Buku - Linear Dynamical Quantum Systems - Analysis, Synthesis, and Control 2022

Communications and Control Engineering
Hendra I. Nurdin
Naoki Yamamoto
Linear
Dynamical
Quantum
Systems
Analysis, Synthesis, and Control
Series editors
Alberto Isidori, Roma, Italy
Jan H. van Schuppen, Amsterdam, The Netherlands
Eduardo D. Sontag, Piscataway, USA
Miroslav Krstic, La Jolla, USA
More information about this series at http://www.springer.com/series/61
Hendra I. Nurdin Naoki Yamamoto
•
Linear Dynamical Quantum

Systems
Analysis, Synthesis, and Control
123
Hendra I. Nurdin Naoki Yamamoto
School of Electrical Engineering Department of Applied Physics
and Telecommunications and Physico-Informatics
UNSW Australia Keio University
Sydney, NSW Yokohama
Australia Japan
ISSN 0178-5354 ISSN 2197-7119 (electronic)

ISBN 978-3-319-55199-9 ISBN 978-3-319-55201-9 (eBook)
DOI 10.1007/978-3-319-55201-9
Library of Congress Control Number: 2017934052
MATLAB® is a registered trademark of The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA
01760-2098, USA, http://www.mathworks.com.
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The main theme of this research monograph is modeling and control of a special
class of open quantum systems that shall be referred to as linear dynamical quantum
systems. They are also referred to in the literature as linear quantum stochastic
systems with the qualifier “stochastic” added because these systems are coupled to
and driven by quantum stochastic processes (i.e., “quantum noise”). For brevity, we
will often refer to them simply as linear quantum systems. Such systems are
frequently encountered in fields such as linear quantum optics and opto-mechanics,
to name a few, with applications in areas including continuous-variable quantum
communication and quantum precision sensing. They are essentially a quantum
analogue of the distinguished class of classical (non-quantum) linear systems that
have played a foundational role in the development of modern and post-modern
systems and control theory.
The reader may wonder, is there a place for a research monograph dedicated to
this class of quantum systems? Certainly, this monograph would not be the first text
to discuss linear dynamical quantum systems. Indeed, the now standard reference
text “Quantum measurement and Control” by Wiseman and Milburn [1] already has
sections devoted to linear dynamical quantum systems (see Sects. 5.6 and 6.6
therein). The present monograph complements existing textbooks that treat the
subject of quantum feedback control, as the text of Wiseman and Milburn and the
more recent “Quantum Measurement Theory and its Applications” by Jacobs [2], in
terms of its emphasis and approach. It focuses on the class of linear quantum
systems and provides an in-depth treatment of the system-theoretic foundations of
such systems, and the control theoretic formulation of quantum versions of familiar
problems from the classical (non-quantum) setting, including estimation and
filtering, realization theory, and linear quadratic control. Some of these topics may
be unfamiliar to researchers in quantum control coming from outside of the control
theory discipline. It is our hope that the monograph can at least partly bridge this
gap and serve as good introduction to these topics as a stepping stone towards
understanding more general nonlinear quantum stochastic systems.
The monograph is aimed at graduate students and researchers with a background
in control theory, physics, or mathematics who have an interest in the topic
v
vi Preface
of quantum feedback control theory, and linear quantum systems in particular.

It assumes that the reader has working knowledge of quantum mechanics, at least at
the undergraduate level, such as in D.J. Griffith’s text “Introduction to Quantum
Mechanics” (Pearson Education). It also assumes that the reader from a non-control
theory background has working knowledge of modern control theory, in particular
with the state-space representation of linear dynamical systems. Some exposure to
continuous-time stochastic filtering and control theory would also be valuable.
With the 2012 Nobel Prize in physics being awarded to experimental physicists
Serge Haroche and David Wineland for their groundbreaking contributions to
experimental quantum control techniques, and continuously improving experi-
mental capabilities in controlling quantum systems, it would not be an exaggeration
to say that it is currently exciting times for quantum control. Linear quantum
systems model a wide range of quantum systems that are taking part in the broader
ongoing quantum revolution, which promises to significantly increase technological
capabilities in sensing, communication, and computing. The quantum future looks
bright, and it is our hope that this monograph can play a part in that future.
HN thanks Clare Donald for proofreading the language in some parts of the
monograph, and Zhan Shi for her assistance in providing some figures for Sect. 6.2.
HN and NY also thank Oliver Jackson from Springer, for inviting them to write this
monograph and facilitating the writing process, and Karin de Bie, for technical
assistance in preparing the monograph prior to her retirement.
Sydney, Australia Hendra I. Nurdin

Yokohama, Japan Naoki Yamamoto
November 2016
References
[1] H.M. Wiseman, G.J. Milburn, Quantum Measurement and Control (Cambridge University
Press, Cambridge, 2010)
[2] K. Jacobs, Quantum Measurement Theory and its Applications (Cambridge University Press,
Cambridge, 2014)
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Quantum Feedback Control: A Brief History . . . . . . . . . . . . . . . . . 3
1.2 Classical Linear Systems and Control Theory. . . . . . . . . . . . . . . . . 6
1.2.1 Classical Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Linear Systems and Control Theory . . . . . . . . . . . . . . . . . . 8
1.2.3 Toward Systems and Control Theory for Linear
Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 10
1.3 Closed Linear Quantum Systems . . . . . . . . . . . . . . . . . . . . . . .... 11
1.4 Open Quantum Systems, the Markov Approximation,
and Quantum Langevin Equations . . . . . . . . . . . . . . . . . . . . . .... 12
1.4.1 Open Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . .... 12
1.4.2 Illustration of the Markov Approximation
and Markov Open Quantum System Dynamics . . . . . . .... 13
1.5 Linear Dynamical Quantum Systems:
Description and Physical Examples . . . . . . . . . . . . . . . . . . . . .... 16
1.5.1 Optical Cavities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 17
1.5.2 Non-degenerate Optical Parametric Amplifiers . . . . . . .... 19
1.5.3 Degenerate Parametric Amplifiers/Optical Parametric
Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.4 Opto-mechanical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5.5 Large Atomic Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2 Mathematical Modeling of Linear Dynamical Quantum Systems . . . 35
2.1 Quantum Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.1 The Boson Fock Space, Exponential Vectors,
and Fundamental Processes on the Fock Space . . . . . . . . . . 36
2.1.2 Adapted Processes and Quantum Stochastic Integrals . . . . . 37
2.1.3 The Quantum Itō Table in Vacuum
and the Quantum Itō Rule . . . . . . . . . . . . . . . . . . . . . . . . . . 39
vii
viii Contents
2.1.4 The Hudson–Parthasarathy Quantum Stochastic

Differential Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2 Linear Dynamical Quantum Systems: Joint Unitary Evolution
of Oscillators and Boson Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3 Equations of Motion: Real Quadrature Form and Complex
Mode Form. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.1 Real Quadrature Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.3.2 Complex Mode Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.3 Transfer Function of Linear Dynamical Quantum
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4 Inclusion of Idealized Static Transformations on Bosonic
Fields: The Bogoliubov Transformation . . . . . . . . . . . . . . . . . . . . . 50
2.4.1 Completely Passive Linear Dynamical Quantum
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.5 Physical Realizability Conditions and Parameterizations
for Linear Dynamical Quantum Systems . . . . . . . . . . . . . . . . . . . . 55
2.5.1 Physical Realizability Conditions for Linear QSDEs . . . . . . 55
2.5.2 Parameterization of Linear Dynamical Quantum
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.5.3 Linear Dynamical Quantum Systems
with Less Outputs Than Inputs . . . . . . . . . . . . . . . . . . . . . . 58
2.6 Stability of Linear Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . 59
2.7 Gaussian States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.7.1 Gaussian State of a Collection
of Single-Mode Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.7.2 Gaussian States of the Field and Their Fock Space
Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.7.3 Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.7.4 Coherent States of a Single-Mode Oscillator . . . . . . . . . . . . 68
2.7.5 Coherent States of a Bosonic Field . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3 Realization Theory for Linear Dynamical Quantum Systems . . .... 73
3.1 Architecture for Strict Realization . . . . . . . . . . . . . . . . . . . . . . .... 75
3.1.1 The Concatenation and Series Product and Reducible
Quantum Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.1.2 Main Synthesis Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.1.3 Systematic Synthesis of Linear Quantum Systems . . . . . . . 80
3.1.4 Illustrative Synthesis Example . . . . . . . . . . . . . . . . . . . . . . . 93
3.2 Architecture for Strict Realization Using Quantum Feedback
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 96
3.2.1 The Model Matrix and Concatenation
of Model Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 96
3.2.2 Edges, Elimination of Edges, and Reduced Markov
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 97
Contents ix
3.2.3 Main Synthesis Results . . . . . . . . . . . . . . . . . . . . . . . . .... 99

3.2.4 Synthesis of Completely Passive Systems . . . . . . . . . . .... 104
3.3 Transfer Function Realization . . . . . . . . . . . . . . . . . . . . . . . . . .... 106
3.3.1 Pure Cascade Realization of the Transfer Function
of Linear Quantum Systems . . . . . . . . . . . . . . . . . . . . .... 106
3.3.2 Conditions for Realizability by a Pure Cascade
Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 107
3.3.3 Transfer Function Realization of Completely Passive
Linear Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.3.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Appendix A: Adiabatic Elimination of Coupled Cavity Modes . . . . . . . . 114
Appendix B: Proof of Theorem 3.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Appendix C: Proof of Lemma 3.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Appendix D: Proof of Corollary 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4 Quantum Filtering for Linear Dynamical Quantum Systems. . . . . . . 123
4.1 Quantum Conditional Expectations . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1.1 Quantum Probability Space . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1.2 Conditional Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.2 Quantum Filtering Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.2.1 Quantum Filtering: The Idea . . . . . . . . . . . . . . . . . . . . . . . . 128
4.2.2 Quantum Filter: Multiple-Input Multiple-Output Case . . . . . 131
4.2.3 Stochastic Master Equation . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2.4 QND Interaction and the Projection Postulate . . . . . . . . . . . 136
4.3 Quantum Kalman Filter for Gaussian Linear
Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.3.1 Example in Quantum Optics . . . . . . . . . . . . . . . . . . . . . . . . 137
4.3.2 Quantum Filtering: Multiple-Input and Multiple-Output
Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.4 Robust Linear Quantum Observers . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.4.1 Guaranteed-Error Robust Observer . . . . . . . . . . . . . . . . . . . 144
4.4.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.4.3 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5 Feedback Control of Linear Dynamical Quantum Systems . . . . . . . . 153
5.1 Measurement-Based Quantum Feedback Control . . . . . . . . . . . . . . 153
5.1.1 Controlled Quantum Evolution and Quantum Filter . . . . . . 153
5.1.2 Measurement-Based LQG Control . . . . . . . . . . . . . . . . . . . 157
5.2 Coherent Feedback Quantum LQG Control . . . . . . . . . . . . . . . . . . 163
5.2.1 Reformulating the Quantum LQG Problem
into a Rank-Constrained LMI Problem . . . . . . . . . . . . .... 167
x Contents
5.2.2 Numerically Solving the Rank-Constrained

LMI Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.2.3 An Extension of the Numerical Procedure . . . . . . . . . . . . . 172
5.2.4 Quantum LQG Control Design Examples . . . . . . . . . . . . . . 175
5.2.5 Quantum LQG Controller Design Example II . . . . . . . . . . . 178
5.2.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.3 Coherent Feedback H 1 Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.3.1 Dissipation Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.3.2 H 1 Controller Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.3.3 H 1 Synthesis in Quantum Optics . . . . . . . . . . . . . . . . . . . . 191
5.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Appendix A: Proof of Theorem 5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Appendix B: Proof of Lemma 5.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
6 Linear Systems and Control Theory for Quantum Information . . . . 203
6.1 Dissipative Generation of Pure Gaussian States . . . . . . . . . . . . . . . 204
6.1.1 General Condition for Pure Gaussian State Generation . . . . 205
6.1.2 Synthesizing a Dissipative Gaussian System . . . . . . . . . . . . 206
6.1.3 Gaussian Cluster State Generation via Dissipation . . . . . . . 207
6.1.4 A QSDE Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6.1.5 Remarks and Further Reading . . . . . . . . . . . . . . . . . . . . . . . 212
6.2 Enhancing Continuous-Variable EPR Entanglement . . . . . . . . . . . . 212
6.2.1 Stability Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.2.2 The Ideal Lossless Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
6.2.3 Effect of Losses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6.2.4 Comparison with Conventional Schemes . . . . . . . . . . . . . . 218
6.2.5 Effect of Nonzero Transmission Delays . . . . . . . . . . . . . . . 224
6.3 Force Sensing and Back-Action Evasion . . . . . . . . . . . . . . . . . . . . 226
6.3.1 Back-Action Evasion and the Standard Quantum Limit . . . 227
6.3.2 System Theoretical Characterization of BAE . . . . . . . . . . . 228
6.3.3 Coherent Feedback for BAE . . . . . . . . . . . . . . . . . . . . . . . . 229
6.3.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
6.4 Quantum Memory with Decoherence-Free Subsystem . . . . . . . . . . 232
6.4.1 General Schematic of an Ideal Quantum Memory . . . . . . . . 233
6.4.2 The Zero Dynamics Principle . . . . . . . . . . . . . . . . . . . . . . . 235
6.4.3 Perfect State Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
6.4.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.5 Robust Quantum Amplification via Coherent Feedback . . . . . . . . . 239
6.5.1 The Phase-Preserving Amplifier . . . . . . . . . . . . . . . . . . . . . 240
6.5.2 Coherent Feedback Control for a Quantum Amplifier . . . . . 242
6.5.3 Example: Non-degenerate Optical Parametric Amplifier . . . 244
6.5.4 Added Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Contents xi
6.6 Feedback Control Experiments . . . . . . . . . . . . . . . . . . . . . . . . .... 246

6.6.1 Coherent Feedback for Optical Squeezing
Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 247
6.6.2 Measurement-Based Feedback for Spin Squeezing
in Atomic Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . .... 251
6.6.3 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 253
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 254
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Notation
pffiffiffiffiffiffiffi
ı 1
R The set of real numbers
Rþ The set of nonnegative real numbers
C The set of complex numbers
H A Hilbert space
L2 ðI ; HÞ The set of H-valued square integrable functions on a
Lebesque measurable set I
BðHÞ The space of bounded linear operators mapping H to itself
ker(X) The kernel or null space of an operator X
range(X) The range of an operator X
½ a1 a2 . . . an A row vector with elements a1 ; a2 ; . . .; an
ða1 ; a2 ; . . .; an Þ A row vector with elements a1 ; a2 ; . . .; an
ðÞ The conjugate of a complex number, and the adjoint of a
Hilbert space operator or a matrix of Hilbert space operators
ðÞ# For a matrix of complex number or Hilbert space operator X,
denotes the elementwise conjugation operation X # ¼ ½Xij
ðÞ> For a matrix of complex number or Hilbert space operators
X, denotes the transposition operation X > ¼ ½Xji
<fg For a matrix of complex number or Hilbert space operators
X, denotes the real part, <fX # g ¼ 1=2ðX þ X # Þ
=fg For a matrix of complex number or Hilbert space operators
X, denotes the imaginary part, =fX # g ¼ 1=2ðıX þ ıX # Þ
[A, B] Commutator AB BA of two operators acting on the same
Hilbert space
½x; y Commutator matrix xy ðx# y> Þ> between two column
vector x and y of operators acting on the same Hilbert space.
It becomes ½x; y> ¼ xy> ðxy> Þ> when x and y have
self-adjoint entries
xiii
xiv Notation
jî A ket, a normalized vector on a Hilbert space

hˆj A bra, the conjugate transpose of a normalized vector jî on
a finite dimensional Hilbert space or the continuous linear
functional on the Hilbert space associated with jî
hˆj jî Quantum expectation with respect to the pure state jî
hi Quantum expectation
Tr Trace of a trace-class operator. If M ¼ ½Mij is matrix of
(trace-class) operators, then TrðMÞ can mean either TrðMÞ ¼
Pn
i¼1 Tr Mjj (only if M is square) or TrðMÞ ¼ Tr Mij ,
depending on the context
TrH Partial trace over H
In An n n identity matrix, subscript n can be dropped if
dimension of matrix can be inferred from context
0n ; 0mn An n n and m n zero matrix, respectively, subscript n or
m n can be dropped if dimension of matrix can be inferred
from context
diagðM1 ; . . .; Mn Þ A block diagonal matrix with M1 ; . . .; Mn on its diagonal
blocks
diagn ðMÞ In M
0 1
J The skew-symmetric matrix
1 0
Jn The skew-symmetric matrix In Jn
Cs ðHÞ Boson/symmetric Fock space over H
eðf Þ Coherent/exponential vector in Cs ðHÞ
Wðf Þ Weyl operator on a Boson Fock space
q Amplitude/position quadrature of a single mode resonator
p Phase/momentum quadrature of a single mode resonator
»i ðtÞ Quantum white noise for incoming field i
·i ðtÞ Quantum white noise for outgoing field i
Rt
Ai ðtÞ Input annihilation operator for field i, Ai ðtÞ ¼ 0 »i ðsÞds
Y i ðtÞ Output annihilation operator for field i
»ðtÞ »ðtÞ ¼ ð»1 ðtÞ; »2 ðtÞ; . . .; »m ðtÞÞ> ; m denotes number of fields
·ðtÞ ·ðtÞ ¼ ð·1 ðtÞ; ·2 ðtÞ; . . .; ·m ðtÞÞ> , m denotes number of fields
AðtÞ AðtÞ ¼ ðA1 ðtÞ; A2 ðtÞ; . . .; Am ðtÞÞ> ; m denotes number of
fields
YðtÞ YðtÞ ¼ ðY 1 ðtÞ; Y 2 ðtÞ; . . .; Y m0 ðtÞÞ> ; m0 denotes number of
output fields (m m0 )
Kij ðtÞ Gauge process from channel j to channel i
KðtÞ KðtÞ ¼ ½Kij ðtÞi;j¼1;...;m , m denotes number of fields
/ Series product of Markov input–output systems
Concatenation product of Markov input–output systems
Notation xv
vNfyðsÞ; 0 s tg von Neumann algebra generated by the spectral projections

of a self-adjoint process fyðsÞ; 0\s\tg
Yt von Neumann algebra generated by the spectral projections
of a self-adjoint process fyðsÞ; 0\s\tg
Chapter 1
Introduction
Abstract This chapter gives a brief overview of and introduction to quantum

feedback control and linear quantum systems. It begins with a brief history of quan-
tum feedback control, in particular the development of quantum analogues of ideas
from stochastic control theory, and discusses the notion of closed and open quantum
systems, the Markov property, and quantum Langevin equations. This chapter then
proceeds with an introduction to linear quantum systems and several physical exam-
ples thereof, including optical cavities, non-degenerate and degenerate parametric
amplifiers, and large atomic ensembles.
This monograph is concerned with the modeling and control of a special class of
quantum systems that shall be referred to as linear dynamical quantum systems or
simply linear quantum systems for short. It is also referred to in the literature as
linear quantum stochastic systems, with the qualifier stochastic added to emphasize
the role of quantum stochastic processes in the dynamics of these systems. This
class [1–4] represents multiple distinct open quantum harmonic oscillators that are
coupled linearly to one another and also to external fields such as Gaussian coherent
laser beams and single/multiple photon fields. Indeed, many important devices can
be modeled by a linear quantum system, ranging from optical cavities, mesoscopic
mechanical resonators, optical and superconducting parametric amplifiers, to linear
atomic quantum memories; see, e.g., [5, Chaps. 7 and 10] and [6–14]. In analogy
with the distinguished classical linear state-space models from modern control the-
ory, see, e.g., [15, 16], the dynamics of a linear quantum system in the Heisenberg
picture of quantum mechanics can be completely described in terms of a quartet of
matrices A, B, C, D. Note that throughout the monograph, we will use the quali-
fier “classical” to refer to objects that are not quantum mechanical (i.e., not subject
to the laws of quantum mechanics). However, the matrices A, B, C, D in a linear
Sections 1.5.2 and 1.5.3 contain some materials reprinted, with permission, from [44] © 2012
IEEE.
© Springer International Publishing AG 2017 1

H.I. Nurdin and N. Yamamoto, Linear Dynamical Quantum Systems,
Communications and Control Engineering, DOI 10.1007/978-3-319-55201-9_1
2 1 Introduction
quantum system cannot be arbitrary but must satisfy a set of constraints imposed by
quantum mechanics for them to represent a physical system, a restriction not encoun-
tered in the classical setting. As first demonstrated in [2], for the case where D is
of the form D = [ I 0 ], with I denoting an identity matrix, A and B must satisfy a
certain nonlinear equality constraint, while B and C satisfy a linear equality con-
straint. These constraints on the A, B, C, D matrices are referred to as physical
realizability constraints [2]. In analogy with classical linear systems, linear quantum
systems present themselves as rather nice and relatively tractable class of quantum
systems with which to explore and potentially develop fundamental ideas and princi-
ples of quantum control, possibly emulating the fundamental role of classical linear
systems in the early development of systems and control theory.
A growing number applications of linear quantum systems have been proposed
or experimentally demonstrated in the literature. From a theoretical point of view,
they have been studied as sensors for extremely weak signals [17, 18] such as
gravitational-wave forces [19–23] and have been proposed as coherent feedback con-
trollers [2, 24] (treated in Chap. 5 of this monograph) for cooling of opto-mechanical
systems [7], for back-action evasion in linear quantum sensors [25, 26] and realiz-
ing quantum non-demolition measurements [25], and for creating decoherence-free
subsystems [25, 27]. They have also been proposed as optical filters for modify-
ing the wave packet shape of single- [28, 29] and multi-photon [30] sources, and
as a means for dissipatively generating Gaussian cluster states [31, 32], which are
of interest for continuous-variable one-way quantum computers [33]. Applications
of measurement-based feedback control strategies to linear quantum systems (see
Chap. 5 of this monograph) have also been investigated for the purpose of cooling
the motion of a particle [1, 34] and enhancing entanglement between separated nodes
in a linear network [35, 36]. More recently, linear quantum models have been shown
to be highly relevant for analyzing linear quantum memories [9, 12–14, 37]. Quan-
tum memory schemes using switched linear quantum systems have been proposed
based on atomic ensembles [38] and by combining passive linear optics with coher-
ent feedback [27]. Experimentally, linear quantum systems have been demonstrated
for broadband disturbance attenuation in bulk quantum optics [39] as well as silicon
photonics [40]. They have also been experimentally shown to be able to modify the
characteristics of squeezed light produced by an optical parametric oscillator [41–45],
and to reshape the dynamics of an electromechanical circuit [8]. In [8], it is high-
lighted that multiple-input multiple-output linear state-space-like modeling of linear
quantum networks, which are extensively employed in modern control theory for
classical linear systems, provides a powerful tool for analyzing these networks when
they involve many components and inputs and outputs. Venturing beyond quantum
information technologies, linear quantum systems are also of interest as a component
for classical signal processing on quantum devices, for instance for processing light
when photons, rather than electrons, are employed to transport information between
cores on a chip and between chips [46]. Together with proposed nonlinear ultra-low
power optical logic gates, such as developed in [47], they form the building blocks
for envisaged ultra-efficient all-optical classical information processing circuits of
the future.
1 Introduction 3
The rest of this introductory chapter is organized as follows. Section 1.1 provides
a brief historical overview of the development of quantum feedback control theory
from the 1970s up to the present, while Sect. 1.2 gives an overview of classical linear
systems and control theory and the prospects for a linear quantum systems and control
theory. This is then followed by a discussion of closed linear quantum systems in
Sect. 1.3 and open quantum systems, in particular the special class of Markov open
quantum systems, in Sect. 1.4. Finally, Sect. 1.5 introduces linear dynamical quantum
systems as a special type of Markov open quantum systems and illustrates some
concrete physical examples of such systems along with the applications in which
they are employed.
1.1 Quantum Feedback Control: A Brief History
We begin with a brief historical overview of the field of quantum feedback control,
especially from the systems and control theory perspective. Of course, this will only
be a part of the whole story, as relevant for the theme of this monograph, see [48] for an
additional overview. We should also remark that in the 1980s seminal work by Tarn,
Huang, and Clark on control of closed quantum systems, represented by a controlled
(time-dependent) Schrödinger equation, appeared, see, e.g., [49, 50]. However, due
to the closed nature of the system, the type of control for these systems is open-loop
control rather than feedback control. That is, the control signal is predesigned and
does not rely on extracting (classical) information from the system by observing
it, or exchanging quantum information with another quantum system in a feedback
interconnection. A comprehensive treatment of open-loop control of closed quantum
systems can be found in D’Alessandro’s text “Introduction to Quantum Control and
Dynamics” [51]. However, the concern of this monograph is linear quantum systems
and feedback control on such systems; therefore, open-loop control schemes are not
discussed further in the monograph.
Returning to our brief historical overview, it would be fair to say that the late
Viacheslav “Slava” Belavkin pioneered the development of quantum mechanical
analogues of notions and tools from stochastic systems and control theory. We find
his seminal works in this area going back to the 1970s.1 Belavkin was a student of the
renowned Russian probabilist and stochastic control theorist Ruslan Stratonovich and
the extension of ideas from classical stochastic systems and control theory (recall that
classical refers to being non-quantum) to the so-called observable quantum systems
(in the terminology of Belavkin) was a natural progression of the research themes
being pursued in the Stratonovich school. One of the main questions that Belavkin
aimed to address was how to describe the time evolution of the quantum state of an
open system under continuous-time measurements/observations, starting with a study
of quantum harmonic oscillators attached to a quantum transmission line. Belavkin’s
work in this direction is a seminal contribution toward a mathematical quantum
1 Most,if not all, of Belavkin’s works have have been archived on his memorial homepage at the
University of Nottingham, https://www.maths.nottingham.ac.uk/personal/vpb/.
4 1 Introduction
systems and control theory. Using conceptual tools from the theory of open quantum
systems that were available at the time, such as the notion of a “quantum instrument”
to describe quantum-state evolution [52], Belavkin succeeded in deriving what is
now called the Belavkin quantum filtering equation or the stochastic master equation
(SME) in the quantum optics community. This equation describes the evolution of
a special class of open quantum systems (including a large class of quantum optical
systems) under continuous observations. Here, quantum filtering, as in the classical
setting of nonlinear filtering theory, is fundamentally connected to a quantum notion
of conditional expectations. The derivation of the quantum filtering equation marks
the first step toward establishing what is now referred to as measurement-based
quantum feedback control, a problem which Belavkin also subsequently solved,
see, e.g., [53–56], using techniques that can be viewed as essentially the quantum
analogue of approaches for solving partially observable stochastic control problems,
albeit in a more abstract form that relies heavily on concepts from mathematical
physics and the theory of operator algebras.
Around the mid-1980s physical modeling of quantum systems interacting with
propagating optical fields was undertaken by Gardiner and Collett. The starting point
was to obtain an accurate dynamical model of several open optical devices, such as
a cavity (resonator) containing a nonlinear crystal [57, 58] (to be elaborated further
in Sects. 1.5.1–1.5.3). In the case of the optical cavity, the system is open since the
internal optical field inside the cavity couples to external incoming and outgoing
optical fields. The major significance of their work is to derive a mathematical model
for the cavity as a dynamical system having inputs and outputs. Input–output models
have been central in systems and control theory, and this point in time marks the
emergence of analogous quantum input–output models. Notably, it later turned out
that this specific input–output modeling method for quantum optics is applicable to
many physical devices interacting with external fields, such as an atomic ensemble
and a nano-mechanical oscillator [5]. The key tool enabling such accurate modeling
is quantum stochastic calculus. Though such a calculus was obtained independently
by Gardiner and Collett in [58] through physical intuition and arguments, a more
general and mathematically rigorous version of quantum stochastic calculus had in
fact been established earlier by Hudson and Parthasarathy in their landmark paper
[59]. However, although Hudson and Parthasarathy devised a more complete calcu-
lus, they did not formulate the important notion of input–output quantum systems
with this calculus, as Gardiner and Collett had done in their work. The input–output
formulation of Gardiner and Collett and its subsequent developments and related
research are summarized in the now classic Gardiner and Zoller quantum optics text
“Quantum Noise” [5]. This book comprehensively describes several quantum sys-
tems having inputs and outputs with dynamical equations, described using quantum
stochastic differential equations (QSDEs) as a quantum analogue of ordinary sto-
chastic differential equations. Thus, the notion of a “quantum input–output system”
could be established, paving the way for a systems and control theory for quantum
systems. Of course, this input–output formulation can be formulated in the rigorous
framework of the Hudson–Parthasarathy quantum stochastic calculus, see, e.g., [60].
1.1 Quantum Feedback Control: A Brief History 5
In the early 1990s, a version of Belavkin’s filtering theory was independently

rediscovered simultaneously by several researchers in quantum optics [61–63], being
obtained from a completely different point of view, using a different approach, and
derived for the purpose of stochastic simulation of open quantum systems rather
than for estimation. In the quantum optics literature, the theory goes by the name
of “quantum trajectory theory” and a comprehensive treatment of this theory can
be found, e.g., in the text [63] of Carmichael, one of its originators. As mentioned
above, Belavkin’s filtering equation in the latter is referred to as the SME. Roughly
speaking, the SME describes the dynamics of the quantum state of a system condi-
tioned on measurement results, which are obtained by detecting the system’s output
field. In other words, it models the evolution of the system as it is continuously
monitored in time. Since the measurement outcomes can be processed and fed back
to influence the system, the introduction of the SME spurred the development of
quantum feedback control for optical systems in the physics literature in the 1990s,
in particular through the influential contributions of Wiseman and Milburn; see their
text [48] for a discussion of various important developments in this direction. This
body of work had quite a distinctive flavor from the work of Belavkin in that system
analysis and derivation of feedback laws relied on rigorous and extensive use of
physical principles rather than mathematical ones. At around this period, the disci-
pline of quantum information science [64] was in its early days, and it would later
be widely recognized that some form of quantum control would be indispensable for
the realization of any practical quantum information processing systems [48, 65].
This is simply because such systems tend to be very fragile and can easily interact
with their environment (acting as a disturbance to the system).
The advent of the Hudson–Parthasarathy quantum stochastic calculus opened a
path toward a modern formulation of Belavkin’s work in a more accessible form and
in a much closer analogy to classical stochastic control theory, including through
the contributions of Bouten and van Handel, see, e.g., [60, 66–68]. Beside their
own important original contributions to the subject, they translated major ideas in
Belavkin’s more abstract work into a language that is familiar to control theorists,
probabilists, and engineers with a working knowledge of stochastic control theory.
This brings out much more explicitly the role of the quantum filtering equations as
a quantum analogue of classical filtering equations. In the classical case, the filter-
ing equations represent an algorithm for real-time mean-square optimal estimation
of a system variable. More precisely, it is a real-time Bayesian update rule based
on the measurement record. This clarity helps in formulating quantum feedback
control problems fully within the framework of quantum probability theory (the
non-commutative extension of classical probability theory), and one can pose many
important questions, such as, how can we utilize estimates from a quantum filter to
engineer a feedback controller? Is it possible to formulate and solve optimal feedback
control problems? If the system contains some uncertainty, can we design a robust
estimator/controller? Can we replace a classical feedback controller composed of an
estimator and actuator by a fully quantum device?
Many of the questions mentioned above have been fully or partly resolved. More
recent developments have witnessed the emergence of quantum versions of electri-
cal network synthesis theory, measurement-based/coherent feedback LQG control,
6 1 Introduction
H ∞ control, robust estimation and control, stability analysis, system identification,

to name a few. They are not mere analogues of classical theories, but have a role
to play in applications to quantum mechanical systems, particularly in quantum
information science, some of which will be highlighted in Chap. 6. Moreover, exper-
imental demonstrations unimaginable a decade ago are within the scope of current
technologies. The long-time dream of being able to control and manipulate quantum
systems has now become a reality. Indeed, the Nobel Prize in Physics in 2012 was
jointly awarded to experimental physicists Serge Haroche and David Wineland with
the citation “for ground-breaking experimental methods that enable measuring and
manipulation of individual quantum systems.”
Due to its practical importance for quantum information processing, quantum
feedback control theory has attracted extensive attention from various research dis-
ciplines. To bring together the interdisciplinary communities working in quantum
control theory, and to accelerate the development of this emerging field, the workshop
series “Principles and applications of control in quantum systems” (PRACQSYS)
was established and held for the first time on the Caltech campus in 2004 [69]. Since
the successful first meeting, the PRACQSYS series of workshops has been succes-
sively held almost every year in various locations in the USA, Europe, Australia, and
Asia, playing the role of a venue to stimulate exchange of ideas among researchers
working in the field.
To conclude, we emphasize that, from both theoretical and experimental view-
points, the special class of linear quantum systems as the focus of this monograph
can play a role in further accelerating developments in the field of quantum control
theory for reasons that will be elaborated in Sect. 1.2.3 and illustrated in Chap. 6.
1.2 Classical Linear Systems and Control Theory
Here, we briefly describe some important results developed in systems and control
theory for classical linear systems. Indeed, they remind us of the strength and mer-
its of the theory, thereby serving as a motivation to study linear quantum systems
encountered in the quantum regime.
1.2.1 Classical Linear Systems
There are several reasons why linear systems are of particular importance. The first
is that many practical systems of interest have a dynamical time evolution that can be
modeled quite faithfully by a linear dynamical system described by linear (determin-
istic or stochastic) differential equations. Sometimes the system contains a nonlinear
term that can be well approximated by a local linear term via linearization.
Let us examine here a typical example of such classical linear systems: a classical
mechanical oscillator. The basis of the physical modeling is, of course, the Newtonian
1.2 Classical Linear Systems and Control Theory 7
equation of motion, or more generally the Hamilton equation
dX ∂H ∂X ∂X ∂H
= {H, X } = − , (1.1)
dt ∂ p ∂q ∂ p ∂q
where H is the Hamiltonian and (q, p) is a pair of canonical variables having the
meaning of position and momentum of a single mechanical degree of freedom. In
particular, let us consider the following quadratic Hamiltonian:
mω 2 2 1 2
H= q + p − qu.
2 2m
This governs the dynamics of an object in a harmonic potential with resonance
frequency ω. The second and third terms represent the kinetic energy, where m is the
mass of the object and u(t) is a time-varying driving (control) force, i.e., an input to
the system. Then, (1.1) immediately yields the following linear equation:

d q 0 1/m q 0 q
= + u, y = [1 0] . (1.2)
dt p −mω 2 0 p 1 p
Here, we have added an output equation y(t) = q(t), under the assumption that q(t)
can be observed directly. Note that even in the case where the system contains some
nonlinear effects, the model (1.2) could be still effective; for instance, the dynamical
equation q̇ = p/m + q 3 can be well approximated using q̇ = p/m if the position q
is confined in a small region satisfying |q| 1.
Other than the above example, there are numerous physical systems that can be
modeled using linear equations, such as the motion of an aircraft at equilibrium and an
atomic force microscope; see [70]. They usually contain multiple variables, inputs,
and outputs, which are sometimes of the order of hundreds, thousands, or more.
Nonetheless, even large-scale multiple-input and multiple-output (MIMO) systems
can often still be modeled by a linear equation, which can be of the following general
form:
ẋ = Ax + B1 u + B2 w1 , y = C x + D1 u + D2 w2 . (1.3)
Here, x ∈ Rn denotes the vector of the system’s state variables, u is the vector of
control input, and y is the vector of observed output. Further, in the above equation,
additional noise variables w1 and w2 are included to represent a more general open
system under environmental effects (w1 and w2 are external signals from the envi-
ronment that disturb the system). Note however that, as we will see later, adding a
noise term is essential in the quantum case if one wishes to do feedback control.
(A, B1 , B2 , C, D1 , D2 ) are matrices representing the system dynamics. The strength
of systems and control theory is that it can deal with a general model of the above
form as long as matrices A, B1 , B2 , C, D1 , D2 satisfy the conditions required by the
particular technique being employed. Lastly, we remark that the above differential
equation can be analytically solved as follows:
8 1 Introduction
t
x(t) = e At x(0) + e At e−As B1 u(s) + B2 w1 (s) ds.
0
This simple fact makes the analysis and controller design for linear systems more
tractable and underlies some of the success of linear systems and control theory.
1.2.2 Linear Systems and Control Theory
Once a linear model of the system of interest is obtained, then linear systems and
control theory provides the solution to several important problems. Let us return to
the oscillator system (1.2). For this system, clearly we cannot directly control q(t)
with the input u(t). But can we do that indirectly through p(t)? Also, we cannot
directly observe p(t). However, can we observe it indirectly through the output
y(t) = q(t)? One may also think about the important problem of designing optimal
control signals. For example, by appropriately designing the input signal u(t) as a
function of y(t), we want to drive q(0) = 0 to q(T ) = 0 at a specific time t = T ,
under a certain optimality criterion; but how can we formulate this problem and
solve it? Moreover, it is often the case that some parameters contained in the system
are unknown; for instance, if the driving force u(t) is not a controllable force but an
unknown signal, then how can we use the output y(t) to estimate u(t)? Linear systems
and control theory provides rigorous and systematic methodologies to solve these
problems, analytically or numerically. Note that, for a nonlinear system, the problems
can become significantly more difficult and in general no analytical solutions are
available. Below we briefly highlight some important results.
(i) Let us consider the controllability problem of determining which components
of x(t) can be directly/indirectly controlled by the input u(t) and the observability
problem of determining which components of x(t) can be observed directly/indirectly
from the output y(t). These notions are completely characterized by the following
controllability matrix C and the observability matrix O:
C = [B1 , AB1 , A2 B1 , . . . , An−1 B1 ],

O = [C , A C , (A )2 C , . . . , (A )n−1 C ] .
First consider the controllability problem. Let us assume rank(C) = m < n. Then,
there exists a coordinate transformation x → x = [x1 , x2 ] , where the dynamics
part of (1.3) can be represented as

d x1 A11 A12 x1 B11 B21
= + u + w1 , (1.4)
dt x2 O A22 x2 O
B22
where x1 (t) ∈ Rm . Thus x1 (t) constitutes a controllable subsystem with respect to
(w.r.t.) u(t), while x2 (t) cannot be controlled by u(t).
Next let us consider the observability problem and assume rank(O) = < n.
Then, as in the above case, the system dynamics has the following representation in
a certain coordinate:

d x1 A11 O x1
= + B1 u + B2 w1 ,
dt x 2 A 21 A 22 x 2

x
y = [C1 , O] 1 + D1 u + D2 w2 . (1.5)
x2
This clearly shows that x2 (t) cannot be observed from the output y(t), while x1 (t)
can; hence x1 (t) constitutes an observable subsystem while x2 (t) is not.
(ii) A particularly important technique related to the observability property is that
of filtering [71]; that is, we wish to estimate the system variable x(t) only through the
output y(t), in the presence of the noisy environment. Let us denote such an estimator
by z(t), which should be a function of the set of outputs, Yt = {y(s) | 0 ≤ s ≤ t}.
Then, if the noises w1 and w2 are independent Gaussian white noises, the most
celebrated and powerful result of linear stochastic systems theory is that the optimal
estimate

x̂(t) = argminz E x(t) − z(t) x(t) − z(t) Yt
can be explicitly calculated by the following Kalman filter:
d
x̂(t) = A x̂(t) + B1 u(t) + V (t)C y(t) − C x̂(t) , (1.6)
dt
d
V (t) = AV (t) + V (t)A − V (t)C C V (t) + B2 B2 , (1.7)
dt
where V (t) represents the estimation error covariance matrix:

V (t) := E x(t) − x̂(t) x(t) − x̂(t) . (1.8)
In the above formula, E denotes the mean with respect to the Gaussian distribution.
Also for simplicity, we have assumed D1 = 0 and D2 = I , and the covariance matrix
of w1 and w2 are both the identity matrices. Equation (1.7) is called a Riccati differ-
ential equation. The initial conditions x̂(0) and V (0) are determined according to an
“initial guess,” which of course should be carefully considered.
(iii) In the linear case, several control problems can be well formulated and, further,
some of them have analytical or partly analytical solutions. A particularly beautiful
and useful theory is the theory of Linear Quadratic Gaussian (LQG) control [72,
73]. The problem is, for the linear system (1.3), to find a feedback control law u(t)
that is a function of Yt and that minimizes the following quadratic-type cost function:
T
J [u(t)] = E x(t) P x(t) + u(t) Qu(t) dt + x(T ) Rx(T ) , (1.9)
0
10 1 Introduction
Fig. 1.1 Diagram of Dynamics

classical LQG control x
y
Controller Filter
u x̂
where (P, Q, R) are the weighting matrices of the appropriate dimensions (Q is

positive definite, and (P, R) are positive semidefinite). Let us assume again that
w1 and w2 are independent Gaussian white noises, and D1 = 0, D2 = I . Then, the
optimal feedback input minimizing J [u(t)] is given by u opt (t) = −Q −1 B1 K (t)x̂(t),
where the time-varying matrix K (t) satisfies
K̇ (t) + K (t)A + A K (t) − K (t)B1 Q −1 B1 K (t) + P = O, (1.10)
with terminal condition K (T ) = R. A notable fact is that the optimal feedback control
input u opt (t) is a function of the optimal estimate x̂(t), the latter being generated
recursively from the Kalman filter equations (1.6) and (1.7). This means that u opt (t)
is obtained by computing K (t) and x̂(t) independently of each other. This fact is
known as the separation principle; see Fig. 1.1.
1.2.3 Toward Systems and Control Theory for Linear

Quantum Systems
In the previous sections, we have seen that, in the classical case, linear systems and
control theory provides a number of powerful tools that can be applied in many
practical situations. It is surprising that many quantum system models of interest, in
particular arising in quantum optical and quantum opto-mechanical systems, can be
described or well approximated by a linear dynamical equation, and, further, there
exists a quantum version of Kalman filtering and LQG control. One of the main
purposes of this monograph is, as mentioned at the beginning of this chapter, to
detail this class of quantum systems, and describe the quantum linear systems and
control theory. There is certainly a beautiful underlying theory that is still undergoing
development and finding natural applications.
Classical linear systems have played a major role in the development of modern
control theory, where many crucial systems and control concepts were first discovered
for these classes of systems first before being extended to other classes of dynamical
systems. Linear quantum systems could potentially also play a role as a class of
quantum models with which to investigate and discover fundamental concepts for
quantum systems and control theory. This will not only be of theoretical but also of
practical interest. Various physical systems that can be modeled as linear quantum
systems, as alluded to above, are relevant for applications such as nano-mechanical

cooling [7, 74], high-precision sensing [19–23], and quantum memories [9, 12–14,
37, 38].
1.3 Closed Linear Quantum Systems
We will now describe linear quantum systems, starting with a quantum version of
the classical harmonic oscillator (1.2). The Hamiltonian has exactly the same form:
mω 2 2 1 2
H= q + p − qu. (1.11)
2 2m
But in quantum mechanics, the position q and the momentum p must be represented
by self-adjoint operators satisfying the canonical commutation relation (CCR)
[q, p] = qp − pq = 2ı. (1.12)
For convenience, here we assume physical units in which the Planck constant is = 2,
a commonly used convention in quantum optics and continuous-variable quantum
information that we shall adopt for the remainder of the monograph without further
comment; see, e.g., [75, 76]. The dynamical equation of (q, p) is described by the
Heisenberg equation of motion rather than the Hamilton equation (1.1); i.e., given a
Hamiltonian H , the dynamics of any quantum operator X evolves in time according
to the following differential equation:
dX
= ı[H, X ]. (1.13)
dt
Hence, substituting the quadratic Hamiltonian (1.11) for (1.13), we find that X = q
and X = p obey the following linear equation:

d q 0 1/m q 0
= + u. (1.14)
dt p −mω 2 0 p 1
Surprisingly, this has completely the same form as the dynamical equation of the
classical harmonic oscillator, (1.2). But recall that q and p are operator-valued vari-
ables, which always satisfy the CCR (1.12). Actually, when u ≡ 0 for simplicity, the
above dynamical equation has a solution of the following form:
q(t) = cos(ωt)q(0) − sin(ωt) p(0), p(t) = sin(ωt)q(0) + cos(ωt) p(0),
thus, [q(t), p(t)] = [q(0), p(0)] = 2ı holds for all t.

12 1 Introduction
Another important fact is that, although (q(t), p(t)) evolve in time in a deter-
ministic way, they are non-commutative random variables or observables, with their
spectra (eigenvalues) corresponding to possible measurement outcomes (if they are
measured). The mean values of these observables are hence well defined; let q(t)
and p(t) be those mean values, then they satisfy

d q 0 1/m q 0
= + u. (1.15)
dt p −mω 2 0 p 1
As now the means q(t) and p(t) are scalar variables, this is a classical dynamical
equation. This can be analyzed in the same way as with (1.2). For instance, it is
possible to plot a trajectory of the dynamics (1.15) in the phase space of ( q , p ).
We now highlight an important difference between the classical and quantum
systems considered above. Unlike the classical case (1.2), there is no output equation
accompanying the quantum dynamics (1.14). This is because the quantum harmonic
oscillator is closed in that it is perfectly isolated from any external influence; that
is, roughly speaking, it evolves “inside a shielded box.” This means that neither
the operator-valued dynamics (1.14) nor the mean dynamics (1.15) represent the
real-time evolution of the system under a monitoring (observation) process. Hence,
the input u(t) appearing in the equations cannot depend on a quantity obtained
from observation of the system (e.g., by performing continuous measurement), and
feedback control is not possible. The input u(t) can only be a signal that is designed
independently of the real-time dynamics of the system, and the only type of control
possible for a closed system is open-loop control. Therefore, if we want to manipulate
the system via real-time feedback control, the system should be open, as described
in the next section.
1.4 Open Quantum Systems, the Markov Approximation,

and Quantum Langevin Equations
1.4.1 Open Quantum Systems
Real-world quantum systems are never completely decoupled from their environ-
ment. In reality, they will, to some degree, be interacting with their environment.
Quantum systems interacting with an external environment are said to be open quan-
tum systems. They are indispensable for modeling various important physical phe-
nomena, for instance, to model the decay of the energy of an atom by spontaneous
emission of photons (particles of light). The environment is often modeled to be
another quantum system that can be viewed as a heat bath to which an open quantum
system can dissipate energy or from which it can gain energy. To describe this inter-
action between an open quantum system and an external heat bat, often idealizations
are introduced to derive approximate but tractable mathematical models. One such
1.4 Open Quantum Systems, the Markov … 13
idealization is the so-called Markov assumption: The coupling of the system and bath
are essentially “memoryless” in the sense that future evolution of the dynamics of the
system given its present state is independent of its past states and those of the bath.
Open quantum systems with such a property are said to be Markov. Markov open
quantum systems are important as they provide effective and accurate approximate
models for various practically important open quantum systems that are encountered
in fields such as quantum optics, opto-mechanics, and microwave superconducting
circuits. As models, their power derives from their relatively tractability for analysis
as their dynamics can be written in terms of first-order operator differential equations.
In this section, we give an illustration of the Markov approximation in a concrete
physical setting. This is followed by an introduction to the evolution equation (in the
Heisenberg picture) for a special class of Markov open quantum dynamics, known
as the quantum Langevin equations.
1.4.2 Illustration of the Markov Approximation and Markov

Open Quantum System Dynamics
Here, we shall derive the quantum Langevin equation, which describes the evolution
of a large class of Markov open quantum system, for a single quantum harmonic
oscillator with mode c (satisfying the CCR [c, c∗ ] = 1) coupled to a single external
electromagnetic bath; for a more detailed discussion, see, e.g., [5, Chaps. 5 and 7].
The essential ideas are summarized as follows:
(i) First compute the Heisenberg equation of both the system and the bath.
(ii) Make a Markov approximation on the system and bath coupling, which leads to
a dynamical equation for the system observables.
Let us begin with a description of the bath. It is composed of a continuum of modes
with annihilation operator bω , where ω denotes the frequency for that mode and
which is also used to label it. As will be shown later, the continuum of frequencies
will range from −∞ and ∞. Technically, negative frequencies are unphysical, but in
the model that we discuss here, typically arising in quantum optics and related fields,
negative frequencies arise through the application of a rotating-wave approximation
(RWA), which we assume has already been applied (we refer to [5] for the details).
The fields at distinct frequencies do not couple with one another and commute with
each other, satisfying the CCR
[bω , bω ] = 0, [bω , bω∗ ] = δ(ω − ω ), (1.16)
where δ(·) denotes the Dirac delta function. The Hamiltonian for the oscillator is
Hsys = ωo c∗ c, where ωo is the oscillator’s resonance frequency. The idealized bath
Hamiltonian is given by ∞
Hbath = ωbω∗ bω dω,
−∞
14 1 Introduction
indicating explicitly the modeling of the bath as a collection of independent infinitesi-

mal harmonic oscillators oscillating at distinct frequencies in a continuum of values.
Next, the system and the bath are assumed to be coupled through the following
interaction Hamiltonian:
∞
Hint = ı κ(ω) bω∗ c − c∗ bω dω.
−∞
Let us now write the Heisenberg equations of both the system and the bath. The bath
equation is given by
dbω (t)
= ı[Hsys + Hbath + Hint , bω (t)] = −ıωbω (t) + κ(ω)c(t),
dt
where (1.16) was used and bω (t) and c(t) are the Heisenberg picture evolution of bω
and c, respectively. This immediately yields the solution
t
−ıωt −ıωt
bω (t) = e bω + e eıωs κ(ω)c(s)ds. (1.17)
0
Likewise, an arbitrary system variable X (t) obeys
d X (t)
= ı[Hsys + Hint , X (t)]
dt ∞
= ı[Hsys , X (t)] + κ(ω) [c(t)∗ , X (t)]bω (t) − bω (t)∗ [c(t), X (t)] dω.
−∞
Then substituting (1.17) into this equation, we obtain the dynamical equation of
the system observables. But as readily seen, the dynamics is not Markov since the
time evolution of X (t) in the future [t, ∞) depends on the past history in [0, t].
Nevertheless, in several typical situations such as the case where the bath is a coherent
laser beam, we are allowed to make a valid approximation that converts the system to
√ In our case, this means that κ(ω) can be approximated to be constant,
a Markov one.
say κ(ω) = γ/2π; i.e., the system-bath coupling does not depend on the frequency.
∞
Then, due to the identity −∞ e−ıωt dω = 2πδ(t), the above system equation reduces
to
d X (t) √
= ı[Hsys , X (t)] + γ [c(t)∗ , X (t)]bin (t) − bin (t)∗ [c(t), X (t)]
dt
1 1
+ γ c(t)∗ X (t)c(t) − c(t)∗ c(t)X (t) − X (t)c(t)∗ c(t) . (1.18)
2 2
This is called a quantum Langevin equation, describing the (approximate) Markov
time evolution of the system. Here, we have defined the input field:
1.4 Open Quantum Systems, the Markov … 15
∞
1
bin (t) = √ e−ıωt bω dω. (1.19)
2π −∞
That is, bin (t) is the Fourier transform of bω . Note the following important commu-
tation relations:
[bin (s), bin (t)] = 0, [bin (s), bin (t)∗ ] = δ(s − t) ∀s, t ≥ 0.
Thus, bin (t) can be regarded as the quantum analogue of classical white noise. Hence
the input to the system (1.18) is precisely this operator-valued quantum white noise
process. In particular, under the Markov approximation, the interaction Hamiltonian
√
takes the form of a time-varying Hamiltonian, Hint (t) = ı γ(bin (t)∗ c − c∗ bin (t)).
Moreover, we can also immediately see that the optical field
∞
1
bout (t) = √ bω (t)dω (1.20)
2π −∞
satisfies the following equation (interpreted as a boundary condition):

√
bout (t) = γc(t) + bin (t). (1.21)
This is the portion of the optical light field bin (t) immediately after its interaction
with the system, which contains some information about the system as indicated by
the first term on the right-hand side of (1.21). Also, as will be shown in the examples,
the phase or intensity of bout (t) can be measured by using a photo detector. That is,
bout (t) represents the output field of the system, and thus, (1.21) is called the output
equation. For instance, if the input is a laser beam, then the output is the part of the
beam that departs from the oscillator after the former’s interaction with the latter.
We can write down a particular form of the Langevin equation when the system
has m input–output fields:
dX m
1 1
= ı[Hsys , X ] + L ∗j X L j − L ∗j L j X − X L ∗j L j
dt j=1
2 2

m
+ [L ∗j , X ]ξ j − ξ ∗j [L j , X ] , (1.22)
j=1
η j = L j + ξ j , ( j = 1, . . . , m), (1.23)
√
where γc, bin , and bout are now replaced by L, ξ j , and η j , respectively. The multiple-
input white noise processes ξ j and multiple-output white noise processes η j satisfy
[ξ j (s), ξk (t)∗ ] = δ jk δ(t − s) and [ξ j (s), ξk (t)] = 0, and [η j (s), ηk (t)∗ ] = δ jk δ(t −
s) and [η j (s), ηk (t)] = 0, respectively. Also note that the interaction Hamiltonian
between the system and the jth bath is now given by
16 1 Introduction
( j)
Hint (t) = ı(ξ j (t)∗ L − L ∗ ξ j (t)). (1.24)
Note that the quantum Langevin equation above does not represent the most general
Markov open quantum dynamics as it does not include the so-called gauge process
which will be introduced and discussed in Chap. 2.
Finally, we remark that justifying the Markov approximation and Langevin equa-
tion for general open quantum systems and coupling operators L is in general subtle.
A different set of physical assumptions from the ones we have described above for
√
the single model oscillator with L = γc, and a different approximating procedure,
may be required to justify employing a Markov approximation. We refer the reader
to [77, 78] and the references therein for the details.
1.5 Linear Dynamical Quantum Systems: Description

and Physical Examples
Armed with the background materials from the preceding section, we are now in
a position to introduce the class of linear dynamical quantum systems, with their
precise description and detailed mathematical modeling being deferred until the next
chapter. To make the exposition of this section more concrete, the introduction will be
followed by a discussion of several examples of physical systems that can be modeled
within this class: optical cavities, non-degenerate parametric amplifiers, degenerate
parametric amplifiers, opto-mechanical systems, and large atomic ensembles.
Linear dynamical quantum systems arise in practice as idealized models of a
collection of coupled distinct open quantum harmonic oscillators that are in a Markov
interaction with their surrounding heat baths (acting as a quantum noise source) with
the following properties:
(i) The system Hamiltonian is of a quadratic form, quadratic here being in the
position and momentum operators of the oscillators, such as Hsys = (q 2 + p 2 )/2.
(ii) The system’s canonical position and momentum operators are linearly coupled
to one or more external (quantum) heat baths, e.g., via the coupling operator
L = (q + ı p)/2.
This results in a linear quantum Langevin equation for the position and momentum
operators of the oscillators driven by stochastic heat baths. For instance, consider
the treatment in [34] of an atom trapped in an optical cavity. The atomic dipole is
strongly coupled to light inside the cavity that results in random mechanical forces
acting on the atom as it absorbs and emits light. When the parameter values are in
the right regime, the details of the dynamics of the optical and atomic dipole become
unimportant, and the optical field can be viewed as an environment for the atomic
motion. The observables relating to the motion of the trapped atom, its position and
momentum operators, can then be treated like those of an open quantum harmonic
oscillator linearly coupled to a bosonic bath. Linear Markov open quantum models
are particularly prominent in quantum optics. They are used to describe various
1.5 Linear Dynamical Quantum Systems: Description and Physical Examples 17
linear quantum optical devices and are also employed to describe counterparts of
these optical devices in other physical platforms such as superconducting microwave
circuits. In the following, we will present several concrete physical examples of
linear quantum systems, taken from quantum optics, nano-mechanics, and atom
optics; these are very active research fields playing important roles for both a purely
academic purpose, for understanding and testing quantum mechanics, as well as for
an applied purpose, the engineering of quantum information processing systems. In
each example, we write down the system equations without detailed derivation, but
connections to the general Langevin equation described in the previous section will
be discussed so that the physical background of the dynamics can be reasonably
conceived.
1.5.1 Optical Cavities
A simple open linear system is an empty optical cavity as illustrated in Fig. 1.2. The
dynamical behavior of this optical device is represented by only one variable; that
is, a standing optical wave with fixed resonance frequency ωo is generated inside the
cavity, and its quantized complex amplitude, i.e., the annihilation operator, aωo (t)
with “mode” ωo is the variable characterizing the system dynamics. We call this
single variable the cavity mode and in what follows will be denoted by a(t) (in the
Heisenberg picture). This mode satisfies the CCR [a(t), a(t)∗ ] = 1 for all t ≥ 0.
The oscillation dynamics of this optical field is described by the system Hamiltonian
Hsys = ωo a ∗ a, with ωo as before. Now, the meaning of “open” is as follows. As
shown in the figure, the cavity is composed of two partially reflecting mirrors and one
perfectly reflecting mirror. Through the partially reflecting mirrors, the intracavity
mode a(t) interacts with two external fields. Let ξ1 (t) and ξ2 (t) be the continuous-
mode annihilation process of those input optical fields, which satisfy [ξi (s), ξ j (t)∗ ] =
δi j δ(s − t). As described later, in this setting, ξ1 serves as a coherent laser field
with amplitude ξ1 (t) = β(t), while ξ2 is a vacuum field (i.e., a field containing
no photons, just vacuum fluctuations). The interaction Hamiltonian of a and ξ j at
( j) √
time t, which corresponds to (1.24), is now given by Hint = ı κ j (a ∗ ξ j − aξ ∗j ).
√
Therefore, L j = κ j a. Here, κ j represents the interaction strength of these two
optical fields; more precisely, it is given by κ j = cT j / with c the speed of light, T j
Fig. 1.2 Mode-cleaning Controller

cavity
a
Piezo
18 1 Introduction
the dimensionless transmissivity of the jth mirror, and the optical path length in
the cavity (hence, κ j has dimension of [1/s]). In this setting, the quantum Langevin
equation for a is obtained from (1.22) as
da κ1 + κ2 √ √
= −ıωo a − a − κ1 ξ1 − κ2 ξ2 .
dt 2
It is often convenient to move to a “rotating frame” where the new variables describing
the dynamics are taken to be ã(t) = eıω p t a(t) and ξ˜ j (t) = eıω p t ξ j (t), with ω p the
center frequency of the driving laser field. With this transformation, where we again
denote the new rotating frame variables by a(t) and ξ j (t) rather than ã(t) and ξ˜ j (t),
the above Langevin equation can be rewritten as
da κ1 + κ2 √ √
= −ıa − a − κ1 ξ1 − κ2 ξ2 . (1.25)
dt 2
The frequency difference = ωo − ω p is called the detuning. For high-quality quan-
tum optical experiments, should be tuned to be zero.
The external field operators also change due to the interaction; as shown in (1.23),
under the Markov approximation, the output fields are given by
√ √
η1 = κ1 a + ξ1 , η2 = κ2 a + ξ2 . (1.26)
As a consequence, it becomes apparent that the optical cavity can be modeled as a

linear dynamical system with input ξ j and output η j . Note that the above equations
are more rigorously described by a quantum stochastic differential equation (QSDE),
which in the Itō form is given by
κ 1 + κ2 √ √
da = − ı + adt − κ1 dA1 − κ2 dA2 ,
√ 2 √
dY1 = κ1 adt + dA1 , dY2 = κ2 adt + dA2 .
The QSDE will be discussed in more detail in the next chapter.

Typically, this system functions as an optical low-pass filter, see [79]. In real-
ity, any single-mode coherent laser field is not an ideal monochromatic source that
only contains a single center frequency, but it is composed of several (continuous)
modes of frequency; in our case, the amplitude of the coherent input field ξ1 , i.e.,
ξ1 (t) = β(t), is not an ideal (sinusoidal) function with frequency ω p . Since this
imperfection usually results in serious degradation of the laser’s performance in
quantum information processing, mode-cleaning of the field is important. The cav-
ity system considered here can actually erase unwanted optical frequencies; that is,
the noisy sideband of the function β(t) is suppressed and a mode-cleaned output
with amplitude η2 (t) appears. To see how this works, we first assume exact mode-
matching, meaning = ωo − ω p = 0. In practice, this condition can be achieved
by measuring η1 to detect the error signal, which is then fed back to a piezoactuator
mounted at one of the mirrors for locking the optical path length in the cavity; see
Fig. 1.2. Then, the resulting system dynamics is given by
da √ √ √
= −κa − κξ1 − κξ2 , η2 = κa + ξ2 ,
dt
where we set κ = κ1 = κ2 . Now, as in the classical case, for a linear quantum system
we can formally define the transfer function to examine the input–output relationship.
The transfer function from ξ1 to η2 is given by 1→2 [s] = −κ/(s + κ), and its gain
in the Fourier domain s = ıω (i.e., essentially, the power spectral density) is obtained
as
κ2
|1→2 [ıω]|2 = 2 .
ω + κ2
Thus, input components with higher frequencies are attenuated in the output field
(note that ω = 0 corresponds to the laser center frequency ω p since we are working
in the rotating frame). This specific type of cavity is known as a mode-cleaning cavity
(MCC) and is often used in quantum optics experiments [6, 79].
1.5.2 Non-degenerate Optical Parametric Amplifiers
We would now like to describe a type of linear optical parametric amplifier, called
a non-degenerate optical parametric amplifier (NOPA). This device is of significant
importance in the area of continuous-variable quantum information where the NOPA
is used to generate entangled beams of light. We here follow the standard treatment of
the NOPA, see, e.g., [80]. As depicted in Fig. 1.3, one realization of a NOPA is as an
optical device consisting of mirrors, forming an optical ring cavity, and a nonlinear
χ(2) optical crystal. The three mirrors M2 , M3 , and M4 are fully reflecting while the
mirror M1 is partially transmissive. An external coherent light is used as a pump
beam to supply quanta to the NOPA. Inside the nonlinear crystal, the pump beam
interacts with vacuum fields in a so-called parametric down conversion process to
generate two photons of the same frequency but of orthogonal polarization, known
as signal and idler photons (in other implementations, the two photons generated
can have the same polarization but different resonance frequencies). We will work
in the scenario where the pump beam is strongly coherent and assumed to remain
more or less undepleted during this interaction, so that it can be treated as a classical
field. Through the mirror M1 , the orthogonally polarized photons circulating inside
the cavity can escape to the outside. The escaping photons form a continuous-mode
Gaussian beam with correlations in the polarization modes that can be regarded as
Einstein-Podolsky-Rosen (EPR)-like entanglement [80].
The NOPA can be mathematically modeled as follows. We basically have two
cavity modes a1 and a2 with the same resonance frequency ωc but are orthogonally
polarized with respect to one another, so that they are distinguishable. They satisfy the
20 1 Introduction
M3 M4 Pumping field
a1 ξ1
a2 ξ2 η1
ξloss M2 M1
η2
Fig. 1.3 A non-degenerate optical parametric amplifier composed of a bow-tie type optical cavity
and a χ(2) nonlinear crystal implementing the parametric amplifier. The mirror M1 is partially
transmissive, while the other mirrors are perfectly reflective for the input–output field. Now the
input field is set to the vacuum. The internal optical loss is modeled by the fictitious vacuum input
with mode ξloss . The crystal is pumped by a strong classical field, which is perfectly transmissive
for the mirrors M3 and M4 . Reprinted with permission from [44] © 2012 IEEE
CCR [ai , a ∗j ] = δi j and [ai , a j ] = 0 and are coupled by an interaction Hamiltonian

Hsys = ı( /2) e−ıωp t a1∗ a2∗ − eıωp t a1 a2 , where is an effective amplitude, related to
the amplitude of the pump beam and the χ(2) coefficient of the nonlinear crystal,
and ω p is the frequency of the pump beam. A stand-alone NOPA has four inputs
(continuous-mode or broadband) fields, all in the vacuum state [5]. In terms of the
Langevin equation (1.22), the mode a1 is coupled to the incoming fields ξ1 and ξ2 by
√ √
the coupling operators L 1 = γa1 and L 2 = γa2 , respectively, where γ denotes
the coupling strength. Furthermore, to model optical losses, the mode is also coupled
√
√vacuum input fields ξloss,1 and ξloss,2 with coupling operators L 3 = κa1
to additional
and L 4 = κa2 , where κ denotes the loss rate. These fields carry photons away
which then become completely lost. We consider the matched case where ωc = ωp /2.
From (1.22), the quantum Langevin equation of the NOPA modes a1 and a2 in the
Heisenberg picture is given by
da1 γ+κ √ √
= −ıωc a1 − a1 + e−2iω p t a2∗ − γξ1 − κξloss,1 ,
dt 2 2
da2 γ+κ √ √
= −ıωc a2 − a2 + a1∗ − γξ2 − κξloss,2 . (1.27)
dt 2 2
Also from (1.23), we obtain the output equations:
√ √
η1 = γa1 + ξ1 , η2 = γa2 + ξ2 . (1.28)
The equations above are time-dependent due to the dependence on the oscillating term
eıωp t . Typically, the equation is transformed to a time-independent one by moving to
rotating frame at half the pump frequency ωp /2. In this rotating frame, one makes the
substitutions a j (t)e−ıω p t/2 → a j (t), ξ j (t)eıω p t/2 → ξ j (t), and η j (t)eıω p t/2 → η j (t)
for j = 1, 2 and all t. After the substitution is made, the equations become time-
independent,
da1 γ+κ √ √
=− a1 + a2∗ − γξ1 − κξloss,1 ,
dt 2 2
da2 γ+κ √ √
=− a2 + a1∗ − γξ2 − κξloss,2 , (1.29)
dt 2 2
√ √
η1 = γa1 + ξ1 , η2 = γa2 + ξ2 . (1.30)
It must be kept in mind that all quantities are now rotating frame quantities although
we have retained the same notation a j , ξ j , η j as in the original (non-rotating frame)
equation. Hence, as with the MCC from the previous example, the NOPA can be
modeled as a linear time-invariant dynamical quantum system. But an important
difference from the case of the MCC is that the equations are now expressed in terms
of both annihilation and creation operators, a j and a ∗j . To analyze the NOPA, it is
convenient to transform it into quadrature form (to be discussed in more detail in
Chap. 2) by defining the following vector of operators:
z = (q1 , p1 , q2 , p2 ) ,
ξ = (ξ1 , ξ1 , ξ2 , ξ2 , ξloss,1 , ξloss,1 , ξloss,2 , ξloss,2 ) ,
q p q p q p q p
y = (η1 , η1 , η2 , η2 ) ,
q p q p
where q j = a j + a ∗j and p j = (a j − a ∗j )/ı. Likewise, for the noise operators we

have defined ξ j = ξ j + ξ ∗j and ξ j = (ξ j − ξ ∗j )/ı, etc. The dynamics of the NOPA
q p
can then be expressed as
dz
= Az + Bξ, y = C z + Dξ,
dt
where
⎡ ⎤
−(γ + κ)/2 0 /2 0
⎢ 0 −(γ + κ)/2 0 − /2 ⎥
A =⎢⎣
⎥,
⎦
/2 0 −(γ + κ)/2 0
0 − /2 0 −(γ + κ)/2
√ √
B = −[ γ I4 κI4 ],
√
C = γ I4 ,
D = [ I4 04 ].
In the operation of a NOPA, we require that the matrix A be Hurwitz (all its eigen-
values lie in the left half plane) so that the system is stable. Stable here is in the sense
that the average number of photons contained in the modes a1 and a2 remains finite
at all times, corresponding to the NOPA operating in what is known as the below
threshold regime.
As alluded to earlier, the type of entanglement produced by a NOPA is a so-
called EPR entanglement, which is characterized by two-mode squeezing between
22 1 Introduction
the pair of output fields η1 and η2 . A two-mode squeezed state between a pair of
continuous-mode output fields such as η1 and η2 is characterized by correlation
and anticorrelation between quadratures of the two fields. As we will see below, in
p p
the model of the NOPA above this is correlation between quadratures η1 and η1 ,
q q
and anticorrelation between η1 and η1 . The amount of correlations, or two-mode
squeezing, between the two beams can be quantified in the frequency domain via
some squeezing spectral densities that will be introduced next.
q q p p
Let η̃ q (t) = η1 (t) + η2 (t) and η̃ p (t) = η1 (t) − η2 (t). Also, let ˜ q [ıω],
˜ [ıω], Z [ıω], and [ıω], respectively, be the Fourier transforms of η̃ (t), η̃ p (t),
p q
∞
z(t), and ξ(t), defined via the transform relation F [ıω] = √12π −∞ f (t) e−ıωt dt,
with F the Fourier transform of f . Therefore,
˜ q [ıω] = C1 Z [ıω] + D1 [ıω] ,

˜ p [ıω] = C2 Z [ıω] + D2 [ıω] ,

where C1 = [1 0 1 0]C, C2 = [0 1 0 −1]C, D1 = [1 0 1 0]D, and D2 = [0 1 0 −1]D.

The two-mode squeezing spectral densities V+ and V− can be defined through the
following identities:
˜ q [ıω]∗
˜ q [ıω ] = V+ (ıω)δ(ω − ω ),
˜ p [ıω]∗
˜ p [ıω ] = V− (ıω)δ(ω − ω ), (1.31)
where · denotes quantum expectation with respect to the initial state of the oscillator
modes and the vacuum state of the incoming fields. The quantity V+ is called the
two-mode amplitude squeezing spectrum and V− the two-mode phase squeezing
spectrum. The two outgoing fields η1 and η2 are entangled at frequency ω rad/s if,
see, e.g., [81],
V = V+ (ıω) + V− (ıω) < 4. (1.32)
A perfect maximally entangled continuous-mode EPR state would have V+ (ıω) =

V− (ıω) = V+ (ıω) + V− (ıω) = 0 for all ω. However, this ideal EPR state is not
physically achievable as it would require an infinite amount of energy to produce
and no losses can be permitted. Thus, in practical devices the goal is to make
V+ (ıω) and V− (ıω) small over a frequency range (bandwidth) such that one has
V+ (ıω) + V− (ıω) < 4 in this frequency range. This yields EPR-like states of the
outgoing fields that are practical approximations of the ideal continuous-mode EPR
state. For the model considered here, the symmetry in the parameter values implies
that V+ (ıω) = V− (ıω), so entanglement at frequency ω is verified by V± (ıω) < 2.
Moreover, for low nonnegative values of ω (note that since we are working in a rotat-
ing frame, the actually frequency is around half the pump frequency, ω p /2) we have
that, to an excellent approximation, V+ (ıω) ≈ V+ (0) and V− (ıω) ≈ V− (0). There-
fore, V+ (0) = V− (0) = V± (0) can be employed as a figure of merit for the degree
of two-mode squeezing or entanglement between the output fields η1 and η2 .
Define H j [ıω] = C j (ıω I − A)−1 B + D j ( j = 1, 2), then it is straightforward

to show that

V+ (ıω) = Tr H1 [ıω]∗ H1 [ıω] , (1.33)

V− (ıω) = Tr H2 [ıω]∗ H2 [ıω] . (1.34)
When κ = 0, we have that [80]
2(1 − k)2
V± (0) = , (1.35)
(1 + k)2
with k = /γ taking a value 0 ≤ k < 1 in the typical below threshold operation.2

Note that (1.32) is a sufficient condition for EPR entanglement, with the two fields
squeezed in amplitude and phase quadratures. However, in general, they may instead
be squeezed in other quadratures. Hence, we now give the following general notion
ψ ψ
of EPR entanglement. Let η1 1 = eıψ1 η1 and η2 2 = eıψ2 η2 with ψ1 , ψ2 ∈ (−π, π].
ψ ψ
Denote the corresponding two-mode squeezing spectra between η1 1 and η2 2 as
ψ ,ψ
V± 1 2 (ıω, ψ1 , ψ2 ), following the definition of V± given above but making the sub-
ψ
stitutions η j j → η j , j = 1, 2.
Definition 1.1 The fields η1 and η2 are EPR entangled at the frequency ω rad/s if
∃ ψ1 , ψ2 ∈ (−π, π] such that
ψ ,ψ2 ψ ,ψ2
V+ 1 (ıω, ψ1 , ψ2 ) + V− 1 (ıω, ψ1 , ψ2 ) < 4. (1.36)
EPR entanglement is said to vanish at ω if there are no values of ψ1 and ψ2 satisfying

the above criterion. Unless otherwise specified, EPR entanglement will refer to the
case with ψ1 = ψ2 = 0.
1.5.3 Degenerate Parametric Amplifiers/Optical

Parametric Oscillators
Mathematically, the degenerate parametric amplifier (DPA) or optical parametric

oscillator (OPO) is just a special case of the non-degenerate one from the preceding
section, in the sense that the two modes a1 and a2 become “degenerate” as depicted
in Fig. 1.4; see [5, 57]. In this case, the two modes are identical in all respects (fre-
quency and polarization) and cannot be distinguished. As a result, again expressing
all quantities in a rotating frame at half the pump frequency as inthe NOPA, the
system Hamiltonian inside the cavity becomes Hsys = ı a ∗2 − a 2 /2, where a is
q q p p
2 Notethat [80] and the formula (1.35) take ξout,a − ξout,b and ξout,a + ξout,b as the squeezed two-
q q p p
mode quadratures. To obtain squeezing in the quadratures ξout,a + ξout,b and ξout,a − ξout,b , one
need only take < 0 and the formula (1.35) is modified to be V± (0) = 2(1 + k) /(1 − k)2 .
2
24 1 Introduction
Fig. 1.4 Optical parametric M3 M4 Pumping field

oscillator. Reprinted with
permission from [44] © 2012
IEEE a ξ
ξloss M2 M1
η
the cavity mode. As in the NOPA case, a couples with the external vacuum field ξ
√
through the partially reflected mirror M1 with coupling constant γ; thus, L = γa.
Consequently, the system’s dynamical equation is given by
da γ+κ √ √ √
=− a + a ∗ − γξ − κξloss , η = γa + ξ, (1.37)
dt 2 2
where again the fictitious vacuum field ξloss is added to model the internal optical
loss. The above equation contains both a and a ∗ ; hence, as before it is sometimes
more convenient to work with the quadrature form equation in terms of the amplitude
quadrature q = a + a ∗ and phase quadrature p = (a − a ∗ )/ı:
q
d q 1 −γ + 0 q √ 10 ξ
= − γ ,
dt p 2 0 −γ − p 01 ξp
q √ q
η γ 0 q ξ
= √ + p ,
ηp 0 γ p ξ
where ξloss is omitted for simplicity. Like the NOPA, the DPA also plays a very
important role in quantum optics and information [6, 82]; in the laboratory, this
device is utilized as a generator of a squeezed Gaussian optical field.
Let us briefly describe how the DPA produces a squeezed Gaussian optical field;
a precise treatment of Gaussian fields will be given in the next chapter. First, we
mention that although the input is vacuum, the output is not a vacuum field (but
does have zero amplitude). The non-trivial property of the output field can be easily
seen by examining the transfer function matrix from the input (ξ q , ξ p ) to the output
(η q , η p ):
s − (γ + )/2 s − (γ − )/2
[s] = diag , .
s + (γ − )/2 s + (γ + )/2
Now, the input is simply the vacuum field; hence, the power in each quadrature
component is |ξ q [ıω]|2 = |ξ p [ıω]|2 = 1; this is called the quantum noise limit
(QNL), which corresponds to a scenario in which quantum fluctuations are equally
distributed in the quadratures q and p, while keeping the minimum uncertainty
relation, |ξ q [ıω]|2 |ξ p [ıω]|2 = 1. (Note that in the Fourier domain, the Heisenberg
uncertainty relation is represented by |ξ q [ıω]|2 |ξ p [ıω]|2 ≥ 1.) On the other hand,
the power of the output fields can be calculated to be
Fig. 1.5 Phase-space picture

for a a vacuum state and b a
(a) p (b) p
squeezed vacuum state q q
ω 2 + (γ + )2 /4 ω 2 + (γ − )2 /4
|η q [ıω]|2 = , |η p [ıω]|2 = .
ω 2 + (γ − )2 /4 ω 2 + (γ + )2 /4
They satisfy the relation |η q [ıω]|2 |η p [ıω]|2 = 1 for all ω, so that the output field
is also in a minimum uncertainty state. But importantly, the quadratures have different
fluctuation variances; for instance when = γ/2, then at the center frequency ω = 0
they are given by |η q |2 = 9 and |η p |2 = 1/9. That is, the fluctuation of the p-
quadrature is reduced below the QNL, while the q-quadrature fluctuates above the
QNL. Hence, the output field is “squeezed” in that there are non-uniform fluctuations
in η q and η p , as depicted in Fig. 1.5b. Squeezed light fields are important in various
situations in quantum mechanics and its applications; in particular, they are used for
generating entangled light beams and for high-precision detection of a tiny force
such as in gravitational-wave interferometers [19, 22].
1.5.4 Opto-mechanical Systems
Nano-mechanical devices are also promising platforms for testing quantum mechan-
ics as well as carrying out quantum information processing. A typical one is an
opto-mechanical oscillator as depicted in Fig. 1.6. This system is composed of a
two-sided optical cavity wherein one of these mirrors is movable (not fixed); see [21,
23, 83, 84]. Thus, the moving mirror can serve as a mechanical oscillator. Let q1 and
p1 be the oscillator’s (i.e., the moving mirror’s) position and momentum operators
satisfying the CCR [q1 , p1 ] = 2ı. Also define q2 = a2 + a2∗ and p2 = (a2 − a2∗ )/ı
with a2 the annihilation operator of the cavity mode. The oscillator is in a harmonic
potential with Hamiltonian Hhar = mωo2 q12 /2 + p12 /2m where m and ωo represent
the mass and the resonant frequency of the oscillator, respectively. The oscillator
and the intracavity field interact with one another through a radiation pressure force,
Fig. 1.6 Opto-mechanical q1, p1

oscillator. Figure adapted Circulator
from [25] a2
ξ
f
η
26 1 Introduction
described by the Hamiltonian Hrp = γ q1 a2∗ a2 . Roughly speaking, the oscillator is

pushed by an optical force whose strength is proportional to the photon number inside
the cavity. This Hamiltonian can be approximated, around the equilibrium point of
the light intensity in the cavity by Hrp = γ q1 (α∗ a2 + αa2∗ ) where α ∈ C denotes
the mean amplitude of the cavity mode, i.e., α = a2 . This approximation is valid
when |α| is much larger than the motion of q1 . Moreover, the cavity field couples
to the incoming optical field ξ(t) through the other√partially reflective fixed mirror,
via the singular interaction Hamiltonian Hint = ı κ(a2∗ ξ − a2 ξ ∗ ). The dynamical
equation of the composite system √ is given by the quantum Langevin equation (1.22)
with Hsys = Hhar + Hrp and L = κa2 , leading to
⎡ ⎤ ⎡ ⎤⎡ ⎤
q1 0 1/m 0 0 q1
d ⎢ p ⎥
⎢ 1⎥ = ⎢
⎢ −mω 2
0 γ 0 ⎥ ⎢
⎥ ⎢ p1 ⎥
⎥
⎣
dt q2 ⎦ ⎣ 0 0 −κ/2 0 ⎦ ⎣ q2 ⎦
p2 γ 0 0 −κ/2 p2
⎡ ⎤ ⎡ ⎤
00 0
√ ⎢ 0 0 ⎥ ξq ⎢1⎥
− κ⎣ ⎢ ⎥ ⎢
+⎣ ⎥ f, (1.38)
1 0⎦ ξp 0⎦
01 0
q √ q
η κ √0 q2 ξ
= + p (1.39)
ηp 0 κ p2 ξ
where (ξ q , ξ p ) and (η q , η p ) are the quadratures of the input field mode ξ and the
output field mode η, respectively. Also, we have assumed α ∈ R and set γ = γ α.
Note that, because of the uncertainty principle, both η q and η p cannot be measured
simultaneously; quantum mechanics allows us to measure only one of them.
In the above system dynamics, we have added a classical force f that acts only on
the oscillator. Typically, f is a very tiny unknown force such as a gravitational-wave
force, and this opto-mechanical oscillator system can act as a sensor for estimating
f . What should be emphasized here is the fact that systems and control theory offers
a number of powerful tools for general dynamical estimation problems or, more
broadly, quantum metrology. This fact is indeed applicable to the above sensing
problem for f . In Sect. 6.3 we will see a feedback control method to construct an
effective sensor.
Before closing this section, we want to demonstrate that a simplified model of
Eqs. (1.38) and (1.39) can be obtained via a technique known in physics as adiabatic
elimination and in mathematics and engineering as singular perturbation; this is
indeed an important topic and will be encountered again in Chap. 3. Here, we only
provide a very rough heuristic treatment that is commonly employed in the physics
literature. The idea is based on the fact that the cavity dynamics is much faster than
that of the oscillator. Hence, one imagines that the cavity state goes immediately to
equilibrium and can thus be eliminated from the√dynamical equation. To proceed, √
we set q̇2 = ṗ2 = 0, which yields q2 = −2ξ q / κ and p2 = 2γq1 /κ − 2ξ p / κ.
Substituting them into (1.38), we end up with
√
d q1 0 1/m q1 0 q 0
= − λ ξ + f, (1.40)
dt p1 −mω 2 0 p1 1 1
where λ = 4γ 2 /κ. Also substitution into (1.39) yields

q
ηq 0 0 q1 ξ
= √ − p .
ηp λ0 p1 ξ
Since η q clearly does not contain any information about the oscillator and, accord-
ingly, the unknown force f , we should measure η p . Hence, the output equation
describing the informative measurement result is given by
√
q
y = [ λ 0] 1 − ξ p . (1.41)
p1
Equations (1.40) and (1.41) are quantum versions of the classical open harmonic
oscillator (1.2). Note that (1.2) is valid even without disturbance noise. Also recall
now that (1.14) has the same form, but it is a closed system having no associated
real-time output signal. In the quantum case, the coupling of the cavity mode to the
external optical field turns the system into an open one, in which information can be
extracted from the system with the optical field serving as a quantum probe.
1.5.5 Large Atomic Ensemble
The last example is taken from the field of atom optics. Let us consider a large atomic
ensemble trapped in a cavity that is coupled to an external optical field as illustrated
in Fig. 1.7. In an ultimately cooled cavity, as a good approximation we can treat each
atom as an ideal “qubit” composed of two energy levels: the ground and excited
states; the observable representing the energy of the kth atom is given by

1 1 0
σz(k) = .
2 0 −1
Trapped
Probe atoms
laser Detector
PBS ηq
Fig. 1.7 Atomic ensemble trapped in a cavity. PBS indicates a polarized beam splitter, which is
used for detecting the polarization of the output field. Figure adapted from [25]
28 1 Introduction
Hence, measuring the energy gives the result +1/2 or −1/2, corresponding to
“excited” and “ground” states, respectively. Besides σz(k) there are also two other
especially important observables that do not commute with σz(k) ,

1 01 1 0 −ı
σx(k) = (k)
, σy = .
2 10 2 ı 0
Here, the observables of interest are those of a collection of the above two-state
system, in particular the total energy of the atomic ensemble, which is mathematically
represented by

N
Jz = σz(k) = (σz(1) ⊗ I ⊗ · · · I ) + · · · + (I ⊗ I ⊗ · · · σz(N ) ), (1.42)
k=1
where I denotes the 2 × 2 identity matrix and N is the number of atoms. Likewise,
we define
N N
Jx = σx(k) , Jy = σ (k)
y . (1.43)
k=1 k=1
A notable fact is that they satisfy the spin CCR [Ji , J j ] = ı i jk Jk , where i jk denotes
the Levi-Civita symbol.
The particular system configuration we are interested in here is one that facilitates
quantum non-demolition (QND) measurement of Jz ; see e.g., [85, 86]. This atomic
ensemble couples with the optical field, having a polarization degree of freedom,
via the Faraday interaction. Through this interaction, the polarization of the output
field rotates depending on the energy level, but ideally the observable representing
the energy level does not change; as a result, Jz can be measured without being
disturbed, and this is the essence of QND measurement of Jz . In terms √ of the Langevin
equation (1.22), this interaction is described by the operator L = M Jz , from which
the dynamics of the system’s observables (Jx , Jy , Jz ) is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
J −M/2 0 −b Jx √ −Jy
d ⎣ x⎦ ⎣
Jy = 0 −M/2 0 ⎦ ⎣ Jy ⎦ + M ⎣ Jx ⎦ ξ p , (1.44)
dt J b 0 0 Jz 0
z
where ξ p is the phase quadrature of the input field’s white noise process, and M
represents the coupling strength. Further, we assume here that the atomic ensemble
is subjected to a magnetic field, which corresponds to a term of the form Hsys =
b Jy in the system Hamiltonian, where b denotes the magnetic field strength. The
output equation is also obtained, which now has a non-trivial component only in the
amplitude quadrature: √
y = η q = 2 M Jz + ξ q . (1.45)
In particular, when b = 0, Jz does not change while Jz still appears in the output
equation. This indicates that we can indeed extract some information about Jz without
disturbing it; such an observable is in general referred to as a QND observable.
A typical problem setting is that the magnetic field is very tiny and its strength
b is unknown; we are interested in estimating this tiny magnetic field [17]. This
magnetometry problem is also an important parameter estimation problem, like the
force estimation problem described in Sect. 1.5.4. To solve this problem, the system
theoretic approach based on the linearized model is again very powerful. Now note
that, when M is small, the system variables obey a skew-Hermitian dynamics, imply-
ing that they preserve Jx2 + Jy2 + Jz2 . Hence, in the large ensemble limit N 1 and
over a short time period, their dynamics are constrained in the tangent space of the
manifold of the dynamics. This is the situation where we are allowed to linearize
the system dynamics. In particular, let us assume that the system is nearly along the
x-axis; thus, we set Jx to be a constant rather than a quantum observable. Then due to
Jx = J = N /2, the only dynamics that is important to us is J˙z = b J . In particular, if
we know that b is a classical random variable obeying a certain stochastic dynamics,
the problem boils down to one that can be √attacked with a system theoretic approach.
Let us assume that b obeys ḃ = −γb + σξ b , where γ is the dissipation rate and ξ b
is a standard Gaussian white noise (ξ q and ξ b are independent); then, as a result the
system’s dynamics is described by the following linear equation:

d Jz 0 J Jz 0 √
= + √ ξ b , y = 2 M Jz + ξ q . (1.46)
dt b 0 −γ b σ
Thus, clearly we can gain access to the unknown value b indirectly through the
dynamics of Jz . A specific scheme for estimating b was established in [17].
Finally, we remark that, although the above discussion has been focused on an
atomic ensemble system in a dispersive regime, another important linear approx-
imation
√ arises in a dissipative regime, where the coupling operator is given by
L = √ M(Jy + ı Jz ). In this case, L is approximated by an annihilation operator
L = Ma, which as a result yields a strictly stable linear system of the same form
as the cavity dynamics discussed in Sects. 1.5.1–1.5.3. (Note that the linear approxi-
mated system considered here, (1.46), is marginally stable.) A typical application of
such a strictly stable linear system is for quantum memories; that is, we are interested
in transferring a quantum state encoded in an input optical field to the atomic state;
see [9, 12–14, 37, 38].
References
1. V.P. Belavkin, S.C. Edwards, Quantum filtering and optimal control, in Quantum Stochastics
and Information: Statistics, Filtering and Control (University of Nottingham, UK, 15–22 July
2006), ed. by V.P. Belavkin, M. Guta (World Scientific, Singapore, 2008), pp. 143–205
30 1 Introduction
2. M.R. James, H.I. Nurdin, I.R. Petersen, H ∞ control of linear quantum stochastic systems.
IEEE Trans. Autom. Control 53(8), 1787–1803 (2008)
3. H.I. Nurdin, M.R. James, A.C. Doherty, Network synthesis of linear dynamical quantum sto-
chastic systems. SIAM J. Control Optim. 48(4), 2686–2718 (2009)
4. J.E. Gough, M.R. James, H.I. Nurdin, Squeezing components in linear quantum feedback
networks. Phys. Rev. A 81, 023804 (2010)
5. C.W. Gardiner, P. Zoller, Quantum Noise: A Handbook of Markovian and Non-Markovian
Quantum Stochastic Methods with Applications to Quantum Optics, 3rd edn. (Springer, Berlin,
2004)
6. A. Furusawa, P. van Loock, Quantum Teleportation and Entanglement: A Hybrid Approach to
Optical Quantum Information Processing (Wiley-VCH, Berlin, 2011)
7. R. Hamerly, H. Mabuchi, Advantages of coherent feedback for cooling quantum oscillators.
Phys. Rev. Lett. 109, 173602 (2012)
8. J. Kerckhoff, R.W. Andrews, H.S. Ku, W.F. Kindel, K. Cicak, R.W. Simmonds, K.W. Lehnert,
Tunable coupling to a mechanical oscillator circuit using a coherent feedback network. Phys.
Rev. X 3, 021013 (2013)
9. M. Hush, A.R.R. Carvalho, M. Hedges, M.R. James, Analysis of the operation of gradient echo
memories using a quantum input-output model. New J. Phys. 15, 085020 (2013)
10. A.A. Clerk, M.H. Devoret, S.M. Girvin, F. Marquardt, R.J. Schoelkopf, Introduction to quantum
noise, measurement, and amplification. Rev. Mod. Phys. 82, 1155 (2010)
11. N. Bergeal, F. Schackert, M. Metcalfe, R. Vijay, V. Manucharyan, L. Frunzio, D.E. Prober,
R.J. Schoelkopf, S.M. Girvin, M.H. Devoret, Phase-preserving amplification near the quantum
limit with a Josephson ring modulator. Nature 465, 64–68 (2010)
12. Q.Y. He, M.D. Reid, E. Giacobino, J. Cviklinski, P.D. Drummond, Dynamical oscillator-cavity
model for quantum memories. Phys. Rev. A 79, 022310 (2009)
13. I. Novikov, A.V. Gorshkov, D.F. Phillips, A.S. Sorensen, M.D. Lukin, R.L. Walsworth, Optimal
control of light pulse storage and retrieval. Phys. Rev. Lett. 98, 243602 (2007)
14. A.V. Gorshkov, A. Andre, M.D. Lukin, A.S. Sorensen, Photon storage in lambda-type optically
dense atomic media I. Cavity model. Phys. Rev. A 76, 033804 (2007)
15. W.L. Brogan, Modern Control Theory, 3rd edn. (Prentice-Hall, Upper Saddle River, 1991)
16. K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control (Prentice-Hall, Upper Saddle
River, 1995)
17. J.K. Stockton, J. Geremia, A.C. Doherty, H. Mabuchi, Robust quantum parameter estimation:
coherent magnetometry with feedback. Phys. Rev. A 69, 032109 (2004)
18. M. Guta, N. Yamamoto, System identification for passive linear quantum systems. IEEE Trans.
Autom. Control 61(4), 921–936 (2016)
19. V.B. Braginsky, F.Y. Khalili, Quantum Measurement (Cambridge University Press, Cambridge,
1992)
20. C.M. Caves, K.S. Thorne, R.W.P. Drever, V.D. Sandberg, M. Zimmermann, On the measure-
ment of a weak classical force coupled to a quantum mechanical oscillator. I. Issues of principle.
Rev. Mod. Phys. 52, 341–392 (1980)
21. M. Tsang, C.M. Caves, Coherent quantum-noise cancellation for optomechanical sensors. Phys.
Rev. Lett. 105, 123601 (2010)
22. H. Miao, Exploring Macroscopic Quantum Mechanics in Optomechanical Devices (Springer,
Berlin, 2012)
23. Y. Chen, Macroscopic quantum mechanics: theory and experimental concepts of optomechan-
ics. J. Phys. B: At. Mol. Opt. Phys. 46, 104001 (2013)
24. H.I. Nurdin, M.R. James, I.R. Petersen, Coherent quantum LQG control. Automatica 45, 1837–
1846 (2009)
25. N. Yamamoto, Coherent versus measurement feedback: linear systems theory for quantum
information. Phys. Rev. X 4, 041029 (2014)
26. Y. Yokotera, N. Yamamoto, Geometric control theory for quantum back-action evasion. EPJ
Quantum Technol. 3(15), 1–22 (2016)
References 31
27. H.I. Nurdin, J.E. Gough, Modular quantum memories using passive linear optics and coherent
feedback. Quantum Inf. Comput. 15(11–12), 1017–1040 (2015)
28. G.J. Milburn, Coherent control of single photon states. Eur. Phys. J. 159, 113–117 (2008)
29. G. Zhang, M.R. James, On the response of quantum linear systems to single photon input fields.
30. G. Zhang, Analysis of quantum linear systems’ response to multi-photon states. Automatica
50(2), 442–451 (2014)
31. K. Koga, N. Yamamoto, Dissipation-induced pure Gaussian state. Phys. Rev. A 85, 022103
(2012)
32. N. Yamamoto, Pure Gaussian state generation via dissipation: a quantum stochastic differential
equation approach. Philos. Trans. R. Soc. A 370, 5324–5337 (2012)
33. N.C. Menicucci, P. van Loock, M. Gu, C. Weedbrook, T.C. Ralph, M.A. Nielsen, Universal
quantum computation with continuous-variable cluster states. Phys. Rev. Lett. 97, 110501
(2006)
34. A.C. Doherty, K. Jacobs, Feedback-control of quantum systems using continuous state-
estimation. Phys. Rev. A 60, 2700 (1999)
35. N. Yamamoto, H.I. Nurdin, M.R. James, I.R. Petersen, Avoiding entanglement sudden-death
via feedback control in a quantum network. Phys. Rev. A 78, 042339 (2008)
36. H.I. Nurdin, N. Yamamoto, Distributed entanglement generation between continuous-mode
Gaussian fields with measurement-feedback enhancement. Phys. Rev. A 86, 022337 (2012)
37. J. Yoshikawa, K. Makino, S. Kurata, P. van Loock, A. Furusawa, Creation, storage, and on-
demand release of optical quantum states with a negative Wigner function. Phys. Rev. X 3,
041028 (2013)
38. N. Yamamoto, M.R. James, Zero dynamics principle for perfect quantum memory in linear
networks. New J. Phys. 16, 073032 (2014)
39. H. Mabuchi, Coherent-feedback quantum control with a dynamic compensator. Phys. Rev. A
78, 032323 (2008)
40. M. Sarovar, D.B.S. Soh, J. Cox, C. Brif, C.T. DeRose, R. Camacho, P. Davids, Silicon nanopho-
tonics for scalable quantum coherent feedback networks. EPJ Quantum Technol. 3(14), 1–18
(2016)
41. M. Yanagisawa, H. Kimura, Transfer function approach to quantum control-part i: dynamics
of quantum feedback systems. IEEE Trans. Autom. Control 48(12), 2107–2120 (2003)
42. M. Yanagisawa, H. Kimura, Transfer function approach to quantum control - part ii: control
concepts and applications. IEEE Trans. Autom. Control 48(12), 2121–2132 (2003)
43. J.E. Gough, S. Wildfeuer, Enhancement of field squeezing using coherent feedback. Phys. Rev.
A 80, 042107 (2009)
44. S. Iida, M. Yukawa, H. Yonezawa, N. Yamamoto, A. Furusawa, Experimental demonstration
of coherent feedback control on optical field squeezing. IEEE Trans. Autom. Control 57(8),
2045–2050. Reprinted, with permission, © 2012 IEEE (2012)
45. O. Crisafulli, N. Tezak, D.B.S. Soh, M.A. Armen, H. Mabuchi, Squeezed light in an optical para-
metric amplifier oscillator network with coherent feedback quantum control. Opt. Express 21,
18371–18386 (2013)
46. R.G. Beausoleil, P.J. Keukes, G.S. Snider, S.-Y. Wang, R.S. Williams, Nanoelectronic and
nanophotonic interconnect. Proc. IEEE 96, 230–247 (2007)
47. H. Mabuchi, Nonlinear interferometry approach to photonic sequential logic. Appl. Phys. Lett.
99, 153103 (2011)
48. H.M. Wiseman, G.J. Milburn, Quantum Measurement and Control (Cambridge University
49. T.J. Tarn, G. Huang, J.W. Clark, Modelling of quantum mechanical control systems. Math.
Model. 1(1), 109–121 (1980)
50. G.M. Huang, T.J. Tarn, J.W. Clark, On the controllability of quantum-mechanical systems. J.
Math. Phys. 24(11), 2608–2618 (1983)
51. D. D’Alessandro, Introduction to Quantum Dynamics and Control. Applied Mathematics and
Nonlinear Science Series (Chapman & Hall/CRC, London, 2008)
32 1 Introduction
52. E. Davies, Quantum Theory of Open Systems (Academic Press, New York, 1976)
53. V.P. Belavkin, Optimal measurement and control in quantum dynamical systems. Institute of
Physics, Nicolaus Copernicus University, Torun, preprint 411 (1979)
54. V. Belavkin, On the theory of controlling observable quantum systems. Autom. Remote Control
44(2), 178–188 (1983)
55. V.P. Belavkin, Nondemolition measurements, nonlinear filtering, and dynamic programming of
quantum stochastic processes, in Modelling and Control of Systems in Engineering, Quantum
Mechanics, Economics, and Biosciences, ed. by A. Blaquiere (Springer, New York, 1988), pp.
245–265
56. V.P. Belavkin, Continuous non-demolition observation, quantum filtering and optimal estima-
tion, Quantum Aspects of Optical Communication, vol. 45, Lecture Notes in Physics (Springer,
Berlin, 1991), pp. 151–163
57. M.J. Collett, C.W. Gardiner, Squeezing of intracavity and traveling-wave light fields produced
in parametric amplification. Phys. Rev. A 30(3), 1386–1391 (1984)
58. C. Gardiner, M. Collett, Input and output in damped quantum systems: quantum stochastic
differential equations and the master equation. Phys. Rev. A 31, 3761–3774 (1985)
59. R.L. Hudson, K.R. Parthasarathy, Quantum Ito’s formula and stochastic evolution. Commun.
Math. Phys. 93, 301–323 (1984)
60. L. Bouten, R. van Handel, M.R. James, An introduction to quantum filtering. SIAM J. Control
Optim. 46, 2199–2241 (2007)
61. J. Dalibard, Y. Castin, K. Mölmer, Wave-function approach to dissipative processes in quantum
optics. Phys. Rev. Lett. 68(5), 580–583 (1992)
62. R. Dum, P. Zoller, H. Ritsch, Monte Carlo simulation of the atomic master equation for spon-
taneous emission. Phys. Rev. A 45(7), 4879–4887 (1992)
63. H. Carmichael, An Open Systems Approach to Quantum Optics (Springer, Berlin, 1993)
64. M.A. Nielsen, I.L. Chuang, Quantum Computation and Quantum Information (Cambridge
University Press, Cambridge, 2000)
65. J.P. Dowling, G.J. Milburn, Quantum technology: the second quantum revolution. Philos. Trans.
R. Soc. Lond. A 361, 1655–1674 (2003)
66. L. Bouten, Filtering and control in quantum optics. Ph.D. dissertation, Catholic University of
Nijmegen (2004)
67. L. Bouten, R. van Handel, Quantum filtering: a reference probability approach (2006).
arXiv:math-ph/0508006 (arXiv preprint)
68. L. Bouten, R. van Handel, On the separation principle of quantum control, in Quantum Stochas-
tics and Information: Statistics, Filtering and Control (University of Nottingham, UK, 15–22
July 2006), ed. by V.P. Belavkin, M. Guta (World Scientific, Singapore, 2008), pp. 206–238
69. H. Mabuchi, N. Khaneja, Principles and applications of control in quantum systems. Int. J.
Robust Nonlinear Control 15(15), 647–667 (2005)
70. K.J. Astrom, R.M. Murray, Feedback Systems: An Introduction for Scientists and Engineers
(Princeton University Press, Princeton, 2008)
71. B.D.O. Anderson, J.B. Moore, Optimal Control: Linear Quadratic Methods (Prentice-Hall,
Englewood Cliffs, 1990)
72. O.L.R. Jacobs, Introduction to Control Theory (Oxford University Press, Oxford, 1993)
73. P. Whittle, Optimal Control (Wiley, Chichester, 1996)
74. D.J. Wilson, V. Sudhir, N. Piro, R. Schilling, A. Ghadimi, T.J. Kippenberg, Measurement-based
control of a mechanical oscillator at its thermal decoherence rate. Nature 524, 325–329 (2015)
75. S. Pirandola, A. Serafini, S. Lloyd, Correlation matrices of two-mode bosonic systems. Phys.
Rev. A 79, 052327 (2009)
76. J. Laurat, G. Keller, J.A. Oliveira-Huguenin, C. Fabre, T. Coudreau, A. Serafini, G. Adesso,
F. Illuminati, Entanglement of two-mode Gaussian states: characterization and experimental
production and manipulation. J. Opt. B: Quantum Semiclass. Opt. 7, S577–S587 (2005)
77. L. Accardi, J. Gough, Y.G. Lu, On the stochastic limit for quantum theory. Rep. Math. Phys.
36(2), 155–187 (1995)
References 33
78. L. Accardi, Y.G. Lu, I. Volovich, Quantum Theory and Its Stochastic Limit. Series, Physics
and Astronomy (Springer, Berlin, 2002)
79. H. Bachor, T. Ralph, A Guide to Experiments in Quantum Optics, 2nd edn. (Wiley-VCH,
Weinheim, 2004)
80. Z.Y. Ou, S.F. Pereira, H.J. Kimble, Realization of the Einstein–Podolski–Rosen paradox for
continuous variables in nondegenerate parametric amplification. Appl. Phys. B 55, 265–278
(1992)
81. D. Vitali, G. Morigi, J. Eschner, Single cold atom as efficient source of EPR-entangled light.
Phys. Rev. A 74, 053814 (2006)
82. S.L. Braunstein, P. van Loock, Quantum information with continuous variables. Rev. Mod.
Phys. 77, 513 (2005)
83. C.K. Law, Interaction between a moving mirror and radiation pressure: a Hamiltonian formu-
lation. Phys. Rev. A 51, 2537–2541 (1995)
84. G.J. Milburn, M.J. Woolley, An introduction to quantum optomechanics. Acta Physica Slovaca
61, 483–601 (2011)
85. L.K. Thomsen, S. Mancini, H.M. Wiseman, Continuous quantum nondemolition feedback and
unconditional atomic spin squeezing. J. Phys. B: At. Mol. Opt. Phys. 35, 4937–4952 (2002)
86. R. Inoue, S. Tanaka, R. Namiki, T. Sagawa, Y. Takahashi, Unconditional quantum-noise supres-
sion via measurement-based quantum feedback. Phys. Rev. Lett. 110, 163602 (2013)
Chapter 2
Mathematical Modeling of Linear Dynamical
Quantum Systems
Abstract This chapter provides a review of the mathematical theory of linear

quantum systems, which is based on the Hudson–Parthasarathy quantum stochastic
calculus as a mathematical tool for describing Markov open quantum systems inter-
acting with external propagating quantum fields. A precise definition of linear quan-
tum systems is given as well as quantum stochastic differential equations representing
their linear equation of motion in the Heisenberg picture. The important notion of
physical realizability for linear quantum stochastic differential equations is intro-
duced, and necessary and sufficient conditions for physical realizability reviewed.
Complete parameterizations for linear quantum systems are given, and transfer func-
tions defined. Also, the special class of completely passive linear quantum systems
is introduced and the notion of stability for linear quantum systems is developed.
Our aim here is to give a brief exposition of the key ideas behind quantum stochas-
tic calculus, as required for the purposes of this monograph. For a more detailed
exposition, we refer the reader to two excellent and comprehensive texts on the sub-
ject, K.R. Parthasarathy’s An Introduction to Quantum Stochastic Calculus [1] and
P.-A. Meyer’s Quantum Probability for Probabilists [2]. We also mention the remark-
ably well-written tutorial paper [3]. Once the basis for quantum stochastic calculus
has been laid out, we proceed to use it to formulate linear dynamical quantum systems
and their stochastic dynamics.
Sections 2.3, 2.4, and 2.7 contain reprinted excerpt with permission from [18]. Copyright (2010)
by the American Physical Society.
Section 2.5.3 contains some materials reprinted, with permission, from [25] © 2014 IEEE.
Section 2.7.2 contains some materials reprinted from [19] © 2014 with permission of Springer.
36 2 Mathematical Modeling of Linear Dynamical Quantum Systems
2.1 Quantum Stochastic Calculus
2.1.1 The Boson Fock Space, Exponential Vectors,

and Fundamental Processes on the Fock Space
The linear quantum systems that are the main subject of this monograph consist of
linearly coupled quantum harmonic oscillators that are in turn coupled to one or more
distinct freely propagating optical fields. The freely propagating fields are bosons
and each field can contain an indefinite number of bosons. A bosonic field can be
mathematically described by a special type of Hilbert space known as a (symmetric)
Fock space. Let H denote a complex Hilbert space with complex inner product ·, ·H
linear in the second slot and antilinear in the first, referred to as the one-particle Hilbert
space. Then the symmetric Fock space over H, denoted by s (H), is defined as
∞

s (H) = C ⊕ H ⊗s j ,
j=1
where H⊗s j = H ⊗s H ⊗s · · · ⊗s H, and ⊗s denotes the symmetric tensor product.

j−times
For any j elements f1 , f2 , . . . , fj ∈ H,the symmetric tensor product of these elements
is given by f1 ⊗s f2 ⊗s · · · ⊗s fj = j!1 π∈Pj fπ(1) ⊗ fπ(2) ⊗ · · · ⊗s fπ(j) , where ⊗ is the
ordinary tensor product on Hilbert spaces, and Pj denotes the set of all permutation
maps π of {1, 2, . . . , j} to itself (there are j! such maps). The Hilbert space H⊗s j is
referred to as the j-particle subspace of s (H). Thus, the symmetric Fock space is
the infinite direct sum of finite particle Hilbert spaces H⊗s j (including the 0-particle
space corresponding to C), representing the fact that a bosonic field can contain an
indefinite number of bosons. The symmetric nature of the Fock space, captured by
the symmetric tensor product space on each finite particle subspace, reflects the fact
that bosons are a class of indistinguishable particles (the other being fermions) with
a wavefunction which is symmetric with respect to an interchange of any pair of its
arguments (fermions, on the other hand, have wavefunctions that are antisymmetric
with respect to an interchange of any pair of its arguments). Every vector ψ ∈ s (H)
can be expressed as an infinite-dimensional vector
ψ = (ψ0 , ψ1 , ψ2 , . . . , ψj , . . .),
with ψ0 ∈ C and ψj ∈ H⊗s j for j ≥ 1. For any two elements ψ, φ ∈ s (H), we have
the inner product ∞

ψ, φ = ψ0∗ φ0 + ψj , φj ⊗s j ,
j=1
with ·, ·⊗s j denoting the inner product on H⊗s j . This inner product is defined via
the identity
u1 ⊗s · · · ⊗s un , v1 ⊗s · · · ⊗s vn ⊗s n = Perm [ui , vj H ]i,j=1,2,...,n ,

2.1 Quantum Stochastic Calculus 37
where Perm(·) denotes the permanent of a square matrix.

An important class of vectors in s (H) is the class of exponential or coherent
vectors e( f ) that is parametrized by f ∈ H. It is defined as
1 ⊗s 2 1
e( f ) = 1, f , f , . . . , f ⊗s k , . . . ,
2! k!
(note that f ⊗s j = f ⊗j ) with inner product
e(g), e( f ) = exp(g, f H ).
In particular, we have the norm
e( f ) = exp( f H ),
2

where · H is the norm on H, f H = f , f H .
Let H1 , H2 , . . . , Hr be Hilbert spaces. The symmetric Fock space has the prop-
erty that s (⊕rj=1 Hj ) ∼ = ⊗rj=1 s (Hj ) for any integer r ≥ 1, where ∼ = denotes
that the two spaces are unitarily equivalent. Therefore, s (⊕rj=1 Hj ) can be iden-
tified with ⊗rj=1 s (Hj ). This identification can be made via the correspondence
e((f1 , f2 , . . . , fr )) ↔ e(f1 ) ⊗ e(f2 ) ⊗ · · · ⊗ e(fr ), for any fj ∈ Hj .
In this monograph, we will be working exclusively with a boson Fock space over
H = L 2 (R+ ; Cm ), with R+ denoting the set of nonnegative real numbers, which we
denote throughout the rest of the monograph as Fm , i.e., Fm = s (L 2 (R+ ; Cm )).
Let Fm (I) = s (L 2 (I; Cm )) for any (Lebesque) measurable set I ⊂ R+ . If
I1 , I2 , . . . , Ir are disjoint subsets of R+ , then Fm (I1 ∪ I2 ∪ · · · ∪ Ir ) = ⊗rj=1 Fm (Ij ).
In the following, we will frequently use the shorthand notation t] ≡ [0, t], t) ≡ [0, t),
[t ≡ [t, ∞), and (t ≡ (t, ∞).
2.1.2 Adapted Processes and Quantum Stochastic Integrals
Let E denote the complex linear space spanned by the exponential vectors on the
Fock space Fm , E = span{e( f ) | f ∈ L 2 (R+ ; Cm )}. For any g ∈ L 2 (R+ ; Cm ) and
bounded self-adjoint operator : L 2 (R+ ; Cm ) → L 2 (R+ ; Cm ), define the operators
A( f ), A∗ ( f ), and () with domain E via their action on the coherent vectors:
∞
A(g)e( f ) = g(s)∗ f (s)ds e( f ),
0
d
A∗ (g)e( f ) = e(f + tg)t=0 ,
dt
()e( f ) = A∗ (f )e( f ).
In particular, we have the relation A∗ ( f )e(h), e(g) = e(h), A( f )e(g) and the
commutation relations
[A( f ), A(g)] = [A∗ ( f ), A∗ (g)] = 0,

∞
[A( f ), A∗ (g)] = f (s)∗ g(s)ds,
0
[(1 ), (2 )] = ([1 , 2 ]),
[A( f ), ()] = A(∗ f ),
[A∗ ( f ), ()] = −A∗ (f ).
Also, A∗ ( f ) is the adjoint of A( f ) with domain Dom(A∗ ( f )) ⊃ E, and () is

self-adjoint on a domain Dom(()) ⊃ E.
Let ei be a standard basis vector in Cm (as a column vector) that is 0 everywhere
except at the i-th element which is 1, and let 1I be the indicator function on the set
I. Define the so-called fundamental processes Aj (t), A∗j (t), and jk (t) acting on E
as
Aj (t) = A(ej 1t] ),

A∗j (t) = A∗ (ej 1t] ),
jk (t) = (ej e
k 1t] ),
with the indices j and k ranging from 1 until m. The subscript j for Aj (t) and
A∗j (t) indicates that they are field annihilation and creation processes on the j-th
bosonic field, respectively, while the subscript jk on jk (t) indicates that it is a pho-
ton exchange operator from the k-th boson field to the j-th field. We may then identify
(and we shall do so for the rest of the monograph without further comment) A(t)
and A∗ (t) as vectors of operators on Fm :
A(t) = [ A1 (t) A2 (t) . . . Am (t) ] ,

A∗ (t) = [ A∗1 (t) A∗2 (t) . . . A∗m (t) ] ,
and identify (t) with the matrix of operators,
(t) = [jk (t)]j,k=1,2,...,m .
From the properties of the creation process, we will freely make the identification
A∗j (t) = Aj (t)∗ , and A∗ (t) = A(t)∗ on E. We also note the fact that (t) = (t)∗
on E.
The process (t) is known as the counting or gauge process. Let h denote the
underlying Hilbert space of a quantum system that is coupled to bosonic fields on the
Fock space Fm . A process {X(t), t ≥ 0} defined on h ⊗ Fm is said to be adapted if
X(t) ψ(t) ⊗ e(f 1[t ) = φ(t) ⊗ e(f 1[t ) for any f ∈ L 2 (R+ ; Cm ), ψ(t) ∈ h ⊗ Fm (t]),
and some φ(t) ∈ h ⊗ Fm (t]). That is, an adapted process acts trivially (i.e., as the
identity operator) on the future factor Fm ([t) of Fm .
Observe that (A(t) − A(s))e( f ) = e(f 1s] ) ⊗ A(1[s,t] )e(f 1[s,t] ) ⊗ e(f 1[t ), and
similarly, (A∗ (t) − A∗ (s))e( f ) = e(f 1s] ) ⊗ A∗ (1[s,t] )e(f 1[s,t] ) ⊗ e(f 1[t ), and
((t) − (s))e( f ) = e(f 1s] ) ⊗ (1[s,t] )e(f 1[s,t] ) ⊗ e(f 1[t ). With these properties,
one can define quantum stochastic integrals with respect to adapted processes:
t
(X1 (s)dA(s) + X2 (s)dA(s) + X3 (s)d(s))
0
t
= (X1 (s) ⊗ dA(s) + X2 (s) ⊗ dA(s) + X3 (s) ⊗ d(s)) , (2.1)
0
where X1 , X2 , X3 are adapted processes, and the quantum stochastic integral is defined
on h ⊗ Fm . The quantum stochastic integral then also defines an adapted process.
These quantum stochastic integrals can be constructed in a fashion that is similar to
the construction of classical Itō stochastic integrals; see [1, 2, 4] for details. Also,
note the important property that any adapted process X(t) commutes with Aj (1[τ1 ,τ2 ) ),
A∗j (1[τ1 ,τ2 ) ), and jk (1[τ1 ,τ2 ) ) for any j, k and any t ≤ τ1 < τ2 ≤ ∞.
From a physical point of view, the processes Aj (t), A∗j (t), and jk (t) arise as
integrated version of idealized quantum white noise processes ξj (t), j = 1, 2, . . . , m,
satisfying the singular commutation relations [ξj (t), ξk (s)∗ ] = δjk δ(t − s) [5, 6],
t
introduced earlier in Chap. 1. That is, we may formally write Aj (t) = 0 ξj (s)ds,

A∗j (t) = 0 ξj (s)∗ ds, and jk (t) = 0 ξj (s)∗ ξk (s)ds. Due to the singular nature of the
t t
quantum white noise processes, it is mathematically simpler to work with the more
regular integrated processes Aj (t), A∗j (t), and jk (t) as these can be rigorously and
explicitly constructed as processes on the Fock space Fm , as we have already briefly
elaborated upon. However, from the perspective of describing the physics involved
in the interaction between the system and the bosonic environment, the white noise
picture is more meaningful, and we shall often employ this in the discussion that
ensues. That is to say that the more fundamental processes from the underlying
physics are the quantum white noise processes rather than the mathematically more
convenient “derived” integrated processes. It should be noted that quantum white
noise processes arise as a consequence of making a Markov assumption regarding
the interaction between the system and the bosonic environment, and it is only within
this approximation that the quantum white noise processes can be given a sensible
physical interpretation. The modeling of a bosonic environment as a quantum white
noise process without the Markov approximation can be problematic and gives rise
to physical inconsistencies, see, e.g., [7] for a discussion.
2.1.3 The Quantum Itō Table in Vacuum and the Quantum

Itō Rule
The fundamental processes Aj (t), A∗j (t), and jk (t) are quantum stochastic
processes on the Fock space Fm . The processes Aj (t) + A∗j (t) and −ıAj (t) + ıA∗j (t)
are non-commuting processes that are each isomorphic to a classical standard Wiener
processes, that is, each can be viewed as Fock space representations of the standard
Wiener process. On the other hand, jj (t) is isomorphic to a classical Poisson process,
that is, the latter is a realization of the former on the Fock space. As a quantum
mechanical system, the bosonic fields have a quantum state that determines their
statistics under measurement. We now assume that the bosonic fields are in the vac-
uum state | = e(0). This is a state in which the fields do not contain any photons.
In this state, we have that A( f )e(0) = 0 for all f ∈ L 2 (R+ ; Cm ), and the forward-
pointing differentials dAj (t) = Aj (t +dt)−Aj (t), dA∗j (t) = A∗j (t +dt)−A∗j (t), and
djk (t) satisfy the quantum Itō product rule, as a quantum adaptation of the clas-
sical Itō product rules: dAj (t)dA∗k (t) = δjk dt; dA∗j (t)dAk (t) = dAj (t)dAk (t) =
dA∗j (t)dA∗k (t) = 0, and
djk (t)dj k (t) = δkj djk (t), dAj (t)dkl (t) = δjk dAl (t),
djk dA∗l (t) = δkl dA∗j (t).
Let X(t) and Y (t) be two adapted processes on h ⊗ Fm that can be expressed as
quantum stochastic integrals with respect to the fundamental processes Aj (t), A∗j (t),
and jk (t), as in (2.1). The quantum Itō product rule holds for the forward-pointing
differential of the product process X(t)Y (t):
d(X(t)Y (t)) = (dX(t))Y (t) + X(t)dY (t) + dX(t)dY (t).
Note the third term that serves as a quantum analogue of the second-order correction
term in the classical Itō stochastic calculus. Since X and Y are quantum stochastic
integrals with respect to the fundamental processes, the correction term can be cal-
culated using (2.1) together with the quantum Itō product rule for the fundamental
processes given above.
2.1.4 The Hudson–Parthasarathy Quantum Stochastic

Differential Equation
Using the quantum stochastic integrals and the quantum Itō rules, one can define
adapted processes as solutions of quantum stochastic differential equations (QSDEs).
A particularly important class of QSDEs that describe the physical scenario of
Markov open quantum systems coupled to a bosonic environment is the Hudson–
Parthasarathy QSDE of the form:
dU(t) = Tr (S − I) d(t) + dA(t)∗ L − L ∗ SdA(t)
−(ıH + 1/2L ∗ L)dt U(t), (2.2)

with initial condition U(0) = I. Here, H is the Hamiltonian of the system and is a
self-adjoint operator on the system Hilbert space h, L = [ L1 L2 . . . Lm ] is a vector
of operators Lj on h, and S ∈ Cm×m ⊗ h is a unitary operator (i.e., S ∗ S = SS ∗ = I),
with entries Sjk being operators on h, called a scattering matrix. They are the operator-
valued “coefficients” or parameters of the QSDE. The form of the equation is such
that the solution U(t) of the QSDE is a unitary adapted process. That is, U(t) is an
adapted process and is unitary for each t: U(t)∗ U(t) = U(t)U(t)∗ = I.1
Following [9], we denote a quantum system G, whose dynamics are governed by
the unitary solution of a Hudson–Parthasarahy QSDE (2.2) with parameters S, L,
and H, by the shorthand G = (S, L, H).
We will now elaborate on properties of the Hudson–Parthasarathy QSDE that
will be crucial for the (measurement-based) feedback control theory of open quantum
systems. Let θt : L 2 (R; Cm ) → L 2 (R; Cm ) denote the one-particle left shift operator,
defined by θt f (·) = f (· + t) ∀t ∈ R+ . Let t be the so-called second quantization of
the one-particle left shift operator, defined as an operator on Fm− = s (L 2 (R; Cm ))
by its action on the exponential vectors as t e( f ) = e(θt f ). It is straightforward
to verify that ∗t e( f ) = e(f (· − t)) and t can be uniquely extended to a unitary
operator on Fm− , ∗t t = t ∗t = I. The solution U(t) of a Hudson–Parthasarathy
QSDE, with U(t) extended from h ⊗ Fm to h ⊗ Fm− as I ⊗ U(t), has the property
that it is a left cocycle with respect to t , meaning that the following holds for all
0 ≤ s ≤ t,
U(t) = ∗s U(t − s)s U(s).
We note that V (t, s) = ∗s U(t − s)s acts non-trivially only on the portion h ⊗
Fm ([s, t]) of h ⊗ Fm . Note that U(t) = V (t, 0) and the cocycle property can be
equivalently expressed as V (t, s) = V (t, τ )V (τ , s) for all 0 ≤ s ≤ τ ≤ t. Indeed,
we have that
V (t, τ )V (τ , s) = ∗τ U(t − τ )τ ∗s U(τ − s)s

= ∗τ U(t − τ )τ −s U(τ − s)s
= ∗s ∗τ −s U(t − τ )τ −s U(τ − s)s
= ∗s V (t − s, τ − s)U(τ − s)s
= ∗s U(t − s)s
= V (t, s),
where in the second and third lines we have used the fact that τ ∗s = τ −s .
A consequence of this property is that Ũ(t) = t U(t) defines a strongly continuous
one-parameter semigroup of unitary operators on h ⊗ Fm− , Ũ(t)Ũ(s) = Ũ(t + s)
for all s, t ≥ 0 with Ũ(0) = I. Therefore there exists a densely defined essentially
self-adjoint operator K such that Ũ(t) = e−ıKt . The operator K is the Stone generator
1 Technically, the existence of a unique solution of the Hudson–Parthasarathy QSDE that is unitary is
guaranteed whenever the coefficients S, L, H are all bounded operators. When they are unbounded
then additional technical assumptions need to be assumed, see, e.g., [8].
of the strongly continuous semigroup {Ũ(t), 0 ≤ s ≤ t}. The difficult and long-
standing problem of characterizing this generator was finally resolved by Gregoratti
[10, 11] building on the work of Chebotarev [12].
Let us now turn to the Heisenberg picture of quantum mechanics, where states
are fixed and operators evolve in time. This will be the natural setting for studying
quantum filtering and control problems that will be considered later in the monograph.
Let jt (·) = U(t)∗ · U(t) and let Z(t) denote any adapted process on the boson Fock
space Fm . Then by the cocycle property of V (t, s) we find that, for any t ≥ s ≥ 0,
jt (Z(s)) = U(t)∗ Z(s)U(t)

= U(s)∗ ∗s U(t − s)∗ s Z(s)∗s U(t − s)s U(s)
= U(s)∗ Z(s)U(s)
= js (Z(s)).
Using this property, it follows that for any (bounded linear) system operator X on h
and 0 ≤ s ≤ t,
jt (X)js (Z(s)) = jt (X)jt (Z(s))

= U(t)∗ XZ(s)U(t)
= U(t)∗ Z(s)XU(t)
= jt (Z(s))jt (X)
= js (Z(s))jt (X).
Therefore, [jt (X), js (Z(s))] = 0 for any system operator X and any adapted process
Z(t) on Fm . In other words, {jt (X), js (Z(s)), 0 ≤ s ≤ t} forms a commutative family
of operators, a property of input–output Markov open quantum systems known as
the non-demolition property. Now, let Z(t) be an adapted process on Fm with the
additional property that [Z(t), Z(s)] = 0 for all s, t ≥ 0. For example, Z(t) =
e−ıφ A(t) + eıφ A(t)∗ (φ ∈ R) and Z(t) = (t), all have these properties. It follows
that
jt (Z(t))js (Z(s)) = jt (Z(t))jt (Z(s))

= U(t)∗ Z(t)Z(s)U(t)
= U(t)∗ Z(s)Z(t)U(t)
= jt (Z(s))jt (Z(t))
= js (Z(s))jt (Z(t)).
Hence, when Z(t) has the stipulated properties, [jt (Z(t)), js (Z(s))] = 0 for all s, t ≥ 0
and {jt (Z(t)), js (Z(s)), 0 ≤ s ≤ t} forms a commutative family of operators. This
is a property of input–output Markov open quantum systems known as the self non-
demolition property.
Now, let Z(t) be an observable for each t, Z(t) = Z(t)∗ , the self non-demolition
property means that Zt = { js (Z(s)), 0 ≤ s ≤ t} describes a continuous out-
put process of the quantum fields that may be continuously monitored/measured
(depending on what Z(t) actually is), to yield a real-valued measurement record.
When Z(t) is self non-demolition, the non-demolition property means that informa-
tion about jt (X) can be inferred from an observation record of Zt . This inference
process is known as quantum filtering, the quantum analogue of classical nonlinear
filtering theory [13], and will be treated in Chap. 4.
2.2 Linear Dynamical Quantum Systems: Joint Unitary

Evolution of Oscillators and Boson Fields
We now begin to introduce and specialize to the class of linear quantum stochastic
systems as the central theme of this monograph. This section describes the joint
evolution of a set of coupled oscillators that are also coupled to external bosonic
fields, via a certain form of interaction, which results in linear dynamics of the
oscillators’ position and momentum operators in the Heisenberg picture. The linear
dynamics is a defining feature of linear quantum systems.
Let there be n independent quantum harmonic oscillators. The j-th quantum har-
monic oscillator has position and momentum operators qj and pj acting on elements
of the underlying Hilbert space L 2 (R; C), the space of all square integrable complex-
valued function on R. These operators satisfy the canonical commutation relations
(CCR)2 :
[qj , pk ] = 2ıδjk , [qj , qk ] = 0 , [pj , pk ] = 0. (2.3)
The position and momentum operators of the n oscillators can be collected in a single
column vector of operators x defined by x = (q1 , p1 , q2 , p2 , . . . , qn , pn ) . We can
then express the CCR more compactly as,
[x, x ] = xx − (xx ) = 2ıJn ,
with Jn = In ⊗ J, with
0 1
J= .
−1 0
Also note in passing that xx = (xx ) since some elements of x do not commute
with one another. The composite system of n quantum harmonic oscillators has a
quadratic Hamiltonian H given by H = (1/2)x Rx, with R a real symmetric 2n×2n
2 We take the common convention that units are taken such that = 1. Also, we will take (2.3) as
the default CCR for the position and momentum operators of multiple distinct oscillators. However,
it is easy to adapt the results of this chapter for a different definition of these operators that satisfy
a different set of commutation relations; see Remark 2.2.
matrix. The quantum harmonic oscillators are also coupled to m distinct external
quantum bosonic fields. They are coupled to the k-th quantum field via a singular
interaction of the form Hamiltonian Hk = ı(Lk ξk∗ (t)−Lk∗ ξk (t)) [5, 6], where Lk = Kk x
(with Kk ∈ C1×2n ) is a linear coupling operator describing the coupling of the position
and momentum operators to ξk (t). Here ξk (t) is a quantum white noise process
as discussed in Sect. 2.1.2. We can collect the coupling operators L1 , L2 , . . . , Lm
together in one linear coupling vector L = [ L1 L2 . . . Lm ] = Kx, with K =
[ K1 K2 . . . Km ] . The joint evolution of the oscillators and quantum fields is then
given by a unitary adapted process U(t) satisfying the Hudson–Parthasarathy QSDE
(2.2), with S being a fixed unitary matrix in Cm×m , i.e., the entries of S in this case
are complex numbers rather than operators on h.
Using the shorthand notation introduced earlier, a linear quantum stochastic
system with parameters S, L = Kx, H = (1/2)x Rx as described above can be
expressed as G = (S, Kx, (1/2)x Rx). In a later section, we will slightly generalize
the notion of a linear quantum system by allowing so-called Bogoliubov transfor-
mations to replace the scattering matrix S.
2.3 Equations of Motion: Real Quadrature Form

and Complex Mode Form
Using the quantum Itō rule and the quantum Itō products, and exploiting the canonical
commutation relations between the operators in x, the Heisenberg evolution
⎡ ⎤
U(t)∗ q1 U(t)
⎢ U(t)∗ p1 U(t) ⎥
⎢ ⎥
⎢ U(t)∗ q2 U(t) ⎥
⎢ ⎥
⎢ ∗ ⎥
x(t) = U(t)∗ xU(t) = ⎢ U(t) p2 U(t) ⎥
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎣ U(t)∗ qn U(t) ⎦
U(t)∗ pn U(t)
of the vector x can be obtained [14, 15]. This is given by the QSDE,
dx(t) = d(U(t)∗ xU(t))

d A(t)
= Ao x(t)dt + Bo d A(t)# ; x(0) = x,
dY(t) = d(U(t)∗ A(t)U(t))

= Co x(t)dt + Do dA(t), (2.4)
2.3 Equations of Motion: Real Quadrature Form and Complex Mode Form 45
with
Ao = 2Jn (R + {K ∗ K}),

Bo = 2ıJn [ −K † S K S # ],
Co = K,
Do = S.
Here,
Y(t) = (Y1 (t), . . . , Ym (t)) = U(t)∗ A(t)U(t)
is a vector of output fields that is produced by the interaction of the quantum harmonic
oscillators and the incoming quantum fields A(t). Note that the Heisenberg picture
dynamics of x(t) is linear, and Y(t) has a component which is a linear combination
of elements of x(t). The index n in the above will be referred to as the number of
degrees of freedom or simply the degree of the linear quantum stochastic system.
2.3.1 Real Quadrature Form
In certain circumstances, it is convenient to write the dynamics (2.4) in the so-called

(real) quadrature form as in [15–17]:
dx(t) = Ax(t)dt + Bdw(t); x(0) = x,

dy(t) = Cx(t)dt + Ddw(t), (2.5)
with
w(t) = 2({A1 (t)}, {A1 (t)}, . . . , {Am (t)}, {Am (t)}) ,

y(t) = 2({Y1 (t)}, {Y1 (t)}, . . . , {Ym (t)}, {Ym (t)}) .
Here,
A = 2Jn (R + {K ∗ K}), (2.6)

B = 2ıJn [ −K ∗ K ]diag(S, S # )m , (2.7)

K + K#
C = Pm , (2.8)
−ıK + ıK #
D = m−1 diag(S, S # )m , (2.9)
where Pm denotes a 2m × 2m permutation matrix acting as
Pm [ a1 a2 . . . a2m−1 a2m ]
= [ a1 a3 . . . a2m−1 a2 a4 . . . a2m ] ,
and
1 1 ı
m = Pm Im ⊗ .
2 1 −ı
Note that the matrices A, B, C, D in the quadrature form are all real and are in a
one to one correspondence with the matrices Ao , Bo , Co , Do , and the quantum noise
vector w(t) satisfies the Itō relationship dw(t)dw(t) = (I + ıJm )dt.
Some remarks are in order about the nature of the quadrature quantum noise w.
From the self non-demolition property we have that [wj (s), wj (t)] = 0 (wj denotes
the j-th component of w) for all s, t ≥ 0 and all j. Indeed,
[wj (s), wj (t)]

= [Aj (t) + A∗j (t), Aj (s) + A∗j (s)]
= [Aj (t), Aj (s)] + [Aj (t), A∗j (s)] + [A∗j (t), Aj (s)] + [A∗j (t), A∗j (s)]
= min(t, s) − min(t, s)
= 0.
When the state is the vacuum state, wj is isomorphic to a standard Wiener process,
in the sense that it is a realization of a standard Wiener process as an operator-
valued process on a Fock space. This will be demonstrated in Sect. 2.7.2 when
Gaussian states of a bosonic field are discussed. However, for any j = 1, 2, . . . , m,
w2j−1 (t) and w2j (t) do not commute for any t, meaning that they cannot be realized
jointly on a common classical probability space. On the other hand, we do have that
[w2j−1 (s), wk (t)] = 0 for all k = 2j and all s, t ≥ 0, so that the collection of processes
{w2j−1 , wk , k = 2j} is equivalent to a collection of independent standard Wiener
processes that can be jointly defined on a common classical probability space. The
non-commutativity among components of w(t) is a crucial difference from the usual
vector of classical independent Wiener processes.
2.3.2 Complex Mode Form
In other situations, it may be more convenient to represent the dynamics of a lin-

ear quantum stochastic system in the (complex) mode form. In this form, instead
of writing down the Heisenberg evolution of amplitude-phase quadrature pairs
(q1 , p1 , q2 , p2 , . . . , qn , pn ) , we write down the Heisenberg evolution of the
annihilation–creation pairs (a1 , a1∗ , a2 , a2∗ , . . . , an , an∗ ) , where aj = 21 (qj + ıpj ) and
aj∗ = 21 (qj − ıpj ). In terms of these operators, the quadratic Hamiltonian H and
coupling operator L can be expressed as [18]
1 1
H = a∗ − a + a∗ + a# + a #+ a,
2 2
where + = [ωjk+ ] and − = [ωjk− ] are complex n × n matrices possessing the

symmetries ∗− = − and + = + , and
L = C− a + C+ a# ,
where C± ∈ Cm×n . Using these expression for H and L one can write down the
QSDE for the Heisenberg evolution of a and a# . Let
a(t) = U(t)∗ aU(t)

= (U(t)∗ a1 U(t), U(t)∗ a2 U(t), . . . , U(t)∗ an U(t)) ,
and introduce the doubled-up matrix

E− E+
(E− , E+ ) ,
E+# E−#
for any complex matrices E± of the same dimension. We also introduce the notation

Im 0
Km , (2.10)
0 −Im
and define the · involution operator that acts on a 2k × 2l matrix X as
X = Kl X ∗ Kk . (2.11)
Then the QSDE for (a(t) , a(t)∗ ) and the output Y(t) can be expressed in the
doubled-up form [18]

da(t) a(t) dA(t)
= Ã dt + B̃ ,
da(t)# a(t)# dA(t)#

dY(t) a(t) dA(t)
= C̃ dt + D̃ , (2.12)
dY(t)# a(t)# dA(t)#
where
1
A∓ = − (C−∗ C∓ − C+ C±# − ı∓ ),
2
1
− = − (A− − A∗− ),
2ı
1
+ = − (A+ + A + ),
2ı
Ã = (A− , A+ ), (2.13)
B̃ = −(C− , C+ ) (S, 0), (2.14)
C̃ = (C− , C+ ), (2.15)
D̃ = (S, 0). (2.16)
Note that it is necessary to adopt a doubled-up form since the evolutions of a(t) and
a(t)# are, in general, coupled. The following example illustrates the description of a
system in mode form.
Example 2.1 Consider the nondegenerate optical parametric amplifier (NOPA) from
Sect. 1.5.2. We consider a lossless version of this device with κ = 0. Therefore, the
device is only coupled
to two fields √ in the vector A(t), and has parameters
collected
0 ı /2 γ 0
− = 02×2 , + = , C− = √ , C+ = 02×2 , and S = I2 . The
ı /2 0 0 γ
doubled-up form of the evolution of the cavity modes in the vector a = [ a1 a2 ] is
⎡ ⎤
−γ/2 0 0 /2
da(t) ⎢ 0 −γ/2 /2 0 ⎥ a(t)
= ⎢ ⎥ dt
da(t)# ⎣ 0 /2 −γ/2 0 ⎦ a(t)#
/2 0 0 −γ/2
⎡√ ⎤
γ 0 0 0
⎢ 0 0 √
γ 0 ⎥ ⎥ dA(t)# ,
−⎢ ⎣ 0 √
γ 0 0 ⎦ dA(t)
√
0 0 0 γ

dY(t) √ a(t) dA(t)
= γ dt + .
dY(t)# a(t)# dA(t)#
In the physics literature, it is more common for the quantum stochastic dynamics to
be written in the heuristic form of a quantum Langevin equation (introduced
t earlier
in Chap. 1) driven by a quantum white noise process ξ(t), where A(t) = 0 ξ(s)ds.
For the mode form of the equation of motion, this quantum Langevin equation takes
the form

ȧ(t) a(t) ξ(t)
= Ã + B̃ ,
ȧ(t)# a(t)# ξ(t)#

η(t) a(t) ξ(t)
= C̃ + D̃ ,
η(t)# a(t)# ξ(t)#
where η is the output process that propagates from the system after the incoming white
noise ξ interacts with the latter. The process η satisfies the same time commutation
relations as ξ.
Remark 2.1 Note that for notational expediency, in the monograph we will often not
explicitly write the time dependence in the QSDEs describing the evolution of the
system.
2.3.3 Transfer Function of Linear Dynamical Quantum

Systems
Whether one is working with a linear quantum system in the quadrature form (2.5)
or the mode form (2.12), as with classical linear systems, one can define a transfer
function. We start with a system G = (A, B, C, D) in the quadrature form (2.5).
Following [18], we can define the Laplace
∞ transform of an adapted quantum stochastic
process such as w(t) as w[s] = 0 e−st dw(t). The “particular” solution of (2.5),
that does not depend on the initial condition x(0) = x, is given by
t
yp (t) = CeA(t−s) Bdw(s) + Dw(t).
0
Now, taking Laplace transforms of both side of the equality gives
yp [s] = (C(sI − A)−1 B + D)w[s],
and from this we can define the transfer function G (s) from w[s] to yp [s] of G just
as for a classical linear system,
G [s] = C(sI − A)−1 B + D.

∞
Similarly, if we are ∞ working in the modeform (2.12) we define A[s]
= 0 e−sτ
∞ ∞
dA(τ ), Y[s] = 0 e−sτ dY(τ ), ξ[s] = 0 e−sτ ξ(τ )dτ , η[s] = 0 e−sτ η(τ )dτ .
t t
Since A(t) = 0 ξ(τ )dτ and Y(t) = 0 η(τ )dτ , note that A[s] = ξ[s] and Y[s] =
η[s]. Moreover, A# [s] = A[s∗ ]# and analogous identities hold for Y[s], ξ[s] and η[s].
The transfer function ˜ G from (A[s], A∗ [s]) to (Y[s], Y ∗ [s]) or, equivalently,
from (ξ[s], ξ ∗ [s]) to (η[s], η ∗ [s]) , for G in the mode form G = (Ã, B̃, C̃, D̃) is
given by
˜ G [s] = C̃(sI − Ã)−1 B̃ + D̃.
The two transfer functions G and ˜ G for the quadrature and mode forms of G,
respectively, are related by the identity
G [s] = m−1 ˜ G [s]m ,
where m is as defined before.

In analogy with the classical case, the transfer function is important in analyzing
the input–output behavior of linear quantum systems.
2.4 Inclusion of Idealized Static Transformations

on Bosonic Fields: The Bogoliubov Transformation
When H = 0 and L = 0, a linear unitary transformation of the input field A(t) is

realized,
Y(t) = SA(t).
Such a transformation preserves the differential commutation relations, that is,

∗ ∗
Y(t) Y(t) A(t) A(t)
d , d = d , d = Km dt, (2.17)
Y(t)# Y(t)# A(t)# A(t)#
where for a (column) vector of operators v and w,
[v, w ∗ ] = vw ∗ − (w # v ) .
However, a unitary S is not the only matrix that can result in preservation of the
differential commutation relations. More generally, one may consider a so-called
Bogoliubov transformation implemented by a Bogoliubov matrix W = (W+ , W− )
of the form
Y (t) A(t)
= W , (2.18)
Y (t)# A(t)#
such that Y (t) satisfies the same differential commutation relation as A(t). This is
guaranteed by the definition of a Bogoliubov matrix, that will now be given.
Definition 2.1 Let · denote the involution defined via (2.11). A complex 2m × 2m
matrix W is a Bogoliubov matrix if it is of a doubled-up form W = (W+ , W− ) for
some complex m × m matrices W+ , W− , and W is -unitary,
W W = W W = I2m .
Notice that, by definition, W = W −1 and W reduces to a unitary matrix when

W− = 0. Linear transformations of the form (2.18) preserve the differential commu-
tation relation, so that the output field Y (t) is a valid output field. However, when
W− = 0 it cannot be modeled within the Hudson–Parthasarathy QSDE framework.
That is, there does not exist a unitary solution U(t) of the QSDE such that Y (t) in
(2.18) can be written as Y (t) = U(t)∗ A(t)U(t).
From a physical point of view, when W− = 0 the “pathology” discussed above
can be understood from the fact that transformations such as (2.18) represent the
2.4 Inclusion of Idealized Static Transformations on Bosonic Fields … 51
action of an idealized infinite-bandwidth squeezing device acting on an incoming

boson field A(t). Squeezing here means that the device reduces the quantum fluctua-
tions on some quadratures of the incoming field A(t) while increasing them in other
quadratures. We will shortly illustrate this point by demonstrating how a Bogoliubov
transformation arises as the infinite-bandwidth limit of a degenerate parameteric
amplifier. Such an idealized device is not physical since it produces uniform squeez-
ing across the continuum of frequencies contained in A(t), a process that would
require an infinite amount of energy. From a mathematical physics perspective, the
linear transformation (2.18) represents a transformation between two fields A(t) and
Y (t). The former arises from a Fock space representation of field operators satisfying
∞
the canonical commutation relations [A( f ), A∗ (g)] = 0 f (s)∗ g(s)ds with respect
to vacuum state of the field, while the latter arises from a Fock space representation
with respect to another state induced by W (e.g., a squeezed field state as postulated
in [6, Chap. 6]); see, e.g., [20] for a more in-depth treatment of these representations
in the case of squeezed field states. When W− = 0 the two representations are not
unitarily equivalent, since the transformation involved violates the necessary and
sufficient conditions of the Shale’s Theorem [1, Theorem 22.11].
Although the transformation (2.18) with W− represents an idealized non-physical
scenario, it is convenient to include in the modeling repertoire as it allows some
simplification in describing the system. That is, it allows high-bandwidth dynamic
linear devices with fast dynamics to be approximated as static linear transforma-
tions without any dynamics. This would be appropriate in instances where the fast
device is connected to other linear devices with appreciably lower bandwidth (slower
dynamics). We illustrate this in the next example.
Example 2.2 Recall the degenerate parametric amplifier (DPA) introduced in

Sect. 1.5.3. Under the assumption of losslessness (κ = 0), the DPA is modeled
as a single oscillator coupled to a single field with S = I, − = 0, + = ı , > 0,
√
C− = γ, and C+ = 0. The matrix Ã is Hurwitz (all its eigenvalues are in the left
half plane) and the system is stable (in the sense that the oscillator’s mean photon
number is bounded at all times) if we take < γ as we shall do from now on. The
mode form transfer function of the system is
2
˜ DPA [s] =
1 s − (γ 2 + 2 )/4 − γ/2
,
P(s) − γ/2 s2 − (γ 2 + 2 )/4
where P(s) = (s +γ/2)2 − 2 /4. The zeros of P are the poles of the transfer function,
namely s = ± /2 − γ/2. In the frequency domain, the output white noise field is
1 1
η[s] = s2 − (γ 2 + 2
)/4 ξ[s] − γξ ∗ [s],
P[s] 2P(s)
where ξ ∗ [s] = ξ[s∗ ]∗ . In terms of quadratures η q = η + η ∗ and η p = (η − η ∗ )/ı, we

find that
q p
η q [s] = DPA [s]η q [s], η p [s] = DPA [s]η p [s],
where
q s − (γ + )/2 1
DPA [s] = = p .
s + (γ − )/2 DPA [s]
In this example we consider the idealized case where γ, → ∞ (approximated in

practice by suitably large values of these parameters) while maintaining the ratio γ
to be constant; see [6, 10.2.1.g] for a more complete discussion. Rescaling γ = kγ0
and = k 0 is equivalent to the substitution of γ by γ0 and by 0 and the rescaling
of s as ks :
˜ DPA [s; γ = kγ0 , = k 0 ] = ˜ DPA [s/k; γ0 , 0 ].
We now take the limit k → ∞ in which the cavity has an instantaneous response.
Thus the cavity’s internal dynamics are essentially eliminated. This yields the output
η[s] given by
η[s] = − cosh(r0 ) η[s] − sinh(r0 ) η ∗ [s],
where
γ0 + 0
r0 = ln ,
γ0 − 0
as a Bogoliubov transformation of the input. In this limit the DPA can thus be viewed
as a static device that outputs a squeezed white noise field from a vacuum white
noise source. The transfer function becomes a constant Bogoliubov matrix that is
independent of frequency. That is,
˜ DPA static [s] = lim ˜ DPA [s/k; γ0 , 0 ]

k→∞
= −(cosh r0 , sinh r0 ), ∀s ∈ C,
and the quadrature transfer functions are the constant functions
= −e−r0 .
q p
DPA static [s] = −er0 , DPA static [s]
Therefore, we see that in the high bandwidth regime, the DPA can be approximated
by a Bogoliubov transformation implemented by the Bogoliubov matrix

cosh(r0 ) sinh(r0 )
W =− .
sinh(r0 ) cosh(r0 )
When a linear quantum system (2.12) with D̃ = I is driven by vacuum noise that
has passed through a device modelled as a static Bogoliubov transformation W =
(W− , W+ ), as shown in part (a) of Fig. 2.1, then the equation of motion for the
system is again of the form (2.12) but now with the matrix D̃ = (S, 0) being
replaced by D̃ = W in (2.13)–(2.16). In particular, we have that
(a)
G0 = (I, Kx, x Rx/2) W
(b)
G = (W, Kx, x Rx/2)
Fig. 2.1 a A static device performing a linear transformation W driving a linear quantum system
G 0 with scattering matrix S = I, and b Equivalent representation of the cascaded system in part a
as a single linear quantum system G = (W, L, H)
1
˜
Ã = − C̃ C̃ − ı , (2.19)
2
B̃ = −C̃ D̃, (2.20)
C̃ = (C− , C+ ), (2.21)
D̃ = W, (2.22)
with
˜ = −ı(ı− , ı+ ).

Note the property that ˜ We denote systems of the type depicted in Fig. 2.1
˜ = .
part (a) with the shorthand notation G = (W, L, H), as shown in part (b) of the
figure.
2.4.1 Completely Passive Linear Dynamical Quantum

Systems
This section will introduce a special class of linear quantum systems that shall be
referred to as completely passive linear quantum stochastic systems, for reasons that
will be explained below. This class will appear in several contexts in the monograph
and are of interest in applications such linear optical quantum memories. Further
discussions relating to this class of systems can be found in, e.g., [21–23] and in
Chap. 3 of this monograph.
For k = 1, . . . , n let ak = qk +ıp 2
k
be the annihilation operator for mode k and
define a = (a1 , . . . , an ) . The vector a satisfies the CCR
#
a ∗ a ∗
# a a − a a
a a
= diag(In , −In ),
and note that,

a
= x,
a# #
where ⎡1 ⎤
2
ı 21 0 0 0 ... 0
⎢0 0 21 ı 21 0 ... 0 ⎥
⎢ ⎥
=⎢. .. .. .. .. .. .. ⎥.
⎣ .. . . . . . . ⎦
0 ... ... ... 0 1
2
ı 21
−1
∗

Moreover, we also have # = 2 and, from the relation # 2
∗
[ ] = I, the identities,
∗ ∗
= I/2 = #
, =0= #
. (2.23)
Therefore, −1
a ∗
a
x= # =2 .
a# a#
We say that a linear quantum system G = (S, Kx, (1/2)x Rx) with n degrees of
freedom is completely passive if H = (1/2)x Rx = (1/2)a∗ R̃a + c and L = Kx =
K̃a for some real constant c, some complex n × n Hermitian matrix R̃, and some
complex m × n matrix K̃. The terminology is motivated by physical considerations.
Notice that the Hamiltonian H contains no terms of the form c1 aj2 , c2 ak∗2 , c3 aj ak
and c4 aj∗ ak∗ and, likewise, the coupling operator L contains no terms of the form
c5 ak∗ (where c1 , . . . , c5 denote arbitrary complex constants), with the indices j and k
running over 1, . . . , n. These terms correspond to interactions that require an external
source of quanta (e.g., an external pump beam) to implement. Therefore, they cannot
be realized using only passive optical components like phase shifters, beam splitters
and mirrors, devices that will be discussed in more detail in Chap. 3. This is the
physical motivation for referring to these class of systems as completely passive.
In a completely passive system, the Heisenberg picture evolution of the system
annihilation operator a(t) is not coupled to a(t)# or to A(t)# . In the complex mode
form, the evolution of a(t) is given by
da(t) = (ı R̃ − (1/2)K̃ ∗ K̃)a(t)dt + K̃ ∗ SdA(t),

dY(t) = K̃a(t)dt + SdA(t).
The transfer function [s] from A[s] to Y[s] given by,

−1
[s] = K̃ sI − (ı R̃ − (1/2)K̃ ∗ K̃) K̃ ∗ + I S,
has the special property that [ıω] is unitary for all for all ω for which ıω is not
an imaginary eigenvalue of ı R̃ − (1/2)K̃ ∗ K̃ [21]. It is easy to see that ı R̃ − (1/2)K̃ ∗ K̃
cannot have any eigenvalues in the (open) right half plane, with a positive real part.
Moreover, if this matrix has all its eigenvalues in the (open) left half plane (all
eigenvalues have negative real part), meaning that the system is stable in a sense that
will be discussed in Sect. 2.6, then [s] is in fact lossless bounded real [22]. This
means that [s] is analytic for all s with {s} > 0, [ıω] is unitary for all ω ∈ R
and, moreover, [s] satisfies,
[s]∗ [s] ≤ I,
for all s with {s} > 0.
2.5 Physical Realizability Conditions

and Parameterizations for Linear
Dynamical Quantum Systems
2.5.1 Physical Realizability Conditions for Linear QSDEs
We have now seen that the dynamics of linear quantum stochastic systems in the
Heisenberg picture leads to an equation of motion in the form of linear QSDEs, i.e., as
given by Eq. (2.5) in the quadrature form and (2.12) in the mode form. These equations
can be viewed as a quantum analogue of the equations that describe classical linear
stochastic systems. However, unlike classical linear stochastic systems, the matrices
A, B, C, D in (2.5) or the matrices Ã, B̃, C̃, D̃ in (2.12) cannot be arbitrary, but are
constrained by quantum mechanics so that these equations represent the evolution of
a valid (open) quantum system. When the linear QSDE represents a valid quantum
system, it is said that the QSDE or the system represented by the QSDE is physically
realizable. To avoid unnecessary repetition, a formal definition will be given below
for systems in the quadrature form. This definition can be directly adapted to systems
in the mode form due to the one-to-one relationship between the two forms.
Definition 2.2 (Physical realizability) A system of linear quantum stochastic dif-

ferential equations of the form (2.5) is said to be physically realizable if there exists
a quadratic Hamiltonian H = (1/2)x Rx, a coupling operator L = Kx, and a Bogoli-
ubov matrix W such that
A = 2Jn (R + {K ∗ K}), (2.24)

B = 2ıJn [ −K ∗ K ]W m , (2.25)

K + K#
C = Pm , (2.26)
−ıK + ıK #
D = m−1 W m . (2.27)
Essentially, the definition states that if a system of QSDEs (2.5) is physically realiz-
able then there is a linear quantum system G = (W, L, H) for which the Heisenberg
evolution of x is given by (2.5). G is, of course, by definition, a physical system. We
have the following result [16, 24, 25]
Theorem 2.1 A system of linear quantum stochastic differential equations of the

form (2.5) is physically realizable if and only if
AJn + Jn A + BJm B = 0, (2.28)

Jn C + BJm D = 0, (2.29)
DJm D = Jm . (2.30)
Proof Note that the condition (2.30) is equivalent to the statement that D is a sym-
plectic matrix (hence it is invertible with D−1 also symplectic). Without loss of
generality, we may reduce the proof to the case with D = I (i.e., W = I) for the fol-
lowing reasons. Let dyo = D−1 Cx(t)dt + dw(t), then y(t) = Dyo (t). Let C = D−1 C
and consider the linear quantum stochastic system,
dx(t) = Ax(t)dt + Bdw(t),

dyo (t) = C x(t)dt + dw(t).
We will show that this system is physically realizable if and only if (2.28)–(2.30)
is satisfied with C replaced with C and D = I. From this it easily follows that the
original system (2.5) is physically realizable if and only if (2.28)–(2.30) holds. This
is because y is a symplectic transformation of yo , so if the system with output y is
physically realizable so is the system with output yo , and vice-versa. Thus for the
remainder of the proof we identify C with C and D with I.
To this end, let us first show that physical realizability of the system G =
(A, B, C, I) implies (2.28)–(2.30). Indeed, physical realizability implies that the
canonical commutation relation for x(t),
x(t)x(t) − (x(t)x(t) ) = 2ıJn ,
and the differential commutation relation for y(t),
dy(t)dy(t) − (dy(t)dy(t) ) = 2ıJn dt,
must hold for all times t ≥ 0. This follows from the fact that the joint evolu-
tion of the system and the bosonic fields is unitary (given by the solution of a
Hudson–Parthasarathy QSDE). Since x(0) = x satisfies xx − (xx ) = 2ıJn ,
for the commutation relation to hold for all x(t), t ≥ 0, we must have
d(x(t)x(t) − (x(t)x(t) ) ) = 0.
2.5 Physical Realizability Conditions and Parameterizations … 57
The condition (2.28) follows from this by substituting the expression dx(t) =
Ax(t)dt + Bdw(t) and using the quantum Itō product rule, and the Itō table to
explicitly compute the differential. By the same procedure and using the expres-
sion dy(t) = Cx(t)dt + dw(t), we obtain (2.29). The details of the calculations can
be found in [16, 26].
We now consider the converse implication, that (2.28) and (2.29) imply that there
exist R = R ∈ R2n×2n and K ∈ Cm×2n such that (2.24)–(2.27) holds. Indeed,
define R = 14 (−Jn A + A Jn ) and K = − 2ı [ 0m×m Im×m ]m−1 B Jn ; see [16, 26]
for details of the origin of these expressions. Using these definitions and (2.28),
we can directly calculate that Jn (R + {K ∗ K}) = Jn R + Jn {K ∗ K} = A and
2ıJn [ −K ∗ K ]n = B. Using the last expression for B and (2.29), we then also
K + K#
get that C = Pm . This completes the proof.
−ıK + ıK #
It should be noted from the above theorem that, remarkably, physical realizability
is equivalent to preservation of canonical commutation relations for x(t) and y(t).
The notion of physical realizability and the physical realizability constraints on the
systems matrices were first introduced in [16, 26] for the quadrature form.
2.5.2 Parameterization of Linear Dynamical Quantum

Systems
From the exposition in the preceding sections, we see that there are two sets of
parameters that can be used to parameterize linear quantum stochastic systems with
the same number of inputs and outputs. The first is the set of three parameters
(W, K, R), with W a 2n × 2n Bogoliubov matrix, K an m × 2n complex matrix,
and R a real symmetric 2n × 2n matrix, that describe the physical parameters of the
system including the Hamiltonian H and the linear coupling operator L. The other
parameterization is via the system matrices (A, B, C, D) that appear in the QSDE for
the Heisenberg evolution (2.5) of the system.
The two parameterizations (W, K, R) and (A, B, C, D) are equivalent, and can
be used interchangeably, according to which one may be more convenient for the
purpose at hand. This is because of a bijective correspondence between the two para-
meterizations: to any given (W, K, R) parametrization there corresponds a unique
(A, B, C, D) parameterization, and vice-versa. The (A, B, C, D) matrices for a given
set of (W, K, R) matrices are given by (2.6)–(2.9) (recall that the class of symplectic
matrices have a bijective correspondence with the class of Bogoliubov matrices). On
the other hand, the (W, K, R) parameterization can be obtained from a given set of
(A, B, C, D) matrices following the proof of Theorem 2.1.
2.5.3 Linear Dynamical Quantum Systems with Less Outputs

Than Inputs
In general, not all the outputs of a system may be of interest or can be observed
or utilized. In many situations, one may only be interested in certain pairs of the
output field quadratures in y(t) [16]. In the most general scenario, one can consider
y(t) having an even dimension 2ny < 2m and D is a 2ny × 2m matrix satisfying
DJm D = Jny . That is, we can consider outputs of the form:
dy(t) = Cx(t)dt + Ddw(t), (2.31)
with C ∈ R2ny ×2n , D ∈ R2ny ×2m with ny < m. Generalizing the notion discussed in
Sect. 2.5.1 (for the case where there are as many outputs as there are inputs), a linear
quantum system with output (2.31) is physically realizable if and only if there exist
matrices C ∈ R2(m−ny )×2n and D ∈ R2(m−ny )×2m such that the system
dx(t) = Ax(t)dt + Bdw(t); x(0) = x,

C D
dy (t) = x(t)dt + dw(t), (2.32)
C D
is physically realizable with the same number of inputs and outputs. Therefore the
matrices A, B, [ C (C ) ] , and [ D (D ) ] satisfy the constraints (2.28)–(2.30)
when C and D in (2.29) and (2.30) are replaced by [ C (C ) ] and [ D (D ) ] ,
respectively. We can easily obtain the following necessary and sufficient condition
for physical realizability of linear quantum systems with less outputs than inputs see,
e.g., [25].
Theorem 2.2 A linear quantum stochastic system with less outputs than inputs is
physically realizable if and only if
AJn + Jn A + BJm B = 0, (2.33)

Jn C + BJm D = 0, (2.34)
DJm D = Jny . (2.35)
Proof The necessity of (2.33)–(2.35) is immediate from the definition of physically

realizable systems with less outputs than inputs given above and the physical real-
izability conditions for systems with the same number of input and outputs. For the
sufficiency, first note that for D satisfying (2.35), it follows from a construction used
in the proof of [27, Lemma 6] that a matrix D ∈ R2(m−ny )×2m can be constructed such
that the matrix D̃ = [ D (D ) ] is symplectic. Now, define C = D Jm B Jn and
C̃ = [ C (C ) ] . Consider now a system G̃ with an equal number of inputs and
outputs and system matrices (A, B, C̃, D̃). From the physical realizability conditions
(2.33)–(2.35) and the definition of C and C̃, it follows that G̃ satisfies (2.28)–(2.30)
and is therefore physically realizable. It now follows from definition that the original
2.5 Physical Realizability Conditions and Parameterizations … 59
system with output y(t), of smaller dimension than w(t), is physically realizable.
This completes the proof.
Thus the physical realizability constraints for systems with the same number of
outputs as inputs, or less inputs than outputs, have essentially the same form, so the
following corollary is immediate.
Corollary 2.1 An arbitrary linear quantum stochastic system is physically realiz-
able if and only if
AJn + Jn A + BJm B = 0, (2.36)

Jn C + BJm D = 0, (2.37)
DJm D = Jny . (2.38)
Remark 2.2 All the results relating to linear quantum systems in this chapter depart
from the assumption that the position and momentum operators satisfy the default
CCR given by (2.3). However, these operators may be defined differently and satisfy
a different CCR. For instance, one may √ wish to take √ as the position and momentum
pair for the j-th oscillator qj = qj / 2 and pj = pj / 2 satisfying [qj , pj ] = ı, and
defining x = (q1 , p1 , q2 , p2 , . . . , qn , pn ) as the canonical internal operator that will
satisfy the CCR [x , (x ) ] = ıJn . More generally, one may take x = Tx satisfying
[x , (x ) ] = 2ıTJn T , with an invertible real 2n × 2n matrix T that would typically
be diagonal. All of the structural results for linear quantum systems that have been
obtained can be adapted to this different choice of CCR by making the substitutions
(given here for the real quadrature form):
TAT−1 → A, TB → B, CT−1 → C,
TJn T → Jn , T− RT−1 → R, KT−1 → K,
while the system matrix D and scattering matrix S remain the same.
2.6 Stability of Linear Quantum Systems
As with classical linear systems, one can define a notion of stability for linear quantum
systems. We will do this for systems in quadrature form but analogous notions and
results also hold for systems in complex mode form. Consider a linear quantum
system in quadrature form driven by fields that are all in the vacuum state. Then
quantum expectation of x(t) is given by the ODE
˙ = Ax(t).
x(t)
The system is said to be:

• Asymptotically stable or simply stable if |x(t)| → 0 as t → ∞ for any initial
state of the system.
• Marginally stable if |x(t)| does not go to 0 as t → ∞, but remains bounded at
all times t ≥ 0 for any initial state of the system.
• Unstable if there exists some initial state of the system such that |x(t)| → ∞ as
t → ∞.
From standard linear state-space theory, we immediately have necessary and suf-
ficient conditions for a system to be (asymptotically) stable, marginally stable and
unstable. These are:
• A system is stable if and only A is Hurwitz. That is, all its eigenvalues have a
negative real part.
• A system is marginally stable if and only if A has one or more eigenvalues on
the imaginary axis with geometric multiplicity equal to its algebraic multiplicity
while all other remaining eigenvalues have negative real part.
• A system is unstable if and only if A has at least one eigenvalue with positive real
part or one eigenvalue on the imaginary axis with geometric multiplicity less than
its algebraic multiplicity.
One can associate a natural energy functional to a linear quantum system. This energy
functional E(x(t)) is defined by

n
E(x(t)) = qj (t)2 + pj (t)2 = Tr(P(t)),
j=1
where P(t) = (1/2) x(t)x(t) + (x(t)x(t) ) . Note that P(t) = P(t) and
P(t) ≥ 0. Suppose that the system has some initial energy but not receiving any
energy from the external input fields (all are in the vacuum state). Then P(t) satisfies
the Lyapunov differential equation:
Ṗ(t) = AP(t) + P(t)A + BB ,
with initial condition P(0) = (1/2)xx + (xx ) . Notice that when the system
is stable, A is Hurwitz and limt→∞ P(t) = P∞ exists, is unique, and given by the
unique solution to the algebraic Lyapunov equation,
AP∞ + P∞ A + BB = 0.
Thus, when the system is stable the system’s energy in the long time settles to a
steady-state value. This is a necessity in applications as unbounded energy growth
leads to the system being damaged or destroyed, so ensuring stability is an important
goal in the control of linear quantum systems.
2.7 Gaussian States 61
2.7 Gaussian States
The oscillators and the bosonic fields to which they can be coupled are defined on
an infinite-dimensional Hilbert space. For the oscillator this is the Hilbert space
⊗nj=1 L 2 (R; C) ≡ L 2 (Rn ; C), where n is the number of oscillators, while for the
bosonic fields it is the Fock space Fm = s (L 2 (R+ ; Cm )), where m is the multiplicity
space of the field, i.e., the number of distinct bosonic fields. As such, on these fields
there exist observables that have a continuous spectra which are the quantum analogue
of real-valued random variables taking values in R. Moreover, one can define states
on the oscillator or field that are the quantum analogue of the Gaussian states in
classical probability theory.
Gaussian states are defined differently for single-mode oscillators (a collection of
oscillators with a single well-defined frequency) and bosonic fields, since the latter
supports an infinite number of modes (in frequency). In this section, we provide a
definition of Gaussian quantum states, beginning with the case of a collection of
single-mode oscillators and then proceeding to define Gaussian states for bosonic
fields. The treatment here closely follows [18, 19].
Gaussian states have a special status with respect to linear quantum systems. Due
to their linear dynamics, linear quantum systems have the property that they preserve
Gaussian states. This means that if the oscillators are initialized in a Gaussian state
and input fields are in a Gaussian state, then the joint state of the oscillators and
output fields will be in Gaussian state at all times in the system’s evolution. This will
be discussed in more detail in Sect. 4.3.2 when quantum Kalman filters for linear
quantum systems are introduced.
2.7.1 Gaussian State of a Collection of Single-Mode

Oscillators
Consider a collection of n independent harmonic oscillators on the collective Hilbert

space L 2 (Rn ; C) = L 2 (R; C)⊗n , with annihilation operators a1 , a2 , . . . , an . As
before, we let a = [ a1 a2 . . . an ] , and introduce Sn to denote the set of all anni-
hilation operators for n independent oscillators, so a ∈ S(n). A quantum state on
L 2 (Rn ; C) is said to be Gaussian if
!
a
exp ı [ u∗ u ] #
a

u a
= exp −1/2[ u∗ u ]F # + ı[ u∗ u ] ,
u a#
for all u ∈ Cn and for a positive semidefinite Hermitian matrix F ≥ 0 given by

!
a − a ∗
I + N M
F= (a − a) (a − a) = (2.39)
(a − a)# M∗ N
with
" #
Njk = (aj − aj )∗ (ak − ak ), Mjk = (aj − aj )(ak − ak ) . (2.40)
We have the properties that N = N ∗ and M = M , and F ≥ 0 implies that N ≥ 0.

F is referred to as the “covariance matrix” of the Gaussian state. Note that the
definition of a Gaussian state given above is based on a quantum generalization of
the characteristic function of a classical Gaussian probability distribution, which
uniquely identifies the distribution.
Let us now consider zero-mean states satisfying aj = 0 for all j. The joint
vacuum state of the n oscillators is a special state corresponding to N = 0, M = 0,
with F = Fvac given by
I 0
Fvac = . (2.41)
00
Here the subscript vac stands for “vacuum.”

For a fixed N ≥ 0, M is constrained by the condition that F ≥ 0. When n =
1, N and M are scalars and the condition F ≥ 0 is equivalent to N ≥ 0 with
|M|2 ≤ N (N + 1). More generally, we can perform the diagonalization V ∗ N V =
diag (N1 , . . . , Nn ) for some unitary V , in which case new modes a = V a can be
defined. Then Nj can be interpreted as the average number of quanta in the mode aj .
Note that, in general, one cannot simultaneously diagonalize N and M.
2.7.1.1 Generalized Araki–Woods Representation
Given a zero-mean Gaussian state characterized by the covariance matrix F in (2.39),

modes having that state can be constructed via canonical transformations of vacuum
modes. That is, we will show that for any Gaussian state for which a has covariance
F given by (2.39), there exists a 2n × 4n matrix S̃0 such that

a a0
= S̃0 (2.42)
a# a0#

a
where a0 = 1 ∈ S(n + n) has vacuum statistics, and
a2

S̃0 S̃0 = I. (2.43)
Indeed, we will construct S̃0 = (E−0 , E+0 ) for some n × 2n matrices E−0 , E+0 . This
can be viewed as a generalization of a construction by Araki and Woods [28] for
non-squeezed thermal states, see [20, 29].
2.7.1.2 Construction of Araki–Woods Vacuum Representation
Step 1: Diagonalize N. Determine a unitary matrix V ∈ Cn×n such that V ∗ N V =

diag (N1 , . . . , Nn ), with the eigenvalues ordered such that N1 ≥ · · · ≥ Nn ≥ 0. We
can restrict our attention to the case with N diagonalized in this manner.
Step 2: Ignore zero eigenvalues. Take the first n+ eigen values to be strictly positive,
with the remaining n0 = n − n+ being zero. With respect to the eigen decomposition
Cn = Cn+ ⊕ Cn0 , decompose F as
⎡ ⎤
I + N++ 0 M++ M+0
⎢ 0 I M0+ M00 ⎥
F=⎢
⎣ M++
M+0
⎥
N++ 0 ⎦ .

M0+ M00 0 0
However, if a positive semidefinite matrix has a zero on a diagonal then every entry
on the corresponding row and column must vanish3 so that in fact,
⎡ ⎤
I + N++ 0 M++ 0
⎢ 0 I 0 0⎥
F≡⎢
⎣ M++
⎥.

0 N++ 0⎦
0 0 0 0
We can thus restrict our attention to the case where N is diagonal and positive definite
(thus invertible).
Step 3: Explicit Construction. We begin by noting the constraint I + N ≥ MN −1 M ∗ ,
which follows from
∗
I −MN −1 I −MN −1 I + N − MN −1 M ∗ 0
0≤ F = .
0 0 0 0 0 0
Now, introduce the matrices X, Y , Z defined by:

X= I + N − MN −1 M ∗ ,
√
Y = N = diag N1 , . . . , Nn ,
3 To see why this is the case, consider a complex k × k matrixE ≥ 0 with E 11 = 0. Taking
u (t) = (tx1 , x2 , . . . , xk ) we have that 0 ≤ u (t)∗ Eu (t) = 2tRe{ j>1 x1∗ E1j xj + j,k>1 xj∗ Ejk xk }.

However, for this inequality to hold for all real t it must be that { j>1 x1∗ E1j xj } = 0. Now,
∗
replacing t with ıt shows that { j>1 x1 E1j xj } also vanishes. As this is true for any xj , it must be
the case that E1j = Ej1 ∗ = 0 for all j > 1.
Z = MY −1 .
Note that Y = Y and from Z = MY −1 we have that YZ = M = M = ZY = ZY .

The matrices satisfy
XX ∗ − YY ∗ + ZZ ∗ = I and YZ = ZY . (2.44)
Now take b1 and b2 to be any two distinct modes in S (n) (hence they are commuting).
Fix the state to be the joint vacuum state for these two modes. Then we have the
representation
a = Xb1 + Yb2# + Zb2 . (2.45)
Indeed, it can be straightforwardly verified that

" # #
a a = Y 2 = N,
" #
aa = ZY = ZY = M.
We have thus constructed the representation (2.42) with S̃0 = (E−0 , E+0 ) and

E−0 = X Z , E+0 = 0 Y .
Property (2.43) follows from (2.44).
2.7.2 Gaussian States of the Field and Their Fock Space

Representation
Consider m continuous-mode bosonic fields indexed by j = 1, 2, . . . , m with

annihilation field operators bj (t) satisfying the field commutation relations [bj (t),
bk (t )∗ ] = δjk δ(t − t ) and [bj (t), bk (t )] = 0. Let us introduce the shorthand nota-
tion,
b
b̆ = # .
b
A Gaussian state of a continuous-mode bosonic field is state with characteristic

function of the form
∞ !
exp ı ŭ (t)∗ b̆ (t) dt
0
∞ ∞
= exp −1/2 ŭ (t)∗ F ŭ (t) dt + ı ŭ(t)∗ b̆(t)dt ,
0 0
for any u ∈ L 2 (R+ ; Cm ). Here, as in the single-mode case, F is a 2m × 2m Hermitian

matrix that is again of the form (2.39) with the entries of N and M specified by the
correlation functions
∗ !
bj (t) − bj (t) bk (t ) − bk (t ) = Njk δ(t − t ),
!
bj (t) − bj (t) bk (t ) − bk (t ) = Mjk δ(t − t ).
That is, ∗!

b̆(t) − b̆(t) b̆(t ) − b̆(t ) ≡ Fδ t − t . (2.46)
Following the treatment in [18], we now consider a zero-mean Gaussian state

ωN,M (·) = · of n continuous-mode bosonic fields with b̆(t) = 0 for t ≥ 0.
An important special case is a joint vacuum Gaussian state, when N = 0 and M = 0.
This is the state when the fields are all empty (contain no photons). The vacuum state
for the field is completely specified by the characteristic function,
∞ ! ∞
exp ı ŭ (t)∗ b̆ (t) dt = exp −1/2 ŭ (t)∗ Fvac ŭ (t) dt ,
0 vac 0
for any u ∈ L 2 (R+ ; Cm ). For concreteness, let us use this definition to show that any
component wj (t) of the quadrature noise w(t) is a standard Wiener process under
the vacuum state, as claimed in Sect. 2.3.2. First, from its definition we immediately
have that (i) wj (0)2 = 0 and wj (t) = 0. Now, take uλ = λej 1[t1 ,t2 ] with λ ∈ R and
0 ≤ t1 ≤ t2 (recall that ej is a unit column vector with all entries 0 except for a 1 as
the j-th component). From the definition of a vacuum state given above,
" #
exp ıλ(wj (t2 ) − wj (t1 ) vac
t2 !
= exp ı dwj (t)
t1
∞ vac
!

= exp ı uλ (t) (b(t) + b(t) )dt
#
0
∞ ! vac
= exp −1/2 uλ (t) 2 dt

0 vac
= exp −1/2λ2 (t2 − t1 ) , ∀λ ∈ R.
Therefore, (ii) the increment wj (t2 ) − wj (t1 ) has the characteristic function of a zero-
mean Gaussian random variable with variance t2 − t1 . Moreover, by adaptedness
we have that (iii) (w(t4 ) − w(t3 ))(w(t2 ) − w(t1 )) = (w(t4 ) − w(t3 ))(w(t2 ) −
w(t1 )) = 0 for all t1 < t2 < t3 < t4 , so that increments at disjoint time intervals
are uncorrelated. Since [wj (t), wi (s)] = 0 for all s, t ≥ 0, from (i)–(iii) we conclude
that wj (t) has precisely the statistics of a standard Wiener process and is therefore a
Fock space realization of the latter.
For field operators ∞b1 (t),∗b2 (t), . . . , bm (t) in a zero-mean jointly Gaussian state,
we define B (
j∞ f ) = 0 f (s) b j (s)ds for any f ∈ L 2
(R+ ; C) and its adjoint process
Bj∗ ( f ) = 0 f (s)bj (s)∗ ds as they are well-defined and more regular mathemati-
cal objects, and can be manipulated using the Hudson–Partharathy quantum sto-
chastic calculus. We introduce these notations to distinguish them from the earlier
notations Aj ( f ) and A∗j ( f ) that were assigned for the special case of a vacuum
field state, if field j is in the vacuum state then one can simply set Bj ( f ) to be
Aj ( f ). The operators Bj ( f ) and Bj∗ ( f ) satisfy the same commutation relations as
∞
A( f ) and A∗ ( f ), [Bj ( f ), Bk∗ (g)] = δjk 0 f (s)∗ g(s)ds, and concrete realizations of
B( f ) = (B1 ( f ), B2 ( f ), . . . , Bm ( f )) and B( f )# = (B1∗ ( f ), B2∗ ( f ), . . . , Bm∗ ( f ))
on a suitable Hilbert space are dependent on the state of the field. However, for arbi-
trary Gaussian states one can relate the associated realizations of B( f ) and B ∗ ( f ) to
the vacuum state representation of these operators, via the generalized Araki–Woods
representation from the previous section. In particular, any operator B( f ) associated
with a zero-mean Gaussian state ωN,M has a Fock space representation of the form
B( f ) = XA1 ( f ) + Y A2 ( f )# + ZA2 ( f ), (2.47)
with complex m × m matrices X, Y , and Z as constructed in the previous section,

determined by the values of the parameters N and M of ωN,M , and where A1 ( f )
and A2 ( f ) are two independent vacuum annihilation processes that can each be
realized on a distinct copy of the Fock space Fm . Note that by construction
t X, Y , Z
guarantee that the commutation relations [Bj ( f ), Bk∗ (g)] = δjk 0 f (s)∗ g(s)ds hold
for any f , g ∈ L 2 (R+ ; C).
The Itō table for a jointly Gaussian state of the fields can be directly constructed
by exploiting the generalized Araki–Woods representation (2.47) and the vacuum Itō
table. Recall that A1 ( f ) and A2 ( f ) in (2.47) are vacuum representations on distinct
copies of the Fock space Fm . The extended Itō table for the integrated operators
Bj (t) = Bj (1[0,t] ) and Bj∗ (t) = Bj∗ (1[0,t] ) when the field is in an arbitrary Gaussian
state is then
× dBk∗ dBk
dBj (δjk + Nkj )dt Mjk dt (2.48)
dBj∗ Mkj∗ dt Njk dt.
Note that in general Gaussian states, the counting process need not be defined. We
can also define a QSDE of the Hudson–Parthasarathy type but in which the vacuum
field operators A and A# are replaced by field operators B and B # corresponding to
a non-vacuum zero-mean jointly Gaussian state of the field. This yields the QSDE
(without the counting process ),
dU(t) = −(ıH + 1/2(L ∗ (I + N )L + L NL # − L ∗ ML # − L M # L))dt

+ dB(t)∗ L − L ∗ dB(t) U(t) (2.49)
with initial condition U(0) = I. As with the vacuum case, after interaction with
the system, B is transformed to B out according to B out (t) = U(t)∗ B(t)U(t). Using
(2.47), the QSDE can be expressed in terms of the vacuum operator A(t) =
[ A1 (t) A2 (t) ] ,
∗
dU(t) = (−(ıH + LN,M LN,M /2)dt + dA(t)∗ LN,M − LN,M
∗
dA(t))U(t),
(2.50)
with
X ∗L
LN,M = .
Z ∗ L − Y L#
Example 2.3 An example of a non-vacuum zero-mean Gaussian state of a single

bosonic field is a squeezed vacuum field state where the parameters M = 0 and
N = 0 satisfy the identity |M|2 = N(N + 1). Such a field can be approximately
produced in the laboratory using the DPA that was treated in Example 2.2. In the
limit discussed in that example, where the DPA becomes idealized with an infinite
bandwidth, its output is an ideal squeezed Gaussian field state satisfying the quantum
Itō rule with
4κ0 0
N = sinh2 r0 = ,
(κ20 − 0)
2 2
2κ0 0 (κ20 − 20 )
M = cosh r0 sinh r0 = .
(κ20 − 20 )2
It can be easily verified that the parameters M and N above satisfy |M|2 = N(N + 1).
In practice, treating the output of the DPA as a squeezed bosonic field will be valid as
long as it is driving a quantum system with sufficiently slower dynamics (sufficiently
lower bandwidth) than the DPA itself.
2.7.3 Coherent States
The Gaussian states that we have discussed so far have a common feature, they all
have zero mean. We shall now discuss an important class of nonzero-mean Gaussian
states that model the output beam from a laser and can facilitate quantum feedback
control. These are the coherent states, and we will now give describe the coherent
states of single-mode oscillators and bosonic fields.
2.7.4 Coherent States of a Single-Mode Oscillator
A coherent state of a single-mode oscillator is a pure state of the oscillator that

is indexed by a complex number α and denoted by |α. The coherent state |α is
a normalized eigenvector of the oscillator’s annihilation operator corresponding to
the eigenvalue α, a|α = α|α. In a coherent state, the annihilation operator has
mean a = α|a|α = α and mean photon number of α|a∗ a|α = |α|2 , while the
complex “covariance” matrix F defined in (2.39) corresponds to that of a vacuum
state, with N = M = 0. Therefore, the vacuum state of an oscillator is a particular
coherent state with α = 0.
A coherent state with α = 0 can be generated form the vacuum state |0 by apply-
ing a unitary displacement operator D(α) = exp (αa∗ − α∗ a), |α = D(α)|0. We
have that D(α)∗ aD(α) = α + a, and we notice the duality between the Schrödinger
D(α)
picture state transformation |0 → |α and Heisenberg picture operator transfor-
D(α)∗ (·)D(α)
mation b → b + αI, α|a|α = 0|a + α|0 for all α ∈ C. Therefore, α is
also referred to as the amplitude of the coherent state.
2.7.5 Coherent States of a Bosonic Field
Coherent states of a bosonic field are directly related to the exponential vectors and
can be interpreted in analogous way to the coherent states of single-mode oscillator.
The coherent state of a single bosonic field is a pure state of the field that is indexed
by a function f ∈ L 2 (R+ ; C) and denoted by | f . It is simply a normalized version
of the exponential vector e( f ),
e( f )
|f = = e( f ) exp − f 2
/2 .
e( f )
It follows that that A(g)|f = g, f |f , thus |f is an eigenvector of A(g) for any
g ∈ L 2 (R+ ; C) corresponding to the eigenvalue g, f . More formally, in terms of the
annihilation field operator b(t) (satisfying [b(t), b(t )∗ ] = δ(t −t ) and [b(t), b(t )] =
0), we have that b(t)|f = f (t)|f for all t ≥ 0, mirroring the property of coherent
states of a single-mode oscillator.
The analogue of the displacement operator D(α) for bosonic fields is the Weyl
operator W (g) defined through its action on the exponential vectors,
W (g)e( f ) = exp(−ıg, f − f 2
/2 − g 2 /2)e(f + g).
From its definition, W ( f ) is a unitary operator on s (L 2 (R+ , C)) and we have that
W ( f )| = |f , where | = e(0) is the vacuum state of the bosonic filed. Therefore,
any coherent state can be generated from e(0) via W ( f ). In the bosonic case, we also
W (g)
have the duality between the Schroedinger state transformation | → |f , with
the Heisenberg picture transformation
W (g)∗ (·)W (g)

A(g) → W (g)∗ A( f )W (g), f |A(g)|f = 0|W (g)∗ A( f )W (g)|0.
Note that W ( f ) = W (ft] ) ⊗ W (f[t ). The Weyl operator W (ft] ) with time index t is
an adapted process and satisfies the QSDE,
dW (ft] ) = −1/2|f (t)|2 dt + f (t)dA(t)∗ − f (t)∗ dA(t) W (ft] ), W (f0] ) = I.
Thus we have that W ( f ) = limt→∞ W (ft] ). The time-indexed Weyl operator is

realized in the laboratory by an electro-optic modulator (EOM) to produce the state
W (ft] )|. In the Heisenberg picture, the EOM implements the displacement b(t) →
b(t) + f (t)I, where b(t) is a vacuum bosonic annihilation operator. Coherent states
of a single-mode oscillator can be created in the steady-state by driving a stable
linear quantum system with a coherent bosonic field. In the laboratory, this would
be implemented by driving the inputs of the system with EOMs.
The treatment above can be easily extended to multiple quantum fields that are
each in a coherent state. If there are m fields each in a coherent state with amplitudes
f1 , f2 , . . . , fm then we replace the Hilbert space L 2 (R+ ; C) with L 2 (R+ ; Cm ), replace
f with the vector f = (f1 , f2 , . . . , fm ) , and generalize the Weyl operator W (ft] )
according to the solution of the QSDE,
dW (ft] ) = −1/2 f (t) 2 dt + dA(t)∗ f (t) − f (t)∗ dA(t) W (ft] ), W (f0] ) = I.
Here A(t) is now a vector of m annihilation processes, one for each field as before.
Since coherent fields can be generated from the vacuum, a system driven by
coherent fields can be analyzed by referring it back to the vacuum. Suppose the
system G = (S, L, H) is driven by coherent fields with amplitude vector f . Then we
can take all fields to be in the vacuum state and consider the joint unitary evolution
of the system and vacuum fields given by Uf (t) = U(t)W (ft] ), where U(t) is the
unitary evolution of G given by the Hudson–Parthasarathy QSDE. By applying the
quantum Itō stochastic calculus, it is easy to show that Uf (t) satisfies the QSDE,
dUf (t) = − ı(H + {L ∗ Sf (t)}) + 1/2(L + f (t)I)∗ (L + f (t)I) dt

+dA(t)∗ (L + f (t)I) − (L + f (t)I)∗ SdA(t)
+Tr (S − I)d(t) Uf (t),
with initial condition Uf (0) = I. The vacuum driven system is thus G vac = (S, L +
fI, H + {L ∗ Sf }). Let jt (·) = U(t)∗ · U(t) as in Sect. 2.1.4 and also let jt (·) =
f
∗
Uf (t) · Uf (t). The equivalence between G and the vacuum referred system G vac is
in the sense that
f
f |jt (X)|f = |jt (X)|, ∀X ∈ B(h) and ∀t ≥ 0.
That is, for all t ≥ 0, the quantum expectation of jt (X) under the coherent field |f
f
coincides with the quantum expectation of jt (X) under the vacuum state, for any
bounded operator X on h.
Finally, we remark that the QSDEs for the Weyl operator and G vac can have time-
dependent parameters, S(t), L(t) and H(t). For instance, in the case of G vac , we have
L(t) = L + f (t)I, and H(t) = H + {L ∗ Sf (t)}, while S is not time-dependent. This
is a slightly more general model than the one presented in Sect. 2.1.4. However, this
extension is valid and defines a unitary solution of the QSDE when S(t), L(t), and
H(t) are adapted processes that are bounded for each t, H(t) self-adjoint and S(t)
unitary for all t ≥ 0 [30].
References
1. K.R. Parthasarathy, An Introduction to Quantum Stochastic Calculus (Birkhauser, Berlin, 1992)

2. P.A. Meyer, Quantum Probability for Probabilists, 2nd edn. (Springer, Berlin, 1995)
Optim. 46, 2199–2241 (2007)
4. R.L. Hudson, K.R. Parthasarathy, Quantum Ito’s formula and stochastic evolution. Commun.
Math. Phys. 93, 301–323 (1984)
5. C. Gardiner, M. Collett, Input and output in damped quantum systems: quantum stochastic
differential equations and the master equation. Phys. Rev. A 31, 3761–3774 (1985)
2004)
7. A. Silberfarb, I. Deutsch, Continuous measurement with travelling-wave probes. Phys. Rev. A
68, 013817 (2003)
8. F. Fagnola, On quantum stochastic differential equations with unbounded coefficients. Probab.
Theory Relat. Fields 86, 501–516 (1990)
9. J. Gough, M.R. James, The series product and its application to quantum feedforward and
feedback networks. IEEE Trans. Autom. Control 54(11), 2530–2544 (2009)
10. M. Gregoratti, The Hamiltonian operator associated to some quantum stochastic differential
equations. Commun. Math. Phys. 254, 489–512 (2001)
11. M. Gregoratti, Erratum: The Hamiltonian operator associated with some quantum stochastic
evolutions. Commun. Math. Phys. 264, 563–564 (2006)
12. A.M. Chebotarev, Quantum stochastic differential equation is unitarily equivalent to a sym-
metric boundary problem for the Schrödinger equation. Math. Notes 61(4), 510–518 (1997)
13. E. Wong, B. Hajek, Stochastic Processes in Engineering Systems (Springer, New York, 1985)
14. S.C. Edwards, V.P. Belavkin, Optimal quantum filtering and quantum feedback control (2005),
arXiv preprint, http://arxiv.org/pdf/quant-ph/0506018
16. M.R. James, H.I. Nurdin, I.R. Petersen, H ∞ control of linear quantum stochastic systems. IEEE
Trans. Autom. Control 53(8), 1787–1803. Reprinted, with permission, © 2008 IEEE (2008)
17. H.M. Wiseman, G.J. Milburn, Quantum Measurement and Control (Cambridge University
References 71
18. J.E. Gough, M.R. James, H.I. Nurdin, Squeezing components in linear quantum feedback
networks. Phys. Rev. A 81, 023804. Reprinted, with permission, © (2010) by the American
Physical Society (2010)
19. H.I. Nurdin, Quantum filtering for multiple input multiple output systems driven by arbitrary
zero-mean jointly Gaussian input fields. Russ. J. Math. Phys. 21(3), 386–398. © 2014 Springer.
Reprinted with permission of Springer (2014)
20. H. Hellmich, R. Honegger, C. Köstler, B. Kümmerer, A. Rieckers, Couplings to classical and
non-classical squeezed white noise as stationary Markov processes, vol. 38 (Publ. RIMS, Kyoto
Univ, 2002), pp. 1–31
21. J. Gough, R. Gohm, M. Yanagisawa, Linear quantum feedback networks. Phys. Rev. A 78,
061204 (2008)
22. A.I. Maalouf, I.R. Petersen, Bounded real properties for a class of annihilation-operator linear
quantum systems. IEEE Trans. Autom. Control 56(4), 786–801 (2011)
23. J. Gough, G. Zhang, On realization theory of quantum linear systems. Automatica 59, 139–151
(2015)
chastic systems. SIAM J. Control Optim. 48(4), 2686–2718 (2009)
25. H.I. Nurdin, Structures and transformations for model reduction of linear quantum stochastic
systems. IEEE Trans. Autom. Control 59(9), 2413–2425. Reprinted, with permission, © 2014
IEEE (2014)
26. H.I. Nurdin, Topics in classical and quantum linear stochastic systems, Ph.D. dissertation, The
Australian National University (2007)
27. H.I. Nurdin, Network synthesis of mixed quantum-classical linear stochastic systems, in Pro-
ceedings of the 2011 Australian Control Conference (AUCC). Engineers Australia (2011), pp.
68–75
28. H. Araki, E.J. Woods, Representations of the canonical commutation relations describing a
nonrelativistic infinite free Bose gas. J. Math. Phys. 4, 637–662 (1963)
29. J. Gough, Quantum white noise and the master equation for Gaussian reference states. Russ.
J. Math. Phys. 10(2), 142–148 (2003)
Chapter 3
Realization Theory for Linear Dynamical
Quantum Systems
Abstract This chapter presents a realization/network synthesis theory for linear

quantum systems. This theory addresses the question of how one can go from an
abstract description of a linear quantum system to a concrete realization of the sys-
tem using quantum optical devices. Two distinct types of realization problems are
introduced and treated: the strict (or hard) realization problem and the transfer func-
tion (or soft) realization problem. The system to be realized is decomposed into
simpler subsystems, and how these subsystems can be realized, at least approxi-
mately, in the quantum optical setting is developed. In particular, it is shown that
simpler realizations can be obtained for completely passive linear quantum systems.
In the previous chapter, the class of linear quantum systems was introduced, and
the notion of physical realizability and parameterizations for this class has been
elaborated upon. A natural question that follows from the theory of Chap. 2, and
which is the theme of this chapter, is, given an abstract mathematical description
of a linear quantum system in terms of linear QSDEs, what physical system would
it correspond to? In other words, what physical system would realize an abstractly
defined linear quantum system? To motivate this question further, we begin with a
brief overview of aspects of linear electrical network synthesis that are relevant to
our purposes in this chapter.
Well known in electrical engineering, and in systems and control theory in par-
ticular, is that a classical (continuous time, causal, linear time invariant) electrical
network described by a set of (coupled) linear ordinary differential equations can
be represented in different ways, for example with a frequency domain or trans-
fer function representation, a modern state-space representation and, more recently,
using Willems’ behavioral representation [1]. In the case where the system is ini-
tially at rest (all initial conditions are zero), one can switch between the transfer
The Introduction, Sect. 3.1 and the associated appendices contain materials adapted from [3]
Copyright © 2009 Society for Industrial and Applied Mathematics. Reprinted with permission.
All rights reserved.
Section 3.2 and the associated appendices contain some materials reprinted, with permission,
from [13] © 2010 IEEE.
Section 3.3 contain materials reprinted, with permission, from [16] © 2010 IEEE.

74 3 Realization Theory for Linear Dynamical Quantum Systems
function and state-space representations for any given network. However, although
one can associate a unique transfer function representation to a state-space represen-
tation, the converse is not true; for a given transfer function, there are infinitely many
state-space representations. The state-space representation can be made to be unique
(up to a similarity transformation of the state-space matrices) by requiring that the
representation be of minimal order (i.e., the representation is both controllable and
observable). The synthesis or realization problem in linear electrical networks the-
ory deals with the inverse scenario where one is presented with a transfer function
or state-space description of a linear system and would like to synthesize or build
such a system from various linear electrical components such as resistors, capaci-
tors, inductors, and op-amps. A particularly advantageous feature of the state-space
representation, since it is given by a set of first-order ordinary differential equations,
is that how a system can be systematically synthesized can be inferred directly from
the representation. For instance, consider the following state-space system,

d x(t) 2 5 1
= x(t) + u(t) (3.1)
dt −2 −4 0.1

y(t) = 0 1 x(t) + u(t),
where x(t) is the state, u(t) is the input signal, and y(t) is the output signal. In an
electrical circuit, u(t) could be the voltage at certain input ports of the circuit and
y(t) could be the voltage at another set of ports of the circuit, different from the
input ports. This system can be implemented according to the schematic shown in
Fig. 3.1. This schematic can then be used to implement the system at the hardware
level, Fig. 3.2 [2, Chap. 13]. However, linear electrical network synthesis is a mature
subject that deals with much more than just how one can obtain some realization of a
particular system. For instance, it also addresses the fundamental issues such as how
a passive network, a network that does not require an external source of energy, can
also be synthesized using only passive electrical components. The primary objective
in this chapter is to develop an analogously systematic method for synthesizing
arbitrarily complex linear quantum systems described by a linear QSDE. A general
-4
u(t) x1(t) x2(t)

+ -2 +
0.1
y(t)
+
Fig. 3.1 Schematic for the implementation of the classical system (3.1). Figure adapted from [3]
© 2009 Society for Industrial and Applied Mathematics. Reprinted with permission. All rights
reserved
3 Realization Theory for Linear Dynamical Quantum Systems 75
1 1
1/3
1
u(t) 1
1/2 1
1/2
10
1 1/4
1
1
1
1
1 y(t)
Fig. 3.2 Hardware implementation of the schematic diagram shown in Fig. 3.1. Figure adapted
from [3] © 2009 Society for Industrial and Applied Mathematics. Reprinted with permission. All
rights reserved
theory for static linear quantum networks that do not contain dynamical components
was developed in [4], whereas the synthesis theory for linear dynamical quantum
systems was initiated in [3]. In the context of linear quantum systems, there are two
distinct types of realization problems that can be considered:
1. Strict or hard realization. This is the problem of realizing the parameters (S, K , R)
of a given linear quantum system G.
2. Transfer function or soft realization. This is the problem of realizing the transfer
function G [·] of a quantum system G.
Notice that a solution to the strict realization problem also solves the transfer function
realization problem, but the converse is false. In the soft problem, there are additional
degrees of freedom that can be exploited to achieve the realization. Namely, there
is freedom to transform to a different collection of internal degrees of freedom that
may simplify the realization problem. This will be made clear in Sect. 3.3. We begin
by considering the strict realization problem for linear quantum systems, following
the development in [3].
3.1 Architecture for Strict Realization
3.1.1 The Concatenation and Series Product and Reducible

Quantum Networks
We first recall the formalisms of concatenation product, series product, and reducible
networks (with respect to the series product) that were introduced in [5] for the
manipulation of networks of general open Markov quantum systems (in the limit of
zero time delays for fields to propagate between network components).

Let G 1 = (S1 , K 1 x1,0 , (1/2)x1,0 R1 x1,0 ) and G 2 = (S2 , K 2 x2,0 , (1/2)x2,0 R2 x2,0 )
be two linear quantum systems, where xk,0 = xk (0). The concatenation product
G 1 G 2 is defined as

G 1 G 2 = S12 , (K 1 x1,0 , K 2 x2,0 ) , (1/2)x1,0

R1 x1,0 + (1/2)x2,0 R2 x2,0 ,
where

S1 0
S12 = .
0 S2
We emphasize here that the definition allows the possibility that x1,0 = x2,0 or that
some components of x1,0 coincide with those of x2,0 . If G 1 and G 2 are distinct
oscillators (i.e., the components of x1,0 act on a distinct Hilbert space to that of
the components of x2,0 ), the concatenation product simply groups together the vari-
ables of two non-interacting linear quantum systems to form a larger linear quantum
system.
One can also pass the output of a system G 1 as the input to system G 2 , as long
as both systems have the same number of input and output channels. This kind of
operation, which is the cascading or loading of G 2 onto G 1 , is represented by the
series product G 2 G 1 defined by

G2 G1 = S2 S1 , K 2 x2,0 + S2 K 1 x1,0 , (1/2)x1,0 R1 x1,0
1
+ (1/2)x2,0 R2 x2,0 + x (K ∗ S2 K 1 − K 2 S2# K 1# )x1,0 .
2ı 2,0 2
The resulting system G 2 G 1 is again a linear quantum system with scattering matrix,
coupling operator, and Hamiltonian as given by the above formula.
With concatenation and series products having been defined, we now arrive at
the notion of a reducible network with respect to the series product, which we
shall henceforth refer to more simply as just a reducible network. Suppose that
a network consists of l linear quantum systems G k = (Sk , L k , Hk ), with L k =

K k xk,0 and Hk = (1/2)xk,0 Rk xk,0 , k = 1, . . . , l. A reducible quantum network N
is simply a network that can be constructed from G 1 , G 2 , . . . , G l by specifying a
direct interaction Hamiltonian H d = j k= j+1 x j,0 R jk xk,0 (R jk ∈ R2×2 ) and a list
S = {G k G j } of series connections among G j and G k , j = k, with the condition
that a list of connections containing algebraic loops, such as {G 2 G 1 , G 3 G 2 , G 1
G 3 }, are disallowed. Here, algebraic loops refer to loops in which the field entering
a system G j as a particular input channel can return via some path as a component
of the same input to G j . The network N is again a linear quantum system and is
denoted by N = {{G k }k=1,...,l , H d , S}. If N0 = {{G k }k=1,...,l , 0, S} = (S0 , L 0 , H0 )
is a reducible network, then N is simply N0 equipped with the direct interaction
3.1 Architecture for Strict Realization 77
Hamiltonian H d , N = N0 (0, 0, H ) = (S0 , L 0 , H0 + H d ). Algebraic loops are

subtle and are beyond the scope of the theory in [5]. They can, however, be treated in
the more general framework of quantum feedback networks developed in [6]. This
section exploits the tools in [5], so the discussion is restricted to reducible networks.
Fortunately, this is in fact sufficient to develop a network synthesis theory of linear
quantum systems as will be shown in Sect. 3.1.2. A network synthesis theory can
also be developed following the more general theory of [6] and will be pursued in
Sect. 3.2.
Two important decompositions of a linear quantum system based on the series
product that will be exploited in the ensuing development are
(S, L , H ) = (I, L , H ) (S, 0, 0), (3.2)

(S, L , H ) = (S, 0, 0) (I, S ∗ L , H ), (3.3)
where (S, 0, 0) represents a static passive linear network implementing the unitary
matrix S.
3.1.2 Main Synthesis Theorem
Let us temporarily restrict our attention to the case of two distinct linear quantum
systems coupled to m independent bosonic fields, with m output channels: an n 1
degree-of-freedom linear quantum system G 1 = (S1 , L 1 , H1 ) with canonical oper-
ators x1 = (q1,1 , p1,1 , . . . , q1,n 1 , p1,n 1 ) , Hamiltonian operator H1 = (1/2)x R1 x1 ,
coupling operator L 1 = K 1 x1 , and scattering matrix S1 , and, similarly, an n 2 degree-
of-freedom linear quantum system G 2 = (S2 , L 2 , H2 ) with canonical operators
x2 = (q2,1 , p2,1 , . . . , q2,n 2 , p2,n 2 ) , Hamiltonian operator H2 = (1/2)x2 R2 x2 , cou-
pling operator L 2 = K 2 x2 , and unitary scattering matrix S2 .
Consider now a reducible quantum network N12 given by N12 = {{G 1 , G 2 }, H12 d
,
G 2 G 1 }, as shown in Fig. 3.3, where H12 is a direct interaction Hamiltonian term
d
between G 1 and G 2 given by
G H12
d
A(t) Y(t)
G1 = (S1 , L1 , H1 ) G2 = (S2 , L2 , H2 )
d . Figure adapted from

Fig. 3.3 Cascade connection of G 1 and G 2 with indirect interaction H12
[3] © 2009 Society for Industrial and Applied Mathematics. Reprinted with permission. All rights
reserved
1 ∗
d
H12 = (1/2)x1 R12 x2 + (1/2)x2 R12

x1 − (L S2 L 1 − L ∗1 S2∗ L 2 ),
2ı 2
1
= x2 R12

x1 − (L ∗2 S2 L 1 − L
2 S2 L 1 ),
# #
2ı
1
= x2 R12 − (K 2∗ S2 K 1 − K 2 S2# K 1# ) x1 ,
2ı
with R12 ∈ R2×2 and the second equality following from the fact that elements of L 1
commute with those L 2 . Also, note that the matrix 2ı1 (K 2∗ S2 K 1 − K 2 S2# K 1# ) is real.
Some straightforward algebraic manipulations (see [5] for details) then show that
f
N12 = (S2 S1 , S2 L 1 + L 2 , H1 + H2 + H12 + H12
d
),
where H12 = 2ı1 (L ∗2 S2 L 1 − L ∗1 S2∗ L 2 ). Now, let us inspect the Hamiltonian term of
f
f
N12 . After substituting in the definition of H1 , H2 , H12 d
, and H12 , we have that

R1 R12 x1
(1/2)[ x1 x2 ]
f
H1 + H2 + H12 + d
H12 = .
R12 R2 x2
Letting x = (x1 , x2 ) , S2←1 = S2 S1 , and defining

R1 R12
R= , (3.4)
R12 R2
K = [ S2 K 1 K 2 ], (3.5)
we see that
1
N12 = S2←1 , K x, x Rx . (3.6)
2
Therefore, N12 = (S2←1 , L 2←1 , H2←1 ), with S2←1 = S2 S1 , L 2←1 = K x, and

H2←1 = (1/2)x Rx. By repeated application of the above construction, we can
prove the following synthesis theorem.
Theorem 3.1 Let G be an n degree-of-freedom linear quantum system with Hamil-

tonian matrix R ∈ R2n×2n , coupling matrix K ∈ Cm×2n , and unitary scattering
matrix S ∈ Cm×m . Let R be written in terms of blocks of 2 × 2 matrices as
R = [R jk ] j,k=1,...,n , where the R jk ’s are real 2 × 2 matrices satisfying Rk j = R jk
for all j, k, and let K be written as
K = [ K 1 K 2 . . . K n ],
where, for each j, K j ∈ Cm×2 . Let G j = (S j , K̃ j x j , (1/2)x j R j j x j ) for j = 1, . . . , n

be independent linear quantum systems with canonical operators x j = (q j , p j ) , m
output fields, Hamiltonian matrix R j j , coupling matrix K̃ j , and scattering matrix
←
−k
S j . Also, define Sk j for j ≤ k + 1 as Sk j = l= j Sl = Sk · · · S j+1 S j for j < k,
Skk = Sk , and Skk+1 = Im×m , and let H d be a direct interaction Hamiltonian
given by
n−1 n
1
Hd = xk R jk − ( K̃ k∗ Sk j+1 K̃ j − K̃ k Sk
#
j+1 K̃ j ) x j .
#
j=1 k= j+1
2ı
(3.7)
∗
If S1 , . . . , Sn satisfies Sn Sn−1 · · · S1 = S and K̃ k satisfies K̃ k = Snk+1 K k for
k = 1, . . . , n, then the reducible network of harmonic oscillators N given by
N = {{G 1 , . . . , G n }, H d , {G 2 G 1 , G 3 G 2 , . . . , G n G n−1 }} is equivalent to G.
That is, G can be synthesized as a series connection G n . . . G 2 G 1 of n one
degree-of-freedom linear quantum systems, along with a suitable bilinear direct
interaction Hamiltonian involving the canonical operators of these systems. In par-
ticular, if S = Im×m (no scattering), then Sk can be chosen to be Sk = Im×m and K̃ k
can be chosen to be K̃ k = K k for k = 1, . . . , n.
Proof Let H j = (1/2)x j R j j x j , L j = K̃ j x j , and
k j−1 j−1

L ∗j L l∗ S ∗jl+1 L j
f
Hk = S jl+1 L l − , k ≥ 2.
j=2 l=1 l=1
We start with the series connection G 12 = G 2 G 1 . It is given by

f
G 12 = (S2 S1 , S2 L 1 + L 2 , H1 + H2 + H2 ).
Repeating this calculation recursively for G 123 = G 3 G 12 , G 1234 = G 4 G 123 , . . . ,

G 12...n = G n G 12...(n−1) , we finally obtain that
n n

G 12...n = Sn1 , Snk+1 L k , Hk + Hnf .
k=1 k=1
f
Now, noting that Hn may be rewritten as
n−1 n
1
Hnf = (L ∗k Sk j+1 L j − L ∗j Sk
∗
j+1 L k )
2ı j=1 k= j+1
n−1 n
1
= (L ∗k Sk j+1 L j − L
k Sk j+1 L j )
# #
2ı j=1 k= j+1
n−1 n
1
= xk ( K̃ k∗ Sk j+1 K̃ j − K̃ k Sk
#
j+1 K̃ j )x j ,
#
2ı j=1 k= j+1
where the second equality holds since L j commutes with L k whenever j = k, we

have that
n n n
Hk + Hnf + H = (1/2)
d
x j R jk xk = (1/2)x Rx,
k=1 j=1 k=1
x = (x1 , x2 , . . . , xn ) .
Therefore, if S1 , . . . , Sn and K̃ 1 , . . . , K̃ n satisfy the conditions stated in the theorem,

we find that N = {{G 1 , . . . , G n }, Hd , {G 2 G 1 , G 3 G 2 , . . . , G n G n−1 } is given
by

N = S, K x, (1/2)x Rx .
That is, N is a linear quantum system with Hamiltonian matrix R, coupling matrix
K , and scattering matrix S, and is therefore equivalent to G. This completes the proof
of the synthesis theorem.
Thus, according to the theorem, synthesis of an arbitrary n degree-of-freedom linear

dynamical quantum system is in principle possible if the following two requirements
can be fulfilled:
1. Arbitrary one degree-of-freedom linear quantum systems G = (I, L , H ) can be
synthesized. It follows from this that general one degree-of-freedom linear quan-
tum systems G = (S, L , H ) can be synthesized as G = (I, L , H ) (S, 0, 0).
2. The bilinear interaction Hamiltonian H d as given by (3.7) can be synthesized.
3.1.3 Systematic Synthesis of Linear Quantum Systems
This section details the construction of arbitrary one degree-of-freedom linear quan-
tum systems and implementation of bilinear direct interactions among the canonical
operators of these systems, at least approximately. The focus is on realization in the
quantum optical domain, using various linear and nonlinear quantum optical com-
ponents, though realizations in other physical platforms are also possible so long as
they support linear dynamics in the Heisenberg picture.
We begin with a description of some key quantum optical components that will
be required for the synthesis. This is followed by a discussion of general synthesis
of one degree-of-freedom linear quantum systems and finally by a discussion of
the implementation of bilinear direct interaction Hamiltonians between distinct one
degree-of-freedom linear quantum systems.
3.1.3.1 Essential Quantum Optical Components
Optical Cavities
An optical cavity is depicted in Fig. 3.4. It is essentially a system of fully reflecting or

partially transmitting mirrors in which light can be trapped, either bounced repeatedly
from the mirrors to form a standing wave or circulating inside the cavity (as in a ring
cavity). If partially transmitting mirrors are present, then light can escape from the
cavity, producing an output field and also acting as source of loss.
An optical cavity has a Hamiltonian of the form Hcav = ωcav a ∗ a, where ωcav is
the resonance frequency of the cavity and a = q+ı 2
p
is the (non-self-adjoint) cavity
annihilation operator (or cavity mode) satisfying [a, a ∗ ] = 1. Here, q = a + a ∗ is
the position operator of the cavity mode (also called the amplitude quadrature of
the mode), and p = −ıa + ıa ∗ is the momentum operator of the cavity mode (also
called the phase quadrature of the mode). If M is a partially transmitting mirror, then
photons escaping the cavity through this mirror can be modeled as the interaction of
M with a vacuum bosonic noise field A(t) incident at this√mirror via the idealized
Hamiltonian Hint of the form (1.24), taking j = 1 and L = κa, where κ is a positive
constant called the mirror decay rate or coupling coefficient. When there are several
leaky mirrors, then the collective escape of photons through these mirrors can be
modeled by a sum of such interaction Hamiltonians, one for each mirror and with
each mirror interacting with its own distinct incident vacuum bosonic field. The total
cavity Hamiltonian is then just the sum of Hcav and all the interaction Hamiltonians.
Note that, in general, the field incident at a transmitting mirror need not be a vacuum
field, but can be other types of fields, such as a coherent laser beam. Nonetheless, the
interaction of the cavity mode with such fields via the √ mirrors will still be governed
by (1.24) with a coupling operator of the form L = κa.
M3
M1
M2
M2 M1
Standing wave (Fabry-Perot) cavity

Ring cavity
Fig. 3.4 Two types of optical cavities: a standing wave or Fabry–Perot cavity (left) and a (three
mirror) ring cavity (right). Arrows indicate the direction of propagation of light in the cavity.
Black rectangles denote fully reflecting mirrors, while white rectangles denote partially transmitting
mirrors. Figure adapted from [3] © 2009 Society for Industrial and Applied Mathematics. Reprinted
with permission. All rights reserved
Degenerate Parametric Amplifier
An optical cavity can be equipped with a χ(2) nonlinear optical crystal and a classical
pump beam in the configuration of a degenerate parametric amplifier (DPA), follow-
ing the treatment in [7, Sect. 10.2] and as introduced earlier in Sect. 1.5.3. However,
in this chapter, DPA will be used to refer to the internal dynamics of the device in
Sect. 1.5.3 rather than its input-output dynamics. The pump beam serves as a source
of quanta that facilitates an interaction in the nonlinear crystal in which photons of
the pump beam are annihilated to create photons of the cavity mode. In an optical
cavity, such as the ring cavity shown in Fig. 3.5, a crystal can be positioned in one
arm of the cavity (e.g., in the arm between mirrors M1 and M2) and shone with
a strong coherent pump beam of (angular) frequency ω p = 2ωr , where ωr is some
reference frequency. The mirrors at the end of the arms should be chosen such that
they fully transmit light beams of frequency ω p . A schematic representation of a
DPA (a nonlinear crystal with a classical pump) is shown in Fig. 3.6.
Remark 3.1 In the remaining figures, black rectangles will be used to denote mirrors
that are fully reflecting at the cavity frequency and fully transmitting at the pump
M3
M2
M1
Nonlinear
crystal
Classical pump
beam C lassical pump
beam
Fig. 3.5 A DPA consisting of a classically pumped nonlinear crystal in a three-mirror ring cavity.
Figure adapted from [3] © 2009 Society for Industrial and Applied Mathematics. Reprinted with
permission. All rights reserved
Fig. 3.6 Schematic representation of a DPA. The white rectangle symbolizes the nonlinear crystal,
while the diagonal arrow into the rectangle denotes the pump beam. Figure adapted from [3] © 2009
Society for Industrial and Applied Mathematics. Reprinted with permission. All rights reserved
frequency (whenever a pump beam is employed), while white rectangles denote

partially transmitting mirrors at the cavity frequency.
Let a = q+ı 2
p
be the cavity mode, and let the cavity frequency be ωcav = ωr + ,
where is the detuning from the reference frequency ωr . Under the assumption
that the pump beam is intense and undepleted in the interaction, it may be treated as
being classical, in which case the crystal-pump-cavity interaction can be modeled by
the (time-varying) Hamiltonian H (t) = ωcav a ∗ a + 2ı (e−iω p t (a ∗ )2 − ∗ eıω p t a 2 ) [7,
Eq. 10.2.1], where is a complex number representing the effective pump amplitude
(i.e., the amplitude of the beam scaled by a constant of the nonlinear crystal). By
moving to a rotating frame with respect to ωr = ω p /2 (i.e., applying the transforma-
tion a → ae−ıω p t/2 ; see [7, Sect. 10.2.1] for a derivation of the equations of motion
of the DPA in the rotating frame), H can be rewritten as
ı
H = a ∗ a + ((a ∗ )2 − ∗ a 2 )
2
and expressed compactly as H = (1/2)x0 Rx0 + c (recall x0 = (q, p) ), where

⎡ ⎤
ı ∗ 1 ∗
⎢ + ( − ) 2 ( + ) ⎥
R = 1/2 ⎣ 1 2 ı ⎦, (3.8)
( + ∗ ) − ( − ∗ )
2 2
and c is a real number. Since c merely contributes a phase factor that has no effect on
the overall dynamics of the system operators, it can simply be ignored. That is, in the
rotating frame, the DPA can be treated as a harmonic oscillator with fixed quadratic
Hamiltonian (1/2)x0 Rx0 .
Two-Mode Squeezing
Beams propagating inside two cavities can be made to interact by having their paths
intersect inside a χ(2) nonlinear optical crystal and pumping the crystal with a pump
beam as a source of quanta. For instance, in a χ(2) optical crystal in which two
cavities interact with an undepleted classical pump beam, as depicted in Fig. 3.7, the
interaction can be modeled by the Hamiltonian
ı
H (t) = (e−ıω p t a1∗ a2∗ − ∗ eıω p t a1 a2 ),
2
where is a complex number representing the effective amplitude of the pump
beam and ω p is the pump beam frequency. Moving to a rotating frame at half
the pump frequency by applying the transformation a1 → a1 eı(ω p /2)t and a2 →
a2 eı(ω p /2)t , H (t) can be expressed in this new frame in the time-invariant form H =
2
ı
(a1∗ a2∗ − ∗ a1 a2 ). This kind of Hamiltonian is called a two-mode squeezing Hamil-
tonian, as it affects simultaneous squeezing in the quadratures of two independent
Fig. 3.7 Two cavity modes

interacting in a single
classically pumped nonlinear
crystal. The diagonal arrow
into the crystal denotes the
pump beam. Figure adapted a1
from [3] © 2009 Society for
Industrial and Applied Nonlinear
Mathematics. Reprinted with crystal
permission. All rights
reserved
a2
Fig. 3.8 Schematic

representation of a two-mode
squeezer. Figure adapted
from [3] © 2009 Society for a1
Industrial and Applied
Mathematics. Reprinted with
permission. All rights a2
reserved
modes that are linear combinations of a1 and a2 . It will play an important role later
on in this chapter. For further details on the physical theory of two-mode squeezing,
see [8, 9] (Fig. 3.8).
Remark 3.2 It is implicitly assumed that the equations of motion for a collection
of linear quantum systems are given with respect to a common rotating frame of
frequency ωr , including the transformation of all bosonic fields Ai (t) according to
Ai (t) → Ai (t)eıωr t , and that all classical pumps employed are of frequency ω p =
2ωr . This is a natural setting in quantum optics where a rotating frame is essential
for obtaining linear time-invariant QSDE models for active devices that require an
external source of quanta. In a control setting, this means both the quantum plant
and the controller equations have been expressed in the same rotating frame.
Static Linear Optical Devices and Networks
Static linear optical devices implement a constant frequency-independent linear

transformation of fields. Such transformations are represented by a complex square
matrix and can be applied to a set of independent incoming single-mode fields a =
(a1 , a2 , . . . am ) to produce an equal number a = (a1 , a2 , . . . am ) of independent
outgoing fields. The incoming and outgoing fields satisfy the CCR for single-mode
oscillators. However, the incoming fields may also be vacuum bosonic fields A(t) =
(A1 (t), A2 (t), . . . , Am (t)) with outgoing bosonic fields (that need no longer be
in the vacuum state) Y(t) = (Y1 (t), Y2 (t), . . . , Ym (t)) . For bosonic fields, the
commutation relations are [dA j (t), dAk (t)] = 0 and [dA j (t), dAk (t)∗ ] = δ jk dt.
To avoid cumbersome and unnecessary repetitions, in the following we restrict
the discussion to the operation of a static linear optical device on single-mode
fields. The operation is completely analogous for bosonic incoming and outgoing
fields, requiring only the obvious substitutions a → A(t), a → Y(t), [a j , ak ] =
0 → [dA j (t), dAk (t)] = 0, and [a j , ak∗ ] = δ jk → [dA j (t), dAk (t)∗ ] = δ jk dt, etc.
The action of a static linear optical device can mathematically be expressed as

a a Q1 Q2
=Q ; Q = ,
a# a# Q #2 Q #1
where Q 1 , Q 2 ∈ Cm×m , and Q is a Bogoliubov matrix as elaborated in Chap. 2. As

a consequence, the output fields a satisfy the same commutation relations as a.
In the special case where Q 2 = 0, the device does not mix creation and annihilation
operators of the fields, and it follows that Q 1 is necessarily a complex unitary matrix.
Such devices are said to be static passive linear optical devices by the fact that they
do not require an external source of quanta for their operation. It is well known that
any static passive network can be constructed using only beam splitters and mirrors
(e.g., see references 2–4 in [10]). In all other cases, the devices are static active.
Specific passive and static devices that will be of interest will be discussed in the
sections to follow.
Phase Shifter
A phase shifter is a device that imparts a phase shift on the incoming field. If the input
field is a, then a phase shifter outputs the field a = eıθ a for some real number θ,
called the phase shift. This device is schematically represented by the symbol shown
in Fig. 3.9. A phase shifter with a single input field is a static passive device with
transformation matrix Q P S given by

eıθ 0
QPS = .
0 e−ıθ
a a
eıθ
Fig. 3.9 Phase shifter with a phase shift of θ radians. Figure adapted from [3] © 2009 Society for
Industrial and Applied Mathematics. Reprinted with permission. All rights reserved
Beam Splitter
A beam splitter is a static passive device that takes two input fields a1 and a2 and
linearly combines them to produce two output fields a1 and a2 such that the num-
ber of photons is conserved: a1∗ a1 + a2∗ a2 = (a1 )∗ a1 + (a2 )∗ a2 . The transformation
performed by a beam splitter is given by

B 0
Q BS = ,
0 B#
where B is a unitary matrix given by

eı/2 0 cos()/2 sin()/2 eı/2 0
B = eı/2 .
0 e−ı/2 − sin()/2 cos()/2 0 e−ı/2
Here, , , , are real numbers. is called the mixing angle of the beam splitter
and the most important parameter. and introduce a phase difference in the two
incoming and outgoing modes, respectively, while introduces an overall phase
shift on both modes.
The operation of a beam splitter with = = = 0 can be modeled by an
effective Hamiltonian H B0 S = ı(a1∗ a2 − a1 a2∗ ) (see [4, Sect. 4.1] for details), in the
sense that
a a
Q B S # = exp(ı H B0 S ) # exp(−ı H B0 S ),
a a
θ+
where a = (a1 , a2 ) . By considering phase-shifted inputs a1 → a1 eı 2 and a2 →
θ−
a2 eı 2 (θ being an arbitrary real number), it follows that a beam splitter with
= 0 and = − will have the effective Hamiltonian H B S = ı(e−ı a1∗ a2 −
eı a1 a2∗ ) = αa1∗ a2 + α∗ a1 a2 , with α = ıe−ı . This is the most general type of
beam splitter that will be employed in the ensuing realization theory. A beam splitter
with a Hamiltonian of the form H B S is represented schematically by the symbol in
Fig. 3.10.
Fig. 3.10 Schematic a1

representation of a beam
splitter. Figure adapted
from [3] © 2009 Society for
Industrial and Applied a2
permission. All rights
reserved
Squeezer
Squeezing of a field is an operation in which the variance of one quadrature of a

mode a, either q or p, is squeezed or attenuated at the expense of increasing the
variance of the other quadrature. This type of operation is performed with a device
called a squeezer. An ideal squeezer realizes the transformation

cosh(s) eıθ sinh(s)
Q squeezer = −ıθ ,
e sinh(s) cosh(s)
where s and θ are real parameters. The parameter s is referred to as the squeezing
parameter, while θ is called the phase angle. For s < 0, the squeezer squeezes the
amplitude quadrature of e−ıθ/2 a (a phase-shifted version of a), while if s > 0, it
squeezes the phase quadrature and then shifts the phase of the squeezed field by θ/2.
It is easy to see that Q −1
squeezer is given by

cosh(s) −eıθ sinh(s)
Q −1 = .
squeezer −e−ıθ sinh(s) cosh(s)
Single-mode squeezing can be performed on the cavity mode inside a DPA, while a
squeezer for a bosonic field is an idealized device of infinite bandwidth that can be
realized approximately by a high bandwidth DPA, see the discussion in Example 2.3.
A squeezer with parameters s, θ is schematically represented by the symbol shown
in Fig. 3.11.
Static Optical Linear Networks
It is well known that an arbitrary static linear optical network has a decomposition
as a cascade of simpler static networks. In particular, any Bogoliubov matrix Q can
be constructively decomposed as [10]:

A1 0 0 D A3 0
Q = exp exp exp
0 A#1 D 0 0 A#3

exp A1 0 cosh D sinh D exp A3 0
= ,
0 exp A#1 sinh D cosh D 0 exp A#3
a a
s, 0
Fig. 3.11 Schematic representation of a squeezer. Figure adapted from [3] © 2009 Society for
Industrial and Applied Mathematics. Reprinted with permission. All rights reserved
a a
e−ıθ2 s, 0 eıθ2
Fig. 3.12 Implementation of a squeezer with arbitrary phase angle employing a squeezer with a
zero phase angle and two phase shifters. Figure adapted from [3] © 2009 Society for Industrial and
Applied Mathematics. Reprinted with permission. All rights reserved
where A1 and A3 are skew symmetric complex matrices and D is a real diagonal
matrix. The first and third factors in the decomposition correspond to passive sta-
tic networks implementable by beam splitters and mirrors, while the second factor
corresponds to an independent collection of squeezers (with trivial phase angles)
each acting on a distinct field. Thus, fields going through a static linear optical net-
work undergo a sequence of three operations: They are initially mixed by a passive
network, then they undergo squeezing, and finally, they undergo a final passive trans-
formation. In the special case where the entire network is passive, the D matrix is
zero and the second factor becomes an identity matrix. As a simple illustration, a
squeezer with arbitrary phase angle θ can be built by sandwiching a squeezer with
phase angle 0 between a −θ/2 phase shifter at its input and a θ/2 phase shifter at its
output, respectively. This is shown in Fig. 3.12.
3.1.3.2 Synthesis of One Degree-of-Freedom Linear Quantum Systems
We will now discuss the systematic realization of one degree-of-freedom linear quan-
tum systems with a trivial scattering matrix, S = I . Such systems are completely
described by a real symmetric Hamiltonian matrix R = R ∈ R2×2 and complex
coupling matrix K ∈ Cm×2 . Therefore, to realize these systems, one needs to be able
to realize both R and K . Below, we will describe the quantum optical realization
of one degree-of-freedom linear quantum systems that is based around a ring cavity
structure, such as shown in Fig. 3.4, utilizing fully reflecting and partially transmit-
ting mirrors, and placing beam splitters and nonlinear optical crystals at appropriate
locations between the mirrors.
The Hamiltonian of the system is H = (1/2)x Rx. In a one degree-of-freedom
setup, such a quadratic Hamiltonian can be realized with a DPA as discussed in
section “Degenerate Parametric Amplifier”. It is easy to see from (3.8) that by suitably
choosing the complex effective pump amplitude parameter and the cavity detuning
parameter of the DPA, one can realize any real symmetric matrix R. In fact, for
any particular R, the choice of parameters and for its realization is unique. For
instance, to realize
1 −2
R= ,
−2 0.5
one solves the set of simultaneous linear equations
− {} = 2, {} = −4, + {} = 1
for , {}, {} to yield the unique solution = 3/2 and = −4 − ı/2.
We now turn our attention to the realization of the coupling operator L = K x0 .

We begin by writing K = [ K 1 . . . K m ] , where K l ∈ C1×2 for each l = 1, . . . , m.
Each K l represents the coupling of the oscillator to the bosonic field Al (t), and so
it suffices to study how to realize the coupling to just one field. To this end, let us
now consider only a single bosonic field A(t) that is coupled to the oscillator via
the coupling operator L = K x0 for some K ∈ C1×2 . It will be more convenient to
express L = αq + β p as L = α̃a + β̃a ∗ , with α̃ = α−ıβ 2
and β̃ = α+ıβ
2
, and a and
∗
a being the oscillator’s annihilation and creation operators, respectively.
Partly taking inspiration from a scheme of Wiseman and Milburn for quantum
non-demolition measurement of the position operator of a cavity [11], we consider
the scheme shown in Fig. 3.13. The scheme relies upon using additional mirrors to
realize an auxiliary cavity mode b with a resonance frequency coinciding with the
reference frequency ωr = ω p /2. In a rotating frame at frequency ωr , the auxiliary
mode b interacts with the mode a via a two-mode squeezer with pump beam frequency
ω p and a beam splitter that collectively realize an interaction Hamiltonian Hab of the
form
ı ı
Hab = (1 a ∗ b∗ − ∗1 ab) + (2 a ∗ b − ∗2 ab∗ ), (3.9)
2 2
where 1 is the effective pump amplitude of the two-mode squeezer and 2 is given
by 2 = 2e−ı , where is the mixing angle of the beam splitter and is the
relative phase introduced between the input fields by the beam splitter. If we take
A(t)
A(t)
eıπ
e ıπ
Y(t)
b Auxiliary
cavity
Hab Y(t)
a a b
Fig. 3.13 Scheme for (approximate) implementation of a coupling L = α̃a + β̃a ∗ to cavity mode a
using an auxiliary cavity b (whose dynamics is adiabatically eliminated), a two-mode squeezer, and
a beam splitter with the appropriate parameters. The left figure is a block diagram showing the fast
mode b interacting with the slow mode a via the direct interaction Hamiltonian Hab , implemented
by the two-mode squeezer and the beam splitter, and also interacting with a 180◦ phase-shifted input
field A(t) to produce the output field Y (t). The right figure details the physical implementation of
the block diagram. Figure adapted from [3] © 2009 Society for Industrial and Applied Mathematics.
Reprinted with permission. All rights reserved
the coupling coefficient γ2 of the partially transmitting mirror M on b to be large,

such that b loses photons at a much faster rate than a and settles to the vacuum
state very quickly, one can take the limit γ2 → ∞ to adiabatically eliminate b and
obtain a reduced dynamics that only involves a. As shown in Appendix “Appendix
A: Adiabatic Elimination of Coupled Cavity Modes” using the singular perturba-
tion theory for QSDEs developed in [12], after eliminating b the resulting coupling
operator to a is given by
1
L = √ (−∗2 a + 1 a ∗ ).
γ2
Thus, if we now choose 1 , 2 , γ2 with γ2 large and such that
∗ 1
α̃ = − √ 2 , β̃ = √ , (3.10)
γ2 γ2
we can approximately realize any coefficients α̃ and β̃ in a linear coupling operator

L = α̃a + β̃a ∗ . Note that the π radian phase shifter in front of A(t) in Fig. 3.13 is
required to compensate for the scattering term in the unitary model that arises after
the adiabatic elimination procedure (detailed in Appendix “Appendix A: Adiabatic
Elimination of Coupled Cavity Modes”).
For the special case where α̃, β̃ satisfy |α̃| > |β̃| ≥ 0, there is an alternative real-
ization of the linear coupling that is based on pre- and postprocessing with squeezed
bosonic fields. Let γ = |α̃|2 − |β̃|2 > 0, and consider the interaction Hamiltonian
Hint (t) = ı(Lη(t)∗ − L ∗ η(t))

= ı((α̃a + β̃a ∗ )η(t)∗ − (α̃a ∗ + β̃ ∗ a)η(t)).
We can rewrite this Hamiltonian in the following form:

Hint (t) = ı a(α̃η(t)∗ − β̃ ∗ η(t)) − a ∗ (α̃η(t) − β̃η(t)∗ )
√
= ı γ(aη (t)∗ − a ∗ η (t)),
t
where η (t) = √1 (α̃η(t)
γ
− β̃η(t)∗ ). Letting Z(t) = 0 η (s)ds, we have that Z(t) =
√1 (α̃A(t) ∗
γ
− β̃A(t) ), and
⎡ ⎤

Z(t) A(t) √α̃ − √β̃γ
=Q , Q= ⎣ γ∗ ⎦.
Z(t)∗ A(t)∗ − √β̃ γ √α̃
γ
√
The key idea is reinterpreting the Hamiltonian in the form Hint (t) = ı γ(aη (t)∗ −
∗
a η (t)), which describes the interaction of the oscillator with the new field Z(t).
√ √
Since |α̃|2 − |β̃|2 = γ > 0, we have that |α/ γ|2 − |β/ γ|2 = 1; therefore, Q
is Bogoliubov. In fact, Z(t) can be obtained from A(t) by passing the latter
Z(t) A(t)
s, 0
M
Zout (t) Y(t)
−s, θ
Fig. 3.14 Scheme for implementation of a coupling L = α̃a + β̃a ∗ with α̃ > 0 and α̃
> |β̃|. Here,
√
s = −arccosh(α̃/ γ), θ = arg(β̃) and the mirror M’ has coupling coefficient γ = α̃2 − |β̃|2 .
through a squeezer with the appropriate parameters (section “Squeezer”). Thus,

Z(t) is a squeezed version of A(t). The required squeezer parameters are s =
√
−arccosh(α̃/ γ) and θ = arg β̃. The field Z(t) satisfies [dZ(t), dZ(t)∗ ] = dt and
the Itō rules,
dZ(t) 01
∗
dZ(t) dZ(t) = Q Q dt.
dZ(t)∗ 00
The interaction Hamiltonian Hint can be implemented in one arm of a ring cavity
with a fully reflecting mirror M and a partially transmitting mirror M’ with coupling
coefficient γ, with Z(t) incident on M’. After the interaction, an output field Zout (t)
is produced by M’ given by
Zout (t) = U (t)∗ Z(t)U (t)

α̃ β̃
= √ U (t)∗ A(t)U (t) − √ U (t)∗ A(t)∗ U (t).
γ γ
However, the actual output of interest is the field Y(t) = U (t)∗ A(t)U (t) that is
produced when the oscillator interacts directly with A(t). Since Q and its inverse
Q −1 are both Bogoliubov, Y(t) can be recovered from Zout (t) by exploiting the
relation,
Y(t) −1 Zout (t)
= Q ,
Y(t)∗ Zout (t)∗
which follows directly from the fact that (Z(t), Z(t)∗ ) = Q(A(t), A(t)∗ ) . The
complete implementation of this linear coupling is shown in Fig. 3.14.
3.1.3.3 Engineering the Interactions Between One-Dimensional Open

Quantum Harmonic Oscillators
Following Theorem 3.1, the second ingredient for synthesizing a general linear quan-
tum system is being able to realize a direct interaction Hamiltonian H d given by
(3.7) between several quantum harmonic oscillators. The only case where this is
not required, and field-mediated interactions suffice, is the instance when R jk and
L j and S j , j, k = 1, . . . , n, are such that H d = 0. H d consists of the sum of direct
interaction Hamiltonians between pairs of one-dimensional harmonic oscillators of
the form Hkl = xk Ckl xl (k = l) with Ckl a real matrix. Under the assumption that the
time it takes for the light in a ring cavity to make a round trip is much faster than the
timescales of all processes taking place in the ring cavity, it will be sufficient to only
consider how to realize Hkl for any pair of one-dimensional harmonic oscillators and
realizing all of them simultaneously in a network. Now, let a j = ( p j + ıq j )/2 and
a ∗j = ( p j − ıq j )/2 for j = k, l, and express Hkl as
Hkl = 1 ak∗ al + ∗1 ak al∗ + 2 ak∗ al∗ + ∗2 ak al
for some complex numbers 1 and 2 . The first component Hkl1 = 1 ak∗ al + ∗1 ak al∗ can
be realized using a beam splitter with mixing angle = |1 |, = − arg(1 ) + π2 ,
= −, and = 0 (see section “Beam Splitter”). The second component Hkl2 =
2 ak∗ al∗ + ∗2 ak al can be realized by interacting ak and al in a two-mode squeezing
process as described in section “Two-Mode Squeezing”, inside a χ(2) nonlinear crys-
tal driven by a classical pump beam of frequency 2ωr and effective pump amplitude
−2ı2 . The overall Hamiltonian Hkl is then realized by positioning the arms of the
two ring cavities (with canonical operators xk and xl ) such that their circulating light
beams intersect at two locations. At one location, a beam splitter is placed to realize
Hkl1 and a pumped nonlinear crystal is placed at the second to realize Hkl2 , respectively.
An example of this scheme is depicted in Fig. 3.15.
Fig. 3.15 Example

ak
implementation of the total
direct interaction
Hkl = Hkl1 + Hkl2 between
the modes ak and al of two
ring cavities. Figure adapted
from [3] © 2009 Society for ak
al
Industrial and Applied
permission. All rights al
reserved
3.1.4 Illustrative Synthesis Example
We now illustrate the network synthesis theory by applying the complete synthe-
sis procedure to two degree-of-freedom linear quantum system. The system under
consideration is coupled to a single external bosonic noise field A(t) and given by

G= (I4×4 , K x, x diag(R 1 , R2 )x), with x = (q1 , p1 , q2 , p2 ) , K = [
3/2 ı/2 1 ı ],
2 0.5 10
R1 , and R2 = .
0.5 3 01
Let G 1 = (I2×2 , K 1 x1 , (1/2)x1 R1 x1 ) and G 2 = (I2×2 , K 2 x2 , (1/2)x2 R2 x2 ) be
two independent one degree-of-freedom linear quantum systems with x1 = (q1 , p1 ) ,
x2 = (q2 , p2 ) , K 1 = [ 3/2 ı/2 ], and K 2 = [ 1 ı ]. From Theorem 3.1, we have that
G can be realized as the reducible network G = {{G 1 , G 2 }, H12 d
, G 2 G 1 }, with the
d
direct interaction Hamiltonian H12 between G 1 and G 2 given by (see (3.7))
1 ∗
d
H12 =− x (K K 1 − K 2 K 1# )x1
2ı 2 2
1 0 −1
= x2 x1 .
2 3 0
This network is depicted in Fig. 3.3.

d
We shall now demonstrate how G 1 , G 2 , and H12 can be individually realized to
synthesize the overall system G.
3.1.4.1 Synthesis of G 1 and G 2

2 0.5
From Sect. 3.1.3.2, we have that R1 = can be realized as a DPA with
0.5 3
parameters = 5 and = 1 + ı, while L 1 = K 1 x1 can be (approximately) realized
by the first scheme proposed in Sect. 3.1.3.2 and shown in Fig. 3.13. To achieve the
desired coupling parameter, we can set the coupling coefficient of the mirror M to
be γ2 = 100, = 10, and the beam splitter to have a mixing angle of −10, with all
other parameters equal to 0 (recall that we require γ2 to be sufficiently larger than the
other parameters so that the auxiliary mode can be adiabatically eliminated). Overall,
G 1 can be implemented around the ring cavity structure depicted in Fig. 3.16. G 2
can be realized similarly. The Hamiltonian H2 = (1/2)x2 R2 x2 can be realized in
the same manner as H1 with the choice = 2 and = 0. Since = 0, to realize
R2 it suffices to have a cavity that is detuned from ωr , the reference frequency in
Remark 3.2, by an amount = 2. The coupling operator L 2 = q2 + ı p2 = 2a2 ,
where a2 is the annihilation operator/cavity mode of G 2 , can be realized simply with
a partially transmitting mirror with coupling coefficient κ = 4, through which an
impinging field A2 (t) interacts with the cavity mode a2 to produce an outgoing field
Y2 (t). The implementation of G 2 is shown in Fig. 3.17.
Fig. 3.16 Realization of G 1 .

Figure adapted from [3] ©
2009 Society for Industrial
and Applied Mathematics.
Reprinted with permission.
All rights reserved
Fig. 3.17 Realization of G 2 .

Figure adapted from [3] ©
2009 Society for Industrial
and Applied Mathematics.
Reprinted with permission.
All rights reserved
a2
Y2 (t)
A2 (t)
d
3.1.4.2 Synthesis of H12
d
Finally, we come to consider
the realization of the direct interaction Hamiltonian H12
0 −1
given by H12 d
= 21 x2 x1 . We can rewrite H12d
in terms of the cavity modes a1
3 0
and a2 as H12d
= −ı(a1∗ a2 − a1 a2∗ ) + 2ı(a1∗ a2∗ − a1 a2 ). Let H12,1
d
= −ı(a1∗ a2 − a1 a2∗ )
∗ ∗
and H12,2 = 2ı(a1 a2 − a1 a2 ) so that H12 = H12,1 + H12,2 . The component H12,1
d d d d d
=
∗ ∗
−ı(a1 a2 − a1 a2 ) can be realized using a beam splitter with mixing angle = −1
and all other parameters equal to 0 (section “Beam Splitter”). The second component
d
H12,2 = 2ı(a1∗ a2∗ − a1 a2 ) corresponds to two-mode squeezing between ak and al in
(2)
a χ nonlinear crystal using a classical pump beam of frequency ω p = 2ωr and
effective amplitude = 4.
3.1.4.3 Complete Realization of G = {{G 1 , G 2 }, H12

d
, G2 G1}
The system G can now be realized by (i) positioning the arms of the two (ring)
cavities of G 1 and G 2 such that their internal light beams intersect at two points
G H12
d
A(t) Y(t)
G1 = (S1 , L1 , H1 ) G2 = (S2 , L2 , H2 )
A(t)
G2
eıπ
G1 Auxiliary Y1 (t)
cavity
Y(t)
Y1 (t)
Fig. 3.18 Realization of G. The block diagram at the top shows how G is realized by a series
d between the canonical operators of
connection of G 1 into G 2 and a bilinear direct interaction H12
G 1 and G 2 . The bottom figure shows the physical implementation of G based on the block diagram.
d
which are occupied by a beam splitter and two-mode squeezer realizing H12,1 and
d
H12,2 , respectively, and (ii) passing the output Y1 (t) of G 1 as input to G 2 . The
realization is shown in Fig. 3.18.
3.2 Architecture for Strict Realization Using Quantum

Feedback Networks
We have shown that any linear quantum system can, in principle, be synthesized
by a cascade of simple one degree-of-freedom linear quantum subsystems together
with a direct interaction Hamiltonian between the canonical operators of the system.
However, we have also seen that direct bilinear interaction Hamiltonians between
subsystems are challenging to realize experimentally, possibly requiring some com-
plex spatial arrangement and orientation of the subsystems. Thus, it becomes impor-
tant to investigate alternative synthesis methods, at least approximate ones. Here, we
describe one such method by exploiting the quantum feedback network formalism
in [6], following the original proposal in [13]. In this scheme, the direct interaction
Hamiltonians are approximately realized by suitable field interconnections among
the systems. The approximation relies on the assumption that the time delays for
establishing field interconnections between systems are negligible compared to the
timescale of the dynamics of the individual systems.
3.2.1 The Model Matrix and Concatenation of Model

Matrices
We will introduce an alternative representation of the system G = (S, L , H ) called

a model matrix [6], which will be particularly useful for the goal at hand. For a
given system G, the model matrix representation M(G) is given by (the partitioned
matrix):
−ı H − 21 L ∗ L −L ∗ S
M(G) = . (3.11)
L S
If L is partitioned as L = (L
1 , L 2 , . . . , L n out ) , with L j ∈ m j × 1 and
n out
j=1 m j =
m, and S is correspondingly partitioned as S = [S jk ] j=1,...,n out ,k=1,...,n in with S jk ∈
Cm j ×m k and nk=1in
m k = m, then the model matrix above can be expressed with
respect to this partition as:
3.2 Architecture for Strict Realization Using Quantum Feedback Networks 97
M(G)
⎡ n out ∗ n out ∗ n out ∗ ⎤
−ı H − 1
2 j=1 L j L j − j=1 L j S j1 ... − j=1 L j S jn in
⎢ L1 S11 ... S1n in ⎥
⎢ ⎥
=⎢ .. .. .. .. ⎥ . (3.12)
⎣ . . . . ⎦
L n out Sn out 1 ... Sn out n in
Given a partitioning of M(G), such as the one above, we can assign a unique label
to each row and column of the partition. For instance, for the partitioning (3.12)
we may attach the labels s0 , s1 , . . . , sn out for the first, second, ..., n out + 1-th row of
M(G), respectively, and r0 , r1 , . . . , rn in for the first, second, ..., n in + 1-th column of
M(G), respectively. With respect to this labeling scheme, block elements of M(G)
can be identified as,
n out n out
1
Ms0 r0 (G) = −ı H − L ∗j L j ; Ms0 rk (G) = − L ∗j S jk , k > 0;
2 j=1 j=1
Ms j r0 (G) = L j , j > 0; Ms j rk = S jk , j, k > 0.
Additionally, the original physical system G can be readily identified from the entries
of its model matrix M(G). For the sake of brevity, we will often omit G and write
M(G) simply as M and denote its entries as Mαβ , with α ranging over row labels
and β ranging over the column labels. Also, the parameters (S, L , H ) of G can be
equivalently viewed as the parameters of the model matrix M.
When there are several model matrices, they can be concatenated together to form
the model matrix of a larger system. This concatenation is the analogue, in terms
of model matrices, of the concatenation product for open Markov quantum systems
and is again denoted by the symbol . If G 1 = (S1 , L 1 , H1 ) and G 2 = (S2 , L 2 , H2 ),
then the concatenation M(G 1 ) M(G 2 ) is defined as:
⎡ ⎤
−ı H1 − ı H2 − (1/2)L ∗1 L 1 − (1/2)L ∗2 L 2 −L ∗1 S1 −L ∗2 S2
M(G 1 ) M(G 2 ) = ⎣ L1 S1 0 ⎦
L2 0 S2
= M(G 1 G 2 ).
3.2.2 Edges, Elimination of Edges, and Reduced Markov

Models
Following [6], one can associate a row partition labeled sk with k > 0 with an output
port sk (having multiplicity n sk ), while a column partition r j with j > 0 can be asso-
ciated with an input port r j (having multiplicity n r j ). If a system has a decomposition
as the concatenation of several subsystems, an output port sk from one subsystem
can be connected to an input port r j of another subsystem (possibly the same sub-
system to which the output port belongs). This connection forms what we shall call
an internal edge and denote by (sk , r j ). Of course, for the connection to make sense,
the ports sk and r j are required to have the same multiplicity. Such an edge then
represents a channel from port sk to port r j . All ports that are part of an internal
edge or channel are referred to collectively as internal ports, and fields entering or
leaving such ports are called internal fields. All other input and output ports that do
not belong to an internal edge are viewed as having semi-infinite edges (that do not
terminate at either an input port or an output port) and are referred to as external
ports. The associated semi-infinite edges are referred to as external edges. Fields
entering or leaving external ports are called external fields.
Consider an internal edge (sk , r j ). In such an edge, there is an associated finite
time delay required for the signal from port sk to travel to port r j . Due to these time
delays, concatenated systems possessing internal edges are longer Markov, despite
each element in the concatenation being Markov. Fortunately, as shown in [6], in
the limit that all time delays on internal edges go to zero, the non-Markov system
converges to a reduced Markov model. Thus, when the time delays are negligible
compared to the timescale of the dynamics of the systems connected by the internal
edges, the reduced Markov model is an effective approximation of the non-Markov
model for the network. We recall the following result:
Theorem 3.2 ([6, Theorem 3.1 and Lemma 16]) Let τ(sk ,r j ) , j, k > 0, be the time
delay for an internal edge (sk , r j ) and assume that I − Sk j is invertible. Then in the
limit that τ(sk ,r j ) ↓ 0, M(G) with the edge (sk , r j ) connected reduces to a simplified
model matrix Mred with input ports labeled r0 , r1 , . . ., r j−1 , r j+1 , . . ., rn out and output
ports labeled s0 , s1 , . . ., sk−1 , sk+1 , . . ., sn in (i.e., the connected ports r j and sk are
removed from the labeling and the associated row and column removed from M(G)).
The block entries of Mred are given by:
(Mred )αβ = Mαβ + Mαr j (1 − Sk j )−1 Msk β ,
with α ∈ {s0 , s1 , . . . , sn out }\{sk } and β ∈ {r0 , r1 , . . . , rn in }\{r j }. Mred is the model
matrix of a linear quantum stochastic system G red with parameters:
(Sred ) pq = S pq + S pj (I − Sk j )−1 Skq ,

(L red ) p = L p + S pj (I − Sk j )−1 L k ,
n out
Hred = H + {L ∗p S pj (I − Sk j )−1 L p },
p=1
for all p ∈ {1, 2, . . . , n out }\{k} and q ∈ {1, 2, . . . , n in }\{ j}.
Internal edges may be sequentially eliminated, by allowing the time delays along
these edges to go to zero one at a time, leading to a corresponding sequence of
reduced model matrices. The order in which the edge eliminations are performed is
irrelevant so that the final reduced model matrix is in fact unique [6, Lemma 17].
Suppose that M can be partitioned as:
⎡ ⎤
−ı H − (1/2)L ∗i L i − (1/2)L ∗e L e −L ∗i Sii − L ∗i Sei −L ∗e Sie − L ∗i Sii
⎣ Li Sii Sie ⎦,
Le Sei See
(3.13)
where the subscript i refers to “internal” and e to “external”, meaning that parameters
with subscript i or ii pertain to internal ports, those with subscript e or ee pertain to
external ports, while Sie and Sei pertain to scattering of internal fields to external
fields and vice versa, respectively. A so-called adjacency matrix can be introduced
to conveniently encode the interconnection among internal input and output ports.
Let n i denote the total multiplicity of internal input and output ports. We may also
view a port with multiplicity k as k distinct ports of multiplicity 1, as necessary.
Suppose that we number these multiplicity 1 ports consecutively starting from 1.
Then, an adjacency matrix η is an n i × n i square matrix whose entries are either 1 or
0 with η( j, k) = 1 ( j, k ∈ {1, 2, . . . , n i }) only if the jth output port and the kth input
port form a channel or internal edge. Clearly, there can be at most one element of any
row or column of η that can take on the value 1. Internal edges can be simultaneously
eliminated as follows:
Theorem 3.3 ([6, Sect. 5]) Suppose that M has a partitioning based on internal and
external components as in (3.13) and that connections between internal ports have
been encoded in an adjacency matrix η. If (η − Sii )−1 exists, then the reduced model
matrix Mred after simultaneous elimination of all internal edges has the parameters:
Sred = See + Sei (η − Sii )−1 Sie ,

L red = L e + Sei (η − Sii )−1 L i ,
Hred = H + {L ∗j S ji (η − Sii )−1 L i }.
j=i,e
3.2.3 Main Synthesis Results
For j = 1, . . . , n, let G jk = (S jk , L jk , 0) for k = 1, . . . , n, k = j, and G j j = (S j j ,

L j , H j ), with S jk ∈ Cc jk ×c jk , c jk = ck j and c j j = m, L jk = K jk x j , and H j = (1/2)
x j R j x j with R j = R j ∈ R2×2 . Here x j = (q j , p j ) is as defined before. Let G j =
nk=1 G jk for j = 1, . . . , n, and note that G j = (S j , L j , H j ) with S j = diag(S j1 ,
S j2 , . . . , S jn ), L j = (L j1 , L j2 , . . . , L jn ) , and H j as already defined.
Consider the model matrix M for the concatenated system G = nj=1 G j (see
Fig. 3.19). Partition M correspondingly according to the given concatenation decom-
position of G. We label the first n + 1 rows of M as s00 , s11 , . . ., s1n , the next n rows
G1 G2 G3
Ar Ys Ar Ys Ar Ys
11 11 21 21 n1 n1
r11 G11 s11 r21 G21 s21 rn1 Gn1 sn1
Ar Ys12 Ar Ys22 Ar Ysn2

12 22 n2
r12 G12 s12 r22 G22 s22 rn2 Gn2 sn2
Ar Ys Ar Ys Ar Ys
1n 1n 2n 2n nn nn
r1n G1n s1n r2n G2n s2n rnn Gnn snn
Fig. 3.19 The concatenation decomposition of G as G = nj=1 nk=1 G jk . Figure adapted

from [13] © 2010 IEEE
as s21 , . . ., s2n , and so on until the last n rows are labeled sn1 , sn2 , . . ., snn . In an
analogous fashion, we label the first n + 1 columns of M as r00 , r11 , . . ., r1n , the
next n columns as r21 , . . ., r2n , and so on until the last n columns rn1 , rn2 , . . ., rnn .
For notational clarity, we will sometimes write a bracket around one of both of the
subscripts of r or s (e.g., as in s(n−1)k or r(k+1)(k+1) ).
Theorem 3.4 Let the output port s jk be connected to the input port rk j to form an
internal edge/channel
e jk = (s jk , rk j ) for all j, k = 1, . . . , n, j = k. Assuming that
−S jk I
is invertible ∀ j, k = 1, . . . , n, j = k, then the reduced model matrix
I −Sk j
Mred obtained by allowing the delays in all internal edges {e jk , j = k} go to zero
has parameters given by:
Sred = diag(S11 , S22 , . . . , Snn ),

L red = (L
11 , L 22 , . . . , L nn ) ,
n n−1 n −1
I −S jk L jk
Hred = Hk + [ L ∗jk L ∗k j ] .
−Sk j I Lkj
k=1 j=1 k= j+1
The proof of the theorem is given in the Appendix, “Appendix B: Proof of

Theorem 3.4”. We also exploit the following lemma:
Lemma 3.1 For any real 2 × 2 matrix R and unitary complex numbers S12 and
S21 satisfying S12 S21 = 1, there exist 1 × 2 complex matrices K 1 and K 2 such
that R − { 1−SS1212 S21 K 1∗ K 2 + 1−SS1221 S21 K 1 K 2# } = 0. In fact, a pair K 1 and K 2 sat-
isfying this is given by K 1 = [ κ ıκ ] with κ an arbitrary nonzero real number
S −S ∗
and K 2 = 2ı[ 1 0 ][ −K 1∗ ∗ K 1 ]−1 R, where = 2 |1−S
21 12
12 S21 |
2 . Or, alternatively,
∗ ∗ −1
K 2 = [ κ iκ ] and K 1 = 2ı[ 1 0 ][ K 2 −K 2 ] R .
See “Appendix C: Proof of Lemma 3.1” in the appendix for a proof of the lemma.
As a corollary to the above theorem and lemma, we have the following result:
Corollary 3.1 Let c jk = ck j = 1 whenever j = k, and c j j = m for all j, k = 1, . . . ,
n. Also, let S j j = Im , Sk j = eıθk j and S jk = eıθ jk with
θk j , θ jk ∈ [0, 2π) satisfying
n 1 ∗

θk j + θ jk = 0, K j j = K j and R j = R j j − 2sym k=1,k= j 1−S jk Sk j K jk K jk
(K j and R j j given), where sym(A) = (1/2)(A + A ), and the pair (K jk , K k j )

( j = k) be given by:
K jk = [ κ jk ıκ jk ],
−1 #
K k j = 2ı[ 1 0 ][ −K ∗jk ∗jk K
jk jk ] (R jk − {K j K k }),
(3.14)
or
K k j = [ κ jk ıκ jk ],
K jk = 2ı[ 1 0 ][ K k∗j jk −K kj ∗jk ]−1 (R jk − {K #
j K k }) ,
(3.15)
Sk j −S ∗
where R jk = Rkj ∈ R2×2 , jk = 2 |1−Sk j S jkjk |2 and κ jk is an arbitrary nonzero real
constant for all j, k. Then the reduced Markov model G red = (Sred , L red , Hred )
has the decomposition G red = nk=0 G red,k with G red,0 = (0, 0, Hred ) and G red,k =
(Skk , L kk , 0) for k = 1, . . . , n. Moreover, the network G net = (Snet , L net , Hnet )
formed by forming the series product of G red,n . . . G red,2 G red,1 within the con-
catenated system G red and defined by G net = G red,0 (G red,n . . . G red,2 G red,1 )
is a linear quantum stochastic system with parameters given by:
Snet = Im ,
L red = K x, K = [ K 1 K 2 . . . K n ],
Hred = (1/2)x Rx, R = [R jk ] j,k=1,...,n .
In other words, G net realizes a linear quantum stochastic system with the above
parameters.
Remark 3.3 The series connection G red,n . . . G red,2 G red,1 is equivalent to form-
ing and subsequently eliminating the internal edges {(skk , r(k+1)(k+1) ); k = 1, . . . ,
n − 1}.
The proof of this corollary is given in Appendix “Appendix D: Proof of
Corollary 3.1”. It shows that any linear quantum system can be realized by the quan-
tum network G net constructed following the prescription of the corollary. Finally,
any system (S, K x, (1/2)x Rx) can be constructed as (S, K x, (1/2)x Rx) =
(I, K x, (1/2)x Rx) (S, 0, 0). The latter is the cascade of a static network realizing
the unitary scattering matrix S and the system (I, K x, (1/2)x Rx).
Example 3.1 Consider a two degree-of-freedom linear quantum system G sys =
(I, K x, (1/2)x Rx) (x = (q1 , p1 , q2 , p2 ) ) with
K = [ K 1 K 2 ] = [ 3/2 ı/2 1 ı ] (K j ∈ C1×2 , j = 1, 2);

⎡ ⎤
2 0.5 1 1
R11 R12 ⎢ 0.5 3 −1 −1 ⎥
R= =⎢ ⎥
⎣ 1 −1 1 0 ⎦ (R jk ∈ R , j, k = 1, 2).
2×2
R12 R22
1 −1 0 1
Let x1 = (q1 , p1 ) and x2 = (q2 , p2 ) . Define

K 11
G 1 = diag(Im , S12 ), x , (1/2)x1 R1 x1 ,
K 12 1
and
K 21
G 2 = diag(S21 , Im ), x , (1/2)x2 R2 x2 ,
K 22 2
with parameters that will be determined in the following. Set θ12 = 0 and θ21 = π/2,
so that θ12 + θ21 = 0 as prescribed in Corollary 3.1. Then, set S12 = eıθ12 = 1,
S21 = eıθ21 = ı, K 11 = K 1 = [ 3/2 1/2 ], and K 22 = K 2 = [ 1 ı ]. Calculate 12 =
∗
S21 −S12
2 |1−S 2 = −1 + ı and set κ12 = 1. By Corollary 3.1, we then set K 12 = [ κ12 ıκ12 ]
12 21 |
S
= [ 1 ı ] and compute K 21 = [ 1.25 − 0.25ı 1.75 + 0.75ı ],

1 0.5
R1 = ,
0.5 2
and
0.625 2
R2 = − .
2 2.625
Therefore, all the parameters of G 1 and G 2 have now been determined. Label the
ports of G 1 and G 2 according to the convention adopted in this section. Concatenate
G 1 and G 2 and form the internal edges (s12 , r21 ) and (s21 , r12 ). Then, eliminate
these edges to form G red . Finally, eliminating the edge (s11 , r22 ) from G red (see
Remark 3.3) yields G net as an approximate realization of G sys . The realization is
depicted in Fig. 3.20. G 1 and G 2 can be physically realized in the quantum optic
domain following the constructions in Sect. 3.1.3. A quantum optical circuit as a
physical realization of G sys is shown in Fig. 3.21.
Gred
Ys
11
Ar r11 s11 r21 s21
11
G1 G2 Ys
22
r12 s12 Ar r22 s22
22
Gsys
Ar r11 s11 r21 s21 Ys

11
G1 G2 22
r12 s12 r22 s22
Fig. 3.20 Realization of G sys via a quantum feedback network. In the top figure, the internal edges
(s12 , r21 ) and (s21 , r12 ) are formed and eliminated to obtain a reduced Markov model G red . Then
in the bottom figure, a series (cascade) connection is formed by the eliminating the internal edge
(s11 , r22 ) to realize G sys . Figure adapted from [13] © 2010 IEEE
A(t) Y(t)
Gsys r11
s22
G1 eıπ G2
s11 r22
Auxiliary
cavity a2 Auxiliary
cavity
a1 r21
eıπ+ıθ21
eıπ+ıθ12
r12 s21
s12
Fig. 3.21 A quantum optical circuit that realizes G sys according to the quantum feedback network
q +ı p
in Fig. 3.20. Here a j = j 2 j is cavity mode of the optical cavity around which the physical
realization of the oscillator G j is based, j = 1, 2. The dashed box labeled G j is the part of the
circuit realizing G j , and the input and output ports of G j are indicated by their respective labels.
Figure adapted from [24]
3.2.4 Synthesis of Completely Passive Systems
We now consider the synthesis of the class of completely passive systems, introduced
in Sect. 2.4.1, via quantum feedback networks. Recall that for completely passive
systems, the Hamiltonian takes the form (1/2)a ∗ R̃a for some complex Hermitian
matrix R̃ and the coupling operator takes the form L = K̃ a for some complex matrix
K̃ . Let us write (1/2)a ∗ R̃a and K̃ a in the following way:

(1/2) R̃ 0n×n a 1
n
(1/2)a ∗ R̃a = (1/2) a ∗ a # − R̃ j j
0n×n (1/2) R̃ # a 4 j=1
n
1
= (1/2)x { ∗ R̃}x − R̃ j j
4 j=1

a
K̃ a = K̃ I 0 = K̃ x.
a#
Therefore, H = (1/2)x Rx with R = { ∗ R̃}, and L = K x with K = K̃ .

From the above, it easily follows that for any completely passive system, the block
diagonal elements R j j of R = [Ri j ]i, j=1,...,n must be diagonal of the form λ j I2 for
some λj ∈ R, the off-diagonal
block elements R jk are real 2 × 2 matrices of the form
α jk β jk
R jk = for some real numbers α jk and β jk , and the coupling matrix K j
−β jk α jk
to x j is of the form K j = [ γ j ıγ j ] for some complex number γ j . It is reasonable to
expect that a completely passive system can be realized using only purely passive
components. The next theorem states that this is exactly the case.
Theorem 3.5 Let G sys = (S, L , H ) be completely passive. Then the systems {G j ;
j = 1, . . . , n} constructed according to Corollary 3.1 are all also completely passive.
Proof By the complete passivity of G sys , we immediately see that K j j = K j is

already of the form required for complete passivity of G j , and the matrix R jk −

# α jk β jk
{K j K k } is a 2 × 2 real matrix of the form for some α jk , β jk ∈ R,
−β jk α jk
whenever j = k. Let κ jk , S jk , and Sk j be chosen according to Corollary 3.1. Set K jk
according to the first equality of (3.14). Then, some straightforward algebra shows
that K k j given by the second equality of (3.14) is of the form K k j = [ γk j ıγk j ] for
some γk j ∈ C, just like K jk . The same holds true if one chooses the alternative of
setting K k j according to the first equality of (3.15) and computing K jk according
to the second equalityof (3.15). Finally, given this special form of K jk , it is eas-
ily inspected that sym { 1−S1jk Sk j K ∗jk K jk } is a diagonal matrix of the form λ jk I2
for some λ jk ∈ R for all j, k, j = k (recall that sym(A) = (1/2)(A + A )), and
since R j j is also diagonal of thisform (again from the passivity of G sys ), we have
that R j = R j j − 2 nk=1,k= j sym { 1−S1jk Sk j K ∗jk K jk } is again of the same form, for
all j. Therefore, it now follows that each subsystem G j of Sect. 3.2.3 with parameters
determined according to Corollary 3.1 is completely passive.
We conclude the discussion with an example synthesis of a completely passive sys-

tem.
Example 3.2 Consider a completely passive two degree-of-freedom linear quantum

system G sys = (I, K x, (1/2)x Rx) (x = (q1 , p1 , q2 , p2 ) ) with
K = [ K 1 K 2 ] = [ −3 + ı −3ı − 1 1 ı ] (K j ∈ C1×2 , j = 1, 2);

⎡ ⎤
2 0 1 4
R11 R12 ⎢ 0 2 −4 1 ⎥
R= =⎢ ⎥
⎣ 1 −4 1 0 ⎦ (R jk ∈ R , j, k = 1, 2).
2×2
R12 R22
4 1 0 1
Setting κ12 = 1, S12 = 1, S21 = ı, and K 12 = [ κ12 ıκ12 ] = [ 1 ı ], from

Corollary 3.1 we obtain K 11 = K 1 , K 22 = K 2 , K 21 = [ 0.5 − 0.5ı 0.5 + 0.5ı ], R1 =
02×2 , and R2 = 0.5I2 . A passive optical circuit that realizes G sys is illustrated in
Fig. 3.22.
A(t) Y(t)
Gsys r11
s22
G1 eıθ11
s11
G2
r22
e−ıθ11
a2
r21
a1 eıπ/2+ıθ12
e−ıθ12
r12
s21
s12
Fig. 3.22 A quantum optical circuit that realizes the passive system G sys of this example. The
circuit consists of only passive optical components: mirrors and phase shifters. Here, θ11 = π −
arctan(1/3) and θ21 = − π4 , and the values of all other parameters of devices in this circuit can be
determined according to Sect. 3.1. Figure adapted from [24]
3.3 Transfer Function Realization
In Sect. 3.1, we have considered a strict type of realization problem for linear quan-
tum systems. It is concerned with synthesizing a linear quantum system described by
a given triplet of parameters, {S, L = K x, H = (1/2)x Rx}. Strict realizability is
relevant to the situation where the dynamics of x(t) is tied to an implementation of a
quantum information processing algorithm and has to be realized as given. However,
in some linear quantum control problems, including robust disturbance attenuation
[14] and LQG synthesis [15] (to be discussed in Chap. 5), the internal dynamics are
unimportant. In such problems, what is important is the transfer function of the system
as determined by the system matrices (A, B, C, D). Thus, there is freedom to modify
the internal dynamics without modifying the transfer function since the latter is invari-
ant under the similarity transformation (A, B, C, D) → (V AV −1 , V B, C V −1 , D)
for any invertible matrix V . However, for linear quantum systems, the transformation
matrix V is restricted to be a real symplectic matrix. That is, V satisfies the condi-
tion V Jn V = Jn . This is to guarantee that the transformed variable Z (t) = V X (t)
satisfies the CCR for quantum oscillators: Z (t)Z (t) − (Z (t)Z (t) ) = 2ıJn , so
that the transformed system is again a physical linear quantum system. Recall that
the set of all symplectic matrices of a fixed dimension form a group. In particular,
V −1 is again symplectic. A similarity transformation with a symplectic matrix V
in linear quantum systems corresponds to replacing G = (S, K x, (1/2)x Rx) with
G = (S, K V −1 x, (1/2)x V −T RV −1 x). This motivates the following definition:
Definition 3.1 Let G = (S, K x, (1/2)x Rx) and G = (S , K x, (1/2)x R x) be

two linear quantum systems. Then, G is said to be transfer function equivalent to
G (or a transfer function realization of G) if S = S and there exists a symplectic
matrix V such that R = V − RV −1 , K = K V −1 (or, equivalently, R = V R V
and K = K V ). G is then said to be transfer function realizable by G , and vice
versa.
Remark 3.4 Two transfer function equivalent systems G and G need not generate
the same input-output dynamics (A(t), Y(t)) for all t ≥ 0. This is because they start
at a nonzero initial condition x(0) = (q1 , p1 , q2 , p2 , . . . , qn , pn ) but have different
parameters (S, K , R). Therefore, their transient dynamics may not be the same.
However, if the A matrix of G is Hurwitz, the input-output dynamics of G and G
converge asymptotically as t → ∞.
3.3.1 Pure Cascade Realization of the Transfer Function

of Linear Quantum Systems
Compared to the strict realization, transfer function realization allows the additional
freedom of performing a symplectic similarity transformation on the internal vec-
tor x(t). Thus, a question that we may ask is, what kind of simplification can we
3.3 Transfer Function Realization 107
obtain in a transfer function realization of a given system? In particular, are there

systems whose transfer functions can be realized by only a cascade of one degree-
of-freedom linear quantum systems, without requiring any direct Hamiltonian inter-
actions between them or approximate realizations thereof? An affirmative answer is
given here for the latter question, following the treatment in [16]. Moreover, it will
be shown that the class of completely passive systems is such a class of systems and
their cascade realization will be explicitly given.
3.3.2 Conditions for Realizability by a Pure Cascade

Connection
We begin by introducing the following terminology: A square matrix F is said to be

lower 2 × 2 block triangular if it has a lower block triangular form when partitioned
into 2 × 2 blocks:
⎡ ⎤
F11 02×2 02×2 . . . 02×2
⎢ F21 F22 02×2 . . . 02×2 ⎥
⎢ ⎥
F =⎢ . . ,
⎣ .. . . . . . . . . ... ⎥
⎦
Fn1 Fn2 . . . . . . Fnn
where F jk , j ≤ k, is of dimension 2 × 2. We first have the following:
Lemma 3.2 The cascade connection G n G n−1 · · · G 1 of one degree-of-freedom

linear quantum systems G i = (Si , K i xi , (1/2)xi Ri xi ) (i = 1, . . . , n) realizes a
linear
quantum system G = (S, K x, (1/2)x Rx) with S = Sn1 , K =
Sn2 K 1 Sn3 K 2 . . . K n , R = [Ri j ]i, j=1,...,n , where R j j = R j , Rk j = {K k∗
Sk j+1 K j } whenever k < j and R jk = Rkj whenever j > k. In particular, R +
{K ∗ K } is lower 2 × 2 block triangular.
Proof The proof proceeds along the lines of the proof of Theorem 3.1. By the series
product formula from Sect. 3.1.1, for the cascade of two one degree-of-freedom lin-
ear quantum systems G 1 = (S1 , K 1 x1 , (1/2)x1 R1 x1 ) and G 2 = (S2 , K 2 x2 , (1/2)x2
R2 x2 ), we get the oscillator G (2) = G 2 G 1 = (S2 S1 , S2 K 1 x1 + K 2 x2 , (1/2)x1
R1 x1 + (1/2)x2 R2 x2 + x2 {K 2∗ S2 K 1 }x1 ). Letting x(2) = (x1 , x2 ) , the latter may

be compactly written as: G (2) = (S(2) , K (2) x(2) , (1/2)x(2) R(2) x(2) ) with S(2) =
S21 = S2 S1 , K (2) = [ S2 K 1 K 2 ], and

R1 {K 2∗ S2 2 K 1 }
R(2) = {K 2∗ S2 2 K 1 } R2
.
Repeating the computation for G (k) = G k G (k−1) iteratively for k = 3, . . . , n − 1

and writing x(k) = (x1 , x2 , . . . , xk ) and

G (k) = (S(k) , K (k) x(k) , (1/2)x(k) R(k) x(k) )
at each iteration k, we arrive at the desired result with G = G (n) , S = S(n) , K = K (n)
and R = R(n) as stated in the lemma.
To see that R + {K ∗ K } is lower 2 × 2 block triangular, we note that K ∗ K may
be expressed as follows:
⎡ ⎤
K 1∗ K 1 K 1∗ S2∗ K 2 K 1∗ S3∗ 2 K 2 ... K 1∗ Sn∗ 2 K n
⎢ K 2∗ S2 K 1 K 2∗ K 2 K 2∗ S3∗ K 3 ... K 2∗ Sn∗ 3 K n ⎥
∗ ⎢
K K =⎣ .. .. .. .. .. ⎥.
. . . . .
⎦
∗ ∗
K n Sn 2 K 1 K n Sn 3 K 2 ... ∗
K n Sn K n−1 ∗
Kn Kn
Note that since K ∗ K is by definition a Hermitian matrix, the 2 × 2 block elements

above the diagonal blocks are the Hermitian transpose of the corresponding elements
below the diagonal blocks. It follows therefore that the imaginary part of the block
(K ∗ K ) jk at block row j and block column k must satisfy the relation: {(K ∗ K ) jk } =
−{(K ∗ K )k j } . However, from the expression for R derived above and its symmetry,
we already have that if k > j:
R jk = Rkj = {K k∗ Sk j+1 K j } = {(K ∗ K )k j } .
Therefore, the off-diagonal upper block elements of R cancel those of {K ∗ K } when
they are summed, and we conclude that the matrix R + {K ∗ K } is lower 2 × 2 block
triangular.
Recall again the partitioning of R as R = [R jk ] j,k=1,...,n with R jk ∈ R2×2 and of K
as K = [ K 1 K 2 . . . K n ] with K k ∈ Cm×2 . We may now state the following result:
Theorem 3.6 A linear quantum system G = (S, K x, (1/2)x Rx) with n degrees
of freedom is realizable by a pure cascade of n one degree-of-freedom linear quan-
tum systems (without a direct interaction Hamiltonian) if and only if there exists a
permutation matrix P such that
P x = (qπ(1) , pπ(1) , qπ(2) , pπ(2) , . . . , qπ(n) , pπ(n) ) ,
with π some permutation map of {1, 2, . . . , n} to itself, such that the matrix P A P =
2PJn (R + {K ∗ K })P is lower 2 × 2 block triangular. If this condition is satisfied
then G can be explicitly constructed as the cascade connection G n G n−1 . . .
G 1 with G 1 = (S, K 1 x1 , (1/2)(x1 ) R11 x1 ), and G k = (I, K k xk , (1/2)(xk ) Rkk xk )
for k = 2, . . . , n, where xk = (qπ(k) , pπ(k) ) , R = P R P = [Ri j ], K = K P =
[ K 1 K 2 . . . K n ], with Ri j ∈ R2×2 and K j ∈ Cm×2 .
Proof The proof of the only if part is a consequence of Lemma 3.2, as follows.
If G can be realized by a pure cascade connection of n one degree-of-freedom
linear quantum systems, there must exist a permutation map π and an associ-
ated permutation matrix P as given in the statement of the theorem such that
G = (S, K x , (1/2)(x ) R x ) and A = P A P is lower 2 × 2 block triangular,

where x = P x.
Conversely, the if part of the proof can be shown by explicitly constructing
a pure cascade connection of n one degree oscillators that realizes G. Suppose
there is permutation matrix P such that A = P A P is lower 2 × 2 block trian-
gular. Let G = (S, K x , (1/2)(x ) R x ) be as defined in the theorem. Then, the
matrix (1/2)J−1 ∗
n A = −(1/2)Jn A = R + {(K ) K } is also lower 2 × 2 block
triangular. As we already saw in the proof of Lemma 3.2, this structure implies
that R jk = {(K j )∗ K k } whenever k > j and Rk j = {(K j )∗ K k } if k < j. Now,
using the notation of Lemma 3.2, let us define the one degree-of-freedom lin-
ear quantum systems G k for k = 1, . . . , n as G 1 = (S, K 1 x1 , (1/2)(x1 ) Rkk xk ),
and G k = (I, K k xk , (1/2)(xk ) Rkk xk ) for k > 1. It follows from Lemma 3.2 that
G n G n−1 · · · G 1 = (S, K x , (1/2)(x ) R x ). That is, this cascade connection
realizes G.
The following corollary is a direct consequence of Theorem 3.6 on the transfer
function realization of a linear quantum system:
Corollary 3.2 A linear quantum system G = (S, K x, (1/2)x Rx) is transfer func-
tion realizable by a pure cascade connection of one degree-of-freedom linear quan-
tum systems if and only if there is a symplectic transformation matrix V such that
the linear quantum system
G = (S, K V −1 x, (1/2)x V − RV −1 x)
has an A matrix which is lower 2 × 2 block triangular.

Proof By Definition 3.1, G is transfer function realizable by a pure cascade
connection if and only if there exists a symplectic matrix V such that G =
(S, K V −1 x, (1/2)
x V − RV −1 x) is realizable by a pure cascade connection. But from Theorem 3.6,
noting that the matrix P in the theorem can be absorbed into V , this is true if
and only if the A matrix associated with G (i.e., A = 2Jn (R + {K ∗ K }), with
R = V − RV −1 and K = K V −1 ) is lower 2 × 2 block triangular.
3.3.3 Transfer Function Realization of Completely Passive

Linear Quantum Systems
We will now show that any completely passive linear quantum system is transfer
function realizable by a cascade connection. Moreover, each component in the cas-
cade will also be completely passive. The first result in this direction was obtained
by Petersen [17, 18] who constructively showed that a “generic” subclass of such
systems, which satisfy assumptions on the distinctness of the eigenvalues and the
invertibility of certain matrices, is transfer function realizable by pure cascading. It
will be shown in this section, following [16], that the result in fact holds true for all
completely passive systems.
Recall from Sect. 2.4.1 that we can express (1/2)a ∗ R̃a and K̃ a in the form
n
1
(1/2)a ∗ R̃a = (1/2)x { ∗ R̃}x − R̃ j j ,
4 j=1
and K̃ a = K̃ x. Therefore, we may set R = { ∗ R̃} and K = K̃ . Also recall
that the 2 × 2 block diagonal elements {R j j ; j = 1, . . . , n} are of the form R j j =
λ j I2 for some λ j ∈ R for all j. We shall now establish some properties of A and
then prove that there exists a unitary symplectic matrix that transforms it into a lower
2 × 2 block triangular matrix.
Lemma 3.3 [ ∗ ] A [ ∗ ] = diag(M, M # ), where
M = (1/2)Jn ∗ ( R̃ − ı K̃ ∗ K̃ ).
Proof For the proof, we exploit the identities (2.23) as well as the following eas-
ily verified identities: Jn = 0 = # Jn ∗ . Using these identities, we have the
following:
[ ∗ ] A [ ∗ ]
= 2[ ∗ ] Jn (R + {K ∗ K }) [ ∗ ]
= 2[ ∗ ] Jn ({ ∗ R̃} + { ∗ K̃ ∗ K̃ }) [ ∗ ]

= 2[ ∗ ] Jn (1/2)( ∗ R̃ + T R̃ # # )
ı
− ( ∗ K̃ ∗ K̃ − K̃ K̃ # # ) [ ∗ ]
2

= 2[ ∗ ] Jn (1/4) ∗ R̃ − 4ı ∗ K̃ ∗ K̃ (1/4) R̃ # + 4ı K̃ K̃ #

ı ı
= diag (1/2)Jn ∗ R̃ − Jn ∗ K̃ ∗ K̃ , # Jn R̃ # + # Jn K̃ K̃ # .
2 2

Then, we have the following theorem:

Theorem 3.7 Let U be the complex unitary matrix in a Schur decomposition of the
matrix M of Lemma 3.3: M = U ∗ M̂U , where M̂ is a lower triangular matrix. Then,
the matrix
V = 2 [ ∗ ] diag(U, U # )[ ∗ ]
is a real, unitary, and symplectic matrix that transforms A into a lower 2 × 2 block tri-
angular matrix: V AV ∗ = Â, where Â is a real lower 2 × 2 block triangular matrix.
Therefore, every completely passive linear quantum system has a transfer function
realization by pure cascading and such a realization is obtained by applying the
construction of Theorem 3.6 to G = (S, K V x, (1/2)x V RV x). Moreover, each
one degree-of-freedom linear quantum system in the cascade will also be completely
passive.
Proof The existence of U is guaranteed by the well-known result that every com-
plex matrix M has a Schur decomposition of the form M = U ∗ M̂U with M̂ lower
triangular. Note then that we also have M̂ # = U # M # U . Let V be as defined in the
theorem. Then by Lemma 3.3, the following is true:
V AV ∗ = 4 [ ∗ ] diag( M̂, M̂ # )[ ∗ ]
= 4( ∗ M̂ + M̂ # # )
= 8{ ∗ M̂}.
Now, since M̂ is a lower triangular matrix, it follows by inspection (using the special
structure of ) that Â = 8{ ∗ M̂} is lower 2 × 2 block triangular, as claimed.
That V is real follows from the fact that we may write V = 2( ∗ U √+ U # # ) =
∗
∗ ] and
4{ U }. That it is unitary follows from the observation that 2[
√
2 are unitary (as a consequence of (2.23)) and that diag(U, U # ) is also
#
unitary. To see that V is also symplectic define b = U a and z = V x. By the unitarity
of U , we have that b and b# satisfy the same the commutation relations as a and a #
(i.e., b is again an annihilation operator). Then, we have

a a
z = V x = V 2[ ∗ ] = 2[ ∗ ]diag(U, U # ) #
a# a

b
= [ ∗ ] # .
b
Now, this implies that z consists of the canonical position and momentum operators
associated with the modes in b and satisfies the same CCR as x. But since z = V x
and V is real, preservation of the CCR implies that V is necessarily a symplectic
matrix (see, e.g., [19, Sect. III]).
Using the fact V −1 = V = V ∗ established above, it follows from Theorem
3.6 that the completely passive quantum system G is transfer function equivalent
to G = (S, K V x, (1/2)x V RV T x) whose A matrix is lower 2 × 2 block tri-
angular. Let K = K V = [ K 1 K 2 . . . K n ] and R = V RV = [R jk ] j,k=1,...,n . By
Theorem 3.6, we have G = G n G n−1 . . . G 1 with G k = (Sk , K k xk , (1/2)xk
Rkk xk ), S1 = S and Sk = I for k > 1. We now show that each G k is passive.
Recall that K = K̃ and write K̃ = [ K̃ 1 . . . K̃ n ] with K̃ k ∈ Cm×1 . Using (2.23),
we have that K x = K V ∗ x = K̃ U ∗ a. By expanding both sides of the equality
K x = K̃ U ∗ a and collecting and equating terms of the same index, it follows that
K k xk = ( K̃ U ∗ )k ak , where ( K̃ U ∗ )k is the k-th Cm×1 block component of K̃ U ∗ .
On the other hand, since G is completely passive we have that Rkk = λk I2 for
some λk ∈ R, and recalling that R = { ∗ R̃}, it follows by inspection (after
some algebraic manipulations using (2.23)) that Rkk = λk I2 for some λk ∈ R. Thus,
xk Rkk xk = λk ak∗ ak + c for some constant c and we conclude that each G k is also
completely passive.
We conclude the exposition with an example of a cascade realization for a completely

passive linear quantum system.
Example 3.3 Let G = (I, K̃ a, (1/2)a ∗ R̃a + 45 ) be a completely passive system with

2 1+ı 1 + 0.5ı −2 + ı
R̃ = 1 − ı 3 and K̃ = −5 − 2ı 3 − 4ı . Therefore for this system,

0.5 + 0.25ı −0.25 + 0.5ı −1 + 0.5ı −0.5 − ı
K = −2.5 − ı 1 − 2.5ı 1.5 − 2ı 2 + 1.5ı
,
and
0.5I2 0.25(I2 − J )
R= 0.25(I2 + J ) 0.75I2
.
From Lemma 3.3 and Theorem 3.7, we find that

−0.6933 + 0.0039ı 0.2244 − 0.6849ı
U = 0.7204 + 0.0209ı 0.2312 − 0.6536ı ,
and
−14.8390 − 0.7912ı 0
M̂ = 0.6344 − 0.2225ı −0.2235 − 0.4588ı
.
Moreover, from the formula for V stated in Theorem 3.7, we have

⎡ ⎤
−0.6933 0.0039 0.2244 0.6849
⎢ −0.0039 −0.6933 −0.6849 0.2244 ⎥
V = ⎣ 0.7204 −0.0209 0.2312 0.6536 ⎦ .
0.0209 0.7204 −0.6536 0.2312
Let K = K V −1 = [ K 1 K 2 ], then

−0.9144 − 0.7441ı 0.7441 − 0.9144ı −0.1926 − 0.3684ı 0.3684 − 0.1926ı
K = 3.4433 + 1.2621ı −1.2621 + 3.4433ı −0.1679 − 0.1501ı 0.1501 − 0.1679ı
,
and
⎡ ⎤
0.7912 0 0.1113 0.3172
R11 R12 ⎢ 0 0.7912 −0.3172 0.1113 ⎥
R = V RV −1 = = ⎣ 0.1113 −0.3172 0.4588 0 ⎦
.
(R12 ) R22
0.3172 0.1113 0 0.4588
Therefore, G is transfer function realizable by G = (I, K x, (1/2)x R x) as a con-

sequence of Theorem 3.7. It is straightforwardly verified that R + {K ∗ K } is lower
2 × 2 block triangular,
⎡ ⎤
0.7912 14.8390 0 0
∗ ⎢ −14.8390 0.7912 0 0 ⎥
R + {K K } = ⎣ 0.2225 −0.6344 0.4588 0.2235 ⎦
,
0.6344 −0.2225 −0.2235 0.4588
from which it follows that G can be realized by a pure cascade of one degree-of-
freedom linear quantum systems. Applying Theorem 3.6 yields G = G 2 G 1 with
G 1 = (I, K 1 x1 , (1/2)x1 R11 x1 ) and G 2 = (I, K 2 x2 , (1/2)x2 R22 x2 ) and inspection
then shows that G 1 and G 2 are both completely passive. Figure 3.23 depicts a quantum
optical realization of G .
A1 (t) A2 (t)
G1
M11 M12
G2
M21 M22
Y1 (t) Y2 (t)
Fig. 3.23 Realization of G as the cascade connection of G 1 and G 2 . G 1 and G 2 are each realized
by an optical cavity and a phase shifter. Here, θ11 = −2.4585, θ12 = 0.3513, θ21 = −2.0525,
and θ22 = −2.4121, and the partially transmitting mirror M jk in the optical cavity G j , j, k =
1, 2, have the coupling coefficients γ11 = 1.3898, γ12 = 13.4492, γ21 = 0.1728 and γ22 = 0.0507,
respectively. The resonance frequencies of the optical cavities of G 1 and G 2 have a detuning of
0.7912 − 0.4588 = 0.3324, with G 1 having the higher resonance frequency. Figure adapted from
[16] © 2010 IEEE
3.3.4 Further Reading
A surprising recent result on the transfer realization of linear quantum systems is

that it turns out that generic linear quantum systems have a cascade realization, thus
including generic active systems (those that require an external source of quanta to
operate) [20]. This is a result of practical importance as it allows the transfer func-
tion of generic linear quantum systems to be realized by a cascade connection of one
degree-of-freedom linear quantum systems. Thus, the transfer function realization
problem for generic linear quantum systems can be reduced to that of realizing one
degree-of-freedom linear quantum systems. Some approaches for the latter realiza-
tion have been elaborated upon in this chapter.
While the Schur decomposition played a critical role in proving these results
for the completely passive case, in the generic case it was shown that there is a
symplectic Schur decomposition which decomposes a generic real square matrix A of
even dimension into the product A = V J V −1 , with V symplectic and J lower 2 × 2
block triangular. This Schur decomposition was based on an algorithm for symplectic
QR decomposition, and necessary and sufficient conditions for a generic real square
matrix A with even dimension to possess the latter decomposition were obtained.
Using techniques from differential topology, it was shown that these conditions are
satisfied generically by real square matrices of a fixed even dimension, say 2n × 2n,
in the sense that it is valid for matrices in an open and dense subset of R2n×2n .
In general, the realization of linear quantum systems can be complex with many
degrees of freedom. To reduce the dimension of the transfer function realization
of linear quantum systems, model reduction of such systems has been considered,
with the goal of reducing the number of internal degrees of freedom required in the
realization. Methods that have been proposed include an adaptation of the well-known
balanced truncation paradigm to linear quantum systems [21, 22] and tangential
interpolatory projection methods [23]. The physical realizability condition for linear
quantum systems places a stringent restriction on the feasibility of performing model
reduction. In particular, it is found that generic linear quantum systems do not possess
a balanced realization. However, it has been shown that completely passive linear
quantum systems always possess a balanced realization and they are a special class
of systems for which model reduction can often be successfully applied, either by
balanced truncation or by tangential interpolatory projection.
Appendices
Appendix A: Adiabatic Elimination of Coupled Cavity

Modes
In this section, we shall derive formulas for two coupled cavity modes in which one
of the cavity has very fast dynamics compared to the other and can be adiabatically
Appendix A: Adiabatic Elimination of Coupled Cavity Modes 115
eliminated, leaving only the dynamics of the slow cavity mode. The cavities are each
coupled to separate bosonic fields and are interacting with one another in a classically
pumped nonlinear crystal. A mathematically rigorous theory for the type of adiabatic
elimination/singular perturbation that we are interested in here was developed in [12].
The two cavity modes will be denoted by a and b, each defined on two distinct
copies of the Hilbert space 2 of square-integrable sequences (Z+ denotes the set
of all nonnegative integers). Thus, the composite Hilbert space for the two cavity
modes is H = 2 (Z+ ) ⊗ l 2 (Z+ ). The interaction in a nonlinear crystal is given, in
some rotating frame, by an interaction Hamiltonian Hab of the form Hab = αa ∗ b +
βa ∗ b∗ + α∗ ab∗ + β ∗ ab, for some complex constants α and β. The mode a is coupled
to a bosonic field A1 , while b is coupled to the bosonic field A2 , both fields in the
vacuum state. The fields A1 and A2 live the boson Fock space F2 = F1 ⊗ F1 . We
take a to be the slow mode to be retained and b to be the fast mode to be eliminated.
(k)
We consider a sequence of linear quantum systems G k = (I, L̃ (k) , Hab ) with
√ √ (k)
L̃ (k) = ( γ1 a, k γ2 b) and Hab = 1 a ∗ a + k 2 2 b∗ b + k(αa ∗ b + βa ∗ b∗ +
α∗ ab∗ + β ∗ ab) each evolving according to the unitary Uk satisfying the left
Hudson–Parthasarathy (H–P) QSDE (as opposed to the conventional right H–P
QSDE in (2.2)):

dUk (t) = Uk (t) L̃ (k)∗ (dA1 (t), dA2 (t)) − L̃ (k) (dA1 (t), dA2 (t))∗

(k)
+ ı Hab − (1/2) L̃ (k)∗ L̃ (k) dt ,
Here, we are using the left QSDE following the convention used in [12] (see Remark
3 therein) so that the Heisenberg picture dynamics of an operator x is given by
x(t) = Uk (t)xUk (t)∗ . We shall use the results of [12] to show, in a similar treatment
to Sect. 3.2 therein, that in the limit as k → ∞:
lim sup Uk (t)∗ φ − U (t)∗ φ = 0 ∀φ ∈ H0 ⊗ F2 (3.16)

k→∞ 0≤t≤T
for any fixed time T > 0, where H0 is an appropriate Hilbert subspace of H (to be
precisely specified in the next paragraph) for a limiting unitary U (t) (again as a left
H–P QSDE) satisfying:

ı22 + γ2 √ √
dU (t) = U (t) − 1 d 22 + γ1 a ∗ dA1 (t) − γ1 adA1 (t)∗
ı22 − γ2
√ γ2
− ı γ2 (ı2 − )−1 (αa ∗ + β ∗ a)dA2 (t) +
2
√
2 γ2 γ1
ı (α∗ a + βa ∗ )dA2 (t)∗ + (ı1 − )a ∗ adt +
ı22 − γ2 2
γ2 −1
(ı2 − ) (αa ∗ + β ∗ a)(α∗ a + βa ∗ )dt , (3.17)
2
on H0 ⊗ F. Note that (3.17) is a left H–P QSDE corresponding to the right form in
Sect. 2.1.4 by noting that we may write:
γ1 ∗ γ2
(ı1 − )a a + (ı2 − )−1 (αa ∗ + β ∗ a)(α∗ a + βa ∗ )
2 2
∗ 2 ∗ ∗ ∗ ∗ 1 ∗ ∗
= ı 1 a a − 2 γ2 2 (αa + β a)(α a + βa ) − ( L̃ 1 L̃ 1 + L̃ 2 L̃ 2 ),
2 + ( 2 ) 2
√ √
with L̃ 1 = γ1 a and L̃ 2 = ı γ2 (−ı2 − γ22 )−1 (α∗ a + βa ∗ ). As such it satisfies
the H–P Condition 1 of [12].
Let φ0 , φ1 , . . . be the standard orthogonal bases of 2 , i.e., φl is an infinite sequence
(indexed starting from 0) of complex numbers with all zeros except a 1 in the lth
place. First, let us specify that H0 = 2 ⊗ Cφ0 , and this is the subspace of H where
the slow dynamics of the system will evolve. Next, we define a dense domain
D = span{φ j ⊗ φl ; j, l = 0, 1, 2, . . .} of H. The strategy is to show that [12,
Assumptions 2–3] are satisfied from which the desired result will follow from [12,
Theorem 3].
(k)
From the definition of Hab , L̃ (k) and Uk given above, we can define the
operators Y, A, B, G 1 , G 2 , and W jl ( j, l = 1, 2) in [12, Assumption 1] as:
Y = (i2 − γ22 )b∗ b, A = i(αa ∗ b + α∗ ab∗ + βa ∗ b∗ + β ∗ ab), B = (i1 − γ21 )a ∗ a,
√ √
G 1 = γ1 a ∗ , G 2 = 0, F1 = 0, F2 = γ2 b∗ , W jl = δ jl . Then, we can define the
operators K (k) , L (k)j in this assumption as:
K (k) = k 2 Y + k A + B; L (k)
j = k F j + G j ( j = 1, 2).
Let P0 be the projection operator to H0 . Let us now address Assumption 2. From

our definition of H0 , it is clear that we have that (a) P0 D ⊂ D. Any element of
P0 D is of the form f ⊗ φ0 for some f ∈ span{φl ; l = 0, 1, 2, . . .}; therefore, since
Y = (ı2 − γ22 )b∗ b and bφ0 = 0, we find that (b) Y P0 d = 0 for all d ∈ D. Define
the operator Ỹ on D by Ỹ f ⊗ φ0 = 0 and Ỹ f ⊗ φl = l −1 (ı2 − γ22 )−1 f ⊗ φl for
l = 1, 2, . . . (Ỹ can then be defined to all of D by linear extension). From the definition
of Y and Ỹ , it is easily inspected that (c1) Y Ỹ f = Ỹ Y f = P1 f for all f ∈ D, where
P1 = I − P0 (i.e., the projection onto the subspace of H complementary to H0 ).
Moreover, because of the simple form of Ỹ , it is also readily inspected that (c2) Ỹ
has an adjoint Ỹ ∗ with a dense domain that contains D. Since F1 = 0, we have that
√
(d1) F1∗ P0 = 0 on D, while since F2∗ f ⊗ φ0 = γ2 b f ⊗ φ0 = 0 ∀ f ∈ 2 , we also
have (d2) F2∗ P0 = 0 on D. Finally, from the expression for A and the orthogonality
of the bases φ0 , φ1 , . . ., a little algebra reveals that (e) P0 A P0 d = 0 for all d ∈ D.
From (a), (b), (c1–c2), (d1–d2), and (e), we have now verified that Assumption 2 is
satisfied.
Finally, we verify that the limiting operator coefficients K , L 1 , L 2 , M1 , M2 ,
N jk (i, j = 1, 2) (as operators on H0 ) of Assumption 3 coincide with the correspond-
ing coefficients of (3.17). These operator coefficients are defined as K = P0 (B −
Appendix A: Adiabatic Elimination of Coupled Cavity Modes 117
AỸ A)P0 , L j = P0 (G j − AỸ F j )P0 , M j = − r2=1 P0 W jr (G r∗ − Fr∗ Ỹ A)P0 and

N jl = r2=1 P0 W jr (Fr∗ Ỹ Fl + δrl )P0 . From these definitions and some straightfor-
ward algebra, we find that for all f ∈ span{φl ; l = 0, 1, 2, . . .}:
γ1 γ2
K f ⊗ φ0 = (ı1 − )a ∗ a + (ı2 − )−1 (αa ∗ + β ∗ a)(α∗ a + βa ∗ ) f ⊗ φ0 ,
√ 2 2
L 1 f ⊗ φ0 = γ1 a ∗ f ⊗ φ0 ,
√ γ2
L 2 f ⊗ φ0 = −ı γ2 (i2 − )−1 (αa ∗ + β ∗ a) f ⊗ φ0 ,
√ 2
M1 f ⊗ φ0 = − γ1 a f ⊗ φ0 ,
√ γ2 −1 ∗
M2 f ⊗ φ0 = γ2 ı2 − (α a + βa ∗ ) f ⊗ φ0 ,
2
and
N11 f ⊗ φ0 = f ⊗ φ0 , N12 f ⊗ φ0 = 0, N21 f ⊗ φ0 = 0,

γ2 + ı22
N22 f ⊗ φ0 = f ⊗ φ0 .
−γ2 + ı22
Therefore, we see that U (t) may be written as:

⎛ ⎞
2 2 2
dU (t) = U (t) ⎝ (N jl − δ jl )d jl + M j dA∗j (t) + L j dA j (t) + K dt ⎠ .
j,l=1 j=1 j=1
Since we have already verified that (3.17) is bona fide left QSDE equation, it now
follows that Assumption 3 of [12] is satisfied. Now (3.16) follows from [12, Theorem
3] and the proof is complete.
Moreover, we can observe from the derivation above that the coupling of a to
A2 (t) after adiabatic elimination will not change if a is also coupled to other cav-
m ∗
ities modes b3 , . . . , bm via an interaction Hamiltonian of the form i= j (α j1 ab j +
∗ ∗ ∗ ∗ ∗
α j1 a b j + α j2 a b j + α j2 ab j ), and each additional mode may also be linearly cou-
pled to distinct bosonic fields A3 , . . . , Am , respectively, as long as these other
modes are not interacting with b and with one another (this amounts to just intro-
ducing additional operators F j , G j , j ≥ 3, etc.). Moreover, under these conditions,
one can also adiabatically eliminate any of the additional modes and the only effect
will be the presence of additional sum terms in U (t) that do not involve b, A1 (t) and
A2 (t).
Appendix B: Proof of Theorem 3.4
For this proof, it will be convenient to interchange some rows and columns of the
model matrix M to form another model matrix M̃ to avoid complicated bookkeeping
and thus reduce unnecessary clutter. Hereby, “rows” and “columns” we mean, respec-
tively, block rows and block columns of M formed with respect to its specified
partitioning. This interchange is as follows.
First, we permute rows of M such that the first 2n − 1 rows from top to bottom
are the rows labeled (while column labels are kept fixed as they are) s00 , s12 , s21 , s13 ,
s31 , . . ., s1n , sn1 , the next 2(n − 2) rows, respectively, are the rows labeled s23 , s32 ,
s24 , s42 , . . ., s2n , sn2 , and so on in the same pattern until we get to the last n rows that
are, respectively, those rows labeled s11 , s22 , . . ., snn . Call the intermediate matrix
resulting from this row permutation M̂. Then fixing the row labels of M̂, we permute
its columns such that the first 2n − 1 columns from left to right are, respectively, the
columns of M̂ labeled r00 , r12 , r21 , r13 , r31 , . . ., r1n , rn1 , the next 2(n − 2) columns
are, respectively, the columns labeled r23 , r32 , r24 , r42 , . . ., r2n , rn2 , and so on in the
same pattern until the final n columns that are, respectively, the columns labeled r11 ,
r22 , . . ., rnn . The resulting matrix after this permutation of columns is M̃.
It is important to note here that since the same permutation is applied to the
rows and columns, M and M̃ are model matrix representations of the same physical
system. That is to say that if M is the model matrix of G = (S, L , H ), then M̃ is the
model matrix of G̃ = (P S P , P L , H ) for some suitable constant real permutation
matrix P, while it is clear that G and G̃ are representations of the same physical
system. Thus with the same internal connections made, a reduced model matrix for
M̃ is also a reduced model matrix for M, up to a possible relabeling of uneliminated
ports.
Let L̃ = P L and S̃ = P S P . Then, L̃ can be partitioned as L̃ = ( L̃
i , L̃ e ) ,
where L̃ i is the first n(n − 1) + 1 rows of L̃, while L̃ e is the last n rows of L̃. They
are of the form:
L̃ i = (L
12 , L 21 , L 13 , L 31 , . . . , L 1n , L n1 , L 23 , L 32 , L 24 , L 42 , . . . , L 2n , L n2 , . . . ,
L
(n−2)(n−1) , L (n−1)(n−2) , L (n−2)n , L n(n−2) , L (n−1)n , L n(n−1) ) ,

L̃ e = (L
11 , L 22 , . . . , L nn ) .

S̃ii S̃ie
Similarly, S̃ can be partitioned as S̃ = , with S̃ii and S̃ee being block
S̃ei S̃ee
diagonal:
S̃ii = diag(S12 , S21 , S13 , S31 , . . . , S1n , Sn1 , S23 , S32 , S24 , S42 , . . . , S2n , Sn2 , . . . ,
S(n−2)(n−1) , S(n−1)(n−2) , S(n−2)n , Sn(n−2) , S(n−1)n , Sn(n−1) ),
S̃ee = diag(S11 , S22 , . . . , Snn ),
and S̃ei and S̃ie both being zero matrices. Then, M̃ has a partitioning of the form
(3.13) by identifying Sii , Sie , Sei , See , L i and L e with S̃ii , S̃ie , S̃ei , S̃ee and L̃ i , and L̃ e ,
respectively. The reduced model matrix resulting from the subsequent simultaneous
elimination of all internal edges (s jk , rk j ) j, k = 1, . . . , n, j = k, can be conveniently
determined by using the adjacency matrix η defined by:
Appendix B: Proof of Theorem 3.4 119

0 Ic12 0 Ic13
η = diag , ,...,
Ic21 0 Ic31 0

0 Ic1n 0 Ic23 0 Ic24
, , ,
Icn1 0 Ic32 0 Ic42 0

0 Ic2n 0 Ic(n−1)n
..., ,..., .
Icn2 0 Ic(n−1)n 0
(Recall that c jk = ck j ). Hence, according to Theorem 3.3, the reduced model matrix
M̃red obtained after elimination of the internal edges {(s jk , rk j ); j, k = 1, . . . , n, j =
k} has parameters given by (recalling that S̃ei and S̃ie are zero matrices):
S̃red = S̃ee + S̃ei (η − S̃ii )−1 S̃ie = S̃ee ,

L̃ red = L̃ e + S̃ei (η − S̃ii )−1 L̃ i = L̃ e ,
n
H̃red = Hk + { L̃ ∗j S̃ ji (η − S̃ii )−1 L̃ i }
k=1 j=i,e
n
= Hk + { L̃ ∗i S̃ii (η − S̃ii )−1 L̃ i }
k=1
n
= Hk + { L̃ ∗i η(η − S̃ii )−1 L̃ i }
k=1
n n−1 n −1
∗ ∗ I −S jk L jk
= Hk + [ L jk L k j ] .
−Sk j I Lkj
k=1 j=1 k= j+1
Since M and M̃ are model matrix representations of the same physical system and
the external fields have the same ordering and labeling in both representations, the
reduced model matrix of Mred and M̃red of M and M̃, respectively, after elimination of
internal edges (s jk , rk j ), coincide. Hence, also the linear quantum stochastic systems
G red and G̃ red associated with M and M̃, respectively, coincide. This completes the
proof.
Appendix C: Proof of Lemma 3.1
We begin by noting that

S12 S21 1 K2
K 1∗ K 2 + K 1 K 2# = [ −K 1∗ ∗ K 1 ] ,
1 − S12 S21 1 − S12 S21 2ı K 2#
∗
S12 S −S ∗ ∗
with = S21
1−S21 S12
− = 2 |1−S
∗ ∗
1−S21 S12
21 12
21 S12 |
2 (exploiting the fact that S12 S12 = 1 =
∗
S21 S21 ). Now, set K 1 = [ κ ıκ ] for an arbitrary nonzero real constant κ, and note
that S21 S12 = 1 implies that = 0 and:
−1
−κ∗ κ
[ −K 1∗ ∗ K 1 ]−1 =
ıκ∗ ıκ

1 ıκ −κ
=− ,
2ıκ2 ||2 −ıκ∗ −κ∗

Z
and therefore for any real matrix V , 2ı[ −K 1∗ ∗ K 1 ]−1 V = for some com-
Z#
plex row vector Z . Therefore, given any R, we see that we may solve the equation

K2
[ −K 1∗ ∗ K 1 ] = 2ı R,
K 2#
for K 2 and this solution is as given in the statement of the corollary.

Alternatively, we could also have started by setting K 2 = [ κ ıκ ] and analogously
solving for K 1 for a given R. It is then an easy exercise that the solution for K 1 in
this case is as stated in the corollary.
Appendix D: Proof of Corollary 3.1
With c jk , S jk , R jk , and K jk , j, k = 1, . . . , n, as defined in the statement of the

corollary, from Theorem 3.4 and Lemma 3.1 we have that Sred = Inm , L red =
(L
11 , L 22 , . . . , L nn ) with L j j = K j x j , and
n n−1 n −1
1 −S jk L jk
Hred = Hj + [ L ∗jk L ∗k j ] .
−Sk j 1 Lkj
j=1 j=1 k= j+1
Expanding, we have:
n n−1 n
1
Hred = (1/2) x j R j x j +
j=1 j=1 k= j+1
1 − S jk Sk j

×(L ∗jk L jk + S jk L ∗jk L k j + Sk j L ∗k j L jk + L ∗k j L k j )

n n
K ∗jk K jk
= (1/2) x j R j + 2sym { } xj
j=1 k=1,k= j
1 − S jk Sk j
Appendix D: Proof of Corollary 3.1 121
n−1 n
S jk Sk j
+ x j { K ∗ Kk j + K K # }xk
j=1 k= j+1
1 − S jk Sk j jk 1 − S jk Sk j jk k j
n n−1 n

= (1/2) x j R j j x j + x j R jk − {K
j K k } xk
#
j=1 j=1 k= j+1

n−1 n
= (1/2)x Rx − x j {K
j K k }x k ,
#
j=1 k= j+1
where sym(A) = (1/2)(A + A ), and R = [R jk ] j,k=1,...,n and Rk j = R jk . From this,

it is clear that using the concatenation product we can decompose G red as G red =
(0, 0, Hred ) nj=1 (Im , L j j , 0). Let G red,0 = (0, 0, Hred ) and G red, j = (Im , L j j , 0),
j = 1, . . . , n. Now, using the series product rule, we easily compute that
G net = G red,0 (G red,n . . . G red,2 G red,1 )

n−1 n
= 0, 0, (1/2)x Rx − x j {K j K k }x k
#
j=1 k= j+1
n−1 n
Im , [ K 1 K 2 . . . K n ]x, x j {K
j K k }x k
#
j=1 k= j+1

= Im , [ K 1 K 2 . . . K n ]x, (1/2)x Rx .
Therefore, G net realizes a linear quantum stochastic system with parameters Snet ,
L net and Hnet , as claimed.
References
1. J.C. Willems, J.W. Polderman, Introduction to Mathematical Systems Theory: A Behavioral

Approach. Texts in Applied Mathematics, vol. 26 (Springer, Berlin, 1998)
2. B.D.O. Anderson, S. Vongpanitlerd, Network Analysis and Synthesis: A Modern Systems The-
ory Approach. Networks Series (Prentice-Hall Inc, Upper Saddle River, 1973)
chastic systems. SIAM J. Control Optim. 48(4), 2686–2718 (2009), © 2009 Society for Indus-
trial and Applied Mathematics. Reprinted with permission. All rights reserved
4. U. Leonhardt, Quantum physics of simple optical instruments. Rep. Prog. Phys. 66, 1207–1249
(2003)
6. J. Gough, M.R. James, Quantum feedback networks: Hamiltonian formulation. Commun.
Math. Phys. 287, 1109–1132 (2009)
2004)
8. C.M. Caves, B.L. Schumaker, New formalism for two-photon quantum optics. I. Quadrature
phases and squeezed states. Phys. Rev. A 31, 3068–3092 (1985)
9. B. Schumaker, C.M. Caves, New formalism for two-photon quantum optics. ii. Mathematical
foundation and compact notation. Phys. Rev. A 31, 3093–3111 (1985)
10. U. Leonhardt, A. Neumaier, Explicit effective Hamiltonians for linear quantum-optical net-
works. J. Opt. B 6, L1–L4 (2004)
11. H.M. Wiseman, G.J. Milburn, Quantum theory of field-quadrature measurements. Phys. Rev.
A 47(1), 642–663 (1993)
12. L. Bouten, R. van Handel, A. Silberfarb, Approximation and limit theorems for quantum
stochastic models with unbounded coefficients. J. Funct. Anal. 254, 3123–3147 (2008)
13. H.I. Nurdin, Synthesis of linear quantum stochastic systems via quantum feedback networks.
IEEE Trans. Autom. Control 55(4), 1008–1013 (2010). Reprinted, with permission, © 2010
IEEE. Extended preprint version available at arXiv:0905.0802
1846 (2009)
16. H.I. Nurdin, On synthesis of linear quantum stochastic systems by pure cascading. IEEE Trans.
Autom. Control 55(10), 2439–2444 (2010) © 2010 IEEE
17. I.R. Petersen, Cascade cavity realization for a class of complex transfer functions arising in
coherent quantum feedback control, in Proceedings of the 2009 European Control Conference
(Budapest, Hungary, 23–26 August 2009), pp. 190–195
18. I.R. Petersen, Cascade cavity realization for a class of complex transfer functions arising in
coherent quantum feedback control. Automatica 47(8), 1757–1763 (2011)
19. G. Linblad, Brownian motion of harmonic oscillators: existence of a subdynamics. J. Math.
Phys. 39(5), 2763–2780 (1998)
20. H.I. Nurdin, S. Grivopoulos, I.R. Petersen, The transfer function of generic linear quantum
stochastic systems has a pure cascade realization. Automatica, pp. 324–333 (2016)
systems. IEEE Trans. Autom. Control 59(9), 2413–2425 (2014)
22. O. Techakesari, H.I. Nurdin, On the quasi-balanceable class of linear quantum stochastic sys-
tems. Syst. Control Lett. 78, 25–31 (2015)
23. O. Techakesari, H.I. Nurdin, Tangential interpolatory projection for model reduction of linear
quantum stochastic systems. IEEE Trans. Autom. Control 62(1), 5–17 (2017)
24. H.I. Nurdin, On synthesis of linear quantum stochastic systems by pure cascading (2010).
arXiv:0905.0802 (arXiv preprint)
Chapter 4
Quantum Filtering for Linear Dynamical
Quantum Systems
Abstract This chapter introduces quantum filtering theory as a quantum analogue

of stochastic filtering theory. When applied to linear quantum systems, this leads
to a quantum version of the Kalman filter. Concepts from quantum probability the-
ory that are relevant to quantum filtering theory are introduced, including a suit-
able notion of quantum conditional expectations. The chapter also briefly introduces
robust observers for linear quantum systems to deal with modeling uncertainties for
the purpose of estimation.
As discussed in Sect. 1.2.2, for classical linear stochastic systems with partial obser-
vation, we can construct a dynamical estimate called the Kalman filter, which gener-
ates an optimal estimator for the system’s state vector in the least mean square sense.
In order to emphasize the importance of the Kalman filter, we highlight the fact that
it can be used for constructing LQG controllers based on the separation principle, as
illustrated in Sect. 1.2.2. This chapter discusses the quantum analogue of the Kalman
filtering theory for linear dynamical quantum systems, as well as a more general
setting of the quantum filtering theory developed by Belavkin (see Chap. 1).
This chapter is structured as follows. First, we begin with a general overview
of a version of quantum conditional expectations, which gives a least mean square
estimate as in the classical case. Then, we use the quantum conditional expectation
to derive a special case of the quantum filtering equation. The Schrödinger picture
representation of this equation will be given, which is called the stochastic master
equation or the quantum trajectory equation in the physics literature. Next, the general
theory is applied to the special linear setup where the state is limited to coherent states,
and the quantum version of the Kalman filter is given. Finally, we consider a practical
aspect of estimation for linear quantum systems. Quantum filtering is a model-based
approach for estimation; as a consequence, the optimal estimate can be constructed
only under the assumption that the system dynamics (more precisely the system’s
(A, B, C, D) matrices) are perfectly known. Of course in any realistic situation, the
model is only an approximation. Hence, we need a robust filter that can estimate the
Section 4.1 contains some materials reprinted, with permission, from [2] © 2009 IEEE.
Section 4.4 contains reprinted excerpt with permission from [8]. Copyright (2006) by the Amer-
ican Physical Society.
124 4 Quantum Filtering for Linear Dynamical Quantum Systems
system’s variable, even when the model is imperfect. A quantum version of robust
filters will be discussed.
4.1 Quantum Conditional Expectations
4.1.1 Quantum Probability Space
In the framework of quantum filtering, we will only consider a restricted notion

of quantum conditional expectations, as linear projective mappings from the space
of observables to an algebra of commutative observables. Hence, we first need to
specify such an algebra of observables and the notion of a state on this algebra. For the
purpose of illustration and to emphasize the basic underlying idea, in the following
we discuss only the case of finite-dimensional systems, wherein the dimension of
the Hilbert space is finite-dimensional and an observable can be represented by a
complex matrix. Quantum harmonic oscillators are thus, strictly speaking, excluded
because they are defined on an infinite-dimensional Hilbert space. Nonetheless, the
notions and results below have an infinite-dimensional analogue, which readers can
find in, e.g., [1] and the references cited therein.
The content we describe here partly follows [2]. First recall that a quantum
mechanical random variable is represented by a linear self-adjoint operator on a
Hilbert space. This axiom means that quantum random variables in general do not
commute with one another, and thus, the conventional classical probability space
(, F, P) needs to be replaced by a quantum probability space, which is composed
of a ∗-algebra and a state defined as follows.
Definition 4.1 (∗-algebra) Let H be a finite-dimensional complex Hilbert space.
A ∗-algebra A is a set of linear operators H → H such that I, α A + β B, AB, A∗ ∈
A for any A, B ∈ A and α, β ∈ C. A is called commutative if [A, B] = 0 for any
A, B ∈ A.
In the above definition, the adjoint A∗ is defined as the unique operator satisfying
A∗ x, y = x, Ay, ∀x, y ∈ H. The ∗-algebra introduced above corresponds to a
σ-algebra F in a classical probability space. The next is the quantum version of
the expectation E, which is referred to as a “state”; this notion has a one-to-one
correspondence with the standard notion of a state in quantum mechanics, which is
represented by a ket vector or a density matrix.
Definition 4.2 (State) A state on A is a linear map P : A → C that is positive:
P(A∗ A) ≥ 0, ∀A ∈ A, and normalized P(I ) = 1.
In the finite-dimensional context, the state can be given explicitly as follows. Let
(e1 , . . . , ed ) be an orthonormal basis of the d-dimensional
d Hilbert space H. Also
recall that the trace can be defined as Tr(A) = i=1 ei , Aei for all A ∈ A; note
that Tr(A) does not depend on the choice of basis. Then, the state P can be expressed
in terms of the trace and an associated matrix. That is, there exists a unique matrix
ρ that satisfies
4.1 Quantum Conditional Expectations 125
P(A) = Tr(ρA), ρ = ρ∗ ≥ 0, Tr(ρ) = 1. (4.1)
ρ is called the density matrix; due to the equivalence between ρ and P, we some-
times also refer to ρ as the state. Note that, from the latter two conditions in (4.1),
P(A∗ A) ≥ 0 ∀A ∈ A and P(I ) = 1 are satisfied. Also especially when ρ is a pure
state, meaning that ρ is a rank-one projection matrix ρ = |bb|, |b ∈ H, then P(A)
can be represented as
P(A) = Tr(|bb|A) = b|A|b. (4.2)
Now a quantum probability space can be defined as follows.
Definition 4.3 (Quantum probability space) Let A be a ∗-algebra of operators on

a finite-dimensional complex Hilbert space H and P be a state on A. Then, (A, P)
is called a (finite-dimensional) quantum probability space.
For a quantum probability space (A, P), if A ∈ A is self-adjoint, then we call A

a quantum random variable or an observable; this is the definition of observables
in the scenario of quantum probability theory. Now, let us consider a commutative
∗-algebra A, in which all quantum random variables in A commute with each other as
in the classical case. In this case, (A, P) is called a commutative quantum probability
space. Now, we can state the well-known Spectral Theorem (Theorem 4.1 below),
emphasizing that a commutative quantum probability space is equivalent to a classical
one. This is based on the following notion:
Definition 4.4 (∗-isomorphism) Let be a set and let F be a σ-algebra on .

A ∗-isomorphism between a commutative ∗-algebra C and the set of bounded
F-measurable functions ∞ (F) on is a linear bijection ι : C → ∞ (F) such that
ι(A∗ )(i) = ι(A)(i)∗ and ι(AB)(i) = ι(A)(i)ι(B)(i) for all A, B ∈ C and i ∈ .
Theorem 4.1 (Spectral Theorem) Let (C, P) be a finite-dimensional commutative

quantum probability space. Then, there exists a classical probability space (, F, P)
and a ∗-isomorphism ι : C → ∞ (F) such that P(A) = EP [ι(A)], ∀A ∈ C.
Proof The theorem can be proven by construction. First, let H = Cn and =

{1, . . . , n}. Since [A, B] = 0 and [A, A∗ ] = 0 ∀A, B ∈ C, all the elements in C can
be diagonalized simultaneously. Hence, we can set A = diag{a1 , . . . , an } and define
a classical random variable ι(A) : → C by ι(A)(i) = ai . Let P be a projection in
C, i.e., P = P ∗ = P 2 , then ι(P) is the indicator function of a subset S P of . We
define F as the set of subsets S P of where P runs through the projections in C.
Furthermore, we define a probability measure P on F by P(S P ) = P(P), ∀P ∈ C.
As a result, we have constructed a classical probability space (, F, P). It is then
easy to verify EP [ι(A)] = P(A).
Thanks to the above spectral theorem, a single observable A can always be repre-
sented as a classical random variable ι(A) on a classical probability space (, F, P),
where the measure P is determined from the state. Then, a measurement on A gen-
erates one of the values that ι(A) can take, with probability according to P. Here, we
pose a very important remark; if two observables do not commute with each other,
then these observables cannot be represented as classical random variables on the
same probability space. Such observables are called incompatible, meaning that they
cannot be measured simultaneously in a single realization of an experiment.
Example 4.1 Let us now see an example, a quantum two-level system, like a single
atom with two distinguishable two energy levels. The Hilbert space is H = C2 . The
quantum probability space of a two-level system is composed of the ∗-algebra of
2 × 2 complex matrices M and a state ψ on M. The state ψ can be represented as
ψ(A) = Tr(ρA), ∀A ∈ M for some matrix ρ satisfying ρ = ρ∗ ≥ 0 and Tr(ρ) = 1.
Let us then study a commutative ∗-subalgebra D = {D = diag(d1 , d2 ) : d1 , d2 ∈
R} ⊂ M. From Theorem 4.1, there always exists a classical probability space that
is in one-to-one correspondence with (D, ψ). Such a classical space is composed of
the sample space = {1, 2}, the set of events F = {∅, {1}, {2}, }, and the proba-
bility measure constructed as follows. Let d1 and d2 be eigenvalues of an observable
D in D and define a classical random variable ι(D) on (, F) as ι(D)(1) = d1
and ι(D)(2) = d2 . Now, D ∈ D has a spectral decomposition D = di Pi with the
projection matrices P1 = diag(1, 0) and P2 = diag(0, 1). Hence, according to the
proof of Theorem 4.1, a probability distribution for ι(D) can be defined through
P({1}) = ψ(P1 ) = Tr(ρP1 ) = ρ11 and P({2}) = ρ22 .
Lastly, let us define the composite quantum probability space of two quantum
probability spaces (A1 , P1 ) and (A2 , P2 ), which are defined on the finite-dimensional
Hilbert spaces H1 and H2 , respectively. First, for two vectors |a1 ∈ H1 and
|a2 ∈ H2 , we define the tensor (Kronecker) product |a1 ⊗ |a2 . Then, the Hilbert
space H1 ⊗ H2 can be introduced by defining an inner product a1 ⊗ a2 , b1 ⊗ b2 :=
a1 , b1 a2 , b2 . Now, the composite quantum probability space (A1 ⊗ A2 , P1 ⊗ P2 )
can be defined on H1 ⊗ H2 as follows. First, an arbitrary element of A1 ⊗ A2 is rep-
resented as a linear combination of product-type elements of the form A1 ⊗ A2 ∈
A1 ⊗ A2 . Second, the state P1 ⊗ P2 is introduced through the relation (P1 ⊗
P2 )(A1 ⊗ A2 ) = P1 (A1 )P2 (A2 ) and its linear extension to all of A1 ⊗ A2 .
4.1.2 Conditional Expectations
Let us consider a quantum probability space (A, P) and two commuting self-adjoint
elements A and B on A. From the Spectral Theorem (Theorem 4.1), A and B can be
represented as classical random variables ι(A) and ι(B) on a single classical prob-
ability space (, F, P). This allows us to specify the classical conditional expec-
tation EP [ι(A) | ι(B)]. Then, the quantum conditional expectation P(A | B) can be
defined as

P(A | B) = ι−1 EP [ι(A) | ι(B)] .
4.1 Quantum Conditional Expectations 127
Next, we aim to extend the above discussion to the case where the variable A can be
conditioned on a commutative ∗-subalgebra C of A, rather than a single variable B.
In fact, as long as A commutes with every element in C, Theorem 4.1 is applicable
to the commutative ∗-algebra generated from A and C. In this case, we can define

P(A | C) = ι−1 EP [ι(A) | σ(ι(C))] , (4.3)
where σ(ι(C)) denotes the σ-algebra generated from ι(K ), K ∈ C. Moreover, this
argument means that the quantum conditional expectation onto C can be defined for
every self-adjoint element in the commutant of C: for a commutative ∗-subalgebra
C, the commutant of C is given by
C := {A ∈ A : [A, C] = 0, ∀C ∈ C}.
Then, we have the formal definition of the conditional expectation as shown below
(note that the general treatment of quantum conditional expectations based on oper-
ator algebras can be found in, e.g., [3]).
Definition 4.5 (Quantum conditional expectation) Let (A, P) be a quantum prob-
ability space, and let C be a commutative ∗-subalgebra of A. Then, the map
P( · | C) : C → C is called (a version of) the quantum conditional expectation from
C to C if P(P(A | C)K ) = P(AK ), ∀A ∈ C , ∀K ∈ C.
Note that an arbitrary element A in the commutant can be uniquely expressed
as A = A1 + ı A2 with A1 and A2 self-adjoint. Hence, the conditional expectation
of A onto C can be defined as simply a linear extension of (4.3). Finally, we pro-
vide some basic properties of both the classical and quantum conditional expec-
tations as follows: (i) P(A | C) is unique with probability one, (ii) P(P(A | C)) =
P(A), (iii) P(C A | C) = CP(A | C) if C ∈ C and A ∈ C (module property), and (iv)
P(P(A | B) | C) = P(A | C) if C ⊂ B (tower property). Note from the tower property
that P( · | C) is a projection. A particularly important fact deduced from the above
properties is that, as in the classical case, P(A | C) is a least mean square estimate of
A given C in the sense that
A − P(A | C)P ≤ A − P(A | C)P + P(A | C) − BP

= A − BP , ∀B ∈ C, (4.4)
where we have defined X 2P := P(X ∗ X ).
4.2 Quantum Filtering Theory
In this section, in the setting of the quantum probability theory elaborated in the
previous section, we illustrate a dynamical estimation method for open quantum
systems via quantum filtering.
4.2.1 Quantum Filtering: The Idea
We begin by describing the idea of quantum filtering in a simple setup where the
system is only coupled to a single probe field, which is in the vacuum state. The case
of multiple probe fields is discussed in the following subsection.
The system dynamics considered here is generated by the following special class
of Hudson–Parthasarathy QSDEs (2.2):
1 ∗
dU (t) = −ıH − L L dt + LdA(t)∗ − L ∗ dA(t) U (t), U (0) = I, (4.5)
2
where again H is the system Hamiltonian and L represents the coupling to the
probe field with annihilation process A(t). For simplicity, we further assume that
the internal scattering process is identity: S = I . Also recall that the operator A(t)
satisfies the quantum Itō rule dA(t)dA(t)∗ = dt and all other products between
dA(t), dA(t)∗ , and dt being zero. Now, let us see the time evolution of an observable
X , i.e., jt (X ) := U (t)∗ XU (t). Using the quantum Itō calculus, we end up with
d jt (X ) = jt (LX )dt + jt ([L ∗ , X ])dA(t) + jt ([X, L])dA(t)∗ , (4.6)
where
1 ∗ 1
LX := ı[H, X ] + L ∗ X L − L L X − X L∗ L .
2 2
The quantum Langevin equation (1.18) derived for an oscillator is a special class of
(4.6), though 1.18 is represented in the Stratonovich form with white noise operators.
Note also that the quantum probability space of interest is generated by jt (X ) with
fixed X and P, the latter now being the composite of the initial state of the system
and the vacuum state of the field.
Next, let us discuss the measurement process. Here, we assume that the real
part of the output field operator is measured by a homodyne detector. This means
that we measure the output field operator ym (t) := Y(t) + Y(t)∗ = U (t)∗ (A(t) +
A(t)∗ )U (t). Again by using the quantum Itō rule, the time evolution of ym (t) can be
obtained as
dym (t) = jt (L + L ∗ )dt + dA(t) + dA(t)∗ . (4.7)
Recall from Sect. 2.1.4 that the output observable ym (t) has the self-non-demolition
property:
[ym (s), ym (t)] = 0, ∀s, t. (4.8)
This means that we find a commutative ∗-algebra (the von Neumann algebra) gen-
erated by ym (s), denoted by Ym,t = vN{ym (s), 0 ≤ s ≤ t}. Therefore, there exists a
∗-isomorphism ι defined in Definition 4.4 such that Ym,t is equivalent to a classical
stochastic process representing the observations. Moreover, from Sect. 2.1.4 we also
4.2 Quantum Filtering Theory 129
πt (X)
jt (X)
A(t) Y(t) y m (t)
System Filter
Fig. 4.1 Setup of quantum filtering
have that ym (t) satisfies the quantum non-demolition (QND) condition,
[ym (s), jt (X )] = 0, ∀s ≤ t, (4.9)
for all system observables X .

Our goal is to construct the best estimate of the system observable jt (X ) based
on the observations Ym,t ; see Fig. 4.1 for the setup. In particular, if the criterion is
given by the least mean-square error, P[( jt (X ) − Z )2 ] with Z an element of Ym,t ,
we can have an explicit solution; that is, because now Ym,t is a commutative algebra
and jt (X ) lives in the commutant of Ym,t , from (4.4) we find that the mean-square
error is minimized when Z is given by the quantum conditional expectation:

πt (X ) = argmin P ( jt (X ) − Z )2 = P( jt (X ) | Ym,t ).
Z ∈Ym,t
Consequently, the optimal filter for the system dynamics (4.6) is given by the time
evolution of πt (X ) as follows:
dπt (X ) = πt (LX )dt

+ πt (X L + L ∗ X ) − πt (X )πt (L + L ∗ ) dym (t) − πt (L + L ∗ )dt . (4.10)
This is called the quantum filtering or Belavkin equation. A brief explanation about
how to derive (4.10) is given at the end of this subsection. Note that (4.10) is an
equation on Ym,t , hence it is essentially a classical stochastic differential equation on
the equivalent classical σ-algebra. In particular, as in the classical case, the last term
dν(t) := dym (t) − πt (L + L ∗ )dt is a classical Wiener increment and is called the
innovations process or simply the innovation. As in the classical case, it is zero-mean;
t t
P(ν(t)) = P dym (s) − P πs (L + L ∗ )ds
0 0
t t
=P js (L + L ∗ )ds + dA(s) + dA(s)∗ − P P js (L + L ∗ ) | Ym,s ds
0 0
t t
=P js (L + L ∗ )ds − P js (L + L ∗ )ds = 0,
0 0
where the tower property mentioned in Sect. 4.1.2 is used. Moreover, it is a Ym,t -
martingale. This means that P(ν(t) | Ym,s ) = ν(s) for any t ≥ s ≥ 0. Equivalently,
following the definition of quantum conditional expectation adopted here, P[(ν(t) −
ν(s))Z ] = 0 for any Z ∈ Ym,s . The martingale property follows straightforwardly
from the definition of ν(t) as follows:
t
∗
P [(ν(t) − ν(s))Z ] = P ym (t) − ym (s) + πτ (L + L )dτ Z
s
t
= P (A(t) − A(s)) Z + jτ (L + L ∗ ) − πτ (L + L ∗ ) Z dτ
s
= P [(A(t) − A(s)) Z ] = P [A(t) − A(s)] P(Z ) = 0,
where we have used the fact that

P πτ (L + L ∗ )Z = P jτ (L + L ∗ )Z ,
for every τ ∈ [s, t] and for every Z ∈ Ym,t in going from the second to third line.
Also note that all elements of Ym,s act non-trivially only on h ⊗ F1 (s]), where h
denotes the Hilbert space of the system and, from Chap. 2, F1 (s]) is the factor of the
Fock space of multipliticy one up to time s, while A(t) − A(s) acts non-trivially only
on F1 ([s, t]) (in particular commuting with any element of Ym,s ). The martingale
property implies that the increment ν(t) − ν(s) is independent of any element of Ym,s
for any t > s. That is, roughly speaking, the increment ν(t) − ν(s) is independent
of any function of {ym (τ ), 0 ≤ τ ≤ s}.
There are several approaches to deriving the quantum filtering equation (4.10),
including the reference probability approach [1, 4, 5] and the martingale method [1].
Here, we will take the simplest route using a technique employed in [6]. We begin
with the ansatz that the quantum filter is of the form
dπt (X ) = H1 dt + H2 dym (t), (4.11)
where H1 and H2 are functions of Ym,t . The goal is to derive explicit forms of H1
and H2 . Now, we introduce
t
1 t
h̄(t) = exp h(s)dym (s) − h(s)2 ds ,
0 2 0
where h(t) ∈ Ym,t and thus h̄(t) ∈ Ym,t as well. Using the Itō rule we find that
d h̄(t) = h̄(t)h(t)dym (t). To determine H1 and H2 , we calculate dP(h̄(t) jt (X )) in
two ways as follows. First, from (4.6) we have that

dP h̄(t) jt (X ) = P d h̄(t) jt (X ) + h̄(t)d jt (X ) + d h̄(t)d jt (X )

= P h̄(t)h(t) jt (L + L ∗ )dt · jt (X ) + h̄(t) jt (LX )dt

+ h̄(t)h(t)dA(t) · jt ([X, L])dA(t)∗

= P h̄(t)h(t) jt (L X + L ∗ X ) + h̄(t) jt (LX ) + h̄(t)h(t) jt ([X, L]) dt

= P h̄(t)h(t) jt (X L + L ∗ X ) + h̄(t) jt (LX ) dt,
where P[dA(t)] = P[dA(t)∗ ] = 0 and the quantum Itō rule dA(t)dA(t)∗ = dt are
used. Now, from the tower property of the quantum conditional expectation, the
above equation can be further calculated as
d

P h̄(t) jt (X ) = P P h̄(t)h(t) jt (X L + L ∗ X ) + h̄(t) jt (LX ) Ym,t
dt

= P h̄(t)h(t)P jt (X L + L ∗ X ) Ym,t + h̄(t)P jt (LX ) Ym,t

= P h̄(t)h(t)πt (X L + L ∗ X ) + h̄(t)πt (LX ) .
In the second way, we first apply the tower property and have

dP h̄(t) jt (X ) = dP P h̄(t) jt (X ) Ym,t = dP h̄(t)P( jt (X ) | Ym,t )

= dP h̄(t)πt (X ) = P d h̄(t)πt (X ) + h̄(t)dπt (X ) + d h̄(t)dπt (X )

= P h̄(t)h(t) πt (L + L ∗ )πt (X ) + H2 + h̄(t) H1 + H2 πt (L + L ∗ ) dt.
Comparing the above two equations and noting that h(t) can be chosen arbitrarily in
Ym,t , we end up with
πt (L + L ∗ )πt (X ) + H2 = πt (X L + L ∗ X ), H1 + H2 πt (L + L ∗ ) = πt (LX ),
which from (4.11) leads to (4.10) by some straightforward rearrangement and

regrouping of terms.
4.2.2 Quantum Filter: Multiple-Input Multiple-Output Case
Now, we consider a general quantum system coupled to m bosonic fields A =

(A1 , . . . , Am ) in the vacuum state, with Hamiltonian H and coupling operators
L = (L 1 , . . . , L m ) , and scattering matrix S ∈ Cm×m . (Note that this setup is not
the most general, as explained later.) When the system is driven by coherent fields,
then, as elaborated in Chap. 2, the Hudson–Parthasarathy QSDE is given by

dU (t) = Tr (S − I ) d(t) + dA(t)∗ L(t) − L(t)∗ SdA(t)
1
− ı H (t) + L(t)∗ L(t) dt U (t), (4.12)
2
with initial condition U (0) = I . Here we have
L(t) = L + S f (t), H (t) = H + {L ∗ S f (t)},
where f (t) is the time-dependent amplitude of the coherent input fields, taking a
value in Cm (see Sect. 2.7.5). Recall that the time evolution of a bounded operator X
in the Heisenberg picture, i.e., X (t) = jt (X ) = U (t)∗ XU (t), is given by
1 1
d jt (X ) = jt ı[H (t), X ] + L(t)∗ X L(t) − L(t)∗ L(t)X − X L(t)∗ L(t) dt
2 2
∗ ∗ ∗
+ dA(t) jt (S [X, L(t)]) + jt ([L(t) , X ]S)dA(t),
with initial condition j0 (X ) = X [5]. Also, the field output defined by Y(t) =
U (t)∗ A(t)U (t) obeys the following equation:
dY(t) = jt (L(t))dt + SdA(t). (4.13)
Next, we describe the measurement process. In the previous section, we consider

the measurement of the real part of Y(t), which leads to the measurement output
equation (4.7), but there are in fact infinitely many choices. For instance, we could
also measure the imaginary part of Y(t). Those output fields can be represented in
the quadrature form as

y(t) = 2 {Y1 (t)}, {Y1 (t)}, . . . , {Ym (t)}, {Ym (t)} ,
as defined in Sect. 2.3.2. We consider a measurement ym (t) of the form ym (t) =

Fm y(t), with Fm ∈ R p×2m . We require Fm to be full row rank and such that
[ym (t), ym (s) ] = ym (t)ym (s) − (ym (s)ym (t) ) = 0, ∀s, t ≥ 0.
Let Ym,t = vN{ym (s), 0 ≤ s ≤ t} denote the von Neumann algebra generated by
ym, j (s), j = 1, 2, . . . , p and 0 ≤ s ≤ t. Here ym, j is the j-th element of the vector ym .
As we saw above, the assumption on Fm ensures that Ym,t = vN{ym (s), 0 ≤ s ≤ t}
is a commutative von Neumann algebra. Moreover, since the output y(t) sat-
isfies [y(s), jt (X )] = 0 for all 0 ≤ s ≤ t (recall Sect. 2.1.4), we also have that
[ym (s), jt (X )] = Fm [y(s), jt (X )] = 0 for all 0 ≤ s ≤ t and for any bounded sys-
tem operator X . Therefore, ym (t) is isomorphic to a classical stochastic process by
an infinite dimensional version of Theorem 4.1, and the quantum conditional expec-
tation πt (x) = P( jt (X ) | Ym,t ) is well defined (recall that P denotes the quantum
expectation with respect to the initial state of the system and field). To simplify the
exposition in what will follow, we impose an additional structure on Fm . Namely,
we require that p ≤ m and Fm can be decomposed as,
Fm = M[ I p 0 p×(2m− p) ][ V V # ]m , (4.14)
with M some invertible real p × p matrix, V some complex m × m unitary matrix,

and m is as defined in Sect. 2.3.2. What these conditions mean is that effectively,
we can write
ym (t) = M[ I p 0 p×(2m− p) ](Ỹ(t) + Ỹ(t)# )

= M[ I p 0 p×(2m− p) ]U (t)∗ (Ã(t) + Ã(t)# )U (t)
for an input field annihilation process Ã(t) = V A(t) and the corresponding output
field annihilation process Ỹ(t) = V Y(t). Recall that A(t) has been taken to be in the
vacuum state. Since V is unitary, Ã(t) is also in the vacuum state and satisfies the
same vacuum Itō products as A(t). Moreover, if we define (t)˜ = V # (t)V then
˜
(t) satisfies the same vacuum Itō products as (t). The imposed structure sim-
plifies the derivation of the associated quantum filtering equation, using the results
of [5] extended to the multiple-input multiple-output case. However, we emphasize
that the additional conditions on Fm can be relaxed and more general linear measure-
ments performed on y(t) can be analyzed, but requiring more laborious algebraic
manipulations. For details, see [7].
Let
ỹm (t) = M −1 ym (t) = [ I p 0 p×(2m− p) ](Ỹ(t) + Ỹ(t)# )
be another measurement process. Since ỹm is linearly related to ym and M is invertible,

ỹm and ym generate the same von Neumann algebra, we have that
πt (X ) = P( jt (X ) | vN{ym (s), 0 ≤ s ≤ t}) = P( jt (X ) | vN{ ỹm (s), 0 ≤ s ≤ t}).
Then, define L̃ = V L and S̃ = V SV ∗ , and the operators
L̃(t) = L̃ + S̃ f (t), H̃ (t) = H + { L̃ ∗ S̃ f (t)}. (4.15)
Note that H̃ (t) = H (t). In terms of these operators and the processes Ã(t), Ã(t)∗
˜
and (t) defined above, the QSDE (4.12) can be rewritten as:

˜
dU (t) = Tr ( S̃ − I ) d (t) + d Ã(t)∗ L̃(t) − L̃(t)∗ S̃d Ã(t)
1
− ı H̃ (t) + L̃(t)∗ L̃(t) dt U (t). (4.16)
2
It then follows, by a straightforward extension of the results of [5] to multiple inputs

and multiple outputs, that the quantum filtering equation is given by
dπt (X )
1
m
= πt ı[ H̃ (t), X ] + L̃ j (t)∗ [X, L̃ j (t)] + [ L̃ j (t)∗ , X ] L̃ j (t) dt
j=1
2
p

+ πt (X L̃ j (t) + L̃ j (t)∗ X ) − πt ( L̃ j (t) + L̃ j (t)∗ )πt (X ) d ν̃ j (t), (4.17)
j=1
t
where ν̃ j (t) = ỹm, j (t) − 0 πs ( L̃ j (s) + L̃ j (s)∗ )ds is the j-th component of the inno-
t
vations process ν̃(t) = ỹm (t) − [ I p 0 ] 0 πs ( L̃(s) + L̃(s)∗ )ds. The process ν̃(t) is
a Ym,t -martingale Wiener process [5]. This can be shown by an analogous argument
to the proof of the martingale property of the innovations process for a system cou-
pled to a single-input field given in Sect. 4.2.1. In particular, ν̃(t) is zero-mean and
is independent of any function of Ym,t .
4.2.3 Stochastic Master Equation
Let us return to the simple setup where the system is coupled to a single probe
field and also the internal system scattering process is identity (i.e., S = I ). First,
we remark that the quantum filter (4.10) describes the time evolution of an observ-
able, which is projected onto the commutative algebra; that is, (4.10) is the Heisen-
berg picture representation of the estimate. Here, we switch to the Schrödinger
picture to see the time evolution of the conditional state as an estimate of the
state. To do this, let us recall that, by defining ι to be a ∗-isomorphism from
Ym,t to the σ-algebra corresponding to the classical measurement record, we have
ι(πt (X )) = ι(P(U (t)∗ XU (t) | Ym,t )); this means that ι(πt (·)) is a nonnegative, lin-
ear, and identity-preserving functional of X . Thus, ι(πt (·)) is a state possessing the
properties of Definition 4.2 almost surely. Hence, we can connect ι(πt (X )) with a
stochastic operator ρ(t) as ι(πt (X )) = Tr(X ρ(t)). As described in (4.1), then ρ(t)
belongs to the following convex set:
S := {ρ : ρ = ρ∗ ≥ 0, Trρ = 1}. (4.18)
The state ρ(t) is called a conditional state, because it is essentially the quan-
tum analogue of a classical conditional probability density. Applying the relation
ι(πt (X )) = Tr(X ρ(t)) to (4.10), we obtain the time evolution of ρ(t) as follows:

dρ = L∗ ρdt + Lρ + ρL ∗ − Tr(Lρ + ρL ∗ )ρ dym (t) − Tr(Lρ + ρL ∗ )dt ,
1 ∗ 1
L∗ ρ := −ı[H, ρ] + LρL ∗ − L Lρ − ρL ∗ L , (4.19)
2 2
where we identify ι(ym (t)) with ym (t) for simplicity. This equation is called
the stochastic master equation (SME). It is the quantum version of the classical
Kushner–Stratonovich equation, which is the stochastic partial differential equation
describing the time evolution of a conditional probability density of a stochastic
system. Note again that (4.19) is a (matrix-valued or operator-valued) classical
stochastic differential equation driven by the zero-mean Wiener process dν(t) =
dym (t) − Tr(Lρ(t) + ρ(t)L ∗ )dt. We also point out that this ideal SME (4.19) pre-
serves a pure state; that is, Tr(ρ(t)2 ) = Tr(ρ(0)2 ) holds if ρ(0) is a pure state.
Now, by averaging the SME and exploiting the fact that the innovation is a mar-
tingale, we obtain the dynamics of ρ̄ := E(ρ):
d ρ̄ 1 1
= L∗ ρ̄dt = −ı[H, ρ̄] + L ρ̄L ∗ − L ∗ L ρ̄ − ρ̄L ∗ L . (4.20)
dt 2 2
This is called the quantum master equation (ME), which describes the uncondi-
tional time evolution of the quantum system interacting with a coherent field via
the coupling operator L. Note that (4.20) can be derived directly from the QSDE
(4.6) as follows. First, the expectation of jt (X ), i.e., P[ jt (X )], yields dP[ jt (X )]/dt =
P[ jt (LX )], due to P[dA(t)] = P[dA(t)∗ ] = 0. Second, because P[ jt (X )] =
P[U (t)∗ XU (t)] is a nonnegative, linear, and identity-preserving functional of X ,
it can be expressed as P[ jt (X )] = Tr(X ρ̄(t)). Combining these two relations, we
end up with (4.20).
Finally, it is straightforward to generalize the SME and ME to the multi-input
setup discussed in Sect. 4.2.2. That is, if the system couples m bosonic input fields
and the first p output fields are measured by homodyne detectors, then the time
evolution of the conditional state ρ(t) is given by the following SME:
m
1 1
dρ = − ı[ H̃ (t), ρ] + L̃ j (t)ρ L̃ j (t)∗ − L̃ j (t)∗ L̃ j (t)ρ − ρ L̃ j (t)∗ L̃ j (t) dt
2 2
j=1
p

+ L̃ j (t)ρ + ρ L̃ j (t)∗ − Tr( L̃ j (t)ρ + ρ L̃ j (t)∗ )ρ
j=1

× dym, j (t) − Tr( L̃ j (t)ρ + ρ L̃ j (t)∗ )dt ,
where H̃ (t) and L̃(t) are defined in (4.15), and ym, j (t) represents the measurement
record corresponding to the j-th output field. Also, the ensemble average of the
measurement results yields the following ME:
d ρ̄ m
1 1
= −ı[ H̃ (t), ρ̄] + L̃ j (t)ρ̄ L̃ j (t)∗ − L̃ j (t)∗ L̃ j (t)ρ̄ − ρ̄ L̃ j (t)∗ L̃ j (t) .
dt j=1
2 2
4.2.4 QND Interaction and the Projection Postulate
As an important application of the quantum filter, here we study a special case

where the coupling operator L is Hermitian, i.e., L = L ∗ . For instance, the Faraday
interaction for
√ an atomic ensemble described in Sect. 1.5.5 satisfies this assumption,
where L = M Jz with M the coupling strength. We further assume that the system
Hamiltonian is zero, H = 0; this is because we are working in a rotating frame with
respect to H satisfying [H, L] = 0. The point of this setting is that, from (4.6), the
time evolution of the observable L is given by
1 1
d jt (L) = jt L ∗ L L − L L ∗ L − L ∗ L L dt + jt ([L ∗ , L])dA + jt ([L , L])dA∗ = 0.
2 2
That is, the system observable L does not change in time, i.e., jt (L) = j0 (L) = L;
in other words, this special interaction does not disturb the physical property of L,
hence it is called a quantum non-demolition (QND) interaction. On the other hand,
when the amplitude quadrature of Y(t) is measured by a homodyne detector, then
the measurement output equation is given by
dym = jt (L + L ∗ )dt + dA + dA∗ = 2 jt (L)dt + dA + dA∗ = 2Ldt + dA + dA∗ .
Therefore, by measuring this quadrature component of the output probe field, we can
extract the information about L without disturbing this observable.
The statistics of the information we obtain can be explicitly evaluated by exam-
ining the corresponding filtering equation,
1 1
dπt (X ) = πt L X L − L 2 X − X L 2 dt + πt (X L + L X ) − 2πt (X )πt (L) dν.
2 2
It is then immediate to see that

dπt (L) = 2 πt (L 2 ) − πt (L)2 dν = 2(L − Lc )2 c dν, (4.21)
where ·c := πt (·). That is, the conditional expectation is a pure diffusion process
weighted by the conditional variance L 2 c = (L − Lc )2 c , thus P[dπt (L)] = 0
and so P[πt (L)] = P[π0 (L)] = P(L), which is consistent with the claim above that
the QND interaction does not disturb L. Moreover, this conditional variance obeys

dL 2 c = d πt (L 2 ) − πt (L)2 ]

= 2 πt (L 3 ) − πt (L 2 )πt (L) dν − 2πt (L)dπt (L) − [dπt (L)]2

= 2 πt (L 3 ) − 3πt (L 2 )πt (L) + 2πt (L)3 dν − 4L 2 2c dt
= 2L 3 c dν − 4L 2 2c dt, (4.22)
where we have used the Itō rule dν 2 = dt. This leads to the following important
relation:
d
P[L 2 c ] = −4P[L 2 2c ].
dt
In general, it is known from classical stochastic stability theory that, if dE(x)/dt ≤ 0

for a nonnegative random variable x(t) contained in a finite set, then limt→∞ dE(x)/
dt = 0 holds almost surely. Hence, in our case, we have P[L 2 2c ] → 0 almost
surely as t → ∞. This means that the conditional state converges to one of the
eigenvectors of L, which is referred to as quantum state reduction. In particu-
lar for the finite-dimensional case (i.e., the case where L is a Hermitian matrix),
very importantly,
the probability of this event can be exactly calculated as follows.
Let L = k k |k k | be the spectral decomposition of L, with k the kth eigen-
value and |k the corresponding eigenvector. Then, as in the above calculation,
we have dπt (|k k |) = 2πt (|k k |)[k − πt (L)]dν, leading to P[π∞ (|k k |)] =
P[π0 (|k k |)] = k |ρ(0)|k . Now, we know from the above result that the condi-
tional state converges into one of the eigenvectors of L, and thus, the left-hand side
of this equation is exactly the probability for the state ρ(0) to converge to |k k |.
As a result, we have
P({ρ(0) → |k k |}) = k |ρ(0)|k .
This is the renowned Born’s projection postulate: continuous measurement with

QND interaction realizes the Born’s projection postulate.
4.3 Quantum Kalman Filter for Gaussian Linear

Quantum Systems
Following the theme of this monograph, we will now focus on linear dynamical quan-
tum systems. First, a simple example is provided to show the idea behind quantum
Kalman filtering and how it is used. Then, based on that intuition, a general Kalman
filter dynamics is given for a class of linear measurements applied to the system.
4.3.1 Example in Quantum Optics
The quantum Kalman filter provides the best estimate of observables of a linear
Gaussian quantum system (i.e., a linear quantum system where all the noise and
the initial state is Gaussian). A very nice structure of the Kalman filter is that it
enables an exact computation of the estimate, while in the general case such an
exact computation is not possible because infinitely many conditional moments need
to be computed. In fact, thanks to the linear and Gaussian nature of the system,
the conditional mean and variance are enough to compute the quantum Kalman
filtering equation. Here, we focus on a special example taken from quantum optics
and illustrate how the Kalman filter can be used.
The example we consider here is a degenerate parametric amplifier (DPA) dis-
cussed in Sect. 1.5.3. This system is characterized by Hamiltonian H = ı (a ∗2 −
√
a 2 )/2 (in a rotating frame) and coupling operator L = γa. Here, we ignore the
optical loss. The QSDE of this system is given by
γ √ √
d jt (a) = − jt (a)dt + jt (a ∗ )dt − γdA(t), dY(t) = γ jt (a)dt + dA(t),
2 2
which is also given by (1.37) with κ = 0 in the white noise form. Here, we consider
the measurement of the amplitude quadrature of the output field Y(t) by a homodyne
detector, which generates the measurement output ym (t) satisfying
√
dym (t) = γ jt (a + a ∗ )dt + dA(t) + dA(t)∗ .
As described before, this is identified as a classical stochastic signal ι(ym (t)) and is
processed by the quantum filter, which yields the best estimate of both the quadra-
tures jt (q) = jt (a + a ∗ ) and jt ( p) = jt (−ı(a − a ∗ )). The entire setup is depicted
in Fig. 4.2. Note that in the scenario of quantum optics, we usually do not intend to
estimate the internal cavity modes, but rather we are interested in the external output
field and its squeezing level. But this simple example can be used to model several
important systems such as an opto-mechanical oscillator, in which case estimating
those internal degrees of freedom (such as the position and momentum) is important.
Also from this example, we will see the necessity of feedback control to produce an
unconditional squeezed state, which will be discussed in Sect. 6.6.2.
The filtering equation of jt (q) and jt ( p) can be derived directly by substituting
X = q and X = p to (4.10). Let us begin by deriving the equation for πt (q). First, it
√
can be seen immediately that πt (Lq) = ( − γ)πt (q)/2 and πt (L + L ∗ ) = γπt (q).
The remaining part (coefficient of the innovation part) can be calculated as follows;
Pumping field
jt (q), jt (p) A(t) πt (q), πt (p)
γ
Y(t)
ym (t) Filter
Fig. 4.2 Schematic of the quantum filter for an DPA. The homodyne detector placed at the output of
the DPA measures the amplitude quadrature of the output field Y (t) and generates the measurement
output ym (t), which is used for producing the estimate πt (x) = (πt (q), πt ( p))
4.3 Quantum Kalman Filter for Gaussian Linear Quantum Systems 139
√
γ √
πt (q L + L ∗ q) − πt (q)πt (L + L ∗ ) = πt (2q 2 + ı[q, p]) − πt (q) · γπt (q)
2
√ √
= γπt (q 2 ) − γ,
where q = q − πt (q). Hence the filtering equation for q is
−γ √ √
dπt (q) = πt (q)dt + γ πt (q 2 ) − 1 (dym − γπt (q)dt). (4.23)
2
This implies that we need to have an equation to compute πt (q 2 ). To obtain this
equation, here we show only the critical part: the coefficient of the innovation part
of the filtering equation for q 2 can be calculated to be
√
πt (q 2 L + L ∗ q 2 ) − πt (q 2 )πt (L + L ∗ ) = κ πt (q 3 ) − πt (q 2 )πt (q) − 2πt (q) ,
suggesting that the third-order moment may be necessary. However, with the assump-
tion that the initial state is a quantum Gaussian state such as a coherent or squeezed
state, the third moment is actually zero, i.e., πt (q 3 ) = 0 (further discussions in
relation to this will be given in Sect. 4.3.2), which leads to πt (q 3 ) − 3πt (q 2 )πt (q) +
2πt (q)3 = 0. Using this relation, we end up with the equation of πt (q 2 ) as follows:
+γ
dπt (q 2 ) = πt (q 2 )dt − γπt (q 2 )2 dt. (4.24)
2
That is, this is an ordinary differential equation in the variable πt (q 2 ), hence it can
be solved numerically if not explicitly. The filtering equation of p can also be derived
in the same way:
+γ √ qp + pq √
dπt ( p) = − πt ( p)dt + γπt (dym − γπt (q)dt).
2 2
(4.25)
Also the conditional variance of p obeys
+γ
dπt (p 2 ) = − πt (p 2 )dt + γdt. (4.26)
2
The above set of Eqs. (4.23)–(4.26), can be summarized as follows:
dπt (x) = Aπt (x)dt + (P(t) − I )C (dym (t) − Cπt (x)dt), (4.27)
d P(t)
= A P(t) + P(t)A + γ I − (P(t) − I )C C(P(t) − I ), (4.28)
dt
where

πt (q) ( − γ)/2 0 √
πt (x) = , A= , C = [ γ 0]
πt ( p) 0 −( + γ)/2
and P(t) is the estimation error symmetrized covariance matrix defined by

πt (q 2 ) πt (qp + pq)/2
P(t) = .
πt (qp + pq)/2 πt (p 2 )
The set of Eqs. (4.27) and (4.28) is called the quantum Kalman filtering equations
(for the DPA). A notable feature of these equations is that (4.28) is a deterministic
differential equation, implying that the estimation error does not depend on the obser-
vation record; (4.28) is called the Ricatti differential equation, and this deterministic
property holds more generally, as will be outlined in the next subsection.
Now, let us examine our filtering equation under the assumption that the ini-
tial state is set to a pure coherent state with mean x = (2, 2) and covariance
matrix xx + (xx ) /2 = I with x = x − x. We set π0 (x) = x and
P(0) = I in the filtering equation. The time evolution of the filtering equation (4.27)
and (4.28) is illustrated in Fig. 4.3. From this, we observe several interesting proper-
ties of the estimate for the DPA system, which are summarized as follows.
• πt (q) changes probabilistically, while πt ( p) obeys a deterministic time evolution
and converges to zero. This is because the off-diagonal elements of P(t) is always
zero due to P(0) = I . This fact implies that we do not obtain any information
about p from the observation; in fact, the dynamics of πt ( p) is now dπt ( p)/dt =
(− − γ)πt ( p)/2, and this is exactly the same as the unconditional time evolution
of p, i.e., P( p), which is generated from the master equation (4.20).
π t( q ), π t( Δ q 2 ) π t( p ), π t( Δ p 2 )
3
−1
−2
0 2 4 6 8 10 0 2 4 6 8 10
time [s / κ ] time [s / κ ]
Fig. 4.3 The time evolution of πt (q) (left) and πt ( p) (right). The estimation errors πt (q 2 ) and
πt (p 2 ) are depicted in blue along with the trajectories of πt (q) and πt ( p), respectively. The
squeezing strength is chosen as = −0.6γ, which guarantees the stability of the system (i.e., A is
Hurwitz) (color figure online)
πt (p) πt (p)
(a) (b)
πt (q) πt (q)
Fig. 4.4 Phase-space picture of the a conditional and b unconditional states of the DPA. The
conditional state fluctuates along the πt (q) axis. The ensemble average of this fluctuation yields the
unconditional state, which is squeezed but not pure
• In this case, the steady-state solution of (4.28) can be explicitly obtained as
+γ γ
π∞ (q 2 ) = , π∞ (p 2 ) = .
γ +γ
Hence π∞ (q 2 )π∞ (p 2 ) = 1 holds. Thus, if = 0, the conditional state is a pure
squeezed state; in fact as discussed in Sect. 4.2.3, the equivalent SME preserves the
purity of the conditional state. In particular, in the simulation, we chose = −0.6γ,
in which case π∞ (q 2 ) = 2/5 and π∞ (q 2 ) = 5/2, hence it is a q-squeezed state.
From the second observation above, the conditional state (when t 1) can be
illustrated in the (πt (q), πt ( p)) phase space. As shown in Fig. 4.4a, it is a pure
squeezed state fluctuating along the πt (q) axis. However, due to this probabilistic
nature, the ensemble average of this conditional squeezed state is not a pure state
anymore, as illustrated in Fig. 4.4b. In fact, the unconditional variance of q at the
steady state can be calculated as P(q 2 ) = π∞ (q 2 ) = γ/(γ − ); hence together
with the fact P(p 2 ) = π∞ (p 2 ) = γ/( + γ) we have that P(q 2 )P(p 2 ) =
γ 2 /(γ 2 − 2 ) > 1; hence it is not a pure squeezed state. Therefore, to obtain a pure
squeezed state in a deterministic way, we need to compensate the fluctuation along
the πt (q) axis, and this can be carried out by feedback. This topic will be discussed
in Sect. 6.6.2.
4.3.2 Quantum Filtering: Multiple-Input

and Multiple-Output Case
We now turn to the case of linear quantum systems with multiple inputs and outputs,
with coherent input fields, with dynamics given by
d x(t) = Ax(t)dt + B( f (t)dt + dw(t)); x(0) = x.

dy(t) = C x(t)dt + D( f (t)dt + dw(t)), (4.29)
where the matrices (A, B, C, D) are given by (2.6), (2.7), (2.8), and (2.7), respec-
tively, and f (t) is a R2m -valued vector of the phase and amplitude quadratures of
the coherent input fields.
We follow the setting of Sect. 4.2.1 and restrict the treatment as before to the case
where the measurement process ym (t) is of the form
dym (t) = Fm dy(t) = Cm x(t)dt + Dm ( f (t)dt + dw(t)), (4.30)
where Cm = Fm C, Dm = Fm D, and Fm is as given by (4.14). Now, with x =

(q1 , p1 , q2 , p2 , . . . , qn , pn ) as before, let
πt (x) = (πt (q1 ), πt ( p1 ), πt (q2 ), πt ( p2 ), . . . , πt (qn ), πt ( pn )) .
We now also restrict ourselves to the case where the state of the oscillators to be
jointly Gaussian, while the input fields, being coherent states, are already Gaussian.
This will be required in the following to obtain the general form of the quantum
Kalman filtering equation for linear Gaussian quantum systems with multiple inputs
and outputs. The linear evolution of x(t) and the Gaussian state of the system imply
that x(t) is a Gaussian observable in sense that,
t
P eıλ x(t) = P eıλ ((t)x+ 0 (t−τ )B( f (τ )dτ +dw(τ )))
t t
= eıλ ((t)x+ (t−τ )B f (τ )dτ )

(t−τ )Bdw(τ )
0 eıλ 0
t
ıλ ((t)x+ 0 (t−τ )B f (τ )dτ )−(1/2)λ R(t)λ
=e ,
for all λ ∈ Rn , where (t) = e At and

t

R(t) = (t) (x − x)(x − x) + (−τ )B B (−τ ) dτ (t) .

0
In the second line above, the first quantum expectation · on the right is with
respect to the oscillators, while the second quantum expectation is with respect to the
fields. For any integer k > 0, let 0 ≤ s1 < s2 < · · · < sk . By a similar calculation,
ym (s1 ), ym (s2 ), . . . , ym (sk ) are jointly Gaussian observables in the sense that,

P eıTr(χ ym (s1 ,...,sk )) = eı (vec(χ) μ(s1 ,...,sk ))−( /2)vec(χ) Q(s1 ,...,sk )vec(χ) ,
1
for any χ ∈ R p×k , where ym (s1 , . . . , sk ) = [ ym (s1 ) . . . ym (sk ) ], and vec(χ) ∈ R p×k
is a column vector formed by stacking the columns of χ from left to right,
⎡ s1 τ ⎤
Dm f (τ ) + Cm 0 (τ − τ )B f (τ )dτ dτ
⎢ s2 D f (τ ) + C τ (τ − τ )B f (τ )dτ dτ ⎥
0
⎢ m m 0 ⎥
μ(s1 , s2 , . . . , sk ) = ⎢ 0 .. ⎥,
⎣ ⎦
sk τ .
0 Dm f (τ ) + Cm 0 (τ − τ )B f (τ )dτ dτ
and
Q(s1 , s2 , . . . , sk )
si sj
= Cm (τ )dτ (x − x)(x − x) (τ ) dτ Cm
0 0
∞ si
+ Dm 1[0,si ] (τ ) + Cm 1[0,si ] (τ )(τ − τ )Bdτ
0 τ

sj
× Dm 1[0,s j ] (τ ) + Cm 1[0,s j ] (τ )(τ − τ )Bdτ dτ .
τ
i, j=1,2,...,k
Extrapolating from the above, it is easy to see that the observables x j (t) ( j =
1, 2, . . . , n) and ym (s1 ), …, ym (sk ) are jointly Gaussian for any integer k > 1 and
any 0 ≤ s1 < s2 < · · · < sk ≤ t, in the sense that any linear combination of these
observables is a Gaussian observable (extending the definition above in the obvious
way).
We know that λ x(t) commutes with any element of the von Neuman algebra Ym,t
for any λ ∈ Rn , and the conditional expectation πt (λ x) = λ πt (x) is an element of
this algebra. Thus, by the Spectral Theorem, λ πt (x) and {ym (s); 0 ≤ s ≤ t} can be
viewed as classical random variables that can be realized on the same classical prob-
ability space. Even more than that, since λ x(t) and {ym (s); 0 ≤ s ≤ t} are jointly
Gaussian observables, as discussed in the preceding paragraph, it follows that the
conditional expectation λ πt (x) must be (equivalent to) a classical Gaussian ran-
dom variable for any λ ∈ Rn . In particular, {πt (x j ); j = 1, 2, . . . , n} are essentially
jointly Gaussian random variables.
Now, using the definitions of H̃ (t) and L̃(t) to explicitly calculate all commutators
on the right-hand side of (4.17), and taking the initial state of the system to be
Gaussian, so that the conditional expectations are jointly Gaussian random variables
and their third-order covariances vanish, we have
dπt (x) = Aπt (x)dt + B f (t)dt

+ [P(t)C Fm + B](Dm Dm
−1
) [dym (t) − Fm Cπt (x)dt]
= Aπt (x)dt + B f (t)dt + [P(t)C Fm + B](Dm Dm
−1
) dν(t), (4.31)

with initial condition π0 (x) =
t x, where = D Fm . Also ν(t) is the innovations
process ν(t) = ym (t) − Fm 0 (Cπs (x) + Du(s))ds, and P(t) is the symmetrized
covariance matrix,
1
P(t) = πt (x − πt (x))(x − πt (x)) + ((x − πt (x))(x − πt (x)) ) ,
2
satisfying the deterministic matrix Riccati differential equation
Ṗ(t) = A P(t) + P(t)A + B B

− (P(t)C Fm + B)(Dm Dm
−1
) (P(t)C Fm + B) , (4.32)
with initial condition

1
P(0) = xx + (xx ) , (4.33)
2
with x = x − x.
4.4 Robust Linear Quantum Observers
A central problem of the quantum filter is that, because it is a model-based estimator,

the exact system parameters must be known for the optimal estimation to be carried
out (in the Kalman filtering case, this means that all the matrices (A, B, C, D) must
be exactly known). In fact, a nominal filter with the wrong parameters is fragile in
the sense that the estimation error can be large or even diverge. Of course, this is
not a problem appearing only in the quantum case, but rather it is a long-standing
problem that has been extensively investigated in classical systems and control theory.
Actually, in the literature for classical systems, we find numerous treatments of robust
filtering, which generates a guaranteed estimate even under system uncertainties. This
section is devoted to discussing a quantum version of these classical robust filters [8].
4.4.1 Guaranteed-Error Robust Observer
For simplicity, let us deal with a general single-input and single-output linear quan-
tum system with single-mode variable x(t) = (q(t), p(t)) = ( jt (q), jt ( p)) . The
system couples to a vacuum field, represented by the coupling operator L = K x with
x = (q, p) . Here, we consider the case where the system Hamiltonian H contains
uncertainty of the following form:
1
H= x (R + R(t))x − x JFu(t), R(t)2 ≤ g I.
2
u(t) is a control input, which can be a function of the measurement result (i.e., a
feedback control input). The real symmetric matrix R(t) represents a time-varying
4.4 Robust Linear Quantum Observers 145
uncertainty bounded by the known constant g > 0. The other matrices, R, F, and K ,
are known. By defining A(t) := JR(t), the dynamics of x(t) is given by
d x(t) = (A + A)x(t)dt + Fu(t)dt + ıJ[K dA∗ (t) − K ∗ dA(t)]. (4.34)
Also here, we consider measuring the amplitude quadrature by homodyne detection,

generating the measurement output
dym (t) = (K + K ∗ )x(t)dt + dA(t) + dA(t)∗ . (4.35)
Note that we can consider a more general case where K also contains some uncer-
tainty; see [8].
Now, motivated by the structure of the Kalman filter (4.31), we aim to design a
classical linear observer of the form
d x (t) = A x (t)dt + Fu(t)dt + B dym (t), (4.36)
where A and B are a matrix and vector to be determined so that the variance
of the estimation error is guaranteed to be within a certain bound. The vector of
classical random variables, x (t) = (q (t), p (t)) ∈ R2 , represents an estimate of
x(t), which is not necessarily an optimal estimate of x(t). (Hence, we refer to a
system with dynamics (4.36) as an “observer” rather than a “filter”.) Furthermore,
the control input u(t) is fixed to a linear function of the observer variable, i.e.,
u(t) = x (t), where is a fixed real column vector with dimension 2. We now
offer two important remarks about the setting considered above. First, note that we
are now considering a hybrid system composed of the quantum system (4.34) and
the classical system (4.36). This could lead to some difficulty in making sense of
the system of differential equations that describe the two systems simultaneously,
especially when they are interacting and there is a measurement involved, the latter
causing the state of the quantum part to collapse. This kind of hybrid system can
be given a consistent interpretation by embedding the classical linear dynamics as a
commutative subdynamics of an appropriate linear quantum system. Details of this
technique of embedding classical dynamics into a quantum system, along with some
examples, can be found in [9, Appendix, pp. 2542–2543] and, for the linear case,
[10, Sect. 3], [11–13]. The second remark is regarding the feedback control input u(t).
It is not an essential point of the theory that will be described; the robust observer
method introduced below holds even if = 0.
Now, an explicit form of (A , B ) that enjoys a guaranteed estimation error bound
is provided in the following theorem.
Theorem 4.2 Suppose there exist positive scalars δ1 , δ2 , and such that the fol-
lowing two coupled Riccati equations have positive definite solutions P1 > 0 and
P2 > 0:
1
(A + F )P1 + P1 (A + F ) + D + P12 + ( g + δ1 )I = 0, (4.37)
AP2 + P2 A + D − (P2 F + J (K ) )(FP2 + (K )J)
− P2 (F P1−1 + P1−1 F )P2 + ( g + δ2 )I = 0, (4.38)
where the matrices A, D and the vector F are defined by
A = A + (D + g I )P1−1 , D = J(K ∗ K )J , F = K + K ∗ + (K )JP1−1 .
Then, the observer
d x (t) = (A − P2 F P1−1 )x (t)dt + Fu(t)dt

+ P2 F + J (K ) (dym (t) − Fx (t)dt) (4.39)
generates the estimate x (t) = (q (t), p (t)) that satisfies
lim (q(t) − q (t))2 + ( p(t) − p (t))2 ≤ Tr(P2 ), (4.40)

t→∞
for all admissible uncertainties contained in R(t).
Proof We consider the augmented variable z̄(t) = (x(t), x(t) − x (t)) , where x(t)
and x (t) satisfy (4.34) and (4.36), respectively. Then, z̄(t) obeys the following linear
QSDE:
d z̄(t) = ( Ā + Ā(t))z̄(t)dt + b̄dA(t)∗ + b̄∗ dA(t),
where

A + F −F ıJK
Ā = , b̄ = ,
A − A − B (K + K ∗ ) A ıJK − B

A(t) 0 J
Ā(t) = = R(t)[I 0] =: M̄R(t) N̄ .
A(t) 0 J
Now, the symmetrized covariance matrix of z̄ is given by V̄nm = z̄ n z̄ m + z̄ m z̄ n /2,

(n, m = 1, . . . , 4). This satisfies the following generalized uncertain relation:

ı JJ
z̄(t)z̄(t) = V̄ (t) + J̄ ≥ 0, J̄ = ,
2 JJ
and the time evolution of V̄ (t) is calculated as
d
V̄ (t) = ( Ā + Ā(t))z̄(t)z̄(t) + z̄(t)z̄(t) ( Ā + Ā(t)) + b̄∗ b̄
dt ı ı
= ( Ā + Ā(t)) V̄ (t) + J̄ + V̄ (t) + J̄ ( Ā + Ā(t)) + b̄∗ b̄
2 2
= ( Ā + Ā(t))V̄ (t) + V̄ (t)( Ā + Ā(t)) + D̄, (4.41)
where D̄ is given by

DD 0 [B (K )J]
D̄ = − .
DD B (K )J B (K )J + [B (K )J] − B B
Our goal is to design A and B such that the matrix inequality
∃ X̄ > 0 s.t. ( Ā + Ā(t)) X̄ + X̄ ( Ā + Ā(t)) + D̄ < 0 (4.42)
is satisfied for all admissible uncertainties: then, it is immediate to see that (4.41)
yields limt→∞ V̄ (t) ≤ X̄ . In order to obtain a sufficient condition for (4.42) that does
not involve Ā(t), we use the following relation:
1 √ 1 √
√ X̄ N̄ − M̄R(t) √ X̄ N̄ − M̄R(t) ≥ 0,
where > 0 is a free parameter. Then, combining the above relation with the assump-
tion on the uncertainty, R(t)2 ≤ g I , we obtain
Ā(t) X̄ + X̄ Ā(t) = ( M̄R(t))( N̄ X̄ ) + ( N̄ X̄ ) ( M̄R(t))

1
≤ X̄ N̄ N̄ X̄ + M̄(R(t))2 M̄
1
≤ X̄ N̄ N̄ X̄ + g M̄ M̄ .
Therefore, the condition (4.42) holds for all admissible uncertainties if there exists
a positive definite matrix X̄ > 0 such that the following Riccati inequality holds:
1
¯ := Ā X̄ + X̄ Ā + D̄ + g M̄ M̄ + X̄ N̄ N̄ X̄ < 0.
Especially here we aim to find a solution of the form X̄ = diag(P1 , P2 ) with P1 and
P2 denoting 2 × 2 positive definite matrices. Then, partitioning the 4 × 4 matrix ¯
into ¯ = ( i j ) with 2 × 2 matrices i j , we obtain
11 = (A + F )P1 + P1 (A + F ) + D + g I + P12 / ,

21 = (A − A − B K − B K ∗ )P1 + D + g I − B (K )J − P2 F ,
22 = A P2 + P2 A + D + g I − B (K )J − [B (K )J] + B B .
Let us now assume that the Riccati equation (4.37), which is equal to 11 =
−δ1 I < 0, has a solution P1 > 0. Then, the equality 21 = 0 yields A = A − B F −
P2 F P1−1 . Moreover, 22 is then calculated as
22 = AP2 + P2 A + D

+ B − P2 F − J (K ) B
− FP2 − (K )J
− (P2 F + J (K ) )(FP2 + (K )J)
− P2 (F P1−1 + P1−1 F )P2 + g I.
Hence, the optimal B that minimizes the maximum eigenvalue of 22 is given by
B = P2 F + J (K ) .
Then, the existence of a solution P2 > 0 in (4.38) directly implies 22 = −δ2 I < 0.
As a result, we obtain ¯ = diag(−δ1 I, −δ2 I ) < 0, which leads to the objective
condition (4.42). Therefore, as mentioned above, we have limt→∞ V̄ (t) ≤ X̄ . Then,
as the third and fourth diagonal elements of the matrix V̄ (t) are, respectively, given
by V̄33 = z̄ 32 = (q(t) − q (t))2 and V̄44 = z̄ 42 = ( p(t) − p (t))2 , we obtain the
assertion (4.40).
The quantum robust observer constructed above has the following important prop-
erty, which one would expect it to have: if the uncertainties are small (or zero), the
robust observer is close (or identical) to the optimal quantum Kalman filter. The
formal statement of this fact is given as follows (for the proof, see [8]).
Proposition 4.1 Consider the case where the uncertainties converge to zero: i.e.,
R(t) → 0. Then, there exist parameters δ1 , δ2 , and such that the robust observer
(4.39) converges to the stationary Kalman filter associated with the system (4.34)
with A(t) = 0.
This proposition also states that we can find the parameters (δ1 , δ2 , ) such that
the robust observer (4.39) approximates the stationary Kalman filter when the uncer-
tainties are small, because the solutions of the Riccati equations (4.37) and (4.38)
are continuous with respect to the above parameters.
4.4.2 Example
Here, we will demonstrate that, in the presence of uncertainty, the robust observer
actually outperforms the nominal Kalman filter, which is no longer optimal for
uncertain systems. The system considered here is a single particle with position
q and momentum p, which is trapped around the origin of the antiharmonic poten-
tial field V (q) = −0.05q 2 . For instance, this can be interpreted as an approximate
potential for the double-well potential V (q) = q 4 − 0.05q 2 . Then, the corresponding
Hamiltonian is given by H free = 2 p 2 − 0.05q 2 = x Rx with R = diag(−0.05, 2).
Apart from this, the system is driven by the control Hamiltonian

0
H control
= −u(t)q, i.e., F = , (4.43)
1
where u(t) is the control input that will be described

√ later. In addition, the system is
subjected to an uncertain Hamiltonian H = − d(t)q 2 , where d(t) is an unknown
time-varying parameter bounded by the known constant g ≥ 0, i.e., d(t) ∈ [0, g] for
all t. For this system, we consider the continuous QND measurement of q, corre-
sponding to L = q or equivalently K = [1 0].
With the above setup, we perform a numerical comparison between the developed
robust observer and the nominal Kalman filter. For the robust observer, we evaluate
the guaranteed upper bound of the estimation error Tr(P2 ) in (4.40). To calculate
Tr(P2 ), we set the parameters δ1 and δ2 to be both 0.1 and also chose that minimizes
Tr(P2 ). For the nominal Kalman filter, on the other hand, we evaluate the stationary
mean-square error between the “true” system and the estimator for the “nominal”
system corresponding to d(t) = 0 (see Appendix C of [8]). As for the control part,
here we take the feedback of the form u(t) = x (t) in the robust observer case and
u(t) = πt (x) in the nominal Kalman filter case. The control gain is set to that
of the stationary LQG controller for the nominal system (see Chap. 5 for details of
LQG control for linear quantum systems).
Let us describe the simulation results. In Table 4.1, the two estimation errors men-
tioned above are provided for several values of g. The uncertainty d(t) is now set to
the worst case value d(t) = g for each g. In the first row of the table, “N/A” means
that the nominal Kalman filter fails in the estimation. That is, the error dynamics
between the true system and the nominal Kalman filter diverges, and it has no sta-
tionary solution. On the other hand, the second row in Table 4.1 shows that the robust
observer is not very sensitive to the uncertain parameter g and eventually works as a
tolerant estimator.
Table 4.1 The stationary estimation error of the nominal Kalman filter, denoted by “KAL”, and
the guaranteed upper bound of the estimation error of the robust observer (i.e., Tr(P2 )), denoted by
“ROB”. Note that the robust observer does not coincide with the nominal Kalman filter even when
g = 0, because the parameter δ2 is now set to a non-zero value. Reprinted with permission from [8]
© (2006) by the American Physical Society
g 0.00 0.20 0.38 0.60 0.80 0.97
KAL 1.43 2.38 40.88 N/A N/A N/A
ROB 1.73 3.32 4.74 7.04 10.12 14.13
We have discussed a robust estimation method that guarantees the estimation error
to be within a known bound. A general framework in this direction is to design
an estimator or controller so that a given cost is guaranteed to be bounded by a
known constant even under unknown perturbations. In particular, it is important to
guarantee stability of the system under unknown perturbations, and there are actually
several methods to design such robust estimators/controllers; see e.g., [14]. Another
direction pursuing the robustness property is the risk-sensitive approach, which has
also been studied extensively in the classical case [15, 16]; in this method, rather
than minimizing a cost function V (x) ≥ 0 with x a system variable, we consider the
problem minimizing e V (x) . In fact, we find that several quantum versions of these
risk-sensitive estimators or controllers have been developed [2, 8, 17, 18].
References
Optim. 46, 2199–2241 (2007)
2. N. Yamamoto, L. Bouten, Quantum risk-sensitive filtering and robustness. IEEE Trans. Autom.
Control 54(1), 92–107. Reprinted, with permission, © 2009 IEEE (2009)
3. S. Sakai, C*-Algebras and W*-Algebras (Springer, Berlin, 1998)
4. L. Bouten, R. van Handel, Quantum filtering: a reference probability approach (2006). arXiv
preprint, http://arXiv.org/abs/math-ph/0508006
6. R. van Handel, J.K. Stockton, H. Mabuchi, Feedback control of quantum state reduction. IEEE
Trans. Autom. Control 50(6), 768–780 (2005)
7. H.I. Nurdin, Quantum filtering for multiple input multiple output systems driven by arbitrary
zero-mean jointly Gaussian input fields. Russ. J. Math. Phys. 21(3), 386–398 (2014). Reprinted,
with permission, © (2006) by the American Physical Society
8. N. Yamamoto, Robust observer for uncertain linear quantum systems. Phys. Rev. A 74, 032107.
Reprinted, with permission, Copyright (2006) by the American Physical Society (2006)
11. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Quantum optical realization of classical linear
stochastic systems. Automatica 49(10), 3090–3096 (2013)
12. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Network Synthesis for a Class of Mixed Quantum-
Classical Linear Stochastic Systems. arXiv: 1403.6928 (2014, arXiv preprint)
13. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Synthesis and structure of mixed quantum-
classical linear systems, in Proceedings of the 51st IEEE Conference on Decision and Control
(Maui, Hawaii, Dec 10–13, 2012) (2012), pp. 1093–1098
14. I.R. Petersen, V. Ugrinovskii, M.R. James, Robust stability of uncertain linear quantum systems.
Philos. Trans. R. Soc. A 370, 5354–5363 (2012)
15. P. Dupuis, M. James, I. Petersen, Robust properties of risk-sensitive control. Math. Control
Syst. Signals 13, 318–332 (2000)
References 151
16. R.K. Boel, M.R. James, I.R. Petersen, Robustness and risk sensitive filtering. IEEE Trans.
Autom. Control 47(3), 451–461 (2002)
17. M.R. James, A quantum Langevin formulation of risk-sensitive optimal control. J. Opt. B:
Quantum Semiclass. Opt. 7, 198 (2005)
18. C. D’Helon, A.C. Doherty, M.R. James, S.D. Wilson, Quantum risk-sensitive control, in Pro-
ceedings of the 45th IEEE Conference on Decision and Control (CDC) (2006), pp. 3132 –
3137
Chapter 5
Feedback Control of Linear Dynamical
Quantum Systems
Abstract This chapter introduces and treats the topic of quantum feedback control
on linear quantum systems. Two distinct approaches to feedback control of quantum
systems are considered: measurement-based feedback control and coherent feed-
back control. For measurement-based feedback control, the notions of controlled
Hudson–Parthasarathy QSDEs and controlled quantum filtering equations are intro-
duced. Three control problems are formulated and solved: measurement-based linear
quadratic Gaussian control, coherent linear quadratic Gaussian control, and coherent
feedback H ∞ control. Examples are provided to illustrate each control method.
In this chapter, we shall formulate and discuss the solutions to three classes of quan-
tum feedback control problems for linear quantum systems and illustrate physical
settings where these control problems are relevant. They are the measurement-based
(quantum) linear quadratic Gaussian (LQG) control problem, coherent quantum LQG
problem, and coherent feedback H ∞ control problem. The exposition in this chapter
will build upon the theoretical foundations that have been developed in the previ-
ous sections, especially the use of quantum filters in the measurement-based LQG
control problem.
5.1 Measurement-Based Quantum Feedback Control
5.1.1 Controlled Quantum Evolution and Quantum Filter
We first set up the theoretical framework for measurement-based (quantum) feed-

back control. In this type of control, continuous linear measurement is performed on
the output Y(t) of a linear quantum system to gain information about the system that
is then used to drive a classical controller. The controller processes the measurement
to produce a control signal that is fed back to the system to control its evolution.
Section 5.2 contains materials reprinted from Automatica [12], with permission from Elsevier.
Section 5.3 and the associated appendices contain materials reprinted, with permission, from
[34] © 2008 IEEE.
154 5 Feedback Control of Linear Dynamical Quantum Systems
Describing a quantum system undergoing measurement and feedback is non-trivial,

and to properly describe such feedback in the quantum setting requires setting up
a modified Hudson–Parthasarathy QSDE that depends on the control signals, the
so-called Hudson–Parthasarathy QSDE for controlled quantum Markov flows
[1, 2]. We shall introduce this QSDE in this chapter and derive the solution to the
measurement-based LQG control problem, following the treatment in [2].
Following Sect. 4.2.2, we consider an n-degree of freedom linear quantum sys-
tem coupled with m bosonic fields, with quadratic Hamiltonian H = (1/2)x Rx,
coupling operator L = K x (K ∈ Cm×2n ), and scattering matrix S ∈ Cm×m . We
will reuse the notation introduced in Sect. 4.2.2. Thus, let ym (t) be the output of
a linear measurement performed on the quantum output y(t) of a linear quantum
system in real quadrature form (2.5), with the same number of inputs and outputs m.
We take the measurement ym to be of the form ym (t) = Fm y(t), with Fm ∈ R p×2m
satisfying the conditions set out in Sect. 4.2.2. We assume that we have two control
signals u 1 (t) ∈ Rc1 and u 2 (t) ∈ Rc2 (with c1 and c2 non-negative integers, and the
convention that u j (t) is dropped if c j = 0), which at each time t ≥ 0 are functions
of R p -valued stochastic processes on the time interval [0, t]. That is, at each time
t ≥ 0, u j (t) u j ({d(s), 0 ≤ s ≤ t}) for some R p -valued stochastic process d(t).
We assume that control can be applied in two ways:
1. Adding a control Hamiltonian term of the form Hc (t) = x N1 u 1 (t) to H , that
depends on the R p1 -valued control signal u 1 (t) with N1 ∈ R2n×c1 .
2. Driving the system with coherent input fields with amplitudes given by the ele-
ments of the vector N2 u 2 (t), with a given N2 ∈ Cm×c2 and u(t) an Rc2 -valued
control signal for each t.
The measurement-based feedback control loop is depicted in Fig. 5.1.
Let z(t) = Fm w(t). Since w(t) satisfies the same differential commutation rela-
tion as y(t), we have that [z(t), z(s) ] = 0 for all s, t ≥ 0. Thus, z is isomorphic to
some classical stochastic process, and the definition u j ({z(s), 0 ≤ s ≤ t}) makes
sense. To reduce notational clutter, let us introduce the notation z [0,t] = {z(s), 0 ≤
Fig. 5.1 Measurement-

based feedback control Quantum
loop plant
u2 u1
(Input field (Hamiltonian Measurement
modulation) modulation)
Classical
controller
5.1 Measurement-Based Quantum Feedback Control 155
s ≤ t} and write u j ({z(s), 0 ≤ s ≤ t}) simply as u j (z [0,t] ). We can now define the
controlled Hudson–Parthasarathy QSDE U (t),

dU (t) = Tr((S − I ) d(t)) + dA∗ (t)(L + S N2 u 2 (z [0,t] ))
− (S ∗ L + N2 u 2 (z [0,t] ))dA(t)

− ı H + Hc (t) + {L ∗ S N2 u 2 (z [0,t] )}

+ 1/2(L + N2 u 2 (z [0,t] ))∗ (L + N2 u 2 (z [0,t] ) dt U (t), (5.1)
with initial condition U (0) = I . Since u j acts on

a commutative process
z, it is
important to note that U (t)∗ u j z [0,t] U (t) = u j U (t)∗ z [0,t] U (t) = u j (ym )[0,t]
(here, following the convention above, (ym )[0,t] = {ym (s), 0 ≤ s ≤ t}). That is, in
the Heisenberg picture, u j becomes a function of the measurement ym , as it should be.
Define the time-dependent operators,
L(t) = L + S N2 u 2 (z [0,t] ) and H (t) = H + Hc (t) + {L ∗ S N2 u 2 (z [0,t] )}.
For any bounded operator X of the quantum system, let X (t) = jt (X ) = U (t)∗ XU (t)
be the time evolution of X in the Heisenberg picture. Then by an application of the
quantum Itō rule, we have that jt (X ) satisfies the QSDE [2, 3],
d jt (X )

∗ ∗ ∗
= jt ı[H (t), X ] + L(t) X L(t) − /2(L(t) L(t)X + X L(t) L(t)) dt
1
+dA(t)∗ jt (S ∗ [X, L(t)]) + jt ([L(t)∗ , X ]S)dA(t)
with initial condition j0 (X ) = X . Constructing the vector x(t) of the quadrature

form of the linear quantum system as in Sect. 2.3, we have the equations for the
controlled linear quantum dynamics,
d x(t) = Ax(t)dt + Bc u(t)dt + Bdw(t), (5.2)
with A and B as given by (2.6) and (2.7), respectively,

u 1 (ym )[0,t]
u(t) = ,
u 2 (ym )[0,t]

and Bc = Bc,1 Bc,2 with Bc,1 ∈ R2n×c1 and Bc,2 ∈ R2n×c2 given by
Bc,1 = 2Jn N1 ,
Bc,2 = −4Jn {ı K ∗ S N2 }.
We also have the linear QSDE for the measurement ym ,

dym (t) = Fm dy(t)

= Fm (C x(t)dt + Dc u(t)dt + Ddw(t)), (5.3)
with C and D given by (2.8) and (2.9), respectively, and

0m×c1 2 {S N2 }
Dc = Pm .
0m×c1 2{S N2 }
˜
Let V , Ã(t) = V A(t), Ỹ(t) = V Y(t) = V U (t)∗ A(t)U (t), and (t) = V # (t)V
be as in Sect. 4.2.2, and
ỹm (t) = M −1 ym (t) = [ I p 0 p×(2m− p) ](Ỹ(t) + Ỹ(t)# ),
be another measurement process. Recall the notation Ym,t = vN{ym (s), 0 ≤ s ≤ t}

from Chap. 4 for the von Neumann algebra generated by ym, j (s), j = 1, 2, . . . , p
and 0 ≤ s ≤ t, where ym, j is the j-th element of the vector ym . We similarly let
Y˜m,t = vN{ ỹm (s), 0 ≤ s ≤ t} denote the von Neumann algebra generated by the
components of ỹm (s), 0 ≤ s ≤ t. Since ỹm is linearly related to ym and M is invertible,
they generate the same von Neumann algebra. Therefore, πt (X ) = P( jt (X ) | Ym,t ) =
P( jt (X ) | Y˜m,t ). Define L̃ = V L, Ñ2 = V N2 , and S̃ = V SV ∗ , and the operators
L̃(t) = L̃ + S̃ Ñ2 u 2 (z [0,t] ) and H̃ (t) = H + Hc (t) + { L̃ ∗ S̃ Ñ2 u 2 (z [0,t] )}.
Note that H̃ (t) = H (t). In terms of these operators and the processes Ã(t), Ã(t)∗ ,
˜
and (t), the controlled QSDE can be rewritten as:
˜
dU (t) = Tr(( S̃ − I ) d (t)) + d Ã(t)∗ L̃(t) − L̃(t)∗ S̃d Ã(t)
− ı H̃ (t) + 1/2 L̃(t)∗ L̃(t) dt U (t). (5.4)
It then follows by a straightforward extension of the results of [2] to multiple inputs

and multiple outputs that the controlled quantum filtering equation is given by
dπt (X )
⎛ ⎞

m
= πt ⎝ı[ H̃ (t), X ] + 1/2 L̃ j (t)∗ [X, L̃ j (t)] + [ L̃ j (t)∗ , X ] L̃ j (t) ⎠ dt
j=1

p
+ πt (X L̃ j (t) + L̃ j (t)∗ X ) − πt ( L̃ j (t) + L̃ j (t)∗ )πt (X ) d ν̃ j (t),
j=1
(5.5)
t
where ν̃ j (t) = ỹm, j (t) − 0 πs ( L̃ j (s) + L̃ j (s)∗ )ds is the j-th component of the
t
innovations process ν̃(t) = ỹm (t) − [ I p 0 ] 0 πs ( L̃(s) + L̃(s)∗ )ds. The process ν̃(t)
is a Ym,t -martingale zero-mean Wiener process [2], following the same line of proof
sketched earlier in Chap. 4. Thus, ν̃(t) − ν̃(τ ) is independent of any function of Ym,τ
for any t ≥ τ ≥ 0.
With x = (q1 , p1 , q2 , p2 , . . . , qn , pn ) as before, let
x̂(t) = πt (x) = (πt (q1 ), πt ( p1 ), πt (q2 ), πt ( p2 ), . . . , πt (qn ), πt ( pn )) .
Then, using the definitions of H̃ (t) and L̃(t), explicitly calculating all commutators
on the right-hand side of (5.5), and taking the initial state of the system to be Gaussian
(analogous to the calculations leading to (4.31) and (4.32) in Chap. 4), yields:
d x̂(t) = A x̂(t)dt + Bc u(t)dt + (P(t)C Fm + B)

−1
×(Dm Dm ) (dym (t) − Fm (C x̂(t) + Dc u(t))dt)
= A x̂(t)dt + Bc u(t)dt
−1
+(P(t)C Fm + B)(Dm Dm

) dν(t). (5.6)

with initial condition x̂(0) = x , where t = D Fm and Dm = Fm D, ν(t) is
the innovations process ν(t) = ym (t) − Fm 0 (C x̂(t) + Dc u(s))ds, and P(t) is the
symmetrized covariance matrix,
1
P(t) = πt (x − x̂(t))(x − x̂(t)) + ((x − x̂(t))(x − x̂(t)) ) ,
2
satisfying the deterministic matrix Riccati differential equation
Ṗ(t) = A P(t) + P(t)A + B B

−(P(t)C Fm + B)(Dm Dm
−1
) (P(t)C Fm + B) , (5.7)
with initial condition

1
P(0) = P xx + (xx ) , (5.8)
2
with x = x − x .
In treatment of the measurement-based LQG control, coherent quantum LQG
control, and coherent feedback H ∞ control that will now follow, the initial state of
the system will always be assumed to be a Gaussian state and the bosonic fields
coupled to the quantum control system are also Gaussian.
5.1.2 Measurement-Based LQG Control
In the measurement-based LQG control problem, we wish to find a control law u(t)
of dimension c = c1 + c2 , with components belonging in Ym,t for each t ≥ 0, to
minimize the quadratic cost function
T
JT (u) = P (C x x(t) + Du u(t)) (C x x(t) + Du u(t))dt
0

+x(T ) Qx(T ) , (5.9)
for a fixed terminal time T > 0, where C x , Du , and Q are some prespecified constant
matrices in Rk×2n , Rk×c , and R2n×2n , with k some positive integer and Q = Q ≥ 0.
Problems for linear quantum systems with quadratic costs of this type are commonly
encountered for cooling and confinement; see, e.g., [4–6] and the references therein.
To be concrete, take, for instance, a single quantum harmonic oscillator (n = 1),
C x = [ I2 0 ] , Du = [ 0 Ic ] and Q = I , then we have that
P((C x x(t) + Du u(t)) (C x x(t) + Du u(t)) + x(T ) Qx(T ))

= P(q(t)2 + p(t)2 + u(t) u(t) + q(T )2 + p(T )2 ).
For this particular cost, cooling is quantified as minimizing the cumulative kinetic
T
energy 0 P( p(t)2 )dt of the system over the interval [0, T ] and its mean kinetic
energy P( p(T )2 ) at the terminal time T , and confinement is quantified as minimiz-
T
ing the cumulative displacement of the system 0 P(q(t)2 )dt away from the origin
over the interval [0, T ] and its mean displacement from the origin P(q(T )2 ) at the
T
terminal time T . The remaining term 0 P(u(t) u(t))dt represents a penalty on the
cumulative control effort required to achieve the cooling and confinement task.
Using the properties of the quantum conditional expectation, the cost function
can be expressed as:
JT (u)

T
=P
P C x x(t) + Du u(t)) (C x x(t) + Du u(t)) Ym,t dt
0

+P P x(T ) Qx(T ) Ym,t
T

=P πt C x x(t) + Du u(t)) (C x x(t) + Du u(t)) dt
0

+P πt x(T ) Qx(T )
T

=P πt x(t) C x C x x(t) + πt (x) C x Du u(t)
0

+u(t) Du C x πt (x)
+ u(t)
Du Du u(t) dt

+P πt x(T ) Qx(T ) .
t
Since the innovations process ν(t) = ym (t) − Fm 0 (C x̂(s) + Dc u(s))ds is a
Ym,t -martingale zero-mean Wiener process, it follows that (5.6) can be viewed as
essentially a controlled fully observable classical linear stochastic system with
deterministic initial condition x̂(0) = x driven by a Wiener noise process ν(t).

Moreover, we note that (5.6)–(5.8) are precisely the Kalman filtering equation that
would be obtained if one ignores the quantum mechanical origin of (5.2) and viewing
them as describing a classical linear stochastic state-space model with state vector
x(t) [7, 8]. The upshot of this observation is that via the quantum conditional expec-
tation, the measurement-based LQG control problem can be mapped to an equivalent
classical LQG control problem, and so the solution of the former problem is of an
analogous form to the latter. That is, the measurement-based LQG control law is of
the form u opt (t) = −K (t)x̂(t), where K (t) ∈ Rc×2n is an optimal time-dependent
gain such that u opt solves the deterministic linear quadratic regulator (LQR) problem
of minimizing the deterministic quadratic cost [9, 10]
T
JLQR,T (u) = (C x x (t) + Du u(t)) (C x x (t) + Du u(t))dt + x (T ) Qx (T ),
0
with x (t) a classical state vector satisfying the deterministic state-space equation,
ẋ (t) = Ax (t) + Bc u(t).
Using this analogy, the solution to the infinite horizon measurement-based LQG
problem with T → ∞, Q = 0, and infinite horizon cost function
T
1
J∞ (u) = lim sup P (C x x(t) + Du u(t)) (C x x(t) + Du u(t))dt
T →∞ T 0

+ x(T ) Qx(T ) , (5.10)
can also be obtained from the solution of the analogous infinite horizon classical
LQG problem.
In the above, we have taken a shortcut to the solution of the measurement-based
LQG problem by exploiting the linearity of the dynamics and quadratic forms of
the cost function to set up an analogy with the classical LQG control problem. The
problem can also be solved in a general framework for measurement-based optimal
control of quantum systems. One can formulate the quadratic cost-to-go, pose the
measurement-based LQG problem as a dynamic programming problem, and write
down the associated Bellman equation that must be satisfied by the associated value
function. As in the classical setting, for a wide range of problems a separated law
holds that is a functional of the stochastic density operator ρ(t) of the quantum filter,
defined via the identity Tr(ρ(t)X ) = ι(πt (X )) for all bounded system operators
X . Roughly speaking, if there is a separated law satisfying the associated Bellman’s
equation, then this policy will be optimal. This is the case for the measurement-based
LQG control problem. For the complete details, we refer the reader to [2, 11], with
[11] focusing on the measurement-based quantum LQG problem. This framework
can be applied to a wide range of optimal control problems for measurement-based

feedback control of quantum systems, beyond the quantum LQG control problem.
We conclude the exposition of measurement-based LQG control with a discussion
of an example application to a trapped atom system, as proposed in [5].
Example 5.1 Consider the off-resonant interaction between a two-level atom and an
optical cavity in which it is trapped. The interaction between the atom and the optical
cavity is given by the Hamiltonian,
g02 ∗
H = Ha − a a cos(k0 qat )2 ,

where a is the annihilation operator of the cavity mode, ω0 is the resonance frequency
of the cavity mode, k0 = ω0 /c is the wave number of the cavity mode (here, c denotes
the speed of light), the detuning between the cavity mode and the two-level atom
(i.e., = ωa − ω0 , where ωa is the atomic transition frequency), qat the atomic
position operator, g0 the cavity-QED coupling constant giving the strength of the
interaction between the cavity mode and the atom, and Ha is the Hamiltonian for the
mechanical motion of the atom.
The work [5] is interested in estimating the position of the atom in the cavity based
on continuous measurement of the phase quadrature of the output light leaking out
of the cavity (i.e., −ıY(t) + ıY(t)∗ ) via homodyne detection. In order to do this, it
is assumed that the atom is trapped in a region of a size which is small compared to
the wavelength of the light, about a region halfway between a node and an antinode
of the standing field inside the cavity, so that the following approximation can be
made:
cos(k0 qat )2 = cos(k0 qat,0 + k0 qat )2 ≈ 1/2 + k0 qat .
Renaming qat as qat , the Hamiltonian H then takes the approximate form,
g02 ∗ g2
H = Ha − a a − k0 0 a ∗ aqat ,
2
state | f with f a constant function

The optical cavity is driven by a field in a coherent
γP
equal to α = 2E for all times, where E = ω 0
and P are the driving laser power.
The decay rate of the cavity is taken to be γ, so that the coupling operator of the
√
cavity to the laser is L 1 = γa. Measurement will be performed on the output field
from the cavity; however, this measurement may be lossy (photons are lost in the
measurement process), and this is modeled by mixing the output from the cavity at
a beam splitter with an auxiliary vacuum field to which photons will be lost. The
beam splitter is assumed to have transmission rate of ≥ 0 and is represented by the
unitary matrix √ √
√ 1−
S= √ .
− 1−
If bin,1 is the input to the first port of the beam splitter and bin,2 the input to the second
port, then the corresponding outputs bout,1 and bout,2 , respectively, would be given by

√ √
bout,1 b √ 1− bin,1
= S in,1 = √ .
bout,2 bin,2 − 1− bin,2
In our case, bin,1 will come from the output field of the cavity, and measurements
will be performed on the output bout,1 . Overall, using the (S, L , H ) notation and
series product operation from Chap. 3, the coherently driven atom and cavity system
followed by the beam splitter is given by
√
γa+α ∗
G = (S, 0, 0) I, , H + { /2(a − a )}
α
0
√ √ √ √
√ 1− √ γ a + √α
= √ , ,
− 1− − (1 − )γ a − 1 − α

H + {α/2(a − a ∗ )}
The measurement will be that of the phase quadrature of the first output of the beam
splitter and takes the form (using the fact that α is real)
√ √
dym (t) = β γ γ(−ıa(t) + ıa(t)∗ )dt − ıdB(t) + ıdB ∗ (t) ,
where β is a real constant relating to the strength of the local oscillator and the
reflectivity
√ √ the beam splitter in the homodyne detection setup, and B(t) =
of
A1 (t) + 1 − A2 (t). Here, A1 (t) is the vacuum annihilation operator for the
field coupled to the cavity, and A2 (t) is the vacuum annihilation operator for the
field entering the second port of the beam splitter. In the context of the theory lead-
√
ing to this example, we have that p = 1, V = diag(−ı, 1) and M = β γ. Let
√ √ √ √
L1 = γa+ α, L 2 = − (1 − )γ a − 1 − α. The quantum filtering equa-
tion for any atomic operator X is given by
⎛ ⎞

2

dπt (X ) = πt ⎝ı/[H, X ] + 1/2 L ∗ [X, L ] + [L ∗ , X ]L ⎠ dt
j j j j
j=1

+ πt (−ı X L 1 + ı L ∗1 X ) − πt (−ı L 1 + ı L ∗1 )πt (X ) dν1 (t)

= πt ı[H, X ] + γ/2 a ∗ [X, a] + [a ∗ , X ]a dt
√ √
+ γ πt (−ı Xa + ıa ∗ X ) − πt (−ıa + ıa ∗ )πt (X ) (β γ)−1 dν1 (t),
where ν1 (t) is the innovations process given by,


√ t
ν1 (t) = ym (t) − β γ πs (−ı L 1 + ı L ∗1 )ds
0
t
= ym (t) − β γ πs (−ıa + ıa ∗ )ds.
0
In [5] the quantum filtering equation is simplified further by transforming to a “dis-

placement picture,” using the
fact
that without the atom the cavity state would con-

verge to the coherent state −2 γωP
0
, where the state of the optical cavity becomes
close to the vacuum, and making the assumption
k0 g02 (| − 2E/γ|2 + 1)| x |

= δ 1.
γ
Under these assumptions, the cavity has very fast dynamics and goes to a steady state
very quickly, relative to the timescale of the dynamics of the atom. The cavity can
thus be adiabatically eliminated. It is shown that the quantum filtering equation can
be approximated as,

dπt (X ) = πt ı/[H , X ] + κ (qat [X, qat ] + [qat , X ]qat ) dt
√ √
+ 2 κ (πt (Xqat + qat X ) − 2πt (qat )πt (X )) (β γ)−1 dν1 (t),
with the modified innovations process ν1 (t) given by,
t
ν1 (t) = ym (t) − 2β 2κγ πs (qat )ds.
0
In the above, H = Ha − |α|2 qat and κ = 2k02 g04 |α|2 /(γ2 ). If we assume that
Ha is quadratic, then the filtering equation above becomes a linear quantum Kalman
filtering equation. This will be the case for a harmonically trapped atom as considered
pat + m at2ω qat2 . Here, m at is the mass of the atom and ω its
2
in [5], where Ha = 2m 1 2
oscillation frequency.
Given the reduced approximate atomic quantum filtering equation, let x =
(qat , pat ) , where pat is the atomic momentum operator and we have the com-
mutation relation [qat , pat ] = ı. Consider a control Hamiltonian of the form
Hc (t) = x Ju 1 (t), with u 1 having two components, and set u 2 (t) = 0. That is, the
system will only be controlled through Hc (t) alone. This corresponds to Bc,1 = I2
(and of course, Bc,2 = 0). One can consider the LQG problem of minimizing the
cost function (5.10) with
⎡√ ⎤ ⎡ ⎤
m at ω 0 0 0
√
⎢ 0 1/ m at ⎥ ⎢
⎥ , Du = r ⎢ √ 0 0 ⎥
Cx = ⎢
⎣ 0 ⎦ ⎣
⎥,
0 m at ω 0 ⎦
√
0 0 0 1/ m at
where r is a positive weighting constant. In the harmonically trapped scenario consid-

ered here, the symmetrized covariance matrix satisfies the matrix Riccati differential
equation,

Ṗq Ṗq p
Ṗ =
Ṗq p Ṗp

2Pq p /m at − 8κ Pq2 Pp /m at − m at ω 2 Pq − 8κ Pq p Pq
= ,
Pp /m at − m at ω 2 Pq − 8κ Pq p Pq −2m at ω 2 Pq p − 8κ Pq2p + 2κ2
with a steady-state solution that can be analytically computed by setting the left-hand
side of the above equation to zero and solving for constants Pq , Pq p and Pp (explicit
expressions for these can be found in [5]). In this case, it turns out that the infinite
horizon measurement-based LQG problem has an analytical solution minimizing
the cost (5.10) given by u opt,1 (t) = K ∞ πt (x), with K ∞ a constant diagonal matrix,
K ∞ = (1/r )I2 .
5.2 Coherent Feedback Quantum LQG Control
An alternative approach to quantum feedback control is to dispense of the measure-

ment process and taking the controller to be another quantum system that is attached
to the plant in a feedback interconnection, as illustrated in Fig. 5.2. This control par-
adigm where a quantum system is controlled by another quantum system is often
referred to as coherent (quantum) feedback control. The interconnection is achieved
by passing some output fields of the plant to the controller and vice versa. All the
information that is being exchanged in the loop are quantum information. There are
several reasons for considering coherent feedback controllers, among them:
Fig. 5.2 Coherent feedback

control loop Quantum
plant
Quantum Quantum
control signal information
Quantum
controller
• Being a quantum system, such a controller can match the speed of the response of
the quantum plant.
• A coherent feedback controller could potentially be fabricated on the same physical
platform as the quantum plant.
• Avoids issues of hardware mismatch between a quantum plant and the measure-
ment apparatus. The former is typically “cold,” operating at very low temperatures,
while the latter is typically “hot,” operating at room temperature.
• The quantum nature of the controller has the potential to offer a performance
improvement over classical measurement-based controllers by exploiting quantum
effects.
A coherent feedback version of the quantum LQG control problem was first formu-
lated in [12] and referred to as coherent quantum LQG control. This type of control
is the main content of this section.
We consider plants that are linear quantum systems described by the linear QSDE,

du(t)
d x(t) = Ax(t)dt + [ B Bw ] ,
dw(t)
dy(t) = C x(t)dt + Dw dw(t),
z(t) = C z x(t) + Dz βu (t). (5.11)
Here, w(t) is a quantum Wiener disturbance vector, βu (t) is an adapted, self-adjoint

process commuting with x(t) (i.e., [βu (t), x(t) ] = 0), and u(t) is a control input
of the form
du(t) = βu (t)dt + d ũ(t). (5.12)
In the decomposition of u(t), βu (t) is identified as the “signal” part while ũ(t)
is the noise part. The vectors w(t) and ũ(t) are assumed to be distinct vacuum
quantum noises, though this assumption is straightforward to relax. The dimensions
of x(t), u(t), w(t), and y(t), are 2n, 2n u , 2n w , and 2n y , respectively. Also, z(t) has
been introduced as a performance variable through which a cost function that will
be optimized will be defined. We note that z(t) need not correspond to a physical
observable.
In coherent quantum LQG control, we take as the controller another linear quan-
tum system with the same degree as the plant, described by the linear QSDE:
⎡ ⎤
dw K 1 (t)
d x K (t) = A K x K (t)dt + [ B K 1 B K 2 B K 3 ] ⎣ dw K 2 (t) ⎦ ,
dw K 3 (t)
du(t) = C K x K (t)dt + dw K 1 (t). (5.13)
Here, x K represents the position and momentum quadratures of the quantum har-
monic oscillators that form the controller, B K 2 is a square matrix of the same dimen-
sion as A K , and B K 1 has the same number of columns as there are rows of C K . The
5.2 Coherent Feedback Quantum LQG Control 165
vectors w K i , i = 1, 2, are noise vectors of non-commutative Wiener processes (in

vacuum states) with dimensions 2n w K i , and y is an input to the controller coming from
the plant (5.11). The plant and controller are interconnected in a fully quantum feed-
back loop by the identification w K 3 (t) ≡ y(t), βu (t) ≡ C K ξ(t), and ũ(t) ≡ w K 1 (t),
resulting in a closed-loop system that is another linear quantum system with internal
dynamics given by the QSDE,
d x̃(t) = Ã x̃(t)dt + B̃dwcl (t),

z(t) = C̃ x̃(t), (5.14)
where x̃(t) = [ x(t) x K (t) ] , and

⎡ ⎤
w(t)
A BC K
wcl (t) = ⎣ w K 1 (t) ⎦ ; Ã = ;
BK 3 C A K
w K 2 (t)

Bw B 02×2
B̃ = ; C̃ = C z Dz C K .
B K 3 Dw B K 1 B K 2
We can associate with the closed-loop system (5.14) a quadratic performance index
T
JT (u) = z(t) z(t) dt. (5.15)
0
Here, · refers to quantum expectation with respect to the initial state of the composite
system consisting of the plant, controller, and all quantum noises. Let us define the
symmetrized covariance matrix P(t) by
P(t) = 1/2 x̃(t)x̃(t) + (x̃(t)x̃(t) ) . (5.16)
We will now express the performance index in terms of P(t) and its dynamical
equation. Applying the quantum Itō rule, we have
d P(t) = 1/2( d x̃(t) x̃(t) + (d x̃(t) x̃(t) ) +

x̃(t) d x̃(t) + (x̃(t) d x̃(t) ) + ( B̃ B̃ + ( B̃ B̃ ) )dt)
= ( Ã P(t) + P(t) Ã + B̃ B̃ )dt.
Therefore, P(·) satisfies the differential equation
Ṗ(t) = Ã P(t) + P(t) Ã + B̃ B̃ ; P(0) = P0 . (5.17)

From the symmetry of C̃ C̃ and P, it follows that
z z = x̃ C̃ C̃ x̃
= Tr(x̃ C̃ C̃ x̃)
= 1
2
Tr(C̃ C̃[x̃ x̃ + (x̃ x̃ ) ])
= Tr(C̃ C̃ P).
Hence, we may express (5.15) as

T
JT (u) = Tr(C̃ C̃ P(t))dt, (5.18)
0
where P(t) solves (5.17). We now focus on the infinite horizon case by passing to the
limit T ↑ ∞. If the matrix Ã is Hurwitz, standard results on Lyapunov equations give
us limt→∞ P(t) = P, where P is the unique symmetric positive definite solution to
the Lyapunov equation:
Ã P + P Ã + B̃ B̃ = 0. (5.19)
Following [12], we consider the infinite horizon coherent quantum LQG problem
with an infinite horizon cost function given by
T
1
J∞ (u) = lim sup z(s) z(s) ds. (5.20)
T →∞ T 0
From (5.19) and standard methods of analysis, we have
J∞ (u) = Tr(C̃ C̃ P) = Tr(C̃ P C̃ ).
To represent a linear quantum system, the controller is required to be physically

realizable; thus, the matrices A K , B K 1 , B K 2 , B K 3 , and C K must satisfy the physi-
cal realizability constraints (2.28)–(2.30). Since in this case (2.30) is automatically
satisfied, this gives the following two constraints,

A K Jn + Jn A
K + B K j Jm j B K j = 0, (5.21)
j=1
Jn C K + B K 1 Jm 1 = 0, (5.22)
where m j = n w K j . The cost-bounded coherent quantum LQG problem can now be

formulated.
Problem 5.1 Given a cost-bound parameter γ > 0, find system matrices A K , B K 1 ,

B K 2 , B K 3 , and C K for the controller (5.13) such that the following requirements are
satisfied.
F1. There exists a symmetric matrix P > 0 satisfying (5.19).

F2. J∞ = Tr(C̃ P C̃ ) < γ.
F3. The physical realizability constraints (5.21) and (5.22) are satisfied.
Problem 5.1 is a formidable non-convex optimization problem in the variables A K ,
B K 1 , B K 2 , B K 3 , and C K since the physical realizability constraints are already
non-convex in these variables. Unlike the measurement-based LQG control prob-
lem, there is currently no known analytical solution to the coherent quantum LQG
problem. The first treatment of this problem in [12] reformulated it into a rank-
constrained linear matrix inequality (LMI) problem and proposed numerical methods
to solve the reformulation using specialized optimization algorithms. Several other
approaches to solve this problem have since been proposed in the literature, e.g., [13–
15]. In particular, [13] also considers the possibility of allowing a direct quadratic
Hamiltonian coupling between the plant and the controller besides the interconnec-
tion via the quantum fields that is considered here. We will now describe the original
approach developed in [12] to solving the cost-bounded coherent quantum LQG
control problem.
5.2.1 Reformulating the Quantum LQG Problem

into a Rank-Constrained LMI Problem
The purpose of this section is to show how Problem 5.1 can be transformed into a rank-
constrained LMI problem. The latter problem can then be treated with appropriate
numerical methods. The key tool will be a matrix lifting and linearization technique
for matrix-valued polynomials in matrix variables. To keep the exposition simple,
we shall focus, for the purpose of illustration, on a plant and controller of degree
n (recall that we seek a controller which is of the same order as the plant) with
n y = n u = n and B K 1 , B K 2 , B K 3 , C K all of dimension 2n × 2n. The matrix-lifting
technique is too complicated to describe in a general form. Besides this, the choice
of matrix-lifting variables is not unique and, for efficiency, the choice needs to be
considered on a case-by-case basis and tailored to the particular plant at hand, in
order take advantage of existing structures in a particular problem.
Consider now a plant (5.11) of degree n with n y = n u = n and a linear quantum
controller (5.13) of the same degree with n w K 1 = n w K 2 = n (hence B K 1 , B K 2 ∈
R2n×2n ). In this case, we have that P will be a symmetric 4n × 4n matrix. We now
transform the constraints (5.19) and J∞ < γ into an LMI constraint. We do this
by exploiting a nonlinear change of variables that was proposed in [16, Eq. (35)].
However, this requires a suitable redefinition of the plant and controller without
altering the closed-loop equations. To this end, we redefine our plant equations as,
d x(t) = Ax(t)dt + Bβu (t)dt + Bw dw (t); x(0) = x,

dy (t) = C x(t)dt + Dw dw (t),
z(t) = C z x(t) + Dz βu (t), (5.23)
with w = [ w w
K 1 w K 2 ] , Bw = [ Bw B 02n×2n ], C = [ 02n×2n 02n×2n C ] ,
and ⎡ ⎤
02n×2n w I2n 02n×2n
Dw = ⎣ 02n×2n w 02n×2n I2n ⎦ .
Dw 02n×2n 02n×2n
Here, y is the output equation for the modified plant. It contains the quantum noise
w K 2 which entered in the original controller equations but not in the original plant
equations. This redefinition means that all noises can be viewed as coming from the
modified plant, as in standard classical LQG problems. We now also redefine our
controller equations as,
d x K (t) = A K x K (t)dt + B K dy (t),

βu (t) = C K x K (t), (5.24)
with B K = [ B K 1 B K 2 B K 3 ]. A simple inspection shows that interconnecting (5.23)

and (5.24) gives the same closed-loop equation (5.14). We are now in the setup of
[16], with D K = 0 in [16, Eq. (2)].
Introduce the auxiliary variables N , M, X, Y, Q ∈ R2n×2n , with X, Y, Q sym-
metric. Application of the nonlinear change of variables from [16, Sect. IV-B] with
D̂ = D K = 0) yields:
A = N A K M + N B K C X + YBC K M + Y AX, (5.25)

B = N BK , (5.26)
C = C K M . (5.27)
In terms of the newly defined variables above, the constraints (5.19) and J∞ < γ
can be recast as the following LMI constraint [16, Eq. (14)]:
⎡
AX + X A + BC + (BC)
⎣ A + A
(YBw + BDw )
⎤
A + A Bw
A Y + Y A + BC + (BC ) YBw + BDw ⎦ < 0,
(Bw ) −I
(5.28)
⎡ ⎤
X I (C z X + Dz C)
⎣ I Y C z ⎦ > 0. (5.29)
C z X + Dz C C z Q
Tr(Q) < γ. (5.30)
Since the controller is of the same degree as the plant, the matrices N and M in the
above may be freely chosen to be any pair of (invertible) square matrices satisfying
M N = I − XY.
If there exists matrices A, B, C, X, Y, Q verifying the LMIs (5.28)–(5.30) as well
as matrices N and M satisfying M N = I −XY, the controller matrices A K , B K , C K
can be recovered from the transformed variables (5.25)–(5.27) as [16, Eq. (40)]:
C K = CM − , (5.31)
B K = N −1 B, (5.32)
A K = N −1 (A − N B K C X − YBC K M − Y AX)M − . (5.33)
Let us now multiply the left- and right-hand sides of (5.21) with N and N , respec-
tively, and define the new variables N̆ = N Jn , Ă K = N A K , and B̆ K i = N B K i ,
i = 1, 2, 3. In terms of the new variables, the physical realizability constraints (5.21)
and (5.22) can be cast as,
(−AM − + ( B̆ K 3 C + Y A)XM − + YBC K ) N̆

+ N̆ (AM − − ( B̆ K 3 C + Y A)XM − − YBC K )
3
+ B̆ K i Jn B̆ Ki = 0, (5.34)
i=1
B̆ K 1 = N̆ C K Jn . (5.35)
Conversely, if A K , B K 1 , B K 2 , B K 3 , C K solve Problem 5.1 and P solves (5.19), then

(5.28)–(5.30) are satisfied for some pairs M and N of square matrices satisfying
M N = I − XY; see [16, Sect. IV-B]. Furthermore, since the associated controller
solving Problem 5.1 is physically realizable by hypothesis, the constraints (5.34) and
(5.35) are thus automatically satisfied. Therefore, we may state following result:
Theorem 5.1 Under the assumptions of this section, Problem 5.1 has a solution for
a given γ > 0 if and only if there exist matrices A, B̆ K 1 , B̆ K 2 , B̆ K 3 , C, X, Y, N̆ ,
M, N , C K satisfying the LMIs (5.28)–(5.30) (with B = [ B̆ K 1 B̆ K 2 B̆ K 3 ]) and the
constraints (5.34) and (5.35), N̆ = N Jn , N M = I − YX and C = C K M .
It can be seen that the constraints (5.34) and (5.35) are essentially polynomial matrix
equality constraints in the (non-commuting) matrix variables A B̆ K 1 , B̆ K 2 , B̆ K 3 ,C,X,
Y, N̆ , and M − . That is, (5.34) and (5.35) are equality constraints involving matrix-
valued multivariate polynomials with non-commuting matrix-valued variables. One
approach to deal with these constraints is to turn them into a collection of scalar
multivariate polynomial equality constraints with scalar decision variables, by tak-
ing the decision variables to be the elements of the matrix variables (we need only
take the upper triangular elements of symmetric variables such as X). It is known
that a collection of scalar multivariate polynomial equality and/or inequality con-
straints can be “linearized” and converted into a set of linear equality and inequality
constraints in some symmetric positive definite matrix variable Z together with the
rank constraint rank(Z ) = 1. This can be done by introducing auxiliary variables
known as lifting variables together with auxiliary equality constraints on these lift-
ing variables [17, 18]. However, this conversion into scalar variables may not be
desirable since converting polynomial matrix constraints into a collection of scalar
polynomial constraints can result in scalar multivariate polynomials of orders much
higher than the order of the original matrix polynomial. This is impractical for the
case that we are considering here since there will be many scalar decision variables
involved, leading to a scalar polynomial program that may become too large to handle
numerically. Indeed, if we are to scalarize (5.34) we would end up with 34 decision
variables and constraints in polynomials of order 4, a substantially large problem.
Therefore, we would like to keep the matrix structure of our problem by pursuing
the idea of defining suitable matrix-lifting variables instead. This will proceed in a
very similar fashion to the scalar situation but being mindful of the fact that matrices,
unlike scalars, do not in general commute with one another.
We will now look at linearizing (5.34) and (5.35) by defining suitable matrix-
lifting variables and related equality constraints, and subsequently transforming the
coherent quantum LQG problem into an LMI problem with a rank 2n constraint. To
this end, let us now set M = I2n and N = I − YX. This removes one free matrix
variable, namely M − , to reduce the complexity of the problem. We will use 14
matrix-lifting variables W1 , W2 , . . . , W14 ∈ R2n×2n defined as follows: Wi = B̆ K i J,
i = 1, 2, 3, W4 = YB, W5 = B̆ K 3 C + Y A, W6 = N̆ C , W7 = N̆ X, W8 = A N̆ ,
W9 = YX, W10 = W4 W6 , W11 = W5 W7 , W12 = W1 B̆ K1 , W13 = W2 B̆ K2 and
W14 = W3 B̆ K3 . Now, let Z be a 46n × 46n symmetric matrix,
Zi, j = [Z kl ]k=2in+1,(i+1)2n,l=2 jn+1,( j+1)2n ,
x = (x1 , . . . , x8 ) = (1, 2, . . . , 8),
and
v = (v1 , . . . , v14 ) = (9, 10, . . . , 22).
We impose that Z satisfies the constraints,

⎫
Z ≥0 Zv6 ,1 − Zx8 ,x5 = 0 ⎪⎪
⎪
Z0,0 − I2n = 0 Zv7 ,1 − Zx8 ,x6 = 0 ⎪⎪
⎪
⎪
Z1,x6 − Zx6 ,1 = 0 Zv8 ,1 − Zx1 ,x8 = 0 ⎪⎪
⎪
⎪
Z1,x7 − Zx7 ,1 = 0 Zv9 ,1 − Zx7 ,x6 = 0 ⎪⎪
⎪
⎬
Zv1 ,1 − Zx2 ,1 Jn = 0 Zv10 ,1 − Zv4 ,v6 = 0
(5.36)
Zv2 ,1 − Zx3 ,1 Jn = 0 Zv11 ,1 − Zv5 ,v7 = 0 ⎪
⎪
⎪
Zv3 ,1 − Zx4 ,1 Jn = 0 Zv12 ,1 − Zv1 ,x2 = 0 ⎪
⎪
⎪
⎪
Zv4 ,1 − Zx7 ,1 B = 0 Zv13 ,1 − Zv2 ,x3 = 0 ⎪
⎪
⎪
⎪
Zv5 ,1 − Zx4 ,1 C − Zx7 ,1 A = 0 Zv14 ,1 − Zv3 ,x4 = 0 ⎪
⎪
⎪
⎭
Zx8 ,1 − Jn + Zv9 ,1 Jn = 0.
Terms of the form Za,b with a, b ∈ {x1 , . . . , x8 }∪{v1 , . . . , v14 } in the above should be
identified with Za,1 (Zb,1 ) . We can then express the LMI constraints (5.28)–(5.30)
in terms of Z by replacing A, B, C, X, Y with Zx1 ,1 , Zx2 ,1 Zx3 ,1 Zx4 ,1 , Zx5 ,1 , Zx6 ,1 ,
Zx7 ,1 , respectively. Meanwhile, the physical realizability constraints (5.34) and (5.35)
become the following linear equality constraints:
−Zv8 ,1 + Z
v8 ,1 + Zv11 ,1 − Zv11 ,1 + Zv10 ,1 − Zv10 ,1
+Zv12 ,1 + Zv13 ,1 + Zv14 ,1 = 0,
Zx2 ,1 − Zv6 ,1 Jn = 0. (5.37)
Finally, we also demand that Z satisfies a rank 2n constraint:
rank(Z ) ≤ 2n. (5.38)
Now, to understand the relationship of the rank-constrained LMI to our original

constraints, suppose that we can find Z satisfying (5.36)–(5.38) as well as the LMI
constraints (expressed in terms of block elements of Z ) and the rank constraint. Due
to the rank constraint, we have the factorization Z = V V , where V ∈ R46n×2n
and satisfy [Vi j ]i, j=1,...,n = I2n . Using (5.36), we can recover A, B̆ K i (i = 1, 2, 3),
C, X, Y, N̆ as, respectively, Zx1 ,1 , . . . , Zx8 ,1 , and Wi = Zvi ,1 , i = 1, . . . , 14. It then
follows that N = N̆ Jn and B K i = N −1 B̆ K i (i = 1, 2, 3). The controller matrices
A K , B K , C K can then be determined using (5.31)–(5.33) and by construction they
will satisfy (5.28)–(5.30), (5.21) and (5.22). Therefore, we have obtained a solution
to Problem 5.1.
We now make the remark that due to the simplifying assumptions M = I and N =
I −YX introduced earlier, the solvability of the rank-constrained LMI problem above
is only sufficient for solvability of the coherent quantum LQG problem. However, it
is not difficult to see that we can remove these simplifying assumptions to produce a
rank-constrained LMI problem that is necessary and sufficient for the solvability of
Problem 5.1. It can be inspected that this can be done by (i) introducing additional
variables M, N , C K ≡ CM − , Ă ≡ AM − , X̆ ≡ XM − , (ii) additional lifting
variables W15 = N M , W16 = ĂM , W17 = X̆M , W18 = C K M and associated
constraints W15 −I +W9 = 0, A−W16 = 0, X−W17 = 0, N̆ −N Jn = 0, C−W18 = 0,
(iii) redefining W6 = N̆ C K , W7 = N̆ X̆ and W8 = Ă N̆ , and (iv) enlarging and
redefining Z as well as the set of constraints (5.36) and (5.37) accordingly. However,
this of course comes at the expense of solving a larger problem.
5.2.2 Numerically Solving the Rank-Constrained LMI

Problem
The rank-constrained LMI problem formulated above can be treated numerically

using iterative algorithms that attempt to directly search for a solution to the prob-
lem. Most of these algorithms are based on alternating projections (see [19] and
the references therein for details). Compared to converting the problem to a scalar
polynomial programming problem and employing LMI relaxations techniques based
on the theory of moments and the dual theory of sum of squares (SOS) polynomi-
als [20–23], these algorithms are much more efficient for the problem at hand in
terms of size, i.e., the number of variables and constraints that have to be considered.
However, alternating projection algorithms have a drawback that they do not neces-
sarily converge to a solution from arbitrary starting points, even if a solution to the
rank-constrained LMI problem exists.
Following [12], we will employ an algorithm that was developed in [19]. This
algorithm has been implemented in the freely available LMIRank MATLAB toolbox
[24]. This toolbox can be called and executed from within the Yalmip optimization
prototyping environment [25]. The algorithm belongs to the family of alternating
projection algorithms but is equipped with a built-in Newton step that could poten-
tially accelerate convergence. We will later employ this toolbox to numerically solve
some example coherent quantum LQG control problems.
The LMIRank algorithm needs an initial starting point. For a given γ > 0, as a
heuristic choice of starting point for the LMIRank solver, one can first solve (5.28)–
(5.30) to obtain A, B, C, X, Y, Q. Then set M = I2n and N = I − YX and calculate
B̆ K 1 , B̆ K 2 , B̆ K 3 , N̆ as well as the matrix-lifting variables W1 , . . . , W14 following the
definitions given in Sect. 5.2.1. Let
V0 = [ I2n A B̆ K1 B̆ K2 B̆ K3 C X Y N̆ W1

. . . W14 ] .
Then, a heuristic starting point for the LMIRank algorithm is Z = V0 V0 .
5.2.3 An Extension of the Numerical Procedure
We will now develop an extension of the rank-constrained LMI approach by for-

mulating a problem that is more natural for the coherent quantum LQG problem
but will involve the optimization of additional matrix variables. The starting point
is that for the LQG cost, it is the transfer function that matters rather than the sys-
tem matrices (A, B, C, D) (recall the discussion in Sect. 3.3). Application of the
symplectic similarity transformation (A, B, C, D) → (V AV −1 , V B, C V −1 , D) for
any invertible matrix V does not change the LQG cost. However, with this trans-
formation the new internal variable x̃ K = V x K satisfies the commutation relation
[x̃ K (t), x̃ K (t) ] = 2ı K , with K = V Jn V . Therefore, we can formulate a relaxed
optimization problem by introducing K as a new matrix variable with the constraint
that it is skew symmetric, invertible, and can be expressed as K = V Jn V for some
invertible matrix V . We thus arrive at the following generalized formulation of the
coherent quantum LQG problem:
Problem 5.2 Given a cost-bound parameter γ > 0, find real matrices A K , B K 1 ,

B K 2 , B K 3 , C K , and an invertible skew symmetric matrix K such that:
1. There exists a symmetric matrix P > 0, satisfying (5.19).
2. J∞ = Tr(C̃ P C̃ ) < γ.
3. The resulting controller is physically realizable. That is, A K , B K 1 , B K 2 , B K 3 , C K ,
and K satisfy (5.21) and (5.22) with Jn replaced by K .
The lemma below is straightforward and shows that solving Problem 5.2 also solves
Problem 5.1.
ˆ K solve
Lemma 5.1 Suppose that the matrices Â K , B̂ K 1 , B̂ K 2 , B̂ K 3 , Ĉ K , and
ˆ
Problem 5.2, and K = V Jn V for some real invertible matrix V . Then, the matrices
A K = V −1 Â K V ; B K i = V −1 B̂ K i i = 1, 2, 3; C K = Ĉ K V, (5.39)
solve Problem 5.1.

Proof Since ˆ K is real skew symmetric and invertible, we can find an invertible
matrix V such that ˆ K = V Jn V . Now, we have that

3
ˆK +
Â K ˆ K Â
K + B̂ K j Jn B̂ K j = 0, (5.40)
j=1
ˆ K Ĉ K Jn u .
B̂ K 1 = (5.41)
After substitution of (5.39) and ˆ K = V Jn V into (5.40) and (5.41), and some
algebraic manipulations, it is easily obtained that A K , B K , and C K satisfy

3
A K Jn + Jn A
K + B K j Jn B K j = 0, (5.42)
j=1
B K 1 = Jn C K Jn u . (5.43)
Since the LQG cost is invariant under a similarity transformation of the controller
state-space matrices, these matrices solve Problem 5.1, as claimed.
The following corollary is then immediate.
Corollary 5.1 Suppose that A K , B K 1 , B K 2 , B K 3 , C K solve the standard
cost-bounded LQG problem for a given γ > 0 (i.e., (5.19) and J∞ < γ are satisfied),
and there exists an invertible real skew symmetric 2n × 2n matrix Z satisfying

3
A K Z + Z A
K + B K j Jn B K j = 0, (5.44)
j=1
B K 1 = ZC K Jn u . (5.45)
Moreover, let V be a real invertible matrix such that V Z V = Jn . Then, the matrices
Â K , B̂ K 1 , B̂ K 2 , B̂ K 3 , Ĉ K given by:
Â K = V A K V −1 ; B̂ K i = S B K i i = 1, 2, 3;
Ĉ K = C K V −1 , (5.46)
solve Problem 5.1.
The corollary states that a solution A K , B K , C K to the standard LQG problem is also
a solution to the coherent quantum LQG problem (Problem 5.1) if and only if there
exists a matrix Z satisfying the equalities stated in the corollary.
Problem 5.2 can also be treated using the rank-constrained LMI procedure devel-
oped in Sect. 5.2.1. Let us do this again under the simplifying assumptions M = I
and N = I − YX (that can easily be removed if desired). Introduce the additional
variable K and substitute A K , B K i (i = 1, 2, 3), C K with, respectively, Â K , B̂ K i
(i = 1, 2, 3), Ĉ K (see Lemma 5.1). We also redefine N̆ = N K , x = (x1 , . . . , x10 ),
and Z to be a real symmetric matrix of dimension 50n × 50n, and replace (5.36)
with the following constraints:
⎫
Z ≥0 Zx8 ,1 + Zx9 ,x10 = 0 ⎪
⎪
⎪
⎪
Z0,0 − I2n = 0 Zv4 ,1 − Zx7 ,1 B = 0 ⎪
⎪
⎪
⎪
Z1,x6 − Zx6 ,1 = 0 Zv6 ,1 − Zx8 ,x5 = 0 ⎪
⎪
⎪
⎪
Z1,x7 − Zx7 ,1 = 0 Zv7 ,1 − Zx8 ,x6 = 0 ⎪
⎪
⎪
⎪
Zv9 ,1 − Zx7 ,x6 = 0 Zv8 ,1 − Zx1 ,x8 = 0 ⎪
⎪
⎪
⎪
Zv10 ,1 − Zv4 ,v6 = 0 Zx10 ,1 + Z1,x10 = 0 ⎬
Zv11 ,1 − Zv5 ,v7 = 0 Zv13 ,1 − Zv2 ,x3 = 0 (5.47)
⎪
⎪
Zv12 ,1 − Zv1 ,x2 = 0 Zv14 ,1 − Zv3 ,x4 = 0 ⎪
⎪
⎪
⎪
Zv1 ,1 − Zx2 ,1 Jn = 0 ⎪
⎪
⎪
⎪
Zv2 ,1 − Zx3 ,1 Jn = 0 ⎪
⎪
⎪
⎪
Zx9 ,1 − I2n + Zv9 ,1 = 0 ⎪
⎪
⎪
⎪
Zv3 ,1 − Zx4 ,1 Jn = 0 ⎪
⎪
⎭
Zv5 ,1 − Zx4 ,1 C − Zx7 ,1 A = 0
In the above, Zx10 ,1 + Z1,x10 = 0 corresponds to the constraint that K = − K,

while Zx8 ,1 + Zx9 ,x10 = 0 corresponds to the constraint N̆ = N K . The rest of the
constraints from Sect. 5.2.1, (5.37), and rank(Z ) ≤ 2n are unchanged. Notice that
now the variable N is an independent variable, whereas in Sect. 5.2.1 N was linearly
related to N̆ by the identity N = N̆ Jn . As the last step, we formulate a heuristic for
the initial guess 0K for K . Possible choices for 0K include 0K = 0 or 0K = Jn .
When 0K has been selected, set N = I − YX, N̆ 0 = N 0K , and
V0 = [ I A B̆ K1 B̆ K2 B̆ K3 C X Y ( N̆ 0 ) N

(0K ) W1 . . . W14 ] .
The LMIRank routine can then be executed with Z = V0 V0 as the starting point.
We conclude the discussion by noting that the requirement for K to be invertible
has not been built into the above numerical procedure. Therefore, the procedure could,
in principle, potentially return a solution with a singular K (up to errors due to finite
machine precision). Although such a solution can be interpreted as a controller with
mixed quantum-classical dynamics (see the discussion in Sect. 5.2.6), an additional
constraint can in principle be added to enforce the invertibility condition on K .
Next, we present some numerical examples of coherent quantum LQG design using
the methods presented here.
5.2.4 Quantum LQG Control Design Examples
We will now apply the approach Sect. 5.2.1 to find a cost-bounded quantum LQG
controller that asymptotically stabilizes a marginally stable quantum plant. The
numerical implementation was carried out in MATLAB using the Yalmip prototyp-
ing environment and LMIRank. The semidefinite program solver used for LMIRank
was SeDuMi Version 1.1 Release 3; see [26]. For comparison, we also compute a
classical LQG controller that controls the plant by driving it with a modulated optical
beam.
The quantum plant under consideration is a one degree of freedom linear quantum
system with Hamiltonian matrix R and coupling matrix K given by
⎡√ ⎤
κ1 0
0 √
R = 1/2 , K = ⎣ κ2 0 ⎦ ,
0 √
κ3 0
with = 0.1 and k1 = k2 = k3 = 10−2 . Its evolution is governed by the linear

QSDE

0 0 0√
dx = xdt + du +
− 0 0 −2 k1

0 0√ 0 0√ dw1
,
0 −2 k2 0 −2 k3 dw2
√
2 k2 0
dy = xdt + dw1 . (5.48)
0 0
Notice that for the given coupling matrix K , the quantum noise inputs couple only
with the position operator of the system. Such a coupling is typically sought in
schemes for quantum non-demolition continuous measurement of position. This
yields a marginally stable plant having two mutually conjugate eigenvalues on the
imaginary axis.
5.2.4.1 Quantum LQG Controller Design Example I
Let us seek another linear quantum system as a quantum LQG controller to asymptot-
ically stabilize the given plant. We choose the performance variable to be z = x +βu ;
i.e., C z = I2×2 = Dz . Selecting γ = 5.75 and solving Problem 5.1 numerically fol-
lowing Sects. 5.2.1 and 5.2.2, LMIRank yields the physically realizable controller
(numerical results are presented only up to four digits behind the decimal point),

−2.3907 0.8420 −0.3029 0.5042
d xK = x dt + dw K 1
−5.5518 1.9380 K −0.6603 1.0819

0.0241 −0.0471 3.3626 2.1470
+10−10 dw K 2 + dy,
0.0576 −0.1136 7.6699 5.0302

−1.0819 0.5042
du(t) = xdt + dw K 1 . (5.49)
−0.6603 0.3029
This controller asymptotically stabilizes the closed-loop system, and the closed-loop
LQG cost achieved is J∞ = 5.7382.
We can see that elements of B K 2 are very small, of the order 10−10 . We note that
B K 2 is the coefficient for the quantum noise w K 2 , which only enters in the controller
and did not originate from the plant. This additional controller noise contributes to
the LQG cost, but since the control goal is to bound this cost the algorithm finds a
controller with a small B K 2 term in order to minimize the impact of the variance of
w K 2 on the LQG cost. Up to the numerical precision of MATLAB, the numerical
results give

3
A K J + JA
K + B K i JB Ki
k=1

−13 0 0.2896
= 10 , (5.50)
−0.2896 0
while removing the term B K 2 JB K2 (i.e., setting B K 2 = 0) returns an identical numer-
ical result on the right-hand side of (5.50). This indicates that the contribution of B K 2
to (5.50) is less than the numerical precision of MATLAB, and to obtain a simpler
controller we may simply set B K 2 = 0.
5.2.4.2 Classical LQG Controller Design
Now that we have obtained a fully quantum controller, we can pose a natural ques-
tion: Does this controller offer any improvement over a classical measurement-based
feedback controller driven by continuous measurements of a quadrature of the plant
output y (e.g., by homodyne detection as discussed in Sect. 4.4.1)? To address this
question, we consider performing continuous measurements of one quadrature of
y (in this case, its first element). This entails replacing the output y in (5.48) with
another output y (another classical signal) defined as:
√
y = [ 2 k2 0 ]xdt + [ 1 0 ]dw1 .
We now wish to seek a classical controller of the form:
d x K = A K x K dt + B K dy ;
βu = C K x K (5.51)
whose output will modulate an optical beam and produce the control signal u:
du = βu dt + dw K 1
= C K x K dt + dw K 1 . (5.52)
To do this, we can apply the standard LQG machinery (with C z = Dz = I2 as before)

to the following modified plant (to account for the presence of the noise w K 1 in the
controller output u, recall the discussion in Sect. 5.2.1) with βu being viewed as the
“control signal”:

0 0 0√
dx = xdt + β dt
− 0 0 −2 k1 u
⎡ ⎤
dw K 1
0 0√ 0 0√ 0 0√ ⎣ dw1 ⎦ ,
+
0 −2 k1 0 −2 k2 0 −2 k3
dw2
√
dy = [ 2 k2 0 ]xdt + [ 1 0 ]dw1 .
The optimal classical controller was then found to be:

−0.0658 0.1 0.3291
d xK = x dt + dy ;
−0.1217 −0.2 K 0.1083

−1 0
du = x dt + dw K 1 ,
0 1 K
and the optimal LQG cost achieved is J∞ = 4.8468.

We find that the cost obtained by this classical controller is actually lower than
the one achieved by the quantum LQG controller, compare J∞ = 4.8468 for the
former with J∞ = 5.7382 for the latter. Moreover, this may not yet be the best
performance that can be achieved by a classical linear controller because there is still
the freedom to select arbitrary rotated quadratures of y and performing heterodyne
detection on both quadratures. This can be implemented by passing y through a phase
shifter and mixing the phase shifter output with an additional vacuum noise at a beam
splitter [27, 28], followed by homodyne detection of the amplitude quadrature of one
beam splitter output and phase quadrature of the other output. This supplies noisy
information about both quadratures of y the controller. To describe this measurement
process, denote the vacuum noise going into the beamsplitter by w0 = (w0,1 , w0,2 )
(in quadrature notation). The beam splitter divides power in its incoming signals
between its two output ports according to the ratio 2 : δ 2 , where 2 + δ 2 = 1,
0 ≤ , δ ≤ 1. The classical measurement signal y at the output of this measurement
scheme is given by,

0 cos(θ) − sin(θ) δ0
dy = dy + dw0 ,
0 −δ sin(θ) cos(θ) 0
where θ is the phase shift introduced by the phase shifter.

We now compute classical LQG controllers of the form (5.51) and (5.52) with y
replaced by y for θ taking values in (−π, π] and in [0, 1]. Notice that measurement
of y is a special case of measurement of y corresponding to setting θ = 0 and = 1.
A surface plot of the LQG cost J∞ achieved by these controllers versus the value
of θ and is shown in Fig. 5.3. The minimum value of J∞ over θ and is about
4.8468 achieved at θ ≈ 0 and ≈ 1, so the general measurement scheme with y
for this example offers no noticeable advantage over just measuring y . In fact, it
is reasonable to conjecture that measuring y is already optimal for this example.1
Thus, it is now natural to ask, does there exist at all a fully quantum LQG controller
that can beat the LQG cost achievable by a classical controller for this example? It
turns out that the answer to the question is affirmative and a fully quantum controller
that beats all classical LQG controllers can be found using the extended numerical
procedure developed in Sect. 5.2.3. This will be shown next.
5.2.5 Quantum LQG Controller Design Example II
We now turn our attention to numerically solving Problem 5.2 to find a fully quantum
controller that achieves a cost J∞ < 4.8468. We will do this by applying the extended
numerical scheme in Sect. 5.2.3 with γ = 5 and 0K = 02×2 . The LMIRank routine
returns the following solution:

0 −0.1820 −0.2125 0.0666
K = ; Â K = ;
0.1820 0 −0.3789 0.0257

0.0642 −0.0547
B̂ K 1 = ;
0.0480 −0.2556
1 We note that [12] suggested that a lower LQG cost of 4.444 can be attained by a classical linear
controller by measuring y with the setting θ = 0 and ≈ 0.715. This is incorrect and the
discrepancy arose out of a coding error therein. However, this error does not negate the findings
of [12]. In fact, it only strengthens the conclusion that there can exist a coherent quantum LQG
controller which outperforms all classical LQG controllers for the same cost function and actuation
structure.
Fig. 5.3 Plot of J∞ versus θ and α

−0.0468 0.0255
B̂ K 2 = 10−10 ;
−0.1160 0.0114

0.3522 −0.0215 −1.4044 0.3008
B̂ K 3 = ; Ĉ K = ,
0.4393 −0.0842 −0.2639 0.3526

√ 01
with the cost J∞ = 4.1793. Since K = V J V with V = 0.182 , from
10
Lemma 5.1 we find the matrices A K , B K i (i = 1, 2, 3), and C K of the controller
solving Problem 5.1 are

0.0257 −0.3789 0.1126 −0.5992
AK = ; BK 1 = ;
0.0666 −0.2125 0.1504 −0.1283

−0.2721 0.0272
BK 2 = 10−10 ;
−0.1096 0.0601

1.0297 −0.1974 0.1283 −0.5992
BK 3 = ; CK = . (5.53)
0.8255 −0.0503 0.1504 −0.1126
As with design example I, we again find that B K 2 has negligibly small entries. So
we may similarly set B K 2 = 0 to obtain a simpler controller. That we could find
a fully quantum controller that outperforms all classical LQG controllers, compare
J∞ = 4.1793 in the former against J∞ ≈ 4.8468 in the latter, indicates the intriguing
potential for a coherent feedback controller to offer performance advantage over
classical controllers, to be discussed further below.
The coherent quantum LQG formalism has been adopted by Hamerly and Mabuchi
in [6] to compare, on an equal footing, the ability of classical linear controllers and
linear coherent feedback controllers to minimize the LQG cost of an optomechanical
system consisting of two optical cavities, each with one movable mirror. The move-
able mirrors move together as a single mechanical component, and their motion is
coupled to a thermal bath with noise power kn . The LQG cost corresponds to cool-
ing the mechanical component, that is, minimization of the number of phonons in
this component. Equal footing here is in the sense that the signals used to drive the
controllers were obtained from the same output port of the plant and the controller
in turn drives/actuates the same input port of the plant. Several different types of
classical and coherent feedback controllers were compared and, similar to what has
been demonstrated in the examples above, it was found that in the regime of small
noise power kn , where quantum effects are significant, a linear coherent feedback
controller can be found that outperformed the linear classical controllers. The work
[6] attributes this advantage, for the system considered therein, to the ability of the
controller to simultaneously process non-commuting quadratures of the output field.
A deeper analysis of why a linear coherent feedback controller can beat a linear
measurement feedback controller was given more recently in [29].
The coherent feedback LQG control theory was originally developed in [12]
in a more general setting that allowed controllers to possibly have mixed quantum-
classical dynamics. The treatment here has been restricted to controllers that are linear
quantum systems with no classical dynamics. The numerical approach described
herein can be straightforwardly adapted to this more general setting.
5.3 Coherent Feedback H ∞ Control
Control laws are usually designed based on a mathematical model for the dynamics
of the plant to be controlled. These models are necessarily imperfect. There can be
many sources of imperfections, for instance, uncertainties about the model parameter
values, and unmodelled dynamics. Such uncertainties or errors in the assumed model
may lead to instability in the closed feedback loop when the feedback controller is
designed based on the imperfect model. A natural question that can then be asked is,
how can controllers be designed so that they can maintain closed-loop stability and
acceptable control performance in the presence of uncertainties and disturbances?
One framework for achieving such a robust design of controllers is H ∞ control
synthesis. This method was initially developed for classical linear time-invariant
systems based on minimizing the H ∞ norm of the closed-loop transfer function
and was solved using the state-space formulation in the landmark paper [30]. A
first principles approach to the solution of the so-called non-singular H ∞ control
problem was presented in [31], and the LMI approach can be found in [16, 32].
5.3 Coherent Feedback H ∞ Control 181
The H ∞ methodology has subsequently been extended to nonlinear systems, see,

e.g., [33].
The adaptation of the H ∞ control method to the quantum setting was presented in
[34] as a framework for designing robust linear quantum controllers that can tolerate
unwanted disturbances/noise and modeling uncertainties. The approach builds on
a formulation of H ∞ control based upon powerful concepts from modern control
theory, including dissipation inequalities for open classical dynamical systems that
originated in the work of Willems [35], and generalizations of this idea to the stochas-
tic setting, classical [36] and quantum [37]. In this section, we will give an exposition
of the H ∞ control approach to coherent feedback quantum control, following [34].
We begin by formulating the dissipation properties of linear quantum systems.
5.3.1 Dissipation Properties
We now recall some important dissipation properties that are frequently exploited
in control engineering, suitably adapted to the quantum setting. These properties
quantify the effect of disturbance inputs on energy transfers and stability. In particular,
a quantum adaptation of the Strict Bounded Real Lemma (Corollary 5.3) will be
presented and used for quantum H ∞ controller synthesis. We will be working with
linear quantum systems of the form
d x(t) = Ax(t)dt + [ B G ][ dw(t) dv(t) ] ,

dz(t) = C x(t)dt + [ D H ][ dw(t) dv(t) ] . (5.54)
The input to the quantum system has two components w(t) and v(t); w(t) represents
disturbance signals, while v represents additional noise sources. The input w(t) is
composed of two components,
dw = βw dt + d w̃. (5.55)
Here βw is an adapted process that is a component of dw which could, say, be

passed to the system as input from another quantum system, while w̃ is a vector of
non-commutative quantum Wiener noise processes. The vector v is also a vector of
non-commutative quantum Wiener noise processes and is independent of w̃(t). In
the treatment of H ∞ control in this chapter, the noises are not necessarily vacuum
noises but may be arbitrary zero-mean Gaussian noise as elaborated in Chap. 2. In
(5.54), we do not write down the output equation, as it is not yet required in this
section, and z(t) is a new vector that is a linear combination of x(s), w(s), and
v(s), 0 ≤ s ≤ t. As in the coherent feedback LQG problem, the vector z(t) is not
necessarily a physical observable but it will be relevant for formulating the control
objective in the ensuing H ∞ control design.
In the following, we will identify a Gaussian state of oscillators or bosonic fields

with its density operator ρ. We will also use the notation ρ, X = Tr(ρX ) = X .
Definition 5.1 Given an operator-valued quadratic form

x
r (x, βw ) = [x βw ]R
βw
where
R11 R12
R=
R
12 R22
is a given real symmetric matrix, the system (5.54) is said to be dissipative with supply
rate r (x, βw ) if there exists a positive operator-valued quadratic form V (x) = x X x
(where X is a real positive definite symmetric matrix) and a constant λ > 0 such
that
t
V (x(t)) + r (x(s), βw (s)) ds ≤ V (x(0)) + λt, ∀t > 0, (5.56)
0
for all initial Gaussian states ρ of the system.

The system (5.54) is said to be strictly dissipative if there exists a constant > 0
such that inequality (5.56) holds with the matrix R replaced by the matrix R + I .
In the definition above, the term V (x(t)) can be interpreted as an abstract notion
of internal “energy” for the linear quantum system at time t. Meanwhile, the term
r (x(t), βw (t)) generalizes to quantum systems the notion of abstract energy flow
in and out of a system at time t. These notions have been extensively used in the
stability analysis of classical deterministic dynamical systems [35, 38]. The dissi-
pation inequality (5.56) was introduced for classical stochastic systems in [36] and
subsequently adapted to linear quantum networks in [37].
The following result relates the dissipation properties to certain LMIs, with the
proof given in the appendix to this chapter.
Theorem 5.2 Given a quadratic form r (x, βw ) defined as above, then the quantum
stochastic system (5.54) is dissipative with supply rate r (x, βw ) if and only if there
exists a real positive definite symmetric matrix X such that the following LMI is
satisfied:
A X + X A + R11 R12 + X B
≤ 0. (5.57)
B X + R12 R22
Furthermore, the system is strictly dissipative if and only if there exists a real positive
definite symmetric matrix X such that the following LMI is satisfied:

A X + X A + R11 R12 + X B
< 0. (5.58)
B X + R
12 R22
Moreover, if either of (5.57) or (5.58) holds, then the required constant λ ≥ 0 can
be chosen as λ = λ0 , where

B
λ0 = Tr X B G F . (5.59)
G
Here, the matrix F is defined by via the relation:

dw
Fdt = dw dv . (5.60)
dv
Remark 5.1 We refer to F defined in (5.60) as an Itō matrix. It can be defined

more generally for any 2m × 1 vector ν(t) of adapted quantum processes such that
dν(t)dν(t) = Fν dt, where Fν would be the Itō matrix for ν(t). When Fν = I2m +
ıJm , then we say that F is canonical. This means that dν(t) = βν (t)dt + Sν (t)dwν (t)
with βν , Sν adapted and the latter satisfying Sν (t)Sν (t) = I2m for each t, and wν
a vacuum quantum Wiener noise vector. The notion of canonical and non-canonical
noises was introduced in [34] to treat rather general linear non-commutative systems
that can have mixed classical and quantum linear dynamics, driven by possibly non-
vacuum Gaussian input fields. Here, we do not consider systems with mixed quantum-
classical dynamics, see Sect. 5.4 for an additional discussion.
The corollaries below state the application of the theorem above to a special case of
the matrix R, defined in terms of error output operator βz (t) = C x(t) + Dβw (t).
This operator is the adapted dt component of the performance variable z(t).
Definition 5.2 The quantum stochastic system (5.54) is said to be Bounded Real
with disturbance attenuation g if the system (5.54) is dissipative with supply rate
r (x, βw ) = βz βz − g 2 βw βw

C C C D x
= [x βw ] .
DC D D − g2 I βw
Also, the quantum stochastic system (5.54) is said to be Strictly Bounded Real with
disturbance attenuation g if the system (5.54) is strictly dissipative with this supply
rate.
With the above definition in hand, we can state following corollary of Theorem 5.2
(see [39] for the corresponding classical result).
Corollary 5.2 The quantum stochastic system (5.54) is bounded real with distur-
bance attenuation g if and only if there exists a positive definite symmetric matrix
X ∈ R2n×2n (n denoting the degrees of freedom of (5.54)) such that the following
matrix inequality is satisfied:

A X + X A + C C C D + X B
≤ 0.
B X + DC D D − g2 I
Furthermore, the quantum stochastic system is strictly bounded real with disturbance
attenuation g if and only if there exists a positive definite symmetric matrix X ∈
R2n×2n such that the following matrix inequality is satisfied:

A X + X A + C C C D + X B
< 0.
B X + DC D D − g2 I
Moreover, in both cases the required constant λ ≥ 0 can be chosen as λ = λ0 , where

λ0 is defined by (5.59).
We can now combine the corollary with the standard Strict Bounded Real Lemma
from modern control theory (e.g., see [31, 40]) to arrive at the next corollary. Note
the terminology used that a square matrix A is said to be stable or is a stability matrix
if it is Hurwitz (all its eigenvalues have a negative real part).
Corollary 5.3 The following statements are equivalent

(i) The quantum stochastic system (5.54) is strictly bounded real with disturbance
attenuation g.
(ii) A is a stable matrix, and C(s I − A)−1 B + D∞ < g.
(iii) g 2 I − D D > 0, and there exists a positive definite matrix X̃ > 0 such that
A X̃ + X̃ A + C C + ( X̃ B + C D)(g 2 I − D D)−1 (B X̃ + D C) < 0.
(iv) g 2 I − D D > 0 and the algebraic Riccati equation
A X + X A + C C + (X B + C D)(g 2 I − D D)−1 (B X + D C) = 0
has a stabilizing solution X ≥ 0.

Furthermore, if these statements hold then X < X̃ .
5.3.2 H ∞ Controller Synthesis
We will now consider the problem of H ∞ controller design for linear quantum
systems. We begin by defining the controller and closed-loop plant–controller system
in Sect. 5.3.2.1. This is followed in Sect. 5.3.2.3 by the application of the Strict
Bounded Real Lemma to the closed-loop system to obtain the main results. Physical
realizability of the resulting linear quantum controllers is addressed in Sect. 5.3.2.4.
5.3.2.1 The Closed-Loop Plant–Controller System
We shall consider linear quantum systems in the real quadrature form (2.5). Systems
of this form will be the prototype for the interconnection of components in the
quantum control system. Let us now describe the plant and controller models and
the closed-loop system formed by the plant and controller.
Plants are taken to be linear quantum systems of the following form,
d x(t) = Ax(t)dt + [ B0 B1 B2 ][ dv(t) dw(t) du(t) ] ;

x(0) = x0 ,
dz(t) = C1 x(t)dt + D12 du(t),
dy(t) = C2 x(t)dt + [ D20 D21 02n y ×2n u ]
×[ dv(t) dw(t) du(t) ] . (5.61)
Here, x0 is a vector of the position and momentum operator pairs of the single-mode
oscillators forming the plant. The input w(t) represents a disturbance signal of the
form (5.55). The signal u(t) is a control input of the form
du(t) = βu (t)dt + d ũ(t), (5.62)
where ũ(t) is the noise part of u(t), and βu (t) is a vector of adapted, self-adjoint
processes. Also, dv(t) represents any additional quantum noise inputs that may enter
the plant. The vectors v(t), w̃(t), and ũ(t) are quantum noises with Itō matrices Fv ,
Fw̃ , and Fũ which are all non-negative Hermitian (see Remark 5.1).
Controllers are taken to be linear quantum systems of the form
d x K (t) = A K x K (t)dt + [ B K 1 B K ][ dv K (t) dy(t) ] ;

x K (0) = x K ,0 ,
du(t) = C K x K (t)dt + [ B K 0 02n u ×2n y ][ dv K (t) dy(t) ] ,
(5.63)
where x K (t) = [ x K 1 (t) . . . x K n K (t) ] is a vector of self-adjoint controller variables.

The vector v K (t) = [ v K 1 (t) . . . v K K v (t) ] is a quantum noise vector consisting of
non-commutative Wiener processes in the vacuum state, and x K ,0 is a vector of the
position and momentum operator pairs of the single-mode oscillators forming the
controller.
The closed-loop system can be formed by making the identification βu (t) =
C K x K (t) and interconnecting (5.61) and (5.63) to give

A B2 C K
d x̃(t) = x̃(t)dt +
B K C2 A K

B0 B2 B K 0 dv(t) B1
+ dw(t),
B K D20 B K 1 dv K (t) B K D21

dv(t)
dz(t) = C1 D12 C K x̃(t)dt + 0 D12 B K 0 , (5.64)
dv K (t)
where x̃(t) = [ x(t) x K (t) ] . We can write the closed-loop system in the form
d x̃(t) = Ã x̃(t)dt + B̃dw(t) + G̃d ṽ(t)

dw(t)
= Ã x̃(t)dt + B̃ G̃ ,
d ṽ(t)
dz(t) = C̃ x̃(t)dt + H̃ d ṽ(t)

dw(t)
= C̃ x̃(t)dt + 0 H̃ , (5.65)
d ṽ(t)
where

v(t) A B2 C K B1
ṽ(t) = ; Ã = ; B̃ = ;
v K (t) B K C2 A K B K D21

B0 B2 B K 0
G̃ = ; C̃ = C1 D12 C K ; H̃ = 0 D12 B K 0 .
B K D20 B K 1
The closed-loop system (5.65) is therefore a linear quantum system of the quadrature
form (2.5).
5.3.2.2 H ∞ Control Objective
In the H ∞ controller synthesis problem, we have a disturbance attenuation parameter

g > 0 as a design parameter that represents the minimum accepted attenuation of the
disturbance signal βw in the performance variable z(t), in a sense that will be defined
below. Typically, it is desired that the parameter g be as small as possible. Given this
parameter, the goal of H ∞ controller synthesis is to find a controller (5.63) such that
the closed-loop system (5.65) satisfies:
t
βz (s) βz (s) + x̃(s) x̃(s) ds
0
t
≤ (g − )
2
βw (s) βw (s) ds + μ1 + μ2 t, ∀t > 0 (5.66)
0
for some real constants , μ1 , μ2 > 0. Here βz (t) = C̃η(t) is the dt component of the
performance variable z(t) in the closed-loop system (5.65). The design criteria can
be interpreted as the controller bounding the “energy” transfer from the disturbance
signal βw (t) to the error signal βz (t) as a component of z(t).
The closed-loop system (5.65) will meet the objective (5.66) if it is designed to
be strictly bounded real with disturbance attenuation g. This follows directly from
Definition 5.2. According that definition, (5.65) will be strictly bounded real with
disturbance attenuation g if and only if there exists a real positive definite symmetric
matrix X and a constant λ > 0 such that
t
x̃(t) X x̃(t) + βz (s) βz (s) − g 2 βw (s) βw (s) + x̃(s) x̃(s)
0
+ βw (s) βw (s) ds ≤ x̃(0) X x̃(0) + λt, ∀t > 0. (5.67)
Inequality (5.66) now follows with μ1 = x̃(0) X x̃(0) and μ2 = λ.

In the following, we will give necessary and sufficient conditions for the existence
of a controller achieving the H ∞ control objective (5.66) for a given g. Explicit
formulas for the controller matrices A K , B K , and C K will be obtained. These results
mirror the corresponding well-known results for classical linear systems as given in,
e.g., [30, 31].
5.3.2.3 Necessary and Sufficient Conditions
In order to establish necessary and sufficient conditions for the existence of an H ∞

controller, we need to impose some assumptions on the plant (5.61). They are the
following.
Assumption 5.1

1. D12 D12 = E 1 > 0.

2. D21 D21 = E2 > 0.
A − jω I B2
3. The matrix is full rank for all ω ≥ 0.
C1 D12

A − jω I B1
4. The matrix is full rank for all ω ≥ 0.
C2 D21
The results will be stated in terms of the following pair of algebraic Riccati equations:
(A − B2 E 1−1 D12

C1 ) X + X (A − B2 E 1−1 D12

C1 )
+X (B1 B1 − g 2 B2 E 1−1 B2 )X
+g −2 C1 (I − D12 E 1−1 D12

)C1 = 0; (5.68)
−1 −1
(A − B1 D21 E 2 C2 )Y + Y (A − B1 D21 E 2 C2 )
+Y (g −2 C1 C1 − C2 E 2−1 C2 )Y
−1
+B1 (I − D21 E 2 D21 )B1 = 0. (5.69)
We assume that the solutions to these Riccati equations satisfy the following.
Assumption 5.2
(i) A − B2 E 1−1 D12

C1 + (B1 B1 − g 2 B2 E 1−1 B2 )X is a stability matrix.
−1
(ii) A − B1 D21 E 2 C2 + Y (g −2 C1 C1 − C2 E 2−1 C2 ) is a stability matrix.
(iii) The matrix X Y has a spectral radius strictly less than one.
We are now ready to state the main result on H ∞ controller synthesis.
Theorem 5.3 Necessity. Consider the system (5.61) and suppose that Assumption
5.1 is satisfied. If there exists a controller of the form (5.63) such that the resulting
closed-loop system (5.65) is strictly bounded real with disturbance attenuation g,
then the Riccati equations (5.68) and (5.69) will have stabilizing solutions X ≥ 0
and Y ≥ 0 satisfying Assumption 5.2.
Sufficiency. Suppose the Riccati equations (5.68) and (5.69) have stabilizing solu-
tions X ≥ 0 and Y ≥ 0 satisfying Assumption 5.2. If the controller (6.53) is such
that the matrices A K , B K , C K are constructed as
A K = A + B2 C K − B K C2 + (B1 − B K D21 )B1 X,

B K = (I − Y X )−1 (Y C2 + B1 D21

)E 2−1 ,
C K = −E 1−1 (g 2 B2 X + D12

C1 ), (5.70)
then the resulting closed-loop system (5.65) will be strictly bounded real with dis-
turbance attenuation g. Also the constant λ ≥ 0 in Definition 5.1 can be chosen
as in (5.59), λ = λ0 with B, G, F replaced by the corresponding matrices of the
closed-loop system.
Proof Using the Strict Bounded Real Lemma Corollary 5.3, the theorem follows
directly from the corresponding classical H ∞ result; e.g., see [10, 31, 41].
Notice that Theorem 5.3 does not specify the remaining controller matrices B K 0 ,
B K 1 , and the noise vector v K . This is because they are free as far as the H ∞ objective
is concerned. However, they will play a role when considering physical realizability
of the resulting controllers.
5.3.2.4 Physical Realization of Controllers
The matrices A K , B K , and C K resulting from the H ∞ controller synthesis of

Theorem 5.3 do not in general define or completely specify a physically realiz-
able linear quantum system. Fortunately, due to the freedom of choosing B K 0 , B K 1 ,
and adding the noise v K without affecting the H ∞ control objective, it will be shown
in the next theorem that we can always choose these remaining matrices to yield a
physically realizable H ∞ controller.

Theorem 5.4 Assume Fy = D20 Fv D20 + D21 Fw D21 is canonical (Remark 5.1). Let
{A K , B K , C K } be an arbitrary triple (such as given by (5.70)). Then, there exists
controller parameters B K 0 , B K 1 , and the controller noise v K such that the controller
(5.63) is physically realizable.
The theorem follows directly from the next lemma.
Lemma 5.2 Let Fy be canonical and {A K , B K , C K } be such that A K ∈ R2n K ×2n K ,

B K ∈ R2n K ×2n y , and C K ∈ R2n u ×2n K . Then, there exist an integer n v K ≥ n u and
B K 1 ∈ R2n K ×2n v K such that the system (5.63) is physically realizable with (taking K
to be the coupling matrix in the coupling operator L)
B K 0 = [ I2n u 02n u ×2(n v K −n u ) ],

R = 1/2(Z + Z ), (5.71)

B K 1 = B K 1,1 B K 1,2 , (5.72)

I
K = 21 C K Pnu
Kb1
Kb2 , (5.73)
ıI
B K 1,1 = Jn K C K Jn u , (5.74)

Kb2 = −ı In y ×n y 0n y ×n y Pn y (I2n y ⊗ M)B K Jn K , (5.75)
∗
B K 1,2 = 2ıJn K −Kb1 Kb1 Pn v K −n u diagn v −n u (M), (5.76)
K
where Z = − 21 Jn K A K and n v K ≥ n u + 1, Pm is a permutation matrix as defined in

Sect. 2.3.2, and
1 ı
M = /21 .
1 −ı
Here, Kb1 is any complex (n v K − n u ) × 2n K matrix such that

∗

Kb1 Kb1 = W + ı 1/2(Z − Z )

0 I
−1/4C K PNu ∗
Pn u C K − (Kb2 Kb2 ) , (5.77)
−I 0
where W is any real symmetric 2n K × 2n K matrix such that the right-hand side of
(5.77) is non-negative definite.
The proof of Lemma 5.2 is given in the appendix, Sect. 5.4.
Remark 5.2 Note that for (5.63) and (5.61) to be compatible, it is required that the
Itō matrices of u and v K satisfy the condition
Fu = B K 0 Fv K B K0 . (5.78)
5.3.2.5 Robust Stability
We will now develop further results that show how the H ∞ synthesis results of
Sect. 5.3.2 can be used to guarantee stability robustness in the presence of real para-
meter uncertainties. We assume that the true closed-loop quantum system corre-
sponding to the system (5.65) is of the form
d x̃(t) = Ā x̃(t)dt + G̃d ṽ(t), (5.79)
where Ā = Ã + B̃C̃, and is a constant but unknown uncertainty matrix satisfying
1
≤ I. (5.80)
g2
The stability robustness will be a consequence of the fact that the H ∞ control design
leads to a closed-loop quantum system which is strictly bounded real with disturbance
attenuation g. We start with the following definition.
Definition 5.3 The closed-loop quantum system (5.79) is said to be mean square
stable if there exists a real positive definite matrix X > 0 and a constant λ > 0 such
that
t

x̃(t) X x̃(t) + x̃(s) x̃(s) ds ≤ x̃(0) X x̃(0) + λt ∀t > 0,
0
for all Gaussian initial states ρ of the plant and controller.
The next lemma and theorem connect the robust stability of the above system to its
H ∞ properties. The proofs for these results can be found in the appendix.
Lemma 5.3 The quantum system (5.79) is mean square stable if and only if the
matrix Ā is a stable matrix.
Proof We first observe that the system (5.79) is mean square stable if and only if it is
dissipative with a supply rate defined by the matrix R = diag(I, 0). It follows from
Theorem 5.2 that the system (5.79) is mean square stable if and only if there exists
a real positive definite symmetric matrix X such that Ā X + X Ā + I ≤ 0. Hence,
using a standard Lyapunov result (e.g., see [10]), it follows that the system (5.79) is
mean square stable if and only if the matrix Ā is asymptotically stable.
Theorem 5.5 If the closed-loop quantum system (5.65) is strictly bounded real with
disturbance attenuation g, then the true closed loop system (5.79) is mean square
stable for all satisfying (5.80).
Proof It follows from Corollary 5.3 that the closed-loop quantum system (5.65) is
strictly bounded real with disturbance attenuation g, then Ã is a stable matrix and
C̃(s I − Ã)−1 B̃ + D̃∞ < g. From this, it follows using the standard small gain
theorem (e.g., see Theorem 9.1 on page 212 of [10]) that the matrix Ā = Ã + B̃C̃
is stable for all satisfying (5.80). Hence using Lemma 5.3, it follows that the true
closed-loop system (5.79) is mean square stable for all satisfying (5.80).
5.3.3 H ∞ Synthesis in Quantum Optics
We will now illustrate some H ∞ controller designs for a quantum optical plant in the
form of an optical cavity. The type of control discussed in the first example has been
experimentally demonstrated in [42] in bulk quantum optics and in [43] in silicon
photonics, in the context of broadband attenuation of a coherent light that is injected
into one port of an optical cavity at the output of another port of the cavity.
5.3.3.1 Quantum Controller Synthesis
The plant that will be considered is an optical cavity which is coupled to three optical
inputs v, w, and u, as shown in Fig. 5.4. As our control objective, we wish to attenuate
the effect of the signal w as a disturbance input to the system on the output z. In
physical terms, this means to dim the light emerging from z resulting from light
shone in at w across a broad range of frequencies. Thus, most of the optical power in
w will be redirected to output fields other than z; again, see [42, 43] for the details
of some physical realizations of this example.
Fig. 5.4 An optical cavity

(plant). Reprinted, with w
permission, from [34]
© 2008 IEEE v y
κ1 = 2.6 κ2 = 0.2
a
κ3 = 0.2
z u
The dynamics of this optical cavity system is given in the complex mode form by:
γ √ √ √
da(t) = − a(t)dt − κ1 dA1 (t) − κ2 dA2 (t) κ3 dA3 (t);
2
a(0) = a,
γ √ √ √
da ∗ (t) = − a ∗ (t)dt − κ1 dA∗1 (t) − κ2 dA∗2 (t) − κ3 dA∗3 (t);
2
a ∗ (0) = a ∗ , (γ = κ1 + κ2 + κ3 ),
√
dY3 (t) = κ3 a(t)dt + dA3 (t),
√
dY2 (t) = κ2 a(t)dt + dA2 (t), (5.81)
where A1 (t), A2 (t), and A3 (t) are, respectively, the annihilation processes of the
input fields in channels v, w, and u (with v and u in the vacuum state, and w in a
coherent state), while Y3 (t) and Y2 (t) are, respectively, the output fields of channels
u and w.
Let us introduce the following quadrature components of (5.81), x1 (t) = q(t) =
a(t) + a ∗ (t), x2 (t) = p(t) = (a(t) − a ∗ (t)/ı, v(t) = (v1 (t) = A1 (t) +
A∗1 (t), v2 (t) = (A1 (t) − A∗1 (t))/ı) , w(t) = (w1 (t) = A2 (t) + A∗2 (t), w2 (t) =
(A2 (t)−A∗2 (t))/ı) , u(t) = (u 1 (t) = A3 (t)+A∗3 (t), u 2 (t) = (A3 (t)−A∗3 (t))/ı) ,
z(t) = (z 1 (t) = Y3 (t) + Y3∗ (t), z 2 (t) = (Y3 (t) − Y3∗ (t))/ı) and y(t) = (y1 (t) =
Y2 (t)+Y2∗ (t), y2 (t) = (Y2 (t)−Y2∗ (t))/ı) . This leads to the quadrature form (5.61)
with the following system matrices:
γ √ √ √
A = − I ; B0 = − κ1 I ; B1 = − κ2 I ; B2 = − κ3 I ;
√2 √
C1 = κ3 I ; D12 = I ; C2 = κ2 I ; D21 = I.
In this example, suppose that the coupling coefficients of the cavity are κ1 = 2.6,
κ2 = κ3 = 0.2, so that the total cavity decay rate is κ = κ1 + κ2 + κ3 = 3. Setting
g = 0.1, the Riccati equations (5.68) and (5.69) were found to have stabilizing
solutions satisfying Assumption 5.2 given by X = Y = 02×2 . If a controller of the
form (6.53) is employed with matrices A K , B K , C K determined as in (5.70), then
following Theorem 5.2 the resulting closed-loop system will be strictly bounded real
with disturbance attenuation g. In this instance, the matrices are given by
A K = −1.1I, B K = −0.447I, C K = −0.447I.
From the resulting controller matrices, it can be seen that (6.53) can be realized as
another optical cavity with annihilation operator a K (with quadratures x K 1 = q K =
a K + a K∗ , x K 2 = p K = (a K − a K∗ )/ı, x K = (q K , p K ) ), connected at the output
to a 180o phase shifter. The cavity realizing the controller has coupling coefficients
κ K 1 = 0.2, κ K 2 = 1.8, κ K 3 = 0.2, and κ K = 2.2. For suitable choices of the
matrices B K 0 and B K 1 , the H ∞ controller takes the quadrature form:

d x K (t) = A K x K (t)dt + [ B K 1 B K ][ dv
K dy ] ,

d ũ(t) = −C K ξ(t)dt + [ I2 02×4 ][ dv
K dy ] ,
Fig. 5.5 An optical cavity

quantum realization of the
controller for the plant u y
shown in Fig. 5.4. Reprinted,
with permission, from [34]
ıπ
© 2008 IEEE e
vK1
κK1 = 0.2 κK3 = 0.2

aK
κK2 = 1.8
vK2
where v K (t) = (v K 11 (t), v K 12 (t), v K 21 (t), v K 22 (t)) are the quadratures of two dis-
tinct quantum noises, B K 1 = [ −0.447I −1.342I ], and ũ(t) is the output of the
cavity. Note that B K 1 (whose choice does not affect the H ∞ performance) has been
specifically chosen so that this controller can be realized as another optical cavity.
The output of the controller is u(t), given by u(t) = K s ũ(t), where K s = −I2 . Here,
K s models the 180o phase shift at the cavity output. Thus, the controller is of the form
(5.63) with B K 0 = [− I 0 ] and B K 1 as given above. This controller is illustrated in
Fig. 5.5.
5.3.3.2 Robust Stability in Quantum Optics
To illustrate the application of H ∞ controller synthesis for stability robustness, we

modify the previous example by allowing for uncertainty in the value of the coupling
coefficient to the input v. That is, we assume the actual value of the coupling coef-
ficient is κ1 + δ, where δ is an unknown parameter modeling the uncertainty. For
this modified example, the equations (5.61) describing the optical cavity now have
matrices
γ+δ √
A=− I ; B0 = κ1 + δ I ; B1 = − κ2 I ;
√2 √
B2 = − κ3 I ; C1 = κ3 I ; D12 = I ;
√
C2 = κ2 I ; D21 = I. (5.82)
We first derive an upper bound of the uncertainty in the matrix A. To do this, we

assume that the uncertainly δ is bounded in magnitude, |δ| < μ for some known
positive constant μ. Then for any non-singular matrix S, we have that − 2δ I = B̃1 C̃1
where B̃1 = μ2 S, C̃1 = S −1 and = − μδ I satisfies ≤ I . Consider a family of
systems of the form (5.61) with system matrices
γ
A = − I + B̃1 C̃1 ; B0 = κ1 + δ I ;
2√ √ √
B1 = − κ2 I ; B2 = − κ3 I ; C1 = κ3 I ;
√
C2 = κ2 I ; D12 = I ; D21 = I, (5.83)
with ≤ I . This family of systems includes the true system as a member.

In order to apply the H ∞ theory together with the results of Sect. 5.3.2.5 to this
example, we consider a related H ∞ problem defined by a system of the form (5.61)
with
γ √
A = − I ; B0 = κ1 + δ I ; B10 = − κ2 I ;
2
√ √
B1 = B10 B̃1 ; B2 = − κ3 I ; C10 = κ3 I ;

C10 D120
C1 = ; D120 = I ; D12 = ;
g C̃1 0
√
C2 = κ2 I ; D210 = I ; D21 = D210 0 . (5.84)
Here, g is the disturbance attenuation parameter in the H ∞ control problem to be

considered. We note that B0 depends on the unknown parameter δ, but this quantity
will not be involved in the calculation of the H ∞ controller.
We set the parameter values for the nominal plant cavity to be the same as in the
previous example, κ1 = 2.6, κ2 = κ3 = 0.2. Also, suppose that μ = 0.1. With a
disturbance attenuation constant of g = 0.1 and S = 1.5I , the Riccati equations
(5.68) and (5.69) were found to have stabilizing solutions satisfying Assumption 5.2
given by X = 0.1733I , Y = 0.0022I . Thus, the corresponding controller matrices
are as follows:
A K = −1.0997I, B K = −0.4464I, C K = −0.4464I. (5.85)
As in the previous example, the controller defined by the matrices (5.85) can be
realized by another optical cavity. In this case, the controller parameters are κ K 1 =
0.1993, κ K 2 = 1.8008, κ K 3 = 0.1993, and γ K = 2.1993. The physical realization
of the controller is analogous to the one shown in Fig. 5.5.
The resulting closed-loop system is of the form (5.65), with

A B2 C K
Ã = ;
B K C2 A K

B1 B10 B̃1
B̃ = = ;
B K D21 B K D210 0

B0 B2 B K 0
G̃ = ;
B K D20 B K 1

C10 D120 C K
C̃ = C1 D12 C K = ;
g C̃1 0

D̃ = 0 D12 B K 0 ,
and, by Theorem 5.2, satisfies the strict bounded real condition with disturbance
attenuation g. As a consequence of Corollary 5.3, we have that C̃(s I − Ã)−1 B̃∞ <
g, and find that
! !
! !
! C10 D120 C K (s I − Ã)−1 B10 ! <g (5.86)
! B K D210 !∞
and ! !
! !
! C̃1 0 (s I − Ã)−1 B̃1 ! < 1. (5.87)
! 0 ! ∞
By Corollary 5.3, (5.86) implies that the nominal closed-loop system is strictly
bounded real with disturbance attenuation g, while (5.87) implies that the closed-loop
system

A B2 C K
d x̃(t) = x̃(t)dt +
B K C2 A K

B0 B2 B K 0 dv(t) B̃1
+ dw(t),
B K D20 B K 1 dv K (t) 0

dz(t) = C̃1 0 x̃(t)dt,
is strictly bounded real with unity disturbance attenuation. Therefore, by Theorem

5.5, we conclude that the closed-loop uncertain system

A + B̃1 C̃1 B2 C K
d x̃(t) = x̃(t)dt +
B K C2 AK

B0 B2 B K 0 dv(t)
,
B K D20 B K 1 dv K (t)
is mean square stable for all matrices such that ≤ I . Hence, the true closed-
loop system must also be mean square stable. Furthermore, for this example, it is
also possible to verify that the true closed-loop system must in fact also be strictly
bounded real with disturbance attenuation g.
5.4 Further Reading
The exposition of H ∞ control for linear quantum systems in this chapter has been lim-
ited to coherent feedback controllers in the form of linear quantum systems. However,
the original treatment in [34], on which the exposition here is based, was quite gen-
eral allowing for linear classical and mixed quantum-classical H ∞ controllers. The
notion of physical realizability for systems with mixed quantum-classical dynamics
and the associated conditions were introduced in [34, 44], based on the notion of a
fully quantum augmentation of mixed quantum-classical linear systems, and devel-
oped further in [45–47]. Application of the theory to the construction of quantum
optical realizations of classical linear stochastic systems was developed in [48].
The H ∞ controller design allowed for some free parameters that can be used to
yield physically realizable controllers. However, this may entail adding additional
noise channels to satisfy the physical realizability constraints, which is undesirable
and needs to be minimized. The minimal number of additional quantum noise chan-
nels that are required to make a linear quantum controller physically realizable for
any triplet (A K , B K , C K ) has been investigated in [49, 50], where the minimal num-
ber can be characterized in terms of rank of a certain matrix that depends upon
(A K , B K , C K ).
Appendices
Appendix A: Proof of Theorem 5.2
The proof of Theorem 5.2 will use the following lemma.

Lemma 5.4 Consider a real symmetric matrix X and corresponding operator-
valued quadratic form x X x for the system (5.54). Then the following statements
are equivalent:
(i) There exists a constant λ ≥ 0 such that ρ, x X x ≤ λ for all Gaussian states ρ.
(ii) The matrix X is negative semidefinite.
Proof (i) ⇒ (ii). To establish this part of the lemma, consider a Gaussian state
ρ which has mean x̄ and covariance matrix Y = [Tr(ρxi x j )], with symmetrized
covariance 1/2(Y + Y ) ≥ 0 and Y − Y = 2ıJn . Then, we can write
Appendix A: Proof of Theorem 5.2 197

n
n
ρ, x X x = X i j ρ, xi x j
i=1 j=1

n
n
= X i j [Yi j + x̄i x̄ j ]
i=1 j=1
= x̄ T X x̄ + Tr[X Y ]. (5.88)
Now for any constant α > 0, consider the inequality of part (i) where ρ is a Gaussian
state with mean α x̄ and covariance matrix Y . Then, it follows from the bound
ρ, x X x ≤ λ and (5.88) that α2 x̄ X x̄ + Tr[X Y ] ≤ λ for all α > 0. From
this, it immediately follows that x̄ X x̄ ≤ 0. However, x̄ can be chosen arbitrarily.
Hence, we can conclude that condition (ii) of the lemma is satisfied.
(ii) ⇒ (i). Suppose that the matrix X is negative semidefinite and let ρ be any
Gaussian state and suppose that ρ has mean x̄ and covariance matrix Y ≥ 0. Then, it
follows from (5.88) that ρ, x X x = x̄ X x̄ + Tr[X Y ]. However, X ≤ 0, Y + Y ≥
0, and Y − Y = 2ıJn implies x̄ X x̄ ≤ 0, and Tr[X Y ] = 1/2Tr[X (Y + Y )] ≤ 0.
Hence, ρ, x X x ≤ 0 and condition (ii) is satisfied with λ = 0.
Proof of Theorem 5.2. Let the system be dissipative with V (x) = x X x. By Itō’s
rule, the product table (5.60), and the quantum stochastic differential equation (5.54),
we have
d V (x(t))
= d x (t)X x(t) + x (t)X d x(t) + d x (t)X d x(t)
= x (t)(A X + X A)x(t) + βw (t)B X x(t) +
x (t)X Bβw (t) + λ0 dt, (5.89)
where λ0 is given by (5.59). We now note that V (x(t)) = ρ, E 0 [V (x(t))] , where

E t (t ≥ 0) denotes the conditional expectation map2 with respect to the vacuum state
| of the field (e.g., see [51, p. 215]), and ρ is an initial Gaussian state of the system.
Combining this with the integral of (5.89) and (5.56) we find that
" t
ρ, E 0 [x (s)(A X + X A)x(s) + βw (s)B X x(s)+
0
#

x (s)X Bβw (s) + λ0 + r (x(s), βw (s)]ds ≤ λt.
Let t → 0 to obtain
"
ρ, x (A X + X A)x + βw B X + x X Bβw + λ0 +
#
x
[x βw ]R ≤ λ.
βw
2 This map is distinct from the quantum conditional expectation used in quantum filtering theory.
Here, x and βw denote the initial conditions. An application of Lemma 5.4 implies
(5.57). Also, (5.58) is a straightforward consequence of this inequality when R is
replaced by R + I where > 0.
To establish the converse part of the theorem, we first assume that (5.57) is satis-
fied. Then with V (x) = x X x, it follows from (5.89) that
t
V (x(t)) − V (x(0)) + r (x(s), βw (s)) ds ≤ λ0 t
o
for all t > 0 and all βw (t). Hence, inequality (5.56) is satisfied with λ given by
(5.59).
If matrix inequality (5.58) is satisfied, then it follows by similar reasoning that
there exists an > 0 such that
t
V (x(t)) − V (x(0)) + r (x(s), βw (s)) +
0
(x(s) x(s) + βw (s) βw (s)) ds ≤ λ0 t.
Hence, inequality (5.56) is satisfied with λ = λ0 given by (5.59) and with R replaced
by R + I .
Appendix B: Proof of Lemma 5.2
The proof of Lemma 5.2 will use the following lemma.

Lemma 5.5 If S is a Hermitian matrix, then there is a real constant α0 such that
αI + S ≥ 0 for all α ≥ α0 .
Proof Since S is Hermitian, it has real eigenvalues and is diagonalizable. Hence,
S = V ∗ E V for some real diagonal matrix E and orthogonal matrix V . Now let
α0 = −λ, where λ is the smallest eigenvalue of S. The result follows since αI + S =
V ∗ (αI + E)V while αI + E ≥ 0 for all α ≥ α0 .
Proof of Lemma 5.2. The main idea is to explicitly construct matrices R ∈ R2n K ×2n K ,
K ∈ C(n v K +n y )×2n K , B K 1 ∈ R2n K ×2n v K , and B K 0 ∈ R2n u ×2n v K , with n v K ≥ n u , such
that (see Chap. 2)
A K = 2Jn K (R + {K∗ K}), (5.90)

BK 1 B K = 2ıJn K [ −K∗ K ]n v K +n y , (5.91)
C K = [ I2n u ×2n u 02n u ×2(n v K +n y −n u ) ]Pnv +n y
K
K + K#
× , (5.92)
−ıK + ıK#

B K 0 02n u ×2(n v K +n y ) = [ I2n u ×2n u 02n u ×2(n v K +n y −n u ) ], (5.93)
Appendix B: Proof of Lemma 5.2 199
are satisfied. To this end, let Z = 21 J−1 n K A = − 2 Jn K A. We first construct matrices

1
Kb2 , Kb1 , B K 1,1 , and B K 1,2 according to the following procedure:

1. Construct the matrix Kb2 according to (5.75).
2. Construct a real symmetric 2n K × 2n K matrix W1 such that the matrix
W2 = W1

Z − Z 1 0 I ∗
+ı − C K Pn u Pn u C K − (Kb2 Kb2 ) ,
2 4 −I 0
is non-negative definite. It follows from Lemma 5.5 that such a matrix W1 always
exists.
∗
3. Construct a matrix Kb1 such that Kb1 Kb1 = W2 , where Kb1 has at least 1 row.
This can be done, for example, using the singular value decomposition of W2 (in
this case, Kb1 will have 2n K rows).
4. Construct the matrices B K 1,1 and B K 1,2 according to Eqs. (5.74) and (5.76),
respectively.
Let R = 21 (Z + Z ). We now show that there exists an integer n v K ≥ n u such
conditions (5.90)–(5.93) are satisfied with the matrix R as defined and with B K 1 =
[ B K 1,1 B K 1,2 ] and
⎡ ⎤
1/2 I ı I Pn u Ck
K=⎣ Kb1 ⎦. (5.94)
Kb2
First note that necessarily n v K ≥ n u +1 > n u since B K 1 has at least 2n u +2 columns.

Also, by virtue of our choice of Kb1 , we have
∗
(Kb1 Kb1 ) = {W2 }

1/4C P
0 I ∗
= 1/2(Z − Z )− Pn u C K − {Kb2 Kb2 },
K nu −I 0
and hence

∗ ∗ ∗ 1/4C P
0 I
(K K) = (Kb1 Kb1 ) + (Kb2 Kb2 ) + Pn u C K
K nu −I 0
= 1/2(Z − Z ).

Since R = Z +Z
2
, we have R + (K∗ K) = Z . Therefore, (5.90) is satisfied.
Now, we observe that ıJn K B K diagn y (M ∗ )Pny = [ T −T # ] for some 2n K × n y
complex matrix T . But by taking the conjugate transpose of both sides of (5.75)
∗
which defined Kb2 , we conclude that T = −Kb2 . Hence,
∗
B K = 2ıJn K [ −Kb2 Kb2 ]Pn y diagn y (M). (5.95)
From (5.74) which defined B K 1,1 , we obtain
B K 1,1 = Jn K C K diagn u (J)

= Jn K C K diagn u (J)(2diagn u (M ∗ ))diagn u (M)

−1 1
= ıJn K C K diagn u diagn u (M)
ı ı

−I I
= ıJn K C K Pnu Pn u diagn u (M). (5.96)
ıI ıI
Combining (5.76), (5.95) and (5.96) gives us
[ B K 1,1 B K 1,2 B K ]

−I I
∗
= 2ıJn K /2C K Pn u
1 Pn u −Kb1 Kb1 Pn v K −n u
ıI ıI

−Kb2∗
Kb2 Pn y Pnw Pn w K diagn w (M)
K K

I
= 2ıJn K −1/2C K Pnu −Kb1 ∗ ∗
−Kb2
−ı I

1/2C P
I
K nu ı I Kb1 b2 Pn w K diagn w K (M)
K

I
= 2ıJn K −1/2C K Pnu −Kb1 ∗ ∗
−Kb2
−ı I

1/2C P
I
K nu ı I Kb1 b2
K

= 2ıJn K −K∗ K .
Therefore, (5.91) is also satisfied. Moreover, it is straightforward to verify (5.92) by

substituting K as defined by (5.94) into the right-hand side of (5.92). Finally, since
n v K ≥ n u , it follows that [ B K 0 02n u ×2(n v K +n y −n u ) ] is precisely the right-hand side of
(5.93). This completes the proof of Theorem 5.4.
References
1. L. Bouten, Filtering and control in quantum optics, Ph.D. dissertation, Catholic University of
Nijmegen (2004)
References 201
4. D.A. Steck, K. Jacobs, H. Mabuchi, S. Habib, T. Bhattacharya, Feedback cooling of atomic

motion in cavity QED. Phys. Rev. A 74, 012322 (2006)
5. A.C. Doherty, K. Jacobs, Feedback-control of quantum systems using continuous state-
estimation. Phys. Rev. A 60, 2700 (1999)
6. R. Hamerly, H. Mabuchi, Advantages of coherent feedback for cooling quantum oscillators.
Phys. Rev. Lett. 109, 173602 (2012)
7. A.H. Jazwinski, Stochastic Processes and Filtering Theory (Academic Press, New York, 1970)
8. A. Bagchi, Optimal Control of Stochastic Systems. International Series in Systems and Control
Engineering (Prentice Hall, Upper Saddle River, 1993)
9. W.L. Brogan, Modern Control Theory, 3rd edn. (Prentice-Hall, Upper Saddle River, 1991)
10. K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control (Prentice-Hall, Upper Saddle
River, 1995)
1846 (2009) © 2009 Reprinted, with permission from Elsevier
13. G. Zhang, M.R. James, Direct and indirect couplings in coherent feedback control of linear
quantum systems. IEEE Trans. Autom. Control 56(7), 1535–1550 (2011)
14. I.G. Vladimirov, I.R. Petersen, A gradient descent approach to optimal coherent quantum LQG
controller design, in Proceedings of the 2015 American Control Conference (ACC) (2015), pp.
1487–1492
15. A.K. Sichani, I.G. Vladimirov, I.R. Petersen, A gradient descent approach to optimal coherent
quantum LQG controller design, in Proceedings of the 2015 American Control Conference
(ACC) (2015), pp. 1487–1492
16. C.W. Scherer, P. Gahinet, M. Chilali, Multiobjective output-feedback control via LMI opti-
mization. IEEE Trans. Autom. Control 42(7), 896–911 (1997)
17. Y. Nesterov, A. Nemirovskii, Interior-Point Polynomial Algorithms in Convex Programming
(SIAM, Philadelphia, 1994)
18. S. Boyd, L. Vanderberghe, Semidefinite programming relaxations of non-convex problems in
control and combinatorial optimization, in Communications, Computation, Control and Signal
Processing - A Tribute to Thomas Kailath, ed. by A. Paulraj, V. Roywhcowdhury, C. Schaper
(Kluwer Academic Publishers, New York, 1997)
19. R. Orsi, U. Helmke, J.B. Moore, A Newton-like method for solving rank constrained linear
matrix inequalities. Automatica 42(11), 1875–1882 (2006)
20. J.B. Lasserre, Global optimization with polynomials and the problem of moments. SIAM J.
Optim. 11(3), 796–817 (2001)
21. M. Kojima, Sums of squares relaxations of polynomial semidefinite programs. Department of
Mathematics and Computer Science, Tokyo Institute of Technology, Tokyo, Japan. Technical
report B-397 (2003)
22. C.W.J. Hol, C. Scherer, Sum of squares relaxations for polynomial semidefinite programming,
in Proceedings of the 16th International Symposium on Mathematical Theory and Networks
(MTNS) 2004 (Leuven, Belgium, July 5–9, 2004) (2004)
23. D. Henrion, J.B. Lasserre, Convergent relaxations of polynomial matrix inequalities and static
output feedback. IEEE Trans. Autom. Control 51(2), 192–202 (2006)
24. R. Orsi, LMIRank: Software for Rank Constrained LMI Problems (2005), http://rsise.anu.edu.
au/~robert/lmirank/
25. J. Löfberg, YALMIP: a toolbox for modeling and optimization in MATLAB, in Proceedings of
the CACSD Conference, Taipei, Taiwan (2004), http://control.ee.ethz.ch/~joloef/yalmip.php
26. Advanced Optimization Lab, McMaster University: SeDuMi v1.1R3 (2006), http://sedumi.
mcmaster.ca/
27. H. Bachor, T. Ralph, A Guide to Experiments in Quantum Optics, 2nd edn. (Wiley-VCH,
Weinheim, 2004)
2004)
29. K. Jacobs, H.I. Nurdin, F.W. Strauch, M.R. James, Comparing resolved-sideband cooling and
measurement-based feedback cooling on an equal footing: analytical results in the regime of
ground-state cooling. Phys. Rev. A 91, 043812 (2015)
30. J.C. Doyle, K. Glover, P.P. Khargonekar, B. Francis, State-space solutions to the standard H2
and H∞ control problems. IEEE Trans. Autom. Control 34(8), 831–847 (1989)
31. I. Petersen, B. Anderson, E. Jonckheere, A first principles solution to the non-singular H ∞
control problem. Int. J. Robust Nonlinear Control 1(3), 171–185 (1991)
32. C.W. Scherer, Multiobjective H2 /H∞ control. IEEE Trans. Autom. Control 40(6), 1054–1062
(1995)
33. J. Helton, M. James, Extending H ∞ Control to Nonlinear Systems: Control of Nonlinear
Systems to Achieve Performance Objectives, Advances in Design and Control, vol. 1 (SIAM,
Philadelphia, 1999)
IEEE Trans. Autom. Control 53(8), 1787–1803 (2008) Reprinted, with permission, © 2008
IEEE
35. J. Willems, Dissipative dynamical systems - Part I: general theory. Arch. Rational Mech. Anal.
45, 321–351 (1972)
36. P. Dupuis, M. James, I. Petersen, Robust properties of risk-sensitive control. Math. Control
Syst. Signals 13, 318–332 (2000)
37. C. D’Helon, M.R. James, Stability, gain, and robustness in quantum feedback networks. Phys.
Rev. A 73, 053803 (2006)
38. A. van der Schaft, L 2 -Gain and Passivity Techniques in Nonlinear Control (Springer, New
York, 1996)
39. S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix Inequalities in System and
Control Theory (SIAM, Philadelphia, 1994)
40. K. Zhou, P. Khargonekar, An algebraic Riccati equation approach to H∞ optimization. Syst.
Control Lett. 11, 85–91 (1988)
41. M. Green, D. Limebeer, Linear Robust Cotrol (Prentice-Hall, Englewood Cliffs, 1995)
42. H. Mabuchi, Coherent-feedback quantum control with a dynamic compensator. Phys. Rev. A
78, 032323 (2008)
43. M. Sarovar, D.B.S. Soh, J. Cox, C. Brif, C.T. DeRose, R. Camacho, P. Davids, Silicon nanopho-
tonics for scalable quantum coherent feedback networks. EPJ Quantum Technol. 3(14), 1–18
(2016)
44. H.I. Nurdin, Topics in classical and quantum linear stochastic systems, Ph.D. dissertation, The
Australian National University (2007)
45. H.I. Nurdin, Network synthesis of mixed quantum-classical linear stochastic systems, in Pro-
ceedings of the 2011 Australian Control Conference (AUCC), Engineers Australia (2011), pp.
68–75
46. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Synthesis and structure of mixed quantum-
classical linear systems, in Proceedings of the 51st IEEE Conference on Decision and Control
(Maui, Hawaii, Dec. 10–13, 2012) (2012), pp. 1093–1098
47. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Network synthesis for a class of mixed quantum-
classical linear stochastic systems (2014), arXiv:1403.6928 (arXiv preprint)
48. S. Wang, H.I. Nurdin, G. Zhang, M.R. James, Quantum optical realization of classical linear
stochastic systems. Automatica 49(10), 3090–3096 (2013)
49. S. Vuglar, I.R. Petersen, How many quantum noises need to be added to make an LTI system
physically realizable? in Proceedings of the 2011 Australian Control Conference (AUCC)
(2011), pp. 363–367
50. S. Vuglar, I.R. Petersen, Quantum implemention of an LTI system with the minimal number
of additional quantum noise inputs, in Proceedings of the 2013 European Control Conference
(ECC) (2013), pp. 2724–2727
51. K.R. Parthasarathy, An Introduction to Quantum Stochastic Calculus (Birkhauser, Berlin, 1992)
Chapter 6
Linear Systems and Control Theory
for Quantum Information
Abstract This chapter illustrates several example applications of the theory of

linear quantum systems to the analysis of problems of interest in quantum informa-
tion processing and discusses two experimental demonstrations of real-time coherent
feedback and measurement-based feedback control from the literature. The problems
covered are dissipative generation of Gaussian states of single-mode oscillators, effi-
cient enhancement of entanglement between traveling Gaussian fields, back-action
evasion, perfect state transfer in a linear quantum network, and robust quantum
amplification. The two experiments are demonstrations of enhancement of optical
squeezing via static coherent feedback and generation of a spin-squeezed state in an
atomic ensemble via measurement-based feedback control.
In this chapter, we discuss several applications of linear systems and control theory
for continuous-variable (CV) quantum information processing. Here, the meaning
of CV is that the spectrum of the observables q or p is continuous in R, in contrast
to the discrete variable case where, for instance, the spectrum of σz = diag(1, −1)
is discrete. The topics discussed in this chapter are as follows:
• Section 6.1: Dissipative generation of arbitrary pure Gaussian states,
• Section 6.2: Entanglement enhancement via coherent feedback networks,
• Section 6.3: Back-action evasion via feedback for force sensing,
• Section 6.4: Quantum memory with decoherence-free subsystem,
• Section 6.5: Robust quantum amplification via coherent feedback control, and
• Section 6.6: Quantum feedback experiments for squeezing enhancement.
The topics in Sects. 6.1 and 6.4 do not deal with feedback control problems but
discuss problems of system synthesis so that the engineered system achieves state
generation in the former case and state transfer in the latter. Both cases fully uti-
lize system theoretic notions/tools such as controllability and zero dynamics. On the
Section 6.1 contains reprinted excerpt with permission from [10]. Copyright (2012) by the
American Physical Society.
Section 6.2 contains some materials reprinted from [22] with permission of Springer.
Section 6.5 contains reprinted excerpt with permission from [65]. Copyright (2016) by the
American Physical Society.
Section 6.6 contains some materials reprinted, with permission, from [21] © 2012 IEEE.
204 6 Linear Systems and Control Theory for Quantum Information
other hand, in Sects. 6.2, 6.3, and 6.5, we elaborate upon coherent feedback networks
for enhanced CV entanglement generation, back-action evasion, and robust quantum
amplification, respectively. Lastly, Sect. 6.6 is devoted to showing two experimental
demonstrations of quantum feedback for optical and spin-squeezing enhancement.
It is hoped that these examples will illustrate to the readers that systems and control
theory provides powerful and pertinent tools for synthesizing and controlling quan-
tum systems as quantum information processors.
6.1 Dissipative Generation of Pure Gaussian States
Our first topic deals with the problem of generating a target Gaussian state in a lin-
ear quantum system. In general, a Gaussian state is useful for quantum information,
such as for quantum teleportation and secure communication [1, 2]. In particular,
it is known that a Gaussian cluster state [3, 4] serves as an essential resource for
measurement-based quantum computation. The method for generating a Gaussian
state presented here belongs to the so-called environment engineering (or reservoir
engineering) approach [5] described as follows. Recall now that for a closed system,
generation of a target final state |ψ f is implemented by applying a unitary operation
to a given initial state |ψi so that |ψi → |ψ f = U |ψi ; but in such an implemen-
tation, the final state |ψ f is very sensitive to the initial state |ψi and the parameters
of the unitary operator U , which includes the duration time. The environment engi-
neering approach overcomes these issues. The point of this approach is to exploit
open systems, rather than closed systems. The dynamics of the statistics of an open
dissipative system is modeled by the master equation (4.20):
dρ m
1 1
= −i[H, ρ] + L k ρL ∗k − L ∗k L k ρ − ρL ∗k L k , (6.1)
dt k=1
2 2
where ρ is the (unconditional) quantum state, H is the Hamiltonian, and L k is the

Lindblad operator representing the kth system–bath coupling. Note again that the
second Lindblad term on the right-hand side of (6.1) represents the dissipation. In
the environment engineering approach, we aim to design H and L k so that ρ evolves
toward a target pure state: ρ(t) → |ψ f ψ f | as t → ∞. The notable advantage
of this method is that the dissipation-induced final state |ψ f is independent of the
initial state ρ(0), and also, we do not need to stop the evolution at a certain precise
time. That is, this method can realize robust generation of |ψ f . Thus, what we need
to use the environment engineering method is the condition for this desired time
evolution to occur. In the finite-dimensional case, this has been given in [6, 7], and
its application to quantum information processing has been proposed [8, 9].
In this section, we discuss the infinite-dimensional Gaussian version of the envi-
ronment engineering method [10]. The discussion begins with a necessary and suffi-
cient condition for the master equation (6.1) to have a unique pure Gaussian steady
state, followed by a complete parameterization of the linear quantum system having
this property. Then, a case study is given to demonstrate a pure Gaussian cluster state
6.1 Dissipative Generation of Pure Gaussian States 205
generation. Finally, a QSDE formulation [11] is provided, showing the general pas-
sivity property of the desired dissipative dynamics and a specific trade-off appearing
in the problem of pure Gaussian cluster state generation.
6.1.1 General Condition for Pure Gaussian State Generation
In this section, we define x as x = (q , p ) = (q1 , . . . , qn , p1 , . . . , pn ) , which

satisfies

0 In
x x − (x x ) = ı J̄n , J̄n = J ⊗ In = .
−In 0
As described in Sect. 2.7.1, a quantum Gaussian state can be completely char-

acterized by only the mean vector x and the symmetrized covariance matrix
V = xx + (xx ) /2, x = x − x. For an open linear system with
a Gaussian state, the time evolution equations of x and V can be obtained in a
very simple form as follows. That is, for a linear quantum system with Hamiltonian
H = x Rx/2 and coupling operator L = K x, where R = R ∈ R2n×2n and
K ∈ Cm×2n , the time evolution of x and V is given by
dx dV
= Ax, = AV + V A + D, (6.2)
dt dt
where A = J̄n (R + {K ∗ K }) and D = J̄n {K ∗ K }J̄ n . The state ρ(t) is Gaussian

with mean x(t) and covariance matrix V (t) for all t.
The environment engineering problem for Gaussian state generation can be for-
mulated as follows. Now, when A is Hurwitz, the steady state of the above linear
quantum system is given by a Gaussian state with mean x(∞) = 0 and covariance
matrix Vs , which is the unique solution to the algebraic Lyapunov equation
AVs + Vs A + D = 0. (6.3)
Therefore, our goal is to synthesize the system matrices R and K so that Vs coincides
with the covariance matrix of a target pure Gaussian state. To attack this problem,
the following result is useful:
Theorem 6.1 Let Vs be a covariance matrix corresponding to a pure Gaussian
state. Then, this is the unique steady state of a linear quantum system characterized
by (R, K ), if and only if
ı
ker Vs + J̄n = range(K ), (6.4)
2
where
K = [K R J̄
n K

. . . (R J̄
n)
2n−1
K ] .
For the proof, see [10]. Note that this result does not state whether a given system
has a unique pure steady state; for this purpose, Theorem 6.1 in [10] can be used.
6.1.2 Synthesizing a Dissipative Gaussian System
Here, we show that Theorem 6.1 leads to a complete parametrization of (R, K ),

which as a consequence provides us with an explicit procedure for synthesizing a
desired open Gaussian system. Let us start from the fact that in general, the covariance
matrix Vs of a pure Gaussian state can be represented as follows [3, 4, 12]:

1 Y −1/2 0
Vs = SS , S = , (6.5)
2 X Y −1/2 Y 1/2
where X and Y are n × n real symmetric and real positive definite matrices (i.e.,
Y = Y > 0), respectively. Note that S is J̄n -symplectic, satisfying S J̄n S = J̄n .
Then, defining Z = X + ıY leads to

ı 1 I
Vs + J̄n = Y −1 [I Z ].
2 2 Z∗
Now, because the rank of [I Z ] is n, we have

ı −Z
ker Vs + J̄n = range .
2 I
Hence, (6.4) is equivalent to

−Z
range = range K R J̄
n K

. . . (R J̄
n)
2n−1
K .
I
To satisfy this condition, it is required that range(K ) is contained in range [−Z I ]

and range [−Z I ] is invariant under R J̄ n . These conditions are expressed as

−Z −Z −Z
K = P, R J̄
n = Q, (6.6)
I I I
where P and Q are some n × m and n × n complex matrices. From these equations,
K can be expressed as

−Z
K = [P Q P . . . Q 2n−1 P].
I
Consequently, the necessary and sufficient condition for range(K ) to coincide with
range [−Z I ] is that there exist P and Q satisfying (6.6) and the rank condition
rank [P Q P . . . Q n−1 P] = n. (6.7)
Now, let us decompose R into the following form:

R1 R2
R= ,
R2 R3
where R1 and R3 are real n × n symmetric matrices and R2 a real n × n matrix. Then,
the latter equation in (6.6) leads to
R1 + R2 X + X R2 + X R3 X − Y R3 Y = 0,
(R2 + X R3 )Y + Y (R2 + X R3 ) = 0.
The second equation equivalently means that there exists a real skew symmetric
matrix (i.e., + = 0) satisfying (R2 + X R3 )Y = . Hence, by writing
P = R3 , we find that R2 can be expressed as R2 = −X P + Y −1 . Then, it follows
from (6.6) that Q is of the form Q = −ı P Y − Y −1 . As a consequence, the
complete parameterizations of K and R are obtained as follows:
K = P [−Z I ],

X P X + Y P Y − Y −1 X − X Y −1 −X P + Y −1
R= . (6.8)
−P X + Y −1 P
Again, P (complex), P (real symmetric), and (real skew) are the parameterizing
matrices. Thus, a procedure to synthesize a linear quantum system generating a target
pure Gaussian state has now been constructed; given a target pure Gaussian state
with covariance matrix (6.5), which determines X and Y , choose the free matrices
P, P , and so that both the system–field coupling L = K x and the Hamiltonian
H = x Rx/2 with the system matrices (6.8) would be physically implementable,
while P and Q = −ı P Y − Y −1 satisfy the rank condition (6.7).
Lastly, note that there always exists a triplet (P, P , ) satisfying the rank con-
dition (6.7), if no additional constraint is required for those matrices; that is, for an
arbitrary given pure Gaussian state, we can always find a Gaussian dissipative system
that produces that state as its unique steady state.
6.1.3 Gaussian Cluster State Generation via Dissipation
Menicucci et al. [4] developed a unified graphical calculus applicable to all pure
Gaussian states, characterized by the matrix Z = X + ıY . A particularly important
result is that the canonical CV cluster state can be represented in a general form as
Z = X + ıe−2r I, (6.9)
(a) (b) L3 (c) 2

L4
1 3
2 4 1 3
L1 L1
L2
4
Fig. 6.1 a Shape of a one-dimensional harmonic chain with 4 oscillators. b Dissipative channels
(indicated by the arrows) for generating a harmonic chain state. c A quasi-local dissipative system
generating the same harmonic chain state in b; the dotted circle represents the system Hamiltonian.
Reprinted with permission from [10] © (2012) by the American Physical Society
where X is the symmetric adjacent matrix representing the structure of the cluster
state. Also, Y = e−2r I (r ∈ R) represents the approximation error of the actual state
with covariance matrix (6.5) to the ideal cluster state corresponding to X ; note that
the ideal state corresponds to the case with r → ∞ (usually interpreted as infinite
squeezing) and is thus not physically realizable.
Let us now apply the procedure obtained in the previous section to synthesize
a linear quantum system generating a typical cluster state, a one-dimensional har-
monic chain state with equal weight: This is the state of a set of quantum oscillators
connected in a chain form as depicted in Fig. 6.1a. In particular, here we consider
the case n = 4. Now, the adjacent matrix X and the resultant graph matrix (6.9) are,
respectively, given by
⎡ ⎤ ⎡ −2r ⎤
01 ıe 1
⎢1 0 1 ⎥ ⎢ 1 ıe−2r 1 ⎥
X =⎢ ⎥ ⎢
⎣ 1 0 1⎦, Z = ⎣
⎥,
1 ıe −2r
1 ⎦
10 1 ıe−2r
The above X represents that for instance, the first and the second subsystems are
coupled, with weight 1. A simple choice of the matrices in (6.8) is P = I4 , P = 0,
and = 0, leading to K = [−Z I ] and R = 0. In this case, the system is driven only
by the dissipation term characterized by the following Lindblad operator L = K x:
L 1 = (−ıe−2r q1 + p1 ) − q2 , L 2 = −q1 + (−ıe−2r q2 + p2 ) − q3 ,

L 3 = −q2 + (−ıe−2r q3 + p3 ) − q4 , L 4 = −q3 + (−ıe−2r q4 + p4 ). (6.10)
Figure 6.1b depicts how the system couples with the dissipation channels. It is
observed from this figure that each dissipative channel acts on at most three sub-
systems; in general, if the channel L acts only on a few subsystems, then it is called
quasi-local [7].
Of course, the way to design a desired system is not unique. For instance, let us
try to find a system that generates the same chain state by a quasi-local dissipative
process acting on at most two adjacent subsystems. This goal is simply motivated by
the fact that engineering such a dissipative environment might be easier to do than
the previous case, apart from that we may need an additional Hamiltonian. Here, we
take P = (1, 0, 0, 0) , implying that the system has one dissipative channel L 1 in
(6.10), which indeed acts on only two subsystems. To determine the Hamiltonian
matrix R in (6.8), we have some freedom, but let us choose
⎡ ⎤
0 1 0 −1
⎢ 1 0 0 0 ⎥
P = X −1 =⎢
⎣ 0
⎥ , = 0.
0 0 1 ⎦
−1 0 1 0
This leads to Q = −ı P Y − Y −1 = −ıe−2r X −1 , and it is readily verified that the

matrix [P Q P Q 2 P Q 3 P] is of full rank. Then, R is given by

X + e−2r X −1 −I
R= .
−I X −1
The loci of nonzero entries of this matrix show that the Hamiltonian H = x Rx/2
has a ring-type structure where the 1–2, 2–3, 3–4, and 4–1 subsystems are connected.
Therefore, we now see that the four-mode harmonic chain state can also be generated
in the dissipative system whose system–field couplings and the Hamiltonian are both
quasi-local, as shown in Fig. 6.1c.
6.1.4 A QSDE Formalism
Let us begin with the fact that a pure Gaussian state |ψ Z having the covariance
matrix (6.5) always satisfies
⎤ ⎡ ⎤ ⎡
p1 q1
⎢ .. ⎥ ⎢ .. ⎥
N |ψ Z = 0, N := [−Z I ]x = p − Zq = ⎣ . ⎦ − Z ⎣ . ⎦ . (6.11)
pn qn

That is, for Z = (z i j ), this equation means that ( pi − j z i j q j )|ψ Z = 0 for
all i. Conversely, if a pure Gaussian state |ψ satisfies N |ψ = 0, it follows that
|ψ = |ψ Z . The vector of operators, N , is called the nullifier associated with the
pure Gaussian state |ψ Z .
Clearly, it should be worth examining the time evolution of N in the Heisenberg
picture, when the conditions given in Theorem 6.1 are imposed. For this purpose, let
us recall the following QSDE with system variable x = (q , p ) :
d x(t)
= Ax(t) − ı J̄n K ∗ ξ(t) + ı J̄n K ξ(t)# , (6.12)
dt
where A = J̄n (R + {K ∗ K }). Recall that the output equation of the system is
given by
η(t) = K x(t) + ξ(t). (6.13)
Now, from (6.6), we have
1
[−Z I ]A = [−Z I ]J̄n R + [−Z I ]J̄n (K ∗ K − K K # )
2ı
1 −Z #
−Z

#
= Q [−Z I ] − [I Z ] P P [−Z I ] − P P ∗ [−Z # I ]
2ı I I
Z −Z #
#
= Q [−Z I ] + P P [−Z I ] = (Q − Y P # P )[−Z I ],
2ı

−Z #
[−Z I ](−ı J̄n K ∗ ) = −ı[−Z I ]J̄n P # = ı(Z − Z # )P # = −2Y P # ,
I

−Z
[−Z I ](ı J̄n K ) = ı[−Z I ]J̄n P = 0.
I
Thus, multiplying both sides of (6.12) by [−Z I ] from the left, we have
dN (t)
= (Q − Y P # P )N (t) − 2Y P # ξ(t). (6.14)
dt
Regarding the output process (6.13), because K x = P [−Z I ]x = P N , it can be

represented by
η(t) = P N (t) + ξ(t). (6.15)
The above two Eqs. (6.14) and (6.15) reveal several important properties of the
system that uniquely generates a pure Gaussian state in a dissipative way. First of all,
it must be a completely passive linear system as defined in Sect. 2.4.1; in fact, as a
property of completely passivity, the system equations contain only the annihilation-
driving noise ξ. Recall that the system equations of an optical empty cavity, which
is a very simple completely passive system, are given by (1.25) and (1.26). Hence,
we now find that the linear quantum system focused in this section is a non-trivial
extension of those simple passive optical systems, in the sense that they have a
non-vacuum pure steady state.
The above result also clarifies the meaning of rank condition (6.7), which is
indeed hard to see without the QSDE formalism discussed here. That is, we have the
following fact (the proof is given in [11]): The matrix Q −Y P # P is Hurwitz if and
only if the rank condition (6.7) is satisfied. That is, the rank condition is necessary for
the target pure Gaussian state |ψ Z to be a unique stable steady state of the system.
This fact can be seen by multiplying all the entries of (6.14) by the composite state
vector | = |ψ ⊗ | from the right (| is the vacuum state on the field); that is,
due to the relation ξ| = 0, we have
d
N (t)| = (Q − Y P # P )N (t)|.
dt
Therefore, if the rank condition (6.7) is satisfied, or equivalently Q − Y P # P is

Hurwitz, the nullifier vector N (t)| = U (t)∗ N U (t)| converges to zero. Hence,
in the Schrödinger picture, U (∞)| is the common zero eigenstate of all the ele-
ments of N , implying that from the reason given at the beginning of this section, the
system state becomes the target |ψ Z as t → ∞. As mentioned above, our system is
an extension of simple passive optical systems; but note again that while the steady
state generated in such an optical system is merely a vacuum state, the steady state
of the system (6.14) and (6.15) can be assigned to the highly non-trivial state |ψ Z .
Based on the above discussion, we here provide a remark. That is, the system
(6.14) and (6.15) with the Hurwitz property is all-pass, meaning that the output
equation at steady state does not have any information about the system, which is
indeed the reason why a pure state is generated in the end. This property can be seen
immediately in the Fourier domain; the linear transformation from the input ξ[ıω]
to the output η[ıω] is given by
η[ıω] = F[ıω]ξ[ıω], F[ıω] := I − 2P (ıω − Q + Y P # P )−1 Y P # .
Then, we can see that the transfer function matrix F[ıω] is unitary, i.e., F[ıω]
F[ıω]∗ = I , for all ω. This means that the output is a white noise process with flat
power spectrum density, i.e., η[ıω]∗ η[ıω] = ξ[ıω]∗ ξ[ıω] = I . Thus, as expected,
the output at steady state does not contain any information about the system state.
Lastly, let us examine the case where the target state is assigned to the canonical
CV cluster sate (6.9). In this case, the nullifier dynamics (6.14) is of the form
dN
= (Q − e−2r P # P )N − 2e−2r P # ξ.
dt
The real part of the eigenvalue of Q − e−2r P # P is calculated as {λ} =

−e−2r P x , where x is the corresponding eigenvector. Thus, if we make r bigger,
or equivalently if we assign the final state closer to the ideal cluster state, then the
eigenvalue moves toward the imaginary axis, and thus, the stability of the system
becomes worse (i.e., the system is closed to a marginally stable system). We can
make this interesting observation more concrete, by defining the convergence time
to the target as T = 1/min (λ) and the approximation error of the state to the ideal
cluster state as = e−2r . It is then straightforward to obtain the relation T ≥ c with
c a constant. Therefore, there is a clear trade-off between the convergence time and
the approximation error; in other words, to obtain a Gaussian state that is very close
to an ideal cluster state by dissipation, we need a lot of time.
6.1.5 Remarks and Further Reading
As remarked at the end of Sect. 6.1.2, if there is no constraint imposed on the system,
then there always exists a linear quantum system that uniquely generates a given tar-
get Gaussian state by dissipation. However, of course in practice several constraints,
which mainly stem from the structure of the system, have to be satisfied. The dis-
cussion at the end of Sect. 6.1.3 is one such example. In the literature, there have
been further progress in this direction; for instance, [13] provides a condition for
dissipation-induced generation of a given pure Gaussian state, under the condition
that the system–field coupling has to be exactly local.
An alternative parameterization of linear quantum systems generating a pure
Gaussian state at steady state was obtained in [14], based on a parameterization of
completely passive linear quantum systems and the group of real symplectic matri-
ces. This result was motivated by the introduction of the notion of quasi-balanceable
linear quantum systems in the context of model reduction of linear quantum systems
by balanced truncation [15].
Lastly, we remark that the dissipation-induced Gaussian state considered here is
a system’s internal state, such as an intracavity state in optics and a motional state
of a nano-mechanical oscillator. That is, it is not the state of an external bosonic
field. Thus, in order to perform an optical quantum information processing with the
generated Gaussian state, it must be extracted to the outside; for this purpose, the
method developed in [16] can be used. Note that if a desired Gaussian cluster state is
generated and extracted to the outside, then universal quantum computation [2] can
in principle be carried out, by employing some non-Gaussian operation such as the
cubic-phase gate or photon counting on the extracted Gaussian state. On the other
hand, if the goal is precision measurement for quantum metrology, we do not need to
extract the dissipation-induced internal Gaussian state to the outside. For instance, a
spin-squeezed state of a large atomic ensemble can be directly used for ultra-precise
magnetometry [17].
6.2 Enhancing Continuous-Variable EPR Entanglement
Recall the non-degenerate parametric amplifier (NOPA) introduced in Sect. 1.5.2. It

is a device that is used to generate EPR entanglement (two-mode squeezed states)
between two of its output beams η1 and η2 in Fig. 1.3. Here, we will show that
two identical NOPAs can be interconnected in a coherent feedback configuration to
enhance the two-mode squeezing while achieving this in a power-efficient manner.
This is inspired by earlier theoretic [18–20] and experimental results [20, 21] that
have shown that single-mode squeezing (recall Sect. 1.5.3) can be improved and/or
spectrally shaped by employing static feedback (using static optical devices) [18, 19,
21] or dynamic coherent feedback [20]. The exposition below largely follows [22].
6.2 Enhancing Continuous-Variable EPR Entanglement 213
Fig. 6.2 Dual-NOPA (G 1 and G 2 ) coherent feedback network. The system contains transmission
time delays τ and transmission losses represented by beam splitters with vacuum noise inputs ξBS,1
and ξBS,2 . This figure does not show output fields associated with losses, which are not observable.
Figure adapted from [22]. With permission of Springer
Consider the coherent feedback network shown in Fig. 6.2, consisting of two
identical NOPAs labeled G 1 and G 2 , Following Sect. 1.5.2, each NOPA G j has two
resonator modes a j and b j that are non-degenerate.
The two modes are coupled by
an interaction Hamiltonian H j = 2ı a ∗j b∗j − a j b j , where is the effective pump
amplitude. Each NOPA has four input fields. The mode a j is coupled to the incoming
√
√ and additional noise ξloss, j,1 via the coupling operators L j,1 = γa j and
field ξin, j,1
L j,3 = κa j , respectively. In a similar fashion, the mode b j is coupled √ to the
√
input ξin, j,2 and noise ξloss, j,2 via the operators L j,2 = γb j and L j,4 = κb j ,
respectively. The constant γ is the decay rate (coupling coefficient) of the mirror
through which a mode interacts with its input, and κ is the coupling coefficient of a
mode to the additional noise, assumed to be the same for a j and b j . The noises ξloss, j,1
and ξloss, j,2 represent channels through which photons created in modes a j and b j
are lost during the interaction, respectively. The outputs corresponding to these two
sources of loss are typically not observable (photons once lost are lost forever).
Due to the separation of G 1 and G 2 , there are transmission losses and time delays
present in the quantum network of Fig. 6.2. For each path, the transmission loss is
effectively modeled by a beam splitter with (positive-valued) transmission rate α and
reflection rate β satisfying α2 + β 2 = 1. The rates are determined by the distance
traversed in a path [23]. The beam splitters have a port for additional vacuum noise
inputs ξBS,1 and ξBS,2 (see Fig. 6.2) that represent channels which carry photons away
as losses. For the NOPAs, a reference value of γr = 7.2×107 Hz is used for the decay
rate of the mirrors, following the values reported for the experiment in [21]. Since
the pump amplitude and coupling coefficient γ are both adjustable quantities, we
define = xγr Hz and γ = γr /y Hz, where x and y are dimensionless parameters
satisfying 0 < x, y ≤ 1 (in practice, y cannot be too small so that multiple resonance
frequencies of the cavity are not excited). The dynamics of the dual-NOPA coherent
feedback network is given by

γ+κ √ √
ȧ1 (t) = − a1 (t) + b1∗ (t) − γξin,1,1 (t) − κξloss,1,1 (t) ,
2 2

γ+κ √
ḃ1 (t) = − b1 (t) + a1∗ (t) − αγb2 (t − τ ) − α γξin,1,2 (t − τ )
2 2
√ √
−β γξBS,2 (t) − κξloss,1,2 (t) ,

γ+κ √
ȧ2 (t) = − a2 (t) + b2∗ (t) − αγa1 (t − τ ) − α γξin,2,1 (t − τ )
2 2
√ √
−β γξBS,1 (t) − κξloss,2,1 (t) ,

γ+κ √ √
ḃ2 (t) = − b2 (t) + a2∗ (t) − γξin,2,2 (t) − κξloss,2,2 (t) , (6.16)
2 2
with outputs
√ √
ηout,1,2 (t) = γb1 (t) + α γb2 (t − τ ) + αξin,2,2 (t − τ ) + βξBS,2 (t) ,
√ √
ηout,2,1 (t) = γa2 (t) + α γa1 (t − τ ) + αξin,1,1 (t − τ ) + βξBS,1 (t) .
(6.17)
We are primarily interested in the entanglement between these two output Gaussian
fields. In particular, we would like to investigate how transmission losses, amplifi-
cation losses, and time delays impact the system’s stability and the entanglement
between the two output fields. Let Alice and Bob be two spatially separated inter-
acting parties. The output field ξout,1,2 can be on Alice’s side, while the other output
field ξout,2,1 can be at Bob’s location. Thus, Alice and Bob can share entanglement
over a distance via the given quantum network. Define the quadratures
z = (a1 , a1 , b1 , b1 , a2 , a2 , b2 , b2 ) ,
q p q p q p q p
q p q p q p q p q
ξ = (ξin,1,1 , ξin,1,1 , ξin,2,2 , ξin,2,2 , ξloss,1,1 , ξloss,1,1 , ξloss,1,2 , ξloss,1,2 , ξloss,2,1 ,
ξloss,2,1 , ξloss,2,2 , ξloss,2,2 , ξBS,1 , ξBS,1 , ξBS,2 , ξBS,2 ) .
p q p q p q p
Then, we have
ż (t) = Az (t) + Bξ (t) ,

q q
ηout,1,2 (t) + ηout,2,1 (t) = C1 z (t) + D1 ξ (t) ,
p p
ηout,1,2 (t) − ηout,2,1 (t) = C2 z (t) + D2 ξ (t) , (6.18)
where
⎡ ⎤
− γ+κ
2
0 2
0 0 0 0 0
⎢ 0 − γ+κ 0 −2 0 0 0 0 ⎥
⎢ 2 ⎥
⎢ γ+κ
0 − 2 0 0 0 −αγ 0 ⎥
⎢ 2 ⎥
⎢ 0 −2 0 − γ+κ 0 0 0 −αγ ⎥
⎢
A=⎢ 2 ⎥, (6.19)
⎢ −αγ 0 0 0 − γ+κ
2
0 2
0 ⎥ ⎥
⎢ 0 −αγ 0 0 0 − γ+κ 0 −2 ⎥
⎢ 2 ⎥
⎣ 0 0 0 0 0 − γ+κ 0 ⎦
2 2
0 0 0 0 0 −2 0 − γ+κ
2
⎡ ⎤
√
0
⎢ γ √0 0 0 0 0 0 ⎥
⎢ 0 γ 0 0 0 0 0 0 ⎥
⎢ √ √ ⎥
⎢ 0 0 α γ 0 0 0 β γ 0 ⎥
⎢ √ √ ⎥
⎢ 0 0 0 α γ 0 0 0 β γ⎥
B=⎢ √
⎢α γ 0

√
√

⎥,
⎢ 0 0 κI8 β γ √
0 0 0 ⎥
⎥
⎢ 0 α √γ 0 0 0 β γ 0 0 ⎥
⎢ √ ⎥
⎢ 0 0 γ 0 0 0 0 0 ⎥
⎣ √ ⎦
0 0 0 γ 0 0 0 0

√
C1 = γ α01010α0 ,
√
C2 = γ 0 −α 0 1 0 −1 0 α ,

D1 = α 0 α 0 01×8 β 0 β 0 ,

D2 = 0 −α 0 α 01×8 0 −β 0 β .
Define H j [ıω] = C j (ıω I − A)−1 B + D j ( j = 1, 2). The two-mode squeezing

spectra V± for the two output fields, introduced in Sect. 1.5.2, can be computed as in
the case of a single NOPA using the formulas (1.33) and (1.34), with A, B, Ci , and
Di as defined above for i = 1, 2.
6.2.1 Stability Condition
The dual-NOPA coherent feedback system is required to be stable to operate properly.

The next theorem provides a general stability condition for the system as a function
of the parameters x, y, κ, and α.
Theorem 6.2 The coherent feedback interconnection of two NOPAs is stable if and
only if
κ
x y < −α + α2 + (1 + y )2 (6.20)
γr
with 0 < x, y ≤ 1.
Proof The determinant of the matrix A given in (6.19) is

⎧ 2 ⎫2
⎨ γ + κ 2 2
α 2 2 2⎬
γ
det(A − λI ) = +λ − − .
⎩ 2 4 4 ⎭
Replacing , γ with = xγr , γ = γr /y and solving the inequality

γ+κ 2 α γ
max( {eig(A)}) = − + + <0
2 4 2
we obtain the theorem.
Thus, for the same decay rate γr for the mirrors in a NOPA, in the ideal lossless
scenario the dual-NOPA coherent feedback system has a lower stability threshold
than a single NOPA, compare (6.20) with the condition 0 < x y < 1 for a single
NOPA.
6.2.2 The Ideal Lossless Case
In the lossless scenario with α = 1 and κ = 0, from (1.33) and (1.34), we find
2
V± (ıω) = (P(ıω)∗ P(ıω) + Q(ıω)∗ Q(ıω)
T (ıω)∗ T (ıω)
+R(ıω)∗ R(ıω) + S(ıω)∗ S(ıω)), (6.21)
where
P(ıω) = α (2ıω + κ)4 − 2[α 2

+ 1 + α2 γ + αγ 2 ] (2ıω + κ)2
−4 1 − α2 γ 2 (2ıω + κ) + α 4
+ 2α 2 γ 2 + 2 1 + α2 3
γ
−2 1 + α γ + αγ ,
2 3 4
√
Q(ıω) = 2 γκ[−α (2ıω + κ)3 − (αγ + ) (2ıω + κ)2
+ α 2 + αγ 2 − 2 1 − α2 γ (2ıω + κ)
+ 3 + αγ 3 − 1 + 2α2 γ 2 + α 2 γ],
√
R(ıω) = 2 γκ[− (2ıω + κ)3 − (3γ + α ) (2ıω + κ)2 + 2
− 3γ 2 (2ıω + κ)
+α 3
+ α γ 2 + 1 + 2α2 2
γ − γ 3 ],
S(ıω) = β[(2ıω + κ)4 + 2γ (2ıω + κ)3 − 2 2
+ α γ (2ıω + κ)2
−2 γ 3 +γ (2ıω + κ) + 4 + 2α 3 γ + 2α γ 3 − γ 4 ],
2
2
T (ıω) = 2 − (2ıω + κ + γ)2 − 4α2 2 γ 2 .
Recall from Sect. 1.5.2 that we will use the two-mode squeezing spectra V± (0) at
ω = 0 as a figure of merit for entanglement. We have the following result.
Theorem 6.3 In the absence of transmission and amplification losses, the two-mode
squeezing spectra of the two outputs in the dual-NOPA coherent feedback system are
2
1 − 2x y − x 2 y 2
V± (0) = 2 . (6.22)
1 + 2x y − x 2 y 2
For a fixed 0 < x ≤ 1,√the two-mode squeezing V± (0) at ω = 0 decreases to 0 as y

approaches the value 2−1x
, the threshold of stability, from below.
Proof When α = 1 and κ = 0, the condition of stability (6.20) becomes x y <

√
2−1. Based on (6.21), we can easily get (6.22), which
√ is a monotonically decreasing
function with respect to x y over the interval (0, 2 − 1), under which the stability
condition of√Theorem 6.2 is satisfied. It is straightforwardly verified that V± (0) goes
to 0 as y ↑ 2−1
x
.
Thus, the dual-NOPA coherent feedback displays a behavior that is analogous to that
of a single NOPA operating below threshold, namely that the two-mode squeezing is
maximized as the pump amplitude is increased to the threshold value of instability.
For the rest of this section, all two-mode squeezing spectra V± (ıω) will be
presented in log–log plots. The squeezing spectra are thus presented in dB unit,
V± (ıω)(dB) = 10 log10 V± (ıω) and V+ (ıω) + V− (ıω)(dB) = 10 log10 (V+ (ıω) +
V− (ıω)). The more two-mode squeezing there is at a frequency ω, the more negative
the plot of V± (ıω)(dB) at that frequency. Perfect two-mode squeezing at this fre-
quency corresponds to V± (ıω)(dB) = −∞. Also, since equations are expressed in a
rotating frame at half of the common pump frequency ω p of the NOPAs, 0 frequency
corresponds to ω p /2, while the “low-frequency” region in these plots corresponds
to positive frequencies around ω p /2.
Figure 6.3 shows the two-mode squeezing spectra V± (ıω) (dB) and V+ (ıω) +
V− (ıω)(dB) versus the frequency ω for x = 0.4 and y = 1 and ignoring amplification
and transmission losses (black solid lines). From (1.32), there is entanglement present
when V+ (ıω) + V− (ıω)(dB) < 10 log10 6 dB = 6.0206 dB. It can be seen that in this
ideal lossless case, V± (ıω)(dB) is about −29 dB and V+ (ıω) + V− (ıω)(dB) has a
practically constant value of around -26 dB for frequencies ω roughly up to 105 rad/s.
6.2.3 Effect of Losses
The presence of transmission and amplification losses inevitably limits the two-mode
squeezing that can be produced by the dual-NOPA network. Here, we will illustrate
how the losses affect the two-mode squeezing while neglecting time delays in the
network. The effect of time delays will be discussed later on.
Fig. 6.3 Log–log plots of V± (ıω) (left) and V+ (ıω) + V− (ıω) (right), when α = 1 (black solid
line), α = 0.97 (red dashed line), α = 0.95 (blue dash-dot line), κ = 0, x = 0.4, y = 1. Figure
adapted from [22] (color figure online). With permission of Springer
Figure 6.4 shows the two-mode squeezing spectra when amplification losses are
present, but there are no transmission losses and time delays. Based on [21], it is
6
assumed that κ ∝ | | with a reference value of κ = 3×10
√
2
for = 0.6 × γr . Therefore,
6
κ = M , with M = √ 3×10
2×0.6×γr
. The figure shows how V and V± have higher values in
5
the frequency range up to 10 rad/s compared to ideal case (black solid line), and this
value steadily worsens as κ increases. Figure 6.5 illustrates the two-mode squeezing
spectra V± (ıω) when both transmission and amplification losses are present. It can
be seen that entanglement becomes worse when both losses are present compared to
when only one of them is present.
6.2.4 Comparison with Conventional Schemes
We will now compare the EPR entanglement generated at two spatially separated
locations, say A (Alice) and B (Bob), by the dual-NOPA coherent feedback system
to that of a single NOPA or two-cascaded NOPAs1 placed midway between A and
B, say, at C (Charlie), when there are no amplification losses in any of the NOPAs
(κ = 0) and transmission delays are neglected (see Fig. 6.6a, b, and c1). The purpose
is to assess how the presence of transmission losses as the sole imperfection in the
1 See [24, 25] for experiments on the entanglement produced by cascading two or more NOPAs.
Fig. 6.4 Log–logplots of V± (ıω) (left) and V+ (ıω)+V

− (ıω)
(right), when κ = 0 (blacksolid line),

6 6 6
κ = 0.1 3×10√ x (red dashed line), κ = 0.2 3×10√ x (blue dot-dash line), κ = 3×10√ x
0.6× 2 0.6× 2 0.6× 2
(pink dotted line), α = 1, x = 0.4, y = 1. Figure adapted from [22] (color figure online). With
permission of Springer
Fig. 6.5 Log–log plots of V±(ıω) when α = 1 (left), α = 0.97 (middle), and α = 0.95 (right),
6
κ = 0 (black solid line), κ = 3×10√ x (red dashed line), x = 0.4, y = 1. Figure adapted from
0.6× 2
[22] (color figure online). With permission of Springer
Fig. 6.6 a Entangled pairs generated by a single NOPA. b Entangled pairs generated by two-
cascaded NOPAs. c1 and c2 Entangled pairs generated by a dual-NOPA coherent feedback system. In
configuration (c1), the NOPAs are distributed in separate locations (A and B), while in configuration
(c2) both NOPAs are in one location, at C. Figure adapted from [22]. With permission of Springer
system will impact the performance of these different schemes. For fairness of the
comparison, it will be made for the situation where the total pump power consumed
by all three systems is the same.
For the coherent feedback (CFB) system with the NOPAs placed at the endpoints
(see Fig. 6.6c1), the transmission rate is α, and for each NOPA, we set y = 1, the
pump power CFB = xγ√ r , and decay rate of mirror γ
CFB
= γr , where x is a real
number between 0 and 2 − 1 (to guarantee stability), and γr is, as before, the
reference decay rate γr = 7.2 × 107 Hz. For the single NOPA and the two-cascaded-
NOPA systems located √ at C, the transmission rate along each transmission channel
√
for the output signal is α (see Fig. 6.6a, b). The higher transmission rate α for
these two schemes is due to the fact that the output signals only have to√travel half
of the distance between A and B. For a single NOPA, we set single = 2xγr and
γ single = γr , while for the two-cascaded-NOPA system, we set cascaded = xγr and
γ cascaded = γr for each NOPA. Therefore, all three systems use the same total pump
power, given by P = 2x 2 γr2 . We choose x so that in the absence of any losses,
the two-mode squeezing spectra of the CFB system at ω = 0 are V±CFB (0) = 2e−r ,
where r ∈ [0, ∞) denotes the degree of squeezing; the more two-mode squeezing, the
larger the value of r . We compare the systems for three values of r , r = − ln(0, 1),
r = − ln(0.05), and r = − ln(0.01). The results are shown in Fig. 6.7. It can be
seen that for each setting of r , above a certain threshold value of the transmission
rate α, the CFB scheme produces more two squeezing than the other two schemes.
However, below this threshold value, the two-mode squeezing is worse than the other
two schemes. We conclude that when all three schemes consume the same amount
of total pump power, the CFB scheme is advantageous over the two other schemes
when transmission losses are not too large.
Rather than placing the two NOPAs at spatially separated locations A and B in a
decentralized configuration, it is also possible to have both NOPAs located at C, as in
configurations (a) and (b). This yields the centralized dual-NOPA configuration (c2)
Fig. 6.7 Plots of V± (0) (dB) as a function of α, when all the systems consume the same total pump
power (P = 2x 2 γr2 ) and κ = 0, with respect to a single NOPA (black dashed line), two-cascaded-
NOPA system (blue dotted line), and a dual-NOPA coherent feedback system in configuration (c1)
(red solid line). We set values of x so that V CFB (0) = 2e−r with r = − ln(0.1) (left), r = − ln(0.05)
(middle), and r = − ln(0.01) (right), when α = 1 (color figure online)
Fig. 6.8 Plots of V± (0) (dB) as a function of α and power, P CFB = P cascaded = P single , when
κ = 0, with respect to a single NOPA (dashed line), two-cascaded-NOPA system (dotted line), and
a dual-NOPA coherent feedback system in configuration (c2) (solid line)
in Fig. 6.6. In all of the centralized configurations, one output beam will be sent from
C to A and the other from C to B; thus, both beams travel a distance
√ that is half of that
traveled in configuration (c1), with a transmission rate of α. Figure 6.8 compares
the entanglement generated from the three configurations (a), (b), and (c2), for the
same values of α and total pump power. The figure shows that the coherent feedback
network in the centralized configuration (c2) generates more two-mode squeezing
than the other two configurations for all 0 ≤ α ≤ 1 and as the parameter x ranges in
value from 0.01 to 0.414 (thus within the stability threshold).
Finally, we compare the dual-NOPA coherent feedback system in the central-
ized configuration (c1) and decentralized configuration (c2), when both configura-
tions consume the same amount of pump power and ignoring amplification losses.
Figure 6.9 shows the entanglement generated by the two configurations for various
Fig. 6.9 Plots of V± (0) (dB) as a function of α, for 0.93 ≤ α ≤ 1 and κ = 0, r = − ln(0.1) (left),
r = − ln(0.05) (middle), r = − ln(0.01) (right), with respect to a dual-NOPA coherent feedback
system placed in the middle (black dashed line) and the two NOPAs distributed at two endpoints
(red solid line). Reprinted from [22] (color figure online). With permission of Springer
Fig. 6.10 Plots of V± (0) (dB) as a function of α, for 0 ≤ α ≤ 1, κ = 0, r = − ln(0.1) (left),

r = − ln(0.05) (middle), r = − ln(0.01) (right), with respect to a dual-NOPA coherent feedback
system placed in the middle (black dashed line) and the two NOPAs distributed at two endpoints
(red solid line) (color figure online)
values of transmission rate α. It can be seen that configuration (c1) has a noticeably
higher degree of EPR entanglement compared to configuration (c2) for higher trans-
mission rates α, but this diminishes as r is increased. Figure 6.10 extends Fig. 6.9 to
compare the two systems over the whole range of 0 ≤ α ≤ 1. It can be seen that the
centralized configuration (c2) produces better entanglement for lower α.
To summarize, when the same total pump power is consumed, and amplification
losses and time delays are neglected, the centralized coherent feedback configuration
(c2) is superior to the single NOPA and two-cascaded-NOPA system for any value
of the transmission rate α, while the distributed coherent feedback configuration
(c1) beats the centralized configuration (c2) for higher transmission rates α. Thus,
for lower transmission losses, one should employ a distributed coherent feedback
network for entanglement generation; for higher transmission losses, one should
switch from the distributed to a centralized coherent feedback scheme.
The results in this section indicate how feedback can, in certain situations, pro-
vide a performance advantage for entanglement generation, even when there are
Fig. 6.11 a Quantum repeater structure with two single NOPAs. There are two endpoints (A and
B) and a repeater station that has two nodes (middle C). Each node contains a quantum memory
that can store a quantum state. NOPA1 and NOPA2 create two pairs of continuous-mode entangled
photons between endpoints and repeater nodes (blue pair and green pair). The two outer quantum
memories can be entangled via entanglement swapping. b Quantum repeater structure with two
dual-NOPA systems. NOPA1 and NOPA2 and NOPA3 and NOPA4 are connected to get entangled
pairs between two points A and C (green nodes) and C and B (blue nodes), respectively. A quantum
repeater in the middle (C) is used to generate entanglement between A and B. Figure adapted from
[22] (color figure online). With permission of Springer
losses present in the transmission channels. A potential application of the dual-NOPA

scheme is to extend the distance for entanglement distribution in a power-efficient
manner. As suggested in Fig. 6.11b, for the same given amount of total pump power,
a dual-NOPA system including NOPA1 and NOPA2 could potentially transport the
same amount of entanglement over a distance L2 that is longer than the distance
L1 achievable by a single NOPA1 in Fig. 6.11a. Given the same distance between
the two endpoints A and B in this figure, the coherent feedback scheme could thus
potentially reduce the number of quantum repeaters required compared to the system
consisting of a single NOPA between every two quantum repeaters.
6.2.5 Effect of Nonzero Transmission Delays
In this section, we now investigate the stability and entanglement generation of the
coherent feedback configuration (c1) in the presence of non-negligible transmission
time delays. We assume that the loss rate across the transmission channel is 0.2
dB/km, which is typical in optical fiber [26]. At this loss rate, α = 0.97 corresponds
to a distance of 1.3230 km and a time delay of τ = 4.41 × 10−6 s, while α = 0.95
corresponds to a distance of 2.2276 km and a time delay of τ = 7.4255 × 10−6 s.
We note that to model interconnections with time delays, it is necessary to first
return to the original model of the NOPA in the non-rotating frame given by (1.27)
and (1.28). Interconnections including time delays are then made to obtain a time-
varying QSDE for the interconnected model, before going to a rotating frame for the
entire interconnected system to remove the time-varying quantities. In the rotating
frame, the time delays will introduce addition phase shifts of eıω p τ /2 , where ω p is
the pump frequency, at the outputs of each NOPA. We assume that these time delay-
induced phase shifts are canceled by placing a constant phase shift of e−ıω p τ /2 at the
output of each NOPA.
To analyze the stability of a linear system with time delays, we can employ an
eigenvalue-based approach as detailed in [27]. This approach has been implemented
in the freely available DDE-BIFTOOL toolbox for MATLAB [28, 29], which is used
to compute the eigenvalues of the coherent feedback network with time delays. The
calculation by DDE-BIFTOOL shows that all roots of the characteristic equation
for the coherent feedback system with time delays have negative real parts when
6
α = 0.97, 0.95, x = 0.4, y = 1, and κ = 3×10 √ x. Therefore, for these parameter
0.6 2
values, the dual-NOPA coherent feedback configuration (c1) is stable in the presence
of the time delays.
The two-mode squeezing spectra V± (ıω) in the presence of time delays is again
given by the identities (1.33) and (1.34). Since the quantum Langevin equations of the
network are linear, the non-rational transfer functions (due to the time delays) H1 [s]
and H2 [s] appearing in (1.33) and (1.34) can be numerically computed in MATLAB,
from which V± (ıω) can then be computed. Figure 6.12 shows the EPR entanglement
generated in the presence of time delays for a specific setting of the parameters of
the dual-NOPA coherent feedback system in configuration (c1). It can be seen that
the time delays have caused the frequency range (bandwidth) of entanglement to
be reduced to less than 105 rad/s. Moreover, the degree of entanglement can also
be observed to decrease, with a larger decrease for longer time delays. However,
the latter effect is not due to the delays per se but is a consequence of the losses
experienced by signals as they propagate through the transmission channels. Longer
delays are associated with longer distances and hence with larger losses and a larger
drop in the entanglement. Finally, in Fig. 6.12, we can also see erratic looking sharp
peaks and dips in the squeezing spectra. We note that this is not a feature which
is peculiar to the dual-NOPA scheme but is common in the frequency response of
systems with internal time delays (see, e.g., [30, p. 182]).
Fig. 6.12 Log–log plots of V± (ıω) (left) and V+ (ıω) + V− (ıω) (right), when τ = 0 (black solid
line), τ = 4.41 × 10−6 (red dashed line), τ = 7.4255 × 10−6 (blue dash-dot line), x = 0.4, y = 1,
6
κ = 3×10 √ x. Decrease in entanglement (increase in dB value of V± (0)) for higher delays is not due
0.6 2
to the delays per se (see the text for details). Figure adapted from [22] (color figure online). With
permission of Springer
6.2.5.1 Further Reading
The properties of the dual-NOPA coherent feedback scheme have been further ana-
lyzed in [31–33]. This includes studying the effect of phase shifts on the transmission
channels on the performance of the scheme [31] and establishing some local optimal
properties [33].
Related experiments have been reported that demonstrate how coherent feedback
can enhance EPR entanglement in continuous-variable systems [34, 35]. However,
these schemes consider the use of a static coherent feedback as in [18, 19, 21] to
enhance EPR entanglement rather than interconnecting two NOPAs together in a
coherent feedback network to improve entanglement.
The dual-NOPA scheme can be extended to a multiple-NOPA scheme involv-
ing multiple NOPAs G 1 , G 2 , . . ., G N , connected in linear coherent feedback chain
equipped with adjustable phase shifters at two of its outputs, ηout,1 and ηout,2 , as
depicted in Fig. 6.13 and analyzed in [36, 37]. Similar conclusions to the dual-NOPA
scheme emerge in the multiple-NOPA scenario, in that the same degree of two-mode
squeezing between the outputs ηout,1 and ηout,2 can be achieved with much lower
pump power and that the scheme produces a stronger degree of two-mode squeezing
with more NOPAs in the chain for the same amount of total pump power and when
only transmission losses are present [36]. Explicit formulae for EPR entanglement in
Fig. 6.13 A linear coherent feedback chain of multiple NOPAs. Figure adapted from [36]
the multiple-NOPA chain, in the idealized lossless scenario and infinite bandwidth
limit, have been derived for an arbitrary number of NOPAs in the chain [37]. The
lossless scenario sets the ultimate EPR entanglement that can be achieved by this lin-
ear chain of NOPAs, by selecting appropriate quadratures of the output fields with a
suitable choice of the output phase shifters, while the infinite bandwidth limit simpli-
fies the analysis but gives an accurate approximation to the EPR entanglement at low
frequencies of interest (i.e., frequencies around half the common pump frequency of
the NOPAs).
6.3 Force Sensing and Back-Action Evasion
In this section, we move to another very important research field in quantum informa-
tion science, i.e., quantum metrology. The goal of this subject is, in a broad sense, to
detect a very weak (quantum-level) signal such as a gravitational wave force [38–41]
and an ultra-small deviation of frequency in an atomic clock [42], with the full use
of quantum mechanical properties such as squeezing and entanglement. In particu-
lar, as mentioned in Sect. 1.5.4, the linear force sensing problem, which is a special
subclass of quantum metrology, can be formulated within the framework of systems
and control theory. More specifically, a linear open quantum system coupled to a
probe field functions as a sensor to detect the signal, where the information about the
signal can be extracted by measuring the probe output field. However, as immedi-
ately expected, this probe field unavoidably introduces a noise in the input port and
eventually degrades the signal-to-noise ratio. In particular, the so-called back-action
(BA) noise places the standard quantum limit (SQL) on the detection sensitivity.
Fortunately, as has been proven in a number of papers, the SQL can be beaten by
some back-action evasion (BAE) techniques, which achieve better detection sensitiv-
ity over SQL. The aim of this section is to describe a general force sensing problem
within the framework of linear systems and control theory [43, 44] and demonstrate
a coherent feedback-controlled system that achieves BAE.
6.3 Force Sensing and Back-Action Evasion 227
6.3.1 Back-Action Evasion and the Standard Quantum Limit
First, let us recall the opto-mechanical oscillator described in Sect. 1.5.4, to see the
meaning of SQL and why BAE is necessary. The simplified system dynamics, which
is obtained by adiabatically eliminating the cavity mode, is given by (1.40); again, f
is the force we want to detect, m is the mass, and ωm is the resonance frequency
√ of the
oscillator. From this dynamics, we find that the measurement output η p = λq1 −ξ p
brings some information about f . However, notably, η p contains both of the probe
input noise ξ q and ξ p ; the unavoidable noise ξ p is called the shot noise, while ξ q
implicitly appears in the measurement output η p through the dynamics of q1 and this
is the BA noise. This fact can be explicitly seen in the Laplace domain:
√
λ √ q √
η [s] =
p
λξ [s] + 2 f [s] + ξ p [s],
m(s 2 + ωm )
2
showing that η p is indeed affected by both the noise quadratures. Now, instead of the
force magnitude f , we focus on the change of oscillator’s position induced by f ; let
us denote this change by g, and then in the Fourier domain s = ıω, this is related to
f as f [ıω] = −m Lω 2 g[ıω], where L is the optical path length of the cavity. Hence,
under the assumption ω ωm , the normalized signal containing g is given by
√
η p [ıω] λ 1
η̃ [ıω] = √
p
= g[ıω] − √ ξ q [ıω] + √ ξ p [ıω].
2λL 2m Lω 2 2λL
This shows that the BA noise ξ q becomes dominant in the low-frequency region
ω ≈ 0, while the shot noise ξ p sets the constant noise floor; also, we see that for
precise detection, both the mass m and the optical path length L should be large,
which is the reason why a gravitational interferometer of a kilometer-scale size was
constructed in order to detect a gravitational wave [39, 45]. Now, the noise power of
η̃ p is bounded from below by the following SQL:
λ 1
S(ıω) = |η̃ p [ıω] − g[ıω]|2 = |ξ q [ıω]|2 + |ξ p [ıω]|2
2m L ω
2 2 4 2λL 2

|ξ q [ıω]|2 |ξ p [ıω]|2 1
≥2 ≥ = SSQL (ıω). (6.23)
4m L ω
2 4 4 m L 2 ω2
The last inequality comes from |ξ q [ıω]|2 |ξ p [ıω]|2 ≥ 1, i.e., the Heisenberg
uncertainty relation. (Note that rigorously, this should be defined in terms of the
power spectral density.) From the above relation, it is clear that SQL appears because
the measurement output η p contains the BA noise ξ q in addition to the unavoidable
shot noise ξ p . Thus, for high-precision detection of g, a special system configuration
should be devised so that the BA noise ξ q is evaded and η p is free from ξ q ; this is
the meaning of BAE. In fact, if BAE is achieved, then by injecting a ξ p -squeezed
probe light field, we can possibly reduce the noise power below the SQL and may
have chance to detect g with better accuracy.
6.3.2 System Theoretical Characterization of BAE
Here, we consider a general linear quantum system given by (2.5):
d x(t) = Ax(t)dt + Bdw(t), dy(t) = C x(t)dt + Ddw(t), (6.24)
and describe the notion of BAE in this general setting [43]. First, let us assume that
the signal to be detected has some correlation with the measurement output ỹ, which
is a part of the quadrature components:
d ỹ = M1 dy = M1 C xdt + M1 Ddw = M1 C xdt + d w̃. (6.25)
Again, the signal is already incorporated in ỹ. Here, M1 is a m × 2m projection

matrix representing which quadrature is to be measured, satisfying M1 Jm M1 = 0
and M1 M1 = Im ; these conditions ensure that all the elements of ỹ(t) are classical
signals commuting with one another as well as with those of ỹ(t ) for all times t, t ,
i.e.,
[ ỹi (t), ỹ j (t )] = 0, ∀i, j, ∀t, t .
Note that w̃ = M1 Dw is a generalized version of shot noise, which must be present

in the measurement output.
Now, the BA noise defined by w̃ = M2 Dw, i.e., observables conjugate to w̃. Here,
M2 is also a m × 2m projection matrix and is chosen so that the matrix [M1 M2 ] is
a symplectic and orthogonal matrix, which as a result leads to
M2 Jm M2 = 0, M2 M2 = Im , M1 Jm M2 = Im ,

M1 M2 = 0, M1 M1 + M2 M2 = I2m . (6.26)
Actually, these conditions ensure the CCR [d w̃(t), d w̃ (t )] = ımin(t, t )dt. In

particular, the last equality M1 M1 + M2 M2 = I2m leads to Dw = M1 w̃ + M2 w̃ .
Hence, the dynamical equation in (6.24) can be rewritten as
d x = Axdt + B D −1 (M1 d w̃ + M2 d w̃ ). (6.27)
We now arrive at the situation of writing the BAE condition in this general setup; for
this, we need that the output ỹ in (6.25) does not contain the BA noise w̃ . That is,
BAE is achieved if and only if the transfer function from w̃ to ỹ is always zero:
BAE: w̃ → ỹ [s] = M1 C(s I − A)−1 B D −1 M2 = 0, ∀s. (6.28)

Moreover, as is well known in systems and control theory, this condition means that
there is no subsystem that is controllable w.r.t. w̃ and observable with respect to ỹ.
This is equivalently represented by the condition
M1 C Ak B D −1 M2 = 0, ∀k ≥ 0. (6.29)
Under this condition, the system Eqs. (6.25) and (6.27) are represented in a trans-
formed coordinate by

x A11 0 x1 B11 0
d 1 = dt + d w̃ + d w̃ ,
x2 A21 A22 x2 B21 B22

x
d ỹ = [ C1 0 ] 1 dt + d w̃.
x2
This implies that the input signal w̃ has no effect on the output ỹ. Finally, note
that even if the above BAE condition (6.28) or (6.29) is satisfied, this by itself does
not necessarily mean that the signal sensitivity is improved; in particular, squeezing
operation on the input field w̃, in addition to the BAE property, is necessary to realize
such an improvement.
6.3.3 Coherent Feedback for BAE
In this subsection, based on the BAE theory formulated in the previous subsection, we
demonstrate a coherent feedback (CFB) control approach for BAE of a mechanical
oscillator driven by an unknown force.
First, we address a general framework for CFB configuration for the linear quan-
tum system
d x(t) = Ax(t)dt + Jn C Jm dw(t), dy(t) = C x(t)dt + dw(t), (6.30)
which is a special case of (2.5) with D = I (see Theorem 2.1). For CFB, at most all
the components of the output field y can be used for feedback, and also, at most all
the components of the input probe field w can be controlled. Here, we take the CFB
control configuration depicted in Fig. 6.14, where the controller dynamics is given
by
d x K = A K x K dt + Jn C1 Jm dw1 + Jn C2 Jm dw2 ,

dy1 = C1 x K dt + dw1 , dy2 = C2 x K dt + dw2 , (6.31)
where n is the number of degrees of freedom of the controller. Note that the controller
has two input–output ports, to avoid the self-interaction of optical fields. Now, y is
connected to the controller’s input w2 , and the controller’s output y1 is connected to w,
Fig. 6.14 A general Plant

configuration of the CFB w y
control. Figure adapted
from [43]
y1 w1
y2 w2
Controller
without involving any measurement process. That is, the CFB control is constructed
by setting
w2 = y, w = y1 . (6.32)
From this condition, the size of C1 and C2 has to be equal, although they are not nec-
essarily of full rank. Combining (6.30), (6.31), and (6.32), we find that the dynamical
equation of the closed-loop system is given by:
d xe = Ae xe dt + Jn+n Ce Jm dw1 , dy2 = Ce xe dt + dw1 , (6.33)
where xe = [x , x K ] , Ae = Jn+n (G e + Ce Jm Ce /2), Ce = [C, C1 + C2 ], and

G C Jm C1 /2 − C Jm C2 /2
Ge = ,
G K + C1 J
m C 2 /2 + C 2 Jm C 1 /2
where denotes the symmetric elements of G e . Here, let us assume that a CFB
controller satisfying C1 + C2 = 0 is implementable. Then, the closed-loop system
(6.33) takes the following form:

A Jn C Jm C1 Jn C Jm
d xe = x dt + dw1 ,
Jn C1 J
mC Jn G K e
0
dy2 = [C 0]xe dt + dw1 . (6.34)
It seems from the structure of this equation that the controller directly couples to
the plant, but not to the external field. Hence, this type of plant–controller coupling
is called a direct interaction. In this sense, (6.34) implies that direct interaction is
realized within the CFB configuration.
Now, let us study the opto-mechanical oscillator from Sect. 6.3.1 as the plant of
interest. To achieve highly precise detection of an unknown force f driving this
system, we aim to apply the above-described CFB control configuration so that the
closed-loop system (6.34) satisfies the BAE condition. The controller is a single-
mode linear quantum system with variable x K = (q3 , p3 ) , with two input fields
(ξ1 , ξ1 ) = dw1 /dt and (ξ2 , ξ2 ) = dw2 /dt. The controller’s system matrices are
q p q p
assumed to satisfy

0 0 −ωm
JC1 J C = , JG K = ,
0 g ωm
which leads to

g 00 −ωm 0
C1 = −C2 = √ , GK = . (6.35)
2γ 10 0 −ωm
We will discuss later about how to physically implement these system matrices. In
this setting, the closed-loop system dynamics (6.34) is given by
d xe = Ae xe dt + Be dw1 + b f dt, dy2 = Ce xe dt + dw1 , (6.36)
where
⎡ ⎤
1/m 0
⎢ −mωm2 ⎥ κ
⎢ ⎥
⎢ 0 ⎥
0 −γ
Ae = ⎢
⎢ κ
⎥ , Be = C ,
⎥
⎢ −γ g ⎥
e
⎣ 0 −ωm ⎦
g ωm

" 0 1 0
Ce = 2γ , b = [0 1 0 0 0 0] .
0 1 0
The output field is (η2 , η2 ) = dy2 /dt; clearly, η2 does not contain any information
q p q
p
about f , and thus, we need to measure η2 , implying that the output signal is given
by "
p p p
η2 = cxe + ξ1 = 2γ[0 0 0 1 0 0]xe + ξ1 . (6.37)
√
Here, we apply the BAE condition (6.29) and find g = κ/ mωm . In fact, with this
q p
choice, the transfer function from the BA noise ξ1 to the output η2 is zero, so we
have √
p 2γκ/m p s−γ
η2 [s] = ξ [s] + f [s].
(s + γ)(s 2 + ωm2 ) 1 s+γ
p p
Thus, by injecting a ξ1 -squeezed light field, which reduces the variance of ξ1 , we
can perform a better detection of f below SQL.
Lastly, let us discuss how to physically implement the above CFB controller. The
point is that the structure of C1 (or C2 ) in (6.35) represents a QND-type interac-
tion between the controller and the field w1 ; for instance, Ref. [2] provides an optical
method to implement this interaction. On the other hand, the matrix G K in (6.35) sim-
ply corresponds to an optical phase shifter. Therefore, the CFB controller designed
here can be realized as a detuned optical cavity coupled to two input–output fields
Fig. 6.15 a Direct (a) (b)

interaction scheme achieving
a2 ( q 1 , p 1) ξ1 ( q 1 , p 1)
BAE for the opto-mechanical
oscillator, proposed by a2
Tsang and Caves [40]. a3 a3
b Equivalent realization via f f
the CFB control. Figure η2
adapted from [43]
η
ξ
via QND interactions, which is depicted in Fig. 6.15b. Note that the set of Eqs. (6.36)
and (6.37) is exactly the same as that of the modified opto-mechanical oscillator
proposed by Tsang and Caves [40], which is shown in Fig. 6.15a.
Synthesizing a system achieving BAE is an old problem, particularly in the field of

gravitational wave detection (see for instance [38–41]). In contrast to their approach
of looking at a specific optical interferometer for gravitational wave detection, the
system theoretic characterization given in Sect. 6.3.2 can potentially yield a gen-
eral approach for synthesizing a system having the BAE property. In fact, in [43]
another coherent feedback controller realizing BAE for the optical interferometer
was proposed, based on the general condition (6.28) or the equivalent (6.29).
In addition to BAE, in [43] system theoretic characterizations of quantum non-
demolition (QND), interactions and decoherence-free subsystems (DFS) are given,
which also provide a concrete coherent feedback controller achieving QND inter-
action and DFS. A crucial fact proven in [43] is that while we can find a coherent
feedback controller achieving BAE, QND, or DFS for a given linear quantum sys-
tem, it is generally not possible to construct a measurement-based feedback con-
troller achieving these three goals. That is, there exists a no-go theorem showing the
distinct superiority of coherent feedback over a measurement-based one, for those
design goals in quantum information science. Again, it should be emphasized that
this result is obtained based fully on a system theoretic approach.
6.4 Quantum Memory with Decoherence-Free Subsystem
A quantum memory is a device that can preserve a quantum state to realize tasks
in computation and communication. Surprisingly, or not surprisingly, systems and
control theory provides a unified methodology for constructing a linear quantum
memory architecture [46]; more precisely, for linear quantum systems, we obtain
6.4 Quantum Memory with Decoherence-Free Subsystem 233
a general procedure for perfect writing and storage of a single-photon or coherent

state, with the full use of the controllability–observability properties and the notion
of zero dynamics. This section is devoted to reviewing this theory, focusing on the
case where the input state is a coherent state, for simplicity.
6.4.1 General Schematic of an Ideal Quantum Memory
Let us begin by describing the general and ideal schematic of a quantum memory.
First of all, the system needs to contain a subsystem that preserves a quantum state
for a long time, which is named here the memory subsystem. During the storage
time, a memory subsystem must be decoupled from other subsystems as well as
all dissipative environments; i.e., it is decoherence-free (DF) [47]. Note that in the
storage stage, the memory subsystem must be disconnected with the transportation
channel that transfers an input state or retrieves the memory state, while in the writing
or reading process, we need to connect the memory subsystem with the transportation
channel. Therefore, the system involves a switching mechanism that opens or closes
the memory subsystem. Concrete examples are for instance an atomic ensemble with
electromagnetic induced transparency (EIT) effect [48–50], a photonic crystal array
[51], and an optical cavity [52].
The above basic function of a general quantum memory system is illustrated in
Fig. 6.16. Let aM denote the mode of the memory subsystem. aB is the mode of the
buffer subsystem, which mediates the optical field and the memory subsystem. As
shown in the figure, these two subsystems can be coupled or decoupled. In the writing
stage (a), we transport an input state of the optical field with mode ξ to the memory
subsystem. Here, let us assume that with the use of a “nice procedure,” the input state
is perfectly transferred to the memory subsystem. Then, the two modes aB and aM
are decoupled, and aM becomes DF. In the storage stage (b), the memory subsystem
preserves the state indefinitely. Finally, in the reading stage (c) at a later time, by
coupling the modes again we can retrieve the state over the output channel with
mode η.
Let us describe the scheme in a more detailed fashion, for the case of linear
quantum systems. The system is given by a completely passive linear quantum system
discussed in Sect. 2.4.1. That is, the vector of system variables a = (a1 , . . . , an )
obeys the following dynamical and output equations:
(a) Writing (b) Storage (c) Reading
aB aB aB
ξ η
aM aM aM
Fig. 6.16 Basic function of an ideal quantum memory. Figure adapted from [46]
ȧ(t) = Aa(t) − C ∗ ξ(t), η(t) = Ca(t) + ξ(t). (6.38)
Again, the system matrices are related to the Hamiltonian and the coupling operator
in the form H = a ∗ a and L = Ca. Also, recall that A = −ı − C ∗ C/2. In
particular, we assume that there is only one external channel, meaning that C is a
1 × n complex row vector; hence, without loss of generality the scattering matrix S
is now set to be S = 1. Now, let aB = (a1 , . . . , am ) and aM = (am+1 , . . . , an ) be
the modes of the buffer and memory subsystems, respectively. Then, in the writing
stage, our task is to devise a procedure (as mentioned above) such that the state of ξ
is perfectly transferred to that of aM . Once this process is completed, the system is
switched to the storage mode, where (6.38) takes the form
∗
d aB (t) AB O aB (t) CB
= − ξ(t),
dt aM (t) O O aM (t) O
η(t) = CB aB (t) + ξ(t). (6.39)
Hence, aM is decoupled from both the input field ξ and the output field η. This
implies that aM is actually DF and its state can be stored. Importantly, (6.39) shows
that in the storage stage, the memory subsystem is uncontrollable and unobservable
with respect to the input and output channels (see Sect. 1.2.2). Thanks to this system
theoretic characterization of the memory subsystem, a convenient necessary and
sufficient condition for such a DF subsystem to exist has been developed in [53].
As a typical setup, we show an atomic ensemble system composed of N -type
atoms confined in an optical cavity [50]. Each atom contains two ground states |s
and |g, and an excited state |e. The transition between |g and |e occurs due to
the coupling of atoms and the cavity mode a1 , with strength g. Also, a classical
field with controllable Rabi frequency ω R induces the transition
N between |s and
|e. Now, define the collective lowering operators σge = k=1 (|ge|)(k) and σgs =
N (k)
k=1 (|gs|) (see 1.42); in the large ensemble limit N √1, these operators√can
be well approximated by annihilation operators as a2 = σge / N and a3 = σgs / N .
The ideal dynamics of a = (a1 , a2 , a3 ) is then given by
⎡ ⎤ ⎡√ ⎤
−κ ig 0 2κ √
da(t) ⎣
= ig 0 iω R ⎦ a(t) − ⎣ 0 ⎦ ξ(t), η(t) = 2κa1 + ξ(t),
dt 0 iω R 0 0
where κ denotes the cavity decay rate. Hence, aB = (a1 , a2 ) represents the buffer
mode and aM = a3 does the memory mode. In fact, when the state transfer from ξ
to a3 is completed, then by setting ω R = 0 the system dynamics has the form (6.39)
and can preserve the transferred state.
6.4.2 The Zero Dynamics Principle
As mentioned above, our task is to develop a procedure that perfectly transfers the
input state of ξ to the state of aM . Now, note that the input state should be brought
over a time temporal pulsed field rather than a continuous-wave one (for instance an
optical field whose amplitude changes in time according to a sinusoidal function),
because obviously the subsystem can absorb only the finite amount of energy. Then,
the problem is to find an appropriate pulse shape for achieving perfect state transfer.
As shown below, the zero dynamics principle gives us a clear solution to this problem.
First, let us assume that the initial state is given by the product of coherent states,
|φ S ⊗ | f F , with f (t) the amplitude of the field coherent state, corresponding to the
pulse function to be determined. Here, the subscripts S and F are used to emphasize
system and field states, respectively. Then, noting that S φ| F f |ξ(t)|φ S | f F = f (t),
we see that the vector of mean amplitudes, m(t) = (a1 (t), . . . , an (t)) with
ai (t) = S φ| F f |ai (t)|φ S | f F , follows
ṁ(t) = Am(t) − C ∗ f (t), f˜(t) = Cm(t) + f (t), (6.40)
where f˜(t) = η(t). Now, for the completely passive linear quantum system (6.38),
the following energy balance identity [54] holds:
# t # t
η ∗ (s)η(s)ds + a ∗ (t)a(t) = ξ ∗ (s)ξ(s)ds + a ∗ (t0 )a(t0 ). (6.41)
t0 t0
Thus, the total energy of the system and the field is preserved for all time. Actually,
combined with (6.40), this equation leads to
# t # t
| f˜(s)|2 ds + a ∗ (t)a(t) = | f (s)|2 ds + a ∗ (t0 )a(t0 ),
t0 t0
where the mean is taken for the state |φ, f . Now, we assume that the system initial
state is vacuum, |φ S = |0 S , which yields a ∗ (t0 )a(t0 ) = 0. Then, in order to
achieve perfect transportation of the energy from the input to the system, we need
that the output field must be vacuum, or equivalently, the amplitude of the output field
is always “zero” because in this case the complete passivity means that the output
is also a coherent state; this is the zero dynamics principle. Therefore, f˜(t) = 0 for
all t ∈ [t0 , t1 ] with t1 , the stopping time of the writing process. Then, the condition
f˜(t) = Cm(t) + f (t) = 0 leads to C ∗ Cm(t) + C ∗ f (t) = 0, and we thus have
1
ṁ(t) = (A + C ∗ C)m(t) = − ı + C ∗ C m(t) = −A∗ m(t). (6.42)
2
This corresponds to zero dynamics, meaning that if the system variable obeys the
dynamics (6.42), the output is always vacuum. Now, the solution of (6.42) is given by
∗
m(t) = e−A (t−t1 ) m 1 , with m 1 a fixed vector. Thus, the condition Cm(t) + f (t) = 0
yields
f (t) = −Cm(t) = −m(t) C = −m −A (t−t1 )

C .
#
1e
Note that the input is transferred in the period t ≤ t1 (writing stage) and f (t) = 0
for t1 ≤ t (storage and reading stages). As a result, we have the following proper
expression:
f (t) = −m 1e
−A# (t−t1 )
C (t1 − t), (6.43)
where (t) is the Heaviside step function taking the value 1 for t ≥ 0 and 0 for
t < 0. Note that the real parts of all the eigenvalues of −A# are strictly positive,
because A is Hurwitz. Thus, f (t) grows exponentially, and hence, it is called a rising
exponential function.
6.4.3 Perfect State Transfer
First, we discuss the concrete procedure to encode (classical) information on the

coherent field state with pulse function (6.43), which is perfectly transferred to the
system; this procedure determines the parameter vector m 1 contained in (6.43). Note
that a coherent field state | f F with temporal pulse shape f (t) can be expressed as
# ∞
B ∗ ( f )−B( f )
| f F = e | F = exp f (t)ξ ∗ (t) − f ∗ (t)ξ(t) dt | F ,
−∞
$∞ $∞
where B( f ) = −∞ f ∗ (t)ξ(t)dt. The power −∞ | f (t)|2 dt is assumed to be finite.
Now, assume that f (t) is characterized, in terms of the orthonormal functions
{γk (t)}k=1,...,n , by
n
f (t) = αk γk (t), (6.44)
k=1
where αk ∈ C represents the classical

$∞ information
to be stored in the memory. Then,
the power of | f F is given by −∞ | f (t)|2 dt = k |αk |2 , and the coherent field state
is described by
% &
| f F = exp αk B ∗ (γk ) − αk∗ B(γk ) | F . (6.45)
k
Next, let us turn our attention to the writing process, which actually transfers the
classical information {αk } to the memory subsystem. For this purpose, first recall
that the system dynamics (6.38) has the solution
# t
A(t−t0 )
a(t) = e a(t0 ) − e At
e−As C ∗ ξ(s)ds.
t0
Since A is Hurwitz, we can take the limit t0 → −∞, which yields

# t1
A# t1
e−A s C ξ ∗ (s)ds,
#
a (t1 ) = −e
#
−∞
where again t1 is the stopping time of the writing process. We now introduce the
following vector of rising exponential functions:
ν(t) = −e−A (t−t1 )

C (t1 − t).
#
(6.46)
Then, a # (t1 ) can be represented by

# ∞

a # (t1 ) = ν(t)ξ ∗ (t)dt = I S ⊗ B ∗ (ν1 ), . . . , I S ⊗ B ∗ (νn ) . (6.47)
−∞
This is a vector of creation operators ak∗ (t1 ) satisfying the canonical commuta-
tion
$ ∞ relation aa ∗ − (a # a ) = I , which means that {νi (s)} are orthonormal;
∗
−∞ νi (t)ν j (t)dt = δi j . Recall now that at the initial time t0 → −∞, the field
input state is the coherent field state (6.45) and the system initial state is the ground
state |0, . . . , 0 S . Then, after the interaction, the whole state at t = t1 is given by
∗ ∗
|(t1 ) = U (−∞, t1 )|0, . . . , 0 S ⊗ e k [αk B (γk )−αk B(γk )] | F
∗ ∗
= U (−∞, t1 )e k [αk B (γk )−αk B(γk )] U ∗ (−∞, t1 )|0, . . . , 0 S | F .
Thus, by setting γk (t) = νk (t) and noting that (6.47) can be expressed as ak∗ (t1 ) =
U ∗ (−∞, t1 )ak∗ (−∞)U (−∞, t1 ) = I S ⊗ B ∗ (νk ), we obtain
∗ ∗
|(t1 ) = e k [αk ak (−∞)−αk ak (−∞)] |0, . . . , 0 S | F
' ∗ ∗
= eαk ak (−∞)−αk ak (−∞) |0, . . . , 0 S | F
k
= |α1 , . . . , αn S | F .
Hence, the system state is transformed to the product of coherent states |αk . That
is, the coherent field state with classical information {αk } encoded in the rising
exponential pulse functions (6.46) can be perfectly transferred to the system.
Now, we describe the ideal memory schematic in a more detailed way. Recall that
the system can tune the memory subsystem to the DF mode in the storage stage or
to the non-DF mode in the other two stages. Hence, after the perfect state transfer
in the writing process has been finished at time t = t1 , the system structure is
instantaneously modified so that the system dynamics takes the form (6.39):
∗
d aB AB O aB CB
= − ξ(t), η(t) = CB aB (t) + ξ(t).
dt aM O O aM O
(a) (b) (c)

t1 t2 t1 t2 t1 t2
5 5 5
2 2 2
4 4 4
f (t) 3 3 3
1 1 1
|0 > |0 >
(d) (e)
t1 t2 t1 t2
5 5
2 2
4 4
3 f˜(t) 3
1 1
|0 >
Fig. 6.17 The perfect memory procedure in a 5-node linear network. Figure adapted from [46]
Clearly, aM constitutes a DF subsystem. Now, for simplicity, we assume n = 5 and

consider the specific case aB = [a1 , a2 ] (the buffer mode) and aM = [a3 , a4 , a5 ]
(the memory mode). Hence, the whole state |α1 , . . . , α5 cannot be stored, but only
its (3, 4, 5) components can be. This means that to achieve the perfect state transfer
and storage, the input field must be prepared to a specific coherent field state | f (t) F
with α1 = α2 = 0 and sent over the rising exponential pulse function (6.46);
more precisely, the input pulse shape should be synthesized by first multiplying the
classical information (α3 , α4 , α5 ) with the basis functions (ν3 (t), ν4 (t), ν5 (t)) and
then generating f (t) = α3 ν3 (t) + α4 ν4 (t) + α5 ν5 (t). In fact, as described before,
the whole state when the writing process is finished is then given by
|(t1 ) = |0, 0 ⊗ |α3 , α4 , α5 ⊗ | F ,
and consequently, by decoupling the buffer and memory subsystems, the state
|α3 , α4 , α5 can be preserved. The above writing and storage processes together
with the reading process are illustrated in Fig. 6.17, and now, those can be explicitly
described as follows. (a) The coherent field state with α1 = α2 = 0 is transferred over
the pulse shape f (t). (b) At time t = t1 , the field state has been perfectly transferred
to the memory subsystem. Then, by switching the system architecture, the memory
subsystem is decoupled. (c) In the storage state in the time period [t1 , t2 ], the trans-
ferred state is preserved. (d) At t = t2 , the memory subsystem is coupled again to
the buffer subsystem and accordingly the output field. (e) The output field with pulse
shape f˜(t), which contains the perfect copy of the information {α3 , α4 , α5 }, can be
retrieved (see [46] for the detailed time evolution of this retrieved state).
In this section, based on the zero dynamics principle, we derived the rising expo-
nential function (6.43) as the desired input pulse realizing perfect state transfer to a
completely passive linear quantum system. This fact can be derived by another sys-
tem theoretic approach as follows. Let f˜[s] = G[s] f [s] be the input–output relation
of (6.40) and z j be a zero of the transfer function G[s], i.e., G[z j ] = 0 holds. Then,
it is known in linear systems and control theory that an input of the form f (t) = e z j t
can yield an output f˜(t) = G[z j ]e z j t = 0 under appropriate conditions. Now, we can
proceed to prove (6.43) as a linear combination of those functions {e z j t }. That is, the
condition for perfect state transfer can be completely characterized by the zero of the
system’s transfer function; this is a special case of the general zero dynamics prin-
ciple. Importantly, as mentioned before, this fact can be extended beyond coherent
states to include single-photon fields, which is an important class of non-Gaussian
states which is of much interest in quantum information science (see [46]). More-
over, we can generalize the theory of quantum memory discussed in this section to
general completely passive linear quantum systems with multiple inputs and outputs,
in which case the notion of zeros is extended to transmission zeros or particularly
blocking zeros [55].
The rising exponential function was found in some specific settings as the desired
pulse function achieving complete absorption of the state by a quantum system
[56–58], and recently, we find several experimental proposals to generate a simple
(monotonous) rising exponential wave packet of a single-photon field [59–62]. How-
ever, the pulse function (6.43) has a complicated non-monotonous form in the general
setup (see [46]), and producing such a complicated pulse function in experiments is
challenging. In this case, an alternative approach proposed in [49, 50, 80] can be
used for the purpose of high-efficiency state transfer. In this approach, an additional
open-loop control input is applied on the system so that an input single-photon field
with a given pulse shape is transferred to the system more efficiently, in the sense
that the energy of the internal system (set to the ground state at the initial time) is
maximized.
6.5 Robust Quantum Amplification via Coherent Feedback
Let us begin with a classical amplifier, which is doubtless a key component incorpo-
rated in almost all current electrical devices. The basic function of an autonomous
amplifier is to transform an input signal u to y = Gu with gain G > 1. However,
usually the system parameters contained in G are fragile, and thus, the output y can
change easily and eventually become distorted. Fortunately, this serious issue was
finally resolved in 1927 by Black [63, 64]. The central idea is in the use of feedback
as shown in Fig. 6.18; that is, an autonomous amplifier (plant) is combined with a
passive system (controller) in a feedback interconnection. Then, the output of the
closed-loop system is given by
Fig. 6.18 Classical feedback Plant

amplification scheme: G is +
the gain of an autonomous G
u -
amplifier, and K is the gain
of a passive controller. K
y
Reprinted with permission
from [65]. © (2016) by the Controller
American Physical Society
G
y = G (fb) u, G (fb) = ,
1 + GK
where K < 1 is the attenuation level of the passive controller. Then, a large gain
G 1 leads to G (fb) ≈ 1/K > 1. That is, the gain of the controlled system
is determined only by the controller’s parameters. This is a very important result,
because a passive device such as a resistor is very robust and its parameters contained
in K can be made to be very stable. Thus, thanks to the feedback, the controlled system
functions as a robust amplifier.
Now, a pertinent quantum version of the above-described commercial classical
amplifier is the quantum phase-preserving amplifier (PPA). This amplifier plays
a central role in, for instance, quantum communication and weak-signal detection
[66–70], but as in the classical case, it is always fragile with respect to the system
parameters. In [65], a general theory was provided to resolve this issue; that is,
a coherent feedback with a completely passive linear quantum system makes the
amplifier significantly robust. This section is devoted to describing this theory, with
a concrete optical example.
6.5.1 The Phase-Preserving Amplifier
The general idea of a PPA is simple. That is, it is a linear transformation of a bosonic
mode b1 to the output
b̃1 = g1 b1 + g2 b2∗ , (6.48)
where b2 is an auxiliary mode. Here, the point is that b2 is necessary in order to

satisfy the CCR of the output mode, [b̃1 (t), b̃1 (t )∗ ] = δ(t − t ). As a consequence,
the coefficients have to satisfy |g1 |2 − |g2 |2 = 1, and hence, the output b̃1 is an
amplified mode of b1 with gain |g1 | > 1. Usually, b2 is set to be a vacuum state;
hence, we have b̃1 = g1 b1 , implying that the phase of b1 is preserved. Note
that contrary to this, a squeezing operation does not preserve the phase of the input;
hence, it is sometimes referred to as phase-sensitive amplification.
A typical physical realization of a PPA is provided by a NOPA as discussed in
Sect. 1.5.2. Recall that this is an optical cavity system with two internal orthogonally
polarized fields a1 and a2 . The cavity contains a χ(2) optical crystal that facilitates
interaction between a1 and a2 via the quadratic Hamiltonian H = ı (a1∗ a2∗ − a1 a2 ),
6.5 Robust Quantum Amplification via Coherent Feedback 241
where ∈ R is assumed for simplicity (note that in Sect. 1.5.2, we discussed the
NOPA with coefficient /2). Also, the cavity couples to two external traveling fields
ξ1 and ξ2 ; the dynamics is given by Sect. 1.29:
da1 γ √
= − − i1 a1 + a2∗ − γξ1 ,
dt 2
da2∗ γ √
= − + i2 a2∗ + a1 − γξ2∗ , (6.49)
dt 2
where 1 and 2 are detuning parameters. Also, for simplicity, optical losses are
ignored. The output equations (boundary conditions) are given by (1.30):
√ √
η1 = γa1 + ξ1 , η2∗ = γa2∗ + ξ2∗ . (6.50)
In the standard scenario of quantum amplification, we only consider the first output
mode η1 and ignore the second one η2 , although in the scenario of entanglement
generation via NOPA both the output modes are of course kept, as discussed in
Sects. 1.5.2 and 6.2. Hence, let us focus on η1 , the Laplace transform of which is
given by
η1 [s] = g1 [s]ξ1 [s] + g2 [s]ξ2∗ [s],

(s − γ/2 + ı1 )(s + γ/2 − ı2 ) − 2
−γ
g1 [s] = , g2 [s] = ,
D[s] D[s]
γ γ
D[s] = s + + ı1 s + − ı2 − 2 .
2 2
The stability condition is obtained by examining the characteristic equation D[s] = 0;
particularly, when 1 = 2 = 0, it is γ 2 /4 − 2 > 0. The amplification process is
described in the Fourier domain s = ıω with ω, the frequency; that is, we consider
the linear transformation at steady state, η1 [ıω] = g1 [ıω]ξ1 [ıω] + g2 [ıω]ξ2∗ [ıω].
Note that g1 and g2 satisfy |g1 [ıω]|2 − |g2 [ıω]|2 = 1 for all ω. In particular, when
1 = 2 = 0, the amplification gain at the resonance frequency of a1 and a2 , ω = 0
(in the rotating frame, ω = 0 corresponds to ω p /2 with ω p , the pump frequency) is
given by
γ2 + 4 2
|g1 (0)| = 2 .
|γ − 4 2 |
Hence, it takes a large value by setting ≈ γ/2 while keeping | | < γ/2.
The above example can be generalized to any PPA; it is modeled as a linear system
with two inputs and two outputs as follows:

η1 [s] ξ1 [s] G 11 [s] G 12 [s]
= G[s] , G[s] = , (6.51)
η2∗ [s] ξ2∗ [s] G 21 [s] G 22 [s]
where ξ1 [s] is the Laplace transform of ξ1 , etc. The transfer function matrix G[s] at
s = ıω (called the scattering matrix in this field) satisfies
|G 11 [ıω]|2 − |G 12 [ıω]|2 = |G 22 [ıω]|2 − |G 21 [ıω]|2 = 1,

G 21 [ıω]G ∗11 [ıω] − G 22 [ıω]G ∗12 [ıω] = 0, ∀ω. (6.52)
Thus, |G 11 [ıω]| represents the amplification gain.
6.5.2 Coherent Feedback Control for a Quantum Amplifier
As will be shown later, and as in the classical case, an autonomous amplifier is

fragile against parameter fluctuation. Hence, we need feedback control. Here, we
take a coherent feedback strategy, because a measurement-based control inevitably
introduces extra classical noise. In particular, the controller is given by a completely
passive linear quantum system with two inputs ξ3 , ξ4 and two outputs η3 , η4 ; note that
a single-input and single-output passive system has gain equal to 1, and therefore,
it does not work as an attenuator. The controller is given in the Laplace domain as
follows: ∗ ∗
η3 [s] ξ3 [s] K 11 [s] K 12 [s]
= K [s] ∗ , K [s] = , (6.53)
η4∗ [s] ξ4 [s] K 21 [s] K 22 [s]
where the creation operator representation is taken to make the notation simple.
Note that the transfer function matrix K [s] is unitary in the Fourier domain; i.e.,
K [ıω]∗ K [ıω] = I holds for all ω [71]. We then consider the following feedback
connection, as shown in Fig. 6.19:
η2 = ξ3 , ξ2 = η4 , (6.54)
which is of course equivalent to η2∗ = ξ3∗ and ξ2∗ = η4∗ . By combining (6.51), (6.53),
and (6.54), we find that the controlled system, with inputs ξ1 , ξ4∗ and outputs η1 , η3∗ ,
has the following input–output relation in the Laplace domain:
Fig. 6.19 Coherent Plant

feedback configuration ξ1 η1
composed of the autonomous
amplifier (plant) G and the
G
controller K . Reprinted with ξ2 η2
permission from [65]. ©
(2016) by the American η3 ξ3
Physical Society K
η4 ξ4
Controller

η1 [s] G (fb) (fb)
11 [s] G 12 [s] ξ1 [s]
= ,
η ∗ [s] G (fb) (fb)
21 [s] G 22 [s]
ξ4∗ [s]
where
G (fb)
11 = [G 11 − K 21 (G 11 G 22 − G 12 G 21 )]/(1 − K 21 G 22 ),
G (fb)
12 = (G 12 K 22 )/(1 − K 21 G 22 ),
G (fb)
21 = (G 21 K 11 )/(1 − K 21 G 22 ),
G (fb)
22 = [K 12 + G 22 (K 11 K 22 − K 12 K 21 )]/(1 − K 21 G 22 ).
The matrix entries satisfy a condition corresponding to (6.52), i.e., |G (fb)

11 [ıω]| −
2
(fb)
|G 12 [ıω]| = 1 ∀ω.
2
Let us now discuss the effect of feedback control on the amplified signal in the
Fourier domain:
η1 [ıω] = G (fb) (fb) ∗

11 [ıω]ξ1 [ıω] + G 12 [ıω]ξ4 [ıω].
Our focus is on the amplification gain |G (fb) 11 [ıω]| especially when the original gain
|G 11 [ıω]| is large. The point is the use of (6.52); first, from |G 21 ||G 11 | = |G 22 ||G 12 |
together with the other two equations, we have |G 11 | = |G 22 | and |G 12 | = |G 21 | (for
the moment, we omit the variable ıω). Moreover, G 11 G 22 − G 12 G 21 = G 22 /G ∗11
holds. Then, in the limit |G 11 | → ∞, it follows that
|G 11 G 22 − G 12 G 21 | |G 22 | 1
= = → 0.
|G 11 | |G 11 |2 |G 11 |
This implies that (G 11 G 22 − G 12 G 21 )/|G 11 | converges to zero in this limit. As a

consequence, we have
G /|G | − K (G G − G G )/|G |
11 11
|G (fb)
11 21 11 22 12 21
| =
11
1/|G 11 | − K 21 G 22 /|G 11 |
G /|G |
11 11 1
→ = .
−K 21 G 22 /|G 11 | |K 21 |
Hence, in the frequency range where |G 11 [ıω]| 1 holds, the amplification gain
of the controlled system is given by |G (fb)
11 [ıω]| ≈ 1/|K 21 [ıω]| > 1. Therefore,
because the completely passive controller is much more robust compared to the
original amplifier, even if G 11 changes while still maintaining a large value, the
controlled system functions as a very robust amplifier with gain 1/|K 21 |. Hence, a
pertinent quantum counterpart of the classical feedback amplification theory is now
constructed.
30
= 5γ (a) β = 0.0 (No Feedback) (b)
= 3γ β = 0.05
β = 0.1
Gain [dB]
20
=γ
10
0
10−2 10−1 100 101 10−2 10−1 100 101
Frequency [ γ rad/s]
Fig. 6.20 a Gain profile of the specially detuned NOPA without feedback. b Gain profile of the
feedback-controlled system with parameters = 5γ and various β. Reprinted with permission from
[65]. © (2016) by the American Physical Society
Plant
ξ4
ξ1 η2 ξ3 η3
η1 ξ2 η4 Controller
Fig. 6.21 Coherent feedback of NOPA (plant) implemented with a beam splitter (controller).
Reprinted with permission from [65]. © (2016) by the American Physical Society
6.5.3 Example: Non-degenerate Optical Parametric Amplifier
Let us consider again the NOPA given by (6.49) as an example. In particular, we here
consider the non-standard parameter choice 1 = 2 = , which circumvents the
so-called gain–bandwidth constraint as shown below. The transfer function matrix
in this case is then given by
2
1 s − γ 2 /4 + ıγ −γ
G[s] = .
(s + γ/2)2 −γ s 2 − γ 2 /4 − ıγ
"
The maximum gain is |G 11 [0]| = 1 + 16 2 /γ 2 , which becomes larger by increas-
ing the pump strength . Figure 6.20a depicts the three cases corresponding to
= γ, 3γ, 5γ. Remarkably, the bandwidth (i.e., the frequency range where nearly flat
amplification gain is realized) does not depend on the gain, which is nearly γ/10; that
is, the amplification is free from the gain–bandwidth constraint. Note that because
the pole of the transfer function matrix is fixed to −γ/2, the system is always stable
irrespective of ; this is a clear contrast to the standard NOPA that requires | | < γ/2
for stability. As a consequence, in a proper parameter regime such that the linearized
model (6.49) is valid, there is no upper limit on and accordingly on the gain |G 11 [0]|
as well.
Now, let us consider coherent feedback control of this amplifier. Here, as illustrated
in Fig. 6.21, we take a beam splitter with transmissivity α and reflectivity β as a
controller, in which case the transfer function matrix is constant:

α β
K [s] = , α, β ∈ R.
β −α
Thus, K 21 = β represents the attenuation level. The amplification gain of the con-
trolled system is then
(1 − β)s 2 + βγs − (1 + β)γ 2 /4 + ıγ

G (fb)
11 [s] = .
(1 − β)s 2 + γs + (1 + β)γ 2 /4 + ıβγ
As expected from the general theory discussed above, we find that the limit
|G 11 [ıω]| → ∞ (i.e., → ∞) leads to |G (fb) 11 [ıω]| → 1/β in a certain fre-
quency bandwidth. Also, by examining the denominator √ of G (fb)
11 [s], we find that
the controlled system is stable when | | < (γ/2β) (1 + β)(1 − β). This yields
|β| < γ/2| | when β 2 1; hence, let us choose = 5γ, meaning |β| < 0.1.
Figure 6.20b shows the gain |G (fb)11 [ıω]| for the cases β = 0.1, β = 0.05, and β = 0.
From this, we observe that the gain of the controlled NOPA is reduced from the value
of the autonomous one; as shown later, by sacrificing this gain, the controlled NOPA
obtains a fine robustness property against parameter fluctuation. Note also that the
controlled NOPA has the gain–bandwidth constraint. This is in fact a general feature
of a controlled amplifier: The gain and bandwidth of the controlled amplifier can be
easily modified by simply tuning the controller.
Finally, let us discuss the robustness property of the controlled amplifier, which is
in fact the main advantage of using feedback. To see this, let us consider the following
imperfect case. First, the system parameters are fragile; the pump strength can
change as = (1 + 0.1e0 ) 0 where 0 is the nominal value; similarly, the detunings
1 and 2 can slightly deviate from , which is modeled by 1 = (1 + 0.001e1 )
and 2 = (1 + 0.001e2 ) . Here, (e0 , e1 , e2 ) are independent random variables drawn
from the uniform distribution on [−1, 1]. The blue lines in Fig. 6.22 are 50 sample
values of the autonomous gain |G 11 [ıω]| in the case 0 = 5γ. This shows that the
parameter fluctuations above lead to significant fragility of the amplification gain.
Hence, this is the situation where we should apply the feedback. The red lines in
30
25
Gain [dB]
20
15
10
5
0
1
10−2 10−1 10 0 10
Frequency [ κ rad/s]
Fig. 6.22 The upper blue lines represent the gain profile of the autonomous NOPA, |G 11 [ıω]|,
(fb)
while the lower red lines correspond to the controlled case, |G 11 [ıω]| with β = 0.1. In both cases,
0 = 5γ. Reprinted with permission from [65]. © (2016) by the American Physical Society (color
figure online)
Fig. 6.22 are 50 sample values of the controlled gain |G (fb)

11 [ıω]| with β = 0.1. This
clearly demonstrates that gain fluctuation in the amplifier is greatly suppressed by
feedback.
6.5.4 Added Noise
In this section, we have discussed only the amplification of the mean, but not of the
higher-order moments, especially the variance. In the general form (6.48), certainly
the mean b̃1 = g1 b1 is amplified, but note that the noise variance defined by
1
|b|2 := bb∗ + b∗ b, b = b − b
2
must also be amplified as follows. If b2 is the vacuum mode, (6.48) leads to |b̃1 |2 =
|g1 |2 |b1 |2 + |g2 |2 /2. This directly implies the degradation of the signal-to-noise
ratio as follows:
= |b̃1 | = |b1 |2 |b1 |2

2
(S/N) < = (S/N).
|b̃1 |2 |b1 |2 + A |b1 |2
Hence, the added noise

|g2 |2 |g1 |2 − 1
A := =
2|g1 |2 2|g1 |2
quantifies how well the signal information is amplified [72, 73]. In particular, in the
large amplification limit |g1 | → ∞, we find A → 1/2, and this is called the quantum
noise limit.
In [65], it was proven that in general, the controlled amplifier functions as a PPA
achieving the quantum noise limit A → 1/2 in the large amplification limit, even if
several imperfections are present in the feedback loop. This effect was demonstrated
in [65] in the numerical simulation for the same NOPA example studied in the
previous subsection. This means that precise fabrication of the feedback control is
not necessary; hence, the presented theory is of practical interest.
6.6 Feedback Control Experiments
We have seen in the previous sections and chapters that there are two types of quantum
feedback schemes: coherent feedback and measurement-based quantum feedback.
Here, we describe two actual experiments that demonstrate the effectiveness of those
feedback schemes in different physical setups. The first one is a fully optical system
that contains a coherent feedback architecture, and the second is a large atomic
6.6 Feedback Control Experiments 247
ensemble controlled by measurement-based feedback. In both cases, the control

goal is to create an enhanced (optical or atomic) squeezed state.
6.6.1 Coherent Feedback for Optical Squeezing

Enhancement
As discussed before, a squeezed state is a genuine non-classical state such that the
fluctuation is reduced below the vacuum noise variance in one of the quadrature
observables q and p. Coherent feedback has the ability to further suppress this fluc-
tuation compared to the case without feedback, as theoretically proven in [18, 19] in
the quantum optical setting. Figure 6.23a illustrates the proposed coherent feedback
configuration composed of an optical parametric oscillator (OPO; see Sect. 1.5.3) as
the plant and a beam splitter (BS) as the controller; the output squeezed state of OPO
is then enhanced by the feedback and produced at the output of BS. This theoretical
proposal was experimentally demonstrated in [21], and this is what we discuss in
detail in this subsection.
We begin by describing the theoretical model of this feedback method, which is
depicted in Fig. 6.23b. First, let us recall from Sect. 1.5.3 that the dynamics of OPO
is given, in a rotating frame, by
da γ+κ √ √ √
=− a + a ∗ − γξ1 − κξloss,1 , η1 = γa + ξ1 , (6.55)
dt 2 2
where γ = cT1 / and κ = cL 1 / with the optical path length in the OPO, T1 the
transmissivity of the coupling BS, L 1 the transmissivity of a fictitious BS (which
models the optical loss), and c the speed of light. In the quadrature form, the Fourier
input–output relation is given by
(a) (b) System

Output Input η1 ξ1
OPO OPO
η3 Controller
Beam CBS
Splitter ξ3 η2
Coherent Enhanced
State Squeezed ξ loss ξ2
State
Fig. 6.23 a Abstract schematic of coherent feedback control for optical field squeezing. b Config-
uration of the coherent feedback loop. Reprinted with permission from [21] © 2012 IEEE
q q q
η1 [ıω] = G[ıω] + g[ıω] ξ1 [ıω] + Ḡ[ıω] + ḡ[ıω] ξloss,1 [ıω],
p p p
η1 [ıω] = G[ıω] − g[ıω] ξ1 [ıω] + Ḡ[ıω] − ḡ[ıω] ξloss,1 [ıω],
where
√
γ 2 − (κ − 2ıω)2 + 2 2 γκ(γ + κ − 2ıω)
G[ıω] = , Ḡ[ıω] = ,
(γ + κ − 2ıω)2 − 2 (γ + κ − 2ıω)2 − 2
2 γ
g[ıω] = ,
(γ + κ − 2ıω)2 − 2
√ q p
and ḡ[ıω] = κ/γg[ıω]. The variance of η1 [ıω] and η1 [ıω] is simply given by
q q p p
the power spectrum S1 (ıω) = |η1 [ıω]|2 and S1 (ıω) = |η1 [ıω]|2 . In particular,
q
when the input field is a vacuum state, we have S1 (ıω) = |G[ıω] + g[ıω]|2 +
p
|Ḡ[ıω] + ḡ[ıω]|2 and S1 (ıω) = |G[ıω] − g[ıω]|2 + |Ḡ[ıω] − ḡ[ıω]|2 . If = 0, then
q p
S1 (ıω) = S1 (ıω) = 1, ∀ω, which is the vacuum noise variance.
Next, we describe the feedback configuration. The controller is called the control
BS (CBS); its transmissivity T2 is tunable. The coherent input field ξ2 (t) is sent
to the CBS. Next, one of the outputs of the CBS, η3 (t), is sent to the OPO. The
output of the OPO, η1 (t), is then sent back to the CBS. This closed-loop system
outputs the enhanced squeezed field η2 (t). The input and output fields of the CBS
are connected by
" " " "
η2 = 1 − T2 ξ2 + T2 1 − L 2 ξ3 + L 2 ξloss,2 ,
" " " "
η3 = − 1 − T2 1 − L 2 ξ3 + L 2 ξloss,2 + T2 ξ2 ,
where ξloss,2 is a vacuum field entering through a fictitious BS with reflectivity L 2 ;

this models the losses in the CF loop. Now, let τa = la /c (τb = lb /c) be the time
delay resulting from the optical path length la (lb ) from (to) the CBS to (from) the
OPO. Then, the feedback interconnection is represented by ξ1 (t) = η3 (t − τa )eiω0 τa
and ξ3 (t) = η1 (t − τb )eiω0 τb , where ω0 is the resonance frequency of the OPO.
Combining these equations with (6.55), we obtain the input–output relation of the
whole closed-loop system in the Fourier domain as follows:
" √
T2 1 − L 2 α± [ıω]
η2± [ıω] = 1 − T2 + √ ξ ± [ıω]
1 + α± [ıω] (1 − T2 )(1 − L 2 ) 2
√
T2 (1 − L 2 )β ± [ıω]
+ √ ξ ± [ıω]
1 + α± [ıω] (1 − T2 )(1 − L 2 ) loss,1
" √
T2 (1 − L 2 )(1 − T2 )L 2 α± [ıω] ±
+ T2 L 2 − √ ξ [ıω],
1 + α± [ıω] (1 − T2 )(1 − L 2 ) loss,2
α± [ıω] = (G[ıω] ± g[ıω])eı(ω+ω0 )(τa +τb ) ,
β ± [ıω] = (Ḡ[ıω] ± ḡ[ıω])eı(ω+ω0 )τb ,
Fig. 6.24 The transmissivity 0

T2 of the CBS versus the
-1
Normalized noise power (dB)

squeezing levels for various
normalized pumping strength -2
x = /γ. The blue, red, and -3
green lines correspond to
x = 0.1, 0.35, and 0.6, -4
respectively. The circles -5
indicate the values at T2 = 1
-6
and L 2 = 0, corresponding
to the uncontrolled OPO. -7
Reprinted with permission -8
from [21] © 2012 IEEE
(color figure online) -9
0.5 0.6 0.7 0.8 0.9 1.0

Transmissivity T 2 of the CBS
where + and − correspond to q and p, respectively. Here, the input is assumed to be

a vacuum state (this is indeed the case in the experiment); then, the power spectrum
of the output field can be calculated as S2± (ıω) = |η2± [ıω]|2 . Also, we assume that
the coherent feedback loop is on resonance, i.e., eıω0 (τa +τb ) = −1.
Now, we conduct a numerical simulation with the following parameters, to exam-
ine the performance of the feedback control for squeezing enhancement: T1 = 0.12,
L 1 = 5.0 × 10−3 , L 2 = 5.0 × 10−2 , l = 0.5 m, and la = lb = 0.25 m.
q q
Figure 6.24 depicts the normalized noise power 10 log10 (|η2 [ıω]|2 /|ξ2 [ıω]|2 ) =
q
10 log10 S2 (ıω) dB versus T2 , at frequency ω/2π = 1 MHz, for several normal-
ized pumping strength x = /γ. Note that the horizontal axis at 0 dB indicates the
vacuum noise variance. The circles represent the values at T2 = 1 with L 2 = 0,
i.e., the squeezing levels of the uncontrolled OPO. This figure shows that if the
pumping strength is weak (x = 0.1 or x = 0.35), there exists an optimal point of
T2 that lowers the variance than that in the uncontrolled case; that is, the coherent
feedback certainly enhances the squeezing. However, if a large pump field is applied
(x = 0.6), the autonomous OPO without control is the best system that minimizes the
variance; the CFB does not enhance the squeezing at all. This is simply because, in
general, a highly squeezed optical field is fragile to optical loss, leading to a trade-off
between the squeezing level and the optical loss. Thus, for a strongly pumped OPO,
the feedback introduces further optical loss rather than enhancing the squeezing.
This is a clear limitation of coherent feedback control for the purpose of squeezing
enhancement.
Finally, let us see the experimental demonstration of the above. The setup is
illustrated in Fig. 6.25 (see [21] for the detailed configuration). A notable point is that
the CBS is constructed so that it is involved in a Mach–Zehnder (MZ) interferometer,
which allows us to tune the controller parameter T2 . In this experiment, the parameters
are chosen as x = 0.106, T1 = 0.20, L 1 = 9.0 × 10−3 , L 2 = 0.12, l = 0.5 m, and
la = lb = 0.25 m. Figure 6.26 shows the squeezing and anti-squeezing levels in the
CW
Fig. 6.25 Experimental configuration. MCC: mode cleaning cavity, PD: photodetector, PZT: piezo-
electric transducer, PBS: polarized beam splitter, HWP: half-wave plate, and LO: local oscillator.
The blue dashed line indicates the CFB loop. The green dashed line indicates a Mach–Zehnder
interferometer, corresponding to the CBS. Reprinted with permission from [21] © 2012 IEEE (color
figure online)
5
Normalized noise power (dB)
-1
-2
0 2 4 6 8
Frequency (MHz)
Fig. 6.26 Frequency dependence of the squeezing and anti-squeezing levels. The blue lines rep-
resent those without feedback (T2 = 1), while the green, red, and pink lines correspond to the
case with feedback under the condition T2 = 0.7, 0.8, and 0.9, respectively. Dark noise (variance
of the number of electrons in the detector) is subtracted. Dashed lines indicate theoretical values.
Reprinted with permission from [21] © 2012 IEEE (color figure online)
frequency domain, with several T2 ; the blue solid lines indicate those without the
coherent feedback (i.e., T2 = 1), while the green, red, and pink solid lines correspond
to the feedback case with T2 = 0.7, 0.8, and 0.9, respectively. The large noise found
at lower frequencies comes from the laser noises and modulation signals that are
used for cavity locking mechanisms. Nonetheless, at higher frequency regime, the
experimental result agrees well with the theoretical values shown with dashed lines.
Note that while better squeezing enhancement is realized if we take a smaller value of
T2 , the effective squeezing bandwidth becomes narrower. This additional constraint
imposed on the feedback control is mainly because the OPO and the feedback loop
suffer from the time delays.
6.6.2 Measurement-Based Feedback for Spin Squeezing

in Atomic Ensemble
Next, we consider a particular experimental demonstration of measurement-based

feedback control as reported in [74]. As before, let us begin with the theoretical
background.
The system is a large atomic ensemble√system as described in Sect. 1.5.5, which is
subjected to the Faraday interaction L = M Jz and the spin rotation implemented by
the magnetic field H = b Jy . Let us recall that this setup realizes QND measurement
of Jz , through the dynamics (1.44) and the measurement output (1.45). In particular,
we here assume the large ensemble limit J = N /2 1 and that the x-component
is much larger than the others, leading to Jx ≈ J and the CCR relation [Jy , Jz ] =
ı Jx ≈ ı J ; further, assume that the initial state is the spin-coherent state, which is
exactly the same as the optical coherent state in the two-dimensional tangent space
at (x, y, z) = (J, 0, 0) in the Bloch-like sphere depicted in Fig. 6.27a. Under these
assumptions, the state is always Gaussian as long as it is confined in the tangent
space. Now, we construct a quantum filter as described in Sect. 4.2.1, to estimate
jt (Jz ) = U (t)∗ Jz√
U (t) using the measurement output (1.45), which in Itō form is
given by dym = 2 M jt (Jz )dt + dA + dA∗ . Then, as seen in (4.21) and (4.22), the
conditional mean and variance of jt (Jz ) change in time as
√
dJz c = J b(t)dt + 2 MJz2 c dw, dJz2 c = −4MJz2 2c dt, (6.56)
where Jz c = πt (Jz ) is the least mean square estimate of jt (Jz ). Also, note that
Jz3 c = 0, because of the Gaussianity of the state. The above equations show
that the variance of jt (Jz ) deterministically decreases, while the mean fluctuates
probabilistically; then, due to the Heisenberg uncertainty relation Jy2 c Jz2 c ≥
J 2 /4, the variance of Jy must increase. Hence, a squeezed state is generated, with
its mean probabilistically fluctuating, which is illustrated in Fig. 6.27b. Note that in
the general setting without the Gaussian approximation, the state finally converges
to one of the eigenstate of Jz if b(t) = 0, as discussed in Sect. 4.2.4. This state is
called a spin-squeezed state, in analogy to the optical squeezed state.
Now, note that without any control (i.e., b(t) = 0), ensemble averaging of the spin-
squeezed states over all the atoms leads to Jz2 ≥ J/2, because of the fluctuation
of the mean. Thus, the unconditional state is not anymore a spin-squeezed state, as
(a) z (b) z
y y
x x
Fig. 6.27 a Spin-coherent state and b spin-squeezed state
1 2 pL 3
FB
pA
qL
qA
pL
qL
FB
4 5 pL 6
FB
pA
qL
qA pL
qL
FB
Fig. 6.28 Measurement-based feedback protocol for spin squeezing
discussed at the end of Sect. 4.3.1. Measurement-based feedback control is applicable

to resolve this issue and eventually generate a spin-squeezed state in a deterministic
way, as follows. The point is that now we have a real-time optimal estimate of jt (Jz ),
namely Jz c = πt (Jz ), by updating the quantum filter (6.56). Hence, the control
input signal b(t) can be set to a function of Jz c ; in particular, let us here take
a simple choice b(t) = −b̄Jz c with b̄ > 0. Then, (6.56) yields dE[Jz c ]/dt =
−J b̄E[Jz c ], leading to E[Jz c ] → 0 in the long time limit. Then, because decrease
of Jz2 c is still maintained under the control due to (6.56), the spin-squeezed state
with mean zero is deterministically generated.
The above measurement-based feedback strategy was experimentally demon-
strated in [74]. Let us now see the idea behind this experiment. Figure 6.28 illus-
trates the control procedure for deterministically
√ √generating a spin-squeezed state.
In the figure, we denote (q A , p A ) = (Jy / J , Jz / J ), and (q L , p L ) are the pair of
normalized (q, p) components of the polarization of the output light field η. The
procedure is to send a polarized beam into the atomic ensemble followed by apply-
ing a magnetic field depending on the measurement result, and in the experiment,
this action was performed twice; that is, rather than the above-described continuous-
time monitoring and feedback, what was demonstrated in the experiment is actually
discrete-time feedback control. The detailed control procedure is as follows (the steps
below correspond to the numbering in Fig. 6.28).
1. Initialization: The atomic ensemble and the light field are prepared to a spin-
coherent state (red circle) and a vertical polarized state (green arrow), respectively.
2. After the atom–field Faraday interaction, the polarization rotates depending on
the probabilistic shift of the atom state, which is now squeezed. The angle of
polarization can be determined by measuring q L by a polarization beam splitter;
this can be used to estimate the shift of the spin-squeezed state, and this shift is
compensated by the feedback control implemented by a magnetic field.
3. As a result, a spin-squeezed state is deterministically generated.
4. Second initialization: The atom state is prepared in the spin-squeezed state gen-
erated through the Steps 1–2–3.
5. The same action as in the Step 2 is applied. Due to the conditional state evolution,
the atom state is further squeezed. Also, the rotation angle of the polarization is
reduced compared to the case in the Step 2.
6. As a result, a further squeezed state is generated.
As implied from the above procedure, the variance of the polarization angle of the
output field depends on the variance of p A . Therefore, the squeezing level of the
atomic state can be evaluated by using the variance of the output q L = κ p A + q L ,
where q L represents the normalized q-component of the√ polarization of the input light
field ξ and κ is the interaction√ strength proportional to M. (Note that this equation
corresponds to dym = 2 M jt (Jz )dt + dA + dA∗ .) In particular, the squeezing
level can be defined by ξsq 2
= (p (2) (1) 2 (i)
A ) /(p A ) , where p A denotes the initial
2
p-component of the atomic state in the ith experiment. Then, from (q L (i) )2 =
κ2 (p A (i) )2 + (q L(i) )2 = κ2 (p A (i) )2 + 1/2. Therefore, the squeezing level
of the atom is represented as
(p A (2) )2 (q L (2) )2 − 1/2

ξsq
2
= = .
(p (1)
A )
2 (q L (1) )2 − 1/2
In the experiment, it was observed that the best squeezed level (with respect to the
feedback gain) is ξsq
2
≈ 0.912, i.e., about 0.8 dB spin squeezing was realized.
There are several experimental demonstrations of quantum feedback control beyond

linear quantum systems; for instance, measurement-based feedback has been exper-
imentally implemented to stabilize photon number states in a microwave cavity [75]
and Rabi oscillations in a superconducting qubit [76]. As for coherent feedback, we
find its application to stabilizing an entangled state of two superconducting qubits

[77]. Here, in accordance with the theme of this monograph, we briefly highlight the
experiments reported in [78, 79], which both experimentally demonstrate feedback
control for linear quantum systems.
The first one [78] experimentally implemented a coherent feedback network for
an electromechanical device, aimed at realizing efficient control and readout of the
mechanical state. While these two operations cannot be simultaneously and effi-
ciently achieved in the standard setup without a feedback architecture, the developed
coherent feedback allows very fast switching between the two. Such very fast manip-
ulation using a feedback controller, which is comparable to the system’s decay rate,
would be quite difficult to implement with a measurement-based feedback controller.
Thus, the experiment illustrates one of the major advantages of coherent feedback
control in an actual physical setup.
The next experiment, [79], demonstrates the use of measurement-based feedback
for cooling a mechanical oscillator to achieve its motional ground state and showed
a significant improvement over the conventional approach. There are two important
system components to accomplish this goal. One is an ultra-precise position sensor
that can resolve the zero-point fluctuation of a nano-mechanical oscillator, and the
other is an actuator powered by radiation pressure force. This illustrates the require-
ments of having a very precise sensor and actuator in order to attain a real advantage
of measurement-based feedback control in the quantum regime. As described in this
work, this is now within reach of current state-of-the-art experiments in the field of
opto-mechanics.
References
1. S.L. Braunstein, P. van Loock, Quantum information with continuous variables. Rev. Mod.
Phys. 77, 513 (2005)
2. A. Furusawa, P. van Loock, Quantum Teleportation and Entanglement: A Hybrid Approach to
Optical Quantum Information Processing (Wiley-VCH, Berlin, 2011)
3. N.C. Menicucci, P. van Loock, M. Gu, C. Weedbrook, T.C. Ralph, M.A. Nielsen, Universal
quantum computation with continuous-variable cluster states. Phys. Rev. Lett. 97, 110501
(2006)
4. N.C. Menicucci, S.T. Flammia, P. van Loock, Graphical calculus for Gaussian pure states.
Phys. Rev. A 83, 042335 (2011)
5. J.F. Poyatos, J.I. Cirac, P. Zoller, Quantum reservoir engineering with laser cooled trapped ions.
Phys. Rev. Lett. 77, 4728 (1996)
6. N. Yamamoto, Parametrization of the feedback Hamiltonian realizing a pure steady state. Phys.
Rev. A 72, 024104 (2005)
7. B. Kraus, H.P. Buchler, S. Diehl, A. Kantian, A. Micheli, P. Zoller, Preparation of entangled
state by quantum Markov processes. Phys. Rev. A. 78, 042307 (2008)
8. F. Verstraete, M.M. Wolf, J.I. Cirac, Quantum computation and quantum-state engineering
driven by dissipation. Nat. Phys. 5, 633–636 (2009)
9. K. Vollbrecht, C.A. Muschik, J.I. Cirac, Entanglement distillation by dissipation and continuous
quantum repeaters. Phys. Rev. Lett. 107, 120502 (2011)
10. K. Koga, N. Yamamoto, Dissipation-induced pure Gaussian state. Phys. Rev. A 85, 022103
(2012), Reprinted, with permission, © (2012) by the American Physical Society
References 255
11. N. Yamamoto, Pure Gaussian state generation via dissipation: a quantum stochastic differential
equation approach. Philos Trans. R. Soc. A 370, 5324–5337 (2012)
12. R. Simon, E.C.G. Sudarshan, N. Mukunda, Gaussian pure states in quantum mechanics and
the symplectic group. Phys. Rev. A 37, 3028–3038 (1988)
13. Y. Ikeda, N. Yamamoto, Deterministic generation of Gaussian pure states in a quasilocal dis-
sipative system. Phys. Rev. A 87, 033802 (2013)
14. O. Techakesari, H.I. Nurdin, On the quasi-balanceable class of linear quantum stochastic sys-
tems. Syst. Control Lett. 78, 25–31 (2015)
systems. IEEE Trans. Autom. Control 59(9), 2413–2425 (2014)
16. T. Tufarelli, A. Ferraro, A. Serafini, S. Bose, M.S. Kim, Coherently opening a high-Q cavity.
Phys. Rev. Lett. 112, 133605 (2014)
17. J. Ma, X. Wang, C.P. Sun, F. Nori, Quantum spin squeezing. Phys. Rep. 509, 89–165 (2011)
18. M. Yanagisawa, H. Kimura, Transfer function approach to quantum control - part ii: control
concepts and applications. IEEE Trans. Autom. Control 48(12), 2121–2132 (2003)
19. J.E. Gough, S. Wildfeuer, Enhancement of field squeezing using coherent feedback. Phys. Rev.
A 80, 042107 (2009)
20. O. Crisafulli, N. Tezak, D.B.S. Soh, M.A. Armen, H. Mabuchi, Squeezed light in an optical
parametric amplifier oscillator network with coherent feedback quantum control. Opt. Express
21, 18371–18386 (2013)
21. S. Iida, M. Yukawa, H. Yonezawa, N. Yamamoto, A. Furusawa, Experimental demonstration
of coherent feedback control on optical field squeezing. IEEE Trans. Autom. Control 57(8),
2045–2050. Reprinted, with permission, © 2012 IEEE (2012)
22. Z. Shi, H.I. Nurdin, Coherent feedback enabled distributed generation of entanglement between
propagating Gaussian fields. Quant. Inf. Process. 14, 337–359 (2015). © 2014 Springer.
Reprinted, with permission of Springer
23. C.C. Gerry, P.L. Knight, Introductory Quantum Optics (Cambridge University Press, Cam-
bridge, 2005)
24. W.P. He, F.L. Li, Generation of broadband entangled light through cascading nondegenerate
optical parametric amplifiers. Phys. Rev. A 76, 012328 (2007)
25. Z. Yan, X. Jia, X. Su, Z. Duan, C. Xie, K. Peng, Cascaded entanglement enhancement. Phys.
Rev. A 85, 040305(R) (2012)
26. B.C. Jacobs, T.B. Pittman, J.D. Franson, Quantum relays and noise suppression using linear
optics. Phys. Rev. A 66, 052307 (2002)
27. W. Michiels, S.I. Niculescu, Stability and Stabilization of Time-Delay Systems: An Eigenvalue-
Based Approach. Advances in Design and Control (Society of Industrial and Applied Mathe-
matics, 2007)
28. K. Engelborghs, T. Luzyanina, G. Samaey, DDE-BIFTOOL vol. 2.00: A Matlab package
for bifurcation analysis of delay differential equations. Department of Computer Science,
Katholieke Universiteit Leuven, Belgium, Technical Report TW-330 (2001)
29. K. Engelborghs, T. Luzyanina, D. Roose, Numerical bifurcation analysis of delay differential
equations using DDE-BIFTOOL. ACM Trans. Math. Softw. 28(1), 1–21 (2002)
30. M. Di Loreto, M. Dao, L. Jaulin, J.-F. Lafay, J.J. Loiseau, Applied interval computation: a
new approach for time-delays systems analysis, in Applications of Time Delay Systems, ed. J.
Chiasson, J.J. Loiseau (Springer, Berlin 2007)
31. Z. Shi, H.I. Nurdin, Effect of phase shifts on EPR entanglement generated on two propagating
Gaussian fields via coherent feedback, in Proceedings of the 53rd IEEE Conference on Decision
and Control (CDC) (Dec 15–17, 2014), pp. 5813–5818
32. Z. Shi, H.I. Nurdin, Optimization of distributed entanglement generated between two Gaussian
fields by the modified steepest descent method, in Proceedings of the 2015 American Control
Conference (ACC) (Jul 1–3, 2015), pp. 2697–2702
33. Z. Shi, H.I. Nurdin, Local optimality of a coherent feedback scheme for distributed entan-
glement generation: the idealized infinite bandwidth limit, in Proceedings of the 54th IEEE
Conference on Decision and Control (CDC) (Dec 15–18, 2015), pp. 7755–7760
34. Y. Zhou, X. Jia, F. Li, J. Yu, C. Xie, K. Peng, Quantum coherent feedback control for generation
system of optical entangled state. Sci. Rep. 5, 11132 (2015)
35. D. Wang, C. Xia, Q. Wang, Y. Wu, F. Liu, Y. Zhang, M. Xiao, Feedback-optimized extraordinary
optical transmission of continuous-variable entangled states. Phys. Rev. B 91, 121406(R) (2015)
36. Z. Shi, H.I. Nurdin, Entanglement in a linear coherent feedback chain of nondegenerate optical
parametric amplifiers. Quant. Inf. Comput. 15(13–14), 1141–1164 (2015)
37. Z. Shi, H.I. Nurdin, Formulae for entanglement in a linear coherent feedback network of mul-
tiple nondegenerate optical parametric amplifiers: the infinite bandwidth case, in Proceedings
of the 2016 American Control Conference (ACC) (Jul 6–8, 2016), pp. 4769–4774
38. V.B. Braginsky, F.Y. Khalili, Quantum Measurement (Cambridge University Press, Cambridge,
1992)
39. C.M. Caves, K.S. Thorne, R.W.P. Drever, V.D. Sandberg, M. Zimmermann, On the measure-
ment of a weak classical force coupled to a quantum mechanical oscillator. I. Issues of principle.
Rev. Mod. Phys. 52, 341–392 (1980)
40. M. Tsang, C.M. Caves, Coherent quantum-noise cancellation for optomechanical sensors. Phys.
Rev. Lett. 105, 123601 (2010)
41. H. Miao, Exploring Macroscopic Quantum Mechanics in Optomechanical Devices (Springer,
Berlin, 2012)
42. E.M. Komar, M. Kessler, L. Bishof, A.S. Jiang, J. Sorensen, Ye, M.D. Lukin, A quantum
network of clocks. Nat. Phys. 10, 582–587 (2014)
43. N. Yamamoto, Coherent versus measurement feedback: linear systems theory for quantum
information. Phys. Rev. X 4, 041029 (2014)
44. Y. Yokotera, N. Yamamoto, Geometric control theory for quantum back-action evasion. EPJ
Quantum Technol. 3(15), 1–22 (2016)
45. B.P. Abbott et al., Observation of gravitational waves from a binary black hole merger. Phys.
Rev. Lett. 116, 061102 (2016)
46. N. Yamamoto, M.R. James, Zero dynamics principle for perfect quantum memory in linear
networks. New J. Phys. 16, 073032 (2014)
47. D.A. Lidar, K.B. Whaley, Decoherence-free subspaces and subsystems, in Irreversible Quan-
tum Dynamics. Lecture Notes in Physics, vol. 622 (2003), p. 83
48. D.F. Phillips, A. Fleischhauer, A. Mair, R.L. Walsworth, M.D. Lukin, Storage of light in atomic
vapor. Phys. Rev. Lett. 86, 783 (2001)
49. I. Novikov, A.V. Gorshkov, D.F. Phillips, A.S. Sorensen, M.D. Lukin, R.L. Walsworth, Optimal
control of light pulse storage and retrieval. Phys. Rev. Lett. 98, 243602 (2007)
50. A.V. Gorshkov, A. Andre, M.D. Lukin, A.S. Sorensen, Photon storage in lambda-type optically
dense atomic media I. Cavity model. Phys. Rev. A 76, 033804 (2007)
51. Q. Xu, P. Dong, M. Lipson, Breaking the delay-bandwidth limit in a photonic structure. Nat.
Phys. 3, 406–410 (2007)
52. J. Yoshikawa, K. Makino, S. Kurata, P. van Loock, P,A. Furusawa, Creation, storage, and on-
demand release of optical quantum states with a negative Wigner function. Phys. Rev. X 3,
041028 (2013)
53. N. Yamamoto, Decoherence-free linear quantum subsystems. IEEE Trans. Autom. Control
59(7), 1845–1857 (2014)
54. M. Hush, A.R.R. Carvalho, M. Hedges, M.R. James, Analysis of the operation of gradient echo
memories using a quantum input-output model. New J. Phys. 15, 085020 (2013)
55. N. Yamamoto, H.I. Nurdin, M.R. James, Quantum state transfer for multi-input linear quantum
systems, in Proceedings of the 55th IEEE Conference on Decision and Control (CDC) (Dec
15–17, 2014) (2016)
56. C.A. Muschik, K. Hammerer, E.S. Polzik, J.I. Cirac, Efficient quantum memory and entangle-
ment between light and an atomic ensemble using magnetic field. Phys. Rev. A 73, 062329
(2006)
57. Q.Y. He, M.D. Reid, E. Giacobino, J. Cviklinski, P.D. Drummond, Dynamical oscillator-cavity
model for quantum memories. Phys. Rev. A 79, 022310 (2009)
References 257
58. Y. Wang, J. Minar, G. Hetet, V. Scarani, Quantum memory with a single two-level atom in a
half cavity. Phys. Rev. A 85, 013823 (2012)
59. S.A. Aljunid, G. Maslennikov, Y. Wang, H.L. Dao, V. Scarani, C. Kurtsiefer, Excitation of a
single atom with exponentially rising light pulses. Phys. Rev. Lett. 111, 103001 (2013)
60. M. Bader, S. Heugel, A.L. Chekhov, M. Sondermann, G. Leuchs, Efficient coupling to an
optical resonator by exploiting time-reversal symmetry. New J. Phys. 15, 123008 (2013)
61. G.K. Gulati, B. Srivathsan, B. Chng, A. Cere, D. Matsukevich, C. Kurtsiefer, Generation of
an exponentially rising single-photon field from parametric conversion in atoms. Phys. Rev. A
90, 033819 (2014)
62. H. Ogawa, H. Ohdan, K. Miyata, M. Taguchi, K. Makino, H. Yonezawa, J. Yoshikawa, A.
Furusawa, Real-time quadrature measurement of a single-photon wavepacket with continuous
temporal-mode-matching. Phys. Rev. Lett. 116, 233602 (2016)
63. H.S. Black, Inventing the negative feedback amplifier. IEEE Spectr. 14, 55 (1977)
64. H.S. Black, Stabilized feedback amplifiers. Proc. IEEE 72, 716–722 (1984)
65. N. Yamamoto, Quantum feedback amplification. Phys. Rev. Appl. 5, 044012 (2016), Reprinted,
with permission, © (2016) by the American Physical Society
66. A.A. Clerk, M.H. Devoret, S.M. Girvin, F. Marquardt, R.J. Schoelkopf, Introduction to quantum
noise, measurement, and amplification. Rev. Mod. Phys. 82, 1155 (2010)
67. N. Bergeal, F. Schackert, M. Metcalfe, R. Vijay, V. Manucharyan, L. Frunzio, D.E. Prober,
R.J. Schoelkopf, S.M. Girvin, M.H. Devoret, Phase-preserving amplification near the quantum
limit with a Josephson ring modulator. Nature 465, 64–68 (2010)
68. H.M. Chrzanowski, N. Walk, S.M. Assad, J. Janousek, S. Hosseini, T.C. Ralph, T. Symul, P.K.
Lam, Measurement-based noiseless linear amplification for quantum communication. Nat.
Photonics 8, 333–338 (2014)
69. F. Hudelist, J. Kong, C. Liu, J. Jing, Z.Y. Ou, W. Zhang, Quantum metrology with parametric
amplifier-based photon correlation interferometers. Nat. Commun. 5, 3049 (2014)
70. A. Metelmann, A.A. Clerk, Quantum-limited amplification via reservoir engineering. Phys.
Rev. Lett. 112, 133904 (2014)
71. J. Gough, R. Gohm, M. Yanagisawa, Linear quantum feedback networks. Phys. Rev. A 78,
061204 (2008)
72. H.A. Haus, J.A. Mullen, Quantum noise in linear amplifiers. Phys. Rev. 128, 2407–2413 (1962)
73. C.M. Caves, Quantum limits on noise in linear amplifiers. Phys. Rev. D 26, 1817–1839 (1982)
74. R. Inoue, S. Tanaka, R. Namiki, T. Sagawa, Y. Takahashi, Unconditional quantum-noise supres-
sion via measurement-based quantum feedback. Phys. Rev. Lett. 110, 163602 (2013)
75. C. Sayrin, I. Dotsenko, X. Zhou, B. Peaudecerf, T. Rybarczyk, S. Gleyzes, P. Rouchon, M.
Mirrahimi, H. Amini, M. Brune, J.-M. Raimond, S. Haroche, Real-time quantum feedback
prepares and stabilizes photon number states. Nature 477, 73–77 (2011)
76. R. Vijay, C. Macklin, D.H. Slichter, S.J. Weber, K.W. Murch, R. Naik, A.N. Korotkov, I. Siddiqi,
Stabilizing Rabi oscillations in a superconducting qubit using quantum feedback. Nature 490,
77–80 (2012)
77. S. Shankar, M. Hatridge, Z. Leghtas, K.M. Sliwa, A. Narla, U. Vool, S.M. Girvin, L. Frunzio,
M. Mirrahimi, M.H. Devoret, Autonomously stabilized entanglement between two supercon-
ducting quantum bits. Nature 504, 419–422 (2013)
78. J. Kerckhoff, R.W. Andrews, H.S. Ku, W.F. Kindel, K. Cicak, R.W. Simmonds, K.W. Lehnert,
Tunable coupling to a mechanical oscillator circuit using a coherent feedback network. Phys.
Rev. X 3, 021013 (2013)
79. D.J. Wilson, V. Sudhir, N. Piro, R. Schilling, A. Ghadimi, T.J. Kippenberg, Measurement-based
control of a mechanical oscillator at its thermal decoherence rate. Nature 524, 325–329 (2015)
80. H. Nakao, N. Yamamoto, Optimal control for perfect state transfer in linear quantum memory,
J. Phys. B: At. Mol. Opt. Phys. 50, 065501 (2017)
Index
A EPR entanglement, 212

Adapted process, 38 Non-degenerate parametric amplifier
Added noise, 246 (NOPA), 212
Adiabatically eliminate optical squeezing enhancement, 247
network synthesis, 90 quantum amplifier, 242
Adiabatic elimination, 26 time delays, 224
Algebraic loops Coherent state
network synthesis, 76 bosonic field, 68
Annihilation process, 38 single-mode, 68
Coherent vectors, 37
Commutant, 127
B Completely passive linear quantum network
Back-action (BA), 226 synthesis, 109
back-action evasion (BAE), 226, 228 Completely passive systems, 104
Back-action evasion (BAE) Concatenation product, 76
coherent feedback control, 229 Conditional state, 134
Beam splitter, 86 Continuous-variable, 203
Belavkin equation, 129 Controllability
Below threshold, 21 matrix, 8
Bogoliubov, 50 subsystem, 8
Bogoliubov transformation Counting process, 38
Shale’s Theorem, 51 Creation process, 38
C D
Canonical commutation relation (CCR), 11 Decoherence-free (DF), 232, 233
Canonical CV cluster state, 207 Degenerate parametric amplifier (DPA), 23,
Cascade realization 82, 138
network synthesis/transfer function real- Density matrix, 125
ization, 107 Detuning, 18
Cavity, 17 Direct interaction, 91, 230
mode, 17 Direct interaction Hamiltonian, 76
mode-cleaning cavity (MCC), 19
Center frequency, 18
χ (2) optical crystal, 19 E
Coherent feedback (CFB), 229 Einstein–Podolsky–Rosen
Coherent feedback control, 164, 181 EPR entanglement, 23
Communications and Control Engineering, DOI 10.1007/978-3-319-55201-9
260 Index
Einstein–Podolsky–Rosen (EPR), 19 K
Energy balance identity, 235 Kalman filter, 9
Environment engineering, 204
Exponential vectors, 37
L
Lindblad operator, 204
F
Linear matrix inequality, 167, 180
Faraday interaction, 28, 251
Linear quantum system, 43
Feedback control, 153
asymptotically stable, 60
coherent feedback control, 163, 181
Bogoliubov transformation, 50
coherent feedback H ∞ control, 181
completely passive, 53
coherent feedback LQG control, 164
complex mode form, 44, 46
controlled filtering equation, 156
Lyapunov differential equation, 60
controlled Hudson–Parthasarathy
marginally stable, 60
QSDE, 155
parameterization, 57
measurement-based feedback control,
physical realizability, 55
153
physical realizability conditions, 56
measurement-based LQG control, 157
real quadrature form, 44, 45
Filtering, 9, 156
controlled filtering equation, 156 stability, 59
quantum, 128 transfer function, 49
Fock space, 36 unstable, 60
annihilation process, 38 LQG control
counting process, 38 classical LQG control, 9
creation process, 38 coherent feedback LQG control, 164
fundamental processes, 38 measurement-based LQG control, 157
gauge process, 38 Lyapunov equation, 205
Fundamental processes on boson Fock
space, 38
M
Markov, 13
G assumption, 13
Gain–bandwidth constraint, 244 quantum Langevin equation, 14
Gauge process, 38 quantum white noise process, 15
Gaussian cluster state, 204 Markov assumption, 39
Gaussian state Master equation, 135
bosonic field, 64 Measurement, 126
generalized Araki–Woods representa- Measurement-based feedback control, 153,
tion, 62 157
pure Gaussian state, 206 quantum Kalman filtering, 140
single-mode, 61 separated law, 159
Model matrix, 96
Mode-matching, 18
H
H ∞ feedback control, 181
Hard realization, 75
Heat bath, 12 N
Heisenberg equation, 11 Network synthesis, 73, 91
Heisenberg uncertainty relation, 24, 227 completely passive systems, 104
Hudson–Parthasarathy QSDE, 132 direct interaction Hamiltonian, 76
Hurwitz, 21 hard realization, 75
model matrix, 96
quantum feedback network, 96
I reduced Markov models, 97
Incompatible, 126 singular perturbation, 26, 115
Index 261
soft realization, 75 interaction, 136

static linear optical devices, 84 QND observable, 29
strict realization, 75 Quantum probability space, 125
transfer function realization, 75, 106, 109 composite, 126
Non-degenerate optical parametric amplifier Quantum state reduction, 137
(NOPA), 19, 212 Quantum stochastic calculus, 36
Nullifier, 209 adapted process, 38
annihilation process, 38
coherent states, 67
O coherent vectors, 37
Observability counting process, 38
matrix, 8 creation process, 38
subsystem, 9 exponential vectors, 37
Observable, 125 Fock space, 36
Open-loop control, 12 fundamental processes, 38
Open quantum systems, 12 gauge process, 38
Optical cavity, 81 Gaussian states, 61
Optical low-pass filter, 18 quantum Itō product rule
Optical parametric oscillator (OPO), 23 fundamental processes, 40
Opto-mechanical oscillator, 25 quantum stochastic integrals, 40
quantum stochastic differential equa-
tions, 40
P quantum stochastic integral, 39
Phase-preserving amplifier, 240 Quantum stochastic differential equation
Phase-sensitive amplification, 240 (QSDE), 18, 40
Phase shifter, 85 left cocycle, 41
Pump beam, 19 non-demolition property, 42
second quantization of the one-particle
left shift operator, 41
Q self non-demolition property, 42
Quantum amplification, 239 Stone generator, 41
Quantum conditional expectation, 126, 127 Quasi-local, 208
Quantum feedback control, see feedback Qubit, 27
control
Quantum feedback network, 96
Quantum filtering, 42, 129 R
conditional state, 134 Radiation pressure force, 25
degenerate parametric amplifier (DPA), Rank-constrained LMI, 167
138 Realization problem, 74
innovations process, 129 Realization theory, 73
martingale, 130 Reduced Markov models, 97
master equation (ME), 135 Reducible network, 76
quantum Kalman filtering, 140 Reservoir engineering, 204
quantum non-demolition (QND), 129 Ricatti differential equation, 9, 140
self-non-demolition, 128 Rising exponential function, 236
stochastic master equation (SME), 135 Risk-sensitive estimators, 150
von Neumann algebra, 128, 132 Robust filtering, 144
Quantum Itō product rule Rotating frame, 18
fundamental processes, 40 Rotating-wave approximation (RWA), 13
Quantum stochastic integrals, 40
Quantum Langevin equation, 14
Quantum metrology, 226 S
Quantum noise limit (QNL), 24, 246 Separation principle, 10
Quantum non-demolition (QND), 28 Series product, 76
262 Index
Shot noise, 227 Stochastic master equation, 135

Singular perturbation, 26, 90, 115 Strict realization, 75
Soft realization, 75 Synthesis problem, 74
Spectral theorem, 125
Spin-coherent state, 251
Spin-squeezed state, 251 T
Squeezed Gaussian optical field, 24 Transfer function realization, 75, 106, 109
Squeezer, 87 Two-level system, 126
Standard quantum limit (SQL), 226, 227 Two-mode squeezed state, 22
∗-algebra, 124 Two-mode squeezing, 83
∗-isomorphism, 125 Two-mode squeezing Hamiltonian, 83
State, 124 Two-mode squeezing spectra, 215
pure, 125
State-space representation, 74
Static linear optical devices, 84 Z
Static linear optical network, 87 Zero dynamics principle, 235

Buku - Linear Dynamical Quantum Systems - Analysis, Synthesis, and Control 2022

Uploaded by

Copyright:

Available Formats

Buku - Linear Dynamical Quantum Systems - Analysis, Synthesis, and Control 2022

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Buku - Linear Dynamical Quantum Systems - Analysis, Synthesis, and Control 2022

Uploaded by

Copyright:

Available Formats

Communications and Control Engineering

Linear Dynamical Quantum

ISSN 0178-5354 ISSN 2197-7119 (electronic)

Library of Congress Control Number: 2017934052

© Springer International Publishing AG 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

of quantum feedback control theory, and linear quantum systems in particular.

Sydney, Australia Hendra I. Nurdin

2.1.4 The Hudson–Parthasarathy Quantum Stochastic

3.2.3 Main Synthesis Results . . . . . . . . . . . . . . . . . . . . . . . . .... 99

5.2.2 Numerically Solving the Rank-Constrained

6.6 Feedback Control Experiments . . . . . . . . . . . . . . . . . . . . . . . . .... 246

jˆi A ket, a normalized vector on a Hilbert space

vNfyðsÞ; 0  s  tg von Neumann algebra generated by the spectral projections

Abstract This chapter gives a brief overview of and introduction to quantum

© Springer International Publishing AG 2017 1

1.1 Quantum Feedback Control: A Brief History

In the early 1990s, a version of Belavkin’s filtering theory was independently

H ∞ control, robust estimation and control, stability analysis, system identification,

1.2 Classical Linear Systems and Control Theory

1.2.1 Classical Linear Systems

equation of motion, or more generally the Hamilton equation

1.2.2 Linear Systems and Control Theory

C = [B1 , AB1 , A2 B1 , . . . , An−1 B1 ],

can be explicitly calculated by the following Kalman filter:

where V (t) represents the estimation error covariance matrix:

Fig. 1.1 Diagram of Dynamics

where (P, Q, R) are the weighting matrices of the appropriate dimensions (Q is

K̇ (t) + K (t)A + A K (t) − K (t)B1 Q −1 B1 K (t) + P = O, (1.10)

1.2.3 Toward Systems and Control Theory for Linear

systems, as alluded to above, are relevant for applications such as nano-mechanical

1.3 Closed Linear Quantum Systems

[q, p] = qp − pq = 2ı. (1.12)

q(t) = cos(ωt)q(0) − sin(ωt) p(0), p(t) = sin(ωt)q(0) + cos(ωt) p(0),

thus, [q(t), p(t)] = [q(0), p(0)] = 2ı holds for all t.

1.4 Open Quantum Systems, the Markov Approximation,

1.4.1 Open Quantum Systems

1.4.2 Illustration of the Markov Approximation and Markov

[bω , bω ] = 0, [bω , bω∗  ] = δ(ω − ω  ), (1.16)

indicating explicitly the modeling of the bath as a collection of independent infinitesi-

Likewise, an arbitrary system variable X (t) obeys

satisfies the following equation (interpreted as a boundary condition):

1.5 Linear Dynamical Quantum Systems: Description

1.5.1 Optical Cavities

Fig. 1.2 Mode-cleaning Controller

As a consequence, it becomes apparent that the optical cavity can be modeled as a

The QSDE will be discussed in more detail in the next chapter.

1.5.2 Non-degenerate Optical Parametric Amplifiers

CCR [ai , a ∗j ] = δi j and [ai , a j ] = 0 and are coupled by an interaction Hamiltonian

where q j = a j + a ∗j and p j = (a j − a ∗j )/ı. Likewise, for the noise operators we

can then be expressed as

˜ q [ıω] = C1 Z [ıω] + D1 [ıω] ,

where C1 = [1 0 1 0]C, C2 = [0 1 0 −1]C, D1 = [1 0 1 0]D, and D2 = [0 1 0 −1]D.

V = V+ (ıω) + V− (ıω) < 4. (1.32)

A perfect maximally entangled continuous-mode EPR state would have V+ (ıω) =

Define H j [ıω] = C j (ıω I − A)−1 B + D j ( j = 1, 2), then it is straightforward

When κ = 0, we have that [80]

with k = /γ taking a value 0 ≤ k < 1 in the typical below threshold operation.2

vNfyðsÞ; 0 s tg von Neumann algebra generated by the spectral projections

K̇ (t) + K (t)A + A K (t) − K (t)B1 Q −1 B1 K (t) + P = O, (1.10)

[bω , bω ] = 0, [bω , bω∗ ] = δ(ω − ω ), (1.16)

described by the Hamiltonian Hrp = γ q1 a2∗ a2 . Roughly speaking, the oscillator is

A(t) = [ A1 (t) A2 (t) . . . Am (t) ] ,

dU(t) = Tr (S − I) d(t) + dA(t)∗ L − L ∗ SdA(t)

V (t, τ )V (τ , s) = ∗τ U(t − τ )τ ∗s U(τ − s)s

[x, x ] = xx − (xx ) = 2ıJn ,

Ao = 2Jn (R + {K ∗ K}),

w(t) = 2({A1 (t)}, {A1 (t)}, . . . , {Am (t)}, {Am (t)}) ,

A = 2Jn (R + {K ∗ K}), (2.6)

where + = [ωjk+ ] and − = [ωjk− ] are complex n × n matrices possessing the

and define the · involution operator that acts on a 2k × 2l matrix X as

Notice that, by definition, W = W −1 and W reduces to a unitary matrix when