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 ownworkor 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 Technologies1 1.2 Quantum Computing3 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...... 5 1.2 Solid-state qubit diagram...... 6 1.3 Quantum gate timing...... 9 1.4 Quantum gate decomposition...... 10 1.5 Surface code layout diagram...... 13 1.6 Quantum circuit to measure surface code syndromes...... 13 1.7 Logical CNOT gate braiding diagram...... 14 1.8 Logical CNOT gate by braiding holes in surface code...... 15 1.9 Active and passive QEC diagram...... 19 1.10 Diagram of a typical quantum filtering set-up...... 20 1.11 SLH network connections...... 21 1.12 Diagram of beamsplitter in series with cavity QED...... 22 1.13 Sample QHDL code...... 23 1.14 Sample QHDL schematic...... 24 1.15 Measurement feedback block diagram...... 25 1.16 Coherent feedback block diagram...... 26 1.17 Overall architecture of a universal quantum computer...... 27
2.1 Routing parameters...... 32 2.2 Diagram of surface code lattice with embedded readout devices...... 34 2.3 Diagram of a generic surface code array unit cell...... 35 2.4 Metal 1 pitch scaling roadmap from ITRS...... 36 2.5 Ring-by-ring routing...... 39 2.6 Layer optimal routing...... 40 2.7 Illustration of 2-D qubit lattice surface gate fanout...... 42 2.8 Fanout scalability at fixed interconnect length...... 43 2.9 Qubit fanout scalability in terms of number of routing layers...... 45 2.10 Fan-out scalability vs. interconnect length for SSI and ECI schemes...... 46
3.1 Surface code layout diagram...... 52 3.2 Distance-2 surface code...... 56 3.3 Surface code continuous error correction diagram...... 58 xx LIST OF FIGURES
3.4 Syndrome estimation trajectories...... 62 3.5 Continuous feedback error correction fidelity comparison...... 63 3.6 Distance-3 surface code diagram...... 64 3.7 Time-domain simulation of distance-3 surface code under continuous QEC. 66 3.8 Fidelity vs. feedback strength plot...... 67
4.1 Potential function plot...... 80 4.2 Lyapunov function scalability...... 85 4.3 State evolution simulation of repetition error correction code...... 97 4.4 Repetition error correction code under dissipative couplings...... 98 4.5 Coherent feedback network for DFS generation...... 100 4.6 Open loop setup for DFS generation...... 101 4.7 Coherent plant-observer network...... 109 4.8 Coherent feedback network for DFS generation...... 111 4.9 Two-cavity system diagram...... 115
5.1 AQASM example...... 122 5.2 AQASM QFT example...... 123 5.3 PyAQASM QFT example...... 124 5.4 Quantum simulator diagram...... 125 5.5 Path integral diagram...... 127 5.6 Quantum circuit simulation objects...... 128 5.7 Benchmark of Hadamard gates...... 129 5.8 Comparison with Liquid...... 130 5.9 Comparison with other simulators...... 130
6.1 QHDL design flow...... 135 6.2 Single-photon generating filter cascading realization...... 136 6.3 Gradient settings for GEM operations...... 137 6.4 Engineered dissipation by direct coupling...... 139 6.5 Quantum circuit of cascading Hamiltonian unitary...... 141 6.6 Discretized single-photon GEM model...... 142 6.7 Amplitude damping LIQUi code snippet...... 143 |⟩ 6.8 Histogram of photon detection events...... 144 6.9 Photon absorption vs. photon bandwidth...... 145
B.1 OpenCL implementation of complex arithmetic...... 177 B.2 OpenCL implementation of multiplication and summation kernels...... 180 List of tables
1.1 Types of experimental quantum bits (qubits)...... 7 1.2 Hadamard gate implementation steps...... 10
2.1 Gate count configurations for spin shuttling interconnect (SSI) and end control interconnect (ECI) protocols...... 37 2.2 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...... 126 5.2 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 modelling132 [ ], machine learning and artificial intelligence24 [ , 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. In fact, classical probabilistic state spaces also scale exponentially in dimension, hence require specialized computational 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 + q 1 , where p and q are real numbers in the unit interval that represent the | ⟩C | ⟩C probability of finding that bit in either state, e.g. the bit has a probability p of being in the 0 state. We always have p + q = 1, and the state 1 0 + 1 1 represents a | ⟩C 2 | ⟩C 2 | ⟩C 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.
α ψ = α 0 + β 1 , α, β C (1.1) | ⟩ | ⟩ | ⟩ ≡ β ∈ 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 the sphere; (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 J A
P P
D
1 Fig. 1.2 Diagram of the Phosphorous in Silicon donor spin qubit. The donor atom has one extra electron whose spin is used as qubit. The surface A gate on top of the atom controls the hyperfine interaction which can be used together with a magnetic fieldto perform addressable single-qubit operations. The J gate in between two adjacent atoms controls 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. Coherence time: hours Long coherence time but gate operations are also slow. Gate fidelity: >99.9% Scalability issues w.r.t. laser control. Max number of qubits: 10-50
Superconducting resonators [33] Electrical LC oscillators with zero resistance. Coherence time: 10-100 µs Electrically-controlled by microwave signals. Gate fidelity: >99.5% Operates at cryogenic temperature. Max number of qubits: 50+
Quantum dots [101, 10] Gate-defined nano structures in semiconductor materials. Coherence time: µs - ms Small (nano-scale) qubits, electrically controlled. Gate fidelity: >99% Using the same fabrication technology as classical Max number of qubits: 2-5 chips. Relatively few interacting qubits can be demonstrated.
Donor atoms [48, 47] Coherence time: up to secs Single dopant atom site in semiconductor. (nuclear spins) Both nuclear and electron spins can be used as qubits. Gate fidelity: >99% Single-atom doping is challenging. Max number of qubits: 2-5
Diamond vacancies (NV centers) [192, 203] Nitrogen atom in diamond and a vacant lattice site. Coherence time: secs Laser-controlled and can operate at room temperature. Gate fidelity: >99% Difficult to couple multiple NV qubits for quantum Max number of qubits: 5-10 computing.
Topological qubits [117, 2] Quasi non-local particles (majorana fermions). Can have extreme long coherence time due to Coherence time: unknown non-locality. Gate fidelity: unknown Less developed theoretical and experimental capabili- Max number of qubits: 1 ties.
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: 1 1 1 H = . (1.2) √2 1 1 − Another widely-used set of single-qubit gates are called Pauli gates whose matrices are: 0 1 0 i 1 0 X = , Y = − , Z = . (1.3) 1 0 i 0 0 1 − Given a single-qubit gate which is represented by a unitary matrix U, a controlled-U gate is a two-qubit gate as follows:
1 0 0 0 0 1 0 0 C(U) = . (1.4) 0 0 u u 00 01 0 0 u10 u11 where u00 u01 U = . (1.5) u10 u11 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
Ai=1 6 Target: R (2π) I Target: R (θ) x ≡ x Observer: R (2π + θ) R ( θ) Observer: R (θ) x ≡ x − x
Fig. 1.3 Timing diagram to implement X-rotation on donor spin electron. A1 is the A-gate contact of the target qubit.
Similar techniques can be used to rotate the qubit around an arbitrary axis. Indeed, ˆ ˆ the Hadamard is nothing but a π rotation around the axis mˆ = i+k (iˆ, jˆ, and kˆ are basis √2 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) HRx(α) ≡ 2 Rx(2π) IRx( 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,
I X I X CNOT = (H I)exp iπ − − (H I) ⊗ 2 ⊗ 2 ⊗ π π = (H I) Rx Rx (1.6) ⊗ 2 ⊗ 2 π h i exp i X X (H I). × 4 ⊗ ⊗ The only two-qubit operation in Eq. (1.6) is exp i π X X , which can be implemented 4 ⊗ straightforwardly using the coupling H = J(σ σ) term, where σ are the Pauli spin ⊗ matrices (with eigenvalues 1), as ± π π exp i X X = (X I)exp i σ σ 4 ⊗ ⊗ 8 ⊗ (1.7) π (X I)exp i σ σ . × ⊗ 8 ⊗ 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.
π a H X X Rx H a | i 2 | i π π U 8 U 8 b R π a b | i x 2 | ⊕ i π π Fig. 1.4 Quantum circuit decomposition of a CNOT gate. U 8 = exp i 8 σ σ . Each of the decomposed elements can be realized by properly-timed pulsing of⊗ the surface gates (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), itisthe 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 aqubit, 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 ofa 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, whicheveris 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 X 2 X 3 X 4
Z 5 Z 6 Z 7 Z
8 X 9 X 10 X 11
Z 12 Z 13 Z 14 Z
15 X 16 X 17 X 18
Fig. 1.5 Surface code layout: white circles represent data qubits; filled circles are syndrome qubits (X stabilizers in green and Z stabilizers in yellow). Each internal stabilizer acts on four adjacent data qubits, while boundary stabilizers act on either two or three data qubits.
0 M 0 H H M | i Z | i Z a a | i | i b b | i | i c c | i | i d d | i | i (a) Z-syndrome measurement (b) X-syndrome measurement
Fig. 1.6 Quantum circuit to measure surface code syndromes. The first line represents the 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 running horizontally.
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]
d+1 εphysical ⌊ 2 ⌋ εlogical C1 C2 , (1.8) ≥ ε threshold 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 by expanding and shrinking a hole. In order to maintain the code distance d throughout this process, 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.
2 2 Nphysical ∝ d (4 Nlogical) , (1.9) × × p 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 ∗(ρ(t)) (1.10) − ∑ L 1.4 Quantum Control Engineering 19
Active Error Correction
Correction
Passive Error Correction/Noise Cancellation
Dynamical decoupling (open-loop) Coherent/classical feedback control
Syndrome Measurement
Quantum Error Correction Qubit Network
Fig. 1.9 Combination of passive and active techniques for quantum error correction in which the passive (autonomous) layer improves the stability and/or error rejection to the 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
† 1 † 1 † L ∗(ρ) = LρL L Lρ ρL L. (1.11) L − 2 − 2
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.
System Measure before after measurement interaction interaction signal estimate filter B(t) Bout(t) Y (t) Xˆ(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) {− − † † 1 † + dA (t)Li L SdAi(t) L Lidt U(t), ∑ i − i − 2 i i ) 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 C G1
2 G G1 G2
Fig. 1.11 SLH network connections: (left) Concatenation and (right) Cascading.
Mathematically, these two connection rules can be expressed as:
S1 0 L1 G2 ⊞ G1 = , ,H1 + H2 , (1.13) 0 S2 L2 † G2 G1 = (S2S1,L2 + S2L1,H1 + H2 + Im L S2L1 ). (1.14) ◁ { 2 } 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.
γ
A˜2 κ A1 B1
B˜1 A2
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:
β α M = − ,0,0 , α β √κa C = I, ,Hc , √γσ − † † 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 IS PORT ( A1 : in fieldmode; A2 : in fieldmode; Vacin : in fieldmode; B1 : out fieldmode; B2 : out fieldmode; Vacout : out fieldmode); END MachZehnder;
-- Secondary unit ARCHITECTURE netlist OF BeamSplitterNetwork IS
COMPONENT Beamsplitter GENERIC ( theta : real := theta_value); PORT ( In1 : in fieldmode; In2 : in fieldmode; Out1 : out fieldmode; Out2 : out fieldmode); END COMPONENT ;
COMPONENT SingleSidedJaynesCummings GENERIC ( 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 part A_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 thisto generate Heisenberg or Schrödinger type dynamical equations to perform simulation or verification as desired.
B2 = B˜2 = A˜2
Cavity BeamSplitter
A1 Out2 Out1 B1 In1 ( ) In1 −
In2 Out1 B˜1
VacIn UOut 2 A In Out Vac Vac
Fig. 1.14 Sample schematic using QHDL schematic capture tool: A QED cavity is connected 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
Target: ρ Quantum Device
I/V Est.:ρ ˆ + Meas. Filter
F
Fig. 1.15 Measurement feedback block diagram.
signals based on pre-defined control algorithms. A typical block diagram ofa 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) unu Field Quantum
Direct Coupling Quantum Field Quantum Controller
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 Software (Algorithms)
Quantum Error Correction Logical Qubits
Quantum Gate Operations
Superconductors Ion Traps Quantum Dots Donors NV Centers Topological
Physical Qubits Quantum Hardware
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 un Ordinateur 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 qubits177 [ , 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:
p d s C = − − (2.1) ⌊ w + s ⌋
• Diagonal channel: the routing space available between diagonal pads. In a square array, its routing capacity is calculated by:
√2p d s D = − − (2.2) ⌊ w + s ⌋
p
grid channel d s
diagonal channel
w
Fig. 2.1 Routing parameters: pad pitch (p), wire width and spacing (w and s), pad diameter (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 thatthe 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 X 2 X 3 X 4
Z 5 Z 6 Z 7 Z
8 X 9 X 10 X 11
Z 12 Z 13 Z 14 Z
15 X 16 X 17 X 18
Data Qubits Syndrome Qubits Readout
Fig. 2.2 Diagram of surface code lattice with embedded readout devices. There are two types of qubits: data qubits and syndrome qubits (X and Z types). Neighboring qubits can interact with each other in order to perform CNOT gates. In this model, each qubit has its own readout device. Dashed lines (black) represent quantum interconnects between neighboring qubits. Dotted lines indicate qubits to which readout devices are associated.
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).
Nc
Nq
READOUT
Nc
Nr
Fig. 2.3 Diagram of a generic surface code array unit cell. Each qubit (circle) has a certain number of surface gates (Nq) to define qubit confinement potential and to perform single-qubit rotations. Between any pair of neighboring qubits, we have Nc coupling gates that are used to control qubit interconnect coupling. At the center of the cell, we have 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 of20 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.
80 ITRS 2013 Production Data
60
40 Metal 1 Pitch (nm) 20
2015 2020 2025 Year Fig. 2.4 Metal 1 pitch scaling roadmap from ITRS 2013 [196]; production data is taken from 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 control interconnect (ECI) protocols. The interconnect node count (Nnodes) is the number of intermediate 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: unit cell dimension(L) p (2.3) RL , ≤ Ntotal(L) where N is the total number of gate contacts in a unit cell and p is the contact total p RL 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
20nm N p nodes 35nm (2.4) RL × > , ≈ 1 + 4Nnodes + 3 p 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 absolute distance for spin shuttling interconnect (SSI) and end control interconnect (ECI) proto- cols. Qubit-qubit pitch is 20 nm for atomically confined qubitN ( q = 1) and 50 nm for electrostatically confined qubitN ( q = 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 unconnected pads are connected first, then inner rings are connected using grid channels of theouter ring 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 routing on the second layer, and so on. The overall procedure is depicted in the right diagram where 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 by proceeding triangularly inward. In this way, it can deploy the diagonal channels, which always have higher routing capacity and take advantage of empty spaces resulted from previously 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 routing approach.
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 padsin 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 routing layers. The bottom layer is a semiconductor material (Si or GaAs) with top gates for control and readout. On the same substrate lies classical integrated electronics used for signal generation, multiplexing, and sensing. In order to bring connections to the surface gates, multi-layer routing is needed. After surface gates are redistributed into a square-grid array of contacts, as shown in the first metal layer, the fan-out routing procedure is carried out layer-by-layer using a specific routing algorithm. This figure demonstrates ring-by-ring routing, which requires three metal layers for this particular grid 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
Number of routing layers comparison for Nq = 1 (L=15, 300nm) Number of qubits routable with 20 layers (L=300 nm)
SSI Ring-by-ring Ring-by-ring 103 8000 SSI Optimal Optimal ECI Ring-by-ring 7000 ECI Optimal 6000 102 5000
20-layer limit 4000
101 3000 Number of layers
Number of physical2000 qubits
1000 464 100 27 27 58 0 1 2 3 4 5 10 10 10 10 10 Nq = 1 Nq = 3 Nq = 1 Nq = 3 Number of physical qubits End Control Interconnect Spin Shuttling Interconnect
Fig. 2.8 Comparison between different interconnect protocols and routing methods at fixed interconnect length: (left) Plot of the number of routing layers vs. number of physical qubits for atomically confined qubits (Nq = 1) at L = 300nm; and (right) scalability comparison between electrostatically confined qubitsN ( q = 3) and atomically confined qubitsN ( q = 1) using the same number of routing layers (20) and interconnect length (L = 300nm, 15 nodes for Nq = 1 and 6 nodes for Nq = 3). Dimension parameters are (see Methods): d = 10nm, w = 5nm, and s = 25nm. The red dashed horizontal line on the left figure represents the technological limit of 20 metal layers that can be fabricated reliably 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 (from2x 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
Number of routing layers for SSI scheme (Nq = 1) Number of routing layers for ECI scheme (Nq = 1) 103 103 L = 15 (300nm) L = 15 (300nm) L = 20 (400nm) L = 20 (400nm) L = 30 (600nm) L = 30 (600nm)
2 102 10
20-layer limit 20-layer limit
1 1 10 10 Number of layers Number of layers
100 100 101 102 103 104 105 101 102 103 104 105 Number of physical qubits Number of physical qubits
Number of qubits routable with 20 layers - SSI Number of qubits routable with 20 layers - ECI 700 50000 Nq = 1 Nq = 1 Nq = 3 Nq = 3 600 Nq = 5 Nq = 5 Hybrid 40000
500
30000 400
300 20000
200 Number of physical qubits Number of physical qubits
100 10000 100 20 20
0 0 L = 300 nm L = 400 nm L = 600 nm L = 300 nm L = 400 nm L = 600 nm
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 atomically confined qubits (Nq = 1) under different interconnect lengths; and (bottom) number of routable qubits comparison between electrostatically confined qubitsN ( q = 3 and Nq = 5), atomically confined qubits (Nq = 1), and hybrid SSI (donors as qubits and dots as 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 20 metal 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
Number of qubits routable with 20 layers vs SSI interconnect length Number of qubits routable with 20 layers vs ECI interconnect length 700000 1200 Nq = 1 Nq = 1 Nq = 3 Nq = 3 600000 1000
500000 800 400000 600 300000
400 200000 Number of physical qubits Number of physical qubits 200 100000
0 0 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 SSI interconnect length (µm) ECI interconnect length (µm)
Fig. 2.10 Fan-out scalability vs. interconnect length for (left) SSI and (right) ECI schemes. The inter-qubit interconnect lengths are given in absolute unit (µm). The number of interconnect qubit nodes can be inferred by noting that the qubit-qubit distance is 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 S(ρ), is a linear, trace preserving, and com- 7−→ pletely positive map1. Theorem 3.1.1 [Kraus (operator-sum) decomposition] S is a completely positive trace preserving map iff † † S(ρ) = ∑k AkρAk, with ∑k AkAk = 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:
p S(ρ) = (1 p)ρ + (XρX +YρY + ZρZ), − 3
which can also be expressed using the representation:
4p 4p I S(ρ) = (1 )ρ + . − 3 3 2 Note: the identity operator (I) is the balanced mixture, i.e. I ρ 1 1 1 2 = 4 + 4 XρX + 4YρY + 4 ZρZ 3. General Pauli channel
S(ρ) = (1 p p p )ρ + p XρX + p YρY + p ZρZ − x − y − z x y z 4. Amplitude damping channel
† † S(ρ) = A0ρA0 + A1ρA1, 1This is equivalent to the Master equation formulation in (1.10). 3.1 Quantum Errors 51
1 0 0 γ where A0 = and A1 = . 0 1 γ2 0 0 − p In the above models, p the parameter 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:
2 n L = ψ (C )⊗ : 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 X 2 X 3 X 4
Z 5 Z 6 Z 7 Z
8 X 9 X 10 X 11
Z 12 Z 13 Z 14 Z
15 X 16 X 17 X 18
Fig. 3.1 Surface code layout: white circles represent data qubits; filled circles are syndrome qubits (X stabilizers in green and Z stabilizers in yellow). Each internal stabilizer acts on four adjacent data qubits, while boundary stabilizers act on either two or 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-flipZ ( ) 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 × d 1 of distance N. By definition, a distance-d code can correct arbitrary − -qubit errors. ⌊ 2 ⌋ 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 Pazand 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 code153 [ ]. Oreshkov and Brun studied continuous quantum error correction for non-Markovian decoherence using the model of a bit-flip code131 [ ]. 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. X X X X , and identity operators (I) are assumed in all places where no other Pauli 1 2 ≡ 1 ⊗ 2 operator applied, such as X X X I I X . 1 4 ≡ 1 ⊗ ⊗ ⊗ 4 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:
1 D[L]ρ = LρL† (L†Lρ + ρL†L), − 2
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 willbe 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 ina 2 2 configuration as shown in × Fig. 3.2. The code stabilizer generators are Z1Z2Z3Z4, X1X2, X1X3, X2X4, and X3X4.
X 1 X
2 Z 3
X 4 X
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 (1 p)ρ + p (XρX +YρY + ZρZ). 7→ − 3 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 channelsL ( E) 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:
j [1,N] γ j ∈ L = σ ( ) . E 3 i r 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:
T LS = √κ [Z1Z2Z3Z4,X1X2,X1X3,X2X4,X3X4] . 58 Chapter 3. Continuous Quantum Error Correction
dA ,L dAE,out E E Surface Code ˆ dAS,LS dAS,out S(t) F(t) Array Estimator Controller
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. The syndrome outputs are measured to estimate the syndrome conditional expectation values (Sˆ(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:
† † 1 † dU(t) = dAE(t)LE LEdAE(t) LELEdt − − 2 (3.1) n 1 + dA†(t)L L†dA (t) L†L dt U(t), S S − S S − 2 S S o where we assume no internal dynamics (H = 0). The QSDE for an arbitrary observable X on the qubit network is given by:
d j X j X X dt t( ) = t(LLE ( ) + LLS ( )) † † + jt([LE,X])dAE(t) + jt([X,LE])dAE(t) (3.2) † † + jt([LS,X])dAS(t) + jt([X,LS])dAS(t).
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) + dAS(out)(t). Assuming that the input fields are in vacuum state, the five output equations canbe 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
† 2 where dWi(t) = dASi(t) + dASi(t) is equivalent to a classical Wiener process , 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:
d X X X dt πt( ) = πt LLE ( ) + LLS ( ) (3.4)
+ ∑ πt( LSi,X ) 2πt(LSi)πt(X) dWi(t), i { { } − }
† where we have used the Hermitian property of syndrome operators (LSi = LSi) 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 g the set of syndrome generators, by using (3.4) we can get their filtering { i} equations:
(4w)γ dπ (g ) = π (g )dt t l − 3 t l k (3.5) + 2√κ ∑ πt(glgi) πt(gl)πt(gi) dWi, i=1{ − } 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 ofsyndrome 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:
4 4 z y F(t) = ∑ λi (t)Zi + ∑ λi (t)Yi (3.6) i=1 i=1
z y where λi (t) and λi (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:
λ λ z = 0 (1 X )(1 X )(1 + X )(1 + X )(1 + Z[), (3.7) 1 32 − 12 − 13 24 34 1234 λ λ z = 0 (1 Xc )(1 + Xc )(1 Xc )(1 + Xc )(1 + Z[), (3.8) 2 32 − 12 13 − 24 34 1234 λ λ z = 0 (1 + Xc )(1 Xc )(1 + Xc )(1 Xc )(1 + Z[), (3.9) 3 32 12 − 13 24 − 34 1234 λ λ z = 0 (1 + Xc )(1 + Xc )(1 Xc )(1 Xc )(1 + Z[), (3.10) 4 32 12 13 − 24 − 34 1234 y λ λ = 0 (1 Xc )(1 Xc )(1 + Xc )(1 + Xc )(1 Z[), (3.11) 1 32 − 12 − 13 24 34 − 1234 y λ λ = 0 (1 Xc )(1 + Xc )(1 Xc )(1 + Xc )(1 Z[), (3.12) 2 32 − 12 13 − 24 34 − 1234 y λ λ = 0 (1 + Xc )(1 Xc )(1 + Xc )(1 Xc )(1 Z[), (3.13) 3 32 12 − 13 24 − 34 − 1234 y λ λ = 0 (1 + Xc )(1 + Xc )(1 Xc )(1 Xc )(1 Z[), (3.14) 4 32 12 13 − 24 − 34 − 1234 c c c c 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).
c 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 z feedback term λ1. 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 5 (γ). A time step of order 10− 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:
( j) ( j) ΠC = ρ0 + ∑σi ρ0σi . (3.15) i, j
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 atthattime[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
Surface Code Syndrome Estimation 1.0
0.5 Z1Z2Z3Z4 X1X2 0.0 X1X3 X2X4 X3X4 Syndrome Value 0.5 −
1.0 − 0.00 0.02 0.04 0.06 0.08 0.10 Time (1/γ) Fig. 3.4 Syndrome estimation for a single trajectory showing +1 to 1 transitions from − three syndrome operators: X1X3, X3X4, and Z1Z2Z3Z4. This indicates a Y error may occur to qubit 3. Dashed line represents the point in time selected for the fidelity and trace distance comparison in table 3.1. case. Nielsen and Chuang have outlined two effective measures which are the trace distance and fidelity122 [ ]. 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 (ρ0′ = 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) (ρ0′ ,ρ) Trace Distance: D(ρ,σ) = 1 Tr ρ σ 0.999 0.069 2 | − | Fidelity: F(ρ,σ) = Tr ρ1/2σρ1/2 0.018 0.998 p 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 atall, as the feedback loop is correcting the errors in real-time while they are occurring. More 3.4 Results 63
Fidelity Comparison 1.0 0.9 0.8 0.7 0.6 0.5 Fidelity 0.4 No Feedback 0.3 Feedback 0.2 Correctable Overlap 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Time (1/γ) Fig. 3.5 Comparison between continuous feedback error correction fidelity vs. no feedback fidelity and correctable overlap. Feedback enhances the code fidelity overlong period of time. The cross-over between no feedback correctable overlap and feedback fidelity indicates continuous error correction outperforms its discrete counterpart if correction 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:
i γ i i i L( ) = (σ ( ) + σ ( ) + σ ( )), (3.16) e 3 x y z
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
X
1 2 3
Z X Z
4 5 6
Z X Z
7 8 9
X
Fig. 3.6 Diagram of distance-3 surface code: white circles are data qubits, green circles are 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:
1 1 d ψ(t) = iH ψ dt + L + L† L L†L L + L† 2 ψ dt (3.17) | ⟩ − | ⟩ 2 ⟨ ⟩ − − 4⟨ ⟩ | ⟩ 1 + (L L + L† ) ψ dW. − 2⟨ ⟩ | ⟩
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 specific feedback 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.
Fidelity Plot vs. Feedback Strength 1.0
0.9
0.8
0.7 Fidelity
0.6
0.5
0.4 0.0 0.2 0.4 0.6 0.8 1.0 λ/κ
Fig. 3.8 Comparison of the final fidelity of the surface code under varying feedback strength.
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 thegrowing 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 equivalent59 [ ]. 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-author 2significant 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 . − N A finite-level quantum system is defined on a Hilbert space H C . 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
ρ˙t = L (ρt) J † 1 † 1 † = i[H,ρt] + ∑ L jρtL j L jL j ρt ρtL j L j. (4.1) − j=1 − 2 − 2
Here H B(H ) is the system Hamiltonian and L B(H ), j = 1, ,J are system ∈ { j ∈ ··· } operators that characterize the system-environment couplings. ZV is the space spanned by the ground states of V. 1 0 0 1 0 i σz = ,σx = ,σy = − are Pauli operators acting 0 1 1 0 i 0 − 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
1 dU(t) = ( iH L†L)dt+dW †L L†dW U(t), (4.2) − − 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) = j (X) = U(t) (X I)U(t) in the space of operators on t ∗ ⊗ 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] + L (X), − L 1 1 L (X) = L†[X,L] + [L†,X]L L 2 2 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,ak∗] = δ jk, where δ jk is the Kronecker delta. Taking the system Hamiltonian and the coupling operator of the system to be, respectively, T quadratic and linear functions of the vector X = a = [a1,...an] ,
H = a†Ma, L = Ca, (4.5)
m n where M is a Hermitian n n matrix, and C C × , the dynamics and output equations × ∈ become
da = Aadt + BdW dy = Cadt + dW, (4.6) 4.2 Background 75
n n n m m n where the complex matrices A C × , B C × , and C C × satisfy ∈ ∈ ∈ 1 A = iM C†C, B = C†. (4.7) − − 2 −
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 tobe 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
G (V) Z†u + u†Z + λ (4.9) P◁W ≤ 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.
n n n m m n where the complex matrices A C × , B C × , and C C × 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 (C a + D u)†u + u†(C a + D u) + λ. ≤ p p p p
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 0. (4.11) B†P C (D + D† ) ≤ − p − p p In the special case, where V = a†a, D = 0 [204] and C = C, this reduces to the p p − condition A + A† 0 ≤ as the condition for passivity. Clearly this condition is satisfied in the case ofan 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 = trace(Vρ), the evolution of ∈ ⟨ ⟩ρ the operator V(t) can be defined via the relation V(t) = V . Note that V = V(0). ⟨ ⟩ρ0 ⟨ ⟩ρt The generator of this Markov process is given by [83]
G (V(t)) = i[V(t),H(t)] + L(V(t)) − J † = i[V(t),H(t)] + ∑ L j (t)V(t)L j(t) − j=1 1 1 L†(t)L (t)V(t) V(t)L†(t)L (t). (4.12) −2 j j − 2 j j
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 K taking the form V = ∑i=1 Vi [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 V consist of { i} 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]. M The multipartite quantum system is defined on H = m=1 Hm which is a tensor product of Hilbert spaces H (each H is associatedN with a subsystem). Unless { m} m otherwise noted, we will the following assumption throughout this chapter. K Assumption 4.2.1 V can be decomposed as V = ∑i=1 Vi, and Vi is defined on a subset of H . V are orthogonal projections, i.e. V 2 = V and [V ,V ] = 0,i = j. Each V is { m} { i} i i i j ̸ i associated with a set of dissipation controls L . Each L allows the decomposition { j} j L j = ∑i Ui, jVi with Ui, j being a unitary operator. Remark 4.2.1 V can be regarded as quasi-local operators [172] since they are de- { i} fined on a subset of H . Therefore, the stabilizing dynamics in this chapter can be { m} considered as specific realizations for the stabilization of DQLS states. V 2 = V 0 is a natural assumption that holds for many applications, e.g. the i i ≥ dissipation control of stabilizer states [143, 185, 139]. V being commutative is an { i} intuitive assumption which enables V to share common ground states. Physical { i} examples of the decomposition includes the dissipation control of graph states [185]. We also make the following assumption. K Assumption 4.2.2 The system Hamiltonian can be written as H = ∑i=1 Hi, where each H satisfies H = V g I. Here g is the smallest eigenvalue of the Hermitian operator i i i − i − i 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 = X X , with X ,X defined on two Hilbert spaces H ,H , respectively. m1 ⊗ m2 m1 m2 m1 m2 Definition 4.2.3 The generator of the evolution of Vi that is induced by a single dissipation control Li is defined by
1 1 G (V ) = i[V ,H] + L†V L L†L V VL†L . (4.13) i Li − i i i i − 2 i i i − 2 i i
Remark 4.2.3 Eq. (4.13) is the generator of Vi controlled by a single coupling operator L . If V is also affected by other coupling operators L , j = i , then we have G (V ) = i i { j ̸ } i V L†V L 1 L†L V 1V L†L G ( i)Li + ∑ j=i j i j 2 j j i 2 i j 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 † ∑ (ViUi, jViUi, jVi Vi) ciVi, ci > 0, (4.15) j=1 − ≤ − 80 Chapter 4. Quantum Reservoir Engineering