Topics in Quantum Computing
Architecture
Thien Nguyen
Research School of Engineering
The Australian National University
A thesis submitted for the degree of
Doctor of Philosophy
College of Engineering and
Computer Science May 2019
© Copyright by Thien Nguyen 2018
All Rights Reserved
This page intentionally left blank.
To my family:
Dad Thang, Mom Phuong, and especially my wife Quyen.
This page intentionally left blank.
Declaration
I hereby declare that except where specific reference is made to the work of others,
to the best of my knowledge and belief, the contents of this dissertation are original
and have not been submitted in whole or in part for consideration for any other degree
or qualification in this, or any other university. This dissertation is my own work or
partially in collaboration with others. This dissertation contains fewer than 100,000
words exclusive of appendices, bibliography, footnotes, tables and equations.
Thien Nguyen
May 2019
This page intentionally left blank.
Acknowledgements
In no way would this thesis have been possible without the support of many people who
guided and encouraged me along the way. First and foremost, I am much indebted to
my supervisors, Prof. Matthew James, who has been incredibly supportive from even
before my arrival at the ANU and remained so throughout my study. Matt gave me total
freedom to pursue my interests and helped me connect with experts in the field. At the
same time, he was constructively critical of various aspects of my study and professional
development.
I am grateful to Prof. Lloyd Hollenberg for taking an interest in me and giving
me an opportunity to work with him on the interesting problems of quantum computer
architecture. My special thanks also go to Dr Charles Hill. Charles’ eye for detail and
technical expertise has helped me in refining my ideas into research projects.
In the course of my study at the ANU, I have had the opportunity to work with and
learn from incredibly smart people to whom I would like to express my gratitude. They
are Dr Yu Pan, Dr Guodong Shi, Dr Shibei Xue, Dr Zibo Miao, Dr Shuangshuang Fu,
Dr Michael Hush, and Dr Andre Calvalho. I appreciated every conversation which I had
with every one of you, through which I have expanded my knowledge and understanding
of the quantum field.
To my fellow PhD students, Alfred, Oliver, Jessica, Alex, James, and Ruvi, it was
great listening to your talks in our group meeting. Special thanks to those who always
accepted my invitation to give a speech when I was the coordinator.
During my PhD, I have had the opportunity to work as a research intern at the Atos
Quantum Lab in Paris, France. To this, I’d like to thank Dr Cyril Allouche, head of
the Atos Quantum research program, and Sophie Houssiaux, head of R&D, Big Data
and Security, at Atos for the invaluable experience which had a profound impact on my
viii
research. I also want to express my thanks to the whole Atos quantum lab members who
taught me a lot about quantum computing, supercomputers, and especially about French
culture.
I am grateful to my wife who, despite struggling with her PhD study, has always
provided me emotional support whenever I need. Special thanks to my family members
in Vietnam, who have supported me along the way.
Last but not least, to every participant of the Friday lunch Quantum Cybernetics
group meeting, it was pleasant sharing lunch and discussing a wide variety of topics
with all of you during last four years.
I also gratefully acknowledge the generous financial support I have received from
the ANU and the ARC CQC2T for my PhD study and research.
Abstract
Quantum computing, which is considered the next revolution of computing technology,
brings together theories of mathematics, physics, and computer science. Building a
quantum computer thus requires a synthesis of knowledge and skills from multiple
disciplines. In this thesis, we take a step toward bridging and connecting the full “stack”
of quantum computing technology which spans across theoretical foundations, hardware
architecture, software and simulation.
At the foundation level, we study one of the central problems of quantum computing,
namely quantum error correction, from a control-engineering perspective. This approach
not only complements the conventional coding-based interpretation but also provides a
potential pathway to designing self-correcting quantum computers. First, we analyse
the surface-code quantum error correction under a continuous feedback-based protocol.
Second, we study the fundamental question of self-correcting quantum systems using
the control method of reservoir engineering.
Next, we study the scalability of a generic surface code quantum computer based on
spin qubits such as quantum dots and donor atoms. Solid-state qubits (quantum dots and
donor atoms), especially those that are Si-based, share similarities in device structures
and manufacturing processes with advanced semiconductor CMOS industry. However,
scaling up those quantum devices present immense challenges related to connectivity.
By applying tools and methods from the semiconductor industry, we can concretely
estimate the routing limitation of a planar connectivity scheme.
Finally, as part of my PhD education at the Australian National University, I took
part in an industry-based research internship to develop a high-performance quantum
simulator. We provide the full description of the software architecture, implementation
details, and benchmarking results of the simulator.
This page intentionally left blank.
List of Publications
This thesis is based on the following collection of papers which have either been
published in or submitted to peer-reviewed journals or conferences. Some of this work
was completed jointly with other researchers and selections from these joint papers to be
included in this thesis are only those that this author contributed significantly towards.
Journal Publications
1. Thien Nguyen, Charles D. Hill, Lloyd C. L. Hollenberg, and Matthew R. James.
Fan-out Estimation in Spin-based Quantum Computer Scale-up. Scientific Reports
7, Article number: 13386 (2017)
2. Yu Pan and Thien Nguyen. Stabilizing Quantum States and Automatic Error
Correction by Dissipation Control. IEEE Transactions on Automatic Control
(Volume: 62, Issue: 9, Sept. 2017) Page(s): 4625 - 4630
3. Thien Nguyen, Zibo Miao, Yu Pan, Nina Amini, Valeri Ugrinovski, and Matthew
James. Pole placement approach to coherent passive reservoir engineering for
storing quantum information. Control Theory Technology (2017) 15: 193.
4. Shibei Xue, Thien Nguyen, Matthew R. James, Alireza Shabani, Valery Ugri-
novskii, Ian R. Petersen. Modelling and Filtering for Non-Markovian Quantum
Systems. Pre-print: arXiv:1704.00986
5. Yi Li, Thien Nguyen, and Yu Pan. Deterministic Multi-Party Quantum Key
Distribution for Wireless Sensor Networks. In preparation.
xii
Peer-reviewed Conference Publications
1. Thien Nguyen, Charles D. Hill, Lloyd C. L. Hollenberg, and Matthew R. James.
Surface Code Continuous Quantum Error Correction Using Feedback. In Proceed-
ings of IEEE 54th Annual Conference on Decision and Control (CDC): Osaka,
Japan, December 2015
2. Yu Pan, Thien Nguyen, Zibo Miao, Nina Amini, Valeri Ugrinovski, and Matthew
James. Coherent observer engineering for protecting quantum information. In
Proceedings of the 35th Chinese Control Conference (CCC): Chengdu, China,
August 2016
3. Thien Nguyen, Zibo Miao, Yu Pan, Nina Amini, Valeri Ugrinovski, and Matthew
James. Pole placement approach to coherent passive reservoir engineering for
storing quantum information. In Proceedings of American Control Conference
(ACC): Seattle, WA, USA, July 2017
4. Shibei Xue, Thien Nguyen, and Ian R. Petersen. Feedback Control of a Non-
Markovian Single Qubit System. In Proceedings of the 11th Asian Control
Conference (ASCC) 2017: Gold Coast, QLD, Australia, December 2017
Table of contents
List of figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii
List of tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Quantum Technologies 1
1.2 Quantum Computing 3
1.2.1 Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Quantum Gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Physical Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.4 Quantum Error Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Software and Programming 16
1.3.1 Quantum Programming Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.2 Quantum Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 Quantum Control Engineering 17
1.4.1 Quantum Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Quantum Input-Output Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.3 Quantum Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5 Outline 25
1.6 Contributions 28
xiv TABLE OF CONTENTS
2 Solid-state Spin Qubit Control Routing . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1 Introduction 30
2.2 Routing Dimension Parameters 32
2.3 Solid-State Spin Qubit Unit Cell Model 33
2.4 Methods 39
2.4.1 Ring-by-ring Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4.2 Layer Optimisation Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Results 41
2.6 Summary 46
3 Continuous Quantum Error Correction . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Quantum Errors 49
3.1.1 Quantum Error Correction Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1.2 Surface Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.1.3 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.1.5 Outline and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Model 56
3.2.1 Distance-2 Surface Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.2 Continuous QEC in the SLH Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Methods 60
3.4 Results 61
3.4.1 Distance-2 Surface Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4.2 Distance-3 Surface Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5 Conclusions 65
4 Quantum Reservoir Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1 Introduction 70
TABLE OF CONTENTS xv
4.2 Background 72
4.2.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.2 Linear Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2.3 Lyapunov Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.3 Quantum Error Correction by Dissipation Control 81
4.3.1 Scalability of Lyapunov Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3.2 Error Correction Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3.3 Definition of AQEC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.4 Scalability of Dissipation Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3.5 Automatic Quantum Error Correction by Dissipation Control . . . . . . . . . . . 92
4.3.6 Dissipation Control of 3-qubit Repetition Code States . . . . . . . . . . . . . . . . . 95
4.4 Decoherence Free Subsystem Synthesis 98
4.4.1 Open-loop Reservoir Engineering for DFS Generation . . . . . . . . . . . . . . . . 99
4.4.2 Coherent Feedback Reservoir Engineering . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.4.3 Special Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.4.4 Special Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.5 Concluding Remarks 117
5 Quantum Programming and Simulation . . . . . . . . . . . . . . . . . . . . . . . . 119
5.1 Quantum Programming Language 120
5.1.1 Quantum Assembly Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.1.2 High-level Programming and Data Processing . . . . . . . . . . . . . . . . . . . . . . 122
5.2 Simulation Engine 124
5.2.1 Overall Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.2.2 Technical Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.2.3 Benchmark Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
xvi TABLE OF CONTENTS
6 Simulating Input-Output Quantum Systems with LIQUi|⟩ . . . . . . . 133
6.1 Background 134
6.1.1 Single-photon State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.1.2 Compound Gradient Echo Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.1.3 Engineering Dissipation Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.1.4 Trotter Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.2 Models and Results 142
6.2.1 Amplitude Damping Set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.2.2 Toy Example: Single-Atom Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.2.3 Discretized Gradient Echo Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.3 Conclusions and Future Work 146
6.3.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.3.2 The Path Forward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.1 Where We Stand 149
7.2 The Way Forward 150
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
A Appendix Mathematical Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
A.1 General form of the QSDE for continuous error correction in Chapter 3 169
A.2 Proof of the scalability condition (4.44)-(4.45) 172
A.3 Proof of the condition (4.52)-(4.53) 173
B Appendix Feynman Path Integral Simulation With FPGA . . . . . . . 175
B.1 Method Overview 175
B.2 FPGA Implementation 177
B.3 Software Components 179
B.3.1 Host Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
TABLE OF CONTENTS xvii
B.3.2 FPGA Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
This page intentionally left blank.
List of figures
1.1 Qubit Bloch sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.2 Solid-state qubit diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.3 Quantum gate timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.4 Quantum gate decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.5 Surface code layout diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.6 Quantum circuit to measure surface code syndromes . . . . . . . . . . . . . . . . . . . 131.7 Logical CNOT gate braiding diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141.8 Logical CNOT gate by braiding holes in surface code . . . . . . . . . . . . . . . . . . 151.9 Active and passive QEC diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191.10 Diagram of a typical quantum filtering set-up. . . . . . . . . . . . . . . . . . . . . . . . 201.11 SLH network connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211.12 Diagram of beamsplitter in series with cavity QED. . . . . . . . . . . . . . . . . . . . 221.13 Sample QHDL code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231.14 Sample QHDL schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241.15 Measurement feedback block diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251.16 Coherent feedback block diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261.17 Overall architecture of a universal quantum computer. . . . . . . . . . . . . . . . . . 27
2.1 Routing parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322.2 Diagram of surface code lattice with embedded readout devices . . . . . . . . . . 342.3 Diagram of a generic surface code array unit cell . . . . . . . . . . . . . . . . . . . . . . 352.4 Metal 1 pitch scaling roadmap from ITRS . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.5 Ring-by-ring routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.6 Layer optimal routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.7 Illustration of 2-D qubit lattice surface gate fanout . . . . . . . . . . . . . . . . . . . . . 422.8 Fanout scalability at fixed interconnect length . . . . . . . . . . . . . . . . . . . . . . . . 432.9 Qubit fanout scalability in terms of number of routing layers . . . . . . . . . . . . . 452.10 Fan-out scalability vs. interconnect length for SSI and ECI schemes . . . . . . 46
3.1 Surface code layout diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.2 Distance-2 surface code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.3 Surface code continuous error correction diagram . . . . . . . . . . . . . . . . . . . . . 58
xx LIST OF FIGURES
3.4 Syndrome estimation trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.5 Continuous feedback error correction fidelity comparison . . . . . . . . . . . . . . . 633.6 Distance-3 surface code diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.7 Time-domain simulation of distance-3 surface code under continuous QEC . 663.8 Fidelity vs. feedback strength plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 Potential function plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.2 Lyapunov function scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854.3 State evolution simulation of repetition error correction code . . . . . . . . . . . . 974.4 Repetition error correction code under dissipative couplings . . . . . . . . . . . . . 984.5 Coherent feedback network for DFS generation. . . . . . . . . . . . . . . . . . . . . . 1004.6 Open loop setup for DFS generation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.7 Coherent plant-observer network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1094.8 Coherent feedback network for DFS generation . . . . . . . . . . . . . . . . . . . . . . 1114.9 Two-cavity system diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.1 AQASM example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1225.2 AQASM QFT example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1235.3 PyAQASM QFT example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1245.4 Quantum simulator diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1255.5 Path integral diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1275.6 Quantum circuit simulation objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1285.7 Benchmark of Hadamard gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1295.8 Comparison with Liquid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1305.9 Comparison with other simulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.1 QHDL design flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1356.2 Single-photon generating filter cascading realization. . . . . . . . . . . . . . . . . . . 1366.3 Gradient settings for GEM operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1376.4 Engineered dissipation by direct coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . 1396.5 Quantum circuit of cascading Hamiltonian unitary . . . . . . . . . . . . . . . . . . . . 1416.6 Discretized single-photon GEM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1426.7 Amplitude damping LIQUi|⟩ code snippet . . . . . . . . . . . . . . . . . . . . . . . . . . 1436.8 Histogram of photon detection events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1446.9 Photon absorption vs. photon bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
B.1 OpenCL implementation of complex arithmetic . . . . . . . . . . . . . . . . . . . . . . 177B.2 OpenCL implementation of multiplication and summation kernels . . . . . . . 180
List of tables
1.1 Types of experimental quantum bits (qubits) . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.2 Hadamard gate implementation steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1 Gate count configurations for spin shuttling interconnect (SSI) and end control
interconnect (ECI) protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.2 Minimum interconnect length for spin shuttling interconnect (SSI) and end control
interconnect (ECI) protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1 Fidelity and trace distance comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.1 Technical details of the quantum simulator . . . . . . . . . . . . . . . . . . . . . . . . . . 1265.2 Memory capacity and bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
This page intentionally left blank.
Chapter 1
Introduction
Technology is anything that wasn’t around when you were born.
Alan Kay
1.1 Quantum Technologies
The discovery of quantum physics in the early 20th century has fundamentally changed
our understanding of the physical world at the microscopic scales. Quantum mechanics,
which said that objects could be in different states (superposition) at the same time and
correlate over long distance without any direct contact (entanglement), has since become
universally accepted as the most accurate description of physical systems.
The so-called second quantum revolution [37] has started a few decades ago pi-
oneered by both physicists and computer scientists. The novel ideas which brought
about this second wave of quantum revolution are, instead of avoiding quantum effects,
we could transform it from a theory for understanding nature into utilising them as
computing and engineering resources.
2 Chapter 1. Introduction
Quantum Mechanics Primers
• Quantum States (Superposition)
A quantum particle can be in two or more states at the same time. The
quantum state can be described mathematically by a Hilbert space, which is
described by vectors with complex coefficients.
• Composite Systems (Entanglement)
The joint state of multiple quantum systems thus can be described by a
tensor product of their Hilbert spaces. This tensor product results in an
exponential growth in the dimensionality of the composite system. Unlike
classical counterparts, quantum composite systems can be inseparable in
the sense that we can not describe them individually as separate compo-
nents. This is the principle of quantum entanglement which is one of the
fascinating quantum phenomena.
Quantum technologies will lead to significant advances in precision timing, sensors
and computation, have a substantial impact on the finance, defence, aerospace, energy,
infrastructure and telecommunications sectors. For example,
• Ultra-high-precision spectroscopy and microscopy, positioning systems, clocks,
gravitational, electrical and magnetic field sensors.
• Quantum-safe secure networks and quantum key distribution.
• Universal quantum computers: quantum algorithms, modelling materials, quantum
chemistry, biological systems.
One of the emerging quantum 2.0 fields, quantum computing [122], is a significant
breakthrough and paradigm shift in our intuition and understanding of computation. Sci-
entists and industries are developing quantum algorithms and applications which could
potentially speed advancements in materials science [8, 67], financial modelling [132],
machine learning and artificial intelligence [24, 39]. However, realising a quantum
computer is as challenging as the very problems it can solve. There are multitudes of
challenges from the device fabrication up to the software to run quantum algorithms.
Another essential field of quantum technology is quantum control [197, 198, 128, 36]
1.2 Quantum Computing 3
whose goal is to control physical systems whose behaviour is governed by the laws of
quantum mechanics.
We will explore fundamentals of quantum computing, quantum programming lan-
guage and simulation, and quantum control in greater details in the next sections.
1.2 Quantum Computing
Quantum computers make use of a quantum-mechanical phenomenon (e.g. superposition
and entanglement) that allows data to be represented as quantum bits (qubits) - these are
not constrained to conventional 0 or 1 binary values, but instead can be a superposition
of zero and one simultaneously. Hence, a set of qubits can represent exponentially more
values than their classical-bit counterparts1. This makes quantum computing a promising
platform which could potentially solve computational problems unmanageable for even
the most advanced conventional supercomputer.
To be qualified as a quantum computing platform, there are several necessary condi-
tions which are often summarised as the DiVincenzo’s criteria [34] for scalable quantum
computers. They are,
1. well-defined qubits: addressable and coherently manipulable,
2. initializable to a well-defined quantum state,
3. long coherence times relative to gate times of a universal set of gates, and
4. high quantum efficiency measurements.
Building a quantum computer, despite how challenging it is, is not in itself the
end goal of quantum computing research. Realizing quantum computing capability
demands that hardware fabrication and control efforts be augmented by the design and
development of quantum algorithms/software. The latter part has seen tremendous
progress being made in the last few decades following the seminar paper of Peter Shor
[157] which sparks the public interest in quantum computing. Despite not having a real
quantum computer, researchers have come up with a host of quantum algorithms to solve
cryptography, searching, and simulation problems.
1In this comparison, classical bits refer to the unit of information (bit) in information theory. Infact, classical probabilistic state spaces also scale exponentially in dimension, hence require specializedcomputational algorithms such as the Monte Carlo method.
4 Chapter 1. Introduction
Quantum Algorithms
• Shor’s algorithm [157, 158]
This is one of the first applications of quantum computers which states that
there exists an efficient way to factor a product of two prime numbers. This
has a huge implication since the difficulty of prime factorization problem is
the foundation of our public-key cryptosystem.
• Grover’s algorithm [62]
Grover’s quantum search algorithm for unstructured data is quadratically
faster than its classical counterpart.
• HHL algorithm [65]
Named after its authors, Harrow, Hassidim and Lloyd, HHL algorithm gives
us the capability to solve systems of linear equations. This paves the way to
transform optimization and machine learning applications to the quantum
domain.
• Quantum Simulation
This includes the fields of quantum chemistry, material science, and particle
physics. Quantum simulation utilizes the most accurate description of
properties and dynamics of systems at nano-scale. Potential applications
are catalyst design (e.g. carbon/nitrogen capture), drug development, and
superconductivity.
1.2.1 Qubits
A probabilistic classical bit can then be represented by a sum of these state vectors:
p |0⟩C + q |1⟩C, where p and q are real numbers in the unit interval that represent the
probability of finding that bit in either state, e.g. the bit has a probability p of being
in the |0⟩C state. We always have p+q = 1, and the state 12 |0⟩C + 1
2 |1⟩C represents a
uniformly random bit. Of course, when we measure the state, it can only be in the state
of 0 or 1 after measurement.
1.2 Quantum Computing 5
A quantum bit or qubit has a similar structure. It has two output states, |0⟩ and |1⟩,and a general qubit state is represented as a sum of vectors α |0⟩+β |1⟩, except now the
weights α and β are complex numbers so that |α|2 + |β |2 = 1.
The probability of the qubit being in state |0⟩ is given by |α|2, and similarly, |β |2
gives the probability of observing outcome |1⟩. Mathematically we represent qubit states
as vectors whose elements are the coefficients are α and β .
|ψ⟩= α |0⟩+β |1⟩ ≡
α
β
, α, β ∈ C (1.1)
Indeed, a qubit has many more states than a random (probabilistic) classical bit. This
can be demonstrated geometrically by the Bloch sphere: the states of a probabilistic bit
lie on the line between the north and south poles, while the states of a qubit occupy the
whole surface of the sphere. The basis states of |0⟩ and |1⟩ then reside on the north and
south poles.
Fig. 1.1 Qubit Bloch sphere: (left) A qubit state can be in any place on the surface of thesphere; (right) A random classical bit can only exist on the north-south poles line.
Mathematically, multi-qubit states are composed via the tensor product. Hence, the
state space of qubits scales exponentially with the number of qubits. Despite its huge
state-space, we can only ever extract n classical bits of information from an n-qubit
system. Therefore, quantum computers are suitable for applications studying global
properties of functions and data.
6 Chapter 1. Introduction
Experimentally, there are a wide-range of physical systems which can be used to
implement qubits. Some of the more prominent platforms are listed in Table 1.1. The
number of qubits quoted in Table 1.1 is based on the current fabrication technology.
In this thesis, one particular platform which will be studied in great details is the
donor-based qubit system (donor atoms). In fact, doping Si crystal with impurity has
been used in the semiconductor industry to fabricate MOS transistors for decades.
However, doping control down to single-atom level [155, 140, 178] is what enables
quantum computation. Donor atoms, like Phosphorous, have one extra valence electron
which is naturally confined by the atomic potential of the donor atom itself. Therefore,
the electron’s degrees of freedom, such as spin, can be utilised as quantum information
carrier (qubit). The pictorial illustration of a donor-based qubit is shown in Fig. 1.2,
which is similar to the original configuration proposed by Kane [87].
A AJ
P P
D
1Fig. 1.2 Diagram of the Phosphorous in Silicon donor spin qubit. The donor atom hasone extra electron whose spin is used as qubit. The surface A gate on top of the atomcontrols the hyperfine interaction which can be used together with a magnetic field toperform addressable single-qubit operations. The J gate in between two adjacent atomscontrols the inter-qubit coupling which is used to implement two-qubit gates.
1.2.2 Quantum Gates
Quantum algorithms are often described by sequences of quantum operations which
are quantum gates. Mathematically, quantum gates are complex matrices by which the
quantum state vector gets multiplied. This transformation by a single-qubit quantum
gate can be visualised by rotation on the Bloch sphere.
If classical algorithms are represented by boolean circuits, quantum algorithms are
described by a sequence of quantum gates which are referred to as quantum circuits.
1.2 Quantum Computing 7
Table 1.1 Types of experimental quantum bits (qubits)
Physical System Characteristics
Trapped ions [92, 114]Laser-controlled charged ions are used as qubits.Long coherence time but gate operations are also slow.Scalability issues w.r.t. laser control.
Coherence time: hoursGate fidelity: >99.9%Max number of qubits: 10-50
Superconducting resonators [33]Electrical LC oscillators with zero resistance.Electrically-controlled by microwave signals.Operates at cryogenic temperature.
Coherence time: 10-100 µsGate fidelity: >99.5%Max number of qubits: 50+
Quantum dots [101, 10]Gate-defined nano structures in semiconductormaterials.Small (nano-scale) qubits, electrically controlled.Using the same fabrication technology as classicalchips.Relatively few interacting qubits can be demonstrated.
Coherence time: µs - msGate fidelity: >99%Max number of qubits: 2-5
Donor atoms [48, 47]
Single dopant atom site in semiconductor.Both nuclear and electron spins can be used as qubits.Single-atom doping is challenging.
Coherence time: up to secs(nuclear spins)Gate fidelity: >99%Max number of qubits: 2-5
Diamond vacancies (NV centers) [192, 203]Nitrogen atom in diamond and a vacant lattice site.Laser-controlled and can operate at room temperature.Difficult to couple multiple NV qubits for quantumcomputing.
Coherence time: secsGate fidelity: >99%Max number of qubits: 5-10
Topological qubits [117, 2]Quasi non-local particles (majorana fermions).Can have extreme long coherence time due tonon-locality.Less developed theoretical and experimental capabili-ties.
Coherence time: unknownGate fidelity: unknownMax number of qubits: 1
There are complementary frameworks which can be used to describe quantum algorithms
(e.g. tensor networks and graphical flow diagrams [25]). However, the quantum circuit
8 Chapter 1. Introduction
framework is by far the most widely-adopted thanks to its analogy to the classical logical
boolean circuit framework.
A quantum gate or quantum logic gate is a fundamental quantum circuit operating
on a small number of qubits. They are the analogues for quantum computers to classical
logic gates for conventional digital computers. Quantum logic gates are reversible,
unlike many classical logic gates. Mathematically, we represent quantum logic gates
using unitary matrices.
For example, the Hadamard gate, which operates on a single qubit, is represented by
the matrix:
H =1√2
1 1
1 −1
. (1.2)
Another widely-used set of single-qubit gates are called Pauli gates whose matrices
are:
X =
0 1
1 0
, Y =
0 −i
i 0
, Z =
1 0
0 −1
. (1.3)
Given a single-qubit gate which is represented by a unitary matrix U , a controlled-U
gate is a two-qubit gate as follows:
C(U) =
1 0 0 0
0 1 0 0
0 0 u00 u01
0 0 u10 u11
. (1.4)
where
U =
u00 u01
u10 u11
. (1.5)
is the original gate matrix.
1.2 Quantum Computing 9
Universal Quantum Gates
A set of quantum gates is universal if any operation possible on a quantum
computer can be reduced to it.
For example, the set of the Hadamard gate (H), a phase rotation gate (by an
arbitrary angle), and the controlled-NOT (CNOT ) gate is universal.
1.2.3 Physical Implementations
Logical qubit operations (quantum gates) have to be experimentally implemented via
complex instrumental control schemes. For instance, to perform quantum operations on
donor electron spin qubits, we utilise the two available surface gates/contacts (A and J)
as shown in Fig. 1.2 together with the applied magnetic field. Then, the scheme of Hill
et al. [70] can be used to implement single qubit gates as well as the CNOT gate.
As an example, the Pauli X-rotation on solid-state spin qubits can be performed in
two steps. Firstly, the target qubit is rotated by 2π while all other qubits are rotated
2π+θ by applying voltage on their A-gates thus increasing their rotation speed. We have
created a relative rotation of θ as desired between the target qubit and all other “observer”
qubits. However, to correct the overall rotation, we need to perform a correction by
rotating all qubits by θ around X-axis. This time we apply voltage on all A-gates since
we want all qubits to rotate evenly. The procedure is depicted in Fig. 1.3.
A1
Jk
Ai6=1
Target: Rx(2π) ≡ I
Observer: Rx(2π + θ) ≡ Rx(−θ)Target: Rx(θ)
Observer: Rx(θ)
Fig. 1.3 Timing diagram to implement X-rotation on donor spin electron. A1 is theA-gate contact of the target qubit.
Similar techniques can be used to rotate the qubit around an arbitrary axis. Indeed,
the Hadamard is nothing but a π rotation around the axis m = i+k√2
(i, j, and k are basis
vectors), which can be implemented similarly as the X rotation. The steps to perform a
Hadamard gate is summarised in Table 1.2. For more information, see Hill et al. [70].
10 Chapter 1. Introduction
Table 1.2 Hadamard gate implementation steps
Step Target qubit Observer qubits
1 Rm(π)≡ H Rx(α)
2 Rx(2π)≡ I Rx(−α)
We can pretty much apply arbitrary single-qubit rotation using this technique. Two-
qubit gates, such as the CNOT gate, can be achieved by using the following decomposi-
tion,
CNOT = (H ⊗ I)exp(
iπI −X
2⊗ I −X
2
)(H ⊗ I)
= (H ⊗ I)[Rx
(π
2
)⊗Rx
(π
2
)]
× exp(
iπ
4X ⊗X
)(H ⊗ I).
(1.6)
The only two-qubit operation in Eq. (1.6) is exp(iπ
4 X ⊗X), which can be implemented
straightforwardly using the coupling H = J(σ ⊗σ) term, where σ are the Pauli spin
matrices (with eigenvalues ±1), as
exp(
iπ
4X ⊗X
)= (X ⊗ I)exp
(iπ
8σ ⊗σ
)
× (X ⊗ I)exp(
iπ
8σ ⊗σ
).
(1.7)
Using quantum circuit notation, we can express the sequence of gates to implement
a CNOT gate between two qubits as in Fig. 1.4, which is physically executed by pulsing
the A’s and J surface metal gates.
U(π8
)U(π8
)|a〉 H X X Rx
(π2
)H
|b〉 Rx(π2
)
|a〉
|a⊕ b〉
Fig. 1.4 Quantum circuit decomposition of a CNOT gate. U(
π
8
)= exp
(iπ
8 σ ⊗σ). Each
of the decomposed elements can be realized by properly-timed pulsing of the surfacegates (A and J).
1.2 Quantum Computing 11
This type of quantum gate decomposition is a crucial aspect of the quantum software
stack which we will explore in Chapter 5. Since each qubit hardware architecture can
handle a specific set of elementary gates (satisfying the universality condition), it is the
task of the software stack to translate from an hardware-agnostic quantum computing
instructions into the actual physical gate operations which the underlying hardware
platform can support [109, QSh].
This is pretty much similar to what a compiler is doing when you wrote classical
computer codes, i.e. converting from text-based high-level programming languages
into assembly bytecodes for the particular target platform. Developing an effective tool-
chain for quantum programming can potentially improve the productivity of algorithm
development and also lower the entry barrier of quantum computing since most of
the complication of underlying physical hardware can be abstracted away and hence
developers can rely on a uniform computing model at the top of the stack.
1.2.4 Quantum Error Correction
The most prominent drawback of the technology is that qubits are inherently fragile.
Hence, they can only retain quantum information for a tiny amount of time. Interactions
within the system and with a noisy environment are typically the limiting factors of
coherence time: the duration during which the underlying qubit system retains its
quantum properties. Therefore, the best devices require extremely clean control signals
and cryogenic operation to reduce thermal noise.
Those random fluctuations will occasionally flip or randomise the state of a qubit,
potentially corrupting the computation. Hence, to create a “functional” quantum com-
puter, we need a quantum computer that is fault-tolerant in the sense that the quantum
computer must be able to detect failures so that they do not spread in space and time
during computation and hence can be corrected.
One of the most prominent contributions of fault-tolerant quantum computing re-
search is the invention and the continuous improvement of various quantum error
correction (QEC) codes [52, 53, 30, 148, 32, 161, 189, 43, 42, 162, 167, 164, 118],
which provide a potential pathway to achieving universal quantum computing.
12 Chapter 1. Introduction
The goal of quantum error correction is to use redundancy and correction to realise
logical qubits with improved error rates as compared with that of the elementary qubits.
The fault-tolerance threshold [99], which is the qubit failure probability below which
reliable quantum computation becomes feasible, is the standard measure of fault-tolerant
quantum computing, where error propagation is constrained such that error correction
protocols remain effective [54].
Among a wide variety of quantum error correction codes, the surface code [30, 94,
45, 189, 43] has stood out in terms of computational error threshold which is about two
orders of magnitude higher than that of conventional concatenated coding schemes. The
critical feature is that implementing the surface code requires a regular 2-D arrangement
of qubits, where neighbouring qubits interact with each other in a pairwise manner and
in parallel (see Fig. 1.5). Qubits are classified either as data qubits or syndrome (ancilla)
qubits according to their roles in the quantum error correction procedure. Each syndrome
qubit measurement fixes an eigensubspace of a stabilizer operator, which involves all
four neighbouring data qubits. Logical qubits are defined as topological defects on the
qubit lattice where syndromes are not measured. Thus, there are two types of logical
qubits, so-called smooth (Z-cut) and rough (X-cut) logical qubits. The code distance is
defined either by the perimeter of the defects or the distance between them, whichever is
smaller. Interested readers should consult Fowler et al.[43] for an in-depth review.
For surface code, there are only two types of syndromes: Z or X syndromes, which
stand for ZZZZ or XXXX operators acting on the four data qubits. The Z and X
syndromes are measured by performing a sequence of CNOT gates between the ancilla
and its four neighbouring data qubits as shown in the quantum circuit diagrams in Fig.
1.6a and 1.6b, respectively.
Logical qubits are defined as topological holes (defects) on the qubit lattice where
syndromes are not being measured. Thus, there are two types of logical qubits, so-called
smooth (Z-cut) and rough (X-cut) logical qubits. The code distance is then defined
either by the perimeter of the holes or the distance between them, whichever is smaller.
The reason is that an uncorrectable error occurs in surface code whenever physical
qubit errors form a continuous chain surrounding a hole or connecting two holes. It is
worthwhile mentioning that the boundary of the code lattice could also be considered
1.2 Quantum Computing 13
1 2 3 4
5 6 7
8 9 10 11
12 13 14
15 16 17 18
X X X
Z Z Z Z
X X X
Z Z Z Z
X X X
Fig. 1.5 Surface code layout: white circles represent data qubits; filled circles aresyndrome qubits (X stabilizers in green and Z stabilizers in yellow). Each internalstabilizer acts on four adjacent data qubits, while boundary stabilizers act on either twoor three data qubits.
|0〉 MZ
|a〉
|b〉
|c〉
|d〉(a) Z-syndrome measurement
|0〉 H H MZ
|a〉
|b〉
|c〉
|d〉(b) X-syndrome measurement
Fig. 1.6 Quantum circuit to measure surface code syndromes. The first line representsthe ancilla, while the following lines represent the surrounding data qubits.
as a superficial hole. Therefore, crossing the boundary from one end to the opposite
or connecting one hole to the boundary are also irreversible logical errors. Interested
readers should consult [43] for an in-depth explanation.
The bigger and farther apart the surface code holes are, the higher code distance
we can get. Therefore, the ability to fabricate a gigantic lattice of qubits is pivotal in
making the surface code to work. Besides, because the logical CNOT operation in the
14 Chapter 1. Introduction
surface code involves braiding a pair of holes as shown in Fig. 1.7, the actual distance
between two distance-d holes needs to be at least 3d. An example of the CNOT braiding
evolution [43] between an X-cut qubit and a Z-cut counterpart is illustrated in Fig. 1.8. In
the surface code layout, a pair of holes that are 3d apart will serve as a unit cell. Braiding
operations can be optimised at the architectural level with regard to the computation
time or lattice area as shown in [135]. At the micro-architectural level, on the other hand,
we need to be able to keep track of the holes’ locations as well as grow and shrink them
by turning off and on the respective syndrome qubits.
1st pair of holes
2nd pair of holes
Time
Fig. 1.7 Braiding diagram of a logical CNOT gate in surface code with time axis runninghorizontally.
The code distance is calculated by the intrinsic error rate of the underlying qubit
hardware as well as the threshold of the selected error correction code, as following [86]
εlogical ≥C1
(C2
εphysical
εthreshold
)⌊ d+12 ⌋
, (1.8)
where C1 and C2 are constants that are code-dependent. For surface code, according to
[46], the values for C1 and C2 are 0.13 and 0.61, respectively. We denote the error rate
per logical gate operation as εlogical , which depends on the target success probability
of the algorithm and the number of logical gate operations to complete it. An entire
algorithm success rate of εalgo will require εlogical ≈ εalgo/Nlogical , where Nlogical is the
number of logical gate operations required for the entire computation. The physical error
rate εphysical is the experimental fidelity of the qubits, which is compared against the
threshold of the surface code at εthreshold ≈ 1% to determine the minimum code distance
1.2 Quantum Computing 15
3d
(a) Original holes (b) Growth hole A
(c) Shrink hole A (d) Growth hole A
Fig. 1.8 Logical CNOT gate by braiding holes in surface code. The braiding is done byexpanding and shrinking a hole. In order to maintain the code distance d throughout thisprocess, the orginal holes need to be separated by at least 3d.
d. This is the minimum code size that can guarantee the algorithm to be completed with
high probability.
From this code distance estimation, we can asymptotically approximate the number
of physical qubits needed and therefore the expected classical resources. Remember
that a distance d logical qubit needs a hole covering approximately d2 physical qubits.
In addition, all the holes need to be 3d apart to make sure logical CNOT gates can be
readily performed between any pair of logical qubits. A simple calculation can be made
16 Chapter 1. Introduction
to estimate the number of physical qubits as follows.
Nphysical ∝ d2 × (4×√
Nlogical)2, (1.9)
in which, the number of logical qubits Nlogical is algorithm dependent, and the code
distance d is from the previous estimation step.
1.3 Software and Programming
1.3.1 Quantum Programming Languages
A quantum programming language is the means by which quantum algorithms, sub-
routines, or applications can be expressed in a human-comprehensible form yet can be
faithfully and consistently translated into quantum operation sequences which imple-
ments the high-level instructions across various hardware platforms. The abstraction
away from the specific hardware implementation is particularly important at this stage be-
cause we still have a handful of competing platforms which are very different in terms of
low-level operations. For application developers, a consistent and platform-independent
view of quantum computation is the key to innovation and productivity.
Most importantly, the programming language must obey the principles of quantum
computing, namely the no-cloning theorem and the reversibility (unitary) of quantum
operations. On top of that, the language itself needs to be able to cope with emerging
computing models, e.g. the mix of classical and quantum data in a quantum/classical
hybrid programming model.
Early quantum programming language proposals from academic researchers, for
example, QCL (Quantum Computation Language) [208], QGL (Quantum Gate Lan-
guage) [165], Scaffold [3], Chisel-Q [100], or Quipper [60], provide the foundation for
the quantum programming research community. Recently, thanks to the tremendous
interest from industry in quantum computing, we have seen the released of industry-
backed quantum programming languages and development kits/environment. Some of
the most popular are:
• LIQUi|⟩ from Microsoft, embedded in F#
• Q#, also from Microsoft
1.4 Quantum Control Engineering 17
• OpenQASM and Quantum Information Software Kit (QISKit) from IBM
• Quil/Forrest from Rigetti
• AQASM from Atos
1.3.2 Quantum Simulation
Since functional quantum computers will not be available anytime soon, we need to be
able to simulate quantum algorithms on conventional computers to validate and debug
those algorithms. The simulation of quantum computers using first principle linear
algebra approach, i.e. matrix-vector multiplication, is strictly memory-bound. The
memory capacity required grows exponentially with the number of qubits involved in
the algorithms.
For reference, one of the largest quantum circuits that have been simulated using
this first principle approach is a 45-qubit simulation, which used 500 terabytes2 of
memory on a state-of-the-art supercomputer [81]. There are other various approaches
which could help reduce the memory requirement at the cost of computing time (space-
time trade-offs) or the accuracy of the simulation (e.g. tensor product approximation
methods).
The lack of access to real quantum computing hardware necessitates the use of
quantum simulators to test algorithms and programs. In addition, by deploying multiple
computation models (not just the universal first principle linear algebra approach) on the
simulator back-end, we could provide the best performance outcome for each particular
quantum circuit of interest.
1.4 Quantum Control Engineering
Quantum control engineering, which has been evolving in tandem with quantum com-
puting, plays a vital role in the design and realisation of quantum devices. The ideas
of using feedback control to stabilise naturally-unstable systems are the cornerstone of
classical control theory. Various control techniques can be deployed to autonomously
correct quantum error in the same manner as the conventional QEC schemes.
21 terabyte (TB) = 1,024 gigabytes (GB)
18 Chapter 1. Introduction
Observables, Measurement and Decoherence
• Observables and Measurement
The wavefunction is what defines the probability of finding (measuring) the
particle’s property (e.g. velocity or position) at a certain value.
When a measurement is made, the quantum state is settled (collapsed) into
one of the possibilities. The outcome of the measurement is randomised
according to the wavefunction (quantum state). However, after measure-
ment, the quantum state is completely determined, i.e. we will always get
the same measurement outcome if repeating that measurement.
• Decoherence
Since quantum superposition states collapse if measured, unintended in-
teractions with the surrounding environment can lead to the “leaking” of
quantum information.
In this section, I want to summarise the fundamental concepts of quantum control
and the role it may play in the entire fault-tolerant quantum computing system, or
“stack”. This can be visualised in the diagram in Fig. 1.9 where the quantum control
layer sits just above the hardware qubit implementation. This control layer is supposed
to provide fully-autonomous stability enhancement and error rejection to the underlying
qubit network before any active error correction actions are applied, e.g. measuring
syndrome qubits, decoding for potential errors, etc.
1.4.1 Quantum Noise
The dynamics of quantum states (ψ(t)) are often described in terms of the Schrödinger’s
equation. For open quantum systems, the dynamics are described by the Lindblad master
equation [19] (Markovian case) which has the following form:
ρ(t) =−i[H,ρ(t)]+∑L ∗L (ρ(t)) (1.10)
1.4 Quantum Control Engineering 19
Quantum Error Correction Qubit Network
Passive Error Correction/Noise Cancellation
Active Error Correction
SyndromeMeasurement
Correction
Coherent/classical feedback control
Dynamical decoupling (open-loop)
Fig. 1.9 Combination of passive and active techniques for quantum error correction inwhich the passive (autonomous) layer improves the stability and/or error rejection tothe underlying qubit network before any active error correction actions are applied, e.g.measuring syndrome qubits, decoding for potential errors.
where ρ(t) is density matrix, H is the self-energy Hamiltonian operator, and the super-
operator L is defined as
L ∗L (ρ) = LρL† − 1
2L†Lρ − 1
2ρL†L. (1.11)
The coupling operator L in equations (1.10) and (1.11) describes the coupling (inter-
action) between the quantum system of interest and the external fields (environment/bath).
It is worth noting that we could in principle describe the system + bath as a whole using
the interaction Hamiltonian. This so-called quantum noise model [49, 141], which
describes the overall system in terms of plant and interacting fields (hence coupling
operators), is the foundation of the quantum feedback control field [198, 84, 205].
1.4.2 Quantum Input-Output Model
Emerging engineering problem of extracting information about the dynamical state of
quantum systems through measurement created the field of quantum filtering [18] which
20 Chapter 1. Introduction
is another form of the generalised continuous measurement formalism [198]. Front and
centre in this theory is the interaction between the quantum field and system-of-interest
as shown in Fig. 1.10 which in essence has two main effects. Firstly, information about
the system is gained by observing the field output which we will harvest using filtering.
This is the most common mechanism used in any metrology schemes. However, what
distinguishes quantum measurement is the inevitable back-action effect which dictates a
conditional or posterior state of the system consistent which an observed outcome.
Measure
afterinteraction
filter
measurementsignal estimate
beforeinteraction
B(t) Bout(t)
System
Y (t) X(t)
Fig. 1.10 Diagram of a typical quantum filtering set-up.
QHDL (Quantum hardware Description Language) is one of the very first Computer-
Aided-Design (CAD) tools for quantum systems built upon the concept of (S, L, H)
encapsulation. This formulation provides a common thread in the design and verification
flow from schematic capture to an HDL-like description and basic dynamical simulation
with both symbolic and numeric capabilities. Thanks to QHDL, a wide variety of
optical devices can be described in a systematic way that is ready for integration, such
as quantum optical logic gates [107], Set-Reset latch [105], or fully-coherent error-
corrected quantum memory [90].
In the quantum input-output formalism [50], the stochastic evolution of an open
Markov quantum system driven by vacuum noise inputs (dA(t)′s) is given by the Hudson-
Parthasarathy Quantum Stochastic Differential Equation [79]:
dU(t) = −iHdt +(S− I)dΛ(t) (1.12)
+ ∑i
(dA†
i (t)Li −L†i SdAi(t)−
12
L†i Lidt
)U(t),
in which, the unitary evolution U(t) is defined on the combined space of the system
plant and coupling fields. Any system operator dynamics can be derived from that using
the relation X(t) = U(t)†XU(t). The resulting operator-based differential equations
1.4 Quantum Control Engineering 21
are referred to as Heisenberg-Langevin equations and completely equivalent to the
Schrödinger picture master equation.
This Heisenberg-picture dynamical model can be parametrised conveniently by a
triple G = (S, L, H), where H is the internal Hamiltonian, L = Li is a set of coupling
operators (e.g. annihilation operators for amplitude damping), and S is a unitary input-
to-output scattering matrix (e.g. beam-splitters in quantum optics or quantum point
contacts in solid-state). This parametrisation scheme is often referred to as the SLH
quantum network theory [55, 56].
The network part of this model is what important since from the three parameters
and the given network topology we can compute the equivalent SLH model of the entire
network thanks to the two basic rules as shown in Fig. 1.11.
G1
G2
G1 G2
G1 G2
G2 C G1
Fig. 1.11 SLH network connections: (left) Concatenation and (right) Cascading.
Mathematically, these two connection rules can be expressed as:
G2 ⊞G1 =
S1 0
0 S2
,
L1
L2
,H1 +H2
, (1.13)
G2 ◁G1 = (S2S1,L2 +S2L1,H1 +H2 + ImL†2S2L1). (1.14)
This is the backbone of the QHDL toolbox, whereby network topology (can be in the
form of schematics or Verilog-style inputs as shown in the below example) is processed
symbolically to derive the overall network model. The system dynamics can then be
simulated by differential equation solvers.
22 Chapter 1. Introduction
Besides the bottom-up approach, we can also perform a top-down decomposition
in the SLH framework by the network synthesis theory [127]. Given an arbitrary SLH
model which may contain a large number of internal dynamical variables (optical modes
or qubits) and inputs/outputs, one can always faithfully identify a suitable collection of
one degree of freedom oscillator components and to connect them serially with proper
Hamiltonian interaction to build up the prescribed system model.
One recent development of the SLH modelling approach is the effort to extend its
application to a wide variety of input states besides the conventional vacuum inputs,
such as thermal field, single-photon and two-photon states [58, 125, 159, 57].
For demonstration purposes, considering a fundamental system of a beam-splitter
(M) and a cavity QED (C) [187] in series as diagrammatically shown in Fig. 1.12.
A1
A2
A2
B1
B1
γ
κ
Fig. 1.12 Diagram of beamsplitter in series with cavity QED.
Individually, each component can be described by its corresponding SLH parame-
ters [56]3:
M =
β −α
α β
,0,0
,
C =
I,
√κa
√γσ−
,Hc
,
Hc = ∆ f a†a+∆aσ+σ−+ ig(σ+a−σ−a†),
where κ and γ are the cavity and atomic decay rates, respectively; ∆ f and ∆a are the
detunings; g is the coupling constant between the field and atomic transition.3assuming instantaneous coupling
1.4 Quantum Control Engineering 23
-- Structural QHDL generated by gnetlist-- Entity declaration
ENTITY BeamSplitterNetwork ISPORT (
A1 : in fieldmode;A2 : in fieldmode;Vacin : in fieldmode;B1 : out fieldmode;B2 : out fieldmode;Vacout : out fieldmode);
END MachZehnder;
-- Secondary unitARCHITECTURE netlist OF BeamSplitterNetwork IS
COMPONENT BeamsplitterGENERIC (
theta : real := theta_value);PORT (
In1 : in fieldmode;In2 : in fieldmode;Out1 : out fieldmode;Out2 : out fieldmode);
END COMPONENT ;
COMPONENT SingleSidedJaynesCummingsGENERIC (
kappa : real := kappa_value);gamma : real := gamma_value);g : real := g_value);Delta_a : real := Delta_a_value);Delta_f : real := Delta_f_value);
PORT (In1 : in fieldmode;VacIn: in fieldmode;Out1 : out fieldmode;UOut : out fieldmode);
END COMPONENT ;
SIGNAL A_1_in : fieldmode;SIGNAL A_2_in : fieldmode;SIGNAL A_1_out : fieldmode;SIGNAL A_2_out : fieldmode;SIGNAL B_1_out : fieldmode;SIGNAL Vac_in : fieldmode;SIGNAL Vac_out : fieldmode;
BEGIN-- Architecture statement part
B1 : Beamsplitter PORT MAP (In1 => A_1_in,In2 => A_2_in,Out1 => A_1_out,Out2 => A_2_out);
C1: SingleSidedJaynesCummings PORT MAP (In1 => A_1_out,VacIn => Vac_in,Out1 => B_1_out,UOut => Vac_out);
-- Signal assignment partA_1_in <= A1;A_2_in <= A2;Vac_in <= Vacin;B1 <= B_1_out;B2 <= A_2_out;Vacout <= Vac_out;END netlist;
Fig. 1.13 Sample QHDL code of the cavity-beamsplitter system.
24 Chapter 1. Introduction
We can easily translate the diagram in Fig. 1.12 into a functional schematic using
appropriate pre-defined components as shown in Fig. 1.14. This can then be rendered
into a QHDL description as listed in Fig. 1.13 which will specify the ports, components,
connections, as well as any user-defined parameters of the system. One can use this to
generate Heisenberg or Schrödinger type dynamical equations to perform simulation or
verification as desired.
In1
In2
Out1Out
2
(−)
BeamSplitterIn1
Out1
VacIn UOut
Cavity
A1
B1
A2
Vac
In
Vac
Out
B2 = B2 = A2
B1
Fig. 1.14 Sample schematic using QHDL schematic capture tool: A QED cavity isconnected in series with a beamsplitter.
1.4.3 Quantum Control
The quantum input-output model, as captured by the (S, L, H) parameter set, bears
similarities to the classical control theory. In particular, the output field channels carry
information about the state of the quantum system due to the couplings as described by
the L operators. This information can be probed by a controller to gain knowledge about
the plant, hence can control it to the desired state. This closed plant-controller loop is
the basis of quantum feedback control.
Depending on the nature of the controller used in the feedback loop, quantum control
can be classified into two main categories:
1. Measurement feedback
The output field is measured continuously and the classical measurement signal
(e.g. analog electrical signal or digitised bits) is processed by a classical controller
which could be a Digital Signal Processing (DSP) chip, a Field Programmable
Chip Array (FPGA), or an electronic circuit. The controller will then emit control
1.5 Outline 25
Meas.I/V
Filter
F
Est.: ρ
Target: ρ
+
Quantum Device
Fig. 1.15 Measurement feedback block diagram.
signals based on pre-defined control algorithms. A typical block diagram of a
measurement feedback setup is shown in Fig. 1.15.
2. Coherent feedback
In coherent feedback control, the output field is fed to another quantum systems
directly as input signals. One quantum system acts as the controller whilst the
other is the plant. No quantum information is “leaked” to the outside world4 and
thus quantum coherence is preserved in the loop. It is worth noting that besides
field coupling, we could also have direct Hamiltonian coupling between the plant
and controller. This can be described as an interaction Hamiltonian involves
the dynamical variables of both systems. Coherent quantum feedback control is
illustrated in Fig. 1.16.
1.5 Outline
This thesis will study quantum computing at various layers of the quantum “stack” as
depicted in Fig. 1.17, from hardware devices to the error-correction layer as well as
quantum algorithms and applications. More specifically,
• Chapter 2: we investigate the routing scalability of solid-state quantum computers
under 2-dimensional error-correction code, namely the surface code. By applying
the classical electronics know-how regarding interconnect routing to a ubiquitous 2-
D qubit array with independent gate control and readout fan-out, we will develop a
concrete procedure for scalability estimation, which is adaptable to a wide range of
4when forming the feedback loop
26 Chapter 1. Introduction
Quantum System
(plant)
Quantum
Controller
Quantum
Field
Quantum
Field
Direct Coupling
Fig. 1.16 Coherent feedback block diagram.
surface code implementations by adjusting the gate configuration and dimensional
parameters.
• Chapter 3: we look at quantum error correction from a control-engineering centric
standpoint whereby we apply continuous control method to error correction qubit
network. A measurement feedback scheme is used to correct quantum error
continuously in real-time for a surface code 2-dimensional qubit array. The
controller design in this chapter is considered classical since it acquires and
processes classical measurement signals to output control signals.
• Chapter 4: we extend the controller design for quantum memory and quantum
error correction into the quantum realm by using techniques of dissipation control
(reservoir engineering) and coherent feedback control. In this chapter, coherent
feedback control technique is used to synthesise perfectly-isolated quantum infor-
mation storage system, namely the decoherence-free subsystem (DFS). Given the
hardware overhead of quantum error correction regarding the number of qubits
depends on the quality of the underlying physical qubits, utilising control tech-
niques to improve qubit stability even before any active error correction operations
are applied as shown in Fig. 1.9 will potentially reduce the amount of overhead
needed for fault-tolerant quantum computing.
1.5 Outline 27
Quantum Hardware
Superconductors Ion Traps Quantum Dots Donors NV Centers Topological
Quantum Gate Operations
Quantum Error Correction
Quantum Software (Algorithms)
PhysicalQubits
Logical
Qubits
Fig. 1.17 Overall architecture of a universal quantum computer.
• Chapter 5: we discuss the features and the development of a quantum programming
language, namely the Atos Quantum Assembly Language (AQASM) which I was
involved during my research internship at their quantum research lab in Paris.
The new quantum software suite also includes a classical quantum simulator
based on Feynman path integral approach. The simulator was successfully tested
on high-performance computing (HPC) platforms and is complementary to the
linear algebra simulator. A combination of the two is shown to provide the best
performance while not sacrificing the universality of the simulator.
• Chapter 6: we develop a method to simulate quantum open systems based on
the quantum control input-output formalism on a commercial simulator, namely
Microsoft LIQUi|⟩. A combination of theoretical and practical solutions can
provide accurate and consistent simulation results when using a digital gate-based
quantum simulator instead of conventional solver-based simulators.
Lastly, Chapter 7 summarises the content of this thesis with a brief outlook on future
work on quantum control, quantum computing architecture, and quantum programming
language. The follow-up Appendices provides detail proofs and additional information
as referred to in the main text.
28 Chapter 1. Introduction
1.6 Contributions
The aims of this thesis are
(i) to study a scalability aspect (fanout routing) of fault-tolerant quantum computation
from an engineering perspective (Chapter 2),
(ii) to apply quantum control technology to improve the stability of qubit systems
which are the foundation of the quantum computing hierarchy (Chapter 3 and 4),
and
(iii) to develop practical solutions to enhance quantum software development (Chap-
ter 5) and to bridge the gap between the digital gate-based simulation model and
the dynamical open quantum system model which is the cornerstone of quantum
control theory (Chapter 6).
It represents the author’s attempt to contribute to knowledge through academic
research and industry engagement activities.
Specifically, Chapter 2, 3 and 4 are derived from published works in peer-reviewed
journals or conference proceedings.
Chapter 5 summarises the author’s research results accomplished during an industry
internship with Atos Quantum Lab in Paris, France. Not only was the quantum simulator
successfully commercialised, but a patent has also been granted5 for the simulation
method, demonstrating the originality of the work.
Chapter 6 is based on the report that the author produced as an entry to the worldwide
Microsoft Quantum Challenge in which he won the Grand Prize. In this work, the
author has “extended LIQUi|⟩’s capabilities by supplementing the existing Hamiltonian
simulator with an innovative gadget for simulating dissipation. This clever use of
amplitude-damping noise enables a quantum computer to be used to simulate open
quantum systems as well as closed systems, and has important applications in real-world
situations.”6, according to the judges from Microsoft.
5France patent No FR3064380, Cyril Allouche and Thien Nguyen, “Procede de Simulation, Sur unOrdinateur Classique, D’un Circuit Quantique”, September 28, 2018
6https://www.microsoft.com/en-us/research/blog/microsoft-quantum-challenge-results-are-in/
Chapter 2
Solid-state Spin Qubit Control Routing
Every person should have their escape route planned.
Simon Pegg
The need for quantum error correction has a profound implication on the scalability
of future quantum computers. For example, some of the significant problems which
need to be addressed are overhead and complexity in terms of the number of qubits, the
qubit layout design to maintain fault-tolerant properties, and the coordination between
quantum and classical software algorithmically to decode and correct errors as well as
perform the intended quantum algorithm on the quantum computer.
This chapter aims to study the fan-out routing scalability, i.e. the ability to route
planar electrical wires, of surface-code quantum error correction code. To deal with
this question, in this chapter we apply the well-developed routing techniques of modern
semiconductor chip design to solid-state qubit platforms. This chapter proposes a
parametrization scheme which models the qubit layout in terms of the surface gate count
and spacing dimension. The results strongly indicate a bottle-neck which requires novel
architecture designs or technological breakthroughs to provide long-term scalability of
surface code solid-state quantum computers.
30 Chapter 2. Solid-state Spin Qubit Control Routing
2.1 Introduction
Building a large-scale quantum computer which can solve classically intractable prob-
lems is a technologically daunting task. With their close connection to highly scalable
classical electronics [72] solid-state spin qubit platforms, such as donor-based qubits [87,
155, 178, 38, 47, 144, 150, 206, 116] and quantum dots [177, 64, 206, 182, 89, 183],
are emerging as promising candidates [10, 40] for scalable quantum computation. On
semiconducting materials, e.g. Si, SiGe, or GaAs, it is possible in principle to fabricate
a large number of interconnecting qubits for quantum information processing. However,
in designing such a large-scale solid-state quantum chip, there is still a gap between the
quantum computer architecture [133, 28, 73, 147, 32, 35, 108, 86, 71] and the physical
qubit device implementation [177, 64, 206]. Architectures necessarily must incorporate
fault-tolerant quantum error correction in order to perform quantum algorithms [122]
at the logical quantum gate level. The physical implementation generally deals with
individual qubits on the basis of physical quantum gate operations, initialisation, and
readout which are the foundation for higher level quantum logical operations. In the
middle ground, quantum computer micro-architectures [86, 180, 71] attempt to bridge
that gap by providing engineering solutions to issues such as classical control, fan-out
interconnects, and chip layout.
One key advantage of the surface code is its nearest-neighbor interaction scheme
which scales favorably over the concatenation approach. However, this scheme also re-
quires a two-dimensional qubit layout and parallel control. In terms of micro-architecture
considerations, one must account for (a) the spatial/geometrical requirements of a 2D
nearest-neighbor interacting qubit array, and (b) the temporal/control requirements of
parallel/synchronous QEC operations. Broadly, one can identify two approaches. In the
ubiquitous independent control model, each quantum element (qubit, gate, interconnect,
readout) are controlled independently. In principle, this approach has the highest density
of quantum control gates each of which must be carefully characterised and timed to
allow for parallel operation across the qubit array (in a number of steps which does
not depend on the array size). At the other extreme, in the distributed control model
introduced in Hill et al. [71], a high degree of multiplexing allows sufficiently large
groups of qubits to be controlled and readout with the required parallelism.
2.1 Introduction 31
Some authors have attempted to address the problem in the independent control
approach by assuming the qubit lattice can be broken into smaller sparsely linked 2D
arrays [181], however, such tiling schemes in general present significant difficulties in
implementing the full range of logical operations required by the surface code. We
instead focus on the spatial/geometrical challenge of fabricating and scaling up of the
full monolithic surface code under the assumption of the independent (non-distributed)
control model in order to compare with the distributed control scheme. Quantum
interconnect protocols to reduce the qubit density are encapsulated in our study by
adding extra coupling surface gates which drive the transport protocols, and assuming
operational errors can be accommodated in the QEC protocol. In terms of gate density,
the generalised quantum interconnect model effectively captures most interconnect
mechanisms by adjusting the number of control gates per interconnect channel.
Under our generalised independent control model, we apply known techniques in
interconnect routing to analyse the geometrical scaling problem of surface code control
fan-out. We consider two types of solid-state spin qubits: atomically confined qubits
(such as phosphorus donors in silicon) [87, 155, 178, 47, 144, 116] and electrostatically
confined quantum dot qubits [177, 64, 206, 182, 89, 183]. In the non-distributed indepen-
dent control approach, every qubit on the surface code lattice has its own separate control
and readout structures that need to be fanned out. The qubit geometry is parametrised by
a universal unit cell which can be used to represent both donor-based and quantum dot
implementations including the quantum interconnects to neighboring cells by adjusting
the number of gates in the unit cell. Other dimensional parameters are selected based on
experimental and technological considerations. We must also stress that the scalability
of 2D spin qubit arrays depends on multiple factors, not just the control fan-out which
we study in this chapter. In particular, one must also address the various control issues
such as parallelizability, synchronisation, control characterisation, and cross-talk as well
as the overall thermal budget given the system will be required to operate at cryogenic
temperatures.
32 Chapter 2. Solid-state Spin Qubit Control Routing
2.2 Routing Dimension Parameters
In order to perform the routing analysis, we need to define the geometry of the wiring
and via pads. In particular, planar routing on each metal layer depends on the dimension
of the metal wires and the spacing between vias. These geometric parameters, which are
shown in Fig. 2.1, can be defined as followings:
• Pitch (p): the spacing between two neighboring pads after redistribution
• Pad diameter (d): the diameter of the pads
• Line width (w): the width of wires
• Line spacing (s): the spacing between wires or wires and pads
• Grid channel: the routing space available between two horizontal or vertical pads.
Its routing capacity is calculated by:
C = ⌊ p−d − sw+ s
⌋ (2.1)
• Diagonal channel: the routing space available between diagonal pads. In a square
array, its routing capacity is calculated by:
D = ⌊√
2p−d − sw+ s
⌋ (2.2)
d
w
s
grid channel
diagonal channel
p
Fig. 2.1 Routing parameters: pad pitch (p), wire width and spacing (w and s), paddiameter (d). The grid and diagonal channels are also indicated.
The smaller the wires, the better fanout scaling can be achieved. However, nar-
row and closely-spaced interconnects also tend to compromise the signal integrity,
2.3 Solid-State Spin Qubit Unit Cell Model 33
especially at high frequency. In this work, to provide upper-bound for the scalability,
we assume minimal wiring dimension of width(w) = 5 nm and spacing(s) = 25 nm.
Similar nanoscale wires have been fabricated in the lab for nanowire structures [191].
This wiring dimension assumption is also consistent with the International Technology
Roadmap for Semiconductors (ITRS) projection that by 2020 mainstream semiconductor
manufacturing will reach nanowire diameter of 5 nm. We assume that the via contact
diameter will double the wire width, i.e. d = 10 nm. Regarding the interconnect pitch
(p) used for routing, as shown in (2.3), when we implement longer interconnect chains
between qubits, the pitch will be extended.
2.3 Solid-State Spin Qubit Unit Cell Model
In this analysis, we will consider solid-state spin-based quantum computer platforms
with a model that encompasses both donor qubits and quantum dot qubits. To construct
a basic unit cell model for the micro-architectural fan-out routing analysis, the low-level
physics of the quantum devices, as well as high-level quantum computing architecture,
can be abstracted by making the following assumptions:
• Generalised quantum spin interconnects between neighboring qubits,
• Dedicated single-shot spin readout for every qubit,
• Single-sided metallization routing.
• Uniform interconnect dimension and spacing.
The first assumption regards the mechanism by which the qubit-qubit interaction is
realised. In principle, we could only implement direct spin-spin coupling, e.g. by spin
exchange or dipole couplings, however, direct spin couplings require stringent spacing
between qubits which restricts the control and readout routing. By adding interconnects
between qubits, we have some flexibility in arranging the qubits and thus can analyse
the fanout scalability accordingly. Secondly, we assume each qubit in the array has
its own spin read-out device which is usually a Single Electron Transistor (SET [115])
or equivalent [27, 184, 51, 77, 149]. This assumption may appear to be more than
necessary since neighboring qubits can share a common readout device by using some
forms of readout multiplexing, for example, the schemes presented in [124, 71, 12].
However, for our generic fan-out analysis, this serves as a baseline scenario from which
34 Chapter 2. Solid-state Spin Qubit Control Routing
we can straightforwardly adapt to other cases by modifying the gate count per qubit
to reflect other specific configurations with readout multiplexing. We assume that the
metal routing layers are built on a single side of the substrate. This is the predominant
routing technology used by the semiconductor industry. Lastly, we assume that the
feature size of interconnect wires is consistent between metallization layers. A pictorial
representation of the qubit array structure with dedicated readout devices is shown in
Fig. 2.2.
1 2 3 4
5 6 7
8 9 10 11
12 13 14
15 16 17 18
X X X
Z Z Z Z
X X X
Z Z Z Z
X X X
Data Qubits Syndrome Qubits Readout
Fig. 2.2 Diagram of surface code lattice with embedded readout devices. There aretwo types of qubits: data qubits and syndrome qubits (X and Z types). Neighboringqubits can interact with each other in order to perform CNOT gates. In this model, eachqubit has its own readout device. Dashed lines (black) represent quantum interconnectsbetween neighboring qubits. Dotted lines indicate qubits to which readout devices areassociated.
Regarding the interconnect protocols, while there seems to be a plethora of coherent
spin transport/coupling mechanisms [17, 23, 61, 199, 163, 68, 110, 176, 175, 12], for
the purpose of our fanout analysis, the main factor to consider is the number of additional
control gates versus interconnect length. We, therefore, consider two broad categories:
(i) gate count grows linearly with the interconnect length, and (ii) gate count is fixed
and independent of the interconnect length. For instance, SWAP-based interconnect and
spin shuttling protocols [12, 11] belong to the first category since we need surface gates
2.3 Solid-State Spin Qubit Unit Cell Model 35
along the channel to execute the quantum operations for spin swapping or shuttling. On
the other hand, protocols such as CTAP [61, 73, 111], capacitive coupling via floating
gate [176], spin chain [17, 23, 199], microwave line coupling [186, 29, 123], electric
dipole coupling [175], and surface acoustic wave spin transport [163, 68, 110] are some
examples of the second category because in those protocols we only need to have some
additional transport control gates at the ends of the interconnect not along the channel.
In what follows, we will use the terms spin shuttling interconnect (SSI) and end control
interconnect (ECI) for those two interconnect categories, respectively. The overall length
of the interconnect is L (for interconnect schemes based on qubit chains we equivalently
describe the interconnect length in terms of the number of nodes, Nnodes).
READOUT
Nr
Nq
Nc
Nc
Fig. 2.3 Diagram of a generic surface code array unit cell. Each qubit (circle) has acertain number of surface gates (Nq) to define qubit confinement potential and to performsingle-qubit rotations. Between any pair of neighboring qubits, we have Nc couplinggates that are used to control qubit interconnect coupling. At the center of the cell, wehave a readout device that has Nr surface gates.
In terms of physical qubit implementation, we categorise the surface metal gates
that need to be fanned out for controllability and readout into three categories: qubit
confining and control (Nq), interconnect coupling control (Nc), and readout (Nr). The
types of physical spin qubits considered are primarily classified by the confinement
mechanism, i.e. either via an atomic Coulomb potential (e.g. donors) for which we
assume Nq = 1, or via electrostatic gates (e.g. quantum dots), for which we assume
Nq = 3. Since we assume spin coupling based interconnects, the center-to-center distance
(pitch) between qubits needs to be sufficiently small. We use the qubit-qubit pitch of 20
36 Chapter 2. Solid-state Spin Qubit Control Routing
nm and 50 nm for Coulomb-confined and electrostatically-confined qubits, respectively.
Using the above gate classification, the surface code lattice can be decomposed into unit
cells, each of which contains one qubit and one readout device as shown in Fig. 2.3.
When partitioning the surface code lattice as shown in Fig. 2.2, there are four equivalent
interactions, namely along the north-east, north-west, south-east, or south-west direction.
For example, the unit cell in Fig. 2.3 is a south west participation scheme where the
interconnects and readout device on the bottom left of a qubit are associated with that
qubit for analysis purposes.
The latest International Technology Roadmap for Semiconductors (ITRS) [196],
which dictates the cadence of the semiconductor industry, predicts a wire pitch (distance
between two neighbouring wires) of around 20nm for the next ten years as shown in
Fig. 2.4. Indeed, 24nm-pitch copper interconnect has been recently demonstrated [200,
13] to be production-worthy using the next generation lithography process. Smaller
nanowires can be fabricated up to atomic precision using laboratory equipment [191].
Hence, the parameters which we use in this study are not only relevant to the laboratory
experiments but also to future quantum devices fabricated by commercial foundries.
2015 2020 2025
20
40
60
80
Year
Metal
1Pitch
(nm)
ITRS 2013Production Data
Fig. 2.4 Metal 1 pitch scaling roadmap from ITRS 2013 [196]; production data is takenfrom 22nm and 14nm nodes [9, 119].
2.3 Solid-State Spin Qubit Unit Cell Model 37
As indicated, we will categorise the interconnect protocols into two groups: SSI,
where the number of interconnect control gate count grows linearly with interconnect
length and ECI with a fixed number of interconnect surface gates regardless of intercon-
nect length. The gate count assumptions for these two scenarios are listed in Table 2.1.
The surface contacts are placed directly on top of the qubits, interconnect rails and
Table 2.1 Gate count configurations for spin shuttling interconnect (SSI) and end controlinterconnect (ECI) protocols. The interconnect node count (Nnodes) is the number ofintermediate qubit nodes along the interconnect channel.
Interconnect type SSI ECI
Nc 4×Nnodes 4
Nr 3 3
readout devices. In order to facilitate routing, these gate contacts are then redistributed
into a square grid array. The surface code qubit array can then be assembled by placing
unit cells next to each other, thus forms a regular global square grid array used for
fan-out routing.
It is worth noting that qubits (dots or donors) along the interconnect rails in Fig. 2.3
are not counted as physical qubits in the following analyses. Only the corner qubit of
the unit cell which can act as a data or syndrome qubit in the surface code (Fig. 2.2) is
accounted for as a physical qubit in the scalability study. In fact, several ECI schemes
that we mentioned earlier do not require intermediate qubit nodes at all, e.g. microwave
or capacitive coupling. For this scheme, the absolute interconnect distance is the only
relevant parameter.
For ECI and SSI schemes that involve qubit chains, an important issue may arise
which is the loss of qubits during transfer/coupling (due to operational errors or per-
manent manufacturing defects). While acknowledging that there are quantum error
correction methods and techniques [15, 161, 69, 164, 118] to mitigate qubit loss, this
aspect of qubit connectivity is outside the scope of our considerations here. Therefore,
we assume the feasibility of reliable quantum interconnects in order to focus on the issue
of fan-out routing scalability.
38 Chapter 2. Solid-state Spin Qubit Control Routing
The contact pitch after redistribution is related to the interconnect length (L) by the
following inequality:
pRL
≤ unit cell dimension(L)√Ntotal(L)
, (2.3)
where Ntotal is the total number of gate contacts in a unit cell and pRL
is the contact
pitch at the redistribution layer (RL). This total gate count may or may not depend
on the interconnect length. On the other hand, the unit cell dimension is proportional
to the interconnect length L. We can clearly see that by increasing the length of the
interconnect (L), the contact pitch after redistribution is extended since the denominator
is either constant or growing on the scale of square root of L while the nominator grows
linearly with L. In principle, larger pitches will benefit the global fan-out routing as
more interconnect routing space is created. This is explained in the Methods section
where we describe the routing parameters and the two commonly-used fan-out routing
algorithms.
At the redistribution layer the dimension parameters are d = 10 nm, w = 5 nm, and s
= 25 nm whilst the contact-contact pitch equals to the redistributed pitch computed by
(2.3). However, there is a minimum contact-contact pitch which needs to be satisfied,
namely pmin = d + s =35 nm. Thus, there is a lower bound on the interconnect length
to space the contacts sufficiently according to (2.3). The worst-case scenario occurs in
the SSI scheme for Coulomb-confined qubits because of their tight qubit-qubit spacing
and increasing number of coupling gates with interconnect length. We can estimate the
minimum interconnect length by using equation (2.3) in conjunction with the gate count
data in Table 2.1 and the qubit pitch assumption, e.g. for the 20 nm case we have
pRL
≈ 20nm×Nnodes√1+4Nnodes +3
> 35nm, (2.4)
which requires a minimum interconnect length (min(Nnodes)) of 14 nodes (280 nm).
Following the same procedure, we can derive the minimum interconnect length for
all configurations in terms of Nq configurations and interconnect schemes as shown in
Table 2.2.
2.4 Methods 39
Table 2.2 Minimum interconnect length in terms of chain node-count and absolutedistance for spin shuttling interconnect (SSI) and end control interconnect (ECI) proto-cols. Qubit-qubit pitch is 20 nm for atomically confined qubit (Nq = 1) and 50 nm forelectrostatically confined qubit (Nq = 3).
min(Nnodes)/Linterconnect SSI ECI
Nq = 1 (20 nm) 14 / 280 nm 5 / 100 nm
Nq = 3 (50 nm) 3 / 150 nm 3 / 150 nm
2.4 Methods
2.4.1 Ring-by-ring Routing
Fig. 2.5 Conventional ring-by-ring routing approach: the outermost ring of unconnectedpads are connected first, then inner rings are connected using grid channels of the outerring until their capacity exhausted. This procedure is then repeated on upper metal layers.The left is the routing on the first metal layer. Similarly, the middle one is the routingon the second layer, and so on. The overall procedure is depicted in the right diagramwhere each ring denotes the remaining pads after each layer of metallization.
A ring-by-ring router will work as follows:
1. Connect the outermost pads directly,
2. Use the grid channels between outermost pads to route internal pads on a ring-by-
ring basis,
3. When all the grid channels are exhausted, move up to an upper metal layer and
repeat step 1 and 2 until all pads are routed.
This approach is very intuitive, as shown in Fig. 2.5. However, the major drawback
of this scheme is that its boundaries are quickly shrinking layer-by-layer (as illustrated by
the smaller and smaller dotted squares on the rightmost diagram in Fig. 2.5). Therefore,
the routing capacity also decays as we proceed to higher and higher layer. This results in
40 Chapter 2. Solid-state Spin Qubit Control Routing
a higher number of metallization layers required as compared to the layer optimisation
scheme.
2.4.2 Layer Optimisation Routing
A second widely used scheme for escape routing is the so-called triangular routing [190]
that is depicted in Fig. 2.6.
Fig. 2.6 Metal layer optimal routing approach: the routing procedure is performed byproceeding triangularly inward. In this way, it can deploy the diagonal channels, whichalways have higher routing capacity and take advantage of empty spaces resulted frompreviously routed pads. The left diagram shows the pads that are routed in the first layer.The middle is the routing on the next layer. The right is the overview of this routingapproach.
In contrast to the intuitive ring-by-ring approach, this scheme was derived as a
maximum flow optimisation problem whereby the opening space left by routed pads in
lower layers are utilised maximally, as shown in the middle diagram of Fig. 2.6. This
resulted in a minimal number of layers required to route all the pads.
An n×n array will require at least k layers of routing, where k is the smallest integer
that satisfies the below inequality [190]:
−2(D+1)(D+2)k2 +[4(D+1)n−10D+8C]k ≥ n2, (2.5)
where C and D are the grid and diagonal capacities in Eq. (2.1) and (2.2), respectively.
2.5 Results 41
2.5 Results
Generally, in order to supply the electrical signals to the control gates or the readout
devices to perform qubit readout, each and every gate needs to be fanned out to connect
to the classical control systems. In conventional nanoelectronics, this is achieved by
overlaying the qubit array with many metal lines on several layers. Electrical connections
from these metal lines to the surface gates are made by vertical conducting “vias”. The
unique advantage of Si-based solid-state quantum platforms is the compatibility with the
classical CMOS electronics, whereby both can be integrated onto the same silicon chip.
Classical electronics can be fabricated outside the surface code qubit lattice as shown
in Fig. 2.7. At the bottom layer lies the semiconducting material substrate in which
qubits are realised and controlled by surface gates. Therefore, we need to fan the surface
gates out to the peripheral classical electronics area where classical processing tasks
are performed. Under the generic model considered here, regardless of the interconnect
protocols, the gate contacts/vias are redistributed into a square-grid array before global
fan-out routing is performed, as illustrated by metal routing layers shown in Fig. 2.7.
The fan-out scalability of 2D qubit arrays is examined by looking at the number of
routing layers required for complete routability. As shown in Fig. 2.7, multi-layer routing
can potentially provide unlimited fan-out capacity if we let the number of metal layers
be unbounded. However, in practice it is imperative to keep the number of metallization
layers to the absolute minimum - usually in the range of 10-15 layers for the most
advanced semiconductor products [119]. The technological and economic challenges
associated with fabricating many layers of nano-scale interconnects are going to be
similar for the various solid-state quantum computing approaches. In the following
analyses, we stretch to a 20 routing layer limit to benchmark the fan-out scalability of the
various quantum interconnect schemes. We will adopt two standard routing algorithms
from classical electronics, namely the ring-by-ring and the layer optimisation algorithms,
which are described in detail in the Methods section.
First, we look at the raw differences between the two routing algorithms at a fixed
interconnect length. The triangular routing (layer optimisation) algorithm is the most
efficient way to fan-out all contacts in terms of the required number of layers (see
Methods). This is shown in Fig. 2.8, where we examine both ring-by-ring and layer
42 Chapter 2. Solid-state Spin Qubit Control Routing
Fig. 2.7 Illustration of 2-D qubit lattice surface gate fanout using multiple metal routinglayers. The bottom layer is a semiconductor material (Si or GaAs) with top gates forcontrol and readout. On the same substrate lies classical integrated electronics usedfor signal generation, multiplexing, and sensing. In order to bring connections to thesurface gates, multi-layer routing is needed. After surface gates are redistributed intoa square-grid array of contacts, as shown in the first metal layer, the fan-out routingprocedure is carried out layer-by-layer using a specific routing algorithm. This figuredemonstrates ring-by-ring routing, which requires three metal layers for this particulargrid array. More sophisticated routing algorithms can be implemented using EDA(Electronic Design Automation) tools.
optimal routing solutions for the SSI and ECI protocols for L = 300 nm (which satisfies
the minimum interconnect length (14 nodes, 280 nm) for the atomically-confined SSI
scheme). It is worth noting that we use the same interconnect length to compute the
fan-out for electrostatically-confined qubits (Nq = 3). Because the qubit-qubit distance
is different, the number of qubit nodes in the interconnect chain varies across different
qubit configurations in the bar chart comparison (Fig. 2.8, right).
We observe a factor of 5 to 8 increase in the number of routable qubits by using
the optimal router across most of the scenarios (except for SSI, Nq = 1) as shown in
the right chart of Fig. 2.8. This highlights the fact that for large-scale qubit integration
the use of design-automation tool suites is important to achieve better routing solutions
2.5 Results 43
101 102 103 104 105
Number of physical qubits
100
101
102
103
Num
ber
ofla
yer
s
20-layer limit
Number of routing layers comparison for Nq = 1 (L=15, 300nm)
SSI Ring-by-ring
SSI Optimal
ECI Ring-by-ring
ECI Optimal
Nq = 1 Nq = 3 Nq = 1 Nq = 30
1000
2000
3000
4000
5000
6000
7000
8000
Num
ber
ofphysi
calqubits
End Control Interconnect Spin Shuttling Interconnect
27 27 58464
Number of qubits routable with 20 layers (L=300 nm)
Ring-by-ring
Optimal
Fig. 2.8 Comparison between different interconnect protocols and routing methodsat fixed interconnect length: (left) Plot of the number of routing layers vs. numberof physical qubits for atomically confined qubits (Nq = 1) at L = 300nm; and (right)scalability comparison between electrostatically confined qubits (Nq = 3) and atomicallyconfined qubits (Nq = 1) using the same number of routing layers (20) and interconnectlength (L = 300nm, 15 nodes for Nq = 1 and 6 nodes for Nq = 3). Dimension parametersare (see Methods): d = 10nm, w = 5nm, and s = 25nm. The red dashed horizontal line onthe left figure represents the technological limit of 20 metal layers that can be fabricatedreliably and economically on a semiconductor substrate.
compared to more intuitive and direct methods such as the ring-by-ring method. One
exception is in the case of SSI scheme for atomically-confined qubits (Nq = 1) where
both routing methods result in the same number of routable qubits. The reason is that
at this interconnect length (L = 300nm), the contact-contact pitch after redistribution
(eq. (2.3)) is barely above the minimum metal-metal pitch requirement, thus no escape
routing (wires between pads) is allowed.
Another point which can be seen from Fig. 2.8 is the advantage of the ECI scheme
over its SSI counterpart in terms of fan-out scalability. This naturally stems from the fact
that ECI protocols require far fewer surface gates than their SSI counterparts (Table 2.1).
We will later investigate the scaling differences between the two schemes in details by
looking at various interconnect lengths. The difference in qubit-qubit spacing manifests
itself in the opposite trend observed in Fig. 2.8 bar chart: while atomically-confined
qubits are the clear winner in the ECI scheme, the opposite is true if the shuttling
interconnect scheme is assumed. This can be explained by the minimum interconnect
length data in Table 2.2, i.e. while Nq = 1 has shorter minimum interconnect requirement
in the ECI scheme, it has a much longer minimum interconnect length in the SSI scheme.
44 Chapter 2. Solid-state Spin Qubit Control Routing
The scalability difference in the ECI scheme between Nq = 1 and Nq = 3 is narrowed
significantly if we use optimal routing for electrostatically-confined qubits (from 2x
to about 25% different). In all the following analyses, we will only consider triangle
(layer optimised) router since this will better reflect the realistic engineering solution.
To highlight the effect of even more confining gates (e.g. double-dots as qubits), we also
perform the analysis for a hypothetical case of Nq = 5.
All the scenarios that we consider so far are homogeneous in the sense that corner
qubits and interconnect qubit nodes are of the same type. For SSI scheme, we can
implement a hybridisation protocol in which atomically confined qubits are used as
surface code physical qubits, while electrostatically confined qubits are utilised for
interconnect coupling. By doing this, we can achieve the best of both worlds for the SSI
scheme, namely minimising Nq and maximising qubit distance. This approach is only
effective for SSI scheme since for ECI the number of interconnect gates is constant.
Fig. 2.9 shows the fan-out scalability in terms of routing layers (optimised router) for
both the SSI and ECI protocols with interconnect length of 300 nm, 450 nm, and 600 nm
(15/20/30 nodes and 6/8/12 nodes for Nq = 1 and Nq = 3/5 or SSI hybrid, respectively).
In the top graphs, the fan-out scaling of atomically confined qubits (Nq = 1) is analysed
in detail to provide a reference and the horizontal line represents the limit of 20 metal
layers as previously explained. Other qubit configurations (Nq = 3/5 and SSI hybrid)
are compared to this reference in the bottom graphs.
An obvious conclusion which can be drawn from both the left charts in Fig. 2.9 is
that the SSI protocol does not provide a consistent fan-out scaling benefit as compared
to its ECI counterpart, as there is no clear trend in terms of the number of routable
qubits vs. interconnect length. The main contributing factors to this fluctuating trend
are the opposite effects of redistributed pitch extension, the increasing number of gates
per unit cell and the granularity of the routing problem (only full routing channels are
considered). On the other hand, ECI protocols provide a monotonic improvement in
terms of the number of integrated qubits vs. interconnect length because the interconnect
length (thus metal pitch) is Nc-independent. The order of routability vs. interconnect
length for different qubit configurations is reserved for both SSI and ECI schemes. While
the former interconnect scheme favours electrostatically-confined qubits due to their
2.5 Results 45
101 102 103 104 105
Number of physical qubits
100
101
102
103N
um
ber
ofla
yer
s
20-layer limit
Number of routing layers for SSI scheme (Nq = 1)
L = 15 (300nm)
L = 20 (400nm)
L = 30 (600nm)
101 102 103 104 105
Number of physical qubits
100
101
102
103
Num
ber
ofla
yer
s
20-layer limit
Number of routing layers for ECI scheme (Nq = 1)
L = 15 (300nm)
L = 20 (400nm)
L = 30 (600nm)
L = 300 nm L = 400 nm L = 600 nm0
100
200
300
400
500
600
700
Num
ber
ofphysi
calqubits
20 20
100
Number of qubits routable with 20 layers - SSI
Nq = 1
Nq = 3
Nq = 5
Hybrid
L = 300 nm L = 400 nm L = 600 nm0
10000
20000
30000
40000
50000
Num
ber
ofphysi
calqubits
Number of qubits routable with 20 layers - ECI
Nq = 1
Nq = 3
Nq = 5
Fig. 2.9 Qubit fanout scalability in the cases of interconnect protocols of (left) SSI and(right) ECI. (Top) Number of routing layers vs. number of physical qubits for atomicallyconfined qubits (Nq = 1) under different interconnect lengths; and (bottom) numberof routable qubits comparison between electrostatically confined qubits (Nq = 3 andNq = 5), atomically confined qubits (Nq = 1), and hybrid SSI (donors as qubits and dotsas shuttling nodes). Dimensional parameters are: d = 10nm, w = 5nm, and s = 25nm.The red dashed horizontal line on top charts represents the technological limit of 20metal layers that can be fabricated reliably and economically on a solid-state substrate.
long qubit-qubit spacing, the later scheme suits atomically-confined qubits a little bit
better thanks to the reduced number of confining gates needed. The hybrid SSI scheme
outperforms both of its homogeneous SSI counterparts but noted that the best it can
achieve is still an order of magnitude less than that of the ECI scheme.
To assess the fan-out scalability of interconnect protocols over extreme length scale,
we extend the interconnect length further (up to 100 intermediate nodes for Nq = 1,
i.e. 2 µm). The result is shown in Fig. 2.10 for the SSI and ECI schemes. This
analysis provides a concrete example to the scaling bottleneck of the SSI protocols in 2D
qubit lattice implementation (only routable up to about 103 qubits for electrostatically
confined qubits and about 200 for atomically confined qubits). The maximum number
of routable qubits is saturating over long SSI interconnect length for both types of qubits.
46 Chapter 2. Solid-state Spin Qubit Control Routing
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
SSI interconnect length (µm)
0
200
400
600
800
1000
1200
Num
ber
ofphysi
calqubits
Number of qubits routable with 20 layers vs SSI interconnect length
Nq = 1
Nq = 3
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
ECI interconnect length (µm)
0
100000
200000
300000
400000
500000
600000
700000
Num
ber
ofphysi
calqubits
Number of qubits routable with 20 layers vs ECI interconnect length
Nq = 1
Nq = 3
Fig. 2.10 Fan-out scalability vs. interconnect length for (left) SSI and (right) ECIschemes. The inter-qubit interconnect lengths are given in absolute unit (µm). Thenumber of interconnect qubit nodes can be inferred by noting that the qubit-qubit distanceis 20 nm for Nq = 1 and 50 nm for Nq = 3.
Multiplexing schemes for SSI control, e.g. Ref [86, 71], will improve the scalability of
these approaches to a certain extent. On the other hand, the ECI protocol can scale up
(quadratic) to an order of 105 qubits over that length scale. Again, as we have already
seen in Fig. 2.9, there is an incremental improvement in terms of scalability at the same
interconnect length when using electrostatically self-confined qubit structures due to
their gate count efficiency. Further steps can be taken to estimate the number of logical
qubits feasible based on the level of error correction required, namely the code distance.
The latter depends on multiple factors such as the gate fidelity, total number of gates
in the algorithm of interest, and the level of output accuracy required. The analysis
in ref [43] provides estimates of the qubit resource required for surface code quantum
computation.
In addition to quantifying the spatial requirements for scale-up, as we have here,
there are undoubtedly other aspects which may restrict the scalability of solid-state
qubit integration, namely control timing, signal integrity, thermal budget, testability, and
manufacturability . Nevertheless, being able to model and extrapolate the limit of each
of the scaling bottlenecks is good engineering practice.
2.6 Summary
If the advancement in solid-state spin qubit fabrication and control follows that of their
classical counterparts, the number of integrated qubits will soon reach the threshold
2.6 Summary 47
where scaling up becomes the next bottleneck. As the quantum network gets larger and
larger to cope with real-world applications, the amount of routing required to provide
control access to surface gates will soon become the limiting factor. By applying the
classical electronics know-how regarding interconnect routing to a ubiquitous 2-D qubit
array with independent gate control and readout fan-out, we have provided a concrete
procedure for scalability estimation, which is adaptable to a wide range of surface code
implementations by adjusting the gate configuration and dimensional parameters. This
estimation procedure is important for large-scale quantum processor design process
where we need to identify at the very early stages the required specifications (so-called
“landing zones” in classical electronics design) regarding quantum interconnect length
and fidelity, back-end metal interconnect dimensions and the number of fan-out layers.
For architectures where each qubit has its own dedicated control lines and readout
device, we have analysed fan-out scenarios associated with two categories of quantum
interconnects, namely spin shuttling interconnects (SSI) and end control interconnects
(ECI) with high and low gate densities respectively. Both interconnect models help
extend the contacts/vias pitch through redistribution, which potentially aids the fan-out
routing procedure. However, SSI protocols result in a poorly scalable situation since the
added interconnect control gates outweigh the pitch scaling benefit. On the other hand,
ECI protocols provide a more consistent fan-out scaling trend with interconnect length,
however, relatively long interconnects (greater than several microns) are required to scale
the system to the million qubit level, where issues such as interconnect fidelity, charac-
terisation and operation time, will affect the error rate and surface code error correction
performance negatively. Above all, the errors induced in the quantum interconnect must
be correctable by the QEC protocol. Multiplexing schemes [74, 12] alleviate the gate
density bottleneck to some extent, with fully distributed control schemes [71] providing
scalability without the need for quantum interconnects.
This page intentionally left blank.
Chapter 3
Continuous Quantum Error
Correction
Almost any decision is better than no decision – just keep moving.
Danielle LaPorte
Fighting decoherence in quantum mechanical systems, due to their inevitable cou-
pling to the surrounding environment, often requires a combination of multiple ap-
proaches such as fabrication techniques, decoupling pulses, quantum feedback control
methods and coding schemes. In this chapter, we will not target a particular quantum
computing architecture but instead focus on the feedback and control scheme for a
promising quantum error correcting coding scheme, namely the surface code. The
approach we chose is an extension of the continuous quantum error correction scheme
proposed by Ahn and co-workers for simple codes [4], where continuously acquired
syndrome measurement signals are used to correct the errors in real-time via Hamiltonian
feedback.
3.1 Quantum Errors
The merits of quantum error correction are often evaluated concerning a specific error
model. In conventional error models, the Pauli X , Y , and Z errors are applied probabilis-
tically to qubits at discrete time intervals. For example, a noise model M is a sequence
50 Chapter 3. Continuous Quantum Error Correction
of quantum operations St(ρ) on the Hilbert space (density matrix, ρ) of the quantum
computer. These error operations are indexed by the time-step (t) at which each of them
is applied.
A general quantum operation, ρ 7−→ S(ρ), is a linear, trace preserving, and com-
pletely positive map1.
Theorem 3.1.1 [Kraus (operator-sum) decomposition]
S is a completely positive trace preserving map iff
S(ρ) = ∑k AkρA†k , with ∑k A†
kAk = I
These are common quantum error channels [122] that are used to model decoherence
in quantum computers:
1. Dephasing channel:
S(ρ) = (1− p)ρ + pZρZ
2. Depolarizing channel:
S(ρ) = (1− p)ρ +p3(XρX +Y ρY +ZρZ),
which can also be expressed using the representation:
S(ρ) = (1− 4p3)ρ +
4p3
(I2
).
Note: the identity operator (I) is the balanced mixture, i.e.I2 = ρ
4 +14XρX + 1
4Y ρY + 14ZρZ
3. General Pauli channel
S(ρ) = (1− px − py − pz)ρ + pxXρX + pyY ρY + pzZρZ
4. Amplitude damping channel
S(ρ) = A0ρA†0 +A1ρA†
1,
1This is equivalent to the Master equation formulation in (1.10).
3.1 Quantum Errors 51
where A0 =
1 0
0√
1− γ2
and A1 =
0 γ
0 0
.
In the above models, the parameter p represents the error probability at each time
step. Hence, we refer to this as the discrete error model. In physical quantum systems
such as atoms, photons, or spins, however, the amplitude and phase of a qubit will often
fluctuate continuously when it is interacting with the environment or other quantum
systems.
It is worth noting that the parity measurement in quantum error correction scheme
provides a discretisation of the set of errors. Hence, the discrete model of QEC is capable
of modelling general continuous decoherence processes. However, continuous errors
can also happen to parity measurements themselves which makes the fully digitised
picture an oversimplification.
3.1.1 Quantum Error Correction Code
A universal approach in information theory to deal with noise-corrupted data is to intro-
duce auxiliary qubits to carry away extra entropy transferred from the noisy environment
into the system. In essence, we need to encode the data (logical qubits) into a larger
codeword that has extra qubits. These added qubits will assist us in identifying the most
probable candidate error, of several possible errors, occurred, so that corrective action
can be taken to preserve the stored quantum information.
Among many quantum error correction (QEC) schemes, stabilizer codes [53] are pre-
dominantly the most effective and well-studied category, to which the surface code [43]
also belongs. A stabilizer code is described by an Abelian (mutually commuting) set of
Pauli operators whose simultaneous +1-eigenspace defines the code space. This set is
called “stabilizer generators” (S ). Stabilized code words (L ) are defined as:
L = |ψ⟩ ∈ (C2)⊗n : P |ψ⟩= |ψ⟩ ,∀P ∈ S .
Therefore, stabilizers act trivially on the code and can be measured simultaneously.
Their eigenvalues form an error syndrome that can help identify the error coset where
the actual error resides. By applying an effective decoding policy (such as minimum
weight matching), the most probable error candidate will be chosen to be corrected.
52 Chapter 3. Continuous Quantum Error Correction
To build a large scale quantum computer, an error correction code is required through
which the error can be made arbitrarily small. In this regard, not all stabilizer codes scale
equally due to their specific construction. Some codes may become more susceptible
to noise on a larger block than the others. This figure of merit is characterized by
the accuracy threshold which is the bound of error rate that the code can tolerate for
effective scaling. Among all stabilizer codes, surface code has been proven to be one of
the best-performing codes with a threshold two orders of magnitude higher than other
error correction codes [45] along with many desirable features that we will explain in
the next section.
3.1.2 Surface Code
The surface code is a type of topological QEC code where qubits are laid out in a
2-dimensional lattice. This planar layout, in combination with nearest-neighbor interac-
tions, makes it a compelling alternative to the conventional stabilizer codes which often
require long-range interactions.
Similar to other stabilizer codes, a surface code is fully characterized by its stabilizer
generators. The stabilizer generators of the surface code come in two types: the so-called
X and Z stabilizers, as depicted in Fig. 3.1.
1 2 3 4
5 6 7
8 9 10 11
12 13 14
15 16 17 18
X X X
Z Z Z Z
X X X
Z Z Z Z
X X X
Fig. 3.1 Surface code layout: white circles represent data qubits; filled circles aresyndrome qubits (X stabilizers in green and Z stabilizers in yellow). Each internalstabilizer acts on four adjacent data qubits, while boundary stabilizers act on either twoor three data qubits.
These local stabilizers are mutually commuting by construction (X and Z stabilizers
either share none or two qubits) and thus form a valid set of stabilizer generators
3.1 Quantum Errors 53
for a QECC. These stabilizers fix a set of states, the so-called “quiescent” states, the
simultaneous eigenstates of all generators. A codeword is also a quiescent state with all
+1 eigenvalues.
Error decoding is based on the topological properties of the lattice whereby a bit-flip
(X) is detectable because it causes the change of the two neighboring Z syndromes; a
phase-flip (Z) by the change of the two neighboring X syndromes and a bit-and phase-flip
(Y ) by both types of syndrome. This is a simplistic explanation for surface code decoding
where we skip details about space-time correlations to detect qubit vs. measurement
errors and the standard minimum weight perfect matching algorithm used for error
decoding. Fowler et al. provides an excellent review for interested readers [43].
The code distance of a surface code can be quickly inferred from lattice dimension.
When using the optimal lattice structure [75], an N ×N (data qubit) lattice forms a code
of distance N. By definition, a distance-d code can correct arbitrary ⌊d−12 ⌋-qubit errors.
Thus, a distance-of-3 code is the smallest universal error correction code that can correct
any single qubit error.
3.1.3 Previous Work
QEC and fault tolerant quantum computing have been studied extensively in the last two
decades mostly focusing on the discrete time scheme. Discrete QEC bridges the gap
between the conventional computer science and the emerging quantum computing field.
However, it relies on the fact that we can do projective measurements and instantaneous
corrections. Therefore, research has been done regarding the dynamical representation
of the QEC process where errors, syndrome measurement, and correction are performed
continuously with finite strengths.
The very first framework for the continuous QEC was laid down by Paz and
Zurek [142] where continuous error correction is modelled as an infinitesimal limit
of syndrome measurement and error correction. The resulting master equation is of
the type of a cooling process where the ancillas are continuously “cooled” from the
entropy acquired by errors. Sarovar and Milburn used this cooling approach in the
3-qubit flip code [153]. Oreshkov and Brun studied continuous quantum error correction
for non-Markovian decoherence using the model of a bit-flip code [131]. Later, Hsu
54 Chapter 3. Continuous Quantum Error Correction
and Brun also developed a method for continuous-time quantum error correction for any
[[n,k,d]] quantum stabilizer code [78].
This chapter on the other hand is inspired by the measurement feedback scheme that
was first developed by Ahn et al. for a simple 3-qubit code [4]. Later, elaboration on
this continuous scheme was done using classical signal filtering techniques [152], and
instantaneous emission syndromes [5].
Classical hybrid control techniques were also proposed for quantum continuous error
correction, in which the Wonham filter was used for error state probability tracking [179].
A filter dimension reduction scheme was also derived [21] for the Wonham filter with
Hamiltonian feedback. Mabuchi described an optimal hybrid controller for the 3-qubit
flip code [106]. [151] implements a continuous-time error correction protocol for
stabilizer codes and discusses some subtleties of routing for the accumulation of multi-
qubit syndromes and how that interacts with the subsystem nature of some stabilizer
codes.
Previously-mentioned control techniques are all referred to as measurement-based as
they involve quantum to classical signal conversion. The other type is coherent quantum
feedback control whose feedback loop only contains quantum mechanical devices. A
coherent feedback scheme for the 3-qubit error correction was shown in [90], in which
the feedback loop forms a fully autonomous self-correcting quantum memory without
the need for classical processing.
In terms of the topological error correction, discrete error correction schemes were
investigated by Fowler et al. [44] and Stephens [162], where an accuracy threshold was
derived for the surface code. This proved that the surface code is an outstanding candidate
for large scale quantum computer architecture. More sophisticated computational
simulations [174] were recently done for the minimal surface code of distance 3 under
various error models in the discrete framework.
A comprehensive review of QEC is presented in the book [99]. A more focused
review on error correction for quantum memory is presented in [166] where continuous
QEC techniques are also summarized.
3.1 Quantum Errors 55
3.1.4 Contributions
The aim of this chapter is two-fold. Firstly, we want to introduce continuous QEC by
measurement feedback for topological codes such as the surface code. This is done
using fully quantum mechanical representations in terms of stochastic master equations
and quantum stochastic differential equations. It provides a complementary view to the
conventional semi-classical circuit model approach of surface code error correction. In
order to keep the simulations numerically tractable, we also chose a reduced 2×2 lattice
containing only four data qubits affected by a limited error model. This can be regarded
as the smallest 2-D structure that still exhibits the non-trivial topological properties
which we want to investigate. Secondly, we have put this feedback scheme in the SLH
framework that is emerging as a useful tool for control engineering.
3.1.5 Outline and Notations
After the general introduction in Section 3.1, we will proceed to describing our surface
code model in Section 3.2 that includes the description of the qubit network, dynamical
equations, and the SLH model. Section 3.3 will detail the feedback policy for continuous
error correction followed by simulation results in Section 3.4. Section 3.5 will conclude
the chapter with some final words and avenues for future research.
In this chapter, Pauli operators σx, σy, σz are also represented by capital letters X ,
Y , and Z, respectively. To shorten the notation, tensor product signs are omitted, i.e.
X1X2 ≡ X1⊗X2, and identity operators (I) are assumed in all places where no other Pauli
operator applied, such as X1X4 ≡ X1 ⊗ I ⊗ I ⊗X4.
The commutator between operators is denoted by [A,B] = AB−BA, whilst anti-
commutator is A,B = AB+BA. Hermitian conjugate of A is A†. The Heisenberg-
picture evolution of an observable X is denoted by Xt = jt(X).
We will also use a couple of common superoperators such as the dissipation (D),
Lindblad generator (LL,H), and the homodyne measurement (H ), which are defined as:
D [L]ρ = LρL† − 12(L†Lρ +ρL†L),
LL,H(X) = i[H,X ]+D [L]X ,
56 Chapter 3. Continuous Quantum Error Correction
H [L]ρ = Lρ +ρL† −ρTr[Lρ +ρL†].
3.2 Model
In this section, we will first introduce the distance-2 surface code block that will be
used to demonstrate the continuous feedback scheme. Then, we will briefly review key
concepts of the SLH dynamical model before deriving the detail model of our surface
code in the continuous measurement feedback scheme.
3.2.1 Distance-2 Surface Code
The surface code lattice that we will consider in the following sections comprises of
four data qubits and five syndrome measurements in a 2×2 configuration as shown in
Fig. 3.2. The code stabilizer generators are Z1Z2Z3Z4, X1X2, X1X3, X2X4, and X3X4.
1
2 3
4
Z
XX
XX
Fig. 3.2 Distance-2 surface code
This code lattice has a code distance of two which is insufficient to correct arbitrary
error such as the general depolarizing error: ρ 7→ (1− p)ρ + p3 (XρX +Y ρY +ZρZ).
Indeed, X errors cannot be localized by this lattice since all four X errors have the same
syndrome. However, single-qubit Y and Z errors can be identified by combining the five
syndrome measurements.
Unlike discrete QEC, the continuous syndrome measurement does not require a
physical qubit but only a field interacting with the relevant data qubits with some specific
coupling parameters. The output fields will carry both information about the syndrome,
3.2 Model 57
as well as additional noise. This information can be used to estimate the original
syndrome. This input-output process as well as the filter and estimation will be described
by the quantum stochastic differential equations.
3.2.2 Continuous QEC in the SLH Framework
Our continuous QEC feedback network will be described in the Heisenberg-picture
using the Hudson-Parthasarathy quantum stochastic differential equation (QSDE) [79],
which is also known as the input-output formalism [49] in the physics literature. The
QEC network is encapsulated by the SLH parametrization [56], which comprises of
scattering matrix S, coupling vector L, and Hamiltonian H. Unless otherwise stated, we
assume there is no scattering between quantum fields, so S = I.
In the continuous QEC scheme, our surface code is coupled to two separate groups
of channels. The first is a collection of error channels (LE) which depend on the chosen
error model. For generic balanced depolarizing channels, the error coupling vector for a
collection of N qubits with per-qubit error rate of γ is:
LE =
√γ
3σ( j)i
j∈[1,N]
i=x,y,z.
This corresponds to a per time step error probability of γdt for each qubit.
As noted before, we will use a limited error model for our distance-2 surface code
that only includes Y and Z errors but not X . There are therefore a total of eight error
channels coupling to the surface code lattice in Fig. 3.2. Despite having separate channels
for errors, we will assume that all of these are unobservable. Indeed, if we can access the
error channels, there will be no need for syndromes as errors can be corrected directly.
In order to measure the error syndromes, we need to create five syndrome field
channels with equal coupling strength, denoted by κ:
LS =√
κ [Z1Z2Z3Z4,X1X2,X1X3,X2X4,X3X4]T .
58 Chapter 3. Continuous Quantum Error Correction
SurfaceCodeArray Estimator Controller
dAS,LS
dAE ,LE dAE,out
dAS,out S(t) F(t)
Fig. 3.3 Block diagram of surface code continuous error correction in the SLH framework:the surface code array is coupled to two groups of channels, namely error channels (dAE)and syndrome channels (dAS). The coupling strengths are LE and LS, respectively. Thesyndrome outputs are measured to estimate the syndrome conditional expectation values(S(t)). The Hamiltonian feedback (F(t)) is a function of the syndrome estimators.
The unitary evolution of the surface code network can be derived using the Hudson-
Parthasarathy QSDE:
dU(t) =
dA†E(t)LE −L†
EdAE(t)−12
L†ELEdt
+dA†S(t)LS −L†
SdAS(t)−12
L†SLSdt
U(t),
(3.1)
where we assume no internal dynamics (H = 0). The QSDE for an arbitrary observable
X on the qubit network is given by:
d jt(X) = jt(LLE(X)+LLS
(X))dt
+ jt([L†E ,X ])dAE(t)+ jt([X ,LE ])dA†
E(t)
+ jt([L†S,X ])dAS(t)+ jt([X ,LS])dA†
S(t).
(3.2)
The measurement signals are taken to be of the forms of homodyme detection on the
syndrome field channels, i.e. dY (t) = dAS(out)(t)+dA†S(out)(t).
Assuming that the input fields are in vacuum state, the five output equations can be
explicitly derived for the surface code lattice in Fig. 3.2:
dYi(t) = 2√
κ jt(Si)dt +dWi(t), (3.3)
3.2 Model 59
where dWi(t) = dASi(t)+dA†Si(t) is equivalent to a classical Wiener process2, that is a
Gaussian distributed random variable with zero mean and variance of dt.
By observing the output field dYi(t), we can write down the optimal estimate of an
system observable using quantum filtering techniques:
dπt(X) = πt(LLE
(X)+LLS(X))
dt (3.4)
+ ∑iπt(LSi,X)−2πt(LSi)πt(X)dWi(t),
where we have used the Hermitian property of syndrome operators (LSi = L†Si) to simplify
the equation. The notation πt(X) stands for the conditional expectation of the observable
X given all measurement records up to time t.
So far, we have not made any simplification by using assumptions about the error
model or observable operators. Despite using a limited error model for our distance-2
lattice, we derive the filtering estimation equations for general depolarizing noise as this
will definitely be used for larger lattice with full error correction capability.
Denote gi the set of syndrome generators, by using (3.4) we can get their filtering
equations:
dπt(gl) =−(4w)γ3
πt(gl)dt
+2√
κ
k
∑i=1
πt(glgi)−πt(gl)πt(gi)dWi,
(3.5)
where w denotes the Pauli weight of the parity operator gl .
The filtering equations for syndrome measurements are non-linear due to the appear-
ance of high-order terms like πt(gl)πt(gi). Also, they expand beyond the initial stabilizer
generator set because of operator product terms like πt(glgi). However, they will form a
finite set of equations which include all operators generated by the group of syndrome
generators.
Having the estimation of the syndrome state, we can proceed to the next step to
build an estimation-based feedback controller to correct the surface code lattice as
diagrammatically shown in Fig. 3.3.
2assuming homodyne detection scheme
60 Chapter 3. Continuous Quantum Error Correction
3.3 Methods
The most widely used feedback strategy is to use an additional Hamiltonian to control
the system of interest based on information acquired at the outputs. As the errors are
occurring, the feedback Hamiltonian will be used to rotate the surface code back to the
code space continuously. Therefore, codeword fidelity will be preserved.
Each type of error requires a corresponding correction Hamiltonian, thus our feed-
back Hamiltonian has the following form:
F(t) =4
∑i=1
λzi (t)Zi +
4
∑i=1
λyi (t)Yi (3.6)
where λzi (t) and λ
yi (t) are the real-time feedback terms that will depend on the estimated
syndrome signatures.
The topological structure of the code allows us to infer the error by identifying two
neighboring syndrome values which change their signs from +1 to -1. We can use that
logic to define the feedback policy as:
λz1 =
λ032
(1− X12)(1− X13)(1+ X24)(1+ X34)(1+ Z1234), (3.7)
λz2 =
λ032
(1− X12)(1+ X13)(1− X24)(1+ X34)(1+ Z1234), (3.8)
λz3 =
λ032
(1+ X12)(1− X13)(1+ X24)(1− X34)(1+ Z1234), (3.9)
λz4 =
λ032
(1+ X12)(1+ X13)(1− X24)(1− X34)(1+ Z1234), (3.10)
λy1 =
λ032
(1− X12)(1− X13)(1+ X24)(1+ X34)(1− Z1234), (3.11)
λy2 =
λ032
(1− X12)(1+ X13)(1− X24)(1+ X34)(1− Z1234), (3.12)
λy3 =
λ032
(1+ X12)(1− X13)(1+ X24)(1− X34)(1− Z1234), (3.13)
λy4 =
λ032
(1+ X12)(1+ X13)(1− X24)(1− X34)(1− Z1234), (3.14)
where we have dropped the time variable for brevity. The maximum feedback strength
λ0 is assumed to be time-independent and equal for all feedback terms. The hats over
the operators represent the conditional expectations, i.e. X12 = πt(X1X2).
3.4 Results 61
The Z errors are detected by X syndrome measurement alone, while Y errors are
detected by combining X syndromes with the central Z syndrome. For example, a Z
error on qubit number 1 will change both X1X2 and X1X3 signs, thus turning on the
feedback term λz1 . On the other hand, a Y error on the first qubit will also change the
center Z syndrome therefore we got a minus sign in the last term.
The stochastic differential equations are solved numerically using an Euler solver.
The time step, measurement and feedback strength will be normalized to the decay rate
(γ). A time step of order 10−5 is used for the solver which yields good convergence over
wide range of measurement and feedback strengths.
Unless otherwise noted, the default number of Wiener noise realisations is 104,
which will be averaged over. We use codeword fidelity, defined by Tr[ρ0ρ(t)], as the
primary figure of merit to assess the performance of our feedback scheme. A fidelity of
1 signals perfect match, while 0 means orthogonal states.
Also, to judge the continuous scheme against the conventional discrete error cor-
rection, we will also look at the correctable overlap projection which is defined as the
combination of the original state and all single-error subspaces:
ΠC = ρ0 +∑i, j
σ( j)i ρ0σ
( j)i . (3.15)
The expectation of this projector at time t gives us an upper bound estimate of the
recoverable fidelity if a discrete error correction cycle is to be performed at that time [4].
3.4 Results
3.4.1 Distance-2 Surface Code
Using our surface code model as described above, we can simulate the evolution of
the qubit array stochastically using numerical methods. A sample simulation run for
one realization of the noise process, also known as trajectory, is shown in Fig. 3.4. We
can clearly see the variation in our syndrome estimates due to noise. However, it also
indicates a clear transition from +1 to -1 of X1X3, X3X4, and Z1Z2Z3Z4 generators.
By inferring from the trajectory, we suspect that a Y3 error is the most probable error.
However, we need some quantitative measures to assess the claim that it is indeed the
62 Chapter 3. Continuous Quantum Error Correction
0.00 0.02 0.04 0.06 0.08 0.10
Time (1/γ)
−1.0
−0.5
0.0
0.5
1.0Sy
ndro
me
Val
ueSurface Code Syndrome Estimation
Z1Z2Z3Z4
X1X2
X1X3
X2X4
X3X4
Fig. 3.4 Syndrome estimation for a single trajectory showing +1 to −1 transitions fromthree syndrome operators: X1X3, X3X4, and Z1Z2Z3Z4. This indicates a Y error mayoccur to qubit 3. Dashed line represents the point in time selected for the fidelity andtrace distance comparison in table 3.1.
case. Nielsen and Chuang have outlined two effective measures which are the trace
distance and fidelity [122]. The first measures how far apart the two states are, whilst the
latter measure the opposite. In our simulator, we can extract the actual state at the time
right after the transition as denoted by dotted line in Fig. 3.4. We call this inferred state
ρ and we will compare it against the initial state ρ0 as well as the Y3-rotated version
of it (ρ ′0 = Y3ρ0Y3). The comparison results are summarized in table 3.1. This shows
that the latter time state (ρ) is much more matched to the Y3-rotated initial state than to
the ρ0 itself. We can then apply our feedback correction policy laid out in Section 3.3.
Measure (ρ0,ρ) (ρ ′0,ρ)
Trace Distance: D(ρ,σ) = 12Tr|ρ −σ | 0.999 0.069
Fidelity: F(ρ,σ) = Tr√
ρ1/2σρ1/2 0.018 0.998
Table 3.1 Fidelity and trace distance comparison
The performance of our continuous error correction against no correction as well as
correctable overlap (discrete correction bound) is shown in Fig. 3.5.
The continuously corrected fidelity is clearly much better than no correction at all,
as the feedback loop is correcting the errors in real-time while they are occurring. More
3.4 Results 63
0.0 0.1 0.2 0.3 0.4 0.5
Time (1/γ)
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Fide
lity
Fidelity Comparison
No FeedbackFeedbackCorrectable Overlap
Fig. 3.5 Comparison between continuous feedback error correction fidelity vs. nofeedback fidelity and correctable overlap. Feedback enhances the code fidelity over longperiod of time. The cross-over between no feedback correctable overlap and feedbackfidelity indicates continuous error correction outperforms its discrete counterpart ifcorrection period is longer than that time.
importantly, over long period of time3, the continuous error correction scheme even
outperforms its discrete counterpart. Note that in order for the discrete scheme to become
more preferable, we need to perform correction cycles frequently enough while still
keep the idealized projective measurement and fast correction assumptions.
3.4.2 Distance-3 Surface Code
To simulate a generic error model, we extend the surface code lattice to a code distance
of three. The qubit lattice is shown in Fig. 3.6.
In the following simulation, we will introduce depolarisation noise source to each
qubit. The noise coupling operators have the form:
L(i)e =
γ
3(σ
(i)x +σ
(i)y +σ
(i)z ), (3.16)
in which we assume the noise strength is the same for all qubits and the noise is balanced
among all Pauli operators. These assumptions can be relaxed when considered specific
physical systems.
3in terms of error correction cycles
64 Chapter 3. Continuous Quantum Error Correction
7 8 9
4 5 6
1 2 3
Z X
X ZZ
X
Z
X
Fig. 3.6 Diagram of distance-3 surface code: white circles are data qubits, green circlesare X-stabilizer channels, and yellow circles are Z-stabilizer channels.
The Hilbert space of this qubit network is quite large. It requires a complex matrix
of the size of 29 ×29 to represent the density matrix. Fully stochastic master equation
simulation at this scale, as performed for the distance-2 lattice, demand significant
computing resources. Thus, to simplify the simulation, we can convert the stochastic
master equation into its counterpart stochastic Schrödinger equation. For a system of
size N, the latter only requires 2N complex elements. The drawback of this approach is
that we are unable to capture inefficient detection since it will induce mixed states.
A stochastic Schrödinger equation for our scheme has the following form:
d |ψ(t)⟩ = −iH |ψ⟩dt +12
(⟨L+L†⟩L−L†L− 1
4⟨L+L†⟩2
)|ψ⟩dt (3.17)
+ (L− 12⟨L+L†⟩) |ψ⟩dW.
We can verify the equivalence between this and the previous stochastic master equa-
tion by using: dρ = d(|ψ⟩⟨ψ|) = d(|ψ⟩)⟨ψ|+ |ψ⟩d(⟨ψ|)+ d(|ψ⟩)d(⟨ψ|) following
Ito rule with the convention: dt2 = 0 and dW 2 = dt.
The surface code in Fig. 3.6 is a degenerate code, which means that more than
one errors are associated with a specific syndrome signature. Nevertheless, errors with
3.4 Results 65
the same syndrome will constitute a stabilizer operator, thus acts trivially on the code.
Therefore, we can correct the error up to modulo of stabilizer operator, e.g. Z1 and Z4
errors have confounding syndrome signature as Z1Z4 is a stabilizer, but correcting one
of them is enough to preserve codeword fidelity.
Using the feedback policy similar to the distance-2 case, we can simulate the perfor-
mance of the continuous QEC for the surface code under general depolarisation noise
as in (3.16). The average code-word fidelity over time of the surface code is shown
in Fig. 3.7, which includes 1000 random noise realisations each. In all the range of
feedback rates considered, the feedback loop, in the long run, has always delivered
some improvement in fidelity. This proves the robustness of the scheme over feedback
parameters.
Fig. 3.7 Time-domain simulation of the distance-3 surface code under continuous QEC.Each curve is the average fidelity over 1000 stochastic trajectories with a specificfeedback correction strength.
From the above simulation, we can also see that the fidelity performance varies
considerably with regard to the feedback strength. This indicates an optimisation
opportunity of the parameter to deliver the best performance. This can be considered as
66 Chapter 3. Continuous Quantum Error Correction
a classic feedback control problem where there is an optimum operating point. Too weak
or too strong feedbacks are both harmful to the performance as can be seen in Fig. 3.8.
0.0 0.2 0.4 0.6 0.8 1.0
λ/κ
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Fide
lity
Fidelity Plot vs. Feedback Strength
Fig. 3.8 Comparison of the final fidelity of the surface code under varying feedbackstrength.
The steep increase in the fidelity shows the significant gain when we have the feed-
back loop in place. Also, the feedback is relatively weak with regard to the measurement.
This makes perfect sense since the controller needs to average over long, noisy syndrome
records to get an accurate estimation of the errors.
3.5 Conclusions
In summary, we have shown a continuous measurement feedback scheme that can be
used for topological codes such as the surface code. The feedback correction preserves
the codeword fidelity over long duration against a decoherence process. Syndrome
values are estimated from measurement signals without the need for physical ancilla
qubits as well as quantum gates. This proves the feasibility of using feedback and control
methods in topological QEC. The QSDE formalism and SLH model provide us with an
3.5 Conclusions 67
efficient tool suite to tackle quantum feedback and control problems. Given the growing
interest in surface code among other alternative codes, this deems a promising method
to be used in quantum computers whereby low-level feedback control loop maintains
the fidelity of quantum memory.
We have just examined our feedback control policy on a minimised surface code
lattice under idealistic conditions. Besides, non-ideal effects, such as loop delay, limited
measurement bandwidth, measurement error, which are always present in reality will
also need to be investigated in the future to assess the robustness of our scheme. The
ultimate goal is to envision and design a coherent feedback controller for the surface
code which can make our surface code a self-correcting qubit lattice in the quantum
domain.
This page intentionally left blank.
Chapter 4
Quantum Reservoir Engineering
The sky is filled with stars, invisible by day.
Henry Wadsworth Longfellow
The environment within which the quantum system operates typically has a contin-
uous degrading effect, i.e. decoherence, on evolution of quantum particles. Reservoir
engineering, also known as quantum dissipation control, is the term used in quantum con-
trol and information technologies to describe manipulating the environment within which
an open quantum system operates. Reservoir engineering is essential in applications
where storing quantum information is required.
When a quantum system possesses a subsystem isolated from the detrimental influ-
ence of the environment and probing fields, the quantum information associated with
dynamics of such a system is preserved and can be used for quantum computation when
needed. In a sense, decoherence-free subsystems (DFS) can play the role of memory
elements in quantum information processing. This has motivated a significant interest in
the synthesis of quantum systems with the desired DFS structure.
On the other hand, when quantum error correction is in use, dissipation control
mechanisms derived using the control-theoretical formation of Lyapunov methods can
also be applied to stabilizes the code states. This forms a “passive” error correction loop
without the need for an active controller.
70 Chapter 4. Quantum Reservoir Engineering
4.1 Introduction
Reservoir engineering refers to the process of determining and implementing coupling
operators L = [L1; ...;Ln] for an open quantum system such that desired behavior is
achieved. Examples of common objectives include quantum computation by dissi-
pation [185], entanglement [97], state preparation [171], and protection of quantum
information [26, 137]. Typically open systems have some unavoidable couplings to
the environment, and such channels may lead to loss of energy and quantum coher-
ence. However, in many systems couplings can be engineered at the fabrication stage,
providing a resource for tuning the behavior of the system.
The scheme of quantum computation by dissipation [185] suggests that the com-
putation result could be encoded into the steady states of a carefully-designed system
Hamiltonian. For instance, measurement-based quantum computation [146] can be
realized provided that we can prepare a giant entangled state often called graph state [66]
or cluster state [121]. After the coding step, the system is coupled to an engineered
environment, and the dissipative dynamics would drive the system to the ground states
which accomplish the computation. This is in contrast to the conventional quantum
computation scheme [122] in which the computation is executed by applying a sequence
of unitary gates.
The dissipative dynamics has been well studied in classical control engineering.
The so-called Lyapunov method [63] plays a vital role in establishing the stability of a
dissipative system. To be specific, a Lyapunov function is proposed as a cost function
and the control is designed such that it decreases the value of the Lyapunov function.
Efforts have been made for the purpose of applying Lyapunov techniques to quantum
systems [112, 98, 188, 76, 169, 7]. These works employ a Schrödinger-picture analysis
which facilitates the engineering of the stability of quantum states [154].
Stabilizing ground states is critical for quantum state engineering and quantum
computation [185, 139]. In this chapter, we apply the Lyapunov methods to establish the
ground-state stability of an operator-sum representation whose ground state encodes the
quantum information. The Lyapunov analysis can thus be conveniently embedded into
the framework of quantum computation by dissipation control. In particular, we focus
on a specific decomposition of the dissipation control, based on which we can derive the
4.1 Introduction 71
conditions for the Lyapunov method to be scalable. Also we will show that the Lyapunov
formalism is closely related to the stabilizer formalism for quantum error correction, in
the sense that the dissipation control can be used for passive error correction in addition
to stabilization.
In Section 4.3, an extended scalability condition which is applicable to a wider
range of applications is proposed to achieve the ground-state stability for a class of
multipartite quantum systems which may involve two-body interactions, and an explicit
procedure to construct the dissipation control is presented. Moreover, we show that
dissipation control can be used for automatic error correction in addition to stabilization.
We demonstrate the stabilization and error correction of three-qubit repetition code states
using dissipation control.
Without active error correction, from the control theory perspective, a quantum sys-
tem is capable of storing quantum information if it possesses a so-called decoherence free
subsystem (DFS). Section 4.4 explores pole placement techniques to facilitate synthesis
of decoherence free subsystems via coherent quantum feedback control. We discuss
limitations of the conventional ‘open loop’ approach and propose a constructive feedback
design methodology for decoherence free subsystem engineering. It captures a quite
general dynamic coherent feedback structure which allows systems with decoherence
free modes to be synthesized from components which do not have such modes.
The problem of DFS synthesis has been found to be nontrivial - it has been shown
in [202] that conventional measurement feedback is ineffective in producing quantum
systems having a DFS, however certain coherent controllers can overcome this limitation
of the measurement-based feedback controllers. The objective of Section 4.4 is to put
this observation on a solid systematic footing, by developing a quite general constructive
coherent synthesis procedure for generating quantum systems with a DFS of desired
dimension.
Our particular interest is in a class of quantum linear systems [129] whose dynamics
in the Heisenberg picture are described by complex quantum stochastic differential
equations expressed in terms of annihilation operators only. Such systems are known to
be passive [83]. Passivity ensures that the system does not generate energy. In addition,
in such systems the notion of system controllability by noise and that of observability
72 Chapter 4. Quantum Reservoir Engineering
from the output field are known to be equivalent [59]. Also, one can readily identify
uncontrollable and unobservable subspaces of the passive system by analyzing the
system in the Heisenberg picture [202]. These additional features of annihilation only
passive systems facilitate the task of synthesizing decoherence free subsystems by means
of coherent feedback.
The focus on a general coherent feedback synthesis is the main distinct feature of our
work which differentiates it from other works of a similar kind, notably from [202, 126].
The paper [202] presents an analysis of quantum systems equipped with coherent
feedback for the purpose of characterizing decoherence free subsystems, quantum
nondemolished (QND) variables and measurements capable of evading backaction;
in [202] all these characteristics are expressed in geometric terms of (un)controllable
and (un)observable subspaces. In contrast, we propose constructive algebraic conditions
for the synthesis of coherent feedback to equip the system with a DFS. These conditions
are expressed in terms of linear matrix inequalities (LMIs) and reduce the DFS synthesis
problem to an algebraic pole assignment problem. We build our technique using the most
general type of dynamic linear passive coherent feedback. We show that the controller
structures from [202] are in fact special cases of our general setting. In addition, we
discuss the conventional open-loop approach to reservoir engineering and show the
shortcoming of such approach.
This chapter is based on the work presented in [137]1 and [120]2 where we derived
the condition for automatic quantum error correction by dissipation control (Theo-
rem 4.3.3) and the DFS synthesis condition (Theorem 4.4.2).
4.2 Background
4.2.1 Notations
Given an underlying Hilbert space H and an operator x : H→ H, x∗ denotes the operator
adjoint to x. In the case of a vector of operators, the vector consisting of the adjoint
components of x is denoted x#, and x† = (x#)T , where T denotes the transpose of a vector.
1the author contributed equally with the co-author2significant contributions in deriving the equations, providing examples demonstrating the synthesis
scheme
4.2 Background 73
Likewise, for a matrix A, A# is the matrix whose entries are complex conjugate of the
corresponding entries of A, and A† = (A#)T . [x,y] = xy− yx is the commutator of two
operators, and in the case where x,y are vectors of operators, [x,y†] = xy† − (y#xT )T .
A finite-level quantum system is defined on a Hilbert space H ≃ CN . Denote the
space of bounded operator on H as B(H ). A quantum state is characterized by a
density operator ρ ∈ B(H ) satisfying trace(ρ) = 1 and ρ ≥ 0. In many cases, the
interaction between the quantum system and the environment is described by a Markov
process, and the dynamical equation of the quantum state ρt can be written as
ρt = L (ρt)
= −i[H,ρt ]+J
∑j=1
L jρtL†j −
12
L jL†jρt −
12
ρtL†jL j. (4.1)
Here H ∈B(H ) is the system Hamiltonian and L j ∈B(H ), j = 1, · · ·,J are system
operators that characterize the system-environment couplings. ZV is the space spanned
by the ground states of V .
σz =
1 0
0 −1
,σx =
0 1
1 0
,σy =
0 −i
i 0
are Pauli operators acting
on a two-level system called qubit. Accordingly, σzi,σxi,σyi are the Pauli operators
defined on the i-th qubit. The vectorization of a matrix A is denoted as vec(A), which is
a column vector obtained by stacking the columns of the A on top of one another.
4.2.2 Linear Quantum Systems
Open quantum systems are systems that are coupled to an external environment or
reservoir [19]. The environment exerts an influence on the system, in the form of vectors
W (t), W †(t) consisting of quantum Wiener processes defined on a Hilbert space F
known as the Fock space. The unitary motion of the passive annihilation only system
governed by these processes is described by the stochastic differential equation
dU(t) =
((−iH− 1
2L†L)dt+dW †L−L†dW
)U(t), (4.2)
U(0) = I,
74 Chapter 4. Quantum Reservoir Engineering
where H and L are, respectively, the system Hamiltonian and the coupling operator
through which the system couples with the environment. Then, any operator X : H→ H
generates the evolution X(t) = jt(X) =U(t)∗(X ⊗ I)U(t) in the space of operators on
the tensor product Hilbert space H⊗F,
dX = G (X)dt +dW †[X ,L]+ [L†,X ]dW, (4.3)
where
G (X) = −i[X ,H]+LL(X),
LL(X) =12
L†[X ,L]+12[L†,X ]L
are the generator and the Lindblad superoperator of the system, respectively [198]. The
field resulting from the interaction between the system and the environment constitutes
the output field of the system
dY = Ldt +dW. (4.4)
Linear annihilation only systems arise as a particular class of open quantum systems
whose operators ak, k = 1, . . . ,n, describe various modes of photon annihilation resulting
from interactions between the environment and the system. Such operators satisfy the
canonical commutation relations [a j,a∗k ] = δ jk, where δ jk is the Kronecker delta. Taking
the system Hamiltonian and the coupling operator of the system to be, respectively,
quadratic and linear functions of the vector X = a = [a1, . . .an]T ,
H = a†Ma, L =Ca, (4.5)
where M is a Hermitian n×n matrix, and C ∈ Cm×n, the dynamics and output equations
become
da = Aadt +BdW
dy = Cadt +dW, (4.6)
4.2 Background 75
where the complex matrices A ∈ Cn×n, B ∈ Cn×m, and C ∈ Cm×n satisfy
A =−iM− 12
C†C, B =−C†. (4.7)
The following fundamental identity then holds [102, 103]
A+A† +C†C = 0. (4.8)
According to [83], passivity of a quantum system P is defined as a property of the
system with respect to an output generated by an exosystem W and applied to input
channels of the given quantum system on one hand, and a performance operator Z of the
system on the other hand. To particularize the definition of [83] in relation to the specific
class of annihilation only systems, consider a class of exosystems, i.e., open quantum
systems with zero Hamiltonian, an identity scattering matrix and a coupling operator
u which couples the exosystem with its input field. The exosystem is assumed to be
independent of P in the sense that u commutes with any operator from the C∗ operator
algebra generated by X and X†. The time evolution of u is however determined by the
full interacting system P◁W, and therefore may be influenced by X , X†.
If the output of the exosystem W is fed into the input of the system P in a cascade or
series connection, the resulting system P◁W has the Hamiltonian HP◁W = H+ Im(u†L),
the identity scattering matrix and the field coupling operator LP◁W = L+u [83]. The
resulting system (P◁W) then has the generator GP◁W.
Definition 4.2.1 — [83]. A system P with a performance output Z is passive if there exists
a nonnegative observable V (called the storage observable of P) such that
GP◁W(V )≤ Z†u+u†Z +λ (4.9)
for some constant λ > 0. The operator
r(W) = Z†u+u†Z
is the supply rate which ensures passivity.
76 Chapter 4. Quantum Reservoir Engineering
Now suppose P is a linear annihilation only system (4.5). Also, consider a perfor-
mance output for the system P◁W to be
Z =Cpa+Dpu.
Taking X = a in (4.3), the system P◁W can be written as
da = (Aa+Bu)dt +BdW, (4.10)
dY = (Ca+u)dt +dW,
Z = Cpa+Dpu.
where the complex matrices A ∈ Cn×n, B ∈ Cn×m, and C ∈ Cm×n are the coefficients of
the annihilation only system P.
Without loss of generality, we further take the storage observable V having the form
V = a†Pa, r(W) = Z†u+u†Z, then it can be shown that the system P is passive with a
storage function V and a supply rate r(W) if for some constant λ > 0,
a†(PA+A†P)a+u†BPa+a†PBu
≤ (Cpa+Dpu)†u+u†(Cpa+Dpu)+λ .
This condition is equivalent to the positive realness condition stated in Theorem 3
of [204] (letting Q = 0 in that theorem):
PA+A†P PB−C†
p
B†P−Cp −(Dp +D†p)
≤ 0. (4.11)
In the special case, where V = a†a, Dp = 0 [204] and Cp = −C, this reduces to the
condition
A+A† ≤ 0
as the condition for passivity. Clearly this condition is satisfied in the case of an
annihilation only system P in the light of the identity (4.8). Hence the annihilation only
4.2 Background 77
system (4.10) is passive with respect to performance output Z =−Ca, with the storage
function V = a†a.
4.2.3 Lyapunov Methods
The stationary states of Eq. (4.1) have been studied extensively for the purpose of state
stabilization [160, 171, 170, 154, 6]. For a multipartite quantum system, the method of
using dissipative dynamics to engineer quantum states has been generalized to the notion
of dissipatively quasi-locally stabilizable (DQLS) states [172, 173]. The theory of DQLS
states proposes a systematic approach to determine whether a given multipartite state is
asymptotically stabilizable if local dissipation controls can be engineered. Furthermore,
if the quantum state satisfies the DQLS condition, the required multipartite system
Hamiltonian and the system-environment coupling operators can be constructively
derived.
The aforementioned results are based on (4.1). Alternatively, the desired states can
be stabilized by studying the evolution of certain operators. Since the expectation of an
operator V ∈B(H ) at the state ρ is calculated by ⟨V ⟩ρ = trace(V ρ), the evolution of
the operator V (t) can be defined via the relation ⟨V (t)⟩ρ0= ⟨V ⟩ρt
. Note that V =V (0).
The generator of this Markov process is given by [83]
G (V (t)) = −i[V (t),H(t)]+L(V (t))
= −i[V (t),H(t)]+J
∑j=1
L†j(t)V (t)L j(t)
−12
L†j(t)L j(t)V (t)− 1
2V (t)L†
j(t)L j(t). (4.12)
A large class of quantum states, including graph states and cluster states (which are
DQLS states as well), can be encoded as the ground states of a multipartite operator
taking the form V = ∑Ki=1Vi [143, 185, 172]. As a result, state stabilization can be
achieved by engineering the dissipation such that the system converges to the ground
states asymptotically. The merit of this formulation is that the ground-state stability of
V can be established by Lyapunov-type operator inequalities. This scenario has been
considered before in [139], where a scalability condition is used to prove the ground-state
stability of V when each Vi is stabilized individually. However the scalability condition
78 Chapter 4. Quantum Reservoir Engineering
proposed in [139] does not hold for certain applications, especially when Vi consist of
two-body interactions. An illustrative example can be found in Section 4.3.
Note that ground-state stability does not necessarily guarantee that a particular
ground state is stable against errors, because the erroneous state may return to a different
ground state under the dissipative dynamics. However, we can prove in Section 4.3 that if
certain types of errors occur to one of the ground states, the dissipation control can steer
the erroneous state back to the initial ground state exactly, without any measurement or
active feedback. In this regard, this result can be considered as an addition to the existing
physical literature on automatic quantum error correction (AQEC) [14, 90, 26, 88, 82].
The results of Section 4.3 are intended to deal with a specialized case, where we
apply algebraic methods to achieve the stabilization of the states by imposing scalable
Lyapunov-type conditions on the operators. If these conditions are satisfied, then the
ground states of the system are DQLS and V is frustration-free[185, 173, 139].
The multipartite quantum system is defined on H =⊗M
m=1 Hm which is a tensor
product of Hilbert spaces Hm (each Hm is associated with a subsystem). Unless
otherwise noted, we will the following assumption throughout this chapter.
Assumption 4.2.1 V can be decomposed as V = ∑Ki=1Vi, and Vi is defined on a subset of
Hm. Vi are orthogonal projections, i.e. V 2i = Vi and [Vi,Vj] = 0, i = j. Each Vi is
associated with a set of dissipation controls L j. Each L j allows the decomposition
L j = ∑iUi, jVi with Ui, j being a unitary operator.
Remark 4.2.1 Vi can be regarded as quasi-local operators [172] since they are de-
fined on a subset of Hm. Therefore, the stabilizing dynamics in this chapter can be
considered as specific realizations for the stabilization of DQLS states.
V 2i = Vi ≥ 0 is a natural assumption that holds for many applications, e.g. the
dissipation control of stabilizer states [143, 185, 139]. Vi being commutative is an
intuitive assumption which enables Vi to share common ground states. Physical
examples of the decomposition includes the dissipation control of graph states [185].
We also make the following assumption.
Assumption 4.2.2 The system Hamiltonian can be written as H = ∑Ki=1 Hi, where each
Hi satisfies Hi =Vi −giI. Here −gi is the smallest eigenvalue of the Hermitian operator
Hi.
4.2 Background 79
Remark 4.2.2 It is experimentally possible to engineer Hamiltonian on a multipartite
quantum system, e.g. [93].
As shown in the next section, the two assumptions allow a concise and scalable
stability analysis based on the generator (4.17). In addition, we have two definitions as
follows.
Definition 4.2.2 Vi is said to be a two-body operator if it can be decomposed as Vi =
Xm1⊗Xm2
, with Xm1,Xm2
defined on two Hilbert spaces Hm1,Hm2
, respectively.
Definition 4.2.3 The generator of the evolution of Vi that is induced by a single dissipation
control Li is defined by
G (Vi)Li=−i[Vi,H]+L†
i ViLi −12
L†i LiVi −
12
V L†i Li. (4.13)
Remark 4.2.3 Eq. (4.13) is the generator of Vi controlled by a single coupling operator
Li. If Vi is also affected by other coupling operators L j, j = i, then we have G (Vi) =
G (Vi)Li+∑ j =i L†
jViL j − 12L†
jL jVi − 12ViL
†jL j.
We also recall one theorem from [83]:
Theorem 4.2.1 If an operator X ≥ 0 satisfies the following inequality
G (X)≤−cX , c > 0, (4.14)
then the system will asymptotically converge to ZX .
Remark 4.2.4 The algebraic condition (4.21) uses X = X(0), H = H(0) and L j =
L j(0). The satisfaction of this condition implies that limt→∞⟨X(t)⟩ = 0. The other
algebraic conditions of this chapter also use the operators at the initial time. For the
details of the Heisenberg-picture stability theory, please refer to [136].
The following theorem can be derived using Theorem 4.2.1, Assumption 4.2.1 and
4.2.2.
Theorem 4.2.2 [139] If the following condition
J
∑j=1
(ViU†i, jViUi, jVi −Vi)≤−ciVi, ci > 0, (4.15)
80 Chapter 4. Quantum Reservoir Engineering
• • •
Fig. 4.1 Each Vi has 0 as its lowest energy. The system may be stabilized to the groundstate of V1 under a local dissipation control. The local dissipation control is scalableif it maintains or decreases the energy of Vi, i = 1, i.e. the system is steered towardsthe ground states of Vi, i = 1 under the local dissipation control of V1. However forinstance, if the local dissipation control of V1 increases the energy of V2, then Li, i = 1must be able to compensate this increase in order to stabilize V2.
holds, then G (Vi)≤−ciVi and Vi is asymptotically ground-state stable under the dissipa-
tion control of L j. In particular, we say Ui, j stabilize Vi if Eq. (4.30) holds.
Now we recall the scalability condition derived in [139].
Theorem 4.2.3 Suppose for each Vi, there exists Li, j such that G (Vi)≤−ciVi,ci > 0
holds. The ground-state stability of V can be implied if the intuitive scalability condition
∑i′ =i
∑j
G (Vi)Li′ , j
≤ 0, (4.16)
can be established for all i.
Remark 4.2.5 V is asymptotically ground-state stable if the local dissipation controls
of Vi′ , i′ = i do not increase the expectation of Vi (Fig. 4.1). However, Eq. (4.23)
is easily violated if Vi involve two-body interactions. If Vi is a two-body operator,
the local control Li, j that stabilizes Vi may not act on Vi′ , i′ = i trivially. To be more
specific, if Vi = Xm1⊗Xm2
, then there exists at least one Vi′ , i′ = i which is defined on
Hm1, otherwise the resulting ground state cannot be entangled in Hm1
. As a result, the
local control Ui, j defined on Hm1⊗Hm2
may act nontrivially on this Vi′ .
4.3 Quantum Error Correction by Dissipation Control 81
The result of this chapter shows the way to construct the dissipation control for the
stabilization of V when Vi can be locally stabilized but the local dissipation controls
do not satisfy the strong scalability condition Eq. (4.23).
4.3 Quantum Error Correction by Dissipation Control
A quantum Lyapunov operator V is an observable on a Hilbert space H for which the
following properties hold [136]:
1. V ≥ 0.
2. G (V )≤ 0.
V ≥ 0 means V is positive semidefinite with its smallest eigenvalue being 0. By the
definition, V could be understood as a virtual energy operator. We define the virtual
energy as the expectation tr(V ρt) = ⟨V ⟩ρt= ⟨V (t)⟩ρ0
= tr(V (t)ρ0). ρ0 is the initial state
of the evolution. The second property G (V )≤ 0 suggests that the virtual energy will not
increase in time, i.e. the virtual energy is stable. However, V being stable is not enough
in the context of this Section. In order to stabilize the system to the ground states of V ,
we need to engineer the dissipation control such that ⟨V (t)⟩ρ0exponentially converges
to zero.
In the above definition, we have used the superoperator notation G (·) to represent
the generator of the Heisenberg-picture evolution
G (X(t)) =−i[X(t),H(t)]+L(X(t)) (4.17)
for a Heisenberg-picture operator X(t). H is the system Hamiltonian and
L(X(t)) =K
∑k=1
L†k(t)X(t)Lk(t)
− 12
L†k(t)Lk(t)X(t)− 1
2X(t)L†
k(t)Lk(t) (4.18)
82 Chapter 4. Quantum Reservoir Engineering
is the Lindblad dissipation term. Lk,k = 1, · · ·,K are K coupling operators character-
izing the interaction between the system and the environment. Also we define
G (X(t))Lk=−i[X(t),H(t)]+L†
k(t)X(t)Lk(t)
−12
L†k(t)Lk(t)X(t)− 1
2X(t)L†
k(t)Lk(t) (4.19)
which represents a generator with regard to a specific coupling operator Lk. We use
ZX := ρ : ⟨X⟩ρ = 0 to denote the subspace spanned by the ground states of X . Then,
V is said to be asymptotically ground-state stable if
⟨V (t)⟩ρ0= ⟨V ⟩ρt
→ 0, as t → ∞. (4.20)
A sufficient condition for the asymptotic ground-state stability of V is given by the
following operator inequality [83]
G (V )≤−cV, c > 0. (4.21)
Integration of the inequality (4.21) yields ⟨V (t)⟩ρ0≤ e−ctV (0).
The inequality (4.21) can be interpreted using the ground-state spaces of the operators.
If ZA is contained in ZB for two operators A,B ≥ 0, then there must exist a constant c > 0
such that A ≥ cB. This can be seen from the decomposition of A and B into
A =
A+ 0
0 0
, B =
B 0
0 0
, (4.22)
with A+ being positive definite and B ≥ 0. Then A+− ε1I > 0 for some ε1 > 0 and
I ≥ ε2B for some ε2, which yields a solution c = ε1ε2 to A ≥ cB. As a result, the
inequality (4.21) can be established if Z−G (V ) is contained in ZV . This property will be
useful in the next section.
4.3.1 Scalability of Lyapunov Method
For multipartite quantum systems, directly solving (4.21) for Lk is computationally
hard, not to mention the physical realizability issues with the resulting multipartite
4.3 Quantum Error Correction by Dissipation Control 83
coupling operators. Indeed, the local dissipation controls designed for the subsystems
can effectively control the overall system, as long as they satisfy certain scalability
conditions. This idea was firstly conjectured in [185]. As a matter of fact, this idea is
similar to the decomposition-aggregation technique [16, 207] for the classical control
engineering of large-scale systems. We can decompose the total Lyapunov operator into
a sum of component Lyapunov operators and design environmental couplings for each
of them accordingly. If the total Lyapunov operator is asymptotically ground-state stable
under the collective action of the individually-designed couplings, we say the Lyapunov
method is scalable (Figure 4.1).
Consider a finite number of Lyapunov operators Vi, i= 1, · ··,N and the stabilization
of the system to the ground states of V = ∑iVi. We call Vi the component Lyapunov
operator. We omit the range of the summation where no confusion is raised. Suppose
for each Vi, there exists a set of local coupling operators Lik such that ∑k G (Vi)Lik≤
−ciVi, ci > 0 holds. Obviously, the ground-state stability of V = ∑iVi can be implied if
∑k
G (Vi)L jk≤ 0, j = i, (4.23)
holds for all i, j. That is, V is asymptotically ground-state stable if the dissipation controls
do not increase the energy of the other component Lyapunov operators. Eq. (4.23) is a
strong scalability condition. Generally speaking, it is possible that some of the individual
dissipation controls indeed increases the energy of the other component Lyapunov
operators but V is still asymptotically ground-state stable.
Example 4.1 (3-qubit repetition code) A logical qubit (|0L⟩ , |1L⟩) is encoded using three
physical qubits, i.e.
|0L⟩= |000⟩ , |1L⟩= |111⟩ . (4.24)
This code can be used to protect quantum information in the subspace spanned by
|0L⟩ , |1L⟩ against single bit-flip noise, with the error operators given by E = σ(i)x , i =
1,2,3. σ(i)x denotes the Pauli operator acting on qubit i, resulting in a flip of the qubit
state.
In the stabilizer formalism [53], the codewords (4.63) are characterized by stabilizers.
In particular, codewords are invariant under the action of the stabilizers. For the 3-qubit
84 Chapter 4. Quantum Reservoir Engineering
repetition code, the group of stabilizers are given by
S = σ(1)z σ
(2)z ,σ
(2)z σ
(3)z ,σ
(1)z σ
(3)z . (4.25)
Lyapunov operator can be defined from the stabilizers as
V =12[(I −σ
(1)z σ
(2)z )+(I −σ
(2)z σ
(3)z )+(I −σ
(1)z σ
(3)z )]
= V1 +V2 +V3, (4.26)
which is a sum of component Lyapunov operators V1,V2,V3 ≥ 0. For V1 = 12(I −
σ(1)z σ
(2)z ), we can easily see that L1 =
12σ
(1)x (I−σ
(1)z σ
(2)z ) is a solution to the inequality
(4.21) due to
G (V1)L1= L†
1V1L1 −12
L†1L1V1 −
12
V1L†1L1 =−V1. (4.27)
However, L1 is not scalable. This is clear by calculating
∑j =1
G (Vj)L1= diag(0,0,1,−1,−1,1,0,0)≰ 0, (4.28)
which violates the scalability condition (4.23). Without loss of generality, we consider one coupling operator (K = 1) for each Vi. We
assume Vi are commuting and satisfying V 2i =Vi, i.e. Vi are projectors. Motivated
by [185], we use the coupling operators of the form Li = UiVi , where Ui is a unitary
operator. Also, we assume the Vi is defined by a displacement of the system Hamiltonian
Hi by Vi = Hi + gi. By this way we can ignore the unitary dynamics. Substituting
Li =UiVi into Eq. (4.17) yields
G (Vi) =ViU†i ViUiVi −Vi. (4.29)
As a result, the unitary operations Ui must satisfy
ViU†i ViUiVi ≤ (1− ci)Vi, ci > 0, (4.30)
4.3 Quantum Error Correction by Dissipation Control 85
1V1U U
6 H11 2
H1
1 2U
H1
3
H1
V V2V 3V
U 3U
Fig. 4.2 The schematic representation of three qubits governed by V1,V2,V3. Each Vi
acts non-trivially on two of the subsystems. Ui = σ(i)x is the unitary rotation on the i-th
subsystem.
in order to establish G (Vi)≤−ciVi. Fig. 4.2 depicts the systems and dissipation controls
as proposed in Example 4.1. The Li =UiVi with Ui satisfying Eq. (4.30) fails to satisfy
the general scalability condition Eq. (4.23).
Therefore, in order to derive dissipation controls that can deal with Example 4.1,
we will need weaker scalability condition. Still, we may suppose there exists a unitary
rotation Ui which stabilizes each component Lyapunov operator Vi by Eq. (4.30). Denote
V (i)n as the set of component Lyapunov operators that commute with Ui. The other
component Lyapunov operators are denoted as V (i)d ,d = i. The sufficient condition for
the asymptotic ground-state stability of V is
V (i)d U†
i V (i)d UiV
(i)d ≤V (i)
d , (4.31)
∑i(Vi ∏
dV (i)
d )≥ λ ∑i
Vi, λ > 0. (4.32)
The resulting dissipation controls are Li =UiVi ∏d V (i)d . Under these controls, ⟨V (t)⟩ρ0
will converge to 0 as t → ∞. For proof of the sufficient condition please refer to
Appendix A.2.
86 Chapter 4. Quantum Reservoir Engineering
The dissipation control can be written as Li = UiV′i = UiVi ∏d V (i)
d . Obviously,
V′i can be regarded as the updated Lyapunov operators with the unitary rotations Ui.
As shown before, the condition (4.45) is satisfied if Z∑i V
′i
is contained in Z∑i Vi
. That
is, the common ground states of the updated Lyapunov operators are also the common
ground states of the original Vi. Particularly, if the common ground state of Vi is
unique, then V′i must share the same unique ground state.
Now we can solve for the correct dissipation control for Example 4.1 using the
weaker scalability condition (4.44)-(4.45). Obviously, the updated Lyapunov operators
V′1,V
′2,V
′3 are V ′
1 = V1V3, V ′2 = V1V2, V ′
3 = V2V3. The dissipation control for the first
Lyapunov operator is given by L1 =U1V ′1 =
12σ
(1)x (I −σ
(1)z σ
(3)z )(I −σ
(1)z σ
(2)z ). We can
derive L2 and L3 similarly. V is asymptotically ground-state stable when the three-qubit
system is coupled to a dissipative environment via L1,L2,L3.
4.3.2 Error Correction Condition
The ground-state stability concerns with driving system to the ground state of V from any
initial state. However, ground-state stability is not necessarily related to error correction.
For instance, suppose we encode the information into a steady ground state of V . If
an error Ea occurs, the perturbed state can be stabilized to ZV . However, we cannot
guarantee that the dissipation control will restore the system to the initial state since the
ground states may be degenerate. In addition, it is not clear whether the coherence of the
initial state will be preserved when subjected to the dissipation.
In this section, we will show that the controls Li =UiVi can passively correct the
errors induced by U†i if Vi satisfy a simple stabilizer-type condition.
Denote |p⟩ as basis vectors of ZV . Suppose the Kraus operators [96] describing the
error process are given by Ea. We consider a two-step passive error correction. In the
first step, the system is coupled to the noise. In the second step, the system is subjected
to the dissipation controls. The controls will return the system to the initial state if any
error occurs in the first step. If the dissipation strength is strong, it is reasonable to ignore
the effect of the noise in the second step.
4.3 Quantum Error Correction by Dissipation Control 87
The state dynamics is governed by the master equation
ρt = L (ρt), (4.33)
where L (·) is in Lindblad form
L (ρt) =−i[H,ρt ]+∑k
LkρtL†k −
12
ρtL†kLk −
12
L†kLkρt . (4.34)
An arbitrary pure state in ZV is expressed by ρ0 =∑p αp|p⟩(∑p αp|p⟩)† =∑p,q αpα∗q |p⟩⟨q|.
The erroneous states are thus given by Eaρ0E†a. The dissipative dynamics stabilizes
the erroneous states back to the initial ones if
L (|p⟩⟨q|) = 0, (4.35)
L (Ea|p⟩⟨q|E†a)
=−κpqEa|p⟩⟨q|E†a +κpq|p⟩⟨q|, κpq > 0 (4.36)
holds for all p,q. The proof of the condition is put in Appendix A.3. Eq. (4.52)-(4.53)
basically says that each of the elements of the erroneous density operator including
coherence terms can be restored to the initial value. Please note this implicitly guarantees
that the orthogonal states evolve into orthogonal erroneous states, which is a sufficient
condition for error correction [95, 14, 82].
With the controls of the form Li =UiVi, it is easy to verify that Eq. (4.52) is satisfied
since Vi|p⟩= 0 for all i, p. In addition, Eq. (4.53) can be explicitly written as
L (Ea|p⟩⟨q|E†a)
= ∑i
LiEa|p⟩⟨q|E†a L†
i
− 12
Ea|p⟩⟨q|E†a L†
i Li −12
L†i LiEa|p⟩⟨q|E†
a
= ∑i
UiViEa|p⟩⟨q|E†aViU
†i
− 12
Ea|p⟩⟨q|E†aVi −
12
ViEa|p⟩⟨q|E†a , (4.37)
88 Chapter 4. Quantum Reservoir Engineering
if the erroneous states Ea|p⟩ are eigenstates of the Hamiltonian H. Recall that Vi is
obtained by displacing the system Hamiltonian and the nonzero eigenvalue of Vi is 1.
Hence, the requirement that Ea|p⟩ are eigenstates of H is equivalent to Ea|p⟩ are
eigenstates of Vi. Obviously, a sufficient condition for (4.62) and (4.53) to be equal is
Ea =U†i , (4.38)
ViU†i |p⟩= 1, (4.39)
VjU†i |p⟩= 0, j = i, (4.40)
for all p. Eq. (4.59)-(4.60) imply a one-to-one correspondence between each error
operator Ea =U†i to a Lyapunov operator Vi. The erroneous states Ea|p⟩=U†
i |p⟩ are
eigenstates of a unique Vi with the positive eigenvalue 1, while remaining as the ground
states of Vj for j = i. Therefore, the component Lyapunov operators are just like the
syndrome projectors proposed in the stabilizer formalism [53].
Consider Example 4.1. We can check the condition Eq. (4.59)-(4.60) by
V′i σ
(i)x |0L⟩ = V
′i σ
(i)x |1L⟩= 1,
V′jσ
(i)x |0L⟩ = V
′jσ
(i)x |1L⟩= 0, j = i. (4.41)
As a result, the dissipation controls Li proposed at the end of Section 4.3.1 can
passively correct the errors defined by the operators (σ (i)x )†= σ
(i)x , which is exactly
the single bit-flip error E .
4.3.3 Definition of AQEC
Next we will introduce the definition of AQEC [14] that is adapted to the context of this
Section.
Definition 4.3.1 The set of error operators is denoted as Ea [96]. In the error process,
an error may occur with a probability. When the error occurs, the corresponding error
operator Ea transforms the initial state ρ0 ∈ ZV to the erroneous state as Eaρ0E†a . In the
correction process that follows the error process, AQEC is defined as the dissipative
dynamics that stabilizes the system to an arbitrary initial state ρ0 ∈ ZV , when an error
4.3 Quantum Error Correction by Dissipation Control 89
modelled by an arbitrary error operator from Ea occurs, i.e.,
limt→∞
(Eaρ0E†a)t = ρ0, (4.42)
holds for arbitrary ρ0 ∈ ZV and Ea. In this case, the errors are said to be correctable by
the dissipation control.
For example, Ea can be the bit-flip error operator σx which may flip the state of the
qubit as σx|0⟩= |1⟩ with certain probability. AQEC will automatically correct the error
and return the system state to |0⟩ by dissipation control.
Note that the AQEC condition (4.42) implies that the system is ground-state stable
as every ρ0 ∈ ZV is an invariant state. As a result, if no error occurs, the ground states
will be stable under the dissipation control. However, ground-state stability does not
necessarily imply error correction capability. For instance, suppose the erroneous state is
Eaρ0E†a . The erroneous state can be automatically steered back to the invariant subspace
ZV if the system is asymptotically ground-state stable, but we cannot guarantee that
the dissipation control will restore the system to the initial state ρ0 if the ground states
are degenerate, i.e. ZV is more than one-dimensional. It is also worth mentioning that
Definition 4.3.1 corresponds to an ideal case that the error process and correction process
take place consecutively.
The following theorem is concerned with the existence of dissipation control.
Theorem 4.3.1 There always exists a set of unitary operators Ui, j, j = 1, · · ·,J that
stabilize Vi.
Proof. We provide a constructive method to prove the existence. Since Vi ≥ 0, we can
write the spectral decomposition of Vi as Vi = ∑n hn|n⟩⟨n|,hn ≥ 0, with |n⟩ being the
basis vectors of the Hilbert space. Denote one of the ground states (with eigenvalue 0)
as |0⟩⟨0|. Therefore, if for each h j > 0 we choose Ui, j = | j⟩⟨0|+ |0⟩⟨ j|+∑n= j,0 |n⟩⟨n|,then ViU
†i, jViUi, jVi = ∑n= j h3
n|n⟩⟨n|. Since hn equals either 0 or 1 for a projector Vi, we
have ∑n= j h3n|n⟩⟨n|= ∑n= j hn|n⟩⟨n|. We can verify that
J
∑j=1
(ViU†i, jViUi, jVi −Vi) =−
J
∑j=1
h j| j⟩⟨ j|=−Vi, (4.43)
90 Chapter 4. Quantum Reservoir Engineering
where h j = 1, j = 1, · · ·,J is the set of positive eigenvalues. The condition (4.30) is
satisfied with ci = 1.
Remark 4.3.1 Theorem 4.3.1 provides a specific solution to the exponential stabilization
problem. Approaches to choose the set of unitaries have also been discussed in [185].
4.3.4 Scalability of Dissipation Control
Theorem 4.3.2 Suppose Ui stabilizes Vi. For each Ui, Vj are separated into two sets,
namely, V (i)n and V (i)
d . The definitions of the two sets are as follows
• [V (i)n ,Ui] = 0,
• [V (i)d ,Ui] = 0.
A sufficient condition for the asymptotic ground-state stability of V is given by
V (i)d U†
i V (i)d UiV
(i)d ≤V (i)
d , (4.44)
∑i(∏
dV (i)
d )≥ λV, λ > 0. (4.45)
The dissipation control is constructed as Li =Ui ∏d V (i)d .
Remark 4.3.2 By definition, V (i)d is the operator that satisfies the condition (4.44) but
does not commute with Ui. Ui acts trivially on V (i)n . In contrast, Ui acts nontrivially on
V (i)d . However, in order for the dissipation controls to be scalable V (i)
d should satisfy
the stability condition G (V (i)d )Ui
=V (i)d U†
i V (i)d UiV
(i)d −V (i)
d ≤ 0. Note that Vi ∈V (i)d since
Ui stabilizes Vi by assumption.
Proof. By assumption of Theorem 4.3.2, we have Ui satisfying Eq. (4.30) for each Vi.
Using Vi ∈V (i)d and the dissipation control Li =Ui ∏d V (i)
d we have
G (Vi)Li= ∏
dV (i)
d U†i ViUi ∏
dV (i)
d −Vi ∏d
V (i)d
= ∏d
V (i)d ViU
†i ViUiVi ∏
dV (i)
d −∏d
V (i)d
≤ (1− ci −1)∏d
V (i)d =−ci ∏
dV (i)
d , ci > 0,
(4.46)
4.3 Quantum Error Correction by Dissipation Control 91
where we have made use of the properties [Vi,Vj] = 0 and V 2i =Vi from Assumption 4.2.1.
For j = i, either [U j,Vi] = 0 which results in
G (Vi)L j= ∏
dV ( j)
d U†j ViU j ∏
dV ( j)
d −Vi ∏d
V ( j)d = 0, (4.47)
or Vi does not commute with U j and satisfies Vi ∈ V ( j)d . According to (4.44), this
implies
ViU†j ViU jVi ≤Vi. (4.48)
Consequently, we can obtain
G (Vi)L j
= ∏d
V ( j)d U†
j ViU j ∏d
V ( j)d −Vi ∏
dV ( j)
d
= ∏d
V ( j)d ViU
†j ViU jVi ∏
dV ( j)
d −Vi ∏d
V ( j)d
≤ 0, (4.49)
which further leads to
G (∑i
Vi) ≤ −∑i
ci ∏d
V (i)d ≤−cmin ∑
i∏
dV (i)
d
≤ −cminλ ∑i
Vi. (4.50)
cmin is the smallest positive number in ci. Here we have used the condition (4.45). By
Theorem 4.2.1, V = ∑iVi is asymptotically ground-state stable.
Remark 4.3.3 The strong scalability condition (4.23) is just a special case of Theo-
rem 4.3.2 with V (i)d =Ui. Furthermore, Eq. (4.44) implies
G (V (i)d )
L′i=UiV
(i)d
≤ 0. (4.51)
Therefore, ∏d V (i)d is the product of the operators in Vi whose energies are not increased
by the local controls L′i =UiV
(i)d . However, if we implement the dissipation control as
L′i =UiVi, the dissipation control may not satisfy the scalability condition Eq. (4.23).
92 Chapter 4. Quantum Reservoir Engineering
The updated dissipation control Li =Ui ∏d V (i)d is still in the form of Li =UiV
′i
with V′i = ∏d V (i)
d . Note that V′i are still projectors. The reason we preclude the
operators V (i)n from the updated dissipation controls is that ∏d V (i)
d ∏nV (i)n = ∏ j Vj
leads to the same V′i for each dissipation control and so often results in the violation of
sufficient condition (4.45) in practical implementation.
4.3.5 Automatic Quantum Error Correction by Dissipation Control
In this section, we derive the condition such that the dissipation control Li can
automatically correct certain types of errors.
Denote |p⟩ ∈ H as the complete basis vectors of ZV . Then we can write the basis
of the bounded operators on ZV as |p⟩⟨q|, where p and q have the same range. An
arbitrary density state in ZV can thus be expanded on the basis as ρ0 = ∑p,q αpq|p⟩⟨q|.Lemma 4.3.1 The dissipation controls Li are error-correcting with respect to the error
operators Ea if
L (|p⟩⟨q|) = 0, (4.52)
L (Ea|p⟩⟨q|E†a)
=−κpqEa|p⟩⟨q|E†a +κpq|p⟩⟨q|, κpq > 0 (4.53)
hold for all p,q.
Proof. The dynamical equation Eq. (4.1) can be written as a linear system equation
˙vec(ρt) = Avec(ρt), (4.54)
using the vectorization of ρt ,H,Li and the relation vec(B1B2B3) = (BT3 ⊗B1)vec(B2).
As a result, A is determined by H and Li. The solution of the above equation is given
by
vec(ρt) = eAtvec(ρ0). (4.55)
The condition (4.52) ensures the invariance of the initial state under the dissipation
control, if no error occurs. If an error Ea occurs, the erroneous state ∑p,q αpqEa|p⟩⟨q|E†a
4.3 Quantum Error Correction by Dissipation Control 93
needs to be steered back to ρ0. Since we have
vec((Ea|p⟩⟨q|E†a)t) = eAtvec(Ea|p⟩⟨q|E†
a), (4.56)
and so
˙vec((Ea|p⟩⟨q|E†a)t)
= eAtAvec(Ea|p⟩⟨q|E†a)
= −κpqeAtvec(Ea|p⟩⟨q|E†a)+κpqvec(|p⟩⟨q|)
= −κpqvec((Ea|p⟩⟨q|E†a)t)+κpqvec(|p⟩⟨q|). (4.57)
Eq. (A.9) is an ordinary first-order differential equation which can be easily integrated
to be
vec((Ea|p⟩⟨q|E†a)t)
= e−κpqtvec(Ea|p⟩⟨q|E†a)
+ κpqvec(|p⟩⟨q|)∫ t
0e−κpq(t−r)dr
= e−κpqtvec(Ea|p⟩⟨q|E†a)+vec(|p⟩⟨q|)[1− e−κpqt ].
(4.58)
This proves (Ea|p⟩⟨q|E†a)t → |p⟩⟨q| as t → ∞. Due to the linearity of the dynamical
equation Eq. (4.54), we can deduce that (∑p,q αpqEa|p⟩⟨q|E†a)t → ρ0 for any error
operator Ea. The system is restored to the initial state exactly.
Eq. (4.53) guarantees that every element of the erroneous density state, including
the non-diagonal terms which characterize the quantum coherence, can be restored
to the initial value. In this sense, Eq. (4.52)-(4.53) are more like the definition of
quantum error correction, as compared to the sufficient conditions for the errors to be
correctable [95, 82, 14].
Based on Lemma 4.3.1, we can prove the following theorem.
Theorem 4.3.3 Suppose the dissipation control takes the form Li = UiVi, with Vibeing projectors and Ui being unitary operators. The sufficient conditions for the set
94 Chapter 4. Quantum Reservoir Engineering
of error operators U†i to be correctable are
ViU†i |p⟩=U†
i |p⟩, (4.59)
VjU†i |p⟩= 0, j = i, (4.60)
for all |p⟩.Remark 4.3.4 The correctable errors are determined by the available dissipation controls.
For example, U†i can be the bit-flip error operator σx. Note that there exist local
dissipation controls Li = σxiVi for the stabilization of the graph states [185, 139].
Therefore, in this case the bit-flip errors caused by σ†xi= σxi are correctable if the
sufficient conditions are satisfied.
Proof. It is easy to verify that Eq. (4.52) is satisfied since ZV is an invariant subspace
under the dissipation control and |p⟩ is the basis vector of ZV . Recall that Vi is obtained
by displacing the system Hamiltonian Hi according to Assumption 4.2.2. Since the
erroneous state vector U†i |p⟩ is an eigenstate of Vi with the eigenvalue 1 and an eigenstate
of Vj, j = i with eigenvalue 0, we can conclude that U†i |p⟩ is also an eigenstate of
the system Hamiltonian H. Based on this fact, we can remove the unitary dynamics in
Eq. (4.53) due to
−i[H,U†i |p⟩⟨q|Ui] =−i[V,U†
i |p⟩⟨q|Ui]
= −i[Vi,U†i |p⟩⟨q|Ui] = 0, (4.61)
4.3 Quantum Error Correction by Dissipation Control 95
and obtain
L (U†i |p⟩⟨q|Ui)
=K
∑j=1
L jU†i |p⟩⟨q|UiL
†j
− 12
U†i |p⟩⟨q|UiL
†jL j −
12
L†jL jU
†i |p⟩⟨q|U†
i
=K
∑j=1
U jVjU†i |p⟩⟨q|UiVjU
†j
− 12
U†i |p⟩⟨q|UiVj −
12
VjU†i |p⟩⟨q|Ui
= |p⟩⟨q|−U†i |p⟩⟨q|Ui. (4.62)
Therefore, Eq. (4.53) holds with κpq = 1. By Lemma 4.3.1, the dissipation control
Li =UiVi can automatically correct errors induced by the error operators U†i .
Eq. (4.59)-(4.60) imply a one-to-one correspondence between each error operator
U†i to Vi. For each error operator U†
i , the erroneous state vector U†i |p⟩ is an eigenvector
of Vi with the positive eigenvalue 1. At the same time, U†i |p⟩ remains as the ground
state of Vj for j = i. For this reason, Vi are similar to the error syndrome projectors as
proposed in the stabilizer formalism [53].
4.3.6 Dissipation Control of 3-qubit Repetition Code States
Let us re-consider the three-qubit code introduced in Example 4.1, i.e.
|0L⟩= |000⟩, |1L⟩= |111⟩, (4.63)
where |0⟩, |1⟩ denote the basis vectors of the physical qubit and |000⟩= |0⟩⊗ |0⟩⊗ |0⟩.For this 3-qubit repetition code, the set of stabilizers is given by
S = σz1σz2,σz2σz3,σz1σz3. (4.64)
96 Chapter 4. Quantum Reservoir Engineering
Based on the stabilizers, we can define V as
V =12[(I −σz1σz2)+(I −σz2σz3)+(I −σz1σz3)]
= V1 +V2 +V3. (4.65)
Note that V1,V2,V3 are two-body operators. |0L⟩ and |1L⟩ are two ground states of V . It
is easy to verify that L1 =12σx1(I −σz1σz2) satisfies
G (V1)L1= L†
1V1L1 −12
L†1L1V1 −
12
V1L†1L1 =−V1, (4.66)
with U1 = σx1. Similarly, we can obtain the local dissipation controls for V2,V3 as
L2 =12σx2(I −σz2σz3),L3 =
12σx3(I −σz1σz3), respectively. However, we have
∑j=2,3
G (V1)L j= σz1σz2V2 ≰ 0, (4.67)
which violates the scalability condition (4.23).
Instead, we can construct the correct dissipation control for V using Theorem 4.3.2.
Since U1 stabilizes both V1 and V3, the updated dissipation control would be designed
as L1 =U1V ′1 =
12σx1(I −σz1σz3)(I −σz1σz2) with V ′
1 =V1V3. L2 and L3 can be derived
similarly with V′2 =V1V2, U2 =σx2, V
′3 =V2V3, U3 =σx3. Furthermore, we have V
′1+V
′2+
V′3 =V1V3+V1V2+V2V3 =
12(V1+V2+V3) =
12V . By Theorem 4.3.2, V is asymptotically
ground-state stable if the 3-qubit system is coupled to a dissipative environment via
L1,L2,L3.
Next, we can verify the error correction condition Eq. (4.59)-(4.60) in Theorem 4.3.3
by
V′i σxi|0L⟩= σxi|0L⟩, V
′i σxi|1L⟩= σxi|1L⟩,
V′jσxi|0L⟩= 0, V
′jσxi|1L⟩= 0, j = i. (4.68)
As a result, the dissipation control Li can automatically correct the errors induced
by the set of error operators U†i = (σxi)
† = σxi, which is exactly the set of the
single bit-flip errors E . Fig. 4.3 is the numerical demonstration of the error correction
4.3 Quantum Error Correction by Dissipation Control 97
0.0 0.1 0.2 0.3 0.4 0.5
Time (γ−1)
0.0
0.2
0.4
0.6
0.8
1.0
Projection
Coherent evolution under Lyapunov-designed dissipation
Tr(ρtρtarget)
Tr(ρtρinitial)
Fig. 4.3 Numerical simulation of the state evolution with the coupling operators L′1,L
′2,L
′3
for 3-qubit repetition code state. Here L′1,L
′2,L
′3 are renormalized by a coupling strength√
κ , i.e. L′i =
√κLi. The initial state is an error-corrupted state ρinitial =
|001⟩+i|110⟩√2
,
which is different from the code state ρ0 = ρtarget =|000⟩+i|111⟩√
2due to the single bit-flip
error σx3. The system converges to the target state with unit probability, which meansthat no coherence is lost during this dissipative process. κ and the time scale γ are set to1 in this example.
performance of the dissipation control. An erroneous state is shown to be restored to
the code state. In particular, the coherence of the initial state is preserved under the
dissipation.
A more realistic automatic error correction scheme is applying the error correction
control in parallel with the noises [82]. In this case, the system can be modelled
as subjected to the noise and the engineered couplings simultaneously. As we have
demonstrated in Fig. 4.4, the derived system-environment couplings for the 3-qubit
repetition code states indeed can be used in parallel with the noises. The bit-flip errors
are modelled by coupling the system to the environment via the additional coupling
operators Lnoisei =
√γσxi, i = 1,2,3, with γ being the coupling strength. As a result,
the system is associated with 6 coupling channels in total. When we increase the strength
of the dissipation controls, a nearly perfect code state preservation can be achieved.
98 Chapter 4. Quantum Reservoir Engineering
Remark 4.3.5 Coherent feedback loop in [90] is essentially implementing the environ-
mental couplings adiabatically [90, Eq. (2)]3. Physical implementation of dissipation
control has also been experimentally demonstrated for superconducting qubit systems,
such as in [156]. Recently, Cohen et al. has demonstrated a scheme which uses the
dissipative gadgets to implement automatic quantum error correction [26].
0.0 0.1 0.2 0.3 0.4 0.5
Time (γ−1)
0.75
0.80
0.85
0.90
0.95
1.00
Tr(ρtρ
0)
Ground state fidelity under errors and engineered dissipations
κ
γ= 1
κ
γ= 5
κ
γ= 10
κ
γ= 50
κ
γ= 100
Fig. 4.4 Numerical simulation of state evolution when the system is subjected to thedissipative couplings L
′1,L
′2,L
′3 and the noise operators Lnoise
i = σxi, i = 1,2,3 simul-taneously. γ = 1. The code state is defined by ρ0 = |000⟩+i|111⟩√
2. Depending on the
coupling strength κ , we achieve different levels of state preservation. In the strongcoupling regime κ/γ ≥ 50, the code state can be continuously preserved against thebit-flip noise.
4.4 Decoherence Free Subsystem Synthesis
As mentioned, a decoherence free subsystem represents a subsystem whose variables
are not affected by input fields and do not appear in the system output fields; this makes
the DFS isolated from the environment and inaccessible to measurement devices, thus
preserving the quantum information carried by the variables of the DFS. In relation to the
annihilation only system (4.6), with a = [a1, . . . ,an]T , a component a j is a decoherence-
free mode if the evolution of a j is independent of the input W and if the system output
3These couplings are obtained if we adiabatically eliminate the coherent controller dynamics.
4.4 Decoherence Free Subsystem Synthesis 99
Y is independent of a j. The collection of decoherence-free modes forms a subspace,
called the decoherence-free subspace4. According to this, decoherence free subspaces
are uncontrollable and unobservable subspaces of a quantum system.
An important fact about the existence of a decoherence-free subsystem for linear
annihilation only systems follows from the results established in [59]:
Proposition 4.4.1 The linear annihilation only system (4.6) has a decoherence-free sub-
system if and only if the matrix A has some of its poles on the imaginary axis, with the
remaining poles residing in the open left half-plane of the complex plane.
Proof. According to [59, Lemma 2], for the system (4.6), the properties of controllability,
observability and Hurwitz stability are equivalent. The statement of the proposition then
follows by contraposition, after noting that being passive, the system (4.6) cannot have
eigenvalues in the open right hand-side of the complex plane due to (4.8).
4.4.1 Open-loop Reservoir Engineering for DFS Generation
With reference to Fig. 4.5, we investigate conditions to enable the synthesis of a quantum
coherent controller-system network to generate a DFS in the interconnected system
through interactions between the principal quantum system and the controller.
The quantum linear passive system in Fig. 4.5 is the system of the form (4.6), and its
input fields are further partitioned as W = [wT ,uT , f T ]T . Here, w represents a ‘natural’
environment for the system, and f and u represent an open-loop and feedback engineered
fields, respectively. According to this partitioning, the system evolution is described as
dap = Apapdt +B1dw+B2du+B3d f , (4.69a)
dy =Capdt +dw. (4.69b)
Accordingly, the matrices of the system have dimensions as follows: Ap ∈ Cn×n,
B1 ∈ Cn×nw , B2 ∈ Cn×nu B3 ∈ Cn×n f , and C ∈ Cnw×n (n,nw,nu,n f ∈ N). We also use
the notation ap for the vector ap (t) =[ap1
(t) , . . .apn(t)]T of the system annihilation
operators defined on its underlying Hilbert space Hp.
4A decoherence-free subspace exists if we can partition the Hilbert space into two subspaces HS =HDF ⊕HN , where the subspace of the Hilbert space not affected by decoherence is denoted by HDF andHN denotes the decoherence-affected, “noisy” subspace.
100 Chapter 4. Quantum Reservoir Engineering
w
Quantumsystemf s
y
Quantumcontroller
SW
y’u’
z
v
z’u1’
ru
u1
Fig. 4.5 Coherent feedback network for DFS generation.
In terms of the Hamiltonian and coupling operators, the system has the Hamiltonian
Hp = a†pMap where M is an n×n complex Hermitian matrix, and is linearly coupled to
the input fields via the coupling operators Lp1= α1ap, Lp2
= α2ap, Lp3= α3ap where
α1 ∈ Cnw×n, α2 ∈ Cnu×n, α3 ∈ Cn f×n are complex matrices. Then the relations (4.7)
specialize as follows:
Ap =−(
iM+12
α†1 α1 +
12
α†2 α2 +
12
α†3 α3
),
B1 =−α†1 ,
B2 =−α†2 ,
B3 =−α†3 ,
C = α1.
The starting point of the discussion that follows is the assumption that under the
influence of its natural environment w alone, (i.e., in the absence of the engineered
fields f and u), the system does not possess a DFS. Mathematically, this assumption
corresponds to the assumption that(Ap,B1
)is controllable and
(Ap,C
)is observable,
4.4 Decoherence Free Subsystem Synthesis 101
since these properties rule out the existence of a DFS in the plant (4.69) when B2 = 0,
B3 = 0; see Proposition 4.4.1 and [202, 59].
In many cases, system couplings can be engineered at a fabrication stage to reduce
unavoidable loss of energy due to decoherence [145, 185]. The process of tuning the
system at the fabrication stage does not involve feedback, and we let Lp2= 0, which
corresponds to α2 = 0 and B2 = 0 in (4.69); see Fig. 4.6.
w
Quantumsystemf s
y
Fig. 4.6 Open loop setup for DFS generation.
Then the system (4.69) reduces to that of the form
dap = Apapdt +B1dw+B3d f , (4.70a)
dy =Capdt +dw (4.70b)
Here, w and f symbolize the natural environment and the fabricated open-loop field,
respectively. Accordingly, the coupling operator Lp1corresponds to a fixed coupling
with the natural environment, while the coupling Lp3corresponds to the engineered
coupling. The physical realizability requirement imposes the constraint that
Ap +A†p +B1B†
1 +B3B†3 = 0, (4.71)
cf. (4.8). Recall [102] that a quantum stochastic differential equation of the form (4.70)
is said to be (canonically) physically realizable if it preserves the canonical commutation
relations, [ap,a†p] = apa†
p − (a∗paTp )
T = I, and is a representation of an open harmonic
oscillator, i.e., it possesses a Hamiltonian and a coupling operator. The satisfaction of
the identity (4.71) is a necessary and sufficient condition for physical realizability [102,
Theorem 5.1].
102 Chapter 4. Quantum Reservoir Engineering
Theorem 4.4.1 Suppose (−iM,B1) is controllable. Then a DFS cannot be created by
coupling the system to an engineered environment.
Proof. To prove the theorem we will show that the matrix Ap has all its eigenvalues
in the open left half-plane of the complex plane, and therefore it cannot have a DFS,
according to Proposition 4.4.1; see [59, Lemma 2].
First consider the system with a fixed coupling with the environment, i.e., Lp3= 0.
For this system, the physical realizability properties dictate that
Ap1 +A†p1 +B1B†
1 = 0, (4.72)
with Ap1 =−iM− 12B1B†
1; see (4.7).
Recall that for an arbitrary n×n matrix Φ and an n×m matrix B, the pair (Φ,B) is
controllable if and only if (Φ+ 12BB†,B) is controllable. Applying this fact to the pair
(−iM,B1) which is controllable by the assumption of the theorem, we conclude that
(Ap1,B1) is controllable. Thus, equation (4.72) can be regarded as a Lyapunov equation
Ap1P+PA†p1 +B1B†
1 = 0
with controllable (Ap1,B1), which has a positive definite solution P = I. Since B1B†1 ≥ 0,
according to the inertia theorem [22, Theorem 3], the above observation about the
existence of a positive definite solution to the Lyapunov equation implies that Ap1 must
have all its eigenvalues in the open left half-plane of the complex plane. As a result, if
(−iM,B1) is controllable, the corresponding passive quantum system with fixed coupling
cannot have a DFS, according to Proposition 4.4.1.
Next consider this system when it is coupled to an engineered environment, i.e.,
Lp3= 0 and B3 = 0. Since Ap1 has been shown to have all eigenvalues in the open
left half-plane of the complex plane, there exists a positive definite Hermitian matrix
P = P† > 0 such that
A†p1P+PAp1 < 0.
On the other hand, according to Corollary 4 of [134], the matrix 12B3B†
3P cannot have
eigenvalues in the open left half-plane of the complex plane, and therefore −12B3B†
3P−
4.4 Decoherence Free Subsystem Synthesis 103
12PB3B†
3 ≤ 0. This implies that
(Ap1 −12
B3B†3)
†P+P(Ap1 −12
B3B†3)< 0
and therefore Ap = Ap1 − 12B3B†
3 must have all its eigenvalues in the open left half-plane
of the complex plane. According to Proposition 4.4.1, this rules out the possibility for
the system with engineered coupling to have a DFS.
Next, suppose that (−iM, [B1 B3]) is not controllable5, therefore (−iM,B1) is not
controllable either. Theorem 4.4.1 does not rule out a possibility for a DFS to exist in
this case. It is easy to show that
ker(C T ) = ker(C Tw )∩ker(C T
f ),
where Cw, C f are the controllability matrices with respect to the inputs w and f , respec-
tively. From this observation, it follows that the dimension of the DFS of system (4.70)
is less or equal to the dimension of each of the decoherence free subsystems arising
when the quantum plant is coupled with the fixed and engineered fields only. This leads
to the conclusion that coupling the system with additional engineered fields can only
reduce the dimension of the DFS. In the remainder of the paper, we will show that using
coherent feedback, on the other hand, does allow to create or increase dimension of a
DFS.
4.4.2 Coherent Feedback Reservoir Engineering
In this section we consider a system of the form (4.69). To simplify the notation we
will combine two static channels w and f into a single channel, which will again be
denoted as w. More precisely, we combine the coupling operators Lp1and Lp3
into a
single operator Lp1. Then the system (4.69) reduces to a system of the form
dap = Apapdt +B1dw+B2du, (4.73a)
dy =Capdt +dw, (4.73b)
5Here, [B1 B3] is the matrix obtained by concatenating the rows of B1 and B3.
104 Chapter 4. Quantum Reservoir Engineering
where the new matrix B1 is composed of the previous matrices B1 and B3, so that using
the new notation we have
Ap =−(
iM+12
α†1 α1 +
12
α†2 α2
),
B1 =−α†1 ,
B2 =−α†2 ,
C = α1. (4.74)
For a coherent quantum controller for the quantum plant (4.69), we will consider
another open quantum linear annihilation only system. Such a system will be assumed
to be coupled with three environment noise channels, y′, z′ and v. The fields y′, z′
are to produce output fields which will be used to form the feedback, and the channel
v will be used to ensure that the constructed observer is physically realizable. As is
known [102], once physical realizability of the observer is ensured, one can readily
construct a scattering matrix, a Hamiltonian and a collection of coupling operators
describing the quantum evolution of the controller in the form of a quantum stochastic
differential equation (4.3). Alternatively, a physically realizable coherent controller can
be represented in the form of the quantum stochastic differential equation (4.6) [102],
i.e., in the form
dac =Acacdt +G1dy′+G2dz′+G3dv, (4.75a)
du′ =Kacdt +dy′, (4.75b)
du′ =Kacdt +dz′, (4.75c)
where for physical realizability, the following constraints must be satisfied [102, Theo-
rem 5.1]:
Ac +A†c +G1G†
1 +G2G†2 +G3G†
3 = 0, (4.76)
K =−G†1, (4.77)
K =−G†2. (4.78)
4.4 Decoherence Free Subsystem Synthesis 105
Interconnection between the controller and the plant are through scattering equations
relating the output fields of the plant with the input channels of the controller and vice
versa. Specifically, the scattering equation
y′
z′
= S
y
z
, (4.79)
links the output field of the plant y and the controller environment z with the input
controller channels y′, z′. Here, S is a unitary matrix partitioned as
S =
S11 S12
S21 S22
. (4.80)
Likewise, feedback from the controller (4.75) is via a unitary matrix W ,
u
u
=W
u′
u′
, W =
W11 W12
W21 W22
. (4.81)
The matrices Ac, G1 = −K†, G2 = −K†, G3, and the scattering matrices S, W are
regarded as the controller design parameters. Our objective in this paper is to find a
procedure for selecting those parameters so that the resulting coherently interconnected
quantum system in Fig. 4.5 possesses a decoherence free subsystem.
To devise the DFS synthesis procedure, we first note that the control system governed
by y,z,v and output u can be represented as
dac = (Acac − (G1S11 +G2S21)B†1ap)dt
+(G1S11 +G2S21)dw
+(G1S12 +G2S22)dz+G3dv, (4.82)
du = (−(W11G†1 +W12G†
2)ac
−(W11S11 +W12S21)B†1ap)dt
+(W11S11 +W12S21)dw
+(W11S12 +W12S22)dz
106 Chapter 4. Quantum Reservoir Engineering
Also, the closed loop system is described by the quantum stochastic differential equation
d
ap
ac
= Acl
ap
ac
dt +Bcl
dw
dz
dv
, (4.83)
with block matrices Acl , Bcl partitioned as follows
Acl =
Ap −B2(W11S11 +W12S21)B
†1 −B2(W11G†
1 +W12G†2)
−(G1S11 +G2S21)B†1 Ac
,
Bcl =
B1 +B2(W11S11 +W12S21) B2(W11S12 +W12S22) 0
G1S11 +G2S21 G1S12 +G2S22 G3
.
Let
Ac = Ap − B2(W11S11 +W12S21)B†1
+(G1S11 +G2S21)B†1
−B2(W11G†1 +W12G†
2) (4.84)
Then for Acl to have all eigenvalues on the imaginary axis or in the left half-plane of the
complex plane it is necessary and sufficient that the following matrices
A = Ap −B2(W11S11 +W12S21)B†1
−B2(W11G†1 +W12G†
2), (4.85)
A = Ap −B2(W11S11 +W12S21)B†1
+(G1S11 +G2S21)B†1 (4.86)
have all eigenvalues on the imaginary axis or in the left half-plane of the complex plane.
4.4 Decoherence Free Subsystem Synthesis 107
Proof. The matrix Acl has the same eigenvalues as the matrix
I 0
I −I
Acl
I 0
I −I
=
Ap −B2W11(S11B†1 +G†
1)
−B2W12(S21B†1 +G†
2)B2(W11G†
1 +W12G†2)
Ap −Ac
−B2(W11S11 +W12S21)B†1
+(G1S11 +G2S21)B†1
−B2(W11G†1 +W12G†
2)
B2(W11G†1 +W12G†
2)+Ac
.
Hence the lemma follows, due to the definition of Ac in (4.84).
Theorem 4.4.2 Suppose matrices S, W are given. Let G1, G2 be such that
1. The following linear matrix inequality (LMI) in G1, G2 is satisfied
R G1 G2
G†1 −I 0
G†2 0 −I
≤ 0, (4.87)
where
R =−B1B†1 −B2B†
2 −B2(W11S11 +W12S21)B†1
−B1(S†21W †
12 +S†11W †
11)B†2 +(G1S11 +G2S21)B
†1
+B1(S†11G†
1 +S†21G†
2)−B2(W11G†1 +W12G†
2)
−(G1W †11 +G2W †
12)B†2; (4.88)
2. The matrices A and A, defined in equations (4.85) and (4.86) respectively, have
all their eigenvalues in the closed left half-plane, with at least one of them having
eigenvalues on the imaginary axis.
Then a matrix G3 can be found such that the closed loop system (4.83) admits a DFS.
Proof. Via the Schur complement, (4.87) is equivalent to
R+G1G†1 +G2G†
2 ≤ 0.
108 Chapter 4. Quantum Reservoir Engineering
Therefore one can find G3 such that
R+G1G†1 +G2G†
2 +G3G†3 = 0.
From this identity and the expression (4.84), the identity (4.76) follows. This shows
that the feasibility of the LMI (4.87) ensures that the controller system (4.75) can be
made physically realizable by appropriately choosing G3. As a result, the closed loop
system, being a feedback interconnection of physically realizable systems, is a physically
realizable annihilation only system. Also, condition (b) and Lemma 4.4.2 ensure that
Acl has eigenvalues on the imaginary axis. Then it follows from Proposition 4.4.1 that
the closed loop system (4.83) has a DFS.
Note that matrices A and A can be rewritten as
A = Ap −B2(W11S11 +W12S21)B†1
−B2[W11 W12]
G†
1
G†2
, (4.89)
A = Ap −B2(W11S11 +W12S21)B†1
+[G1 G2]
S11
S21
B†
1 (4.90)
A necessary condition to ensure that an eigenvalue assignment can be carried out
for these matrices by selecting G1, G2, is that the pair (Ap,B2[W11 W12]) is control-
lable and the pair (Ap,
S11
S21
B†
1) is observable; the latter condition is equivalent to
the controllability of the pair (A†p,B1
[S†
11 S†21
]). Indeed, these controllability and
observability conditions imply that (Ap −B2(W11S11 +W12S21)B†1,
S11
S21
B†
1) is ob-
servable and (Ap −B2(W11S11 +W12S21)B†1,B2[W11 W12]) is controllable. Therefore, if
(Ap,B2[W11 W12]) and (A†p,B1
[S†
11 S†21
]) are controllable, one can always select G1
and G2 so that the matrices A, A have a required eigenvalue distribution. Thus the condi-
4.4 Decoherence Free Subsystem Synthesis 109
tions of Theorem 4.4.2 boil down to solving a simultaneous pole assignment problem
under an LMI constraint.
We next demonstrate that our pole assignment problem captured quantum plant-
controller DFS architectures considered in [138, 202].
4.4.3 Special Case 1: DFS synthesis using a coherent observer [138]
In [138], the DFS synthesis was carried out using a quantum analog of the Luenberger
observer for a class of linear annihilation only systems as described by (4.5); see Fig. 4.7.
This controller structure is a special case of the architecture in Fig. 4.5, when
S =
I 0
0 I
, W =
0 I
I 0
With this choice of S and W , we have from (4.84)
Ac = Ap +G1B†1 −B2G†
2 (4.91)
w
Quantumsystem
y
Quantumcontroller
u
z
v
Fig. 4.7 Coherent plant-observer network considered in [138].
110 Chapter 4. Quantum Reservoir Engineering
Corollary 4.4.3 Suppose the pair (Ap,C) is observable and the pair (Ap,B2) is control-
lable. Let G1, G2 be such that
1. The following linear matrix inequality (LMI) is satisfied
R G1 G2
G†1 −I 0
G†2 0 −I
≤ 0, (4.92)
where
R =−B1B†1 −B2B†
2
+G1B†1 +B1G†
1 −B2G†2 −G2B†
2 (4.93)
2. The matrices
A = Ap −B2G†2, (4.94)
A = Ap +G1B†1 (4.95)
have all eigenvalues on the imaginary axis or in the left half-plane of the complex
plane, with at least one of them having eigenvalues on the imaginary axis.
Then the closed loop system admits a DFS.
Proof. Via the Schur complement, condition (4.92) is equivalent to the condition
−B1B†1 −B2B†
2 +G1B†1 +B1G†
1
−B2G†2 −G2B†
2 +G1G†1 +G2G†
2 ≤ 0.
This ensures that
Ac +A†c +G1G†
1 +G2G†2 ≤ 0.
Therefore, one can find G3 such that the controller is physically realizable. The claim
then follows from Theorem 4.4.2.
4.4 Decoherence Free Subsystem Synthesis 111
4.4.4 Special Case 2: Coherent feedback DFS generation model from [202]
Consider a system of Fig. 4.5 in which S = I, W = I, and let G2 = 0, G3 = 0. This
corresponds to the system shown in Fig. 4.8; which was considered in [202]. In this
case, the controller matrix becomes
Ac = Ap −B2B†1 +G1B†
1 −B2G†1. (4.96)
w
Quantumsystem
y
Quantumcontroller
r
u
Fig. 4.8 Special Case 2: Coherent feedback network for DFS generation consideredin [202].
Corollary 4.4.4 Suppose the pair (Ap,C) is observable and the pair (Ap,B2) is control-
lable. Let G1 be such that
1. The following equation is satisfied
−(B1 +B2)(B1 +B2)† +G1(B1 −B2)
†
+(B1 −B2)G†1 +G1G†
1 = 0; (4.97)
112 Chapter 4. Quantum Reservoir Engineering
2. The matrices
A = Ap −B2B†1 −B2G†
1, (4.98)
A = Ap −B2B†1 +G1B†
1 (4.99)
have all eigenvalues on the imaginary axis or in the left half-plane of the complex
plane, with at least one of them having eigenvalues on the imaginary axis.
Then the closed loop system admits a DFS.
Proof. Condition (4.97) ensures that
Ac +A†c +G1G†
1 = 0.
Next, A and A have eigenvalues on the imaginary axis or in the open left half-plane,
hence the statement of the corollary follows from Theorem 4.4.2.
4.4.5 Examples
Example 1
To illustrate the DFS synthesis procedure developed in the previous section, consider
a system consisting of two optical cavities interconnected as shown in Fig. 4.7. The
system is similar to those considered in [126].
The cavity to be controlled is described by equation (4.73), with all matrices becom-
ing complex numbers
Ap = −iM− κ1 +κ22
, B1 =−√κ1, B2 =−√
κ2,
C = −B∗1 =
√κ1. (4.100)
Here, κ1, κ2 are real non-negative numbers, characterizing the strength of the couplings
between the cavity and the input fields w and u, respectively, and M characterizes the
Hamiltonian of the cavity.
Clearly, the pair (Ap,C) is observable and the pair (Ap,B2) is controllable, therefore
the optical cavity cannot have a DFS unless the cavity is lossless. To synthesize a DFS,
let us connect this cavity to another optical cavity with the same Hamiltonian, as shown
4.4 Decoherence Free Subsystem Synthesis 113
in Fig. 4.7. This corresponds to letting the controller have the coefficients
Ac = −iM− κ3 +κ42
, G1 =−√κ3, G2 =−√
κ4,
K =√
κ4, G3 = 0. (4.101)
and letting the scattering matrices S and W be
S =
1 0
0 1
, W =
0 1
1 0
. (4.102)
We now apply Corollary 4.4.3 to show that the parameters κ3,κ4 for the controller cavity
can be chosen so that the two-cavity system has a DFS. It is readily verified that the
matrices A and A in (4.94), (4.95) reduce to
A = −iM− κ1 +κ22
−√κ2κ4, (4.103)
A = −iM− κ1 +κ22
+√
κ1κ3. (4.104)
From Corollary 4.4.3, we need either A or A to have poles on the imaginary axis in
order to create a DFS within the closed-loop system. Clearly, for the two-cavity system
under consideration this can only be achieved by placing the pole of A at the origin. For
this, the coupling rate κ3 of the controller must be set to
κ3 =(κ1 +κ2)
2
4κ1. (4.105)
Also we must satisfy the LMI condition (4.92). The matrix R in this example reduces to
R =−κ1 −κ2 +2√
κ1κ3 −2√
κ2κ4. (4.106)
Hence, using (4.105) and (4.106), the LMI condition (4.92) reduces to the two following
inequalities:
Re[−κ1 −2κ1
√κ2κ4 ±
√D]≤ 0, (4.107)
114 Chapter 4. Quantum Reservoir Engineering
where
D = κ21 +κ
31 +2κ
21 κ2 +κ1κ
22 +4κ
21 κ4
+4κ21 κ2κ4 −4κ
21√
κ2κ4.
The inequality (4.107) is the only constraint for the remaining coupling parameter κ4
to be determined. Notice that there is an obvious solution to this inequality in the case
where κ1 = κ2 = κ . The solution is κ3 = κ4 = κ which satisfies both (4.105) and (4.107).
The above calculations demonstrate that by placing the pole of the controller on the
imaginary axis, one can effectively create a DF mode which did not exist in the original
system. This fact has been established previously in [126] by calculating the system
poles, whereas we have arrived at this conclusion from a more general Corollary 4.4.3,
as a special case.
Example 2
We now present an example in which, the DFS is created which is shared between
the controlled system and the controller. The controlled system in this example consists
of two cavities as shown in Fig. 4.9.
Denote the matrices associated of the Hamiltonians corresponding to the each cavity
internal dynamics as M1, M2. Also for the convenience of notation, define the complex
numbers
γ j =√
κ j, j = 1, . . . ,4,
associated with the coupling strengths within the cavities. All four constants are assumed
to be nonzero.
4.4 Decoherence Free Subsystem Synthesis 115
w
u
y
r
k1
k2
k4
Cavity 1
k3
Cavity 2
Fig. 4.9 The two-cavity system for Example 2.
Then the equations governing the dynamics of the two-cavity system have the form
of (4.73) with
Ap =
−
(iM1 +
|γ1|2+|γ2|22 + γ∗1 γ2
)−γ2γ∗3
−γ∗1 γ3 −(
iM2 +|γ3|2
2
) ,
B1 =
−(γ1 + γ2)
−γ3
, B2 =
−γ4
0
,
C = −B†1 =
[γ1 + γ2 γ3
]. (4.108)
116 Chapter 4. Quantum Reservoir Engineering
To verify observability of the pair (Ap,C), we observe that
det
C
CAp
=
12
γ∗3(|γ3|2(γ∗1 − γ
∗2 )
+(γ∗1 + γ∗2 )(|γ1|2 −|γ2|2)
+2i(γ∗1 + γ∗2 )(M1 −M2)) .
Suppose γ1 = −γ2, then det
C
CAp
= γ∗3 γ∗1 |γ3|2, and we conclude that the matrix
C
CAp
is full rank. This implies that in the case γ1 = −γ2, the pair (Ap,C) is
observable. Also, the pair (Ap,B2) is controllable, since
det[B2 ApB2
]=−γ
24 γ3γ
∗1 = 0.
These observations allow us to apply Corollary 4.4.3 to construct a DFS by interconnect-
ing the two-cavity system with a coherent quantum observer, which we now construct.
For simplicity, choose
G1 =
g1
0
, G2 =
0
g2
.
With this choice of G1 and G2 and under the condition γ1 = −γ2, the matrices A =
Ap −B2G†2 and A = Ap +G1B†
1 take the form
A =
−iM1 −γ2γ∗3 + γ4g∗2
−γ∗1 γ3 −(
iM2 +|γ3|2
2
) ,
A =
−iM1 −γ2γ∗3 −g1γ∗3
−γ∗1 γ3 −(
iM2 +|γ3|2
2
) .
Letting g2 =γ∗2 γ3γ∗4
, g1 =−γ2 allows us to conclude that each of the matrices A and A have
one imaginary eigenvalue and one eigenvalue with negative real part, −(
iM2 +|γ3|2
2
).
4.5 Concluding Remarks 117
It remains to show that the LMI condition (4.92) is satisfied in this example. Noting
that with the above choice of g1, g2,
R =
−|γ4|2 2γ2γ∗3
2γ∗2 γ3 −|γ3|2
,
R < 0 holds provided |γ4|2 > 4|γ2|2. Next, the LMI (4.92) in this example requires that
−|γ4|2 2γ2γ∗3 −γ2 0
2γ∗2 γ3 −|γ3|2 0 γ∗2 γ3γ∗4
−γ∗2 0 −1 0
0 γ2γ∗3γ4
0 −1
< 0 (4.109)
Using the Schur complement, this requirement is equivalent to
−|γ4|2 2γ2γ∗3
2γ∗2 γ3 −|γ3|2
+
|γ2|2 0
0 |γ2|2|γ3|2|γ4|2
< 0.
The latter condition holds when |γ4|2 > (3+√
10)|γ2|2.
4.5 Concluding Remarks
In this chapter, we have applied methods from quantum reservoir engineering and
coherent control to address one of the most challenging aspects of quantum computing,
namely decoherence.
Firstly, we derive an algebraic procedure based on Lyapunov formalism for the
engineering of dissipation control. The scalability and error correction capacity of
the dissipation control of the form L = UV has been investigated in detail. Hence,
the Lyapunov formalism may facilitate the application of the classical decomposition-
aggregation technique to a large scale quantum system.
Secondly, we have proposed a general coherent quantum controller synthesis proce-
dure for generating decoherence-free subspaces in quantum systems. Decoherence-free
118 Chapter 4. Quantum Reservoir Engineering
components capable of storing quantum information are regarded to be essential for
quantum computation and communication, as quantum memory elements [126].
Chapter 5
Quantum Programming and
Simulation
Let the computer itself be built of quantum mechanical elements which obey
quantum mechanical laws.
Richard Feynman
The first publicly-available quantum computer was announced in 2016 by IBM [20],
which provides a prototype platform for researchers around the world to test their
quantum computing programs on a real quantum computer. Despite offering an excellent
experience for quantum programmers, the platform is still only a proof-of-concept and
could not provide the scale and reliability needed for algorithm discovery or development.
Indeed, from the evolution of conventional computer science, we have learned that
software is as important as hardware. Software innovations can even lead and guide
hardware development. For quantum computing, the current set of applications is very
limited. Hence, computer simulation and emulation of quantum computing algorithms
on classical computers will still be the most accessible and reliable way that we can rely
on to explore quantum software and applications in the foreseeable future. Thanks to
the constant innovation of the microelectronic industry, we now have high-performance
computing platforms which are capable of providing simulation and emulation of sizable
quantum computers. Furthermore, innovative mathematical approaches could be used to
120 Chapter 5. Quantum Programming and Simulation
extend the capability of those simulators in specific use cases beyond the memory-bound
limit of the Hilbert space.
The work presented in this chapter involves the definition of a quantum assembly
language; the complete tool-chain which supports compilation from high-level procedu-
ral wrapper language to the assembly instructions and the serialization of the assembly
instructions into a binary data structure which could be sent to a simulator or a real
quantum computers; and lastly a simulator engine which is based on Feynman path
integral that provides unique performance advantages under various test cases as well
as could be integrated with a conventional linear algebra simulators to achieve optimal
time-to-result for all cases. These were done as part of my research internship at Atos
Quantum, a research lab of the Big Data and Security division of Atos SE.1
5.1 Quantum Programming Language
A quantum computer, as described by the quantum gate model, can execute arbitrary
quantum programs (algorithms) by applying a specific sequence of gate operations.
This computing model is essentially similar to classical computing devices where those
low-level operations are described by machine code instructions.
These instructions are directly mapped to a human-readable form which is often
called assembly languages. In principle, every computer programs (e.g. games, simula-
tion, web browsers, etc.) can and in fact must be programmed by the assembly language
specific to the hardware that they are running on. The tedious process of writing com-
puter programs instruction-by-instruction is thus not very practical and is best reserved
for very specific use cases (e.g. low-level firmware). Most computer programmers
learned to code in the so-called high-level programming languages (e.g. C/C++, Python,
Java, Javascript, etc.) which are both hardware agnostic and algorithmically descriptive
(as supposed to be operationally descriptive).
Hence, a quantum programming paradigm could mimic that of its classical coun-
terpart, i.e. providing multi levels of abstraction from the low-level gate operations to
the high-level algorithmic descriptions. This includes an assembly-language definition
which closely matches (modulo trivial gate transformations) and directly executable on
1This work does not reflect the views or opinions of Atos SE or any of its employees.
5.1 Quantum Programming Language 121
quantum computer hardware. On top of that, a higher level language and the correspond-
ing compiler is ultimately necessary for practical quantum programming purposes. A
comprehensive solution for quantum programming language also needs to be flexible
and agnostic in the sense that it can interface with both quantum hardware as well as
simulators/emulators.
5.1.1 Quantum Assembly Language
The quantum assembly (QASM) language is an abstract description of operations of
a general-purpose quantum computer. This includes a set of quantum instructions as
well as the syntax and semantics rules. In designing our own QASM language, the
Atos Quantum Assembly Language (AQASM), we take into consideration the following
criteria:
• Fully quantum: Although the first generation of AQASM will only interact with a
classical simulator/emulator back-end, the long-term goal is to be able to use it on
a real quantum computer.
• Low-level: AQASM is designed around the quantum gate model. The basic
instructions correspond to elementary quantum gates.
• Extendable: AQASM also supports user-defined quantum gates and arbitrary gate
extensions by adding control or gate parameterisation.
• Standard: AQASM is the universal language for all future ATOS-developed
quantum simulators to maintain interoperability.
• Complete toolchain: We also plan to develop high-level wrappers together with
proper compilation programs in order to provide a high-level abstraction to devel-
opers. Also, it is essential to provide parsing/translating support for other existing
quantum programming languages to AQASM.
The AQASM has native support for arbitrary quantum gates (including user-defined
unitary gates and gate parameterisation), expanding quantum gates (e.g. complex
conjugate and adding control), and basic quantum operations (e.g. quantum Fourier
transform). An example of AQASM programs is given in Fig. 5.1.
The compilation of AQASM into quantum circuit data structures is implemented in
Coco/R [113], which is a widely-used framework for syntax validation and compiler
122 Chapter 5. Quantum Programming and Simulation
DEFINE GATE qft2 = QFT[2]DEFINE GATE qft10 = QFT[5]
BEGINqubits 3
QFT q[0],q[1],q[2]H q[0]S q[1]T q[2]CNOT q[0],q[1]GATE ex q[1]X q[2]CNOT q[1],q[2]IQFT q[0],q[1],q[2]DAG(CTRL(RX[PI/2])) q[2],q[0]
END
Fig. 5.1 An example of arbitrary quantum circuit described by the Atos quantum assem-bly language.
generation. By defining the compiler descriptions, which includes the list of keywords
(tokens) as well as all the syntax and semantics rules, we can use Coco/R to generate the
corresponding scanner and parser to process the AQASM inputs.
5.1.2 High-level Programming and Data Processing
The advantages of assembly languages are the strong correspondence between the
language, i.e. the instructions, and the physical operations, i.e. quantum gates, on
quantum hardware. Experimentalists can import such a program as in Fig. 5.1 and
readily generate the control sequence for lab instruments to implement the AQASM
program. Or as in the case of our simulation back-end, it can take those AQASM
programs and perform the simulation accordingly.
However, the main drawback of low-level languages, e.g. AQASM, is its verbosity,
which often hides the algorithmic structure and hierarchy of the program. For example,
let’s take a look at an 18-qubit quantum Fourier transform program written in AQASM2
as depicted in Fig. 5.2. There are approximately 170 AQASM instructions in that
program which has been abbreviated for brevity.2This is an explicit AQASM representation of the QFT algorithm. The language does have native
support for QFT and inverse QFT (IQFT) as demonstrated in Fig. 5.1.
5.1 Quantum Programming Language 123
BEGINqubits 18H q[0]CTRL(PH[PI/2]) q[1], q[0]H q[1]CTRL(PH[PI/4]) q[2], q[0]CTRL(PH[PI/2]) q[2], q[1]H q[2]CTRL(PH[PI/8]) q[3], q[0]CTRL(PH[PI/4]) q[3], q[1]CTRL(PH[PI/2]) q[3], q[2]H q[3]CTRL(PH[PI/16]) q[4], q[0]// Skipping...//H q[17]END
Fig. 5.2 An example of 18-qubit QFT program written in AQASM. The body of theprogram has been abbreviated for clarity.
Most people will find it difficult to comprehend a program as in Fig. 5.2. More
importantly, it is extremely tedious and error-prone for developers/programmers to write
quantum programs in that manner. Hence, upon the complete definition of AQASM
in terms of language features, semantics and syntax rules, we embarked on creating
high-level wrappers for AQASM in commonly-used classical programming languages.
For instance, for Python, we developed PyAQASM package 3 which supports looping,
subroutine declarations, function calls, classical processing of intermediate data, etc. For
comparison, the same 18-qubit QFT program in Fig. 5.2 can be expressed in PyAQASM
using the program listed in Fig. 5.3.
The looping and parametrisation supported by PyAQASM not only shorted the
program significantly but also allow novice readers to comprehend the overall structure
of the algorithm. In addition, by having Python support with PyAQASM, we can
integrate the programming workflow more seamlessly. For example, a web interface
can be used to input the PyAQASM program; the user’s computer can then generate
3The author and other researchers of the ATOS Quantum Lab collaborated on defining the AQASMlanguage itself. The implementation of the Coco/R-based compilation unit was developed by anotherresearcher. The author supported the testing of the compiler and implemented the first high-level Python-based wrapper as described in this section.
124 Chapter 5. Quantum Programming and Simulation
from AQASM import *
nbqubits = 18
AQASM_prog = AQASM("QFT.aqasm", nbqubits)
for i in range(nbqubits - 1):AQASM_prog.apply(qGate("H"), qubit(i))for j in range(i+1):
phase = "PI/" + str(2**(i-j+1))AQASM_prog.apply(
qGate("PH", CTRL = ’true’, PHASE = phase),[qubit(i+1), qubit(j)])
AQASM_prog.apply(qGate("H"), qubit(nbqubits - 1))
AQASM_prog.finish()
Fig. 5.3 An example of 18-qubit QFT program written in PyAQASM. PyAQASM isa Python-based wrapper for AQASM which can export the corresponding AQASMprogram for simulation and experiment purposes.
the AQASM and serialise it via the Internet to the simulator (e.g. a supercomputer in a
remote server room) or real quantum hardware in a remote laboratory; the result data of
that algorithm is returned to the user for further data analysis and visualisation.
In the next session, we will discuss the ATOS quantum simulator which is a High-
Performance Computing (HPC) software program designed specifically to simulate
quantum programs written in AQASM.
5.2 Simulation Engine
The exponential size of the Hilbert space which describes the computation space of
quantum computers makes the quantum simulation on classical computers impossible
in general. Indeed, a conventional linear-algebra based approach will be limited firstly
by the computer memory to store the quantum state. In this section, we present an
alternative approach to the classical simulation which employs the space-time trade-
off 4 to extend the ability to simulate a larger number of qubits and achieve shorter
time-to-result in specific use cases.
4memory footprint vs. runtime
5.2 Simulation Engine 125
The program code was written by the author in C++ programming language (includ-
ing OpenMP [130] extension), with the primary goal is to reach high performance on
High-Performance Computing (HPC) systems.
5.2.1 Overall Architecture
The modular representation of our simulation suite is depicted in Fig. 5.4, in which
the top right-hand side is the core simulation engine running on high-performance
computer clusters. The left-hand side represents user modules (i.e. running on a personal
computer/laptop) which handle user interfaces for code editing or importing (from other
programming languages) as well as other post-simulation data analysis and visualisation.
The ATOS compiler, which also runs on the user side, compiles and serialise AQASM
Fig. 5.4 The functional diagram of ATOS quantum simulator.
code into a binary file for remote execution.
The technical specifications of the computing hardware and software packages are
listed in Table 5.1.
On the hardware side, the ATOS Bullion server is built with an extremely large
amount of high-bandwidth memory (per CPU socket as well as in total) which is to
126 Chapter 5. Quantum Programming and Simulation
Table 5.1 Technical details of the ATOS path-integral-based quantum simulator.
Programming Language C++Compiler gcc (6.2.1)Multithreading OpenMP
Bullion (8 sockets, 6TB of main memory)Platform Intel Xeon CPU E7-8890 v4 (2.20GHz)
24 cores/socketLibraries Thrift (serialise/deserialise)
Coco/R (AQASM compiler)
address the memory-bound nature of quantum mechanics simulation. The memory
specifications of the specific machine that is used for this project is listed in Table 5.2.
It is worth noting the drop in bandwidth (by more than an order of magnitude) when a
CPU wants to access memory from another node. This is one of the reasons why we
want to explore the path integral approach since it does not require access to the global
state vector during the computation. On the other hand, a linear algebra simulator, which
constructs a global state vector and distributes it among multi-node clusters, will need
node-to-node data exchange in order to perform the matrix-vector multiplication since
each node keeps a portion the state vector of the size (2m) locally and performs state
update on its part (which requires inter-process communication from time to time).
Table 5.2 Computer memory capacity and bandwidth specifications of the simulationhardware platform.
Memory per socket 768GBMemory per node 1.5TBNumber of nodes 4Memory bandwidth 58 GB/sInter-socket BW 30 GB/sInter-node BW 4.9 GB/s
5.2.2 Technical Implementation
In the Feynman path integral approach, the simulation of a quantum operation is divided
into individual paths whose results are then combined as shown in equation (5.1). The
term “path” here refers to the fact that we have an intermediate node (z) which can
5.2 Simulation Engine 127
take any random values. Each assigned value of z represents a path from x to y and
by summing the contributions of all possible paths, we can compute the propagation
amplitude from x to y through the quantum operation U .
⟨x|U |y⟩= ∑z⟨x|U1 |z⟩⟨z|U2 |y⟩ , U =U1U2 (5.1)
In the context of quantum computing simulation, Fig. 5.5 demonstrates a simple
example where we have a three-qubit register with the initial state of |010⟩. The paths
generated by applying quantum gates on the qubit register are explicitly visualised which
highlights the fact that in this path integral simulation, each path “walker” does not
need access to the global state vector. In an HPC multi-threading environment, each
computing thread can handle each path independently hence can be beneficial when we
have a massively parallel computer.
Fig. 5.5 Example of a path integral diagram on a 3-qubit register starting at the state
|010⟩ and applying the a sequence of two gates[
α γ
γ∗ β
]and
[a bb∗ c
]on the second
qubit.
In terms of resources, since each path computation is simply a traversal from a
classical input and classical output (n-bit string), the memory requirement scales linearly
with the number of qubits (O(n)). On the other hand, the upper-bound for run-time is
O(2nx2depth), where the depth is the number of quantum gates. Therefore, there are only
a few scenarios where path integral is the preferred method of simulation, e.g. classical
128 Chapter 5. Quantum Programming and Simulation
input state (initial state is classical), or fat-short quantum circuits (many qubits but only
a few gates), etc. More importantly, by analyzing the input circuit, we can devise the
optimal simulation strategy in terms of runtime or memory consumption.
At the high level, the simulation program is designed using the object-oriented model.
Some key objects classes and their relative relationship is depicted in Fig. 5.6.
Fig. 5.6 Object-oriented design: Quantum circuit simulation objects
The serialised byte-code from the front-end will be translated into the internal data
structures from which the simulator further derives an execution model in terms of paths.
Those paths are distributed among computing threads and recollected at the end to form
the final simulation result.
5.2.3 Benchmark Results
In order to validate the simulator, we performed some initial tests using basic quantum
circuits such as all Hadamard gates (on each qubit) and the Quantum Fourier Transform
(QFT). The main purpose of those tests is to confirm the threading scalability of the
path integral approach, i.e. we get a linear reduction in runtime when more computing
resources (CPU cores) are added. This is the main difference between the path integral
and the linear-algebra methods where the latter is often constrained by memory traffic.
5.2 Simulation Engine 129
First, we ran a scalability test in which we gradually increased the number of
computing threads available for the simulator. The data for a test suite consists of all
Hadamard gates on a 30-qubit register is shown in Fig. 5.7.
Fig. 5.7 Run-time vs. number of computing thread benchmark of the simulator runningall Hadamard gates on a 30-qubit register.
Fig. 5.7 clearly demonstrates the linear scaling that we expect from a path-integral
simulator, especially when the thread count is in the range from 16 to 256 threads.
Outside that range of core count, there are a few hardware-related effects kick in which
changes the expected scaling behaviours. In the low core count range (8 vs. 16), since the
number of cores per CPU socket is 24, running only a few cores, e.g. 8 cores, will allow
the CPU to throttle the clock frequency of those executing cores hence the difference in
runtime is minimal when we double the core count in this case. On the other hand, the
total number of physical hardware cores that are available is 384. Hence, using more
threads than that will not result in any significant performance improvement as we can
see in the tail-end of Fig. 5.7.
Next, we compare the ATOS path-integral simulator against the publicly available
Microsoft LIQUi |⟩ (Fig. 5.8) and a generic linear algebra simulator such as Intel qHiP-
STER (Fig. 5.9). In both cases, we use Quantum Fourier Transform as the test case.
130 Chapter 5. Quantum Programming and Simulation
Fig. 5.8 Run-time comparison between the ATOS simulator and LIQUi |⟩ running theQuantum Fourier Transform using the same number of threads (128). The initial state is|00...0⟩.
Fig. 5.9 Run-time comparison between the simulator and linear algebra (LA) basedsimulator. We implemented a generic LA quantum simulator using the same computingmodel as qHiPSTER so that we can run on the local Bullion server. The runtime data ofthe LA simulator matches the public data of qHiPSTER (after compensating the memorybandwidth differences) as shown by the red triangles in the cases of 29 and 30 qubits.The ATOS path-integral simulator provides faster time-to-result in this case as comparedto the LA-based simulator.
When comparing to Microsoft LIQUi |⟩, Fig. 5.8 shows a significant performance
gap in favour of the ATOS simulator. This can be explained partly because the ATOS
simulator is heavily optimised for this particular hardware configuration while the
LIQUi |⟩ simulator is designed mainly for general PC-based setups. Secondly, the path-
5.2 Simulation Engine 131
integral based simulator can take advantage of the fact that the input state is classical
hence eliminating a lot of configuration paths.
Fig. 5.9, on the other hand, shows a much closer gap between an optimised LA-based
simulator and a path integral based counterpart. The path-integral simulator is still faster
in this case stemming from the fact that the input state is singular. One thing that we
can observe in Fig. 5.9 is the almost doubling of run time for the path integral simulator
when a qubit is added. This is due to the fact that the number of paths which need to be
explored doubles.
Preliminary data from the ATOS quantum simulator has been promising thanks
to its innovative approach to maximising low-level parallelism to achieve speed-up
in various use cases. The performance advantage is also derived from software and
hardware co-design for the high-end Bull Bullion HPC platforms. A proposal to study
the feasibility of using FPGA as an accelerator for this type of simulation is summarized
in Appendix B. The simulator is further enhanced and maintained by the ATOS Quantum
Lab after the author finished the research internship there.
This page intentionally left blank.
Chapter 6
Simulating Input-Output Quantum
Systems with LIQUi|⟩
Never try to solve all the problems at once — make them line up for you
one-by-one.
Richard Sloma
Quantum computation is usually described by unitary operations which is based
on closed quantum system dynamics. In contrast, open quantum dynamical systems
are often modeled by the master equation which takes into account the environmental
interaction. Using the complementary Heisenberg-picture approach, we can derive the
quantum stochastic differential equations that describe the dynamics of system operators
of an open quantum plant. This is commonly referred to as the quantum input-output
formalism, which underpins the burgeoning field of quantum feedback control.
The inclusion of dissipation processes, allows us to passively and autonomously
perform a wide variety of quantum operations, such as enhancing adiabatic quantum
computation, and state preparation/stabilization. For instance, Chapter 3 and 4 have
demonstrated the application of open quantum system models in developing protocols
to improve the stability of qubit systems. More importantly, fully coherent feedback
loops mediated by field coupling without measurement have been shown to have major
advantages such as high operating speed for quantum optics applications and the non-
demolition nature of coherent operations.
134 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
In this chapter, we describe the use of Microsoft Language-Integrated Quantum
Operations framework (LIQUi|⟩[194]) in simulating complex engineered open quantum
systems by discretizing the physical system composition, and introducing a dissipation
emulator gadget. The Hamiltonian simulator module of LIQUi|⟩ is used for time
evolution simulation of the Trotter decomposed unitary.
In particular, we will apply this method to compound gradient echo quantum memory
which is capable of efficiently capturing and storing a quantum state of light. In these
experiments, memory hardware consists of a continuous solid-state material (rare-earth-
ion-doped crystal) which will be represented in simulation by a network of qubits. This
not only provides analysis results for such a discrete quantum memory model, but can
also potentially be used as a blueprint for a quantum digital signal processor capable of
processing incoming continuous variable quantum information.
6.1 Background
As introduced in Chapter 1, quantum network theory has been developed based on
the quantum stochastic differential equation model. In particular, the modular and
system-theoretic SLH parametrization scheme [56] provides the mathematical frame-
work for quantum network description and analysis. The SLH-based Quantum Hardware
Description Language (QHDL)[168], which is illustrated in Figure 6.1, has been suc-
cessfully deployed to design a number of quantum optics devices and applications such
as quantum NAND logic gate and latch[107], optical Set-Reset flip-flop[105], and a
coherent error correction loop[91]. Backend differential equation solver based tools such
as QuTiP[85] and XMDS[31] are often used to simulate the system dynamics described
by its SLH parameters.
The Microsoft Language-Integrated Quantum Operations framework (LIQUi|⟩[194]),
which is based on quantum gate formalism, is not specifically designed for this type
of low level dynamical simulation. However, since LIQUi|⟩ is capable of faithfully
simulating a general purpose quantum computer, porting the dynamical SLH model onto
LIQUi|⟩ will pave the way for the implementation of discrete/digital quantum emulators
of open quantum systems.
6.1 Background 135
Functional Spec.(e.g. gate/latch/memory/control)
Quantum OpticsCircuit Editor
Component Netlist
QHDL Definition
DynamicalSimulation
ComponentLibrary
Symbolic AlgebraNumerical Simulation
Fig. 6.1 QHDL design flow which incorporates quantum optics schematic editor, netlist-ing tool, Hardware Description Language (HDL) parser, and a full quantum dynamicalsimulation backend.
One recent development of the SLH modeling approach is the effort to extend its
application to a wide variety of input states besides the conventional vacuum inputs,
such as thermal field, single-photon and two-photon states. Next, we will introduce the
SLH model for a system driven by a continuous-mode single-photon wave packet.
6.1.1 Single-photon State
Depending on its source, a photon can be single or multi-modal. In this work, continuous-
mode single photon state is concerned. We can decomposed such state in the time and
frequency domains as followings,
∣∣∣1ξ
⟩=∫ +∞
0ξ (t)dB†(t) |0⟩ ,
∣∣∣1ξ
⟩=∫ +∞
−∞
dωξ (ω)a†(ω) |0⟩ . (6.1)
136 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
Plant(SLH)
λ(t)σ−
Fig. 6.2 Single-photon generating filter cascading realization.
The function ξ (t) is the time-domain wave packet shape of the photon, which can be
interpreted as the probability amplitude of detecting the photon at time t. ξ (ω) is the
Fourier transform of ξ (t).
In the following simulation, we will consider incoming single-photon wave packets
which have Gaussian shape parametrized by:
ξ (t) =(
Ω2
2π
)1/4
exp[−Ω2
4(t − tc)
2], (6.2)
where tc specifies the peak arrival time and Ω is the frequency bandwidth of the pulse.
Using a pre-“signal-generating filter” which is driven by vacuum, we can derive
the SLH model for the augmented system. Equations (6.3) and (6.4) show the SLH
parameters of the pre-filter ancilla and the overall network, respectively.
Ganc = (I,λ (t)σ−,0), λ (t) =ξ (t)√∫
∞
t |ξ (s)|2ds, (6.3)
GT =(
S, L+λ (t)Sσ−, H +λ (t)Im(L†Sσ−)), (6.4)
in which Gs = (S,L,H) is the original system parameters.
The signal-generating filter (Ganc) is a vacuum-driven two-level atom initially pre-
pared in the excited state which will decay in time to the ground state via spontaneous
emission of a photon into the light field. That output field is then channeled to the system
of interest as shown in Figure 6.2.
6.1 Background 137
Fig. 6.3 Gradient settings for GEM operations
6.1.2 Compound Gradient Echo Memory
In the conventional quantum computation paradigm, there is little or no use of a “memory”
device since in most cases the problems are formulated as isolated computing tasks
with classical information inputs, e.g. factoring, optimization, etc. Nevertheless, if
networking between distant quantum computers are required, photons are the best
candidate for quantum information carriers. In that scenario, a quantum memory is
needed to capture and release photons on-demand.
One of the most promising schemes for quantum memory is the so-called Controlled
Reversible Inhomogeneous Broadening (CRIB). CRIB can be implemented on any
“controllable” absorbing physical system including NV centers, quantum dots, atomic
vapors and rare-earth-metal-ion-doped solids. Gradient echo memories (GEM) fall
under the CRIB category, in which an electric field induced Stark shift is deployed for
broadening. Two different linear gradients are applied for writing and reading phases
that causes the rephasing effect as shown in Figure 6.3.
While semi-classical approaches [41] can be used to simulate the memory operations,
no direct access to individual quantum state of each ion can be attained that way. On
the other hand, using the SLH formalism, Hush et al. [80] modeled the memory as a
series of optical cavities. Each cavity represents a collection of atomic absorbers at the
corresponding location in the memory, which all absorb light of a single frequency. The
number of excitations within a given cavity may be thought of as the number of excited
atoms in the ensemble at the corresponding location. In this framework, the SLH model
138 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
of a gradient echo memory is given by
GGEM = G1 ▷G2...▷Gn (6.5)
=
(SGEM = I, LGEM = ∑
k
√βak,
HGEM = ∑k
ξka†kak +
β
2i
N
∑j=2
j−1
∑k
(a†
jak −h.c))
.
The above model is a vacuum input model which has not taken into account the source
term. To simulate non-classical input such as the single photon state, the generating
filter approach mentioned previously becomes handy and can be apply directly to get the
complete model as followings,
GT = Ga ▷GGEM (6.6)
=
(I,LGEM +λ (t)σ−,HGEM +
12i
(λ∗(t)σ+LGEM −h.c
)),
in which λ (t) term dictates the photon wave packet shape as given by Eq. (6.3).
The cascading cavity approach is extremely accurate in the case that there are high
concentrations of light absorbers (ions). What we want to investigate is the performance
of such memory protocols in the discrete settings where only a countable number of
two-level qubits involved as absorbers. More importantly, the dynamics is simulated on
a discrete-time basis which is executable on a general-purpose quantum computer. The
analogy could be using a Analog-to-Digital converter (ADC) with limited resolution and
sampling rate to digitize continuous signals for further processing. Not only will this
provide new insights into such discrete memory operations, with a synthesis flow to a
target quantum hardware, we can implement such protocol on a real quantum computer.
The last piece of this simulation/emulation approach is a method of implementing
engineering dissipation which is given by the L operator in the SLH model. This will be
introduced in the next section.
6.1 Background 139
SystemEnv.
AncillaHL
√γσ−
HL = ω(σ−L† + σ+L)
EquivalentModel
L
Fig. 6.4 Engineered dissipation by direct coupling to an amplitude damping ancillarysystem.
6.1.3 Engineering Dissipation Control
At first glance, the input-output open quantum system approach can be easily imple-
mented on LIQUi|⟩ by a set of suitably designed Kraus super-operators. While being
ideal for simulation purposes, hardware synthesis to real physical systems will be
extremely challenging. An alternative approach could be implementing a universal
dissipative gadget which is just an amplitude-damped qubit as proposed by Verstraete
et al.[185] The damped ancillary qubit is directly coupled to the other qubits (such
as the memory qubits) to provide the desired dissipation. The engineering dissipation
configuration is illustrated in Figure 6.4.
The coupling Hamiltonian between the ancilla and the system has the form Hc =
ω(σ−L† +σ+L). In the limit that the damping rate, γ , of this ancilla is much greater
than the strength ω of the coupling Hamiltonian, one can adiabatically eliminate the
excited level of the ancilla, thus reconstructing the desired dissipation L for the system.
The upper bound for error of this adiabatic approximation is ∥ε∥∼ 1γ2 .
The coupling Hamiltonian can be easily transformed into a sequence of discrete
quantum gates executable on a quantum computer while there are many physical systems
that may act as a dissipative gadget, such as low-Q optical or superconducting cavities.
Therefore, a LIQUi|⟩ program written in this way is more likely to be ready for hardware
synthesis.
140 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
6.1.4 Trotter Decomposition
One key aspect of simulating quantum dynamics is the time evolution simulation, i.e.
applying exp(−iHt) to the quantum state. Unlike differential equation solvers which
directly solve for this time evolution, LIQUi|⟩ uses the quantum gate model where the
complex exponential term can be approximated by[195]:
exp(−iHt) =
(∏
kexp(−ihk∆t)
)t/∆t
+O(∆t), (6.7)
in which, the original Hamiltonian has the summation form H = ∑k hk and ∆t is the
Trotter step size.
The term exp(−ihk∆t) is in turn implemented by a sequence of quantum gates. This
is known as first-order Trotter decomposition which is often used in quantum chemistry
simulation[193].
For open quantum systems with engineered dissipation coupling, we can use the
“dissipative gadget” that was introduced in the previous section to adiabatically emulate
the environmental coupling. As shown in Eq. (6.6), the GEM Hamiltonian consists
of two classes of terms. The first one is the number operator term (a†a) which can
be implemented directly by a T (θ) gate. The coupling Hamiltonian Hc required for
engineering dissipation, Hc = ω(σ−L† +σ+L), has the form of the excitation operator
(hpq(a†paq +a†
qap)) which can be realized by the sequence of gates as shown in FIG. 5
(ii) in Wecker et al.[193]
Last but not least, the cascading term, i.e. β
2i ∑Nj=2 ∑
j−1k
(a†
jak −h.c)
, which is not
commonly seen in the quantum chemistry model, can also be decomposed in the same
manner, i.e.
a†jak −h.c. ≡ σ
( j)x − iσ ( j)
y
2σ(k)x + iσ (k)
y
2−h.c.
= iσ(k)y σ
( j)x −σ
(k)x σ
( j)y
2. (6.8)
The resulting sequence of quantum gates (rendered by LIQUi|⟩) which can be used to
implement complex exponential of the terms in Eq. (6.8) is shown in Figure 6.5 which
6.1 Background 141
H H Y Y †
Y θZ/2 Y † H −θZ/2 H
Fig. 6.5 Quantum circuit that represents the unitary generated by the cascading Hamilto-nian term.
involves change-of-basis gates (H and Y ) and the standard unitary exp[−i(θ/2)(σz⊗σz)].
The LIQUi|⟩ code snippet that is used to run this Trotter decomposed term is shown as
below.
//Trotterization of Cascading
Hamiltonian
let HamiltonianCasOp(theta:float, p:
int, q:int) (qs:Qubits) =
let a_qubit = qs.[p..]
let s_qubit = qs.[q..]
Ybasis s_qubit
H a_qubit
CNOT !! (qs,q,p)
Rz(-theta) 1.0 ’’’’ a_qubit
CNOT !! (qs,q,p)
YbasisAdj s_qubit
H a_qubit
H s_qubit
Ybasis a_qubit
CNOT !! (qs,q,p)
Rz(theta) 1.0 ’’’’ a_qubit
CNOT !! (qs,q,p)
H s_qubit
YbasisAdj a_qubit
142 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
Qubit 1Single-photonAncilla
Qubit N
Dissipation Ancilla
√γσ−
Fig. 6.6 Discretized GEM model driven by single-photon source used for LIQUi|⟩simulation.
6.2 Models and Results
6.2.1 Amplitude Damping Set-up
Using the dissipation gadget, the discretized gradient echo memory model used for
LIQUi|⟩ simulation is shown in Figure 6.6 below. The dissipation ancilla at the bottom
is a qubit which is subjected to general amplitude damping. This is done by running idle
gate at each Trotter step with the error probability of (1− exp(−γ∆t)). A LIQUi|⟩ code
snapshot for the amplitude damping set-up is also shown in Figure 6.7.
The LIQUi|⟩ simulation is performed on a Windows 7 desktop machine using Intel
Core i7-4770 3.4 GHz (Haswell) with 16 GB of main memory. In the case of running
4-qubit gradient echo memory model (6 qubits in total), the runtime for 1,000 trajectories
each with 10,000 Trotter steps is 3 hours.
6.2.2 Toy Example: Single-Atom Memory
Before delving into the discretized GEM simulation, we will test the LIQUi|⟩ set-up
in a much simpler situation, namely a single-atom memory. Here we consider a single
two-level atom driven by a Gaussian continuous-mode single photon as given by Eq.
(6.2). This problem has been investigated thoroughly by both conventional and quantum
network approaches.
The coupling operator of the atom is L =√
Γσ−. The atom is assumed to be free of
internal Hamiltonian dynamics and field scattering for simplicity, i.e., H = 0 and S = I .
Here Γ > 0 is the normalized coupling rate (Γ = 1).
6.2 Models and Results 143
//Probability of amplitude dampinglet probDamp_E = 1.0 - Math.Exp(-
gamma*dt)let noise = Noise(circ,ket,models)ket.TraceRun <- 0noise.LogGates <- falsenoise.TraceWrap <- falsenoise.TraceNoise <- falsenoise.DampProb(0) <- 0.0 //single-
photon ancillanoise.DampProb(1) <- 0.0 //qubit 1noise.DampProb(2) <- 0.0 //qubit 2...noise.DampProb(Nq+1) <- probDamp_E
//dissipation ancilla
Fig. 6.7 LIQUi|⟩ code snapshot for the amplitude damping set-up.
By using Trotter decomposition technique and LIQUi|⟩, we were able to simulate
the dynamics of this open quantum system. In our simulation, dissipation is induced by
an ancilla which undergoes amplitude damping. In Figure 6.8, we plot the “detection
time” histogram and the excited-state population for the case of Ω = 1.5Γ, which is
known to be the optimal Gaussian wave packet for atomic excitation. The detection time
is triggered by the state-collapse event of amplitude damping noise and histogram is
plotted from 10,000 LIQUi|⟩ runs.
This is consistent with theoretical analysis which shows that that the maximum
excitation probability of an atom driven by a Gaussian-shaped single photon is about
80%[189]. Since we are essentially running a stochastic simulation on the input-output
model, we have the benefit of being able to simulate the conditional dynamics and output
observable statistics (photon detection in this case) if we choose to observe the output.
By sweeping the Gaussian bandwidth (larger Ω means broader spectrum in frequency
domain and vice versa), we can get the maximum excitation (absorption) rate of our
single atom model, as shown in Figure 6.9 (black curve). There is a very narrow range
of bandwidth where the absorption peaks. The optical “depth” created by cascading
144 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
0 5 10 15 20
Time
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Fre
qu
ency
Single-photon pulse shape
0 5 10 15 20
Time
0.0
0.2
0.4
0.6
0.8
1.0
PE
Fig. 6.8 Histogram of photon detection events from a two-level atom driven by aGaussian-shape single photon state. Inset: the excited state population of the atom. Thephoton wave-packet is included for reference (red curve in main and dotted curve ininset). Trotter step is chosen to be 2π×10−4, and the number of trajectories (realizations)is 10,000.
absorbers of gradient echo memory can enhance the photon absorption as shown in the
below section.
6.2.3 Discretized Gradient Echo Memory
Using the full GEM model with signal generating filter as in Eq. (6.7) and Figure 6.6, we
can investigate the performance of the discrete memory in terms of photon absorption
(efficiency). Due to computing resource and runtime constraints, only models of two-
and four-atom GEM are considered in this report (4 and 6 qubits in total, respectively).
Despite minimal system size, the performance enhancement is significant as shown in
Figure 6.9.
Almost perfect absorption (100% excitation) can be achieved for some Gaussian
wave packet shapes by using just 4 atoms in series. More importantly, a monotonic
improvement over the full range of bandwidth is achieved by introducing more qubit
resources. Notice that in the case of the multi-atom memory, the single excitation from
6.2 Models and Results 145
0 2 4 6 8 10
Ω/Γ
0.0
0.2
0.4
0.6
0.8
1.0
Pmax
E
Single Ion
Two-ion GEM
Four-ion GEM
Fig. 6.9 Single photon absorption probability as a function of photon bandwidth. Thebaseline (black) is the single atom model which has the max absorption of 80% for avery narrow range of bandwidth. The more atoms involve, the better absorption and alsowider range of bandwidth can be achieved. Each data point is a LIQUi|⟩ run of 1,000trajectories. Solid line is polynomial fit.
the photon is distributed among all memory qubits in a Fourier-transformed (frequency)
basis. Therefore, the choice of gradient is of importance and can be optimized for a given
target input wave packet. Since we use a linear symmetric gradient and the number of
atoms are even (thus missing a zero point), the absorption rate has bi-modal behavior, as
is clearly seen in the four-atom case. If, as for continuous solid-state GEM, the gradient
is a smooth continuous function, the memory can effectively capture the full Fourier
spectrum of incoming photons.
Besides the fact that this discrete memory model can be implemented on a general
purpose quantum computer, one major advantage is that we can potentially extend the
range of gradient. Inhomogeneous broadening (by the electric field induced Stark shift,
for instance) has difficulty in achieving the bandwidth required for telecommunication
wavelength photon absorption. If implemented on a digital quantum computer, the
gradient term is controlled by the Trotter decomposed unitary thus can be changed freely
provided that the time step can be reduced for sufficient error bound.
146 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
This discretized memory model has close analogy to classical digital signal pro-
cessing technology. The gradient quantum memory program running on the a quantum
computer is equivalent to an Analog-to-Digital Converter (ADC), where (i) the Trotter
step serves as the sampling rate; (ii) the resolution is dictated by the number of qubits;
and (iii) the sampling method is frequency gradient mapping (classical ADC’s have a
wide variety of sampling methods such as integrating and sigma-delta.)
6.3 Conclusions and Future Work
6.3.1 Summary
In this chapter, we have adopted LIQUi|⟩ to simulate quantum open systems using the
input-output formalism. A combination of theoretical (Trotter decomposition, dissipation
gadgets) and practical (Hamiltonian simulator and the noise models available in LIQUi|⟩)solutions can provide accurate and consistent simulation results with other tool suites as
demonstrated in the “single atom driven by single photon” case.
Simulation-wise, the use of LIQUi|⟩ for this sort of numerical simulation struggles
to compete with more specialized direct differential solver based tools such as XMDS
or QuTiP due to (i) extra efforts needed to perform the Trotter decomposition and set-
up the dissipation ancilla; (ii) simulation errors associated with these approximations
(could be trade-off by additional runtime). However, if we look at the grand scheme of
quantum computation with reference to the classical computing paradigm, this approach
is essentially bridging the analog and digital worlds of the quantum framework. The
fact that LIQUi|⟩ is a quantum simulator which guarantees a direct “plug-and-play” on a
future digital quantum computer. Being able to simulate this system on LIQUi|⟩ shows
that we can readily emulate this sort of quantum dynamics on a general purpose quantum
computer. This is what we have shown in the discrete quantum memory example where
a small collection of qubits running synchronously to perform photon quantum state
storage. An initial dry run of this hypothetical memory demonstrates great improvement
in terms of the absorption rate even with an unoptimized gradient scheme.
Using this discretized method to demonstrate open quantum system approach, we
can get new insights about hardware requirements and potential errors associated with
emulating continuous “analog” quantum functions and inputs. This is somewhat missing
6.3 Conclusions and Future Work 147
in the current quantum computing picture where computing problems are often formu-
lated as an algorithm started on a given initial state without the need for external quantum
inputs. The SLH formalism provides a modular and systematic approach toward this
modeling problem thus worthwhile integrating to the LIQUi|⟩ framework, especially
when quantum hardware synthesis is taken into account since the SLH framework has
native models for a lot of quantum optics, opto-mechanical and cavity QED devices.
6.3.2 The Path Forward
In the discretized GEM simulation, we have introduced an analogy between a GEM
with a finite number of qubits and a classical A/D converter. There are some immediate
considerations that warrant further investigation such as error bound estimation regarding
the system size and Trotter step size; whether aliasing effect exists in this context; some
optimization techniques that we can do to improve the performance.
As this approach can be used to treat a large variety of quantum input-output dynam-
ical models besides the GEM, some future research avenues which can be considered
are listed below.
• Non-Markovian quantum noise model: Recent work[201] has shown that the
amplitude damping dissipation ancilla, when treated non-adiabatically, is equiva-
lent to a non-Markovian (colored) Lorentzian noise source. By combining multiple
ancillas, we can construct an arbitrary noise spectrum. We can thus use this model
to evaluate the performance of a quantum computer under influence of colored
noise, especially regarding the performance of quantum error correction codes.
The advantage of this dynamical approach is that an experimentally measured
noise spectrum can be fed directly into the model by adjusting the parameters
of the dissipation ancillas. For instance, quantum dot - cavity QED is a good
candidate, since the noise in this platform is intrinsically colored.
• Coherent digital quantum controller: Another direction is to look at the use of
a quantum computer in quantum control settings which usually involve continuous
input-output models. Especially in the case of coherent feedback control whereby
control objectives are achieved by field coupling quantum systems together. This
removes the measurement step and can perform control actions at the intrinsic
148 Chapter 6. Simulating Input-Output Quantum Systems with LIQUi|⟩
system speed (e.g. for quantum optics systems). A coherent digital controller can
perform equally well in this setting if equipped with appropriate quantum coherent
A/D and D/A converters at its input and output ports. By allowing quantum
programing, a coherent digital quantum controller can potentially achieve much
better performance than just wiring up quantum components (e.g. cavities, beam-
splitters, Kerr non-linear media.)[104]
• Quantum synthesis: As LIQUi|⟩ is part of a larger effort to create a full quantum
design flow from software to target hardware implementation, it is worthwhile
considering the integration of the SLH network model at the low-level physical
description. It is one of the most accurate and systematic dynamical descriptions
for quantum optics devices and circuits.
In comparison, a classical VLSI design flow always has a so-called “parasitic
extraction and analysis” step after “place and route” during which the full RLC
model of the entire circuit plus routing is compiled and simulated. Besides
simulations at the logic and transistor levels, it is deemed necessary to perform such
low level verification especially for complex and high speed designs. The approach
taken by classical integrated circuit design is by no mean the absolute solution
for quantum computer design. However, by incorporating this “analog” model
of quantum devices, we can definitely improve the versatility of the synthesis
process especially in cases where continuous time is the most natural description
(e.g. continuous variable quantum computation).
Given the “Big Hairy Audacious Goal” of creating a functional quantum computer in
the next one or two decades, a solid foundation for quantum system design is imperative.
And LIQUi|⟩ created by Microsoft is the first bold step in that direction. Augmenting
the quantum gate model of LIQUi|⟩ with the SLH dynamical model either by emulating
or intrinsic integration will open up new applications as well as cater to a larger research
community.
Chapter 7
Concluding Remarks
Any sufficiently advanced technology is indistinguishable from magic
Arthur C. Clark
Thanks to the first-wave of quantum innovation in the 20th century, we have achieved
unimaginable technologies, especially in the electronics industry. The second wave of
quantum revolution that we are witnessing is set to experience exponential growth in
a foreseeable future. Building a functional quantum computer which can outperform
any of our current computing devices and systems is one of the grand challenges of the
quantum technology sector.
Hence, the primary goal of this thesis is to study the scalability and engineering
aspects of quantum computing architectures. That includes the estimation of the overhead
incurred in quantum error correction, the potential use of advanced quantum control
techniques in migrating quantum errors as well as enhancing existing quantum error
correction protocols, and the development of quantum software simulation tools to aid
early quantum programmers.
7.1 Where We Stand
• In Chapter 2, we applied the classical electronics know-how regarding interconnect
routing to provide a concrete procedure for fan-out scalability analysis. The
150 Chapter 7. Concluding Remarks
results not only indicate the potential scalability bottleneck but, more importantly,
highlight the need for innovative solutions w.r.t. the scale-up of quantum circuits.
• In Chapter 3, we derived a continuous measurement feedback scheme that can be
used for topological quantum error codes. Quantum control modelling techniques,
namely the Quantum Stochastic Differential Equation (QSDE) and the (S, L,
H) model are demonstrated in designing the controller to tackle the problem of
correcting surface code qubit lattice automatically.
• In Chapter 4, we take a step back and look at the problem of quantum decoherence
at a fundamental/theoretical level. In particular, we demonstrated a pathway to
the synthesis of the so-called decoherence-free subsystem (DFS) via quantum
coherent feedback. Despite not being universal, the technique is nonetheless
applicable to a wide variety of passive quantum systems. Furthermore, we adopted
the Lyapunov formalism to the design and synthesis of passive quantum error
correction.
• In Chapter 5, we summarised the quantum programming suite which includes the
definition of a gate-based quantum assembly language, the design of a quantum
programming work-flow, and most importantly the implementation of a high-
performance quantum simulator. The framework that we derived bridges the gap
between computer science and quantum physics hence assists the development
and testing of quantum algorithms and application.
• In Chapter 6, we demonstrated that a digital quantum simulator, such as LIQUi|⟩is capable of simulating quantum open systems using the input-output formalism.
This bridges the gap between gate-based quantum computation and open quantum
system approaches.
7.2 The Way Forward
Quantum computing architecture is a rapidly-evolving field which takes advantages of
both the academic and industry research. For example, the surface-code quantum error
correction, which is currently state-of-the-art, could be surpassed by other codes. Also,
all the assumptions about the qubit devices in terms of geometry, operating conditions,
7.2 The Way Forward 151
and characteristics could dramatically change in the near future given the substantial
investment of technology companies around the world.
Hence, the study of quantum computing architecture needs to keep up with the rate of
development in the field. In particular, the fan-out scalability study in Chapter 2 has laid
the foundation for other types of scalability analysis where we could predict in advance
potential scaling issues and hence adjust the design to circumvent such limitation.
Control and coherence of large quantum systems remain one of the most challenging
aspects of quantum technologies. Quantum control engineering is still an emerging field,
yet very promising. Indeed, we can translate many quantum computing problems into
control problems as what we demonstrated in this thesis with the quantum error cor-
rection. Quantum control could provide a scalable approach, given its autonomous and
passive nature, toward the construction of large quantum networks which is crucial for
quantum computation. The problem of quantum decoherence suppression or elimination
is particularly well-suited to quantum control engineering.
Last but not least, the software stack will play a vital role in the roll out of any future
quantum computing platforms. At this stage, the lack of real quantum hardware is the
limiting factor in quantum application development. However, as the field progresses, the
need for an end-to-end quantum software development environment will become more
prevalent. This area, in particular, requires a strong collaboration between academic and
industry. The ultimate goal is to provide quantum software engineers with the tools they
need to explore and fully-utilise the capability of quantum computers.
This page intentionally left blank.
References
[QSh] Microsoft quantum development kit. https://docs.microsoft.com/en-us/quantum/?view=qsharp-preview. Accessed: 2019-04-27.
[2] Aasen, D., Hell, M., Mishmash, R. V., Higginbotham, A., Danon, J., Leijnse, M.,Jespersen, T. S., Folk, J. A., Marcus, C. M., Flensberg, K., et al. (2016). Milestonestoward majorana-based quantum computing. Physical Review X, 6(3):031016.
[3] Abhari, A. J., Faruque, A., Dousti, M. J., Svec, L., Catu, O., Chakrabati, A., Chiang,C.-F., Vanderwilt, S., Black, J., and Chong, F. (2012). Scaffold: Quantum program-ming language. Technical report, PRINCETON UNIV NJ DEPT OF COMPUTERSCIENCE.
[4] Ahn, C., Doherty, A. C., and Landahl, A. J. (2002). Continuous quantum errorcorrection via quantum feedback control. Physical Review A, 65(4):042301.
[5] Ahn, C., Wiseman, H. M., and Milburn, G. (2003). Quantum error correction forcontinuously detected errors. Physical Review A, 67(5):052310.
[6] Altafini, C. and Ticozzi, F. (2012). Modeling and control of quantum systems: anintroduction. IEEE Transactions on Automatic Control, 57(8):18981917.
[7] Amini, H., Somaraju, R. A., Dotsenko, I., Sayrin, C., Mirrahimi, M., and Rou-chon, P. (2013). Feedback stabilization of discrete-time quantum systems sub-ject to non-demolition measurements with imperfections and delays. Automatica,49(9):26832692.
[8] Aspuru-Guzik, A., Dutoi, A. D., Love, P. J., and Head-Gordon, M. (2005). Simulatedquantum computation of molecular energies. Science, 309(5741):17041707.
[9] Auth, C., Allen, C., Blattner, A., Bergstrom, D., Brazier, M., Bost, M., Buehler, M.,Chikarmane, V., Ghani, T., Glassman, T., et al. (2012). A 22nm high performanceand low-power cmos technology featuring fully-depleted tri-gate transistors, self-aligned contacts and high density mim capacitors. In VLSI Technology (VLSIT), 2012Symposium on, page 131132. IEEE.
[10] Awschalom, D. D., Bassett, L. C., Dzurak, A. S., Hu, E. L., and Petta, J. R.(2013). Quantum spintronics: engineering and manipulating atom-like spins insemiconductors. Science, 339(6124):11741179.
[11] Baart, T., Jovanovic, N., Reichl, C., Wegscheider, W., and Vandersypen, L.(2016a). Nanosecond-timescale spin transfer using individual electrons in a quadruple-quantum-dot device. Applied Physics Letters, 109(4):043101.
154 REFERENCES
[12] Baart, T. A., Shafiei, M., Fujita, T., Reichl, C., Wegscheider, W., and Vandersypen,L. M. K. (2016b). Single-spin ccd. Nature nanotechnology.
[13] Bao, T., Yakimets, D., Ryckaert, J., Ciofi, I., Baert, R., Veloso, A., Boemmels,J., Collaert, N., Roussel, P., Demuynck, S., Raghavan, P., Mercha, A., Tokei, Z.,Verkest, D., Thean, A.-Y., and Wambacq, P. (2014). Circuit and process co-designwith vertical gate-all-around nanowire fet technology to extend cmos scaling for 5nmand beyond technologies. In Solid State Device Research Conference (ESSDERC),2014 44th European, page 102105.
[14] Barnes, J. P. and Warren, W. S. (2000). Automatic quantum error correction. Phys.Rev. Lett., 85:856859.
[15] Barrett, S. D. and Stace, T. M. (2010). Fault tolerant quantum computation withvery high threshold for loss errors. Physical review letters, 105(20):200502.
[16] Bellman, R. (1962). Vector lyanpunov functions. SIAM Journal on Control andOptimization, 1(1):32.
[17] Bose, S. (2003). Quantum communication through an unmodulated spin chain.Phys. Rev. Lett., 91:207901.
[18] Bouten, L., Van Handel, R., and James, M. R. (2007). An introduction to quantumfiltering. SIAM Journal on Control and Optimization, 46(6):21992241.
[19] Breuer, H.-P. and Petruccione, F. (2002). The theory of open quantum systems.Oxford University Press on Demand.
[20] Castelvecchi, D. et al. (2017). Quantum cloud goes commercial. Nature,543(7644):159159.
[21] Chase, B. A., Landahl, A. J., and Geremia, J. (2008). Efficient feedback controllersfor continuous-time quantum error correction. Phys. Rev. A, 77:032304.
[22] Chen, C.-T. (1973). A generalization of the inertia theorem. SIAM Journal onApplied Mathematics, 25(2):158161.
[23] Christandl, M., Datta, N., Ekert, A., and Landahl, A. J. (2004). Perfect state transferin quantum spin networks. Phys. Rev. Lett., 92:187902.
[24] Ciliberto, C., Herbster, M., Ialongo, A. D., Pontil, M., Rocchetto, A., Severini, S.,and Wossnig, L. (2018). Quantum machine learning: a classical perspective. Pro-ceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences,474(2209):20170551.
[25] Coecke, B. and Duncan, R. (2011). Interacting quantum observables: categoricalalgebra and diagrammatics. New Journal of Physics, 13(4):043016.
[26] Cohen, J. and Mirrahimi, M. (2014). Dissipation-induced continuous quantumerror correction for superconducting circuits. Phys. Rev. A, 90:062344.
REFERENCES 155
[27] Colless, J. I., Mahoney, A. C., Hornibrook, J. M., Doherty, A. C., Lu, H., Gossard,A. C., and Reilly, D. J. (2013). Dispersive readout of a few-electron double quantumdot with fast rf gate sensors. Phys. Rev. Lett., 110:046805.
[28] Copsey, D., Oskin, M., Impens, F., Metodiev, T., Cross, A., Chong, F. T., Chuang,I. L., and Kubiatowicz, J. (2003). Toward a scalable, silicon-based quantum com-puting architecture. Selected Topics in Quantum Electronics, IEEE Journal of,9(6):15521569.
[29] Deng, G.-W., Wei, D., Li, S.-X., Johansson, J., Kong, W.-C., Li, H.-O., Cao, G.,Xiao, M., Guo, G.-C., Nori, F., et al. (2015). Coupling two distant double quantumdots with a microwave resonator. Nano letters, 15(10):66206625.
[30] Dennis, E., Kitaev, A., Landahl, A., and Preskill, J. (2002). Topological quantummemory. Journal of Mathematical Physics, 43(9):44524505.
[31] Dennis, G. R., Hope, J. J., and Johnsson, M. T. (2013). Xmds2: Fast, scalablesimulation of coupled stochastic partial differential equations. Comput Phys Commun,184(1):201208.
[32] Devitt, S. J., Fowler, A. G., Stephens, A. M., Greentree, A. D., Hollenberg, L. C.,Munro, W. J., and Nemoto, K. (2009). Architectural design for a topological clusterstate quantum computer. New Journal of Physics, 11(8):083032.
[33] Devoret, M. H. and Schoelkopf, R. J. (2013). Superconducting circuits for quantuminformation: an outlook. Science, 339(6124):11691174.
[34] Divincenzo, D. P. (2000). The physical implementation of quantum computation.Fortschr. Phys, 48:2000.
[35] DiVincenzo, D. P. (2009). Fault-tolerant architectures for superconducting qubits.Physica Scripta, 2009(T137):014020.
[36] Dong, D. and Petersen, I. R. (2010). Quantum control theory and applications: asurvey. IET Control Theory & Applications, 4(12):26512671.
[37] Dowling, J. P. and Milburn, G. J. (2003). Quantum technology: the secondquantum revolution. Philosophical Transactions of the Royal Society of London A:Mathematical, Physical and Engineering Sciences, 361(1809):16551674.
[38] Dreher, L., Hoehne, F., Stutzmann, M., and Brandt, M. S. (2012). Nuclear spins ofionized phosphorus donors in silicon. Phys. Rev. Lett., 108:027602.
[39] Dunjko, V. and Briegel, H. J. (2018). Machine learning & artificial intelligence inthe quantum domain: a review of recent progress. Reports on Progress in Physics,81(7):074001.
[40] Dzurak, A. S. (2016). Spin-based quantum computing in silicon CMOS-compatibleplatforms. In 2016 IEEE International Electron Devices Meeting (IEDM), page13.2.113.2.2.
156 REFERENCES
[41] Fleischhauer, M. and Lukin, M. D. (2000). Dark-state polaritons in electromagneti-cally induced transparency. Phys. Rev. Lett., 84:50945097.
[42] Fowler, A. G. (2013). Coping with qubit leakage in topological codes. Phys. Rev.A, 88:042308.
[43] Fowler, A. G., Mariantoni, M., Martinis, J. M., and Cleland, A. N. (2012a). Surfacecodes: Towards practical large-scale quantum computation. Phys. Rev. A, 86:032324.
[44] Fowler, A. G. and Martinis, J. M. (2014). Quantifying the effects of localmany-qubit errors and nonlocal two-qubit errors on the surface code. Phys. Rev.A, 89:032316.
[45] Fowler, A. G., Stephens, A. M., and Groszkowski, P. (2009). High-thresholduniversal quantum computation on the surface code. Phys. Rev. A, 80:052312.
[46] Fowler, A. G., Whiteside, A. C., and Hollenberg, L. C. L. (2012b). Towardspractical classical processing for the surface code. Phys. Rev. Lett., 108:180501.
[47] Fuechsle, M., Miwa, J. A., Mahapatra, S., Ryu, H., Lee, S., Warschkow, O.,Hollenberg, L. C., Klimeck, G., and Simmons, M. Y. (2012). A single-atom transistor.Nature Nanotechnology, 7(4):242246.
[48] Fuhrer, A., Fuchsle, M., Reusch, T., Weber, B., and Simmons, M. (2009). Atomic-scale, all epitaxial in-plane gated donor quantum dot in silicon. Nano letters,9(2):707710.
[49] Gardiner, C. and Zoller, P. (2004). Quantum noise: a handbook of Markovian andnon-Markovian quantum stochastic methods with applications to quantum optics,volume 56. Springer Science & Business Media.
[50] Gardiner, C. W. and Collett, M. J. (1985). Input and output in damped quantumsystems: Quantum stochastic differential equations and the master equation. Phys.Rev. A, 31:37613774.
[51] Gonzalez-Zalba, M., Barraud, S., Ferguson, A., and Betz, A. (2015). Probing thelimits of gate-based charge sensing. Nature communications, 6.
[52] Gottesman, D. (1996). Class of quantum error-correcting codes saturating thequantum hamming bound. Phys. Rev. A, 54:18621868.
[53] Gottesman, D. (1997). Stabilizer codes and quantum error correction. arXivpreprint quant-ph/9705052.
[54] Gottesman, D. (2010). An introduction to quantum error correction and fault-tolerant quantum computation. In Quantum information science and its contributionsto mathematics, Proceedings of Symposia in Applied Mathematics, volume 68, page1358.
[55] Gough, J. and James, M. (2009a). Quantum feedback networks: Hamiltonianformulation. Communications in Mathematical Physics, 287(3):11091132.
REFERENCES 157
[56] Gough, J. and James, M. R. (2009b). The series product and its application toquantum feedforward and feedback networks. Automatic Control, IEEE Transactionson, 54(11):25302544.
[57] Gough, J. E. and James, M. R. (2017). The series product for gaussian quantuminput processes. Reports on Mathematical Physics, 79(1):111133.
[58] Gough, J. E., James, M. R., Nurdin, H. I., and Combes, J. (2012). Quantum filteringfor systems driven by fields in single-photon states or superposition of coherent states.Phys. Rev. A, 86:043819.
[59] Gough, J. E. and Zhang, G. (2015). On realization theory of quantum linearsystems. Automatica, 59:139151.
[60] Green, A. S., Lumsdaine, P. L., Ross, N. J., Selinger, P., and Valiron, B. (2013).Quipper: a scalable quantum programming language. In ACM SIGPLAN Notices,volume 48, page 333342. ACM.
[61] Greentree, A. D., Cole, J. H., Hamilton, A. R., and Hollenberg, L. C. L. (2004).Coherent electronic transfer in quantum dot systems using adiabatic passage. Phys.Rev. B, 70:235317.
[62] Grover, L. K. (1996). A fast quantum mechanical algorithm for database search.In Proceedings of the twenty-eighth annual ACM symposium on Theory of computing,page 212219. ACM.
[63] Haddad, W. M. and Chellaboina, V. (2008). Nonlinear dynamical systems andcontrol: a Lyapunov-based approach. Princeton University Press.
[64] Hanson, R., Kouwenhoven, L. P., Petta, J. R., Tarucha, S., and Vandersypen, L.M. K. (2007). Spins in few-electron quantum dots. Rev. Mod. Phys., 79:12171265.
[65] Harrow, A. W., Hassidim, A., and Lloyd, S. (2009). Quantum algorithm for linearsystems of equations. Phys. Rev. Lett., 103:150502.
[66] Hein, M., Eisert, J., and Briegel, H. J. (2004). Multiparty entanglement in graphstates. Phys. Rev. A, 69:062311.
[67] Hempel, C., Maier, C., Romero, J., McClean, J., Monz, T., Shen, H., Jurcevic, P.,Lanyon, B. P., Love, P., Babbush, R., Aspuru-Guzik, A., Blatt, R., and Roos, C. F.(2018). Quantum chemistry calculations on a trapped-ion quantum simulator. Phys.Rev. X, 8:031022.
[68] Hermelin, S., Takada, S., Yamamoto, M., Tarucha, S., Wieck, A. D., Saminadayar,L., Brle, C., and Meunier, T. (2011). Electrons surfing on a sound wave as a platformfor quantum optics with flying electrons. Nature, 477(7365):435438.
[69] Herrera-Mart. A., Fowler, A. G., Jennings, D., and Rudolph, T. (2010). Photonicimplementation for the topological cluster-state quantum computer. Phys. Rev. A,82:032332.
158 REFERENCES
[70] Hill, C. D., Hollenberg, L. C. L., Fowler, A. G., Wellard, C. J., Greentree, A. D.,and Goan, H.-S. (2005). Global control and fast solid-state donor electron spinquantum computing. Phys. Rev. B, 72:045350.
[71] Hill, C. D., Peretz, E., Hile, S. J., House, M. G., Fuechsle, M., Rogge, S., Simmons,M. Y., and Hollenberg, L. C. L. (2015). A surface code quantum computer in silicon.Science Advances, 1(9).
[72] Hoefflinger, B. (2012). Chips 2020: a guide to the future of nanoelectronics.Springer.
[73] Hollenberg, L., Greentree, A., Fowler, A., and Wellard, C. (2006). Two-dimensionalarchitectures for donor-based quantum computing. Physical Review B, 74(4):045311.
[74] Hornibrook, J., Colless, J., Mahoney, A., Croot, X., Blanvillain, S., Lu, H., Gossard,A., and Reilly, D. (2014). Frequency multiplexing for readout of spin qubits. AppliedPhysics Letters, 104(10):103108.
[75] Horsman, C., Fowler, A. G., Devitt, S., and Meter, R. V. (2012). Surface codequantum computing by lattice surgery. New Journal of Physics, 14(12):123011.
[76] Hou, S. C., Khan, M. A., Yi, X. X., Dong, D., and Petersen, I. R. (2012). Optimallyapunov-based quantum control for quantum systems. Phys. Rev. A, 86:022321.
[77] House, M., Kobayashi, T., Weber, B., Hile, S., Watson, T., van der Heijden,J., Rogge, S., and Simmons, M. (2015). Radio frequency measurements of tunnelcouplings and singlet-triplet spin states in Si: P quantum dots. Nature communications,6.
[78] Hsu, K.-C. and Brun, T. A. (2016). Method for quantum-jump continuous-timequantum error correction. Phys. Rev. A, 93:022321.
[79] Hudson, R. L. and Parthasarathy, K. R. (1984). Quantum ito’s formula and stochas-tic evolutions. Communications in Mathematical Physics, 93(3):301323.
[80] Hush, M., Carvalho, A., Hedges, M., and James, M. (2013). Analysis of theoperation of gradient echo memories using a quantum input-output model. New J.Phys, 15(8):085020.
[81] Hr, T., Steiger, D. S., Smelyanskiy, M., and Troyer, M. (2016). High performanceemulation of quantum circuits. In Proceedings of the International Conference forHigh Performance Computing, Networking, Storage and Analysis, page 74. IEEEPress.
[82] Ippoliti, M., Mazza, L., Rizzi, M., and Giovannetti, V. (2015). Perturbativeapproach to continuous-time quantum error correction. Phys. Rev. A, 91:042322.
[83] James, M. R. and Gough, J. (2010). Quantum dissipative systems and feed-back control design by interconnection. IEEE Transactions on Automatic Control,55(8):18061821.
REFERENCES 159
[84] James, M. R. and Nurdin, H. I. (2015). A tutorial introduction to quantum feedbackcontrol. In 2015 IEEE Conference on Control Applications (CCA), page 112.
[85] Johansson, J., Nation, P., and Nori, F. (2012). Qutip: An open-source pythonframework for the dynamics of open quantum systems. Comput Phys Commun,183(8):17601772.
[86] Jones, N. C., Van Meter, R., Fowler, A. G., McMahon, P. L., Kim, J., Ladd, T. D.,and Yamamoto, Y. (2012). Layered architecture for quantum computing. Phys. Rev.X, 2:031007.
[87] Kane, B. E. (1998). A silicon-based nuclear spin quantum computer. Nature,393(6681):133137.
[88] Kapit, E., Chalker, J. T., and Simon, S. H. (2015). Passive correction of quantumlogical errors in a driven, dissipative system: A blueprint for an analog quantum codefabric. Phys. Rev. A, 91:062324.
[89] KawakamiE, ScarlinoP, Ward, D. R., Braakman, F. R., Savage, D. E., Lagally,M. G., Friesen, M., Coppersmith, S. N., Eriksson, M. A., and Vandersypen, L. M. K.(2014). Electrical control of a long-lived spin qubit in a Si/SiGe quantum dot. Naturenanotechnology, 9(9):666670.
[90] Kerckhoff, J., Nurdin, H. I., Pavlichin, D. S., and Mabuchi, H. (2010). Design-ing quantum memories with embedded control: Photonic circuits for autonomousquantum error correction. Phys. Rev. Lett., 105:040502.
[91] Kerckhoff, J., Pavlichin, D. S., Chalabi, H., and Mabuchi, H. (2011). Design ofnanophotonic circuits for autonomous subsystem quantum error correction. NewJournal of Physics, 13(5):055022.
[92] Kielpinski, D., Monroe, C., and Wineland, D. J. (2002). Architecture for a large-scale ion-trap quantum computer. Nature, 417(6890):709.
[93] Kienzler, D., Lo, H.-Y., Keitch, B., de Clercq, L., Leupold, F., Lindenfelser, F.,Marinelli, M., Negnevitsky, V., and Home, J. (2015). Quantum harmonic oscillatorstate synthesis by reservoir engineering. Science, 347(6217):5356.
[94] Kitaev, A. Y. (2003). Fault-tolerant quantum computation by anyons. Annals ofPhysics, 303(1):230.
[95] Knill, E., Laflamme, R., and Viola, L. (2000). Theory of quantum error correctionfor general noise. Phys. Rev. Lett., 84:25252528.
[96] Kraus, K., Bhm, A., Dollard, J., and Wootters, W. (1983). States, effects, andoperations fundamental notions of quantum theory. In States, Effects, and OperationsFundamental Notions of Quantum Theory, volume 190.
[97] Krauter, H., Muschik, C. A., Jensen, K., Wasilewski, W., Petersen, J. M., Cirac,J. I., and Polzik, E. S. (2011). Entanglement generated by dissipation and steady stateentanglement of two macroscopic objects. Physical review letters, 107(8):080503.
160 REFERENCES
[98] Kuang, S. and Cong, S. (2008). Lyapunov control methods of closed quantumsystems. Automatica, 44(1):98108.
[99] Lidar, D. A. and Brun, T. A. (2013). Quantum error correction. CambridgeUniversity Press.
[100] Liu, X. and Kubiatowicz, J. (2013). Chisel-q: Designing quantum circuits with ascala embedded language. In Computer Design (ICCD), 2013 IEEE 31st InternationalConference on, page 427434. IEEE.
[101] Loss, D. and DiVincenzo, D. P. (1998). Quantum computation with quantum dots.Physical Review A, 57(1):120.
[102] Maalouf, A. I. and Petersen, I. R. (2011a). Bounded real properties for a classof annihilation-operator linear quantum systems. IEEE Transactions on AutomaticControl, 56(4):786801.
[103] Maalouf, A. I. and Petersen, I. R. (2011b). Coherent h∞ control for a class ofannihilation operator linear quantum systems. IEEE Transactions on AutomaticControl, 56(2):309319.
[104] Mabuchi, H. (2008). Coherent-feedback quantum control with a dynamic com-pensator. Phys. Rev. A, 78:032323.
[105] Mabuchi, H. (2009a). Cavity-qed models of switches for attojoule-scale nanopho-tonic logic. Phys. Rev. A, 80:045802.
[106] Mabuchi, H. (2009b). Continuous quantum error correction as classical hybridcontrol. New Journal of Physics, 11(10):105044.
[107] Mabuchi, H. (2011). Nonlinear interferometry approach to photonic sequentiallogic. Appl. Phys. Lett., 99(15):153103.
[108] Mariantoni, M., Wang, H., Yamamoto, T., Neeley, M., Bialczak, R. C., Chen,Y., Lenander, M., Lucero, E., O’connell, A., Sank, D., et al. (2011). Implementingthe quantum von neumann architecture with superconducting circuits. Science,334(6052):6165.
[109] McClean, J. R., Kivlichan, I. D., Sung, K. J., Steiger, D. S., Cao, Y., Dai, C., Fried,E. S., Gidney, C., Gimby, B., Gokhale, P., et al. (2017). Openfermion: the electronicstructure package for quantum computers. arXiv preprint arXiv:1710.07629.
[110] McNeil, R., Kataoka, M., Ford, C., Barnes, C., Anderson, D., Jones, G., Farrer, I.,and Ritchie, D. (2011). On-demand single-electron transfer between distant quantumdots. Nature, 477(7365):439442.
[111] Menchon-Enrich, R., Benseny, A., Ahufinger, V., Greentree, A. D., Busch, T., andMompart, J. (2016). Spatial adiabatic passage: a review of recent progress. Reportson Progress in Physics, 79(7):074401.
[112] Mirrahimi, M. and Van Handel, R. (2007). Stabilizing feedback controls forquantum systems. SIAM Journal on Control and Optimization, 46(2):445467.
REFERENCES 161
[113] Moessenboeck, H. (1990). Coco/R: A Generator for Fast Compiler Front-Ends.Citeseer.
[114] Monroe, C. and Kim, J. (2013). Scaling the ion trap quantum processor. Science,339(6124):11641169.
[115] Morello, A., Dzurak, A., Huebl, H.-G., Clark, R. G., Van Beveren, L. H. W.,Hollenberg, L. C. L., Jamieson, D. N., and Escott, C. (2013). Control and readout ofelectron or hole spin. US Patent 8,507,894.
[116] Muhonen, J. T., Dehollain, J. P., Laucht, A., Hudson, F. E., Kalra, R., Sekiguchi,T., Itoh, K. M., Jamieson, D. N., McCallum, J. C., Dzurak, A. S., and Morello, A.(2014). Storing quantum information for 30 seconds in a nanoelectronic device.Nature nanotechnology, 9(12):986991.
[117] Nadj-Perge, S., Drozdov, I. K., Li, J., Chen, H., Jeon, S., Seo, J., MacDonald,A. H., Bernevig, B. A., and Yazdani, A. (2014). Observation of majorana fermions inferromagnetic atomic chains on a superconductor. Science, 346(6209):602607.
[118] Nagayama, S., Fowler, A. G., Horsman, D., Devitt, S. J., and Van Meter, R.(2017). Surface code error correction on a defective lattice. New Journal of Physics,19(2):023050.
[119] Natarajan, S., Agostinelli, M., Akbar, S., Bost, M., Bowonder, A., Chikarmane,V., Chouksey, S., Dasgupta, A., Fischer, K., Fu, Q., et al. (2014). A 14nm logictechnology featuring 2nd-generation finfet, air-gapped interconnects, self-aligneddouble patterning and a 0.0588 µm2 SRAM cell size. In Electron Devices Meeting(IEDM), 2014 IEEE International, page 37. IEEE.
[120] Nguyen, T., Miao, Z., Pan, Y., Amini, N., Ugrinovskii, V., and James, M. R.(2017). Pole placement approach to coherent passive reservoir engineering for storingquantum information. In American Control Conference (ACC), 2017, page 234239.IEEE.
[121] Nielsen, M. A. (2004). Optical quantum computation using cluster states. Phys.Rev. Lett., 93:040503.
[122] Nielsen, M. A. and Chuang, I. L. (2010). Quantum computation and quantuminformation. Cambridge university press.
[123] Nigg, S. E., Fuhrer, A., and Loss, D. (2017). Superconducting grid-bus surfacecode architecture for hole-spin qubits. Phys. Rev. Lett., 118:147701.
[124] Nowack, K., Shafiei, M., Laforest, M., Prawiroatmodjo, G., Schreiber, L., Reichl,C., Wegscheider, W., and Vandersypen, L. (2011). Single-shot correlations andtwo-qubit gate of solid-state spins. Science, 333(6047):12691272.
[125] Nurdin, H. I. (2014). Quantum filtering for multiple input multiple output systemsdriven by arbitrary zero-mean jointly gaussian input fields. Russian Journal ofMathematical Physics, 21(3):386398.
162 REFERENCES
[126] Nurdin, H. I. and Gough, J. E. (2015). Modular quantum memories using passivelinear optics and coherent feedback. Quantum Information & Computation, 15(11-12):10171040.
[127] Nurdin, H. I., James, M. R., and Doherty, A. C. (2009a). Network synthesis of lin-ear dynamical quantum stochastic systems. SIAM J Control Optim, 48(4):26862718.
[128] Nurdin, H. I., James, M. R., and Petersen, I. R. (2009b). Coherent quantum lqgcontrol. Automatica, 45(8):18371846.
[129] Nurdin, H. I. and Yamamoto, N. (2017). Linear Dynamical Quantum Systems.Springer International Publishing.
[130] OpenMP (2018). Openmp. http://www.openmp.org/.
[131] Oreshkov, O. and Brun, T. A. (2007). Continuous quantum error correction fornon-markovian decoherence. Phys. Rev. A, 76:022318.
[132] Ors, R., Mugel, S., and Lizaso, E. (2019). Quantum computing for finance:overview and prospects. Reviews in Physics, page 100028.
[133] Oskin, M., Chong, F., and Chuang, I. (2002). Overhead reduction in a architecturefor quantum computers. IEEE Computer, 35(1):7987.
[134] Ostrowski, A. and Schneider, H. (1962). Some theorems on the inertia of generalmatrices. Journal of Mathematical Analysis and Applications, 4(1):7284.
[135] Paetznick, A. and Fowler, A. G. (2013). Quantum circuit optimization by topo-logical compaction in the surface code. arXiv preprint arXiv:1304.2807.
[136] Pan, Y., Amini, H., Miao, Z., Gough, J., Ugrinovskii, V., and James, M. R. (2014).Heisenberg picture approach to the stability of quantum markov systems. Journal ofMathematical Physics, 55(6):062701.
[137] Pan, Y. and Nguyen, T. (2017). Stabilizing quantum states and automatic er-ror correction by dissipation control. IEEE Transactions on Automatic Control,62(9):46254630.
[138] Pan, Y., Nguyen, T., Miao, Z., Amini, N., Ugrinovski, V., and James, M. (2016).Coherent observer engineering for protecting quantum information. In ControlConference (CCC), 2016 35th Chinese, page 91399144. IEEE.
[139] Pan, Y., Ugrinovskii, V., and James, M. R. (2015). Ground-state stabilization ofopen quantum systems by dissipation. arXiv preprint arXiv:1502.05593.
[140] Park, S.-J., Persaud, A., Liddle, J., Nilsson, J., Bokor, J., Schneider, D., Rangelow,I., and Schenkel, T. (2004). Processing issues in topdown approaches to quantumcomputer development in silicon. Microelectronic engineering, 73:695700.
[141] Parthasarathy, K. R. (2012). An introduction to quantum stochastic calculus,volume 85. Birkher.
REFERENCES 163
[142] Paz, J. P. and Zurek, W. H. (1998). Continuous error correction. Proceedingsof the Royal Society of London. Series A: Mathematical, Physical and EngineeringSciences, 454(1969):355364.
[143] Perez-Garcia, D., Verstraete, F., Wolf, M. M., and Cirac, J. I. (2008). Peps asunique ground states of local hamiltonians. Quantum Information & Computation,8(6):650663.
[144] Pla, J. J., Tan, K. Y., Dehollain, J. P., Lim, W. H., Morton, J. J., Jamieson, D. N.,Dzurak, A. S., and Morello, A. (2012). A single-atom electron spin qubit in silicon.Nature, 489(7417):541545.
[145] Poyatos, J. F., Cirac, J. I., and Zoller, P. (1996). Quantum reservoir engineeringwith laser cooled trapped ions. Phys. Rev. Lett., 77:47284731.
[146] Raussendorf, R. and Briegel, H. J. (2001). A one-way quantum computer. Phys.Rev. Lett., 86:51885191.
[147] Raussendorf, R. and Harrington, J. (2007). Fault-tolerant quantum computationwith high threshold in two dimensions. Phys. Rev. Lett., 98:190504.
[148] Raussendorf, R., Harrington, J., and Goyal, K. (2007). Topological fault-tolerancein cluster state quantum computation. New Journal of Physics, 9(6):199.
[149] Rossi, A., Zhao, R., Dzurak, A., and Gonzalez-Zalba, M. (2016). Dispersivereadout of a silicon quantum dot with accumulation-mode gate sensor. arXiv preprintarXiv:1610.00767.
[150] Saeedi, K., Simmons, S., Salvail, J. Z., Dluhy, P., Riemann, H., Abrosimov, N. V.,Becker, P., Pohl, H.-J., Morton, J. J., and Thewalt, M. L. (2013). Room-temperaturequantum bit storage exceeding 39 minutes using ionized donors in silicon-28. Science,342(6160):830833.
[151] Sarma, G. and Mabuchi, H. (2013). Gauge subsystems, separability and robustnessin autonomous quantum memories. New Journal of Physics, 15(3):035014.
[152] Sarovar, M., Ahn, C., Jacobs, K., and Milburn, G. J. (2004). Practical scheme forerror control using feedback. Physical Review A, 69(5):052324.
[153] Sarovar, M. and Milburn, G. (2005). Continuous quantum error correction bycooling. Physical Review A, 72(1):012306.
[154] Schirmer, S. G. and Wang, X. (2010). Stabilizing open quantum systems bymarkovian reservoir engineering. Physical Review A, 81(6):062306.
[155] Schofield, S. R., Curson, N. J., Simmons, M. Y., Rue F. J., Hallam, T., Oberbeck,L., and Clark, R. G. (2003). Atomically precise placement of single dopants in Si.Phys. Rev. Lett., 91:136104.
[156] Shankar, S., Hatridge, M., Leghtas, Z., Sliwa, K., Narla, A., Vool, U., Girvin,S. M., Frunzio, L., Mirrahimi, M., and Devoret, M. H. (2013). Autonomouslystabilized entanglement between two superconducting quantum bits. Nature,504(7480):419422.
164 REFERENCES
[157] Shor, P. W. (1994). Algorithms for quantum computation: discrete logarithmsand factoring. In Proceedings 35th Annual Symposium on Foundations of ComputerScience, page 124134.
[158] Shor, P. W. (1997). Polynomial-time algorithms for prime factorizationand discrete logarithms on a quantum computer. SIAM Journal on Computing,26(5):14841509.
[159] Song, H., Zhang, G., and Xi, Z. (2016). Continuous-mode multiphoton filtering.SIAM Journal on Control and Optimization, 54(3):16021632.
[160] Spohn, H. (1976). Approach to equilibrium for completely positive dynamicalsemigroups of n-level systems. Reports on Mathematical Physics, 10(2):189194.
[161] Stace, T. M. and Barrett, S. D. (2010). Error correction and degeneracy in surfacecodes suffering loss. Phys. Rev. A, 81:022317.
[162] Stephens, A. M. (2014). Fault-tolerant thresholds for quantum error correctionwith the surface code. Phys. Rev. A, 89:022321.
[163] Stotz, J. A., Hey, R., Santos, P. V., and Ploog, K. H. (2005). Coherent spintransport through dynamic quantum dots. Nature materials, 4(8):585588.
[164] Tang, Y.-C. and Miao, G.-X. (2016). Robust surface code topology againstsparse fabrication defects in a superconducting-qubit array. Physical Review A,93(3):032322.
[165] Technologies, R. B. (2018). Quantum gate language (qgl). https://bbn-q.github.io/QGL/.
[166] Terhal, B. M. (2013). Quantum error correction for quantum memories. arXivpreprint arXiv:1302.3428.
[167] Terhal, B. M. (2015). Quantum error correction for quantum memories. Rev. Mod.Phys., 87:307346.
[168] Tezak, N., Niederberger, A., Pavlichin, D. S., Sarma, G., and Mabuchi, H. (2012).Specification of photonic circuits using quantum hardware description language.Philosophical Transactions of the Royal Society of London A: Mathematical, Physicaland Engineering Sciences, 370(1979):52705290.
[169] Ticozzi, F., Lucchese, R., Cappellaro, P., and Viola, L. (2012). Hamiltoniancontrol of quantum dynamical semigroups: Stabilization and convergence speed.Automatic Control, IEEE Transactions on, 57(8):19311944.
[170] Ticozzi, F., Schirmer, S. G., and Wang, X. (2010). Stabilizing quantum states byconstructive design of open quantum dynamics. IEEE Transactions on AutomaticControl, 55(12):29012905.
[171] Ticozzi, F. and Viola, L. (2009). Analysis and synthesis of attractive quantummarkovian dynamics. Automatica, 45(9):20022009.
REFERENCES 165
[172] Ticozzi, F. and Viola, L. (2012). Stabilizing entangled states with quasi-localquantum dynamical semigroups. Phil. Trans. R. Soc. A, 370(1979):52595269.
[173] Ticozzi, F. and Viola, L. (2014). Steady-state entanglement by engineeredquasi-local markovian dissipation: Hamiltonian-assisted and conditional stabilization.Quantum Information & Computation, 14(3-4):265294.
[174] Tomita, Y. and Svore, K. M. (2014). Low-distance surface codes under realisticquantum noise. Phys. Rev. A, 90:062320.
[175] Tosi, G., Mohiyaddin, F. A., Tenberg, S. B., Rahman, R., Klimeck, G., andMorello, A. (2015). Silicon quantum processor with robust long-distance qubitcouplings. arXiv preprint arXiv:1509.08538.
[176] Trifunovic, L., Dial, O., Trif, M., Wootton, J. R., Abebe, R., Yacoby, A., andLoss, D. (2012). Long-distance spin-spin coupling via floating gates. Phys. Rev. X,2:011006.
[177] van der Wiel, W. G., De Franceschi, S., Elzerman, J. M., Fujisawa, T., Tarucha,S., and Kouwenhoven, L. P. (2002). Electron transport through double quantum dots.Rev. Mod. Phys., 75:122.
[178] Van Donkelaar, J. A., Greentree, A. D., Alves, A. D., Jong, L. M., Hollenberg,L. C., and Jamieson, D. N. (2010). Top-down pathways to devices with few andsingle atoms placed to high precision. New Journal of Physics, 12(6):065016.
[179] van Handel, R. and Mabuchi, H. (2005). Optimal error tracking via quantumcoding and continuous syndrome measurement. arXiv preprint quant-ph/0511221.
[180] Van Meter, R. and Horsman, C. (2013). A blueprint for building a quantumcomputer. Communications of the ACM, 56(10):8493.
[181] Vandersypen, L., Bluhm, H., Clarke, J., Dzurak, A., Ishihara, R., Morello, A.,Reilly, D., Schreiber, L., and Veldhorst, M. (2016). Interfacing spin qubits in quantumdots and donors-hot, dense and coherent. arXiv preprint arXiv:1612.05936.
[182] Veldhorst, M., Hwang, J., Yang, C., Leenstra, A., de Ronde, B., Dehollain, J.,Muhonen, J., Hudson, F., Itoh, K., Morello, A., et al. (2014). An addressable quantumdot qubit with fault-tolerant control-fidelity. Nature nanotechnology.
[183] Veldhorst, M., Yang, C., Hwang, J., Huang, W., Dehollain, J., Muhonen, J.,Simmons, S., Laucht, A., Hudson, F., Itoh, K. M., et al. (2015). A two-qubit logicgate in silicon. Nature, 526(7573):410414.
[184] Verduijn, J., Vinet, M., and Rogge, S. (2014). Radio-frequency dispersivedetection of donor atoms in a field-effect transistor. Applied Physics Letters,104(10):102107.
[185] Verstraete, F., Wolf, M. M., and Cirac, J. I. (2009). Quantum computation andquantum-state engineering driven by dissipation. Nat. Phys., 5(9):633636.
166 REFERENCES
[186] Viennot, J., Dartiailh, M., Cottet, A., and Kontos, T. (2015). Coherent couplingof a single spin to microwave cavity photons. Science, 349(6246):408411.
[187] Walls, D. F. and Milburn, G. J. (2007). Quantum optics. Springer Science &Business Media.
[188] Wang, D., Fowler, A., Stephens, A., and Hollenberg, L. (2010). Thresholderror rates for the toric and planar codes. Quantum Information & Computation,10(5):456469.
[189] Wang, D. S., Fowler, A. G., and Hollenberg, L. C. L. (2011). Surface codequantum computing with error rates over 1%. Phys. Rev. A, 83:020302.
[190] Wang, R., Shi, R., and Cheng, C.-K. (2006). Layer minimization of escape routingin area array packaging. In Computer-Aided Design, 2006. ICCAD ’06. IEEE/ACMInternational Conference on, page 815819.
[191] Weber, B., Mahapatra, S., Ryu, H., Lee, S., Fuhrer, A., Reusch, T., Thompson,D., Lee, W., Klimeck, G., Hollenberg, L. C., et al. (2012). Ohm’s law survives to theatomic scale. Science, 335(6064):6467.
[192] Weber, J. R., Koehl, W. F., Varley, J. B., Janotti, A., Buckley, B. B., Van de Walle,C. G., and Awschalom, D. D. (2010). Quantum computing with defects. Proceedingsof the National Academy of Sciences, 107(19):85138518.
[193] Wecker, D., Bauer, B., Clark, B. K., Hastings, M. B., and Troyer, M. (2014).Gate-count estimates for performing quantum chemistry on small quantum computers.Phys. Rev. A, 90:022305.
[194] Wecker, D. and Svore, K. M. (2014). Liqui|>: A software design architecture anddomain-specific language for quantum computing. arXiv:1402.4467.
[195] Whitfield, J. D., Biamonte, J., and Aspuru-Guzik, A. (2011). Simulation of elec-tronic structure hamiltonians using quantum computers. Mol. Phys., 109(5):735750.
[196] Wilson, L. (2013). International technology roadmap for semiconductors (itrs).
[197] Wiseman, H. M. and Milburn, G. J. (1993). Quantum theory of optical feedbackvia homodyne detection. Phys. Rev. Lett., 70:548551.
[198] Wiseman, H. M. and Milburn, G. J. (2009). Quantum measurement and control.Cambridge University Press.
[199] Wk, A., Łuczak, T., Kurzynski, P., Grudka, A., Gdala, T., and Bednarska,M. (2005). Unmodulated spin chains as universal quantum wires. Phys. Rev. A,72:034303.
[200] Xu, K., Souriau, L., Hellin, D., Versluijs, J., Wong, P., Vangoidsenhoven, D.,Vandenbroeck, N., Dekkers, H., Shi, X., Albert, J., Tan, C. L., Vertommen, J.,Coenegrachts, B., Orain, I., Kimura, Y., Wiaux, V., and Boullart, W. (2013). Keycontributors for improvement of line width roughness, line edge roughness, andcritical dimension uniformity: 15m half-pitch patterning with extreme ultraviolet
REFERENCES 167
and self-aligned double patterning. Journal of Micro/Nanolithography, MEMS, andMOEMS, 12(4):041302.
[201] Xue, S., James, M. R., Shabani, A., Ugrinovskii, V., and Petersen, I. R. (2015).Quantum filter for a class of non-markovian quantum systems. arXiv:1503.07999.
[202] Yamamoto, N. (2014). Coherent versus measurement feedback: Linear systemstheory for quantum information. Phys. Rev. X, 4:041029.
[203] Yao, N. Y., Jiang, L., Gorshkov, A. V., Maurer, P. C., Giedke, G., Cirac, J. I.,and Lukin, M. D. (2012). Scalable architecture for a room temperature solid-statequantum information processor. Nature communications, 3:800.
[204] Zhang, G. and James, M. R. (2011). Direct and indirect couplings in coherentfeedback control of linear quantum systems. IEEE Transactions on Automatic Control,56(7):15351550.
[205] Zhang, J., Liu, Y.-x., Wu, R.-B., Jacobs, K., and Nori, F. (2017). Quantumfeedback: theory, experiments, and applications. Physics Reports, 679:160.
[206] Zwanenburg, F. A., Dzurak, A. S., Morello, A., Simmons, M. Y., Hollenberg,L. C. L., Klimeck, G., Rogge, S., Coppersmith, S. N., and Eriksson, M. A. (2013).Silicon quantum electronics. Rev. Mod. Phys., 85:9611019.
[207] iljak, D. D. (1978). Large-scale dynamic systems: stability and structure, vol-ume 2. North Holland.
[208] er, B. (1998). A procedural formalism for quantum computing.
This page intentionally left blank.
Appendix A
Mathematical Proofs
A.1 General form of the QSDE for continuous error correction in Chap-
ter 3
Conditional master equation for continuous error correction monitor:
dρc =γ
3
n
∑m=1
∑j
D [σ(m)j ]ρcdt +κ
k
∑i=1
D [gi]ρcdt +√
κ
k
∑i=1
H [gi]ρcdWi
In which,
• σ(m)j represent Pauli operators (error channels), i.e. σ
(1)x = X1, etc.
• gi is set of stabilizer operators: for our distance-2 surface code, they are
Z1Z2Z3Z4, X1X2, X1X3, X2X4, and X3X4
We want to derive a set of stochastic differential equations for π(gi) = gi, which is
the expectation of stabilizer operator (syndrome value)
gi = π(gi) = Tr[giρc];dgi = Tr[gidρc]
dgl =γ
3
n
∑m=1
∑j
Tr[glD [σ(m)j ]ρc]dt +κ
k
∑i=1
Tr[glD [gi]ρc]dt +√
κ
k
∑i=1
Tr[glH [gi]ρc]dWi
170 Chapter A. Mathematical Proofs
• First term
Tr[glD [σ(m)j ]ρc] = Tr[gl(σ
(m)j ρcσ
(m)j −ρc)]
= Tr[glσ(m)j ρcσ
(m)j ]−Tr[glρc]
= Tr[σ (m)j glσ
(m)j ρc]−Tr[glρc]
=
0 if gl(m) ∈ I,σ j
−2gl if gl(m) /∈ I,σ j
The gl(m) in the last equation denotes the position m Pauli operator in the stabilizer
gl . In deed, if the stabilizer gl contains σ(m)j or has an identity operator at that
position, σ(m)j glσ
(m)j = gl . Otherwise, σ
(m)j glσ
(m)j =−gl , eg. X1(Z1Z2Z3Z4)X1 =
−Z1Z2Z3Z4 or Y1(X1X2)Y1 =−X1X2
For a depolarizing channel, we can generalize the results as following:
γ
3
n
∑m=1
∑j
Tr[glD [σ(m)j ]ρc]dt =−(4w)γ
3gl,w = Pauli weight of gl
• Second term
κ
k
∑i=1
Tr[glD [gi]ρc] = κ
k
∑i=1
(Tr[gl(giρcgi)]−Tr[glρc])
= κ
k
∑i=1
(Tr[(giglgi)ρc]−Tr[glρc]) = 0
This is the basis property of stabilizer operators: commuting and unitary.
• Last term
A.1 General form of the QSDE for continuous error correction in Chapter 3 171
√κ
k
∑i=1
Tr[glH [gi]ρc]dWi =√
κ
k
∑i=1
Tr[gl(giρc +ρcgi −ρcTr[giρc +ρcgi])]dWi
=√
κ
k
∑i=1
Tr[gl(giρc +ρcgi)]−Tr[glρc]Tr[giρc +ρcgi]dWi
= 2√
κ
k
∑i=1
Tr[glgiρc]−Tr[glρc]Tr[giρc]dWi
= 2√
κ
k
∑i=1
glgi − gl gidWi
Notice in this equation that:
– The product glgi can become a new operator outside the set of initial stabilizer
set. However, stabilizer code properties tell us that glgi is also a stabilizer of
the code and they will be closed.
– The resulting QSDE is non-linear as it involves second-order terms like gl gi
In conclusion, we have derive the set of QSDE’s tracking the parity syndrome of the
surface code which have the following form:
dgl =−(4w)γ3
gl +2√
κ
k
∑i=1
glgi − gl gidWi
with w is the weight if gl and l is the index of all members of the set generated by giki=1
For example, the distance-2 surface code has 5 stabilizer operators: Z1Z2Z3Z4,
X1X2, X1X3, X2X4, and X3X4. With that, we can form a close set of all stabilizer opera-
tors: Z1Z2Z3Z4, X1X2, X1X3, X2X4, X3X4, X2X3, X1X4, X1X2X3X4, Y1Y2Z3Z4, Y1Z2Y3Z4,
Z1Y2Z3Y4, Z1Z2Y3Y4, Z1Y2Y3Z4, Y1Z2Z3Y4, Y1Y2Y3Y4.
There are a total number of 15 equations, which is a great reduction from propagating
the full density matrix having 4n elements (256 for the distance-2 four qubit code).
However, there is also a caveat associated with this QSDE approach. We have not
yet included the feedback in our conditional master equation. Adding feedback will then
break the closeness of our stabilizer QSDE’s.
172 Chapter A. Mathematical Proofs
Let’s consider a specific type of feedback which is Hamiltonian (unitary) feedback,
i.e. adding the commutation term −i[F,ρc]dt into the conditional master equation, where
F is the feedback Hamiltonian.
Using a similar approach, we will have an additional term: Tr(−igl[F,ρc]) which
can be rewritten as Tr(−i[gl,F ]ρc) using the cyclic property of the trace. At this point,
we can see that additional operators will appear namely the commutator [gl,F ]. If we
choose F = σ(m)j to correct Pauli errors, the commutator will certainly be outside the
stabilizer set. Thus, a reduced QSDE model as before cannot be realized. Approximation
methods are required to reduce the dimension of the syndrome filter in order to tackle
real surface codes as they contain a big number of qubits.
A.2 Proof of the scalability condition (4.44)-(4.45)
Proof : By assumption, we have Ui satisfying Eq. (4.30). Let Li =UiVi ∏d V (i)d and we
have
G (Vi)Li= Vi ∏
dV (i)
d U†i ViUiVi ∏
dV (i)
d −Vi ∏d
V (i)d
≤ (1− ci −1)Vi ∏d
V (i)d =−ciVi ∏
dV (i)
d , ci > 0.
(A.1)
For j = i, either [U j,Vi] = 0 which results in
G (Vi)L j= Vj ∏
dV ( j)
d U†j ViU jVj ∏
dV ( j)
d −ViVj ∏d
V ( j)d
= 0, (A.2)
or U j satisfies (4.44) as
ViU†j ViU jVi ≤Vi. (A.3)
A.3 Proof of the condition (4.52)-(4.53) 173
Note that Eq. (A.3) implies Vi ∈ V ( j)d according to the assumption. Based on this fact,
we obtain
G (Vi)L j
= Vj ∏d
V ( j)d U†
j ViU jVj ∏d
V ( j)d −ViVj ∏
dV ( j)
d
= Vj ∏d
V ( j)d ViU
†j ViU jViVj ∏
dV ( j)
d −ViVj ∏d
V ( j)d
≤ 0, (A.4)
which further leads to
G (∑i
Vi) ≤ −∑i
ciVi ∏d
V (i)d ≤−cmin ∑
iVi ∏
dV (i)
d
≤ −cminλ ∑i
Vi. (A.5)
cmin is the smallest number in ci. Here we have used the condition (4.45).
A.3 Proof of the condition (4.52)-(4.53)
Proof. The Liouville-space representation of the master equation (4.33) is
vec(ρt) = Avec(ρt), (A.6)
whose solution is
vec(ρt) = eAtvec(ρ0), (A.7)
where vec(ρt) is a column vector obtained by piling up the column vectors of ρt , and A
is matrix determined by H,Lk.
The condition (4.52) ensures the invariance of the state if no error occurs in the first
step. If the error Ea occurs, the erroneous state Ea|p⟩⟨q|E†a needs to be steered back to
|p⟩⟨q|. Since we have
vec((Ea|p⟩⟨q|E†a)t) = eAtvec(Ea|p⟩⟨q|E†
a), (A.8)
174 Chapter A. Mathematical Proofs
and so
˙vec((Ea|p⟩⟨q|E†a)t)
= eAtAvec(Ea|p⟩⟨q|E†a)
= eAt [−κpqvec(Ea|p⟩⟨q|E†a)+κpqvec(|p⟩⟨q|)]
= −κpqeAtvec(Ea|p⟩⟨q|E†a)+κpqvec(|p⟩⟨q|)
= −κpqvec((Ea|p⟩⟨q|E†a)t)+κpqvec(|p⟩⟨q|). (A.9)
Eq. (A.9) is an ordinary first-order differential equation which can be easily integrated
to be
vec((Ea|p⟩⟨q|E†a)t)
= e−κpqtvec(Ea|p⟩⟨q|E†a)+κpqvec(|p⟩⟨q|)
∫ t
0e−κpq(t−r)dr
= e−κpqtvec(Ea|p⟩⟨q|E†a)+ vec(|p⟩⟨q|)[1− e−κpqt ]. (A.10)
This proves (Ea|p⟩⟨q|E†a)t → |p⟩⟨q| as t → ∞. Due to the linearity of the Liouville
equation, finally we have (∑p,q αpα∗q Ea|p⟩⟨q|E†
a)t → ρ0.
Appendix B
Feynman Path Integral Simulation
With FPGA
B.1 Method Overview
The Feynman path integral method can be used to calculate the propagation amplitude
from the input state (classical) to a particular classical output state. To illustrate this,
considering a quantum algorithm f operating on N qubits:
f : |i1i2...iN⟩ 7→2N−1
∑k=0
αk |k⟩ , (B.1)
In the Feynman path integral framework, we theoretically need to compute the
coefficients αk’s iteratively for all possible outcomes (2N). Then, this is completely
equivalent to the linear algebra approach which represents the output as a tall column
vector of length 2N . An unique advantage of the path integral method is its memory
efficiency and distributed-computing compatibility. Firstly, since we do not store the
whole state vector at any given time, we do not have the memory “hard” limit (space-time
trade-off). Secondly, path calculation is an independent workload, thus can be distributed
to a large number of processing units without worrying about data dependency. For
example, each unit compute the αk amplitude for a specific value of k.
176 Chapter B. Feynman Path Integral Simulation With FPGA
Let’s look at the method in more details. Consider a quantum circuit with d layers
of quantum gate and N qubits. Assuming only 1- and 2-qubit gates are used (satisfying
quantum computing universality).
G1
G2
G3
G4
Gd
The input is a classical state (|x⟩) that we load to the qubits. We calculate the
amplitude αk to get a readout of |y⟩= |k⟩, k ∈ [0,2N −1] as
⟨x| f |y⟩= ∑internal nodes
⟨x1|G1 |a1⟩⟨x2x3|G2 |b1c1⟩ ..Gd |y4y5⟩ , (B.2)
in which we used the notations a to h to represent the 8 qubit lines and the subscript
numbers to denote the sequence of gates on that line. We shall call a connection inside the
circuit between gates as “wire”. The idea is to sum over all possible binary combinations
of internal wires.
In order to explain this concept better, let’s look at a simple example which is a
3-qubit quantum Fourier transform (QFT) circuit.
We have labeled all the internal wires which need to be summed over. Since we have 4
internal wires, each path calculation will require a sum of 24 = 16 terms. Each term is
computed from a product of 6 complex numbers (equal to the number of gate layers).
B.2 FPGA Implementation 177
And to get the complete final state vector, we need to repeat this task 23 = 8 times to
cover all possible outcomes of |y1y2y3⟩. Explicitly, we have:
⟨x1x2x3|QFT3 |y1y2y3⟩ = ∑c1c2b1c3
(⟨x1|H |c1⟩⟨x2c1|CR(π/2) |x2c2⟩⟨x2|H |b1⟩ (B.3)
× ⟨x3c2|CR(π/4) |x3c3⟩⟨x3b1|CR(π/2) |x3y2⟩⟨x3|H |y1⟩).
For each layer of gate, we only deal with the local gate which is applied at that layer.
The operands for the multiplier are elements of the corresponding gate matrix indexed
by the current classical bra and ket states.
B.2 FPGA Implementation
As shown in the previous example, the calculation of Feynman path integral can be
efficiently simulated on the FPGA by:
1. Streaming the complex matrix elements, i.e. ⟨a1b1|G |a2b2⟩, into the FPGA. The
host CPU will serve as a parser for the input quantum circuit and dispatch the
coefficients onto a continuous stream sending to the FPGA.
// Define complex number operation functions
//multiplyfloat2 comp_mult(float2 a, float2 b)
float2 res;res.x = a.x * b.x - a.y * b.y;res.y = a.x * b.y + a.y * b.x;return res;
//addfloat2 comp_add(float2 a, float2 b)
float2 res;res.x = a.x + b.x;res.y = a.y + b.y;return res;
Fig. B.1 OpenCL implementation of complex arithmetic.
178 Chapter B. Feynman Path Integral Simulation With FPGA
2. Implement a pipeline of sum operations with enough depth to hide the latency in-
volved in the data loading and multiplication (multiple copies of the multiplication
unit). The high-level structure can be seen as following:
+ + + +
m-way
Complex
Multiplier
m-way
Complex
Multiplier
m-way
Complex
Multiplier
m-way
Complex
Multiplier
RESULT STORE
LOAD LOAD LOAD LOADCOMPUTE UNIT
3. Multiple compute units can be mapped onto the FPGA so long as the hardware
complexity is permitted. Many path can be calculated independently on the same
FPGA or even across multiple FPGA’s. For a very large quantum circuit, we can
choose between fast single amplitude calculation (latency optimized) or parallel
multiple amplitudes (throughput optimized). Data access can be organized in
SIMD fashion for more efficient bandwidth.
4. Perform quantum circuit optimization before the simulation, e.g. combining gates
together for less memory transfer between FPGA and CPU, reducing internal
nodes, etc. Innovative data structure for gate coefficient storage to quickly identify
multiplying sequences which contain zero element(s), thus don’t send to the
FPGA.
B.3 Software Components 179
B.3 Software Components
B.3.1 Host Program
Task: Manage the overall execution of the simulator, including: reading the input
quantum gate sequence and configurations; assembling the required data to send to the
FPGA; collecting the results returned by all the compute units of the FPGA; performing
any circuit optimization if required.
Components:
• Quantum circuit parser: parsing the quantum assembly language. For example,
the sequence for the above 3-qubit QFT circuit is shown below.
num_operations: 6
// Quantum Fourier Transform on 3-qubit register x
H x[1]
cR_Piover2 x[2], x[1]
H x[2]
cR_Piover4 x[3], x[1]
cR_Piover2 x[3], x[2]
H x[3]
• FPGA control and management for initialization, send and receive data, etc. This
part is closely related to the kernel implementation, e.g. the assembled data stream
needs to match the work-group size of the kernel and the returned amplitude data
needs to be mapped to the corresponding state, etc.
• Quantum circuit optimization: reorganizing and combining gates to reduce internal
nodes; eliminating zero or insignificant paths; optimizing data transfer between
host and kernel; etc.
B.3.2 FPGA Kernel
Task: Fetching data in main memory sent by the host, computing the complex mul-
tiplication, adding the amplitudes of multiple paths. Depending on the problem size
(number of gates/circuit depth), multiple compute units can be instantiated in one FPGA
for computing multiple output states or segmenting the sum into chunks to optimize for
low latency in case only one amplitude is concerned.
180 Chapter B. Feynman Path Integral Simulation With FPGA
// Define the work group size__attribute__((reqd_work_group_size(NUM_WORKGROUPS, 1, 1)))
// Multiplication kernel__kernel void Feynman_Mult (global const float2 * restrict src)
//get global IDuint gid = get_global_id(0);
//temp var for running productfloat2 running_prod;running_prod.x = 1.0f;running_prod.y = 0.0f;
// Compute the multiplication of corresponding elements// in the data stream.#pragma unrollfor (unsigned a = 0; a < MUL_POINTS; a ++)
running_prod =comp_mult(running_prod, src[gid*MUL_POINTS + a]);
// Write the result to the FIFO channel// for the next kernel (sum) to processwrite_channel_altera (chanin, running_prod);
// Summation kernel__kernel void Feynman_Sum(global float2 * restrict result)
//temp var for running sumfloat2 running_sum;running_sum.x = 0.0f;running_sum.y = 0.0f;
float2 tempVar;for (unsigned i = 0; i < NUM_WORKGROUPS; i++)
// read data from the channeltempVar = read_channel_altera(chanin);
// cumulative sumrunning_sum = comp_add(running_sum, tempVar);
//get global IDuint gid = get_global_id(0);//send back the sum result with respect to the assigned IDresult[gid] = running_sum;
Fig. B.2 OpenCL implementation of multiplication and summation kernels.
B.3 Software Components 181
Internally, the compute unit has a couple of kernels passing data through FPGA FIFO
channel as shown in the figure below. This will be more efficient and flexible in terms of
memory access (potential bottleneck). Data input and output to and from the kernel are
fetched and assembled before accessing the memory.
Components:
• Fetcher: fetch data stream from memory. Optimize for efficient data access.
• Multiplier: need to be sufficiently wide and balance for fast execution.
• Sum: deeply pipelined to mask the latency of the fetcher and multiplier. For real
circuits with a sufficient number of internal nodes (many terms in the sum), this
strategy is feasible since after the initial delay to do the first fetch and multiplica-
tion, sum operation can be done in every cycle.
This page intentionally left blank.
Index
automatic quantum error correction, 78
Bloch sphere, 5
coherence time, 11
coherent feedback, 25
decoherence-free subsystem, 69
dissipatively quasi-locally stabilizable, 77
DiVincenzo’s criteria, 3
donor-based qubits, 6, 30
entanglement, 1, 2
fault-tolerance threshold, 12
Fock space, 73
Hilbert space, 2
Lindblad superoperator, 74
logical qubits, 12
master equation, 18
measurement feedback, 24
open quantum system, 73
open quantum systems, 18
quantum assembly language, 121
quantum circuits, 6
quantum computer architecture, 30
quantum computer micro-architectures, 30
quantum computers, 3
quantum computing, 2
quantum dissipation control, 69
quantum dots, 30
quantum error correction, 11, 12
quantum feedback control, 24
quantum filtering, 19
quantum gates, 6
quantum input-output model, 20
quantum noise, 19
quantum programming language, 16
quantum stochastic differential equation,
20
quantum Wiener process, 73
qubit, 3
spin qubits, 30
superposition, 1, 2
surface code, 12
tensor product, 2, 5
universal quantum gates, 9