University of Calgary PRISM: University of Calgary's Digital Repository
Graduate Studies The Vault: Electronic Theses and Dissertations
2016-01-26 Machine Learning for Designing Fast Quantum Gates
Zahedinejad, Ehsan
Zahedinejad, E. (2016). Machine Learning for Designing Fast Quantum Gates (Unpublished doctoral thesis). University of Calgary, Calgary, AB. doi:10.11575/PRISM/26805 http://hdl.handle.net/11023/2780 doctoral thesis
University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission. Downloaded from PRISM: https://prism.ucalgary.ca UNIVERSITY OF CALGARY
Machine Learning for Designing Fast Quantum Gates
by
Ehsan Zahedinejad
A THESIS
SUBMITTED TO THE FACULTY OF GRADUATE STUDIES
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE
DEGREE OF DOCTOR OF PHILOSOPHY
GRADUATE PROGRAM IN PHYSICS AND ASTRONOMY
CALGARY, ALBERTA
January, 2016
c Ehsan Zahedinejad 2016 Abstract
Fault-tolerant quantum computing requires encoding the quantum information into logical qubits and performing the quantum information processing in a code-space.
Quantum error correction codes, then, can be employed to diagnose and remove the possible errors in the quantum information, thereby avoiding the loss of information.
Although a series of single- and two-qubit gates can be employed to construct a quan- tum error correcting circuit, however this decomposition approach is not practically desirable because it leads to circuits with long operation times. An alternative ap- proach to designing a fast quantum circuit is to design quantum gates that act on a multi-qubit gate. Here I devise quantum control schemes to design high-fidelity single-shot multi-qubit (up to three) quantum gates.
Quantum control task is to steer quantum dynamics towards closely realizing specific quantum operation by varying the external control parameters (external field) such that the resultant evolution closely approximates the desired evolution. A set of instructions that determines the control parameters, and hence the e↵ectiveness of the control scheme, is called a policy. Machine learning algorithms can be employed to find successful policies for designing quantum gates. In particular, we employ supervised machine learning techniques to generate these successful policies.
Finding successful policies is a feasibility problem for which optimization algo- rithms can be employed. Greedy algorithms are at the heart of machine learning tech- niques. They converge faster onto a successful policy and require less-computational resource than non-greedy algorithms. However, there is no guarantee that greedy algorithms succeed to a feasible solution when there exist constraints on i) gate op- eration time ii) computational resources, and iii) experimental resources.
Our results show the failure of standard greedy machine learning algorithms and
ii the superiority of non-greedy machine learning algorithms over greedy ones for de- signing quantum logic gates, when there exist constraints on the quantum system.
We have also observed the failure of existing greedy and non-greedy techniques for designing high-fidelity three-qubit gates. Hence, we devised our machine learning technique called Subspace-Selective Self-adaptive Di↵erential Evolution (SuSSADE).
Each three-qubit gate designed by SuSSADE operates as fast as an entangling two- qubit gate under the same experimental constraints. Preface
We have two published papers [1, 2], the content of which are used in the appropriate sections of my thesis. The main portion of the papers are used in Chapters 5 and
6eitherverbatimorwithsomerequiredmodifications.Inordertokeeptheflowof di↵erent subjects smooth in my thesis, I have used the background and introduction parts of the papers in the early chapters of my thesis. Here I list the two papers which are published based on this work:
1. Ehsan Zahedinejad, Sophie Schirmer, and Barry C. Sanders. Evo-
lutionary algorithms for hard quantum control. Physical Review A,
90:032310, Sep 2014. arXiv.org:1403.0943
2. Ehsan Zahedinejad, Joydip Ghosh, and Barry C. Sanders. High-fidelity
single-shot To↵oli gate via quantum control. Physical Review Letters,
114:200502, May 2015. arXiv.org:1501.04676
In order to make the material taken from our publications and used in my thesis more informative, I have modified and introduced some changes into the context of previously published work. The changes that I made to those materials taken from our publications and used in the body of my thesis are listed as below:
Chapter 2: Section 2.4.3 contains some sentences verbatim (but not • explicitly marked) from the introduction sections of [1, 2].
Chapter 3: Sections 3.3 and 3.11 transcribed from [1]. • Chapter 5: Mostly contains material from [1]. The following changes • are made to the content to make the chapter more informative:
– Whenever found appropriate, a sentence is added to refer
to the introduction and background chapters. All figures
iv in this chapter are cited to the corresponding figures in
the paper.
– In Section 5.1 the first paragraph is completely modified.
The gate operation time T is changed to ⇥ to be consis-
tent with the rest of the thesis. The rest of this section
is unchanged.
– Subsection 5.3.1 is modified to clarify our choice of ex-
ternal pulse. A sentence is added to refer to Section 3.5
for more clarity.
– In Subsection 5.3.2 we have removed the details of evo-
lutionary algorithms. A sentence is added to refer to
Secs. 3.11 and 3.10 in which we discussed the optimiza-
tion algorithms in detail.
– Subsection 5.3.3 is modified slightly by adding a sentence
which refers to the evolution equation 3.4.
Chapter 6 contains some of the material from [2]. All the result for the To↵oli gate are reproduced from [2] with proper citations to the original publication. Other figures are based on our new results which will be published in [3]. The following are the list of the changes made to the content to make the chapter more informative:
Section 6.1: The first paragraph includes the first paragraph of [2] with • some modifications to include two other three-qubit gates (i. e. CNOT-
NOT and Fredkin gates).
Section 6.1: The fourth paragraph copied from [2]. • Subsection 6.4.6 transcribed from [2] and modified extensively to be • more clear. We added a sentence to refer to original DE algorithm in Section 3.11.
Section 6.5 is copied verbatim from [2]. • Subsection 6.6.1: The second paragraph is transcribed from [2]. • Subection 6.8.3 discussion on distortion of control pulses is transcribed • from [2]. Acknowledgements
IwouldliketoexpressmyspecialappreciationtomyadvisorBarrySanders,forhis endless support, enthusiasm, and insight during my PhD studies. I would like to thank him for allowing me to follow my interest in research and to become an independent researcher. Your advice on both research and my career have been priceless. I also like to thank my co-supervisor Dennis Salahub for his subtle guidance on my research on molecular dynamics Simulation.
I am deeply indebted to Joydip Ghosh whose insight and valuable ideas helped me to complete this work. I am also grateful for the expertise of my recent scientific collaborators, Sophie Schirmer, Nathan Babcock. I specially thank Doug Phillips for helping me to become familiar with the world of supercomputers during the first year of my PhD program. For insightful discussions and words of encouragement, I wish to thank Jonny Johannes, Pantita Palittapongarnpim, and Hamidreza Kaviani. I am grateful to Lucia Wang, Nancy Jing Lu, Tracy Korsgaard, and Gerri Zannet for their administration assistance during my PhD studies.
IacknowledgeWestgridComputeCanadaforprovidingcomputationalresources to enable this work. I acknowledge the Murray Fraser Memorial Graduate program and Eyes High International Doctoral Scholarship and support from Natural Sciences and Engineering Research Council of Canada.
I would not be here without the moral and physical support of my great parents
Ghorban Zahedinejad and Nosrat Zahedi. Words cannot express my feelings, nor my thanks for all your sacrifice in my life. I would like to thank my brothers and sister
Ali, Mohammad and Elahe. Just saying thank you will never repay your kindness.
Last but not the least I would like to thank my beautiful wife, Samira Hafezi for her support and care during the stressful time of my PhD studies. Your support and
vii encouragement was in the end what made this dissertation possible. I cannot imagine what would I do without you in my life! Table of Contents
Abstract ...... ii Preface ...... iv Acknowledgements ...... vii TableofContents...... ix List of Tables ...... xii ListofFigures...... xiii List of Symbols ...... xvii 1Introduction...... 1 1.1 Motivation ...... 1 1.2 Research problem ...... 3 1.3 Research objective ...... 4 1.4 Physical models ...... 4 1.5 Approach ...... 5 1.6 Summary of the research achievements ...... 9 1.7 Overview of chapters ...... 10 2QuantumInformationProcessing:APrelude...... 13 2.1 Quantum information theory ...... 13 2.2 Classical and quantum bits ...... 14 2.3 Classical logic gates ...... 16 2.4 Quantum logic gates ...... 17 2.4.1 Single-qubit gates: Pauli matrices and the Hadamard . . . . . 18 2.4.2 Controlled-NOT(CNOT)gate...... 19 2.4.3 Three-qubit gates: To↵oli and Fredkin ...... 20 2.5 Universalquantumgates ...... 22 2.6 Quantum error correction ...... 22 2.7 Fault-tolerant quantum computing and threshold theorem ...... 24 2.8 Quantum noise ...... 24 2.8.1 Decoherence-induced noise ...... 25 2.8.2 Quantum noise and quantum operation ...... 26 2.8.3 Quantumoperationsandenvironment ...... 27 2.8.4 Operator-sum representation ...... 27 3QuantumControl:Background...... 29 3.1 Optimal control theory: Application in quantum systems ...... 29 3.2 Control fields: Time- vs. frequency-domains ...... 31 3.3 Quantum control ...... 32 3.4 Machine learning: A quantum control tool ...... 33 3.5 Controlpulsesaslearningparameters ...... 35 3.6 Learning algorithm confidence: Intrinsic fidelity ...... 37 3.7 Averagestatefidelity ...... 38 3.8 O✏ine vs. online noise consideration ...... 38 3.9 Threshold-fidelity gate: A feasibility problem ...... 39 3.10 Greedyoptimizationalgorithms ...... 40
ix 3.10.1 Nelder-Meadtechnique ...... 40 3.10.2 Quasi-Newton...... 42 3.10.3 Krotov ...... 43 3.11 Evolutionary algorithms ...... 45 3.11.1 GeneticAlgorithm ...... 46 3.11.2 Particle Swarm Optimization ...... 47 3.11.3 Di↵erential Evolution ...... 48 4 Superconducting Artificial Atoms and Gate Design: Background . . . 50 4.1 Quantum computing with superconducting devices ...... 50 4.1.1 JosephsonJunction...... 51 4.1.2 Charge qubit ...... 53 4.1.3 Transmon ...... 54 4.2 Single-Qubit gate design for superconducting artificial atom . . . . . 56 4.3 Two-Qubit (CZ) gate design for superconducting artificial atoms . . . 56 5 HardQuantumControlProblems ...... 60 5.1 Introduction ...... 60 5.2 Criteriatoevaluatealgorithmperformance ...... 62 5.3 Methods ...... 64 5.3.1 Type of control function or hypothesis ...... 64 5.3.2 Optimizing the control function ...... 64 5.3.3 Evaluating the objective function ...... 66 5.4 Qutrit phase gate ...... 66 5.4.1 Qutrit phase gate: Physical Model ...... 67 5.4.2 Qutrit phase gate: Results ...... 69 5.4.3 Qutrit phase gate: Discussion ...... 70 5.5 Controlled-NOT(CNOT)gate...... 71 5.5.1 CNOTgate: PhysicalModel...... 72 5.5.2 CNOTgate:Results ...... 72 5.5.3 CNOTgate:Discussion ...... 75 5.6 Discussion on algorithms performance ...... 76 5.7 Conclusion ...... 78 6 Designing High-Fidelity Single-Shot Three-Qubit Gates ...... 79 6.1 Introduction ...... 79 6.2 Physical model ...... 82 6.2.1 ProjectedHamiltonian ...... 83 6.2.2 Unitaryevolution ...... 83 6.2.3 Phase compensation ...... 84 6.3 Three-Qubit gate design: An avoided-crossing level approach . . . . . 84 6.4 Quantum control ...... 86 6.4.1 Three-Qubit gate design: A quantum control approach . . . . 86 6.4.2 Supervised machine learning: A quantum control tool . . . . . 87 6.4.3 Controlpulsesaslearningparameters ...... 87 6.4.4 Confidence or fitness functional ...... 87 6.4.5 Machine Learning and optimization algorithms ...... 88 6.4.6 Subspace-Selective Self-Adaptive Di↵erential Evolution . . . . 89
x 6.5 Noisemodel...... 92 6.5.1 Amplitudedamping...... 92 6.5.2 Phase damping ...... 93 6.6 Three-Qubit logical gates ...... 93 6.6.1 To↵oli...... 94 6.6.2 Fredkin ...... 95 6.6.3 CNOTNOT ...... 96 6.7 Results ...... 97 6.7.1 To↵oli...... 99 6.7.2 Fredkin ...... 102 6.7.3 Controlled-NOT-NOT (CNOTNOT) ...... 105 6.8 Discussion ...... 109 6.8.1 Controlpulses...... 109 6.8.2 Intrinsicfidelity...... 110 6.8.3 Noise...... 111 6.9 Conclusion ...... 114 7ConcludingRemarks...... 116 7.1 Summary ...... 116 7.2 Hardquantumcontrolproblems ...... 116 7.3 Threshold-fidelity To↵oli gate ...... 117 7.4 Threshold-fidelity Fredkin and CNOTNOT gates ...... 118 7.5 Future work ...... 119 ASubspace-selectiveSelf-adaptiveDi↵erential Evolution (SuSSADE) source code in C++ ...... 143
xi List of Tables
3.1 The truth table representation of CZ gate. T and C denote the control (first) and target (second) qubits, respectively. A CZ gate applies a Pauli-Z operator on the target qubit if the state of the control qubit is 1 and leaves the state unchanged otherwise ...... 35 | i 5.1 Median, best-case, worst-case, and }t for logarithmic intrinsic infi- delity L for the qutrit phase gate with ⇥ =2.5⇡, K =10,andR =80 (R =40)forgreedy(evolutionary)algorithms...... 70 5.2 Median, best-case, worst-case, and }t for logarithmic-infidelity L for the CNOT gate with ⇥ =3.2, K =4,andR =80(R =40)forgreedy (evolutionary) algorithms...... 74
6.1 The truth table representation of CCZ gate. C1 and C2 denote the control qubits and T represents the target qubit. The columns under the Output and Input columns show the states of the three qubits before and after applying CCZ ...... 95 6.2 The truth table representation of Fredkin gate. C1 and C2 denote the control qubits and T represents the target qubit. The column under the Output and Input columns show the state of the three qubits before and after applying the Fredkin gate...... 96 6.3 The truth table representation of CZZ gate. C1 denotes the control qubit and T1 and T2 represent the target qubits. The columns under the Output and Input show the states of the three qubits before and after applying CZZ ...... 98
xii List of Figures and Illustrations
2.1 Bloch sphere representation of a qubit. Each point on the surface of the sphere denotes a pure state and a rotation of the Bloch sphere about any of the axes represents a single-qubit Pauli X, Y , Z operation. 15
4.1 A schematical view of a inductor-capacitor (LC) circuit with L and C denote the inductance and capacitance of inductor and capacitor, respectively...... 51 4.2 (a) A schematic view of a Josephson Junction (JJ) with two super- conducting (S) electrodes separated by a thin insulator (I). (b) Circuit model of a JJ, and (c) equivalent circuit representation which consists of a non-linear element and a capacitive element...... 52 4.3 (a) Circuit representation of a Charge qubit. The dotted red line denotes the isolated charge island. The bias-voltage Vg is used to tune the qubit frequency. (b) The circuit representation of a transmon which is a descendants of Charge qubit. Transmon overcomes the problem of o↵set charge noise by shunting the JJ by an external capacitance. The higher external capacitance in transmon is shown by bigger capacitor plates (red plates) ...... 53 4.4 (a) The energy (E) spectrum of the system Hamiltonian of two capacitively-coupled transmons. The frequency of the first transmon is fixed at 6.5 GHz. The frequency of the second qubit, "2(t)varies from 7.5to6.5GHz...... 58
5.1 Logarithmic intrinsic infidelity L vs iteration number ı for the qutrit gate using the quasi-Newton method with ⇥ =10⇡ (red, solid lines), ⇥ =4⇡ (blue lines with ’+’ markers), and ⇥ =3⇡ (green lines with ’ ’ markers) such that K = 50 in all cases. [reproduced from ref. [1], Fig.⇥ 1a] 69 5.2 (a) Median-run performance (b) Best-run performance for the qutrit phase gate (⇥ =2.5⇡ and K =10).LogarithmicintrinsicinfidelityL vs iteration number ı for (⇤)GA,( )DE,()CommonPSO, (I)PSO1,()PSO2,( )PSO3,( ⇥) quasi-Newton, (V)simplex ⇤ ⌃ B and (O)KrotovwithR =80(R =40)forgreedy(evolutionary) algorithms. [reproduced from ref. [1], Figs. 2a and 2b]...... 70 5.3 Logarithmic intrinsic infidelity L vs iteration number ı for the CNOT gate using the quasi-Newton method with T =30andK =30((red, solid lines), T =10andK =10(bluelineswith’+’markers)and T =4andK =4(greenlineswith’ ’ markers) [reproduced from ref.[1],Fig.1b]...... ⇥ 73
xiii 5.4 (a) Median-run performance (b) Best-run performance for the CNOT (⇥ =3.2andK =4).LogarithmicintrinsicinfidelityL vs iteration number ı for ( ) GA, ( )DE,()CommonPSO,(I) PSO1, ⇤ ⇥ ( )PSO2,(⌃)PSO3,(B) quasi-Newton, (V)simplexand(O)Krotov with⇤ R =80(R =40)forgreedy(evolutionary)algorithms[reproduced fromref.[1],Figs.2cand2d]...... 74
6.1 The energy (E)spectrumofthreenearest-neighbor-coupledtransmons. The first and third transmon frequencies are fixed at 4.8and6.8 GHz respectively. The frequency of the second transmon varies from 4.5to 7.5GHz ...... 86 6.2 The quantum circuit representation of the To↵oli (CCNOT) gate (left), which is equivalent to the CCZ gate up to a local transformation of two Hadamard gates (right). The horizontal solid black lines are
circuit wires, .418pt• shows the control qubits and denotes the Pauli-X operator acts on the target qubit. The boxes with Z and H denotethePauli-ZandHadamardgates...... L ...... 94 6.3 The quantum circuit representation of Fredkin (Controlled-Swap)
gate. The horizontal solid back line is the circuit wire, .418pt• denotes the control qubit and the big cross sign shows the SWAP gate which acts on the target qubits (second and third qubits)...... 96 6.4 The quantum circuit representation of the Controlled-NOT-NOT (CNOTNOT) gate (left), which is equivalent to the CZZ gate up to local Hadamard gates (right). The horizontal solid black lines are
circuit wires, .418pt• represents the control qubit and denotes the Pauli-X operator acting on the target qubit. The boxes with Z and H denote the Pauli-Z and Hadamard operations, respectively.L ...... 98 6.5 Optimal pulses for designing To↵oli gate with the resultant fidelity better than 0.999 and the gate operation time of 26 ns. Transmon frequencies, "i, are varied from -2.5 to 2.5 GHz, which are within the experimental constraint of transmon implementation. The black dots denote the learning parameter for SuSSADE. A) The piecewise- constant pulses for each transmon frequency. B) The piecewise- error-function pulses for each transmon frequency [reproduced from ref.[2],Fig.1]...... 100 6.6 A) The dependence of intrinsic fidelity for the To↵oli gate on the evolution time ⌧ for various values of g. The discretized values show the actual numerical data with ⌃, , ,and⇤ corresponding to the values of g to be 20, 30, 40, 504 MHz, respectively. A cubic interpolation fits the curves to the data. B) The relation between the inverse of the gate operation time and coupling strength between transmons, where the dots denote the actual numerical results for various values of g 20, 30, 40, 50 . A linear-fit interpolates the data points.[reproducedfromref.[2],Fig.2].2 { } ...... 101
xiv 6.7 The fidelity versus the coherence time for the To↵oli gate. The dots denote the actual numerical data and the blue solid line shows a cubic- fit interpolation on the actual data. [reproduced from ref. [2], Fig. 3]. 101 6.8 Intrinsic fidelity of the To↵oli gate as a function of " for CCZ gate. The vertical red dotted-line denotesF the threshold, such that > 0.9999 on the left of the dotted line...... 102 6.9 OptimalF pulses for designing Fredkin gate with the resultant fidelity better than 0.999 and the gate operation time of 26 ns. System frequencies, "i, are varied from -2.5 to 2.5 GHz, which are within the experimental requirements of transmon implementation. The black dots denote the learning parameters for SuSSADE. A) The piecewise-constant pulse for each transmon frequency. B) The piecewise-error-function pulses for each transmon frequency...... 103 6.10 A) The dependence of intrinsic fidelity of the Fredkin gate on the evolution time ⌧ of the system for various values of g. The discretized values show the actual numerical data with , ⇤, , ⌃ corresponding to the values of g to be 20, 30, 40, 50 MHz,4 respectively. A cubic interpolation fits the curves to the data. B) The relation between the inverse of the gate operation time and coupling strength between transmons where the dots denote the actual numerical results for various values of g 20, 30, 40, 50 . A linear fit interpolates the pointstotheactualdata.2 { . . . . .} ...... 104 6.11 The fidelity versus the coherence time for the Fredkin gate. The dots denote the actual numerical data and the red solid line shows a cubic-fit interpolation on the actual data...... 105 6.12 Intrinsic fidelity versus the random noise applied on the optimal pulse of the FredkinF gate. The vertical red dotted-line denotes the threshold, such that on the left side of the line > 0.9999...... 106 6.13 Optimal pulses for designing CZZ gate with the resultantF fidelity better than 0.999 and the gate operation time of 31 ns. System frequencies, "i, vary from -2.5 to 2.5 GHz which are within the experimental constraints of transmon implementation. The black dots denote the learning parameter for SuSSADE. A) The piecewise-constant pulses for each transmon frequency. B) The piecewise-error-function pulses foreachtransmonfrequency...... 107 6.14 A) The dependence of the intrinsic fidelity of the CZZ gate on the evolution time ⌧ of the system for various values of g. The discretized values show the actual numerical data with ⌃, , ,and⇤ corresponding to the values of g to be 20, 30, 40, 50 MHz, respectively.4 A cubic interpolation fits the curves to the data. B) The relation between the inverse of the gate operation time and coupling strength between transmons where the dots denote the actual numerical results for various values of g 20, 30, 40, 50 . A linear-fit interpolates the pointstotheactualdata.2 { ...... } ...... 108
xv 6.15 The fidelity versus the coherence time for the CZZ gate. The dots denote the actual numerical data and the red solid line shows a cubic-fit interpolation on the actual data...... 108 6.16 Intrinsic fidelity versus " for CZZ gate. The vertical red dotted-line denotes the threshold,F such that on the left side of this line > 0.9999.109 6.17 A schematic view of the Closed-Loop Learning Control (ClLC)F technique which uses SuSSADE as a quantum control scheme. Based on this control procedure, the optimal pulses which are generated using SuSSADE fed into a noisy quantum system. if the output met the target fidelity, the procedure aborts otherwise a new set of control pulses is chosen using SuSSADE, and the iteration continues until the target is met...... 114
xvi List of Symbols, Abbreviations and Nomenclature
Symbol Definition U Unitary operator
CNOT Controlled-NOT gate
CNOTNOT Controlled-NOT-NOT gate
K Number of control (learning) parameters
⇥ Gate operation time
GA Genetic Algorithm
PSO Particle Swarm Optimization
DE Di↵erential Evolution SuSSADE Subspace-selectiveSelf-adaptiveDi↵erential Evolution
CZ Controlled-Z gate
ClLC Closed-loop Learning Control
CCNOT Controlled-Controlled-NOT H Hadamard gate
CZZ Controlled-Z-Z gate
CCZ Controlled-Controlled-Z gate
T1 Relaxation time
T2 Dephasing time ⌧ Time evolution of a quantum system
"(t) Amplitudeoftime-dependentcontrolfield
Hˆ dr Drift hamiltonian
Hˆ c Control hamiltonian t Timestep
Utarget Target unitary operator
xvii U[⇥, "(⌧)] Approximatedunitaryoperator
"i Control (learning) parameters Intrinsic fidelity F ¯ Average state fidelity F JJ Josephson Junction
L Logarithmic intrinsic infidelity
CC Cooperative Coevolution
OCT Optimal Control Theory
CPB Cooper Pair Box
DRAG DerivativeRemovalbyAdiabaticGate
xviii Chapter 1
Introduction
“When we get to the very, very small world – say circuits of seven atoms
– we have a lot of new things that would happen that represent completely
new opportunities for design. Atoms on a small scale behave like nothing
on a large scale, for they satisfy the laws of quantum mechanics. So,
as we go down and fiddle around with the atoms down there, we are
working with di↵erent laws, and we can expect to do di↵erent things.
We can manufacture in di↵erent ways. We can use, not just circuits,
but some system involving the quantized energy levels, or the interactions
of quantized spins, etc.”
— Richard P. Feynman, 1960 [4]
1.1 Motivation
Quantum computing [5] enables e cient algorithms for problems that are intractable on non-quantum computers. The power of quantum computing [6] arises by represent- ing and processing information in superpositioned states. Two of the most important quantum algorithms are the Shor factorization [7, 8] and the Grover search [9] algo- rithms with the former yielding an exponential and the latter yielding a quadratic speed-up over their classical counterparts. The extraordinary e ciency of quantum algorithms motivates the e↵ort of building a universal quantum computer to leverage the parallelism power of the quantum algorithms.
One way to perform quantum computing is the circuit-based approach, which is proposed by David Deutsch [10]. A circuit-based quantum computer employs quan- tum logic gates [11] to process quantum information. Quantum logic gates are the
1 building block of a circuit-based quantum computer. They act on qubits (quantum bits) to process quantum information [12]. Designing fast and high fidelity quan- tum gates is the key element to enable scalable quantum computing and quantum error correction. Fast quantum logic gates are also required to design fast quantum processor units [13, 14].
A quantum computing algorithm can be represented by a unitary matrix U of size 2n with n being the number of qubits (quantum bits). Designing a procedure for calculating the output by applying U on the input state involves decomposition into a concatenation [15] of a universal set of quantum gates [16]. The universal class of quantum gates eases the practical implementation of quantum algorithms. Pauli single-qubit gates and Controlled-NOT two-qubit gate comprise one class of universal quantum gates. Experimental progress towards a high-fidelity universal set of gates comprising single- and two-qubit operations has been impressive, exceeding 99.9% for single-qubit [17] and 99% for an entangling two-qubit gates [11].
Quantum information is very fragile against external noise [18]. One way to protect the quantum information against gate imperfection or external noise is to employ quantum error corrections [19]. One can then perform quantum information processing in a code-space and use the quantum error correcting codes to diagnose and remove the errors, thereby avoiding the loss of information. Although error correcting codes can be decomposed into a sequence of single- and two-qubit gates [20], this decomposition approach is not preferable in practice because it leads to quantum circuits [21, 22] with long operation time. Designing quantum gates that act on more than two qubits leads to faster quantum processor units.
2 1.2 Research problem
There are two desirable criteria for any quantum gate design proposal. First, a pro- posed scheme has to enable the design of quantum gates with the shortest possible operation time. Second, it has to be experimentally implementable with less exper- imental resources. The former criterion enables processing of quantum information within timescales smaller than the decoherence time of qubits, thereby maintaining the quantum information. The latter eases the experimental realization of quantum gates with high-fidelity. In the first part of my PhD thesis, I show that, having two constraints of time and experimental resource on the schemes for designing quantum gates, will turn the gate design problems into hard problems. I propose algorithms to overcome this di culty and to enable fast quantum gates that can be implemented with less experimental resources.
Thus far most experimental e↵orts to build a circuit-based quantum computer have been limited to designing high-fidelity single- and two-qubit gates and to per- forming the quantum information processing. However, fault-tolerant quantum com- puting requires encoding the quantum information into logical qubits and performing quantum computing in a code-space [19]. Quantum error correcting codes [23, 24] then must be employed to diagnose and remove the errors. Three-qubit gates are the building blocks of many quantum error correction codes [23, 24, 25, 26]. Designing fast high-fidelity three-qubit gates enables the implementation of fast quantum error correcting circuits, thereby leading to fast quantum processor units.
Although a universal set of single- and two-qubit gates can be employed to decom- pose three-qubit gates into a series of single and two-qubit gates [27, 28], in practice, this decomposition-based approach is undesirable because it leads to quantum cir- cuits with long operation time [21, 22] . In the second part of my thesis, I devise algorithms to design single-shot high-fidelity three-qubit gates without resorting to
3 ine cient decomposition approaches.
1.3 Research objective
In this thesis two objectives are pursued. First, we show that hard quantum gate design problems exist. In particular, we bring two examples of gate design problems including: CNOT [28] and qutrit phase gates [29], which are key elements of standard quantum computing instructions sets for qubits and for qutrits, respectively. It is well- known that, given no constraints on the quantum evolution process, either there is always an evolution path toward a target gate or there will be no path to the target gate [30]. Here our aim is to show that having time and experimental resources as the constraints on the evolution of the quantum system, there will be no guarantee of success for standard schemes to find a successful evolution path. Hence, one should explore alternative approaches or devise novel schemes to designing fast quantum gates upon the failure of existing methods.
The second objective is to devise quantum gate schemes for designing three-qubit gates. The three-qubit gates that we consider here are the To↵oli, Fredkin and CNOT-
NOT, which are the typical three-qubit gates up to a local transformation. Here our aim is to first show that the existing methods fail to design high-fidelity three-qubit gates in the hand, then we propose a new technique to enable the design of fast three-qubit gates.
1.4 Physical models
Our goal to generate policies for designing CNOT and qutrit phase gates is to show the existence of hard quantum gate design problems. Thereby any physical system can be chosen to implement these gates. For the qutrit phase gate one can choose a spin-1 system or single atom with three pertinent electronic levels. For the CNOT
4 gate we have chosen the one-dimensional linear Ising-ZZ model [31]. Physically, the
Ising-ZZ model corresponds to a one-dimensional spin chain, which was originally studied in the context of explaining ferromagnetism.
We limit the design of high-fidelity single-shot three-qubit gates to a specific phys- ical system, namely superconducting artificial atoms. Recent progress in supercon- ducting artificial atoms [32, 33, 34] has made them appealing for quantum information processing, especially for gate-based quantum computing [35]. In particular our su- perconducting artificial atoms are working in the regime of a transmon [36], which is adescendantofachargequbit[37,38].Inordertodesignthree-qubitgates,weonly consider the first four energy levels of the transmons.
1.5 Approach
One approach to designing fast quantum gates is to formulate the task as a control problem and devise policies that successfully achieve the desired outcome. In the context of quantum gate design, the quantum control [39] task is to steer the quantum dynamics towards closely realizing specific unitary operations. Control is achieved by varying the external control parameters such that the resultant evolution closely approximates the desired evolution. A set of instructions that determine the control parameters, and hence, the e↵ectiveness of the control scheme, is called a policy. We use machine learning [40] techniques to generate policies for designing three-qubit gates.
The task of machine learning is to develop algorithms that can learn from system behaviour and predict the future behaviour of the system based on their past evo- lution. Machine learning algorithms have already been applied to various problems in quantum information science, such as phase estimation [41], asymptotic state esti- mation [42], discriminating quantum measurement trajectories and improving read-
5 out [43], and quantum gate design [44]. One can classify the machine learning al- gorithms in three distinct categories [40] namely; supervised learnings, unsupervised learnings and reinforcement learnings. Here we focus on supervised machine learning algorithms [45] as quantum control tools.
The supervised learning task is to infer a function (hypothesis) from the labelled data (training set). The training data comprises an input vector along with its corre- sponding output vector. The supervised learning algorithm trains the hypothesis on the training data to construct an inferred hypothesis. This inferred hypothesis can be further used to label novel data. The idea of supervised learning can be generalized to develop quantum control schemes that deliver successful policies for quantum gate design. In, particular, the learning problem for designing quantum logic gates are re- gression problems for those we don’t require to check the learnt hypothesis for novel data. We only construct the hypothesis to provide the maximum possible confidence for our learning procedure.
A quantum logic gate is a map between an input and an output state. One can always represent the action of any quantum logical gate on the basis elements in terms of a truth table. There is a one-to-one correspondence between the input and output elements in this truth table, as the quantum logic gates are themselves reversible.
In the context of supervised learning problem, we consider this truth table as the training set and consider qubits frequencies to be the hypothesis. Loosely speaking, we train our hypothesis on the truth table data as the training set to generate a successful policy that determines the shape of the external pulses.
Although our machine learning technique can generate any type of external pulses, here we only consider two type of pulses including the piecewise-constant [1] and piecewise-error-function [46]. We chose the piecewise-constant function because it is computationally less-expensive as opposed to expensive realistic piecewise-error func-
6 tions. We use the piecewise-constant pulse when we analyze the e↵ectiveness of our quantum control schemes against the physical parameters of the system. We employ the piecewise-error function to provide a realistic implementation of the policy for de- signing three-qubit gates. We provide the numerical evidence that the performance of our quantum control schemes do not depend on the shape of the pulse but on the number of learning (control) parameters.
In the case of a piecewise-constant pulse, we discretize each control pulse (hypoth- esis) and express it as the sum of K orthogonal constant functions over the interval
[0,⇥]. In our machine learning approach, K denotes the number of learning param- eters or experimental resources to shape the external pulse. We construct a smooth pulse (piecewise-error-function) by connecting each of the control parameters using an error function. We measure the performance of our learning algorithm by taking the distance between the target and approximated operators. Minimizing this distance will increase the confidence of our learning procedure.
According to the threshold theorem [19], if the accuracy in designing quantum gates exceeds the threshold for fault-tolerant quantum computing, quantum error correction codes can be used to perform arbitrarily long quantum computation. The bounded error on the accuracy of designed quantum gates implies that the generated policies for designing quantum gates do not need to lead to a perfect target gate.
The generated policy only need to approximate the target gate within some error which is acceptable by fault-tolerant quantum computing. This bounded error turns the problem of finding a successful policy for designing high-fidelity quantum gates into a feasibility problem. Any policy that results in approximated unitary operation within threshold fidelity will be a successful (feasible) policy. Therefore, I use the term feasible policy and successful policy interchangeably throughout this thesis. One can employ optimization algorithms to find feasible policies for designing fast quantum
7 gates.
To search for a successful policy, we have used several optimization routines in- cluding the greedy and nongreedy (stochastic) algorithms. For greedy algorithms we employ three well-known optimization techniques, namely Simplex [47], quasi-
Newton [48] and Krotov [49, 50, 31]. We choose these three algorithms because of their superior performance over other greedy algorithms in many quantum control problems [51, 52, 48, 31, 53]. There is a large class of non-greedy algorithms but we choose to test just the three most common or promising evolutionary algorithms, namely the traditional Genetic Algorithm (GA) [54] (a commonly used algorithm) and the modern particle swarm optimization (PSO) [55] and Di↵erential Evolution
(DE) [56] algorithms (promising for this type of problem). The promising nature of
PSO and DE is based on many studies that have shown the superiority of DE and
PSO over other evolutionary algorithms [57, 58, 59].
For all the gate design problems in this thesis we first try the greedy algorithms for finding a successful policy. Greedy algorithms are fast in converging on a fea- sible policy when performing a local search. Greedy algorithms are not guaranteed to succeed (i) when the search domain is non-convex or (ii) when the computational resources (model dimensionality) or the time for performing the control task is con- strained [60, 61, 1]. Upon the failure of greedy algorithms on specific quantum control tasks we try the three most promising evolutionary algorithms. Generally, evolution- ary algorithms fail when the number of learning parameters increases, or decoherence and loss are included. We have devised a new optimization routine that succeeds to generate a successful policy when existing optimization algorithms fail.
For realistic implementation of quantum gates, we follow the standard practice of gate design by first ignoring the noise and generating a successful policy to designing gates for closed quantum systems. We then evaluate the performance of the generated
8 policies under the random noise on the learning parameters. We also evaluate the performance of successful policies in the presence of decoherence-induced noise. In order to incorporate decoherence into our systems evolution, we model each qubit in the physical model as a harmonic oscillator su↵ering from an environmental e↵ect.
1.6 Summary of the research achievements
We have succeeded to generate successful policies to design CNOT and qutrit phase gates [1]. For these two problems, we have shown that if one evolves the quantum sys- tem over a long time with a large number of control parameters, the greedy machine learning algorithms can always find a successful policy to design threshold-fidelity gates. Reducing the control parameters and shortening the operation time turn the problem of designing CNOT and qutrit phase gates into hard quantum control prob- lems. Under the constraints of operation time and experimental resources (i. e. control parameters), greedy machine learning techniques fail to generate a successful policy for designing high-fidelity quantum gates. We then used the non-greedy machine learning techniques as quantum control tools to search for an feasible policies. We have shown the superiority of evolutionary machine learning algorithms, in particular, the DE-based machine learning, in finding the successful policies even in the presence of constraints on operation time and experimental resources.
We have proposed a new machine learning algorithm, which is suited for high- dimensional learning problems [2]. Our machine learning algorithm is based on a variant of DE, which is called Subspace-Selective Self-Adaptive DE (SuSSADE). SuS-
SADE has been developed as a modular optimization engine, which takes the objective function and constraints on the learning parameters as input and searches for an opti- mal solution for the problem. SuSSADE is capable of solving non-convex optimization problems, because it globally searches for the optimal solution. We have developed
9 our machine learning algorithm such that it runs on single or multiple processors.
We have employed SuSSADE to generate policies for designing high-fidelity single- shot three-qubit gates for a linear chain of three capacitively coupled transmons. The three-qubit gates that we designed are To↵oli, Fredkin and CNOTNOT. Each of the three-qubit gate operates as fast as an entangling two-qubit CZ gate under the same experimental constraints. Here in order to design the high-fidelity three-qubit gates, we first test the standard machine learning algorithms. We have observed that existing machine learning techniques failed to achieve any fidelity better than
99%. We then proposed our evolutionary machine learning algorithm to overcome this challenge. The generated policies from SuSSADE led to the three-qubit gates with fidelity better than 99.9%.
We have tested the generated policies for designing three-qubit gates in the pres- ence of noise. Our analyses show that our policies are resilient against the random noise on the learning parameters within the experimentally prescribed random noise on the external pulse [62]. We have also tested our policies against decoherence- induced noise. The resultant fidelity remains above the threshold even in the presence of decoherence-induced noise. Our policies to design threshold-fidelity gates account for the leading order of noise that causes distortion on the learning parameters. We have proposed a Closed-loop Learning Control (ClLC) to suppress the higher order noise on the quantum system.
1.7 Overview of chapters
This chapter briefly explained the objectives pursued in the course of my PhD stud- ies. I have summarized the problem statement, the approach that I took to tackle the problems and the achievements of solving each problems. This thesis includes seven chapters. The interdisciplinary nature of my PhD research requires the knowledge in
10 four fields including quantum information, quantum control, superconducting artifi- cial atoms, and machine learning. Therefore, the next three chapters establish some familiarities with these topics.
In order to establish the framework for quantum gate design, Chapter 2 gives an introductory on quantum information. I introduce qubits which are the basic unit of information in quantum computing. I discuss the action of quantum gates on qubits basis and explain the importance of designing fast quantum logic gates. I explain the importance of quantum error correction for fault-tolerant quantum computing. I conclude the chapter by giving the mathematical model for noise, which we used to incorporate decoherence into our quantum systems.
Chapter 3 includes the mathematical foundation of quantum control. In this chapter I explain how to formulate a gate design problem into a quantum control problem. Then I formulate a quantum control problem into a supervised machine learning problem. Improving the performance of a learning procedure is equivalent to
finding a feasible solution for a feasibility problem. Hence, I discuss the possible choice of optimization algorithms that can be employed to tackle the feasibility problem.
In Chapter 4, I introduce the superconducting artificial atoms, which are the physical system for designing three-qubit gates. I start with the earliest proposed superconducting qubit, the charge qubit, and then discuss the transmon, which is the improved version of charge qubit. I review the existing proposals for designing high-
fidelity single- and two-qubit gates for the architecture of superconducting artificial atoms.
In Chapter 5, I present the results of our early work [1] on designing two important quantum control gates including CNOT and qutrit phase gates. I show that hard quantum control problems exist upon the existence of constraints on the quantum system. I show the performance of several machine learning techniques for generating
11 these two quantum gates.
Chapter 6 presents our results on generating successful policies for designing high-
fidelity single-shot three-qubit gates including: To↵oli, Fredkin and CNOTNOT [2,
3]. We show the success of our machine learning technique in designing three-qubit gates, which are as fast as an entangling two-qubit gate under the same experimental constraints. We discuss the performance of our policies in the presence of di↵erent types of noise in Chapter 6.
In Chapter 7, I summarized the works that I have accomplished in this thesis. I suggest some of the further works that can branch out from this thesis.
12 Chapter 2
Quantum Information Processing: A Prelude
“The entire universe must, on a very accurate level, be regarded as a
single indivisible unit in which separate parts appear as idealisations
permissible only on a classical level of accuracy of description. This
means that the view of the world being analogous to a huge machine,
the predominant view from the sixteenth to nineteenth centuries, is now
shown to be only approximately correct. The underlying structure of
matter, however, is not mechanical. This means that the term ‘quan-
tum mechanics’ is very much a misnomer. It should, perhaps, be called
quantum “nonmechanics”.”
— David Bohm, 1951 [63]
2.1 Quantum information theory
The theory of quantum computing has been established initially by a number of pi- oneers in the field. Richard Feynman [64] was the first who proposed his concept of a Quantum Computer. A machine which can e ciently simulates a quantum sys- tem as opposed to a classical Turing machine [65] which faces an exponential slow down. David Albert contributed to the field by showing how a quantum mechani- cal automaton remarkably outperforms a classical automaton [66]. However David
Deutsch [67, 10] was the one who made the key discovery by proving that “There exists or can be built a universal quantum computer that can be programmed to perform any computational task that can be performed by any physical object.”
The quantum computer proposed by David Deutsch [10] is a circuit-based quan- tum computer which is one of the approaches to performing the quantum comput-
13 ing. A circuit-based quantum computer employs quantum logic gates to process the quantum information. Quantum logic gates are the building blocks of a quantum computer. Designing fast quantum gates is the key factor to designing fast quantum processor units. The primary goal of my thesis is to devise quantum control schemes to design fast quantum logic gates. A prelude of quantum information comes in the following sections to establish some familiarity with the subject.
2.2 Classical and quantum bits
Data are the distinct pieces of information. In classical information theory, a bit is the smallest distinguishable unit of data. In practice, a classical bit is not a mathematical abstract of information but any binary physical device that can exist in two well- defined, coherent, and mutually exclusive states – either 0 or 1. The state of the physical systems can be found by performing measurement on the device, without a↵ecting the current state of the system (Or the system can be fully initialized to its initial state upon a measurement on the system).
In the quantum information theory, a quantum bit which is usually refereed to as a qubit is the smallest distinguished unit of data. Just as a classical bit, a qubit has two states – 0 and 1 . In quantum mechanics ‘ ’ is called the Dirac notation that | i | i | i is the standard notation to represent the state of the quantum system. Qubits have privilege over the classical bits because they do not need to be in mutually exclusive states. In other words, it is possible to have a linear combinations of states, called superposition:
= ↵ 0 + 1 (2.1) | i | i | i with ↵ and being complex numbers. It is clear from 2.1 that the qubit is a vector in a two-dimensional complex vector space with states 0 and 1 known as computational | i | i basis states.
14 ˆz 0 | i
| i
✓
yˆ
xˆ
1 | i
Figure 2.1: Bloch sphere representation of a qubit. Each point on the surface of the sphere denotes a pure state and a rotation of the Bloch sphere about any of the axes represents a single-qubit Pauli X, Y , Z operation.
Unlike a classical bit that can be examined through a measurement with certainty, we cannot perform measurement on a qubit such that the output of the measurement gives us both ↵ and . This is because a measurement projects the qubit state into one of its basis with certain probabilities. Upon a measurement on a qubit we get either 0 with probability ↵ 2 or 1 with probability 2 with the sum of these | i | | | i | | probabilities equal to one.
Similar to a classical bit, a qubit is not a mathematical abstract but a quantum- mechanical system (physical device) that can exist in two well-defined states. Many physical systems can be employed to realize qubits. For example, the vertical and horizontal polarization of a photon can represent a qubit [68]. Nuclear spin in a magnetic field is another physical system that can be used to realize a qubit [69].
The quantum state of a two-level system (qubit) can mathematically be best represented by a Bloch sphere (Fig. 2.1) which is a unit 2-sphere. It is convenient to consider the north and south poles of the sphere to correspond to unit vectors 0 | i
15 and 1 , respectively. Then one can write the qubit state as: | i ✓ ✓ =cos 0 + ei sin 1 (2.2) | i 2 | i 2 | i with 0 ✓ ⇡ and 0 2⇡ and
1 0 0 = 2 3 , 1 = 2 3 (2.3) | i 0 | i 1 6 7 6 7 4 5 4 5 are two dimensional real vectors.
Qubits are very fragile against external noise. Usually in quantum computing, the time that a qubit maintains its initial information without being a↵ected by the en- vironment is called the coherence time of that qubit. The coherence time for a qubit depends on the physical system implementation of that qubit. For example, super- conducting qubits, in particular transmons, loose their information over a timescale of
60µs (See 4.1.3 for more detail). Decoherence places the time constraint on quantum computing devices to perform computational tasks in timescales much shorter than the coherence time of a qubit. Therefore, e↵orts are divided into two main domains to avoid information loss resulting from decoherence, a) Designing qubits with long decoherence times and b) designing fast quantum logic gates which operate in faster time than the qubit coherence time. The latter is the topic of the current thesis.
2.3 Classical logic gates
George Boole was a British mathematician who started to explore the concept of logic gates in his e↵orts to formalize the laws of thought [70]. His main interest was to establish the concept of reasoning in mathematics. He introduced the Boolean algebra in his book of ‘The Mathematical Analysis of Logic’ [71]. Boolean algebra is to perform algebra not in a quantitative method but in terms of variables that take only discrete binary values like ‘0’ or ‘1’.
16 The primary application of Boole’s algebra is in fabricating classical electronic devices called logic gates, which can perform Boolean logic on classical bits. A logic gate is a physical device, like a transistor, which takes one or more inputs as 0s and
1s and output the result as a single or more logical variables. The AND and OR are two examples of logical gates.
The AND, and OR gates are examples of irreversible logic gates. The term ir- reversible stands for those logical gates whose inputs cannot be retrieved from their corresponding output. This loss of information is one of the sources of heat genera- tion in the current classical computers. On the other side there is a set of reversible classical gates that for those the initial information processed can be retrieved via the output of the reversible logic gates. The NOR gate and the To↵oli gate are examples of reversible logical gates [72].
In order to make the engineering design of a computational unit simpler, there exists a specific set of logic gates, “universal gates” for which the logic circuit of any computational task can be decomposed into a series of universal logic gates. The
NAND gate or (AND, OR, NOT) gates are examples of classical universal logic gates which means any classical logical function can be computed with a series of these gates.
2.4 Quantum logic gates
Given an introduction of what a logic gate does in a classical computer now we can have a better grasp of what a quantum logic gate does in a quantum computer! A quantum logic gate acts on qubits to process quantum information. They take the state of one or more qubits as input and return the output as the state of one or more qubits. For example, if the qubit information is encoded in an atomic or nuclear spin, the external magnetic field is being used to manipulate the qubit information.
17 Another example would be the energy levels of a atom which can serve as a qubit system. One can manipulate the encoded information via emitting external pulses on the atomic system. Therefore, similarly to a classical logic gate, a quantum logic gate is a physical implementation of a quantum system which evolves based on quantum mechanical laws.
Since the quantum logical gate itself is a quantum system, we can provide a mathematical representation of its evolution based on the laws of quantum mechanics.
Schr¨odinger’s equation is the principal equation that governs the transformation of a quantum system from some initial to a final state. Converting the external | ii | if forces acting on the quantum system into energy quantities we can represent the total energy of a quantum system by Hˆ , the Hamiltonian of the system. Hence, we can write the time-dependen, Schr¨odinger equation as follow:
@ i~ | i = Hˆ (2.4) @t | i with ~ denoting the reduced Planck’s constant. Although there are many quantum gates in the context of quantum information processing, here I mostly focus on those gates that enable universal quantum com- puting and play a significant role in quantum error correction. I discuss the physical action of each gate and show their mathematical representation in the qubit basis.
2.4.1 Single-qubit gates: Pauli matrices and the Hadamard
Single-qubit gates are those logical gates that act on one qubit. A special set of single- qubit gates is the set of Pauli matrices (1, X, Y , Z), which are unitary and Hermitian operators. Pauli matrices are special because any single-qubit Hamiltonian can be written as the weighted sum of these matrices. Pauli matrices can be represented in
18 the basis of a single qubit ( 0 , 1 )asfollows[19]: | i | i
10 01 0 i 10 1 = 0 1 ,X= 0 1 ,Y= 0 1 ,Z= 0 1 . (2.5) 01 10 i 0 0 1 B C B C B C B C @ A @ A @ A @ A The Hadamard gate is another single-qubit gate. It maps 0 1 ( 0 + 1 )and | i! p2 | i | i 1 1 ( 0 1 ). There is no classical counterpart for the Hadamard gate, because | i! p2 | i | i it generates a quantum state, which is a superposition of 0 and 1 upon its operation | i | i on these states. Based on the action of the Hadamard gate on a qubit basis one can write its mathematical representation as follows:
1 11 H= . (2.6) 2 0 1 1 1 B C @ A Here I only discussed those single-qubit gates that I will use in the following chapters.
An interested reader should refer to [19] for further discussion on other single-qubit gates.
2.4.2 Controlled-NOT (CNOT) gate
Controlled-NOT (CNOT) is an entangling two-qubit gate which operates on two qubits. A CNOT gate applies the Pauli-X operator on the second (target) qubit if the first (control) qubit is at state 1 , otherwise leaves the state of the target | i qubit unchanged. A CNOT gate is equivalent to a Controlled-Z (CZ) gate under a
Hadamard transformation on the target qubit. A CZ gate applies a Paul-Z operator on the target qubit, if and only if the state of the control qubit is 1 . The mathemat- | i ical representation of CNOT and CZ gates in the computational basis of two qubits
19 ( 00 , 01 , 10 , 11 )canbewrittenasfollow: | i | i | i | i
1000 100 0 0 1 0 1 0100 010 0 CNOT = B C , CZ = B C . (2.7) B C B C B0001C B001 0C B C B C B C B C B0010C B000 1C B C B C @ A @ A 2.4.3 Three-qubit gates: To↵oli and Fredkin
AquantumTo↵oli gate is an essential element for (non-topological) quantum error correction [6, 7] and the key component for reversible quantum computing [2, 15].
The quantum To↵oli gate is to e↵ect a three-qubit controlled-controlled-NOT (CC-
NOT) gate, which means that the third (target) qubit is flipped only if the first two (control) qubits are in the 1 state and not flipped otherwise. Constructing the | i mathematical representation of the To↵oli is simple given the action of the gate on three qubits. One can write this mathematical representation in the computational basis ( 000 , 001 , 010 , 011 , 100 , 101 , 110 , 111 )ofthreequbitsasfollow: | i | i | i | i | i | i | i | i
10000000 0 1 01000000 B C B C B00100000C B C B C B00010000C B C CCNOT = B C (2.8) B C B00001000C B C B C B00000100C B C B C B C B00000001C B C B C B00000010C B C Similar to the equivalence relation@ between two-qubit CNOTA and CZ gates, the
CCNOT and CCZ operations are equivalent under the local transformation on the
20 target qubit:
CCNOT = [1 1 H] CCZ [1 1 H] (2.9) ⌦ ⌦ ⌦ ⌦ where H is the single-qubit Hadamard gate. A CCZ gate applies a Pauli-Z operator on a target qubit, if and only if the state of the control qubits is 11 and leaves the | i target state unaltered otherwise.
The Fredkin gate is another important three-qubit gate which enables reversible quantum computing [19]. The Fredkin gate is also called a controlled-swap operator, because it applies a swap gate on the quantum state of the second and third qubits if the state of the first qubit is 1 , and leaves the quantum state unchanged otherwise. | i The Fredkin gate is an excitation-number-preserving operator which means the output state has the same number of excited states as the input state. It is also a self- inverse operation which means applying two consecutive Fredkin operations gives the same output as the input. The Fredkin gate can be configured to function as other single logical quantum qubits [73]. The logical action of the Fredkin gate can also be represented as a matrix in the computational basis of three qubits as follow:
10000000 0 1 01000000 B C B C B00100000C B C B C B00010000C B C Fredkin = B C (2.10) B C B00001000C B C B C B00000010C B C B C B C B00000100C B C B C B00000001C B C @ A
21 2.5 Universal quantum gates
As I explained earlier in 2.3, there is a small set of classical gates (AND, OR, NOT) that are universal for classical computation. A similar universality is true for quantum computation. A set of gates is universal for quantum computation if any unitary operation can be decomposed into a series of only those gates. In quantum computing there are a few sets of universal gates, with the three following are the most important ones [19]:
1. Controlled-phase (CZ), and single-qubit Pauli X, Y and Z spin–half
matrices.
2. To↵oli and Hadamard gates.
3. Fredkin and Hadamard gates.
In principle one set of universal gates such as single-qubit and CZ gates is su cient for quantum computing. However an outstanding problem is that many quantum algorithms [21, 22, 7] require a gate acting on at least three qubits. Although three- qubit gates can be constructed by a decomposition into a series of single- and two- qubit gates [27, 28], in practice, this approach is not desirable since it leads to quantum circuits with long operation times [21, 22]. Designing high-fidelity single-shot three- qubit gates is one of the goals of this thesis.
2.6 Quantum error correction
Noise is a great challenge for quantum information processing. Ideally we want to build the quantum system such that noise can be avoided completely. In practice, this ideal situation fails to exist and one should devise methods to protect the system against the e↵ects of noise.
22 The basic idea of protecting information against noise is to encode the information by adding some redundant information to the initial information. Under the condition that some of the encoded information being corrupted by the noise, there exists enough redundancy in the encoded information that the initial information can be retrieved successfully. Protecting quantum information against noise is based on the same ideology.
The basic idea of quantum error correction is inspired by the ideas introduced by Shor [24]. In general, the procedure of error correction includes three distinct steps [19, 24]. First, one encodes the quantum information (initial information) into a quantum error-correcting code. This step is usually done by using the unitary operation defined by the error correcting algorithm. The encoded information resides in a larger Hilbert space than the Hilbert space for the initial information to ensure enough redundancy for restoring the initial information. The space in which encoded information resides is called code-space and the eigenstates of the code-space are called logical states. Quantum information processing is being performed in the code-space on the logical states. In the second step, after logical states are subjected to noise, one should perform the syndrome measurement to diagnose the type of the error on logical states. In the last step we perform a recovery operation to retrieve the initial information and transform the quantum system into its original state of the code.
Two examples of well-known quantum error correcting codes are Calderbank–
Shor–Steane algorithm [19] and Laflamme et al. error-correcting code [23]. The former encodes each physical qubit (initial information) into a block of seven qubits, which are called logical qubits. The latter encodes each physical qubit into a block of five logical qubits. Both algorithms account for an arbitrary error on a single qubit, with the Laflamme et al. error-correcting code proposing the minimum number of logical qubits for performing error correction.
23 2.7 Fault-tolerant quantum computing and threshold theorem
Of course my discussion on quantum error correction codes in previous section is based on this assumption that encoding the quantum information into the code-space and decoding the information out of it can be done perfectly. In practice this simplified assumption can not be made, because quantum gates, which are the building blocks of quantum error correcting circuits can not be designed perfectly. However the theory of fault-tolerant quantum computing [24] comes to the rescue by removing the assumption of having perfect encoding and decoding for having fault-tolerant quantum computation.
As the quantum information dynamically undergoes computation in the code- space, quantum error correcting codes have to be recursively applied to protect the quantum information against decoherence. Given the fact that quantum logic gate are the building block of quantum algorithms and perfect design of quantum gate is challenging, the main question is how the quantum computation can be performed correctly (without the error accumulated over the computation) when the quantum error-correcting codes themselves are faulty? Remarkably there is a celebrated theory called “threshold theorem” which says provided only that the error probability per gate is below a certain constant threshold, quantum error correction can be employed to perform any arbitrarily long quantum computation [19].
2.8 Quantum noise
Closed quantum systems, are those quantum systems that do not su↵er from un- wanted interactions with environment. Although studying closed quantum systems deepen our understanding of such ideal systems, in practice, a quantum system cannot be isolated from interaction with outside world. In fact, there is no closed quantum system except the universe itself. The unwanted interactions of the quantum system
24 with environment appears as noise in quantum information processing systems. In order to build practical and useful quantum processor unit, one needs to understand noise on the quantum system. There are several types of noise which can a↵ect the quantum system. Here we discuss the decoherence-induced noise and the existing approaches that can model this type of noise.
2.8.1 Decoherence-induced noise
The decoherence-induced noise comes from the fragility of quantum information against the interaction with the environment. In general if a qubit is initialized in a superposition of its eigenstates ( 0 , 1 ), there is no guarantee that it stays in | i | i this superposition forever because of external noise acting on the qubit system. This fragility of information to external noise is also present in classical information pro- cessing. This is the reason that several layers of classical error correction codes are applied on classical computing and storage devices like CDs, Hard Drive and RAMs to protect the loss of information against the external noise. Quantum information processing devices can also exploit error correcting codes to protect their information; however, a deep knowledge of various types of noise is essential for designing e↵ective quantum error correction schemes. Here I discuss the decoherence-induced noise and its e↵ect on quantum information.
In the quantum information processing the decoherence-induced processes are labelled by their e↵ective timescales. The Bloch sphere representation (Fig. 2.1) of qubit is the best picture to explain the e↵ects of these processes. In the Bloch representation, T1 processes that move the state of the qubit toward the poles are called relaxation time, and T2 processes that rotate the Bloch vector around the vertical axis (axis connecting the ground and excited levels) are called dephasing.
In the following we review some backgrounds on the noise model that we use to incorporate decoherence-induced into our quantum system.
25 2.8.2 Quantum noise and quantum operation
The quantum operations [19] provides a solid mathematical formalism to describe the dynamics of open quantum system. In the context of quantum operations one
first describes the initial state of the quantum system in terms of a density operator
(density matrix), ⇢ and follow the system evolution according to the following map:
⇢0 = (⇢)(2.11) E
The map, , is a quantum operation and (⇢)isthefinalstateaftertheprocess E E occurs (up to some normalization factor).
There exist three distinct methods of understanding the quantum operations, all of which are equivalent [19]. In the first method we study the dynamics that caused by the interaction of a system with an environment. While this method is a physically sound approach to understand the dynamics of the system under noise, it is not mathematically convenient. Operator-sum representation is the second method that can be employed to study the quantum operation. In contrast to the first method, the operator-sum representation is an abstract method, however, it provides a solid mathematical framework for theoretical study of an open quantum system. The last and third method is based on a set of physically motivated principles under which a dynamic map in quantum mechanic evolves. I will not discuss the third method, as it lacks the mathematical strength of the second method and the physical concreteness of the first method.
In order to incorporate the e↵ect of noise in our simulation of quantum system,
I employ the operation-sum representation method. However before providing the detail of this method, an understanding of how the system and environment interact with each other in a open quantum system is required. Then, in the following, I first explain the mathematical representation of this interaction and discuss the operation- sum representation afterward.
26 2.8.3 Quantum operations and environment
One way of considering noise in an open quantum system is to follow the evolution of the system as it interacts with environment as another system. Since we consider the environment as another system, for the sake of clarity, I call our system of interest as the principal system, which with environment comprises a closed quantum system.
We represent the evolution of this closed-system by U. Let assume that the principal system and the environment are initially at a product state ⇢ ⇢ . Thereby, after ⌦ env the unitary transformation U acting on the system, the principal system will no longer interact with environment. Thus, one can perform a partial trace (Trenv)overthe environment to obtain the state of the principal system as follow:
(⇢)=Tr U(⇢ ⇢ U † (2.12) E env ⌦ env ⇥ ⇤ Equation 2.12 represents the quantum operation in terms of the interaction of the system with an environment, thereby giving the first of three equivalent definitions of quantum operation. One issue with using (2.12) as a representation of the quantum operation is how to specify U when the environment has many degrees of freedom. It seems the operation-sum representation can address this concern in an e↵ective way.
2.8.4 Operator-sum representation
One can represent the quantum operations in terms of the operator-sum represen- tation [2]. In order to present the quantum operation in terms of operator-sum, let e be an orthonormal basis for the environment with an infinite number of modes. | ii We consider U as the unitary operation of the environment and the principal system, which together forms a closed quantum system. Assuming the initial state of the en- vironment is ⇢ = e e (if not we are free to introduce an extra system purifying env | 0ih 0| the environment), one can decompose (2.12) in terms of the operators (which act Qi
27 on the state space of the principal system) as follow:
(⇢)= †⇢ (2.13) E Qi Qi i X where e U e . Equation 2.13 is called the operator-sum representation of Qi ⌘hi| | 0i and are known as operation elements for the quantum operation . The E {Qk} E operation elements satisfy the completeness relation as follow:
† i = 1 (2.14) Qi Q i X with 1 representing the identity matrix. Equation 2.14 is satisfied by the quantum operations which are trace-preserving.
It can be shown that the operation-sum approach for considering noise is equiva- lent to the first method that I explained in Subsec. 2.8.3 (See [19] for more details).
Therefore, I employ the operation-sum approach to perform numerical simulation of noise in our quantum system. Performing this numerical calculation is straightfor- ward once we know the mathematical representation of the operation elements. In chapter 6, I give the explicit mathematical formulas of operation elements when I discuss decoherence-induced noise on three-qubit gate quantum systems.
28 Chapter 3
Quantum Control: Background
“People worry that computers will get too smart and take over the world,
but the real problem is that they’re too stupid and they’ve already taken
over the world.”
— Pedro Domingos, 2015 [74]
In this chapter I first start with the application of Optimal Control Theory (OCT) in quantum control. Then I clarify the choice of our control field to control the quantum system. I also discuss the quantum control in the context of gate design problems. Then I introduce the supervised machine learning approach as a quantum control tool. I explain how a gate design problem can be turned into the learning problem for finding a successful policy. The problem of finding a successful policy is afeasibilityproblemforwhichoptimizationalgorithmsmustbeusedtotacklethe problem. I discuss several optimization algorithms that I used in the latter chapters to generate the successful policies for designing fast quantum logical gates.
3.1 Optimal control theory: Application in quantum systems
The task of OCT is to redirect the control process by maximizing an objective func- tional via some external control fields, which are usually expressed as functions in time or frequency-domains [75]. OCT has been successfully applied in controlling many quantum systems including the selective breaking of chemical bonds [76], the creation of ultrafast semiconductor optical switches [77], electron transfer in biological systems [78], and quantum information processing [79].
Depending on the type of control problem, there is a variety of objective functional
29 choices that can be defined for controlling a quantum system. However, in most cases studied so far [30], the objective is to maximize the transition probability of the system
2 Pi f = Uif , (3.1) ! | | with
U = f U i . (3.2) if h | | i Here, U = U("(t)) is a time-dependent evolution operator, which is a functional of the control function "(t). In (3.1), complete controllability of the control system implies that Pi f =1andPi f < 1denotesthatthesystemisonlypartiallycontrollable. ! ! Typically, an OCT problem involves an optimization over a large number of (from ten to a thousand) control parameters which result from the discretization of the con- trol function over the time- or frequency-domain. Searching the optimal solution over such a large dimensional space of control parameters requires e↵ective optimization routines. Greedy (local) algorithms are largely used in the context of OCT, because they require less computational resources and converge faster towards a successful solution as opposed to non-greedy (global) algorithms, which usually require more computational resources with slower convergence rate.
In spite of the high dimensionality of the quantum control problems and the com- plexity of the map between the system evolution and the control function, greedy algorithms are remarkably e↵ective in delivering almost perfect control schemes for controlling the quantum system. This remarkable performance gives rise to this ques- tion: under what conditions can greedy algorithms be used reliably for solving the quantum control problem?
It is shown in [30] that greedy algorithms can easily yield excellent-quality control schemes for a quantum control problem under the absence of any constraint on the external control field. Under this assumption “no suboptimal local extrema exist as
30 a trap when seeking optimal quantum control” [30]. Although there exist various classes of extrema in the quantum control landscape with distinct ease of discovery and robustness properties, they all have peaks with the same height, all resulting in excellent-quality control schemes (See for example Fig. 1 in [30]).
Naturally, a variety of issues, such as quantum noise, the precision of the control devises, random noise on the control field and physical constraints on the control parameters could all invalidate the assumption that all the quantum control problems can be solved using the local optimization routines. Inspired by the above discussion we aim to explore the performance of greedy algorithm on quantum gate design under the constraints on system evolution and control field [60, 61]. Upon the failure of standard greedy algorithms we propose state-of-the-art evolutionary algorithms to enable the design of fast high-fidelity quantum gates.
3.2 Control fields: Time- vs. frequency-domains
Application of OCT in quantum system was initially motivated by controlling chemi- cal reactions [76, 78]. The initial idea was to control the output of a chemical reaction by tuning the magnitude of an external control function "(t). In terms of the type of the control function, there are two di↵erent but equivalent approaches, which are followed in the field [80, 81]. The first one is having the control function in the time-domain and the second one is in the frequency-domain.
Time- and frequency-domains are two ways of studying the same dynamic system.
They can be interchangeably used to control the quantum system. The interchange- ability means that no information will be lost upon transformation from one domain to another. In the time-domain we are interested to studying the evolution of a pro- cess element over time, whereas the frequency-domain we study how fast or slow a process occurs. Therefore these two di↵erent domains provide the tools to look at the
31 same physical phenomena di↵erently.
The control procedure in the time-domain was first proposed by Tannor and
Rice in 1985 [80]. The control procedure can be explained based on a two-color pump/probe scheme, where the control is the time delay between the two produced pulses. This time delay between pulses designed such that the second (dump) pulse comes into resonance with the evolution of a pump-excited molecule and directs the whole system toward a target product Frank-Condon region. Experimental imple- mentations of the Tannor-Rice pump-dump scheme have been successfully performed by several groups [82, 83, 84]
The control procedure in the frequency-domain was first proposed by Shapiro and
Brumer in 1986 [81]. This scheme is mainly inspired by the coherent control of chem- ical reaction. It employs two laser pulses with known phase di↵erence and distinct frequencies to change the phases of interfering dynamical pathways in order to ma- nipulate the dynamical processes [85]. There are several experimental demonstrations of the Shapiro-Brumer scheme such as coherent control of product distribution [86] and interference between ionization process [87].
In this thesis we employ the control field in the time-domain, which is consistent with the current experimental implementations and theoretical proposals for designing quantum gates [11, 88, 46].
3.3 Quantum control
Quantum control aims to steer quantum dynamics towards closely realizing specific quantum states or operations [75] with applications to femtosecond lasers [89, 90], nu- clear magnetic resonance and other resonators [91, 92, 93, 94, 95], laser-driven molec- ular reactions [96, 80], and to quantum-gate synthesis for quantum computing [97].
Control is achieved by varying the external control parameters such that the resultant
32 evolution closely approximates the desired evolution. The set of instructions that de- termine the control parameters, and hence, the e↵ectiveness of the control scheme, is called a policy. We employ the quantum control schemes to generate policies for designing quantum gates.
In any quantum control problem, the goal is to decompose the system’s Hamilto- nian into a controllable and an uncontrollable part, and steering the dynamics towards adesiredevolutionthroughvaryingthecontrollablepartofthesystem.Foraclosed system, the Hamiltonian
L Hˆ ["(⌧)] = Hˆ dr + "(⌧) Hˆ c = Hˆ dr + " (⌧)Hˆ c, (3.3) · ` ` X`=1 acts on Hilbert space H [98] with drift Hamiltonian Hˆ dr describing free (uncontrolled) evolution, which we treat as being time-independent here. The control Hamiltonians, represented by the vector operator Hˆ c(t)=(Hˆ c)(for ` the control field labels) ` { } should steer the system towards the desired evolution with time-varying control am- plitudes contained in the vector "(⌧):= " (⌧) . { ` } The resultant unitary operation over the time interval, ⇥ writes as follow:
⇥ U[⇥,"(⌧)] = exp i Hˆ ("(⌧))d⌧ (3.4) T ⇢ Z0 with the time-ordering operator [99]. Machine learning techniques can be employed T to evolve (3.4) towards a target operation.
3.4 Machine learning: A quantum control tool
The task of machine learning [40] is to develop algorithms that can learn from system behaviour and predict the future behaviour of the system based on their past evolu- tion. Machine learning algorithms have already been applied to various problems in quantum computing such as phase estimation [41], asymptotic state estimation [42], and discriminating quantum measurement trajectories and improving readout [43].
33 One can classify the machine learning algorithms into three distinct categories in- cluding, supervised and unsupervised learnings, and reinforcement learning.
The supervised learning task is to infer a function (hypothesis) from the labelled data (training set). The training data is comprised of the input vector along with their corresponding output vectors. The supervised learning algorithm trains the hypothesis on the training data to construct an inferred hypothesis. This inferred hypothesis can be further used to label novel data. Examples of supervised learning problems are regression or classification problems [100, 101].
Here I first explain our strategy to transform a gate design problem into a learning problem and discuss the specific learning problem we are dealing with. A quantum logic gate is a map between an input and an output state. One can always represent the action of any quantum logical gate on the basis elements in terms of a truth-table.
There is a one-to-one correspondence between the input and output elements in this truth-table, as the quantum logic gates are themselves reversible. The quantum gate design can be turned into a regression problem where the main task is to construct a hypothesis that maps the input state of qubits to its corresponding output state after the action of the desired gate.
Our strategy to transform a gate design problem into a learning problem leads to amulti-valuedregressionproblem.Typicalregressionproblemsinmachinelearning consist of finding a hypothesis that maps an n-dimensional input vector to a single- valued output parameter. A single-valued regression problem poses a much easier learning problem in contrast to our multi-valued regression problems. Novel machine learning techniques must be devised to tackle these multi-valued regression problems.
A simple example of quantum gate that can be transformed into a learning problem is a CZ gate. Given the action of the CZ quantum logic gate on the first and second qubits, it is straightforward to construct a truth-table which corresponds to the action
34 Input Output T C T C 0 0 0 0 |0i |1i |0i |1i |1i |0i |1i |0i |1i |1i |1i -| 1i | i | i | i | i Table 3.1: The truth table representation of CZ gate. T and C denote the control (first) and target (second) qubits, respectively. A CZ gate applies a Pauli-Z oper- ator on the target qubit if the state of the control qubit is 1 and leaves the state unchanged otherwise | i of a CZ gate (See Table 3.1). In Table 3.1 C and T denote the control (first) and target (second) qubits. In the context of supervised learning problem, we consider the truth-table as the training set. Loosely speaking, we train our hypothesis on the truth-table data as the training set such that the learnt hypothesis maps input states to the final states with high confidence.
Having clarified that the truth-table represents the training set, we now discuss what the hypothesis is. In the context of quantum gate design, the hypothesis is represented by the external pulses. Therefore, we train the parameters of the external pulses on the truth table data to shape the external pulses such that the system evolution approximates the target gate. If the hypothesis is learnt successfully, it will generate a policy which determines the shape and strength of the external pulses.
Thus far, we have explained the training set and hypothesis in the context of quantum gate design. However, we still need to know how to measure the success of our learning procedure as well as the explicit form of the hypothesis (external field), which we elaborate in the next two sections.
3.5 Control pulses as learning parameters
Our learning algorithm uses the external pulses as the hypothesis and a successful policy will determine the shape of the optimal pulse. Although our quantum control
35 approach can generate any type of pulses (hypothesis) for "(t), here we only consider two types of pulses: piecewise-constant and piecewise-error-function. In the case of piecewise-constant function, we discretize each control pulse, "(t), and express it as sum of K orthogonal constant functions over the interval [0,⇥]. A "(t)thencanbe shown as:
"1 0 1 "2 "(t):=B C (3.5) B . C B . C B C B C B"K C B C @ A Each "l in (3.5) is a constant number over the sequential and equally-spaced time steps: