,

+ + 2-)4 = 0, (4.236

is easily solved to yield the fundamental solutions

G 1 (t) = [ 2 sin ut + cos vt] e t, (4.237

1
G2(t) = — sin vte -14 (4.238 r

where

= VW°2 - 7 2V , < w0, (4.2391

is the characteristic frequency in the underdamped case. In the overdamped case

v becomes imaginary and we write

(4.240'

such that the fundamental solutions take the form

1 (t) = [ 2_ siG nh It + cosh r)t] e,

G2 (t) = sinh V- te -7t (4.242,

 CALDEIRA LEGGETT MODEL
- 267

We use the spectral representation of the noise kernel (4.225) and carry out the
s-integration in eqn (4.224) to get
+Q
(12 (t)) = 7 f &ow coth ( 2 :137,) h(c.o), (4.243)
2 7 m1v1 - -0

with the function

2
(4.244)

These expressions are valid for both the overdamped and the underdamped case.
Inserting eqn (4.243) and eqns (4.237), (4.238) or (4.241), (4.242) into (4.226)
yields the decoherence function for the harmonic oscillator (Caldeira and Leggett,
1985). The obtained expression is valid for arbitrary coupling strengths and
temperatures. It turns out, however, that a general analysis of the frequency
integral in eqn (4.243) is extremely difficult. In particular, the integral depends
logarithmically on St for large cutoff frequencies. However, simple statements
may be obtained for certain limiting cases. For the following discussion we set
al, = 1/2mw4 corresponding to an initial superposition of coherent states.
In the weak damping limit we have -y < wo and vP.,, wo . The function h(w) has
thus two sharp peaks at w -A.,-' ±wo and in the limit -y 4 0 we may approximate
—

( pm) ,.,..,., 7(2N + 1) +

27rmwo
f00 dwh(w), (4.245)
-00
where we have introduced the Planck distribution N = N (w o ) through the re-
lation 2N(wo ) + 1 = coth(w0 /2kBT). The remaining frequency integral may be
determined with the help of the method of residues which leads to
2N + 1
(1-2 (t)) P,, 2mw [1 - exp(-2-yt)] . (4.246)
o
Substituting into the general decoherence formula (4.226) and noting that
1 1
0-(3G(t) + 2 G(t) = 2mwo exp ( -2-yt) (4.247)
4m2o-0
in the weak damping limit we see, as expected, that the expression for the deco-
herence function reduces to (4.161) which was obtained in the quantum optical
limit. Note that the relaxation constants are related through -yo E 2-y.
Let us also discuss the high-temperature limit for arbitrary couplings. Em-
ploying eqn (4.230) yields

sin 2vt + 2-v2 sin2 ut

(I2 (t)) r-z,, kBT2 [1 - e -2-Yt (1 + - -- -
7 (4.248)
fftwO V
The corresponding high-temperature decoherence function F(t) is provided by
inserting this expression into the general formula (4.226). The result is valid for
268 DECOHERENCE

0.75 -

410

0.5 -
<

`Icl°
co
0.25-

0 -1 0
10 10

FIG. 4.11. Semilogarithmic plot of the underdamped high-temperature deco-

herence function I'm as given by eqns (4.226) and (4.248) (solid line), and of
the corresponding quantum optical result (4.161) (dashed line) as a function
of T = 2-yt. Parameters: 2kBT/tao0 = 10, -y/wo = 0.1.

both the underdamped and the overdamped case. As far as the underdamped
case is concerned we observe, as expected, that the exact decoherence function
differs from the quantum optical one by terms of order 7/wo. Figures 4.11 and
4.12 show I'mfor two different values of 7/wo and a comparison with the cor-
responding quantum optical result (4.161) in the high-temperature limit. The
exact decoherence function oscillates around the quantum optical one for short
times. With growing time these oscillations die out and I'm
converges to the
quantum optical limit.
The weak damping limit leads to

F(t)
Ax2 + 44,42 2kB T 2t
,y
(4.249)
8ag coo

for times satisfying 27t < 1 and (2kBT)(27t)/wo < 1. Accordingly, the ratio of
the decoherence time TD to the relaxation time TR = 1/27 is

TD
(4.250)
TR = AX 2 44,42 2kB T .

Let us compare this result with the strongly overdamped case. This case is defined
by the limit -y » wo such that i) 7. The decoherence function is determined
by substituting y = in eqn (4.248). Let us investigate times t satisfying
 DECOHERENCE AND QUANTUM MEASUREMENT 269

FIG. 4.12. The same as Fig. 4.11 but for 7/wo = 0.01.

w2
2(7 - ) 7t« 1, 2(7 + i))t 4-yt >> 1. (4.251)
7
The decoherence function for the strongly overdamped particle then becomes
Ax2 + 44,42 2kB T co g 27t (4.252)
sag Wo 272
We note that this expression differs from the weak damping result (4.249) by the
factor wg /2-y 2 . It must be noted, however, that the relaxation rate is TR = 1/27
in the weak damping case, while it takes the form
1 7 272
TR = r-
r"' = TR (4.253)
2(7 - f)) W0
in the strong damping limit. It follows that the strong damping decoherence time
I) satisfies

(4.254)
TR TR

Thus we arrive at the remarkable result that the quotient of decoherence time
and relaxation time is the same for the weak and for the strong damping case,
and that it coincides with the result (4.170) found in the quantum optical limit.

4.7 Decoherence and quantum measurement

The quantum theory of measurement provides an important application of envi-
ronment-induced decoherence. The destruction of quantum coherence through
270 DECOHERENCE

the influence of an environment singles out a specific basis set, known as the
pointer basis, in the Hilbert space of some quantum apparatus M which is em-
ployed to measure a quantum object S. Since decoherence, as it is understood
here, is ultimately linked to the tracing over degrees of freedom of the environ-
ment, it cannot, of course, solve the measurement problem. This means that
decoherence cannot be used to deduce the reduction of the state vector and the
statistical interpretation of quantum mechanics from the unitary evolution given
by the Schrödinger equation. However, in any realistic measurement scheme the
coupling of a macroscopic apparatus to its environment can be shown to lead,
under quite general physical circumstances, to a dynamical selection of a specific
pointer basis and, thus, to a unique definition of what is being measured by the
apparatus (Zurek, 1981, 1982).

4.7.1 Dynamical selection of a pointer basis

We investigate a quantum system S which is coupled to an apparatus or meter
M through an interaction Hamiltonian Hsm(t). The meter degree of freedom
will be described fully quantum mechanically such that the Hilbert space of the
combined system is the tensor product space = 'Hs m . We further denote
the interaction Hamiltonian in the interaction picture by

= e in t Hsm (t) e —tHot (4.255)

H M (t)

where Ho = Hs + Hill is the sum of the self-Hamiltonians of system and meter.

As in Section 4.1 the system—meter interaction is supposed to lead, in the ideal
case, to perfect correlations between certain orthogonal basis states Sn ) of S
and meter states Mn) such that the time evolution over a time interval T takes
the form (see eqns (4.7) and (4.8))

losm(0)) =E 0 1m) —4 losm(T)) = E cnds.) 1Mn). (4.256)

This is precisely the type of dynamics underlying an indirect, QND measurement

scheme: The system S represents the quantum object, while the meter M pro-
vides the quantum probe. The initial probe state is _it4-) and the quantity being
measured is

= ank9n)(Snl, (4.257)

or some function f () thereof. In eqn (4.256) it is assumed that the back-action

evasion condition holds,

[§, 11-L m (t)] = 0. (4.258)

 DECOHERENCE AND QUANTUM MEASUREMENT 271

This ensures, as we know, that the basis states 1Sn ) are unaffected by the system—
meter interaction. Let us suppose that the meter states are, at least approxi-
mately, orthogonal (see also the example treated in Section 4.7.2), such that we
have"

(511 1Sn' ) = (5 mm, (Mn1 114-m) = (5nrn• (4.259)

The application of the projection postulate to the meter quantity

b IMO (-114-711, (4.260)

n

then yields that the readout bn is found with probability 1c,n 1 2 and that the
system's wave function conditioned on this event is subsequently given by 1Sn).
Within the framework of an indirect measurement scheme we may thus say that
a measurement carried out on the meter system leads to a measurement of a
system's quantity f(§) and induces the reduction of the system's state vector
in accordance with the projection postulate. After the measurement the reduced
density matrix of S takes the form,

Ps (T ) = 1 2 Isn)(snl, (4.261)
n

describing the measurement on the non-selective level.

Although being strongly suggested by the decomposition of 17,bsm (T)) in
(4.256), the above interpretation is incomplete for the following reason. After
the system—meter interaction the combined system ends up in an entangled state
l'Osm(T)), describing perfect correlations between the system states 1 5n ) and the
meter states 1 Mn ). However, 1Ipsm (T)) is still a superposition of these correlated
states. The latter coexist in 1Ip sm (T)) and without application of the reduction
postulate to the meter system M there is no a priori reason of accepting only one
of the states 1Sn ) 0 1Mn) as physically real. In fact, without making a definite
decision on an observable being measured on M we could consider another set
of basis states,

liffn) Imnixmn , licio, (4.262)

n,

and write the state vector of the combined system after time T alternatively as

lO s m( T )) = nn ) 0 1-Tlin ), (4.263)
n

where we have introduced new normalized system states through

"To facilitate the following discussion we assume here that the interaction is non-degenerate,
i.e. that it leads to an indirect measurement of a non-degenerate system observable.
272 DECOHERENCE

znrs-7,,) E sn 1)(1Vin Hun o. (4.264)

n'
With the same right as before we could now claim that the measurement of the
meter observable

Tw= En -6-n l ivin x kn I (4.265)

leads to the measurement of the system observable

'--5-i = E and§n)(§711, (4.266)

n
or of some function f (S). The problem is that, although we did not change the
system—meter interaction and although the reduced system's density matrices
obviously coincide,

ps(r) = E lc„,12 ,sfn)(sn = E -J,,i2ls'n)(§ini, (4.267)

n n
the to-be-measured system observable is ambiguous since, in general, the quan-
tities § and §' do not commute,

[ 5 ,:'4-1 0 (4.268)

The question is thus: Do we have a measurement device which measures S— or :§?

Once a basis of meter states Mn ) is given, the corresponding states ISO in
a decomposition of the form (4.263) are denoted as relative states. They are.
in general, not orthogonal to each other but may be taken to be normalized
in which case they are defined uniquely up to a phase factor. The concept of
relative states has been used by Everett (1957) in his relative state formulation
of quantum mechanics. We may assume the En in eqn (4.264) to be real and
non-negative which yields

En = 2 (4.269)
E Icni1 2 1(-Tin Mn') 1

ni
and
1
ISO = cn (4.270)
n■

Hence, the scalar product of the relative states ISO is given by

(§4-§n) = _ ILcn nE, 1 C 12 071-

Cm
71,1 0 (Mn i I lVini)*
n1 111 (4.271)

We note that for non-orthogonal relative states eqn (4.266) still defines a self-
adjoint operator, where, however, the an are not its eigenvalues. Moreover, eqn
 DECOHERENCE AND QUANTUM MEASUREMENT 273

(4.267) represents the system's density matrix after the measurement in two al-
ternative ways as a mixture of pure states, where in the second case these pure
states are not orthogonal.
The above problem becomes particularly acute if all non-zero coefficients en,
n = 1, 2, ... , D, in the superposition of the initial state 10sm (0)) have the same
absolute value, that is if lOsm (0)) is maximally entangled. In this case ps (r) be-
comes proportional to the identity in the subspace spanned by the corresponding
states ISn ),

D
PS(T) = 1'971)(94 (4.272)
n=1

Assuming further that the corresponding basis states 1Mn ) and 1Mn ) span the
same subspace we immediately find with the help of eqn (4.271) that the relative
states are orthogonal, that is

(,§ni gn) = (5mn (4.273)

for any choice of meter basis states (with the above restriction, of course). By
varying the meter basis we can measure any system basis. Thus we are led to
the surprising conclusion that our measurement device is capable of measuring
any system's observable.
It is important to realize that the problem discussed above does not provide a
contradiction to the orthodox interpretation of quantum mechanics, for it arises
only if one refuses to make a definite decision on the meter observable and to
apply the reduction postulate. However, the situation is somehow dissatisfying
since according to experience a certain measuring device which has been designed
to measure, for example, the momentum of a quantum system, does measure
momentum, and not position. The ambiguity in the measured system observable
is obviously due to the fact that the system—meter interaction does not fix a
unique basis of states 1Mn ) in the meter's Hilbert space 1-im (or in a certain
subspace thereof). This ambiguity can only be avoided if for some reason a
specific basis is singled out, that is if only a specific physical quantity M =
En bn iMn)(Mrt (or some function of it) can be measured on M. The system
observable S measured by the device is then determined by the corresponding
set of relative states 1Sn ).
At this point it must be taken into account that the meter M is usually
assumed to represent a macroscopic degree of freedom, which, in turn, is coupled
to a huge (usually infinite) number of further degrees of freedom. Environment-
induced decoherence then leads to a dynamical selection of a specific basis. This
means that the meter's Hilbert space 7-im is decomposed into coherent subspaces
(which are supposed to be one-dimensional here) in such a way that a local
observer cannot observe coherences between different subspaces. Superpositions
of the form (4.262) are thus effectively destroyed and a unique basis {1Mn)}
274 DECOHERENCE

System Meter

:ow •A i,•••

FIG. 4.13. Schematic picture of the measurement scheme. In addition to the

system-meter interaction the meter is coupled to a large environment which
dynamically selects a specific pointer basis through decoherence.

emerges. This basis is often referred to as the pointer basis. The environment
E acts like a measuring apparatus: While the total system S + M + E evolves
unitarily, of course, the reduced system S+M behaves as if its state continuously
collapses into one of the correlated states 1S„)® Ma). The reduced state of Sd-M
then behaves, with regard to local measurements, as an incoherent statistical
mixture of the states 1 Sn) 0 M n) •

Thus, we consider a measurement scheme of the type depicted in Fig. 4.13.

The total interaction Hamiltonian in the interaction picture will be supposed to
be of the form
TT f\ = I-1,6(t) + (4.274)

Here, we assume for simplicity that HsmE = 0, that is there are no triple
interaction terms in the total Hamiltonian, and that HsE = 0, which means
that the direct interaction between system and environment is zero. The whole
measurement process can then be decomposed into two phases as follows (Zurek,
1981).
1. As before 1-/,6(t) acts over a time interval [0, 7 in which HM (t)
]

HfuE(t). During this time interval the desired system-meter correlations

are built up under the influence of HL,/ (t),

enIsn) 1m) 0 1E) E enIsn) 0 imn) 1E), (4.275)

where 1E) is the initial state of the environment. As before, the back-
action evasion condition (4.258) is assumed to hold, which guarantees that
the system states 1S„) are not affected by the system-meter interaction.
This first phase of the process may be called pre-measurement.
2. For t > T the meter-environment interaction dominates, that is HiviE (t) >>
11,6(0. This yields the dynamical selection of the pointer basis through
 DECOHERENCE AND QUANTUM MEASUREMENT 275

environment-induced decoherence. Assuming also the back-action evasion

condition for the meter-environment interaction,

[R, HLE(t)] = 0 , (4.276)

we may write the second phase of the measurement process as

enisn) mn) 1E) E cn ® IM) l En). (4.277)

We note that the most general solution of condition (4.276) reads

HL,E (t) =E (mn1 Bn(t), (4.278)

where the B(t) are arbitrary operators of the environment. Condition

(4.276) guarantees that the eigenbasis of M is not affected during the sec-
ond phase of the process. The specific system-meter correlations created
during the pre-measurement are thus not destroyed through the environ-
ment; they are still present in the reduced system-meter density matrix,

PSM (t) = E icni2 Sn )(S7,1 IM,,)(Mn I. (4.279)

This is, of course, only a schematic picture of what is going on in a realistic

measurement scheme. In particular, it turns out that in a more realistic descrip-
tion the ideal back-action evasion condition (4.276) cannot, usually, be fulfilled
exactly. The pointer basis can, however, be still approximately orthogonal, which
is fully sufficient for the measuring device to work. This point is illustrated with
the help of a specific model in the next subsection.

4.7.2 Dynamical model for a quantum measurement

Let us discuss here a dynamical model for a quantum measurement process which
has been introduced and investigated by Walls, Collett and Milburn (1985). It
consists of two coupled oscillator modes a and b, where a is the to-be-measured
quantum object and b represents the meter. The system-meter interaction is
taken to be of the form (the entire discussion will be performed in the interaction
picture and we omit for ease of notation the index /)

Hsm = - - ata (bE* - btE) . (4.280)

2

The physical background for this type of coupling is a four wave mixing interac-
tion (Walls and Milburn, 1994) which leads to a back-action evasion interaction,
276 DECOHERENCE

satisfying [f (ata), Hsiv] = O. Disregarding the coupling of mode b to the envi-

ronment we get the following pre-measurement time evolution:

losm(o)) = EcnIn) s O) _> losm(t)) = E cn1n)s Int/2)i. (4.281)

Thus, if the meter mode b is initially in the vacuum state 10)m, the number
states n,)5, of mode a become correlated with the coherent states

Mn) = InEt/2)m, (4.282)

which will be seen to play the rôle of the pointer basis states. We note that these
states are approximately orthogonal since their overlap is exponentially small for
t » 1,
6 1: t2 n)2]
(MrnlMn) = exp[ 1 (4.283)

The environment is represented by a collection of modes ci which are linearly

coupled to the meter mode b,

HME = t + EK•C•J• (4.284)

33

Tracing over the environment and performing the Born-Markov approximation

in the quantum optical limit we get the following master equation for the system-
meter density matrix,

1
c
ipsm(t) = -- [ata (be - btE) ,psm(t)]
dt 2
1 1
+70 (bpsm(t)bt - -btbpsm(t) - -psm(t)btb) , (4.285)
2 2
with the initial condition

P s m (0) = s m (0)) s m (0)1 = emcn ongros (10)(01)m • (4.286)

mm

We note that HmE violates the back-action evasion condition (4.276). We know,
however, that the master equation (4.285) leads to a rapid destruction of co-
herences between coherent states. We thus expect an approximately orthogonal
pointer basis to emerge, consisting of the coherent states (4.282).
To solve the master equation we employ the ansatz

psm(t) = > fnni(t) onxmos 0 onmornmom (4.287)

mm
 DECOHERENCE AND QUANTUM MEASUREMENT 277

where fnni (t) and 0 (t ) are c-number functions with initial conditions f,,,,(0) =
* en and On (0) = 0. Inserting this ansatz into the master equation (4.285) one
cm
is led to the differential equations

= — 1Sn +I
I , (4.288)

inni 1 d —ne Me
OM On +—— (10nd 2 + Ifimr) — (4.289)
.1
,rnrn = 70 2 dt 2 On 2 13;7%,

which are easily solved to yield

l'2' (1 — C'Y0t/2) , (4.290)

13 n(t) = 7o
— } . (4.291)
1E1 2 (n — m) 2 {3 —Nt — 4e-0t/2 + eyot
f nm(t) = C*rnen exp [
270

The solution of the master equation can thus be written in the following two
alternatives forms,
6.1 2 2
Psm(t) = Ec rvnexp [ 2,y1 02 (n — m) {3 — -yo t — 4e -1' 0t/ 2 + e—yot }]
mm
x (1n)(ml) s 0 (10n(t))(0m(t))m
2 70t
* 1 E1 y/2
(n — m) 2 1 — —2 _ e —ot
71 exp
=Ecmc 'To
nm

x (1n)(m) s 0 (
1fin(t))( 37n(t)l) (4.292)
(0m(t)1fin(t)) ) m .
Let us first discuss the limit Nt < 1, that is the limit of times which are
short compared to the relaxation time -y4 '. Then we have to lowest order in Nt,
nEt (4.293)
2
3— Nt — 4e- ,rot/2 _ (70 0 3 , (4.294)

and, hence,

psm(t),E c7n en exP [- - 4 (1E10 2 7ot — m) 2 1

nm
x ( n (m s
) ) (n Et /2 ) (mEt / 21) m • (4.295)

The measurement will be the more accurate the smaller the overlap (4.283) of
the pointer states (4.282). Thus we consider the limiting case,

lElt oo, 7ot —4 0, (lElt) • (Tot) = fixed. (4.296)

278 DECOHERENCE

Equation (4.295) shows that in this limit psm(t) becomes diagonal in the pointer
basis IMO = TI,EtI2),

psm(t)P-' E Icn12 onxn s ) ® Imn )(mnl. (4.297)

The decohering influence of the environment has thus singled out an approxi-
mately orthogonal basis of pointer states. This happens on a time scale which,
under the given conditions, is small compared to the relaxation time. Therefore,
damping effects play a negligible rôle and the pointer basis is independent of the
relaxation rate.
We finally remark on another interesting property of the model. The second
equation in (4.292) gives the following reduced density matrix for mode a,
2
ps(t) = E
nm
c,n c„ exp 1E1 (n — m) 2
7o
\

{1
— 22. — e —Y0t /2 11 (In)(rnDs • (4.298)
2

For -yo t < 1 we have to lowest order

_ _ e —yo t/2 ,,,..,,, __1 (2,0 02 .

1 _ -Yot (4.299)
2 8
It follows that the off-diagonals of ps(t) in the number state basis of mode a
decay exponentially with an exponent which is proportional to t2 , indicating the
non-Markovian character of the model for short times. For large times, -yo t » 1,
we find

ps(t) = E em,„,, rL
nm
1E ' (n — m) 2 ] 0 700nDs .
2-yo1
(4.300)

On differentiating this equation with respect to time we thus obtain

d 1E21
Ps(t) = -- [ata [ata ps(t)]] . (4.301)
2-yo ' '
This is a Markovian master equation for the reduced system S and its form is
that of a master equation in the singular coupling limit. The model thus shows a
transition from non-Markovian behaviour for short times to Markovian dynamics
for long times.

References
Bouwmeester, D., Ekert, A. and Zeilinger, A. (eds.) (2000). The Physics of
Quantum Information. Springer-Verlag, Berlin.
Breuer, H. P., Dorner, U. and Petruccione, F. (2001). Preparation and deco-
herence of superpositions of electromagnetic field states. Eur. Phys. J., D14,
377-386.
 REFERENCES 279

Brune, M., Hagley, E., Dreyer, J., Maitre, X., Maali, A., Wunderlich, C., Rai-
mond, J. M. and Haroche, S. (1996). Observing the progressive decoherence
of the "meter" in a quantum measurement. Phys. Rev. Lett., 77, 4887-4890.
Caldeira, A. O. and Leggett, A. J. (1985). Influence of damping on quantum
interference: An exactly soluble model. Phys. Rev., A31, 1059-1066.
DiVincenzo, D. P. (1995). Two-bit gates are universal for quantum computa-
tion. Phys. Rev., A51, 1015-1022.
Everett III, H. (1957). "Relative state" formulation of quantum mechanics.
Rev. Mod. Phys., 29, 454-462.
Friedman, J. R., Patel, V., Chen, W., Tolpygo, S. K. and Lukens, J. E. (2000).
Quantum superposition of distinct macroscopic states. Nature, 406, 43-46.
Gallis, M. R. and Fleming, G. N. (1990). Environmental and spontaneous
localization. Phys. Rev., A42, 38-42.
Joos, E. and Zeh, H. D. (1985). The emergence of classical properties through
interaction with the environment. Z. Phys., B59, 223-243.
Joos, E. (1996). Decoherence through interaction with the environment. In
Decoherence and the Appectrence of a Classical World in Quantum Theory
(eds. Giulini, D., Joos, E., Kiefer, C., Kupsch, J., Stamatescu, I.-0. and Zeh,
H. D.), Springer-Verlag, Berlin, pp. 35-136.
Maitre, X., Hagley, E., Dreyer, J., Maali, A., Wunderlich, C., Brune, M., Rai-
mond, J. M. and Haroche, S. (1997). An experimental study of a Schrödinger
cat decoherence with atoms and cavities. J. Mod. Opt., 44, 2023-2032.
Myatt, C. J., King, B. E., Turchette, Q. A., Sackett, C. A., Kielpinski, D., Itano,
W. M., Monroe, C. and Wineland, D. J. (2000). Decoherence of quantum
superpositions through coupling to engineered reservoirs. Nature, 403, 269-
273.
Palma, G. M., Suominen, K.-A. and Ekert, A. K. (1996). Quantum computers
and dissipation. Proe. R. Soc. Lond., A452, 567-584.
Savage, C. M. and Walls, D. F. (1985a). Damping of quantum coherence: The
master-equation approach. Phys. Rev., A32, 2316-2323.
Savage, C. M. and Walls, D. F. (1985b). Quantum coherence and interference
of damped free particles. Phys. Rev., A32, 3487-3492.
Schrödinger, E. (1935). Die gegenwdrtige Situation in der Quantenmechanik.
Naturwissenschaften, 23, 807-812,823-828,844-849.
Steane, A. (1998). Quantum computing. Rep. Prog. Phys., 61, 117-173.
Tegmark, M. (1993). Apparent wave function collapse caused by scattering.
Found. Phys. Lett., 6, 571-590.
Turchette, Q. A., Myatt, C. J., King, B. E., Sackett, C. A., Kielpinski, D., Itano,
W. M., Monroe, C. and Wineland, D. J. (2000). Decoherence and decay of
motional quantum states of a trapped atom coupled to engineered reservoirs.
Phys. Rev., A62, 053807,1-22.
Unruh, W. G. and Zurek, W. H. (1989). Reduction of a wave packet in quantum
Brownian motion. Phys. Rev., D40, 1071-1094.
Unruh, W. G. (1995). Maintaining coherence in quantum computers. Phys.
280 DECOHERENCE

Rev. A, 51,992-997.
Walls, D. F., Collet, M. J. and Milburn, G. J. (1985). Analysis of a quantum
measurement. Phys. Rev., D32, 3208-3215.
Walls, D. F. and Milburn, G. J. (1994). Quantum Optics. Springer-Verlag,
Berlin.
Zurek, W. H. (1981). Pointer basis of quantum apparatus: Into what mixture
does the wave packet collapse? Phys. Rev., D24, 1516-1525.
Zurek, W. H. (1982). Environment-induced superselection rules. Phys. Rev.,
D26, 1862-1880.
Zurek, W. H. (1991). Decoherence and the transition from quantum to classical.
Phys. Today, 44, 36-44.
Zurek, W. H. (1998). Decoherence, einselection and the existential interpreta-
tion (the rough guide). Phil. Trans. R. Soc. Lond., A356, 1793-1821.
 Part III
Stochastic processes in Hilbert
space
 5
PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

We saw in Section 2.1.3 that the statistical properties of a quantum mechanical

ensemble are fully characterized in terms of a density matrix. However, if selec-
tive measurements of one or several observables are carried on the ensemble, it
will split into a number of sub-ensembles, each sub-ensemble being conditioned
on a particular outcome of the measurements. The theoretical description of the
collection of sub-ensembles thus created leads to a new kind of quantum statisti-
cal ensemble which differs from the one encountered in Section 2.1.3. The central
subject of the present chapter will be to introduce these new ensembles. They
will enable us in Chapter 6 to look at the dynamics of open quantum systems
from a new perspective, namely from the viewpoint of stochastic processes in
Hilbert space.
The considerations of Section 5.1 serve to give a precise physical interpre-
tation of the new type of ensembles. The required mathematical framework of
functional integration in Hilbert space will be developed in Section 5.2. There we
introduce probability density functionals on Hilbert space which allow a general
characterization of the new kind of ensembles and enable the construction of ap-
propriate stochastic time-evolution equations for the state vector. As a further
generalization required in later chapters we are going to study in Section 5.3
probability distributions on the space of density matrices.

5.1 The state vector as a random variable in Hilbert space

We introduce in this section a new type of quantum statistical ensemble which
may be characterized by means of a random state vector in Hilbert space. The
general concept will also be illustrated with the help of a simple example.

5.1.1 A new type of quantum mechanical ensemble

In Section 2.1.3 we studied a certain type of ensembles 6' the statistical properties
of which are completely characterized by means of a density matrix p. In order
to distinguish such ensembles from the new ensembles introduced below they
will be denoted by E = Ep in the following.
Recall that an ensemble Ep was constructed by mixing M ensembles E,,
a = 1, 2, ... , M, of a certain quantum mechanical system with the weights wa =
NΠ/N. Each ec, represents a pure ensemble describable by a normalized state
vector 0,. The number Na denotes the number of systems in E„ and
284 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

N= Na (5.1)
(1=1

is the total number of systems in E. Denoting by , the indi-

st IV,,)
(t

vidual systems of Ea we may write the result of the mixing process symbolically
as
{ Q(1) q(N1) Q(1) g(N,) Q(1) NI)1
S(
(- 9 6-)1 ,." "1 7 " ' " a ," • )"ct 7 " 7 " M ' - 7 "M ' (5.2)
1

This notation serves to emphasize that the ensemble Ep is simply a collection of

N quantum systems.
The important point to note is the following: Saying that the statistical prop-
erties of Ep are completely described by the density matrix

P = z_d W al 'Oct)(Oct (5.3)

(1=1

one presupposes that the N systems SV making up the ensemble Sp are in-
distinguishable in the sense that no information on the original grouping into
the subsets Ea is available. An experimenter performing measurements on Ep
only knows that, taking at random a member from the ensemble, this particular
system is to be described with probability wa with the help of the state vector
It is certainly true that an experimenter with only this information to hand
can by no means predict anything other than the probabilities and expectation
values determined with the help of the density matrix p. It is for this reason
that one calls Ep a mixture, that is, a totally disordered collection of quantum
systems.
As discussed in Section 2.1.3.2 a given density matrix p can be expressed
in an infinite number of ways in the form (5.3) as a convex linear combination
of pure states '‘,/)a ) (0a 1 which need not be orthogonal. There is thus always an
infinite number of ensembles of the type Ep which all lead to the same density
matrix. Consider two such ensembles C, and Pp such that

M'
P= E waloa )(
oal = E wi,174)(741, (5.4)
,3=1

where wa , Oa are the weights and states of ensemble Ep , and w ,13 , 0 ,13 are those
of EP' ' Although both ensembles could have been prepared in entirely different
ways, they are described by the same density matrix p. Physically, the meaning
of this statement is that there is no way for an observer to distinguish these
ensembles by means of any experimental setup. Of course, the observer could
decompose both ensembles into a number of pure sub-ensembles by measuring
some discrete, non-degenerate observable R (or else, by measuring a complete
set of commuting observables). However, not only are the probabilities for the
 THE STATE VECTOR AS A RANDOM VARIABLE IN HILBERT SPACE 285

e)
many identically
prepared systems

6.,

FIG. 5.1. Picture of an ensemble of type 4 (left) and of an ensemble of type

Sp (right).

measurement outcomes the same for Ep and Pp , but also the collections of sub-
ensembles thus created. In this sense it is justified to say that an ensemble of
type Ep is completely characterized by a density matrix p.
As will now be demonstrated, one can, however, design an entirely different
type of ensemble the statistical properties of which are not fully characterized
by a density matrix. We denote such ensembles by the symbol Sp. To construct
them we again start from the ensembles Ea describable by pure states ipc,. One
may imagine that each of these states has been prepared by the measurement of
some complete set of commuting observables. The various 0, need not, however,
to be orthogonal. We again combine the S, with respective weights w a . However,
this time we would like to keep the information that a particular quantum system
belongs to a particular ensemble E. For this purpose we consider Na identically
, 6,,{,2) , . . . , en)
prepared copies d ) of E, for each a. The new ensemble ep is
then the collection of these ensembles, that is, an ensemble of ensembles:
gp { e ( l ) , . . . , El NO , . . . , Ei(xl) , . . . 7 6ra ) , 7 7 7 , EV 7 . 7 ., ""( N

t'Al
M ) 1

• ( 5.5 )

Note the decisive difference between (5.2) and (5.5): 4 is a disordered set of N
elements each of which represents an individual quantum system prepared in one
of the states O a . On the other hand, Ep is a set whose elements are again sets,
namely the ensembles EV. The distinction between both types of ensembles is
illustrated in Fig. 5.1.
The advantage of this construction is that Sp can now be regarded as a
sample which has been drawn from the sample space

12 = VI , — , Ea , ... , em } (5.6)

furnished with the probability measure

wa. (5.7)
286 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

Thus we have constructed a well-defined classical probability space: The sample

space Q is a set of ensembles Ea , the algebra A of events is provided by the
subsets of Q, and the probability measure p assigns the probabilities wa to the
elements of II
The physical significance of the sample space i4, or else, of the new type of
ensembles Ep stems from the following fact: An experimenter confronted with
an ensemble of the type Ep can always find out by an appropriate measurement
which particular ensemble Ea is realized, that is, he can determine to any desired
precision the state /P a describing Ea . In addition, he can do so in such a way
that the ensemble Ea is only negligibly disturbed by the measurement. To achieve
this he merely measures the state on a sub-ensemble of Ea , leaving the remaining
part of Ea completely unchanged. Thus we see also that the elements of 5-2 can
be considered as classical objects such that the rules for combing the various E„
are just the rules of ordinary, classical logic.
The probability measure (5.7) assigns to each Ea a probability wa . Since Ect
is represented by a state vector Oa the measure (5.7) gives rise to a probability
,

distribution on the space of state vectors. The conclusion is that the state vector
becomes a random variable in Hilbert space which is given by the assignment

Oa . (5.8)
The corresponding probability distribution is given by

P(Oct) = wa. (5.9)

This equation yields a probability distribution on the underlying Hilbert space
I-1 or, more precisely, on the space of rays in 7-1.
Having defined the random state vector (5.8) characterized by the distribu-
tion (5.9) we can interpret the corresponding density matrix (5.3) as the covari-
ance matrix of the random variable, that is, we may set

p= E ()OM , (5.10)
where E denotes the expectation value defined through the probability distribu-
tion (5.9). A more general definition will be given in the next section.
It is very important to realize that the concept of a random state vector is
by no means a hidden-variable theory. On the contrary, the above construction
is fully consistent with the statistical interpretation of quantum mechanics. The
reason for this fact is that the stochastic state vector is defined as a map on
the sample space Q whose elements are again ensembles. In complete agreement
with the statistical interpretation of quantum mechanics, a particular realization
Oa of the random state vector thus represents a pure statistical ensemble Ea of
quantum systems, all typical quantum correlations and interference effects being
embodied in this particular Oa .

It is obvious that an Ep-ensemble contains more information than a corre-

sponding En -ensemble. Thus, there exist observable quantities for Ep which are
 THE STATE VECTOR AS A RANDOM VARIABLE IN HILBERT SPACE 287

not observables for ep and which cannot be expressed in terms of the density
matrix. To give an example we consider some self-adjoint operator R. On ep we
can measure, of course, the variance

Var(R) = tr R2 p} — [tr {Rp}] 2

{

2
= Ew.(o.1/720.) — [Ewa(?PaIRIO(/)] (5.11)

of R in the usual way: Take a sufficiently large sample from ep , and estimate
from it the dispersion of R. The same quantity can be measured also on Sp, of
course. Consider now the following decomposition of Var(R),

Var(R) = Vari (R) + Var2(R), (5.12)

where

Val. ' (R) = Ewa [(vial/121o.) — oPaIR002] , (5.13)

and
2
Var2 (R) = Ewa(oaLRoa)2 — [Ewa(oalRoal (5.14)
a a

Both Val.' (R) and Var2 (R) represent non-negative quantities and their sum
equals the usual quantum variance (5.11) which is determined by the density
matrix p. Note, however, that Val. ' (R) and Var2 (R) cannot, in general, be ex-
pressed as density-matrix expectation values of some self-adjoint operator, which
means that these quantities are not observables for E. But they do represent
measurable quantities for ep-ensembles.
A definite prescription for the measurement of both Var. ' (R) and Var2(R) on
ep can be given. For this purpose we take a sample of ensembles 4,i) from Cp.
To measure Var i (R) we first determine the dispersion ( Octl R 2 lOct) (OaR 2
l 100
of R in each of the individual ensembles d) by carrying out a sufficiently large
number of measurements on the systems making up C,ç,i ) . The obtained values
for the dispersion of R will, in general, be different for the various EV . The
dispersion is thus a real random variable X whose classical statistical average
(X) equals Van i (R) . To obtain Var2 (R) one could, of course, simply subtract
Vari (R) from Var(R). But Var2 (R) can also be measured directly as follows. One
determines the expectation value (a O 1R1 O a ) of R in each individual ensemble
CV . In general, this expectation value takes on different values for the different
d) . Thus, the quantity Y = (Oa RIO a ) represents a real random variable whose
classical statistical variance Var(Y) = (Y 2 ) — (Y) 2 is equal to Var2 (R).
We infer from eqn (5.13) that Var i (R) vanishes if and only if the dispersion
of R vanishes for all 0,, which occur with non-zero probability, that is if and only
288 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

if all these ?P a are eigenstates of R. Thus, if Var. ' (R) is found to be zero we can
immediately conclude that Sp is an ensemble which consists of pure ensembles in
eigenstates of R. The variance Vari (R) can therefore be considered as a measure
for the distance of the Sp-ensemble to an ensemble made up of eigenstates of
R. On the other hand, definition (5.14) shows that Var 2 (R) vanishes if and only
if the random variable ( RNPŒ) is sharp, i.e. takes on a single value. In the
general case Var 2 (R) is the dispersion of the pure state quantum expectation
value and represents a measure of the statistical fluctuations of (0,1R '11),) over
the ensemble.
A partition of the quantum statistical variance Var(R) similar to (5.12) has
been considered by Wiseman in the context of the examination of a continu-
ously monitored laser (1993), while the variance Var2(R) has been introduced by
Wilmer, Castin and Dalibard (1993) as a measure for the statistical fluctuations
of the stochastic wave function. Examples of the physical significance of these
variances for the dynamics of quantum stochastic processes will be discussed in
Sections 6.7.2 and 8.2.

5.1.2 Stern-Gerlach experiment

Two Er -ensembles belonging to the same density matrix p cannot be distin-
guished by any experimental procedure. Two Ep-ensembles, however, can be
distinguished even if they lead to one and the same density matrix. Let us illus-
trate this point with the help of a simple example. We consider two observers.
called A and B, each with a Stern-Gerlach apparatus. A uses her apparatus to
prepare a large sequence of atomic beams consisting of spin- atoms, whereas B
analyses the beams with his own apparatus.
The Hilbert space belonging to the internal degree of freedom of the spin-4
atoms is 71 = C2 . As we know from Section 3A.2, any density matrix p in this
space can be written as a unique linear combination of the unit matrix I and
the Pauli spin matrices a l , a2, and a3 ,
1
1)01 = (I + V &')-
,
(5.15)

where V denotes the Bloch vector, satisfying 1'0 = 1 for a pure state. Thus, a pure
state is uniquely represented by a point V on the surface of the Bloch sphere.
Consequently, the probability distribution (5.9) describing an ensemble of the
type ep can be represented by means of a probability density P(11) which is
concentrated on the surface of the Bloch sphere, that is which vanishes outside
this surface and satisfies the normalization condition

f d3 v P(f)') = 1 (5.16)

Thus, the Bloch vector V becomes a random unit vector following the distribution
P(V). Note that P(11) is, in fact, a distribution on the space of rays in 'H. Of
course, one could characterize the distribution more explicitly as a density P =
 THE STATE VECTOR AS A RANDOM VARIABLE IN HILBERT SPACE 289

P(0,c,o) introducing spherical coordinates (0, ço) on the surface of the Bloch sphere
(Wiseman and Milburn, 1993a, 1993b). We prefer, however, to work here with
the above general representation.
Let us illustrate how observer A can generate an Sp-ensemble. We suppose
that A chooses a fixed unit vector 1. Employing her Stern—Gerlach
apparatus, she then prepares by the measurement of fi et a large sequence of
atomic beams. Each individual beam, consisting of a large number of atoms,
represents an ensemble in a definite pure state which is an eigenstate of 4-6 with
eigenvalue +1 or —1 and which is therefore given by one of the antipodal points
±fi on the Bloch sphere. Furthermore A ensures that in the sequence of beams
she produces both eigenstates occur with an equal weight of 1. The resulting
ensemble of the type Ep is therefore described by the following probability density
on the surface of the Bloch sphere,
11
P (if) = d (fi — Fi)
— 5(+ 71). (5.17)
2

This ensemble is now passed to observer B who performs measurements on it.

Let us first note that the Ep-ensemble given by (5.17) leads to the density
matrix

f d3 v p(V)P(F1) = ( 61- , (5.18)

as is easily verified with the help of (5.15). This density matrix corresponds to
an unpolarized atomic beam and is independent of ft. Thus, different choices
for fi obviously lead to different Sp-ensembles but to one and the same density
matrix p. The question is therefore, can B find out with the help of his own
Stern—Gerlach apparatus which particular direction has been chosen by A, that
is, can he determine the direction of the unit vector if which is obviously not
contained in the density matrix p? The answer to this question is affirmative, of
course.
Picking up a particular beam from the ensemble, B does not know by which
state it should be described. But he does know that each beam is in some definite
pure state and since each particular beam consists of many atoms he can do
statistics with it. To determine the direction of fi observer B can proceed as
follows. First he picks a sample {EV } of atomic beams. Each element EV is
an atomic beam describable by a pure state of the form (5.15). To determine
the Bloch vector fi observer B rotates his Stern—Gerlach magnet until there
is no splitting of the atomic beam into two parts. The direction defined by
this condition is the direction of the Bloch vector V. Observer B then knows
that the beam is in a pure state given by an eigenstate of if • ê, and he can
choose the orientation of V in such a way that the eigenvalue is +1. The beam
is thus determined to be in the state (5.15) with this Bloch vector V. Repeating
this procedure for his sample 14P1 of beams B will find out that only two
Bloch vectors V = ±fi occur in the sample with equal probabilities of 4. This
290 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

information enables B to reconstruct the probability distribution (5.17) created

by A and to fix the direction of the unit vector /I.
It should be clear from these considerations that B can also reconstruct, to
any desired degree of accuracy, any distribution P(V) in a similar manner. For
example, A could have created, at least in principle, the ep-ensemble given by
the distribution
1
P(IT) = —
47r 6 — 1) (5.19)

To generate this distribution A prepares atomic beams in the pure states p(V).
where the Bloch vector V is uniformly distributed over the surface of the Bloch
sphere. The preparation of this Ep-ensemble requires that A performs prepara-
tion measurements of all observables V - 6. We note that for different V these
observables do not commute, in general, and that they are measured on different
atomic beams. We also note that the normalization condition (5.16) is satisfied
and that (5.19) again yields the density matrix (5.18) of an unpolarized beam.
Following the procedure described above, B will now find, within the usual sta-
tistical uncertainties caused by the finiteness of his sample, that the Bloch vector
is uniformly distributed over the whole surface of the Bloch sphere and concludes
that the distribution is given by (5.19).
The fact that an Sp-ensemble contains more information than a correspond-
ing ep-ensemble may also be seen by determining the variances Var i (R) or
Var2 (R). Choosing a fixed unit vector ñ observer B may take R to be R = • .6.
for example. The expression for Vari (R) then becomes

Vari (rli • 6- ) = f d3 vP(fi) (tr { (tit • 6- ) 2 p(V)} [tr { ( fit' ) (v) )

• 6r. p

f d3 vP(V) (1 — (Tit V) 2 ), (5.20)

which is valid for all distributions P(t7). For the distribution (5.17) this expression
reduces to

Vari (fit' • 5-) = sin2 x, (5.21)

where x E [0, 7r) denotes the angle between the directions of ñ and ff. Corre-
spondingly, we find

Var2 (ili • 6) = cos2 x , (5.22)

such that the decomposition (5.12) of the variance Var(rii, • 6-) takes the form

Var(rn • 6) = sin 2 x + cos2 X = 1. (5.23

We observe that the variances Var 1 , 2 (rn • 6), which are measurable on E p , de-
pend on the angle between the directions chosen by A and by B. By contrast.
 PROBABILITY DENSITY FUNCTIONALS ON HILBERT SPACE 291

since p describes an unpolarized beam, all quantities which are measurable on a

corresponding Er -ensemble, for example the variance Var(rn • 6), are independent
of .x.
According to eqn (5.21) the variance Vari(ril • 6) vanishes if ni, is parallel
or antiparallel to rt. In this case B concludes that Ep is an ensemble which
consists of ensembles in eigenstates of fil - 6. This is an example of the general
property of Var. ' (R) discussed at the end of the preceding subsection, namely
that Var i (R) = 0 if and only if the ep-ensemble consists of eigenstates of R.

5.2 Probability density functionals on Hilbert space

The example discussed in the preceding section clearly demonstrates the physi-
cal significance of ensembles of type Ep. They result from selective preparation
measurements of a set of (not necessarily commuting) observables and give rise
to certain probability distributions on projective Hilbert space. In this section
we develop an appropriate mathematical framework which enables the general
description of such a distribution. A rigorous mathematical treatment of prob-
ability measures and functional integration in Hilbert space may be found in
Skorohod (1974) and Gihman and Skorohod (1980). The rôle of probability dis-
tributions on Hilbert space in the formulation of quantum mechanics has been
discussed by Bach (1979, 1980, 1981) (see also Bach and Wenning, 1982; Cyran-
ski, 1982). We emphasize, however, that the physical interpretation developed
above is entirely different from that suggested by Bach.

5.2.1 Probability measures on Hilbert space

Following the construction of a probability space given in Section 1.1 we consider
a system A of subsets of the Hilbert space 7-1 which plays the rôle of the sample
space. 12 The system A is assumed to form a a-algebra. Thus, each element A E A
is a subset of 7-1 and the system A of subsets satisfies the conditions of a a-algebra
formulated in Section 1.1.1. According to Section 1.1.2 a set function p = p(A)
defined for all A E A is called a probability measure if it satisfies the Kolmogorov
axioms (1.2), (1.3) and (1.5).
To give an example, we consider the simplest measures, i.e. the Dirac mea-
sures p = (5.0,. Taking an arbitrary fixed state vector 0 0 the corresponding Dirac
measure (5.00 is defined by

1, if 00 EA,
(5/P0(A) { 0, if 00 Ø A ,

for all A E A. As is easily verified this definition yields a probability measure

satisfying the Kolmogorov axioms.
In physical applications it is customary to work with densities. We thus in-
troduce a probability density functional P = P[0] of the state vector 0 and
write

12 More precisely speaking, the sample space is the projective Hilbert space, see Section 5.2.2.
292 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

,u(A) = f DODO*P[O] (5.25)

A
for the probability of A. Accordingly, the probability density functional of the
Dirac measure (5,/, will be denoted by P[O] = (5[0 — ?Po] such that we have the
relation

(5.0c, (A) = f DODO*(5[0 'Oo ]. (5.26)

A
The integration in these formulae is carried out over a subset A of the Hilbert
space. They thus represent a multidimensional or even an infinite-dimensional
integral, that is a functional integral. Our next task will be to give an explicit
construction for the corresponding functional volume element DODO*
in Hilbert
space
We take a fixed orthonormal basis {On } in 7-i and write the decomposition of
E 74 with respect to the basis vectors as follows,
— E ZnOn• (5.27)

The probability density functional P = P[0] can thus be considered as a function

P = P[zn , zn*] of the complex variables zn , zn*. Alternatively, we can regard P as
a function P = P[an , bn] of the real variables a n , bn defined by
Zn— a + ib. (5.28)
An appropriate expression for the volume element in Hilbert space is the usual
Euclidean volume element in a real space with coordinates an ,
bn , that is we set
DODO* = H dan dbn . (5.29

Writing the differentials da n and dbn as

1
da n = — (dz n dz n*), (5.30)
2
1
dbn = — (dz n — dz n*), (5.31
2i
we can write the volume element (5.29) in terms of the coordinates z n and z.

al,GDO* = —2i dz n dz. (5.32

rt
Summarizing we have the following explicit formulae for functional integration
in Hilbert space

f DODV)* P[11)] = f II dandbnP[an, bn] — f 11 Ldzndz

2
n*P[zn, z n* ] . (5.33 )
A A n A n

The variables zn and zn* must be treated as independent integration variables

corresponding to the two independent real integration variables a n and bn.
 PROBABILITY DENSITY FUNCTIONALS ON HILBERT SPACE 293

In accordance with definition (5.29) the density of the Dirac measure

belonging to the state

oo = Ezn°on 3 Zn
° = an0 + tOn0 ,
•I
(5.34)
n

is defined to be

00] = 11 6(an — a cn' ) 6 (bn — C. (5.35)

n

This is a functional Dirac function given by a product of ordinary 6-functions.

If one works in the position representation, for example, one may proceed
as follows. We consider a finite volume V and discretize it into cells of equal
volumes AV centred at x i , / = 1, 2, ... , M. The continuous wave function 0(x)
is thus approximated by the finite-dimensional vector

(5.36)
/
In this case the volume element takes the form

DODO* = H doi dot. (5.37)

/
In the continuum limit this will be written formally as

DODO* = H d0(x)d0* (x), (5.38)

x
where the product is to be extended over all space coordinates.
In the following we will often make use of an important property of the func-
tional volume element on Hilbert space constructed above, namely its invariance
under linear unitary transformations

0' = UO (5.39)
of the Hilbert space. This invariance may be expressed by the formula,
DO'DO1* = DODO*. (5.40)
To prove this formula we regard U as a unitary matrix in the basis introduced
in eqn (5.27) and decompose it into real and imaginary parts,

U = R (U) + i:1(U). (5.41)

The unit arity of U leads to the relations
R(U)R(U) T + :1'(U)(U) T = 1, (5.42)
— R(U):1(U) T = 0, (5.43)
where T denotes the transposed matrix. In the chosen representation the matrix
U describes a unitary transformation z n i— zn1 from the coefficients z n in the
294 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

basis decomposition of ti) to the coefficients zn in the basis decomposition of

= U. The corresponding transformation of the real coefficients an , bn defined
by z = an + ibn is provided by the real matrix

_ [R(U) —a(U)) (5.44)

R(U)

It is now easy to check with the help of eqns (5.42) and (5.43) that U is an or-
thogonal matrix satisfying Idet U = 1. Thus, as was to be expected, the unitary
transformation U of the Hilbert space 7-1 induces an orthogonal transformation
Ü of the real variables an , bn which were introduced to define the volume element
in Hilbert space. We can now apply the transformation formula for multidimen-
sional integrals to conclude that

da71 ,db1,2 =Idet 01 dandbn = dandbn , (5.45)

which proves the unitary invariance (5.40) of the volume element.

An important consequence of this result is that our definition (5.29) for the
volume element DOD** is independent of the choice of the basis {On }. Moreover.
we immediately deduce from the transformation formula (5.40) that the Dirac
density functional 6[0] is invariant under unitary transformations,

(5.46)

and, hence,

(5kG — U00] = (5[Ü—L O — Oo]. (5.47}

These relations are useful for calculations with probability density functionals.

5.2.2 Distributions on projective Hilbert space

We have already met the simplest examples of probability measures on Hilbert
space, namely the Dirac measures (5,/,, . The corresponding density functionals
6[0 — IN] satisfy

(5.48)
f DO.W*(5[0 — 0 0 ] = 1.
The general normalization condition for the probability density functional P[vj
reads

f D'ODO*P[0] = 1,
where the integration is extended over the whole Hilbert space.
According to the general principles of quantum mechanics the physical state
of a pure ensemble is uniquely described by a ray in projective Hilbert space (see
 PROBABILITY DENSITY FUNCTIONALS ON HILBERT SPACE 295

Section 2.1.3.2). In other words, pure states are characterized by normalized wave
functions and wave functions which differ by a phase factor are equivalent. Since
we want to characterize quantum statistical ensembles of the type gp with the
help of a probability density functional P[0] we postulate that the probability
distribution is concentrated on the surface of the unit sphere in Hilbert space.
That is we demand the existence of a functional Q[0] such that

P i) ]
[.
= 6(101 1 - 1 )(2 [0 ]. (5.50)

This means that only normalized wave functions occur with non-zero probability.
The equivalence of wave functions which differ by a phase factor leads to the
requirement that the density functional does not depend on the phase of the
wave function, i.e. for all x E [0, 27) we must have

P[exp(iX)0] = P[0]. (5.51)

By virtue of the conditions (5.50) and (5.51) the functional P[0] can be re-
garded as a probability density on projective Hilbert space (Breuer and Petruc-
cione, 1995). In fact, a representation of projective Hilbert space is obtained by
taking the surface of the unit sphere in 7-1 defined by 11011 = 1, and by identifying
those points on it that differ by a phase factor. Since P[0] is constant along the
rays in 7-1 it can be considered as a functional on projective Hilbert space. Sum-
marizing, an ep-ensemble is characterized by a probability density functional
P = P[0] satisfying the normalization condition (5.49) and the requirements
(5.50) and (5.51) of a functional on projective Hilbert space.
The simplest example for a density functional on projective Hilbert space is
provided by the expression

27
PM = f—
dX
27r (5[0 — e ix 0o], (5.52)
o
where 00 is a normalized state. As is easily checked this functional satisfies the
normalization condition (5.49), as well as the postulates (5.50) and (5.51). In
fact, since 00 is supposed to be normalized the functional (5.52) is concentrated
on the unit sphere in 7-1. Moreover, P[0] is invariant under changes of the phase
of 0 by virtue of the x -integration in eqn (5.52).
Another simple situation occurs, if the experimental preparation yields an
ensemble which consists of a discrete set of ensembles Ea labelled by an index a,
each Ea being describable by a normalized state vector 0„. This is the situation
that was investigated at the beginning of this chapter, and also encountered
in the example of the Stern-Gerlach experiment. Again we denote by w, the
statistical weight of the ensemble ea, such that w, > 0, and a
w, , 1. The
probability density P[0] which describes the corresponding ensemble of the type
ep consists of a sum of Dirac densities which are concentrated around the various
296 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

Oa , each Dirac density being weighted with the corresponding factor w a . Thus,
in this case we have
27r
P[7pi = E wo,
a
f
0
(5 [0 eiX Ocd. (5.53)

Again, it can be immediately verified that our basic requirements (5.49), (5.50)
and (5.51) are fulfilled for this probability density functional.
The densities P[0] encountered so far are all representable by sums of Dirac
measures. Of course, one can also easily construct continuous densities. An ex-
ample of such a density has already been given in Section 5.1.2. The density
P(e) given in eqn (5.19) is constant on the surface of the Bloch sphere iJ = 1
the projective Hilbert space of a spin-1 particle. To generalizewhicrepsnt
this idea we consider an arbitrary s-dimensional linear subspace V of 1-1 which
is spanned by an orthonormal set 0 i , i = 1,2, ... , s,

V = span{017 02, • • • (5.54)

Let us assume that the preparation process yields a mixture of all normalized
states in V with equal weights (von Neumann, 1955). The correct probability
density describing the corresponding Sp-ensemble takes the form
1 f (5.55)
P ['P] = K J ciw ( À) 6['P

Here we have introduced the surface K of the unit sphere in V. This surface
represents a manifold of real dimension (2s — 1) given by

K {(/)(A) E o po E Appi , Ai E C, E 1 Ai 1 2 = 1 } (5.56)

Furthermore, dw(A) denotes the usual Euclidean surface element of K and

1 -1( = f dw(A) (5.57)

is the total volume of K. Hence we see that eqn (5.55) provides a uniform prob-
ability density which is concentrated on the unit sphere K C V. Note that the
integration over the sphere K in eqn (5.55) ensures that the probability density
functional P[0] does not depend upon the special choice of the basis functions
Oi. Equation (5.55) makes sense also for the case s = 1. In that case K is iso-
morphic to the unit circle and the dw(A)-integration reduces to the integration
over the phase x in eqn (5.52).
More generally, we may consider an ensemble ep which results from the
combination of ensembles of the above type with weights w, > 0 satisfying
 PROBABILITY DENSITY FUNCTIONALS ON HILBERT SPACE 297

= 1.Using an obvious generalization of the above notation this leads to

the density functional

P[0] = E wa f du),(A) (5[0 — o a (A)]. (5.58)

a K,
It should again be clear that this density functional (5.58) satisfies the basic
properties (5.49), (5.50), and (5.51).

5.2.3 Expectation values

With the help of the probability density functional P[0] corresponding to an
Ep-ensemble we can now define the expectation value of a functional F[0] of the
random state vector 0 through the relation

E (F[0]) E f DODO* P[01F[]. (5.59)

Taking a self-adjoint operator R we may consider the functional F[0] = (0140)

and its expectation value

E &AO) = f DODO * P[0] (040- (5.60)

The quantity

( '7) = E ((01/7 10)) = tr IRO (5.61)

is the expectation value of R in an 4-ensemble corresponding to the Sp-ensemble
described by P[0]. This expectation value for the measurement of R on 4 is ob-
tained by averaging (01/0) over the probability density functional P[0]. Hence,
it is evident that the density matrix p describing Ep is equal to the covariance
matrix

p= E 0)(01) E f D 'OD O* P[0] ION

( (5.62)

of the random state vector 0 defined in terms of the density functional P[0]
(compare with eqn (5.10)). This shows again that the statistics of an 4-ensemb 1 e
is completely determined by a probability density functional P[0]• The converse
is of course not true, in general, since a probability density functional is not fixed
uniquely by giving only its covariance matrix.
Let us evaluate expression (5.62) for the density functional given in eqn (5.53).
Substituting the expression for P[0] we get
27r
* f dx
p = E wa f Dop o 27r 6[0 — e x 0cd10)(01,
— (5.63)
a
0
298 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

and, carrying out the functional integration with the help of the properties of
the Dirac density, we arrive at
27r
f dx
opa
= Ewalool, (5.64)
a
a 0

which is recognized as the density matrix of the corresponding 4-ensemble. As

another example we substitute the density functional (5.55) into (5.62) which
yields
1 1
p = f DODO* f dw(À) (5[0 ON] 0)(0 dw(A)10(A))(0(A)1
K =IKIf

(5.65)

where we have again carried out the functional integration over the Hilbert space
and used the representation (5.56) for the surface K of the unit sphere. Perform-
ing finally the dw(A)-integration one finds

I' 1 1
p— 1
1K1J
dw(A) El Ai 10i)(0i1 = 2_, 10i)(0i1 = (5.66)

As expected, the corresponding 6'p -ensemble is described by a density matrix p

which is proportional to the identity Iv in the subspace V spanned by the
All quantities measurable on an 4-ensemb 1 e are expressible in terms of the
covariance matrix p of P[0] which involves moments of of second order. Higher-
order moments of the random state vector represent quantities which are only
measurable on ep. Examples are the variances Var i (R) and Var2 (R) introduced
already in eqns (5.13) and (5.14). With the help of the probability density func-
tional P[0] these equations can now be written as (Breuer and Petruccione.
1996)

vari (R ) = f Dopo*P[0] (( 0 1 R2 10) — (5.67)

and
2
Var2(R) = f DODO* P[0] (01R10)2 [f DODO * P[0] (01R10)] (5.68)

These expressions involve fourth-order moments of the state vector. Var i (R)
represents the dispersion of R averaged over the pure states contained in the
ensemble and is thus a measure of the average intrinsic quantum fluctuations.
On the other hand, Var 2 (R) is the variance of the real random variable (OW).
It is thus a measure of the classical statistical fluctuations of this quantity.
 ENSEMBLES OF MIXTURES 299

5.3 Ensembles of mixtures

In the foregoing discussion we have considered ep as an ensemble of various
ensembles ec, which are describable by pure quantum mechanical states
This idea led us to the introduction of probability density functionals P[0] on
projective Hilbert space. An even more general type of ensembles arises, if one
takes into account that, as a result of the preparation process for example, the
various E, might be in true statistical mixtures described by density matrices
p,. This idea then forces us to consider probability density functionals on the
space S(7L) of density matrices.

5.3.1 Probability density functionals on state space

We denote probability density functionals on the state space of density matrices
by P = P[a],where a E S(R) is a density matrix. They characterize the statis-
tical properties of the random density matrix a. To construct the corresponding
volume element we proceed in close analogy to the case of distributions P[0] on
the Hilbert space R.
Taking an orthonormal basis {O n } in 74 we write the matrix representation
of an operator a as

CY = ElooGrnm,(0m.1- (5.69)
nm

Next we decompose the matrix elements anm into real and imaginary parts,

Cfnm = Unm + iVntn • (5.70)

An appropriate volume element may now be defined by

Do- E II dun dv nm .
.,.._ m,,,..... (5.71)
n,m

A functional F = F[a]can be considered as a multidimensional function F[a] =

F[urim , vrim ] of the variables u, and v„ and we define the functional integra-
tion by means of

f D a-F[u] = f II dunin dv nm F[unin,

n,m
Vnmi- (5.72)

The corresponding Dirac density is given by

6[a] = H 6 (u nn, )6( vnin ) - (5.73)

n,m

A probability density functional on the space of density matrices is a non-

negative functional P = P[a] which vanishes if any of the following conditions
are violated,
300 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

li mn = U nrn , Vrnn = —Vnrn , (5.74)

tr a = Eunn 1, (5.75)
n
a > 0. (5.76)
The first condition ensures that only Hermitian matrices occur with non-zero
probability and implies, in particular, that v, = 0. The second condition is the
normalization condition for the random density matrix a, while the third condi-
tion requires that P[a] must be identically zero outside the convex region defined
by the positivity of a. Finally, the normalization condition for the probability
density functional takes the form

f Do-P[a] = 1. (5.77)

As for the volume element DODO* one can easily prove that the volume
element Du is invariant under unitary transformations given by
cr' = UaUt ) (5.78)
where U is an unitary operator in R, i.e. that we have
Da' = Du. (5.79)
In particular, this invariance implies that our definition (5.71) of Du does not
depend on the basis {On chosen.
}

The expectation value of some self-adjoint operator R which is obtained by

measuring on a corresponding ensemble of the type Ep is given by the first mo-
ment of P[a],

(R) = f Da-P[a] tr {Ro-} . (5.80)

Consequently, the density matrix describing 4 takes the form,

p = E(a) E f Da- P[a] a. (5.81)

Of course, the different types of variances defined in the preceding section can
be introduced in an analogous way for density functionals P[a] on the space of
density matrices. More precisely, we have again the decomposition (5.12) of the
variance Var(R) into two variances Var i (R) and Var2 (R), where now

Var i (R) = f Do-P[a] (tr {R2 o- } — [tr {Ra } ?) , (5.82)

and
2
Var2(R) = f Do-P[a] [tr {Ro } ? — [f Da-P[a]tr {Ra}] . (5.83)

The physical interpretation and the measurement prescription for these quanti-
ties are analogous to those given in Section 5.1.1.
 ENSEMBLES OF MIXTURES 301

5.3.2 Description of selective quantum measurements

As an example of the use of the probability density functionals P[a] we consider
the selective measurement described by effects Fm and operations ck,i (see Sec-
tion 2.4.2). Performing such a measurement on an ensemble Ep described by the
density P[a] yields a new ensemble Eip which is described by the density Pqoa]
given by

cb in ()

P10-1 f D o-P[a] tr {Fm o- } 6 [ tr {4),72(0.)} al. (5.84)

7 2,1
Obviously, the new density functional is normalized in view of condition (2.148),

f Do-'13 [0-1= E f Do-P[o-]tr {Fm } = f Do- P[a] = 1. (5.85)

On using the first representation theorem of quantum measurement theory (see

eqns (2.157) and (2.158)) we can rewrite (5.84) as

1=)/[oi] E f Dap[a] tr {E n mtkt2mku}

in tr
Ek Q rrik a9mtk
Ek Qtmk Q niko- }
0-1 1.
(5.86)

In the simplest case the index k in the last equation takes on a single value only.
This situation occurs, for example, if one considers an indirect measurement
with a quantum probe which is in a pure state initially. The last equation then
simplifies to

E f Do -P[a]tr {5-2Qm o- }6[9"1°1"n (5.87)

rn tr {545-2,,o- }

It follows from eqn (5.87) that a distribution P[0] of pure states transforms
again into a distribution PTV] of pure states: If we perform the measurement on
an ensemble in the state V), the new state conditioned on the outcome in is given
by = Qrn0/119?-n011• Thus we get the following relation between the initial
and the final density functional on Hilbert space,

= E f DoDo* Ipmo126 11 9 rn0

0/1 P[0 ]. (5.88)

The above relations describe the change of a probability density functional

induced by the selective measurement on ensembles of the type Ep. They consti-
tute the starting point for our derivation of quantum stochastic processes which
will be presented in the next chapter.
302 PROBABILITY DISTRIBUTIONS ON HILBERT SPACE

References
Bach, A. (1979). Quantum mechanics and integration in Hilbert space. Phys.
Lett., A73, 287-288.
Bach, A. (1980). A probabilistic formulation of quantum theory. J. Math.
Phys., 21, 789-793.
Bach, A. (1981). Aspects of a representation of quantum theory in terms of
classical probability theory by means of integration in Hilbert space. J. Phys.
A: Math. Gen,., 14, 125-132.
Bach, A. and Wenning, T. (1982). A probabilistic formulation of quantum
theory. II. J. Math. Phys., 23, 1078-1081.
Breuer, H. P. and Petruccione, F. (1995). Stochastic dynamics of open quantum
systems: Derivation of the differential Chapman—Kolmogorov equation. Phys.
Rev., E51, 4041-4054.
Breuer, H. P. and Petruccione, F. (1996). Quantum measurement and the
transformation from quantum to classical probabilities. Phys. Rev., A54.
1146-1153.
Cyranski, J. F. (1982). Quantum theory as a probability theory on Hilbert
space. J. Math. Phys., 23, 1074-1077.
Gihman, I. I. and Skorohod, A. V. (1980). The Theory of Stochastic Processes
I. Springer-Verlag, Berlin.
Molmer, K., Castin, Y., and Dalibard, J. (1993). Monte-Carlo wave function
method in quantum optics. J. Opt. Soc. Am., B10, 524-538.
von Neumann, J. (1955). Mathematical Foundations of Quantum Mechanics.
Princeton University Press, Princeton.
Skorohod, A. V. (1974). Integration in Hilbert Space. Springer-Verlag, New
York.
Wiseman, H. M. (1993). Stochastic quantum dynamics of a continuously mon-
itored laser. Phys. Rev., A47, 5180-5192.
Wiseman, H. M. and Milburn, G. (1993a). Quantum theory of field-quadrature
measurements. Phys. Rev., A47, 642-662.
Wiseman, H. M. and Milburn, G. (1993b). Interpretation of quantum jump
and diffusion processes illustrated on the Bloch sphere. Phys. Rev., A47.
1652-1666.
 6
STOCHASTIC DYNAMICS IN HILBERT SPACE

A quantum master equation describes the dynamics of an open system by an

equation of motion for its reduced density matrix PS (t). In the language of the
previous chapter this provides the evolution of an Er -ensemble. On the other
hand, Ep-ensembles are characterized through density functionals P[0] on the
underlying Hilbert space. Drawing on the analogy to the theory of classical
stochastic processes developed in Chapter 1, one could ask the question as to
whether the unfolding in time of an Ep-ensemble may be formulated in terms
of an appropriate dynamics for a time-dependent density functional P[O, t]. It
will be shown in the present chapter that such a formulation is indeed possible.
It leads to a stochastic process 0(t) in the open system's Hilbert space which
reproduces the density matrix through its covariance matrix, that is through the
expectation value ps (t) =
This is the idea underlying the so-called unravelling of the master equation.
Instead of representing the dynamics of an open system by a quantum master
equation for its density matrix, it is formulated in terms of a stochastic process
for the open system's wave function. The structure of the Lindblad generator
for a quantum Markov process leads to a close relation between dynamical semi-
groups and piecewise deterministic processes in Hilbert space. This chapter gives
a detailed account of this relation and of its physical interpretation and signifi-
cance. The theory will also be illustrated with a number of applications.
The stochastic representation of quantum Markov processes already appeared
in a fundamental paper by Davies (1969), and was applied by Srinivas and Davies
(1981) to a derivation of the photocounting formula. While the theory was origi-
nally formulated in terms of a stochastic process for the reduced density matrix, it
was recognized by Barchielli and Belavkin (1991), Dalibard, Castin and Molmer
(1992) and by Dum, Zoller and Ritsch (1992) that it can be formulated also as
a stochastic process for the state vector in the reduced system's Hilbert space
and that it leads to efficient numerical simulation algorithms. At the same time,
Carmichael (1993) developed the idea of the unravelling of the master equation
in terms of an ensemble of quantum trajectories which are the realizations of the
underlying stochastic process.
The physical basis for the stochastic state vector evolution is provided by
continuous measurement theory: The stochastic dynamics of an Sp-ensemble
pertaining to an open system results from a continuous monitoring of certain
observables of its environment. The process OM thus embodies the random
changes of the state vector conditioned on the outcomes of this monitoring.
304 STOCHASTIC DYNAMICS IN HILBERT SPACE

Accordingly, different detection schemes yield different stochastic processes for

the conditioned state vector, as was demonstrated by Wiseman and Milburn
(1993a, 1993b). Microscopic derivations of various stochastic processes will be
given in this chapter.

6.1 Dynamical semigroups and PDPs in Hilbert space

This section serves to introduce, on a formal mathematical level, the relation
between dynamical semigroups and a certain class of piecewise deterministic
processes in Hilbert space. We further discuss some important mathematical
properties of this class of PDPs. The physical meaning of the stochastic repre-
sentation as well as its derivation from microscopic models are shown in the next
section. Specific physical examples are investigated in Sections 6.3-6.7.

6.1.1 Reduced system dynamics as a PDP

6.1.1.1 Closed systems Let us first consider a closed system whose dynamics
over the time interval from time t o to time t is described by a unitary time-
evolution operator

U(t, to) = exp [—iH(t — to)] , (6.1)

where H denotes the Hamiltonian of the system. In order to construct the prob-
ability density P[», t] for the process OM describing the evolution of a corre-
sponding Ep-ensemble (see Chapter 5) we have to introduce an initial probabil-
ity distribution /30[0] which represents the state of the ensemble at time to . The
probability density for 0(0 to take the value V) at time t is then given by

P[11), t ] = f DOoD O 8Po[005 [11) — U(t, t o) 0 0] = Po[ U-1 ( t,t o)0 .

] (6.2)

In the second equality we have made use of the definition (5.35) of the Dirac
density in Hilbert space and of eqn (5.47). The above equation simply expresses
the fact that any 00 drawn from the initial distribution Po evolves according to
the Schrödinger equation, that is the density at 1/) at time t is equal to the initial
density at the corresponding initial value U-1 (t, to )'. In the language of the
theory of stochastic processes V) is a deterministic Markov process (see Section
2.4.3) and the 6-function in (6.2) is nothing but the transition probability of the
process.
Equation (6.2) can be written in differential form. To this end we introduce
the functional derivatives 6/60(x) and 6/60*(x) with the properties

6 6
0(Y) = (6.3)
60(x)
6 6
O. (6.4)
60(x) 0*(y) = 60* (x) 0(y) ----
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 305

V (t ,to)
Ps (to) Ps ( t )
E[
XHÏ E[10(t)) OP Mil
p[il TP,b ,tk 41 > p t]
to]

FIG. 6.1. The unravelling of the master equation leads to a commutative di-
agram: Starting from P[O, to] one can either form the covariance matrix to
get ps(to) and propagate with the help of the super-operator V(t, to) which
represents the time evolution according to the master equation (6.7), or else
one can first propagate the stochastic process to get P[O, t] and form the
covariance matrix. Both ways lead to the same density matrix ps(t).

If F[0(t)] is any functional of a time-dependent wave function the chain rule

may be expressed in terms of the functional derivatives as follows

6F d11)(x,t) 6F ch1)*(x,t)1
= f d x {0
5( (6.5)
dt x) dt 60* (x) dt f•

With the help of these relations we immediately obtain from eqn (6.2) the dif-
ferential form
6
— p[o t],
at '
if dx { 60(x) (H0)(x) 6.
0* (x)
(H0)* (x)} P[O,t]. (6.6)

This is just the Liouville equation corresponding to the flow induced by the
Schrödinger equation /-p(t) = iH0(t). It is easy to check that (6.6) preserves
—

the basic conditions (5.49), (5.50) and (5.51) required for a probability density
functional on projective Hilbert space.

6.1.1.2 Unravelling the quantum master equation Let us now turn to an open
system S. The dynamics of the density matrix ps(t) of S is assumed to be
describable by a quantum dynamical semigroup satisfying a Markovian master
equation in Lindblad form (see Section 3.2.2),

d
Ps(t) = — 41 I, Ps(t)] + Epyi (Aips(t)A ti — Ati A i ps(t) — - ps(t)Ati A,)
(6.7)

Our aim is to demonstrate that the dynamics given by this equation can be
represented as a piecewise deterministic process 11)(t) in the Hilbert space of the
open system S in the following sense. The process will be defined through an
appropriate time-evolution equation for its probability density functional P[5, t].
306 STOCHASTIC DYNAMICS IN HILBERT SPACE

The density functional at time t is related to the density functional at time to

by

-13 [0,t] = f DV )- D:O. 710,0,t01-13 W),t01, (6.8)

where T [ b,q,t0] is the conditional transition probability of the process. The

process 0(t) gives rise to a certain equation of motion for the covariance matrix
which was defined in eqn (5.62) and will be written explicitly as follows,

ps (x,x', t) E [(x, t)* (z', t)] =-- f DODO* P[O,t]0(x)11)* (x'). (6.9)

The covariance matrix is identified with the density matrix ps(x, x', t) of S. The
basic requirement is then that the expectation value (6.9) satisfies the Lindblad
equation (6.7). Thus, the process OM reproduces the reduced system's density
matrix through its covariance matrix, as is illustrated in Fig. 6.1. A process which
satisfies this requirement is sometimes called unravelling of the master equation
(Carmichael, 1993).

6.1.1.3 The Liouville master equation As discussed in Section 1.5.1 a PDP

may be defined in terms of a Liouville master equation for its probability density
(Breuer and Petruccione, 1995a, 1995 b). We consider here a Liouville master
equation for a PDP in Hilbert space which takes the following form analogous
to eqn (1.150),

N P[O,t] = i fdx { 6.0(5(x) G(0)(x) 6.0*(5 (z) G()* (x)} P[5, t]

+f {W[01V-)]P[V;,t] — W[010]P[O,t]}. (6.10)

The first term on the right-hand side provides the Liouville part. It represents
the generator corresponding to the deterministic time-evolution equation
d
— = —iG(0(0), (6.11)
dt
where G(0) is a non-linear operator defined by

G(0) = + —2i El'illAi011 20. (6.12)

The linear operator fi is given by

H — -yiAti Ai . (6.13)
2

Thus, k consists of two parts, namely of the Hamiltonian H which appears in

the Lindblad equation (6.7), and a non-Hermitian part which is defined in terms
of the Lindblad operators A.
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 307

The jump part of eqn (6.10) (given by the second term on the right-hand
side) describes the rate of change of P[, t] due to discontinuous jumps of the
wave function. The gain term represents the total rate for all transitions from
any state into the state 0, whereas the loss term gives the total rate for all
transition leaving the state O. The corresponding transition rate is defined to be

Aii7)
W[010 ] = E7illAv0112 6[ (6.14)
1Av011
Deterministic evolution The periods of deterministic
6.14 evolution of the
process are given by eqn (6.11) which has the form of a non-linear Schrödinger
equation. The formal solution corresponding to the normalized initial value
0(0) = /7) can be written in terms of the non-Hermitian Hamiltonian fi as
follows,

(t) _ gt (z);) exp ( — if/0

(6.15)

To prove this equation we differentiate it with respect to time,

d O(t) = — iftO(t) dllet0112.

—10(t)11e—mt11-2— (6.16)
dt 2 dt
The time derivative of the norm is found to be
d -
(fit _ fi)
—
dt ile —dit-b11 2 = 41eifitt
= (6.17)

and, hence,

2_ (6.18)
11e - lktv3 E'Yi(0(t)1Ati,Ai10(0).

Substituting this into eqn (6.16) immediately yields eqn (6.11).

Equation (6.18) shows that the time evolution generated by H leads to a
monotonic decrease of the norm of the wave function. The non-linear part of
eqn (6.11) compensates this decrease, such that the full non-linear Schrödinger
equation preserves the norm of the wave function.

6.1.1.5 Jump process and waiting time distribution According to eqn (6.14)
the total rate for transitions from a given state L to some other state is given by

F [0]= f D 5 DO * W[010] = FrillAiV;112. (6.19)

Let us assume that the normalized state was reached through a jump at time
t. Due to the continuous time evolution between the jumps, the total rate for
308 STOCHASTIC DYNAMICS IN HILBERT SPACE

the next jump depends on the time T elapsed since time t. With the help of the
flow ''(r) gr() defined in eqn (6.15) the time-dependent total transition rate
may be written

F[g4)] =Elii(O(T)1A it Ail0(T)) = — Tr

ci lnllexp(—ifir)'0112, (6.20)

where we have used eqn (6.18) in the second step. According to the general theory
of PDPs (see Section 1.5) the distribution function for the random waiting time
T is thus given by

F[V, , 7-] = 1 — exp ( — j. ds F[gs(V;)] = 1— IlexP( — iii' '0'011 2 - (6.21)

o
This quantity yields the probability for the next jump to occur somewhere in
the time interval [t, t + T] . We see that the waiting time distribution is simply
determined by the decrease of the norm of exp(—i1-/- 7- )0- , that is by the decrease
of the norm of the state vector that obeys the linear part of the non-linear
Schrödinger equation.
Following our general discussion of the waiting time distribution function of
PDPs in Section 1.5.2, we first note that the limit

exp(—i f1 4 11 2 q (6.22)

exists and thus F[17), oc] = 1 — q. The number q is the defect and satisfies 0 <
q < 1. For vanishing defect it follows that F[', cc] = 1. This means that the
next jump occurs with probability 1 in some finite time. However, if fi has a
zero-mode it is possible that q > 0. In this case, the defect q is the probability
that after time t no further jumps occur. A physical example will be given in
Section 8.2.
The quantity 147[010] (eqn 6.14) denotes the probability density per unit of
time for a jump from 1-/) to 0. Since it is given by a discrete sum of functional
6-functions we have a discrete set of possible transitions. Under the condition
that the state just before the jump is given by /-/), the particular jump

1/3 Ai 1/)
(6.23)

takes place with probability

(6.24)
pi r[V3] •
Note that these probabilities sum up to 1, that is E i pi = 1 by virtue of eqn
(6.19). The transitions (6.23) will be referred to as quantum jumps.
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 309

FIG. 6.2. A single realization of the PDP for a one-dimensional harmonic oscil-
lator. The oscillator is coupled to a low-temperature heat bath and is driven
by a time-dependent force. The picture shows Izi)(x, 01 2 as a function of z
and t. The initial state is the ground state of the oscillator. In addition to
smooth evolution periods, one observes the wave function performing sudden,
discontinuous jumps.

Again, one easily verifies that the Liouville master equation (6.10) preserves
the basic conditions (5.49), (5.50) and (5.51). In particular, the conservation of
the norm with probability one follows from the fact that both the deterministic
evolution and the jumps preserve the norm of the wave function. The Liouville
master equation thus defines a stochastic process in projective Hilbert space.
As an example of a PDP in Hilbert space we show in Fig. 6.2 the square
10(x, t)1 2 of a single realization of an unravelling of the master equation (3.307)
for the damped harmonic oscillator discussed in Section 3.4.6. The jump opera-
tors are just the creation and annihilation operators a, at, and we have added a
time-dependent force to the Hamiltonian part of the dynamics.

6.1.1.6 Stochastic differential equation As we know from Section 1.5, a PDP

defined by a Liouville master equation of the form (6.10) can also be formulated
in terms of an equivalent stochastic differential equation. In the present case the
310 STOCHASTIC DYNAMICS IN HILBERT SPACE

latter is given by (compare with eqn (1.197))

do(t) = _iG(0(0)dt E ( 10
A (t )
( ) 1 OM) dNi(t), (6.25)

where the Poisson increments dNi(t) satisfy

dNi(t)dNi(t) = Si j dNi (t), (6.26)

E [dNi(t)] = Pyi Ai0(t)11 2 dt. (6.27)

This shows that the process Ni(t) counts the number of jumps of type i, i.e.
the number of jumps (6.23) with Lindblad operator A. The processes Ni(t) are
inhomogeneous since the expectation values of the increments dNi(t) depend on
time through the time-dependence of the state vector.

6.1.1.7 Quantum master equation To prove that the PDP defined by the Li-
ouville master equation (6.10) indeed provides an unravelling of the quantum
master equation (6.7) we have to derive the equation of motion governing the
covariance matrix (6.9). On differentiating (6.9) with respect to time we find

—aat ps(x,xi,t)= f a
ati)D0*— P[0,t10(x**(x')
at
a
E N ps . (6.28)

Here, we have decomposed the total rate of change of ps into the rate of change
induced by the deterministic Liouvillean part and the rate of change induced by
the jump part of (6.10). The first contribution is obtained from the first term on
the right-hand side of (6.10),

a
— ps

at
= f DO DO* f dy 1/)
j (x )* (x 1)

x 606(y) G(0)(Y) 60* (y) G(0) * (Y)} P{, t]

= f D'OD11)* {G() (x)5* (x') — 11)(x)G* (0)(x`)} P[11), t]. (6.29)

In the second step we have performed a functional integration by parts and used
the properties (6.3), (6.4) of the functional derivatives. On substituting (6.12)
we thus get

= (ftps _ psftt) 1 2 0(x)0 * (x i )] (6.30)

Note that the expectation values on the right-hand side represent certain fourth-
order correlation functions of O.
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 311

With the help of the second term on the right-hand side of (6.10) the jump
contribution to the rate of change of ps is found to be
a
— ps

at
=f f ND0.0(x)0* (xi) fivroAPro,t] - wtomPro,til
(6.31)

We insert (6.14) and carry out one of the two functional integrations in each
term. This yields

-yi A i psA ti — -yiE Ai0 1 2 0(x)b * (4] • (6.32)

J

If we now add eqns (6.30) and (6.32) we see that the contributions from the
fourth-order correlation functions cancel each other and we are left with a closed
equation for the covariance matrix. The latter is easily seen to take the form of
the Lindblad equation (6.7), which concludes the proof.

6.1.2 The Hilbert space path integral

As for any PDP the propagator T[0,07), to ] can be represented in terms of a
sum over all possible realizations of the process. An immediate adaptation of the
formalism developed in Section 1.5.3 shows that this yields a path integral taken
over all possible paths in the Hilbert space which connect the state 17) at time t o
the state 0 at time t. Since the stochastic process is homogeneous in timewith
we may set to E- 0 without restriction in the following.
To begin with, we write the Kolmogorov forward equation for the process as
follows (compare with eqn (1.167))
T[0, 00, 0] = T (°) [0, q, 0] (6.33)

+1 ds f D 1 DV4 f D 52DO;T0[O,Opi,s]T47 [01102]11 [02,s1V 1 0].

The physical interpretation of this equation is the same as the one given in
Section 1.5.3. In particular, the quantity

T (°) [0,0, 0] = ( 1- — 0,t]) (5 [0 — gt(01 (6.34)

is the contribution to the propagator from the path without jump, that is the
contribution that stems from pure deterministic evolution. More generally, we
write T( N) [0, t '0,0] for the contribution to the propagator which involves the
paths with exactly N jumps. The full propagator can then be written as an
expansion in the number of jumps as follows,
CKD

T[0,0,0] = (6.35)
N=0
312 STOCHASTIC DYNAMICS IN HILBERT SPACE

The N-jump contributions satisfy the recursion relation

T (N) [11) , , 0] = f ds f D0113114 f D (6.36)

x T (0) [0, OP], W [01 *2]T (N-1) [02, 8117), 0]

which holds for N > 1. It is convenient to introduce the non-unitary time-

evolution operator

(t) = exp(—ikt). (6.37)

On using eqn (6.14) and

tI ( t — s) 7h.
To) [0, 0 1, — s*111 26 [v) (6.38)
— s)6111 '
and carrying out the integration over 01 in eqn (6.36), we find that the recursion
relation can be cast into the form

0,0] = f ds f DD* -yi Û(t — s)Ai0 12 (6.39)

(I(t s)A i 0
—

T (N— ) s 1 0,
Û(t —

N-fold iteration of the recursion relation (6.39) finally leads to

tN t2
T (N) p,t173, = f dt N f f EEE (6.40)
0 0 0 iN iN -1 ii

t f# ,; Ot (tl ; • • • ; tN, i N)
X PNV, 1 1 ul; • • • ; tNI iN)S
110t(ti/ii; • • • ; t Nl i N)II]

This relation can easily be verified by induction over N with the help of the
recursion relation (6.36). The quantity

'ON (ti ,ii;... ; tNI iN) (6.41)

= Û(t — tN )A iN 0(tN — . Û(t2 — ti )A i t 0- (t i )1:6

is the state vector at time t conditioned on the following event: A quantum jump
with jump operator A t , occurs at time t1, a jump with operator At, at time
t2 , , a jump with operator Ai N at time tN , and no further jumps take place
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 313

in the time interval [0, t]. The corresponding probability density for this event,
known as the multi-time exclusive probability, is given by

PtAti,il; • • • ; iN) 'YiN • • • 7i1110N(t1 il; • • • ; N, iN)11 2 • (6.42)

According to eqn (6.40) the N-jump propagator is found by integrating the

contributions from the OtN over all jump times t 1 , t2 , , t N , and by a summation
over all jump types ii, i2, • • • , iN•
Substituting (6.40) into (6.35) yields the Hilbert space path integral (Breuer
and Petruccione, 1996a). It must be emphasized that this path integral is entirely
different from the Feynman—Vernon path integral that was introduced in Section
3.6.4.2. The latter is a sum over paths in configuration space, each path being
weighted by a complex factor given through an effective action functional. By
contrast, the Hilbert space path integral is a sum over paths OM in Hilbert
space, where each path contributes a real and positive weight factor which is
just the probability of the path. It thus provides a mathematical formulation of
Carmichael's idea of quantum trajectories (Carmichael, 1993) and of the Monte
Carlo wave function method (Molmer, Castin and Dalibard, 1993; Dum et al.,
1992; Gardiner, Parkins and Zoller, 1992; Castin and Molmer, 1995; Ms/Amer and
Castin, 1996)
If we have only a single jump operator A with corresponding rate -yo eqn
(6.40) reduces to the simpler form
t tN t2

T" ) ['/', 0, = f dtN f dtN-1 • • • f (6.43)

o
,tN)
x p tN(ti, • • • ,tN)6[11)
t (t1, ... ,tN)11]
NN

where

t ( ti, • .. ,
ON tN) = Û(t — tN)AÛ(tN — — ti)AtI(ti) (6.44)

and

PtN ( ti,... ,t N ) = -Z1 10tAt i , ,tN)11 2 . (6A5)

If we integrate T (N) [1», t, 0] over V) we get the probability pN (t) for exactly N
quantum jumps in the time interval [0, t],
t tN t2
PN(t)= f dtN f dtN_1... f dtiP tN(ti, • • • ,tN)• (6.46)

These expressions will be used in the examples of the following sections to de-
termine the statistics of jump events.
314 STOCHASTIC DYNAMICS IN HILBERT SPACE

6.1.3 Diffusion approximation

Under certain conditions the diffusion limit of the Liouville master equation
(6.10) exists and yields a Fokker—Planck equation for the probability density
functional which, in turn, is equivalent to a stochastic Schrödinger-type equation.
The diffusion limit of the PDP thus gives rise to an unravelling of the quantum
master equation in terms of a diffusion process in Hilbert space.
For the sake of simple notation we assume in the following that we have only
one jump operator A with corresponding rate N. According to the general theory
of stochastic processes a diffusion expansion of a given master equation can be
performed if the size of the transitions among the states becomes arbitrarily small
and if, at the same time, the number of transitions in any finite time interval
becomes arbitrarily large. In order to formulate these conditions we introduce a
small (dimensionless) parameter E and write the Lindblad operator as

A=I+EC, (6.47)

where I denotes the identity operator and the operator C is independent of E.

Our aim is to investigate the behaviour of the Liouville master equation (6.10)
in the limit E 0. On using eqn (6.47) we obtain to second order in E

G(0) = IN — Pyci fi + E(ct + c) + E 2 ctc} 0

i
+ - 'Yo 11 + E(C t + C) + E2(CtC)0 V) (6.48)
2
and

W[1140] = 'Yo (1 + E(Ct ± C),i,- + E2 (CfC) ) W — '0 + EM(0) +

(6.49)

where we have introduced the abbreviation

(o) v, E ( 010 1 4 (6.50)

and defined the non-linear operators

M(0) E {C — 1 (Ct + C),/, } 0 , (6.51)

1 3
N(0) E: — — { (CtC)v, — — (Ct ± Cg ± (Ct ± C)v,C} z». (6.52)
2 4

Inserting these expressions into the Liouville master equation we obtain to second
order in E
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 315

N P[o, f dx (x) K ()(x) 80 . (x) K (V))* (x)} Pk! , (6.53)

62
,y0 E2 f dx f dx / { (x)60 (x l ) M (0) (x) I (x'
(52
+
(50* (x) irk* (x')
(52
+2 (x) M (0 )* (x') P[O,
So (x) SO* (x')
This is a Fokker—Planck equation for the probability density functional. The
non-linear drift operator K(0) takes the form

K(0) = + 70E {C - Ct
1 1 1
±i -y0E 2 - (Ct C) /pC - - (0 + 0 2 — — CtC} (6.54)
2 8 2

As can be seen from eqn (6.53) the diffusion part of the Fokker—Planck equation,
involving the second-order functional derivatives, scales as NE 2 . Thus, in order
to obtain a non-vanishing and finite diffusive contribution in the limit E 0,
we assume that the E-dependence of the relaxation time -yo reads

'To = E 2 'TOI (6.55)

where is independent of E. On the other hand, the drift operator (6.54) contains
a term which is proportional to Py0E = Fy0E -1 . This term diverges in the limit
E 0 unless we impose the condition that the operator C is self-adjoint, that
is C = C. Using this condition we obtain for the drift operator

K () = H + i 0 (0C — —21 (C ) — —21 C2 } (6.56)

while the operator M(0) takes the form

M(0) = (C (C)/p) (6.57)

With these expressions the Fokker—Planck equation (6.53) is equivalent to the

following stochastic Schrödinger equation in Itô form,

4(0 = —iK(0(t))dt + -I y- o M(tp(t))dW(t), (6.58)

where dW (t) is the increment of a real Wiener process.

To get a well-defined diffusion limit we have assumed above that the jump
operator is self-adjoint. Of course, it is possible to formulate other conditions and
to perform a different expansion of the Liouville master equation. In particular,
if we have several Lindblad operators the conditions leading to a diffusion limit
may be different. A physical example will be discussed in Section 6.4.2.
316 STOCHASTIC DYNAMICS IN HILBERT SPACE

(a) Unitary time evolution

U(t, to )
10 0) minie••••m immineop.
0)
(01/1 10)
00 ) min imimono. I
U(t, to )
(b) Stochastic time evolution

11(0, 0, ti(00, \/, to]

00) 10)
2 E [(01 A1 0)]
100) ‘,./.

FIG. 6.3. Determination of matrix elements of Heisenberg picture operators for

(a) a closed system, and (b) for an open system.

6.1.4 Multi time correlation functions

-

The representation of the dynamics of a Markovian quantum master equation in

terms of a stochastic process in the underlying Hilbert space 14 can be generalized
to lead also to a stochastic formulation of multi-time correlation functions (see
Section 3.2.4). Here we show that this can be achieved by the construction of
stochastic processes in a suitably enlarged Hilbert space, namely in the doubled
Hilbert space I-1 e 1-I (Breuer, Kappler and Petruccione, 1997; Breuer, Kappler
and Petruccione, 1998).

6.1.4.1 Reduced Heisenberg picture operators According to Section 3.2.3 we

can write the matrix elements of an arbitrary reduced Heisenberg picture oper-
ator A(t) (we suppress the index H) in the following way

(001A(00o) = (OolVt(t,0)A10 0 ) = trf AV(t, 0 )00(001/, (6.59)

where A is the corresponding Schrödinger picture operator at time t o = 0 and

V(t, 0) is the super-operator describing the time evolution according to the mas-
ter equation (6.7). Equation (6.59) can be interpreted in the following way. For
the calculation of the matrix element (00 1A(t)100 ) start with the initial 'density
matrix'100)(0 0 1 and propagate it up to the time t. Then calculate the expectation
value of A with respect to the propagated 'density matrix'. However, 100(00 1
is, in general, not a true density matrix for it is neither Hermitian nor positive.
Thus, it cannot be characterized as the covariance matrix Eirti))(01] of a distri-
bution in the Hilbert space 1-1 of the open system. Hence a direct application of
the stochastic unravelling developed in the preceding sections to the calculation
of Heisenberg picture operators is not possible.
 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 317

In a closed system where the time evolution of states is given through a uni-
tary evolution operator U(t, 0) we can calculate the matrix element (001A(t)Wo)
in the following way (see Fig. 6.3): Propagate 00 and 00 independently to ob-
tain 0= U(t, 0)00 and //) = U(t, 0)00 , and evaluate the scalar product (AO).
This suggests the following method for the determination of the matrix elements
(6.59) for an open system: Instead of propagating independently the state vectors
co E 7-t and 00 E R, we design a stochastic process 9(t) in the doubled Hilbert
space, that is the direct sum
171 = (6.60)
and use this process to propagate the normalized pair of state vectors

(6.61)

This is illustrated in Fig. 6.3.

To formalize the idea we introduce the propagator T[0, t100 , to ] of the process
0(t) in the doubled Hilbert space, where we write

0(t)

Again, it is assumed that the process is normalized,

0(t) 12 = 110(011 2 + 110(0112 E- (6.63)

We can now formulate our basic definition as follows. The process OW is said to
be a stochastic representation of the reduced Heisenberg picture if the condition

(000(0100) = 2 f DODO* (0A101 -7[0,t100,0]E 2E [(0(t)12410(t))1 (6.64)

holds for all 00 , 00 , and for all operators A. Thus, once we have constructed such
a process 0(0 we can determine any matrix element of the Heisenberg operator
A(t) with the help of the expectation value of the quantity (0(t)1A10(t)).
The question is now, how can we construct a process 9(t) which satisfies eqn
(6.64)? To answer this question we consider the following master equation in the
doubled Hilbert space,

( t) = 75(0] + -yi — —21 ;lit ;I'd-J(0 — —21 j)(t)Ati , (6.65)

where the Hamiltonian and the Lindblad operators in the extended space are
defined as
_ H _ Ai 0
(6.66)
0 H) ' 0 Ai ) •
We take 0(t) to be an arbitrary unravelling of the master equation (6.65) in the
doubled Hilbert space with the initial condition 0(0) = 00 . Thus, 0(t) may be,
318 STOCHASTIC DYNAMICS IN HILBERT SPACE

for example, a PDP of the type discussed in Section 6.1.1, or else a diffusion
process of the type given in Section 6.1.3. We now claim that the process 8(t)
is a stochastic representation of the reduced Heisenberg picture, that is it fulfils
condition (6.64). Thus, any unravelling of the master equation in the doubled
Hilbert space gives rise to a stochastic representation for the matrix elements of
Heisenberg operators in the reduced space.
To prove this claim we write the density matrix in 1-1 as follows,
Fj(t) = i512 (t)
(6.67)
i521 (t) i522 (t) ) •
By assumption this is a solution of the extended master equation (6.65) corre-
sponding to the initial condition
1 ( 10o ) (63 I 1 0 0) (6.68)
P-(0 ) = 1 00) (00
100)(001 l'Oo)(00 )
By virtue of the block-diagonal structure of the operators (6.66) the master
equation (6.65) yields four independent equations such that all components i5ii (t)
separately solve the original master equation (6.7). In particular, i521 (t) solves
this equation with the initial value
1
F)21 (0 ) = (00 l• (6.69)
Thus, we have

)-621(t) = V (t, 0) (-
21100)(001) (6.70)

and, hence, by virtue of (6.59)

(0o IA(0100 2tr{ 24. /-521 (t)}. (6.71)

On the other hand, since OW is an unravelling of the master equation (6.65)

with initial condition 0(0) = 80 we have

f- i(t) = f DODO* 0)(01 110 , 3, 0i, (6.72)

and, consequently,

)-521(t) = f DODO* 10) (01 17 [0 ,t100 , (6.73)

By inserting eqn (6.73) into eqn (6.71) we arrive at

(00A(0 10o) = 2 f DODO* tr {A10 (01}T[ 0, t100 ,

= 2 f DODO* (01AMI[0 , tO o , 0], (6.74)

which concludes the proof.

 DYNAMICAL SEMIGROUPS AND PDPS IN HILBERT SPACE 319

Thus we have shown that matrix elements of reduced Heisenberg picture

operators are calculated correctly if the stochastic process in the doubled Hilbert
space unravels the extended quantum master equation (6.65). It is important to
note that the above proof does not rely on a specific unravelling of the quantum
master equation (6.65). On the contrary, it is valid for any stochastic process the
covariance matrix of which is governed by eqn (6.65).

6.1.4.2 Multi time correlations The process 0(t) in the doubled Hilbert space
-

14 can now be used to construct a stochastic representation of multi-time cor-

relation functions. Consider for example the two-time correlation function (see
eqn (3.86))

g(t,$) = (001-13(t)C(s)163), (6.75)

where t > s > 0 and we may assume, without restriction, that the initial state
is a pure state. The stochastic process which represents this correlation function
may then be defined by the following algorithm.
1. Start in the state 00 at time t = 0 and use the stochastic time evolution
in the Hilbert space 14 to obtain the stochastic wave function 0(s).
2. Propagate the normalized state
1 ( 0(s)
0(s) = (6.76)
110(s),C0(s)) 1 COO) ER

using the stochastic time evolution in the doubled Hilbert space 74 to obtain
the normalized state vector

0(t) = q:,(( t )
) ) (6.77)
e .

3. The above correlation function is then obtained by computing the expec-

tation value

g(t, s) = E [11(0(s), C0(s))11 2 (0(01B*(0)] •

(6.78)

This scheme may be generalized to the treatment of arbitrary time-ordered

multi-time correlation functions of the form (see eqn (3.87))

g(ti ...,tn; si, sm) = (001131(si)... Bm(sm)Cn(tn).. Ci(ti)10o). (6.79)

We use the same notation as in Section 3.2.4, and define the Schrödinger opera-
tors (compare this definition with the one given in eqn (3.90))

F1 = I, G1 = Ci , if ri = ti 0 83 for some i and all j,

= Bti , Gi = I, if ri= si ti for some j and all i, (6.80)
= Ci , if ri =ti = si for some i and j.
The multi-time correlation function is then obtained in the following way:
320 STOCHASTIC DYNAMICS IN HILBERT SPACE

1. Start with the state 00 at time t = 0 and propagate it to time r 1 to obtain

O (r1 ) with the help of the process in W.
2. Employing the process in the doubled Hilbert space, propagate the state

1 Fp b(r
0(ri) (6.81)
II (FiO(r , G O(ri))11 GiO(r1)
to obtain the state

0(r2 = (0(r2)) (6.82)

z )(r2)
)

Then, jump to the state

1 F20(r2)
0(7.2) = (6.83)
ll(F20(r2),G20(r2))1 G211)(r2)
and propagate it to time 7-3 , and so on.
3. Finally, the multi-time correlation is found from

(6.84)
(F10(ri),G10( 7. 1)) 1 2
(F20(r2), G21,1) (r2))11 2 (Fq -10(r —1) G b (r q — 1 )) 11 2
x (0(r 014 G q10 ( 1'0)] •
It should be remarked that this algorithm enables the determination of corre-
lation functions of arbitrary order with the help of a stochastic process in the
doubled Hilbert space. An example is shown in Fig. 6.5.
We finally note that the algorithm developed above is related to the method
proposed by Dum, Zoller and Ritsch (1992) (see also the discussion by Marte
et al., 1993a, 1993b; Molmer and Castin, 1996). A further method has been
proposed by Dalibard, Castin and Wilmer (1992).

6.2 Stochastic representation of continuous measurements

The stochastic unravelling of the quantum master equation introduced in the
preceding section yields an intuitive physical picture which may be very helpful
for the identification of the basic mechanism underlying the reduced system
dynamics. This will be illustrated by means of a number of examples in this
and the next two chapters. The method also leads to an efficient tool for the
numerical simulation of the open system's dynamics (see Chapter 7).
However, it must be realized that the stochastic representation of a given
master equation in terms of a PDP in Hilbert space is not unique. Suppose we are
given a Lindblad equation with a certain Hamiltonian H and Lindblad operaton,
Ai with corresponding rates -yi . We may then transform to a new H', and new
 STOCHASTIC REPRESENTATION OF CONTINUOUS MEASUREMENTS 321

-yii by any of the transformations given by eqns (3.72) and (3.73), without
changing the Lindblad generator. However, the corresponding PDP does change,
in general, under these transformations. For a precise physical interpretation and
for a microscopic derivation of a unique process one therefore needs an additional
input which does not enter the derivation of the corresponding quantum master
equation.
Thus, in this section we investigate the following question: What is the phys-
ical basis for the description of the reduced system dynamics in terms of a
stochastic wave function OM, and can one give a microscopic derivation of such
dynamics which is in agreement with the basic principles of quantum mechanics?
It will be demonstrated that the theory of continuous measurements provides an
appropriate framework to answer these questions, as was demonstrated by several
authors (Barchielli and Belavkin, 1991; Hegerfeldt and Wilser, 1991; Wiseman
and Milburn, 1993a, 19931); Breuer and Petruccione, 1997) (see also the review
article by Plenio and Knight (1998) and the references cited therein).

6.2.1 Stochastic time evolution of Ep-ensembles

The physical situation studied here is similar to that of an indirect measurement
scheme as already investigated in Section 2.4.6. Here, the quantum object, living
in some Hilbert space 7-is, represents an open system whose evolution we wish
to describe in terms of a stochastic wave function OM. The quantum object is
coupled to another quantum mechanical system, the environment, whose Hilbert
space is denoted by 7LB The environment is continuously monitored by some
measuring device. The environmental state thus acts as a quantum probe. The
continuous monitoring of the latter yields certain information on the object sys-
tern which can be deduced from the correlations between the object and probe
system.
In Chapter 3 the dynamics of the reduced density matrix ps of the object
system S was obtained from the density matrix p of the combined system by
taking the partial trace over the degrees of freedom of the environment B. Fixing
an orthonormal basis of state vectors (p a in 7-1B we may write the reduced density
matrix as

Ps = trBP =E(40aPkoct). (6.85)

a

The trace over the degrees of freedom of the environment can thus be viewed as a
non-selective measurement of an environmental observable with eigenvectors (pa .
In a non-selective measurement the information on the measurement outcomes is
thrown away and the sub-ensembles conditioned on the measurement outcomes
are re-mixed during the evolution. As a result the reduced density matrix ps
does not depend on the basis cp a , that is, it does not depend in any way on the
measured observable of the environment. Therefore, the time development of Ps
describes the unfolding of an ensemble of type Ep.
322 STOCHASTIC DYNAMICS IN HILBERT SPACE

However, a time evolution for an ensemble of type ep is obtained if we de-

scribe the process on the selective level, that is, if we keep the information ob-
tained by the measurement record and if we take into account the splitting into
the various sub-ensembles conditioned on it. This leads to a stochastic dynamics
for the object system's wave function /P(t). Thus, the stochastic wave function dy-
namics describes the evolution on the level of an indirect, selective measurement
which is performed on the environment and which leads to a stochastic back-
action on the state vector of the object system. This back-action has already been
discussed in connection with the general scheme of an indirect measurement. In
view of this picture it must be expected that an average over the realization of
the process OM leads to the dynamics on the non-selective level, that is to the
corresponding equation of motion for Ps (t).

6.2.2 Short-time behaviour of the propagator

To formulate these ideas in mathematical terms we take some initial time t o and
some later time t E to + 7- , where T > O. The conditional transition probability
T[5, t/, to] of the process OM is defined through eqn (6.8). Suppose that at time
t an ideal quantum measurement of a non-degenerate environmental observable
with eigenvectors (pa is carried out. The corresponding operation is then given
by the operators (compare with eqn (2.211))

9 a —= (Pal U( t) t 0)100)) (6.86)

where the environment is assumed to be in a pure state given by 00 , and U(t, to )

is the time-evolution operator of the combined system.
The situation under study is the same as the one considered in the preceding
chapter. In fact, on comparing eqn (5.88) with the definition (6.8) we immediately
see that the conditional transition probability is given by the following exact
expression,

T[0 , 0--p, to] , E llo„,-61126 [ 9a° 01 (6.87)

a 11 9aa
This is the probability of finding after a complete orthogonal measurement of
the environment at time t the reduced system in state V) under the condition
that at time t o the state 7,G was given. According to eqn (6.8) the probability
density functional P[O, t] describes the ensemble of type Ep that results from the
complete measurement on the environment at time t. The rôle of the operators SI,
is to describe the operation corresponding to the back-action on the object system
S induced by the measurement. Note that T is considered to be a conditional
transition probability in projective Hilbert space. Namely, if P[O, to] is a phase-
invariant density functional, then the density functional P[O, t] according to eqn
(6.8) is automatically phase invariant for a propagator of the form (6.87). This
can be easily checked with the help of the invariance properties of the Dirac
 STOCHASTIC REPRESENTATION OF CONTINUOUS MEASUREMENTS 323

measure. It also follows from eqns (6.8) and (6.87) that the total probability for
being in any state at time t is equal to 1, namely

f D/PDV)* T [7/), OP, to] = 1, (6.88)

and that for t to only the probability of being in the initial state is different
from zero, that is,

li m
t— *to
T O, t 11/), to = 6 [71)
[ ]
gid • (6.89)

In order to derive a PDP describing the continuous measurement one now

has to invoke the Markov approximation. For a sufficiently small time increment
T (but still T » TB) the short-time behaviour for the conditional transition
probability takes the general form pertaining to a PDP, namely

T[I1), tIO , to ] = (1 — TF[]) [z-p — iTG(0) — 71)] + 7 -T47 [714]. (6.90)

The strategy to derive the PDP is thus simply this: One first derives the short-
time behaviour of the conditional transition probability given by eqn (6.87) em-
ploying the Markov approximation. A comparison with the general structure
(6.90) then yields the transition rate W[010] for the quantum jumps, as well as
the generator G(0) for the deterministic pieces of the process.
We close this section with a few remarks. Involving the variables of the envi-
ronment and the exact time-evolution operator U(t, to ), eqn (6.87) for the con-
ditional transition probability is an exact expression. However, in most physical
applications one is interested in, the environment constitutes a system with a
large (practically infinite) number of degrees of freedom with a quasi-continuous
spectrum of frequencies, e.g. the continuum of modes of the electromagnetic
field. In such cases, it is of course impossible in practice to design a measure-
ment scheme that corresponds to a complete and orthogonal decomposition of
unity. However, as will be seen in the following examples, the Markov approxima-
tion enables one to decompose the conditional transition probability into a small
number of terms that can be interpreted as the different alternative outcomes of
an incomplete measurement scheme. The latter is easily realized physically since
it corresponds to a resolution of the identity in terms of projection operators
that project, in general, onto high-dimensional subspaces of 7-LB •
The process defined by the short-time behaviour (6.90) of the conditional
transition probability has to be interpreted as resulting from a continuous mea-
surement of the environment according to the measurement scheme that is de-
fined by the operators Q a . This interpretation is necessary because each applica-
tion of the infinitesimal generator of the process implies a state reduction which
is fixed by the measurement scheme. Furthermore, it is clear that the term con-
tinuous has to be understood in the sense of a coarse graining in time which
enables the Markovian approximation of the dynamics (compare the discussion
324 STOCHASTIC DYNAMICS IN HILBERT SPACE

of the quantum Zeno effect in Section 3.5.1). It should also be clear that one
uses at each time the assumption of (approximate) statistical independence of
the system and environment. In other words, it is assumed that the probe is only
weakly disturbed by the object system and that after each time interval of the
order of T one can apply anew the short-time expression (6.90) for the propa-
gator. It should be clear that this assumption of weak disturbance can only be
valid if the environmental state 00 is only weakly influenced by the measurement
process. This condition is satisfied, for example, if the bath state is contained in
the reduction basis (p a that fixes the measurement scheme. The examples treated
below are precisely of this type.
Above, we have assumed the Markov process to be homogeneous in time
since this is sufficient to treat the examples that will be given below. A non-
homogeneous Markov process can result if, for example, the reduced system is
coupled to time-dependent external fields, or if the state of the environment
depends on time through the time evolution generated by the bath Hamiltonian
HB. For such a non-homogeneous Markov process the generator G as well as the
transition rate W may depend explicitly on time. In the examples below we shall
treat the case of time-dependent external fields. However, since the Hamiltonian
will still be time independent in the interaction picture, the Markov processes
are time homogeneous in that picture. Essential modifications occur, however.
in the case of strong external fields (see Section 8.4).

6.3 Direct photodetection

As our first example we study in this section the driven two-level atom considered
in Section 3.4.5. At .zero temperature eqn (3.271) yields the vacuum optical
master equation
d ift
— ps(t) = — [0-± o-_, p(t)] (6.91)
dt 2

+-ro (a-Ps(t)a+ — -

2
a+a–Ps(t) –2 Ps(t)a-Fa–)

for the atomic density matrix PS (t).

The measurement scheme underlying the derivation of the PDP is the di-
rect, continuous detection of the fluorenscence photons emitted by the two-level
source. Thus, the environment is provided by the radiation modes and the vac-
uum of the electromagnetic field acts as the probe state. The continuous mon-
itoring of the photons radiated gives rise to a stochastic process for the atomic
state vector.

6.3.1 Derivation of the PDP

We write the Hamiltonian of the radiation field as follows,

HB = Ewi b-,!bi , (6.921

 DIRECT PHOTODETECTION 325

where b 3t.1 b •3 are the creation and annihilation operators of the field modes labelled
by j. The frequency of the mode j is denoted by wi . The interaction picture
Hamiltonian Hi (t) takes the form

Hi(t) = e iw° t o-+B(t) e -iw °t o- _ Bt (t) L (6.93)

The first two terms describe the coupling of the atomic dipole moment to the
radiation field, where

B (t) = (6.94)

and the kj are coupling constants. The third term yields the interaction with
the resonant driving field,
St
HL - - (o-_ a+) . (6.95)
2
Suppose that the photons of the modes bi emitted by the atomic source are
observed through direct detection by a photocounter. This means that the basis
vectors (p a must be taken to be the Fock states created by the operators bti out
of the field vacuum,

Pa E
(
(6.96)

Here, the index a stands for a complete set {Ni of occupation numbers of the
}

modes bi In particular, we define the vacuum of the electromagnetic field

(6.97)

and the one-photon states

E 1j) E bjt 1°) (6.98)

The two-level atom constitutes the quantum object, while the electromagnetic
field plays the rôle of the quantum probe. The pure probe state will be taken
to be the vacuum state, that is, we set 00 = (po E O) . This choice satisfies the
requirement of weak disturbance, since the demolition measurement of the field
quanta puts the electromagnetic field back into the vacuum state.
Our task is now to derive the operators SI, which describe the operation for
the detection of the photons. Employing second-order perturbation theory we
have

f f f
t ti

Ui(t, to ) I — i (ti) — dt"Ill(e)H1(t"). (6.99)

to to to

and, hence, eqn (6.86) leads to

326 STOCHASTIC DYNAMICS IN HILBERT SPACE

— 6a,0 + fa + gcx) (6.100)

—E(fifa,) 0 } + (fcva),, (6.101)

a'

where we have introduced the operators

fa =f de ( (pa VI A t i) 10),
to
(6.102)

=—ff
t t '

ga dt" ((p a (e)H (0)10) . (6.103)

to to

For (p a = (vacuum state) this yields

= (6.104)

and for cpa = 1j) (one-photon state) we get

f
t

';
fj = —.K; dtici(w3—wo)ti a_ (6.105)
to

Note that fa = 0 if (p a is any N-photon state with N > 2. Thus, to second

order we have to keep only the first-order term in 1l3 which corresponds to the
projection onto the one-photon sector, and the second-order term in fto which
corresponds to the projection onto the field vacuum. This yields, on using eqns
(6.100), (6.101), (6.104), and (6.105),

9./ = fig- , (6.106)

HRI 17) 11 2 = (6.107)

and

cl0 = / — iTHL + go, (6.1081

11 90` 3 2 = 1 412(a±a_4- (6.1091

3

Equation (6.106) shows that ft i is the product of two factors: The first one is f )
which is simply the probability amplitude for the observation of a photon in the
mode j; the second factor is the atomic lowering operator cr_ which represents
the back-action on the object system conditioned on this observation.
 DIRECT PHOTODETECTION 327

Collecting these results we see that the conditional transition probability may
be split into two parts

T[0,07), to] = NO, t3, to] + T1[010,13 to ],

where

To[0, t 0,to] = mo-b1126

11Q001
and

Top,o),toi = E 11Qp-b112,5 [ 1:3 a'-b- (6.112)

1.fd Ho-411
The factor h/ h1 under the argument of the (5-function is a pure phase factor. Re-
membering that the propagator represents a transition probability in projective
Hilbert space, we see that this factor drops out if the propagator is convoluted
with any phase-invariant density functional. Thus we can write

Ti [0, orP, to] = (E 1Q,i'0112

1 10-- 01 1
(6.113)
i
This expression shows that the operation conditioned on the measurement of a
photon does not depend on the mode j. The reason for this fact is, obviously,
that S--t i depends on j only through a scalar factor. Physically, this means that
the back-action on the quantum object is the same for all field modes j.
The decomposition (6.110) of the conditional transition probability corre-
sponds to two disjoint classical alternatives. T1 describes the detection of pre-
cisely one photon. The factor in front of the 6-functional in (6.113) is the prob-
ability for the detection of a photon, whereas the argument of the 6-functional
shows that the state
Gr_77)
= (phase factor) x 1g) ( 6. 1 14 )

is the new state of the atom conditioned on that measurement result. Note that
this state is just the ground state 1g) of the two-level system, that is the detection
of a photon puts the atom into the ground state.
On the other hand, To gives the contribution from no photodetection. The
probability for this event is given by eqn (6.109) and the state of the reduced
system conditioned on that result is given by

Q00
- • (6.115)
11Q 0 011
328 STOCHASTIC DYNAMICS IN HILBERT SPACE

Next we perform the Markov approximation. Under the assumption that the
electromagnetic field may be approximated by a continuum of field modes we
obtain for r» 1/wo from eqns (6.105) and (6.103)

E rrd, TN, (6.116)

1
go — — 7—You+ Gr— (6.117)
2
where we have neglected the Lamb shift, and
+00
-Yo = f dtezw 0t (01B(t)Bf (0)10) = 27rD(wo)1K(w0)1 2 (6.118)
-CC

is the atomic damping rate, D(w) being the density of field modes.
Thus we have from eqn (6.116) and eqn (6.107)

11Q.i112 T70(Gr+Gr_),z-,,
3

and from eqn (6.116) and eqn (6.109)

11 90a 2 1 — T-Yo(a+a—),
and, hence, inserting (6.119) into (6.113)

T1[0,07), to] T-Yo(g+Gr ) — [ IkuT 1,V)b-- 11

From eqn (6.117) and eqn (6.100) we get

1
= I + fo + go rrz-',- I — iTHL — (6.122)

and therefore it follows with eqn (6.120) that

QV)
o 1
{/ — iTHL — 'T`To Vf-Fa— (a+g—)4 't P- • (6.123)
dçW11
Thus, we find the contribution To of no photodetection by inserting (6.120) and
(6.123) into (6.111),

To [ b, tlz-P, to] rrz-,'- { 1 — TN (a+Gr—)il, (6.124)

1
x6 — — — (g+a- -)4} d•
If we finally add eqns (6.124) and (6.121) we obtain the short-time behaviour
of the conditional transition probability T[0,0), to]. The latter is seen to be pre-
cisely of the general form (6.90) characteristic of a PDP in Hilbert space. In fact.
 DIRECT PHOTODETECTION 329

from a comparison with (6.90) we infer that the generator of the deterministic
evolution periods reads

G(0) = fro + H Gr_ 2 7/),

with the non-Hermitian Hamiltonian

I/ = HL —
2
The transition functional takes the form

W[]= 'y0 2 (6.127)

11g-011
Accordingly, the corresponding Liouville master equation (6.10) describes a PDP
with only one type of quantum jump

= (6.128)
I la-011
which occur with a rate, i.e. probability per unit of time, which is given by
N 1 2 . These jumps represent the back-action on the object system: Condi-
tioned on the detection of a photon, they put the atom into its ground state.
Finally, the equivalent stochastic differential equation (see eqn (6.25)) now
reads explicitly,

d(t) = (HL — 1 -1-

2° — 1 ,7-0(t)11 2 1) (t)dt

71)(0) dN(t), (6.129)

where, as usual, dN(t) 2 = dN(t) and E [dN(t)] = -Yo la— V) ( 011 2 dt.
Following the reasoning of Section 6.1.1 we conclude that the PDP con-
structed above leads to a covariance matrix

Ps(t) = E 00(0)(0(01) (6.130)

which satisfies the Bloch equation (6.91) for the density matrix. As an ex-
ample we depict in Fig. 6.4 the simulation results for the stochastic quantity
(0/0)(0(t) le). The figure shows the average of this quantity taken over a sam-
ple of many realizations of the PDP. The average provides a statistical estimate
for the off-diagonal element (glps(t)le) = (o-+ (t)) of the density matrix, as is
demonstrated by a comparison with the solution of the Bloch equation.
As a further example we show in Fig. 6.5 the results of a numerical simula-
tion of the stationary four-point correlation function (o-±o-±(T)o-_(T)o-_). This
correlation has been estimated from a simulation of the corresponding stochastic
process 0(t) in the doubled Hilbert space IC as explained in Section 6.1.4.
330 STOCHASTIC DYNAMICS IN HILBERT SPACE

0.3

0.2

0.1

—0.2

5 10 15 20 25 30 35 40
f2 t

FIG. 6.4. Simulation of the PDP for direct photodetection of a driven two-level
system interacting with the field vacuum. The solid curve represents the
imaginary part of the density matrix element (gdps(t)1e) = (o-±(t)) obtained
from the analytical solution of the quantum master equation (see eqn (3.289))
for -yo gt = 0.1. The symbols give the average taken over an ensemble of 104
realizations of the PDP.

6.3.2 Path integral solution

The Liouville master equation for the PDP of direct photodetection may be
solved by means of the path integral technique of Section 6.1.2.
Let us concentrate in the following on the underdamped case St > -y0 /2.
and consider the special initial condition 0(0) = 1g). Using eqn (6.126) in the
definition (6.37) we find

= e--yot/4
- cos pt 71 sin pt
— 1 iZ sin pt
(6.131)
ie sin pt cos pt +—
7°
LIA sin pt

where we haye introduced the parameter p which is defined by

(6.132)

With the help of eqns (6.21) and (6.131) the waiting time distribution function
is easily found to be
 DIRECT PHOTODETECTION 331

0.6

0.0
0 1 2 3 4 5
'MT

FIG. 6.5. The correlation function (ci-±c)-±(7)a_ (7)0-_) for a driven two-level
atom on resonance with Rabi frequency St = 10-yo . The figure shows the
simulation results for a sample of 105 realizations of the PDP in the doubled
Hilbert space (thin line and error bars), and the analytical solution (thick
line).

F[10 t1 = 1- 1101011 2
, (6.133)
-rd sin-2 tit +'---(°,t sin tit cos pt) .
1 - e -0t l2 (1+ -8/7 -

The corresponding density f(t) is given by

d 02
7Q
f(t) E - F{19),ti
dt
= 70 11u-0(01 0 11 2 = 4112 e -7ot/2 sin 2 pt. (6.134)

These are the expressions for the waiting time distribution for direct photode-
tection of resonance fluorescence (Carmichael et al., 1989). Note the oscillating
behaviour of the density f(t), as shown in Figs. 6.6 and 6.7. The comparison
with the corresponding exponential distribution shows that the probability of
small waiting times is strongly suppressed. This is obviously due to the fact that
the atom is in the ground state immediately after a quantum jump. It therefore
takes some time to get excited again with appreciable probability and to be able
to emit another photon.
Let us investigate the path integral representation (6.43) for the N-jump
contribution to the propagator. First, we infer from eqn (6.44) that
332 STOCHASTIC DYNAMICS IN HILBERT SPACE

0.09

0.08 ■

0.07 -

0.06

0.05
..-
0.04 -

0.03 ■

0.02 •■ • ■

0.011

0
0
)

10 15 20 25 30 35
/40
S2t

FIG. 6.6. Plots of the density f (t) of the waiting time distribution according
to eqn (6.134) with 'yo/ 1 = 0.1 (solid curve), and of the corresponding den-
sity for an exponential waiting time distribution with the same mean value
(dashed curve).

ON (t 1 , — - , tN) Û (t — tN)19)
(6.135)
kb N
t (tl, . - . , tN)ll HO (t — tN)19)11
apart from an irrelevant phase factor. Moreover, eqn (6.45) gives the multi-time
exclusive probabilities

PtN(ti, t2 , - - - , tN) = PO (t — tN)f(tN — tN-1) • • • f (t2 — h.) f (ti), (6.136)

where f (t) is given by eqn (6.134) and

po (0 = 1 — F[1g) , t] = HO (t)1011 2 (6.137)

denotes the no-jump probability, that is the probability that no count is observed
in the time interval [0, t]. Thus, with the help of eqn (6.43) we find the following
compact form for the full propagator

U(t)g)
T[Olt19)10] = po (t)6 [0 1 (6.138)
11U(t)Ign

±E
co

N=1 0
f f dtN
0
dtN_, .. . f
0
dtiPtN(ti. , - - - ,tN)6 k ,
0 (t — tN)Ig)
HU (t — t AO
I
 DIRECT PHOTODETECTION 333

0.35
\

0.3 ■

0.25

0.2
,
0.15
\
\
\

0.1 \
\
\
,
,

,
0.05

10 15
2t

FIG. 6.7. The same as Fig. 6.6 but for -yo /ft = 1.

It is evident from these formulae that the complete statistics of the counting
events can by reconstructed from the exclusive probabilities pt (t 1 ,t2,... ,tN). In
particular the probability of N counts in the interval [0, t] may be found from
eqn (6.46). If we introduce the Laplace transformed quantities
00
fiN (A) =
f
o
dt exp( At)p N (t),
— (6.139)

and

f () = f dt exp( At) f (t), — (6.140)

o
eqn (6.46) together with (6.136) leads to

= 1 if(A) [] N , (6.141)

where we have used the fact that the Laplace transform of the no-counting
probability is given by (see eqn (6.137))

/50(A) = 1 (1 — 1(A)) . (6.142)

334 STOCHASTIC DYNAMICS IN HILBERT SPACE

0.2

0.18 -

0.16

0.14 -

0.12 - -

--.
:z 0.1
--
- ^

.
\
0.08 \
-

0.06 - -

0.04 - -

0.02 -

% 5 10 15 20 25 30 35 40
N

FIG. 6.8. The probability p N (t) for N photon counts in the time interval [0, t] for
the driven two-level system. The figure shows the simulation results obtained
from an ensemble of 10 3 realizations of the PDP for direct photodetection
(histogram). The solid line represents the Poisson distribution with the same
mean value as p N (t) . Parameter: -y0/S2 = 1 and St t = 50.

The Laplace transform of the density f may be found with the help of (6.134),

--- 70Q2
J(À) = (6.143)

where we have introduced the abbreviation A = A + -y0/2. Thus, the Laplace

transform of the probability p N (t) can be written as

(1,yos-2 2)N [v ± .i.,_yoA ± s-22]

fiN (A) = 2 N+1 ' (6.144)
[A (A 2 + (2p)2)]

This is the photocounting formula for direct photodetection of flourescence ra-

diation derived by Mollow (1968, 1969, 1975).
Figure 6.8 shows the photocounting probability p N (t) for a fixed time t as a
function of N. The picture has been obtained from a numerical simulation of the
PDP by counting the number of quantum jumps in a sample of realizations of
the process. The picture also shows the corresponding Poisson distribution with
the same mean value. The sub-Poissonian character of p N (t) is clearly visible. It
may be characterized with the help of the Mandel Q paramter defined by
 HOMODYNE PHOTODETECTION 335

1HL(t )

b3

source

local oscillator

FIG. 6.9. Experimental setup of homodyne photodetection. The source is a

two-level atom with frequency coo which is driven by an external interac-
tion HO). The light emitted by the source (modes bi ) traverses a beam
splitter and is detected at detectors D I and D2. The emitted light interferes
with the light field from the local oscillator (modes ei ). The detectors register
the quanta of the modes ch i and d2i defined in the text.

Var(N(t))
Q= 1. (6.145)
(N (t))

For a Poisson distribution we have Q = 0, while Q < 0 characterizes a sub-

Poissonian distribution. For the simulation shown in the figure the Mandel Q
parameter was found to be Q R.,- —0.69 which is in good agreement with the
analytical result (Mandel and Wolf, 1995).
The sub-Poissonian character may be understood from the waiting time dis-
tribution shown in Fig. 6.7 for the same parameters. The density f(t) for the
PDP exhibits a pronounced bump such that the variance of the random waiting
time is less than that of the corresponding exponential distribution. In particular,
the photon emission events exhibit a strong antibunching effect. The distribution
pN(t) of the photon counts is therefore sharper than the corresponding Poisson
distribution which would be obtained from an exponential waiting time distri-
bution.

6.4 Homo dyne photo detection

As a second example we consider in this section a driven two-level atom whose
emitted light is detected by homodyne photodetection (Walls and Milburn, 1994;
Wiseman and Milburn, 1993a, 1993b). A schematic picture of the setup is shown
in Fig. 6.9.
336 STOCHASTIC DYNAMICS IN HILBERT SPACE

6.4.1 Derivation of the PDP for homodyne detection

The Hamiltonian HB of the environment is taken to be

HB = E wi bti bj E wj ci ci . (6.146)

Again, the bi-modes are the radiation modes that couple directly to the atom.
Furthermore, we have added a term that contains the modes e3 of the local
oscillator whose light interferes with the radiation emitted by the atom. We take
into account the presence of the local oscillator by a coherent driving of the
ci -modes. The interaction picture Hamiltonian now takes the following form,
Hi (t) = e -iw° t o- _Bt (t) + e iwc'to-±B(t) + Oe -iw°tCt + 0* eiw° t C(t) + HL,
(6.147)
where HL is defined in (6.95) and B(t) is given by (6.94). The quantity j3 repre-
sents the amplitude of the local oscillator field and

= E
Ki e _tw•t
3 Ci (6.148)

As for direct photodetection, the electromagnetic field serves as the quantum

probe, the pure probe state being the field vacuum Oc, = 1 0).
The crucial point of the derivation of the stochastic process for homodyne
photodetection is the correct choice of the basis (pa in the reduction formula.
What is measured at the detectors D I and D2 (see Fig. 6.9) is the superposition
of the light emitted from the atom and the light of the local oscillator. Under
the assumption that the beam splitter has transmittivity and including a
phase shift of r/2, the field quanta that are detected at D 1 and D2 are defined,
respectively, by the annihilation operators
1
d 1,3 = 3 + ici),
(b• ic•) (6.149)
1
d2,, = (bi — (6.150)

Note that these operators satisfy the usual boson commutation relations

= 5 kki 5jj' (6.151)

where the index k = 1,2 labels the two detectors D I and D2. Thus, d 3
be interpreted as the creation operator for a photon in mode j at detector Dk•
Consequently, we take as the basis (pa the Fock basis which is generated by the
creation operators d out of the field vacuum, that is, we write

( Pa = 1{1\711) ; {N1 2) } ).
} (6.152)
 HOMODYNE PHOTODETECTION 337

The index a now represents two sets of occupation numbers: { e ) } is the set
of occupation numbers of the dk J .-modes, where k = 1, 2. In particular, we have
two types of one-photon states given by

(Pk,i E ik, i) d1 0 ), k = 1, 2. (6.153)

We can now proceed in close analogy to Section 6.3. The first step is to
determine the operators ft c, which yield the operation corresponding to the mea-
surement scheme. These operators are given by eqn (6.86), where for the states
'Pa expression (6.152) must be used. We shall use again the definitions (6.102)
and (6.103) for the operators fa and ga . We then find that for the vacuum state
fo = —iTHL. For a one-photon state (Pa = ik,i), however, we have
,

fk,3 = —i f de c i wot i {(kilBt(ti)p)a_ + (k, (t1 )1 0 )0} (6.154)

to

The matrix elements in the above equation are easily calculated to be

jiBt (t` )10) = , (6.155)

and

(k, (t1 )10) – 1 e eiwit' (6.156)

'4 3
where we use the convention that the upper sign corresponds to k = 1 and the
lower sign to k = 2. Hence we obtain

==h07- (6.157)

where

E — de e -w0)t'
(6.158)
3
to

On using these results we can decompose the conditional transition proba-

bility into three parts as follows

T[0, 0 , t o ] = TO[O, tO1 T1[O, 01411 ± T2[0, til-P1t01 , (6.159)

where
338 STOCHASTIC DYNAMICS IN HILBERT SPACE

Tokb, 0- , to] = 11 901-P11 26 [ 11990111 01 (6.160)

Ti[lp,tk-b,to] = I 91,j02) [11((cru-- ++ i /30))011

(6.161)

[ (u_ - (6.162)
T2[0, tki, , to] = (E11 9 2,11 2 ) 6
[I1(a-
The factors in front of the 6-functions are given by

imo/A2 = 1 — (6.163)
k,j

EII1i,jII = = E 1.61 2 ((a_ + ifi) t (Gr- + ifi))7„ (6.164)

Elp2,i112 = =E _ is) t(c,_ _ (6.165)

3

The above decomposition of T into three terms expresses three classically dis-
joint alternatives. Either no photon is detected (contribution To), or a photon
is detected at D 1 (contribution T1 ), or a photon is detected at D2 (contribution
T2
In the Markov approximation we find in the same way as in the previous
section that

E fi I 2 r''j 770 (6.166)

and, therefore,

E ii3 ) t (Gr-± 0)4, (6.167)

IIoii 2 rrz,- 1 — r o {(cr±g_),ÇT + (6.168)

Moreover, we have
1 r
go 770 ig-Fa- +101 2 1 (6.169)
2
Thus we get

Qoi4 1
{i - iTHL - - 77o (a+a- (6.170)
11 9011 2

Note that the terms involving 1012 cancel each other and that we therefore find
the same expression (6.123) as in the previous section.
 HOMODYNE PHOTODETECTION 339

Summarizing these results, we finally obtain

Tokblikp- ,toi = {1 — r-ro [11a-11 2 + IsIl } (6.171)

xs [{I — iTHL — - -/--yo (cr+ o- _ -

Ti [0,tk - >to] = 'T-yoll(Gr_ + is)cP112 6r (a- ±ifi)'-' 01, (6.172)

LII(g- +i/3)J11

T2[0,G,t0] = y-roii(Gr- — is)-bil2(5 (6.173)

[11( Gru - — i )6 )2iI1 d
We add these three contributions to get the short-time behaviour of the propa-
gator and compare with the general form (6.90). This shows that the generator
G(0) for the deterministic evolution periods of the PDP is of the same form as
in eqn (6.125). However, the transition functional is now given by

+i
T47[01/-P] = l'011(0-- + ifi)-b1126
[ii((gor- + i )310))11 d

+ -1'Yoli(a-- - ifi)' -bii 2 (5 r (Gr- 01 . (6.174)

2 L - 0)01i j
The Liouville master equation thus describes a piecewise deterministic process
with two types of quantum jump, corresponding to the detection of photons at
DI or D2,

(g- 1 ifi) -b 1-yoll(o-_ ± iSi611 2 . (6.175)

ii(g- 1 i0)011 2
The deterministic pieces of the PDP are obtained from the same differential
equation as for direct photodetection.
The equivalent stochastic differential equation for the PDP describing homo-
dyne detection takes the form

d(t) = -i (f/ + llia_0(t)11

i' 2 ) 0 (t)dt (6.176)

+ ( (a- ± OM) dNi(t) + ( (a- - 'OM) dN2 (t).

\ii(a- ± 0*(t)ii 1 g- - i0)0(t)ii
The processes Nk(t) count the events at detector Dk and satisfy
dNk (t)dNk, (t) = (5kk, dNk (t), (6.177)
and

E[dNi(t)] = 1.11(0-- + 0)0(011 2 dt, E[dN2(t)] = 22 11(0-- - i0)0(011 2 dt.

(6.178)
340 STOCHASTIC DYNAMICS IN HILBERT SPACE

0.5 -0.5

-0.5 0.5
V
V 2
1

FIG. 6.10. Diffusion on the surface of the Bloch sphere. The picture shows a
single realization of the random Bloch vector 11 (0 = (0(t) 16- 111)(t)) for the pro-
cess OM given by the stochastic Schrödinger equation (6.181). Parameters:
0.1 and 0= 0.

6.4.2 Stochastic Schrödinger equation

The PDP for homodyne photodetection yields a certain stochastic Schrödinger
equation in the diffusion limit. To derive it we write

i3 = iifileia (6.179)

and investigate the limit E E VIA -* 0 of a strongly excited local oscillator.

This can easily be done by invoking the results of Section 6.1.3 on the diffusion
limit of the Liouville master equation. We first observe that, instead of u_ ± ifi,
we may use the two jump operators

I ± EC E I ± 5 e -ie o- -, (6.180)

if we scale the damping constant by a factor of 01 2 , that is if we replace 70 —*

'Yo1012. Then we have a situation which is similar to the one discussed in Section
6.1.3, where, however, we now have two jump operators. Since C E exp(-i9)o-_
 HOMODYNE PHOTODETECTION 341

V V
1 2

FIG. 6.11. 103 realizations of the random Bloch vector 17(t) =

plotted at three fixed times given by Qt = 1.5 (dots), Ot = 3.0 (crosses), and
Ot = 4.5 (stars). All trajectories started out at the south pole of the sphere,
Parameters: -yo /O = 0.1 and 60 --= 0.

appears with a positive and with a negative sign, the linear part (in e) of the
drift generator (6.54) of the Fokker—Planck equation drops out and a well-defined
diffusion limit is obtained without restriction on C.
The resulting stochastic Schrödinger equation for homodyning, first derived
by Wiseman and Milburn (1993a, 1993b) and by Carmichael (1993), may be
written as follows,
d(t) = —i,K(0(t))dt + Ar-W1 M(0(t))dW(t), (6.181)
where dW(t) is a real Wiener increment. The drift operator K(0) reads

K(0) = Ha + il- {(C + CC — Ct C — ±11 (C + C1 ) 20 } 0, (6.182)

while the noise term is given by

1
M() , { c - -2 (C +00} 0, (6.183)
342 STOCHASTIC DYNAMICS IN HILBERT SPACE

By construction, the stochastic Schrödinger equation (6.181) exactly conserves

the norm of the wave function. It leads to a stochastic differential equation for
the random Bloch vector

6(0 = (0(01 6 10 (0), (6.184)

describing a diffusion process on the surface of the Bloch sphere. A single real-
ization of this process is depicted in Fig. 6,10. We observe that the noise induced
by the homodyne measurement scheme drives the Bloch vector out of the plane
defined by y1 = 0, Satisfying the Bloch equation, the expectation of e(t), how-
ever, remains in this plane as illustrated in Fig. 6.11 which shows the evolution
of an ensemble of realizations.

6.5 Heterodyne photodetection

The measurement scheme for heterodyne photodetection is the same as that for
homodyne detection, the only difference being that the frequency of the local
oscillator is detuned from the system's frequency wo . In the notation of the
previous section this means that we have to replace the amplitude of the local
oscillator fi by

fit = 0 exp(-iOt), (6.185)

where 0 is the finite detuning of the local oscillator (not to be confused with the
Rabi frequency).

6.5.1 Stochastic Schrödinger equation

We perform this replacement directly in the stochastic Schrödinger equation
(6.181) for homodyne detection, which yields the drift operator

K(0) = HL O- i - (e -24"9-ç2t) u - + u+)00" (6.186)

1 fi,, \ 2 , -- 240 — Qt) ± ( 0.420e 2i(O — C2t) ±
_L _ 2(u ) 0 (a+ ) 01 } 0,
m 4 i\ u i '°

while the noise term becomes

M(0) - {e -i( "w) o- - 1 (e -m) o-_ + ei(0-42t) o-± ) v, 0. (6,187)

Following Wiseman and Milburn we assume that the detuning is much higher
than the system's characteristic damping rate which is of the same order as the
interaction strength with the driving field, i.e. we assume

(6,188)

Hence, we may consider a time interval At such that I2 At » 1 but 'Yo At < 1.
Now we integrate eqn (6,181) with (6.186) and (6,187) over the interval At, and
 HETERODYNE PHOTODETECTION 343

neglect terms of second order in -yo At. We consider the drift and the diffusion
terms separately. The integration of the drift term leads to
t+At
f dSK(0) (6.189)

= HI* At - ill) {a a- At - t f
+Atds(e -2i(6-118) o- ) 0 cy_ - t+At ds(0-±),0"_
2 ± -
t t
td-At t+At
1
± f ds[(o-_) 20 e
4 + /5+ ,2 240—Os)1 ± _1
k hpe
2
f ds( cr—)0( 0-40 1/5 -

Since the first and the third time integral in this expression are of order -yo /O
they can be omitted since by assumption It » 70 . So we are left with
t+At

f dSK(0) = [Ha — i 1 (0-±a_ — (0"±)00-_ (0- (0"±),0) 0] At.

(6.190)
Accordingly, the integration of the noise term leads to
t+At t+At

f M(tP)dW(S) = f dW (s)e — i(Û—Ç2s) cr_O (6.191)

t+At

f dW (s)(e — ' ( " s) a_

Introducing the two quadratures

(e ° a_ + io
1

X 0 E. — (6.192)
2
170 E —2 ( o-± — o-_) , (6.193)

we can write
t+At t+At
dW (s) cos(0s) { C'9 cr_ — (X0)0} 0
f M()dW(S) = f (6.194)

id-At

f dW(S)sin(0,3) 07,94}
344 STOCHASTIC DYNAMICS IN HILBERT SPACE

The integrals occurring in the above equation suggest we define two new Gaussian
variables
t+At
8W(t) E N7.- f dW (s) cos(0s),
,

t+At
8W(t) E — V dW (s) sin(f2,9).
t
It is easy to show that to zeroth order in Sr i , 8147x and 8 W y have zero mean
and satisfy (Wiseman, 1994)

E [8Wq (t)(5Wq, (e)] = 8q , q 1 (At - It - el)0(At - It - el), (6.197)

where q and q' stand for x and y and e denotes the Heaviside step function. When
taking the continuum limit -y0 /0 0 the quantity At is infinitesimal on the
system's time scale, i.e. At —> dt , so 8W q (t) can be regarded as an infinitesimal
Wiener increment dW q (t) satisfying

dT47q (t)dW q i (e) = 8q , q , dt. (6.198)

For the averaged noise contribution (6.194) we may therefore make the replace-
ment
t+At
1 .
f MetP)dW(s) — { e-i° a _ - ( (,9),p 1 OdWx (t) (6.199)
t
1
± — {-ie -i° o- _ - (Ye) 0 } OdWy(t)-
-
Summing up the results obtained so far the stochastic Schrödinger equation
pertaining to heterodyne detection reads
1
4)(0 = -41-ladt - _ - @+ )a- + - (a_ )0 (o-+ )) 0 dt (6.200)
2 2

+ V eyo=2 (e .
° a_ - (X 0 ) 0 ) OdWx (t) - \17° (ie-i° o-_ + (Y0 ),p)odwy (t),
2
As is easy to verify this equation is norm preserving. It is also straightforward
to check that it is equivalent to the following stochastic equation for the unnor-
malized wave function

cirCb = {-iHL - lo-± a_ + -yo (o-± ) o u_ } Odt + NrfoCT_OdW(t), (6.201)

where d'/3 is the increment of the unnormalized wave function and dW (t) is the
differential of a complex Wiener process
 HETERODYNE PHOTODETECTION 345

1
dW(t) :_--7_ — (dWx (t) – idWy (t)). (6.202)
V2
F-7
Note, that the right-hand side of eqn (6.201) involves the normalized wave func-
tion 0 (t).
The examples considered in the present and the previous sections clearly
illustrate the general concept behind the stochastic representation of continuous
measurements. As we have discussed already the stochastic state vector describes
the dynamics of an ensemble of type Ep for the object system, namely the atomic
source. The output light field is observed by different detection schemes which,
consequently, yield different stochastic processes for the atomic wave function:
Information on the quantum object is extracted in different ways by monitoring
the environment and, thus, also the back-action on the object turns out to be
different. On the other hand, the evolution of the corresponding ensemble of type
Ep is given, in all cases considered above, by one and the same density matrix
equation, namely by the vacuum optical master equation (6.91). This reflects
the fact that in all cases we have the same local coupling between atom and
radiation field.

6.5.2 Stochastic collapse models

It might be important to mention at this point that stochastic Schrödinger equa-
tions of the type of eqns (6.181) and (6.200) have been proposed, in an entirely
different physical context, by Pearle (1976, 1989), Ghirardi, Rimini and Weber
(1986), Ghirardi, Pearle and Rimini (1990), Gisin (1984, 1989), and by Gisin and
Percival (1992, 1993). Namely, these authors suggest modifying the Schrödinger
equation on a fundamental level by the addition of non-linear and stochastic
terms. Various models have been proposed, such as the continuous spontaneous
localization model (eqn 6.181) and the quantum state diffusion model. The aim
is to explain the non-existence of the superposition of certain macroscopically
distinct states. The random terms in the Schrödinger equation lead to a dy-
namical destruction of such superpositions in a way that macroscopic objects
are practically always in definite localized states. In these theories the destruc-
tion of superpositions is a real process going on in physical space and time. It
is not, as in environment-induced decoherence, a result of the tracing over an
external world that became entangled with an open system. Thus, the funda-
mental modifications of quantum mechanics based on random Schrödinger-type
equations provide dynamical collapse models in which the stochastic state vector
describes an individual quantum system and enables some kind of macro-realistic
interpretation.
Stochastic collapse theories of this kind are not the interpretation followed
here. As discussed in Chapter 5 stochastic Schrödinger equations of the form
derived above admit a clear interpretation as the evolution of an Ep-ensemble.
Nevertheless, it might be instructive to show that a stochastic Schrödinger equa-
tion of the type suggested in the quantum state diffusion model, which was de-
veloped by Gisin and Percival (Percival, 1998), can be derived from eqn (6.200)
346 STOCHASTIC DYNAMICS IN HILBERT SPACE

by an appropriate phase transformation.

We begin by noting that eqn (6.200) can be rewritten with the help of the
complex Wiener differential as

do(t) . — iHL odt — ( Ct C — ( C t )0 C + ( C)0 (00) 0dt

+.\/-)T, (C — (C)0)0dW(t)
\F—
To
+ — ((C)odW(t) — (Ct) o dW*(t)) 0, (6.203)
2
where we again use C ,_=_ exp(-0)o-_. The first noise term in this equation is
already of the form of the noise term in the stochastic Schrödinger equation
proposed in quantum state diffusion (see below). Therefore, we will now try to
eliminate the second noise term with the help of an appropriate phase transfor-
mation. To this end, we try the ansatz

(t) --,-- exp[iço(t)]0(t) (6.204)

with the stochastic phase

t
i
(PM -= —V5T) f [(C)0(ti)dW(t i ) — ( 0 )0(t9c/W * (t/ )] - (6.205)
2 0

Since the phase is a real random number the phase transformation preserves the
norm of the wave function and leaves unchanged the expression for the density
matrix. We now determine the stochastic Schrödinger equation governing the
dynamics of zi) using the rules of the Itô calculus.
Following our previous discussion the differential of the phase is

i
4)(0 --= — Nfr—
Yo ((C) oodW (t) — ( 0 )0(t)dW * (t)) - (6.206)
2

Accordingly, the differential of the transformed wave function '/' is given by

d't" (t) _,--- exp[iv(t + dt)]0(t + dt) — exp[iso(t)]0(t)

--, exp[i(v(t) + dso(t))](0(t) + &NO) — exp[io(t)]0(t)
--= exp[îo(t)] lexp[idv(t)](0(t) + th/'(t)) — OW} (6.207)

and, hence,

d(t) , exp[iv(t)] {(exp[idv(t)] — 1) 0(t) + exp[idv(t)]4(t)} . (6.208)

Since the differential dcp is of order VT/t we have to expand (exp[idso] —1) to second
order. Conversely, since c/0 already contains terms of order NATt it is sufficient
 HETERODYNE PHOTODETECTION 347

to expand the exponential function in the term exp[i4]chP to first order. To the
desired order eqn (6.208) therefore reads

1
cksb = (i4 — (ciço) 2 ) 't + exp[iv] (1 + i4) cbt,b . (6.209)

Upon inserting eqn (6.203) into the above equation we get

cirt-p (id p — 1 (d(p) 2 ) ,Cb — illv t'spdt (6.210)

—'1 (C 1 C — (C1 )C + dt

+V-ITIe-ie (c - ( C ) 't- dW(t)

)

\1 7) ((n ly-) dW - ( 0)0W*) ,Ip- + ,'' icLiod,O.

+4
The terms appearing in the above equation are easily evaluated. It follows from
(6.206) that

i d c,o0 = — 2 ((C)c/W — (Ct)dW ) (6.211)

and this term is seen to compensate the second and the third noise term in eqn
(6.210). Equation (6.206) leads to

170
-- (ct,o)Vp = --(Ct) -(C) dt, (6.212)
2 4 0 0
because of the properties of the complex Wiener increment. Furthermore, we
find

exp[ic,o]i4d0 = 1(0 ) CI- dt. (6.213)

Inserting eqns (6.211), (6.212) and (6.213) into eqn (6.210) we finally obtain

_ _21 ct c 21
ct(b = —ilivt p- dt + -yo ((Ct 4C (Ct),-p-(C)) z-pdt

— (C)0) 1 P- clW, (6.214)

which is just the stochastic differential equation of the quantum state diffu-
sion model involving a single Lindblad operator C. We thus conclude that the
stochastic Schralinger equation of quantum state diffusion, which has been pro-
posed originally as a model describing dynamic state vector localization, appears
here as an equation which represents the stochastic dynamics of heterodyne pho-
todetection in the diffusion limit.
348 STOCHASTIC DYNAMICS IN HILBERT SPACE

6.6 Stochastic density matrix equations

The stochastic processes considered so far describe the dynamics of ep-ensembles
of pure states 0(0. In a completely analogous manner we can also derive the pro-
cesses for the corresponding ensembles of mixtures. This will lead us to equations
of motion for the time-dependent distributions P[o-, t] on the space of density ma-
trices a or, equivalently, to stochastic density matrix equations.
Probability distributions on the space of density matrices have already been
introduced in Section 5.3. We have also discussed there the description of selective
measurements. Our starting point will be eqn (5.84) which, applied to the present
case, yields the new distribution P[o-, to +-7-] of the object system at time t = to-FT
as it results from the measurement on the probe system after the interaction
between object and probe during the time interval T. Thus, eqn (5.84) leads to

P[o- , to + r = f 14),,(6-)1 6 [ (6.215)

tr f.t m (a)} a P[6- '
]

In view of this relation the conditional probability for the density matrix at time
to + T to be o- under the condition that the density matrix at time t o is 6- takes
the form
ck m (&) al
T[o, to + TI6- , to ] tr {m(&)} (6.216)
[tr
{.21),,(5-)1

The next step is again the determination of the operation ci)„, pertaining to
the measurement scheme under consideration. As an example, let us consider
direct photodetection. As in Section 6.3 we have to distinguish two cases, that
is the index m takes on two values m = 0, 1. The operation

'to (51 = trB flroUi(t, to) ( 6- 0 1 0 X0 1) (t, to)70} = ooa-oto (6.217)

corresponds to the case that no photon is detected, where 70 = 10)(01 is the

projection onto the vacuum state which serves as the probe state. The other
case is that a photon is detected (note that in second-order perturbation theory
there are no contributions from sectors with a higher number of photons). The
corresponding operation is

(6") = trB Ui(t, to) ( 6- 1 0 ) (0 1) Uit (t, to)7ri = (6.218)

where 71- 1 = E3 Ii)(il

denotes the projector onto the one-photon subspace. In
the same manner as before (see eqns (6.106), (6.116) and (6.122)) we find that
in the Born Markov approximation 13
—

13 To avoid confusion with the random density matrix o- we write here A for cr_ and At for
o-+.
 STOCHASTIC DENSITY MATRIX EQUATIONS 349

1
Oo r& I — irf-IL — — r-yoAtA, (6.219)
2
Qi r& JA, where E 1412 ,., ryo . (6.220)
i
The second equation shows that the operation for the measurement of a photon
can be written

Li2A6-At ry0 A5- At. (6.221)

Hence we find that the propagator (6.216) is again a sum of two terms, T
To + T1 , expressing disjoint classical alternatives: To corresponds to the event
of no photon detection, and T1 to the event that a photon is detected. These
contributions are given by

00'W°
To[o-, to + to] = tr Ot0 005- }
tr {05- }
(1 — 7---yotr {AtAei- }) (5 [a- - - , (6.222)
and

[c r to +
, to] tr {(1) 1 (&)}S
[ tr {(1).; 1(6-(5-) )1 a]
[ A6-At
r-yotr{AtA6- } S (6.223)
tr {AtA 6-- } a] -
Here, we have introduced the non-linear super-operator

G(0- ) E flu — crilt + iyotr {iltilo- } (6.224)

where (see eqn (6.126))

f-I = (6.225)
2
It is clear that the above form for the propagator yields a PDP for the
stochastic density matrix a(t), where g(a) generates the deterministic parts
of the evolution. The equivalent stochastic differential equation for the PDP is
given by

Aio- (t)At
do- (t) = — iG (o- (t))dt + (t) dN (t), (6.226)
tr { At Au (t)}
where the Poisson increments have the properties,
dN (t) 2 = dN (t), E [dN(t)] = Pyo tr {AtAo- (t)} dt. (6.227)

The stochastic process a(t) is easily seen to have the following property. If
we suppose that we have a pure state a(t) = 10(t))(0(t)1 at some time t, then
350 STOCHASTIC DYNAMICS IN HILBERT SPACE

purity is preserved under the stochastic time evolution. For such a case, the
PDP thus describes a stochastic state vector evolution which is precisely the one
derived in Section 6.3, namely the one given by eqn (6.129). This property of
the stochastic density matrix equation is due to the fact that we have just one
Lindblad operator and, thus, the operation 4. 1 corresponding to the detection of
a photon transforms pure states into pure states, as can be seen directly from
eqn (6.221).
The above conservation of purity must be carefully distinguished from the
irreversible nature of the process on the level of the corresponding ensemble of
type ep . Namely, the mean density matrix

ps(t ) -E f D uP[Gr, ti 0- (6.228)

obeys the irreversible dynamics given by the vacuum optical master equation
(6.91), as is easily verified by taking the average over eqn (6.226).

6.7 Photodetection on a field mode

6.7.1 The photocounting formula
As our last example, we consider in this section the quantum measurement of
the photons in a mode a of the electromagnetic field by a photon detector. It
will be assumed that the operation describing the back-action on the field mode
is proportional to the annihilation operator of the mode, that is we have a single
jump operator given by A = a. If we denote the damping constant of the mode
by -yo we thus have a PDP which is defined by the non-Hermitian generator

H = — i'Y° ata (6.229)

2 '
with the corresponding evolution operator

0(t) = exp(—if/t) = exp [-21)P at al (6.230)

2 ,

and by the transition functional

W[Ok'b-] = 70 I laz-P11 2 (5 [ a/1-1 0] - (6.231)

ila0ii
For this process the path integral representation developed in Section 6.1.2
can immediately be written down. First, we note that

(I(t)a --, exp(Pyo t/2)at(t). (6.232)

This relation can be used in eqn (6.44) to bring all jump operators to the left
which gives
 PHOTODETECTION ON A FIELD MODE 351

, t N ) = exp [N-70 t/2 — -yo (t i + + tN )12]aN ts1(t). (6.233)

The full propagator T[0,0,0 ] can therefore be written as follows,

T[0,t1z--k,0] = pN(t)6 aNÛ(t) (6.234)

N=0 ON (tiP- 11
The probability p N (t) for N counts in [0, t] is found from eqn (6.46) with the
help of the multi-time exclusive probabilities

ptN(ti , , tN) = ,tN)II 2

= YC)v exP [Noiot — (ti + + tN)] liaNÛ(N11 2 . (6.235)

Carrying out the time integrations one arrives at

= [t)1N
Nt exp(N-yo t)lia N Ci(t)/-P11 2 . (6.236)
PN(t)

Here we have introduced the quantum efficiency

11(t) E- 1 — exp(--yo t). (6.237)

This quantity is the probability of counting a photon in the time interval [0, t ]
from a one-photon state = ati0). In fact, for a one-photon state we have
obviously pi (t) = p(t), po (t) = 1— p(t) and p N (t) = 0 for N > 2.
The above representation of the process has been obtained for an initial pure
state 'Cb. If the field mode is initially in a mixed state given by /5 we have to
average (6.236) over a corresponding initial distribution P[,5, t = 0 ] which leads
to

[//(01 N
PO) = Nt exP(N'yot)tr Ot(t)(at) N aN tsi(t)fil . (6.238)

As is easily demonstrated the above result is equivalent to the photon counting

formula derived by Mollow (1968), Scully and Lamb (1969), and Selloni et al.
(1978), which is usually written in the form

[p(t)at cd N
P N(t) = tr N! exp [— p(t)at : (6.239)

where f (at, a) : denotes the normal ordered expression of some function f of

the annihilation and creation operators. In the remainder of this subsection we
are going to analyse the above expansion of the propagator (6.234) for different
initial state vectors /7).
352 STOCHASTIC DYNAMICS IN HILBERT SPACE

6.7.1.1 Fock state Let us first take an initial Fock state, i.e. b' .- = In). Then we
have the trivial relations
aN Û(t)in)
= In — N), (6.240)
i1aNtI(t)in)11
and
n!
IlaN Û(t)in)11 2 = exP(—nNt) (n — N)!' (6.241)

which yield
00

T[0,0,0] = E pN(t)(5[0 - in - N)l- (6.242)

N=0
By virtue of (6.236) the counting probability is found to represent a binomial
distribution,

PN (t) = ( N
n ) [P(t)] N [ 1 — 11 (t) ]
n—N ' (6.243)

If the initial state is a mixed state 13 we have

PO) = n ) [ti(t)] N [ 1 —
(nifiln) ( N (6.244)
n=1V

which corresponds to the expression given by Srinivas and Davies (1981).

6.7.1.2 Coherent state For an initial coherent state, ' -'b = la), we have (omit-
ting an irrelevant phase factor)
aN Û(t)la) = la exp(--yot/2)), (6.245)
liaNÛ(t)la)11
and
ii aNty- (01012 ____ i ct i2N exp (i_ N'yot) exP [ — (6.246)
larti(t)] -
This yields the propagator
00

Tio , tk-b, 01 = E pN(t)8 [o - la exP( - 70/ 2 ))] = 6 [0 — la exP( - 70/ 2 ))1,

N=0
(6.247)
showing that a pure coherent state remains a pure coherent state under time evo-
lution which, as a consequence of damping, shrinks continuously to the vacuum
state. The counting probability now represents a Poisson distribution,
((t))N
pN(t) =N NI exp [ — (N(t))] , (6.248)

with mean value (N(t)) = 1(11 2 p(t).

 PHOTODETECTION ON A FIELD MODE 353

6.7.1.3 Schradinger cat initial states We now consider a Schrödinger cat type
initial state (Garraway and Knight, 1994; Goetsch, Graham and Haake, 1995)
given by a symmetric superposition of two coherent states,
1
(6.249)
N+(a) (la) + 1 — a)) -

The corresponding antisymmetric superposition is written as

1
la, ) -7- N_(a)(la) — 1 — a)) - (6.250)

The normalization constants are

N+ (a) = 2e-1 ' 12 /2 -Vcoshla1 2 , (6.251)

N_(a) = 2e-1 ' 12 /2 Vsinh la1 2 . (6.252)

With the help of

aN CI(t)la,+) = N+1(a) a N exp(—N7o t/2)exp(-1a1 2 /1(t)/2

x { la exP(—Nt/2)) + (-1) N 1 — a exP( — Nt/ 2 )} (6.253)

it is easy to show that for N even we have

12NeXp(_N,y0t) COSh (1a12e--yot)

I1aNCT(t)la,±)112 ,iICE1 , (6.254)
cosh la1 2

and

aN tsi(t)lia, +)
= la exp(—Nt/2), +), (6.255)
11aNO(t)la,+)11
while for N odd we have

iia llo (o i ce,+ )1 1 2_ i1ct1i2N exp(—N-yot) sinh (1a12e—Tot) (6.256)

cosh la1 2 '

and

aN Û(t)la, +)
= la exp(—eyo t/2), —) (6.257)
IlaNCT(t)la,+)11
This yields the counting formulae

[1a1 2 p(t)]v cosh (1a1 2 e—Y0 9

) (6.258)
PN(t _AR cosh la1 2
354 STOCHASTIC DYNAMICS IN HILBERT SPACE

for even N, and

po) [Ial 2 p(t)] N sinh (kk1 2 e --yot)

(6.259)
N! cosh lal 2

for odd N. Summing separately over the even and the odd values of N we
immediately get the following expression for the propagator,

7[0,01-P, 0] = P""(t)8 [0 — la exP( —Nt/ 2), +)]

±edd (t)(5 [0 — la exp(--yo t/2), —)1, (6.260)

where

cosh (larp(t)) cosh (1a 1 2 11 —

Peven (t) = E pN(t) = 1 (6.261)
Neven
cosh kW

and

sinh 01 2 /1(t)) sinh (1a1 2 [1 — ft(t)])

edd (t) = E po) = cosh i ct i 2
(6.262)
No dd

It is clear from eqn (6.260) that the initial Schrödinger cat remains a Schrödinger
cat. The effect of the quantum jumps is to switch from the even superposition to
the odd superposition. As a consequence of damping both cats move towards the
vacuum state, while the coherence of the superposition is completely maintained.
This requires, however, that the photon detection provides a complete record.
As we saw in Section 4.4.1 the coherence is effectively destroyed already by a
single photon that escapes undetected.

6.7.2 QND measurement of a field mode

In Section 4.7 we have analysed a specific model for the quantum measurement
process developed by Walls, Collet, and Milburn (1985). This model consists of a
field mode a (the to-be-measured system), and a field mode b (the meter), which
are coupled by a four-wave mixing interaction. The meter mode b is coupled via
amplitude coupling to a zero-temperature environment.
Assuming that the environment acts as a perfect photoelectron counter, we
investigate here this model from the viewpoint of continuous monitoring of the
photons in the meter mode b. The corresponding unravelling of the master equa-
tion (4.285) gives rise to a PDP for the state vector Osm of system-plus-meter.
We determine the photon statistics for this measurement scheme and demon-
strate that the quantum demolition measurement of the photons of mode b leads
to a QND measurement of a certain system observable, namely of the square
(ata) 2 of the photon number in mode a.
 PHOTODETECTION ON A FIELD MODE 355

For the present model, the PDP is defined through the jump operator b, with
corresponding rate -yo , describing the demolition measurement of the quanta of
the meter mode, and by the non-unitary time-evolution operator

Û(t) = exp(—ifismt) = exp [— cita (b 6* — bi . 6) — Nt bi- b (6.263)

2 ]'

corresponding to the generator

Hsm = Hsm — fb (6.264)

b '
where Hsm is given by (4.280). As in eqn (4.281) we investigate an initial state
of the form

z-p•sm = E enin)s 0 low. (6.265)

n
Let us construct the propagator of the process. To this end we note that

bÛ(t) = Û(t) [e-7°t I 2 b — L. (e — °t/ 2 — 1) ata] . (6.266)

'Yo

This relation can be used in eqn (6.44) to bring all the O's to the left. Since
mode b is initially in the ground state we are left with the expression
N
, t N ) = (=:.1.) (e— l'otN /2 _ 1) ( e -70tN-1/ 2 _ 1) ... (e - 0t1/ 2 _ 1)
OtAr(ti,..-
'Yo
x 0.( t) (at a) N sm • (6.267)

From this we conclude that the full propagator of the process reads
co

T[Osm, qsm, 0] = E pN(t)(5[0sm (I(t) (at a)sm

-
N z-
P
IIU (t)(ata)NI-P.smil
j
, (6.268)
N=0

where the probability pN (t) for N counts is found from eqns (6.45) and (6.46).
Performing the time integrations we get

1 2,
pN(t) = yi A(t) N1 10(t)(ata)N sm 11 (6.269)

with
1 6 12
A(t) = 1 ' (yo t + 4e—Y0t/ 2 — e -70t — 3) . (6.270)
1
With the help of the relation
356 STOCHASTIC DYNAMICS IN HILBERT SPACE

Û(t)(atez p- sm E cne-mon2/27iNiros 1,3n (t))m, (6.271)

where 071 (0 is defined in eqn (4.290), we can determine the norm in the expression
for p N (t) , which finally yields

p N (t) = E 1,12 tkr-, [pn (t)]N

e— (6.272)

We observe that the distribution p N (t) for the counting events represents a
weighted sum of Poisson distributions with mean values

p(t) A(t)n 2 . (6.273)

In the limit -yo t cc, lei fixed, as well as in the limit (4.296) the quantity
A(t) becomes arbitrarily large. The Poisson distributions in the sum (6.272) are
thus strongly peaked around the average values pn (t), the relative widths being
1/ Vp n (t) 0. Since further the Poisson distribution with mean value pn (t)
occurs with the relative weight len 1 2 we conclude that the measurement of the
number N of photon counts practically always yields one of the values in the
vicinity of N = p(t) with the respective probabilities le 7 1 2 . For a given outcome
N we conclude from eqn (6.273) that

n2 =— (6.274)
A(t)
that is we infer a measurement value for the observable
A = ( a t a )2 . (6.275)

This measurement of the square of the photon number in mode a through

the measurement of the quanta in mode b provides a QND measurement as may
be seen from eqn (6.271). The function
f( n) e- A(t)n 2 /2 TI N (6.276)

is sharply peaked around the value n given by eqn (6.274). The record N thus
yields the conditional state vector

Û (t)(at a) N s m
In)s Ifin(t))m (6.277)
110(t)(ata)N1 p-. smii

where n is related to the outcome N by eqn (6.274). Summarizing, in the limits

given above the propagator of the process becomes

Tiosm,tk b•sm,oi
- E lenrs[osm - iTos ifin)m] (6,278)
 PHOTODETECTION ON A FIELD MODE 357

100

BO Var(A) ---- Var (A) + Var2 (A)

60 Var2 (A)

40

20

0 1 2 3 4

FIG, 6.12. The variances Var i (A), Var2 (A), and Var(A) for the observable
A = (ata) 2 . We depict the simulation results for 104 realizations of the
PDP describing the monitoring of the meter mode b. The simulation was
performed with Yo = 1 and 6 = 2 and the initial state Osm was taken to be
a superposition of the states n = 1, 2, 3, 4, 5 with equal amplitudes.

Equation (6.278) expresses the transformation from quantum mechanical am-

plitudes to classical probabilities during the measurement process. The initial
pure state Osm is transformed into a statistical mixture consisting of definite
states in)s 0 if3n)M- This mixture is an ensemble of type (T), where the various
pointer states In)s 1,30m represent classical alternatives.
The transition (6.278) may be illustrated further with the help of the vari-
ances Van i (A) and Var2 (A) introduced in eqns (5.67) and (5.68). Since A com-
mutes with Hsm the quantum mechanical variance

Var (A) = Var i (A) + Var2 (A) (6.279)

stays constant. Since further the initial state is pure, it follows that

Var 2 (A(0)) = 0, Var.' (A(0)) = Var(A). (6.280)

In the long time limit -yo t co, lei fixed, Var. ' (A(t)) decreases to zero, that is

lim Var. ' (A(t)) = 0, (6.281)

showing that the final ensemble is an ensemble of eigenstates of A. This may be

shown directly with the help of the Liouville master equation pertaining to the
process Osm(t). In fact, one finds
358 STOCHASTIC DYNAMICS IN HILBERT SPACE

0
Var i (A(t)) = — NE (Q[Osm] 2 ) < 0, (6.282)

where we have introduced the functional

1
QflPsitil = ((bosm Imo s m ) — 1 101) s m11 2 (Os mlAlOs m)) . (6.283)
11b0smil
Since the expectation on the right-hand side in (6.282) is taken over a non-
negative functional Q 2 we see that Var i (A(t)) decreases monotonically and,
therefore, approaches a definite value in the limit t 0. On the other hand, it
is easy to show that the functional Q[Osm] vanishes if and only if Osm is one of
the correlated states 1n)s 010 I fiii )m. These states are eigenstates of A and, thus,
eqn (6.281) follows.
Thus we see that the states in)s 0 ifi ii )m are the stable states of the process
Osm M. The stochastic dynamics drives the initial state (6.265) with probability
ic„1 2 into the state in)s 0 1070m. On the other hand, the variance Var 2 (A(t))
approaches Var(A), demonstrating that the statistical fluctuations of the random
variable (OsmiAlOsm), being zero initially, increase monotonically until they
become equal to the quantum fluctuations given by Var(A). This dynamical
behaviour of the variances expresses the fact that the potential outcomes of the
measurement contained in the amplitudes of the initial state are made objective
during the process of measurement. These features are illustrated in Fig. 6.12.
where we show the time dependence of the variances obtained from a numerical
simulation of the PDP (Breuer and Petruccione, 1996b).

References
Barchielli, A. and Belavkin, V. P. (1991), Measurements continuous in time
and a posteriori states in quantum mechanics. J. Phys. A: Math. Gen., 24,
1495-1514.
Breuer, H. P. and Petruccione, F. (1995a), Reduced system dynamics as a
stochastic process in Hilbert space. Phys. Rev. Lett., 74, 3788-3791.
Breuer, H. P. and Petruccione, F. (1995b). Stochastic dynamics of quantum
jumps. Phys. Rev., E52, 428 441. -

Breuer, H. P. and Petruccione, F. (1996a). Hilbert space path integral repre-

sentation for the reduced dynamics of matter in thermal radiation fields. J.
Phys. A: Math. Gen., 29, 7837 7853. -

Breuer, H. P. and Petruccione, F. (1996b). Quantum measurement and the

transformation from quantum to classical probabilities. Phys. Rev., A54.
1146-1153.
Breuer, H. P. and Petruccione, F. (1997). Stochastic dynamics of reduced wave
functions and continuous measurement in quantum optics. Fortschr. Phys., 45.
39-78.
Breuer, H. P., Kappler, B. and Petruccione, F. (1997). Stochastic wave-function
approach to the calculation of multitime correlation functions of open quantum
systems. Phys. Rev., A56, 2334-2351.
 REFERENCES 359

Breuer, H. P., Kappler, B. and Petruccione, F. (1998). Heisenberg picture

operators in the stochastic wave function approach to open quantum systems.
Eur. Phys. J., Dl, 9-13.
Carmichael, H. (1993). An Open Systems Approach to Quantum Optics, Volume
m18 of Lecture Notes in Physics. Springer-Verlag, Berlin.
Carmichael, H. J., Singh, S., Vyas, R. and Rice, P. R. (1989). Photoelectron
waiting times and atomic state reduction in resonance fluorescence. Phys.
Rev., A39, 1200-1218.
Castin, Y. and Molmer, K. (1995). Monte Carlo wave-function analysis of 3D
optical molasses. Phys. Rev. Lett., 74, 3772-3775.
Dalibard, J., Castin, Y. and Molmer, K. (1992). Wave-function approach to
dissipative processes in quantum optics. Phys. Rev. Lett., 68, 580-583.
Davies, E. B. (1969). Quantum stochastic processes. Commun. Math. Phys., 15,
277-304.
Dum, R., Parkins, A. S., Zoller, P. and Gardiner, C. W. (1992). Monte Carlo
simulation of master equations in quantum optics for vacuum, thermal, and
squeezed reservoirs. Phys. Rev., A46, 4382-4396.
Dum, R., Zoller, P. and Ritsch, H. (1992). Monte Carlo simulation of the atomic
master equation for spontaneous emission. Phys. Rev., A45, 4879-4887.
Gardiner, C. W., Parkins, A. S. and Zoller, P. (1992). Wave-function quantum
stochastic differential equations and quantum-jump simulation methods. Phys.
Rev., A46, 4363-4381.
Garraway, B. M. and Knight, P. L. (1994). Evolution of quantum superpositions
in open environments: Quantum trajectories, jumps, and localization in phase
space. Phys. Rev., A50, 2548-2563.
Ghirardi, G. C., Rimini, A. and Weber, T. (1986). Unified dynamics for micro-
scopic and macroscopic systems. Phys. Rev., D34, 470-491.
Ghirardi, G. C., Pearle, P. and Rimini, A. (1990). Markov processes in Hilbert
space and continuous spontaneous localization of systems of identical particles.
Phys. Rev., A42, 78-89.
Gisin, N. (1984). Quantum measurements and stochastic processes. Phys. Rev.
Lett., 52, 1657-1660.
Gisin, N. (1989). Stochastic quantum dynamics and relativity. Helv. Phys.
Acta, 62, 363-371.
Gisin, N. and Percival, I. C. (1992). The quantum-state diffusion model applied
to open systems. J. Phys. A: Math. Gen., 25, 5677-5691.
Gisin, N. and Percival, I. C. (1993). Quantum state diffusion, localization and
quantum dispersion entropy. J. Phys. A: Math. Gen., 26, 2233-2243.
Goetsch, P., Graham, R. and Haake, F. (1995) Schrödinger cat states and single
runs for the damped harmonic oscillator. Phys. Rev., A51, 136-142.
Hegerfeldt, G. C. and Wilser, T. S. (1991). Ensemble or individual system,
collapse or no collapse: A description of a single radiating atom. In Classical
and Quantum Systems: Foundations and Symmetries, (eds. Doebner, H. D.,
Scherer, W. and Schroeck, F., Jr.), pp. 104-115, World Scientific, Singapore.
360 STOCHASTIC DYNAMICS IN HILBERT SPACE

Mandel, L. and Wolf, E. (1995). Optical Coherence and Quantum Optics. Cam-
bridge University Press, Cambridge.
Marte, P., Dum, R., Taïeb, R., Lett, P. D. and Zoller, P. (1993a). Quantum
wave function simulation of the resonance fluorescence spectrum from one-
dimensional optical molasses. Phys. Rev. Lett., 71, 1335-1338.
Marte, P., Dum, R., Taïeb, R. and Zoller, P. (1993b). Resonance fluorescence
from quantized one-dimensional molasses. Phys. Rev., A47, 1378-1390.
Mollow, B. R. (1968). Quantum theory of field attenuation. Phys. Rev., 168,
1896-1919.
Mollow, B. R. (1969). Power spectrum of light scattered by two-level systems.
Phys. Rev., 188, 1969-1975.
Mollow, B. R. (1975). Pure-state analysis of resonant light scattering: Radiative
damping, saturation, and multiphoton effects. Phys. Rev., Al2, 1919-1943.
Winer, K. and Castin, Y. (1996). Monte Carlo wavefunctions in quantum
optics. Quantum Semiclass. Opt., 8, 49 72.
-

Molmer, K., Castin, Y. and Dalibard, J. (1993). Monte Carlo wave function
method in quantum optics. J. Opt. Soc. Am., B10, 524-538.
Pearle, P. (1976). Reduction of the state vector by a nonlinear Schrödinger
equation. Phys. Rev., D13, 857-868.
Pearle, P. (1989). Combining stochastic dynamical state-vector reduction with
spontaneous localization. Phys. Rev., A39, 2277-2289.
Percival, I.C. (1998). Quantum State Diffusion. Cambridge University Press.
Cambridge.
Plenio, M. B. and Knight, P. L. (1998). The quantum-jump approach to dissi-
pative dynamics in quantum optics. Rev. Mod. Phys., 70, 101-144.
Scully, M. O. and Lamb, W. E., Jr. (1969). Quantum theory of an optical maser.
III. Theory of photoelectron counting statistics. Phys. Rev, 179, 368-374.
Selloni, A., Schwendimann, P., Quattropani, A. and Baltes, H. P. (1978). Open-
system theory of photodetection: Dynamics of field and atomic moments. J.
Phys. A: Math. Gen., 11, 1427 1438.
-

Srinivas, M. D. and Davies, E. B. (1981). Photon counting probabilities in

quantum optics. Opt. Acta, 28, 981-996.
Walls, D. F., Collet, M. J., and Milburn, G. J. (1985). Analysis of a quantum
measurement. Phys. Rev., D32, 3208-3215.
Walls, D. F. and Milburn, G. J. (1994). Quantum Optics. Springer-Verlag.
Berlin.
Wiseman, H. M. (1994). Quantum Trajectories and Feedback. PhD Thesis.
Department of Physics, University of Queensland.
Wiseman, H. M. and Milburn, G. J. (1993a). Quantum theory of field-
quadrature measurements. Phys. Rev., A47, 642-662.
Wiseman, H. M. and Milburn, G. J. (1993b). Interpretation of quantum jump
and diffusion processes illustrated on the Bloch sphere. Phys. Rev., A47.
1652-1666.
 7
THE STOCHASTIC SIMULATION METHOD

Most stochastic processes encountered in science cannot be solved by analytical

methods. One therefore has to rely on numerical methods to derive predictions
from a given stochastic dynamics. Numerical algorithms can be employed to solve
the equation of motion governing the probability distribution for the stochastic
variable itself, the master equation or the Fokker—Planck equation, for example.
Another approach is to derive a system of deterministic differential equations
for the moments of the process and to solve it by numerical integration. How-
ever, these methods are often not feasible in practice due to the high degree
of complexity of the underlying problem: The probability distribution may be
represented by a density in a very high (or even infinite) dimensional space,
while the dynamics of the moments may lead to an infinite hierarchy of coupled,
non-linear equations. In both cases the numerical integration by deterministic
algorithms is extremely difficult.
An alternative and very powerful tool in the numerical analysis of stochastic
processes is provided by the so-called Monte Carlo or stochastic simulation meth-
ods. The basic idea of these methods is to generate by a numerical algorithm
independent realizations of the underlying stochastic process and to estimate
with the help of statistical means all desired expectation values from a sam-
ple of such realizations. A stochastic simulation thus amounts to performing an
experiment on a computer. It yields the outcomes of single runs with their cor-
rect probabilities and provides, in addition to the mean values, estimates for the
statistical errors of the quantities of interest.
In the present chapter we discuss in some detail the application of Monte
Carlo methods to the stochastic processes in Hilbert space that were derived in
Chapter 6. Depending on whether the stochastic process is a piecewise determin-
istic process or a diffusion process in Hilbert space the corresponding individual
realizations consist of intervals of deterministic evolution periods interrupted by
sudden jumps, or of continuous, nowhere differentiable paths. Appropriate sim-
ulation techniques for both types of processes will be described in Sections 7.1
and 7.2 and illustrated by specific examples in Section 7.3.
To determine numerically the density matrix of an open quantum system one
can either integrate the density matrix equation directly or else simulate the pro-
cess for the stochastic wave function and estimate the covariance matrix. In gen-
eral, the density matrix equation leads to a system of linear equations involving
N2 complex variables, N denoting the effective dimension of the Hilbert space.
By contrast, stochastic simulation only requires the treatment of N complex van-
362 THE STOCHASTIC SIMULATION METHOD

ables characterizing the state vector. For large N, that is for high-dimensional
Hilbert spaces, one thus expects Monte Carlo simulations to be numerically more
efficient than the integration of the density matrix equation, provided the size of
the required sample of realizations does not increase too strongly with N. This
point will be illustrated in Section 7.4 which deals with a detailed comparison
of the numerical performance of the stochastic wave function method and of the
integration of the corresponding density matrix equation.

7.1 Numerical simulation algorithms for PDPs

At several places in Chapter 6 we have already shown results obtained from nu-
merical simulations of PDPs in Hilbert space. Here we discuss in some detail the
simulation technique and special features of appropriate numerical algorithms. A
general account of the use of Monte Carlo methods in statistical and condensed
matter physics may be found in Binder (1995) and Landau and Binder (2000).

7.1.1 Estimation of expectation values

A stochastic simulation algorithm serves to generate a sample of independent
realizations of the stochastic process OM for the wave function. Let us denote
these realizations by Or(t), r = 1, 2, ... , R, where R is the number of realiza-
tions in the sample. The aim of the simulation is, in general, to estimate the
expectation values

A E E (F[0, t]) = f DODO * P[O, tiF[0 , t] (7.1)

of real functionals F[o,t] of the random state vector which may depend explicitly
on time. An unbiased and consistent estimator for the expectation value Mt is
provided by the sample average
R
-
Mt= yi
1
E F[or(t), t]. (7.2)
r=1
Here and in the following a hat is used to indicate an estimator. Of particular
interest are the expectation values

mt E E 0(t)1 13 10(t))) = f DODIP*P[O, ti(01/3 10) (7.3)

of the observables B of the open system, for which the estimator takes the form
1 R
112It = fi E (Or (t)IBIO T (0)- (7.4)
r=1

It is clear that the estimated expectation value is subjected to statistical

errors. A natural measure for the statistical fluctuations of the random variable
F[0, t] is the variance
 NUMERICAL SIMULATION ALGORITHMS FOR PDPS 363

= E (F[0, t] 2 ) — {E (F[0, t])} 2 . (7.5)

In the case F = (01.1310) this is just the variance

= Var2(B), (7.6)
which was already introduced in eqns (5.14) and (5.68). An appropriate estimator
for the statistical errors in the determination of Mt from a finite sample of size
R is given by

1
^2
at = R(R E (F[or(t),t]
1)r
=1.
- 11.4) 2 - (7.7)

The quantity ?it is known as the standard error of the mean value M. For the
estimation of the mean value of an observable B it is given by

1
(7.8)
R(R — 1) rE ((O r (t)IBIO r (t)) — 11'4) 2
If the realizations in the sample are statistically independent the standard error
eft decreases with the square root of the sample size R,

1
(7.9)

This provides a relation between the standard error of the mean value and the
number of realizations of the sample.

7.1.2 Generation of realizations of the process

Let us consider a PDP in Hilbert space of the type studied in Section 6.1.1 which
can be represented through the stochastic differential equation (see eqn (6.25))

A itP(t)
do(t) = (ft + E ,yiiiitio(t)112) ocodt + OW) dNi (t).

(7.10)
The Poisson increments dNi (t) satisfy eqns (6.26) and (6.27), while the non-
Hermitian Hamiltonian fi is given by (see eqn (6.13))

ft = H — 2 i
(7.11)

We also recall from Section 6.1.1 that the cumulative waiting time distribution
(6.21) for the quantum jumps is given by
F[0, = 1— exP( — ifiT)CP11 2 , (7.12)
where T represents the random waiting time between successive jumps.
364 THE STOCHASTIC SIMULATION METHOD

From the discussion of Sections 1.5.2 and 6.1.1 we immediately see that a
sample of realizations Or(t) of the process defined by eqn (7.10) in the time
interval O, tf] can be generated by means of the following algorithm:
[

1. Assume that the normalized state Or (t) was reached through a jump at
time t and set Or(t) = O. If t is the initial time t = 0, 't-i) is the initial state of
the process which must be drawn from the initial distribution P [1/), t = 0].
2. Determine a random waiting time T according to the distribution function
(7.12). This can be done, for example, by drawing a random number n
which is uniformly distributed over the interval [0, 1] and by determining
T from the equation

= F[77),T] = exp( — ikr)011 2 - (7.13)

For n > q there exists a unique solution, q being the defect of the waiting
time distribution defined in eqn (6.22). For n < q we set T = Do in which
case there will be no further jumps. Within the time interval [t, t + 7] the
realization follows the deterministic time evolution given in (6.15),

exp(—iks)
Or (t + s) = 0 < S < T. (7.14)
I exP( — i-f1411 '

3. At time t+T (if T is finite and t-FT < t1) one of the possible jumps labelled
by the index i in eqn (7.10) occurs. Select a specific jump of type i with
probability

7illAiOr (t + r)11 2 (7.15)

pi =
Ei fiYillAiO r (t + 7)11 2
and replace

A iOr t +
(

Or t + (7.16)
I lAiOr + '011
(

4. Repeat steps 1 to 3 until the desired final time t1 is reached, which yields
the realization Or (t) over the whole time interval [0, t1].
5. Once a sample of realizations Or (t), r = 1, 2, , R, has been generated ac-
cording to this algorithm any statistical quantity can be estimated through
an appropriate ensemble average, as described in the previous subsection.
Let us now discuss the various parts of the algorithm in more detail.

7.1.3 Determination of the waiting time

An essential part of the simulation algorithm is the determination of the random
waiting time.
 NUMERICAL SIMULATION ALGORITHMS FOR PDPS 365

7.1.3.1 Exponential waiting time distributions The easiest case is provided by

an exponentially distributed waiting time,
F [ , T] = 1 — eXp (—Fr) . (7.17)
Such a waiting time distribution always arises when the total jump rate does not
depend on time, i.e. Ffi-p, F = constant > 0. If n is an uniformly distributed
random number in the interval [ 0, 1], the waiting time T is easily determined
with the help of (7.13) which yields
1
T= ln (7.18)
The development of appropriate random number generators has been discussed,
for example, by Press, Flannery, Teukolsky and Vetterling (1992) and by Knuth
(1981).
7.1.3.2 Multi exponential waiting time distributions A waiting time distribu-
-

tion which is a sum of exponential functions arises if the jump operators Ai are
eigenoperators of the Hamiltonian H, which leads to (see eqn (3.125))

=0. (7.19)

This was the case, for example, in the weak-coupling master equation (3.140)
without external driving fields. It follows that H and the positive operator
Ei Ai have a common eigenbasis {la», i.e.

HIct) = Ela) (7.20)

and

= Fa lcx), (7.21)

with real eigenvalues Ea and Fa > 0. Hence, also H- is diagonal in this basis and
we may write

Ria) = (Ea — —;ra) la). (7.22)

The waiting time distribution is therefore a sum of exponential functions whose

exponents are given by the imaginary parts of the eigenvalues of H. Accordingly,
the waiting time T is determined by the equation

= exP( — ifir)'011 2 = YT I( 101 2 exp(—F a r). (7.23)

a

If î is proportional to one of the la) we recover the exponential waiting time

distribution (7.17). In general, eqn (7.23) can no longer be solved analytically
366 THE STOCHASTIC SIMULATION METHOD

for T. However, it may be solved numerically provided the basis la) and the
corresponding eigenvalues Fa are known. Equation (7.23) also shows that we
will have a defect if F a = 0 for some a with (a l't/..)) O.

7.1.3.3 General waiting time distributions Step 2 of the simulation algorithm

described in Section 7.1.2 rests on the assumption that the waiting time distri-
bution F[:0 ,t] is known explicitly such that one can solve eqn (7.13) for T, either
analytically or numerically. If F[5, T] is known the algorithm described is usually
the most efficient one, in particular, if the random time steps T become large
and if an analytical expression for exp(—ifir) is also known.
However, in the general case neither the deterministic evolution nor the wait-
ing time distribution are known explicitly. The waiting time T must then be
determined along with the numerical solution of the deterministic dynamics. To
achieve this one starts from the normalized initial state irb and solves numerically
the equation

d
— b(t) = —i1-10(t), (7.24)
dt

which yields the non-normalized state vector OM = exp(—ifit) the norm of

which decreases monotonically. According to eqn (7.13) the waiting time T can
be determined by checking after each integration step whether the square of the
norm has decreased to the value 27 . Thus, step 2 of the algorithm given in Section
7.1.2 may be replaced by the following procedure:
2. Draw a random number n uniformly distributed in [0, 1] and determine
the non-normalized solution OW of the linear Schrödinger equation (7.24)
corresponding to the initial value 0(0) = 0. The waiting time T is then
determined by the condition

= (7.25)

The deterministic evolution between the jumps takes the form

(t + s) = "0(s) 0< s < T. (7.26)

A ' HO(S -

This form of the algorithm should be preferred when performing simulations

in the general case. For the numerical solution of eqn (7.24) between the jumps it
is often convenient to expand the wave function with respect to an eigenbasis of
the Hamiltonian H and to bring (7.24) into a linear system of ordinary differential
equations for the components of 0. The integration can then be performed by
employing standard numerical routines appropriate for deterministic differential
equations. We remark finally that the above form of the algorithm, in particular
condition (7.25) and eqn (7.26), also applies, of course, if the generator in eqn
(7.24) depends on time, that is if H = H (t)
 ALGORITHMS FOR STOCHASTIC SCHR6DINGER EQUATIONS 367

7.1.4 Selection of the jumps

According to step 3 of the algorithm described in section 7.1.2 we have to select
a specific quantum jump of type i with probability p i given by eqn (7.15) and to
carry out the corresponding transition (7.16). The index i specifying the jump
is a random number with values in some index set / = {1, 2, ... , K } following
the distribution p i .
Essentially, there are two ways to implement numerically the selection of the
jumps. If the index set I is small, then i may be drawn with the help of the
method of linear search (Gillespie, 1976, 1992). To this end, one subdivides the
interval [0, 1] into K sub-intervals Ji of lengths pi and draws a uniform random
number x G [0, 1]. The index i is then found by searching for the interval Ji
which contains z. When applying this algorithm, it is important to check that
the probabilities pi are larger than the resolution of the random number generator
(Hanusse and Blanché, 1981).
If the values of the probabilities p i are approximately equal to each other
it may be more efficient to use the following rejection method (Press, Flan-
nery, Teukolsky and Vetterling, 1992): Draw two independent uniform random
numbers z and y in [0, 1]. The first random number is used to select an index
= int(Ks) + 1 from the index set I. If the second random number satisfies
Y < pilpmax this index i is accepted, where pmax denotes the maximum of the
pi . Otherwise i is rejected and the procedure is repeated.
The advantage of the rejection method over the linear search is that the
former is numerically efficient even if the number K of elements in the index
set becomes large. A condition is that the product of pmax and K should not be
much larger then 1.
The linear search and the rejection method get inefficient if the index set I
is very large and if the distribution pi is very inhomogeneous. In these cases it
might be helpful to cluster possible jumps into logarithmic classes (Fricke and
Schnakenberg, 1991) or to make use of search trees (Maksym, 1988; Blue, Beichl
and Sullivan, 1995). However, the time which is saved due to improved search
algorithms might be lost again in the actualization of the data structures (Breuer,
Huber and Petruccione, 1996).

7.2 Algorithms for stochastic Schraclinger equations

In this section we develop numerical algorithms for the simulation of stochastic
Schriidinger equations (SSEs) of diffusion type. Such equations were encountered
in Section 6.1.3 as diffusion approximations of a corresponding PDP in Hilbert
space. They also arise in the context of fundamental dynamical collapse models
(see Section 6.5.2).
Below, we are going to discuss four different algorithms to solve SSEs, namely
the well-known Euler scheme, an adaptation of the Heun and the Runge—Kutta
scheme to SSEs and a second-order weak scheme for stochastic differential equa-
tions proposed by Platen (Kloeden and Platen, 1992). Special emphasis is laid
on the order of convergence of these schemes using as an example a stochastic
368 THE STOCHASTIC SIMULATION METHOD

Schrödinger equation of the type encountered in Section 6.4.2 (Breuer, Dorner

and Petruccione, 2000). The methods discussed here can of course be applied to
other kinds of SSEs. In particular, they may easily be generalized to simulate
SSEs involving complex instead of real Wiener increments.
7.2.1 General remarks on convergence
Let us consider a stochastic Schrödinger equation of the following general Itô
form,

&PM = D1 (OM) dt + D2 (0 (t)) dW (t), (7.27)

with the drift term

D 1 (0) = —iHO + -;f- ((A + At)0A — AtA — 41 (A + At)) 0, (7.28)

and the diffusion term

1
D2 OM = Nry (A_ —2 (A ± At) o ) 0, (7.29)

where we have made use of the shorthand notation

(A) 0 (0 1A 10 ), (7.30)

and dW (t) is a real Wiener increment. An equation of the form (7.27) was ob-
tained in Section 6.4.2 by performing the diffusion limit of homodyne photode-
tection (see eqn (6.181)). For simplicity we consider the case of a single Lindblad
operator A. The corresponding density matrix equation is given by

d 1 1
— p(t) = —itH, p(t)] + 7 (Ap(t)At — — At Ap — — pAt A) L p(t). (7.31)
dt 2 2
We note that eqn (7.27) is a stochastic differential equation with multiplicative
noise which means that the diffusion term D 2 (21) (t)) which multiplies the Wiener
increment depends on the stochastic variable OM. The noise is called additive
if the diffusion term does not depend on the stochastic variable.
It must be emphasized that OM is a process in the Hilbert space lis of an
open quantum system. It thus represents, in general, a process in an infinite-
dimensional vector space. However, in many cases of interest the dissipative
character of the equations of motion ensures that the dynamics is confined to a
finite-dimensional subspace of 7- Is. It is therefore justified physically to restrict
the discussion to finite-dimensional spaces, that is, we may regard eqn (7.27) as
defining a process in an effectively finite-dimensional space.
Consider now a numerical scheme to integrate eqn (7.27) over the time in-
terval [t o , to + 7 ] . Such a scheme leads to a discrete time approximation V) k for
the exact process 0(tk ) at times tk E to ± k At, where k = 0, ... ,n = T I At.
 ALGORITHMS FOR STOCHASTIC SCHR6DINGER EQUATIONS 369

The discretization does not have to be equidistant but to simplify the notation
this will be supposed here. In this section, Ok will always denote a numerically
generated approximation, while OM denotes the exact process defined by eqn
(7.27). Analogously, we denote by

Pk = E{I0 k) (0 kl] (7.32)

the discrete time approximation of the density matrix, whereas

P(t) = EtIO(t))(0(t)ii (7.33)

is the exact density matrix satisfying the Lindblad equation (7.31).
In order to characterize the convergence behaviour of the numerical schemes
to be discussed below we will compare the discrete approximation pk with the
Taylor expansion of the exact density matrix p(t) which is given by

p(t + At) = p(t) + rp(t)At + L 2 p(t)At 2 + 0(t3 ).

—
2
1 (7.34)

Without loss of generality, we will always assume that a deterministic (i.e. sharp)
initial state 00 E tp(to ) is given. The single-step error of a certain numerical
scheme may then be expressed through the difference

pl - p(ti) = 0(t'), (7.35)

which means that the scheme reproduces the Taylor expansion of p(t) including
terms of order 0' — 1 in At.
It is straightforward to prove that the integration over a finite time interval
[t o , to + T] decreases the order of convergence by one since we need n = 711 At
time steps to calculate the density at time to + T = tn , i.e.

pn — p(t1 ) = O(Lt) (7.36)

with = fi' 1. If this equation is satisfied the numerical scheme will be said to
be a scheme of order fi in the following.
Equation (7.36) is a special case of weak convergence of order 0. The degree of
an approximation of a stochastic differential equation can also be characterized
by the notions of strong convergence (Kloeden and Platen, 1992). If the numerical
algorithm leads to an approximation On
for 0(W satisfying

E {II On — 0(t 11 ) Ill - AV3 (7.37)

for sufficiently small At, it is said that the scheme converges strongly with order
0. This definition imposes a condition on the closeness of the random variables
On and 0(t12 ) at the end of the integration interval. By contrast, weak conver-
gence only requires that the probability distributions of 61 and 0(t11) are close
to each other which is a much weaker criterion. In practice, one is often only
interested in this weaker form of convergence, for example, when considering the
approximation of functionals of the stochastic variable.
370 THE STOCHASTIC SIMULATION METHOD

7.2.2 The Euler scheme

The Euler scheme is probably the simplest recursive algorithm to get an approx-
imation for the state vector. It takes the form

Ok+1 =11) k D1( 114)At D2(0k)AWk- (7.38)

The Wiener increments AWk, k = 0, 1, 2, ,n - 1, are independent Gaussian
random variables with zero mean and variance At, the index k indicating different
realizations at each time step. Thus, we can write

A Wk = (7.39)
where k is a Gaussian distributed random variable with mean zero and unit
variance, that is a random variable following a standard normal distribution. It
should be mentioned that it is not necessary to use Gaussian distributed random
variables if one requires convergence in the sense of eqn (7.36). One may use other
random variables instead which coincide, for example, only in the first and the
second moment with those of a standard normal distribution.
It is easy to demonstrate that eqn (7.36) is satisfied with fi = 1. The Euler
scheme thus provides a weak scheme of order 1. In spite of the low order of
convergence the Euler algorithm is useful because it is very easy to implement
numerically and often yields a reasonable degree of approximation. In general,
the scheme does not conserve the norm of the state vector, by contrast to the
SSE (7.27). The state vector should therefore be normalized after every iteration.

7.2.3 The Heun scheme

The following method, which we shall call the Heun scheme, is a generaliza-
tion of the Heun method known from the numerical integration of deterministic
differential equations. It is defined by the recursion relation

4, 1 n (4, \ n( A n \A
Ok+1 = (Pk ± 0,bk) '-1 Wk , (7.40)

with

= k + D1(5k)Lt + D2(0k)AWk- (7.41)

A Wk again takes the form of eqn (7.39). For vanishing diffusion one recovers
the Heun method for deterministic differential equations. It turns out that, in
general, the Heun algorithm again leads to relation (7.36) with = 1. It is
therefore a scheme of order 1, similar to the Euler method. This is in contrast
to SSEs with additive noise, where the Heun method is actually of order two.

7.2.4 The fourth-order Runge-Kutta scheme

For deterministic differential equations there is a great variety of numerical in-
tegration methods and one is easily misled when heuristically modifying these
 ALGORITHMS FOR STOCHASTIC SCHR6DINGER EQUATIONS 371

10

A N

V
N

0.0 0.5 2.0 2.5

FIG. 7.1. One realization of the process (7.50) for the damped harmonic os-
cillator with initial condition Ino) = 19) (continuous line). The dashed line
represents the analytical expression (7.56) for the expectation value of the
number operator.

methods in order to integrate SSEs. The following example shows that it is indeed
wise to proceed cautiously.
A well-known numerical scheme for deterministic differential equations is
the fourth-order Runge—Kutta method. A heuristic modification to approximate
solutions of SSEs is obtained if one integrates in each time step the drift with the
Runge—Kutta method and the diffusion term of the SSE with the Euler scheme.
This leads to the following scheme,
1
Ok+1 = k + —6 1011, + 24, + 2 + tp;',} At + D2 (0k)AWk (7.42)

with

(7.43)
012, = D i (Ok + —21 Atz/4, ), (7.44)

= D1(0k + —21 Atq), (7.45)

'14 = Di (Ok + (7.46)
The order of convergence can again be determined by calculating the difference
in eqn (7.36). The surprising result is that the scheme is of order fi = 1. The
stochastic generalization of the Runge—Kutta method converges with order 1, by
372 THE STOCHASTIC SIMULATION METHOD

10 ,

A
V

0.0 0.5 10 1.5 2.0 2.5

yt

FIG. 7.2. The average over 100 realizations of the damped harmonic oscillator.
The initial condition is Ino ) = 19). The dotted line gives the analytical result.

contrast to the Runge—Kutta method used for deterministic differential equations

which is of order 4.
From this example one can see that in general simple heuristic generalizations
of higher-order Runge—Kutta schemes do not necessarily lead to higher-order
methods for SSEs. The above method converges of course but, in general, there
will be no advantage over the Euler method, for example.

7.2.5 A second order weak scheme

-

The final algorithm we are going to consider has been proposed by Platen (Kloe-
den and Platen, 1992). In our notation this scheme is provided by the recursion
relation

1f
k+1 = ± — V--)1(0k)
2
+ &'k)) At
1,
+ kr312 P k )+D2(ok )+ 2 D2(tPk)) Wk
(

1
+ — (132 (01-,H) D2 {(Wk ) 2 At} At-1 /2
()) (7.47)
4

where

ZPIc Dl(Pk)At D2(5k)Wk, (7.48)

'11) = k Dl(tPic)At ± D2(0k) - t (7.49)
 EXAMPLES 373

A
V

0.0 0.2 0.4 0.6

FIG. 7.3. A single realization of the mean number operator of the damped har-
monic oscillator. The thin line corresponds to the second-order weak scheme
and the bold line to the Euler, the Heun and the Runge—Kutta schemes (not
distinguishable on the scale of the figure). The smooth line represents the
analytical result (7.56).

Here, the increments AWk must fulfil certain conditions. These are satisfied if eqn
(7.39) is assumed. For vanishing noise the scheme reduces to the Heun scheme
for deterministic differential equations.
Performing again the error analysis one finds that Platen's scheme converges
with order 0 = 2. Thus, contrary to the algorithms studied before, Platen's
scheme is really a higher-order scheme in the sense defined above. This point
will be illustrated in the examples below.

7.3 Examples
In this section we apply the simulation algorithms for PDPs and stochastic
Schrödinger equations to some specific examples.
7.3.1 The damped harmonic oscillator
7.3.1.1 Simulating the PDP We study the following process for the damped
harmonic oscillator (see Section 6.7),

&PM = (ata — (0(t)latal0(t))) + OW) dN (t). (7.50)

Let us simulate this equation for the simplest initial condition, i.e. for a number
state 0(0) = In). In this case the deterministic evolution periods of the process
are trivial since
374 THE STOCHASTIC SIMULATION METHOD

0.50
Runge-
Kutta
0.40
Heun
Euer-
0.30
A
V 0.20

0.10

Platen
0.0C)

—0.10 „ I

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35

At

FIG. 7.4. A plot of the mean photon number f/T. relative to the analytically
calculated value (n(T)) at ey1-1 = 1.2 versus the step size At for different
methods. The solid lines correspond to linear and quadratic fits.

exp(—ifit)In) = exp(--ynt/2)1n). (7.51)

Also the jumps are trivially given by

In)
am) = In — 1). (7.52)
lain) ii
Since the waiting time distribution is

F[In), T] = 1 — I I exp(—ifiT)In)11 2 = 1 — exp(-7nr), (7.53)

we find that the random waiting time may be obtained from the equation
1
T = — — ln (7.54)

where is a uniformly distributed random number in the interval [0, 1].

?)

Figure 7.1 shows one realization of the process for the initial condition 0(0) —
Ino) = 19). As an example of the calculation of expectation values we also show in
Fig. 7.2 the mean of the number operator determined from a sample of R = 100
realizations through (see eqn (7.3))

1
= E(Or(t)latalOr(t))- (7.55)
r=1
 EXAMPLES 375

1 .0
a Euler
0.8 A Heun
Runge—Kutta
E 0.6 - Platen

a_ 0.4
o
0.2

0 0
t I t 111111L[111111-L/t [111

0.00 0.10 0.20 0.30 0.40 0.50

<n>(T)1

FIG. 7.5. CPU time (normalized to one) versus absolute error for the data points
of Fig. 7.4.

The analytical expression for this quantity is, of course, given by

(ata(t)) = no exp(—yt). (7.56)

According to eqn (7.4) the standard error for this mean value is provided by the
expression

1 \2
^2
= R(R ((Or Miat aiOr (0) – (7.57)

The values for the standard error are indicated in Fig. 7.2 by error bars.

7.3.1.2 Simulating the SSE The stochastic Schn5dinger equation (7.27) for the
damped harmonic oscillator reads

1
7 ((a + at) 0(t) a — at a — — (a + at) 20 (0 ) 0(t)dt
d(t) = —
2 4
1
-H\Fy (a — — (a + at) 0 (0) 0(t)dW(t). (7.58)
2
In the numerical simulations the state vector has been represented in the basis
of number states In). Again, the initial state is taken to be zp(o) = Io = 9 ) and
the Hilbert space has been truncated at nmax = 12, that is the simulation was
performed in a subspace of dimension N = 13.
376 THE STOCHASTIC SIMULATION METHOD

1.0

0.8

0.6

0.4

0.2

0.0
o 5 10 15 20
Tt

FIG. 7.6. One realization of the PDP (7.59) for the driven two-level atom. The
continuous line shows pli computed from the realization. The dashed line rep-
resents the analytical solution for pli according to eqn (3.289). Parameters:

A single realization of the process is depicted in Fig. 7.3. The solid line shows
the analytical curve (7.56). The thin line represents the solution calculated with
the second-order weak scheme of Platen and the bold line the solutions calculated
with the other methods, which are not distinguishable in this example. The size
of the time steps is At = 0.01. Note that the same sequence of pseudo-random
numbers was used for each run to stress the differences or similarities between
the methods.
In order to study the convergence behaviour, we compute the mean value
(7.55) of the number operator at a fixed time T for R = 105 realizations and for
different step sizes At. The results are displayed in Fig. 7.4. Error bars are not
drawn in this figure because they would be roughly of the size of the symbols.
Obviously Platen's second-order weak scheme converges very well already for
quite large step sizes compared to the other schemes. The heuristic Runge—Kutta
method shows the worst convergence behaviour.
Based on the knowledge of the order of convergence it is possible to fit lin-
ear functions (in the case of the Euler, Heun and Runge—Kutta methods) or
quadratic functions (in the case of Platen's methods) to the data points. These
fits (also shown in Fig. 7.4) are in good agreement with the data points and
confirm the results derived in Section 7.2.1.
Another criterion for the assessment of numerical algorithms is, of course, the
CPU time which is required to achieve a given accuracy of the results. Figure 7.5
 EXAMPLES 377

1 .0 1_

yt

FIG. 7.7. The average over 1000 realizations of the PDP (7.59) for the driven
two-level atom. The continuous line shows pu computed from the average
over the realizations. The dashed line represents the analytical solution for
Pli according to eqn (3.289). The standard error of the mean is indicated by
error bars. The parameters are the same as in Fig. 7.6.

displays the CPU time (normalized to one) versus the error for the data shown
in Fig. 7.4. We see that Platen's scheme shows the best performance, followed
by the Euler scheme.

7.3.2 The driven two-level system

As a second example we consider the PDP given by the equation

0-_0(t)
dW(t) = illo--0(0112) 0(t)dt+ ( Ilia 0011 0(0) dN(t). (7.59)
(11

As shown in Section 6.3 this equation describes the direct photodetection of

a driven two-level system (see eqn (6.129)), while the stochastic Schn5dinger
equation

4(0 = il/L0(t)dt
1
+ — ((cr_. + C/±)ip0"__ - ca_ - - (a_ +a + )) 0(t)dt
2 4
1
+ Nry + o-±)0)0(t)d-W(t) (7.60)
378 THE STOCHASTIC SIMULATION METHOD

Yt

FIG. 7.8. A single realization for p H_ computed from eqn (7.60) according to
the different numerical schemes of Section 7.2 with step size At = 0.01. The
results are not distinguishable on the scale of the figure. The smooth line gives
the analytically solution (3.289). The inset shows a magnification of a small
part (thin line: Platen's scheme; bold line: Euler, Heun and Runge—Kutta
schemes). Parameters: fl = -y = 0.4.

corresponds to homodyne photodetection (see Section 6.4). In the numerical

simulations we start with the atom in its ground state 1g) and compute the
probability

Pu(t) = (e1P(010) (7.61)

of finding the atom in the excited state le). From a sample of realizations this
probability is estimated by determining the average

1
ii = (7.62)
r=i

while the corresponding standard error ô-t is given by

-2 1
(7.63)
at
R(R — 1) TE=1(1(elOr (t))1 2 — kt) 2

The discussion of the algorithm for the generation of realizations of the piece-
wise deterministic process may be kept brief because we have already collected
all relevant formulas for the waiting time distribution and the jumps in Section
 EXAMPLES 379

0.010

0.008

0.006

0.004

0.002

0.000

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35

Gt

FIG. 7.9. Estimated means of 1(e1 (T))1 2 minus the exact values for different
step sizes and methods. The parameters are Si = 0.4, 7 = 0.4 and 7T = 8.4.

6.3. According to eqns (7.13) and (6.133) the random waiting time 7- can be
determined by solving the equation
2
7 sin2 /1-7- _I._ ; in- cos lay) (7.64)
= exP( - 7T/2) (1 + 8112 sin

for T. We have used a numerical routine to find the root of this equation. In
Fig. 7.6 we show 1(e 10 (0)1 2 for one realization of the PDP. The corresponding
average over an ensemble of 1000 relizations can be seen in Fig. 7.7 which clearly
reveals the convergence of the algorithm.
Let us now turn our attention to the simulation of the stochastic Schrödinger
equation (7.60). Single realizations calculated with the different numerical meth-
ods introduced in Section 7.2 are shown in Fig. 7.8. In the inset, showing an
enlarged part of the image, it is seen that the weak scheme of order two differs
from the other schemes which are still hardly distinguishable on the enlarged
scale.
The results of simulations with different step sizes and R = 5 x105 realizations
per point are shown in Fig. 7.9. For the sake of clarity we display only error bars
at points which are not too close to other points. They have been calculated
according to eqn (7.63). The error bars omitted are of about the same size. In
any of the four cases the numerical results are in very good agreement with
the analytical predictions. The Runge—Kutta method seems to converge better
than the Euler and the Heun scheme for bigger step sizes. However, if one takes
into account the corresponding CPU times, as is done in Fig. 7.10, it becomes
380 THE STOCHASTIC SIMULATION METHOD

1 .0
EWer-
0.8 L Heun
Runge—Kutta
E 0.6 + Paten

o_ 0.4
o

0.2

0.0

0.000 0.002 0.004 0.006 0.008 0.010

PT — pl,(T)1

FIG. 7.10. The CPU time (normalized to one) versus the absolute error for the
data points of Fig. 7.9.

obvious that these three methods are more or less equal with respect to the
ratio of accuracy and speed. Anyhow, Platen's scheme shows again the best
performance. This changes if one chooses parameters such that 'y < 0, i.e. if
the deterministic part of the SSE prevails over the noise part. The Runge—Kutta
method then shows the best performance, while Platen's scheme behaves like the
Heun method. This is, of course, to be expected since for < fl the deterministic
part of the equation of motion dominates and, therefore, the fourth-order Runge-
Kutta method becomes the most efficient one.

7.4 A case study on numerical performance

Monte Carlo algorithms are often the only way to study high-dimensional sys-
tems, for which deterministic calculations based on the numerical integration of
the equations of motion are beyond the capacity of any computer. On the other
hand, deterministic methods may be preferred for low-dimensional systems. In
this section we want to present a systematic analysis of the numerical efficiency
of the Monte Carlo wave function method and compare it with that of the inte-
gration of the corresponding density matrix equation. The main interest lies in
the dependence of the time-consumption on the system size N. It will be shown
that the CPU time consumption for the two approaches may be expressed in
terms of simple power laws.
 A CASE STUDY ON NUMERICAL PERFORMANCE 381

7.4.1 Numerical efficiency and scaling laws

The results of a Monte Carlo simulation are always subject to statistical errors.
These errors in turn are related to the number of realizations that are generated
in the simulation. Hence the total CPU time required by a Monte Carlo simula-
tion depends crucially on the desired accuracy. A typical task of the numerical
investigation could be, for instance, 'estimate the expectation value of the energy
with a relative error of 1 %', or: 'calculate the density matrix with a precision of
better than 10 -4 in each of its elements'. The question to be answered is then, of
course: Under which circumstances should a stochastic simulation be preferred?
The crucial quantity for the relative performance is therefore the number of
realizations R. Hence, the dependence of the errors of the simulation with the
sample size deserves a more detailed discussion. To this end let us denote by N
the number of complex variables which are used for the numerical representation
of the wave function, that is, the number of basis states. Accordingly, the number
of complex variables which are necessary to represent the density matrix is then
N 2 /2. The relation (7.9) allows us to write down the following equation

-2 AB )
( N
(7.65)
a-t = R '

in which the factor AB (N) takes into account the dependence of the statistical
error on the observable B and on the system size N, but does not depend on
the sample size R. Using a sufficiently large sample of realizations, AB (N) can
be determined by fitting eqn (7.65) to the simulated data. Then, eqn (7.65) can
be solved for R,

AB(N)
R E R(N) = (7.66)
at

This is the number of realizations that is necessary to achieve a given accuracy 6 -t

for observable B and system size N. If AB (N) varies as a power of the system size,
AB (N) — N — a , the following classification can be made (Ferrenberg, Landau and
Binder, 1991):
1. If x = 1, the observable B is strongly self-averaging.
2. If 0 < s < 1, the observable B is self-averaging.
3. If x = 0, the observable B is not self-averaging.
With these considerations we can now investigate the scaling behaviour of the
CPU time required by the deterministic and by the stochastic approach to the
numerical estimation of expectation values with a prescribed accuracy.
In the case of the density matrix equation approach, the part of the numerical
integration routine that dominates the CPU time consumption is the calculation
of the generator fp of the Lindblad equation. One such calculation requires, for
large enough N, an amount of CPU time proportional to a power of N, and
382 THE STOCHASTIC SIMULATION METHOD

the CPU time needed to integrate the density matrix over a given physical time
interval is, to leading order in N,

TDME = ki Si (N) AT' 3 . (7.67)

Here, Si (N) is the number of times the quantity rp has to be evaluated, ki and
)3 depend on the type of the specific problem, but not on N. Besides, ki depends
on the particular implementation on a computer.
Analogously, in many examples the time-critical part of the stochastic sim-
ulation is the calculation of the generator ko,and the CPU time required for
the simulation is

Tsts = k2 R(N) s 2 (N) Na. (7.68)

R(N) is the number of realizations of the process that are generated to treat the
system of size N, 82(N) is the number of evaluations of kii) for one realization,
and k2 is analogous to ki .
In many situations, s i (N) and 5 2 (N) will be roughly equal. Provided that
similar numerical integration routines are used, this is the case if the smallest
time scale of the dynamics of the stochastic wave function is about equal to that
of the density matrix. Since we want to separate the effects of system size from
dynamical phenomena, this case is the one of interest for us, and the example
presented in Section 7.4.2 illustrates that case.
Let us briefly note that there are also situations where s i (N) and s2 (N)
are quite different. In general their ratio might depend on N and they do not
necessarily grow in the same way with the physical time over which the system
is studied. Consider the case where the time scale of the dynamics of a single
realization of the stochastic process varies, during its temporal evolution, in a
wide range, as for example, in laser cooling (see Section 8.3). The simulation of
one realization will then contain stretches with very long time steps, interrupted
by phases of more rapid development and short time steps. The integrator of
a density matrix equation, on the other hand, which describes the dynamics of
the whole ensemble, must always adapt to the short time scale. Clearly, in such
cases the stochastic wave function method is the preferred choice.
The number of floating point operations to calculate rp and to calculate kV)
differ by about a factor N, and one expects

e& a + 1. (7.69)
This relation will be verified in the example below. Concluding these general
considerations, we can write eqns (7.67) and (7.68) in a more succinct form

TDME = k i. NŒ+ 1 (7.70)

Tsts --= k 2 Na — x . (7.71)
Here we have assumed that the numbers of steps s i and 82 are roughly equal
and can be absorbed into the constants a, le i and k2 . The performance of the
 A CASE STUDY ON NUMERICAL PERFORMANCE 383

stochastic wave function method versus that of the density matrix integration
can be measured by the difference between the exponents, which is 1 in the
non-self-averaging (x 0) and 2 in the strongly self-averaging case (x =1).
7.4.2 The damped driven Morse oscillator
Let us illustrate the scaling laws (7.70) and (7.71) for the CPU time consumption.
To this end, we consider a non-trivial example, namely a damped Morse oscillator
which is driven by a time-dependent force. It will be seen that the scaling laws
are very well satisfied for this non-linear problem.

7.4.2.1 Description of the model The Hamiltonian describing the coherent

part of the dynamics is taken to be

Hs(t) = Hm + (7.72)

where
1 no

Hm = + V (q) (7.73)

and

V (q) = D[1— exp(—bq)1 2 (7.74)

is the Morse potential. The Morse Hamiltonian (7.73) may be used to model,
e.g. a molecular degree of freedom within a single electronic potential energy
surface. With an appropriate choice of the parameters it yields a fairly realistic
description of, for example, the vibrational dynamics of the local O-H bond in
the water molecule. The external driving is described by the interaction term

HL(t) = pqFo s(t) sin cu Lt. (7.75)

Here, fig is the relevant component of the molecular dipole moment, F0 is the
maximum field strength and s(t) is the envelope of the driving pulse. To be
specific, we take

s(t) = sin2 (7.76)

tp
in the following, where tp represents the pulse length. For the physical back-
ground of this model see Manz and W6ste (1995).
The Morse Hamiltonian Hm has a finite number N of bound states which
depends on the parameters of the Morse potential V(q). The following inves-
tigation will be restricted to the case where the dynamics is well confined to
the bound state sector of the Morse oscillator. The simulation will therefore be
performed in the energy eigenbasis 1j), j = 0, 1, 2, ... , N — 1, of Hm with the
corresponding energy eigenvalues Ei which are known analytically.
384 THE STOCHASTIC SIMULATION METHOD

The jump operators are taken to be

k j, k = 0,1, . ,N — 1, (7.77)

with corresponding rates 7jk. These rates are assumed to be given by the expres-
sions for the weak-coupling master equation describing the dissipative dynamics
in a thermal reservoir. The operators Ajk are eigenoperators of Hm belonging
to the eigenvalues (Ej — Ek ). The effect of the jump operator Aik on the system
wave function may thus be interpreted as a transition describing the emission
or absorption of a vibration quantum of energy lEj — Ek I. In the energy repre-
sentation the Markovian master equation for the matrix elements pik = (jlpsik)
therefore reads

dpik .
dt
= - 1(E - Ek) Pik - if (t) E (Q.j1 Plk Qlk Pii)

1
+6ik 7i1 MI) —(
— + rk)Pjk• (7.78)
2

The corresponding PDP is governed by the following equation for the components
of the normalized state vector 0,

1
dt .)
= —i(EljOi + f(t) EQikok) (E -yki - E-ykiloir) Oi
k k kl

-FE ( - oi) dNk, (0. (7.79)

kl

In these equations f (t) is the time-dependent external force (cf. eqn 7.75)

1(0 = 1.c.Fo s(t) sin(c Lt), (7.80)

and the Qj k = (j lql k) are the matrix elements of the dipole operator. The matrix
(Qjk) is real and symmetric. ri is the total rate of all jumps away from I i),

N-1
= E (7.81)
m.0

Finally, the Poisson increments satisfy

dNki (t)dNki (t) = 6kki 6u , dNki (t), (7.82)

E [dNki(t) =`Tiakbil 2 dt. (7.83)
 A CASE STUDY ON NUMERICAL PERFORMANCE 385

CPUtime per time step tsj 10.000

1.000

0.100

0.010

0.001 , i

32 ao 48 56 64 72 80
System Size N

FIG. 7.11. The CPU time per step for the integration of the density matrix
equation (A) and for the propagation of the wave vector V) (o). The dotted
lines indicate how the exponents a and /3 were calculated from the slope; the
continuous lines simply connect the data points.

7.4.2.2 Simulation results In the following we solve eqns (7.78) and (7.79) and
compare the numerical performance for both equations. The initial state is always
taken to be the ground state of the Morse oscillator. In order to study the effect
of the system size on the time consumption of the numerical routines, a series of
similar oscillators with varying number N of bound states was investigated. The
system size N was varied in the range N = 12, . . . , 78. In order to not just blow
up the number of states, with the actual dynamics always staying in the same
number of low-lying states, it is necessary to appropriately scale the driving field
as well. For the following study two parameter combinations are employed, one
corresponding to weak damping, the other to strong damping. We also note that
the same Runge—Kutta routine is used for the integration of the density matrix
equation and for the integration of the deterministic pieces of the PDP. The
details of the physical parameters used and of the numerical implementation are
described by Breuer, Huber and Petruccione (1997).
Let us first look at the exponents a and O' which were introduced in eqns (7.67)
and (7.68). Figure 7.11 shows the CPU time per time step of the numerical in-
tegrator as a function of the system size N. Measuring the slope of the lines one
finds

= 3.0 ± 0.1, a ------ 2.0 ± 0.1, (7.84)

which confirms eqn (7.69). These values can be easily understood: In the case
386 THE STOCHASTIC SIMULATION METHOD

-
Number ofsteps

--I

5.0•103
-

1 I I I

16 24 32 40 48 56 64 72 80
System Size N

FIG. 7.12. The triangles (A) display 8 1 , the number of integrator steps to cal-
culate one pulse using the density matrix equation. The diamonds (o) show
82 , the average number of steps to calculate one realization of the process.

of the stochastic simulation, the most time-consuming part is the multiplication

of V) with the dipole matrix Q (cf. eqn (7.79)), which requires 0(N 2 ) floating
point operations. Analogously, for the density matrix integration the calculation
of the right-hand side of eqn (7.78) for all j and k involves 0(N 3 ) floating point
operations.
Figure 7.12 displays the number of integrator steps 81 and 82 that are nec-
essary to calculate a whole pulse. As we can see s i and s2 increase with N, but
remain roughly equal. Their increase is due to particular properties of the pre-
sented example. The systems of differential equations which have to be solved
for the density matrix calculation and for the stochastic simulation both become
stiffer with increasing N. In particular, the ratio between the highest and the
lowest eigenenergy of the oscillator grows about linearly with increasing N.
In order to investigate the behaviour of R(N), the number of Monte Carlo
realizations that have to be generated to treat a system of size N, it is necessary
to look at the standard error of the observables of the system. Let us consider.
for example, the energy Hm of the Morse oscillator, whose standard error will
be denoted by H. According to eqn (7.66) we have

A ll-(N)
R(N) = 6,- 2 • (7.85)
H

Figure 7.13 shows the behaviour of the function AH(N) for two different environ-
ment parameter sets. Up to statistical fluctuations, A(N) and, therefore, R(N) is
 A CASE STUDY ON NUMERICAL PERFORMANCE 387

0.004

0.003 -

0.002 -

:
,

0.001

20 40 60 80
N

FIG. 7.13. The function All-(N) measures the self-averaging property of the
mean energy of the Morse oscillator Hm, cf. eqn (7.65). The triangles A
correspond to weak dissipation, the diamonds o to strong dissipation.

a non-increasing function of N. Similar results are obtained for other observables

and for the matrix elements of p. This means that, in order to achieve a constant
statistical error in the simulation results when the system size N is increased,
the number of realizations need not be increased. It follows that the stochastic
simulation will eventually, for large system sizes, be always faster than solving
the density matrix equation.
This result is exemplified in Fig. 7.14. The plots show the CPU times needed
to integrate the density matrix equation and to generate as many realizations
of the stochastic process as are necessary to obtain the standard error ô-H =
4.10 -3 of the mean oscillator energy. The number R of realizations was calculated
with the help of eqn (7.85). According to Fig. 7.13, we chose AH(N) = 10 -3
independently of N. The curves follow different power laws and at some point
No they intersect. In the present example, No r-t: 35 for the weak dissipation
case and No ',:ss.-, 55 for the case of strong dissipation. Above No , the stochastic
simulation is faster.
The main result of the numerical study is that the CPU time for the density
matrix integration TDmE and the time for the stochastic wave function simula-
tion Tst s scale with the system size N as

TDME ' N 3 / (7.86)

TStS ' N 2 . (7.87)
388 THE STOCHASTIC SIMULATION METHOD

Weak Disst anon Weak Dissipation

1.5 lo

Z'
Q 1.0
ct
E.

0.5

40 20 40
N N

1.0

20 40
N

FIG. 7.14. CPU times needed to integrate the density matrix equation (A) and
to generate as many realizations of the stochastic process (o) as are necessary
to obtain the standard error 6-H = 4. 10' for the mean oscillator energy.
The plots on the right-hand side cover the full variation of the system size
while the plots on the left side zoom in at lower N.

Although the numerical study was performed on a specific example, the consid-
erations made in Section 7.4.1 are far more general. In particular, whereas the
absolute values of the exponents in eqns (7.86) and (7.87) depend on specific
properties of the system under study, their difference is of more general signif-
icance. Under general conditions, the exponent in the expression for TDmE is
expected to be larger by 1 to 2 than the exponent of Tsts.
The quantities TDmE and Tsts that have been investigated above stand for
the time that the programs run on a single processor. When comparing the effi-
ciency of numerical codes that are designed for processing on a parallel computer
with many processors, other important criteria are speed-up and scalability. The
speed-up is defined as the ratio between the wall-clock times needed to do the
job on a single processor and on the parallel computer. Scalability means that
the speed-up is close to the number of processors of the parallel machine, for a
wide range of numbers of processors. This implies that little time is spent on
the communication between processors and that the synchronization overhead
is small. Whereas an efficient, scalable parallelization of the density matrix in-
 REFERENCES 389

tegration might be a complicated task, the stochastic wave function method is

an intrinsically parallel and very well scalable algorithm. Since the individual
realizations of the sample are generated independently, the only communication
needed is for the final averaging or archiving, and for parameter control. Monte
Carlo methods are therefore ideally suited for parallel processing.

References
Binder, K. (ed.) (1995). The Monte Carlo Method in Condensed Matter Physics.
Springer-Verlag, Berlin.
Blue, J. L., Beichl, I. and Sullivan, F. (1995). Faster Monte Carlo simulations.
Phys. Rev., E51, R867—R868.
Breuer, H. P., Huber, W. and Petruccione, F. (1996). Fast Monte Carlo algo-
rithm for nonequilibrium systems. Phys. Rev., E53, 4232-4235.
Breuer, H. P., Huber, W. and Petruccione, F. (1997). Stochastic wave-function
method versus density matrix: a numerical comparison. Comp. Phys. Com-
mun., 104, 46-58.
Breuer, H. P., Dorner, U. and Petruccione, F. (2000). Numerical integration
methods for stochastic wave function equations. Comp. Phys. Commun., 132,
30-43.
Ferrenberg, A. M., Landau, D. P. and Binder, K. (1991). Statistical and sys-
tematic errors in Monte Carlo sampling. J. Stat. Phys., 63, 867-882.
Fricke, T. and Schnakenberg, J. (1991). Monte Carlo simulation of an inho-
mogeneous reaction-diffusion system in the biophysics of receptor cells. Z.
Phys., 83, 277-284.
Gillespie, D. T. (1976). A general method for numerically simulating the
stochastic time evolution of coupled chemical reactions. J. Comp. Phys., 22,
403-434.
Gillespie, D. T. (1992). Markov Processes. Academic Press, Boston.
Hanusse, P. and Blanché, A. (1981). A Monte Carlo method for large reaction-
diffusion systems. J. Chem. Phys., 74, 6148-6153.
Kloeden, P. E. and Platen, E. (1992). Numerical Solutions of Stochastic Dif-
ferential Equations. Springer-Verlag, Berlin.
Knuth, D. E. (1981). Seminumerical Algorithms. Addison-Wesley, Reading,
Mass.
Landau, D. P. and Binder, K. (2000). A Guide to Monte Carlo Simulations in
Statistical Physics. Cambridge University Press, Cambridge.
Maksym, P. A. (1988). Fast Monte Carlo simulation of mbe growth. Semicon-
ductor Science and Technology, 3,594 596.
-

Manz, J. and Wiiste, L. (eds.) (1995). Femtosecond Chemistry. VCH, Weinheim,

New York.
Press, W. H., Flannery, B. P., Teukolsky, S. A. and Vetterling, W. T. (1992).
Numerical Recipes. Cambridge University Press, Cambridge.
 8
APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

Chapter 6 provided a number of simple examples for stochastic processes in

Hilbert space and their derivations in the framework of continuous measurement
theory. The present chapter deals with a presentation of the theory on the basis of
the microscopic equations of quantum electrodynamics (QED). The emphasis lies
on the derivation of the quantum operations and the corresponding stochastic
processes for various detection schemes directly from the Hamiltonian which
describes the interaction of the matter degrees of freedom with the quantized
electromagnetic field. Throughout this chapter we will treat the matter degrees
of freedom non-relativistically, having in mind applications to quantum optics.
After a brief survey of the quantization of the electromagnetic field we start
with the derivation of the general expression for the QED operation which de-
scribes the back-action on the matter degrees of freedom induced by the measure-
ment on the field variables. From the general expression we deduce the stochastic
representation for the multipole radiation of the matter current. The stochastic
dynamics will be formulated in terms of a stochastic equation of motion for the
reduced density matrix of the source. For the case of a complete measurement
of the quantum numbers of the radiated photons a stochastic evolution equation
for the state vector of the source is obtained. We shall also deal with incom-
plete measurements. They lead to an unravelling of the dynamics in the form
of a stochastic density matrix equation which does not preserve the purity of
quantum states.
The chapter includes a detailed discussion of some important physical prob-
lems for which the interplay of coherent quantum evolution and dissipative pro-
cesses plays an important rôle. First, we investigate the emergence of dark states
in the interaction of atoms with coherent laser fields. In addition, we study the
mechanism of coherent population trapping in the sub-recoil cooling of atoms
in laser fields. The corresponding piecewise deterministic process for the quan-
tized motion of the atoms is derived. It provides an interesting example of the
emergence of long-range Lévy-type distributions resulting from a quantum in-
terference effect. Finally, we examine dissipative phenomena in the dynamics of
systems in strong periodic driving fields. Appropriate stochastic wave function
dynamics can be derived by employing the representation of the reduced system's
state vector in a basis consisting of Floquet states.
 CONTINUOUS MEASUREMENTS IN QED 391

8.1 Continuous measurements in QED

8.1.1 Constructing the microscopic Hamiltonian
To begin with, we briefly recall the quantization of the free radiation field (see,
e.g. Bjorken and Drell, 1965) which may be represented by the free field operator
t) for the vector potential. Imposing the Coulomb gauge condition

• ;i ( , t) = o, (8.1)

electric and magnetic field operators are given by"

1 0 -
(8.2)
Ê(-1' t) =
fi(Z, t) = x (8.3)

The field operators are Heisenberg picture operators satisfying the wave equation

( 182 A)
o. (8.4)
e2 ot2

For the following discussion it will be helpful to decompose the radiation

field into a complete set of mode functions Ûx (x) labelled by some index A. They
satisfy the Helmholtz equation
2
(A ± ) (Z) = 0 (8.5)

with suitable boundary conditions, and are normalized such that

i.
f d3x ÛX() • tfA, ( 2)- = (8.6)

The quantity cdJA denotes the frequency of the mode A. We further impose the
condition that the modes are transverse,

f • CA (Z) = 0 1 (8.7)

such that the completeness relation for the modes takes the form

= (5y ( i, it ) EE ( & ii — — (8.8)

A
A

Here, (5,? ("Z , r) is the transverse 6-function which, applied to any vector field,
projects onto the transverse component of the field.
14 1n this chapter we write explicitly all physical constants such as Planck's constant h, the
speed of light c, electron charge e and mass m. Further, we use Gaussian units such that the
fine structure constant is given by a = e 2 /he 1/137.
392 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

On using the mode functions we can represent the vector potential as,

A4(Y, t) = E tiac2 (OA (Z)e —iwA t b), + C/(x)e iwA t btx ) , (8.9)
WA
A

while the electric field operator takes the form

f(,t) = iE v27rhw, (ri, ( y)e—twAtb,„ _ c(i)etwAtbtA ) (8.10)

A

The operators btA and b), are the creation and annihilation operators for the
photons of mode A, satisfying bosonic commutation relations,

[bA l btv ] = (5 AA 1 / (8.11)

[bA, bAil -= [btA , btv ] = 0. (8.12)

In particular, we denote by

IA) = btAIO) (8.13)

the one-photon states, where IC» is the vacuum state of the field. Exploiting these
commutation relations one obtains the equal-time commutation relations for the
field operators, e.g.

[Ei (x , t),Ai(P,t)] = 47rihc 4. (4 i').

- (8.14)

The projection onto the transverse component embodied in the transverse 5-

function on the right-hand side of this equation ensures that the commutation
relations are compatible with Gauss's law t • f = 0 for the free field. Finally, in
the radiation gauge the free Hamiltonian of the transverse degrees of freedom of
the field can be written as

= 1 1
f d3 x [g2 ± fp] = 87r
HB f d3x [f2 ± A.(_,A)11]
8ir

--_, E tiWA (btA bA ± —1 ) (8.15)

2
A

The degrees of freedom of the electromagnetic radiation field will now be cou-
pled to the electronic degrees of freedom of a bound quantum system, an atom.
for example, whose self-Hamiltonian is denoted by H. The Hamiltonian H1 de-
scribing the interaction between the electrons of the system and the radiation
field can be written as

Hi(t) = — f d3 x { .-./( i,- t) • if(i, t) + e AZ,

• t) fi(Z, t)} . (8.16)
c me
 CONTINUOUS MEASUREMENTS IN QED 393

We treat the matter degrees of freedom in the non-relativistic approximation and

neglect the diamagnetic term which is proportional to the square of the vector
potential. The quantity j(, t) is the paramagnetic electron current density,
1
je (
,t), E—
27n Va ) (t)6 (I — I ( a) (t)) + 6 (.I — 20) (0) /5(a) (t) } , (8.17)
a

whereas gri, t) is the electron spin density,

Az, t) = E 6 (Z — I(a) (t)) —h 6"(ct) (t)

•
(8.18)
a
2

Later on we shall also use the expressions for the electron density

Q(, t) = E 6 (I — i(a) (t)) , (8.19)

a

and for the orbital angular momentum density

r(Z, t) ---, na x jtv, t)

= E 1{ (a) (06 (I — Y(a) (t)) + o (I — Z(a) (t)) Da) (t) } . (8.20)
a

In these formulae the index a runs over the electrons of the system, and i(a) (t),
0'0 (0, i ( a) (t), and hei( a) (t)/2 denote, respectively, position, linear momentum,
orbital angular momentum, and spin operators of the particles in the interaction
picture. Equation (8.16) thus describes the interaction of the electromagnetic
field with the paramagnetic electron current density j and with the density of
the magnetic moment associated with the electron spin density. Note that the
interaction Hamiltonian Hi (t) as given in (8.16) involves the interaction picture
operators for the field and for the matter degrees of freedom, such that H1 (t) is
the interaction picture Hamiltonian.

8.1.2 Determination of the QED operation

Let us now derive the general expression for the operation pertaining to a con-
tinuous monitoring of the electromagnetic field variables. To this end, we first
take an ideal case and suppose that we can perform a complete, orthogonal
measurement of the one-photon states IA). This means that we assume some
measurement device which is able to detect the quanta radiated by the source
together with a measurement of a complete set of quantum numbers A of the
photon.
Following our discussion in Sections 6.2 and 6.3 we consider again the field
vacuum 10) acting as the probe state. In accordance with our previous notation
we denote by 7- the considered time interval of the interaction between object and
probe system, that is, between source and radiation field. The arbitrarily chosen
394 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

initial time is denoted by to , such that t = to + T is the final time at which the
measurement on the radiation field is performed. Also, the initial density matrix
of the source at time to is written as 5- .
Employing second-order perturbation theory we then find with the help of ex-
pression (8.16) for the interaction Hamiltonian that the corresponding operation
is given by

= QA&Q tA , (8.21)
where the operator ft takes the form (cf. eqn (6.86))

,f
S-2 )
to
(A1111- (e)10 =
27c2 f
hci.JA
to
dtteiwAti

x f d 3 x {:iV,t') • 0;
- (I) + --e—grY,e) • [t* x 0]}
- . (8.22)
rue
The operation for the event that no photon is detected over the time interval r
can be written as

(8.23)
where
t'
= —
1
f f
to to
dt"(01H1(e)Hi(t")10). (8.24)

Inserting a complete set of states, which may be restricted to the one-photon

sector here, we get the equivalent expression
t

J
to
fdt"Epo-Lt(e)10) -t(AIH,r(t")14
to

These relations will be used below to determine the operation for the detection
(8.25)

scheme.
In order to deal with the time integration involved in the operations (8.22)
and (8.25) we shall decompose the various densities into eigenoperators of the
Hamiltonian Hs of the source. These eigenoperators are defined through (com-
pare eqn (3.120))

, w) = E II(E) 0) II(E'), (8.26)

E' E=hw

with analogous relations for the particle density p, the spin density and the
orbital momentum density r.
The operators f i, 0), QV, 0), 0) and
( 0) 4.z,
 CONTINUOUS MEASUREMENTS IN QED 395

denote the corresponding Schrödinger picture operators taken at a fixed time

t = 0. Furthermore, II(E) denotes the projector onto the eigenspace of Hs
corresponding to the energy eigenvalue E and the sums are extended over those
energies E and E' which belong to a fixed energy difference of hw.. We assume
that H8 describes a bound system such that the interaction picture densities are
given sums taken over a discrete set of system frequencies co,

( t) , E e-iwvry, w) + h.c., (8.27)

w
C iw t o(Z, w) + h.c. , (8.28)
Lo
AZ , t) = E e —i 't g(Z, w) + h.c., (8.29)
w
t) =E ciwtrri, w) + h.c. (8.30)
w

On using this decompositions into eigenoperators we can write the operator

S-2À which determines the operation for the event that a photon with quantum
numbers A is detected as follows,

2 c2 exp[i(coA—co)7-1 — 1
!IA = E exp[i(wA co)tol
A co,), — co
w

xf d3 x { .-i ( i- , c.4.)) • 03: (Z) + mee ,§(i , c.4.1) • [ f x ti'; (Z)] } , (8.31)

Since the mode frequencies co), are positive the sum over the system frequencies
w may be restricted to run over the positive frequencies only, which will always
be assumed in the following.
We see from eqn (8.31) that the operation defined by f2,), leads, in general,
to complicated short-time behaviour of the object system. In the following we
assume that the discrete system frequencies co are well separated with a minimal
distance which is large compared to A — 1/7 - , the energy uncertainty associated
with the interaction time r. Physically, this means that over interaction times of
order 7- we can distinguish the frequencies radiated by the system and identify
uniquely the corresponding transition of the object system. In other words, we
are considering a spectral detection of the object system with a finite resolution
which is small compared to the widths of the transitions but large in comparison
to the distance between their frequencies.
Under this assumption, we see that for a given frequency cox of the field modes
there is exactly one term which dominates the sum over co in eqn (8.31). This is
the term which fulfils the condition

wA E To E [co — A, cd.., + ,A]. (8.32)

396 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

Thus, by keeping only the secular term in the operation we obtain for the QED
operation (omitting irrelevant phase factors),

27c2 exp[i(wA — w)7] — 1

hCO À C 4., A - CA1

x f d 3 x {—e :IrY , co) • (TX (i) + —e--.§4ri,w) - ['- x 043:

C ITLC
rid} , (8.33)

where the system frequency w is determined by condition (8.32).

8.1.3 Stochastic dynamics of rnultipole radiation

As an example let us discuss the case of the spectral and angular momentum
measurement of the emitted photons. The complete set of quantum numbers A
measured by the device consists of the frequency of the photons, the square of
the total angular momentum, the component of the angular momentum in a
fixed direction éz , and the parity quantum number. Hence, the index A is given
by the set of quantum numbers

A = (k, J, M, ii) , (8.34)

where J = 1, 2, 3, ... , M = J, J + 1, . . . , J — 1, J, and it = 0, 1. The one-

— —

photon states IA) thus describe free photons with definite frequency WA --=_L- wk =
ck. They are eigenstates of the square of the total angular momentum of the
field, of the z-component of the total angular momentum, and of the parity
operator with respective eigenvalues h2 J(J + 1), hM , and (-1).1-7 . Our aim
is to derive a stochastic differential equation for the density matrix a(t) of the
source corresponding to this measurement scheme.
As is well known the mode functions OA(Z) represent electric multipole fields
for it = 1 and magnetic multipole fields for it = 0 (Akhiezer and Berestetskii,
1965). Coupling the orbital angular momentum and the spin of the photons, one
can express the mode functions in terms of superpositions of products of vector
spherical harmonics and Bessel functions, corresponding to their angular and
radial part, respectively. They form a complete set and can be chosen to satisfy
the normalization condition

f d3 x 0;,`
- (i) • tiv () = (5(k — (8.35)

We do not need here the explicit expressions for the mode functions; they
may be found, for example, in Shore (1990). In the range of optical frequencies
we may use the approximation kd < 1, that is we assume that the wavelength is
large compared to the linear extension d of the source. Consequently, only the be-
haviour of the mode functions in the vicinity of the origin is relevant. Substituting
the explicit expressions into (8.33) one obtains for the operation corresponding
to the detection of a photon with quantum numbers A = (k, J, M, it),
 CONTINUOUS MEASUREMENTS IN QED 397

CA kJ j +1 eXP[iPlc — WM — 1 IQ j m7r(w).
f2 1cJA17z- (8.36)
rh (2J — 1)!! J(2J +1)

Here we have introduced the frequency components of the electric 2J-pole mo-
ments,

\I f
Qjm ,7=1(W) = 47 d3 x ep( 2,w)rl
- y; 1 (Z), (8.37)
2J + 1

and of the magnetic 2 .1 -po1e moments,

Qm,7r=o(w) =1/ 2 7:F 1 f d3x inee [ j +1 1 r( -2, w) + W'(Z-, cd

) •
(8.38)

In these formulae the Yjm(4 denote the ordinary spherical harmonics, r E IA •

and n!! . 1 • 3 • 5 • for n odd.
It follows from eqn (8.36) that the spectral angular momentum measurement
of the radiated photons leads to a PDP with the quantum jumps

Chm7(w)eiCifi1,(w) (8.39)
tr s

For each particular set of quantum numbers (co, J, M, 7r) we thus have a certain
jump which is given by the application of the corresponding transition frequency
component IQ jm 7r (0) of the 2 .1 -pole moment.
The rate corresponding to the jump (8.39) is given in the Markov and rotating
wave approximation by the expression
co
1
— f dk trs fi - 21 jm „f2kJm71-6- } ,rz--,' 7,7(w)tr s {Q tJA4-7,-(w)C2J m7r(w)6- } , (8.40)
T
o
where
2(J + 1) k 2J + 1
(8.41)
7.7(w) = J(2J +1)[(2J — 1)!!]2 h

These are the well-known expressions for the transition rates of electric and
magnetic multipole radiation.
Finally, we need to derive the expressions for the event of no photodetection.
Invoking again the Markov and the rotating wave approximation, eqn (8.25)
leads to

(8.42)
c..)J M7
398 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

where we have neglected the Lamb shift which yields a contribution to the Hamil-
tonian Ils of the source. Therefore, we get the following expression for the op-
eration of the zero photon detection event,

Qo 6- Q ot
s'ss I — iG(a)dy, (8.43)
trs h

where the super-operator

ih
g (a) = — -y./(w)(c24-„(w)(2.7m7,-(coa+.0-QtJ„,(w)(2,/m7r(w))
ca M7r.
+ih E 7,,(w)trs {Qtjm .n.(w)Q jm„(w)o- } a (8.44)
c.0,1 Mir

represents the generator for the deterministic parts of the PDP.

Summarizing these results we have found that the spectral and angular mo-
mentum monitoring of the source results in a PDP which obeys the following
stochastic differential equation for the density matrix (7(0 in the interaction
picture,

do- (t) = —
h
G (0- ( t))dt (8.45)

±
,Jmn- trs{w,„,„r (w)(2Jm7r(w),,-(01

The complete jump statistics of the PDP is embodied in the Poisson increments
1
a(t) dNwjm „(t).

which satisfy

dNuj j17r (t) dNuj i ji mi ir, (t) =7 (5wc,) , (5 jj , (5m m/ (571- 7 , dNw jm 7 (t), (8.46)
E [dN„,jm„ (0] = -yj(w)trs {W I m ,r (w)Q jm 7 (c.o)o- (t)} dt. (8.47)

It is immediately clear that eqn (8.45) leads to a stochastic process for pure
states. As already discussed in Chapter 6, this is directly connected to the fact
that we have performed a measurement of a complete system of observables of
the photon, and because the back-action induced by these measurements depends
on the photon frequency c,., /, only through the frequency cd..) of the transition of
the source.

8.1.4 Representation of incomplete measurements

Let us consider the specific case of electric dipole radiation (J = 1, 7 = 1). The
jump operators may then be written as

(2 1 m,„= 1 (w) E-. éli- • Â(w), M = 0, ±1, (8.48)

 CONTINUOUS MEASUREMENTS IN QED 399

where
1
•ect = e±i = [ix ± iiy] (8.49)

and

= f d3 x e, (8.50)

is an eigenoperator of the total dipole moment

B(t) = f d 3 x e Z, trX' Ee—uotif(w)± h.c. (8.51)

corresponding to the transition frequency w.

A complete measurement for electric dipole radiation thus encompasses the
detection of the transition frequency w and of the angular momentum component
hM of the photon along the fixed direction ez . The stochastic evolution equation
(8.45) then directly yields the following stochastic equation for pure states,

[ Oil • 24*(w)0(t)
do(t)=_G(i,(0)dt+E E 0(01 corw,m(t),
m=0,± IICM • if(w)0(t)ii
(8.52)

where the generator of the deterministic pieces of the PDP is

ih
G(0 )
2 E-y(w)(41(w)•if(w)—(ift(w)-24*(w ,b)//) ))

ih
EH 2 E (c4) ) (Al GO • if(w)) ,0
+ — (8.53)

The expectation values of the Poisson increments dN,,,, m (t) are given by

E [dN„,m (0] =- y(w) in, • if(c0)0(011 2 dt, (8.54)

where
\ 4 w3
7(w./ = 3 he3 (8.55)

which gives the well-known transition rate for electric dipole radiation.
We now suppose that the measurement of the photon spin component hM
is carried out on the non-selective level, that is the information on the photon
angular momentum along ez is thrown away. The stochastic process then loses the
property of transforming pure states into pure states under the time evolution.
400 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

The dynamics must therefore be described in terms of a stochastic evolution

equation for the density matrix a(t) which takes the form

do- (t) = — —hi G(o- (t))dt

A(w)a(t)At
(w)o- Mill (w)
± ,.., [tr s {At (w) • if(w)0-(0}
1
o- (t) dN,(t).

As can be seen the operation pertaining to the detection of a photon with fre-
(8.56)

quency c.,.) now involves an incoherent sum over the angular momentum compo-
nents,

24*(w)(7211(w) =E M=0, ±
(e11`1 (w))un • 2 4*(w)) f .
• 24+ (8.57)

This expresses the fact that no information on the angular momentum of the
photons is obtained during the monitoring of the field. As a result the statistics
of the Poisson increments depends only on the mean transition rate obtained by
averaging over the angular momentum components,

E [dN, (t)] = -y(w)tr s {Al (w) • 2i(c4))0-(01 dt. (8.58)

One can easily imagine a situation in which even the information on the
frequencies of the photons is thrown away. In this case we obviously have only one
type of quantum jump. These jumps are counted by a single Poisson increment
dN (t) which satisfies

E [dN(t)] = E-y( w)tr s {Al ( w ) • if(w)0- ( 01 dt. (8.59)

w
The stochastic differential equation for the resulting PDP thus reads

do- (t) =

±
[ E, 7(4))24*(c4))o- (t)iff (w)
o- (t) dN (t). (8.60)
trs { E, 7(w)ift (co) • A(w)a(t)}}

In summary, we observe that by successively reducing the number of ob-

servables measured selectively in the detection scheme one obtains different
stochastic evolution equations, involving different jump super-operators which
contain an increasing number of summands. This results from the smaller amount
of information which is extracted from the system: The coarse graining of the
detection scheme leads to a back-action on the object system which describes a
stronger increase of entropy.
 DARK STATE RESONANCES 401

le, me = 0)
CT+
o-_

■•■ ••1■11111.

z
lg, m g = —1) Ig,mg = +1)

FIG. 8.1. Level scheme leading to a dark state resonance by a quantum in-
terference effect. The figure shows the coupling between two J, = Jg = 1
angular momentum manifolds through two counter-propagating laser beams
with different helicities o-±.

On the other hand, as can be seen from the equations derived above, the
generator g for the deterministic pieces of the PDPs is always the same. This is
due to the fact that the deterministic part of the process corresponds to the event
of zero photon detection and is, thus, not influenced by the specific form of the
detection scheme. This can also be deduced directly from eqn (8.25) which shows
that the operation for the zero detection event depends only on the one-photon
subspace, but not on the specific basis of one-photon states chosen. The same
situation already occurred in the case of homo dyne photo detection (see Section
6.4).

8.2 Dark state resonances

In order to illustrate the foregoing discussion we investigate in this section a
transition between two angular momentum manifolds (Molmer, Castin and Dal-
ibard, 1993). This example leads to the emergence of a certain trapping state
of the process, known as a dark state, which will play an important rôle in the
following section.

8.2.1 Waiting time distribution and trapping state

We consider a two-level atom with transition frequency w o . Both levels are three-
fold degenerate forming manifolds with total angular momentum J, = 1 (excited
level) and Jg = 1 (ground level). The level scheme is depicted in Fig. 8.1. We
introduce the energy eigenstates g, m g ) and e, m e ) which are simultaneously
eigenstates of the z-component of the atomic angular momentum operator with
eigenvalues hrng and hrne , respectively. The frequency component A(w 0 ) .,21_4 of
the atomic dipole operator _6 can then be written as
402 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

ii4 = E 19, Tng)(9, mg Iblejne)(e, me'. (8.61)

mg,me ,
=0 ±1

Since b' is a vector operator, .244. can easily be expressed in terms of the reduced
matrix element of the dipole operator, which will be denoted by d, and the
Clebsch-Gordan coefficients (1m e 1M 1m9 ) which describe the coupling of the
atomic angular momentum and the photon angular momentum with components
M = 0, ±h.
In addition to the coupling to the vacuum of the radiation field the atom
is subjected to a resonant laser field with frequency coi, = wo , which is linearly
polarized in the y-direction. One observes that the transition m e = 0 mg = 0 is
forbidden since the corresponding Clebsch-Gordan coefficient (1010110) vanishes.
If we take some initial state in the manifold spanned by the states 1g, m y = ±1)
we then find that the dynamics is confined to the subspace spanned by the basis
states

le, m e = 0), 1g, m g = +1), g,mg = - 1). (8.62)

The representation for the frequency component of the dipole operator in this
subspace is found to be
d
Ax = (8.63)
id
A y = — -y (1g, +1)(e,01 - 1g, -1)(e,01), (8.64)
A, = 0. (8.65)
On using the resonance condition the Hamilton operator describing the linearly
polarized laser field can be written in the interaction picture as

HL = -Eo (Ay + Aty ) = ih9 -1 0 0 , (8.66)

2 (
+1 0 0)
where 9 = dEo /h denotes the Rabi frequency and E0 the laser field amplitude.
The second equation shows the matrix representation in the basis (8.62). If we
note further that

,214t • A= AtAx + Aty Ay = 41 2 1e, 0)(e, 0 1, (8.67)

we can write the non-Hermitian Hamiltonian ii in eqn (8.53)

ih ( 7 9 —9 )
k = HLL, —
ih;Y le,o)(e, 0 = - 2— -9 0 o), (8.68)
9 0 0
where the atom-laser interaction has been added to ft. Note that we have ab-
sorbed the square of the reduced matrix element into -y which thus has the
dimension of inverse time.
 DARK STATE RESONANCES 403

On using these results it is easy to write down analytical expressions for the
non-Hermitian evolution and the cumulative distribution F[0, t ] for the random
waiting time t for the PDP defined by the general equation (8.52). Since after
any jump the state always ends up in the manifold spanned by the states g, ±1),
it suffices to determine these quantities for the states lying in this subspace, that
is, for the states of the form
o
(to+ ) •
(8.69)
(
(P-
On diagonalizing k one easily finds that for these states

1 2 (V+ - V-) 9(0

exp t)
--fi co = - (p+ (1 — h(t)) + cp_ (1 + h(t)) (8.70)
(
h

whereas the waiting time distribution is given by

F [co,t] =1— Ilexp(—ikt/h)(P12 (8.71)

1
= 1 - - (1+ h2 (t) + 2g2 (t)) - 1? () (1- h2 (t) - 292 (t)) .
2
Here we have introduced the abbreviations

g(t) = _ .it/ 4 sin (1) ,

tt
h(t) = _ e — t,'4 [cos (1) + iii sin (a
t ,

with
,y 2
2112 - ' (8.72)
4 •
In the following this quantity is assumed to be real and positive.
The important property of this example is that the non-Hermitian Hamilto-
nian (8.68) has a zero mode which is given by the state

o
I. 1
O ne = — (19,+1 ) +1g, -1 )) = r-(1), (8.73)

that is, we have

fhP ric = o. (8.74)

It follows that F[0,,,, t] E 0. The state O n, is therefore a stable state of the pro-
cess for any measurement scheme, which means that, once this state is reached,
404 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

yt

FIG. 8.2. Four realizations of the PDP for a two-level system driven by a res-
onant laser field polarized in the y-direction. For each realization OM the
figure shows the population r(t) = 1(0nci0(t))I2 of the trapping state Onc•
The stochastic process corresponds to the measurement of the photon angular
momentum along the z-axis. The initial state is 0(0) = 1g, 1) and 0 = -y.

no further jumps take place and the state vector becomes independent of time.
Moreover, since the other eigenvalues of ft have negative imaginary parts, the
state 0, is attractive under deterministic evolution and may therefore be called
a trapping state of the process.
From the physical point of view the emergence of the trapping state results
from an interference effect as is easily seen by looking at the form of the non-
Hermitian Hamiltonian. The linearly polarized laser field may be represented by
the superposition of two counter-propagating fields in the z-direction (see Fig.
8.1). One component is right-circular polarized and one is left-circular polarized.
The right-circular component induces the transitions 1g, —1) -4 le, 0), whereas
the left-circular component yields the transitions 1g, +1) 4 le, 0). As can be see
-

from k the amplitudes for these transitions interfere destructively and cancel
exactly for the linear combination of the ground state manifold given the trapping
state (8.73). Therefore, 0, does not couple to the external laser field, thus
representing a dark state.
 DARK STATE RESONANCES 405

-
0.8
-
0.6 -
u-
0.4 - -

0.2

0
o 2 3 4 5 6 7 8 9 10
yt

1 2 3 4 5 6 7 8 9 10
yt

FIG. 8.3. The waiting time distribution function F[1g, 1),t] and the corre-
-

sponding density f (t) = dF[Ig , 1), t] / dt for the two-level system with a
-

dark state and 12 = -y.

8.2.2 Measurement schemes and stochastic evolution

Following Molmer, Castin and Dalibard (1993) let us discuss two different mea-
surement schemes, namely the measurement of the photon angular momentum
along the z-axis, and along the y-axis. For the measurement along the z-axis one
finds from eqns (8.63)-(8.65) two non-vanishing jump operators

. d
él` • A = ± g, +1)(e, Ok
— (8.75)

where we have taken the polarization vectors éci = éz and e'± = +Vs ±iey )/ 4.
The jump operator él` • 1 implies that the photon carries away the angular
momentum ±h along the z-axis, such that the atomic transition involves a change
of the z-component of its total angular momentum given by Th. We observe that
the trapping into the state 07„ results essentially from the non-unitary evolution
generated by H. This is illustrated in Fig. 8.2 where we depict the occupation

7r (t) E I(Onc10 ( t))1 2 (8.76)

of the trapping state for four realizations of the process corresponding to this
measurement scheme. As can be seen, the quantum jumps put the state vec-
tor into either 1g, +1) or 1g, -1), for which r(t) = One of the realizations
406 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

0.8 -

0.6

0.4 -

0.2

5 10 15
yt

FIG. 8.4. The same as Fig. 8.2, but now the photon angular momentum is
measured along the y-direction. The quantum jumps put the state vector
either directly into the trapping state, in which case 7r = 1, or into the state
perpendicular to it, such that r = 0.

shown in the figure evolves continuously into the dark state 0,,,. For the initial
state 0(0) = lg, 1) chosen in the figure such a realization occurs with a finite
—

probability which is given by the corresponding defect

1
lim F{Ig , — 1), t] = . - (8.77)
t-00
of the waiting time distribution. Figure 8.3 shows a plot of the cumulative dis-
tribution function F[1g, 1), t] and its density.
—

The situation is markedly different if we monitor the photon angular momen-

tum along the y-axis. For this case we take the polarization vectors 6 = gy and
è+± = +(é4z, + iex)/ -f . to obtain the jump operators

è-'6' . Â = (8.78)

él . Â = + -i4Onc)(e, 13 1, (8.79)
2
where we have introduced the state
1
— 191 -1 )), (8.80)
vh
 DARK STATE RESONANCES 407

0.25

0.2

Var(R)
Var (R)

0.05

Var2(R)

5 1 10 15
yt

FIG. 8.5. The different variances of the observable R = One) (Onel as a function
of time. The measurement scheme records the photon angular momentum
along the z-axis. Var(R) has been determined from the solution of the density
matrix equation, whereas Var i (R) and Var 2 (R) have been estimated from a
sample of 104 realizations.

which is perpendicular to the trapping state. The jump operator è-'8' • Â represents
the back-action in the case that the photon angular momentum component along
the y-axis turns out to be zero. It puts the atom into the state 1/),. By contrast,
the expression for el • Â shows that both measurement results +h for the an-
gular momentum component put the atom into the trapping state 'One . Thus,
the trapping state may be reached directly through a quantum jump which is
illustrated in Fig. 8.4.
Both these measurement schemes lead, of course, to the same density matrix
equation and cannot be distinguished if one looks only at the corresponding
ensemble of type Ep . As discussed extensively in Chapter 5, the two measurement
schemes can, in fact, be distinguished if the measurement is performed on the
selective level, that is, if we investigate the corresponding ensembles of type Sp.
To demonstrate this we plot in Figs. 8.5 and 8.6 the variances introduced in eqns
(5.12), (5.13) and (5.14) pertaining to the observable

R = iOnc)(Onci, (8.81)

which is just the projection onto the trapping state. The population 7r(t) defined
in eqn (8.76) is a random variable, which may be written as
408 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

0.25

0.2

0.05

yt

FIG. 8.6. The same as Fig. 8.5, but with the measurement scheme recording the
photon angular momentum along the y - axis.

7 (0 = (0(t)1R 10( 0). (8.82)

Thus we have
Var. ' (R) = E [7r (0] E [r2 (t)] , (8.83)
Var2(R) = E [7 2 (0] — (E [ 7 (t)]) 2 (8.84)
[7(t)] ) 2 . Var(R)=E[it]— (8.85)
Obviously, the sum of Var i (R) and Var2(R) is equal to Var(R) and is indepen-
dent of the measurement scheme. The figures clearly exhibit that after some
time Var i (R) becomes considerably smaller for the measurements along the y-
axis than for the measurement along the z-axis. This is easily understood in the
context of quantum measurement theory. As we have seen above the measure-
ments along the y-axis put the state vector into the trapping state On, or, else,
into the orthogonal state /P c . This measurement scheme therefore corresponds
to a measurement of R, that is, the scheme tends to generate a basis in which
R is diagonal. After all realizations have performed their first jump the mea-
surement scheme has produced an ensemble Ep consisting of eigenstates of R
such that Var. ' (R) = 0. On the other hand, the measurements of the angular
momentum along the z-axis project the state vector onto the states Ig, ±1) for
which Var. ' (R) = . In this case it therefore takes much longer for Var i (R) to
decrease to zero.
 LASER COOLING AND LEVY PROCESSES 409

The variance Var 2 (R) is a measure of the dispersion of 7(0 over the ensemble
Ep which, for the simulation of the figures, was taken to consist of 104 realiza-
tions. Since Var.' (R) is larger for the measurement along the z-axis, Var2(R) is
smaller for this measurement scheme. The conclusion is that, in order to mini-
mize the statistical fluctuations of 7(0 = (0(t)140(t)), one should try to obtain
as less as possible information on the quantum observable R. If one is only in-
terested in a numerically efficient simulation method, then one should use the
measurement along the z-axis to get better statistics.

8.3 Laser cooling and Lévy processes

The dynamics involved in the cooling of the translational motion of atoms by
means of appropriately designed laser fields provides important and interesting
examples for stochastic processes in Hilbert space. In fact, this is one of the first
examples for which the stochastic wave function dynamics has been used with
great success (Cohen-Tannoudji, Bardou and Aspect, 1992; Castin and Molmer,
1995). It turns out that the stochastic method enables one to deal with a series
of problems in the deep quantum regime (Cohen-Tannoudji, 1992a) where an
enormous variability of relevant time scales is involved, and where a treatment
on the basis of the optical Bloch equations is difficult.
Let us consider for example the so-called Doppler cooling which employs the
Doppler shift of the laser frequency as seen from the rest frame of the moving
atoms. The basic physical principle underlying this cooling mechanism (Cohen-
Tannoudji, 1992 b) is the following. We take two laser beams counter-propagating
along the z-axis. The laser frequency wL, of both beams is tuned below the rest
frame transition frequency w of a simple two-level atom in the laser fields, such
that the detuning 5 E wL, w becomes negative. If the atom is moving into the
positive z-direction, for example, it gets closer to resonance by the blue shift
of the frequency of the laser field propagating in the negative z-direction. The
atom is thus excited by the resonant absorption of photons from that beam
and, by conservation of linear momentum, experiences a radiation pressure in
the direction of the laser beam. The subsequent spontaneous emission of fluores-
cence radiation, however, is essentially isotropic in space which leads effectively
to a dissipation of energy and a net friction force exerted on the atom. This
friction force has the same direction as the laser beams. Thus, the two counter-
propagating beams yield a damping of the z-component p = pz of the atomic
momentum.
The dissipation of the atomic momentum is accompanied by fluctuating forces
which are due to the random momentum kicks of size hk = ha;/ c caused by the
spontaneously emitted photons. It thus appears that the variance Ap 2 of the
atomic momentum cannot be smaller than a fundamental limit which is given in
terms of the recoil energy
h2 k2
1
ERE = - kBTR (8.86)
2M 2
410 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

by the relation

Ap2 > 2MER, (8.87)

where M denotes the atomic mass. The recoil energy may be associated with a
certain temperature, known as the recoil temperature
h2 k2
TR - 7 . (8.88)
NI KB

As an example, for the transition 23 S1 23 Pi of helium one finds that TR

4pK.
This single photon recoil limit provided by the recoil temperature TR seems
to be a fundamental limit for any cooling mechanism which involves spontaneous
photon emissions. In order to overcome this limit one has to suppress in some way
spontaneous emission processes. As we have seen in the preceding section this
is indeed possible if the atomic wave function is driven into a dark state which
does not couple to the laser fields as a result of a quantum interference effect.
This is precisely the sub-recoil cooling scheme which will be discussed in this
section (Cohen-Tannoudji, 1992a, 1992b; Bardou et al., 1994). By using variants
of this scheme or by invoking carefully designed pulses it has been demonstrated
in a series of beautiful experiments that cooling temperatures in the nanokelvin
range can be achieved (Kasevich and Chu, 1992; Lawall et al., 1994; Reichel
et al., 1995; Lawall et al., 1995).
As will be discussed below the sub-recoil cooling mechanism leads to sev-
eral interesting properties of the statistics of quantum jumps. The destructive
interference of the transition amplitudes for the atom—laser interaction gives rise
to long-tail waiting time distributions which behave asymptotically like certain
stable Lévy distributions. In addition to important practical applications, the
sub-recoil cooling scheme thus provides a nice example of the emergence of Lévy
statistics as a result of quantum interference phenomena (Bardou, Bouchaud,
Aspect and Cohen-Tannoudji, 2001).

8.3.1 Dynamics of the atomic wave function

Up to now we have completely neglected translational degrees of freedom. This
means that the centre of mass 11 of the atom has been fixed at the origin of the
coordinate system where, making use of the dipole approximation, the electric
field is evaluated.
To derive the stochastic dynamics of the state vector for the cooling scheme
we must take into account the motion of the centre of mass of the atoms (Cohen-
Tannoudji, Dupont-Roc and Grynberg, 1998). The translational degrees of free-
dom will be treated fully quantum mechanically, that is, the conjugated variables
11 and atomic momentum P are regarded as operators satisfying canonical com-
mutation relations,

[Pi, = (8.89)
 LASER COOLING AND LEVY PROCESSES 411

The Hamiltonian Hs describing the free atom now consists of two parts, namely
the kinetic energy of the centre-of-mass motion and the energy of the internal
degrees of freedom. For simplicity we consider a two-level atom and denote by
II, and 119 the projections onto the manifolds of excited and ground level, re-
spectively. Hence we have
p2
Hs = + hr..oll, , (8.90)
2M
where c4.) is the atomic transition frequency. The ground state energy has been
taken to be zero.
The atom is subjected to two interactions: First, the interaction with a co-
herent laser field which will be described by a classical c-number field EL,(Y,t)
and, second, the interaction with the quantized radiation field whose electric
Schrödinger picture field operator will be denoted by E(ã). Invoking the dipole
approximation we may thus write the atom-laser interaction as

Ho), - 15 • EL (ii,t), (8.91)

whereas the interaction of the atom with the radiation field takes the form

HI (8.92)

The quantity :64 is the atomic dipole operator. Note that both electric fields are
evaluated at the centre of mass of the atom. The total Hamiltonian for the atom
interacting with the coherent laser and the quantized radiation fields therefore
reads

H(t) = Hs + HL(t) + H.!. (8.93)

To be definite we consider in the following the cooling scheme sketched at

the beginning of this section, namely, the electric field EL (ii, t) describes two
counter-propagating laser beams along the z-direction (see Fig. 8.1). The beam
propagating into the positive z-direction is right-circular polarized, the beam
propagating into the negative z-direction is left-circular polarized. Introducing
the corresponding polarization vectors,

(8.94)

we can write
E0
L41(fl,t) = — (4 exp{+ikLz} + é_ exp{-ikLz}) e'L t + c.c., (8.95)
2
where E0 is the field amplitude. As in Section 8.2 we consider the case of a
J, = 1 --> Jg = 1 transition. The lowering part  of the atomic dipole operator
412 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

is thus given by eqns (8.63)—(8.65). On using these expressions we find for the
atom—laser interaction in the rotating wave approximation

HL(t) = 2_ /— (l e 0)(9 +11e—i(wLt+kLz) — e, 0) (g, —id e —i(wLt—kLz)) + h.c.,

v2
(8.96)
where 1 = Eo d/h is the Rabi frequency.
Since the centre-of-mass coordinate fi is canonically conjugated to the atomic
momentum P the exponential exp(—ik fi) acts as a shift in momentum space,
that is

exp( • fi ) ,f d3 p 1154)(1T + hk41, (8.97)

where the states 123) are momentum eigenstates of the centre-of-mass motion
which are normalized as

(A Y ') = (8.98)
Hence we find

d3p le,045) ((g,+1,15+ hkL — ( 1 ,71— e —i wL t

1-10) =
+h.c., (8.99)

- E kLez and we have introduced the product states

where 4,

19,±1)01/5), le, 045) le, 0) (8.100)

These are simultaneous eigenstates of the kinetic energy of the centre-of-mass
motion and of the Hamiltonian for the internal degrees of freedom.
Equation (8.99) has an obvious physical interpretation: The atom—laser in-
teraction induces transitions from the ground to the excited state manifold. By
the conservation of angular momentum, in the transition g 1) 4 e 0) the , — - ,

atom absorbs a right-circular polarized photon from the beam propagating in

the positive z-direction. The conservation of linear momentum thus implies that
the atomic momentum changes by an amount +h-4,. Likewise, in the transition
+1) 4 le, 0) a left-circular polarized photon is absorbed from the beam prop-
-

agating in the negative z-direction such that the atomic momentum changes by

The form (8.99) for the laser interaction suggests introducing, in analogy to
the procedure of the preceding section (cf. eqns (8.73) and (8.80)), the states

Onc(P1 ) = + 1 , 15 + + 1 45 — )) ( 8.1 01 )
1
10c(A) — 1 .9,+ 1-,/1+ hiL) - ig -145 - (8.102)
v2
 LASER COOLING AND LEVY PROCESSES 413

which satisfy the normalization conditions

= 6iii 6 ( 54- /5"), (8.103)

where j and ii stand for 'Tv' or V. In terms of these states we can write the
atom-laser interaction Hamiltonian as
hf2
Hi, (t) = —
2 f d 3 p le, 045)(0, (fi) 1 e -i'L t + h.c. (8.104)

This form shows directly that the laser field couples only to the state 10 ( )),
whereas One (p)) is a zero mode of this interaction,
h("2
110)10c(15)) = —2 le, 0 ,25)e -iwLt , (8.105)
HL(t)Inc(p) = O. (8.106)

Let us now turn to the interaction of the atom with the quantized radiation
field. Using plane wave field modes normalized in a box of volume V with periodic
boundary conditions,
1
UÀ V) = (8.107)
VT/

we may write the electric field operator evaluated at the position ./1 of the centre
of mass as

27rru.ok ( m( k) exp
mil) ,iE - i (w kt - k4 . fi )] bm(k)
V

- qi (ij) exp [+i(wkt - k • lid btm (k)) . (8.108)

Here h k- ' is the photon momentum, Wk = ck the frequency, and the vectors ém (k)
denote the corresponding polarization vectors. We take here M = ±1, where
M = +1 corresponds to a right-circular, and M = -1 to a left-circular polarized
photon (helicity +1 and 1, respectively).
-

The operators bm(k) and btm (k) denote the annihilation and creation opera-
tors for photons with wavenumber k and polarization M. The operation describ-
ing the measurement of such a photon over the time interval T = t — to is found
to be
t
i ,
- h f e(ii
d MIlii(e)10)
to
i(Wk -007 _ 1
244
e it,,, k _w)t o e
71%
(k) • 2 4-4.e -4.R . (8.109)
hV cok - (-0
414 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

Thus we deduce that the back-action for the photon measurement is given by
(omitting an irrelevant phase factor)

g'74(k) • ife —ik.44 0

(8.110)

We see that the operation 1 kM acting on the atomic wave function V) involves
the operator exp(—a•fi) which induces a shift of the atomic momentum by --hi
(see eqn (8.97)). This shift describes the recoil of the atomic momentum caused
by the emission of a photon with momentum +hi.
Thus we observe that the measurement of the linear momentum of the photon
amounts to an indirect measurement of the recoil momentum of the centre of
mass of the atom. This is also seen directly from the explicit form for the back-
action which is obtained by making use of eqns (8.97) and (8.63). We find

—Md
(k) ife' R = f d3
p (cos 0 — M)Ig, +1,11 — hk)(e, 0,231
2 A/2
+ 0'4' cos + M)1g , 141 hi) (e, 0,p). (8.111)
( — —

This equation shows that the excited state decays spontaneously into a linear
superposition of the ground states of the atom whereby the atomic momentum
changes by —hk'. The amplitudes of the ground states depend on the direction
of the photon momentum hk through the polar coordinates (0, cp) of k.
With the help of eqn (8.111) we get for the rate dF of the emission of a photon
into the solid angle c/1'2(k) with polarization M

dF(k,M) = — E imp,m0112 (8.112)

T fi'EcIS-2(I;)
3
= 32r 1 'Y(+ cos 2 11,b)aNk), (8.113)

where -y = dI 2 / 3 he3. Equation (8.113) shows that the rate of spontaneous

radiation is independent of M and of the azimuth (p. It only depends on the
z-component

u hk = hk cos° (8.114)

of the photon momentum. According to eqn (8.113) the random variable u follows
the distribution
31 1 u )2 ]
(8.115)
=
(

q(u) *

which is normalized as
 LASER COOLING AND LEVY PROCESSES 415

±hk

f hk
du q(u) = 1. (8.116)

The above results yield the full three-dimensional structure of the process.
An effective one-dimensional description is obtained if we average over the un-
observed x and y-components of the atomic momentum and over the azimuth
cp and the polarization M. This means that we now describe the process on the
non-selective level with regard to these degrees of freedom. The jump operators
describing the back-action on the atom then depend only on the z-component u
of the photon momentum and project either on 1m +1) or on 1g, — 1), that is we
have the following two types of jump operator

4(u) = 1
12-
f dplg 1 +11 p — (8.117)

J_ (u) = 1 f dpl. q, —1,p — (8.118)

V-2-
where p E la, denotes the z-component of the atomic momentum.
Collecting our results we obtain the following stochastic differential equation
for the PDP describing the dynamics of the internal degree of freedom and of
the z-component of the translational motion of the atom,

f,
±hk
d(t) = --Ki G(0)dt + du E [ J8(*) dNu (t), (8.119)
s_, ivs(u)011 0]
, s

where the Poisson increments satisfy

E[dNu, s (0 ] = 2 I1J8(u)0(t)11 2 q(u)dt, (8.120)

dN„,s (t)dN„, ,s , (t) = 6,,, 6(u — OdNu , s (t). (8.121)

One should note that dN, s is a field of Poisson increments (see Section 1.5.4)
indexed by a continuous variable u for the jump sizes. The generator G(0) of
the deterministic motion is given by

G(0) = II 0 + — (01 1140, (8.122)

2
and the non-Hermitian Hamiltonian takes the form
, p2
H= ± h (-6 — ) 11, + HL. (8.123)
2M
The quantity 6 = wi, — w is the detuning between atomic transition frequency
co and laser frequency coL, which will be taken to be zero in the following, i.e.
O = 0. Here we have also performed a canonical transformation with the unitary
416 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

operator exp (—icoLll e t) in order to remove the explicit time dependence of the
Hamiltonian, such that the atom—laser interaction reads

hQ
I/L, = -- dl*,0,P)(0(p)1+ h.c. (8.124)

These equations describe the PDP for the cooling mechanism.

8.3.2 Coherent population trapping

To understand the essential features of the stochastic dynamics of the cooling
mechanism we introduce the family of the manifolds

M(p)= span fle, 0,p), 1g, +1,p+ hkL),Ig, 1,1) — hki,)}

- -

= span {1e, 0 ,P), 10724)), 104))} , (8.125)

which are labelled by the momentum p. We denote by II(p) the orthogonal pro-
jection onto the manifold M(p). It is then easily verified that this projection
commutes with the Hamiltonian ft, that is,

= 0. (8.126)

This fact is connected to the conservation of the total linear momentum of the
atom—laser system and implies that the manifolds M (p) are invariant under the
deterministic evolution of the process, which means that, once the state vector
has reached .A4 (p) through a jump, it remains in M (p) until the next jump
occurs. If we rewrite the jump operators as
1
./± (u) = —2 f dP (lOnc(P — u ± hkï, )) 1 1'04 — u+ hkL))) (e, 0,p (8.127)

we see that a jump with J±(u) leads to a change of the manifold given by

.A4 (P) ---4 M (p — u ± hkL). (8.128)

Remember that u is a random number in the interval [—hk, +hk], which is dis-
tributed according to the density q(u) given in eqn (8.115). Physically, u describes
the random recoil due to spontaneous emissions.
Let us take the states le, 0,p), lOne(P)), lOc(P)) as basis states of a certain
manifold .A4 (p). The non-Hermitian Hamiltonian ft(p) which generates the de-
terministic evolution periods in M(p) can then be written as
/ n2
ft(p) = 2I'm + ER) / + f/(P), (8.129)

where I denotes the 3 x 3 unit matrix, ER is the recoil energy (8.86) and we
have introduced the non-Hermitian 3 x 3 matrix
 LASER COOLING AND LEVY PROCESSES 417

c.)

FIG. 8.7. Plot of the widths F(p) and Fe (p) of the ground state manifolds as
a function of the atomic momentum p. The parameters chosen correspond to
the transition 23 S1 23 P1 of helium.

( — ER — ih7/2 0 hQ/2
1-7'(p) = 0 0 hkplM (8.130)
hf2/2 hkplM 0

Note that the off-diagonal elements describe the various couplings, namely the
atom—laser interaction

(e,O,p`IftiOc(P)) = (e, O I PI IHL 10(p)) (8.131)

and the coupling between the ground state levels,

(041 01- 10nc(P)) = (041 )1P2 / 2MOne(P)) = hm

kP 11), (8.132)

leading to the Doppler energy hkpIM. It is this last term which yields a coupling
between the internal degree of freedom and the translational motion, called the
motional coupling.
For simplicity we set h = 1 in the following. As a specific example we shall
investigate below the laser cooling on the transition 23 Si 23 P1 of helium for
which we have 27r/k = 1.083 p,m, 7/27r = 1.6 MHz, and ERIh-y = 0.027. The
Rabi frequency is chosen to be 1 = 0.37. This was the first system for which the
cooling mechanism described below has been demonstrated experimentally. All
numerical values and simulations given below refer to this example.
418 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

The key point of the sub-recoil cooling process is the spectrum of the complex
eigenvalues of the non-Hermitian Hamiltonian '1'7(p) and the resulting structure
of the waiting time distribution
F[0, T] = 1 — i f/ (P)T)0 1 2 • (8.133)
If p = 0 we immediately read off from eqn (8.130) that ionc (o))
is a zero mode of
17(0). This is precisely the dark state which was already encountered in Section
8.2. Due to the atom—laser interaction and the motional coupling all three levels
develop a finite width for non-vanishing momentum p. These widths are defined
as twice the negative imaginary parts of the complex eigenvalues Ee(p), Enc (p ),
and e c (p) of f7 (P),
Fe (p) = (8.134)
F„(p) = —2a 6„,(p), (8.135)
Fe (p) (8.136)
The corresponding exact eigenvectors of the non-Hermitian Hamiltonian 1 7 4)
will be denoted by 1.t e (p)) , (1)nc(P))1 and (1>c(p)). Here we have labelled the
eigenstates such that their continuously connected states at p = 0 have the largest
weight on the states 10,(0)), 10n ,(0)), and 10,(0)), respectively. In particular,
F(p) -4 0 and 14 nc(P)) -4 One (0)) for p 4 0. In fact, it is easily found with
-

the help of the characteristic eigenvalue equation that F(p) is given for small
p by the expression
4ER ) 2 (p 2 -

Fn, (p) = 7 ( —) + 0(p4 ). (8.137)

Q k
The width of the state 143,2 (p)) which has a large contribution from the state
10/2 (p )) thus vanishes as /32 . This is illustrated in Fig. 8.7 which shows a plot of
F, (p) and Fc (p) as a function of the momentum p. The third width (p) may re
be found from the condition that the sum of the widths is identically equal to 7.
We observe a dip in the curve for F(p) and a peak of the width F(p) around
p = 0. For large values of the momentum both widths are of the same order and
decrease due to the Doppler shift.
The plot of the decay rates shows that around zero momentum F(p) and
F(p) differ by orders of magnitude. This fact has a drastic influence on the
properties of the waiting time distribution (8.133). As we have seen the atomic
state vector enters some manifold .A4 (p) (eqn (8.125)) through a jump. If p is
exactly equal to zero, then with a probability of 1/2 the atom remains trapped
forever in the manifold .A4 (p = 0). This is due to the fact that in both states
1g, +1, +k) the amplitude for being in the non-coupling state 17,G,(p = 0)), which
is an exact zero mode of 1 7 (p = 0), is equal to 1/Na We illustrate this in
-

Figs. 8.8 and 8.9 which show the waiting time distribution (8.133) for the initial
states 1g, +1,p + k) and 10,(p)) as a function of the waiting time T and of the
momentum p.
 LASER COOLING AND LEVY PROCESSES 419

100 -2
Y p/k

FIG. 8.8. The waiting time distribution F{Ig , +1, p + k), r] as a function of the
waiting time T and of the momentum p. As a result of the quantum interfer-
ence in the laser excitation, with probability an atom which is initially in
the state I g, +1, k) never leaves the manifold M(0).

For small but non-zero momentum p the atomic state vector leaves the man-
ifold M (p) with certainty. However, the waiting time T for the next jump can be
extremely long: Both states I g, ±1, p ± k) have a large contribution (of approx-
imately 50%) from the state 14',,,(p)) which is an exact eigenstate of 'f/(p) with
an extremely small width F„(p). What happens is that during the deterministic
evolution period following the jump into M (p) the atomic state vector is rapidly
driven into the state 14'724)) where it remains trapped for a long time. The
approach to the state 14.„,(p)) is very fast since the widths Fe (p) and Fe (p) are
by orders of magnitude larger than the width F(p).
The above picture is confirmed by numerical simulations of the process de-
fined by the stochastic differential equation (8.119). We show in Fig. 8.10 a single
realization of the PDP describing the cooling process. The realization involves a
total number of 104 quantum jumps. The picture shows a remarkable fact which
is typical for the sample paths of the process, namely that the total evolution
time is dominated by the waiting time T for a single quantum jump of about
7T = 5 107 . Note that the momentum is constant between the jumps. As will
be demonstrated in the next subsection the statistics of the waiting times is
governed by a long-range distribution similar to a stable Lévy distribution.
It is also obvious from the figure that the long waiting times are interrupted by
a series of very small ones. This is illustrated in Figs. 8.11 and 8.12 where we have
420 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

FIG. 8.9. The waiting time distribution F[1(p)), 71 for the initial state
1624)). Due to the quantum interference in the laser excitation, the prob-
ability of a jump out of the state I0 7„(p)) is strongly suppressed for small
momenta p.

depicted enlarged sections of Fig. 8.10. It is this proliferation of vastly different

time scales which makes the dynamics so difficult for a treatment based on the
optical Bloch equations. In the stochastic simulations, however, the treatment
of the long-range waiting time distribution does not cause any problems: On
using the inversion method, for example, even the longest waiting time period
can be calculated in a single step. It is thus not necessary to adapt the numerical
time step to the smallest time scale of the problem; the stochastic simulation
algorithm chooses an appropriate time step automatically.
Since I 4) 72c (P)) is very close to 10,24)) for small p we can thus say that the
atoms that fall into a small region around p = 0 can be trapped coherently with
a high probability for a long time and that an extremely large fraction of those
atoms is with certainty in the state IOnc (p)). In order to have really a cooling
process we need, however, a further mechanism which brings at least a part of the
atoms into the vicinity of p = 0. Such a mechanism is provided by the momentum
diffusion due to the random kicks caused by the spontaneous emission processes.
Thus we see that the cooling of the atomic momenta results from the interplay
of two basic processes, namely from the diffusion of momentum, which is due
to spontaneous emissions, and from the coherent population trapping, which
stems from a quantum interference effect. One may interpret this by saying that
the quantum interference effect gives rise to a diffusion in in space
 LASER COOLING AND LEVY PROCESSES 421

60

40

20

-20

- 40 -

- 60

0 1 2 3 4 5
it
X 107

FIG. 8.10. Stochastic simulation of sub-recoil laser cooling. The plot shows the
momentum 1)(0 for a single realization of the stochastic process defined by
the differential equation (8.119) involving 104 quantum jumps.

involving a diffusion coefficient which strongly depends on the momentum. In

contrast to the usual scheme of Doppler cooling the basic physical mechanism is
not an effective friction force, but rather a velocity selective coherent population
trapping.

8.3.3 Waiting times and momentum distributions

To get further insights into the statistics of the random waiting times we fix
a region Ipl < pt of momenta smaller than some trapping momentum p t . It is
assumed here that Pt is chosen so small that the escape rate rnc(P) from the
states 14)„,(p)) behaves to a good approximation as p2 for Ipl < Pt (see eqn
(8.137)), and that Pt < k. Suppose that an atom lands in this region through
a quantum jump starting outside the trapping region with momentum II > Pt.
Since the size of the jumps is of the order k it may be assumed further that
the atoms land somewhere in the trapping region with a uniform momentum
distribution.
As discussed in the preceding subsection, with a high probability the atomic
state after the jump is driven nearly immediately into the state (1.„,(p), from
which it may escape with a rate Fric (p). Under the condition that a specific
momentum p is given the density of the waiting time distribution for the next
quantum jump is approximately
422 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

1.5 ■

0.5 =

—1.5

- ■

_L ,

2 4 6 8 10 12 14 16
yt
X 105

FIG. 8.11. Enlarged section of Fig. 8.10 showing a succession of a large number
of jumps with small waiting times. These jumps finally drive the atom into
the vicinity of zero momentum where it remains trapped for a long time.

f (p, T) = F(p) exp (—F„, (p)y) . (8.138)

If we combine this with the above assumption of a uniform momentum distribu-
tion we get the following expression for the unconditioned probability density of
the waiting times,

+pt rnc(P)T
\ 1 1
s( r)) =2.Pt
—f f (p, r)dp = 2rnc(pt)1/27-3/2 f due. (8.139)
— Pt 0

For rne(Pt)T » 1 the integral converges to r(3/2) = Vi/2 which leads to the
asymptotic expression

fi —1 r) 3 / 2
s(r P...,) —
) — , (8.140)
47 7

where we have defined the characteristic trapping time Tt associated with pt ,

Tt = {r nc(pt)] -1 • ( 8.141)

By virtue of eqn (8.137) this yields

 LASER COOLING AND LEVY PROCESSES 423

X 10 3

1.5

0.5

1.0514 1.0516 1.0518 1.052 1.0522 1.0524 1.0526 1.0528

yt 6
X10

FIG. 8.12. Enlarged section of Fig. 8.11 showing that the atom lands in the
trapping region with a very small momentum of less than 0.5 • 10 -3 k.

'rrt (8.142)
( 4E
9 R) 2 ( Pt)
k2

The distribution (8.140) for the waiting time shows a slowly decaying power
law behaviour s(r) T -3 / 2 , which is characteristic of a stable Lévy distribution
with scaling exponent a = By an appropriate choice for the scaling parameter
we find that the Lévy distribution

1 (Tt)3/2 2 )
Tt
SL(T) = eXp (-- — (8.143)
Tt T 16 7

has the same asymptotic behaviour as the waiting time distribution s(r). The
distribution (8.143) has already been encountered in eqn (1.264) as an example
of a stable Lévy distribution whose explicit analytical expression is known. We
have thus found that the quantum interference effect which leads to a vanishing
width of the non-coupling state „,c (p = 0)) gives rise to a long-range waiting
time distribution (Bardou et al., 1994).
The power law decay of the waiting time distribution is characteristic of the
asymptotic behaviour of a random variable which follows asymptotically a stable
Lévy distribution. Figure 8.13 shows the result of a numerical simulation of the
waiting time distribution and compares it with the distribution (8.143). One
clearly observes the slow decay of the distribution. In the simulations we have
424 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

x
5

4.5

3.5

w2.5

1.5

0.5

2000 4000 6000 8000 10000

7

FIG. 8.13. Distribution s(r) of the waiting times T for initial states with mo-
mentum inside the trapping region II < pt . The figure shows a histogram of a
simulation with 106 realizations with Pt = 0.1k corresponding to a trapping
time of about -yrt = 771. For comparison the figure also shows the stable
Lévy distribution with scale parameter a = (solid line). The last bin of the
histogram contains all waiting times larger than those shown.

used pt = 0.1k corresponding to a trapping time of about -yrt = 771 by virtue

of eqn (8.143). As expected the simulation data are in good agreement with the
distribution (8.143) for times much larger than this trapping time Tt.
As explained in Section 1.6.3 the distribution (8.143) is infinitely divisible
and stable with a scaling exponent of a = It follows that the random variable
following SL (r) has the following property. We take N independent copies
T2 , TN of T and define the scaled sum

1 1
7-
N 1 /a E Ti
i=1
N2 E Ti.
i=1
(8.144)

The new random variable f- is then again distributed according to the Lévy
distribution (8.143). Note that the scaling factor is 1/N 2 and not 1/N as it
would be in those cases where the central limit theorem could be applied. The
quantity E ri /N scales as Nt which shows that the ordinary mean value of
a large number N of copies of T is with a high probability larger than any given
term of the sum. This demonstrates the drastic departure of the waiting time
 LASER COOLING AND LEVY PROCESSES 425

FIG. 8.14. Illustration of the renormalization transformation (8.145) of the ran-

dom waiting time in laser cooling for N = 5. The figure shows double log-
arithmic plots for a simulation with a total number of 106 realization with
Pt = 0.1k (symbols) and the corresponding stable Lévy distribution (solid
lines). Top: original data, middle: first application of .F5, bottom: second
application of .F5.

statistics from the usual behaviour described by the central limit theorem which
is not applicable here since the moments of T diverge.
The above scaling relations are exactly valid only for the Lévy distribution
8L (r) and not for the true waiting time distribution s(r). However, we can invoke
the stability property of the Lévy distribution to derive universal scaling rela-
tions. To this end, we consider an infinite sequence r j = 1, 2, 3, ... , of copies of
,

a random variable following the distribution s(r) and take a fixed N. Equation
(8.144) can then be viewed as defining a transformation

Ira (8.145)
to a new sequence , j = 1, 2, , of random numbers given explicitly by
Nj
1
= Tk. (8.146)
Nl/a
k=l+N(j-1)

The map -FN can be regarded as a renormalization transformation: The sequence

ri is grouped into sets of N elements each, the elements of each set are summed
426 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

1,8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0
-2 -1.5 -0.5 o 0.5 1.5
p/k

FIG. 8.15. The distribution w(p) of the atomic momentum at time Pyt = 104
obtained from a simulation with 104 realizations. The distribution shows two
sharp peaks at p = 1k with a width Sp which is much smaller than the
recoil momentum k. The initial momenta have been drawn from a Gaussian
distribution with a standard deviation of 3k (solid line).

and the sum is scaled by a factor of N -1 /a. This yields a new, coarse-grained
sequence of random variables 'T-j .
It can be shown that under successive application of the renormalization
transformation .FN the distributions of the coarse-grained sequences converge to
a corresponding a-stable Lévy distribution, which is (8.143) in the present case.
The renormalization transformation thus drives the waiting time distribution
s ( r) into the i-stable Lévy distribution (8.143).
The mathematical reason for this behaviour is that the attractive fixed points
of the renormalization transformation .FN, considered as a map in the space of
probability distributions, are given by the a-stable distributions. In this context
the central limit theorem can be understood as expressing the existence of a fixed
point characterized by a = 2. The waiting time distribution s(r) discovered
above belongs to the basin of attraction of the fixed point characterized by
a = 1/2. This is the origin of a universal behaviour in the statistics of the
waiting times. The renormalization transformation is illustrated in Fig. 8.14 for
the case N = 5.
Let us finally look at the resulting distributions of the atomic momenta (As-
pect et al., 1989). Such a distribution may be defined by
 LASER COOLING AND LEVY PROCESSES 427

3.5

2,5

1.5

0.5

0 •
-2 -1.5 -0,5 0 0.5 1 1.5 2
p/k

FIG. 8.16. The distribution v(p) (eqn 8.150) at time ryt = 10 4 obtained from
a simulation with 104 realizations. The distribution shows a single peak at
p = 0 with a width much smaller than the recoil momentum k. The initial
momenta have been drawn from a Gaussian distribution with a standard
deviation of 3k (solid line).

w(P) = (Pltrint (P)1P), (8.147)

where p = Ell 0) (0 is the atomic density matrix and trint denotes the trace over
the internal degrees of freedom of the atom. This normalized density describes the
distribution of p as it will be found if a momentum measurement is performed
on a sample of atoms. Since in the trapping region an appreciable amount of
the total number of atoms will be in states 117bnc(P')) with p' small and since
(Pi One (P')) consists of two 6-peaks at p = p' k we expect that the distribution
w(p) contains two peaks at positions p = +k over a broad background.
For a rough estimate of the width 6p of these peaks we fix some interac-
tion time t and ask for the range 'Pi < 6p of momenta such that the atoms
remain trapped with appreciable probability during time t. Clearly, this range of
momenta is found from the relation F„c (6p)t 1, which yields on using (8.137)

6p Çl 1
(8.148)
k 4ER .-15(i •

This estimate shows that (5p scales as SI/Vi. It implies that there is, at least
in principle, no lower limit for the width of the momentum distribution; for
428 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

sufficiently large interaction times the width 6p becomes much smaller than the
recoil momentum k. Consequently, the associated cooling temperature T, defined
by

1 61)2
— kBT, = (8.149)
2 2M
is much smaller than the recoil temperature TB. This is illustrated in Fig. 8.15
which shows the distribution 11(p) of the atomic momenta at time -yt = 10 4
obtained from a stochastic simulation with 104 realizations.
Another momentum distribution may be defined as

v(P ) = (On c(P)Ip lOrt e(P)), (8.150)

which yields the fraction of atoms that will be found on measurement in the
state 10(p)). Here we expect a single peak at p = 0 which is clearly seen in
Fig. 8.16.
Concluding, it must be emphasized that it makes no sense to define the
cooling temperature in terms of the mean kinetic energy of the atoms. The
reason for choosing definition (8.149) for 11, is that the trapped atoms are with a
high probability in the states 10(p)) which represent coherent superpositions of
momentum eigenstates. The definition of Te thus takes into account the quantum
correlations between the internal and the translational degrees of freedom of the
atoms.

8.4 Strong field interaction and the Floquet picture

The situation encountered in the preceding sections was that of a bound quantum
system which is coupled to an environment consisting of a continuum of electro-
magnetic field modes, and to a coherent external driving field. The procedure to
treat this situation was simply to add the Hamiltonian describing the external
driving to the coherent part of the equation of motion. However, following the
derivation of the quantum optical master equation, one observes that this proce-
dure is only justified if the external driving field represents a small perturbation
and may be treated on an equal footing with the coupling between the reduced
system and its environment. Adding the Hamiltonian of the external field to the
coherent part of the dynamics without also changing the dissipative part of the
dynamics, amounts to making the rotating wave approximation for the interac-
tion between system and driving field. It is obvious that these conditions are
violated if the external field is strong.
An appropriate strategy to deal with strong periodic driving fields is to treat
the interaction between the open system and external driving field exactly by
employing the Floquet basis of the open system, rather than the stationary eigen-
states of the unperturbed system Hamiltonian. This idea leads to a Markovian
quantum master equation for the reduced density matrix which is in Lindblad
form when written in the Floquet basis. Such an approach has been applied to
 STRONG FIELD INTERACTION AND THE FLOQUET PICTURE 429

the description of Rydberg atoms in strong microwave fields under the influence
of external noise (Bliimel et al., 1991), and to a parametrically driven harmonic
oscillator (Kohler, Dittrich and Hânggi, 1997).
In the present section we shall connect the idea of treating dissipative quan-
tum systems in strong driving fields in the Floquet picture with the stochastic
representation of the continuous monitoring of the environment. We derive the
PDP describing the spectral detection in the presence of strong external driving.
As will be seen the jumps of the PDP occur between manifolds spanned by Flo-
quet states of the system, reflecting the back-action resulting from the emitted
radiation. The observed frequencies of the radiation spectrum turn out to be the
differences between Floquet eigenvalues. We will also discuss a simple example
which serves to demonstrate the connection of the Floquet theory to the dressed
atom picture of resonance flourescence.

8.4.1 Floquet theory

The free evolution of the reduced quantum system is generated by some time-
independent Hamiltonian M. The coupling to the driving field is represented
by a Hamiltonian HL(t) which is assumed to be periodic in time, that is

HL (t + TL) = HL(t), (8.151)

where TL 271-1c.oL and wi, denotes the driving frequency. Thus, the Hamiltonian
that describes the reduced system without its interaction with the environment
is given by the TL-periodic operator

Hs(t) = H,., + HL(t). (8.152)

The strategy is now to treat exactly that part of the dynamics which is
described by the system Hamiltonian H(t). By this procedure one avoids the
rotating wave approximation for the interaction between system and external
driving, as well as any perturbative treatment for this part of the dynamics.
To this end, we invoke the Floquet theory of quantum mechanical systems with
a time-periodic Hamiltonian (Shirley, 1965; Zeldovich, 1967). The Schrödinger
equation

d i
hHs(t)0(t) (8.153)
Tft11)(t) --
is a differential equation with time-periodic coefficients. According to Floquet's
theorem there exists a complete set of solutions 0,40 of eqn (8.153) which are
labelled by an index r and which can be written in the form

Or (t) = Ur (t) exp {—iE rt/h} . (8.154)

The form of these solutions is quite similar to that of the stationary states of a
system with a time-independent Hamiltonian. The quantities 6 r appearing in the
430 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

phase factors exp(—k rt/h) are time independent and are called quasi-energies
or Floquet indices. The state vectors ur(t) depend periodically on time,

u r (t + TL ) = ur (t), (8.155)

and are referred to as Floquet states. The completeness of the Floquet solutions
is expressed by the relation

E iur(oxurn = /-, (8.156)

which shows that, for each fixed time t, the Ur(t) form a complete set of basis
states. Moreover these states can be chosen to be orthogonal,

(U r (t) 'Ur' (t)) = (8.157)

Thus, the Floquet states form a time-dependent basis in the Hilbert space of
the system under consideration. It follows that any solution 'OM of the time-
dependent Schrödinger equation (8.153) can be decomposed as

0(t) , E ar ur (t)exp {—iE rt/h} . (8.158)

The important point to note is that according to Floquet's theorem the ampli-
tudes ar are time independent.
Inserting eqn (8.154) into the time-dependent Schrödinger equation we ob-
serve that the Floquet wave functions U r (t) may be obtained from the eigenvalue
equation

Hur (t) E {Hs(t) — i ha} 74(0 = erur (t) (8.159)

by imposing periodic boundary conditions of the form (8.155). The operator H

in eqn (8.159) is the Floquet operator which has to be considered as an operator
on the extended Hilbert space

HF - 'HS 0 '1"-1 71 , ( 8.160)

where 7 la is the Hilbert space of the open system, and WT I denotes the space
- ,

of square-integrable TL -periodic functions. The scalar product in the extended

Hilbert space is given by

(NV)) E TL C
f - ( 0)11)(0), (8.161)
T1,
0

such that the expression — ihOt in the eigenvalue eqn (8.159) becomes a Hermitian
operator on IF.
 STRONG FIELD INTERACTION AND THE FLOQUET PICTURE 431

It is important to remark the following fact. Given some solution 74(0 of the
Floquet eigenvalue equation (8.159) with quasi-energy E r , then for any integer n
also

tt r, n (t) E Ur (t) exp{iwLnt}

is a solution with quasi-energy

Er,n, E Er + hwLn.

On the other hand, for all integers n the states (8.162) lead to one and the same
solution of the time-dependent Schrödinger equation since

ur,„(t)exp{ —iE r,„t/h} = ur (t)exp{—iE rt/h} . (8.164)

The class of states (8.162) belonging to a fixed r and to different integers n are
therefore physically equivalent.
The time-evolution operator Us(t, t') corresponding to the Schrödinger equa-
tion (8.153) obeys

o us(t,e) , – iH-Au s (t,e), us (ti ,e) = I.

— (8.165)
at h
By virtue of the Floquet representation (8.158) it takes the form

us(t, ti) = E iu,(0)(ur(tt)lexP {–iEr(t — t i )/h} 1 (8.166)

r

as is easily checked with the help of eqn (8.159) and the completeness relation
(8.156).

8.4.2 Stochastic dynamics in the Floquet picture

In order to derive a stochastic process for the dynamics of the reduced system
coupled to the continuum of radiation modes we consider the dipole operator

_6 (t) = u,f, (t, o)us (t, o). (8.167)

This is the exact Heisenberg picture dipole operator for the system plus external
driving field. We are seeking a decomposition of /5(t) into frequency components
A(w) such that we can write

15 (t) = E exp{—iwt}A(w) + h.c., (8.168)

w
where the sum is extended over a certain set of positive frequencies w. Employing
the Floquet decomposition (8.166) of the time evolution operator Us(t, 0) one
finds
432 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

id(t) E lur(0))(ur(t)i13iuri(t))(ur,(0)1 expl-i(Er, - Er)00. (8.169)

r,r/

Here, the dipole matrix element depends periodically on time and we may use
the Fourier decomposition to get

(ur(t)1Blur'(t)) = E((ur,71113lue)) exp{iwLnt}, (8.170)

where we have used the states (8.162) and the scalar product (8.161) in the
extended Hilbert space. Thus, we can write the dipole operator as

b(t) = E iur(0))((ur,noluri))(uri(0)1expl-i(er, — Er —
r,r',n

which finally leads to

4 (w) E lur(0))((ur,nolurmur , (0)1.

r,r 1 ,n

The sum in this expression runs over all sets (r, r', n) of quantum numbers that
satisfy the condition

Er/ — Er — hWL11 = hr..0 > 0. (8.173)

For a given frequency w there are, in general, several sets (r, r', n) that fulfil this
constraint. The frequency component 4.(w) of the dipole operator is different
from zero if at least one of the corresponding matrix elements is different from
zero, that is, if

(04,7101 24 i )) E
f
TL
dt
TL
n
"/L t (tir(t)IblUr F (t)) O. (8.174)
o
This means that the n-th Fourier component of the oscillating matrix element
of the dipole operator in the Floquet basis must be non-zero. Let us denote by
grad the set of positive frequencies w given by eqn (8.173) and the additional
selection rule (8.174). The set (A rad determines the positions of the peaks in the
radiation spectrum emitted by the system.
At this stage we can now proceed in a similar manner as in Section 8.1,
whereby the frequency decomposition (8.51) of the interaction picture dipole
operator is to be replaced by the decomposition (8.168) of the Heisenberg picture
dipole operator. In order to be able to apply the rotating wave approximation
with respect to the system—environment coupling the condition

lw w'l » 7(w)1((ur,n1 13 lur , ))1 2 (8.175)

must be satisfied for the different frequencies of the radiation spectrum. The
result is then that the spectral detection leads to a PDP for the state vector of
 STRONG FIELD INTERACTION AND THE FLOQUET PICTURE 433

the source which is given by (Breuer and Petruccione, 1997; Breuer, Huber and
Petruccione, 2000)

d(t) = — (0(t))dt Y' [ e'54 • -,zr (w)0(t) dN w ,m(t).

h, 21(w)0(011
ivitd,± •

(8.176)

The expressions for the generator G(0) and for the expectation values of the
Poisson increments dAT,,m(t) are formally identical to the ones derived in Section
8.1 (see eqns (8.53) and (8.54)). However, the new feature is that the jump
operators Â(c.,.)) are defined in terms of the Floquet representation, involving the
exact time evolution due to the periodic driving force. According to eqn (8.172)
the 1.(w) are lowering operators pertaining to the positive transition frequencies
E grad of the source and describe quantum jumps between manifolds spanned
by Floquet states.
Summarizing, we find that the positions of the peaks of the fluorescence
spectrum are determined by the differences of the quasi-energies modulo inte-
ger multiples of ricoL. A given frequency co appears in the radiation spectrum
if the matrix elements ((u r Iblur, )) do not vanish, since otherwise the corre-
sponding jump operator would be zero, according to the selection rule (8.174).
Besides the usual selection rules for these matrix elements caused by symmetry,
those selection rules that arise from the Fourier content of the Floquet states
are important. For example, the condition for the generation of high harmonic
radiation (L'Huillier, Schafer and Kulander, 1991), i.e. large n in eqn (8.173),
is that high Fourier modes of the Floquet states are considerably excited. This
condition is fulfilled in particular in the vicinity of near degeneracies or avoided
crossings in the quasi-energy spectrum plotted as a function of the driving field
amplitude (Breuer, Dietz and Holthaus, 1988).
The equation of motion for the reduced density matrix of the source is easily
obtained by determining the equation of motion for the covariance of the PDP
derived above. Transforming back to the Schrödinger picture we get the master
equation

d
Ps(t) = [HO), ps(t)] + Dt(Ps(t)), (8.177)

where

vt (ps) E (w) OP, Opslf ( A), t) (8.178)

1
_ (w, t) • Z(A.r, t)ps — — psift ( A), t) t))
2 2

is the dissipator involving the time-dependent operators

434 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

A.(w,t)= E iu,comur,nifilur,murqoi. (8.179)

r,r' ,n

Again, the sum runs over those sets (r, r', n) that correspond to the given fre-
quency w of the radiation spectrum. For each fixed t the dissipator Vt is in
Lindblad form. However, by contrast to the case of the quantum optical master
equation, it depends explicitly on time. The physical reason for this fact is easily
understood. The external driving leads to a strong distortion of the dipole mo-
ment. Since the system couples to the environment via its dipole moment, the
driving field also strongly influences the dissipation mechanism.

8.4.3 Spectral detection and the dressed atom

Let us consider a two-level atom which is strongly driven by a resonant driving
field. Assuming the rotating wave approximation to be valid, we may write the
system Hamiltonian as

Hs(t) = wo o-± 0- — ( 0-±e–iwLt a_ e iwLt) (8.180)

2
where wo denotes the level spacing, Ç1 is the Rabi frequency and we have set
h= 1. At resonance, that is for wL = wo , a basis of Floquet states is given by
1 ( + e –zw.o )
U±(t) = (8.181)
1

with corresponding quasi-energies

1
6± = – Q. (8.182)
2
Note that the Floquet states have been chosen such that their corresponding
quasi-energy difference is equal to the Rabi frequency O. The Schrödinger picture
dipole operator is given in the present notation by

D= -F a . (8.183)
Thus, in the weak driving case we have just one jump operator which is given
by the lowering operator a_. The resulting PDP defined by means of this jump
operator is precisely the process that has been used to simulate the quantum
optical Bloch equation in Section 6.3.
To determine the jump operators in the strong driving case we employ the
Floquet representation (8.166) of the time-evolution operator to obtain the Hei-
senberg picture dipole operator,

D(t) = *t,O)DUs (t, 0) = I u±( 0))(u±(01-13 1u+(t))(u40)1 (8.184)

-Hu_(0))(u_(t)IDIu_ (t)) (u (0)1
+itt+ (0)) (u + (t)IDIu _ (0) (u_ (0)le )t
+lu_(0))(u_MIDItt±(t))(u±(0)le(E--E+)t.
 STRONG FIELD INTERACTION AND THE FLOQUET PICTURE 435

Since

(u±(t)1131u±(0) = + -1 (e iwLt Lt) , (8.185)

(it + (01-Diu —(t)) = (eiu)Lt e—t), (8.186)

we get

1
D(t) = —2 (e twi• t + et) + (0 )) + — — ( 0)) —
1
±- (e i( u) L+n) t —
2
1 (ei(wL _s-2)t e—i(wL+s-2)t)iu_(0))(u+(0)1.
(8.187)
2
From this relation we immediately infer that we have three jump operators be-
longing to the positive frequencies wL and coL Q, namely
1
APL) = au ± (0))(u ± (0)1— 1u_(0))(u_ (0)1) , (8.188)
1
A(wL + = lu—(0))(u-F(0)i, (8.189)

.i4(c.4.)L — = --21 1u± (0))(u—(0)1. (8.190)

Thus, instead of one jump operator a _ we get three jump operators in the strong
driving case. The condition for strong driving is provided by (8.175) which leads
here to the condition 1 » y.
To write the corresponding PDP defined by eqn (8.176) more explicitly we
decompose the state vector in the Floquet basis as follows,

0(0 0+(t)lu-F(0)) + 0-(01u-(0)), (8.191)

and represent it through a two-component vector,

0(t) —= ( °±(t) ) (8.192)

0_ (t) •

The jump operators given in eqn (8.188) are represented by Pauli matrices,

1
APL ). APL (8.193)
2 °' 2
Equation (8.176) therefore leads to the stochastic differential equation

d0(t) = [
a3 0(t)
110_30(011
( t) dN3(t) E o(t)ag
q=± I laq0(011
0(t)1 dNg (t). (8.194)
436 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

We observe that the generator of the deterministic evolution periods of the PDP
vanishes, such that the process becomes a pure jump process. The statistics of
these jumps is determined by the Poisson increments which satisfy,
'Y 'Y
E [dN3 (t)] = –dt, E [d/V±(t)j = I(/)
–(t)1 2 dt. (8.195)
4 4
It follows from the relative weight of the jump rates that, provided the system is
known to jump at some time t, the jumps with o-± occur with the probabilities
I OT (t)12, while the jump with o-3 takes place with probability 12 .
The physical meaning of the PDP can be elucidated by drawing on the cor-
respondence between the Floquet picture and a full quantum treatment of the
driving mode (Shirley, 1965). This correspondence leads directly to an interpre-
tation of the process in terms of the dressed atom picture (Cohen-Tannoudji and
Reynaud, 1977, 1979). In accordance with the notation used in eqn (8.162) we
define
= u±(t)einwL t . (8.196)
These Floquet states have quasi-energies
1
E±,n = +- O + nwL. (8.197)
2
For strong driving fields the Floquet states correspond to the dressed states of the
atom, that is to the stationary eigenstates of the combined system consisting of
atom and quantized driving mode. For large photon numbers N the eigenenergies
E±, N of the dressed states 1+, N) of the atom are approximately given by
1
E±,N P.J., +- 11 + NwL • (8.198)
2
Comparing this expression with the quasi-energies (8.197) of the Floquet states
(8.196) we see that we have the following correspondence
u(t) .4-----? l +)N + n) (8.199)
between the Floquet states and the dressed states. On using this correspondence
we find that the transition with jump operator A(wL + SI) corresponds to the
transition

I+, N + 1 ) —> (8.200)

whereas the jump with the operator A(wL – S2) corresponds to the transition
1–, N + 1) I+ , N) (8.201)
between the dressed atomic states. Likewise, the transition induced by the jump
operator A(wL) can be interpreted as the transition

al+, N + 1) + 131– , N + 1) —> al+,N) – M – , N). (8.202)

Thus we conclude that the Floquet representation yields a simple jump process
describing quantum jumps between dressed atomic states. It is also obvious that
 REFERENCES 437

the three types of quantum jump correspond to the three peaks of the strong-
driving Mollow spectrum (3.306): We have a central peak at frequency coL and
two sideband peaks at frequencies coL + O. Since in the stationary case both
Floquet states are found with probability it follows from (8.195) that the
integrated intensities of the two sideband peaks coincide and are equal to
times the integrated intensity of the central peak.

References
Akhiezer, A. I. and Berestetskii, V. B. (1965). Quantum Electrodynamics. In-
terscience Publishers, John Wiley, New York.
Aspect, A., Arimondo, E., Kaiser, R., Vansteenkiste, N. and Cohen-Tannoudji,
C. (1989). Laser cooling below the one-photon recoil energy by velocity-
selective coherent population trapping: Theoretical analysis. J. Opt. Soc.
Am., B6, 2112-2124.
Bardou, F., Bouchaud, J.-P., Emile, O., Aspect, A. and Cohen-Tannoudji, C.
(1994). Subrecoil laser cooling and Lévy flights. Phys. Rev. Lett., 72, 203-206.
Bardou, F., Bouchaud, J.-P., Aspect, A. and Cohen-Tannoudji, C. (2001). Lévy
Statistics and Laser Cooling. Cambridge University Press, Cambridge.
Bjorken, J. D. and Drell, S. D. (1965). Relativistic Quantum Fields. McGraw-
Hill, New York.
Bliimel, R., Buchleitner, A., Graham, R., Sirko, L., Smilansky, U. and Walther,
H. (1991). Dynamical localization in the microwave interaction of Rydberg
atoms: The influence of noise. Phys. Rev., A44, 4521-4540.
Breuer, H. P., Dietz, K. and Holthaus, M. (1988). Strong laser fields interacting
with matter I. Z. Phys., D10, 13-26.
Breuer, H. P., Huber, W. and Petruccione, F. (2000). Quasi-stationary distri-
butions of dissipative nonlinear quantum oscillators in strong periodic driving
fields. Phys. Rev., E61, 4883-4889.
Breuer, H. P. and Petruccione, F. (1997). Dissipative quantum systems in
strong laser fields: Stochastic wave-function method and Floquet theory. Phys.
Rev., A55, 3101-3116.
Castin, Y. and Molmer, K. (1995). Monte Carlo wave-function analysis of 3d
optical molasses. Phys. Rev. Lett., 74, 3772-3775.
Cohen-Tannoudji, C., Bardou, F. and Aspect, A. (1992). Review on fundamen-
tal processes in laser cooling. In Proceedings of Laser Spectroscopy X (eds.
Ducloy, M., Giacobino, E. and Camy, G.), pp. 3-14, World Scientific, Singa-
pore.
Cohen-Tannoudji, C. and Reynaud, S. (1977). Dressed-atom description of res-
onance fluorescence and absorption spectra of a multi-level atom in an intense
laser beam. J. Phys., B10, 345-363.
Cohen-Tannoudji, C. and Reynaud, S. (1979). Atoms in strong light-fields:
Photon antibunching in single atom fluorescence. Phil. Trans. R. Soc.
Lond., A293, 223-237.
438 APPLICATIONS TO QUANTUM OPTICAL SYSTEMS

Cohen-Tannoudji, C. (1992a). Laser cooling and trapping of neutral atoms:

Theory. Phys. Rep., 219, 153-164.
Cohen-Tannoudji, C. (1992b). Atomic Motion in Laser Light, Les Houches
Session LIII on Fundamental Systems in Quantum Optics, pp. 1 164. North-
-

Holland, Elsevier, Amsterdam.

Cohen-Tannoudji, C., Dupont-Roc, J. and Grynberg, G. (1998). Atom Photon
—

Interactions. John Wiley, New York.

Kasevich, M. and Chu, S. (1992). Laser cooling below a photon recoil with
three-level atoms. Phys. Rev. Lett., 69, 1741-1744.
Kohler, S., Dittrich, T. and HAnggi, P. (1997). Floquet—Markovian description
of the parametrically driven, dissipative harmonic oscillator. Phys. Rev., E55,
300-313.
Lawall, J., Bardou, F., Saubamea, B., Shimizu, K., Leduc, M., Aspect, A. and
Cohen-Tannoudji, C. (1994). Two-dimensional subrecoil laser cooling. Phys.
Rev. Lett., 73, 1915-1918.
Lawall, J., Kuhn, S., Saubamea, B., Bigelow, N., Leduc, M. and Cohen-
Tannoudji, C. (1995). Three-dimensional laser cooling of helium beyond the
single-photon recoil limit. Phys. Rev. Lett., 75, 4194-4197.
L'Huillier, A., Schafer, K. J. and Kulander, K. C. (1991). Theoretical aspects
of intense field harmonic generation. J. Phys., B24, 3315-3341.
MOlmer, K., Castin, Y. and Dalibard, J. (1993). Monte Carlo wave-function
method in quantum optics. J. Opt. Soc. Am., B10, 524-538.
Reichel, J., Bardou, F., Ben Dahan, M., Peik, E., Rand, S., Salomon, C. and
Cohen-Tannoudji, C. (1995). Raman cooling of cesium below 3 nK: New
approach inspired by Lévy flight statistics. Phys. Rev. Lett., 75, 4575-4578.
Shirley, J. H. (1965). Solution of the Schredinger equation with a Hamiltonian
periodic in time. Phys. Rev., 138B, 979-987.
Shore, B. W. (1990). The Theory of Coherent Atomic Excitation. Volume 2.
Multilevel Atoms and Incoherence. John Wiley, New York.
Zeldovich, Ya. B. (1967). The quasienergy of a quantum-mechanical system
subjected to a periodic action. Sou. Phys. JETP, 24, 1006-1008.
 Part IV
Non-Markovian quantum processes
 9
PROJECTION OPERATOR TECHNIQUES

As discussed in Chapter 3 the laws describing the dynamics of open quantum

systems can be derived from the unitary dynamics of the total system. In general,
the reduction of the degrees of freedom in the effective description of the open
system results in non-Markovian behaviour. It is the aim of this part to introduce
the reader to some powerful techniques which allow a systematic description of
the non-Markovian features of the dynamics of open systems.
A general framework to derive exact equations of motion for an open system is
provided by projection operator techniques. These techniques were introduced by
Nakajima (1958) and Zwanzig (1960) and independently by the Brussels school
(Prigogine, 1962). They are widely used in non-equilibrium statistical mechanics
(Haake, 1973; Balescu, 1975; Grabert, 1982; Kubo, Toda and Hashitsume, 1985).
The basic idea underlying the application of projection operator techniques to
open quantum systems is to regard the operation of tracing over the environment
as a formal projection pl Pp in the state space of the total system. The super-
—,

operator P has the property of a projection operator, that is P 2 = P, and the

density matrix Pp is said to be the relevant part of the density p of the total
system. Correspondingly, one defines a projection p J Qp onto the irrelevant
—

part Qp, where P + 2 is equal to the identity map. The aim is then to derive a
closed equation of motion for the relevant part Pp.
We are going to discuss in this chapter two variants of projection operator
techniques, the Nakajima—Zwanzig and the time- convolutionless technique. Both
methods lead to an exact equation of motion for the relevant part Pp. In the
case of the Nakajima—Zwanzig method this is an integro-differential equation in-
volving a retarded time integration over the history of the reduced system, while
the time-convolutionless equation of motion provides a first-order differential
equation which is local in time.
The time-convolutionless projection operator technique leads to a time-local
expansion of the equation of motion with respect to the strength of the system—
environment coupling. It thus supports an investigation of non-Markovian effects
beyond the Born approximation. To each order in the coupling the equation of
motion involves a time-dependent but local generator. The rules for the per-
turbation expansion of the convolutionless generator will be developed. We are
mainly concerned in this chapter with the derivation of the most important gen-
eral results; specific physical applications will be studied in the next chapter.
442 PROJECTION OPERATOR TECHNIQUES

9.1 The Nakajima—Zwanzig projection operator technique

We consider the general physical situation of an open system S coupled to an
environment B. The dynamics of the density matrix p(t) of the combined system
is specified by some microscopic Hamiltonian of the form

H = Ho + celli, (9.1)
where Ho generates the uncoupled time evolution of the system and environment.
Hi- describes their interaction, and a denotes a dimensionless expansion param-
eter. When working in the interaction representation, the equation of motion for
the density matrix reads
a p(t) = —ice[Hi (t), p(t)]
— (9.2)
at
where we have set h = 1 and the interaction picture representation of the inter-
action Hamiltonian is defined by

H1 (t) = exp(illo t)HI exp(—iHot). (9 .3)

The Liouville super-operator is denoted by f(t).

9.1.1 Projection operators

In order to derive an exact equation of motion for the reduced density matrix
ps of the open system it is convenient to define a super-operator P according to
p 1— P p = trB {p} 0 pB E PS 0 pB, (9.4)
where pB is some fixed state of the environment. This super-operator projects on
the relevant part of the density matrix p in the sense that Pp gives the complete
information required to reconstruct the reduced density matrix ps of the open
system. Accordingly, a complementary super-operator Q,

2P = P — P P, (9.5)
may be introduced, which projects on the irrelevant part of the density matrix.
The super-operators P and 2 are maps in the state space of the combined system.
that is in the space of density matrices of the total Hilbert space R = 7-is 0 RB•
They have the obvious properties

(9.6)
p2 _ p, (9.7)
Q2 _ Q, (9.8)
P2 = 2P = 0, (9.9)
which can be easily checked using the definitions (9.4) and (9.5) and assuming
pB to be normalized, trB pB = 1.
 THE NAKAJIMA—ZWANZIG PROJECTION OPERATOR TECHNIQUE 443

The density matrix pB used in definition (9.4) is an operator in 7-1B . It may

represent a quite arbitrary, but known environmental state, called the reference
state. The choice of pB strongly depends on the specific application one has
in mind. In the following we shall suppose this state to be time independent.
Typically, it is taken to be the stationary Gibbs state of the environment. In
many cases it may also be assumed that the odd moments of the interaction
Hamiltonian with respect to the reference state vanish

trB {1/ 1 (t1 )Hi(t2 )...H/(t 2n+1 )pB 1 = 0, (9.10)

which leads to the relation
Pr(t1 ),C(t2)... f(t2)P = 0 (9.11)
for n = 0, 1, 2, .... This technical assumption is not required for the derivation
of the equation of motion. It will however be used later on in order to simplify
the expressions of the perturbation expansion. It is important to remark that
we do not demand any particular form for the initial conditions at this point. In
particular we do not assume factorizing initial conditions.

9.1.2 The Nakajima-Zwanzig equation

Our aim is now to derive a closed equation for the relevant part Pp(t), i.e.
for the density matrix ps(t) = trB p(t) of the open system. By applying the
projection operators P and Q to the Liouville-von Neumann equation (9.2)
and by invoking the time independence of the reference state the following set
of coupled differential equations for the relevant and the irrelevant part of the
density matrix is obtained,
a a
— Pp(t) = P — p(t) = aP f(t)p(t), (9.12)
at at
a a
— Qp(t) = Q— p(t) = a2r(t)p(t). (9.13)
at at
On inserting the identity / = P + 2 between the Liouville operator and the
density matrix p this may also be written as
a
— Pp(t) = aPE(t)Pp(t) + aP,C(t)Qp(t), (9.14)
at
a
— Qp(t) = aQ,C(t)Pp(t) + a2r(t)Qp(t). (9.15)
at
To get a closed equation for the relevant part of the density matrix we solve eqn
(9.15) and insert the solution into eqn (9.14). The formal solution of eqn (9.15)
corresponding to a given p(to ) at some initial time to may be expressed as
t
Qp(t) = g(t,to )Qp(to ) + a f dsg(t,$)Qr(s)Pp(s), (9.16)
to
444 PROJECTION OPERATOR TECHNIQUES

where we have introduced the propagator

t
g(t, 8) E T , exp [ce f ds'2,C(s1 . (9.17)
s
As usual, the operator T, describes the chronological time ordering: It orders
any product of super-operators such that the time arguments increase from right
to left. The propagator G(t, s) thus satisfies the differential equation

a
— G(t 8) = aQ,C(t)G (t, s) (9.18)
at '
with the initial condition

g (8 , 3) = I. (9.19)

Inserting the expression (9.16) for the irrelevant part of the density matrix
into the equation of motion (9.14) for the relevant part we obtain the desired
exact equation for the time evolution of the relevant part of the density matrix.
a Pp(t) = al)
c(t)g(t, to)2p(to) + al) r(t)P p(t)
t
+02 f dsP f(t)G(t, .9)2,C(s)P p(s). (9.20)
to
This equation is known as the Nakajima—Zwanzig equation. It is an exact equa-
tion for the relevant degrees of freedom of the reduced system. The right-hand
side involves an inhomogeneous term Pr(t)o, to ) 2p(to) depending on the ini-
tial condition at time t o , and an integral over the past history of the system in the
time interval [to , t]. It thus describes completely non-Markovian memory effects
of the reduced dynamics. If condition (9.11) is satisfied for n = 0, the second
term in the Nakajima—Zwanzig equation vanishes and we may cast it into the
compact form
t
ata P p(t) = f ds (t, s)P p(s) + aPCMG (t, to)2 p(to). (9.21)
to

The convolution or memory kernel

/C(t,$) = ct 2 P r(t)G(t, s) 2 C(s)P (9.22)

represents a super-operator in the relevant subspace.
The integro-differential equation (9.21) is exact and holds for all initial con-
ditions and for almost arbitrary systems and interactions. Unfortunately, the
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 445

Nakajima—Zwanzig equation is usually as difficult to solve as the Liouville equa-

tion describing the dynamics of the total system. This means that perturbation
expansions are needed in order to discuss the relevant dynamics in a way acces-
sible to analytical or numerical computations. Obviously, the equation may be
expanded in the coupling constant a, i.e. in powers of the interaction Hamilto-
nian HI. Alternatively, it may be expanded around t in powers of the memory
time, i.e. in the width of the kernel 1C(t, s), where, of course, for 1C(t, s) ,c:-.', (5(t — s)
in the absence of memory effects we obtain the Markovian description. Some-
times it might also be convenient to perform the perturbation expansion for the
Laplace transform of Ps (t) in the Schrödinger picture.
For a factorizing initial condition p(to ) = ps (to) 0 pB we have P p(to) = P(to)
and, therefore, Qp(to ) = O. Hence the inhomogeneous term of the Nakajima-
Zwanzig equation (9.21) vanishes and the exact equation for the relevant part of
the density matrix reduces to
t
a p(t) f ds1C(t, s)P p(s). (9.23)
N TI =
to

To second order in the coupling strength a we obtain

/C(t,$) = a2 2.C(t)2,C(s)P + 0(a3 ), (9.24)

which leads to an equation of motion of second order for Pp(t)
t
a f dsPE(t),C(s)Pp(s),
= a2
NPp(t) (9.25)
to

where we again made use of P,C(t)P = O. If we now introduce the explicit

expressions for the projection operator P and for the generator f(t) we get the
Born approximation of the master equation
t
a _2 f ds tr B [Hr (0, [Hr (8), Ps(s) 0 PB1 (9.26)
— ps(t)=
at
to

which we already met in eqn (3.116).

This approach to the non-Markovian dynamics of open quantum systems has
some practical disadvantages. The perturbative approximation of the memory
kernel simplifies the derivation of the equations of motion, but unfortunately
not their structure. The approximate equation of motion is again an integro-
differential equation, whose numerical solution may be quite involved.

9.2 The time-convolutioniess projection operator method

In practice the time convolution in the memory kernel of the Nakajima—Zwanzig
equation is difficult to treat. In this section we show how to remove the time
446 PROJECTION OPERATOR TECHNIQUES

convolution in the master equation. This is achieved through a method which is

known as the time-convolutionless projection operator technique. This technique
has been developed by Shibata et al. (Shibata, Takahashi and Hashitsume, 1977;
Chaturvedi and Shibata, 1979; Shibata and Arimitsu, 1980) and we are going
to apply it here to the microscopic theory of an open quantum system which
is coupled to an environment. The method yields a systematic expansion of the
dynamics of the system of interest in terms of the coupling strength. In particular,
we will develop expressions for the quantum master equation up to fourth order
in the coupling for factorizing and for non-factorizing initial conditions.

9.2.1 The time local master equation

-

The idea of the time-convolutionless projection operator technique is to eliminate

the dependence of the future time evolution on the history of the system from the
Nakajima—Zwanzig master equation and thus to derive an exact master equation
for the open system which is local in time. In order to achieve this objective we
proceed in the following way: The density matrix p(s) on the right-hand side of
eqn (9.16) is replaced by the expression
p(s) = G(t, s)(P + 2) p(t), (9.27)
where G(t, s) is the backward propagator of the composite system, i.e. the inverse
of the unitary time evolution of the total system. Formally, we may write
t
G (t, s) = T , exp [ a f ds'
— (9.28)
s
where T„ indicates the antichronological time-ordering.
With the help of the relation (9.27) the equation (9.16) for the irrelevant part
of the density matrix may now be written as
t

QP(t) = g(t, to)2P(to) + a f ds G(t, 8)2 r(s)P G (t, s)(P + 2)p(t). (9.29)
to
Introducing the super-operator
t
E(t) = a f ds G(t , s)2,C(s)PG(t, s), (9.30)
to
we can express the irrelevant part of the density matrix through
[1 — E (0] Qp(t) = G(t, to ) Qp(to ) + E(t)Pp(t). (9.31)
Note that the super-operator E(t) contains both propagators g and G, so that
it does not specify a well-defined chronological order. E(t) has the obvious prop-
erties E(t0 ) = 0 and E(t)1„ 0 = 0. Hence, 1 — E(t) may be inverted for not too
large couplings and in any case for small t to . Thus, we get
—
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 447

2p(t) = [1- E(t)] -1 E(t)Pp(t) 4- [1- E(t)F 1 g(t,to)QP(to). (9.32)

This equation states that the irrelevant part Qp(t) of the density matrix can in
principle be determined from the knowledge of the relevant part Pp(t) at time
t and from the initial condition Qp(to ). The dependence on the history of the
relevant part which occurs in the Nakajima-Zwanzig equation has thus been
removed by the introduction of the exact backward propagator G (t, s). It must
be noted, however, that for strong couplings and/or large time intervals t - to it
may happen that eqn (9.31) cannot be solved uniquely for Qp(t) such that the
inverse of 1 - E(t) does not exist. We are going to exemplify this situation in
Section 10.1.2.
To complete the derivation of the time-convolutionless master equation, we
insert eqn (9.32) into the equation of motion for the relevant part (9.14) and
obtain the following exact time-convolutionless (TCL) form of the master equa-
tion,

8
— Pp(t) =1C(t)Pp(t) +I(t)2P(to), (9.33)
at
with the time-local generator, called the TCL generator,

/C(t) = a'Pr(t) [1 - E(01 -1 7) , (9.34)

and the inhomogeneity

1(t) = (ARCM [1 - E (O]' g(t,t0 )Q. (9.35)

The equation of motion (9.33) is exact and local in time. Although the super-
operators /C(t) and 1(t) are, in general, extremely complicated objects, eqn (9.33)
can be used as a starting point of a systematic approximation method by ex-
panding 1C(t) and 1(t) in powers of the coupling strength a. This will be shown
in the following subsections.

9.2.2 Perturbation expansion of the TCL generator

Of course, the super-operator k(t) only exists when it is possible to invert the op-
erator [1- E(t)]. Let us assume then that E(t) may be expanded into a geometric
series
00

[1 - E(t) ] -1 = E [(t)]. (9.36)

On substituting this into the expression (9.34) one gets

DO 00

/C(t) = a E RCM [E(t)] P = E an/Cn (t). (9.37)

rt-------0 n-_-_-1
448 PROJECTION OPERATOR TECHNIQUES

To determine the contribution kri (t) of n-th order in a to the TCL generator
1C(t) we also expand E(t) in powers of a,

E anEn (t),
DO

E(t) = (9.38)

insert this into eqn (9.37), and sort equal powers of a. For example, to fourth
order in a this gives:
K: 1 (t) = (9.39)
1(2(0 = (9.40)
1(3(0 = PL(t) { [El MP + E2(01 7) , (9.41)
/C4 (t) = RCM {[Ei (t) ]3 + Et (0E2 (t) + E2 (t) E1 (t) + 3 (t)} P. (9.42)

Finally, the contributions En (t) are found with the help of eqns (9.30) and (9.38)
by expanding also the propagators 0,8) and G(t, s) defined in eqns (9.17) and
(9.28) in powers of a.
Let us determine more explicitly the first four terms of the expansion. To
simplify the expressions we use condition (9.11) and take to = 0. Equation (9.39)
immediately gives
1( 1 (0 = RCMP = 0. (9.43)

The first-order term E i (t) is given by

t
= f dt i Qr(ti )P, (9.44)
o
which yields
t
/(2 (t) = f dtiP,C(t),C(ti YP. (9.45)
o
The second-order term E2 (t) is found to be
i ti
E2 (t) = f dt i f dt2 [Q,C(4)2,C(t2)P — 2E(t2)Pr(t1)] . (9.46)
o o
Since P2 = 0 we conclude from eqn (9.44) that [Ei (O] 2 = 0 and, therefore,
t ti
1C3 (t) = PC(t)E2(OP = f dti f dt2P,C(t)C(t i )C(t2 )P = 0, (9.47)
o o
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 449

where we made use of condition (9.11) for n = 0 and n = 1. To find /C4 (t) we
first note that [El(t) ] 3 = Ei (t)E2(t) = 0 because of P2 = 0. Thus we have from
eqn (9.42)

1C4 (t) = P r(t) [ 2 (t) 1 (t) + E3 (t)] P. (9.48)

Invoking eqns (9.44) and (9.46) the first term is found to be

t ti t2
P r(t)E2(t)El (t)P = — f dt i f dt2 f dt 3 (9.49)
o o o
xPr(t)[r(t2)PE(ti),C(t3)P + 03yp,02),c(t1)P + r(t3)P r(ti) 02)P1 •

Note that to get this expression the triple time integral has been brought into
time-ordered form, t > t 1 > t2 > t3 > O. Similarly, one finds

Pr(t)E3(t)P (9.50)
t t2
= f dt i f dt 2 f dt3P.C(t)[f(t1)2f(t2),C(t3)P+ r(t3)7) ,C(t2),C(ti)Pl.
o o o

Summarizing, the fourth-order contribution to the TCL generator takes the form,

t2
(t) = f dtl f dt2 f dt3 (Pf(t)r(t1 )r(t2),C(t3)P Pr(t),C(ti)P,C(t2),C(t3)P
0 0 0

—Pr(t),C(t2)P,C(t 1 ),C(t3)P —Pf(t),C(t 3 )P,C(ti ),C(t2)P).

(9.51)

The second-order generator /C2 (t) of the TCL master equation leads to the
following equation for the reduced density matrix ps(t),

a (9.52)
— ps(t)= —a 2 f ds tr B [I / (t), [1-1/(8),Ps(t) pg,
at
o

which should be contrasted to the corresponding second-order approximation

(9.26) of the Nakajima—Zwanzig equation: Both equations are of second order
and it is therefore to be expected that they approximate the exact dynamics
with the same accuracy. This point will be illustrated in Section 10.1.2 with the
help of a simple example. In practice, the TCL form is to be preferred because
it involves a time-local generator instead of a convolution kernel.
450 PROJECTION OPERATOR TECHNIQUES

The expressions (9.45) and(9.51) can be made more explicit if one writes the
interaction Hamiltonian as a sum of products of Hermitian operators Fk and Qk
which act in the system's and the environment's Hilbert space, respectively,

1/1- = Fk Qk. (9.53)

Such a decomposition of the interaction into Hermitian operators is always possi-

ble. It is usually identical to the physically given form of the system—environment
interaction.
Let us assume that the reference state pB of the environment is a Gaussian
state. This implies that all moments of Hi- with respect to pB can be expressed
in terms of moments of second order. Thus, we define the correlation functions

vii (ti t2) = ai trB {Qi(t 1 )gi (t2)PB} (9.54)

t2) = trB {Qi(ti)Qi(t2)PB} (9.55)
where and denote the real and imaginary part, respectively. It will be con-
venient to introduce the following shorthand notation,

Ô= (t), î = F 1 (t1), = Fi2(t2), ••• • (9.56)

and

Vol = P12 = ( t l, t2), •. (9.57)

1101 = 1)01 (t, t 1 ), 1) 12 = (67 t 2)7 ••• • (9.58)
With these definitions any moment of Hi- may be expressed in terms of the
system operators Fi and of the correlation functions vii and ?pi in a compact
way. For example, we have for the second-order moments,

trB {H/(t)H/(t i )ps 0 pB} = E (vol + iqoo ups, (9.59)

io,ii
and, by virtue of the Gaussian property, for the fourth-order moments,

trB fli1(t)Hr(t1)H1(t2)H1(t3)ps 0 PB}

= E R u01 + i 1)01 ) ( v23 + i 1) 23) + (u02 + ii/02)(v13 + i 1) 13)

i0 ,i3

+(v03 + i1)03)(v12 + i2-1 12)] 0:4)s. (9.60)

Invoking the above shorthand notation we can write the second-order contri-
bution (9.45) to the TCL generator as follows,

K2(ops pi3 = — dt, (poi [6, ri,psll + [6, {i,Ps}] ) PB, (9 .61)
io,ii
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 451

while the fourth-order contribution (9.51) becomes

t ti t2

1C4 (t)PS 0 PB = E f chi f f dt2 dt3

io,i,,i27i 0 0 0

x (vo2 vi3[6, [[î,] , [', ps]]] + i/o27713 [6, [[î,] , {A, Ps }]]
+ii/o2v13 [O, { [ i-,*] , [A, Ps] }] — qogin [6, { [i, '] , { A, Ps} 1]

+vo31'12 [6, [[î,] , [, Ps]]] + ivo37/12 [6, [[î,] , {*, Ps }]]

+inovin [6, { [i, 'A] , [, Ps] }] —1/o3 î 12 [6, { [i, "A] , {2, p OH
+(vo3v12 —1)031)12) [6 , [i, [[, 'Al , Ps] ]]

+i(v0312 + n03v12) [6, PI, { [, A] ,Ps } ]] ) ® p B. (9.62)

These formulae will be employed in Chapter 10 to derive the TCL generators of

quantum master equations for specific microscopic models.

9.2.3 The cumulant expansion

A general formula for the n-th-order contribution .C(t) to the generator of
the T,CL master equation can be derived by employing a technique which was
originally designed by van Kampen (1974a, 1974b) for the perturbation expan-
sion of stochastic differential equations. Let us restrict to the homogeneous case
Pp(0) = p(0) and introduce the notation

(X) E P XP (9.63)

for any super-operator X. The formal solution of the Liouville—von Neumann

equation (9.2) then leads to the representation

t
P p(t) = (T, exp [a f dsr(s)1 ) Pp(0) (9.64)
o
for the relevant part of the density matrix. Of course, this is just a shorthand
notation for
t t tl
Pp(t) = 1 + 0 f dti(r(ti)) + 02 f dt i f dt2 (f(t i ),C(t 2 ))
o o o
t ti t2
+03 f dt i f dt 2 f dt3(G(t1),C(t2),C(t3)) + • • • Pp(0). (9.65)
o o o

Differentiating this equation with respect to time we get

452 PROJECTION OPERATOR TECHNIQUES

a
—P p(t) = a(r(t)) + a2 f dt i (f(t),c(to)
at o
t
+a3 f dt i f dt 2 (f(t),C(t 1 ),C(t2)) + • • • 'Pp(0). (9.66)
o o
The trick is now to invert the expansion on the right-hand side of eqn (9.65) to
express Pp(0) in terms of Pp(t) and to insert the result into eqn (9.66). As was
shown by van Kampen this procedure may be carried out in a systematic fashion
to yield an expansion for the equation of motion in powers of a. Comparing this
expansion with the TCL master equation one is led to the following result for
the n-th-order contribution of the generator K(t),
t ti

= f dt i f dt2 f dt n _ i (r(t),C(t i ),C(t2 )... r(tn _ i » 0,, (9.67)

I o
where the quantities

(0)01)02) f(tn---1))oc (9.68)

are called ordered cumulants. They are defined by the following rules.
First, one writes down a string of the form 2L. f`P with n factors of f in
between two Ps. Next one inserts an arbitrary number q of factors P between the
Ls such that at least one f stands between two successive P factors. The resulting
expression is multiplied by a factor (-1)q and all Es are furnished with a time
argument: The first one is always G(t). The remaining Ls carry any permutation
of the time arguments t 1 , t 2, • • • tn-1 with the only restriction that the time
arguments in between two successive Ps must be ordered chronologically. In eqn
(9.68) we thus have t > > t i , ti > . > t k , t i > . > tin , etc. Finally, the
ordered cumulant is obtained by a summation over all possible insertions of P
factors and over all allowed distributions of the time arguments.
For commuting Cs the ordered cumulants reduce to the ordinary cumulants.
The reader may easily check that under the condition (9.11) all odd contributions
K2n-hi (t) vanish and that these rules immediately yield the expressions (9.45) and
(9.51) for the second and the fourth-order contribution of the TCL generator.
9.2.4 Perturbation expansion of the inhomogeneity
As in the Nakajima-Zwanzig equation the inhomogeneity1(t)Qp(0) in the time-
convolutionless quantum master equation (9.33) depends on the density matrix
p(0) at the initial time to = O. For factorizing initial conditions Qp(0) vanishes
and the resulting exact equation of motion is homogeneous. In this section we
discuss the effect of non-factorizing initial conditions of the form
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 453

p(0) = E OkPeqPk E APeq, (9.69)

k
where the operators Ok and Pk act in the Hilbert space Rs of the open system
and peq denotes the equilibrium density matrix of the combined system. This
type of initial condition arises for example in the determination of equilibrium
correlation functions of system observables. It can also be used to describe the
preparation of the system through a quantum measurement since the super-
operator A defined in eqn (9.69) takes the form of a quantum operation if one
chooses Pk = O Section 2.4.2).
The perturbation expansion of the inhomogeneity 1 (t ) Qp(0) may be per-
formed in two steps: First, one expands the super-operator 1(t) given in eqn
(9.35) in powers of the coupling strength a, and, second, one determines Qp(0).
The expansion of 1(t),
DO

1(t) = E anlii (t), (9.70)

is very similar to the expansion of the generator 1C(t) and takes the form (Chang
and Skinner, 1993)

11 (t) = PL(t)Q, (9.71)

t
12(t) = f dt1P,C(t)f(t1)2, (9.72)
0
t ti
13 (0 = f dt i f dt 2 [P,C(t)r(t1)Qr(t2)Q — Pr(t)r(t2)P,C(ti )2], (9.73)
o o
where we again use condition (9.11).
The next step consists in the determination of Qp(0) = 2Ap eq . Since the
super-operators A and Q commute, this amounts to the computation of 2peq
which can be achieved by the following method (Breuer, Kappler and Petruc-
cione, 2001). We define for any fixed T > 0 the density matrix [Mt) to be the
density of a system which has been prepared in such way that the state p,( 7- )
-

at time t = - T is some factorizing state. Thus, we have 7 'p7 ( r) = pr (Y) and

- - —

2p,-(-7- ) = 0. Assuming the total system to be ergodic, we get

peq = lirn p,-(0). (9.74)

T -} CO

This enables one to express Qpeq as

Qpeq = lim Qp,(0). (9.75)

T -> CO

On the other hand, eqn (9.32) yields a relation between the relevant part P1),(0)
and the irrelevant part Qpy (0). Using this relation for to = - T we obtain
454 PROJECTION OPERATOR TECHNIQUES

2p,(0) = [1 — E T (0)] -1 ET (0)PpT (0), (9.76 )

where
o
ET (0) = a f ds g(o, s)2E(s)P G(0, s). (9.77)
- T

Combining eqns (9.75) and (9.76), we get

2Peq = R PPeq, (9.78)

where the operator R. is defined as
DO

R. = lim [1 — E T (0)] -1 ET (0) = lim Y (E T (0)) n . (9.79)

T -} CO
n=1
The relevant part Ppeg of the equilibrium density matrix and the irrelevant part
QPeq are therefore related by the exact equation (9.78).
Also the operator R. may be expanded in powers of a, namely
DO

'R. = E anR.ri . (9.80)

n:.----1

Again, this expansion is accomplished by expanding ET (0) in powers of a, which

yields
o
R, 1 = f dt i r(ti )P, (9.81)

0 ti
R,2 = f chi f dt2Q,C(t1 ),C(t2)P, (9.82)

0 tl t2
R,3 = f dti f dt2 f dt3 [r(t1)Q(t2)r(t3) (9.83)

—,C(t 2 )7) ,C(t 1 ),C(t3) —,C(t 3 )P,C(ti ),C(t2 )]P.

The last step which completes the expansion of the inhomogeneity is the
combination of the expansions for the operators 1(0 and R. which finally leads
to

/(t)Qp(0) = IMARY peg E J (OP peg . (9.84)

Explicitly, the second and the fourth-order contributions to the super-operator
J(t) are given by
 THE TIME-CONVOLUTIONLESS PROJECTION OPERATOR METHOD 455

0
J2(t) = MARI dt1P,C(t)A,C(t1)P, (9.85)
= f
-DC

and by

( t) = li (t)AR,3 +12 M AR, 2 +13 (t)AR, 1 , (9.86)

with
0 tl t2

11 (t)AR3 dti f dt2 f dt3 [Pf(t)Ar(t1)Qr(t2).03)P

= f

—PG(t)A,C(t2 ) PE(t 1 ),C(t3)P — `1) ,C( t )A.C(t3)P01)02)P]

t 0 t2

I2 (t )A 1,2 =
f dti f dt 2 f dt3 Pf(t) ,C (t i )A2r(t 2 )r(t3 )P,
—cc —cc
t o
13 ( t )A1?, 1 = dti f dt 2 f dt 3 [P r(t)r(ti )2r(t2 )Ar(t 3 )P
0
f 0 - DC
—P(t)r(t 2 )PL(t i )AL(t 3 )P]. (9.87)

Using the form (9.53) for the interaction Hamiltonian and employing the short-
hand notation introduced in Section 9.2.2 we obtain for the second-order contri-
bution,

(t) PS 0 PB = E f dt, (vol [6, A [i, Ps]] + , A {i, ps}]) ®PB,

(9.88)

which is similar to the expression for 1C2 (t) (compare eqn (9.61)). Section 10.2.4
gives an example of the application of the above method to the calculation of
equilibrium correlation functions.

9.2.5 Error analysis

The perturbation expansion of the generator IC(t) and of the inhomogeneity J(t)
to various orders in the coupling strength a can be used to obtain a computable
estimation for the error introduced by a certain approximation. To this end, we
consider the relative error

Ig o _ E a2k,c2k(t)
(2n) k=1
er = (9.89)
Iligt)11
456 PROJECTION OPERATOR TECHNIQUES

of the approximation of order 2n, where 111(11 denotes any appropriate norm for
the super-operator /C. To leading order in a this can be written as
e (2,) = a 2n IIK2n+2 (9.90)
IIK2
Thus, the error of the approximation to second order can be related to the
fourth-order approximation by
(2) 2 I IK 4(011 (9.91)
er (o w
If we assume that the coefficients /C2 and /C4 have the same order of magnitude ,
then this relation can be used to estimate the formal expansion parameter a by
means of the known relative error, i.e.
e (2) (9.92)
Employing the same argument again, we find for the error of the fourth-order
approximation

e (4) ce4 (t)

(9.93)
r IlK: 2 (t) Ir
However, if the approximation to sixth order is not known, we cannot compute
e 4) directly. To construct a computable error estimate we use the assumption
that the orders of magnitude of 1C 2 , 1C4 and 106 are the same, and we obtain
e (4) a4 (e (2)) 2 (9.94)

Thus, the relative error of the approximation to fourth order can be estimated by
computing the square of the relative error of the approximation to second order.
This procedure yields for a certain order of approximation a crude error estimate
which is easy to compute since it only relies on the actual approximation and not
on the evaluation of higher-order terms. A number of examples will be studied
in the next chapter.
The same arguments can of course also be applied to the matrix elements
of the super-operator k(t) and of the inhomogeneity J , yielding computable
error estimates for these quantities as well.

9.3 Stochastic unravelling in the doubled Hilbert space

The most general master equation for the reduced density matrix p(t) which
results from the time-convolutionless projection operator technique takes the
following form in the homogeneous case,
a +E ci (ops (t)Dti (t),
— ps(o= A(t)ps(t)+ ps(t)Bt(t) (9.95)
at

with some time-dependent linear operators A(t), B(t), Ci(t) and D,(t). This
equation is linear in p(t) and local in time, but it need not be in Lindblad form.
 STOCHASTIC UNRAVELLING IN THE DOUBLED HILBERT SPACE 457

Nevertheless, it can be represented by a certain stochastic process for the wave

functions of the reduced system (Breuer, Kapp ler and Petruccione, 1999).
In order to construct an unravelling of eqn (9.95) we follow the strategy
which was already applied to the calculation of multi-time correlation functions
in Section 6.1.4. Namely, we describe the state of the open system by a pair of
stochastic wave functions

4)(0
0(t) = (9.96)

such that 0(t) becomes a stochastic process in the doubled Hilbert space 7-I =
I-Is e 7-is. Denoting the corresponding probability density functional by P[0, t ],
we can define the reduced density matrix as

ps(t) = f DODO* P[0,t] (9.97)

Consider now the following stochastic differential equation for the process
0(t) in the doubled Hilbert space,

d0(t) = —iG (0 , t)dt + E v1, :1 (0 0(0 11 (00(t)

t
_ 0 (t ) ) dNict), (9.98)

where the Poisson increments dNi (t) satisfy

11Ji (t)0(t)11 2
E[dNi (t)] = dt (9.99)
11 0(t)112 '
dNi(t)dN (t) = dNi (t), (9.100)

and the non-linear operator G(0, t) is defined as

J i(t)0 (OW
G (0 ,t) = (t) + 110(0 112 ) °V , (9.101)
'

with the time-dependent operators

(A(t) (t) )
F( t) = Ji( t) = (9.102)
0 B(t) D(t)

Equation (9.98) describes a PDP whose deterministic pieces are solutions of

the differential equation

(t) = G (0 , t), (9.103)

and whose jumps take the form

458 PROJECTION OPERATOR TECHNIQUES

0(t)
110(011 (t)0(t) =
110(t)ii Ci(t)0(t)
(9.104)
11J0(t)1I 11Ji0(t)11 Di(t)0(t)

Employing the rules of the calculus for PDPs one easily demonstrates that the
covariance matrix

f5(t) = f DODO* 13 [0 , t]l0) (01 (9.105)

of the process 0(t) satisfies the equation of motion

a
— i5(t) = F(t)i5(t) + i5(t)Ft (t) + E (t)j)(t)Jit (t). (9.106)
at
Note that, in general, this equation is also not in Lindblad form. The important
point is, however, that the component (9.97) of i5(t) obeys eqn (9.95), as is easily
verified using the block structure of the operators F(t) and Ji (t). This provides
the sought unravelling of eqn (9.95) in terms of a process in the doubled Hilbert
space. A specific example will be studied in Section 10.1.3.

References
Balescu, R. (1975). Equilibrium and Nonequilibrium Statistical Mechanics. John
Wiley, New York.
Breuer, H. P., Kappler, B. and Petruccione, F. (1999). Stochastic wave-function
method for non-Markovian quantum master equations. Phys. Rev., A59, 1633—
1643.
Breuer, H. P., Kappler, B. and Petruccione, F. (2001). The time-convolutionless
projection operator technique in the quantum theory of dissipation and deco-
herence. Ann. Phys. (N. Y.), 291, 36-70.
Chang, T.-M. and Skinner, J. L. (1993). Non-Markovian population and phase
relaxation and absorption lineshape for a two-level system strongly coupled to
a harmonic quantum bath. Physica, A193, 483-539.
Chaturvedi, S. and Shibata, F. (1979). Time-convolutionless projection operator
formalism for elimination of fast variables. Application to Brownian motion.
Z. Phys., B35, 297-308.
Grabert, H. (1982). Projection Operator Techniques in Nonequilibrium Statis-
tical Mechanics, Volume 95 of Springer Tracts in Modern Physics. Springer-
Verlag, Berlin.
Haake, F. (1973). Statistical treatment of open systems by generalized master
equations. Springer Tracts in Modern Physics, 66, 98-168.
Kubo, R., Toda, M. and Hashitsume, N. (1985). Statistical Physics II. Nonequi-
librium Statistical Mechanics. Springer-Verlag, Berlin.
Nakajima, S. (1958). On quantum theory of transport phenomena. Progr.
Theor. Phys., 20, 948-959.
Prigogine, I. (1962). Non-Equilibrium Statistical Mechanics. Interscience Pub-
lishers, New York.
 REFERENCES 459

Shibata, F. and Arimitsu, T. (1980). Expansion formulas in nonequilibrium

statistical mechanics. J. Phys. Soc. Jap., 49, 891-897.
Shibata, F. and Takahashi, Y. and Hashitsume, N. (1977). A generalized
stochastic Liouville equation. Non-Markovian versus memoryless master equa-
tions. J. Stat. Phys., 17, 171-187.
van Kampen, N. G. (1974a). A cumulant expansion for stochastic linear differ-
ential equations. I. Physica, 74, 215-238.
van Kampen, N. G. (1974b). A cumulant expansion for stochastic linear differ-
ential equations. IL Physica, 74, 239-247.
Zwanzig, R. (1960). Ensemble method in the theory of irreversibility. J. Chem.
Phys., 33, 1338-1341.
 10
NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

In this chapter we apply the time-convolutionless (TCL) projection operator

technique developed in the preceding chapter to some typical examples. The first
example is that of the spontaneous decay of a two-level system into a reservoir
with arbitrary spectral density. For this simple problem the TCL expansion can
be carried out to all orders in the system—reservoir coupling. The model therefore
serves to examine the performance of the TCL method.
As specific examples we treat the damped Jaynes—Cummings model and the
spontaneous decay into a photonic band gap. The damped Jaynes—Cummings
model on resonance is also used to exemplify the breakdown of the TCL expan-
sion in the strong coupling regime. In the off-resonant case the model leads to a
TCL generator which is not in Lindblad form and involves a time-dependent de-
cay rate that can take on negative values. The corresponding unravelling through
a stochastic process in the doubled Hilbert space designed in Section 9.3 will be
illustrated by means of numerical simulations.
The last two sections are devoted to quantized Brownian motion, where non-
Markovian effects play an important rôle for certain parameters. The first exam-
ple is the damped harmonic oscillator, the second one is the spin-boson model.
The purpose of this chapter is to illustrate that the TCL projection operator
technique has several advantages in practical applications. Although the tech-
nique is clearly perturbative, it does not rely on an expansion of a given physical
quantity, but rather on an expansion of the equation of motion for the reduced
density matrix which may then be solved either analytically or numerically. From
this solution the desired physical quantities are determined. In many cases the
range of validity of this procedure is much larger than that achieved by a direct
perturbation expansion of the desired quantity to the same order.
The TCL technique is applicable to a large class of physical systems: It does
not rely on a specific form of the interaction, of the initial state of the combined
total system, or of the spectral density of the environment, nor does it employ
specific symmetry or scaling properties. Moreover, the perturbation expansion
of the TCL generator provides a systematic way to go beyond the Markovian
approximation. For many systems that arise in the theory of open quantum sys-
tems the analysis of the Markovian approximation is relatively easy to perform
and often gives a first qualitative picture of the long-time dynamics. The expan-
sion technique then offers the possibility of studying non-Markovian phenomena
in a systematic way, or to judge whether or not a non-Markovian treatment is
necessary.
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 461

10.1 Spontaneous decay of a two-level system

The first example of the application of the TCL projection operator technique
is an exactly solvable model of a two-level system which decays spontaneously
into the field vacuum. This system has been studied already in Section 3.4.2
employing the Born—Markov approximation which led to an equation of motion
for the reduced density matrix in Lindblad form. Here we treat the model in the
rotating wave approximation to all orders in the coupling between system and
reservoir. The exact solutions for various forms of the spectral density will be
compared with the results of the time-convolutionless master equation derived
in Section 9.2.1 and with the corresponding perturbation expansion of the TCL
generator 1C(t) developed in Sections 9.2.2 and 9.2.3.
The Hamiltonian of the total system is given by

H Hs + HB ± HI = HO + (10.1)

where

Ho = Ewkblbk, (10.2)

=- a+0 B + a ®Bt with B =-- Egkbk• (10.3)

The transition frequency of the two-level system is denoted by wo , and a+ are

the raising and lowering operators. The index k labels the different field modes
of the reservoir with frequencies wk, creation and annihilation operators btk , bk
and coupling constants gk.

10.1.1 Exact master equation and TCL generator

Let us introduce the states (Garraway, 1997)

= 1 0) s 0 10)B, (10.4)
=- 11)s ® (10.5)
7/4 i o) s ®1k)B, (10.6)
where 10)s = a_ 11)s and 11)s =- o-± 10)s indicate the ground and excited state of
the system, respectively, the state 10)B denotes the vacuum state of the reservoir,
and k)B = b0)B denotes the state with one photon in mode k. In the interac-
tion picture the state 0(0 of the total system obeys the Schrödinger equation

d
— 0(0 =-- —i_111(t)0(t), (10.7)
dt
where

HIM = 0-+(t)B(t) + a_ (t)Bt (t) (10.8)

462 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

is the interaction picture Hamiltonian with

0-±(t) = a+ exp (±iwot), (10.9)

and

B(t) , E gk bk exp(—iC.Jkt). (10.10)

k

It is easy to check that the total Hamiltonian (10.1) commutes with the
'particle number' operator

N , or+o- _ + E btk bk, (10.11)

k

i.e. we have

[H, NJ = 0. (10.12)
Thus, N is a conserved quantity. It follows that any initial state of the form

0(0) — coo° + cl (o)01 + E ck(0)11)k (10.13)

k

evolves after time t into the state

o(t) = coot, + el (t)lpi + E c i (t)/, j . (10.14)

k

The amplitude co is constant since H1(t)7,b0 =- 0, while the amplitudes c1 (t) and
ck (t) are time dependent. The time development of these amplitudes is governed
by a system of differential equations which is easily derived from the Schrödinger
equation (10.7),

èi(t) = — i E gk exp[i(wo — wk)t]ck (t), (10.15)

k
è k (t) =- — igi*, exp[—i(c.J0 — cuk)t]c i (t). (10.16)
Assuming that ck (0) = 0, i.e. there are no photons in the initial state, we solve
the second equation and insert the solution into the first to get a closed equation
for ci (t),
t
6 (t) = — f dti f (t — t 1) 6 (t i ) • (10.17)
0
This procedure was already used in the classic paper by Weisskopf and Wigner
(1930) on the determination of the natural line width. The kernel f (t t 1 ) is
—

given by the correlation function

 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 463

f (t – t 1 ) = trB {B(t)13 1. (t1 )pB lexp[iw o (t — t1)], (10.18)

where p B = (10) (OD B is the vacuum state of the reservoir. As usual, this kernel
may be expressed in terms of the spectral density J(w) of the reservoir as follows,

f (t – t i ) = f du.a (w) exp[i(wo – w)(t – t1)]. (10.19)

The exact equation (10.17) can now be solved by means of a Laplace transfor-
mation. In subsequent sections we will present solutions for different forms of
the spectral density.
With the help of the probability amplitudes co, ci (t) and ck(t) we can now
express the reduced density matrix ps (t) of the two-state system as
ecoic(i,t)( ti 2) 1 c; 1 ((tt)) 12 ) .
Ps (t) = trB { I OM) (OM I } = ( l (10.20)

Differentiating this expression with respect to time and recalling that o = 0 we

get
d A l ei (t)I 2 c'6êi (t)
(10.21)
di P s (t) = 1 dte0 é; (t) - 11C1 (t )1 2 ) •
-

Introducing further the quantities

S(t) , {- t) } ,
-2a ci((t (10.22)

-y(t) = cil ( t ) } , (10.23)

we can rewrite eqn (10.21) as

d i
— Ps(t) = – – S(t)[(7-0--,Ps(t)] (10.24)
dt 2
1 1
+-y(t) {o-_ ps (Oa+ – – o-± a_ ps (t) – – Ps (t)a+a– }
2 2
This is an exact master equation for the reduced system dynamics. Obviously,
S(t) plays the rôle of a time-dependent Lamb shift and -y(t) that of a time-
dependent decay rate.
The remarkable point of the master equation (10.24) is that it is already of
the form of a time-convolutionless master equation,
d
Ps(t) = ks(t)Ps(t), (10.25)

where the generator AC(t) is connected to the TCL generator 1C(t) by means of
the relation
ks (t)ps (t) = trB VC (t)ps (t) 0 p B} . (10.26)
We observe that the structure of )C(t) is similar to that of a Lindblad generator.
However, due to the time dependence of the coefficients 8(t) and -y(t) eqn (10.24)
464 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

does not provide, in general, a quantum dynamical semigroup. We will also

see below that the time-dependent rate -y(t) may even become negative which
violates the complete positivity of the generator.
In Sections 9.2.2 and 9.2.3 a perturbation expansion for the TCL generator
in powers of the coupling strength a was developed. Let us demonstrate that
this expansion produces in all orders a master equation of the form (10.24) and
that it generates a perturbation expansion of the coefficients -y(t) and 8(t) in the
master equation which takes the form

-y (t) E a2n,y27, ,
n=1
(10.27)

8(t) =- E a
oo
(t) . 2ns2n
(10.28)
n=1

To get this expansion we introduce an expansion parameter a by replacing the

correlation function f (t) by a 2 f (t), since the spectral density is to be considered
as a quantity of second order in the coupling.
We first note that o-± =- (11) (01)s is an eigenoperator of the generator K(t).
i.e. we have
1
= — — [-y(t) + iS(t)jo-+. (10.29)
2
On the other hand, with the help of (10.26) the general formula (9.68) for the
expansion of the TCL generator into ordered cumulants leads to
t t1 t2rt -2
00

ks( t )g+ E a2n f dt i f dt 2 f dt 21 1 (10.30)

n=1 0 0 0
xtrB {(f(t),C(t i ) f(t 2n_ i )) 0 c0-+ 0 p B}

Note that only even orders of a appear in the expansion because relation (9.11) is
satisfied here. Employing (t) p = i[H (t) , p] one easily verifies that o-± 0 p B is
—

an eigenoperator of the super-operator r (t) ,C (t 1 ) corresponding to the eigenvalue

- f (t - t 1 ) , i.e.

L(0000 -± O pB = —f(t — ti)o-+ 0 pB. (10.31)

Since further Pa+ 0 pB = u+ O pB the expansion (10.30) can be restated as

-

follows,
t il i2n -2
00

KS (Oa+ = E a2n f dti f dt2 f dt2„ 1 (10.32)

n=1 o o
X ( -1 ) 71 (f tl (t2 t3) f(t27-2 t27-1 )) De g +-
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 465

Comparing the last expression with eqn (10.29) we immediately get the TCL
expansion for the coefficients of the exact quantum master equation

t1 t2n —2

2'2n (t) iS2n (t) f dt i f dt 2 . . f dt27-1 (10.33)

0 0 0
X 2 ( -1-) n±l (f (t (t2 t3) • • f (t2n-2 t2n - 1))0c •

The rules for the construction of the quantity (f (t t 1 ) .. . f (t22 t2 7 -1))0c — —

can be stated as follows: Write down the corresponding expression for the or-
dered cumulant of order 2n according to the rules given below eqn (9.68). Omit
all factors of P and replace the pairs G(ti )r(ti ) of successive L-factors by the
correlation function f (t i t). —

Following these rules eqn (9.45) immediately yields the second-order contri-
bution for the coefficients of the master equation,

-Y2(t) iS2(t) = 2 f f (t - ti) (10.34)

0

while (9.51) leads to the fourth-order contribution,

t t2

- Y4 (t) iS4(t) = 2 f dt i f dt 2 f dt3[f (t - t2) f (ti - t3) f t3)f (t1 t2 )] .

0 0 0
(10.35)

It is convenient to define the real functions CO and 'I'(t) through

2f (t) = (t) (t). (10.36)

In terms of these functions the second-order contributions to the decay rate -y(t)
and to the energy shift S(t) determining the Born approximation read

-y2 t f dt i (1)(t — ti ), (10.37)

492(t) = f dti xl/(t — t 1 ), (10.38)

and hence the Markovian decay rate -ym and the Markovian Lamb shift Sm are
found by extending the integration to infinity,
466 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

FIG. 10.1. Schematic representation of the damped Jaynes-Cummings model.

00
Ym = f ds ( s),
- (10.39)
o
00
Sm , f ds W(s). (10.40)
o
The fourth-order TCL contributions to the decay rate and to the energy shift
may be expressed as
t tl t2
1
= f dt1 f dt2 f dt3 [ 4:1) (t — t2) (1) (t1 — t3) ± 41 (t — t3) (1) (t1 — t 2)
0 0 o
-111(t - t 2 )111(t 1 - t3) - ‘11 (t - t3)‘11 (t1 - t2)] , (10.41)

and
t t2 ti
S4(t) =-1 f dti f dt2 f dt3 [T(t - t2)(t1 - t3) + 1:11 (t - t2)W(t1 - t3 )
2
o o o
+T(t - t 3 )(1.(t 1 - t2) + 4)(t - t3)T(t1 - t2)] • (10.42)

Thus we see that the form of the exact master equation (10.24) is preserved
in all orders of the coupling and that the TCL expansion amounts to an ex-
pansion of the coefficients S(t) and -y(t). In the following subsection we study
the non-Markovian behaviour of some specific models involving different spectral
densities (Breuer, Kappler and Petruccione, 1999).

10.1.2 Jaynes-Cummings model on resonance

The damped Jaynes-Cummings model describes the coupling of a two-level atom
to a single cavity mode which in turn is coupled to a reservoir consisting of har-
monic oscillators in the vacuum state. A schematic representation of the Jaynes-
Cummings model is depicted in Fig. 10.1. If we restrict ourselves to the case of a
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 467

1.5
TCL 4
— exact
Markov
1.0

GME 2
0.5 exact -
-- TCL 4
TCL 2 0.2
Markov
0.O L. 0.0
0 1 2 3 4
-yot -yot
(a) (b)

0.05 1.0 \
— exact
TCL 4
1 - GME 2
0.5

o TCL 2 0.0
TCL 4 -
GME 2
Markov
-0.15 -0.5
0 1 2 3 4 0 10 15 20
'yot -yot
(c) (d)

FIG. 10.2. The damped Jaynes—Cummings model on resonance. Exact solution

(exact), time-convolutionless master equation to second (TCL 2) and fourth
order (TCL 4), generalized master equation to second order (GME 2), and
the Markovian quantum master equation (Markov): (a) Decay rate of the
excited state population, (h) the population of the excited state, and (c)
deviation of the approximate solutions from the exact result, for TR = 5-TB
(moderate coupling). The symbols show the results of a stochastic simulation
of the time-convolutionless quantum master equation using 105 realizations.
(d) Population of the excited state for TR =- 0•27-B (strong coupling).

single excitation in the atom—cavity system, the cavity mode can be eliminated
in favour of an effective spectral density of the form

\ 1 -yo
(10.43)
44)) — 27r (woo — w) 2 + A2 '

where coo is the transition frequency of the two-level system. The parameter A
defines the spectral width of the coupling, which is connected to the reservoir
correlation time TB by the relation TB = A-1 and the time scale TR on which the
state of the system changes is given by TR
468 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

In order to compute the exact probability amplitude c1 (t) we evaluate the

reservoir-correlation function f (t) using the spectral density (10.43),

1
At) = — -y0 A exp(—Alti). (10.44)
2
For this f (t) the differential equation (10.17) for the probability amplitude cl (t)
can easily be solved to give the exact solution

()
dt , _A . , (dt
c1 (t) = c1 (0)e —At / 2 [cosh 2 + d smn —2--)] , (10.45)

where d =- /A2 — 2-y0 A. This yields the time-dependent population of the excited
state

P1 1 (t) = pi 1 (0)e — At cosh

[
( dt_ . A ., dt 2
(10.46)
2 + d sum —2- -

Using eqns (10.22) and (10.23) we therefore obtain a vanishing Lamb shift, 8(t)
0, and the time-dependent decay rate
270 A sinh(dt/2)
(10.47)
d cosh(dt 12) + A sinh(dt/2) .

In Fig. 10.2(a) we show this time-dependent decay rate -y(t) ('exact') together
with the Markovian decay rate -ym = -yo ('Markov') for TR = 5TB. For short times,
i.e. for times of the order of TB, the exact decay rate grows linearly with t, which
leads to the correct quantum mechanical short-time behaviour of the transition
probability. In the long-time limit the decay rate saturates at a value larger
than the Markovian decay rate, which represents corrections to the golden rule.
The population of the excited state is depicted in Fig. 10.2(b): For short times,
the exact population decreases quadratically and is larger than the Markovian
population, which is simply given by P11 (0) exp (--y0 t), whereas in the long-time
limit the exact population is slightly less than the Markovian population.
Next, we determine the solution of the generalized quantum master equation
(9.26) in the Born approximation. To this end, we insert the spectral density of
the coupling strength (10.43) into eqn (10.36) to obtain W (t) E- 0 and

(1)(0 = -y0 A exp(—At). (10.48)

This leads to the following form of the generalized master equation,

t
d
Ps(t) = lioA f ds e —À(i— s) [a_ps(s)o-± — —
1
2
o-±a_ps(s) —
1

o
(10.49)
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 469

We differentiate with respect to t and obtain

1 1
iis(t) — /Vs(t) + -yox [a_ ps (t)o-± — —2 u+ o-_ ps (t) — —2 ps (t)o-±H . (10.50)
Due to the exponential memory kernel, this equation of motion is an ordinary
differential equation which is local in time, and contains only ps(t), 1s(t) and
lis (t). The solution leads to the following time evolution of the population of the
upper level,
()
d't , (d't
(0)e —At / 2 [cosh 2 + —idS11111 T)] (10.51)

where (1' = V A 2 — 4-yo A. From this expression, we can determine the time-
dependent decay rate

/i1 1 (t) 2-y0 A sinh(dt/2)

-Y(t) =
-
--

= (10.52)
P11 (t) d' cosh(d't/2) + sinh(d't/2)
the structure of which is similar to the exact decay rate (10.47). Note, how-
ever, the difference between the parameters d and d' which can also be seen in
Fig 10.2(a) where we have also plotted the decay rate (t) (`GME 2'): For short
times, the decay rate "-y(t) is in good agreement with 7(0, but in the long-time
limit, -y(t) is too large.
Finally, the time-convolutionless decay rate can be determined from eqn
(10.41). The second and fourth-order contributions are given by
72(0 = yo (1 — e') , (10.53)
and

'-y4(t) = [sinh(At) — At] e - At . (10.54)

This corresponds to a Taylor expansion of the exact decay rate -y(t) in powers
of the expansion parameter a2 = '-yo/A =-- TB/TR. Thus we see that the TCL
expansion provides an expansion in the ratio of the reservoir correlation time TB
to the system's relaxation time TR. Figure 10.2(a) clearly shows that the second-
order contribution -y2 (t) to the TCL expansion as well as the contribution up
to fourth order 7 (4) (0 = -y2(t) + '-y4(t) approximate the exact decay rate very
well for short times, and that the fourth-order contribution -y (4) is also a good
approximation in the long-time limit.
The time evolution of the population of the excited state can be obtained by
integrating the rate -y (4) (t) with respect to t. This yields

p(141) (t) = pli (0) exp (— f ds -y (4) (s)) (10.55)

o
In order to compare the quality of the different approximation schemes, we show
the difference between the approximated populations and the exact population
 470 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

in Fig. 10.2(c). In addition to the analytical solutions of the generalized master

equation and the time-convolutionless master equation, we have also performed
a stochastic simulation of the time-convolutionless quantum master equations
with 10 5 realizations. Since the time-dependent rates -y2(t) and -y (4) (t) are pos-
itive for all t, the corresponding generators are in Lindblad form, and we can
use the stochastic unravelling developed in Section 6.1.1. Figure 10.2(c) shows
that the stochastic simulation is in very good agreement with the corresponding
analytical solutions. Moreover, we see that the difference between the solution of
the time-convolutionless master equation to fourth order and the exact master
equation is small (see also Fig. 10.2(b)), whereas the errors of the generalized
master equation and of the time-convolutionless master equation to second order
are larger and of the same order of magnitude. In fact, the Markov approxima-
tion even leads to a slight improvement of the accuracy, compared to the Born
approximation, which is surprising if we consider the heuristic derivation of the
quantum master equation.
The approximation schemes used here are perturbative and hence rely on the
assumption that the coupling is not too strong. But what happens if the system
approaches the strong coupling regime? We will investigate this question by
means of the damped Jaynes—Cummings model on resonance, where the explicit
expressions of the quantities of interest are known.
First, let us take a look at the exact expression for the population of the
excited state (10.46): In the strong coupling regime, that is for -yo > A/2 or
TR < 2TB, the parameter d is purely imaginary. Defining d = —id we can write
the exact population as

ci2t ± Aci sin _2_

cit 2,
p11 (t) = pii (0)e —At COS (10.56)

which is an oscillating function that has discrete zeros at

2 ii
t, mr — arctan — (10.57)
d A) '

where n =-- 1, 2, .... Hence, the rate -y(t) diverges at these points (see eqn (10.23)).
Obviously, -y (t) can only be an analytical function for t E [0, t o [, where to is the
smallest positive zero of p ii (t).
On the other hand, as we have just seen, the time-convolutionless quantum
master equation corresponds basically to a Taylor expansion of -y (t) in powers of
-Yo, and the radius of convergence of this series is given by the region of analyticity
of -y(t). For -yo < A/2, this is the whole positive real axis, but for -yo > A/2 the
perturbation expansion only converges for t < to. This behaviour can be clearly
seen in Fig. 10.2(d), where we depict p ii (t) and ik(t) for TR = TB/5, i.e. for
strong coupling: The perturbation expansion converges to p11 (t) for t < to
6.3/ yo, but fails to converge for t > to.
-
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 471

0 10 15 20
Yot
-

FIG. 10.3. The damped Jaynes-Cummings model on resonance. Exact popu-

lation for the three different initial conditions p» (0) = 1.0, 0.5, 0.0 in the
strong-coupling regime (7-R = 0.21-B).

The solution of the generalized master equation to second order shows a quite
distinct behaviour, but also fails in the strong-coupling regime: For -yo > A/4 the
population P11 (t) starts to oscillate and even takes negative values, which is
unphysical (see Fig. 10.2(d)).
The 'failure' of the time-convolutionless master equation at t = to can also
be understood from a more intuitive point of view. The time-convolutionless
equation of motion states that the evolution of the reduced density matrix only
depends on the actual value of ps (t) and on the TCL generator. However, at
t = to the time evolution also depends on the initial value of the density matrix.
This fact is illustrated in Fig. 10.3, where we have plotted the population p1 1 (t)
for three different initial conditions, namely p i]. (0) = 1.0, 0.5, 0.0. At t = to,
the corresponding density matrices coincide, regardless of the initial condition.
However, the future time evolution for t > to is different for these trajectories.
It is therefore intuitively clear that a time-convolutionless form of the equation
of motion which is local in time ceases to exist for t > to . The formal reason for
this fact is that at t = t o the operator 1- E(t) (see Section 9.2.1) is not invertible
and hence the generator 1C(t) does not exist at this point.

10.1.3 Jaynes Cummings model with detuning

-

In this section we treat the damped Jaynes-Cummings model with detuning,

i.e. the same setup as in the preceding example but the centre frequency of the
cavity is detuned by an amount A against the atomic transition frequency w o .
In this case the spectral density takes the form
472 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

f \ 1 -yo A 2
(10.58)
4Â)) 27 (c.o0 - A - co) 2 + A2 '
and thus the functions (DM and W(t) are given by

4.(t) =- -yoAe -At cos(t), (10.59)

Ili(t) = -yo Ae -At sin(At). (10.60)
With these functions, the time-dependent Lamb shift S (4) (t) = S2 (t) ± S4(t) and
decay rate (4) (t) = -y2(t)+N(t) to fourth order in the coupling can be calculated
using eqns (10.42) and (10.41). The integrals can be evaluated exactly and lead
to the expressions

s( 4)(t) , -yo AA
A2 + A2 [1 - e -At (COS(At) + 1 sin(At))] (10.61)
-yd A2 A3 C AL
{ [1 - 3 (*) 2 ] (e At — e -A1 COSPAO)
2(A2 ± A2 ) 3

I
- 2[1 - (A-) 4 ]Atsin(At) +4[1 + (1) 2 At COS(At)

- 1- [ 3 - ( 1) 2 IC At sin(2At) 1
and

-YoÀ 2
= [1 — e — At (cos(At) — sin(At))1 (10.62)
A 2 + A2
+ ydA5e—A { [
-

(°,-
3 , ) 2 ] (e At e - At cos(2At))
2(A2 + A 2)t 31
- -

I
_2[1 - () 4 At cos(At) + 4 [1 + ('-) 2 ] At sin(At)

+1-: [3 - (1'-) 2 ] e -At sin(2At)}.

In Fig. 10.4(a) we have depicted -y(4)(t) together with the exact decay rate,
which can be calculated by solving the differential equation for the probability
amplitude e1 (t) for A = 8A and A =- 0.3-yo. Note that the spontaneous decay rate
is severely suppressed compared to the spontaneous decay on resonance. This
can also be seen by computing the Markovian decay rate -ym which is given by

-yo A2
- Ym = A2 + A 0.015-Yo• (10.63)
2
However, this strong suppression is most effective in the long-time limit. For short
times, -y(t) oscillates with a large amplitude and can even take negative values.
which leads to an increase of the population. This is due to photons which have
been emitted by the atom and are reabsorbed at a later time. Hence, the exact
quantum master equation as well as its time-convolutionless approximation are
 SPONTANEOUS DECAY OF A TWO-LEVEL SYSTEM 473

0.15 r exact
1.00 ,

TCL 4
0.10 Markov

eS' 0.05
r
Zi 0.90
0.00 \ /-\\

-0.05

-0.10 0.80
0 6 12 18 6 12
-Yot "YOt
(a) (b)

FIG. 10.4. The damped Jaynes—Cummings model with detuning. Exact solu-
tion (exact), time-convolutionless master equation to fourth order (TCL 4),
and the Markovian quantum master equation (Markovian): (a) Decay rate
of the excited state population, and (b) the population of the excited state,
including a stochastic simulation of the time-convolutionless quantum master
equation with 10 5 realizations (symbols), for A = 0.3-yo and A = 8A.

not in Lindblad form, but conserve the positivity of the reduced density matrix.
This is of course not a contradiction to the Lindblad theorem, since a basic
assumption of this theorem is that the reduced system dynamics constitutes a
dynamical semigroup. This assumption is obviously violated here.
Although the time-convolutionless master equation is not in Lindblad form
and involves negative transition rates, it can be represented through an ap-
propriate stochastic process as was shown in Section 9.3. The dynamics of the
stochastic wave function 0(t), which is an element of the doubled Hilbert space
= "1-ts ED1-1s, is governed by the stochastic differential equation (9.98), where
the operator F(t) is given by
0
F(t) = —Ty (t) 0 (10.64)
,

and we have a single jump operator

(y (4 )(t)o-_ 0
J(t) _ (10.65)
0 o-

The jumps induce instantaneous transitions of the form

0(t) ii1j190( (t)t )i 111 J0(t) ( 7(4)1 0( t))sl0)s (10.66)

If the rate -y( 4) (t) is positive these transitions lead to a positive contribution to
the ground state population p00 (t), while a negative rate results in a decrease of
Poo (t) •
474 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

In Fig. 10.4(b), we show the results of a stochastic simulation for 105 realiza-
tions, together with the analytical solution of the time-convolutionless quantum
master equation and the exact solution. Obviously, the agreement of all three
curves is good and the stochastic simulation algorithm works excellently even
for negative decay rates. In addition, we also show the solution of the Markovian
quantum master equation which clearly underestimates the decay for short times
and does not show oscillations.

10.1.4 Spontaneous decay into a photonic band gap

As our final example, we treat a simple model for the spontaneous decay of a
two-level system in a photonic band gap which was introduced by Garraway
(1997). To this end, we consider a spectral density of the form
0 2
W1 fi
(
W2 f2
J (w ) _ " 0 (10.67)
27r (a) — c.o0) 2 ± (r 1 /2)2 (c.,) — w0 )2 + (r2 /2)2)'
where 118 describes the overall coupling strength, 1' 1 the bandwidth of the 'flat'
background continuum, F2 the width of the gap, and W1 and 1172 the relative
strength of the background and the gap, respectively. In Fig. 10.5 we show the
excited state's decay rate -y(t) for the parameters F1/110 =- 10, F 2 /110 = 1,
W1 = 1.1, and W2 = 0.1. For short times, -y(t) increases linearly on the time scale
171 and then attains a maximum value. This phase of the dynamics stems from
transitions into the 'flat' background continuum. For longer times, i.e. t » F 2-1 ,
the system resolves the structure of the spectral density and transitions are
strongly suppressed due to the presence of the gap. The decay rate therefore
becomes smaller and smaller until it reaches its final value. Thus, the population
of the excited state decreases rapidly for times of order IV , and slowly in the
long-time limit (see Fig. 10.5).
The time-dependent Lamb shift S (4) (t) and the decay rate (4) (t) of the time-
convolutionless quantum master equation to fourth order can be computed by
inserting the spectral density of the coupling strength J(w) into eqn (10.36).
This yields xlf(t) E 0 and

(1) (t) = 2 11 8 ( Wi e -rit/2 - W2 e -r2t/2 ) , (10.68)

which can be inserted into eqns (10.42) and (10.41). Since W(t) E:- 0 the Lamb
shift S 4 (t) vanishes; the time-dependent decay rate -y (4) (t) can be computed
explicitly, and is in good agreement with the exact decay rate for our choice of
parameters (see Fig. 10.5).

10.2 The damped harmonic oscillator

The damped harmonic oscillator has already been studied at various places in
Chapter 3, where we discussed the master equation (3.307) in the quantum opti-
cal limit and the high-temperature Brownian motion master equation (3.410). It
 THE DAMPED HARMONIC OSCILLATOR 475

0.4 T

— exact x TCL 4
TCL 4 exact

0.0
0 2 4 6 8 10 0 2 4 10
clo t Qo t

FIG. 10.5. Spontaneous decay in a photonic band gap: Exact solution (exact),
and time-convolutionless master equation to fourth order (TCL 4). Left: De-
cay rate of the excited state population. Right: The population of the excited
state, including a stochastic simulation of the time-convolutionless quantum
master equation with 10 5 realizations (symbols), for W1 = 1.1, W2 = 0.1,
ri/O0 = 10, and F 2 /f2 0 = 1.

provides one of the few open systems which can be solved analytically — for ex-
ample by solving the exact Heisenberg equations of motion (Section 3.6.3) or by
means of the Feynman—Vernon path integral technique (Section 3.6.4). There-
fore, the damped harmonic oscillator is particularly suited to investigate the
parameter regime where the perturbation expansion of the time-convolutionless
master equation yields reliable results beyond the Born—Markov limit. We will
show that the perturbative treatment is in good agreement with the exact so-
lution for sufficiently high temperatures at any coupling strength and for low
temperatures at weak or moderate couplings.

10.2.1 The model and frequency renormalization

We consider the Caldeira—Leggett model for a harmonic oscillator, linearly cou-
pled to a bath of harmonic oscillators (see Chapter 3.6). The Hamiltonian of the
composite system thus takes the form

H = H + HB H1, (10.69)

where

HS = —21 P2 + 1 2 2
(10.70)
E (2 1
1/B = 7-2 (10.71)
Ti

HI= —XEK n x 7, = ® B. (10.72)

476 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

The Hamiltonians Hs and _HB generate the free time evolution of the system
and the bath, respectively, and H1 denotes their interaction. Note that we have
chosen units such that h = 1 and that we have rescaled position and momentum
variables according to XR/Fn and p 1) .0/71, so that the mass of the
oscillator is normalized to one in the subsequent expressions.
The frequency cob which enters the definition of the system Hamiltonian
(10.70) is not the physically observable frequency of the oscillator but the bare
frequency, since the coupling to the environment induces a frequency shift which
depends on the cutoff frequency () (see Section 3.6.2.1). The bare frequency is
connected to the renormalized frequency w0 by the relation
2 2
= 2+ (10.73)

with
2
= Ink 'w 2 2 fcx) dco jw
(w) = 21/Q, (10.74)
n n
0

where J(w) denotes the Ohmic spectral density with a Lorentz-Drude cutoff
function (see eqn (3.392)).
The frequency w on the coupling strength and is of second order
a2 in the coupling. Hence, the bare frequency cob is a function containing terms
of order a2 . On the other hand, the generator 1C(t) = 1C(t;cob) and the inho-
mogeneity (t) = J(t; wb ) (see eqn (9.84)) depend, among other parameters,
on the bare frequency cob . In order to obtain a consistent expansion in terms of
the coupling strength, we thus have to take into account explicitly the potential
renormalization. To second order this yields

/C2(t) = /C2(t; woo) and J2 (t) (t; WO), (10.75)

while the fourth-order contributions read

\ 8 k 2 ( t; wo)
J 4 (t)K4(t; wo) 2wco (10.76)
awo
w 2 0,72(t; wo)
J4(t) = j4(t ; w°) (10.77)
+ 2c:0 awo
where the expressions )C 7 (t) and 1(t) denote the n-th order contributions to
the generator and the inhomogeneity as derived in Sections 9.2.2 and 9.2.4.
Thus, for the computation of the second-order contribution we simply use
the renormalized frequency w0 instead of the bare frequency cob. To fourth order,
we have to consider two terms: The first term is again obtained by replacing the
bare frequency with the renormalized frequency in the usual expressions for K4
and J. The second term is given by differentiating K2 and J2 with respect to
the renormalized frequency.
 THE DAMPED HARMONIC OSCILLATOR 477

10.2.2 Factorizing initial conditions

In this section we investigate the relaxation of an initially factorizing state into
the equilibrium state. To this end, we determine the generator 1C(t) to fourth or-
der using eqns (9.61) and (9.62). After some algebra we obtain in the Schrödinger
representation the following quantum master equation with time-dependent co-
efficients:

d i
Ps(t) = — i [I Is, Ps(t)] — A(t) [X 2 , Ps(t)]
—i\(t) [X, {P, ps (0 }] — Dpp(t) [X, [X, ps(t)]i
+2D px (t) [X, [P, Ps (t)]] - (10.78)

This equation has precisely the same structure as the exact quantum master
equation (Haake and Reibold, 1985; Hu, Paz and Zhang, 1992; Karrlein and
Grabert, 1997) and it can be shown that it is not in Lindblad form (see, e.g.
Sandulescu and Scutaru, 1987).
The coefficient A(t) leads to a time-dependent energy shift, A(t) is the clas-
sical damping term and Dpp(t) and D p x (t) are diffusion terms. The physically
observable frequency of the oscillator is given by

w2 (t) =- 4 + A (t) = c4 + 4 + A (t). (10.79)

As we will see below, in the long-time limit 4 and A (t) tend to compensate
each other, such that the observable frequency w(t) is close to the renormalized
frequency wo.
To second order in the coupling strength, the coefficients entering the master
equation read

t
A (2 ) (t) = ds D(s) cos(w0s), (10.80)
_Ids
o
t
1
A(2) (t) =
2wo fo ds D (s) sin(coos), (10.81)

t
1
=
D (t)
4w0 fo ds D 1 (s) sin(wos) , (10.82)

t
1
D(t) = —2 f ds Di (s) cos(w o s) , (10.83)
O
which is in agreement with the results obtained by Hu, Paz and Zhang. We recall
that the dissipation and the noise kernel D(s) and D 1 (s) were introduced in eqns
478 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

(3.393) and (3.394), respectively. The Markovian limit is obtained in the limit
t —> oo, yielding for Ohmic dissipation

Am _ 2-y12 3
- 92 (10.84)

AM - -
1792 (10.85)

D PX = QkBT
17 Y,
()° 4 - 9 1 1)711 (10.86)
M
2 Q2
1 -2 ,_,
-r "l0 n=_00 (v i ± w 02 )(1Vn
1 ± 9 )'

D PP
meYw°92 w°
=- Ç2 2, w2 cot h ( 2kBT) (10.87)
1- o '
In the high-temperature limit kBT » 9 » wo this yields the coefficients of the
well-known Caldeira-Leggett master equation (3.410), namely
A CL _ (10.88)
Act., = (10.89)
17,
D lyk _ 'YkBT (10.90)
9 '
DPP = 2-ykBT. (10.91)
Note that the coefficient Da is small in comparison to the other coefficients
and may thus be set equal to zero. The explicit expressions for the coefficients
to fourth order can be found in (Breuer, Kappler and Petruccione, 2001). The
energy shift A(t) and the damping coefficient A(t) only depend on the dissipation
kernel D (t) . Hence, these quantities are independent of the temperature of the
reservoir. In Fig. 10.6 we show the time dependence of the physically observable
frequency w(t) and of the coefficient A(t) for moderate coupling (-y = w o , () =
20w0). Both functions decay on a time scale which is of the order of magnitude of
the inverse cutoff frequency 9 -1 and approach a constant value for long times.
In the limit () oc the fourth-order contribution becomes negligible and the
constants are given by the Markovian limits: wp wo and A(t) _), AcL .
The time dependence of the temperature-dependent diffusion coefficients
D pp(t) and Dpx(t) to fourth order is depicted in Fig. 10.7 for three different
temperature regimes. For high temperatures, i.e. k B T » Q, the diffusion coeffi-
cients vary on a time scale which is of the order of the inverse cutoff frequency
(top). For intermediate temperatures, i.e. Q » kBT » wo , the coefficients vary
on a time scale of the order of the thermal correlation time 1/kB T (middle), and
for low temperatures the time dependence of the diffusion coefficients is gov-
erned by the inverse system frequency w o-1 (bottom). From this behaviour, we
expect the perturbation approximation to be in good agreement with the exact
solution for high temperatures if -y < SI, for intermediate temperatures in the
regime -y < kB T and for low temperatures in the regime -y «w0. This point will
 THE DAMPED HARMONIC OSCILLATOR 479

2Q —I

0 0.1 0.2 0.3 0.4 0.5

wot

2Q —I

0 0.1 0.2 0.3 0.4 0.5

Loot

FIG. 10.6. Time dependence of the physically observable frequency w(t) and
of the damping coefficient A(t) of the harmonic oscillator to fourth order in
the coupling strength. The parameters are: -y = wo, Q = 20w0 .

be discussed in more detail in Section 10.2.3, where we compute the stationary

density matrix of the reduced system.
Let us now consider the relative error of the coefficients of the quantum
master equation, which we define for A(t) as

A (2 n) (t) — A(t)
e r(2 n) = lim (10.92)
t—f oc A(t) ,

and in a similar way for the other coefficients. This type of error was discussed
in Section 9.2.5. By using the expansion of the coefficients to fourth order we
obtain the following estimates for these errors:
480 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

PP
2Q-1

k B T >> Q » wo kB T » Q » w o

0 0.1 0.2 0.3 0.4 05 0 0.1 0.2 0.3 0.4 05

t

Q» kBT >> wo Q» k B T >> c.,)0

1
27rkBT

1
27r kB T

DIA

0.02 0.04 0.06 0.08 01 0.02 0.04 0.06 0.08 01

wot wot

Q >> wo » k B T » wo » kB T

2wo

2wo
DA

1 2 3 0 1 2 3
Wot Wo t

FIG. 10.7. Time dependence of the diffusion terms D p p (t) and Dpx(t) to fourth
order in the coupling strength. We distinguish three parameter regimes.
High temperatures, intermediate coupling (top): -y = wo , = 20w0,
kBT = 1000w0. Intermediate temperatures, intermediate coupling (middle):
y = Ct)0 = 200 coo, kBT = 10 wo. Low temperatures, weak coupling (bot-
-

tom): -y = 0.1w0 , 12 = 20w0 , kB T = 0.01 wo•

AO) (t) A( 2)(t) -(4) _

e-(2) = urn
-

r er — (10.93)
t—>00 A ( 2 ) (t)

and similar expressions can be found for the other coefficients. These estimates
..(2)
are compared with the actual errors in Fig. 10.8. The figure shows that er
provides a very good estimate of the actual error eP ) of the perturbation ex-
pansion. This is not surprising since this estimate is based on the fourth-order
contribution. On the other hand, the estimate el) does not involve higher-order
terms but only relies on the second and fourth-order contribution. Nevertheless
 THE DAMPED HARMONIC OSCILLATOR 481

1 1

lini t „(.4,p2 (t) — Log

0.01 0.01

0.0001 0.0001

le-06 le-06

le-08 le-08
0.01 1 0.01 0.1 1
7/W0
1 1

0.01 0.01

0.0001 0.0001

le-06 le-06

le-08 le-08
0.01 0.1 1 0.01 0.1 1
'/wo ')//wo

FIG. 10.8. Exact relative errors eP ) and e 4) and their estimates 6 2) and
to second and fourth order for the coefficients A(t), w(t) 2 — w D p p (t) and ,

D p (t)

it provides a good means for getting the order of magnitude of the actual error
which is enough for most applications. Thus, except for the second order, the
error of the perturbation expansion can be reliably estimated without computing
higher-order terms of the expansion.

10.2.3 The stationary state

In this section we investigate the stationary state of the time-convolutionless
master equation in order to find the parameter regime in which the perturba-
tion expansion yields reliable numerical results. Since the stationary state of the
harmonic oscillator is Gaussian, it is completely determined by the first and sec-
ond moments of the observables X and P. The dynamics of these moments is
governed by the adjoint master equation

= j ([11s OD t A(t) ( [X 2, (-9 ] )t

+iA(t) ({P, [X, t — D pp(t) ([X, [X, O ]] ) t
+2D px (t) , [P, t, (10.94)
482 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

where 0 is some arbitrary system operator and ( (9 ) t denotes the time-dependent

expectation value trs{Ops(t)} of the observable 0 with respect to the reduced
density matrix ps(t).
For the observables X and P eqn (10.94) yields the classical equations of
motion of a damped harmonic oscillator, namely

d
(X)t = (P) (10.95)
dt
d
4(0 (X)t — 2A(t)(P)t, (10.96)

and for the second moments we obtain

({X, P}), (10.97)

d (p2)t
- (t) ( { X, P}) t — 4A(t) (P2 ) + 2D p p (t) (10.98)
dt
d
= —24(t)(X 2 ) t + 2 (P2 )1,
—2A(t)({X,P}) 1 4Dp.v (t). (10.99)
The stationary solution of these equations of motion reads (S:andulescu and Scu-
taru, 1987)

(10.100)

—
1
(x2 ) = tHilloo 24(0 [ D pp (t)
I_ A(t) + 4D p x (01 (10.101)
D pp(t)
(P2 ) (10.102)
t—> DO 2A(t)
In the weak coupling limit, i.e. in the limit 7 0, the stationary state can be
computed by inserting the Markovian coefficients defined in eqns (10.84)—(10.87).
which lead to the expressions

(X 2 = —1
2wo coth ( 21:T ) , (10.103)

(P2 ) =—
w2° coth ( 2ZT ) . (10.104)

The variances AX 2 and AP2 are (up to a scaling factor coo ) identical and thus.
the state of the harmonic oscillator is not squeezed. In the high-temperature
regime the variances are proportional to the temperature, i.e. the oscillator be-
haves like a classical oscillator. On the other hand, in the low-temperature regime
the variances are bound by the Heisenberg uncertainty relation and the oscillator
behaves quantum mechanically.
 THE DAMPED HARMONIC OSCILLATOR 483

For -y > 0 we can approximate the stationary solution by using the expres-
sions for the coefficients of the master equation in fourth order. This yields
corrections to the second moments (X2 ) and (P2 ) which are linear in 2/. In Fig.
10.9 we compare the approximated with the exact variances for -y = 0.25 coo and
-y = 0.5 wo in the low-temperature regime. Both solutions show qualitatively the
same behaviour: The variance in the position decreases, while the variance in the
momentum increases for increasing temperatures — the state of the oscillator is
squeezed in this parameter regime. Figure 10.9 also indicates that the approxi-
mation of the master equation to fourth order yields reliable numerical results
for -y < c.03 /2.
A more detailed quantitative analysis of the error of the approximation is
depicted in Fig. 10.10. For a fixed value of the cutoff frequency 0 this figure shows
the regions in the parameter space where the relative error in the variance of the
momentum is less than 5% and 1%, respectively. Obviously, the perturbation
expansion is in good agreement with the exact results for -y < kB772 in the
high-temperature regime, and for -y < wo /2 in the limit T .— 0. Numerical
investigations show that for 2, < 0/2 this behaviour may be summarized by the
condition
ado
< en(Q,er) (10.105)
wa + (kBT) 2 —
where the constant en, depends on the cutoff frequency 0, the desired accuracy
e r , and the order n of the approximation, but is independent of the temperature.
For 0 = 20 coo and er = 5% we have, for example, e2 = 0.042 and c4 = 0.84,
while for er = 1% one gets e2 = 0.0068 and c4 = 0.25. In Fig. 10.10 we have
also marked the boundary of the parameter regime for which eqn (10.105) holds,
which is in very good agreement with that region where the relative error is, in
fact, less than the prescribed value of er .
The fact that the perturbation expansion also yields numerically reliable re-
sults in the low-temperature regime seems to contradict the fact that the thermal
correlation time TB r•i 1/kBT becomes infinite in the limit T .— 0. On the other
hand, as discussed in Section 10.2.2, the time dependence of the temperature-
dependent diffusion coefficients is dominated by the time scale wo-1 . Thus, the
memory time in the low-temperature regime is the inverse system frequency and
not the thermal correlation time. In fact, as we saw in Section 3.4 a perturbative
treatment is extremely accurate in the quantum optical regime even in the limit
T -- 0. This explains why the Markovian master equation is particularly useful
in that regime and why it provides an approximation which is uniform in T and
holds even in the vacuum optical case (T = 0).

10.2.4 Non-factorizing initial conditions

In this section we want to determine the equilibrium position autocorrelation
function (X(t)X) which is defined by

(X (t)X) E tr { e ifft Xe —iHt Xpeq } , (10.106)

484 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

0
0 0.5 1 1.5 2
kBT/wo
3 T

/ 17 2\ exact
WO Y1 / TCL4 xxxx

0
0 0.5 1 1.5 2
kBT/wo

FIG. 10.9. Low-temperature behaviour of the stationary variances (X 2 ) and

(P2 ) for -y = 0.25 wo (top) and for -y --- 0.5w0 (bottom). `TCL 4' denotes
the solution of the quantum master equation to fourth order and 'exact' the
solution given in (Grabert, Schramm and Ingold, 1988). The cutoff is chosen
to be Q = 20w0 .

where the trace is taken over the system and reservoir. This definition can be
transformed into the interaction picture,

(X(t)X) = tr {X(t)V(t)}, (10.107)

where X(t) = exp(iHst)X exp(—iHst) is the interaction picture position opera-

tor. The operator V(t) introduced above is a solution of the equation of motion

d
— V (t) = — 0I1- (t) , V (t)] (10.108)
dt
 THE DAMPED HARMONIC OSCILLATOR 485

40 50 40 50

10 10

0.1

e-
0.01

0.001 0.001

0.0001 0.0001
0 1 5 0 2 3 4 5
kB T/wo

FIG. 10.10. The regions in parameter space, where the relative error of the
stationary value of (P2 ) is less than 5% (left side) and less than 1% (right
side). TCL 2 (TCL 4) denotes the solution of the time-convolutionless master
equation to second (fourth) order. The crosses mark the boundaries of the
regions where eqn (10.105) holds. The bottom shows an enlarged section of
the low-temperature and weak-coupling regime. The cutoff is chosen to be
20 wo .

corresponding to the initial condition

V(0) = XPe q E-- APeq• (10.109)

In general, this initial value neither provides a true density matrix nor does
it represent a factorizing state. However, it does belong to the class of initial
conditions (9.69) discussed in Section 9.2.4.
Since the operator X only acts in the Hilbert space of the open system it
commutes with the projection operator P and we can write the position auto-
correlation function as follows,

(X(t)X) = tr {X(t)PV(t)} . (10.110)

The equation of motion for the correlation function is now obtained by differ-
entiating both sides of this equation with respect to t and making use of eqns
(9.33) and (9.84),
486 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

d
— (X (t)X) = i tr { [Hs , X (t)]PV (t)} (10.111)
dt
+tr -(X(t)k(t)PV(t)} + tr {X(t)J(t)Pp eq f .

In a similar way we can derive an equation of motion for the momentum-position

correlation function, namely
d
— (P(t)X) = i tr {[H s , P(t)1PV (Of (10.112)
dt
+tr {P(t)K(t)PV(t)} + tr {P (t) J (t)P Peg } .
Inserting the expression for k(t) for the harmonic oscillator derived in Section
10.2.2 we obtain a closed set of differential equations for (X(t)X) and (P (t)X):
d
(P (t)X) + Ix x (t), (10.113)
1t- (X (t)X) =
d
— (P(t)X) = -4(t)(X(t)X) - 2A(t)(P(t)X) + Ipx (t), (10.114)
dt
where the inhomogeneities are defined as
/xx (t) = tr {X(t),7(t)Pp eci l , (10.115)
1-p x (t) = tr {P(t)J(t)Ppec i l • (10.116)
The equation of motion for the correlation function (P(t)X) is, apart from the
inhomogeneity, identical to the equation of motion for the observables X and P
(see eqn (10.96)). This is the basic assertion of the quantum regression theorem.
Thus, in the weak coupling limit, where the effect of the inhomogeneity can be
neglected, the quantum regression theorem holds to the same level of accuracy.
as the master equation. For stronger couplings the equation of motion has to
be supplemented by the inhomogeneous term. It is important to note that this
result is not in contradiction to the reported failure of the quantum regression
theorem (Grabert, 1982; Talkner, 1995; Ford and O'Connell, 1996), since the
description of the reduced dynamics by means of a convolutionless equation of
motion does not constitute a quantum Markov process possessing the semigroup
property.
By using the definition of the super-operator 1(t) (see eqn (9.35)) it can be
shown that the inhomogeneous term /xx (t) vanishes to all orders in the coupling
strength. This can also be seen directly from the exact Heisenberg equation of
motion 5CH(t) = PH(t) which leads to /xx (t) E 0 in eqn (10.113). The term
/px (t) is evaluated by making use of the perturbation expansion of the super-
operator J(t) given in Section 9.2.4. To second order in the coupling strength
we find
o o
1
I (X 2) f ds D(t - s) cos(wo s) - f ds D1 (t - s) sin(c.00 s). (10.117)
-00 - 00

The fourth-order contribution PA is given by Breuer, Kappler and Petruccione

(2001).
 THE DAMPED HARMONIC OSCILLATOR 487

2-}{2 (X ) 27Q(X 2 )

o 2 3
coot

o
0.1 0.2 0.3 0.4 05 O 0.1 0.2 0.3 0.4 05
LA) at
(x ,

exact
xxx TCL 4

o (

10 20 30 40 50 0 2 4 6 8 10
coot coot

FIG. 10.11. Time dependence of the inhomogeneous term Ipx (t) (top) and the
real part of the position autocorrelation function (Xii(t)X) (bottom) for low
temperatures and weak couplings, i.e. kBT = 0.01 coo and -y = 0.1 wo (left)
and for high temperatures and intermediate couplings, i.e. kBT = 1000 C•00
and 7 = (4) 0 (right). The cutoff is chosen to be 0 = 20w 0 . The inset shows
the long-time behaviour of Ipx (t) for low temperatures after the initial jolt
has decayed (note the different units).

The time dependence of ipx (t) is depicted in Fig. 10.11. In the low-temper-
ature regime, there are two different contributions to Ipx (t): For small times,
i.e. for times of the order of the inverse cutoff frequency, Ipx (t) takes large
positive values. For intermediate times, i.e. for times of the order of the inverse
system frequency, Ipx (t) is negative. In the limit t Do the inhomogeneous term
/px(t) vanishes and the master equation becomes homogeneous. In contrast to
this behaviour, /px (t) simply decays exponentially on a time scale 0 -1 in the
high-temperature regime.
This behaviour has important consequences for the time evolution of the po-
sition autocorrelation function. To see this, we consider the equation of motion
for (P(t)X), eqn (10.114). The first term, —w p2 (t)(X(t)X), gives rise to a large
negative contribution in the short-time behaviour due to the initial jolts in the
physically observable frequency w(t) (see Fig. 10.6), which would lead to a fast
decay of (P(t)X) on a time scale 0'. However, this contribution is compen-
sated by the inhomogeneous term Ipx (t) which takes large positive values for
all temperatures. Thus the somewhat artificial initial jolts, which are induced by
488 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

the short-time behaviour of the generator K(t), are compensated by the inho-
mogeneity J(t).
Figure 10.11 also shows the real part of the autocorrelation function (X(t)X)
in the underdamped and in the overdamped regime. The lines indicate the ex-
act solution as given by Grabert, Schramm and Ingo ld (1988), and the symbols
denote the solution of the perturbation expansion of the equations of motion
(10.114). The agreement of both expressions is very good for the considered pa-
rameters. A systematic investigation of the range of validity of the perturbation
expansion to second and fourth order is depicted in Fig. 10.12. This figures show
the regions in parameter space where the relative error
maxt>0 fre .(t) — C(t)J} (10.118)

for a certain order of approximation n is less than 5% and 1%, respectively.

The function Gx (t) denotes the exact autocorrelation function and Cri (t) its
nth order approximation. In this figure we have also included the boundaries of
the regions where eqn (10.105) holds for c2 = 0.0055, c4 = 0.3 (e r = 5%) and
c2 = 0.0013, c4 = 0.15 (e r = 1%). As in the case of the relative error of the sta-
tionary variance (P 2 ), this gives a good estimate of the error of the perturbation
expansion. Note, however, that the quality of this estimate also depends on the
ratio of the coupling strength ty to the cutoff frequency 0, since the approxima-
tion of the physically observable frequency w(t) and of the classical damping
coefficient A(t) depend on this ratio.

10.2.5 Disregarding the inhomogeneity

The explicit computation of the inhomogeneity J(t) for non-factorizing initial
conditions can become rather tedious since it consists of more terms than the
generator 1C(t). Thus, there naturally arises the question if this effort is really
necessary or if it is possible to disregard the inhomogeneity. In this section we
will briefly discuss two proposals of this kind which yield the correct results in
the weak-coupling limit and show their failure for certain parameters.
The first possibility we want to investigate is to simply ignore the correlations
between the system and the environment and to assume a factorizing initial state.
Thus, the equation of motion for the correlation function (P(t)X) would take
the form
d
(13 (t)X) = —c.4(t)(X(t)X) — 2A(t)(P(t)X). (10.119)
dt
However, as we discussed in Section 10.2.4 the short-time dynamics of the phys-
ically observable frequency cup is dominated by a large initial jolt which leads to
an enhanced decay of the autocorrelation function. This behaviour is depicted
in Fig. 10.13 (Trop 1').
The second proposal we want to discuss avoids this difficulty. In this approach
we also ignore the initial correlations, but in contrast to the first approach the
 THE DAMPED HARMONIC OSCILLATOR 489

40 50

FIG. 10.12. The regions in parameter space where the relative error of the
position autocorrelation function (X(t)X) is less than 5% (left side) and
less than 1% (right side). TCL 2 (TCL 4) denotes the solution of the
time-convolutionless master equation to second (fourth) order. The crosses
mark the boundaries of the regions where eqn (10.105) holds. The cutoff is
chosen to be 1 = 200w 0 .

time evolution of the master equation is generated by the stationary value of

K(t). Thus the equation of motion for (13 (0X) would take the form
d
— (P(t)X) = —4,(X(t)X) — 2 . (P(t)X), (10.120)
dt
where
cDp = lim w(t), A = lim A(t).
t—›.00
(10.121)

As demonstrated in Fig. 10.13 (see 'Prop 2') this approach also leads to significant
deviations from a calculation which takes into account the inhomogeneity. These
deviations are due to the negative contribution to ipp(t) (see the inset of Fig.
10.11).
Thus, it is in general inevitable to explicitly include the inhomogeneity in the
case of non-factorizing initial states. An attempt to approximate the initial state
by a factorizing state can lead to additional errors, which are of the order of the
corrections one makes by expanding the generator to higher orders.
490 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

Loot

FIG. 10.13. Approximation of the correlation function (P(t)X) determined by

neglecting the inhomogeneity (see text for details). The parameters read:
kBT/c.o0 = 0.0 1 , 7/w0 = 0.1, and S2 = 20w0 .

10.3 The spin boson system

-

The investigation of the spin-boson system, i.e. a two-level system interacting

with a bath of harmonic oscillators, is of particular interest in the theory of
open quantum systems. First of all many physically interesting systems can be
approximated by a two-level system in the low-temperature limit by neglecting
higher excitations. This point has been discussed in detail by Leggett et al.
(1987) in the context of the problem of tunnelling between two potential wells.
Another reason for the wide interest in the spin-boson system is that it provides
an excellent model for the investigation of certain approximation methods.

10.3.1 Microscopic model

The spin-boson system describes a two-level system which is linearly coupled
to a bath of harmonic oscillators. The essential difference to the model treated
in Section 10.1 is that the system—reservoir coupling is not treated within the
rotating wave approximation. The Hamiltonian of the composite system thus
takes the form

H = Hs + HB + 111, (10.122)
1
Hs = —woo - z, (10.123)
2
HB ,E (

1
± 1 Tnriwr2i x n2 ,
(10.124)
ri
2m,-, 2
1
Hi- = — (ix 0 B, (10.125)
2
 THE SPIN-BOSON SYSTEM 491

where Hs and Hg generate the free time evolution of the system and the bath,
respectively, and III denotes their interaction. As in the case of the harmonic
oscillator (see Section 10.2) the reduced system dynamics is characterized by the
dissipation kernel D(t) = i[B(t), _131 and by the noise kernel D i (t) = ({BM, B})
(see eqns (3.385) and (3.386), respectively). The spectral density J(w) of the
model is again taken to be of the Lorentz-Drude form,
1, w 02
J(w) = 71 Wo 112 (10.126)
+ w2 '
With this definition of the damping constant 7 the dimension of D(t) and D1 (t)
is equal to the square of an energy as it should be for an interaction Hamiltonian
of the form (10.125).

10.3.2 Relaxation of an initially factorizing state

In this section we compute the perturbation expansion of the generator of the
quantum master equation for the spin-boson system and determine its stationary
solution. The quantum master equation for the reduced density matrix is most
conveniently written in terms of the Bloch vector (6(t)) (see eqn (3.223)) as
follows,
d
— (6(t)) = A(t)(6(t)) + 2(t), (10.127)
dt
where A(t) is a time-dependent 3 x 3 matrix and the inhomogeneity g(t) is
a 3-vector. These quantities can be computed by using the expressions for the
generator kW given in Section 9.2.2. In the Schrödinger picture the perturbation
expansion of A(t) and E (t) takes the form
( 0 - wo 0 0 )
A(t) = w o + a x (t) a(t) 0 , ( t) = 0 (10.128)
0 0 azz (t) bz (t)
To second order in the coupling strength we obtain
t
1
a yx
(2) (t) = — f ds Di(s)sin(wos), (10.129)
2
0
t
a(t) = a (z2z) (t) = --1 Ids Di(s)cos(wos), (10.130)
2
o
t
1
b 2 (t) = - - f ds D(s) sin(wo s), (10.131)
4
o
which is similar to the coefficients of the quantum master equation of the har-
monic oscillator (see Section 10.2.2). Note however that the term A(t) which
492 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

0.1 0.2 0.3 0.4 05 0.1 0.2 0.3 0.4 05

wo t wo t

bm

0 0.1 0.2 0.3 0.4 05 0 0.1 0.2 0.3 0.4 05

Wot Wot

FIG. 10.14. Time dependence of the dephasing coefficients a(t) and a(t) and
of the relaxation coefficients a„(t) and b(t) to fourth order in the coupling
strength in the intermediate-coupling regime, i.e. 7 = 0.3 wo , l = 20 coo ,
kBT = 10w0.

leads to a large cutoff dependent frequency shift is not present in this model.
Hence, there is no need for an explicit frequency renormalization procedure in
this case. The fourth-order contributions can be found in (Breuer, Kappler and
Petruccione, 2001).
The coefficients ayx (t) and a yy (t) are responsible for the dephasing of the
two-level system and the coefficients a zz (t) and b z (0 describe the relaxation to
the equilibrium state. The time dependence of these coefficients to fourth order
in the coupling is depicted in Fig. 10.14 for an intermediate-coupling strength.
Note that the stationary values of ayx (t) and b (t) show a significant deviation
from the Markovian limit whereas the approximation of a(t) and a zz (t) to
second order is quite accurate.
In the Markovian regime and in the high-temperature limit kB T >> 12 »
the matrix A and the inhomogeneity g can be approximated by
0 —wo
Am = I
(
wo + rykoBT -ykBT (1
0 ) ,

gM (10.132)
wo
7kBT
0 0
wo
It is important to note that in this limit the relaxation rate as well as the de-
phasing rate are proportional to the temperature, in contrast to the harmonic
 THE SPIN-BOSON SYSTEM 493

FIG. 10.15. Temperature dependence of the stationary value of (az)s for

different coupling strengths. The population of the excited state is
Pu i = (1 + (a,),)/2.

oscillator where the temperature of the environment only has an effect on the
diffusive behaviour of the system but not on dissipation.
The stationary state of the two-level system is readily obtained by making
use of eqns (10.127) and (10.128),

2b(t)
( a x), = (ay ), ,-- 0, (az)s = — lim (10.133)
t—>co a zz (t) .

The temperature dependence of (az ), is depicted in Fig. 10.15. For 7 = 0 we

obtain the thermodynamic limit in which the reduced density matrix is deter-
mined by the Boltzmann distribution. For 7 > 0 we find deviations from the
Boltzmann distribution. In particular the population of the excited state does
not vanish in the limit T 0 in contrast to the predictions of a calculation to
second order. These results are qualitatively similar to the results obtained by
De Raedt and De Raedt (1984).
The order of magnitude of the relative error of the stationary state can be
determined by estimating the relative error of the coefficients azz and bz using
the procedure outlined in Section 9.2.5. Thus the relative error to second and
fourth order can be estimated by
494 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

10

cq
0

0.01
0 10 20 30 40 50
kBT/wo
10

, • • • .• • •

0. 1

0.01
0 10 20 30 40 50
kBT/wo

FIG. 10.16. The regions in parameter space where the estimated relative errors
éP ) and '6 4) of the stationary value (o-z ), are less than 5% (top) and less
than 1% (bottom). The cutoff is chosen to be SI = 20w0 .

a2 (t) — a 2( ) (t) b(z 4) (t) — 14 2) (t)

6 ,(2 ) lim (10.134)
t—+D3 a (2) (t) (2)
bz (t)
2
(t) — cti29(t) b(z 4) (t) — 14 2) (t)
(2) . (10.135)
a(z 2z) (t) bz (t)
2 )

As in the case of the harmonic oscillator, we expect that the estimate for the
relative error to second order provides a quantitatively correct result, whereas
the error estimate to fourth order only yields the correct order of magnitude.
In Fig. 10.16 we show the regions in parameter space where the error esti-
mates are lower than 5% and 1%, respectively. For low and intermediate tem-
peratures the behaviour of the error estimate is essentially the same as for the
harmonic oscillator. However, in the high-temperature limit kBT >> fl the max-
imum coupling strength is proportional to 1/11 . This is due to the fact that the
 THE SPIN-BOSON SYSTEM 495

relaxation coefficient a„ (t) and the dephasing coefficient a(t) are proportional
to the temperature in this limit. Thus, when we consider the high-temperature
limit we have to fix the product IT in order to obtain meaningful results (Chang
and Skinner, 1993). This behaviour is very different from that of the harmonic
oscillator, where the temperature has an influence only on diffusion and not on
damping.

10.3.3 Equilibrium correlation functions

In this section we compute the equilibrium correlation function (ax (t) a- x ) To this
end, we consider the equations of motion for the correlation functions (ax (t)o-x )
and (ay (t)o-x ), which can be derived from the adjoint master equation. Following
the strategy described in Section 10.2.4 we obtain
d
wo( 0- y x) + x x (t), (10.136)
Tt (Grx(t)a x)

dt ça ylt
= ayx(t)i((T x(0 0- x) + ay y (t) (ay (t)u x) + y x (t), (10.137)
where the inhomogeneities are defined as
/xx (t) = tr {a x (t) J (OP peg } , (10.138)
Iy x (t) = tr fay (t).1(t)PPecil • (10.139)
Using the appropriate expansion for J(t)'P peq (see Section 9.2.4) we obtain
-Txx (t) E 0 to all orders in the coupling strength. This is, of course, an im-
mediate consequence of the exact Heisenberg equation of motion (3rx = - wogy •
The second-order contribution to the inhomogeneity x (t) is found to be

i
IY(2)X = 2 (a. z ) f ds D (t - s) cos(wos)- 2 f ds D(t - s) sinp o s). (10.140)
-00 -co
The fourth-order contribution to Iy .,y may be found in (Breuer, Kappler and
Petruccione, 2001).
Figure 10.17 shows the time dependence of the correlation function (a(t)o- )
in the low and high-temperature regimes. The Markovian approximation yields
quantitatively good results for the real part of the correlation functions, whereas
the fourth-order contribution introduces significant corrections in the imaginary
part. Note that in the low-temperature regime the parameters are chosen in
such a way that the system is underdamped, whereas in the high-temperature
regime we have chosen a coupling strength for which the system is overdamped.
In the latter case, the Markovian approximation reveals that the dynamics of the
correlation function is hi-exponential where the fast and slow decay rates can be
approximated by
3
-ykBT WO
rfast = rsiow (10.141)
WO -ykBT
496 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

0.25 0.25

TCL 4 —
Markov xxxx

-0.25 o
0 10 20 30 40 50 o 5 15 20
wo t

TCL 4 —
Markov xxxx

5 10 15 20
t

FIG. 10.17. Real and imaginary parts of the correlation function ((T x (t)o-x ) in
the low-temperature regime, i.e. kBT/wo = 0.1, 'y = 0.5w0 (left) and in the
high-temperature regime, i.e. kBT/wo = 50, ey = 0.2(.00 (right). The cutoff is
= 20w0 .

10.3.4 Transition from coherent to incoherent motion

Especially in the context of macroscopic quantum coherence the cross-over from
coherent to incoherent motion is very important: quantum coherence can only
be observed for a coupling strength which is well below the critical damping -y,
where this cross-over takes place.
As pointed out by Egger, Grabert and Weiss (1997) the critical damping
strength -y, depends on the criterion used to define coherent motion. Since we
are only interested in the order of magnitude of the critical damping strength we
take a simple criterion which is independent of the initial preparation, i.e. which
is valid for the expectation value (a(t)) as well as for the correlation function
(ax (t)o -x (0)). This criterion is based on the eigenvalues of the matrix

A —wo (10.142)
t_>D0 (wo + ayx (t) a yy (t)

which determines the long-time dynamics of (o-x (t)) and (o-x (t)ax (0)) (see eqns
(10.127) and (10.136)). If both eigenvalues are real, then the quantities under
consideration decay exponentially in the long-time limit whereas they oscillate
 REFERENCES 497

TCL 4
- - - TCL 2
1.5 -

0.5

1 2 3 4 5
kBT/wo

FIG. 10.18. The cross-over from coherent to incoherent motion for the pertur-
bative approximation to second and fourth order. The cutoff is chosen to be
0 = 20w0 .

if both eigenvalues are complex. Thus we define the critical damping ry, through
the condition

lim [a2 (t) - 4w0(wo + ayx(t))1 = 0. (10.143)

Figure 10.18 shows the critical damping strength for the approximations to sec-
ond and fourth order as a function of temperature. In the limit T -> 0 the
estimated relative error of 44x) (t) is less than 1% whereas the estimated error of
4j (t) is 18%. The relative error of -y, is expected to be of the same order of
magnitude.

References
Breuer, H. P., Kappler, B. and Petruccione, F. (1999). Stochastic wave-function
method for non-Markovian quantum master equations. Phys. Rev., A59, 1633-
1643.
Breuer, H. P., Kappler, B. and Petruccione, F. (2001). The time-convolutionless
projection operator technique in the quantum theory of dissipation and deco-
herence. Ann. Phys. (N. Y.), 291, 36-70.
Chang, T.-M. and Skinner, J. L. (1993). Non-Markovian population and phase
relaxation and absorption lineshape for a two-level system strongly coupled to
a harmonic quantum bath. Physica A, 193, 483-539.
De Raedt, B. and De Raedt, H. (1984). Thermodynamics of a two-level system
coupled to bosons. Phys. Rev., B29, 5325-5336.
Egger, R., Grabert, H. and Weiss, U. (1997). Crossover from coherent to inco-
herent dynamics in damped quantum systems. Phys. Rev., E55, R3809-R3812.
498 NON-MARKOVIAN DYNAMICS IN PHYSICAL SYSTEMS

Ford, G. W. and O'Connell, R. F. (1996). There is no quantum regression

theorem. Phys. Rev. Lett., 77, 798-801.
Garraway, B. M. (1997). Nonperturbative decay of an atomic system in a cavity.
Phys. Rev., A55, 2290-2303.
Grabert, H. (1982). Nonlinear relaxation and fluctuations of damped quantum
systems. Z. Phys., B49, 161-172.
Grabert, H., Schramm, P. and Ingold, G.-L. (1988). Quantum Brownian motion:
The functional integral approach. Phys. Rep., 168, 115-207.
Haake, F. and Reibold, R. (1985). Strong damping and low-temperature anoma-
lies for the harmonic oscillator. Phys. Rev., A32, 2462-2475.
Hu, B. L., Paz, J. P. and Zhang, Y. (1992). Quantum Brownian motion in
a general environment: Exact master equation with nonlocal dissipation and
colored noise. Phys. Rev., D45, 2843-2861.
Karrlein, R. and Grabert, H. (1997). Exact time evolution and master equations
for the damped harmonic oscillator. Phys. Rev., E55, 153-164.
Leggett, A. J., Chakravarty, S., Dorsey, A. T., Fisher, M. P. A., Garg, A. and
Zwerger, W. (1987). Dynamics of the dissipative two-state system. Rev. Mod.
Phys., 59, 1-85.
SAndulescu, A. and Scutaru, H. (1987). Open quantum systems and the damp-
ing of collective modes in deep ineleastic collisions. Ann. Phys. (N. Y.), 173,
277-317.
Talkner, P. (1995). The failure of the quantum regression hypothesis. Ann.
Phys. (N. Y.), 167, 390-436.
Weisskopf, V. and Wigner, E. (1930). Berechnung der natiirlichen Linienbreite
auf Grund der Diracschen Lichttheorie. Z. Phys., 63, 54-73.
 Part V
Relativistic quantum processes
 11
MEASUREMENTS IN RELATIVISTIC QUANTUM
MECHANICS

In the relativistic domain the quantum theory of measurement exhibits several

new and interesting features which are not encountered in the non-relativistic
theory (Landau and Peierls, 1931; Bohr and Rosenfeld, 1933; Bloch, 1967). The
present chapter is devoted to a systematic study of measurement and state re-
duction in relativistic quantum mechanics. An elementary introduction to the
problem may be found in (Breuer and Petruccione, 1999).
The basic idea underlying the presentation of this chapter is to consider the
state of a quantum system as a functional on the set of spacelike hypersurfaces
in Minkowski space. This idea has been developed in the early days of relativistic
quantum field theory by Dirac, Schwinger and Tomonaga with the aim to give a
manifest covariant formulation for the time-evolution equation of the state vec-
tor, known as the Schwinger—Tomonaga equation. The same concept turns out
to be very useful in a systematic treatment of measurement and state reduction
in relativistic quantum mechanics, as has been pointed out in a series of papers
by Aharonov and Albert (1980, 1981, 1984a, 1984b). In fact, a covariant state
reduction postulate is obtained if one regards the state reduction as taking place
along those spacelike hypersurfaces which cross the classical chance event pro-
vided by the readout of a local measurement. As is demonstrated in the present
chapter, this concept leads to a consistent formulation for the state evolution of
a quantum system which is conditioned on a single or on multiple local mea-
surements arbitrarily distributed in space-time, and for the corresponding joint
probability distributions containing arbitrary local and non-local quantum corre-
lations. We also develop in this context the relativistic formulation of continuous
measurements in terms of a covariant PDP for the state vector.
An important point which must be taken into account in the relativistic
quantum theory of measurement is the possibility of non-local measurements.
It has been shown by Aharonov, Albert and Vaidman (1986) that it is possible
to carry out the measurement of certain non-local quantities without measur-
ing local properties of the system. Such measurements are performed as indirect
measurements invoking entangled quantum probes. We shall discuss here in de-
tail the measurement of non-local observables and the verification of non-local,
entangled quantum states.
As will be seen the possibility of the measurement of non-local quantities is
strongly restricted by the causality principle. These restrictions can be formu-
lated with the help of a theorem due to Popescu and Vaidman (1994) which
502 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

states, essentially, that non-local measurements must necessarily erase local in-
formation in order to be compatible with the causality principle. The present
chapter contains a proof of this theorem and a detailed discussion of its physical
implications. In particular, one can deduce that only a certain class of non-local
operators allows a quantum non-demolition measurement which is in agreement
with the von Neumann—Liiders postulate.
The quantum theory of measurement thus leads to important consequences
for the notions of observables and states if it is combined with the require-
ments of special relativity. Several examples for the application of the general
theory will be presented. Furthermore, the preparation of states, the notion of
exchange measurements and the instantaneous transfer of a coherent quantum
state, known as quantum teleportation, will be discussed.

11.1 The Schwinger Tomonaga equation

—

In the following we write x 4 = (x ° ,±÷) for the space-time coordinates of a point

x in Minkowski space R4 , where p, runs from 0 to 3. The first component x ° = t
is the time coordinate and I = (x', x 2 , x 3 ) denotes the space coordinates. We
choose units such that h = c = 1, where c is the speed of light. The Lorentz
invariant inner product of two 4-vectors x and y is defined by

xy = x/iy t, = gi„x4 yv = x ° y ° — x' • 94, (11.1)

where gii, denotes the metric tensor, and repeated indices are summed.
In the interaction picture the time evolution of the state vector I CO) is
governed by the Schrödinger equation

a in-,( 00(t)),
at 1 `11(t)) = —
(11.2)

where

Hi (t) = f d 3 x 7-1(t, ±4) (11.3)

is the interaction picture Hamiltonian and W(x) = 71(t, ±*) denotes the Hamil-
tonian density of the theory. For simplicity we investigate here theories without
derivative couplings and the Hamiltonian density is assumed to transform as a
scalar under Lorentz transformations.

11.1.1 States as functionals of spacelike hypersurfaces

Given a fixed coordinate system, I T(t)) characterizes the state of a quantum me-
chanical system at a fixed time x° = t and thus allows the evaluation of the ex-
pectation values for all observables which are localized on the three-dimensional
hypersurface given by x ° = t = constant. The relativistically invariant general-
ization of this concept is that of a state vector which is associated with a general,
three-dimensional spacelike hypersurface cr. Such a hypersurface is defined to be
 THE SCHWINGER-TOMONAGA EQUATION 503

a three-dimensional manifold in Minkowski space which extends to infinity in all

spacelike directions and which has at each point x E a a unit, timelike normal
vector n» (x) satisfying

np,(x)n. a (x) = 1, n° (x) > 1. (11.4)

The state vector then becomes a functional

l ip) = IT(a)) (11.5)

on the space of such hypersurfaces. The same holds for the density matrix of the
system which again yields a functional

P = P(a) (11.6)

on the set of spacelike hypersurfaces.

The corresponding generalization of eqn (11.2) is the Schwinger-Tomonaga
equation for the state vector (Tomonaga, 1946, 1947; Schwinger, 1948; Schweber,
1961),

SIT ( a )) = —a-0441(u)), (11.7)

60- (x)

or, else, for the density matrix,

= —7, [7-1 (x), P(a)] • (11.8)

The Schwinger-Tomonaga equation is a functional differential equation. It may

be considered as a differential equation in a continuous family of time variables.
Each point x E a may be represented in a special coordinate system as xi' =
(x ° (i), Y), such that each space point Y has its own time variable x ° = z 0 (i).
Equations (11.7) and (11.8) involve a variation 60- of the hypersurface a in which
these time variables are varied independently, with the restriction that a + Sa is
again a spacelike hypersurface.
Formally, the functional derivative 6/(50- (x) in eqns (11.7) and (11.8) is de-
fined as follows (see Fig. 11.1). Consider some point x e a and an infinitesimal
variation a -4 a + Sa of the hypersurface around x. The volume of the four-
dimensional space-time region enclosed by a and a + So- will be denoted by
12(x). Then we define for any functional F(a),

SF (o-) _ F (o- + So- ) — F(o- )

uni (11.9)
60- (X) - S2(x)-HD 12(x)

To give an explicit example for this kind of derivative we consider some vector
field rii (x) and define the functional
504 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

a + 6a

,,,„,,,,,„„_.7.. _. . .I
. _. . _______■- -
..
/ .

Q(x)
FIG. 11.1. Illustration of definition (11.9) of the functional derivative (5/(50- (x).
The figure shows some small variation au of the spacelike hypersurface a
around the point x such that a and a + au enclose a four-dimensional
space-time volume 12(x).

F(u) E f do- (x)n 4 (x)F 12 (x).

a
Here, the integration is performed over the hypersurface a with the help of the
Lorentz-invariant surface element da(x) on a which is induced by the Lorentz
metric. Explicitly one has

d3 x
do- (x) = (11.11)
n° (x)'
where d3 x = dx 1 dx 2 dx 3 and n°(x) is the time component of the unit normal
vector n 4 (x) at the point x on u. Thus, F(u) is just the flow of the vector field
174 (X) through the hypersurface a. Using definition (11.9) we find the functional
derivative,

OF(u) _ 1 . 1
f do- (4n4 (x l )Ftt (x 1 ) — f do - (4n 4 (x`)F i,(x 1 )
(50- (x) S-1(x11 —HD Q(x)
)n

[o. +8o- cr
1
= lim f d 1 x' al-T (x')
l(x) -+O
q) S2
Q(x)
= 84114 (X), (11.12)
where we have used Gauss's theorem in the second step. The four-dimensional
space-time integral is extended over the region enclosed by the hypersurfaces
a and a + Ou. This obvious result states that the change in the flow through
a which results from an infinitesimal variation of a around x is given by the
divergence of the vector field at x. For a vector field satisfying attr„(x) 0 this ,

means that the corresponding flow F(u) is independent of a.

 THE SCHWINGER—TOMONAGA EQUATION 505

.. ' ....... ft...„

/
\
-..------------4—."-------....
°
.T y

FIG. 11.2. Sketch of the second-order variation involved in the integrability

condition (11.13). Lying on the same hypersurface a, the points x and y are
separated by a spacelike distance. For the Schwinger—Tomonaga equation to
be integrable the order of the variations around x and y must be immaterial
which is ensured by microcausality.

The Schwinger—Tomonaga equation (11.7) or (11.8) is obviously covariant

since both the Hamiltonian density 71(x) and the functional derivative (51 6o- (x)
transform as scalars under Lorentz transformations. Being an equation in a con-
tinuous family of variables, the Schwinger—Tomonaga equation must fulfil a cer-
tain integrability condition in order for a solution to exist (Schweber, 1961). In
direct analogy to the case of a partial differential equation in a finite number
of dimensions, a necessary integrability condition for the Schwinger—Tomonaga
equation takes the form
452p(a) (52p(u)
/ =
(50- (x)So - (y) (5 o- (y)(50- (x)
where the points x and y are located on the same hypersurface a and are thus
separated by a spacelike interval (see Fig. 11.2). The integrability condition is
a direct consequence of the requirement of microcausality for the Hamiltonian
density, which states that 7-1(x) and 7-1(y) commute for spacelike separations
(Weinberg, 1995),
[7-1(x),1-1(y)] = 0, for (x — y) 2 < 0. (11.14)
Namely, using the Jacobi identity we immediately get from the microcausality
condition
62 p (0) ep(u )
(11.15)
a- (x)o- (y)
(5(5 (5 o- (y)6a- (x) = [[1-1 (x),W(y)] , P(a)] = O.
The integrability condition ensures that the Schwinger—Tomonaga equation has
a unique solution p(a) once an appropriate initial density matrix p(o-o ) has been
specified on an initial hypersurface ao . Formally, this solution can be written as
p(a) = U (o-, o-o )p(o-o )Ut (o-, o-0 ), (11.16)
where we have introduced the unitary evolution operator
,
U(a, ao ) = T, exp [—i f d ix 7-1(x) ] . (11.17)
up

As usual, T, denotes the chronological time-ordering operator.

506 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

11.1.2 Foliations of space-time

A foliation of Minkowski space is defined to be a smooth one-parameter family
T = {a(r)} (11.18)
of spacelike hypersurfaces a(r) with the property that each space-time point x
is located on precisely one hypersurface of the family. If we denote the latter by
ax , this means that we have crx = CY (T) for exactly one parameter value r.
A given foliation a(r) gives rise to a corresponding family of state vectors

I 4' (7 )) —
(11.19)
The Schwinger—Tomonaga equation can then be formulated as an integral equa-
tion,
0- (7)
1CT)) = I 41 ( 0) ) — i f d4 x (x) 1 111 (0-x )). (11.20)
co

The four-dimensional integration is extended over the region enclosed by an

initial hypersurface o-0 = o- (r = 0) and the hypersurface a(r) of the family
which lies entirely in the future of ao.
The hypersurfaces CT (T) of a foliation can be conveniently defined with the
help of an implicit equation of the form
f (x,r) = 0, (11.21)
where f (x, r) is a smooth scalar function. With an appropriate normalization of
f we may assume that the unit normal vector n 0 (x) at the point x E a(r) is
given by

n(x) =
af (x, r) (11.22)
axii
It follows from eqns (11.20), (11.21) and (11.22) that 1111(T)) obeys the equation
of motion
d af
dT IT(T)) = —i f do- (x)
--- —il I (T)IT (T)) , (11.23)
ar 71 (x)1 4 ' (7 ))
0.(7)
where the integration is performed over the hypersurface a- (r) of the foliation.
This is a manifest covariant form of the Schrödinger equation (11.2). To prove
it we first note that for two hypersurfaces of the foliation corresponding to two
infinitesimally separated parameter values T and T + dT eqn (11.20) yields
cr(r+dT)
di(T)) = IT(T + dr)) — 1 41 (7 )) = — i f d4 x/i(x)R(T)). (11.24)
0-(T)
On using d4 x = da(x)Inoaxo /arldr = da(x)ia f 1 ark& one is immediately led
to eqn (11.23).
 THE MEASUREMENT OF LOCAL OBSERVABLES 507

As an example, consider a foliation by means of a family of parallel hyper-

surfaces a(r) given through the equation

f (x, 7 -) E nx — T= 0, (11.25)

with a constant unit normal vector n1 . Such a foliation can be associated with
an observer 0 moving along the straight world line y(T) = ny with constant
velocity V such that

dy
(11.26)
n = -TT = (7 ' 777)
is the 4-velocity of 0 with
1
'Y = (11.27)
V 1 — 11 2
The parameter T denotes the proper time of the observer, that is, the time of a
clock attached to O. At each fixed T the time axis in an observer's rest frame is
given by the unit vector n, whereas instantaneous 3-space at that time is given
by the flat, spacelike hypersurface u(r) which is orthogonal to n and contains
the point VT) , i.e. which is defined by the equation

n(x — y(7- )) E nx — T = 0. (11.28)

The hypersurface a(r) is therefore the set of those space-time points x to which
observer 0 assigns one and the same time coordinate T. We have la 1 par' = 1
and, hence, eqn (11.23) takes the form (Jauch and Rohrlich, 1980)

ddr i l F
--- er» = —i f do- (x) W(441 (7)) E —iH(T)IXF(T)). (11.29)
a(r)

In particular, in the special coordinate system in which the unit normal vector
n coincides with the time axis, n4L = (1, 0, 0, 0), this equation becomes identical
to the Schrödinger equation (11.2). Analogous equations hold, of course, for the
density matrix p(a).

11.2 The measurement of local observables

This section deals with the measurement of local quantities. We begin our discus-
sion by constructing the operation for an indirect measurement at a single point
(or at a single localized space-time region) and formulate an appropriate rela-
tivistic state reduction postulate. The requirements of relativistic covariance lead
to the prescription that the state reduction must be performed on all spacelike
hypersurfaces which pass the classical event given by the measurement outcome.
This prescription for the state vector reduction yields the result that the prob-
ability amplitude becomes a multivalued function on the space-time continuum.
508 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Thus, in the relativistic domain the state vector (or the density matrix) must be
regarded, in general, as a functional on the set of spacelike hypersurfaces. This
concept has already been used in the previous section for the formulation of the
Schwinger—Tomonaga equation.
Having discussed single measurements, we turn to the description of multiple
measurements carried out at a set of points which may be arbitrarily distributed
in space and time. The evolution of the state vector conditioned on the readouts
of the various measurements will be formulated in terms of a relativistically
covariant stochastic process. It is shown further that the measurement outcomes
can be described by a consistent family of Lorentz-invariant joint probability
distributions. The latter are shown to contain all local and non-local quantum
correlations. As an example we briefly discuss EPR-type correlations which are
embodied in non-local, entangled quantum states. We further discuss here the
formulation of relativistic PDPs describing continuous measurements.

11.2.1 The operation for a local measurement

We consider some observable 24(o-,i ) which is associated with a spacelike hyper-
surface U rn, and which is given in terms of an integral over some local, Hermitian
field ço(x),

A(u m ) = f do- (x)G(x)(p(x). (11.30)

Um

Here, G(x) denotes some smooth function with compact support around some
point x m located on urn,. In the following we shall describe the local measurement
of A(a m ) as an indirect measurement. More specifically, we assume in analogy
to the procedure in Section 2.4.6 that the quantum field ço(x) is coupled linearly
to the generalized momentum P of some probe system. The interaction between
the field and the quantum probe is further assumed to be localized in some small
space-time region containing the support of G(x) (see Fig. 11.3). After the inter-
action the generalized coordinate Q, canonically conjugated to P, is measured
on the probe system. The corresponding readout q of the Q measurement and
the initial probe state 10) then lead to an, in general approximate, measurement
Of A (um ).
We consider two hypersurfaces a and o- ' which are separated from the inter-
action region as indicated in Fig. 11.3. Our basic assumption will be that locally
any observable can be measured with the help of such an indirect measurement
scheme and that, at least in principle, the interaction time can be made arbitrar-
ily small such that the free evolution of both the field and the quantum probe
can be neglected over this time. We then get the following expression for the
unitary evolution operator which takes the state of the total system (object plus
quantum probe) from the hypersurface a' to the hypersurface a,

V(o -, o- `) = exp [—iA(cr n,)P] . (11.31)

 THE MEASUREMENT OF LOCAL OBSERVABLES 509

FIG. 11.3. Schematic picture for the measurement of an observable A(am ) with
compact support on the hypersurface a-, . The forward light cone of xi and
the backward light cone of xf are tangential to the interaction region as
indicated. The unitary operator V(o -, a') maps the state vector on a' to the
state vector on u.

As indicated in Fig. 11.3 the hypersurface a' is taken to intersect the backward
light cone V_ (xi ) based at xi which is defined by

V_ (xi .{xelie 1(x — s i ) 2 > o, z0 <x},

) (11.32)

while a intersects the forward light cone V+ (xi) based at X f• ,

V+(xf) = Ix E R4 I (x — X f ) 2 > 0, X° > x}. (11.33)

The points xi and xf are chosen such that these light cones are tangential to
the interaction region. It is clear that a' and a do not intersect the interaction
region. The Q measurement is carried out on a surface such as u which crosses
the forward light cone based at X.
Likewise, the initial state of the field and probe must be given on a hypersur-
face such as a' which crosses the backward light cone of xi . This state is given
by a product state IT(a'))010), where 1111(o-')) denotes the state of the quantum
field. The final state after the interaction between quantum field and probe and
the subsequent ideal Q measurement with the result Q = q then takes the form

lq) OW (a, ai ) (1 41 (ai )) 010) E 10(q)I tlf (C71 )) - (11.34)

Here we have applied the state projection postulate to the ideal Q measurement
which projects the state vector onto the eigenstate 10 of Q corresponding to the
510 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

)) 1 Q(a)0(ur1))
.VP(a)

FIG. 11.4. Idealized picture for the measurement at a single space-time point
x m • The hypersurface u m, is chosen to cross S m , whereas u' crosses the back-
ward light cone of x, and u its forward light cone. The operation Q(a)
describes the reduction of the state vector conditioned on the outcome a of
the indirect measurement.

eigenvalue q. The above expression implies that the indirect measurement device
is described by the operation

Q(q) = (qi exP [ —iA(am)P]10) = (q A(um)) (11.35)

which acts on the states of the quantum field. Here we have introduced the wave
function O(q) = (qP) of the initial probe state in the Q representation.
As an example we take the initial wave function of the probe in the Q repre-
sentation to be a Gaussian function (Dibsi, 1991) with variance 712 ,
2
0(q) = (2771 2 ) -1/4 exp ----1
[-- . (11.36)
- 47) 2

The operation is then found to be

(q — A(u n,)) 2 1
S2(q) = (277/ 2 ) 1/4 exp [ (11.37)
4772
This demonstrates that a probe which has been prepared initially in a Gaussian
state with variance 712 enables the approximate measurement of the observable
A(o-m ) to an error of the order q. We also observe that the readout q leads to
an inferred value for the observable A(o-m ) which is given by q = a. This is due
to the fact that the above Gaussian has zero mean value and, therefore, q = a is
a bias-free estimate for the observable. In the following we shall always assume
that the mean value of the probe states is zero (which can always be achieved by
subtraction of the bias, of course) and identify the readout q with the inferred
value a.
 THE MEASUREMENT OF LOCAL OBSERVABLES 511

11.2.2 Relativistic state reduction

The whole situation developed so far can now be idealized by shrinking the
total interaction region to a single point xm as illustrated in Fig. 11.4. The
device then acts as an indirect measurement of the local observable A(x, i ).
Since all hypersurfaces crossing the forward (backward) light cone based at x m
are unitarily equivalent we are thus led to the following state reduction postulate
for local measurements. Consider some foliation .F = fo- (7- )1 of space-time. Then
the state reduction
1
1 41 (ant)) Q(a)1+(um)) (11.38)
Vp(a)
occurs along that hypersurface um of the foliation which crosses the point 5 m .
The operation for the local measurement at x m now takes the form

S-2(a) = q5 (a — A(x, n )) . (11.39)

The state reduction (11.38) is conditioned on the readout q = a which is dis-
tributed according to the probability density

P(a) = P(a)R(am))11 2 = (11.40)

We recall that Qt(a)52(a) is a Hermitian and positive operator,

Slt(a)S2(a) = I0(a — A(xrn))I 2 > 0, (11.41)

which satisfies

f daS2t (a)S2(a) = f dq10(a — A(x,,))1 2 = I, (11.42)

such that the density 'P(a) is normalized,

(11.43)
f daP(a) = f da(Co-7,0t (a) 0 (a)1+(an-t)) = (111 (um)1 41 (am)) = 1.
Analogous relations are of course valid for mixed states p(o-m ).
For different foliations, the corresponding surfaces that intersect the point 5,„,
can, of course, be different. The above state reduction postulate thus amounts
to the prescription that the state vector reduction occurs along all spacelike
hypersurfaces which cross the point xm in which the local measurement is being
performed. This is the state reduction postulate first formulated by Aharonov
and Albert (1984b). A dynamical model for this relativistic reduction postulate
has been developed by Breuer and Petruccione (1998). For an extended region
given by the support of G(x) we can say that the state reduction takes place
along all spacelike hypersurfaces which coincide on the support of G(x) (see
Figs. 11.5 and 11.6).
512 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

O3

Xm

0- 2

0- 1

FIG. 11.5. Illustration of the state vector reduction postulate. Different folia-
tions lead to different hypersurfaces crossing the point x m . The state reduc-
tion thus occurs along all spacelike hypersurfaces passing S m .

The above state reduction postulate is obviously covariant since it is for-

mulated without reference to any specific coordinate system. Additionally, the
probability distribution P (a) for the readout a as well as for the operation Q(a)
applied to the state vector does not depend on the specific surface a as long
as it crosses the point x m . Being a local observable, A(x,) commutes with the
Hamiltonian density 1-1(x) for spacelike separations, and, hence, 71(x) commutes
with the operation:
[ 7 1(x), 5)(a)] =0, for (x — xfli ) 2 <0. (11.44)
This is the condition of microcausality for the operation. Considering then some
hypersurface which crosses 5, and varying this surface around any point x on
it, keeping x,,,, fixed, we get with the help of eqns (11.40) and (11.44),
6
P(a) = i(111(a)i [1-1(x), Se (a)S)(a)] 111/(0- )) = 0. (11.45)
6a(x)
This proves that P(a) is, in fact, independent of a as long as a crosses x m • The
same conclusion holds for an extended region given by the support of G(x), with
the restriction, of course, that the variations leave invariant the common support
of C(s).
In view of the above state vector reduction postulate, a given foliation .F =
{o- (r)} now gives rise to a corresponding stochastic process I 4/(o- 0-M for the state
vector. Starting from a state 1*(o-0)) on an initial hypersurface ao = o- (r = 0) the
state vector evolves continuously according to the Schwinger—Tomonaga equation
until the (uniquely determined) hypersurface am which crosses x in, is reached.
Then, conditioned on the readout a, the state reduction (11.38) occurs along ain,
and the state vector evolves continuously again. If (Ter) crosses the future light
cone of x in, the state vector on o- (r) resulting from the readout a can be written
as
-

0- (7)

1
1 41 (a(T))) = T Q(a) exp —i f crI x 7-1(x) ) I III (Gro ) )• (11.46)
-VP (a)
co -
 THE MEASUREMENT OF LOCAL OBSERVABLES 513

°.s..... ft.
■ _CT3
.
. .
suppG —

..-
.

— , ....
..
.......i■l ■

0- 2

FIG. 11.6. For the measurement of a local observable with compact support on
some hypersurface the state reduction takes place along all spacelike hyper-
surfaces which coincide on the support of the measured quantity.

Here, the chronological time ordering operator T, acts on the exponential in-
volving the local Hamiltonian density 7-1(x) as well as on the operation Q(a)
which contains the local observable A(x,i ).
The stochastic state vector evolution formulated above is obviously Marko-
vian. Note also that different foliations lead to different state vector evolutions.
However, for identical readouts the state vector (11.46) depends only on the
initial surface ao and on the final surface o- ('r). The process is therefore integrable
in the same sense as the purely unitary evolution according to the Schwinger-
Tomonaga equation: We take two different foliations Ti. = {al ('ri )} and .F2 =
{0-2(72)} with common initial and final hypersurfaces ao and a, respectively.
These foliations yield corresponding state vectors III/1(0 -1(7-i))) and I +2 ( 0-2 (T2 ))).
Integrability then means that both foliations will lead to one and the same state
vector on the common final surface a, namely we have

= (11.47)

provided that we start from the same initial state vector on ao and that we have
identical readouts a. The integrability of the process is an immediate consequence
of the property of microcausality. It implies that the state vector is, in fact, a
functional on the set of spacelike hypersurfaces.
It is important to realize that the state vector history as given in eqn (11.46)
is conditioned on the outcome a of the measurement at xm . It thus depends on
the classical event that the observable Q takes on the value a. Let us suppose that
the measurement of Q is performed immediately after the interaction between
the field and the quantum probe and that the result is communicated via a
classical light signal. The readout a is then available everywhere in the forward
light cone based at x m . The latter is defined as the set of points x satisfying
(x — X 70 2 > 0 and x° > 4, and will be denoted by V+ (xi,) (see eqn (11.33)).
Consider an observer 0 moving along a world line y = y(T) where T denotes
the proper time. If that world line intrudes into V+ (x n,) at the proper time 'r1,
say, the observer knows the readout a of the measurement on the probe and can
thus set up the state vector history (11.46) depending on a. It is clear that the
514 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

state reduction (11.38) takes place in the observer's past, that is for Tm < ri .
which means that it occurs (possibly a long time) before the observer gets the
information on the readout.

11.2.3 Multivalued space time amplitudes

-

As we have seen the state vector reduction expressed by eqn (11.38) yields a
covariant stochastic state vector history (11.46) associated with each foliation
of space-time. This fact has been demonstrated to be a simple consequence of
the causality principle. However, it must be emphasized that the state reduc-
tion postulate leads to the important conclusion that the probability amplitudes
which are determined by the functional ItIf(a)) need not represent single-valued
functions on Minkowski space (Aharonov and Albert, 1984b).
We illustrate this point with the help of an example. To this end, we first
construct a simple device for the effective position measurement carried out on
a single-electron state. Consider the observable

A = f d 3 xG(±4)0 1 (x)0(x), x = (t,„ ±s), (11.48)

where Ot,„(x) and Oa (x) denote the field operators of the electron field, a being
a spinor index. They create and annihilate, respectively, an electron at x and
satisfy Fermionic anticommutation relations,

= fOta(t,Y), 143(t,I 1 )} = 0 ) (11.49)

Na(t,±4),Oti3(t, * ')} = 6 (±4 — ±")60. (11.50)
Furthermore, G(Y) is a smooth function with compact support. We suppose that
G(Y) is equal to 1 in a small region g of space around x rn, and falls rapidly to
zero outside G.
Taking 1111(o-m )) to be a one-electron state, we consider the amplitude

x(x) = (0 10 (x)1 41 ( 0-m)) (11.51)

for the electron to be at the space-time point x, where 10) denotes the ground
state vacuum of the electron field. Our aim is to determine how the operation
f/(q) pertaining to the measurement of A acts on this amplitude x(x). We find

(0 10(x)Q(q)R(am)) = (q1( 0 10(x)e -zAP IT(um))10)

_ mo ieiAP 0 (x)e -JAPI T /urn
MO), ( 11 . 52 )
where we have used A10) = 0 in the second step. The anticommutation relations
yield

[A,/p(x)] = —G(Y)/p(x), (11.53)

from which it follows that
 THE MEASUREMENT OF LOCAL OBSERVABLES 515

e iA (x ) e —iAP = e — iG(x ) .
(11.54)

Therefore we get

( 0 10(x) 1) (q)iCum)) = (qi(Oie -2G(g)P7P(x)iCam))10)

= (q G(Y))X(x). (11.55)
This shows that the operation acts as follows on the one-electron amplitude,

x(x) 0(q — G(Y))x(x), (11.56)

and that the probability density for the readout q becomes

P (q) = f d3 x10(q G (Y))1 2 1X(x)1 2 (11.57)

In order to enable a sufficiently accurate measurement the probe wave function

0(q) must be a function which is sharply peaked around q = 0, with a width
much smaller than 1. Equation (11.57) clearly shows that our device allows the
approximate measurement of the random variable q =
Suppose now that the amplitude x(x) consists of two localized wave packets
x 1 (x) and '( 2) (x), such that the support of x( 1 ) (x) is contained in the region g,
where G() is equal to 1. The support of (2) (x) and that of G(Y) are assumed
to be disjoint. We then have

P(q) = 146(q — 1 )1 2 f d 3 x 1X (1) (x)1 2 + 10(012 f d 3 x 1X (2) (x)1 2 . (11.58)

This means that we get the readout q 1 with probability f d3x I1(x)I2,
showing that the particle is in g, and the readout q 0 with probability
f d3 x lx(2)(x)1 2 , showing that the particle is not in g. As can be inferred from
eqn (11.56), in the first case the operation projects the amplitude x(x) onto
x (1 ) (x), in the second case onto ( 2 ) (x),

X(x) X (1) (x) 111X (1) 11, for q 1, (11.59)

x(x) x(2) (x)/11x (2) II/ for q 0. (11.60)
Let us now consider the following situation (Aharonov and Albert, 1984b)
which will lead to the conclusion that the amplitude x (x) is not a single-valued
function on the space-time continuum (see Fig. 11.7). Suppose that we have
prepared on an initial hypersurface cro some one-electron state lilf(a0 )) such that
the amplitude (010(x)14/(o-0 )) represents a superposition of two wave packets
x (1)( x ) and x (2)( x ) ,

(0 10(x)1 41 (a0)) = x 1 (x) X(2 ) (x), x e cro. (11.61)

The wave packets are supposed to be localized in small regions of space around
Y • - ) and Y(2), respectively, and follow their world tubes with zero mean velocity.
516 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

P4
Iw 2) )

I 1li (0-1))
al

a2

I w(0-0))
0.0

FIG. 11.7. The figure illustrates that space-time amplitudes are, in general.
multivalued functions: The state vectors associated with two different hyper-
surfaces a l and a2 may yield different amplitudes at a common point Q if at
the point P a position measurement is carried out.

For simplicity we may neglect the extension as well as the spreading of the wave
packets. These effects can easily be taken into account, but do not change the
argument.
At some space-time point P a position measurement is performed with the
help of a device of the type considered above. We assume that the measurement
leads to the result that the electron is at P (the case q = 1 above). Given such
a situation we may consider two flat spacelike hypersurfaces a i and a2 which
intersect at the space-time point Q. Both hypersurfaces appear as equal-time
hypersurfaces in appropriately chosen coordinate frames K 1 and K2, that is
there are observers 0 1 and 0 2 at rest in K 1 and K2, respectively, such that u 1 is
an equal-time hypersurface for 0 1 , and u2 is an equal-time hypersurface for 09.
The important difference between both observers is that for 02 the measurement
has already taken place, whereas for 0 1 it has not. Consequently, both observers
assign different amplitudes to one and the same objective space-time point Q.
Namely, the state reduction following the measurement of the electron at P yields

( 01 0(0141 (g 2 ) ) = 0 (11.62)

on a2 . On the other hand, on al the state reduction has not yet occurred so that
we have

(0 10(01 41 (0-1)) = x 2 (Q) O. (11.63)

Thus we find that the amplitudes differ at the point Q where both hypersurfaces
intersect,

(010(Q) 1 41(0-0 ) 0 0 10(01*(0-2)), (11.64

 THE MEASUREMENT OF LOCAL OBSERVABLES 517

This relation clearly demonstrates our claim, namely that the one-particle am-
plitude (0[0(x)141(a)) is a multivalued function on space-time: The value of this
amplitude at x depends, in general, on the hypersurface a which crosses x. In
other words, the amplitude depends on the foliation and, thus, on the complete
state vector history.
Although there is obviously a Lorentz transformation which maps the plane
al to a2 , the corresponding one-particle amplitudes are clearly not related by
a unitary transformation. This is due to the fact that the measurement on the
system makes it an open system such that in situations like the one considered
above there is no unitary representation of the Lorentz group for the reduced
system. The same conclusion holds also if we consider the corresponding non-
selective measurement: In that case we have a mixed state on a-2 , whereas we
have a pure state on ai.

11.2.4 The consistent hierarchy of joint probabilities

Generalizing the foregoing analysis we now consider a sequence of local measure-
ments performed at the points

X
(1)
, X
(2)
,... , x (IC) , (11.65)

which may be arbitrarily distributed in space and time. At each x ( k ) , where k =

1, 2, ... , K, a local observable A( k ) = A(x(k)) is measured indirectly through the
coupling to a quantum probe in the initial state 10(k )). The coupling between the
field and the k-th quantum probe involves the generalized momentum P(k) and
a direct measurement of the canonically conjugated observable Q (k) is performed
after the interaction. We assume here that the K quantum probes act completely
independently on the field, that is, we may write

10) = ® 10(2)) ® ... ® i O K)) (11.66)

for the total state of the K probe particles. We also assume as before that the
mean values (0(k) 1Q ( k)10(k)) vanish, such that the inferred values a(k) for the
observables A(k) are simply given by a( k ) = q( k) . Writing (/)(k)(q(k)) for the Q(k)
representation of the initial state of the k-th probe particle we thus obtain the
operations

Q(k) ( a(k)) _ 0(k) ( a(k)

(11.67)

which describe the change of the system state conditioned on the outcome a(k)
at X (k )
As mentioned before, the points x(k ) may be distributed arbitrarily in space
and time. Thus, we may have spacelike as well as timelike separations between
them. Since the A( 1') are assumed to be local observables, the causality principle
518 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

ensures that for spacelike separations the operations commute with themselves
and with the Hamiltonian density, that is we have
[Q (k)( a (k)) , wt)( a (0)] = 0, for (x (k) — x (1) ) 2 <0, (11.68)

[71(x), ft (k) (a (k) )] =0, for (x — x (k) ) 2 < 0. (11.69)

The results of the previous subsection are now readily generalized to the above
collection of local measurements (Breuer and Petruccione, 1999). Taking some
foliation with initial surface ac, we find for the state vector on a(r) (compare
with eqn (11.46)),

K o(T)
IT ( 0 - (T))) =V • T,
T ( 11 S-2 (k) (a(k) ) exp —i f d 4xl -i(x) 1 111 (uo))• (11.70)
k=1 (70
Again, the time-ordering operator acts on the exponential as well as on the
operations ç1(k). The normalization factor .11/ is given by
Ar _ [pK(a(i),... ,a(K))] -1/2 ,
(11.71)

where 'PK (a (1 ), ... , a (K) ) denotes the joint probability of the readouts,
-
2

Pic (a(1) , • • • , a(K) ) — T,

(K
IT Q(k) ( a (k) )
(7 ( 7 )

exp[_i f ct I X 7 -1(x)
k=1 (70
-

(11.72)
The surface (Ter) may be chosen arbitrarily with the only restriction that all the
x( k ) must be located in the past of it, which means that a(r) must cross the
forward light cones of all the
Equation (11.70) associates with each foliation a unique stochastic process
Each realization of the process represents a state vector history which
is conditioned on the readouts that follow the joint probability (11.72). It should
be clear that this process is Markovian and integrable. In particular, the joint
probability PK (a( 1 ), .. . , a(K )) represents a Lorentz-invariant expression and does
not depend on the foliation. The reason is that the time-ordering operator T, is
defined in an invariant fashion by virtue of the causality conditions (11.68) and
(11.69). We also note that due to the completeness relations

f da(k)Q(k)t (a(k))Q(k)(a(k)) = I (11.73)

the joint probability is normalized,

f da(1) . . . f da(K) KT, (a (i) , . . .

,a) _ 1. (11.74)
 THE MEASUREMENT OF LOCAL OBSERVABLES 519

x( 2)

Q(..F1) = Q(3 )Q( 2)Q( 1 ) QGF2) = Q( 1 )Q(3 )s2 )

(2

x(n•
•x ( 2 )

-----------

Q ( F3 ) _ Q(3)Q(1)Q(2)

FIG. 11.8. An example foreqn (11.70) involving three measurement points with
two spacelike and one timelike distance. The figure shows three different
foliations which all lead to one and the same total operation.

Let us illustrate in Fig. 11.8 how eqn (11.70) works for the specific case
of three measurement points (a similar situation has been considered by Bloch
(1967)). The distances between x( 1 ) and x( 2) and between x( 1 ) and x( 3 ) are
spacelike, whereas x (2) and x( 3) are separated by a timelike interval. The figure
shows three different foliations 1.1 , .F2 , and J. According to eqn (11.70) these
foliations lead to the three total operations
= Q(3 )Q( 2)Q( 1 ) (11.75)
s--2(y2 ) _ (11.76)
QGF3 ) , (11.77)
which describe the state vector on the final hypersurface of the corresponding
foliation (for simplicity, the Hamiltonian density may be set equal to zero). f2(1)
520 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

P3 (a(1), a(2), a(3))

-p2 ( a(1) , a m)

FIG. 11.9. The hierarchy of joint probabilities (11.78) seen by an observer mov-
ing along some world line y(r): Each time the observer intrudes into a new
light cone of the measurement points x (k ), a new member of the family is
generated.

commutes with 12( 2) and with SP), but 11 (2 ) does not commute with S2( 3 ), of
course. As is easily seen all three total operations are identical. This shows the
integrability of the process as well as the relativistic invariance of the joint prob-
ability distribution.
It must be emphasized that the state vector history (11.70) is conditioned
on the total readout (a( 1 ), ... , a(K)). We consider again an observer 0 moving
along some world line y = y(r). As before we suppose that the results of the
measurements at the points x(k) are communicated via light signals such that
the outcomes of the single measurements are available in the respective future
light cones V+(x (k )). Moving along y(T) the observer 0 successively intrudes into
these light cones. Suppose that at the proper time 7- observer 0 has intruded
into the light cones of the points x( k i), where 1 = 1, 2, ... , L, and L < K.
Observer 0 then defines for any foliation the state vector history which is given
by eqn (11.70) with the only modification that the product Fik now extends only
over the points x(ki) in the forward light cones of which 0 is located, that is only
over those measurements for which 0 knows the outcome.
It should be clear that in order to set up the state vector history observer 0
need not know the outcomes of the measurements in advance, nor does 0 have to
know in advance which observable is measured at x( k ) or if any measurement at
this point has been performed at all. The information on what is measured and
where the measurements took place can be communicated, of course, together
with the corresponding readouts.
Let us, for ease of notation, label the measurement points in such a way that
observer 0 intrudes first into the light cone of x( 1 ), then the light cone of
and finally the light cone of x(K). Moving along, observer 0 thus generates. ,
 THE MEASUREMENT OF LOCAL OBSERVABLES 521

a family of joint probabilities (see Fig. 11.9),

pi (a( 1 )) , p2 (a( 1 ) ,a( 2)) , ... pK (a( 1 ) , ... ,a(K)) , (11.78)

which are defined by

, a (k ) ____
)

H
T+__ (n S2M (am) exp —i d4 x 71(x) I 4' co
( ))

,.i CO

(11.79)

The points x( 1 ), ... , x(k) are located in the space-time region enclosed by the hy-
persurfaces o-0 and o-(r). As we have already noted all joint probabilities (11.78)
are normalized and do not depend on the specific chosen foliation. The latter
property implies that any two observers with the same information on the read-
outs agree completely on the corresponding joint probabilities. The following
consistency condition can also be verified easily,

(k) = f da(k+i) Pk+1 (a(1) , ...

, a) (11.80)

This shows that each new readout a(k+1) which is communicated to observer 0
when she or he intrudes into the light cone 174.(x (k+ 1 )) is compatible with the
joint probabilities of lower order in the hierarchy. Thus we conclude that (11.78)
forms a consistent hierarchy of joint probability distributions.

11.2.5 EPR correlations

It is important to note that the above formulae have been derived without re-
strictions on the initial state vector 1 klf (o-0 )). They are thus capable of describing
arbitrary quantum correlations between timelike and spacelike separated points.
Quantum correlations at spacelike separated points arise if Illi(o-0 )) represents
a non-local, entangled state. As an example we consider Bohm's formulation
(Bohm, 1951) of the famous EPR gedanken experiment developed by Einstein,
Podolsky and Rosen (1935): A two-particle system in a state of total spin 0,

ocao» = .--,i_
li (1+)(1)1—)(2) — 1—)(1)1+)(2))
disintegrates into two particles with spin 1 flying with opposite velocities along
the x-direction. In the initial state (11.81) the states I±)( k) are eigenstates of
the z-components olk) of the spin operator at two points x ( k), k = 1,2, with
eigenvalues ±1. At the points x (k ), being spacelike separated, the spin projections
a • 6-(1) and 3 • 5 (2) are measured. Here 6 and g denote unit vectors lying in the
522 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

(y, z)-plane and forming angles a and with the z-axis. Neglecting any particle
interaction we now have for the joint probability of the measurements
= a (0 )!-2 (2) f a (2) (11.82)
)1 41 (c0))11 2 .
Assuming that the measurements represent ideal measurements of the spin pro-
jections we have two possible readouts a(k ) = ±1, whereas the operations are
given by the projections on the corresponding eigenvectors. We then find for the
single measurements at x (1 ) and x (2) the mean values

( 41 (0-0)Ic .d (1) 141 (0- 0)) = OP(uo)Ig • d (2) 1 41 (0-0)) = 0, (11.83)

and the variances

Var (Ci .6 (1) )= (4/(Gro)l (6. d (1) ) 2 1, (11.84)

Var (d. 6 (2 ) ) ( 4quo)1 (ig • 6(2) ) 2 quo)) 1. (11.85)

The correlation coefficient (see eqn (1.40)) is easily found to be

E a (n a(2),p2 ( a(i) ,Cor(a,3)= a(2))
a (t) ,a (2)

(W(Gro) 6r' (1) ) d (2) ) 1 41 (0-0))

= — cos(a — /3). (11.86)

The correlation coefficient describes the quantum correlations embodied in the

joint probability distribution P2(a (1) ,a (2) ). For a = ,3 we have Cor(a, fi) = —1
from which it follows that the readouts are perfectly anticorrelated, namely that
a(1)
For general measurements on the particles we have the relations

Pi (a') =10 1) (a (1) )1 41 (uo))H 2 3 (11.87)

(a (1) ) = f da(2) P2(a (1) a(2) ). (11.88)

Equation (11.87) reveals that the probabilities for the single, local measurement
at x (1) are completely independent of the measurement at x (2) : They do not de-
pend on the readout a( 2), nor do they depend on the observable being measured
there, that is on the angle /3, for example. The probabilities for the single mea-
surement at x( 1 ) do not even depend on where a possible second measurement
is performed, as long as the latter is spacelike separated from x( 1 ), of course. As
pointed out already, eqn (11.88) means that the unconditioned probabilities for
the single measurement at x( 1 ) are not changed when the local observer gets any
information on the measurement at a second point x(2 ), which clearly reveals the
consistency of the joint probabilities for the present example.
 THE MEASUREMENT OF LOCAL OBSERVABLES 523

11.2.6 Continuous measurements

With the help of the idea of the state vector as a function on the set of spacelike
hypersurfaces, the stochastic wave function representation of continuous, selec-
tive measurements can be given a relativistically covariant form. As an example
we construct here a piecewise deterministic process for the state vector of the
source of an optical cavity which is covariant under Lorentz transformations and
which describes the stochastic dynamics induced by a continuous monitoring of
the radiated photons through a moving detector. The generalization to the gen-
eral case of a quantum dynamical semigroup generated by an arbitrary number
of Lindblad operators is straightforward; further details of the theory may be
found in (Breuer and Petruccione, 2001).
We consider the output signal of a two-level atom in an optical cavity which
is detected by a photocounter. The photodetector moves with velocity 71 = vé
relative to the cavity in the direction é of the output signal, thereby following a
world line y(r) in Minkowski space. Again, the parameter T denotes the proper
time of the detector, that is, the time of a clock attached to the detector. In the
following we allow for an accelerated motion for which the 4-velocity

dy _,
(11.89)
n(T) EE Tr = (eY ' 7v)
is not a constant, where nl-i(T)n ji (T) ---- 1, and 7 :-E-- (1 — v 2 ) -1 / 2 . For each 'T the
equation

n(r)(x — y(T)) , 0 (11.90)

defines a fiat, spacelike hypersurface o- (T). This hypersurface is the set of those
space-time points x to which an observer 0 moving with the detector assigns
one and the same time coordinate 'r. The family of hypersurfaces .T = {a(T)}
determined by the path y(T) represents a foliation of a certain space-time region.
More precisely, this is the space-time region in which different hypersurfaces of
the family do not intersect. The corresponding condition is /(x) < g — ', where
/(x) is the distance of a point x from the world line y(T) and g is the accel-
eration of the detector measured in its own rest frame (Misner, Thorne and
Wheeler, 1973). The family { a(T)} then provides a foliation of the set of space-
time points x which fulfil this condition. As we shall see, beyond a distance
g-1 fr om the detector's world line a complete, continuous measurement
and,
therefore, a stochastic representation of the source dynamics ceases to exist, in
general.
Our aim is now to construct a relativistically covariant, stochastic representa-
tion of the source dynamics in terms of a state vector I TM) = I T(o- (T))) which
results from the continuous monitoring of the radiated photons through the mov-
ing photodetector. Each time a photon is detected the wave function of the source
undergoes an instantaneous change. We assume, as in the non-relativistic theory
(see Section 6.3), that this change is obtained through the application of some
524 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Lindblad operator L to the state vector. Due to the lack of an absolute time the
term instantaneous change of the wave function is not a relativistically covariant
one. However, by the very principles of quantum mechanics, source, radiation
field, and detector have to be regarded as a whole. The important conclusion is
that a Lorentz transformation affects the quantum object and the probe as well
as the detector and the hypersurfaces a(T) of the foliation of observer O. Thus
we do get a covariant prescription for the state vector reduction if we postu-
late that the state vector reduction occurs instantaneously in the detector's rest
frame, that is, along a certain spacelike hypersurface cr('r) of the foliation of the
observer 0 associated to it.
As a consequence the jump operator now becomes a function L(T) = L(o- (7-))
of the hypersurfaces of the foliation. In direct analogy to the non-relativistic
formulation we obtain on the basis of our state reduction postulate the following
Markovian stochastic state vector equation describing a piecewise deterministic
process:
MI (r)) , —iH (7 - )141(7- ))dr
— 1 (Lit (r)L(T) — (Lt (r)L(T))) 14 ler))da

( L(T)I4 1 (T))
+
V (Lit (r)Ler))
1 41 (T))) dN (a),

where we have introduced the abbreviation

(Lt(T)L(r)) E01/ ( T)Iii t (T)L(T)1 11/ (T))•
The first term on the right-hand side of eqn (11.91) represents the unitary dynam-
ics (see eqn (11.23)), while the second and the third term yield the irreversible
part of the evolution induced by the continuous monitoring of the quantum ob-
ject. The structure of these terms is similar to the one encountered in Section
6.1.1. The differential da plays the rôle of an invariant time increment which will
be determined below. The quantity dN (a) is the increment of a Poisson process
which obeys
E [dN (a)] , (Lt (r)L(r))da, [dN(a)1 2 = dN (a), (11.93)
where E denotes the expectation value of the process. The second relation in
(11.93) tells us that dN (a) takes on the values 0 or 1. As long as no photon
is detected we have dN (a) --,-- 0 and, thus, the second term of eqn (11.91) gives
the evolution of the state vector conditioned on the outcome that no photon is
detected. If a photon is detected we have dN (a) :,--- 1, such that the third term
of eqn (11.91) yields the corresponding jump of the state vector given by
L(o- ) IT (a)
(11.94)
\AV (a)L(a))
Thus, the Poisson process N (a) simply counts the number of photon detection
events.
 THE MEASUREMENT OF LOCAL OBSERVABLES 525

According to eqn (11.94) the stochastic jumps of the state vector occur along
the hypersurfaces u :=--- a- (r) of the foliation given by the detector path. If a
photon has been detected at a certain proper time T the state vector reduction
has to be performed at the corresponding retarded proper time
R
fret :=--- T — — , (11.95)
c

taking into account the time required for the light signal to propagate from
the source to the detector, where R denotes the instantaneous distance from
the source to the detector. This follows directly from the fact that the detected
signal yields information on the state of the source at the retarded time. Thus,
the precise prescription for the state vector reduction takes the following form:
The reduction of the state vector occurs along the spacelike hypersurface ( )i
a \ Tret
at the retarded proper time T„t which corresponds to the proper time 'T of the
actual measuring event. Thus, eqn (11.91) gives rise to a stochastic equation of
motion for the source wave function IT( Tr e t ) ) :=-- IT (a ( Tref ) )) •
In order to determine the invariant time parameter a used in eqn (11.91) we
first observe that, according to eqn (11.93) , the photocurrent as measured in the
rest frame of the detector is given by
da
J ---- (V erg(r)) . (11.96)
J is the average number of photon counts per unit of the proper time interval
dT. Due to the Lorentz invariant nature of the scalar product we may simply
set (Lt(r)L(r)) = -ye, where -yo is an invariant emission rate characteristic of the
source. Thus we have
da
(11.97)
j — 7° Tr .
If the detector is at rest with respect to the source (y = 0), the detected pho-
tocurrent must be Jo = -ye . Thus, a must be equal to the proper time of the
source, i.e. a is the time of a clock fixed at a position in the vicinity of the
source. To see that this conclusion is correct we consider the case of a moving
detector (y 0 0). It is easy to show with the help of the transformation laws for
the electromagnetic field strength tensor that the photocurrrent J as measured
in the rest frame of the detector is given by

,11. —y
J = l'o (11.98)
1+v'
where we do not assume that y is constant. Now, with the above choice for the
quantity a we find

(11.99)

which proves our claim.

526 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

As mentioned earlier the source must lie within a distance 1 < g-1 from
the world line of the detector. This condition implies that da I drret, > 0 and
that, therefore, eqn (11.91) represents a sensible stochastic state vector equation
with a positive increment da. If this condition is violated, obviously no complete
continuous monitoring by the detector is possible. As an example one might
think of a detector in hyperbolic motion, in which case the observer can outrun
the photons radiated by the source.
The covariance of our stochastic state vector equation (11.91) under Lorentz
transformations is obvious. Since da is an invariant, the quantity dN(a) is an
invariant stochastic process. Furthermore, since the quantum expectation value
(Lt(a)L(a)) transforms as a Lorentz scalar both the dissipative and the stochas-
tic term of (11.91) transform covariantly. It is important to emphasize that the
transformation laws also involve a transformation of the jump operator, namely

L'(0) --,--- UL(o-)Ut , (11.100)

where U denotes the unitary representation of the Lorentz transformation. Phys-

ically, this means that the quantum object as well as the environment and the
measuring apparatus have to be Lorentz transformed in order to obtain covari-
ance of the stochastic process. In view of the physical meaning of the process
as a continuous measurement, this is a plausible prescription which is in full
agreement with both quantum mechanics and special relativity.

11.3 Non-local measurements and causality

In the analysis of the preceding section we have assumed that the total initial
state of the quantum probe is given by a simple product state of the form (11.66).
A more general class of measurements can be constructed with the help of quan-
tum probes which are in entangled initial states. We are going to demonstrate in
the present section that entangled probe states enable one to perform the mea-
surement of non-local observables and states as first shown by Aharonov and
Albert (1980, 1981, 1984a, 1984b).
The possibility of non-local measurements has several important consequences
for the notions of states and observables in relativistic quantum theory. As before
we will assume that only local interactions between object and probe system are
involved. The dynamics of the total system will therefore be compatible with
the causality principle, of course. However, if causality is combined with certain
conditions on the properties of the measuring device and on the behaviour of
the state vector of the object system alone, one is led to important restrictions
on the measurability of observable and states (Aharonov, Albert and Vaidman.
1986; Sorkin, 1993).
The conditions imposed by the causality principle will be derived in this
section. For example, we shall demonstrate below that the projections on non-
local, entangled states are not measurable in the conventional sense of quantum
mechanics. It will also be shown that only specific classes of entangled, non-local
 NON-LOCAL MEASUREMENTS AND CAUSALITY 527

states and only certain types of non-local observables with entangled eigenstates
allow a quantum non-demolition (QND) measurement.
Although the possibility of QND measurements of non-local observables and
states is strongly restricted by causality, the preparation of non-local states is
not. Indeed, as will be shown any non-local entangled state could in principle
be prepared without conflict with causality. There is also a quite different type
of measurement, known as an exchange measurement, which never leads to any
contradiction to causality and which will also be discussed. Finally, as an inter-
esting application we investigate the transmission of an unknown quantum state
with the help of a classical communication channel and a quantum channel pro-
vided by an EPR entangled quantum state. This so-called quantum teleportation
nicely illustrates some of the features of relativistic quantum measurements.
11.3.1 Entangled quantum probes
We consider the case that the probe system has been prepared initially in some
entangled state 10). In the Q ( k) representation we write for the wave function of
the total probe system
(* (1) ,... ,q (K)) = ((Jo) , ... , q (K)10) . (11.101)
Our previous result (11.72) on the joint probability distribution can then be
immediately generalized to yield
PK(a (1) , ... ,a (K) ) (11.102)
2
o- ( 7)

, T, f2(a (1) , ... ,a(K) ) exp —i f d l x 71(x) 0(o-0 )) ,

cro
where, again, the x (1 ), ... , X (K) are located in the space-time region enclosed by
the hypersurfaces uo and a('r). The total operation may be written
,a (K) ) = 0(a (i) _ A(i) ,.. , , a (K) _ Am). (11.103)
It should be noted that, in general, this is only a formal expression since any
two operators A(k) and A (1 ) need not commute if the corresponding points x ( k )
and x(1 ) are separated by a timelike distance. However, under the time-ordering
operator T, in eqn (11.102) the operation is unambiguously defined.
The usage of an entangled quantum probe leads to decisive consequences for
the conditioned evolution of the state of the object system. The most important
feature is that for entangled probe states the evolution does not, in general, trans-
form pure states into pure states and that the process becomes non-Markovian.
This can be seen with the help of the simplest case, namely that of a measure-
ment at two spacelike separated points x( 1 ) and x( 2 ). Since the corresponding
operators AM and A(2) commute, the operation
= q5(a(1) — — 4 (2) ) (11.104)
is unambiguously defined without the time-ordering operator, of course.
528 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

preparation of
probe state

IT)
FIG. 11.10. General scheme for an indirect measurement device which uses an
entangled probe state 10).

We suppose that the probe state 10) is an entangled state. To be specific we

may regard 10) as the state of a two-particle system which is prepared in an
entangled state in some localized region. These probe particles are then brought
to the space-time points x (1) and x (2) , respectively, where they interact with the
object system. Afterwards, the observables Q( 1 ) and Q( 2) are measured on the
probe particles to yield the readouts a( 1 ) and a( 2 ), as indicated in Fig. 11.10.
Consider now two different foliations Y and Y' by parallel surfaces as in-
dicated in Fig. 11.11. As before we may regard the hypersurfaces belonging to
these foliations as equal-time hypersurfaces of two observers 0 and 0' moving
with different 4-velocities n and n'. Clearly, n is the unit normal vector of Y and
n' is that of .P.
In the case that the probe represents an entangled state one finds that 0
and 0' associate completely different state histories to their foliations. Let us
first look at the situation from the viewpoint of O. Observer 0 starts from some
initial state 141(a0 )) on ao . As indicated in the figure, there is some surface al
of the foliation .7- which crosses both points x (1) and x (2 ), which means that
both measurements are simultaneous for observer O. This implies that the state
reduction in Y occurs along a l and that the readout a(1 ), a (2) gives the final
state

141(0-2)) = Ar 1l(a(i),a(2))141(0-0)) (11.105)

on the surface a2 of the foliation of 0, A( being some normalization factor.

Thus, observer 0 describes the state history as an evolution of pure states, the
reduction taking place at a single instant of time in the rest frame of O.
 NON-LOCAL MEASUREMENTS AND CAUSALITY 529

X(1) X(2)

0- 0

F(observer 0)
.P(observer 0')

FIG. 11.11. Two different foliations Y and

corresponding to two observers
0 and 0' who describe the non-local measurement depicted in Fig. 11.10 in
their respective coordinate frames.

The state history associated with the foliation .T' is completely different. In
this foliation there is no hypersurface which crosses both points xW and x( 2),
that is for observer 0' the two measurements are not simultaneous: There exists
an intermediate region of time in which the first measurement at x (2) and the
corresponding state reduction already took place whereas the one at x(1) has not.
For an entangled quantum probe this leads to the result that in this intermediate
region object and probe are, in general, in an entangled total state which is given
by

= f da' a') 0 l a (2 )) (a (1) ,a (2) e -iA( 2 )p( 2 ) 10)1111(4)). (11.106)

By contrast, if the probe is in a direct product state kb) = 10(1) ) 0 10(2)), the
intermediate state also becomes a direct product

1 4qui))= -V10 (1) ) 0 10/ (2) ) Ø (a(2) 1e — " 2(2) 10(2) )1 41 (0- ) ), (11.107)

as it was in those cases considered in the preceding section. It is also clear that
after the second measurement at x(1) the total object-plus-probe system is always
in a product state again which takes the form

1 4) (c4)) = Ar1a (1) ) 1a (2) ) 0 (11.108)

The entanglement of the object—probe system shows that the state referring
to the observables of the object system represents, in general, a mixture in the
intermediate region of the foliation .P,

f
p(o-D da(1) q5(a(1) , a (2) A(2) )p(o-D0t(a (1) , a (2) — A (2) ), (11.109)
731(a1(2) )
530 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

where

Pi (a(2) ) --=-- f da (1) NI ()10t 0 (a(i) , a(2) — A (2) )11I1 (cr,') )) (11.110)

is the unconditioned probability density for the readout a( 2 ) of the measurement

at x( 2). Thus we see that observer 0' associates the following history for the
object state with her/his foliation P,

p(o- ) (pure state) p(o) (mixture) p(o) (pure state). (11.111)

In the first step the pure state p(o) is transformed into a mixture p(o-'1 ) by
the first measurement at x(2 ). In the second step this mixture then transforms
into a pure state p(0-) again: The destruction of the pure state by the first
measurement is completely undone by the second measurement.
The above description clearly reveals the non-Markovian character of the
state history. The final state p(a) of the object system after the second mea-
surement is obtained by applying the operation 1l(a( 1 ), a( 2 )) to the initial state
p(o), that is by applying it to an object state which was given at a finite time
interval AT prior to the second measurement. We remark that this time interval
corresponding to the intermediate region of time between the two measurements
can be made arbitrarily large. For example, if x (1) = ( 0, f(1) ) and x(2 ) = (0, Y(2))
are the coordinates of the measurement events in the rest frame of 0 and if 0'
moves in the direction of Z( 1 ) — ±.( 2) the time interval between the first and the
second measurement is found to be

AT = -yvAx, (11.112)

where 7 = (1—y2 ) -1 / 2 , y is the speed of 0' relative to 0, and Ax = 40) _ 2)•

The non-Markovian feature of the process is due to the fact that the infor-
mation which is required to restore the pure object state in the second step is
contained in the probe state during the time interval AT . This information is lost
by tracing over the variables of the quantum probe. As we have seen, the pro-
cess describing the behaviour of the total object-plus-probe system is, of course.
Markovian.
11.3.2 Non local measurement by EPR probes
-

In this subsection we shall demonstrate that the use of entangled probe states al-
lows the quantum non-demolition measurement of certain non-local observables.
As in the preceding subsection our quantum probe constitutes a two-particle
system with canonical coordinates Q (1) and Q(2 ) and corresponding conjugated
momenta P( 1 ) and P( 2). The initial probe state is taken to be an EPR-type
entangled state which may be defined with the help of the relations
(p(1)+ p(2)) 10) ___ 0 , (11.113)

(Q(1) - Q( 2 )) 10) = 0 . (11.114)

 NON-LOCAL MEASUREMENTS AND CAUSALITY 531

Note that the total canonical momentum

p ___ p(1) + p(2) (11.115)
commutes with the relative coordinate
Q ___ Q (1) _ Q (2) , (11.116)

such that P and Q can take simultaneously sharp values. Introducing also the
average of the coordinates
1
Q = (Q( 1 ) + Q(2)) , (11.117)
2
as well as the relative momentum
I:, 1 ( p(1) _ p(2) ) (11.118)
2
we have two new pairs of canonically conjugated coordinates and momenta,

[-P,(2] = [15,C2]= ---i, (11.119)

[P,C2]=[P,Q]= 0. (11.120)

Using the mixed (P, Q) representation we can thus define the probe state by
means of
(11.121)

Let us determine the operation describing the readouts a( 1 ) = q( 1) and

a (2) = q(2) for the local measurement of Q(') and Q( 2) after the object—probe
interaction. Employing the above relations we find
_ a(2) lexp[—iA(1) P(1) — iA(2)P(2)]10)

:=--- (a (1) , a(2) 1ex1){ — i(A (1) — A(2) ) 13 — —2i (A (1) + A (2) )Pi10)
4 (11.122)
= (a (1) ,a(2) 1exP[— i(-4(1) — (2))15L0),
where we have used the fact that P commutes with P and that PP) = 0.
Consider now the non-local operator
A = A(1) — 4 (2) , (11.123)

and its spectral decomposition

A = E a HA(a), (11.124)
a

which is assumed to be discrete for simplicity. Here and in the following we write
HA (a) for the projection onto the eigenspace of the operator A belonging to the
eigenvalue a.
532 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

If we introduce the spectral decomposition of A into the expression (11.122)

for the operation we get

= E(a(1), a(2) le — ' 15 1p = 0,4. = 0) IIA(a)

a

= E(a(1), a(2) 1P = 0, 4 = a) 11A (a)• (11.125)

a

In the second step we have used the fact that exp(—iaï") is a translation operator
which shifts 4 by the amount a, since P is canonically conjugated to Q. Now,
the remaining matrix element vanishes unless a( ' ) — a( 2) -=-- a,
0 , 4 = a ) = 6(a(i) — a2 — a) , (11.126)
which gives

Q(a(1) ,a (2) ) = E (5(a (1) — a(2) — a) IIA(a). (11.127)

a

This equation clearly shows that the operation depends only on the difference
of the readouts and that the possible values for this difference coincide with the
eigenvalues of A. The correctly normalized operation pertaining to the outcome
a = a( 1 ) — a(2 ) can therefore be written as follows,

12(a) -=-- HA (a), a E spec(A), (11.128)

showing that the operation is just equal to the corresponding projection of the
spectral family of the operator A.
Thus, we see that the EPR entangled probe state allows an ideal quantum
non-demolition measurement of the non-local observable A. The operation de-
scribes the back action on the quantum object which is in full agreement with the
von Neumann—Liiders projection postulate: The readouts yield the eigenvalues
of A and if the initial state of the object system was is an eigenstate of A then
this state is not changed by the measurement.
To appreciate what has been achieved it is important to realize the following
properties of the above measurement scheme. First, it must be emphasized that
the measurement device does not yield any information on the local observables
A( 1 ) and A(2 ). This is connected to the fact that neither Q( 1 ) nor Q( 2) are sharply
defined in the initial probe state, since these observables do not commute with P.
Only the relative coordinate "0 which does commute with P is sharply defined.
The entangled probe state 0.) is thus not appropriate for a measurement of the
local observables A( 1 ) and A( 2) . Consider, for example, the local measurement at
x (2) and the foliation .7' of the previous subsection (Fig. 11.11). If we introduce
the spectral decomposition of the local observable
4( 2 ) _ E a(v2 ) ,t2 ) (11.129)
,
 NON-LOCAL MEASUREMENTS AND CAUSALITY 533

we find for the density matrix describing the quantum object after the first
measurement at

pcati) ----
-
Ellpp(ovv. (1 1.1 30 )
v

This is precisely the density matrix as it would be obtained in a non-selective

measurement of the local observable A( 2 ). It follows that the measurement of
Q( 2 ) does not give any information on A(2 ): The quantum object behaves as
if we measure A( 2 ) and immediately erase that information. In particular, the
possible outcomes a( 2 ) are not given by the eigenvalues of A( 2 ), but are uniformly
distributed and independent of the initial object state.
Thus we see that the device allows the measurement of A ---- 4( 1 ) — 4( 2 )
without measuring the local quantities A (1 ), 4 (2 )• Of course one can also measure
any linear combination of the form

A .=-- aA (1 ) + 34 (2) . (11.131)

This is achieved by an appropriate replacement of the local observable, A( 1 ) -4

ctA( 1 ), A( 2 ) -4 —fiA( 2 ). Recall that our basic assumption is that, at least in
principle, we can measure any local observable. Similarly, one can also measure
any product

A -= A(1) • 4 (2) (11.132)

provided both local observables have a definite sign, for example A( 1 ) > 0,
A( 2 ) > 0. Namely, in that case we may replace A( 1 ) -4 ln A( 1 ), A (2 ) -4 ln A (2 ).
However, as we shall demonstrate later, it is not possible to measure all observ-
ables belonging to the algebra of operators of the object system.
The second important point to be noted is that the device allows the mea-
surement of non-local observables which cannot be measured locally. Consider
the observable A = A( 1 ) + A( 2 ). Suppose first that A is non-degenerate. In that
case we can measure A also locally by simply measuring A( 1 ) and A( 2 ) separately
(with the help of a probe which is in a direct product state). Such a measurement
will then also be a QND measurement of A: The corresponding readouts
a(2 ) are the eigenvalues of A (1 ) and A( 2 ), respectively, and a = a( 1 ) + a( 2 ) is an
eigenvalue of A. Moreover, all eigenstates of A are also eigenstates of A( 1 ) and of
A(2 ), which follows from the fact that the local observables commute and that A
is non-degenerate. Therefore, the separate measurement of A(1 ) and A( 2 ) leaves
unchanged all eigenstates of A.
The situation changes completely if A is degenerate. The separate measure-
ment will then, in general, not be a QND measurement of A. To see this we
consider some degenerate eigenvalue a of A and two corresponding orthogo-
nal eigenstateslxi) and lx2). The non-local measurement procedure constructed
above clearly has the property that it leaves invariant the total subspace spanned
534 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

by x i ) and 1x2 ). However, we can always choose these states such that they are
simultaneous eigenstates of AM and A(2). We suppose that the corresponding
eigenvalues are different and consider the initial state

Ix) = 1 OW +1X2)) (11.133)

What happens then during separate local measurements of A( 1 ) and A(2 ) is that
this initial state goes over with probability to either x i ) or 1x2 ). The separate
measurement of the local quantities is therefore not a QND measurement of their
sum.
As an example, which will be used several times later on, we consider a system
of two particles with spin One particle interacts locally with the device at
the other at x (2). The total Hilbert space is, of course, given by the tensor product
7-1 = As our local observables we take the spin components of the particles
along the z-direction,
1 4 (2 )
A (1 ) = —o- (1 ) (11.134)

and consider the measurement of the sum of the spins along that direction

1/
A = — (o-(1 ) u (2)) E L. ( 11.135 )
2 '
A basis of eigenstates of Jz may be written in an obvious notation as
1
1j = 0, in = 0) = ( i+ )(1)1_)(2) _i_)(1)1+)(2)), (11.136)

1,7n = 0) = ( H) (1) H (2) +1 41)142)) , (11.137)

j = i , rn = + 1) i+)(1)1+)(2), (11.138)
Ii = i l m = -1) = 14 1 )14 2 ), (11.139)
where in denotes the eigenvalue of J. These eigenstates have been chosen to be
simultaneous eigenstates of the square of the total spin

1_ 1 ( 6(1) _.(2)
(11. 1 40)
2 u '

that is fli,m) = + imi,m). The states 1i = 0, m = 0) and 1i = 1, m = 0)

span the two-fold degenerate eigenspace belonging to the eigenvalue 0 of J.
The non-local measurement using an EPR-type probe state allows the QND
measurement of L. In a separate measurement of the local spins of the particles.
however, the singlet state 1j = 0, in = 0), for example, will be transformed into
either 1 +)(1)I_ ) (2) or 41) +)(2).
 NON-LOCAL MEASUREMENTS AND CAUSALITY 535

Let us now demonstrate that besides linear combinations and certain prod-
ucts we can also measure arbitrary modular sums of local observables. The fact
that this type of non-local observable is measurable at least in principle, will
play an important rôle in the further development of the theory.
Obviously, it suffices to construct an explicit measuring device for the quan-
tity

B = A mod z E (A (1 ) — A(2) ) mod z, (11.141)

where z > 0 is some real number. To measure B we employ the following initial
probe state,

10) = E o, = nz), (11.142)

where the sum runs over an appropriate subset of the set of integers and AI
is some irrelevant normalization factor. The probe state is a superposition of
EPR entangled states. In contrast to the probe used before, not only are the
local coordinates Q(') and Q( 2 ) undetermined in this state, but also the relative
coordinate (2 is only determined up to a multiple of z, that is (0 mod z)10) = O.
Proceeding in precisely the same manner as above we now get the following
expression for the operation,

1l(a(1) ,a (2) ) = E E(a (1) , a(2) LI) --= 0 , a + nzglA(a). (11.143)

n a

Correctly normalized, the operation takes the form

(a ' , a 2 ) IT4(a). (11.144)

Here the sum extends over all a and n with the constraint of a fixed value for
the quantity a +nz = a(') —a( 2 ). This shows that the operation depends, in fact,
only on the quantity a mod z b and we may write the operation as

Q(b) = E HA(a) 1-1B(b). (11.145)

a mod z=b

Hence, the operation Q(b) is equal to the projection onto the eigenspace belonging
to the eigenvalue b of the modular sum B. The use of the probe state (11.142)
therefore allows the measurement of the modular sum of local observables. We
remark that it is not necessary that the sum over n in the initial probe state
extends over all integers. In the following we shall apply the measurement of
modular sums only to bounded operators. In that case it suffices if the sum runs
over a finite number of integers, namely those which project the eigenvalues of
A into the fundamental interval [0, z).
536 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

11.3.3 Quantum state verification

The discussion of the preceding subsection raises the question of whether we can
measure all non-local observables by an appropriate design of quantum probes
and by employing various local interactions between probe and object system.
Surprisingly, the answer to this question is that we cannot measure all non-local
observables, that is we cannot design an ideal, quantum non-demolition mea-
surement for each observable belonging to the algebra of operators of the object
system. In fact, as we shall see the measurement of most non-local observables
is incompatible with the causality principle. In the following sections we shall
formulate the restrictions on the measurability of observables which are imposed
by causality and investigate some classes of non-local operators and states which
are measurable.
For a clear and systematic treatment of the subject the concept of a so-called
state verification measurement turns out to be extremely useful. This concept is
more general than the usual notion for the measurement of states which is used
in quantum mechanics. In the present subsection we define state verification
measurements and describe them in terms of operations and effects within the
framework of generalized measurement theory developed in Section 2.4.2.
A state verification measurement of a given state Iwo) is defined to be a mea-
suring device which performs a Yes/No-decision experiment with the following
property. Let IT) be any state of the underlying Hilbert space and

1 41 ) = Olio) + 01 41 ±) (11.146)
its decomposition into the component parallel to Iw o ) and the component per-
pendicular to it, such that ( 4101 41 ± )--=-- 0 and liar + 101 2 = 1. If prior to the
measurement the system is in the state III)) then the measuring device responds
with the result Yes with probability I1 2 , and it responds with the result No with
probability 1012 .
Analogously we define a QND state verification as a state verification mea-
surement with the following additional property. If the result is Yes then the
system is in the state 141 0 ) after the measurement. This implies that the initial
state Iw o ) is left unchanged by the measurement since the result Yes is then ob-
tained with certainty. In the case that the measurement yields No the system is
in a state 'WI) which is orthogonal to IWO, but which is not necessarily equal to
the orthogonal component I W1) of the initial state. The latter property implies
that on repeating the experiment we again find No with certainty.
It should be clear that quantum state verification measurements do not, in
general, represent measurements of the projection

HO = 1 410)( 11j01 (11.147)

onto the initial state in the conventional sense of quantum mechanics. In the
case of a QND state verification the orthogonal component is allowed to change
during the measurement, and for a general state verification measurement no
 NON-LOCAL MEASUREMENTS AND CAUSALITY 537

assumption at all is made regarding the behaviour of the state vector of the
system.
We now describe the form of a general state verification using the language
of operations and effects. According to its definition the device has a set fal
of Yes-readouts and a set {b} of No-readouts. Correspondingly we have a col-
lection of Yes-operations l yes (a) and of No-operations N0 (b). The behaviour
of the density matrix of the object system as a result of the non-selective state
verification can then be described by

p E Qyes (a)pqes (a) E 1N 0 (b)/A2tN0 (b). (11.148)

a

As we saw in Section 2.4.2 this is, apart from the assumption of discreteness of
the sets of the readouts, the most general setting. In particular it includes the
possibility of quantum probes in mixed states and those of incomplete, approxi-
mate, or non-selective measurement of the final probe states.
On introducing the Yes- and the No-effect,

Fyes„Eqes(a)1lyes(a), (11.149)

FN0 E Q-Ï,,o (b)Q,,,„(b), (11.150)

we get from the conservation of probability

FYes FN0 = I. (11.151)

Let us consider the initial state IT) and its decomposition (11.146). We can write
the probability for the outcome Yes of the state verification of Iwo) as follows

PYes(T) = ( 41 1 11YesI T )
= Ict1 2 (W0dFYes1 41 0) 1/3 1 2 ( T _LIFYes1 41 ±) (a0 * ( 41 "IFYesH11 0) + c.c)
ictr. (11.152)

The last equality expresses the condition for a state verification measurement
and must be true for all a, 0, and for all 14/±). Setting first 13 = 0 and then
a = 0 one easily deduces the relations

(410117yes 14/0) = 1, (4/"Wyes IxIii) = 0. (11.153)

If these relations are inserted into eqn (11.152) one also finds that

(T_L1FYes00) = 0. (11.154)

It follows that for any state verification measurement the Yes-effect is equal to
the projection onto the verified state,
538 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Fyes = Ho E 1W0)(Tol. (11.155)

With the help of eqns (11.153) one is led to the following equations,

flyes (a)0 ± ) = 0, (11.156)

and

1 rsTo(b)1 410) = O. (11.157)

Equation (11.156) holds for all a and for all states perpendicular to the verified
state, while eqn (11.157) is valid for all b.
Summarizing we see that for any state verification measurement all Yes-
operations annihilate the No-states, all No-operations annihilate the Yes-state,
and the Yes-effect is equal to the projection onto the Yes-state. With the help of
the conservation of probability it follows also that the No-effect is equal to the
projection onto the orthogonal complement,

FNo Ho E H 1. (11.158)
In order for the state verification to be a QND measurement two further condi-
tions must be satisfied. First, when applied to the state IWO all Yes-operations
must yield a state which is proportional to Iwo). Hence, only a single Yes-
operation is required which must be equal to the projection onto the verified
state,

QYes — Ho. (11.159)

The second additional condition for a QND state verification is that all No-
operations must map the subspace orthogonal to Iwo) into itself.
These results will be applied in the next subsection to the state verification
of non-local states.

11.3.4 Non local operations and the causality principle

-

In the present and the following subsection we study the measurement of some
quantum system which consists of two localized parts (1) and (2) described by
the Hilbert spaces 7-1 (1 ) and 71 (2 ), respectively. The total Hilbert space of the
quantum object is the tensor product 74 = 7 ((') 7l( 2 ). We investigate the state
-

verification of an entangled state IWO in this space. Introducing appropriate

local bases lx, 1) ) and IV ) ) in both parts of the system a given state Iwo) can be
represented in terms of its Schmidt decomposition as follows (see Section 2.2.2),
D

iTo) E ix(,2 )). (11.160)

We assume here and in the following that Wo ) is normalized and that its Schmidt
number, denoted by D, is finite. Thus, the Schmidt decomposition consists of a
 NON-LOCAL MEASUREMENTS AND CAUSALITY 539

finite number of terms and we assume that the complex numbers ai, i = 1,... , D,
are different from zero,

E 1a/1 2
i=1
= 1, c (1 1. 161)

For D > 2 the state IT O ) is an entangled state. We recall that T o ) is called

maximally entangled if the absolute values of all non-vanishing coefficients in
the Schmidt decomposition are equal to each other, that is if la,d = lArb- for
all i = 1, , D.
We further denote by ( 1 ) and 74(2) the local subspaces spanned by those
basis vectors that occur in the Schmidt decomposition of ITO with a non-zero
coefficient,

= span x(. 1) )1 (11.1 62)

J ,D

971 (2) = span x(. 2) ) ( 11. 16 3)

k i=1,...,D

The tensor product 3--i (1 ) fi (2 ) is thus a D2 -dimensional space.

11.3.4.1 Formulation of the causality principle We consider the setup sketched

in Fig. 11.12. The initial state of the system may be any state IT). At two
spacelike separated points x (1) and x(2 ) a state verification measurement of the
state IT0) is performed. In the localized region belonging to part (I) of the
system we perform a measurement of some observable B( 1 ) referring to the local
variables in that part. In the following we denote by E [B( 1) 141] the expectation
value of B( 1 ) for this local measurement performed after the state verification,
under the condition that the initial state is IT).
Suppose now that some local interaction may be applied on part (2) of the
system prior to the state verification measurement. This can be described by
some unitary operator U(2 ) acting on the Hilbert space 7/( 2) • For example, if the
quantum object constitutes a spin system one may regard U(2) as describing the
interaction of the system with a magnetic field by which an experimenter can
flip the spin of the particle in part (2). Since we assume that anything can be
measured locally we may also assume that any unitary transformation can be
realized by an appropriately chosen local interactions in each local part of the
system.
We assume that the measurement of B( 1 ) is carried out in a space-time region
which is separated from the interaction region of U(2) by a spacelike interval. The
causality principle then implies that the result of the local measurement carried
out in part (1) after the state verification, does not depend in any way on the local
interaction which is applied to the system in part (2) before the state verification.
This means that the expectation value of any observable B (1 ) determined after
540 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

local measurement of B( 1)

non-selective
-I,-
verification of ITO

local interaction U (2)

liP)

FIG. 11.12. Illustration of the causality principle expressed by eqn (11.164).

After a verification measurement of xli o ) a local observable B( 1 ) is measured
on part (1) of the system. Prior to the verification measurement a local inter-
action described by the unitary operator U( 2 ) may be applied to part (2) of
the system. The measurement of B( 1 ) and the interaction U (2 ) are separated
by a spacelike distance such that the expectation value of B (1 ) must be in-
dependent of whether the interaction U( 2 ) is or is not applied to the system.
The state verification itself is described on the non-selective level.

the state verification must be the same regardless of whether U (2 ) is applied to

the system or not, that is we must have

EPP) 024]
( =
(11.164)
This equation expresses the requirement of the causality principle on the state
verification measurement.
It should be clear that the state verification must be described here as a non-
selective measurement. The reason is that any local observer in part (1) does
not know the result of the verification of the non-local state 14/0) at the time
when B( 1 ) is measured. However, if eqn (11.164) were not always true a local
observer in part (2) could transfer information from part (2) to part (1) with a
superluminal speed. Thus, the expectation value is given by the expression,

E[B (1 )14i] ---- E(41191es(a)B("Yes(a)41)

a

+ E(WAtro(b)B (1) SIN.(b)1 41 ). (11.165)

b

The causality principle formulated by eqns (11.164) and (11.165) leads to

important consequences for the measurability of non-local quantities. It must be
 NON-LOCAL MEASUREMENTS AND CAUSALITY 541

emphasized that eqn (11.164) is trivially satisfied if we assume, as we of course

do, that only local object—probe interactions are involved in the measurements.
In fact, as is easily seen eqn (11.164) becomes trivial if we write it in terms of
the state vector (or density matrix) for the total object—probe system. However,
important conclusions can be drawn from the above formulation of the causality
principle if it is combined with the conditions for a non-local state verification
measurement and with certain properties of the underlying Hilbert space of the
combined object system.

11.3.4.2 Theorem on the erasing of local information In the following we prove

a theorem by Popescu and Vaidman (1994) which can be derived from the causal-
ity principle (11.164) and which leads to several interesting conclusions on the
possibilities of non-local measurements. This theorem states that

E[B (1) I T] = E[B (1) 013] for all IT) E "ii (1) 0 W (2) . (11.166)

This means that the expectation values for all local observables in part (1) are
independent of the initial state IT) prior to the state verification measurement.
In other words, after the state verification a local observer in part (1) cannot find
out by the measurement of local observables BO) the initial state of the system
prior to the state verification. After the state verification there is thus no trace
of the initial state in the local density matrix referring to the variables of part
(1) of the system. This property is called the erasing of local information by the
state verification measurement.
It must be stressed that the erasing of local information refers to initial
states IT) which belong to the subspace 74 (1) 0 ?-1 (2) . For an arbitrary 4/) E
10.) e, W2) eqn (11.166) is, in general, wrong. As an example we take T o ) =
1X(1))01X(2)). This product state can obviously be verified by performing separate
local measurements of the projections onto the states lx( 1 )) and lx (2 )) in part
(1) and part (2) of the system. Consider further the state IT) = 10 1) ) 0
where 10(1)) is orthogonal to Ix (1) ). Since the chosen state verification is an
ideal quantum measurement we have the conditional expectations E[B( 1 )I T o ] =
(o w 1B(1)10(1)%) which are, in general, not equal
(X(1)1B(1)1X(1)) and E[B(1)I4'] __,
to each other.
To prove (11.166) we first evaluate the expression for the conditional expec-
tation value (11.165). Using the decomposition (11.146) of IT) and taking into
account the properties (11.156) and (11.157) of the state verification measure-
ment we obtain,

E[B (1) I T]
= ice E(4'019Yes (a)B (1) S2 y„(a)1 410) +101 2
b
== I a l 2 Ea[B(1) I 41 01 ± 10I 2 E[B (1) I 41-Li• (11.167)

Thus we find
 542 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

E[B (1 )I 41] = E[B (1 ) I To] I01 2 (E[B(1)1411] E[B (1) R o ] ) . (11.168)

If IT) has a non-vanishing orthogonal component, that is if 0 in the decom-

position (11.146), the last equation tells us that E[B( 1 )IT] is equal to E[13 (1 )1To ]
if and only if E[B( 1 )IT1] is equal to E[B( 1 )IT ob
E[B (1) IT] = E[B (1) 141 0] < E[B (1) ITij E[B (1) 1410]. (11.169)
For the proof of eqn (11.166) the following strategy is used. In a first step one
demonstrates that any state IT) E 12 (1) 0 71 (2) can be represented in the form,

IT) = Ecro.2)00), (11.170)

r=1
where the cr are c-numbers and the 0. 2) are unitary operators which act on the
local space 1-1 (2) of part (2) of the system. This means that any such state vector
can be represented as a superposition of a finite number of states which can
be generated from 'T o ) by the application of unitary transformations which act
only on part (2) of the system. In the second step of the proof one demonstrates
that any state of the form (11.170) obeys eqn (11.166). In this second step one
uses the causality principle (11.164) as well as the property (11.169).
The existence of the representation (11.170) will be demonstrated by an ex-
plicit construction of the local, unitary operators Ur(2). The space 14 4 ) ® 'HP) is
spanned by the basis vectors

1)d° ) g N 2 ), 1 , 1, 2, ... . (11.171)

Note that the restriction k < D stems from our requirement that IT) belongs to
the space 74 (1 ) 7-1 (2) . As we have remarked already the theorem is not true for
all state vectors belonging to WO 'H (2 ). Obviously, it suffices to demonstrate
that any basis vector (11.171) can be represented in the form (11.170). This is
achieved with the help of two unitary transformations (4 2) and U2(2) which act
in the space 7/ (2) and which are defined by

(42) 1x(k2) ) = 1X i( 2) ))
U1 2) Ix 2) ) = IX k( 2) ),
U1 2) 1Xi(2) ) = for i
and by

ti 2) 1X k( 2) )
[J 2) 1Xi 2) ) =lx k( 2) ), for k 1,
[J 2) 1V ) ) = Ix 2 for iX k,l.
(2) .
Thus, U1 simply exchanges the states lxk(2)) and lxi(2)), and U2(2)exchanges
these states and introduces an additional minus sign. For k = 1 the operator
 NON-LOCAL MEASUREMENTS AND CAUSALITY 543

(2) ( 2)
Ui sequal
i to the identity, whereas U2 multiplies the k-th basis vector by —1.
Applying these operators to Iwo) we get for k Xl

14 2) 00) =dh 1 lkiX k( 1) )1Xi2) ) d- crilXi i) )1)(2 ) ) dh ailX i( 1) )1X(i 2) ), (11.178)

1.1 2) R 0) == — CviciX k( 1) )1Xi2) ) ailXi i) )14 ) ) dh ailX(i1))1X(i2)). (11.179)

It follows that

u 2 [ 'I'0 ) -U2 1 0) = 2ak 141 ) ® le), (11.180)

which gives a representation of the required form, namely

IX k( 1) ) ® le) = ( —L ) Iwo). (11.181)

2ak U 2)
2ak W ) —1---
As is easily seen, the case k =1 leads to the same equation. Note that ak 0 for
k =1,... ,D due to our assumptions on IT) and Iw o ). This concludes the proof
of eqn (11.170).
We now have to show that any IT) of the form (11.170) satisfies eqn (11.166).
This is done by induction over N, the number of terms involved in the represen-
tation. For N = 1 we may set c1 1 since IT) is normalized. Equation (11.166)
then follows immediately from the causality principle (11.164).
Assume that (11.166) holds for all states with a decomposition of the form
(11.170) involving N terms. Consider some normalized state whose representa-
tion of this form contains N + 1 terms,
N+1
IT) = E cro2)14'0). ( 11. 1 8 2)
r=1

For brevity we write E [T] instead of E[B (1 )I4'] in what follows. According to the
causality principle (11.164) we have

E[T] = E [(4)±1 ) 4' ] (11.183)

where
—1 —1
([11)
+ ) = Ecr (2)
(uN±, ) Ur(2) I* ) CN+114 0) (11.184)
r=1

If the sum over r in (11.184) is zero we have already proved the theorem. Let us
therefore assume that it is non-zero and introduce the normalized state

= .1V Cr ((4)±1 ) 0 2) 0) (11.185)

T=1
544 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

with an appropriate normalization factor .111-. The induction assumption then

gives

E[T] E[4/ 0]. (11.186)

Next we decompose the state lit),

= To) + 01 41 ±), (11.187)

where we may assume that 0 0 since otherwise the theorem follows immedi-
ately. Equation (11.186) together with the property (11.169) yields

E[TL] E[4'0]. (11.188)

In view of eqns (11.184), (11.185), and (11.187) we also have the decomposition,
-1 a
((42+) 1 ) IT) = ( + eN+1) Iwo) + :A7R-L) ai lT0) + 01 4'1). (11.189)

Since 0' 0 (see above) we may now use eqn (11.188) to conclude with the help
of (11.169) that

E [(Ur+i ) 1 xli] E[4,0], (11.190)

from which the theorem follows by making use of eqn (11.183).

It should be noted that the theorem has been proven under quite general
circumstances. As physical assumptions we have merely used the existence of
the operations and effects describing a state verification measurement as well as
the causality principle.

11.3.5 Restrictions on the measurability of operators

Let us draw some important conclusions from the theorem on the erasing of
local information with regard to the measurability of non-local operators in the
combined Hilbert space of the system.
The first conclusion is that the projections 110 = 14/0)(4/01 onto entangled
states Iwo) are not measurable. To prove this statement we derive a contradiction
to the causality principle. Let

I T 0) := Ecfilx(i 1 ))® be) (11.191)

be the Schmidt decomposition of Iwo ). Since we assume that this state is en-
tangled, at least two of the coefficients ai are different from zero, that is we
have ak,a1 0 0, for some pair of indices k 0 1. A QND measurement of the
projection 110 means that one has just two operations, namely Sly„ = 110 and
 NON-LOCAL MEASUREMENTS AND CAUSALITY 545

No H 1 -= I — Ho. It follows that not only 00) but also any state IT_L)
orthogonal to it must be left unchanged by the measurement, that is

E[B (1) 141o] ( 4'olB (1) 1 41o), (11.192)

E[B(1)Ri] = ( F-LIB(1)1411). (11.193)

The erasing of local information expressed by eqn (11.166) then implies that

111±1B (1) 141 ± ) = I -13 ( 1) 1 41 0) (11.194)

must hold for all states which are orthogonal to Ro ) and which belong to the
space i-i (1) 1-1 (2) . Now we take IT") IX k( 1) ) Ixi 2) ) and choose B( 1 ) =
This leads to

(T-L1 13(1) 1 41 ") = (41) IB (1) I41) ) 0, (11.195)

( 41 01B (1) 1 41 0) Elaii2(x 1 )1B(i)ix,(i 1 )) = icf,12 0, (11.196)

which is incompatible with eqn (11.194). This shows that a measurement of the
projections onto an entangled state would contradict the causality principle.
We note that a contradiction to causality only emerges if at least two of
the Schmidt coefficients ai are non-vanishing, that is only if ) is an entangled
state. Of course, all projections onto product states Nk(1))01X/(2)) are measurable.
As an example one may think of the square of the total spin operator f (see
eqn (11.140)) for a non-local system of two spin4 particles. The operator f 2
has a three-fold degenerate eigenspace belonging to the eigenvalue 2, and a non-
degenerate eigenstate corresponding to the eigenvalue 0. The non-degenerate
eigenstate, namely the singlet state I i = 0, in = 0), is an entangled state. A
measurement of f 2 is thus equivalent to a measurement of the projection onto
an entangled state. Thus, the operator f 2 is not measurable.
These considerations can be generalized as follows. Let A be some measurable
observable in the space 1-1 =11 (1) 01-1 (2) with at least one non-degenerate eigen-
value a and a corresponding entangled eigenstate IAN of the form (11.191) with
Schmidt number D > 2. As before we define 14( 1 ) and 171( 2 ) to be the subspaces
(1) (2 )
spanned by the respective basis vectors Ix i ) and Ix i ), where i = 1, . . . , D. If
we further suppose that A leaves invariant the subspacel---1 (1) 014 (2) the causality
principle leads to the following conclusion: The non-degenerate eigenstate ITO
as well as all other eigenstates of A in the space 71i (1) ® ?I-( (2) must be maximally
entangled.
To prove this statement we introduce a basis IT,), v = 0, , D2 — 1, of
eigenvectors of A belonging to the space 14 (1) 011 (2). The state given by y = 0
is just the eigenstate IT() ) corresponding to the non-degenerate eigenvalue a.
Now, any measurement of A is necessarily a state verification of Ro ). To
cast the causality condition into an appropriate form we introduce the density
546 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

p(1) ( F) which is defined to be the local density matrix describing the variables
of part (1) after the state verification, under the condition that the initial state
was Ii11). This density is obtained by taking the partial trace over part (2) of the
density matrix of the total system after the state verification,

p(l) (T) = tr (2) E 9yes(a)R) intyes (a) + EQN.(b)o)(4fotNo(b)} •

a
(11.197)

The erasing of local information expressed by eqn (11.166) can then be re-
formulated as
p (i) (4, ) p(i) (4,0 ,) , for all 1 4') C 7:1(1) 0 7-1(2) . (11.198)

Since all I AF) belong to the space 14 (1 ) 014 (2) we have for all

= (11.199)

The A measurement leaves unchanged all eigenstates. Thus we conclude

tr (2) {1 4/0)( 4/01} = tr (2) fi lli v)( 41 u II • (11.200)

Summing this equation over u and taking into account that the I ■liv ) form a basis
in the space 14 (0 014 (2 ) we immediately get,

D2 -1
D2 • tr (2) {141o)(410 = tr (2) E ixpo (Tv'
v=.
tr (2) { (1 ) 0/.7t (2) }
= D • If_1( 1 ) , (11.201)

from which it follows by using the Schmidt decomposition of 141 0 ),

D
tr (2) {141 0)(4, 01} = E 1 4(1).
b- (11.202)

This proves that ad = 1/V7D for all i = 1,... ,D, that is, that the non-
degenerate eigenstate ITO is maximally entangled. Equation (11.200) now tells
us that this must be true also for the other eigenstates II'). Thus we see that.
in fact, all eigenstates of A in the subspace 14 (1) 01( (2) have a Schmidt decom-
position of the form
D
1
ov) E fiv,iix(v1,)i) ® ix(v2,b,
i=1
where I= (11.203)

with appropriate local basis vectors Ix) and lx v2 ).

 NON-LOCAL MEASUREMENTS AND CAUSALITY 547

As our final application of the causality principle we shall give a complete

characterization of all non-degenerate, measurable observables for our previous
example, namely for a non-local system of two spin 4 particles (Popescu and
Vaidman, 1994). Thus we let A be a non-degenerate observable in the composite
Hilbert space of these particles. Obviously, there are two possible cases: Either all
eigenstates are direct products, or else there is at least one entangled eigenstate.
If all eigenstates are direct products one can easily verify that it is possible
to choose appropriate local bases such that the eigenstates of A take on the
following form,

IT1) = (11.204)
1 412) - I - z)1 + z'), (11.205)
ITO = I + zn i - z'), (11.206)
1 414) =1 (11.207)

where we denote by I ±n) the eigenstate of the spin component along the direction
n, with corresponding eigenvalue ±1. The causality principle then leads to the
conclusion that the direction z" must be parallel or antiparallel to z, which means
that the eigenstates of A can always be cast into the form,

liTY = I + ,z)1 + zi), (11.208)

1 4' 2 ) = I - z )1 + z'), (11.209)
IT3) = I + ,01 - z'), (11.210)
IT4) - I - z) I - z'). (11.211)

For the proof of this statement we take the local observable B( 1 ) = I - z) (-z1.
Since IT') is an eigenstate of A,

E[B (1) 14/1] --= (T1lB (1) 141 1) --= 0. (11.212)

For the state VD) --= I + z)I — z') the causality principle yields

E[B (1 )I(D] --= E[B (1) 14/ 1 ] --= 0. (11.213)

On the other hand, since A is assumed to be a measurable operator we must

have
4
E[B (1) 01 ' E oitForoFilB(1)141i)
= i(-Ez i + z" ) 1 2 . I(-zi + z")1 2 +1(+zI — z")1 2 ' K — zi — z")1 2
--= 0, (11.214)

showing that indeed z" -= ±z, as claimed.

548 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Let us now consider the case that at least one eigenstate is an entangled
state. Our general considerations then reveal that all four eigenstates of A must
be maximally entangled. By a suitable choice of local basis vectors one can
therefore always cast these eigenstates into the following form,
1
ITO = (I + z)I z') — I — z)I + z')) (11.215)
1
1 412) = —2 (I + z)I z i ) + I — z)I + z 1 )) (11.216)
1
IT3) = (I + z)I + z') — I — z)I z')) (11.217)
1
= —2 + 1 + z') + I z) — z')) (11.218)
A non-degenerate operator with eigenstates of this form is called a Bell-state
operator.
Thus we have shown that the set of all measurable, non-degenerate opera-
tors decomposes into two classes: Either all eigenstates are direct products of
the form (11.208) or else all eigenstates are maximally entangled states of the
form (11.215). What we have demonstrated is that a measurement of any non-
degenerate operator which does not belong to one of these two classes would
contradict causality. It remains to be shown, however, how those operators whose
eigenstates are of the above form can be measured.
It is obvious that any non-degenerate operator with eigenstates of the form
(11.208) can be measured. In fact, this is achieved by two separate measure-
ments carried out locally in both parts of the system. We now demonstrate that
also any Bell-state operator, that is any non-degenerate operator A with eigen-
states of the form (11.215) can be measured. This will be done with the help of
the measurement of two non-local observables A1 and A2 (see Fig. 11.13). We
construct these observables and verify that they represent operators which we
already know to be measurable.
The A measurement must be a QND state verification of all eigenstates 1 .
We denote by no the subspace spanned by 1 if 1) and 14/2), and by H 1 the subspace
spanned by 14/ 3 ) and 14/ 4 ). With the help of the first measurement of A1 we want
to find out whether the initial state belongs to Ho or to H. To this end, consider
the operator
1
(11.219)
In the subspace Ho it has the eigenvalue 0, whereas II 1 is spanned by the eigen-
states corresponding to the eigenvalues ±1. Thus we have J? -= 0 in Ho, and
4 1 in 11 1 . The relation J? J", mod 2 tells us that J? belongs to the class
of measurable operators. Thus, our first measurement is the measurement of the
quantity

A1 J, mod 2. (11.220)
 NON-LOCAL MEASUREMENTS AND CAUSALITY 549

A2 --= Jx mod 2

=J mod 2

FIG. 11.13. A measurement device which enables the measurement of

the Bell-state operator with the non-degenerate eigenstates given by
eqn (11.215). One uses two quantum probes by which the non-local quan-
tities A 1 = Ji, mod 2 and A 1 = Jx mod 2 are measured.

Through the second measurement we must be able to distinguish the states

1 4'1,2) in Ho as well as the states 14/3,4) in IL. To find an appropriate observable
to achieve this we first note that

0-1) 1 1 z) = i T z), a x( 2)I 1 z i ) = I + z i), (11.221)

(1) (2)
which shows that the operator ax (Tx, flips the spin of both particles. Now, I T 1 )
and 111'3 ) are odd, whereas 1412) and 14/4) are even under a spin flip of both par-
ticles. Thus, a measurement of o-V-) crx(2, ) together with the previous measurement
of A1 enables us to verify all four eigenstates, which constitutes a QND measure-
ment of the Bell-state operator A. To show that o-V-) o-x(2, ) is indeed measurable
we write it as follows,
=2 1, (11.222)
J
where

Jx = v-
1 l (1)
x
(2)
+ g,x
■
)• (11.223)

Thus, the measurement of a-1) o-12, ) is equivalent to the measurement of J. Again,

we have 4 = , J- x mod 2 and, hence, the second measurement to be performed is
that of the observable
550 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

C AN

(
I
I
, ' A1

I W)
FIG. 11.14. General measurement device for a QND state verification mea-
surement. The scheme involves the measurement of non-local quantities
A1 , A2, . . . AN described by the operations S2 1 (ai), 12 2(a2), • • • , nN(aN)•
,

A2 = ,ix mod 2. (11.224)

Thus we have constructed explicitly a scheme which allows the measurement

of the Bell-state operator A. It is clear that the device constitutes a QND mea-
surement of A: Take any eigenspace of A1 and any eigenspace of A2. Then both
eigenspaces are two-dimensional, orthogonal to each other, and have a common
one-dimensional subspace which represents an eigenstate of A. It follows that all
eigenstates of A are left unchanged by the measurement. This can also be seen
by noting that the non-locally measured quantities commute, [ A 1 , A2] = 0, as is
easily verified.

11.3.6 QND verification of non local states

-

In the previous section we have discussed the measurability of operators in the

tensor product space 1-1 (1) 0 1-1 (2 ). Let us turn to the QND verification of entan-
gled states in this space, that is we now consider the verification of single state
vectors, rather than the measurement of operators. We would like to answer the
question of which type of entangled states allows a QND verification and how
such measurements could be carried out. Again we invoke the causality princi-
ple which enables one to exclude the measurability of a large class of entangled
states.
We consider the following general measurement scheme (see Fig. 11.14). The
measurement consists of a sequence of N measurements of the (in general non-
local) observables A1 , A2, . . . , AN. Each A, is measured with the help of an
appropriate (possibly entangled) quantum probe as described in Section 11.3.1.
We suppose that all A n, are of the form of those operators which were shown to
 NON-LOCAL MEASUREMENTS AND CAUSALITY 551

be measurable. Denoting the operation corresponding to the nth readout a n, by

S2,(a11 ) the total operation for the device takes the form
f/(a i , , aN) = 52N(aN)9N_1(aN-1) ••• Q1 (cti), (11.225)
where the time-ordering is understood. As for any measurement we have

E , aN ) = I, (11.226)
al ,aN

which follows immediately from the time-ordering and with the help of the prop-
erty,

E 1en (an )1,,(an ) = I. (11.227)

an

For all measuring procedures considered here we require that the operation
is a function of the measured quantity. This implies that all operations commute
with their adjoint,

),1-42 (a, )1 = 0. (11.228)

This equation, in turn, yields that in addition to (11.226) we also have

E ft(a i ,, aN) = I, (11.229)

al ,••• >aN

as is easily verified. Physically, this means that an initial density matrix which
is proportional to the identity is not changed by the measurement.
We now regard the measurement device as a state verification of some entan-
gled state It o ). Introducing the corresponding Yes/No-operations we thus have,
in addition to eqn (11.151),

E 12y„(a)f/ ty„(a) + E 9No (b) 12L0 (b) = I. (11.230)

a

We need one further condition. We will assume here that the operations
pertaining to the measurement leave invariant the space 1:1 (1 ) "i1 (2), that is
the space which is spanned by the basis vectors with a non-vanishing coefficient
in the Schmidt decomposition. In physical terms this means that the quantities
being measured do not excite new basis vectors, beside those already present in
the Schmidt decomposition. In view of this condition we may restrict the whole
discussion in the following to the subspace 94 (1) '14 (2) .
Our goal is to prove the following statement. For any entangled state IWO
which allows a QND state verification of the type considered above the causality
principle requires that all coefficients ai of its Schmidt decomposition (11.160)
must necessarily be equal to each other, that is ail = 1/N/75 . This means that
552 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

only maximally entangled states can be verified by such QND measurements

without contradiction to causality. In addition we are going to demonstrate that
and how these maximally entangled can be measured.
For the proof of the above statement we use the causality condition in the
form (11.198). Explicitly, that condition can be written with the help of the
Yes/No-operations as

tr(2) {I*0)(4101} (11.231)

tr (2) {E f2yes (a)1 41 )( 41 1 91 es(a) + E f/No(b)141)(41gro(b)}

a

Our strategy is similar to the one used in the previous Section 11.3.5: We choose a
set of orthonormal basis vectors ITO, u = 0, 1,... , D2 — 1, which span the space
'14 (1 ) 0 74 (2) , insert these basis vectors into eqn (11.231), and sum over v. The
essential difference to the case of a QND operator measurement is that we cannot
assume here that all basis vectors 14/,) are unchanged through the measurement:
In general, only ITO does not change in the QND state verification. Thus, instead
of eqn (11.201) we now have

D2 tr(2) {1410)(1oil (11.232)

= tr (2) E Yes (a)I ( l ) 04-i(2) qes (a) + E (b)4( 1 ) cot- (2)14\10 (b) .
a

However, in view of eqn (11.230) and of our requirement that the operations
leave invariant the space /71 (1 ) 014( 2 ) we can conclude from the last equation,

D 2 . tr (2) fig/0( 41 0i} tr (2) {4 ( 1 )074 ( 2 ) 1 --= D • 4(1), (11.233)

from which we obtain

1
tr (2) {Io)( 41 01} = Tko• (11.234)

This proves our statement, namely that in order to be measurable without con-
tradicting the causality condition, the state ITO must be maximally entangled.
Stated differently, after the QND state verification the local density matrices
referring to the local variables in part (1) of the system must necessarily be
proportional to the identity in the space 14 (1) . This clearly expresses once again
the erasing of local information: After the state verification there is no trace of
the initial state in the local mixtures, and the latter describe states of maximal
entropy.
What has to be demonstrated finally is that the maximally entangled states
can, at least in principle, be verified by some QND measurement. This will be
done by an explicit construction of a measuring device (Aharonov, Albert and
Vaidman, 1986). To this end, we first note that by an appropriate choice of the
 NON-LOCAL MEASUREMENTS AND CAUSALITY 553

phases of the local basis vectors one can always put a maximally entangled state
into the form
D
1
I xF 0) 0)) ® ixi( 2)). (11.235)
Nr.5 z
A verification of this state can be carried out with the help of two successive, non-
local measurements, similar to the device constructed to measure the Bell-state
operator.
The first measurement is that of the observable
A A (1 ) + A (2) , (11.236)

where the local observables A( 1 ) and A (2) are defined through

A mi x ,(i ' ) ) = (11.237)

4 (2) lx ,(i 2) ) = (11.238)

with i 1, 2, ... , D. Obviously, 14/0) is an eigenstate of A with eigenvalue 0.

Moreover, the vectors I xli) belonging to the D-dimensional eigenspace A = 0
have the general form
D

IT) = Eoilx(ii)) (11.239)

i=1

This shows that by the outcome A = 0 we verify that the state has a Schmidt
decomposition in the given local basis vectors Ix 1) )) IV ) ).
By the second measurement we want to verify that the coefficients Oi in
(11.239) are all equal to each other. To find an appropriate observable we consider
the local unitary operators U (1) , U (2) defined through

u'1 )Ix 4( 1) ) = i = 1, , D — 1, (11.240)

u( 1) 1 Xn = 1 (11.241)
U(2) I X (i 2) = 1X (i 2+)i) i 1, . . . , D 1, (11.242)
U (2) 1X i( ) 1X (12 5 (11.243)

These operators induce a cyclic shift of the index i. Introducing the unitary
operator U = U(1 ) U (2) we find that for any state (11.239) in the space A 0,
D
uixp) = E IV)) ix(i2)), (11.244)
• =1
where (30 OD . Thus we see that UlT) = I 4') if and only if all Oi are equal to
each other. It follows that I W0 ) is the only eigenstate of U corresponding to the
eigenvalue 1 in the space A = 0.
554 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Since U (1 ), U(2 ) are local unitary operators we can introduce local observables
B (1 ), B( 2 ) by means of

U(1) = exp {iB (1) } , U (2) = exp iB (2) } , (11.245)

such that, since BO) and B( 2) commute,

U = exp {i (B (1) + B (2) ) . (11.246)

From the above property of U we deduce that ITO is the only eigenstate of the
operator

B = (B (1) + B (2) ) mod 27r (11.247)

belonging to the eigenvalue 0 in the space A = O. As was demonstrated in Section

11.3.2 such modular sums are measurable.
This completes the construction of the measuring device: The state ITO is
verified by the outcomes A = 0 and B 0 of two non-local measurements. The
initial state Io) is obviously left unchanged by the measurement. Moreover,
since A and B commute it is easy to show that the device constitutes a QND
state verification measurement.

11.3.7 Preparation of non local states

-

We have seen above that causality imposes strong restrictions on the measur-
ability of non-local observables and states. In particular, we have found that a
normalized, entangled and non-local state with a Schmidt decomposition of the
form
D
IT ) Eadx(i 1 )) ® (11.248)

is not measurable by our QND state verification device, unless all Icd are equal
to each other, that is unless the state is maximally entangled. However, it is
important to observe that all states of the form (11.248) can be prepared by an
appropriate, non-local measuring device without contradiction to causality.
A device for the preparation of the states (11.248) can be designed as follows
(Aharonov, Albert and Vaidman, 1986). First we prepare locally the (normalized)
states

Po) ) =

=
 NON-LOCAL MEASUREMENTS AND CAUSALITY 555

which are then combined to form the initial state

, D
101) ) 0 10(2) ) 1 E aiiX (. 1) ) IX 3(.2) ) (11.251)

in the tensor product space li( 1 ) 71 (2 ). Both parts (1) and (2) of the combined
system may be separated by a spacelike distance.
In the second step we now measure the non-local observable A = A(') +
where the local quantities AP) and A(2 ) are given through eqn (11.237). The
outcome A = 0 of that measurement projects the initial state IT) onto the state
IT) whose Schmidt decomposition is of the desired form in the given local bases
(2)
i ) and lx, ), namely we have
I X (1)
D
-> 0) E aiix,(i1)) ix. (11.252)

Of course, there is no conflict with the causality principle since the measurement
of A involves only local interactions. However, the preparation of IT) is successful
only with a certain probability which is given by
D
1
P(A = = 1( 41 0)1 2 = 1
b- E lad (11.253)
T5*
The important conclusion to be drawn from the above considerations is that
for any measurement device one has to distinguish carefully between the states
that are measurable and those states that can be prepared by the device. It is
this possibility of preparation which justifies regarding the vectors (11.248) in
the tensor product space really as states of the combined system.

11.3.8 Exchange measurements

At several places we have derived contradictions to the causality principle by
means of the assumption that a QND state verification measurement leaves in-
variant the state I To ) which is to be verified, as well as its orthogonal comple-
ment. It may be seen from the formulation of the causality condition expressed
by eqn (11.166) that no contradiction will arise if one performs a demolition
measurement such that in all cases the object system will end up in one and
the same state. Denoting the latter by l' ) and assuming for simplicity that we
have only a single Yes-operation we have the following Yes/No-operations for the
device,

12 Yes = 1 41f )(To I, (11.254)

sIN0 = I Tf)(T_L (b)I. (11.255)
The states I T_L(b)) are orthogonal to I To ) and may, of course, depend on b, the
only necessary restriction being the normalization condition
556 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Fyes + FNo 4es 9Yes Enjko(b)f-No(b)

IT 0) Ohl E Ri(b))(4/±(01 =/. (11.256)

The above Yes/No-operations satisfy all conditions of a state verification mea-

surement. The final state Ixlif) may or may not be equal to 1‘110). For example,
it may be given by the vacuum state 10) of the electromagnetic field. This case
occurs, e.g. if the device measures the quanta of the field which are annihilated
after registration.
An important example for such demolition measurements are so-called ex-
change measurements. We define an exchange measurement as a device involving
local interactions with each part of the system. It is assumed that the Hilbert
(k)
space 7-1 (k) of the local part (k) of the object system and the Hilbert space 7 1 g -

of the corresponding part of the probe are isomorphic. Introducing local basis
(k)
vectors Ixi ) and I(/) k) ) in 1-1 (k ) and le) ,
respectively, the local object-probe
interaction in part (k) is supposed to take the form

Ix(ik ) ) io(j.k) )> ix(i k) ) (11.257)

The interaction thus acts simply by exchanging the labels of the basis vectors of
object and probe.
It is clear that such an interaction if applied to any non-local state of the
quantum object leads to a final object state which is isomorphic to the initial
probe state and vice versa. For example, taking the initial object state

oo) = E ctilx(i')) ® he) (11.258)

and the initial probe state

Io ) = 10(i2) )
(11.259)

we find as a result of the interactions in both parts (1) and (2) of the total system

Ro) 1 4'0) —> (E OilX i( 1) ) ®1)42) )) (E aikb i( 1) ) 10 2) )) . (11.260)

This shows that the final state IT f) of the object system is isomorphic to the
initial probe state I(1.0 ), namely

of) = E oilx(j1 )) (11.261)

whereas the final probe state is isomorphic to the initial object state. Thus the
rôles of object and probe state have been exchanged as a result of the local
interactions in both parts of the system.
 QUANTUM TELEPORTATION 557

FIG. 11.15. Schematic space-time picture of quantum teleportation. The incom-

ing state 10)) in the Hilbert space of Alice (A) is transferred to an isomor-
phic copy 10( 3)) of the Hilbert space of Bob (B). The quantum channel is
provided by a maximally entangled state 10(2,3)). The bold line represents
a classical communication channel through which Alice communicates her
measurement result to Bob. If Alice carries out a QND measurement she is
left with an entangled state 10(1 '2 )) which is isomorphic to 10(2 ,3)), provided
her measurement has been successful.

It must be noted, however, that with the local exchange interactions the
measurement is not yet complete: After the local interactions the various parts
of the probe state must be brought together to one place, at which a local
measurement on the probe state can be carried out. Any non-local measurement
on the probe state is, of course, subjected to the same causality restrictions as
discussed in the previous subsections.

11.4 Quantum teleportation

Non-local entangled states describe quantum correlations which are expressed
through the joint probability for measurements performed on separated parts
of a combined quantum system. An interesting application is to employ the
properties of entangled states for the coherent transfer of an unknown quantum
state from one part of a system to another, spacelike separated part. We shall
study here this transfer of a quantum state, known as quantum teleportation
(Bennett et al., 1993; Vaidman, 1994).

11.4.1 Coherent transfer of quantum states

Let us first define precisely what is meant by the teleportation of a quantum state
(see Fig. 11.15). We consider a system which is composed of three Hilbert spaces
7-1 (1) , 71 (2) and 7-/ (3) . In the following these spaces are supposed to be isomorphic,
558 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

that is to have the same dimension D. The total space of the composite system
is the threefold tensor product

_ R (1) -H( 2 ) (11.262)

The spaces WO and li( 2 ) describe the degrees of freedom of some localized part
of the total system which will be referred to as part (A). The third space 71 (3)
represents a second localized part of the system, denoted by part (B), which
is separated from part (A) by a spacelike distance. We may associate a local
observer, called Alice, with part (A), and a local observer, called Bob, with
part (B). Thus, Alice can measure all local observables pertaining to the space
W 1) OR (2 ), whereas the local observables in the space 1-1( 3 ) are available to Bob.
Consider now some state
D-1

10 (1) ) = E
i=0
(11.263)

belonging to the first Hilbert space "H( 1 ). 15 Alice aims to transfer this state in-
stantaneously and coherently to Bob, without measuring or destroying it: The
absolute values as well as the phase relations of the amplitudes ai are left un-
changed after the transfer process.
Such a transfer process can in fact be achieved with the help of an entangled
state 0 (2 ' 3 )) in the space '1-1 (2 ) 9-I( 3 ). To be definite we take this state to be a
certain maximally entangled state,

D-1

10(2'3) ) = E Ix(2)) 3
ix(i 3)). (11.264)

The state 10( 2,3 )) can be prepared in some localized space-time region. It is
then separated into two parts, described by the Hilbert spaces 1-1 (2) and li (3) .
which are brought to Alice and Bob, respectively. Thus, the quantum correlations
embodied in 10(2 ' 3 )) may be detected through joint measurements carried out by
Alice and Bob. In the following the state 10 (2 ' 3) ) will be referred to as a quantum
channel.
The initial state of the combined system is now given by

D-1

1 410) = 10 (1) ) ® 10 (2 '3) ) = E ® ix(i2)) ix(i 3)). (11.265)

v ,j=0
Suppose then that Alice performs a certain verification measurement of the state

15 Notethat in contrast to our previous convention the index i which labels the basis vectors
now runs from 0 to D — 1.
 QUANTUM TELEPORTATION 559

D-1

=
D
E Ix(:)) ®
k-O
(11.266)

which belongs to the Hilbert space available to Alice. If the state verification
yields a positive result the initial state 'T o ) of the total system is projected onto
the state 10( 1,2 )). As is easily deduced from the above equations we have
D-1
0
(0(1,2)14j
E
D i_o
131
)) (11.267)

This shows that Alice finds a positive result with a probability which is given by

= M0(12) 1 4'011 2 = D12 . (11.268)

Conditioned on this result the state prior to the measurement is transformed

into the final state after the measurement as follows,

1 410) = 10 (1) ) 0 10 (2 ' 3) ) —> 10 (1 '2) ) 010(3)). (11.269)

Thus we see that the state reduction conditioned on a positive outcome for the
measurement of Alice leads to the desired teleportation process: The final state
103)) in Bob's Hilbert space 1-1( 3) given by eqn (11.267) represents a copy of
Alice's initial state OM) (eqn (11.263)) in 7-1(').
The following facts should be noted. Being induced by a state reduction
following the measurement in one part of the system, the teleportation pro-
cess (11.269) can be viewed as occurring instantaneously in all Lorentz frames.
Namely, as we have already discussed it can be regarded as taking place along
all spacelike hypersurfaces which cross the classical event produced by Alice's
measurement.
Since the teleportation of a state involves only local interactions in one part
of the system there is, of course, no conflict with causality and no transfer of
information with a speed greater than the speed of light. Indeed, the reduced
density matrix which refers to Bob's Hilbert space fl(3) is proportional to the
unit matrix in this space, which is due to the fact that the quantum channel
(11.264) represents a maximally entangled state. Since the operations describing
Alice's measurement are local and act only ink') 0?-1 (2 ), Bob's reduced density
matrix is not affected by Alice's measurement, if the latter is described on the
non-selective level. Bob's density matrix which describes a state of maximal
ignorance is thus unchanged during Alice's measurement.
However, if the result of Alice's measurement is communicated via a classical
channel to Bob, he can use that information to conclude that his state 10(3)) is
isomorphic to P (1 )), provided the result of Alice's measurement was positive.
Note that in this case neither Alice nor Bob have measured the transferred
state. After her measurement Alice's state 10(1,2)) is an entangled state which is
isomorphic to the quantum channel ,( 23 teleportation of the given state
works although it remains completely unknown to both Alice and Bob.
560 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

11.4.2 Teleportation and Bell-state measurement

If the teleportation process is carried out on an ensemble of initial states in the
manner described above it is, in general, successful only for a fraction 1/D2 of
the ensemble. The reason for this fact is that a negative outcome of Alice's state
verification destroys, as a rule, the initial state ,Ip (i )). However, Alice can, at least
in principle, always design a device which yields a teleportation rate of 100%.
For this purpose Alice performs a measurement which projects the initial state
onto an orthogonal basis which consists of maximally entangled states, that is
Alice performs a measurement of a Bell-state operator in 1-1 (1) 0 1-1 (2) .
To explain this point in more detail we first note that an orthogonal basis of
maximally entangled states in 1-1 (1) 011 (2) is provided by the following expression,
13_1
10(2,2))„, E o , (1)
(11.270)
lu r,k IX(k+n) mod D) O IX 1(02) )•
k=0

The index v = 0, 1, ... , D2 — 1 labels the basis vectors 10(2' 2) ). Each v has a
unique representation of the form

v = r + nD, (11.271)
where

r = v mod D == 0, I, . . . , D — 1, (11.272)
and n --= 0, 1, ... , D — 1. The amplitudes Or,k in eqn (11.270) are defined to be
1 kr
r— exp (-27ri— ) . (11.273)
VD D
Let us demonstrate that eqn (11.270) yields, in fact, a basis of the desired
form. First, it is immediately clear that the
i 0)12)
,.) are maximally entangled.
Namely, by an appropriate relabelling of the basis vectors in )1(1), corresponding
to the first factor in the tensor product, we can always achieve the condition
that (11.270) takes on the standard form of the Schmidt decomposition. Since
all N,k are equal in absolute value,
1
Pr ki = ,---, (11.274)
' vD

we find that all I q5 )1 ' 2 ) ) are maximally entangled.

It therefore remains to be shown that the 10 (1 '2) ) are orthogonal and normal-
ized. To this end, one first verifies
D-1

E Or*/
k=0
,kfir,k = 6r1r, (11.275)
 QUANTUM TELEPORTATION 561

which is easily done with the help of the summation formula for the geometric
series. Setting v = r + TtD, and v' = r' + ni D we find for the scalar product,
D-1
(0 (1,2) 10( ( , 2) E / ( 1)
■
,kf - r+1 X(k±n9 mod D IX{11)H-1) mod D)(X k(2) X, )
k,1=0
D-1
(1) (1)
E 07' ,k0r,k (X (k±n' ) mod D IX(k+n) mod D)
k=0
D-1

—E 07, ,k Or,k 6n' n = 8r'r6n'n• (11.276)

k=0

In the second and the third step we have used the orthogonality of the basis
vectors 141)) and 1X k( 2 ), whereas in the last step eqn (11.275) has been used.
Thus we finally obtain the desired orthogonality of the basis vectors
(0 ( 12) (1 ,2)) = 6 v, v, (11.277)
which concludes the proof. Applying eqn (11.270) to the special case D = 2
(spin-1 particles) one finds the Bell states given by eqn (11.215).
Returning to quantum teleportation we now decompose the total initial state
(11.265) of the device as follows,

E
D 2 -1
(11.278)
v=0
From eqns (11.265) and (11.270) we get
D-1
(( 1 03 )
1 2) 'P) = D D,3*r,(i—n)
2•-\[.— mod D I A-(i—n) mod D)' (11.279)
i=0

For each v we introduce a corresponding unitary operator UP) which acts in

Bob's Hilbert space 71(3) . We define these operators by their action on Bob's
local basis,

UP) ) mod D1X ((i3 --n)

) mod D))
i = 0, 1, ,D — 1. (11.280)
By virtue of eqn (11.274) these operators are indeed unitary. In addition to the
multiplication by phase factors, they simply induce a cyclic permutation of the
basis states
This enables us to rewrite eqn (11.278) as follows,
D 2 -1
12
liP0)= l'Ip(1)) E lo, ))0u 3 )1 ,1p(3)), (11.281)
v=0
where 10 (3) ) is given by eqn (11.267). From this equation we read off that a
teleportation rate of 100% is indeed possible if Alice carries out a measurement
562 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Alice Bob

FIG. 11.16. Sketch of the setup used for an experimental realization of quantum
teleportation. Alice's state analysis is realized by the superposition of the two
field modes b1 and b2 at a beam splitter BS and the subsequent measurement
of the outgoing photons in modes c1 , c2 through detectors D1 and D2. The
successful teleportation to the field mode b 3 has been demonstrated by a state
analysis of that mode (not shown).

of the complete basis 10P-2) ). All D 2 possible outcomes of that measurement

occur with probability 1/D 2 . Receiving the outcome v of Alice's measurement
via the classical communication channel, Bob can then apply the inverse of the
corresponding unitary operator UP) to his state which yields in all cases an
isomorphic copy 10(3) ) of Alice's initial state IOW).
As an example, let us apply the above construction to the case D = 2, that
is, to the spin-1 example studied previously. We then find that Bob has to apply
one of the following unitary operators

20) = ax 3(3) 1°- Y1 (11.282)

depending on the outcome of Alice's measurement.

11.4.3 Experimental realization

We conclude our discussion with the discussion of an experimental realization
of quantum teleportation performed by Bouwmeester et al. (1997). The experi-
mental setup is sketched in Fig. 11.16. It works with three electromagnetic field
modes b1, b2 and b 3 . We write b biA for the corresponding creation and
annihilation operators, where i = 1, 2, 3 and A .<-4, t denotes the photon polar-
 QUANTUM TELEPORTATION 563

ization. The quantum channel is provided by an EPR source which generates the
entangled two-photon state
1 (bt bt _ bt bt )10\
0 244 31 21 344 (11.283)

by parametric down-conversion (Michler et al., 1996). A careful design of the de-

vice guarantees that all three modes contain a single photon. The total Hilbert
space is then isomorphic to the tensor product of the Hilbert spaces correspond-
ing to the three modes.
The state to be transferred is given by

= (abt ++ + fib i ) 10), 1a1 2 +01 2 ------ 1, (11.284)

such that we have the following three-photon initial state,

\ 1 t ht n
— vr2 ((Ai"• + fibt, ) .6++bt3I (11.285)

Alice performs a state verification measurement on her photons in modes bi

and b2 using a beam splitter of transmittivity (see the discussion on homodyne
photodetection in Section 6.4). She measures with the help of her detectors D1
and D2 the quanta annihilated by the operators ci, and c2A. With a suitable
choice of the phases for these operators we have

CIA = (NA + b2),), (11.286)

1,
C2A = —
0 01À NA) (11.287)

A crucial point of the experiment is to ensure that the photons in modes 1)1
and b2 arrive simultaneously at the beam splitter, since otherwise they could
be distinguished through their arrival times. A more refined theoretical analysis
(Braunstein and Mann, 1995) is also required which takes into account the space-
time structure of the states in the various modes. We neglect such considerations
since they do not modify the principal statements.
In order to find the possible coincidence events measured through detectors
D1 and D2 we solve relations (11.286) and (11.287) for bo, and b2), and substitute
into eqn (11.285), which leads to
1 fvt ct _ t t
1*o) = + cit c2++ ) (ab 3t H + $bt3t ) 1 0)

\ir
-8, ( e4eti++ _ c2t tc2t ++) (ab 3t ++ — /303 ) 10)

10 4 12 r41 2
Vg l_c 'lli — L c4i ) )3 4+41 0 )
1 I 2\
+ .7g [Ctl++] 2 — [C t2 ++ ] ) ab 0). (11.288)
564 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

From this equation we infer that we may distinguish four different cases of
coincidence events. In case 1, which is described by the first term on the right-
hand side of eqn (11.288), both detectors register a single photon. Note that
in this case the polarizations of the photons are necessarily different from each
other, since terms proportional to ct ++ 4 ++ or c4ct2I do not appear in the above
decomposition. This is an interference effect: Due to the beam splitter there are
two different paths which contribute to the event of registering a single photon in
both detectors. The amplitudes for these paths interfere destructively such that
only that component of the initial two-photon state contributes which contains
different polarizations in the modes b1, b2. It follows from the first line that case
1 occurs with a probability of 1 and leads to the following state of mode b3,

10 (3) ) = (ab 3f ,+ +$b) 10), (11.289)

which represents a copy of the state 10 (1) ) given by eqn (11.284). This case
is investigated in the Zeilinger experiment (Bouwmeester et al., 1997) where
the teleportation process is demonstrated with the help of an additional state
analysis carried out on mode b3 .
Case 2, which also occurs with probability / is described by the second line
of eqn (11.288). It consists of those events where two photons are detected by
either D1 or by D2 with different polarizations. Case 3 (third line) and case 4
(fourth line) represent those events where either D1 or D2 registers two photons
with both polarizations equal to t or both equal to ++, respectively. The total
probability for case 3 is equal to 1012/2 and equal to la1 2 /2 for case 4.
In case 2 Bob gets the final state (ab. — ,3bt3I)10) in mode b3 . By a simple
application of a, he again obtains a copy of10( 1 ) ). Thus, if Alice could distinguish
faithfully the cases 1 and 2 from the cases 3 and 4, a teleportation rate of 50%
could be achieved in principle with the above scheme. This is the upper limit
since in cases 3 and 4 the (normalized) final state of mode b3 is either 1)1'3++ 0) or
t
b 31 10) such that the phase relations of Alice's initial state OW) are irreversibly
lost.
Quantum teleportation offers completely new perspectives for experiments on
the foundations of quantum mechanics. It could be used, for example, to transmit
quantum states over large distances without destruction by decoherence effects.
Note also that the teleported state itself could be part of an entangled state
in some larger Hilbert space. It is thus possible to transfer the entanglement
between particles and to carry out tests of the Bell inequalities (Bell, 1964;
Clauser and Shimony, 1978) on particles which did not directly interact in the
past.
In addition to new experiments on the foundations of quantum mechanics
several further applications of teleportation devices in quantum information pro-
cessing and communication have been suggested (Bouwmeester et al., 1997). For
example, they could be used to store the information carried by photons on
trapped ions, or to teleport a quantum state over large distances even if the
 REFERENCES 565

available quantum channels are of very poor quality, employing entanglement

purification. Moreover, quantum teleportation schemes could also serve as links
between quantum computers or to protect states in quantum computers from
their environment.
We remark finally that the teleportation scheme studied above provides a nice
example of the concept of states as functionals on the set of spacelike hypersur-
faces, which was introduced in Section 11.1.1. Let us denote by x (1) and x (2)
the space-time points corresponding to the photodetection events at detectors
D1 and D2, respectively. Above we have tacitly assumed that these events occur
simultaneously in Alice's coordinate frame. However, in principle D1 and D2
could be arbitrarily far apart and we may consider some spacelike hypersurface
a which crosses the forward light cone based at x( 1 ) and the backward light cone
based at x(2) . If a is flat, it may therefore be viewed as an equal-time hypersur-
face in the coordinate frame of a moving observer 0 for whom the detection at
DI takes place earlier than that at D2. According to our discussion in Sections
11.2.1 and 11.2.3 we find that observer 0 associates a state to the hypersurface
a which is not related by any unitary representation of the Lorentz group to any
state in the foliation of Alice. In fact, according to eqn (11.288) we have on a a
state vector which is conditioned on the outcome of the first measurement at D1
and which is given by one of six possible states, namely two two-photon states,

ct21 (ab 3t ++ + fibt3t) 0), ct2+, (abt3+, + Obt31 ) 1 0) , (11.290)

three one-photon states,

b41 0), bt 3++10 ), (abt3 ,, — )3030 10), (11.291)

and one three-photon state,

1 (0 [C2$ ] 2 bt3++ — a
0 [4 ÷+ ] 2 b3t t ± Ct2++ Ct2$ [abt3++ — Obtn ] ) 0). (11.292)

Note also that the mixture which describes the measurement on the non-selective
level has no definite photon number. This illustrates once again that a state
vector (or a density matrix) must be represented as a functional on the set
of spacelike hypersurfaces, which means that it must be linked to the various
foliations of space-time.

References
Aharonov, Y. and Albert, D. Z. (1980). States and observables in relativistic
quantum field theories. Phys. Rev., D21, 3316-3324.
Aharonov, Y. and Albert, D. Z. (1981). Can we make sense out of the measure-
ment process in relativistic quantum mechanics? Phys. Rev., D24, 359-370.
Aharonov, Y. and Albert, D. Z. (1984a). Is the usual notion of time evolution
adequate for quantum-mechanical system? I. Phys. Rev., D29, 223-227.
566 MEASUREMENTS IN RELATIVISTIC QUANTUM MECHANICS

Aharonov, Y. and Albert, D. Z. (1984b). Is the usual notion of time evolu-

tion adequate for quantum mechanical system? II. Relativistic considerations.
Phys. Rev., D29, 228-234.
Aharonov, Y., Albert, D. Z. and Vaidman, L. (1986). Measurement process in
relativistic quantum theory. Phys. Rev., D34, 1805-1813.
Bell, J. S. (1964). On the Einstein-Podolsky-Rosen paradox. Physics, 1, 195-
200.
Bennett, C. H., Brassard, G., Crépeau, C., Jozsa, R., Peres, A. and Wootters,
W. K. (1993). Teleporting an unknown quantum state via dual classical and
Einstein-Podolsky-Rosen channels. Phys. Rev. Lett., 70, 1895-1899.
Bloch, I. (1967). Some relativistic oddities in the quantum theory of observation.
Phys. Rev., 156, 1377-1384.
Bohm, D. (1951). Quantum Theory. Prentice-Hall, New York.
Bohr, N. and Rosenfeld, L. (1933). Zur Frage der Meflbarkeit der elek-
tromagnetischen Feldgriiflen. Det Kgl. Danske Videnskabernes Selskab.,
Mathematisk-fysiske Meddelelser XII, 8, Kopenhagen.
Bouwmeester, D., Pan, J.-W., Mattle, K., Eibl, M., Weinfurter, H. and Zeilinger,
A. (1997). Experimental quantum teleportation. Nature, 390, 575-579.
Braunstein, S. L. and Mann, A. (1995). Measurement of the Bell operator and
quantum teleportation. Phys. Rev., A51, R1727 1730.
-

Breuer, H. P. and Petruccione, F. (1998). Relativistic formulation of quantum-

state diffusion. J. Phys. A: Math. Gen., 31, 33-52.
Breuer, H. P. and Petruccione, F. (1999). State vector reduction in relativistic
quantum mechanics: An introducion. In Open Systems and Measurement in
Relativistic Quantum Theory (eds. Breuer, H. P. and Petruccione, F.), Volume
526 of Lecture Notes in Physics, pp. 1-30. Springer-Verlag, Berlin.
Breuer, H. P. and Petruccione, F. (2001). Relativistic theory of continuous
measurements. In Chance in Physics. Foundations and Perspectives (eds. Bric-
mont, J., Diirr, D., Galavotti, M. C., Ghirardi, G., Petruccione, F. and Zanghi,
N.), Volume 574 of Lecture Notes in Physics, pp. 195-212, Springer-Verlag,
Berlin.
Clauser, J. F. and Shimony, A. (1978). Bell's theorem: Experimental tests and
implications. Rep. Prog. Phys., 41, 1881-1927.
Diàsi, L. (1991). Covariant formulation of multiple localized quantum measure-
ments. Phys. Rev., A43, 17-21.
Einstein, A., Podolsky, B. and Rosen, N. (1935). Can quantum-mechanical
description of physical reality be considered complete? Phys. Rev., 47, 777-
780.
Jauch, J. M. and Rohrlich, F (1980). The Theory of Photons and Electrons.
Springer-Verlag, New York.
Landau, L. and Peierls, R. (1931). Erweiterung des Unbestimmtheitsprinzips
fiir die relativistische Quantentheorie. Z. Phys., 69, 56 69.
-

Michler, M., Mattle, K., Weinfurter, H. and Zeilinger, A. (1996). Interferometric

Bell-state analysis. Phys. Rev., A53, R1209-R1212.
 REFERENCES 567

Misner, C. W., Thorne, K. S. and Wheeler, J. A. (1973). Gravitation. Freeman,

San Francisco.
Popescu, S. and Vaidman, L. (1994). Causality constraints on nonlocal quantum
measurements. Phys. Rev., A49, 4331-4338.
Schweber, S. S. (1961). An Introduction to Relativistic Quantum Field Theory.
Row, Peterson and Company, Evanston, Illinois.
Schwinger, J. (1948). Quantum electrodynamics. I. A covariant formulation.
Phys. Rev., 74, 1439-1461.
Sorkin, R. D. (1993). Impossible measurements on quantum fields. In Direc-
tions in General Relativity (eds. Hu, B. L. and Jacobson, T. A.), Volume 2.
Cambridge University Press, Cambridge.
Tomonaga, S. (1946). On a relativistically invariant formulation of the quantum
theory of wave fields. I. Prog. Theor. Phys., 1, 27-42.
Tomonaga, S. (1947). On a relativistically invariant formulation of the quantum
theory of wave fields. II. Prog. Theor. Phys., 2, 101-116.
Vaidman, L. (1994). Teleportation of quantum states. Phys. Rev., A49, 1473—
1476.
Weinberg, S. (1995). The Quantum Theory of Fields, Volume I. Cambridge
University Press, Cambridge.
 12
OPEN QUANTUM ELECTRODYNAMICS

In quantum electrodynamics the matter degrees of freedom are coupled to the

radiation field A 4 (x) through a local, gauge-invariant interaction density of the
form ji-t(x)A p (x), where j(x) is the conserved current density of the charged
matter. Due to the linear structure of this coupling, the problem of constructing
a complete formal representation of the reduced matter dynamics can be solved
exactly if the electromagnetic radiation field is initially in a state describable by
a Gaussian characteristic functional. An appropriate strategy to achieve this goal
is the use of functional methods from field theory. In this chapter we combine
such methods with the super-operator technique to derive an exact, relativistic
representation for the reduced density matrix of the matter degrees of freedom.
The reduced dynamics involves a certain time-ordered functional of current
density super-operators which completely describes the influence of the elec-
tromagnetic radiation field on the matter dynamics. This functional has already
been employed in our study of quantum Brownian motion in Section 3.6.4 for the
derivation of the Feynman—Vernon influence functional and of the path integral
representation for the propagator function of the Caldeira—Leggett model.
The functional super-operator technique developed in this chapter can be
used as a starting point for the derivation of the master equations encountered
in previous chapters. Here, we illustrate this technique by means of a specific
application. Namely, we investigate the suppression of the quantum coherence of
charged particles through the emission of radiation in typical interference devices.
To this end, the degree of decoherence will be characterized through a relativis-
tic, gauge-invariant decoherence functional. An appropriate technique allowing
the explicit determination of the decoherence functional for simple interference
devices will be developed. It turns out that the relative motion of two interfer-
ing wave packets leads to a loss of coherence which is caused by the emission
of soft, low-frequency bremsstrahlung. This decoherence mechanism provides a
quite fundamental process in quantum electrodynamics since it dominates for
short times and because it is at work even in the electromagnetic field vacuum
at zero temperature.
Finally, we investigate the possibility of the destruction of coherence of the
superposition of many-particle states. It will be argued that, while the decoher-
ence effect is small for single electrons at non-relativistic speed, it is strongly
amplified for certain superpositions of many-particle states.
 DENSITY MATRIX THEORY FOR QED 569

12.1 Density matrix theory for QED

In contrast to the convention in previous chapters we shall use in this chapter
rationalized Heaviside—Lorentz units such that h = c = 1 and the fine structure
constant is given by
e2 1
(12.1)
47r 137'
where e is the electron charge and c the speed of light. As in Chapter 11 space-
time coordinates of 4-vectors in Minkowski space are denoted by xP = (x° ,
and we use a metric tensor with signature (+1, —1, —1, —1). Accordingly, the
Lorentz scalar product of two 4-vectors xi2 = (x°, and y' = (y° , y) is written
as
xy = xii yo = yiivx p y v = x 0 y0 (12.2)
where we use the summation convention. Occasionally, we will reintroduce factors
of c and h.

12.1.1 Field equations and correlation functions

To be specific we perform our study in the Coulomb or radiation gauge. In
this gauge the elimination of the radiation degrees of freedoms is easily car-
ried out. Although we lose manifest Lorentz covariance by taking a fixed gauge,
the Lorentz covariance and the gauge invariance of the final results are easily
established.

12.1.1.1 Maxwell equations and commutation relations In the Coulomb gauge

the free radiation field is described by the vector potential If satisfying the
Coulomb gauge condition
t' • A = O. (12.3)
The electric and magnetic fields are given in terms of  by16

Â
(12.4)
Ot
= x (12.5)
The free Maxwell equations,

• ET = 0, (12.6)
•.6 = 0, (12.7)

x fa, = — , (12.8)

xij= aÉ'
atT ' (12.9)

16 We put an index T on the electric radiation field ET to emphasize its transverse character.
570 OPEN QUANTUM ELECTRODYNAMICS

then lead to the wave equation for the transverse vector potential,
( n2
E]if X-- — A)
ot 2 = o. (12.10)

The wave equation (12.10) may be solved with the help of the plane wave
decomposition of the vector potential,

A(x) = d3k E [4(k)bx(k)e _ikx 4(,-i)q(k)e+2,kx} (12.11)

V2(27) 3 w A=1,2J

As usual, the plane wave modes are characterized by a wave vector k and two
unit polarization vectors FAA, such that

ex (k) = ol (12.12)
(k) (k) = (12.13)
E e)ci(k) = _ kik3 = 1,2,3. (12.14)
A=1,2 1k1 2

The tensor Pii(k) is the projector onto the transverse component in k-space. In
position space it takes the form

P2 i (it)
aia •
A (12. 1 5)

The field operators b),(k) and 14(k) represent the destruction and creation of a
photon with wave vector k and polarization ex (k), satisfying bosonic commuta-
tion relations

[ex (k), ex, (P)] = 0, (12.16)

[b),(k), OA, (re )] = 6)0,, S(k — (12.17)

The momentum 4-vector of the photon is written as

= (k° ,k) = (w,rc), w= (12.18)

and kx lex o =wx° rc •

—

With the help of the above relations it is easy to determine the commutation
relation for the free vector potential at arbitrary times,

DT (x — x')ii E i[A i (x),A i (x 1 )] = Pii D(x — — x'). (12.19)

The function D(x — x') is the Pauli—Jordan commutator function of the electro-
magnetic field. Its Fourier representation takes the form
 DENSITY MATRIX THEORY FOR QED 571

d3 k —ik(x—x') e ik(x —x i ))
(x — x') = il (12.20)
2(27) 3 w (e
Explicitly, one finds
, 1
D(x — x ) = 47rr [6(r — t) — 6(r + t)] (12.21)
1
= — —sign(xo — x 10 )6[(x — x') 2 ], (12.22)
27r
where

r t Eixo (12.23)
We observe that D (x — x') is an antisymmetric function and satisfies
D(x — x0=x,0 = 0, (12.24)
00D(x — x') 0 , = —) (12.25)

Moreover, D (x — x') solves the homogeneous wave equation

OD (x — x') = 0, (12.26)
and is obviously invariant under Lorentz transformations'

(12.27)
that is, we have

D(A[x — x']) = D(x — x'). (12.28)

The commutator function D (x x') is a singular function on Minkowski space
—

which vanishes outside the surface of the light cone defined by (x x') 2 = 0 and —

has a 6-type singularity on it. It is connected to the retarded Green function

Dret(x — x') and to the advanced Green function Dadv (X —XI) of the wave equation
through the relations

D (x — x') = — (D„ t (x — x') — Dadv (X — x')) , (12.29)

1
sign(xo — x'0 )D(x — x') = — 6[(x — ) 2 ] (12.30)
27r
= (Dret(x x') Dadv(x x i )) •
Finally, we remark that the commutation relation (12.19) immediately yields by
virtue of eqn (12.25) the equal-time commutation relation between the vector
potential and the transverse electric field operator,

[ffir (t, A i (t , 0)1 = +iP1i6(Z — 0) i(5:31 (Y — 0). (12.31)

17 More precisely, it is invariant under orthochronous Lorentz transformations which satisfy

A° 0 > 1.
572 OPEN QUANTUM ELECTRODYNAMICS

12.1.1.2 The anticomniutator function As we have already seen in our study

of quantum Brownian motion, in addition to the commutator function a further
correlation function plays an important rôle in the description of the reduced sys-
tem dynamics. This is the anticommutator correlation function which is defined
by
(x — x') ii E ({A i (x), A i (x 1 )1) f
- (12.32)
The angular brackets denote the average with respect to the radiation field in a
thermal equilibrium state at a certain temperature T,
in 1 r L
(0)f = trf 2-f- exp l —il f/ ri6 EITJ } 7

where

Hf = f c13 k E wbtx b), (1-i)

A=1,2
represents the Hamiltonian of the free radiation field, trf denotes the trace over
the radiation degrees of freedom, and Zf is a normalization factor, the partition
function of the field.
With the help of

fbA(k), } f = { btA (k), VA, (P)}) f = 0, (12.35)

(lb), ,(k l )1)f = 6,6(1; - ki) coth (w/2kBT), (12.36)

one finds
(x — x')ii -=. Pii D i (x — (12.37)
with the anticommutator function

(x --z iÇ d3 k
2(2703w
(e —ik(x—x`) e ik(x—x' )/) coth (w/2kB T)

Dre (x — x') (x — x'). (12.38)

In the last line we have decomposed D 1 into a vacuum and a thermal part. The
vacuum part is easily found to be
d3 k (e —ik(x—x') e ik(s —x i )
2(27) 3 b.;
1 [ 1 11
+P
47r2 r r t r tj
—

1 1
(12.39)
27 2 (x — 42 *
P denotes the Cauchy principal value. Like the commutator function D(x — x')
the vacuum contribution Drc (X — x') of the anticommutator function is Lorentz
 DENSITY MATRIX THEORY FOR QED 573

invariant and satisfies the free wave equation. While D (x x') is antisymmetric,
-

DIrac ( x x l\ is obviously symmetric. We note that Dl'ac(x - x') does not vanish
)
outside the light cone and exhibits a principal value singularity on the light cone.
To determine the thermal part of the anticommutator function one first car-
ries out the angular integration of the k-integral in (12.38) to obtain
CO
1
D ti.11 (x - s') = 4 f du) [coth (u; I 2kBT) - 1] [sin w (r - t) + sin w (r + t)] .
71_2r
o
(12.40)
With the help of the formula (4.57) this yields
1 1 ( r + t) 1 1
[ 1 coth ( I' t ) 1 + coth
472 r TB TB r—t TB TB j r + t .1 •
(12 41)
This expression involves the thermal correlation time
1
TB -1.- (12.42)
7 kBT
which was already introduced in eqn (4.58).
One observes that Mh (x - x') is a regular function on Minkowski space. If
we combine the above expression for D1 h (x - x') with expression (12.39) for the
vacuum contribution we see that the contributions involving 1/(r ±t) cancel each
other. With the convention that the singularity on the light cone (x-x') 2 = 0 is to
be treated as a principal value singularity we can write the total anticommutator
function as follows,
1 r-t
Di (x - x') = coth (—) + coth (—
r ±t)
472 rTB TB TB

1 sinh(2r/TB)
-
_ (12.43)
4712 rTB sinh[(r - ON] sinh[(r + ON] .
It must be noted that the anticommutator function is not Lorentz invariant,
which is connected to the fact that the thermal radiation field singles out a
certain frame. Of course, the full anticommutator function is symmetric and
satisfies the free wave equation. According to eqn (12.43), D 1 (x x') changes -

sign if one crosses the surface of the light cone, that is D i (x - x') > 0 for
space-like distances ((x - x') 2 < 0) and D i (x - s') < 0 for time-like distances
((x - 4 2 > 0).
The dipole approximation of the anticommutator function is obtained by
taking the limit r 0 in eqn (12.43) which yields
1
Di (t, 0) = (12.44)
27 2 713 sinh2 (t/TB) .
In the limit It < TB this function diverges as t -2 , while it decays exponentially
as exp(-210-B) for times It > TB.
574 OPEN QUANTUM ELECTRODYNAMICS

On the other hand, by taking the limit t 0 of eqn (12.43) we find

1
D i (0,r) coth(r/TB). (12.45)
27-7-7-B

This shows that D 1 (0, r) diverges as r -2 for small distances (r < TB) but decays
only algebraically as T. for large distances (r >> TB).

12.1.1.3 The correlation function of the electric field A simple but interesting
application is the determination of the correlation function of the transverse
electric field ET (s). Invoking homogeneity in space and time, one can see that
it suffices to calculate the quantity

(1E7'4), ET MD f =
== 80 D1 (t,
(t, r), (12.46)

where t = x° and r In the last step we have used the fact that D 1 satisfies
the homogeneous wave equation. By use of the explicit expression (12.43) this
gives

= (8ii — " i " i ) (12.47)

—1 [ coshRr — t)/7-B] cosh[(r + ON] 1
. I + \/ 1
272 dr Lsinh 3 [(r t) 1 TB]
—
sinh3 [(r + t)1 Ta J

where " i = xdr. Summing over i = j we are led to the expression

—1 [ coshRr — t)/TB] cosh[(r + ON] I

f 727.br Lsinh 3 [(r t)/TB]
—
sinh3 [(r + t)17-B 1
(12.48)

Let us investigate first the limit r 0 of eqn (12.48) which is given by

1 6 + 4 sinh2 (t/TB)
({ÊT(t,0),ÉT(0)1) f = (12.49)
71-2 713 sinh4 (t/TB)

For times which are short compared to the thermal correlation time this yields

\ 6
ffT, (t, 0), ÊT ( 0 ) } ) f 720 ' « Ta (12.50)

which corresponds to the vacuum contribution of the correlation function. Thus,

the vacuum part dominates for short times and diverges as
 DENSITY MATRIX THEORY FOR QED 575

The thermal contribution to the correlation function of the electric field is

given by

1 [ 6+ 4 sinh2 (t/TB) 6 1
{ .É'T (t,0),f `r (°) (12.51)
}) th 71-24L sinh4 (t/TB) (t/TB) 4 ]

For It1/7-B 0 the term within the square brackets approaches the value 2/15,
such that we may write

2 ( t
({ÊT(t,o),ÊT(o)}) = (12.52)
th 4 g TB
157T 2 TB

where the function

15 [ 6 + 4 sinh2 T 6
g(r)
y4
(12.53)
2 sinh4 T

satisfies g(0) = 1. The energy density of the field in thermal equilibrium is thus
found to be
1 72 (kBT) 4
Uth = KfET(0),ET(0)/ =
/ th 1572 7-4 15

which is the Stefan—Boltzmann law of black-body radiation.

On the other hand, in the case t » TB eqn (12.49) yields

16 e-2Iti/TB.
{ÈT(t,0),ÉT(0)} , f 71-2 TB4 (12.54)

The correlation function thus decays exponentially for times large compared to
the correlation time TB.
Finally, let us investigate the limit 0 in eqn (12.48). This limit gives

r), ÉT (0) = — 2 cosh(r/1-13)

-

(12.55)
K J/ f 71-2 Tbr sinh3 (r/TB)

so that we have

7 2 r4
for r< TB, (12.56)
16 - 2 r7-13
/
({É(o
r,r),ÊT(0)})f 71-2TB3r e for r >> TB. (12.57)

The correlation function of the electric field thus diverges as r -4 for short dis-
tances, r < TB. By contrast to the behaviour of the anticommutator function
D 1 in this limit, it decays exponentially for distances large compared to TB.
576 OPEN QUANTUM ELECTRODYNAMICS

12.1.2 The reduced density matrix

We now turn to the reduced dynamics of the matter degrees of freedom which
are coupled to the radiation field. In the Coulomb gauge the Hamiltonian density
in the interaction picture takes the form (Weinberg, 1995; Jauch and Rohrlich,
1980)

7-1(x) =7-lc(x) + 7-1T (x).

Here,

' HT (X) = --/( x) • X(X) E j P (X)Aiz (X)

represents the density of the interaction of the matter current density j°(x) with
the transverse radiation field A° = (0, if). The matter current density is assumed
to satisfy the continuity equation

= 0, (12.60)

expressing the local conservation of charge. The Coulomb energy density is given
by

= _21 d3 y i° (x° , Y):7 ° (x° , (12.61)

471- IY —

such that
f xf y j o (x o y)i o (x o
Hc (x ° ) = (12.62)
4 7 1Y —
is the instantaneous Coulomb energy of the charge distribution given by j° (x).
For notational convenience, we have set in eqn (12.59) the scalar potential of
the radiation field equal to zero, A° = 0. This allows us to write the correlation
functions of the field as

DT (x — x') = i [A (x), A v (x 1 )] = — P D (x — (12.63)

Drir (x — x') --=. ({A (x) , A v (x')}) f +.13/„D1(x — x i ), (12.64)

where 1341) is defined by eqn (12.15), and by -Poo = Poi -=- I = 0 for j = 1, 2, 3.
It must be remembered, however, that these and all other correlation functions
to be introduced later involve the transverse projection P.
As in the preceding section all fields are taken to be in the interaction picture.
The interaction picture density matrix p(t) for the coupled system then satisfies
the Liouville—von Neumann equation which we write as

a
— p(t) = f d 3 x L(t ,)p(t) . (12.65)
Ot
 THE INFLUENCE FUNCTIONAL OF QED 577

The Liouville super-operator r(x) is given by r(x) = £c(x) + .CT(x), where

we introduce the Liouville super-operators pertaining to the densities of the
Coulomb field and of the transverse field,

EC (X)P [RC (X)) P], (12.66)

-CT (X)P E ['HT (X), P]. (12.67)

Integrating the Liouville—von Neumann equation formally with the help of the
chronological time-ordering operator and taking the trace over the radiation field,
we obtain the following equation,
tf
pm (t f) = trf T, exp f di x r(x)1 p(ti)} . (12.68)
ti

This equation relates the density matrix pm (t f ) describing the matter degrees
of freedom at some final time tf to the density matrix p(ti ) of the combined
matter—field system at some initial time t i . It will be the starting point for the
derivation of the influence functional in the next section.

12.2 The influence functional of QED

Employing functional techniques of field theory, we derive in this section a super-
operator representation for the influence functional of QED. The influence func-
tional will be given in the form of a functional of super-operators of the matter
current density, involving certain Green functions of the radiation field. The re-
sulting expression will be discussed physically. In particular, it will be related
to well-known expressions for the vacuum-to-vacuum amplitude of the electro-
magnetic radiation field in the presence of a classical current density, and to
the classical formulation of a system of charged particles in terms of a non-local
action principle (Breuer and Petruccione, 2001).

12.2.1 Elimination of the radiation degrees of freedom

Our aim is to eliminate the variables of the electromagnetic radiation field to
obtain an exact representation for the reduced density matrix pm of the matter
degrees of freedom. To this end, we invoke the formal representation (12.68).
The derivation is performed in two steps. First, we eliminate the time-ordering
of the electromagnetic variables and, second, we determine the remaining trace
over the field variables employing the Gaussian property of the field state.

12.2.1.1 Eliminating the time ordering of the field variables Our first step is
a decomposition of the chronological time-ordering operator T„ into a time-
ordering operator T for the matter current and a time-ordering operator TA ,_
for the electromagnetic field variables as T, = Ti,TA,. This enables one to write
eqn (12.68) as
578 OPEN QUANTUM ELECTRODYNAMICS

tf

Prn(t f) T3, (trf 111,1 _ exp f (Cc (x) + .CT(x))1 P(t)}) .

(12.69)

The currents jA(x) commute under the time-ordering Ti„. We may therefore
treat them formally as commuting c-number fields under the time-ordering sym-
bol. Since the super-operator Lc (x) contains only matter variables, the corre-
sponding contribution can be pulled out of the trace. Hence, we have
[t f tf

pni (t f ) = V, exp
( f dlx re (x)] trf {V_ exp [ f dlx .CT (x)] p(t i )})
ti ti
(12.70)
We now proceed by eliminating the time-ordering of the A fields. With the
help of the Wick theorem (Itzykson and Zuber, 1980) we get

T I,_I exp f dix LT (X) ] (12.71)

ti
[ tf tf t
[ f
1
= exp — f di x f dix i [LT(x), .CT(x')]0(t — exp f dIx .CT (x) .
2
ti
In order to determine the commutator of the Liouville super-operators we invoke
the Jacobi identity which yields for an arbitrary test density p,

[CT (x) , .CT (x' )]p = ,CT (x) LT p - ,CT (x') LT (x)p
= (x), [7-1T (x i ), Pfl + HT (x') [1T (x),
['

= - [[RT (x), 'HT (41, 14. (12.72)

The commutator of the transverse energy densities may be simplified to read

[RT (x), 7-iT (x')] jt` (x)jv (x')[A0 (x), A, (xi)] , (12.73)
since the contribution involving the commutator of the currents vanishes by
virtue of the time-ordering operator Ti,. Thus, it follows from eqns (12.72) and
(12.73) that the commutator of the Liouville super-operators may be written as

[rT (x), (x 1 )]p = — [A (x) , A, (x 1 )][f (x) jv (x i ) , (12.74)

It is useful to introduce current super-operators 4(x) and J _(x) by means

of

(x)P (x)P, J—t` (x)P Pi"(x). (12.75)

Thus, J+(x) is defined to be the current density acting from the left, while J_ (x)
acts from the right on an arbitrary density. With the help of these definitions
 THE INFLUENCE FUNCTIONAL OF QED 579

we bring the expression for the commutator of the Liouville super-operators into
the form

[CT (x), (x i )] = — (x), A, (x 1 )] (,n(x), fF (x') — J1-` (x)Jv (x')) . (12.76)

Inserting this result into eqn (12.71), we can write eqn (12.70) as
tf

pni (t f) (exp f di x rc (x)

ti
tf tf

f d4 x f d 4x 1 0(t — ti )[A 4 (x), Av (x')]4` (x)Jfk (x i )

tf tf

f
ti
di x f d 4 x 1 6(t — t')[A 4 (x), A v (4],/Lt (x)J'' (x I )
ti

x trfexp
{ f di x ,CT(x)Ip(t i )}) . (12.77)
t

This is an exact formal representation for the reduced density matrix of the
matter variables. The time-ordering of the radiation degrees of freedom has been
removed and the latter enter eqn (12.77) only through the functional
tf
W[4, ,1_] trf {exp f d i x ,CT (x)1 P(t)} , (12.78)
t,

since the commutator of the A fields is a c-number function.

12.2.1.2 The influence super operator The functional (12.78) involves an aver-
-

age over the field variables with respect to the initial state p(ti ) of the combined
matter— fi eld system. It therefore contains all correlations in the initial state of
the total system. In the following our central goal is to investigate how corre-
lations are built up through the interaction between matter and radiation field.
We therefore consider an initial state of low entropy which is given by a product
state of the form

P(ti) = p(t i ) pf, (1 2. 79)

where Pm (t i ) is the density matrix of the matter at the initial time and the
density matrix p f of the radiation field describes the thermal equilibrium state at
temperature T. The influence of the special choice (12.79) for the initial condition
may be eliminated by pushing ti to —oo and by switching on the interaction
580 OPEN QUANTUM ELECTRODYNAMICS

adiabatically. This is the usual procedure used in quantum field theory in order
to define asymptotic states and the S-matrix. The matter and the field variables
are then described as in-fields, obeying free field equations with renormalized
mass.
For an arbitrary initial condition p(ti ) the functional W[4, J_] can be de-
termined by means of a cumulant expansion (see Section 9.2.3). Since the initial
state (12.79) is Gaussian with regard to the field variables and since the Liouville
super-operator ,CT (s) is linear in the radiation field, the cumulant expansion ter-
minates after the second-order term. In addition, a linear term does not appear
in the expansion because of (44 (X)) f = 0. Thus we immediately obtain
tf
tf
1
= exp di x f
t,
f
t,
(CT (x),CT (x'» i ]
pm (ti ). (12.80)

Inserting the definition of the Liouville super-operator .CT (s) into the exponent
of this expression one finds after some algebra,
tf tf
1
di X d4 X 1 (CT (X)LT ( XI ))f Prn ( t i) (12.81)
2
ti ti

tf tf

= 1 f dix f trf LHT(x), NT(x 1 ),P(ti)]]

2
t2 ti
tf tf
1
= f dix f (12.82)
2
ti ti
[(Ay (X))f cL'LF (x),IfF (s') + (A 1,(x)A, (x' )) f J (x)J (s')
— (A, (x')A 4 (x)) f J_PF (x)J(s') — 4,(x)A, (4) f ,11 x)J(4]
` ( (ti )

On using this result together with eqn (12.80), we can cast eqn (12.77) into the
following form,

tf

PM( t f) = exp [ f dix.Cc (x) (12.83)

t .f tf
1
f di x f d 4 x'{ — (8(t — A, (s')] + A ii (x)) f)J_PF (x)JfF (s')
t2 ti
+(t9(t — t')[A(x), A v (x i )] — (A 4 (x)A,(4) f ) (x),P'(x')

+0,(xi )A 4 (x)) f J(x)J''(x') + (A Ii (x) A, (4) f (x)J fF (x')}] p rn (t i)

 THE INFLUENCE FUNCTIONAL OF QED 581

At this point it is useful to introduce some new notation for the correlation
functions of the electromagnetic field, namely the Feynman propagator and its
complex conjugate (T_, denotes the antichronological time-ordering),

iDF(x — (T,[A„.,(x)Av(x i )i)f (12.84)

= 0(t — t')[A p (x), A v (x1 )] + (Ay (x')A p (x» f ,
iD*F (x — — (T _,[A m (x)A v (x')]) f (12.85)
= 9(t — t')[A m (x), Av (x i )] — (A m (x)A v (4) f,

as well as the two-point correlation functions

D+(x — x'), (A p (x)A 11 (4) f (12.86)

(A v (x')A m (x)) f (12.87)

As is easily verified these functions are related through the identity

— iDF(x + iD(x x'),„„ + D±(x — + D _(x — = 0.

(12.88)
With the help of this notation the density matrix of the matter can now be
represented in the compact form

pm(t1) = T exp (ic1.[J+, J_]) prn (ti), (12.89)

where we have introduced the influence phase functional

t1 t1 tf
1
i(1)[4, J_] f
= d4 x fc(x) + — d4 x d4 x i (12.90)
2

x { —iDF (x — x 1 ),„„J'!-Ft (x)f ±v (x') + iD F* (x —

+D_(x — x').E FL (x)Jv (xl) + D + (x —

Equation (12.89) provides an exact representation for the matter density matrix
which takes on the desired form: It involves the electromagnetic field variables
only through the various two-point correlation functions introduced above. One
observes that the dynamics of the matter variables is given by a time-ordered
influence super-operator, that is a time-ordered exponential function whose ex-
ponent is a bilinear functional of the current super-operators 4(x).
It should be remarked that the influence phase c13[4, J_] is both a functional
of the quantities 4(x) and a super-operator which acts in the space of density
matrices of the matter degrees of freedom. There are several alternative methods
which may be used to arrive at an expression of the form (12.90) as, for example,
path integral techniques (Feynman and Vernon, 1963) or Schwinger's closed time-
path method (Chou, Su, Hao and Yu, 1985; Diôsi, 1990).
582 OPEN QUANTUM ELECTRODYNAMICS

For the study of decoherence phenomena another equivalent formula for the
influence phase functional will be useful. To obtain this formula we use the
commutator and the anticommutator functions which have been introduced and
determined explicitly in Section 12.1. These functions are related to the above
correlation functions through
1 m
D±(x - = -D (X - XI ) - DT (X - xi ) (12.91)
2 1 Pv 2
1
D _(x - x')= - DT (x x') + DT (x - ) /iv , (12.92)
IL 2
1
iDF(x s i ) = - DT (x x') - sign(t - tI )DT (x x') Iiv, (12.93)
"v 2 1 I"' 2
1 ,
- iD (x - = - D (x x') „„ + - sign(t t')D T (x - x,. (12.94)
" 2 1 2

Correspondingly, we define a commutator super-operator Je (z) and an anticom-

mutator super-operator Ja (x) by means of

J L (x)P E [im (x), (12.95)

ffit (x)P E te(x), PI, (12.96)

which are related to the previously introduced super-operators J (z) by

(x) = (x) J (x), (12.97)

Jf`i (x) = (x) + J (x). (12.98)
In terms of these quantities the influence phase functional becomes
tf

i4:13[Je , Ja ] = f d4 x fc (x) (12.99)

ti
tf

+ f d4 x f d 4 x' D T (x x') (x)Jau (x') - -1 Dx

T( xi 4‘ (x } •
2 2 1
ti ti

This form of the influence phase functional will be particularly useful later on. It
represents the influence of the radiation field on the matter dynamics in terms of
the two fundamental two-point correlation functions D(x - x') and D 1 (x - x').
Note that the double space-time integral in eqn (12.99) is already a time-ordered
integral since the integration over xio extends over the time interval from t i to
t = X0
The result expressed by eqns (12.89) and (12.99) has been used in Section
3.6.4 for the determination of the influence functional of the Caldeira-Leggett
model (see eqns (3.507) and (3.508)). In fact, for the Caldeira-Leggett model we
have to use a coupling of the form 1/1 B (see eqn (3.377)). Going through
 THE INFLUENCE FUNCTIONAL OF QED 583

the above derivation we see that in this case the influence functional (12.99)
takes on precisely the form given in (3.508): The commutator and anticommu-
tator super-operators Jg, Jfit must be replaced by the super-operators x, and
Sa defined in eqn (3.510), while the counter-term (3.509) plays a similar rôle to
the Coulomb term .Cc (x). Furthermore, the correlation functions DT (x
and DT(x - x')„„ must be replaced by the correlation functions of the Caldeira-
Leggett model defined in eqns (3.385) and (3.386).

12.2.2 Vacuum to vacuum amplitude

- -

It is instructive to compare eqns (12.89) and (12.90) with the structure of the
Markovian quantum master equation in Lindblad form containing a set of Lind-
blad operators A i (see Section 3.2.2). One observes that the terms of the influence
phase functional involving the current super-operators in the combinations J+J -

and J_ 4 correspond to the gain terms in the Lindblad equation having the form
Ai pni Ati . These terms may be interpreted as describing the back-action on the
reduced system of the matter degrees of freedom induced by 'real' processes in
which photons are absorbed or emitted. The presence of these terms leads to a
transformation of pure states into statistical mixtures. Namely, if we disregard
the terms containing the combinations 4J_ and J_ J+ the right-hand side of
(12.89) can be written as U (t f, ti)p m (ti )Ut (t f , ti ), where
tf
U(tf , ti ) = exp f d 4 x 7 ic (x)
- (12.100)
t,
tf tf

- - f d 4 x f d 4 5'DF(x x i LiviAL(x)iv(x 1 )

This shows that the contributions involving the Feynman propagators and the
combinations J+ J+and J_ J_ of super-operators preserve the purity of states.
Taking the limits ti -co and t oo the operator U(tf , ti) becomes
-

the time-ordered product of the functional

A[j] exp [-i f d4 x c (x) - f d 4 x f d 4 D F -

E eXp [i (s (1) iS(2) )] . (12.101)

At zero temperature this expression is the vacuum-to-vacuum amplitude in the

presence of a classical current density ji2 (x) (Feynman and Hibbs, 1965). It yields
the amplitude for the field to start in the vacuum at ti = -oc and to end up in the
vacuum at t + oo . In the second equation of (12.101) we have decomposed the
exponent into real and imaginary parts, using the corresponding decomposition
of the Feynman propagator as given in eqn (12.93).
584 OPEN QUANTUM ELECTRODYNAMICS

The exponent of A[j] provides a complex action functional S 5 (1) + j5(2).

Its imaginary part is found to be

s(2) = f d4 x f d'x'Di(x - (x) • I-T (x i )

= f f d4 x' (x — x i )j1,(x)im (x'). (12.102)

In the first line jT denotes the transverse component of the current density,
satisfying tr* • j'T = 0. The coupling to the transverse component is due to the
fact that DT (x x') p,„ contains the transverse projection Pp,. In the second line
of (12.102) we have used the current conservation (12.60) to cast the expression
into covariant form. To see how this form emerges we transform into Fourier
space, using (12.38) and the Fourier transform of the current density,

(k) = f d4 xe -d" (x). (12.103)

This leads to

s(2) 1f d3 k (ij • j(k))(k# • 1(k)) *

coth(w/2kBT) (Pk) • j* (k)
J 2(270 3 w
1 f d3k
coth(co 12kBT) RV:1(k)) (12.104)
J 2(270 3 w
where we have used

(k" • 1 (k)) (lc* • (10)*

1(k) • j* (k) == 1- (k)'P(k) -- jo(k)iô(k)

= (12.105)

which follows from current conservation, km ?' (k) = 0. Thus we have

exp [-28 (2) ] (12.106)

d3 k
= exp {—
j 2(27 )3w
coth(w 12kBT) (k))1
This is the no-photon emission probability, that is the probability that the cur-
rent density does not emit any photons (Itzykson and Zuber, 1980). We see that
for finite temperatures, T > 0, the no-photon emission probability is reduced in
comparison to the vacuum case. This is just the effect of induced emission and
absorption processes.
 THE INFLUENCE FUNCTIONAL OF QED 585

By use of the relation (12.30) and of current conservation the real part of the
action functional if found to be

8117 f d 4 s f d4 si { 6(1 2_ zx, 1°I ) jo(x)jo(x i ) — 6 [(x — x /) 2 ] (x) • ji, (x 1 )}

SO) — _-

f d4 x f d I x'6[(x x') 2 ] jii (x)Y -1 (x`). (12.107)

This is the classical Wheeler—Feynman action (Wheeler and Feynman, 1945,

1949; Rohrlich, 1965). It allows the description of the interaction of a system of
charged classical particles by means of a non-local action principle. The degrees
of freedom of the electromagnetic field have been completely eliminated under
the boundary condition of no net emission of radiation (which is known as the
complete absorber theory). As a result one is left with a non-local action func-
tional which contains both the retarded and the advanced Green function (see
eqn (12.30)).

12.2.3 Second-order equation of motion

The representation (12.89) immediately yields the following second-order equa-
tion of motion for the density matrix of the matter degrees of freedom,
d
— Pm(t) = [H(t), Pin(t)] (12.108)
dt

f d3 f d 3 s' f ds'o D(x — s i )[-IT(x),{ 1- T(x 1 ),pni(t)}]

ti

1
—
2
f d3 x f d3 X' f &Pi (X — X') {JT (X), {jr (x 1), Pm (t)} ,
ti
where s = (t, i) and x' = -±# ') This equation is just the time-convolutionless
master equation to second order in the coupling. As we know it provides a non-
Markovian master equation since it involves the non-local dissipation and noise
kernels D(x s') and D i (x — s'), respectively. The various master equations
encountered in the previous chapters can be derived from this equation. For
example, the quantum optical master equation is obtained from it by performing
the Markovian and the rotating wave approximation.

12.2.3.1 The quantv,m optical limit It might be instructive to sketch how the
quantum optical master equation can be derived from eqn (12.108). To this
end, we insert the Fourier representations of the Green functions D(x — y) and
D i (x — y) (see eqns (12.20) and (12.38)) into eqn (12.108). Using then the Fourier
components of the transverse current density which are defined by

jT (X0 , 1-j) = f d 3s jT (so, Z)e -ik.Y (12.109)

 ▪
586 OPEN QUANTUM ELECTRODYNAMICS

the space integrals over i and i t can easily be evaluated. Next, one performs
the Markov approximation which consists in the replacement

f dx`o
f
-00
dsio = f dr, (12.110)

where we have substituted T E t x ot . Finally, we introduce the decomposition

of the current into eigenoperators of the system Hamiltonian, such that

an (X0, = Ee JT(Wrt,k). (12.111)

con
The sum runs over the frequencies co n of the system defined by the energy differ-
ences of the unperturbed system. This representation enables us to employ the
rotating wave approximation in the quantum optical limit. The latter consists
in keeping only the secular terms in double sums over the system frequencies w
and is equivalent to an averaging procedure over the rapidly oscillating terms.
Putting it all together we obtain the following master equation,
d
— pm (t) = [He (t), pm (0] i [His + H ss, pm (t)] (12.112)
dt
1
+ E -y- (Wn) f (k) (irr (con, k*) Pm itT (‘-‘)n — {Pr (wn, (con, TO, Pm })
wn>0
r
+E (wn) f cin (k) (Wn, k)Pin.ir(con,k') liT(wn , k4)74(wn , Pm})

where dQ(k) denotes the element of the solid angles in the direction of the
wavevector k.
The last two terms on the right-hand side in (12.112) represent the dissipator
of the master equation describing incoherent processes, namely induced absorp-
tion processes, the rates of which are proportional to -y ± (wn ) = wn N(con )/870,
as well as induced emission and spontaneous processes taking place with a rate
which is proportional to 7_ (w ii ) = wn [N(w n ) + 1]/87r 2 .
The coherent Hamiltonian part of the dynamics contains the Coulomb in-
teraction H. Additionally, it is modified through the presence of the radiation
field which leads to contributions from the Lamb shift Hamiltonian

HLS 'EP f d3k 1 -

2 (270 3 w w n - w h(wn,k) 1T(C'in,k)
(12.113)
W71,

and from the Stark shift Hamiltonian

f d3 k N
Hss = (Pr (wn, rc).7T (wn , — (wn, NtT (wn, k))
2(27) 3w wn -w
con
(12.114)
 THE INFLUENCE FUNCTIONAL OF QED 587

It is easy to check that the density matrix equation (12.112) leads to the various
forms for the quantum optical master equation used in previous chapters (note
the change of units here).

12.2.3.2 Mass renormalization and Lamb shift Let us have a closer look at the
Lamb shift Hamiltonian (12.113). The contribution from IILs is formally infinite
and must be renormalized according to the renormalization procedure of QED
(see, e.g. Weinberg, 1995). For our purposes it suffices to make a few remarks on
the non-relativistic treatment of this term.
Invoking the dipole approximation and introducing the spectral decomposi-
tion of the system Hamiltonian through Hs = En En in)(ni, we find for the shift
AE, of the n-th level induced by HLS:
i-2 in ax
e2 w
AE rt = 67 2 77-1 2 EP
mOn f
0
„b.,
w _ Wmn
I ( 71 1/51 74 2 , (12.115)

where comn = Em — En and ft max is an ultraviolet frequency cutoff. To leading

order, the frequency integral diverges linearly with the cutoff. To split off the
linearly divergent part we write the integrand as
(.4.) Wmn
=1+
W — Wmn W — Wmn

Correspondingly, the energy shift consists of two parts,

AEn = AE nt +
The first part is given by

E f( I
in ax
e2 e
20
%) g, max • 1 -.2
kJ (n1/51 111) 1 2 67 2 7112 MP In). (12.118)
67 2 m2
mn 0

This term may be interpreted as the matrix element which stems from a renor-
malization of the mass in the kinetic energy of the system Hamiltonian. In fact,
if we write the physical mass m = m0 + 6m as a sum of a bare mass m0 and an
electromagnetic mass contribution 6m, the total kinetic energy reads
,--.2 ,-*2 ,--.2
.1 ' Y •-...• Y 6m 42
(12.119)
-

— = •-...•
2m 2(m o + 6m) 2m0 2m0 .
Comparing this with (12.118) we see that the mass correction is given by

e 2 0 max
6m = 372 . (12.120)

This is the same mass renormalization as found from the classical Abraham-
Lorentz equation for the electron (see Section 12.3.4). If we take the cutoff Qmax
588 OPEN QUANTUM ELECTRODYNAMICS

to be on the order of the electron mass, we find that the mass correction is small,
Im = 0/37 2 0.0031. We note that in a full relativistic treatment Sin only
depends logarithmically on the cutoff, that is we have (Bjorken and Drell, 1964)

3e 2771 Cmax
ôni = 872 ln (12.121)

The second part of the energy shift in (12.117) is given by

t-2 1-111:16X

= 6
67 2m2
2
EP dw Winn 1(nlfr1)1 2 '
— W inn,
(12.122)
rnOn

In our present non-relativistic calculation this expression gives the main contri-
bution to the observed atomic level shift caused by the vacuum fluctuations of
the electromagnetic field. The sum over intermediate states can be worked out
and yields the famous Bethe formula for the Lamb shift in hydrogen (Bjorken
and Drell, 1964).

12.3 Decoherence by emission of bremsstrahlung

As a further application of expression (12.99) for the influence phase functional
we investigate in this section the destruction of quantum coherence through the
emission of bremsstrahlung. To this end, we investigate a simple, prototypical
interference device for a charged particle and ask for the loss of coherence which
is induced by the interaction with the radiation field. The interference device
involves two possible paths of the charged particle: The initial wave packet is split
into two components which first move apart and which are then recombined to
measure locally their capability to interfere. This type of interference experiment
thus involves a relative motion of two spatially separated components of the wave
function. It turns out that the possibility of the emission of radiation leads to a
(partial) destruction of quantum coherence which results in a reduction of the
interference contrast.
An appropriate measure for the decoherence in the interference experiment
is a certain relativistically covariant and gauge-invariant functional, the decoher-
ence functional, which will be derived and determined explicitly in the present
section. It will also be demonstrated that the obtained decoherence functional in-
volves the typical features of the emission of bremsstrahlung. Explicit analytical
expressions for the vacuum and the thermal contribution to the decoherence func-
tional and for the corresponding coherence lengths are determined. These expres-
sions reveal that bremsstrahlung leads to a fundamental decoherence mechanism
which dominates for short times and which is present even in the electromag-
netic field vacuum at zero temperature. The influence of bremsstrahlung on the
centre-of-mass coordinate of a systems of many identical charged particles is also
studied and shown to lead to a strong suppression of quantum coherence.
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 589

W2 )

FIG. 12.1. Sketch of a prototypical interference device which is employed to

introduce the decoherence functional. An electron emerges from the source
Q and can follow two possible paths Yi and y2, which leads, in general, to an
interference pattern observed on a screen at S. The two quantum alternatives
may be described through the wave packets IT 1 ) and ITO.

12.3.1 Introducing the decoherence functional

Let us consider an interference device of the type sketched in Fig. 12.1. A charged
particle, say an electron, is emitted by the source Q and can move along two dif-
ferent world lines yi and y2 to reach a screen at S, where an interference pattern
is observed. These paths represent two quantum alternatives whose probability
amplitudes may be described by two wave packets I (ti )) and 11P 2 (ti )). With
the help of the superposition principle we find that the wave function

= Ri(t)) + i 4'2(ti)) (12.123)

describes the physical situation depicted in the figure. Alternatively, the electron
state can be represented in terms of the density matrix pm (ti) = IT(ti))(T(ti)I
which may be written as

Pm(ti) = + p22(ti) + p 12 (ti ) + p21 (ti ), (12.124)

where pafi(t i ) = IT a (t2 ))(T o(t i )I with o43 = 1, 2. One observes the emergence
of the interference term p 12 (ti ) + P21 (t i ). Remember that we are working in the
interaction picture and that we therefore have pm (t) = pm (t i ) for all times in
the case of a vanishing coupling between matter and electromagnetic field.
590 OPEN QUANTUM ELECTRODYNAMICS

Our aim is to determine with the help of the influence super-operator the
structure of the electron density matrix pm in the presence of the electromag-
netic radiation field. An essential simplification is achieved if the matter current
density can be treated as a classical current. This approximation can be justified
under the following conditions. First, we assume that the wavelength A = c/w
of the photons emitted by the currents is large compared to the Compton wave-
length Ac of the electron,
h
A » Ac = (12.125)
me ,
and, thus, also large in comparison to the classical electron radius re = ac r^:J
2.8 x 10 -15 m. This requirement is equivalent to hw < mc2 . In this low-energy
regime one may neglect pair creation and annihilation amplitudes and treat
the matter current density as a given classical field (Jauch and Rohrlich, 1980;
Cohen-Tannoudji, Dupont-Roc and Grynberg, 1998). The same procedure is
used, for example, in the non-perturbative analysis of radiative corrections in
the low-frequency limit (see below). In an experiment of the type sketched in
Fig. 12.1 the paths involve an acceleration of the electron through a certain field
of force. This force gives rise to a certain characteristic acceleration time Tp . We
define Tp as the inverse of the highest frequency in the power spectrum of the
force acting on the electron. In the following we call Tp the preparation time since
it can be interpreted as the time required to set into motion the interfering wave
packets. As a consequence of the existence of such a characteristic time we have
a natural upper cutoff Rna,, for the frequency spectrum of the emitted radiation
which is of the order
1 c
9 max r - = -
.....,
,
(12.126)
"rp go
where the length scale o-0 represents the order of the minimal wavelength of the
radiation. Our above requirement thus takes the form

ao >> Ac. (12.127)

This also implies that the characteristic acceleration time Tp is large compared
to re/c. It is known from classical electrodynamics that this condition ensures
that the energy radiated is small compared to the kinetic energy of the particle
and that therefore radiative damping effects are small (Jackson, 1999) (see also
Section 12.3.4).
The second condition is that the motion of the current can be reasonably
described within a semiclassical approximation. This leads to the requirement
Ay/y < 1, where y is a typical velocity and Ay the velocity uncertainty. Assum-
ing that the wave packets represent states of minimal uncertainty with spatial
width Ax one is led to the condition
Ay h
,
<1 (12.128)
y myAx '
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 591

or, equivalently,

AdB
(12.129)
<<
where Ads = h/mv is the de Broglie wavelength. This is the typical condition
for a semiclassical treatment.
In view of these conditions we now assume that p(ti) represents a state
which is an approximate eigenstate of the current density. Thus, if pm (ti) is a
pure state,

we suppose that

where .91 (x) is a classical current density. Hence, we also have to the same degree
of accuracy,

f(x pin ti) = Li"(x),

) ( pm (tt)] O. (12.132)

The initial state p(t i ) does not necessarily have to be a pure state. It suffices
to require (12.132), where
(j/.L W) = trm (x)p m (ti)} = e(x) (12.133)

is the expectation value of the current density. In any case we immediately obtain
with the help of (12.132) and expression (12.99) for the influence phase

pm(tf) pm(ti). (12.134)

This equation states that the system is essentially unaffected by the radiation
field, i.e. by virtue of our assumption that the initial state is an approximate
current eigenstate, the dynamics of the density matrix is nearly the same as that
of the free system. The same conclusion can be drawn from an investigation of
the dynamics of a Gaussian wave packet under the influence of the radiation
field by use of the exact analytical expression for the propagator function in the
dipole approximation (see Section 12.3.4).
Let us now return to the interference device and assume that the superposi-
tion (12.123) consists of two approximate current eigenstates,

iNx)iTi(ti)) '^dd si(x)Ri(ti)), (12.135)

iNx)IT2(ti)) k'dd 4(x)IT2(ti)), (12.136)

where s i (x) and 8 2( 5 ) are classical current densities. These currents are assumed
to be concentrated within two world tubes around the paths Iii and y2 of the
592 OPEN QUANTUM ELECTRODYNAMICS

C =o

FIG. 12.2. A closed current world tube indicating the support of the current
difference density c# = sn/Nd The world tube is located around the
closed path C = y l — y2 formed by following Yi in the positive and y 2 in the
negative direction (see Fig. 12.1).

interference device, respectively (see Fig. 12.2). By virtue of eqn (12.135) we

have

4t (x)P11(ti) ;zid 4L (x)p22(ti) 0, (12.137)

(x) ion (ti ) (8'11 (x) — (x)) p12(ti), (12.138)
Jfit(x) p i 2 (ti) (4(x) + 4 (x)) pi2(ti), (12.139)

and

f (x)p i (4) c( x )P2 2 ( t ) 0, (12.140)

re (x)P12 (ti) (x) 9-1c2(x)) )912 (t i ), (12.141)

where

f
Rc i , 2 (x) = 12 d3y (12.142)
471i— yi
are the Coulomb energy densities associated with the current densities (x) and 4
$'IL (x), respectively. We may suppose that the corresponding Coulomb energies
 DECOHERENCE BY EMISSION OF BREMSSTRAHLLING 593

for both possible paths are equal to each other. The expression (12.99) for the
influence phase functional now immediately leads to

Pm(tf) pli (ti) + P22 (ti) exp(i4)p 12 (ti ) + exp(—i43*)p 2i (ti ), (12.143)

where

i(13 821 (12.144)

tf
= f d4 x f d 4 (sm (x) — 4(x)) (4(x') +
DT (x — x') pv (x'))
2
ti ti
tf tf
— 1- f d's f d4xiDrir(x — (4(x) — 4(x)) (4(x') — s12'(x')) .
4
ti ti

We observe that the electromagnetic field affects the interference terms through
a complex phase 4. = 41[8 1 ,82] which is a functional of the two possible classical
paths Yi and y2, or, more precisely, of the associated current densities 4(x) and
4(4 The real part of 4. leads to a distortion of the interference pattern. The
imaginary part of 4), on the other hand, yields a suppression of the interference
terms in (12.143) given by the factor

lexp(i4)) I exp F. (12.145)

The exponent F will be referred to as the decoherence functional (compare the

discussion in Section 4.1). From the second term in eqn (12.144) we infer that F
represents a bilinear functional of the current difference

(x) = (4(x) — 4(x)) (12.146)

and may be written as follows,

tf tf
1
F[c] = -- di x f d 4 (x — x 1 )c# (x)ev (xi (12.147)
2
ti ti

12.3.2 Physical interpretation

To bring F[c] into a more convenient form let us define to to be the time cor-
responding to Q (see Fig. 12.1), that is the time at which the wave packet is
separated into two components, while tf denotes the final time when both pack-
ets are recombined at S. It is obvious that the current difference c4 (x) vanishes
for times prior to t o and for times later than the final time tf . In fact, the support
594 OPEN QUANTUM ELECTRODYNAMICS

of the current difference cA(x) lies in the interior of a closed world tube around
the loop

C = yi — y2 (12.148)

which is formed by following yl in the positive and y2 in the negative direc-

tion, as illustrated in Fig. 12.2. Current conservation, OA & = 0, therefore en-
ables us to write the decoherence functional as (compare the discussion following
eqn (12.102))

F[c]
21 f d f d4 x / D1 (x — s') [ — di (x)cli (4] -
.4 x
(12.149)

We observe that F[c] represents a relativistically covariant and gauge invariant

functional. Moreover, F[c] is Lorentz invariant in the vacuum case as may be
seen explicitly from the fact that Dr"(x — x') is given by the invariant function
(12.39).
Equation (12.149) suggests several interesting physical interpretations. With
the help of the Feynman propagator (12.84) and the two-point correlation func-
tions (12.86) and (12.87) we can express the complex phase factor exp(i(1)) in
eqn (12.143) as follows,
exp(i43) = A[81]A[s2] * (12.150)
1
x exp [- f di x f dix'D_ (x — x') Av .4(x)s 2v (x i )
2
+ 1- f d4x f d 4x 1 D+(x —
2
where A[j] is defined in eqn (12.101). This form for the phase factor was first
derived by Ford (1993) for the case of zero temperature with the aim of studying
the influence of conducting boundaries on electron coherence. As we have seen,
at zero temperature A[j] is the vacuum-to-vacuum amplitude in the presence of a
classical current density jA(x). The first term on the right-hand side in (12.150)
is thus the product of the vacuum-to-vacuum amplitudes in the presence of
the current densities slit and .5 -'. This contribution describes virtual processes
in which photons are emitted and reabsorbed by either the currents sii. and 4.
Correspondingly, the exponential on the right-hand side of eqn (12.150) is the
contribution of the emission of real photons. These processes also contribute to
the decoherence functional since a photon can be emitted by both currents and
carries away information on the path taken by the electron. Moreover, at finite
temperatures thermally induced emission and absorption processes occur.
Expression (12.150) also answers the question as to whether virtual vacuum
processes or real photon emissions are responsible for the decoherence effect
expressed by the functional F[c]. We see that the decoherence factor (12.145) de-
pends on virtual processes through the Feynman propagator DF(s -- x l ) Av (which
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 595

is contained in the vacuum-to-vacuum amplitudes) as well as on real processes

described by the Green's functions D±(s—xi),,„. Thus, it is the combined effect
of virtual and real processes which leads to decoherence. The physical reason
is that the mere possibility of real photon emissions also reduces the vacuum-
to-vacuum amplitudes which can already lead to a reduction of the interference
contrast. For example, one can think of an interference device in which the path
Yi describes a uniform motion of the electron, whereas the other path y2 is
strongly bent, involving a large acceleration of the electron. Suppose we observe
the photons emitted in the experiment. If we know that no photon has been
emitted it is then very likely that the electron has taken the path y i , which
results in a suppression of the interference pattern formed by the electrons of
the corresponding sub-ensemble. This situation is similar to that of a two-level
atom which is initially in a superposition of the excited state and the ground
state. If we find that the atom did not emit a photon during a large time in-
terval (large compared to the inverse of the emission rate) we have effectively
measured the state of the atom to be the ground state. Thus, the off-diagonal
terms of the atomic density matrix approximately vanish without the emission
of real photons.
It is also interesting to express the decoherence factor directly in terms of the
amplitude A[c]. Invoking (12.91) we find

exp (F[c]) = IA[c]1 2 . (12.151)

Obviously, we have F[c] < 0, and F[c] = 0 for s ii' = si t , that is for a vanishing
current difference, c11 = O. Equation (12.151) gives rise to another interesting
interpretation: The decoherence factor which multiplies the interference term
is given by the no-photon emission probability in the presence of the current
density 02 . This current is the same as the current which would be created
by two particles with opposite charges ±e1-4, one moving along y i and the
other along y2 . The smaller the vacuum-to-vacuum amplitude for this current
density the larger the reduction of the interference contrast. This must have
been expected since it is the difference between the currents s i and s 2 which
determines the extent to which the two possible paths can be distinguished, and,
thus, the degree of the loss of coherence.
These interpretations in terms of the emitted photons must be taken, how-
ever, with some care. The reason is that we consider here processes on a finite
time scale and not transitions between asymptotic states. It is well known that
certain matter currents emit an infinite number of long-wavelength (soft) pho-
tons whose frequencies approach zero, while their total energy adds up to a fi-
nite value. This is the so-called infrared catastrophe (Weinberg, 1995; Jauch and
Rohrlich, 1980) which arises in the perturbative calculation of radiative correc-
tions to any process involving charged matter. The complete removal of infrared
divergences requires a non-perturbative treatment in which the amplitudes for
the emission of real and virtual soft photons are summed to all orders, such that
the processes involving real and virtual photons become indistinguishable in the
596 OPEN QUANTUM ELECTRODYNAMICS

low-frequency limit. Infrared divergences can be shown to cancel provided a finite

resolution Ruin for the photodetection is introduced: Insisting on the perturba-
tive picture, one could say that there is always an infinite number of quanta,
namely those whose frequency is lower than Stmin , which escapes undetected and
cannot be observed in principle.
Our analysis treats the matter current classically but it is non-perturbative
(Jauch and Rohrlich, 1980). In view of the above considerations it is obvious
that the decoherence functional F[e] does not lead to infrared divergences since
it describes a process taking place in the finite time interval between the splitting
of the wave packet at t o and the recombination at t1 . This gives rise to a natural
frequency resolution of the order
1
Qmin (12.152)
tf — to

The emergence of this effective infrared cutoff will be seen explicitly in the cal-
culations of the next subsection, where it will be demonstrated that the arising
integrals over the photon frequencies converge at the lower limit w —4 0. In
addition we also have an ultraviolet cutoff Stmax which has already been intro-
duced in eqn (12.126). This cutoff can be accounted for by the introduction of a
finite width ao characterizing the current world tube, as will be seen in the next
subsection.

12.3.3 Evaluation of the decoherence functional

We wish to evaluate here the decoherence functional for some specific situations.
On using eqn (12.149) and the Fourier representation (12.38) for the anticom-
mutator function D 1 (s — s') we find

f d3k
F[c] = — 3w coth(w/2kBT) [—d-i(k)eil(k)] , (12.153)
2(27 )

where we have introduced the Fourier transform

c(k) f d i xexp(—ikx)e(x) (12.154)

of the current difference c1-' (s).

Let us first show explicitly how a finite width of the current world tubes
gives rise to an ultraviolet cutoff scale. To this end, the currents s(x) and 4(x)
are taken to be concentrated within world tubes of spatial extent cro around the
world lines yi = (7- ) and y2 = y 2 (7.). The world lines are parametrized by their
proper time T such that

= dY2(T) (12.155)
di-
are the corresponding 4-velocities. To be specific we write
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 597

1,2 (x) = e f dy 1, 2 (7)6 cro (x - y i , 2 (T)), (12.156)

where
1 [ (- gi)21
Sao (X — x l ) = 6(X0 x) exp (12.157)
u 2o-ci
is a smeared 6-function described by a Gaussian with width oro . We remark that
we neglect the spin contribution to the current density. This is justified as long as
the length scales involved in the problem under consideration are large compared
to the Compton wavelength (Ford, 1993).
On using eqn (12.156) in eqn (12.154) the Fourier transform of the current
difference is found to be

cl (k) = — [f dr (r) exp(-ikYi (7)) - f dr u(r) exP( -ikY2(r))1

x exp [__1 0.02 w 2] (12.158)

2
We see that the finite width ao of the current world tubes yields an effective
ultraviolet cutoff Q. - a-0-1 as given in eqn (12.126). Our main interest is
an estimation of the decoherence functional for some specific situations. We
therefore ignore in the following the specific form of the cutoff function in (12.158)
and work with a sharp cutoff at the maximal frequency co = Q max . It will be seen
below that the final expression for F[c] depends on Q inax only through ln Qmax .
This extremely weak logarithmic dependence shows that the precise value of o-0
or of the preparation time Tp is rather irrelevant. The important point to note
here is that the emergence of an effective ultraviolet cutoff has a clear physical
origin.
Thus, we now write eqn (12.153) as

C2 max
1
f cico coth(co/2kBT) f (100 [-c1L (k)e7,(k)] , (12.159)
F[c] 167r3

where c/Q(k) denotes the element of the solid angle in the direction of the unit
vector k Oki. The Fourier transform of the current difference can be expressed
as a loop integral over the closed loop C = yi - y2)

exp( -ikx). (12.160)

For simplicity let us consider the case that the loop C consists of four straight
world line segments (see Fig. 12.3(a)). The vertices of the loop are denoted by
ao , al , a2, a3 , whereas the corresponding 4-velocities are given by
598 OPEN QUANTUM ELECTRODYNAMICS

a) b)

a()

FIG. 12.3. (a) The closed loop C = Yi — y2 used for an explicit determination
of the decoherence functional F[c]. The loop consists of four straight world
line segments with 4-velocities u i , u 2 , u 3 , u4. The vertices are located at the
space-time points ao , a l , a2 , a3 . As indicated, the loop corresponds to the
total time 2tf. (b) A single line segment with initial point a, endpoint b and
4-velocity u.

ulri, = 712, - ( 1 _ liln i2) - 112 . (12.161)

We further assume that the arrangement is symmetric, that is we suppose that
u 1 = u4, tt2 = u3, and that a l — ao = a2 — a3, a2 — al = a3 — ao. For a single
line segment with initial point a, endpoint b and 4-velocity u (see Fig. 12.3(b))
we obtain
b
f dx exp(—ikx) i_u re —ikb
ku L
_ e —ikal (12.162)
a

With the help of this formula the Fourier transform (12.160) of the current
difference is found to be,

C(k) =
ie
+_
111 re —ikai e —ikaol ± U2
_ [e 2 e —ikail
ku 1 L
—

kU2

—
u3 r _ika3
— Le - —e ol _ U4 [ e _ika 2 _ e —ika31] . (12.163)
ku3 ku4
It should be noted that kli cP(k) = 0 as required by current conservation. Note
further that ell (k) shows the correct behaviour under Lorentz transformations. In
particular, one finds that 1L (k) transforms into cA (k) exp( — ikb) under a space-
time translations by the 4-vector b. If we now use the symmetry properties of
the loop we arrive at
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 599

ie
c(k) = — u
2

\/ kU2 kU4
U4 I (12.164)

where we have introduced

Ç(k) = e —ika2 (1_ e ik(a2—a3)) ( 1 e

ik(a
(12.165)

Using these results one is led to the following expression for the decoherence
functional,

Oma.
2
dw
F[c] = — u2 — u4
w coth(w/2kBT) f dft(k)w 2 [ ku2 ig(k)1 2 . (12.166)
ku4
o

We denote the time interval associated with a single line element of the loop by
t f , that is we set to = —t f and t to = 2t (see Fig. 12.3(a)). Then we have
—

k(a2 — a3) = cotf (1 — k • 174), (12.167)

k(a2 — ) = wtf (1 — k • 172)• (12.168)

In order to estimate further the expression (12.166) we approximate

k(a2 a3) k(a2 al ) wt1 , (12.169)

which leads to

1
Ig(k)1 2 =8 [(I — coswtf)— :4 (1 — cos 2wtf)1 . (12.170)

This allows us to estimate the decoherence functional (12.166) as follows,

max

a dw tot. TT\ f 1 fi
F[c] f — z&B. — cos wtf) — — y_ — cos 2 wtf )1
4
o
[ U2 U4 2
x f dft(k)w 2 (12.171)
kU2 kU4

Let us first concentrate on the integral over the photon frequencies w, that is
on the first integral on the right-hand side of (12.171). The integrand is propor-
tional to which is a typical signature for the spectrum of bremsstrahlung
(Jackson, 1999). In addition to vacuum bremsstrahlung there may be thermally
induced emission and absorption processes (Itzykson and Zuber, 1980), which
are embodied in the factor coth(w/2kB T). At zero temperature (vacuum field)
this factor may be replaced by 1.
600 OPEN QUANTUM ELECTRODYNAMICS

In order to calculate the frequency integral it turns out to be useful to decom-

pose it into a vacuum contribution and a thermal contribution which vanishes
for T = 0. We therefore write
o max
— coth(w/2kB T) (1 — cos wtf)
FE
f
0
(lbw,
= Fvac ± Fth ) (12.172)

where

(1 — coswtf) (12.173)

is the vacuum contribution, while

u max
Fth E — [coth(w/2kBT) — 11(1 — cos wtf) (12.174)
f t
o
is the thermal contribution. The frequency integral F,a, can be evaluated as
follows. Substituting x = wtf we get

QMEtX tf

FVaC (1 — COS X) = ln (02,,,ax tf) ± 0 1 (12.175)

x ( 91-naxtf ) 1

o
where lng ',:,-,' 0.577 is Euler's constant (Gradshteyn and Ryzhik, 1980). For
Qmaxtf >> 1 we thus have asymptotically
In (gRnaxtf) . (12.176)

This relation demonstrates that the vacuum integral over the photon frequen-
cies converges at the lower limit w --> 0 and that it gives rise to an effective
infrared cutoff of order ft mir, — Of, as discussed in the previous subsection
(see eqn (12.152)). We also observe that the vacuum frequency integral increases
weakly with the logarithm of Qmax tf . . As indicated in eqn (12.176) we keep for
simplicity in the following only the leading contribution in our expressions. It
should be kept in mind, however, that one can include without difficulty the
terms of higher order which vanish in the limit Qmax tf —4 oc. The integral
(12.173) has already been evaluated in Section 4.2.2, where we have used, how-
ever, an exponential cutoff. A comparison with the formula (4.54) shows that a
change of the cutoff shape introduced only a small change in the value of the
integral.
The thermal contribution Fth has also been evaluated in Section 4.2.2 (see
eqn (4.55)) under the condition kBT < hRua„, which will also be assumed here.
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 601

To give an example for this condition we take the ultraviolet cutoff 12„,,, --,
10 19 s -1 , corresponding to a length scale of order 100Ac. The above condition
then means that T < 108 K. Thus we obtain with the help of (4.55)

(sinh (tf /TB) )

Fth ln (12.177)
tf /TB J .

On using the results (12.176) and (12.177) we can now determine the frequency
integral in eqn (12.171),

St rna.
dw 1
J
o
—
c,,.,
coth(w / 2) [(1 - cos wt ) - - (1 - cos 2wtf)]
f 4
(12.178)

(sinh(t f I TB)) 1 (sinh(2t/ /TB) ) .

ln (0/ 1,,,x tf) + ln ln ri i
4 tf /TB 4 htf / TB

It remains to calculate the frequency-independent angular integral in eqn

(12.171), that is we have to evaluate integrals of the form
U nU rn
/(nn, urn) :=_--
_ f ciSICk)co2 (12.179)
(kun)(kurn)

= f dn(k) ,1 — V • il
n M,,
(1 — k • fln)(1 - k • f1m )'
(12.180)

where n, m = 2,4. We note that the combination c/1(k)w 2 is an invariant quan-

tity, such that gu n , um ) is a Lorentz-invariant integral. To determine this in-
tegral we may therefore transform to a coordinate system in which the second
velocity is equal to zero, that is Vni = 0. In this system the magnitude vn = izinl
of the first velocity is equal to the relative velocity

(12.181)

which is, by definition, a relativistic invariant quantity. Thus we arrive at

47r 1 m v„ m .
/(Un , U rn ) = f c8-2(k) 1„ = — tar (12.182)
1 - k • fl'n VTLM

This formula is correct also for the case u n = um , giving /(un , ?in ) = 47r, as may
be seen directly from the expansion of tanh -1 (s) for small arguments,

1, 1
tanh -1 x = x + -x- + - x 5 ±.. (12.183)
3 5
Thus we find,
602 OPEN QUANTUM ELECTRODYNAMICS

2
f A2(k)W 2 LU2 7,1.14 = gU2, 2/2) ± /(2/4 2/4) 2 I(U2) U4)
riAt2 rtoU4
( 1

=-87r — tanh -1 v24 - 1) . (12.184)

v24

Substituting (12.184) and (12.178) into eqn (12.171) we finally obtain

r[c] = rvac rth, (12.185)

where
6c 1
1' vac ln (02,, ax t f) tanh -1 1)24 - 1) (12.186)
V24

is the vacuum decoherence functional and

8a [in (sinh(tf/TB)) 1 ln (sinh(2tf /TB) )1
rth
tf/TB 7r{ 4 2t f /TB
1
tanh 1 v24 - 1)
( -

(12.187)
V24

is the thermal contribution to the decoherence functional. As expected, we see

from these expressions that F[c] strongly depends on the relative velocity V24
which is due to the fact that the decoherence is caused by the emission of
bremsstrahlung. The larger 1,24 the larger is the involved acceleration of the
charged particle which creates the radiation field.
An important result is that bremsstrahlung leads to a partial destruction of
coherence even at zero temperature. The magnitude of the vacuum contribution
rva, is seen to increase as the logarithm of the time tf if the relative velocity is
held fixed. This weak dependence is connected to the effective infrared resolu-
tion SImin Of of the interference device: For increasing tf photons of lower
and lower frequencies could in principle be detected and, thus, more and more
information is lost on tracing over the photon field. On the other hand, the term
within the square brackets in (12.187) approaches tf/2T B for tf » TB. Thus,
keeping fixed the relative velocity the magnitude of the thermal contribution
rth increases linearly with t1 for times tf >> TB. This describes the decohering
influence of absorption and emission processes induced by the thermal field. It
follows that for short times the vacuum contribution dominates, while decoher-
ence is mainly due to thermally induced processes for large times. The time ef
corresponding to the cross-over between these two regimes is determined by the
relation
2 t*f
ln (0/maxt*) = (12.188)
f 3 TB .

Taking S./ max - 10 19 s -1 and T = 1 K we find from this condition that the
cross-over time is of order
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 603

t f* ,--, 30 TB ^-, 10 -10 s. (12.189)

This means that for the given example the vacuum decoherence dominates for
times small compared to 10' s.
To facilitate the further discussion, let us investigate the case of opposite
velocities with equal magnitude, that is VI = gl = - V2 = -V3 (see Fig. 12.3(a)).
The relative velocity is then found to be

2v
V2 4= (12.190)
1 + v2 '
where

v=
la4 --5,31
(12.191)
2tf '

This situation corresponds to the case of two wave packets in a superposition

which first move apart with opposite mean velocities fA and 6 = —VI, respec-
tively, and, having reached their maximal distance icli — c13I, approach each other
again with velocities t72 and fl4 = —V2 . For non-relativistic velocities we have
v24 ,=.-., 2v and we may use the expansion (12.183) to obtain
16ct tf 2
rth -- - V , t f » TB • (12.192)
TB 37r

One can then ask the following question: Given a fixed electron energy, that is
a fixed velocity v, how far can we coherently separate the components of the
electronic state without exceeding a given threshold i Fui for the decoherence?
Provided the thermal contribution dominates, eqn (12.192) leads to the condition

16cv tf 2
37 V= iro i, (12.193)
TB

from which we obtain the maximal possible separation

37 i rol crB .
ma. = 2vt1 = d (12.194)
8a v/c
Choosing iro i = 0.01, which corresponds to a threshold of 1% decoherence, we
find that the maximal distance at T = 300 K is given by

4 ium
dmax ,-.)
-- (12.195)
v /c '
This shows that one can achieve rather large coherent separations for non-
relativistic electrons. For example, in the experiment performed by Hasselbach,
Kiesel and Sonnentag (2000) an electronic beam was coherently separated by
a lateral distance of about d = 100 sum. To compare this experiment with our
604 OPEN QUANTUM ELECTRODYNAMICS

results we take an electron energy of 1 keV and use a fixed value of 10 cm for
the distance corresponding to the time interval from t o to tf. For 1F0 1 = 0.01
condition (12.193) then yields dmax -,-:,' 4.5 cm at T = 1K and dmax P..)-' 0.26 cm at
T = 300 K. Note that the quantity y in (12.193) represents the magnitude of the
lateral component of the electron velocity in the experiment, which is due to the
fact that it is the relative velocity that enters the formula for the decoherence
functional. The values obtained for dmax are large compared with the lateral
distance d of the experiment, demonstrating that our theory is in full agreement
with experiment.
The result expressed through eqns (12.186) and (12.187) can also be dis-
cussed from another point of view. Namely, instead of keeping fixed the velocity
y (eqn (12.191)), we consider a fixed maximal spatial distance A', 1 — a3 1 between
the paths. Thus, for increasing tf the velocity y becomes smaller and smaller and,
consequently, the decoherence effect through bremsstrahlung becomes smaller
and smaller. For large enough times y is non-relativistic such that the vacuum
and the thermal contribution to the decoherence functional are given by

2a
r vac – — ln (01 maxtf) lai – d31 2 (12.196)
7r t2f '

and
8cv [ (sinh(tf /7-B)
) 1 ln (sinh(2tf
) /TB )] lai — a312
-,-,: –
ii 37
— ln rth 2 . (12.197)
tf /7-B 4 2tf /7-B tf

According to eqn (12.196) the magnitude of r,, decreases essentially as t f-2 ,

while eqn (12.197) shows that the magnitude of the thermal contribution rth
decreases as t f-1 for tf » TB,

4a lai — 5,31 2
rth Re, q . (12.198)
07r t f TB
We again observe the cross-over between two regimes of times: For short times
the vacuum decoherence dominates, whereas the thermally induced decoherence
dominates for large times. This can be seen in Fig. 12.4 where we have plotted
the expressions (12.186) and (12.187) as a function of tf for a fixed value of
11 — cI31.
The expressions (12.196) and (12.198) suggest we define a vacuum and a
thermal coherence length by means of

ld1 – 7/31 2
rvac (12.199)
2 L(ti ) v ac 1
571 — 163 1 2
rth = (12.200)
2L ( t f ) t2h •

This leads to (reintroducing factors of c)

 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 605

x 10-4
o

-0.2 -

-0.4-

-0.6 - -

, , -

4:0 60 80 100 120 140

ti/TB

FIG. 12.4. The vacuum contribution r vac and the thermal contribution rth
-

of the decoherence functional F according to eqns (12.186) and (12.187).

For a fixed maximal distance 151 — c7,3 1 = c7-.8 between the paths the two
contributions are plotted against the time tf which is measured in units of
the thermal correlation time TB. Parameters: T = 1 K and ftmax = 10 19 5 -1 .

7r 10.4
L(tf)„a, = 4a ln (02maxtf)
c tf r-:....', c • tf, (12.201)
On (Anaxtf)

and

t11/2 • T -1 /2 .
v
L(t f) th ,----- 37 C2 TBt f `,:.`.,' 12.7\I C2 TBt f CX (12.202)
8ct

Equation (12.201) implies that the vacuum coherence length is roughly of order

L(t i ) vac — c • t f. (12.203)

This simple result means that for a given time t1 the radiation field does not
destroy quantum coherence on length scales which are small compared to the
distance that light travels during this time. This also explains why the radiation
field is quite ineffective in destroying quantum coherence of single, localized
electrons.
We close this section by considering briefly another interference device which
is depicted in Fig. 12.5: Here we suppose that the wave packets brought to
606 OPEN QUANTUM ELECTRODYNAMICS

a2

U2 I U4

211

FIG. 12.5. A current loop involving two line segments which meet in the infinite
past (compare with Fig. 12.3(a)). This situation corresponds to an interfer-
ence device in which the relative velocity v13 between the interfering wave
packets vanishes initially.

interference are at rest initially. As indicated in the figure this case corresponds
to a situation in which the line segments described by the 4-velocities u 1 and
u3 meet each other in the infinite past. Thus we set u 1 = u3 in eqn (12.163) to
obtain,

ie{ u1 [ e —ikai _ e—ika

c(k) = e_'"3] (12.204)
.N/2 kul
U2 [ e —ika2 _ e —ikail _ ______
U4 [ —ika2 _ e3 ]}
e
ku2 ' ku4
Performing here the approximation (12.169) we are again led to expression
(12.166), where now

0012 = 2(1 - cos wtf), (12.205)

which immediately yields the expression

r[ci 2a [ in olmaxtf ln (sin

h (tf /TB) )1 ( 1
— tanh-1 v24 - 1) . (12.206)
j 7r [ tf/TB V24

In the non-relativistic limit we use the expansion (12.183) to obtain

sinh(t /TB) )] 2
F[c] - [ln (02maxtf) + ln v , (12.207)
37r t f / TB

where we have again assumed '62 = —774.

Decoherence through bremsstrahlung exhibits a highly non-Markovian char-
acter since the decoherence functional depends on the whole paths of the in-
terference device. This can be illustrated by a comparison of eqn (12.206) with
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 607

eqns (12.186) and (12.187): After the time corresponding to the maximal distance
between the wave packets we have in both cases two wave packets approaching
each other with opposite velocities of the same magnitude v. The difference
between the decoherence functionals obtained in the two cases shows that the
suppression of quantum coherence through bremsstrahlung depends on the total
history of the process and that the memory time is of the order of the total time
t1 of the experiment.

12.3.4 Path integral approach

An exact analytical representation for the problem of a single electron interact-
ing with the radiation field which takes into account the finite width as well
as the spreading of the wave packets can be obtained in the (non-relativistic)
dipole approximation (Barone and Caldeira, 1991; Diirr and Spohn, 2000). This
simple case is of special interest for it allows an analytical treatment for several
interesting cases which already exhibit the basic physical mechanism leading to
the decoherence mechanism through bremsstrahlung (Breuer and Petruccione,
2000).
Invoking the dipole approximation we may replace the commutator function
(eqns (12.19) and 12.20)) and the anticommutator function (eqns (12.37) and
(12.38)) with their space-independent expressions as follows,
00

DT (x — x') fSii D(t t') t5j sin c.4.; (t — t'), (12.208)

co
— x' ) ii SiiD i (t — t') öj f cica(w) coth (w/2kBT) cos w(t — t'),
(12.209)

where the spectral density is given by

e2
J(w) 3 7 .6 00/max Ci) ). (12.210)

As before we have introduced here an ultraviolet cutoff nmax . One observes

that the spectral density increases with the first power of the frequency w. Had
we used the dipole form • ET for the coupling of the electron coordinate
to the electric field (which is obtained through a canonical transformation),
the corresponding spectral density would be proportional to the third power of
the frequency. This means that the coupling to the radiation field in the dipole
approximation may be described as a special case of the Caldeira—Leggett model
discussed in Section 3.6. In the language of the theory of quantum Brownian
motion the radiation field constitutes a super-Ohmic environment (Barone and
Caldeira, 1991; Anglin, Paz and Zurek, 1997). Note also that we now include the
factor e2 into the definition of the correlation function.
608 OPEN QUANTUM ELECTRODYNAMICS

Within the non-relativistic approximation we may replace the current density

by

j(t, 12 6(1 Z(t)) + (5(g — I(t)) 1-3.(t) , (12.211)

ma ?no
where j(t) and I(t) denote the momentum and position operator of the electron
in the interaction picture with respect to the Hamiltonian
772
Hm = ± V (4 (12.212)
2mo
for the electron, V(i) being an external potential. The bare electron mass is
denoted by mo (see below).
We are thus led to the following non-relativistic approximation (neglecting
the spin degree of freedom) for the influence functional representation of the
single electron density matrix (compare with eqn (12.99)),

p (t) p (t
Prn.(t f) = T, ( exp I dt 1 clti {—i D(t — t i ) -'e -'a ') (12.213)
2 mo mo
ti ti
1
-- D i (t — t i ) 15'e(t)15c(e) } ]) prn (ti ).
2 mo mo
In accordance with the definitions (12.95) and (12.96) fie is a commutator super-
operator and 75,, an anticommutator super-operator.

12.3.4.1 Path integral representation and classical equations of motion As we

know from Section 3.6.4.2 the reduced density matrix given in eqn (12.213)
admits an equivalent path integral representation. We introduce new coordinates
through the relations d' and = (i ± it ), and set, for simplicity, the
initial time equal to zero, t i = 0. The propagator function J is then defined by

prri (ilf,d'f,tf)= f d3 ri f d 3 qiJ(rf,q'f,tf;ri,)p,,,(4,6,0), (12.214)

and admits a path integral representation of the form (see eqn (3.527))

J(Ff, 67f, tf;77;, = fDrf Dd'expfiAV, (12.215)

This is a double path integral taken over all paths fqt), d'(t) which satisfy the
boundary conditions,

r.(0) = i , f(t 1 ) = 4 (0 ) = (At f ) = qf . (12.216)

The weight factor for the paths is defined in terms of an effective action functional
A given by (see eqn (3.531))
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 609

tf

= f dt (n-1 0 ;14. — (71 + (772) + V(71 — (772))

0
tf

f f
tf

dt 9(t t')D t(t)(t')

o 0
tf tf

+ - f dt f D i (t — )(t)(t i ). (12.217)
4
o o
It is instructive to investigate first the classical equations of motion deter-
mined by the effective action functional. The first variation of A with respect to
f(t) and (At) leads to the classical equations of motion,

1 d
m)t) + — Vf:(V(e+ (772) + V(71 - (772)) + f D(t — )7(t i ) —

2 dt

0
tf

2
d f dt' D — t i )(tf ), (12.218)
o

and
tf

mo(t) + 2t 4.(v(F+ (772)+ v(F-- (772)) + f .13(e — t)Ot i ) = 0.

(12.219)

The real part of the equation of motion (12.218), which is obtained by setting
the right-hand side equal to zero, yields the famous Abraham-Lorentz equation
for the electron (Jackson, 1999). It describes the radiation damping through the
damping kernel D (t t'). To see this we write the real part of eqn (12.218) as
-

moi(t) + — t
dt
d
f D( '. (
)77 ti ) = P (t), (12.220)
o

where P(t) denotes the external force derived from the potential V. The damping
kernel can be written as
Sl ma ,

D(t — t') = f dco

0
e2

37r
2W sinw(t — t')
372
e2 d

dti f ti) '

(12.221)

where we have introduced the function

610 OPEN QUANTUM ELECTRODYNAMICS

sin Qmax t
f (t) t (12.222)

To be specific the UV-cutoff ftmax is taken to be Wmax = mc2 , which yields

Qmax 0.78 x 1021 s -1 . The term of the equation of motion (12.220) involving
the damping kernel can be written as follows,
t t
2 d d •
a
d f de D (t - t i);1(t` ) =-e—= f de [— f (t - t' ) ] 7- ( t')
0
3R-2 dt dt'
0
(12.223)

t
2 d •• • •
= -e- - f de f (t - t i Mt') + f (0)r(t) - f (t)f'(0) .
372 dt
o
For times t such that Qmax t » 1, i.e. t » 10' s, we may replace f (t) by 7(5(0 ,
and approximate f (t) 0, while f (0) = Q max. Thus we obtain,

d
f D(t t')F.(ti ) = 3e722 ddt [ (t) + Qinax f•'.(0] (12.224)
o

which finally leads to the equation of motion,

e2
e 2 f/max )•
( mo + 372 /At) - — v(t) = P(t), (12.225)
Gir

where ti = ;7.' is the velocity. This equation is known as Abraham-Lorentz equa-

tion (Jackson, 1999). The above derivation, which is similar to the one given
by Barone and Caldeira (1991), shows that the damping kernel leads to two
contributions. The first one provides a dressing of the electron mass by an elec-
tromagnetic mass contribution Sm = e 21 max /37 2 . This renormalization of the
electron mass is the same as the one found in Section 12.2.3 (see eqn (12.120)).
The second contribution introduced by the damping kernel is proportional to
the third derivative of 77'(t) and describes the damping of the electron motion
through the emitted radiation. This term is independent of the cutoff.
The equation of motion (12.225) can be obtained heuristically by means of the
Larmor formula for the power radiated by an accelerated charge. More rigorously,
it has been derived by Abraham and by Lorentz from the conservation law for
the field momentum, assuming a spherically symmetric charge distribution and
that the momentum is of purely electromagnetic origin (Jackson, 1999). The
decomposition in = mo + Sm of the electron mass into a bare mass mo and an
electromagnetic mass contribution (5m is unphysical since the electron is never
observed without its self-field and the associated field momentum. In other words,
we identify m with the observed physical mass which enables us to write eqn
(12.225) as
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 611

m (il(t) — 11,'(t)) =
To - (12.226)

where we have defined the characteristic time constant

e2
TO = —2 r e
= 67rm 0.6 x 10 -23 s. (12.227)
3
As is well known, eqn (12.226), being a classical equation of motion for the
electron, leads to the problem of exponentially increasing runaway solutions and
to an apparently acausal classical behaviour known as pre-acceleration (Jackson,
1999). The time 7-0 represents a characteristic radiation time scale of the electron
motion in the following sense (see also the corresponding discussion in Section
12.3.1). Suppose the electron is at rest initially. If the external force f(t) acts
upon the electron for a short period of time T the condition T >> TO implies
that the kinetic energy of the electron is large in comparison to the radiated
energy according to Larmor's formula. This implies that the influence of radiative
damping may be neglected provided P(t) changes only slightly over times of the
order of To . As in Section 12.3.1 we shall use this condition in the following.
It allows us to discard the term involving the damping kernel altogether in the
classical equations of motion. In the language of Brownian motion we could say
that the short-time behaviour is that of a strongly underdamped particle and
that decoherence is entirely due to the noise kernel in the electron propagator
function.
It might be useful to formulate this condition also for an electron which moves
in a harmonic potential
1 2 _i2
V(x) = -mc.,./o x . (12.228)
2
In this case the classical equation of motion reads

d2 613
P+ 471- To 77 ' = 0. (12.229)
dt3
With the help of the ansatz ?At) = 6 exp(zt) we are led to a cubic equation for
the characteristic frequencies,

Z2 1-, Wo2 — ToZ 3 =

0. (12.230)

For vanishing coupling to the radiation field (TO = 0) the solutions are located at
z+ = ±iwo , describing the free motion of a harmonic oscillator with frequency
(43 . For TO > 0 eqn (12.230) has three roots, one is real and the other two are
complex conjugated to each other. The real root corresponds to the runaway
solution and must be discarded. Let us assume that the period of the oscillator
is large compared to the radiation time,
1
To < —• (12.231)
Lao
612 OPEN QUANTUM ELECTRODYNAMICS

Because of (12.227) this assumption is well satisfied even in the regime of optical
frequencies. We may thus determine the complex roots to lowest order in w0 g-0 ,
1 2
Z+ = ±iWo — Toc06. (12.232)
The purely imaginary roots ±ic.00 of the undisturbed harmonic oscillator are thus
shifted into the negative half-plane under the influence of the radiation field. The
negative real part describes the radiative damping. In fact, we see that r(t) decays
as exp(--yt/2), where

2 2 traiô
'Y = Tocuo = — a 2 (12.233)
3 mc
is the damping constant for the radiative damping of the oscillator. If we consider
time intervals T of the order of magnitude of one oscillator period, coo- - 1, we
have -yr = (wo ro )(wo r) wo ro < 1. Thus we see again that we may neglect
effects of radiative damping provided the radiation time TO is small compared to
the typical time scale wo-1 of the undisturbed mechanical motion.

12.3.4.2 Determination of the propagator function We can now determine the

propagator function for the electron density matrix explicitly for an arbitrary
quadratic potential. The procedure is essentially the same as the one used in
Section 3.6.4.2 for the Caldeira-Leggett model. We therefore sketch only briefly
the result for a free electron moving in the radiation field. The details of the
calculation may be found in (Breuer and Petruccione, 2000).
Under the above conditions the propagator function for the electron is found
to be
3 f.
m
ArflTf)tfif;14 .i) =
27rtf
exp
tf
V.( — (Tf+ — r(Tf,Ct
f)} .
(12.234)
As expected, depending only on the difference f'f - i the propagator function is
,

invariant under space translations. Furthermore, one easily recognizes that the
contribution
3
m,
Go (17:f 6,tf) ( exP {— lrf 17.i)(6t f (12.235)
27rtf tf
is simply the propagator function for the electron's density matrix in the case of
a vanishing coupling to the radiation field.
The function T(Tir, j tf) in eqn (12.234) describes the influence of the radi-
,

ation field and may be written as

[ln 0Imax tf + ln
(sinh(tf/TB ))1 (qf Ti ) 2
r(Tf,Ti,tf)
t f I TB )J (ct f) 2
- 6) 2
(12.236)
2L(tf ) 2
 DECOHERENCE BY EMISSION OF BREMSSTRAHLUNG 613

Here, we have introduced the quantity L(t1) which is defined by

37r sinh(tf/TB))]
L(t f) 2 — [ln afl t f + ln ( —1 • (ct1) 2 , (12.237)
4a max tf /TB

and which can be interpreted as a time-dependent coherence length.

12.3.4.3 Wave packet dynamics With the help of eqn (12.234) we can study
the time-evolution of electronic wave packets in order to estimate the influence
of the finite width and of the spreading on the decohrence mechanism. To this
end, we investigate an initial state given by the superposition of two Gaussian
wave packets separated by a distance 2a. We assume that both packets have a
width ao and that they are centred initially at Y = = ±(a, 0, 0). The packets
are supposed to approach each other with the average speed y = ko /m > 0. For
simplicity the motion is assumed to occur along the x-axis. Thus we have the
initial state
3/4
5o (x) = 111 ( 1 exp [ 6)2t iko (Y E )]
27ra 6 4a6
1 )3/4 [ dy2
(
± A2
271-0-6
exp
4a-02
iko(+ et)] , (12.238)

where ic'o = (ko , 0,0) and A 1 , A2 are complex amplitudes. Our aim is to determine
the interference pattern that arises in the moment tf = a/v of the collision of
the centres of the packets at Y = 0. The position space density at the time tf is
found with the help of the formula

Prn(f.f,tf) E Prn(rf,q.f = 01t1) (12.239)

3 2
= f d3 ri f d3 qi m
27rtf
exp (r1 77'04; qt
2L(t f) 2
11
0(i + — aii4-,;( 17/ -
2
Performing the Gaussian integrations we get the result
3/2
1 \
Prn(Ff ) t f) = (12.240)
exp 1 2 )2 j
[ 2 :(ftf
(
27 (3- (t f ) 2 )
x 11A 1 1 2 + 021 2 + (A i A; exp [icp(Ff) + r(t f )i)}.

We recognize a Gaussian envelope centred at rf = 0 with width a- (tf). an inco-

herent sum lAi 1 2 ± 021 2 , and an interference term proportional to _4 1 _4. The
interference term involves the phase

cpVf ) = —2k.oFf (1 — (12.241)

614 OPEN QUANTUM ELECTRODYNAMICS

as well as the decoherence function

(12.242)

The quantity E is given by

-1
( Li (b)2 n.12 0.4 (t f )2 )
E=1+ j , + (12.243)
44 t !f

The term —24rf of the phase cp(f'f) describes the interference pattern as it
would be obtained for a free Schrödinger particle, while the term 2k'orf E leads to
a modification of the period of the pattern. Moreover, without the contribution
proportional to E the decoherence function (12.242) exactly coincides with the
expression (12.207). Thus we see that the influence of the finite width and of the
spreading of the wave packets brought to interference may indeed be neglected
provided the condition E < 1 holds. This condition is always satisfied in the
present approximation. Since E attains the maximum value
1
Em ax = (12.244)
1 + mL(t f) 2 It f

one is led to the requirement L(t f) 2 » ht f / rn, which is always fulfilled for times
tf and temperatures T satisfying ctf » A c and kBT < mc2 .

12.4 Decoherence of many particle states

-

In the previous section we have obtained several expressions for the decoherence
functional which describes the loss of coherence in an interference device. We
address here the question as to whether these results could explain the absence
of the coherence of certain superpositions of states that can be considered as
macroscopically distinct.
For single electrons the vacuum decoherence through bremsstrahlung turns
out to be small at non-relativistic speeds. For example, taking (A max — c/Ac and
t1 of the order of 1 s, and using a velocity y which is already as large as 1/10
of the speed of light, one finds that iFyac i — 10 -2 , which corresponds to a 1%
suppression of interference. By virtue of the weak logarithmic dependence on the
cutoff scale this estimate is true also for other particles carrying an elementary
charge.
Matters could however be different for many-particle states. To investigate
this case it is important to specify clearly the structure of the superposition
under consideration. We distinguish two extreme classes of many-particle states
(Joos, 1996). If (,) and bp2) are two states containing only a small number of
particles, one may consider a superposition of the form

i tF = (cd(P1) + 3 1S02» N . (12.245)

 DECOHERENCE OF MANY-PARTICLE STATES 615

As our notation indicates, the state IT) is an N-fold tensor product of a superpo-
sition of the few-particle states lyt ) and Iç 22 ). Therefore, the decoherence factor
for IT) is expected to be very small.
The situation changes, however, substantially if one considers another class
of states, namely those of the form

IT) = ctl(Pi N + 01(P2r. (12.246)

In contrast to (12.245) this is a superposition of two N-fold tensor products. It

is this class of states that will be investigated in the following regarding their
decoherence properties. Our aim is to derive the dependence of the decoherence
functional on the particle number N
Let us consider a system which is composed of N identical particles with
mass m and charp e. Our aim is to construct an effective master equation for
the density pcm (R, fi') of the centre-of-mass coordinate

1
(12.247)
i=1

for such a system. Here, the x are the particle coordinates and we suppress, for
simplicity, the spin degree of freedom. We introduce the relative coordinates ei
through

fi + 4 (0. (12.248)

Since the sum up to zero, they are functions of 3N — 3 internal variables which
we denote collective by q. Let us suppose that the state of the N-particle system
is described in the position representation by a density matrix of the form

Pm = Pcm(fi, fi')Pint(q, q`), (12.249)

where pcm and pint are separately normalized to 1,

I.
d3 R pcni (14, fi) = f dq pi nt (q,q) = 1. (12.250)

The density matrix pcm describes the centre-of-mass coordinate, while pint rep-
resents the state of the internal degrees of freedom. For example, one finds that
the quantity

1
W - R) = f dq E fi 71(q)) pint (q,q) (12.251)
j=1

is the density of finding a particle at Y under the condition that the centre-of-
mass coordinate is at R. This function is obviously normalized as
616 OPEN QUANTUM ELECTRODYNAMICS

d3 x w(i) = 1.
f (12.252)

If the system described by the state (12.249) performs a translational motion

it is reasonable to assume that its total current density can be approximated by
an effective current density of the form

(x) = —
Ne {/5w(I fi) + fi)P} , (12.253)
2M
where 1 = -ialafi is the total momentum canonically conjugated to the centre-
of-mass coordinate fi, Ne is the total charge and M = Nm the total mass. The
expression (12.253) implies that the current density of the internal degrees of
freedom vanishes. In particular, it excludes the possibility that the whole systems
is in a rotational state which would require us to introduce three further collective
coordinates as, for example, the three Euler angles.
Equation (12.253) shows that the case of an N-particle system can be dealt
with by using the replacements e -4 Ne and m M = Nm, and by interpret-
ing the length scale cro , which appears in the UV cutoff scale ftmax 1/a-0, as the
linear extension of the one-particle density w(Z). A representation for the density
matrix pcm (fi, 11`) is then obtained from (12.89) by substituting the effective cur-
rent (12.253) into the functional (12.99). Invoking the non-relativistic (dipole)
approximation one is led to the following representation for the centre-of-mass
density,
tf tf
(t) P(ti)
Pcm (tf ) = T, exp
( f dt f de {—2i D(t — t') PM M
(12.254)
ti ti
1 fic(t) Pe (t' )
-- D1(t — ti) Pcm (ti),
2 MM

where 15,(t) and 1(t) denote the interaction picture commutator and anticom
mutator super-operators for the total momentum. The dissipation and the noise
kernel are given by the expressions (12.213), where the spectral density now takes
the form
N 2e2
J(w) 32 ci.; 0(Q max — (12.255)
The results of Section 12.3.4 can now immediately be transferred to the
present case with the help of the replacements given above. It follows that the
vacuum decoherence functional for an N-particle state scales with the square N2
of the particle number,

rvac —N 2 -
8Q ln (gnmaxtf) V 2 . (12.256)
37r
This scaling with the particle number obviously leads to a strong amplification of
the decoherence effect. To give a very extreme example we take N = 1022 which
 REFERENCES 617

corresponds to u0 — 1 cm for typical free electron densities in metals. Let us ask

for the maximal speed y leading to 1% decoherence. With the help of (12.256)
we find y 10 -14 m s -1 . For a distance of 1 m this implies, for example, that a
successful interference experiment would take about three million years!
The scaling (12.256) of F with the square N 2 of the particle number can be
traced back to two facts. First, the radiative back-action is proportional to the
square of the total charge since the emitted radiation adds coherently in the
limit of long wavelength. Second, the decoherence functional only depends on
the logarithm of the cutoff 1max, which means that it depends only very weakly
on the total mass or on the spatial extent of the N-particle state. In the cases
discussed here one must expect, of course, a large radiative damping in addition
to the decoherence effect. It is of great fundamental interest to investigate also
the influence of these phenomena for composite neutral objects.

References
Anglin, J. R., Paz, J. P. and Zurek, W. H. (1997). Deconstructing decoherence.
Phys. Rev., A55, 4041-4053.
Barone, P. M. V. B. and Caldeira, A. 0. (1991). Quantum mechanics of radia-
tion damping. Phys. Rev., A43, 57-63.
Bjorken, J. D. and Drell, S. D. (1964). Relativistic Quantum Mechanics.
McGraw-Hill, New York.
Breuer, H. P. and Petruccione, F. (2000) Radiation damping and decoher-
ence in quantum electrodynamics. In Relativistic Quantum Measurement and
Decoherence (eds. Breuer, H. P. and Petruccione, F.), Volume 559 of Lecture
Notes in Physics, pp. 31-65. Springer-Verlag, Berlin.
Breuer, H. P. and Petruccione, F. (2001). Destruction of quantum coherence
through emission of bremsstrahlung. Phys. Rev., A63, 032102-1(18).
Chou, K.-c., Su, Z.-b., Hao, B.-l. and Yu, L. (1985). Equilibrium and nonequi-
librium formalisms made unified. Phys. Rep., 118, 1-131.
Cohen-Tannoudji, C., Dupont-Roc, J. and Grynberg, G. (1998). Atom-Photon
Interactions. John Wiley, New York.
Dièsi, L. (1990). Landau's density matrix in quantum electrodynamics. Found.
Phys., 20, 63-70.
Diirr, D. and Spohn, H. (2000). Decoherence through coupling to the radiation
field. In Decoherence: Theoretical, Experimental, and Conceptual Problems
(eds. Blanchard, P., Giulini, D., Joos, E., Kiefer, C. and Stamatescu. I.-0.).
Volume 538 of Lecture Notes in Physics, pp. 77-86. Springer-Verlag, Berlin.
Feynman, R. P. and Vernon, F. L. (1963). The theory of a general quantum
system interacting with a linear dissipative system. Ann. Phys. (N. Y.). 24.
118-173.
Feynman, R. P. and Hibbs, A. R. (1965). Quantum Mechanics and Path Inte-
grals. McGraw-Hill, New York.
Ford, L. H. (1993). Electromagnetic vacuum fluctuations and electron coher-
ence. Phys. Rev., D47, 5571-5580.
618 OPEN QUANTUM ELECTRODYNAMICS

Gradshteyn, I. S. and Ryzhik, I. M. (1980). Table of Integrals, Series, and

Products. Academic Press, New York.
Hasselbach, F., Kiesel, H. and Sonnentag, P. (2000). Exploration of the funda-
mentals of quantum mechanics by charged particle interferometry. In Decoher-
ence: Theoretical, Experimental, and Conceptual Problems (eds. Blanchard, P.,
Giulini, D., Joos, E., Kiefer, C. and Stamatescu, 1.-0.), Volume 538 of Lecture
Notes in Physics, pp. 201-212. Springer-Verlag, Berlin.
Itzykson, C. and Zuber, J.-B. (1980). Quantum Field Theory. McGraw-Hill,
New York.
Jackson, J. D. (1999). Classical Electrodynamics (third edition). John Wiley,
New York.
Jauch, J. M. and Rohrlich, F. (1980). The Theory of Photons and Electrons.
Springer-Verlag, New York.
Joos, E. (1996). Decoherence through interaction with the environment. In
Decoherence and the Appearence of a Classical World in Quantum Theory
(eds. Giulini, D., Joos, E., Kiefer, C., Kupsch, J., Stamatescu, 1.-0. and Zeh,
H. D.), pp. 35-136. Springer-Verlag, Berlin.
Rohrlich, F. (1965). Classical Charged Particles. Addison-Wesley, Reading,
Massachusetts.
Weinberg, S. (1995). The Quantum Theory of Fields, Volume I. Cambridge
University Press, Cambridge.
Wheeler, J. A. and Feynman, R. P. (1945). Interaction with the absorber as
the mechanism of radiation. Rev. Mod. Phys., 17, 157-181.
Wheeler, J. A. and Feynman, R. P. (1949). Classical electrodynamics in terms
of direct interparticle action. Rev. Mod. Phys., 21, 425-433.
 INDEX

if-algebra, 3 state verification measurements, 540

Chapman-Kolmogorov equation, 16, 117
Abraham-Lorentz equation, 587, 610 differential, 17
acceleration time, 590 characteristic function, 10
additive process, 46 Lévy process, 46
adiabatic elimination of fast variables, 208 observable, 72
adiabatic theorem, 256 Wigner, 74, 228
Alice, 558 Chebyshev inequality, 10
annihilation operator, 392 classical communication channel, 559
anticommutator function, 572 closed system, 110
approximation coherence length, 613
Born, 131, 136, 441, 445, 465, 468 coherent population trapping, 420
Born-Markov, 132, 136, 146, 174, 176 coherent state, 162, 242
Markov, 131, 136, 141, 470 coherent state representation, 162
mean-field, 203 coherent subspace, 224, 231
rotating wave, 132, 139, 144, 172, 586 commutator function, 570
singular coupling, 169 complete absorber theory, 585
weak-coupling, 131, 206 complete positivity, 88, 123
autocorrelation function, 189, 487 composite quantum system, 74
Compton wavelength, 590
back-action, 220, 322 concavity, 80
evasion condition, 104, 270, 275 conditional probability, 5
backward propagator, 446 conditional transition probability, 15
bare frequency, 476 PDP in Hilbert space, 306
basin of attraction, 426 short-time behaviour, 322, 323
bath, 116 continuity condition, 28
Bayes's theorem, 5, 88 convergence
Bell inequalities, 564 strong, 369
Bell states, 561 weak, 369
Bell-state operator, 548, 553, 560 convex linear combination, 68
Bethe formula, 588 convexity, 81
black-body radiation, 575 cooling temperature, 428
Bloch equation, 154 correlation coefficient, 10
Bloch vector, 148, 170, 288, 491 EPR correlations, 522
Bob, 558 correlation function
Bochner theorem, 136 bath, 174
Boltzmann distribution, 138, 161, 493 electric field, 574
Boltzmann equation, 201 equilibrium, 453
Borel set, 6 multi-time, 125, 316
Brownian motion, 30, 460, 607 stochastic representation, 319
free, 181, 186, 199 Poisson process, 26
quadratic potential, 185 reservoir, 134
Wiener process, 31
Caldeira-Leggett model, 172, 263, 607 Coulomb gauge, 391, 569
Cauchy process, 57 counter-term, 173, 184, 193
causality principle, 517, 539 counting probability, 313
local density matrix, 546 covariance matrix, 10
measurement of entangled eigenstates, 545 covariant state vector history, 514
measurement of projections, 544 CPI] time, 377, 379, 380
620 INDEX

scaling behaviour, 381 Doppler energy, 417

creation operator, 392 dressed states, 255, 436
cross-section, 240 dynamical map, 118
cumulant expansion, 451, 464, 580 dynamical phase, 256
cumulative distribution function, 7 dynamical semigroup, 18
current density, 23, 393, 576
current super-operator, 578, 582 effect, 86
cutoff frequency, 483, 487 Ehrenfest equation, 113, 181
eigenoperator, 132, 142, 147, 394
damping kernel, 184 einselection, 225
dark state, 404, 410 electron density, 393
de Broglie wavelength, 234, 241, 591 elementary events, 3
decay rate, 122, 472, 474 ensemble, 63
time-dependent, 463, 472 of type ep, 285, 321
decoherence, 166, 220 of type ep , 283
bremsstrahlung, 568, 588 entangled state, 78, 521
Brownian motion, 234 basis, 560
electromagnetic field states, 251 maximally, 78, 223, 273, 539, 552
internal transitions, 235 preparation, 554
many-particle states, 614 QND verification, 550
quantum measurement, 269 verification, 538
scattering, 237 entropy, 552
decoherence function, 222 flux, 129
Caldeira-Leggett model, 264 linear, 82, 223
damped harmonic oscillator, 249 production rate, 128
free Brownian motion, 265 quantum measurement, 92
short-time behaviour, 225 relative, 81, 128
strongly overdamped, 269 Shannon, 79
system-size dependence, 231 von Neumann, 79, 129, 203
thermal part, 229 environment, 115
vacuum optical, 243 EPR correlations, 521
vacuum part, 229 erasing of local information, 541, 544
decoherence functional, 588, 593 ergodic system, 453
decoherence rate, 233 error
Brownian motion, 234 estimate, 456
internal transitions, 237 relative, 456, 479, 483, 488
scattering, 241 estimator
decoherence time, 222, 233 expectation value, 362
defect, 35, 308, 364, 366, 406 statistical error, 363
density matrix, 66 Euler scheme, 370
as covariance matrix, 286, 297, 310 event, 3
equilibrium, 454 expectation value, 9, 297
irrelevant part, 441
QED, 576 Feynman propagator, 581
reduced, 76, 116, 460 field modes, 391
relevant part, 441 field operators, 391
detailed balance, 138 field vacuum, 392
deterministic drift, 30 as quantum probe, 325
deterministic process, 19, 304 Floquet states, 430
Dicke states, 214 Floquet theory, 429
diffusion matrix, 30 fluctuation-dissipation theorem, 190
diffusion process, 28 Fock states, 325
dissipation kernel, 174, 477 Fokker-Planck equation, 29, 164
dissipator, 123, 148, 169, 178, 202, 586 diffusion in Hilbert space, 314
Doob's theorem, 31 foliation, 506, 511
 INDEX 621

Gauss's law, 392 infinitely divisible, 47

Gauss's theorem, 504 influence functional
Gaussian process, 31 Feynman-Vernon, 192, 193, 475
Gaussian state, 450, 481, 580 path integral representation, 195
generating function, 11 QED, 581, 582
cumulant, 46 path integral representation, 608
Gibbs state, 129, 137, 443 super-operator representation, 193
golden rule, 23, 138, 146, 468 influence super-operator, 581
infrared catastrophe, 595
H-theorem, 203 initial condition
Hamiltonian factorizing, 446, 477
atom-laser interaction, 411 non-factorizing, 446, 452, 489
Caldeira-Leggett, 172, 475 initial jolts, 188, 487
density, 502 integrability condition, 505, 513
electromagnetic field, 392 interaction picture, 113
Jaynes-Cummings, 253 interference contrast, 247
non-Hermitian, 306 invariant manifolds, 416
QED, 576 irreversibility, 19
spin-boson, 225, 490 isolated system, 110
harmonic oscillator, 160, 242, 266, 373, 460, Itô calculus, 41
474, 491 diffusion process, 39
Hartree equation, 203 PDP, 42
Heisenberg equation of motion, 112 Itô formula, 41
Brownian motion, 182
Heisenberg picture Jaynes-Cummings model, 460
closed system, 112 on resonance, 470
open system, 124 photonic band gap, 474
stochastic representation, 316 with detuning, 471
Helmholtz equation, 391 joint probability distributions, 12, 70
Heun scheme, 370 consistent hierarchy, 521
high harmonic generation, 433 family of, 13
Hilbert space, 59 relativistically invariant, 518
Dirac measure, 291 jump process, 21
doubled, 316, 457
extended, 430 Klein inequality, 81
functional derivatives, 304 KMS condition, 137
functional integration, 292 Kolmogorov axioms, 4
projective, 68 Kolmogorov consistency conditions, 13
volume element, 292 Kolmogorov forward equation, 311
Hilbert space path integral, 311 Kolmogorov theorem, 14
direct photodetection, 332
photodetection on a field mode, 351 Lévy-Khinchine theorem, 47
Hilbert-Schmidt operator, 76 Levy measure, 48
homogeneous process, 16 stable, 53, 410, 423
hypersurface, 502, 508 Lévy process, 45
unit normal vector, 503 characteristic exponent, 46, 54
variation, 503 generator, 47
Lamb shift, 136, 145, 148, 151, 468, 474,
increments 477, 586
complex Wiener process, 165 time-dependent, 463, 472
field of Poisson, 43 Larmor formula, 610
independent, 46 laser cooling, 409
Poisson process, 27, 363, 415 momentum distributions, 426
stationary, 46 sub-recoil, 410
Wiener process, 32, 368 laser equations, 207
622 INDEX

Lieb theorem, 82, 128 Nelson theorem, 71

Lindblad operator, 122 noise
Lindblad theorem, 122, 473 additive, 368
Liouville equation, 21, 305 multiplicative, 368
Liouville master equation, 33 noise kernel, 174, 264, 477
PDP in Hilbert space, 306 non-Markovian dynamics, 445, 460
Liouville operator, 111, 577 normal ordering, 163
Liouville space, 76
Liouville—von Neumann equation, 111 observable, 60
Lorentz equations, 207 mean value, 65
not self-averaging, 381
macroscopic quantum coherence, 496 self-averaging, 381
Mandel Q parameter, 334 strongly self-averaging, 381
Markov condition, 14 observable frequency, 488
Markov process, 14 one-step process, 24, 161
quantum, 117 open system, 115
master equation, 22, 109 operation, 86, 168, 322, 453
adjoint, 125, 162, 481 direct photodetection, 325
Caldeira—Leggett, 173, 178, 199, 478 entangled probe state, 527
recoilless limit, 200 homodyne photodetection, 337
Floquet representation, 433 indirect measurement, 322
generalized, 470 laser cooling, 413
Hu—Paz—Zhang, 477 multipole radiation, 396
Lindblad, 122 QED form, 396
low-density, 140 relativistic, 510
Markovian, 119 representation theorem, 89
mean-field, 203 state verification measurements, 537
non-linear, 201 orbital angular momentum density, 393
Pauli, 138, 161 ordered cumulants, 452, 464
quantum optical, 145, 585 Ornstein—Uhlenbeck process, 31
Redfield equation, 132 complex, 165
singular-coupling, 139 overdamped, 157, 266, 488
time-convolutionless, 447, 470, 474, 585 strongly, 268
unravelling, 306
doubled Hilbert space, 457, 473 P-representation, 162
weak-coupling, 135 parallel processing, 388
Matsubara frequencies, 175 Pauli operator, 146
memory kernel, 444, 469 phase flow, 20
micro-reversibility, 202 phase space methods, 162
microcausality, 505 photocounting formula, 351
operation, 512 coherent state, 352
mixture flourescence radiation, 334
improper, 76 Fock state, 352
statistical, 68 QND measurement, 356
moments, 9 Schrödinger cat, 353
Monte Carlo method, 361 photodetection, 596
numerical efficiency, 380 direct, 324
Monte Carlo wave function method, 313 field mode, 350
Morse oscillator, 383 heterodyne, 342
multi-time exclusive probability, 313 homodyne, 335, 368
multipole moment photon antibunching, 335
electric, 397 physical frequency, 477, 478
magnetic, 397 piecewise deterministic process, 32
continuous measurement, 323
Nakajima—Zwanzig equation, 444 density matrix, 349, 400
 INDEX 623

direct photodetection, 329 Lindblad generator, 122, 137, 139, 141,

doubled Hilbert space, 457, 473 145, 152, 178, 463, 470, 473, 477
Floquet representation, 432 parametric Lindblad generator, 201
Hilbert space, 305 relaxation to equilibrium, 137
homodyne photodetection, 339 quantum efficiency, 351
Kolmogorov forward equation, 38 quantum jump, 308
laser cooling, 415 selection algorithms, 367
path integral representation, 37 quantum measurement
sample paths, 35 approximate, 93, 537
state vector, 399 Bell-state operator, 548
stochastic calculus, 39 continuous, 166, 321, 391, 523
Planck distribution, 152 dynamical model, 275
Platen scheme, 372 entropy, 92
Poincaré recurrence times, 135 exchange, 556
pointer basis, 270, 274 generalized, 85
Poisson process, 23, 49 ideal, 85, 166, 270
non-homogeneous, 25 incomplete, 398
polarization vectors, 411 indirect, 96, 168, 508
positive operator-valued measure, 87 local, 508
pre-acceleration, 611 non-ideal, 167
pre-measurement, 274 non-local, 526
preferred basis, 222 non-local observables, 530
preparation, 83 non-local, modular sums, 535
non-local states, 554 non-selective, 84, 168, 321, 533, 537
preparation time, 590 on ensembles of type ep, 301
QND, 102, 532, 544
probability current density, 29
QND state verification, 536
probability density functional
selective, 84, 301
Hilbert space, 291
state verification, 536, 539
space of density matrices, 299
quantum object, 96
probability distribution, 6
quantum probe, 96, 508
projective Hilbert space, 286, 295
entangled, 526
probability measure, 4
quantum regression theorem, 127, 158, 486
projection operator technique
quantum teleportation, 557
Nakajima-Zwanzig, 442
quantum trajectories, 313
time-convolutionless, 446, 460
quasi-energy, 430
projection postulate, 83, 502, 509, 532
qubit, 226, 231
relativistic, 511
projection-valued measure, 65
Rabi frequency, 154, 171, 254
propagator, 15
radiation
propagator function, 193, 608
dipole, 398
pure state, 68
multipole, 397
radiation damping, 609, 612
quantization of radiation field, 391 random variable, 5, 63
quantum Boltzmann equation, 202 Hilbert space, 286
quantum channel, 558, 563 multivariate, 8
quantum computer, 226 ray, 68
quantum dynamical semigroup, 119, 304, Rayleigh scattering, 160
473, 486 realization, 6, 12, 362
contracting, 123 recoil energy, 409, 416
first standard form, 121, 136, 139, 153, recoil temperature, 410
178 reduced system, 115
generator, 119 reference state, 443
Heisenberg picture, 122 relative states, 272
irreversibility, 128 renormalization transformation, 425
624 INDEX

renormalized frequency, 476, 477 statistical operator, 66

reservoir, 116, 130 Stefan-Boltzmann law, 575
reservoir correlation time, 467 Stern-Gerlach experiment, 288
response function, 189 stochastic collapse model, 345
runaway solutions, 611 stochastic differential equation, 39, 181
Runge-Kutta scheme, 371 density matrix, 348
diffusion process, 41
S-matrix, 238 PDP, 42
sample, 362 PDP in Hilbert space, 309
average, 362 stochastic process, 11
sample path, 12 continuity condition, 28
sample space, 3 generator, 18
scalability, 388 multivariate, 12
scale invariance, 52 non-Markovian, 527
scaling exponent, 54 relativistic, 513, 518
scattering amplitude, 239 stochastic simulation method, 361, 470, 474
Schmidt decomposition, 77, 538, 551 algorithm for PDPs, 364
Schmidt number, 78 algorithm for SSEs, 367
Schrödinger cat, 242, 251, 353 linear search, 367
Schrödinger equation, 110 numerical efficiency, 380
non-linear, 209, 210, 214, 307 rejection method, 367
stochastic, 315 strong coupling regime, 470
continuous spontaneous localization, 345 sub-ensemble, 68
heterodyne detection, 344 subadditivity, 80
homodyne detection, 342 strong, 82
quantum state diffusion, 347 super-operator, 111
Schwinger-Tomonaga equation, 503 super-radiance, 210
self-adjoint operator, 60 superselection rule, 225
single-step error, 369 superselection sector, 224
singular value decomposition, 78
spectral correlation tensor, 143, 150 T-matrix, 239
spectral decomposition, 61 TCL generator, 447, 461, 486, 491
spectral density, 174, 228, 264, 463 perturbation expansion, 447, 451
Jaynes-Cummings model, 467, 471 TCL inhomogeneity, 447, 486, 488, 495
Lorentz-Drude, 175, 191, 491 perturbation expansion, 453
Ohmic, 175, 185, 186, 191, 265, 476 teleportation rate, 560
photonic band gap, 474 tensor product, 75
QED, 607 thermal correlation time, 229, 573
super-Ohmic, 607 time inversion, 18
spectral family, 60 time ordering
spectral measure, 64 antichronological, 446
spectrum chronological, 444
equilibrium fluctuations, 189 time-evolution operator, 110
resonance fluorescence, 158, 437 transition rate, 21
self-adjoint operator, 61 transverse delta function, 391
speed-up, 388 trapping state, 34, 404
spin density, 393 trapping time, 422
spin-boson model, 226, 460, 490 tunnelling, 490
squeezed vacuum, 134, 149 two-level system, 169, 377, 461, 467
stable distribution, 52
stable process, 52 ultraviolet cutoff, 607
standard error of the mean, 363 ultraweak continuity, 123
Stark shift, 145, 148, 151, 586 underdamped, 157, 266, 488
stationary process, 16
statistical independence, 5 vacuum-to-vacuum amplitude, 583, 595
 INDEX 625

variance, 10, 362 PDP, 33

dark state population, 407 PDP in Hilbert space, 308
ensembles of type Ep, 287, 298, 300, 357 pure jump process, 34
observable, 65 wave equation, 391, 570
von Neumann equation, 111 Weisskopf—Wigner theory, 462
Wheeler—Feynman action, 585
waiting time distribution, 363 Wick theorem, 578
dark state, 401 Wiener process, 30
exponential, 365 multivariate, 32
fluorescence radiation, 331 Wigner distribution, 73, 245, 254
general, 366
laser cooling, 421 Zeno effect, 139, 166, 167 '
multi-exponential, 365 Zeno time, 169, 170