You are on page 1of 250

QUANTUM COMPUTING WITH NUCLEAR SPINS

IN SEMICONDUCTORS

a dissertation
submitted to the department of applied physics
and the committee on graduate studies
of stanford university
in partial fulfillment of the requirements
for the degree of
doctor of philosophy

Thaddeus D. Ladd
June 2005

c Copyright by Thaddeus D. Ladd 2005



All Rights Reserved

ii

I certify that I have read this dissertation and that, in my opinion,


it is fully adequate in scope and quality as a dissertation for the
degree of Doctor of Philosophy.

Yoshihisa Yamamoto
(Principal Adviser)

I certify that I have read this dissertation and that, in my opinion,


it is fully adequate in scope and quality as a dissertation for the
degree of Doctor of Philosophy.

Alexander Fetter

I certify that I have read this dissertation and that, in my opinion,


it is fully adequate in scope and quality as a dissertation for the
degree of Doctor of Philosophy.

Mark Kasevich

Approved for the University Committee on Graduate Studies.

iii

iv

Abstract
The successful implementation of a scalable, fault-tolerant quantum computer would introduce a type
of information processing more powerful than any available today. Reciprocally, the discovery of a
fundamental obstacle to such a system would be an important advance in the foundations of quantum
theory. No such fundamental obstacles are currently known, but neither has any architecture been
shown to be experimentally scalable.
Many technologies have been considered for finding such an architecture; in this work I focus on
nuclear spins in semiconductors. Semiconductors provide promising optical means for polarizing and
measuring small nuclear spin ensembles, which are tasks that pose critical challenges to quantum
computers based on nuclear magnetic resonance (nmr). At the same time, semiconductor nuclei
are sufficiently coherent quantum oscillators to allow complex information processing using resonant
radio-frequency pulse sequences. In particular, the isotopically clean and magnetically quiet environment of pure, high quality, bulk single-crystal silicon provides a nuclear environment allowing what
may be the longest absolute coherence time of any solid-state qubit currently under consideration.
I have experimentally tested this claim using high-power nmr pulse sequences to eliminate inhomogeneous dephasing and dipolar evolution among an ensemble of

29

Si nuclei in isotopically modified

silicon crystals. Intrinsic decoherence processes are only observed in polycrystalline silicon, where
1/f charging noise processes are likely to blame. In high-quality single crystal samples, nuclear
coherence persists for over 25 seconds, a timescale limited only by pulse sequence imperfections.
I will discuss an architecture that takes advantage of this clean nuclear environment, but I will
also address its scalability limitations due to silicons poor optical characteristics. These limitations
will suggest new experiments employing nuclear spins in optically controlled semiconductor quantum
dots, which may hold more promise for future scalable quantum computer architectures.

Acknowledgements
This is a dissertation and not a research paper. Were it the latter, the text would be shorter, of
course, but the list of authors would be much, much longer. Many minds contributed to the material
I will present here, and many helped with support and encouragement. I would like to thank some
of them now.
One could not ask for a better adviser than Yoshi Yamamoto. His research program is expansive
and dynamic, and yet he is able to be a very supportive manager to all of his students. And although
it took me some time to fully realize it, he is above all else a great educator. He has guided me
toward successes but has also let me make mistakes; I understand now that both were required to
allow the kind of professional growth through which he has guided me. His approach to solving
problems reminds us that great challenges are great challenges and small issues are small issues, and
that throughout all technical problems, basic physics and engineering principles usually provide the
answers. His clear manner of thinking, his vision, and his ambition have all been inspirational.
Maintaining clear thinking, vision, and ambition in the darkness of the lab can sometimes be
exhausting, and here the other students and researchers of the Yamamoto group have been guiding
lights for me. I must first thank the resourceful Jonathan Goldman, with whom I have worked
most closely in these seven years. Jonathan and I arrived to the Yamamoto group together and
undertook the daunting task of building a new lab for nmr, micromagnetics, magnetic resonance
force microscopy, and magneto-optics, starting with no equipment, no labspace, and no technical
expertise. For the expertise, we both spent an educational summer in Japan. In my case, Atsushi
Goto, Tadashi Shimizu, and others at the National Research Institute of Metals in Tsukuba, were
extremely helpful for teaching me the basics of low-temperature solid-state nmr. Recently, Jonathan
and I have built yet another new laboratory from scratch, something no graduate student should
have to do twice. Jonathans hard work deserves more recognition that it has so far received.
Many of the actual experiments described in this dissertation were principally aided by the
industrious Denys Maryenko. Denys arrived as an undergraduate student at the same time as
vi

another helpful individual worthy of acknowledgement, Ernest Yeung. Motivated by their own
ambition, these two kicked off the new nmr lab with early relaxation experiments in insulating
crystals, including fluorapatite samples grown by Professor Ian Fisher, who was extremely gracious
to offer so much time and expertise to our project. Denys returned for his Masters thesis and
began the experiment to measure long decoherence times in silicon. He did much of the work in
this experiment, including the construction of the nmr probe that made the results possible and
the programming of the pulse sequences. More importantly, his enthusiasm for the work pushed it
forward in times of difficulty.
Although few in the Yamamoto group have worked directly on nmr, everyone in the group has
taught me something. In particular, Will Oliver taught me much about high frequency electronics,
and Na Young Kim has also been a great help in this arena. Charlie Santoris expertise in singlequantum-dot optical measurements has been a great resource; he is a true scientist, a die-hard
skeptic capable of bringing out the optimist in the rest of us. Since Charlie left the group, Stephan
Gotzinger has nicely filled his shoes. Kai-Mei Fu is a dynamo in the lab, bringing energy and ideas
to the most challenging of experiments. When Anne Verhulst joined the lab, she brought useful nmr
knowledge that was quite valuable. I must acknowledge Edo Waks for always having the answer to
any question I ask him, and always taking the time to provide it. Most recently Andrei Faraon has
started his graduate career partly in our laboratory, where his help has been much appreciated. Our
recent work with quantum dots would also not be possible without the growth expertise of Glenn
Solomon and Bingyang Zhang.
Much of my thesis project, however, has been theoretical, and discussions with group members
in this arena deserve acknowledgement. The whole idea of crystal lattice quantum computation was
initiated by Fumiko Yamaguchi, whose open-minded ideas have seeded a large portion of my work.
Mike Jura and Matt Terrel were visitors to our group early in their careers where they helped me
work on the theory of optical polarization of nuclei in semiconductors. And Cyrus Master has always
been a brilliant theoretical consultant.
I have benefited from all the other members of the Yamamoto group as well, whether from
discussions in the hallway, inspirational group-meeting talks, pointed questions, or just good times.
So I extend thanks to Oliver Benson, Aykutlu Dana, Hui Deng, Leo Di Carlo, Eleni Diamanti,
David Fattal, Dehuan Huang, Robin Huang, Kyo Inoue, Jungsang Kim, Shinichi Koseki, Prab
Kuniyil, Debbie Leung, Xavier Matre, Matt Pelton, Jocelyn Plant, David Press, Patrik Recher,
Kaoru Sanaka, Barry Sanders, Masa Shirane, Mitsuro Sugita, Jelena Vuckovic, and Gregor Weihs.
The administrative help in the Yamamoto group is also world-class, thanks to the hard work of
vii

Yurika Peterman, Mayumi Hakkaku, and Rieko Sasaki.


Much of this research, however, has benefited from collaborations outside of Stanford. In particular, my collaboration with the group of Kohei Itoh at Keio University in Japan has been extremely
fruitful. Kohei and his students Eisuke Abe and Rodney Van Meter have helped me not only by
providing isotopically engineered silicon, but also with their ideas and their extraordinary vision.
Naveen Khaneja at Harvard has been a very helpful resource; he suggested mrev-16 as a way to
remove effective fields in decoupling experiments, and as I will discuss this suggestion turned out
to be more helpful than either of us anticipated. At Berkeley, Anant Paravastu and Pat Coles in
the group of Jeff Reimer have taught me much about optical polarization in semiconductors, and I
am impressed by their intelligence and patience in the face of the challenge of understanding their
data. Several people at the ibm Almaden Research Center were helpful for various reasons: Ike
Chuang, for both his long-term vision and his practical advice; Nino Yannoni for sharing his infinite reservoir of nmr expertise; and Bruce Gurneys research group, including Jeff Childress and
Matt Carey, for providing me an enjoyable summer internship. Finally, the group of David Cory at
mit, especially Chandrasekhar Ramanathan, Jonathan Baugh, and HyungJoon Cho, have recently
brought my understanding of nmr to the next level.
I must thank the many, many people who wrote the code for LATEX2e and provided it for free.
Without this typesetting software the development of a document of this complexity and length
would have been much, much harder.
I would also like to thank the members of my oral defense and reading committees, not only for
patiently hearing and reading about my ideas but for their own inspirational work. Mark Kasevich
taught me much in my three quarters as his teaching assistant, and his research program is inspirational. Sandy Fetters class in electromagnetism was the best I took at Stanford, and perhaps some
element of the style of his notes has carried on into this dissertation. Walter Harrison has written
nearly all of the papers and books in solid-state theory that actually increased my comprehension of
the subject. I also thank my chair, Jody Puglisi, for bringing his nmr expertise to the committee.
I also owe a huge debt of gratitude to the Fannie and John Hertz Foundation. Their monetary
support was extremely generous, and the discussions with directors (Lowell Wood in particular)
and other fellows at interviews and symposia helped me maintain my direction, especially in the
beginning.
I must also acknowledge Henry Chin, who has provided kind friendship amongst much advice,
both technical and fuzzy. And finally, my extremely brilliant wife, Sharon Ungersma, for her love,
intelligence, guidance, support, and encouragement that, ultimately, make all of this worthwhile.

viii

Contents
Abstract

Acknowledgements

vi

1 Introduction
1.1

1.2

1.3

Is Quantum Mechanics Complete? . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.1

Einstein and Local Realism . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.2

Schr
odinger and Verschr
ankung . . . . . . . . . . . . . . . . . . . . . . . . . .

Theoretical Quantum Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.1

The Deutsch-Josza Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.2

Shors Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.3

Other Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12

1.2.4

Quantum Error Correction and Fault Tolerance . . . . . . . . . . . . . . . . .

12

Experimental Quantum Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

2 Physical Resources for Quantum Computing


2.1

2.2

2.3

21

Three Essential Requirements for Fault Tolerant Quantum Computation . . . . . . .

21

2.1.1

Scalable Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

21

2.1.2

Universal Quantum Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

2.1.3

Initialization/Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

Opening the Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

2.2.1

Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

24

2.2.2

Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

2.2.3

Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

Quantum Memory and Quantum Repeaters . . . . . . . . . . . . . . . . . . . . . . .

34

2.3.1

34

The Reasons for Quantum Communication . . . . . . . . . . . . . . . . . . .


ix

2.4

2.3.2

The Quantum Repeater . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

35

2.3.3

Physical Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

Quantum Computing Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

2.4.1

Atomic and Molecular Optics Implementations . . . . . . . . . . . . . . . . .

41

2.4.2

Solid-state Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43

2.4.3

Linear Optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53

3 NMR Quantum Computers

55

3.1

NMR Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

57

3.2

Liquids vs. Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

3.2.1

Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

3.2.2

Distinguishing Qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

72

3.2.3

Quantum Logic and Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . .

75

3.2.4

Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

3.2.5

Are Liquid State Quantum Computers Really Quantum? . . . . . . . . . . .

82

Theory of Indirect Nuclear Interactions . . . . . . . . . . . . . . . . . . . . . . . . .

84

3.3.1

Effective Hamiltonians from Second Order Perturbation Theory . . . . . . . .

84

3.3.2

One Electron Spin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

86

3.3.3

Two Coupled Spins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

86

3.3.4

Lattice of Coupled Spins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

88

The All-Silicon Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

91

3.4.1

Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

3.4.2

Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

96

3.4.3

Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

3.3

3.4

3.5

Quantum-Dot NMR Quantum Computers . . . . . . . . . . . . . . . . . . . . . . . . 102


3.5.1

Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

3.5.2

Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

3.5.3

Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4 Decoupling: Theory

107

4.1

Ensemble vs. Single-Spin Measurement

. . . . . . . . . . . . . . . . . . . . . . . . . 108

4.2

Theory of Dipolar Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110


4.2.1

Multiple Pulse Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

4.2.2

Magic Angle Spinning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118


x

4.3

4.4

Spin Echoes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119


4.3.1

The Effects of -pulse Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

4.3.2

Spin-echo Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

4.3.3

Even-Odd Asymmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

Recoupling Pulse Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

5 Decoupling: Experiment
5.1

5.2

131

Experimental Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131


5.1.1

Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

5.1.2

Samples and Coils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5.1.3

The Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

5.1.4

Small Angle FID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

5.1.5

Pulse Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

5.1.6

Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5.1.7

Spin Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

6 General Relaxation Theory

153

6.1

Fermis Golden Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

6.2

General Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

6.3

6.2.1

System and Bath . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

6.2.2

Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

6.2.3

Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

6.2.4

Observable Averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

Spin Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160


6.3.1

Spin-Boson Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

6.3.2

General Formulae for T1 and T2 . . . . . . . . . . . . . . . . . . . . . . . . . . 164

6.3.3

Two Hyperfine Coupled Spins . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

7 Nuclear Relaxation in Silicon


7.1

7.2

177

Free Carriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177


7.1.1

Conduction Electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

7.1.2

Valence Holes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

Trapped Carriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182


7.2.1

Stationary Impurity Donors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182


xi

7.2.2
7.3

Optically Excited Bound Excitons . . . . . . . . . . . . . . . . . . . . . . . . 186

1/f Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189


7.3.1

General Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

7.3.2

Decoherence Due to 1/f Noise During Decoupling . . . . . . . . . . . . . . . 190

8 Prospectus

197

A Notation

199

A.1 Vectors, Matrices, and Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199


A.2 Quantum Mechanics and NMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
A.3 Liouville Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
A.4 Many-body notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A.5 The Hyperfine Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
B Spin Algebra

207

B.1 Single Spin Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207


B.2 Two Spin Commutators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
B.3 Diagonalizing a 2 2 Hermitian Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 210
Bibliography

211

Index

229

xii

List of Figures
1.1

A comparison of execution times of classical and quantum factoring algorithms vs.


the size L, in bits, of the number to be factored. The curve labelled nfs represents
the expected scaling of the classical nfs algorithm with the total scale set by the
current world-record largest implementation of it, in which 104 personal computers
running in parallel factored a 576-bit number in one month in 2003, according to
rsa Security, Inc. The curves labelled Shor C represent an implementation of Shors
quantum algorithm in which logical qubits may only couple to nearest neighbors. The
curves labelled Shor E represent an implementation in which any logical qubit may
couple to any other logical qubit. These implementations are described by Van Meter
and Itoh [13]. Both algorithms use 100L logical qubits. For comparison, clock speeds
of 1 Hz or 1 MHz are shown; these clock speeds must include the time for extra
processing for quantum error correction. . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

11

A quantum circuit representing the 3-qubit bit-flip code. The qubit represented by the
top line is encoded into three qubits with two controlled-not operations on two ancilla
qubits. A single qubit error occurs somewhere in the dashed box. The syndrome and
recovery operations are implemented with two controlled-not operations and two
polarization measurements; the top qubit is flipped if both ancilla are measured in
state |1i.

2.1

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14

The number of pure qubits n which may be extracted at finite polarization as a


function of the polarization p. Each curve is labelled by the total sensitivity, which is
the number of qubits one is capable of measuring divided by the number of qubits in
the ensemble (or the number of repetitions of the experiment). . . . . . . . . . . . .
xiii

29

2.2

A modification of the quantum error correction circuit shown in Fig. 1.2, in which the
measurements with feed-forward control have been replaced by a coherent Toffoli gate
for recovery. After the recovery, the ancilla qubits must be discarded or reinitialized.

2.3

31

Schematic of an optical repeater. Optical beams a and b meet at a beam splitter,


and the output beams are sent to distant qubits A and B. Each qubit is an identical
-system, as shown in the center of the figure. The two qubit states are the two
lower states of the system; the excited state is optically accessible only to state |0i.
Many photons are lost at A and B, but some photons reflect off A or B if they are in
state |0i. These reflected photons are combined at a second beam-splitter, completing
a Mach-Zender interferometer; a detection event at port c may project the distant
qubits into an entangled state. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.1

38

The dipolar magnetic field BD (r) generated by one nucleus is seen by a random
assortment of other nuclei. Only the z-component of the dipole field is important at
high applied fields B0 ; this is the secular component. This component vanishes at the
magic angle M shown.

3.2

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

62

Sample geometries for generating a large magnetic field gradient across a regular
crystalline lattice. The magnetic field in the macromagnet causes planes of nuclei to
have unique Larmor frequencies j . Calculations of expected magnetic field gradients
for these geometries appear in Ref. 141. . . . . . . . . . . . . . . . . . . . . . . . . .

3.3

74

The basic geometry for a two-dimensional magnetic field gradient. The gradient field
G points at a slight angle from the x-axis. The gradient causes each nucleus in the
grid to have a unique Larmor frequency j .

3.4

. . . . . . . . . . . . . . . . . . . . . .

75

An illustration of the geometry of nuclei discussed in the text. The nuclei form a
periodic lattice. The applied magnetic field and magnetic gradient are assumed to
point in the z-direction, so horizontal rows of nuclei have the same Larmor frequency
j . The nearest-neighbor qubit-qubit distance is a and the nearest-neighbor distance

among ensembles is 1 a. If the angle is the magic angle, then 1 = 2. . . . . . .

3.5

78

A schematic for a fluorapatite crystal lattice quantum computer. The fluorapatite


crystal is shown mounted on a silicon cantilever, whose oscillations provide mrfm
readout. The cantilever and crystal are aligned with the micromagnet which generates
the field gradient. The inset shows the one-dimensional structure of

19

F nuclei in

fluorapatite. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xiv

80

3.6

A schematic for an all-silicon quantum computer. The figure shows the integrated
micromagnet and bridge structure needed for distinguishing qubits and for read-out.
The bridge has length l = 300 m, width w = 4 m, and thickness t = 0.25 m. The
micromagnet has D = 400 m, L = 4 m, and W = 10 m, and produces a field
gradient of B z /z = 1.4 T /m, uniform over a 100 m by 0.2 m region inside the
bridge. The insert shows the structure of the silicon matrix and the terrace edge. The
darkened spheres represent the

29

Si nuclei, which preferentially bind at the edge of

the Si step. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7

93

Images of single atom chain growth in silicon. A stm image appears on the left with
corresponding ball-and-stick atomic model on the right; the arrows on the bottom
compare critical atomic locations between the two images. The images labelled (a)
show a terrace-step on the Si(111)77 surface before deposition of additional silicon;
this structure would be made from spin-0

28

Si. The images labelled (b) show that

deposited silicon atoms form a straight atomic chain on the terrace step; these extra
atoms would be the spin-1/2

29

Si nuclei. This figure is taken from Ref. 171, which

explains the experimental conditions and the modelling in more detail.


3.8

. . . . . . .

95

Energy diagram for the neutral donor (P0 ) and its bound exciton (P0 ,X) in a magnetic
field. The (P0 ,X) state is populated via capture of a resonantly excited free exciton.
The

4.1

31

P nuclear state can be determined by the energy difference of a and b. . . . .

99

Simulated dipolar decoherence versus time. The broken lines represent individual spin
measurements Re{Gj (t)} on an 8-spin simulation of dipolar evolution with uniformly
random coupling constants with range [D0 /2, D0 /2]. The solid line is the magnitudesum |M (t)|. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

4.2

The wahuha pulse sequence. All pulses are broadband /2 pulses of the indicated
phase. Measurements are made at the times marked S; here the toggling and rotating
reference frames coincide. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4.3

Schematic of the cpmg-mrev-16120 pulse sequence with spin-echo data. The echoes
shown in the upper left and expanded in the upper right are data from an isotopically
natural single crystal of silicon. These are obtained by first exciting the sample with
a /2-pulse of arbitrary phase , decoupling with the mrev-16 sequence shown in
detail on the bottom line, and refocusing with -pulses of phase = + /2 every
120 cycles of mrev-16. The magnetization is sampled once per mrev-16 cycle in the
windows marked with an S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
xv

4.4

A sample pulse sequence incorporating Hadamard decoupling and wahuha. This is


constructed directly from H(4) by translating each change of sign in the ith row to
a soft RX () pulse, labeled , at frequency i . The first, all positive, row of H(4)
has been removed to eliminate inhomogeneous broadening. Single qubit rotations, as
indicated by , must occur between full cycles of the sequence. Recoupling between
qubits i + 1 and i + 2 may be achieved by inserting a pulse where indicated by the
dotted line. The top row represents a broadband wahuha sequence for homonuclear
decoupling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

5.1

The series parallel tank circuit with an extra Q-reducing capacitor C . . . . . . . . . 135

5.2

A typical data time series and Fourier transform showing clear spin locking effects.
This data was taken with polycrystalline silicon under mrev-8. The raw time series
on the left shows the first second of the real (in-phase) and imaginary (out-of phase)
amplitudes; the waterfall plot on the right shows the result of applying the dft to
each echo block as described in Sec. 5.1.6. . . . . . . . . . . . . . . . . . . . . . . . . 144

5.3

The magnitude of the center and side peaks from a waterfall plot when -pulses
are applied frequently (N = 5) without the cpmg convention under mrev-8 with
tc = 1.03 ms, again taken with polycrystalline silicon. These correlated oscillations
in the magnitude are due to -pulse errors. . . . . . . . . . . . . . . . . . . . . . . . 146

5.4

Oscillations observed in the imaginary part of the waterfall side peak when using
alternating -pulse phases with an orthogonal preparation pulse, except where labelled
cpmg where the pulses are of constant phase. The oscillation frequencies vary with
the sparsity of the -pulses (N ) as shown in the plot on the left and with the frequency
offset as shown in the plot on the right, as expected for -pulse errors. The oscillations
are removed by using the cpmg convention. This data was taken with polycrystalline
silicon under mrev-8 with tc = 0.681 ms. . . . . . . . . . . . . . . . . . . . . . . . . 147

5.5

Pulsed spin locking observed in heavily doped silicon wafers under cpmg-mrev-16N
with tc =2.46 ms. The magnitudes of the side-peak echoes are shown. . . . . . . . . . 148

5.6

The cpmg sequence in heavily doped silicon wafers without decoupling. The different
traces correspond to different -pulse spacings 2 . Long-lived echoes are likely due
to spin-locking effects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
xvi

5.7

Coherence time vs. cycle time in single-crystal silicon. The solid line is a fit showing
the exponent 2.09 0.07 for the isotopically enhanced sample (left) and 2.00
0.2 for the isotopically natural sample (right). The insets show the integrated logmagnitude of the spin-echoes decaying in time for a few cycle times. . . . . . . . . . 149

5.8

The decay of the spin-echo peaks under mas for several rotation speeds , with
exponential fits (left), and the observed decay times T2 plotted against (right).

5.9

. 150

Spin echoes for isotopically depleted silicon under cpmg-mrev-16120. The solid
line shows exp(t/8 sec), for comparison. . . . . . . . . . . . . . . . . . . . . . . . . 151

5.10 Echo decay curves for pure polycrystalline silicon of natural isotopic abundance. No
significant variation in the data is observed as tc is changed. The solid line is a fit to
the function described in Sec. 7.3.2.
7.1

. . . . . . . . . . . . . . . . . . . . . . . . . . . 152

A schematic of the magnetic noise spectral density 2 J() in silicon. (The direction
of the magnetic noise vector is neglected in this plot). The Lorentzian magnetic noise
spectral density due to a very fast fluctuator (white noise) is represented by the dashed
line. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

7.2

Residuals versus time. The deviation of the data of Fig. 5.10 from the fitting function
of Eq. (7.48), with a histogram of those residuals on the right, consistent with Gaussian
noise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

xvii

xviii

Chapter 1

Introduction
The more success the quantum theory has, the sillier it looks.
Albert Einstein
I dont like it, and Im sorry I ever had anything to do with it.
Erwin Schr
odinger

1.1

Is Quantum Mechanics Complete?

The theory of quantum mechanics tells us that the world is much, much more complex than what
we observe. It posits the bizarre notion that the results of physical measurements do not exist
until the actual measurements are made. In the meantime, our universe is described by its wave
function, a very large mathematical description allowing the universe to simultaneously coexist in
contradictory configurations.
The superb agreement between quantum electrodynamic calculations and experimental measurements of simple systems such as energy levels of the hydrogen atom might suggest that, insofar as
energies are low and only electromagnetic forces are at play, this theory is final, a closed book.
Frontiers in the quest for a more complete theory are then pushed to the other forces, to extremely
high energies, and to gravity, where general relativity must somehow be fused with quantum field
theory. Physics does not yet have a complete theory to explain the processes during the big bang
or in a black hole, and those who seek to work at the cutting edge of theory often find themselves
in these cramped, uncomfortable times and places.

Chapter 1. Introduction

But if we go as far as believing that the predictive power of the wave function indicates that
it represents physical reality, then there must be another frontier in physical theory: finding the
edge between the quantum world and the classical world. How can we reconcile the multifarious
possibilities of the wave function, verified in observations of microscopic quantum systems, with the
much simpler observations of the macroscopic world?

1.1.1

Einstein and Local Realism

One approach to this question is to suggest that quantum theory is simply incomplete. This approach
was taken by Einstein, whose famous 1905 paper on the photoelectric effect and support of the
ideas of de Broglie played critical roles in the establishment of quantum theory. In 1935, Einstein,
with coworkers Podolsky and Rosen [1], argued that quantum mechanics must be missing hidden
variables. The route to contradiction was to show that the act of quantum measurement would
violate the principles of relativity. The resulting paradox is known as the epr paradox.
The act of measuring a quantum system inevitably changes that system, an interpretation held
from the early beginnings of quantum theory in Copenhagen. Before a measurement, we have the
wave function, a mathematical representation of the quantum state, but prior values of particular
measurements are not contained in this wave function in any deterministic sense. For each measurement, these values are randomly selected with a probability distribution that may be calculated
given the wave function. After the measurement, the wave function collapses into the measured
state, and as the theory is written this collapse happens instantaneously over the entire wave
function, even if the wave function describes spatially distant, non-interacting objects. In this sense,
quantum theory is notlocal. If wave function collapse is a physical action, Einstein argued, it
allows instantaneous action at a distance, a notion prohibited by the tenets of relativity. The act
of quantum measurement was therefore contradictory to Einstein, and he argued that additions to
the theory would be required to save it. Einstein preferred a theory in which the results of any
measurement were physically present before the measurement. The apparent randomness in this
theory would be a result of hidden variables currently unknown to the theory. Such a theory would
be realistic or objective.
This line of thinking developed substantially after Einsteins time. The crucial step came in 1964
when Bell [2] developed quantitative inequalities to differentiate the statistics of measurements of a
system described by a strictly local, realistic theory from the measurements predicted by standard
quantum mechanics. In the subsequent decades, multiple experiments violating Bells inequalities

1.1 Is Quantum Mechanics Complete?

showed, shy of obscure loopholes, that quantum mechanics is indeed either non-local or not realistic. Einsteins famous intuition failed him in this case; the modern researcher brands Einstein, quite
ironically, as classical, and insists on a new, more open-minded intuition. Each of us who accepts
quantum mechanics must choose whether we prefer to abandon either locality or realism, for
quantum mechanics cannot have both.

1.1.2

Schr
odinger and Verschr
ankung

Schr
odinger might be considered the inventor of the wave function, although to him the wave function
represented a real, physical wave, not a generalized probability distribution. This latter interpretation is due to Born, Heisenberg, and especially Bohr. Schr
odinger sided with Einstein in finding fault
with the Copenhagen interpretation. He attempted to bring the Physical Review paper of Einstein
et al. [1] to popular attention with an article in the German magazine Naturwissenschaften (Natural Sciences) [3]. In this article, Schr
odinger identified the heart of the epr paradox, the existence
of states featuring multiple quanta with a non-classical kind of non-local correlation. He called this
concept Verschr
ankung, an ordered intertwining, which has been translated as entanglement. To
popularize the strangeness of entanglement, Schr
odinger asked the reader to imagine a macroscopic
entangled object. His choice was a cat, brought to an entangled state in which, due to a carefully
contrived series of events beginning with a single nucleus in a superposition state, the cat persists
in a state simultaneously alive and dead. This was supposed to be so absurd that, like Einstein,
the reader would assume something was wrong, or at least incomplete, about quantum mechanics.
Schr
odingers famous misfortuned cat challenges a naive notion about quantum mechanics: that it
applies only to physical systems whose action is of order ~. In fact this is not built into the theory,
and there is no fundamental limit in quantum mechanics to the size of entangled states, allowing
such bizarre possibilities as Schr
odingers cat.
More recently, the concept of decoherence has provided a functional resolution of Schr
odingers
dilemma. Decoherence may be summarized as the following process: when information about degrees
of freedom of a large system are ignored or lost, quantum superpositions (including entanglement)
lose phase coherence, causing descriptions of states to revert to classical probability distributions.
The mathematical part of this definition is straightforward, as I will discuss in Sec. 2.2.1 as the partial
trace over a density operator. The premise of it, however, simply moves Schr
odingers concerns about
the theory to a different place, rather than eliminating them. When exactly is information lost or
ignored ? Disturbingly metaphysical is the following question: Does decoherence require a conscious
observer?

Chapter 1. Introduction

Many interpretations of quantum mechanics answer these questions in different ways. Some
theories posit new physical effects that introduce a fundamental decoherence as the size of the
system grows; a particularly tempting line of thought connects this fundamental decoherence to
gravity. A useful review of the state of this line of questioning is given by Leggett [4]. I believe that
these questions cannot be answered strictly theoretically. Currently, the interpretation of quantum
mechanics is a matter of taste. However, an experimental route to further explore these questions
is available, just as measurements of Bells inequalities have answered the epr paradox in favor of
quantum mechanics without local realism. If quantum mechanics is complete, at least on laboratory
length and time scales, then it should be possible to build arbitrarily large entangled states of matter,
such as Schr
odingers cat. Introduced as a notion of absurdity in 1935, Schr
odingers cat has become
an experimental goal pursued by physicists around the globe.
Certainly, states exhibiting coherent quantum entanglement among many quanta have been
observed in various systems. Obvious examples are bosonic condensates such as superfluid helium
and Bose-Einstein condensates of optically trapped atoms. Coherent states of macroscopic numbers
of fermions are also observable; superconductors and the quantum Hall effect provide examples.
However, these systems exhibit entanglement only for a small subset of states, which are either close
to the energetic ground states of the quantum system or are stabilized by measurement. Closer
analogs to Schr
odingers cat have been approached in various experiments; examples include a
superposition of distinct optical coherent states in a microwave cavity [5] and in distinct states of
superconducting Josephson junction circuits, which I discuss further in Sec. 2.4.2. Even in these
experiments, the available number of quantum states and the timescale during which they survive
before extrinsic decoherence processes wash them out is limited. A system to truly satisfy the qualms
of Schr
odinger should have access to the entirety of its Hilbert space.
How large is Hilbert space? There is truly a gestalt to Verschr
ankung, as demonstrated by the
following argument. For this, we enumerate the state space of n quantum two-state systems, known
as quantum bits or qubits. A single qubit is not impressive; it is easily simulated by a single classical
oscillator. The number of states available to a qubit can be enumerated by considering all operations
that can be performed on a qubit in its ground state. This is the group of 2 2 unitary matrices with
unity determinant, SU(2). Physically, this is the group of norm-preserving rotations in 3 dimensions.
It has exactly 3 generators, I z , I x , and I y . These may respectively correspond to 3 measurements
of a classical oscillation: the dc bias, the in-phase amplitude, and the out-of-phase amplitude. In
the quantum case, measurement of the mean of the binomial random variable corresponding to
these three generators would provide enough information to completely reconstruct the state. If we

1.2 Theoretical Quantum Computing

neglect all entangled states, the number of states available to n qubits corresponds to the number
of operations in the group [SU(2)]n ; the direct product of SU(2) with itself n times. This group
has 3n generators, and is no larger than the space available with n classical oscillators, each with
relative phase, amplitude, and bias. But the postulates of quantum mechanics dictate that Hilbert
space is much, much larger than this. Nothing in the theory prohibits arbitrary unitary operations
on n qubits, so the group of available operations is SU(2n ), a group much larger than [SU(2)]n .
The group SU(2n ) has 4n 1 generators, which means an exponentially large number of statistical
measurements would be needed to collect all the quantum information in an unknown state of n
qubits!
The quantum computer is a system that exploits the inherent non-classicality of a gigantic
Hilbert space; it is a generalized Schr
odingers cat. It is a device designed to take a reasonably
large number of quanta and put them into an arbitrary highly entangled state. Then, once these
macroscopic entangled states are available, their existence can be demonstrated without a decohering
measurement, by using them to compute information in ways no classical computer could. If such a
quantum computer is built, it would act as a new variety of Bell inequality measurement, not testing
local realism but rather testing whether quantum mechanics holds up to complexity.
Building very large, persistent entangled states requires the suppression of decoherence, and by
extension the suppression of measurement. This would be extremely impractical if it were not for the
possibility of quantum error correction, a process whereby quantum information is protected from
certain kinds of decoherence. The discovery of algorithms possible on quantum computers with
quantum error correction is a recent and revolutionary development of quantum mechanics. It has
shifted the direction of research in the foundations of quantum theory from Is quantum mechanics
complete? to If quantum mechanics is complete, what can we do with it?

1.2

Theoretical Quantum Computing

The quantum computer did not arise out of the considerations described in the previous section,
but rather out of fundamental considerations in computer science. It was at about the same time
as Einsteins and Schr
odingers attacks on quantum theory that Turing [6] introduced what has
become known as the Universal Turing Machine, and suggested that any algorithmic process can be
performed on such a machine, a suggestion that has become known as the Church-Turing thesis. As
digital computers became more and more powerful in the 1960s and 70s, an important modification
to this thesis arose, which is that the efficiency of an algorithm is also important. The strong

Chapter 1. Introduction

Church-Turing thesis says essentially that if one physical device can perform an algorithm, then a
Universal Turing Machine can perform the same algorithm without a substantial difference in the
number of computational steps.
One kind of algorithm that was noted by many to be inefficient on any classical computer was the
simulation of large quantum systems. This is because the wave-function has exponentially growing
complexity for a linearly growing number of quanta. In 1982, Feynman [7] turned this situation on
its head by suggesting that the efficient simulation of large quantum systems should be possible with
a quantum mechanical computer [7].
A logical extension to this idea is that a quantum computer may efficiently perform an algorithm
that a Universal Turing Machine cannot, challenging the strong Church-Turing thesis. It was Deutsch
[8] who first sought and found a specific example of such an algorithm, which was subsequently
expanded with Josza into the Deutsch-Josza algorithm [9].

1.2.1

The Deutsch-Josza Algorithm

The Deutsch-Josza algorithm is the simplest of quantum algorithms, and will therefore provide
an illustrative example. Despite its simplicity, it features most of the key elements that a quantum
computer will be expected to perform. It is certainly a quantum algorithm that completes its task in
exponentially fewer steps than its classical analog. However, it presumes the presence of an oracle,
a black-box part of the algorithm whose implementation-time is not considered in the complexity
measure, and which must itself be quantum mechanical. The algorithm must use this oracle in order
to find out what the oracle does.
It is known before the algorithm begins that the oracle takes in a string of n-bits, x, and for
each of the 2n possibilities it returns a single bit corresponding to function f (x). It is promised that
either f is constant, meaning f (x) returns zero for all x or one for all x, or f is balanced, meaning
f (x) returns zero for exactly half of the 2n input strings x and one for the other half. The goal of
the algorithm is to deduce whether the function f is constant or balanced while calling the oracle a
minimum number of times. This artificial problem was invented for the sake of the algorithm; it is
not known to have practical application.
A classical computer cannot deterministically evaluate whether f is constant or balanced without
calling the oracle 2n1 + 1 times. If after 2n1 different queries f always returns a single value, both
possibilities are still viable; one more call is needed to eliminate the possibility of the constant
function1 . A quantum computer, however, assumes that the oracle is linear, so that if you may
1 Note,

however, that if the a priori probability of each possibility is equal, then it becomes exponentially less

1.2 Theoretical Quantum Computing

input one bit-string to it, you may insert all strings of qubits at once.
A linear superposition of all bit strings is obtained by applying the single Hadamard gate, H, to
each bit. This is sometimes referred to as the Walsh-Hadamard gate. This gate takes a single qubit

from |0i to (|0i + |1i)/ 2 or from |1i to (|0i |1i)/ 2, hence converting the polarization of the

qubit to the phase of the qubit2 . It is easily seen that if the Hadamard gate is applied one-by-one
to a register of qubits, it takes any bit string represented by those qubits to a superposition of all
possible bit strings, with the information about the original string encoded as the relative phases of
the terms in the superposition. For example, if we apply this gate one-by-one to each qubit in the
string |010i, we obtain
H 3 |010i = H |0i H |1i H |0i

 
 

1
|0i + |1i |0i |1i |0i + |1i
=
8


1
= |000i + |001i |010i |011i + |100i + |101i |110i |111i .
8

The generalization to n-qubits may be written


1 X
H n |xi =
(1)xy |yi .
2n y

(1.1)

The Deutsch-Josza algorithm begins with the register of n qubits all in the |0i state and one
ancillary qubit in state |1i. To this we apply the (n + 1)-qubit Hadamard gate:
H

(n+1)

|0i

|1i =
2n+1

X
x



|xi |0i |1i .

(1.2)

We now have our superposition of all possible states which we feed to the oracle Uf . The oracle has
the form of a reversible many-qubit gate, so the initial information must be preserved. It therefore
has the form
Uf |x, yi = |x, y f (x)i ,

(1.3)

where indicates bitwise addition (i.e. addition modulo 2). When Uf is applied to the linear

probable after each random call that the function is balanced. In this sense the Deutsch-Josza algorithm shows little
improvement over a probabilistic classical algorithm.
p
2 To be precise, the polarization of the qubit is p = hI z i/I and the phase is = i log[hI + i/ hI + I ihI I + i]. In a
Bloch sphere representation, p is the sine of the latitude and is longitude.

Chapter 1. Introduction

superposition of Eq. (1.2), the result is


n

Uf H (n+1) |0i

|1i =

1
2n+1

X
x



(1)f (x) |xi |0i |1i .

(1.4)

Here we see Schr


odingers Verschr
ankung. Although only one call has been made to the oracle, the
information about an exponentially large number of calls to f is stored in the phases of a manybody entangled wave function. For the quantum computer to be universal, any such entangled wave
function must be physically possible.
The algorithm is not yet complete, though, because the relative phases contained in the wave
function are not directly measurable. A polarization measurement made at this stage of the computation would show the computer in any of its possible states with equal probability! To be useful,
the phases of the wave function must be converted back to polarization, which may again be done
with the (n + 1)-qubit Hadamard gate:
H (n+1) Uf H (n+1) |0in |1i =

1 XX
(1)xyf (x) |yi |1i .
2n x y

(1.5)

Now, the polarization measurement of the register has the following probability of yielding the initial
n

state |0i

|1i:


2
1 X

1, f is constant


n
n
|h1| h0| H (n+1) Uf H (n+1) |0i |1i|2 = n
(1)f (x) =
2

0, f is balanced.
x

(1.6)

Even though the final measurement is stochastic, the initial promise that f is constant or balanced
assures that if any single qubit of the first n qubits is measured in the state |1i, the function must have
been balanced. If this promise is not made, then the algorithm offers no deterministic information
about the function.

1.2.2

Shors Algorithm

The Deutsch-Josza algorithm provided some idea that quantum computers were powerful. Key to
the algorithm was the initial conversion of the polarization of each qubit to the phase of that qubit,
performed by the Hadamard transformation. This is essentially the idea of the Fourier transform.
The classical Discrete Fourier Transform (dft) takes a vector of N complex values g to a vector of

1.2 Theoretical Quantum Computing

N complex values G with components


N 1
1 X
Gk =
gj e2ijk/N .
N j=0

(1.7)

The Hadamard transformation is a single-qubit Quantum Fourier Transform (qft). Consider a


qubit in state g0 |0i + g1 |1i. The dft of its amplitudes gives
1
G0 = (g0 + g1 )
2
1
G1 = (g0 g1 ).
2
The Hadamard transformation applied to our qubit is G0 |0i + G1 |1i.
The dft is routinely used to find periodicity in data, and indeed, in 1994 Simon [10] showed an
algorithm using Hadamard transformations for finding the period of a function. The truly dramatic
leap in the field of theoretical quantum computation, however, came when Shor [11] generalized from
the Hadamard transformation to the n-qubit qft, and in doing so demonstrated an algorithm that
takes only a composite number N as an input and determines with high probability a non-trivial
prime factor of N in a polynomial number of steps.
The n-qubit qft is easily defined. First, it is useful to convert from our prior notation of bitvectors x to the binary number representation x = x0 + 2x1 + 4x2 + + 2n xn . The qft is defined
as
Uqft

n
2X
1

x=0

gx |xi =

n
2X
1

y=0

Gy |yi ,

(1.8)

where G is the 2n -component vector of complex values calculated as the dft of the 2n -component
vector g.
Naively, it might seem to take 22n operations to compute the dft of the 2n complex amplitudes
gx , since for each of the 2n values Gy we must multiply all 2n values of gx by phase factors and then
perform a summation. A step of critical importance in classical algorithms, however, was the 1965
discovery of the Fast Fourier Transform (fft) by Cooley and Tukey [12]. The fft exponentially
speeds up calculation of the dft to only order n2n operations without loss of accuracy. The critical
step for quantum algorithms was the discovery that the qft could be done in exponentially fewer
steps than even the fft, requiring only O(n2 ) operations! This does not mean that the qft can
replace the ubiquitous fft, as the qft is fundamentally quantum mechanical, operating in the full,
unmeasurable Hilbert space of n qubits.

10

Chapter 1. Introduction

The qft is useful for finding the eigenvalue of a unitary operator U corresponding to a known
eigenstate |ui. I will now describe the algorithm for doing this starting from the answer. Since U is
unitary, the desired eigenvalue has the form exp(2i), where 0 < 1. Let us represent with
t-bit accuracy as x/2t . The algorithm seeks to generate the t-qubit state |xi. The qft of |xi is
t

2 1
2 1
1 X 2ixy/2t
1 X 2iy
Uqft |xi =
e
|yi =
e
|yi .
2t y=0
2t y=0

This state may be generated by starting with one register of qubits in the superposition state
P
2t/2 y |yi, obtained with Hadamard transformations as in the Deutsch-Josza algorithm, and a
second register of qubits in the known eigenstate |ui. The desired state results from the operation
|yi |ui |yi U y |ui .

(1.9)

The algorithm using this operation and the inverse qft to find is known as the phase estimation
algorithm. This is the quantum mechanical part of Shors algorithm.
The rest of Shors algorithm is classical number theory. To summarize, suppose we wish to find
the m factors of the number N . A classical computer can efficiently eliminate the simple possibilities
that N is even or N = ab for integers a and b; let us assume that N is neither. A random integer
x < N is chosen; a classical computer can quickly check to see that x and N are coprime. Then, the
unitary operator U used for the phase estimation algorithm is modular multiplication by x. That is,
for an input ket |yi, U outputs the ket corresponding to x multiplied by y, modulo N . The operation
corresponding to Eq. (1.9) is therefore modular exponentiation, and the specific implementation of
it may be by a variety of classical (but reversible!) subalgorithms. Which subalgorithm takes the
least amount of time depends on the resources available and the size of N , as recently summarized
by Van Meter and Itoh [13]. It may readily be shown that the modular multiplication operator U
has r eigenstates |us i with eigenvalues exp(2is/r), where the integer r satisfies xr = 1 mod N .
Here the integer s runs from 0 to r 1. The number r is the order of x modulo N and it is
the desired output of the phase estimation algorithm. Although calculation of the eigenstates |us i
would require knowledge of r, it may also be shown that the equal superposition of these states
yields the binary-representation ket |1i. Using this known state instead of one of the |us i states
in the phase estimation algorithm outputs a fraction s/r with a random s. If s/r is known with
sufficient precision, the denominator r may be efficiently calculated. With probability 1 2m , this

r is even and xr/2 6= 1 mod N , and if this is the case, then the greatest common denominator of

1.2 Theoretical Quantum Computing

11

100 years

,
rC

z
1H

Sh

Hz
,1
E
r
Sho
Hz

NFS

1 year
1 month
1 hour

M
,1
C
r

Sh

MH
,1
E
r
Sho

1 minute
1 second

100

1000
L

10,000

Figure 1.1: A comparison of execution times of classical and quantum factoring algorithms vs. the size L, in bits, of the number to be factored. The curve labelled nfs
represents the expected scaling of the classical nfs algorithm with the total scale set by
the current world-record largest implementation of it, in which 104 personal computers
running in parallel factored a 576-bit number in one month in 2003, according to rsa
Security, Inc. The curves labelled Shor C represent an implementation of Shors quantum algorithm in which logical qubits may only couple to nearest neighbors. The curves
labelled Shor E represent an implementation in which any logical qubit may couple to
any other logical qubit. These implementations are described by Van Meter and Itoh
[13]. Both algorithms use 100L logical qubits. For comparison, clock speeds of 1 Hz
or 1 MHz are shown; these clock speeds must include the time for extra processing for
quantum error correction.

N and either xr/2 + 1 or xr/2 1 is a factor of N . If r is odd, or xr/2 = 1 mod N , the algorithm
fails and should be repeated until it succeeds.
Most of this algorithm may be efficiently performed on a classical computer; the key quantum
mechanical steps are to feed a linear superposition state into the modular exponentiation operator
and the inverse qft. Details of both these quantum aspects and the classical aspects of the algorithm
may be found in Nielsen and Chuang [14].
Shors algorithm is the proving ground for quantum computation. Since the fastest known algorithm for factoring on a classical computer, the number field sieve (nfs), is super-polynomial
in number of steps, and Shors algorithm is polynomial, a quantum computer seems to offer dra-

12

Chapter 1. Introduction

matic speed-up over classical computers, challenging the strong Church-Turing Thesis3 . Figure 1.1
compares the speed of potential quantum computers for factoring n-bit numbers against optimized
networks of classical computers as a function of n. More practically, however, the slow speed
at which classical computers factor numbers is a critical assumption of the heavily used RivenShamir-Adleman (rsa) encryption system [15]. This system is frequently used for security in digital
communications today, and a working quantum computer would compromise that security. Shors
algorithm has therefore extended the interest in quantum computers from physics and computer
science academia to defense and financial sectors.

1.2.3

Other Algorithms

Shors algorithm is currently the most important quantum algorithm, but it is by no means the only
one. The quantum Fourier transform has other applications; these generalize to finding hidden
subgroups of a group on which the function implemented by a unitary operation is defined. Solving
the discrete logarithm problem is another example of an algorithm in this class.
Another important algorithm is Grovers algorithm [16], which searches an unstructured list
for a target bit string with polynomial speed-up over classical search algorithms. Although the
improvement in computation time is not as dramatic as in the case of Shors algorithm, the ubiquity
of searching gives this algorithm its importance.
Specific algorithms have also been devised for using quantum computers to simulate other quantum systems of interest, as proposed by Feynman. This is potentially one of the most important
uses for quantum computers in the future of technological development. A long-term dream is to
build a device that allows quantitative prediction for systems such as biological macromolecules,
high-temperature superconductors, and even atomic scale classical computers.

1.2.4

Quantum Error Correction and Fault Tolerance

None of these quantum algorithms will be useful unless the reality of decoherence is suppressed.
Quantum computers would not have any experimental promise if they were limited by the decoherence processes observed for even the most long-lived laboratory qubits. Fortunately, Shor and others
showed soon after the development of Shors algorithm that errors in quantum systems (including
decoherence) can be corrected, even if the tools used to correct them are themselves faulty!
The continuous Hilbert space in which quantum states live would seem to require error correction
3 It is important to note that a more efficient classical algorithm for factoring could exist; extensive efforts by great
minds to find it or prove its nonexistence have, however, met no success.

1.2 Theoretical Quantum Computing

13

similar to that needed for analog signals, which is notoriously more difficult than digital error
correction. However, just as the act of measurement stochastically digitizes the quantum state
of the quantum computer, so may it digitize a continuous set of errors, and then error correction
techniques similar to those for classical digital error processing are available.
The simplest example to demonstrate this point is a 3-qubit bit-flip code. Begin by considering
a classical digital computer, where it is possible for a bit to erroneously flip with small probability
< 1. A simple error correction technique is to copy this bit two times, and then periodically
correct any possible errors by examining the three copies and taking a majority vote. Then the
probability of error is reduced to the probability of two simultaneous bit flips, which, if the errors
are uncorrelated, has much smaller probability 2 .
We would encounter two problems if we attempted to naively implement this on a quantum
computer. The first is the no-cloning theorem, which tells us that we cannot simply copy the
unknown state of a qubit to another qubit. The no-cloning theorem is very simple to prove, but its
introduction to quantum theory has been a crucial advance for understanding the nature of quantum
information. The proof is by contradiction. Suppose a quantum cloning device C existed, which
could take an arbitrary, unknown quantum state |i and a second, initialized system |0i and copy
the unknown state into that second system, i.e.
C |i |0i = |i |i .

(1.10)

Already we see that C cannot possibly be unitary; the information about the initial state of the
second system has been lost. As long as the cloning works, however, we will allow C to increase
total entropy. But we may also see that C cannot be linear, for if we apply it to |i = |i we obtain
|i |i, and if we apply it to |i = |i =
6 |i we obtain |i |i. If C is linear then applying it to
|i = |i + |i yields |i |i + |i |i, which is different from the desired |i |i. Therefore such a
cloning device cannot exist in quantum mechanics, where allowable operations are non-negotiably
linear. This no-cloning theorem prevents us from simply copying a qubit in an arbitrary state and
taking a majority vote4 .
The second problem we would encounter in carrying our classical majority-vote bit-flip code over
to quantum mechanics is the fact that the error might not be a full bit flip, but rather a partial bit
flip. A qubit is equivalent to a spin-1/2 particle with angular momentum ~I; this angular momentum
may point along any direction of the Bloch sphere. For example, the state of qubit j may begin as
4 The

no-cloning theorem does not prevent us from generating an ensemble of qubits in the same state; this may be
done by initializing every qubit in a known state and processing them identically. Ensemble-based quantum computing
in this sense exhibits a simple form of majority-vote error correction.

14

Chapter 1. Introduction

Figure 1.2: A quantum circuit representing the 3-qubit bit-flip code. The qubit represented by the top line is encoded into three qubits with two controlled-not operations
on two ancilla qubits. A single qubit error occurs somewhere in the dashed box. The
syndrome and recovery operations are implemented with two controlled-not operations
and two polarization measurements; the top qubit is flipped if both ancilla are measured
in state |1i.

|1i, with density operator = 1/2 + Ijz , but the small error process, rather than flipping the qubit
to the other pure eigenstate |0i, simply depolarizes the qubit to some degree yielding the density
operator = 1/2 + (1 2)Ijz , for unknown < 1. We can describe this error process as
(1 ) + 4Ijx Ijx .

(1.11)

It would seem at first that we need an error correcting procedure that restores the qubit for any
continuous value of ! This plays the same role as the bit-flip probability in the classical digital
case only if we projectively measure the polarization of our qubit, which we do not wish to do if we
want to maintain its coherence for quantum information processing.
Shor [17] pointed out that these two problems are fairly easily overcome. Rather than copying
our unknown qubit, we entangle it with two extra ancillary qubits. This may be done by applying
controlled-not operations to each ancillary 
qubit, which flip
 the ancillary qubit only if the unknown
qubit is in state |1i, converting the state |0i + |1i |00i to |000i + |111i. The quantum
circuit describing this operation is shown in Fig. 1.2. The error process then occurs on each qubit
individually, taking
= |2 | |000ih000| + | 2 | |111ih111| + |000ih111| + |111ih000|

1.2 Theoretical Quantum Computing

15

to


|2 | (1 3) |000ih000| + |100ih100| + |010ih010| + |001ih001|


+| 2 | (1 3) |111ih111| + |011ih011| + |101ih101| + |110ih110|


+ (1 3) |000ih111| + |100ih011| + |010ih101| + |001ih110|


+ (1 3) |111ih000| + |011ih100| + |101ih010| + |110ih001| .

(1.12)

The key thing to notice about this quantum state is that all new terms are orthogonal to the
original density operator. These orthogonal states may therefore be projected out by a suitable
syndrome detection and recovery operation. In this case, the syndrome detection is accomplished
by attempting to disentangle the two ancillary qubits from the original qubits, using the same
controlled-not operations. This results in the state


| | (1 3) |000ih000| + |111ih111| + |010ih010| + |001ih001|


+| 2 | (1 3) |100ih100| + |011ih011| + |110ih110| + |101ih101|


+ (1 3) |000ih100| + |111ih011| + |010ih110| + |001ih101|


+ (1 3) |100ih000| + |011ih111| + |110ih010| + |101ih001| .
2

(1.13)

Note that the states in the original density operator are preserved, and all error terms have resulted
in a bit flip of one or both of the ancillary qubits. Moreover, if we have measured both of the
ancillary qubits flipped from |00i to |11i, then we know a bit-flip has occurred on our initial qubit,
which we may correct with a simple bit-flip gate. Our initial qubit is therefore recovered in all cases.
The reason our recovery from a continuous error is accomplished with a digital bit flip is the use of
projective measurement to collapse the Hilbert space occupied by our qubit to subspaces where the
qubit either completed its erroneous flip or did not undergo any error at all.
This procedure has completely eliminated error for the error model chosen in Eq. (1.11). Such
an error model might arise due to high-temperature spin-relaxation processes, where it is taken as
a discretization of a master equation of the form

X
d
x
x
=
4Ij Ij .
dt
j

(1.14)

16

Chapter 1. Introduction

In some small time t, the evolution of three qubits each evolving by this equation is iterated
(t + t) = (t) +

t+t

X

4Ijx (t )Ijx (t ) dt
j

(1 3t)(t) + 4t

X
j

Ijx (t)Ijx + 82 t2

Ijx Ixk (t)Ijx Ikx + . . . .

jk

This resembles the error model we have corrected with = t, but we have neglected correlated
errors, which occur to order 2 in this master equation approach. Our three-qubit code fails to
protect against these terms. They may be ignored given a sufficiently fast error correction period
t. In the end, our quantum error correction process has performed the same as our classical digital
error correction, eliminating bit-flip errors that are first order in with the use of two extra qubits
to redundantly store the initial information.
This example discussed only one type of error, the longitudinal depolarization of individual qubits
as described by Eq. (1.14). For other types of error, different encodings, syndrome measurements,
and recovery operations are possible, and these may be concatenated to protect against many types
of error simultaneously. Shors original nine-qubit code concatenated the above three-qubit code,
protecting against longitudinal depolarization, with a similar three-qubit code in an orthogonal basis
to simultaneously protect against transverse depolarization of individual qubits. An independent
approach to quantum error correction by A. Steane [18] was also developed at about the same time.
More efficient codes can be derived using techniques from classical digital error correction. One of
these so-called Calderbank-Shor-Steane (CSS) codes [19, 20] uses only 5 qubits to correct against
the same errors as the 9-qubit Shor code. The useful stabilizer formalism [21] has been a crucial aid
in the development of codes.
Another type of protection against decoherence operates on a different principle [22, 23]. This
is the use of decoherence-free subspaces. This type of protection may be succinctly summarized
as the exploitation of the symmetry of the error model to encode qubits in subspaces that are
invariant to the error operation. For example, suppose the qubits are spin-1/2 magnetic dipoles in
an environment with a global fluctuating magnetic field. The symmetry group for global rotations
of the qubits admits multiple irreducible representations, some of which correspond to angular
momentum singlet states. For two spins, there is a one-dimensional singlet subspace corresponding
to the Bell state |i |i. As a one-dimensional subspace, this state undergoes no evolution
under global rotations, and therefore the noisy global field cannot change it. If the fluctuating field
is strictly in the z-direction, a further invariant subspace for the two-qubit system is the m = 0
state of the triplet subspace, the Bell state |i + |i; in this high symmetry case these two states

1.2 Theoretical Quantum Computing

17

may provide a decoherence-free qubit. If the fluctuating magnetic field has components in all spatial
directions, only the singlet-states are invariant to it; fortunately, among a large number n of spin-1/2
dipoles there is a large degeneracy of singlet states. To be precise, there are
n!
2n1

(n/2 + 1)!(n/2 1)!


n
singlet states, which is most of the states! Asymptotically, then, one may encode O(n log n)
noiseless qubits. Decoherence-free subspaces have already proven to be a valuable error-protection
tool in a variety of experiments, especially with trapped ions [24, 25] but their applicability is
ultimately limited due to the requirements of symmetry in the noise model.
A questionable assumption in these quantum error correction schemes is that the resources they
require can be reliably collected without making the initial error probability worse. Suppose that
we hope to correct one type of error. This could occur due to an ongoing physical noise process but
could also be due to erroneous state preparation, faulty logic, or imperfect quantum measurement.
To correct this error, we must prepare more states, perform more logic, and make more quantum
measurements. The time to perform these additional resources or the ongoing imperfection of our
system means that in fixing the original error, we could make the problem worse by causing more
errors. What is truly required for large-scale quantum computation is fault tolerance, a construction
in which state preparation, quantum logic, and even quantum measurement are performed so that
each add a sufficiently small amount of error per use that the error-correcting protocol may be
concatenated with itself indefinitely to achieve an arbitrarily accurate quantum computer. It was
again Shor [26] that introduced this notion of fault tolerance into quantum computation, and a
substantial number of researchers have developed constructions to show that fault-tolerant quantum
computation is possible, provided the errors in state preparation, logic, and measurement fall below
threshold values.
Suppose that our controlled-not gate is faulty, so that a bit-flip error occurs on the target
qubit with probability every time we use the gate. If we did no error correction at all, then
every controlled-not gate in our algorithm would accumulate an error and accurate computation
would not be possible, so we decide to employ the 3-qubit code discussed above. We have shown
that this encoding corrects bit-flip errors to order 2 . During the act of encoding of this code, we
use one controlled-not gate on each ancilla, and the syndrome measurement requires one more
controlled-not gate on each ancilla, so the very act of encoding introduces an error of order 2,
which is corrected to order (2)2 . As long as 1/2 this may be acceptable, but perhaps we
desire even more accuracy. We could improve our error-correction by encoding our encoded qubits

18

Chapter 1. Introduction

one more time. This means 2 controlled-not gates for each of our encoded qubits. For this to be
done fault-tolerantly, we must make sure that errors dont propagate; if a physical controlled-not
gate causes only a single bit error with probability , then an encoded controlled-not gate should
not cause more than one error in the encoded bit. In particular, a controlled-not gate between
two encoded qubits in the 3-bit code can happen qubit-by-qubit, preventing correlated errors from
accumulating during the gate. The encoded controlled-not therefore introduces an error no larger
than the residual error for a bit flip on the encoded qubits, (2)2 . The total residual error after two
k

levels of encoding is therefore 2[(2)2 ]2 . After k levels of encoding, the residual error is (2)2 /2, a
number which quickly vanishes as k increases, as long as is smaller than the threshold of 1/2. If
all controlled-not gates are performed fault-tolerantly on the encoded qubits, the total error after
N gates is this small number multiplied by N . In this case, we need 3k qubits and just as many
operations per controlled-not gate. The encouraging conclusion is that if we can tolerate a total final
error , then the number of physical qubits and logic gates needed grows only polylogarithmically
in N and 1/.
This simple example hides the daunting resources actually required for fault-tolerant quantum
computing. In our example, we supposed that our only error was the random bit flip during a
faulty controlled-not; in reality there can be many types of error occurring during many types of
operations. We also only discussed the controlled-not gate for the 3-qubit bit-flip code, a gate which
is easy to make fault tolerant. In reality, many quantum gates do not have a natural fault-tolerant
architecture, in the sense that a naive construction causes low-level errors to propagate excessively
among encoded blocks of qubits. Creating fault-tolerant gates may require large additional resources
in ancillary qubits, projective measurements, and low thresholds depending on the gate, the error
model, and the encoding. More realistic threshold values for typical error models and fault-tolerant
architectures are of order 105 or worse, with individual fault-tolerant procedures requiring more
qubits and physical logic gates than have been implemented in any actual experiment to date. I
offer a brief discussion of the physical resources required in Chapter 2; a reader interested in the
further details of the theory of quantum error correction and fault-tolerant quantum computation
is advised to begin with Chapter 10 of Nielsen and Chuang [14].
Despite the physical difficulty of reaching the threshold for fault-tolerant quantum computation,
the very existence of a threshold gives hope that quantum computers could be realistically scaled to
arbitrarily large size, even in the presence of inevitable decoherence processes. Were it not for the
developments in quantum error correction and fault tolerance, there would be no hope for quantum
computation.

1.3 Experimental Quantum Computing

1.3

19

Experimental Quantum Computing


Either it is possible to build a quantum computer, or it isnt.

This trite little sentence is more profound than it seems, and as such has motivated the research
I present in the remainder of this dissertation. The key point is that if we show that quantum
computers are possible, which we can only do by actually building one, then mankind will have
an information processing tool unlike any it has possessed before. However, if by some discovery
or theory we are able to conclusively show that quantum computers are not possible, then the
argument against them will add a crucial chapter to the development of quantum mechanics and
the foundations of physics.
But grim reality must not be forgotten. We are not currently able to build quantum computers,
but the barriers that prevent us from doing so are far from fundamental. Those who build quantum
computers from trapped ions have trouble with random magnetic fields; those who build them from
superconducting Josephson junctions have impurities in oxide layers; those who build them from
photons and beamsplitters have spurious reflections from the surfaces of their photodetectors. Any
physical apparatus used to build a quantum computer must be extremely clean and efficient, and
ordinary technical problems are currently the only clear obstacle between reality and a world of
quantum computers.
If progress is to be made, we can only attempt to chip away at this mountain of technical
problems. There are many possible paths to the quantum computer, and until one succeeds we
cannot know whether quantum mechanics will provide us with a tool for more and more complex
technology in the distant future, or whether new physics will be found along the way to show us why
we are doomed to macroscopic classicality. Either way, however, the technical problems solved in
the quest will provide valuable knowledge with unknown future benefit to both our understanding
of the universe and our ability to manipulate it.
This dissertation focusses on one class of paths toward quantum computing; these are paths
making use of nuclei in the solid state, in particular without the help of electron wave function
engineering. In terms of the amount of money and the number of researchers and published papers,
the efforts described here represent a quiet minority in the race to quantum computing, but no one
can easily say which efforts will bear the most important fruit in the final harvest of technologies
for quantum computers.
In the following chapter I describe the needed resources for quantum computation, and then
without much technical detail the various ideas for meeting these resources in quantum computing

20

Chapter 1. Introduction

research around the globe. In Chapter 3 I focus on technologies for quantum computers using solidstate nuclei. The advantage of nuclei are their extremely long relaxation times, and I discuss the
origins of nuclear relaxation processes in ensuing chapters, especially in silicon which has particular
promise for quantum computing architectures. In Chapter 4 and Chapter 5 I describe an experimental exploration of the decoherence properties of silicon nuclei. Relaxation processes in nmr are
outlined in Chapter 6 and applied to silicon in Chapter 7. The notation used in all of these chapters
is explained in Appendix A.
The focus of the material in this thesis is on physical theory and experiment toward quantum
computing. There is little further discussion of algorithms, error correction, or communication
protocols. In other words, this work deals strictly with the hardware, and leaves the software
to others. Readers interested in algorithms and protocols are invited to begin with Nielsen and
Chuangs textbook on the subject [14].

Chapter 2

Physical Resources for Quantum


Computing
Somewhere around the place Ive got an unfinished short story about Schrodingers Dog;
it was mostly moaning about all the attention the cat was getting.
Terry Pratchett

2.1

Three Essential Requirements for Fault Tolerant Quantum Computation

There is much room for creativity in the construction of quantum computing proposals. DiVincenzo
[27] has summarized the physical implementation of quantum computers and has listed five requirements for a quantum computer, although even these as originally stated may be too restrictive, and
are certainly not independent. In this chapter, I regroup these into three requirements, and focus
on the one which is especially challenging for nmr-based approaches to quantum computation.

2.1.1

An exponentially large and coherent Hilbert space, without exponentially large space, time, or energy.

The easiest and by far most common example of a large Hilbert space is a collection of separate
qubits. A set of N qubits each with energy separation ~0 and effective volume V can be assembled

22

Chapter 2. Physical Resources for Quantum Computing

in volume N V and energy N ~0 . Even if these qubits are arranged in a one-dimensional line and
are only allowed to interact with their nearest neighber, the time required for universal quantum
computation (that is, the time to implement sufficiently many logic gates to approximate any desired
unitary operator on the qubits to a desired accuracy) is polynomial in N . However, the number
of physical logical gates required to allow fault-tolerant logic among arbitrary encoded qubits is
so large that fault-tolerant thresholds are extremely low for such one-dimensional, nearest neighbor
geometries; an architecture in which arbitrarily distant qubits can rapidly interact is much preferred.
Physical qubits are not absolutely essential. Quantum information processing can in principle
be performed with n-state systems instead of 2-state systems, or even with continuous quantum
variables. The only important consideration is that multiple quanta are a critical component of
establishing an exponentially large Hilbert space. As an example that abandons this principle,
suppose we had sufficient control to access the infinity of bound states of the hydrogen atom; for
an n-qubit computation we use 2n of its infinite states. The energy is clearly bound, since none of
these states exceed the 13.6 eV binding energy. This is certainly a large Hilbert space contained in
non-exponential energy. However, the effective volume of the atom grows exponentially, as the 2n th
state has roughly a radius of 2n Bohr radii. If we used the energy states of an electron trapped
in a hard box, the size would always be limited, but the energy would grow exponentially. These
requirements have been quantified by Blume-Kohout et al. [28].
Of course, decoherence can also limit the effective size of Hilbert space available. Decoherence
prohibits the existence of those coherent quantum states that make up most of Hilbert space and
limits the long-time dynamics to the small subspace of classical states. Decoherence must be sufficiently slow in comparison to the rate at which we may correct for it. Decoherence times must be
preserved as the size of the system is increased.
Physical systems that allow the addition of logical qubits without exponential increases in physical size or energy, and without drastic change in the ability to reach fault-tolerant thresholds, are
said to be scalable. Although many proposals for quantum computers have been called scalable,
none have yet shown experimental evidence of this difficult property.

2.1.2

Universal Quantum Logic

Actually performing the computation requires sufficient control resources. In the standard model,
these resources are single-qubit unitary gates (Hadamard, /8 rotations, and phase-gates) and twoqubit unitary gates (controlled-not), although the choice of specific gates is not universal [29, 30].
Further, the controlled-not gate may be efficiently constructed from almost any two-bit interaction

2.2 Opening the Box

23

in collaboration with arbitrary one-qubit gates [31]. In this sense, universal quantum logic has been
shown to be the easiest criterion to obtain. However, performing quantum logic with sufficient
accuracy to allow fault-tolerant operation is a challenge for every proposal.
Another important result shows that even the necessity for one-bit gates may be dropped. Bacon
et al. [32] and Levy [33] have shown that universal logic for qubits encoded in decoherence-free
subspaces may be performed using only controllable exchange interactions.
Quantum logic may also be enacted in models other than the standard gate model. In particular,
non-unitary operations (that is, operations involving measurement) can also be used. In Sec. 2.4.3
we discuss quantum computing with linear optics, which employs no direct nonlinearity between
photons but uses the inherent nonlinearity of quantum measurement. A related idea is the clusterstate quantum computing model, in which the entire computation is accomplished by assorted
measurements of a large initial entangled state of qubits. These types of architectures and the
relations between them have recently been summarized by Childs et al. [34].

2.1.3

Initialization to a Fiducial Initial State and Measurement of an


Output State

The initialization of a computer and the ability to read out the state of arbitrary qubits may seem
like two requirements, but in fact they are closely related. These are both non-unitary operations,
coupling the quantum information inside the computer to the outside world. Indeed, if we can
projectively measure the current state of the system without destroying it, and we can apply universal
logic, we can initialize to any state we want following the measurement. In this sense the ability
to perform rapid, nondestructive, projective measurement on arbitrary qubits implies the ability to
initialize. The converse is also reasonable; the ability to rapidly initialize individual qubits could
remove the need for rapid projective measurement during quantum error correction. The only
measurement then needed would be a read-out of the computational result, which could be slow
and even non-projective (as in a weak measurement of an ensemble.) These considerations are not
obvious, but they are of crucial importance for nmr-based quantum computers. I therefore discuss
them at greater length in the next section.

2.2

Opening the Box

In Chapter 1 I likened a quantum computer to Schr


odingers cat. A key difference is that Schr
odingers cat lived and died inside a closed box, with all quantum coherence trapped within. After

24

Chapter 2. Physical Resources for Quantum Computing

opening the box to check on the cats health, no evidence of the quantum coherence remains. In
the case of the scalable quantum computer, we may gain evidence of macroscopic coherence after
opening the box in the end, but this is not sufficient; we need selective coupling of the boxs interior
to the outside world in order to allow error correction, either using localized selective measurements
or rapid initialization of individual qubits.
What is difficult about initialization is that we must erase the information that is already present
in the system, an action tantamount to cooling the system or subsystem to absolute zero temperature.
In other words, the entropy before initialization must be completely removed from the qubit system.

2.2.1

Entropy

The von Neumann entropy of a quantum system is


S = Tr { log } .

(2.1)

If the logarithm is base 2, the entropy is measured in bits; if a natural log is used, the unit is nats.
In the basis in which is diagonal, the diagonal elements form a classical probability distribution;
in this case Eq. (2.1) gives the Shannon entropy, usually denoted H. Indeed it was von Neumann
who recommended that Shannon term his information measure entropy.
The entropy is a measure of uncertainty about a quantum system or probability distribution.
The entropy may be used, for example, to quantify decoherence. An ideal quantum computer is a
closed system, but environmental noise and operational error inevitably raise the entropy. As the
simplest example, consider a single qubit (spin operator I), perhaps a nuclear spin, coupled to a
two-state system in the environment with spin operator S, perhaps an impurity electron spin. The
I, and the environment is in thermal equilibrium,
qubit begins in an arbitrary state I (0) = 1/2 + n
S (0) = 1/2 + pS z , where p 1 is the polarization of the environment. This definition implies that
(1 + p)/2 is the probability of finding the environment in the |i state. The initial entropy of the
qubit is zero and the initial entropy of the environment is
H(p) =

1+p
1+p 1p
1p
log

log
.
2
2
2
2

(2.2)

Suppose our qubit is coupled to the environment via an interaction of the form AI z S z ; after a time

2.2 Opening the Box

25

t the combined state of the system is


z

IS (t) = eiAtI S I (0) S (0)eiAtI S









1
At
At
At
At
1
= + nx I x cos
+ 2I y S z sin
+ ny I y cos
2I x S z sin
+ nz I z
+ pS z .
2
2
2
2
2
2
(2.3)
The interaction of the environment with the qubit has entangled the two systems, but the total entropy is not increased; it remains SIS = H(p). Decoherence occurs when we discard the information
about the environment; the mathematical procedure is to trace over the environment variables. All
terms in Eq. (2.3) with an odd number of S z operators therefore vanish, since this operator indicates
dependence on the environmental spin and as such is traceless. The result after this operation is
I (t) =





1
At
At
At
At
+ nx cos
pny sin
I x + ny cos
+ pnx sin
I y + nz I z .
2
2
2
2
2

(2.4)

The transverse components of hIi have apparently shrunk; this is decoherence. The entropy of the
qubit has increased from zero to H(r), where r is the length of the qubits vector in the Bloch sphere:
r = 1 (1 p2 ) sin2 sin2

At
.
2

Here, is the initial polar angle of the qubit. Notably for this model the entropy of the environment
has remained constant; the coupling dynamics did not affect it. The total entropy has therefore
increased:
SI + SS = H(r) + H(p) SIS = H(p).

(2.5)

This will always be the case; it is the second law of thermodynamics. One notable extension to this
model is to consider the coupling to a large ensemble of unpolarized (p = 0) spins in the environment,
assuming environment spin j has coupling strength Aj . In this case the transverse components of
hIi given by the reduced density operator shrink by the factor
Y
j

X A2j
Aj t
cos
exp
t2 .
2
8
j

For large numbers of environment spins, the removal of transverse components appears to be irreversible. This is illusory, however, as it results from our neglect of the quantum information stored in
the environment when we traced over its variables. The information may in many cases be restored

26

Chapter 2. Physical Resources for Quantum Computing

to the qubits by a spin echo. This and similar techniques will be discussed in Chapter 4. The
above model for decoherence is a very simple example; a more sophisticated model for analyzing
such processes will be developed in Chapter 6.
Let us consider an example of entropy accounting when quantum error correction acts to preserve
the entropy of a qubit. For this we return to the simple bit-flip error model introduced in Eq. (1.11)
on page 14 and the protocol we discussed to correct for it. The bit-flip error process clearly increases
the entropy of the system. For example, the entropy of the state in Eq. (1.12) may calculated to be
S = (1 3) log(1 3) 3 log .

(2.6)

This is the same as the Shannon entropy for a 2-bit probability distribution
P00 = 1 3,

P01 = P10 = P11 = .

(2.7)

This probability distribution corresponds to the probabilities that no error happened (the 00 case)
or that one of the three qubits was flipped. Our syndrome measurement for this code extracts
exactly this much entropy from the quantum computer; the syndrome measurement (decoding and
polarization measurements) distinguishes between the measurement projectors
E00 = |000ih111| + |111ih000|
E10 = |010ih010| + |101ih101|
E01 = |001ih001| + |110ih110|
E11 = |100ih100| + |011ih011| .

(2.8)

The probability distribution of Eq. (2.7) with entropy S given by Eq. (2.6) is found by Px = hEx i. By
making the projective measurement corresponding to the four projectors Ex , an entropy of exactly
S is removed from the system. If the result x is measured, the resulting density operator
=

Ex Ex
,
Tr {Ex Ex }

is, of course, the initial density operator of the original qubit, except with a spin-flip in the case
x = 11. This density operator has zero entropy. (If the initial qubit had some entropy, the final
density operator after the error correction would also have exactly the same amount).
It may appear that the entropy created by the spin-flip error was obliterated, as in the famous

2.2 Opening the Box

27

paradox of Maxwells demon. In reality, there is no Maxwells demon. Rather, the entropy was
transferred outside the quantum computer, where the random measurement result increased the
entropy of some external classical information processor.
Entropy is useful as a measure of information and its conservation, but it is also the usual entropy
from statistical mechanics; these are the same. If a system is in thermal equilibrium with a large
heat bath, the density operator is found by maximization of the entropy at constant average energy,
leading to the Boltzmann distribution. The entropy in nats and temperature T of the system then
satisfy
S=

hEi E0
+ ln Z,
kB T

(2.9)

where E0 is the ground state energy and Z is the partition function.


Initialization of a quantum computer could mean bringing the system to the ground state, in
which case hEi = E0 and ln Z = 0. However it can also mean bringing the system to any other
fiducial pure state, which may be out of thermal equilibrium, but still resulting in S = 0. Crucially,
the initial state does not need to be a pure state at all. Quantum computing is still possible with
mixed states, if the structure of the probability distribution is well known. Solution nmr quantum
computing is a prime example of this; it is not inherently the mixed initial states that prevents these
schemes from being unscalable, as we discuss in the next section. Although I will not discuss it here,
an especially provocative demonstration that quantum computers beginning with large amounts of
entropy may still outperform classical computers was given by Knill and Laflamme [35].
The equivalence of the thermodynamic entropy and the entropy as information measure reminds
us that computational error is equivalent to heating. Initialization and error correction are correspondingly equivalent to cooling. In the next two sections, I discuss algorithmic methods for both.
They both resemble the way we cool samples to very low temperatures in the laboratory: we put
them in contact with gas molecules from a liquid helium bath. These very cold molecules interact
with the sample, allowing entropy to transfer from the sample to the gas. In principle it could
transfer back, bringing the sample to the temperature of the helium bath, but we cool the sample even further by pumping away the hot gas molecules. Algorithmic cooling and quantum error
correction without measurement are very similar processes, in which entropy is transferred to cold
ancilla qubits; the resulting hot qubits are then removed from the computer.

2.2.2

Initialization

In Sec. 2.4 I will discuss how various architectures prepare their qubits in a pure state, or close to
it. In this section, I approach the question of how close to a pure state the physical architecture

28

Chapter 2. Physical Resources for Quantum Computing

must bring the qubits in order for quantum computation to still be practical. These arguments are
not limited to nmr quantum computers, although the low energies of nuclear spin splittings make
these quantum computers the most difficult to initialize. For this reason, and to connect with the
notation elsewhere in the dissertation, I use nmr notation, labelling qubits each as spin-1/2 with
spin operator I.
We suppose the physical cooling applied to our qubits brings them to a Boltzmann distribution
with some temperature T , and hence to the state
T =


Y1
exp(~H/kB T )
=
+ pIjz ,
Tr {exp(~H/kB T )}
2
j

where
p = tanh

~0
2kB T

(2.10)

(2.11)

For this analysis, only the energy of each qubit ~0 is important. Here, we are focusing on one
quantum computer in the ensemble; the sum over j is the sum over n qubits. Even though this
state is mixed at finite T , quantum computation is still possible by a variety of means. For example,
Q
imagine that we could arrange our experiment to measure the observable operator j Ijz . After
we measure
performing some arbitrary computation superoperator C,
Y
j

Y Y
Y Y 


Ijz C T =
Ijz C pIkz = pn
Ijz C Ikz ,
j

(2.12)

since all other terms are traceless. If the computation has been arranged so that the eigenvalues of
Q z
j Ij reveal the desired information, then this process has yielded the same result as if the qubits had
begun in a pure state, except that the strength of the signal has been uniformly reduced by pn . This

uniform reduction in signal is severe, as p is of order 105 at reasonable temperatures and magnetic
fields. To maintain a measurable signal, the size of the ensemble N must grow exponentially to keep
N pn above the measurement threshold.
This example was not the ideal choice of measurement. The best choice is
M=

Y 1

which yields a signal of size


S=

Y 1


Ijz ,

(2.13)

n

n

(2.14)

+ Ijz

1+p
2

1p
2

We consider the asymptotics of this expression in two limits. If polarization begins very low, we

2.2 Opening the Box

29

1000
n

10

-20

100
10

-8

10

-3

0.1

10

0.5

1
0%

p
20%

40%

60%

80%

100%

Figure 2.1: The number of pure qubits n which may be extracted at finite polarization
as a function of the polarization p. Each curve is labelled by the total sensitivity, which
is the number of qubits one is capable of measuring divided by the number of qubits in
the ensemble (or the number of repetitions of the experiment).

have
S np/2n1 ,

p 1,

(2.15)

which is still exponentially bad. The inevitable exponential loss of signal in high-initial-entropy
quantum computers was pointed out in the first proposals for liquid-state nmr architectures [36, 37]
and raised skepticism for the future of quantum computation based on nmr [38]. However, if high
levels of nuclear polarization can be obtained, the scaling of Eq. (2.14) changes. This function is
plotted in Fig. 2.1, where it may be seen that for p & 70%, the scaling switches to a nearly linear
dependence:
S 1 n(1 p),

p 1.

(2.16)

Bringing the polarization close to unity is essential for scalability.


Usually, initialization will involve physical cooling, meaning the coupling of the quantum computer to some very large, cold bath. In many architectures, simply placing the device in a bath
of liquid helium or a dilution refrigerator will do. However, algorithmic techniques have also been
considered for cooling. An algorithm cannot reduce the entropy of a system; ideal reversible computation techniques may at best conserve the entropy of a system. However, entropy may be localized

30

Chapter 2. Physical Resources for Quantum Computing

in a subregister of qubits, and these qubits may be removed from the computer. Suppose we begin
with n qubits, each with polarization p and therefore each with entropy H(p), as given in Eq. (2.2),
measured in bits. The total entropy is then S = nH(p). Suppose this system is closed, so the total
entropy cannot change. An ideal algorithm cools a subsystem of k qubits to zero entropy and puts
as much entropy as possible into the remaining n k qubits; these therefore each have a full bit
of entropy, and their combined n k bits of entropy must be the initial entropy S. It follows that

k/n = 1 H. For small p, we may approximate Eq. (2.2) as 1 H p2 /2 ln 2 + O(p4 ). Therefore if


p begins rather high, say 30%, then one in fifteen qubits may be cooled with an ideal algorithm.
A nearly ideal algorithm was proposed by Schulman and Vazirani [39]. It relies upon the following
simple boosting procedure. Note that this algorithm is strictly classical. Consider two bits; the
bit in location 1 begins in state A and the bit in position 2 begins in state B. Here A and B
are taken as random binomial variables. Initially both bits have polarization p0 , so for example
P (A = 0) = (1+p0 )/2. The two bits are uncorrelated, so P (A = a|B = b) = P (A = a). We begin by
correlating the two bits with a controlled-not operation, changing the states to [A , B ] = [A, B A].
Due to the new correlation,
P (A = 0|B = 0) =

P (A = 0)
1 + p0
=
,
P (A = B)
1 + p20

P (A = 1|B = 0) =

P (A = 1)
1 p0
=
,
P (A = B)
1 + p20

(2.17)

and therefore in the case B = 0 the polarization of bit 1 is 2p0 /(1 + p0 )2 . The polarization appears
to have doubled in this case, and therefore, we move this colder bit to a higher position by swapping
it the bit in position 3. If B = 1, then no further action is performed. Of course we do not
measure B ; this entire operation may be implemented by a unitary operator employing conditional
operations. If the bit in position 3 also began with initial polarization p0 , the resulting final total
polarization has been cooled to (3/2)p0 (1/2)p30 ; that is, for small p0 its polarization has increased
by 50%! The bit in position 2, meanwhile, has been heated; its final polarization is reduced to
p0 /2 + (1/2)p30. The first bit ends up hottest of all, with the very small polarization p20 . Repetition
of this boosting procedure can bring a subregister of bits nearly the size of the entropy-limited
subregister to a polarization that approaches unity exponentially as the number of initial bits n
grows, in only O(n log n) steps, as shown by Schulman and Vazirani [39]. An initial experimental
demonstration of this boosting procedure in a liquid-state nmr quantum computer was implemented
by Chang et al. [40].
Subsequent research along this line has considered the possibility of improving the procedure
in the following way. In the ideal algorithm mentioned above, the n k = nH discarded bits
were heated to their full entropy of 1-bit each. Suppose we could selectively cool every bit in

2.2 Opening the Box

31

Figure 2.2: A modification of the quantum error correction circuit shown in Fig. 1.2,
in which the measurements with feed-forward control have been replaced by a coherent
Toffoli gate for recovery. After the recovery, the ancilla qubits must be discarded or
reinitialized.

this entire subregister back to its initial polarization p0 [and therefore their initial entropy H(p0 )],
while maintaining polarization of the cold k = [1 H(p0 )]n bits. Applying the same ideal cooling

procedure, the total entropy in the second stage is reduced to nH 2 (p), so we may now fully cool
k2 to zero entropy, where (n k2 ) = nH 2 (p). After m stages of this, the number of cold bits is

n[1H m (p)], which could approach the entire register as m increases! It may be physically unrealistic
to dynamically change which bits are physically cooled and which are preserved, however; a more
realistic design is to imagine one refrigerant bit which may be cooled at any time to some bias
polarization p, and n logical bits which we hope to cool algorithmically. A recent such openbath algorithmic cooling protocol has shown that as long as p 2n , a fraction of the n bits only

logarithmic in 1/p need be discarded to polarize all spins to a final polarization of 1O((1log p)1 ).

The maximum number of steps to achieve this result is 4np2 (1 log p), a nearly optimal result [41].

2.2.3

Measurement

The ability to rapidly and selectively cool a subset of qubits while conserving the state of the
remaining qubits is not only important for initialization; it may also reduce the requirements on
measurement. Consider as an example the simple bit-flip error correction protocol discussed in
Sec. 1.2.4. In this protocol, a single logical qubit was encoded in three physical qubits; two of those
qubits were measured (syndrome measurement) and the third may have been rotated depending on
the result of that measurement (recovery). In this model, it is assumed that we may repeat the
encoding, syndrome measurement, and recovery processes repeatedly, which would seem to require

32

Chapter 2. Physical Resources for Quantum Computing

many projective measurements during the quantum algorithm. The state of the ancilla qubits is
known after measurement, and therefore they may in principle be rapidly re-used. However, it is not
hard to see that for this and similar protocols, it is sufficient to simply refresh (that is, re-initialize)
those ancilla qubits. To demonstrate, consider the circuit shown in Fig. 2.2 as a modification to the
error-correction circiut of Fig. 1.2. The polarization measurement and recovery operation has been
replaced with a coherent, or reversible, logic gate, in this case a not gate controlled by two qubits,
known as a Toffoli gate. The density operator after the decoding step was given by Eq. (1.13); if we
now apply the Toffoli gate to this density operator we obtain


| | (1 3) |000ih000| + |011ih011| + |010ih010| + |001ih001|


2
+| | (1 3) |100ih100| + |111ih111| + |110ih110| + |101ih101|



+ (1 3) |000ih100| + |011ih111| + |010ih110| + |001ih101|





+ (1 3) |100ih000| + |111ih011| + |110ih010| + |101ih001|




 

= |0i + |1i h0| + h1| (1 3) |00ih00| + (|11ih11| + |10ih01| + |01ih01|) . (2.18)
2

The first qubit is back to its original state and is completely separable from the other two qubits
which now carry the full entropy S of the error, given by Eq. (2.6). These qubits may simply
be discarded. They cannot be reused for the next cycle of error correction unless that entropy is
removed by the initialization procedure.
Replacing measurement and recovery operations with refreshed ancilla qubits may be readily

generalized. Suppose we wish to do a projective measurement on a general state , which may


include an arbitrary number of qubits. For example, may be an encoded pure state which has

been corrupted by noise. We suppose there are m possible measurement results with measurement
superoperators Ei . In the example above, m = 4, the measurement operators are given by Eq. (2.8).

If we achieve syndrome result i, then our recovery unitary is Ui . We may introduce an m-state

ancilla system (that is, log m qubits) initialized to 0 A and apply the following unitary operator to
the combined system:

m1
m1
X
X
 
 
 
0 =
j E
j j =
U E E U j .
U
U
j j
j j
A
A
A
j=0

(2.19)

j=0

That this operator is unitary may be seen by taking the trace of the right-hand side and noting

2.2 Opening the Box

33

that it preserves the trace of the original state. For this one needs the completeness relation for the
measurement operators
m1
X

Ej Ej = 1.

(2.20)

j=0



If this process was an error-correcting step, the syndrome-detected and recovered state Uj Ej Ej Uj

is proportional to the original state ; the proportionality constant is the probability of obtaining

syndrome j, Pj = Ej . Thus the resulting state following this unitary is
X 
 
 m1
0 =
U
Pj j .
A

(2.21)

j=0

If the error correction was perfect, the ancilla qubits carry the full entropy of , which evidently was
H(P).
This observation must be made with caution, however, especially in view of fault tolerance. This
error-correcting unitary operator U exists and can usually be constructed with polynomial resources,
but it may be difficult to implement U in a naturally fault-tolerant way. In particular, fault-tolerant
implementations of U may themselves require measurement, which therefore require more refreshed
ancilla and conditioning unitaries. The error threshold for fault tolerance in a scheme without
measurement may be extremely difficult to obtain, or it may not exist at all. Whether fault-tolerant
quantum computation is possible without projective measurement is still an open question, and
certainly the answer depends critically on the error model and the encoding scheme used.
There are two purposes of measurements in architectures for quantum computation that cannot
be replaced with conditioning unitaries and fresh ancilla. The first is the use of measurement
as verification of the generation of particular ancilla states. Many architectures for fault-tolerant
processes require ancilla qubits to be prepared in entangled states with high fidelity, despite faulty
procedures for state preparation. This is possible by generating the states in a faulty way and then
introducing a non-destructive measurement which verifies that they are in the correct state. If the
verification fails, a recovery unitary is not implemented as in the error correction case; rather, the
qubits are refreshed and preparation is attempted again until verification succeeds.
The second use of measurement which is obviously unavoidable is the final readout of the computer output. Fortunately, however, this output is measured in the computational basis, and the
measurement can in principle be much slower than coherence times in the system. For most algorithms, and in particular Shors algorithm, a weak ensemble measurement that measures the mean of
an observable without wave function collapse is sufficient. However, this form of measurement may

34

Chapter 2. Physical Resources for Quantum Computing

still be a difficult component to physically implement, as each qubit must be individually measured
without depolarizing other qubits.

2.3

Quantum Memory and Quantum Repeaters

Large fault-tolerant quantum computing may not be possible in a single device. Scaling considerations may limit the number of qubits in a device, or certain resources such as efficient measurement
schemes may require communication between a register of qubits implemented with one physical
hardware to another. But how may quantum information be reliably communicated? I now only
briefly discuss this large field of study.

2.3.1

The Reasons for Quantum Communication

There are three important motivations for the ability to receive, remember, and transmit a quantum
state faithfully. The first is the important application of quantum cryptography, in which the intrinsic
randomness of the first projective measurement of a single quantum state is exploited to share secrecy
between two parties. Several schemes have been devised to assure that the communicating parties
are indeed making the first projective measurements, so that those states previously measured by
an eavesdropper make a sufficiently negligible contribution to the resulting secret key. Another
important application for quantum memory and communication is for fault-tolerant linear optics
quantum computing architectures which rely almost entirely upon it.
The third motivation, which will be the focus here, is to alleviate the resources required for
quantum computing. The ability to move quantum states from one quantum computing device to
another enables those devices to be very small, perhaps only tens to hundreds of qubits, but to
still have the ability to perform large quantum algorithms as part of an extended network of such
devices. A further extension is the possibility of incorporating different technologies into a quantum
network; for example, superconducting devices with their high clock speed could be used for rapid
processing tasks while quantum information is stored in semiconductor nuclei or trapped ions, which
have slower operation times but much longer coherence times.
Indeed, the modern classical computer would not function well if the capacitive dynamic random
access memory were the same as the magnetic hard drive, or if the hard drive attempted to use
the same technology as the dynamic random access memory. Consider further that for classical
computers, the modern internet has revolutionized society for business, government, and leisure.
An individuals personal computer in his living room becomes orders of magnitude more powerful

2.3 Quantum Memory and Quantum Repeaters

35

once it is connected to the internet with a reasonably fast rate of communication; it now has the
potential to be as large as all the worlds libraries, and to obtain amounts of data which could
not be contained in that living room were it full of hard drives. Likewise, if the problem of faithful
quantum communication and quantum memory can be tackled, the problem of making large quantum
computers may be reduced to the problem of making small quantum computers.
The only qubit relied upon to transmit information over long distances is the photon. Massive
particles such as electrons, ions, or neutral atoms have very short effective coherence lengths in
any realistic laboratory, and while transport of these qubits may be useful in some architectures
employing them, it will not be sufficient for an extended quantum network. Photons, in particular
infrared photons in fused-silica optical fibers, may travel many kilometers without any substantial
loss or decoherence1 . Photons are not perfect, however, and the loss of photons from optical fibers
places the most serious bottleneck on quantum communication.

2.3.2

The Quantum Repeater

In classical optical communication, the problem of loss from fibers is solved by adding optical amplifiers. Erbium-doped fiber amplifiers, for example, have enabled fiber-optic communication to extend
across oceans. An amplifier essentially copies the information carried by the photon into many
photons. An amplifier intended to compensate for loss is called a repeater.
For quantum information, this kind of repeater is impossible, due to the no-cloning theorem,
discussed in Sec. 1.2.4. This theorem also prevents a single quantum bit from communicating any
more information than a classical bit, for if we could clone a quantum bit we could measure its wave
function to arbitrary detail. In conjunction with the epr paradox, this would allow communication
faster than the speed of light. The linearity of quantum mechanics has apparently saved the theory
of relativity.
A quantum repeater must be a very different device from its classical analog. While the
classical repeater trades entropy (and energy) for distance, providing fresh photons to replace those
that have been lost, the quantum repeater trades time (and energy) for distance.
The quantum repeater is based on the principle of quantum teleportation [42]. Quantum teleportation allows a quantum state to be transported across a classical communication channel, assuming
a generic entangled quantum state may be shared across that channel. To quickly summarize the
procedure, Alice wishes to send an unknown qubit in state |i to Bob. Alice and Bob first generate
1 The

ultimate flying qubit would be the helicity of the neutrino! A neutrino has near unity probability to fly
straight through the earth without a single interaction. While a crazy scheme to create such a flying qubit may come
to mind, it is unlikely that there is any reasonable scheme to receive it.

36

Chapter 2. Physical Resources for Quantum Computing

a generic entangled state between them, for example the singlet state |0iA |1iB |1iA |0iB . Alice
may perform a Bell-state analysis on the combination of her qubit and her half of the entangled
pair. This may be done by a single controlled-not gate followed by a two-qubit measurement in
the computational basis, or a special correlated measurement. The resulting two bits of classical
information she obtains may be sent to Bob, who may then use them to perform an appropriate
rotation on his half of the entangled pair to generate |i .

This allows the following protocol for a quantum repeater. Alice has a qubit |i in a relatively
long-lived memory which she wishes to send to Bob, but the channel between them is extremely
lossy. So she generates many generic pairs of entangled states. She may do this by generating
entangled photons, which must be caught by both her memory device and Bobs; by generating
photonic qubits as emissions from her memory device which must be caught by Bob; or by observing
photons emitted or scattered from stationary qubits with erasure of which-path information, as
I will describe in the next section. These entangled pairs carry no information about her unknown
quantum state and may be made at will. If they are lost, no information about |i is lost. She
sends these entangled states to Bob until Bob successfully receives one. This may take a long time,
depending on the probability of loss in the channel. However, once Bob has received the photon, he
then notifies Alice, and the quantum teleportation protocol may proceed. Bob could then send the
information to the next repeater, and so on.

We cannot expect this communication process to proceed perfectly. Perhaps the transmission of
the entangled pair of is imperfect, or the entanglement between them was imperfect when they were
generated. This reduces the fidelity of the transmission of |i. Fortunately, there is an analog to
quantum error correction here. Instead of generating just one faulty pair of entangled states, Alice
and Bob generate a pair of pairs of faulty entangled states. They each perform a simple controllednot gate between their two received qubits and measure the target bit. If these two entangled pairs
were perfect, this measurement would always yield the same value for both Alice and Bob. It may
readily be shown, however, that if these entangled pairs are not perfect, then in those instances when
Alice and Bob do measure the same value, the pair of qubits that they have not measured is in a
better entangled state. Therefore Alice and Bob repeatedly follow this protocol while communicating
their results over a classical channel; by keeping those pairs that had improved entanglement, and
by iterating the protocol on these improved pairs, they may eventually distill entangled states of
arbitrarily high fidelity. This procedure is called entanglement purification [43].

2.3 Quantum Memory and Quantum Repeaters

2.3.3

37

Physical Implementation

Theoretical work in quantum repeaters and quantum communication is quite developed, and I
will discuss it no further here. I instead discuss the physical tools required for this simple idea.
First, Alice and Bob are using small quantum computers. They are applying controlled-nots and
rotations to a set of quantum coherent qubits. Second, Alice needs to remember her qubit state |i,
as well as her half of the entangled pair, for enough time to complete the various acts of classical
communication required for the protocol. Likewise Bob must remember his half of the entangled
photon pairs for the duration of the entanglement purification. How long this time is depends on
the range of communication and how many entanglement purification steps are needed. Third, Alice
and Bob must be able to quickly generate entanglement over a long distance. Finally, Alice and Bob
must know when they have received their entangled states without destroying them.
One way to imagine a quantum repeater is to have a source of entangled pairs of photons. Then,
both Alice and Bob must be able to convert their photonic qubits into their quantum memory and
know when they have done so. This amounts to a quantum non-demolition measurement of single
photon states, which has been an experimental challenge in quantum optics for many years. Still,
experimental efforts to implement this type of quantum repeater have been pursued; an example is
summarized by Yablonovitch et al. [44].
A more concise implementation of these technically challenging requirements has recently been
proposed by Childress et al. [45] based on photon scattering from solid-state -type emitters. The
basic principle, illustrated in Fig. 2.3, is as follows. Alice and Bob have identical stationary qubits
(electron spins) which are themselves long-lived or which may be transferred to long-lived qubits
(nuclear spins). The qubit states are two energetically split metastable ground states, but it is
another, much higher energy state that is optically active. We presume, however, that due to a
selection rule or due to energetic separation, the photon modes which couple the state |0i to the
excited state are entirely orthogonal to the photon modes which couple the state |1i to the excited
state. Therefore, if the qubit is in state |1i, it is transparent, but if it is in state |0i it has some
probability of scattering a photon (via real or virtual absorption and reemission.) Let us presume
that, in this case, the amplitude for Alice to receive a photon in a specific mode and scatter it into
a desired output mode, given that her qubit is in state |0i, is . Bobs corresponding scattering
amplitude is . The probability of such a scattering event, ||2 for Alice, may be very low since

either the single-photon cross section or the collection efficiency will inevitably be limited.
These two distant qubits are then placed as conditional mirrors in opposite paths of a large
Mach-Zender interferometer, as shown in Fig. 2.3. The annihilation operators for the optical field

38

Chapter 2. Physical Resources for Quantum Computing

c
A
d

B
a

Figure 2.3: Schematic of an optical repeater. Optical beams a and b meet at a beam
splitter, and the output beams are sent to distant qubits A and B. Each qubit is an
identical -system, as shown in the center of the figure. The two qubit states are the
two lower states of the system; the excited state is optically accessible only to state
|0i. Many photons are lost at A and B, but some photons reflect off A or B if they are in
state |0i. These reflected photons are combined at a second beam-splitter, completing a
Mach-Zender interferometer; a detection event at port c may project the distant qubits
into an entangled state.

at the two outputs of the interferometer are then given by




1
|0ih0|A |0ih0|B a +
2


1
d=
|0ih0|A + |0ih0|B a +
2
c=



1
|0ih0|A + |0ih0|B b
2


1
|0ih0|A |0ih0|B b.
2

(2.22)
(2.23)

This is most rigorously analyzed using coherent states, as they remain coherent during the evolution
provided by the beam splitters and the attenuation at the scatterers. Suppose we input an optical
coherent state into mode a, |a i = exp(a ||2 /2) |0i . The resulting wave function for photons
at the output of the interferometer is then

|i = |00ih00|iAB
+ |01ih01|iAB
+ |10ih10|iAB




 !
+



+
d
2 c
2




E
E
1

c + d
2
2

2



 !

1

c
+ d
2
2 2

+ |11ih11|iAB |0i .

(2.24)

2.3 Quantum Memory and Quantum Repeaters

39

Suppose we detect n > 0 photons at port c. This detection projects the wave function into

|n i =

h
h0| cn |i
2
n

|00ih00|iAB ( )n e|()| /8 +
=
n
Pn
2n!Pn 2
|01ih01|iAB n e||

/8

+ |10ih10|iAB () e||

/8

(2.25)

where the probability of detecting n photons is


Pn =

i
||2n h
2n |()|2 /4
2
2n ||2 /4
2
2n ||2 /4
2
|

|
e
|h00|i|
+
||
e
|h01|i|
+
||
e
|h10|i|
.
2n!4n
(2.26)

Note that if both qubits are in state |0i and if = , this probability vanishes due to destructive
interference, meaning that if a photon is observed one of the qubits must have been in state |1i, but
since = there is no way to know which one. Of course the probability of measuring a photon in
output arm c is very low if and are very low; however, this limits the speed of the repeater and
not its ultimate fidelity.
Like most quantum information processing protocols, the entanglement procedure begins with
Hadamard gates on all qubits. Alice and Bob therefore each prepare their qubit in the superposition
state
|iA = |iB =

|0i + |1i

.
2

(2.27)

Then we find that our final projected state is


2

|n i =

n e|| /8 |01i + ()n e|| /8 |10i + ( )n e|()| /8 |00i


p
.
||2n e||2 /4 + ||2n e||2 /4 + | |2n e|()|2 /4

(2.28)

This is clearly an entangled state. A measure of its entanglement can be found by tracing over
Bobs qubit and finding the entropy of Alices resulting qubit. The degree of entanglement is then
a smooth function of x = /. The ideal case is = , meaning the scattering amplitudes for
both Alices and Bobs qubits are identical. In this case the resulting state is an exact Bell state,
immediately suitable for quantum teleportation. If and are different, a situation difficult to
avoid in reality, then the fidelity between |n i and the singlet state |01i |10i is
2

|1 (x)n es(1|x| ) |
.
Fn = p
2 1 + |x|2n e2s(1|x|2 ) + |1 x|2n e2s|1x|2

(2.29)

where s = ||2 /8. For odd n, this fidelity is greater than 1/2 for any x > 0, especially when the
amplitude of the coherent state is very large. Therefore, in principle, entanglement purification is

40

Chapter 2. Physical Resources for Quantum Computing

always possible.
This result has assumed that the detector is able to fully resolve how many photons it has counted.
No physical photon detector can do this perfectly, and many cannot do it all. Even if = , this
creates a problem, since |n i generates a different Bell state for even or odd n. Suppose an avalanche
photodetector is used which only resolves whether one or more photons arrived, without distinction
as to how many arrived. The probability of detecting an event is now
P =

Pn =

n=1

i
2
2
2
1h
3 e|()| /4 e|| /4 e|| /4 .
8

(2.30)

The resulting qubit state becomes the mixed state


=

and the fidelity becomes F =


as

pP

X
Pn
|n ihn | ,
P
n=1

2
n (Pn /P )Fn .

F =

(2.31)

In the case = , this may be explicitly evaluated

sinh(2s) s
e ,
2 sinh(s)

(2.32)

which is still larger than 1/2 for any s > 0. If and are too different and is very large, however,
the resulting mixed state may not even be entangled. One can ensure entanglement in this case even
when 6= by using single photon states as an input to the interferometer, but then the probability
of measuring the photon that projects into the desired entangled state scales as ||2 , in contrast to

the potentially enhanced ||2 for a high-amplitude coherent state as given by Eq. (2.30).

The above analysis has, of course, ignored various sources of error. The homogeneous broadening
of the scatterers causes an error akin to 6= . In general, the presence of any which-path
information indicating which qubit scattered the photon will reduce the fidelity of the resulting
entangled state. A more detailed discussion of this particular proposal with error estimates is
provided by Childress et al. [45]. Important additional considerations include balancing the long
interferometer and accounting for the probability of detector dark counts and multiple spontaneous
emission events from the scatterers.
The conclusion here is that the key resource needed for this kind of quantum communication is
stationary qubits with nearly identical (and hopefully not too low) photon scattering cross sections.
Various optical emitters are under consideration for this, such as semiconductor quantum dots in
optical cavities. Another requirement, however, is that these qubit systems must be able to remember
quantum states for the potentially long time it takes to successfully complete the repeater protocol

2.4 Quantum Computing Technology

41

enough times for entanglement purification. As I will discuss in the following chapter, it may be
that among solid-state implementations, only nuclear spins have sufficiently long coherence times
for this.

2.4

Quantum Computing Technology

Any attempt today to comprehensively survey the history and status of the global effort to physically
implement quantum computers would be futile, as it would become out-of-date during editing and
printing. Since the breakthrough theoretical results of the mid-1990s, the interest in quantum
computing has been large and new proposals and experiments are constantly published. The present
discussion attempts to be neither historical nor comprehensive. Rather, I will attempt to classify
attempts to build quantum computers according to the physical principles involved, with references
to only some of the relevant proposals and experiments.
I will first divide the many proposals for quantum computing into those using solid-state devices
and those that dont. This division is more related to the engineering and expertise required for
experiment than the physical principles involved in the imagined computer. The preference toward
solid-state devices is prejudiced by the success of solid-state classical computers, and this prejudice
may be justified by the technological infrastructure currently existing for solid-state device processing. However, single isolated atoms and molecules outside of solids have always led the way to
experiments showing quantum coherence, and so we discuss these first.

2.4.1

Atomic and Molecular Optics Implementations

The first image that should come to mind when thinking of quantum dynamics is the atom. The
isolated atom and its optical spectrum are the genesis and the testing ground of quantum mechanics.
There are several proposals for quantum computing based on isolated atoms or molecules, differing
primarily by how those atoms or molecules are spatially isolated.

Trapped Ions
Historically first [46] and perhaps most promising are quantum computers based on electromagnetically trapped laser-cooled ions. The long coherence times of optical states of ions make them
oscillators with record high Q-values, leading to their promise as time standards [47]. For quantum
computing, the qubit may be hyperfine states of a single ion manipulated with stimulated Raman

42

Chapter 2. Physical Resources for Quantum Computing

transitions [24, 4850], or states directly connected by an optical transition [25, 5153]. Both systems
have been successfully used for two-qubit quantum algorithms and error correction techniques.
Initialization and readout with laser cooling and spectroscopy are now standard practice, at least
for relatively small chains of ions. Laser cooling can typically allow polarizations larger than p = 0.9;
algorithmic techniques as discussed in Sec. 2.2.2 work very well in this regime.
Cirac and Zoller [54] showed one way in which the collective motional modes of chains of such ions
can be used for universal logic; several variations on this idea have also been devised. A summary
may be found in Ref. 55. A key advantage to this kind of logic is that it allows any two ions in the
trap to be coupled, since the gate is mediated by a non-local phonon mode. This greatly improves
the prospects for fault-tolerant operations. However, putting arbitrarily large numbers of cold ions
in a single trap is not possible; the vibrational states become too densely spaced, complicating
both initialization and logic. Large quantum computers may still be feasible by moving ions in
and out of traps with electrically controlled conveyer belts, as proposed by Kielpinski et al. [56], or
sharing entanglement between distant traps using linear optics with projectively measured photons
generated by or scattered from the ions [57]. These latter schemes follow the same principle as that
discussed in Sec. 2.3.3, and initial experimental measurements of entanglement between a stationary
trapped ion and a flying photon have been reported by Blinov et al. [58].
Decoherence may still be limited by electrical noise in the rf trap, especially once complicated
architectures are built for moving ions around a lithographically fabricated electrical chip. It has
been found in several experiments that the primary decoherence channels are global electromagnetic
noise, since ions with coherence times on the order of milliseconds might have increased coherence
times of seconds when encoded into decoherence-free subspaces [24, 25]. The qubit in this case may
be the spin of the ions nucleus, but this is strongly coupled at all times to the electrons charge,
which itself is always weakly coupled to the classical world of the rf trap. Greater coherence could
therefore be possible with neutral atoms.

Neutral Atoms
Several early proposals considered the use of neutral atoms coupled via strong interactions with
single photons in a high-Q cavity [5962]. The atoms could pass through this cavity one by one.
The scalability of such schemes is limited, but the ideas raised in these proposals have been extremely
valuable.
A more scalable architecture may be considered by trapping neutral atoms in optical standing
waves, forming an optical lattice [6365]. Motion-controlled dipole-dipole interactions of the atoms

2.4 Quantum Computing Technology

43

could then provide quantum logic. Such an architecture is at present less experimentally advanced
than electromagnetically trapped ions. There may even be some advantage to using laser-trapped
polar molecules in a similar architecture [66].
Liquid NMR
Nuclear magnetic resonance of molecules in liquid solution solves the problem of isolating qubits
resonantly. In such schemes [36, 37], the qubits are nuclei in a molecule that have different resonances
due to local magnetic field differences induced by the chemical structure of the molecule. Scaling
this architecture to hundreds of qubits may at first seem to be limited by the ability to generate
a sufficiently complicated molecule with differentiable resonant frequencies; however this scheme is
actually limited by the need to operate at high temperature and insufficient ability to polarize the
nuclear qubits, points which will be discussed in Sec. 3.2. Nonetheless, liquid-state nmr quantum
computers remain the most technologically advanced; to date they represent the only system to
have experimentally demonstrated enough qubits (7) and enough coherence ( 1 sec) to perform a
version of Shors algorithm [67].

2.4.2

Solid-state Implementations

The reality of trapped ion, atom, and molecule experiments at the present time is a small number
of qubits. Many liken such architectures to the early days of classical digital computers, when the
hardware components such as vacuum tubes were inserted by hand in a gigantic room; the physical
size of the room might limit the size of ones classical computer. The exponential increase in classical
computing power was made possible by the solid-state transistor and semiconductor microfabrication
techniques for assembling millions of such transistors in parallel. In the case of ions and atoms, the
amount of hardware required to simply hold these qubits in place is substantial. It is hoped that
solid-state implementations of quantum computers might be able to take advantage of similar largescale integration techniques that have enabled powerful classical computers. The combination of
these techniques and the requirements of quantum coherence continues to be a daunting challenge
to physicists and engineers. Many different approaches to solid-state quantum computers have been
taken.
Optically Driven Quantum Dots
An immediate solid-state analog to trapped ion or atom schemes are schemes using optically driven
quantum dots, which have been proposed for quantum computation in a large variety of ways [6872].

44

Chapter 2. Physical Resources for Quantum Computing

A quantum dot is a solid-state analog to an atom, showing similar discrete energy levels resolved by
optical spectroscopy. Proposals generally employ charged dots, in closer analogy to an ion. They
may be cooled to their ground state by physical cooling of their substrate and they may be measured
optically. The difficulties in such proposals, however, include the short coherence times of quantum
dot states, the difficulty of reliably fabricating suitable structures, and the challenges of controlling
the interactions between quantum dots.
I will explain one scheme [68] in detail here, as its advantages will relate to further architectural
considerations in Sec. 3.5. Previous schemes involving quantum dots as qubits used local, shortrange interactions to solve the difficult problem of qubit-qubit coupling. These interactions can be
slow and difficult to control, and since they only couple neighboring qubits, bit-swapping is needed.
Interactions mediated by a common mode coupling to all qubits, such as the phonon bus in trapped
ions, are more promising. In neutral cavity qed schemes, the atoms are coupled via a common
cavity field. This proposal combines the solid-state scalability of semiconductor quantum dots with
the long-range, cavity-field-mediated interactions of cavity qed.
The two-state system which provides the qubit in this scheme is the spin of a single electron,
trapped in the conduction band of a quantum dot embedded in a solid-state optical cavity. The dot
is doped so that the valence band is full and only this single electron exists in the conduction band.
The dot is sufficiently small to ensure that this single electron is trapped in the conduction-band
ground state at low temperature. The motivation for using the spin of the electron is the generally
accepted fact that the decoherence times for spins in semiconductors are much longer than other
relevant time scales, such as the electron-hole recombination time or the decay rate of the first
conduction-band excited state.
The calculation and measurement of spin decoherence times in quantum dots of various kinds
is an ongoing field of study. Numerous measurements have shown T1 times of spins in several
kinds of quantum dots in the 10 ms range, but decoherence measurements are more challenging.
Measurements of the exciton spin dephasing time in ensembles of chemically synthesized CdSe
dots [73] and self-assembled InAs dots [74] have yielded maximum values of several nanoseconds,
but these experiments were limited by inhomogeneous effects, primarily differences in the g factors of
dots due to their varying sizes. It is thought that the T2 time constant for electron spins in quantum
dots may be as high as a microsecond in the absence of these inhomogeneous effects. Experiments
measuring exciton spin lifetimes in single quantum dots have shown only that the spin lifetimes
are substantially longer than the very short exciton lifetime [75]. Hanle effect measurements on
singly charged quantum dots formed by GaAs/AlGaAs quantum well interface fluctuations measure

2.4 Quantum Computing Technology

45

single electron dephasing times of 16 ns, probably limited by spectral diffusion from nuclear spin
fluctuations during the long measurement [76]. This timescale would be extended if the nuclear
spins were fully polarized, or in a quantum dot architecture in which the nuclear spins are absent,
which is possible with group-iv or group-ii-vi semiconductors.
The two states of the electron spin in these quantum dots are separated in energy primarily by
applying a small magnetic field. It is convenient to use the Voigt geometry, in which the magnetic
field is applied parallel to the plane of the substrate in which the quantum dot is embedded. We
label this magnetic field axis as the x-direction, so the eigenstates for the spin of the conduction
electron are |mx = 1/2i and |mx = 1/2i, which I notate simply as |i and |i, respectively. When
an exciton is optically excited, the trapped conduction electron and the electron part of the exciton
form a spinless singlet state, and the total angular momentum of this trion is therefore given by the
angular momentum of the valence hole. For the valence states, however, the dominant spin-splitting
term is the spin-orbit coupling, which leaves as eigenstates the heavy-hole states |mz = 3/2i and
the light-hole states |mz = 1/2i. These states are only slightly perturbed by the magnetic field
in the x-direction if the field is sufficiently small and the z-confinement is sufficiently tight. The
numbers mz in these states are the z-projections of the total angular momentum of the valence
holes, and hence they represent linear combinations of the x-basis spin-up hole and spin-down hole
states. Both heavy-hole states have optical matrix elements with both conduction band states. We
assume the field is sufficiently small that we can ignore mixing between the heavy-hole and light-hole
states. The only transitions used in this scheme are between the |i and |i states and the nearly
degenerate heavy-hole upper-valence band states | 3/2i, which provide the upper excited energy
level in a -type system, as employed in the optical repeater scheme in Sec. 2.3.3 and illustrated
in Fig. 2.3. Fabricating a quantum dot in which all the optical transitions in this diagram are
well-resolved at a sufficiently low field to avoid heavy-hole/light-hole mixing may be a challenge,
although some quantum dots have nearly lifetime-limited linewidths and very large splittings. The
mixing of light-hole states complicates but does not rule out the present scheme. A theoretical and
experimental study of these splittings for self-assembled InAs quantum dots in varying magnetic
field was conducted by Bayer et al.. [77].
Crucial to this scheme is the ability for stimulated Raman adiabatic passage (stirap) between
the two spin states of the conduction electron for arbitrary single qubit rotations. For the stimulated
Raman transitions, two optical fields are applied. The first is resonant with the transition between
the spin-down conduction-band state (|i) and the upper valence-band states (| 3/2i). The second
is resonant with the transition between the spin-up conduction-band state (|i) and the upper

46

Chapter 2. Physical Resources for Quantum Computing

valence-band states (| 3/2i). Loosely speaking, one of these fields induces a transition in which
the hole in the conduction band is transferred to the valence band, and the second of the fields
returns the hole to the conduction band, except in the other spin state. (In fact, both fields are
detuned somewhat from the conduction-band/valence-band energy difference in order to ensure that
the valence hole is only virtually excitedthat is, to minimize the probability that a hole will be
excited in the valence band and remain there). The net result is that the hole and the electron in
the conduction band state have been coherently transferred from one spin state to the other, and
indeed the duration of the pulses may be chosen so as to achieve any superposition of the |i and |i
states. This may be done without initial knowledge of the state of the qubit, but still maintaining
an adiabatic condition in which loss from the cavity and relaxation of the upper valence-band states
are severely reduced. Detailed schemes have been developed by Chen et al. [78].
Imamoglu et al. [68] propose that each quantum dot has its own dedicated laser from a fiber tip,
and that each laser is tuned to its dot, compensating for any differences between dot parameters. This
may not be needed for small cavities, since large variation between dots may allow every resonance
to be spectrally isolated. These single qubit rotations may be carried out in parallel, and they can
be very fast ( 10 ps) in comparison to the expected spin-decoherence time. It is this single dot
addressability that limits the scalability for one of these devices: the resolution afforded by tapered
fiber-tips in the near field limits resolution to about 1000
A, and too many spectrally addressed dots
would inevitably overlap. However, it is exactly the quantum communication proposal in Sec. 2.3.3
that can allow multiple such devices to form a large quantum network.
The same principle of Raman transitions is used in order to achieve qubit-qubit coupling within
a single cavity. For this, one of the optical fields of the stimulated Raman transition is a cavity
field seen by all qubits. Imamoglu et al. [68] propose to accomplish such a field by embedding the
quantum dots in a microdisk cavity with very high Q. The cavity loss rate must be much less than
the frequency splitting between the |i and |i states. With such a cavity in place, each qubit sees
two fields, one provided by its laser and the other from the cavity. Initially, suppose that these two
fields are detuned from the frequencies needed to achieve the Raman transition described above.
That is, suppose their difference-frequency corresponds to the |i/|i energy splitting plus a small
detuning which may be varied from dot to dot. When these detunings are made the same for two
dots, then these two dots are coupled.
The physical picture for this coupling may be simplified as follows. When the dots are decoupled,
the detuning is sufficiently large to prevent any transition. The combined energies of the cavity
photon and the laser photon are insufficient to both stimulate a downward transition to a valence

2.4 Quantum Computing Technology

47

state and an upward transition to the other spin state. When the dots are coupled, the detunings
of two dots are made equal, so that the missing energy at one dot is compensated for by the next.
For example, although a dot in the lower-energy state |i does not have enough energy in its local
laser and cavity fields to make a Raman transition to |i, another dot elsewhere in the cavity may
make a simultaneous transition from |i to |i, releasing extra energy into the cavity field which is
exactly tuned to lift the first dot to |i.
This type of indirect interaction stems from second-order perturbation theory; I will show methods for similar calculations in Chapter 3.3. Interactions of this type have been included in a variety
of quantum computation schemes. For example, another proposal for long-range interactions between semiconductor quantum dots is similarly mediated by optically excited free excitons in the
semiconductor substrate [79].
The final two elements for quantum computation are initialization and measurement. Optical
pumping schemes may allow initialization. Imamoglu et al. [68] propose single spin measurement by
the presence or absence of a single Raman-scattered cavity photon. That is, a laser field is applied
whose frequency is chosen for exact two-photon resonance with the cavity mode on the |i to |i
Raman transition. But the cavity is empty, so that if the electron is in state |i, no transition occurs.
If the electron is in |i, a cavity photon may be created and a laser photon annihilated in order to
complete the transition. This single cavity photon will eventually leak out of the cavity. However,
the feasibility of this scheme in view of inevitably imperfect collection efficiency is questionable.
The most important piece of missing technology in this scheme is sufficiently high-Q cavities. In
fact, although the spin decoherence time may be as high as 1 s, the lowest cavity loss rate available
in the microdisk structures discussed in Ref. 68 is on the order of 10 ps. Of course, the cavity
photons which are used during quantum computation are only virtually excited, and so it may be
estimated that an effective decoherence time caused by cavity loss is on the order of 1 ns. Although
the Q is not crucial for measurement, it is important that the primary source of loss is transmission
to the photodetector, so that the lost photon may be counted.
This proposal is one of many ideas for using optically-driven semiconductor quantum dots for
quantum computation. Semiconductor quantum dots are convenient because they are optically
bright and their characteristics can be somewhat manipulated by growth conditions. However,
many of the ideas of these schemes could also be implemented in optically active impurities in other
host materials. For example, Shahriar et al. [80] have assembled a proposal for using spectral-holeburning techniques with a heavily inhomogeneously broadened ensemble of optical impurities in
an optical cavity. The favored impurity is the nitrogen-vacancy center in diamond, whose coher-

48

Chapter 2. Physical Resources for Quantum Computing

ence characteristics, optical properties, and interactions with substrate nuclei have already led to a
number of ground-breaking single-emitter experiments, including demonstrations of single photon
generation [81], single nuclear spin detection, and a logic gate between the electron spin and a single
13

C nucleus [82]. This impurity is a top candidate for the kind of quantum network scheme discussed

in Sec. 2.3.3.

Electrically Driven Quantum Dots


In closer analogy to classical computers, quantum dots may be driven by electromagnetic gates.
A key proposal of this type was made by Loss and DiVincenzo [83], although a number of other
proposals for electrically driven quantum dots have appeared. These quantum dots are defined by
electrostatic gates which deplete regions of a 2-dimensional electron gas (2deg) and define a potential
to trap one and only one electron using Coulomb blockade. The direct control over the electron wave
functions allows controlled electron spin couplings by increasing the wave function overlap for nearby
dots. When wave functions strongly overlap, a fast Heisenberg exchange interaction is switched on;
when they are kept separate, the electrons are well decoupled. Initialization may occur through
either thermalization at a large magnetic field or via various implementations of spin-filter devices,
which could also be employed for measurement. The general principle for the most promising such
devices is to convert spin to charge by taking advantage of the Pauli exclusion principle, which
will only allow two electrons in the spin-singlet state to occupy a single quantum dot ground state.
Measuring single electron charges may then be performed with single electron transistors (sets).
The existing technology for mesoscopic physics with 2deg devices and the expected speed of the
switchable exchange interaction make such schemes quite promising. The substantial theoretical
and experimental effort that has gone into them has been recently reviewed in Ref. 84.
A related idea was proposed by Platzman and Dykman [85]. Instead of trapping an electron in
a quantum dot, a less noisy solution may be to trap an electron on the surface of liquid helium,
which provides a relatively quiet environment. Lateral motion on the surface is then prevented by
electromagnetic gates. This scheme may have improved coherence properties over more traditional
quantum dot architectures.

Electron Impurity States in Semiconductors


Similar to a quantum dot, a single impurity in an appropriate substrate may act as a coherent
atomic system. The Si/Ge-based proposal of Vrijen et al. [86] is an example. In spirit the proposal is
similar to that of Loss and DiVincenzo [83]. However, the element of band-structure engineering may

2.4 Quantum Computing Technology

49

improve control. The possibility of coupling these spin qubits to photons for quantum communication
is also more promising.
There are many challenges to wave-function engineering, however. An important question is the
spin decoherence time in these structures. An encouraging measurement showed spin decoherence
times approaching 60 ms at T = 4 K for phosphorous-bound-electron spins in isotopically purified
silicon [87]. It is unclear whether such long decoherence times will be preserved upon introduction
of the semiconductor alloying and electrical gating proposed by Vrijen et al. [86]. Recent spin echo
measurements of free electrons in Si/SiGe quantum wells have revealed decoherence times of 3 s
[88]; this timescale is limited by spin-orbit coupling effects and will certainly lengthen if the electron
is laterally confined. Still, it is worthwhile to consider an architecture that uses the electron spin of
these experimentally characterized phosphorous impurities directly, without the addition of alloying
or gates. Such a proposal was recently made by Sousa et al. [89] with a very similar spirit and
construction to the all-silicon nuclear spin proposal I discuss in Sec. 3.4.

Superconducting Josephson-Junction Circuits


If you tried to make a quantum computer using classical electronics, you would find that the resistance of normal metals would constantly leak the quantum information into heat, causing rapid
decoherence. This problem may be alleviated using superconducting circuits. Defining qubits with
such circuits is best done using the nonlinearities afforded by the Josephson effect.
The rudimentary physics of superconducting qubits is most easily explained by analogy to mechanical systems, although the reader is warned of a dangerous lack of rigor in this discussion. In
quantum electrodynamics, the electric and magnetic fields are conjugate variables. In particular,
if a current I0 proceeds around a loop, the magnetic flux through the loop and the electric flux
parallel to the loop Q have the commutator
[, Q] = i~,

(2.33)

and therefore and Q are analogous to the position and momentum of a particle. When a thin
insulating layer separates sections of the superconductor, tunnelling may still occur, but a phase occurs between the superconducting wave functions on either side of the barrier. This is the Josephson
effect, and it creates an effective potential term


V () = EJ 1 cos

2
0



(2.34)

50

Chapter 2. Physical Resources for Quantum Computing

in terms of the Josephson energy EJ = I0 0 /2 and the flux quantum 0 = h/2e. The Hamiltonian
is further modified by capacitance C, which introduces a term Q2 /2C analagous to the kinetic
energy, and an inductance L, which introduces a harmonic oscillator-like potential term 2 /2L.
A flux qubit or persistent current qubit [90] is biased to create a double-minimum potential from
the combined parabolic inductance and oscillatory Josephson terms. The two minima correspond
to the two qubit states of the circuit. Roughly speaking, these two minima correspond to current
going in one direction of the loop and current going in the other. These flux qubits are controlled
by applying flux administered from other superconducting elements in the circuit, which alter this
potential to allow the qubit to move from one state to another. The flux from one qubit could then
affect the dynamics of the next, allowing entanglement. Qubit measurement is performed by directly
measuring the flux with a superconducting quantum interference device (squid). Spin echoes in such
a device operating at 5.71 GHz were detected, revealing a decoherence time of 20 ns [91]. A less
direct measurement yielded a coherence time of 2.5 s in a 0.9 GHz device [92]. An inviting idea is
to couple the magnetic flux due to a nuclear spin qubit to such a system, as the use of squids for
nuclear detection is well established.
Another kind of superconducting qubit is the phase qubit. Unlike the flux qubits potential with
two minima, the phase qubits potential is that of a single anharmonic potential; the lowest two
states of this potential provide the qubit. These qubits may be capacitively coupled for quantum
logic [93]. An early experiment on a 16 GHz phase qubit showed Rabi oscillations decaying on a
5 s timescale [94], although the meaning of this number is subject to debate. Most experiments
show coherence times in the 20 ns range [95].
The charge qubit [96] relies ultimately on the quantization of charge into individual Cooper pairs,
which becomes a dominant effect when a sufficiently small box is defined by multiple Josephson
junctions. For these qubits the Josephson energy energy EJ is on the same order as the capacitance,
so the periodicity of potential of Eq. (2.34) leads to Bloch-like states in . It is clearer in this
case to work in the analogy of the reciprocal lattice, where instead of k-states we have quantized
charge states corresponding to different integer numbers of Cooper pairs in the box; two such states
provide the qubit. Early implementations of this qubit [96] showed distressingly short coherence
times of only a couple nanoseconds; appropriately biased variations qubits operating at 16.5 GHz
later showed coherence times measured via spin-echo in the vicinity of 0.5 s [97]. These qubits have
been capacitively coupled to form a simple quantum logic gate [98] and have recently been coupled
to a microwave transmission line resonator for cavity-qed-type experiments [99]. This may be a first
experimental step to coupling these fast qubits to longer-lived qubits such as trapped ions or even

2.4 Quantum Computing Technology

51

nuclei; a relevant proposal was considered by Tian et al. [100].


The differences between the various kinds of superconducting qubits are largely practical. For
example, there is a large difference in scale; the phase qubit may be up to 100 m2 in area, the
flux qubit only 1 m2 , and the charge qubit 0.01 m2 . This varying range of scale is accompanied
by the varying effective impedances of the circuits; the phase qubit may be close to 50 , while
the flux qubit has k impedance and the charge qubit has nearly M impedance. These are very
relevant numbers when working with microwave frequencies; it can greatly change the difficulty of
engineering the circuit. In general, the phase qubit is the easiest to engineer but shows the poorest
performance, while the charge qubit is the most difficult to handle but with the best results so far.
Of course, this simple viewpoint certainly oversimplifies the situation.
The key advantage to superconducting implementations is that they are very fast, easy to initialize (thermal initialization is sufficient at the 10 mK temperatures at which these devices best
perform), and easy to measure individually. However, it is not clear whether their coherence properties are sufficient, especially when they are coupled in large networks.

Solid-State NMR
The final solid-state implementation we consider lives at the opposite extreme from the superconducting implementations. Nuclear spins are very slow, extremely difficult to initialize (thermal
initialization will never be sufficient), and notoriously difficult to measure individually. However,
the need for long-lived quantum coherence may justify these difficulties.
The largest difference between the many proposals for solid-state nmr quantum computers is the
coupling between the nuclei. We take a chronological approach to describe the various schemes for
nuclear-nuclear couplings that have been considered.
The possibility of employing the long coherence times of solid-state nuclei for quantum computation dates at least to 1995 when a very speculative paper by DiVincenzo [29] pointed out the
similarity of double magnetic resonance experiments to quantum logic and the possibility of using
magnetic resonance force microscopy [101, 102] to probe individual electron spins. There was no
indication for a method to selectively couple nuclear spins in this early paper. Then, several proposals appeared in 1998, shortly after the proposals for liquid-state NMR2 . The first solid-state
proposal came from Privman et al. [103] which considered using indirect nuclear-nuclear couplings
mediated by excitons in a 2deg in the quantum Hall regime. While this proposal mentioned many
ideas that would be employed by later proposals, it did not develop a complete architecture, and
2 This

year was also the year of my arrival to Stanford University.

52

Chapter 2. Physical Resources for Quantum Computing

remained highly speculative. In particular, the means of controlling interactions and distinguishing,
initializing, and measuring nuclei in this proposal were only unsubstantiated guesses. The idea of
using nuclei in quantum hall effect semiconductors remains an interesting possibility, however, and
several subsequent papers have further developed the ideas [104106].
Much more attention was paid to a more concrete proposal published by B. Kane [107]. In particular, this proposal was especially influential because its silicon-based architecture was reminiscent
of classical silicon-based electronics. The key component to this scheme is taking advantage of the
large hyperfine coupling between spin-1/2

31

P nuclei in silicon and donor electrons bound to those

nuclei. Given sufficient local control of the electron and nuclear spin, polarizing and measuring
the individual nuclei could be achieved by polarizing and measuring single electron spins, which
is possible by spin-to-charge conversion and single charge detection with a set. This scheme also
employs indirect nuclear interactions mediated by pairwise exchange-coupled donor electron spins,
which I will discuss in Sec. 3.3.2. Like the proposal of Loss and DiVincenzo [83], this exchange
coupling between donors is switched on and off by manipulating the electron wave function overlap
via electrostatic potentials generated by lithographically fabricated metallic gates. Such indirect
interactions become weaker at higher magnetic fields; Kane [107] proposes operation at fields of 2 T,
at which temperatures of order 100 mK are required to polarize the electron spins.
Kane [107] indicated that the ultimate coherence time for these

31

P qubits is uncertain, since it

is likely that charging noise generated by the electronic environment created by the control circuitry
would lead to nuclear decoherence of uncertain scale. Theoretical calculations are usually insufficient
for estimating such noise sources, as they depend crucially on the quality of device fabrication. The
fabrication of such a device itself also poses great challenges. The

31

P nuclei would need to be

placed with high precision at a spacing of approximately 10 nm; subsequent work showed that
effects resulting from interference among the six conduction band valley minima in silicon would
require the

31

P nuclei to be placed with atomic precision [108], although the presence of strain

may help alleviate the interference effects [109]. The metallic gates for controlling the electron
wave function would need to be on an even smaller scale than the nuclear spacing and with similar
precision of placement, a difficult task for modern electron-beam lithography. However, progress
toward fabrication [110] and measurement [111] in this scheme has been made in recent years. The
development of a device such as this is at the cutting edge of nanotechnology, and we can expect
further development of these technologies in the years to come.
Shortly after Kanes proposal, the early successful experiments in solution nmr led several researchers to wonder whether an ensemble-based approach to solid-state nmr might be more practical.

2.4 Quantum Computing Technology

53

Proposals were developed that were more closely connected to the small quantum computers afforded
by liquid-state nmr, using molecular crystals with analagous ideas for distinguishing, coupling, and
measuring nuclei [112]. Experimental demonstration of quantum information processing in such a
molecular single-crystal solid was recently reported by Leskowits et al. [113].
These proposals still offered no means for scaling above approximately 30 qubits. A proposal published by Yamaguchi and Yamamoto [114] suggested new technologies to attempt to bring ensemblenmr technologies into a scalable regime. I will discuss aspects of this proposal and later proposals
that developed from it in Chapter 3.

2.4.3

Linear Optics

Finally, an important development in the implementation of quantum computers has been the implementation of logic gates via postselection from projective measurements. An example motivating
this possibility is the quantum repeater implementation discussed in Sec. 2.3.3. This protocol entangles two distant qubits after measurement of a photon in one output arm of the interferometer.
When exactly did the nonlinear interaction affording this entanglement occur? Were they entangled before measuring a photon in port c? Suppose we had measured a photon from port d of the
interferometer in the ideal = limit: in this case the resulting wave function for the two qubits
is entirely separable. This might lead us to believe that the two qubits were not entangled before
the measurement. In fact, the lack of local realism in quantum mechanics forbids an answer to the
question of prior entanglement. Before the measurement, the spins are in a coherent superposition
of entangled and unentangled states; the reality of their entanglement is not determined until after
the photon measurement is made.
A similar idea can be employed to entangle dual-rail implementations of photons [115]. A dualrail implementation is simply a pair of optical modes, such as two polarization states or the modes
of two separate single-mode fibers. An appropriate Fock-state encoding, choice of beamsplitters,
and projective measurement can allow an entangling gate between photons with a finite probability
of success dependent on the number of ancillary photonic qubits used. This is a remarkable result,
since many years of attempts to implement reversible computation using nonlinear optics met little
success due to the low optical nonlinearities available in nature. The proposal of Knill, Laflamme,
and Milburn (klm) [115] was the article in the journal Nature with the most citations in 2002, and its
ideas represent a key direction of quantum computation research today. The technical challenges for
implementation are still considerable, as initialization requires the generation of specific qubit Fock
state superpositions and measurement requires high-efficiency number-resolving photon counters.

54

Chapter 2. Physical Resources for Quantum Computing

The probabilistic nature of the post-selected quantum gates is not a large problem when combined with ideas from Gottesman and Chuang [116] to allow universal quantum computation with
generalized teleportation protocols and single qubit gates. Essentially, the desired entangled states
may be made prior to encoding valuable data in the quantum information processor. Attempts to
make the entangled states may be repeated until there is success; these may be used when desired
in combination with Bell-state measurement and single qubit operations (which are especially easy
in linear optics) to perform deterministic quantum information processing.
However, fault-tolerant architectures for this idea require one important element: the ability to
remember general entangled states and generate them on demand. In principle, quantum memory
could also be given an all-linear-optics implementation by creating a long fiber loop with periodic
quantum error correction achieved with high-probability nonlinear klm gates, although the scalability of such a scheme is questionable. In particular, it would be better to have a high-fidelity memory
based on a long-lived stationary qubit, and for this the nuclear spin is the obvious choice.
The employment of nuclear spins as a quantum memory in an optical network provides one
example of interfacing different quantum information technologies. As each technology has its own
advantages and disadvantages, interfacing them will likely be an important step in the development
of large quantum computers. In the final harvest for quantum technologies, I believe many of these
systems will have a role, and it is likely that nuclear memory in particular will be important. For
the remainder of this dissertation, therefore, I focus on nuclei and the means to process quantum
information with them.

Chapter 3

NMR Quantum Computers


Some of these other things, quantum dots and nanotechnology and that kind of thing
I will admit to being a skeptic around those things replacing mainstream digital silicon.
You can clearly make a tiny little transistor by these techniques with potentially great high
frequency, but can you connect a billion of them together? Thats really the problem.
Gordon Moore
The use of spin-1/2 nuclear spins as qubits poses a distinct advantage over other qubit systems.
These two-state systems couple only to magnetic fields. The coupling of these nuclei to photons is
negligible; the matrix element directly coupling a nuclear dipole to an rf photon is of order
40 2 ~
1020 sec1 .
3 3

(3.1)

Spontaneous decay in vacuum would only occur on astronomical timescales. The direct coupling
of nuclei to lattice or molecular vibrations is also non-existent, although such vibrations can play
indirect but critical roles in the thermal relaxation of nuclei. Nuclear spins do couple to the magnetic fields of nearby electrons, although electrons paired in closed valence shells exhibit very little
magnetic noise. All of this isolation leads to the possibility of nuclei maintaining quantum coherence
for an extremely long time. As a few dramatic examples of long coherence times observed for nuclei,
consider the recent results in Table 3.1. These results are limited not by fundamental relaxation processes, but rather by imperfections in the experimental arrangement, so that improved engineering
could in principle extend T2 to even longer times in each case.
Such long coherence times do make nuclei seem very promising for building a quantum computer.

56

Chapter 3. NMR Quantum Computers

Nucleus and its medium


29
Si nuclei in solid single-crystal silicon
9
Be nucleus in a Penning ion trap
129
Xe nuclei in liquid xenon
3
He nuclei in optically polarized gas

25
600
1300
2000

T2
sec
sec
sec
sec

0
60 MHz
303 MHz
37.5 kHz
10 kHz

Reference
Ladd et al. [117]
Bollinger et al. [118]
Romalis and Ledbetter [119]
Chupp et al. [120]

Table 3.1: Extreme coherence times measured for nuclear spins.

However, they come at great cost. Nuclei are so well isolated from their environment that they are
notoriously difficult to measure. The signal-to-noise ratio (snr) is typically the limiting factor in
most nmr-based applications, such as multi-dimensional Fourier transform nmr spectroscopy for
analytical chemistry and medical resonance imaging (mri). Typically, between 1014 and 1018 nuclei
form an ensemble for measurement in these applications, and improvement of the snr by only one or
two orders of magnitude is considered a major engineering advance. In most models of fault-tolerant
quantum computation, arbitrary single qubits must be available for rapid, projective measurement
at any time. In this sense, there are approximately 15 orders of magnitude between practical nmr
and quantum computation.
A number of technologies may be introduced to bridge the gap. First, a single qubit need not
be a single nucleus; an ensemble of nuclei may act as one qubit if these nuclei are in identical
environments and uncoupled to one another. Second, measurement techniques different from those
used in routine nmr spectroscopy or mri may be employed. Single nuclear spin measurement is
performed, for example, in ion-trap experiments, where the fluorescence between one hyperfine state
and a shelved state rapidly indicates the state of the ion with high fidelity, including the nuclear
spin; this identification is the reason an ion trap experiment appears in Table 3.1. Measurement
of small ensembles or even single nuclei in the solid-state may be possible with optical techniques
under special circumstances. A very recent example is the measurement of a single 13 C nuclear spin
near a charged nitrogen-vacancy center (NV ) in a synthetic diamond crystal via optically detected
electron-spin magnetic resonance [82]. Another important route to sensitive spin measurement is
magnetic resonance force microscopy, in which the mechanical force between nuclear spins and a
magnetic field gradient is detected as a deflection of an extremely sensitive mechanical oscillator
[101, 102].
Although there are many ideas for solving the snr problem, integrating these ideas with the other
requirements for quantum computing in a realistic architecture is very challenging, even on a theoretical basis. In the following sections, we consider several ideas for nuclear-spin-based architectures
and comparatively explore their advantages and disadvantages.

3.1 NMR Physics

3.1

57

NMR Physics

We begin by reviewing the basic physics of nmr. We will mostly limit the discussion to spin-1/2
nuclei, which are generally preferred to higher-spin nuclei due to the fact that their interactions are
limited to magnetic dipole interactions. In this simple case, the entire nmr Hamiltonian can be
written in the form
H =

X
j

B(rj , t) Ij ,

(3.2)

where the sum is over all the nuclei in the sample and B(rj , t) is the total magnetic field at the nuclear
site rj . Of course, there are many sources of magnetic fields that contribute to B(rj , t). Usually, an
nmr experiment is designed so that the largest magnetic field, by far, is the one we directly apply
with an external, usually superconducting magnet. By convention, this magnetic field points in the
z-direction. A typical magnetic field (and the one corresponding to the nmr magnet used for the
experiments described in this dissertation) is 7 T. At such a field, the energy splitting of protons (the
Larmor frequency) is B0 /2 = 300 MHz. It is worth pointing out immediately that this frequency
is already lower than the clock speed of modern digital computers and is the slowest oscillation
frequency of any proposed qubit. The exponential speed-up gained by quantum algorithms will
guarantee that an nmr quantum computer will operate faster than the fastest classical computer for
some problem size, but we must be aware that if the quantum computer becomes too slow, the time
for either type of computer to solve the problem may be prohibitively long, as indicated in Fig. 1.1.
The local environment of the nucleus shields it from the applied magnetic field, so the magnetic
field at the nuclear site is usually smaller by a few parts per million. This may be due to a variety of
effects; among them is the diamagnetic shielding from valence orbitals near the nucleus, known as the
chemical shift. In metals, the Knight shift due to the contact hyperfine coupling to the conduction
electrons is important. If there are unpaired electrons in the sample, the average dipolar or contact
hyperfine couplings to them may lead to an additional frequency shift. Further, the bulk magnetic
susceptibility of the sample may lead to a diamagnetic shielding which depends on the geometry of
the sample. In this dissertation, I lump all of these time-independent, sample-dependent shifts into
the applied magnetic field B0 , an action which sweeps under the rug everything that is interesting
about nmr spectroscopy! Those interested in the physics behind spectroscopic shifts in nmr are
advised to consult the textbooks of Abragam [121] and Slichter [122]. In the end, we have the largest
term in the Hamiltonian, the Zeeman term
HZ = 0

X
j

Ijz = 0 M z ,

(3.3)

58

Chapter 3. NMR Quantum Computers

where 0 = B0 for the applied field B0 including any uniform shielding effects or time-independent
magnetic fields. The use of the total magnetization operator M will be convenient for homogeneous
ensembles, since if only homogeneous fields are applied this operator acts as a single spin vector.
Typical nmr experiments are dominated by the eigenstates of M z with E close to zero; these eigenstates are each separated by the integer energy ~0 . For an ensemble of N 1 nuclei, the degeneracy
of the mz eigenstate is approximately the asymptotic Shannon information carried by N bits with

polarization 2mz /N , namely 2N H(2m

/N )

2N exp(2m2z /N ). This Gaussian degeneracy function

dominates the Boltzmann statistics for small polarizations. For example, the average thermal magnetization (which is exactly N p/2 = (N/2) tanh 0 /2, for = ~/kB T ) may be calculated from the
partition function as an integral
1
1
hM i =
ln Z
ln
0
0
z

e2mz /N e0 m dmz =

N 0
.
4

(3.4)

This is a useful method to obtain higher-order thermal expectation values.


The degeneracy of the Zeeman Hamiltonian may be broken by inhomogeneous broadening. This
occurs because the magnetic field across the sample is not uniform, due to accidental or deliberate
inhomogeneities in the applied field or due to variations in the local magnetic field such as varying
chemical shifts or non-uniform bulk susceptibility response fields. Non-uniform time-independent
magnetic fields across the sample are placed into a separate shift term
H =

X
j

[B(rj ) B0 ]Ijz .

(3.5)

The next largest term is the hyperfine coupling to the electrons in the system, whose form is
discussed in Sec. A.5. The size of this term may vary from hundreds of kHz to more than 100 MHz
depending on the wave function of the electrons. Actually, second-order effects due to this term
have already been lumped into the Zeeman Hamiltonian, as this term is ultimately the source of
the chemical shift. This term will further be important for polarizing nuclei, which we discuss in
Sec. 6.3.3, and for electron-mediated nuclear-nuclear interactions, which I discuss in Sec. 3.3.
The next largest term is the resonant or near-resonant radio frequency (rf) field, typically applied
using a large coil of wire arranged orthogonal to the applied field in a tuned rf circuit. An rf current
in the coil induces a magnetic field 2B1 (rj , t) cos[t (t)] in the x-direction, by convention. The
magnitude 2B1 (rj , t) and the phase (t) of the rf are externally controlled parameters, as indicated
by their time dependence. A few words on the classical description of this field are worthwhile.
Quantum effects with the rf are not easily observed, and indeed it has been shown convincingly

3.1 NMR Physics

59

[123] that the physics of the interaction of nuclei with an inductive coil is not the same as that of an
atom with a cavity, in the sense that the interaction between the two is not mediated by radiation
and absorption of real photons. Rather, the physics is near-field exchange of virtual photons, and
the full quantum mechanical description of these high-order processes is neither simple nor useful,
so our semiclassical approach is both reasonable and sufficient.

We now make the first of several dynamic coordinate transformations that we will consider in this
dissertation. This one is common to nearly every analysis of high-field nmr: the rotating frame. The
rotating frame picture of nmr is very similar to the usual interaction picture of orthodox quantum
mechanics, but an important distinction sometimes leads to confusion. The rotating reference frame
does not rotate at the resonant frequency of the nucleus, 0 , but rather at the frequency of the rf
signal, . The transformation to this reference frame is therefore

H exp it

X
j

Ijz H exp it

X
j

Ijz +

Ijz .

(3.6)

I will use this rotating frame so pervasively that it will be given no special notation. Generally,
analysis of pulse sequences for a closed nmr system will take place in the rotating frame by default,
while analyses of interactions between nuclei and electrons for calculations of relaxation, dynamic
polarization, and measurement will take place in the laboratory frame unless noted otherwise. The
rf Hamiltonian in the rotating frame is written
Hrf =

X B1 (rj , t) 
j

j (t)
=
2

2
X
j

iti(t)

Ij+ ei(t)

+e

it+i(t)

Ij ei(t)



Ij+ eit

Ij eit



+ 2it+i(t)
2iti(t)
+ Ij e
+ Ij e
.

(3.7)

The term in the curly braces is the counter-rotating term, and corresponds to the component of
the linearly polarized rf field which rotates in the opposite orientation from the rotation of the
nuclei in a magnetic field. As long as the Larmor frequency 0 = B0 is much larger than the Rabi
frequency j = B1 (rj ), this counter-rotating term has very little effect on the nuclei due to its
rapidly oscillatory character in the rotating frame. This kind of term is called non-secular, and its
omission is an approximation which is only invalid in the rare situtations that the Rabi frequency
approaches the Larmor frequency. With this term absent, the rf Hamiltonian in the laboratory

60

Chapter 3. NMR Quantum Computers

frame is rewritten
Hrf =

X
j



j (t) cos[t (t)]Ijx sin[t (t)]Ijy ,

(3.8)

showing that our rf field is rotating at frequency . For this reason the omission of the counterrotating term is an approximation known as the rotating wave approximation (rwa).
We now note that in the rotating reference frame, the combination of the Zeeman, shift, and rf
Hamiltonians is
HZ + H + HRF =


X
j (t) + i(t)
j Ijz +
(In e
+ In ei(t) ) ,
2
j

(3.9)

where
j = B(rj ) .

(3.10)

This Hamiltonian causes a rotation for spin j about the unit vector
=
n

[j cos , j sin , j ]
q
,
2j + j2

(3.11)

with generalized Rabi frequency


=

q
2j + j2 .

(3.12)

This is the way all nuclear spin control is obtained in nmr. We may choose the angle of rotation
via the phase (t) of our rf, and the duration of the rotation by pulsing the Rabi frequency j (t).
A few important terms should now be established. A pulse with = 0 is termed an X-pulse,
and if it lasts a duration j (t) = we denote the subsequent rotation RX (). The choice = /2
is a Y -pulse, = is an X pulse, = /2 is a Y pulse, and rotation with a general choice of
might be notated R (). A typical nmr experiment is to apply a single /2 pulse to spins initially
pointing in the z-direction. After the rotation R (/2) the transverse oscillating magnetization is
observed; this experiment is termed free induction decay (fid).
A broadband or hard pulse is one chosen so that j j for all j, with a further assumption
that the rf field is roughly constant across the sample so that j = . As a consequence, the
is negligible. Therefore,
generalized Rabi frequency is for all spins, and the z-component of n
all spins undergo the same rotation about some transverse vector corresponding to . A selective
or soft pulse is one which is resonant only with a particular spin or spin ensemble, say spin k, so

3.1 NMR Physics

61

that k = 0. It excludes spins of different frequency with a weak Rabi frequency j (t) j ,
where j 6= k. Soft pulses are frequently shaped, meaning j (t) is also a function of time whose
Fourier transform is band-limited to less than the lowest j . Then, the resonant spin undergoes a
rotation about a transverse angle, while non-resonant spins see only a rotation about the z-axis. This
spurious z-axis rotation can be further compensated for if needed; methods for this have recently
been discussed in the context of nmr quantum computation [124]. We will return to the uses of
hard and soft pulses in Chapter 4.

Let us now return to the hierarchy of magnetic fields witnessed by nuclei. After the externally
applied and rf fields, the next largest magnetic field witnessed by nuclei is typically the magnetic field
caused by other nuclei in the sample. These magnetic fields may be effective magnetic fields due to
distortion of nearby electron wave functions by neighbor nuclei; these are electron-mediated couplings
such as the J-coupling which are especially important in nmr of molecules in liquid solution. A
general theory of indirect nuclear couplings will be discussed in Sec. 3.3. Another coupling is much
more pervasive than these indirect couplings in solid-state nmr: the direct magnetic field due to the
dipole moment of a neighbor nucleus. This leads to the general dipole coupling in the laboratory
frame

0 X ~2 2
HD =
3
4
rjk
j<k

(Ij rjk )(Ik rjk )


Ij Ik 3
2
rjk

This dipole coupling Hamiltonian is subdivided into several terms as HD =


(0)
HD
(1)

HD

(2)

HD

(3.13)
P2

m=2

(m)

HD

where




0 2 X 1 3 cos2 jk z z 1 +
+
=
~
Ij Ik
I I + Ij Ik ,
3
4
rjk
4 j k

(3.14)

(3.15)

j<k


3 0 2 X sin jk cos jk eijk  z
~
Ij Ik + Ijz Ik ,
3
2 4
rjk
j<k

3 0 2 X sin2 jk e2jk
~
Ij Ik ,
3
4 4
rjk

(3.16)

j<k

where we have introduced standard polar coordinates for the vector rjk connecting nuclei j and k.
These components may also be written in terms of spherical harmonics Ylm (ij ), where ij is the
angle between the z-axis and the vector connecting spins i and j, and spherical tensor operators
Tijm , as
(m)
HD

0 2
=
~
4

4 X
Y m (jk ) m
(1)m 2 3
Tjk .
5
rjk
j<k

(3.17)

62

Chapter 3. NMR Quantum Computers

B0

Figure 3.1: The dipolar magnetic field BD (r) generated by one nucleus is seen by a
random assortment of other nuclei. Only the z-component of the dipole field is important
at high applied fields B0 ; this is the secular component. This component vanishes at the
magic angle M shown.

This subdivision is useful because of the relation




(m)
(m)
z
M , HD
= mHD .

(3.18)

More generally,


m
m
= mTjk
M z , Tjk


p
m1
m
M , Tjk
= 6 m(m 1)Tjk
.

(3.19)
(3.20)

(0)

This implies, in particular, that HZ and HD share a set of eigenstates, which I label |mz , i , where
(0)

is the eigenvalue of HD . The eigenfrequencies are thus written hmz , | HZ |mz , i = mz 0 and
(0)

hmz , | HD |mz , i = .
(m)

In moving to the rotating frame and the rwa, it follows that in this frame the term HD

is

(0)

proportional to exp(imt), and therefore all terms besides HD are dropped. Unless otherwise
noted, only this m = 0 term will be considered, and the superscript notation will be reserved for
those rare occasions when the nonsecular terms become important.
A more geometric way to understand the truncation of the non-secular terms is shown in Fig. 3.1.
The magnetic field of a neighboring nucleus BD (r) is not strictly in the z-direction; it also has
transverse components. All components are much, much smaller than the applied magnetic field, so
define = BD (r)/B0 as a vector whose components are all much less than one. The magnitude of

3.1 NMR Physics

63

the total magnetic field seen by a nucleus may then be written as


|Btotal | = B0

p
(1 + z )2 + (x )2 + (y )2 .

Using a 1, we retain only the term first order in , leaving


|Btotal | = B0 [1 + z + O(2 )].
(0)

This term z yields exactly the secular Hamiltonian term HD .


For quantum computers, the dipolar coupling will play the dual role of being a nuisance leading
to decoherence in spin ensembles and being a tool for allowing universal quantum logic. A useful
visualization of the dipolar coupling as a source of decoherence is to regard it as a large bath
of oscillators, as in the spin-boson model. In this picture, we make the approximation that the
eigenstates of the Zeeman and dipolar Hamiltonians may be written as a direct product space,
|mz , i = |mz i |i. Of course, this is not exactly true; the state space of dipolar eigenstates
|i changes when mz changes, but for mz 0 this picture allows reasonable approximations to

the dynamics. Then the part of this product space corresponding to the Zeeman eigenstates is the
system in which we attempt to maintain populations and coherences, while the product space
corresponding to the dipolar eigenstates is the bath which may exchange energy (at low field)
and entropy (at any field) with the system. We will return to this picture several times. Of course
this system/bath picture will mislead us when we use rf pulse sequences to reverse the dipolar
Hamiltonian, which will be the subject of Chapters 4 and 5.
The many-body nature of the dipolar coupling prevents analytic calculation of its eigenstates
|i. A sufficient approximation for the dynamics due to this Hamiltonian in this dissertation is the
Van Vleck method of moments, which is well described in various textbooks [121, 122, 125, 126].
The method may be succinctly summarized in the Liouville notation. The expectation value of the
transverse magnetization of the density operator in the rotating frame due to the secular dipolar


Hamiltonian is exp(iHD t) (0) , but we cannot analytically exponentiate HD because we do not

know its eigenstates. So we first ask only about the short time evolution by expanding the exponential
to lowest order in t to find


(t)





1 2 2 i 3 3
1 iHD t HD t + HD t + (0) .
2
6

(3.21)

Suppose we seek the average decay of the phase of an ensemble of spins due to the dipolar coupling.

64

Chapter 3. NMR Quantum Computers





Hence we seek, for example, the normalized expectation value M x (t) / M x M x . Now, since

we have applied the rwa, the secular dipolar Hamiltonian will not alter the coherence number of

any term in the density operator, since it commutes with M z . In other words, it conserves energy;
it cannot flip one spin up in the ensemble without flipping another down. As a consequence, the


only term in (0) that contributes to this trace is M x . Another consequence of the conservation


of coherence number is that terms with odd powers of t vanish from M x (t) . Therefore we have
2 x
4 x



M
M
M x eiHD t M x
1 M x HD
1 M x HD
2




 t +
 t4 +
1
x
x
x
x
x



2
6
M M
M M
M Mx
1
1
= 1 + M 2 t2 + M 4 t4 + .
2
6

(3.22)

These Taylor coefficients Mn are easily seen to be the nth moments of the Fourier transform of




M x (t) / M x (0) . Evidently we may calculate the 2nth moment as
n

M2n = (1)

2n x 
M
M x HD

 .
M x M x

(3.23)

Calculating high order moments becomes challenging due to the large number of commutators hidden
in this expression. If the shape of the resonance function were Gaussian, then the second moment
1/2

would characterize the entire decay function. In reality, this is not the case, but M2
as an order of magnitude for

T21 .

More precise means for estimating

T21

may suffice

in terms of functions of

higher moments may be found in Refs. 125, 126.

Aided by the spherical tensor relations of Eq. (3.20), we find that the second moment is
M2 =


2 X (1 3 cos2 )2
1
3
0 2
jk
Tr {[HD , M x ][HD , M x ]} =
I(I + 1)
~
.
6
N
2N
4
rjk

(3.24)

This is the Van Vleck formula, valid for any I. This expression will come in useful as an estimate
for unwanted effects of the dipolar coupling.

A final addition to the magnetic fields witnessed by nuclei are those local magnetic fields caused
by electrons in the sample that fluctuate in time. These fluctuating magnetic fields lead to spin
relaxation and decoherence, as will be discussed at length in Chapter 6. For now, it will suffice
to write the phenomenological equations used to describe the combination of rf pulses and spin
relaxation in the rotating frame, written down by Bloch in his seminal paper on nuclear induction

3.1 NMR Physics

65

in 1946 [127]:

x
d x

hIj i = hIj i n
dt

y
d y

hIj i = hIj i n
dt

z
d x

hIj i = hIj i n
dt

1 x
hI i
T2 j
1 y
hI i
T2 j


1
hIjz i hIjz iT ,
T1

(3.25)
(3.26)
(3.27)

is given by Eq. (3.11) and hIjz iT is the polarization at thermal equilibrium. These equations
where n
are not always strictly valid, but they hold true in a remarkably large number of situations and
provide both the historical and conceptual introduction of the two important relaxation timescales.
The first is T1 , the rate of spin relaxation to thermal equilibrium by inelastic scattering with the
environment, and the second is T2 , the rate of decoherence by elastic scattering. As discussed at
the beginning of the chapter, the very long decoherence rates T2 of nuclei, relative to their resonant
frequencies 0 , are the reason nuclei hold so much interest for quantum computers in the first place.
The origins of T1 and T2 may be conceptually understood as resulting from magnetic noise. We
presume a noise term in the interaction Hamiltonian of the form
Hint (t) =

X
j

b(rj , t) Ij ,

(3.28)

where by assumption the small fluctuating field b(r, t) has zero mean. Noise from this field that is
resonant with the nuclei will cause the nuclei to rotate about random axes in the xy-plane, causing
the nuclei to both dephase and repopulate their states back to thermal equilibrium. This picture
provides some understanding that
2
1
=
T1
2


hbx (t)bx (0)i + hby (t)by (0)i cos(0 t)dt = 2 J (0 ).

(3.29)

Likewise, random magnetic fields in the z-direction will change the frequency of nuclei by a random
amount, leading to decoherence by spectral diffusion. In addition to the T1 effect, then, transverse
decoherence is written
1
1
2
=
+
T2
2T1
2

hbz (t)bz (0)idt =


2 
J (0 ) + Jk (0) .
2

(3.30)

This equation for T2 only applies for decoherence due to noise from memory-less processes of a large
reservoir of states. Its applicability to small environments or to effects with memory such as the

66

Chapter 3. NMR Quantum Computers

dipolar coupling among nuclear ensemble members is very limited. We will return to these issues
and derive these two equations in Chapter 6.

3.2

Liquids vs. Solids

We begin the approach to solid-state nmr quantum computers by describing liquid-state nmr quantum computers. Liquid nmr computers have had a technological head start over competing architectures. Their early success for implementing small algorithms and error-correcting protocols prior to
2001 was summarized by Cory et al. [112]; since then they have been applied to many applications,
including Shors algorithm [67]. We will discuss how each aspect of operation must change when
moving from these liquid-based computers to solid ones; some changes will be advantageous, and
others will not be.
Liquid-state nmr quantum computers operate with an unscalable architecture not only because
the number of measurable qubits is limited, but also because of the way an algorithm must be structured. These computers initialize the nuclei thermally in a first stage, undergo coherent evolution to
perform the given algorithm, and are then subjected to a slow, weak, ensemble measurement. The
experiment may then be repeated several times and the results combined in post-processing. Ironically, the coherent evolution where the actual quantum computation occurs is usually the shortest
part of the experiment! As discussed in Chapter 2, this three-stage initialize-compute-measure
model of computation is difficult to make scalable, as fresh qubits must constantly be supplied to
remove the entropy created by decoherence and faults in the computation. If only one global initialization step takes place, then this means the supply of qubits is quickly exhausted. For molecules in
a liquid, the number of spins in a given molecule is quite limited. While solid-state nmr quantum
computers with thousands of qubits may be able to work in this model, even these will run out of
fresh qubits too quickly to allow long algorithms. Nonetheless, in this chapter we will limit ourselves
to considering architectures that operate using this three-stage model, and return to this important
point at the end of the chapter.

3.2.1

Initialization

The easiest way to build up enough signal to measure nuclear spins is to form a large ensemble of
N nuclei. Doing so may seem to solve the measurement problem, but actually it raises the related
and almost equally difficult problem of initialization. In liquid nmr computers, there is no dynamic
cooling in the initialization phase. The nuclei are simply allowed to reach thermal equilibrium at

3.2 Liquids vs. Solids

67

room temperature, for which polarizations p are of order 106 .

Pseudo-pure states
The limitations presented by the low polarizations in liquid-state nmr were recognized in the first
proposals for nmr quantum computers [36, 37], but it was also noted that while these problems
prevent scalable quantum computation, they do not prevent quantum computers of limited size.
The novel content of these proposals were methods for generating pseudo-pure states. I alluded to
these methods in Sec. 2.2.2 where I showed that measuring highly correlated observables can allow
the post-selected extraction of a highly polarized subensemble that undergoes the desired quantum
algorithm. In practice, directly measuring such highly correlated observables is very difficult, yet
various very efficient ways to go about isolating the desired subensemble have been developed. I
summarize only the two simplest techniques below.
The first is state labelling, as proposed by Gershenfeld and Chuang [36]. Here you throw away
some qubits in order to label the active ones. This is best described by a trivial example. The high
temperature, thermal equilibrium density operator for two spins of similar Larmor frequency 0 can
be written to lowest order in the polarization p
 1 
T 1 +
4


p
1 2
2


p
1 2 .
2



(A clarification of notation is in order; the notation 1 2 is the Liouville ket for the projector
operator |1 2 ih1 2 |; this notation is described in Appendix A.3.) In order to achieve the one-qubit

effective pure state 1 , the final measurement is conditioned on the second spin. Run the quantum
computation on the first spin C1 on the thermal state (no preprocessing) and measure at the end
 p

I1z 2 C1 T = I1z C1 1 ,
2

where the equality results from the observations that




I1z 1 = 2 2 = 0,


2 2 = 1.

We apparently obtain the result of a computation performed on one completely pure qubit, with an
overall scale factor of p/2.
An example with one more qubit indicates how this procedure may scale upwards. In the case

68

Chapter 3. NMR Quantum Computers

of three similar qubits, the thermal density operator is approximately


 1  3p

T 1 + 1 2 3
8
8

3p

1 2 3 +
8


p
1 2 3
8

p
1 2 3 +
8


p
1 2 3
8

p
1 2 3 +
8


p
1 2 3
8

p
1 2 3 .
8

A certain unitary preparation P may easily rearrange these states in a desired way. The quantum
circuit is not hard to construct, but it is more easily described in words: flip spin 1 if spins 2 and
3 are in opposite states. The resulting density operator is
 1  3p

P T 1 + 1 2 3
8
8

3p
1 2 3 +
8


p
1 2 3
8

p
1 2 3 +
8


p
1 2 3
8

p
1 2 3 +
8


p
1 2 3
8

p
1 2 3 .
8

We may now perform our computation on qubits 2 and 3 conditioned on qubit 1 to find



 p

1 I2z I3z C23 P T = I2z I3z C23 2 3 .
2
The general principle is to look for states that are equally populated in the density operator, and
treat this as a uniform background for the subspace that is to be pure. This subspace plus the fully
polarized state form the 2k states of the pseudo-pure subspace. Fortunately, most of the states of
a thermal density operator are equally populated; for n qubits with equal Larmor frequencies, the
number of logically labelled pure qubits which may be obtained is

k = log2 1 +

n!
(n/2)!2

for example k = 37 for n = 40.


It may not be obvious how to apply the projection operators selecting the state of an ancilla bit
when working with an ensemble. In practice, one measures the state without any post-processing
operations, obtaining a sum of the result for all ancilla states. Then one repeats the computation
and adds to the end a post-processing step consisting of a series of controlled-not gates to invert the
answer for each undesired ancilla state. Summing the two results cancels the blocks of the density
operator in which we are not interested. An experimental demonstration of logical labelling was
performed by Vandersypen et al. [128].
The problem with state labelling, above, is that it costs qubits, which are especially expensive
in liquid molecules. A more commonly used approach which does not sacrifice the use of qubits

3.2 Liquids vs. Solids

69

is exhaustive temporal averaging [129], in which the terms of the density operator which do not
represent the desired pure state are shuffled in several experiments in order to cancel. Once again,
a two-qubit example illustrates the idea. Consider the initial density operator and the following
preprocessed density matrices:
 1 
P0 T = 0 =
3
 1 
P1 T = 1 =
3
 1 
P2 T = 2 =
3


1 1 
1 +
3 4

1 1 
1 +
3 4

1 1 
1 +
3 4


p
1 2
2

p
1 2
2

p
1 2
2



p
1 2
2


p
1 2
2


p
1 2 .
2

These preparation operators Pb are simply single spin rotations with a scalar multiplication. No
post-processing is needed. The initial effective density operator from adding the three is
 X  1
T =
Pb T =
4
b



 4

4
1 p 1 + p 1 2 ,
3
3

which is exactly the effective pure state needed. The problem with this method, of course, is that
one needs to run the quantum algorithm 2n 1 times, so the time needed for any computation scales
exponentially with the number of qubits.
Other techniques to establish pseudo-pure state techniques have been developed [129], including
the use of gradient coils to spatially label the desired subensemble [112]. In principle, most of
these techniques can be used with resources that grow subexponentially as the quantum computer is
scaled up, but all would require an exponentially large number of averages over many experiments to
maintain snr. Another crucial detail is that they require knowledge of the probability distribution of
the initial state. This is not an issue in the case of thermal equilibrium, but could become important
if non-equilibrium polarizations are used.
Dynamic Nuclear Polarization
Inevitably, scaling to ensemble-based nmr computers larger than about 10 qubits must require
enhanced polarization or more sensitive measurement techniques; in all realistic situations it will
require both. This is the principle reason to move from liquid-state nmr to solid-state nmr, for in
the latter both enhanced polarization and ultra-sensitive measurement have been demonstrated in a
variety of experiments. Fortunately, the long thermal relaxation times of nuclei in solids mean that
if the nuclei can be cooled independent from the lattice, they will remain cold for sufficiently long
periods of time to be useful.

70

Chapter 3. NMR Quantum Computers

Several attempts to use enhanced polarization in liquid nmr quantum computers have been
attempted, by cross relaxation with easily polarizable nuclei such as the protons in para-hydrogen
[130] or hyper-polarized xenon gas [131]. While these and other efforts increased the snr of the
experiments by some factor, the enhancement is not nearly sufficient for scalable operation, even
using the algorithmic cooling techniques discussed in Sec. 2.2.2.
Polarization is not feasible by simply cooling the temperature of the lattice. Static laboratory
magnetic fields generated by superconducting solenoids are limited to about 10 T. Referring to
Eq. (2.11), p becomes order unity only for T . 5 mK. Even if the lattice is cooled and maintained
at such cold temperatures, the question remains as to whether the nuclei will equilibrate with the
lattice on a reasonable timescale. At low temperatures, electron spin fluctuations freeze out and
increase the isolation of nuclei. For example,

19

F nmr has been performed in CaF2 in a dilution

refrigerator [132]. At T = 10 mK, a thermal relaxation time (T1 ) of about 4 months was observed
at B0 = 0.15 T. This already long timescale was observed to become exponentially longer as the
magnetic field was increased. And this empirical measurement was considered unexpectedly rapid
in comparison to the theoretical expectation. Of course this behavior is not universal; for example
nuclei in metallic samples may still have reasonable T1 times at very low temperatures, and nmr
in metals is even used for ultra-low temperature thermometry. However, even if a sample could be
sufficiently polarized by cooling the lattice in high field, other architecture requirements such as the
use of high-power rf pulses for nuclear-spin control are difficult at such low temperatures.
Dynamic polarization of nuclei has a long and established history, beginning with the prediction
of Overhauser [133] that saturating the electron spin resonance in a metal would lead to a 1000-fold
improvement in the nuclear polarization. An explanation of the Overhauser effect is aided by the
definition of spin temperature. The meaning of this concept seems conceptually obvious, but the
elucidation of certain subtleties may benefit from a few comments. The obvious definition is simply
the inverse of the Curie law:
1
1 1 z 
1
3
=
B
hI i
hI z i.
kB TS
~0 I
~0 I(I + 1)

(3.31)

If I > 1/2 there may not be a well defined spin temperature, since the 2I + 1 > 2 states may
not be distributed according to a Boltzmann distribution, but if they are then Eq. (3.31) applies.
However, when discussing ensembles of nuclei, this definition of spin temperature may be misleading.
A recurring theme in solid-state nmr is that even though the nuclear system may appear to have
irreversibly heated according to this definition of spin-temperature, that heating is entirely
reversible! The first example of this was Hahns spin-echo [134], where the heating was due

3.2 Liquids vs. Solids

71

to inhomogeneous broadening; later, magic echoes [135] reversed the heating due to dipolar
coupling. Another key consideration for spin-temperature of ensembles is the issue of spin diffusion.
Dynamic nuclear polarization may at first be localized to limited regions of the sample where the
polarizing electrons are located. The system does not truly equilibrate to a spin temperature until
the nuclear polarization is allowed to diffuse throughout the ensemble, a process which occurs via
the dipolar coupling. The spin-temperature hypothesis is that nuclear polarization effects in the
presence of spin diffusion lead to a well-defined, homogeneous distribution of states truly following
the expectations of a Boltzmann distribution at TS ; this hypothesis is not self-evident and not always
correct, but has provided a very useful conceptual framework in solid-state nmr for many years.
Its utility is that the final density operator of a system after polarization processes proceed may be
assumed to be a Boltzmann distribution corresponding to the total Hamiltonian. This point of view
is well discussed in several texts, for example in Slichter [122].
Let us return to the physical origins of the Overhauser effect. This effect relies on those processes
between electron and nuclear spins that tend to bring these two systems into thermal equilbrium. By
artificially manipulating the spin temperature of the electrons, the nuclear spin temperature is likewise manipulated, since there are no other appreciable pathways for nuclear relaxation. Overhauser
predicted that this principle would polarize nuclei in a metal by artificially eliminating electron spin
polarization using rf saturation. It was shown that this effect was neither unique to metals nor
to rf electron spin saturation. Early generalizations of the Overhauser effect which will be important for this dissertation were observed in semiconductors, where the electron spin temperature
was maintained out of equilibrium either optically [136] or by passing a direct current through the
sample [137]. Combined with modern silicon heterostructures, these techniques may yield very large
nuclear polarizations.
Other means of dynamic nuclear polarization exist as well, such as the solid-state effect in which
forbidden transitions that allow electron spin polarization to be directly transferred to nuclear
polarization are resonantly excited with appropriate radio or microwave frequencies [121]. Unlike
the various Overhauser effects, which generally work best with delocalized electrons, the solid-state
effect works with fixed impurities interacting with nuclei via the dipolar coupling.
An idea was proposed by Yamaguchi and Yamamoto [114] for transferring the polarization of spins
in an antiferromagnet to lattice nuclei using the solid-state effect. The problem with this scheme
is that it neglects the dynamics of the strongly-coupled electron spin system, instead assuming
that individual electrons see a fixed exchange field. This mean-field theory assumption can be
useful for certain calculations in the theory of magnetism, but resonant excitation of a localized

72

Chapter 3. NMR Quantum Computers

electron spin mode is not likely to be possible. Rather, the application of microwave frequencies to
antiferromagnets excites delocalized spin-wave modes across the entirety of the spin-lattice.
These spin waves surely have an effect on the thermodynamics of nuclei. They lead to nuclear
relaxation, and indeed to decoherence. A calculation of nuclear decoherence timescales in the proposed system of Yamaguchi and Yamamoto [114] due to spin-wave modes is presented in Ref. 138.
These calculations show how spin-waves might impede the progress of a quantum computer using
antiferromagnetic electron spins for initialization. However, later research considered the possibility
of using spin-waves to provide a controlled means of rapidly coupling nuclei [139, 140]. I will briefly
discuss the use of spin-waves to couple nuclei in Sec. 3.3.

3.2.2

Distinguishing Qubits

A glance at Table 3.1 provides a very encouraging view of liquid-state nmr, since the coherence
times for nuclei in liquid xenon, for example, are extremely long. It would be difficult to imagine
making a quantum computer with liquid xenon, however, since there would be no convenient way of
distinguishing one xenon nucleus from the next. Any means used to spatially distinguish the nuclei
would be quickly circumvented by the rapid motion of the atoms in the liquid.
For this reason, liquid-state nmr quantum computation is performed with molecules of low symmetry. Of course, nmr spectroscopy sees its most extensive use identifying the chemical environments
of nuclei in molecules. The different chemical shifts afforded to each nucleus in a molecule are a
reflection of the different electronic wavefunctions from the chemical bonds in the molecule. Using
different spin-1/2 nuclei in differing chemical environments has allowed liquid-state nmr quantum
computers with increasing numbers of nuclear spins. For example, the 7-qubit molecule used for
Shors factoring algorithm [67] was the result of considerable chemical engineering. Finding more
complex molecules is a challenge for scaling liquid-state nmr quantum computers, although nmr
spectroscopy is routinely performed on complex molecules with hundreds to thousands of resolvable
nuclear resonances.
In principle, this method of resolving nuclei can be carried over to solid-state nmr using molecular
crystals, as discussed in Sec. 2.4.2, but the scalability of such ideas is extremely limited. Another
possibility is motivated by the consideration that nuclei are routinely distinguished by position in
magnetic resonance imaging (mri). This is accomplished using magnetic field gradients generated
with current-carrying coils in suitable geometries. The resolution of mri is limited by the snr of
inductive pick-up. If a solid-state nmr system had extremely high snr, it could in principle employ
much larger gradients in order to further improve the spatial resolution, leading to the highly desired

3.2 Liquids vs. Solids

73

goal of non-invasive atomic-scale microscopy.


The dream of imaging single nuclear sites has had its greatest experimental push in mrfm
[101, 102], where field gradients of order 1 T/m are routinely achieved using small ferromagnetic
particles located on the end of high-Q mechanical cantilever oscillators. (Generating field gradients
this high with conducting coils is more difficult.) If the detection sensitivity is sufficient, gradients
this large are sufficient for resolving atomic planes in typical materials.
However, further geometry considerations must also be made. In mrfm, the gradient induced
by the ferromagnetic particle brings into resonance a slice of the sample in the rough shape of an
elliptical shell. For atomic resolution, a single spin must be isolated within this slice, or else the
gradient must be moved around the sample and the data deconvolved to reconstruct an image, as in
computer-aided tomography. In a solid-state nmr quantum computer, the option of simply isolating
a single nucleus is not available, since multiple nuclei are inevitably desired for computation. The
option of mechanically moving a gradient is also not practical, since this would limit the operation
time of the architecture to the slow mechanically-induced gradient switching time.
A more practical architecture would be to have a fixed magnetic field gradient arranged in a
suitable geometry to create resonant slices that coincide with the atomic structure of the nuclear
system. For example, if nuclei are arranged in a one-dimensional chain, it is sufficient that the
profile of the slice be orthogonal to the chain. However, if ensembles of chains are used, then this
orthogonal slice must be sufficiently flat that all of the ensemble members see nearly the same field
z
across the slice. This means that the gradient field BG
(r) must be one-dimensional, meaning the
z
vector G(r) = BG
(r) points in the same direction for all r. Strictly speaking, this is not allowed

by Maxwells equations, but it may be well approximated given that the z-component of BG (r) is
much more important than the transverse components in light of the rwa, and that we only need
the one-dimensional gradient to persist over a very small sample, possibly consisting of only 105 to
108 atomic sites. Further, the ensemble members do not need to see exactly the same field, as small
amounts of inhomogeneous broadening generated by small field curvature can be refocussed with
spin-echo techniques. It is sufficient that the inhomogeneous linewidth caused by this curvature be
much smaller than the frequency separation induced by the field gradient.
The constraints on field gradient design, then, are not intractable, and specific designs for generating gradients for solid-state nmr computers using lithographically designed ferromagnets were
presented in Ref. 141. Most of these designs use a simple box geometry, as shown in Fig. 3.2. If the
ferromagnetic box is very small, then it generates a dipole field which varies with distance r as 1/r3 .
If, on the other hand, the ferromagnetic box is infinitely large, then it generates a perfectly uniform

74

Chapter 3. NMR Quantum Computers

B0

4
3
2
1

1 2 3 4

Magnet

Crystal

Crystal
Magnet

Figure 3.2: Sample geometries for generating a large magnetic field gradient across a
regular crystalline lattice. The magnetic field in the macromagnet causes planes of nuclei
to have unique Larmor frequencies j . Calculations of expected magnetic field gradients
for these geometries appear in Ref. 141.

field with no field gradient. Finding a sufficiently homogeneous one-dimensional gradient is then a
matter of finding a middle ground between these extremes. The calculations in Ref. 141 determine
the field from a particular geometry by simply integrating over the magnetic dipoles contained in
the material. For such a calculation to be valid and not drastically altered by self-demagnetizing
fields, the ferromagnet must either be extremely hard or be well saturated by a large externally
applied field. In particular, Ref. 141 suggests using dysporsium, whose saturation-moment is among
the highest of elemental metals, in a large magnetic field (> 3.5 T).
One must bear in mind the fabrication difficulties of making such magnetic field gradients.
Besides the ability to fabricate the desired structures, they require the relative placement of the
micromagnet and the quantum computer lattice to have atomic scale resolution. This criterion is
not difficult to meet in terms of the distance between the micromagnet and the computer lattice,
but rather because of the angle between the gradient direction and the relevant crystal axis.
If sufficient control over the angle of the gradient is available, then one-dimensional gradients

3.2 Liquids vs. Solids

75

12

11

10

G
8

b
6
3

2
y

Figure 3.3: The basic geometry for a two-dimensional magnetic field gradient. The
gradient field G points at a slight angle from the x-axis. The gradient causes each
nucleus in the grid to have a unique Larmor frequency j .

are further generalizable to two-dimensional geometries, simply by tilting the one dimensional
and y
directions to be parallel to the lines of
field gradient. Referring to Fig. 3.3, define the x
nuclei in the grid. The gradient in the z-component in the magnetic field, G, points approximately
, but making a slight angle . We consider this gradient to be constant across the lattice.
along x
Correspondingly, two nearest-neighbor nuclei separated by the vector b
x see magnetic fields varying
by bG cos , while two nearest-neighbor nuclei separated by the vector b
y see magnetic fields varying
and
by bG sin . We define an active lattice of finite extent as a n m, with n nuclei along x
. All nuclei will be distinguished as long as bG cos > mbG sin . This condition
m nuclei along y
assures that all nuclei in the active area see a different Larmor frequency. Nuclei outside the active
area have frequencies equal to those within; these nuclei may be neglected if they are out of range of
the measuring apparatus. They will unfortunately couple into the active area due to their dipolar
coupling, and avoiding this kind of problem is a key element in solid-state nmr quantum computer
design. We return to this issue only in Sec. 3.5.

3.2.3

Quantum Logic and Decoupling

Another crucial difference between liquids and solids in ensemble based nmr is the problem of
decoupling. Suppose we have an ensemble of N identical spin-1/2 nuclei, and suppose that this

76

Chapter 3. NMR Quantum Computers

ensemble is perfect in the sense that every spin sees exactly the same environment (i.e. the same
magnetic field, the same rf control fields, the same initial polarization, etc.). Although the Hilbert
space of this system has 2N basis states, the high symmetry effectively limits the space to the 2N fully
symmetric product states. These states are merely N -fold copies of the states that would be accessed
through single spins subject to the same environment and control operations. The density operator
Q
for a qubit may at all times be written = j j , where j enumerates the members of the ensemble,
 Q

Q
and all operators may be written C = j Cj , so that C = j Cj j . The computation begins,

proceeds, and ends in the symmetric product state. Mathematically, the ensemble spin operators
M act effectively as single spin operators.
Now suppose that there is some coupling between the ensemble members. Unless each ensemble

member couples to every other ensemble member with the same strength, this Hamiltonian will
not be contained in the fully symmetric product space, and in general will allow the system to
wander through all 2N states of the full Hilbert space. When the system leaks out of the symmetric
subspace, the fidelity of the computation has decreased. This may be considered a decoherence
process. Decoherence is always a result of losing information about the internal degrees of freedom
to a larger environment; in this case, the internal degrees of freedom are the 2N states of the
symmetric subspace, and the larger environment is represented by the remaining 2N 2N states
(the space of eigenstates of the coupling Hamiltonian).
Ensemble members in nmr are always coupled by virtue of the dipole-dipole coupling between
nuclei. Leakage from the symmetric subspace into the other states due to this coupling will be
discussed at length in this dissertation; we refer to it as dipolar decoherence. This effect must
either be minimized or reversed for quantum computing.
Solution nmr has a built-in minimization of the intra-ensemble dipolar coupling. In this case,
each ensemble member is a single molecule, floating in a dilute solution. The dipolar coupling to
the next ensemble member is very low in part because these molecules are spatially separated, but
mostly because the relative motion of these molecules averages the dipolar coupling to a very small
perturbation. The advantage of this system is that the rate of dipolar decoherence is very slow,
often exceeding many seconds. The disadvantage is that it is uncontrolled; the random motion of
the molecules means that information lost to dipolar decoherence is mostly irretrievable.
Decoupling is an especially difficult barrier to coherent nuclear spin manipulation in the solid
state; the lack of relative motion between ensemble members results in the timescale for dipolar
decoherence to be reduced from many seconds to less than a millisecond. However, solid-state nmr
presents a new opportunity: since the dipolar couplings are now temporally and spatially fixed in

3.2 Liquids vs. Solids

77

place, it becomes possible to time-reverse the dipolar Hamiltonian, a process known as decoupling.
We discuss the means for doing so in Chapter 4. With a combination of spatial separation of
ensemble members and decoupling techniques, the timescale of dipolar decoherence can be extended
by many orders of magnitude.
The long coherence times achieved by decoupling are useful for spectroscopy, but when we desire
quantum logic then we find we must restore the dipolar couplings between nuclei in order to construct
logic gates. The selective removal of decoupling is called recoupling, and nmr development has
produced a large variety of means for recoupling. Some of these will be also be discussed in Chapter 4.
However, a fundamental limitation to these decoupling and recoupling techniques is that they are
only sensitive to frequency. This has an important consequence for the geometry of nuclear lattices
to be employed in an ensemble-based solid-state nmr quantum computer.
Let us elaborate on this geometry constraint by supposing that we have ensembles of magnetically
distinct spins (meaning one ensemble differs from the next due to a large magnetic field difference,
created by a magnetic field gradient or a chemical shift). We might therefore notate each spin
operator with two subscripts, the first indicating which ensemble, and the second indicating which
member of that ensemble. A consequence of the large magnetic field shift between spins is that the
dipolar coupling between different qubit ensembles no longer conserves energy:
[H , HD ] 6= 0.

(3.32)

This is because the field gradient imposes an energy cost ~ to an inter-ensemble flip-flop

+
process, where is the frequency separation between ensembles. Therefore, for the Iij
Ikj terms of

the inter-ensemble dipolar coupling to be important, this energy must have a source or destination.
The lengthy T1 times which are typical of solid-state nmr indicate that processes in which this
energy is exchanged with the lattice are extremely improbable. It is known that this energy can be
absorbed into the dipolar bath [142], but such processes become less probable as the square of the
small ratio of the field gradient energy difference to the dipolar coupling energy. Neglecting such
terms, the dipolar Hamiltonian term coupling ensemble j to a different ensemble k is
HD,jk =

0 2 X 1 3 cos2 jl,km z z
~
Ijl Ikm .
3
4
rjl,km

(3.33)

lm

This is the interaction we will use for two-qubit logic gates between qubits j and k.. The sufficiency of
this interaction for universal quantum logic has been demonstrated [36]; we discuss how in section 4.4.
But here is the problem: even if we are able to decouple the dipolar coupling between members of the

78

Chapter 3. NMR Quantum Computers

4
z
3
a
2a
a

1a

Figure 3.4: An illustration of the geometry of nuclei discussed in the text. The nuclei
form a periodic lattice. The applied magnetic field and magnetic gradient are assumed
to point in the z-direction, so horizontal rows of nuclei have the same Larmor frequency
j . The nearest-neighbor qubit-qubit distance is a and the nearest-neighbor
distance

among ensembles is 1 a. If the angle is the magic angle, then 1 = 2.

same ensemble and decouple ensembles, when we desire to recouple ensembles j and k, the entirety of
couplings described by Eq. (3.33) are in place. The first ensemble member in ensemble j couples to
the first ensemble member in ensemble k, as desired, but unfortunately it also couples to the second
member of ensemble k, and the third, and so on. This creates a large network of cross-couplings
that short-circuits the intra-ensemble decoupling. The terms in Eq. (3.33) with l 6= m lead to an
error or decoherence term, whose decoherence timescale I will notate T2cc.

In liquid-state computers, this is not a problem because the decoupling due to molecular motion
is not a frequency selective decoupling, and the ability to achieve recoupling is due to the presence
of indirect molecular-bond mediated nuclear couplings (J-couplings) which are not averaged by this
motion. But in solid-state computers we must rely upon the 1/r3 dependence of the dipolar coupling
to lengthen T2cc by spatially separating ensemble members while maintaining a small qubit-qubit
distance. In a molecular computing approach, this may be achieved by assuring that the active
molecules are very dilute in the crystal, a possibility allowed by isotopic manipulation. A more
scalable geometry by which this is possible is for each quantum computer to be a straight atomic
chain, and for the chains comprising the ensemble to be well separated, as illustrated in Fig. 3.4.
Thus, different qubits are relatively densely spaced in a one-dimensional array, whereas the ensemble
corresponding to a single qubit is a large plane of relatively diffuse sites. Physical examples of such

3.2 Liquids vs. Solids

79

one-dimensional materials will be discussed below. For now, let us focus on the geometry where
such straight atomic chains are parallel to the applied magnetic field and to the field gradient,
although this is not a necessary condition.
To simplify notation, we write the nearest-neighbor qubit-qubit distance as rij,(i+1)j = a. The
nearest-neighbor desired qubit-qubit coupling strength is then
D=

0 2 ~
.
2a3

(3.34)

We also notate the distance between chains as rjl,j(l+k) = ak , as illustrated in Fig. 3.4. Then
Eq. (3.33) for the coupling between qubit j and another m atoms down the chain may then be
rewritten as
HD,j, j + m = D

X
1 X z z
2k /m2 2
I
I
+
D
Iz Iz
.
jl j+m,l
m3
2(2k /m2 + 1)5/2 jl j+m,l+k
l

(3.35)

kl

The second sum represents the unwanted cross-coupling terms. The effective decoherence caused by
these terms may be roughly estimated by the method of moments. We find


1
cc
T2 (m)

2

1
16

D
m3

2 X
k

(2k /m2 2)2

(2k /m2 + 1)5

2
D
(m)
.
m3

(3.36)

The function (m) characterizes the amount of error present due to parasitic couplings between
ensemble members when attempting to couple two qubits m planes away from each other. For a
very one-dimensional crystal, (m) can be quite small until m exceeds 1 , the ratio of the nearestneighbor chain-chain distance to the nearest-neighbor qubit-qubit distance. An important parameter
for a crystal lattice quantum computer is therefore 1 , which we desire to be much greater than
one. If this number is sufficiently large, then quantum gates between nearest-neighbor qubits can
be performed with acceptable levels of error.
One idea worth mentioning here is to use the dependence of the dipolar coupling on jl,km instead
of the dependence on rjl,km . For example, imagine a geometry where the desired coupling is strictly
along the z-axis, so jl,kl = 0, but the nearest-neighbor cross-coupling occurs along a vector which
makes the magic angle with the z-axis, defined by
1 3 cos2 M = 0
and illustrated in Fig. 3.1. This geometry may be achieved by putting 1 =

(3.37)

2. Although the idea

80

Chapter 3. NMR Quantum Computers

a=3.44

B0
L
W

i,j

9.42

i+1,j+1
i,j+1

1a

z
x

Figure 3.5: A schematic for a fluorapatite crystal lattice quantum computer. The fluorapatite crystal is shown mounted on a silicon cantilever, whose oscillations provide mrfm
readout. The cantilever and crystal are aligned with the micromagnet which generates
the field gradient. The inset shows the one-dimensional structure of 19 F nuclei in
fluorapatite.

of the magic angle is used very effectively in most decoupling techniques, as I discuss in Chapter 4,
it is not sufficient to help us here. This geometry would only eliminate the 1 terms in Eq. (3.36),

pushing the figure of merit to 2 . As illustrated in Fig. 3.4, this geometry limits 2 to 2 2, which
limits gate errors to order 3
0.04, an intolerable amount. A one-dimensional structure with
2
1 1 is more desirable than using the angular dependence of the dipolar coupling.
A natural crystal with a fairly high 1 is fluorapatite, Ca5 F(PO4 )3 , whose 19 F structure is shown
in the inset of Fig. 3.5. Quantitatively, we find that 1 =2.74 for this crystal, so (m) is only small for
m 1. However, if during a computation we always suppress all but nearest neighbor couplings, the
errors in quantum gates will be on the order of (1) = 1/58. Although not perfect, this is nearly an
order of magnitude improvement over simpler crystal structures such as the cubic structure of CaF2 .
The one-dimensional nuclear structure of fluorapatite was recognized in early NMR experiments
[143]. It has provided a one-dimensional testing ground for lineshape analysis investigations [144]
and, more recently, experiments studying multiple quantum coherence [145, 146], a topic intimately
related to quantum computation.
Could a quantum computer be made from fluorapatite? Like the well-studied CaF2 , this crystal is
an ionic insulator with a high Debye temperature. In such a material, interactions of the nuclei with
the intrinsic electrons can be completely ignored, and the effect of phonon-induced displacements

3.2 Liquids vs. Solids

81

leads to nuclear relaxation estimates that are astronomically long. Instead, it is observed that
nuclear relaxation in fluorapatite and similar crystals is entirely caused by paramagnetic impurities
[147]. In very pure crystals at low temperature, T1 can range from minutes to weeks1 .
While a long T1 removes the worry of thermal relaxation causing errors in a quantum computation, the lack of electron spins leaves very few practical avenues for nuclear polarization. One
possible method of nuclear polarization in fluorapatite is to employ the solid-state effect with deliberately introduced paramagnetic impurities. Of course, the introduction of paramagnetic impurities
has severe consequences for decoherence. If we assume a spin-1/2 electron impurity which obeys the
Bloch equations, we find that a nearby nucleus experiences a T2 of
 2  g 2  1 3 cos2 2
1
0
B
=
T1e (1 p2e ),
T2
4
2
r3

(3.38)

where r is the length and the polar angle of the vector connecting the nucleus to the impurity,
and pe is the electron polarization. If r is in units of nanometers and T1e in units of seconds, this
equation estimates that T2,ext 10 ps r6 /T1e (1 p2e ). At very low temperature, pe may be high
but T1e is long; for example, Cr5+ impurities in fluorapatite show a T1e of 0.1 s at 4 K [149]. If

fluorapatite were to be used for quantum computation, those qubits close to impurities would have
to be discarded. However, one can imagine an architecture in which these qubits are used only for
open-bath algorithmic cooling, as discussed in Sec. 2.2.2, while no impurities interact strongly with
the remainder of the qubit register.
Another approach for initialization might be to grow the fluorapatite crystal on a semiconductor
substrate, employ optical Overhauser effects to polarize the nuclei in that semiconductor, and allow
cross-polarization across the semiconductor/fluorapatite surface. Such an approach remains highly
speculative, and indeed it may be more promising to simply use the semiconductor substrate itself
for all the quantum computation, an approach we will consider in Sec. 3.4.

3.2.4

Measurement

In the end, the key reason for moving from liquids to solids is one of measurement. The prospect
of direct measurement of the magnetic moment of a single nucleus is bleak. The usual method of
measuring nuclear transitions in liquids is to directly measure the Faraday induction of precessing
nuclei with a macroscopic coil. The very small magnetic moment of a single nucleus can barely
1 Synthetic

crystals were grown at Stanford with the collaboration of Ian Fisher via the KF-flux growth techniques
described by Oishi and Kamiya [148]. These crystals showed a 19 F T1 of 3.5 minutes at room temperature and very
similar lineshape data to those of Lugt and Caspers [143].

82

Chapter 3. NMR Quantum Computers

compete with thermal noise processes in such a macroscopic object, even at very low temperatures.
Basic equations for the snr of nmr appear in Abragam [121]; a clear method of calculation was
presented by Hoult and Richards [150]. The most sensitive direct detection methods for magnetic
moments are mrfm and the use of squids. The latter unfortunately only work well at low magnetic
field, although they are capable of measuring small numbers of electron spins [151]. Recently, mrfm
has been successful at measuring the relaxation properties of a single electron spin [152]. Both of
these technologies are still several orders of magnitude shy of measuring single nuclear spin states,
although they may be useful for the detection of small nuclear ensembles.
Indirect measurement techniques in the solid state have, however, been capable of measuring
even single nuclear spins by optical means. The general idea is to couple the nucleus to an electron
spin with a strong hyperfine interaction, and then observe the effect of this hyperfine interaction on
the optical spectrum of the electron system. The optical measurement of the electron system is a
quantum non-demolition measurement, which allows many photons to carry the same information
about the eigenstate of the electron spin. The first such experiment in a solid-state system was
performed in pentacene molecules in p-terphenyl [153], in which the magnetic resonance of a single
pair of protons was detected after many hours. A more rapid measurement detecting the state of a
nuclear spin was performed more recently for a

13

C nucleus near a NV center in diamond [82].

These and other ultra-sensitive nuclear measurement technologies offer hope that some solidstate nmr architecture for a quantum computer may exist. However, those systems in which nuclei
have been sensitively detected are not the same as those systems in which nuclei have been well
polarized, and these are not the same as those where nuclear decoupling has been feasible. It can
be very frustrating to attempt to integrate these technologies. In Sec. 3.4, I overview one of the
attempts to do so.

3.2.5

Are Liquid State Quantum Computers Really Quantum?

As a final consideration of interest in moving from liquid-state nmr to solid-state nmr computers,
some researchers have asked whether the liquid-state computers are truly using quantum resources.
An important argument suggesting that they might not be was raised by Braunstein et al. [154],
where it was shown that the states described by the density operators of liquid-state nmr qubits are
separable, meaning they do not show entanglement. Without entanglement, it may be argued, the
state space taken by the n qubits can be simulated by n classical oscillators with a local realistic
description. The qubits could still see correlations in classical probability distributions, and these
probability distributions may be directly deduced through the large ensemble measurements used.

3.2 Liquids vs. Solids

83

One way of responding to this argument is to complain that the separability of the density operators is merely a mathematical observation, and if the nuclei in the molecules began in a pure state,
then surely the entangling gates employed for the quantum algorithms would generate entanglement.
This argument may seem sufficient if you interpret the density operator as indicating that molecules
actually in various pure states are distributed in the solution according to the distribution described
by the density operators diagonal representation. In this case you might say that some of the
molecules in the solution were in the correct ground state, and for these surely the computation
process was quantum.
However, a counter to this point of view is that it is overly realistic, ascribing real quantities
to quantum particles in a general quantum state before measurement. A unifying view upon which
much of quantum information theory relies is that there is no specific distinction between the kind
of probability in the density operator indicating our lack of knowledge about an existing feature of
the quantum system and the kind that describes quantum uncertainty, except for certain measures
such as the entropy, which is so high in liquid nmr that quantum uncertainty plays almost no role.
Definite progress in this debate would be obtained if a strictly classical model were constructed
to explain all aspects of solution nmr experiments. Attempts to do so, however, require either
an exponential loss of signal [155] as multiple entangling gates are applied, which is not observed
in experiments, or an exponentially growing number of classical hidden variables to describe the
dynamics [156]. The suggestion from this work is that it is not the state space that is important
for quantum computation, but rather the number of unitary implementations that can, in principle,
be implemented on that state space. This point of view is reflected by the Hilbert-space counting
argument I presented in Sec. 1.1.2, where I estimated the size of Hilbert space by quantifying the
number of operations one can apply in that space. This leads to the notion that the power of
quantum computation arises from the ability to traverse an exponentially large space in polynomial
time, but with much freedom as to where in that space you operate. In other words, the essential
ingredient for quantum computation may be the ability to entangle rather than entanglement itself.
This intriguing point of view, however, goes too far. Linden and Popescu [157] argued convincingly
that for some algorithms at least, some actual state entanglement is indeed a necessary but not
sufficient resource.
This line of debate was worthwhile for many researchers, as it forced us to think more carefully
about the true power of quantum computation. In the practical matter of building nmr quantum
computers, however, it has little bearing, since approaching scalable quantum computers will require
cooling qubits into states sufficiently far from maximal entropy that density operator entanglement

84

Chapter 3. NMR Quantum Computers

will be generated, removing the original point of contention. After all, everyone agrees that the
liquid-state nmr computers require an exponential resource to operate (the exponentially large
ensemble to maintain snr), and that a scalable architecture will have to be relieved of this resource.

3.3

Theory of Indirect Nuclear Interactions

Before proceeding to discuss a specific implementation for a solid-state nmr quantum computer, I
will briefly overview the basic theory for considering different physical means for coupling nuclei in
order to generate quantum logic. I have already discussed the dipolar coupling; this coupling is a
sufficient resource for universal logic in solid-state nmr quantum computers and will be the basis
for quantum logic in the architecture described in Sec. 3.4. However, this coupling is quite slow,
limiting quantum computers to nearest-neighbor interactions with entangling gate speeds slower than
a kilohertz. I have already alluded to a large range of other interactions between nuclei, including
the J-coupling in molecules, which is mediated by electrons in chemical bonds; the RudermanKittel-Kasuya-Yosida (rkky) interaction which is mediated by free conduction electrons in metallic
systems; and the Suhl-Nakamura interaction which is mediated by spin-waves in ferromagnets and
antiferromagnets. In this section, we describe such interactions in general, and consider their use in
possible architectures.

3.3.1

Effective Hamiltonians from Second Order Perturbation Theory

In this section we develop a general formalism for deriving an indirect nuclear interaction. In
standard perturbation theory (as opposed to, say, the Brillouin-Wigner expansion), the nth order
(n)

correction to the eigenstate |k i, which I notate |k i, is


(0)

|k i = |k i
X
hj | H1 |k i
(1)
|k i =
|j i
k j

(3.39)
(3.40)

j6=k

(2)

|k i =

X
i6=k
j6=k

|j i

X | hj | H1 |k i |2
hj | H1 |i i hi | H1 |k i 1
|k i
,
(k j )(k i )
2
(k j )2

(3.41)

j6=k

(2)

and so on. Note that the second term of |k i arises due to normalization, and is usually omitted.
The goal is to now calculate the second-order energy shift in the electron subsystem given a

3.3 Theory of Indirect Nuclear Interactions

85

general nuclear state. This may be done by leaving nuclear spin operators as operators, but expanding the electron terms out in the eigenenergy basis, and then allowing the eigenstates to develop
perturbative terms.
For this I want to generalize the second-order perturbative shift for the case of a superposition
of energy eigenstates, either coherent or incoherent. We thus write without approximation

(2)

(0)

(0)

(1)

Ek = ~hk |H1 |k i =

(1)

(1)

(0)

hk |H1 |k i + hk |H1 |k i
)
(2
(1)
(0)
(0)
(1)
n
o
|k ihk | + |k ihk |
(1)
= ~Tr H1
Tr H1 k ,
2

(3.42)

thereby defining a pure, first-order perturbed density operator. Now consider the generalization of
the density operator in the energy eigenstate basis:
(0)

X
ij

(0)

(0)

(1)

(0)

Mij |i ihj |,

(3.43)
(0)

(1)

|i ihj | + |i ihj |
2
ij


1X X
|k i hk | H1 |i i hj | |i i hj | H1 |k i hk |
=
Mij
+
2 ij
i k
j k

(1)

Mij

(3.44)
(3.45)

k6=i,j

1X
|i ihi | |j ihj |
Mij hj | H1 |i i
.
2 ij
i j

The symmetrization is performed to keep Hermitian. The crucial assumption, as alluded to above,
is that the coefficients of the superposition Mij do not change with the perturbation. From this it
follows that
E (2) = ~

X X

ij k6=i,j

Mij hj | H1 |k i hk | H1 |i i

1
2

1
1
+
i k
j k

1 X
hi | H1 |i i hj | H1 |j i
~
Mij hj | H1 |i i
.
2 ij
i j

(3.46)

Now we have a generalization which allows both coherent and incoherent mixtures. Such a generalization can be used to derive the original rkky results corresponding to an incoherent (thermal)
collection of electrons in a non-interacting electron gas [158].
For a diagonal wave packet, such as a thermal state with Pi = Mii = Z 1 exp(i ), we have

86

Chapter 3. NMR Quantum Computers

simply
E (2) = ~

X
k6=i

For a coherent wave packet |i =

k ck

Pi

|hi | H1 |k i|2
.
i k

(3.47)

|k i, we use Eq. (3.46) with Mij = ci cj .

We now use this formalism to calculate effective nuclear-coupling Hamiltonians due to fixed
electron spins.

3.3.2

One Electron Spin

The first, simplest example is the coupling due to a single electron in a spin-independent localized
spatial state coupled to several nuclei. For this we use the simple hyperfine interaction of Eq. (A.25)
on page 206. There are only two electron states to consider, |i and |i. We suppose a finite electron
polarization pe , so Eq. (3.47) becomes
+
+
X
S |i 1 pe h| Il+ S |ih| Im
S |i
E (2)
1 + pe h| Il S + |ih| Im
=
Al Am
+
~
2
e
2
e
l6=m

= pe

X Al Am 

+

= Heff . (3.48)
Il Im
+ Il+ Im
e

l<m

Essentially, pe Al Am /e is the second-order rate of probability that a flip-flop between one nuclear
spin and the electron is accompanied by a flip-flop between the electron and a second nucleus. This
interaction could be generated by an exciton in a semiconductor quantum dot or an unpaired electron
in a semiconductor impurity, effecting an optically controlled rkky interaction. It may easily be of
order 10 kHz. This simple calculation is only a precursor to more involved calculations for optically
controlled rkky interactions; a more useful calculation must treat the electrons which couple to the
nuclei in a many-body formalism.

3.3.3

Two Coupled Spins

The next example we will consider is two nuclei coupled by two coupled electron spins. Let us
assume a general bilinear interaction between the electron spins:
Hee =

XX
ij

ab b
Sia Jij
Sj ,

(3.49)

ab

ab
ab
where Jij
is a coupling tensor. For the nearest-neighbor exchange interaction, Jij
is a constant

J ab whenever i and j correspond to nearest neighbor electrons. For the dipolar coupling at high

3.3 Theory of Indirect Nuclear Interactions

field,
ab
Jij
=

87

0 2 2 1 3 cos2 ij ab
g B
( 3zab ).
3
2h
2rij

(3.50)

We now limit the discussions to two spins, i and j. The eigenstates depend on the symmetry
ab
of Jij
. However, because J ab is symmetric, there exists a rotation which diagonalizes it. In zero

magnetic field, we may simply work in the coordinate system in which J is diagonal. In high magnetic
field gB B0 ~|J|, any off-diagonal elements represent non-secular terms and may therefore be
neglected. In intermediate magnetic fields, gB B0 ~|J|, then a more complicated diagonalization
may be necessary. For the present discussion, let us presume a large magnetic field in which J ab is
diagonal. Then the eigenstates of the system are
|0 i = cos

|i + sin |i ,
2
2

0 = e sec
Jx + Jy + Jz
4
Jx + Jy Jz
+ =
4

1
| i = (|i |i) ,
2
1
|+ i = (|i + |i) ,
2

|4 i = cos |i sin |i ,
2
2

4 = e sec .

(3.51)
(3.52)
(3.53)
(3.54)

Here, tan = (Jy Jx )/4e . Of course, the ordering of states may change depending on the sign of
the components of J. We presume the electrons occupy the polarized ground state.

We now presume a contact hyperfine interaction of the form


H1 = A

X
j

Sj Ij .

(3.55)

There are several matrix elements to calculate. Between the ground and states,
h0 | H1 | i = A

X
k



A
+

+

Ik h0 | Sk | i = cos (I1 I2 ) sin (I1 I2 ) ,


2
2
2 2

and therefore
|h0 | H1 | i|2
=
0

A2
y
z
x
2J12 2(J12 + J12
) 8e sec

(I1+

I2+ )(I1

I2 )


1
+
+ 2

2
sin [(I1 I2 ) + (I1 I2 ) ] .
2

88

Chapter 3. NMR Quantum Computers

Between the ground state and state 4 the matrix element is


h0 | H1 |4 i = A
giving

X
k

Ik h0 | Sk |4 i =

A
sin [I1z + I2z ] ,
2

|h0 | H1 |4 i|2
A2 sin 2 z
2
=
[I1 + I2z ] ,
E0 E4
16e

To find the total interaction, we sum over all excited states, yielding

Heff =

 +

A2 (J x + J y )
A2 sin 2 z z
I1 I2 + I1 I2+ sin (I1+ I2+ + I1 I2 )
I1 I2 .
z
2
x
y
2
(4e sec J ) + (J + J )
8e

(3.56)

Let us estimate this numerically in a simple case. Assume cylindrical symmetry, = 0, since
anisotropic exchange interactions are difficult to engineer in non-ferromagnetic systems. Let us also
assume e J a . Then we simply have
Heff =

A2 J +
[I I + I1 I2+ ].
8e2 1 2

(3.57)

This is the interaction suggested by Kane [107]. It can be of order 10100 kHz at small magnetic
fields.

3.3.4

Lattice of Coupled Spins

It would be naive to assume the above result continues to apply to a lattice of N coupled electron
spins. If many electrons are coupled in an extended lattice, one must first calculate the spectrum of
eigenstates of this system, an intractable problem except in special cases. Having done this, these
eigenstates must be used as the virtual states.
One approximation that has been useful is to presume that the electrons are very close to their
known ground state. This results in the Suhl-Nakamura indirect nuclear interaction [159, 160]. As
an example of this, we consider the simplest case of an isotropic S = 1/2 Heisenberg ferromagnet
ab
in a magnetic field, which has a simple non-degenerate ground state2 . Hence we use Jij
= Jij ab .

In the ground state, all electron spins are antiparallel to the field. (In zero field, we note that all N
symmetric spin states form the degenerate ground state manifold.)
The basic assumption is that deviations from the ground state, if sufficiently small, may be
2 A more complicated example is in the frustrated antiferromagnet CeP, for which calculations of the Suhl-Nakamura
interaction appear in Ref. 138.

3.3 Theory of Indirect Nuclear Interactions

89

mapped onto a harmonic oscillator. A common way to do this is with the Holstein-Primakov
transformation,
q
1 aj aj aj ,
q
= aj 1 aj aj ,

Sj =

(3.58)

Sj+

(3.59)

Sjz = S + aj aj ,

(3.60)

which models deviations from the polarized ground state as bosonic excitations. The square root
terms are needed to correctly replicate spin commutation relations. In an approximation linear in
ground-state deviations, however, the square root terms may be ignored and the exchange Hamiltonian may be approximated as
Hee (N e S + JS 2 N 2 ) + Se

nj + S

X
ij

Jij (ni + nj ai aj ai aj ) + O(n2 ).

(3.61)

Collective modes across the periodic crystal are then defined by


1 X iqrj
e
aj .
aq =
N j

(3.62)

The q vectors are reciprocal lattice vectors; that is


q=

2 X
ak al
mj
,
N j
aj (ak al )

(3.63)

where the mj s are integers and aj s are lattice vectors. Using these modes and dropping the constant
ground state energy term, the first order deviation Hamiltonian is written
Hee
where
J(q) =

[e + J(q)]aq aq ,

(3.64)


X
1 cos[(rm rn ) q] Jmn ,

(3.65)

which is independent of n for a periodic crystal.


We now see the essence of this approximation. Near the bottom of the potential created by
the exchange interaction and the magnetic field, we treat the spectrum of states as harmonic with

90

Chapter 3. NMR Quantum Computers

frequencies
q = e + J(q).
We model N such states, for each of the reciprocal lattice vectors q. These are spin-wave modes.
The state of nq spin-waves of wave vector q has energy n[q + J(q)] above the ground state energy,
and spin angular momentum
hnq | ~

X
j

Sjz |nq i = (nq N S)~.

(3.66)

Intrinsic to this derivation is that hnq i 1. In the present problem of virtual excitation of spin
waves in the ground state, this approximation is valid.

Now the flip-flop terms of the contact hyperfine interaction may be written, in the same linear
approximation, as
Hen =

AX +
I a + Ij aj ,
2 j j j

(3.67)

which we will revisit as the spin-boson model discussed in Sec. 6.3.1. In the second-order approach
to effective nuclear couplings and in the linear approximation for spin-waves, Hen cannot access
states of more than one spin-wave. The effective Hamiltonian for ground state electron spins may
therefore be written
Heff =

X |h0| H1 aq |0i|2
q

= A2
=

X
j6=l

q1

1 X + iq(rl rj )
Ij Il e
N

(3.68)
(3.69)

jl

D(rj rl )[Ij+ Il + Ij Il+ ].

(3.70)

For example, consider a one-dimensional chain with lattice constant a and nearest-neighbor couplings
only. Then q = 2m/aN , J(q) = 2J[1 cos(2m/N )], and
D(na) = A2

N 1
1 X
cos(2mn/N )
.
N m=0 e + 2J[1 cos(2m/N )]

(3.71)

Most of the contribution to the sum comes from m N , so we expand the denominator to lowest

3.4 The All-Silicon Approach

91

order in m/N and treat this small number as a continuous variable x to find
2

D(na) A

cos(2nx)
A2 en e /J

dx
.
e + 4 2 Jx2
4
Je

(3.72)

This is a very different result from Eq. (3.57)! We see that the presence of collective modes fundamentally affects the scaling of the effective Hamiltonian. In particular, the strength of the indirect
coupling between nuclei spaced by n lattice constants no longer grows monotonically with J. For
J > n2 the indirect coupling decreases with increasing J, for in this limit the lattice becomes too
strongly coupled to allow the interaction to take place.
A similar approximation in two dimensions with square symmetry leads to
D(r)

A2
K0
2J

 r 
r e
,
a J

(3.73)

where K0 (z) is a Bessel function of the second kind. Again, the scaling with respect to J is completely
different. The asymptotic limits for r a are
p
A2 exp((r/a) e /J)

p
D(r)
,
2 2
Je (r/a)
!
r
A2
a J
D(r)
ln
,
2J
r e

e J

(3.74)

e J.

(3.75)

We see a similar dependence on J only in the high field case, where it is the applied field that mostly
determines the energy of the virtual excitation.

3.4

The All-Silicon Approach

The above provides only a brief glance at indirect nuclear interactions. More detailed calculations
have provided the basis for most solid-state nmr proposals for quantum computing. I have discussed
these various proposals in previous sections. All of them, however, require a hyperfine coupling to
electrons, and these electrons in turn may limit the coherence of the nuclei. An architecture with no
hyperfine field during the computation, but still with the capability for initialization and measurement, may therefore be desirable. Quantum logic is still possible without indirect interactions by
sufficient control over the direct dipolar Hamiltonian, as I will discuss in Chapter 4. The question
remains as to what material might allow such an architecture.
Silicon is best known for its use in classical computing microchips. It turns out that it has many

92

Chapter 3. NMR Quantum Computers

qualities which are useful for quantum computing, but these qualities have very little to do with
the reasons we use silicon for microchips! In this section, based on Ref. 161, I detail a complete
proposal for using silicon for quantum computing.
The possibility of optical pumping for initialization is an important reason for considering a silicon implementation. In addition, from an engineering perspective, there are many other advantages.
Perhaps the most important motivation is that the crystal growth and processing technologies for
silicon are highly matured. Although methods exist for the growth of small single crystals of fluorapatite, for example, the technology to incorporate this material in micrometer-scale devices has not
yet been developed. With its wealth of fabrication technologies and its favorable material properties,
pure silicon has proven to be an excellent material for micrometer-scale mechanical oscillators, which
will be important when we discuss force detection in Sec. 3.4.3.
Another advantage of silicon is that, unlike group-iii-v semiconductors, the family of stable
nuclear isotopes is quite simple: 95.33% of natural silicon is
4.67% is

29

28

Si or

30

Si, which are both spin-0, and

Si, which is spin-1/2, perfect for the qubit. Thus, silicon is well suited for nuclear-spin

isotope engineering [162].


A crucial element of the all-silicon proposal we present here is that, following optical polarization,
there is no need for electrons for the control or measurement of nuclear spins, unlike a number of
other silicon-based proposals which require bound donor electrons and (usually) nanometer-scale
metallic gates [104, 107, 163, 164]. The absence of any spurious spins, nuclear or electronic, should
leave the

29

Si nuclei well decoupled from their environment, a hypothesis whose experimental test

will be the subject of the following two chapters.

3.4.1

Fabrication

Of course, the diamond crystal structure of silicon is not one-dimensional, as in the case of fluorapatite; especially if one considers the randomly placed 29 Si nuclei in a sample with natural abundance.
To make an effective crystal lattice quantum computer, we must somehow fabricate long, straight
atomic chains of the active

29

Si isotopes in a single-crystal matrix of the other, spin-0 isotopes. We

emphasize that, considering the wide range of silicon processing methods available, there may be
more than one way to fabricate such a system; in the following we consider one particular method.
We note that this procedure is similar to one already proposed for a Kane-like implementation [165].
We start with a nearly isotopically pure

28

Si(111) (< 1%

29

Si) silicon-on-oxide wafer with the

surface miscut by an angle of about 1 towards (112).


The buried oxide layer is needed to release
the silicon structure to make a free bridge of silicon. This limits the active area of qubits within the

3.4 The All-Silicon Approach

93

Dy micromagnet

D
t

L
s
w
Optical

fiber

Figure 3.6: A schematic for an all-silicon quantum computer. The figure shows the
integrated micromagnet and bridge structure needed for distinguishing qubits and for
read-out. The bridge has length l = 300 m, width w = 4 m, and thickness t = 0.25 m.
The micromagnet has D = 400 m, L = 4 m, and W = 10 m, and produces a field
gradient of B z /z = 1.4 T /m, uniform over a 100 m by 0.2 m region inside the
bridge. The insert shows the structure of the silicon matrix and the terrace edge. The
darkened spheres represent the 29 Si nuclei, which preferentially bind at the edge of the
Si step.

homogeneous region of the magnetic field gradient, and may also be used as a mechanical oscillator
for mrfm detection, as will be discussed in Sec. 3.4.3. It has been demonstrated [166, 167] that such
a miscut, followed by a simple multi-step annealing sequence, leads to atomically straight terracesteps on the vicinal Si(111)77 surface. These steps run along the (110) direction for up to 2 104
lattice sites, with an average terrace width of 15 nm. The lack of kinks in these steps is due to the
high energy cost of interrupting or misplacing the 77 domains.
This miscut wafer is processed in a multi-chamber molecular beam epitaxy (mbe) machine
equipped with a scanning tunneling microscope (stm) and a pre-heating stage. The stm is used for
verification of the step edges; afterwards, the wafer is transferred to the growth chamber to form
atomic chains of

29

Si using the step-flow mode at temperatures T 600C [168]. Arrival of

isotopes at the substrate surface induces a surface transition from 77 to 11 [169], so that

29

Si

29

Si

isotopes travel to the step edges and form atomically straight lines. We terminate the evaporation

94

of
28

Chapter 3. NMR Quantum Computers

29

Si when the atomic chains are one atom wide and run completely along the step. Accidental

Si or

30

Si isotopes in the

29

Si source will lead to chain interruptions; this will reduce the effective

ensemble size unless the impurity level of the source is much less than 1/n, where n is the number
of qubits. A 28 Si capping layer of 15 nm is grown on top of the

29

Si chains and we repeat the same

sequence of multi-step annealing, 29 Si chain growth, and capping to produce replicas of parallel 29 Si
chains. The last step in the growth process is the capping of ensembles of 29 Si chains by a thick 28 Si
layer. The diffusion of the buried silicon chains is negligible at the 600 C growth temperature,
according to the measurement of the silicon self-diffusion constant by Bracht et al. [170].
Initial steps in this fabrication procedure have been carried out at Stanford University and at
Keio University [171]. In particular, the researchers at Keio have made progress elucidating the
detailed crystal structure of the terrace step, so that the addition of single silicon atoms can be
confirmed by stm. Figure 3.7 shows a single chain of silicon atoms grown in the vicinity of one of
these steps by the Keio group.
The atomic chains resulting from this procedure lead to a highly one-dimensional structure; on
average, we find 1 15 nm/2
A = 75. In particular, (1) 106 , so the error in the nearestneighbor gate operation is tolerable. We also note that the atomic chains, while very regular, are
not straight rows of atoms once incorporated into the diamond structure. As shown in Fig. 3.6,
the atoms zig-zag3 with polar angle satisfying cos2 ij,(i+1)j = 2/3, resulting in a nearest-neighbor

dipolar coupling coefficient D = 2 0.4 kHz.


The magnetic field gradient is obtained with a dysprosium micromagnet. This is fabricated using
evaporative deposition and lift-off. Its range of operation is limited by the bridge structure, which is
created with e-beam lithography, plasma etching, and selective chemical etching of the oxide layer.
Calculations similar to those in Ref. 141 show that the magnetic field gradient due to the micromagnet is B z /z = 1.4 T/m, which is persistent over the thickness of the bridge and is superposed
with a large homogeneous field B0 of 7 T. The distance in the z-direction between two

29

Si nuclei

in an atomic chain, which we notate a, is 1.9


A, so the gradient leads to a qubit-qubit frequency
difference of j+1 j = aBz /z = 2 2 kHz. The active region is a 100 m by 0.2 m area in

the center of the bridge, containing N = 105 chains persisting over the bridge thickness. This active

region is sufficiently small and the magnetic field sufficiently homogeneous that all equivalent qubits
in an atomic plane lie within a bandwidth upper bounded by 0.6 kHz, which is substantially less than
the qubit frequency separation of 2 kHz. However, it is of comparable order to the dipolar coupling
between qubits, a difficulty which could complicate decoupling and recoupling pulse sequences in
this architecture.
3 This

zig-zag means the 1/m3 dependence in Eq. (3.35) is not precisely correct for small qubit displacements m.

3.4 The All-Silicon Approach

(a)
b

95

(b)
d

Figure 3.7: Images of single atom chain growth in silicon. A stm image appears on the left with
corresponding ball-and-stick atomic model on the right; the arrows on the bottom compare critical
atomic locations between the two images. The images labelled (a) show a terrace-step on the
Si(111)77 surface before deposition of additional silicon; this structure would be made from spin-0
28
Si. The images labelled (b) show that deposited silicon atoms form a straight atomic chain on the
terrace step; these extra atoms would be the spin-1/2 29 Si nuclei. This figure is taken from Ref. 171,
which explains the experimental conditions and the modelling in more detail.

96

Chapter 3. NMR Quantum Computers

3.4.2

Initialization

The ability to perform optical pumping in silicon [136] aids the initialization problem, though such
techniques are unlikely to yield 100% polarization. Fortunately, the techniques of algorithmic cooling
and the use of pseudo-pure states, both discussed in Sec. 2.2.2, may be employed. These techniques
introduce scaling limitations which are only restrictive if the physical polarization provided by optical
pumping is too low.
A solution of simple rate equations for the optical Overhauser effect yields the steady state
nuclear polarization in the strong-pumping limit as
hI z i

T1e G+ G
,
+ T1e
2

(3.76)

where is the lifetime of the photoexcited electrons and G is the transition probability for generating a spin mz = 1/2 with circularly polarized light.
The indirect bandgap of silicon limits the nuclear polarization for two reasons. First, the two
transition probabilities G are affected by the band structure; the indirect bandgap leads to a
reduced value of (G+ G ) relative to direct bandgap semiconductors. Second, the ratio of the
conduction electron spin-lattice relaxation time T1e to its recombination time is small, since the
indirect processes necessary for electron recombination lead to a long . As a result, low-field (1 G)
experiments at 77 K [136, 172] have not exceeded 0.1% nuclear polarizations. Improved nuclear polarization in silicon may be observable in higher magnetic fields (10 T) and lower temperatures
(1 K); in these regimes, the transition probabilities G may be more favorable, and T1e is substantially longer [173]. A small sample will also help, since rapid recombination of electrons via surface
states can reduce .
Progress in optical pumping has been made at Stanford. Verhulst [174] has shown optically
pumped polarizations of order p = 0.25% using a titanium:sapphire laser tuned to = 1024 nm
with an optical power of 300 mW applied for thirty minutes at T = 4 K and B0 = 7 T. This is the
largest polarization experimentally obtained in silicon to date, although it is still quite small. It is
in part limited by the long relaxation time of silicon; in the dark, T1 is about 1200 hours, and it
reduces to 4060 hours in the presence of the optically pumped electron-spin bath. Irradiation for
longer times would therefore improve the polarization; also, there is evidence that excitation above
the band-gap leads to higher polarizations, but only for a small number of nuclei within the small
skin-depth of the light. Probably the most important advance for optical pumping will be improved
understanding of the role of impurities and subsequent optimization of the impurity content to

3.4 The All-Silicon Approach

97

maximize nuclear polarization. Verhulst [174] speculates that polarizations as high as 10% could
be obtained with these improvements. Improved nuclear polarization could also be obtained by
injecting a spin-polarized current, perhaps from an epitaxially grown, optically pumped SiGe alloy
layer.
Optical pumping in silicon is not nearly as well studied as GaAs, where the direct bandgap leads
to larger enhancements obtained in shorter times. I will briefly discuss the current understanding
of optical polarization of nuclei in GaAs, and how it may relate to silicon, in Sec. 6.3.3.

3.4.3

Measurement

The measurement of the small ensemble of nuclear spins in the homogeneous region of the magnetic
field gradient is the most challenging part of the proposal.
Magnetic Resonance Force Microscopy
The presence of a large magnetic field gradient provides a natural means for performing mrfm
on a magnetization ~M z , since this technique is sensitive to the gradient force given by F z =
~M z B z /z. Here silicon once again proves the most advantageous material, as its structural
properties are optimal for micromechanical oscillators [175].
The experiment is performed in high vacuum (< 105 torr) and at low temperatures (4 K). A
coil is used to generate the rf pulses for logic operations and decoupling sequences; it also generates
the continuous-wave radiation for readout. An optical fiber-based displacement sensor is used to
monitor deflection of the bridge using interferometry. Sub-
Angstrom oscillations can be detected;
larger oscillations can be damped with active feedback which avoids additional broadening while
maintaining high sensitivity [176].
Readout is performed using cyclic adiabatic inversion [121], which modulates the magnetization
of a plane of nuclei at a frequency near or on resonance with the bridge. The spins of resonant
frequency i are irradiated with the rf field B x = 2B1 cos{i t (/m ) cos(m t)}, where m is
the modulation frequency chosen to be near the resonance of the bridge oscillations, and is the
frequency excursion, which should be much smaller than the frequency separation between qubits
[102]. The z component of the ith planes magnetization is deduced from the phase of the resulting
bridge oscillation. Simultaneous detection of signals from multiple planes is possible if the different
planes to be measured are driven at distinct modulation frequencies m .
The force resolution for mrfm is limited by thermal fluctuations of the mechanical oscillator [177].

Force resolutions of 5.6 1018 N/ Hz have been reported for single-crystal silicon cantilevers at

98

Chapter 3. NMR Quantum Computers

4 K [175]. The thermal noise threshold of the bridge structure in Fig. 3.6 is estimated to be

1.2 1017 N/ Hz based on a lumped harmonic oscillator model4 and assuming a modest quality
factor Q of 104 [178]. Using a gradient of 1.4 T/m and the gyromagnetic ratio for

29

Si, this

force sensitivity corresponds to a snr of one when measuring an ensemble of about 200 nuclei. An
ensemble of approximately 105 nuclei may fit into the bridge, leading to a total sensitivity of 0.002.
The number of measurable qubits will depend on this sensitivity and the degree of polarization as
shown in Fig. 2.1.
A possible concern is that thermal fluctuations of the micromechanical bridge in the presence
of a large field gradient will lead to decoherence due to spectral diffusion. By directly combining
Eq. (3.30) with the thermal noise statistics of a high-sensitivity mechanical oscillator [177], the T2
timescale due to the bridges thermal drift is T2c = kc Qa2 / 2 kB T 25 s. Active feedback
stabilization is expected to increase this timescale by as many as four orders of magnitude, bringing
it close to the lengthy T1 timescale5 .

Optical detection
The implementation of mrfm requires a substantial amount of additional engineering into the architecture. An optical detection technique, described here and in Ref. 179, not only has the potential
to be more sensitive, but also to remove the constraints imposed by the need to generate a high-Q
mechanical oscillator. The cost is the requirement of an impurity bearing an unpaired electron spin.
The desired impurity is a

31

P nucleus, which binds a single donor electron at low temperature. The

spin of this donor electron is strongly hyperfine-coupled to the 31 P nuclear spin. The effective coherence time of the

31

P nucleus and electron spin pair might be no longer than the decoherence time of

the electron spin. Recent measurements show that if this impurity is located in isotopically depleted
silicon, its own spin coherence time is quite long [87], but still short in comparison to what may
be expected for isolated nuclei. This nucleus should therefore only be used for measurement (and,
possibly, initialization); it may be coupled and decoupled from active

29

Si nuclei using standard

heteronuclear coupling techniques in nmr.


The optical detection technique employs the ability of the

31

P neutral donor to bind a free

exciton, forming a bound exciton (be) state. When a be decays radiatively to the neutral impurity
state, its linewidth is characteristically narrow due to the localization of the exciton [180]. The
hyperfine interaction between the 31 P nucleus and the neutral-impurity bound electron is sufficiently
4 This

model yields an estimated spring constant of k 0.0042 N/m and a resonance frequency c /2 23 kHz.
similar calculation allows an estimate for the T2 due to motion fluctuations caused by lattice phonons, which
is negligible.
5A

3.4 The All-Silicon Approach

99

ms=0 mh

mI
1/2
1/2

+3/2
+1/2
-1/2
-3/2

1/2
1/2

ms

b
+1/2
-1/2
-1/2
+1/2

+1/2
-1/2

(P0,X)

P0

Hyperfine Splitting
Figure 3.8: Energy diagram for the neutral donor (P0 ) and its bound exciton (P0 ,X) in
a magnetic field. The (P0 ,X) state is populated via capture of a resonantly excited free
exciton. The 31 P nuclear state can be determined by the energy difference of a and b.

strong that the nuclear state of the impurity can be determined via the energy of the resulting be
photoluminescence (pl).
In a magnetic field, the ground state of the 31 P be, notated (P0 ,X), is split into four hole Zeeman
levels. The ground state of the neutral donor, P0 , is split into two electron Zeeman levels [181]. These
levels are illustrated in Fig. 3.8. We assume a large applied magnetic field such that the electron/hole
Zeeman interaction is much greater than the hyperfine interaction.
The P0 state is described by effective mass theory (emt), where the total wavefunction e (r)
of the donor electron is given by the product of the emt envelope function F (r) and the Bloch
wavefunction (r) = u(r)eikr . Both the Bloch and envelope functions are s-like, so the dominant
part in the hyperfine interaction is the Fermi contact term, given by Eq. (A.25). When an exciton
is bound to the neutral-donor site, the two electrons form an antisymmetric spin singlet state.
Consequently, the hyperfine splitting of the be is determined only by the spin of the bound hole.
Since the hole Bloch function is p-like, the Fermi contact term is negligible, and assuming that the
envelope function is s-like, the orbital and dipolar terms will be much smaller than the P0 contact
hyperfine splitting. Thus, the energy difference between the transitions a and b of the (P0 ,X) pl,
shown in Fig. 3.8, is determined entirely by the hyperfine splitting of the P0 state, which we notate
HF,P0 . The electron density at the

31

P (r = 0) site has been determined via electron spin

100

Chapter 3. NMR Quantum Computers

resonance to be 0.44 1024 cm3 , corresponding to a HF,P0 /2 of 60 MHz [182]. This splitting
has also been calculated with reasonable accuracy using a corrected envelope function to account
for the discrepancy in the observed and emt ionization energies [183, 184].
The (P0 ,X) state in Si decays primarily via a non-radiative Auger process with a lifetime of
300 ns. It may also decay radiatively with a lifetime of 2 ms [185]. If the pl linewidth of a single
31

P donor impurity is lifetime broadened, it would be on the order of 3 MHz, which is much

smaller than the hyperfine splitting of 60 MHz. Experimentally, the pl linewidth from an ensemble
of

31

P impurities was measured to be less than 150 MHz at 2 K, which includes an inhomogeneous

broadening effect and was limited by spectrometer resolution [180]. At worst, then, optical detection
requires resolving a 60 MHz shift under a linewidth of 150 MHz.
A critical requirement for sensitive measurement is a large number of emitted photons. After
the frequent Auger recombination events, the electron/hole recombination energy is imparted to the
second electron, which is ionized. To ensure a fast recapture process of the donor electron, one can
optically excite conduction electrons and free excitons which will be captured at the neutral-donor
site. The be will relax thermally within its lifetime [181], and in a magnetic field of 10 T at 4 K,
the probability for it to occupy the lowest Zeeman level is 0.8. Thus, approximately 400 photons/s
will be emitted at the desired transitions a and b in Fig. 3.8.
However, the extraction of an emitted photon out of the high-refractive-index Si substrate as a
well-collimated beam for optical detection is difficult. For this purpose, one can incorporate a planar
distributed Bragg reflector dbr cavity, at the center of which a 31 P impurity is embedded. The cavity
modifies the radiation pattern and concentrates the emitted power in the normal direction. With
high and low dbr refractive indices n1 = 3.0 and n2 = 1.5, the output coupling efficiency (-factor)
into a beam emitted in the normal direction can be as high as 0.8 for a random dipole orientation.
This efficiency is achieved at the cost of a radiative decay rate decreased by a factor of 3. The overall
number of pl photons available for optical detection would then be 400 0.8/3 100 photons/s. A
more advanced solution to the photon collection problem is to employ a two-dimensional photonic
crystal structure that can simultaneously enhance the output coupling efficiency and the radiative
decay rate. A detailed analysis based on the finite-difference-time-domain method predicts that
an optical mode volume of 0.5(/n)3 , a Q value of 30,000, a spontaneous emission decay rate
enhancement (Purcell) factor of 100, and a -factor of 1 can be achieved [186]. The expected photon
flux in this case is 4 104 photons/s.
In order to find the measurement time needed to determine the nuclear spin state, we calculate
the signal-to-noise ratio for direct frequency detection. The be pl is collected and sent to the

3.4 The All-Silicon Approach

101

input of a Mach-Zender interferometer. The photons in the two outputs of the interferometer are
detected by single photon detectors. The photocurrents are subtracted, and the resulting current is
time-integrated.
Assuming an input state with mean optical frequency op and a Lorentzian spectral lineshape
with full-width-half-max , the average integrated single-photon current is proportional to
hIi = e /2 cos(op ).

(3.77)

The variance of the current is proportional to


hI 2 i = hI 2 i hIi2 = 1 e cos2 (op ).

(3.78)

The two input frequencies of the be pl, corresponding to the two nuclear spin states, are =
op HF,P0 /2. The difference in the integrated current between the nuclear states is the signal,
so the amplitude signal-to-noise ratio for such a frequency detection scheme is given by
v
u
u sin
S
=t
N

1
1
0
0
2 HF,P sin 2 HF,P
1
2

cos 2 HF,P0 + e 1

2

N,

(3.79)

in which the interferometer is biased such that op = (m + 12 ) and N is the total number of
photons collected. The maximum signal-to-noise ratio is 0.29 for one photon for a pl linewidth of
150 MHz and a delay time = 2 ns. Assuming 100 photons per second are collected, the minimum
integration time needed for a signal-to-noise ratio of unity is approximately 0.1 sec. If we assume
an enhanced spontaneous emission decay rate with a 2d photonic structure mentioned above and a
collection efficiency of 0.5, the minimum integration time is reduced to approximately 103 s.
Any practical detection scheme will suffer from finite detector efficiency and dark count rates.
However, negligible dark count rates and a detection efficiency of 0.4 can be obtained at = 1.1 m
with a superconducting transition edge sensor [187]. The effect of finite detector efficiency d is
to decrease the signal-to-noise ratio by d , extending the measurement time. Spurious noise from
interferometer drift may be averaged away by using cyclic adiabatic inversion on the nucleus, just as
in the mrfm case, with lock-in amplification of the interferometer output modulated at frequency
m .
Sufficient integration of the luminescence signal measures the nuclear spin state only if that state
is stable over a sufficient number of excitation/luminescence cycles. The T1 of isolated

31

P nuclei,

however, exceeds 10 hours at 1.25 K and 0.8 T [173]. Cross relaxation with the phosphorous-bound

102

Chapter 3. NMR Quantum Computers

electron, induced by the hyperfine interaction with the emission of a phonon, is also a slow process.
It was measured to be 5 hours at 1.25 K and 0.8 T and should theoretically scale to 30 seconds at 4 K
and 10 T [173]. Since the P0 ground state only lasts nanoseconds before each optical re-excitation
in this scheme, these already long equilibrium relaxation times are negligible. The nucleus may only
be appreciably destabilized by the rapid optical excitations we introduce. Calculations for T1 due
to optical excitations will have to wait for Sec. 7.1.2. The result may be readily guessed without the
detailed calculation, however. It must be a second-order perturbative effect, for which the probability
scales as the matrix element over the energy denominator, squared. In this case the matrix element
is the hyperfine coupling and the energy denominator is the energy of an electron/nuclear spin flipflop, which is dominated by the electron Zeeman energy. Therefore, the probability of an event must
be roughly (~HF,P0 /gB B0 )2 . For reasonable magnetic fields, this probability is on the order of
107 .
It is unfortunate that the predominant decay mechanism is non-radiative, since each Auger
process increases the probability of nuclear randomization without providing a signal photon. After
2 106 excitations, at which point the probability of nuclear randomization exceeds about 1/10, we

can only expect to have collected and detected 25 photons with a dbr planar cavity or 5 103 with

a 2d photonic crystal, respectively. This still yields a usable signal-to-noise ratio of 2 or 20 for the
respective geometries. Thus, we expect the measurement-induced lifetime of the single nuclear spin
to be long enough to allow for the measurement of the nuclear spin state.

3.5

Quantum-Dot NMR Quantum Computers

Although initialization and measurement are possible using only silicon, it is unclear whether such
an approach can be made scalable. Silicons poor optical efficiency has led to two difficulties: optical
polarization of nuclei takes a very long time (on the order of 60 hours) and the measurement of
single nuclear spins as discussed in the previous sections, when including the realities of detector
and overall collection effiencies, might take seconds to minutes. Add the time to swap the desired
bit to a nearby

31

P impurity for measurement, and we see that this technique cannot be efficiently

used to measure ancilla qubits during error correction.

3.5.1

Measurement

What if the semiconductor quantum computer were more optically efficient? For example, suppose
the nuclei are contained inside an optically active quantum dot. Could individual nuclei be more

3.5 Quantum-Dot NMR Quantum Computers

103

rapidly measured? The answer comes from a closer look at Eq. (3.79). In a quantum dot, the
hyperfine splitting to a single nucleus is substantially reduced and the optical linewidth, , is
substantially increased. For any /, there exists some which maximizes the signal; for large
/ this is almost exactly 2/. Therefore, for large /, the signal to noise ratio scales as
S

2

N.
N

0.5
(e/)2 1

(3.80)

The number of photons collected N scales as r B, where B is the bandwidth of the measurement
(that is, the inverse of the integration time). If the emitter is nearly lifetime limited, the radiative

lifetime r 1 , and in this case the total snr scales as r B. In this sense, if you decrease
by a factor f by moving to a lifetime-limited quantum dot, the optical snr is preserved if the
radiative lifetime decreases by a factor of more than f 2 .
Suppose the quantum dot is a self-assembled InAs semiconductor quantum dot or a GaAs quantum dot defined by edge fluctuations of a GaAs/AlGaAs quantum well. Then the single-nucleus
hyperfine splitting of the exciton linewidth is on the order of a megahertz; it has been reduced
from silicon by f = 60. The radiative lifetime, however, is very short, typically several hundred
picoseconds; this is a decrease over silicons 2 ms optical lifetime by seven orders of magnitude, so it
is decreased by about f 4 . Therefore, even though the linewidth of the quantum dot is much, much
larger and the hyperfine coupling is smaller, the greatly increased number of photons means that
the optical snr is actually increased by a factor of about f = 60.

3.5.2

Initialization

What about initialization? Gammon et al. [188] have shown that the 69 Ga and 75 As nuclear spins in
single GaAs quantum dots defined by edge fluctuations of GaAs/AlAs quantum wells are up to 60%
polarized in about 3 seconds by application of above-band circularly polarized light. This is over 200
times better polarization than silicon, and it occurs about 50,000 times faster. The theory presented
by Gammons group to explain the phenomenon [189] suggests that a similar effect should occur
in self-assembled InAs dots. A recent experimental study in charge-tunable dots suggest that the
same effect can occur in charged quantum dots [76]. The nuclear polarization is optically detected
rather quickly by observing the effective magnetic field seen by the quantum dot exciton due to the
hyperfine-coupled nuclei. This effective magnetic field enhances or detracts from the Zeeman shift
of the optical exciton spectrum in a magnetic field; this shift is called the Overhauser shift. The
Overhauser shift is comparable to the linewidth when the nuclei are well polarized, and therefore

104

Chapter 3. NMR Quantum Computers

the nuclear polarization can be spectrally deduced, without the interferometric techniques described
above.
Recent experiments at Stanford have suggested that these effects are more difficult to observe
in InAs quantum dots. Certainly a shift similar to the Overhauser shift is easily observed in InAs
quantum dots, but preliminary experiments suggest that this shift is not nuclear in origin, since the
application of resonant rf which should depolarize the nuclei has no effect on the shift, and it appears
on a time scale faster than a few milliseconds. There are two key differences between the GaAs dots
used by Gammon [Naval Research Lab (nrl) dots] and the InAs dots used at Stanford (Stanford
dots). First, Stanford dots are isolated by etching lithographic mesa structures in which a small
number of dots are contained. Small mesa structures introduce surface states close to quantum dots,
in which electron charges may accumulate. These electron charges are suspected to contribute to the
linewidth of the quantum dot spectra, and their spin polarization may explain the false Overhauser
shifts. In contrast, the nrl dots are isolated by small holes in a metal masking layer. Also, the nrl
dots are in close proximity to a GaAs/AlAs quantum well, where the continuum of exciton states
exhibit strong Overhauser effects. The analog in the Stanford dots is the InAs wetting layer, which
is less regular. If the continuum of states in the quantum well of the nmr dots plays a crucial role,
the polarization available in InAs dots may be quite different. Certainly, Gammon observes very
different behavior from dot to dot; the variance of Overhauser shifts is quite large, which may be a
result of fluctuating dot size or of local changes to the structure of the quantum well in the vicinity
of the dot.
Clearly, more study is needed to fully understand nuclear polarization and detection in quantum
dots, but already experimental evidence favors these systems over bulk silicon and 31 P impurities in
silicon.
Quantum dots based on group-iii-v semiconductors, however, have a distinct disadvantage over
silicon. Namely, all isotopes of nuclei in these semiconductors have non-zero spin, and in particular
they all have spin I > 1/2. Obviously this will not allow the arrangement of atomic chains of active
nuclei in a spin-less substrate as is possible in silicon. Since one-dimensional structures are not
available, the ensemble-based architecture must be abandoned, and instead a planar architecture
with a two-dimensional gradient may be considered. However, spurious couplings throughout the
substrate will always place high demands on decoupling pulse sequences. Further, all nuclei have
quadrupole moments, allowing these nuclei to couple directly to electric field gradients created by
strain inside the quantum dot. These electric field gradients fluctuate due to charging noise and
phonons. Even if decoupling is well-performed, intrinsic decoherence inside these dots will be far

3.5 Quantum-Dot NMR Quantum Computers

105

worse than in the I = 1/2 case.


Two possibilities have been considered. One is to grow a monolayer of GaP inside a GaAs quantum dot using mbe techniques. Only the

31

P nuclei are considered optically active; the remaining

nuclei are decoupled. Alternatively, one can move to quantum dots based on group-ii-vi semiconductors, combining the advantages of isotope engineering with direct bandgap semiconductors. Both
of these avenues require further research.

3.5.3

Scalability

Another concern is the limited size of the quantum dot. Depending on its size, it may only have 104
to 105 active nuclei, a number which will further decrease once qubits are discarded for initialization
techniques. It is tempting to ask for use of all the nuclei in the substrate, but the constraints of
the magnetic field gradient discussed in Sec. 3.2.2 must be remembered. A two-dimensional field
gradient may only have a finite extent, as shown in Fig. 3.3. The angle must be chosen so that
bG sin is sufficiently large to resolve nuclei in the y-direction, an impossibility if the gradient is
less than or on the same order as dipolar couplings, and hence sin D/bG. But distinguishing
all nuclei in the lattice requires bG cos > mbG sin , so for small , m must be upper bounded by
about bG/D. Nuclei outside of this active area must be ignored, for we cannot distinguish them.
The gradient must be designed to be sufficiently large that dipolar couplings of order D/m3 may be
ignored; otherwise nuclei outside the active area will inevitably couple into the quantum computer,
causing decoherence.
Further, shining light on the quantum dot inevitably affects every nucleus inside it, making the
measurement or initialization of ancilla qubits during a quantum computation quite questionable.
A scalable, fault-tolerant architecture using nuclei in quantum dots will inevitably require a
large, coupled network of quantum dots. I have already discussed two ways that electron spins in
two charged quantum dots may be entangled; the first is the cavity-qed proposal of Imamo
glu et al.
[68], discussed in Sec. 2.4.2. The second is the quantum repeater analyzed in Sec. 2.3.3. Both of
these require the construction of quantum dots strongly coupled to cavities; indeed this a current
direction of development for many research groups around the world [186]. The relatively shortlived electron spins in these quantum dots may be transferred to the nuclear spins using a variety
of techniques developed over the last 50 years in electron-nuclear double resonance. I have already
mentioned an example of such an experiment in a NV center in diamond [82].
It is tempting to imagine a large network of quantum dots prepared with a large number of
entangled nuclei between them. Coherent computation performed by rf pulse sequences may happen

106

Chapter 3. NMR Quantum Computers

in one dot while another is undergoing measurement or initialization. Moving a qubit from one dot to
the next could be done by teleportation; however, teleportation requires a Bell-state measurement,
and the light introduced for measurement of one qubit could destroy the coherence of other qubits
in the dot. Even for this a fix may be imagined, for if some nuclei are supposed to retain coherent
information while laser light is introduced for initialization or measurement, these nuclei may be
encoded into decoherence-free subspaces relying on the mostly uniform hyperfine field of the quantum
dot electron.
Such an architecture has become complicated, and a detailed account would easily fill a second,
entirely theoretical dissertation. All of these ideas hinge, however, on the supposed coherence of
semiconductor nuclei, emphasizing the importance of experimental tests of this coherence. One such
experimental test will be the subject of the next two chapters.

Chapter 4

Decoupling: Theory
It takes less time to do a thing right than it does to explain why you did it wrong.
Henry Wadsworth Longfellow
In this chapter, I describe methods for preserving coherence times in solid-state nmr in a sufficiently
controlled way to allow quantum logic. The goal is twofold: first, to lengthen dipolar-limited coherence times and experimentally measure lower bounds of coherence times for possible solid-state
nuclear qubits, and second, to develop techniques and ideas for employing the dipolar coupling for
quantum computing.
A word of warning must be made on the first goal, however. Here, I use the term coherence
to refer to the coherence of single nuclei, or in other words the preservation of single quantum
coherences. In a typical experiment, one begins by placing many nuclei in the equal superposition
state

1
|i = |i + ei |i ,
2

(4.1)

and then the shifts of over time between different nuclei are tracked. This dephasing is revealed
by an ensemble measurement; if inhomogeneous broadening is eliminated, the ensemble result is
similar to measuring a single nucleus repeatedly and averaging the results. Differences between
ensemble and single-spin measurements are discussed in Sec. 4.1. This single-spin definition of
coherence is not always appropriate in solid-state nmr because nuclei evolve under dipolar couplings
into coherent, highly correlated, many-body states. These states may be observed in experiments
designed to measure multiple quantum coherences; such experiments show that even when the phase
information of single nuclei has become lost to the ensemble, the ensemble as a whole has not lost that

108

Chapter 4. Decoupling: Theory

information to the environment. The recent experiments described in Ref. 146, for example, reveal
quite convincingly that these multiple quantum coherence terms also remain coherent. Whether
a process is treated as coherent or not depends on where the boundary of the coherent system is
defined; as discussed in Sec. 3.2.3, experiments attempting to measure the coherence of single nuclei
limit the system to the symmetric subspace of the nuclear ensemble, while experiments with multiple
quantum coherences consider all states of the ensemble as a single system.

4.1

Ensemble vs. Single-Spin Measurement

I discuss the results for measurement of the T2 decay of an ensemble of nuclei in terms of decoherence
of a single qubit, but there are differences between ensemble measurements and single qubit measurements. Important differences include the practicality of each measurement method, the effect
of back-action, and the sensitivity to initial conditions. Even if measurement artifacts due to such
factors are neglected, a fundamental difference between the two measurement types remains.
By definition, single-qubit decoherence is the uncontrolled decay of off-diagonal elements of the
qubit density operator in the logical basis. To be precise, let us denote by j (t) the density operator
of the system after tracing over all degrees of freedom other than the j th qubit. The off-diagonal
components, normalized by their initial value, are given by the function


Tr Ij+ j (t)

.
Gj (t) =
Tr Ij+ j (0)

(4.2)

The magnitude of this function will decay in the presence of decoherence. If single-nuclear-spin
measurement were possible, then |Gj (t)| is the decay function that we would construct by initializing
a single spin in the same state many times, measuring its x and y spin projections (in separate
experiments) at different times t, and averaging the results of an ensemble of experiments.
When we make a heterodyne nmr measurement of N spins that evolve according to some pulse
PN
sequence, our observable is M + = j=1 Ij+ . In this discussion, we neglect the small variation in

measurement strengths for each spin due to rf inhomogeneity. When we examine the normalized
magnitude |M (t)|2 = hM + (t)ihM (t)i/hM + (0)ihM (0)i of the measured magnetization, we obtain
2

|M (t)| =

PN

PN
j=1 Gj (t)
k=1
PN
PN
G
(0)
j=1 j
k=1

Gk (t)
Gk (0)

(4.3)

If we assume that each qubit begins in the same initial state, as occurs in our sequence, this may be

4.1 Ensemble vs. Single-Spin Measurement

109

simplified to
|M (t)|2 =

1 X
1 X
|Gj (t)|2 + 2
Gj (t)Gk (t).
2
N j
N

(4.4)

j6=k

The first term may be recognized as an average of single-spin measurement results. The second term
may recognized as an interference term. As a simple example of this formalism, suppose there is
only inhomogeneous broadening with no refocusing -pulses (so that each qubit oscillates at its own
frequency j ), but otherwise no decoherence. Then Gj (t) = exp(ij t) and
2

|M (t)| =

1 + N 1

j6=k

ei(j k )t

e2t/T2 .

(4.5)

The presence of inhomogeneous broadening is the most common difference between an ensemble and
single spin measurement. Hence, the -pulses used to refocus such effects are crucial.
The principal question now is whether the nuclear dynamics cause constructive or destructive
interference in the second term of Eq. (4.4). We expect constructive interference in a system of high
symmetry such as dipolar coupling in CaF2 ; here, the ensemble average is equivalent to a series
of single-spin measurements, and the oscillatory character of the dipolar dynamics is revealed in
experiments [126]. In a random environment such as isotopically natural silicon, the nuclear-nuclear
couplings differ from site to site. When dipole-dipole couplings are the source of decoherence, as
in the experiments described in the following chapter, each spin undergoes different oscillations in
its own dipolar environment, and destructive interference is seen. Therefore, the dipole-limited
coherence times observed are underestimates for what one might observe through a series of singlespin measurements on a typical nucleus, for which more oscillatory decay curves would be expected.
To illustrate this argument, we show in Fig. 4.1 the result of a simulation of eight spins evolving
according to a dipolar coupling with coupling strengths drawn randomly from a uniform distribution
between D0 /2 and D0 /2. While the initial decay observed in |M (t)| represents the average of the
initial decay of the individual spin measurements Gj (t), destructive interference reduces the longerlived oscillations observable in individual spin measurements.
When spectral diffusion is the leading cause of decoherence, single-spin decoherence would also
be interpreted from the ensemble measurement if each qubit witnesses the same noise spectral
density. A canonical example of such homogeneous broadening is Doppler broadening in the optical
spectroscopy of gases. In contrast, we will develop a model for spectral diffusion in polycrystalline
silicon, discussed in Sec. 7.3.2, which proposes that the random fluctuating environment seen by
each nucleus is different, depending on both the lifetime of the electronic fluctuation near each
nucleus and the frequency shift it causes. Hence a spin-echo experiment on a single spin might

110

Chapter 4. Decoupling: Theory

1.0
0.8
0.6
0.4
0.2
0.0
-0.2
-0.4

tD 0
0

Figure 4.1: Simulated dipolar decoherence versus time. The broken lines represent individual spin measurements Re{Gj (t)} on an 8-spin simulation of dipolar evolution with
uniformly random coupling constants with range [D0 /2, D0/2]. The solid line is the
magnitude-sum |M (t)|.

result in a decay from one echo to the next given by Eq. (7.41), whereas a sum over many such
spins yields the different function of Eq. (7.48). In this case, the more rapidly diffusing spins make
a lesser contribution to the ensemble average, and therefore the observed coherence times represent
overestimates for what one might observe through a series of single-spin measurements on a typical
nucleus.
I will make no further discussion of these distinctions, and instead refer only to the dipolar
decoherence of single nuclei, remembering that a true hypothetical single spin measurement might
yield different results.

4.2

Theory of Dipolar Decoupling

Discussion of decoupling techniques, especially as applied to spectroscopy, can be found in standard


nmr textbooks [122, 125]. In this section, I review only the pertinent details required to understand
experiments attempting to lengthen observable coherence times. I begin by discussing multiple pulse
sequence (mps) decoupling, and the related technique of magic angle spinning (mas).

4.2 Theory of Dipolar Decoupling

4.2.1

111

Multiple Pulse Sequences

In mps decoupling, rapid rotations are applied to the spin system by short, broadband rf pulses
in a periodic cycle. The key concept for comprehension of mps decoupling is the toggling reference
frame. This reference frame follows the pulses; if a /2 pulse about the x-axis occurs in the
pulse sequence, then simultaneously the toggling reference frame rotates by /2 about the x-axis.
Correspondingly, an isolated nucleus periodically rotated by the mps would be seen as stationary in
the toggling reference frame. To describe this, we use a shorthand for the terms of the Hamiltonian
that are constant in time: Hc = H0 + HD , where H0 = HZ + H are the combined Zeeman offset
and inhomogeneous broadening terms. The secular dipolar Hamiltonian given by Eq. (3.14) is also
abbreviated
HD =

X
j6=k



Djk 2Ijz Ikz Ijx Ikx Ijy Iky .

(4.6)

The rf Hamiltonian in the rotating frame, Eq. (3.7), with phases and amplitudes chosen to cause
the desired pulses and neglecting the counter-rotating term (rwa), is time-dependent and periodic
due to the pulse sequence. It therefore generates a periodic unitary operator
 Z t

Urf (t) = T exp i
Hrf (t )dt .

(4.7)

In the toggling frame, the total internal Hamiltonian is therefore also time-dependent, and its terms
are written

0 (t) = Urf
H
(t)H0 Urf (t)

(4.8)

D (t) = U (t)HD U (t),


H
rf
rf

(4.9)

where Urf (t) describes the unitary evolution due to the rf control sequence. Likewise, the coupling
to the environment takes on an additional time dependence, leading to
env (t) = ~
H

X
i

b(ri , t) Urf
(t)Ii (t)Urf (t).

(4.10)

During some intervals of the multiple pulse sequence, the toggling frame and the rotating reference
frame will be the same; it is only during these intervals that we measure the nuclear spin dynamics.
Measurements timed this way are called stroboscopic observation.
Calculation of the toggling frame Hamiltonian is best demonstrated by example. For this example we use the simplest symmetric decoupling sequence, the wahuha sequence [190] illustrated in

112

Chapter 4. Decoupling: Theory

S
0

Y
2

S
6

Figure 4.2: The wahuha pulse sequence. All pulses are broadband /2 pulses of the
indicated phase. Measurements are made at the times marked S; here the toggling and
rotating reference frames coincide.

Fig. 4.2. During the first half of the first application of this sequence, the toggling frame Hamiltonian
is
c (0 < t < ) = Hc
H
X
X


=
j Ijz
Djk 2Ijz Ikz Ijx Ikx Ijy Iky ,

c ( < t < 2 ) =
H

k6=j

 



k6=j

RX

Hc RX
2
2
X
X


y
=
j Ij
Djk 2Ijy Iky Ijx Ikx Ijz Ikz ,
 

 
 


c (2 < t < 3 ) = R
H
R

H
R

R
c
X
Y
Y
X
2
2
2
2
X
X


y y
x
x x
=
j Ij +
Djk 2Ij Ik Ij Ik Ijz Ikz .
j

(4.11)

k6=j

In the second half of the sequence, the rotations of the first half are undone:



 
  
c (3 < t < 4 ) = R R Hc R R
H
Y
Y
X
X
2
2
2
2
X
X


y y
x
x x
=
j Ij +
Djk 2Ij Ik Ij Ik Ijz Ikz ,
j

k6=j

 



c (4 < t < 5 ) = R
H
Hc RX
X
2
2
X


=
n Iny Djk 2Ijy Iky Ijx Ikx Ijz Ikz ,
n

c (5 < t < 6 ) = Hc
H
X
X


=
j Ijz
Djk 2Ijz Ikz Ijx Ikx Ijy Iky .
j

(4.12)

k6=j

Note that the vector components of the terms starting with I z spin operators in this sequence follow
the progression Z, Y, X, X, Y, Z. It makes a palindrome in time by design. A compact description

4.2 Theory of Dipolar Decoupling

113

known as the Mansfield notation [125] for this sequence is [[Z, Y, X]]. For times greater than 6 , we
simply repeat the pulse sequence. For this sequence we say that its cycle time tc is 6 .

Analysis of the effect of this pulse sequence is aided by Flocquets theorem, which tells us that
the unitary time-evolution operator generated by the periodic, time-dependent internal Hamiltonian
c (t) = H
0 (t) + H
D (t) may be written
H


U (t) = Urf (t)T exp i

Hc (t )dt Up (t) exp(iF t),

(4.13)

where Up (t) is periodic with the same period tc . This theorem is analogous to Blochs theorem in solid-state theory, except applying in time rather than in space. If we measure the spin
magnetization only once per period tc (stroboscopic observation), then we are interested only in
Up (mtc )[exp(iF tc )]m for integers m. The prefactor Up (mtc ) is constant and may be neglected. In
average Hamiltonian theory (aht) [191], F tc is expanded in orders of tc , the cycle time of the mps,
using the Magnus expansion:
F tc =

n=0

(n) tc .
H

(4.14)

c (t). The first


The nth order average Hamiltonian are written as time-integrals of commutators of H
three terms are
Z tc
(0) = 1
c (t1 )
H
dt1 H
(4.15)
tc 0
Z tc
Z t2
h
i
(1) = i
c (t2 ), H
c (t1 )
H
dt2
dt1 H
2tc 0
0

Z tc
Z t3
Z t2
h
i 
h
i
1
(2) =
c (t3 ), H
c (t2 ), H
c (t1 ) + H
c (t1 ), H
c (t2 ), H
c (t3 )
H
dt3
dt2
dt1 H
.
6tc 0
0
0
c (t). Terms of order n have magnitude |tc Hc |n ,
The zeroth order term is simply the time average of H
which become progressively smaller if tc is made sufficiently small. From Eqs. (4.12), we see that the
dipolar part of the zeroth order term is exactly zero. This is precisely the purpose of the sequence.
c (t) has time-reversal symmetry, as in the design of the wahuha sequence
It may be shown that if H
[[Z, Y, X]], then all odd-order terms also vanish.

A zeroth-order term does survive. By averaging Eqs. (4.12), we find the zeroth order term
(0) = 1
H
0
3

X
n

n (Inx + Iny + Inz ),

(4.16)

114

Chapter 4. Decoupling: Theory

= (1, 1, 1)/ 3 axis. In a typical nmr experiindicating that on average, nuclei rotate about the n
ment, however, only motion transverse to z is observed. Therefore under evolution of this sequence,
will be
a complicated transverse motion will be observed, and any nuclei pointing parallel to n
direction.
observed to be static. We say that wahuha creates an effective field in the n

A formal treatment of aht may be found in Mehring [125] and elsewhere. The discussion in
this textbook and others also shows why it is advantageous to use the sequence [[Z, Y, X]][[Z, Y , X]]
instead of just the wahuha sequence. Although this sequence no longer results in time-reversal
c (t), its ability to correct for pulse errors makes it strongly preferred over wahuha.
symmetry for H
This sequence is the important mrev-8 sequence [192, 193]. It has zeroth order effective field
(0) = 1
H
0
3

n (Inx + Inz )

(4.17)

and a first order effective field


X
(1) =
n2 (Inx Inz ).
H
0
3 n

(4.18)

Higher order effective field terms exist as well. These imply that the effective field is approximately

= (1, 0, 1)/ 2 direction for this sequence. The time is the space between pulses; in the
in the n
case of mrev-8, 12 = tc .

The lack of time-reversal symmetry means dipolar effects enter with a first-order dipolar crossterm:
(1) = i
H
D0
3

X

j Ijy ,

Djk (Ijx Ikx

Ijz Ikz )

j6=k

(4.19)

With the help of Appendix B, this may be simplified to


X
(1) = 2
H
Djk (j k )Ijx Ikz .
D0
3

(4.20)

j6=k

Note that this term would vanish if we had no inhomogeneous broadening. This may be considered
an effective dipolar-like interaction left by mrev-8. Recall, however, that we omitted nonsecular
terms of the dipolar Hamiltonian in the presence of large magnetic fields. If we apply our pulse
direction.
sequence sufficiently far off resonance, then we must consider the effective field in the n
Suppose we treat this effective field as very large. Then what is the secular component of Eq. (4.20)?

4.2 Theory of Dipolar Decoupling

115

To find out, we introduce a generalization of the raising and lowering operators, defined by
Ij =

p
inz ny nx y iny + nz nx z
1 n2x Ijx + p
Ij p
Ij ,
1 n2x
1 n2x

Ij .
Ij0 = n

(4.21)

, this choice of transformation is not unique, and in particular this one is ill-behaved
For a given n
=x
. However, it will suffice for this calculation. The inverse transformation is
for n

1p
1 n2x Ij+ + Ij ,
2

1
y
0
Ij = ny Ij + p
Ij+ Ij ,
2
2i 1 nx
X

1
p
Ijz = nz Ij0
nx (1 n2x ) iny I .
2
2nz 1 nx

Ijx = nx Ij0 +

(4.22)

= (1, 0, 1)/ 2, corresponding to the zeroth-order effective field, Eq. (4.20)


Using this definition with n
is rewritten


X

1 +
 + +
+
0 0
(1) =
H
D
(

)
I
I

(I
I
+
I
I
)

Ij Ik + Ij Ik .
jk
j
k
j k
j k
j k
D0
3
4
12

(4.23)

j6=k

The term in the square braces is secular ; the spherical tensor coupling between spins j and k
2
0
has exactly the form of Tjk
. The term in the curly braces is non-secular, with the form of Tjk
.

These latter terms will average out if the offset term is large enough, an effect known as second
averaging [194].
The presence of a first-order secular dipolar term in the presence of inhomogeneous broadening
limits the effectiveness of mrev-8, an effect noted first during the experiments to be described in the
next chapter. In the course of these experiments I also tried a variety of other pulse sequences which
are more sophisticated than mrev-8 for higher order dipolar decoupling and pulse correction, including br-24 [195], Cory-48 [196], and sme-16 [197]. These sequences are state-of-the art constructions
which show the best dipolar decoupling performance in samples with small inhomogeneous broadening. Although some decoupling is observed with all of these sequences, all of them suffer from the
same first order cross-term in the presence of heavy inhomogeneous broadening, and therefore none
showed superior performance to a new sequence we call mrev-16, shown at the bottom of Fig. 4.3.
The construction of mrev-16 makes only a minor modification to mrev-8. By examination of
Eq. (4.20), we see that if the effective field were in the z-direction, this term would be strictly

116

Chapter 4. Decoupling: Theory

0.1

Time (sec)
0

10

20

30

40

50

S MREV16
59tc

2 MREV
0
60tc

0.4

61tc

0.5

MREV

MREV

120tc

180tc

240tc

300tc

MREV16 S MREV16 S

60tc

0.3

tc=24

62tc

MREV8 [101]
X

0.2

MREV8 [101]
X

Figure 4.3: Schematic of the cpmg-mrev-16120 pulse sequence with spin-echo data. The echoes
shown in the upper left and expanded in the upper right are data from an isotopically natural single
crystal of silicon. These are obtained by first exciting the sample with a /2-pulse of arbitrary phase
, decoupling with the mrev-16 sequence shown in detail on the bottom line, and refocusing with
-pulses of phase = + /2 every 120 cycles of mrev-16. The magnetization is sampled once
per mrev-16 cycle in the windows marked with an S.

nonsecular. Recall that the wahuha sequence [[Z, Y, X]] effectively rotates the nuclei around the

(1, 1, 1)/ 3 axis, and that the mrev-8 sequence [[Z, Y, X]][[Z, Y , X]] effectively rotates the nuclei

around (1, 1, 1)/ 3 followed by (1, 1, 1)/ 3, averaging to rotation about (1, 0, 1)/ 2. The mrev16 cycle effectively rotates the nuclei in succession about all of the upper diagonal vectors in the unit
cube. The sequence achieving this in Mansfield notation is [[Z, Y, X]][[Z, Y , X]][[Z, Y, X]][[Z, Y , X]]. As
a result, the effective field is parallel to (0, 0, 1), as desired.
The first three effective field terms are
X
(0) = 1
H
i Iiz ,
0
3 i
X
(1) =
H
i2 (Iix 2Iiy ) ,
0
3 i


2 X 3
381
(2)
y

H0 =
i 3Ii
Iz .
144 i
2
For this sequence, tc = 24 as illustrated in Fig. 4.3.

(4.24)
(4.25)
(4.26)

4.2 Theory of Dipolar Decoupling

117

The first order dipolar-offset cross-term under mrev-16 works out to


X
(1) = 2
H
Djk (j k )(Ijx Ikz 2Ijy Ikz ).
D0
3

(4.27)

j6=k

Although this has an additional term not present in the mrev-8 result of Eq. (4.20), the new term is
also non-secular. To see this, we may use the usual operators Ij , since the effective field is parallel
to z. In terms of these, we find
X
(1) = 1 + 2i
Djk (j k )Ij+ Ikz + hermitian conjugate.
H
D0
3

(4.28)

j6=k

The terms in this first-order average Hamiltonian have the form of the non-secular Tij1 spherical
tensor operators; for sufficiently large offsets, they will average to zero due to second averaging.
The lowest-order secular terms of mrev-16 appear in second order. There are many of these
terms, and they can become quite complicated. For example, the second order dipolar Hamiltonian
for all wahuha and mrev sequences is [125]
i
2 h
(2) = H(x) H(y) , [H(x) , H(y) ] ,
H
D
D
D
D
D
18

(4.29)

using the notation


(a)

HD =

X
j6=k


Djk 3Ija Ika Ij Ik .

(4.30)

We may evaluate Eq. (4.29) using Eq. (B.15) on page 210. Using the convention that Djk = 0 when
j = k, we first derive
(x)

(y)

,H
]
[H
D
D
X


= 4i
Dij Dik 4xyz Iiz Ijx Iky 2xzy Iiy Ijx Ikz + yzx Iix Ijy Ikz + yxz Iiz Ijy Ikx + zxy Iiy Ijz Ikx 2zyx Iix Ijz Iky
ijk

= 12i

X
(Dik Dkj + Dij Djk + Dji Dik )Iix Ijy Ikz .
ijk

This result is quite simple and, notably, isotropic. Then we have


i
2 h (x)
(y) , [H
(x) , H
(y) ]
HD H
D
D
D
18
X


= 4 2
Djl (Dik Dkj + Dij Djk + Dji Dik ) Iiy Ily (Ijx Ikx Ijz Ikz ) + Iix Ilx (Ijy Iky Ijz Ikz )
ijkl

(4.31)

118

Chapter 4. Decoupling: Theory

2 2

X
ijk



Dij (Dik Dkj + Dij Djk + Dji Dik ) 3Iiz Ijz Ii Ij . (4.32)

This equation indicates the kind of effective interaction the nuclei undergo in the presence of a pulse
sequence, at least in the absence of pulse errors and inhomogeneous broadening. In general, cross
terms between the dipolar coupling and these other effects will also be present in second order.
The use of mrev-16 has another advantage, which is that the effective field is in the z-direction.
This simplifies the introduction of additional control pulses and the data analysis, because dominant
motion of the nuclei is in the same pattern as if there were no pulse sequence at all, a point we will
return to in the next chapter. However, the cost for this convenience is a reduction of spectroscopic
resolution. To explain, not only the direction but the magnitude of the effective field is important
for spectroscopy, since the effective field is the desired spectroscopic frequency shift. For wahuha,

frequency shifts are reduced by a factor of 3. For mrev-8 the reduction factor is 3/ 2, and for
mrev-16 the reduction factor is 3.

4.2.2

Magic Angle Spinning

mas decoupling operates on a slightly different mechanism than mps. Rather than using rf irradiation to induce nuclear rotations, the sample is physically rotated about an axis at angle m from the
z-axis. This is analogous to wahuha, except that the spinning is occurring in real space rather than
in the spin space. In the samples reference frame, the dipolar coupling constants Djk become time
dependent; the only component of Djk (t) constant in time is proportional to 3 cos2 m 1, which
is eliminated by choosing m to be the magic angle that eliminates this term. The effective-field
term of wahuha, for example, makes this angle with the z-axis. Other terms of Djk (t) oscillate
at multiples of the sample spinning speed. Again, one may expand the Flocquet Hamiltonian F in
D (t) averages to
powers of the spinning cycle period and find that the spinning dipole Hamiltonian H
zero in zeroth order but not in higher orders. In particular, a first-order correction is present [191].
An important question is whether mps experiments can outperform mas experiments in the
presence of inhomogeneous broadening. A rough comparison, using aht, may be made as follows.
For heavily inhomogeneously broadened samples, the leading-order average Hamiltonian contribut (1) , which is of order
ing to decoherence under mas is the first order dipolar/offset cross term, H
D0
tc T21 (T2 )1 . Here, T2 is the undecoupled decoherence rate, and T2 is the dephasing time due

to inhomogeneous dephasing. Under mrev-16, the leading-order term after second averaging is
(2) t2 T 1 (T )2 , as discussed above. If the cycle time for the mps is comparable to the rotaH
c 2
2
D00
tion period for mas, and if the field inhomogeneity is approximately the same for both experiments,

4.3 Spin Echoes

119

then mps may be expected to outperform mas if its cycle time is substantially faster than T2 . This
condition seems to be met in the experiments described in the next chapter, although a more careful
experimental comparison would have to be performed in the same apparatus, so that inhomogeneities
would be the same in both experiments.

4.3

Spin Echoes

Even if we suppose that the dipolar decoupling works extremely well, the inhomogeneous offsets
(0) term still cause static dephasing, which we desire to periodically
described by the dominant H
0
refocus. This may be done by applying -pulses every N cycles of mrev-16, creating periodic
trains of Hahns spin-echoes [134]. The basic principles of the spin-echo are well described in many
textbooks [122, 126], so I will not repeat these basic principles here. I will only focus on the effects
of -pulse errors and how to avoid them. The popular Carr-Purcell-Meiboom-Gill (cpmg) phase
convention is the easiest way to correct for these errors [198] and the one which has yielded the
best results in my experiments. It should be noted that the cpmg convention changes the sequence
according to the initial phase of the nuclear qubit, which is incompatible with the memory of
unknown or entangled quantum states. However, more complex nmr techniques such as composite
pulses [199] are known to allow phase-independent pulse correction. Refocusing -pulses are not only
useful for removing the effects of inhomogeneous magnetic fields; they would likewise be employed in
an nmr quantum computer for decoupling and recoupling of multiple dipolar-coupled qubits [200],
as we discuss in Sec. 4.4. We refer to the combined mrev-16 and cpmg -pulse sequence as cpmgmrev-16N ; its experimental implementation in silicon will be the subject of the next chapter.
Use of the cpmg convention for -pulses turns out to be very important for these experiments. I
found experimentally that using different phase conventions for echo sequences causes oscillations in
the echo amplitudes, due to -pulse errors. These sorts of effects are common to many areas of nmr
Here I present a formalism to analyze these errors, which will serve to explain why the heavily-used
cpmg sequence works. In this section, I focus on a single spin with offset and ignore the dipolar
coupling. The effect of inhomogeneous broadening will be seen by averaging over .

4.3.1

The Effects of -pulse Errors

The -pulse error of greatest importance is the deviation in the rotation angle by a small amount
. This error can never be eliminated for all spins in the ensemble due to the presence of rf
inhomogeneity, and so we focus on this error. Errors due to variations of the phase or shape of the

120

Chapter 4. Decoupling: Theory

pulse are less critical.


Let us consider the unitary operator U+ corresponding to waiting a time , applying a -pulse
with phase X and angle + , waiting a time 2 , applying a second ( + )-pulse with the same
phase, and then waiting another . This is the standard periodic part of a cpmg sequence. Another
possibility is to apply the same sequence, except to make the phase of the second pulse X. This
unitary operator will be written U . Let us suppose that free evolution for a single test spin corresponds only to precession around the z-axis by offset-frequency , and that the -pulse is infinitely
short. Then, making heavy use of Eq. (B.10) on page 208, we find
z

U = ei I ei(+)I e2i I ei(+)I ei I


h
i
z
x
x
x
z
x
x
z
= ei I eiI e(ii)I eiI e2i I eiI eiI ei I
z

= ei I eiI e2i I eiI ei I


i
h
ih
z
x
z
z
x
z
= ei I eiI ei I
ei I eiI ei I
= ei(I

cos I y sin ) i(I x cos +I y sin )

iI
eiIm
e n.

Let us now find the representation of this rotation operator in the usual 3 3 rotation basis. We
represent any single-spin unitary U (t) as
U

ij





I i U I j U
3
 =
=
Tr U I i U I j .
z
z

I(I + 1)(2I + 1)
I I

(4.33)

The prefactor will be abbreviated 1 . This representation may be fundamentally understood as


resulting from representing our single-spin density operator as




= x I x + y I y + z I z .

(4.34)



corresponds to converting i P U ij j . The convenience of this
The evolution U
j

representation is that it recalls the familiar 3 3 rotation matrices for real vectors2. That is,

1 These
2 For

eiI 0 cos

0 sin

sin ,

cos

(4.35)

results hold for any I; for I = 1/2 insert = 2.


single spins of any representation, classical rotations provide a complete description of all unitary evolution.

4.3 Spin Echoes

121

eiI

cos

0 sin

sin

cos

cos

eiI sin

sin

(4.36)

(4.37)

cos 0 .

0
1

In this representation, our double-( + )-pulse unitary operators become




ij
iI

= Tr I i eiIm
e n I j eiIn eiIm
U

 j


(1 cos ) + I n
sin eiIm
I cos + nj I n
.
= Tr I i eiIm

(4.38)

We now focus on each of the three terms in this expression. The first term is



cos ) I m
sin
Tr I i cos I j cos + mj I m(1

= ij cos2 + mi mj (1 cos ) cos ijk mk cos sin . (4.39)

The second term is




cos + (

m
sin ]
Tr I i nj (1 cos ) [I n
n m)(I
m)(1
cos ) I n

i nj (1 cos ) sin . (4.40)


= ni nj (1 cos ) cos + mi nj (
n m)(1
cos )2 (
n m)

The third term is




cos + I m(
m
n
)(1 cos ) I ( n
m)
sin ]
Tr I i sin [I n

j (1 cos ) sin [ni mj ij (


sin2 . (4.41)
= ijk nk cos sin + mi (
n m)
n m)]

Let us now check this expression in a couple of extreme limits. First of all, if = 0, it is clear that
ij
only the first term survives and gives U
= ij ; both sequences give the identity in the absence of

pulse errors, as they should. Perfect pulses lead to perfect spin-echoes. Now let us suppose there
=m
=x
, and from the
are pulse errors, but no inhomogeneous broadening ( = 0). In this case n
definition of U we should have U+ = exp(2iI x ) and U = 1. We indeed find
ij
U
= ij cos2 + xij (1 cos ) cos ijx cos sin + xij (1 cos ) + ijk cos sin (xij ij ) sin2

122

Chapter 4. Decoupling: Theory

= ij (cos2 sin2 ) + (1 1)ijx cos sin + xij (1 cos2 sin2 )

cos
2
sin
2
1

cos
2
+ ijx
+ xij
.
= ij
1
0
0
As a final important limit, let us consider the height of the second echo in the case of high inhomogeneous broadening. In this case, we average over an infinite uniform distribution. Hence, in
this limit,
= h
hmi
ni = 0

1 ij
( + yij )
2 x
1
hmi nj i = hni mj i = (xij yij )
2
hmi mj i = hni nj i =

= h
=0
h
n mi
n mi
= hm(
n m)i
=0
h
n(
n m)i
=
hmi nj (
n m)i

1 ij
( + yij ).
4 x

Thus we have
1
1
ij
hU
i = ij cos2 + (xij + yij )(1 cos ) cos + (xij + yij )(1 cos )2 (xij yij ) sin2
4
2

 1 ij
= (xij + yij ) cos + sin4
(x yij ) sin2 + zij cos2 .
(4.42)
2
2
ij
yy
yy
xx
xx
zz
zz
We note that hU
i is diagonal. Moreover, hU+
i = hU
i, hU
i = hU+
i and hU+
i = hU
i.

ij
ij
Hence we focus our attention on hU+
i, noting that hU
i can be recovered by switching the x and

ij
y components. Writing hU+
i as a matrix and expanding to fourth order in , we find

hU+ i
4

3 + 2 cos cos2

1 4
16

0
0

3 cos2 + 2 cos 1

0
0

1 2 +
0

0
13 4
48

0
1 2 + 31 4

4 cos2

+ O(6 ).

(4.43)

The meaning of these equations is as follows. If we use constant phase X -pulses, we consider the
xx
+ case and note that only hU+
i lacks terms of second order in . If we therefore begin with all

4.3 Spin Echoes

123

xx
spins pointing along X, our echo amplitude will scale as hU+
i and pulse errors will be cancelled

to order 2 so we should initialize with a Y or Y pulse. Either choice corresponds to the cpmg
sequence. On the other hand, if we use alternating pulses ( case, in which the xx and yy

components are switched from the + case), then we should not initialize with a Y pulse; rather,
we should initialize with an X pulse to start the spins in the Y -direction! Either method works for
correcting errors, but they cannot be combined. Either pulse-error compensation scheme depends
on the initial nuclear phase.
When there is another decoupling sequence causing an effective field in some direction besides
the z-direction, then the above analysis holds true as long as the pulses and initial spin directions
are orthogonal to that effective field. For the mrev-16 sequence, any initial transverse direction will
do, making the implementation much simpler.

4.3.2

Spin-echo Oscillations

In the experiments to be described in the next chapter, we will see how the second order errors
which occur with the improper choice of initial phase manifest themselves: they lead to oscillations
in the echo amplitude whose frequency depends on the average offset and the -pulse spacing. We
now simplify our expression for U in order to explain these oscillations. Collecting terms and using
the identities
m
= cos 2,
n

m
= z sin 2,
n

(4.44)

we obtain
ij
U
= ij (cos2 cos 2 sin2 ) + (mi mj + ni nj )(1 cos ) cos + ijk (nk mk ) cos sin

(mi jz iz mj ) sin 2 (1 cos ) sin + mi nj cos 2 (1 cos )2 ni mj sin2 . (4.45)


In the cases where oscillations are observed, there is an error to order 2 . We therefore expand this
expression to second order only:
ij
U
= ij + ijk (nk mk ) +


2  i j
m m 2ni mj + ni nj 2 ij (1 cos 2 ) + O(3 ).
2

(4.46)

124

Chapter 4. Decoupling: Theory

Written as matrices, this equation is

U+

2 sin 2

1
2
2

sin 2

1 22 cos2
2 cos

1 22 sin2
2
2

0
2 cos
1 22 cos2

+ 2 sin 2

2 sin

1 22 sin2

sin 2

2 sin

(4.47)

(4.48)

To order 2 , these are rotation matrices. To find the orientation of the rotation, we calculate the
eigenvectors. The characteristic equation for U+ is


(1 ) (1 22 cos2 )2 + 42 cos2 = 0,
so one eigenvalue is 1, as it must be, and the others solve, to O(2 ),
(1 )2 + 42 cos2 = 0.
The solutions, again to O(2 ) are
= 1 22 cos2 2i cos .
These have magnitude 1 + O(2 ). They represent ei+ , where
+ tan1

2 cos
2 cos
1 22 cos2

is the angle of rotation. Of course, this could be done more exactly: the character of a rotation
matrix is given by its (representation-independent) trace, which is equal to 1 + 2 cos . The trace of
ij
U
is easily calculated as

X
m


mm
U
= cos2 2 + 1 + cos2 2 cos2 + 2 sin2 2 cos 2 cos 2 sin2
= 3 22 (1 cos 2 ) + O(4 )

= 1 + 2 cos ,

4.3 Spin Echoes

125

a result which clearly shows the correct limits when = 0 or = 0. From this equation we clearly
derive
cos = 1 22 (1 cos 2 ) + O(4 ),
from which it follows again that
+ = 2 cos + O(3 )

(4.49)

= 2 sin + O(3 ).

(4.50)

Therefore, when the incorrect initial nuclear phase is chosen, each pair of echoes rotates each nuclear
spin by angle , which depends on the pulse error , the average frequency offset , and the -pulse
spacing .
About which axes do these rotations occur? The axes of rotation are given by the eigenvectors
corresponding to = 1, which are given by

cos
0

sin(2 )

0
sin

sin 2

now only to O(2 ). We see again that, in the + case, if the spins begin with the incorrect initial
phase Y , then they are orthogonal to the effective rotation axis + and they will be most sensitive
to pulse errors. Once again we see that we can best avoid these oscillations by using the standard
cpmg prescription of beginning the spins in the X-direction with a Y -pulse. The spins are then
only rotated for each pair of -pulses by the small angle around an axis tipped slightly above the
x-axis in the xz-plane.
If one uses phase-alternating pulses ( case) and the Y -pulse to prepare, the spins begin orthogonal to the rotation axis and rotate around it by angle every pair of pulses. The period of
the oscillation is roughly 2/ = / sin , so if the homogeneous linewidth is sufficiently narrow
and the offset is sufficiently large, this oscillation is readily observed. We will see it experimentally
in Sec. 5.1.7 as Fig. 5.4.

4.3.3

Even-Odd Asymmetry

So far we have considered the unitary operator due to a -pulse pair. However, one typically observes
also the echo occurring in the middle of the pair, and finds it to have reduced magnitude than the

126

Chapter 4. Decoupling: Theory

following echo. This aspect of the cpmg sequence is well discussed in Cowan [126]. We may quickly
use the formalism we have already developed to show how much that echo is reduced quantitatively.
Thus we want to find the unitary rotation corresponding to waiting a time , applying a single
-pulse of phase X, and waiting another duration . This rotation is
n
o
z
x
z
z
x
z
ij
U
= Tr ei I ei(+)I ei I I i ei I ei(+)I ei I I j .

(4.51)

Using Eq. (B.10) and a few trigonometric identities yields


ij
U
=

cos2 2 + sin2 2 cos 12 sin sin

sin sin2 2
cos2

sin 2 sin

sin sin2

+ sin2

sin 2 sin

cos

1
2

sin sin

cos 2 sin

cos 2 sin , (4.52)

cos

where = 2 . When averaged over a uniform distribution of frequencies , this becomes

cos2 /2

ij
hU
i=
cos2 /2

cos

1 2 /8

.
1 + /8

1 + 2 /2
2

(4.53)

So the odd echoes are reduced to second order in , while the even ones are only reduced to fourth
order, an effect frequently observed in experiments.
These pulse sequence effects are common to many subfields of nmr and mri, and a vast literature
exists analyzing errors and solutions. The mathematics we have developed in this section may prove
helpful for understanding this literature.

4.4

Recoupling Pulse Sequences

The previous section considers the theoretical problem of turning off the dipolar coupling and inhomogeneous broadening. However, in order to make a quantum computer employing this interaction,
we need a way to turn it back on between desired qubits.
The first step to doing this is to distinguish qubits with a large magnetic field gradient, as
discussed in Sec. 3.2.2. Then, as discussed in Sec. 3.2.3, qubits in different ensembles have a different
form of coupling, given by Eq. (3.35) on page 79. Coupled operations such as a controlled-not acting
on the j th and the j + mth spins can be performed by letting this coupling evolve for a time m3 /D.

4.4 Recoupling Pulse Sequences

127

The result, neglecting cross-coupling terms, is the unitary operation


P

z
z
UD,j,j+m m3 /D = ei k Ijk Ij+m,k .

(4.54)

The use of this unitary to form a controlled-not was discussed by Gershenfeld and Chuang [36].
As discussed in Chapter 2, the existence of this gate and all single qubit rotations assures universal
quantum logic.
The introduction of homonuclear dipolar decoupling pulse sequences as discussed in the previous
section will change the heteronuclear coupling; for example wahuha would change it into an exchange interaction. This has no effect on the universality of the interaction. The mrev-16 sequence
maintains the form of the heteronuclear coupling, but reduces its speed by a factor of 3.
The remaining requirement, however, is a scheme where the heteronuclear coupling may be
turned altogether off, achieving heteronuclear decoupling, and then reintroduced only between two
arbitrary qubits, achieving selective recoupling. Such a scheme was devised in Ref. 200; I briefly
summarize the scheme here.
The essential principle is as follows. Focus on qubit j. Then qubit k causes an effective magnetic
field of size D/|k j|3 . This field may be treated as inhomogeneous broadening, and hence may
be refocussed by a spin-echo sequence. A soft -pulse resonant to qubit j but far off-resonance
from the remaining qubits, with equal time duration before and after the pulse, will refocus the
coupling. Alternatively, a soft -pulse resonant to qubit k with equal time duration before and after
the pulse will also refocus the coupling, this time by inverting the sign of the effective field. These
work because in the toggling frame created by the pulses, the Hamiltonian of Eq. (3.35) changes
sign, leading to a zeroth order average Hamiltonian of zero. Further, the toggling frame Hamiltonian
commutes with itself at all times, so if the -pulses are perfect, every order term of the Flocquet
Hamiltonian vanishes.
We then consider schemes which concatenate a certain number of equal-time intervals and use
soft pulses to control the signs of Iiz for each spin. The essential information on the signs can be
represented by a sign matrix defined as follows. The sign matrix of a pulse scheme for n spins
with m time intervals is the n m matrix with the (i, a) entry being the sign of Iiz in the ath time
interval. We denote any sign matrix for n spins by Sn . For example, the inversion of the sign for a
single spin in a pair can be represented by the sign matrix

S2 =

+ +
+

(4.55)

128

Chapter 4. Decoupling: Theory

WAHUHA

i
i+1
i+2

Figure 4.4: A sample pulse sequence incorporating Hadamard decoupling and wahuha.
This is constructed directly from H(4) by translating each change of sign in the ith row
to a soft RX () pulse, labeled , at frequency i . The first, all positive, row of H(4)
has been removed to eliminate inhomogeneous broadening. Single qubit rotations, as
indicated by , must occur between full cycles of the sequence. Recoupling between
qubits i + 1 and i + 2 may be achieved by inserting a pulse where indicated by the
dotted line. The top row represents a broadband wahuha sequence for homonuclear
decoupling.

Each column represents a time interval; each row thus represents a sequence of m intervals for a
particular spin. An entry in the ith row represents an interval that is preceded and followed
by -pulses. Therefore, each sign matrix corresponds to a sequence of events for the whole system.
Decoupling is achieved whenever any two rows in the sign matrix disagree in exactly half of the
entries (all couplings are negated for exactly half of the time). The general construction of the
decoupling scheme is now reduced to finding sign matrices satisfying the above criteria. The kind
of sequence generated in this way is illustrated in Fig. 4.4.
The above scheme can be generalized to decouple n spins with m time intervals as follows:
Construct the n m sign matrix Sn , with entries + or , such that any two rows disagree in exactly
half of the entries. For each sign in the ith row and the ath column, apply -pulses before and
after the ath time interval.

A useful formalism for constructing sign matrices exists: the Hadamard matrix of order n,
denoted by H(n). This is an n n matrix with entries 1, such that H(n)H(n)T is n times the
identity matrix. The rows are pairwise orthogonal; therefore any two rows agree in exactly half of
the entries. Columns are also pairwise orthogonal.
Whenever H(n) exists, there is a decoupling scheme for n spins concatenating only n time
intervals. However, H(n) may or may not exist for a given n. For an arbitrary integer n, let n be
the smallest integer that satisfies n n with known H(n). To construct a decoupling scheme for
n spins when H(n) does not necessarily exist, we start with H(n) and take Sn to be any n n

4.4 Recoupling Pulse Sequences

129

submatrix of H(n). In other words, Sn is formed by choosing n rows from H(n), which still achieves
decoupling because subsets of rows of H(n) are still pairwise orthogonal. The resulting decoupling
scheme for n spins requires n time intervals.
This decoupling scheme also removes any frequency offsets caused by inhomogeneous broadening.
To see how, note that such frequency offsets correspond to terms in the Hamiltonian linear in Iiz , so
negating Iiz for half of the time results in no net evolution for these terms. Therefore, such evolution
for all spins can be removed if the sign matrix has identically zero row sum. We note that complete
negations of rows or columns of Hadamard matrices leave their essential properties intact; thus we
may always perform such negations to generate a Hadamard matrix which has only +s in the first
row and column. The sign matrix with zero row sum can be made by then removing this first,
all-positive row. All other rows have zero row sums by orthogonality. Such construction is possible
unless n = n, in which case construction should start with H(n + 1).
To implement selective recoupling between the ith and the k th spins, the sign matrix should have
equal ith and k th rows, but any other two rows should be orthogonal. The coupling term Iiz Ikz never
changes sign and that coupling is implemented selectively, while all other couplings are removed.
The needed sign matrix can be obtained from H(n) by replacing the k th row with a copy of the ith
row. This scheme also removes frequency offsets and requires no more than n time intervals.
The decoupling and recoupling schemes require n time intervals. They require at most nn pulses,
since Xi Xi = I and the X pulses are only used in pairs. The remaining question is: how does n
depend on n? Some Hadamard matrices are missing, either because no construction methods are
known or they simply cannot exist. H(n) can only exist for n = 1, n = 2 or n 0 mod 4.
Hadamard conjectured long ago that H(n) exists for every n 0 mod 4. This famous conjecture
has been verified for all n < 428. Even if this conjecture is false, the asymptotic scaling of n is linear
with n, as argued in Ref. 200.
The result of this scaling is that, when using the Hadamard scheme, the time to perform each
gate in a computation scales linearly with the number of qubits n. As qubits are added, the amount
of time needed to recouple them grows. For very large n, however, some qubits in the chain become
so distant that the 1/m3 factor in Eq. (3.35) renders their interaction negligibly small. In this
case the Hadamard pulse scheme can be truncated to decouple qubits only in sets of l, and a new
t
decoherence process is introduced with T2l
D = l3 [1+F 2 (l)]1/2 . If we compare the total decoherence
t 1
time T2 = (1/T20 + 1/T2l
) , where T20 combines all other decoherence processes, to the clock speed

tc , we find that there is an optimum l at which to truncate the Hadamard pulse scheme. The effective
number of logic gates T2 /tc , therefore, at first decreases as n and then flattens once n reaches this

130

Chapter 4. Decoupling: Theory

optimum l.
The implementation of a spin-echo sequence simultaneous with broadband homonuclear dipolar
decoupling, as in the cpmg-mrev-16N sequence, employs most of the experimental techniques
that will be required for quantum logic. There is one crucial difference: for effective quantum logic,
soft pulses must be used for the single bit rotations and the -pulses in the Hadamard scheme. The
implementation of simultaneous soft pulses for quantum logic has been nicely developed in liquidstate nmr quantum computers [124], but the presence of broadband homonuclear dipolar decoupling
complicates matters. The problem is that soft pulses take time to implement, and simultaneous
mrev-16 pulses sequences may not perform as expected during the soft -pulses.
An important avenue of research is to develop pulse sequences that combine broadband dipolar
decoupling with soft pulses, which also escape the Hadamard-schemes property of slowing down the
computer by O(n) with the number of qubits n. Preliminary work in this direction is in progress [201].

Chapter 5

Decoupling: Experiment
In theory there is no difference between theory and practice. In practice there is.
Yogi Berra
It is a common occurrence in nmr that pulse sequence ideas which look good on paper fail
in practice. It is likewise an unfortunate aspect of quantum physics that predictions of quantum
behavior are rendered useless by unexpected decoherence. In this chapter, based on Ref. 117, I
discuss experiments testing our ability to control the dipolar coupling in silicon, and attempts to
measure limits to the coherence times of silicon nuclei.

5.1

Experimental Details

Although mps and mas experiments are now routine in solid-state nmr, the application of these
techniques to silicon introduces new challenges. The small gyromagnetic ratio of 29 Si and its sparse
isotopic abundance (4.7% in isotopically natural silicon) is an advantage for observing long coherence
times; these imply that the dipolar coupling is low. However, the snr of an nmr measurement using
inductive pickup may be expressed [121]

2 n B0 Q V
snr
N,
T 3/2

(5.1)

where n is the nuclear density, V is the volume of the sample (assumed to fit tightly in the rf
coil), T is the temperature, N is the number of signal averages, and Q is the quality factor of the rf

circuit. Something in this expression must compensate for the low 2 n . Going to a magnetic field

132

Chapter 5. Decoupling: Experiment

B0 higher than 7 T is very expensive. Signal averaging is impractical because the time for nuclei to
recover after an experiment is 5T1 , which in the case of pure silicon is nearly a full day. The long
T1 suggests a magnetically quiet environment for the nuclei, which is promising for observing a long
T2 time, but the turnaround of an entire day between acquisitions makes the experiment difficult.
Lowering the temperature T might improve the signal, but then T1 quickly becomes much longer.
Raising the Q of the circuit too high is problematic because of the introduction of excessive phase
transients, which are glitches in the rf phase during pulse ramping due to too narrow a transfer
function in the rf circuit. These phase transients cause pulse errors which reduce the performance
of decoupling pulse sequence. The easiest option to raise the snr is to increase the volume of the
sample, and therefore also the volume of the coil in which it fits.
However, the duration of a /2 pulse may be expressed [122]

t/2

B0 V
.
P Q

(5.2)

Short /2 pulses are desirable; in the previous chapter we treated each rotation as happening
instantaneously. The effects of finite pulse width may be analyzed in aht [125] and under most
conditions they appear as a second-order correction. However, as the pulse width approaches the
pulse spacing the sequence must begin to fail. One wants t/2 to be as small as possible so that
and tc may be as small as possible, which in turn means that higher order terms in aht will be
as small as possible. If the volume of the sample is increased for improved snr, and if B0 and Q
are kept constant, the rf power must be made quite high. In our1 experiments, total rf powers of
800 W were used; this is much higher than in most nmr experiments.

5.1.1

Apparatus

The mps experiments were performed using an 89-mm bore 7 T superconducting solenoid nmr
magnet (Oxford) and a commercial nmr spectrometer (Tecmag) operating at 59.575 MHz. Pulses
were amplified with an 800 W broadband power amplifier (cpc). All experiments were performed
at room temperature.
The critical component of the experiment was the probe, which was homemade to accomodate
the large samples and high rf powers of about 800 W.
1 The

principle players in these experiments were myself and Denys Maryenko.

5.1 Experimental Details

5.1.2

133

Samples and Coils

We used a variety of samples with different rf coils designed to maximize filling factor and rf
homogeneity. The silicon samples we studied were:
An isotopically enhanced (96.9%

29

Si) cylindrical sample of single-crystal silicon; this sample

and its nmr properties were discussed by Verhulst et al. [202].


A stack of polycrystalline silicon cylinders purchased from Alfa Aesar; these samples are free
of impurities at the level of 0.1 ppm, with natural isotopic abundance (4.7%

29

Si). This stack

was 2 cm long and 0.95 cm wide.


A smaller sample of isotopically depleted silicon, with between 0.98% and 1.3%

29

Si, varying

across the sample. This sample, grown using the techniques described in Ref. 203, was a
cylinder 6 mm in diameter and 2 cm long. It featured aluminum impurities at a level of
1017 cm3 introduced to shorten T1 to allow signal averaging.
A high-quality single-crystal of silicon purchased from Marketech, also with natural isotopic
abundance. This sample featured less than 5 1013 cm3 n-type impurities. It was cut into
a sphere of diameter 1.5 cm.
Heavily doped commercial wafers of metallic n-type silicon. These samples were convenient
because the T1 was only 50 seconds, unlike the purer samples for which we measured a T1
of 4.5 hours. However, the snr for the metallic silicon was always substantially lower, and
decoupling sequences always performed poorly, even for powdered samples. We speculate that
this is because of skin-depth effects.
With all but the single crystal samples, we used a 6 cm-long, 1 cm-diameter coil wound using 2 mmdiameter bare copper wire with variable pitch to improve rf homogeneity [204]. The rf homogeneity
was important for the longer samples; we characterized the homogeneity by measuring

29

Si Rabi

oscillations using a similarly-sized sample of solidified grease containing dimethyl siloxane (dmso).
We found the fid intensity for the antinode at pulse angle 450 to be 92% as strong as that at 90 ,
indicating moderate rf homogeneity. The /2-pulse duration was 9 s for this coil. Experiments
with the single crystal sample used a constant-pitch coil approximately 6 cm long, in which the
sample fit very tightly. The /2-pulse duration was 15 s for this coil. The rf homogeneity for this
coil over the spherical sample was roughly the same as for the variable pitch coil over the cylindrical
sample. Both coils were held firmly by coil forms made of Delrin plastic.

134

Chapter 5. Decoupling: Experiment

5.1.3

The Circuit

The nmr circuit was a series-parallel tank circuit, which is extremely common and is analyzed in
several textbooks. The text by Fukushima and Roeder [205] is particularly helpful2 . Generally, the
tuning and matching of such a circuit is most easily performed by choosing the variable capacitors
which closely achieve the desired resonance given the inductance of the measurement coil, and
then adding fixed ceramic capacitors to various places in the circuit to perfect the tuning near the
middle range of both variable capacitors. Glass, teflon, quartz, and sapphire variable capacitors were
obtained from Voltronics, Inc. The sapphire capacitors appear far more robust at high power; the
glass and quartz capacitors have at times cracked and the teflon capacitors burn. For the extremely
high power work, we obtained variable vacuum capacitors from Jennings. These variable capacitors
are extremely robust but also quite large. Fixed ceramic capacitors consistent with high power
operation were obtained from hec.
We designed and built the probe to be as rigid and robust as possible for high power rf, embedding the variable vacuum capacitors in Delrin plastic and placing them as close to the coil as possible.
The Delrin dielectric and the short high-voltage leads minimized arcing through air. Using these
very large capacitors near the coil allowed high snr and flexible tuning for frequency and impedance
matching, but resulted in a very bulky probe that could not fit in a cold cryostat. Coil heating was
a large concern, as the plastic coil holder would melt after about a minute of a high power mps;
however, the probe tuning remained roughly constant during long pulse sequence experiments as
checked by continuous monitoring of the rf power reflected from the probe.
The transmission line from the tank circuit to the preamplifier was made from coaxial copper
pipes designed for 50 impedance. This design allows the option of omitting capacitors from
the tank circuit altogether, and instead using the low-loss transmission line to transform the coil
impedance to 50 by bulky elements outside the magnet. However, snr is always reduced in such
a mode since the effective volume of the resonator includes the long transmission line. We did not
obtain favorable results without the capacitors in the magnet.
A small open bnc connector provided an antenna for direct monitoring of the rf-field, which
allowed us to symmetrize phase transients without nmr detection. We found that our results did
not noticeably vary with asymmetric versus symmetric phase transients, as long as the quality factor
Q of the circuit was kept reasonably low. The Q for a coil with inductance L and resistance RL is
Q = 0 L/RL . The thick copper wire used in our coil and the coils large size lead naturally to a
high Q, but the effective Q of the circuit was reduced with an added capacitor, as I now explain.
2 If

you are a new student working on nmr, do not hesitate to find this book!

5.1 Experimental Details

135

I0
C

RL
ZL

C'

C''

Figure 5.1: The series parallel tank circuit with an extra Q-reducing capacitor C .

Although the tank circuit analysis is common, we empirically found that the addition of a
capacitor, C , between the coil and ground aided greatly in the performance of the circuit in terms
of its tunability and its characteristics with phase transients. This capacitor effectively lowered the
circuit Q without substantial reduction of the B1 field. I now review the usual tank circuit analysis
with the possible inclusion of this capacitor to show why.
Suppose we have an impedance ZL in parallel to a tunable capacitor C and in series to a tunable
capacitor C . The impedance ZL certainly contains the coil inductance L, but it also contains the
resistance RL and possibly the added capacitance C . The impedance of the whole circuit, shown
in Fig. 5.1, is
Z() =

ZL
1
+
.

iC
1 + iZL C

(5.3)

The reactance (the imaginary part of the impedence) is therefore


X() =

1
XL C|ZL |2
+
.
C
|1 + iZLC|2

(5.4)

We presume that RL ZL , and so we find the resonance frequency r by solving X(r ) = 0 when
ZL = iXL . We therefore find simply that
r (C + C )XL (r ) = 1.

(5.5)

136

Chapter 5. Decoupling: Experiment

When ZL = iL, this leads to the usual analysis of the series-parallel tank circuit. If we allow ZL
to contain both a capacitor C and an inductor, i.e. that XL () = L 1/C , then we find
r2

1
=
L

1
1
+
C + C
C

(5.6)

What of the impedance? From (5.3) we have simply


1
R()
=
.
RL
|1 + iZLC|2

(5.7)

We are interested in the real part of this expression, at r , for small RL :


R(r )
1

.
RL
(1 r XL C)2

(5.8)

We have simply
r XL = r2 L
leading to
R(r )

RL

1
1
=
,
C
C + C
C
1+
C

2

(5.9)

which is independent of C . Whether the C capacitor is present or not, we must tune the capacitors
C and C so that R(r ) matches the impedance of the transmission line and subsequent amplifiers,
so we set R(r ) = R0 = 50 . If we assume that C and C are always adjusted so that the resonant
frequency is set to near the desired Larmor frequency and the impedence is set to R0 , then the
extra capacitor C serves only to decrease the effective inductance of the coil. In so doing, it also
reduces the total reactance of the coil, essentially lowering its quality factor, defined here as simply
the phase of ZL . Decreasing this effective Q lessens the probes sensitivity to tuning and reduces
phase transient effects.

The cost for this improved performance, of course, is that the B1 field at resonance is slightly
reduced. The current passing through the inductor can be seen by direct calculation. The current
through ZL and the tuning capacitor C must sum to the total current in the circuit, which has

magnitude |I0 | = P R0 , where P is the total power delivered to the circuit and R0 is the total
impedance of the circuit on resonance, which is again designed to be 50 . The capacitor current
at resonance is ir C VL , and the current through ZL is IL = VL /ZL (r ), where VL is the common
voltage across these two elements. Since these sum to I0 , we must have I0 /VL = (ir C + 1/ZL(r )),

5.1 Experimental Details

137

which means the current through ZL is


IL = I0

1
1
= I0
.
1 + ir C ZL (r )
1 C /(C + C ) + ir RL C

(5.10)

We now ask the effect of adding the capacitor C , assuming that L, RL , and r are fixed and
that the circuit is tuned and matched. The factor C /(C + C ) is fixed by virtue of Eq. (5.9)
p
to (1 + R0 /RL )1 , and therefore the only factor which changes is the imaginary part of the
denominator. From the resonance and matching conditions, it may be easily be found that the
addition of C requires C to be tuned to
p
RL /R
C =
,
2
Lr 1/C

(5.11)

and therefore as C is reduced C must be suitably increased. This means the total current in
the inductor is decreased. However, for low resistance coils it is not decreased by much, since for
p
R0 RL the real part of the denominator of Eq. (5.10), which does not change, scales as R0 /RL ,
p
while the imaginary part, which only slightly increases, scales as RL /R0 . The full expression may
be simplified to

p
p
R0 /RL
RL /R0
I0
p
+i
=

IL
Q 1/RL r C
1 + R0 /RL

RL
R0

i
+
Q

RL
R0


1+

1
2
r LC

(5.12)

showing once again that this capacitor has served only to lower the effective Q of the inductor. In
these large-coil decoupling experiments, there is plenty of Q to spare. The improved pulse sequence
performance is well worth the very small reduction in signal and in the rf field.

5.1.4

Small Angle FID

Before one can perform decoupling experiments, samples need to be characterized and pulses need
to be tuned. The methods for these standard nmr procedures are many. Here, I only discuss a few
useful concepts which are particularly helpful when dealing with long T1 samples such as silicon.
Suppose T1 is extremely long and the only goal of the experiment is to observe an fid spectrum,
perhaps to measure the resonant frequency of the nuclei, deduce T2 , or measure T1 in a saturationrecovery experiment. In order to get a high signal without waiting for T1 recovery, the following
question is important: is it better to use a /2 pulse in a single shot, or average together many shots
with a smaller angle? The answer is the latter.
This is quickly seen as follows. Suppose you use tipping angle . Then the first fid has a signal

138

Chapter 5. Decoupling: Experiment

proportional to sin , and the transverse magnetization decays during the measurement to leave a
residual polarization proportional to cos . Consequently, the next shot has signal sin cos , and
the polarization decays to cos2 . If we repeat this N times and sum, we obtain the signal

S(N, ) =

N
1
X

1 cosN
sin
N
sin
.
1 cos
1 cos

sin cosn =

n=0

(5.13)

This is maximized at = 0, where it goes as 21 . If T1 relaxation is neglected, then, the maximal


signal results from the lowest possible angle . This, however, neglects noise. If is brought to zero

and an infinite number of measurements are made, the N accumulation of incoherent noise will be

disastrous. Therefore we may generalize by assuming a noise term proportional to N . The signal
to noise ratio is then
R(N, ) S(, )

1 cosN

(5.14)

which is maximized when


R
N
R

cosN (1 2N ln cos ) 1 = 0,

=0

= 0 (1 cos )[N sin2 cosN 1 + cosN 1] = 0.

(5.15)

These equations may be combined into the single relation


1 cos
1 2N ln cos

sin2
+ 2 ln cos
cos

= 0.

(5.16)

This still turns out to be satisfied only at = 0. However, it is interesting to note that near equality
is achieved for many values of as long as it is small. The equation is exactly solved for all near
zero to order 3 . This suggests that there is a large region in the range of N and where R(N, )
is near its maximum. For given , the optimal N is at
N

1.25643
.
ln sec

(5.17)

For any finite , any more acquisitions than this will add more noise than signal. If N is chosen as

this maximum value, the snr depends on as ln sec S(, ). To lowest order in , this function

is 2(1 + 4 /96); a very flat function indeed for all /2. Therefore almost any small could be
chosen, and then N appropriately chosen according to Eq. (5.17).

5.1 Experimental Details

5.1.5

139

Pulse Tuning

Pulse errors degrade the performance of both mrev and cpmg pulse sequences; they only selfcompensate for errors to a limited degree. A good method to tune the /2 pulse is therefore
necessary.
When tuning the /2 pulse for a high T1 sample, one might try to include in the coil another
sample whose T1 is short and whose resonance does not overlap the sample of interest. The pulse
may therefore be tuned using this sample. When doing experiments with pure silicon, therefore,
we include in the inhomogeneous region of the coil some dmso, whose T1 is 23 seconds and whose
chemical shift is such as to shift its resonance 3.477 kHz from the pure-silicon resonance. Although
this shift is large enough to resolve the two materials completely, it is negligible in comparison to
the Rabi frequency. The effective field created by off-resonance rf may be written
Beff = ( cos , sin , ) =

p
2 + 2 (sin cos , sin sin , cos )

where = /2 + tan1 (/). A typical /2 pulse duration in our high-power probe is 14 s


on resonance, implying /2 17.9 kHz. So the magnitude of the effective Rabi frequency is

3.4772 + 17.92 kHz = 18.2 kHz, a change of less than 2%. The axis of rotation tips away from the
xy-plane by only tan1 (7.9/3.477) = 11. It is easily shown that the magnetization is rotated
into the xy-plane for an off-resonance pulse at the time
 cos1 ( 2 /2 )
1

cos1 cot2 =

2
2 + 2
2 + 2



4 2
1+
= 14.07 s.
2 2

This 70 ns addition to the /2 pulse width is beneath the resolution of our pulse-programmer, so
we may neglect it.
However, in practice this idea does not work so well, for the simple reason that there is not
enough room in our coil for both the silicon sample of interest and the dmso. Whichever sits in the
middle of the coil sees a nice, homogeneous rf field and shows clean Rabi oscillations. That which
sits in the edges sees an inhomogeneous rf field and very low-visibility Rabi oscillations.
Although other techniques have been tried, to date it seems the best tuning sequence is to saturate
the sample, wait 5T1 , pulse with a duration , and repeat while incrementing . This is painfully
slow, but it is reliable. In principle, one obtains the highest resolution for such a tuning sequence
around , since it is here that the sine function has its largest first derivative. However, this
is also where the sine function is close to zero, so in the presence of noise there is very little signal
with which to tune! Hence in practice we find the peak fid signal near /2. This peak is very

140

Chapter 5. Decoupling: Experiment

broad; the full-width-half-maximum (fwhm) of its absolute value is fwhm = 1 2 sin1 (1/2) =
1.0471. A sharper peak can be found by examining the second spin-echo. Let us suppose that
both nearly--pulses have the same phase as the initial pulse. The nearly--pulses have angle + ,
and the initial pulse has angle ( + )/2. Then, our initial spin vector is (0, cos(/2), sin(/2)),
and in the high inhomogeneity limit, this is multiplied by hU+ ()i as given by Eq. (4.42). It is easy
to see that it is only the yy-component of this matrix which matters, since we only measure the
x and y components, hU+ i is diagonal, and the initial x component is zero. The amplitude of the

yy
second echo therefore goes as A() = hU+
()i cos(/2). Let us as a final manipulation write that

the correct /2 pulse duration is 0 = /(2), so that = ( /0 1). We may then write that the
amplitude of the second echo is given by
A() = sin


2 0




3 cos2 ( /0 ) 1 cos ( /0 )

.
4
2

(5.18)

The peak here is somewhat narrower than the simple sine function; the fwhm of its absolute value is
0.741 . Using echoes further out than the second would be sharper still, but unless T2 is very long
one runs again into the problem of diminished signal. Note that using alternating -pulse phases
(and hence using hU ()i) would be foolish for tuning purposes, since these alternating phases
deliberately make the echo height less sensitive to tuning errors. The factor of 3 cos2 ( /0 ) 1

would become 3 cos2 ( /0 ), and the resulting peak is broader.

5.1.6

Data Analysis

Before proceeding with further experimental details, I will explain how data is processed when
analyzing the results of a cpmg-mrev-16N experiment.
The magnetization is measured as an oscillatory voltage after inductive pick-up and amplification;
the reading is regarded as a complex number whose real part is the component in-phase with the
local rf oscillator and whose imaginary part is out-of-phase with the local rf oscillator. This
signal is sampled at the end of each mrev-16 sequence. A single sampling actually measures the
magnetization for a time duration of approximately , but is so short that only a few points exist
in this time series. This short time signal is simply time-integrated, unless the sequence is applied
so far off resonance that the average offset frequency is larger than 1/ , in which case these time
points are Fourier-transformed and the region around the spectral peak is integrated.
The result is time series data with one point every tc , except for those locations occurring every
N tc where a pulse is applied. This time series shows a train of echoes which decay. For examining

5.1 Experimental Details

141

an echo train, we could simply integrate the magnitude of each echo and follow the decay of the
magnitude area. Improved decoupling, however, results from off-resonance pulses due to second
averaging, and an important processing step is to assure that the nuclear spin oscillations occur
at the expected frequency. Therefore I now describe in detail the way a train of echoes is Fourieranalyzed.
We notate our signal before dephasing as S(t), and presume that if we could measure it continuously, it would be something like
S(t) = eit et/T2 .

(5.19)

The effect of inhomogeneous broadening and the cpmg refocusing sequence is to multiply this
continuous signal by the echo-envelope function
Z
X

G(t) =

n=

g(t t )(t nN tc )dt .

(5.20)

The function g(t) has an approximately Gaussian shape; it looks something like g(t) = exp[(t/T2)2 ].
Note that G is defined for all t and is perfectly periodic. On the other hand, S(t) is zero for t < 0.
We denote the continuous time signal with dephasing as H(t) = S(t)G(t).
Let us denote the discrete Fourier transform (dft) as
B/

DA,B


f (t) (n ) =

imn

f (m ) =

m=

m=A/

f (t)ein t (t m )dt.

(5.21)

This is to be recognized as the Fourier transform of the product of three functions:


f0 (t) = f (t),

1, A x B,
f1 (t) =

0, otherwise,

f2 (t) =

m=

(t m ) =

1 X 2ikt/
e
,

k=

f0 (t)eit dt = f(),

f1 (t)eit dt = (B A)sincA,B (),

f2 (t)eit dt =

2 X
( 2k/ ).

k=

For convenience, we have defined


sincA,B () = e

i(B+A)/2



BA
sinc
,
2

(5.22)

142

Chapter 5. Decoupling: Experiment

where sinc(x) = sin(x)/x. Hence we may write

i


BA X h
DA,B f (t) (n ) = 2
f sincA,B (n 2k/ ),

(5.23)

k=

where indicates convolution.


To acquire spectral information, we apply this dft on the N 1 points between the nth and

(n + 1)th -pulses. We call this a block of points. There is no data point at the location of the

pulse, but a null point is inserted so that the time series is evenly spaced. The points with data
corresponding to the nth block run from nlow = nN N/2 + 1 to nhigh = nN + N/2 1. Therefore
the nth spectrum is


H(t) (j )
h
X

 
i
einN tc sinc(T ) (j 2k/tc),
= 4(T /tc )
S G
n

hn (j ) = Dtclow

tc ,nhigh tc

(5.24)
(5.25)

k=

where T = (N/21)tc is the analog to the usual timescale in spin-echo experiments. To calculate

G(),
let us first expand G(t) in a Fourier series. It has period N tc , so it may be written
G(t) =

Aj e2ijt/N tc ,

(5.26)

G(t)e2ijt/N tc dt.

(5.27)

j=

where
1
Aj =
N tc

N tc /2

N tc /2

If echo functions are strictly non-overlapping, meaning T2 T, then Aj is nothing more than the
Fourier transform of a single echo [with function g(t)] at frequency 2j/N tc . This remains the case
even when the echoes overlap if we use the definition of G(t) given by Eq. (5.20). We thus write in
all cases that
Aj =
so that

1
g (2j/N tc ) ,
N tc

2 X

G()
=
g(2j/N tc )( 2j/N tc ).
N tc j=

(5.28)

(5.29)

Correspondingly,


X

() = 2
2j/N tc )
S G
S(
g (2j/N tc ),
N tc j

(5.30)

5.1 Experimental Details

143

and
h
 
i
einN tc sinc(T ) =
S G

Z


2 X
2j
2j/N tc )ei( )nN tc sinc ( )T d .
g
S(
N tc j
N tc

(5.31)

Here we find it convenient to make our first approximation. We assume that because T2 T , the

signal spectrum S()


is very strongly peaked around in comparison to sinc(T ). Hence the

2j/N tc )
sinc[( )T ] part of the integral is roughly constant over the small range where S(

is appreciable. This range is where 2j/N tc is close to . Hence we find

h
 
i
einN tc sinc(T )
S G
Z


2 X
)einN ( )tc d
S(
g (2j/N tc ) sinc ( )T j(1 2/N )
N tc j
=



4 2 X
g (2j/N tc ) sinc ( )T j(1 2/N ) einN tc S(nN tc ).
N tc j

(5.32)

We then have the nth spectrum


hn (j ) =



 
X  2k 
k T
g
sinc (j )T 2 l +
.
N tc
N tc

T
16 3 2 einN j tc S(2nN tc )
tc
kl

(5.33)

To make sense of this, consider our second approximation: T T2 . The sinc function in this case
has a width in the (discrete) space of k-indices that is much less than g(2k/N tc ). Correspondingly,
this sinc may be taken as a discrete -function, forcing l + k/N to be (j 0 )tc /2. Therefore,
under these approximations the spectrum is
hn (j ) = 16 3

X
T inN j tc
e
S(2nN
t
)
g(j 2l/tc ).
c
t2c
l

Note that we may still have aliasing if T2 tc , the sampling time. If T2 tc , only the l = 0 term

in the sum will matter. The cycle time must be made much shorter than T2 not only to eliminate
this possibility, but also to minimize higher order terms in aht.

The conclusion is that the full set of approximations


T2 T T2 tc ,

(5.34)

144

Chapter 5. Decoupling: Experiment

Imaginary Part
0

0.2

0.4 0.6 0.8


Time (sec)
Freq. (Hz)

1.0

-400

Tim

e (s
ec)

Real Part

400

10

Figure 5.2: A typical data time series and Fourier transform showing clear spin locking
effects. This data was taken with polycrystalline silicon under mrev-8. The raw time
series on the left shows the first second of the real (in-phase) and imaginary (out-of
phase) amplitudes; the waterfall plot on the right shows the result of applying the dft
to each echo block as described in Sec. 5.1.6.

results in the simple spectrum


hn (j ) einN j tc S(2nN tc )
g (j ).

(5.35)

Generally we show the absolute magnitude of these plots, so the overall phase factor does not matter.
This results in the spectra shown in the waterfall style plot in Fig. 5.2.

5.1.7

Spin Locking

For the perfect signal of Eq. (5.19), we expect a waterfall plot to appear as a simple spectrum
showing the Fourier transform of the echo function decaying in time. And indeed, on the right-hand
side of Fig. 5.2, this is what we see. However, there is also a long train of spikes at zero frequency,
and a reflection of the spectrum on the opposite side of these spikes.
This data was taken using mrev-8, and these extra features are a result of the effective field
discussed in Sec. 4.2.1. The spike in the middle is due to spins that begin parallel to the effective
field. In the toggling frame, these spins are polarized along this effective field, and cannot become

5.1 Experimental Details

145

depolarized without exchanging energy. Their lifetime is therefore not T2 , but rather a kind of T1 ,
which could perhaps be called T1 , for T1 in the toggling frame. This is a kind of spin locking,
where in the usual case using continuous-wave rf, the locked spins decay according to T1 , T1 in
the rotating frame. Note that this spin-locked component is still rotated by the -pulses; its phase
switching is clearly visible in the corresponding time-series plot of Fig. 5.2.
The wide lobe on the right-hand side of the waterfall plot corresponds to spins transverse to the
effective field, and so does the lobe on the left-hand side. This lobe is sometimes called the quad
echo.
The origins of these spectral features can be deduced by analyzing the result of a rotation about

= (1, 0, 1)/ 2. The Hamiltonian of Eq. (4.17) describes rotation about this axis at average offset
n

frequency = 2/3. In the standard 3 3 representation, this results in the rotation matrix

2 sin t

1 + cos t

1
U (t) = 2 sin t
2
1 cos t

2 cos t

2 sin t

1 cos t

2 sin t .

1 + cos t

(5.36)

The measured signal is S(t) = x (t) + iy (t), which for this evolution yields

S(t) =
x (0) + z (0)
+
2

[x (0) z (0)][1 +
4

2]

+ i (0) e

i t

[x (0) z (0)][1 +

2]

ei t .
(5.37)

The first term corresponds to spins that are spin-locked to the effective field. The second and third
terms are the transverse terms which may undergo decoherence, and they appear at the frequencies
, corresponding to the left and right lobes seen in the data. The relative sizes of the lobes
depend on the initial spin projections j (0) created by the preparation pulse.

In experiments measuring physical decoherence times, it is crucial to separate the coherently


oscillating side-lobes from the spin locking. To some degree, this problem is alleviated by the mrev16 sequence. However, due to finite pulse-width effects and higher order average Hamiltonian terms
[see Eqs. (4.25) and (4.26)], the effective magnetic field witnessed by the nuclei during mrev-16 is
still not exactly parallel to the z-axis. Consequently, a magnetization may still be spin-locked to
this effective field with a small component in the xy-plane. The spectral separation of the spinlocked center peak from the sidelobes is therefore still an important step in the data analysis. For

146

Chapter 5. Decoupling: Experiment

1
Magnitude of center peak

0.1

0.01

Magnitude
of side peak
0.2

0.4

Time (sec)
0.6

0.8

Figure 5.3: The magnitude of the center and side peaks from a waterfall plot when pulses are applied frequently (N = 5) without the cpmg convention under mrev-8 with
tc = 1.03 ms, again taken with polycrystalline silicon. These correlated oscillations in
the magnitude are due to -pulse errors.

the results in Sec. 5.2, the rf was detuned about 120 Hz from the center of the nuclear resonance
frequency. The transverse field was then seen to oscillate with a center frequency of /2 40 Hz,
as expected from the zeroth-order average Hamiltonian. Decay curves were generated by integrating
the detuned side peaks between half-maxima. This procedure also eliminated the influence of pulse
ringdown effects, since any measurement artifacts from our applied rf also appear in the center of
each spectrum.
In principle, if pulses are applied orthogonal to the effective field, the spin-locked and transverse
components should remain independent. However, in the presence of -pulse errors, these two
components can become mixed and the magnetization may oscillate between them. Such an effect
is shown in Fig. 5.3.
This data was generated using mrev-8 with frequent -pulses (applied every N = 5 cycles) not
following the cpmg convention. In this case, the preparation /2-pulse was chosen to have phase
X, and subsequent -pulses were applied alternatingly in the Y and Y directions. As discussed in
Sec. 4.3.1, this sequence does not correct for -pulse errors like the cpmg convention. In principle, it
should initialize spins transverse to the effective field, and subsequent -pulses should be transverse
to the effective field. Nonetheless, a spin-locked component is observed, and as shown in Fig. 5.3 it
undergoes slow oscillations correlated to similar observations observed in the transverse components.
This is due to errors in the -pulses. Figure 5.3 shows the magnitude of the data for N = 5; if N
is reduced and one phase component of the data is observed, a more regular oscillation is observed,
as shown in Fig. 5.4. These oscillations were described in Sec. 4.3.2. Once again, the magnetization

5.1 Experimental Details

0.4

147

Imaginary Part

0.2
0.0
-0.2

Offset=400 Hz
N=60
N=120
N=180

-0.4
-0.6
-0.8
-1.0

10
15
Time (sec)

20

25

1.0
0.8
0.6
0.4
0.2
0.0
-0.2
-0.4
-0.6

Imaginary Part
CPMG, Offset=300 Hz

N=60
Offset=400 Hz
Offset=300 Hz
0

10
15
Time (sec)

20

25

Figure 5.4: Oscillations observed in the imaginary part of the waterfall side peak when
using alternating -pulse phases with an orthogonal preparation pulse, except where
labelled cpmg where the pulses are of constant phase. The oscillation frequencies
vary with the sparsity of the -pulses (N ) as shown in the plot on the left and with
the frequency offset as shown in the plot on the right, as expected for -pulse errors.
The oscillations are removed by using the cpmg convention. This data was taken with
polycrystalline silicon under mrev-8 with tc = 0.681 ms.

is not oscillating around the z-axis; as a result the oscillations in the real and imaginary components
are not directly out of phase, accounting for the oscillations in the magnitude.
Another important effect is observed in this data, which is that after the initial, oscillatory
decay, even the transverse component has a long-lived, spin-locked component. This occurs when
N is chosen too low (that is, when -pulses are applied too quickly.) This is due to the effect of
pulsed spin locking [206, 207]. The data in Fig. 5.5 show how the apparent decay times vary with
N using mrev-8. The spin-locked decay time is very long when -pulses are applied every N = 5
cycles, but it rapidly decreases as the rate of -pulses is reduced. The fact that this is a spin-locking
effect and not true coherence can be verified by careful observation of the phase of the signal after
many -pulses; when pulsed spin locking is present, the phase of the signal near the tail of the
decay is uncorrelated to the initial phase of the spins, as determined by the preparation pulse. We
found pulsed spin locking to be present both when we used constant phase -pulses, as in the cpmg
sequence, and when we used -pulses of alternating phase. For -pulses applied every 120 cycles,
the effects of pulsed-spin locking seem to be absent, although a very small tail in the echo decay is
still present, presumedly due to this effect.
As a final note on spin locking, these pulsed spin locking effects may still be observed without
decoupling, as shown in Fig. 5.6. This behavior is also observed if the -pulse phases are alternated.

148

Chapter 5. Decoupling: Experiment

1
Echo Magnitude

N=5
N=10
N=15
N=20
N=30
N=40
N=50
N=60

0.1
0

10

15

20

Time (sec)
Figure 5.5: Pulsed spin locking observed in heavily doped silicon wafers under cpmgmrev-16N with tc =2.46 ms. The magnitudes of the side-peak echoes are shown.

Very similar data was observed by Dementyev et al. [208] and called anomalous; in their work
pulsed spin locking was dismissed as the cause of the effect since long-lived echoes, albeit small
ones, are still observed even when -pulses are applied very sparsely. These echoes are extremely
weak, however, and show the pronounced even-odd asymmetry discussed in Sec. 4.3.3, indicating
definite pulse errors. A thorough explanation of these echo magnitudes would require a complex
simulation of dipolar dynamics in the presence of inhomogeneous static and rf magnetic fields, but
conceptually the observations do not fall far from the expectations of pulsed spin locking due to
imperfect cpmg sequences applied in the in the T2 T2 regime.

5.2

Results

Let us now return to decoupling experiments, and discuss the results of cpmg-mrev-16120 sequences after pulses have been tuned and spin-locked components have been eliminated.
As shown in Fig. 4.3, the cpmg-mrev-16120 sequence allows the observation of hundreds
of spin-echoes. Figure 5.7 shows the resulting T2 of decoupling the single-crystal samples. The
insets show the magnitude decay of the detuned echo; the data for both samples fits reasonably
well to an exponential decay, as shown, and the resulting least-squares fit for each tc is plotted.
For the isotopically enhanced sample, the T2 before decoupling is 450 s for the [001] orientation,

5.2 Results

149

1
Echo Magnitude

0.4638 ms
0.9758 ms
1.4878 ms

0.1

1.9998 ms
4.0478 ms
0

20

60
40
Time (ms)

80

100

Figure 5.6: The cpmg sequence in heavily doped silicon wafers without decoupling. The
different traces correspond to different -pulse spacings 2 . Long-lived echoes are likely
due to spin-locking effects.

97%

29

Si

4.7%

29

Si

20

T2 (sec)

10
7
5

0.3
0

0.7

1.392 ms
1.636 ms
1.876 ms
2.596 ms
3 Time (sec)

1.568 ms
2.288 ms
3.008 ms
4.688 ms
10 20 30 Time (sec)

tc (ms)

Figure 5.7: Coherence time vs. cycle time in single-crystal silicon. The solid line is a
fit showing the exponent 2.09 0.07 for the isotopically enhanced sample (left) and
2.00 0.2 for the isotopically natural sample (right). The insets show the integrated
log-magnitude of the spin-echoes decaying in time for a few cycle times.

150

Chapter 5. Decoupling: Experiment

2.8

2.6

1.0 kHz
1.5 kHz
2.5 kHz
3.5 kHz
4.0 kHz
4.5 kHz
5.0 kHz

Echo Magnitude

2.2
2.0

Time (sec)

1.2

0.1

0.01

2.4

T2 (sec)

/2

1.8
1.6
1.4

10

/2 (kHz)
1

Figure 5.8: The decay of the spin-echo peaks under mas for several rotation speeds ,
with exponential fits (left), and the observed decay times T2 plotted against (right).

as reported previously for this sample [202]. The cpmg-mrev-16120 sequence extends the T2 in
this sample to nearly 2 seconds. For long tc , the coherence time reduces as t2
c , indicating that
decoherence is dominated by second order terms in the average Hamiltonian, the result expected
from the discussion in Sec. 4.2.1. For short tc , finite pulse width effects become more important,
and the sequence fails [125]. In isotopically natural silicon single crystals, the coherence time is
even longer due to the scarcity of

29

Si in the lattice. As shown in Figs. 4.3 and 5.7, the spin-echoes

in the sample last for as long as a minute, showing a T2 of 25.0 0.2 sec. The effective Q of this

qubit, then, is 0 T2 = 109 , exceeding the Q of any other solid-state qubit by at least four orders of
magnitude.
The T2 times in single-crystal silicon were also observed under mas experiments performed by
Eisuke Abe at Keio University. In these experiments T2 was not as long as in the Stanford mps
experiments. The observed decay was exponential with T2 = 2.6 sec at the fastest spinning speeds.
This decay was independent of the -pulse timing, as expected for dipolar couplings. The T2 as a
function of rotating speed is shown in Fig. 5.8. The T2 varies roughly linearly with , as expected
from first order aht and consistent with typical mas results [209].
As the strength of the dipolar coupling is further decreased by isotopic depletion, the dipolar
coupling constants Djk become much smaller than the frequency offsets j . The dominant second-

151

Magnitude

5.2 Results

10
Time (sec)

15

20

Figure 5.9: Spin echoes for isotopically depleted silicon under cpmg-mrev-16120. The
solid line shows exp(t/8 sec), for comparison.

order dipolar average Hamiltonian term leading to decoherence is then the dipolar/offset cross term,
(2) t2c |Djk || 2 |. For isotopic percentage p less than about 10%, we would expect
which scales as H
j

T21 to be proportional to the dipolar coupling constants, which vary as the inverse cube of the
distance between

29

Si isotopes. Correspondingly, we expect T21 p, approaching T11 as p 0.

However, our attempt to observe this isotope effect in depleted samples was not successful. In the
depleted sample, the same decoupling sequence which led to T2 = 25 seconds in isotopically natural
silicon led to a decay time not exceeding 8 seconds, as shown in Fig. 5.9. This noisy data results
from 10 averages in one experiment lasting a week. We believe the reduced T2 is due to the presence
of lattice defects in the sample.
Similar data is observed in the sample of pure, polycrystalline silicon. Although the shallow
impurity content of this sample is very low, leading to T1 = 4.5 hr, the cpmg-mrev-16120 sequence
leads to a decoherence timescale of approximately 8 seconds. The higher snr for this isotopically
natural sample allowed us to study this decay more carefully. There are two unusual features of the
data, both revealed in Fig. 5.10. First, the decay curve is neither exponential nor Gaussian. Second,
this decay curve retains its shape as tc is altered. If this decay were due to residual dipolar coupling
terms of the average Hamiltonian, some change in shape would be expected. We conclude that this
decoherence is due to low-frequency noise intrinsic to the sample. Our data in isotopically depleted
silicon could be due to the same type of decoherence. In Sec. 7.3.2 we will present a theory that the
same thermal processes at defects which lead to 1/f charge noise are responsible for this unusual
data, but first, in the next chapter, I will review the formalism needed for this theory.

152

Chapter 5. Decoupling: Experiment

Echo Magnitude

Cycle time (tc)


1.338 ms
1.914 ms
2.490 ms
3.306 ms

0.1

Time (sec)
0

10

15

20

25

30

Figure 5.10: Echo decay curves for pure polycrystalline silicon of natural isotopic abundance. No significant variation in the data is observed as tc is changed. The solid line is
a fit to the function described in Sec. 7.3.2.

Chapter 6

General Relaxation Theory


Truth is much too complicated to allow anything but approximations.

John von Neumann

A model for the relaxation times observed in silicon and in any potential lattice used for quantum
computation must be derived carefully. Many of the approximations employed in the literature are
appropriate in a regime quite different from the experimental reality.
In this chapter, we will present a very general presentation of the equations originally put forth
by Kubo and Tomita [210] and apply them in the important cases of nuclear spin-lattice relaxation,
intrinsic nuclear decoherence processes, and the Overhauser effect.

6.1

Fermis Golden Rule

Before proceeding to develop a general (and fairly involved) theory, let us begin with an informal
derivation of Eqs. (3.29) and (3.30) using Fermis golden rule. Throughout this section I will be
assuming spin-1/2 nuclei.
Fermis Golden Rule, first derived by Dirac as a result of first order, time-dependent perturbation
theory, is most familiarly written as stating that the rate of transition from state i to state f is
Wif = h|hf | H1 |ii|2 (Ei Ef ).

(6.1)

154

Chapter 6. General Relaxation Theory

The hyperfine interaction Hamiltonian can be summarized as


H1 =

Dlpq
Il cp cq ,

(6.2)

lpq

where the subscripts p and q can carry momentum, spin, band, and other information. In the case

of electrons interacting via the contact hyperfine interaction, the factors Dlpq
may be read from

Eq. (A.23); however, Eq. (6.2) is sufficiently general to describe dipolar interactions as well. The
index is 0 for the z component, and 1 for the raising and lowering operators.

Using this

notation, the rate for nuclear spin l flipping between state |i i and |f i due to nuclear spin operator

Il may be written, after some algebra,


l
Wif
=h

X
X
2
hnj ih1 nk i
|Dljk
| | hi | Il |f i |2 (k j ~0 ),

(6.3)

jk

where the sum over j and k is the sum over all electron or hole states, which have energies j and k .
One further notational simplification may now be introduced. A density of states for the electrons
or holes is defined as
g(E) =

X
j

(E j ).

(6.4)

Using this definition, Eq. (6.3) can be written as


l
Wif
=h

XZ

g(E)g(E +~0 )hn(E)ihn(E +~0 )+1i|Dl (E, E +~0 )|2 | hi | Il |f i |2 dE. (6.5)

Fermis golden rule may now be applied to calculate T1 for the lth nucleus by writing
1
l
l
= W
+ W
.
T1

(6.6)

I now show how this formula relates back to Eq. (3.29). For this I will drop the l subscript to
consider only a single nucleus. We first note from Eq. (6.2) that the th component of the effective
P


local field may be simply related to Dpq
as b = (1 + 2 ) pq Dpq
cp cq . We then take a couple steps

back in the derivation of Fermis golden rule to write that

X
jk

2
hnj ih1 nk i|Dljk
| (k j + 0 )

1
= lim
T T

1 + 2

2 Z

n
o

ei0 (t t) Tr eiH0 t b eiH0 t e eiH0 t b eiH0 t dtdt

6.1 Fermis Golden Rule

1
= lim
T T

155

1 + 2

2 Z

1
= lim
T T

T
0

ei0 (t t) Tr b (t)e B (t ) dtdt

1 + 2

2 Z

ei0 (t t) hb (t )b (t)idtdt . (6.7)

We now treat hb (t )b (t)i as a correlation function. We assume the noise process described by
b (t) is statistically stationary, and thus this correlation function may be written hb (t t)b (0)i.

Make the variable switch t t t and transform the integrals as follows:


lim

1
f (t t )dt dt = lim
T T

T /2

T /2

T /2

T /2

1
T T

= lim

f (t t )dt dt
T /2

T /2

T /2+t

T /2+t

f (t)dt dt

f (t)dt. (6.8)

Including this manipulation in Eq. (6.7) and comparing to Eq. (6.3) yields
X

Wif =

| hi | I |f i |

1 + 2

2 Z

ei0 t hb (t)b (0)idt.

(6.9)

For the T1 case, then, we have


1
2
= W + W =
T1
4


ei0 t hb+ (t)b (0)i + ei0 t hb (t)b+ (0)i dt

Z 

2  x
hb (t)bx (0)i + hby (t)by (0)i cos(0 t)+
=
2



x
y
y
x
hb (t)b (0)i hb (t)b (0)i sin(0 t) dt.

(6.10)

The sine term is usually neglected because the cross-correlation term is presumed even. We are thus
left with the expression for T1 in Eq. (3.29).
To derive a similar expression for T2 , we calculate the probability of a transition between the
states
1
|x i = (|i + |i) ,
2

1
|x i = (|i |i) .
2

(6.11)

Using Eq. (6.9), then, we find

Wxx =

2
1 X
h| I |i h| I |i + h| I |i h| I |i
4

2 Z

1
ei0 t hb (t)b (0)idt = W + (W + W ) .
2 0
4

(6.12)

156

Hence

Chapter 6. General Relaxation Theory

1
1
2
= Wxx + Wxx = 2W + (W + W ) =
T2
2
2

hbz (t)bz (0)idt +

1
,
2T1

(6.13)

the same as Eq. (3.30).


This informal derivation has hidden many assumptions about the physics of the local field b(t),
but it allows us a quick derivation. We now approach these equations more carefully with a more
detailed analysis.

6.2

General Equations

In this section, I develop an open-bath formalism well suited to nmr relaxation calculations. There
have been many approaches to the topic, and in general different formalisms work better for different
systems. The formalism presented here is especially well suited to noisy solid-state environments with
nuclei at arbitrary polarizations. Once this formalism is complete, we will again derive Eqs. (3.29)
and (3.30) with more careful consideration of the approximations used. This formalism will allow
us to naturally generalize these equations to the situation where nuclear spins are rapidly rotated
by a multiple pulse sequence, as in the experiments described in the previous chapter.

6.2.1

System and Bath

The basic situation for any open-bath formalism is that there is a small system with operators
Xi and a very large bath or reservoir (I will use bath) with operators Bi . In the Heisenberg
picture, the equal-time commutators of Xi and Bi , denoted [Xi (t), Bi (t)], all vanish at all times,
meaning that the system and bath are separate systems. Of course, these Heisenberg operators may
not necessarily vanish at different times, however, because we suppose that the system and bath are
coupled.
The Hamiltonian is written
Htotal = Fsys (X) 1 + 1 Fbath (B) + Fint (X, B)
{z
} | {z }
|
{z
} |
Hsys

Hbath

Hint

with the important observation that [Hsys , Hbath ] = 0. We require that [Hsys , Hint ] 6= 0 and
[Hbath , Hint ] 6= 0; otherwise there is no energy flow between system in bath. The density operator
for the combined system and bath will be denoted , and the reduced density operator, averaged
over the environment, will be denoted .

6.2 General Equations

157

We work in the interaction picture, in which our density operator is defined as




= ei(Hsys +Hbath )t ,

(6.14)

Hint (t) = ei(Hsys +Hbath )t Hint ei(Hsys +Hbath )t .

(6.15)

where is the density operator in the Schr


odinger picture. The time dependent interaction Hamiltonian is

The equation of motion of is then


i
which is solved by





(t) = Hint (t) (t) ,
t

Z t
Z


 X
(t) =
(i)n
dt1 Hint (t1 )
0

n=0

t1

dt2 Hint (t2 )

(6.16)

tn1



The nth term of the sum will be notated n (t) .

6.2.2



dtn Hint (tn ) (0) .

(6.17)

Time Domain

We begin by deriving the basic master equation in this notation.


For this section, the average of an operator hOi refers to an average over the bath. We assume
that, by construction of the interaction Hamiltonian, hHint (t)i is zero at all times, and therefore

n = 0. The lowest order term in our series is therefore


2 (t) =

dt1

t1



dt2 Hint (t1 )Hint (t2 ) (0) .

We assume that Hint Hspin , so that under the usual assumptions of perturbation theory we may
take this equation as the time evolution for . We also employ the Born approximation, stating that
the state of the system and bath is nearly separable at all times. We may then average over the
environment to obtain

d
2 (t) =
dt



dt2 hHint (t)Hint (t2 )i 2 (t) .

Another key assumption comes into play, which is the separation of time-scales. Although the bath
operators may be rapidly fluctuating, the evolution of 2 (t) (which always has the environment
traced out) is much slower than these fluctuations. This is tantamount to assuming a Markov (i.e.

158

Chapter 6. General Relaxation Theory

memory-less) process. This allows us, after averaging over the bath, to replace 2 (0) with 2 (t). We
also manipulate the integral with the substitution = t t2 to obtain
Z


d
2 (t) =
dt

This equation is a general master equation.



d hHint (t)Hint (t )i 2 (t) .

(6.18)

For a bath in thermal equilibrium, the autocorrelation of Hint (t) is




hHint (t)Hint (t )i = Tr Hint (t)Hint (t )Z 1 (T ) exp(Hbath /kB T ) .

(6.19)

We must be careful, because Hbath usually will not commute with Hint . The consequence of this is
that the noise spectral density corresponding to this autocorrelation function will be asymmetric in
. We will encounter this situation in Sec. 6.3.3 and see that it results in the principle of detailed
balance.

6.2.3

Frequency Domain

Rather than using the time-domain differential operator equation we have derived, analysis of nuclear
relaxation is often simplified in the frequency domain. This will be especially useful when we will
make the first simplifying assumption, which is that the time dependence of Hint (t) can be written
as just a few Fourier components. Thus we write
1
Hint (t) =
2

eit Hint () =

ei t H ,

(6.20)

where we used
Hint ()

2( )H .

(6.21)

However, for the time being we will keep our results completely general by just using the Fourier
transform of Hint (t), which is always well defined.
Applying this Fourier transformed Hamiltonian to Eq. (6.17) results in


n (t) =

1
2i

n Z

d1 Hint (1 )

d2 Hint (2 )

dn1 Hint (n1 )


!
Z t
Z t1
Z tn1
X


dt1
dt2
dtn exp
il tl n (0) . (6.22)

6.2 General Equations

159

The time integrals can be written as [210]

gn (t; 1 , . . . , n )

dt1

1
=
2i

t1

dt2

tn1

dtn exp i

n
X

l tl

l=1

" 



#1
n
X
e
p p i1
p i1 i2 p i
l
dp,
pt

yielding


n (t) =

6.2.4

(6.23)

l=1

1
2i

n Z

dn n gn (t; n )

n
Y
l=1



Hint (l ) (0) .

(6.24)

Observable Averages

Equation (6.24) is an exact result. However, it is rarely convenient to solve this expression for all
n, nor is it usually a good approximation to simply use the first term and treat this as the actual
evolution of the density operator. Rather, we shall be interested in some spin observable A. In
particular, we seek the deviation of A from its equilibrium value, which we notate as A. After
averaging over the lattice, this observable is written






hA(t)i = A (t) A T = A eiHspin t (t) T .

(6.25)

where T represents the thermal density operator and is a resonant frequency. In the presence
of the relaxing perturbation, the resonant frequencies will be shifted and any oscillations will be
damped. To model this behavior, we make the ansatz

where (0, t) = 0.





A (t) T = A (0) T exp[(, t)]

(6.26)

The key step, known as the cumulant expansion, is to expand in orders of , and compare the
lowest terms to our perturbative expression for the density operator. This is a nontrivial expansion
for anything beyond the lowest non-vanishing order, but the non-vanishing lowest order terms are
usually sufficient. Thus we write
(, t) (0, t) +

2
(0, t) + . . .
2

(6.27)

160

Chapter 6. General Relaxation Theory

We now compare Eqs. (6.24) and (6.26):


Z







d1
A (t) T = A (0) T i
g1 (t; 1 ) A Hint (1 ) (0) T
2
Z
Z



d1
d2
2
g2 (t; 1 , 2 ) A Hint (1 )Hint (2 ) (0) T + O(3 )
2
2






 2 
+ 2 + O(3 ).
= A (0) T + A (0) T (0, t) + A (0) T
2

We make one simplifying assumption: if the first order term is non-zero, this means that there
is a static field applied by the lattice onto the spin, such as a chemical shift. We have usually
already incorporated this first-order term in Hspin , so we presume it is zero. We thus keep only the
second-order term, leading to the approximation


hA(t)i A (0) T
" Z


#
Z

A Hint (1 )Hint (2 ) (0) T


d1 d2


exp
. (6.28)
g2 (t; 1 , 2 )
A (0) T
2
2

6.3

Spin Relaxation

We now apply these general equations to the specific problem of the relaxation of nuclear spins.
We begin with a simple, common example, the spin-boson model, and then generalize to arbitrary
magnetic noise. The generalization will bring us back to Eqs. (3.29) and (3.30) but with conditions
under which these equations are appropriate.

6.3.1

Spin-Boson Model

As a simple example of using these equations, then, let us first calculate relaxation times using
this formalism for the popular spin-boson model. We have seen an example of this model in the
interaction of nuclei with spin-waves under the linear approximation in Sec. 3.3. In the spin-boson
model, the lattice is modelled as a collection of harmonic oscillators. Thus
Hlattice =

X
n



1
n an an +
.
2

(6.29)

6.3 Spin Relaxation

161

The interaction is
Hint =

Dnz an an I z + Dn+ an I + Dn an I + .

(6.30)

For simplicity, we presume the spin Hamiltonian describes only a static magnetic field. We also
include any first-order shifts due to the boson bath:
"

Hspin = B0

X
n

Dnz han an i

I z = 0 I z .

(6.31)

The time-dependent interaction is


Hint (t) =

X
n


Dnz an an han an i I z + ei(n 0 )t Dn+ an I + ei(n 0 )t Dn an I + .

(6.32)

This is a sum of three terms, with frequencies 0, (n 0 ). We thus use Eq. (6.20) and eliminate

the frequency integrals immediately. We notate each term as Hn , where = 0 for the I z term and
= 1 for the I terms. The exponent in Eq. (6.28) thus becomes



XX
A Hn Hn (0) T
1


.
=
g2 (t; n , n )
2
A (0) T

(6.33)

n n

Let us first calculate T1 for nuclei with low polarization using this model. Our spin observable is
A = I z . Suppose the spin begins saturated, and the lattice begins in thermal equilibrium. Then
X
1 exp(~0 I z /kB T )
(0) T
Z 1 exp
~n an an /kB T
2I + 1
n

I z T ,

where = ~0 /(2I + 1)kB T and T is the reduced thermal density operator of the lattice. Each
term of the numerator of becomes






I z Hn Hn I z T = [I z , Hn ] Hn I z T


= , (1 0 ) [Hn , Hn ] I z T

 
 

z
= |Dn |2 , nn (1 0 ) I z I z 2a
I I 1 T
n an T + I



= |Dn+ |2 , nn (1 0 ) 1 + 2han an i I z I z



~n
+ 2
= |Dn | , nn (1 0 ) coth
I z I z .
2kB T

162

Chapter 6. General Relaxation Theory

The only non-zero terms here are those for which = 6= 0, meaning that = . Therefore,
g2 (t; , ) is the sum of the residues of the function
f (p) =

ept
p2 (p i )

on the imaginary axis. Thus


g2 (t; , ) =

1 + i t ei t
.
2

Performing the sum over then gives




X 2 2 cos(n 0 )t
1
~n
+ 2
=
|Dn | coth
2 z
(n 0 )2
2kB T
n




X
~n
2 (n 0 )t
+ 2
2
= t
sinc
|Dn | coth
.
2
2kB T
n
We thus find
(



)
X
I(I + 1)
(

)t
~
n
0
n
hIz i hIz iT =
exp t2
sinc2
|Dn+ |2 coth
.
3
2
2kB T
n

(6.34)

For very short times, this is characterized by Gaussian decay with a timescale determined by the full
spectrum of bosons. If this spectrum is broad-band and continuous, however, this Gaussian decay
will be very shortlived. As the timescale gets to be long we call upon the identity
lim sinc2 (t) =

(),
t

(6.35)

in which case we find


hIz (t)i hIz iT =





I(I + 1)
~0
exp t2|D0+ |2 coth
g(0 ) ,
3
2kB T

(6.36)

where g() is the density of phonon states at frequency . We thus find


1
= 2|D0+ |2 coth
T1

~0
2kB T

g(0 ).

(6.37)

In this limit, this result is the same as what would have been obtained from Fermis golden rule. Our
new formalism, however, has done more than Fermis golden rule; it has provided us a more physical
theory for T1 decay which shows Gaussian behavior at short times making a smooth transition to

6.3 Spin Relaxation

163

exponential decay at longer times.


We can perform a similar calculation for T2 . In this case, we presume our initial deviation density
operator is
(0) = I x T .

Noting that I x T = 0, we find


s 
s
X
XX

I
I
H
H


n
n

.

hI x i =
eis0 t I s I s exp
g2 (t; , )
4
I s I s

s,s =

(6.38)

n n

When or are zero, we are using the shorthand notation


a0n = an an han an i.
The first elimination of terms comes from the trace

 2

I s I s = I(I + 1) ss .
3

Next, we note that

(6.39)







I s Hn Hn I s T = [I s , Hn ] Hn I s T




s+
=(s )(1)s+ Dn a
Hn I s T
n I



s+
s
, a
=(s )(1)s+ Dn Dn [a
n I
n I ] I T

=, (s )(1)s+ Dn Dn
h

i

s s
+ nn I I s+ I s .
(s + 2)(1)s ha
n an i I I

For 6= 0, we get an answer identical to the T1 case




I Hn 6= 0 Hn , I s T = nn |Dn+ |2 coth
s

~n
2kB T


I z I z .

For these terms the derivation follows that for T1 exactly. For = 0, we obtain

h
i




I s Hn0 Hn 0 I s T = Dnz Dnz han an an an i han an ihan an i I s I s


1 z 2
~n
2

= nn |Dn | csch
.
4
2kB T

(6.40)

164

Chapter 6. General Relaxation Theory

For these terms, the time dependence is g2 (t; 0, 0), the residue at p = 0 of the function ept /p3 . This
third-order pole leaves a residue of t2 /2. Thus,
"

#
I(I + 1)
~n
1 z (t) t2 X z 2
2
hI i =
cos(0 t) exp

|Dn | csch
.
3
2 2
8 n
2kB T
x

(6.41)

This model predicts a Gaussian, rather than exponential, decay of transverse oscillations.

6.3.2

General Formulae for T1 and T2

These results are readily generalized to relaxation caused by almost any kind of magnetic noise. The
most general linear interaction for a spin-1/2 is a Zeeman coupling to an arbitrary magnetic field,
already introduced as Eq. (3.28):
Hint (t) = b(t) I(t) = bz (t)I z


 +
b (t)ei0 t I + b (t)ei0 t I + .
2

(6.42)

There is no need to limit our discussion to spin-1/2, except that for I > 1/2, there could be couplings
nonlinear in the spin operators. We will not discuss these quadrupole-related relaxation processes
here. The only assumption we make about the three components of the random field b is again that
they fluctuate about 0:
hb(t)i = 0.
This field can be a classical field or a quantum mechanical operator. We may compactly write the
Fourier transform of the interaction Hamiltonian as
Hint () =

1
X

=1

b
(

)I
=
H (),
0
1 + 2

(6.43)

where b () is the Fourier transform of b (t).


Longitudinal Relaxation (T1 )
To calculate T1 , we consider the time evolution of I z following saturation. We thus have to work
out the following braket:



b (1 0 )[I z , I ] H (2 )
1 + 2

2


=
[b
(

)I
,
b
(

)I
]
.
1
0
2
0
(1 + 2 )(1 + 2 )



I z H (1 )H (2 ) =

6.3 Spin Relaxation

165

For this term to remain non-zero, cannot be zero, and = . This braket thus becomes

 2

I z H (1 )H (2 ) =
h{b (1 0 ), b+ (2 + 0 )}i I z
4

2
+ h[b (1 0 ), b+ (2 + 0 )]i I(I + 1) I z I z
4
2 z 
I C(1 , ).

(6.44)

A small amount of inspection reveals that this expression is invariant to the operation of swapping
simultaneous with 1 2 . We now treat the correlators:

Z Z 

b (t)b+ (t ) ei(1 0 )t ei(2 +0 )t




b+ (t)b (t ) ei(2 +0 )t ei(1 0 )t dtdt



Z Z 


=
b (t)b+ (t + ) ei(2 +0 ) b+ (t)b (t + ) ei(1 0 ) ei(1 +2 )t dtd.


b (1 0 ), b+ (2 + 0 ) =

We now assume b(t) is stationary to complete the t-integral.



Z 

+
i(1 0 )
+

i(1 0 )
d.
= 2(1 + 2 )
hb (0)b ( )ie
hb (0)b ( )ie
The -function allows us to eliminate one of our integrals. We thus find
Z
1
2 X
d1
z =
g2 (t, 1 , 1 )C(1 , )
2
4
2
=
Z
2 X
d 1 + it eit
=
C(, ),
4
2
2
=

where C(, ) is the sum of correlators satisfying C(, ) = C(, ). When we sum over , then,
we take for = 1, thus writing both terms with C(, 1). The sum over thus yields
1
2
z = t2
2
4

d
t
sinc2
2
2

Z 

hb (0)b+ ( )iei(0 )t + hb+ (0)b ( )iei(0 )t d

Z 
 #
I(I + 1) I z I z

+
i(
)t
+

i(
)t
0
0

+
hb (0)b ( )ie
hb (0)b ( )ie
d .
I z

166

Chapter 6. General Relaxation Theory

To obtain T1 , we take the long time limit and find


1
2
z t
2
4

Z 


+
i0 t
+

i0 t
hb (0)b ( )ie
+ hb (0)b ( )ie
d

Z 
 #
I(I + 1) I z I z

+
hb (0)b+ ( )iei0 t hb+ (0)b ( )iei0 t d .
I z

We thus find a sum of two terms. The first is


1
2
=
T1
4


Z 

+
i0 t
+

i0 t
hb (0)b ( )ie
+ hb (0)b ( )ie
d.

(6.45)

This may be rewritten in terms of the components of b as


2
1
=
T1
2


hbx (0)bx ( )i + hby (0)by ( )i cos(0 t)+



hbx (0)by ( )i hby (0)bx ( )i sin(0 t) d.

(6.46)

It is difficult to imagine a case where the second term is nonzero, and thus it is omitted, leaving us
once again with Eq. (3.29).
To this we may need to add a second term that depends on I and the temperature by the function

F (I, ) = I(I + 1) I z I z .

(6.47)

For = 1/(2I + 1) T , this works out to



2

I +1 4
F (I, ) = (I + 1) coth BI ()
I(I + 1) 1 2 ,
3
2
30
3

(6.48)

a very small number at most temperatures and zero when I = 1/2.


Let us apply these equations in a couple of examples. The first we have already encountered:
the spin-boson model, where
b+ (t) =

2 X + in t
D a e
.
n n n

Plugging directly into Eq. (6.45) yields


X
1
=
|Dn+ |2 2(n 0 )h{an , an }i = 2|D0+ |2 coth
T1
n

~0
2kB T

g(0 ),

in agreement with Eq. (6.37). The correction term, which we omitted previously under the guise of

6.3 Spin Relaxation

167

a high-temperature approximation, works out, in this case, to


F (I, )

X
n

|Dn+ |2 2(n 0 )h[an , an ]i = 2F (I, )|Dn0 |2 g(0 ).

For I > 1/2, then, the complete spin-boson model predicts


h

i
1
= 2|D0+ |2 g(0 ) coth
+ F (I, ) .
T1
2
A more important example is when the fluctuating field is itself another spin, perhaps an electron
with spin vector S, which itself fluctuates and whose coupling constant fluctuates:
b+ (t) = A(t)S + (t),
for which
Z


1
2
=
hA(0)A( )i hS S + iei(e 0 ) + hS + S iei(e 0 ) e| |/T2e d
T1
4
Z
2
=
h(S x )2 + (S y )2 i
hA(0)A( )i cos[(e 0 ) ]e /T2e d.
4
0

(6.49)

The correction term is now quite important. It is



Z
I(I + 1) I z I z 2 z

hS
i
hA(0)A( )i cos[(e 0 ) ]e /T2e d.
2
I z
0

(6.50)

This is the crucial coupling term for Overhauser effects, where no longer describes saturated
nuclear polarization. We will return to this term using a master equation approach in Sec. 6.3.3.

Transverse Relaxation (T2 )


The above formalism works well when we are considering the dynamics of only nuclei coupled to
an equilibrium environment. When considering T2 in the solid-state, the dominant mechanism is
the dipolar coupling. The cumulant expansion formalism still has bearing for nuclear spins close
to maximal entropy. As discussed in Sec. 3.1, most of the nuclear states are near mz = 0, where
the Hilbert space of eigenvalues of M z and HD may be approximately treated as separable. Thus
the space of M z eigenstates is chosen as the system and the space of HD eigenstates is chosen as

the bath. In this case the dipolar energies are so low that at finite temperature we can consider
the bath to be in the maximally mixed state . With these assignments, a calculation of T2 due

168

Chapter 6. General Relaxation Theory



to the dipolar Hamiltonian would assume an initial deviation density operator (0) M x .

The transverse component M x is zero in thermal equilibrium, and the secular dipolar Hamiltonian

is time independent. Resultingly, the cumulant expansion yields




!
 2

x
x
t
t2 M x HD HD M x
x



hM (t)i = M M exp
= M M exp M2 ,
2
2
M x M x
x

(6.51)

where M2 is exactly the second moment derived in Sec. 3.1 as Eq. (3.24). Thus we see that the
cumulant expansion reproduces Van Vlecks method of moments in the Gaussian approximation.
In the case that the bath is separate from the nuclei, we may use the cumulant expansion method
to derive Eq. (3.30). In this case we are again interested in the average of the transverse components
of the nuclear spin with the initial deviation density operator = I x T , so the important


braket is I x H (1 )H (2 ) . For the case 6= 0, this becomes



2
[b (1 0 )I z , b (2 0 )I ]
2
2
(1 + )(1 + )



2 h
=
h{b (1 0 ), b (2 0 )}i [I z , I ] +
8
i


h[b (1 0 ), b (2 0 )]i {I z , I } .



I x H (1 )H (2 ) =

The assumption = I x made a high



temperature approximation, which makes this term proportional to {I z , I } I x = 0. However,

The second term here deserves some clarification.

if we had not made this high-temperature approximation and had instead used = exp(I x ) 1,

then this term is only exactly zero if I = 1/2. For I > I/2, this term is present but very small, of
order 3 , smaller than even the F (I, ) terms in the T1 calculation. We will now neglect this term.
If we just consider the first term, then, we find




2
I x H (1 )H (2 ) = h{b (1 0 ), b (2 0 )}i I x .
8

The terms with = are the same as the first term in Eq. (6.44), except smaller by a factor
of 2. If we proceed with the calculation for T2 , we find this term generates 1/2T1 , the so-called
lifetime broadening term. However, the terms with = also make a contribution, which will
involve the magnetic spectral noise density at 20 . These terms could be measurable in situations
where T1 T2 , but for the situations described in this dissertation all of these 0 dependent terms
are negligible.

6.3 Spin Relaxation

169

For the case = 0, only terms with = 0 are nonzero, and the braket reduces to



I x H0 (1 )H0 (2 ) = i 2 [bz (1 )I y , bz (2 )I z ]
=



2  z
h{b (1 ), bz (2 )}i I x + h[bz (1 ), bz (2 )]i {I y I z } .
2

(6.52)


Again, the term I y I z is negligible for small polarizations, and exactly zero for I = 1/2. The

remaining term may be formally seen, then, to lead to the result of Eq. (3.30) on page 65. However,
it is more convenient to leave the braket in the present form,


Z
Z
t
2 d1 d2
hI x (t)i hI x (0)i exp

g2 (t; 1 , 2 )h{bz (1 ), bz (2 )}i ,


2T1
2 2 2

(6.53)

and apply it to several important examples.


The first example was the spin-boson model, where
bz () = 2() 1

X
n

Dnz (an an han an i).

This case was treated above, and shown to give a Gaussian decay. This is an example of effectively
inhomogeneous broadening for which formula Eq. (3.30) does not apply.
The second example is the case where bz is the hyperfine field due to a spin-1/2 electron,
bz (t) = A(t)[S z (t) hS z i],
in which case we may complete the integrals to find
1
1
2
=
+
T2
2T1
2



hA(t)A(0)i hS z (t)S z (0)i hS z i2 dt.

(6.54)

If the electron is itself obeying the Bloch equations, this reduces to


1
1
2
=
+ (1 p2e )
T2
2T1
8

hA(t)A(0)iet/T1e dt.

(6.55)

A third example is one where bz (t) is a classical, stationary noise process. In this case we again
use the fact that bz (t) is stationary to make the manipulation
z

hb (1 )b (2 )i = 2

hbz ( )bz (0)i cos(1 )(1 + 2 )d.

(6.56)

170

Chapter 6. General Relaxation Theory

Completion of the 2 integral removes (1 +2 ) and puts the time dependence of as g2 (t; 1 , 1 ).
The remaining terms of the 1 integral are strictly even function of 1 , as seen in the above equation,
so the only important contribution of g2 (t; 1 , 1 ) is its component which is even with respect to
1 . We have already seen that this component is t2 sinc2 (1 t). We thus encounter the integral
integral =

d1 t2
sinc2 (1 t)hbz ( )bz (0)i cos 1 .
2 2

(6.57)

Letting c be the correlation time for the random variable bz (t), we identify two important limits.
The first is t c ; this is the example of decoherence when the bath is mostly constant in time, as in
the case of the spin-boson model. In this case the sinc function is broader than the autocorrelation
function of bz and it may be ignored; the integral of over the cosine gives (1 ), cancelling the
1 integral, and so the time dependence of this function goes as t2 /2. The second limit is t c ,
in which case the sinc function is very narrow and approaches the limiting function (/t)(1 );
in this case the time dependence looks to be linear in t. These limits are both generated by the
approximation
integral

(t )hbz ( )bz (0)id.

(6.58)

Again, in the limit t c , the autocorrelation function is roughly constant over the integral and
a t2 /2 time dependence results. For t c , the factor t is effectively t over the range of the
integral and may be pulled out of the integral, leaving a time dependence which goes as t.
Using this approximation and neglecting lifetime broadening, Eq. (6.53) may be summarized as
 2Z t


z
z
hI (t)i hI (0)i exp
(t )hb ( )b (0)id ,
2 0
x

(6.59)

a result we may now use to calculate T2 in the presence of pulse sequences.

The Cumulant Expansion in the Toggling Frame


env (t), contains
The environment-coupling Hamiltonian in the rotating, toggling reference frame, H
two kinds of terms, which vary by the frequency at which the constituent spin operators oscillate due
to the rotating and toggling frame transformations. There are the longitudinal terms, proportional
to Ijz , and the transverse terms, proportional to Ij (t). (We specify an individual nucleus j in
order to later consider a model where each nucleus sees a different fluctuating field b(rj , t).) All
components oscillate according to the periodic rotations incurred by the pulse sequence; for these
oscillations, we imagine expanding each spin operator in a Fourier series. For example, Ijz in the

6.3 Spin Relaxation

171

rotating, toggling frame is written

URF
(t)Ijz URF (t)

2nit/tc
Az
Ij .
n e

(6.60)

n= =x,y,z

The transverse spin components in the rotating frame, however, still oscillate at 0 , and thus
the transverse components are written

URF
(t)Ij (t)URF (t) = eit

2nit/tc
A
Ij ,
n e

(6.61)

n= =x,y,z

where A = Ax + iAy . Thus, Eq. (6.59) may be expanded as


hIjx (t)i

!
Z t

s
i(s0 +2n/tc )t
Asr
2 X
n
r
= exp
(t ) b (rj , t)b (rj , 0) e
d . (6.62)
2 n= s,r=0,1 1 + s2 0

We now assume that the components of b(rj , t) are uncorrelated, and assume cylindrical symmetry
about the magnetic field, simplifying the sum over coordinates to

hIjx (t)i

"

Z t

2 X
zz
= exp
A
(t ) hbz (rj , )bz (rj , 0)i e2nit/tc d +
2 n= n 0
#
X As,s Z t

s
i(s0 +2n/tc )t
n
s
(t ) b (rj , )b (rj , 0) e
d
.
2
0
s=

(6.63)

As we have seen previously, the second term may be recognized as a T1 term, sensitive only to
magnetic noise around 0 .
We will use this result in the next chapter to discuss decoherence due to 1/f noise during
decoupling.

Dipolar Longitudinal Relaxation (T12 )


Before concluding this section, it is worth considering one final case. At low fields, there is a
contribution to T1 which is often neglected. In certain applications in crystalline solids where spinlattice relaxation is very slow, this process can dominate the spin-noise fluctuations. The process
is summarized by the flip of a spin with the Zeeman energy compensated for by dipolar coupling
energies and angular momentum compensated by the negligible net motion of the sample. The
(1)

process consists of transitions between dipolar eigenstates caused by HD

. (The m = 2 terms

also contribute, but their influence is less important due to the larger energy compensation which

172

Chapter 6. General Relaxation Theory

must occur). Using the cumulant expansion, we find the autocorrelation function of M z (t) =
M z (t) hM z i due to this noise process as

2 sin2 [( )t/2]
X
hM z (t)M z (0)i

n
n
0
(1)

=
exp
+
1,
n
|
H
|m
,
ni

hm

z
z
D
hM z (0)2 i
[(n n 0 )/2]2

mz ,n,n




z
z
2
hmz + 1, n | M |mz + 1, n i hmz , n| M |mz , ni /hM (0) i .
(6.64)
In the infinite temperature limit the second line is 1 = [N I(I + 1)/3]1 . Finite temperature
corrections are of order ~0 /kT . For short times, we see a Gaussian function with
1 n (1) (1) o
1
= Tr HD HD
2
221

(6.65)

and in the long time limit, when the sinc function becomes a -function, we see an exponential decay
with memory time
1
2
=
T21

2
X

(1)
hmz + 1, n | HD |mz , ni (n n 0 ),

(6.66)

mz ,n,n

(0)

consistent with Fermis Golden Rule. Since the eigenvalues and eigenstates of HD are difficult to
calculate for many spins, we follow Ref. 211 to combine this formula with the method of moments
as follows:
1
1
=
T21

mz ,mz ,n,n

(1)

(1)

hmz + 1, n | HD |mz , nihmz , n| HD

|mz + 1, n i ei(n n 0 )

n (0)
o
(0)
(1)
(1)
d Tr eiHD HD eiHD HD
ei0



Z
i
2 h (0)
1
(1) (1)
(0)
(1)
(1)
=
d Tr HD HD
HD , [HD , HD ] HD + ei0

2
Z

n
o
 r
2
1
(0)
(1)
(1)
2 2
0
02 /421
M21

d
exp

Tr
[H
,
H
][H
,
H
]

=
,
0
12
D
D
D
D
2
2 M e
221
4
2

12

(6.67)

(0)

(1)

where, in the last step, M21 = |[HD , HD ]|2 . This time scale is of the order 1/|HD | at vanishing
applied fields, but it rapidly increases when the applied field exceeds |HD |/ 100 G. This timescale
is therefore important in adiabatic demagnetization experiments, where it quantifies the time for
relaxation from Zeeman order to dipolar order at vanishing fields [212, 213].

6.3 Spin Relaxation

6.3.3

173

Two Hyperfine Coupled Spins

In this section, we suppose that we have both nuclear and electron spins, neither of which are
necessarily in equilibrium, but which nonetheless undergo those dynamical processes that tend to
take the system into equilibrium. With these simple assumptions, we will derive the basic equations
describing Overhauser effects. For this problem we leave the cumulant expansion approach and
develop rate equations for the deviation of from T , . For this, it is worth returning to the
time-domain master equation of Eq. (6.18). The assumptions are well discussed in Abragam [121],
although Abragam introduces a high-temperature (low-polarization) assumption which is suspended
in the following derivation.
We assume the hyperfine interaction again takes the form
Hint (t) = A(t)[I + S eit + I S + eit ].

(6.68)

The I z S z term is considered part of the unperturbed (system) Hamiltonian.



The master equation, Eq. (6.18), for I z is therefore

 

[I z , A( )I + S ei + A( )I S + ei ], A(0)(I + S + I S + )]
0

Z t 

1
=
d ei hA( )A(0)i + ei hA(0)A( )i I z S + S I + I S z
2 0

+i sin( )h[A( ), A(0)]i I I + S + S .

d z 
1
I =
dt
4



Note that we have made no assumptions about the temperature or about the nuclear spin. The first
simplifying assumption is that S = 1/2, so that
S+S =

1
+ Sz.
2

(6.69)

This makes no presumptions about the temperature. Frequently, one sees a high temperature
approximation made for the nuclei:
hI + I i

2
2
I(I + 1) + I z I(I + 1).
3
3

(6.70)

This assumption, however, fails for high-I nuclei at low temperature. Instead, we simplify slightly

174

Chapter 6. General Relaxation Theory

by writing
I + I = I x I x + I y I y + I z = I(I + 1) I z I z + I z ,

(6.71)

I + I = I x I x + I y I y I z = I(I + 1) I z I z I z ,

(6.72)

yielding
d z 
1
I =
dt
2


Iz
z z
z
d cos h{A( ), A(0)}i
(I(I + 1) I I )S
2


I(I + 1) I z I z
z z
+ i sin h[A( ), A(0)]i
I S . (6.73)
2

Let us analyze this equation more carefully in the case of I = 1/2. Then we have four states, ordered
in energy, from lowest to highest as
|1i = |e n i
|2i = |e n i
|3i = |e n i
|4i = |e n i .
To convert from the population expectations ni = hi| |ii to the spin operators I z and S z , we have
the equations
n3 + n4 = hS z i + 1/2,

(6.74)

n2 + n4 = 1/2 hI z i,

(6.75)

n2 + n3 = 2hI z S z i + 1/2.

(6.76)

We then obtain the equations


n 1 =

1
2

d h{A( ), A(0)}iei n1 + h{A( ), A(0)}iei n4 ,

(6.77)

n 4 = n 1 , n 2 = n 3 = 0. This leads to the natural definition for the rate constants


G(, T ) =

1
2

d h{A( ), A(0)}iei .

(6.78)

We have extended the limits of the intergral to under the assumption that we are studying

6.3 Spin Relaxation

175

dynamics on timescales much slower than the correlation time of the modulation of the hyperfine
field given by A(t). We now derive, by example, the principle of detailed balance. This example
could be easily extended to a more general set of bath operators. In this example, we use the form
Eq. (A.23) for the hyperfine interaction, and assume our electron is free with wave vector k and
remains within its band. We then find
A(t) =

1 X
(k k)ei(k k)R ck ck ei(k k )t .
N

(6.79)

Thus we find

G(, T ) =

2 X
|(k k)|2 ( k + k )hnk (1 nk ) + nk (1 nk )i
N
k
2 X
|(k k)|2 ( k + k )hnk (1 nk )i(1 + e ).
=
N

(6.80)

For thermal fermionic distributions, the key observation from this expression is that
G(, T ) = e G(, T ).

(6.81)

This is the principle of detailed balance. For convenience, then let us define
G(, T ) =

K
.
1 + e

(6.82)

Returning to Eq. (6.73), we find a commutator term which, in this model, may be written
Z

2 X
|(k k)|2 ( k + k )hnk nk i
N
k
2 X
=
|(k k)|2 ( k + k )hnk (1 nk )i(e 1)
N
k


~
= G(, T ) tanh
,
2kB T

h[A( ), A(0)]iei d =

and thus we have


 
 z 



 


d
I
~
I(I + 1) I z I z
.
I z = K
I(I + 1) I z I z S z tanh
Iz Sz

dt
2
2kB T
2
(6.83)


The equation for S z may be found by noting that [I z + S z , Hint ] = 0, and therefore the sum

176

Chapter 6. General Relaxation Theory

of the two spins is strictly conserved. Thus


d z 
d z 
S =
I .
dt
dt

(6.84)

These are the general Solomon equations for the coupled electron and nuclear magnetizations.
The simplest treatment is to assume I = 1/2 and, in a high-temperature approximation, that

I S z = 0. Then the steady state solution is
z


1
I = hS z i + tanh
2
z

~
2kB T

hS z i hS z iT ,

(6.85)

showing that if the electron spin polarization is held away from equilibrium, then the nuclear spin
polarization will be held away from equilibrium by the same amount. The traditional Overhauser
effect holds hS z i at zero, leading to a nuclear polarization equal to the thermal electron polarization,
which is a factor of roughly one thousand more than the nuclear polarization due to the electrons
larger magnetic moment.
While this simple theory has been very successful at describing the dynamics in metals, it is
insufficient to describe the optical Overhauser effect in bulk semiconductors. Optical pumping
experiments in silicon and GaAs have showed that multiple electron species are involved [214]. In
particular, the growth of the nuclear polarization in GaAs under laser illumination does not seem to
require spin-diffusion. In addition to other evidence, this strongly suggests that nuclear polarization
in bulk semi-insulating GaAs is principally caused by delocalized carriers. However, almost no
nuclear polarization is observed in very pure samples of GaAs. This suggests that some impurity in
semi-insulating GaAs is needed for nuclear polarization. This impuritys role is further demonstrated
by the fact that the nuclear polarization changes sign upon tuning the laser energy through 1.495
eV or applying a large electric field. Rate equations beyond the equations I have presented here
are needed to explain this sign change. The detailed modelling of the complicated dynamics in real
optical pumping experiments is an ongoing direction of research [215].

Chapter 7

Nuclear Relaxation in Silicon


Experience without theory is blind, but theory without experience is mere intellectual play.

Immanuel Kant

In this chapter, I will survey the theory and experiment of T1 and T2 in silicon. This will involve
several related calculations, all involving the interactions between nuclei and electrons or holes.
These electrons and holes could be in thermal equilibrium or optically excited. They could be
trapped by shallow or deep impurities or delocalized in their corresponding band. There are thus
many cases, but I hope to summarize many such calculations using a common formalism. I will
conclude the chapter by discussing a model whereby the tc -independent decoherence measurements
shown in Chapter 5 are explained by 1/f charging noise.

7.1
7.1.1

Free Carriers
Conduction Electrons

The easiest place to begin is with the discussion in Abragam [121], which follows the important
papers of Shulman and Wyluda [216] and Bloembergen [217]. The purpose here is to calculate T1
due to free conduction electrons from ionized n-type donors (i.e. phosphorous). In this case it is
clearly the Fermi contact term of the hyperfine coupling that will provide the local field.
For this calculation we may use Fermis golden rule directly, as in Sec. 6.1. In this case, the

178

Chapter 7. Nuclear Relaxation in Silicon

matrix elements are given by the hyperfine coupling described in Appendix A.5:

Dlkk
=

ee (k k)ei(k k)Rl
.
2N

(7.1)

For this metallic case, the e subscripts in ee (k k) refer to the electrons at the Fermi surface.
As usual, we ignore the k dependence of ee (k) and the exponential factor plays no role in this first
order calculation. Hence these numbers are independent of k, and by extension, of E. The thermal
occupation factors for Fermi-Dirac statistics are
hnk ih1 nk i = f (k )[1 f (k )] f (k )[1 f (k )] kB T (k EF ).

(7.2)

The energy due to both the electron spin and the nuclear spin are ignored in comparison to the
Fermi energy. Since all expressions are functions only of energy, we use Eq. (6.5) on page 154, which
yields
1
= 4~
T1

g 2 (E)kB T (E EF )

1
~
2
2
|ee (0)| dE = 2 g 2 (EF ) |ee (0)| kT.
4N 2
N

(7.3)

For free electrons, the density of states at the Fermi energy is g(EF ) = 3N/4EF , yielding
2

1
9~ |ee (0)| T
=
,
T1
16 kB TF TF

(7.4)

showing the characteristic metallic T dependence, about which more details may be found in
2

Abragam [121]. Note that |ee (0)| varies as the square of the conduction electron concentration
2/3

2/3

ne , and TF ne . Consequently T1 ne

, a relation we will revisit shortly.

The important difference between the calculation above, which is applicable to metals, and a
similar calculation for n-type silicon is the occupation statistics of electrons in the conduction band.
The key to moving from a metal to a semiconductor is to approximate the electrons statistics, for
small spin-flip energy deviations , in the following form:


ne E
P (E) = f (E) 1 f (E )
e
,
Nc

(7.5)

and likewise for P (E); we will notate them all simply by P (E). Here, f + (E) = hnk i, ne is the
volume density of donors, and the relative density Nc is a normalization defined so that
Z

P (E)g(E)dE =

1
V ne ,
2

(7.6)

7.1 Free Carriers

179

where the 1/2 is because the integral counts only one electron spin.
Paying careful attention to the spin degrees of freedom, Eq. (6.5) becomes
W
W
W

Z
~
2
|ee (0)|
g(E)g(E ~e + ~0 )f + (E)[1 f (E ~e + ~0 )]dE
=
2N 2
Z
~
2
=
|
(0)|
g(E)g(E + ~e ~0 )f (E)[1 f + (E + ~e ~0 )]dE
ee
2N 2
Z
X
~
2
2
|
(0)|
g
(E)
f (E)[1 f (E)]dE.
=
ee
8N 2

(7.7)
(7.8)
(7.9)

Incorporating Eq. (7.5) yields


Z
~
2 ne
|
(0)|
g(E)g(E ~e + ~0 )eE dE
ee
2N 2
Nc
Z
~
2 ne
=
|ee (0)|
g(E)g(E + ~e ~0 )eE dE
2N 2
Nc
Z
~
2 ne
=
|
(0)|
g 2 (E)eE dE.
ee
4N 2
Nc

W =

(7.10)

(7.11)

(7.12)

In order to proceed, we need an expression for g(E). Let us presume that the number of dopants
is sufficiently high that we may use the free electron density of states, but sufficiently low that there
is no impurity band. Then we use the average density of states
g(E) = l

V
(2m1 m2 m3 E)1/2 KE 1/2 .
2 2 ~3

(7.13)

The factor l comes from the influence of the l = 6 conduction band minima in silicon, and the masses
m1 , m2 , m3 are the anisotropic effective masses. The normalization is then given by
2
Nc =
V

g(E)e

2K
dE =
V

1/2 E

K
l
dE =
(kB T )3/2 =
(mk
B T )3/2 . (7.14)
V
2~3

The W integral, which directly yields T2 , is readily performed:


W = ne l

2
2 |ee (0)| p
2m1 m2 m3 kB T .
8 2 ~2

(7.15)

It is not hard to see that for B0 0, W = W = 2W , and that in this limit


1
1
2 |ee (0)|2 p
|ee (0)|2 2
=
= 4W = ne l
2m
m
m
k
T
=
~
ne N c ,
1
2
3
B
T1
T2
2 2 ~2
kB T

(7.16)

which is exactly the expression given by Abragam [121], Ch. IX, Eq. (62), albeit with different

180

Chapter 7. Nuclear Relaxation in Silicon

notation. For silicon,


2
h
i2 h
i2
2 |ee (0)|
0
0
2
=
g0 B V |ue (0)|
=
g0 B
2
(2)
3
3

2
2
3

2
4
1 T cm
60 MHz
24 J

2 9.27 10
2
186 = 2.5 1020 cm3 MHz .
3 10
J
T
7T

(7.17)

In this equation, I introduced the parameter1 = V |ue (0)|2 as discussed in Abragam [121]. At room
temperature,
h
1
=
,
kB T
6.2 THz

Nc = 2.8 1019 cm3 ,

(7.18)

yielding
T1 = 5 1019 sec cm3 n1
e .

(7.19)

Using these numbers, Eq. (7.16) agrees well with the data of Shulman and Wyluda [216] as well as
the data we have taken in our own lab.
At higher field, T1 will become longer, but T2 will remain unchanged. It follows that T2 may be
estimated by T1 at low field in this case. (This will generally be true when the local field is due to
the contact hyperfine of electrons which may freely scatter off nuclei). One enters the high field
regime, where T1 > T2 , when c 1/0 . Therefore, for this system, the high field regime is not
available in reasonable magnetic fields and temperatures. Even the first order correction in ~e /kB T
for 1/T1 vanishes.
In an early paper on the subject, N. Bloembergen [217] derived a similar expression for T1 in doped
semiconductors and interpreted the result in the following semi-classical picture. The relaxation rate
for nuclei may be regarded as the probability of a nuclear spin-flip per collision with an electron,
multiplied by the number of collisions per unit time. The number of collisions per unit time is on the
p
order of ne d2 vth , where d = 1/3 is the atomic distance and vth = 3kB T /m is the average thermal

velocity of a Maxwell-Boltzmann gas. The probability of a collision is roughly the square of the
interaction strength ee (0) times the time of an interaction. The time of the interaction is d/v, but

Bloembergen argues that this v is not the thermal velocity but rather a quantum-limited velocity for
an electron passing an atomic site. Since d is an atomic scale, the momentum of a particle occupying
1 A further note about this constant: I have always used the approximation that u (r) is well localized in a unit
e
cell. Many derivations approach this approximation by replacing the volume factor V , using
Z
Z
Z
1
1=
d3 r |ue (r)|2 N
d3 r |ue (r)|2 = N
d3 r |ue (r)|2 V hu2e i ,

V

so that V is replaced by the inverse of the average of ue (r) about the unit cell.

7.1 Free Carriers

181

a particular site is on the order of ~/d, and therefore the relevant velocity is on the order of ~/md.
Thus the collision (or correlation) time is md2 /~, and therefore the total relaxation rate is

2

1
md2
ne d2 vth ee (0)
,
T1
~

an expression with the correct dependencies on the physical parameters.


Both Abragam et al. [218] and Sapoval and Lepine [219] made the unexplained experimental
observation that T1 shortens drastically at very low fields. Abragam et al. [218] examined a sample
with ne = 5 1016 phosphorous dopants per cubic centimeter, and found that T1 remained roughly
constant at 5 min between 50 and 10,000 Gauss at 77 K [218]. However, T1 reduced to only 30 sec
in the Earths magnetic field ( 0.5 Gauss). Sapoval and Lepine [219] examined a very pure sample,
with ne < 1014 cm3 , and found T1 = 290 min to be due to unknown paramagnetic impurities.
They found T1 to be constant as a function of field until the field became as low as the local dipolar
field of 0.176 Gauss, at which point T1 reduced to 103 min. This is probably due to the T12 effect
discussed in Sec. 6.3.2.
The data of Shulman and Wyluda [216] show that T1 1/ne at room temperature for ne

10

17

1018 cm3 , as predicted above. In contrast, the data of Sundfors and Holcomb [220] show

that T1 2/3 for higher dopant ranges at low temperature (1.6 K), characteristic of metals

as discussed above. Moreover, Sundfors and Holcomb report that T1 1/T , corresponding once
again to metallic behavior. The threshold between metallic and semiconductor behavior is near
ne 5 1018 cm3 , although at such densities the system may still reasonably obey Boltzmann
statistics at room temperature. A summary of the use of nuclear T1 studies in silicon to elucidate
the semiconductor/metal transition is given by Jerome [221].

7.1.2

Valence Holes

The physics of the correlation time for hyperfine field fluctuations due to valence holes is the same
as for conduction electrons. The only difference is in the magnitude of the hyperfine field.
To quantify the difference, we follow Grncharova and Perel [222]. The assumption is made, as
usual, that we can ignore the contact hyperfine coupling for the holes. However, the dipole coupling
terms of Eq. (A.13) still couple the nuclei to the holes. Hence the matrix elements for use in Fermis
golden rule, as in Sec. 6.1, will be

Dnkn
k




L S
0
1
r (S r)

=
g0 B
nk
+3
n k .
4
1 + 2
r3
r5

(7.20)

182

Chapter 7. Nuclear Relaxation in Silicon

The quantum number n indicates the band, which carries the spin and orbital angular momentum
information. It is argued that we may neglect the light hole and split-off bands because the density
of states in those bands is much lower. For the heavy-hole band with k k h001i, the holes have
J = 3/2, mJ = 3/2. For an arbitrary direction of k, we may simply rotate the state as |hh, ki =
exp(iJz k ) exp(iJy k ) |hh, k
z i. The result is that the effective field is reduced by a factor
hB 2 iheavy hole
hB 2 ielectron

16
=
45

1
r3

(0 g0 B )2

hh

|ee (0)|

1 hr3 ihh
,
5 |e (0)|2

in addition to the corrections to the effective masses.


While ee (0) may be estimated via the hyperfine splitting, a more indirect measure is needed to
know hr3 i. Grncharova and Perel report that decent experimental agreement is had with the T1
data of Shulman and Wyluda [216] using the Goudsmit expression


1
r3

,
320 Z

where Z = 11 for silicon and is the spin-orbit splitting of the split-off band, 4 meV for silicon.
Grncharova and Perel also report a discrepancy in the low temperature data of Sundfors and
Holcomb [220], possibly due to partial localization of holes. In fact, this scattering calculation
is not correct for lightly doped silicon at low temperature; in that case the paramagnetism of the
localized impurity states plays a more important role. The data of Sundfors and Holcomb looks
metallic at low temperature only because the doping levels are extremely high.

7.2
7.2.1

Trapped Carriers
Stationary Impurity Donors

The T1 due to a trapped impurity is calculated using the methods of Sec. 6.3.2. We presume the
impurity donor to be trapped in the strictly spatial wavefunction
(r) =

X
k

n (k)un (k, r)eikr

ue (0)F (r).

Note that the envelope function F (r) in silicon is anisotropic and has six minima.
Again we presume the contact hyperfine term dominates, in which case the effective magnetic

7.2 Trapped Carriers

183

field seen by the nucleus at position r is


b(r, t) = 1 ee (0)|F (r)|2 [S(t) hSi] .
The electron spin fluctuation function is related to its spin relaxation times, so we write for the
magnetic noise spectral density [see Eqs. (3.29),(3.30),(6.49) and (6.55)]
Ji (r, ) =


1
nd
i
2
|ee (0)| 2 |F (r)|2
1 ik p2e
,
2
4(~)
ne
1 + [( i e )i ]2

(7.21)

where the factor nd /ne is the probability that the impurity at r = 0 has trapped its carrier. The
inverse correlation time i1 is a sum of the thermal excitation rate exc and the electron spin-flip
1
1
rate T1e
for i =k or the electron phase-flip rate T2e
for i =. The spin-lattice relaxation time for

a nucleus at position r with respect to the impurity is therefore


1
1
nd
T2e
= 2 |ee (0)|2 2 |F (r)|4
,
T1 (r)
4~
ne 1 + (e T2e )2

(7.22)

where the approximation e 0 e has, as usual, been made.


A sidenote is in order here, regarding the following question: restricting our discussion to the
coupled spin system at zero temperature, is there still nuclear relaxation due to frozen electron
spins? In the case of T2 , the answer is clearly seen as no, due to the factor 1 p2e . Frozen electrons
always provide the same longitudinal field at the nuclei. However, Eq. (7.22) would indicate that
even at zero temperature, there can still be a fluctuating transverse field. How can this be if all
spins are aligned to the magnetic field? The answer is that there is zero-point motion of the spins.
p
The full angular momentum of the electron is ~ S(S + 1), but only the projection ~S is parallel to
the z-direction. For general spin S, the mean-square transverse component is 1/2(S + 1) of the full

moment in the ground state; for spin-1/2, this fraction is maximal at 1/3. Still another question is
whether T2e is finite at zero-temperature. The answer to this is probably yes as well, but that is
beyond the scope of this note.

To estimate the temperature dependence of the probability nd /ne given a concentration nexc of
thermally or optically excited electrons, we follow Bagraev et al. [223] to introduce a pseudo-Fermi
energy EF such that
nexc = 2

g(E)e(EEF ) dE = Nc eEF .

(7.23)

184

Chapter 7. Nuclear Relaxation in Silicon

The number of electrons trapped by impurities is simply


nd =

g(E)f (E) dE =

ne (E Ed )

1
ne
dE =

,
1 + e(EEF )
1 + e(Ed EF )

(7.24)

where the donor ground state Ed is negative. (We have been setting the zero of energy at the bottom
of the conduction band). Note that there is no factor of two for zero-field spin degeneracy, since
Coulomb repulsion prevents two spins from occupying a single impurity. Solving for EF yields
nd
nexc
=
.
ne
nexc + Nc eEd

(7.25)

The spin relaxation times and the thermal relaxation time will compete for intermediate temperatures (e.g. 77 K). Let us begin with the probability for thermal excitation. According to Bagraev
et al., in thermal equilibrium the rate of excitation must be equivalent to the rate of capture, given
by
nd exc = nexc hvicap (ne nd ) =

nexc
hvicap Nc eEd = nd hvicap Nc eEd .
nexc + Nc eEd

(7.26)

With these dependences in place, Bagraev et al. proceed to use the solution to a free spin-diffusion
equation, assuming uniformly distributed nuclei separated by a, to obtain the result
"

2 #
1
a2 1

= ne aB D 1 + 1 + ln
,
T1
D T1 (0)

(7.27)

with T1 (0) given by Eq. (7.22). This is clearly an awkward expression and does not agree quantitatively with their experimental results of T1 vs. nexc , the latter being modified by optical excitation.
There is qualitative agreement, however.
For now, let us leave aside the issue of spin diffusion and consider only a nucleus near an impurity.
For this case the fundamental relation Eq. (7.22) may be trusted. To develop a relation between T1
vs. temperature T , we need to know the value and temperature dependence of T2e . In this case, by
T2e we mean the phase memory time of the electrons independent of the nuclei. There is evidence,
to be discussed below, that the measured T2e is usually affected by the nuclei. The relaxation
formula of Sec. 6.1, however, implicitly assumes that the local field is fluctuating independently of
the nuclear spins; in other words, that any nuclear spin information is lost in the infinite electron
reservoir. This is rarely actually the case. In the limit of isolated nuclei coupled to trapped, isolated
impurity electrons and interacting only via the contact hyperfine interaction, there are no I z S

7.2 Trapped Carriers

185

or I S z terms, so only flip-flop processes need be considered. Energy conservation requires that
such processes can happen only when the electron phase is advanced or retarded due to external
influences such as phonons or other impurities. Therefore, the most suitable value to take for T2e is
that due to electrons isolated from nuclei.

An upper estimate of T2e is T1e ; dilute nuclei do not play much role in the spin-lattice relaxation
of electrons since they cannot accommodate the energy of an electron spin-flip. Therefore T1e is
a measure of the timescale of interactions between the electron and external degrees of freedom.
However, those external degrees of freedom, like the nuclei, may not have a large density of states
at the electron Zeeman energy, and for this reason T2e is expected to be much lower. A lower
estimate is the local field due to nearby electrons, be they trapped or free. The strongest coupling
to nearby trapped carriers will be the dipolar interaction; however, a fast exchange interaction is
known to couple trapped impurities with free carriers. Values for some of these timescales may
be found in the papers of Feher et al. [173, 224, 225]. The esr linewidths of the trapped carriers
studied in those papers are dominated by inhomogeneous broadening from the nuclear hyperfine
interaction, but some exchange narrowing is seen indicative of exchange coupling between multiple
electron impurities. Until recently, the best measure available for the phase memory time of electrons
relatively isolated from each other and nuclei was given by Gordon and Bowers [226], where they
used spin-echos to measure T2e = 0.52 ms in isotopically enriched

28

Si at 1.4 K.

Suppose that for a quantitative estimate of T1 (r), we use this measured value of T2e , while for
high temperatures we estimate T2e T1e and use the T1e data of Feher [224]. Do such estimates
give reasonable numbers? For comparison, we know from the work of Feher and Gere [173] that T1
for phosphorous nuclei is on the order of 10 hours, and that ee (0)F (0)/2 = 60 MHz at the

31

site. At low temperature, then


1
1

T1 (31 P)
4

60 MHz
28 GHz

2

1T
1
1
=
,
B0 0.52 ms
452 sec

a result which is too fast by a factor of about 80. Either the formalism is incorrect, or the effective correlation time used is too short. For T1 to be 10 hours, one requires T2e = 40 ms,
a number still substantially smaller than T1e (3000 seconds at 1.25 K [173].)

How does this

number compare to the electron dipolar field? Consider the density at which Feher reports a
lack of concentration dependence of electronic relaxation ne 1016 cm3 . The atomic density
of silicon is 1 = 2/(2.2
A)3 = 1.9 1023 cm3 , so the average electron-electron distance is

186

Chapter 7. Nuclear Relaxation in Silicon

5.5
A (ne )1/3 = 1500
A, for which the dipolar coupling, in SI units, is
fee-dip

0
1
(28 109 )2
1
(gB )2 3 107 6.626 1034

,
4h
r
(640 1010 )3
80 ms

a number of the correct order of magnitude to explain such a correlation time. It is more likely
that the data of Gordon and Bowers [226] was not able to isolate this longer T2 . In fact, these
numbers for T2e , between 40 and 80 ms, correspond almost exactly to the more recent T2e spin-echo
measurements of Tyryshkin et al. [87] for phosphorous donors in higher quality isotopically depleted
samples of silicon, where a value of 60 ms is estimated at low temperature.
In summary, a very reasonable estimate for T1 due to trapped carriers may be found from the
transverse fluctuations of the electron spins, which are themselves modified due to electron-electron
dipolar couplings.
The situation is different for T2 . One must be a little careful with the formula. If one naively
estimates T21 h(bz )2 iT1e and uses the typically long T1e times, one can obtain outrageously short
estimates for T2 . In particular, suppose the motion of the local electron is completely frozen; then
T1e and T2 0. This is because of how we have defined T2 . If the electron is frozen, its
effective field at the nucleus is constantly present, and therefore it constantly advances the phase of
the nucleus; thus the nucleus very quickly goes out of phase. An ensemble of nuclei would undergo
rapid dephasing, but this dephasing is inhomogeneous and can be therefore refocussed with a simple
pulse.
As an example of a careless calculation, consider the data of Feher and Gere [173] taken at
0.38 T and 1.25 K. Under these conditions, 1 p2e = 0.97, meaning that the electrons are still mostly
random and this polarization factor can be neglected. However, T1e is a very long 3000 sec. Using
again a hyperfine coupling constant for the
2

as ( 60 10

12

31

P nuclei of 60 MHz, one might therefore estimate T2

3000 sec) = 8 ns, a number which is ridiculously low. This may instead be

considered a fast inhomogeneous dephasing time T2 which could be refocussed with a pulse.

The conclusion that we draw is that, in defining T2 , we must seek those processes whose correlation times are smaller than the experiment we are describing, a critical point to which we will
return when we discuss 1/f noise.

7.2.2

Optically Excited Bound Excitons

In Sec. 3.4.3, I discussed a method for optical detection of

31

P nuclei in silicon. An important issue

for this proposal is the probability that the optical excitations used in the scheme will themselves

7.2 Trapped Carriers

187

scatter the nucleus before its state may be deduced.


I first argue that nuclear spin relaxation due to optically excited, delocalized conduction electrons
is not expected to be a significant effect. The number of conduction electrons required for neutralization of the Auger-ionized donors in 1 ns is approximately 1013 cm3 , assuming a 4 K electron
capture cross section of 4 1011 cm2 [227]. At this temperature, electrons are expected to thermalize quickly, in which case the T1 should resemble the calculation for intrinsic donor electrons that
we considered at the beginning of this chapter. For the low densities created by optical excitation,
this theory predicts a T1 greater than 108 seconds. Even 29 Si in heavily doped (> 1017 donors/cm3)
silicon has a measured T1 in excess of 200 minutes at 4 K [220, 221]; the T1 corresponding to
2

31

nuclei would only be shorter by a factor of 29 Si /31 P 4, still leaving this timescale unimportant.
An argument that T1 due to free excitons should also be negligibly long follows a similar reasoning,
since only the spin of the s-like electron interacts appreciably with the nucleus, and Boltzmann
statistics may still be assumed.
Of greater concern is the probability of a nuclear spin flip during the capture of a free electron
following the Auger process. Such a nuclear flip arises in second order perturbation theory, in
which a virtual electron capture and a virtual electron/nuclear spin flip-flop in the neutral-donor
state occur concurrently with the energy compensated by the real emitted phonon. To estimate the
probability of such an event, we note that the energy cost of a hyperfine-induced electron/nuclear
spin flip-flop (approximately the 1.2 meV Zeeman energy of the neutral-donor spin) is substantially
smaller than the donor binding energy (45 meV). We may therefore assume that the density-of-states
factors in Fermis golden rule are unchanged between the first and second order processes, and that
they are independent of the initial spin state. We also assume that the optical excitation of free
carriers is not spin selective, and since T1 for these carriers exceeds the capture time, this implies
that the initial spin polarizations are approximately equal. It follows from these assumptions that
the probability ratio between the first and second order processes is well approximated by the ratio
between their matrix elements. The second order matrix element for an electron/nuclear flip-flop
process, assuming without loss of generality that the nucleus begins in the |i state, may be written
as




2
P+ H P0
P0 H P0 2
C
1
(2)



VP+ P0 =
,


EP0 EP0

(7.28)





in which P+ describes the ionized donor and P0 describes the neutral donor with electron


spin up and nucleus spin down. This initial state and the intermediate state P0 are the only


important ones for nuclear destabilization, since the P0 state is unperturbed by H1 . The un-

188

Chapter 7. Nuclear Relaxation in Silicon

perturbed eigenenergy difference between the intermediate and final state, EP0 EP0 , is a sum
of the electron Zeeman term, the nuclear Zeeman term, and HF,P0 ; the electron Zeeman term
dominates this sum. The Hamiltonian term HC refers to the interaction leading to the capture of
the free electron; thus, a first-order matrix element for a neutralization process without a flip-flop
is written as

2



2
(1)

VP+ P0 = P+ HC P0 .

(7.29)

We assume that this Coulombic process has no spin selectivity. It follows that the probability per
transition of a flip-flop may be written as

2

(2)



2

2
VP+ P0
1 HF,P0
1 P0 H1 P0
.
=

2
2
2
2 EP0 EP0
2 ge B B0
(1)

(1)

(2)

VP+ P0 + VP+ P0 + VP+ P0
(7.30)

This probability can be seen from an alternative viewpoint: if we use second-order time-independent




perturbation theory to calculate the mixing of the P0 and P0 states due to the flip-flop terms

of H1 , and presume that capture rates to these perturbed states are the same as to the unperturbed

states, the same probability is obtained.


The free exciton capture processes leads to a similar second-order probability for nuclear randomization. In this case, however, the binding energy of the free exciton to the neutral donor is only
somewhat larger than the electron Zeeman energy, so the density-of-states factors in the transition
rates can become important. However, we do not expect this correction to alter the per-transition
probability by more than a factor of order unity. We thus estimate that at 10 T, the probability of
a nuclear flip for each free-exciton-capture and free-electron-capture process is approximately twice
the result of Eq. (7.30): (60 MHz/280 GHz)2 = 5 108. The consequences of this probability were
discussed in Sec. 3.4.3.
Nuclear spin flip due to the P0 hyperfine coupling during radiative decay has a similar order
to that during free electron capture, but since radiative decay is 7000 times less frequent, this
probability may also be neglected. In a comparable detection scheme in a more optically active
material such as a neutral semiconductor quantum dot, a similar spin-flip probability is possible for
each radiative transition. A crucial difference is that in this case, the electron spin exists in the
exciton state, and if it was excited with polarized light, the probabilities of the two initial electron
spin states are quite different. Rather than simply randomizing the nucleus, these second-order
spin-flip processes may polarize the nucleus. It is exactly this type of calculation that has been used
to explain the emergence of nuclear polarization in GaAs quantum dots [189].

7.3 1/f Noise

7.3

189

1/f Noise

So far, we have only discussed relaxation processes in which the correlation time c is very short,
corresponding to white magnetic noise. However, low frequency noise sources exist as well, and
experimental evidence for such low frequency magnetic noise was discussed in Sec. 5.2.
In the case of electric noise, the low frequency spectrum is dominated by charging processes
with uniformly distributed activated energies. These charge lead to so-called 1/f noise, a rich field
of study in its own right. A good summary of the field, with good discussion of silicon, is given
by Weissman [228]. Presumedly, these charges have magnetic properties as well, and therefore 1/f
charging processes in bulk silicon and silicon surfaces can lead to nuclear relaxation.

7.3.1

General Considerations

Suppose there are charge traps in the bulk of the sample, perhaps due to deep impurities, dangling
bonds at crystal interfaces, or likewise. Suppose that in their ground state (which may be charged
or uncharged) they are diamagnetic, but in their excited state they exhibit a paramagnetic dipole
moment.
Presumedly, these impurities would also affect T1 , and so we attribute them to the T1 in undoped silicon observed in previous work [216, 219], which is limited to about 300 minutes at room
temperature. It is notable that this timescale seems to be independent of the growth conditions of
the silicon sample (in particular, it seems to be common to both single-crystal and polycrystalline
samples) and to magnetic fields. Sapoval and Lepine [219] measured a temperature dependence for
this timescale as
1
1

eW/kB T ,
T1
10 min

(7.31)

where W 70 meV. One could imagine explaining this equation with a model whereby charge traps,
with charging energy Ec 280 meV, are scattered throughout the lattice with density N , and when
excited the spin (assume S = 1/2) of their trapped charge fluctuates with c = T1e 1/0 . Nuclei

interact with these local cites via the dipole interaction, in particular the I S z term which is usually
responsible for T1 in the presence of dipolar fields from dipolar impurities [121]. Correspondingly
the T1 for the ith nucleus is

2  sin2 cos2 


1
0
i
i
=9
gB
T1e hS 2 i hSi2
6
T1i
4
ri
 


1
gB B0
2
Ec /kB T
(gG)
T1e e
cosh
,
ri6
2kB T

(7.32)

190

Chapter 7. Nuclear Relaxation in Silicon

where in the approximation G =

p
3/50 B /4 and the angular dependence has already been

averaged. The factor eEc /kB T comes about as the probability of finding S = 1/2; when discharged
the impurity has no spin to interact with the nucleus. Of course the polarization local to this impurity
will be carried to the rest of the nuclei in the bulk by spin-diffusion, which is a very slow process in
silicon. In Abragam [121], it is shown that under a few assumptions, (including an assumption that
the dynamics at the impurity are much faster than spin-diffusion), spin-diffusion will lead to a bulk
T1 of

1/4


p
1
gB B0
3/4
= 4N D
eEc /4kB T ,
gG T1e cosh
T1
2kB T

(7.33)

where D is a spin diffusion constant related to the dipolar T2 . This model is a tempting one to
explain the empirically observed Eq. (7.31). If, however, these are the same charges that contribute
to 1/f noise, their dynamics are very slow and must be incorporated into the model for spin-diffusion.
The likelihood of a slow change in the impurity populations was alluded to by Sapoval and Lepine
[219] to explain their observed T1 times. More detailed modelling of this system is required to fully
support the idea that 1/f charging noise can explain the T1 observed in undoped silicon.
Note that we chose c to be the very fast T1e despite the presence of a slow fluctuation, because
1
1
1
1
the effective correlation time if a fast and slow process are superposed is eff
= fast
+ slow
fast
.

7.3.2

Decoherence Due to 1/f Noise During Decoupling

In the case of T2 due to 1/f noise, we must be somewhat careful. Any magnetic noise process with
a correlation time c 1/0 cannot explain why T2 < T1 , unless the noisy magnetic field is highly
anisotropic. In particular, the electron spin fluctuating on its room temperature T1e timescale is
this fast and its dipolar field is insufficiently anisotropic to lead to a short T2 and not a comparably
short T1 . However, the average longitudinal field of the electron spin is nonzero. Therefore, if the
spin is undergoing a slow charging fluctuation, then a slow longitudinal noise process is created by
its average longitudinal dipolar field.
Mathematically stated, we presume that the longitudinal noise field before substracting the
average field is
B z (t) n(t)S z (t),

(7.34)

where n(t) is the occupation of the impurity state, which we assume is uncorrelated to the spin.
(Such an assumption would be invalidated if the charging process were spin-dependent). To get the

7.3 1/f Noise

191

zero-mean noise field, we subtract off the mean, leaving


bz (t) n(t)S z (t) hnihS z i,

(7.35)

and therefore the autocorrelation function is


hbz (t)bz (0)i hn(t)n(0)ihS z (t)S z (0) hni2 hS z i2




z
z
z 2
z 2
2
= hn(t)n(0)i hS (t)S (0)i hS i + hS i hn(t)n(0)i hni


p2e
1/T1e
2
=
hn(t)n(0)ie
+ hn(t)n(0)i hni .
4

(7.36)

The term decaying as e1/T1e is the broadband noise term that should lead to T1 , as discussed in
the last section, and therefore cannot explain a shorter T2 . The second term can, however.
We presume that for a single impurity, this term describes a Markov process with correlation
time j , so we define
2 hbz (rj , )bz (rj , 0)i = 2j ej ,

(7.37)

where 2j is the variance of the frequency shift due to this fluctuating charge. A formula for j
would include the dipolar coupling to the impurity and the thermal electron polarization.
This is not the only possible model for explaining low frequency magnetic noise, however. Another
possibility is that when charges fluctuate, they change the local diamagnetic shielding around nearby
nuclei, effectively changing the chemical shift. This will also lead to a small frequency shift j
correlated to the charging noise of the impurity. For the remainder of this discussion, we simply
treat j as an unknown random variable.
To see what T2 we expect during our decoupling pulse sequence, we combine this autocorrelation
function with the cumulant expansion in the toggling frame developed in Sec. 6.3.2. The result is
2j X zz
j
1
=
A
,
T2j
2 n n 2j + (2ni/tc)2

(7.38)

where the Fourier coefficient Azz


n is found by

Azz
n

1
=
tc

tc
0

n
o

Tr URF
(t)Ijz URF (t)Ijz
I(I + 1)/3

e2int/tc dt.

(7.39)

Equation (7.38) indicates to us that if most spins see a correlation time j that is smaller or
of the same order as t1
c , then the observed T2 should depend on tc , in contrast to our data for

192

Chapter 7. Nuclear Relaxation in Silicon

2J()

10 Hz

1/f
100 mHz

2j/j

1 mHz

n/tc

2k /k

H
z

z
1

M
H

z
10
0

M
10

M
H
z
1

kH

10
0

kH
z
10

kH
z
1

10

100 Hz

Figure 7.1: A schematic of the magnetic noise spectral density 2 J() in silicon. (The
direction of the magnetic noise vector is neglected in this plot). The Lorentzian magnetic
noise spectral density due to a very fast fluctuator (white noise) is represented by the
dashed line.

polycrystalline silicon shown in Fig. 5.10. Our data would seem to be explained by processes with
j much larger than t1
c , in which case T2j j . This point is illustrated in Fig. 7.1.
The assumption j t1
is the common motional narrowing or white-noise limit for this
c
T2 noise process. However, as shown in Fig. 7.1 the noise cannot strictly be white at frequencies
higher than t1
c , as in the case of the dashed line in Fig. 7.1. The magnitude of this noise at 0 is
the same as at each frequency n/tc for all integers n, so if this component alone existed then we
should have T2 T1 . To explain T2 T1 , low frequency noise which cuts off at frequencies lower
than 0 must be present.
In summary, to explain the independence of T2 on the decoupling cycle time, the noise is assumed
to be caused by Lorentzian noise distributions that are flat (at noise power 2j /j ) over the range
of n/tc where the Fourier operators of the pulse sequence are significant. To explain the longer T1 ,
this same noise must cut off at a frequency lower than 0 . Several such Lorentzian curves are shown
in Fig. 7.1; the specific model we suppose is a collection of such Lorentzian curves that sum to a
total spectral noise density scaling as 1/f , a point to which we return shortly.
The cumulant expansion approach leading to Eq. (7.38) is not appropriate for fluctuations which
are very slow in comparison to the measurement time; however, the tc -independence of the data
suggest that the dominant source of this decoherence is processes much faster than tc . These faster
P
processes are well described by the cumulant expansion approach, and since n Azz
n = 1, we presume

7.3 1/f Noise

193

they lead to a local decoherence rate given by


2j
1

.
T2j
2j

(7.40)

Although Eq. (7.40) is sufficient for our purposes, we also clarify the issue of timescales with another, non-perturbative approach [229] which neglects the fast mrev-16 pulse sequence but accounts
for the slower refocusing effects of the -pulses. In this approach, the decay function with a single
refocusing pulse (Hahn echo) at the 2e echo peak and with frequency shifts fluctuating according
to a Poisson random pulse train, is given by
D

ei(2e ) = e2e j



q
cosh 2e j 1 gj2 gj2
1 gj2



q
sinh 2e j 1 gj2
,
q
+
1 gj2

(7.41)

where gj = j /j . This result assumes a single -pulse; if we examine the more complicated
equations for a train of -pulses, we find results comparable to assuming that unrefocussed coherence
is forgotten every echo cycle, so that
D

E D
En
ei(2ne ) ei(2e ) .

(7.42)

Although time is measured discretely at the peak of each echo, we will set t = 2ne and treat it
as continuous. However, the value of e = 120tc in comparison to other time scales is critical, as
evidenced by the following argument. Suppose nucleus k is coupled to a fluctuator that is slow in
comparison to the chemical shift, so that k k . In this gk limit, Eqs. (7.41) and (7.42)
give
he

ik (t)



42k e2 k
i exp
t .
3

(7.43)

Now suppose nucleus j is coupled to a very fast fluctuator, with j j . In this gj 0 limit we
find
he

ij (t)

!
2j
i exp
t ,
2j

(7.44)

the same result as Eq. (7.40) derived using the toggling frame cumulant expansion in the j t1
c
limit. We now ask whether nucleus j or k contributes more heavily to the observed ensemble signal.
In our model, nuclei dephase independently, and the sum of their decay functions provides the
signal. Consequently, those spins which decay the slowest contribute to the signal the most2 . Thus
2 This is in contrast to models where a single electron spin or Josephson Junction qubit couples to many bistable
fluctuators, and the fastest decaying components contribute most to the final signal, which is calculated as a product

194

Chapter 7. Nuclear Relaxation in Silicon

we compare decay rates for our two spins:


T2j
8 2k
=
j k e2 .
T2k
3 2j

(7.45)

If e is larger than the geometric average of the two fluctuator rates, we find that the nuclei coupled
to fast oscillators will dominate the sum, as these decay the slowest. The datas independence of
e = 120tc indicates that we are working in this regime, i.e. that our signal is dominated by spins
close to fast oscillators causing rapid spectral diffusion. This is the limit where the perturbative
cumulant expansion approach agrees with the non-perturbative approach of Eq. (7.41). In the
following, then, we assume that our signal is dominated by nuclei decaying according to Eq. (7.40).
We now introduce a distribution of j across the sample by assuming that the charging/discharging
processes leading to this nuclear decoherence are the same as those which are well known to lead to
1/f noise near silicon surfaces. The standard model for 1/f noise supposes that across the sample,
j is randomly distributed according to the probability density function

D () =




ln(high /low ) 1 , low < < high ,

0,

(7.46)

otherwise.

It may be easily seen that this distribution leads to 1/f charge noise for low f high . The
details of the physical processes leading to this distribution in silicon are discussed in Ref. 228.
We presume our nuclei are dephased by a random selection of bistable oscillators with lifetime
j . We also presume the nuclei undergo random shifts j , and since isotope placement is diffuse and
random, we assume j will be mostly uncorrelated with j (corresponding to roughly one impurity
per nucleus). We therefore arrive at the ensemble decay function
X
j

heij (t) i

dD ()

 2  Z

E1 ( 2 t/2high) E1 ( 2 t/2low )
dD () exp t = dD ()
,
2
ln(high /low )

where E1 (x) is the exponential integral

(7.47)

R
x

/y dy. The E1 ( t/2low ) term is much smaller than

the E1 ( 2 t/2high) if high low , as appears to be the case for 1/f noise observed in silicon, so
we neglect this term. For the distribution of frequency shifts D (), we assume that is peaked
around some average . We thus expand the integrand about this average to lowest order, allowing

over fluctuators rather than a sum.

7.3 1/f Noise

195

0.10

Residual

0.05
0.00
-0.05
Cycle time (tc)
1.338 ms
1.914 ms
2.490 ms
3.306 ms

-0.10

Time (sec) 0 5 10 15 20 25 30 0

Counts
50

Figure 7.2: Residuals versus time. The deviation of the data of Fig. 5.10 from the fitting
function of Eq. (7.48), with a histogram of those residuals on the right, consistent with
Gaussian noise.

us to complete the integral without detailed knowledge of the distribution:


X
j



heij (t) i N 1 E1 (t) + (1 + 2t) exp(t) ,

(7.48)

where = /2high and = 2 / 1. (The normalization constant N would be a free parameter


for any model, since the magnitude of our data shifts from experiment to experiment due to differing
initial magnetizations and probe temperatures). This function has the correct shape for our data;
we also reproduce its shape with computer simulations of the decoherence model described.
Figure 5.10 shows this theoretical curve, fit to the data by the Levenberg-Marquadt method for
least-squares fitting. We find insignificant difference between fitting all four experimental curves
separately or fitting all the data simultaneously. The curve shown fits all the data assuming shared
constants , but independent normalization constants N for a total of 6 fitting parameters over
493 data points. The residuals are shown in Fig. 7.2, where they are checked against Gaussian noise
with a 2 test. The first two echoes are slightly weaker than the theoretical curve, and systematically
deviate from the model. This is not surprising, since we omitted the effects of those spins which
decay more rapidly due to slower spectral diffusion in the derivation of Eq. (7.48). Otherwise, we
find that the residuals are consistent with = 0.02 Gaussian noise with a 2 of 1.09 over 16 bins,
leaving no indication of systematic disagreement between the data and the model.

196

Chapter 7. Nuclear Relaxation in Silicon

The parameter fit is optimized at = 22 2 mHz, = 0.20 0.05. This value of would be
consistent, for example, with an average chemical shift of 0.5 ppm and a cutoff rate constant high
300 kHz. In the Dutta-Horn model for 1/f noise [228], we would expect high exp(E/kT ),
where E is an energy barrier for the fastest charge traps in the sample. This model thus predicts an
exponential temperature dependence for this decoherence rate. Some extra temperature dependence
could also result from a changing average electron spin polarization.
In summary, we find that our data in polycrystalline silicon is consistent with our model for
decoherence induced by 1/f charging processes superimposed over unbiased Gaussian noise. This
decoherence source should diminish in single crystals, as shown in our single-crystal data, and at
low temperatures.

Chapter 8

Prospectus
It would appear that we have reached the limits of what it is possible to achieve with
computer technology, although one should be careful with such statements, as they tend
to sound pretty silly in 5 years.
John von Neumann, said in 1949
I offer now a prospectus rather than a conclusion, as the ideas, theories, and experiments presented in this dissertation represent only the very beginning of a long course of research.
I do not know whether, 50 years from now, the quantum computer will be a research dream, an
expensive laboratory apparatus owned only by a wealthy government agency, or a standard feature
of commercial cell phones. However, experimental quantum information processing in the form of
quantum cryptography and simple, few-qubit devices have already come far enough in the last 10
years to ensure the place of quantum information in mankinds toolbox of technologies. For this
technology to grow, quantum memory and the ability to manipulate it will be crucial, and I have
argued in this dissertation that for these tasks semiconductor nuclei may be the most promising
physical qubit system.
I have focused on silicon, whose convenient isotopes and advanced fabrication technology make
it a likely material for these advanced quantum technologies. I have pushed the limits of nmr
decoupling techniques to experimentally seek a limit to the quantum behavior of nuclei in silicon in
the form of decoherence due to sources outside of the nuclear system. A theoretical treatment of
nuclear relaxation in silicon suggests that I have found such a limit only in polycrystalline silicon
samples, where well-known sources of electronic charging noise may be the cause of decoherence.
The ability to grow high quality single crystals of very pure silicon may prevent this noise source

198

Chapter 8. Prospectus

from being a problem in future silicon-based devices; I have only been able to place a modest lower
bound of 25 seconds on the coherence times available to isolated nuclei in this material. I have
developed ideas for designing quantum computer architectures based on bulk single-crystal silicon
in order to take advantage of this long coherence time.
However, the limitations of silicon must also be acknowledged. Although silicons optical capabilities allow some degree of nuclear spin polarization and sensitive nuclear spin measurement, its
poor optical efficiency may ultimately limit its applicability in large scale technologies. I cannot say
whether future progress will teach us ways to circumvent these problems, or whether other materials
will be developed to avoid them. This observation echoes the larger field of optoelectronics, where
the two diverging directions being pursued in the effort to combine computers and fiber optics are
to improve or to abandon silicon-based devices.
What can replace silicon? I have discussed the possibility of quantum dots based on group-iii-v
semiconductors such as GaAs; here the optical capabilities are strong, but coherently controlling the
noisy nuclear environment is a challenge. Other group-iv materials such as diamond with its NV
centers or isotopically engineered group-ii-vi materials may provide more promising routes, but here
the technology for fabrication is lagging. Experiments in my laboratory and others all around the
globe will continue to work on these questions for some time to come.
Perhaps the future will bring the observation of some new fundamental physics that will prevent
making large quantum coherent devices, and as a result our understanding of our universe will
deepen. Or perhaps we will have completely mastered the strangeness of quantum theory, silencing
the concerns expressed years ago by Einstein and Schr
odinger. The von Neumann quote that begins
this chapter reminds us how dim our vision of the future of technology must be, and how many
possibilities must exist to surprise us.

Appendix A

Notation
Maths should be about notions, not notations.
Carl Friedrich Gauss

A.1

Vectors, Matrices, and Tensors

Vectors, matrices, and tensors are notated by boldface roman characters, such as v. If they have
. Vector components are generally notated by
unit length, they may be written with a hat as v
v. Subscripts are usually reserved for vector labels. The vector name in italics
superscript: v x = x

without superscripts refers to its magnitude: v = v v = |v|.


P
When higher rank tensors are useful, I switch to a component notation. For example, jk v j Tjk v k
is the same as v T v.

One matrix is special, and that is the Kronecker-delta, ij . This is the identity matrix; it is equal

to 1 if i = j and 0 otherwise. The Kronecker-delta is also used with subscripts; the rules are the
same. If a single subscript is used with two superscripts, kij , then this symbol is a projection matrix
in its superscripts; it is zero unless i = j = k, in which case it is 1.
I also make use of the Levi-Cevita antisymmetric tensor density jkl . 123 = 231 = 312 = 1,
321 = 213 = 132 = 1, and all other components are zero.

A.2

Quantum Mechanics and NMR

Dirac kets are always written in this work with an angled bracket, i. Usually in this dissertation
I use normalized states satisfying h|i = 1 for any state description , although occasionally I

200

Appendix A. Notation

suppress the normalization factor for a simple superposition of states when it is obvious, such as the

omission of 1/ 2 from |i = |i + |i.


Operators are not given special notation in this work; any symbol may be considered an operator.
Those symbols which represent classical numbers (c-numbers) may be considered operators which
commute with every operator. An operator in a particular basis is referred to as a matrix.
Several operators occur repeatedly and are worth special mention.
The notation H is used for the Hamiltonian except in Chapter 6, Chapter 7, and Appendix A.5
where many-body effects are at play, in which case it represents the Hamiltonian density while the
Hamiltonian is notated H (not to be confused with Shannon entropy or Hadamard matrices). When
discussing many-body physics, H is the Hamiltonian seen by a particular particle, and H is the total
many-body Hamiltonian. Both by convention have units of angular frequency; in other words each
Hamiltonian in this dissertation has one factor of ~ missing from the usual Hamiltonian with energy
units.
It was John von Neumann who introduced the crucially important density operator . For
quantum mechanical states without classical randomness (pure states), the density operator is
simply = |ih|. Even in this simple case, the density operator offers some notational advantage
over the usual ket notation. The general pure state of a spin-1/2 with spin operator I may be written
=

1
I,
+n
2

(A.1)

is the unit vector pointing along the direction with maximal probability. The sphere traced
where n
is the surface of the Bloch sphere.
out by all possible n
The density operator is most important when classical statistical mechanics is involved, in which
case it may always be written
=

X
j

Pj |j ihj |

in some basis of states j . Such states are mixed. The most general state of a spin-1/2 system
may now be written
=

1
+ p I,
2

(A.2)

where p covers beth the surface and the interior of the Bloch sphere. If p < 1, the state is mixed.
One basis for is special, and that is the basis of eigenenstates of the Hamiltonian. In thermal
equilibrium, the density operator is diagonal in this basis, with hj| |ji = Pj = Z 1 exp Ej /kB T ,
where Z is the partition function. I frequently use the notation = ~/kB T . In the eigenenergy

A.2 Quantum Mechanics and NMR

201

basis, diagonal elements of quantify populations and off-diagonal elements quantify coherences.
Decoherence is the gradual disappearance of off-diagonal elements in this basis. When the density
operator is expanded as a product-sum of spin operators, the coherence number of a particular term
is the number of raising operators in a term minus the number of lowering operators.
The expectation value of any operator A is
hAi = Tr {A} ,

(A.3)

where Tr {} is the trace. Recall that the trace is independent of basis. If the state is pure, = |ih|
so this reduces to h| A |i.
Spin operators also occur repeatedly. I use I for nuclei and S for electrons by convention. For
nuclear spin-1/2, I = 1/2, and I a = I a , where x,y,z are the usual Pauli matrices in the z-basis.
direction is ~h
The expectation value for angular momentum in the n
n Ii. Spin operators usually
occur with a subscript indicating which nucleus in the lattice or molecule carries that spin (or, more
precisely, which nuclear wave function carries that spin). When a product of two such operators
appears, such as Iiz Ijx , the implied operation is the tensor product, Iiz Ijx . In this regard, an
expression such as 1/4 + Iiz + Iix Ijy is shorthand for

1
1i 1j + Iiz 1j + Iix Ijy ,
4
where 1 is the identity operator for SU(2) in the representation corresponding to I.
For any vector, whether it be spin or position, we define z-basis raising and lowering elements
as v = v x iv y .
I use si units throughout. The magnetic moment of a nucleus is n = ~hIi. I only refer to
g-factors in reference to electrons, which have magnetic moment e = gB hSi.
When discussing an ensemble of nuclei, I employ the notation
M=

Ij

(A.4)

where the summation is over all ensemble members. The eigenstates of the M z operator are mz and
they correspond to the Clebsch-Gordon decomposition for any total angular momentum manifold
J. The letter M is chosen to be reminiscent of magnetization, but it does not include the factor ~.

202

A.3

Appendix A. Notation

Liouville Notation

I occasionally slip into Liouville notation which seems to be popular in modern analyses of solid-state
nmr [125] but is also useful for abbreviating notation in certain quantum information calculations.

For each operator A, we associate a Liouville vector A , which is always notated with a parenthesis
). An important shorthand is to put a state label in a Liouville ket to mean the projection

operator for that state:




= |ih| .

(A.5)


An operator B generally acts on a Liouville ket A as



B A = [B, A] .

Crucial here is the easy-to-show corollary that



eiB A = eiB AeiB .

Hence, in this notation, what looks like a unitary operator acting on a vector is really an operator
acquiring time dependence by being sandwiched between a unitary operator and its inverse. The
equation of motion for the density operator in the Schr
odinger picture is then simply
i



= H ,
t

(A.6)

and the equation of motion for an operator O in the Heisenberg picture is

We also have for the inner product



O = iH O .
t

(A.7)




A B = Tr A B .
which is an operator which acts on operators. For
We also introduce the superoperator, O,


A =
example, the unitary operator U = exp iHt acting on a Liouville ket A would be notated U


U AU . Superoperators need not be unitary.

A.4 Many-body notation

A.4

203

Many-body notation

Already I have used notation such as |i + |i for the state of two spin-1/2 particles. This notation
is very much a shorthand. Suppose we are discussing two nuclei at two different locations in a crystal
or molecule. The positions of these nuclei are therefore described by two separate wavefunctions,
1 (r) and 2 (r). The full state must be antisymmetric because these nuclei are fermions. Therefore
the wavefunction would be written
|i =


1
|1 , i |2 , i + |1 , i |2 , i |2 , i |1 , i |2 , i |1 , i .
2

(A.8)

This may be abbreviated using standard anti-commuting fermionic creation operators as


|i = (c1, c2, + c1, c2, ) |0i
Z
Z


3
= d r d3 r 1 (r)2 (r ) (r) (r ) + (r) (r ) |0i .

(A.9)
(A.10)

The advantage of using such operators is that wavefunction symmetry is built-in to the algebra.
For fixed nuclei, this notation is unnecessarily cumbersome, especially when discussing mixed states,
and it reveals no new information about measurable observables since the separate nuclei are always spatially distinguishable. However, the electrons which might interact with those nuclei in a
semiconductor could be subject to Pauli-exclusion effects, and therefore this many-body formalism
is appropriate.
In these cases I use a fairly standard notation from solid-state physics, in which the sth component
of the one-electron operator-spinor is
s (r) =

X
n,j

ns (r Rj )cnRj s =

uns (k, r)eikr cnks ,

(A.11)

n,k

where ns (r Rj ) is the Wannier function corresponding to the j th lattice site, band n and spin
(or, sometimes, total angular momentum projection) s, and ns (k, r) = uns (k, r) exp(ik r) is the
corresponding Bloch function; the two are related by
1 X
ns (r Rj )eikRj .
ns (k, r) =
N j

(A.12)

204

A.5

Appendix A. Notation

The Hyperfine Interaction

The general interaction between an electron and a nucleus has been beautifully summarized by
Abragam [121] as
H1 =



L S 3r(S r) 8
0
g0 B I
+
+
S
(r)
,
4
r3
r5
3

(A.13)

where L and S are the orbital and spin angular momenta of the electron, respectively. The vector r
connects the nucleus and the electron. More precisely, L = r pe ; one cannot simply use the atomic
L if the nucleus in question is not at the center of the atom! If the orbital angular momentum is
not quenched by a crystal field, in the case of an atom or, for example, a 4f electron in a rare-earth
crystal, this may be rewritten in a manifold of the electrons total angular momentum J = L + S as
H1 = g0 B n ALJ I J,

(A.14)

where
ALJ =

20
(r) for L = 0,
3

ALJ =

0 1 L(L + 1)
for L 6= 0.
4 r3 J(J + 1)

(A.15)

I will begin by considering only s-electrons in semiconductors, whose hyperfine interaction with the
nucleus is typically much larger than the corresponding dipolar coupling of an electron with orbital
angular momentum.

Let us therefore assume that we can ignore the orbital angular momentum of the electrons, and
correspondingly we can ignore spin-orbit coupling. It follows that the Bloch and Wannier functions
of Eq. (A.11) are independent of spin, and hence we may drop the s subscripts. Summing over
lattice nuclei and electrons, then, we have in the Wannier representation,

H1 =

dr (r)H1 (r) =

XX
1 0
g0 B
n (Rl Rj )n (Rl Rj )
2 6

n,n j,j ,l

h
i
Ilz (cnRj cn Rj cnRj cn Rj ) + Il (cnRj cn Rj ) + Il+ (cnRj cn Rj ) . (A.16)
For a periodic crystal, the product n (Rl Rj )n (Rl Rj ) should depend on neither j nor l, but

rather on the two differences j l and j j . This symmetry is made more apparent by using the

A.5 The Hyperfine Interaction

205

Bloch representation:

H1 =

XXX

1 0
g0 B N
un (k, Rl )un (k , Rl )ei(k k)Rl
2N 6
nn kk l
h
i
Ilz (cnk cn k cnk cn k ) + Il (cnk cn k ) + Il+ (cnk cn k ) , (A.17)

where the function


N

0
g0 B un (k, Rl )un (k , Rl ) nn (k k)
6

(A.18)

depends only on the difference of wave-vectors by virtue of the symmetry of the lattice.
A few comments on nn (k k) are in order. If we look at the magnitude of nn (0), which
will primarily concern us throughout this work, we see the factor
un (0, Rl )un (0, Rl ) =

1 X
n (Rl Rj )n (Rl Rj ),
N

(A.19)

jk

which appears also in Eq. (A.16). The Wannier functions n (R) are normalized like atomic functions
at each site, whereas the Bloch functions un (0, R) are normalized across the entire crystal. The
former therefore give a better interpretation of the hyperfine coupling, and the factor N in Eq. (A.18)
is therefore intended to cancel the 1/N on the right-hand side of Eq. (A.19). This normalization
also corresponds to that used in Paget et al. [230].
Another note is worth pointing out. Using the so-called k p method, a perturbative expression
for un (k, r) is [231]
Z
~2 X
un 0
un (k, r) = un (0, r) + i
dr un (0, r )k un (0, r ).
m En0 En 0

(A.20)

n 6=n

For group-iv and group-iii-v semiconductors, the symmetry properties are such that, in making a
calculation for ue (k, r) (that is, for the lowest conduction band), only the contribution of uh (0, r),
the Bloch-function of the uppermost valence band, is important, and the energy difference between
these two bands is the bandgap E0 . We may therefore write, approximately,
ue (k, r) ue (0, r) + i

~k P
uh (0, r),
mE0

(A.21)

where the matrix element P is of order h/aB . Very roughly, then, for small k,
ue (k, Rl )ue (k , Rl ) ue (0, Rl )ue (0, Rl ) + i

~
P (k k)uh (0, Rl )ue (0, Rl ).
mE0

(A.22)

206

Appendix A. Notation

The second term kicks in when |k k| is on the order of a 40th of a lattice vector. In most of this
paper, however, the p-like holes will have uh (0, Rl ) so much smaller than the ue (0, Rl )s of the s-like
electrons that I will ignore the second term and with it all k-dependence.
For ease of reading, then, I repeat Eq. (A.17) as the final hyperfine coupling for spin-1/2 electrons:

H1 =

1 XXX
nn (k k)ei(k k)Rl
2N
nn kk
l
h
i
Ilz (cnk cn k cnk cn k ) + Il (cnk cn k ) + Il+ (cnk cn k ) . (A.23)

To make the algebra simpler, I will sometimes notate the terms in the square brackets as

Il cnk cn k

(A.24)

where is 1, 0 or 1 corresponding to lowering, z, or raising Pauli matrices, with components , ,


respectively.
When limited to a single electron in a single state described by k-space wave function Cn (k)
independent of electron spin, this coupling may be written
H1 =

X
l

Al S Il ,

(A.25)

where
Al =

XX
nn kk

Cn (k )Cn (k)ei(kk )Rl nn (k k )

(A.26)

and S is the spin-operator for the electron. If we neglect the k-dependence of the Bloch functions,
the prefactor is simplified to

2
X

0


Al = N g0 B
un (0)Fn (Rl ) ,


6
n

(A.27)

where Fn (Rl ) is the Fourier transform of Cn (k) at the location of nucleus l. When speaking of a
single localized electron, such as that of an exciton in a quantum dot or a donor bound to
silicon, this is the form of the hyperfine coupling used.

31

P in

Appendix B

Spin Algebra
The most important characteristic of spin operators is their commutation relation:
[I i , I j ] = iijk I k ,

(B.1)

independent of representation, which may be rewritten as


, I m]
= iI n
m.

[I n

(B.2)

Another convenient formula is for the raising and lowering operators, I = I x iI y . Let s = +1, 0,
or 1, where I 0 = I z . Then

B.1

[I s , I s ] = (1)s+s (s s )I s+s .

(B.3)

Single Spin Rotations

This commutation relation characterizes rotation as seen by the following argument. First, recall
that for complex and arbitrary operators A, B
eA BeA =

X
n
[A, B]n ,
n!
n=0

where [A, B]0 = B and [A, B]n+1 = A[A, B]n [A, B]n A. This is easily proven by taking derivatives
with respect to and comparing terms order by order in .

208

Appendix B. Spin Algebra

Hence,
iIn =
eiIn I me

X
(i)j
j=0

j!

, I m]
j.
[I n

(B.4)

We now simplify by powers of two. First we note that


, I m]
j =Iu
[I n

(B.5)

for some complex vector u. Then we have


, I m]
j+2 = [I n
, [I n
, I u]] = i[I n
, I n
u]
[I n

(B.6)

un
I
=n

(B.7)

.
= I u (
n u)I n

(B.8)

, I m]
j for even j:
Let us consider [I n
, I m]
0 =Im

[I n
, I m]
2 =Im
(
n

[I n
n m)I
n
m
n, so n
u = 0. It follows that all operators
To continue the list, we note that u for j = 2 is m
, I m]
j for even j > 0 are equal to [I n
, I m]
2 . Let us now consider [I n
, I m]
j for odd j.
[I n
, I m]
1 = iI n
m,
for which we write u = i
Once again, n
u = 0,
For j = 1 we have [I n
n m.
, I m]
j for odd j are equal to [I n
, I m]
1 . We may therefore rewrite our sum
so all operators [I n
in (B.4) as
iIn = (
n
+ [I m
(
n
]
eiIn I me
n m)I
n m)I

X (i)j
X (i)j

iI (
n m)
. (B.9)
j!
j!
j even
j odd

From this we find our final result


iIn = I m
cos + (m
n
)(I n
)(1 cos ) I m
n
sin .
eiIn I me

(B.10)

From this, an important consequence is


z

eiI I eiI = ei I .

(B.11)

B.2 Two Spin Commutators

B.2

209

Two Spin Commutators

In evaluating the theoretical dynamics of coupled spins, it is useful to know the commutator between
products of arbitrary spins. For this we will consider only spin-1/2, which satisfies
{Iia , Iib } =

1 ab

(B.12)

in addition to Eq. (B.1). In particular, this implies


Iia Iib =

 1
1
[Iia , Iib ] + {Iia , Iib } =
2
2


1
iabc Iic + ab .
2

(B.13)

We obtain
[Iia Ijb , Ikc Ild ] = Iia [Ijb , Ikc Ild ] + [Iia , Ikc Ild ]Ijb
= Iia Ikc [Ijb , Ild ] + Iia [Ijb , Ikc ]Ild + Ikc [Iia , Ild ]Ijb + [Iia , Ikc ]Ild Ijb

= i jl bde Iia Ikc Ije + jk bce Iia Ije Ild + il ade Ikc Iie Ijb + ik ace Iie Ild Ijb .

(B.14)

Here, we obviously consider i 6= j and k 6= l. Suppose that any two spins are equal; for example
i = k and j = l. Then we have
[Iia Ijb , Iic Ijd ] = ibde Iia Iic Ije + iace Iie Ijd Ijb




i bde
1 ac e i acf f
1 db
acf f
bde e
=
i Ii +
Ij + Ii i Ij +
2
2
2
2

i bde ac e
=
Ij + ace bd Iie .
4
This is nonzero if and only if one and only one of the two species (i or j) has the same operator on
both sides of the commutator, in which case it may have been removed from the commutator in the
first place!
In particular, we find that [Iia Ija , Iib Ijb ] = 0, a very relevant result for calculating commutators of
dipolar Hamiltonians. More generally,

[Iia Ija , Ikb Ilb ] = iabe jl Iia Ikb Ije + jk Iia Ije Ilb + il Ikb Iie Ija + ik Iie Ilb Ija .
Evaluating commutators of many-spin Hamiltonians of the form

i6=j

a a a
Dij
Ii Ij , where of course

210

Appendix B. Spin Algebra

a
a
Dij
= Dji
, is usually no more difficult than keeping track of sums:

XX
XX
a a a
b b b

Dij
Ii Ij ,
Dkl
Ik Il =
i6=j

=i

k6=l

X
ab

abe

XX X
i

a
b a b e
a
b a e b
a
b b e a
a
b e b a
Dji
Dki
Ij Ik Ii + Dji
Dik
Ij Ii Ik + Dij
Dki
Ik Ii Ij + Dij
Dik
Ii Ik Ij

j6=i k6=i,j

= 4i

abe

ab

B.3

XX X
i

a
b e a b
Dij
Dik
Ii Ij Ik .

(B.15)

j6=i k6=i,j

Diagonalizing a 2 2 Hermitian Matrix

As a final note, a common algebraic task is the diagonalization of a 2 2 Hermitian matrix. This
can be done very elegantly using the spin-1/2 formalism above. A general Hermitian matrix with
real components a, b, c, and can be written

M =

a
ce

ce

= a + b + (a b) sec m
I,
2
b

(B.16)

where and are the polar angles of m, and tan = 2c/(a b). This is made diagonal by using
=x
sin y
cos . Then
Eq. (B.10). Consider the vector n
iIn = I (m
cos m
n
sin .) = I z,
eiIn I me

(B.17)

so
eiIn M eiIn =

a+b
+ (a b) sec I z ,
2

(B.18)

the latter of which is diagonal with the expected eigenvalues


a+b ab
=

2
2

1+

4c2
.
(a b)2

(B.19)

The diagonalizing matrix is evidently

T = eiIn =

cos 2
sin 2 ei

sin 2 ei
cos 2

the normalized eigenvectors are the columns of this matrix.

(B.20)

Bibliography
[1] A. Einstein, B. Podolsky, and N. Rosen, Can quantum-mechanical description of physical
reality be considered complete?, Phys. Rev. 47, 777 (1935).
[2] J. S. Bell, On the Einstein-Podolsky-Rosen paradox, Physics 1, 195 (1964).
[3] E. Schr
odinger, Die Gegenw
artige Situation in der Quantenmechanik (The present situation
in quantum mechanics), Naturwissenschaften 48, 807; 49, 823; 50, 844 (1935).
[4] A. J. Leggett, Testing the limits of quantum mechanics: motivation, state of play, prospects,
J. Phys.: Condens. Matter 14, R415 (2002).
[5] M. Brune et al., Observing the progressive decoherence of the meter in a quantum measurement, Phys. Rev. Lett. 77, 4887 (1996).
[6] A. M. Turing, On computable numbers, with an application to the Entscheidungsproblem,
Proc. Lond. Math. Soc. 2 42, 230 (1936).
[7] R.P. Feynman, Simulating physics with computers, Int. J. Theor. Phys. 21, 467 (1982).
[8] D. Deutsch, Quantum theory, the Church-Turing principle, and the universal quantum computer, Proc. R. Soc. London A 400, 97 (1985).
[9] D. Deutsch and R. Josza, Rapid solution of problems by quantum computation, Proc. R. Soc.
London A 439, 553 (1992).
[10] D. Simon, On the power of quantum computation, In Proceedings, 35th Annual Symposium
on Foundations of Computer Science. IEEE Press (1994).
[11] P. W. Shor, Algorithms for quantum computation: discrete logarithms and factoring, In
Proceedings, 35th Annual Symposium on Foundations of Computer Science. IEEE Press (1994).

212

Bibliography

[12] J. W. Cooley and O. W. Tukey, An algorithm for the machine calculation of complex Fourier
series, Math. Comput. 19, 297 (1965).
[13] R. Van Meter and K. M. Itoh, Fast quantum modular exponentiation, Phys. Rev. A 71 (2005),
in press.
[14] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, Cambridge
University Press (2000).
[15] R. Rivest, A. Shamir, and L. Adleman, A method for obtaining digital signatures and publickey cryptosystems, Communications of the ACM 21, 120 (1978).
[16] L. Grover, Quantum mechanics helps in searching for a needle in a haystack, Phys. Rev. Lett.
79, 325 (1997).
[17] P. Shor, Scheme for reducing decoherence in quantum computer memory, Phys. Rev. A 52,
2493 (1995).
[18] A.M. Steane, Error correcting codes in quantum theory, Phys. Rev. Lett. 77, 793 (1996).
[19] A. R. Calderbank and P. W. Shor, Good quantum error-correcting codes exist, Phys. Rev. A
54, 1098 (1996).
[20] A. M. Steane, Multiple particle interference and quantum error correction, Proc. R. Soc.
London A , 2551 (1996).
[21] D. Gottesman, Class of quantum error-correcting codes saturating the quantum Hamming
bound, Phys. Rev. A 54, 1862 (1996).
[22] P. Zanardi and M. Rasetti, Noiseless quantum codes, Phys. Rev. Lett. 79, 3306 (1998).
[23] D. A. Lidar, I. L. Chuang, and K. B. Whaley, Decoherence-free subspaces for quantum
computation, Phys. Rev. Lett. 81, 2594 (1998).
[24] D. Kielpinski et al., A decoherence-free quantum memory using trapped ions, Science 291,
1013 (2001).
[25] C. F. Roos et al., Bell states of atoms with ultralong lifetimes and their tomographic state
analysis, Phys. Rev. Lett. 92, 220402 (2004).
[26] P. Shor, Fault-tolerant quantum computation, In Proceedings, 37th Annual Symposium on
Fundamentals of Computer Science page 56. IEEE Press (1996).

Bibliography

213

[27] D. P. DiVincenzo, The physical implementation of quantum computation, Fortschr. Phys. 48,
771 (2000).
[28] R. Blume-Kohout, C. M. Caves, and I. H. Deutsch, Climbing mount scalable: physical resource
requirements for a scalable quantum computer, Foundations of Physics 32, 1641 (2002).
[29] D. P. DiVincenzo, Two-bit gates are universal for quantum computation, Phys. Rev. A 51,
1015 (1995).
[30] A. Barenco et al., Elementary gates for quantum computation, Phys. Rev. A 52, 3457 (1995).
[31] J. L. Dodd, M. A. Nielsen, M. J. Bremner, and R. Thew, Universal quantum computation and
simulation using any entangling Hamiltonian and local unitaries, Phys. Rev. A 65, 040301R
(2002).
[32] D. Bacon, J. Kempe, D. Lidar, and K.B. Whaley, Universal fault tolerant quantum computation on decoherence-free subspaces, Phys. Rev. Lett. 85, 1758 (2000).
[33] J. Levy, Universal quantum computation with spin-1/2 pairs and Heisenberg exchange, Phys.
Rev. Lett. 89, 147902 (2002).
[34] A. M. Childs, D. W. Leung, and M. A. Nielsen, Unified derivations of measurement-based
schemes for quantum computation, Phys. Rev. A (2005), in press.
[35] E. Knill and R. Laflamme, The power of one bit of quantum information, Phys. Rev. Lett.
81, 5672 (1998).
[36] N. A. Gershenfeld and I. L. Chuang, Bulk spin-resonance quantum computation, Science 275,
350 (1997).
[37] D. G. Cory, A. F. Fahmy, and T. F. Havel, Ensemble quantum computing by nmr spectroscopy,
Proc. Natl. Acad. Sci. 94, 1634 (1997).
[38] W. S. Warren, The usefulness of nmr quantum computing, Science 277, 1688 (1997).
[39] L. J. Schulman and U. V. Vazirani, Molecular scale heat engines and scalable quantum computation, in Proceedings of the 31st stoc, p. 322, acm press, New York (1999).
[40] D. E. Chang, L. M. K. Vandersypen, and M. Steffen, nmr implementation of a building block
for scalable quantum computation, Chem. Phys. Lett. 338, 337 (2001).

214

Bibliography

[41] L. J. Schulman, T. Mor, and Y. Weinstein, Physical limits of heat-bath algorithmic cooling,
Phys. Rev. Lett. 94, 120501 (2005).
[42] C. H. Bennett et al., Teleporting an unknown quantum state via dual classical and EinsteinPodolsky-Rosen channels, Phys. Rev. Lett. 70, 1895 (1993).
[43] C. H. Bennett et al., Purification of noisy entanglement and faithful teleportation via noisy
channels, Phys. Rev. Lett. 76, 722 (1996).
[44] E. Yablonovitch et al., Optoelectronic quantum telecommunications based on spins in semiconductors, Proc. IEEE 91, 761 (2003).
[45] L. Childress, J. M. Taylor, A.S. Sorensen, and M. D. Lukin, Fault-tolerant quantum repeaters
with minimal physical resources, and implementations based on single photon emitters, e-print:
quant-ph/0502112 (2005).
[46] C. Monroe, D. M. Meekhof, B. E. King, W. M. Itano, and D. J. Wineland, Demonstration of
a fundamental quantum logic gate, Phys. Rev. Lett. 75, 4714 (1995).
[47] R. J. Rafac et al., Sub-dekahertz ultraviolet spectroscopy of 199 Hg+ , Phys. Rev. Lett. 85, 2462
(2000).
[48] B. E. King et al., Cooling the collective motion of trapped ions to initialize a quantum register,
Phys. Rev. Lett. 81, 1525 (1998).
[49] D. J. Winelend et al., Experimental issues in coherent quantum-state manipulation of trapped
atomic ions, J. Res. Natl. Inst. Stand. Technol. 103, 259328 (1998).
[50] D. Liebfried et al., Experimental demonstration of a robust, high-fidelity geometric two ionqubit phase gate, Nature 422, 412 (2003).
[51] F. Schmidt-Kaler et al., Realization of the Cirac-Zoller controlled-not quantum gate, Nature
422, 408 (2003).
[52] F. Schmidt-Kaler et al., The coherence of qubits based on single Ca+ ions, J. Phys. B: At.
Mol. Opt. Phys. 36, 623 (2003).
[53] S. Gulde et al., Implementation of the Deutsch-Jozsa algorithm on an ion-trap quantum
computer, Nature 421, 48 (2003).

Bibliography

215

[54] J. I. Cirac and P. Zoller, Quantum computations with cold trapped ions, Phys. Rev. Lett. 74,
4091 (1995).
[55] A. M. Steane and D. M. Lucas, Quantum computing with trapped ions, atoms and light,
Fortschr. Phys. 48, 839 (2000).
[56] D. Kielpinski, C. Monroe, and D. J. Wineland, Architecture for a large-scale ion-trap quantum
computer, Nature 417, 709 (2002).
[57] L.-M. Duan, M. Lukin, J. I. Cirac, and P. Zoller, Long-distance quantum communication with
atomic ensembles and linear optics, Nature 414, 413 (2001).
[58] B. B. Blinov, D. L. Moehring, L.-M. Duan, and C. Monroe, Observation of entanglement
between a single trapped atom and a single photon, Nature 428, 153 (2004).
[59] A. S. Parkins, P. Marte, P. Zoller, and H. J. Kimble, Sythesis of arbitrary quantum states via
adiabatic transfer of Zeeman coherence, Phys. Rev. Lett. 71, 3095 (1993).
[60] Q. A. Turchette, C. J. Hood, W. Lange, H. Mabuchi, and H. J. Kimble, Measurement of
conditional phase shifts for quantum logic, Phys. Rev. Lett. 75, 4710 (1995).
[61] T. Sleator and H. Weinfurter, Realizable universal quantum logic gates, Phys. Rev. Lett. 74,
4087 (1995).
[62] T. Pellizzari, S.A. Gardiner, J.I. Cirac, and P. Zoller, Decoherence, continuous observation,
and quantum computing: A cavity qed model, Phys. Rev. Lett. 75, 3788 (1995).
[63] G. K. Brennen, C. M. Caves, P. S. Jessen, and I. H. Deutsch, Quantum logic gates in optical
lattices, Phys. Rev. Lett. 82, 1060 (1999).
[64] D. Jaksch, H.-J. Briegel, J. I. Cirac, C. W. Gardiner, and P. Zoller, Entanglement of atoms
via cold controlled collisions, Phys. Rev. Lett. 82, 1975 (1999).
[65] V. Giovannetti, D. Vitali, P. Tombesi, and A. Ekert, Scalable quantum computation with
cavity QED systems, Phys. Rev. A 62, 032306 (2000).
[66] D. DeMille, Quantum computation with trapped polar molecules, Phys. Rev. Lett. 88, 67901
(2002).
[67] L. M. K. Vandersypen et al., Experimental realization of Shors quantum factoring algorithm
using nuclear magnetic resonance, Nature 414, 883 (2001).

216

Bibliography

[68] A. Imamoglu, D.D. Awschalom, G. Burkard, D.P. DiVincenzo, D. Loss, M. Shermin, and
A. Small, Quantum information processing using quantum dot spins and cavity qed, Phys.
Rev. Lett. 83, 4204 (1999).
[69] N. H. Bonadeo, J. Erland, D. Gammon, D. Park, D. S. Katzer, and D. G. Steel, Coherent
optical control of the quantum state of a single quantum dot, Science 282, 1473 (1998).
[70] G. D. Sanders, K. W. Kim, and W. C. Holton, Scalable solid-state quantum computer based
on quantum dot pillar structures, Phys. Rev. A 60, 4146 (1999).
[71] T. A. Brun and H. Wang, Coupling nanocrystals to a high-Q silica microsphere: Entanglement
in quantum dots via photon exchange, Phys. Rev. A 61, 032307 (2000).
[72] J. H. Reina, L. Quiroga, and N. F. Johnson, Quantum entanglement and information processing via excitons in optically driven quantum dots, Phys. Rev. A 62, 012305 (2000).
[73] J.A. Gupta, D.D. Awschalom, X. Peng, and A.P. Alizisatos, Spin coherence in semiconductor
quantum dots, Phys. Rev. B 59, 10421 (1999).
[74] R. J. Epstein, D. T. Fuchs, W. V. Schoenfield, P. M. Petroff, and D. D. Awschalom, Hanle
effect measurements of spin lifetimes in InAs self-assembled quantum dots, Appl. Phys. Lett.
78, 733 (2001).
[75] M. Paillar et al., Spin relaxation quenching in semiconductor quantum dots, Phys. Rev. Lett.
86, 1634 (2001).
[76] A. S. Bracker et al., Optical pumping of the electronic and nuclear spin of single charge-tunable
quantum dots, Phys. Rev. Lett. 94, 047402 (2005).
[77] M. Bayer et al.,

Fine structure of neutral and changed excitons in self-assembled

In(Ga)As/(Al)GaAs quantum dots, Phys. Rev. B 65, 195315 (2002).


[78] P. Chen, C. Piermarocchi, L. J. Sham, D. Gammon, and D. G. Steel, Theory of quantum
optical control of a single spin in a quantum dot, Phys. Rev. B 69, 75320 (2004).
[79] C. Piermarocchi, P. Chen, L. J. Sham, and D. G. Steel, Optical rkky interaction between
charged semiconductor quantum dots, Phys. Rev. Lett. 89, 167402 (2002).
[80] M. S. Shahriar, P. R. Hemmer, S. Lloyd, P. S. Bhatia, and A. E. Craig, Solid-state quantum
computing using spectral holes, Phys. Rev. A 66, 032301 (2002).

Bibliography

217

[81] C. Kurtziefer, S. Mayer, P. Zarda, and H. Weinfurter, Stable solid-state source of single
photons, Phys. Rev. Lett. 85, 290 (2000).
[82] F. Jelezko, T. Gaebel, I. Popa, M. Domhan, A. Gruber, and J. Wrachtrup, Observation of
coherent oscillation of a single nuclear spin and realization of a two-qubit conditional quantum
gate, Phys. Rev. Lett. 93, 130501 (2004).
[83] D. Loss and D. P. DiVincenzo, Quantum computation with quantum dots, Phys. Rev. A 57,
120 (1998).
[84] V. Cerletti, W. A. Coish, O. Gywat, and D. Loss, Recipes for spin-based quantum computing,
Nanotechnology 16, R27 (2005).
[85] P.M. Platzman and M. I. Dykman, Quantum computing with electrons floating on liquid
helium, Science 284, 1967 (1999).
[86] R. Vrijen, E. Yablonovitch, K. Wang, H. W. Jiang, A. Balandin, V. Roychowdhury, T. Mor,
and D. DiVincenzo, Electron-spin-resonance transistors for quantum computing in silicongermanium heterostructures, Phys. Rev. A 62, 012306 (2000).
[87] A.M. Tyryshkin, S. A. Lyon, A. V. Astashkin, and A. M. Raitsimring, Electron spin-relaxation
times of phosphorous donors in silicon, Phys. Rev. B 68, 193207 (2003).
[88] A. M. Tyrishkin, S. A. Lyon, W. Jantsch, and F. Sch
affler, Spin manipulation of free twodimensional electrons in Si/SiGe quantum wells, Phys. Rev. Lett. 94, 126802 (2005).
[89] R. De Sousa, J. D. Delgado, and S. Das Sarma, Silicon quantum computation based on
magnetic dipolar coupling, Phys. Rev. A 70, 052304 (2004).
[90] J.E. Mooij et al., Josephson persistent-current qubit, Science 285, 1036 (1999).
[91] I. Chiorescu, Y. Nakamura, C.J.P.M. Harmans, and J.E. Mooij, Coherent quantum dynamics
of a superconducting flux qubit, Science 299, 18691871 (2003).
[92] E. Ilichev et al., Continuous monitoring of Rabi oscillations in a Josephson flux qubit, Phys.
Rev. Lett. 91, 097906 (2003).
[93] A. J. Berkley et al., Entangled macroscopic quantum states in two superconducting qubits,
Science 300, 15481550 (2003).

218

Bibliography

[94] Y. Yu, S. Han, X. Chu, S-I. Chu, and Z. Wang, Coherent temporal oscillations of macroscopic
quantum states in a Josephson junction, Science 296, 889892 (2002).
[95] J.M. Martinis, S. Nam, J. Aumentado, and C. Urbina, Rabi oscillations in a large Josephsonjunction qubit, Phys. Rev. Lett. 89, 117901 (2002).
[96] Y. Nakamura, Yu. A. Pashkin, and J.S. Tsai, Coherent control of macroscopic quantum states
in a single-Cooper-pair box, Nature 398, 786 (1999).
[97] D. Vion et al., Manipulating the quantum state of an electrical circuit, Science 296, 886889
(2002).
[98] T. Yamamoto, Yu. A. Pashkin, O. Astaflev, Y. Nakamura, and J. S. Tsai, Demonstration of
conditional gate operation using superconducting charge qubits, Nature 425, 941 (2003).
[99] D. I. Schuster et al., ac stark shift and dephasing of a superconducting qubit strongly coupled
to a cavity field, Phys. Rev. Lett. 94, 123602 (2005).
[100] L. Tian, P. Rabl, R. Blatt, and P. Zoller, Interfacing quantum-optical and solid-state qubits,
Phys. Rev. Lett. 92, 247902 (2004).
[101] J. A. Sidles, Folded Stern-Gerlach experiment as a means for detecting nuclear magnetic
resonance in individual nuclei, Phys. Rev. Lett. 68, 1124 (1992).
[102] D. Rugar, C. S. Yannoni, and J. A. Sidles, Mechanical detection of magnetic resonance, Nature
360, 563 (1992).
[103] V. Privman, I. D. Vagner, and G. Kventsel, Quantum computation in quantum-hall systems,
Phys. Lett. A 239, 141 (1998).
[104] D. Mozyrsky, V. Privman, and M. L. Glasser, Indirect interaction of solid-state qubits via
two-dimensional electron gas, Phys. Rev. Lett. 86, 5112 (2001).
[105] D. Mozyrsky, V. Privman, and I. D. Vagner, Nuclear-spin qubit dephasing time in the integer
quantum Hall effect regime, Phys. Rev. B 63, 85313 (2001).
[106] R. G. Mani, W. B. Johnson, and V. Narayanamurti, Nuclear spin based quantum information
processing at high magnetic fields, Nanotechnology 14, 515 (2003).
[107] B.E. Kane, A silicon-based nuclear spin quantum computer, Nature 393, 133 (1998).

Bibliography

219

[108] B. Koiller, X. Hu, and S. Das Sarma, Exchange in silicon-based quantum computer architecture, Phys. Rev. Lett. 88, 027903 (2002).
[109] B. Koiller, X. Hu, and S. Das Sarma, Strain effects on silicon donor exchange: Quantum
computer architecture considerations, Phys. Rev. B 66, 115201 (2002).
[110] J. L. OBrien et al., Towards the fabrication of phosphorous qubits for a silicon quantum
computer, Phys. Rev. B 64, 161401 (2001).
[111] T. M. Buehler, D. J. Reilly, R. Brenner, A. R. Hamilton, A.S. Dzurak, and R.G. Clark,
Correlated charge detection for readout of a solid-state quantum computer, Appl. Phys. Lett.
82, 577 (2003).
[112] D. G. Cory et al., nmr based quantum information processing: achievements and prospects,
Fortschr. Phys. 48, 875 (2000).
[113] G. M. Leskowits, N. Ghaderi, R. A. Olsen, and L. J. Mueller, Three-qubit nuclear magnetic
resonance quantum information processing with a single-crystal solid, J. Chem. Physics 119,
1643 (2003).
[114] F. Yamaguchi and Y. Yamamoto, Crystal lattice quantum computer, Appl. Phys. A: Mater.
Sci. Process. 68, 1 (1999).
[115] E. Knill, R. Laflamme, and G. J. Milburn, A scheme for efficient quantum computation with
linear optics, Nature 409, 46 (2001).
[116] D. Gottesman and I. L. Chuang, Demonstrating the viability of universal quantum computation using teleportation and single-qubit operations, Nature 402, 390 (1999).
[117] T. D. Ladd, D. Maryenko, Y. Yamamoto, E. Abe, and K. M. Itoh, Coherence time of decoupled
nuclear spins in silicon, Phys. Rev. B 71, 14401 (2005).
[118] J.J. Bollinger, D.J. Heinzen, W.M. Itano, S.L. Gilbert, and D.J. Wineland, A 303-MHz
frequency standard based on trapped Be+ ions, IEEE Trans. Insrum. Measurement 40, 126
(1991).
[119] M. V. Romalis and M. P. Ledbetter, Transverse spin relaxation in liquid 129 Xe in the presence
of large dipolar fields, Phys. Rev. Lett. 87, 67601 (2001).
[120] T. E. Chupp, E. R. Oteiza, J. M. Richardson, and T. R. White, Precision frequency measurements with polarized 3 He,

21

Ne, and

129

Xe atoms, Phys. Rev. A 38, 3998 (1988).

220

Bibliography

[121] A. Abragam, Principles of Nuclear Magnetism, Clarendon Press (1961).


[122] C. P. Slichter, Principles of Magnetic Resonance, Springer, third edition (1990).
[123] D. I. Hoult and N. S. Ginsberg, The quantum origins of the free induction decay signal and
spin noise, J. Magn. Reson. 148, 182 (2001).
[124] M. Steffen, L. M. K. Vandersypen, and I. L. Chuang, Simultaneous soft pulses applied at
nearby frequencies, J. Magn. Reson. 146, 369 (2000).
[125] M. Mehring, Principles of High Resolution NMR in Solids, Springer-Verlag (1983).
[126] B. Cowan, Nuclear Magnetic Resonance and Relaxation, Cambridge (1997).
[127] F. Bloch, Nuclear induction, Phys. Rev. 70, 460 (1946).
[128] L. M. K. Vandersypen, C. S. Yannoni, M. H. Sherwood, and I. L. Chuang, Realization of
logically labeled effective pure states for bulk quantum computation, Phys. Rev. Lett. 83,
3085 (1999).
[129] E. Knill, I. Chuang, and R. Laflamme, Effective pure states for bulk quantum computation,
Phys. Rev. A 57, 3348 (1998).
[130] P. H
ubler, J. Bargon, and S. J. Glaser, Nuclear magnetic resonance quantum computing
exploiting the pure spin state of para hydrogen, J. Chem. Physics 113, 2056 (2000).
[131] A. S. Verhulst, O. Liivak, M. H. Sherwood, H. M. Vieth, and I. L. Chuang, Non-thermal
nuclear magnetic resonance quantum computing using hyperpolarized xenon, Appl. Phys.
Lett. 79, 2480 (2001).
[132] P. L. Kuhns, P. C. Hammel, O. Gonen, and J.S. Waugh, Unexpectedly rapid

19

F spin-lattice

relaxation in CaF2 below 1 K, Phys. Rev. B 35, 4591 (1987).


[133] A. W. Overhauser, Polarization of nuclei in metals, Phys. Rev. 92, 411 (1953).
[134] E. L. Hahn, Spin echoes, Phys. Rev. 80, 580 (1950).
[135] W-K. Rhim and H. Kessemeier, Transverse-magnetization recovery in the rotating frame,
Phys. Rev. B 3, 3655 (1971).
[136] G. Lampel, Nuclear dynamic polarization by optical electronic saturation and optical pumping
in semiconductors, Phys. Rev. Lett. 20, 491 (1968).

Bibliography

221

[137] W. G. Clark and G. Feher, Nuclear polarization in InSb by a dc current, Phys. Rev. Lett. 10,
134 (1963).
[138] T. D. Ladd, J. R. Goldman, F. Yamaguchi, and Y. Yamamoto, Decoherence in crystal lattice
quantum computation, Appl. Phys. A: Mater. Sci. Process. 71, 27 (2000).
[139] A. Goto, T. Shimizu, K. Hashi, H. Kitazawa, and S. Ohki, Decoupling-free nmr quantum
computer on a quantum spin chain, Phys. Rev. A 67, 22312 (2003).
[140] A. Khitun, R. Ostroumov, and K. L. Wang, Spin-wave utilization in a quantum computer,
Phys. Rev. A 64, 62304 (2001).
[141] J. R. Goldman, T. D. Ladd, F. Yamaguchi, and Y.Yamamoto, Magnet designs for a crystallattice quantum computer, Appl. Phys. A: Mater. Sci. Process. 71, 11 (2000).
[142] A. Z. Genack and A. G. Redfield, Nuclear spin diffusion and its thermodynamic quenching in
the field gradients of a type-II superconductor, Phys. Rev. Lett. 31, 1204 (1973).
[143] W. Van Der Lugt and W. J. Caspers, Nuclear magnetic resonance line shape of fluorine in
apatite, Physica 30, 1658 (1964).
[144] M. Engelsberg, I.J. Lowe, and J. L. Carolan, Nuclear-magnetic-resonance line shape of a linear
chain of spins, Phys. Rev. B 7, 924 (1973).
[145] G. Cho and J. P. Yesinowski,

H and

19

F multiple-quantum nmr dynamics in quasi-one-

dimensional spin clusters in apatites, J. Phys. Chem. 100, 15716 (1996).


[146] H. Cho, T. D. Ladd, J. Baugh, D. G. Cory, and C. Ramanathan, Multi-spin dynamics of the
solid-state nmr free induction decay, Phys. Rev. B (2005), in press.
[147] N. Bloembergen, On the interaction of nuclear spins in a crystalline lattice, Physica 15, 386
(1949).
[148] S. Oishi and T. Kamiya, Flux growth of fluorapatite crystals, Nippon Kagaku Kaishi 9, 800
(1994).
[149] I. N. Kurkin and E. A. Tsvetkov,

Local debye temperature of Cr5+ impurity ions in

Ca5 (PO4 )3 F single crystals, Sov. Phys. Solid State 20, 870 (1978).
[150] D. I. Hoult and R. E. Richards, The signal-to-noise ratio of the nuclear magnetic resonance
experiment, J. Magn. Reson. 24, 71 (1976).

222

Bibliography

[151] D. D. Awschalom, J. F. Smyth, G. Grinstein, D. P. DiVincenzo, and D. Loss, Macroscopic


quantum tunneling in magnetic proteins, Phys. Rev. Lett. 68, 3092 (1992).
[152] D. Rugar, R. Budakian, H. J. Mamin, and B. W. Chui, Single spin detection by magnetic
resonance force microscopy, Nature 430, 329 (2004).
[153] J. Wrachtrup, A. Gruber, L. Fleury, and C. von Borczyskowski, Magnetic resonance on single
nuclei, Chem. Phys. Lett. 267, 179 (1997).
[154] S. L. Braunstein et al., Separability of very noisy mixed states and implications for nmr
quantum computing, Phys. Rev. Lett. 84, 1054 (1990).
[155] R. Schack and C. M. Caves, Classical model for bulk-ensemble nmr quantum computation,
Phys. Rev. A 60, 4354.
[156] N. C. Menicucci and C. M. Caves, Local realistic model for the dynamics of bulk-ensemble
nmr information processing, Phys. Rev. Lett. 88, 167901 (2002).
[157] N. Linden and S. Popescu, Good dynamics versus bad kinematics: Is entanglement needed
for quantum computation?, Phys. Rev. Lett. 87, 047901 (2001).
[158] M. A. Ruderman and C. Kittel, Indirect exchange coupling of nuclear magnetic moments by
conduction electrons, Phys. Rev. 96, 99 (1954).
[159] H. Suhl, Effective nuclear spin interactions in ferromagnets, Phys. Rev. 109, 606 (1958).
[160] T. Nakamura, Indirect coupling of nuclear spins in antiferromagnet with particular reference
to MnF2 at very low temperature, Prog. Theor. Phys. 20, 542 (1958).
[161] T.D. Ladd, J. R. Goldman, F. Yamaguchi, Y. Yamamoto, E. Abe, and K.M. Itoh, All-silicon
quantum computer, Phys. Rev. Lett. 89, 17901 (2002).
[162] K. M. Itoh and E. E. Haller, Isotopically engineered semiconductors new media for the
investigation of nuclear spin related effects in solids, Physica E 10, 463 (2001).
[163] G. P. Berman, G. W. Brown, M. E. Hawley, and V. I. Tsifrinovich, Solid-state quantum
computer based on scanning tunneling microscopy, Phys. Rev. Lett. 87, 97902 (2001).
[164] G. P. Berman, G. D. Doolen, P. C. Hammel, and V. I. Tsifrinovich, Magnetic resonance force
microscopy quantum computer with tellurium donors in silicon, Phys. Rev. Lett. 86, 2894
(2001).

Bibliography

223

[165] I. Shlimak, V. I. Safarov, and I. D. Vagner, Isotopically engineered silicon/silicon-germanium


nanostructures as basic elements for a nuclear spin quantum computer, J. Phys: Condens.
Matter 13, 6059 (2001).
[166] J. Viernow, J. L. Lin, D. Y. Petrovykh, F. M. Leibsle, F. K. Men, and F. J. Himpsel, Regular
step arrays on silicon, Appl. Phys. Lett. 72, 948 (1998).
[167] J. L. Lin, D. Y. Petrovykh, J. Viernow, F. K. Men, D. J. Seo, and F. J. Himpsel, Formation
of regular step arrays on Si(111)7 7, J. Appl. Phys. 84, 255 (1998).
[168] C. Ke-Ming, J. Gao-Long, S. Chi, and Y. Ming-Ren, The growth dynamics of Si(111) mbe
studied by rheed intensity oscillations, Acta. Phys. Sin. 39, 1945 (1990).
[169] A. V. Latyshev, A. L. Aseev, A. B. Krasilnikov, and S. I. Stenin, Initial stages of silicon
homoepitaxy studied by in situ reflection electron microscopy, Phys. Status Solidi A 113, 421
(1989).
[170] H. Bracht, E. E. Haller, and R. Clark-Phelps, Silicon self-diffusion in isotope heterostructures,
Phys. Rev. Lett. 81, 393 (1998).
[171] T. Sekiguchi, S. Yoshida, and K. M. Itoh, Self-assembly of parallel atomic wires and periodic
clusters of silicon on a vicinal Si(111) surface, Phys. Rev. Lett. (2005), in press.
[172] N. T. Bagraev, L. S. Vlasenko, and R. A. Zhitnikov, Inversion of nuclear magnetization of
compensated silicon in interband absorption of light in weak magnetic fields, JETP Letters
25, 190 (1977).
[173] G. Feher and E. A. Gere, Electron spin resonance experiments on donors in silicon. II. Electron
spin relaxation effects, Phys. Rev. 114, 1245 (1959).
[174] A. S. Verhulst, I. G. Rau, Y. Yamamoto, and K. M. Itoh, Optical pumping of

29

Si nuclear

spins in bulk silicon at high magnetic field and liquid helium temperature, Phys. Rev. B 71,
235206 (2005).
[175] T. D. Stowe et al., Attonewton force detection using ultrathin silicon cantilevers, Appl. Phys.
Lett. 71, 288 (1997).
[176] U. D
urig, H. R. Steinauer, and N. Blanc, Dynamic force microscopy by means of the phasecontrolled oscillator method, J. Appl. Phys. 82, 3641 (1997).

224

Bibliography

[177] T. B. Gabrielson, Mechanical-thermal noise in micromachined acoustic and vibration sensors,


IEEE Trans. Electron Devices 40, 903 (1993).
[178] F. R. Blom, S. Bouwstra, M. Elwenspoek, and J. H. J. Fluitman, Dependence of the quality
factor of micromachined silicon beam resonators on pressure and geometry, J. Vac. Sci.
Technol. B 10, 19 (1992).
[179] K-M. C. Fu, T. D. Ladd, C. Santori, and Y. Yamamoto, Optical detection of the spin state of
a single nucleus in silicon, Phys. Rev. B 69, 125306 (2004).
[180] D. Karaiskaj, M.L.W. Thewalt, T. Ruf, M. Cardona, H.-J.Pohl, G.G. Deviatych, P.G. Sennivok, and H. Riemann, Photoluminescence of isotopically purified silicon: How sharp are
bound exciton transitions?, Phys. Rev. Lett. 86, 6010 (2001).
[181] U. O. Ziemelis, R. R. Parsons, and M. L. W. Thewalt, Temperature dependent zeeman study
of bound exciton and bound multiexciton complex photoluminescence in Si(P) and Si(Li),
Can. J. Phys. 60, 222 (1982).
[182] G. Feher, Observation of nuclear magnetic resonances via the electron spin resonance line,
Phys. Rev. 103, 834 (1956).
[183] W. Kohn, Shallow impurity states in silicon and germanium, in Solid State Physics, vol. 5, p.
257, Academic (1957).
[184] G. Feher, Nuclear polarization via hot conduction electrons, Phys. Rev. Lett. 3, 135 (1959).
[185] W. Schmid, Auger lifetimes for excitons bound to neutral donors and acceptors in Si, Phys.
Status Solidi B 84, 529 (1977).
[186] J. Vuckovic and Y. Yamamoto, Photonic crystal microcavities for cavity quantum electrodynamics with a single quantum dot, Appl. Phys. Lett. 82, 2374 (2003).
[187] B. Cabrera et al., Cryogenic detectors based on superconducting transition-edge sensors for
time-energy-resolved single-photon counters and for dark matter searches, Physica B 280, 509
(2000).
[188] D. Gammon et al., Nuclear spectroscopy in single quantum dots: Nanoscopic Raman scattering
and nuclear magnetic resonance, Science 277, 85 (1997).
[189] D. Gammon et al., Electron and nuclear spin interactions in the optical spectra of single GaAs
quantum dots, Phys. Rev. Lett. 86, 5176 (2001).

Bibliography

225

[190] J. S. Waugh, L. M. Huber, and U. Haeberlin, Approach to high-resolution nmr in solids,


Phys. Rev. Lett. 20, 180 (1968).
[191] U. Haeberlin and J. S. Waugh, Coherent averaging effects in magnetic resonance, Phys. Rev.
175, 453 (1968).
[192] P. Mansfield, M. J. Orchard, D. C. Stalker, and K. H. B. Richards, Symmetrized multipulse
nuclear-magnetic-resonance experiments in solids: Measurement of the chemical-shift shielding
tensor in some compounds, Phys. Rev. B , 90 (1973).
[193] W-K. Rhim, D.D. Elleman, and R.W. Vaughan, Analysis of multiple pulse nmr in solids, J.
Chem. Physics 59, 3740 (1973).
[194] U. Haeberlin, J. D. Ellett, and J. S. Waugh, Resonance offset effects in multiple-pulse nmr
experiments, Phys. Rev. 55, 53 (1971).
[195] D. P. Burum and W-K. Rhim, Analysis of multiple pulse nmr in solids. III, J. Chem. Physics
71, 944 (1979).
[196] D. G. Cory, J. B. Miller, and A. N. Garroway, Time-suspension multiple-pulse sequences:
applications to solid-state imaging, J. Magn. Reson. 90, 205 (1990).
[197] G. S. Boutis, P. Cappellaro, H. Cho, C. Ramanathan, and D. G. Cory, Pulse error compensating symmetric magic-echo trains, J. Magn. Reson. 161, 132.
[198] S. Meiboom and D. Gill, Modified spin-echo method for measuring nuclear relaxation times,
Rev. Sci. Instrum. 29, 6881 (1958).
[199] M. H. Levitt and R. Freeman, Compensation for pulse imperfections in nmr spin-echo experiments, J. Magn. Reson. 43, 65 (1981).
[200] D. W. Leung, I. L. Chuang, F. Yamaguchi, and Y. Yamamoto, Efficient implementation of
coupled logic gates for quantum computation, Phys. Rev. A 61, 042310 (2000).
[201] F. Yamaguchi, T. D. Ladd, C. P. Master, Y. Yamamoto, and N. Khaneja, Efficient decoupling
and recoupling in solid-state nmr for quantum computation, (2005), in preparation.
[202] A.S. Verhulst, D. Maryenko, Y. Yamamoto, and K. M. Itoh, Double and single peaks in nuclear
magnetic resonance spectra of natural and
68, 054105 (2003).

29

Si enriched single crystal silicon, Phys. Rev. B

226

Bibliography

[203] K. Takyu, K. M. Itoh, K. Oka, N. Saito, and V. I. Ozhogin, Growth and characterization
of the isotopically enriched

28

Si bulk single crystal, Jpn. J. Appl. Phys., Part 2 38, L1493

(1999).
[204] S. Idziak and U. Haeberlin, Design and construction of a high homogeneity rf coil for solidstate multiple-pulse nmr, J. Magn. Reson. 50, 281 (1982).
[205] E. Fukushima and S. B. W. Roeder, Experimental Pulse nmr: A Nuts and Bolts Approach,
Addison-Wesley (1981).
[206] E. D. Ostroff and J. S. Waugh, Multiple spin echoes and spin locking in solids, Phys. Rev.
Lett. 16, 1097 (1966).
[207] D. Suwelack and J. S. Waugh, Quasistationary magnetization in pulsed spin-locking experiments in dipolar solids, Phys. Rev. B 22, 5110 (1980).
[208] A. E. Dementyev, D. Li, K. Maclean, and S. E. Barrett, Anomalies in the nmr of silicon:
Unexpected spin echoes in a dilute dipolar solid, Phys. Rev. B 68, 153302 (2003).
[209] H. Kessemeir and R. E. Norberg, Pulsed nuclear magnetic resonance in rotating solids, Phys.
Rev. 155, 321 (1967).
[210] R. Kubo and K. Tomita, A general theory of magnetic resonance absorption, J. Phys. Soc.
Japan 9, 888 (1954).
[211] N. Bloembergen, S. Shapiro, P. S. Pershan, and J. O. Artman, Cross-relaxation in spin systems,
Phys. Rev. 114, 445 (1959).
[212] J. Philippot, Spin-spin relaxation and spin temperatures, Phys. Rev. 133, A471 (1964).
[213] J. Jeener, H. Eisendrath, and R. Van Steenwinkel, Thermodynamics of spin systems in solids,
Phys. Rev. 133, A478 (1963).
[214] A. K. Paravastu et al., Optical polarization of nuclear spins in gaas, Phys. Rev. B 69, 075203
(2004).
[215] A. K. Paravastu, P. Coles, T. D. Ladd, R. S. Maxwell, and J. A. Reimer, Photocurrentmodulated optical nuclear polarization in bulk GaAs, (2005), in preparation.
[216] R. G. Shulman and B. J. Wyluda, Nuclear and magnetic resonance of
silicon, Phys. Rev. 103, 1127 (1956).

29

Si in n- and p-type

Bibliography

227

[217] N. Bloembergen, Nuclear magnetic relaxation in semiconductors, Physica 20, 1130 (1954).
[218] A. Abragam, J. Combrisson, and I. Solomon, Dynamic polarization of

29

Si in silicon, C.R.

Acad. Sc. Paris 246, 1035 (1958).


[219] B. Sapoval and D. Lepine, Examination of a system of dilute nuclear spins: spin-lattice
relaxation in very pure silicon, J. Phys. Chem. Solids 27, 115 (1966).
[220] R. K. Sundfors and D. F. Holcomb, Nuclear magnetic resonance studies of the metallic transition in doped silicon, Phys. Rev. 136, A810 (1964).
[221] D. Jerome, Effects on the nmr properties of the metal-nonmetal transistion in impure semiconductors, Rev. Mod. Phys. 40, 830 (1968).
[222] E. I. Grncharova and V. I. Perel, Relaxation of nuclear spins interacting with holes in
semiconductors, Sov. Phys. Semicond. 11, 997 (1978).
[223] N. T. Bagraev, L. S. Vlasenko, and R. A. Zhitnikov, Optical orientation of

29

Si nuclei in

n-type silicon and its dependence on the pumping light intensity, Sov. Phys. JETP 44, 500
(1976).
[224] G. Feher, Electron spin resonance experiments on donors in silicon. I. Electronic structure of
donors by the electron nuclear double resonance technique, Phys. Rev. 114, 1219 (1959).
[225] D. K. Wilson and G. Feher, Electron spin resonance experiments on donors in silicon. III.
Investigation of excited states by the application of uniaxial stress and their importance in
relaxation processes, Phys. Rev. 124, 1068 (1961).
[226] J. P. Gordon and K. D. Bowers, Microwave spin echoes from donor electrons in silicon, Phys.
Rev. Lett. 1, 368 (1958).
[227] P. Norton, T. Braggins, and H. Levinstein, Recombination of electrons at ionized donors in
silicon at low temperatures, Phys. Rev. Lett. 30, 488 (1973).
[228] M. B. Weissman, 1/f noise and other slow, nonexponential kinetics in condensed matter, Rev.
Mod. Phys. 60, 537 (1988).
[229] R. de Sousa and S. Das Sarma, Theory of nuclear-induced spectral diffusion: Spin decoherence
of phosphorus donors in Si and GaAs quantum dots, Phys. Rev. B 68, 115322 (2003).

228

Bibliography

[230] D. Paget, G. Lampel, B. Sapoval, and V. I. Safarov, Low field electron-nuclear spin coupling
in gallium arsenide under optical pumping conditions, Phys. Rev. B 15, 5780 (1977).
[231] P. Y. Yu and M. Cordona, Fundamentals of Semiconductors, Springer (2001).

Index
-system, 37, 45

dimethyl siloxane (dmso), 133, 139

-pulses, 119, 126

dipolar bath, 63

2-dimensional electron gas (2deg), 48, 51

dipolar coupling, 61, 84


dipolar decoherence, 75, 109, 110

average Hamiltonian theory (aht), 113, 118


Bells inequalities, 2
bit-flip code, 13
bits, 24

discrete Fourier transform (dft), 8, 141


discrete logarithms, 12
DiVincenzo criteria, 21
dynamic nuclear polarization, 69

Bloch equations, 64

effective field, 113, 144

boosting, 30

Einstein, A., 2

br-24 pulse sequence, 115

Einstein-Podolsky-Rosen (epr), 2, 35

broadband pulse, 60, 111, 130

electron spin decoherence, 44

bulk susceptibility, 57

entangled photon pairs, 37

chemical shift, 57, 72


Church-Turing thesis, 5, 11
coherence number, 63

entanglement, 3, 7
entanglement purification, 36
entropy, 24

controlled-not, 22

fast Fourier transform (fft), 9

Cory 48-pulse sequence, 115

fault-tolerant quantum computation, 12

counter-rotating term, 59

Fermis golden rule, 153, 177, 187

decoherence, 3, 12, 22, 153


decoherence-free subspaces, 16, 42, 106
decoupling, 75, 94, 107, 126, 190

Feynman, R. P., 6
Flocquet Hamiltonian, 113, 127
fluorapatite (Ca5 F(PO4 )3 ), 80
free induction decay (fid), 60, 137

dephasing, see inhomogeneous broadening


Deutsch, D., 6

gestalt, 4

Deutsch-Josza algorithm, 6

gravity, 1, 3

diamond, see NV centers

Grovers algorithm, 12

230

Index

Hadamard gate, 7

multiple pulse sequence (mps), 131, 190

Hadamard matrix, 128

multiple pulse sequences (mps), 111

hard pulse, see broadband pulse

multiple quantum coherences, 107

hidden subgroup problem, 12


hidden variables, 2
Hilbert space, 4, 5, 21, 156
Hydrogen atom, 22

nats, 24
neutral atoms, 42
no-cloning theorem, 13, 35
nuclear magnetic resonance (nmr), 55

inhomogeneous broadening, 58, 107, 109, 115,


118, 119, 127
interaction Hamiltonian, 65
J-coupling, 78, 84

T1 , see thermal relaxation


T2 , see decoherence
decoherence, 55, 63, 64, 72, 153, 167, 177
in liquids, 43, 66
initialization, 66

Kane proposal, 52, 86, 88

proposals for quantum computing, 43, 51

Knight shift, 57

signal-to-noise ratio (snr), 55, 72, 81, 131

Larmor frequency, 57
laser cooling, 42
linear optics quantum computers, 23, 34, 53
locality, 2
Mach-Zender interferometer, 37, 100
magic angle, 62, 79
spinning (mas), 118, 131, 150
magic echoes, 70
sme-16, 115
magnetic field gradient, 72, 94, 126
magnetic noise, 65
magnetic resonance force microscopy (mrfm),
56, 73, 82, 97
magnetic resonance imaging (mri), 56, 72, 126
Magnus expansion, 113
Maxwells demon, 26

spectroscopy, 55, 57, 118


thermal relaxation, 64, 153, 164, 177
number field sieve (nfs), 11
NV centers, 47, 56, 82, 105
optical fibers, 35
oracle, 6
Overhauser effect, 70
optical, 71, 92, 96
Overhauser shift, 103
period finding, 9
perturbation theory, 84
phase estimation algorithm, 9
phonons, 80, 98
pseudo-pure states, 67
state labelling, 67
temporal averaging, 68

modular exponentiation, 10
mrev, 114

quad echo, 145

Index

231

quantum bits, see qubits

Simons algorithm, 9

quantum communication, 34

single electron transistor (set), 48, 52

quantum cryptography, 34

single nuclear spin detection, 47, 98, 108, 187

quantum dots, 40, 43, 102

single photon generation, 47

quantum error correction, 5, 12

soft pulse, 60, 127, 130

entropy accounting, 26

solid-state effect, 71, 81

recovery, 15

spectral diffusion, 109, 192

syndrome detection, 15, 26

spectral hole burning, 47

without measurement, 31

spherical tensor operators, 61, 115

quantum Fourier transform (qft), 8

spin boson model, 90

quantum Hall effect, 4, 51

spin diffusion, 70

quantum repeater, 35

spin echoes, 70, 119

quantum simulation, 6, 12

spin locking, 144

quantum teleportation, 35

spin temperature, 70

qubits, 4, 22

spin waves, 71, 88


spin-boson model, 160

Rabi frequency, 59
radio frequency (rf) Hamiltonian, 60
realism, 2
recoupling, 94, 126
Riven-Shamir-Adleman (rsa), 11
rotating frame, 59

stimulated Raman transitions, 41, 45


stroboscopic observation, 111
Suhl-Nakamura interaction, 72, 84, 88
superconducting quantum interference device
(squid), 50, 82
superconductors, 4, 34, 49, 49

rotating wave approximation (rwa), 59

charge qubit, 50

Ruderman-Kittel-Kasuya-Yosida (rkky) inter-

flux qubit, 50

action, 84, 85
scalability, 22
Schr
odingers cat, 3, 23
Schr
odinger, E., 3
second averaging, 115
second moment, 63, 168

phase qubit, 50
tank circuit, 134
Toffoli gate, 31
toggling frame, 111, 127, 190
trapped ions, 17, 41, 56
Turing, A., 5

selective pulse, see soft pulse


Shors algorithm, 8

universal quantum logic, 22

silicon, 71, 91, 96, 109, 133, 177

universal Turing machine, 5

232

Van Vleck formula, 64


Verschr
ankung, see entanglement
wahuha, 111
Walsh-Hadamard Gate, see Hadamard gate
wave function, 1
Zeeman Hamiltonian, 57

Index

You might also like