Q UA N T U M I N F O R M AT I O N A N D Q UA N T U M O P T I C S

WITH SUPERCONDUCTING CIRCUITS

Superconducting quantum circuits are among the most promising solutions for the devel-
opment of scalable quantum computers. Built with sizes that range from microns to
tens of meters using superconducting fabrication techniques and microwave technology,
superconducting circuits demonstrate distinctive quantum properties such as superposition
and entanglement at cryogenic temperatures. This book provides a comprehensive and
self-contained introduction to the world of superconducting quantum circuits and how
they are used in current quantum technology. Beginning with a description of their basic
superconducting properties, the author then explores their use in quantum systems, showing
how they can emulate individual photons and atoms and ultimately behave as qubits within
highly connected quantum systems. Particular attention is paid to cutting-edge applications
of these superconducting circuits in quantum computing and quantum simulation. Written
for graduate students and junior researchers, this accessible text includes numerous
homework problems and worked examples.

j ua n j o s é g a r c í a r i p o l l is Senior Researcher at the Spanish Research Council

(CSIC), and a leading researcher in the design of quantum hardware for quantum technolo-
gies. He has been at the forefront of superconducting quantum circuits research for several
years and has published more than a hundred papers on related topics.

Published online by Cambridge University Press

Published online by Cambridge University Press
 Q UA N T U M I N F O R M AT I O N A N D
Q UA N T U M O P T I C S W I T H
SUPERCONDUCTING CIRCUITS

J UA N J O S É G A R C Í A R I P O L L
Institute of Fundamental Physics (IFF), CSIC

Published online by Cambridge University Press

 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107172913
DOI: 10.1017/9781316779460
© Juan José García Ripoll 2022
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2022
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: García Ripoll, Juan José, author.
Title: Quantum information and quantum optics with superconducting circuits /
Juan José García Ripoll.
Description: New York : Cambridge University Press, 2022. |
Includes bibliographical references and index. |
Summary: “The dawn of the 20th century brought us the birth of quantum mechanics
and a deeper understanding of the miscroscopic world. The new theory describing photons,
atomic spectra and many other physical processes, postulates that the microscopic world is, in its
truest essence, probabilistic. Particles such as electrons or photons move or “propagate” as probability
waves to be detected at a given position, or in a given state. However, those waves or wavefunctions are
very different from a mere representation of our ignorance about the world”– Provided by publisher.
Identifiers: LCCN 2021061918 (print) | LCCN 2021061919 (ebook) |
ISBN 9781107172913 (hardback) | ISBN 9781316779460 (epub)
Subjects: LCSH: Quantum theory. | Quantum optics. | BISAC: SCIENCE / Physics / Quantum Theory
Classification: LCC QC174.12 .G359 2022 (print) | LCC QC174.12 (ebook) |
DDC 535/.15–dc23/eng20220521
LC record available at https://lccn.loc.gov/2021061918
LC ebook record available at https://lccn.loc.gov/2021061919
ISBN 978-1-107-17291-3 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

Published online by Cambridge University Press

 Contents

List of Figures page xi

List of Tables xiii
Notation xiv
1 Introduction 1
1.1 The Book 4
1.2 Acknowledgments 6
2 Quantum Mechanics 7
2.1 Canonical Quantization 7
2.1.1 Hamiltonian Equations 9
2.1.2 Quantum Observables 10
2.1.3 Unitary Evolution 12
2.2 Two-Level Systems 13
2.3 Density Matrices 14
2.4 Measurements 16
3 Superconductivity 19
3.1 Microscopic Model 22
3.2 Macroscopic Quantum Model 24
3.3 Superfluid Current 25
3.4 Superconducting Phase 26
3.5 Gauge-Invariant Phase 27
3.6 Fluxoid Quantization 29
3.7 Josephson Junctions 30
4 Quantum Circuit Theory 34
4.1 Introduction 34
4.1.1 What Makes a Circuit Quantum 34

Published online by Cambridge University Press

 vi Contents

4.1.2 How Do We Work with Quantum Circuits? 38

4.2 Circuit Elements 40
4.2.1 Capacitor 41
4.2.2 Inductor 42
4.2.3 Josephson Junctions or “Nonlinear Inductors” 43
4.2.4 Other Elements 44
4.3 Quantization Procedure 44
4.4 LC Resonator 49
4.5 Transmission Line 50
4.6 Charge and Transmon Qubits 52
4.7 SQUIDs 53
4.7.1 rf-SQUID 53
4.7.2 dc-SQUID 55
4.8 Three-Junction Flux Qubit 57
4.9 Number-Phase Representation 59
Exercises 61
5 Microwave Photons 63
5.1 LC Resonator 63
5.1.1 Energy Quantization and Photons 63
5.1.2 Hamiltonian Diagonalization 65
5.1.3 Phase Space Dynamics 67
5.1.4 Are There Real Photons? 68
5.2 Transmission Lines or Waveguides 69
5.2.1 Periodic Boundary Conditions 70
5.2.2 λ/2 and λ/2 Microwave Cavities 71
5.2.3 Tunable Cavities 73
5.3 Three-Dimensional Cavities and Waveguides 75
5.4 Photon States 76
5.4.1 Fock States 76
5.4.2 Thermal States 77
5.4.3 Coherent States 78
5.4.4 Schrödinger Cat States 79
5.4.5 Single-, Two-, and Multimode Squeezed States 80
5.4.6 Wigner Functions and Gaussian States 82
5.5 Gaussian Control of Microwave Photons 85
5.5.1 Coherent Drivings and Displacement Operations 85
5.5.2 Coupling to an Environment 87
5.5.3 Cavity Spectroscopy 90
5.5.4 Losses and Heating 92

Published online by Cambridge University Press

 Contents vii

5.5.5 Beam Splitters and Circulators 93

5.5.6 Amplification 95
5.5.7 Photon Quadrature Measurements 99
5.6 Conclusion 103
Exercises 103
6 Superconducting Qubits 106
6.1 What Is a Qubit? 106
6.1.1 From Logical to Physical Qubits 106
6.1.2 Qubit Hamiltonian 110
6.1.3 Interaction Picture 111
6.1.4 Single-Qubit Gates 112
6.1.5 Decoherence and Dephasing 112
6.1.6 Relaxation and Heating 115
6.2 Charge Qubit 116
6.2.1 Coulomb Blockade 116
6.2.2 The Actual Superconducting Charge Qubit 117
6.2.3 Qubit Hyperbola 120
6.2.4 Charge Qubit History 121
6.3 Transmon Qubit 122
6.3.1 Moving Particle Picture and Energy Bands 123
6.3.2 Transmon as Anharmonic Oscillator 125
6.3.3 Josephson Junctions and the Mathieu Equation 126
6.3.4 Transmon as Qubit 128
6.4 Flux Qubit 130
6.4.1 Frustration and Current States 130
6.4.2 rf-SQUID Qubit 132
6.4.3 Persistent Current Qubit 135
6.4.4 General Operation 138
6.5 Qubit–Qubit Interactions 140
6.5.1 Dipolar Magnetic Interaction 141
6.5.2 Dipolar Electric Interaction 142
6.5.3 Coupling Tunability 144
6.5.4 Mediated Interactions and Tunable Couplers 145
6.6 Qubit Coherence 147
Exercises 149
7 Qubit–Photon Interaction 156
7.1 Qubit–Line Interaction Models 157
7.1.1 Dipolar Interaction 157
7.1.2 Spin-Boson Hamiltonian 159

Published online by Cambridge University Press

 viii Contents

7.1.3 Spectral Function and Spin-Boson Regimes 160

7.1.4 Rotating Wave Approximation 163
7.2 Waveguide-QED 164
7.2.1 Wigner–Weisskopf Approximation 165
7.2.2 Input–Output Relations 166
7.2.3 Spontaneous Emission Spectrum 167
7.2.4 Single-Photon Scattering 169
7.2.5 Quantum Links 172
7.3 Cavity-QED 173
7.3.1 Quantum Rabi and Jaynes–Cummings Models 175
7.3.2 Jaynes–Cummings Ladder 177
7.3.3 Vacuum Rabi splitting 178
7.3.4 Rabi Oscillations: Weak and Strong Coupling 179
7.3.5 Ultrastrong Coupling 182
7.3.6 Multiple Qubits 183
7.3.7 Off-Resonant Qubits and Dispersive Coupling 184
7.4 Circuit-QED Control 185
7.4.1 Direct Cavity Spectroscopy 185
7.4.2 Qubit Dispersive Measurement 187
7.4.3 Two-Tone Spectroscopy 190
7.4.4 Single-Photon Generation 190
7.4.5 Qubit Reset 191
7.4.6 Cavity Fock States Superpositions 192
7.4.7 Cavity Schrödinger Cats 193
Exercises 194
8 Quantum Computing 198
8.1 Quantum Circuit Model 198
8.2 Quantum Registers 202
8.2.1 Measurements 202
8.2.2 Qubit Reset 204
8.2.3 Architectural Decisions 204
8.3 Gate Toolbox 206
8.3.1 Universal Set of Gates 206
8.3.2 Two-Qubit Exchange Gates (iSWAP) 207
8.3.3 Two-Qubit Tunable Frequency CZ Gate 209
8.3.4 Two-Qubit Tunable Coupling CZ Gate 212
8.4 Tomography and Error Characterization 213
8.4.1 Classes of Errors 213
8.4.2 Error Models: Completely Positive Maps 214

Published online by Cambridge University Press

 Contents ix

8.4.3 Error Quantification: Fidelity 217

8.4.4 Randomized Benchmarking 219
8.5 Fault-Tolerant Quantum Computers 221
8.5.1 Local Errors and Global Qubits 222
8.5.2 Passive versus Active Error Correction 223
8.5.3 Stabilizer Codes 224
8.5.4 Surface Code 225
8.5.5 Fault-Tolerant Thresholds and Outlook 230
8.6 Near-Term Intermediate Scale Quantum Computers 232
8.6.1 What Is NISQ? 232
8.6.2 Hybrid Quantum Computers 233
8.6.3 Quantum Volume 234
8.7 Outlook 234
Exercises 236
9 Adiabatic Quantum Computing 239
9.1 Adiabatic Evolution 239
9.1.1 Landau–Zener and Qubit Adiabatic Control 241
9.1.2 The Adiabatic Theorem 243
9.1.3 Circuit-QED Applications of Adiabatic Theorem 244
9.2 Adiabatic Quantum Computing Model 245
9.2.1 The Adiabatic Quantum Computing Algorithm 246
9.2.2 Resource Accounting 247
9.3 The Choice of Hamiltonian 248
9.3.1 A Primer on Complexity Classes 248
9.3.2 QUBO and NP-Complete Hamiltonian Problems 250
9.3.3 QMA-Complete Problems 251
9.3.4 Scaling of Resources 252
9.4 D-Wave’s Quantum Annealer 253
9.4.1 D-Wave’s Architecture 254
9.4.2 Device Operation 257
9.4.3 Performance Analysis 260
9.5 Summary and Outlook 266
Exercises 267
Appendix A Hamiltonian Diagonalizations 270
A.1 Tridiagonal Matrix Diagonalization 270
A.1.1 Periodic Boundary Conditions 270
A.1.2 Open Boundary Conditions 271
A.2 Harmonic Chain Diagonalization 272

Published online by Cambridge University Press

 x Contents

A.3 Schrieffer–Wolff Perturbation Theory 273

A.3.1 Nondegenerate Perturbation Theory 273
A.3.2 Degenerate Perturbation Theory 274
A.3.3 Considerations 275
Appendix B Open Quantum Systems 277
B.1 Nonunitary Evolution 277
B.2 Master Equations 278
B.2.1 Lindblad Equation 279
B.2.2 Linear System–Bath Coupling 280
B.2.3 System in a Thermal Bath: Cooling and Heating 280
B.2.4 Perturbations and Generalizations 281
B.2.5 Strong Nonlinearity and Multilevel systems 282
B.3 Input–Output Theory 282
B.3.1 Memory Function 283
B.3.2 Markovian Approximation: Decay Rate and Lamb Shift 284
B.3.3 Input–Output Relations 284
B.3.4 Spectroscopy 286
References 287
Index 299

Published online by Cambridge University Press

 Figures

2.1 Typical microwave measurement setup. page 17

2.2 Quantum circuit for a generalized quantum measurement. 18
3.1 Josephson junction circuit and quantum model. 30
4.1 Thermal fluctuations in a quantum device. 37
4.2 Linear circuit elements and notation. 41
4.3 Circuit quantization procedure. 45
4.4 Quantization of the LC resonator equivalent circuit. 49
4.5 Quantization of the equivalent circuit for a microwave guide. 51
4.6 Equivalent circuit for the charge qubit. 52
4.7 Equivalent circuit for the rf-SQUID qubit. 54
4.8 Equivalent circuit for a dc-SQUID. 55
4.9 Washboard potential for a dc-SQUID. 57
4.10 Equivalent circuit for a three-junction flux qubit. 58
5.1 Photographies of superconducting resonators. 64
5.2 Sketch of a waveguide and equivalent circuits. 70
5.3 Dispersion relation and eigenmodes of a transmission line. 72
5.4 Eigenmodes of a λ/2 and λ/4 transmission line resonators. 73
5.5 Examples of Wigner functions. 84
5.6 LC-resonator coupled to semi-infinite transmission line. 87
5.7 Beam splitters and circulators. 94
5.8 Quadrature measurements with mixers and IQ-mixers. 102
6.1 Anharmonic spectrum of a physical qubit. 109
6.2 Picture and equivalent circuit of charge and transmon qubits. 118
6.3 Energy levels of a charge qubit. 119
6.4 Energy levels and eigenfunctions of the transmon qubit. 124
6.5 Photography and equivalent circuits of two flux qubits. 131
6.6 Energy levels of the rf-SQUID. 132
6.7 Inductive energy and eigenstates of the three-junction flux qubit. 137

xi

Published online by Cambridge University Press

 xii Figures

6.8 Qubit–qubit dipolar interactions. 141

6.9 Qubit coherence times. 149
7.1 Qubits interaction with a transmission line. 158
7.2 Spontaneous emission and scattering. 165
7.3 Spontaneous emission of a photon by a transmon qubit. 168
7.4 Scattering of a propagating photon by a qubit. 171
7.5 Photography of a transmon in a microwave cavity. 173
7.6 Energy levels of the Jaynes–Cummings model. 177
7.7 Transmission spectra for a qubit-cavity setup in the strong and ultrastrong
coupling regimes. 179
7.8 Dynamics in the weak and strong coupling regimes. 181
7.9 Spectroscopy schemes for a qubit in a cavity. 185
7.10 Two-tone spectroscopy of a charge qubit. 191
8.1 Quantum computing circuit, schematic notation. 200
8.2 Superconducting qubit quantum register. 203
8.3 Phase gate with two transmons. 210
8.4 Surface code layout and circuits. 226
8.5 Two-qubit gate performance. 231
9.1 Energy gaps, adiabatic evolution, and Landau–Zener processes. 240
9.2 D-Wave architectural details. 254

Published online by Cambridge University Press

 Tables

3.1 Critical temperatures of various superconducting materials. 20

4.1 Equivalence between frequencies and temperatures. 37
4.2 Mathematical representations of the flux and charge operators. 61
6.1 Different types of qubits. 108
7.1 Regimes of the ohmic spin-boson model. 162
8.1 Most common single-qubit and two-qubit unitaries. 201
8.2 Completely positive maps for usual error models. 215

xiii

Published online by Cambridge University Press

 Notation

 2
Ô Uncertainty of operator, (Ô)2 = 12 Ô Ô † + Ô † Ô −  Ô  .
Variance of observable, Ô † = Ô, as (Ô)2 = (Ô − Ô)2  .
C Capacitance
EJ Josephson energy
H.c. Hermitian conjugate, as in a + H.c. = a + a †
L Inductance
L Lagrangian or Lindblad operator
x, s . . . Vectors of numbers such as x = (x1,x2, . . . ,xN )
h Planck constant, 6.626070040(81) × 10−34 J · s
h̄ = 2π h
Reduced Planck constant
φ Electric flux on a node or a branch of a circuit (Section 4.3)
0 = 2e h
Magnetic flux quantum, 2.067833831(13) × 10−15 Wb
ϕ0 = 0
2π
= h̄
2e
Flux-to-phase conversion
σ (H ) Spectrum or collection of eigenvalues of an operator H
σ x, σ y, σ z Pauli matrices
σ Vector of Pauli matrices σ = (σ x ,σ y ,σ z )
Linear capacitor
Linear inductor
Nonlinear inductance associated to a Josephson junction
Josephson junction: nonlinear inductor and capacitor
in parallel
Constant voltage source

xiv

Published online by Cambridge University Press

 1
Introduction

The dawn of the twentieth century brought us the birth of quantum mechanics and a
deeper understanding of the microscopic world. The new theory describing photons,
atomic spectra, and many other physical processes postulates that the microscopic
world is, in its truest essence, probabilistic. Particles such as electrons or photons
move or “propagate” as probability waves to be detected at a given position, or in a
given state. However, those waves or wavefunctions are very different from a mere
representation of our ignorance about the world.
Quantum mechanics is the mathematical language that we use to describe the
microscopic world. Quantum mechanical states – our wavefunctions – go beyond
our classical understanding of the world, accepting the possibility of a system to
coexist in a quantum superposition of two distinguishable configurations – i.e., an
atom in two positions, a photon in two polarizations, or a neutron moving in two
opposite directions. Only when we measure the state of the quantum mechanical
object, it collapses to a well-defined configuration, which may be different on each
realization of the experiment. Even more dramatically, multiple particles may be
in a collective superposition or entangled state, allowing arbitrary measurements
on these particle to produce random but perfectly correlated outcomes, irrespective
of the space and time separation between those measurements. These and other
predictions of quantum mechanics made some physicists such as Albert Einstein
deem the theory as “incomplete” or even inconsistent.
The end of the twentieth century and the beginning of the twenty-first have
witnessed an incredible collective effort to challenge the wildest predictions of
quantum mechanics in the broadest variety of experimental systems possible, from
photons, to atoms, all the way to macroscopic solid-state devices and circuits. Great
experimentalists of our time have revealed quantum properties in objects that we
can see with the naked eye, created complex quantum states of tens to thousands of
particles, and entangled photons and spread them to different points of the planet,

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 2 Introduction

to test their quantum mechanical correlations. In the process, not only has quantum
mechanics been validated, but we have become incredibly good at controlling
quantum systems.
Trapped ions were one of the first systems to enter this controlled quantum regime
(Leibfried et al., 2003). Using electromagnetic traps and laser cooling, experimen-
talists may isolate one or more charged particles, cooling some of their properties
close to the absolute. This makes it possible to observe the discrete or quantized
nature of the atomic excitations, not only at the level of the electronic states, but also
with respect to quanta of energy stored in the motional states of the trapped atom.
If instead of charged particles we use neutral atoms, we can trap more particles
and enter the regime of mesoscopic quantum systems. Experiments in 1995 showed
how to to trap and evaporatively cool a cloud of 106 alkali atoms in an almost perfect
vacuum down to few nanokelvins. At these temperatures, the ultracold atoms form
a singular state of matter called Bose–Einstein condensate (BEC) (Townsend et al.,
1997), in which all bosonic atoms are described by the same wave function, φ(x),
whose dynamics and quantum properties can be engineered at will.
Starting from a perfectly controlled state of matter, such as the condensate, we can
build extremely sophisticated quantum states and physical devices. We can split the
atoms of a BEC into the pockets of an optical lattice, to recreate a crystal of bosons
and simulate quantum magnetism and even Luttinger liquid physics. Bose–Einstein
condensates can be used to cool down fermionic atoms and create Fermi seas,
study Bardeen–Cooper–Schrieffer (BCS) superconductivity or even implement the
Hubbard model for electrons in a solid. And using time-dependent controls, we can
create large amounts of entanglement and squeezing that can be used for sensing or
for foundational experiments in quantum information.
The beauty of atomic, molecular and optical (AMO) physics’ bottom-up approach
can hardly be exported to solid-state devices. Condensed-matter objects are simply
too large. We cannot control all atoms in a piece of metal and convince them
to adopt a perfect state. Moreover, solid devices rarely exist in perfectly isolated
environments. They are always in contact with other elements – contacts, substrates,
measurement apparatuses – at higher temperatures than the AMO setups we just
discussed.
In a way, once we overcome the barriers associated to cooling and trapping, it is
not surprising that we can reveal quantum mechanical phenomena in large atomic
setups – it just confirms that the rules of quantum mechanics extend to very large
composite systems. The million dollar question, however, was whether those rules
extend all the way to solid-state and condensed matter systems!
This was indeed the great challenge for condensed matter systems around the
1990s. Can we prepare a perfect solid-state quantum system? Can we build a
simple, atomic-like device, with a perfectly controlled state? The approach to

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 Introduction 3

this problem is subtly different from AMO. Starting from a macroscopic object –
a gated or self-assembled quantum dot, impurities in diamond, nanoresonators,
or superconducting circuits – we select one or two degrees of freedom in which
we seek evidence of quantum phenomena. The candidate for a quantum degree of
freedom may be the charge of a small superconducting island, the electrical current
along a metallic loop, or the wavefunction of electrons in a dot or a color center,
for example. Experimentalists focus on one property and work toward isolating
it from their environments, cooling them down, and engineering better and better
quantum states.
This challenge, the quest for an artificial atom – what we now call a qubit; see
Chapter 6 – was still ongoing when I first learned about the field of superconduct-
ing circuits. The review article by Makhlin et al. (2001) showed some promising
candidates in the forms of charge qubits (Section 6.2) or three-Josephson junction
qubits (Section 4.8), but quantum superpositions were fragile and easily killed by
the environment.
The surprise came in year 2004, when the Yale team (Wallraff et al., 2004) rad-
ically improved the lifetime of charge qubits by placing them inside microwave
resonators. That work, and subsequent works at various groups in the University of
California, Santa Barbara (UCSB), Saclay, Karlsruhe Institute of Technology (KIT),
etc., sparked the beginning of a productive intersection between superconducting
circuit technology, quantum optics, and quantum information, which we now call
circuit-QED.
We live in the second quantum revolution (Dowling and Milburn, 2003), in which
quantum mechanics becomes a tool for industrial applications and practical devices.
These applications are generally known as quantum technologies, because they use
the unique properties of quantum mechanics – quantum superpositions, entangle-
ment, noncontextuality, etc. – to create actual technologies for quantum communi-
cation, quantum cryptography, quantum sensing, quantum simulation, and quantum
computing.
Superconducting circuits are at the center the quantum revolution. Supercon-
ductors are the basis for the most accurate single-photon detectors, which are
used in advanced setups for quantum communication. We rely on superconducting
qubits for some of the most powerful and accurate quantum computers to date
(see Chapter 8), and also to simulate large and complex problems from quantum
magnetism (Chapter 9). Superconducting circuits have also found their way,
as controls and interfaces for other microwave-based quantum technologies,
from quantum dots to nanomechanical resonators. Understanding how these
circuits work, how they are designed, and how they are operated is a natural
path for both theoreticians and experimentalists alike. This book is a self-contained
undergraduate manual designed to help you gain this understanding.

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 4 Introduction

1.1 The Book

The book is structured in three parts. The first part introduces the platform, super-
conducting circuits, and a common language, formed by quantum mechanics and
circuit quantization. The second part introduces the two fundamental objects we
build using superconductors: these are photonic devices and artificial atoms. They
are studied separately, and then combined to develop the theory of circuit quan-
tum electrodynamics, or circuit-QED. The third part concerns the physics that can
be explored and the technologies that can be built using superconducting circuits.
Along separate chapters and sections, we learn about how to build quantum com-
puters, quantum simulators, quantum optimizers, and other quantum technologies
using circuits or hybrid setups.
The structure and the level of the book also allow different applications. This text-
book was born and has been used as an introductory course in quantum optics. Using
qubits and microwave photons, it is possible to review the basic concepts in light–
matter interaction, the Rabi, Jaynes–Cummings and Dicke model, superradiance,
and subradiance, and even understand sophisticated models of low-dimensional
photonics. But the same book can be repurposed to focus on the design and con-
struction of practical quantum computers or to learn about applications in the field
of quantum simulation and many-body physics.
The book is mostly self-contained. It assumes familiarity with the formalism of
quantum mechanics, including key concepts such as pure states, density matrices,
observables, and the Schrödinger equation, but these concepts are reviewed as they
are used. At different stages, we will also introduce minimal concepts in second
quantization and field theory, such as the notion of modes, creation and annihilation
operators, commutation relations, bosonic statistics, etc. We also review important
concepts from quantum information theory, such as qubits, entanglement, quantum
gates, etc. The most important ideas are reinforced through selected exercises that
students are encouraged to do.
Building on these prerequisites, the reading order of this book would be the fol-
lowing one. In the first part of the book, Chapter 2 provides an overview of quantum
mechanics, focusing on concepts that are used throughout the book. In Chapter 3, we
introduce a minimal understanding of the theory of superconductivity, with a focus
on London’s mesoscopic models and the electrical properties of superconducting
circuits. Chapter 4 introduces the theory of quantum circuits. This is a set of rules
that allow us to find the mathematical model that describes a given superconducting
circuit. We derive these rules using as practical examples all the elementary circuits
that we will later study, such as qubits, microwave resonators, or superconducting
quantum interference devices (SQUIDs).

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 1.1 The Book 5

In the second part of the book, Chapter 5 focuses on the linear models of super-
conducting circuits, such as microwave LC resonators and waveguides (the equiva-
lent of coaxial cables), and showing how the quantization of the circuit gives birth
to microwave photons that can be created, manipulated, and measured. Chapter 6
studies superconducting circuits in the highly nonlinear regime, in which the circuit
acts as an artificial atom or qubit. We review the most popular qubit designs, such as
the transmon or the flux qubit, and develop tools to prepare and characterize those
qubits. Chapter 7 combines artificial atoms with microwave photons, developing
the theory of light–matter interaction. This chapter is a primer on concepts from
quantum optics – models for low-dimensional atom–light interaction, the physics
of spontaneous emission, photon absorption, and open quantum systems. We then
focus on the simpler setup of qubit–resonator interactions, introducing the Rabi and
Jaynes–Cummings models, and showing how qubits can control the state of light
and vice versa, cavities can be used to control and measure qubits.
The third part of the book focuses on the applications of superconducting circuits
to various quantum technologies. We begin in Chapter 8 studying universal quantum
computers in the circuit model. We offer a checklist of ingredients that are needed
to build a practical quantum computer, and investigate how those elements are built
and characterized in the superconducting platform. We also discuss the roadmap
toward fault-tolerant, error-corrected, and fully scalable quantum computers, and
what we can do with their small-scale, faulty versions in the near term. Finally,
Chapter 9 builds on the previous formal developments to introduce a design for an
adiabatic quantum computer, also known as quantum annealer. This chapter starts
with a formal description of this computational model, where a physical system is
adiabatically coerced into a configuration that represents the solution to a mathe-
matical problem. It then moves on to the actual superconducting architecture for a
quantum annealer using superconducting flux qubits, along the line of the D-Wave
machine, but referencing other later designs.
If the book is to be used with a strong focus on quantum optics or quantum
circuits, the student or teacher should at least cover Chapters 1–7. This should give
students an overview of the most relevant experiments in the field from 2004 to this
day, providing them with the language and tools to access more complex literature
and explore their own ideas. Chapters 8 and 9 in the third part of the book are
independent from each other, but require a good understanding of the first part of
the book. A minimal set of exercises is provided within each chapter. Solutions and
errata will be posted on the author’s webpage.1

1 http://juanjose.garciaripoll.com.

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 6 Introduction

1.2 Acknowledgments
This book would not be possible without the help of many colleagues and friends
who supported me both before and during the writing of the text. I was introduced
to the exciting field of superconducting circuits by my friend and colleague Enrique
Solano, who himself laid out the foundations to many ideas that are explored in the
realm of circuit-QED and photon measurements.
I have learned a lot interacting with fellow experimentalists, including Frank
Deppe, Achim Marx, Adrian Lupascu, Chris Wilson, Gerhard Kirchmair,
Mathieu Juan, and Aleksei Sharafiev. I am indebted to Pol Forn-Díaz, whose
curiosity and perseverance has strongly driven our interest in ultrastrong qubit-
microwave interactions and quantum annealing.
This book would not be born without the spark and motivation to write it, which
was ignited by my friend and outstanding physicist Oriol Romero-Isart. My visits to
Innsbruck – always a place for inspiration at various inflection points in my career –
and the lectures I gave there helped me move on and were the seed for this book.
I obviously owe many thanks to the students who have suffered various iterations of
this work, and also my former PhD student Borja Peropadre, whose thesis inspired
the first iterations of this project.
Special thanks must go to Carlos Navarrete, Guillermo Romero, Manuel Pino,
Paula García-Molina, Pol Forn Díaz, and Alp Sipahigil for carefully reading various
iterations of this manuscript.
None of this would be possible without the support of a family that has always
been there, even when I was not. I owe to them, not this book, but the privilege of a
life as scientist.
And finally, this book exists because of you, the reader. I hope it will help you
approach this field, learn about its beauty and its challenges, and inspire you to be
a better “quantum mechanic.”

https://doi.org/10.1017/9781316779460.001 Published online by Cambridge University Press

 2
Quantum Mechanics

As mentioned in the introduction, this book assumes some basic understanding of

quantum mechanics at the undergraduate level. This understanding is essential for
many reasons: not only to be able to write down realistic models, compute things,
and make predictions, but also because quantum physics is a world full of subtleties
and unexpected phenomena which may cause major confusion and headaches when
seen for the first time.
Consequently, here is a friendly warning: If you never took a course on quantum
physics and the formalism of quantum mechanics, take it, or grab some books that
can help you overcome this initial barrier. There are many good books to choose
from, such as the two volumes of Cohen-Tannoudji et al. (1977) or the wonderful
and complete book by Ballentine (1998). Simply take the one that resonates best
with you.
This said, I felt the need to write down a chapter where I could summarize many of
the concepts and tools from quantum mechanics that are repeatedly invoked in this
book. Some of these are pretty basic, such as the relation between the Schrödinger
equation and unitary operators, but others are a bit more subtle and extend beyond
a typical course on quantum mechanics, such as the notion of density matrices and
master equations. Please take this therefore as a “unifying” chapter that provides
a common language and definitions, and which you may freely skip if you have a
graduate or postgraduate level in the study of quantum physical systems.

2.1 Canonical Quantization

Quantum physics is an old field of research, whose birth we attribute to Planck’s
theory for the black-body radiation. A black body is an ideal object that can absorb
energy at any frequency of the spectrum, which incidentally implies that it is the
body that can emit energy most efficiently at any given frequency. In 1900, the

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 8 Quantum Mechanics

physicist Max Planck showed that experiments measuring the radiation from black
bodies could be explained by assuming that these objects – which at the time were
just perfect cavities with a tiny hole – could only exchange energy in fixed amounts
or quanta, determined by the frequency ν of the light emitted or absorbed by the
black cavity.
In his treatment, Planck models the excitations of the black body as a collection of
harmonic oscillators with frequencies that cover the measured spectrum – i.e., ν(k),
labeled by wave vectors k. The energy of those oscillators is quantized, which means
that the oscillators equilibrate to the same temperature by exchanging discrete units
of energy or quanta with the environment. If we could measure the state of the
black body, its energy would be a sum of the quanta nk that are stored in each
electromagnetic mode k:

E= hν(k) × nk, nk ∈ {0,1,2, . . .}. (2.1)
k

In this model, each oscillator has associated a quantum of energy hν(k) determined
by the frequency1 ν(k) and Planck’s constant h  6.62607004(81) × 10−34 J/Hz.
The collection of all integers |nk1 ,nk2 , . . . is a unique configuration of the black
body, which we call the quantum state, and the collection of all states is used to
develop a statistical model of the black body’s spectrum.
Despite his success, Planck was very wary of extending the idea of quanta to
the actual electromagnetic field. It was Einstein who made the connection between
Planck’s quanta and the existence of a particle of light, the photon. With this
particle, Einstein could explain in 1905 the photoelectric effect: Some materials
may convert light into an electrical current, but when the intensity of light is
lowered enough, this current becomes a series of discrete random bursts, which
Einstein associated with the absorption of photons. This successful explanation was
shortly followed by Bohr and Rutherford’s model of the atom, based on quantized
electronic orbits that explained the discrete spectra of light-emitting atoms. Barely
a decade later, Schrödinger (1926) and Heisenberg (1925) replaced all ad hoc
quantization ideas with two equivalent formulations of quantum mechanics based
on wave and matrix equations. In both theories, the discretization of energies is
a mathematical consequence of the discrete spectra of the operators that govern
the evolution of light and matter. Put to work, the newly born theory provided
quantitative explanations for the spectra of atoms, molecules, solids, and the
electromagnetic field itself, consolidating nonrelativistic quantum mechanics as

1 Ordinary frequencies are typically denoted by the letter ν and are measured in the S.I. unit of Hertz (Hz). Quite
often we will also use angular frequencies ω = 2π ν, sometimes denoted in rad/s or s−1 . In the first case, the
quantum of energy is given by hν, while in the second case it is given by Planck’s reduced constant h̄ω with
h̄ = h/2π .

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 2.1 Canonical Quantization 9

the tool for understanding chemistry, solid-state physics, nano-electronics, and

photonic devices, to name a few examples.
Schödinger’s and Heisenberg’s formulations are powerful theories that explain
the microscopic behavior of Nature in a bottom-up fashion: starting from elemen-
tary components – the electron, the proton, and the neutron – and fundamental
interactions – for instance, the Coulomb attraction between protons and electrons
mediated by the electromagnetic field – one builds a many-body equation whose
solution accounts for all the physics we observe in the laboratory. Unfortunately, the
bottom-up approach does not always scale well as we move on to larger systems. In
this book, we are concerned with solid-state superconducting devices that include
more than 1024 atoms, all collectively exhibiting quantum mechanical phenomena.
It is unfathomable to even think of writing an equation for all those particles, and
we are forced to seek effective descriptions that are consistent with the principles
and rules of the underlying quantum mechanical theory.
Shortly after the publication of Schrödinger’s and Heisenberg’s work, Paul Dirac
developed an alternative derivation of quantum mechanics, known as canonical
quantization, that establishes a link between the quantum model for a given object
(particle, field, etc.) and the dynamics that we would expect from it in a classical
world. In Chapter 4, we will apply this procedure to the quantization of an electrical
circuit, developing a quantum theory of superconducting circuits. This theory will
be consistent with the microscopic description introduced in Chapter 3, and it will
provide the appropriate limit of the circuit when temperatures are high enough that
superconductivity is lost, or quantum phenomena are masked.

2.1.1 Hamiltonian Equations

For simplicity, we will describe how canonical quantization works for a simple
object: a point-like particle2 with position x and momentum p = mẋ, moving in
an external potential V (x). The particle’s trajectory is governed by a set ordinary
differential equations – Newton’s equations – which we write in terms of the parti-
cle’s acceleration ẍ and the force ∇V (x) experienced by the particle
1
∇V (x).ẍ = − (2.2)
m
Newton’s equation can be derived from
 t a stationary action principle, as the trajec-
tory that minimizes the action S = t12 L(ẋ,x)dt. The functional S maps orbits x(t)
to real numbers according to the Lagrangian

2 As we will see in Chapter 4, this is not a futile exercise, because the harmonic potential V (x) = 1 mω2 x2 is
2
formally analogous to the simplest electrical circuit, an LC resonator, and describes how this circuit is actually
quantized.

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 10 Quantum Mechanics

1
L(ẋ,x) = mẋ2 − V (x). (2.3)
2
According to the stationary principle, a small perturbation of the particle’s true tra-
jectory xε (t) = x(t) + ε(t) should leave the action unperturbed up to second-order
corrections S[xε ] = S[x] + O(ε 2 ). This stationary principle produces Lagrange’s
equations
d ∂L ∂L
= , (2.4)
dt ∂ ẋn ∂xn
which are equivalent to the original Newtonian equations (2.2).
We now introduce a Hamiltonian formulation, where the functional that generates
the dynamical equations is a function of two canonically conjugate variables, x and
p, with the prescription
∂L
H (x,p) = pẋ − L(ẋ,x), . with p =(2.5)
∂ ẋ
This Legendre transform establishes a link between the particle’s velocity ẋ and
its canonical momentum p, which now replaces the former in all equations. The
transform also produces an object, the Hamiltonian H (x,p), governing the orbits
of the particle. More precisely, any observable O(x,p,t) that we can construct as a
function of the canonical variables and time evolves according to the Hamiltonian
equation
d ∂O
O = {O,H } + , (2.6)
dt ∂t
with the classical Poisson brackets
  ∂A ∂B ∂B ∂A

{A,B} = − . (2.7)
j
∂xi ∂pi ∂xi ∂pi

In particular, since {xi ,pj } = δij this prescription trivially recovers Newton’s equa-
tions, but now expressed as a set of first-order differential equations:
d ∂H 1 d ∂H
x= = p, p=− = −∇V (x). (2.8)
dt ∂p m dt ∂x

2.1.2 Quantum Observables

Dirac’s canonical quantization describes the transition from a Hamiltonian theory
of a classical particle to a quantum mechanical theory that preserves the particle’s
dynamical equations (2.6). First of all, following the axioms of quantum mechanics,
it introduces a Hilbert space of vector states describing our system. Next, it replaces
any measurable quantity, including x, p, and any other function thereof O(x,p), with

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 2.1 Canonical Quantization 11

linear Hermitian operators x̂, p̂, Ô acting on this Hilbert space. Each observable will
have a spectrum of eigenvalues and eigenstates determining all possible measure-
ment outcomes.
In our toy model, the quantum state of a particle that is at position r ∈ RN
is associated a vector |r in the Hilbert space, with the property x̂n |r = rn |r.
Generic states are constructed as quantum superpositions of different measurement
outcomes, such as the wavefunction ψ(r):

|ψ = ψ(r) |r dN r. (2.9)

The weights of the wavefunction ψ(r) ∈ C are complex numbers whose modulus
gives the probability distribution P (r) = |ψ(r)|2 that the particle is found at the
position r, if the observable x̂ is ever measured. States  are normalized, so that the
total probability adds up to one, 1ψ = ψ|ψ = |ψ(r)| dr = 1, and we can
2

define the expectation values of measurements:


x̂ψ = ψ|x̂|ψ = ψ(r0 )∗ ψ(r1 ) r0 | x̂ |r1  dN r0 dN r1 (2.10)

= r0 |ψ(r0 )|2 dr0 .

Note how, by using the orthogonality of position eigenstates r0 |x̂|r1  = r1 r0 |r1  =
r1 δ(r0 − r1 ), we recovered the formula for the average over the probability
distribution P (r).
Canonical quantization includes one final prescription that makes the algebra
of operators and states consistent with the classical limit of these equations. We
replace everywhere the Poisson brackets for classical variables with the commutator
between the respective observables {A,B} → −i[Â, B̂]/h̄, where [Â, B̂] = ÂB̂ −
B̂ Â. For the isolated particle, this prescription transforms {xn,pm } = δnm into

[x̂n, p̂m ] = i h̄δnmj . (2.11)

If our space of positions is continuous and contained in the region  ⊂ RN , we can

build our Hilbert space using integrable functions ψ(r) ∈ L2 (), associating posi-
tion and momentum with operators x̂n ψ(r) = rn ψ(r) and p̂n ψ(r) = −i h̄∂rn ψ(r)
that satisfy the commutation relations (2.11).
Irrespective of the implementation of our Hilbert space, (2.11) implies that canon-
ically conjugate operators are incompatible: They cannot share a common basis
of eigenstates. Thus, a state with a well-defined position |x0  cannot have a well-
defined value of the momentum. This leads to the Heisenberg uncertainty principle:
1
xn pm ≥ h̄δnm, (2.12)
2

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 12 Quantum Mechanics

which relates the variances of pairs of observables, xn = x̂n2  − x̂n 2 and pm .
As we will see later, this uncertainty has also a physical manifestation in the world of
quantum circuits, where position and momenta are replaced by voltage and intensity,
quantities that cannot be measured simultaneously with absolute precision.

2.1.3 Unitary Evolution

We have introduced quantum states and observables as two mathematical objects
that together predict the statistics of measurement outcomes. This information is
bound to change in time, as observables and states evolve. In canonical quantiza-
tion, the identification of Poisson brackets with commutators translates the classical
equation (2.6) into the Heisenberg equation:

dÔ i ∂ Ô
= − [Ô, Ĥ ] + . (2.13)
dt h̄ ∂t
In this model, the dynamics is generated by a Hamiltonian operator that results from
replacing the canonical variables with the corresponding observables
Ĥ = 2m p̂ + V (x̂). In the Heisenberg picture, observables change in time starting
1 2

from a well-known initial condition O(t0 ) = O0 . The states |ψ0  remain stationary,
and they are regarded as objects that map the changing observables to their
expectation values Ō(t) = Ô(t) = ψ0 |Ô(t)|ψ0  .
The Heisenberg equation is rather inconvenient: We have to work with big
and complex operators, and extracting the measurement statistics becomes a very
convoluted process. In many situations, we would rather work with an equation
that determines how states evolve from, say, an initially localized configura-
tion ψ0 (x) = δ(x − x0 ), spreading to other meaurement outcomes. This infor-
mation is provided by the Schrödinger equation or Schrödinger picture, whereby
observables have an immutable representation, but wavefunctions change in time
Ō(t) = Ôψ(t) = ψ(t)|Ô0 |ψ(t), with

i h̄∂t |ψ(t) = Ĥ |ψ(t) , with |ψ(t0 ) = |ψ0  . (2.14)

For our isolated particle in an external potential V (x), using the position represen-
tation, where p̂ = −i h̄∇, this results into a simple wave equation for the complex
amplitude of probability ψ(x):
1
i h̄∂t ψ(x) = (−i h̄∇)2 + V (x) ψ(x). (2.15)
2m
Even if they look very different, the Schrödinger and Heisenberg representa-
tions give the same predictions because they are both solved by a common unitary

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 2.2 Two-Level Systems 13

transformation called the evolution operator. This operator is the solution of an

enlarged Schrödinger equation:
d
i h̄ Û (t,t0 ) = Ĥ Û (t,t0 ), with Û (t0,t0 ) = 1. (2.16)
dt
In the case of constant Hamiltonians, the unitary operator U is a Lie rotation in the
Hilbert space, generated by the Hamiltonian:

Û (t,t0 ) = exp[−i(t − t0 )H /h̄]. (2.17)

This operator is unitary U U † = U † U = 1. It can be inverted Û (t2,t1 )−1 = Û (t1,t2 )

and solves both the Schrödinger |ψ(t) = Û (t,t0 ) |ψ0  and the Heisenberg equations
Ô(t) = Û (t0,t)Ô0 Û (t,t0 ), as mentioned before.

2.2 Two-Level Systems

Not all physical systems have continuous degrees of freedom. We are going to
work with smaller systems that only have two or three configurations that are active
in a given experiment. These discrete systems have smaller Hilbert spaces, with
wavefunctions defined in complex vector spaces. For a quantum system with two
possible states |0 and |1, the wavefunctions in the two-dimensional Hilbert space
are described by two complex amplitudes:
 
ψ1
|ψ = ψ0 |0 + ψ1 |1 ↔  = ∈ H = C2, (2.18)
ψ0

with the usual normalization |ψ0 |2 + |ψ1 |2 = 1.

Two-dimensional Hilbert spaces are very common. They are a natural represen-
tation of the spin s = 1/2 states of an electron, a proton, or a neutron: |0 and |1
correspond to spin down and up along a given direction; they are sometimes used
for describing the polarization states of a photon, horizontal versus vertical; and
they appear most frequently in quantum optics when modeling atomic transitions –
i.e., ground state |0 versus excited state |1 – and light–matter interaction. Nowa-
days, two-dimensional quantum systems are also called qubits, because, in analogy
to the classical bit, they represent the minimal quantum object where information
can be stored and processed. We show in Chapter 6 that it is possible to build
superconducting circuits that are accurately described as qubits, and discuss in later
chapters how these circuits are applied to quantum computing and simulation.
The algebra of two-level systems is analyzed using a complete set of observables,
called the Pauli matrices:
     
0 1 0 −i 1 0
σ =
x
, σ =
y
, σ =
z
. (2.19)
1 0 i 0 0 −1

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 14 Quantum Mechanics

Note that in this representation, |0 and |1 are the two eigenstates of σz , and we write
σz = |11| − |00|. If we enlarge the set of Pauli matrices to include the identity,
σ α |3α=0 = {1,σ x ,σ y ,σ z }, we have a basis where we can expand any observable in
this Hilbert space:

1
3
Ô = tr Ô σ̂ α σ̂ α . (2.20)
4 α=0

In particular, a general qubit Hamiltonian has the form

Ĥ = E + B n · σ̂, (2.21)

where n is a direction in the three-dimensional space, σ̂ = (σ̂ x , σ̂ y , σ̂ z ) and E and

B are constants. For spins, the Hamiltonian is interpreted as the coupling between
the qubit’s dipole moment ∝ σ̂ and the magnetic field Bn along the given direction.
Interestingly, this form has two important properties. First, because n · σ̂ has
the properties of a Pauli matrix, the eigenvalues of this Hamiltonian are simply
λ± = E ± B. Second, and for the same reasons, we can use the Pauli expansion to
compute the evolution operator of the qubit:

it Ĥ  
Û (t,0) = exp − = e−iEt/h̄ cos(Bt) 1 − i sin(Bt) n · σ̂ . (2.22)
h̄

We will use this formula when studying superconducting qubits and the implemen-
tation of single-qubit gates in Section 6.1.4.

2.3 Density Matrices

The formalism of unitary evolution for a quantum system assumes that the system
under study is perfectly isolated from other quantum or classical objects and subject
to error-free control. Even if systematic errors are greatly reduced, no physical sys-
tem can be perfectly isolated: At the very least, there will always be the omnipresent
electromagnetic field, carrying the cosmic background radiation and putting our
system in contact with the noisy classical world.
It is therefore safe to say that experiments never prepare pure states |ψ. Instead,
real quantum systems must be described using an ensemble operator, also known as
density matrix ρ̂: a nonnegative Hermitian operator ρ̂ = ρ̂ † ≥ 0, with proper nor-

malization tr(ρ̂) = 1(∼ n n|ρ̂|n), which gives expectation values of operators

as Ô = tr(Ô ρ̂) ∼ n ρnn n|O|n .
Pure states |ψχ  written as density matrices become projectors ρ̂χ = |ψχ ψχ |.
General states, however, are mixed states, because they can be reconstructed as

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 2.3 Density Matrices 15

classical ensembles of pure states created with different classical probabilities p(χ )
(Ballentine, 1970):

ρ̂ = |ψχ ψχ | p(χ )dχ. (2.23)

This equation (2.23) describes the output of an experiment where parameters have
some uncertainty. Mixed states can also arise when a system enters in contact with
another system, called the environment. In principle, we should consider the global
wavefunction of the system plus its environment, allowing both to be correlated

|global  = s,E s,E |s ⊗ |E ∈ Hsystem ⊗ Henvironment . However, since we will not
have access to all degrees of freedom of the environment, we must trace out all the
information that we ignore, obtaining a much smaller ensemble that only describes
our system:

ρ̂sys = trenvironment |global  global | = s,E s∗,E |ss  | . (2.24)
E s,s 

Density matrices are particularly simple in the case of two-level systems, where
they can be expanded in the qubit basis ρij := i|ρ̂|j , or as a combination of Pauli
operators (2.20):

   1 
1
ρ00 ρ01 1
ρ̂ = = ρij |ij | = 1 + Sα σ̂ α . (2.25)
ρ10 ρ11 2 2
i,j =0 α=x,y,z

The set of all physical states with S = (Sx ,Sy ,Sz ) falls inside the Bloch sphere
|S| ≤ 1. The surface of this sphere is formed by pure state (|S| = 1), and its center
is the completely depolarized state S = 0 or ρ̂ = 12 1.
The completely depolarized state is an example of classical state, diagonal den-
∗
sity matrices – ρ10 = ρ01 = 0 – which may be constructed as a convex combination
ρ̂ = ρ00 |00| + ρ11 |11| of preparing state |0 with probability P0 = ρ00 and
preparing state |1 with probability P1 = 1 − P0 . Compare this now with the super-
position state, |+ = √12 (|0 + |1). This state exists only in the quantum model for
a two-level system. This particular superposition maximizes the off-diagonal ele-
ments ρ01 and ρ10 , also known as coherences, recognized as a true signature of
quantumness in states.
Sometimes a quantum system is sufficiently isolated and the timescales of study
are short enough that we can approximate its evolution with a Hamiltonian that
involves just that system. In that case, the unitary evolution of vectors in the Hilbert
space dictates a recipe for updating the density matrix ρ̂(t) = Û (t,t0 )ρ̂0 Û (t,t0 )† .
More generally, a quantum system will get entangled with its environment, suffering

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 16 Quantum Mechanics

an incoherent evolution. Under certain physically reasonable assumptions (cf.

Appendix B), the equation that describes this dynamics is the Lindblad master
equation:
i
∂t ρ̂ = − [Ĥ, ρ̂] + Lt (ρ̂). (2.26)
h̄
The linear superoperator Lt (α ρ̂1 + β ρ̂2 ) = αLt (ρ̂1 ) + βLt (ρ̂2 ), called the Lindblad
superoperator, contains the information about the noise or the environment-induced
decoherence. There are no general prescriptions to write down or even solve this
kind of equation, but in a few cases the coupling is so weak and the environment so
big that it instantaneously loses all memory about the system’s dynamics. This is
the so-called Markovian limit, in which L is independent of time and of the system’s
initial conditions. This limit provides a very simple and very accurate description
of the loss of energy and of quantum coherence for many of the superconducting
circuits that we will study – see, for instance, Section 5.5.2 or 6.1.5.

2.4 Measurements
The axioms of quantum mechanics prescribe the behavior of a quantum system
under a complete measurement of any observable O, also known as Von Neu-
mann or projective measurements. Each observable is associated to a different
Hermitian operator Ô. The eigenvalues that result from diagonalizing this operator
on correspond to the possible measurement outcomes of the measurement. Let

P̂n = m |on,mon,m| be the projector onto all quantum states for which the
observable Ô has the value on . This projector is built using eigenstates of the
observable Ô |on,m = on |on,m, understanding that one measurement outcome
may be given by many different quantum states, which differ in other generic
quantum properties, here denoted as m. According to quantum mechanics, an ideal
projective measurement of the observable Ô onto a state ρ will produce the outcome
on with probability p(on ) = tr(P̂n ρ̂). If the measurement is also nondestructive, the
quantum state after the measurement will be projected onto a new density matrix:

1
ρ̂ → P̂n ρ̂ P̂n . (2.27)
p(on )

We can highlight other properties and types of measurements. First, we may

realize that the ideal measurement from (2.27) is an instance of a quantum non-
demolition (QND) measurement, one which, if repeated twice on the same system,
always produces the same expected value Ô. Ideal QND measurements are one
of the targets for quantum computing setups. In those experiments, we need to
determine the state of the superconducting qubit with certainty, through a projective

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 2.4 Measurements 17

measurement of the qubit’s polarization σ z, that leaves the qubit in a well-defined

state |0 or |1. This way, we can use the outcome of the measurement as input to
the following steps of quantum algorithms or to error-correction protocols.
Qubit measurements, such as used in those quantum computers, should ideally
be single shot. This means that every time we run the experiment, we obtain a real
value on associated to one measurement outcome, without errors. Note, however,
that even if we obtain a meaningful value every time we measure, the estimation
of Ô or the probabilities p(on ) can still be quantum limited, and we may need
to repeat the experiments many times to obtain such estimates with high accuracy.
For instance, the unbiased estimator of the average using M experimental measure-

ments Ōest = M1 M i=1 oni is itself a random variable with a standard deviation that
approaches the quantum uncertainty:
Ô
Ōest = √ . (2.28)
M
Computing a good estimate means bringing this deviation down to zero, which we
do by repeating the experiment again and again, until Ōest lays below our desired
tolerance.
In experiments, we rarely find direct projective measurements. More generally,
experiments are designed so that we measure an auxiliary quantum object that
has interacted with and extracted the information from the system we want to
measure. The reason to operate this way is to reduce decoherence. If we connect an
oscilloscope directly to a microwave resonator, the big classical object will quickly
deteriorate the quantum state of the photons that are inside the cavity. It is therefore
more convenient to create a setup such as the one in Figure 2.1, in which we perform
a weak connection between the resonator and a superconducting waveguide that
extracts only a tiny fraction of the photons, which are amplified and fed into a
detector.

Figure 2.1 A resonator or cavity stores microwave photons whose state we wish to
measure. Instead of connecting the measurement apparatus to the cavity, we make
a weak connection between the cavity and an outgoing superconducting cable that,
after passing through an amplifier, brings the signal to the measurement apparatus.
The signal bout is proportional to the cavity signal acavity , but contains vacuum
noise. The measured signal, Gbout +ε, is amplified with a gain G > 1, but contains
additional noise ε from the amplifier.

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 18 Quantum Mechanics

Figure 2.2 Quantum circuit for a generalized quantum measurement.

There are two consequences to using such a setup. First of all, we no longer have
a single-shot measurement. Amplification introduces noise in the signal, meaning
that our oscillator will retrieve a measurement on + ε that is affected by random
fluctuating photon noise ε. In absence of systematic errors, those errors average

out ε = 0, and we can still produce meaningful estimates Ō  M1 m om , but
the presence of large noise prevents us from determining the quantum state of our
system after the measurement – i.e., the measurement is no longer quantum limited.
The second consequence of indirect measurements is that we need a broader
framework to understand both the measurement statistics and the state of the quan-
tum system after a given measurement outcome. This framework is provided by
generalized quantum measurements, depicted in Figure 2.2.
(1) The system ψ is put in contact with the auxiliary quantum object φ.
(2) Both systems interact through some unitary evolution, Û .
(3) We measure the auxiliary object using a projective measurement Ô.
The generalized measurement or positive operator valued measurement (POVM)
associates measurement outcomes om to operators M̂n that are no longer projectors,
but still satisfy some completeness relation:

M̂n† M̂n = 1. (2.29)
n

The POVM operators determine the statistics of the measurement outcomes

p(om ) = M̂n† M̂n , and, if the measurement is nondestructive, also the post-
measurement state of the quantum system:
M̂n |ψψ| M̂n†
|ψ → . (2.30)
ψ|M̂n† M̂n |ψ

https://doi.org/10.1017/9781316779460.002 Published online by Cambridge University Press

 3
Superconductivity

A conductor is a material that can transport electric charge. A typical example of a

conductor is a metal, such as copper, iron, or gold. These materials are formed by
a somewhat regular lattice of positively charged ions and a sea of mobile electrons.
Subject to an external force, the valence electrons will accelerate toward the regions
of lowest potential energy, creating an electrical current. As opposing agent to this
acceleration, the electrons experience a friction due to the collisions with ions in
the lattice. A typical electron will move an average distance, called the mean free
electron path, before it hits an ion, exchanges kinetic energy with it, and randomly
changes both the velocity and direction of motion.
What I described in the last paragraph is the usual microscopic model for Ohmic
resistance. At the macroscopic level, it implies that, in order to sustain a current I
flowing through a conductor, we have to continously exert a force on the charge
carriers, establishing a potential energy difference or voltage V between two points
of the conductor. The larger the current we wish to establish, the faster the material
dissipates energy and the larger the voltage V required. The proportionality constant
between current I and voltage V is the device’s resistance R:
V = I × R. (3.1)
The resistance R is an extensive property that depends on the size of the material
through which charge carriers propagate, as well as the rate of collisions that we
mentioned before, the so-called resistivity ρ.
From our microscopic interpretation of the resitivity ρ, it is obvious that there will
be many contributions to this value. Part of the resistivity will be due to intrinsic
sources. These include lattice defects – dislocations, boundaries between grains in
the material, different orientations in the crystal – and impurities or atoms from
different materials that penetrate the conductor. In addition to these, we also have to
consider the influence of temperature. When the metal heats up, the ions move faster
and farther away from their equilibrium positions in the lattice. This increases both

19

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 20 Superconductivity

Table 3.1. Critical temperatures of various superconducting materials.

Element or compound Symbol SC type Tc Gap 2

Aluminum Al Type I 1.175 K 82.2 GHz
Lead Pb Type I 7.20 K 660 GHz
Mercury Hg Type I 4.15 K 399 GHz
Niobium Nb Type II 9.2 K 738.5 GHz
Yttrium barium copper oxide YBa2 Cux Oy High-Tc 70–90 K –

the probability of collisions with electrons as well as the amount of energy that they
can exchange. When we combine both effects, we typically find an experimental fit
of the form
ρ ∼ ρ0 [1 + α max{T − T0,0}], (3.2)
with a temperature-dependent contribution and an intrinsic plateau ρ0 that depends
on the material and even on the specific sample.
The study of the temperature-dependent resistance in metals experienced a break-
through with the discovery of liquid helium by Kamerlingh Onnes in 1908. One of
the first applications of liquified helium was to study the conductivity of metals
(mercury, tin, and lead) under temperatures ∼ 4.2 K where the thermal contribution
to the resistivity should be negligible. In 1911, Kamerlingh Onnes submerged a
wire of mercury in helium and observed that the resistivity of the metal suddenly
dropped to zero. In other words, the previous plateau disappeared, ρ0  0, even
though nothing special had been done to improve the material’s purity and lattice
perfection. After these initial discoveries, further materials were shown to enter
this new superconducting phase at sufficiently low temperatures, summarized in
Table 3.1.
Later studies have shown that a superconducting material cannot be just charac-
terized as a conductor without resistance. As we will soon see, superconductivity is
an intrinsically quantum effect that has other unexpected consequences:

– The first one is the possibility of establishing persistent currents. If we build

a superconducting ring and induce a current, for instance by passing a magnet
through the loop or using an external inductor, this current can persist forever
without any energy penalty on the material.
– The current that flows around a hole or in a loop can only take certain dis-
crete values that are compatible with the quantization of the magnetic flux
that traverses the loop. This fluxoid quantization happens in units of the flux
quantum 0 = h/2e, a universal magnitude governing the operation of SQUIDs
(Section 4.7) and superconducting qubits (Chapter 6).

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 Superconductivity 21

– Superconductors exhibit another surprising phenomenon known as the Meissner

effect, which is the ability of the superconductor to “expel” magnetic fields. If
we take a superconducting object and switch on a magnetic field, the mate-
rial develops superconducting currents on the surface such that they cancel the
magnetic fields inside the object. This phenomenon is similar to how charges
on the surface of a conductor arrange to cancel all electric fields inside the
bulk and lays at the heart of different levitation experiments with magnets and
superconductors.
– Finally, superconductivity also leads to counterintuitive behavior in super-
conductor–insulator–superconductor interfaces. The quantum nature of the
superconducting charge carriers allows them to tunnel through thin insulating
barriers, establishing currents that behave in unexpected ways in presence of dc
and ac potentials. This Josephson effect makes it possible to develop Josephson
junctions, a nonlinear inductor that makes superconducting circuits useful for
quantum technology applications.
This phenomenology, which is universal across the board of all superconduct-
ing materials, can be explained with a mesoscopic theory that combines quantum
mechanics and electromagnetism. However, not all superconducting materials are
identical: They can be distinguished by other properties, such as the conditions
under which superconductivity appears or is destroyed.
– Type I superconductors is the denomination for the first family of supercon-
ducting materials that were discovered. This includes most elementary metals,
such as mercury, tin, or aluminum. Superconductivity manifests at low critical
temperatures, between 1–4 K, and is destroyed at relatively low critical mag-
netic fields. Above such fields, the Meissner effect abruptly disappears and the
material experiences a first-order phase transition into an ordinary metal.
– Type II superconductors includes niobium and a plethora of alloys with larger
critical temperatures and superconducting gaps. These materials survive stronger
magnetic fields through the creation of magnetic vortices: thin “tubes” of ordi-
nary metal that allow the field to cross the material, while the rest of the electrons
remain in a superconducting state.
– High-Tc superconductors is yet another family of superconducting materials,
discovered late in the twentieth century. These superconductors are rare-earth
ceramic alloys with a complex, quasi-two-dimensional structure that allows
superconductivity at temperatures above nitrogen’s boiling point (70 K). Since
we must work at lower temperatures to engineer microwave quantum circuits
(see Section 4.1.1), and since these ceramic materials are difficult to fabri-
cate, they are not very interesting for superconducting quantum information
technology.

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 22 Superconductivity

After this brief overview of superconducting materials, we will introduce both a

microscopic interpretation of superconductivity, as well as a macroscopic theory –
the London theory or macroscopic wavefunction model – that explains much of the
superconducting phenomenology. A distilled version of this model will be the basis
in Chapter 4 to develop an effective theory of superconducting circuits.

3.1 Microscopic Model

Which mechanism allows some metals to become superconductors? This is a
Nobel Prize–winning question that was answered by physicists John Bardeen,
Leon Cooper, and John R. Schrieffer in 1957 (Bardeen et al., 1957a,b). Together,
they explained many of the superconducting phenomena mentioned previously,
as well as other many-body properties of the superconducting materials, which
include the following:
– The evidence of a phase transition and of some type of energy gap to break
superconductivity, as suggested by the existence of a critical temperature Tc and
a critical magnetic field above which superconductivity disappears.
– The exponential decrease of the superconductor’s heat capacity with temperature
C ∝ exp(−1.5Tc /T ). This property was consistent with a many-body theory in
which there is an energy gap, a minimum excitation energy per particle of order
1.5kB Tc .
– Further evidence of some excitation gap, as provided by the electromagnetic
absorption spectrum: the minimum photon energy required to locally excite or
break the superconducting state lays somewhere around 2 × 1.5 × kB Tc .
– Finally, the magnetostatic properties of the superconductor, including supercon-
ducting currents and the Meissner effect could be explained by introducing a
coherence length – i.e., quantum correlations at short distances. As it is now well
known in condensed matter physics, the existence of finite coherence lengths is
usually an indicator of a gapped model.
The answer by Bardeen and collaborators is known as the BCS theory. This theory
proposes that the superconductor is actually a Bose–Einstein condensate of charge
carriers. As it was already known from studies of Bose–Einstein condensation and
early models of 4 He superfluidity, a weakly interacting Bose–Einstein condensate
can support superfluid currents that never stop and that are immune to small imper-
fections, impurities, and collisions that do not carry too much energy. In order to
justify the existence of Bose–Einstein condensation, the BCS theory introduces
an effective attraction between the metal’s valence electrons with opposite spin.
This attraction is mediated by the phonons of the crystalline structure that forms
the metal. At low temperatures, it gives rise to the BCS instability, in which the

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.1 Microscopic Model 23

Fermi theory breaks down, and electrons join into stable bound particles called
Cooper pairs (Cooper, 1956). Because of the spin-statistics connection, the pair
of two bound electrons is a particle with a bosonic statistics, which condense into a
superfluid state.
The BCS theory provides a very elegant and also very straightforward frame-
work for studying the superconductor, with only a few parameters that account for
all the physics. The main parameter is the BCS or superconducting gap, usually
denoted by (k). The gap is the binding energy of every two electrons that form
a pair. It depends mildly on the electron’s momenta, and the smallest value (0)
explains many quantiative properties of the superconducting phase and the phase
transtion. For instance, as we increase the temperature, we can expect that processes
in which pairs are broken become relevant, destroying superconductivity. The quan-
titative answer is a bit more complicated, as the gap itself depends on temperature
√
(T )  1.74(0) 1 − T /Tc , and the critical temperature Tc  (0)/1.76kB is the
point at which pairing becomes energetically trivial.
The superconducting gap also explains some features in the interaction of the
superconductor with electromagnetic fields. A superconductor can absorb photons
through two different mechanisms. Low-energy microwaves in the range of 1–20
GHz can excite plasmons of the charged superfluid. These processes create quantum
excitations that behave very much like photons (see Chapter 5) or like artificial
atoms (see Chapter 6), and which we can route, confine, and operate using super-
conducting circuits. The second mechanism involves stealing a Cooper pair from
the condensate and breaking it into two separate electrons. The energy required
for this is h̄ω  2(0)  3.52kB T . For the case of the widely used material in
superconducting circuits, aluminum, this energy lays in the range of 100 GHz.
Therefore, we can suppress this type of event by sufficiently cooling our circuits
and isolating them from the environment, with filters that prevent the injection of
highly energetic photons.
The BCS theory has other important consequences, including studies of heat
capacity, impurities, quasiparticle excitations, Andreev states and normal super-
conductor interfaces. Overall, this theory applies very well to type I supercon-
ductors, and to some extent to type II, but it does not explain high-temperature
superconductivity.
Fortunately, we are not so much interested in complex superconducting mate-
rials or sophisticated excitations. Rather, we would like an effective model of
the superfluid condensate in the simple materials, Al or Nb, which are used
in the quantum circuit experiments. As explained by Gor’kov (1959), the BCS
model of superconductivity predicts an effective nonlinear theory for the con-
densate order parameter. This is the Ginzburg–Landau model or, in the simpli-
fied linear version that we introduce here, the macroscopic wavefunction model

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 24 Superconductivity

(Orlando, 1991). Without many complications, this intuitive model will provide a
solid and approachable foundation to the engineering of quantum circuit in later
chapters.

3.2 Macroscopic Quantum Model

The BCS model for superconductivity proposes that electrons group into a larger
unit, the Cooper pair, that has the properties of both being a charged particle, with
charge q = −2e, as well as being a boson. The bosonic nature of the particle is what
makes it possible for all Cooper pairs to condense, sharing a common superfluid
state that is insensitive to defects in the material and any other drag force. The
theory that we are about to explain is very similar to other models that have been
put forward and successfully used, for instance, in the study of weakly interact-
ing Bose–Einstein condensates of alkali atoms (Pitaevskii and Stringari, 2016),
and even BCS superfluids built from fermionic atoms. Our formulation of the the-
ory follows closely Orlando (1991), a book we encourage you to read for a better
understanding of superconducting properties, magnetostatics, and other interesting
phenomenology.
The macroscopic wavefunction theory is based on the assumption that the many-
body state is described by a collective wavefunction that is a product state of the
same wavefunction for each of the N Cooper pairs:

(x1,x2 . . . xN ;t) = ξ (x1,t)ξ (x2,t) · · · ξ (xN ,t). (3.3)

This type of macroscopic accumulation of particles into the same quantum state is
what we expect from a Bose–Einstein condensate well below its critical temper-
ature. However, we are also allowing for this accumulation, which is typically a
property of a ground state, to also describe the dynamics of the collective system
in time, as it reacts to external perturbations from electromagnetic fields, currents,
etc. This is a conceptual extension that is only justified by the agreement with
experiments and the exact simulations of small systems.
The macroscopic wavefunction theory leads us to introduce new fields ns and θ ,
which respectively describe the charge density and, as we will soon see, the flow of
particles:
√ 
ψ(x,t) = N ξ (x,t)  ns (x,t)eiθ(x,t) . (3.4)
We typically assume a constant and approximately uniform density of carriers
throughout most of the material, ns (x,t)  n̄s . This assumption is approximately
valid in many situations because matter tends toward charge neutrality. It will not
apply when we consider the charge trapped on a capacitor or in a superconducting

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.3 Superfluid Current 25

island. Those deviations will be studied as perturbative corrections to the back-

ground of superconducting particles in the macroscopic theory.
Assuming that the macroscopic wavefunction is a viable model, we now postulate
a very general model for its dynamics:

1
i h̄∂t ψ = (−i h̄∇ − qs A)2 + qs v(x,t) ψ. (3.5)
2ms

This model is inspired by the Schrödinger equation for a charged particle moving
in an electromagnetic field with scalar and vector potentials v(x,t) and A(x,t),
respectively. The model introduces two effective parameters ms and qs to describe
the mass and charge of the Cooper pair. As we have seen, the charge is precisely
known:
qs = −2e = −2 × 1.60217662 × 10−19 C. (3.6)
However, ms = 2m∗e contains the effective mass of the electrons moving through
the solid lattice m∗e , which depends on the band structure and has to be determined
experimentally for each material.

3.3 Superfluid Current

From the previous equation, we can already obtain two important properties that
we need for studying real superconducting circuits. The first property is the charge
distribution, which is given by

ρ(x,t) = qs |ψ(x,t)|2 . (3.7)

This superfluid charge includes a very large background that compensates the
charge of the ions structuring the lattice of the metal or alloy. From the point of
view of circuit theory, it is more interesting to work with the superfluid current,
a vector field J(x,t) describing the flow of charges. The evolution of the electric
charge Q confined in a volume , is related to the supercurrent flowing across the
boundary of that volume ∂:
 
d
Q= ∂t ρd x = −
3
J · dn. (3.8)
dt  ∂

Here n is the unit vector normal to the surface ∂ at each point of the boundary.
The same physics is described by the continuity equation:

∂t ρ = −∇ · J. (3.9)

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 26 Superconductivity

As Fritz London conjectured, the superfluid current may be derived from the
Schrödinger equation (3.5) as a combination of the macroscopic wavefunction cur-
rent and the electromagnetic field:
   
∗ h̄ qs
J = qs × Re ψ −i ∇ − A ψ . (3.10)
ms ms
This combination explain the fluxoid quantization and the Meissner effect, as
described by London et al. (1935).
Note that we can obtain the electric current intensity of a circuit I by integrating
the charge current J(x,t) across any section S of any cable in the circuit:

I (t) = 
J(x,t) · dS. (3.11)
S

The current intensity can be different at different points of a large superconducting

circuit. However, the conservation of charge – Cooper pairs are not destroyed in
our simple, conservative model – implies that the current coming into a supercon-
ducting element must balance with the current going out into other circuit elements.
This will be key in our analysis of circuits and derivation of quantitative models in
Chapter 4.

3.4 Superconducting Phase

The superconducting wavefunction contains information about the charge distribu-
tion and the electrical current. We will now argue that most of the information is
actually hidden in the phase of the wavefunction. We will also relate this phase to a
macroscopically observable quantity, the flux.
The first statement is rather obvious. We have already discussed that the density
of charged particles must be a rather uniform property, dependent only on the prop-
erties of the material – i.e., how many electrons the atoms donate to the conduction
band where Cooper pairs are formed. If we assume that ns (x,t) = |ψ|2 is constant
and uniform, currents are divergence-free:
∇ · J = 0, (3.12)
and all information about the superconductor must actually reside in the phase of
the wavefunction. Indeed, working with (3.10), we obtain
h̄ qs
J = qs ns ∇θ − A . (3.13)
ms ms
The second statement is more subtle. Let us assume the Coulomb gauge ∇ · A = 0,
and specialize the macroscopic quantum model (3.5) for a wavefunction with
uniform density. The Coulomb gauge implies θ = 0, and

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.5 Gauge-Invariant Phase 27

1
− h̄∂t θ  J2 + qs v. (3.14)
2ns
Notice the new constant, the isotropic London coefficient1  = ms /qs2 ns .
In the absence of currents, J = 0, (3.14) becomes the so-called phase-voltage
relation
qs
∂t θ  − v. (3.15)
h̄
This relation is an obvious consequence of unitary evolution. For quasi-stationary
states, the wavefunction remains constant up to a global phase, determined by
the energy of the system ψ(x,t) = exp(−iEt/h̄)ψ(x,0). Since the energy of the
charged particle in a potential is E = qs v, we obtain ∂t θ = −E/h̄ = −qs v/h̄. The
problem with relation (3.15) is that it is not gauge invariant – it is only valid in the
Coulomb gauge – and it has been derived under the condition of no persistent
currents, J = 0. We have to complete our derivation to regard more general
conditions!

3.5 Gauge-Invariant Phase

In order to correct (3.15), we will separate the superconducting phase into a term that
is always the same, and a contribution that depends on our choice of electromagnetic
gauge. The gauge-invariant phase ϕ(x,t) is defined by removing the contribution
of the vector potential, taking as reference one (arbitrary) location of the supercon-
ductor x0 :

qs x
θ(x,t) − θ (x0,t) = ϕ(x,t) − ϕ(x0,t) + A(r,t) · dl. (3.16)
h̄ x0
In the definition of ϕ(x,t), the choice of path from x0 to x is arbitrary, but (i) it must
be unique for each point x, (ii) it must be continuous, (iii) all paths must remain
in the superconductor, and (iv) they must not cross each other.2 Except for a set of
points of zero measure – the discontinuities of ϕ – we can define a gauge-invariant
wavefunction:
−i qs
x
A(r,t)·dl

ψGI (x,t) = e h̄ x0 ψ(x,t) = eiϕ(x,t) ns (x,t), (3.17)

1 As mentioned before, superconducting currents shield magnetic fields out of the material. The London
coefficient is related to the penetration depth of magnetic fields into the superconductor, a fact that can also be
derived from this theory (Orlando, 1991).
2 This choice of path is one of the essential steps in working with any superconducting circuit, as we will see in
Chapter 4. However, in that chapter we will assume quasi-one-dimensional structures, where it is easy to
believe that such paths do exist, at least in the form xl = xl + x0 (1 − l) for l ∈ [0,1].

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 28 Superconductivity

which satisfies a Schrödinger equation without vector potential:

1
i h̄∂t ψGI = (−i h̄∇)2 + qs v(x,t) ψGI . (3.18)
2ms
We now reach a gauge-independent relation between the phase and the
electric field:
  
∂ qs x ∂
ϕ(x,t) = −∇v − A · dl (3.19)
∂t h̄ x0 ∂t

qs x qs
= E(r,t) · dl = [V (x0 ) − V (x)] .
h̄ x0 h̄
This gauge-invariant phase is related to the voltage difference V across the super-
conducting circuit,3 defined as the energy required to transport a unit of charge from
x0 to x. It is convenient to refer the voltage to V (x0 ) := 0 and work instead with the
electric flux, defined up to an irrelevant offset as
 t
φ(x,t) = V (x,τ )dτ . (3.20)
0

Introducing also the magnetic flux quantum,

h h
0 = =  2.067833758(46) × 10−15 Wb, (3.21)
|qs | 2e
we identify the gauge-invariant phase with the electric flux
2π
∂t ϕ(x,t) = ∂t φ(x,t), (3.22)
0
As we will see in Chapter 4, this identity makes the electric flux one of the two
preferred variables when working with superconducting circuits, the other one being
the charge q that accumulates on a superconducting element. Another important
property of the gauge-invariant phase is that it determines the superconducting
current:
 
h̄ qs qs h̄qs ns
J = qs n ∇ϕ + A − A = ∇ϕ. (3.23)
ms h̄ ms ms
This makes sense, because the superconducting current, just like the circuit voltage,
cannot depend on the choice of gauge.

3 V and v are not the same observable. Only the former is gauge independent.

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.6 Fluxoid Quantization 29

3.6 Fluxoid Quantization

There is a third consequence of the phase-current relation. This one relates to the
allowed values of the current and of the magnetic flux trapped inside superconduc-
tors. Let us multiply (3.13) by  = ms /ns qs2 and integrate it around a closed-loop
C around a simply connected region, C = ∂S:
  
h̄
(J) · dl + A · dl − ∇θ · dl = 0. (3.24)
C C C qs

We use Stokes’s theorem to transform the second line integral into a surface integral
over the region S. The contour integral of the phase must produce a multiple of
2π × h̄/qs = −0 , as otherwise the wavefunction ψ would be discontinuous. This
results in the quantization equation
 
(J) · dl + B · dS + 0 × m = 0, with m ∈ Z. (3.25)
C S

This equation states that the magnetic flux trapped in a loop


loop = B · dS, (3.26)
S

plus the flux due to the induced supercurrents must be an integer multiple of the
magnetic flux quantum (3.21). This is a very powerful result, used by Deaver and
Fairbank (1961) to estimate 0 and demonstrate that the superconducting particles
qs = −2e are formed by two electrons.
There is a simpler derivation of fluxoid quantization that is of more interest to
us, and which is based on the gauge-invariant phase (3.16). This definition can be
reformulated as
qs
∇θ = ∇ϕ + A. (3.27)
h̄
Integrating around a closed loop, we find

2π
∇ϕ · dl = 2π × m + loop . (3.28)
C 0
Using the phase-voltage relation, this can be rewritten as a condition for the
flux differences along different segments of a superconducting circuit, or fluxoid
quantization:

∇φ · dl = 0 × m + loop . (3.29)
C

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 30 Superconductivity

This condition is one of the constituent relations in the theory of superconducting

circuits from Chapter 4. Note that unlike the wavefunction, the flux and the gauge-
invariant phase need not be continuous, as their definition depends on a choice of
paths that we make before studying the circuit. However, our derivation reveals that
all choices are consistent, and that the existence of loops in a circuit reduces the
number of independent variables, because fluxes are not completely independent
from each other.

3.7 Josephson Junctions

The last topic that we cover in this chapter is a simple, yet very powerful device that
was postulated by Josephson (1962) and verified experimentally shortly thereafter
by Anderson and Rowell (1963). The device in question is called a tunnel Josephson
junction.4 It is a superconductor–insulator–superconductor “sandwich,” where the
insulating area is so thin that it allows quantum tunneling of Cooper pairs. As
Josephson predicted, the tunneling of pairs creates unexpected relations between
the applied voltage (dc or ac) on the superconducting leads and the intensity that
circulates through the junction. We can use the macroscopic quantum model to
derive those relations.
Figure 3.1b shows the schematics of a Josephson junction, with three separate
regions: two superconducting leads of arbitrary size and an insulating barrier. Our
toy model for the junction is a one-dimensional5 potential barrier of height U0
and width d that hinders the propagation of the macroscopic wavefunction ψ(x,t).

Figure 3.1 (a) Circuit schematics for a Josephson junction connected to an inten-
sity source. (b) The Josephson junction is made of two superconducting leads
(white) separated by a thin insulating barrier (gray). (c) We model the junction as a
barrier energy U0 , which is thin enough, d, that it allows some quantum tunneling
of Cooper pairs.

4 We can obtain a similar physics through other physical devices, such as constrictions, point contacts, and
normal interfaces. However, the theoretical picture is simpler in this case.
5 The one-dimensional model is sufficient for describing the type of small junctions that appears in typical
superconducting circuits. A more detailed model that takes into account the transverse dimensions is found in
Orlando (1991).

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.7 Josephson Junctions 31

We seek stationary solutions of the Schrödinger equation for the gauge-invariant

wavefunction
1
EψGI (x) = (−i h̄∂x )2 + v0 (x) ψGI (x), (3.30)
2ms
with the potential barrier

0, |x| > d/2
v0 (x) = (3.31)
U0, otherwise.
As a boundary condition, we impose a current intensity I flowing through the junc-
tion. Thanks to the relation between current intensity and superconducting current,
our boundary condition can be written as

I = J · dS = J × A, (3.32)
S

where the proportionality constant A is the area perpendicular to the junction – i.e.,
the cross-section size.
What do the solutions of the Schrödinger equation look like with these condi-
tions? Since far away from the insulator the current is fixed, we expect plane-wave
solutions such that ∂x ϕ ∝ J . More precisely, we write
√
ψGI (x) ∝ neikx , |x| > d/2. (3.33)

The sign and magnitude of the superconduction current may be derived from
the momentum k of the wavefunction as J = qs ns h̄k/ms . The momentum also k
determines the energy of the solution
h̄2 k 2 1 2 2
E= = J . (3.34)
2ms 2n
When the current is very small J  0, this energy will lay well below the barrier
U0 . Inside the insulator, we will have an equation of the form
ms 1
∂x2 ψGI = 2
(U0 − E)ψGI = 2 ψGI, |x| ≤ d/2, (3.35)
h̄ ξ
which can only be satisfied with exponentially decreasing or increasing solutions,
ψGI ∝ exp(±x/ξ ). The final solution then reads

ψGI (x) = α+ cosh(x/ξ ) + α− sinh(x/ξ ), |x| ≤ d/2, (3.36)

with the parameters

√ eiϕL (−d/2) ± eiϕR (d/2)
α± = ns . (3.37)
ed/2ξ ± e−d/2ξ

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 32 Superconductivity

The superfluid current is uniform across the circuit. Along the insulator
qs ∗
J = Re(−ψGI i h̄∂x ψGI ) = Jc sin[ϕR (d/2) − ϕL (−d/2)], (3.38)
ms
it is proportional to the Josephson junction’s critical current
qs h̄ ns
Jc = − ≥ 0. (3.39)
ms ξ sinh(d/ξ )
Currents above this value cannot be captured by the exponential solution, but instead
correspond to plane waves with E > U0 .
The previous derivation has shown us that small values of the current have a
nonlinear relation with the gauge-invariant phase jump δϕ = ϕR − ϕL across the
junction:
I = Ic sin(δϕ). (3.40)
As clearly explained by Orlando (1991), the first Josephson relation still holds in
presence of magnetic and electric fields, provided we still work with the gauge-
invariant phase.
Assume now that we establish a voltage difference V among the junction’s
leads. Since we work with metals, the voltage will be approximately uniform
along each lead. We then expect  t a flux difference on both sides of the insulator
δφ(t) = φ(d/2) − φ(−d/2)  0 V (τ )dτ . Using the connection between flux and
phase (3.22), we obtain the second Josephson relation:
2π dV
δϕ = . (3.41)
0 dt
The Josephson relations combine into an equation connecting flux and current:
 
2π
I = Ic sin δφ . (3.42)
0
The dc Josephson effect is a consequence of this equation: a constant voltage bias V
induces an oscillating current due to the linearly growing flux:
 
2π
I (t) = Ic sin V t + δϕ(0) . (3.43)
0
However, (3.42) has more general implications, as a constituent equation, allow-
ing us to include Josephson junctions in the general theory of superconducting
circuits from Chapter 4. For instance, we can derive the inductive energy stored
in the junction as we switch on the voltage – and the flux – to a finite value:
 t
E= I (t)V (t)dt. (3.44)
0

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 3.7 Josephson Junctions 33

With the change of variables V (t)dt = dφ, we can substitute the expression (3.42)
and integrate between the initial and final value of the flux
 δφ    
2π Ic 0 2π
E= Ic sin φ dφ = − cos δφ . (3.45)
0 0 2π 0
From the point of view of superconducting circuit theory, the junction behaves as
a nonlinear inductive element. The inductance may be derived from the current–
voltage relation or from an expansion of the inductive energy just derived. In both
cases, we obtain a similar expression:
V 0 1
LJ = dI
= . (3.46)
dt
2π Ic cos(2πδφ/0 )
This nonlinear inductance and the constituent equation (3.42) will be repeat-
edly used in Chapter 4 when studying superconducting qubits, dc-SQUIDs, and
rf-SQUIDs.

https://doi.org/10.1017/9781316779460.003 Published online by Cambridge University Press

 4
Quantum Circuit Theory

4.1 Introduction
This chapter introduces a quantum mechanical theory of electrical circuits, devel-
oped in the spirit of the original work by Yurke and Denker (1984), and building
on the more pedagogical explanations by Devoret (1995). The chapter begins
discussing the temperature and power requirements needed to observe quantum
phenomena in electrical circuits. Section 4.2 recalls the three lumped circuit
elements – the capacitor, the inductor, and the Josephson junction that we use to
build most superconducting circuits, describing the classical equations that govern
them. Section 4.3 introduces an effective quantum theory of circuits, formulated
as a set of rules to transform a circuit network into the quantum Hamiltonian that
describes its dynamics. Sections 4.4–4.8 apply this theory to the most important
circuits in this book – a simple LC resonator, a transmission waveguide, a nonlinear
resonator, and different types of Josephson junction loops. You are strongly advised
to work out these examples and exercises on your own, to build a solid foundation
for later chapters where we use these circuits to engineer photons (Chapter 5),
superconducting qubits (Chapter 6), and light–matter interfaces (Chapter 7).

4.1.1 What Makes a Circuit Quantum

A circuit is “quantum” when we need a quantum mechanical theory to describe it.
Quantum circuits exhibit phenomena that are exclusive of quantum physical sys-
tems, such as energy quantization, reversible unitary evolution, quantum superposi-
tions, and entanglement. In a quantum circuit, electric current, voltage, charge, and
flux are quantum observables. Measurements of these quantities result in probabilis-
tic outcomes, governed by uncertainty relations that limit how much information we
can obtain in a single experiment or multiple experiments. Lastly, quantum circuits
are modeled with quantum mechanical equations, such as the Schrödinger equations

34

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.1 Introduction 35

for pure states, or the master equations for mixed states. All this affects the way we
design, control, and operate superconducting quantum circuits.
Like all solid-state devices, electrical circuits are built from myriads of quantum
systems – a sea of electrons floating through a lattice of ions, covalently bound
molecules, oxidized metal, etc. A circuit will exhibit quantum phenomena depend-
ing both on how it is built and how it is operated. Common wisdom states that there
are three causes for a system to act in a purely classical way, hiding its quantum
nature:
(1) When we combine very large number of quantum objects into a many-body
system, the properties of the macroscopic object tend to be well defined, with
negligible quantum uncertainties. These objects can be safely described with
classical equations.
(2) Phenomena such as superpositions and entanglement are destroyed by inter-
action with the environment – e.g., by surrounding electromagnetic fields and
particles. The typical decoherence time in which this happens tends to be shorter
than the duration of common experiments.
(3) Preparing good quantum states requires cooling the system to a point in which
thermal fluctuations do not obscure quantum fluctuations. The temperature at
which this happens varies with the physical model.
Let us discuss these challenges in greater detail:

Macroscopicity. To understand the first obstacle, consider the Heisenberg uncer-

tainty principle for two complementary observables such as position and momentum
xp ≥ h̄/2. Learning the position of a particle with a very low uncertainty
x leads to a large uncertainty in the momentum p. However, since the Planck
constant is very small, h̄ ∼ 1.055 × 10−34 J · s, the actual uncertainty in the velocity
of a macroscopic object can be very small. Realistically, x ∼ 800 nm is limited
by the wavelength of the light we use to observe objects; assuming a small particle
of 1 gram, then v = p/m ∼ h̄/mx ∼ 10−25 m/s, which is unobservable for
us humans.
A solution to this problem is to work with selected degrees of freedom that
can be brought down to a “microscopic” limit. In this book, we study electrical
circuits; these are macroscopic objects – with sizes that range between millimeters
and meters – but within these objects we manipulate the charge or current degrees of
freedom, manifesting their quantum properties. We may create quantum states such
as the superposition √12 (|N+|N + 1), in which a small superconducting island has
either N or N + 1 Cooper pairs – a microscopic difference on top of a macroscopic,
classical background. Similarly, we can prepare superpositions of left- and right-
moving currents, involving a small number of Cooper pairs that rotate along a

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 36 Quantum Circuit Theory

superconducting loop. Probing these degrees of freedom and small differences is

difficult, but we have the technology to do it.

Thermal fluctuations. Once we have selected a degree of freedom out of

a macroscopic object, the next obstacle is to prepare it in a state that reveals
quantum phenomena. If we do nothing special to a quantum object, its quantum
superpositions or entangled states will progressively deteriorate, until the object
achieves a thermal state in equilibrium with its surrounding environment.1 Thermal-
ization brings our quantum system progressively closer to a diagonal mixed state

ρ = n pn |nn| with the classical probability distribution {pn } of configurations
(see Exercise 4.1). As the quantum system approaches a termal state and thermal
fluctuations begin to dominate the experiment, they completely mask the quantum
uncertainties and correlations that reveal the quantum nature of our system.
We quantify the effect of thermal fluctuations using statistical physics. Let us
consider a quantum system with eigenstates |n that have regularly spaced eigenen-
ergies En  h̄ωn. This is the case of superconducting qubits – two-level systems
where n can be 0 or 1; see Chapter 6 – and of microwave photons – where n can be
any nonnegative integer n ∈ {0,1,2,3, . . .}; see Chapter 5. In a typical experiment
we want to first cool the system as close to the ground state |0 as possible. This
is a reference state from which other states and superpositions can be created in a
predictable way. If we let our system thermalize, it will have some overlap with the
ground state, but it will also spread over other slightly more energetic configurations
|1 , |2 . . . . A perfectly thermalized state will follow the Boltzmann distribution,
pn ∝ exp(−h̄ωn/kb T ). As shown in Figure 4.1, the probability of being in any
excited state pexcit = 1 − p0 = exp(−n̄) decreases exponentially with the ratio
n̄ = h̄ω/kB T both for photons and two-level systems.
Thermal fluctuations are strongly suppressed once the temperature drops well
below a value h̄ω/kB dictated by the characteristic energy scale of excitations.
Table 4.1 shows relevant energy scales and temperatures for various experi-
mental platforms. For a typical circuit that works at microwave frequencies
ω = 2π × 8 GHz, the temperature reference2 is T = h̄ω/kB  0.38K. For tem-
peratures well below this limit, of 400,100,50, and 10 mK, we predict an excitation
probability of 38%, 2%, 0.04%, and 210−15 %. Thus, placing the electrical circuit in
a good dilution refrigerator that operates in the 20–50 mK guarantees that thermal
fluctuations will be negligible.

1 There is always a reservoir in contact with our system, even if that reservoir is just the electromagnetic field at
the temperature of the cosmic background radiation!
2 A good rule of thumb to remember this calculation is the equivalence between 20 GHz and 1 K, which makes
the excitation probability e−1 ∼ 36%. Every time we halve the temperature, we divide this probability by ∼3.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.1 Introduction 37

Table 4.1. Equivalence between the frequencies

needed to excite a quantum system and the
temperature at which thermal excitations become
relevant.

1015 Hz ∼ 50 000 K Ultraviolet

300 THz ∼ 15 000 K Near infrarred
1 THz ∼ 50 K High-energy microwaves
1−20 GHz ∼ 50 mK−1 K Microwaves

100
50

10
5

1
0.5

0 1 2 3 4 5 6

Figure 4.1 Thermal fluctuations in a quantum device. We study a device with either
two possible configurations separated by an energy h̄ω (dashed), or a ladder of
states with energies {0, h̄ω,2h̄ω, . . .} (solid). Introducing the ratio n̄ = h̄ω/kB T ,
we plot the probability that the system is found in an excited state. The solid line
follows the law exp(−n̄).

Environmental influence. Even if we ensure a low-temperature surrounding, our

circuit will still couple to the environment: a mixture of surrounding electromag-
netic fields, control cables, antennaes, materials and impurities of the substrate, etc.
The sample will interact with this environment, radiate energy, absorbe fluctuations,
and experience jitter in its eigenenergies. We identify these three generic processes
with the terms dissipation, heating, and dephasing. They all lead to a generic phe-
nomenon known as decoherence, which is the destruction of quantum coherence or
quantum fluctuations in our system, steering it toward a classical state.
There are trade-offs to be made when dealing with decoherence. The materials
forming our circuits carry uncontrollable sources of errors that limit our total coher-
ence time. For instance, in ordinary metals, impurities, defects, phonons, etc., cause
some electric resistance that cannot be neglected, even at low temperatures. This
is one intrinsic source of error that can be eliminated by working with supercon-
ductors, but many others persist: trapped charges in defects of a material, two-level

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 38 Quantum Circuit Theory

fluctuators affecting the inductance of Josephson junctions, radiation to free space,

dirt and fabrication defects, etc. On top of this, we must consider the coupling of
our quantum systems to measurement devices and controls: this coupling is also a
source of decoherence, which we can diminish but never completely suppress.
Altogether, environment and imperfections both limit the time during which we
can operate a circuit coherently. In times shorter than this value, we may prepare a
quantum state, perform quantum gates, create entanglement, and measure quantum
properties with enough accuracy. Past this time, evolution will deviate from a pure
dynamics, quantum superpositions will deteriorate, and the whole experiment –
unless corrected in some way, as discussed in Section 8.5 – will behave as a noisy
classical system.
So, summing up, it is theoretically and experimentally possible to develop elec-
trical circuits that work in a quantum regime. These circuits will operate at very low
temperatures, at which some materials already become superconductors. The oper-
ation of those circuits can be done during moderately long times without significant
errors, preserving quantum coherence throughout the process. Let us now discuss
what are the theoretical models that describe the coherent part of evolution.

4.1.2 How Do We Work with Quantum Circuits?

An electrical circuit is a network of metallic, insulating, and semiconducting ele-
ments, such as capacitors, inductors, resistors, and transitors, which, combined with
appropriate power soures, control the behavior of current I and voltage V . Classical
network theory is a systematic framework to analyze complex electrical circuits,
finding out the independent variables that describe the circuit – e.g., current or
voltage on select nodes of the network – and producing a set of ordinary differ-
ential equations that describes the dynamics of those variables. If the circuit is
conservative – it has no resistance in any element – the theory can be brought to
a Hamiltonian formulation, quantifying the total energy of the system as a function
of those variables. As explained in Section 2.1.1, both descriptions contain the same
information: We obtain the Hamiltonian from the differential equations, and we can
recover the dynamical equations from the Hamiltonian through differentiation.
Let us illustrate this with the help of an LC resonator, a model circuit that will
be used frequently throughout this book. A minimal LC resonator, sketched later in
Section 4.3, consists of a capacitor C connected in parallel to an inductor L. The pair
of variables (I,V ) define a phase space similar to that of position and momentum
(x,p). They satisfy the ordinary differential equations:
dV I dI V
= , and =− . (4.1)
dt C dt L

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.1 Introduction 39

Alternatively, we can write a second-order differential equation:

d2 V
= −ω2 V . (4.2)
dt 2
From your classical mechanics course, you may remember this as the equation for
a harmonic oscillator. It describes the periodic
√ oscillations of the voltage (and of
the current), with a frequency ω = 1/ LC that is defined by the capacitance C
and the inductance L of the circuit. The total energy of the circuit is conserved (see
Exercise 4.4):
1 1
H = CV 2 + LI 2 . (4.3)
2 2
It is the sum of the capactive and inductive energy stored in the circuit, and also as
the Hamiltonian associated to (4.1), where V and I play the role of position and
momentum x and p.
Quantum circuits are really not that different. Because we are interested in sup-
pressing thermal fluctuations and decoherence, these circuits are built from super-
conductors – alluminum and niobium, typically – and operated at temperatures of
tens of milli-Kelvin. We design these circuits to only include capacitors, inductors,
and Josephson junctions. There are no transistors, and the resistance of the junctions
is usually neglected.
As before, the circuit elements induce relations between observables, which
result in two possible formulations of a quantum network theory for a circuit, both
of which rely on the notions from Section 2.1.3. We can construct a set of ordinary
differential equations. These are Heisenberg equations for the quantum observables
of current and voltage, which become operators with associated expectation values,
uncertainties, and correlations that can be probed in experiments. Alternatively, we
can build a Hamiltonian, a quantum operator that is a function of those observables.
This new operator can be used to recover the same Heisenberg equations, or to
define a Schrödinger equation for the wavefunction of the circuit.
In the Heisenberg formulation, we convert (4.1) or (4.2) to equations for the
infinite-dimensional operators V̂ and Iˆ:
d2
V̂ (t) = −ω2 V̂ (t). (4.4)
dt 2
Interestingly, we can compute the expectation value ψ0 |V̂ (t)|ψ0  of the voltage
V̂ (t) at time t, for a circuit that is initially in quantum state |ψ0 . We find that it
reproduces the dynamics of a classical circuit:
d2
V̂ (t) = −ω2 V̂ (t) . (4.5)
dt 2

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 40 Quantum Circuit Theory

In the Schrödinger formulation, we build a quantum Hamiltonian Ĥ = 12 C V̂ 2 +

1 ˆ2
2
LI using (4.3) and replacing variables with now static operators. The Hamilto-
nian defines a Schrödinger equation for the quantum state of the resonator circuit
|ψ(t), as
d
|ψ = Ĥ |ψ ,
i h̄ (4.6)
dt
with initial condition |ψ(0) = |ψ0 . As explained in Section 2.1.3, when we work
in the Schrödinger picture, observables do not change in time, but states do. Now the
expectation value of the voltage across the capacitor will be ψ(t)|V̂ |ψ(t), which
consistently follows the same equation (4.5).
We leave the study of Schrödinger or Heisenberg equations for later chapters. In
this one, the goal is to take any quantum circuit and derive equations such as (4.2)
or (4.4), together with their Hamiltonians. This requires first some basic notions of
electrical network theory. In Section 4.2, we provide definitions and conventions for
voltage and current, as well as the relations between them imposed by the circuit
elements. Section 4.3 adds Kirchoff’s laws to the mix, formulating the complete
quantum network theory.

4.2 Circuit Elements

Our study of quantum circuits is a generalization of the lumped-element circuit
theory. This is a very basic model for electrical circuits where we disregard the
internal structure of capacitors, inductors, or nonlinear devices, treating them as
point-like elements. Each element is described by two branch variables: the electric
current intensity I flowing through the element, and the voltage drop V experienced
by the charged particles that cross it. The lumped-element approximation is justified
when voltage and current oscillate at frequencies whose wavelength is much longer
than the size of the circuit elements. Under such conditions, V and I have no spatial
dependence and are related by the constitutive equations of the element under study.
This point of view is illustrated in Figure 4.2. A generic circuit element has two
connection points, A and B. When placed in a circuit network, a certain amount
of current I passes through the element. We assign a direction to the element and
associate the positive sign of the current I > 0 to currents that flow along that
direction. In this particular drawing, a positive current implies a flow of positively
charged particles from B to A. In ordinary circuits, the current originates from
a potential difference between the elements’ leads. Positive charges flow from a
region of high potential energy to a region of lower potential energy. This sets the
convention that when there is current going from B to A, there will be an associated
voltage difference V = VB − VA > 0 growing in the opposite direction.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.2 Circuit Elements 41

Figure 4.2 Circuit elements, together with the usual convention for current and
voltage. From left to right: generic element, capacitor, inductor, and battery. Note
that particles flow from the point of the circuit with the largest potential energy to
the area with the lowest potential energy. Hence, V and I have opposite directions.

Note that these conventions still work when the current is formed by negatively
charged particles, such as our Cooper pairs. Negatively charged particles move
along the direction of growing electric potential, from A to B, because their total
energy E = −2e × V is therefore decreased. However, if we understand the current
intensity as the time derivative of the charge I = dQ/dt, a flow of negative particles
from A to B is equivalent to a flow of positively charged particles from B to A.
Using these conventions, we will now analyze two linear elements, the capacitor
and the inductor, and then move on to the nonlinear Josephson junction. For all
three circuits, we will establish the voltage-current relations that will be used in
Section 4.3 to build the circuit’s dynamical equations.

4.2.1 Capacitor
We model a generic capacitor using two parallel plates, as shown in Figure 4.2. We
charge the capacitor by establishing a (positive) current that drains charges from
plate A and induces a positive charge in the opposite plate. As we know from our
physics textbooks, the charge imbalance creates a uniform electric field E directed
from the positive to the negative plate. This electric field is the gradient of a poten-
tial E = −∂V /∂x that decreases from the positively charged plate B down to
the plate A. Since E is proportional to the charge stored in the plates, we have a
proportionality relation:
 t
Q 1
V = VB − VA = = I (τ )dτ, (4.7)
C C 0

where the constant C ∝ (area×plate distance) is the total capacitance of the circuit.
We frequently use this relation in its differential form:

dV
I =C = C φ̈. (4.8)
dt

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 42 Quantum Circuit Theory

Charging a capacitor requires some work. Moving charges from potential VA to

VB without friction demands an energy
1 Q2 1 1
Ecap = QV = = CV 2 = C φ̇ 2 . (4.9)
2 2C 2 2
This is the electrostatic energy stored in the capacitive element. It will appear in this
form in all sorts of circuits, from microwave resonators to Josephson junctions and
superconducting qubits.
If the capacitor is made of a superconducting material, the transfer of n extra
Cooper pairs between the plates will change the energy by an amount
e2
Ecap = 4EC n2, . with EC = (4.10)
2C
Typical superconducting circuit elements will have small capacitances, ranging
from the picofarad to the femtofarad – 1 pF = 10−12 F and 1 fF = 10−15 F in SI
units – with a charging energy3 EC taking values between 10 μeV and 10 meV –
or frequencies from 40 MHz to 40 GHz. As with many other properties of a circuit,
the values the capacitances in a circuit will be typically fixed, and very accurately
predicted by the geometry of the metallic elements and dielectrics in the circuit.

4.2.2 Inductor
The inductance is the inertia that a conductor exhibits when we try to change the
current that traverses it. As sketched in Figure 4.2, a change in the current I passing
through a cable induces a change in the magnetic field B around it. By Maxwell’s
equations, an electric field will appear that opposes the change in magnetic field
∇ × E = −∂B/∂t. This electric field is equivalent to a potential drop parallel and
proportional to the time variation of the electric current:
dI
, or φ = LI .
V =L (4.11)
dt
The proportionality constant L is called the inductance, and (4.11) is the constitutive
relation for this lumped-element circuit.
Similar to the capacitance, the inductance relates the magnetic energy stored in
a circuit to the current intensity that passes through it. In an linear inductor L, we
need to spend an energy
1 1 2
Eind = LI 2 = φ (4.12)
2 2L
to raise the current intensity – or the electric flux – from zero to I .

3 Remember the factor 4 in the definition of E , which originates from the fact that the charging energy is a
C
concept that is also found in quantum dots and other electrical objects where the charge unit is an electron, not
a Cooper pair.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.2 Circuit Elements 43

The inductance of a circuit element depends on its geometry and the material with
which it is constructed. In the case of a superconducting material, we may also need
to consider the kinetic inductance. This accounts for how the current flow is affected
by microscopic details, such as the penetration depth of the material and the actual
thickness of the superconductor film (Orlando, 1991). This source of inductance is
particularly interesting when we wish to enhance the magnetic interaction between
circuit elements (see Section 7.3.4).

4.2.3 Josephson Junctions or “Nonlinear Inductors”

So far we have discussed circuit elements that appear in classical circuits. We now
introduce the Josephson junction, a circuit that is unique to the superconducting
world. If you haven’t done so yet, I recommend that you read Chapter 3 and in
particular Section 3.7. There we explain that a junction consists of a thin insula-
tor sandwiched in between two superconducting leads. The quantum tunneling of
Cooper pairs along the junction establishes a nonlinear relation between the current
I flowing through the junction and the flux difference φ between the leads:
   
2π φ
I = Ic sin φ = Ic sin . (4.13)
0 ϕ0
The constant Ic is the critical current of the junction, and we have introduced the
flux-phase relation ϕ0 = 0 /2π .
For small fluxes, a Josephson junction behaves as a linear inductor. Expanding
(4.13) to first order in the flux difference φ  0 , we obtain φ  LJ I with the
Josephson inductance
1 Ic
= . (4.14)
LJ ϕ0
In general, however, the Josephson junction is a nonlinear device that makes quan-
tum circuits more interesting, departing from traditional circuit theory, and enabling
new devices such as superconducting qubits and SQUIDs.
The nonlinear inductance adds new contributions to both the inductive and the
capacitive energies:
 
q2 2π
EJJ = − EJ cos φ . (4.15)
2CJ 0
The Josephson junction energy EJ = Ic /ϕ0  L−1 J will be the dominant term,
but we will often see also small capacitive contribution, mediated by charge
differences between the superconducting leads and their interaction through the
insulating dielectric. To account for this, junctions are usually drawn in parallel
with a fake capacitor CJ that captures this physics (see Section 4.6 and the lumped-
element circuits therein). Keep in mind, however, that this is just a notational

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 44 Quantum Circuit Theory

convenience: The capacitor and the nonlinear inductor are just one entity, and the
loop formed by these two lumped elements does not exist.
The values of the Josephson critical current and of the Josephson energy depend
very much on the actual application. For instance, the work by van der Wal et al.
(2000) uses three junctions with critical currents in the range Ic ∼ 400−600 nA.
This implies Josephson energies EJ ∼ 13.2−19.8 nJ or EJ /h̄ ∼ 1 250−1 870 ×
2π GHz. However, you will also find many qubits with Josephson energies that
have lower values, in the 100s of GHz.

4.2.4 Other Elements

General circuits also include voltage and current sources. We can treat these ele-
ments as boundary conditions that fix either the voltage V (t) or the current I (t),
leaving the other variable free to satisfy any other dynamical constraints in the
circuit. Figure 4.2 sketches a battery, a constant voltage source. The potential energy
is larger in the positive electrode A, and lower in the negative one, and the difference
between them is approximately constant, so that VA − VB = Vbat is independent of
time. An ideal battery sets no limit how much current it can supply to the circuit:
The electric current is determined by the dynamical equations of the whole circuit.
We will also find dc- and ac-current sources that provide either a constant I = I0 or
oscillating current I = I0 cos(ωt + φ), leaving the potential V free to accommodate
all other circuit restrictions.
General circuits also contain resistive elements. Establishing a current I through
such element demands a potential difference V = RI that compensates for the loss
of energy across the resistor R. There may appear some resistance when we combine
superconducting circuits with metallic parts, or as part of the Josephson junction
equations. In general, we try to minimize these resistances until they become negli-
gible, because of two problems. First, they remove energy in an incoherent way. This
prevents them from fitting into the unitary evolution formalism from Section 4.3.
And second, such resistive elements may have other negative effects, such as trans-
forming the absorbed energy into heat that affects the whole sample.

4.3 Quantization Procedure

Circuit quantization is a mathematical process aimed at building a quantum mechan-
ical model for a superconducting quantum circuit formed by linear elements –
capacitors, inductors, Josephson junctions, and voltage and current sources. As
discussed before, the theory applies at low temperatures, and also for excitation
energies well below the superconducting gap. The quantization is a multistep
process that constructs classical equations for the electrical circuit, then derives a

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.3 Quantization Procedure 45

Figure 4.3 Quantization steps. (a) Open tree and node fluxes; (b) identify all
oriented branch fluxes; (c) define loop fluxes and use flux quantization to find
independent branches. The direction of the loop defines the orientation of the
enclosed surface S with which we compute the flux (3.26).

classical Hamiltonian associated to those equations, and concludes with a quantum

version of the model. The process is similar to the LC resonator example in
Section 4.1.2 and to the canonical quantization of moving particles from Section 2.1.
I will present you with a version of the theory that is very similar to the quantization
rules described by Devoret (1995), discussing a few subtleties that appear in the
application and interpretation of the rules.
Assuming that we start from a lumped-element circuit scheme, such as the one
in Figure 4.3, the quantization steps are as follows:
(1) Labels: Identify all distinct nodes in the circuit and associate a flux variable
to each of them φj (cf. Figure 4.3a). A node is an intersection where two or
more circuit elements connect. All points in the same line are equivalent, and
a line can only have one distinct node.
(2) Circuit tree: Select a node – e.g., a ground plane – and build an open and
directed tree of branches T that runs through all nodes, without any loops.
Create a different set of closing branches that complete the loops and the
circuit C. In Figure 4.3b, T = {a,b}, and C = {c,d}.
(3) Branch fluxes: Associate a flux difference
φx = φj − φi , x = i → j ∈ T, (4.16)
to each branch x that includes a circuit element (cf. Figure 4.3b). This variable
is the difference between the fluxes at the end (j ) and beginning (i) of the
branch. For convenience, we can consider redundant orientations, using the
convention φ−x = φj →i = −φi→j = −φx .
(4) Flux quantization: For all branch elements that close loops x  ∈ C, deter-
mine the flux differences φx  using fluxoid quantization to set their value (see
Section 3.6). Given a series of directed branches {x} that form a properly
oriented loop, the sum of flux differences is equal to the number n of flux

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 46 Quantum Circuit Theory

quanta trapped in the superconducting ring4 and the magentic flux induced
by external fields loop, computed with the right orientation of the enclosed
surface (cf. Figure 4.3c):

φx = n0 + loop . (4.17)
x∈loop

When all branches belong to the main tree, except for branch x  = i → j, we
have
φx  = n0 + loop + (φj − φi ), x  = i → j ∈ C, (4.18)
to be compared with (4.16). Note that capacitors actually implement a physical
discontinuity in the conductor. When capacitive elements close a loop, n = 0,
and there will be no trapped fluxes.
(5) Branch currents: Compute the currents traversing each circuit element. For a
circuit on a branch b = i → j , the current will flow from j to i with a value I :
Capacitor Ij →i = Cb V̇b = Cb φ̈b See Section 4.2.1
Inductor I˙j →i = Vb /Lb ⇒ I = φb /Lb See Section 4.2.2
Junction Ij →i = IJ,b sin(φb /ϕ0 ) See Section 4.2.3
The branch flux φb is either (4.16) or (4.17), depending on whether b ∈ T or
b ∈ C. The capacitance Cb , inductance Lb , or critical current IJ,b are deter-
mined by the circuit element and do not depend on the branch orientation.
(6) Current conservation: Write down one equation per node, establishing that
the sum of currents coming in and out of the node add up to zero:

Ij →i = 0. (4.19)
j

This will produce one ordinary differential equation per node i in the tree, with
a general form
    
b 1 φb
si Cb φ̈b +
b
si φb + b
si Ib sin = 0, (4.20)
b∈capacitor b∈inductor
Lb b∈junction
ϕ0

where sb = ∂φb /∂φi = ±1 is a sign determining whether the branch b leaves

or enters the ith node.
(7) Holonomic constraints: Circuits are typically connected to ground planes,
batteries, or microwave sources. These fix either the voltage of a node φ̇k =
V (t) or the current along a branch. Use those constraints to eliminate variables
and equations that are no longer needed.

4 This value should be ideally zero, if the system is properly annealed. Note that discrete changes in the trapped
fluxes along an experiment may cause decoherence in the quantum state of the circuit.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.3 Quantization Procedure 47

(8) Lagrangian: Rewrite the differential equations (4.20) in the form

d ∂L ∂L
= , (4.21)
dt ∂ φ̇i ∂φi
and identify the Lagrangian L from which they are derived. If we follow the
preceding conventions, the Lagrangian will be a function of the node fluxes
φ = (φ0,φ1, . . .)T , containing the capacitive and inductive energies
L(φ, φ̇ i ) = Ecap (φ, φ̇ i ) − Eind (φ, φ̇), with (4.22)
 1
Ecap (φ, φ̇ i ) = Cb φ̇b2, and
b∈capacitor
2
 −φb2 
Eind (φ, φ̇ i ) = + EJ,b cos(φb /ϕ0 ). (4.23)
b∈inductor
2Lb b∈junction

We keep the branch fluxes φb because they make the expression more com-
pact, but these values should ultimately be replaced with the independent node
variables φi and φ̇i . Note also that, in order to have a proper identification of
energies, we cannot rescale (4.20) or introduce new variables before deriving
the Lagrangian.
(9) Canonical variables: Define the node charges as the canonically conjugate
momenta of the node fluxes:
∂L
qi = . (4.24)
∂ φ̇i
Use these variables and a Legendre transform to construct a Hamiltonian

H (q,φ) = qi φ̇i − L(φ, φ̇), (4.25)
i

where we use (4.24) to write flux derivatives as functions of charges and fluxes,
φ̇(q,φ). With this Hamiltonian, the evolution equations read
dφ ∂H dq ∂H
= , =− . (4.26)
dt ∂q dt ∂φ
(10) Quantization: As in Section 2.1, replace the canonical variables (φi ,qi ) with
operators (φ̂i , q̂i ) that satisfy commutation relations
[φ̂i , q̂j ] = i h̄δij . (4.27)
Define a Hamiltonian operator Ĥ using (4.25) with the flux and charge
operators as variables. Due to the separation between capacitive and inductive
energies, this operator will be Weyl-ordered and Hermitian. As explained in
Section 2.1.3, this operator can be used to derive the Heisenberg equations
for charge and flux, or to study the Schrödinger equation for the circuit
wavefunction (4.6).

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 48 Quantum Circuit Theory

In this whole treatment, we have worked with fluxes φi on the nodes of the tree
because they are the canonical variables and because one can derive the general form
of the Lagrangian (4.22) from the current equations on these nodes (4.20). However,
this is not a unique way to work with the circuit: Any change of coordinates that
satisfies the constraints of our system will lead to a valid quantization procedure.
We can work with node charges, branch charges,5 or branch fluxes as variables
instead. We will discuss this last case because it illustrates the structure of the
resulting Hamiltonians and because it is convenient for some Josephson junction–
based circuits.
Nodes and tree branches are largely interchangeable. On every connected tree
there will be N + 1 nodes and N branches. When the tree includes voltage sources
or ground planes, the number of independent variables is identical in both repre-
sentations. Each node flux can then be written as a reference potential plus various
branch fluxes from on the tree. The change of coordinates from branch variables
φ b to node fluxes φ = Aφ b is performed by a matrix of ±1’s and 0’s. We use this
matrix to express the Lagrangian (4.22) as a function of the branch fluxes L(φ b, φ̇ b ).
In doing so, we must remember that closing branches are sums of branch variables

on the tree, due to flux quantization φc = n0 +  − b φb . We may also need to
reduce the number of variables in φ b to account for voltage or current constraints.
The inductive terms remain similar in structure and do not need to be discussed.
The capacitive terms adopt a quadratic form:
1 T
Ecapacitive (φ̇ b ) = φ̇ b C φ̇ b + Cg VT φ̇. (4.28)
2
This includes a capacitance matrix Cbb and a voltage source term V connected
to the circuit through the coupling capacitances Cg . The capacitive term fixes the
definition of branch charges Qb = ∂L/∂ φ̇ b = C φ̇−Cg V, and gives the Hamiltonian
1
H (Qb,φ b ) = (Qb − Cg V)T C −1 (Qb − Cg V) (4.29)
2
 φ2 
+ b
− EJ,b cos(φb /ϕ0 ).
b∈inductor
2Lb b∈junction

This Hamiltonian is mathematically identical to the one derived for φ. We can

now adopt canonical commutation relations for the new flux and charge operators
[φ̂b, Q̂b ] = i h̄δbb , and work with them in the same way as with the node operators.

5 These are, for instance, the variables used in the works by Yurke and Denker (1984) or Blais et al. (2004) to
derive equations for transmission lines and resonators.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.4 LC Resonator 49

It is important to remark how the capacitance matrix C appears in the final

Hamiltonian. The original equivalent circuit was based on a series of nearest-
neighbor interactions, given by the capacitive energy across different circuit
elements, 12 C(φ̇i − φ̇j )2 . This typically means that C and Cg are both sparse
matrices and vectors reproducing a simple, typically planar topology. However,
once we derive the canonical variables, we find that the inversion of the matrix C −1
establishes long-range, dipolar-type interactions between arbitrary pairs of charges.
On the one hand, this is one of the sources of cross-talk that has to be taken care of
when building complex quantum circuits – e.g., in large-scale quantum computers
with hundreds of qubits and resonators, we would like to have no residual interaction
between faraway elements. On the other hand, such long-range interactions are
very similar to what we find in many-body systems, and they can be harnessed
for the purpose of simulating complex condensed matter or theoretical physics
models in the lab!

4.4 LC Resonator
Let us apply the quantization formalism to the LC circuit shown in Figure 4.4a,
which consists of a superconducting resonator coupled to a constant voltage source.
The four lumped elements in this circuit are the resonator’s (i) capacitor C and
(ii) inductor L, (iii) the voltage source, and (iv) the gate capacitor that connects
it to the resonator, Cg . Using this information, please try to develop the quantum
Hamiltonian yourself, checking the result with the following derivation – one of
many equivalent ones.
In Figure 4.4b, we have constructed one possible circuit tree. The illustration
marks the three different nodes {φ0,φ1,φ2 }, three tree branches {φa ,φb,φc }, and a

Figure 4.4 (a) Equivalent circuit for an LC resonator consisting on an inductor

in parallel with a capacitor, subject to an external potential V (t). (b) Nodes, node
fluxes, and branch fluxes for the circuit description and quantization.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 50 Quantum Circuit Theory

closing branch φd . From left to right, the nodes include a ground plane – common
to the inductor, the capacitor, and the battery – the intersection between the inductor
and the two capacitors φ1 , and the cathode φ2 . We have also chosen directions for
all branches, defining the sign of potential differences:

φa = φ1 − φ0, (4.30)
φb = φ2 − φ1, (4.31)
φc = φ2 − φ0 . (4.32)

The ground plane is a constant potential reference, which allows us to elimi-

nate one variable φ0 = 0. The battery fixes the branch fluxes φ̇c = φ̇2 − φ̇1 = V ,
which implies φ̇2 = V and φ̇b = V − φ̇1 . Since the LC cannot support a supercur-
rent, the voltage difference along the inductor L (solid line) and along the capaci-
tor C (dashed line) are identical, φa = φd = φ1 . This leaves only one independent
variable, φ1 .
With the conventions of voltage and current from Section 4.2, the flux difference
φa is associated to the left-moving currents φ̇a /L and C φ¨a along the capacitor and
inductor. These currents equal the current coming from the battery through Cg ,
1
C φ̈1 + φ1 = Cg (V̇ − φ̈1 ). (4.33)
L
Introducing C = C + Cg,
d 1
C φ̇1 − Cg V = − φ1 . (4.34)
dt L
This equation is conservative and derived via (4.21) from the Lagrangian
1 1 2
L = C φ̇12 − Cg φ̇1 V − φ . (4.35)
2 2L 1
The charge variable is the conjugate momentum of the node flux q1 = ∂L/∂ φ̇1 =
C φ̇1 −Cg V . Using a Legendre transform H = q1 φ̇1 −L and φ̇1 = (q1 +Cg V )/C ,
we obtain the Hamilltonian
1 1 2 1 1 2
H = (q1 + Cg V )2 + φ1 = (q1 − qg )2 + φ, (4.36)
2C 2L 2C 2L 1
with the externally induced charge qg = −Cg V .

4.5 Transmission Line

We will now study a model for circuits that transport microwave energy without
significant losses. Many physical objects can do this: coaxial cables that bring the

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.5 Transmission Line 51

Figure 4.5 Equivalent circuit for a transmission line or a coplanar waveguide.

The line is discretized as a set of coupled resonators, each of size x and with
capacitances and inductances given by (4.37). The tree starts in the ground plane
and spans as a star to all nodes through the vertical capacitors.

TV signal to your home, superconducting lines printed on a two-dimensional chip,

or copper and aluminum tubes. Most of these objects are described with the equiv-
alent circuit from Figure 4.5. This lumped-element model is an array of coupled
LC resonators. The capacitors account for the electrostatic energy stored in the
waveguide, while the inductors model the inertia against changes in the electric
current. The circuit is a discretized version of the guide where each capacitor and
inductor accounts for a small segment x that is much smaller than the guided
wavelengths. The properties of these elements depend on the capacitance and induc-
tance per unit length:
Ci = ci x, Li = li x. (4.37)
We will assume that the quantities ci = c(xi ) and li = l(xi ) are smooth functions of
the position along the waveguide xi .
The circuit tree is an open star with the ground plane at the center φ0 , and branches
along the capacitors that connect the ground to the individual node fluxes φi . As in
the LC, we can take φ0 = 0 as reference. The branch fluxes along the inductors
are leftward oriented and φi+1→i = φi+1 − φi . Current conservation on the nodes
balances the flow of charges between inductors with the current flowing in and out
of the ground plane:

1 1
Ci φ̈i = (φi+1 − φi ) − (φi − φi−1 ). (4.38)
Li Li−1

This equation can be derived from the Lagrangian:


N
1 
N−1
1
L= Ci φ̇i2 − (φi+1 − φi )2 . (4.39)
i=1
2 i=1
2Li

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 52 Quantum Circuit Theory

The charges follow the expected definition qi = ∂L/∂ φ̇i = Ci φ̇i . The Legendre
transform produces the Hamiltonian:

N N
1 2  1
N−1
H = qi φ̇i − L = qi + (φi+1 − φi )2 . (4.40)
i=1 i=1
2Ci i=1
2Li

We can quantize the theory replacing (φi ,qi ) with operators satisfying [φ̂i , q̂j ] =
i h̄δij . The continuum limit of this theory is obtained by inserting the actual size of
the discretization:
  2
N
x q̂i 2  x
N−1
φ̂i+1 − φ̂i
H = + , (4.41)
i=1
2c(xi ) x i=1
2l(xi ) x

defining the charge density ρ̂(xi ) = q̂i /x, and replacing sums with integrals to
obtain a quadratic field theory:
  
1 1
H = dx ρ̂(x) +
2 2
[∂x φ̂(x)] . (4.42)
2c(x) 2l(x)

These fields satisfy new commutation relations [φ̂(x), ρ̂(y)] = i h̄δ(x − y).

4.6 Charge and Transmon Qubits

We will now study a circuit that, depending on the parameter regime, describes
either a charge qubit and a transmon qubit. The circuit, shown in Figure 4.6, is a
simple nonlinear version of the LC circuit: It consists of a superconducting island
coupled to a ground plane via a Josephson junction and externally controlled by a
voltage source. We model this qubit using four lumped elements: (i) a junction with
critical current Ic ; (ii) the total capacitance C between the island and the ground

Figure 4.6 (a) Equivalent circuit for the charge qubit, with (b) the actual network
on the right. The superconducting island is the region contained between the two
capacitors, CJ and Cg . The nonlinear element EJ represents the channel by which
pairs can tunnel into the ground plane.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.7 SQUIDs 53

plane – both direct and mediated by the junction; (iii) a gate capacitor Cg ; and (iv)
voltage source V , which accounts for any external control on the island.
The quantization procedure copies the one in Section 4.4, using the same inde-
pendent variables and constraints. We simply replace the linear inductor with the
junction’s nonlinear current-voltage relation
 
d Cg
C φ̇1 − V̇ = −Ic sin(φ1 /ϕ0 ). (4.43)
dt C
The Lagrangian’s potential energy is now the nonlinear inductive energy of the
junction
1
L = C φ̇12 − Cg φ̇1 V + EJ cos(φ1 /ϕ0 ), (4.44)
2
with EJ = Ic ϕ0 . The charge operator has the same expression and the Hamiltonian
becomes a nonlinear function
1
H = (q − qg )2 − EJ cos(φ1 /ϕ0 ), (4.45)
2C
where the external potential sets the equilibrium charge qg = −Cg V in the ground
state of the qubit.

4.7 SQUIDs
We will now discuss the circuits for two different superconducting quantum inter-
ference devices (SQUIDs). These are superconducting circuits that make use of flux
quantization and Josephson tunneling. This feature makes the SQUID so sensitive
to magnetic fields that they can be used as high-precision magnetometers in biology,
medicine, or geology, among other fields. In a way, SQUIDs are the peak of super-
conducting circuit technology, a vast field of research6 that predates and enables
modern-day quantum circuits. In this book, we do not have space to go deep into
this field, but will frequently rely on the SQUIDs to implement superconducting
qubits, add tuneability to circuits, and build quantum amplifiers.

4.7.1 rf-SQUID
The rf-SQUID consists of a Josephson junction that is shunted by a linear inductor,
as shown in Figure 4.7. The device is called a radio-frequency SQUID because
it relies on the ac Josephson junction effect. It is a cheap magnetometer and also

6 See, for instance, The SQUID Handbook by Clarke and Braginski (2004).

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 54 Quantum Circuit Theory

one of the earliest superconducting qubits, which still finds applications in modern
quantum annealers (see Section 9.4.1).
Compared to previous devices, the rf-SQUID introduces new formal challenges
that were not discussed in circuit quantization: (i) We have a loop without reference
potential; (ii) the SQUID couples to external current sources via a mutual induc-
tance M; and (iii) the external currents induce a net magnetic flux in the loop  = 0,
which affects the dynamical equations.
The equation for the node variable φ0 balances the currents on the junction and the
linear inductor. The current in the inductor is partially caused by the self-inductance
of the branch φL /L, and partially by the mutual inductance. According to our sign
conventions, the downward external current I induces an upward moving current
on the SQUID −MI , so as to reduce the total magnetic interaction between both
circuits. Hence,
1
CJ φ̈J + Ic sin(φJ /ϕ0 ) =
φL − MI . (4.46)
L
We replace the branch fluxes φJ and φL with their definitions in terms of flux nodes,
using the fluxoid quantization. Assuming no trapped vortices and with our choice of
tree, we have φL = φ1 −φ0 and φJ = − + (φ0 − φ1 ) = −( + φL ). Note that the
total magnetic flux  is created both by the external currents I and by the current
supported by the loop itself φL /L. Far away from the regime of bistability, we can
neglect this last contribution and approximate ˙ = 0. The equation on the lower
node

d   φ0 − φ1 −  φ1 − φ 0
CJ (φ̇0 − φ̇1 ) = −Ic sin + − MI (4.47)
dt ϕ0 L

is identical to that of the one on the upper node, up to a global sign. Both equations
have the form 4.21 with the Lagrangian

Figure 4.7 Equivalent circuit for the rf-SQUID qubit in full version (a) or restricted
to the SQUID (b). A Josephson junction (CJ ,EJ ) is shorted by a large inductor L
that may couple to an external magnetic field generated by a current source. The
mutual inductance M is introduced in the text.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.7 SQUIDs 55

1 1 φ0 − φ1 − 
L = CJ (φ̇1 − φ̇0 )2 − (φ1 − φ0 )2 − M(φ1 − φ0 )I − EJ cos ,
2 2L ϕ0
(4.48)

and the Josephson energy EJ = Ic ϕ0 .

These equations do not set any value for φ0 + φ1 , which is a constant of motion
d
dt
(φ0 + φ1 ) = 0 decoupled from the flux difference φL . Following Section 4.3, we
use φL as independent variable. Expressing the Lagrangian in terms of this variable,
we obtain the conjugate momentum QL = CJ φ̇L and the Hamiltonian

1 1 2 1
H  Q2L + φL − MI φL + EJ cos (φL + ) . (4.49)
2CJ 2L ϕ0

If we look at the inductive terms, the role of MI is to shift the flux φL at which the
minimum of energy is reached. However, this shift MI L will have to compete with
the tendency of the nonlinear potential to trap the flux around φL ∝ 2πϕ0 + . In
a regime of maximum frustration, where /ϕ0 ∼ π , we can obtain an energy land-
scape with two local minima that can be applied to create a qubit, as discussed in
Section 6.4.

4.7.2 dc-SQUID
The “interference” in the word SQUID is revealed by circuits with multiple junc-
tions, such as the dc-SQUID from Figure 4.8a. In this lumped-element model, there
are two junctions connected in parallel, forming a loop. The incoming supercurrents
split into two distinct paths, interfering at the exit port. The outcome of this inter-
ference and the total dc current that passes through depend on the magnetic flux 
injected through the SQUID.

in out

Figure 4.8 (a) A dc-SQUID is a device consisting on two junctions in parallel,

threaded by some magnetic flux. (b) A more detailed version of the circuit must
take into account the capacitive and inductive energy of the junctions if the pair is
unbalanced.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 56 Quantum Circuit Theory

In a conservative scenario, where the resistance of the junctions is negligible, the

total incoming and outgoing currents must be the same and must split equally on
both nodes: I = Id − Iu = Iin = Iout . As in the rf-SQUID, there is only one inde-
pendent equation that summarizes the conservation of current:

CJ 2 φ̈d + IJ 2 sin(φd /ϕ0 ) = CJ 1 φ̈u + IJ 1 sin(φu /ϕ0 ) + I .

Using flux quantization and our choice of main tree, the upper and lower branch
fluxes are φu = φ1 − φ0 and φd =  + φ0 − φ1 , with the total external and trapped
fluxes . We assume identical junctions and use the identity sin(a) ± sin(b) =
2 sin a±b
2
cos a∓b2
to write
   
φd − φu φd + φu
CJ (φ̈d − φ̈u ) = I − 2Ic sin cos x. (4.50)
2ϕ0 2ϕ0

Let us make a second change of variables φ± = 12 (φd ± φu ). The sum φ+ = −/2

is the circulating flux in the SQUID, while the difference φ− defines the total
current passing through the device. The effective Lagrangian and corresponding
Hamiltonian are
 
1 φ−
L = (2CJ )φ̇− + EJ () cos
2
− I φ−, (4.51)
2 ϕ0
 
Q2 φ−
H = − + I φ− − EJ () cos . (4.52)
4CJ ϕ0
This Hamiltonian describes a new type of circuit element: the tunable Josephson
junction. This is an effective junction with twice the original capacitance 2CJ , where
the Josephson energy can be adjusted between −2Ic ϕ0 and +2Ic ϕ0 using an external
magnetic field :

EJ () = 2Ic ϕ0 cos(/2ϕ0 ) = EJ (0) cos(/2ϕ0 ). (4.53)

This element is used to make other quantum circuits tunable. For instance, we
can upgrade a charge qubit (Section 6.2), replacing its junction with a SQUID to
adjust the qubit’s spectrum of energies. We can also use the SQUID in place of a
linear inductor, such as in an LC resonator or transmission line, to modulate the LC
resonance in time and create a parametric amplifier (see Section 5.5.6).
dc-SQUIDs can be used to measure magnetic fields with high precision. The crit-
ical current Ic () can change from 0 to hundreds of nanoamperes over the span of a
flux quantum 0 ∼ 2 × 10−15 T · m−2 . This means that if we can detect changes in
Ic (), we can use a SQUID to measure a magnetic field (or flux) with sensitivities
close to attoteslas. In order to measure Ic , we can rely on the dynamics of a current-
biased SQUID:

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.8 Three-Junction Flux Qubit 57

Figure 4.9 Washboard potential of the dc-SQUID (4.54) (a) below and (b) on the
verge of switching to a voltage state.

 
φ−
Eind = −EJ () cos − I φ− (4.54)
ϕ0
For small biases, this function has the shape of a washboard potential (cf.
Figure 4.9a). The infinitely many local minima are metastable configurations of
the flux φ− with V ∝ φ̇− = 0. When the bias current I exceeds the critical value
Ic () = 2Ic cos(/2ϕ0 )/ϕ0 , all such minima disappear (cf. Figure 4.9a), and the
flux “flows” down the slope, linearly accelerating in time.7 This translates into a
measurable potential difference between the SQUID’s ports, which can be detected.
In the world of quantum circuits, SQUIDs are used to measure the current state
of other circuits. For instance, a flux qubit has two states |0 and |1 that correspond
to different orientations of the supercurrent. These states can inject different values
of the flux 0 and 1 into a nearby SQUID. We can bias the SQUID to a point in
between the critical currents for those two states, say Ic (0 ) < I < Ic (1 ). When
doing so, the SQUID will either remain the same or switch to a voltage state, depend-
ing on whether the qubit is in the |1 or |0 states. These are almost projective
measurements, limited in precision by noise in the SQUID and the finite duration
of the measurement.

4.8 Three-Junction Flux Qubit

The last circuit in this chapter is a superconducting loop with three Josephson junc-
tions, shown in Figure 4.10a. The lumped-element circuit has three independent
node variables φ0,1,2 , which combine to create two independent branch variables

7 This is a formally ill-defined problem, which stabilizes to an asymptotically stable solution once we consider
the resistance of the junctions, or add a resistor in parallel to the circuit.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 58 Quantum Circuit Theory

Figure 4.10 (a) The qubit is a loop with three Josephson junctions, two identical
(1,2) and a smaller one (3). (b) The equivalent circuit where we focus only on the
branch fluxes.

φa = φ1 − φ0 and φb = φ2 − φ1 , plus a closing branch φc = − − φa − φb =

− + φ0 − φ1 . There are two current equations
IJ 3 sin(φc /ϕ0 ) + CJ 3 φ̈c = IJ 1 sin(φa /ϕ0 ) + CJ 1 φ̈a (4.55)
= IJ 2 sin(φb /ϕ0 ) + CJ 2 φ̈b,
which derive from the Lagrangian:
1
L = (CJ 1 + CJ 3 ) φ̇a2 + φ̇b2 + CJ 3 φ̇a φ̇b (4.56)
2      
φa φb  + φa + φb
+ EJ 1 cos + EJ 2 cos + EJ 3 cos .
ϕ0 ϕ0 ϕ0
These qubits are designed with two identical junctions CJ 1 = CJ 2 = Cj and
EJ 1 = EJ 2 = EJ , and a junction with an area α times smaller, so that EJ 3 = αEJ ,

CJ 3 = αCJ . As in the rf-SQUID, the sum of node fluxes i φi is a constant of
motion,8 which leaves only two branch fluxes as independent variables. We now
introduce a change of coordinates
φ± = φa ± φb . (4.57)
With the choice of the preceding parameters and the new variables, we find the
charges Q+ = 12 CJ (1 + 2α)φ̇+ , and Q− = CJ φ̇− , and the Hamiltonian:
1 Q2+ 1 Q2−
H = +
2 CJ ( 12 + α) 2 CJ
     
 + φ+ φ− φ+
− EJ α cos + 2 cos cos . (4.58)
ϕ0 2ϕ0 2ϕ0
This energy functional has a complicated landscape of equilibrium states, depending
on the ratio of areas α and the external flux , which we will discuss in Section 6.4.3.

8 We can set this variable to zero by grounding one of the nodes.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 4.9 Number-Phase Representation 59

4.9 Number-Phase Representation

In the last sections, we have derived quantum mechanical Hamiltonians for various
circuits, specifying no more than a generic commutation relation between flux and
charge operators [φ̂b, q̂b ] = i h̄δbb . We will now develop two mathematical rep-
resentations of these operators. These are descriptions of the flux φ̂ and charge q̂
observables, as operators acting on a Hilbert space of quantum states.
Unfortunately, even though flux and charge satisfy a canonical commutation rela-
tion, they are not completely equivalent to position and momentum. First of all,
unlike momentum, the charge is a discrete operator that is proportional to the num-
ber of Cooper pairs q̂ = −2en̂. Second, the flux is related to the superconductor’s
phase ϕ̂ = 2π φ̂/0 . This results in a periodic representation for the phase operator,
remarkably different from the usual position operator x̂. Both differences imply that
we cannot simply use the representation of position and momenta in terms of plane
waves, and we must derive a specific phase-number representation.
There exist two representations of the phase-number or charge-flux commutation
relations9 that are operationally useful and can be applied to solve most problems:
Phase representation: We construct the Hilbert space using the eigenstates of the
flux operator φ̂ or the equivalent phase operator ϕ̂ = 2π φ̂/0 . In this represen-
tation, the circuit wavefunction
 2π  2π
| = dϕ1 · · · dϕN (ϕ1, . . . ,ϕN ) |ϕ1, . . . ,ϕN 
0 0

expands over states with a well-defined value of the node phases |ϕi . This
continuous representation transforms the Schrödinger equation into a partial-
differential equation for (ϕ1, . . . ,ϕN ), which can be solved analytically in
some important cases, such as the transmon qubit (Section 6.3).
Number representation: We construct the Hilbert space using the eigenstates of
the charge operator q̂, which are states with a given number of excess Cooper
pairs q̂ = −2en̂. These states form a discrete basis {|n , n ∈ Z}. Operators such
as n̂, q̂, or the Hamiltonian H are represented as infinitely large matrices, which in
most cases can be truncated to a reasonable cutoff |n| ≤ nmax , either for analytical
approximation – see the charge qubit in Section 6.2 – or for some numerical
diagonalization method – see the three-junction flux qubit in Section 6.4.3.
In each representation, we have one operator that is diagonal and simple, but
we ignore the conjugate observable. In the number representation, the charge

9 Our notation is consistent with Pegg and Barnett (1989) and with the original definition of the phase operator
by Susskind and Glogower (1964).

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 60 Quantum Circuit Theory

operator is a sum of projectors onto well-defined numbers of Cooper pairs,


q̂ = n (−2en) |nn| for n ∈ Z. We know that all states in the Hilbert space
can be expressed as linear superpositions of number states | = n n |n, but
we ignore how to write flux φ̂ and phase ϕ̂ in this basis.
In order to find a faithful representation of the flux and phase, we must look at
the Lie algebra they generate. We use the canonical commutation relations between
φ̂ and q̂ to find an alternative equation:

ei φ̂/ϕ0 q̂ = (q̂ − 2e)ei φ̂/ϕ0 . (4.59)

Note how the phase operator ϕ̂ = φ̂/ϕ0 is the generator of displacements in the
space of charges. Using the number representation, we recover a dimensionless
version of this identity:
n̂ei ϕ̂ = ei ϕ̂ (n̂ − 1), or (4.60)
e |n = |n − 1 ,
i ϕ̂
(4.61)
and conclude that the exponential of the phase is a ladder operator:

ei ϕ̂ = |n − 1n| . (4.62)
n

The Josephson energy combines two exponentials, allowing processes in which a

Cooper pair tunnels in or out of the superconducting island:
1 1
cos(ϕ̂) = |n + 1n| + |nn + 1| . (4.63)
2 n 2 n
We can perform a similar calculation in the basis of eigenstates of the phase
operator. Now we have to study how the phase and number operators act on a generic
wavefunction (ϕ):
   
n̂ ei ϕ̂ (ϕ) = eiϕ (n̂ − 1)(ϕ) . (4.64)

Notice that we have safely replaced the ϕ̂ operator with its representation, the phase
degree of freedom ϕ. The previous equation is solved by the identification n̂ = i∂ϕ
and we can derive the phase eigenstates from the equation
m|ei ϕ̂ |ϕc  = m + 1|ϕc  = eiϕc m|ϕc  , (4.65)
which implies
1  iϕc m
|ϕ = √ e |m . (4.66)
2π m∈Z
This discussion completes a dictionary for working with flux and charge, sum-
marized in Table 4.2. Using these tools, we can take the Hamiltonian of a Josephson

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 Exercises 61

Table 4.2. Mathematical representations of the flux and

charge operators.

Continuous Discrete
Variable ϕ∈R n∈Z

Charge q = −i2e∂ϕ q = −2e n n |nn| = −2e n̂
0 
Flux φ̂ = 2π ×ϕ ei2π φ̂/0 = n |n − 1n|.

junction (4.45) and write it in different representations. In the phase variables, the
state is given by a periodic wavefunction (ϕ,t) that evolves in time with a partial-
derivative equation:

∂ ∂ 2
i h̄ = −4EC 2 − EJ cos(ϕ). (4.67)
∂t ∂ϕ
In Chapter 6, we solve analytically a stationary version of this equation, but in
practical applications with two or more variables, or with additional terms, it is not
possible to do it. We could in principle address those complicated examples using a
method such as finite differences. However, a simpler approach is to write the same
Hamiltonian in the number representation. For instance, in the Josephson junction
case (4.67) we have

H = [4EC n |nn| − EJ (|n + 1n| + |nn + 1|)] . (4.68)
n

In many real-world applications, where EJ /EC ≤ 100, the probability that a state
with large number of Cooper pairs is occupied is very low. This means we can
reduce the complexity of the Hamiltonian, truncating the matrices and vectors to
a smaller subspace, of |n| ≤ nmax ∼ 10 or less. In some cases, such as the charge
qubit 6.2, this truncation is even more dramatic, and we end up in a two-dimensional
subspace, such as n ∈ {0,1}.

Exercises
4.1 Discuss the differences between the thermal state ρ∞ = 12 |00| + 12 |11|,
and the quantum superposition |ψ = √12 (|0 + |1). What is the density
matrix ρq associated to the latter state? How can we distinguish both states
experimentally if we have access to measurements of the σ z and σ x operators?
4.2 Using the commutation relation [φn,qm ] = i h̄δnm, prove (4.59). Hint: Write
down the Taylor expansion of the exponential, and use the commutator on
each term.

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 62 Quantum Circuit Theory

4.3 Reproduce the derivations of all the circuit models and effective Hamiltonians
from Sections 4.4 to 4.8.
4.4 Using (4.1), show that the energy of the LC circuit H = 12 CV 2 + 12 LI 2 is
conserved, that is, dH /dt = 0. What happens if you add a resistor in parallel
to the circuit?
4.5 Draw a circuit of two capacitively coupled LC resonators. The circuit should
be the result of replacing the battery in Figure 4.4 with another resonator. Find
out the quantum model for this circuit. Find out the canonical variables Q̃n and
φ̃n that diagonalize the Hamiltonian, and bring it to a form
 Q̃2 1
Heff = n
+ Ln φ̃n2 . (4.69)
n
2Cn 2
4.6 Similar to Problem 4.5, derive the circuit and the Hamiltonian for two identical
charge qubits connected via a capacitor. Show that, in the limit of weak cou-
pling, the interaction between both qubits is modeled by the product of charges
∝ q1 q2 from both qubits.

4.7 Compute analytically the value of the bias current at which the washboard
potential from Figure 4.9 loses all its local minima.
4.8 Think how you would write the Hamiltonian for the 3-JJ flux qubit (4.58) as
a matrix in the number-phase representation. Remember to use the decompo-
sition of trigonometric functions into exponentials that have a simple matrix
representation cos(ϕ) = 12 (eiϕ + e−iϕ ). Also realize that eiϕ+ eiϕ− will translate
into a Kronecker or tensor product of two matrices A ⊗ B, each representing
the action of the exponential on a different Hilbert space – such as A for eiϕ+
and B for eiϕ− .

https://doi.org/10.1017/9781316779460.004 Published online by Cambridge University Press

 5
Microwave Photons

In this chapter, we begin a journey through the practical applications of supercon-

ducting quantum circuits. Our first stop is linear circuits, such as the LC resonators
and transmission lines introduced in Sections 4.4 and 4.5. We discuss how the
energy of these circuits is quantized in units that we call photons – more precisely,
microwave photons – and how this reflects in the Hamiltonian and the representation
of observables. Microwave photons can be stored in LC resonators and microwave
cavities for a long time, or they can travel through transmission lines to distant cir-
cuits and detectors. Microwave photons may be prepared in very different quantum
states, from well-defined number states, to entangled states and sophisticated quan-
tum superpositions. We explain how to create and characterize these states using
linear or “Gaussian” operations, which include microwave sources, inductive and
capacitive interactions, parametric quantum amplifiers, and devices that meassure
voltage and power.

5.1 LC Resonator
5.1.1 Energy Quantization and Photons
Figure 5.1a shows a very small LC circuit made of two superconducting elements: a
very visible interdigitated capacitor with many fingers, shunted by a linear inductor
in the form of a small superconducting loop. The LC circuit is so small that the
lumped-element circuit approximation is well justified. To be more precise, we can
tune global properties, such as the total charge, flux, voltage, or intensity, but we
cannot control these properties’ distribution across the circuit, because this “struc-
ture” or mode is fixed in a state of minimum energy. If we want to alter the mode
structure, we need to introduce changes over distances d ∼ 100 μm. The energies
associated to these tiny details is hundreds of GHz, well above the superconducting
gap and far from our experiments’ operation regime.

63

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 64 Microwave Photons

0.2 mm

Figure 5.1 (a) Zero-mode LC resonator formed by an interdigitated fingers capac-

itor and a short line that acts as an inductor. Courtesy of Pol Forn-Díaz. (b) Equiva-
lent circuit. (c) λ/2 microwave resonator formed by a waveguide intersected by two
capacitors, C1 and C2 . Reprint from Pechal et al. (2014). (d) Three-dimensional
microwave cavity formed by carving two aluminum blocks that will be joined
together, forming a closed hull. Reprinted figure with permission from Paik et al.
(2011). Copyright (2011) by the American Physical Society.

The resonator in Figure 5.1a not only a very good example of the lumped-
element approximation, it even resembles the graphical representation of the LC
circuit, shown in Figure 5.1b, which we introduced in Section 4.4. As discussed
then, charge, flux, voltage, and current all follow a similar oscillatory equation,
φ̈ = −ωLC 2
φ, solved as φ(t) ∼ cos(ωt + ϕ) with a single frequency that depends
on the electrical properties of the circuit:

1
ωLC = √ . (5.1)
LC

Because this circuit only supports a fixed spatial distribution of charge and flux,
it only has a single mode or frequency, which is why it is also called a zero mode
resonator.
The LC resonator is conceptually similar to the ideal oscillators that Planck used
in his model of the black-body’s electromagnetic radiation (see Section 2.1). Fol-
lowing Planck’s and Einstein’s ideas, we could propose that our superconducting
resonator is a quantum mechanical object that only exchanges discrete units of
energy with the environment – our quanta or microwave photons. The states of the

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.1 LC Resonator 65

resonator would then be labeled by the number of photons that it stores |n, and the
total energy would be a sum of quanta En ∼ h̄ω × n, n ∈ {0,1,2, . . .}.
This qualitative reasoning is essentially correct. It can be justified using Dirac’s
quantization and the quantum circuit models from Section 4.3. The “photons” that
we find in the superconductor have a mixed nature: They are excitations of the
electromagnetic field that are accompanied by waves of charge and current on the
surface of the metal. This hybrid nature would make it more appropriate to call
them microwave plasmons, but we keep the name “photon” to distinguish these
plasmonic excitations from others that have a more anharmonic spectrum, which
we use to build qubits.
The microwave photons stored in a resonator do not interact with each other,
at least not until we reach high powers that excite nonlinear contributions to the
inductance and capacitance. Also, when placed on extended larger systems, such
as a superconducting waveguides (cf. Figure 5.1c), the photonic nature of these
plasmons evidences in a propagation without dispersion, at a fraction of the speed
of light.
However, the microwave photons do have a matter-like component. This com-
ponent allows us to control the properties of the photon – frequency, speed, wave-
length, etc. – by a proper design of the circuit: capacitance, inductance, thickness
of the material, etc. Moreover, the matter component in the wave enables the strong
interaction between these microwave photons and other circuits, atoms, molecules,
NV-centers, etc., opening regimes of extreme light–matter interactions that were
previously unreachable in quantum optics.
Unfortunately, working with microwave photons has some drawbacks. In particu-
lar, the small energy contained in a microwave photon makes it very difficult to build
a microwave photodetector or photon counter. The lack of such devices prevents
us from directly verifying the quantization of the cavity’s excitations En = h̄ωn, a
fact that only becomes evident when we introduce other, more matter-like nonlinear
circuit elements, such as qubits.

5.1.2 Hamiltonian Diagonalization

We already have a quantum mechanical model for the LC resonator (4.36):
1 1 2
H = (q̂ − qg )2 + φ̂ . (5.2)
2C 2L
This model is analogous to the problem of a quantum particle in a harmonic poten-
tial. In absence of charge bias qg = 0, we identify
1 2
2C
q̂ + 12 φ̂ 2 ⇔ 1 2
2m
p̂ + 12 mω2 x̂ 2 ,
[φ̂, q̂] = i h̄ ⇔ [x̂, p̂] = i h̄.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 66 Microwave Photons

The capacitance is the oscillator’s mass C ∼ m, while the inductance acts as a

1
restoring potential, 2L ∼ 12 mω2 . Together they produce the natural frequency of
the resonator ω = (LC)−1/2 .
This dictionary also maps the textbook methods1 for bringing the harmonic oscil-

lator into diagonal form Ĥ = n En |nn|, with a complete set of eigenstates |n.
Specifically, we know that the oscillator eigenstates are Fock states |n labeled with
all nonnegative integers2 n ∈ {0,1,2, . . .}. The excitations in these states are created
and annihilated by the bosonic Fock operators â and â † :
√ √
â |n = n |n − 1 , â † |n = n + 1 |n + 1 , ∀n ∈ {0,1,2, . . .}. (5.3)
These operators satisfy bosonic commutation relations [a,a † ] = 1 and together
define the number operator n̂ = â † a. Using this representation, we write the Hamil-
tonian and the flux and the charge operators as
⎧ 
  ⎨ φ̂ = h̄Z
(â + â † ),
1 2
Ĥ = h̄ω â â + 2 = h̄ω n̂ +
† 1
, with  (5.4)
2 ⎩ q̂ = h̄ i(â † − â).
2Z

The spectrum of eigenenergies of Ĥ is discrete and numerable, En = h̄ω(n + 12 ),

and each energy level corresponds to a different number of photons n, as anticipated.
The diagonalization gives expressions for the flux and charge operators in the Fock
space, but also for voltage and intensity:3

d i h̄Z † q̂
V̂ = φ̂ = − [φ̂, Ĥ ] = ω i(â − â) = , (5.5)
dt h̄ 2 C

d i h̄ φ̂
Iˆ = q̂ = − [q̂, Ĥ ] = −ω (â + â † ) = − .
dt h̄ 2Z L
√
In these formulas, the impedance Z = L/C plays the role of√ an oscillator
wavepacket size. This parameter defines a unit of flux φvac = h̄Z, and the
“strength” of quantum fluctuations of observables in the vacuum state |0. Due
to the canonical commutation relations [φ̂, q̂] = i h̄, those fluctuations are also
constrained by the Heisenberg uncertainty principle (2.12):
1
V̂ Iˆ ∝ φq ≥ h̄. (5.6)
2
In other words, the more precisely we measure φ̂ (or q̂), the more uncertain is our
estimate of complementary observable q̂ (φ̂).

1 See, e.g., chapter 6 from Ballentine (1998).

2 Even though we use the same symbol as in Section 4.9, here n represents the quanta of excitation, and is only
indirectly related to the number of Cooper pairs.
3 These relations, derived with the Heisenberg equation, hold also in the Schrödinger picture!

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.1 LC Resonator 67

5.1.3 Phase Space Dynamics

We can solve the Schrödinger equation for a time-independent Hamiltonian Ĥ as
|ψ(t) = e−i Ĥ t/h̄ |ψ(0). In the case of an LC circuit, we write

|ψ(t) = exp(−iEn t/h̄) n|ψ(0) × |n . (5.7)
n

This method no longer works when the Hamiltonian depends on time, such as when
the resonator is connected to an oscillating potential V (t) in (5.2). This problem is
best analyzed in the Heisenberg picture (2.13).
The Hamiltonian for a resonator coupled to an oscillating potential V (t) reads
Ĥ = h̄ωa † a + i h̄(t)(a † − a) + Eglobal (t). (5.8)
√
The linear displacement h̄(t) = h̄/2ZCg V (t)/C contains the external poten-
tial. The energy shift Eglobal (t) ∝ V (t)2 will be ignored, since it commutes with all
operators. The Heisenberg equations for the canonical and Fock operators are linear,
first-order differential equations:

d
φ̂ = 1
[ q̂ − Cg V (t)], dâ
dt C
⇔ = −iωa + (t). (5.9)
d
dt
q̂ = − L φ̂,
1 dt
The Fock operators experience a phase and a complex displacement α(t):
 t
−iωt
â(t) = e â(0) + e−iω(t−τ ) (τ )dτ = e−iωt â(0) + α(t), (5.10)
0
and provide the solution for flux and charge:
 
h̄Z +iωt † h̄
φ̂(t) = (e a + H.c.) + φext (t), q̂(t) = (ie+iωt a † + H.c) + qext (t),
2 2Z
√ (5.11)
with α = [φext (t) + iZqext (t)]/ h̄Z.
In absence of external drive, the solutions in the phase space (φ̂, q̂) are elliptical
orbits with angular frequency ω:
φ̂(t) = cos(ωt)φ̂(0) + sin(ωt)Z q̂(0), (5.12)
1
q̂(t) = cos(ωt)q̂(0) + sin(ωt) φ̂(0). (5.13)
Z
As in the classical theory, the impedance is the ratio between the positive (or nega-
tive) frequency components of the flux and charge, or voltage and current intensity.
This can be seen from a Fourier transform, or directly, taking the positive frequency
terms from (5.11):
√
φ̃(+ω) h̄Z/2 a † (0) Ṽ (+ω)
=√ = = −iZ. (5.14)
q̃(+ω) h̄/2Z a (0)
†
I˜(+ω)

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 68 Microwave Photons

Here and throughout this book, we often write the expected value of a non-
Hermitian operator a † . Even though a † is not an observable, it can be decomposed
as the addition of two measurable quantities:
 
2 2Z
a † (t) = φ̂ − i q̂. (5.15)
h̄Z h̄
Therefore, the expectation value of a † , a, a † a, or even higher moments a 2 a † , etc.,
can be inferred from the composition of two separate measurements. This will be
relevant in later sections, where we study the amplification of light and the simul-
taneous measurement of quadratures (Section 5.5.7).

5.1.4 Are There Real Photons?

Chapter 2 mentioned direct evidences of light quantization in the photelectric effect.
In these experiments, light carries enough energy – 0.1−10 eV in the infrarred to UV
range – that can excite a short-lived but measurable electric current. The discreteness
of the detection events and the distribution of energies in the generated currents are
evidences that light is absorbed by the material as individual quanta of energy, the
photons.
Unfortunately, we cannot produce similar evidence using microwave photons. In
a typical frequency range of 1–10 GHz, light particles carry a very small amount
of energy, 4−40 μeV, insufficient to trigger a conventional photodetector.4 We must
therefore seek other ways to explore the quantization of light and of observables in
superconducting resonators.
One approach is to stick to linear circuits and devices, extending them with ampli-
fiers and voltage-current measurement devices (cf. Section 5.5.7). Such a setup
allows us to measure the quadratures of the field, including the energy operator
itself â † â. Unfortunately, these are noisy processes that do not yet discriminate
individual photons and only provide good ensemble estimates of the field operators.
This is enough for many tasks, including state tomography, entanglement witnesses,
or squeezing, but not for a single-shot readout of photons.
The other approach is to combine resonators with nonlinear devices. As explained
in Chapter 6, there are superconducting circuits with discrete, highly anharmonic
spectra, called qubits. When paired with a resonator, these saturable devices can
only emit or absorb a single quantum of energy at a given frequency. Using qubits,
we can create photons one by one, engineering complex superpositions of one-,
two-, and more-photon states predictably. We can also do accurate tomography

4 Note the word conventional. A microwave photon can still trigger other elements, such as superconducting
circuits Romero et al. (2009a,b), as a basis for microwave photodetection.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.2 Transmission Lines or Waveguides 69

of the number of photons and reconstruct the wavefunction of the resonator or in

any other bosonic mode. This kind of protocol, introduced later in Chapter 7, is
also single-shot: Each experiment provides a measurement outcome that does
not destroy the photon state and can be use for feedback and control of the
photon’s state.

5.2 Transmission Lines or Waveguides

The zero-mode resonators are a theorist’s dream: They have a simple design, they
can be described by a one pair of variables (φ̂, q̂), and the two parameters L and
C can be accurately engineered to produce the desired resonance ωLC . However,
zero-mode resonators are rarely used in quantum experiments. They are necessarily
small, so that it becomes difficult to pack the antennas needed to control them,
together with other circuits such as superconducting qubits. The limitation is not
so much on the fabrication part, but on the operation part. Increasing the density of
components leads to errors in the form of undesired crosstalks between circuits, bad
isolation, etc.
Photons can also be injected and stored in other linear circuits that are generically
called waveguides. Like the strings of a guitar, they accommodate many different
photonic modes, in the form of stationary waves with distinct wavelengths and
frequencies. In an extreme limit, the work by Sundaresan et al. (2015) resolved
spectroscopically 75 photonic modes within the same waveguide. However, when
so many modes participate in the dynamics, it is preferable to disregard the notion
of standing waves and work with propagating photons, as discussed in Chapter 7.
There are various elements that transport photons without distortion or signifi-
cant losses. The simplest one is the coaxial cable that brings the TV signal to our
receivers. These cables have a conducting core surrounded by a grounded metal-
lic shell. The electromagnetic waves move longitudinally in the free space among
conductors, dragging with them charge and current excitations on both surfaces.
The equivalent of a coaxial cable in a 2D superconducting microchip would be
either a stripline or a coplanar waveguide. The stripline consists of two lines of
superconducting material sitting above and below the substrate, and one of them
connected to ground, as seen in Figure 5.2b (bottom). The field moves in between
both lines, confined in the substrate. The coplanar waveguide is an alternative and
more popular design where two ground planes surround a thin superconducting line.
Now the electromagnetic field runs also outside the substrate, making it a bit more
robust under perturbations of the material.
All 2D superconducting waveguides are typically modeled with the same
lumped-element circuit, discussed in Section 4.5. The circuit includes a capacitive
term for the interaction between the core conductor and the ground planes, and an

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 70 Microwave Photons

Figure 5.2 (a) A coplanar waveguide. The left and right superconducting planes
are grounded to a fixed potential while the charge and electromagnetic waves run
through the middle section. (b) Other planar waveguides, such as the stripline
(top), differ from the coplanar waveguide (bottom) on the position of ground
planes relative to the main conductor. (c) Equivalent circuit for the waveguide (see
Section 4.5).

inductive energy describing the “inertia” of the conductors to change the current.
Introducing the capacitance and inductance per unit length, c(x) and l(x), we
derived two alternative models: a discrete theory (4.40) and a continuous field
theory (4.42). We will now diagonalize and study these models in three different
scenarios with different boundary conditions and line terminations.

5.2.1 Periodic Boundary Conditions

Let us take the discrete model for a transmission line (4.40):
 1 2  1 2 
N N
H = q̂i + φ̂i+1 − φ̂i+1 φ̂i . (5.16)
i=1
2C i=1
L

The line has length d and is divided into N segments with the same capacitance and
inductance, C = cx and L = lx. We assume that the line closes onto itself, so
that φ̂N+1 = φ̂1 .
The second term in this model is a quadratic form, φ T Bφ/2L, with a nonnegative,
symmetric matrix B (cf. Section A.1.1). Following the mathematical tools from
Appendix A.2, we diagonalize the matrix B and obtain N normal modes to write
the Hamiltonian as a sum of N independent oscillators:
  
† 1
Ĥ = h̄ωk b̂k b̂k + . (5.17)
k
2

Each collective degree of freedom is associated to a pair of Fock operators

[b̂k, b̂k† ] = δkk , with a frequency and a momentum:
1  2π  π π 
ωk = √ 2 − 2 cos(kx) , kn = ×n∈ − , . (5.18)
CL d x x

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.2 Transmission Lines or Waveguides 71

There are a total of N modes, which we identify with left- and right-moving waves,
k < 0 and k > 0, labeled by the integer numbers n ∈ (−N/2,N/2]. We can recon-
struct the flux and charge fields along the waveguide’s positions xm = x × m as

 h̄  
φ̂(xm ) = uk (x)b̂k + uk (x)∗ b̂k† , (5.19)
k
2cωk

 h̄cωk  
q̂(xm ) = x iuk (x)∗ b̂k† − iuk (x)b̂k ,
k
2
√
using the normalized eigenmode wavefunctions ux (x) = eikxm / d.
For small momenta, the dispersion relation is approximately linear:
1 1
ωk  v|k|, v=√ , |k|  . (5.20)
c0 l0 x

The dispersion relation, the speed of light v, and the operators φ̂ and q̂/x are all
defined in terms of intensive quantities and scale-independent normal modes. As
anticipated in Section 4.5, this facilitates taking the limit x → 0, where flux and
charge are replaced with continuous fields [φ̂(x), ρ̂(y)] = i h̄δ(x − y). The new
Hamiltonian
 d  2 
1 1 
H = dx ρ̂(x) +
2
∂x φ̂(x) (5.21)
0 2c 2l
corresponds to a massless Klein–Gordon theory. These fields satisfy a wave equa-
tion with periodic boundary conditions φ̂(x + d,t) = φ̂(x,t):
 
1
∂t −
2
∂ φ̂(x,t) = 0,
2
(5.22)
c(x)l(x) x

whose stationary solutions (ωk2 − ∂x2 )uk (x) = 0 are the continuous version of the
discretized eigenmodes that we found before.

5.2.2 λ/2 and λ/2 Microwave Cavities

Transmission lines are often cut to form a one-dimensional open cable of finite
length d, coupled to the environment through two capacitors C1 and C2 . As shown
in Figure 5.1b, photons are confined between those two capacitors, which act as
partially reflective mirrors and create the microwave equivalent of a Fabry–Perot
cavity.
Instead of working with the discrete model, it is more convenient to start from the
continuous theory (5.21) and change the boundary conditions. Initially we neglect
the coupling to the outer waveguides C1,2 = 0 and focus on the eigenmodes of an

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 72 Microwave Photons

(a) (b)

k xd

Figure 5.3 (a) Dispersion relations for a transmission line. The dashed line plots
the relations for x = 1,1/2,1/4, and the solid line plots the limiting relation for
x = 0. When the transmission line has a finite length d = 6, the momenta and
frequencies are discrete (circles). (b) First flux modes for an open trasmission line
of finite length.

isolated line. In this limit, no charge can escape the line and the currents must vanish
at the boundaries. The zero-current property translates into Neumann-type boundary
conditions for the flux ∂x φ|x=0,d = 0. The Hamiltonian (5.21) with these boundary
conditions is diagonalized by an infinite set of standing waves:
  πn 
2
u (x) =
(n)
cos x , n ∈ N = {1,2,3, . . .} (5.23)
d d

with quasimomentum kn = πn/d and wavelength λn = d/2n. Figure 5.3a shows

the first three eigenmodes, all with zero derivative at the boundaries. Note how the
length of the resonator is half the wavelength of the fundamental mode d = λ1 /2.
Due to the boundary conditions, the edges of a λ/2 resonator have no current, but
reach a maximum of the charge on all modes:

 h̄c0 ω
ρ(x) = (n)
u (x) i bn† − bn . (5.24)
n
2

As show in Figure 5.1c, we can use this fact to couple capacitively the resonator
to a semi-infinite transmission line. The coupling, whose strength depends on the
capacitors C1 or C2 used, is our means to control the resonator, injecting photons or
measuring its internal state (see Section 5.5).
If we only need one control access, we may want to connect the other boundary to
a ground plane, as shown in Figure 5.4b. This connection replaces one of the Neu-
mann boundary conditions with a fixed value for the voltage at that point φ̂(0) = 0,

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.2 Transmission Lines or Waveguides 73

(a)

(b)

(c)

Figure 5.4 First φ0 (x) and second φ1 (x) eigenmodes of (a) λ/2 and (b) λ/4 trans-
mission line resonators. (c) A transmission line resonator with a SQUID becomes
an adjustable resonator. The external flux  affects the SQUID’s inductance, the
lowest energy eigenmodes, and the effective length of the resonator d(). Note that
the drawing is not on scale: A typical SQUID is much smaller than the waveguide.

while the other condition remains the same ∂x φ̂(d) = 0. The new set of modes that
satisfies these boundary conditions are the sinusoidal functions
  
2 π (2n − 1)
u =
(n)
sin x . (5.25)
d 2d
The length of the waveguide is a quarter of the wavelength of the fundamental mode,
d = λ1 /4. Such λ/4 resonators are interesting because close to the edge x = 0 the
current ∂x φ/ l is maximal. This is beneficial for coupling inductive elements to the
resonator and to place inductive elements that tune the properties of the cavity, as
we will now see.

5.2.3 Tunable Cavities

In theory, a microwave resonator can be tuned by changing the capacitance or the
inductance of the elements that form it. In practice, adjustable capacitors require
mobile parts to change their geometry,5 which is impractical to operate within a
fridge and may introduce noise and heat. Tuning the inductance, on the other hand,
is a simpler task.

5 Does anyone remember old radio receivers with a frequency selection wheel that adjusted a big capacitor? In
this device, the tuning wheel changed the opposing surface between two metallic plates, adjusting the
capacitance.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 74 Microwave Photons

We have already seen that a dc-SQUID recreates a tunable Josephson junction,

where the inductive energy depends on the external magnetic field EJ () =
2EJ (0) cos(/ϕ0 ). We can shunt the dc-SQUID with a big enough capacitor to
create a tunable, zero-mode LC resonator. The only requirement is to operate the
circuit in a regime of small flux, such that the SQUID’s inductive potential remains
approximately harmonic:

φ̂ 1 2 ϕ02
− EJ () cos  φ̂ , with LJ () = . (5.26)
ϕ0 2LJ EJ ()

If we do not want to use a zero-mode resonator, we can place the dc-SQUID inside
a transmission line, as shown in Figure 5.4c. When the SQUID is much smaller than
the wavelength of the photons that we excite, we can treat it as a local perturbation
in the Klein–Gordon model (5.21):

 d
1
Ĥind  φ̂(x)2 δ(x − x0 )dx + O(φ 4 ). (5.27)
0 2LJ ()

This term affects the eigenfrequencies and eigenmodes of the guide. In particular,
the normal modes become discontinuous around the SQUID: The flux jumps by an
amount δφ ∼ ∂x φ × δx, which would correspond to extending the waveguide by an
additional length δx = LJ / l0 – a ficticious length that remains “folded” inside the
SQUID.
The inductance of a junction can be much larger than that of typical coplanar
waveguide. The change in the length δx/λ can be comparably large, affecting the
location of the nodes and antinodes of the resonator. If we connect other circuits –
superconducting qubits, resonators, etc. – to those points, we can use a dc-SQUID
to bring them in and out of resonance, coupling and decoupling the resonator from
them (Forn-Díaz et al., 2017).
The change in resonator length can also be much faster than the time d/v for a
photon traverse the waveguide (Sandberg et al., 2008). In this regime of driving,
it is as if the “mirrors” of the resonator moved at close to the speed of light. Such
conditions simulate a gedanken experiment known as the dynamical Casimir effect,
which results in the production of entangled photon pairs out of the vacuum. This
experiment would be impractical with real mirrors and cavities, but has been repeat-
edly demonstrated by Wilson et al. (2011) and later works.
Finally, in Section 5.5.6 we will see that a periodic modulation of the inductance
can be used to achieve an effect known as parametric amplification. When we send
a quantum signal to this resonator, it is reflected back with one or two quadratures

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.3 Three-Dimensional Cavities and Waveguides 75

that are enlarged or shrunk. Quantum amplification is a very strong and robust
mechanism, instrumental for verifying the quantum properties of a photonic field
(see Section 5.5.7).

5.3 Three-Dimensional Cavities and Waveguides

Our description of microwave resonators would not be complete without a discus-
sion of three-dimensional resonators. Unlike transmission lines and LC resonators,
which are built on a two-dimensional chip, a 3D cavity is an actual cavity inside
block of metal. As shown in Figure 5.1c, 3D cavities are fabricated into two or
more blocks that join together to fully enclose a three-dimensional vacuum space.
The blocks are normally made of aerospace-grade aluminum, but larger cavities
and waveguides have been constructed using plain copper. As the transmission line,
the 3D cavity is an extended object with a complicated mode structure, where the
electromagnetic field propagates in free space, supported by charge and current
excitations on the inner surface of the “tube.” This produces multiple eigenfre-
quences and quantized bosonic excitations that we can group together in a single

Hamiltonian H = n ωn b̂n† b̂n , as we did for the waveguides.
Three-dimensional cavities are no newcomers to the quantum information world.
They were used in the pioneering experiments by the Nobel Prize winner Serge
Haroche (Raimond et al., 2001) because they can trap photons for very long times.
Three-dimensional cavities are usually combined with other 3D circuits. Supercon-
ducting qubits, antennas, flux injectors, coaxial waveguides, etc., are all components
that can be fabricated independenly and later inserted at different points in the cavity.
The result is a kind of superconducting Lego structure that can be adjusted and
reused for different purposes over time. This is an interesting paradigm shift from
traditional 2D circuits, which can only be fabricated for a single task, with all other
circuits permanently attached. Moreover, it must be remarked that 3D waveguides
and coaxial cables can transport the photons along relatively long distances. As of
this writing, the record was set by the ETHZ group, who extended the setup from
Magnard et al. (2020) to connect two separate cryostats by waveguides 30 m long,
operating at tens of millikelvins.
Three-dimensional cavities pose a minor theoretical inconvenience. Simulating
the behavior of microwaves, reproducing the resonator mode structure and frequen-
cies, or predicting the interaction with other embedded circuits becomes a great
numerical challenge. We have to use finite-element methods to solve Maxwell’s
equations, and apply a sophisticated framework known as black-box circuit quanti-
zation (Nigg et al., 2012) to reverse-engineer a lumped-element model that describes
the cavity with all the embedded components. In this case, the equivalent circuit is
really a mathematical convenience: We cannot ascribe the effective capacitances,

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 76 Microwave Photons

inductances, and nonlinearities to any particular region of the superconductor – they

are just models for the dynamics of the combined system.

5.4 Photon States

Microwave cavities and resonators are used to explore quantum superpositions,
macroscopic quantum states, entanglement, or the statistical properties in quantum
measurements. The superconducting circuit literature is rich in examples of cre-
ation, manipulation, and detection of photonic quantum states. Some great works
are the methods by Hofheinz et al. (2009) to engineer arbitrary superpositions
states of up to 15 photons, the experiments by Vlastakis et al. (2013) demonstrating
long-lived Schrödinger cats with up to 100 photons, or the generation of two-mode
squeezing by Mallet et al. (2011). In this section, we discuss families of quantum
states that are often used in the literature. We close the section with a mathematical
representation of quantum states, the Wigner function, that is both experimentally
and theoretically convenient to investigate the photonic field.

5.4.1 Fock States

The Fock states are the eigenstates of the number operator. They are obtained from
the vacuum |0, climbing upward in the number of photons with the creation oper-
ator |n = √1n! a † n |0. Fock states can be expanded in the eigenbasis of the dimen-
sionless quadratures:
1 φ̂ i q̂
x̂ = √ (â + â † ) = √ , p̂ = √ (â † − â) = √ . (5.28)
2 h̄Z 2 h̄/Z
For instance, in terms of position eigenstates x̂ |x0  = x0 |x0 , the Fock states are
Gaussians modulated by Hermite polynomials6 Hn (x):
1 −x 2 /2
x0 |n =  √ Hn (x)e . (5.29)
2n n! π
In particular, the vacuum state |0 has a purely Gaussian wavefunction.

The Fock basis is a useful representation for pure |ψ = n=0 ψn |n and for

mixed states ρ = nm ρnm |nm|. In these expansions, |ψn |2 = | n|ψ |2 and
ρnn = n|ρ|n are the probabilities of finding n photons in the resonator. These
expansions are often truncated at a cutoff Nmax consistent with the maximum energy
injected in the resonator. This simplifies both the experimental characterization of

6 We use the definition H (x) = 1, H (x) = (x − ∂ )H

0 n x n−1 (x).

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.4 Photon States 77

quantum states – we ignore high-order moments that are hard to measure – as well
as the theoretical simulations of practical circuits.
Fock states are simple and useful, but difficult to prepare. If we drive a linear
resonator with an external field, the coupling (t)(â † + â) spreads the wavefunction
over all Fock states. To create a state of one photon, we need a single photon source:
a saturable quantum device that is always excited to the same state, and which
always decays emitting just one photon. The word “saturable” means that the device
can take a finite amount of energy: typically, it will jump from some ground state |g
to the same excited state |e, and it will stop there. When we stop driving the source,
the photon source decays back to |g, emitting a single photon. These nonlinear two-
level circuits are the superconducting qubits that we introduce in Chapter 6.
Using qubits, we can play some magic tricks. For instance, we can prepare the
qubit in a superposition of ground and excited state α |0 + β |1, so that when
it exchanges energy with the microwave resonator it implements the operation
(α + β â † ). Multiple iterations of this trick lead to arbitrary superpositions of Fock
states, as demonstrated by Hofheinz et al. (2009). Moreover, the same qubit can also
be used as photon counter, to reconstruct the full wavefunction of the microwave
resonator – see Section 7.3.2 and works by Schuster et al. (2007) and Hofheinz
et al. (2009).

5.4.2 Thermal States

A saying goes that real experiments do not have pure states: Since we never have
100% control of all experimental conditions, states in the lab are always mixed
states – classical superpositions of experiments with slightly varying conditions.
This is in particular true when we consider quantum circuits at thermal equilibrium
with their environment. These circuits are modeled with a canonical ensemble of
quantum states |n, excited with Boltzmann weights exp(−En /kB T ). Applied to
the single-mode resonator, this creates a Bose–Einstein distribution of photons:

 ∞
e−βH
ρ= = e−β h̄ωn (1 − e−β h̄ω ) |nn| . (5.30)
Z(β) n=0

This ensemble gives a nonzero probability of finding photons in the resonator pn>0
and a nonzero average occupation n̄:

pn>0 = cρnn = 1 − ρ00 = e−h̄ω/kB T , (5.31)
n>0
1
n̄ = a † a = . (5.32)
eh̄ω/kB T − 1

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 78 Microwave Photons

Recalling that h̄ × 2π × 20 GHz  kB × 1 K, the average population of photons

ranges between 1.5 photons for a 1 GHz resonator at 100 mK, to 10−21 photons for
a 10 GHz resonator at 10 mK – i.e., from something that may hinder the fidelity of
quantum operations to a negligible correction.
For some applications in quantum simulation, it might be desirable to prepare
nonequilibrium thermal states at a temperature higher than that of the cryostat.
This can be done by pumping the oscillator with narrow-band white noise; this
results in a master equation at an artificially high temperature (cf. Section 5.5.4)
that equilibrates to the desired state, without “breaking” the sample or the super-
conducting state.

5.4.3 Coherent States

The coherent states are eigenstates of the annihilation operator a |α = α |α. There
is one coherent state for every complex eigenvalue α = αx + iαp ∈ C. The vacuum
is a coherent state with α = 0. The displacement operator is a unitary transforma-
tion D(α)† = D(−α) = D(α)−1 that shifts the Fock operators:
D(α)† âD(α) = â + α, D(α)† â † D(α) = â † + α ∗ . (5.33)
Hence, all coherent states result from displacements of the vacuum:
∗
 αn
|α = D(α) |0 = eαâ −α â |0 = √ e−|α| /2 |n ,
† 2
(5.34)
n n!

and can be uniquely identified by their displacement α = αre + iαim = √1

2
α|x̂|α +
√i α|p̂|α, which is their “position” in the phase space defined by (x̂ , p̂). As
2
they evolve in time e−iH t/h̄ |α ∝ |e−iωt α, coherent states describe elliptical orbits
in phase space, similar to a classical oscillator. Coherent states have Gaussian
wavefunctions in position, momentum, or any other direction in phase space. The
width of the Gaussian
1 √
x|α = 1/4 e−(x− 2αre ) /2+iαim x
2
(5.35)
π
is the same along any direction and saturates the Heisenberg uncertainty relation,
that is, x̂ = p̂ = √12 and φ̂q̂ = 12 h̄. For this reason, coherent states are also
called minimum uncertainty states.
Resonators are prepared in coherent states by feeding them with coherent
microwave signals, via the coupling term in (5.8) (see Section 5.5.1). Unlike
visible/IR/UV coherent fields, which require lasers, there are many ways to generate
coherent microwaves: from high-power masers, magnetrons, and tubes, to low-
power solid-state electronic devices. A typical coherent microwave source is formed

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.4 Photon States 79

by a chain of electronic devices. An ultrastable crystal oscillator – referenced to an

atomic clock for phase and frequency stability – creates a signal at a fixed frequency
around hundreds of MHz. The output signal is frequency-multiplied in a nonlinear
circuit, doubling its frequency until it reaches the desired range of GHz. The last
fine-tuning is done with additional (digital) electronics that shift the frequency to
the desired value, in resonance with the LC circuit.

5.4.4 Schrödinger Cat States

Quantum mechanics predicts the existence of superpositions of “macroscopically”
distinguishable states. The textbook example is Schrödinger’s cat experiment: a live
cat is placed inside a box with a poisoning device that is activated by a decaying
radioactive particle. At any given point in time, the state of the cat in the box must
be described as a quantum superposition of being alive (with the radioactive atom
still active), and being dead (with a decayed atom and free poison). Schrödinger’s
gedanken experiment challenges the limits of quantum mechanics, setting a bound-
ary between a somewhat loosely defined “macroscopic” world, where we do not
experience quantum superpositions, and the realm of coherent quantum mechanics.
Scientists have accepted Schrödinger’s challenge, exploring how to create large
quantum superpositions using microwave resonators.7
Photonic Schrödinger cats are non-Gaussian states and require nonlinear opera-
tions to be created. As explained in Chapter 7 and in the work by Vlastakis et al.
(2013), a qubit with a dispersive coupling can push a microwave resonator into a
superposition of two coherent states with opposite displacements:
1
|cat (θ ) = √ (|α + eiθ |−α). (5.36)
2
In state-of-the-art experiments, the cat has as many as |α|2 ∼ 100 photons. At this
point, the two states |±α can be macroscopically distinguished through quadrature
measurement – i.e., the voltage or current leaking from the resonator. This distin-
guishability also makes the state all the more fragile, decreasing its lifetime linearly
with the number of photons.
Interestingly, it is possible to extend the lifetime of a photonic Schrödinger
cat using an error-correcting code that detects if the state has lost any pho-
ton. When this happens, the cat state is not immediately destroyed, but sim-
ply changes phase, jumping around in the Hilbert space a |cat (θ ) ∝ √12 (|α −
eiθ |−α) = |cat (θ + π ). This new state is orthogonal to the original

7 Also on other circuits, such as superconducting current loops (van der Wal et al., 2000), and on trapped ions,
optical lattices, etc.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 80 Microwave Photons

state: cat (θ)|cat (θ + π ) ∝ exp(−2|α|2 ). However, using qubits we can detect

those tiny phase changes, either correcting the state or bookkeeping the accumulated
phase θn and correcting later measurements. This technique can stabilize a cat state
to a time that is around the resonator lifetime (Ofek et al., 2016). Further extensions
would need to account for dephasing – difference in the relative phases of |α and
|−α – using a larger space of states and error-detection strategies.

5.4.5 Single-, Two-, and Multimode Squeezed States

Squeezed states are Gaussian states where the uncertainty of one or more quadra-
tures is reduced at the expense of increasing the uncertainty in the conjugate
quadratures. There are two important motivations for working with squeezed states:
seeking a metrological advantage and creating entanglement. The metrological
advantage appears already in the single-mode squeezed state. The reduction in the
uncertainty of one quadrature, such as x̂, means that an experimental estimate of
x̂ will need a lower number of measurements M to achieve the same precision –
see (2.28) in our discussion of unbiased estimators.
Single-mode squeezing is a unitary transformation, parameterized by the squeez-
ing strength r and angle θ :
  
1 −i2θ 2
Usq-1 (r) = exp re i2θ † 2
â − re â . (5.37)
2
In the Heisenberg picture,8 this Bogoliubov transformation combines annihilation
and creation operators:
†
â(r) := Usq-1 (r)â(0)Usq-1 (r) = cosh(r)â(0) − sinh(r)â † (0)ei2θ . (5.38)

Introducing the rotated quadrature x̂φ = √1 (â † eiφ + âe−iφ ), we find two observables
2
with opposite behavior
x̂θ (r) = e−r x̂θ (0), p̂θ (r) := x̂θ+π /2 = e+r p̂θ (0). (5.39)
The operator p̂θ and its uncertainty p̂θ are both amplified with a gain factor
√
G = e+r , while the operator √ x̂θ (r) and its variance x̂θ (r) are both contracted
−r
by the inverse factor e = 1/ G.
Experimentally, this means that measuring x̂θ (0) on the squeezed vacuum
Usq-1 (r) |0 produces
√ results with an uncertainty below the standard quantum limit
x̂θ = e−r / 2. Squeezing is measured in decibels rdB := 10 × log10 (Onew /Oold ).

8 Instead of studying how the states are transformed, we study how the observables that we want to measure are
transformed.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.4 Photon States 81

A value rdB ∼ 3 means a 50% reduction in uncertainty e−r ∼ 0.5. Not too long ago,
such values were considered the state of the art in quantum optics – for instance,
the Laser Interferometer Gravitational-Wave Observatory (LIGO) uses between 2.7
and 3.2 dB of squeezing in their interferometer. These days, much larger values of
10−12 dB (0.1–0.06 reductions) are available in the superconducting lab, thanks to
the parametric amplifiers described in Section 5.5.6.
The second motivation for engineering squeezed states is to create entanglement
and correlations between two or more bosonic modes (Laurat et al., 2005). In the
typical two-mode squeezed state, we compress a quadrature that is a linear com-
bination of two oscillator modes b1 + b2 , using the general two-mode squeezing
Gaussian operation:
 
Usq-2 (r) = exp rei2θ bj† bi† − r −i2θ bi bj . (5.40)

This operator mixes the annihilation and creation of different modes, as in

b̂1 (r) = cosh(r)b̂1 (0) − sinh(r)e2iθ b̂2† (0), (5.41)

−2iθ
b̂2† (r) = cosh(r)b̂2† (0) − sinh(r)e b̂1 (0).

Applying this transformation onto a thermal state, we achieve squeezing and ampli-
fication of the joint quadratures x̂θ± = 12 (b̂1† ± b̂2† )e−iθ + H.c. and p̂θ± = 2i (b̂1† ± b̂2† ) +
H.c. similar to before:

x̂θ± (r) = e∓r x̂θ± (0), p̂θ± (r) = e±r x̂θ± (0). (5.42)

We will discuss the creation of two-mode squeezing with linear amplifiers and these
gain/contraction relations in Section 5.5.6.
The two-mode squeezed state is an entangled state because it cannot be written
as product of two independent states for each of the modes ρsq-2 = ρ1 ⊗ ρ2 . This is
qualitatively appreciated when we write the Schmidt decomposition of the state in
√  n/2
the Fock basis |0r  := Usq-2 (r) |0 = 1 − λr n λr |n,n, with λr = tanh(r)2 .
We can also compute the reduced density matrix that results from tracing out or

ignoring one of the modes ρ1 = tr2 (|0r 0r |) = (1 − λr ) n λnr |nn|. The von
Neumann entropy of this matrix is a measure of the entanglement between the
modes S = −tr(S log(S)) ∼ 2r and diverges with the amount of squeezing r.
It is possible to create entangled states of multiple degrees of freedom. In particu-
lar, the transmission line is an example of multimode squeezed state. From (5.19), we
see that the fields {b̂k, b̂k† } are macroscopic superpositions of quadratures at different
positions. Thus, while the vacuum state may seem a rather trivial state with no
photons b̂k | = 0, it is in fact a highly correlated state from the point of view of

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 82 Microwave Photons

the field operators along the line, {φ̂(x), q̂(x)}. If we try to estimate the uncertainty
of the local observables, we will find strong divergences:
 h̄eik(x−y)
φ̂(x)φ̂(y) − φ̂(x) φ̂(y) = bk bk†  , (5.43)
k
2c0 xω k

both in the infrared and ultraviolet limit.9 This manifestation of entanglement in the
1D field theory is but one example of other quantum field theory concepts that can
be explored using superconducting circuits – particle localization, propagation of
entanglement and causality, vacuum energy and Casimir effect, etc.

5.4.6 Wigner Functions and Gaussian States

Coherent, thermal, and squeezed states are representatives of the Gaussian state
family. Defining this family requires a tiny bit of quantum information theory.10
Let us assume a system of N bosonic modes, described with dimensionless posi-
tion and momenta (5.28). We group all operators into a vector R̂ = (x̂1, . . . , x̂N ,
p̂1, . . . , p̂N )T , introducing the commutator matrix iij = [R̂i , R̂j ] and the displace-
ment operator:
     0 1 x 
0
D(R0 ) = exp i R̂ R0 = exp i x̂ , p̂
T T T
. (5.44)
−1 0 p0
Note that for a single mode
1
† −α ∗ â
D(R0 ) = ei(p0 x̂−x0 p̂) = eαâ
with α = √ (x0 + ip0 )., (5.45)
2
The Wigner function is the Fourier transform of the characteristic function
χ (R) = tr [ρD(R)]:

1 T
Wρ (R) = 2N
eiY R tr [ρD(R)] d2N y. (5.46)
(2π) R2N
The marginals of the Wigner function give the probability distributions of the
position and momentum operators, x|ρ|x = Wρ (x,p)dN p, and p|ρ|p =
N
Wρ (x,p)d x. We can also use Wρ to compute any expected values involving
the quadratures:

F (x̂, p̂)ρ = tr[F (x̂, p̂)ρ] = Wρ (x,p)F (x,p)dN xdN p. (5.47)
R2N

9 Infrared because the denominator approaches zero for k = 0, and ultraviolet because the fraction is
proportional to 1/n for |n| = 1, . . . ,N/2, and the sum diverges logarithmically with the discretization size.
10 Notation follows Adesso et al. (2014) and Olivares (2012).

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.4 Photon States 83

If Wρ ≥ 0, the Wigner function represents a true probability distribution. In this

case, the experimental measurement outcomes can be classically simulated by sam-
pling this distribution. However, Wρ in general is just a quasiprobability distribu-
tion, because it can reach negative values. The states for which this happens are
usually referred to as nonclassical states.
A Gaussian state is one whose Wigner function is a Gaussian:
1  T −1

W (R) = √ exp −(R − r)  (R − r) . (5.48)
π N det()

The center of the Gaussian r := R̂i ρ and its covariance matrix 

ij := {R̂i − ri , R̂j − rj }ρ = R̂i R̂j + R̂j R̂i ρ − 2 R̂u  R̂j ρ , (5.49)

are uniquely determined by the first and second moments of the quadratures.
As we anticipated, all single- and multimode coherent states |α are Gaussian
states.
√ Their Wigner functions look like displaced vacua: They are centered on
d = 2(αre,αim ) and – because they are minimal uncertainty states – they have the
same width on all directions. Figure 5.5a and b illustrate the Wigner function for a
vacuum state and for a coherent state at α = 5/2.
A squeezed state is a Gaussian state with unequal widths along two or more
mutually orthogonal directions. Figure 5.5e shows the Wigner function of a single-
mode squeezed vacuum state – state U (r) |0 with θ = 0 in (5.37). As expected,
the quadrature x̂ is reduced by a factor e−r = 1/2, while the canonically conjugate
momentum is enlarged by the inverse factor. The result is a Gaussian √ with elliptical
equiprobability contours, and a minimum width x̂ = 1/2 2 below the Heisen-
berg limit.
The family of Gaussian states includes also thermal states. Moreover, all Gaus-
sian states can be written as a finite temperature state of one quadratic model,
ρ = e−βHeff /tr e−β Ĥeff with Ĥeff = 12 Hij (R̂i − ri )(R̂j − rj ). The process of determin-
ing the model Ĥeff and the temperature β = 1/kB T is equivalent to finding the
simplectic transformation that diagonalizes .
Gaussian operations are those that preserve the Gaussian nature of a state. This
includes all linear transformations between modes R = U † RU = OR generated
by quadratic Hamiltonians U = exp(−iHeff t). Examples of those coherent transfor-
mations are the single-mode and two-mode squeezing operations from Section 5.4.5.
However, Gaussian states are also preserved by incoherent operations that change
the area covered by the Wigner function, enlarging the uncertainty of one or more
quadrature. We will see examples of those operations when we study the coupling
of a Gaussian model to a linear environment in Section 5.5.4.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 84 Microwave Photons

(a) (b)

(c) (d)

(e)

Figure 5.5 Wigner functions of (a) the vacuum state |0, (b) a coherent state
|α = 5/2, (c) a Fock state |1, (d) a Schrödinger cat ∝ |5/2 + |−5/2, and (e) a
squeezed state with e−r = 1/2. Note how the Wigner function for the |0 and |α
state look identical, up to a displacement. The Wigner function becomes negative
for the “nonclassical” states (c) and (d).

The Wigner function is a useful tool to reconstruct and identify bosonic quantum
states, even if they are not Gaussian. Figure 5.5c and d show the Wigner function
of a Fock state and of a Schrödinger cat (5.36). The Fock state spreads over a ring
of radius |α|2 ∼ n, while the Schrödinger cat shows the interference between to
coherent states |±α.
The Wigner function is an infinite-dimensional object that can only be approx-
imately reconstructed under realistic assumptions. The following are main recon-
struction methods:
Gaussian state verification. This method only works for Gaussian states. One mea-
sures moments of the quadratures (a † )m a n  up to a certain order n + m = Nmax > 2.
These moments are used to verify that the state is Gaussian, showing that higher-
order moments n + m > 2 are related to the first and second moments by Wick’s
theorem (Menzel et al., 2012). Once the Gaussian nature is verified, the covariance
matrix  and the mean values r can be derived from the moments with n + m ≤ 2.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 85

Parity measurements. There is a formula that relates the Wigner function to the
†
parity operator  = (−1)a a :
2 2
Wρ (α) =
tr(ρα ) = D(−α)† D(−α)ρ . (5.50)
π π
This formula enables the computation of the Wigner function at every point of phase
space, through a three step process: (i) prepare the state ρ, (ii) displace the bosonic
mode using some external microwave drive to construct ρα := D(α)† ρD(α), and
(iii) measure the parity over this new state tr(ρα ), for instance with the help of a
qubit (Hofheinz et al., 2009).
Truncated moments. For states where there is a maximum number of photons,
higher-order moments above Nmax can be vanishingly small. This allows us to recon-
struct the Wigner function from the smaller set of nonzero moments (Eichler et al.,
2011):
 (−λ∗ )m λn 1 ∗ ∗
(â † )m â n ρ e− 2 |λ| +αλ −α λ .
2
Wρ (α = x + ip) = d2 λ 2 (5.51)
n,m C
π n! m!

Radon transform. This is a method for unbiased tomography with first-order

quadrature measurements. It reconstructs W (x,p) from the marginals P (xθ ) =
dpθ W (cos(θ )xθ + sin(θ)pθ , cos(θ)pθ − sin(θ)xθ ) along various directions θ
in phase space. The statistics of P (xθ ) is recovered from binning measurements
of a rotated quadrature Q̂ω,θ ∝ xθ , such as voltage or intensity, with an angle
determined by the phase of the reference oscillator (see Section 5.5.7). The inverse
Radon transform approximates Wρ (x,p) from a set of p(xθ ) over the different
angles (Mallet et al., 2011).

5.5 Gaussian Control of Microwave Photons

We have discussed some linear superconducting circuits – microwave resonators,
transmission line resonators, and three-dimensional cavities – and the quantum
states that they support. We have suggested linear operations – displacements,
heating, amplification and squeezing, measurement – that can be formally used
to create and reconstruct those states. In this section, we explain how all those
Gaussian operations are implemented in actual experiments, together with the
Hamiltonian models that describe them.

5.5.1 Coherent Drivings and Displacement Operations

The simplest way to control an LC circuit is to connect it to a voltage source,
providing a time-dependent potential bias V (t), as in Figure 4.4. The canonical

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 86 Microwave Photons

quantization produces the linear, driven harmonic oscillator from (5.8), with the
√
external field (t) = h̄/ZCg V (t)/C . This external drive induces a displacement
transformation onto the field â → D(α)† âD(α) = â + α. If we switch from the
Heisenberg to the Schrödinger picture, the same operation interpreted as preparing
the resonator, from a vacuum state, into a coherent state with adjustable displace-
ment and phase |α = D(α) |0 (see Problems 5.4 and 5.5).
Let us study an oscillating microwave field (t) = 0 cos(ωd t + ϕ) that res-
onates with the circuit for a brief period of time. The solution of (5.10) predicts a
displacement of the resonator with two parts:
 
â(t) = e−iωt â(0) + αRWA (t) + αnon-RWA (t) . (5.52)

The rotating term oscillates in synchrony with the resonator:

 t
0 −i(ωd τ +ϕ) 0 e−iϕ ei(ω−ωd )t − 1
αRWA (t) = eiωτ e dτ = . (5.53)
0 2 2 i(ω − ωd )
Close to resonance ωd − ω → 0, it creates a linearly growing displacement
|αRWA | ∼ 0 t. The counter rotating term has a rapidly oscillating integrand:
 t
0 +i(ωd τ +ϕ) 0 eiϕ ei(ω+ωd )t − 1
αnon-RWA (t) = eiωτ e dτ = , (5.54)
0 2 2 i(ω + ωd )
which averages to a smaller displacement |αnon-RWA | ∼ |0 /(ω + ωd )|  1. The
rotating wave approximation (RWA) consists in neglecting this displacement,
assuming that the Rabi frequency is smaller than the characteristic frequen-
cies |0 |  ω + ωd and that we are interested in the long-time dynamics
t  1/(ω + ωd ).
Once we know the outcome of the Heisenberg equations, we can develop a qual-
itative argument to apply the RWA directly onto the Hamiltonian. This argument
goes as follows: Since the cavity field oscillates as â(t) ∼ e−iωt , we can neglect
any term of the form e−iωd t â(t) because such terms will lead to small corrections
O((ω + ωd )−1 ). Applied to the Hamiltonian (5.8), this produces an RWA model:
0 −i(ωd t+ϕ) †
HRWA = ωa † a + i e â − ei(ωd t+ϕ) â , (5.55)
2
which produces the displacement (5.52) with αnon-RWA = 0.
Since our ideal resonator has no losses, a resonant driving ωd = ω injects an
unlimited amount of energy, which grows quadratically in time a † a ∝ t 2 . In
Section 5.5.3, we correct this absurdity, collecting the experimental losses from
actual resonators – both intrinsic, as well as due to a coupling to the voltage source
and the transmission line – in a decay rate κ. Qualitatively, the losses are equivalent

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 87

to a complex shift of the resonator frequency ω → ω − iκ/2 (0 < κ  ω). With

this change, the resonator occupation saturates at around ∼ |0 |2 /κ 2 , a number that
can be still very large. For instance, state-of-the-art experiments use this method to
engineer displacements with hundreds of photons (Vlastakis et al., 2013) as the
starting point for creating macroscopic Schrödinger cats.

5.5.2 Coupling to an Environment

Actual resonators are connected to the outer world. We have already considered
the possibility of driving the resonator with microwaves. These signals must be
brought into the circuit and into the resonator by long coaxial cables (Figure 5.1d),
coplanar waveguides (Figure 5.1c) or some other type of microwave guide. These
“cables” introduce classical and quantum microwaves from far away sources (b̂in )
to control the circuit and collect signals from the circuit (b̂out ) into measurement
devices. However, they also connect the resonator to an infinite number of quantum
microwave modes that have their own dynamics.
Figure 5.6 sketches two models of an LC circuit connected to a semi-infinite
transmission line. Following Sections 4.4 and 4.5, the Hamiltonian of the capaci-
tively coupled LC circuit is
1 1 2
Ĥ = Ĥt-line + (Q̂ − Cg V̂ ) + ˆ .
 (5.56)
2C 2L
ˆ are the resonator’s charge and flux, Ĥt-line is the Hamiltonian for a
Here, {Q̂, }
semi-infinite transmission line (4.42), and V̂ = ∂t φ̂(x0 ) is the electric potential

(a)

(b)

Figure 5.6 A superconducting circuit – in this case, an LC resonator – can be

capacitively (a) or inductively (b) coupled to a semi-infinite transmission line,
which acts as an environment. Input–output theory (5.65) predicts that field
reflected by the resonator b̂out (t) carries part of the incoming field from the line
b̂in (t) together with a signal that leaks from the coupled circuit.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 88 Microwave Photons

created by the transmission line at the interface with the LC circuit. If the coupling
is inductive, we would instead have
1 2 1
Ĥ = Ĥt-line +Q̂ + ( ˆ − MLIˆ)2, (5.57)
2C 2L
where M is the mutual inductive coupling between the LC and the line
(cf. Section 4.7.1), and Iˆ = ∂x φ̂(x0 )/ l0 is the transmission line’s current close
to the place of mutual interaction.
When we quantize either circuit, we obtain a Caldeira–Leggett model for an
oscillator {â, â † } interacting with a bath of bosonic modes, corresponding to the
microwaves in the transmission line {b̂k, b̂k† }. If the coupling is so weak that it does
not change the eigenmodes of the transmission line, we may write
   
H = h̄ωâ † â + h̄ â + â † gk b̂k† + gk∗ b̂k + h̄ωk b̂k† b̂k . (5.58)
k k

Depending on whether the coupling is capacitive or inductive, we obtain slightly

different coefficients:
 
(cap) h̄Z ∗ (ind) h̄ ∂x u∗k (x0 )
h̄gk = vωk × uk (x0 ), or h̄gk = vωk × . (5.59)
2 2Z k
These are written in terms √ of the transmission line’s impedance Z, the speed of
light on the line v = 1/ cl, the dispersion relation ωk  v|k|, and the eigenmode
wavefunction uk (x). The coupling constants strictly vanish on the zeros of the eigen-
modes or their derivatives. For this reason, Figure 5.6 uses different types of trans-
mission lines to engineer a capacitive and an inductive coupling. Moreover, since
the wavefunctions are normalized, the constants decrease with the transmission line
size d as gk ∼ O(d −1/2 ). Fortunately, this prefactor is also the spacing in momentum

space
 gk ∼ gk0 (k)1/2 . This allows replacing all sums with integrals k |gk |2 →
dk|gk0 |2 , producing finite values of decay rates and interaction strengths in the
limit d → ∞.
There are three methods to work with the Caldeira–Leggett Hamiltonian, depend-
ing on the tasks we want to perform:
(1) We neglect the quantum fluctuations from the control line, treating the microwave
in the line as a classical field. The bath operators are replaced by complex
numbers that add up to a coherent drive (t), which we use to control the
resonator, as in Section 5.5.1.
(2) We study the state of the resonator as a density matrix that results from tracing
out the bath modes ρcav = trbath {|ψψ|}. We derive a master equation for ρcav
that describes the leakage of energy from the resonator into the line, as well as
decoherence introduced by the fluctuations in the antenna. This description is
used in Section 5.5.4.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 89

(3) We study the state of the radiation that leaks from the circuit into the antenna.
We derive Heisenberg equations relating the field on the line to the field in the
LC circuit, in what is known as input–output theory. This approach is used to
model the spectroscopy of a resonator in Section 5.5.3 and to develop techniques
for measuring the quantum state of the circuit based on the leaked radiation in
Section 5.5.7.

Input–Output Theory
Our goal is to analyze how the quantum fluctuations in the open transmission line
affect the resonator. We assume that the exchange of energy between both elements
happens at a slow pace, slower than the speed at which the LC circuit evolves, and
the speed at which waves propagate through the line. In this limit, the line can
control and monitor the resonator, without affecting too much its evolution. This
separation of time scales justifies a RWA effective Hamiltonian:
  
HRWA = h̄ωâ † â + h̄gk â b̂k† + h̄gk∗ â † b̂k + h̄ωk b̂k† b̂k . (5.60)
k k

The Heisenberg equations for this model describe a continuous exchange of pho-
tons between the oscillator and the bath (cf. Appendix B.3). Integrating out the bath,
we get a Langevin equation for the resonator:
 t
d
â = −iωâ(t) − i ξ̂ (t) − K(t − τ )â(τ )dτ . (5.61)
dt t0


The second term is the input field ξ̂ (t) = k gk∗ e−iωk (t−t0 ) b̂k (t0 ) transported by the
line. The third term is the field emitted by the resonator into the line in the past â(τ ),
which may be partially reabsorbed.
The memory function K(τ ) is the Fourier transform of the spectral function:
 
∗ −iωk (t−τ ) 1
K(t) = gk gk e = J QO (ω̄)e−i ω̄t dω̄. (5.62)
k
2π

The spectral function of an ideal transmission line grows linearly with the frequency,
acting as an Ohmic bath:

J QO (ω̄) = 2π |gk |2 δ(ω̄ − ωk )  παω1 . (5.63)
k

Whenever the spectral function is a such broad and smooth function, the Fourier
transform K(τ ) becomes extremely concentrated around τ = 0. The Markovian
limit is a regime in which we approximate the memory function with a Dirac delta,

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 90 Microwave Photons

K(t)  (iδLamb − κ2 ) × δ(t). In the Markovian limit, the bath has no memory of the
oscillator’s past, and (5.61) becomes local in time:

d  κ √
â(t) = −iω − â(t) − i κ b̂in (t). (5.64)
dt 2
We have to read this equation as follows. First, the environment introduces a Lamb
shift of the oscillator’s resonance, ω = ω − δLamb . This is a slowdown of the res-
onator caused by “dragging along” the modes of the line as it evolves. However, this
slowdown is rarely discussed, because once the resonator is connected to the line
we can only measure ω , not ω.
The second term −(κ/2)â is an exponential attenuation of the resonator field. It
is caused by photons leaking into the bath at the cavity decay rate κ = J QO (ω ) 
παω . The ratio κ/ω is uniform across most of the spectrum: low- and high-
frequency resonators decay at the same relative speed. Since κ is the rate at which
the bath and the resonator exchange energy, consistency with the RWA in (5.60)
imposes the limitation κ/ω  1 for this whole treatment to be justified.
The last term in the equation represents the injection of photons coming from the
√
line into the resonator, −i κ b̂in . When the transmission line contains a coherent
microwave drive, such as the (t) used to control the resonator, it is customary to
perform a displacement of the input operators, separating this “classical” contribu-
tion b̂in (t) → (t) + b̂in (t), from a truly quantum noise operator b̂in (t) that accounts
for the quantum fluctuations that the bath injects into the resonator.
The Heisenberg equation for the cavity is accompanied by an input–output rela-
tion that connects the field reflected by the cavity b̂out to the input field b̂in and the
field that leaks from the resonator:
√
b̂out (t) = b̂in (t) − i κ â(t). (5.65)

Together with (5.64), this can be used to study cavity spectroscopy and quadrature
measurements, as explained in Sections 5.5.3 and 5.5.7.

5.5.3 Cavity Spectroscopy

Spectroscopy is the study of radiation absorbed or emitted by a system, as a function
of the frequency and intensity of the “light” with which we illuminate it. Many
quantum mechanical systems – atoms, molecules, color centers in diamond, super-
conducting circuits – have a few experimentally relevant states with well-defined
energies. The absorption and emission spectra reveal the transitions that are allowed
between those states, the difference between those energies, hidden symmetries of
the quantum object, and even sometimes the underlying Hamiltonians.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 91

We have all tools to analyze the spectroscopy of an LC resonator or microwave

cavity. Let us begin with a reflected light setup, sketched in Figure 5.6a. The input–
output theory from Section 5.5.2 predicts the spectroscopic signal b̂out as a function
of the light that illuminates the resonator b̂in and the instantaneous state of the circuit
â(t). We can both show how to discriminate the resonator signal â from the total
reflected field b̂out , as well as give predictions on the combined signal.
To simplify this task, we assume that the probe field is a coherent microwave field
accompanied by a thermal microwave background. We split b̂in (t) = (t) + ε̂(t)
with a classical displacement (t) and a noise operator ε̂(t). The thermal noise
averages to zero ε̂(t) = 0 when we study the first moments of the reflected field
√
b̂out  = (t)−i κ â(t), giving us direct access to the state of the resonator. Most
experiments do spectroscopy with monochromatic beams (t) = 0 e−iωd t . The
Heisenberg equation for the resonator (5.64) predicts a quasistationary displacement
of the cavity in the asymptotic limit:
√
−i κ
â(t → ∞) = 0 e−iωd t . (5.66)
i(ω − ωd ) + κ/2

As anticipated in Section 5.5.1, the losses stabilize a coherent state with a number
of photons around 4|0 |2 /κ 2 . The reemitted signal and the incident microwave
interfere, producing the reflected field

i(ω − ωd ) − κ/2
b̂out (t → ∞) = 0 e−iωd t = r(ωd )(t). (5.67)
i(ω − ωd ) + κ/2

The cavity reflects the field (t) with a phase shift r(ωd ) = e−i2 arctan(2δ/κ) . This
shift can be experimentally determined (cf. Section 5.5.7) and used to estimate the
resonance ω – not the bare frequency ω! – with very high precision.
Measuring the reflected signal from a circuit is inconvenient, because the signal
has to be separated from the original drive using circulators. A common alternative is
to work with transmitted microwaves, using the configuration in Figure 5.4a, where
we inject the signal into one side of the resonator, and measure the output signal
on the other side. We can extend the input–output theory to consider the two decay
channels of the λ/2 resonator: to the left âL and to the right âR , with the total decay
κ = κL + κR :

d √ √
â = (−iω − κ/2)â − i κL b̂Lin − i κR b̂Rin, (5.68)
dt
√
out
b̂L,R = b̂L,R
in
− i κL,R â.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 92 Microwave Photons

A signal 0 e−iωd t coming from the right port b̂Rin results in a transmitted signal b̂Lout
and a reflected signal âRout . For instance:
√
κL κR
b̂Lout  = 0 e−iωd t , (5.69)
i(ω − ωd ) + (κL + κR )/2
i(ω − ωd ) − (κR − κL )/2
b̂Rout  = 0 e−iωd t . (5.70)
i(ω − ωd ) + (κL + κR )/2
The reflected signal is now zero when on resonance ωd = ω , and the transmitted
signal only contains the field emitted by the cavity:
|0 |2
n(ωd ) ∝ . (5.71)
(κ/2)2 + (ωd − ω)2
In the transmission spectroscopy setup, the LC circuit is a filter that transmits
frequencies on a narrow band around ωd = ω . The full-half-width (FHW) of this
filter – the separation between the two points at 50% maximum transmitted power –
is given by the decay rate κ and by the inverse of the quality factor Q = ω/κ.
A Q ∼ 105 means that the resonator filters out frequencies that are 0.002% outside
the central frequency ω . Such a cavity or resonator is a very good isolator in which
to embed other circuits.11 When those circuits are off-resonant with the cavity, they
will be effectively shielded from the environment. We can undo this shielding and
allow the circuit talk to the b̂k modes by shifting its resonances, bringing it closer
to the frequencies ω allowed by the resonator. These ideas will be explored in later
chapters about cavity-QED and quantum computing devices.

5.5.4 Losses and Heating

The quantum state of the resonator is described by a reduced density matrix
ρ = trbath |ψψ|, after tracing out the field. The evolution of ρ(t) can be deduced
from the Langevin equation (Gardiner and Zoller, 2004) or directly from the
Schrödinger equation (see Appendix B.2). We need to assume that the bath
remains in a relatively unperturbed state. More precisely, we separate all coherent
microwave fields that propagate through the line, from the quantum fluctuations
that are intrinsic to the bath b̂k → βk (t) + b̂k . The complex numbers βk add up to
create (t). The quantum fluctuations are usually fixed to a Bose–Einstein thermal
equilibrium distribution:

! 1
n̄ω = b̂k b̂k δ(ωk − ω)dω = −h̄ω /k T . (5.72)
e k B +1

11 Note the superconducting qubit attached to the transmission line resonator in Figure 5.1c. In other montages,
the resonator sits in between a transmission line and the quantum register, as seen in Figure 8.2.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 93

The result is a Lindblad equation or master equation:

κ
∂t ρ = −i[Ĥeff,ρ] + (n̄ω + 1) 2âρ â † − â † âρ − ρ â † â (5.73)
2
κ
+ n̄ω 2â ρ â − â â † ρ − ρ â â † .
†
2
The master equation contains an effective Hamiltonian Ĥeff = ω â † â + (t)â † +
(t)∗ â, with the same microwave drive  and renormalized frequency ω as before.
There are also two Lindblad superoperators cooling and heating the density matrix
with rates κ(n̄ω +1) and κ n̄ω , both of which depend on the temperature, through the
thermal occupation of the modes (see Section 4.1.1 and Figure 4.1). For experiments
with resonators around 6–10 GHz and temperatures of 10 mK, we typically neglect
the thermal occupation n̄ω  0 and all the heating terms (cf. Section 5.4.2).
In absence of external driving, the cooling and heating terms reach a balance
(Exercise 5.8):
d †
â â = κ(n̄ω − â † â) (5.74)
dt
and the resonator equilibrates with the bath â † â = n̄ω . We call the timescale for
thermal equibration, 1/κ, the T1 time. During thermalization, the density matrix
decoheres: The off-diagonal terms ρn=m = n|ρ|m or coherences die off at a
slightly lower pace, known as T2 ∼ 2/κ.
We could slow decoherence and dissipation by insulating the resonator, reduc-
ing κ. However, this happens at the expense of our ability to control and mea-
√
sure the resonator, since both b̂out, ∝ κ. For this reason, experiments strike
a balance between decoherence and control, engineering κ in the range of kilohertz
to megahertz for resonator frequencies of gigahertz. These value lead to quality
factors Q = ω/κ between 1 000 and 106 . We may interpret Q as the number of
times a photon bounces between the resonator’s walls before being lost into the
environment. These values are not only competitive with visible range quantum
optics, but they are so good that several groups are exploring the use of microwave
photons instead of qubits to store and manipulate quantum information.12

5.5.5 Beam Splitters and Circulators

There are two devices that are quite common when working with propagating
microwaves, both of which implement linear transformations of propagating
microwaves. The beam splitter is a four-port element dividing the signal of each

12 In the work by Vlastakis et al. (2013), ω ∼ 2π × 8.18 GHz, κ ∼ 2π × 7.2 kHz and Q ∼ 106 . This supports
extremely long-lived Schrödinger cat states and error correction!

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 94 Microwave Photons

Figure 5.7 (a) Microwave beam-splitter with 90◦ hybrid design. (b) Scheme of a
three-port circulator.

port into two different ports. In visible optics, beam splitters are partially reflective
mirrors, but in microwaves we have to use sophisticated designs of waveguides
with different lengths and impedances.
Figure 5.7a shows a 90◦ hybrid, a microwave beam splitter with resonance fre-
quency λ/2. The device is a square with four segments of coplanar waveguides √with
the same length λ/2, and different impedances – Z0 on the vertical arms, Z0 / 2 on
the horizontal segments. The hybrid connects to four waveguides at its four corners,
which we can use to inject and extract microwaves. Right at the frequency resonant
with the hybrid’s wavelength λ, the input and output modes are related by the simple
unitary transformation of a 50–50 beam splitter:
⎛ ⎞ ⎛ ⎞
b̂1out ⎛ ⎞ in
⎜ out ⎟ 0 0 i 1 ⎜b̂1 ⎟
⎜b̂2 ⎟ −1 ⎜0 i⎟ ⎜ in ⎟
⎜ ⎟= √ ⎜ 0 1 ⎟ ⎜b̂2 ⎟ . (5.75)
⎜b̂out ⎟ 2 ⎝i 1 0 0⎠ ⎜ in ⎟
⎝ 3 ⎠ ⎝b̂3 ⎠
b̂4out 1 i 0 0 b̂in
4

If we inject a signal through port 1 or through port 2, it splits into an equal super-
position of ports 3 and 4. For instance, S † b̂1in † S = − √i 2 b̂3out † + − √12 b̂4out † . The scat-
tering matrix of the hybrid depends on the frequency of the incoming microwave.
Away from resonance, not only is the splitter unbalanced, but some the energy
is spread over all four ports – including some undesired back-reflection. How-
ever, these effects are negligible within a broad band of 10% the central resonance
(Schneider, 2014).
Beam splitters, in this or other designs, are used to divide and to merge signals.
For instance, in Section 5.5.7 we divide a signal into two different ports b̂3,4 and
process those ports separately. This allows us to measure two different quadratures
(b̂1† + b̂1 ) and i(b̂1† − b̂1 ) of the original input field. Beam splitters can also be used
combine light from two different bosonic modes and create entangled microwaves
(Menzel et al., 2012).

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 95

The circulator is a curious device that implements a three-port system without

time-reversal symmetry. The scattering matrix of the circulator from Figure 5.7b is
⎛ out ⎞ ⎛ ⎞ ⎛ in ⎞
b̂1 0 0 1 b̂1
⎜ out ⎟ ⎝ ⎠ ⎜ in ⎟
⎝b̂2 ⎠ = 1 0 0 ⎝b̂2 ⎠ . (5.76)
b̂ out 0 1 0 b̂ in
3 3

The circulator maps an input signal from one port straight into the following port,
with a circular order denoted in circuits by an arrow (see Figure 5.7b). A circulator
breaks time reversal symmetry: If we reflect back a signal that exits from port 2, it
does not return to the input port 1, but exits from port 3. To break this symmetry,
circulators are built with permanent magnets, as bulky devices that cannot be placed
on chip – although there is ongoing research in implementing such circulators using
2D electron gases or periodically driven quantum circuits.
Circulators are used to isolate systems or separate signals. In the first type of
applications, we connect our circuit to port 1, ground port 3, and connect a measure-
ment device on port 2. This way our circuit is not affected by the noise generated by
the measurement device – which sinks into port 2. In the second type of application,
the setup is identical but we connect a microwave source to port 3. This way, the
measurement device only collects the light reflected by the circuit at port 1 and is
not affected by the microwave source.13

5.5.6 Amplification
The Gaussian set of operations includes the amplification of one or more quadra-
tures to make them more easily measured. Amplification appears naturally when
look at the single-mode (5.38) and two-mode (5.41) squeezing transformations in
terms of the gain factor G = cosh(r)2 ≥ 1:
√ √
â(r) = Gâ(0) − G − 1ei2θ â † (0), (5.77)
√ √
b̂1 (r) = Gb̂1 (0) − G − 1e2iθ b̂2† (0). (5.78)
The first equation is an example of a degenerate, phase-sensitive linear amplifier.
It is degenerate because it operates using a single mode â and just one frequency
of photons. It is phase sensitive because amplification only takes place along one
direction in phase space, p̂φ (r) = √i 2 (â † eiθ − âe−iθ ), which is the one we should
use in the measurements. This type of process is also called noiseless amplification
because it preserves the relative strength of fluctuations p̂/p̂, without any increase
in noise due to the amplification.

13 See, for instance, the work by Hoi et al. (2011, 2012), where a superconducting qubit operates on microwave
light, sorting out photons and creating non-Gaussian states.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 96 Microwave Photons

The second equation is an example of a nondegenerate, phase-insensitive linear

amplifier. In this paradigm, we distinguish the signal b̂1 from the noise mode b̂2 .
In actual experiments, b̂1 and b̂2 tend to have different frequencies so that they
can be filtered separately, hence the name non-degenerate. The amplifier is phase
insensitive because the signal operator contains two quadratures b̂1 = √12 (x̂1 + i p̂1 )
that are simultaneously amplified. In other words, we use the same setup to measure
x̂1 , p̂1 or any
√ other rotated quadrature and all of them will be amplified with the
same gain G. This simultaneous amplification is performed with a cost: The
fluctuations of b̂2 add up a contribution to the variance of b̂1 , x̂1 , and p̂1 . Assuming
that the amplifier mode is in an uncorrelated thermal state, we write

b̂1 (r)2 = G b̂1 (0)2 + (G − 1) n̄amp + 1

2
. (5.79)

The last term is the number of photons added by the amplification process. A quan-
tum limited amplifier is one which adds the minimum amount of photons allowed
by quantum mechanics n̄amp = b̂2† b̂2  = 0.

Linear Parametric Amplifier

It is possible to build a linear amplifier by periodically modulating the frequency
of a tunable resonator (Abdo et al., 2009). In an ideal realization, we modulate the
inductance of the resonator with a single frequency ∼ 12 L−1 [1 + ε cos(t + θ )]φ̂ 2 .
The oscillating perturbation ∝ ε is written down using the Fock operators of the
unmodulated Hamiltonian (ε = 0). Expanding the cosine term and applying the
RWA leads to the effective model:
 
Ĥ = h̄ωamp â † â + 12 h̄g ei(t+θ) â 2 + e−i(t+θ) â † 2 . (5.80)

We will assume that this oscillator is connected to a waveguide that brings the
signal and collects the amplified output. Using the input–output theory for a single
lead produces two linear equations:

d √
â = (−iωamp − κ/2)â − igei(t+θ) â † − i κ b̂in, (5.81)
dt
√
b̂out = b̂in − i κ â. (5.82)

We can solve these equations and establish a linear transformation from input b̂in
to output b̂out modes in frequency space. Following Roy and Devoret (2016), we
√
express â = (b̂in − b̂out )/ i κ and construct an equation that couples annihilation
and creation operators at different frequencies:

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 97
   
d κ −i(t+θ) out † d κ
+ iωa + b̂ −ige
out
b̂ = + iωa − b̂in −ige−i(t+θ) b̂in † .
dt 2 dt 2
(5.83)
Thisis more obvious when we Fourier transform the equation, using b̂(t) =
√1
2π
e+iωt b̂(ω)dω, to derive an algebraic relation between modes at equidistant
frequencies from the drive  = ω1 + ω2 :
1 −1 in
b̂out (ω1 ) − iρe−iθ b̂out † (ω2 ) = b̂ (ω1 ) − iρe−iθ b̂out † (ω2 ). (5.84)
χ (ω1 ) χ (ω1 )∗
We have introduced the single-mode bare susceptibility χ (ω)−1 = 2i(ωa −ω)/κ +1
and coupling strength ρ = 2g/κ. These equations can be written in scattering form,
as a unitary transformation det(S) = 1 between fields:
 
b̂out (ω1 ) b̂in (ω1 )
= S(ω1,ω2,ρ,κ) in † . (5.85)
b̂out † (ω2 ) b̂ (ω2 )
When the amplifier is operated in non-degenerate mode, we use two different
frequencies ω1 = ω2 and drive close to the parametric resonance   2ωa . The
scattering matrix gives us the direct GD = |S11 |2 and indirect gain GI = |S12 |2 as
a function of the rescaled detuning δ = (ω2 − ω1 )/κ:
1 + δ 2 + ρ 2 in 2ρe−iθ
b̂out (ω1 ) = b̂ (ω 1 ) + b̂in † (ω2 ) (5.86)
ρ 2 + (i + δ)2 ρ 2 + (1 + iδ)2
= S11 b̂in (ω1 ) + S12 b̂in † (ω2 ).
When we operate on direct gain, the signal is inserted and recovered at the same
frequency, and the amplifier is phase preserving. When we use the indirect gain,
the signal is injected in mode b̂in (ω2 ) and recovered from b̂out (ω1 ) ∝ b̂in † (ω2 ). This
is called a phase-conjugating amplifier, because it transforms a coherent field (t)
√
into its complex conjugate GI (t)∗ . The gains in the nondegenerate parametric
amplifier can be quite large: Ideally, both GD and GI diverge when the cooperativity
approaches ρ → 1. This divergence is associated to instabilities in the driven cavity,
which can be cured by shifting  slightly away from the parametric resonance, or
adding internal losses to the resonator.
We can operate the same device as a phase-sensitive degenerate amplifier, taking
the limit ω2 = ω1 = /2. The amplifier implements single-mode squeezing on
b̂(ω1 ) and b̂(ω1 )† . The unitary matrix S has two eigenvalues, λ and 1/λ, with the
gain factor

√ ρ 2 − 2 + 1
λ= G=  , (5.87)
ρ 2 − 2 − 1
expressed in terms of the normalized detuning  = 2(ωa − ω1 )/κ.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 98 Microwave Photons

Josephson Parametric Amplifiers

We have mentioned that the linear parametric amplifier can be implemented using
a tunable cavity, such as the one in Figure 5.4c, by periodically modulating the flux
through a SQUID (Abdo et al., 2009). This design has two potential problems. The
driving mechanism, with an external antenna, can introduce quantum fluctuations
in the cavity, causing decoherence. Moreover, once we amplify the signal enough,
there will be a large number of photons, and the SQUID will stop acting as a linear
inductor.
There exists an alternative setup that embraces this nonlinearity. It is known as
a Josephson parametric amplifier (JPA) and consists of a superconducting cavity
with an embedded chain of SQUIDs (Castellanos-Beltran et al., 2009). This chain
provides a knob to adjust the frequency of the cavity and creates a large nonlinear
contribution to the Hamiltonian. The JPA’s long λ/2 also supports two ports: One
side of the resonator, strongly coupled, receives the signal that will be amplified;
the other port, weakly coupled, is illuminated with a strong coherent microwave or
pump. The pump strongly displaces the cavity and activates the nonlinearity, causing
a periodic shift of the cavity’s resonance. We treat this modulation as a c-number
contribution to the cavity operator â(t) → â(t) + α(t), keeping terms of order |α|
or larger in the Hamiltonian. The result is a linear quadratic Hamiltonian (5.80).
The JPA is modeled as a single-mode nonlinear oscillator with a Kerr-type inter-
action between photons:

K †2 2
H0 = h̄ωa â † â − â â . (5.88)
2
The Kerr nonlinearity K can be traced back to the Josephson junction potential,
cos(φ̂/ϕ0 ) = 1 − 12 (φ̂/ϕ0 )2 + 14 (φ̂/ϕ0 )4 + · · · . The input–output theory for the JPA
with a signal b̂Sin and a pump b̂pump
in
is
 
d κ1 + κ2 √ √ in
â = −iωa − â + iK â † â 2 − i κ1 b̂Sin − i κ1 b̂pump . (5.89)
dt 2

The pump field is replaced with a complex classical drive b̂pump in

∼ βe−i(ωp t+ξ )
that displaces the cavity by ∼ α0 e−i(ωp t+ξ )t ∼ b̂pump
in
/ωa (see Section 5.5.1). A new
−i(ωp t+ξ )t
bosonic operator ĉ = â − α0 e collects the quantum fluctuations around
this displacement, producing a familiar Langevin equation:
 
d κ1 + κ2 √
ĉ  −iωa − ĉ + i2Kα02 ĉ + iKα02 e−i(2ωp t+2ξ ) ĉ† − i κ1 b̂Sin . (5.90)
dt 2

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 99

This model is equivalent to the linear parametric amplifier in (5.81), with the pump
2ωp = , phase θ = 2ξ , intensity-dependent resonator frequency ωamp = ωa −
2Kα02 , and intensity-dependent coupling g = Kα0 . Using these identities, we can
find the regimes of nondegenerate parametric amplification ωp  ωa , where the
JPA acts as a phase-insensitive amplifier, as well as the phase-sensitive degenerate
amplifier regime, ωp  ωS .
JPAs are fantastic amplifiers that power many of today’s experiments. Unfor-
tunately, the stronger the amplification, the narrower the bandwidth over which
the JPA operates. Typical designs of JPA amplify signals of gigahertz with large
gains GD,I ∼ 22 dB, over a relatively narrow frequency band of around megahertz
(Castellanos-Beltran et al., 2009).
There is an alternative to the JPA that supports comparable gains over a broader
range of frequencies. These devices are known as traveling wave parametric
amplifiers (TWPAs). TWPAs were introduced by Cullen (1960), but they have
been adapted to work with a long, open string of chained Josephson junctions or
SQUIDs. The chain provides a 1D model with a Kerr nonlinearity that is capable of
amplifying the waves that propagate through it, with comparable gains of ∼ 20 dB
over a bandwidth of 3 GHz (Macklin et al., 2015). A single amplifier can work
over multiple control signals for a superconducting circuit, and can be used to
amplify the output from many different measurement devices, simultaneously, and
without interference between signals. This is ideal for large and complex quantum
computing setups, where tens and probably hundreds of qubits must be manipulated.
Unfortunately, the theory and details of TWPAs fall outside the scope of this book,
but you are encouraged to dig into the literature to learn about this incredibly useful
technology.

5.5.7 Photon Quadrature Measurements

There are no good microwave photodetectors and photon counters in the market yet.
Instead, superconducting quantum circuits are analyzed using classical devices that
measure the power or voltage. These classical detectors sit at the end of a chain of
amplifiers – cold (inside the fridge) and hot (outside) mixers and filters that boost,
shift around, and clean the signal.
Following input–output theory, the signal that leaves the experiment b̂out contains
information about the instantaneous state of our system, in the form of a small
√
photon flux κ â(t). Hence, we describe all these stages with a simple equation:
√  √
b̂amp = G(b̂out + ε̂ noise ) = Ḡâ(t) + G − 1ĥ†, (5.91)

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 100 Microwave Photons

with a prefactor Ḡ that includes all gains and losses – also the attenuation at the
cavity port κ – and an effective field ĥ summing all classical and quantum noise. This
includes noise from the amplifiers, beam splitters, detectors, and cables themselves.
The added noise is quantified in the number of photons it adds prior to the gain
itself (5.79). This noise can be as low as 1/2 in ideal quantum-limited amplifiers, or
even lower in noiseless phase-sensitive amplifiers. However, when further connec-
tions and mixers are accounted for,14 we find more common values of 1−10 pho-
tons in Josephson-junction-based amplifiers, and 30−50 in cyrogenic high-electron
mobility transistor (HEMT) amplifiers.

Power Measurements
We can feed the signal b̂amp (t) into a device that measures the integrated or average
power. A broadband power meter uses diodes to rectify the microwave current,
creating a dc signal that is calibrated to determine the average, peak, or integrated
power in the signal. In practice, we need to very accurately calibrate the average
power Pnoise of the background noise – i.e., noise introduced by amplification stages,
input lines, etc. – to separate the power of the actual signal Psignal = P −Pnoise , from
the measured power P .
Since the dynamical range of a detector can span multiple orders of magnitude,
the signal is measured using a logarithmic scale. The dBm unit is a log-10 scale
referenced to a standard power of 1 milliwatt. A power P in watts is converted to
this scale as PdBm = 10 log10 (P /10−3 ). A very good broadband power detector
operating in the gigahertz range can have a minimum sensitivity of −70 dBm. This
sensitivity can be improved using spectrometers that work on narrow bands around
the desired frequency. A good device can have a detection threshold of around
−140 dBm/Hz, relative to the bandwidth. If we wish to detect photons generated
by a cavity with a bandwidth κ from 1 kHz to 1 MHz, this means the threshold for
detection will lay around −110 and −80 dBm, respectively.
An important problem is that even in this scale, quantum microwave signals
can take ridiculously small values. When a cavity releases photons of frequency
ω ∼ 2π × 10 GHz at a rate κ = 1 MHz, the output power will be Pphoton = h̄ωκ ∼
6.626 × 10−18 W, or PdBm ∼ −171 dBm. This is ∼ 10−9 or ∼ −90 dB to below
the threshold of commercial power detectors! Hence the need for various stages of
amplification that bridge this gap.
Another problem is obtaining a good signal-to-noise ratio for Psignal , estimated as
the difference of two large quantities, the measurement with signal P , and a calibra-
tion of the noise in the empty line Pnoise . Both quantities have to be measured with

14 See Exercise 5.10 to discover how these noises add up.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 5.5 Gaussian Control of Microwave Photons 101

high precision, over long integration times, and fluctuating experimental conditions.
This challenge in stability and calibration makes this measurement technique less
desirable than quadrature measurement techniques.

Quadrature Measurements
The use of power meters has been superseded by the use of digital techniques.
An analog-to-digital converter (ADC) can periodically sample the voltage of an
incoming quantum signal, producing a high-resolution stream of data from which a
computer – or a field programmable gate array (FPGA) card (Eichler et al., 2012) –
estimates moments â , â † â, etc. The benefit of this technique is that it provides
more information than just a power measurement, but actual implementations face
two obstacles. First, the ADC will only sample one quadrature, the voltage â(t) +
â † (t). Second, we must slow down the signal, because the ADC cannot sample fields
b̂amp (t) ∼ â(t) that oscillate at frequencies of gigahertz.
Theoretically, both problems have easy solutions. The first obstacle is avoided
by dividing the signal b̂amp into two copies – one in phase, and another one with a
90◦ rotation eiπ /2 b̂amp – and feeding each copy to a different ADC. This provides
us with two streams of values, associated to the two quadratures √12 (b̂amp + b̂amp† )
and √i 2 (b̂amp − b̂amp† ).
The second problem is solved by looking at how the cavity operator evolves.
From input–output theory (cf. Section B.3.2), we know that most of the oscillation
in the field is due to “trivial” phases, â(t) = âslow (t)e−iωt . The actual information
is encoded an operator that evolves at a much slower pace âslow (t) = eiωt â(t). We
just need to fabricate this operator by mixing the original signal with the opposite
field eiωt .
Both solutions are jointly solved by a type of microwave circuit called mixers.
A simple mixer is a nonlinear circuit that takes two signals f (t) and i(t), and returns
their product f (t)i(t), as sketched in Figure 5.8a. If we feed our signal â(t) and a
classical oscillating field i(t) = cos(ωm t +θ) into a mixer, it will produce two copies
of the signal at two sidebands. The first sideband oscillates with a lower frequency
and is close to our slow modulation ei(ωm t) â(t)  ei(ωm −ω)tx âslow (t). The second side-
band oscillates at higher frequency ω + ωm and can be filtered out. In addition to
this, note how the sidebands will be phase-shifted an angle θ determined by the
input field. We can use this angle to select which quadrature we wish to explore.
The complete setup is shown in Figure 5.8c. We feed the signal to an amplifier
and chain the output into an IQ mixer. The IQ mixer uses two mixers to combine
the signal with two reference microwaves, cos(ωm t) and sin(ωm t), derived from a
stable local oscillator. The two outputs of the IQ mixer are two quadratures, X̂ and
P̂ , that are sampled by two ADCs. This produces two classical traces X(tn ) and

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 102 Microwave Photons

Figure 5.8 (a) A mixer is a nonlinear device that multiplies a signal f (t) with a
modulation i(t). (b) An IQ mixer divides the input signal, combining it with two
signals in 90◦ opposition cos(ωt) and sin(ωt). The output of the mixer contains
the two orthogonal quadratures of the signal. (c) The combination of a mixer
with a local oscillator, filters, and amplifiers allows us to measure any quadrature
Q̂ω,θ = √1 (aω e−iθ + aω† e−iθ ) in a quantum field φ(t).ˆ The classical signal q(t)
2
is a random sequence of voltages associated to one realization. Averages of many
realizations of q(t) estimate the expectation value Q̂ω,θ (t).

P (tn ) that combine into a complex number Z(tn ) = X(tn ) + iP (tn ). In practical
applications, ωm − ω is never zero, but a finite value around MHz. This is slow
enough that the ADC can reconstruct the signal, but large enough that the signal
is not masked by slow 1/f noise in the electronics. If we take into account this
residual oscillation and calibrate
√ the gains of our circuit, we can build a sequence
S(tn ) = Z(tn )e−i(ωm −ω)t / Ḡ, that is a sampled measurement of our slow field with
some noise ŝ(t) = âslow (t) + ĥ† (t).
The classical values S(tn ) can be used to reconstruct a measurement of resonator
quadratures â. Let us assume that we prepare the LC circuit and stop all external
controls. As soon as we do, the cavity field begins to decay exponentially âslow (t) =
â(0)e−κt/2 . Using the weight function w(t) = κe−κt/2 , we recover the cavity oper-
ator from traces of our ADC chain:
 
w(t)ŝ(t)dt = â(0) + ĥ†w ∼ Sw := w(tn )S(tn )t. (5.92)
n

In other words, in each experimental realization, the weighted average Sw imple-

ments a measurement of the operator â(0) + ĥ†w . If we repeat the experiment many
times, we can estimate the ensemble averages Sw , Sw∗ Sw , Sw2  , . . .. These values
provide the statistics of the cavity because (i) odd moments of the noise cancel out

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 Exercises 103

ĥ†w  = 0, (ii) the statistics of the noise are independent of the signal ĥ†w â 
ĥw  â  0, and (iii) noise moments – ĥw ĥ†w  or higher – can be calibrated from
experiments with an empty cavity. This allows us approximate arbitrary moments
of the cavity field as â(0)  Sw , â † (0)â(0)  Sw∗ Sw  − hw ĥ†w , etc. First,
second, third, and higher moments are used to test whether a state is Gaussian, and
to do Wigner function tomography as explained in Section 5.4.6.
This analysis of bosonic quadratures is a powerful method that has been used to
reconstruct the wavefunctions of squeezed states (Mallet et al., 2011) and of single
photons (Eichler et al., 2011). The quadratures themselves allow the reconstruction
of other properties, such as the energy of a wavepacket with less than one pho-
ton – i.e., values around 10−26 Joules! – or the spectral distribution of a photon
wavepacket (Menzel et al., 2010).

5.6 Conclusion
We close here a long chapter on microwave photons. The chapter began discussing
how linear circuits – and nonlinear circuits operating in a linear regime – store
energy in quantized units, which we called photons. These bosonic excitations
behave as optical photons in all respects, from the linear spectrum of equispaced
eigenenergies, h̄ω,2h̄ω,3h̄ω . . ., to the quantum description of modes and wave-
functions.
The noninteracting nature of photonic excitations poses both advantadges as
well as problems. Among the advantages, we find a relatively simple mathematical
description, access to amplification, and real-time tomography of wavefunctions.
Photons can also be confined in high-quality environments, for long times of
hundreds of microseconds and support long-lived entangled states and quantum
superpositions.
The problems begin when we want to create those interesting states or think about
developing a quantum computer. We cannot use linear circuits to generate or detect
individual photons, and we are constrained to a family of states (Gaussian states)
that can be simulated classically and pose no computational interest. To escape these
constraints, we must introduce nonlinearities. This is the focus of the next chapter,
where we study the design and operation of superconducting qubits.

Exercises
5.1 Compute the expected values and uncertainties of q̂, φ̂, V̂ , and Iˆ for the vac-
uum state |0 of an LC resonator.
5.2 Plot the average occupation of a microwave cavity as a function of the cavity
resonance ω and the temperature of the cryostat, for temperatures ranging from

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 104 Microwave Photons

5–100 mK. What is the minimum frequency at which the occupation number
lays below 10%? And 1%?
5.3 Show that the number of photons in a coherent state |α is n̂ = |α|2 . Compute
the uncertainty of all quadratures, {φ̂, q̂}, on such a state.
5.4 Show that the ground state of Hamiltonian (5.8), for constant drive (t) = 0 ,
is a coherent state with α ∝ . Hint: Use the definition of the coherent state
in terms of displacement operators, and the commutation relations between a,
a † , and D(α), such as âD(α) = D(α)(â + α).
5.5 If we start with the vacuum as initial state |(0) = |0, the coherent drive (t)
in (5.8) will produce a coherent state |(t) ∝ |α(t) with the displacement
from (5.10). Relate the Schrödinger and Heisenberg pictures, showing that
U (t)† â |(t) = â(t) |(0). Use the explicit formula for â(t) to demonstrate
that |(t) is a coherent state. Bonus points if you get the phase of the coherent
state right.
5.6 Compute the Wigner function for a single photonic mode in (i) a coherent
state, (ii) a Fock state with one photon, and (iii) a Schrödinger cat with generic
displacement α. Which of these states is Gaussian? Which of them has a
negative Wigner function? Which of them is broader in phase space and which
of them saturates the Heisenberg uncertainty relation?
5.7 We want to place a magnetic molecule inside a λ/2 and a λ/4 transmission
line resonator. The molecule couples to the magnetic field generated by the
nearby currents in the line.
(1) Which position optimizes the coupling between the molecule and the fun-
damental mode, ω0 ?
(2) Which position optimizes the coupling to the first harmonic of the cavity,
ω1 ?
(3) Where should we place the molecule so that it couples equally well to both
modes, ω0 and ω1 ?
5.8 Cavity spectrum. In Section 5.5.3, we studied how the driven-dissipative res-
onator stabilizes to an average occupation number. We want to solve the same
problem but using master equations:
(a) Starting from the driven model H = a † a + (∗0 e−iωd t a + 0 eiωd t a † ), and
following Section B.2, show that the master equation in the interaction
picture becomes a function of the detuning, δ = ωd − ω:
  κ
∂t ρ = −i δ â † â + 0 a + H.c.,ρ + 2âρ â † − â † âρ − ρ â † â . (5.93)
2

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 Exercises 105

(b) Write and solve the evolution equations for â = tr(âρ) and for the
number of photons â † â. Hint: Use the cyclic property of the trace to
convert terms of the form tr(AρB) = tr(BAρ) = BA.
(c) Show that the FHW linewidth is δ = κ.
5.9 Assume a cavity without driving, coupled to an environment at zero temper-
ature. Use the master equation (5.73) to solve analytically the density matrix
of the cavity, using as initial conditions the pure states ρ(0) = |ψ ψ| in a
superposition of one photon and the vacuum, |ψ = cos(θ ) |0 + sin(θ ) |1.
Show that the average number of photons â † â = tr(ρ(t)â † â) in this state
follows (5.74). Study the decay of the coherence 0|ρ(t)|1. How fast is it,
and how does it relate to κ?
5.10 A microwave signal passes sequentially through two amplifiers with gains
G1,2 and added number of photons n̄1,2 . Each stage is described by an equation
of the form (5.91), with noise operators ĥ†1,2 that are independent from each
other – e.g. ĥn† m† r† k n† m† r† k
1 ĥ1 ĥ2 ĥ2  = ĥ1 ĥ1  ĥ2 ĥ2  – and from the signal. Use this
property to estimate the total gain and total number of added photons of the
device.

https://doi.org/10.1017/9781316779460.005 Published online by Cambridge University Press

 6
Superconducting Qubits

We have already introduced the qubit as the smallest quantum mechanical system
that we can study. In this chapter, we further explore qubits, examining how they
behave, both formally – through Schrödinger and master equations – and through
their implementation in various superconducting circuits. Each design has unique
characteristics, but we can develop a common framework for describing the qubit’s
evolution, its coupling to the environment and to external controls, the qubit’s qual-
ity and coherence properties, etc. We also discuss in detail the most used qubits
in present and past state-of-the-art experiments – charge qubits, transmons, and
flux qubits. This chapter intentionally drops some interesting circuits – fluxoniums,
g-mons, and other curious animals out there – but it should be enough to approach
any qubit architecture critically, deducing their Hamiltonians and coherence prop-
erties, and understanding the pros and cons of each those designs critically.

6.1 What Is a Qubit?

6.1.1 From Logical to Physical Qubits
In classical information theory, the bit is the smallest unit of information and corre-
sponds to a discrete variable with two states, labeled 0 and 1. The qubit – or quantum
bit – generalizes this idea, denoting the smallest useful amount of quantum informa-
tion. A qubit may adopt any possible state in the Hilbert space created by the super-
positions of two orthogonal and physically distinguishable states |0 and |1. The
continuum of states for a qubit includes not only the basis states |0 and |1, but
any other pure state – a linear combination α |0 + β |1 with normalized com-
plex amplitudes α,β ∈ C, |α|2 + |β|2 = 1 – or mixed state – Hermitian matrices

ρ = i,j =0,1 ρij |ij | with trρ = 1, and ρ † = ρ (see Sections 6.1.5 and 6.1.6).

106

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.1 What Is a Qubit? 107

The word bit applies to both the unit of information and to the physical system
that encodes it. Similarly, a qubit would be the smallest unit of quantum information,
as well as any physical system that embodies those two-dimensional superpositions.
Qubits as such do not exist in Nature, because all known physical systems require
more than just two quantum states to be described. A photon has a polarization
degree of freedom, but also frequency and momentum; an electron has a spin,
but also a spatial wavefunction. Creating a physical qubit requires controlling the
dynamics of a quantum system, constraining it to a subset of all physically available
quantum states.
In addition to this simplified dynamics, there other practical requirements, includ-
ing preparation, read-out, and measurement. We can group them in a “shopping
list,” a variation of DiVincenzo’s famous requirements for a quantum computer
(DiVincenzo, 1995):
(1) The physical qubit has two accessible orthogonal eigenstates |0 and |1.
(2) It supports arbitrary, long-lived quantum superpositions α |0 + β |1.
(3) We can reset the qubit to some state, typically |0.
(4) We can do a projective measurent on the computational basis, typically through
the observable σ z = |11| − |00|.
(5) We can perform (or approximate) arbitrary unitary rotations in the qubit’s
Hilbert space.
(6) We can implement at least one type of universal two-qubit operation among
pairs of qubits.
(7) We can engineer devices with medium to very large numbers of qubits.
(8) Tolerances in all operations – unitary gates, measurements, qubit reset – are
tight enough to implement fault-tolerant quantum computation.
Out of this list, items (1)–(5) suffice for many applications in quantum communi-
cation and quantum optics. Combined with some type of qubit–qubit interaction –
which unlike in term (6) does not need to be very precise (1) this even allows us
to build some types of quantum simulators. Conditions (6) and (7) were recently
achieved together in setups with more than 50 qubits (Arute et al., 2019; Wu et al.,
2021). This enables near-term intermediate scale quantum computers – also known
as noisy intermediate-scale quantum (NISQ) computers thanks to Preskill (2018) –
to perform tasks, such as the simulation of random quantum circuits, which are
arguably difficult in the classical computing world. It also opens the door to other
useful applications of imperfect near-term quantum computers, especially in the
study of quantum physical systems.
However, the ultimate goal in the field is to achieve fault-tolerant quantum com-
putation as a means to build quantum computers that run arbitrarily long quantum
algorithms. Fault-tolerant devices need extremely large numbers of qubits to encode

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 108 Superconducting Qubits

Table 6.1. Some qubits and energy scales involved.

Object Degree of freedom Energy scales

Photon Polarization IR to UV
" Frequency: two frequencies "
" Which-way: a photon moves "
through one of two waveguides
Electron Electronic spin GHz to THz
" Position in two quantum wells GHz
Nuclei Angular momentum GHz
Atoms, ions Electronic state 400 THz–1015 Hz
molecules, color Hyperfine state GHz
centers Position in lattice kHz–MHz
Molecules Rotational states 100 GHz
" Vibrational states THz
Superconductors Charge/current states GHz

information redundantly. These qubits are operated with arbitrary numbers of mea-
surements and quantum gates to implement quantum algorithms that are resilient
to environmental and operational errors. We will discuss this requirement further
in Section 8.5. However, as of the writing of this book this is a regime that is far
from being achieved, even if we have promising results and proofs of principle
demonstrators Chen et al. (2021).
The quest for physical qubits has ran in parallel to – and often been the motivation
for – the search for controllable quantum systems in the lab. Table 6.1 enumer-
ates the most successful qubit systems, together with the degrees of freedom that
encode the information and the energy scales involved. Indeed, the energy scales
of the qubit degrees of freedom are extremely important, because they underlay the
general conditions for doing quantum experiments regarding temperature, isolation
requirements, cooling, and preparation times (see Section 4.1.1).
However, our shopping list introduces a new and extremely important considera-
tion: the need of single out two states from an experimental device to encode a qubit.
This is not as easy as it sounds. Take, for instance, the photonic superconducting
circuits from Section 5.4. We could identify the qubit with the states of zero and
one photon, |0 and |1. However, as discussed in Problem 6.2, we cannot rely on
microwaves to implement transitions between |0 and |1 without involving other
photon number states. Linear LC-type circuits therefore do not satisfy the require-
ments for qubit preparation, arbitrary rotations in the qubit space, or even projective
measurements.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.1 What Is a Qubit? 109

Figure 6.1 (a) A general qubit uses two metastable states |0 and |1, out of
a spectrum that is anharmonic – that is, the energy spacing is not uniform. In
particular, the qubit ω01 is not an integer multiple or fraction of the energy
difference to other neighboring states of the system, ω0n or ω1n . (b) Energy levels
of the Ca+ ion, including the two qubit states |0 = |S1/2  and |1 = |D5/2 .

One solution is to focus on a degree of freedom that is naturally reduced to a two-

dimensional space – e.g., the spin of an electron, the polarization of a propagating
photon – while simultaneously “freezing” other properties of the quantum system.
In this book, and in most of circuit-QED, we instead look for physical systems with
an intrinsically anharmonic energy spectrum, such as the one in Figure 6.1a. In this
picture, the |0 is mapped to the ground state – facilitating the reset of the qubit –
and the |1 is the first excited and long-lived state. Anharmonicity means that the
energy gap h̄ω01 between qubit states is not an integer fraction of the gap to other
neigboring states. In other words, ω1n,ω0n cannot be a multiple of ω01 . This implies
that we can drive transitions between our qubit states |0 and |1 using resonant
photons with energy h̄ω01 , without leakage – transitions to other states.
The anharmonicity condition is often fulfilled by matter qubits. Take, for
instance, a trapped ion quantum computer using Ca+ atoms (cf. Figure 6.1b).
The qubit’s zero and one states are associated to the ground state manyfold S1/2
and to the metastable state D5/2 . When we illuminate the ion with a laser at a
wavelength 729 nm, we bridge the energy difference between the S and D states,
inducing coherent rotations in the qubit space. Since the energy difference between
the S and P1/2,P3/2 , or between D3/2 and P , is not a multiple of this wavelength,
the atom rarely transitions outside the qubit space {S1/2,D5/2 }.
Our goal in this chapter is to recreate these conditions, engineering what we call
superconducting articifial atoms. These are nonlinear circuits that can store qubit-
like degrees of freedom, and where the nonlinearity arises either from the Coulomb
repulsion between Cooper pairs or from the nonlinear inductance in a Josephson
junction. This includes three large families of qubits: the charge qubit (6.2), the
transmon qubit (6.3), and the flux qubits (6.4).

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 110 Superconducting Qubits

Note that in addition to the quantum information bit and the physical qubit, there
exists the idea of logical qubit, analyzed later in this book. In this case, the quantum
bit is stored formally in the Hilbert space of a complex quantum system. The qubit is
no longer identified with the eigenstates of that model, but is prepared, operated, and
measured in a sophisticated, error-correcting, or error-suppressing way. Typically,
logical qubits can be synthesized out of systems from many imperfect physical
qubits. However, as demonstrated by Ofek et al. (2016), one can also develop logical
qubits using complex quantum systems, such as Schödinger cats created by the inter-
action between superconducting qubit circuits and superconducting cavities. Here
the superconducting qubit is just a nonlinear circuit that enables the preparation,
operation, measurement, and error correction of the logical qubit, which is no longer
an associated to an eigenstate of the cavity but to a very long-lived quantum state.

6.1.2 Qubit Hamiltonian

All artificial atoms that we study in this chapter have a simple Hamiltonian
representations:

H = En () |nn| + (t)d̂. (6.1)
n

The levels |n are the eigenstates of the circuit that makes our atom. The lowest
energy states |0 and |1 are identified with our qubit. As sketched in Figure 6.1a,
the spectrum is anharmonic and the gap h̄ω01 = E1 − E0 is not an integer fraction
of the spacing to other levels, ω02,ω12 . . .
There will be typically two direct controls on any qubit. Tuning an external
parameter such as a magnetic flux , we can change the energy levels En ()
preserving their population. We can also “illuminate” the qubit with microwaves
(t) that couple to the off-diagonal dipolar moment operator n|d̂|n = 0 and
induce transitions between eigenstates.
By using weak or slow controls, and engineering the spacing between levels,
we can constrain the dynamics of the artificial atom to just two lowest energy
states. This effective two-level system lives in a reduced Hilbert space H2 created
by arbitrary linear superpositions of |0 and |1. Following Section 2.2, we model
the dynamics in this space using the most general real-valued1 qubit Hamiltonian:
h̄(t) z h̄ε(t) x
H = σ + σ + E0 1 = E0 + B(t) · σ . (6.2)
2 2

1 It will become evident, when we introduce actual qubit designs, that we can write all models using two
observables, one diagonal – the qubit energies – and one off-diagonal – the dipole moment operator.
A consistent selection of phases in the definition of states |n allows us to represent those observables as σ x
and σ z (see Problem 6.1).

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.1 What Is a Qubit? 111

The parameter (t) is the qubit gap h̄ω01 = E1 − E0 , associated to σ z = |11| −

|00|. The transverse driving and the dipole moment strength d̂ ∼ σ x = |10| +
|01| combine into ε(t) ∝ (t). The energy offset E0 can be ignored, as it only
produces an undetectable global phase shift of the wavefunction.

6.1.3 Interaction Picture

In experiments, qubits spend most of the time parked close to some reference fre-
quency, such as 0 . During that time, they evolve freely as
 
1
U0 (t) = exp −i 0 σ z t = e−i0 t/2 |11| + ei0 t/2 |00| . (6.3)
2

We can get rid of this “free evolution” and eliminate the dynamical phases by work-
ing in a new rotating frame, called the interaction picture. This involves redefining
the wavefunction and the observables:

|ψI (t) = U0 (t)† |ψ(t) , ÔI = U0 (t)† ÔI U0 (t), (6.4)

so that expectation values and predictions remain unaltered ψI |ÔI |ψI  = ψ|Ô|ψ.
In this frame, diagonal observables such as σ z remain the same and the evolution
of the new wavefunction ψI is dictated by an interaction picture Hamiltonian:

d
i h̄∂t |ψI  = HI (t) |ψI  = U0† (t)H (t)U0 (t) − i h̄U0† (t) U0 (t) |ψI (t) . (6.5)
dt

This transformation helps us in the study of microwave controls:

ε(t) = ε0 cos(ω0 t + ϕ). (6.6)

It is customary to take the microwave field as frame of reference 0 = ω0 , because

it actually provides the experiment with a calibrated time reference. The interaction
Hamiltonian for (6.2) becomes

h̄( − ω0 ) z h̄ε0 −iϕ +

HI (t) = σ + e σ + ei2ω0 t+iϕ σ + + H.c. . (6.7)
2 4

The diagonal part is proportional to the detuning between the qubit frequency and
the drive δ =  − ω0 . Out of the off-diagonal terms, we can single out the static one
e−iϕ σ + , using a rotating wave approximation to neglect the interaction that rotates
with twice the frequency ei2ω0 t+iϕ σ + .

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 112 Superconducting Qubits

6.1.4 Single-Qubit Gates

A physical qubit must implement or approximate arbitrary single-qubit unitary
operations on the qubit Hilbert space H2 . Up to irrelevant global phases, these
single-qubit gates are rotations in the Bloch sphere:
U (θ, n) = exp(iθ n · σ ) = cos(θ)1 + i sin(θ)(
n · σ ), (6.8)
with angle θ around the rotation axis n ∈ R3 on the Bloch sphere.
Evolution with the Schrödinger equation (2.16) already implements a unitary
operation. If we wish to create a particular gate W , we must investigate the combi-
nation of gaps (t) and external drives ε(t) that, after a finite time, give rise to the
desired operation U (T ) = eiξ W , up to an irrelevant global phase ξ . The search for
(t) and ε(t) is a quantum control problem.
Fortunately for us, the interaction picture (6.7) already brought the Hamiltonian
into a form H ∝ n · σ , where the direction n can be fully tuned by changing the
qubit frequency , the amplitude ε0 and the phase ϕ of the drive. The simplest gates
are diagonal in the computational basis, n = ez = (0,0,1). They are called phase
gates because they only impart phases on the qubit states:
 iθ 
e 0
Uθ = e iθσ z
= ∼ e−iθ |00| + e+iθ |11| . (6.9)
0 e−iθ
This operation is generated by tuning away the qubit from its resting frequency,
(t) = δ + ω0 . Choosing the time T and detuning δ allows us to adjust the phase to
some standard values θ = −T /2 = π/2, π/4 or π/8 mod 2π, which correspond
to the Z, S, and T gates:
     
1 0 i π2 σ z 1 0 i π4 σ z 1 0 i π8 σ z
Z= =e ,S= ∼e ,T = iπ /4 ∼ e .
0 −1 0 i 0 e
(6.10)
Alternatively, we can apply a coherent drive (6.6) to the qubit, implementing
 a
rotation around the axis n ∝ (ε0 cos(ϕ),ε0 sin(ϕ),δ), at a frequency ω = 12 δ 2 + ε02 .
In particular, working on resonance δ = 0 with ϕ = 0 or π/2 produces rotations
around σ x or σ y , respectively. The combination of a rotation around σ y with a phase
creates the ubiquitous Hadamard gate:
 
iπ σ z /2 iπ σ y /2 1 1 1
H =e e =√ . (6.11)
2 1 −1

6.1.5 Decoherence and Dephasing

Experiments never create pure states, because of the unavoidable imperfections and
limited precisions – both in the creation of the initial states and in the control of
the experiments – as much as because of the interaction between our setup and the

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.1 What Is a Qubit? 113

environment. Instead, experiments must be described using mixed states: averages

of different pure states prepared and evolve with random or fluctuating conditions,
as sketched in (2.23). This averaging causes decoherence,2 the progressive loss of
the quantum fluctuations, due to the destruction of quantum superpositions and
entangled states.
The simplest type of decoherence is dephasing, a random scrambling of the rel-
ative phase between the eigenstates of our artificial atom – or any other quantum
system. Dephasing is normally caused by fluctuations in the eigenenergies of our
system, which deviate from the ideal values En by small amounts that randomly
fluctuate in time δEn .
Obvious sources of dephasing are the control lines used to stabilize and measure
qubits and cavities. Those cables and antennas may carry fluctuating electromag-
netic fields, white noise that affects the energies of the superconducting circuit.
A similar scrambling is caused by electric charges trapped on the substrate of the
superconducting chip. These charges act as two-level systems or quantum fluctua-
tors that couple to the qubit, shifting their energy levels, typically over longer time
scales and sometimes in a quasistatic fashion – i.e., the electric field induced by the
charges is static over the duration of an experiment, but may change from realization
to realization.
Pure dephasing can be modeled as a Brownian motion of the phase, which accu-
mulates random displacements due to the fluctuations in the qubit’s energy levels,
H  ∼ 12 (h̄ + δE)σ z . In the absence of other perturbations, a single realization of
the experiment is accounted for by a stochastic equation for the phase:
d
|n → e−iϕn (t) |n ,
ϕn = En + δEn (t).
h̄ (6.12)
dt
This equation by itself is somewhat useless, because we do not know the noise
realization δEn (t) of each experiment. However, we may incorporate our ignorance
of the noise, modeled as some probability distribution p(δEn,t), in an ensemble
average description of the experiment using mixed states.
Theoreticians like to work with white (or slightly colored) Gaussian noise mod-
els, which are uncorrelated in time δEn (t)δEn (t  ) = h̄2 σ 2 δ(t − t  ). In this limit
equation (6.12) describes a Brownian particle, with a Gaussian probability distribu-
tion of the phase centered on En t/h̄, with linearly growing variance σ 2 ∝ t. The
evolution of a quantum state ρ(0) is computed as a linear transformation:
 −(ϕ−t)2 /(2σ 2 t)
1 1 ze
ρ(t) = εt (ρ(0))  e−i 2 ϕσ ρ0 ei 2 ϕσ
z
√ dϕ, (6.13)
2πσ 2 t

2 In recent years, a resource theory has been put forward to quantify coherence (Baumgratz et al., 2014) as a
function of the off-diagonal elements in the density matrix. From this theory, it follows that decoherence is the
destruction of such a resource. We are not so rigorous in our description.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 114 Superconducting Qubits

that belongs to the family of a completely positive maps or quantum channels

(see Section 8.4.2). For a white noise model, the integral has an analytical expression
 
ρ11 e(−i−γφ )t ρ10
ρ(t) = . (6.14)
ρ01 e(+i−γφ )t ρ00

The map predicts an exponential decay of the coherences – the off-diagonal ele-
ments of the density matrix – with a dephasing rate γφ = σ 2 /2, and a dephasing
time T2 ∼ 1/γφ .
The dephasing channel (6.14) is reproduced by the dephasing master equation:
d i γφ
ρ = − [H,ρ] + (σ z ρσ z − ρ). (6.15)
dt h̄ 2

The Hamiltonian H = 12 σ z accounts for the noise-free evolution, while the Lind-
blad operator L[ρ] = σ z ρσ z − ρ models the destruction of coherences.
The dephasing master equation can be derived – as we did for the cavity environ-
ment coupling – from a microscopic model where the energy perturbations δEn (t)
arise from a diagonal coupling with a bath, provided the spectrum of fluctuations
in the bath’s Markovian limit reproduces the desired white noise model over all fre-
quencies. However, the white noise model of dephasing is a bit simplistic. Electrical
circuits – and mesoscopic systems in general – are affected by what is known as
1/f -noise or slow noise. When we study the power spectrum of these fluctuations,
the contribution of noise diverges at low frequencies. This implies a slow decay of
noise correlations – e.g., δE(t  )δE(t) ∼ exp(−(t  − t)/Tc ) with a large Tc that
prevents self-averaging. The limit in which the correlation time diverges Tc → ∞
describes a quasistatic noise, which remains more or less constant throughout each
experimental realization, but fluctuates between runs. Such models are described
with variants of (6.13), with other decay forms that depend on the spectral properties
of the noise (Ramos and García-Ripoll, 2018), and do not have a Markovian master
equation (see Exercise 6.3).
However, even if slow fluctuations are theoretically inconvenient, they may be
beneficial, because we can design controls to suppress them. Assume for instance,
that the noise δEn has long time correlations. There is a technique, called spin-echo,
that exactly cancels this fluctuation, and which consists in introducing a spin flip
halfway through the experiment:

ρ(t) = e−iδEt/h̄ σx e−iδEt/h̄ ρ(0)eiδEt/h̄ σx eiδEt/h̄ = ρ(0). (6.16)

If the noise is not quasistatic, but has a finite-time correlation T , we can still apply
spin-echo at regular intervals with a spacing shorter than the fluctuations, time scale
δt  T . This leads to a total or partial cancelation of noise, extending the lifetime

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.1 What Is a Qubit? 115

of quantum superpositions well beyond the intrinsic T2 time of the experiment –

see, for instance, the work of Yan et al. (2015) with flux qubits, and Problem 6.3.

6.1.6 Relaxation and Heating

Short of losing the qubit itself, the worst type of decoherence is the one that scram-
bles the probabilities of the basis states ρii . As in the photonic case, we find such
processes in the interaction between a qubit and a Markovian environment. In this
case, the incoherent evolution is described by the qubit’s equivalent of a cavity’s
master equation in Section 5.5.4
d i γ
ρ = − [H,ρ] + [n̄ + 1] 2σ − ρσ + − σ + σ − ρ − ρσ + σ − (6.17)
dt h̄ 2
γ
+ n̄ 2σ + ρσ − − σ − σ + ρ − ρσ − σ + .
2
This master equation can be derived from a microscopic model of a qubit interact-
ing with a memoryless environment, with a continuum spectrum that is populated by
a thermal distribution of photons n̄ω . The master equation contains relaxation events
σ − |11| σ + that take the qubit from the “excited” state |1 to the lower energy state
|0. It also includes the converse process, or heating σ + |00| σ − → |11|. The
ratio between both processes is dictated by the temperature of the environment:
γheat n̄
= ∝ e−/kB T . (6.18)
γcool n̄ + 1
When the qubit’s gap is larger than the effective temperature, we can neglect
heating n̄  0 and stay with pure relaxation. The exact solution of the master
equation in this limit is
 ∗ 
ρ11 e−t/T1 e−it−t/T2 ρ10
ρ(t) = ∗ . (6.19)
ρ01 e+it−t/T2 ρ00 + (1 − et/T1 )ρ11
Note how the excitation probability decays over a time T1 = 1/γ , while coherences
disappear over a slightly longer timescale T2∗ = 2T1 . If an experiment combines
dephasing and relaxation, the first timescale remains the same, γ = 1/T1 , but the
decoherence time is reduced (see Problem 6.5):
1 1 1
∗ = + , (6.20)
T2 2T1 T2
and denoted by the symbol T2∗ . In this scenario, to avoid confusion, 1/γφ is also
often referred to as Tφ instead of T2 .
It is important to remark that the losses that we are discussing are not incompat-
ible with the superconducting nature of the circuit. While the circuit has negligible

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 116 Superconducting Qubits

resistance, it can still dissipate energy to the environment, to the substrate, or to

other circuit elements – antennas, qubits, resonators – on the same chip. Some of
this radiation can be suppressed by placing the qubits in better environments –
e.g., enclosing them in the 3D cavities from Section 5.3 – to get qubit lifetimes
T1 ∼ 100 μs. Another way to improve the lifetime of the qubit is to use better
fabrication processes with different materials or techniques that minimize the imper-
fections and probability of trapping charges. Recent experiments by Place et al.
(2021) have reported lifetimes of T1 ∼ 300 μs this way. Assuming qubit operation
rates of 100 MHz, this is a very long lifetime that allows for 30 000 operations before
losses kick in!

6.2 Charge Qubit

After this formal introduction to the world of qubits, we can study our first arti-
ficial atoms. We begin with the charge qubit, a small, micro- or nanometer-size
metallic island. The island is so small that the energy grows very rapidly with
every new Cooper pair we add or subtract from it. This anharmonic growth, known
as a Coulomb blockade, allows us to engineer a very good qubit eigenspace that
can be easily controlled by electrostatic potentials. However, as we will discuss in
this section, this facility of control makes the qubit very fragile to dephasing by
interaction with the stray electrostatic field and trapped charges.

6.2.1 Coulomb Blockade

Picture a small piece of metal with a lattice of positively charged ions and a sea of
electrons binding them. In equilibrium, the metal aims toward charge neutrality, a
state in which the density of positive charges balances the negative ones. This state
has some reference energy E0 . If we add one more electron to this island, we have
to pay a small energetic cost  = E1 − E0 , which can be explained by the Coulomb
interaction of the quasiparticle with the negatively charged electron plasma and the
positively charge ion lattice. Because the average charge of the metal is zero, and
because of screening effects in the electron plasma, this interaction is not too large
and causes a moderate gap .
The situation changes if we add a second charge. This new particle does not face
a balanced material, but interacts repulsively with the charge introduced before.
The increase in charging energy Coulomb >  due to this repulsion is called the
Coulomb blockade. Because of this blockade, there exists a potential qubit space
form by the states with zero |0 and one extra charge |1, with a small splitting
among them ω01 ∼ , and separated by a large gap ω02 =  + Coulomb from
higher-charge excitations.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.2 Charge Qubit 117

The Coulomb blockade is an energy scale that depends on properties of the con-
ducting material, such as the density of charges and the geometry of the piece.
It predates the notion of qubits and the study of quantum information: Coulomb
blockade is observed in quantum dots, in transport experiments with molecules,
and, of course, also in the superconducting world.
Fortunately, we do not need to worry about microscopic models to describe a
charge qubit. Since we work with macroscopic devices, with a humongous number
of particles, we can treat the superconducting island as a lumped-element circuit,
describing the total energy as a quadratic function of the total charge, without caring
for how this charge distributes across the island. The work or energy required
to charge the island is given by the self-capacitance of this metallic object, as
expected:
1
E= Q2 . (6.21)
2Cself
If our excitations are electrons with undivisible charge e, we can define a constant
called the charging energy EC = e2 /2Cself and express the total cost in term of the
number of carriers n = Q/e, as in E = EC n2 .
Note the anharmonicity in this expression: The cost of adding one particle is
E1 − E0 = EC , but the second charge involves a larger penalty E2 − E1 = 3EC .
We could define our qubit using the Hilbert space of the neutral state and the state
with a single Cooper pair, |0 and |1. The total Hamiltonian then would read
EC z
H = EC |11| =σ + constants. (6.22)
2
This situation is inconvenient for several reasons: The splitting EC is too large for
a small island, we have no means to modify this splitting, and we also do not yet
know how to add other control terms to the Hamiltonian, such as a dipolar coupling
with σ x = |10| + |01| (see (6.2)). All this requires a slightly more sophisticated
circuit model.

6.2.2 The Actual Superconducting Charge Qubit

The actual superconducting charge qubit represents an upgrade from the ideal super-
conducting island to a more sophisticated device, shown in Figure 6.2a, with the
equivalent circuit analyzed in Section 4.6. The first new ingredient is the capacitive
coupling Cg with a conductor – light blue metallic piece in Figure 6.2a – that can
transport a microwave, represented as an electrostatic potential V (t).
The second ingredient is a pair of Josephson junctions that connect the qubit to
a ground plane, allowing the injection and extraction of charges. The two junctions
form a small dc-SQUID, which is threaded by a small magnetic flux . As it was

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 118 Superconducting Qubits

E
C

C V
10 μm 500 nm

100 μm

Figure 6.2 (a) Superconducting island coupled to the ground plane through two
Josephson junctions. (b) Zoomed-in picture of one junction. Pictures courtesy of
Andreas Wallraff, ETH, Zurich. (c) Reminder of the equivalent circuit discussed
in Section 4.6, with the superconducting island marked in gray. The two junctions
add up to a single effective junction with nonlinear inductance that depends on the
flux  enclosed by the loop. (d) Image of a transmon qubit, a charge qubit where
the capacitance has been dramatically enlarged.

explained in Section 4.7.2, the parallel-junction circuit is equivalent to a Josephson

junction EJ () that can be tuned with the magnetic flux .
Combined together, the external potential and the SQUID give the lumped-
element circuit from Figure 6.2b and the effective nonlinear Hamiltonian:
1 2
H = q − qg − EJ cos(φ/ϕ0 ). (6.23)
2C
The constant C is the total capacitance of the the charge qubit, including the
coupling to the gate voltage, the junction CJ , and the island itself Cself . In practice,
we cannot tell CJ and Cself apart, Cσ  CJ + Cg . In addition to the dc-SQUID, the
external gate potential V introduces our second tuning element, the offset charge
qg = −Cg V , which determines the equilibrium state of the island.
The charge qubit is best described using the number state representation from
Section 4.9. The states {|n} are labeled by the excess (n > 0) or defect (n < 0)
of Cooper pairs with respect to charge neutrality (n = 0). The charge operator is
diagonal in the number basis q̂ |n = −2en |n , n ∈ Z, and the electrostatic energy
is proportional to the charging energy EC :
2e2
En = (n − ng )2 = 4EC (n − ng )2 . (6.24)
C

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.2 Charge Qubit 119

Figure 6.3 Energy levels of the charge qubit without the Josephson junction
(dashed) and with the tunneling amplitude EJ  0.1EC (solid). At zero voltage,
the lowest eigenstates are |0 and |1. At ng = 1/2, both states become almost
degenerate but the tunneling creates new eigenstates that are superpositions of the
former |± = √1 (|0 ± |1).
2

Figure 6.3 shows the energy levels En of the charge states in dashed lines. Around
the point ng = 12 , states |0 and |1 become degenerate3 and well separated from all
other charge states, n = 2,3, . . . . This makes them good candidates for qubit states,
provided we find ways to reproduce the qubit Hamiltonian (6.2) using the external
controls  and V .
The first control is given by the external potential. If we perturb the value of the
offset charge slightly away from ng = 12 , we obtain an energy splitting between
the 0 and 1 charge states that grows as h̄ε = 8EC (ng − 12 ). The second control is
due to the nonlinear inductance of the SQUID. This effective Josephson junction
enables the tunneling of charges in and out of the island (4.63), cos(φ̂/ϕ0 ) =
1
2 n (|n + 1 n| + |n n + 1|). Assuming that the capacitive interaction is the
dominant energy scale EC > EJ , and that we are close to degeneracy ng  1/2,
the Josephson energy may be truncated to the 0 and 1 charge states, as −EJ ()σ x
with the tunneling operator σ x = |10| + |01|. Altogether, we obtain the desired
qubit Hamiltonian in a 2 × 2 matrix form:
(
h̄ε z h̄ x h̄ε = 8EC ng − 12 ,
H = σ − σ , with (6.25)
2 2 h̄ = EJ ().

3 There are other points n = 3 , 5 , . . . , where different pairs of states become degenerate. However, the
g 2 2
physics of those points is similar to that which we discuss later and presents no formal advantage.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 120 Superconducting Qubits

As explained in Problem 6.6, this approximation can be verified in a self-consistent

way, using perturbation theory to study the corrections that appear when we take
into account the coupling between the qubit subspace and higher excited states
|2,|−1,|3,|−2, . . ..

6.2.3 Qubit Hyperbola

We can diagonalize the qubit Hamiltonian by writing it as H = 12 Eσn with a Pauli
matrix operator σn = n · σ along a direction on the XZ plane, n = (cos θ,0, sin θ).
The overall prefactor

E = ε2 + 2 (6.26)
gives the energy ± 12 E of the ground and excited states,

|0n  = cos(θ/2) |0 − sin(θ/2) |1 , and (6.27)

|1n  = sin(θ/2) |0 + cos(θ/2) |1 .

The energy level structure of the qubit is usually referred to as the qubit hyper-
bola, because E0 and E1 satisfy the equation of a hyperbola with asymptotes at
E0,1  ± 12 |ε| when the applied potential ε is very large.
The qubit hyperbola is symmetric with respect to the symmetry point ε = 0,
where the separation between energies is minimal E ≥ . In absence of Joseph-
son energy,  = 0 and the two charge states |0 and |1 would have the same
energy (cf. Figure 6.3 and dashed line at ng = 1/2). However, the introduction
of quantum tunneling has activated quantum fluctuations of the charge and it has
broken the degeneracy, introducing a minimal separation . This separation causes
the new eigenstates of the problem to be the symmetric and antisymmetric charge
superpositions:
1 1
|± = √ |0 ± √ |1 . (6.28)
2 2
The symmetry point ε = 0 is experimentally relevant, as the configuration in
which the qubit experiences the least dephasing. Indeed, if the qubit is surrounded
by stray electric fields that introduce random contributions δV , these contributions
cancel to first order ∂E/∂V = 0 at ε = 0. Thus, the energy levels fluctuate the
least around this point, and the lifetime of our quantum states is maximized. The
notion of a symmetry point for safe operation is a very general concept that appears
also in the flux qubit. In Section 6.3 we will push this idea further, ensuring that the
energy levels become insensitive to perturbations in the electrostatic field, which
gives us the transmon qubit design.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.2 Charge Qubit 121

Finally, a remark on the notational discrepancy between our charge qubit Hamil-
tonian (6.25) and the general qubit model (6.2). This difference does not affect the
physics, as both models are related by a unitary transformation that swaps the σ x
and σ z operators (see also Problem 6.1). This transformation is a relabeling with
new qubit states:

|0̃ = |− h̄ z h̄ε x
⇒ H = σ̃ + σ̃ . (6.29)
|1̃ = |+ 2 2

The diagonal Pauli matrix is now σ̃ z = |1̃1̃| − |0̃1̃|. Some works favor this nota-
tion, because only the states |ĩ – not the |i – can be experimentally detected,
especially when working with qubits inside cavities (cf. Section 7.4.2).

6.2.4 Charge Qubit History

The idea of a small capacitance junction as an implementation of a qubit was put
forward by Shnirman et al. (1997), contemporarily with other proposals of universal
quantum computers – such as trapped ions by Cirac and Zoller (1995), quantum
dots or impurities in solid-state devices by Loss and DiVincenzo (1998), or nul-
cear magnetic resonance (NMR) experiments by Gershenfeld and Chuang (1997)
and Cory et al. (1997). When Shnirman’s work was published in 1997, solid-state
experiments were still behind AMO4 experiments in terms of control and coherence,
but that was soon to change. Contemporary works by Nakamura et al. (1997) and
Bouchiat et al. (1998) validated the two-level model (6.25) for small superconduct-
ing islands. These experiments resolved spectroscopically the eigenenergies of the
island, confirming the quantization of the charge and the qualitative pictures from
Figure 6.3.
Shortly thereafter, Nakamura et al. (1999) provided full evidence of the qubit
structure, engineering quantum superpositions of charge states and implementing
single-qubit rotations. In this work, the authors prepared a well-defined charge state,
in the limit of large external field |ε0 |  ||. They then jumped abruptly to the
symmetry point ε = 0, waited for a short time t, and returned to large field limit, to
measure the charge state. As explained in Problem 6.7, the final state of the qubit
was a quantum superposition cos(t/h̄) |0 + sin(t/h̄) |1, where the angle t/h̄
depended both on the time spent by the qubit at the symmetry point as well as on
the magnetic flux passing passing through the SQUID,  = EJ cos(/2ϕ0 ).
The work by Nakamura is an important milestone in the development of solid-
state quantum information processors. It is arguably one of the works that changed

4 Atomic, molecular, and optical.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 122 Superconducting Qubits

people’s opinion regarding the future of the field, suggesting that it was possible
to engineer complex quantum states and quantum registers (Makhlin et al., 2001).
Later experiments with similar schemes demonstrated two-qubit Hamiltonians and
qubit–qubit interactions (Pashkin et al., 2003; Yamamoto et al., 2003), a crucial
ingredient for scalable computations.
However, Nakamura’s charge qubit implementation was inadequate for scalable
computations. The superconducting islands decohered very quickly T1 ∼ 2 ns,
because of the surrounding antennas and probes that controlled and measured the
qubit. To compensate for this, the qubits were engineered with huge energy gaps,
EJ  84 μeV ∼ 20 GHz, and operated very rapidly, and even then only allowed
for a few oscillations before losing all quantum properties.
In 2004, the Yale group had the idea of fabricating a charge qubit inside
a microwave resonator. As we have seen in Chapter 5, microwave cavitives
only absorb excitations at discrete frequencies, ωn = n × ω0 , filtering out other
electromagnetic fields. By placing the qubit inside the cavity and ensuring that
the qubit’s energy gap E was very different from the allowed photons, they
could effectively isolate and protect the qubit from most of the electromagnetic
fluctuations. The idea of Purcell filtering the environment was put forward in the
theory work by Blais et al. (2004) and immediately demonstrated by Wallraff et al.
(2004) in an experiment that showed times of 200 ns (see also Schuster et al.,
2005). That was orders of magnitude better than the bare charge qubit demonstrated
only a few years before.
The conditions of relative isolation provided by the cavity, together with oper-
ating the qubit close to the symmetry point, made it possible that the coherence of
the superconducting qubit was no longer limited by relaxation but by environment-
induced dephasing. In Yale’s charge qubits, dephasing rates of γφ ∼ 2π × 750 kHz
or T2 ∼ 1μs, allowed testing many quantum optics ideas, as well as some primitive
quantum information protocols. These were wonderful beginnings with fast-paced
developments in the study of the Rabi model, strong coupling, Wigner function
tomography, and large entangled states. However, developing scalable quantum
computers required improving the quality of the qubits even further. The avenue
that has proved most fruitful was to further suppress dephasing, reducing the
curvature of the qubit hyperbola even further. This idea led to the transmon
qubit in 2007, a significant redesign of the charge qubit’s parameter, which
we now discuss.

6.3 Transmon Qubit

The charge qubit dephasing around the symmetry point is dominated by second-
order fluctuations of the energy levels with respect to the fluctuating fields

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.3 Transmon Qubit 123

∝ ∂ε22 E. These corrections can be expressed in terms of perturbations of the

charge offset δng :
 
1 EC 2
E = E(ng = 0) + EJ δng + O δn2g . (6.30)
2 EJ
A strategy to reduce these corrections is to increase the ratio EJ /EC . If we move
from EJ /EC  1 for a charge qubit to a new setup with a ratio EJ /EC  50,
our crude estimate predicts a reduction of the dephasing terms by four orders of
magnitude 0.0004. The transmon is a design of a qubit that achieves this larger
ratio by shunting5 the Josephson junction of the charge qubit with a large capacitor,
as shown in Figure 6.2d. The increase in the capacitance reduces EC and leads to
an exponential reduction in the dephasing terms. The result is a simpler and larger
qubit, easier to fabricate and with greater reproducibility and stability.

6.3.1 Moving Particle Picture and Energy Bands

To understand why increasing EJ /EC decreases the sensitivity to electrostatic fields,
we have to reinterpret the charge qubit Hamiltonian (6.23) using the number-phase
representation from Section 4.9:
1 2
H = 4EC (−i∂ϕ − ng )2 − EJ cos(ϕ) ∼ p + V0 cos(2πx/a). (6.31)
2m
This model resembles a particle with position ϕ moving in a periodic potential
−EJ cos(ϕ), with mass m ∼ h̄2 /8EC . By analogy with the Bloch theory for elec-
trons in a solid, we derive three qubit regimes depending on the relative strength of
the potential, EJ /EC :
(1) When EJ  0, we have an isolated superconducting island. The eigenstates
of the problem have a well-defined number of charges n, which we associate
to plane waves6 ψn (ϕ) ∝ exp(−inϕ). The eigenenergies follow a parabolic
dispersion En = 4EC n2 , reminiscent of the kinetic energy of a free particle (cf.
Figure 6.4a, dashed line).
(2) A weak periodic potential EJ < EC breaks the degeneracy at the symmetry
points ng = 1/2,3/2, . . . by an amount proportional to EJ . From Figure 6.3a we
see that the original parabolas are transformed into Bloch bands En (ng ). These
bands match those of a particle moving in a periodic potential with momentum

5 Shunting means providing an alternative (parallel) path for the charges to flow. The transmon results
from capacitively shunting a Josephson junction, with one or two capacitors that are connected in
parallel to it.
6 A maximal certainty of the quasimomentum n corresponds to the maximum uncertainty of the position ϕ.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 124 Superconducting Qubits

Figure 6.4 (a) Rescaled energy levels of a transmon qubit as a function of the
offset charge ng , for ratios EJ /EC = 0.02 (dashed), 2 (dot-dashed), and 16
(solid). The values plotted lack an irrelevant constant and are scaled to the qubit
gap with no offset charge εn = (En (ng ) − E0 (0))/h̄ω10 . (b) For a transmon with
EJ /EC = 16, we plot the energy levels relative to the ground state εn (0), together
with the nonlinear inductive potential EJ cos(ϕ)/h̄ω01 . On top of each level we
plot the eigenfunctions, shifted and scaled arbitrarily. (c) Relative anharmonicity
of the transmon qubit αr = ω12 /ω01 − 1. (d) Energy levels of a transmon when we
can tune the Josephson energy with a dc-SQUID, as EJ () = EJ cos(/ϕ0 ), for
EJ (0)/EC = 50.

n + ng . The value of n is not the charge, but the index of the band, and ng
represents the quasimomentum of the Bloch wavefunction (cf. Figure 6.4a,
dot-dash line).
(3) When the confinement is deep, one or more bands fall within the periodic poten-
tial EJ  4EC n2 . The splitting between neighboring bands is so large that
they become flat. This limit corresponds to a particle hopping among the min-
ima of the inductive potential cos(φ). The width of the band |En (ng = 1/2) −
En (ng = 0)| is the tunneling amplitude between wells, which decreases expo-
nentially with the barrier height V0 ∼ EJ . In other words, when EJ /EC is very
large, the sensitivity of En (ng ) to the value of ng is exponentially small com-
pared to the original charge qubit (cf. Figure 6.4a, solid line). In this limit, we
can ignore the bias or set it to ng = 0.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.3 Transmon Qubit 125

The transmon qubit implements the third limit, in which EJ /EC  50 and three
or more bands fall inside the Josephson inductive potential (cf. Figure 6.4b). This
ratio decreases exponentially the sensitivity to electrostatic fluctuations, which is
extremely good for the qubit lifetime, but this advantage comes at a cost: It results
in a qubit with a reduced anharmonicity.

6.3.2 Transmon as Anharmonic Oscillator

In the transmon limit, the lowest-energy bands are deep inside the cosine potential.
We can approximate the eigenstates by the eigenfunctions of a slightly anharmonic
resonator:
1 2 1 EJ 2 1 EJ 4
H  q̂ − qg − EJ + 2
φ̂ − φ̂ + O φ 6 . (6.32)
2C 2 ϕ0 24 ϕ04
We diagonalize the quadratic part for qg = 0 using Fock operators:
 
EJ 1/4 i
q̂ = −2en̂ = 2e √ (â † − â), (6.33)
8EC 2
 1/4
8EC 1
φ̂ = ϕ0 √ (â † + â).
EJ 2
The quadratic part becomes an oscilator with equispaced energy levels and a plasma
√
frequency h̄ω01  8EC EJ . We include the external field qg q̂/C and the Duffing
term φ 4 , working up to second order in perturbation theory. The result is a nonlinear
oscillator:
H  h̄ω01 â † â − h̄α â † â † â â + i h̄ε(â † − â), (6.34)
with a small anharmonicity α  ω01 and a driving term ε(t) that includes qg with
all other transmon parameters.
In this nonlinear oscillator, the first and second excitations |1 and |2 are
separated by a gap ω12 := ω01 − α that is slightly smaller than the splitting
between the ground and first excited state ω01 . The spacing between energy levels
gets even smaller as we climb up the ladder, because the cosine potential gets softer
at higher energies, making it cheaper to continue exciting as we move away from
the ground state. Once more, perturbation theory gives an estimate of the relative
anharmonicity:

α ω01 − ω12 8
αr = =  . (6.35)
ω01 ω01 EJ /EC
For a typical ratio of EJ /EC = 50, the anharmonicity is about 5% the transmon’s
fundamental frequency. For a qubit with ω01 ∼ 2π × 6 GHz, this value can be
∼ 2π × 300 MHz.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 126 Superconducting Qubits

The charge operator q̂ plays the role of the dipole moment of the transmon,
connecting neighboring states that differ in parity – i.e., it allows transitions from
0 to 1, 1 to 2, . . . . This operator is explicitely coupled in (6.34) to a field ε(t) that
contains the gate voltage ng or qg . The source of the field ε can be a classical,
coherent pulse, which we treat as a complex, time-dependent pulse with the tools
from Section 6.1.3. However, the field may also have a quantum origin and be
generated by a resonator, a cavity, or a waveguide, giving rise to the cavity-QED
setups in Chapter 7.
It may seem counterintuitive that, since the transmon bands are flat and indepen-
dent of ng , an electric field may have any kind of effect on the system. However,
notice that the linear term can be approximately absorbed with a unitary redefinition
of the Fock operators, â → â +iε/ω01 . This reveals that, up to first order, the energy
levels really do not depend on ng , but this external field may induce a unitary change
in the transmon eigenstates ψn (ϕ). In other words, starting from our model

H = E1 |ψ1 (ng )ψ1 (ng )| + E0 |ψ0 (ng )ψ0 (ng )| , (6.36)

changes in the external potential lead to a perturbation of the form

 !  ! 
H (ng + δng )  H + δng ∂ng ψ1 ψ1 + ∂ng ψ0 ψ1 + H.c. . (6.37)

If we project these corrections back onto the basis {|ψ0 (ng ) , |ψ1 (ng )}, we recover
the term proportional to q̂ or i(â † − â).

6.3.3 Josephson Junctions and the Mathieu Equation

The properties of the transmon qubit may be obtained analytically, by solving the
eigenvalue equation for the Josephson junction Hamiltonian (6.31), with the bound-
ary condition ψ(ϕ+2π ) = ψ(ϕ). We get rid of ng changing the boundary condition,
with a phase induced by the “gauge field” ng :

ψ(ϕ) = eing ϕ ξ (ϕ). (6.38)

The new function ξ is not periodic ξ (ϕ + 2π ) = e−ing 2π ξ (ϕ) and satisfies

 
−4EC ∂ϕ2 − EJ cos(ϕ) ψn (ϕ) = εn ψn (ϕ). (6.39)

This equation is related to the Mathieu differential equation, which in standard

notation (Olver et al., 2010) reads

d2
u(z;q) + [a − 2q cos(2z)]u(z;q) = 0. (6.40)
dz2

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.3 Transmon Qubit 127

Given a value ν, the Mathieu equation admits a complete set of orthonormal solu-
tions labeled by nonnegative integer k such that
 π
uk (z;q)uk (−z;q)dz = δkk (6.41)
0
uk (z + π ) = ei(ν+2k)π uk (z), n ∈ {0,1,2, . . .}. (6.42)
These solutions are the Floquet functions uk (z) = meν+2k (z;q), where ν + 2k is the
Floquet characteristic exponent of the function uk (z) after a period π . Note that these
functions have the required boundary condition if we identify 2zϕ. This suggests
that we rescale and shift the phase variable ϕ = 2z + π , with which 2∂ϕ = ∂z ,
so that
ε EJ
∂z2 + − cos(2z) ξ (ϕ) = 0. (6.43)
EC EC
From here it follows the identification of physical quantities to Mathieu equation
parameters
ε EJ
a= ,q= , and ν = −2ng (6.44)
EC 2EC
and the eigensolutions and eigenenergies
 
eing ϕ ϕ + π EJ
ψn (ϕ) = √ me−2ng +k(n,ng ) ; , (6.45)
2 2 2EC
En = MA (r,q), (6.46)
where MA (r,q) is the Mathieu characteristic function for even functions. It is cus-
tomary to define the index k as a function of n and ng that produces ordered energies:
En ≥ Em for all n > m. The function for implementing this ordering is rather con-
voluted. An expression that works for all values except 2ng ∈ Z is7
1  
k(n,ng ) = −round ng + (−1)floor(ng ) −1 + (−1)n (1 + 2n) . (6.47)
4
Here round(x) is the nearest integer to the real number x, and floor(x) is the integer
immediately below x. The eigenenergies are given by the Mathieu characteristic
function MA (r,q) with the preceding identifications.
These formulas allow us to compute the eigenenergies of the transmon qubit
for different ratios of EJ /EC , as shown in Figure 6.4a. Problem 6.11 shows that
we also have access to the wavefunctions, a fact that is useful for practical studies
of transition matrix elements and coupling strengths. Figure 6.4b illustrates some
eigenstates for a ratio EJ /EC = 16, where the two lowest bands are already deep

7 The formula in Koch et al. (2007) is wrong.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 128 Superconducting Qubits

in the noninductive potential. Notice how ω12 , the excitation energy from the first
excited level, becomes very similar to ω01 , the qubit frequency. In this case, the
anharmonicity is α = 8.2%.

6.3.4 Transmon as Qubit

The transmon Hamiltonian has two controls: We may tune the plasma frequency
ω01 using the dc-SQUID, and we may couple the transmon to an external potential.
Since the first control is simple to understand, let us instead focus on the influence
of a periodically modulated electric potential. We consider a pulse

ε(t) = w(t) cos(ω01 t + φ), (6.48)

with a carrier frequency ω01 and a smooth envelope w(t) that switches the pulse on
and off over a time T . The Fourier components of this pulse spread approximately
over a frequency band ω01 ± 2π/T . Thus, we expect that for long enough pulses,
2π/T  α, our external drive will only be able to couple the states |0 and |1,
suppressing leakage outside the qubit space {|0 , |1} to unwanted states |2,|3, etc.
In the limit of finite bandwidth controls and no leakage, we can project the nonlin-
ear oscillator model onto the qubit manifold spanned by the ground and first excited
state. Up to constants
1
H = h̄ω01 σ z + h̄εσ y . (6.49)
2
This is essentially our qubit model (6.2) up to an irrelevant change of basis σ x → σ y
(see Problem 6.1). Following Section 6.1.4, the transmon is a fully controllable
qubit, in the sense that oscillating electromagnetic fields ε(t) with tunable ampli-
tude, frequency, and phase can be used to implement any qubit operation – even
without tuning the gap ω01 .
As in the charge qubit, frequency tunable transmons rely on a dc-SQUID with
some magnetic flux  to control the effective value of the Josephson junction
EJ () = EJ (0) cos(/ϕ0 ) and adjust the qubit gap (cf. Figure 6.2d). In particular,
the external flux can only decrease the plasma frequency from a fixed, maximum
√
value approximately as ω01 () = ω01 (0) cos(/ϕ0 ).
Frequency tuning finds its use in bringing the transmon qubit on-resonance to
other circuits, such as resonators and other qubits, to implement quantum operations
and gates (e.g., Section 8.3.3). There are two limitations in this type of control.
The first one is that we cannot decrease EJ () too much, because then the ratio
EJ /EC becomes so small that decoherence kicks in and spoils the setup. The second
problem is that the Fock operators and the basis states implicitely depend on the
value of the Josephson energy – see (6.33). Nonadiabatic changes of the Josephson

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.3 Transmon Qubit 129

energy must be taken into account by considering the explicit time dependence of
the Hamiltonian, via ∂t â,∂t â † and induces corrections into the qubit model.
Finally, regarding the detection of the qubit, there are no direct measurement
schemes for the transmon. Instead, transmon qubits will be typically connected to
one or more resonators, which we will use to indirectly probe the state of the qubit
in a nondemolition fashion. Those methods, which are general to all types of qubit,
are discussed in Section 7.4.3.
The transmon, in any of its incantations, is the most used qubit in all kinds
of superconducting experiments, from those working with photons and qubits to
reproduce quantum optics, to quantum information experiments with tens of qubits
by companies such as IBM, Google, or Rigetti Computing. It is a very popular type
of qubit for obvious experimental reasons. First of all, it is a large qubit, with a
simple design and easy to fabricate – experimentalists can basically cut and paste
masks from one lab to another. Second, the qubit has a very reproducible design,
because its properties are not extremely sensitive to the underlying parameters. For
√
instance, the plasma frequency goes as 8EJ EC , so that small deviations in the
Josephson energy – due to fluctuations in the junction’s thickness or size – appear
as small corrections of 5% in the final qubit. Finally, transmon qubits have very
competitive decoherence times, ranging from T2 ∼ 20 μs up to 0.3 ms – and maybe
better, by the time you read this book.
Interestingly, for a transmon bigger is better. We can increase the coherence
time, simplify fabrication and reproducibility, and enhance the Josephson energy
by making the samples larger. This realization led to the development of the three-
dimensional transmon by Paik et al. (2011), which is nothing but a very large trans-
mon, with a size of millimeters, fabricated not on a chip, but on a portable sapphire
substrate. This device is a tiny object that can be moved around and inserted into dif-
ferent experimental setups, including 3D cavities. Because the transmon frequency
can be verified outside the experiment through simple resistive measurements,8 and
because transmons are mobile and can be manually interchanged, the practical need
for tunability may disappear from experiments, opening the door to setups with 3D
transmon qubits that are extremely long lived, with T2∗  0.092 ms in state-of-the-
art setups (Rigetti et al., 2012).
On the negative side, the operation speed of the transmon is ultimately limited
by its anharmonicity. For a qubit gap ω01  6 GHz, anharmonicity α  300 MHz
and T2 = 40 μs, a rough estimate would set the limit on 103 single-qubit rotations
within the qubit’s lifetime. That is reasonably large for to-date algorithms, but it may
fall short for large-scale computations. The situation gets even worse if we consider

8 The resistance of the junction at room temperature gives an idea of E , while E can be well estimated from
J C
the original design due to geometric considerations.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 130 Superconducting Qubits

the interaction speed of the transmon with other circuits, such as cavities and other
qubits. In this limit, we can find operation times of 4−40 MHz, which further lower
the size accurate quantum computations. In the following chapter, we will meet the
flux qubits, a design that solves both problems, enlarging their anharmonicity and
allowing for stronger couplings to other devices.

6.4 Flux Qubit

The charge qubit and the transmon are defined by the distribution of charge accu-
mulated on an effective capacitor, which may be a superconducting island or some
large superconducting plates. Flux qubits are instead defined by a superposition of
current states, describing the flow of charges in a ring. In this context, the name flux
qubits alludes to the magnetic flux that is trapped in the ring, once we activate the
superconducting currents.
We will study two flux qubits with specific applications in quantum simulation
and quantum optics. We begin in Section 6.4.2 with the rf-SQUID, a supercon-
ducting loop with a single Josephson junction. This qubit is used in Chapter 9
when discussing the D-Wave “quantum annealer,” a simulator for quantum Ising
models. We also study the three-junction flux qubit or persistent current qubit in
Section 6.4.3. This is a smaller qubit, with good coherence properties, which has
facilitated new regimes of light–matter interaction, discussed in Chapter 7.

6.4.1 Frustration and Current States

Current-based qubits were a very popular device well before the transmon occupied
the quantum computing space (Makhlin et al., 2001). Phase qubits, rf-SQUIDs,
three- and four-junction loops, or fluxoniums were used to explore quantum super-
positions (van der Wal et al., 2000), qubit-cavity interactions (Chiorescu et al., 2004)
and quantum state engineering (Hofheinz et al., 2009). However, once the transmon
showed its improved coherence and reproducibiliy, it became widely adopted and
other qubits lost ground.
Not everything is rosy in the transmon world. The anharmonicity |ω12 − ω01 | sets
limits on how fast we can operate the transmon and also how strong it can couple
to other devices, such as cavities or other qubits. If we ignore this condition, we
compromise the integrity of the qubit space, causing leakage9 to the excited states
|2,|3, etc.
Flux qubits solve all limitations on coupling strength and speed with a design that,
thanks to the use of frustrated supercurrents, produces very anharmonic spectrum.

9 See Problem 6.10 for some crude estimates.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.4 Flux Qubit 131

Figure 6.5 Two flux qubit designs. (a) Equivalent circuit for an rf-SQUID qubit,
such as the one in Bennett et al. (2007). (b) Three-junction flux qubit circuit (left)
and photograph of the same qubit inmersed in a dc-SQUID for readout. Picture
courtesy of Clarke and Wilhelm (2008), adapted with permissions.

Figure 6.5 shows different incantations of a flux qubit. The device is essentially a
superconducting loop with one or more Josephson junctions. The loop is threaded
by a magnetic field, which modifies the inductive energy, causing the appearance of
two degenerate ground states with distinguishable current distributions. These will
be our qubit states.
The doubly degenerate ground state is explained by the fluxoid quantization.
The magnetic flux that passes through the qubit causes a discontinuity in the flux
variables along the superconducting circuit δφloop , which satisfies

∇φ · dl = δφloop = 0 × Z + ext . (6.50)
C

A magnetic flux with a half-integer number of flux quanta ext = 12 0 , induces

frustration in the circuit: There are two equivalent values of the discontinuity
δφloop = ± 12 0 that fulfill the equation. By virtue of the flux–current relation (3.23),
the two values of δφloop translate into two distinguishable current states with similar
intensity but opposite directions. They are almost our qubit, but not quite yet.
The argument about fluxoid quantization is too general and not tied to any
circuit design. A practical qubit has further requirements: two stable, degenerate
ground states that can be connected by quantum mechanical means. To satisfy
the first requirement, we must stabilize energetically the two values of the flux.
For that, we insert Josephson junctions in the superconducting loop, ensuring that
their sinusoidal potentials produce two identical minima close to the two frustrated
states ± 12 0 .
The second requirement seems to contradict the previous one: we need to “cou-
ple” or enable transitions between the two current states. Part of this requirement
is fulfilled by the capacitance of the circuit, which tends to delocalize the ground
state wavefunction in flux space, coupling |+ 12 0  and |− 12 0 . However, to allow

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 132 Superconducting Qubits

the capacitance to act, we also need to balance the Josephson energies, lowering the
energy barriers just enough to activate quantum tunneling.

6.4.2 rf-SQUID Qubit

Following our discussion, the recipe for building a flux qubit includes (i) a supercon-
ducting loop, (ii) one or more Josephson junctions, and (iii) an external magnetic
field that induces frustration. The rf-SQUID is the oldest and the simplest qubit
design that matches all three demands. As shown in Figure 6.5a, the circuit consists
of one big loop providing some linear inductive potential, in series with a Josephson
junction or a dc-SQUID.
The Hamiltonian of the rf-SQUID was introduced in Section 4.7.1:

1 2 1 2 1
H  q + φ − EJ cos (φ + ) . (6.51)
2CJ 2L ϕ0

As with the transmon, we interpret H as the Hamiltonian for a quantum particle of

finite mass m ∼ h̄2 /8EC moving in the inductive potential created by L and EJ .
While the inductive terms confine the particle, the capacitive term contributes to its
delocalization.
Since the qubit operates in the limit of weak capacitive energies, EC /EJ  1,
we develop a good intuition of the qubit’s dynamics by studying first the induc-
tive potential. The energy landscape for the rf-SQUID is sketched in Figure 6.6.

Figure 6.6 Energy levels of the rf-SQUID (a) exactly at the maximal frustration
point  = 0 /2 and (b) 10% away from this value. Note how the shifted cosine
potential (long dashed lines) conspires with the inductor potential (parabola with
short dashed lines) to create two local minima, L and R in the frustrated case
Note also how a small change in the flux implements a relative energy shift of
the two current states |L and |R. We have exaggerated the plots using a ratio of
inductances L/LJ = 5 of the loop versus the junction.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.4 Flux Qubit 133

It consists of a broad parabolic potential (short dashed lines) on top of which we

superimpose a weaker, periodic potential (long dashed lines) with a finite barrier
height EJ cos((φ + )/ϕ0 ).
When there is no magnetic flux threading the rf-SQUID, the inductor’s and the
Josephson junction’s potential align with each other and define a common ground
state around zero flux, φ = 0. On average, this state has no charge and no current,
and behaves like a transmon.
We increase the external flux until we reach   12 0 . At this point, we achieve
frustration, because the rf-SQUID cannot minimize the energy of the linear and non-
linear terms simultaneously.10 The two degenerate ground states are the solutions
of a transcendental equation:
∂ 1 EJ L L
(ϕ0 ϕ)2 + EJ cos(ϕ) = 0 ⇔ ϕ = 2 sin(ϕ) = sin(ϕ), (6.52)
∂ϕ 2L ϕ0 LJ

where LJ = ϕ02 /EJ is the effective inductance associated to the Josephson junction
(see Section 3.7). Apart from the trivial maximum ϕ = 0, this equation reveals
two absolute minima whenever β = LJ /L < 1. As an example, Figure 6.6a shows
two absolute minima labeled L and R, both close to the ideal frustrated solution
φL,R  ± 12 0 .
Let us complete our model, adding some “quantumness” to the classical picture
of well-defined flux states. We first consider each of the local minima separately. If
they are deep enough, each of them may be approximated by a harmonic potential.
When combined with the capacitive term, the inductive potential gives rise to two
effective LC resonator ground states,
1
HL,R  4EC n̂2 + EJ (ϕ − ϕL,R )2, (6.53)
2
each placed around a different minimum φL and φR . The approximate wavefunc-
tions of these two solutions are Gaussians:
) *
1
ψL,R (ϕ) = ϕ̂|L,R ∝ exp − 2 ϕ − ϕL,R . (6.54)
2σϕ
The wavefunctions of the |L and |R states are centered on the left and right poten-
tial minima, and correspond to the left- and right-moving currents from Figure 6.5a.
√
The width σϕ ∼ 8EC /EJ is both the uncertainty of the phase variable and the
spreading of the wavefunction. Note how it depends on the curvature of the inductive
potential, which is dominated by the contribution from the Josephson energy.

10 As explained in Chapter 9, we talk about frustration when we have a functional that is a sum of terms, but we
cannot find the minimum of that functional by looking at each term separately.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 134 Superconducting Qubits

At low energies, we assume that the qubit remains close to the ground states |L
and |R. We can project the complete Hamiltonian, without approximations, onto
the subspace spanned by these two states. Note that |L and |R are not orthogonal.
The wavefunctions ψL,R (ϕ) spread outside their respective potential minima, over-
lapping on the potential barrier. Consequently, the projected Hamiltonian includes
terms that enable the tunneling of states in one well |R to the other one |L, and vice
versa. We quantify this probability amplitude with the tunneling matrix element:

h̄t = L|H |R = ψL (ϕ)∗ H ψR (ϕ)dϕ ∝ EJ e−(ϕR −ϕL ) /σϕ = 0.
2 2
(6.55)

The amplitude t is proportional to the matrix element of the inductive energy

between flux states and thus grows with EJ . However, the localization of ψL,R also
−1/4
makes this value exponentially sensitive to the wavepacket size σϕ ∝ EJ and to
the barrier height.
When performing this projection, we also have to consider the energy of the
minima themselves, L|H |L and R|H |R. Under perfect frustration, shown in
Figure 6.6a, both terms would be identical. However, when we change the mag-
netic flux away from the symmetric point, 12 0 + δ, the flux perturbation δ
displaces the cosine potential, shifting the energy of the current states in opposite
directions (see Figure 6.6b). This energy shift is proportional to the flux variation
h̄ε = μ × δ = 12 (L|H |L − R|H |R).
When we take both terms into consideration, we arrive at a complete model for
our flux qubit, given by

H  μ × δ(|LL| − |RR|) + h̄t (|RL| + |LR|) . (6.56)

Since states |L and |R are not strictly orthogonal, they do not form a convenient
basis. It is much better to work with new qubit states that are superpositions of the
original ones:
1 1
|0 = √ (|L − |R), and |1 = √ (|L + |R). (6.57)
2 2
These states produce a qubit Hamiltonian in the usual form:

H = h̄tσ z + μδσ x . (6.58)

The model contains a fixed, minimum energy gap  = 2t that depends on the
tunneling, and an adjustable magnetic field bias h̄ε = μδ that is used to implement
single-qubit unitary operations (see Section 6.1.4).
The rf-SQUID has a long history that predates its uses for quantum information
processing (Clarke and Braginski, 2004). The circuit itself was conceived as an

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.4 Flux Qubit 135

ultrasensitive magnetometer (Silver and Zimmerman, 1967). It was only much later
identified as a qubit candidate due to its two-level subspace and relatively large
anharmonicity.
The rf-SQUID is still used in various devices, including the D-Wave quantum
annealer or quantum optimizer. Its large size can be an advantage. By stretching the
superconducting loops (Harris et al., 2007), a single qubit may physically overlap
or come into contact with many other qubits. This is used to create large lattices of
thousands of qubits with large connectivity over long distances. The resulting setups
can simulate complicated spin Hamiltonians and implement sophisticated quantum
optimizations, as explained later in Chapter 9.
The Hamiltonian (6.58) reveals the two main problems of the flux qubit. First, the
qubit gap  ∝ 2t is exponentially sensitive to the Josephson energy. Small changes
in EJ , due to differences in the junction’s area or thickness, or due to trapping of
charges or impurities in the oxidized layer, can cause large fluctuations in the qubit
properties, hampering reproducibility – especially when compared to the transmon.
The second problem is decoherence. The bias field of the qubit is generated by
external magnetic fluxes δ, that are transported close to the qubit by different types
of antennas. Unfortunately, as described before when working with cavities, these
cables not only transport the static current we need to create the magnetic flux.
They also act as bath that couples to the qubit, introducing quantum fluctuations
that must be modeled as dephasing and decoherence (see Sections 6.1.5 and 6.1.6).
As in the charge qubit (Section 6.2.3), the optimal operation point of the qubit to
minimize this environment-induced dephasing lays at the symmetry point δ  0
of maximum frustration.
However, even close to the symmetry point, the decoherence of the rf-SQUID
may be inadequate for quantum information purposes. These qubits are large by
design – to have a large coupling strength μ, to overlap with faraway qubits (see
Chapter 9) and to have a large linear inductance – and this size comes together with
a greater sensitivity to stray fields and decoherence. An obvious solution is to work
with smaller qubits, where the frustrated potential is created using only Josephson
junctions, as explained in the following section.

6.4.3 Persistent Current Qubit

The flux qubit based on three (or four) junctions was introduced by Mooij et al.
(1999) and experimentally demonstrated by van der Wal et al. (2000). The persistent
current qubit replaces the big inductive loop from the rf-SQUID with a smaller loop
where most of the inductance is provided by a ring of Josephson junctions, as shown
in Figure 6.5b. When the loop is threaded by half a flux quantum, the qubit develops
the same degenerate and frustrated current states as the rf-SQUID. However, the

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 136 Superconducting Qubits

two current states are energetically stabilized in the multiwell inductive landscape
generated just by the Josephson junctions, without any linear inductors.
The resulting qubit is smaller than an rf-SQUID – i.e., a few micrometers versus
150 μm structure in Figure 6.5a. This makes it possible to pack more qubits in the
same setup, and to embed them in exotic structures, such as a transmission line
nanoconstriction, to achieve extreme regimes of interactions with the microwave
field.11 There have also been significant improvements in flux qubit fabrication
and experiments by Yan et al. (2015) have demonstrated coherence times that are
comparable to those of the best 2D transmons T2  40 μs. All this makes the three-
junction qubit the second most used qubit in the field, and usually the one that people
refer to when discussing “flux qubits.”
The circuit of the persistent current flux qubit was derived in Section 4.8 and it
is sketched again in Figure 6.5b. It consists of three junctions. Two junctions are
identical and have the same Josephson energy and capacitance EJ and CJ . The
third junction’s area is smaller by a factor α < 1, leading to a reduced Josephson
energy α × EJ and a smaller capacitance α × CJ . The effective Hamiltonian has
the form
1 1 2
H = q+2 + q + EJ V (φ−,φ+ ) (6.59)
2(1/2 + α)CJ 2CJ −
with a nonlinear potential
    
 − φ+ φ+ φ−
V (φ−,φ+ ) = αEJ cos − 2EJ cos cos . (6.60)
ϕ0 2ϕ0 2ϕ0
This Hamiltonian depends on the total flux jump across the two bigger junctions
φ+ = φ1 + φ2 . Note how we have eliminated the flux jump across the smaller junc-
tion, using the fluxoid quantization, φα =  − φ+ .
Once more, the flux qubit Hamiltonian joins two competing potentials. A global
and fixed inductive potential ∝ EJ cos(φ+ /2ϕ0 ) favors a global minimum at
φ+ = φ− = 0. On top of this, we superpose a cosine potential that can be displaced
along the φ+ axis using the external flux . When  = 12 0 , this potential favors
two minima with opposite signs of the flux.
Figure 6.7 shows two landscapes of the inductive energy: (a) without frustration
and (b) right at the maximum frustration point  = 12 0 . The local minima of
energy lay always along the line φ− = 0, where the bigger junctions share the same
current φ1 = φ2 . However, in the frustrated case we find two minima along the φ+
axis. These minima correspond to two directions of a similar current, i.e., positive

11 See, for instance, the experiments by Forn-Díaz et al. (2010) and Niemczyk et al. (2010) based on the ideas by
Bourassa et al. (2009), or more recent experiments by Forn-Díaz et al. (2016) demonstrating ultrastrong
coupling between flux qubits and propagating photons.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.4 Flux Qubit 137

Figure 6.7 Inductive energy and eigenstates of the three-junction flux qubit.
Nonlinear inductive potential (6.60) with (a) α = 0.9 and no external flux and
with (b) α = 0.9 and /ϕ0 = π, together with the respective cuts along φ1 = φ2
(c,d). (e) Energy levels of the three-junction flux qubit with anisotropy α = 0.7. (f)
Energy gap between the “qubit” states E1 − E0 and between the first and second
excited states, E2 − E1 , as a function of the anisotropy, for  = 0 /2.

flux jumps and right-moving currents for φα,1,2.+ > 0, and negative jumps or left-
moving currents for φα,1,2,+ < 0.
The persistent current qubit is rather complicated and does not bend well to
analytical estimates. As explained in Problem 6.13, a numerical method that
works well is to diagonalize the qubit Hamiltonian in a basis of number states,

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 138 Superconducting Qubits

defined on the space of the ϕ± = φ± /ϕ0 variables. Following the techniques from
Section 4.9, and with a moderate number of states, we can get very good estimates
of the eigenenergies and eigenstates of the three-junction qubit for all values of the
frustration parameter .
Figure 6.7e shows the spectrum for a typical junction anisotropy α = 0.7, and
the usual ratio between inductive and capacitive energies EJ /EC = 1/50. Right on
the symmetry point, the spectrum develops two states that are very close and form a
very good candidate for a qubit space, |0 , |1. A proper inspection of those states
reveals that they are superpositions of two current states with opposite directions,
associated to opposite values of φ+ , as in (6.57). These persistent currents are the
ones giving the name to the qubit, whose original motivation was to show a quantum
superposition of two macroscopically distinguishable states (Mooij et al., 1999), in
the spirit of Schrödinger cat–type experiments.
Figure 6.7e shows the splitting between these states h̄ω01 ∼ E1 −E0 , and the gap
separating them from higher excitations h̄ω12 ∼ E2 − E1 . These values may seem
small, but Josephson energies for a multiple-junction qubit are pretty large, around
100 GHz, and a gap in the range 0.05–0.1EJ – or 5–10 GHz – is acceptable for most
experiments. One conclusion from this plot is that the relative anharmonicity of our
qubit is very large. For realistic parameters of the qubit, it is easy to get factors
of ω12 /ω01 ∼ 1000% or larger, orders of magnitude better than a transmon. The
same plots show the exponential decay of qubit energy splitting – i.e., the tunneling
amplitude h̄t – with the barrier between potential wells ∝ αEJ .
Figure 6.7e reveals that we have not only good qubit states, but also a full qubit
Hamiltonian as in (6.2). Note how small changes in the total flux introduce shifts in
the energy levels, recreating the qubit hyperbola curves from Section 6.2.3. A more
careful analysis reveals that the flux qubit behaves similarly to the rf-SQUID and
follows the qubit model:
 
1
 = (E1 − E0 )= 1 0 , and ε  μ  − 0 . (6.61)
2 2
The effective magnetic dipole μ can be estimated from the numerical simulations,
fitting the energy levels to hyperbola, or semiclassically, studying the energy shifts
of the minima as a function of an external magnetic field (see Problem 6.12).

6.4.4 General Operation

Flux qubits can be regarded as large magnetic dipoles with two preferred ori-
entations, associated to the two current states. They couple best to the magnetic
fields – generated by nearby antennas, currents in microwave resonators, or other
flux qubits – and as a result, this is the usual way to control them. The image of

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.4 Flux Qubit 139

the flux qubit as a pseudospin is supported by their highly anharmonic spectrum.

As we have seen, the large nonlinearity of the Josephson potential allows us to
engineer qubit states that have a large splitting, ω01  4 − 8 GHz, protected by an
even larger separation from higher excited states, ω12 > 20 GHz. The result is a
qubit that allows fast operations – i.e., the bandwidth of the pulses that interact
with them can be comparable to the gap ω01 – and that interact strongly with other
superconducting elements – the coupling strength g, as we will soon see, can be
of the order of magnitude of h̄ω01 without yet causing significant leakage to other
excited states.
We can implement a projective measurement of the flux qubits in the current
basis, detecting the total flux trapped in the loop – i.e., the external flux plus the
contribution of the qubit currents. The measurement is performed by a dc-SQUID
that is inductively coupled to the qubit. The SQUID, as we have already seen
in Section 4.7.2, is a sensitive magnetometer. When placed close to a flux qubit
(Figure 6.5a), or completely surrounding it (Figure 6.5b), the SQUID captures some
of the qubit’s magnetic flux inside its own loop. When we apply a current to the
SQUID, it will switch to a voltage state at a critical value of the current. This value
depends very sensitively on the trapped flux and can be used to discriminate the
qubit’s state. Unfortunately, the coupling between the qubit and the SQUID is also
a source of decoherence because it involves a permanent interaction between the
qubit and the readout device. Nowadays, qubits are more often measured through
the cavities they connect to, as in Section 7.4.3, or using some circuit that couples
or decouples the readout device, as in the D-Wave architecture (see Section 9.4.1).
There are two intrinsic challenges in the design and operation of flux qubits:
reproducibility and decoherence. Regarding the first problem, we have seen how
the qubit gap  depends exponentially on the junction energy EJ . Small fabri-
cation errors can take us from a situation where the energy gap between qubit
states ω01 ∼  is on the order of a few gigahertzs, to a situation where it is so
small that the quantum fluctuations are destroyed because of thermal excitations in
the chip. This problem is typically solved by tuning the qubit gap: i.e., replacing
the α−junction with a dc-SQUID that allows us to set the effective value of the
anisotropy (Paauw et al., 2009). However, it is quite difficult to inject flux in a
dc-SQUID without adding some magnetic flux to the flux qubit’s loop itself – unless
one creates intricate designs, known as gradiometric qubits (see Problem 6.14) –
which is another reason why these qubits are not often used.
Regarding coherence times, flux qubits are largely limited by fluctuations in the
magnetic flux. These can be induced by the current sources and stray field that are
trapped by the qubit – even if it is as small as a persistent current qubit – and cause
the dephasing of the qubit. Coherence times of typical persistent current qubits
lay around 1−6 μs, with exceptional values around the tens of microseconds.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 140 Superconducting Qubits

Interestingly, we can improve coherence by an order of magnitude, reducing

dephasing rates to almost zero. The method is to shunt the flux qubit with a
large capacitor. As in the transmon, a large capacitance decreases the sensitivity
to external fields, increases coherence times, and enhances reproducibility, all at
the expense of anharmonicity. The work by Yan et al. (2016) reports capacitively
shunted flux qubits where decoherence is almost all attributed to relaxation T2 
T1 /2  40 μs, with no significant dephasing. The qubit has a smaller but still
significant anharmonicity in the range of 500–900 MHz.
Overall, research on flux qubit design and qubit decoherence is a huge and ongo-
ing effort, motivated by applications in quantum simulation and quantum annealing,
where we need to preserve quantum superpositions of many interacting qubits for
adiabatically long times. However, let us delay those considerations to Chapter 9,
ending here our discussion of qubit designs and moving forward to address other
important questions, such as understanding how qubits and interact and talk to
each other.

6.5 Qubit–Qubit Interactions

Superconducting qubits are usually organized in quantum registers,12 arrays where
qubits can be individually controlled and measured, and where they can interact with
each other. The natural interactions between qubits are two-body couplings.13 Two-
body interactions can be expanded using the Pauli matrices of the involved qubits.
This leads to models often found in textbooks from condensed matter physics and
quantum magnetism:
1    y y y
H = i σiz + i σ x + Jijx σix σjx + Jijz σiz σjz + Jij σi σj , (6.62)
i
2 ij ij ij

which combine longitudinal and transverse “magnetic fields” – i and i – with

x,y,z
two-qubit interactions Jij .
Depending on the active parameters, (6.62) receives different names. When
Hamiltonian (6.62) only contains one type of Pauli operators – e.g., J x ,J y ,ε = 0 –
we have the classical Ising model, a model whose eigenstates are all classical
states with well-defined values of σiz . If the interaction and the magnetic field
appear on different Pauli operators – e.g., J x , = 0 and others zero – we have the
quantum Ising model or Ising model with a transverse field. Since each direction

12 See Figure 8.2 in Chapter 8.

13 As it also occurs in Nature, multibody interactions may appear in some perturbative limits from two-body
interactions, but they will not be discussed in this book.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.5 Qubit–Qubit Interactions 141

of interaction, X, Y, or Z, is associated with a circuit degree of freedom, it is quite

unusual to find two or more types of interactions, but this is possible. For instance,
if J x = J y = 0 or J x ,J z, = 0, we arrive at an XY model, describing the exchange
of quantum excitations between qubits. We also have the well-known Heisenberg
Hamiltonian J x = J z = J y = 0, and the general XYZ model J x ,J y ,J z = 0.
The design of qubit–qubit interactions has three main applications. First, the
implementation of quantum magnetism Hamiltonians (6.62) converts the supercon-
ducting circuit into a very accurate quantum simulator. We can prepare states and
study how they evolve under the effective Ising and XY models to deepen our under-
standing of quantum many-body physics, quantum phase transitions, many-body
localization, termalization, quantum topological order, and many other interesting
phenomena. Second, as we will see later in Chapter 9, preparing the ground state of
these models is a kind of computation that can be used to solve very hard problems –
both classical and quantum. Finally, and most important, activating these qubit–
qubit interactions for brief periods of time, we implement universal two-qubit and
multiqubit unitary operations, which are the building blocks for scalable universal
quantum computation (see Chapter 8).

6.5.1 Dipolar Magnetic Interaction

Superconducting qubits can experience a mutual interaction, mediated solely by
the electromagnetic field. Consider the two flux qubits in Figure 6.8b. Each qubit’s
state is a superposition of currents with left- and right-moving orientations, |L

Figure 6.8 Qubit–qubit dipolar interactions. (a) Flux qubits and (b) transmon
qubits have an associated magnetic or electric dipole moment. These moments
interact with each other through the equations of electromagnetism, ∝ d1 d2 and
y y
∝ μ1 · μ2 , giving rise to effective qubit couplings J σ1x σ2x or J σ1 σ2 . Sometimes,
as in the magnetic case (c), an auxiliary circuit can be used to mediate and tune the
interaction.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 142 Superconducting Qubits

and |R. The superconducting current is a σ x Pauli operator multiplied by an aver-

age current intensity14 I˜c :
Iˆ = I˜c |LL| − I˜c |RR| = I˜c σ x . (6.63)
The current in each qubit creates a magnetic field distribution that is very similar to
the one from a magnetic dipole μ ∝ Iˆe⊥ sitting perpendicular to the qubit’s plane.
From the classical interaction of magnetic dipoles,15 we expect an antiferromagnetic
interaction Hint ∼ +M Iˆ1 Iˆ2 that favors antiparallel orientations of the dipoles – or,
equivalently, currents flowing in opposite directions on both qubits.
We can also derive this interaction in the lumped-element circuit model, using
geometric properties that can be designed experimentally. In particular, we must
introduce, as in the rf-SQUID Hamiltonian (see Section 4.7.1), the mutual induc-
tance M between the superconducting qubits’ loops. The mutual inductance, which
can be estimated from the shape of the circuits, accounts for the change in the mag-
netic flux of one circuit created by another qubit’s currents, and the magnetic energy
that results from it. This change is symmetric and is written as a quadratic correction
to the energy Eint = M12 I1 I2 , in terms of the actual currents I1,2 flowing in both
loops. These currents may have a complicated expression in terms of the circuit’s
fluxes.16 When the perturbation is small compared to the inductive potential, we
can approximate its contribution to the Hamiltonian with a product M12 Iˆ1 Iˆ2 of the
current operators that we estimated in the qubit basis (6.63). When it is large, on the
other hand, M can itself influence the qubit states, leading to renormalizations that
must be studied in a self-consistent way.
For a generic array of flux qubits, the inductive coupling adopts the Ising form
Jij μ̂i μ̂j ∼ Jij σix σjx , with matrix Jij = M I˜ci I˜cj . For a nearest-neighbor and planar
configuration such as Figure 6.8b, Jij > 0 and the interaction is antiferromagnetic –
it favors configurations with opposing signs for σix and σjx . However, other values
are possible through a clever engineering of the qubit’s geometry.

6.5.2 Dipolar Electric Interaction

We can study the superconducting island and the transmon in an similar way to the
flux qubits. When we charge those qubits, we create a charge imbalance concen-
trated over a small region of space. This charge distribution creates an electric field

14 Note that the average value I˜ is not exactly the supercurrent of the junctions themselves. For instance, in the
c
rf-SQUID we find I˜c = |φL,R |/L, in terms of the equilibrium value of the flux for either potential well (see
Problem 6.15).
15 For two dipoles μ , their interaction energy goes as E = −(μ /4π |r|3 )(3(μ · r)(μ · r) − μ · μ ) with
1,2 0 1 2 1 2
their separation r. Since both our dipoles sit on the same plane, we expect E = +μ0 μ1 μ2 /4π |r|3 .
16 For instance, in the three-junction flux qubit I
1,2 ∝ αIc sin(φ+ /ϕ0 ).

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.5 Qubit–Qubit Interactions 143

that is very similar to the one from an electric dipole di ∝ q̂i . From the classi-
cal theory of electrostatic interactions, we expect a contribution to the energy that
grows with the product of those dipoles and decreases with their separation. Unlike
the magnetic interaction, there is great variability on the sign and strength on the
interactions,17 based on the relative orientations of the qubits and their separation,
but we may write something like
( x x
1 J σ1 σ2 , for charge qubit,
Hint = ± q̂1 q̂2 ∼ y y
(6.64)
Ceff J σ1 σ2 , for transmon.
The sign and the effective mutual capacitance Ceff can be determined by the micro-
scopic model. The coupling operator may also change depending on our convention
for the charge basis: In the superconducting island, q̂ ∝ σ x , while for the transmon
with the notation from (6.33), q̂ ∝ i(|10| − |01|) ∼ σ y . Sometimes these dif-
ferences are irrelevant, and may be suppressed with a local unitary transformation,
but other times they are not, leading to unequivalent models – e.g., imagine two
transmons with SQUIDs interacting also inductively through their loops!
We may be a bit more rigorous in our treatment of the capacitive interactions by
considering the effective circuit in Figure 6.8c. The complete model includes both
the self-capacitance of the transmon or charge qubit CJ , and a mutual capacitance
between both qubits Cg . The whole circuit may be described using the flux degrees
of freedom in the two superconducting islands, with a capacitance matrix:
 
CJ + Cg −Cg
C= , (6.65)
−Cg CJ + Cg
where the term 12 Cg (φ̇1 − φ̇2 )2 causes Cg to appear both in the diagonal, renormaliz-
ing the effective capacitance of each qubit, and in the off-diagonal, with something
that resembles an interaction. This structure is preserved in the effective Hamilto-
nian, which depends on the inverse capacitance matrix:
⎛ 1+Cg /CJ ⎞
Cg
2
 2 
CJ CJ Cg
C −1 = ⎝ ⎠+O . (6.66)
Cg 1+Cg /CJ CJ
CJ2 CJ

We have a dipolar interaction, but the mutual capacitance Ceff = CJ × (CJ /Cg )
differs from Cg and we can only ignore the renormalization of the qubits when in
the limit of weak interactions Cg /CJ  1.

17 For two electric dipoles, the electrostatic energy is (d · d − 3(d · r)(d · r)/r 2 /(4π  r 3 ) as a function of
1 2 1 2 0
their separation r.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 144 Superconducting Qubits

As in the case of electric dipoles, interactions between qubits decrease with

the distance, but are never strictly nearest-neighbor. Take for instance a one-
dimensional ring of qubits, coupled with periodic boundary conditions, so that
 
their capacitive energy reads i 21 CJ φ̇i2 + i 21 Cg (φ̇i − φ̇i+1 )2 . When we invert the
capacitance matrix for this circuit in the limit Cg /CJ  1, we obtain an interaction
that decays as Cij  exp(− log(Cg /CJ )|j − i − 1|)/CJ , but if we consider two-
dimensional arrangements, we may obtain Cij that decay polynomially or even
logarithmically up to a critical length (Ortuño et al., 2015). These types of inter-
action are often neglected, but they contribute to the residual cross-talk – all the
unavoidable interactions between qubits are not considered in the model, but which
appear in the experiment, due to the intervening elements, substrate, cables, qubit
separation, etc.

6.5.3 Coupling Tunability

There are scenarios where we only want some qubits to interact with an other. For
instance, as explained in Section 8.3, quantum algorithms are decomposed into
elementary operations, some of which involve the evolution of one pair of qubits
with a given interaction for a brief period of time – e.g., J σ1z σ2z to implement a
control-phase rotation. During this time, those qubits must not talk to other qubits, or
otherwise the gate will be spoiled. This motivates our need to activate or deactivate
interactions at will.
Looking at qubits that interact with each other, there are two approaches to switch
interactions on and off. A hardware-based approach is to change the qubit by modi-
fying its dipole moment. For instance, we could replace principal junction in a flux
qubit – e.g., the α loop in the three-junction design – with a tunable dc-SQUID.
By threading the SQUID with a magnetic flux, we effectively change the critical
current, lowering the value of the interaction. Unfortunately, this also causes us to
lose the qubit itself!
A slightly more robust trick is to change the experimental conditions to make
interaction less effective. In the dipolar coupling limit, we have two-body inter-
actions along directions orthogonal to the natural Hamiltonian of the qubit. For
instance, we find
1 z  2 z
σ +
H = σ + J σ1x σ2x , (6.67)
2 1 2 2
where 1,2 are the qubits’ gaps. The coupling term contains four types of contribu-
tions σ1x σ2x = |0011| + |0110| + |1001| + |1100| , which flip or exchange
excitations between both qubits. If |J |  |1,2 |, we may invoke the RWA to neglect
the interactions with |00 and |11, because those states have very different energies.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.5 Qubit–Qubit Interactions 145

If we further impose a large detuning δ = |1 − 2 |  |J |, we might even ignore

all interaction terms, because all the connected states would be off-resonant from
each other.
This technique is rather general and often used in the laboratory. It has two
shortcomings. The first problem is that we neglect the influence on the dynamics
due to the off-resonant interactions, and which to lowest order in perturbation theory
are (see Problems 6.16 and 6.17)
J2 J2
Hint = (|1111| − |0000|) + (|0101| − |1010|) + · · · .
2(1 + 2 ) 2 −  1
(6.68)
These contributions may accumulate over long times, producing large changes in
the phases of the qubit, which disturb our computations.
The second problem is that this strategy does not scale up nicely to large
setups. We can tune the qubit frequencies only within a finite bandwidth, such
as δ ∼ 1−2 GHz. Thus, if we try to park N qubits separated from each other, at
least two of them will be at a distance O(1/N) in the spectrum. This sets a lower
bound in the corrections that we neglected, which now grow as O(J 2 N/δ). This
situation, called frequency crowding, appears already for setups with a moderate
number of qubits, limiting the fidelity of quantum operations (see Problem 6.16).
As we will see now, all these obstacles may be strongly suppressed or eliminated
when we use other circuits to mediate the interactions between qubits.

6.5.4 Mediated Interactions and Tunable Couplers

In this section, we introduce the notion of tunable coupler, a superconducting circuit
that acts as a mediator of interactions in a way that can be adjusted by external
controls. Mediated interactions are an important element in superconducting circuit
technology. Couplers allow us to connect circuits that are far apart – or which at
least are not immediate neighbors. We use them to control the strength and sign of
interactions, and sometimes they are even used to cancel other residual couplings
or cross-talk.
We will now discuss a rather general framework of interactions mediated by
“fast” circuits, circuits that never leave their ground state, but which can transport
the influence of one circuit to another. We will discuss the particular example of a
dc-SQUID, illustrated in Figure 6.8c. In this configuration, the magnetic field from
one qubit induces a supercurrent into the mediating loop; this current generates its
own magnetic field that influences the far away qubit and vice versa.
When the mediating loop can quickly adapt to the influence of both qubits, it
remains in a quasi-instantaneous equilibrium state, and we describe the coupling

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 146 Superconducting Qubits

scheme with the general framework of adiabatic interactions. To be precise, the

mediating circuit C must satisfy a series of conditions that for our two-qubit setup
become the following ones:
(1) The coupler interacts pairwise with the qubits 1 and 2. The total energy of the
three systems E(x1,x2,XC ;ext ) depends on the states of both qubits x1,2 , on
the state of the coupler XC and on some control parameter .
(2) The coupler evolves faster than the qubits it talks to. We can apply the Born–
Oppenheimer approximation,18 also known as adiabatic elimination, and
assume that the coupler is always at the lowest energy state
XC = Xeq (x1,x2 ;ext ).
(3) We can eliminate the coupler variable, which is fixed, obtaining an effective
interaction Eeff (x1,x2 ;ext ) = E[x1,x2,Xeq (x1,x2 ;ext )].
(4) Using the control parameter ext , we can change the form and strength of the
interaction energy Eeff . We can use the parameter cancel any previous direct
interaction or cross-talk between qubits 1 and 2, or to establish a new effective
coupling between them.
As shown by van den Brink et al. (2005), this design allows the dc-SQUID from
Figure 6.8c to work as tunable inductive coupler. The Hamiltonian for the mediating
SQUID, introduced in Section 4.7.2, depends on the total flux  that passes through
the loop:
  
1 2  φ̂−
H = q̂ + 4EJ cos × cos . (6.69)
4CJ − 2ϕ0 ϕ0
√
We assume that the plasma frequency of the SQUID h̄ωdc ∼ 8EC EJ () is much
faster than the timescales at which  will change, ωdc  /. ˙ The state of the cou-
pler XC ∼ {q−,φ− } adapts to the adiabatic changes in the flux , leaving the SQUID
always in the instantaneous ground state defined by the external parameters. The
energy of the SQUID can be approximated by ESQUID ()  h̄ωdc /2, as expected
from an oscillator in the vacuum state.
We now have to include the effect of the nearby qubits. The flux enclosed by
the loop  is a sum of a classical control field ext and the flux injected by nearby
qubits through their mutual inductance Mi . The combined Hamiltonian for the two
qubits with the coupler reads
Ĥ = Ĥ1 + Ĥ2 + ESQUID  + M1 σ1x + M2 σ2x . (6.70)

18 This approximation or way of thinking is quite common in physics. It describes how electrons mediate an
interaction between atoms in a molecule, or how phonons bind electrons to form Cooper pairs.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 6.6 Qubit Coherence 147

In this Hamiltonian, because of the adiabatic approximation, we replaced the Hamil-

tonian for the SQUID with its instantaneous ground state energy – a classical func-
x
tion of quantum operators σ1,2 . We can rewrite the energy functional as a product
of the dipole moments of both qubits:19
1
Ĥ = Ĥ1 + Ĥ2 + J12 ()σ1x σ2x , (6.71)
2
where Ĥ1,2 are the Hamiltonians for the coupled qubits, and the mutual inductance
J12 is the second-order contribution:20
∂2
J12 ()  M1 M2
ESQUID (). (6.72)
∂2
Many experiments have demonstrated tunable coupling using dc-SQUIDs (Hime
et al., 2006), more complicated circuit loops (van der Ploeg et al., 2007) or even a
qubit (Niskanen et al., 2007) with faster dynamics. A particularly relevant device is
the rf-SQUID coupler developed by Harris et al. (2007), and which together with a
somewhat outdated flux qubit design constitutes the basis of the D-Wave quantum
annealer discussed in Chapter 9.

6.6 Qubit Coherence

Qubit coherence is one of the limiting factors in the design and implementation
of quantum computers and quantum devices. Decoherence destroys the quantum
information that is stored in the qubits. It not only acts during the time in which
the qubits are “resting,” but also affects the qubit during the realization of single-
qubit operations and measurements. In presence of decoherence, our controls are
no longer the idealized models from Section 6.1.4, but include corrections – such
as dephasing, relaxation, or cross-talk – that may limit the quality with which those
operations can be performed.
It is very important, and a first step in any quantum technology application, to
characterize the coherence properties of the system at hand. This is true not only for
fundamental works with new qubits designs – e.g., Nakamura et al. (1997), van der
Wal et al. (2000), or Steffen et al. (2010). Higher-order devices, such as multiqubit
quantum computers (Arute et al., 2019), and quantum computing infrastructures,
such as the IBM Quantum Experience, must continuously calibrate their devices,

19 Higher powers of Pauli matrices are either the identity, or the matrices themselves. Therefore, the function
ESQUID can only produce terms that are proportional to σ1,2 x or to the product σ x σ x , as seen by expanding
1 2
the function to arbitrary powers in a Taylor series.
20 Ideally, ∂E
SQUID /∂  0. If not, this represents a shift of the qubit gaps, which can be included in the theory.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 148 Superconducting Qubits

monitoring the coherence times of the qubits, as well their natural frequencies,
quality of measurements, fidelity of operations, etc.
An ab initio determination of how one or more qubits evolve can be a rather
cumbersome problem. Chapter 8 will introduce methods for a self-consistent tomo-
graphic reconstruction of how a density matrix evolves under a general control.
Fortunately, in absence of external controls, qubits tend to follow a simple decoher-
ence model that mostly contains relaxation and dephasing (6.19). This model can
be fully determined by calibrating the values of the relaxation time T1 and of the
dephasing time, T2 or Tφ , in different experiments.
To calibrate the relaxation time, we study the evolution of the excited state
|1. We reset the qubit to |0, flip the state of the qubit with a π pulse, wait
for a time t, and measure the excited population P1 (t) = 12 1 + σ z . As shown
experimentally (Steffen et al., 2010), even if we make small errors in the preparation
and measurement, the excited population can be fit to an exponential curve
P1 (t) = P1 (0) exp(−t/T1 ), which gives the value of T1 .
To obtain the value of the decoherence time, we must study the evolution of a
quantum superposition, using a protocol known as Ramsey interferometry. We start
with the |0 state, but now implement an incomplete rotation around σ y with the π/2
angle, preparing the state √12 (|0 − |1). We wait for a time t, repeat the rotation,
and measure σ z . The excited state population now follows a damped oscillation
∗
P1 (t) = 12 (1 + cos(t)e−t/T2 ).
The coherence time is always limited by relaxation, T2∗ = 2T1 . In practice, this
time is shorter because of dephasing and other sources of decoherence Tφ , as shown
in (6.20). In those cases, we can further separate and partially cancel the contribu-
tion from dephasing, using spin-echo to obtain a slightly larger effective coherence
lifetime T2echo (see Section 6.1.5).
Figure 6.9 illustrates the experimental coherence times T2∗ (or T2echo , depending
on the work) for a large variety of qubits. There is an overall positive trend, where
coherence times almost double every year,21 starting from the charge qubit in
Nakamura et al. (1999), until recent experiments with 2D transmon qubits (Place
et al., 2021) and photon-encoded qubits (Rosenblum et al., 2018). We see significant
jumps due to the placement of qubits in microwave resonators (Wallraff et al.,
2004), the introduction of large three-dimensional transmons (Paik et al., 2011), the
development of capacitively-shunted flux qubits (Yan et al., 2016), the encoding of
qubits in photonic degrees of freedom (Wang et al., 2016; Rosenblum et al., 2018),

21 This has been deemed Schoelkopf’s law, by analogy with Moore’s law for Silicon-based integrated chips.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 Exercises 149

γ
γ
10−3 E
E
T F t
10−4 T J
T t t t T
t t
J
T t t t t t t
T2∗ (seconds)

t G
t
10−5 F G
t t t
C t t t
10−6 t G
Q C t
f J
10−7 γ P

J J
10−8
C C J
2000 2005 2010 2015 2020
Year

Figure 6.9 Qubit coherence times T2∗ (or T2 measured after spin-echo) across the
years. Letters signify: charge qubit (C), flux qubit (J), fluxonium (f for 2D, F for
3D), phase qubit (P), transmon (t for 2D, T for 3D), photon encoded qubit (γ),
error-corrected bosonic qubit (E), and gatemon (G).

and recent work to improve the surface properties of 2D transmons (Place et al.,
2021). An important question is whether this trend will saturate, or whether we will
discover new and better ways to design, fabricate, and integrate qubits in quantum
circuits.

Exercises
6.1 Show that a qubit Hamiltonian of the form H = 2 σ z + 2ε σ x can be mapped
to other forms, such as H = 2 σ z + 2ε σ y and H = 2 σ x + 2ε σ z , by a choosing
a new basis of qubit states |0̃ = U |0 and |1̃ = U |1. Compute the unitary
transformation U in those two particular cases.
6.2 A single-mode harmonic oscillator with Gaussian Hamiltonian and linear
operations cannot implement a qubit. Assume a general model of the form
H = ωa † a + (f (t)a + f (t)∗ a † ) with a general driving f (t) ∈ C. Prove that
starting from state |0 it is never possible to create a perfect Fock state
|1. Discuss how this changes when we add a nonlinearity U a † a † aa to the
Hamiltonian, with U > 0.
6.3 Let us study the limit of quasistatic dephasing. We will use a diagonal
perturbation to the dynamics H  ∼ h̄2 ( + δω)σ z , but the random perturbation
δω has no dynamics: It is constant through each experimental realiza-
tion, but fluctuates from experiment
√ to experiment with Gaussian distribution
P (δω) = exp(−δω /2σ )/ 2πσ .
2 2 2

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 150 Superconducting Qubits

(1) Compute completely positive map εt (ρ) describing the evolution of the
qubit. Compare the resulting map with (6.13). Show that coherences now
decay as ρ01 (t) ∝ ρ01 (0) exp(−t 2 σ 2 /2).
(2) Show that the evolution of the density matrix ρ(t) can be written in
Lindblad form (6.15), but now the equation is not Markovian, and the
dephasing rate depends on time γφ = tσ 2 .
(3) Assume that we apply a spin-echo refocusing protocol (6.16), which
consists of instantaneous flip operations UBB = σ x , repeated with period
TBB . Show that the density matrix after two pulses ρ(2TBB ) is the one we
would have obtained by evolving with H = 12 σ z .
6.4 Let us solve the dynamics for a qubit with Hamiltonian H = 12 σ z , in contact
with a bath at finite temperature (6.17).
(1) Using the master equation to estimate ∂t ρ, prove that the evolution of an
observable for such a qubit is
d γ
O = −i [O,H ] + σ + [O,σ − ] + [σ +,O]σ −  . (6.73)
dt 2
(2) Next, derive the Bloch equations equations for σ z  and σ ± . Solve
those equations in the zero temperature limit n = 0.
(3) Use those expected values to reconstruct the Bloch vector (2.25). Compute
the norm of the vector, and show that it is contracting and the state remains
inside the sphere.
(4) Using the Bloch vector S(t), prove that the density matrix evolves as

1 − P1 + (e−t/T1 − 1) [P1 − ρ11 (0)] ρ01 (0)eit−t/2T1
ρ(t) = ,
ρ10 (0)e−it−t/2T1 P1 + e−t/T2 [ρ11 (0) − P1 ]
where P1 = n/(1 + 2n) = limt→+∞ ρ11 (t) is the asymptotic excitation
probability of the qubit and T1 = (1 − 2P1 )/γ gives the timescale |T1 |
at which ρ(t) converges to its asymptotic solution. Relate this solution to
the zero temperature limit of pure relaxation.
6.5 Write down a master equation that combines a qubit gap h̄, dephasing with
a rate γφ and relaxation with a rate γ . Solve the evolution for an initial density
matrix ρ0 . Study the evolution of ρ01 (t) = 0|ρ(t)|1 and prove (6.20).
6.6 Charge qubit truncation. Let us consider the effect of small potential
perturbations ng = 12 + x onto a charge qubit with EJ /EC, x  1. We
will rely on the degenerate perturbation theory from Section A.3.2, identifying
our target subspace as the qubit space, P0,α=0 = |00| + |11|. Show that,
up to second order in EJ /EC , it only the coupling to the degenerate subspace
{|−1 , |2}. Using the series expansion (A.36), you should obtain a diagonal
effective Hamiltonian:

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 Exercises 151

2 EJ2
Heff = −EJ σ x + 4EC 1
+x − |00|
2
8EC (1 + 2x)
2 EJ2
+ 4EC 1
−x − |11| .
2
8EC (1 − 2x)
Does this model differ from (6.25)? What happens when x = 0? And when x
is so small that the diagonal terms can be linearized? Can we distinguish the
effective theory from the truncated one?
6.7 Following the experiment by Nakamura et al. (1999), we prepare a charge
qubit at a point ng  0, cooled down to charge neutrality |0. We then abruptly
move the experiment to the symmetry point ng = 1/2 for a time T , after
which we return to ng = 0 and measure the charge in the island. What is
the Hamiltonian of the qubit at ng = 1/2 in the charge basis? How does
the quantum state evolve, and what is the final state of the qubit at time T ?
What is the probability that we find one excess Cooper pair at the end of
the experiment?
6.8 Let us take the charge qubit by Nakamura et al. (1999), where the energy
gap   80 μeV. We are going to analyze its sensitivity to the environment
assuming that there can be trapped charges around the qubit at a distance of
d ∼ 4 μm. As a toy model, we assume that that all of the qubit on average
experiences the potential induced by this electron V = e/4πε0 d. What is the
extra energy acquired by the |1 charge state relative to the |0? How do you
write this correction into the Hamiltonian? How much does ng change by
this external influence? What about the energy levels E? How much do
they change?
6.9 Diagonalize the oscillator part of the transmon Hamiltonian in (6.32), disre-
garding the contribution of the nonlinear terms, so as to estimate the transmon
properties.
(1) Estimate transition frequency h̄ω01 .
(2) Express the charge operator and flux operator in the oscillator basis. What
are the matrix elements of those operators between the ground and excited
state? Compare with Section 6.3.4.
(3) Using nondegenerate perturbation theory, estimate the negative anhar-
monicity due to the quartic term ∝ EJ φ 4 .
(4) Connecting (6.33) to the one for an LC resonator, what is the correspond-
ing impedance of the transmon?
6.10 Let us write down the transmon Hamiltonian for the lowest three levels with
an external, time-dependent perturbation:

H = (ω12 + ω01 ) |22| + ω01 |11| + (ε12 (t) |21| + ε01 (t) |10| + H.c.) .

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 152 Superconducting Qubits

Assume that ω12 =  + α and ω01 =  are related by a small anharmonicity

α, and that ε12  ε01 = ε0 cos(ωt). We are going to study the dynamics and a
possible leakage outside the qubit space:
(1) Perform a unitary transformation with U = exp(−iH0 t/h̄), where H0 =
ω(|22| + ω |11|), to arrive at an equation similar to (6.7). In that equa-
tion, neglect the terms that oscillate with frequencies ω01 + ω, ω12 + ω,
etc. Show that the outcome reads
Heff = (δ + α) |22| + δ |11| + ε0 (|21| + |10| + H.c.) .
(2) Assume a large anharmonicity |δ + α|  |δ|, and ignore the third state.
Solve the dynamics of an initial state that is in the ground state, |ψ(0) =
|0, as a function of time, detuning and coupling. What conditions do you
need to reach the perfect state |1, and how does the time Tflip required
depend on ε0 ?
(3) Consider leakage into the third state perturbatively, while on resonance,
δ = 0. For that, realize that this matrix will have two eigenstates that
will be close to |0 and |1 but will be mixed with a small population
of |2. Show that the excitation probability grows as |ε0 |2 /|α|2 . If the
anharmonicity is 5% of the gap , what does this tell us about the
maximum coupling strength |ε0 | of the transmon to other systems?
(4) For a transmon with an anharmonicity of 2π × 400 GHz, what would be
the leakage error for implementing a gate in Tflip = 10 ns? What is the
shortest pulse we can afford to reduce leakage below 1%?
6.11 Using the relations from Section 6.3.3 and the software package Mathematica,
we can easily compute the transmon eigenenergies and eigenfunctions.22 The
eigenenergies are given by the Mathieu characteristic function, which Math-
ematica labels as MA (r,q) = MathieuCharacteristicA[r,q]. The
eigenfunctions can be reconstructed from the Mathieu sine and cosine func-
tions, seν (z) and ceν (z), as meν+2k (z,q) = ceν+2k (z,q) + iseν+2k (z,q), which
Mathematica expresses as cer (z,q) = MathieuC[r,q,z] and ser (z,q) =
MathieuS[r,q,z].
(1) Plot the function MA (r,q) for different ratios of q = EJ /2EC = 1, 10,
20, and 50 over a range r ∈ [−6,6]. Can you locate the different bands?
(2) Using the index function k(n,ng ) and the characteristic function, plot the
energy levels of the transmon for EJ /EC = 1,10,20,50. Also reconstruct
the anharmonicity plot.

22 As a curious detail, our definition of z and of q makes q positive, and later versions of Mathematica seem to
behave more stably when q ≥ 0.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 Exercises 153

(3) From the wavefunctions, by means of Mathematica’s numerical integra-

tion, compute the expectation values of the charge operator among the
ground state and first excited state:
 2π
q01 = ψ1 (ϕ)∗ (−i2e∂ϕ )ψ0 (ϕ)dϕ. (6.74)
0

Plot its behavior as a function of the EJ /EC ratio both for ng = 0 and
ng = 1/2. Hint: Mathematica also has the derivatives ceν (z) and seν (z).
6.12 Let us look at the three-junction persistent current qubit, from
Section 6.4.3. In the seminal experiment by van der Wal et al. (2000), the
two biggest junctions had critical currents and capacitances Ip  570 nA and
C  2.6 fF; the smallest junction was so by a factor α = 0.82, and the total
qubit had an area of about 10 μm2 . Using the circuit model from Section 4.8
and the preceding theory, answer these questions:
(1) How large is the energy barrier between the two potential wells along
φ− = 0? How does it compare to the total depth of the potential?
(2) Estimate the gauge-invariant phase ϕ across the big junctions and relate
it to the current around the loop.
(3) Write down the inductive energy Eind of the three Josephson junctions
operating at the flux qubit point,  = 0.5ϕ0 + B , where B is the flux
induced by a uniform magnetic field perpendicular to the loop. Estimate
the qubit’s effective dipole as the change in energy with respect to B,
μ ∝ ∂E/∂B.
(4) How does the effective magnetic moment (in Bohr magnetons) compare
to the permanent magnetic moment of atoms such as chromium? Discuss
the reason for the difference.
6.13 We can diagonalize the flux qubit Hamiltonian (6.59) using the number-phase
representation from Section 4.9. Note that we have to use two integer numbers
|n+,n− , one for each flux degree of freedom.
(1) Adimensionalize Hamiltonian H expressing all terms as a function of
ratio EJ /EC .
(2) Write down the Hamiltonian H in the two-dimensional number basis
|n+,n− . Your result should produce diagonal expressions for q± and
introduce charge increments or decrements associated to cos(φ± /2ϕ0 )
and similar terms.
(3) Translate the previous expression into a small Mathematica or Matlab
program that works with a truncated space, n± ∈ {−N, . . . ,N}. Typically,
N  10 will suffice.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 154 Superconducting Qubits

(4) Diagonalize the Hamiltonian to compute the eigenenergies E0,1,2 around

the frustrated point,   12 0 . Try to reproduce the curves from
Figure 6.7e,f, for both the qubit gap and the separation from the second
excited state.
(5) Demonstrate that the Hamiltonian at any other flux point H ( 12 0 + δ)
can be decomposed according to the formula (6.2). For that, compute the
two eigenfunctions |0 and |1 at the symmetry point 12 0 and project the
Hamiltonian at any other value onto those states, computing the effective
matrix:
 
0|Hδ |0 0|Hδ |1
Heff = ,
1|Hδ |0 1|Hδ |1
and expand in the basis of Pauli matrices.
6.14 Take the qubit design by Paauw et al. (2009). In this gradiometric qubit,
the flux is spread over two symmetrically placed loops. Show that the qubit
operation point depends only on the flux difference up − down . The qubit is
therefore insensitive to longer-scale fluctuations of the magnetic field, which
may be favorable for reducing sensitivity to environmental noise.

6.15 Flux qubit current operator. Solve approximately the equations for the
minima of the inductive potential φL,R , in the limit of large L/LJ ratio. Show
that in this limit, the phase ϕ can be written as ±(π −ε), with a small correction
ε ∝ L/LJ . Use the derivation in Section 4.7.1 to write down the current
operator for the rf-SQUID in terms of the flux jump along the inductor φ̂.
Estimate the matrix elements of the current operator on the |L and |R states.
Show that it is approximately diagonal in these states (6.63). Using the change
of basis (6.57), prove that the current operator has approximately the form
I˜c σ x , finding the expression for I˜c = |φL,R |/L.
6.16 The rotating wave approximation also applies to the dynamics of qubits.
Let us consider the interacting qubit model (6.67), in the limit of resonant
qubits 1 = 2 .
(1) Use the decomposition σ x = σ + + σ − to find what would be the RWA
Hamiltonian HRWA associated to this model.

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 Exercises 155

(2) Show that H can be decomposed into two 2 × 2 matrices acting in the
even and odd subspaces, {|00 , |11}, and odd {|01 , |10}. Use this to
compute all eigenvectors and eigenenergies.
(3) Analyze the eigenvectors and show that in the limit |g|  ||, this matrix
has approximately the same eigenstates as the Hamiltonian HRWA that
results from neglecting the counterrotating terms. Estimate the correction
to the states introduced by not neglecting the counterrotating terms and
show that the associated excitation probability scales as |J /|2 .
6.17 Return to the interacting qubit model (6.67), but now assume that the qubits
are far apart in frequency space, δ = 1 − 2  |J |.
(1) Show that the Hamiltonian H still can be decomposed into even and odd
subspaces, as in Problem 6.16.
(2) Show that, in the limit 1 = 2 the spectrum is formed by four different
eigenenergies. Analyze the same problem using second-order nonde-
generate perturbation theory (see Section A.3.1). Find out the effective
model for this Hamiltonian in the strongly off-resonant limit, considering
corrections up to O(J 2 /δ). Based on your exact diagionalizations, at
which limit of the interaction does perturbation theory fail?
(3) Solve analytically the Schrödinger equation for two qubits that start
in the state |01 and evolve with the full Hamiltonian (6.67) and with
an effective Hamiltonian where you neglect all interactions. Call those
solutions |ψ01eff
(t) and |ψ01
exact
(t).
(4) Estimate the fidelity F = | ψ eff |ψ exact  |2 between the states evolved with
the idealized model, where effective interactions have been disconnected,
and the real one. What reasonable values can J, 1, 2 and t take to
achieve an infidelity 1 − F ≤ 10−1,10−2 and 10−4 , respectively?
(5) Repeat the previous calculation using the qubit parameters from the work
by Barends et al. (2014).

https://doi.org/10.1017/9781316779460.006 Published online by Cambridge University Press

 7
Qubit–Photon Interaction

It can be argued that the work by Wallraff et al. (2004) included the two impor-
tant breakthroughs that superconducting circuits needed to become a mainstream
quantum technology. The first breakthrough was the jump in a qubit’s coherence
time that appeared by placing those qubits in the gap of a microwave resonator. The
improvement was so impressive that superconducting qubits regained momentum –
and strongly surpassed other solid state platforms – in the quest for scalable quantum
computers.
The second breakthrough in this work was to show how qubits could interact
coherently with the photons trapped in the resonator that protected the qubit. The
combined qubit–resonator sample evidenced an exchange of photons that was
compatible with the canonical Jaynes–Cummings or Rabi models from quantum
optics. Thus emerged the field of quantum optics with superconducting circuits,
circuit quantum electrodynamics or simply circuit-QED. This field brings to the
lab old ideas from quantum optics, such as quantum superpositions of photons
(Hofheinz et al., 2009), Schrödinger cat states (Vlastakis et al., 2013), and collective
effects in superradiance (Mlynek et al., 2014; Nataf and Ciuti, 2010). A large
body of fundamental results was accompanied with crucial developments in the
preparation, control, and measurement of superconducting circuis. Microwave
resonators became the tool to control and measure superconducting qubits (Schuster
et al., 2005), to create entanglement (Majer et al., 2007) and to assist quantum gates
between qubits (Rigetti and Devoret, 2010). Understanding how this happened is
the goal of this chapter.
We begin this important segment of the book introducing a model for light–matter
interaction between superconducting circuits and microwave photons. Roughly, we
quantize a large circuit with one or more qubits in a microwave transmission line, we
then identify linear and nonlinear excitations – photons and qubits – and establish
parallelisms between the resulting Hamiltonian and atoms in optical environments.
We first apply this model to study the dynamics of a qubit in an open environment,

156

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.1 Qubit–Line Interaction Models 157

explaining how an atom can emit photons, one at a time, and interact with a
propagating microwave field. We then make a drastic change, placing our qubits
inside microwave resonators. We show that photons in a resonator can interact more
strongly and for longer periods of time with our artificial atom. This leads to simpler,
radically more useful models, such as the Jaynes–Cummings Hamiltonian. We
explore applications of these circuit-QED setups to the engineering and tomography
of qubit and photon states, introducing the dispersive coupling, single- and two-tone
spectroscopy and quantum nondemolition (QND) measurements of qubits.

7.1 Qubit–Line Interaction Models

The central idea in circuit-QED is to engineer matter-like and photon-like excita-
tions using different plasmonic excitations of a superconducting quantum circuit.
The fact that such artificial photons and atoms are made of the same material means
we can design them with a good impedance match, favoring an easy, fast, and
lossless exchange of energy and information. In circuit-QED setups, the interaction
between the microwave photons and the superconducting qubits is comparatively
stronger than that of optical photons with atoms and molecules. This allows indi-
vidual qubits to fully reflect photons in transmission lines (Astafiev et al., 2010)
and to reach the ultrastrong coupling regime and break down the RWA (Forn-Díaz
et al., 2010; Niemczyk et al., 2010).
A typical setup, sketched in Figure 7.1, consists of one or more qubits coupled
to waveguides or other microwave resonators. In 2D waveguide setups, qubits sit
between the central conductor (bottom of the figure) and the ground plane (top),
while in 3D waveguides the qubits are inserted into the hollow of the copper or
aluminum tubes. As usual, the qubits exhibit localize excitations with an anhar-
monic spectrum. These atom-like states couple to the electric and magnetic fields
that propagate along the waveguide, and that are bound to the photon-like plasmonic
excitations of the superconducting microwave guide.

7.1.1 Dipolar Interaction

Let us begin with a transmon inside the coplanar coaxial waveguide from Figure 7.2.
As suggested by the sketch, the charge accummulated on the transmon interacts
with the capacitor formed by the central conductor and the ground plane. We can
describe this system using the mutual capacitance between the transmon and the
small segment of waveguide around the qubit. Alternatively, and consistent with
Section 6.5, we can model the transmon as an electric dipole sitting on the electric
field confined inside the waveguide. Both images are equivalent and lead to the same
mathematical description.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 158 Qubit–Photon Interaction

Figure 7.1 Transmon and flux qubit sitting inside a one-dimensional microwave
transmission line – central conductor at the bottom, ground plate on the top. The
qubits interact preferably with either the voltage or the current transported by the
waveguide, through their electric and magnetic dipoles.

In the lumped-element circuit model, the waveguide enters as a potential biasing

the qubit’s equilibrium point. The Hamiltonian is the one in Section 4.6, but the gate
voltage is an operator V̂ (x) that contains the waveguide’s field around the qubit. The
effective model is
1  2
H = q̂ − Cg V̂ (x) − EJ cos φ̂/ϕ0 + Hphotons . (7.1)
2C

We split this Hamiltonian into three contributions, Hqubit + Hphotons + Hint , two
slightly renormalized Hamiltonians for the qubit and the waveguide, and an interac-
tion term. The interaction between charge qubits and photons, when projected onto
the qubit eigenbasis, adopts the dipolar form1 from Section 6.5
Cg
Hint = − q̂ V̂ (x)  εσ x V̂ (x). (7.2)
C
As before, the voltage stimulates transitions between both qubit states.
In addition to the coupling term, the qubit-line interaction has slightly modified
the qubit and transmission line properties. It adds a correction to the transmon’s
capacitance, which is now larger C  CJ + Cqubit-line . It also adds a local term
to the line V̂ (x)2 , causing a renormalization of the transmission line’s properties.
Potentially, this effect can be so large as to expel all plasmons away from the qubit’s
position xqb , zeroing V̂ (xqb ) and decoupling the qubit from the line (García-Ripoll
et al., 2015). In most experiments, however, the qubit-line coupling is small, and
both effects can be neglected – or maybe even undistinguished from a change in the
qubit’s properties during fabrication.
We can build a similar argument for the flux qubit. The qubit’s loop hosts a
superposition of circulating currents and behaves as a small magnetic dipole μ.
This dipole interacts with the magnetic field generated by the curents on the line, as

1 If we follow the same convention as before, the transmon would couple using the σ y operator. However, for all
applications discussed in this chapter, we can use a local unitary transformation to transform σ y → σ x .

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.1 Qubit–Line Interaction Models 159

in gμ · B(x). Another way to reach the same conclusion is to see the currents in the
line as generators of a magnetic field that modifies the flux inside the qubit . As
in Section 6.4, we find this perturbation induces opposite energy shifts of the |L
and |R current states. When we define the qubit states at the symmetry point, the
shift adopts the same dipolar coupling form as for the charge qubit:

Hint = M Iˆqubit Iˆ(x) ∼ εσ x Iˆ(x). (7.3)

The coupling constant ε now accounts for the qubit–line mutual inductance M and
the qubit’s supercurrent intensity I˜c .

7.1.2 Spin-Boson Hamiltonian

To complete our model, we must include the quantum nature of the transmission
line, replacing voltage and current operators with their expressions in terms of pho-
tons. We assume a homogeneous waveguide, with uniform inductance and capac-
itance per unit length l and c. The current is defined as Iˆ(x) = ∂x φ̂(x)/ l, while
the voltage can be derived from the Heisenberg equation V̂ (x) = ∂t φ̂(x,t) =
−i[φ̂, Ĥphotons ]/h̄. Using waveguide modes (5.19),
  
 e−ikxm † eikxm
h̄ωk
V̂ (xm ) = √ b̂k − √ b̂k , (7.4)
k
2c
d d
  
 h̄k 2 k eikxm e −ikxm
Iˆ(xm ) = i √ b̂k − √ b̂k . †

k
2cωk l d d

Inserting these in the qubit–line Hamiltonian, we get the spin-boson model:

 h̄ z  x  
Ĥ = h̄ωk ak† ak + σ + σ gk ak + gk∗ ak† . (7.5)
k
2 k

The Hamiltonian is the same for capacitively and for inductively coupled qubits,
with minor changes in the coupling constants:
 
cap cap h̄ωk ind k h̄
gk = −iG , and gk = iG
ind
. (7.6)
2cd l 2cωk d

The prefactor G encapsulates the microscopic details of the qubit–line coupling. For
the transmon, it is proportional to the qubit–line capacitance Gcap = −2eCg /C .
For a flux qubit, it combines the qubit–line mutual inductance M with the qubit’s
critical supercurrent Gind = M I˜c .

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 160 Qubit–Photon Interaction

Many 2D waveguides exhibit a close to linear dispersion relation ωk = v|k|,

with group velocity v 2 = 1/cl. We can then show that both couplings scale iden-
√
tically with frequency ∼O( ωk ) and have a simple expression in terms v and the
√
waveguide’s impedance Z = l/c, as
 
cap h̄Z vω k h̄ vωk
gk = −iGcap × , gkind = i sgn(k)Gind × . (7.7)
2 d 2Z d
The spin-boson Hamiltonian (Leggett et al., 1987) is one of the canonical mod-
els for studying dissipation in quantum mechanics.2 The spin-boson model is the
simplest mathematical realization of a small and discrete quantum system – e.g.,
an electron in an atom or a molecule – coupled to a big bosonic environment. In
experiments, the environment is so large and evolves so quickly that it is immune to
the qubit’s tiny perturbations. As we did with the resonator, we can trace the envir-
onment out and develop a master equation just for the qubit. The resulting equation,
already introduced in Section 6.1.6, models the relaxation and decoherence of the
qubit accurately, with few parameters to calibrate experimentally.
In this chapter, we plan to dig deeper, using the spin-boson model also as the
fundamental tool in our study of circuit-QED and qubit–photon interactions. Instead
of ignoring or tracing out the environment, we will treat both the qubit and the
waveguide on equal footing. By studying the influence of the qubit on the envir-
onment, we will see that decoherence and losses are a consequence of the qubit’s
capacity to absorb and emit quantua of excitations or photons. These photons can
be used for spectroscopy, monitoring the qubit, transporting information, and medi-
ating interactions between quantum devices.

7.1.3 Spectral Function and Spin-Boson Regimes

The spin-boson model (7.5) has three sets of parameters: the qubit’s gap , the spec-
trum of bosons ωk , and the distribution of coupling constants gk . We can combine the
last two sets of parameters into the spectral function J QO (ω) from (5.63). This was a
mathematical object that we used to analyze the memory function of a bosonic bath,
when coupled to a microwave resonator. We then argued that the spectral function
evaluated at the cavity’s resonance frequency ω provided a good estimate of the
decay rate κ = J QO (ω) at which the cavity equilibrates to the bath’s temperature.
We can extend this intuition to the study of a qubit in a photonic environment, but
we have to consider additional subtleties.

2 The other one, the Caldeira–Leggett model, was used in Chapter 5 to describe losses in a resonator.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.1 Qubit–Line Interaction Models 161

The first subtlety is the distinction between continuous and discrete spectral func-
tions. The former are associated to extremely long waveguides,3 supporting all
frequencies and wavelengths. Discrete functions, on the other hand, appear when
we place the a superconducting qubit inside a λ/2 or λ/4 resonator, and we can
resolve the individual eigenfrequencies – with the qubit typically interacting with
just one or few of those modes.
Even in the continuous case, we can also find gapped spectral functions, which
are zero over an extended range of frequencies. A common example is a 3D
microwave guide made from a long tube of aluminum or copper. The waveguide
only supports modes propagating with a frequency that lays above a minimum
energy ωk ≥ ωcutoff , which is needed to excite the transverse electromagnetic profile
of the waveguide’s modes. Below this frequency, the waveguide has no excitations
that can absorb energy from the qubit. When the qubit is placed inside the bandgap,
 ∈ [0,ωcutoff ], the waveguide acts as a filter that blocks all excitations, protecting
the qubit from decoherence.4
Conventional waveguides on 2D chips are approximately gapless, offering exci-
tations from very low frequencies up to some intrinsic cutoff.5 The spectral function
of such waveguides grows linearly with the frequency of the excitations, describing
an Ohmic spin-boson model:
J QO (ω) = παω1 . (7.8)
This regime is to be contrasted with other polynomial spectral functions J (ω)  ωs
with smaller and bigger exponents, called the sub-Ohmic s < 1 and super-Ohmic
s > 1 regimes.
We can verify (7.8), rewriting the coupling constants from (7.7) as
√
|gk | = h̄ παvωk /2d, and inserting them into the definition of the spectral func-
tion (5.63):
 2πv παωk 
J (ω) = 2×
QO
δ(ω −ωk )  dωk παωk δ(ω −ωk ) ∼ παω1 . (7.9)
k>0
d 2

This assumed a linear dispersion relation, a uniform density of states dωk ∼ vdk 
2πv/d, and two directions of propagation for each frequency.
The Ohmic spin-boson model exhibits a rich phenomenology, depending on
the dimensionless ratio α, describing the qubit–photon interaction strength and
the competition between incoherent processes (decay) and coherent processes

3 Experiments don’t allow “infinite” lengths, but a waveguide shorted by an impedance-matched resistor
behaves as an infinitely long medium, too (Devoret, 1995).
4 The qubit can still radiate to free space and into its substrate. These are other environments that are rarely
modeled, but grouped into intrinsic or nonradiative losses.
5 The superconducting gap is at least one possible cutoff.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 162 Qubit–Photon Interaction

Table 7.1. Regimes of the Ohmic spin-boson model.

0 < α < 1/2 Markovian Exponentially damped oscillations of the

qubit to a state with a small, nonzero
excitation probability.
α = 1/2 Tolouse point Exactly solvable point where exponential
decay is exact.
1/2 < α < 1 Kondo regime Mapping to the Kondo problem.
Non-exponential decay. Problem cannot
be approximated by a master equation.
Strong correlation between the qubit and
the photonic environment.
1<α Localization Quantum tunneling between qubit states is
suppresed. Dynamics amounts to
dephasing due to the bosonic environment.

(qubit rotation with the Hamiltonian). As we summarize in Table 7.1, there are four
distinct phases and behaviors depending on α.
The first region α ≤ 1/2 is where coherent dynamics overcomes decoherence at
short times: The qubit evolves between the excited and ground states, with oscil-
lations that dampen at a rate γ . This dynamics is very well approximated by a
memoryless or Markovian master equation (6.17), with a decay rate that, for small
values α  1/2, is provided by the spectral function γ = J QO ()  πα. Note
how γ / = πα so that α can be regarded as the relative speed of spontaneous
emission, as compared to the qubit’s intrinsic dynamics .
As α approaches 1/2, the dynamics of the qubit follows a slightly corrected
Markovian dynamics, due to the entanglement between the qubit and its electro-
magnetic environment. We regard the qubit as dressed by a cloud of photons that
slows it down, shifting its frequency  to lower values, and changing the decay
rate γ , which is no longer proportional to the spectral function. These corrections
were estimated analytically by Shi et al. (2018) and observed experimentally by
Forn-Díaz et al. (2016).
When the coupling strength increases above α = 1/2, correlations between
the qubit and its surrounding photon cloud prevent any Markovian description.
The dynamics of this strongly correlated system is faster than exponential and
requires very different (mostly numerical) descriptions.6 The theory around this
point describes a qubit decay that is accompanied by the emission of odd numbers
of photons – one, three, five, etc. – with varying combinations of frequencies.

6 Path integral calculations, the Kondo renormalization group, or matrix product states (MPS) (Peropadre et al.,
2013; Shi et al., 2018) provide insight into this region.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.1 Qubit–Line Interaction Models 163

Finally, if we push the coupling strength above α = 1 we cross a quantum phase

transition into the localization phase. In this region, the qubit is effectively frozen
and quantum tunneling is suppressed. Qualitatively, this may be derived from the
spin-boson model in the limit,   0. In this limit, the coupling σx (ak + ak† )
produces two eigenstates with opposite orientations of the qubit, |± ∝ |0 ± |1.
These states are accompanied by displacements of the bosonic field along opposite
directions |±gk /ωk . When g is very large, both ground states are so different that
tunneling between them is virtually impossible and the qubit is frozen, suffering
at most a dephasing in the |± basis due to the evolution of the different photonic
environments.
Approaching or exceeding the Tolousse point α = 1/2 is extremely hard for
atomic and molecular systems. However, experiments by Forn-Díaz et al. (2016)
have showed a very good impedance match between flux qubits and photons that
approaches α = 1/2 and even overflows into the α > 1 region. These experiments –
and also others that take place in cavities under similar conditions (Niemczyk et al.,
2010; Peropadre et al., 2010; Yoshihara et al., 2016) – are generically referred
to as achieving the ultrastrong or even deep-strong coupling regime. They offer
interesting possibilities for studying non-Markovian open quantum systems, strong
light–matter and photon–photon correlation effects, and applications to the design
of quantum simulators (Kurcz et al., 2014; Pino and García-Ripoll, 2018; Shi et al.,
2018). However, before discussing these interesting regimes, we will first focus
on the equally useful regime of α  1/2, where most quantum computing and
circuit-QED experiments take place.

7.1.4 Rotating Wave Approximation

When the qubit–bath coupling is the slowest timescale, α  1/2 and we can make
the rotating wave approximation (see Section 5.5). We neglect counter-rotating
terms of the form σ + ak† + σ − ak , and keep the coherent exchange between the qubit
and the bosonic field σ + ak + σ − ak† . The resulting model is the Jaynes–Cummings
Hamiltonian:
 h̄ z   
Ĥ = h̄ωk âk† âk + σ + gk σ + âk + gk∗ âk† σ − . (7.10)
k
2 k

In the RWA limit, every excitation of the qubit is accompanied by the subtraction
of a photon from the environment. Such dynamics conserves the total number of
excitations given by
 † σz + 1
N̂ = âk âk + . (7.11)
k
2

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 164 Qubit–Photon Interaction

This operator commutes with the Hamiltonian [N̂,H ] = 0, is a constant of motion

d
dt
N̂ = 0, and can be used to classify the eigenstates of the problem. The Hilbert
space splits into subspaces H = H0 ⊕ H1 ⊕ H2 ⊕ · · · , labeled by the integer
number of excitations N̂ = 0, 1, 2, . . . , which are not coupled by the dynamics
with the Hamiltonian.
The ground state manifold contains a single state H0 = {|0,vac} with the qubit
in the unexcited state and an empty waveguide without photons – the vacuum state.
The single excitation manifold H1 contains an infinite family of wavefunctions with
a simple parameterization, describing quantum superpositions of the excited qubit
|1,vac and a propagating photon |0,φ(x) with any possible spatial profile φ(x).
This set of states will be used in Section 7.2 to study the spontaneous decay of a
qubit and the interaction of propagating photons with a two-level system. Finally,
states with two or more photons HN≥2 require more sophisticated techniques – e.g.,
input–output theory, scattering theory, path integral formulations, or numerics – and
will be mostly ignored in this book.
It is important to note that we can derive RWA models that go beyond the two-
level approximation. These models are relevant for studying transmons in open
transmission lines. In this case, the interaction term will be slightly more complex
and involves transitions between more than two states, i.e.,
  i,j
Ĥ = i |ii| + gk |i,j â|k + H.c. + · · · . (7.12)
i i,j

This is still a relatively simple Hamiltonian, with a ground state that consists of the
vacuum in the bosonic environment and an unexcited transmon, |0 ⊗ |vac. Inter-
estingly, if we allow for cyclic transitions, such as g 01,g 12,g 02 = 0, the number of
excitations will no longer be conserved. This prevents using the theoretical methods
in this chapter, but enables interesting side-effects and processes such as splitting
a photon into two other frequencies (Sánchez-Burillo et al., 2016) or facilitating
interactions between photons at different frequencies (Hoi et al., 2013).

7.2 Waveguide-QED
Waveguide-QED is a relatively new subfield of quantum optics that studies the
interaction between propagating photons confined in 1D systems and few-level sys-
tems. The term includes broad families of experiments. The photonic medium could
be nanofibers, plasmons, or photonic crystals that interact with ultracold atoms,
quantum dots, or color centers. Of course, waveguide-QED also describes the super-
conducting setups from Figure 7.1.
In this section, we introduce various tools that are developed in the context of
waveguide-QED, focusing on the RWA limit with excitations and states |ψ ∈ H1

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.2 Waveguide-QED 165

Figure 7.2 (a) An excited qubit relaxes by emitting one photon in a superposition
of directions in an empty line. (b) Conversely, if we send a photon against a qubit,
it will be partially reflected and partially transmitted, with amplitudes R and T .

that satisfy N̂ |ψ = 1 × |ψ. As sketched in Figure 7.2 in this regime we expect
only two relevant physical processes: the spontaneous emission of a photon by an
excited qubit; and the scattering of a propagating photon by a relaxed qubit. Both
processes can be modeled using a quantum superposition of qubit and photon excita-
tions, known as the Wigner–Weisskopf ansatz. Despite its simplicity, the wavefunc-
tion model captures important concepts, including spectroscopy, Markovianity, and
decoherence.

7.2.1 Wigner–Weisskopf Approximation

The Wigner–Weisskopf model is a mathematical ansatz or trial state, built from a
coherent superposition of an excited qubit and a propagating boson:
) *

+ †
|ψ(t) = c1 (t)σ + ψk ak |0,vac . (7.13)
k

This ansatz represents the most generic state within the single-excitation sector
of the RWA Hamiltonian (7.10). We therefore make no mistake when we project
the Schrödinger equation for this Hamiltonian onto this subspace, and obtain a
differential equation for the amplitudes of the qubit’s c1 (t) and one for the boson’s
excitations ψk (t) (see Problem 7.1):

i∂t c1 = c1 + gk ψk, and i∂t ψk = ωk ψk + gk∗ c1 . (7.14)
k

As in our study of the lossy cavity, we integrate the equation for the propagating
photon ψk (t), beginning from a time far away in the past t0 → −∞. The math-
ematical procedure is identical7 to the one used in Section B.3.2, replacing the

7 The solutions are formally identical because we are dealing with linear differential equations!

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 166 Qubit–Photon Interaction

operator a(t) with the complex amplitude c1 (t). This method produces an exact
integro-differential equation for the qubit’s amplitude:
 t
∂t c1 (t) = −ic1 (t) − iξ (0,t,t0 ) − K(t − τ )c1 (τ )dτ . (7.15)
t0

The equation contains a source term that describes the incoming photonic field at
the qubit’s position, derived from the initial condition of the waveguide ξ (x,t,t0 ) =
 ikx−iωk t
k gk e ψk (t0 ): It also contains an integral with the memory function K(t)
of the bath (5.62). In the Markovian limit, where the bath loses all memory of the
qubit’s past, it is replaced with a complex number K(t − τ ) 
(iδLamb − 12 γ )δ(t − τ ):
 
1
i∂t c1 =  − i γ c1 + ξ (x,t,t0 ). (7.16)
2
As in the cavity, the Lamb shift δLamb = − describes the slowdown of the qubit
due to the surrounding photons, while the spontaneous emission or qubit decay rate
γ  J QO () models the relaxation of the qubit to the ground state.

7.2.2 Input–Output Relations

We can establish a relation between the field that illuminates the qubit ψk (t → −∞),
the photons absorbed and emitted by the qubit, and the total field that departs
from the qubit position ψk (t → +∞). The result is analogous to the input–output
relations from Section 5.5.2.
To discriminate incoming from scattered photons, we work in position space,
defining left- and right-moving fields that group positive and negative momenta

± (x,t) = d −1/2 k∈R+ e±ikx−iωk t ψk (t). As usual, d represents the length of the
waveguide, and the open transmission line is recovered in the limit d → ∞. Using
the formal solution for ψk (t), we can express these fields:
  t eikx−iωk (t−τ )
± (x,t) = ± (x,t,t0 ) − i gk∗ √ c1 (τ )dτ . (7.17)
k∈R+ 0
t d

This expressions combines a term ± (x,t,t0 ) that depends on the initial condition
ψk (t0 ), with a photon emitted by the qubit ∝ c1 (t). We distinguish two solutions
depending on whether our boundary condition is set back in the past – an input field
 in (x,t) = ± (x,t, − ∞) – or in the far future – the output field  out (x,t) =
± (x,t, + ∞). To move further, we use our recurrent assumptions:
(1) Without loss of generality, we place the qubit at x = 0.
(2) We assume a uniform spacing of momenta, ∼2π/d.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.2 Waveguide-QED 167

(3) The coupling strength

√ gk will change very little around the qubit frequency,
scaling as gk = g/ d.
(4) The density of states around  is approximately uniform, and we can linearize
the dispersion relation ωk  v |k| with group velocity v.
We can relate the coupling strength g to the spontaneous emision rate:
 
|g|2
J (ω) = 2π
QO
|gk | δ(ω − ωk ) = 2
2
δ(ω − ωk )dωk, (7.18)
k
v
√
which implies that γ = 2 |g|2 /v and gk = γ v/2d. With this parameterization
and the Markovian assumption from earlier sections, we compute the emission of
the qubit:

  t γ v e±ikx−iωk (t−τ )
± (x,t) − ± (x,t, ± ∞) = −i √ c1 (τ )dτ
t 2d d
k∈R+ 0
  
γ v +∞ dωk t
 −ic1 (t) dτ e±iωk x/v−iωk (t−τ )
2 0 2πv t0

1 γ
= −i sign(t0 ) c1 (t ∓ x/v).
2 2v
(7.19)
The photon field that reaches the qubit from the left or from the right ± (x,t) cannot
depend on whether we used the input or output boundary condition to derive it.
Equating both alternatives gives an input–output relation between incoming and
outgoing fields on both directions:

γ
± (0,t) = ± (0,t) − i
out in
c1 (t). (7.20)
2v
We must read this as a relation between the field just before and just after interacting
with the qubit,  in and  out , and the photons absorbed and emitted by the qubit.
When we look at this equation, it may seem as if right-moving (+in,out ) and left-
moving (−in,out ) fields are decoupled. This is not true, because the qubit field c1 (t)
is fed by both sets of photons and present in both equations. This means the qubit
can absorb a left-moving photon and reemit it as a right-moving photon, and vice
versa.

7.2.3 Spontaneous Emission Spectrum

Consider an excited qubit c1 (t0 ) = 1 sitting on an empty waveguide ψk (t0 ) = 0.
The Wigner–Weisskop theory predicts that the qubit will exponentially decay to its

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 168 Qubit–Photon Interaction

Figure 7.3 Spontaneous emission of a photon by a transmon qubit, as observed

in the experiment by Sharafiev et al. (2021). At times t < 0, the transmon is
being constantly driven. The drive stops at t = 0. The transmon then relaxes to
a ground state, emitting a photon. Despite some experimental artifacts caused
by amplification – e.g., dark fringe close around t = 0, one may appreciate the
narrowing of the photon spectrum over a time t  1/γ ∼ 0.3 μs, until it converges
to the Lorentzian profile (7.22).

ground state P1 = |c1 |2 = exp(−γ t), emitting a photon into the waveguide. This
phenomenon is called spontaneous emission, and the constant γ is the spontaneous
emission rate of the qubit.
The input–output relations (7.20) allow us to compute the wavepacket of the emit-
ted photon in terms of the qubit’s dynamics. This photon is a truncated exponential,8
with a maximum amplitude at the wavefront and an exponentially decaying tail:

γ (−i −γ /2)(t∓x/v)
± (x,t) = −i (t)
out
e . (7.21)
2v
Once the photon has been fully emitted, the Fourier transform of the wavepacket
reveals a normalized Lorentzian spectrum,9 centered on the qubit’s frequency :
1 γ /2
|ψk |2 = . (7.22)
π γ /4 + ( − ωk )2
2

The full-half-width (FHW) of this profile – i.e., the bandwidth of the photon or
linewidth of the two-level system – is just the emission rate γ .
The spontaneous emission of a photon has been experimentally observed in fre-
quency and in time by Sharafiev et al. (2021). The spectrum of the emitted photon
changes in time, as shown in Figure 7.3. The Lorentzian prediction is only strictly
true at long times t  1/γ , once the photon has been fully emitted. At very short
times 1/t    γ , the Markovian approximation still does not apply and the
qubit deposits energy in all frequency modes, with a rather homogeneous spectrum

8 With the Heaviside function (t) = 0 for t < 0, (t) = 1 for t > 0.
9 Compare this with the cavity transmission spectrum from Section 5.5.3.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.2 Waveguide-QED 169

|ψk |2 ∝ |gk |2 . At intermediate times, the photon looks like a truncated version of
the exponential wavepacket, shortened to a length t from its wavefront. During
those times, the width of the photon goes as ∼1/t  γ , fulfilling the so-called
time-energy Heisenberg uncertainty relation – i.e., the shorter time we measure, the
longer the uncertainty in the energy we obtain.
During spontaneous emission, the qubit and the waveguide entangle with each
other. For this reason, if we trace out the waveguide we obtain a quantum state with
less information – i.e., a state that has decohered. Consider a qubit that is initially
prepared in the quantum superposition, in an empty waveguide:
|ψ(0) = (α |0 + β |1) ⊗ |vac . (7.23)
This state will decay in time, producing a photon only when the qubit is in the |1
state. We can derive the dynamics of the qubit-waveguide system using the Wigner–
Weisskopf ansatz, including the state that does change |0,vac. When we trace out
the photonic environment, the resulting density matrix ρqb = trphotons (ρ) is exactly
 2 
|α| + (1 − |β|2 e−γ t ) αβ ∗ e−it−γ t/2
ρqb = . (7.24)
α ∗ βe+it−γ t/2 |β|2 e−γ t
The qubit begins in a pure state10 in a quantum superposition with coherences ρ01 ∝
α ∗ β. The superposition deteriorates, as evidenced by the decay of the coherences
and of the excited state’s probability. Asymptotically, as t grows to infinity, the qubit
returns to a pure ground state. At this point, all the quantum information from the
qubit has been transferred to a quantum superposition of a propagating photon and
a vacuum state.
The time evolution of the qubit’s density matrix is analogous to the one predicted
for a qubit in a zero-temperature bath (6.19). It exhibits a decay time T1 = 1/γ
and a decoherence time T2 = 2/γ , given by the qubit’s spontaneous emission rate.
As shown in Problem 7.4, this analysis can be generalized, verifying that both
the Wigner–Weisskopf ansatz and the master equation (6.17) provide equivalent
descriptions of the emission process.

7.2.4 Single-Photon Scattering

Imagine that, as depicted in Figure 7.2c, we send one photon against a relaxed
qubit in a transmission line. Since the number of excitations is conserved and the
qubit always relaxes to the ground state, the outcome of this experiment must be a
transmitted and/or a reflected photon. The analysis of the transmitted and reflected

10 Pure states are projectors, which satisfy ρ (0)2 = ρ (0).

qb qb

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 170 Qubit–Photon Interaction

fractions is called single-photon spectroscopy. Like the coherent spectroscopy from

the cavity in Section 5.5.3, it informs us about the qubit renormalized frequency 
in the open environment, and the combination of the qubit’s spontaneous emission
rate γ with other nonradiative decay and dephasing channels.
We can use the input–output relations to analyze this experiment. We rewrite
(7.16), including the input photons that approach and drive the qubit:

γ  in 
i∂t c1 (t) = + (0,t) + −in (0,t) + ( − iγ )c1 (t). (7.25)
2v
This equation is solved as a convolution of the input field with the exponentially
decaying wavefront of a reemitted photon:
  t
γ  in  
c1 (t) = −i + (0,τ ) + −in (0,τ ) e(−i −γ /2)(t−τ ) dτ . (7.26)
2v tin
As usual, the output field is derived from the qubit’s state by (7.20).
Since the scattering relation is linear, we can decompose the incoming wavepack-
ets into simpler states and analyze their scattering. We choose the basis of quasi-
monochromatic fields that are adiabatically switched on:11
 +
(−iω+0+ )t a e(−iω+0 )t , t < 0
± (0,t) ∝ (−t)a± e
in
= ± (7.27)
0, otherwise.
Such wavepackets are infinitesimally narrow in frequency space. When integrated,
they give the following output fields for t ∈ (−∞,0]:
+ )t + )t γ a+ + a−
±out (0,t) = b± e(iω+0 = e(iω+0 a± − . (7.28)
2 i( − ω) + γ /2


The way we read this equation is that an input field coming from the left (a+ = 1) or
from the right (a− = 1) is reemitted by the qubit in all directions, producing nonzero
amplitudes b+,b− .
The input amplitudes and output amplitudes, a± and b± , are related by a unitary
transformation, called the scattering matrix, which contains the transmission and
reflection coefficients, Tω and Rω :
     (
b+ Tω Rω a+ Tω = 1 + Rω,
= , with (7.29)
ω = − 2 i( −ω)+γ /2 .
γ 1
b− Rω T ω a− R

Figure 7.4 illustrates the transmission and reflection probabilities, |Tω |2 and |Rω |2 ,
and the frequency-dependent phase shifts experienced by the photon. Note how

11 The term e0+ t is a normalized exponential 1 et/2 with  > 0, so that the wave vanishes in the distant past

t → −∞. At the end of the calculations, we take the limit in which  approaches 0 from above to recover
monocromatic fields.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.2 Waveguide-QED 171

(a) (b)
1.0 1.0
|T 2| R/
0.8
0.5
0.6
0.0 T/
0.4
– 0.5
0.2
|R2|
0.0 – 1.0
–4 –2 0 2 4 –4 –2 0 2 4
/ /

Figure 7.4 (a) Tranmitted and reflected probabilities for an incoming photon, as a
function of the photon–qubit detuning δ = ω −  . (b) Phase slip (in units of π )
of the transmitted and reflected photon, ϕT and ϕR , respectively.

|Rω |2 displayw a Lorentzian profile, centered on the qubit’s resonance. This is to be

expected, since the reflected field only contains the energy produced by the qubit’s
reemission of the input field. This Lorentzian peaks exactly on resonance  = ω, a
point at which Rω = −1, Tω = 0, and the qubit acts as a perfect mirror for a single
photon.
The scattering profiles from a qubit, and the fact that it acts as a perfect mirror
for individual photons, were first demonstrated by Astafiev et al. (2010), using flux
qubits. In this experiment, the qubit had some intrinsic dephasing γφ and experi-
enced some intrinsic nonradiative losses γ  , meaning that it could dissipate energy
to other environments, different from the waveguide. The transmission spectrum
under nonideal conditions can be studied with a small modification of the theory,
which treats the qubit using a more general master equation while preserving the
input–output relations. This gives
γ 1
Tω = 1 − . (7.30)
2 i( − ω) + γφ + (γ + γ  )/2


The single-photon spectroscopy reveals the qubit’s renormalized frequency  , but

the linewidth is a function of the combined decoherence rate γφ + (γ + γ  )/2. The
qubit is only a good mirror when it is in strong coupling regime, that is, γ  γ ,γφ .
In the waveguide-QED literature, the strong-coupling regime is characterized by
the ratio β = γ /(γ + γ  )  1. Astafiev et al. (2010) were the first researchers to
simultaneously demonstrate strong coupling between a qubit and an open waveg-
uide, and to implement single-photon scattering for that setup. In contrast to early
attempts with atoms or quantum dots, the superconducting setup benefited from the
good impedance match between the qubit and the line, and from the good isola-
tion provided by the coplanar waveguide. After the work by Astafiev et al. (2010),

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 172 Qubit–Photon Interaction

we have seen other works using the strong-coupling regime to control photons
(Hoi et al., 2012), to engineer strong photon–photon interactions (Hoi et al., 2013),
and to explore all regimes in the Ohmic spin-boson model (Forn-Díaz et al., 2016).

7.2.5 Quantum Links

One important and emerging application of waveguides is the connection of super-
conducting quantum processors in the same or in separate dilution refrigerators. As
we saw in Section 7.2.3, we can map the quantum information stored in a super-
conducting qubit to the state of a traveling photon. The goal of these quantum links
is to do this mapping in a transitive way, mapping the state of a qubit into a photon
that is then absorbed by another qubit at the other side of the waveguide.
In order to achieve perfect state transfer, we need three ingredients. First, we
need to control the coupling between the qubit and the waveguide, achieving a time-
dependent control of γ (t). In particular, we need to switch off the coupling at the
end of the transfer, to ensure that the qubit that absorbed the incoming photon does
not decay again!
Second, we need to control the direction along which the photons propagate, to
ensure that all the information goes from the first to the second qubit. In absence
of chiral waveguides, where photons can only move along one direction, this is
achieved by placing both qubits on the opposite sides of the guide.
Third, we need to control the shape of the photon wavepacket and the control
γ (t) to ensure that the second qubit experiences the time-reversed process of the
first qubit’s emission. As pointed out by Cirac et al. (1999), we need to engineer
the decay rate of the first qubit γ1 (t) to produce a photon that has a time-symmetric
wavepacket. We then design the spontaneous emission rate of the second qubit as
the time-reversed profile γ2 (tprop − t), where tprop = d/v is the time for the photon
to travel along the whole length of the waveguide.
This problem can be modeled as a Wigner–Weisskopf problem with two qubits
instead of the one we have used so far. Doing things properly, we will be able to
transfer a arbitrary quantum state (7.23) from one qubit to another, over relatively
short times T  tprop + O(1/γ ), of a few tens of nanoseconds for medium-size
waveguides. Interestingly, one may show that it is possible to implement perfect
state transfer protocol for photons whose duration O(1/γ ) exceeds the propagation
time between qubits. This is a regime where the input–output formalism does not
really apply, because the spacing between frequency modes in the waveguide –
the free spectral range – is too large for a continuum limit description. However,
Wigner–Weisskopf theory still produces a valid protocol, provided that we tune the
qubit close to one of the waveguide’s mode   ωk .

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 173

As mentioned previously, Magnard et al. (2020) have demonstrated quantum

state transfer between qubits hosted at separate refrigerators, over distances of 5 m.
The experiment used a long waveguide cooled at temperatures of tens of millikelvin.
It also used a sophisticated method to achieve the tunable qubit–photon interac-
tions where a microwave drive controls the transfer of the excitation |1 from each
transmon to a cavity, from which a photon rapidly leaks into the waveguide that
connects both fridges. Despite this complication, the physics still fits into a Wigner–
Weisskopf model, where the amplitude and phase of the microwave control a time-
dependent qubit–photon coupling, which is used to create and absorb the symmetric
photons required for state transfer. We expect that similar designs will enable future
local networks of quantum computers, supporting both quantum communication
and distributed quantum computation.

7.3 Cavity-QED
An artificial atom interacts coherently with the photons in a transmission line, but
these photons quickly fly away, leaving us with a decohered qubit. We can increase
the interaction time between the qubit and the photon by cutting the waveguide to
a shorter length, as shown in Figure 7.5. In this setup, the waveguide becomes a
microwave resonator or a cavity, and we are exploring the field of cavity quantum
electrodynamics, or cavity-QED.

500 mm

Figure 7.5 Photography of a transmon qubit standing near a microwave resonator.

The picture shows the transmission line that forms the cavity (a), a zoom of the qubit
(b), the lines that control the cavity (c) and the transmon’s gap (d), and a line that
interacts with the transmon to control its state (e). Picture courtesy of A. Wallraff,
with minor edits.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 174 Qubit–Photon Interaction

As explained in Chapter 5, the cavity is a quantum filter that only allows certain
microwave photons in. If the resonator is short, those frequencies are broadly spaced
(see Section 5.2.2), opening large gaps in the spectral function. If we park the
frequency of the qubit to lay deep in those gaps, the qubit is prevented from emitting
photons, and we preserve its quantum state for long times, only limited by the qubit’s
other nonradiative decoherence mechanisms. If instead we tune the qubit close to a
cavity’s resonance, the qubit will be able to emit photons that will bounce back and
forth among the resonator’s walls. The photons now interact with the qubit for longer
periods of time, only limited by the cavity’s decay rate. The Purcell enhancement
of the qubit–photon interaction allows the qubit to endow the resonator with highly
nonlinear properties, creating multiphoton states and engineering sophisticated non-
Gaussian states of light. Similarly, when the resonator is connected to the outer
world through a transmission line, we can use the cavity to efficiently control and
measure the state of the qubit.
Figure 7.5 shows a state-of-the-art circuit-QED experimental setup with the most
common ingredients. You should recognize (a) a transmission line resonator and
(b) a transmon qubit. The λ/2-resonator has an approximate length d = λ/2 ∼ 1
cm and supports an infinite number of harmonic resonances, with regularly spaced
frequencies ωn = nω1 , for n = 1,2, . . . (cf. Section 5.2.2 and Figure 5.3). The res-
onator interacts capacitively with a semi-infinite transmission line, used to read the
state of the cavity and to inject energy into it.
The transmon qubit (b) has a size of about 300 μm and sits inside the resonator,
between the core of the transmission line and one of the ground planes. The position
of the qubit is not arbitrary: It is close to the end of the cavity, because that is where
the electric potential of the resonator modes is maximal (cf. Figure 5.3), but it is
not too close so that it remains protected by the surrounding cavity ground planes.
The qubit also has some control and readout lines. Note the (c) line carrying some
current that sinks into the ground plane: The magnetic field generated by this current
affects the transmon’s SQUID inductance change and thereby changes the qubit’s
gap. Another line (d) applies a voltage into the qubit and is used to rotate the qubit
basis, as well as for readout.
Summing up, already this simple-looking circuit exhibits three control knobs and
two measurement setups. We can tune the two parameters of the qubit  and ε
as requested in Section 6.1.1, we can drive and displace the cavity field using the
ideas from Section 5.5.1, and we can perform spectroscopy of the qubit and of the
cavity by sending photons and studing the reflected signals. As we explain now,
these are extremely powerful controls that can be used as much for quantum optics
experiments as to engineer scalable quantum computers.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 175

7.3.1 Quantum Rabi and Jaynes–Cummings Models

Consider a qubit inside a zero-mode microwave resonator, which is itself connected
to a transmission line. This is a special case of the spin-boson model studied in
Section 7.1, where the qubit sees a big Lorentzian resonance:12
4κω0 ω
J QO (ω)  |g|2 2
, (7.31)
ω02 − ω2 + (κω)2
centered on the microwave cavity’s frequency ω0 , with a linewidth proportional
to cavity decay rate κ, and with a parameter g that describes the speed at which the
cavity and the qubit interact.
Typically, we would like to bring the qubit close to resonance with the cavity
  ω0 . We also want the qubit–photon coupling to be in some type of strong-
coupling regime, in which the qubit and the cavity exchange photons faster than the
speed at which photons escape from the cavity, |g|  κ. This way, the qubit can
perform meaningful operations in the photonic field before it decays.
Under these conditions, we begin to see that we cannot treat all photonic degrees
of freedom as a unit. The coherent exchange of excitations between the cavity and
the qubit is a highly non-Markovian situation. This is confirmed by the dramatic
change of the spectral function J QO (ω) across the range of frequencies involved in
the dynamics ∼ (ω0 − g,ω0 + g). Since we cannot approximate J QO as a constant,
the bosonic kernel K(t) becomes a function that oscillates rapidly, with frequency
∼ g; only at long timescales 1/κ the bosonic environment loses all memory about
the interaction with the qubit.
The solution is to consider the qubit with the cavity as a combined quantum sys-
tem, deriving a quantum Hamiltonian for both interacting objects, and only includ-
ing afterward the waveguide as a truly Markovian environment. When we quantize
the qubit in contact with the resonator, we obtain the single-mode limit of (7.5),
which is known as the quantum Rabi model:
h̄ z
σ + h̄ω0 â † â + h̄σ x g â + g ∗ â † .
Ĥ = (7.32)
2
The first and second terms are the qubit and cavity Hamiltonians, which are coupled
by a capacitive or inductive dipolar interaction g.
The quantum Rabi model is named after Austrian physicist I. I. Rabi (1898–
1988), who received the Nobel Prize for his studies on nuclear magnetic resonance.

12 See Zueco and García-Ripoll (2019) for a complete derivation of the cavity-QED model for all regimes of
losses in the cavity and the qubit.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 176 Qubit–Photon Interaction

NMR is the problem of one or more magnetic moments or spin driven by an

oscillating electromagnetic field. We addressed a reduced version of this problem in
Section 6.1.4, when studying the control of a qubit by a classical field.
Equation (7.32) is a more complicated version where the driving field is quantum
mechanical in nature.
In most experiments, the Rabi coupling is slower than the qubit’s and cavity’s
frequencies, |g|  ω,. By the RWA, we can eliminate the counter-rotating terms
σ + a + a † σ − , to produce the Jaynes–Cummings Hamiltonian:13
h̄ z
ĤJC = h̄ωâ † â + h̄ gσ + â + g ∗ â † σ − + σ , (7.33)
2
a model that is simpler and analytically solvable.
Using the theory of open quantum systems and the methods from Chapters 5
and 6, we can derive a master equation that includes the decay of the excitations
from the cavity into the waveguide with rate κ, coherent microwave drives for the
qubit ε(t) and the cavity (t), and the qubit’s intrinsic losses γ and dephasing γφ :
 
∂t ρ = − i ĤRabi/JC + h̄ε(t)σ x + h̄cav (â + â † ),ρ (7.34)
γ
+ 2σ − ρσ + − σ + σ − ρ − ρσ + σ − + γφ σ z ρσ z − ρ
2
κ
+ 2âρ â † − â † âρ − ρ â † â .
2
This master equation is accompanied by an input–output relation between the cavity
and the waveguide, which relates the field leaked into the line with the instantaneous
state of the cavity (5.65).
While it is clear that we need γ and γφ as small as possible – ideally below the
megahertz regime – the cavity decay rate κ requires a delicate compromise between
coherence and controllability. On the one hand, a small value of κ enables the qubit
and the photons to interact for longer times. On the other hand, a small value of κ
reduces our capacity to control and measure the experiment – the signals that leak
from the cavity are weaker (5.65), and it becomes harder to inject energy into the
microwave mode. Finally, for the coherent coupling g, we aim to be in the strong-
coupling regime, which in this case translates into |g|  γ ,γφ ,κ.
As an example, the original setup by Wallraff et al. (2004) used qubits and cavities
with frequencies around 6 GHz, with κ ∼ 0.8 × 2π MHz, γ ∼ 0.7 × 2π MHz, and
g ∼ 5.5 × 2π MHz. The more modern three-qubit mini-quantum-computing setup
by DiCarlo et al. (2010) explores a similar range of qubit and cavity frequencies, but

13 Interestingly, models (7.5), (7.32), and (7.33) were all posed by physicists Jaynes and Cummings to describe
an idealized atom in a quantum electromagnetic field. However, the Jaynes–Cummings or JC Hamiltonian
nowadays refers to RWA version (7.33).

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 177

Figure 7.6 Energy-level structure of a qubit and a cavity without coupling (solid),
as described by the Jaynes–Cummings model (7.33). The interaction couples neigh-
boring states |g,n and |e,n − 1, creating new eigenstates that are superpositions
of the original ones, and whose energies split even further (dashed).

now the cavity–qubit couplings were around 2π × 220 MHz, two orders of magni-
tude larger than the cavity losses, κ = 2π ×2.4 MHz, and γ ∼ 0.8−1.2×2π MHz.

7.3.2 Jaynes–Cummings Ladder

For times shorter than the decoherence times, ∼ 1/γ , 1/κ, we can study the cavity-
QED setup using the Jayness–Cummings (JC) model, without the master equa-
tion (7.34). Since the JC model is a variant of the RWA spin-boson model, it also
conserves the total number of excitations N̂ . In cavity-QED, it is customary to label
the ground and excited qubit states as |g = |0qubit  and |e = |1qubit , which allows
writing
1
N̂ = |ee| + â † â = â † â + (σ z + 1). (7.35)
2
The Hilbert space splits into a Jaynes–Cummings ladder of decoupled subspaces
with different values of N̂ and increasing energy, as sketched in Figure 7.6. At the
bottom of the ladder, we find the zero-dimensional subspace for the ground state
H0 = {|g,0}. All other subspaces are formed by all possible linear combinations
of two basis states, Hn≥1 = lin{|g,n , |e,n − 1}, one with the unexcited qubit and
n photons |g,n, and one where the qubit got excited at the expense of one cavity
photon |e,n − 1.
Since the number of excitations commutes with the Hamiltonian [N̂, Ĥ ] = 0,
the Hamiltonian also decomposes as a box-diagonal matrix, the direct sum of 2 × 2
matrices ĤJC = Ĥ0 ⊕ Ĥ1 ⊕ · · · of 1 × 1, each coupling one or two states within
a rung Hn of the JC ladder. This subdivision means that the whole dynamics is
integrable – in other words, we can analytical solve the evolution of the cavity–
qubit setup in all those subspaces, not just for one photon.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 178 Qubit–Photon Interaction

With the exception of the trivial ground state, H0 = h̄(1−)

2
|g,0g,0|, all other
Hamiltonians can be written as follows:

Hn≥1 = h̄ω n − 1
2
+ h̄n vn · σ̃, (7.36)

using the Pauli matrices

σ̃ x := |e,n − 1g,n|+|g,ne,n − 1| , σ̃ z := |e,n − 1e,n − 1|−|g,ng,n| ,

and the Rabi frequency n and unitary vector v:

  √ √ 
1 Re(g) n Im(g) n  − ω
n = ( − ω) + |g| n,
2 2 vn = , , . (7.37)
2 n n 2n
The two-dimensional Hamiltonians can be diagonalized with the usual tricks from
Section 2.2. Each rung of the ladder is associated two eigenenergies:

En,± = h̄ω n − 1
2
± h̄n, (7.38)

which split symmetrically around h̄ω n − 12 . We call these the lower and upper-
polariton branches, En− and En+ , because the eigenstates are hybridized states
(polaritons) of excitations living on the cavity |g,n and on the qubit |e,n − 1. The
qubit–photon coupling introduces a level repulsion between the polariton branches,
which separate by an amount proportional to the Rabi frequency 2n .
When the cavity and the qubit are straight on resonance  = ω, the splitting is
√
minimal 2 n|g| and the two polariton states become symmetric superpositions:
1
|ψn±  = √ (|g,n ± |e,n − 1) , n ≥ 1. (7.39)
2
√
If, on the other hand, the detuning is very large | − ω|  n|g|, the coupling
strength will be insufficient to hybridize the qubit and the cavity. The two eigenstates
will be then approximately close to the bare states |g,n and |e,n − 1, and will have
close noninteracting eigenenergies En± ∼ {h̄ωn − 12 h̄, h̄ω(n − 1) + 12 h̄ω}, with
small dispersive corrections to be discussed in Section 7.3.7.

7.3.3 Vacuum Rabi splitting

The single-polariton energy gaps E1− − E0 and E1+ − E0 are visible as the smallest
quanta of energy that an empty cavity-QED system can absorb. They were first
probed spectroscopically by Wallraff et al. (2004), measuring which microwaves
were transmitted by a cavity that contained a frequency-tunable charge qubit.
Figure 7.7a sketches the results from a similar experiment, as we move the qubit’s
frequency in and out of resonance with the cavity.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 179

Figure 7.7 (a) Transmission through the cavity as a function of the qubit bias, ε,
for  = 0.7ω and g = 0.05ω. With√the dashed line we plot the qubit energy levels
for g = 0. The linewidths assume γ κ = 0.005ω. (b) Transmission spectroscopy
for a flux qubit interacting with the third mode of a microwave resonator, in the
ultrstrastrong coupling regime g/ω = 16%. (c) Two different fits with the RWA
and Jaynes–Cummings and the full Rabi model.

As explained previously, in the regime of large detuning (points “A” in

Figure 7.7a), the qubit and the cavity cannot exchange excitations. In this situation,
the setup develops one qubit-like and one cavity-like eigenstates, of which only
the latter can be excited by the coherent microwave drive. For instance, if   ω,
the incoming microwaves only connect the ground state |g,0 with the the upper
polariton branch, which consists mostly of |g,1 state. The result is a single peak
in the cavity transmission sitting at E1+  h̄ω, with a linewidth proportional to the
decay rate of the cavity ∼ κ.
If we instead bring the qubit close to resonance with the cavity   ω (points “B”
in the same plot), the energy that we inject in the cavity will be efficiently exchanged
with the qubit. In this case, the two polariton branches will appear as resonances,
symmetrically placed around the cavity’s frequency with a separation proportional
to the qubit–cavity coupling strength h̄ω± ∼ h̄ω ± 2h̄|g|, and a linewidth propor-
tional to the decoherence rate of the combined qubit–cavity system.
This separation of peaks is known as the vacuum Rabi splitting, and it is a smok-
ing gun for coherent qubit–cavity coupling in circuit-QED. Naturally, for the peaks
to be distinguishable, their linewidth must be narrower than the splitting, which
takes us back to the strong coupling condition |g|  κ,γ introduced before.

7.3.4 Rabi Oscillations: Weak and Strong Coupling

Spectroscopy is not a conclusive evidence of the “quantumness” of a circuit – after
all, interacting classical oscillators also exhibit resonances and splittings, even at
finite temperature. A more convincing result would be to reconstruct the dynamics

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 180 Qubit–Photon Interaction

of the qubit in the cavity. Fortunately, the evolution under the full JC model can be
solved analytically for any state |ψ(0) ∈ Hn , using the trick (2.22)
 
e,n − 1|ψ(t) 1  
ψ n (t) = = e−iω(n− 2 )t cos(n t) − i sin(n t)vn · σ̃ ψ n (0).
g,n|ψ(t)
Take for instance the initial state |ψ(0) = |g,n. Our analytical solution predicts
a (partial) transfer of population from the ground to the excited qubit. The period of
the Rabi oscillations is determined by the Rabi frequency n , while the efficiency
of the transfer depends on the relative strength of the detuning | − ω|/n . Indeed,
the transfer is only perfect in the resonant case  = ω,
|ψ(t) = cos(n t) |e,n − 1 − i sin(n t) |g,n , n ≥ 1, (7.40)
√
where it happens with a frequency n = g n that is half the Rabi splitting.
Actual Rabi oscillations are always damped due to decoherence in the cavity
or the qubit. To analyze this, let us project our density matrix onto the different
sectors with a fixed number of excitations ρN (t) = PN ρ(t)PN . Following (7.34),
these matrices satisfy the master equation:
i 
∂t ρN = − Heff ρN − ρN Heff †
+ κ âρN+1 â † + γ σ − ρN+1 σ +, with (7.41)
h̄
h̄γ h̄κ
Heff = HJC − i σ + σ − − i â † â.
2 2
The matrix ρN is fed by excitations from higher sections ρN+1 and itself evolves
with a non-Hermitian “Hamiltonian” Heff , which is excitation number conserving,
and where the frequencies of the qubit and the cavity now contain imaginary correc-
tions. If our initial condition has N or less excitations, ρN+1 (t) = 0, we can solve
analytically the dynamics of ρN as a Schrödinger-like equation:
†
ρN (t) = e−iHeff t/h̄ ρN (0)e+iHeff /h̄ . (7.42)
We then find that the role of γ and κ in Heff is to exponentially reduce in time the
probability PN = tr(ρN ) of having N excitations (see Problem 7.9), as one would
expect from their interpretation as qubit and cavity loss rates.
Let us apply this idea to study the Rabi oscillations of an excited qubit in an
empty cavity ρ1 (0) = |e,0e,0|. The single excitation component of the density
matrix is a projector ρ1 (t) = |ψ(t)ψ(t)|, onto an unnormalized state |ψ(t) =
ce (t) |e,0 + cg (t) |g,1, which follows the equation
 
1 1
¯
c(t) = e−i 2 (+ω)t− 4 (κ+γ )t × cos(t)1 ¯
− i sin(t)v̄ · σ c(0), (7.43)

   
−ω κ −γ 2 g −ω κ −γ
¯
= g + 2 +i , v̄ = ,0, +i .
2 4 ¯
 ¯
2 4¯

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 181

1.0
(a) (b)
0.8

0.6

0.4
0.2

0.1 0.2

0.0 0.0
0 5 10 15 20 25 0 5 10 15 20 25

Figure 7.8 Dynamics in the weak and strong coupling regimes. (a) Excited state
population for the cavity–qubit system starting in the |e,0 state, as a function of
time and coupling strength g. (b) Horizontal cuts for g = 0.05γ ,0.1γ and 0.2γ .
Plots use ω =  = 1, γ = κ = 0.1.

The norm of |ψ decreases exponentially, as does the probability of having one
excitation in the system tr(ρ1 ) = !ψ!2 ∼ exp(−(κ + γ )t/4). By virtue of (7.41),
this probability is transferred to the ground state component of the density matrix
ρ0 (t), through processes of cavity decay and the qubit’s nonradiative losses.
Our analytic solution predicts Rabi-type oscillations between the |e,0 and
|g,1 states, modulated by an exponential envelope. Figure 7.8b explores these
√
oscillations for a broad range of the cooperativity factor g/ γ κ, with select
cuts plotted in Figure 7.8a. In the weak-coupling regime, decoherence is stronger
than qubit–photon interactions g  γ ,κ, and we cannot resolve the oscillations.
This is exemplified by the exponentially decaying solid line in Figure 7.8b. The
opposite regime, called strong coupling, appears when g  γ ,κ. Thanks to a longer
coherence time, we can detect one or several Rabi oscillations in the dashed and
dotted lines from Figure 7.8b.
The strong-coupling regime was considered the Holy Grail of quantum optics for
a long time. One of the difficulties in reaching strong coupling was the need for a
cavity with a large quality factor14 Q = ω/κ. The first AMO experiments in the
strong-coupling regime used superconducting microwave cavities that could trap
photons for a long time. These cavities were made to interact with flying photons
in Rydberg states in a series of experiments that culminated with the measurement
of Rabi oscillations and the creation of Schrödinger cats (Haroche, 2013). In retro-
spect, circuit-QED is the natural evolution of the experiments by Nobel Prize winner

14 Q loosely represents the number of times a photon can oscillate in a cavity before it decays.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 182 Qubit–Photon Interaction

Serge Haroche. We keep the good microwave cavities, but replace the atom with an
even more strongly interacting object – the superconducting qubit.

7.3.5 Ultrastrong Coupling

The strong-coupling regime starts where interaction overcomes losses. Similarly,
the ultrastrong-coupling (USC) regime arises when the light–matter interaction
speed g approaches the qubit and photon frequencies,  and ω, and we can no
longer neglect the counterrotating terms. In the USC, we have to abandon the
RWA and use the full Rabi model (7.32) because the JC model delivers wrong
predictions in the eigenstates, in the associated spectra (cf. Section 7.4.1) and in the
qubit–cavity dynamics.
First of all, in the USC regime, the qubit gets highly entangled with the photonic
field, to a point that even the ground state becomes a superposition of vacuum |g,0
with even-excitation states, such as |e,1, |g,2, etc. These changes in the ground
and excited states are enabled by the Rabi coupling ∼ σ x (g â + g ∗ â † ), a term that
favors a hybridization of the cavity with the qubit, through a state-dependent dis-
placement â → â − g ∗ σ x /ω. This qubit–photon hybridization also activates tran-
sitions between states that do not conserve the number of excitations, making the
dynamics richer and more interesting.
The differences are remarkable when we compare the spectra delivered by exper-
iments, such as Figure 7.7b, with the numerical predictions of the JC and Rabi
model, shown in Figure 7.7c. The existence of non-excitation-conserving processes
modifies the eigenenergies not only around the crossing points   ω, but also
at points where the qubit is highly off-resonant from the cavity. Notice how around
those points – ε = 0 in Figure 7.7c – the JC model predicts a huge dispersive energy
shift of the cavity, which is corrected by the Bloch–Siegert shift from the Rabi model
(Forn-Díaz et al., 2010).
The USC regime with a single quantum emitter was first demonstrated by Niem-
czyk et al. (2010) and Forn-Díaz et al. (2010), using flux qubits in microwave
resonators and reaching relative coupling strengths g/ ∼ 12%. Later experiments
by Yoshihara et al. (2016) have surpassed these values, obtaining ratios g/ω close
to and above 100%. In order to achieve these coupling strengths, those experiments
required the use of flux qubits, the only qubits with a large enough anharmonicity
as not to be “bridged” by the large interactions. The USC regime was possible
because, as mentioned before, superconducting microwave photons and supercon-
ducting qubits are very similar excitations that can interact very efficiently. It must
be remarked that, beyond mere curiosity and a challenging regime, the USC regime
is also a useful tool in the development of quantum simulators, because it enables
the implementation of tunable qubit–qubit Ising Hamiltonians and exploration of

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.3 Cavity-QED 183

interesting quantum phase transitions (Kurcz et al., 2014). However, we will not
have space available to discuss these applications in this book.

7.3.6 Multiple Qubits

Let us now consider an experimental setup where one resonator hosts multiple
qubits, such as DiCarlo et al. (2010). The RWA model for those setups is the Tavis–
Cummings Hamiltonian, an extension of the JC model:

N N 
 
h̄i
HTC = σiz + h̄ωâ † â + h̄gi σi+ â + H.c. , (7.44)
i=1
2 i=1

which can also be upgraded to a full master equation (7.34). In this model, the qubits
talk “collectively” with the resonator. The cavity exchanges excitations with all of
them and mediates interactions among the qubits.
Spectroscopically, the features of this model are very similar to the JC Hamilto-
nian. For simplicity, we will discuss the case of identical emitters, with i =  and
gi = g. In this limit, the TC becomes an RWA version of the Dicke model:
1
H = Ŝ z + h̄ωâ † â + g Ŝ + â + g ∗ Ŝ − â †, (7.45)
2
 N z
with the collective operators Ŝ ± = N ±
i=1 σ and Ŝ =
z
i=1 σ .
The cavity is coupled to a symmetric superposition of excitations in √ the qubit
space. This collective excitation acts as a new quasiparticle b̂† ∼ S + / N with
bosonic statistics, which interacts more strongly than the single-qubit case. In the
limit of many qubits,
√
H  h̄b̂† b̂ + h̄ωâ † â + N |g|(b̂† â + H.c.). (7.46)
This physics is analogous to many other collective phenomena, from lasing to Bose–
Einstein condensation, where a bunch of identical quantum systems act coherently
in a synchronized way, and this way they become “better” at doing something – in
this case, at producing a photon.
The ground state of this model is the trivial vacuum state |0̃ := |g,g . . . ,g,0.
The single-excitation space contains N + 1 states with one excitation on any qubit
{σi+ |0̃} or on the cavity â † |0̃. These states form one upper and one lower polariton
branch, with eigenfrequencies

1
E1± = h̄ω ± ( − ω)2 + g 2 N . (7.47)
2
Note the increased √
splitting between polaritons caused by the bosonic enhancement
of the interaction Ng. In addition to the polariton states, the Hilbert space is

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 184 Qubit–Photon Interaction

completed with N − 1 dark states, excitations that live in the qubits and do not
“see” the cavity. For instance, if we have N = 2 qubits, the only dark state is the
qubit singlet √12 (σ1+ − σ2+ ) |0̃.

7.3.7 Off-Resonant Qubits and Dispersive Coupling

Many experiments work with qubits and cavities in the non-resonant or dispersive
regime. In this regime, the qubit and the photons have very different frequencies,
with a separation or detuning so large that it cannot be bridged by the qubit–photon
coupling δ := | − ω|  n̄ g, suppressing the exchange of excitations between
the qubit and the resonator.
Why is this regime interesting? One reason is that we minimize the qubit deco-
herence by placing its frequency deep in the bandgap created by the resonator. As
explained in near the beginning of Section 7.3, away from the cavity resonance,
the spectral function is smooth and very small J QO (ω)  4κ|g|2 /δ 2 , signaling a
low probability that the cavity extracts energy from the qubit and dumps it into the
waveguide.
The second reason has to do with the possibility that qubits exchange exci-
tations via the resonator. The energy difference between a qubit and the cavity
makes it unfavorable to transfer excitations from a qubit to the cavity and vice
versa. However, it is still possible for excitations to jump from qubit to qubit
via the cavity, if the qubits are close in frequency. As we will soon see, these
virtual processes are terms that appear in perturbation theory and induce both
exchange interactions σi+ σj− as well as cavity-dependent frequency shifts of the
qubits σ z â † â.
The dispersive regime is best studied using degenerate perturbation theory via the
Schrieffer–Wolff transformation. Following Section A.3, we separate our Hamil-
tonian into a free term and an interaction Ĥ = Ĥ0 + g V̂ . Our perturbation is the

Tavis–Cummings term g V̂ ∼ i (σi+ â + â † σi− ). We compute the Schrieffer–Wolff
generator on the bare basis of the qubits and the cavity:
 gi
g Ŝ = σi+ â − â † σi− . (7.48)
i
 i − ω

The complete effective model is

  
 gi2 h̄  gi2
Heff = h̄ ω + σi â † â +
z
i + σiz (7.49)
i
i − ω 2 i i − ω
 h̄gi gj  
+ σi+ σj− + σi− σj+ .
i=j
i − ω

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.4 Circuit-QED Control 185

The first term has been written in a way that evidences a shift of the resonator
frequency. The direction or sign of this shift depends on the state of the qubit σ z
and on the detuning. We will see in Section 7.4.2 how this shift can be used to
perform a nondemolition measurement of the qubit’s state. The dispersive coupling
also introduces a constant dc-Stark shift on the qubits and, when there are two or
more qubits, an exchange interaction σj+ σi− mediated by virtual transitions through
the resonator.15 These exchange terms can be used to generate entangled states,
simulate spin Hamiltonians, and implement two-qubit quantum gates, among other
applications.

7.4 Circuit-QED Control

Circuit-QED setups are the foundation on which many other experimental ideas are
built. Cavities, waveguides, and qubits are the main components of quantum optics,
quantum simulation, and quantum computing setups. It is therefore important to
pair the theory from earlier sections with the actual experiments to understand how
these setups are operated, and also as preparation for later chapters.

7.4.1 Direct Cavity Spectroscopy

Our theoretical treatment of the cavity-QED setup has put great emphasis on the
combined dynamics of the qubit and the resonator. However, the first evidence of
strong coupling between a qubit and a cavity was given by spectroscopy (Wallraff
et al., 2004), a technique that does not require measuring the qubit. Figure 7.9

Figure 7.9 Spectroscopy schemes for a qubit in a cavity. (a) Single-tone spec-
troscopy, consists in driving the cavity with a coherent state at a frequency ωd and
studying the transmitted and reflected light. (b) Two-tone spectroscopy consists in
probing the cavity transmission at a frequency ωp , while driving the qubit with a
different frequency ωd . The driving of the qubit may take place through a separate
antenna or – if enough power is available to compensate for the attenuation –
through the antenna that drives the cavity itself.

15 The mediating role of the cavity is evident when you analyze the second-order Schrieffer–Wolff correction
1 g 2 [S,V ] ∝ σ + â â † σ − which is responsible for these terms.
2 j i

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 186 Qubit–Photon Interaction

sketches the two most common setups for spectroscopy of a cavity–qubit system.
The simplest method, in Figure 7.9a, drives the resonator with a monochromatic
microwave at different frequencies ωd , studying how much power is transmitted
and reflected by the cavity.
The analysis of single-tone spectroscopy mimics the spectroscopy of linear res-
onators from Section 5.5.3. We first establish a relation between the output fields
that are reflected/transmitted by the cavity, and the operator of the cavity itself. In
the transmission case, all the signal is produced by the emission from the cavity,
√
b̂out  = −i κ â(t). We may compute this expectation value using our master
equation for the cavity-QED setup (7.34), with an added term (t)â † + (t)∗ â that
accounts for the probe drive (t) = 0 exp(−iωd t + iφ).
There is a simple trick that works when the strength of the drive is weak enough
that it can only excite transitions from the unique ground state – |g,0 in this case –
to a finite set of excitations that are disconnected from each other – in this setup,
the two polaritonic branches E1,± with states |1± . In this limit, we approximate
our system as a collection of harmonic oscillators, with the vacuum being the
ground state and one oscillator mode for each excited state at frequencies dictated
†
by the allowed transitions – |vac ∼ |g,0 and b̂± |vac ∼ |1±  the two polaritons,
and h̄ω± = E1± − E0 . Each of the oscillators is driven by the microwave with a
†
reduced amplitude (t)β± b̂± + H.c. and experiences also a reduced decay rate
κ± = |β± | κ, where β± ∼ 1± |â † |vac is the matrix element of the cavity operator.
2

Under these conditions, the output field becomes a linear superposition of the fields
absorbed and reemitted through the different allowed transitions:
 κs 1
T (ωd )0 eiφ = βs 0 eiφ . (7.50)
s=±
2 i(ω s − ω d ) + κ s /2

This model shows that a low-power spectroscopic signal will reveal two Lorentzian
resonances associated to the lower and upper polariton branches, centered on the
frequencies ω± = (E1± − E0 )/h̄. Each of these Lorentzian profiles will be very
similar to the ones in Figure 7.4, with the roles of R and T reversed if we are looking
at a transmission spectrum. The height of the resonance is the proportional to the
matrix element β± and is therefore dependent on how much of a photon component
|g,1 is present in each of the eigenstates. Conversely, the width of the Lorentzian
depends on the lifetime of these polaritons: The faster that these states decay due
to κ or γ , the broader the resonance.
In Figure 7.7, we plotted the transmission spectrum of a setup where the gap of
the qubit is tuned using the transverse field εσ x . When the qubit is far away from
the cavity resonance, only one of the two polariton branches has sufficient overlap
with the excited cavity mode |g,1. This mode allows the transmission of photons
around the cavity frequency ωd = ω, with a linewidth given by κ. When the qubit

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.4 Circuit-QED Control 187

approaches the cavity, we begin to see the vacuum Rabi splitting in action. Straight
on resonance, 2 + ε2 = ω2 , the two polariton resonances ωd = ω ± g split with a
gap 2g. In the strong coupling regime, the decay rate of the cavity and of the qubit
are smaller than this splitting, and both resonances can be resolved.
Low-power spectroscopy was also the fundamental tool in identifying the
ultrastrong coupling regime, both in the works by Niemczyk et al. (2010) and
Forn-Díaz et al. (2010). Note that even in the USC regime, the approximation of
independent resonances works, although the computation of the matrix elements
βn is complicated by the strongly hybridized and qubit–photon entanglement in
the low-energy excitations. In Figure 7.7b,c, we reproduce a transmission spectrum
from Niemczyk et al. (2010), illustrating a qubit that comes to resonance with the
third mode of a transmission line resonator. The coupling strength between the
qubit and the resonator is very large, g/ω  12%, thanks to a clever design where
the qubit is embedded in the central conductor of the resonator. As anticipated
in Section 7.3.5, the spectrum from Figure 7.7b is only fitted properly when we
consider the full Rabi model (7.32), with both rotating and counterrotating terms.

7.4.2 Qubit Dispersive Measurement

When we send a signal against a microwave resonator, it is reflected or transmitted
with an amplitude and a phase shift. Both quantities depend, among other things, on
the detuning between the driving field and the resonator. We saw this in Section 5.5.2
for a linear cavity, and it happens again when we have a qubit. The difference now
is that the features of this resonance will be affected by what is inside the cavity.
This idea inspired the quantum dispersive measurement of a superconducting
qubit, developed by Wallraff et al. (2005), to detect the state of a qubit without
destroying it. The technique assumes a qubit and a resonator in the dispersive
regime, far away from each other in frequency space. In this limit, the cavity
resonance experiencies a shift (7.49) that depends on the state of the qubit. More
precisely, as described by the model
h̄
Heff = h̄ ω + χ σiz â † â + ( + χ )σ z, (7.51)
2
the frequency of the resonator is now ω + χ σ z , with χ = g 2 ( − ω)−1 . If   ω,
the resonance is shifted upward when the qubit is in the excited state and downward
otherwise.
Let us assume that we have a microwave cavity sitting at the end of a waveguide,
such that it cannot leak energy elsewhere. This corresponds to a setup like the
one in Figure 7.9 where we eliminate the Tωd port, a configuration also studied in
Section 5.5.3. If we send a microwave drive at any frequency through that guide, we

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 188 Qubit–Photon Interaction

expect all the field to be reflected, possibly affected by a small delay that changes
the phase of the reflected field:
i(ω − ωd ) − κ/2
Rωd = ∼ eiφ(ωd ) . (7.52)
i(ω − ωd ) + κ/2
When we drive on resonance ωd = ω and there is no qubit, the reflected field expe-
riences a π shift and Rωd = −1. However, if there is a qubit, the reflected field on
resonance will produce a slightly different shift:
iσ z χ − κ/2 z
Rωd = ∼ ei(π+δφσ ), (7.53)
iσ z χ + κ/2
which depends on the qubit’s state σ z = ±1:
χσ z
tan(π + δφσ z )  . (7.54)
κ
Detecting this phase is a process that takes some time, during which we
integrate the quantum field reflected by the cavity, using the techniques from
Section 5.5.7. This measurement attempts to distinguish the coherent state created
by the microwave drive inside the cavity, and that depends on the state of the qubits
|α±  := |αei(π±δφ) . A proper discrimination of these states requires us to strike
several balances.

– The power of our coherent microwave probe must be low enough not to create
large numbers of photons inside the cavity – which would invalidate the dis-
persive approximation – but also large enough so that the two coherent states
are distinguishable. In other words, we want a small overlap αe+iδφ |αe−iδφ  ∝
e−|2αδφ| ∼ 0, while keeping |α|2  1.
2

– Similarly, we want the coupling to be large and the detuning to be small, in

order to maximize the dispersive shift χ, achieve a large phase shift, and enable
a strong distinguishability between states. However, χ must be small enough
that the corrections to perturbation theory may be ignored, and the dispersive
Hamiltonian (7.49) remains an accurate description.
– During the integration of the reflected field, we obtain an estimate of the coherent
state with an uncertainty that decreases with the integration time ∼ 1/T . This is
both because of the accuracy in the integration of the oscillating field, and also
because the cavity requires some time to reach the asymptotic states |α± .
– At the same time, we cannot make the measurement too long, because during
measurement the qubit is affected both by its intrinsic relaxation and by addi-
tional decoherence channels enabled by the coupling to the cavity, all of which
add up to γeff ∼ κ|g/( − ω)|2 + γ .

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.4 Circuit-QED Control 189

The dispersive measurement is a type of generalized measurement where we put

our the qubit into contact with a cavity, let them interact, and finally deduce the state
of the qubit from that of the cavity. Since Imα± = |α| sin(δφ)σ z , such discrimina-
tion can be performed by inspecting the sign of a quadrature P̂ = √12 (â † − â), to
deduce the sign of σ z = s. Ideally, this measurement has two outcomes s = ±1,
obtained with probabilities ps = tr(εs (ρ)), from a positive map ε that combines
the qubit–cavity interaction U with the projection  onto the right sign of the
quadrature P̂ :
 
εs (ρqubit ) = trcavity sign(P̂ )=−s U ρqubit ⊗ |vacvac| U † , with (7.55)
U  |gg| ⊗ D(α− ) + |ee| ⊗ D(α+ ).
In this limited framework, the measurement is specified by two probabilities:
Mp = tr(−1 |α− α− |) = tr(+1 |α+ α+ |), (7.56)
Mn = tr(+1 |α− α− |) = tr(−1 |α+ α+ |) = (1 − Mp ).
The smallest probability Mn measures how much of the |α−  coherent state lies in
the region P̂ > 0, and vice versa, how much of the |α+  lies below P̂ < 0. It quan-
tifies the mistake in the discrimination of σ z , so that the combined measurement
reads
ε−1 (ρ) = (1 − Mn )tr(ρqubit |gg|) + Mn tr(ρqubit |ee|), (7.57)
ε+1 (ρ) = Mn tr(ρqubit |gg|) + (1 − Mn )tr(ρqubit |ee|).
From this derivation, we see that in the limit of good discrimination, Mn  1,
the measurement is close to an ideal projective measurement of the qubit on the
basis {|g , |e}. This is a type of quantum nondemolition or QND measurement that
can be repeated multiple times, obtaining almost always the same value. It is also
the type of measurement that we want to have in quantum computers, both for the
application of quantum algorithms as well as for error correction.
The fidelity of the dispersive QND measurement is limited by the discrimina-
tion error Mn and by the decoherence of the qubit – which was not considered
in the previous reasoning, and which modifies the maps ε±1 (ρ). We can improve
both errors, allowing for longer measurements with lower decoherence, by adding
Purcell filters to the measurement device (Reed et al., 2010). The Purcell filter is
an additional cavity that sits between the qubit’s cavity and the waveguide, further
restricting the photons that can leak out of the original setup. The filter must be
broader than χ so that it can discriminate the two displaced frequencies, but it must
be narrow enough so that the qubit’s frequency  falls outside the filter’s linewidth.
This way, the effective decay rate of the qubit is lowered from the value κ|g/δ|2
predicted by (7.31), to a much smaller number. At the same time, the setup allows

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 190 Qubit–Photon Interaction

for stronger drives, faster measurements, and reduced errors, which as of the time
of writing this book lay around 1% (2%) in 88 ns (48 ns) for state-of-the-art setups
(Walter et al., 2017).

7.4.3 Two-Tone Spectroscopy

Having access to the qubit’s instantaneous state of the qubit is a powerful tool that
facilitates a new type of spectroscopy, where we can probe the qubit’s response to
external fields, calibrating its Hamiltonian independently of that of the cavity.
Sketched in Figure 7.7b, two-tone spectroscopy uses two driving fields, one
that will excite the qubit ωd , and a probe drive that is resonant with the bare
cavity frequency ωp = ω. The transmission of the probe beam is continuously
monitored, measuring the phase shift experienced by the photons. Since this is a
continuous process, where we are at the same time directly influencing the qubit,
the study of the phase shift only provides an average of the cavity shift and, by
extension, of σ z .
As in usual spectroscopy, this procedure is repeated for a broad range of qubit
drives ωd . When ωd is far from resonance, the qubit will remain in its approximate
ground state and the phase shift will have a fixed, nonzero value. However, when
ωd approaches the renormalized frequency of the qubit,  + g 2 /( − ω), the qubit
will begin to oscillate. In this case, σ z  will average to zero and the photons will
not experience any phase shift.
In Figure 7.10, we show an excerpt of the results by Schuster et al. (2005).
Note how the method reveals the qubit’s hyperbola with great clarity and contrast.
The two-tone spectroscopy is complementary to the single-tone spectroscopy from
Section 7.4.1. In the latter method, we only get a good signal from the qubit when it
is strongly hybridized with the cavity. In contrast, the two-tone spectroscopy works
best when the qubit is in the dispersive regime and not hybridized at all.

7.4.4 Single-Photon Generation

Once we have a superconducting qubit inside a microwave resonator, we can apply
different microwave and qubit controls to engineer quantum states of the qubit, the
cavity, or both. The simplest application of such protocols is to engineer a one-
photon state in the cavity. For this, the simplest method is to start with a qubit gap
 strongly detuned from the resonance frequency ω of the cavity where the qubit
resides. The energy separation should be large enough |ω − |  |g| that we can
neglect the qubit–cavity coupling. We assume that the qubit starts in this situation
and in the |g state. The qubit is then driven resonantly to acquire one excitation |e,
and we rapidly shift its frequency to bring qubit and cavity into resonance  = ω for

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.4 Circuit-QED Control 191

Figure 7.10 Two-tone spectroscopy of a charge qubit. (b) shows the phase shift
experienced by the transmitted photons as the qubit energy is moved around using
an external voltage and driven with some frequency, νd . As shown in (a), when the
qubit is appropriately excited, the average phase drops significantly to zero. Figure
reproduced from Schuster et al. (2005).

half a Rabi period. This way, the qubit and the cavity exchange excitations, evolving
from |e,0 to |g,1. If at this point we rapidly bring the qubit out of resonance again,
we will have created a single-photon state.
As discussed in Section 5.4.6, this idea was demonstrated by Eichler et al. (2011)
in a seminal experiment that not only showed the generation of the photons, but also
demonstrated the nature of the quantum state by doing Wigner function tomography
of photon’s wavefunction as it left the cavity.

7.4.5 Qubit Reset

Many experimental protocols require that we have qubits in well-defined |g or
|0 states. This requisite may be hard to obtain experimentally. First, if we rely on
cooling, the quality of the states we prepare is limited by the temperature of the
experiment, which may not be enough if the qubit’s frequency is modest. Second,
cooling may take too long a time, especially when we strive to prepare qubits with
long coherence times T1 of hundreds of microseconds or even longer.
Since most experiments place qubits close to different types of cavities, including
the ones for readout, we can use those cavities to enhance the qubit’s relaxation and
prepare almost perfect zero states. The protocol begins with the qubit in a situation
of low gap detuning  < ω − |g|. The gap of the qubit is slowly changed until
it lands above the resonator’s frequency  > ω + |g|. If this shift is done at an
adiabatic pace d/dt  |g × |, the qubit will transfer its quantum state to the

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 192 Qubit–Photon Interaction

resonator in a coherent process (α |g + β |e) ⊗ |vac → |g ⊗ (α + βa † ) |vac .

Once the transfer is complete, we wait a time O(1/κ) for the cavity to relax and
emit any excitations, so that the whole system is projected onto the |0 ⊗ |vac state.
The way to analyze the qubit–cavity state transfer is to realize that we have two
separate subspaces. The state |g,0 does not evolve at all, and it does not matter
whether we write α |g ⊗ |0 or |g ⊗ α |0. The single-excitation subspace, on the
other hand, implements a Landau–Zener process, as described in Section 9.1.1. In
this process, we start from a situation where the combined state |e,0 is an approxi-
mate eigenstate of the qubit–cavity model, which is adiabatically deformed into the
final eigenstate |g,1 at the end of the process.
Interestingly, this strategy may be even used to suppress population that has
leaked into higher excited states, as demonstrated by McEwen et al. (2021). If the
dynamics is implemented in a time longer than a Rabi period, the reset quality over
all energy levels can be very high O(10−4 ), well below what can be experimentally
detected.

7.4.6 Cavity Fock States Superpositions

Using our controls on the qubits, the cavities, and the qubit–cavity coupling, we
can engineer very sophisticated protocols to create arbitrary states of either system,
or both. We have already described how to inject a single photon into the cavity.
As explained by Law and Eberly (1996), we can generalize this idea to create
superpositions of N Fock states in the cavity:

N
|N  = cn |g,n . (7.58)
n=0

The idea is technically complex, although it can be understood qualitatively. One

just needs to realize that, given a state with N photons

N−1
|χN  = N
αe,n−1 |e,n + αg,n
N
|g,n , (7.59)
n=0

we can create a state |χN−1  that has one less excitation. For this, we use two
operations. The first one, Qn , consists in bringing the qubit into resonance with
the cavity  = ω and wait a time tN sufficient for all population from |g,N  to
be transferred to |e,N − 1. The second operation is a rotation of the qubit CN ,
transforming |e,n ↔ |g,n − 1, which converts |e,N − 1 into |g,N − 1 and
leaves us with a |χN−1  state.
Using this technique, we can compute a sequence of gates that undoes the desired
state C0 Q0 · · · CN QN |χN  = |0. If we apply the reversed sequence C0†, Q†0 . . . Q†N ,

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 7.4 Circuit-QED Control 193

we will be able to construct the Fock superposition state experimentally, leaving the
qubit unentangled from the cavity. This technique was demostrated by Hofheinz
et al. (2009) with up to N = 5 photons, in a very interesting work that not only
created the state but used Wigner function tomography to reconstruct the complete
density matrix of the cavity.

7.4.7 Cavity Schrödinger Cats

A much simpler protocol can be used to engineer the Schrödinger cat states from
Section 5.4.4, using the dispersive coupling regime (7.51). For simplicity, we will
study the dynamics of the system in a rotating frame where we eliminate the qubit’s
and cavity’s frequencies, working with the Hamiltonian

H = χ σ z a † a. (7.60)

Assume we prepare the cavity in a coherent state |α and simultaneously excite
the qubit into a superposition √12 (|g + |e). After a time t, the combined state will
have evolved into
1 1
√ e−i(−χ)t |g |e−iχ t α + √ e−i(+χ )t |e |e+iχ t α . (7.61)
2 2
We will choose χ t = π/2, so that both coherent states have opposite signs, and
further displace the cavity along −iα, obtaining the state

1 1
√ e−i(−χ)t |g |−2iα + √ e−i(+χ )t |e |0 . (7.62)
2 2
Interestingly, we can now use the dispersive shift of the qubit’s frequency to
implement a rotation that only is effective in the vacuum state. In particular, we drive
the qubit with a long pulse, resonant with the frequency  + χ , which only excites
the transition |g,0 ↔ |e,0, leaving all |g,n > 0 and |e,n > 0 approximately
unaltered. The outcome of this rotation is a new state where the qubit is decoupled
from the cavity:
1
|g ⊗ √ (|−2iα + |0) . (7.63)
2
Using further displacements, we can bring this into a more symmetric superposition
|−iα + |iα. Note that we can also change the relative phase of the coherent states
by playing with the single-qubit rotation’s phase. Moreover, due to free evolution
with the resonator’s Hamiltonian, the states ±iα rotate in phase space so that we
effectively can prepare superpositions along different directions.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 194 Qubit–Photon Interaction

This type of coherent state superpositions has been demonstrated by Vlastakis

et al. (2013) with coherent states that involve up to |α|2 ∼ 100 photons. Such super-
positions constitute the basis for very interesting encodings of quantum information
and even error-corrected quantum computing protocols (Mirrahimi et al., 2014),
which unfortunately we do not have space to discuss here.

Exercises
7.1 Show how (7.13) can be obtained by projecting both sides of the Schrödinger
equation i∂t |ψ(t) = Ĥ |ψ(t) onto the basis states |1,vac and âk† |0,vac.
7.2 Markovian kernel. Let’s have a look at the qubit’s dynamics:
 t
i
∂t c1 = −ic1 − J (ω)e−iω(t−τ ) c1 (τ )dω dτ, (7.64)
2π 0
assuming that the spectral function is constant J (ω) = 2πα over the whole
real line, ω ∈ R. Prove that the differential equation is exactly Markovian and
only depends on the value of c1 (t). What are the decay rate γ and the Lamb
shift?
7.3 Qubit–photon entanglement. We can use (7.24) to estimate the amount of
entanglement between the qubit and the emitted photon. The von Neumann
entropy of the reduced density matrix ρqb

S = −tr(ρqb log(ρqb )) = − λi log(λi ) (7.65)
i

is a function of the eigenvalues λi of the density matrix, and it is also a quan-

tifier of the entanglement between our qubit and the photonic environment.
Use the Pauli expansions and diagonalization methods from (2.21) to compute
λ± . Show that this entanglement is zero at t = 0 and t = +∞ and find its
maximum.
7.4 Zero-temperature master equation. Assume a qubit in a generic mixed state,
placed in an empty transmission line, ρ = ρqb ⊗ |vac vac|.
(1) Show that ρ can be written as a |1 1 | + b |2 2 |, where
|i  = |ψi  ⊗ |vac and |ψi  are the eigenstates of ρqb .
(2) Show that we can study the evolution of |1  and |2  separately using
the formalism from Section 7.2.
(3) Show that the total density matrix obeys the relaxation equation (6.19) and
that this expression is a solution of (6.17) with n̄() = 0.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 Exercises 195

7.5 A chiral interaction is one that breaks time-reversal symmetry. In our context
of light–matter interactions, a qubit interacting with a chiral medium would
couple differently to photons moving along different directions g+k = g−k .
Let us assume that our waveguide is chiral because the k < 0 modes are totally
absent. Take the following steps to understand how this changes light–matter
interactions:
(1) Assuming gk ∼ gdk 1/2 , integrate the qubit dynamics and compute the rela-
tion between g and the spontaneous emission rate, γ .
(2) Write the new input–output relations: How many fields do you have to
consider? What is the new prefactor in front of c1 (t)?
(3) Repeat the derivation of the scattering coefficients for an incoming photon.
Show that Rω = 0, |Tω | = 1 and that the photon only experiences a phase
shift.
(4) Explain when the chiral model can be used to study the scattering of
photons by a qubit that is placed at the end of a semi-infinite transmis-
sion line.
7.6 Consider a multilevel circuit with N excited levels, interacting with the trans-
mission line. We assume that each level only couples to the ground state via
the propagating photons:


N   
H = i |ii| + ωk âk† âk + βi σi+ gk âk + gk∗ â † . (7.66)
i=1 k k

Here σi+ = |i 0| and the βi measure the relative strength of the different
0 ↔ i transitions.
(1) Derive the evolution equations for the single-excitation ansatz:


N 
|ψ = ci σi+ + ψk âk† |0,vac . (7.67)
i=1 k

(2) Integrate the photons out and apply the Markovian approximation to obtain
N equations for the ci coefficients. How does the excited state |i evolve
in time?
(3) Generalize the input–output relations to consider the N transitions.
(4) Compute the scattering amplitudes Tk and Rk for a single incoming pho-
ton. What does the spectrum look like?

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 196 Qubit–Photon Interaction

7.7 Lambda system and dark states. Let us study the following energy-level
structure, with two degenerate ground states, |a and |b, and one excited state:

What can you say about the spontaneous emission from state |1 and the inter-
action with propagating photons? Hint: Show that this problem is equivalent
to a two-level system plus one noninteracting dark state for any (complex)
values of the two coupling strengths, ga,bk .
7.8 Photodetection. Let us now devise a simple photodetector using the a few-
level system that can switch the state through the absorption of a photon.
A two-level system is not a valid photodetector because even if the photon
is absorbed, the state |1 will subsequently decay. Instead, we introduce the
following three level system:

In this scheme, an incoming photon couples to the qubit and excites the tran-
sition 0 ↔ 1. At the same time, there is an incoherent transition 1 → b that
depletes the excited state population with rate  and deposits it irreversibly
onto the state |b, thus signaling the presence of a photon. We are going to
model this system with the same equations as for the two-level system. The
only change is the addition of an ad hoc incoherent channel:
i∂t c1 = ( − iγ )c1 − ic1 . (7.68)
(1) Assuming an adiabatically switched on wavepacket, compute the proba-
bility of landing onto the state |b
Pb = 1 − |c1 |2, (7.69)
by integrating the dynamics of the qubit under this driving. Show that it is
always below 50%.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 Exercises 197

(2) What are the transmission and reflection amplitudes of the photon, Tω and
Rω ? How does the photodetection process modify the probabilities |T |2
and |R|2 ?
7.9 Given (7.42), prove that the probability of having N excitations, given by
P (N,t) = tr(ρN (t)) decays exponentially. More precisely, prove the equation
d
dt
P (N) ≤ −(κ(N − 1) + γ )P (N ).
7.10 Prove the expressions for the dispersive measurement (7.57) in the ideal case
where decoherence can be ignored.

https://doi.org/10.1017/9781316779460.007 Published online by Cambridge University Press

 8
Quantum Computing

8.1 Quantum Circuit Model

The term quantum computer originated in the 1980s, although with different mean-
ings and perspectives. At this time, computers were on their way to becoming a
commodity and computer science was no longer an emergent field, but there was
already a growing concern about the energy requirements and the efficiency of those
first computers. Not long before, Landauer had connected information processing
and thermodynamics, establishing a lower bound to the energy that is consumed
whenever a bit of information is created or discarded: kB T log 2  0.017eV.
Computers of that era were very inefficient, and scientists were looking for alter-
native computation models that could attain Landauer’s limit. Some focused on
the hardware, looking at computer designs that use superconducting circuits or
nonlinear optical elements, two paradigms that have been ultimately abandoned.
Other, more theoretically minded researchers worked on reversible computation.
These researchers looked at the two-bit gates that are used in all computers – NAND,
XOR, etc. – and realized that each gate drops a bit of information whenever it is
executed. Connecting this idea with Landauer’s principle, they argued that we could
look at reversible operations, such as the Toffoli gate in combination with the NOT
gate, as a universal framework for computation that would be more energy efficient.
You may already realize that quantum mechanics was a natural candidate for
reversible computation. After all, it is a physical framework built on the premise of
unitary and thus reversible time evolution. In 1980, physicist Paul Benioff formal-
ized this notion, establishing a connection between the Turing machine, a universal
model of computation, and the evolution of an idealized quantum physical system
(Benioff, 1980). In other words, he found a quantum Hamiltonian that implements
the Turing machine.

198

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.1 Quantum Circuit Model 199

Around the same years, we also find celebrated physicist Richard Feynman
looking at the relation between computers and quantum mechanics from a different
angle. Coming from a theoretical physics background, and having worked on
the application of computers to solving theoretical physics problems, he chal-
lenged the capabilities of classical computers (both real and abstract, such as the
Turing machine) to simulate the evolution of quantum systems. In a well-known
talk, Feynman (1982), he illustrated the exponential complexity of simulating the
Hilbert space of even very simple quantum models – particles jumping on a lattice –
and introduced the idea of quantum simulation, engineering a quantum physical
device to reproduce the dynamics of a complex quantum model that we do not
know how to solve.
Benioff’s work and Feynman’s ideas converge in the work by Deutsch (1985),
a seminal paper that introduced the concept of a quantum Turing machine. This
is a device that not only can process classical information – and therefore repro-
duce a classical Turing machine – but also information stored in generic quantum
states, including quantum superpositions and entangled states. Deutsch’s quantum
Turing machine provided the theoretical foundation for what we now understand as
a universal quantum computer.
Despite its theoretical value, both the classical and the quantum Turing machines
are inconvenient designs that do not inform the real architecture of computers.
Instead, quantum computers are developed in the lab using more scalable and highly
parallelizable frameworks. In this chapter, we focus on the particular framework of
quantum circuits introduced also by Deutsch (1989). The quantum circuit model
assumes a quantum register formed by near-ideal qubits. The register acts both as a
memory and as a processor.1 Quantum computations are implemented by modifying
this register through a composition of qubit measurements with unitary operations
that are extracted from a universal set.
The quantum circuit model includes also a very elegant graphical representation.
The algorithm’s operations and measurements are placed on a set of horizontal lines,
similar to a musical “pentagram.” Each horizontal line addresses one qubit. A box
in this pentagram is a unitary operation acting on the qubits that it is connected to.
The complete circuit can be read as a set of gates and measurements to be executed,
either in parallel or sequentially, as we read the diagram from left to right.
As an example, Figure 8.1a shows a quantum algorithm to recreate and mea-
sure a Schrödinger cat or Greenberger–Horne–Zeilinger (GHZ) state √12 |00000 +
√1
2
|1111. In this example, we begin with a set of qubits initialized to a zero state
|00000. The first operation is a single-qubit gate, a Hadamard gate H (cf. Table 8.1)

1 In some sense, one could argue that such a design is closer to an abacus than to a Turing machine (a head
moving over an infinite, 1D stream of data) or a von Neumann computer (with CPU, control unit, memory,
external storage, and I/O).

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 200 Quantum Computing

Figure 8.1 (a) Quantum circuits are schematics of quantum algorithms and proto-
cols that combine (i) qubits, represented by circles with horizontal lines; (ii) single-
qubit operations that intersect only one such lines; (iii) multiqubit operations;
and (iv) measurements. (b) Single-qubit operations, as denoted in Table 8.1.
(c) Two-qubit conditional operations for a CZ and CNOT, respectively. Quantum
measurements for feedback operations (d) can be moved to the end of an algorithm,
using two-qubit gates (e).

that sets the first qubit into a superposition state, √12 (|0+|1) ⊗ |0000. Subsequent
two-qubit operations, called CNOT’s, copy the state of the first qubit to all other
qubits, creating a large entangled GHZ state.
The power of the quantum circuit model does not lay in this nice, prescriptive,
and visually appealing representation. The circuit model gives us a minimal set of
logical ingredients that, combined together, can reproduce any classical or quantum
computation we may think of. This includes Shor’s celebrated algorithm for fac-
toring, the phase and amplitude estimation algorithm, the Harrow–Hassidim–Lloyd
(HHL) algorithm for inverting matrices, and many modern applications to real-life
problems in chemistry, finance, logistics, etc.
DiVincenzo (2000) formalized the requirements of the quantum circuit model in a
celebrated paper. The DiVincenzo criteria are a set of five requirements for scalable
quantum computation plus two for quantum communication. Cited verbatim, these
elements are the following:
(1) A scalable physical system with well-characterized qubits
(2) The ability to initialize the state of the qubits to a simple fiducial state, such
as |000 . . .

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.1 Quantum Circuit Model 201

Table 8.1. Most popular single-qubit and two-qubit unitaries, with their
names and their generators – i.e., Hermitian operator O such that
U = exp(iO + iφ) with some irrelevant global phase φ.
   
Hadamard 1 1 Pauli-X 0 1
π y H = √1 π x X=
2σ 2 1 −1 2σ 1 0
   
Pauli-Y 0 −i Pauli-Z 1 0
Y = π z Z=
πσy i 0 2σ 0 −1
   
Phase 1 0 “π /8” 1 0
π z S= π z T =
4σ 0 i 8σ 0 eiπ /4
⎛ ⎞ ⎛ ⎞
Control-NOT 1 0 0 0 Control-Z 1 0 0 0
⎜0 1 0 0⎟ ⎜0 1 0 0 ⎟
(CNOT)
z
⎝0 0 0 1⎠
(CZ)
z z
⎝0 0 1 0 ⎠
2 (σ1 + 1)σ2 4 (σ1 −1)(σ2 −1)
π x π
0 0 1 0 0 0 0 −1
⎛ ⎞ ⎛ ⎞
1 0 0 0 1 0 0 0
√
iSWAP ⎜0 0 i 0⎟ SWAP ⎜0 1+i 1−i
0⎟
⎝0 i ⎜ 2 2 ⎟
+ −
4 (σ1 σ2 + H.c.)
3π 0 0⎠ π
8 σ1 · σ2 ⎝
0 1−i 1+i
0
⎠
0 0 0 1 2 2
0 0 0 1

(3) Long relevant decoherence times, much longer than the gate operation time
(4) A “universal” set of quantum gates
(5) A qubit-specific measurement capability
(6) The ability to interconvert stationary and flying qubits
(7) The ability to faithfully transmit flying qubits between specified locations
These criteria provide an “engineering checklist” that applies to all platforms –
trapped ions, superconducting circuits, quantum dots, etc. If all first five criteria are
met, and if the decoherence times lay above a certain threshold, we can provably
build arbitrarily large quantum computers, capable of any quantum computation
we can think of, using error correction and fault-tolerant principles to make those
computations with arbitrary precision.
In this chapter, we will go through the elements in DiVincenzo’s checklist. We
will use the tools and knowledge gathered in Chapters 6 and 7, and complete them
with additional techniques for implementing gates, architectural decisions, char-
acterizing coherence and fidelity, etc. The roadmap is as follows. In Section 8.2,
we introduce the most common forms of quantum registers, how they are mea-
sured and reset, thus covering checklist items (1), (2) and (5). Section 8.3 dis-
cusses item (4) and how arbitrary computations can be approximated using a small
set of quantum operations, a universal gate set. We show that those gates include

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 202 Quantum Computing

single-qubit operations – already covered in Chapter 6 – and at least one universal

two-qubit operation, of which we offer various designs. Section 8.4 discusses the
characterization of errors in the quantum computer, with an improved version of
item (3). Section 8.5 provides some insight on how having low error rates allows us
to implement error-correction strategies, opening the door to arbitrarily large and
arbitrarily complex quantum computations.

8.2 Quantum Registers

We introduced in Section 6.1.1 the idea of a physical qubit as a real system with two
long-lived quantum states that can exist in an arbitrary superposition of those states.
A quantum register is a collection of qubits – physical if we refer to an experiment,
or logical if we are focused on the algorithmic representation – that can be subject
to the unitary operations and measurements of a quantum circuit model.
A physical implementation of a quantum register is conditioned by various exper-
imental choices:

– The choice of physical system (i.e., circuit) that implements a qubit

– How those qubits are locally addressed and measured
– The quantum operations that we can implement on those qubits
– The topology of those qubits in the experimental setup

The choice of qubit is particularly important, because it conditions all other

options in the computer design. One could, for instance, think about using the flux
qubit, a device with strong anharmonicities, the potential for very strong interactions
and complex qubit–qubit interaction geometries (see Section 9.4.1). However, at
the moment this book is being written, all superconducting quantum computers are
based on the transmon qubit or variations thereof.
The transmon has three nice properties: It has long coherence times, it is a very
reproducible qubit, and it has a flexible design that adapts very well for implement-
ing large quantum registers of 50 and more qubits (Arute et al., 2019), with varying
topologies and designs. Figure 8.2a illustrates a quantum computer with nine X-mon
qubits, a variation of the transmon, so-called because of their shape. Notice that the
qubits are never alone: they are accompanied by lines, microwave resonators, and
shielding ground planes, which are needed to protect, manipulate, and interrogate
the register. Where to place those elements is an important architectural decision,
which these days is coming closer to an industrial secret.

8.2.1 Measurements
The election of the transmon qubit naturally imposes one type of measurement
scheme: We must adopt the cavity-mediated, QND measurement that was presented

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.2 Quantum Registers 203

(a)

(b) (c)

Figure 8.2 (a) Superconducting qubit quantum register built from nine X-mon
qubits, a variant of the transmon qubit with a cross shape. The qubits are close
to each other for nearest-neighbor interactions; they are connected to individual
cavities (wiggly lines on the top) for dispersive qubit readout, and they are tuned by
small flux lines coming from the bottom of the picture. Modified from Barends et al.
(2014) with permissions. (b) Alternative qubit topology, resembling the work by
Arute et al. (2019), with similar qubits, but higher connectivity, and using auxiliary
qubits to mediate interactions. (c) Three-dimensional sandwich packing of two
chips, joined by superconducting indium bonds. One chip contains resonators and
readout channels structures, while the other one packs qubits.

in Section 7.4.2. As shown in Figure 8.2a, each qubit is attached to a readout res-
onator, which is connected to the outer world. The microwave cavity acts as a
narrowband filter that isolates the qubit from the readout cables. However, when
we send a resonant microwave pulse through those cables, the cavity reflects the
light with a phase that carries information about the state of the qubit. This is a
good, nondestructive POVM measurement of the qubit, with a small uncertainty of
a few percent, ∼1–3% in current experiments.
If all the resonators that talk to the qubits have slightly different frequencies,
we can multiplex the readout phase. Instead of sending a monochromatic pulse
that is resonant with one cavity, we send a broadband signal with the frequencies
of all resonators. Analyzing the reflected signal allows us to gather information
about the state of all qubits that were addressed by the probe signal. The readout of

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 204 Quantum Computing

this broadband signal requires the use of sophisticated, broadband, and quantum-
limited amplifiers, such as a TWPA (see Section 5.5.6), that can boost the signal
of all cavities at the same time. As shown by Arute et al. (2019), such multiplexed
measurement schemes can scale up to 53 qubits with significant effort. Interestingly,
the same idea that is used to implement local addressing in the measurements can
be used to scale up and parallelize local operations on the qubits.

8.2.2 Qubit Reset

One could expect that, simply because the quantum register is at a low temperature,
it will thermalize to the ground state – which for a transmon state is just the tensor
product |000 . . ., the fiducial state required by DiVincenzo’s list. Following the
calculations after Section 4.1.1, and assuming a qubit frequency around 6 GHz, the
excitation probability for a fridge between 10–50 mK will range from 3 × 10−11 %
up to 0.3%. These are very good figures, but in order to reach them one needs to
wait longer than decay time T1 of the qubit. For highly scalable computers with
high-quality qubits T1 ∼ 100 μs, that represents an important waste of time.
Actual superconducting quantum computers rely on active ground state prepara-
tion to reset the qubit state to zero. One simple alternative is to measure the state
of the qubit. If the qubit is in the desired state |0, one does nothing; otherwise, we
apply a π pulse to flip the state from |1 to |0. This scheme does not require any
additional tools, but the fidelity of the resulting state is limited by the quality of our
measurement, which in existing setups is not very good.
Other methods rely on the readout cavities, using them to shorten the lifetime
of the qubit T1 . By bringing the qubit in resonance with the cavity, we enable the
transfer of excitations from the |1 state to the resonator, where they may decay into
the readout cables, accelerating the reset time. This is the mechanism used by Arute
et al. (2019) in their 53-qubit quantum processor, where they achieved a reset time
of 10 μs, consistent with the resonator bandwidth. There are other alternatives that
use microwave drives to enable the exchange of excitations between the transmon
and the readout cavity, even in situations when they are off-resonance (Egger et al.,
2018). This type of method achieves good preparation fidelities, around 1% in times
below a microsecond (Egger et al., 2018; Magnard et al., 2018).

8.2.3 Architectural Decisions

Even after we have the qubit, readout, and control mechanisms, there are many
important choices regarding the physical layout and organization of the quantum
register. A very important one concerns the layout of the circuit – how qubits,

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.2 Quantum Registers 205

resonators, and control lines are fabricated onto the chip. This layout will inform the
topology of the circuit, establishing which qubits can interact with others, directly
or through mediating circuits.
Take for instance the one-dimensional layout from Figure 8.2a. Qubits are placed
on a row and are in close proximity to each other. Each qubit has two nearest
neighbors, which are capacitively coupled to it. Resonators are moved away to
a separate row and are collectively addressed, and there is a third row of direct
controls on the qubits. This arrangement is clean, provides direct access to all qubits,
and minimizes the number of qubits that see each other, which is good to lower
cross-talk.
Unfortunately, the one-dimensional design does not scale well in applications.
Quantum algorithms require interactions between arbitrary pairs of qubits. These
interactions can be decomposed into other gates, but the decomposition becomes
very costly in one-dimensional topologies. For instance, we can make a quantum
operation between qubits 1 and 9 only by inserting gates between 1 and 2, 2 and 3,
and so on. If we follow this approach, it becomes clear that we are constrained by
the taxi-driver metric2 of how far two arbitrary qubits lie in a given lattice.
There are also other considerations. For instance, quantum error correcting archi-
tectures have been designed with certain lattice topologies in mind. Figure 8.2b
illustrates an arrangement of physical qubits on a square lattice, where each qubit
has four neighbors. This topology is required for error correcting algorithms based
on the surface code, as we explain in Section 8.5.4. However, moving from a simple
planar architecture (cf. Figure 8.2a) to a dense lattice (cf. Figure 8.2b) raises the
question of where to place all other components – resonators, control lines – that do
not fit in the space between qubits.
The solution to this problem comes from an early work by O’Brien et al. (2017),
which looked at how to shield qubits and protect them from decoherence. This
work introduced a sandwich architecture, where two chips are bonded together
using tiny indium contacts. One chip contains the qubits, the other one acts as a
shield against electromagnetic fields, providing better coherence. Since Rigetti’s
work, there has been steady progress in using the same sandwich architecture, but
transferring resonators and control lines to the top chip.
The idea is sketched in Figure 8.2c, using IBM’s patent as inspiration (Brink
et al., 2019). The qubit is fabricated on one slab, together with ground planes and
control lines – which also need some 3D fabrication, because they would not fit
among qubits. Resonators and readout lines are constructed on a different chip. Both
chips are attached mechanically to each other, and contacted through a small “ball”

2 The taxi-driver metric evaluates the distance between two nodes in a lattice by counting the smallest number of
edges that make a path connecting both nodes.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 206 Quantum Computing

of indium. This superconducting ball deforms and sticks to the aluminum, estab-
lishing an electrical bridge. In this sketch, the indium bond assists in creating a
common ground plane for both chips, but it is thick enough that keeps both layers
well separated – creating what is known as an air gap.

8.3 Gate Toolbox

In the quantum circuit model, a computation is decomposed into three phases: reg-
ister preparation, unitary evolution, and measurement. We have seen how the first
and last step are done. We will now explain how the most complicated phase, an
arbitrary unitary evolution, is decomposed into smaller, manageable operations that
can be implemented in the lab.

8.3.1 Universal Set of Gates

A general quantum computation may require the implementation of an arbitrary
unitary operation in the Hilbert space of the quantum register. Just from the shape
of the unitary matrix itself, we appreciate the exponential number of parameters
that inform the operation – 22N complex amplitudes for N qubits! These arbitrary
numbers are produced by the quantum computing expert almost at random, drawing
from the set of algorithms, without connection to any experimental protocol, and
with no idea of how to implement them in the laboratory.
Fortunately, the circuit model builds on a solid mathematical foundation, which
takes care precisely of this question: the decomposition of arbitrary computations
into elemental gates. This is made possible by the notion of a universal set of quan-
tum gates. We call by this name any finite set of unitary operations {U1,U2, . . . ,UM }
with which we can approximate global unitary W transforming the N -qubit register
+ ε
space to arbitrary precision W  K n=1 Uσn + O(ε).
Fortunately for us, there exists a very powerful theorem that provides two
instances of a universal gate set. This theorem states that any unitary transformation
W acting N qubits can be approximated to arbitrary accuracy using a universal set
of quantum gates made from the following:
(1) Arbitrary unitary rotations of individual qubits
(2) One universal two-qubit
√ gate acting on any pair of qubits, such as the CNOT,
CZ, iSWAP, and SWAP from Table 8.1
The procedure, which can be found in the book by Nielsen and Chuang (2011), may
be further simplified, replacing the first step with

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.3 Gate Toolbox 207

(1) Hadamard gates and π/8 gates acting on individual qubits

Sometimes, these decompositions are easy to find, with patience and a bit of
practice. Take for instance the algorithm in Figure 8.1a for constructing a five-qubit
gigahertz state √12 (|00000 + |11111), or look at the well-known algorithms for a
quantum Fourier transform, phase estimation, factoring, etc. More often, however,
if we try to apply these ideas to general problems, such as simulating the evolution
of a quantum material or a molecule, we may not have a practical decomposition
to use. This may easily happen because we cannot explicitly compute the matrix
elements of the unitary to simulate, or because the decompositions we find require
an exponentially growing number of operations.
With these caveats in mind – which simply highlight the need for more research
in the theory of quantum algorithms – the fact is that we have decomposed DiVin-
cenzo’s fourth requirement into two very well-defined tasks: implementing single-
qubit operations and finding a protocol to create a two-qubit universal gate. We will
now explore the theoretical and experimental solutions to these problems.
We do not need to spend too much time discussing the design of local operations.
This matter was largely addressed in Chapter 6. In Section 6.1.4, we introduced the
continuum of single-qubit transformations in the Bloch sphere (6.8), and how it can
be implemented using an external microwave drive (6.6) with tunable amplitude,
frequency, and phase. Varying these parameters, and adjusting the duration of the
pulse, we showed how to implement the Z, S, T, and H gates from Table 8.1, and
provided evidence that any other local gate can also be implemented.
Single-qubit gates are the fastest and most precise operations that are performed
in superconducting quantum circuits, with fidelities around 0.09% in state-of-the-
art experiments (Arute et al., 2019) – which are impressive, but are still above the
2 × 10−5 fidelities achieved in trapped ions by Brown et al. (2011).
Two-qubit gates, on the other hand, are not so simple and also not so precise.
There are many choices of universal two-qubit gates, some of which are shown in
Table 8.1. We will discuss three designs, which are based on different ways to make
the qubits interact, or to control their dynamics.

8.3.2 Two-Qubit Exchange Gates (iSWAP)

In this section, we discuss the two-qubit gates generated by off-diagonal dipo-
lar interactions, such as those found in capacitively coupled transmons (see
Section 6.5.2). We start from the Ising model with transverse coupling, shown
in (6.67). When the interaction term is usually much weaker than the qubit gap,

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 208 Quantum Computing

|J |  |i |, an RWA-type argument allows us to replace this Hamiltonian with a

simple exchange model (see Problem 6.16):
 1 z 2 z
Heff = σ + σ + J (σ1+ σ2− + σ1− σ2+ ) (8.1)
2 1 2 2
⎛ 1 +2 ⎞
2
0 0 0
⎜ 0 1 −2
0 ⎟
⎜ J ⎟
=⎜ 2
1 −2 ⎟.
⎝ 0 J − 2 0 ⎠
0 0 0 − 1 +
2
2

The two eigenstates, |0,0 and |1,1, are preserved throughout evolution, accu-
mulating the same local phases they would acquire in absence of interaction. When
1 = 2 , the one states |0,1 and |1,0 are resonant and can exchange their only
excitation:
⎛ ⎞
1 0 0 0
z ⎜0 cos(J t) −i sin(J t) 0⎟
U (t) = e−iHeff t = e−it(σ1 +σ2 ) ⎜ ⎟
z

⎝0 −i sin(J t) cos(J t) 0⎠ . (8.2)

0 0 0 1

Evolution with this Hamiltonian for a brief period of time will result in the
joint action of three commuting operations: local rotations with the phases t, an
exchange of probability between the |0,1 and |1,0, and an interaction phase in
the single-excitation subspace, generated by σ1z σ2z . This combination of operations
is particularly useful to simulate fermionic models in the quantum computer, such
as in quantum chemistry applications. A particular instance of this evolution is
J t = 3π/2, which produces a gate equivalent to the iSWAP gate, up to local
transformations:
z z
U (π/2J ) = eiπ (σ1 +σ2 )/2J UiSWAP . (8.3)
The iSWAP is a universal operation – two iSWAP gates can build a CNOT gate
(Williams, 2011) – although it is less used in practical applications.
These exchange-type gates appear in all setups that allow for off-diagonal dipolar
interactions. They can be implemented in quantum computers with directly coupled
transmons – as in Figure 8.2 – but they also appear in devices with circuit-mediated
interactions. The work by Majer et al. (2007) used the dispersive coupling between
qubits in a superconducting cavity from Section 7.3.7 to implement the whole family
of rotations enabled by the exchange term, including the iSWAP. More recently, we
find these gates in the Google Sycamore quantum processor, which relies on detuned
transmons to mediate the interaction between qubits in the quantum register.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.3 Gate Toolbox 209

Their setup can recreate various dipolar interactions, including the exchange term
from (8.1), which they used to implement iSWAP two-qubit gates, as a foundation
for their 53-qubit quantum supremacy experiment (Arute et al., 2019).

8.3.3 Two-Qubit Tunable Frequency CZ Gate

Even though the iSWAP is a very natural choice for superconducting circuits, we
need to explore other gates that are more commonly used. One of them is the
CZ, control-Z, CPHASE or control-phase operation, shown in Table 8.1. This gate
and the CNOT are equivalent up to two local rotations (see Problem 8.2). The CZ
belongs to a the family of two-qubit phase gates, which are diagonal gates defined
by four phases θ (s1,s2 ):

U |s1,s2  = eiθ(s1,s2 ) |s1,s2  . (8.4)

A two-qubit phase gate is a maximally entangling and universal gate when

1 ±π
θ = {θ (1,1) + θ(0,0) − [θ (0,1) + θ (1,0)]} = mod 2π . (8.5)
4 4
All these two-qubit phase gates are equivalent to each other, up to single-qubit
transformation and global phases. A popular gate in this set is the CZ or control-Z
operation from Table 8.1 (cf. Exercise 8.4):
 iπ z
UCZ = e±iπ s1 s2 |s1,s2 s1,s2 | = exp (σ + 1)(σ2z + 1) .
s1,s2
4 1

The CZ may be generated by evolution with an interaction of the form σ1z σ2z . Unfor-
tunately, the coupling terms that we have found in Section 6.5 are not diagonal in
the qubit basis and cannot be directly used for this purpose.
A common way to implement phase gates is to do an adiabatic passage or adia-
batic deformation of the Hamiltonian that enhances the strength of interactions for
some period of time. Our system is built with a control parameter ω, which tunes the
Hamiltonian H (ω). We assume that the control begins and ends at the same point,
ω(0) = ω(T ) = ω0 , at which H (ω0 ) approximates to two noninteracting qubits
with a well-defined basis |s1,s2 . We design the passage so that around the middle
point ω(T /2), the interaction between qubits is dominant.
If the states of the two qubits are never degenerate when we change ω and if we
modify the control ω(t) slowly (adiabatically) enough, every two-qubit eigenstate
|s1,s2  of H (ω0 ) will be mapped to the instantaneous eigenstates |nω  of the Hamil-
tonian and return back to their original configuration. Those states only experience

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 210 Quantum Computing

some dynamical and geometric phases |s1,s2  eiθ(s1,s2 ) , which can be derived from
the instantaneous eigenenergies they experienced during the passage:
 T
1
θ (s1,s2 ) = − dtE(ns1,s2 ,ω(t)). (8.6)
h̄ 0

This idea was first used by DiCarlo et al. (2009) to implement a CZ gate for
transmon qubits in a cavity. It was later perfected by Barends et al. (2014) for
transmons that interacted directly with each other, in the design from Figure 8.2. The
CZ adiabatic gate requires two qubits with slightly different frequencies, 2 > 1 ,
and our control is the frequency of the second qubit 2 = ω(t). The transmons have
an always-on exchange interaction (cf. Section 6.5), either direct or mediated by the
cavity. This interaction is weaker than the anharmonicities α1,2 of both transmons
along the passage.
As explained in García-Ripoll et al. (2020), we can analyze the gate in the reduced
space {|0,0 , |1,0 , |0,1 , |1,1 , |0,2} of the qubit basis plus a state |0,2 with two
excitations in the second transmon. The effective Hamiltonian in this subspace reads
⎛ ⎞
0 0 0 0 0
⎜0 h̄ ⎟
⎜ 1 J 0 0 ⎟
⎜ ⎟
Heff (t) = ⎜ 0 J h̄ 2 0 √0 ⎟. (8.7)
⎜ ⎟
⎝0 0 0 h̄1√+ h̄2 2J ⎠
0 0 0 2J 2h̄2 − h̄α2

The gate protocol starts at a point of large detuning 2 − 1  |J |, shown in

Figure 8.3. The qubits almost do not see each other, we can neglect the couplings

Figure 8.3 Phase gate with two transmons with different energy gaps. (a) We tune
the frequency of the second transmon ω2 (t), from a largely detuned value down to
the resonance ω2 = ω1 + α2 between the |0,2 and |1,1 states. (b) Eigenenergies
of the tuned Hamiltonian, for levels that begin and end at the states |1,1 (solid)
and |0,2 (dashed). The |1,1 state is pushed downward from the eigenenergies of
the noninteracting model (J = 0, dotted) by an amount δEint (t).

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.3 Gate Toolbox 211

J , and the qubit basis is almost an eigenstate of Heff . The gate proceeds by low-
ering down 2 close to the point 2 = 1 + α2 and back to its original value (cf.
Figure 8.3a). If the tuning of the frequency is adiabatic enough and |J |  2 − 1 ,
states |0,0 , |1,0 , |0,1 remain approximately invariant throughout the gate.
The qubit register |1,1, on the other hand, experiences an interaction with
the |0,2 state that cannot be neglected. Throughout the evolution, |1,1 is there-
fore mapped to an adiabatically deformed eigenstate, with a trajectory shown in
Figure 8.3b as a solid line. Halfway through the passage, it reaches the super-
position state √12 (|1,1 − |0,2) around the degeneracy point 2 = 1 + α2 . This
deformation also changes the instantaneous eigenenergy, which √ is shifted downward
h̄1 + h̄2 − δEint (t), with a maximum value δEint ∼ 2 2J right at the avoided
level crossing.
The instantaneous eigenenergies of the various adiabatic passages manifest in
different dynamical phases for each of the two-qubit states. The three lowest energy
states acquire phases that are noninteracting in nature:
 T
θ(s1,s2 ) = − (1 s1 + 2 (t)s2 )dt. (8.8)
0

These are phases similar to the ones generated by the single-qubit Hamiltonian
1
2
h̄1 σ1z + 12 h̄2 (t)σ2z . The |1,1 state also acquires this local phase rotation, but
it also experiences the energy shift caused by the interaction:

i t
θ(1,1) = θ (0,1) + θ(1,0) − δEint (t)dt. (8.9)
h̄ 0
By adjusting the detuning 2 (t) (total duration, waiting time around degeneracy,
ramp profile, etc.), one may engineer a very good CZ gate that satisfies (8.5). In
addition to this, it is also possible to tune the rotation angle to produce other control-
phase gates, which may be useful in algorithms such as the quantum Fourier trans-
form (see Problem 8.4).
As shown experimentally (Barends et al., 2014) and theoretically (García-Ripoll
et al., 2020), this gate can be very fast. It does not need to be implemented adia-
batically. The only requirements for the gate to succeed are (i) minimizing leakage
and (ii) ensuring a large interaction phase. For the first condition, we need to ramp
down the qubit slow enough as to prevent transitions outside the two-excitation
subspace. Fortunately, this is limited by the anharmonicity, and happens already at
the speed of local gates (T  1/|α|).  The second condition only involves tuning
the waiting time, so that the integral δEint becomes large enough. Once more, √this
also happens in a relatively short time, of the order of the energy splitting 2 2J ,
with gates as short as tens of nanoseconds.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 212 Quantum Computing

8.3.4 Two-Qubit Tunable Coupling CZ Gate

The tunable-frequency CZ gate has been successfully demonstrated by various lab-
oratories, but it has three important drawbacks, which arise from our use of the
qubit detuning to modulate effective interactions. The first problem is the frequency
crowding phenomenon discussed in Section 6.5.3, which makes it very hard to
have a large number of qubits, parked at sufficiently distanced frequencies with
negligible interactions. The second problem is that, even for weak couplings, those
off-resonant interactions manifest as residual couplings that fundamentally limit the
achievable fidelity. The final problem is that during the frequency ramp, qubits may
accidentally cross degeneracies with other neighbors, not the ones we want them to
interact with, a phenomenon known as cross-talk.
An elegant solution to this architectural nightmare is the use of tunable couplers,
devices that mediate the interaction and allow us to switch it on and off, without the
need of moving the qubit’s frequencies (see Section 6.5.4). In the work by Chen et al.
(2014), a SQUID mediates the coupling between nearby transmons, in a scheme
that allows not only for a dipolar coupling, but also for canceling spurious residual
interactions.
Using this setup, it is possible to implement a CZ gate that builds on the
|11–|02 interaction, as before. Before and after the gate, the coupling is effectively
canceled, so that the qubit frequencies do not really matter, and can be optimized
for minimum decoherence (Arute et al., 2019). However, during the operation of
the gate, the qubits will be parked at resonant frequencies, 2 = 1 − α2 , and the
coupling will be switched on and off, from J (0) = 0 up to a maximum coupling
Jmax  α2 and back.
The model at hand is identical to the one before (8.7). We neglect the interaction
on the states |0,0 , |0,1, and |1,0, which remain instantaneous eigenstates of
Heff (t) and only acquire local phases (8.8). In the two-excitation subspace, however,
the instantaneous eigenstates are different:

1 1
|± = √ |1,1 − √ |0,2 . (8.10)
2 2

The state |1,1 suffers a different, but still quasi-adiabatic dynamics, that depends
on the energy splitting between these eigenstates 2 × δEint (t):

1
U (t) |1,1 = √ e−i(ω1 +ω2 )t × e−iθint (t) |1,1 − e+iθint |0,2 , (8.11)
2
 t
θint (t) = δEint (τ )dτ .
0

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.4 Tomography and Error Characterization 213

If we tune the passage so that θint (T ) = ±π mod 2π, we obtain a CZ gate. As

demonstrated by Chen et al. (2014), the gate can be implemented very rapidly,
in a time 1/J ∼ 30 ns, with very small leakage outside computational space and
reasonable fidelity.

8.4 Tomography and Error Characterization

We have discussed four out of the five essential ingredients in our shopping list for
a quantum computer, but we have not elaborated much on how errors in those items
impact the scalability of the computer, or even how to detect those errors. In this
section, we are going through that list again, identifying sources of imperfection.
Some of them have already been brought up in earlier chapters, but we will now also
derive general mathematical models for those errors and tools to characterize them.

8.4.1 Classes of Errors

DiVincenzo’s third requirement is actually insufficient. The checklist refers to “long
decoherence times,” assuming that all errors in the design come from decoherence.
Instead, we must consider four types of errors that pop up in a circuit model quantum
computer.

– State preparation. These are the errors we make when initializing the quantum
register. As mentioned earlier, thermalization by itself is not fast enough, and
we need to couple our qubits to other items (cavities, controls, measurement
devices) and apply complex protocols to ensure a fast reset. State preparation
errors account for deviations from the ideal zero state |0⊗N , which, as we saw
earlier, can be around 1–2%.
– Measurement errors. The precision of measurements in quantum computers is
limited. Limited integration time, measurement induced decoherence, and ampli-
fier and detector noise are some circumstances that require us to model measure-
ments as POVMs with intrinsic uncertainties. The measurement fidelity is the
probability that, given a certain outcome of our device, the measured qubit was
in that particular state. Circuit-QEDs’ fast dispersive qubit measurements seem
to have relatively good fidelities, with around ∼ 99–99.8% fidelity or success
probability (see Barends et al., 2014; Jeffrey et al., 2014; Walter et al., 2017). This
is good, but not yet so good that we can use unlimited repeated measurements
on the same qubit.
– Qubit decoherence. We have studied the loss of coherence and quantum
superpositions due to environment-induced relaxation, heating, and dephasing.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 214 Quantum Computing

The models from Sections 6.1.6 and 6.1.5 are good descriptions when these
process are uncorrelated – i.e., when qubits talk to separate environments in
Markovian processes that degrade the purity of the states over time. Those
models are simply characterized by the relaxation and dephasing times, T1 and
T2 , for each qubit. However, sometimes this characterization is insufficient,
because decoherence is correlated – e.g., fluctuations of a surrounding magnetic
field or control cables affect various qubits similarly – or because there is
residual cross-talk, unwanted interactions between the qubits underlying this
loss of information.
– Gate imperfections. Single-qubit and two-qubit unitaries are implemented
using microwave drives in combination with magnetic fields that tune the
frequencies and couplings of qubits. The operation of these gates is affected
by intrinsic errors as well as by our limited capacity on stabilizing and crafting
the controls. A π -pulse flip of the qubit will have errors if we fail to give the
microwave pulse the right amplitude, phase, or duration, and also if the frequency
of the qubit fluctuates slightly due to environment-induced dephasing. The
quality of our gates is measured in terms of a gate fidelity, which is the probability
that we implement the right operation or, alternatively, the mathematical overlap
between the experimental and the desired outcome.

There is nothing fundamentally wrong in quantum states being subject to errors

and imperfections. Our day-to-day electronics and devices are subject to similar
issues. Some causes are fundamentally unavoidable, such as the soft errors induced
by cosmic rays that hit our computers every few hours. However, engineers have
learned to detect and correct those errors, making semiconductor chips a reality
with immense scaling power. It is now well understood, and one of the funda-
mental pillars of the quantum computation theory, that we can follow a similar
roadmap: (i) understanding and modeling the errors that take place in our computer,
(ii) quantifying the intensity of those errors, and (iii) developing error-correcting and
fault-tolerant operation strategies.

8.4.2 Error Models: Completely Positive Maps

We use the formalism of trace-preserving completely positive maps to represent
the errors and imperfect operations in a quantum computer and any quantum infor-
mation process. We introduced this representation when studying qubit dephasing
(Section 6.1.5) and relaxation (Section 6.1.6). These maps are linear transformations
ε(ρ) of density matrices into density matrices (cf. Section B.1), and thus represent
the most general physical operations that are allowed by quantum mechanics.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.4 Tomography and Error Characterization 215

Table 8.2. Completely positive maps for useful error models.

Bit flip ε(ρ) = pρ + (1 − p)σ x ρσ x

Phase flip ε(ρ) = (1 − p)ρ + pσ z ρσ z
Bit flip ε(ρ) = pρ + (1 − p)σ x ρσ x
Depolarizing channel ε(ρ) = (1
 − 3p/4)ρ + p/4σ x
 ρσ
x

Totally depolarizing channel ε(ρ) = 1 − α=x,y,z pα ρ + α=x,y,z pα σ α ρσ α
(asymmetric)
Totally depolarizing channel ε(ρ) = (1 − p) ρ + p 1d
√ √
Amplitude damping A1 = |0 0| + 1 − γ |1 1|, A2 = γ |0 1|
√ √
Phase damping A1 = |0 0| + 1 − γ |1 1|, A2 = γ |1 1|

A trace-preserving completely positive map is a linear superoperator that trans-

forms density matrices into normalized density matrices – i.e., probability is
conserved, and we do not discard events – which admits a decomposition into
Kraus operators:

ε(ρ) = Ai ρA†i . (8.12)
i
† 
The Kraus operators satisfy i Ai Ai = 1, because they produce normalized
density matrices, and we may have up to 2N of them.
In the ideal operation of a quantum computer, a positive map can be just a unitary
transformation ε(ρ) = U ρU † , describing a quantum gate with a single Kraus oper-
ator A = U . In the most general case, a physical map will describe an imperfect
gate, decoherence of a quantum register, or even the loss of information during
measurements.
Quite often, we make physically reasonable assumptions about the underlying
errors. A common one is to separate the ideal gates from the errors, as if they were
independent processes3 εerror (U ρU † ). Quite often, we also assume that those errors
decompose into spatially uncorrelated errors acting on the idle qubits, and some
average error that is independent of the operation, acting only on the qubits that
are manipulated. Moreover, in the description of the spatially uncorrelated errors,
we also resort to a family of models, summarized in Table 8.2, that are physically
appealing and have simple parameterizations.
These simpler models and views are needed to develop intuition, understand the
experiments, and make progress in the development of error-correction strategies.
Without simplification, completely positive maps are humongous objects: For a

3 For unitary operations, this is always possible, at least formally.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 216 Quantum Computing

system of N qubits, the most general map is described by a matrix with about
d 4 = 24N parameters. To see that, let us examine the Bloch sphere representation
of a density matrix, introducing a complete set of observables formed by all tensor
products of Pauli matrices4 Sα ∈ {1,σ x ,σ y ,σ z }⊗N . The 22N operators form a com-
plete orthogonal basis with respect to the matrix scalar product tr(Sα† Sβ ) = dδα,β .
This allows us to expand any density matrix:
d2  
1 1 †
ρ =s ·S=
T
sα Sα, with sα = tr Sα ρ = Sα ρ . (8.13)
α=1
d d
A general completely positive map ε(ρ) is a linear transformation of the generalized
Bloch vector, represented by a matrix M ∈ R4 ×4 with about 16N independent
N N

coefficients:

1 
d 2

ε(ρ) = (Ms) · S = T
Mα,β tr(Sα† ρ)Sβ . (8.14)
d α,β=1

There are other alternative representations for the completely positive maps in an
experiment. The χ matrix (Gilchrist et al., 2005; O’Brien et al., 2004) expresses the
superoperator in a complete basis Sα :

d 2

ε(ρ) = χαβ Sα ρSβ† . (8.15)

α,β=1

For physical maps, the matrix χ is positive semidefinite and can be diagonalized
χ = V V † , producing one of many equivalent definitions for the Kraus operators

A = Vα Sα (Neeley et al., 2008). This representation is linearly related to the
Pauli transformation matrix M:
 d 2
1
Mαβ = tr(Sα Sγ Sβ Sδ )χγ δ, (8.16)
γ ,δ=1
d

and equivalent from any formal or operational point of view.

A third representation is based on the Choi–Jamiolkowski isomorphism between
quantum processes and quantum states. The Choi matrix for a quantum process
ε(ρ) is obtained by applying the positive map onto one subsystem of a maximally

entangled state | = i √1d |ii:
 
ρε := (1 ⊗ ε) (||) = Mαβ SβT ⊗ Sα = χαβ |αβ| . (8.17)
α,β α,β

4 For two qubits, for instance, this would be formed by all operators in the set of 16 combinations: {1, σ x , σ y ,
1 1
y y y y y y y
σ1z , σ2x , σ2 , σ2z , σ1x σ2x , σ1x σ2 ,σ1x σ2z , σ1 σ2x , σ1 σ2 , σ1 σ2z , σ1z σ2x , σ1z σ2 , σ1z σ2z }. Note how these operators are
both Hermitian and unitary.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.4 Tomography and Error Characterization 217

It is linearly reated to two other positive map representations, M (8.14)

√ and χ (8.15),

through the convenient entangled basis |α = i (1 ⊗ Sα ) |ii / d.
Quantum process tomography is the reconstruction of the superoperator associ-
ated to a physical map, ε(ρ) or M. The tomography protocol uses a complete set
of observables Sα together with a complete set of states constructed from them:5
ρk = Bkα Sα . The protocol involves preparing the 22N or more input states ρk and,
for each possible input, measuring the 22N observables Sα multiple times.
The outcome of a tomography experiment is a set of real numbers that are unbi-
ased estimators of the expectation values λα,k  Sα ε(ρk ) that must be related to the
positive map ε. If there was no statistical uncertainty, we establish a linear relation
between the estimators and the matrix M:
λα,k = tr(Sα ε(ρk )) = Mα,β Bβ,k . (8.18)
In practical situations, this equation cannot be solved, because the statistical uncer-
tainties in λα,k create instabilities that produce unphysical maps.
A more robust strategy is to realize that for M to be physically acceptable, the

Choi matrix ρε := α,β Mα,β SβT ⊗ Sα must be positive semidefinite (Merkel et al.,
2013). The technique of maximum likelihood estimate (MLE) is a convex optimiza-
tion method that finds the superoperator M that best approximates all experimental
estimates – e.g., the expectation values λα,k – while satisfying the physical con-
straints for this matrix (see supplementary material for Chow et al., 2012).
Standard quantum process tomography is a costly and delicate process. It has
been used to study the process tomography of free evolution and local gates for
one or two qubits (Chow et al., 2009, 2012; Yamamoto et al., 2010), using both the
M or χ matrix representation. In either case, the tomography does not scale well
for larger setups. It involves multiple repetitions of 24N experiments, and a very
delicate handling of uncertainties and systematic errors. Moreover, as shown by
Merkel et al. (2013), state preparation and measurement (SPAM) errors can make
the reconstruction process unstable, or at least produce unreasonable estimates of
the tomographic errors. This motivates the search for alternative characterization of
errors, which are less general and do not care so much about the error model, but
focus on the intensity or strength of those errors.

8.4.3 Error Quantification: Fidelity

One way to quantify the relevance of noise and decoherence in an experiment is
to use a metric that tells us how far our states and quantum processes are from

5 For instance, a qubit has S ∈ {1,σ x ,σ y ,σ z }, and we could use the overcomplete set of pure states
ρk ∈ { 21 (1 ± σ z ), 12 (1 ± σ x ), 12 (1 ± σ y )} as reconstruction basis (Chow et al., 2012).

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 218 Quantum Computing

their ideal targets. The natural measure for quantum states is the fidelity, a number
between 0 and 1 determining the probability that we prepared the desired quantum
state or reference. Given any two states, described by density matrices ρ and σ , their
mutual fidelity is
  2
F (ρ,σ ) = tr ρ 1/2 σρ 1/2 ∈ [0,1]. (8.19)

Note how this expression simplifies to the overlap squared for pure states
F (|ψ , |φ) = |ψ|φ|2 . It is clear from this particular case that a perfect fidelity
F = 1, read as “100% fidelity,” is only achieved for identical states, up to global
phases. The infidelity  = 1 − F (ρ,σ ) measures the error in approximating σ by ρ
and is related to their separation.6
The notion of fidelity can be extended to work with quantum processes in differ-
ent ways. For instance, say that we wish to compare an ideal unitary operation U
to the actual process that takes place in the lab ε(ρ). Formally, we could apply ε(ρ)
to different pure states |ψ and estimate the average probability that we obtain the
right answer – i.e., the state fidelity between U |ψ and ε(|ψψ|). This leads to the
average gate fidelity:

Favg (ε,U ) = ψ|U † ε(|ψψ|)U |ψ dψ. (8.20)

This integral is computed by averaging over all possible pure states, sampled with
a uniform (Haar) distribution.
The average gate fidelity works with general positive maps, but as argued by
Gilchrist et al. (2005), it lacks stability under composition with other trivial opera-
tions.7 The solution to this problem is given by the process fidelity or entanglement
fidelity, a quantifier based on the Choi–Jamiolkowski isomorphism between pro-
cesses and states (8.17).
We introduce the process fidelity of two completely positive maps, using their
Choi matrices as proxies for the comparison:8
1
Fproc (ε0,ε1 ) := tr ρε†0 ρε1 = tr M0† M1 = tr χ0† χ1 . (8.21)
d2

6 For normalized pure states, !ψ − φ!2 = 2(1 − F 1/2 )   for small errors. Note that traditional definitions in
the literature chose the geometrical convenience (see Nielsen and Chuang (2011)) F 1/2 = |ψ|φ| as fidelity
for pure states. You find this convention in early works in quantum technologies (Steffen et al., 2006), but it
has since been abandoned.
7 Imagine that we put together our system and an uncoupled experiment where we do nothing. We would expect
that the auxiliary system should not modify the average of the fidelity, Favg (1 ⊗ ε0,1 ⊗ ε) = Favg (ε0,ε), but
this is not true in general!
8 Originally, the process fidelity was introduced as F †
proc (ε,U ) = |U ε(||)U |, to compare positive
maps ε with ideal unitary operations (Gilchrist et al., 2005). This formula measures the fidelity of the Choi
matrix for the positive map, with respect to the pure state associated to U .

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.4 Tomography and Error Characterization 219

Interestingly, the average gate fidelity and the process fidelity are related:

dFproc + 1 tr(M0† M1 ) + d
Favg = = . (8.22)
d +1 d2 + d

If an experiment is designed for full process tomography, we can recover both Fproc
and Favg . This idea has been applied to the study of a qubit under no operations (Nee-
ley et al., 2008) to characterize single-qubit operations (Chow et al., 2009) and to
study universal two-qubit gates (Bialczak et al., 2010). Unfortunately, it is difficult
to scale up this quantifier to larger numbers of qubits due to exponential growth in
experimental setups required for complete process tomography ∼ d 4 ∼ 24N . This
limitation has motivated the quest for simpler experimental protocols that exploit
the power of randomized operations to estimate the fidelities of state preparation,
measurements, and single-qubit and two-qubit gates.

8.4.4 Randomized Benchmarking

The term randomized benchmarking (RBM) denotes a family of experimental meth-
ods that estimate the average errors of quantum operations in a quantum register, as
well as the combined SPAM errors. There are various incantations of this technique,
all of them based on these two principles:

– We can approximate the average over states in (8.20) by perturbing a state

|ψ randomly with Pauli operations that act locally on each qubit – e.g. the
set {Sα } we introduced before – and averaging over many realizations of these
perturbations.
– If there are no systematic errors, our experimental realization of the elementary
single- and two-qubit gates is an approximate combination of the ideal gate and
an average error map εavg (U ρU † ). The average error does not depend on the gate,
but may be different for single- and two-qubit gates.

A randomized benchmarking protocol uses random Pauli sampling to provide

an estimate of the average error or average infidelity associated to a set of gates,
Favg (εavg ). There are many variations of randomized benchmarking, but most focus
on the Clifford group C. This is the group that transforms the tensor products of
nonidentity Pauli matrices onto other products of Pauli matrices, up to signs and
global phases. The group is generated by Hadamard and π/2 phase rotations on
each qubit, plus CNOT gates on any pairs of qubits (Gottesman, 1998). It is of par-
ticular interest because it contains the gates that are used in stabilizer-state quantum
error correction, but this group is not, by itself, sufficient for universal quantum
computation.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 220 Quantum Computing

A typical RBM protocol follows this algorithm:

(1) Select a set of natural numbers L = {l1 < l2 < · · · < lm }.
(2) For each integer l ∈ L, generate a random sequence of l gates in the set
{g1,g2, . . . ,gl } ∈ C.
(3) Repeat these steps Np times, to gather sufficient statistics:
(a) Create a random sequence of Pauli operations {S1,S2, . . . ,Sl }.
(b) Compute the state ρx = gl Sl gl−1 Sl−1 . . . g1 S1 |0.
(c) Determine a Pauli Sl+1 and an additional gate gl+1 such that
ρf = gl+1 Sl+1 ρx is an eigenstate of the σiz operators.
(d) Create experimentally the state ρf and measure whether σiz operators cor-
respond to the prediction. If so, increment the counter Cl .
(4) Estimate the probability of success for each length l, using the outcomes of
these experiments, as Pl = Cl /Np .
Typically, the probability of success Pl can be fitted to an exponentially decaying
function:
Pl = A0 pl + B0 . (8.23)
The average error map therefore behaves as the completely depolarizing model
from Table 8.2, with the average depolarization parameter p. The fitting parameters
A0 and B0 then include contributions from preparation and measurement errors. If
d = 2N is the dimension of the Hilbert space for a register with N qubits, the average
fidelity and the average error for a single quantum operation can be estimated as
p d −1
Favg = + (1 − p),  = 1 − Favg = (1 − p). (8.24)
d d
These are averaged quantities, not only over input states, but also over all the possi-
ble Clifford gates that our quantum computer could implement. We expect that this
double-averaging justifies even more the assumptions listed at the beginning of this
section.
The same protocol can be extended to calibrate one specific gate, G, with some
ad hoc changes: First, we insert the benchmarked gate after each Clifford operation:

,
1
ρf = gl+1 Sl+1 Ggl Sl |0 . (8.25)
i=l

This produces a similar exponential fit (8.23), with an updated depolarization

parameter pG . The average error per gate for G is estimated as
 
d −1 pG
G = 1− . (8.26)
d p

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 221

The operation G may be as simple as idling or waiting for a time t after each
Clifford gate. When doing so, we are calibrating the coherence of our qubits, either
individually – when the experiment is done on each qubit separately – or as a whole –
when we consider the Clifford group on more than one qubit. O’Malley et al. (2015)
has taken this idea a step further, developing an error metrology protocol that com-
bines randomized benchmarking with Ramsey-type pulses. The result is a very
sensitive analysis toolbox to explore the decoherence and dephasing properties of
the qubits, including a careful study of the phase fluctuations that bound deviations
from the exponential decay model.
Randomized benchmarking was introduced by Knill et al. (2008) and later gen-
eralized to multiple qubits by Magesan et al. (2011) and Gaebler et al. (2012). It
exploits important connections between the performance of quantum computers in
large, arbitrary calculations and the quality of the ingredients we built into the com-
puter. In particular, it demonstrates in an intuitively appealing way the exponential
accumulation of tiny errors in the quantum processors. For instance, a failure rate
as small as 0.1% error per gate transforms into a success probability of P20 = 80%
after 20 gates, P40 = 66% after 40, and P160 = 20% after 160, and so on. Considering
that useful quantum algorithms must be much longer to compete against classical
methods, it becomes obvious that we need to find ways to cope with this rapid
deterioration of the quantum register, incorporating some type of error correction.
One may argue that randomized benchmarking provides a standard, low-cost,9
and reproducible strategy to map out the quality of a quantum computer’s compo-
nents. In particular, RBM is a very generic algorithm that applies to all quantum
computers based on the circuit model. It therefore enables a systematic comparison
of setups and platforms, such as superconducting circuits versus trapped ions –
Magesan et al. (2012) versus Gaebler et al. (2012) – which is not very different
from how we compare the quality of high-end versus low-end electronics – e.g.,
using global figures such as the failure rates of their hard disks, memories, and
processors. However, keep in mind that by doing so we might be dropping relevant
information about the nature and correlations among errors, which is crucial for the
development of error correction and fault-tolerant computation strategies.

8.5 Fault-Tolerant Quantum Computers

Imagine a hard disk that can store classical bits, with a low probability of corruption
ε after a time T . The probability that one bit is faulty after all this time grows
exponentially p(1,L) = ε(1−ε)L L1 ∼ O(εLL ) with the number of bits L. To avoid

9 At least when compared with full tomography.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 222 Quantum Computing

this scenario, we can use a simple error-correction strategy, which is to copy the
information redundantly, using N physical bits to encode one logical bit, our unit
of information. Because the setup is classical, we may inspect the states of these
logical bits periodically, to see whether one or more bits have changed their state
and fix them.
In this classical scenario and encoding, a simple error-correction strategy is to
implement a majority vote, which copies to all physical bits the state that is occupied
by at least N/2 of them. This majority vote strategy works, because the probability
that n out of N bits have been corrupted decreases exponentially with n. If we sum
the probability that n > N/2 bits have been corrupted – in which case the majority
vote would be wrong! – we obtain the probability that our error-correction fails. This
probability scales with the redundancy as O(N N εN/2 ). The error – correction strat-
egy therefore works provided ε  1/N, which we may achieve if we interrogate
the physical qubits with fast enough frequency.
Unfortunately, this error-correction scheme does not immediately work for quan-
tum information. The no cloning theorem (Dieks, 1982; Wootters and Zurek, 1982)
implies that we cannot copy an arbitrary quantum state of one qubit onto other
qubits, going from |φ |0N−1 to |φN . But, more fundamentally, there is no way
that we can measure a qubit without destroying the quantum superpositions that
might be stored in them. We need other encodings and other strategies that are well
adapted to the quantum world.

8.5.1 Local Errors and Global Qubits

All quantum or classical correction strategies need some assumptions about the
errors that we need to protect ourselves from – the error model. This concerns the
type of perturbations that our system will experience, and the probability that those
perturbations occur in a given time. In our preceding example, we worked with local
bit flips, taking place at rate ∼ /T . We adopted a global strategy, the majority vote,
which was able to find the original information in a global degree of freedom of
our bits.
This distinction – local errors versus globally stored information – maps rather
well to the quantum treatment of information. We typically assume that quantum
errors happen locally on each qubit. The physical intuition is that each qubit
interacts with its immediate environment, which is sufficiently large, complex, and
Markovian, that it cannot correlate this qubit without far away qubits in the register.
In this scenario, which may be confirmed experimentally with RBM or similar
techniques, the error models are compositions of spatially uncorrelated incoherent
transformations – e.g., any error model from Table 8.2, from depolarization to pure
dephasing.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 223

If this assessment is correct, we can develop a redundant encoding of quantum

information into logical qubits, using collective degrees of freedom that are only
mildly affected by the local errors. In this encoding, the Hilbert space of N qubits
is mapped to a collection of states |O,α. The labels O run over 2m collective states
that encode the Hilbert space of m logical qubits.
One expectation is that, by using collective or redundant degrees of freedom, we
will reduce the probability that information is completely lost. Then, depending on
the approach to error correction we take, we may assume that the additional degrees
of freedom α will either absorb the entropy arising from decoherence, or we will
expect to be able to detect, from changes α → α  , which errors took place – and
in that case, undo the error in the relevant degrees of freedom O with appropriate
unitary transformations Uα→α . In both cases, our information and quantum super-
positions will survive, because we never look into the logical quantum register!

8.5.2 Passive versus Active Error Correction

There are two big types of error correction: There is intrinsic fault tolerance, and
there is active error correction and fault-tolerant operation. We say that a quantum
system is intrinsically resilient to errors – intrinsically protected or relies on passive
error correction – when the energy level structure protects the logical qubits from
external perturbations.
Ideally, the logical qubit degrees of freedom Ôn should commute with the
operators that couple the quantum register to the environment [Ôn, ĝm ] = 0,
so that the logical states are not affected by decoherence. More generally, we
will find that the environment is unable change the state of the logical qubits
O|ĝm1 ĝm2 · · · ĝmk |O   = 0, unless we have a large number k ≥ kmin of cou-
pling operators. By considering the effects of the environment perturbatively (see
Section A.3.2), we will confirm an exponential suppression of logical perturbations
that could be sufficient for practical computations – specially if kmin grows with the
size of the system!
One example of this approach is the Majorana excitations in a hybrid
superconductor–nanowire interface. These are excitations of one delocalized
particle onto two spatially separated regions of the wire that cannot be distinguished
locally. Quantum computation is still possible, by braiding Majorana excitations,
in a model of topological quantum computation that is intrinsically fault tolerant –
not only the logical qubits, but all operations take place in the the logical space and
are immune to errors.
Intrinsic protection is rare outside the world of topologically protected models
and topological quantum computation. A more general approach is to engi-
neer active error-correction methods, using the building blocks of the circuit

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 224 Quantum Computing

model – single-qubit gates, two-qubit gates, and measurements – to correct errors in

the information-free degrees of freedom |α, before they accumulate and leak into
the logical qubit space. Shor (1995) first suggested this possibility for a small set of
errors, using a very wasteful repetition code. Later, Steane (1996) and Calderbank
and Shor (1996) introduced a scalable approach to error correction that spreads
the information of a logical qubit over multiple physical qubits with the help of
entanglement.
The works by Steane, Calderbank, and Shor also introduced explicit error
models – i.e., they decomposed the expected errors into fundamental transfor-
mations of the physical qubits – and studied rigorously at which rates those errors
could be corrected. However, their approach to error correction poses difficulties
in the generalization to arbitrary error models, growing to arbitrary code sizes, and
implementing completely fault-tolerant quantum operations – i.e., implementing
quantum computations in the presence of errors, not just correcting the decoherence
in an imperfect qubit. Nowadays, both theoreticians and experimentalists have
shifted their focus to other schemes that are qualitatively simple to understand and
build at scale using the architectures we have discussed.

8.5.3 Stabilizer Codes

The stabilizer formalism is a particular approach to error correction that attempts to
stabilize a quiescent state |α = |0, using a complete set of measurements on the
information-free degrees of freedom. The method identifies a large enough set of
stabilizer operators Pk that commute with each other [Pk,Pk ] = 0. Typically, the
stabilizers have eigenvalues ±1, and the quiescent state is the only eigenstate that
satisfies Pk |0 = + |0 , ∀k.
The error-detection phase consists in measuring all observables Pk to obtain the
syndrome {pk }, a collection of measurement outcomes. When any of these values is
different from +1, we know that an error-occurred and mutated the quiescent state
|0 → |α  .
The error correction phase investigates the values of the syndrome and, through
a classical algorithm, proposes with a unitary transformation to undo the errors
U |α   → |0. As in the majority vote method, there is a maximum amount of
errors that a stabilizer method can correct. Above a certain number, we can no longer
determine which state |α   we have and how to take it back to the quiescent state.
But even below that threshold, the classical computation of U may be extremely
hard when there are many changes in the syndrome. To prevent either problem, we
must repeat the error-detection and correction stages frequently enough to reduce
the average number of perturbations in the syndrome.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 225

The good thing about stabilizer formalisms is that they can be implemented using
the tools from the circuit model. Both the measurements Pk and the corrections U
are small quantum algorithms that can be compiled into elementary single- and two-
qubit gates and single-qubit measurements, possibly with some auxiliary qubits to
assist the syndrome measurements.
Most importantly, the types of stabilizer codes that we care about also allow the
implementation of fault-tolerant computations. These are unitary gates that act on
the logical qubit space without perturbing the stabilizer space or the syndrome.
Such gates are called “fault tolerant” because they work even if we use imperfect
components to reproduce them – i.e., single qubit gates with some residual errors,
imperfect measurements, etc. When we implement the fault-tolerant gate, errors
manifest as perturbations of the syndrome that can be corrected, producing the ideal
operation.
Let us emphasize that we can only implement fault-tolerant gates when the syn-
drome does not show too many changes in between correction steps. To put limits on
this, we set the fault-tolerance threshold, a bound in the infidelity of any operation
from our quantum toolbox – including state preparation, measurement, and gates,10
all of which can be characterized via RBM. In theory, when a platform remains
within the specified thresholds, we can create fault-tolerant computers working
with quantum registers and quantum algorithms of arbitrary size. In practice, this is
largely dependent on assumptions such as noise models, code sizes, and the specific
correction protocols.

8.5.4 Surface Code

There are various stabilizer codes in the literature, but two topological models have
significantly driven change in the quantum computing world: these are the surface
code by Bravyi and Kitaev (1998) and the color code by Bombin and Martin-
Delgado (2006). Both codes have similar properties:
– They are lattice-based codes. They assume that qubits are laid out on a planar
topology, with nearest-neighbor interactions.
– They are stabilizer codes, where the stabilizer operators are multiqubit observ-
ables defined by the plaquettes and edges of the lattice.
– Both codes support fault-tolerant computations, with universal sets of single-
qubit and two-qubit gates.

10 Including periods in which a qubit is idle, which are treated as “identity” operations.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 226 Quantum Computing

Figure 8.4 (a) Surface code layout corresponding to the qubit arrangement from
Figure 8.2b. We show a plaquette and a vertex stabilizer operator, and one of the
ancillary qubits. (b) Quantum circuit associated to the measurement of a σ1z σ2z σ3z σ4z
plaquette stabilizer, using the central auxiliary qubit for the measurement.

– They can encode any number of logical qubits, provided the lattice sizes are large
enough.
– The codes can be implemented in a formally simple way, using the single- and
two-qubit operations from the circuit model.

We will now work out the ingredients of the surface code formalism, using a simple
lattice as an example. For further reference, we recommend the review articles by
Fowler et al. (2009, 2012).

Stabilizer Operators
Figure 8.4a shows a toy realization of a surface code, where the qubits match the
numbered transmons in the quantum register from Figure 8.2b. The surface code
operates on a square lattice, but the qubits sit on the edges, not on the vertices of
the lattice. A whole plaquette contains four edges and four qubits; we also accept
open plaquettes formed by just three edges. The example in Figure 8.4a has four
complete plaquetes and 2 + 2 incomplete ones, distributed along the two rough
edges on the sides.
There are two sets of stabilizers on this lattice. One set is formed by products of
z
σ operators on the qubits of each plaquette; the other set is constructed by products
of four σ x operators from the qubits surrounding a vertex. Figure 8.4a illustrates two
stabilizers: Z2 = σ2z σ5z σ6z σ9z and X6 = σ7x σ10
x x x
σ11 σ14 , assuming we order qubits from
left to right, top to bottom, as in Figure 8.2b.

Logical Qubits and Local Gates

Logical qubits are best described using the collective observables that act as the
qubit’s Pauli matrices – e.g., they can be used to expand the density matrix using a

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 227

y
Bloch vector. These operators implement the usual pseudospin algebra [jx ,k ] =
y
2i z δj k , [j ,kz ] = 2i x δj k , and [jz,kx ] = 2i x δj k . They also commute with
all stabilizers [ x,y,z,Zn ] = [ x,y,z,Xn ] = 0, because the stabilizers operate in the
independent degrees of freedom |α.
Logical qubits in the surface code are created by the boundaries and the holes
in the lattice. Each defect decreases the number of commuting stabilizers, leaving
undetermined degrees of freedom to be used for storing quantum information. The
logical qubit operators for these degrees of freedom are constructed from products
of Pauli matrices of physical qubits across the lattice, along the edges that connect
defects (holes) and borders. One may create an arbitrary number of qubits in large
lattices, provided they accommodate enough independent rough edges and holes
(Fowler et al., 2009, 2012).
The example from Figure 8.4a hosts 18 qubits and 17 stabilizers – eight stabilizers
of type Z and nine of type X. Since the stabilizers can take two values {+1,−1} and
are linearly independent, they jointly determine that the dimension of the subspace
|α is 217 . The remaining degrees of freedom can be grouped in a logical Hilbert
space of dimension 218−17 = 2, one qubit, described by three Pauli-like operators
{ x,y,z }. The diagonal Pauli operator for the logical qubit is any product of σ z along
a string that connects both rough edges – e.g.,  z = σ8z σ9z σ10 z z
σ11 . The real off-
diagonal operator string operator is given by  = σ2 σ9 σ16 , or any topologically
x x x x

equivalent string that crosses plaquettes from top to bottom. With these two trans-
formations, we also have  y = −i z  x , totaling three logical gates  x,y,z , that we
can implement by manipulating the physical qubits.

Error Detection and Correction

The error model for the surface code is based on detecting two elementary types
of errors: spin-flips and phases. Both errors correspond to Pauli unitary transfor-
mations, as in σ x ρσ x or σ z ρσ z . By extension, this error model also detects Pauli
Y errors, which can be decomposed as a phase in combination with a spin-flip,
σ y ρσ y = σ x σ z ρσ z σ x . Consequently, the surface code may correct many of the
noise models from Table 8.2, from pure dephasing, to completely depolarizing
channels.
As mentioned earlier, these errors are detected as changes in two or more values
of the syndrome measurements. Take for instance the ninth qubit on our lattice and
imagine it experiences a bit flip. This perturbation affects two stabilizers that overlap
on this qubit, Z2 and Z6 = σ9z σ12 z z z
σ13 σ16 . Their expectation values are transformed
by the unitary perturbation U as O,α|U † Z2,6 U |O,α, which means that in the
Heisenberg picture the new stabilizers have changed sign σ9x † Z2,6 σ9x = −Z2,6 . If
the error is a π phase, then the affected stabilizers will be X4 = σ5x σ8x σ9x σ12 x
and

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 228 Quantum Computing

X5 = σ6x σ9x σ10 σ13 , which will also change sign, σ9z † X4,5 σ9z = −X4,5 . And finally,
x x

Pauli Y errors on this qubit manifest as joint phase and spin-flip errors, which affect
all four stabilizers that overlap on the affected qubit Z2,6 and X4,5 .
We build our our error-detection phase on the capacity of the surface code to
detect all these single-qubit transformations – σ x ,σ y , and σ z – in isolated or multiple
qubits. Qualitatively, we rely on the expansion of a general completely positive
map as a linear combination of products of Pauli matrices – the Sα operators from
Section 8.4.2. Following this expansion, we interpret generic errors as statistical
superpositions of single-qubit transformation, which can be detected.
The error detection works when changes on the syndrome can be mapped,
unequivocally, to the actual Pauli transformations. For this, we need the expansion
to involve products of few operators, and that these products are localized enough
as not to change the values of one or more logical qubits.11 We achieve this safe
regime by assuming architectures with spatially uncorrelated errors and ensuring
that error monitoring takes place faster than the rate of errors. In this situation, we
can identify not just the errors, but a successful error-correction strategy to bring
back the syndrome to the desired quiescent state Xk,Zj = 0 ∀k,j .
The measurement of the stabilizers, like all other operations in the surface code,
can be implemented the quantum circuit toolbox. The stabilizer operators such as
Z1 , Z2 , or X4 are three- or four-qubit observables that measure the parity12 of
those qubits in either the σ z or σ x basis. This parity measurement requires a single
auxiliary qubit and four two-qubit operations. The algorithm shown in Figure 8.4b
maps the Z-parity of four qubits – i.e., a plaquette – onto an auxiliary qubit that
is placed at the center of it. The extra qubit begins on the |0 state and suffers as
many spin-flips as neighboring qubits are in the |1 state. The outcome is |0 ⊕ P ,
where P = 12 (1 + Z) is 1 if the parity is odd on the plaquette, or 0 otherwise.
A similar algorithm can be constructed for X-parity measurements, with auxiliary
qubits placed on the vertices.
Since the parity measurements demand auxiliary qubits, it is quite common that
these qubits be included in the architecture of the quantum register. For instance,
the surface-code lattice with 18 qubits from Figure 8.4a could be constructed using
the white qubits from Figure 8.2b, with the additional 18 qubits as targets for the
parity measurement.

11 Imagine that the transformations happen to connect the boundaries, as in  x = σ x σ x σ x . This error is
2 9 16
actually flipping the logical qubit, and is no longer localized to the degrees of freedom |α that we can safely
manipulate.
12 Take a set of bits s ,s , . . . ,s ∈ {0,1}. The parity is even or odd depending on whether the number of ones is
1 2 n
even or odd. This definition extends to a qubit basis by determining how many qubits return a positive result
+1 when measuring σ z or σ x .

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 229

The error-correction phase, in which we bring the syndrome back to the original
value, is implemented as a collection of single-qubit quantum gates acting on strings
of physical qubits. These strings of Pauli matrices run along the edges that connect
vertices or plaquettes with incorrect stabilizer values, correcting all of the errors
simultaneously. The identification of the actual errors and the unitary operations that
undo them is a computationally intensive stage that is still subject to investigation
and improvement.
As an illustration of the difficulty in finding the right error-correction strategy, it
is illustrative to read the article by Kelly et al. (2015). This experiment demonstrated
the operation of a repetition code – a simpler 1D stabilizer code that works with the
parities of neighboring qubits Zi = σiz σi+1 z
– discussing strategies for the detection
of errors. More recently, Chen et al. (2021) demonstrated this code working with
up to 21 qubits and 50 layers of correction using Google’s Sycamore quantum
processor, in a work that also showed an implementation of the surface code with
four qubits, for a more limited number of operations.

Universal Gates
From the previous discussion, we could infer how to implement three logical single-
qubit gates,  x ,  z , and  y , and any logical-qubit measurement, in a fault-tolerant
way. The first step is to implement the gate or measure the collective observable, as
prescribed by the definition of the logical operator. The second step involves mea-
suring the syndrome and correct any errors – which in the case of the measurement
might involve updating the measured value to reflect the changes in the physical
qubits.
In practice, the  x,y,z gates are not needed. As explained by Fowler et al. (2012),
we only need to care about building a universal set of gates with the H , S, T , and
CNOTs operating on the logical qubits. The  x,y,z will be “moved” out of the algo-
rithm and converted in corrections of the measurement outcomes.13 However, this
does not really simplify the task of explaining how the new gates are implemented.
Qualitatively, a standard implementation (Fowler et al., 2012) relies on the pos-
sibility to manipulate the qubit lattice in topologically nontrivial ways. First of
all, to create the surface code lattice, we need to measure the stabilizers of a set
of qubits – the ones that form the lattice – correcting those values to the desired
outcome. This already suggests that we can engineer holes or defects, or cut a
lattice into two separate pieces, by ignoring certain qubits, not considering them
part of any stabilizer. Conversely, we can merge lattices by stabilizer measurements

13 Consider a stupid algorithm that combines X , Z , and a CN OT acting on both qubits C . We can
1 2 12
transform (X1 ⊗ Z2 )C12 = (1 ⊗ Z2 )C12 (X1 ⊗ X2 ) = C12 (X1 ⊗ (−Z2 X2 )). These ideas allow us to move
the Pauli matrices to the end, where they only affect the final measurement outcome.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 230 Quantum Computing

along their boundaries, and include new qubits or fill holes by including those
qubits in the definition of the stabilizer.14
If we define the logical qubits as strings that connect holes in the lattice, we can
implement CNOT gates in a completely fault-tolerant operation by braiding holes
around each other. The Hadamard gates H are also implemented in a fault-tolerant
way that involves isolating the region of the lattice that contains the qubit, doing
some key measurements, and merging this patch back to the lattice. In both cases,
errors introduced during either the CNOT or the H are self-corrected by the repeated
measurement and corrections of the stabilizer group.
Unfortunately, the S and T gates are more complicated. They must be imple-
mented using quantum algorithms that rely on H , CNOT, and auxiliary logical
qubits prepared in a given state. The preparation of these auxiliary states is a delicate
operation not covered by the stabilizer formalism, and which may introduce errors.
To make these errors as small as possible, we rely on quantum information recipes
that fall outside the code itself. In particular, we use state distillation, which com-
bines multiple imperfect copies of a state to produce one with higher fidelity. All
this makes the S and T operations the most costly of the universal and fault-tolerant
gate set – and a valuable resource that is thoroughly analyzed when designing a
quantum algorithm.

8.5.5 Fault-Tolerant Thresholds and Outlook

The discussion of stabilizer codes, and the surface code in particular, is much richer
than this summary. We are intentionally leaving out important topics – e.g., what
happens when we “lose” a qubit, or when errors move, or how many errors a code
supports, or how large are the overheads, etc. – to simplify the conversation and
provide you with some qualitative ideas that drive the architectural decisions on
superconducting quantum computing platforms.
One first takeaway message is the fact that there exists a robust formalism for
encoding logical qubits in the quantum states of extended lattices, which are simple
enough to be implemented with superconducting circuits. The second takeaway
is that this encoding provides us with error-correction and fault-tolerant operation
schemes, to store and process quantum information for arbitrarily long periods of
time, provided that error rates are low enough. The key concept here is the term low
enough, which translates into existence of one or more thresholds for fault-tolerant
operation, as mentioned before.

14 In the process, a stabilizer that contained three qubits may pass to contain four, as the lattice has one more
edge.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.5 Fault-Tolerant Quantum Computers 231

101

1 − F (%)
100
tgate (ns)

102

101
2008 2010 2012 2014 2016 2018 2020
Year

Figure 8.5 (a) Two-qubit gate errors and (b) two-qubit gate speeds for transmon
qubits. The type of marker denotes the type of gate: CZ (square), CNOT (dot),
cross-resonance (circle) (Chow et al., 2011), iSWAP (star).

There is a growing literature on the conditions for fault-tolerant computation on

different quantum technologies. In particular, the surface code has been subject
to extensive analysis. Some optimistic results by Fowler et al. (2009) establish
an error threshold of 1% on all operations – single- and two-qubit fidelity and
measurements – for building arbitrarily large codes. Other works set the boundary at
even higher levels, when errors are not generic but are biased toward certain models.
One could argue that we are very close to those thresholds. Single-qubit gates
have very good fidelities, with errors below 0.1% in most works. We also know
how to improve state preparation using active controls. Measurement errors are still
above desired thresholds, but can be improved using Purcell filters and parametric
amplification (Walter et al., 2017). Regarding two-qubit gates, Figure 8.5 shows
a collection of data points from over 30 works with transmon qubits. There are
already several realizations of fast CZ and iSWAP two-qubit gates with less than
1% error (Barends et al., 2014; Arute et al., 2019). The overall tendency is also
favorable, with a 20% decrease in two-qubit error gate every year15 over all gates,

15 Note that qubit coherence already follows a Moore’s type law – with an 80% increase in coherence time every
year (cf. Section 6.6) – this naturally leads to a decrease in gate errors. This said, gate operations are more
complicated than just “waiting,” as they are also sensitive to the quality of the controls, distortions in the
cables that bring the control signals, and many other technical sources of errors that need to be overcome.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 232 Quantum Computing

and this optimistic message is reinforced by recent milestones in error correction

with larger numbers of qubits and operations (Chen et al., 2021).
It therefore seems that fault-tolerant limits are not out of reach, and, just like
trapped ions have done, superconductors may get over the threshold in a few years.
However, as we have seen, this is just the beginning. In order to implement fault-
tolerant quantum computation, we need larger lattices of qubits capable of accom-
modating the qubit overhead that is required for errors to be detected and cor-
rected. This means not only larger quantum computers, but orders of magnitude
increases in the number of controls, filters, and measurement devices, as well as
the hardware to instrument all protocols. It is by no means a simple “upgrade”
from existing technology, but a roadmap that will take quite some time to com-
plete – unless we have comparable breakthroughs in error correction and quantum
computing theory!

8.6 Near-Term Intermediate Scale Quantum Computers

8.6.1 What Is NISQ?
This book has been written in what Preskill (2018) called the noisy interme-diate-
scale quantum (NISQ) era. There exist already working quantum computers ranging
from a few to 50-something qubits. Some of them are being developed in public
laboratories (Rol et al., 2019; Krinner et al., 2020), while others are being developed
in big companies – Google (Arute et al., 2019), IBM (Cross et al., 2019) – and
startups – e.g., Rigetti (Otterbach et al., 2017) – but none of these machines is
capable of scalable, error-corrected fault-tolerant computations.
This is an era in which the control of large quantum mechanical systems is
a reality, but it is far from perfect. State-of-the-art decoherence times, gate and
SPAM fidelities, and system sizes are not yet within the specifications for building
a robust and scalable quantum computer. It is also likely that we will be unable
to achieve scalable and fault-tolerant operation in the very near future, due to
technical demands in many aspects of superconducting and microwave technology.
These involve material science, quantum control, error correction, and electronic
engineering as some of the areas where there is immediate work to do, but it is
also safe to assume that we need important breakthroughs in other areas to move
forward.
In this interim time, we are stuck with devices that lack error correction, but
are steadily growing in size and quality by contrasted metrics. These devices are
by no means “useless.” They are covering important milestones in the path toward
scalable quantum computation. They are also a platform on which we steadily learn
important lessons on what works and what does not – even when these challenges
become more and more difficult over time!

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.6 Near-Term Intermediate Scale Quantum Computers 233

For instance, we already have first evidence of quantum devices and experiments
that can beat what can be simulated with classical computers (Arute et al., 2019)
while ensuring full control of the experimentalist. In this rewarding time, we are see-
ing new developments in the field of quantum algorithms, adopting new paradigms
where the quantum computer works hand in hand with classical algorithms and
postprocessing techniques. This approach has illuminated new solutions to canoni-
cal (toy) problems in quantum chemistry (Kandala et al., 2019; Arute et al., 2020),
even in the presence of significant errors and decoherence! We expect that the same
platforms will continue to support scientists in studying such interesting concepts in
theoretical physics and quantum many-body physics as quantum transport, many-
body localization, and time crystals (Mi et al., 2022; Neill et al., 2021).

8.6.2 Hybrid Quantum Computers

From our first collective steps into the NISQ era, there are already some interesting
lessons learned. First, the circuit model for quantum computation has consolidated
as a useful abstraction for interfacing with the quantum computer. The qubit reset,
measurements, single-qubit operations, and universal two-qubit gate of each plat-
form become the new assembly language with which all computations can be con-
structed. This language is understood by the quantum computer’s firmware. This is
a set of classical computers, microwave generators, and FPGAs for data collection,
which converts those instructions into actual microwave pulses, interprets the sig-
nals from the quantum device as binary measurement outcomes, and performs other
low-level tasks, such as qubit calibration.
On top of this layer, we find higher-level interfaces, such as IBM Qiskit, Google’s
Cirq, Rigetti’s Pyquil, etc. These are classical frameworks, written in friendly
programming languages (Python, C#, etc.) or even in graphical interfaces (IBM’s
Quantum Experience16 ), with which we can write quantum algorithms and operate
the quantum computer. These frameworks do not need to be attached to the quantum
computing infrastructure. They may run separately, sending remote orders to one
or more quantum computers, in our more or less common assembly language, for
the actual firmware to process and execute.
Interestingly, this is not a limitation. We can run very low-level operations this
way – for instance, we can engineer a complete randomized benchmarking protocol
using entirely the high-level interface. We can also benefit from the possibilities
of combining quantum and classical computations, in a hybrid quantum-classical
computing model. This means relying good classical algorithms for certain tasks
(e.g., optimizers, compilers) that are easier for large silicon-based computers, and

16 https://quantum-computing.ibm.com/.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 234 Quantum Computing

delegating to the Hilbert space of the quantum computer other taks that might be too
large or too slow on an ordinary computer – for instance, creating and manipulating
the wavefunction of a complex material, or the complex multivariate probability
from a sophisticated risk model.

8.6.3 Quantum Volume

This level of abstraction means that computations may be, to some extent, agnostic
to the actual hardware on which it runs. There is a natural competition between
platforms – trapped ions, circuits, quantum dots, and photonic systems – for offering
the most powerful backend. In this scenario, an important question arises: How do
we compare those devices? A partial answer could be the one from Section 8.4,
which involves creating a map of the coherence times and gate fidelities of the
device.
Unfortunately, this approach is insufficient, because it does not take into account
that the same algorithm may be compiled very differently on different platforms.
For instance, to make gates between arbitrarily separated qubits may be more costly
on a one-dimensional register than on a 2D lattice. And different algorithms may
have different assembly representations depending on whether the computer uses
an iSWAP or a CNOT or can support a continuum of two-qubit gates!
A partial solution to this problem is the quantum volume, a metric introduced by
Bishop et al. (2017) and perfected by Cross et al. (2019). In the spirit of RMB, the
quantum volume studies random circuits, computing the largest integer N such that
we can execute quantum algorithms involving N qubits and N layers of generic two-
qubit gates17 and still have a decent fidelity. Once we have that Nmax , the quantum
volume is the size 2Nmax of the Hilbert space that can be encoded in it. As of this
writing, the largest quantum volume 32 for superconducting circuits indicates that
we can execute a random circuit using five qubits and five layers of one- or two-qubit
operations and be safe about the outcome.

8.7 Outlook
In this chapter, we have seen how quantum computers are built and operated. The
hardware is conceptually simple. It builds on the tools developed in Chapters 6

17 By generic, we mean any rotation in SU (4) covering all of the Hilbert space for two qubits. We also mean that
those gates can take place among any pair of qubits. In practice, the arbitrary separation means that quantum
gates must be decomposed – transpiled – into a much larger number of assembly gates (Cross et al., 2019).
This is really what makes the quantum volume a challenging metric.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 8.7 Outlook 235

and 7, with additional techniques for two-qubit quantum gates and for characterizing
the computer’s performance.
It is clear we need further developments in error correction and fault-tolerant
computation models. Existing paradigms impose significant overheads – number of
measurements, postprocessing of data, and rapid feedback controls – and demand
orders of magnitude more qubits than what we are likely to supply in the following
decade. We also need a better understanding of quantum circuits, algorithms, and
compilation strategies.
However, there are also many architectural, design, and theoretical challenges
open. We can highlight the following ones:
– We need better qubits, with longer coherence times. There are hints that this
can improve with new materials and the application of state-of-the-art surface
treatment and fabrication (Place et al., 2021). But maybe we also need different
types of qubits, with intrinsic topological protection!
– Complex algorithms and error correction require us to pack more qubits into the
chip. This is an important challenge, because the use of microwave technology
sets a lower bound on the size of qubits and control devices. Maybe the route
to success lays in going beyond existing sandwich architectures to fully three-
dimensional integrated devices (Brecht et al., 2017, 2016).
– If we increase the number of qubits means, we need to increase commensurately
the number of readout and control lines. This requires important developments in
the control and microwave electronics – which already has shifted to customized
FPGA circuitry (Arute et al., 2019) – and clever strategies to optimize the limited
number of microwave lines that can be inserted into the fridge. If at some point
even this becomes an unavoidable limitation, we might need to transition to
distributed quantum computers connected by quantum links, as explained in
Section 7.2.5.
– We need better controls and higher-quality gates. This challenge may be
addressed at the hardware level, looking for new interconnects between qubits,
producing different, more scalable or stronger couplings. Or it may be addressed
at a quantum control level, improving on the design of quantum gates, making
them more robust against external perturbations, or even against the obvious
distortions suffered by microwaves when they are injected into a chip.
These are very important and very difficult challenges. I hope that the information
in this book, combined with existing literature and your ingenuity, will stimulate you
to address them, joining many of the excellent teams working in the field worldwide,
or working to create your own!

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 236 Quantum Computing

Exercises
8.1 Prove that a Hadamard gate acting on each qubit creates a quantum superpo-
sition of all possible binary numbers in a quantum register:
N −1
2
⊗N 1
H |00 . . . 0 = |b . (8.27)
2N/2 b=0

Discuss how to implement this gate for N identical flux qubits.

8.2 Unitary equivalence of universal gates. Prove that the CZ gate and the
CNOT gates in Table 8.1 are equivalent up to local transformations. In
other words, there are single-qubit unitary operations Ui and Wi such that
(U1 ⊗ U2 )UCZ (W1 ⊗ W2 ) = UCNOT .
8.3 Cavity-mediated exchange interactions. Assume two qubits interacting with
the same cavity:
 i
H = σiz + gi (σi+ a + a † σi− ) + ωa † a. (8.28)
i
2

(1) Using second-order perturbation theory in the off-resonant limit,

|ω − i |  |gi |, obtain a dispersive coupling model. Show that this
Hamiltonian in the subspace of no photons is equivalent to
  i 2gi2
 
g1 g2 1 1

Heff = − σi −
z
+ σ1x σ2x . (8.29)
i
2 δi 2 δ1 δ2

(2) Solve the dynamics induced by the effective Hamiltonian exp(−iHeff t)

after a time t. In which limit does this operation become an iSWAP gate?
(3) Suppose that the cavity is in a thermal state with temperature T = 0.050
K, for ω ∼ 6 GHz. How does the excitation probability of the cavity affect
the two-qubit exchange operation? Estimate the infidelity of the gate.
8.4 Phase gate. Take the model for the two interacting transmons (8.7).
(1) Diagonalize the Hamiltonian for all values of J and ω2 and plot the
eigenenergies as a function of ω2 /ω1 ≥ 1.
(2) Show that there are two avoided energy level crossings at ω1 = ω2 and
ω1 + α = ω2 .
(3) Show that in the limit ω2  ω1 , the qubit basis is mapped to four eigen-
states of this problem. Which energy levels do those states follow as we
adiabatically change ω2 ?
(4) Assume that we can perform a perfect adiabatic passage, from a large ω2
down to resonance ω2 = ω1 + α, where each two-qubit state |s1,s2  is
mapped to an instantaneous eigenstate of H (ω2 ). Compute formally the
accumulated phases by all four two-qubit states along the round trip.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 Exercises 237

(5) Prove that it is possible to recreate any control-phase gate, with any angle
exp(iθ σ1z ⊗ σ2z ), assuming arbitrarily long adiabatic passages with no
decoherence. Hint: Show that for a large enough passage time T , there
is a nonzero interaction phase θ. Then prove that the value of θ can
be enlarged arbitrarily by deforming the adiabatic passage ω2 (t × T  /T ),
with T  > T .
(6) Find out the algorithm for the quantum Fourier transform. What phases
θ are required in that algorithm?
8.5 We want to simulate time-evolution with the Ising model, implementing the
unitary operation
U (t) = exp −itσ1z σ2z , (8.30)
but we only have a toolbox with control-NOTs and arbitrary single-qubit oper-
ations. Devise a protocol to approximate U (t) using such gates. Hint: Inves-
tigate what happens when you sandwich a local operation in between two
CNOTs.
8.6 Useful positive maps. We now study a set of very common positive maps.
(1) Completely depolarizing channel. Provide a physical interpretation for
the completely depolarizing channel in Table 8.2. Show that the symmetric
version is recovered from the one in that table by setting px = py = pz =
p/3.
(2) Random flip. Write down a completely positive map to describe this
physical operation: With probability p, preserve the state and elsewhere
flip the state of the qubit, exchanging |0 and |1.
(3) Dephasing. Show that the dephasing of a qubit (6.14) may be written as
a channel that with probability p preserves the state, and elsewhere keeps
only the diagonal elements:

1−p  
ε(ρ) = pρ + tr(ρ)1 + (1 − p)tr(σ z ρ)σ z .
2
(4) In all previous maps, determine the Kraus operators and show that the map
preserves the trace of a density matrix: trε(ρ) = trρ.
8.7 Prove that the set of operators formed by all tensor products of Pauli matrices
Sα ∈ {1,σ x ,σ y ,σ z }⊗N (cf. Section 8.4.2) forms a basis of linearly indepen-
dent and orthogonal elements in which all matrices can be expanded. Hint:
Reinterpret matrices as column vectors, show that (A,B) = tr(A† B) is a good
scalar product for the matrix space, and prove the expansion (8.13).
8.8 Derive the Choi matrices (8.17) using the M and χ representations. Hint: Show

that |α = √1d di=1 |i ⊗ Sα |i forms an orthonormal basis.

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 238 Quantum Computing

8.9 Deduce the form of the matrices M and χ for a perfectly coherent positive
map, ε(ρ) = U ρU † , in terms of the unitary operation U . Particularize it for a
single-qubit rotation U = exp(−iθ n · σ ) and a CNOT.
8.10 Compute the process matrices M and χ for the single-qubit error map from
(6.19). Compute the average fidelity and the process fidelity for this map.
How do those figures relate to T1 and T2∗ ?
8.11 Design a randomized benchmarking protocol for a single qubit. What gates
do you need to use? Simulate this protocol for a qubit that experiences relax-
ation (T1 = 1 μs) and one that experiences dephasing (Tφ = 1 μs), and has
no SPAM errors. In the simulation, assume a fixed duration τ ∼ 10 ns for
all single-qubit gates, and that the error of the gate can be fully attributed to
decoherence, that is, εU (ρ) = εdecoh (U ρU † ). Fit the success probability Pl to
the exponential curve (8.23) and extract the parameter p. Compare the error
estimate p with the estimated average and process fidelity associated to εU (ρ).
8.12 The average fidelity (8.20) results from the fidelity of a physical process over
all pure states. It has been shown by Nielsen (2002) that this definition is
equivalent to replacing the map ε(ρ) with a totally depolarizing channel (see
Table 8.2) with depolarization p. Show that the relation between p and Favg
is given by (8.24).

https://doi.org/10.1017/9781316779460.008 Published online by Cambridge University Press

 9
Adiabatic Quantum Computing

In Chapter 8, we introduced the quantum circuit model for quantum computing, in

which we decompose algorithms into elementary operations – resets, gates, mea-
surements – that can be mapped to local controls and qubit–qubit interactions. In this
chapter, we introduce a very different computation model, based on the simulation
of complex many-body Hamiltonians and the preparation of ground states. This is
an extremely appealing idea, which connects the worlds of quantum simulation and
quantum computation and which has been rediscovered many times. A particularly
relevant work by Farhi et al. (2000) gave it the name of adiabatic quantum compu-
tation, because of the particular choice for ground state preparation. This work had
a tremendous impact in the field and even inspired the creation of large quantum
devices with thousands of qubits, D-Wave’s quantum annealer. We will discuss
the ideas in this work, together with the computational power of the Hamiltonian
model, and also get the differences between ideal adiabatic quantum computers and
the NISQ quantum annealers that are available in the lab. As in the previous chapter,
we try to stay as focused on the superconducting circuit hardware as possible, but
there are unavoidable theoretical concepts that must be well explained in order to
motivate and understand why and how experiments work.

9.1 Adiabatic Evolution

In most of this book, we have studied fast controls and the real-time dynamics they
induced in qubits and cavities. The motivation was, of course, to make quantum
operations as fast as possible, minimizing the role of decoherence. In this chapter,
we switch gears and focus mainly on slow control and adiabatic passages that are
used to prepare low-temperature states.
Imagine, as sketched in Figure 9.1a, a complex physical system with a
Hamiltonian that can be manipulated. The control, denoted ε, represents some
generic parameters of the Hamiltonian that we can change in time, so that

239

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 240 Adiabatic Quantum Computing

Figure 9.1 (a) Sketch of the energies and gaps throughout an adiabatic passage
in which a Hamiltonian Ĥ () is deformed from (t = 0) to time (t = T ). The
Hamiltonian maintains a gap between the excited states and the ground state energy
EGS , which is at least Emin . (b) Qubit hyperbola for a flux qubit that begin
in its ground state and is subject to a linearly growing bias field, from a very
large negative value ε  |min | to a very large positive one. In this Landau–Zener
process, there is a minimum qubit gap min , depending on which the process will
be adiabatic – the qubit remains in the ground state manyfold and |1 is deformed to
|0 – or diabatic – there are transitions to the excited state, leaving some probability
in |1 at the end.

Ĥt = H (ε(t)). We start the adiabatic passage from an eigenstate of the Hamiltonian
Ĥ0 , which we assume is easy to prepare. We change the control in time ε(t),
expecting that our experiment remains as close as possible to the ground state of
the time-dependent Hamiltonian Ĥt .
Implicit in the notion of adiabatic passage is the idea that ε can change slow
enough for the system to adapt and remain in a quasi-equilibrium state without
exciting to higher-energy states. To pose some classical analogs, think of how you
can adiabatically stretch a rubber band. If you stretch it too fast, it will acquire too
much energy, exciting phonons that show as vibrations of the band. Or think about
how you can push a glass full of water across a table. If you push too fast, the surface
of the water will deform and even spill out of the glass, but if you push it gently, the
surface of the liquid will remain unperturbed.
We now address the idea of adiabatic passage from two points of view. We start
with a very specific and simple model, a qubit, where we can solve the problem
analytically and even provide a quantitative answer to the shape of the control. This
problem will give us an intuition about adiabatic and diabatic processes, which
can be applied to many different problems. This intuition then acquires a rigor-
ous mathematical framework through the adiabatic theorems, a set of sufficient
criteria to ensure an adiabatic passage for every given Hamiltonian. This theorem
lays the foundations for the adiabatic quantum computation (AQC) model that we
discuss next.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.1 Adiabatic Evolution 241

9.1.1 Landau–Zener and Qubit Adiabatic Control

The simplest model in which we can study an adiabatic passage is a qubit. Imagine
a flux qubit (cf. Section 6.4) with a tunneling amplitude h̄min , controlled by an
external bias field h̄ε. The Hamiltonian for this qubit adopts the familiar form (6.2),
which we change basis to become
h̄ε z h̄min x
Ĥ (ε) = σ + σ . (9.1)
2 2
We can compute the ground states of the qubit for any value of ε, using the Pauli
matrix decomposition:
⎧ 
1 ⎨ = ε2 + 2 ≥ | |,
min
Ĥ (ε) = h̄(ε) n · σ, with min
(9.2)
2 ⎩n = 1 ( ,0,ε).
 min

This model is equivalent to a magnetic field B = (ε)n interacting with a quantum

spin S = 12 σ .
The ground state eigenenergy − 12 h̄(ε) is separated from the excited state by a
gap h̄(ε). The ground state is a pure state with a Bloch vector anti-aligned with
the magnetic field σ  = −n. It has a wavefunction
   
θ (ε) θ (ε)
|0̃ε  = cos |0 + sin |1 , (9.3)
2 2

that depends on the polar angle n = (sin θ,0, cos θ ). In the limit of large bias
|ε|  |, the ground state of the Hamiltonian approximates the two current
states, (0,0,+1) or |1 = |R, for very negative field and (0,0, − 1) or |0 = |L
for very large positive bias. These are two states that we can prepare, either by
waiting for the qubit to relax to the ground state at those extreme points, or using
some type of measure-and-reset protocol — e.g., we could dispersively measure
the qubit’s state and flip it using a microwave if the qubit was in the excited
current state.
We now wonder what happens if we start from such a well-prepared ground state
and change the control parameter ε, as in Figure 9.1b. Does the qubit remain in the
instantaneous ground state or does it get excited? In the first case, which we call the
adiabatic evolution, the qubit’s wavefunction is always the one in (9.3) up to global
phases. In the second case, which is called the diabatic evolution, the qubit will not
adapt, and it will either get locked in the original current state or, more generally,
become a quantum superposition of both qubit states.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 242 Adiabatic Quantum Computing

To answer this question, we assume that ε(t) is a monotonously growing func-

tion, and write the wavefunction on a rotating frame that brings the instantaneous
eigenstates of Ĥ (ε) to the qubit basis:
1 y
|ψ(t) = e 2 iθ(ε(t))σ |ξ (t) . (9.4)
If we start at a large negative value of ε and we are perfectly adiabatic, |ξ (t)  |0,
and the probability of diabatic transitions Pdia = |1|ξ (t)|2 becomes approximately
zero.1 We write the dynamics of |ξ (t) in a rotating frame, using ε instead of time
as coordinate (see Problem 9.1):
1 min y  !
i∂ε |ξ (ε) = (ε)σ z − 2
σ ξ (ε) . (9.5)
2ε̇ 2(ε)
This Hamiltonian has contributions that describe the adiabatic (σ z ) and countera-
diabatic (σ y ) evolution. In the perfectly adiabatic limit, in which the rate of change
ε̇ → 0 is very slow, the adiabatic term dominates the Hamiltonian, and the wave-
function remains an eigenstate up to a dynamical phase |ξ (t) = eiφ(t) |0, with
φ(t) = (ε)dε/ε̇. In this limit, the counteradiabatic term is a weak perturbation
that changes more slowly than 4/min , and is therefore incapable of bridging the
gap /ε̇. We can estimate the small corrections introduced by this term as being of
order ε̇min /3 < ε̇/2min .
When the speed of change ε̇ is constant, the two-level system admits an analyt-
ical solution called the Landau–Zener formula (Landau, 1932; Zener, 1932). This
formula describes a limit in which we start from the instantaneous ground state
|ψ(−∞) = |1, and sweep the control all the way to ε → +∞. In this problem,
the probability of diabatic excitation is
   
π 2min π
PLZ = exp − = exp − . (9.6)
2 ε̇ 2v̄τadiab
This expression contains two timescales: the rate of change of the Hamiltonian’s
gap v̄ = ε̇/min , and the smallest energy scale for the qubit along the passage,
min = 1/τadiab . When the rate of change v̄ is faster or comparable to the minimum
frequency of the qubit, we are in the diabatic regime, in which PLZ approaches 1.
In the opposite regime, when the rate of change v̄ is slow enough, the qubit is in the
adiabatic regime and the excitation is negligible PLZ = 0.

1 Note that, due to our unitary transformation, |0 and |1 are no longer current states, but an instantaneous
ground state and excited state.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.1 Adiabatic Evolution 243

9.1.2 The Adiabatic Theorem

There are many rigorous generalizations of the Landau–Zener problem, all of which
receive the name adiabatic theorem. They analyze the adiabatic deformation of
eigenstates in a tunable Hamiltonian, assuming that a relevant subspace2 is well
separated from the rest of the spectrum by an energy gap. The theorems provide suf-
ficient conditions for an initial state that begins in this subspace, to remain inside it,
as it is continuously deformed. These theorems also bound the error of the adiabatic
passage – the probability of leaking to other states – by relating the rate of change of
the control and the instantaneous gaps of the Hamiltonian, similar to Section 9.1.1.
The study of adiabatic passages is rather old, but it starts to acquire rigorous
foundations around year 1950, when Kato (1950) derived the first conditions for
the adiabatic deformation of a ground state. Later works have made these results
more useful and rigorous using spectral theory. A full demonstration of any of the
adiabatic theorems falls outside the scope of this book,3 but most versions (Kato,
1950; Galindo and Pascual, 1991; Jansen et al., 2007; Lidar et al., 2009) assume
similar conditions on the tunable Hamiltonian H (ε) and the initial state:
(1) The adiabatic passage happens by deforming a Hamiltonian Ĥ (ε) through a
control ε(t) that takes a total time T .
(2) The Hamiltonian H (ε) and the parameter tuning ε(t) must be differentiable a
number of times.
(3) The spectrum of the deformed Hamiltonian is gapped and separates into two
sectors, σ (Ĥ (ε)) = σground (ε) ∪ σexcited (ε), such that
– it has a ground state manifold with eigenvalues in a set σground ,
– it has a set of excited states, with eigenvalues σexcited ,
– the minimum distance between these sets is h̄min .
(4) In many proofs, the Hamiltonian spectrum is upper-bounded, that is,
Emax := maxε {max σ (Ĥ (ε))} < +∞.
(5) The evolution begins on an eigenstate from the ground state manifold, with
eigenvalue σground .
The conditions are a bit hard to read, but we can map them to our qubit example.
The deformed Hamiltonian Ĥ (ε) was given by (9.1), and it had one ground state

2 The theorems work for one state, or for a collective of states that are very close in energy, degenerate, or
almost degenerate. We do not care much what happens within this collective of states, but we want to preserve
all probability within this subspace.
3 This said, we encourage the reader to have a look at the cited works, which are long, but relatively easy to
follow.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 244 Adiabatic Quantum Computing

σground = {− 12 h̄(ε)} and one excited state σexcited = {+ 12 h̄(ε)}, separated by

a gap h̄min . The Hamiltonian was bounded, Emax = maxε 12 h̄(ε). We used a
linearly growing control, which was therefore differentiable ε̇ = v̄min .
Using these conditions, the adiabatic theorems bound the diabatic excitation or
error probability4 as
 α 2β
1 1 Emax
Pexcitation ≤ × × . (9.7)
T min (h̄min )α
The exponents α,β are integers and depend on the particular proof and assumptions.
Traditional demonstrations (Kato, 1950; Galindo and Pascual, 1991; Teufel, 2003)
use α = 0 and β = 1 and require only one-time differentiability. In these versions,
the maximum energy drops out of the equation and we recover an LZ-like formula:
 2
1
Pexcitation ≤ O . (9.8)
T min
If we increase the differentiability to k ≥ 2, Jansen et al. (2007) derives a slightly
better bound where the error still decays algebraically, but the power has improved:
Pexcitation  O(1/T 2 )2 . Finally, Lidar et al. (2009) showed that sufficient control
over the Hamiltonian, introducing n + 1-times differentiability, allowed improving
the exponent arbitrarily, to β ∼ 1/n. This implies that the adiabatic error Pexcitation
can be made smaller and smaller without significantly changing the total adia-
batic time.
The adiabatic theorem is a powerful tool which gives full rigor to the intuitions
developed in the study of the Landau–Zener processes, and extends it to arbitrary
models, including large setups with thousands of qubits and resonators. This gen-
erality implies two limitations: First, the argument is perturbative, not a closed
formula like (9.6); second, the theorems rely on global quantities, such as the total
adiabatic passage time T and minimum gaps, without regard for the controls. Those
limitations are inconsequential, because in many circumstances we do not have
access to a detailed knowledge of the spectrum. Instead, we work with bounds for
h̄min or Emax , and seek a global scaling of the passage time T to ensure adiabaticity.

9.1.3 Circuit-QED Applications of Adiabatic Theorem

The Landau–Zener processes and the adiabatic theorem give us a fantastic insight
into the manipulation of quantum systems, illustrating how the energy separation
between eigenstates determines the regimes of adiabatic and diabatic evolution.
We can use this to better design the circuit-QED and quantum computing exper-
iments from earlier chapters.

4 Some works study the norm-2 of the error, |δ = |ψ

ideal  − |ψ(T ). This is related to the error probability
1/2
!δ! = δ|δ1/2 = Pexcitation and affects the powers shown in (9.7).

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.2 Adiabatic Quantum Computing Model 245

As an example, imagine we wish to bring a qubit in resonance with a cavity, to

make them interact with the JC Hamiltonian for a brief period of time. We want
this change to be “abrupt” or diabatic. We therefore need to compare the minimum
energy splitting in the qubit-cavity Hamiltonian, given by the Rabi splitting 2|g|,
with the rate of change in the qubit’s gap. For the control to be diabatic, we find
˙
/g > g. Moreover, if we use a transmon, a parametric change of  = ω01 intro-
duces some perturbation into the qubit’s Hamiltonian ∝ ω̇01 . The analysis of this
perturbation gives us another LZ-like condition, ω̇01  ω01 2 2
,ω12 .
Alternatively, the LZ technique also gives us insight in how to adiabatically con-
nect states of a model for different values of a control parameter, to ensure smooth
adiabatic transitions. This idea can be used to create single-photon wavepackets,
starting from excitations in a qubit. This experiments relies on the avoided level
crossing between an excited qubit |e,0 and a cavity with one photon |g,1. The
protocol begins with a qubit that is highly detuned from the cavity  < ω, which
is adiabatically brought into resonance with the cavity and beyond, until  > ω.
If  changes at a rate slower than the qubit–cavity coupling,  ˙  g 2 , this process
preserves the state |g,0 and maps adiabatically |e,0 → |g,1. If we calibrate the
dynamical phases properly, we can even use this protocol to transfer any quantum
superposition from the qubit onto the cavity (|g+β |e)⊗|0 → |g (α |0+β |1).
Finally, Zener (1932) showed that the Landau–Zenner formula is an upper bound
for the excitation probability, and that other bounds are derived from the instanta-
neous speed ε̇(t). This idea has inspired the fast quasi-adiabatic passage (FAQUAD)
method, which advocates fixing the rate of change in the Hamiltonian to be propor-
tional to the minimum energy gap in the controlled quantum system ε̇ ∝ (ε). This
technique enables faster controls by speeding up the dynamics in regions where the
gap to excited states is still large – e.g., large |ε| in Figure 9.1b. One application
discussed by García-Ripoll et al. (2020) is to use these methods for accelerating the
transmon-based two-qubit CZ gates from Section 8.3.3.

9.2 Adiabatic Quantum Computing Model

The adiabatic quantum computation model puts together two ideas:
– All quantum computations may be recasted as finding the ground state of a
gapped5 problem Hamiltonian Ĥp .
– This ground state can be prepared by adiabatic evolution with a quantum simu-
lator of the Hamiltonian Ĥp or a deformation thereof.

5 Here “gapped” means that there is an energy gap between the ground state manyfold that encodes the
computation and all other states.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 246 Adiabatic Quantum Computing

The first connection was formalized by different groups, starting from Kitaev et al.
(2002) until the more recent work by Biamonte and Love (2008), which creates
apparently simple models for arbitrary quantum circuits. We will explore this
mapping in the context of Section 9.3, where we worry about which Hamiltonians
represent useful and hard computations, and which are physically implementable.
But first, let us explore how realistic it is to satisfy the second requirement, which
is building the ground state of Ĥp .

9.2.1 The Adiabatic Quantum Computing Algorithm

Preparing the ground state of a generic complex quantum system is a difficult task.
This is true, even if the simulated Hamiltonian is gapped and we can put the simula-
tor in contact with a bath at a temperature so low, kB T  h̄min , that the Boltzmann
distribution predicts a macroscopic occupation of the ground state. First of all, the
contact with the bath can change the eigenstates of a quantum system, causing it
to thermalize into states that are different from the desired ones. Second, even if
this incoherent coupling does not change the system, we cannot guarantee that all
degrees of freedom will interact similarly with the bath. There may be uncountable
properties and symmetries that have a vanishing coupling to the thermal environ-
ment and cool at impossibly slow rates.
The adiabatic quantum computation algorithm was introduced by Farhi et al.
(2000) as a coherent protocol to create the ground state of complex Hamiltonians,
using a carefully engineered adiabatic passage. The AQC algorithm works with a
gapped problem Hamiltonian Ĥp , whose ground state (or ground states) |ψp  we
wish to prepare. It assumes the following:
(1) There exists a Hamiltonian H0 with an easy to prepare ground state |ψ0 .
(2) The interpolating Hamiltonian

Ĥ (ε) = Ĥ0 (1 − ε) + Ĥp ε (9.9)

can be implemented in some hardware – what we call a quantum simulator –

for any values of the parameter ε ∈ [0,1].
(3) The interpolating Hamiltonian has an energy gap that never drops to zero,
(ε) = 0 (cf. Figure 9.1).
The adiabatic algorithm for preparing |ψp  consists of the following steps:
(1) Initialize the quantum simulator with interactions Ĥ (0) = Ĥ0 and prepare its
ground state |(0) = |ψ0 .
(2) Tune ε(t) from ε(0) = 0 to ε(T ) = 1 at a speed that satisfies the adiabaticity
condition:

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.2 Adiabatic Quantum Computing Model 247

1 d min
(εt )  . (9.10)
min dt h̄
(3) Measure and characterize the final state |(T )  |ψp .
The adiabatic algorithm works by smoothly distorting a state that is easy to
prepare, typically an uncorrelated product state, through an adiabatic passage
that is guaranteed to succeed by the adiabatic theorem. Unlike other quantum
computing paradigms, such as the quantum circuits from Chapter 8, the construction
of |ψp  is a continuous process, not a combination of simple, well-controlled
operations.
One important question in the AQC algorithm is the choice of initial Hamiltonian
H0 . Ideally, we want a model whose ground state |ψ0  has some overlap with the
final solution, because if we do not achieve this, the adiabatic passage might have a
vanishingly small gap. A common choice is to prepare a product state that contains
a quantum superposition of all states in the computational basis:
 ⊗N
⊗N 1 1 1 
|(0) = |+ = √ |0 + √ |1 = N/2 |ss| . (9.11)
2 2 2 s

This state can be prepared as the ground state of a noninteracting model, a collection
of transverse fields:

N
H0 = − σmx . (9.12)
m=1

The adiabatic protocol works by interpolating between both limits, starting with
large external fields and small interactions, and gradually increasing the strength of
the couplings in Hp .

9.2.2 Resource Accounting

We need a careful and fair understanding of the resources needed in a quantum com-
putation to determine whether the problem can be solved with existing resources,
and whether the algorithm and the implementation are competitive with classical
solutions.
In the quantum circuit model, the resources are size and time, quantified by the
number of qubits and the number of elementary operations (gates and measure-
ments). If these resources grow favorably with the problem size, as it does for Shor’s
factoring algorithm, we consider that the quantum circuit is a good solution, and we
talk about a quantum advantage.
In AQC, the accounting of resources also involves size and time, given respec-
tively by the size of the quantum simulator – e.g., number of qubits required to

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 248 Adiabatic Quantum Computing

implement H (ε) – and the duration of the adiabatic passage. In AQC, the size
of the simulator grows gently with the problem size. Time, on the other hand, is
usually quantified as the time to solution, the duration of an adiabatic passage that
guarantees we reach the ground state with a minimum probability p.
As discussed earlier and in the context of the Landau–Zener processes, the time
to solution depends on the minimum energy gap of the problem Hamiltonian. This
gap typically shrinks with growing problem size, with different scalings depending
on the problem’s difficulty – exponential T ∼ e−αN for difficult NP problems,
algebraically for “manageable” problems T ∼ N −α , as discussed in Section 9.3.4.
The relation between time and gap shows how this resource can be distorted by
our choice of Hamiltonian. Imagine we can guarantee an adiabatic passage in time T
for a problem Hamiltonian Ĥ with gap min . If we can implement a simulator for
the rescaled Hamiltonian Ĥ  = 2Ĥ , 3Ĥ , or even larger, we will arbitrarily increase
the gap and arbitrarily contract the time to solution.
In practice, the strength of interactions available in the lab is finite and the total
available energy Emax = maxε !H (ε)! = maxε {max σ (H (ε))}, is bounded. There-
fore, when comparing different adiabatic quantum algorithms it is often useful to
work with a rescaled dimensionless time τ := T Emax /h̄, which is free from this
uncertainty and scales more adequately with the problem size N . The different
accounting in time and size will pop up when we compare AQC’s with universal
gate-based quantum computers, when we discuss the difficulty of adiabatic compu-
tations, and in Section 9.4.3, where we discuss this paradigm critically.

9.3 The Choice of Hamiltonian

The adiabatic algorithm is a global optimizer for problem Hamiltonians. It can
therefore solve arbitrary optimization problems sopt = argminF (s) if we encode
the cost function into the problem Hamiltonian:

Ĥp = F (s) |ss| . (9.13)
s

This suggests that we can explore the relation between problem Hamiltonians and
families of computational problems F (s) with different complexity. We will con-
firm this idea, deriving Hamiltonians for difficult families of classical and quantum
problems, called NP-hard and QMA-complete. But first, we need to know these
classes.

9.3.1 A Primer on Complexity Classes

In order to make this discussion a bit rigorous, we have to introduce some notions
of computational complexity classes. Do not worry! There’s not much that needs

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.3 The Choice of Hamiltonian 249

to be understood to get a grasp of what follows. We begin in the classical world,

using the paradigm of Turing machines or universal classical computers, and define
(qualitatively) the following three relevant classes:
P This class contains all problems that can be solved with polynomial resources
(time and space) by a classical computer.
NP This is a larger class of problems whose solution can be verified on a classical
computer with polynomial resources. Given a solution to an NP problem, testing
whether it is correct is a task in P.
NP-complete This subset gathers universal problems in NP. That is, any problem
in NP can be translated to or solved with an NP-complete problem. Note that this
mapping is efficient (it takes polynomial resoures), but solving the NP-complete
problem is not!
NP-hard This set contains all problems that can be used to solve NP-complete
problems, but which may or may not be so easily verifiable.
You already know problems that are in P. For instance, arithmetic problems or
checking whether a number is prime are tasks that can be done with polynomial
resources in simple computers. You may also know some NP-complete problems,
such as the map coloring problem or the graph isomorphism problem – finding
whether two graphs are identical and can be mapped one to another by just relabeling
the vertices. Finally, there are NP-hard problems, such as the traveling salesperson
problem – finding the optimal route through all the cities in a graph – to which P
and NP-complete problems can be recast, but which cannot be verified so easily.6
A similar hierarchy of classes has been developed for problems that can be
addressed in a quantum computer:
BQP The Bounded Quantum Polynomial class contains all problems that can be
solved with polynomial resources (time and space) on a universal quantum com-
puter, with a fixed success probability p > 1/2.
QMA Quantum Merlin–Arthur is the equivalent of NP. It includes all problems
whose solution can be verified by a universal quantum computer with fixed suc-
cess probability, p > 1/2.
QMA-complete As with NP-complete, this class gathers all the QMA problems
that can be used to solve any other problem in QMA, with at most a polynomial
overhead in resources due to the translation.
Note how the BQP class is the quantum analog of P and contains all the quantum
circuits that we can write with moderate resources and run in polynomial time. Some

6 Indeed, how do you know if the route is optimal? Note, however, that answering the question, “is there a route
of length L?” is NP-complete.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 250 Adiabatic Quantum Computing

well-known examples are the quantum Fourier transform and Shor’s factoring algo-
rithm. From the definitions, using the fact that a quantum computer can reproduce
any classical computation, we can infer the following relations:
P ⊂ NP ⊂ QMA, and P ⊂ BQP ⊂ QMA. (9.14)
We believe that BQP contains problems that are not in P. There is evidence of this
in the difficulty of simulating the evolution of quantum many-body systems, or even
relatively “small” random quantum circuits, where entanglement grows quickly. It
is also widely believed in the academic world7 that NP is not contained in BQP. This
means that quantum computers cannot exponentially speed up the solution of those
problems, even though one may not rule out other less dramatic accelerations.
In the following two subsections, we will show that both NP and QMA problems
can be restated as the task of preparing the ground state of a Hamiltonian. We will
identify specific representatives of those Hamiltonians that are universal, in the
sense that they are NP- or QMA-complete and can be used to solve any problem
in those classes. Interestingly, both will turn out to be models of interacting qubits
that are good candidates to be implemented in the laboratory.

9.3.2 QUBO and NP-Complete Hamiltonian Problems

The spin glass is a deceivingly simple NP-hard problem that we find in the world
of condensed-matter physics:
 
HpNP = Ji,j σiz σjz + hj σjz . (9.15)
i,j j

Barahona (1982) showed that computing the ground state of this model is at least
an NP-complete problem, provided the spins are arranged on a planar graph,8 inter-
actions are antiferromagnetic along the edges that connect qubits Jij = J > 0, and
we can control the sign of the magnetic fields, which have comparable strength to
the interactions |hj | ∼ J .
The spin glass in (9.15) is a diagonal Hamiltonian (9.13). Finding the ground state
of this problem is thus equivalent to finding the minimum of a classical function. If
we move from sign variables si ∈ {−1, + 1}, to boolean variables xi = 12 (si + 1) ∈
{0,1}, the spin glass become an instance of a quadratic binary optimization (QUBO)
problem:

F (x) = xi Qij xj , xi ∈ {0,1}. (9.16)
ij

This is a family of NP-hard problems, to which many other NP problems can be

recast.

7 But apparently not so much in the quantum investment world.

8 A planar graph is one that can be drawn on a plane without edges crossing each other.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.3 The Choice of Hamiltonian 251

We call embedding the transformation of a QUBO problem back to a spin-

glass (9.15) simulator, for instance, to solve the problem using AQC. As you may
have guessed, the Q matrix of general QUBO problems can be more complex than
a planar graph. In this case, the problem is typically mapped in an approximate
way, identifying some variables xi directly to spins, and introducing many other
spins and interactions to reproduce the desired energy landscape.
The spin-glass model (9.15) belongs to the kind of Hamiltonians that can be
simulated using superconducting qubits and resonators. Indeed, as explained in
Section 6.5, the qubit–qubit dipolar interactions and the qubit’s external controls
may, up to suitable rescalings, imitate any kind of problem of the type (9.15). In
Section 9.4, we will discuss the D-Wave quantum annealer, a quantum simulator
for Ising-type spin glasses that follows these ideas, adding ingredients such as long-
distance tunable couplers and large-scale integration.

9.3.3 QMA-Complete Problems

We have mentioned before that AQC and the quantum circuit model are equivalent.
This idea was put forward by Kitaev et al. (2002), engineering a Hamiltonian that
encodes the outcome of a quantum computation with N qubits and M gates, using
an adiabatic quantum computer with N × M qubits. Several works after Kitaev’s
continued the assimilation of the QMA class into Hamiltonians with simpler inter-
actions and lower complexity. Finally, Biamonte and Love (2008) demonstrated that
already an apparently simple two-body Hamiltonian with XX and ZZ interactions
is QMA-complete:
  
ĤpQMA = hi σiz + i σix + Jij σiz σjz + tij σix σjx . (9.17)
i i,j

The local controls and couplings in ĤpQMA can be derived constructively from
a quantum circuit. Ideally, we could simulate the XXZZ Hamiltonian in our setup,
prepare the ground state using the AQC algorithm, measure its energy – this amounts
to measuring Pauli operators, as shown in Problem 9.5 – and verify that energy lies
below a prescribed bound, to determine the success of the computation. Moreover,
using the same setup, we could also address other NP-complete problems.9
In practice, this introduces two technical difficulties. The first one is to ensure a
topology of qubits that maps to the problems we wish to solve, or that at least can
provide suitable embeddings. The second problem is how to simulate the XXZZ
interactions, when the qubits that we have studied only seemed to have one type of
dipolar moment.
One solution to this problem is to use flux qubits, and engineer them to enable
both inductive and capacitive interactions. If we use current superposition as basis

9 Note how (9.17) contains the NP-hard spin-glass model (9.15) in the limit  = t = 0.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 252 Adiabatic Quantum Computing

states, the magnetic and capacitive interactions create σ x σ x and σ y σ y couplings,

respectively. Recent experiments by Ozfidan et al. (2020) have demonstrated simul-
taneous interactions in rf-SQUID qubits, which are corroborated by numerical simu-
lations of the same setup (Consani and Warburton, 2020). However, it is still unclear
to what extent we can control the J and t couplings individually, or whether, as
our work suggests (Hita-Pérez et al., 2022), they appear in combination with other
couplings, such as σ z σ z .

9.3.4 Scaling of Resources

The analysis of time as a resource is a very complicated and controversial matter.
Ideally, the adiabatic theorems provide a sufficient condition for the passage to suc-
ceed. Equations (9.7), (9.8), or (9.10) give a heuristic scaling of T based on global
quantities, such as the minimum gap min . In practice, we rarely know min , but we
may calibrate the difficulty and the scaling of time based on a coarse understanding
of the interpolating Hamiltonian.
A central idea is that the models Ĥ0 and Ĥp are so different that the interpolating
Hamiltonian Ĥ (ε) must experience a phase transition as we increase the parameter
ε from 0 to 1. For infinitely large systems, the energy gap strictly vanishes at the
critical point where we transition between ground states of a different nature. How-
ever, for finite systems such as the ones we study in the quantum simulator, we will
experience some type of finite-size scaling min = f (N ).
A lucky situation would be that, when interpolating between (9.12) and (9.15),
the lattice of qubits experiences only a second-order phase transition, in which the
gap shrinks algebraically with the problem size min ∼ N −α . Those problems are
considered “easy” because the time to solution grows “slowly” with the problem
size, as T ∝ N α .
More generally, for NP-hard problems the spin glass Ĥp will exhibit a collection
of almost degenerate ground states that are macroscopically different from each
other. In those models, the interpolating Hamiltonian passes through a first-order
phase transition, where the gap vanishes exponentially fast min ∝ α −L and the
time-to-solution diverges exponentially T ∝ α L . This growth is consistent with the
hardness of the problem and the scaling of classical algorithms, and reveals that we
cannot obtain an exponential advantage by using the quantum algorithm.
This type of physics is so common that it can already be observed in simple one-

dimensional problems Hp = Ji σiz σi+1
z
. When the couplings are regular, such as
an all-ferromagnetic interaction Ji > 0, the interpolating Hamiltonian will cross
a second-order phase transition from para- to ferromagnetic ground states. At the
critical point on which the transition happens, the gap vanishes algebraically around
the transition point, min  N −ν , and the passage is experimentally “affordable.”

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 253

However, if we make the Ising interaction random,10 then suddenly the phase√tran-
sition from paramagnet to spin glass will scale much worse, min ∝ exp(−g N ).

9.4 D-Wave’s Quantum Annealer

In 1999, a small team of enterpreneurs and academic researchers founded a com-
pany to work on superconducting quantum technologies. This company is known
for one of the longest and more focused efforts toward building an adiabatic quan-
tum processor using superconducting circuits. In 2007, the firm made a modest
and rather controversial debut introducing a superconducting chip with 16 qubits.
After this demonstration and further improvements in the machine’s design, the
company produced the so-called D-Wave One, a larger processor with 128 qubits,
which was later followed by the larger D-Wave Two (2012) with 512 qubits, the D-
Wave 2X (2015) with about 1 152 functional qubits, the D-Wave 2000Q (2017) with
2 048 qubits, and a D-Wave Advantage, with 5 640 qubits, already disclosed as of
this writing.
In this section, we are going to discuss D-Wave’s older architecture. Follow-
ing the literature, we must call D-Wave’s devices quantum annealers and not an
adiabatic quantum computer, because they are not capable of universal quantum
computation – they are not QMA-complete. Instead, D-Wave’s efforts focus on
NP-hard QUBO optimization problems, implementing powerful quantum simula-
tors for Ising models with transverse magnetic fields:
  
H (t) = i (t)σix + Jij (t)σiz σjz + hi (t)σiz, (9.18)
i ij I

Here, you recognize the spin-glass model (9.15) in combination with the initial
bias field H0 ∝ (t) from which we start the adiabatic passage (9.12). D-Wave’s
machines can be programmed to change in time all parameters from the effective
Hamiltonian (9.18), allowing for a good dynamical range and determination of the
sign in all variables. This enables the users to implement an adiabatic passage in
which (t) ∝ (1 − ε(t)) is decreased, while the problem Hamiltonian is enhanced
J (t),h(t) ∝ ε(t).
The quantum simulation, using superconducting devices, of an NP-hard Hamilto-
nian with time-dependent quantum fluctuations (9.18) is a great scientific achieve-
ment and a great technological challenge worth studying. We will discuss the
devices’ architecture, how the annealer works in practice, its performance in
comparison to classical algorithms, and how “quantum” the D-Wave machine
really seems to be.

10 See section 4.1.2 in Bapst et al. (2013) for an overview.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 254 Adiabatic Quantum Computing

(a) (c)

(b) (d)
3

1.5

–1.5

–3

-0.4 0 0.4

Figure 9.2 (a) Tunable coupling circuit for two flux qubits with an interac-
tion mediated by an rf-SQUID. (b) Experimentally measured mutual inductance
between the two flux qubits for this particular circuit. (c) Detail of a flux qubit
design, including elements to tune the qubit frequency (LT), tune the qubit gap
(CCJJ), induce an external field (IPC) and read-out the qubit state (RO). (d) Archi-
tectural design of a cell with eight qubits in the D-Wave chimera graph. All qubits
have local controls, tunable interactions and readout, and talk to each other through
the ICO couplers. Different cells talk to each other via the XCO couplers. Figure
adapted from Harris et al. (2009) and Harris et al. (2010b) with permissions.

9.4.1 D-Wave’s Architecture

D-Wave’s architecture is summarized in various panels from Figure 9.2. The simu-
lator is made of moderately good flux qubits, with controllable tunneling strength
and magnetic bias. Qubits interact with each other through a long-range inductive
coupler with tunable inductance to implement matrix Jij . The whole setup sits on top
of a sophisticated circuitry – partly superconducting, partly semiconducting digital
technology – that provides the dynamical controls of the local and interact on terms.
Let us go through the qubits and couplers in detail.

Highly Tunable Flux Qubits

The first ingredient in Hamiltonian (9.18) is a two-level system with adjustable
“magnetic fields” along the σ z and σ x direction. Until 2017, D-Wave relied on

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 255

slightly modified rf-SQUIDs to implement these two-level systems (Harris et al.,

2010a). Introduced in Section 6.4, rf-SQUID qubits consist of a single Josephson
junction and a large inductor. The qubit is crossed by a half-integer multiple of the
magnetic flux quantum ext = 12 0 , to induce phase frustration. The inductive
potential thus acquires a symmetric double-well structure (cf. Figure 6.6b), with
minima on two states with opposite current, |L and |R. We identify these as our
qubit states, which are coupled by tunneling i , as in Hamiltonian (9.18).
The choice of a flux qubit for an Ising-type quantum simulator is very natural
because of the following:

– We control the magnetic field ∼ σ z through the injected flux.

– We can control the tunneling amplitude ∼ σ x if we design the qubit with a tunable
Josephson junction, as explained later. This allows us to reduce the qubit’s gap
i at the end of the passage, something that would be hard for other qubits.
– Flux qubits have a very large anharmonicity and work very much like a two-level
system, without leakage to other states.
– Flux qubits interact inductively. As explained in Section 6.5.4, this type of inter-
action can be mediated and tuned with SQUIDs or similar devices.
– Because the anharmonicity can be very large, we can afford large qubit–qubit
dipolar interactions J σix σjx . This is required at the end of the phase transition,
where the problem Hamiltonian Ĥp dominates.

To implement all these features and simulate the quantum annealing Hamiltonian,
D-Wave has introduced changes in the qubit design (Harris et al., 2010a,b). The
changes, illustrated in Figure 9.2b, work as follows:
(1) The tunable Josephson junction of the rf-SQUID is implemented by a com-
plicated circuit, called a compound Josephson junction (CCJJ). Introduced in
Harris et al. (2010a), this circuit works like a dc-SQUID, creating a nonlin-
ear inductive potential EJ (xccjj ) cos(φ/ϕ0 ) for the qubit’s flux variable φ. The
Josephson energy EJ depends on the external flux that passes through the CCJJ.
Changing this flux modifies Josephson barrier EJ between potential wells. In
other words, the field xccjj may increase or decrease the tunneling amplitude
(xccjj ) in Hamiltonian (9.18).
(2) Changes in the qubit’s Josephson energy affect the critical current in the loop,
as well as the inductive coupling to the external fields and to other qubits.
The qubit’s current is inversely proportional to the inductance of the loop
IL,R ∼ ±φ/L. We can adjust and compensate for changes in this value by
inserting a dc-SQUID, labeled LT in Figure 9.2c. The dc-SQUID operates in
the linear regime as a tunable inductor L() (see Section 4.7.2), whose value
can be decreased to increase the current, and vice versa.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 256 Adiabatic Quantum Computing

(3) An adjustable inductor – labeled IPC in Figure 9.2c – mediates the interaction
of the qubit with a current source Ig . By adjusting this inductor, we can control
how much flux enters the qubit,11 thereby controlling the effective magnetic
field 12 h̄hi (t)σ z in (9.18).
(4) Another tunable inductor controls the coupling between the qubit and a readout
(RO) device. The RO is a current-biased CCJJ that operates very similarly to a
current-biased dc-SQUID (cf. Section 4.7.2). During a measurement, the tun-
able inductor is activated, so as to inject a small amount of flux into the CCJJ
δφ = δφ0 + ησ z . If the current bias and η are both large enough, the CCJJ
will switch to a voltage state when the qubit is in one of the states σ z = +1,
remaining inert in the other case. As explained in Berkley et al. (2010), this
implements the almost projective measurement described in Section 6.4.4. Nat-
urally, to minimize decoherence, at all other times the mediator is deactivated,
eliminating any residual coupling to the RO device (η = 0), and minimizing
decoherence.
We see in this design an abundant use of tunable inductors as control elements
for qubit gaps, magnetic fields, qubit potentials, measurement, and decoherence. A
simple counting of elements in Figure 9.2c reveals a total of six hardware controls
per qubit – one flux control for the inductor, one for the qubit gap, two for the
measurement, and two for the bias. Some of these controls are stabilized with the
use of the CCJJ, a device that allows better tolerance against fabrication errors – for
instance, small variations in the Josephson energies in the coupler circuit – and a
perfect cancellation of inductance.

Tunable Qubit–Qubit Coupling

The loop of an rf-SQUID qubit is not required to have any particular shape, other
than enclosing the right amount of magnetic flux. This allows experimentalists
to stretch the flux qubits so that they overlap with each other, forming a highly
connected graph. Figure 9.2d sketches how this is done for a unit cell of a D-Wave
quantum processor with eight qubits and 16 crossings.
Wherever two qubits overlap, both qubits experience an enhanced magnetic flux,
influenced by the other qubit’s current state. Similar to Section 6.5.1, the mutual
inductance between qubits creates a Jij σiz σjz dipolar interaction.12 The constant Jij
is nonzero only for connected qubits – e.g., J04 , J14 , etc., a total of 16 connections

11 See Section 6.4.2 and around (6.56), with the convention that now |L and |R are the qubit basis states.
12 Section 6.5 defined the dipolar moment operator in the qubit basis at the degeneracy point. However, in this
chapter we have changed basis, and the dipolar moment operator is defined in term of current states
σ z = |LL| − |RR|.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 257

in Figure 9.2. It is, however, static in nature, because the mutual inductance only
depends on the topology and geometry of the qubits’ intersection.
In order to achieve tunability of the coupling J matrix, D-Wave’s researchers
added tunable inductors that mediate additional couplings. As sketched in
Figure 9.2a and d, D-Wave’s mediator is an rf-SQUID that contains a CCJJ
circuit instead of a junction. The SQUID implements the interaction mechanism
from Section 6.5.4, but now with two controls: the SQUID’s internal flux and the
amplitude of the Josephson energy.
Figure 9.2b illustrates the dynamical range of the qubit–qubit mutual inductance:
One may control not only the amplitude, but also the sign of every qubit–qubit
interaction, reinforcing, canceling, or even reversing the geometric inductance pre-
viously discussed. As for the orders of magnitude, the mutual inductance shown
before ranges around ±1 pH, a very decent value. Considering that the qubit persis-
tent currents are |Ip |  1.0 μA for each of the two possible orientations, this gives a
safe estimate for the interaction energy E = M|Ip |2  10−24 J, which corresponds
to a frequency ω  E/h̄  2π× GHz. This value is consistent with the 4 GHz
antiferromagnetic splitting measured spectroscopically for qubits within a single
cell of the D-Wave architecture (Berkley et al., 2013; Lanting et al., 2014).

9.4.2 Device Operation

The operation of the flux-qubit quantum simulator has four phases: a design
phase, in which we find out the parameters of the Hamiltonian that we want to
program (9.18); a programming stage, in which we schedule the flux biases that
will control the local and interaction terms in the Hamiltonian, determining the
annealing schedule; an execution phase, where the annealing is performed; and a
final measurement stage, in which we detect the state of each qubit in the simulator.

Design and Embedding

Before using the qubit array to implement an adiabatic passage, we need to design
a set of interactions and local fields, and a protocol to change them along the pas-
sage, to simulate the interpolated Hamiltonian Ĥ [ε(t)]. Assuming that Ĥ0 is just a
transverse magnetic field (9.12) with J = h = 0, we can focus on determining the
couplings J and biases h that implement the problem Hamiltonian at the end of the
passage – i.e., the spin glass (9.18) or its equivalent QUBO formulation (9.16).
Unfortunately, as explained before, the topology of the desired problem Hamil-
tonian, given by the QUBO matrix Q, will not coincide in general with the qubit
arrangement in the quantum simulator. In particular, D-Wave’s design separates
qubits into cells of eight elements, with 16 mutual interactions. Each of these cells

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 258 Adiabatic Quantum Computing

then interacts with every neighboring cell via four of its qubits. For instance, in
Figure 9.2d, qubits 4, 5, 6, and 7 connect to neighboring cells on the same column,
while qubits 0, 1, 2, and 3 propagate interactions along a row. This sparse, nonplanar
topology is called the Chimera graph.
Our first goal is to find an embedding of the problem Hamiltonian Hp as an
equivalent model Hp implemented in the Chimera graph. Usually, the embedding
is done by directly identifying the logical qubits of Ĥp as a subset of the physical
qubits in the destination graph. Other qubits within the graph are used to mediate
interactions between the logical bits. The embedding must be such that Ĥp and
Ĥp exhibit similar low-energy eigenenergies, and from every configuration of the
embedded model |S  we can deduce the logical bits |s of the desired model.
In some lucky cases, the graph that we want to simulate can be found as a sub-
graph – a minor – of the Chimera graph. In this case, we need no auxiliary qubits
and the interactions can be mapped directly to couplings in the Chimera lattice. For
instance, the four-qubit model H = J σaz (σbz + σcz + σdz ) can be embedded into the
unit cell from Figure 9.2d, identifying (a,b,c,d) → (q4,q0,q1,q2 ) in the unit cell,
and setting J40 = J41 = J42 = J43 = J in the annealer’s controls.
Other times we need to copy a qubit, duplicating its value, because there is not
an exact one-to-one embedding. For example, think of the triangular lattice cell
H = J (σaz σbz + σbz σcz + σcz σaz ). We can implement this model with the logical qubit
assignment (a,b,c) → (q4,q0,q1 ), and copy the value of q0 to q5 with a strong
ferromagnetic interaction. This means we will have a coupling matrix J40 = J41 =
J51 = J , and J50 = −Jextra , with a large constant |Jextra |  |J |.
More generally, we need to engineer arbitrary interactions. The technique of
perturbative gadgets by Kempe et al. (2006) allows the engineering of rather general
effective interactions starting from an experimental platform. As before, we identify
logical qubits with physical qubits, and add auxiliary qubits and interactions to
reproduce the target model. However, this framework is more systematic and
powerful, enabling the simulation of arbitrary two-, three- and other many-body
interactions between the logical qubits, which are engineered from the simple-
looking Hamiltonian in (9.18).
Embedding and gadgets extend the family of models accessible to a given archi-
tecture, but they have associated costs. First, the use of auxiliary qubits increases the
size of the computation. Venturelli et al. (2015) studied the worst-case scenario of a
fully connected network and found that the Chimera graph requires O(N 2 ) physical
qubits to represent N logical ones. Second, the search for a good embedding is
itself an expensive classical overhead – minor embedding is an NP-hard problem.
Therefore, in many cases we must resort to a suboptimal, but good-enough heuristic
(Cai et al., 2014) for the simulation. Third, some gadgets will produce interactions
that can be perturbatively small, leading to a decrease in the available energies Emax

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 259

and an increase of time. Finally, some embeddings require many auxiliary qubits and
long chains of interactions. The transverse magnetic field combined with these long
chains produces a contraction of the gap13 min and an increase of passage time.

Annealing Schedule and Control

Once we have defined the values of the parameters , Jij , and hi at the beginning
t = 0 and at the end t = T of the passage, we need to find a control that interpo-
lates between both limits. In Section 9.2, we suggested a linear ramp of the fields,
which decreases the transverse field  ∝ (1 − t/T ), at the same rate it increases
the strength of the problem Hamiltonian J,h ∝ t/T . It turns out that this is not
a natural choice for a flux-qubit design, because the values of , J and h are all
related to the critical current of the qubit.
Take, for instance, the tunneling amplitude. It decays exponentially as we
increase the Josephson energy of the qubit, i ∝ exp(−αEJ,i (xccjj,i )). We can
find a schedule for the control flux xccjj (t) that imposes the desired decrease in
the quantum fluctuations i . However, that schedule will also change the critical
current of the qubit14 IJ,i ∝ EJ,i , enhancing the coupling to the external bias fields
hi ∝ IJ,i and the qubit–qubit interactions Jik ∝ Mik IJ,i IJ,k .
These implicit dependencies may cause some terms, such as the transverse fields,
to decrease faster than other terms, as biases, and interactions grow. To some extent,
one may compensate these dependences through the control of the mutual induc-
tances Mik or the qubit’s injected flux. However, it will quite often happen that
both the maximum available energy !Hp (t)! ∼ Emax and the instantaneous gap 
will deteriorate, affecting the actual passage speed and conditions of the adiabatic
theorem. As example, see figure 9 in Albash and Lidar (2018), which shows the
annealing schedules as measured by D-Wave for the 2000Q and 2X devices, along
the dimensionless passage time s = t/T .

Programming and Measurement Overheads

In D-Wave’s devices, the annealing schedule is programmed into a specialized
superconducting circuitry, using an embedded digital-to-analog converter and on-
board superconducting switches (Johnson et al., 2010). The control circuitry also
takes care of the final measurement of all qubits, extracting the information at the
end of the passage. The total execution time of an annealing protocol thus includes
three times: the programming time, the annealing time, and the measurement time.

13 Qualitatively, this contraction is caused by the difficulty of transverse fluctuations  σ z in effecting a change
i i
over a string of qubits that want to be on the same state (see Problem 9.6).
14 See figure 5 in Harris et al. (2010b) and how a change of x by an amount 0.02 causes an increase of I
ccjj 0 J
from 0 to 1.6 μA, and an exponential decrease in  from 109 to 105 Hz.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 260 Adiabatic Quantum Computing

The programming and measurement times are constant overheads that do not
depend so much on the problem at hand. These can be very large, with respective
values around 7 ms and 125 μs reported for a D-Wave 2000Q by Albash and Lidar
(2018). The annealing time is problem dependent, but it is usually the shortest value
in the whole experiment. Due to strong filtering of controls to limit the influence
of high-frequency noise, the shortest annealing time starts at 20 μs in the D-Wave
Two device and 5 μs in D-Wave One, Two X, and 2000Q. This minimal passage
time can obscure or perturb the interpretation of resource scaling, as explained by
Mandrá et al. (2016).

9.4.3 Performance Analysis

D-Wave’s quantum optimizer is probably the largest man-made quantum device
with independently addressable qubits, tunable local fields, and fully adjustable
interactions.15 It is therefore important to understand how coherent or how “quan-
tum” these devices are, and what intrinsic and extrinsic causes limit their perfor-
mance. Unfortunately, the architectural choices that facilitate the operation of these
quantum annealers – highly filtered and slow controls, slow measurements at the
end of the control, no individual qubit addressing or excitation – prevent us from
applying sophisticated quantum information methods to analyze coherence, evolu-
tion, and entanglement. For instance, the tomography protocols from Section 8.4.2,
fidelity measurements, and randomized benchmarks from earlier chapters are out of
question. Instead, we obtain information from the device by studying its operation
as quantum simulator and through multiple adiabatic passages.

Evidence of Quantum Effects

One way to show evidence of quantum effects is to study the ground state and exci-
tations of the interpolating Hamiltonian (9.18) in different regimes of interactions:
from the paramagnetic phase (J,h = 0), to the spin-glass phase ( = 0), passing
through the strongly correlated phase transitions. Unfortunately, D-Wave’s chips are
protected by strong low-pass filters that eliminate any high-frequency microwave.
This means we cannot use microwave spectroscopy (Sections 7.4.1 and 7.4.2) to
probe the simulator.
Berkley et al. (2013) designed a method to circumvent this limitation, which they
called qubit tunneling spectroscopy. This technique relies on isolating a probe qubit

15 In terms of qubits, it may be argued that ultracold atoms in optical lattices have larger systems. However, as of
this writing, they have not exhibited a similar control of both local fields and separate interactions between
neighboring quantum systems.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 261

that interacts with other qubits that implement the desired Hamiltonian (9.18). The
setup is described by the effective model

0 x ε0 z 1
Ĥ = σ + σ0 + J σ0z σkz + εcomp σ0z + Ĥp . (9.19)
2 0 2 2

We assume that the probe qubit σ0z interacts with only one other qubit σkz from the
many-body simulator Ĥp . We control the probe’s bias ε0 and tunneling , as well as
a compensating field acting on the contacted qubit. When this field is εcomp = 2J ,
the eigenstates of H split into two manifolds, with different values of σ0z = ±1,
which are coupled by a weak tunneling 0 :
 
1 1
Ĥ = Ĥp + |10 1| ⊗ ε0 + 2J σk + 0 1 ⊗ σkx .
z
(9.20)
2 2

The system is adiabatically prepared to an eigenstate |10,ψref , using a very large

negative bias ε0  −2J , a large coupling J > kB T that guarantees a large sep-
aration from other states. This state will be our reference energy Eref . We will
then adiabatically shift the bias to a higher value ε0 + δω, until Eref + δω = En
and our reference state becomes resonant with |00,n, an eigenstate of the problem
Hamiltonian Hp |n = En |n. At resonance, the tunneling 0 becomes relevant and
enables oscillations between states, |10,ψref  and |00,n, which can be detected by
measuring the probe qubit. The probability of a probe qubit flip as a function of ε0
will show peaks centered on those resonances, with linewidths the depend on the
decoherence rates of the qubits.
This method was applied by Lanting et al. (2014) to study the eigenenergies and
eigenstates of a Hamiltonian Hp with two and eight qubits from a unit cell in the
Chimera graph. Their spectroscopy was in perfect agreement with the predictions
of an exact diagonalization model. It revealed all regimes of the multiqubit spin
model, the competition between interactions Jij and quantum fluctuations i , and
the reduction of the gap around the phase transition point. Moreover, those studies
showed that, for such a small number of qubits, the gap was large enough that an
adiabatic passage could prepare the ground state with great fidelity F  1.
Another way to probe the quantum properties of the annealer is to study how
well it simulates the Hamiltonian Ĥp , and other equivalent models, from a purely
condensed-matter physics point of view, such as studying phase transitions. In a
remarkable work, Harris et al. (2018) simulated a complete 8 × 8 × 8 3D spin
lattice, embedding the 3D Hamiltonian of 512 qubits into a Chimera architecture
with 2 048 qubits. Engineering the right annealing passages, they could explore
the three-dimensional phase space determined by the effective temperature T , the
spin–spin interaction J , and the transverse magnetic field . They could probe all

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 262 Adiabatic Quantum Computing

paramagnetic, anti-ferromagnetic, and spin-glass phases, as well as find the phase

boundaries predicted by quantum Monte Carlo.
The work by King et al. (2018) explores a different type of physics, looking
at a topological phase transition that appears for 2D Ising models with transverse
field. Once more, they worked with the embedding method to simulate frustrated
spin lattices with 1 800 logical qubits on a 2D cylindrical topology and different
connectivities (triangular lattice and square-octagonal) on a superconducting pro-
cessor with 2 048 physical qubits. Using a very low-temperature adiabatic passage
T  8 mK, they could explore the Berezinksii–Kosterlitz–Thouless phase transi-
tion from a paramagnetic phase – a large transverse magnetic field  – into the
critical region, comparable  and J – where vortices appear. They could probe the
appearance of a long-range, algebraically decaying topological order – the vortex
phase – and compare with simulations based on state-of-the-art quantum Monte
Carlo methods, which matched with the effective temperature T = 8.4 mK.
Both of these works represent a deep exploration of quantum simulation with
large systems, only comparable to contemporary experiments with ultracold atoms
in optical lattices. Moreover, these works demonstrate the value of large quantum
processors for studying fundamental physics, and confirm that the ideas from adia-
batic quantum computing are useful even outside of the computing world.

Decoherence
Despite the evidence of coherent macroscopic quantum tunneling events (Lanting
et al., 2014), and simulation of quantum many-body phase transitions (Harris et al.,
2018; King et al., 2018), D-Wave quantum processors up until 2019 have too large
decoherence rates to be considered an adiabatic quantum processor.
The role of decoherence can be studied at the single-qubit level. Harris et al.
(2010b) analyzed the tunneling between current states in a qubit at  = 0, to reveal
fluctuations in the qubit’s flux and environmentally induced dephasing. Later,
Berkley et al. (2013) used qubit tunneling spectroscopy to analyze the spectra
and coherence of one qubit. Both studies suggested a large decoherence rate, with
qubit linewidths  0.7 GHz, only an order of magnitude away from the qubit–
qubit interactions. This sets the decoherence time T2∗ ∼ 1 ns, three to six orders of
magnitude worse than state-of-the-art superconducting qubits (cf. Section 6.6), and
much shorter than the shortest passage times ∼ 5 μs.
Harris et al. (2010b) further studied the behavior of a small eight-qubit quantum
simulator when subject to a control designed for an adiabatic passage. They showed
that the dynamics of the qubits could be modeled using a master equation that
combines the interpolating Hamiltonian with a finite heating rate – given by the
effective temperature of the chip – and a strong dephasing rate – essentially the

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 263

T2∗ of a single qubit. Even if the low temperatures were unable to bridge the gap
of this relatively small system, and the experiment could prepare ground states
with high fidelity, the study evidences that the superconducting processor does not
implement an adiabatic quantum computation in the sense of Section 9.2. Instead,
we can describe this dynamics as an implementation of quantum annealing.

Quantum Annealing versus AQC

The term annealing originated in the world of metallurgy and material science, and
refers to a smooth change of the temperature of a material, starting from a hot,
noisy, highly “destructive” environment capable of creating many excitations, down
to a cool and quiet environment. When done properly, classical annealing can take
highly excited material down to an almost perfect equilibrium state, where atoms
or molecules arrange into perfect lattices with few defects.
One extension of this concept to the computational world is known as simulated
annealing. The algorithm takes as input a function F (s) to be optimized – e.g., a
classical NP-hard spin glass (9.15) or a QUBO function. It reinterprets this function
as the energy of a classical system that is in contact with a bath at temperature T .
Using Markov-chain Monte Carlo or some other technique, it simulates the thermal
fluctuations of the variables s in a process of continuously decreasing temperature.
After a sufficiently long time – the time to solution – the classical algorithm
will visit with high probability the minimum of F (s) or a good approximation
thereof.
Quantum annealing is a related idea. It was introduced by Kadowaki and Nishi-
mori (1998) as a mean to approximate classically the ground states of quantum
Hamiltonians. Instead of looking just at the temperature, quantum annealing sim-
ulates quantum fluctuations with progressively decreasing strength. The method
emulates, by quantum Monte Carlo techniques, the physics of the interpolating
Hamiltonian (9.18), starting with a Hamiltonian in the regime of large transverse
fields , and crossing over to a regime of dominating interactions, where we recover
the ground state of the problem Hamiltonian Ĥp .
Unlike adiabatic quantum computation or the idea of adiabatic passages, quantum
annealing also includes the possibility of additional sources of fluctuations, such as
temperature. Probably for this reason, we find the term quantum annealer already
in early descriptions of quantum simulators (Harris et al., 2010b), which imple-
ment adiabatic-like passages with interpolating Hamiltonians, in nonideal condi-
tions – interactions with the environment, dephasing, heating or cooling due to finite
temperatures, etc. We will now discuss how theoretical and experimental evidence
supports this analogy, and how even the idea of quantum annealing affects the
computational power of the simulator.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 264 Adiabatic Quantum Computing

Evidence of Quantum Annealing

Quantum annealing, adiabatic quantum computers, and simple Landau–Zener pas-
sages all exploit quantum fluctuations (quantum tunneling) to access other states
and configurations in a more efficient way than classical techniques.
Consider the LZ passage for a flux qubit. When we change the bias field ε from
a large negative value to a large positive one, we are changing from a potential in
which the minimum energy state is the |L current (cf. Figure 6.6b) to the opposite
configuration, with a minimum around |R. Classically, the qubit would only shift
from L to R when the potential is so tilted that the |L state becomes unstable.
However, the quantum mechanical fluctuations introduced by the capacitor enable
the tunneling of the qubit between minima, and already at ε = 0, halfway through
the qubit hyperbola, 50% of the probability has been transferred from one well to
another.
One school of thought supports the idea that quantum fluctuations, when applied
in the right Hamiltonians and annealing passages, can provide a computational
advantage in optimization problems. The argument is that quantum fluctuations
enable tunneling across local minima, avoiding incorrect solutions and reaching the
true ground state faster than through classical optimization techniques. A legitimate
question is whether this advantage is present in existing NISQ devices and for the
type of NP-hard problems that we can simulate (9.18).
We already have evidence that quantum fluctuations are present in the adiabatic
passages with flux-qubit processors. The experiments by Johnson et al. (2011) and
Dickson et al. (2013) demonstrated quantum effects in spin models with 8 and 16
qubits. Boixo et al. (2013) studied also small devices with eight qubits and showed
that the annealing passages were consistent with exact simulations based on master
equations – that is, evolution that combines an adiabatic passage Hamiltonian (9.18)
with some decoherence. The signatures of the passage were also incompatible with
models based on instantaneous thermalization or classical evolution.
Boixo et al. (2014) extended this study to passages with 108 spins. Because the
system was much larger, instead of master equations, they had to simulate the quan-
tum annealing using a stochastic method known as simulated quantum annealing
(SQA), which is an adaptation of quantum Monte Carlo to emulate the annealing
schedule. Their work found that SQA correlated very well with the outcomes of the
D-Wave annealer. Thus, instances which are hard – i.e., low success probability –
for Monte Carlo SQA are also hard for the flux-qubit simulator, and vice versa.
This correlation was not found for classical methods based on simulated annealing.
Nowadays, there is some evidence that both fluctuations and quantum correla-
tions are relevant in the actual device. Albash et al. (2015) showed that passages
with 20 qubits in the D-Wave 2 processor are consistent with correlated models
based on master equations, as in Boixo et al. (2013). This work also disproved

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 9.4 D-Wave’s Quantum Annealer 265

that experiments could be fully explained using classical models of spin-dynamics

(SD) and spin-vector Monte Carlo16 (SVMC), based on the statistics of the output
states and excited state dynamics during the passages. Later, Boixo et al. (2016)
and Denchev et al. (2016) have shown the relevance of finite-range tunneling –
quantum tunneling events where more than one qubit change state simultaneously –
in passages with specially crafted Hamiltonians and up to 945 qubits.

Computational Advantages
One original motivation for constructing the flux-qubit quantum simulator of the
spin-glass Hamiltonian (9.18) was the possibility of solving NP-complete problems,
and whether quantum effects – fluctuations, entanglement – can improve the effi-
ciency in finding those solutions. It is for this reason that many works have analyzed
the scaling of the physical quantum annealer, comparing it to other classical meth-
ods, with various results.
A first step in these analyses is figuring out the right metric. It turns out that
this is the time to solution, which is the shortest execution time, including rep-
etitions, that guarantees a solution with a high success probability, such as 99%
(see Problem 9.7). A second step is to choose the right classical algorithms for
the comparison. Here the focus has been on generic algorithms that work for all
problems, namely spin dynamics, simulated annealing, spin vector Monte Carlo,
and Monte Carlo simulation of quantum annealing passages. This ignores to some
extent the fact that some problems have specific tailored methods that are very
efficient.17
Denchev et al. (2016) is one of the first works to prove a constant speedup in the
physical quantum annealing (QA) as compared to the classical method, with up to
eight orders of magnitude. They found a similar scaling between SQA and QA with
the problem size, but the Monte Carlo method was about 108 times slower in real
time. Later works by Mandrá et al. (2016) and Mandrá and Katzgraber (2018) have
extended the comparison to other sequential and parallelized methods, considering
also some tailored algorithms. These works find also a similar scaling between
SQA and QA, with a constant speedup factor in time to solution, although they
also realize that this speedup could be compromised by minimum passage time
and hint at the need to find the optimal time over all possible passages. Albash
and Lidar (2018) performed the optimal time-to-solution analysis for the first time,
comparing hard problems with a planted solution. This study confirms the speedup

16 This is a method introduced by Shin et al. (2014) that represents the quantum register as a product state of
uncorrelated spins, updated in an easy-to-simulate classical method that accounts for some of the quantum
fluctuations.
17 For instance, there is a very clever and very efficient algorithm that is specific for solving QUBO problems in
the Chimera graph topology, due to Selby (2014).

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 266 Adiabatic Quantum Computing

of QA over classical simulated annealing. It also finds that QA is quantitatively

faster, but it finds for the first time that SQA and SVMC have a slightly better
asymptotic behavior – though in absolute terms it is larger, the optimal TTS grows
slightly slower than QA with the number of qubits.

9.5 Summary and Outlook

In this chapter, we have introduced a new computational model based on the prepa-
ration of ground states from problem Hamiltonians. We have introduced the adi-
abatic quantum algorithm as a specific method to solve this task and presented
sample Hamiltonians that encode useful and universal problems. In the second half
of the chapter, we have set our focus on a superconducting circuit architecture that
simulates adiabatic passages with NP-complete Hamiltonians. We have discussed
the technology and limitations of this architecture and investigated how the device
performs in actual experiments, and what those experiments teach us about adiabatic
computations and quantum annealing.
Unfortunately, this discussion is a bit muddied by the intrinsic difficulties and
challenges posed by the choice of Hamiltonians. Some key messages that we can
distill from the broad literature are the following:

– NP-complete spin-glasses are stoquastic models that can be efficiently simulated

using Monte Carlo methods and similar techniques. As such, even if experi-
ments show that quantum fluctuations have a computational advantage, there is
no exponential quantum advantage, because those fluctuations can be simulated
by classical computers.
– D-Wave’s design and implementation represent a very important technological
milestone in the development of quantum simulators and quantum technologies.
Those devices have still interesting fundamental applications, can be used to
probe interesting physics, and may involve a de facto computational speedup
with respect to other simulation techniques.
– Existing quantum annealers are based on relatively old qubit technology, with
relatively short coherence as compared to modern qubits, long readout times,
slow controls, and sparse, limited connectivity.

There are some obvious paths to move ahead, both in the field of quantum anneal-
ing and in quantum simulation (Hauke et al., 2020):

– Study new annealing schemes, which may involve the use of different schedules
for different qubits, random perturbations, or different interaction schemes. Some
of these techniques may eliminate first-order phase transitions, or allow those to
happen in smaller regions that get connected and produce better solutions.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 Exercises 267

– Achieve better coherence times in qubits that are compatible with quantum
annealer designs. Some examples are recent designs of capacitively shunted flux
qubits, which feature slower dephasing rates (Steffen et al., 2010; Yan et al.,
2016).
– Engineer new interaction schemes between superconducting qubits, to create
nonstoquastic couplings of σix σjx simultaneous with σiz σjz . This is not a simple
task, but there are evidences for such couplings with flux qubits by Ozfidan
et al. (2020), and there are even designs to enable three- and four-body couplings
(Chancellor et al., 2017).
However, if we have learned something from the discussion and research with
D-Wave’s processors, it is that sometimes we can make something very good for the
wrong reasons. Superconducting quantum annealers were born out of commercial
interest around computational applications, which failed at demonstrating a quan-
tum advantage. However, the result is still a unique demonstration of the potential
for scalable quantum simulations with superconducting quantum circuits. Maybe
one possible future of this technology – microwave circuits and qubits – lies pre-
cisely in avenues that explore this potential, researching fundamental and practical
problems at the interface between quantum optics, condensed-matter physics, and
other fields with complex quantum systems. I hope you, as a reader, will be able to
bring this new insight into the field, and use the tools from this book to construct
those new applications.

Exercises
9.1 Let us analyze the Landau–Zener model (9.2).
(1) Show that Ĥ (ε) = 12 (ε)U σ z U † where U = exp(−iθ(ε)σ y /2).
(2) Rewrite the wavefunction of the two-level system in a rotating frame with
|ψ(t) = U (t) |ξ (t) and show that the rotated state |ξ (t) evolves with the
Hamiltonian:
1 min ε̇ y  !
Ĥeff = (ε)σ z − σ ξ (t) . (9.21)
2 22
(3) Implement a change of variables t → ε, using ∂t = ε̇∂ε , to recover the
effective equation (9.5).
(4) Provide an upper bound on the rate of change of the prefactor in front of
σ y and compare it with the σ z term.
(5) Approximately diagonalize the effective Hamiltonian, showing that the
2 ×  ≤ 
|0 eigenstate of σ z gets a correction of order min 2ε̇ 2ε̇
min
.
9.2 Let’s assume that we have a qubit with frequency , interacting with a
microwave cavity of frequency ω through a coupling strength g, as described

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 268 Adiabatic Quantum Computing

by the Jaynes–Cummings model (7.33). Initially the qubit is off-resonant but

with a small frequency   ω, and we wish to move the qubit so that it has a
larger frequency  > ω. How fast do we have to increase  so that the state
of the qubit and the cavity remains unchanged? What determines whether the
speed is adiabatic or not?
9.3 If we implement the spin-glass model (9.15) in one dimension, we have that
the only nonzero elements are of the form Ji,i+1 . Show that in this case the
problem is trivial. Hint: Eliminate the signs in J with a gauge transformation
σiz → σix σiz σix on a subset of spins.
9.4 Aharonov et al. (2004) proved that an adiabatic computation with a sparse
Hamiltonian can be efficiently simulated by a quantum circuit. Essentially, we
can simulate the time evolution by decomposing with circuits that simulate
rotations based on Ĥ0 and Ĥp for brief intervals:
,
L
U= exp(−i Ĥ0 t(1 − ε(tn ))) exp(−i Ĥp tε(tn )). (9.22)
n=1

Naturally, the precision of the simulation increases when we decrease the time
step t = T /L and increase the total passage time T .
(1) Show that evolution with Ĥ0 decomposes into local rotations.
(2) How many quantum gates do we need to simulate evolution with the spin-
glass model (9.15) if the qubits are organized in the 2D and 3D lattices
studied by Barahona (1982)?
9.5 Take the QMA-complete Hamiltonian from (9.17). Show that the energy Ĥ 
of any quantum state can be estimated by repeating M times the measure-
ment18 of either all σiz or all σix Pauli observables√on all qubits. Show that
the uncertainty in these measurements scales as 1/ M. Assume we want to
keep the experimental precision below some bound ε. Prove that the number
of measurements scales at most as O(N 4 /ε), assuming the worst-case scenario
of N 2 terms in (9.17).
9.6 Assume we want to establish an interaction J between two logical qubits,
assigned to positions 1 and N + 1, which are connected by an intermediate
chain of N physical qubits, as in
1
Ĥ = h̄ σ1z + σN+1 z
− Jf σ1x σ2x + · · · + σN−1
x
σNx + J σNx σN+1
x
. (9.23)
2
Compute the minimum gap of this model in the limit of large ferromagnetic
coupling |Jf |. If we make an adiabatic passage with this Hamiltonian, using

18 This is called a two-setup experiment, because experimentalists only have to do two kinds of experiments that
only differ in the final measurement at the end.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 Exercises 269

Ĥ0 (9.12), what is the probability that the state of qubit 1 is copied to the N
intermediate qubits?
9.7 The usual metric in annealing experiments is the time to solution, that is,
the time tTTS required to find the solution with a fixed probability, typically
pTTS = 99%. Assume that we perform passages with a time tramp and find the
true ground state with a probability pramp < 99%. Compute the number of
times M that we have to repeat the annealing experiment to obtain a success
probability p ≥ 99.99%. Discuss what happens to tTTS when the ramp time is
lower-bounded due to the restrictions of the device.

https://doi.org/10.1017/9781316779460.009 Published online by Cambridge University Press

 Appendix A
Hamiltonian Diagonalizations

A.1 Tridiagonal Matrix Diagonalization

A.1.1 Periodic Boundary Conditions
We need to diagonalize the matrix
⎛ ⎞
a b ... 0 b
⎜b a ... 0 0⎟
⎜. ⎟
B=⎜
⎜ ..
.. . .
. . ⎟ ∈ RN ×N .
⎟ (A.1)
⎝0 0 ... a b⎠
b 0 ... b a

The eigenmodes are defined by the eigenvalue equations

b(um−1 + um+1 ) + aum = λum, m = 1 . . . N, (A.2)

where we need the identification um+N = um associated to periodic boundary conditions.

We can solve the eigenvalue equations with plane waves:

1 2π n 2π
m = √ exp(ikn m),
u(n) kn = ∈ × {0,1,2, . . . N − 1}, (A.3)
N N N

that naturally satisfy the boundary conditions, um+N = um exp(ikN ) = um exp(i2π n). The
eigenvalues associated form a band λn = a + 2b cos(kn ).
It is quite often more convenient to define the quasimomenta kn so that they explore the
interval (−π,π ]. For even sizes N = 2M,
π
kn ∈ × {−M + 1, − M + 2, . . . − 1,0,1, . . . M}, (A.4)
2M
while for odd one N = 2M + 1
π
kn ∈ × {−M, − M + 1, . . . − 1,0,1, . . . M}. (A.5)
2M + 1

270

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 A.1 Tridiagonal Matrix Diagonalization 271

We can also use real wavefunctions if we instead diagonalize B with a real orthogonal
transformation, B = O T O, O ∈ RN×N , O T O = 1. The eigenfunctions are then linear
sinusoidal functions:

2
u(even,n)
m = cos(kn m), kn ∈ (0,π ) (A.6)
N

2
(odd,n)
um = sin(kn m),
N
plus the two extra solutions:

1
u(even,0)
m = , k0 = 0, (A.7)
N

1
u(even,n)
m = (−1)m, kn = π . (A.8)
N

A.1.2 Open Boundary Conditions

We need to diagonalize the tridiagonal matrix
⎛ ⎞
a b ... 0 0
⎜b a . . . 0 0⎟
⎜. . . ⎟
B=⎜ ⎜ .. .. .. ⎟ ∈ RN×N .
⎟ (A.9)
⎝0 0 . . . a b⎠
0 0 ... b a

The eigenmodes are defined by the eigenvalue equations:

b(um−1 + um+1 ) + bum = λum, m = 2. . . N − 1 (A.10)

bu2 + au1 = λu1,
buN−1 + auN = λuN .

These equations are equivalent to a periodic boundary condition problem (A.2) of size
2(N + 1), where one imposes uN+1 = u2(N+1) = 0. This solution is constructed from
two degenerate plane waves with opposite momenta and the same eigenenergy:

um ∝ eikm − e−ikm . (A.11)

The final solution, including normalization factors, reads


2 π
um = sin(kn m), kn ∈ × {1,2, . . . N} (A.12)
N N +1
λn = a + 2b cos(kn ).

Unlike the periodic case, we can no longer shift the linear momenta, which are constrained
to the interval [0,π ).

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 272 Hamiltonian Diagonalizations

A.2 Harmonic Chain Diagonalization

We are going to diagonalize a problem of coupled Harmonic oscillators, which we write in
the following form:
1 T 1 T
H = p̂ p̂ + x̂ B x̂. (A.13)
2C 2L
Let us assume that matrix B ∈ RN×N , which is real and symmetric, is diagonalized with a
unitary or orthogonal transformation:

ij = λi δij
B = U † U, with (A.14)
U T U = 1.
We define the position and momentum operators using N Fock operators:
 1  ∗ †

x̂m = an √ Umn b̂n + Umn b̂n , (A.15)
n 2
 h̄ i  
∗ †
p̂m = √ Umn b̂n − Umn b̂n , (A.16)
n
an 2
†
which satisfy the usual commutation relations [b̂n, b̂m ] = δnm . The constants an > 0 are
“length” scales to be adjusted.
∗ U
Since Unm nr = δmr , these operators satisfy the canonical commutation relations
[x̂m, x̂r ] = 0, [p̂m, p̂r ] = 0, [x̂m, p̂r ] = i h̄δmr . Thanks to this orthogonality condition, we
can write
 h̄2  
p̂T p̂ = b̂ †
n b̂ n + b̂ n b̂ †
n − b̂ †2
n − b̂ 2
n (A.17)
m
2an2
 a 2 λn  
n
x̂T B x̂ = b̂n† b̂n + b̂n b̂n† + b̂n† 2 + b̂n2 . (A.18)
m
2
2 a2 λ
2 = 2L = 2 h̄ωn , to cancel the b̂ and b̂
h̄ n 1 2 † 2 terms. This gives us the canonical
We impose 2Ca
n
frequencies and the oscillator lengths:
 
λn h̄
ωn = , an = , (A.19)
CL Cωn
and makes the Hamiltonian diagonal:
  
1
Ĥ = h̄ωn b̂n b̂n +
†
. (A.20)
n
2
In a problem with periodic boundary conditions, with Bij = Bi+1j +1 , the matrix B can
be diagonalized with either sines and cosines or complex exponentials (Section A.1.1). If
we choose the plane waves, we write
 1  ikm 
x̂m = an √ e b̂k + e−ikm b̂k† , (A.21)
k
2N
 h̄ i
p̂m = √ (e−ikm b̂n† − eikm b̂n ), (A.22)
an 2N
k

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 A.3 Schrieffer–Wolff Perturbation Theory 273

where the sum takes place over the allowed values

 of the quasimomentum
 k. The diagonal-
 †
ized Hamiltonian takes the form Ĥ = k h̄ωk b̂k b̂k + 2 , with the degeneracy ωk = ω−k .
1

A.3 Schrieffer–Wolff Perturbation Theory

In this book, and in many physics problems, we often face Hamiltonians that can be decom-
posed into a dominant Hamiltonian and a perturbation:
Ĥ = Ĥ0 + g V̂ . (A.23)
The Hamiltonian Ĥ0 is a simple problem – e.g., a bunch of noninteracting qubits and
microwave resonators – whose eigenstates |n and eigenenergies En we know with complete
accuracy. Perturbation theory is a mathematical strategy to approximate the eigenstates and
eigenenergies of Ĥ as “corrections” of the original states and eigenvalues.
Schrieffer–Wolff or Kato perturbation theory is a family of perturbation theory methods
that have become popular as a technique designing quantum simulators and engineering
quantum interactions. The assumption of this particular theory is that, for weak enough
interactions g V̂ , the eigenspaces of Ĥ0 are only slightly deformed so that they can be related
to those of Ĥ by a unitary transformation, to be approximated perturbatively. We will now
discuss two variants of this situation, with increasing level of complexity.

A.3.1 Nondegenerate Perturbation Theory

Assume that we write a one-to-one mapping between the eigenstates:
 
Ĥ0 = n En |nn|
 , with |n  = Û (g) |n . (A.24)
Ĥ = n Ẽn (g) |n n |
This can only work if the eigenenergies of each Hamiltonian are nondegenerate and they
never cross – e.g., Ẽn (g) < Ẽn+1 (g) ∀n. If g is small enough that this level crossing never
happens, we can study the unitary transformation as being generated by a non-Hermitian
operator Û (g) = exp(−g Ŝ(g)), expanding the function Ŝ(g) as a perturbative series in the
small parameter g.
Our derivation proceeds by analyzing the effective Hamiltonian Ĥeff = eg Ŝ Ĥ e−g Ŝ that
results from mapping Ĥ back into the original unperturbed space Ĥ0 :
    1  
Ĥeff = Ĥ0 + g V̂ + g Ŝ, Ĥ0 + g 2 Ŝ, V̂ + g Ŝ, g Ŝ, Ĥ0 + O(g 3 ). (A.25)
2
We assume that V̂ is off-diagonal on the original eigenbasis. We can always justify this
assumption by incorporating the diagonal terms into Ĥ0 as an effective renormalization of
the qubits and photons in that model:1

Ĥ0 → Ĥ0 + g |nn|V̂ |nn| , (A.26)
n

1 Alternatively, you may use more sophisticated derivations such as the one by Bravyi et al. (2011), on which
this appendix is based.

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 274 Hamiltonian Diagonalizations

V̂od := |nn|V̂ |mm| . (A.27)
n=m

This allows us to assume Vnm := n|V̂od |m = 0 when n = m. We can then expand our
generator g Ŝ = g Ŝ0 + O(g 2 ), with an operator Ŝ0 that cancels all first-order terms in the
expansion of Ĥeff , producing [g Ŝ, Ĥ0 ] = −g V̂od .
In the nondegenerate perturbation theory limit, this equation can be solved formally,
projecting onto the original eigenstate basis:

g(Em − En ) n|Ŝ0 |m = −g n|V̂ |m , (A.28)

which is solved by
 Vnm
Ŝ0 = × |nm| . (A.29)
En − Em
n=m

This gives us the second-order expansion of the effective model:

1
Ĥeff = Ĥ0 + g 2 [Ŝ0, V̂ ] + O(g 3 ) (A.30)
2
  
g2  Vnm Vmr Vnm Vmr
 (En + Vnn ) |nn| + + |nr| .
2 En − Em Er − Em
n n,r=m

We will use these expansions in several qubit designs, such as when justifying the truncation
to a two-level subspace in the charge qubit (Problem 6.6), or when analyzing the residual
interactions in two highly detuned qubits (see Section 6.5.3 and Problems 6.16 and 6.17).

A.3.2 Degenerate Perturbation Theory

Sometimes we want to consider interaction strengths that go slightly beyond the limits we
have mentioned, or we want to study Hamiltonians Ĥ0 with eigenstates that are energetically
very close to each other and are strongly mixed by the perturbation. To make perturbation
theory work in this particular limit, we have to assume that the spectrum of Ĥ0 can be divided
into eigenspaces of states that, even during the perturbation, will remain close to each other
and not mingle with other subspaces. This complicates the notation somewhat, forcing us
to label eigenstates by subspace α, as in |n,α and |n,α, and introduce projectors onto the
original and deformed models:
 
P0,α = n |n,αn,α|
  , with Pα (g) = Û (g)P0,α Û (g)† . (A.31)
Pα (g) = n |n ,αn,α|

As discussed by Bravyi et al. (2011), we can still develop a Schrieffer–Wolff transforma-

tion building on the expansion (A.25), but now Heff has to be treated as a box-diagonal
Hamiltonian:
 
Heff = P0,α Heff P0,α = α
Heff . (A.32)
α n

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 A.3 Schrieffer–Wolff Perturbation Theory 275

This structure changes the equation we have to solve for Ŝ0 and V̂ . We need to split
V̂ = V̂d + V̂od into contributions that “live” within the same energy subspace and that will
be incorporated into Ĥ0 and contributions V̂od that connect different subspaces:
P0,β [Ŝ, Ĥ0 ]P0,α = P0,β V̂ P0,α , α = β. (A.33)
These equations can be solved, producing the anti-Hermitian operator:
 1
Ŝ = |n,αn,α|V̂ |n,βn,β| . (A.34)
E − Em,β
n,m n,αα=β

We can continue further with the perturbation theory argument, substituting this operator
into the series for Ĥeff and inspecting the box within one particular subspace. The result is
a generalization of the nondegenerate perturbation series:
1  
α
Ĥeff = P0α Ĥ0 + g V̂d P0α + g 2 P0α Ŝ, V̂od P0α + O(g 3 ). (A.35)
2
Explicitely in terms of eigenstates,
 
α
Ĥeff  En,α |n,αn,α| + n,α|V̂ |m,α |n,αn,α| (A.36)
n n,m
 
g2   1 1
+ +
2 En,α − Em,β Er,α − Em,β
β=α n,r=m

× |n,αn,α|V̂ |m,βm,β|V̂ |r,αr,α| .

This series combines the matrix elements of the perturbation within a given subspace (see
the first line), with additional terms that result from virtual transitions to other subspaces
(generically labeled β = α in the second and third lines). We use these expansions when
studying the dispersive coupling between cavities and resonators in Section 7.3.7.

A.3.3 Considerations
Equations (A.30) and (A.36) are mathematical recipes for an effective Hamiltonian that is
expressed in the basis of eigenstates of the unperturbed model Ĥ0 . From this point of view,
the appeal of this method to the design of quantum simulators and quantum engineering is
clear: This theory allows us to design effective interactions, simply by studying how various
perturbations to a system or circuit V̂ transform into qubit–qubit or qubit–photon couplings.
In essence, this is what we do in Sections 6.5.1 and 6.5.2 when introducing the dipolar
magnetic couplings. These interactions were treated as diagonal perturbations V̂d that took
place in the qubit space. However, this method allows us to consider the effect of the
interaction via the connection of the qubit eigenspace to higher excited states, or to take
the already developed interaction and understand how it can be approximated by other,
much simpler models – as we did in Section 6.5.3 and (6.67). Finally, perturbation the-
ory also provides us with effective interactions in limits, such as the dispersive qubit–
cavity coupling from Section 7.3.7, where the interactions are only mediated by excited
states and V̂d = 0.

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 276 Hamiltonian Diagonalizations

A word of caution is needed when applying the Schrieffer–Wolff transformation to

develop effective models. This relates to the fact that the eigenstates of the perturbed
problem result from deformations of the original qubit states, and that the Hamiltonian Ĥeff
is a unitary transformation of the Hamiltonian that actually rules the dynamics in the lab.
For instance, consider we have some state | that may be either an eigenstate of the
effective model
α
Ĥeff | = Er (g) | , (A.37)
or the solution of a Schrödinger equation at a given time:
i∂t |(t) = Ĥeff |(t) . (A.38)
In both cases, we will have those states expressed in the basis of non-interacting states:

| = n |n,α . (A.39)

However, this is a “fake” frame of reference. If at some point we need to make predictions
about the observables and properties of the state |, we will need to undo the Schrieffer–
Wolff transformation:

|lab  = e−g Ŝ |  n (1 − g Ŝ0 ) |n,α + O(g 2 ). (A.40)

In doing so, we see that the application or Ŝ or Ŝ0 deforms the basis in which we expanded
our state, from a superposition of the original bare states of the qubit or cavity, to contain
some hybridization with excitations |m,β injected by the Ŝ0 operator. The actual proba-
bility of those excitations grows as O(g 2 ), but it may be something to consider, especially
when doing high-precision experiments, such as quantum gates and quantum computations.

https://doi.org/10.1017/9781316779460.010 Published online by Cambridge University Press

 Appendix B
Open Quantum Systems

B.1 Nonunitary Evolution

The theory of open quantum systems studies the evolution of quantum mechanical systems
that are in contact with other larger systems, which we call environments. This theory
assumes the postulates of quantum mechanics, but understands that when we look at a piece
of a larger object, it is not described by a wavefunction but a reduced density matrix, and
does not satisfy a Schrödinger equation but some other type of linear differential equation
that describes its incoherent evolution.
As with any other treatment, the need to develop a theory of open systems is contingent
to the experimental timescales. We can argue that no quantum object is actually isolated –
for instance, all quantum objects are in contact with the electromagnetic field that permeates
our universe. However, when we perform an experiment with an atom, for example, the time
for that atom to get entangled with the universe – or even the laboratory, or its trap – may
be larger than our observation time. Under such circumstances, the formalism of quantum
mechanics and Schrödinger equations can provide an effective and correct description of
the experiment.
However, this book is full of experimental setups where we connect quantum circuits
to waveguides and measurement devices, which inject and extract energy and information
from qubits and cavities alike. Moreover, those circuits are themselves imperfect objects
that decohere due to interaction with the chip’s substrate, with trapped charges, and with
the surrounding electromagnetic fields. Both timescales, measurement and decoherence,
will be long, but not so long that they can be neglected.
When we cannot ignore the interaction between our experiment and the environment, we
have to describe our system as the reduced density matrix of a larger and more complicated
wavefunction, as explained in Section 2.3. The evolution of these density matrices is no
longer unitary, but described with a trace-preserving completely positive map. These are
linear transformations operating in the space of bounded linear operators that include den-
sity matrices. They are linear ε(αρ + βρ  ) = αε(ρ) + βε(ρ  ). They are positive1 because

1 The term completely positive is a technicality meaning that when we compose our Hilbert space with another
one, the extended operator remains positive. In other words, ε ⊗ 1B is positive for any dimension of the
identity 1B .

277

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 278 Open Quantum Systems

they transform nonnegative operators ρ ≥ 0 into nonnegative operators ε(ρ) ≥ 0. And

they are trace preserving because they maintain the normalization of density matrices,2
tr(ε(ρ)) = tr(ρ).
Most physical operations are trace-preserving completely positive maps that mutate den-
sity matrices into other density matrices. These superoperators – transformations of other
operators, in this case density matrices – admit a Kraus decomposition, a compact expres-
sion of the map as a sum of operators multiplying the density matrix:

n 
ε(ρ) = K̂i† ρ K̂i , with K̂i† K̂i = 1. (B.1)
i i

The matrices {K̂i , K̂i† } are called Kraus operators. The number of Kraus operators in a
decomposition can grow quadratically with the Hilbert space dimension. However, many
transformations require very few terms. These are then efficient representations of noise,
errors, and prototypical decohererence models – such as the ones in Problem 8.6 – that can
be experimentally calibrated and used to engineer error-correcting methods.
In the study of superconducting circuits, we are not so much concerned with specific
instances of a positive map. Our interest is more focused on the non-unitary evolution of
density matrices, given by differentiable maps ρ(t) = ε(ρ(0),t). Here the Kraus decompo-
sition is not so useful, as having a differential equation for ρ(t). Lindblad (1976) studied
such master equations and found that they are linear ordinary differential equations with a
common structure:
−1  
d 2

i 1
∂t ρ = Lρ = − [Ĥ,ρ] + lˆi ρ lˆi† − lˆi† lˆi ρ − ρ lˆi† lˆi . (B.2)
h̄ 2
i=1 i

The generator of the evolution L is the Lindblad superoperator. It separates into a Hamil-
tonian term −i[Ĥ,ρ]/h̄, and up to d 2 − 1 decoherence terms, where d is the dimension of
our system. Both the Hermitian operator Ĥ and the dissipators {lˆi , lˆi† } can depend on time,
but we will focus on the Markovian dynamics with constant dissipators. This is a relevant
situation that appears when the environment is very large and “forgets” very quickly the
history of the system with which it interacts. We will see in Section B.2 how in such cases
we can derive master equations, with very few Lindblad terms, that are both simple and intu-
√
itively appealing. Examples include Pauli operators lˆi ∼ γ σ − modeling the spontaneous
√
relaxation of a qubit (see Section 6.1.6), Fock operators lˆi ∼ κ â that model the cooling
of a microwave resonator in a zero-temperature environment (see Section 5.5.4), etc.

B.2 Master Equations

Our goal in this section is to derive a master equation for the reduced density matrix
of an open system. This equation arises when we trace out the environment in the
Schrödinger equation for the combined system-environment wavefunction. For that, we
assume a weak coupling between the system and the environment coupling, treating the
Schrödinger dynamics up to second order in perturbation theory. The result is an integro

2 The only circumstance where we would accept loss of trace would be when we want to model experiments
with postselection, where some fraction of outcomes are discarded. Then we could accept a decrease
in the norm, interpreting the trace of the density matrix as the total success probability of postselection.

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 B.2 Master Equations 279

differential equation that becomes local in time under a Markovian approximation – that the
environment forgets the history of our subsystem. With a slight massaging, we will end up
with a simple Lindblad equation that applies to microwave resonators and superconducting
qubits. But we will also offer prescriptions to study more complicated problems, such as
composite, driven, or highly nonlinear quantum systems.

B.2.1 Lindblad Equation

Assume a system–bath Hamiltonian that separates into a solvable term Ĥ0 and a small
perturbation Ĥ1 . The unitary operator for the combined system-bath wavefunction U (t)
satisfies a Schrödinger equation,
d
U (t) = (Ĥ0 + Ĥ1 )U (t),
i h̄ (B.3)
dt
with initial condition U (0) = 1. We can split the evolution operator into two factors U (t) =
exp(−i Ĥ0 t/h̄)W (t). The first factor is the exponential of the problem that we know how
to solve Ĥ0 . It becomes the exact solution when Ĥ1 vanishes. The second factor W is a
correction generated by the perturbation Ĥ1 . It satisfies the interaction picture equation:

i h̄Ẇ (t) = ei Ĥ0 t/h̄ Ĥ1 e−i Ĥ0 t/h̄ W (t) =: Ĥ1 (t)W (t). (B.4)
We can apply this interaction picture also to density matrices, splitting ρ(t) =
exp(−i Ĥ0 t/h̄)ρI (t) exp(i Ĥ0 t/h̄). The interaction picture ρ̃I satisfies a master equation
generated by the perturbations dtd ρI (t) = −i[Ĥ1 (t)/h̄,ρI (t)]. Usually, the dynamics gen-
erated by Ĥ1 (t) is much slower than the unitary evolution induced by exp(−i Ĥ0 t/h̄). This
justifies a Dyson series for the solution of the master equation, which we truncate up to
second order:

i t+δt  
ρI (t + δt)  ρ(t) − Ĥ1 (t  ),ρ(t) dt  (B.5)
h̄ t
 t+δt  t 
1   
− 2 Ĥ1 (t  ), Ĥ1 (t  ),ρ(t) dt  dt  .
h̄ t t
The next approximation is to neglect the feedback of the system onto the environment.
We impose that expectation values over the bath and system are uncorrelated, and that the
former are approximately constant.3 For an off-diagonal coupling4 TrB {Ĥ1 (t  )ρ̃(t)} = 0,
 t+δt  t 
1   
ρ̃S (t + δt)  ρ̃S (t) − 2 TrB Ĥ1 (t  ), Ĥ1 (t  ), ρ̃S (t) ⊗ ρ̃B dt  dt  .
h̄ t t

To continue further, we need to impose some structure in Ĥ1 . The following section
introduces a coupling that is typical of linear systems.

3 Traditionally, it has been said that the system and the environment adopt a product state structure where the
bath does not change, ρ̃(t) = ρ̃S (t) ⊗ ρ̃B , but this is not strictly needed.
4 We will see that this is not really a restriction, as typically Ĥ ∼ O ⊗ O , where the system and bath
1 S B
operators OS and OB have zero diagonal elements.

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 280 Open Quantum Systems

B.2.2 Linear System–Bath Coupling

Many superconducting circuits consist of a small system, such as a qubit or a cavity, which
is linearly coupled to a large number of bosonic operators, embodying the electromagnetic
enviroment, a transmission line, an antenna, or any reservoir the circuits thermalize to.
Such system–bath couplings take the form of products between system and bath operators
 and b̂k :
 
Ĥ1 = † gk b̂k +  gk∗ b̂k . (B.6)
k k

The Hamiltonian Ĥ0 typically has a simple effect on these operators:


ei Ĥ0 t/h̄ Âe−i Ĥ0 t/h̄ = eit  
−i
⇒ Ĥ1 (t) = † eit gk b̂k e−iωk t + H.c. (B.7)
e 0 b̂k e 0 = e k b̂k
i Ĥ t/h̄ Ĥ t/h̄ iω t
k

When we insert this expression into the formula (B.6), we obtain four combinations of
{Â, † } and {b̂k , b̂k† } acting on ρ̃S :
 t+δt
- .
ρ̃S = dt IB † B † L + IB † B L† + IBB † L†  + IBB L† † ρ̃S (t),
t
with LX̂Ŷ ρ̃ = 2X̂ ρ̃ Ŷ − Ŷ X̂ ρ̃ − ρ̃ Ŷ X̂. (B.8)
The prefactors are four integrals of expectation values over the bath:
  t+δt
! 
IBB = gp gp b̂p b̂q e−i(ωp −)t−i(ωq −)t = IB∗ † B † (B.9)
pq t

  t+δt ! 
IBB † = gp gp∗ b̂p b̂q† e−i(ωp −)t+i(ωq −)t ,
pq t

  t+δt ! 
IB † B = gp∗ gp b̂p† b̂q e+i(ωp −)t−i(ωq −)t .
pq t

For a large enough bath, these integrals and the superoperators are approximately constant.
The equation can be summarized as the action of one superoperator ρ̃S (t + δt) = ρ̃S (t) +
δt × Lρ̃S (t), which in the limit δt → 0 generates a Lindblad equation:
d
ρ̃S = Lρ̃S . (B.10)
dt

B.2.3 System in a Thermal Bath: Cooling and Heating

If we focus on thermal baths, only IBB † and IB † B survive, because

b̂q b̂p  = b̂q† b̂p†  = 0, b̂q† b̂p  = n̄(ωp )δq,p . (B.11)

All integrals can be expressed in terms of the spectral function (5.63), e.g.,
 t+δt   t+δt
dω 
IB † B = dt  J (ω)n̄(ω)ei(ω−)(t−t ) = dt  F (t − t  ). (B.12)
t 2π t

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 B.2 Master Equations 281

Following Walls and Milburn (1994), the Markov approximation enters by assuming that J
and n̄ are smooth functions, almost constant over a broad range of frequencies. As in
Section B.3.1, the function F (t − t  ) vanishes quickly beyond a short memory time τ  δt.
We can thus take the upper integration limits to ±∞, obtaining a distribution
 ∞  
±iτ 1
dτ e = π δ() ± iPV , (B.13)
0 
where τ = t − t  and  = ω − . Using this distribution, we conclude
γ γ
IB † B = n̄() + iδ , and IBB † = [n̄() + 1] + iδ, (B.14)
2 2
with the decay rate γ = J QO () and two different Lamb shifts δ  and δ that are usually
neglected. Inserting these values in (B.8) gives the usual quantum optics textbook master
equation:
d γ  
ρ̃S = [n̄() + 1] 2Âρ † − † Âρ − ρ †  (B.15)
dt  2 
γ
+ n̄() 2† ρ  − † ρ − ρ † ,
2
with the cooling and heating Lindblad operators. At this point, it is interesting to remark that
this equation can be brought back to the Schrödinger image, to compute the reduced density
matrix of the system in the laboratory frame, ρS (t) = trB ρ(t) = e−i Ĥ0 t/h̄ ρ̃S (t)ei Ĥ0 t/h̄ . In a
compact form,
d i γ γ
ρS = − [Ĥ0,ρS ] + [n̄() + 1] L† ρS + n̄() L†  ρS . (B.16)
dt h̄ 2 2

B.2.4 Perturbations and Generalizations

Our derivation assumed that the coupling operator  has a simple interaction picture (B.7).
This happens for resonators, where Â(t) = eit â is a bosonic operator (see Section 5.5.2),
and two-level systems, where Â(t) = eit σ − (Section 6.1.6). However, there are three
important problems that do not conform to this scheme: a quantum system under a
time-dependent control, a weakly nonlinear system, and a composite quantum system.
The canonical superconducting circuits for each case are the resonator with a coherent
microwave drive (5.8), the transmon qubit (6.34), and the Jaynes–Cummings model
(7.33) for a qubit in a microwave cavity. In these setups, it is too difficult to compute
 as we originally defined it. Instead, we use the part of Ĥ0 that is easy to build the
interaction picture. The corrections – driving (t)(â+ â † ), nonlinearity α â † 2 â 2 , or coupling
g(σ + â + σ − â † ) – are small terms that we incorporate into Ĥ1 and our perturbative
expansion.
We illustrate this with the qubit-resonator model (7.33). The cavity and the qubit are
coupled to two independent baths – the transmission line for the cavity and some source
of intrinsic losses for the qubit. We develop the interaction picture using Ĥ0 = h̄ω†  +
 †
2 h̄σ +
1 z
k h̄ωk b̂k b̂k and group all other terms into the “coupling” part:
 
Ĥ1 = gσ − â † + â † gk b̂k + σ − fk ĉk + H.c. (B.17)
k k

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 282 Open Quantum Systems

This perturbation contains three uncorrelated terms Ĥ1 = Ĥ1a + Ĥ1b + Ĥ1c , such that
tr(H1x H1y ρB ) = 0 if x = y. The Dyson series for ρ̃S has three independent contributions
that originate three Lindblad operators in the master equation dtd ρ̃S = (La + Lb + Lc )ρ̃S ,
one of which, La , will be a Hamiltonian contribution. A long and tedious but technically
simple process confirms this idea. It results in a model (7.34) with the original JC Hamil-
tonian, accompanied by dissipation terms for the cavity and the qubit. Note that the decay
QO QO
rates κ = Jb (ω) and γ = Jc () are evaluated in the spectral functions of the respective
independent environments. If the cavity and the qubit share the environment, the derivation
of the master equation will reveal interesting cross terms, Lâσ + ,Lâ † σ − , but these are out of
the scope of this book and have not been measured in experiments.

B.2.5 Strong Nonlinearity and Multilevel systems

When the system is strongly nonlinear or it is made of strongly coupled components, we
must use an alternative derivation of the master equation using the eigenstates of the system

Hamiltonian ĤS = n En |En En | . We begin expanding the coupling operator in this

basis  = m<n Amn |Em En | , with decay operators that connect states Em < En . The
system–bath coupling will enable all these transitions. In the interaction picture,

Ĥ1 (t) = A∗mn gk einm t−iωk t |En Em | b̂k† + H.c., (B.18)
m=n k

where h̄nm = En − Em > 0 are the positive frequency gaps. The analysis of Ĥ1 leads to a
small explosion of terms in (B.8), but they all support the same approximations. The result
is a generalized master equation:
↓
 mn ↑
d i mn
ρS = − [Ĥ0,ρS ] + Lnm ρS + Lmn ρS . (B.19)
dt h̄ m,n
2 2

The Lindblad operators model decay and excitation:

Lmn ρ = 2 |En Em | ρ |Em En | − |Em Em | ρ − ρ |En En | . (B.20)
The rates satisfy detailed balance and are proportional to the original coupling strengths,
that is,
↓
mn = [n̄(nm + 1)]|Amn |2 J QO (nm ), (B.21)
↑
nm = n̄(nm )|Amn | J 2 QO
(nm ).
Note how (B.19) includes the simpler model from (B.16). However, the new equa-
tion describes the ultrastrong coupling limit of a qubit in a microwave resonator (see
Section 7.4.2), something that the trivially corrected master equation from Section B.2.4
fails to do.

B.3 Input–Output Theory

We derive master equations by focusing on the system and ignoring its effect on the envir-
onment. This is often justified: When a superconducting qubit emits a microwave photon,
the contribution of this photon to the universe’s observables is tiny. This said, we may be
interested in that photon, in capturing it, measuring it, and extracting the information it
carries about the system that created it.

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 B.3 Input–Output Theory 283

In this section, we develop a formalism known as input–output theory that is comple-

mentary to the master equation approach. In this formalism, we study the observables of
a system and of its environment. Using the Heisenberg picture, we develop a stochastic
equation for the system (5.64), where the environment acts as a source of noise. We also
obtain an input–output relation (5.65) between the field of the environment before and
after it interacts with the system. The first equation makes prediction about the systems’s
dynamics, while the second equation tells us how to verify these predictions, measuring the
excitations that are emitted by the system when it decoheres. This is a powerful framework
that describes spectroscopy, amplifiers, and measurement devices, and which applies to
many superconducting circuit experiments.

B.3.1 Memory Function

Our study of input–output formalism starts with a resonator that interacts with a bosonic
environment. We use the RWA Hamiltonian (5.60) to derive Heisenberg equations for the
cavity and bath operators:
d  d
â = −iωâ − i gk∗ b̂k , and b̂k = −iωk b̂k − igk â. (B.22)
dt dt
k

The bath equations are solved implicitely, using initial conditions in the faraway past
t0 → −∞ or in the faraway future tf → +∞. The first type of solutions reads
 t
b̂k (t) = e−iωk (t−t0 ) b̂k (t0 ) − igk e−iωk (t−τ ) â(τ )dτ . (B.23)
t0

It contains the field leaking from the cavity â(t), together with the field that was injected
long ago into the line {bk (t0 )} and that is now reaching the cavity. Using this solution, we
obtain an integrodifferential equation for â(t) (5.61) that depends on (i) the bare properties
of the cavity, ω; (ii) the noise operator for the incoming field,

ξ̂ (t) = gk∗ e−iωk (t−t0 ) b̂k (t0 ); (B.24)
k

and (iii) the memory function of the bath K(t − τ ). The memory function is the Fourier
transform of the spectral function J QO (5.62). This function is broad and changes slowly
around the frequencies of interest. This produces a Fourier transform K(t) that concen-
trates around t = 0 and becomes zero past the memory time of the bath.
Let us illustrate this argument using a superconducting transmission line. The line has
a spectral function that is approximately Ohmic (i.e., linear) around the cavity resonance
J QO (ω̄) ∝ ω̄. In order to avoid ultraviolet divergencies, we regularize this function with an
exponential high-energy cutoff ωc constraining the range of frequencies that pass through
the guide. Apart from mathematical convenience, this is motivated by the filters on the
line and ultimately by the superconducting gap. Using this cutoff, we compute a memory
function:
αωc2
J QO (ω̄) = π α ω̄e−ω̄/ωc ⇒ K(t) = , (B.25)
(itωc + 1)2
which decays on a time of order tmemory  1/ωc . The memory time is the shortest timescale
in the problem – shorter than 2π /ω or than the time for the cavity to emit all its energy.

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 284 Open Quantum Systems

We talk about the Markovian regime as a limit in which the bath immediately loses all
memory of its interaction with the system. In this limit, the Markovian approximation is to
take the system operators as approximately constant during tmemory, replacing the memory
function with a Dirac delta K(t) ∼ δ(t).

B.3.2 Markovian Approximation: Decay Rate and Lamb Shift

The procedure to simplify (5.61) receives the name of separation of time scales. It rewrites
the cavity operator as a product of a slow modulation âslow and a fast oscillation exp(−iω t)
with unknown frequency ω . Introducing the variable u = t − τ ,
 t  t

K(t − τ )â(τ )dτ  âslow (t) K(t − τ )e−iω τ dτ (B.26)
t0 t0
 +∞  
iω u 1
= â(t) K(u)e du = iδω + κ â(t),
0 2
we simplify the whole integral down to a linear term that is local in time and depends on
the Lamb shift δω = ω − ω and the decay rate κ. We can further relate these constants to
the spectral function:
 ∞
γ 
iδω + = K(u)e−iω u du (B.27)
2 0
 ∞
1 
= J QO (ω)ei(ω −ω)u dωdu (B.28)
2π 0
  
1 1
= π δ(ω − ω) + iPV J QO (ω)dω. (B.29)
2π ω − ω
The real and imaginary parts of the last integral give us the decay rate of the cavity as the
spectral function evaluated at the dressed cavity resonance κ = J QO (ω ), and the Lamb
shift as the principal value part:
 −  +∞
1 J QO (ω + x)
ω − ω = − lim + dx, (B.30)
2π →0 −∞ + x
κ =J QO (ω ). (B.31)
These are implicit relations that have to be solved for ω and κ. However, in usual exper-
iments we only have access to ω , which is obtained after fitting spectra, and not to ω.
Hence, many quantum opticians ignore ω and define κ in terms of the “real frequency” of
the resonator.

B.3.3 Input–Output Relations

We still have not discussed the noise operator ξ̂ (t). To give some meaning to this term, we
are going to first solve the Langevin equation (5.64) in the asymptotic limit exp(κt0 ) → 0:
 e−iωk (t−t0 )
â(t) = e(−iω+κ/2)t â(0) − i g ∗ b̂k (t0 ). (B.32)
i(ω − ωk ) + κ/2 k
k

Note how the oscillator feeds only from a narrow band κ around the resonance ω . This band
is so small that we can assume that the coupling strength gk , the spectral function J QO (ω),

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 B.3 Input–Output Theory 285

and the density of states are more or less constant. Applying this idea to the study of the
input operators gives
 
†  2 −iωk (t−t  ) 
[ξ̂in (t), ξ̂in (t )] = |gk | e = dωJ QO (ω)e−iω(t−t ) (B.33)
k


 J QO (ω ) dωe−iω(t−t ) = κδ(t − t  ). (B.34)
√
We replace the input operators with new ones b̂in (t) = ξ̂in / κ that satisfy causal com-
mutation relations, producing the Langevin equation (5.64). Note that the cavity opera-
tor â(t) is dimensionless, but κ b̂in † b̂in has dimensions of photon flux [time]−1 . Hence
[b̂in ] = [time]−1/2 .
We can repeat the whole preceding study, but using initial conditions that are far away
in the future, tf → +∞. The equation is slightly different:
d √
â = (−iω + κ/2)â − i κ b̂out, (B.35)
dt
and now involves the field reflected by the cavity b̂out . The problem with this equation is that
we do not really know the state of the environment in the future, b̂k (+∞), so that we cannot
compute b̂out . Fortunately, we can guess b̂out indirectly, realizing that both boundary condi-
tions, b̂out and b̂in, must give the same value for the field experienced by the cavity (B.23):
  t
∗ √ in
gk bk (t) = κ b̂ (t) − i K(t − τ )â(τ )dτ (B.36)
k −∞
 +∞
√
= κ b̂out (t) + i K(t − τ )â(τ )dτ . (B.37)
0

Combining both equations and using R K(t − τ )â(τ )dτ  κ â(t), we arrive at the input–
output relation:
√
b̂out (t) = b̂in (t) − i κa(t). (B.38)
Let us close this section with some remarks. The first one is to point out that the input–
output relations vary with how we write the coupling Hamiltonian. If instead of (5.60) we
use h̄i(gk abk† −gk∗ bk a † ), this amounts to a phase shift â → i â that suppresses some factors:
d √
â(t) = (−iω − κ/2)â(t) − κ b̂in (t), (B.39)
dt
√
b̂out (t) = b̂in (t) + κ â(t). (B.40)
Another important variation is when we want to consider multiple input and output chan-
nels. This happens in transmission experiments, when the LC resonator is inserted in the
middle of a line, or when it is connected to multiple lines. In that case, the input–output
equations read
   √
d
â(t) = −iω − κ /2 â(t) − i
r
κ r b̂in (t), (B.41)
dt r n
√
b̂rout (t) = b̂rin (t) − i κ r â(t). (B.42)
The index r = 1,2, . . . R runs over all input–output channels, which may have very different
decay rates κ r .

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 286 Open Quantum Systems

B.3.4 Spectroscopy
Throughout this book, we offer many examples of spectroscopy: experiments where we
illuminate a quantum system with microwaves and study the radiation that it absorbs, emits,
and scatters. Such experiments fit naturally in the input–output framework that we just
developed, provided it is extended to treat more general setups, such as qubits, cavity-QED,
and other composite and nonlinear systems.
The generalization of input–output theory parallels Section B.2.5. We consider an arbi-
trary quantum system with discrete eigenstates, coupled to R input–output channels:

R 
†
Ĥ1 = Armn gr,k |En Em | b̂r,k + H.c. (B.43)
r=1 m=n k

Applying the same treatment and approximations as before, we arrive at a series of input–
output relations:

b̂rout (t) = b̂rin (t) − i γmn
r σ̂ ,
mn (B.44)
with system ladder operators σ̂mn (t = 0) = |Em En | that feed the different lines at
different rates γmn r ∝ |Ar |2 J QO ( ).
mn nm
Studying the evolution of the system operators σ̂mn is rather complicated, and it is rarely
done in the Heisenberg picture. Most often, we are just interested in the expectation val-
ues and correlations of the output fields, b̂rout  . In this case, we can get away with any
method to estimate quantities such as σ̂mn (t) . The most popular method is to calculate the
coherences using a master equation such as (B.19), adapted to reproduce the conditions of
a spectroscopy protocol.
The simulations presume that the waveguides are close to zero temperature and host some
input fields b̂rin (t) ∼ r (t). The Hamiltonian part of the master equation accommodates
the coherent driving with which we illuminate our system. The losses are extended to add up
all spectroscopy channels and the intrinsic decoherence of our sample. We typically neglect
heating. In other words,
  
Ĥ0 = Em |Em Em | + γmnr  (t)∗ |E E | + H.c., (B.45)
r m n
m r,m,n

↓ ↑
mn = int
γmn + r
γmn and nm  0.
r

In the common case in wich we drive monochromatically through one channel

0 (t) ∼ 0 exp(−iωt), this model predicts a Lorentzians response in all channels:
)  *
r γ0
γmn
−iωt mn
b̂r (t) ∼ 0 e
out
δr0 − . (B.46)
i(ω − nm ) + mn /2

The Lorentzians are centered on the different transition frequencies mn . The weight and
the linewidth of these resonances give information on the decoherence rates and on the
transition strengths, mn ∝ |Amn |2 . This is a lot of information that can be used to con-
firm and fit our microscopic models, such as the strong coupling and ultrastrong coupling
Hamiltonians from our circuit-QED models (Section 7.4.1).

https://doi.org/10.1017/9781316779460.011 Published online by Cambridge University Press

 References

Abdo, B., Suchoi, O., Segev, E., et al. 2009. Intermodulation and parametric amplification in
a superconducting stripline resonator integrated with a dc-SQUID. EPL (Europhysics
Letters), 85(6), 68001.
Adesso, Gerardo, Ragy, Sammy, and Lee, Antony R. 2014. Continuous variable quantum
information: Gaussian states and beyond. Open Systems & Information Dynamics,
21(01–02), 1440001.
Aharonov, Dorit, van Dam, Wim, Kempe, Julia, Landau, Zeph, Lloyd, Seth, and Regev,
Oded. 2004. Adiabatic quantum computation is equivalent to standard quantum
computation. SIAM Review, 50(4), 755–787.
Albash, Tameem, and Lidar, Daniel A. 2018. Demonstration of a scaling advantage for a
quantum annealer over simulated annealing. Physical Review X 8, 031016 (2018).
Albash, Tameem, Vinci, Walter, Mishra, Anurag, Warburton, Paul A., and Lidar, Daniel A.
2015. Consistency tests of classical and quantum models for a quantum annealer.
Physical Review A 91, 042314 (2015).
Anderson, P. W., and Rowell, J. M. 1963. Probable observation of the Josephson supercon-
ducting tunneling effect. Physical Review Letters, 10(6), 230–232.
Arute, Frank, Arya, Kunal, Babbush, Ryan, et al. 2019. Quantum supremacy using a
programmable superconducting processor. Nature, 574(7779), 505–510.
Arute, Frank, Arya, Kunal, Babbush, Ryan, et al. 2020. Hartree–Fock on a superconducting
qubit quantum computer. Science, 369(6507), 1084–1089.
Astafiev, O., Zagoskin, A. M., Abdumalikov, A. A., et al. 2010. Resonance fluorescence of
a single artificial atom. Science (New York, N.Y.), 327(5967), 840–843.
Ballentine, L. E. 1970. The statistical interpretation of quantum mechanics. Reviews of
Modern Physics, 42(4), 358–381.
Ballentine, Leslie E. 1998. Quantum Mechanics. World Scientific.
Bapst, V., Foini, L., Krzakala, F., Semerjian, G., and Zamponi, F. 2013. The quantum
adiabatic algorithm applied to random optimization problems: the quantum spin glass
perspective. Physics Reports, 523(3), 127–205.
Barahona, F. 1982. On the computational complexity of Ising spin glass models. Journal of
Physics A: Mathematical and General, 15(10), 3241–3253.
Bardeen, J., Cooper, L. N., and Schrieffer, J. R. 1957a. Microscopic theory of superconduc-
tivity. Physical Review, 106(1), 162–164.
Bardeen, J., Cooper, L. N., and Schrieffer, J. R. 1957b. Theory of superconductivity.
Physical Review, 108(5), 1175–1204.

287

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 288 References

Barends, R., Kelly, J., Megrant, A., et al. 2014. Superconducting quantum circuits at the
surface code threshold for fault tolerance. Nature, 508(7497), 500–503.
Baumgratz, T., Cramer, M., and Plenio, M. B. 2014. Quantifying coherence. Physical
Review Letters, 113(14), 140401.
Benioff, Paul. 1980. The computer as a physical system: a microscopic quantum mechanical
Hamiltonian model of computers as represented by Turing machines. Journal of
Statistical Physics, 22(5), 563–591.
Bennett, Douglas A., Longobardi, Luigi, Patel, Vijay, Chen, Wei, and Lukens, James E.
2007. rf-SQUID qubit readout using a fast flux pulse. Superconductor Science and
Technology, 20(11), S445–S449.
Berkley, A. J., Johnson, M. W., Bunyk, P., et al. 2010. A scalable readout system for
a superconducting adiabatic quantum optimization system. Superconductor Science
and Technology, 23(10), 105014.
Berkley, A. J., Przybysz, A. J., Lanting, T., et al. 2013. Tunneling spectroscopy using a
probe qubit. Physical Review B, 87(2), 020502.
Bialczak, R. C., Ansmann, M., Hofheinz, M., et al. 2010. Quantum process tomography of a
universal entangling gate implemented with Josephson phase qubits. Nature Physics,
6(6), 409–413.
Biamonte, Jacob D., and Love, Peter J. 2008. Realizable Hamiltonians for universal
adiabatic quantum computers. Physical Review A, 78(1), 012352.
Bishop, Lev S., Bravyi, Sergey, Gambetta, Jay M., and Smolin, John. 2017. Quantum
Volume. Technical Report.
Blais, Alexandre, Huang, Ren-Shou, Wallraff, Andreas, Girvin, S., and Schoelkopf, R.
2004. Cavity quantum electrodynamics for superconducting electrical circuits: an
architecture for quantum computation. Physical Review A, 69(6), 062320.
Boixo, Sergio, Albash, Tameem, Spedalieri, Federico M., Chancellor, Nicholas, and Lidar,
Daniel A. 2013. Experimental signature of programmable quantum annealing. Nature
Communications, 4(1), 2067.
Boixo, Sergio, Rønnow, Troels F., Isa kov, Sergei V., et al. 2014. Evidence for quantum
annealing with more than one hundred qubits. Nature Physics, 10(3), 218–224.
Boixo, Sergio, Smelyanskiy, Vadim N., Shabani, Alireza, et al. 2016. Computational
multiqubit tunnelling in programmable quantum annealers. Nature Communications,
7(Jan), 10327.
Bombin, H., and Martin-Delgado, M. A. 2006. Topological Quantum Distillation. Physical
Review Letters, 97(Oct), 180501.
Bouchiat, V., Vion, D., Joyez, P., Esteve, D., and Devoret, M. H. 1998. Quantum coherence
with a single Cooper pair. Physica Scripta, T76(1), 165–170.
Bourassa, J., Gambetta, J., Abdumalikov, A., Astafiev, O., Nakamura, Y., and Blais, A. 2009.
Ultrastrong coupling regime of cavity QED with phase-biased flux qubits. Physical
Review A, 80(3), 032109.
Bravyi, S. B., and Kitaev, A. Yu. 1998. Quantum codes on a lattice with boundary. e-print
arXiv:quant-ph-9811052.
Bravyi, Sergey, DiVincenzo, David P., and Loss, Daniel. 2011. Schrieffer–Wolff transfor-
mation for quantum many-body systems. Annals of Physics, 326(10), 2793–2826.
Brecht, T., Chu, Y., Axline, C., et al. 2017. Micromachined integrated quantum circuit
containing a superconducting qubit. Physics Review Applied, 7(Apr), 044018.
Brecht, Teresa, Pfaff, Wolfgang, Wang, et al. 2016. Multilayer microwave integrated
quantum circuits for scalable quantum computing. npj Quantum Information, 2(1),
16002.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 References 289

Brink, Markus, Corcoles-Gonzalez, Antonio, Gambetta, Jay M., Rosenblatt, Sami, and
Solgun, Firat. 2019 (May 28). Low loss architecture for superconducting qubit
circuits. US Patent 10,305,015.
Brown, K. R., Wilson, A. C., Colombe, Y., et al. 2011. Single-qubit-gate error below 10−4
in a trapped ion. Physics Review A, 84(Sep), 030303.
Cai, Jun, Macready, William G., and Roy, Aidan. 2014. A practical heuristic for finding
graph minors. e-print arXiv:1406.2741.
Calderbank, A. R., and Shor, Peter W. 1996. Good quantum error-correcting codes exist.
Physical Review A, 54(Aug), 1098–1105.
Castellanos-Beltran, M. A., Irwin, K. D., Vale, L. R., Hilton, G. C., and Lehnert, K. W. 2009.
Bandwidth and dynamic range of a widely tunable Josephson parametric amplifier.
IEEE Transactions on Applied Superconductivity, 19(3), 944–947.
Chancellor, N., Zohren, S., and Warburton, P. A. 2017. Circuit design for multi-body
interactions in superconducting quantum annealing systems with applications to a
scalable architecture. npj Quantum Information, 3(1), 21.
Chen, Yu, Neill, C., Roushan, P., et al. 2014. Qubit architecture with high coherence and
fast tunable coupling. Physical Review Letters, 113(Nov), 220502.
Chen, Zijun, Satzinger, Kevin J., Atalaya, Juan, et al. 2021. Exponential suppression of bit
or phase errors with cyclic error correction. Nature, 595(7867), 383–387.
Chiorescu, I., Bertet, P., Semba, K., Nakamura, Y., Harmans, C. J. P. M., and Mooij, J. E.
2004. Coherent dynamics of a flux qubit coupled to a harmonic oscillator. Nature,
431(7005), 159–162.
Chow, J. M., Gambetta, J. M., Tornberg, L., et al. 2009. Randomized benchmarking and
process tomography for gate errors in a solid-state qubit. Physical Review Letters,
102(9), 090502.
Chow, Jerry M., Córcoles, A. D., Gambetta, Jay M., et al. 2011. Simple all-microwave
entangling gate for fixed-frequency superconducting qubits. Physical Review Letters,
107(8), 080502.
Chow, Jerry M., Gambetta, Jay M., Córcoles, A. D., et al. 2012. Universal quantum gate set
approaching fault-tolerant thresholds with superconducting qubits. Physical Review
Letters, 109(6), 060501.
Cirac, J. I., and Zoller, P. 1995. Quantum computations with cold trapped ions. Physical
Review Letters, 74(20), 4091–4094.
Cirac, J. I., Ekert, A. K., Huelga, S. F., and Macchiavello, C. 1999. Distributed quantum
computation over noisy channels. Physical Review A, 59(6), 4249–4254.
Clarke, John, and Braginski, Alex I. (eds). 2004. The SQUID Handbook: Fundamentals and
Technology of SQUIDs and SQUID Systems, Volume I. Wiley-VCH.
Clarke, John, and Wilhelm, Frank K. 2008. Superconducting quantum bits. Nature,
453(7198), 1031–1042.
Cohen-Tannoudji, Claude, Diu, Bernard, and Laloë, Franck. 1977. Quantum Mechanics.
Wiley.
Consani, Gioele, and Warburton, Paul A. 2020. Effective Hamiltonians for interacting super-
conducting qubits: local basis reduction and the Schrieffer–Wolff transformation. New
Journal of Physics, 22(5), 053040.
Cooper, Leon N. 1956. Bound electron pairs in a degenerate Fermi gas. Physical Review,
104(4), 1189–1190.
Cory, D. G., Fahmy, A. F., and Havel, T. F. 1997. Ensemble quantum computing by NMR
spectroscopy. Proceedings of the National Academy of Sciences of the United States
of America, 94(5), 1634–1639.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 290 References

Cross, Andrew W., Bishop, Lev S., Sheldon, Sarah, Nation, Paul D., and Gambetta, Jay M.
2019. Validating quantum computers using randomized model circuits. Physical
Review A, 100(Sep), 032328.
Cullen, A. L. 1960. Theory of the travelling-wave parametric amplifier. Proceedings of the
IEE Part B: Electronic and Communication Engineering, 107(32), 101.
Deaver, Bascom S., and Fairbank, William M. 1961. Experimental evidence for quantized
flux in superconducting cylinders. Physical Review Letters, 7(2), 43–46.
Denchev, Vasil S., Boixo, Sergio, Isakov, Sergei V., et al. 2016. What is the computational
value of finite-range tunneling? Physical Review X, 6(3), 031015.
Deutsch, David Elieser. 1985. Quantum theory, the Church–Turing principle and the
universal quantum computer. Proceedings of the Royal Society A: Mathematical,
Physical and Engineering Sciences, 400(1818), 97–117.
Deutsch, David Elieser. 1989. Quantum computational networks. Proceedings of the Royal
Society of London. A. Mathematical and Physical Sciences, 425(1868), 73–90.
Devoret, M. H. 1995. Quantum fluctuations in electrical circuits. Les Houches, Session
LXIII, 351–386.
DiCarlo, L., Chow, J. M., Gambetta, J. M., et al. 2009. Demonstration of two-qubit
algorithms with a superconducting quantum processor. Nature, 460(7252), 240–244.
DiCarlo, L., Reed, M. D., Sun, L., et al. 2010. Preparation and measurement of three-qubit
entanglement in a superconducting circuit. Nature, 467(7315), 574–578.
Dickson, N. G., Johnson, M. W., Amin, M. H., et al. 2013. Thermally assisted quantum
annealing of a 16-qubit problem. Nature Communications, 4(1), 1903.
Dieks, D. 1982. Communication by EPR devices. Physics Letters A, 92(6), 271–272.
DiVincenzo, D. P. 1995. Quantum computation. Science, 270(5234), 255–261.
DiVincenzo, David P. 2000. The physical implementation of quantum computation.
Fortschritte der Physik, 48(9-11), 771–783.
Dowling, Jonathan P., and Milburn, Gerard J. 2003. Quantum technology: the second
quantum revolution. Philosophical Transactions of the Royal Society of London.
Series A: Mathematical, Physical and Engineering Sciences, 361(1809), 1655–1674.
Egger, D. J., Werninghaus, M., Ganzhorn, M., et al. 2018. Pulsed reset protocol for
fixed-frequency superconducting qubits. Physical Review Applied, 10(Oct), 044030.
Eichler, C., Bozyigit, D., and Wallraff, A. 2012. Characterizing quantum microwave
radiation and its entanglement with superconducting qubits using linear detectors.
Physical Review A, 86(3), 032106.
Eichler, C., Bozyigit, D., Lang, C., Steffen, L., Fink, J., and Wallraff, A. 2011. Experimental
state tomography of itinerant single microwave photons. Physical Review Letters,
106(22), 220503.
Farhi, Edward, Goldstone, Jeffrey, Gutmann, Sam, and Sipser, Michael. 2000. Quantum
computation by adiabatic evolution. arXiv, quant-ph:0(jan).
Feynman, Richard P. 1982. Simulating physics with computers. International Journal of
Theoretical Physics, 21(6-7), 467–488.
Forn-Díaz, P., Lisenfeld, J., Marcos, D., et al. 2010. Observation of the Bloch–Siegert
shift in a qubit-oscillator system in the ultrastrong coupling regime. Physical Review
Letters, 105(23), 237001.
Forn-Díaz, P., García-Ripoll, J. J., Peropadre, B., et al. 2016. Ultrastrong coupling of a single
artificial atom to an electromagnetic continuum in the nonperturbative regime. Nature
Physics, 13(1), 39–43.
Forn-Díaz, P., Warren, C. W., Chang, C. W. S., Vadiraj, A. M., and Wilson, C. M. 2017.
On-demand microwave generator of shaped single photons. Physical Review Applied,
8(5), 054015.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 References 291

Fowler, Austin G., Mariantoni, Matteo, Martinis, John M., and Cleland, Andrew N. 2012.
Surface codes: towards practical large-scale quantum computation. Physical Review
A, 86(Sep), 032324.
Fowler, Austin G., Stephens, Ashley M., and Groszkowski, Peter. 2009. High-threshold
universal quantum computation on the surface code. Physical Review A, 80(Nov),
052312.
Gaebler, J. P., Meier, A. M., Tan, T. R., et al. 2012. Randomized benchmarking of multiqubit
gates. Physical Review Letters, 108(26), 260503.
Galindo, Alberto, and Pascual, Pedro. 1991. Time-dependent perturbation theory. Pages
161–199 of: Quantum Mechanics II. Springer Berlin Heidelberg.
García-Ripoll, J. J., Peropadre, B., and De Liberato, S. 2015. Light–matter decoupling and
A2 term detection in superconducting circuits. Scientific Reports, 5(1), 16055.
García-Ripoll, J. J., Ruiz-Chamorro, A., and Torrontegui, E. 2020. Quantum control of
frequency-tunable transmon superconducting qubits. Physical Review Applied, 14(4),
044035.
Gardiner, C. W., and Zoller, P. 2004. Quantum Noise. 3rd ed. Berlin, Heidelberg: Springer-
Verlag Berlin Heidelberg.
Gershenfeld, Neil A., and Chuang, Isaac L. 1997. Bulk spin-resonance quantum computa-
tion. Science (New York, N.Y.), 275(5298), 350–356.
Gilchrist, Alexei, Langford, Nathan K., and Nielsen, Michael A. 2005. Distance measures
to compare real and ideal quantum processes. Physical Review A, 71(6), 062310.
Gor’kov, Lev P. 1959. Microscopic derivation of the Ginzburg–Landau equations in the
theory of superconductivity. JETP 36(9), 1364.
Gottesman, Daniel. 1998. Theory of fault-tolerant quantum computation. Physical Review
A, 57(Jan), 127–137.
Haroche, Serge. 2013. Nobel lecture: controlling photons in a box and exploring the
quantum to classical boundary. Reviews of Modern Physics, 85(3), 1083–1102.
Harris, R., Berkley, A. J., Johnson, M. W., et al. 2007. Sign- and magnitude-tunable coupler
for superconducting flux qubits. Physical Review Letters, 98(17), 177001.
Harris, R., Johansson, J., Berkley, A. J., et al. 2010a. Experimental demonstration of a robust
and scalable flux qubit. Physical Review B, 81(13), 134510.
Harris, R., Johnson, M. W., Lanting, T., et al. 2010b. Experimental investigation of an
eight-qubit unit cell in a superconducting optimization processor. Physical Review
B, 82(2), 024511.
Harris, R., Lanting, T., Berkley, A. J., et al. 2009. Compound Josephson-junction coupler
for flux qubits with minimal crosstalk. Physical Review B, 80(5), 052506.
Harris, R., Sato, Y., Berkley, A. J., et al. 2018. Phase transitions in a programmable quantum
spin glass simulator. Science, 361(6398), 162–165.
Hauke, Philipp, Katzgraber, Helmut G, Lechner, Wolfgang, Nishimori, Hidetoshi, and
Oliver, William D. 2020. Perspectives of quantum annealing: methods and implemen-
tations. Reports on Progress in Physics, 83(5), 054401.
Heisenberg, W. 1925. Über quantentheoretische Umdeutung kinematischer und mechanis-
cher Beziehungen. Zeitschrift für Physik, 33(1), 879–893.
Hime, T., Reichardt, P. A., Plourde, B. L. T., et al. 2006. Solid-state qubits with current–
controlled coupling. Science (New York, N.Y.), 314(5804), 1427–1429.
Hita-Pérez, María, Jaumá, Gabriel, Pino, Manuel, and García-Ripoll, Juan José. 2022.
Ultrastrong capacitive coupling of flux qubits. Physical Review Applied 17, 014028.
Hofheinz, Max, Wang, H., Ansmann, M., et al. 2009. Synthesizing arbitrary quantum states
in a superconducting resonator. Nature, 459(7246), 546–549.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 292 References

Hoi, Io-Chun, Kockum, Anton F., Palomaki, Tauno, et al. 2013. Giant Cross–Kerr effect
for propagating microwaves induced by an artificial atom. Physical Review Letters,
111(5), 053601.
Hoi, Io-Chun, Palomaki, Tauno, Lindkvist, Joel, Johansson, Göran, Delsing, Per, and
Wilson, C. M. 2012. Generation of nonclassical microwave states using an artificial
atom in 1d open space. Physical Review Letters, 108(26), 263601.
Hoi, Io-Chun, Wilson, C. M., Johansson, Göran, Palomaki, Tauno, Peropadre, Borja, and
Delsing, Per. 2011. Demonstration of a single-photon router in the microwave regime.
Physical Review Letters, 107(7), 073601.
Jansen, Sabine, Ruskai, Mary-Beth, and Seiler, Ruedi. 2007. Bounds for the adiabatic
approximation with applications to quantum computation. Journal of Mathematical
Physics, 48(10), 102111.
Jaynes, E. T., and Cummings, F. W. 1963. Comparison of quantum and semiclassical
radiation theories with application to the beam maser. Proceedings of the IEEE, 51(1),
89–109.
Jeffrey, Evan, Sank, Daniel, Mutus, J. Y., et al. 2014. Fast accurate state measurement with
superconducting qubits. Physical Review Letters, 112(19), 190504.
Johnson, M. W., Amin, M. H. S., Gildert, S., et al. 2011. Quantum annealing with
manufactured spins. Nature, 473(7346), 194–198.
Johnson, M. W., Bunyk, P., Maibaum, F., et al. 2010. A scalable control system for a
superconducting adiabatic quantum optimization processor. Superconductor Science
and Technology, 23(6), 065004.
Josephson, B. D. 1962. Possible new effects in superconductive tunnelling. Physics Letters,
1(7), 251–253.
Kadowaki, Tadashi, and Nishimori, Hidetoshi. 1998. Quantum annealing in the transverse
Ising model. Physical Review E, 58(5), 5355–5363.
Kandala, Abhinav, Temme, Kristan, Córcoles, Antonio D., Mezzacapo, Antonio, Chow,
Jerry M., and Gambetta, Jay M. 2019. Error mitigation extends the computational
reach of a noisy quantum processor. Nature, 567(7749), 491–495.
Kato, Tosio. 1950. On the adiabatic theorem of quantum mechanics. Journal of the Physical
Society of Japan, 5(6), 435–439.
Kelly, J., Barends, R., Fowler, A. G., et al. 2015. State preservation by repetitive error
detection in a superconducting quantum circuit. Nature, 519(7541), 66–69.
Kempe, Julia, Kitaev, Alexei, and Regev, Oded. 2006. The complexity of the local Hamil-
tonian problem. SIAM Journal on Computing, 35(5), 1070–1097.
King, Andrew D., Carrasquilla, Juan, Raymond, Jack, et al. 2018. Observation of topo-
logical phenomena in a programmable lattice of 1,800 qubits. Nature, 560(7719),
456–460.
Kitaev, A., Shen, A., and Vyalyi, M. 2002. Classical and Quantum Computation. Graduate
Studies in Mathematics. American Mathematical Society.
Knill, E., Leibfried, D., Reichle, R., et al. 2008. Randomized benchmarking of quantum
gates. Physical Review A, 77(1), 012307.
Koch, Jens, Yu, Terri, Gambetta, Jay, et al. 2007. Charge-insensitive qubit design derived
from the Cooper pair box. Physical Review A, 76(4), 042319.
Krinner, S., Lazar, S., Remm, A., et al. 2020. Benchmarking coherent errors in con-
trolled-phase gates due to spectator qubits. Physical Review Applied, 14(2), 024042.
Kurcz, Andreas, Bermudez, Alejandro, and García-Ripoll, Juan José. 2014. Hybrid quantum
magnetism in circuit QED: from spin-photon waves to many-body spectroscopy.
Physical Review Letters, 112(18), 180405.
Landau, L. 1932. Zur theorie der energieubertragung II. Physik. Z. Sowjet, 2, 46–50.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 References 293

Lanting, T., Przybysz, A. J., Smirnov, A. Yu., et al. 2014. Entanglement in a quantum
annealing processor. Physical Review X, 4(2), 021041.
Laurat, Julien, Keller, Gaëlle, Oliveira-Huguenin, José Augusto, et al. 2005. Entanglement
of two-mode Gaussian states: characterization and experimental production and
manipulation. Journal of Optics B: Quantum and Semiclassical Optics, 7(12), S577–
S587.
Law, C. K., and Eberly, J. H. 1996. Arbitrary control of a quantum electromagnetic field.
Physical Review Letters, 76(7), 1055–1058.
Leggett, A., Chakravarty, S., Dorsey, A., Fisher, Matthew, Garg, Anupam, and Zwerger,
W. 1987. Dynamics of the dissipative two-state system. Reviews of Modern Physics,
59(1), 1–85.
Leibfried, D., Blatt, R., Monroe, C., and Wineland, D. 2003. Quantum dynamics of single
trapped ions. Reviews of Modern Physics, 75(1), 281–324.
Lidar, Daniel A., Rezakhani, Ali T., and Hamma, Alioscia. 2009. Adiabatic approximation
with exponential accuracy for many-body systems and quantum computation. Journal
of Mathematical Physics, 50(10), 102106.
Lindblad, G. 1976. On the generators of quantum dynamical semigroups. Communications
in Mathematical Physics, 48(2), 119–130.
London, F., London, H., and Lindemann, Frederick Alexander. 1935. The electromagnetic
equations of the supraconductor. Proceedings of the Royal Society of London. Series
A - Mathematical and Physical Sciences, 149(866), 71–88.
Loss, Daniel, and DiVincenzo, David P. 1998. Quantum computation with quantum dots.
Physical Review A, 57(1), 120–126.
Macklin, C., O’Brien, K., Hover, D., et al. 2015. A near-quantum-limited Josephson
traveling-wave parametric amplifier. Science (New York, N.Y.), 350(6258), 307–310.
Magesan, Easwar, Gambetta, J. M., and Emerson, Joseph. 2011. Scalable and robust
randomized benchmarking of quantum processes. Physical Review Letters, 106(18),
180504.
Magesan, Easwar, Gambetta, Jay M., Johnson, B. R., et al. 2012. Efficient measurement
of quantum gate error by interleaved randomized benchmarking. Physical Review
Letters, 109(8), 080505.
Magnard, P., Kurpiers, P., Royer, B., et al. 2018. Fast and unconditional all-microwave reset
of a superconducting qubit. Physical Review Letters, 121(Aug), 060502.
Magnard, P., Storz, S., Kurpiers, P., et al. 2020. Microwave quantum link between super-
conducting circuits housed in spatially separated cryogenic systems. Physical Review
Letters, 125(26), 260502.
Majer, J., Chow, J. M., Gambetta, J. M., et al. 2007. Coupling superconducting qubits via a
cavity bus. Nature, 449(7161), 443–447.
Makhlin, Yuriy, Schön, Gerd, and Shnirman, Alexander. 2001. Quantum-state engineering
with Josephson-junction devices. Reviews of Modern Physics, 73(2), 357–400.
Mallet, F., Castellanos-Beltran, M. A., Ku, H. S., et al. 2011. Quantum state tomography of
an itinerant squeezed microwave field. Physical Review Letters, 106(22), 220502.
Mandrá, Salvatore, and Katzgraber, Helmut G. 2018. A deceptive step towards quantum
speedup detection. Quantum Science and Technology, 3(4), 04LT01.
Mandrá, Salvatore, Zhu, Zheng, Wang, Wenlong, Perdomo-Ortiz, Alejandro, and Katz-
graber, Helmut G. 2016. Strengths and weaknesses of weak-strong cluster problems:
a detailed overview of state-of-the-art classical heuristics versus quantum approaches.
Physical Review A, 94(2), 022337.
McEwen, M., Kafri, D., Chen, Z., et al. 2021. Removing leakage-induced correlated errors
in superconducting quantum error correction. Nature Communications, 12(1), 1761.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 294 References

Menzel, E. P., Deppe, F., Mariantoni, M., et al. 2010. Dual-path state reconstruction scheme
for propagating quantum microwaves and detector noise tomography. Physical Review
Letters, 105(10), 100401.
Menzel, E. P., Di Candia, R., Deppe, F., et al. 2012. Path entanglement of continuous-
variable quantum microwaves. Physical Review Letters, 109(25), 250502.
Merkel, Seth T., Gambetta, Jay M., Smolin, John A., et al. 2013. Self-consistent quantum
process tomography. Physical Review A, 87(Jun), 062119.
Mi, Xiao, Ippoliti, Matteo, Quintana, Chris, Greene, et al. 2022. Observation of time-
crystalline eigenstate order on a quantum processor. Nature 601, 531.
Mirrahimi, Mazyar, Leghtas, Zaki, Albert, et al. 2014. Dynamically protected cat-qubits:
a new paradigm for universal quantum computation. New Journal of Physics, 16(4),
045014.
Mlynek, J. A., Abdumalikov, A. A., Eichler, C., and Wallraff, A. 2014. Observation of
Dicke superradiance for two artificial atoms in a cavity with high decay rate. Nature
Communications, 5(1), 5186.
Mooij, J. E., Orlando, T. P., Levitov, L., Tian, L., van der Wal, C. H., and Lloyd, S. 1999.
Josephson persistent-current qubit. Science, 285(5430), 1036–1039.
Nakamura, Y., Chen, C. D., and Tsai, J. S. 1997. Spectroscopy of energy-level splitting
between two macroscopic quantum states of charge coherently superposed by Joseph-
son coupling. Physical Review Letters, 79(12), 2328–2331.
Nakamura, Y., Pashkin, Yu. A., and Tsai, J. S. 1999. Coherent control of macroscopic
quantum states in a single-Cooper-pair box. Nature, 398(6730), 786–788.
Nataf, Pierre, and Ciuti, Cristiano. 2010. No-go theorem for superradiant quantum phase
transitions in cavity QED and counter-example in circuit QED. Nature Communica-
tions, 1(6), 1–6.
Neeley, Matthew, Ansmann, M., Bialczak, Radoslaw C., et al. 2008. Process tomography
of quantum memory in a Josephson-phase qubit coupled to a two-level state. Nature
Physics, 4(7), 523–526.
Neill, C., McCourt, T., Mi, X., et al. 2021. Accurately computing the electronic properties
of a quantum ring. Nature, 594(7864), 508–512.
Nielsen, M. A, and Chuang, I. L. 2011. Quantum Computation and Quantum Information.
Cambridge University Press.
Nielsen, Michael A. 2002. A simple formula for the average gate fidelity of a quantum
dynamical operation. Physics Letters A, 303(4), 249–252.
Niemczyk, T., Deppe, F., Huebl, H., et al. 2010. Circuit quantum electrodynamics in the
ultrastrong-coupling regime. Nature Physics, 6(10), 772–776.
Nigg, Simon E., Paik, Hanhee, Vlastakis, et al. 2012. Black-box superconducting circuit
quantization. Physical Review Letters, 108(24), 240502.
Niskanen, A O, Harrabi, K, Yoshihara, F, Nakamura, Y, Lloyd, S, and Tsai, J S. 2007.
Quantum coherent tunable coupling of superconducting qubits. Science (New York,
N.Y.), 316(5825), 723–726.
O’Brien, J. L., Pryde, G. J., Gilchrist, A., et al. 2004. Quantum process tomography of a
controlled-NOT gate. Physical Review Letters, 93(Aug), 080502.
O’Brien, William, Vahidpour, Mehrnoosh, Whyland, Jon Tyler, et al. 2017. Superconduct-
ing caps for quantum integrated circuits. e-print arXiv:1708.02219.
Ofek, Nissim, Petrenko, Andrei, Heeres, Reinier, et al. 2016. Extending the lifetime of a
quantum bit with error correction in superconducting circuits. Nature, 536(7617),
441–445.
Olivares, S. 2012. Quantum optics in the phase space. European Physical Journal Special
Topics, 203(1), 3–24.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 References 295

Olver, F. W. J., Lozier, D. W., Boisvert, R. F., and Clark, C. W. (eds). 2010. NIST Handbook
of Mathematical Functions. Cambridge University Press.
O’Malley, P. J. J., Kelly, J., Barends, R., et al. 2015. Qubit metrology of ultralow phase noise
using randomized benchmarking. Physical Review Applied, 3(Apr), 044009.
Orlando, Terry P. 1991. Foundations of Applied Superconductivity. Addison-Wesley Pub-
lishing Company.
Ortuño, M., Somoza, A. M., Vinokur, V. M., and Baturina, T. I. 2015. Electronic transport in
two-dimensional high dielectric constant nanosystems. Scientific Reports, 5(1), 9667.
Otterbach, J. S., Manenti, R., Alidoust, N., et al. 2017. Unsupervised machine learning on
a hybrid quantum computer. e-print arXiv:1712.05771.
Ozfidan, I., Deng, C., Smirnov, A.Y., et al. 2020. Demonstration of a nonstoquastic
Hamiltonian in coupled superconducting flux qubits. Physical Review Applied, 13(3),
034037.
Paauw, F., Fedorov, A., Harmans, C., and Mooij, J. 2009. Tuning the gap of a superconduct-
ing flux qubit. Physical Review Letters, 102(9), 090501.
Paik, Hanhee, Schuster, D. I., Bishop, Lev S., et al. 2011. Observation of high coherence in
Josephson junction qubits measured in a three-dimensional circuit QED architecture.
Physical Review Letters, 107(24), 240501.
Pashkin, Yu. A., Yamamoto, T., Astafiev, O., Nakamura, Y., Averin, D. V., and Tsai, J. S.
2003. Quantum oscillations in two coupled charge qubits. Nature, 421(6925), 823–
826.
Pechal, M., Huthmacher, L., Eichler, C., et al. 2014. Microwave-controlled generation of
shaped single photons in circuit quantum electrodynamics. Physical Review X, 4(4),
041010.
Pegg, D., and Barnett, S. 1989. Phase properties of the quantized single-mode electromag-
netic field. Physical Review A, 39(4), 1665–1675.
Peropadre, B., Forn-Díaz, P., Solano, E., and García-Ripoll, J. 2010. Switchable ultrastrong
coupling in circuit QED. Physical Review Letters, 105(2), 023601.
Peropadre, B., Zueco, D., Porras, D., and García-Ripoll, J. 2013. Nonequilibrium and
nonperturbative dynamics of ultrastrong coupling in open lines. Physical Review
Letters, 111(24), 243602.
Pino, Manuel, and García-Ripoll, Juan José. 2018. Quantum annealing in spin-boson model:
from a perturbative to an ultrastrong mediated coupling. New Journal of Physics,
20(11), 113027.
Pitaevskii, Lev, and Stringari, Sandro. 2016. Bose-Einstein Condensation and Superfluidity.
Oxford University Press.
Place, Alexander P. M., Rodgers, Lila V. H., Mundada, Pranav, et al. 2021. New material
platform for superconducting transmon qubits with coherence times exceeding 0.3
milliseconds. Nature Communications, 12(1), 1779.
Preskill, John. 2018. Quantum computing in the NISQ era and beyond. Quantum, 2(Aug.),
79.
Raimond, J. M., Brune, M., and Haroche, S. 2001. Manipulating quantum entanglement
with atoms and photons in a cavity. Reviews of Modern Physics, 73(3), 565–582.
Ramos, Tomás, and García-Ripoll, Juan José. 2018. Correlated dephasing noise in sin-
gle-photon scattering. New Journal of Physics, 20(10), 105007.
Reed, M. D., DiCarlo, L., Johnson, B. R., et al. 2010. High-fidelity readout in circuit
quantum electrodynamics using the Jaynes–cummings nonlinearity, 173601. Physical
Review Letters, 105(17).

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 296 References

Rigetti, Chad, and Devoret, Michel. 2010. Fully microwave-tunable universal gates in super-
conducting qubits with linear couplings and fixed transition frequencies. Physical
Review B, 81(13), 134507.
Rigetti, Chad, Gambetta, Jay M., Poletto, Stefano, et al. 2012. Superconducting qubit in
a waveguide cavity with a coherence time approaching 0.1 ms. Physical Review B,
86(10), 100506.
Rol, M. A., Battistel, F., Malinowski, F. K., et al. 2019. Fast, high-fidelity conditional-phase
gate exploiting leakage interference in weakly anharmonic superconducting qubits.
Physical Review Letters, 123(Sep), 120502.
Romero, G., García-Ripoll, J. J., and Solano, E. 2009a. Microwave photon detector in circuit
QED. Physical Review Letters, 102(17), 173602.
Romero, Guillermo, García-Ripoll, Juan José, and Solano, Enrique. 2009b. Photodetection
of propagating quantum microwaves in circuit QED. Physica Scripta, T137(T137),
014004.
Rosenblum, S., Reinhold, P., Mirrahimi, M., Jiang, Liang, Frunzio, L., and Schoelkopf, R. J.
2018. Fault-tolerant detection of a quantum error. Science, 361(6399), 266–270.
Roy, Ananda, and Devoret, Michel. 2016. Introduction to parametric amplification of
quantum signals with Josephson circuits. Comptes Rendus Physique, 17(7), 740–755.
Sánchez-Burillo, E., Martín-Moreno, L., García-Ripoll, J. J., and Zueco, D. 2016. Full
two-photon down-conversion of a single photon. Physical Review A, 94(5), 053814.
Sandberg, M., Wilson, C. M., Persson, F., et al. 2008. Tuning the field in a microwave
resonator faster than the photon lifetime. Applied Physics Letters, 92(20), 203501.
Schneider, Christian Markus Florian. 2014. On-Chip Superconducting Microwave Beam
Splitter. Ph.D. thesis, TU Munich.
Schrödinger, E. 1926. Quantisierung als Eigenwertproblem. Annalen der Physik, 384(4),
361–376.
Schuster, D. I., Houck, A. A., Schreier, J. A., et al. 2007. Resolving photon number states
in a superconducting circuit. Nature, 445(7127), 515–518.
Schuster, D. I., Wallraff, A., Blais, A., et al. 2005. ac stark shift and dephasing of a
superconducting qubit strongly coupled to a cavity field. Physical Review Letters,
94(12), 123602.
Selby, Alex. 2014. Efficient subgraph-based sampling of Ising-type models with frustration.
e-print, sep, arXiv:1409.3934.
Sharafiev, Aleksei, Juan, Mathieu L., Gargiulo, Oscar, et al. 2021. Visualizing the emission
of a single photon with frequency and time resolved spectroscopy. Quantum, 5(Jun),
474.
Shi, Tao, Chang, Yue, and García-Ripoll, Juan José. 2018. Ultrastrong coupling few-photon
scattering theory. Physical Review Letters, 120(15), 153602.
Shin, Seung Woo, Smith, Graeme, Smolin, John A., and Vazirani, Umesh. 2014. How
“quantum” is the D-Wave machine? e-print arXiv:1712.05771.
Shnirman, Alexander, Schön, Gerd, and Hermon, Ziv. 1997. Quantum manipulations of
small Josephson junctions. Physical Review Letters, 79(12), 2371–2374.
Shor, Peter W. 1995. Scheme for reducing decoherence in quantum computer memory.
Physical Review A, 52(Oct), R2493–R2496.
Silver, A., and Zimmerman, J. 1967. Quantum states and transitions in weakly connected
superconducting rings. Physical Review, 157(2), 317–341.
Steane, Andrew. 1996. Multiple-particle interference and quantum error correction. Pro-
ceedings of the Royal Society of London. Series A: Mathematical, Physical and
Engineering Sciences, 452(1954), 2551–2577.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 References 297

Steffen, Matthias, Ansmann, M., Bialczak, Radoslaw C., et al. 2006. Measurement of
the entanglement of two superconducting qubits via state tomography. Science,
313(5792), 1423–1425.
Steffen, Matthias, Kumar, Shwetank, DiVincenzo, David P., et al. 2010. High-coherence
hybrid superconducting qubit. Physical Review Letters, 105(Sep), 100502.
Sundaresan, Neereja M., Liu, Yanbing, Sadri, Darius, et al. 2015. Beyond strong coupling
in a multimode cavity. Physical Review X, 5(2), 021035.
Susskind, L., and Glogower, J. 1964. Quantum mechanical phase and time operator. Physics,
1, 49–61.
Teufel, Stefan. 2003. Adiabatic Perturbation Theory in Quantum Dynamics. Lecture Notes
in Mathematics, vol. 1821. Springer Berlin Heidelberg.
Townsend, Christopher, Ketterle, Wolfgang, and Stringari, Sandro. 1997. Bose–Einstein
condensation. Physics World, 10(3), 29–36.
van den Brink, Alec Maassen, Berkley, A. J., and Yalowsky, M. 2005. Mediated tunable
coupling of flux qubits. New Journal of Physics, 7(1), 230–230.
van der Ploeg, S. H. W., Izmalkov, A., van den Brink, Alec Maassen, et al. 2007.
Controllable Coupling of Superconducting Flux Qubits. Physical Review Letters,
98(5), 057004.
van der Wal, C. H., ter Haar, A. C., Wilhelm, F. K., et al. 2000. Quantum superposition of
macroscopic persistent-current states. Science, 290(5492), 773–777.
Venturelli, Davide, Mandrá, Salvatore, Knysh, Sergey, O’Gorman, Bryan, Biswas, Rupak,
and Smelyanskiy, Vadim. 2015. Quantum optimization of fully connected spin
glasses. Physical Review X, 5(3), 031040.
Vlastakis, Brian, Kirchmair, Gerhard, Leghtas, Zaki, et al. 2013. Deterministically encoding
quantum information using 100-photon Schrödinger cat states. Science (New York,
N.Y.), 342(6158), 607–610.
Wallraff, A., Schuster, D. I., Blais, A., et al. 2004. Strong coupling of a single photon to
a superconducting qubit using circuit quantum electrodynamics. Nature, 431(7005),
162–167.
Wallraff, A., Schuster, D. I., Blais, A., et al. 2005. Approaching unit visibility for control
of a superconducting qubit with dispersive readout. Physical Review Letters, 95(6),
0650501.
Walls, D. F., and Milburn, G. J. 1994. Quantum Optics. Springer Berlin Heidelberg.
Walter, T., Kurpiers, P., Gasparinetti, S., et al. 2017. Rapid high-fidelity single-shot
dispersive readout of superconducting qubits. Physical Review Applied, 7(5), 054020.
Wang, C., Gao, Y. Y., Reinhold, P., et al. 2016. A Schrodinger cat living in two boxes.
Science, 352(6289), 1087–1091.
Williams, Colin P. 2011. Quantum Gates. Springer London. Pages 51–122.
Wilson, C. M., Johansson, G., Pourkabirian, A., et al. 2011. Observation of the dynamical
Casimir effect in a superconducting circuit. Nature, 479(7373), 376–379.
Wootters, W. K., and Zurek, W. H. 1982. A single quantum cannot be cloned. Nature,
299(5886), 802–803.
Wu, Yulin, Bao, Wan-Su, Cao, Sirui, Chen, et al. 2021. Strong quantum computational
advantage using a superconducting quantum processor. Physical Review Letters 127,
180501.
Yamamoto, T., Neeley, M., Lucero, E., et al. 2010. Quantum process tomography of
two-qubit controlled-Z and controlled-NOT gates using superconducting phase qubits.
Physical Review B, 82(18), 184515.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 298 References

Yamamoto, T., Pashkin, Yu. A., Astafiev, O., Nakamura, Y., and Tsai, J. S. 2003. Demon-
stration of conditional gate operation using superconducting charge qubits. Nature,
425(6961), 941–944.
Yan, F., Gustavsson, S., Kamal, A., et al. 2015. The flux qubit revisited to enhance coherence
and reproducibility. e-print, arXiv:1508.06299.
Yan, Fei, Gustavsson, Simon, Kamal, Archana, Birenbaum, et al. 2016. The flux qubit
revisited to enhance coherence and reproducibility. Nature Communications, 7(1),
12964.
Yoshihara, Fumiki, Fuse, Tomoko, Ashhab, Sahel, Kakuyanagi, Kosuke, Saito, Shiro,
and Semba, Kouichi. 2016. Superconducting qubit–oscillator circuit beyond the
ultrastrong-coupling regime. Nature Physics, 13(1), 44–47.
Yurke, Bernard, and Denker, John. 1984. Quantum network theory. Physical Review A,
29(3), 1419–1437.
Zener, C. 1932. Non-adiabatic crossing of energy levels. Proceedings of the Royal Society
A: Mathematical, Physical and Engineering Sciences, 137(833), 696–702.
Zueco, David, and García-Ripoll, Juanjo. 2019. Ultrastrongly dissipative quantum Rabi
model. Physical Review A, 99(Jan), 013807.

https://doi.org/10.1017/9781316779460.012 Published online by Cambridge University Press

 Index

T1 , 93, 169 charge

T2 , 93 variable, 47
charging energy, 117, 118
adiabatic Chimera graph, 258
passage, 239 chiral, 195
quantum computation, 245 Choi matrix, 216, 218
quantum computation algorithm, 246 Choi–Jamiolkowski isomorphism, 216
theorem, 243 circuit
adiabatic evolution, 241 quantization, 44
adiabatic passage, 209 circuit model, 199
amplifier circuit-QED, 3, 156
degenerate, 95 circulator, 94
linear, 95 classical
phase insensitive, 95 annealing, 264
phase sensitive, 95 Clifford group, 219
quantum limited, 96 coherence, 15
anharmonicity, 108, 138 completely positive maps, 214
annealing, 263 continuum, 52
approximation coplanar waveguide, 69
Markov, 281 Coulomb blockade, 116
rotating wave, 86, 89, 111, 163, 176 counterrotating terms, 86, 89, 176
coupling
bath, 87, 88 capacitive, 142
BCS dispersive, 208
theory, 22 inductive, 141, 256
beam splitter, 93 tunable, 145
Bloch sphere, 15, 216 covariance matrix, 83
Bogoliubov transformation, 272 cross-talk, 144, 212
boundary conditions current, 19
open, 271 critical, 43
periodic, 270 electric, 26
persistent, 20
capacitance, 41 superconducting, 25
capacitance matrix, 48 superfluid, 25
cavity, see resonator
driving, 85 D-Wave, 253
with a qubit, 174 qubit, 254
spectrum, 104 dc-SQUID
three-dimensional, 75 quantization, 55
tunable, 73 decay rate, 166, 284

299

https://doi.org/10.1017/9781316779460.013 Published online by Cambridge University Press

 300 Index

decoherence, 15, 37, 93, 112, 169 single-qubit, 112

dephasing, 113 Z, S, and T, 112
relaxation, 115 Gaussian
density matrix, 14 operation, 83
dephasing, 37, 113, 120, 237 state, 82
dephasing rate, 114
detuning, 184 H.c., xiv
diabatic evolution, 241, 244 Hadamard, 112
dispersive Hamiltonian, 10
coupling, 185 heating, 37
regime, 184 Heisenberg
displacement operator, 78 equation, 12
dissipation, 37 uncertainty principle, 11
dynamical Casimir effect, 74 hybrid, 94

embedding, 258 impedance, 66

energy inductance, 42
charging, 42 Josephson, 33
entanglement, 81 kinetic, 42
entropy mutual, 142, 257
von Neumann, 81 tunable, 146
error input–output, 166
threshold, 230 input–output equation, 90
error correction, 221 input–output theory, 282
errors interaction picture, 279
cross-talk, 214 IQ mixer, 101
decoherence, 214 Ising Hamiltonian, 141
gates, 214
measurement, 213 Jaynes–Cummings
SPAM, 213, 217 Hamiltonian, 163, 176
state preparation, 213 Josephson
effect, 21
FAQUAD, 245 energy, 33, 43
fast quasi-adiabatic-passage, see FAQUAD relation, 32
fault tolerance, 107, 225 Josephson junction, 126
FHW, 92
fidelity Kraus
average, 220 decomposition, 278
average gate, 218 operators, 237, 278
process, 218
state, 218 Lagrangian, 9, 46
flux, 28 Lamb shift, 90, 166, 284
quantum, 20, 28 lambda system, 196
variable, 47 Landau–Zener
fluxoid quantization, 20, 29 formula, 242
Fock process, 241, 264
basis, 76 leakage, 109, 128, 243
states, 76 Lindblad
frequency equation, 16, 278, 280
angular, 8 operator, 16, 278
frequency crowding, 145, 212 localization, 162
frustration, 133 logical bit, 221
full-half-width, 92
macroscopic quantum model, 24
gain, 80 magnetometer, 56
gap map
superconducting, 20, 22, 23 completely positive, 277
gate dephasing, 237
Hadamard, 112 depolarization, 237
phase, 112 trace preserving, 277

https://doi.org/10.1017/9781316779460.013 Published online by Cambridge University Press

 Index 301

Markovian regime, 283 QND, see measurement

master equation, 16, 92, 278, 281 quality factor, 75, 92, 93, 181
cavity, 176 quantization
Mathieu equation, 126 black-body, 8
matrix canonical, 9
tridiagonal, 270, 271 charge qubit, 52
measurement dc-SQUID, 55
dispersive, 187 flux qubit, 57
generalized, 18, 189 LC resonator, 49, 64
nondemolition, 16, 189 rf-SQUID, 53
projective, 16 transmission line, 50
quantum limited, 18 quantum
qubit, 57 advantage, 247
single shot, 17 annealing, 263, 264
uncertainty, 17 register, 140
Von Neumann, 16 simulator, 141
memory volume, 234
function, 89, 166, 194 quantum computing
time, 283 circuit model, 199
microwave quantum control, 112
source, 78 quantum electrodynamics, see circuit-QED
mixer, 101 quantum gate
swap, 208
quantum register, 202
NISQ, 232, 264
quantum simulation, 199
nodes, 45
quantum Turing machine, 199
noise, 95
qubit, 13, 77, 106
quasistatic, 114
3D transmon, 129
white, 113
capacitively shunted flux, 140
NP-hard, 250
charge, 116
flux, 135
Ohmic, 161
gradiometric flux, 139, 154
oscillators, 272
hyperbola, 120
interactions, 140
parametric amplifier, 74 logical, 222
parity, 228 persistent current, 135
Pauli physical, 107
basis, 14 quantization, 52, 57
matrices, 13 reset, 191
periodic boundary conditions, 70 rf-SQUID, 132
perturbation theory, 273 transmon, 122
degenerate, 274
Kato, 273 Rabi
nondegenerate, 274 frequency, 178
Schrieffer–Wolff, 273 Hamiltonian, 175
phase, 59 oscillations, 180
gauge invariant, 27 vacuum Rabi splitting, 178
superconducting, 26 Ramsey interferometry, 148
voltage relation, 27 randomized benchmarking, 219
phase space, 78 reflection, 170
phase stability, 78 representation
photodetection, 196 phase, 59
photon, 64, 68 resistance, 19
Planck resistivity, 19
constant, 8 resonator
polariton λ/2, 71
lower and upper, 178 λ/4, 72
process tomography, 217 driving, 67
Purcell LC, 38, 63
filter, 189 quantization, 49

https://doi.org/10.1017/9781316779460.013 Published online by Cambridge University Press

 302 Index

rf-SQUID single-mode squeezing, 80

quantization, 53 thermal, 77
rotating term, 86 two-mode squeezing, 81
RWA, see approximation, rotating wave vacuum, 76, 81
stripline, 69
scattering strong coupling, 171, 175, 179,
matrix, 170 181, 187
Schrödinger cat, 79, 193 surface code, 225
Schrödinger equation, 12 swap gate, 208
Schrieffer–Wolff transformation, 184 symmetry point, 120, 122
signal, 95 syndrome, 224
single-photon source, 77
SPAM, 217, 219 Tavis–Cummings Hamiltonian, 183
spectral function, 89, 160 temperature
spectroscopy, 90 critical, 20, 22, 23
cavity, 185 thermal state, 36
qubit, 190 threshold
qubit tunneling, 260 fault tolerance, 225
single-tone, 185 time to solution, 248
two-tone, 190 transmission, 170
speed of light, 71 transmission line, 69
spin echo, 114 quantization, 50
spin glass, 250 transmon
spin-boson frequency tunable, 128
Hamiltonian, 159, 160 tunable
Ohmic, 161 coupling, 74
spontaneous emission, 167 inductance, 73
squeezing, 80 tunable coupler, 212
decibels, 80
single-mode, 80 ultrastrong coupling, 163, 182, 187
two-mode, 81 uncertainty principle, 66
stabilizer code, 224 unitary
stabilizer operator, 224 single-qubit, 112, 207
state two-qubit, 209
coherent, 78, 83 universal quantum computer, 199
dark, 196 USC, see ultrastrong coupling
Fock, 84
Gaussian, 83 vacuum energy, 146
minimum uncertainty, 78 vacuum Rabi splitting, 179
mixed, 14 voltage, 19, 28
Schrödinger cat, 84
squeezed, 83 wave equation, 71
thermal, 83 waveguide-QED, 164
state transfer, 172 weak coupling, 181
states Wigner function, 82
Fock, 76 reconstruction, 84
Gaussian, 82 Wigner–Weisskopf, 165
mixed, 77
multimode squeezed, 81 XY Hamiltonian, 141

https://doi.org/10.1017/9781316779460.013 Published online by Cambridge University Press