Sparrow C 2018 PHD Thesis

Quantum Interference in Universal
Linear Optical Devices for Quantum

Computation and Simulation
Christopher Sparrow
Department of Physics
Imperial College London
A thesis submitted in accordance with the requirements for the degree of

Doctor of Philosophy
September 2017
Abstract
It is believed that the exotic properties of quantum systems can be harnessed to perform
certain computational tasks more efficiently than classical theories allow. The production,
manipulation and detection of single photons constitutes a potential platform for perform-
ing such non-classical information processing. The development of integrated quantum
photonics has provided a miniaturised, monolithic architecture that is promising for the
realisation of near-term analog quantum devices as well as full-scale universal quantum
computers.
In this thesis we investigate the viability of these photonic quantum computational
approaches from an experimental and theoretical perspective. We implement the first
universally reconfigurable linear optical network; a key capability for the rapid prototyping
of photonic quantum protocols. We propose and demonstrate the use of these devices
as a new platform for the programmable quantum simulation of molecular vibrational
dynamics.
We then tackle an important outstanding problem in linear optical quantum com-
puting; quantifying how partial-distinguishability amongst photons affects logical error
rates. Finally, we propose a series of schemes aimed at counteracting these distinguisha-
bility errors in order to achieve practical quantum technologies with imperfect photonic
components.
2
Acknowledgements
I would first like to thank my supervisor Anthony Laing for advice, guidance and scien-
tific insight throughout my PhD. I would also like to thank Jeremy O’Brien and Terry
Rudolph for the inspiration to explore the many beautiful and frustrating quirks of pho-
tonic quantum computing.
I am hugely grateful to all those with whom the results in this thesis were obtained.
Whether in the lab or at the whiteboard; Enrique Martı́n-López, Nick Russell, Jacques
Carolan, Chris Harrold, Patrick Birchall, Hugo Cable, Alex Neville and Nicola Maraviglia,
I learned a great deal from all of you.
There are too many people to acknowledge at the Centre for Quantum Photonics (and
latterly QETlabs), it provided a wonderful environment for working, both intellectually
and socially. Similarly, my journey in the world of quantum information science so far
would not have been the same without the learning and laughter shared with Cohort 4 of
the Controlled Quantum Dynamics CDT at Imperial College London.
Of course, the real thesis was the friends I made along the way... I will always look
back with great happiness at my years in Bristol and this is in large part due to the many
close friends and acquaintances with whom many chats, dinners and beers were shared;
Alex, Allison, Beccie, Callum, Chris, Hugo, Javier, Lorraine, Nicki, Patrick, Phil, Raf and
Will to name but a few.
I would like to thank my whole family for a lifetime of support and providing me with
the opportunities and encouragement that have led me here.
Finally, this thesis is dedicated to Kirsten, for her love and for her infinite patience
putting up with me while it was undertaken and written up.
4
Declaration of Originality
I declare that the work in this thesis was carried out in accordance with the requirements
of the University’s Academic Regulations for Research Degree Programmes. The thesis
was written entirely by myself and accurately reflects work carried out during my graduate
studies. Work done in collaboration with, or with the assistance of, others, is indicated
as such.
The copyright of this thesis rests with the author and is made available under a
Creative Commons Attribution Non-Commercial No Derivatives licence. Researchers are
free to copy, distribute or transmit the thesis on the condition that they attribute it, that
they do not use it for commercial purposes and that they do not alter, transform or build
upon it. For any reuse or redistribution, researchers must make clear to others the licence
terms of this work
5
Publications
1. J. Carolan, C. Harrold, C. Sparrow, E. Martı́n-López, N. J. Russell, J. W. Sil-

verstone, P. J. Shadbolt, N. Matsuda, M. Oguma, M. Itoh, G. D. Marshall, M.
G. Thompson, J. C. F. Matthews, T. Hashimoto, J. L. O’Brien, and A. Laing,
‘Universal Linear Optics’, Science, 349, 711-716, (2015).
2. A. Holleczek, O. Barter, A. Rubenok, J. Dilley, P. B.R. Nisbet-Jones, G. Langfahl-

Klabes, G. D. Marshall, C. Sparrow, J. L. O’Brien, K. Poulios, A. Kuhn, and J.
C.F. Matthews, ‘Quantum Logic with Cavity Photons From Single Atoms’, Phys.
Rev. Lett., 117, 023602, (2016).
3. A. Neville, C. Sparrow, R. Clifford, E. Johnston, P. Birchall, A. Montanaro, A.

Laing, ‘Classical Boson Sampling Algorithms with Superior Performance to Near-
term Experiments’, Nat. Phys., 13, 1153-1157 (2017).
4. C. Sparrow, E. Martı́n-López, N. Maraviglia, A. Neville, J. Carolan, C. Harrold, N.

Matsuda, J. L. O’Brien, Y. Joglekar, D. Tew, A. Laing, ‘Simulating the Vibrational
Quantum Dynamics of Molecules with Integrated Photonics’, Submitted, (2017).
5. C. Sparrow*, P. Birchall* et al., ‘Linear Optical Quantum Computing with Partially-

Distinguishable Photons’, in prep, (2017).
* authors contributed equally to this work
6
Contents
Abstract 2
Acknowledgements 4
Declaration of Originality 5
Publications 6
1 Introduction 16
1.1 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Quantum Information and Computation 20

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2 Quantum Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.1 Physical states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.3 Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.4 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.5 Interpretations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.6 Operational quantum theory . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Quantum Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.1 The Qubit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.2 Entanglement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.3 Nonlocality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
8
CONTENTS
2.3.4 Quantum tomography . . . . . . . . . . . . . . . . . . . . . . . . . 34

2.3.5 Distance measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Quantum computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.1 Turing machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.2 Computational Complexity Theory . . . . . . . . . . . . . . . . . . 41
2.4.3 Circuit model quantum computation . . . . . . . . . . . . . . . . . 43
2.4.4 Quantum error correction . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.5 Fault tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.4.6 Stabilizer formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.7 Measurement-based quantum computation . . . . . . . . . . . . . . 53
2.4.8 Other computational models . . . . . . . . . . . . . . . . . . . . . . 55
2.4.9 Physical platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3 Quantum light 62
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 Quantisation of the electromagnetic field . . . . . . . . . . . . . . . . . . . 62
3.3 Linear optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.4 Quantum Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4.1 The Hong-Ou-Mandel effect . . . . . . . . . . . . . . . . . . . . . . 69
3.4.2 The HOM dip and partial-distinguishability . . . . . . . . . . . . . 71
3.5 Coherent states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.6 Linear optical quantum computing . . . . . . . . . . . . . . . . . . . . . . 75
3.6.1 KLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.6.2 MBLOQC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.6.3 Boson sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.7 Quantum photonics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.7.1 Photon sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.7.2 Linear optical networks . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.7.3 Detectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.7.4 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
9
CONTENTS
4 Universal Linear Optics 100

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.2 The Reck et al. Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.3 A Universal Linear Optical Processor . . . . . . . . . . . . . . . . . . . . . 103
4.3.1 Characterisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4 Linear Optical Gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.4.1 Process tomography experiments . . . . . . . . . . . . . . . . . . . 106
4.4.2 Heralded integrated gates . . . . . . . . . . . . . . . . . . . . . . . 111
4.4.3 Characterisation of Linear Optical Networks . . . . . . . . . . . . . 117
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5 Simulating Vibrational Quantum Dynamics of Molecules with Integrated

Photonics 122
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2 Molecular Vibrational Dynamics in the Harmonic Approximation . . . . . 124
5.3 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4 Simulating Four Atom Molecules . . . . . . . . . . . . . . . . . . . . . . . 127
5.5 Energy transfer and dephasing in NMA . . . . . . . . . . . . . . . . . . . . 131
5.6 Vibrational relaxation in H2 O . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.7 Anharmonic Hamiltonian for H2 O . . . . . . . . . . . . . . . . . . . . . . . 137
5.8 Adaptive feedback control in the dissociation of NH3 . . . . . . . . . . . . 140
5.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6 Linear Optical Quantum Computing with Partially-Distinguishable Pho-

tons 146
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.2 Errors in LOQC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.3 Photon Distinguishability . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.3.1 Dual-path states . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.2 Distinguishability models . . . . . . . . . . . . . . . . . . . . . . . . 151
10
CONTENTS
6.4 Generating entanglement with

partially-distinguishable photons . . . . . . . . . . . . . . . . . . . . . . . . 152
6.5 Entangled measurements with
partially-distinguishable photons . . . . . . . . . . . . . . . . . . . . . . . . 159
6.6 Measurement-based resource generation . . . . . . . . . . . . . . . . . . . . 160
6.6.1 Overview and initial state . . . . . . . . . . . . . . . . . . . . . . . 160
6.6.2 GHZ generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.6.3 Fusion gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.7 Error rates on Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7 Mitigating Distinguishability Errors 174

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.2 Hong-Ou-Mandel Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.3 The Distinguishability Tolerance of Bell State Generators . . . . . . . . . . 179
7.4 Generating entanglement from distinguishable photons . . . . . . . . . . . 181
7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8 Conclusion 186
8.1 Key Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
8.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Bibliography 190
A Universal Linear Optics 210

A.1 Transfer Matrix Reconstructions . . . . . . . . . . . . . . . . . . . . . . . . 210
B Simulating the Vibrational Quantum Dynamics of Molecules with Inte-

grated Photonics 212
B.1 Molecule details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
B.1.1 HFHF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
B.1.2 HNCO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11
CONTENTS
B.1.3 N4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
B.1.4 P4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
B.1.5 SO3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
B.1.6 NMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
B.1.7 H2 O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
B.1.8 NH3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
B.2 Additional data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
C Linear Optical Quantum Computing with Partially-Distinguishable Pho-

tons 223
C.1 Trace inequality proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
C.2 Bound on real term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
C.3 Additional transfer matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 226
C.4 Equivalence of pre- and post-fusion corrections . . . . . . . . . . . . . . . . 227
C.5 Fusion Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
C.6 Additional measurement operators . . . . . . . . . . . . . . . . . . . . . . 231
C.7 Second order fusion error maps . . . . . . . . . . . . . . . . . . . . . . . . 235
D Mitigating Distinguishability Errors 238
12
List of Figures
2.1 Bloch sphere representation of pure qubit state . . . . . . . . . . . . . . . 28

2.2 Schematic of CHSH Bell test . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Hierarchy of computational complexity classes . . . . . . . . . . . . . . . . 43
2.4 Circuit model quantum computation . . . . . . . . . . . . . . . . . . . . . 45
2.5 Measurement based quantum computation . . . . . . . . . . . . . . . . . . 54
3.1 Hong-Ou-Mandel dip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.2 Single and dual-rail photonic qubits . . . . . . . . . . . . . . . . . . . . . . 76
3.3 Elementary circuits in the KLM scheme . . . . . . . . . . . . . . . . . . . . 79
3.4 Type I and II Fusion gates . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.5 Schematic of ballistic photonic quantum computation . . . . . . . . . . . . 82
3.6 Boson Sampling schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.7 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.1 The Reck et al. scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.2 Characterisation of phase shifters . . . . . . . . . . . . . . . . . . . . . . . 106
4.3 Process tomography for single qubit gates . . . . . . . . . . . . . . . . . . 108
4.4 Process tomography of postselected CNOT gate . . . . . . . . . . . . . . . 111
4.5 KLM CNOT gate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.6 Experimental Bell state generation . . . . . . . . . . . . . . . . . . . . . . 115
5.1 Simulating the vibrational dynamics of molecules in the harmonic approx-

imation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.2 Evolution of localised vibrational excitations in four atom molecules . . . . 130
13
LIST OF FIGURES
5.3 Quantum energy transfer and dephasing in NMA . . . . . . . . . . . . . . 132

5.4 Vibrational relaxation in H2 O . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.5 Anharmonic model for H2 O . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.6 Optimised success probability for nonlinear phase shift (NPS) gates . . . . 139
5.7 Adaptive feedback control algorithm for selective dissociation in NH3 . . . 142
6.1 Linear optical networks for Bell state preparation and measurement . . . . 154
6.2 Purity and Fidelity of logical Bell states produced with partially-distinguishable
photons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.3 Schematic of an all-optical implementation of ballistic LOQC . . . . . . . . 162
6.4 Summary of LOQC with a distinguishable particle . . . . . . . . . . . . . . 168
6.5 Output state ρG produced in lattice by one distinguishable particle. . . . . 169
6.6 Logical error rates on lattice . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7.1 Hong-Ou-Mandel Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

7.2 Improvement in dominant eigenvalue for HOM Filter . . . . . . . . . . . . 178
7.3 Linear optical networks for heralded bell generation . . . . . . . . . . . . . 179
7.4 Error tolerance of Bell generators . . . . . . . . . . . . . . . . . . . . . . . 180
7.5 Entanglement generation from spectrally distinguishable photons via tem-
poral resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
B.1 Four atom molecule data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

B.2 H2 CS data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
B.3 NMA data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
B.4 H2 O data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
B.5 Cost function and probability of bunching in AFC experiment . . . . . . . 221
B.6 Full distributions for initial and final AFC states . . . . . . . . . . . . . . . 222
C.1 Fusion operations including local unitaries . . . . . . . . . . . . . . . . . . 229

C.2 Error maps for second order errors in fusion measurements . . . . . . . . . 237
14
Chapter 1
Introduction
Classical theories of physics provide sensible and convincing explanations for the world
that humans can observe and interact with. Unfortunately, these explanations fall apart
at the scale of the smallest physical systems. The discovery and development of quantum
theory overturned many of the foundational assumptions underpinning our previous un-
derstanding of nature, resulting in decades of debate amongst scientists and philosophers
alike. However, despite widely spread opinions on interpretational aspects of the theory,
quantum mechanics as a tool for predicting the outcomes of experimental observations
remains unparalleled in scientific history.
An investigation of the theory purely in terms of the probabilities of observing events,
an information-theoretic perspective, has revealed that the unique phenomena that quan-
tum systems exhibit can allow for fundamentally new ways of processing information.
Perhaps the most striking encapsulation of this concept is the quantum computer, a
computer built from quantum mechanical components which can perform certain com-
putational tasks exponentially faster than devices operating under the laws of classical
physics. The global race to realise the first universal quantum computer represents both
a grand engineering challenge with huge potential to solve real world problems, but also
a scientific landmark akin to the creation of a new state of matter.
Light has played a key role in the theoretical and experimental development of quan-
tum information science, from the photoelectric effect [1] through to the first loophole-free
16
1. Introduction
Bell tests [2–4]. The ease of manipulation and low noise properties of quantum states of
light have meant that optical systems have often been ideal systems to investigate funda-
mental quantum physics. Recently however, with the development of integrated quantum
photonics [5–7], devices which produce, manipulate and detect photons on chips have
become a promising platform for future quantum technologies.
The quest to deliver quantum technologies can be tackled via a top-down or bottom-up
approach. On the one hand we can produce blueprints for high-level architectures realising
universal, fault-tolerant optical quantum computation and investigate the experimental
resources and requirements that would be needed to achieve such schemes. From the other
side, we can consider the toolbox of the current state-of-art and near-future experimental
capabilities and investigate specific information processing tasks which can be targeted
for near-term impacts.
In this thesis we will explore both of these approaches towards realising practical
quantum computational devices. Experimentally, we demonstrate the first universal lin-
ear optical device; an important milestone on the road towards non-universal models of
photonic quantum computation. We first illustrate the versatility of this platform by
using the single device as an experimental testbed for implementing and characterising a
series of quantum logic gates, many for the first time in integrated optics. Although it
is known that schematically simple linear optical experiments can solve classically hard
computational problems [8], real world applications for such devices are currently lim-
ited. Here we show that universal linear optical devices can operate as analog quantum
simulators for the dynamics of quantised molecular vibrations. We present experimental
simulations with tens of thousands of device configurations, exploring the dynamics of a
range of molecules and molecular behaviours including their dephasing, thermalisation,
energy transfer and dissociation.
Moving to large-scale universal architectures, a key missing ingredient in our under-
standing of the practicalities of linear optical quantum computing is the existence of good
error models. The link between physical imperfections in realistic devices and errors in
the resulting quantum computation is crucial to assessing the fabrication tolerances and
error correction overheads necessary to achieve fault-tolerance. A major source of logical
17
1. Introduction
error in photonics is the production of photons which are not perfectly indistinguishable
in all their degrees of freedom. Here we develop a model for this error and explore its
resulting effects on leading schemes for universal linear optical quantum computing. The
insights gained from this study allow us to then propose a series of new techniques de-
signed to suppress or bypass these errors without requiring detailed knowledge of the
photonic states.
1.1 Thesis Outline

This thesis is structured as follows:
Chapters 2 and 3 provide the background material. Chapter 2 includes a brief overview
and summary of quantum theory, quantum information and computing. Chapter 3 then
introduces the quantum description of light, photonic quantum computation and finally
the experimental photonic platform we use in the remainder of the thesis.
Chapter 4 describes the demonstration of a universal linear optical processor acting
upon six optical modes. This device is used to implement a series of optical quantum logic
gates including the first demonstrations of a heralded CNOT and a newly-derived Bell
state generator in integrated optics as well as the highest process fidelity recorded for a
postselected CNOT gate so far. The reconstruction of the transfer matrices implementing
these gates demonstrates that the logical error observed in multi-photon experiments is
largely due to imperfect sources and detection.
Chapter 5 proposes and demonstrates universal linear optical devices as a new quan-
tum simulator for molecular vibrational dynamics. The simulation of localised vibrational
excitations in a range of molecular systems is presented including investigations of energy
transfer, open system dynamics and anharmonic corrections via measurement-induced
nonlinearities. We go on to operate the experiment in a closed feedback loop to simulate
a quantum control experiment aimed at maximising the photo-dissociation of ammonia.
Chapter 6 investigates the effects of using partially-distinguishable photons in linear
optical quantum computers. After presenting a new formalism for calculating the logical
information of photonic states containing partially-distinguishable photons, this is then
18
1. Introduction
used to investigate standard linear optical gates and eventually the resulting logical error
rates in an architecture for universal computation are computed for the first time.
Chapter 7 seeks to alleviate the difficulty of producing highly indistinguishable photons
for quantum technologies by proposing a series of schemes which can result in improved
error rates without any direct manipulation of the additional degrees of freedom of the
photons. These include a filter based upon Hong-Ou-Mandel interference, new error-
tolerant gates and the use of high resolution detectors to erase distinguishing information.
Finally, Chapter 8 summarises the work and provides concluding remarks.
19
Chapter 2
Quantum Information and

Computation
2.1 Introduction
In this chapter we introduce the physical and mathematical framework which will be
used in the remainder of the thesis and provide the context in which the work is placed.
We begin by introducing the standard postulates of quantum theory. We then introduce
quantum information theory and quantum computation. Various parts of this section
draw upon notes and presentations from Michael Nielsen & Isaac Chuang, John Preskill,
Robert Spekkens, Robin Blume-Kohout, David Jennings and Dan Browne.
2.2 Quantum Theory

By the early 20th century it was becoming increasingly clear that existing theories of
physics could not explain many observed phenomena such as black body radiation and
the photoelectric effect. Throughout the next century a new theory of physics was de-
veloped by Bohr, Einstein, Heisenberg, Schrödinger and many others to explain these
experiments built upon the postulate that physical systems and properties can be quan-
tised, that is fundamentally discrete. This quantum theory required radical departures
20
2. Quantum Information and Computation
from common-sense assumptions about nature such as the existence of well-defined phys-
ical properties prior to and independent of measurement. Furthermore, it introduced new
physical phenomena such as entanglement and nonlocality, which had no analogue in the
previous classical descriptions. We begin by enumerating the standard postulates of the
theory of Quantum Mechanics followed by a brief discussion of the continuing difficulties
in its interpretation.
2.2.1 Physical states
A complete description of a physical state in quantum mechanics is given by a ray in a

Hilbert space, H. A finite-dimensional Hilbert space is simply a vector space over the real
or complex numbers with an inner product 1 . They are conventionally represented by a
unit vector |ψi which uniquely defines the state up to a global phase. Systems compose
by tensor product, so the composite state of two systems, A and B, is a ray in the space
HAB ≡ HA ⊗ HB .
2.2.2 Observables
Physically measurable properties, known as observables, are described by Hermitian op-

erators, A = A† . They admit a spectral decomposition
X
A= ak Πk (2.1)
k
P
where ak are the eigenvalues of A and Πk = k |φk ihφk | are projectors onto the space
spanned by the vectors |φk i, the eigenstates of A with corresponding eigenvalue ak ,
A|φk i = ak |φk i (2.2)

1
For infinite-dimensional spaces they must have the additional property of completeness, which loosely
means there are no ‘gaps’ in the space.
21
2.2.3 Evolution
Evolution in quantum mechanics is described by a unitary operator, U −1 = U † . In

the Schrödinger picture, states are time-dependent and the evolution is described by
|ψ(t)i = U |ψ(0)i. Infinitesimal evolution is via the Schrödinger equation
∂|ψ(t)i
i~ = Ĥ|ψ(t)i (2.3)
∂t
where Ĥ is the Hamiltonian, the observable corresponding to a system’s energy. When

the Hamiltonian is time-independent U = e−iĤt/~ .
Evolution in quantum mechanics can be equivalently expressed in the Heisenberg pic-
ture, where the time dependence is instead put into the observables, A(t) = U † A(0)U . In
this case infinitesimal evolution is via
dA(t)
i~ = [A(t), Ĥ] (2.4)
dt
assuming no additional time dependence of A.
2.2.4 Measurements
The measurement of a physical observable is a process in which a classical observer learns

information about a quantum system. The outcome of a measurement of observable A
is one of its eigenvalues {ak }. The probability of measuring an outcome ak when the
quantum state prior to measurement is |ψi, is given by the Born rule
Pr(k|ψ) = hψ|Πk |ψi (2.5)
The measurement process changes the state, preparing it in an eigenstate (or eigenspace
if the eigenvalue is degenerate) corresponding to the outcome via the update map
Πk |ψi
|ψi → p . (2.6)
hψ|Πk |ψi
22
2.2.5 Interpretations
These postulates present some problematic aspects. There appears to be friction between
the evolution and measurement postulates. Quantum systems evolve deterministically
and continuously via U , until they are measured by a classical observer when they evolve
probabilistically and discontinuously. The problem is that we are not told what should be
treated as a classical observer and what should be treated as a quantum system, i.e. where
should we make the cut. If a measurement device is treated quantum mechanically, the
system and measurement device will evolve deterministically to a joint state. In contrast,
if we treat the measurement device classically, we ‘collapse’ the system state and receive
a probabilistic outcome. This is usually referred to as the measurement problem and
remains a subject of intense debate and disagreement to this day [9–11].
There have been a number of approaches to ‘fix’ the measurement problem, though
none have garnered universal acceptance, and all seem unsatisfactory to some degree.
One can reject the measurement postulate and posit that all evolution is described by
the evolution postulate, e.g. the many worlds interpretation [12]. The cost is of course
that one must then accept the existence of a perhaps infinite number of unobservable
universes. In addition, there are difficulties in making sense of the role of probabilties and
Born’s rule if evolution is entirely deterministic.
Conversely, one can reject the evolution postulate, for instance collapse models where
systems undergo spontaneous collapse on some characteristic timescale, though in this
case one must accept that energy is no longer completely conserved [13].
Alternatively, one can reject the state postulate, i.e. that |ψi contains a complete
description of the physical state. One such way to do so is to use extra degrees of freedom
which describe the ‘real’, or ontological, state. One example is Bohmian mechanics [14],
where the quantum state is supplemented with the true positions of all particles, the
wavefunction then determines the evolution of these positions in time. The difficulties
of this approach include the special status of position over all other observables, and its
inherent nonlocality.
Finally, one can reject some other aspect of the formalism, for instance by replacing
23
the rules of classical logic [15] or rejecting the notion of an objective reality [16].
The remarkable irony of quantum theory is that although there remains little agree-
ment about how it should be understood, it is arguably the most successfully tested
physical theory ever devised (e.g. it predicts the fine structure constant to at least within
half a part per billion [17]). In the following, we will not concern ourselves with any par-
ticular interpretation and instead adopt the “Shut up and calculate!” [18] approach; using
the well understood mathematical rules of the theory to translate and predict the results
of experiments consisting of state preparations, transformations and measurements on
quantum systems. In order to so we will require a slightly more general formalism, which
can additionally describe classical uncertainties and the study of quantum subsystems.
2.2.6 Operational quantum theory
In this section we briefly describe the most general state preparations, measurements and
evolution allowed in quantum theory.
States
The most general form of quantum state preparation is an ensemble {pi , |ψi i}, where pi
is the probability of preparing the pure state |ψi i. This ensemble defines a density matrix
P
ρ = i pi |ψi ihψi |. If rank(ρ) > 1 it is referred to as a mixed state. A general ρ has the
properties
Hermiticity: ρ = ρ†
Trace one: Tr[ρ] = 1
Positivity: hψ|ρ|ψi ≥ 0 ∀ |ψi ∈ H (2.7)
where the positivity constraint avoids the possibility of negative probabilities for measure-
ment outcomes and is usually denoted as ρ ≥ 0. As well as describing a classical proba-
bilistic mixture of pure states, ρ can also describe the reduced state of a subsystem of a
composite system. For instance, the state of system A in a joint state |ψiAB ∈ HA ⊗ HB
24
can be described by the partial trace over the space HB ,

TrB |ψihψ| = ρA . (2.8)
Conversely, any ρ can be described as the reduced state of a pure state (known as a
purification of ρ) in a larger space.
Measurements
The most general form of quantum measurement is a positive operator valued measure
(POVM). This is a set of operators {E} known as the POVM effects which satisfy the
conditions
Positivity: Ek ≥ 0 ∀k
X
Sum to identity: Ek = 1 (2.9)
k
The Born rule for these generalised measurements and states becomes
Pr(k|ρ) = Tr[Ek ρ] (2.10)
and the state update rule for a measurement with outcome k becomes
Mk ρM†k
ρ→ (2.11)
Tr[ρEk ]
where Ek = M†k Mk . Any POVM can be described by coupling the measured system to an
ancilla system prepared in a fixed state and performing a regular projective measurement
upon the composite system, a procedure known as a Naimark extension.
Evolution
The most general form of quantum evolution is via completely-positive trace-preserving

(CPTP) maps , sometimes known as quantum channels. A CPTP map is a superoperator,
E, which maps density matrices to density matrices, E(ρ) = ρ0 . They must satisfy the
25
following conditions
X X
Linearity: E pk ρ k = pk E(ρk )
k k
Trace-preserving: Tr[E(ρ)] = 1
Complete positivity: [E ⊗ 1](ρ) ≥ 0 ∀ρ ∈ HA ⊗ HB (2.12)
Where the additional requirement of complete positivity is required to ensure that the
output state is valid when the channel is applied to a subsystem of a larger system. All
maps satisfying these conditions can be written in Kraus decomposition as
X
E(ρ) = Ak ρA†k (2.13)
k
A†k Ak = 1. {Ak } are known as the Kraus operators of E, and they are defined
P
where k
up to a global unitary transformation. Any quantum channel can be described by coupling

the system to an ancilla system in a pure state, unitarily evolving the composite system
and throwing away the ancilla.
To describe infinitesimal evolution for more general non-unitary evolutions that are
nevertheless Markovian (ρ(t + dt) can be determined entirely by ρ(t)), the Schrödinger
equation can be generalised to the Lindblad master equation
X 1
Li ρL†i − {L†i Li , ρ}

ρ̇ = −i[Ĥ, ρ] + (2.14)
i
2
where {Li } are known as the Lindblad operators and Ĥ is the natural Hamiltonian of the
system.
2.3 Quantum Information

In 1948, Claude Shannon at Bell Labs published a seminal paper that established the field
of information theory [19]. This work explained how to formalise and quantify the concept
of information and its communication via the use of probability theory and entropy. A
26
‘message’ can be considered as a discrete random variable X which can take one of n
symbols {x1 , x2 , ..., xn }. The information content of the message is then determined by
the probability mass function {Pr(xi )} over these symbols and quantified by the Shannon
entropy
X
H(X) = − Pr(xi ) log2 Pr(xi ) (2.15)
i
Shannon defined the minimal unit of information as a bit, the amount of information
obtained from a message which has an equal probability of being in one of two symbols, or
states, conventionally labelled 0 and 1. Just over half a century later, these contributions
have dramatically shaped the modern information age in which we live.
A profound insight, which has become increasingly popular among scientists in the
last few decades, is that the concept of information can also be fundamental to our
understanding of physics. This position is encapsulated by John Archibald Wheeler’s
doctrine ‘it from bit’ 2 . Quantum mechanics naturally lends itself an information-theoretic
description since, when considered entirely abstractly, it can be thought of as in some
sense a generalised probability theory. The development of quantum information theory
has lead to a much richer understanding of the theory itself, allowing for the differences
between classical and quantum theories to be elucidated and quantified.
2.3.1 The Qubit
As the minimal unit of classical information is the bit, the minimal unit of quantum
information is the qubit, a quantum system that lives in a two-dimensional Hilbert space.
Conventionally, the state of a qubit is often described in the computational basis:
   
1 0
|0i ≡   ; |1i ≡   (2.16)
0 1
2
“It from bit. Otherwise put, every it - every particle, every field of force, even the space-time
continuum itself - derives its function, its meaning, its very existence entirely - even if in some contexts
indirectly - from the apparatus-elicited answers to yes-or-no questions, binary choices, bits”
27
|0i
| i
|+ii
|+i
|1i
Figure 2.1: Bloch sphere representation of pure qubit state. An arbitrary single qubit pure
state |ψi can be represented by a point on the surface of the unit sphere parameterised
by the angles θ and φ.
Since an arbitrary pure qubit state can be parameterised as
θ θ
|ψi = sin |0i + eiφ cos |1i (2.17)
2 2
they have an intuitive geometrical representation as points on the surface of the three-
dimensional unit sphere. Of particular use in what follows will be the eigenstates of the
Pauli matrices {σx , σy , σz } (hereafter simply {X, Y, Z})
Z : {|0i, |1i} (2.18)

1 1
X : {|+i = √ (|0i + |1i), |−i = √ (|0i − |1i)} (2.19)
2 2
1 1
Y : {|+ii = √ (|0i + i|1i), |−ii = √ (|0i − i|1i} (2.20)
2 2
which form the vertices of an octahedron centred on the origin.

Other important aspects of quantum information can be visualised using the Bloch
sphere. Points within the Bloch sphere correspond to mixed states. We can therefore
28
represent an arbitrary qubit state by a vector ~r = (rx , ry , rz ) where
1
ρ = (1 + rx X + ry Y + rz Z). (2.21)
2
p 2
The pure states correspond to unit vectors |~r| = rx + ry2 + rz2 = 1, and the length of
the vector describes how mixed the state is. This mixedness is often characterised by the
purity, P = (1 + |~r|2 )/2 of the state. More generally,
P(ρ) = Tr(ρ2 ). (2.22)
Another important property that is clear from considering the Bloch sphere is that a
general decomposition of a mixed state as a convex combination of pure states is not
unique. For instance, the maximally mixed state, ρ = 1/2, can be decomposed as
1 1
ρ = (|0ih0| + |1ih1|) = (|+ih+| + |−ih−|) (2.23)
2 2
and, indeed, in an infinite number of ways. However, for all non-maximally mixed states, ρ
has a unique spectral decomposition where the constituent pure states are the orthogonal
eigenstates of ρ
X 1 + |~r| 1 − |~r|
ρ= λi |ψi ihψi | = |ψihψ| + |ψ⊥ ihψ⊥ |. (2.24)
i
2 2
Finally, the concept of information as entropy can be carried over to the quantum realm.
The von Neumann entropy is defined as
X
S(ρ) = −Tr(ρ log2 ρ) = − λi log2 (λi ). (2.25)
i
For all pure states S(|ψihψ|) = 0 which can be understood as the fact that there exists
a measurement with a deterministic outcome (the measurement {|ψihψ|, 1 − |ψihψ|}),
whereas S(1/2) = 1 as the result of any measurement is completely random, i.e. it
is equivalent to a fair coin flip and therefore a single bit. The Von Neumann entropy
29
replaces Shannon entropy in quantum information theory and also provides a way to
quantify quantum entanglement, as we will see below.
2.3.2 Entanglement
Although much of the quantum information framework can be understood by the single
qubit example, a key feature of quantum theory that often defies classical description
is entanglement [20] between multiple quantum systems. A pure entangled state is one
which cannot be written as a product state, a tensor product of pure states on each of the
subsystems, e.g.
|ψiAB = |ΨiA ⊗ |ΦiB . (2.26)
The entangled states, which cannot be written in this way, therefore cannot be understood
in terms of independent, local states of each subsystem and are instead only fully described
by a correlated state of the whole system.
The set of entangled states can be defined as those which are not separable. Separable
states are those which can be written as
X
ρsep = pi |Ψi ihΨi | ⊗ |Φi ihΦi |. (2.27)
i
Examples of pure entangled states for two qubits that will be used throughout this thesis
are the Bell states. These are
1 1
|φ± i = √ (|00i ± |11i), |ψ ± i = √ (|01i ± |10i) (2.28)
2 2
and they form a basis for two qubit states.

An alternative way to understand entanglement is to define it as that which cannot
be created using only LOCC: local quantum operations on each subsystem and classical
communication. This definition helps quantify how much entanglement there is in a
state. Although there are many ways to quantify entanglement (see, for example, Ref.
[21]), an appealing general property of an entanglement measure is LOCC-monotonicity,
30
that is functions which do not increase under LOCC. For pure, bipartite systems there
is a natural entanglement measure which most such monotones reduce to known as the
entropy of entanglement
X
E[|ψiAB ] = S(TrB [|ψihψ|]) = −Tr[ρA log ρA ] = − λA A
i log λi (2.29)
i
where S is the von Neumann entropy introduced previously. A quick check verifies that
product states, with pure reduced states will have no entropy of entanglement (e.g.
E[|00i] = 0) whereas maximally entangled states whose reduced states are maximally
mixed maximise E (e.g. E[|φ+ i] = 1).
Unfortunately when considering mixed state entanglement, things get considerably
messier. Consider, for instance, the mixed state formed by taking an equal mixture of
Bell states
1
ρ = (|φ+ ihφ+ | + |φ− ihφ− |) (2.30)
2
this state has an equivalent decomposition as
1
ρ = (|00ih00| + |11ih11|) (2.31)
2
and is therefore a separable, classically correlated state even though it is a mixture of

maximally entangled states. The entropy of entanglement can be generalised to mixed
states as the entanglement of formation
X
EF (ρ) = min pi E[|ψi i] (2.32)
{pi ,ψi }
i
where the minimum is taken over all pure state decompositions of ρ. The entanglement
of formation remains an LOCC-monotone but the form of the minimisation makes its
calculation extremely difficult in general. However, for the simple case of two-qubit states
there is a closed form solution based on another entanglement measure, the concurrence
C(ρ) = max[0, λ1 − λ2 − λ3 − λ4 ] (2.33)
31
where λi are the ordered square roots of the eigenvalues of the matrix ρ(Y ⊗ Y )ρ∗ (Y ⊗ Y ).
This can then be related to EF as
1 + p1 − C 2 (ρ)
EF (ρ) = h (2.34)
2
where
h(x) = −x log2 x − (1 − x) log2 (1 − x). (2.35)
For pure states |ψi = a|00i + b|01i + c|10i + d|11i, the concurrence takes the particularly
simple form
C(|ψi) = 2|ad − bc|. (2.36)
2.3.3 Nonlocality
Entanglement was at the core of one of the greatest foundational developments in modern
physics, the discovery of Bell nonlocality [22]. In their famous paper of 1935 [23], Ein-
stein, Podolsky and Rosen (EPR) argued 3 that quantum theory could not be a complete
description of reality. At the heart of their argument was the implicit assumption of lo-
cal realism; that systems have well-defined states of reality which can only be influenced
by their immediate surroundings. If two systems are in an entangled state |ψiAB , then
the choice of measurement performed upon A can affect the post-measurement state of
system B, |φiB . But if |φiB represents the complete reality of system B, then it would
appear that what happens at A can instantaneously affect the reality of B, even when the
systems are arbitrarily separated. Since such influences violate locality, the conclusion of
EPR was that the quantum state was not a complete description of the true underlying
reality of a system, and when additional ‘hidden variables’ are included, locality can be
restored.
In 1964 John Bell proved the remarkable result that even if the quantum state is an
incomplete description of reality, we still cannot describe the outcomes of all quantum
3
It has been claimed that Einstein did not in fact read the paper prior to publication and that in
correspondence with Schrödinger suggested that it had not completely captured his perspective. See [24]
for his own, simpler argument along the lines of what we present here.
32
x 2 {0, 1} y 2 {0, 1}
Alice Bob
a 2 {+1, 1} b 2 {+1, 1}
Figure 2.2: Schematic of CHSH Bell test. Two separated parties Alice and Bob, possibly
sharing prior correlations λ, are given inputs x and y and output a and b respectively.
The conditional probabilities Pr(ab|xy) allowed in any physical theory which obeys lo-
cal realism are constrained by inequalities. Quantum theory violates these inequalities,
proving that it cannot be described by any local realistic model.
experiments with a local model [25]. The simplest statement of Bell’s theorem can be
made by considering the following scenario depicted in Fig. 2.2: two parties, conventionally
named Alice and Bob are first separated so that they cannot communicate, they are then
each given an input in the form of a single bit, x, y ∈ {0, 1}, and told to produce an
P
output a, b ∈ {−1, +1}. After many trials, the expectations hax by i = a,b abPr(ab|xy)
can be estimated, and the quantity
S = ha0 b0 i + ha0 b1 i + ha1 b0 i − ha1 b1 i (2.37)
calculated. Since Alice and Bob are separated and have no prior knowledge of the inputs
they will be given, we can express the condition of locality via a factorising of the prob-
abilities of their outcomes, i.e. Alice’s output cannot depend upon Bob’s input. Alice
and Bob can share prior correlations however, including the possibility of any underlying
‘hidden variables’ or joint causal influences, and these can be included as an additional
variable λ Z
Pr(ab|xy) = dλq(λ)Pr(a|x, λ)Pr(b|y, λ) (2.38)
where in any given run, the value of λ is determined by some distribution q(λ). Now,
substituting Eq. (2.38) into Eq. (2.37) results in the bound
S≤2 (2.39)
33
known as the Clauser-Horne-Shimony-Holt (CHSH) inequality [26]. Note that so far we

have not made any reference to physics, the CHSH inequality is simply a general bound
upon all physical theories whose statistics satisfy Eq. (2.38).
Now, consider the following scenario in quantum theory: Alice and Bob each take one
qubit of the entangled Bell state |ψ − i = √1 (|01i − |10i)
2
and according to their input bit,
perform measurements of the following observables
√ √
x0 → Z , x1 → X , y0 → −(Z + X)/ 2 , y1 → (−Z + X)/ 2. (2.40)
√
The expectations in Eq. (2.37) in this case become ha0 b0 i = ha0 b1 i = ha1 b0 i = 1/ 2 and
√
ha1 b1 i = −1/ 2 resulting in
√
S = 2 2 > 2. (2.41)
This then proves that no local, realistic model can reproduce the correlations in this
experiment.
Since the first convincing experimental demonstration by Aspect et al. [27], Bell in-
equalities have been violated in many laboratory settings and recently three experiments
[2–4] were performed in a loophole-free manner, meaning that Alice and Bob were sepa-
rated so that their measurement settings were chosen outside of each others light cones
and the detection was efficient enough that no conspiratorial loss could have changed the
results.
Finally we note that entanglement and nonlocality, although related, are not equivalent
concepts and there exist subtleties in their relationship that are still not well understood.
For instance, there are known entangled states which still admit a local hidden variable
model [28] and, in some scenarios, the states which maximally violate Bell inequalities
are not the most entangled states [29].
2.3.4 Quantum tomography
Although any quantum system is described by its state ρ, learning the state of an unknown
system poses fundamentally new challenges.
34
Firstly, in contrast to classical theories, it is not always possible to simultaneously

measure a set of quantum observables. This is because two physical observables Â and B̂
will generally not commute [Â, B̂] = ÂB̂ − B̂ Â 6= 0, i.e. they will have different eigenstates
and so cannot be jointly measured whilst remaining consistent with the measurement
postulate. Furthermore, since measurement collapses the state, in order to infer the results
of incompatible measurements on a system, we require multiple copies of the initial state.
A second reason why multiple copies of a state are required is that our knowledge
is restricted by the inherently statistical nature of quantum measurements. Even if we
believe the state to be |ψi and perform the two-outcome measurement {|ψihψ|, 1−|ψihψ|},
a positive result is still consistent with any state not orthogonal to |ψi.
The task of quantum state tomography is to estimate an unknown state ρ by per-
forming measurements on a set of identically prepared copies of it4 . If the system has a
d-dimensional Hilbert space, then ρ is a d×d Hermitian matrix containing d2 independent
parameters. One of these parameters is fixed by the trace condition Tr(ρ) = 1, and so
d2 − 1 parameters must be estimated.
The simplest example of state tomography is for a single qubit. A natural choice of
measurements are the Pauli observables (plus identity) as they form an orthonormal basis
for 2 × 2 hermitian matrices. From Eq. (2.21) we see that this case is equivalent to
estimating the vector ~r where
rx = hXi = Tr(Xρ), ry = hY i = Tr(Y ρ), rz = hZi = Tr(Zρ). (2.42)
These probabilities are naturally estimated by sample expectation values. If we measure

Z on N copies of ρ, collapsing n0 copies onto |0ih0| and n1 to |1ih1|, then we can estimate
rz by
n0 − n1
= rzest ≈ hZi (2.43)
N
and the state can then be estimated directly via Eq. (2.21). A problem that is apparent
4
Of course in reality the copies %i will not be completely identical. However, as long as the choice
of measurement we perform on a P given copy is not correlated with that position then we can faithfully
N
reconstruct the average state ρ = i %i /N .
35
in the above is that we can only ever estimate hZi to some degree of precision due to
statistical fluctuations. These fluctuations can cause further problems, as if the state that
is being tomographed is approximately pure, the estimates we get {rxest , ryest , rzest } can result
in a reconstructed state which contains negative eigenvalues, i.e. is not a valid density
matrix. To overcome this problem there are several methods of statistical inference that
can be used [30] which restrict to only valid state estimates, the most commonly used
approach is maximum likelihood estimation [31].
The likelihood principle loosely states that all information from a set of data relevant
to inferences about the value of some parameter θ is contained in the likelihood function,
L(θ) ≡ Pr(data|θ). Therefore, by using a parameterised ρ as our model, a tomographic
estimate can be obtained by maximising the likelihood function L(ρ) = Pr(counts|ρ).
Process tomography is a similar task in which from a series of uses of a quantum
channel the transformation it performs is estimated. We begin with an aside regarding
representations of quantum processes. In Sec. 2.2.6 we introduced CPTP maps described
in Kraus form. For performing process tomography it is usually more convenient to use
the process matrix or Choi representation. Here we simply express the channel in a basis
of operators {B}
X
E(ρ) = χij Bi ρBj . (2.44)
i,j
The Kraus representation is recovered by diagonalising the matrix χ. The complete

positivity constraint is also straightforwardly expressed as the condition χ ≥ 0.
The Choi-Jamiolkowski (CJ) isomorphism states that there is a one-to-one correspon-
dence between the set of d-dimensional CPTP maps and a subset of d2 -dimensional states.
It is given by
ρE = (E ⊗ 1)[|ΨihΨ|] (2.45)
√1 |ii|ii is a maximally entangled state in d2 -dimensions. When rep-

P
where |Ψi = d i
resented in the computational basis, the process matrix χ is directly related to the CJ
state via ρE = χ/d. This link means that process tomography can be thought of as state
tomography upon the CJ state. The probability for an experiment in which a preparation
36
ρi is input, and a POVM element Ej is measured can be calculated as
Pr(j|i) = dTr[(ρ>
i ⊗ Ej )ρE ]. (2.46)
This expression can then be used to perform maximum likelihood process tomography.
The only difference is that the trace constraint Tr(ρ) = 1 becomes the trace-preserving
constraint TrA (ρE ) = 1/d. This new constraint means that a CJ state (or CPTP map)
has only d4 − d2 parameters, and not the d4 − 1 of a general state.
2.3.5 Distance measures
In order to characterise the quality or utility of a given state or process, it is important

to have some notion of the similarity of quantum states and channels. Here we give
the main distance measures that are used in this thesis to compare states and processes.
Given that most quantum experiments also involve estimating probabilities from relative
experimental frequencies, we also introduce the distance measures for classical probability
distributions that will be used.
The trace distance (sometimes known as the L1 or total variation distance) between
two probability distributions {pi } and {qi } is defined as
1X
D(~p, ~q) ≡ |pi − qi |. (2.47)
2 i
The trace distance is a metric on probability distributions 5 and has an operational mean-
ing as the probability of successfully distinguishing between the two distributions given
that the optimal event for this task is received. The statistical fidelity is defined as
X√
Fs (~p, ~q) ≡ pi q i . (2.48)
i
Unlike the trace distance, the statistical fidelity is not a metric, but it gives a measure of
the overlap of two distributions and is particularly useful to consider given its quantum
5
I.e. it satisfies the conditions D(x, x) = 0, D(x, y) = D(y, x) and the triangle inequality: D(x, z) ≤
D(x, y) + D(y, z)
37
generalisation.
Analogous measures can be defined for quantum states. The trace distance between
quantum states ρ and σ becomes
1
D(ρ, σ) ≡ ||ρ − σ||1 (2.49)
2
√
where the trace norm is ||X||1 = Tr( X † X). If both states are pure, this reduces to
p
D(|ψi, |φi) = 1 − |hψ|φi|2 . As in the classical case, it is a metric and has an operational
meaning related to a discrimination task. In the scenario in which, with equal probability,
an agent is given either ρ or σ, then the maximum probability of the agent correctly
guessing the state given as single copy (when they perform the optimal measurement) is
1
given by Prmax = 2
+ 12 D(ρ, σ) [32].
The quantum state fidelity of two states ρ and σ is
√ √
F(ρ, σ) ≡ || ρ σ||21 . (2.50)
If σ = |ψihψ| then F(ρ, |ψi) = hψ|ρ|ψi. This is often the case when comparing a realistic
experiment to its ideal counterpart. In such scenarios the fidelity can be understood as
the probability of measuring an ideal target state |ψt i (via the measurement {|ψt ihψt |, 1 −
|ψt ihψt |}) given an experimental state ρ 6 . Furthermore, it can be shown that the fidelity
can more reliably indicate the number of copies required to distinguish two states with
high accuracy, a highly relevant task in tomography experiments [35].
The fidelity and trace distance can be related to each other via the Fuchs-van de Graaf
inequalities [36]
p p
1− F(ρ, σ) ≤ D(ρ, σ) ≤ 1 − F(ρ, σ) (2.51)
where the second bound becomes an equality when ρ and σ are both pure.
These measures can also be promoted to distance measures on channels [34]. Two
natural generalisations of the state fidelity are the process fidelity and average gate fidelity.
6
√
Sometimes fidelity is defined as F, e.g. in [33], however we choose this definition so that it can be
interpreted as a probability in the manner described [34].
38
The process fidelity between two CPTP maps E and Λ is simply the state fidelity between
their CJ states
Fpro (E, Λ) = F(ρE , ρΛ ). (2.52)
The average gate fidelity is defined as the state fidelity between the channel outputs,
averaged over all pure input states
Z
Fave (E, Λ) = dµ(ψ)F[E(|ψihψ|), Λ(|ψihψ|)] (2.53)
where dµ is the uniform measure (the Haar measure) on state space. In the case that one
of the channels is unitary, Λ = U, these measures can be related via
(d + 1)Fave (E, U) = dFpro (E, U) + 1 (2.54)
where d is the Hilbert space dimension. The main reason for using these fidelity measures
is largely experimental convenience. They can be estimated with fewer measurements
than are required for process tomography [37] and the average gate infidelity (1 − Fave )
is the quantity calculated in randomised benchmarking protocols [38], a standard charac-
terisation procedure in matter-based qubits.
Although it is possible to define average and process versions of the trace distance
in a similar manner, a more commonly used measure is the diamond distance [39]. The
diamond distance between maps E and Λ is
||E − Λ|| = max ||(1 ⊗ E)[ρ] − (1 ⊗ Λ)[ρ]||1 (2.55)

ρ
and has an operational meaning which is analogous to the trace distance for states: it is
related to the maximum probability of successfully distinguishing between the channels
E and Λ in a single-shot using the optimal input state (which in general can be half of
an entangled state on a larger system) and making the optimal measurement. Although
less straightforward to calculate than the previous measures, the diamond distance can
be numerically computed via semidefinite programming [40].
39
In the context of quantum computing, the average gate fidelity represents an average-
case error rate whereas the diamond distance represents a worst-case error rate. These
measures can be related to each other via the inequalities
(d + 1)(1 − Fave (E, 1)) 1 p

≤ ||E − 1|| ≤ d(d + 1)(1 − Fave (E, 1)). (2.56)
d 2
2.4 Quantum computing

Having introduced the idea of quantum information and demonstrated ways in which it
can deviate from classical information, we now turn to the central objective of this thesis,
using quantum systems to perform computational tasks.
2.4.1 Turing machines
In a landmark paper of 1937 [41], Alan Turing developed a theory of computation based on
what is now known as a Turing machine. This is a hypothetical device which is comprised
of two elements. A tape which is divided into cells, each of which contains a symbol, Si ,
from some finite alphabet and a controller which is in an internal state Qi from some
finite set. The controller can perform the following actions; move one cell to the left or
right, read the current symbol, overwrite the current symbol, change internal state and
stop running and output an answer. A computer program can then be understood as the
set of rules which determine the action of the controller given {Si , Qi }, and the initial
configuration of the tape defines the instance or input of the problem.
Turing showed that there is a universal Turing machine, i.e. there is a Turing machine
which can take as input a description of any other Turing machine M and its input x and
produce the same output as M would have if given x. This result confirmed the existence
of programmable computers, an important notion that means that you can use the same
computer to check your emails and play DOOM.
In a famous proof, Turing then went on to show that the halting problem, the problem
of deciding if a given program will ever halt, cannot be solved with any Turing machine.
This lead to the Church-Turing thesis, which states that any problem that is computable
40
is computable by a Turing machine. Formally, a decision problem (one with the two
possible outputs ‘accept’ and ‘reject’) can be considered as a language L ⊆ {0, 1}∗ , a
collection of binary strings. For a computable problem there must exist a Turing machine
which accepts all strings in L and rejects all strings not in L.
A variation of the Church-Turing thesis that connects this abstract theory to physics is
the Physical Church-Turing Thesis, which states that all physically computable functions
are computable on a Turing machine. However, this statement does not address the
efficiency of computing. Certain problems can often be ‘practically uncomputable’ given
the resources in terms of time and memory that would be required to solve them. To
address this question, and understand where quantum computers enter this story, we
need the machinery of computational complexity theory.
2.4.2 Computational Complexity Theory
Complexity theory aims to classify computational problems into classes related to the
relationship between the size of the problem instance and the minimal resources required
to solve it. This can be quantified by considering the resources of time and space (i.e.
memory). TIME(f (n)) is the class of languages for which an instance of size n requires
an amount of time (or number of computational steps) that grows as O(f (n)). Similarly,
SPACE(f (n)) is the class of languages for which an instance of size n requires an amount
of space (i.e. number of bits of memory) that grows as O(f (n)). Using these definitions,
the following broader complexity classes can be considered:
• P (Polynomial) is the class of languages that are decidable by a Turing machine in

polynomial time, i.e. it is the class ∪k∈NTIME(nk ).
• NP (Nondeterministic Polynomial) is the class of languages for which, if a Turing

machine accepts then there is a polynomial size witness that allows you to verify
that answer in polynomial time. Formally, a language L is in NP if there exists
a polynomial time Turing machine M , a string x ∈ {0, 1}∗ and a witness w ∈
{0, 1}poly(|x|) such that ∀x ∈ L, M (x, w) accepts and ∀x ∈
/ L, M (x, w) rejects.
41
• PSPACE (Polynomial Space) is the class of languages that are decidable by a Turing
machine in polynomial space, i.e. it is the class ∪k∈NSPACE(nk ).
• EXP (Exponential) is the class of languages that are decidable by a Turing machine
k
in exponential time, i.e. it is the class ∪k∈NTIME(2n ).
and these classes obey the containments
P ⊆ NP ⊆ PSPACE ⊂ EXP. (2.57)
where only the final one is currently provably strict, although all are suspected to be so.
?
The question P = NP remains one of the great open questions in mathematics but it is
widely believed, and often assumed, that P 6= NP. Problems in P include basic numerical
operations, linear programming and primality testing. Problems in NP include factoring,
the travelling salesman problem and Donkey Kong [42]. The former of these is quite
special however. To see why we need to introduce some extra concepts. A problem is
said to be NP-hard if any NP problem can be efficiently reduced to it. A problem is then
said to be NP-complete if it is both a) in NP and b) NP-hard. This means that if an
efficient algorithm was known for any NP-complete problem, then all NP problems would
be efficently solvable. Factoring integers into their prime factors is a rare example of a
natural problem that is believed to be in NP but not NP-complete.
We can also consider a generalisation of P when the additional resource of a source of
randomness is allowed.
• BPP (Bounded-error Probabilistic Polynomial) is the class of languages for which

there exists a polynomial time Turing machine M, a string x ∈ {0, 1}∗ and a string
2
of perfectly random bits r ∈ {0, 1}poly(|x|) such that ∀x ∈ L, Pr[M (x, r)accepts] ≥ 3
/ L, Pr[M (x, r)accepts] ≤ 13 .

whereas ∀x ∈
The probabilities are chosen by convention and in principle could be any values p and
1 − p where p 6= 1/2. The hierarchy of all these complexity classes is depicted in Fig. 2.3.
This finally leads us to the Extended Church-Turing thesis (ECT). It states [8] that
42
EXP
PSPACE
NP
BQP
BPP
Figure 2.3: Hierarchy of computational complexity classes. Solid black lines indicate
known containments. Dashed red line indicates suspected relationship of BQP to other
classes.
“All computational problems that are efficiently solvable by realistic physical

devices, are efficiently solvable by a probabilistic Turing machine.”
The ECT is a statement about the relationship between physics and computation and
its claim is that no problem outside of BPP can be efficiently solved by a physical de-
vice. With the introduction of new ideas from quantum information theory however, this
statement now appears highly likely to be false.
2.4.3 Circuit model quantum computation
The stage is now set to introduce the concept of a quantum computer. The first description
of an analog quantum computing device is usually attributed to Feynman [43] in 1982,
where he suggested that the exponentially large state space of quantum systems could
only be simulated efficiently by a computer which works according to quantum theory
too. This was put onto digital ground in 1985 when David Deutsch defined a universal
Quantum Turing machine [44]. This is a generalisation of a Turing machine where now
the tape and controller can be quantum systems.
A perhaps simpler way to understand a universal quantum computer is in terms of the
43
quantum circuit model. In this model we can think of any computation as three stages:
Initialisation
We start with a system of n qubits. These qubits are prepared in a simple to prepare state
which can be dependent upon the input string x and a set of ‘ancilla’ qubits prepared in
the all zero state, e.g. |xi ⊗ |0...0i.
Gates
A universal quantum computer should be able to apply an arbitrary unitary transforma-

tion to the n qubit input state. For instance, for some computable function f there will
be a unitary U which prepares the state U |xi|0i = |xi|f (x)i. However, this procedure
will generally not be efficient. Instead, in analogy with classical computing where circuits
can be built up by composition of a small gate set of logic gates, we are interested in
quantum circuits which can be constructed by composing a polynomial number of gates
from a small gate set that act on just one or a few qubits.
Examples of common single qubit quantum gates are the Pauli matrices and the
Hadamard and T gate, which can be represented in the computational basis as
   
1 1 1 1 0
H=√  , T =   (2.58)
2 1 −1 i π4
0 e
and the most common two qubit gates are the controlled-NOT (CNOT) and controlled-
Z gates (CZ)
   
1 0 0 0 1 0 0 0
   
   
0 1 0 0 0 1 0 0
CNOT = 
 
 , CZ = 

.
 (2.59)
0 0 0 1 0 0 1 0
   
0 0 1 0 0 0 0 −1
These gates can generate entanglement when acted upon two separable qubits and so are
sometimes referred to as ‘entangling gates’. They can be understood as gates that perform
conditional logic in the computational basis, for instance a controlled-Z gate performs a
44
(a) (b) (c)

|0i |0i U1 |0i H T H
U7
|0i |0i U4 ...
|0i H T ...
U2
|0i |0i |0i H H
|0i
..
U ..
|0i
..
U3
U5
..
|0i
..
H T T
..
.. .. .. .. .. ..
.. .. .. .. .. ..
|0i |0i U6 |0i H T T
Figure 2.4: Circuit model quantum computation. All qubits are initialised in a known
state e.g. |0i⊗n and measured in the computational basis. a) The computation is encoded
in an n-qubit unitary U . b) This is decomposed into single and two-qubit unitaries. c)
Each of these are approximated via a universal gate set e.g. H, T ,CNOT.
Pauli Z unitary upon the second (target) qubit if the first (control) qubit is in the state
|1i; CZ = |0ih0| ⊗ 1 + |1ih1| ⊗ Z.
A universal quantum gate set is a finite set of quantum gates G such that ∀ > 0 there
is a gate sequence of elements of G that approximates (since the space is continuous) any
n-qubit unitary U ∈ SU (2n ) to an accuracy . The overhead required to -approximate a
non-gate set element on a fixed number of qubits U ∈ SU (d) using gate set elements in
SU (d) can be shown by the Solovay-Kiteav theorem (SK) [45] to be O(logc (1/)) with c
some constant. This is an important result as otherwise any potential quantum advantage
could be nullified by the inefficiency of decomposing gates.
Although almost any two qubit gate is universal [46], a conventional gateset that is
often considered is the combination of H, T and CNOT. Via SK all universal gatesets
can be reduced to each other with at most a polynomial overhead in any case. Once we
fix a universal gateset G then we are interested in the set of circuits {Cn } which contain
a polynomial number of gates from G.
Measurement
To read out, all qubits are simply measured in the computational basis. Although more
complicated measurement schemes could be used, they can always be subsumed somehow
into the circuit and ancilla qubits. The outcome of the measurement will therefore be a
bitstring for which the quantum computer then either accepts or rejects.
45
See Fig. 2.4 for a schematic summary of the circuit model of quantum computation.
With this model in place, we can now define the complexity class of problems that are
efficiently solvable on a quantum computer
• BQP (Bounded-error Quantum Polynomial) is the class of languages for which there
exists a family of polynomial-size quantum circuits {Cn } and a string x ∈ {0, 1}n
2
such that ∀x ∈ L, Pr[Cn (x) accepts] ≥ 3
/ L, Pr[Cn (x) accepts] ≤ 31 .
whereas ∀x ∈
Any proof that BQP 6= BPP would then overthrow the ECT and prove the power of
quantum computers exceeds that of a probabilistic Turing machine.
The first result to challenge the ECT was courtesy of Deutsch and Josza [47]. In their
problem, we are given a black-box known as an oracle that implements some function
f : {0, 1}n → {0, 1} and the task is to decide if it is balanced or constant. They showed
an exponential separation between a deterministic quantum algorithm using this oracle
and the corresponding deterministic classical algorithm 7 . However, this problem turns
out to be solvable in BPP when we allow a probabilistic classical algorithm. Inspired
by this, Bernstein and Vazirini [48] and Simon [49] then found new algorithms which
did show BQP 6= BPP relative to an oracle with a quasipolynomial and exponential
separation respectively. These problems both look quite contrived however, and rely on
this black-box oracle model which cast doubts on their practicality.
In 1994, building upon these previous results, Peter Shor then described an algorithm
which remains the best-known in quantum computation [50]. He showed that a quan-
tum computer can be used to solve the factoring problem (and discrete logarithm) in
polynomial time. As introduced previously, this problem is believed to be in NP, and
has immediate real-world applications. Specifically, the algorithm can be used to find
the prime factors a and b of a given large composite number N = ab in O(log N )3 time.
The best known classical algorithm, the general number field sieve, runs in time roughly
1/3
O(2(log N ) ). A scalable implementation of Shor’s algorithm would break much of the
world’s current public key encryption such as the RSA cryptosystem [51].
7
This determinism means that the algorithm actually defines an oracle relative to which EQP (Exact
Quantum Polynomial) is different to P.
46
The key quantum part of Shor’s algorithm does not actually directly solve the factoring
problem. Instead, it can be shown using number theory that with classical pre- and post-
processing, the problem of factoring can be efficiently reduced to order finding. This is the
problem in which we are given integers x and N (x < N ), which are coprime (they share
no common factors) and we wish to find the smallest integer r such that xr mod N = 1.
The quantum algorithm for order finding uses two registers of qubits, a work register
and a control register. The algorithm proceeds by a series of controlled unitary gates
acting on the work register, conditionally controlled by the control register, followed by
a quantum Fourier transform (QFT) and measurement of the control register. The QFT
consists of a network of single and two-qubit gates which performs the quantum analogue
P −1
of a fast Fourier transform, i.e. the mapping |ji → √1N N k=0 exp[i2πjk/N ]|ki.
When considered in the language of Group theory, the order finding routine in Shor’s
algorithm is a special case of the more general hidden subgroup problem (HSP). By
considering the HSP over different groups, similar efficient quantum algorithms have been
found which break other cryptosystems such as those based on the discrete logarithm or
elliptic-curve discrete logarithm problems (see for more details)
Since this breakthrough, many new quantum algorithms have been discovered (see the
Quantum Algorithm Zoo [52]), though often these algorithms fall into a few categories.
For instance, algorithms built around performing the quantum Fourier transform unitary
such as Shor’s algorithm and oracular algorithms such as Simon’s and the unstructured
search algorithm of Grover [53].
Perhaps the most important group of quantum algorithms for future applications
are those which perform quantum simulations. Simulating the dynamics of many-body
quantum systems is a computational bottleneck across many crucial areas of physics,
chemistry and material science. Therefore, the ability to efficiently simulate these systems
promises to contribute to a revolution in modern science. Following the spirit of Feynman’s
original proposal, a digital Hamiltonian simulation algorithm was found by Lloyd [54]. In
general, simulating the evolution of an arbitrary Hamiltonian applied for a time t simply
requires applying the quantum circuit that implements U (t) = e−iHt/~ , however this will
not be efficient unless it can be performed with a polynomial number of gates. Thankfully,
47
many-body quantum systems which occur in nature that we would wish to simulate can
usually be decomposed into a sum of a polynomial number of Hamiltonians which act
P
locally on just a few systems, H = i Hi . If all the Hi commute, then we can perform
the evolution via
Y Y
|ψ(t)i = e−iHt |ψ(0)i = e−iHi t |ψ(0)i = Ui (t)|ψ(0)i (2.60)
i i
and every Ui (t) can be efficiently implemented with our universal gateset. However,
in general the local Hamiltonians will not commute. To get around this, the solution
proposed by Lloyd is to use the Trotter formula
Y n
e−iHt = lim e−iHi t/n (2.61)
n→∞
i
so that evolution can be achieved by a series of repeated small time evolutions of each
local Hamiltonian. Understanding the number of Trotter steps n that would be in practice
required to achieve an accurate simulation of useful systems is the focus of ongoing research
[55, 56] as well as new approaches that use a Taylor series truncation instead of the
Trotter decomposition [57]. There is a subtlety involved in these algorithms, which is
that although the simulation may offer an exponential speed up over current classical
methods for performing the evolution, the result of the algorithm as described is not a
bitstring, but the full state |ψ(t)i. To make the algorithm useful in practice we then
need to identify some physical property of |ψ(t)i that can be efficiently estimated from a
polynomial number of measurements of the state.
2.4.4 Quantum error correction
Although we have seen that processing information that is encoded in pure quantum
states can provide unique advantages, it is important to remember that this information
has to live in the physical world. As such, it is crucial that the information is stable and
robust to the natural complexity and uncontrollability of realistic environments.
Classical bits, which are either 0 or 1 are often naturally robust as they are encoded
48
in macrostates of physical systems (e.g. for a CMOS gate with 5V supply, a logical 0 is
assigned to any signal with < 1.5V and logical 1 is any signal > 3.5V [58]) and if not it
is straightforward to make this information robust via simple error correction schemes. If
there is a probability p of a bit flip error, then as long as p < 1/2, by redundantly encoding
the bit in many copies and taking a majority vote the errors can be straightforwardly
suppressed.
To make quantum information robust, there are many new challenges to consider.
Firstly, quantum states form a continuum and so one might be sceptical that the infinite
set of possible error mechanisms can all be correctable. Secondly, there is a famous
result in quantum information theory which states that quantum information cannot be
perfectly cloned [59] (i.e. there is no unitary transformation which can always perform
U |ψi|0i = |ψi⊗2 for an unknown state |ψi) as we did in the classical example. Finally,
making measurements of quantum systems to identify if an error has occurred disturbs
their evolution, potentially itself corrupting the information. For all these reasons it is
remarkable that quantum error correction is even possible, though for the same reasons
strategies to keep quantum information robust are necessarily more complex and resource
heavy than their classical counterparts.
In order to understand quantum error correction, we need the concept of a codespace
C. This is a subspace of a larger Hilbert space H in which we can encode a ‘logical’ qubit
α|0̄i + β|1̄i redundantly in a log2 (dim(H)) qubit physical state. The simplest example
is Shor’s repetition code [60], the quantum analogue of the classical strategy described
above.
Using three physical qubits, the codespace is C = span(|000i, |111i) and |0̄i = |000i
and |1̄i = |111i. If this state is subject to a Pauli X error (i.e. a bitflip) on a physical
qubit, by measuring the operators Z1 Z2 and Z2 Z3 , where Zi denotes Pauli Z on qubit i,
we can protect against these errors. This is because the measurements do not collapse the
state, but any -1 outcome informs us that the state has left C and the specific outcome
string tells us which qubit has been flipped and therefore requires the application of a
correcting X gate.
A general condition for a codespace C to be correctable under some noise process
49
described by the CPTP map E is the PEEP condition [33]
P Ei† Ej P = αij P (2.62)
where P is the projector onto C, {Ei } are the Kraus operators of E and α is a Hermitian
matrix. If this condition is satisfied then there exists a recovery map R such that for
all states ρ in the codespace, R(E(ρ)) ∝ ρ, i.e. the noise model is correctable on C. A
key insight is that if the PEEP condition is satisfied for a channel with {Ei }, it is also
satisfied for any channel whose Kraus operators are linear combinations of {Ei }. This
√
means that if the depolarising channel { 1 − p1, p3 X, p3 Y, p3 Z} is correctable on
p p p
C, then any single qubit channel is correctable on C. For instance, if we promote the
three qubit Shor code into a nine qubit version where |0̄i = √1 (|000i
2
+ |111i)⊗3 and
|1̄i = √1 (|000i
2
− |111i)⊗3 which also protects against phase flip (Z) errors, then this can
protect against any single qubit error. Error correction codes can be classified by [[n, k, d]]
where n qubits are used to encode k logical qubits with a distance d (the minimum number
of errors which can perform a logical error in the codespace) and so the Shor code is called
a [[9, 1, 3]] code.
2.4.5 Fault tolerance
Of course, simply protecting static quantum information from corruption is not enough.
In order to perform fault-tolerant quantum computation we need to be able to prepare
states, implement gates and perform measurements without errors building up in our
computation or losing any potential speed up by resource overhead. Thankfully, in 1996
it was shown that fault-tolerant computation is indeed possible thanks to the existence
of threshold theorems [45, 61, 62].
To sketch out the spirit of these results we can consider a hypothetical architecture
where it is possible to prepare, transform and measure states with schemes which are able
to correct an error on a single qubit (like the Shor code). If errors occur on the physical
qubits with a probability p, then the probability that the encoding fails is the sum of
all cases where two or more errors occur, which grows like O(cp2 ). The question is then
50
whether this error rate can be further decreased with a reasonable resource overhead? One
way to do so is via code concatenation, where new codes are constructed by iteratively
performing the same encoding on each logical qubit of the previous code. If we concatenate
n
n times, then the probability of the fault-tolerant component failing will become c−1 (cp)2 ,
1
so if p < c
then the error rate will be double-exponentially suppressed with n. Say that
we want to be able to survive t timesteps with a probability of error , then we see
that only n = log(log(c)) levels of concatenation are needed and the size of the circuit
required to implement a fault tolerant component given the original circuit depth is d is
dn = O(poly(log( ct ))). We therefore see that if the probability of errors on physical qubits
p is less than a ‘threshold’, 1c , then increasing the size of the code will continue to suppress
errors in the computation with only a poly-logarithmic overhead. This is the fundamental
result which makes scalable digital quantum computers a practical possibility in principle.
The challenge that remains is therefore to find the codes on the physical qubits which
can protect against errors in state preparation, measurement and gates. In practice, the
latter is often the most challenging since performing gates upon encoded information
has the potential to spread errors to ‘clean’ qubits. A highly desirable property of a
code is therefore that it admits transversal gates, that is gates which can be performed
by operations acting on only a single qubit in each codeblock. Unfortunately, it was
proven that no code can perform a full, universal gateset transversally [63]. Instead, more
complicated schemes are required, such as performing gates by distilling large numbers of
uncorrected, noisy ancilla states into a few pure copies [64].
Although achieving fault tolerance appears a daunting task, the error thresholds re-
quired continue to improve with the current best proven threshold for arbitrary noise
6.7 × 10−4 [65] and up to ≈ 1% for a depolarising noise model with topological codes [66],
which will be briefly discussed below.
2.4.6 Stabilizer formalism
To complete our discussion of quantum error correction and for its use later on, we now
take a short digression to introduce the stabilizer formalism. The heart of this approach is
51
to describe a state not by its state vector, |ψi, but by a set of operators for which |ψi is the
+1 eigenvalue eigenstate. For instance, we can describe the state |0i by the operator Z and
|+i by X. The operators which are considered will be members of the Pauli group. The
Pauli group on a single qubit is given by P1 := {±1, ±i1, ±X, ±iX, ±Y, ±iY, ±Z, ±iZ}
and on N qubits PN := P1 ⊗ P1 ... ⊗ P1 . A stabilizer group S is then a subgroup of the
Pauli group which is said to ‘stabilize’ a subspace VS if
VS = {|ψi : Gi |ψi = |ψi ∀Gi ∈ S} (2.63)
where the Gi are called the stabilizers of VS . An n qubit stabilizer state is one which
can be efficiently represented by a generating set of stabilizers S = hG1 , ..., Gn i. More
generally, if S = hG1 , ..., Gk i with k < n then this defines a stabilized subspace (i.e. a
codespace) with a dimension dim(VS ) = 2n−k . Using this subspace as a code, errors can
be detected by measuring G1 , ..., Gk , producing a string ~s ∈ {1, −1}k of outcomes known
as syndrome measurements where any collection of −1 outcomes indicate an error. The
remaining n−k stabilizers can then be used as logical operators on the encoded codewords.
A particularly elegant application of this approach is in topological codes, the current
leading approach to implementing a fault-tolerant quantum computer. Consider a lattice
of qubits which are stabilized by k-local plaquette operators defined at each lattice site.
The key idea is that if there is a small stabilized subspace then information can be
encoded in the topological features of the whole lattice such that a logical error can only
occur if there is a long chain of errors across the lattice, whereas local clusters of errors
will be detectable by the plaquette syndromes. A physical error threshold can then be
identified, below which, increasing the size of the lattice will reduce the logical error
rate. In order to promote topological codes from quantum memories to computational
platforms also requires extra features which we will not describe, such as encoding qubits
in lattice defects [66] or performing lattice surgery [67]. Due to its high threshold (∼
1%), a popular architecture is the surface code, a 2D array of qubits involving 4−local
stabilizer generators. However, the scale of the task in achieving useful fault-tolerant
computation when error rates are not far below threshold can be seen from estimates of
52
performing Shor’s algorithm to factor a 2000 bit number requiring 107 physical qubits8
[68]. Therefore, along with the experimental efforts in reducing per-component error rates,
the development of better fault-tolerant codes and approaches with higher thresholds and
lower resource overheads are an area of intensive current research.
Another important insight that is gained from the stabilizer formalism involves the
efficient simulation of quantum circuits. The Clifford group of unitaries is the set of
unitaries which map Pauli group elements to Pauli group elements,
{U : U GU † ∈ PN ∀G ∈ PN }, (2.64)
and so can continue to be described within the stabilizer formalism. The Clifford group on
√
n qubits can be generated by H, Z, CNOT gates [33] and so it includes many standard
quantum circuits. Since an n qubit stabilizer state can be efficiently represented by its
n stabilizer generators (rather than its 2n amplitudes), and applying Clifford unitaries
simply updates these generators, any quantum computation which begins in a stabilizer
state, undergoes evolution through a Clifford circuit and Pauli observables are measured
is efficiently simulatable. This is the content of the Gottesman-Knill theorem [69].
A surprising consequence of the Gottesman-Knill theorem is that even circuits which
can generate highly entangled states remain classically tractable (as we will see in the
next section). Quantum computational speed up is therefore a much more subtle resource
than simply having lots of entanglement.
2.4.7 Measurement-based quantum computation
In the circuit model of quantum computation we begin with a product state and then
apply a series of few-qubit gates which typically prepares a large entangled state which is
then measured in the fixed computational basis.
In 2001, Raussendorf and Briegel introduced a completely new way of performing uni-
versal quantum computation [70]. Measurement-based quantum computation (MBQC)
8
And in fact ten times as many just to perform the ‘magic’ state distillation that is required to
implement a universal gateset
53
(a) (b) (c)

|+i
| i Z(✓) H m |+i
1 4
U1 U3
|+i 2 5
|+i X m HZ(✓)| i |+i U2 U4
|+i 3
Figure 2.5: Measurement based quantum computation. a) Simple teleportation circuit

demonstrating the principle of applying transformations via entanglement and measure-
ment. b) The graph state representation. Each vertex represents a qubit in the |+i state
and connected vertices are subject to a CZ gate. c) Converting between the circuit and
measurement based models. Single and two-qubit gates are applied in the measurement
based model via successive measurements upon a graph state. The outcomes of one layer
of measurements inform the measurement basis choices in the next layer.
begins with the preparation of a fixed, highly entangled initial state and quantum cir-
cuits are executed by performing adaptive local measurements on this state. In this
new paradigm, instead of creating entanglement, we consume it as a resource to move
information forward in the computation.
At the heart of MBQC is the notion of quantum teleportation. Teleportation is a sem-
inal result in quantum information which consists of the mapping of a quantum state from
one Hilbert space to another via the use of entanglement [71]. This allows the transmis-
sion of quantum information without having to transmit physical systems themselves. The
standard teleportation protocol consists of an unknown qubit and a Bell state |ψi ⊗ |φ+ i.
9
Performing a measurement in the Bell basis upon the first two qubits then projects
the remaining qubit into the state X s1 Z s2 |ψi, where s = {s1 , s2 } is a classical two bit
representation of the outcome of the measurement. Therefore, if we begin with entangle-
ment shared between two locations then, with the addition of a classical signal, quantum
information can be communicated between them. Teleportation experiments have been
performed between locations separated over 100km [72] and even between Earth and an
orbiting satellite [73].
A simplified version of the teleportation scheme is shown in Fig. 2.5a where Z(θ) =
exp(−iθZ/2). Entangling a logical qubit with an ancilla and measuring it in a basis
determined by θ teleports the logical state to the ancilla with a gate, again determined
by θ, applied to it. To generalise this approach of implementing gates via entanglement
9
One way of doing this is via the circuit consisting of a CNOT followed by H.
54
and teleportation we introduce a family of multi-qubit states known as graph states.

A graph state is defined by a graph G = (V, E) with vertices V and edges E. Every
vertex represents a single qubit prepared in the |+i state and every edge connecting two
vertices represents a CZ gate applied between the corresponding qubits. Since these are
all stabilizer operations, graph states have an efficient description in terms of the stabi-
Q
lizer generators Si = Xi j∈N (i) Zj , where N (i) is the neighbourhood (i.e. all connected
vertices) of i.
The cluster state is the graph state corresponding to a square lattice. It can be shown
that this state is universal for quantum computation [74]. A horizontally connected linear
graph can be considered as a single qubit channel with vertical connections between used
to implement two qubit gates. In this way it is possible to map a measurement-based com-
putation to a standard circuit based implementation. Measurement based computation
is particularly attractive in implementations where performing large numbers of gates on
the same physical systems is harder than producing many physical systems which are all
measured after a fixed depth circuit. As we shall see, photonics is one such architecture.
Finally, we note that the cluster state is by no means unique, various other states have
been found to be universal for MBQC [75], though all share a similar locality and efficient
description in terms of local operators. In fact, it has been found that most arbitary pure
states are too entangled to be measurement based resources [76, 77].
2.4.8 Other computational models
Here we briefly mention other models of quantum computation to give a fuller picture of
the landscape of potential approaches
Adiabatic quantum computation
Yet another paradigm for implementing quantum computing that does away entirely with
quantum gates is via the adiabatic theorem [78]. A computational problem is mapped such
that its solution is encoded in the ground state of a Hamiltonian Ĥ. A physical system
is then prepared in the ground state of an easy-to-implement Hamiltonian Ĥin . This
55
initial Hamiltonian is then tuned to Ĥ adiabatically, such that the system should remain
in the ground state. The speed at which this transition can be performed is related to
the spectral gap of the final Hamiltonian, the energy difference between the ground and
first excited state. Perhaps surprisingly, this model can be shown to be polynomially
equivalent to standard circuit based quantum computing [79].
DQC1
There are other known models of quantum computation that likely cannot exploit the full
power of BQP. In 1998, motivated by the physics of current NMR experiments, Knill and
Laflamme proposed the Deterministic quantum computation with one pure qubit (DQC1)
or one clean qubit model [80]. In this model, just a single pure qubit is prepared in the
|+i, along with many copies of the maximally mixed state. Performing a controlled-
U between the pure qubit and the mixed register, followed by measurement in the X
and Y bases allows an estimate of the unnormalized trace of the matrix U . Although
non-universal, DCQ1 raises some fundamental questions. Since there is no entanglement
generated in this process, this further questions the nature of quantum computational
speedup. Some have attributed the power of this model to quantum discord, a different
form of non-classical correlation [81].
Boson Sampling, IQP, Random circuit sampling
Recently, more non-universal computational models have been the subject of great interest
in relation to the concept of ‘quantum computational supremacy’ [82]. Since the resource
overheads for fault tolerant, universal quantum computing are so large, as sketched out
above, there are few prospects for such a device performing supra-classical computation
soon. This motivates non-universal models which are adapted to current experimental
hardware so that a quantum speed up may be demonstrable in the near term, without
full error correction overheads. Whilst likely to not be computationally useful and not
a definitive disproof of the ECT, such experiments would be convincing evidence that
it is possible to build machines which can exhibit controllable quantum coherence on
56
large enough scales to not be simulable on classical devices. Several such approaches
have recently been proposed, all consisting of short quantum circuits for which classical
sampling of the output distribution is likely to be inefficient. Instantaneous Quantum
Computing (IQP) [83] circuits consist of a set of qubits prepared in |+i, all subject to
gates which are diagonal in the Z basis, then measurement in the X basis. Random circuit
sampling [84] is an approach in which random single qubit unitaries from a universal gate
set are interspersed with CZ gates and the output is sampled. Boson sampling [8] is an
algorithm which is suited to linear optical experiments and will be discussed in greater
length in the next section.
2.4.9 Physical platforms
After our journey through the mathematical and computer science foundations of quantum
computing, we finally turn back toward physics. In order to encode and manipulate
quantum information we need physical quantum systems. In 1996 DiVincenzo formulated
his famous criteria which must be must met by physical implementations of quantum
computing. These are [85] (note that these criteria were formulated specifically for circuit
model computation and more general, architecture indepedent, formulations have since
been presented [86]):
1. A scalable physical system with well characterized qubits.
2. The ability to initialize the state of the qubits to a simple fiducial state, such as
|000...i.
3. Long relevant decoherence times, much longer than the gate operation time.
4. A universal set of quantum gates.
5. A qubit-specific measurement capability.
These all seem necessary requirements however there is an inherent friction at work here.
The ability to maintain long coherence times (3) implies that the system is very well
isolated from its environment, yet we need to be able to precisely interact with the system
57
to prepare, manipulate and measure it (2,4,5). The simultaneous suppression of noise and
increase in control is the key challenge in the quest to realise useful quantum information
processors. One other thing to note is the inclusion of scalabiity (1). It is often claimed
that like the first classical computers the first quantum computers will be very large,
non-practical machines in laboratories. Although this is likely to be true for the first non-
universal machines which demonstrate quantum advantages, if it is true that millions or
more physical qubits will be needed to perform useful, faster-than-classical computation,
then scalability, miniaturisation and manufacturability are likely to be hugely important
properties of the eventual preferred platform.
NMR
Many of the early demonstrations of experimental quantum information protocols were

performed with Nuclear Magnetic Resonance (NMR) systems. NMR uses the Zeeman
levels of spin-1/2 nuclei within a molecule in a magnetic field as its qubits. In contrast to
other platforms, NMR uses large ensembles of molecules to prepare ‘pseudo-pure’ states
e.g. of the form |00ih00| + (1 − )1/4 where is the polarization of the whole sample.
Gates and readout can then be performed using standard NMR techniques. There has
been much discussion on the legitamacy of this approach, due to the seeming lack of
entanglement [87] (this is what motivated the DQC1 model described previously) and the
lack of large-scale scalability. Either way, NMR computing has provided many insights
and tricks which have been subsequently transferred to other architectures.
Ion Traps
Ion traps remain perhaps the most sophisticated platform for quantum computing exper-
iments. In the standard Paul trap setup, an AC RF field and fixed electrostatic potentials
are used to confine single ions and laser cooling techniques are used to cool them to the
quantum regime. An ion qubit is then realised by its electronic energy levels (either
ground and excited or both ground hyperfine states) and can be manipulated by laser
control fields. Two-qubit gates are mediated via the quantised collective vibrations of
58
multiple ions and readout is achieved via detecting fluorescence from a separate transi-
tion. Ion trap experiments have demonstrated the highest reported fidelities [88] and the
maximum number of entangled qubits [89].
A major challenge for ion trap experiments is how to scale up beyond a single linear
trap. Two approaches include using photonic links [90] and shuttling ions around in a 2D
array [91].
Superconducting Qubits
Superconducting qubit systems are a promising platform for large-scale quantum infor-
mation processing. Using integrated circuits with superconducting Josephson Junctions
cooled down to tens of milikelvins, the collective electronic degrees of freedom (such as
current and voltage) enter the quantum regime. Two level systems can be realised in a
number of different configurations based around control of the quantised charge and phase
in such circuits and their states are readout via microwave cavity resonators.
Although initially blighted by low coherence times, superconducting qubits have re-
cently been demonstrated to perform high fidelity one and two-qubit gates [92]. The
current engineering challenges involve the realisation of a two-dimensional, scalable archi-
tecture and all the control and readout circuitry that this entails.
Silicon Quantum Dots
Another architecture which has been making great progress in recent years is based on the
spins of single electrons in quantum dots fabricated in silicon. Following the previously
stated vision that useful quantum technology will need to be immediately scalable, such
devices can be fabricated via standard CMOS techniques [93]. Recently, high fidelity
qubits have been demonstrated and the first two-qubit gates achieved [94].
Linear optics
Finally we turn to photonics, the platform which will be investigated in this thesis. En-
coding quantum information in the degrees of freedom of optical photons rather than
59
matter qubits has the advantage that they do not generally interact with their thermal
environment and are therefore less susceptible to decoherence. The drawback is that pho-
tons also do not interact strongly with each other, and therefore two-qubit gates using
linear optics must be implemented indirectly via measurement and feedforward and are
inherently probabilistic. With the development of integrated quantum photonic devices,
this technology platform is also moving towards scalability and miniaturisation.
The physics of Linear Optical Quantum Computing (LOQC) will be fully described
in the following chapter, beginning with a description of quantum optical states and their
interference properties.
60
Chapter 3
Quantum light
3.1 Introduction
In this chapter we move from the language of quantum information and computational
complexity to the physical platform that this thesis will investigate, quantum photon-
ics. We begin with a description of the quantum theory of light and its application in
linear optical quantum computational schemes. We finish with a review of current exper-
imental progress in quantum photonics and describe the experimental setup used in the
experiments described in Chapters 4 and 5.
3.2 Quantisation of the electromagnetic field

The classical electromagnetic field in a vacuum can be described fully by Maxwell’s equa-
tions
∇.E = 0 (3.1)
∇.B = 0 (3.2)
∂B
∇×E=− (3.3)
∂t
∂E
∇ × B = µ0 0 (3.4)
∂t
62
3. Quantum light
where E and B are the electric and magnetic fields respectively and µ0 and 0 are the
permeability and permittivity of free space (and µ0 0 = 1/c2 , where c is the speed of light
in vaccum). These fields can be described in terms of a vector potential A, and a scalar
potential Φ, such that
B=∇×A (3.5)
∂A
E = −∇Φ − . (3.6)
∂t
Working in the Lorenz gauge, where ∇Φ = 0, and the Coulomb gauge, where ∇.A = 0,
substitution into Eq. (3.4) results in the wave equation for A,
1 ∂ 2A
∇2 A = . (3.7)
c2 ∂t2
This equation can be solved via separation of variables to produce the general solution
X
−iωk t ∗
A(r, t) = a(k, s)e uk (r)(k, s) + a (k, s)eiωk t u∗k (r)∗ (k, s) (3.8)
k,s
such that
X h i
−iωk t ∗
E(r, t) = i ωk a(k, s)e uk (r)(k, s) − a (k, s)eiωk t u∗k (r)∗ (k, s) (3.9)
k,s
where {uk (r)} is an orthonormal set of mode functions labelled by k which will depend
upon the boundary conditions used, a(k, s) are the complex field amplitudes and (k, s)
is a complex polarisation vector where s labels the two independent polarisations.
Using the change of variables
r
1
a(k, s) = ωk x(k, s) + ip(k, s)
4 ω
r 0 k
1
a∗ (k, s) =

ωk x(k, s) − ip(k, s) (3.10)
40 ωk
63
3. Quantum light
it can then be shown that the classical Hamiltonian

Z
1 3
1
H= dr 0 E(r, t).E(r, t) + B(r, t).B(r, t) , (3.11)
2 µ0
where the integral is over all space defined by the boundary conditions, can be reduced
to the simple form
1X 2
H= (p (k, s) + ωk2 x2 (k, s)). (3.12)
2 k,s
We therefore see that the Hamiltonian is a sum of independent simple harmonic oscillators
for each mode.
A formal quantisation of the field can be found in e.g. [95, 96], here we will simply
1
apply the standard quantisation rules by replacing the position and momentum by the
quantum operators x̂(k, s) and p̂(k, s) obeying the commutation relations
[x̂(k, s), x̂(k0 , s0 )] = [p̂(k, s), p̂(k0 , s0 )] = 0 (3.13)
[x̂(k, s), p̂(k0 , s0 )] = i~δkk0 δss0 . (3.14)
The quantum Hamiltonian Ĥ can then be diagonalised by returning to the now quantised
form of Eq. (3.10) (for a single mode)
r
1
â = (ωx̂ + ip̂) (3.15)
2~ω
r
1
â† = (ωx̂ − ip̂) (3.16)
2~ω
which satisfy
[â, â] = [â† , â† ] = 0, [â, â† ] = 1. (3.17)
Applying this transformation results in the Hamiltonian operator describing a sum of

1
Note these do not correspond to physical position or momentum but define an abstract phase space
for describing the state of the optical mode.
64
3. Quantum light
independent quantum harmonic oscillators,
X h 1i
Ĥ = ~ωk â (k, s)â(k, s) + .
†
(3.18)
k,s
2
~ωk /2 just sets the zero-point energy, and so just appears as a constant
P
Since the sum k,s
offset in all expectation values, conventionally it is omitted
X
Ĥ = ~ωk a† (k, s)a(k, s) (3.19)
k,s
and from now on we will drop the hats on a operators for convenience. The Hilbert space
of a quantum harmonic oscillator is an infinite dimensional space whose eigenstates are
the Fock states, {|ni}, which correspond to states of a fixed number, n, of quantised exci-
tations. In the case of the electromagnetic field, these excitations are known as photons.
The operator n̂ = a† a is known as the number operator since its expectation counts the
number of photons in the oscillator’s state
a† a|ni = n|ni (3.20)
and a† and a are known as creation and annihilation operators since they add or subtract
excitations from the oscillator’s state
√
a† |ni = n + 1|n + 1i (3.21)
√
a|ni = n|n − 1i. (3.22)
A Fock state of n photons can therefore be written as
(a† )n
|ni = √ |0i (3.23)
n!
where we will use |0i to denote the vacuum state, i.e. a|0i = 0.
A basis state for a general multimode quantum optical state is therefore a tensor
65
3. Quantum light
product of Fock states for each mode
O
|nk,s i (3.24)
k,s
A fundamental difference between classical and quantum particles is the notion of indis-
tinguishability. In principle, a set of identical classical particles can always be individually
labelled such that their trajectories can be described independently. This is not the case
for quantum particles. If the particles are identical bosons e.g. photons, swapping two
particles does not change the global quantum state. Therefore, these particles can not be
individually labelled, and we say they are indistinguishable.
The field quantisation formalism we have considered here has this indistinguishability
implicitly defined via the use of creation and annihilation operators on Fock space. How-
ever, we can also describe the state of multiple bosons in a basis of the tensor product of
the single-particle states {|ψks i} of the modes. In order to satisfy the indistinguishability
condition, the full wavefunction in this picture must be symmetric under exchange of
particles, i.e.
Y O 1 X hO i
a† (k, s)|0i = |1ik,s ≡ √ σ |ψks i (3.25)
k,s k,s
n! σ∈Sn k,s
where σ is the operator which permutes the systems and Sn is the symmetric group
(whose elements are all the permutations of (1,2,...,n)). In contrast, for a system of
identical fermions, swapping particles results in a π phase shift to the global wavefunction
and therefore fermionic wavefunctions must be antisymmetric under exchange.
In what follows, we will often consider a series of separated but identical locations for
which we can assign the same Hamiltonian of the form Eq. (3.19) to each location (e.g.
a waveguide within an integrated photonic device). We will label these the paths of the
system. The operator which creates a photon in the ith path, in a state defined by the
superposition of complex amplitudes φ(k, s) in each mode will then be denoted as
X
a†i [φ] ≡ φ(k, s)a†i (k, s) (3.26)
k,s
66
3. Quantum light
with [a(k, s), a† (k0 , s0 )] = δkk0 δss0 . When the set of allowed modes is continuous, k →
P
dk and δkk0 → δ n (k − k0 ) becomes a Dirac delta function. If it is assumed that the

R
function φ is the same for all paths, then simply a†i will be used.
3.3 Linear optics

We now introduce the transformations that will largely be the focus of this thesis. Linear
optical transformations are those whose Hamiltonians describe a simple linear coupling
of optical modes, i.e. for a system of n paths as described in the previous section
n
X
Ĥ = Hij a†i aj (3.27)
i,j=1
where H is an n × n hermitian matrix.

The most standard linear optical component is a beamsplitter. Considering two spatial
paths with associated creation operators a†1 and a†2 , a beamsplitter between them can be
described by the Hamiltonian
Ĥ = a†1 a2 + a†2 a1 . (3.28)
This Hamiltonian generates a unitary operator U = e−iθĤ , where θ determines the reflec-
tivity of the beamsplitter. To show how this operator acts on photonic states we need the
Baker-Hausdorff lemma [97]
θ2 θ3
e−iθÂ B̂eiθÂ = B̂ − iθ[Â, B̂] − [Â, [Â, B̂]] + i [Â, [Â, [Â, B̂]]] + . . . (3.29)
2! 3!
Since [Ĥ, a†1 ] = a†2 and [Ĥ, a†2 ] = a†1 we can use Eq. (3.29) to show that
Ua†1 U † = e−iθĤ a†1 eiθĤ = a†1 cos θ − a†2 i sin θ (3.30)
Ua†2 U † = e−iθĤ a†2 eiθĤ = −a†1 i sin θ + a†2 cos θ. (3.31)
67
3. Quantum light
We can summarise these relations as

    
a†1 †
cos θ −i sin θ a†1
U   U =    . (3.32)
† †
a2 −i sin θ cos θ a2
When θ = π/4 this transformation describes a 50:50 beamsplitter operation.

This approach can be generalised to a general m-mode linear optical transformation
of the form of Eq. (3.27), resulting in the identity
   
a†1 a†1
   
 †  †
 a2  † a 
> 2
U  . U = T  . 
  (3.33)
 ..   .. 
   
†
am a†m
Where T = e−iθH is known as the transfer matrix and is the same matrix which
describes the transformation of classical electric fields through the component. Using this
identity, we can describe the Schrödinger evolution of an arbitrary m-mode p-photon state
|ψi under a linear optical transfer matrix T by
|ψ(T )i = e−iθĤ |ψi (3.34)

m
X Y [a†i ]ni
= U(T ) α~n √ |0i
n∈Np
~ i=1
ni !
X Y [U(T )a† U † (T )]ni
= α~n √i U(T )|0i
~
n i
n i !
X Y [ j Tji a†j ]ni
P
= α~n √ |0i (3.35)
~
n i
n i !
where Np is the set of all m element tuples (n1 , ..., nm ) which sum to p and the third line
follows from p uses of 1 = U † U. Therefore, the most concise description of a linear optical
transformation is via the update rules for the creation and annihilation operators
X X
a†i → Tji a†j , ai → Tji∗ aj . (3.36)
j j
68
3. Quantum light
3.4 Quantum Interference
3.4.1 The Hong-Ou-Mandel effect
Although the quantum operators evolve in accordance with classical fields as shown above,
due to the notion of indistinguishable particles explained earlier, single photons can
demonstrate uniquely quantum behaviour, even when there are no interactions between
the photons themselves.
The most famous example of such an effect is the quantum interference demonstrated
in the Hong-Ou-Mandel (HOM) effect [98]. Here, a single photon is incident upon each
input port of a 50:50 beamsplitter. Classically, one would expect that half the time the
photons will be detected at different output ports due to the two possibilities of this
event occurring (both reflect or both transmit). However since these two possibilities
are indistinguishable, their probability amplitudes destructively interfere, resulting in no
coincidental detections at all. We can see this using the transfer matrix for a beamsplitter2
 
1 1 i
BS = √   (3.37)
2 i 1
and the update rule for creation operators,
U(BS)|11i
=U(BS)a†1 a†2 |0i
=U(BS)a†1 U † (BS)U(BS)a†2 U † (BS)U(BS)|0i

1
= (a†1 + ia†2 )(ia†1 + a†2 )|0i
2
1
= i (a†1 )2 + (a†2 )2 |0i

2
1
= √ (|20i + |02i) (3.38)
2
√
Note that physical beamsplitters are often instead described by the transformations a†1 → (a†1 +a†2 )/ 2
2
√
and a†2 → (a†1 − a†2 )/ 2. This asymmetry is due to the dielectric coating used on one side of the
beamsplitter. Here we will be considering a symmetric beamsplitter transfer matrix which represents for
example, a directional coupler in integrated optics.
69
3. Quantum light
where we have ignored the global phase in the last line. This interference between in-
distinguishable single photons is what we will describe as quantum interference. The
phenomenon of the photons always exciting the same port of a balanced beamsplitter is
often referred to as photonic bunching.
We can also understand the HOM effect in the single particle representation described
by Eq. (3.25). Here, the effect appears much less mysterious, since we see explicitly that
the particles do not interact, but their correlation is related to the initial symmetrisation
of the wavefunction [99]
1
U(BS) √ |ψ1 i|ψ2 i + |ψ2 i|ψ1 i (3.39)
2
1
= BS ⊗ BS √ |ψ1 i|ψ2 i + |ψ2 i|ψ1 i (3.40)
2
1
= √ |ψ1 i|ψ1 i + |ψ2 i|ψ2 i (3.41)
2
where |ψi i is the single photon wavefunction in path i.

Finally, we can also think of quantum interference in terms of a Feynman path integral-
like approach. Here we consider the trajectories of distinguishable particles undergoing
evolution under a transfer matrix U . The amplitudes of trajectories which are the same
up to the labelling of the particles are then summed, resulting in the quantum amplitude
for that transition. For the HOM setup, the amplitude for the anti-bunched term can
therefore be expressed as the sum of the amplitudes of the trajectory in which the ordering
of the photons is preserved and the one for which they are swapped
h11|U(U )|11i = U11 U22 + U12 U21 (3.42)
= Per(U ) (3.43)
U =BS 1 1 i i
−−−→ = √ × √ + √ × √ (3.44)
2 2 2 2
=0 (3.45)
70
3. Quantum light
where the permanent of an n × n matrix A is defined as
XY
Per(A) = Ai,σ(i) (3.46)
σ∈Sn i
and is equal to the more common matrix determinant without the alternating sign. This
approach generalises to the result [8, 100]
Per(US,T )
hS|U(U )|T i = Q √ √ (3.47)
i si ! ti !
N N
where |T i = i |ti i, |Si = i |si i are the input and output Fock states respectively and
US,T is the submatrix of U formed by taking ti copies of the ith column of U and si copies
of the ith row of U . This picture again makes it clear that it is not really the photons
themselves that interfere but rather global trajectories of the photons which cannot, in
principle, be distinguished by any measurement.
3.4.2 The HOM dip and partial-distinguishability
In our discussion of quantum interference, we have been implicitly assuming that the
photons are in the same optical modes relative to each path, i.e. a†1 [ψ], a†2 [ψ]. In a practical
scenario however, their internal states (spectral, polarisation etc.) will not be completely
identical. If we assume that the beamsplitter performs the same transfer matrix on all
internal states, we can analyse the HOM setup again, this time where the photons begin
in arbitrary internal states
|ψHOM i =U(BS)a†1 [ψ1 ]a†2 [ψ2 ]|0i (3.48)

1h i
= i(a†1 [ψ1 ])2 + i(a†2 [ψ2 ])2 + a†1 [ψ1 ]a†2 [ψ2 ] − a†2 [ψ1 ]a†1 [ψ2 ] . (3.49)
2
Clearly when ψ1 = ψ2 we see that the final two terms cancel and we obtain the standard
bunching behaviour. However, if the functions ψ1 and ψ2 are not perfectly overlapped,
hψ1 |ψ2 i =
6 1, then a non-zero probability of a concidental detection will occur. Using a
model in which single-photon detectors can measure the number of photons in each path
71
3. Quantum light
(a) (b)
C(⌧ )
a†1 [ 1]
C(⌧ ) C(0)
V=
C(⌧ )
F12 2
= |h 1 | 2 i|
a†2 [ 2 ]
x C(0)
x(µm)
Figure 3.1: Hong-Ou-Mandel dip. a) Standard HOM setup where two photons with in-
ternal states |ψ1 i, |ψ2 i are impinged upon a balanced beamsplitter and the coincidental
countrate is measured, represented by the POVM element F12 , as a function of the dis-
placement of one photon relative to the other. b) Example experimental HOM dip. From
this experiment, the visibility V of the two photons can be measured. Data courtesy of
J. Carolan.
but cannot otherwise resolve the mode, the POVM corresponding to a joint detection is
F12 = F1 ⊗ F2 , where
X
Fi = a†i [ξk ]|0ih0|ai [ξk ]. (3.50)
k
and {ξk } is a basis of functions on the internal state space. The probability of this outcome
is then
1 − |hψ1 |ψ2 i|2
Pr(11) = hψHOM |F12 |ψHOM i = (3.51)
2
so we see that the effective indistinguishability of two photons is determined by the overlap
|hψ1 |ψ2 i|2 and can be measured by the probability of anti-bunching in the HOM setup.
In order to perform a typical HOM experiment, one photon is displaced relative to the
other such that at long delays τ , the countrate of the joint detection is proportional to
the probability of distinguishable particles, C(τ ) ∝ 1/2. When instead the displacement
is set to 0, C(0) ∝ Pr(11). As the relative delay is changed, a dip in countrate is traced
out, known as the HOM dip, see Fig. 3.1. From these countrates, the HOM dip visibility
can then be calculated
C(τ ) − C(0)
V= = |hψ1 |ψ2 i|2 . (3.52)
C(τ )
We can generalise this result to mixed internal states by considering ensembles of pure
states. By introducing superoperators Aj [%] which create a photon in the jth path with
72
3. Quantum light
P
an internal state described by % = i pi |ψi ihψi |, the state in Eq. (3.49) becomes

† †
ρHOM = U(BS) A1 [%1 ] ◦ A2 [%2 ] ◦ |0ih0| U † (BS) (3.53)
resulting in
V = 1 − 2Pr(11) (3.54)

= 1 − 2Tr F12 ρHOM ] (3.55)

= Tr S(%1 ⊗ %2 ) (3.56)

= Tr %1 %2 (3.57)
where S is the swap operator S|ψi i ⊗ |ψj i = |ψj i ⊗ |ψi i. It is interesting to note that for
pure internal states, the notion of photon indistinguishability as measured by the HOM
visibility is equivalent to the previously considered scenario of state discrimination in Sec.
2.3.5, where the distinguishability of states was determined by |hψ1 |ψ2 i|2 . However for
mixed states there is no longer an equivalence, since two photons that are in identical - but
mixed - states % have a HOM visibility equal to their purity V = Tr[%2 ] whereas their trace
distance is clearly zero. The notion of photon indistinguishability here is instead related
to the symmetry of the internal states under exchange as can be explicitly seen in Eq.
(3.56). In fact Eq. (3.56) generalises to arbitrary joint internal states, V = Tr[S%12 ], which
means that for instance for a symmetric state of the form |ψ12 i = √1 (|ψ1 i|ψ2 i + |ψ2 i|ψ1 i),
2
even if hψ1 |ψ2 i = 0, a perfect HOM dip will be observed.
3.5 Coherent states

Although in this thesis we will largely consider photonic states with a fixed number of
photons such that the Hilbert space can be truncated, more general optical states live
in the full infinite dimensional Fock space. An important class of optical states are the
73
3. Quantum light
coherent states. These states are the eigenstates of the annihilation operator
a|αi = α|αi (3.58)
and they can be described in Fock space as
D̂(α)|0i (3.59)
= exp αa† − α∗ a |0i

(3.60)
∞
2 /2
X αn
=e−|α| √ |ni (3.61)
n=0 n!
where D̂(α) is known as a displacement operator. Coherent states are considered the ‘most
classical’ pure states as their expectation for the electric field is equivalent to the classical
result with only the fluctuations of the vacuum. Coherent states are used to represent laser
light3 and the evolution of a set of coherent states |αi = |(α1 , α2 , ..., αm )T i ≡ m
N
i=1 |αi i
input to the m modes of a transfer matrix T evolve as
m
+
O X
U(T )|αi = Tij αj = |T αi. (3.62)
i=1 j
As such, coherent states do not demonstrate the quantum interference of the highly non-
classical Fock states. It is possible however, to see an analogue of the HOM dip, with a
maximum visibility of 0.5, by injecting two coherent states with a fast-varying relative
phase between them upon a beamsplitter [102].
Coherent states are a special case of a more general class of states known as Gaussian
states, which are states that can be described by Gaussian quasi-probability functions in
optical phase space [103]. Although not explored in this thesis, the two-mode squeezed
vacuum state we will describe later in this chapter is another example of a Gaussian state.
3
Physically this is not likely to be true, since phase diffusion processes with optical frequency coherence
times in the laser mean that coherence between the Fock states is quickly supressed i.e. the state is better
2 P |α|2n
described by ρ = e−|α| n n! |nihn| [101]. However, if all measurements on a laser beam are performed
in relation to a reference beam from the same source, then they will produce the same result as |αi.
74
3. Quantum light
3.6 Linear optical quantum computing

Optical systems make natural quantum information carriers due to their accessibility and
ease of manipulation in experiments. Photonic states tend to stay highly pure as they
do not generally interact with their thermal environment, and can be manipulated via
classical optical components. As such, photonic experiments do not require cooling to low
temperatures and can be constructed and calibrated in the classical regime. In principle, a
photonic qubit in an optical device can be defined using any two-level subspace of the full
multi-mode Fock space. The most common encodings however are known as single-rail
and dual-rail and are shown in Fig. 3.2. Single-rail qubits [104, 105] are defined by the
vacuum and single photon states of a single optical mode,
|0iL ≡ |0i, |1iL ≡ a† |0i. (3.63)
This approach however suffers from the difficulty of preparing and measuring superpo-
sitions of the vacuum and single photon states. This is because the phase relationship
between these levels |0i + eiωt |1i oscillates rapidly and so coherence between these states
cannot be practically measured without an additional phase reference, resulting in an
effective photon super-selection rule [106]. A more common approach therefore is to use a
dual-rail encoding [33], where the qubits are defined by a single photon across two modes
(here we shall assume these are the paths considered earlier, though in general polarisation
or more exotic modes could be used)
|0iL ≡ a†1 |0i = |10i, |1iL ≡ a†2 |0i = |01i. (3.64)
The advantage of this encoding is that universal transformations on a single dual-

rail qubit are straightforward. A linear optical component described by the transfer
matrix U acts on a dual-rail qubit |ψi as U(U )|ψi = U |ψi. It is possible to implement
any U ∈ SU (2) with just a variable-reflectivity beamsplitter and two phase shifters,
which perform the evolution eiθn̂ a† e−iθn̂ = eiθ a† on a single mode, via an Euler angle
decomposition.
75
3. Quantum light
(a) (b)
|0i |0i
|1i |1i
Figure 3.2: Single and dual-rail photonic qubits. a) Single-rail qubits are defined by the
vacuum and single photon Fock state of a single optical mode. b) Dual-rail qubits are
defined by a single photon spread across two optical modes.
Since a universal set of single qubit gates can be implemented straightforwardly with
linear optical components, universality can be achieved via just the addition of a CZ gate.
In the dual-rail picture, the only non-trivial evolution that is required is a π phase shift,
conditional upon there being a photon in both |1i modes. Although not possible with just
linear optical transformations, as explained in the proceeding section, this evolution can
be achieved via a nonlinear optical process in a Kerr medium. A cross-Kerr Hamiltonian
of the form
Ĥkerr = φa†2 a2 a†4 a4 (3.65)
results in the operator evolution
† †
a†1 → a†1 , a†3 → a†3 , a†2 → eiφa4 a4 a†2 , a†4 → eiφa2 a2 a†4 (3.66)
and so the logical qubit states evolve as
a†1 a†3 |0i → a†1 a†3 |0i, a†1 a†4 |0i → a†1 a†4 |0i, a†2 a†3 |0i → a†2 a†3 |0i, a†2 a†4 |0i → eiφ a†2 a†4 |0i. (3.67)
When φ = π, this performs the correct CZ gate [107, 108]. Although an attractive
proposal, this approach to optical quantum computation has not seen significant experi-
mental progress. This is because Cross-Kerr nonlinear coefficients are generally so weak
that to produce a π phase shift from a single photon requires an impractical amount of
material (for instance, it would require ∼ 105 km of optical fibre [109]). A more recent
76
3. Quantum light
approach is to use light-matter interaction, such as cavity-QED to achieve large nonlin-

earities though this results in a far higher experimental complexity. In addition, it has
been argued that the above single-mode analysis fails to account for multimode effects
that preclude high-fidelity gates when more fully investigated [110]. However, although
still not demonstrated, recent experimental [111] and theoretical results [112] do suggest
a more promising outlook for a nonlinear optical quantum computer.
3.6.1 KLM
For a long time it was believed that scalable universal quantum computation was not
possible with just single photons, linear optics and photon detection. This is because
without nonlinear optical transformations, there are no photon-photon interactions and
thus useful entanglement cannot be generated. Consider, for instance, the evolution
√
|00i → (|00i + |11i)/ 2. In the dual-rail picture it is easy to see that no linear optical
transformation can perform this transformation, since the Bell state cannot be factored
into a product of linear combinations of creation operators [113]
X X √
U(U )a†1 a†3 U † (U ) = Ui1 a†i Uj3 a†j =6 (a†1 a†3 + a†2 a†4 )/ 2. (3.68)
i j
However, in 2000 Knill, Laflamme and Milburn (KLM) [114] showed the surprising
result that photonic quantum computation was in fact possible without nonlinear optics.
The key insight was that the combination of quantum interference and photon detection
can be used to implement ‘measurement-induced’ nonlinearities which can be used to per-
form a probabilistic CZ gate. They then provided constructions which can convert these
probabilistic gates into an asymptotically deterministic scheme with only a polynomial
resource overhead, therefore proving that universal linear optical quantum computation
(LOQC) was in fact possible using only single photon sources, linear optics (including
quantum memories) and adaptive measurements (the ability to feedforward measurement
outcomes to decide future measurements).
The building block of the KLM CZ is the nonlinear sign-shift (NS) gate. This gate
performs a π phase shift upon the state in a single optical mode conditional upon it
77
3. Quantum light
containing two photons

UNSS c0 |0i + c1 |1i + c2 |2i = c0 |0i + c1 |1i − c2 |2i. (3.69)
The innovation was to use ancilla single photons and measurement such that if a desired
measurement outcome is registered, then the correct operation of the gate is performed.
This process of measuring additional photons to project the remaining photons into a
known logical state is known as ‘heralding’. Using a single ancilla photon, two additional
optical modes, the transfer matrix
 
1 − 21/2 2−1/4 (3/21/2 − 2)1/2
 
TNS =  −1/4 1/2 (3.70)
1/2 − 1/2
 
2 1/2 
 
(3/21/2 − 2)1/2 1/2 − 1/21/2 21/2 − 1/2
and heralding upon detection of a photon in the second mode and vacuum in the third
mode results in the correct application of the NSS gate,
h10|U(TNS )|ψin i|10i ∝ UNS |ψin i (3.71)
where states are represented in the Fock basis. The correct measurement pattern is
obtained with a probability of 1/4. Interestingly, this scheme has provably the highest
probability of success for an NS gate even if arbitrary ancilla resources are available [115].
Using two NS gates in the configuration shown in Fig. 3.3 can then implement the
required CZ gate. When there are 0 or 1 photons in modes 2 and 4, then the NS gates
act as identity such that the beamsplitters (which are taken to be self-inverse) result in a
trivial evolution. However, when the initial state is a†2 a†4 , a HOM bunching occurs after
the beamsplitter such that both NS gates implement π phase shifts, and the overall state
picks up the desired -1 factor.
The immediate problem with this gate is that it only succeeds with a probability of
(1/4)2 = 1/16, such that the probability of successfully applying n two-qubit gates is
exponentially small in n. When the NS gates do not succeed, the logical information is
78
3. Quantum light
(a) (b)
| i UNS | i
|1i
|0i
NS
NS
NS
NS
Figure 3.3: Elementary circuits in the KLM scheme. a) The nonlinear sign (NS) gate
performs a π shift on the |2i state conditional upon the correct heralding measurement.
Two of these gates can then be used to perform a probabilistic but heralded CZ gate
between two dual-rail qubits. b) This gate can be performed via teleportation such that
the logical qubits do not enter the gate unless the NS gates are successful. Conditional
upon the detection outcomes, phase shifts on the output may need to be applied.
lost. The second important result shown in the KLM paper was that the probability of
this gate can be boosted towards unity using quantum teleportation. The trick, as shown
in Fig. 3.3(b), is that a CZ can be commuted through a teleportation circuit such that the
probabilistic gates can be performed ‘offline’. This means the NS gates can be repeated
until successful, and only then are the logical qubits teleported to the output.
Unfortunately, the Bell measurements required for teleportation are themselves prob-
abilistic using only linear optics. KLM therefore gave a generalisation of the teleportation
circuit which uses an increasing number of ancilla resources to boost the probability of suc-
cess of the circuit towards one. In addition, they also provided error correction protocols
which allowed information to be preserved even when the teleportation fails.
Unfortunately, although a striking proof that LOQC is possible, the resource overheads
in the original KLM proposal are highly impractical. To implement a single two-qubit
gate with a 95% success rate would require tens of thousands of optical components [116].
Since the original proposal, a series of improvements and adaptations to the scheme have
been proposed [113], though all still suffer from a steep resource overhead for practical
79
3. Quantum light
implementations.
3.6.2 MBLOQC
Two developments which have pushed LOQC closer to a practical reality are the shift
to measurement-based quantum computation and the development of ballistic schemes.
Photonics is particularly well-suited to the MBQC model as the entangled resource state
generation process can be performed using probabilistic gates offline, leaving the compu-
tation to proceed deterministically via adaptive single qubit measurements.
Using cluster state computation for linear optics was first suggested by Nielsen[116]
whilst still using the existing KLM schemes, and later improved by Browne and Rudolph
[117] by replacing the KLM gates with simpler gates referred to as fusion gates. The
Type-I fusion gate, shown in Fig. 3.4(a), takes two photonic qubits as input and measures
one, performing a logical projection which is unitarily equivalent to the form
1
FI = √ |0ih00| + |1ih11| (3.72)
2
with a probability of 1/2. This transformation can be used to ‘fuse’ together graph states.
Consider two graph states where one qubit from each 1 and 1’, connected to qubits i and
j respectively, are input to the gate. The initial states can be written as
1 Y
0 1 Y
|Gi = √ |0i|G1 i + Zi |1i|G1 i , |G i = √ |0i|G10 i + Zj |1i|G10 i (3.73)
2 i
2 j
where |Gk i is the graph state remaining when qubit k is removed. Now, consider applying
the fusion gate FI
1 Y
FI |Gi|G0 i = √ |0i|G1 i|G10 i + Zi Zj |1i|G1 i|G10 i (3.74)
2 i,j
which is a new graph state joined at the location of one of fused qubits. If a different
detection pattern is measured, then the resulting state will be the same graph state with
80
3. Quantum light
(a) FI (b) FII

|1i |1i
|0i |0i
|1i
|0i
Figure 3.4: Type I and II Fusion gates. a) The type-I gate FI takes two dual-rail photonic
qubits and measures one, preserving the remaining qubit with a probability 1/2. b) The
type-II gate FII takes two dual-rail qubits and measures both, projecting them onto a
Bell state with a probability of 1/2.
local Pauli rotations which can be corrected for.

The Type-II fusion gate, shown in Fig. 3.4(b) is unitarily equivalent to the linear op-
tical Bell state analyser [118]. This gate takes two photonic qubits as input and measures
both, projecting onto a logical bell state, e.g. FII = |φ+ ihφ+ |, with a probability of 1/2.
Again consider the fusion of two qubits 1 and 1’ which are connected to qubits i and j
respectively. The stabilizer generators associated with 1, 1’, i and j are given by
Y Y
S1 = X1 Zi , S10 = X10 Zj , Si = Xi Z1 Ai , Sj = Xj Z10 Bj (3.75)
i j
Where Ai and Bj represent Z on all other qubits connected to i and j. To fuse the
clusters, we first perform a Hadamard rotation on qubit 1’ which results in
Y
S10 = Z10 Zj , Sj = Xj X10 B (3.76)
j
Projecting this state onto the Bell basis of qubits 1 and 1’ then leads to, for the case of a
|φ+ i outcome,
Y Y
S̄1 = Z1 Z10 , S̄10 = X1 X10 , S̄i = Xi Zj Ai , S̄j = Xj Zi Bj (3.77)
j i
forming a new graph state with qubits 1 and 1’ measured out and bonds between all
qubits i and j.
81
3. Quantum light
(a) (b)
..
T .. ..
T ..
. . . .
Figure 3.5: Schematic of ballistic photonic quantum computation. a) A supply of single

photons is used to generate a supply of few-qubit entangled states. These entangled
states are then fused together probabilistically. b) The resulting state is a cluster state
with missing bonds. If the probability of success of the fusions is high enough, then there
is guaranteed to be a spanning path with high probability.
Using a combination of these two fusion gates and a supply of dual-rail Bell states, a
cluster state capable of universal computation can be constructed. The Bell states can
be generated from four single photons with a probability of 3/16. This circuit will be
explored in detail in Chapter 6.
The second paradigmatic shift from the KLM approach was first suggested in [119],
which is to use a ballistic instead of repeat-until-success strategy. Even though the entan-
glement can be generated offline in the MBQC model, when a probabilistic gate fails it will
generally be necessary to re-try the gate, sometimes many times. This results in an archi-
tecture which requires sophisticated fast-switching networks and delay lines. In contrast,
ballistic schemes use each gate just once and then use adaptive post-processing to achieve
universal computation on the resulting non-ideal cluster states. This is achieved by using
ideas from percolation theory. Once the probabilistic fusion gates have been implemented,
the resulting global state is a graph state of a target lattice with a fraction of missing
bonds (entanglement) and sites (qubits). If the probability of missing bonds is below a
critical threshold, then with high probability there will be a spanning path through the
lattice, providing a universal resource state. For a schematic of such schemes, see Fig. 3.5.
One way to then perform MBQC is to ‘renormalise’ blocks of imperfect lattice into single
qubits of a perfect lattice [119].
The best known current scheme for MBQC in a ballistic approach are the protocols
from [120, 121], where ‘boosted’ fusion gates, which trade-off extra ancilla resources for
82
3. Quantum light
higher success probabilities, are used to ballistically generate graph states from a supply of
three-photon entangled states. These schemes will be described in more detail in Chapter
6.
3.6.3 Boson sampling
Despite the huge effort and resources being poured into the global race to build quantum
computers in the hope of solving real world problems, the more fundamental scientific
challenge of demonstrating the inherent computational advantages of quantum systems
over classical systems is yet to be convincingly shown. An obvious route to this goal would
be to use the best known quantum algorithm, Shor’s algorithm, to factor a number that is
not possible with current classical technology. However, this approach has two drawbacks.
Firstly, experimental progress has been limited to a handful of qubits, whereas Shor’s
algorithm requires hundreds of logical qubits (and therefore likely millions of physical
qubits) to outperform current algorithms. Furthermore, it is yet to be conclusively proven
that factoring is a classically hard problem and the existence of an efficient classical
algorithm for factoring would not imply any dramatic theoretical consequences such as
the collapse of complexity classes. A recent trend in quantum information science has
therefore been to develop models of quantum computation which, although non-universal,
perform classically hard tasks under strong complexity arguments and are tailored to
current physical platforms in order to demonstrate a quantum advantage with near-term
experimental resources [83, 84, 122]. The prototypical example of such an algorithm is
Boson Sampling.
Boson Sampling is the following computational problem:
• Take as input an m × m unitary matrix U ∈ SU (m). Consider the m × n matrix A

formed from its first n columns (where we will take m = n2 ).
• Sample from the distribution over tuples S = (s1 , s2 , ..., sm ), where si ∈ {0, 1} and
2
P
i si = n, with the probability mass function Pr(S) = |Per(AS )| , where AS is the
matrix formed by taking the rows of A for which si = 1.
83
3. Quantum light
s1
n
{ ..
U 2 Cn
2
⇥n2
..
s2
s3
s4
|hS|U (U )|1...10...0i|2 = |Per(US )|2
. .
|0i s n2 1
|0i s n2
Figure 3.6: Boson Sampling schematic. n single photons are input into a linear optical
network with transfer matrix U acting on n2 modes. All collision-free detection patterns
S = (s1 , s2 , ..., sn2 ) are measured, sampling from a distribution whose probability mass
function is related to the permanent of submatrices of U .
It can be straightforwardly seen, via Eq. (3.47), that this computational problem
maps onto the following physical problem (see Fig. 3.6):
• Take as input an m × m unitary matrix U ∈ SU (m). Construct a linear optical

network which implements this transfer matrix.
• Sample from the distribution of n-fold photon detection patterns obtained at the
output of the linear optical network if the input is prepared in the Fock state |1n i =
|1i⊗n |0i⊗m−n .
In 2010, Aaronson and Arkhipov [8] showed that if there is an efficient classical algo-
rithm for Boson sampling, then the polynomial hierarchy would collapse to its third level,
a complexity implication thought highly unlikely. In contrast, the quantum experiment
described above scales polynomially in experimental resources and avoids some of the
impracticalities of LOQC, such as probabilistic gates, fast switching and feed-forward.
However, any realistic, non-error corrected experiment will not solve the Boson sampling
exactly. The crucial result proven in [8] was therefore that, modulo some highly plausible
conjectures, this sampling problem is still classically hard for distributions D0 that are
merely close to the Boson sampling distribution DBS in variation distance ||D0 −DBS || ≤ .
We now give a brief summary of this proof.
84
3. Quantum light
We first need to introduce some new computational complexity classes. #P is the class
of function problems of the form “compute f (x)” where f gives the number of accepting
paths of an NP machine. Note that this is not a decision problem as we were considering
in Sec. 2.4.2 but is instead a counting problem. If an NP problem is of the form ‘are
there any Hamiltonian paths with cost less than c’ then the corresponding #P problem
asks ‘how many Hamiltonian paths have cost less than c’.
The polynomial hierarchy, PH, is the union of a set of complexity classes, the 0th level
NP
of which is P and the higher levels take the form NP, NPNP , NPNP , ... 4
where NPNP is
the class of problems that can be solved by an NP machine given an NP oracle. Similar
to P6=NP, It is a fundamental assumption in complexity science that this full hierarchy is
not contained within any specific level, i.e. it does not collapse to this level. These two
complexity classes can be linked via Toda’s theorem [123] which shows that PH ⊆ P#P ,
i.e. that all problems in the entire polynomial hierarchy can be solved in polynomial time
with access to a #P oracle.
The reason that we have introduced the #P class is that the matrix permanent is the
prototypical problem in this class. In 1979, Valiant [124] proved that the problem of com-
puting the permanent of a matrix of 0’s and 1’s is #P-complete. The fact therefore that
the amplitudes of transitions of Fock states through linear optical networks are propor-
tional to permanents (see Eq. (3.47)) implies a striking connection between complexity
theory and the physics of fundamental particles. Interestingly, the same transmission ex-
pression for a system of non-interacting fermions is given by the determinant, a function
that can be computed in polynomial time.
Unfortunately, although the evolution of photons in linear optical networks is governed
by permanents, it is not possible to use this fact directly to compute a #P-hard problem
using a photonic experiment. This can be understood by considering the problem of
estimating the permanent of an arbitrary n × n matrix B. In order for B to be embedded
into a unitary transfer matrix, we can place a rescaled version B/||B||, where ||B|| is the
largest singular value of B, in the top left corner of a 2n × 2n matrix U . In Chapter 4 we
4
PH also contains the similar set of classes with at the first level CoNP. CoNP is the complementary
class to NP i.e. if an NP problem asks whether there exists a string satisfying certain properties then the
problem in CoNP asks whether all strings satisfy certain properties.
85
3. Quantum light
will explore how this is done, and further applications of this concept. The probability of
measuring the photons exciting in the first n modes is
|h1n |U(U )|1n i|2 = ||B||−2n |Per(B)|2 . (3.78)
This probability however, is exponentially small for ||B|| > 1, meaning that an exponential
number of trials would be needed to accurately estimate it. On the other hand, when
||B|| < 1, there exists a classical algorithm, Gurvits’ algorithm [125], for approximating
2
the permanent to additive error ± in time O( n2 ) which does just as well as the quantum
experiment.
The insight of Aaronson and Arkhipov was to consider a more general scenario and
define the Boson sampling problem as above to perform exactly the task that quantum
experiments actually solve. Their argument then relies on two facts; firstly, that approx-
imating |Per(X)|2 to a multiplicative factor is still #P-hard. Secondly, there exists an
algorithm by Stockmeyer [126] that says that a BPPNP machine can always estimate the
probability that a polynomial-time randomised algorithm accepts to within a multiplica-
tive factor, even when it is exponentially small. Therefore, if a polynomial-time algorithm
exists for exact Boson sampling, then this could be fed to Stockmeyer’s algorithm to es-
timate a #P quantity in BPPNP , resulting in P#P ⊆ BPPNP . Then, via Toda’s theorem
and known containments, it can be shown that
P#P ⊆ BPPNP (3.79)
PH ⊆ BPPNP (3.80)
NP
PH ⊆ NPNP (3.81)
and so the polynomial hierarchy collapses to its third level.

Although a surprising complexity result this result is hard to connect to real exper-
iments. This is because realistic experiments can only approximately sample from the
exact Boson sampling distribution. Thankfully, Aaronson and Arkhipov were also able to
show, modulo two conjectures, that the polynomial hierarchy also collapses if there is an
86
3. Quantum light
efficient algorithm for approximate Boson sampling.

Their proof works via the following steps. The first important observation is that
the n × n submatrices of a Haar-random m × m unitary matrix, with m ≥ n6 (though
it is expected to also be true for m ≥ n2 ) are approximately equal to matrices of i.i.d
Gaussians. Therefore, they first consider the following problem
|GPE|2± :Given a matrix X ∼ N (0, 1)Cn×n of i.i.d Gaussians, estimate |Per(X)|2 to
within additive error ±n! in poly(n, 1/) time. (3.82)
and prove that if there exists a classical algorithm which can sample from a probability
distribution D0 such that ||D0 − DBS || ≤ in poly(n, 1/) time then |GPE|2± is solvable in
BPPNP . The next step is then to show that if the Permanent Anti-Concentration Con-
jecture (PACC) holds, which roughly says that the distribution of Gaussian permanents
is not highly concentrated around zero, then |GPE|2± is poly-time equivalent to another
problem GPE×
GPE× :Given a matrix X ∼ N (0, 1)Cn×n of i.i.d Gaussians, estimate Per(X) to
within multiplicative error ±|Per(X)| in poly(n, 1/) time. (3.83)
A final conjecture, the Permanent-of-Gaussians conjecture (PGC), then states that GPE×
is #P-hard. Then, via Toda’s theorem and the same arguments as before we have shown
that if an efficient classical algorithm for approximate Boson sampling exists, then the
polynomial hierarchy again collapses to its third level.
Boson sampling is a particularly attractive approach for current quantum photonic
technologies. It does not require either the measurement-induced nonlinearity or adaptive
measurement schemes of KLM or MBQC, simply quantum interference between photons.
Furthermore, it seems possible that Boson sampling experiments can be scaled with-
out the full machinery of fault-tolerant error correction towards the quantum supremacy
threshold.
A number of adaptations and follow-up results have been suggested since the original
87
3. Quantum light
proposal. These include using inherently probabilistic sources in a so-called ‘scattershot’

approach [127], rather than the |1n i state, more generally using Gaussian states or mea-
surements [128, 129], and allowing for a constant number of lost photons [130].
A number of proof of principle experimental demonstrations of Boson sampling have
been performed [131–135], with the current record 5 photons in 9 modes [136]. The
regime in which quantum experiments would outperform classical machines was initially
predicted to be in the region 20-30 [8] and recently even as low as 7 [137].
However it was recently shown by co-authors and myself [138] that problem sizes of
up to 50-60 can likely be solved by supercomputers with practical runtimes.
Metropolised independence sampling (MIS) is a Markov chain Monte Carlo method
used to sample from distributions for which direct sampling is difficult. The algorithm
works via constructing a Markov chain on the state space of possible outcomes where
the probability of transitioning between a pattern S and S 0 is split into a proposal and
acceptance stage. The proposal distribution g(S) is chosen such that efficient sampling
from this distribution is possible. A new state S 0 is then drawn from this proposal
distribution, and the algorithm accepts this state with a probability given by
h f (S 0 ) g(S) i
T (S 0 |S) = min 1, (3.84)
f (S) g(S 0 )
where f (S) is the distribution one wishes to sample from. Using the Boson Sampling
distribution as f (S), a good proposal g(S) was found to be the equivalent distribution
for classical distinguishable particles whose probability mass function is given by
Pr(S) = Per(|AS |2 ). (3.85)
If one can show that the resulting Markov chain has converged to its stationary distribu-
tion, and using a thinning procedure to eliminate autocorrelation between samples (since,
for instance, if a proposal is not accepted the chain will remain in the same state for
multiple timesteps), this algorithm can be compared to experimental approaches. Strong
numerical evidence suggests that convergence and thinning requires only ∼ 100 steps,
88
3. Quantum light
independent of n, and so can output a sample from the Boson Sampling distribution with
a constant number of permanent evaluations. This is a vast improvement over the direct
2
sampling approach, which would require the calculation of nn permanents.
30 photon Boson Sampling was shown to be possible in tens of minutes on a standard
laptop and, assuming that the algorithm continues to perform equally well on larger
instances, that ∼ 50 photon experiments could be simulated on a supercomputer. This
result was recently reinforced by Ref. [139], where the classical complexity of exact Boson
sampling was shown to be O(n2n + poly(m, n)).
3.7 Quantum photonics

In this section we shall turn to the current experimental state-of-the-art in quantum
photonics, along with a description of the experimental setup used in the experiments
described in this thesis.
3.7.1 Photon sources
Although sources of laser light that can be described quantum optically have been widely
available for many decades, quantum technologies require the generation of optical states
which are ‘truly’ nonclassical such as single photons. Single photon sources (SPS) are
perhaps the most pressing engineering challenge for current quantum photonic technolo-
gies. This is because they require nonlinear processes with trade-offs which are hard to
simultaneously satisfy. We can list the most important properties of single photon sources
as
• Source efficiency: the probability that a single photon is generated by our source on
every attempt.
• Single photon purity: the purity of the full photonic state describing the single
photon (e.g. its spectral purity).
• Multi-photon contamination: the probability of the state to be found in higher
89
3. Quantum light
5
photon number spaces.
• Indistinguishability: how identical the single photons produced are, both from sub-
sequent uses of the same source and between different sources.
• Repetition rate: the rate at which photon generation is attempted.
There are two general approaches to realising single photon sources, all-optical and
matter-based. They each have their own strengths and weaknesses. All-optical sources
tend to produce purer and more indistinguishable photons at the cost of a low efficiency
due to their probabilistic nature and to keep multi-photon contamination minimised. On
the other hand, matter-based sources can generally operate with higher efficiencies and
do not suffer from multi-photon contamination to the same degree. However, producing
pure and indistinguishable photons from multiple sources is generally a greater challenge.
Both sources can operate at high repetition rates however, into the GHz regime, one of
the reasons LOQC and photonic technologies are appealing for information processing.
The most common SPS which has been the workhorse of quantum optical experiments
for several decades is based upon the principle of spontaneous parametric down-conversion
(SPDC). SPDC is an optical process which occurs within nonlinear media. When light
is propagating in a medium, the electromagnetic Hamiltonian from Eq. (3.11) has an
additional term in terms of the macroscopic polarization P
Z
1 1
H= dr3 0 E(r, t).E(r, t) + B(r, t).B(r, t) + E(r, t).P(r, t) . (3.86)
2 µ0
P can be expanded in a power series of the electric field amplitudes
P(r, t) = P(1) (r, t) + P(2) (r, t) + P(3) (r, t) + ... (3.87)
Usually, the terms in this expansion are considered in the frequency domain (rather than
5
Note that often this is considered to be included in the purity. Here we describe these as separate
imperfections and consider them independently.
90
3. Quantum light
time), and their components are given by [140]
(n) (n)
Pj (ω) = χjα1 ...αn (ωσ ; ω1 , ..., ωn )Eα1 (r, ω1 )...Eαn (r, ωn ) (3.88)
Pn
where ωσ = i ωi , χ(n) is the nonlinear susceptibility tensor of rank n + 1 and α ∈
{x, y, z}.
SPDC is a χ(2) , three-wave mixing process in which a single photon from a pump field
is converted to two photons of lower frequency. The standard setup consists of pumping a
nonlinear crystal with a strong laser. In this configuration the pump can be described by
a classical field with a spectral amplitude α(ωp ) and assuming a frequency-independent
χ(2) in the spectral range of interest, it can be shown that the interaction Hamiltonian of
the SPDC process is given by [141]
Z Z
ĤI ∝ dωs dωi α(ωs + ωi )Φ(ωi , ωs )a†1 (ωs )a†2 (ωi ) + h.c. (3.89)
where s, i label the signal and idler modes respectively, the function Φ describes the
phase-matching conditions and the constant of proportionality is a function of χ(2) and
the intensity of the pump field. The function f (ωi , ωs ) = α(ωs + ωi )Φ(ωi , ωs ) is known
as the joint spectrum of the signal/idler fields. Ideally, the joint spectrum is separable,
f (ωi , ωs ) = ψ(ωi )φ(ωs ). In this case, the Hamiltonian simplifies to
HI ∝ a†1 [ψ]a†2 [φ] + h.c. (3.90)
This process results in a unitary operator
Ŝ(ξ) ≡ exp(−ξa†1 [ψ]a†2 [φ] + ξ ∗ a1 [ψ]a2 [φ]). (3.91)
Acting on the vacuum, this generates the two-mode squeezed vacuum state
∞
X 1 †
Ŝ(ξ)|0i = sech(r) (−1)n einθ tanhn (r) (a1 [ψ]a†2 [φ])n |0i (3.92)
n=0
n!
91
3. Quantum light
where ξ ≡ reiθ . When the squeezing parameter r is small this state is given by ∼
P n 2 4
n ξ |ni1 |ni2 . Choosing r such that r r means that if a detector for one mode
clicks, with high probability a single photon has been prepared in the remaining mode.
The cost is that most of the time no photons are generated at all. One promising approach
to overcoming this inherent inefficiency is to multiplex many SPDC sources[142–144], such
that with high probability at least one source clicks. The generated photon can then be
switched to the desired mode via feedforward. The main barrier for current realisations
of such schemes is the lack of availability of high speed, low-loss switches.
Another challenge in SPDC is producing highly pure single photons. This is because
in practice the joint spectrum is rarely separable. Under fairly general conditions it can
be represented in Schmidt decomposition as [145]
N
X √
f (ω1 , ω2 ) = pi ψi (ω1 )φi (ω2 ). (3.93)
i=1
This results in a Hamiltonian which is a sum of terms like Eq. (3.90) and a state which
is a tensor product of independent two-mode squeezers
N
O
Ŝi (ξi )|0i. (3.94)
i=1
In the single pair subspace, at low pump powers, this is then
X√
pi |ψi i1 |φi i2 (3.95)
i
where |χii ≡ a†i [χ]|0i. This means that the generated photons are spectrally entangled.
If the photon in mode 2 is measured, the resulting photon in mode 1 is then described by
the state
X
ρ= pi |ψi ihψi | (3.96)
i
p2i .
P
and so the purity of the source is simply i
For some platforms, such as integrated Silicon photonics, there is no χ(2) coefficient
92
3. Quantum light
due to the centro-symmetry of the crystal lattice. Therefore sources following very similar
principles can be implemented using higher-order nonlinearities such as spontaneous four-
wave mixing (SFWM) [146].
The other main approach to producing single photon sources is to use a single quantum
emitter. This is an (approximately) two-level quantum system with a radiative optical
transition that can be used to generate single photons. Usually such sources work via the
coherent excitation of the system, followed by its spontaneous decay and the emission of
a single photon. If all non-radiative decay is suppressed then such sources can be highly
efficient with minimal multi-photon contamination. Emitters that have been used as SPSs
include single atoms [147], ions [148], molecules [149], defects in crystal structures such
as nitrogen-vacancy centres in diamond [150] and quantum dots [151].
The main shortcomings of these approaches have usually been twofold. On the one
hand, producing high efficiency sources in practice requires highly efficient coupling of
the light from the source into a single optical mode. Secondly, producing high purity
and indistinguishable photons is often difficult due to for instance phonon dissipation and
local environmental effects [152].
Recent improvements in quantum dot sources have demonstrated increased efficiencies
and high indistinguishability between photons produced in successive time-bins [153, 154],
and seem to be the most promising route to producing tens of photons in the near-term.
Integrating such sources into a miniaturised architecture, as all-optical sources can be,
remains an engineering challenge although recent progress coupling integrated gallium
arsenide dots to waveguides is promising [155].
3.7.2 Linear optical networks
As we will see in the next chapter, arbitrary linear optical networks (LON) can be com-
posed by a set of beamsplitter and phaseshifter operations. For example, if a dual-rail
qubit is encoded in the polarisation degree of freedom {|Hi, |V i} of a single photon propa-
gating in free-space then these operations can be implemented via waveplates. Waveplates
are birefringent crystals oriented such that they can produce a phase shift in one polar-
93
3. Quantum light
isation component relative to the other without changing the propagation of the light.
The combination of so-called half-wave plates and quarter-wave plates (differing in their
thickness relative to the wavelength of light) allows for arbitrary unitary transformations
of polarisation encoded qubits.
However, in order to achieve large-scale photonic networks, one needs to construct
complex interferometers between many optical modes. This is difficult to achieve using
bulk optical components such as waveplates and half-silvered mirrors due to the practical
size of the optical setup required and the problem of maintaining precise phase-stability
across the network in the face of fluctuating ambient conditions.
Integrated optical circuits [5, 6] have provided a platform allowing dramatic improve-
ments in experimental capabilities due to their inherent phase stability and small device
footprint. Waveguides which spatially confine light in integrated devices are formed via
the refractive index contrast between a cross-section of a core material and its surround-
ing cladding. By solving Maxwell’s equations within this region, including the boundary
6
conditions, TE (where there is no electric field in the direction of propagation) and TM
(where there is no magnetic field in the direction of propagation) modes can be found.
Waveguide structures are then designed such that in the frequency range of operation only
a single TE and TM mode can propagate, so the mode in a single waveguide corresponds
to a single path of a linear optical network, or a single rail of a photonic qubit.
Beamsplitters in integrated devices are often achieved by two principal mechanisms.
Directional couplers are formed by bringing two waveguides close together such that there
is an overlap between the single waveguide modes. This coupling can be described by
the Hamiltonian from Eq. (3.27). By coupling for a length L, an arbitrary reflectivity
beamsplitter can in principle be achieved. An alternative approach to perform the same
linear optical transformation is to use a multi-mode interference (MMI) device. This is a
device in which two waveguides are fed into a larger region in which many more modes
can propagate. The dimensions of this region are then selected such that light can be
coupled back into single mode waveguides but interferences have changed their relative
6
Interestingly, for a rectangular waveguide, no analytical solutions to Maxwell’s equations exist and
modes must be found by discretising the space and using a numerical mode solver.
94
3. Quantum light
intensities. In practice, MMIs often provide a more precise reflectivity at the cost of being
lossier.
Phaseshifts can be implemented by modulation of the phase in one waveguide relative
to others. There are a number of possible ways of implementing this, such as using
electro-optic modulators which exploit the Pockels or Kerr effect. To date however, most
experiments have used thermo-optic modulators which are much slower, but lower loss.
Increasing the temperature of a waveguide changes its refractive index. This can be
understood by considering the linear term of the polarization, P(1) = 0 (r − 1)E, where
√
n = r , and noting that the dipole moment of the medium, and so r , is increased as the
temperature is increased. The temperature-dependent phase shift due to the thermo-optic
effect relative to another waveguide is
L dn
∆φ = 2π ∆T, (3.97)
λ dT
where L is the length, λ the frequency, ∆T the temperature shift and dn/dT the thermo-
optic coefficient. Thermo-optic phaseshifters are usually achieved by patterning resistive
heaters above the waveguides and applying currents to change the temperature.
Popular material platforms for integrated quantum photonics include Silica, Silicon
Nitride, Indium Phosphide, Galium Arsenide, Lithium Niobate and Silicon. All these
platforms have strengths and weaknesses in terms of their size, loss, nonlinearity and
performance of passive and active optical components. Most current multi-photon exper-
iments, including those reported in this thesis, use a Silica (SiO2 ) cladding and silica-doped
core, similiar to optical fibres [156–160]. These circuits have the advantage of being low
loss in both propagation but also in coupling into and out of the chip. However, the
size of the waveguides and integration density means that such circuits are not appropri-
ate for large-scale applications. More potentially long-term scalable platforms are Silicon
[7], where high component densities can be achieved and classical CMOS-compatible sili-
con photonics fabrication techniques can be used and Gallium arsenide [161], which can
support deterministic quantum dot photon sources.
95
3. Quantum light
3.7.3 Detectors
The final key component of a quantum photonics experiment are the single photon de-
tectors [162]. Again, we can list the most important properties which define the quality
of a single photon detector
• Efficiency: the probability of detecting a photon if one arrives at the detector
• Dark count rate: the probability of measuring a false positive, i.e. a photon when
none is present.
• Timing jitter: the variation in time interval between the absorption of a photon and
the registering of a classical electrical pulse signifying its detection.
• Dead time: the time it takes for the detector to be ready for subsequent measurement
following a detection event.
• Photon number resolving: the ability of the detector to measure the exact number
of photons arriving at the detector.
Again, there are often trade-offs, for instance between efficiency and dark count rate.
The most commonly used detectors in current experiments are single photon avalanche
photodiode (SPAD) detectors. These are semiconductor devices biased such that a single
photon absorption process triggers a self-sustaining avalanche creating a sharp rise in
current. SPADs tend to have efficiencies of around ∼ 50 − 60%.
More recently, detectors with improved performance have been demonstrated and are
commercially available using superconducting technology. Transition edge sensor (TES)
detectors [163] work via the use of a superconducting film at its critical temperature.
When a photon is absorbed, the temperature rises, increasing its resistance and this
change can measured, allowing for number resolution. Superconducting nanowire single
photon detectors (SNSPDs) [164] are another type of superconducting detector which,
although not offering intrinsic number resolving7 , are faster and can be placed onto inte-
grated waveguide structures for scalability. SNSPDs work below the critical temperature
7
Though pseudo-number resolving can be achieved e.g. by the use of multiple nanowires on the same
optical waveguide
96
3. Quantum light
Figure 3.7: Experimental setup. The multiphoton source uses a pulsed Ti:Saph laser to
perform SPDC, preparing two two-mode squeezed states which are coupled into fibre. The
integrated photonic chip consists of six germanium doped silica waveguides with a silicon
cladding with 30 directional couplers and thermo-optic phase shifters. Photon counting
is performed via an array of single photon avalanche photodiodes. The experiment is
interfaced with a computer for real-time photon coincidence counting and setting of all
phase shifters. Figure adapted from [165]. Reprinted with permission from AAAS.
but biased such that the current is close to its critical current. A photon incident upon the
nanowire results in the formation of a ‘hotspot’, a small non-superconducting region with
nonzero resistance, this then produces a voltage pulse which can be measured. Supercon-
ducting detectors have shown efficiencies of over 90% as well as improving timing jitter
and dark count rates. The main obstacle to scalability using these detection systems is
that they both require low temperatures (< 10K), whereas the ability to operate at room
temperature is one of the most appealing aspects of quantum photonic technologies.
3.7.4 Experimental setup
Here we introduce the experimental setup shown in Fig. 3.7 which was used in the exper-
iments reported in this thesis.
97
3. Quantum light
Source
A pulsed, Titanium:Sapphire laser (Coherent Chameleon) is used to generate 140fs pulses

at ∼ 808nm at a repetition rate of 80MHz. A half-wave plate and polarising beamsplitter
(PBS) are used to attenuate the power. Next, a β-barium borate (BBO) crystal is used
to perform second harmonic generation. Dichroic mirrors are then used to remove the
remaining 808nm light and a 0.5mm Bismuth Triborate BiB3 O6 (BiBO) crystal is used to
perform SPDC from the up-converted 404nm pulse. Down-converted photons are emitted
in a cone at opening angle θ = 6◦ and pass through a 3nm interference filter at 808nm.
Prisms are aligned to couple light from four equidistant points on the SPDC cone (at 12,3,6
and 9 o’clock) into polarisation maintaining fibres (PMF). Ideally, this source produces
two independent two-mode squeezed vacuum states (at 12,9 and 3,6 o’clock respectively).
Two of the collection stages have motorised actuators which can be computer controlled
for performing HOM-dip type experiments.
When either pair is connected directly to the detectors, the ratio between the co-
incidence detection rate and the single detection rate is ∼ 12%. Taking into account
the detector efficiency, this translates to a heralding efficiency of around 24%. The in-
distinguishability, measured by Hong-Ou-Mandel dip visibility, for photons generated in
the same pair generation event is ∼ 0.98 while for photons generated in different pair
generation events, the HOM visibility is ∼ 0.91.
These non-unit visibilties result from a number of mechanisms. Since each generated
photon is incident upon the 3nm filter at different positions and angles, their spectral
properties will all be slightly different. When performing interference between photons
from the two independent sources there is the additional effect of non-unit purity caused
by spectral entanglement as discussed in Sec. 3.7.1. Also in this case a further temporal
jitter is introduced from group velocity mismatch resulting from generation of the photon
pairs at different locations in the crystal [166]. Since the downconverted photons travel at
a different speed through the crystal than the pump, the two photon pairs will experience
a different amount of temporal walk-off on a shot-by-shot basis.
98
3. Quantum light
Circuit
The silica-based integrated photonic chip was fabricated at the Nippon Telegraph and
Telephone company (NTT) in Japan. Flame hydrolysis deposition followed by pho-
tolithographic and reactive ion etching was used to fabricate germanium doped silica
(SiO2-GeO2) waveguides with dimensions 3.5µm × 3.5µm with a silica cladding onto
a silicon substrate. Thin-film Tantalum Nitride (Ta2 N) thermo-optic heaters were then
fabricated on top of the circuit with dimensions 1.5mm × 50µm. The circuit is formed of
a cascaded array of 30 directional couplers (each with a length of 500µm) and 30 phase
shifters designed to perform a universally reconfigurable transfer matrix on six waveguide
modes. The design of this circuit will be explored further in the following chapter. The
coupling losses have been estimated as ∼ 9% per facet and the directional couplers at
< 2.3%. The average loss fibre-to-fibre was measured to be ∼ 42%. The device is actively
cooled via a Peltier cooling unit.
The thermo-optic modulators are driven by electronic heater driver boards designed
in-house which can deliver up to 20V with 4.9mV resolution and current up to 100mA.
These are then interfaced with a computer to set all the heaters to implement a given
transfer matrix.
Detectors
The detection system uses 12 SPADs (Perkin Elmer SPCM-AQRH-14), each with effi-
ciencies 50 − 60%, a dark count rate of ∼ 100Hz, timing jitter of ∼ 350ps and a dead time
of 32ns. A coincidence counting card time-tagging all simultaneous channels in a time
window usually set to be around 2ns is used to register detection events (see the thesis
of Peter Shadbolt for more details [167]). For each channel it is possible to set a specic
time delay that is used by the counting card to compensate for the discrepancy in the
signals arrival time introduced in the experiment by optical fibres, detectors, electronics
and coaxial cables.
99
Chapter 4
Universal Linear Optics
4.1 Introduction
Universality is a crucial concept in computation. The ability to reprogram a single device
to solve any possible instance of a problem is a key capability which has driven the modern
computing revolution. For linear optical devices, this notion of universality is captured
by the ability to perform any unitary transformation upon a set of optical modes. The
standard approach to constructing quantum optical experiments for over twenty years
has involved assembling and realigning new setups (or more recently, fabricating new
integrated devices) to implement a given optical circuit. However, it was shown in 1994
by Reck et al. [168] that there is a constructive scheme for implementing an arbitrary
unitary transfer matrix via a set of beamsplitters and phase shifters. If these components
are all fully reconfigurable, then such a design provides a platform for universal linear
optical transformations. In this chapter we report the implementation of the world’s first
universal linear optical chip on six optical modes and describe its use as a testbed for
experimentally exploring a wide range of linear optical quantum information processing
tasks.
100
4. Universal Linear Optics
4.2 The Reck et al. Scheme

It has been known since the 19th century [169] that any m × m unitary matrix in U (m)
can be decomposed into a series of ∼ m2 /2 two-dimensional unitaries. The scheme of
Reck et al. uses a circuit of tunable beamsplitters and phase shifters to physically realise
this decomposition in linear optics. The building blocks are two-mode rotations acting
on modes i and i + 1 of the form
 
1 0
 
0 1
 

 
..
.
 
 
 
 iφ θ iφ θ 
 e sin 2 e cos 2 
Ti (θ, φ) = 
 . (4.1)
θ θ
− sin 2

 cos 2 
 
..
.
 
 
 
 
 1 0
 
0 1
These transformations can be achieved (up to a global phase which can be corrected
for) using two controllable phase shifters via the circuit on the left-hand side of Fig. 4.1(b)
where the Mach-Zehnder interferometer (MZI) with a phase θ acts as a variable reflectivity
beamsplitter and an additional phaseshifter (PS) implements φ.
MZI(θ) = BS PS(θ) BS (4.2)

   
1 1 i  eiθ 0 1 i
=    (4.3)
2 i 1 0 1 i 1
 
(θ+π)
−i 2 
sin 2θ cos 2θ
=e  (4.4)
cos 2θ − sin 2 θ
(θ+π)
PS(φ) MZI(θ) = e−i 2 T (θ, φ) (4.5)
We now consider the norm 1 complex vector (U11 , U21 , ..., Um1 )> which forms the first
column of the unitary matrix U that we wish to implement. This vector of amplitudes
101
(a) TT11(✓1,1
1,1 ,, 1,1))
1,1
TT12(✓
(✓1,1
2,2, , 2,2
1,1 )) TT21(✓ 1,1 ,
(✓1,2 1,2))
1,1
..
.
D3 D2 D1
. .. T3 (✓2,3 , 2,3 ) ..
.
Dm ...
Tm (✓ (✓1,1, ,
T1m,m 1,1 ))
m,m ... Tm (✓m,1 , m,1 )
(b)
T (✓, )
1,1 ✓1,1
✓
BS BS 2,2 ✓2,2 1,2 ✓1,2
3,3 ✓3,3 2,3 ✓2,3 1,3 ✓1,3

4,4 ✓4,4 3,4 ✓3,4 2,4 ✓2,4 1,4 ✓1,4
5,5 ✓5,5 4,5 ✓4,5 3,5 ✓3,5 2,5 ✓2,5 1,5 ✓1,5
Figure 4.1: The Reck et al. scheme. (a) Left: an arbitrary linear optical unitary U can
be decomposed into a product of unitaries Di acting on modes i, ..., m. Right: These
can then be decomposed into a cascade of two-mode rotations T . (b) Left: Physical
implementation of a T rotation as a MZI with phase shift θ followed by a phase shift φ.
Right: Full scheme for a 6 × 6 unitary.
can be prepared from an input in the first mode via the circuit
D1 (1, 0, ..., 0)> = Tm−1 (θ1,m−1 , φ1,m−1 )...T2 (θ1,2 , φ1,2 )T1 (θ1,1 , φ1,1 )(1, 0, ..., 0)>
= (U11 , U21 , ..., Um1 )> . (4.6)
With Uij = uij eiαij , the θ’s are used to set the amplitudes u according to the relation
θ1,1 θ1,2 θ1,i

ui1 = cos cos ... sin i = 1, ..., m − 1 (4.7)
2 2 2
and φ’s set the phases (relative to the phase on the mth element, which is not tunable)
via φ1,i = αi1 − αm1 . We can now consider the new matrix
 
eiαm1 0 ··· 0
 
 
†
 0 
D1 U =  ..
 (4.8)
U m−1
 
 . 
 
0
If the same procedure is performed to prepare the first column of U m−1 given an input
102
in the second mode, we obtain another diagonal cascade of transformations
D2 = Tm−1 (θ2,m−1 , φ2,m−1 )...T3 (θ2,3 , φ2,3 )T2 (θ2,2 , φ2,2 ). (4.9)
†
Continuing in the same way for the remaining columns we produce Dm−1 ...D2† D1† U = Λ
where Λ = Diag[(eiαm1 , eiαm2 , ..., eiαmm )] and can be implemented by m additional phase
shifts. Rearranging we therefore find U = D1 D2 ...Dm−1 Λ. Fig. 4.1(a) shows this circuit
schematically and Fig. 4.1(b) shows the physical layout of the circuit for m = 6 1 . There
are m(m − 1)/2 pairs {θij , φij } in the two-mode rotations plus the m input phase shifts
results in a full parameterisation of U (m).
We note that the Reck scheme results in a triangular network of components. Recently,
a new scheme has been proposed by Clements et al. [170] which uses a new decomposi-
tion of U (m) into a square network with the same number of components. The Clements
scheme benefits from a smaller maximum circuit depth, the number of layers of beam-
splitters, and is less susceptible to path-dependent losses due to its symmetric structure.
On the other hand, the Reck scheme is likely to be preferable when states are only to be
input into a subset of modes due to its reduced effective circuit depth in this case. There
is also some evidence that it has improved resilience to fabrication errors [171].
4.3 A Universal Linear Optical Processor

The device introduced in Sec. 3.7.4 is an experimental realisation of a universal linear
optical network. The circuit is comprised of 30 directional couplers and 30 thermo-optic
phase shifters forming 15 MZIs, plus an additional 15 phase shifters to implement the φ
parameters in the Reck scheme. Silica is used as the platform of choice due to the low
insertion and propagation losses which allow for multi-photon experiments with practical
count rates. To implement a 2π phase shift for a given heater requires a temperature
shift from room temperature to around 74◦ C, applying a power of ∼ 0.8W for ∼ 10ms.
1
The observant reader will note that this circuit is reversed relative to (a). Due to experimental
practicalities the device is operated this way. The Reck decomposition is therefore performed for U > .
Phases on the output modes are also neglected since photon counting measurements are insensitive to
them
103
This switching time is likely not fast enough for the adaptive feedforward required for
true LOQC, but suffices for applications such as testing linear optical circuitry and gates
and the bosonic simulations presented in the next chapter.
4.3.1 Characterisation
Since each heater is slightly different, and the path lengths of the Mach-Zehnder arms are
not perfectly matched, each phase shifter must be calibrated such that its phase-voltage
relationship is known. This phase-voltage relationship is modelled using the parameteri-
sation
φ(V ) = mod a + bV 2 + cV 3 , 2π

(4.10)
where a is the phase difference when the heater is switched off i.e. at 0V, b is the ohmic
response of the resistor and c is a correction for non-ohmic response (b >> c). Phase
shifters in MZIs can then be calibrated by sweeping the voltage when a heralded single
photon2 is input into its top port and the count rate is measured at the top output port.
This count rate is then modelled by

C = A − B cos φ(V ) (4.11)
and A, B and a, b, c are estimated via a least squares fit. The a values were found to
be approximately uniformly distributed between 0 and 2π, and typical values for b and c
were 0.05 and 0.001. The procedure for a full calibration of the device proceeds as follows.
A calibration curve is measured using T1 (θ1,1 , φ1,1 ) to estimate θ1,1 (V ). Then, setting
θ1,1 to divert light to the input of T2 (θ1,2 , φ1,2 ), θ1,2 (V ) can be estimated in the same way.
This procedure is then continued to calibrate all {θi,j }. In order to calibrate the {φi,j }
heaters, the calibrated θi,j ’s are set to form larger MZIs. For instance, the arrangement
T2 (π/2, 0)T3 (π, 0)T1 (π, 0)T2 (π/2, φ2,2 ) (4.12)

2
This can be done equivalently with classical laser light and intensity measurements, though given the
operation will be to some degree mode-dependent, the best calibration in practice uses the same optical
states that are to be used for the quantum experiments.
104
can be used to measure φ2,2 (V ). See Fig. 4.2(b) for a schematic of this procedure and
example calibration curve. Once a calibration is obtained for all heaters, any unitary can
be ‘dialled up’ on the device via its Reck decomposition. This calibration was found to
continue to perform well over a long period of many months without any noticeable drop
in experimental fidelities.
The uncertainty in the phase set at a given voltage can be estimated via the un-
certainties in a, b and c. Averaging across the device we find ∆θ ≈ 0.03 rad, with the
uncertainty predominantly coming from ∆a. The experimental fringes also have a visibil-
ity V = (Cmax − Cmin )/Cmax which is given by 2B/(A + B). The main reason that V < 1
is due to the non-ideal reflectivity of the directional couplers. If the reflectivities of the
two directional couplers are given by 1/2 + δ1 and 1/2 + δ2 then
p
2 (1 − 4δ12 )(1 − 4δ22 )
V = p (4.13)
1 + 4δ1 δ2 + (1 − 4δ12 )(1 − 4δ22 )
and if δ1 = δ2 , which may be the case if fabrication errors result in a constant offset across
a device, then V = 1 − 4δ 2 . The average visibility seen across all heaters is 0.995, so we
can loosely estimate that δ ∼ 0.03. This non-unit visibility in each MZI means that it
is not possible to achieve full transmission or reflection, even if the phase shift θ was set
perfectly.
A final experimental consideration is determining any sources of crosstalk. The two
most likely mechanisms are thermal and electrical. Thermal crosstalk occurs when heat
diffusion from one phase shifter can change the phase in another interferometer. It is
mitigated in the design of the device by the fabrication of trenches between waveguides
(see Fig. 3.7) resulting in no measurable thermal crosstalk. Electrical crosstalk can occur
when several heaters share a common ground. When multiple heaters are driven, the non-
zero resistance of the ground (∼ 10mΩ) results in a raising of the ground voltage and the
wrong phase shifts being set. To solve this problem, all resistances were measured, and
extra classical pre-processing was used to calculate a set of adjusted voltages incorporating
the expected ground shift.
Finally, it was measured that the polarisation of the input light was being rotated at
105
(a) ✓ (c)
7000
6000
5000
Counts
⇡ 4000
(b) 3000
⇡/2 ⇡/2 2000
1000
⇡ 0
0 2 4 6 8 10
Voltage (V)
Figure 4.2: Characterisation of phase shifters. (a) Phases in MZIs can be characterised
by monitoring the countrate modulation when the voltage is sweeped. (b) For additional
phases, larger MZIs can be formed using four calibrated MZIs, with two acting as beam-
splitters and two as identity transformations. (c) Example of calibration curve for one
heater measuring the countrate vs voltage applied.
certain places through the circuit. This would result in non-ideal multi-photon interference
and so is filtered out via the use of fibre PBSs on the output channels.
4.4 Linear Optical Gates

Once calibrated, the device can be used as a testbed for linear optical quantum logic. As
a benchmark of device performance, we demonstrated a set of process tomography exper-
iments on standard linear optical gates. Then, going beyond previous demonstrations,
we implemented the first heralded logic gates in integrated photonics, performing a KLM
CNOT gate and a heralded Bell state generator.
4.4.1 Process tomography experiments
The simplest mapping of quantum logic to the universal device is to use each MZI and
associated phase shifts to implement arbitrary single qubit logic on a single photon in
two modes. In order to perform a process tomography experiment, a complete set of
initial states and measurements must be made. This can be achieved using a chain of
three MZIs, the first to prepare a state, the second to perform the gate and the third to
106
State θ φ Observable φ θ
|0i π 0 Z 0 π
π
|1i 0 0 X 0 2
π π π
|+i 2
0 Y 2 2
π
|−i 2
π
π π
| + ii 2 2
π 3π
| − ii 2 2
Table 4.1: On-chip phase shifts to prepare Pauli eigenstates and measure Pauli observ-
ables.
change the measurement basis (see Fig. 4.3). Using the three T3 (θ, φ) transformations,
process tomography was performed on the set of single qubit gates {X, Y, Z, H, T } using
a set of preparations and measurements consisting of the six Pauli eigenstates. The
phases required to prepare Pauli eigenstates and measure Pauli observables are displayed
in Table 4.1.
An estimate of the Choi-Jamiolkowski state ρest
E corresponding to the gates were found
using maximum-likelihood estimation (MLE). As described in Sec. 2.3.4, the theoretical

probability of measuring an outcome corresponding to the POVM effect Ej when the
state ρi is prepared is given by Prij (ρE ) = dTr[(ρ>
i ⊗ Ej )ρE ]. Therefore, if nij events are
registered in a total of Nij trials, the Likelihood function has the binomial form

Nij
L(nij |ρE ) = Prij (ρE )nij [1 − Prij (ρE )]Nij −nij . (4.14)
nij
The MLE estimate is then given by the maximum over the product of likelihoods for each
experiment labelled by i and j subject to physicality constraints
Y
ρest
E = max L(nij |ρE ) (4.15)
ρE
i,j
subject to: ρE ≥ 0, TrA (ρE ) = 1/d (4.16)
Unfortunately performing this maximisation is generally not computationally practical,

especially when moving to two or more qubits. To alleviate this problem we can make a
107
X Y
Figure 4.3: Process tomography for single qubit gates. Using three consecutive two-mode
unitaries, single qubits can be prepared, manipulated and measured. The real part of the
estimated process matrices χest for the single qubit Pauli and H gates are shown along
with both the real and imaginery part for T . Hollow bars represent the ideal values.
Figure adapted from [165]. Reprinted with permission from AAAS.
Gaussian approximation to Eq. (4.14)
n −N f − Pr (ρ )2 o
ij ij ij E
L(nij |ρE ) ≈ exp (4.17)
fij (1 − fij )
where fij = nij /Nij is the experimental frequency. Since the log of the likelihood has
the same maximum, we can simplify the problem to the minimisation of the negative
log-likelihood
X Nij fij − Prij (ρE ) 2

( )
ρest
E = min
ρE
i,j
fij (1 − fij )
subject to: ρE ≥ 0, TrA (ρE ) = 1/d (4.18)
We now have obtained an optimisation which can be phrased as a semidefinite program

(SDP), a standard form of convex optimisation problem for which many practical solvers
are available. Here we use the CSDP solver [172].
One problematic aspect of MLE is that if the direct linear inversion estimate of ρE is not
positive (as described in Sec. 2.3.4), then the estimate ρest
E will contain zero eigenvalues
108
Gate Fpro ||.||

X 0.997 ± 0.001 0.023 ± 0.001
Y 0.993 ± 0.001 0.031 ± 0.001
Z 0.994 ± 0.001 0.031 ± 0.001
H 0.995 ± 0.001 0.030 ± 0.001
T 0.978 ± 0.001 0.068 ± 0.001
Table 4.2: Process fidelities and diamond distances between ideal single qubit gates and
experimentally reconstructed gates.
[173]. Zero eigenvalues suggest that certain outcomes can never occur, which is not a
justified conclusion given a small sample of data. Furthermore, as can be seen from Eq.
(4.18), when fij = 0 or 1 the objective function is not well-defined. A statistically well-
motivated approach to deal with these problems is to use hedged MLE [174], where the
experimental frequencies are adjusted to
nij + β
fij = (4.19)
Nij + Kβ
where K is the number of outcomes in the measurement of fij and β is a hedging parameter
(we set β = 0.1 as suggested in [175]). The experimentally reconstructed process matrices
in the Pauli basis are shown in Fig. 4.3 and the process fidelities and diamond distances
to the ideal gates are displayed in Table 4.2. Errors in these quantities were estimated
by a boostrapping method, where the raw experimental counts were resampled 100 times
drawn from a distribution with means given by the experimental data and the statistics
of the resulting distributions analysed.
A standard test of linear optical quantum logic is the postselected CNOT gate from
Ref. [176]. This is a gate which, instead of using ancilla photons and heralding, performs
the correct evolution conditional upon the output state remaining within the computa-
tional subspace. The limitation of postselection is that with destructive single photon
detection the gate can only be confirmed to have worked by measuring the computational
photons so that no further processing of the qubits is possible. This approach is therefore
only scalable if highly challenging single photon non-demolition measurements can be
made. On the other hand, these gates demonstrate the basic building blocks of an LOQC
109
architecture, relying on phase stable linear optical networks and the quantum interference
of indistinguishable photons.
Fig. 4.4 shows the gate schematic and results. The CNOT gate is applied to two
photonic qubits (defined by modes 2,3 and 4,5) conditional upon finding a single photon
in each of the qubits at the output, which occurs ideally with a probability 1/9. The
transfer matrix for this gate is given by
√ 
1 2 0 0 0 0
√ 
 2 −1 0 0 0 0
 
 
 
1  0 0 1 1 1 0
Tpost =√   (4.20)
3
 0 0 1 0 −1 1 

 
0 1 −1 0 −1
 
 0
 
0 0 0 1 −1 −1
which are mapped to the following on-chip phases (see Fig. 4.1(b) for reference)
 
(0.00, 1.23) (θ12 , φ12 ) (0.00, 3.14) (θ14 , φ14 ) (0.00, 3.14)
 
(0.00, φ22 ) (θ23 , 1.23) (0.00, 1.57) (3.14, 3.14) 
 

 
R= (0.000, 1.23) . (4.21)
 
(0.00, 3.14) (3.14, 1.57)
 
 
 (0.00, φ44 ) (θ45 , 3.14) 
 
(0.00, 3.14)
where Rij = (θi,j , φi,j ). Phases {θ22 , φ23 } and {θ44 , φ45 } can be set to prepare arbitrary
pure states of qubits 1 and 2 respectively, and {θ12 , φ12 } and {θ14 , φ14 } can be set to
perform arbitrary projective measurements on qubits 1 and 2 respectively.
Process tomography was performed via the same method as before, this time preparing
and measuring states in a basis formed by the tensor product of Pauli eigenstates. This
results in a total of 1296 experimental frequencies (though only 324 unitary configurations
since four projectors can be measured simultaneously). The estimated process was found
to have a process fidelity of Fpro = 0.909±0.001, average gate fidelity Fave = 0.927±0.001
and diamond distance ||.|| = 0.115 ± 0.002. These are better than those previously
110
(a) (c)
(b)
Figure 4.4: Process tomography of postselected CNOT gate. (a) The postselected CNOT
gate works via the interference of two photons followed by measurement of a pattern
within the logical subspace which occurs 1/9 of the time. (b) Optical circuit for this
gate. Couplers represent two mode rotations and their colour represents their reflectivity
according to the key. Orange transformations are arbitrary two mode rotations which are
picked via Table 4.1 to perform process tomography. (c) Real part of the experimentally
estimated process matrix χest of the gate. Hollow bars represent ideal CNOT gate values.
reported[157, 158, 177] in bulk and integrated experiments specifically designed to perform
this gate. This demonstrates a high level of performance of the device along with its
versatility.
4.4.2 Heralded integrated gates
The key mechanism which enables LOQC is the effective interaction between photons
implemented indirectly via quantum interference and photon measurement. This is the
approach which is used in the original KLM CZ gate described in Sec. 3.6.1. Here
we implement a version of this gate with simplified NS circuits courtesy of Ralph et al.
[178]. This gate can then be used to implement a CNOT via the simple circuit identity
(1 ⊗ H)CZ(1 ⊗ H) = CNOT.
A full implementation of this gate would generally require eight optical modes, however
we can note that the |0i rail of the control qubit, c0 , does not need to couple to any others
and so can be implemented via an additional straight alignment waveguide on the chip.
111
Also, the final ancilla mode for the NS gate on the control qubit can be implemented
via simply changing the effective detector efficiency of that mode in post-processing.
Therefore, although full tomography as before is not possible, the gate’s action in the
computational basis can be demonstrated, the first demonstration of this form of photonic
logic in an integrated device.
The unitary transfer matrix required on our six mode circuit to perform the KLM
CNOT gate is given by
 
0.476 −0.622 −0.440 0.440 0 0
 
−0.622 −0.476 0 0 0.622 0 
 
 
−0.383 0.707 −0.383 −0.348
 
0 0.293
TKLM =


 (4.22)
 0.383 0 0.707 0.293 0.383 0.348 
 
0.622 −0.440 0.440
 
 0 0.476 0 
 
0.306 0 0.166 −0.166 0.306 −0.870
corresponding to a set of phases

 
(0.00, 0.99) (4.71, 1.57) (4.54, 4.95) (5.38, 1.79) (3.82, 0.00)
 
(0.00, 1.57) (1.57, 0.99) (2.19, 0.00) (4.71, 0.00)
 

 
R= (1.57, 2.23) .
 
(0.00, 1.57) (5.50, 0.00)
 
 
 (0.00, 3.14) (3.14, 0.00)
 
(0.00, 0.00)
Control and target qubits are implemented via the modes {0, 2}, {3, 4} respectively
(mode 0 representing the additional straight waveguide), and the heralding modes are
{1, 5}. One final experimental difficulty is the preparation of the |1111i state required at
the input (the two computational basis states along with the ancilla photons in modes
1,5). Restricting to the four-photon subspace, the state produced by the source is actually
∼ √1 (|1111i + |2200i + |0022i). A configuration of the input modes is therefore chosen
3
for which, in ideal operation, these extra terms in the input cannot give rise to a correct
output pattern.
112
(a) (c)
(b)
c0 c0
|1i |1i
c1 c1
t0 |0i
t1 t0
|1i t1
|1i
Figure 4.5: KLM CNOT gate. (a) Two ancilla photons are detected to herald a CNOT
acting upon two photonic qubits. (b) Circuit mapped onto the universal linear optical
circuit, colours represent the reflectivities of two mode rotations according to the key. The
control qubit is defined by modes c0 and c1 and target qubit t0 , t1 . (c) Logical truth table
measured in the computational basis. Hollow bars represent the ideal CNOT operation.
Fig. 3.3 shows the truth table obtained from the experiment. Each measurement took
around 5 hours to collect an average of 120 four-photon coincidences. The mean statistical
fidelity between the experimental and ideal probabilitiy distributions, averaged over the
four input states was found to be F̄s = 0.930 ± 0.003.
As described in Sec. 3.6, although the KLM scheme provides a technically scalable
architecture for LOQC it has been superseded by more efficient approaches such as those
based on MBQC. The main resources that these schemes require is a supply of entangled
photonic qubits. The most efficient circuit for producing heralded Bell states from single
photon states, which we will label B8 , is that of Zhang et al [179] which uses four single
photons and measures two to herald a Bell state with a probability3 of psucc = 3/16. Again
however, this circuit requires eight optical modes. Instead, we developed a new scheme for
Bell state generation using just six modes, which we will label B6 . Although less efficient,
succeeding with a probability of psucc = 2/27, this circuit has some interesting properties
3
This can be boosted to psucc = 1/4 with additional circuitry and distillation [180]
113
when compared to the scheme of [179]. These will be explored in Chapter 6.

The circuit schematic is shown in Fig. 4.6 and the transfer matrix for this gate is given
by
 
0.707 0.707 0 0 0 0
 
0.408 −0.408 −0.577 0.577 0 0
 

 
0.408 −0.408 0.289 + 0.289i −0.289 + 0.289i −0.408i −0.408i 
 
B6 =  
0.408 −0.408 0.289 − 0.289i −0.289 − 0.289i
 
0.408i 0.408i 
 
0.333 − 0.471i 0.333 − 0.471i 0.236 − 0.333i 0.236 − 0.333i
 
 0 0
 
0 0 0 0 0.707 0.707
corresponding to a set of phases

 
(0.00, 1.57) (0.00, 1.23) (0.00, 1.57) (0.00, 3.14) (0.00, 3.14)
 
(0.00, 3.14) (0.00, 1.57) (1.57, 1.23) (0.00, 3.14)
 

 
R= (0.00, 1.57) .
 
(0.00, 3.14) (0.00, 3.14)
 
 
 (0.00, 3.14) (0.00, 3.14)
 
(0.00, 3.14)
The qubits are defined by modes {1, 2} and {5, 6}. In order to verify the output of
the circuit, we would ideally perform tomography upon the heralded state. However,
separating the tomography measurements from the state generation is not possible for
the first qubit. The natural measurements that can be made are therefore of operators
Z ⊗ {X, Y, Z}, the output distributions of which are shown in Fig. 4.6c. In order to verify
entanglement, a measurement of the first qubit is needed in another basis.
To obtain statistics for an X measurement on qubit 1, we created a fibre Sagnac loop
between the output of modes 1 and 2 which injects the photon back into the chip with its
phase coherence maintained but with an X gate applied to the logical qubit. Due to the
configuration of the circuit, the photon then sees a 1/3 reflectivity MZI between modes
2 and 3, before a 1/2 reflectivity MZI between modes 1 and 2. This circuit therefore
114
(a) (c)
1
(b) (d) E+
2 E
3 |1i |1i
4 |1i |1i
5
Figure 4.6: Experimental Bell state generation. (a) four input photons are used to gener-
ate a bell pair via measurement of two photons. (b) Circuit for new Bell state generation
scheme. (c) Experimentally measured distributions at the output for Pauli measurements,
furthest right uses the Sagnac loop setup. (d) Sagnac loop setup to measure top qubit
in conjugate basis. The top photon is re-injected into the chip and, in two experiments,
measured in either in mode 2 E− or mode 1 E+ . Figure adapted from [165]. Reprinted
with permission from AAAS.
implements a single qubit POVM with the three elements, {E+ , E− , E1 } where
1 1 1
E+ = |0ih0| + |1ih1| + √ |0ih1| + |1ih0| (4.23)
6 2 2 3
1 1 1
E− = |0ih0| + |1ih1| − √ |0ih1| + |1ih0| (4.24)
6 2 2 3
2
E1 = |0ih0| (4.25)
3
corresponding to detection in modes {1, 2, 3 or 4} respectively. Since it is not possible

to detect in modes 3 and 4 (these are inputs), we restrict to events where the photon is
detected in modes 1 and 2, which occur with relative probabilities given by
Tr(E+ ρ) Tr(E− ρ)
Pr+ = , Pr− = (4.26)
Tr(E+ ρ) + Tr(E− ρ) Tr(E+ ρ) + Tr(E− ρ)
115
allowing for the estimation of the value

√
3hXi
Pr+ − Pr− = (4.27)
h0|ρ|0i + 3h1|ρ|1i
such that, with additional knowledge from the Z measurement, we are able to estimate
the expectation of X.
Since the experiment would require both input and detection in mode 1, we implement
it in two stages, first inputting in mode 1 and detecting four-fold coincidences in modes
{2,3,4,5} and {2,3,4,6}, then inputting in mode 2 and detecting four-fold coincidences in
modes {1,3,4,5} and {1,3,4,6} (note that by inputting in mode 2, the state would pick
3π
up a relative π phase shift so this is offset by setting θ3,5 = 2
). The two data sets are
then normalised with respect to each other by the ratio of total counts in each experiment.
Measurements with the Sagnac loop were taken for 60 hours, whilst measurements without
were taken for 16 hours, collecting in total around 200 four-fold coincidences for each
experiment.
Estimating the expectations hZ ⊗ Zi and hX ⊗ Xi allows the entanglement of the
generated state to be verified. Since
1
Fmin = (Tr[ρZ ⊗ Z] + Tr[ρX ⊗ X]) (4.28)
2
= hφ+ |ρ|φ+ i − hψ − |ρ|ψ − i (4.29)
= F(ρ, |φ+ i) − F(ρ, |ψ − i) (4.30)
≤ F(ρ, |φ+ i) (4.31)
and any state with F(ρ, |φ+ i) > 1/2 is entangled [181], Fmin > 1/2 is an entanglement
witness. We find that Fmin = 0.673 ± 0.031 and therefore certify to a high confidence that
the experiment does indeed generate heralded entanglement.
116
4.4.3 Characterisation of Linear Optical Networks
Comparing the four-photon heralded gates and the single and two photon experiments,
it is clear that the quality of the quantum logic is significantly diminished in the higher
photon number experiments. One unique property of photonic gates however is that they
can be characterised independently of the photonic qubits they act upon and, in principle,
with only classical states of light. Full process tomography of a linear optical gate, though
necessarily inefficient, can be performed using multimode coherent states as probes and
homodyne measurement [182, 183]. However, if the assumption is made that the linear
optical network perfectly preserves coherence and applies the same transfer matrix M to
all relevant optical modes, then this problem can be reduced to the far simpler task of
estimating this transfer matrix.
The transfer matrix can be estimated in a number of ways. It is possible to do so
using coherent state inputs and intensity measurements [184]. These approaches require
additional interferometry before or after the gate however. Here we use an approach which
is motivated by current experimental regimes, where light is generated and detected off-
chip and all interferometry is performed within the integrated circuit itself. Using single
photons and pairs of photons at the input ports the transfer matrix can be reconstructed.
The protocol we use, based on Ref. [185], first estimates the absolute value of each
element of M . In principle |Mij |2 can be directly estimated by inputting a single photon in
mode j and measuring the probability of it exiting from mode i. However, this estimate
includes several sources of loss which we may wish to separate. Firstly, there may be
input-mode dependent loss arising from fluctuations in intensity in the probe state or
coupling into the device. Similarly, there may be output-mode dependent loss arising
from coupling out of the device and differing detector efficiencies. Finally, there is the
possibility of path-dependent loss through the circuit itself.
If the path-dependent loss is negligible in comparison to the input and output losses,
i.e. the LON is well-described by a transformation proportional to a unitary U , then
it is possible to infer the absolute values |Uij | from the matrix of single photon counts.
Sinkhorn’s theorem [186] states that any nonnegative matrix can be decomposed as the
117
4
product of two diagonal matrices and a doubly stochastic matrix
|M |2 = Lout |U |2 Lin (4.32)
where Lin and Lout are diagonal matrices quantifying the input and ouput losses respec-
tively. The Sinkhorn-Knopp algorithm [187], which consists of iteratively normalising the
columns and rows of |M |2 until convergence, can then be used to obtain |U | from the
measured |M |.
However, when using circuits where different paths experience differing circuit depths,
such as in the Reck scheme, this assumption is likely not to be valid. Therefore, we
must attempt to characterise the input and output losses independently. This is done by
estimating the relative detector efficiencies by the intensities measured when configuring
the device to send all light to each detector in turn. The input losses can be estimated
by monitoring the single photon generation rate, e.g. via beamsplitting part of the in-
put beam directly to detectors. Adjusting the single photon count matrix to take these
measured relative losses into account and then renormalising the result by a single scalar
equal to the largest total countrate across all experiments then provides an estimate of
|M |.
In order to learn the phases {αij } of the matrix elements Mij = |Mij |eiαij , pairs of
photons can be used in a HOM-type configuration. Since the input and output phases of
a LON are only defined relative to any pre- or post-interferometry, these phases can be
set such that the first column and row of M are real without loss of generality. Inputting
and detecting one photon in the first mode results in experimental probabilities given by
Pr(1i|1j) = |M11 |2 |Mij |2 + |Mi1 |2 |M1j |2 + 2|M11 ||Mij ||Mi1 ||M1j | cos(αij ) (4.33)
where the photons are input in modes 1 and j and detected in modes 1 and i. Performing
HOM dips in this configuration then allows |αij | to be estimated from the visibility V of
4
A matrix for which all columns and rows sum to 1
118
the dip (or anti-dip) and the known |M | matrix via
h −(|M |2 |M |2 + |M |2 |M |2 )V(1i|1j) i
11 ij i1 1j
|αij | = arccos . (4.34)
2|M11 ||Mij ||Mi1 ||M1j |
Finally, the sign of the phases can be deduced from a further set of two-photon experi-
ments. For an experiment in which photons are input in modes j, j 0 and detected in i, i0
the quantity |αi0 j 0 − αij 0 − αi0 j + αij | can be estimated. The signs can then be calculated
as

sgn(αij ) = sgn ||αi0 j 0 − αij 0 − αi0 j + αij | − |αi0 j 0 − αij 0 − αi0 j − |αij ||| (4.35)

− ||αi0 j 0 − αij 0 − αi0 j + αij | − |αi0 j 0 − αij 0 − αi0 j + |αij ||| . (4.36)
We are free to choose sgn(α22 ) = +1 and then the set of experiments which measure
V(2i|12),V(12|2j) and V(2i|2j) where i, j run [3, m] are sufficient. The full procedure
requires m2 single photon measurements (for which only m experimental configurations are
required) and 2(m − 2)(m − 1) two-photon measurements (for which 2m − 3 experimental
configurations are required).
We performed this characterisation protocol on the three two-qubit gates described in
the above (these reconstructions can be found in appendix A.1). To compare these we
use the circuit fidelity [170]
2
Tr(U † Mest )
Fc = q (4.37)
†
mTr(Mest Mest )
to assess the distance between the desired unitary transformation U and implemented
transfer matrix. This can be thought of as the fidelity between the state produced when
light is input in a single port of the device and the target state, averaged over all ports.
We find Fc = 0.996 ± 0.004, 0.969 ± 0.018 and 0.988 ± 0.040 for the Bell state generator,
the KLM CNOT gate and the postselected CNOT gate respectively. Using these recon-
structions, it is possible to estimate the logical error introduced via the optical circuitry.
Considering for instance, the Bell sate generator circuit, assuming only the coherent error
119
in the transfer matrix

h11|23 U(Mest )|011110i ∝ |ψM i (4.38)
we find a state fidelity of F(|ψM i, |φ+ i) = 0.988 and a concurrence C(|ψM i) = 0.998.
This shows that the circuit itself is capable of producing highly entangled states if acting
upon ideal input states. We further note that all of these circuits were obtained via the
individual phase shifter calibrations and performing further optimisation on the applied
voltages is likely to increase these fidelities [188]. This highlights that the logical error
observed in the four-photon experiments can be largely attributed to a reduction in the
quality of the quantum interference, as well as errors introduced due to accidental counts
and imperfect detection. For instance the HOM visibility between off-pair photons is
∼ 0.9, which we will see in Chapter 6 already limits the concurrence of the Bell state
generator to ∼ 0.8.
4.5 Discussion
In this chapter we have demonstrated the experimental operation of a universal linear
optical circuit and its use as a testbed for photonic quantum logic schemes, including new
entanglement generation schemes and the first demonstration of measurement-induced
nonlinearity in an integrated platform. The versatility of the device allows us to switch
between all these experiments in seconds, providing a platform for rapidly prototyping
new and existing schemes. We demonstrated high performance of the linear optical trans-
formation over different configurations via transfer matrix characterisation, showing the
potential for the device to produce highly entangled heralded states and identifying im-
perfect quantum interference and detection as the main sources of experimental error.
Statement of Work
Much of the work reported in this chapter was published in Carolan et. al. Science,
349, 711 (2015) [165]. The universal linear optical device was conceived by A. Laing
and fabricated by N. Matsuda and colleagues at NTT. Characterisation, calibration and
120
data taking were largely performed by J. Carolan and C. Harrold with input from N.
Matsuda and myself. The Process tomography experiments were designed and analysed
by myself. The KLM gate experiment was designed by myself and analysed by C. Harrold
and J. Carolan. The Bell state generator circuit was conceived by myself, the experiment
designed by myself, J. Carolan and J. Matthews and analysed by myself. The transfer
matrix characterisation methods were initially conceived by A. Laing, developed by J.
Carolan and myself and analysed by myself.
121
Chapter 5
Simulating Vibrational Quantum

Dynamics of Molecules with
Integrated Photonics
5.1 Introduction
In the previous chapter we demonstrated a powerful new capability for quantum photon-
ics, universal linear optical circuits. We went on to demonstrate their use as a testbed for
experimentally investigating linear optical logic gates. Here we propose and experimen-
tally demonstrate another application for such devices as a new platform for quantum
simulations.
Computer simulations that model physical systems have revolutionised modern sci-
ence, technology and engineering. Before the advancement of digital computers, early
analog computers that mapped the controllable equations of motion of one system to
mimic another helped solve many real world problems from simulating ballistics and
rocket science to modelling tidal systems and power networks [189].
Analog quantum computers are devices which can directly simulate the quantum dy-
namics of systems in nature via tuning an equivalent Hamiltonian of controllable quantum
systems in the lab. This analogical approach aims to maximise the complexity and utility
122
5. Simulating Vibrational Quantum Dynamics of Molecules with Integrated Photonics
of naturally evolving quantum systems without the vast challenges and overheads required
to engineer them into individually addressable qubits and gate sequences. Although nec-
essarily limited in scale and versatility, such devices may find uses in the near term as well
as representing an important milestone on the road to universal digital quantum computa-
tion. Here we demonstrate programmable integrated quantum photonics as a simulation
platform for models of vibrational molecular quantum dynamics using the analogies of
optical modes in integrated waveguides for vibrational modes in molecules and photons
for quantised vibrational excitations.
Progress in modelling, controlling, and observing the ultrafast dynamics of molecules,
has revealed the underpinning mechanism of interference among vibrational modes in im-
portant behaviour such as bond selective chemistry [190]. Increasing levels of control over
vibrational wavepackets has allowed selective dissociation governed by a single quantum
of vibrational energy [191], manipulation of individual molecules at ambient conditions
[192], preparations of coherent superpositions on sub-femtosecond timescales [193], and
single vibrational states of ultracold molecules [194, 195]. Molecular dynamics are now
observable on their ultrafast intrinsic timescale [196, 197].
The prospect of more sophisticated control with quantum states of light and larger
molecules brings additional challenges for modelling however. Evolving a multi-excitation
state across many vibrational modes described by independent quantum harmonic oscilla-
tors is equivalent to the Boson sampling model and therefore computationally inefficient.
More sophisticated molecular models, for example with anharmonic corrections to the
potentials, are also likely to be computationally complex.
In this chapter we first demonstrate that integrated photonics makes a platform for
the efficient simulation of vibrational molecular dynamics in the harmonic approximation.
We show how to go beyond this model by simulating thermal relaxation using additional
ancilla optical modes and anharmonicities using measurement-induced nonlinearities. Fi-
nally, we use adaptive feedback control with our simulator in a closed loop for finding
optimised initial states for the dissociation of NH3 , demonstrating the setup as a testbed
for control theory.
123
5.2 Molecular Vibrational Dynamics in the Harmonic

Approximation
A quantum mechanical description of molecular vibrations begins with the Born-Op-
penheimer approximation (BOA) [198], the assumption that the nuclear and electronic
dynamics can be separated due to their difference in masses and therefore timescales. The
full molecular Hamiltonian can be written as
Ĥmol = Tn + Te + Vn,n + Vn,e + Ve,e (5.1)
where T refers to kinetic energy, V potential energy, n the nuclei and e the electrons. The
BOA separates the wavefunction into a nuclear part χ(R), where R is a vector of the
nuclear positions and an electronic part ψ(r, R) where r is a vector of the electron posi-
tions. An adiabatic approximation is then made whereby the electronic time-independent
Schrödinger equation is solved for fixed configurations of the nuclei R resulting in the
electronic potential energy surface (PES) Ee (R). The nuclear Schrödinger equation then
reduces to

Tn + Ee (R) |χ(R)i = Etotal |χ(R)i. (5.2)
To solve this equation, the dependence of the PES is first expressed in mass-weighted
√
cartesian coordinates x̃i = mi xi where i labels both the nucleus ∈ 1, ..., N and coordinate
∈ {x, y, z}. The PES is then expressed as a Taylor series around the nuclear equilibrium
structure R̃ = 0,
3N 3N
X ∂Ee (R̃) 1 X ∂ 2 Ee (R̃)
Ee (R̃) = Ee (0) + x̃i + x̃i x̃j + ... (5.3)
i=1
∂ x̃i R̃=0 2 i,j=1 ∂ x̃i ∂ x̃j R̃=0
Since the energy gradients are zero at the equilibrium (minimum energy) structure, setting
Ee (0) = 0 and using the harmonic approximation in which we truncate to second order,
124
the Hamiltonian in Eq. 5.2 becomes
3N 3N
1 X ∂2 1 X ∂ 2 Ee (R̃)
− + x̃i x̃j . (5.4)
2 i=1 ∂ x̃2i 2 i,j=1 ∂ x̃i ∂ x̃j R̃=0

∂ 2 Ee (R̃)
The Hessian matrix with elements Hij = ∂ x̃i ∂ x̃j
can be found using standard
R̃=0
numerical electronic structure methods from quantum chemistry for large molecules with
hundreds of vibrational modes [199]. Diagonalising the Hessian produces a set of 3N
mass-weighted normal coordinates {Qk } for which the Hamiltonian reduces to that of 3N
independent quantum harmonic oscillators
X ∂2
2
+ ωk Q2k . (5.5)
k
∂Q k
The unitary matrix Q which diagonalises H has six eigenvectors with zero eigenvalues
corresponding to molecular translation and rotation. The remaining 3N − 6 eigenvectors
correspond to the normal modes of the molecule and their eigenvalue the vibrational fre-
quency ωi . Finally, we can express Eq. (5.5) with creation operators that create quantised
vibrational excitations in the vibrational modes; Ĥ = i ~ωi a†i ai (omitting zero point en-
P
ergy). We can finally then map this system of non-interacting bosonic modes to our linear
optical device.
The spatial localisation of vibrational energy is important for understanding many
molecular phenomena such as energy transport and dissociation. We therefore consider
a basis transformation, a†i →
P L †
j Uji aj , from the collective normal modes to a set of
modes which are localised around a single atomic site or chemical bond. These localised
modes are found numerically using a Pipek-Mezey type method [200]. The kinetic energy
contribution from a nucleus p in a normal mode i is given by Kp,i = α=x,y,z Q2pα,i . The
P
localised basis is then found by maximising the sum of squares of these kinetic energy
contributions over all unitary transformations Q0kα,i = j UjiL Qpα,j
P
h 3N −6 X
3N i
0
X
max Kp,i2 . (5.6)
UL
i k
125
Once the localised basis is found, dynamics in this basis can then be tracked by
L † L
= i ~ωi UkiL L
P P
simulating the Hamiltonian ĤL = k,j Hkj ak aj where Hkj Uji . Since the
matrix H L can be exponentiated efficiently in m (roughly in time O(m3 )), reconfiguring
a universal linear optical device to implement the transfer matrix U (ti ) = exp (−iH L ti /~)
for a series of timesteps {ti } allows the efficient simulation of the Hamiltonian ĤL upon
arbitrary multi-mode vibrational input states.
5.3 Experimental Procedure

The procedure for collecting experimental data begins with a set of experiments to measure
detector efficiencies. A heralded single photon is input to the chip which is set to route
the photon to each output mode. Collecting detection events for a fixed period of time for
each output mode then allows us to estimate the relative detector efficiencies by dividing
all the count rates by the largest count rate. The input state is then prepared and the
experiment is performed by logging all coincidental detections at the output with the
integrated chip set to perform the set of unitaries required for the simulation. These
experimental counts are then corrected using the measured efficiencies by dividing the
total counts in a detection pattern by the product of the relative detector efficiencies
of the detectors in that pattern. These corrected counts are then used to estimate the
expectation value of projecting the output state onto a detection pattern as the frequency
of corrected counts in that pattern to the total number of corrected counts.
Finally, the statistical noise in the estimates of these probabilities resulting from the
finite counting statistics is estimated via a bootstrapping approach, where 1000 sets of
data are resampled from Poissonian distributions with means given by the experimentally
measured counts, and the standard deviation of the probabilities estimated from these
sets are used to provide error bars on plots.
Many of the experiments require the ability to resolve multiple occupancy in the out-
put modes. Although number-resolving detectors were not available in this experiment
we are able to circumvent this requirement by performing probabilistic number resolving
between one and two photons using auxiliary fibre beamsplitters (FBS) and detectors.
126
This is achieved by inserting a FBS at each of the relevant output fibres and connecting
both output modes to detectors; when there are two photons incident at the FBS there
is then a probability that both the detectors generate a signal. For a beamsplitter with
reflectivity r connected to detectors with efficiencies η1 and η2 , the probability of detecting
a coincidental detection given a two-photon input is 2r(1 − r)η1 η2 . Performing detector
efficiency measurements as described above, the count rates in each detector will be pro-
portional to their effective efficiencies, rη1 and (1 − r)η2 . As before, relative efficiencies
are then calculated by dividing by the maximum count rate and all coincidence counts are
re-normalised by the product of these relative effective efficiencies. A final correction is
then applied to account for multiple occupancy, for instance in a two photon experiment
counts corresponding to the same output mode are halved, whereas counts corresponding
to different output modes are averaged over the four possible ways of detecting such an
output.
5.4 Simulating Four Atom Molecules

As a first example we consider Thioformaldehyde (H2 CS), a molecule with four atoms
and six vibrational modes which is a prototype for many spectroscopic experiments [201].
For this, and all four atom molecules below, the equilibrium geometries and Hessian ma-
trices were computed using CC2, an approximate coupled-cluster correlated wavefunction
method, and the TVZPP atomic orbital basis sets [199, 202, 203].
The computed wavenumbers of the normal modes of H2 CS are
{3195.45, 3102.74, 1501.77, 1061.45, 1017.08, 1008.04}cm−1 (5.7)
127
and the transformation between the normal modes and localised modes is found to be
 
0.563 0.572 0.332 0.251 0. 0.427
 
 −0.563 0.572 0.332 0.251 0. −0.427 
 
 
 −0.427 −0.392 0.572
 
0.138 0. 0.564 
UL =   (5.8)
−0.196 −0.355 0.914
 
 0. 0. 0. 
 
 
 0. 0. 0. 0. 1. 0. 
 
−0.427 0.392 −0.572 −0.138 0. 0.564
Fig. 5.1(a) visualises these normal modes as well as the localised modes which consist
of two CH stretch modes, two CH bend modes, a wagging mode (in which both H atoms
move out of the plane) and a CS stretch mode which are mapped in this order to the
six waveguide modes of the photonic chip. We note that since the wagging mode is not
coupled to the other modes in UL , the resulting system is effectively five mode.
Injecting the two-mode squeezed vacuum state produced by one SPDC source into the
two waveguides corresponding to the CH stretch modes, we track the combined probabil-
ity of detecting all excitations in these CH stretch modes, in the CH bend modes and split
between these four for the two- and four-photon subspaces. The left panel of Fig. 5.1(b)
shows this evolution restricted to the two-photon subspace, we observe dynamics in which
both excitations oscillate between stretches and bends via the intermediate states where
one is in each. The L1 distance between the theoretical and experimentally obtained dis-
tributions, averaged over all timesteps is D̄ = 0.064 ± 0.028. The right panel of Fig. 5.1(b)
displays the evolution of the four quanta term, where now each of the stretch modes are
initially doubly occupied, again combining the probabilities of detecting all excitations in
stretches, bends and combinations for measurement patterns with up to two excitations
per mode. We see a similar oscillatory behaviour between stretches and bends, but with
an apparent suppression of the probabilities attributable to the state extending into the
combinatorially growing space of multiple excitations. The full four photon distribution
has a distance D̄ = 0.159 ± 0.074 to the ideal distribution.
One of the features of this analog simulator is that time is a programmable parameter.
128
(a)
1. Construct molecular model 2. Transformation to basis of
H2 CS from vibrational eigensystem local vibrational modes
0.8 t
0.6 Local CH stretch mode spectrum
Intensity
0.4
×× 3. Map vibrational evolution
0.2 × ×
0.0 ×× to time dependent transfer
1000 2000 3000 matrix
Wavenumber (cm-1 )
+
4. Implement discrete instance of
5. Collect photon statistics at each transfer matrix on universally
timestep to track evolution reconfigurable photonic chip
(b) (c)
Stretches Bends Stretch & Bend
1.0 1.0 1.0 1.0
occupation
occupation
occupation
occupation
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
Prob ofProb
Prob ofProb
Prob ofProb
Prob ofProb
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0
0 10 20 30 40 0 10 20 30 40 0 5 10 15 20 25 30 0 50 100 150 200 250 300
Time (fs) Time (fs) Time (fs) Time (fs)
Figure 5.1: Simulating the vibrational dynamics of molecules in the harmonic approxima-
tion. (a) Steps for implementing a simulation using H2 CS as an example. (b) Evolution
of a vibrational two-mode squeezed vacuum state in the localised CH stretch modes of
H2 CS. Left panel displays the theoretical and experimentally estimated total probability
within the two-excitation subspace of measuring both excitations in the CH stretch modes
(black), CH bend modes (blue) and a combination of both (grey). Right panel shows the
same for the four-excitation subspace with up to two excitations in each mode. (c) Short
and long timescale evolution of a single vibrational excitation in a CH stretch mode of
H2 CS. Measuring the probability of finding the excitation in the stretch mode by taking
the mean (blue boxes) of measurements taken in narrow windows (left panel) spread over
a longer timescale (right panel).
This allows us to probe dynamics at arbitrary evolution times. In Fig. 5.1(c) we provide
an example of this by observing the probability for detecting a localised excitation in a CH
stretch vibration given it is prepared in this mode at t = 0. As shown in Fig. 5.1(a), this
stretch mode is a superposition of a set of lower and higher frequency components. Using
the source to produce heralded single photons and measuring time steps at an increased
rate in narrow windows across a longer evolution period we are able to observe both the
high frequency behaviour within each window as well as the low frequency periodicity
129
1.0 1.0
(a) P4 (b) HNCO
occupation
occupation
0.8 0.8
0.6 0.6
Prob ofProb
Prob ofProb
0.8 0.8
0.6 0.4 0.6 × 0.4
Intensity
Intensity
× ×
0.4
0.2
0.4 ××
× 0.2
0.2 × × × 0.2 ×
0.0 × × 0.0
0 20 40 60 80 100
0.0 ×× 1000× 2000
×
3000
0.0
400 500 600 0 10 20 30 40
Wavenumber (cm-1 ) Time (fs) Wavenumber (cm-1 ) Time (fs)
1.0 1.0
(c) SO3 (d) HFHF
occupation
occupation
0.8 0.8
0.6 0.6
Prob ofProb
Prob ofProb
0.8 0.8
0.6 0.4 0.6 × 0.4
Intensity
Intensity
0.4
× 0.4 ××
× × 0.2 0.2
0.2 × 0.2 × × ×
× × 0.0 × 0.0
0.0 ×× × 0 20 40 60 80 100
0.0 × ×× 0 10 20 30 40
500 750 1000 1250 0 1000 2000 3000 4000
Wavenumber (cm-1 ) Time (fs) Wavenumber (cm-1 ) Time (fs)
1.0 1.0
(e) N4
occupation
occupation
0.8 0.8
0.6 0.6
Prob ofProb
Prob ofProb
0.8 0.3
0.6 0.4 0.4
Intensity
Intensity
0.2 ×
0.4 × ××
0.2 × ×
0.2 0.1 ×× × ××× ××× 0.2
×
0.0 ×× ×× × × 0.0 0××× ××××× ××× × × 0.0
500 750 1000 1250 0 20 40 60 80 100 1000 2000 3000 0 20 40 60 80 100
Wavenumber (cm ) -1 Time (fs) Wavenumber (cm-1 ) Time (fs)
Figure 5.2: Evolution of localised vibrational excitations in four atom molecules. (a-d)
Evolution of a single excitation in P4 ,SO3 ,HNCO and HFHF between a stretch mode
(black) and another coupled local mode (blue). The modes involved are represented
diagramatically with arrows as well as by their spectral intensities alongside the theoretical
evolution and experimental data. (e) Evolution of one and two-vibrational excitations in
N4 . Left, the relevant single excitation spectrum and data. Right, the relevant two
excitation spectrum and data. In all plots error bars resulting from shot-noise in the
experimental relative frequencies are too small to see.
by tracking the average probability of each window. The mean distance across these
experiments is D̄ = 0.014 ± 0.006. Additional one, two and three excitation data can be
found in Appendix B.2.
Since general four atom molecules have 3×4−6 = 6 vibrational modes1 , our simulator
can explore the dynamics of any such system. To demonstrate this we simulated the
coherent dynamics of a series of four atom molecules which represent a wide range of
molecular structures and energy scales, P4 , SO3 , HNCO, HFHF and N4 . Figs. 5.2(a-d)
show the time evolution of a single excitation initially prepared in a local stretch mode.
The change in the occupation probability to a second, spectrally overlapped (‘coupled’)
local mode is plotted. We observe dynamics with varying characteristic times governed
1
Except for linear molecules which have 3N − 5 vibrational modes, as they have one fewer rotational
mode.
130
by the vibrational spectra of the molecules. Due to the geometry and bonding structure
of P4 , it has the longest period oscillations between opposing stretches with SO3 showing
similar stretch coupling on shorter timescales. HNCO and HFHF in contrast display faster
dynamics with increased mode coupling between hydrogen bond stretches and bends. The
average L1 distance over all these experiments is D̄ = 0.022 ± 0.007. Fig. 5.2(h) shows
the dynamics of both a single excitation and two excitations initially prepared in stretch
modes of N4 . The additional structure in the vibrational spectrum and the introduction
of multi-photon quantum interference result in a more complex time dependence of the
detection probabilities.
The vibrational spectra, UL and additional data for these four atom molecules can be
found in Appendix B.
5.5 Energy transfer and dephasing in NMA

The flow of vibrational energy in molecules is a fundamental process for chemical reac-
tion rates [204] and functionality in biomolecules [205], while control of intramolecular
energy transfer could lead to new nanotechnologies such as molecular electronics [206].
The vibrational quantum dynamics of a molecule within an environment can be described
by the interplay of coherent unitary evolution and incoherent dephasing resulting from
random fluctuations of the vibrational frequencies - a process referred to as spectral dif-
fusion. N-methylacetamide (NMA) is the simplest molecular model (Fig. 5.3(a)) of the
peptide bond in proteins, where quantum coherence may play a role in energy transfer
[207]. In this section, we simulate a model for intramolecular energy transport in NMA
in the presence of dephasing.
2
We consider a subspace spanning 6 (of its 30) backbone vibrational modes (see
Appendix B for full details), which support a basis of approximately localised vibrations,
including two rocking modes (curved arrows in Fig. 5.3) and two stretch modes (straight
arrows in Fig. 5.3). These six local modes are mapped to our photonic chip, such that
2
For this case density functional theory with the B3LYP functional and a split-valence atomic orbital
basis set was used to calculate the Hessian
131
(a) NMA (b)

1.0
occupation
0.8
0.6
Prob ofProb
0.8
×× 0.4
Intensity
0.6 ×
0.4 × 0.2
0.2
× × ×
0.0 ×
600 800
× ×
1000
× ×
1200
0.0
0 500 1000 1500
-1
Wavenumber (cm ) Time (fs)
(c) 1.0 (d) 1.0
Prob prob
occupation
0.8 0.8
Total anti-bunched
0.6 0.6
Prob
0.4 0.4
Prob of
0.2 0.2
0.0 0.0
0 200 400 600 800 1000 1200 50 100 150 200 250
Time (fs) Time (fs)
Figure 5.3: Quantum energy transfer and dephasing in NMA. (a) A six-mode vibrational
subspace of the NMA molecule is considered with the spectral components of three lo-
calised modes (used in (b)) shown in the right panel. (b) Energy transfer of a single local
vibrational excitation. Simulation of a single excitation initially in a local rocking mode
at one end (black) and its transfer to local modes at the opposite ends (blue and grey)
when subject to a dephasing channel with T2 = .53ps. Solid lines represent theory and
points represent experimentally estimated probabilties. (c) Quantum interference and
transport of multiple vibrational excitations. Evolution of a two-excitation state initially
in separate local modes (black) and its probability to be found bunched in the NH stretch
mode for indistinguishable bosons (solid blue) and under the same dephasing channel.
Also included is the theory curve for distinguishable particles (dashed blue). (d) The
total probability for measuring a fully anti-bunched state (solid black) and the same for
an initial state of three excitations (dot-dashed black) initialised in the modes from (b).
the coherent time evolution of an excitation can be tracked, following the same procedure
as described in the previous section.
To simulate uniform dephasing across the six vibrational modes we consider an ad hoc
model described by the Lindblad master equation
5
Γ Xh 1 i
ρ̇ = L(ρ) = −i[Ĥ, ρ] + Ẑ(k)ρẐ † (k) − {Ẑ † (k)Ẑ(k), ρ} (5.9)
6 k=1 2
exp(i2πjk/6)a†j aj , i.e. Z(k) are six dimensional Heisenberg-Weyl ma-

P
where Ẑ(k) = j
trices and Γ−1 , usually referred to as the T2 time, describes the characteristic coherence
time between the modes. Solving this equation results in the channel
132
Mk (t)e−iĤt/~ ρ(0)eiĤt/~ Mk† (t) with time-dependent Kraus operators

P
ρ(t) = k
6
r
1 + 5e−Γt X †
M0 (t) = aj aj
6 j=1
6
r
1 − e−Γt X
Mk>0 (t) = exp{i2πjk/6}a†j aj (5.10)
6 j=1
Including the change of basis to the local vibrational modes, we can therefore simulate
these dephasing dynamics for a set of vibrational modes with arbitrary T2 times by sta-
tistical averaging the expectation values obtained from experiments implementing the
transfer matrices U (t, j) = UL Z(j)e−iHt/~ UL† over j.
Using a single photon, we simulated the probability for a single excitation initialised
in a local rocking mode at one end of the molecule to be transferred to two localised
modes (a rocking mode and a C-C stretch mode) at the opposite end of the molecule, The
experimental results shown in Fig. 5.3b show dynamics that are initially oscillatory, with
vibrational energy transfer between the rocking modes at either end of the molecule via
an intermediate C-C stretch. The increasing effect of the suppression of coherences from
dephasing results in evolution toward a steady state with reduced peak probability for
the excitation to be found in these localised modes. The T2 time constant of coherence
decay used is 0.53ps, though any time constant can be simulated by changing only the
post-processing of data.
Generalising the simulation for initial states of multiple excitations allows us to inves-
tigate the combination of dephasing and quantum interference for multi-excitation energy
transport. Injecting one photon into the waveguide corresponding to the rocking mode
and another in the waveguide corresponding to the C-C stretch mode, which are each
localised at opposite ends of the NMA molecule (black arrows in Fig. 5.3(c)), we sim-
ulated the probabilities for finding the excitations in this state, and the state in which
both excitations bunch in an NH stretch mode (double blue arrows in Fig. 5.3(c)). The
results in Fig. 5.3(c)) show more complex oscillatory transfer between these separated
and localised states, that again tends toward a steady state. However, after full dephasing
has occurred, the probability for two excitations to be bunched in the NH stretch mode
133
is twice as high for excitations that behave as indistinguishable bosonic particles than for
excitations that behave as distinguishable classical particles whose theoretical probability
is also plotted.
For a given molecule, the probability that no bunching occurs (multiple excitations
are not localised) generally decreases as the number of excitations increases [208]3 . In
Fig. 5.3(d) the probability for the subspace of no-bunching events is simulated for two and
three excitations under fully coherent dynamics. The initial state for the two-excitation
evolution is the same as in the previous example; the initial state for the three-excitation
evolution comprises an excitation in each of the local modes shown in Fig. 5.3(b).
The average distances across all single, two and three excitation distributions in these
examples are given by 0.017 ± 0.005, 0.046 ± 0.014 and 0.139 ± 0.068 respectively.
5.6 Vibrational relaxation in H2O

Although our model of a closed system of uncoupled harmonic oscillators accounts for
much of the physics of molecular vibrations, we now consider extensions to better describe
realistic systems by including energy dissipation and potential anharmonicities.
When a vibrational state is excited in a molecule, exchange of this energy via intra- and
inter-molecular coupling to other degrees of freedom will eventually lead to thermalisation
- a process known as vibrational relaxation. The vibrational relaxation pathways of H2 O
remain an area of current investigation [209–211].
Here we simulate the relaxation of H2 O via an amplitude damping model, depicted in
Fig. 5.4a. Consider the Lindblad master equation
X 1 1
ρ̇ = L(ρ) = − i[Ĥ, ρ] + Γ1 |3ihj|ρ|jih3| − {|jihj|, ρ} + Γ2 |0ih3|ρ|3ih0| − {|3ih3|, ρ}
j=1,2
2 2
X 1
+ Γ3 |0ihj|ρ|jih0| − {|jihj|, ρ} (5.11)
j=1,2
2
Where |ji ≡ a†j |0i represent a single excitation in one of the three vibrational modes
3 m m+n−1

Averaged over all possible unitaries this probability goes as n / n for n excitations in m
modes.
134
1 1.0
(a) H2 O 8580cm
1 (c)
8452cm
0.8
occupation
1 3 0.6
Prob ofProb
1.0 × 6292cm 1 0.4
Intensity
0.8
0.6 0.2
0.4 ××
0.2 2
0.0 × ×4000
× 0.0
2000 3000 0 200 400 600 800 1000
1
-1
Wavenumber (cm ) 4665cm Time (fs)
1.0
(b)
occupation
0.8
0.6
Prob ofProb
0.4
0.2
Unitary dilation and 0.0

statistical averaging 0 200 400 600 800 1000
Time (fs)
Figure 5.4: Vibrational relaxation in H2 O. (a) Energy level diagram for single excitation
harmonic levels and ground state of H2 O along with the spectral components of the
two local OH stretch modes (black and grey) and the symmetric bend normal mode
(blue). (b) The open system dynamics resulting from a vibrational relaxation process
can be simulated by statistical averaging of the evolution under a set of linear operators
implemented via unitary dilation. (c) Evolution of probability for a single excitation
initially in an OH stretch mode (black) to be found in the stretch modes (top panel) and
symmetric bend (bottom panel) under relaxation dynamics.
and |0i ≡ |0i represents no vibrational excitations. Solving Eq. (5.11) results in time
dependent Kraus operators
p p
M1 (t) = 1 − e−Γ1 t |3ih1|, M2 (t) = 1 − e−Γ1 t |3ih2| (5.12)
p p
M3 (t) = 1 − e−Γ2 t |0ih3|, M4 (t) = 1 − e−Γ3 t |0ih1| (5.13)
v
p
u
u X5
−Γ
M5 (t) = 1 − e 3 |0ih2|, M0 (t) = 1 −
t t Mi† (t)Mi (t) (5.14)
i=1
The evolution between the non-vacuum states can then be simulated by the set of transfer
matrices K
Ki<3 (t) = UL Mi (t)e−iHt/~ UL† . (5.15)
Since this channel is non-unital (it doesn’t map 1 → 1), it cannot be simulated as a
135
convex sum of unitary evolutions as before. However, we can implement general linear
transformations using the concept of unitary dilation at the cost of a non-unit probability
of success. Any m × m matrix A with spectral norm ||A||2 ≤ 1 can be embedded as the
top-left block of a 2m × 2m unitary matrix UA . All possible unitary dilations of A can
be parameterised using the Cosine-Sine decomposition [212]. By considering a singular
value decomposition of A
~ †
A = U cos (θ)V
~ = Diag(cos θ1 , ..., cos θm ) and U, V † unitaries, we can express the possible

where cos (θ)
dilations of A into 2m modes as
   
U 0 ~ ~
cos(θ) − sin(θ) V †
0
UA =    
0 X ~
sin(θ) ~
cos(θ) 0 Y†
where X, Y ∈ U (m) are arbitary unitaries. In this experiment we used the choice
X = Y = 1. Using this technique, we are then able to simulate each of the set of
transformations K using their dilations UK , and postselecting upon detection in the top
three modes of the device.
The terms involving M3−5 can be simulated via introducing relative losses to the
modes. This could be achieved by coupling to further ancilla modes and this time a
detection in these modes would correspond to the projection |0ih0|. Here instead, we
simply apply a time-dependent effective efficiency to the detected counts on each mode.
Finally, M0 (t) is diagonal and commutes with UL , meaning that it can also be achieved
via a rescaling of the detected counts from the ideal unitary evolution UL e−iHt/~ UL† .
Statistically mixing these experiments results in the correct expectation values and
states ρ(t) for the thermalisation process. We used Γ1 = 0.24ps, Γ2 = 0.26ps and Γ3 =
1.36ps, which are the experimentally measured relaxation times from Ref. [210].
In Fig. 5.4c we simulate the thermalisation of a local OH stretch excitation via the
symmetric bend normal mode to its ground state. The probability of measuring the
excitation in the two local stretch (left panel) modes and the symmetric bend (right panel)
demonstrates the transfer of population from the high energy stretch modes via the lower
136
energy bend mode to the ground state. The average L1 distance in these experiments was
D̄ = 0.024 ± 0.007.
5.7 Anharmonic Hamiltonian for H2O

Going beyond the harmonic approximation we now consider the simulation of vibrational
modes described by coupled anharmonic oscillators. Again considering H2 O, including
all third derivatives and the semi-diagonal quartic derivatives in the PES and applying
vibrational perturbation theory [213, 214] yields an adjusted Hamiltonian
X √
~xij ωi ωj /2 a†i ai + a†j aj + 2a†i a†j ai aj

Ĥa = Ĥ + (5.16)
i≤j
where Ĥ is the harmonic approximation Hamiltonian and xij are the perturbation theory
coefficients.
In principle the higher order nonlinear terms in Ĥa can be simulated directly via weak
optical nonlinearities such as self-phase modulation and cross-Kerr effects. Here instead
we demonstrate an approach based on measurement-induced nonlinearity.
It is possible to implement a conditional phase shift on a two-photon Fock state using
an additional ancilla photon and postselection upon detection of a specific heralding signal
in additional optical modes. If the phase is π, this is the KLM NS gate [114]. Using
generalised nonlinear phase-shift (NPS) gates we are able to implement arbitrary phase
shifts between the zero, one and two Fock states of an optical mode.
In order to implement the NPS gates, we wish to perform diagonal unitaries in the
Fock basis characterised by the evolution
UNPS (φ1 , φ2 ) a0 |0i + a1 |1i + a2 |2i = a0 |0i + a1 eiφ1 |1i + a2 eiφ2 |2i

(5.17)
for a single optical mode. We will search for gates which use a single ancilla photon,
injected in an additional mode. In this case it is possible to then restrict the search to
arbitrary 2 × 2 transfer matrices M which can be extended to unitaries using the principle
137
V(R)
(a) (c) 1.0
occupation
0.8
0.8 0.6
Prob ofProb
0.6
Intensity
0.4
×× × × × × R
0.4
↭
0.2 × × × × 0.2
0.0 ×
7250 7500
×7750
0.0
Wavenumber (cm-1 ) 0 100 200 300 400 500
Time (fs)
1.0
(b)
occupation
0.8
0.6
Prob ofProb
0.4
0.2
Ancilla photon and
modes 0.0
0 100 200 300 400 500
Time (fs)
Figure 5.5: Anharmonic model for H2 O. (a) Spectrum of two excitations in bunched
(black) and anti-bunched (blue) local stretch modes for a harmonic (dashed) and an-
harmonic model (solid). (b) The anharmonic evolution is implemented via measurement-
induced nonlinearity using an ancilla photon and modes. (c) Evolution of the two bunched
local stretch excitations to be found in the anti-bunched state (top panel) and the bunched
state (bottom panel) under both models.
of dilation explained above. By considering the equations that must be satisfied
Per[M(0,1),(0,1) ] = e−iφ1 Per[M(1,1),(1,1) ]

(5.18)
2Per[M(0,1),(0,1) ] = e−iφ2 Per[M(2,1),(2,1) ]
the set of gates can be parameterised as

 √ 
iφ1

1− 1− eiΦ x e 0
M = √   (5.19)
y
x
1 − eiΦ y 0 1
where Φ = φ2 − 2φ1 . Optimal gates can then be found by performing the optimisation
max Psucc
(5.20)
subject to: ||M ||2 ≤ 1
where Psucc = |y|2 is the probability that the gate succeeds in performing the desired
138
0.25
3.0
0.24
2.5
0.23
0.22
2.0
0.21
Psucc
φ2
1.5
0.20
0.19
1.0
0.18
0.5
0.17
0.0 0.16
0.0 0.5 1.0 1.5 2.0 2.5 3.0
φ1
Figure 5.6: Optimised success probability (Psucc ) for a nonlinear phase shift (NPS) gate
mapping performing the transformation using two ancilla modes and one ancilla photon.
The white line is the curve φ2 = 2φ1 , along which Psucc = 1.
transformation. The optimisation is performed by using the sequential least squares

programming (SLSQP) [215] algorithm included in the SciPy [216] python package.
Fig. 5.6 shows the success probabilities for the optimised gates for pairs of phases
{φ1 , φ2 }. Beyond the trivial case of {φ1 , 2φ1 }, where Psucc = 1, globally optimal solutions
occur for {φ1 , φ1 +π} where Psucc = 1/4. These solutions can be seen to be generalisations
of the KLM NS gate. For the set of worst case {φ1 , φ2 } pairs, we find that Psucc = 1/6.
Ref. [115] establishes an upper bound on the probability of success of NPS gates of
the form {0, φ} with arbitrary ancilla states of Psucc ≤ (3 − cos(π − φ))2 /16. For the case
when φ = π this bound is saturated by the KLM NS gate. However, from our numerical
search we found that in restricting to such single photon ancillas, this upper bound was
not generally achievable. Furthermore, we observed that these probabilities were not
improved by considering separable ancilla states containing two photons. We leave it as
an important open problem to establish whether the upper bound from [115] is achievable
and what form of ancilla state is required. If the bound is achievable with reasonable
additional physical resources, this measurement-induced approach may be preferable to a
genuine analog nonlinear photonic simulator.
We simulated both harmonic and anharmonic evolution, again using H2 O, this time
139
restricting to the stretch mode subspace. Two photons injected together into the top
mode of the chip serves to simulate two excitations initialised in a superposition of the
harmonic model eigenstates corresponding to a local OH stretch. As shown in Fig. 5.5(b),
when simulating the anharmonic model, this input state can now be understood as the
same superposition of the new energy eigenstates of Ĥa , while a third photon injected
into the third optical mode serves as the ancillary system.
Fig. 5.5(c) shows the results for simulating the probabilties for these two vibrational
excitations to remain bunched or anti-bunch, under both the harmonic model, as before,
and the anharmonic model using the NPS gate. The difference in the detection patterns
between the two models, Ĥ and Ĥa , is a function of their different spectra, shown in
Fig. 5.5(a). The probability to detect a single excitation in each of the modes (i.e. anti-
bunched), shown in the top panel of Fig. 5.5(c), acquires a simple frequency shift for the
anharmonic evolution corresponding to the adjusted energy levels, as seen in Fig. 5.5(a).
In contrast, the probabilities for the state to remain doubly occupied display markedly
different dynamics between the harmonic and anharmonic cases, as shown in the bot-
tom panel of Fig. 5.5(c). This is a result of the three vibrational eigenstates no longer
being equally spaced in energy, as seen in Fig. 5.5(a), introducing new frequencies in
the evolution. For this set of experiments, the average distances between the ideal and
experimental distributions for the harmonic and anharmonic cases are 0.02 ± 0.01 and
0.06 ± 0.02 respectively.
5.8 Adaptive feedback control in the dissociation of

NH3
We finally consider the application of our photonic simulator as a platform for testing
approaches to quantum control. Adaptive feedback control (AFC) has been proven to be
a practical laboratory technique in finding optimal control fields for selective chemistry. In
AFC, a pulse-shaped control field is iteratively refined based on the evaluation of results
from molecular measurements. An algorithm evaluates each control setting based on its
140
measured outcome with respect to a pre-defined goal, and searches through the available
space of control parameters to move towards a locally optimal solution. This procedure
naturally incorporates the realistic laboratory environment and constraints.
Our example molecule is ammonia, NH3 , a prototype for studying dissociation, in-
cluding vibrationally mediated pathways, in which the states of its products, NH2 + H
depend on the prior vibrational state in the ground electronic state [191]. We therefore
consider the following model; the molecule is initially in its ground electronic state with
some vibrational energy among its normal modes. Through interaction with a light field
an electronic state is excited. We can obtain the resulting vibrational state by projecting
the vibrational ground state modes onto the vibrational modes of the excited state. This
transformation can in general be achieved via single mode squeezing, displacement and
linear optic transformations [217], here we approximate the projection by a unitary trans-
formation between the modes UGE . The transformation used to approximately model the
projection of the vibrational modes of the ground state QG to those of the excited state
QE is the one that maximises the overlap
UGE = min ||QE − QG U ||2 . (5.21)

U
The vibrational state then evolves in time under the Hamiltonian corresponding to the
excited state normal modes4 . Probing the resultant state in a localised basis as before
allows us to identify three local NH stretch modes.
The specific aim of this simulation, schematically depicted in Fig. 5.7(c), is to let an
AFC algorithm find initial states of two vibrational excitations (in the electronic ground
state molecule) which result in a maximal total probability to bunch in any of the three NH
stretch modes (of the electronically excited molecule) over the first 10fs of evolution, which
we associate with a preferred dissociation pathway, whilst suppressing other bunched
events which we associate with other pathways.
The algorithm begins with an initial state of two photons in orthogonal optical modes.
4
The planar equilibrium geometry and Hessian matrix of the first electronic excited state of NH3 were
computed at the CC2/TZVPP level of theory, where the Hessian was computed by finite difference using
the nuclear gradients of the excited electronic state [218]
141
(a) (b) (c)

NH3 | ~
e (✓, t)i
0.5
0.4
Intensity
0.3 ×
××
0.2 AFC
0.1
0.0 × ×
×
1000 1500 2000 2500
Wavenumber (cm-1 ) Adaptive Feedback Control
Calculate C
~ 0)i
| g (✓, Update ✓~i ! ✓~i+1
(d)
NH bunch Non-NH bunch Anti-bunched
1.0
0.5
0.5
0.5
-Cost
C
Prob
0.0
0.5
-0.5
0 100 200 300 400

2 4 6 8 10
Iteration
Iteration
Time (fs)
Figure 5.7: Adaptive feedback control algorithm for selective dissociation in NH3 . (a) A
two-excitation vibrational state, parameterised by θ,~ is initialised in the ground state vi-
brational modes of NH3 . The electronic state is excited and the localisation of vibrational
energy is measured over time. These measurements are then used to feedback to the state
preparation in order to increase energy localisation in NH stretch modes, promoting a
~ a
break-up of the molecule. (b) This scenario is simulated via the state preparation U (θ),
†
transformation between the ground and excited modes UGE , evolution under the excited
state modes and measurement via UL . (c) An example dataset showing (left panel) the
initial probabilities with a random input state for bunching in the NH stretch modes (red),
bunching in the remaining three localised modes (blue) and detection in anti-bunched pat-
terns (yellow) for five timesteps for three iteration numbers. The right panel shows the
measured cost function at every iteration.
We then imagine we are able to prepare a set of states defined by this state subject to a
~
unitary mixing of the vibrational modes, U (θ).
We ignore the sixth mode as this remains uncoupled from the others (see Appendix
B.1) and perform an arbitrary unitary transformation upon the remaining five modes.
There are 14 free parameters, θi , in this state preparation, which are mapped to the
phase shifts and beamsplitter splitting ratios within a linear optical network. These are
the parameters which are optimised in the experiment. This parameterised initial state is
then used as input to a simulation where a vibrational state in the ground state of NH3
is electronically excited, evolves for a time t and we measure the outcome distribution in
the local basis. An AFC algorithm then updates the parameters θ~ in order to maximise
142
the probability of bunching events in any of the three NH stretches within the first 10fs
of evolution whilst suppressing other bunched events.
The feedback algorithm is a type of random restart hill climbing optimisation with
an objective function which depends on observed data, and is implemented as follows.
We begin with a set of uniformly random parameters. We then collect all two-photon
coincidences at the output for a series of time steps separated by ∆t. After correcting for
detector efficiencies and the inefficient detection of multiple photon mode-occupancies,
the probabilities of bunched events
~ t) = |h2i |ψ(θ,
pibunch (θ, ~ t)i|2 (5.22)
can be estimated, where
X
~ t)i =
|ψ(θ, Uj1 (θ, ~ t)a† a† |0i
~ t)Uk2 (θ, (5.23)
j k
j,k
~ t) = UL e−iHt/~ U † U (θ).
and U (θ, ~ Since we wish to selectively promote dissociation via the
GE
three NH stretch modes (corresponding to the first three modes in the localised basis), the
~ t) = P3 pi
quantity we wish to maximise is p1 (θ, ~
i=1 bunch (θ, t), whereas we wish to minimise
~ t) =
bunching in (and therefore dissociation via) the remaining localised modes p2 (θ,
P6 i ~
i=4 pbunch (θ, t). The objective function that we use is
minθ~ C (5.24)
X
C = −α wi ∆pi (5.25)
i
with
i−1
Y
~ j∆t) + p2 (θ,
~ j∆t)]/N

w1 = 1, wi>1 = 1 − [p1 (θ, (5.26)
j=1
~ i∆t) − p2 (θ,
∆pi = [p1 (θ, ~ i∆t)]/N (5.27)
1
α= (5.28)
1− ( NN−1 )N
143
where later times are weighted in order to correct for the possibility of dissociation occur-
ring previously, N is the total number of timesteps and the factor α is used to renormalise
the objective function such that it is constrained to the interval [−1, 1].
The updating of the values of θi is performed using the Nelder-Mead simplex method [219]
contained in the SciPy [216] python package. At each iteration, the unitaries are set, data
is collected and the objective function C evaluated. The parameters θ~ are then updated
and the procedure is repeated until no further progress is detected, or 2000 evaluations
of C have occurred for a given starting set of parameters. The Nelder-Mead method is
preferred to gradient based methods in this situation, as the presence of experimental
noise is likely to cause C to be nonsmooth. The optimisation process is repeated for
many random initial parameter sets (random restarts) in order to increase the probability
of finding the globally optimal solution.
Fig. 5.7c shows an example experimental trial. We obtain a final value of C = −0.771
starting from a random initial state with C = +0.337. In all, this procedure involved the
collection of measurement statistics for almost 2000 unique settings of the photonic chip.
The experiment was repeated with six random initial states, all achieving similar final
values of the cost function with a mean C̄ = −0.845 (see Appendix B.2 for full data set).
These results show that it is possible to highly localise vibrational energy in NH stretches
of electronically excited NH3 by selective initial state preparation in its ground state in
addition to demonstrating the viability of our simulation scheme in this setting.
5.9 Discussion
In this chapter we have demonstrated that quantum states of light propagating in reconfig-
urable integrated photonic devices provide a new platform for the simulation of molecular
vibrational dynamics. In order to scale the experimental demonstration presented here to
tackle real world problems will require some additional questions to be investigated and
settled.
The losses and inefficiencies in quantum light sources, linear optical networks and
detectors will limit the rates at which the photonic simulator can be operated and in
144
turn its advantage over classical technology [138]. One approach to mitigating losses
may be to incorporate quantum memories, for instance by heralding and storing the
successful transformation of photons at each input optical mode sequentially. Considering
the simulation of different classes of input and output configurations such as Gaussian
states [128] and measurements [129] may also help reduce the experimental demands
required for demonstrating a quantum speedup.
Although even approximately sampling from the output distribution of linear optical
experiments like the ones we present can be classically inefficient [8], in order to produce
a useful simulation the linear optical network must be able to perform the correct transfer
matrix to a good degree of precision. Although the precision in the circuit parameters
must necessarily increase with dimension [220, 221], linear optical elements with extinction
ratios of over 60dB have been demonstrated [222, 223].
Finally, the development of programmable nonlinear optics at the quantum level is a
key functionality for quantum technologies and remains a major challenge for the field,
however, we have shown that weak nonlinearities provide a sufficient augmentation to
simulate richer molecular models. With modest progress in these areas, our approach
could yield an early class of practical quantum simulations that operate beyond current
classical limits. It might then be possible to discover new possibilities for molecular
control with quantum states of light, including new dissociation pathways and molecular
conformations.
Statement of Work
The simulation of localised molecular vibrations with universal linear optics was conceived
and developed by A. Laing and E. Martı́n-López. Computation of the Hessian matrix,
normal modes and local basis transformations for all molecules were provided by D. Tew.
Data taking was performed by N. Maraviglia, A. Neville and myself. All experiments
were designed and analysed by myself. Methods for simulating open system dynamics
were conceived and developed by myself. Implementation of NPS gates and the AFC
algorithm were developed by myself, A. Neville and N. Maraviglia.
145
Chapter 6
Linear Optical Quantum Computing

with Partially-Distinguishable
Photons
6.1 Introduction
The preceding chapters investigated current and near-term experimental applications of
photonic quantum information processing. Here we now turn our attention to the longer
term goal of achieving universal quantum computation with linear optics.
As discussed in Sec. 2.4.5, the fundamental challenge in constructing a useful quan-
tum computer is in implementing a set of universal quantum gates which can preserve
and manipulate the information in a quantum codespace with an error rate that is below
the fault-tolerance threshold of that code. Understanding how imperfections in device
fabrication and control result in corruption of the quantum information is therefore es-
sential for assessing the requirements of physical components to achieve fault-tolerance
as well as guiding the development of bespoke error correction strategies. Furthermore,
the natural error rate of a platform can make a significant difference to its practical scal-
ability; a quantum computer running just below the error threshold will require orders
of magnitude more physical qubits than one running significantly below it. Since LOQC
146
6. Linear Optical Quantum Computing with Partially-Distinguishable Photons
architectures already require large resource overheads due to their inherently probabilistic
gates, it is particularly important to suppress their logical error rates.
It is not always clear how physical errors emerge as logical errors however. In LOQC,
HOM interference, when coupled with photon detection, provides a mechanism to gen-
erate entanglement which is unreliant on direct interaction between photons. However,
ideal quantum interference between photons only occurs if the photons are completely
indistinguishable in all degrees of freedom but the one being interfered. Various forms
of filtering are commonly used to achieve improved photon indistinguishability at a cost
of added photon loss. In practice however, no two photons will be perfectly indistin-
guishable and therefore it is important to understand how using partially-distinguishable
photons will affect the quantum information within the linear optical quantum computer
and to find out how close to perfectly indistinguishable photons are required to be for a
fault-tolerant computation.
In this chapter we study the effect of using sources of partially-distinguishable photons
to generate and process photonic entanglement. We describe how states of partially-
distinguishable photons are transformed by linear-optical networks and measured by
photon-counting detectors and how to calculate the logical content of these states. We
then apply our methods to simple examples of entanglement generation and measurement
with linear optics which help elucidate the key features of the general problem. Finally,
we use these tools to calculate the resulting error rates in a recent proposal for LOQC
given in Ref. [120]. This analysis provides several insights into general features of er-
ror models in these architectures as well as specific properties of distinguishability error
models which are explored further in the next chapter.
6.2 Errors in LOQC

The main ways in which errors can occur in the operation of a linear optical quantum
computer are the following:
• Photon loss. This can occur through a number of mechanisms such as the scat-
tering of light into undetected modes, absorption in its propagation medium or
147
inefficiency in detection.
• Circuit error. Errors introduced by imperfect linear optical components. This can
be split into coherent and stochastic errors. Coherent errors are unitary errors which
preserve purity, e.g. an offset in the reflectivity of a directional coupler. Stochastic
errors reduce the purity of the state, e.g. dephasing caused by electrical noise in an
electro-optic phase shifter.
• Photon distinguishability. Errors introduced by the non-ideal interference of

partially-distinguishable photons.
• Multiphoton terms. Errors introduced by a photon source generating multiple

photons instead of a single one.
• Dark counts. Detectors clicking when there is no photon present.
Of course in any realistic device all these errors and more will occur simultaneously and
likely interact in non-trivial ways. It is important however to identify the most dominant
of these errors in terms of their likely influence on the overall error rates in current and
future devices.
Photon loss is perhaps the most considerable practical challenge in LOQC. However,
since it can generally be heralded without affecting the logical state, loss can be considered
largely a problem for scalability rather than fault-tolerance. Furthermore, it is known that
fault tolerant computation can still be achieved in the presence of very high loss rates
[224, 225], in fact it has been argued that in the limit of vanishing logical error, heralded
loss rates approaching 100% could be tolerated in the resource state generation [226].
Circuit errors in linear optical components, such as phase shifters or beamsplitters, are
predominantly of the coherent form due to fabrication imprecision [227]. As we demon-
strated in Chapter 4, experiments can already perform highly accurate multimode trans-
formations. Furthermore, since coherence is not lost, these errors can be corrected by
simple calibration procedures resulting in interferometers with high extinction ratios e.g.
99.999% in Ref. [222].
148
Multiphoton errors and dark counts will contribute to the logical error but in principle
can be greatly mitigated if reliable photon-number resolving detectors with high timing
resolution are available. Any detection events involving too many or few photons can
simply be treated as a heralded loss. In this case, the only errors are caused by second-
order processes where a combination of multiphoton emission and photon loss or photon
loss and dark counts occurs.
The effects of partial-distinguishability are important to consider as they are the great-
est source of logical error in current experiments as we saw in Chapter 4. As discussed
in Sec. 3.7.1, there are practical difficulties in producing identical photon sources using
both all-optical and matter-based systems. Although the deleterious effects of partial-
distinguishability can be reduced via filtering, for instance using narrowband frequency
filters to reduce spectral entanglement in pair generation sources, this process is gener-
ally inefficient and to achieve perfect indistinguishability requires the full control of the
photon’s additional degrees of freedom. The highest HOM-dip visibility between indepen-
dent sources with at least modest efficiencies is ∼ 0.95% [160], emphasising the technical
challenge.
6.3 Photon Distinguishability

Here we introduce the approach that will be used to model partial-distinguishability errors.
The main assumption that will be made is that the distinguishability originates from the
generation of the photons and linear optical networks manipulate only the paths leaving
all additional, internal, degrees-of-freedom unchanged.
Evolution of a photonic state under the LON T is therefore governed by the mapping
X
a†r [ψj ] → Tkr a†k [ψj ]. (6.1)
k
where, as before, a†i [ψj ] is the operator which creates a photon in path r with an internal
149
state specified by the state |ψj i ∈ Hint . Similarly, we can consider mixed states
X
A†r [%int ] ◦ ρ = pi a†r [ξi ]ρar [ξi ] (6.2)
i
where %int is an operator on Hint with eigenvalues and eigenstates pi and |ξi i respectively.
A single coarse-grained photon-number-resolving detector (CG-PNRD) which does not
resolve any information except for the number of photons in a path, r, may be described
by a POVM {Fpr } with elements associated with detecting p photons
" k # " k
#
k
X Y 1 Y 1 XO
Fpr = √ (a†r [ξi ])ni |0ih0| √ (ar [ξi ])ni = |ni ihni |r,i . (6.3)
n∈Np i=1
ni ! i=1
ni ! n∈N i=1 p
Here Np is the set of all tuples (n1 , ..., nk ) which sum to p, {ξi } is an orthonormal set of
functions spanning Hint and k = dim (Hint ). For brevity we define operators associated to
(p ,...,p ) p
detection patterns: Fpr ≡ F(r11,...,rjj) ≡ Fpr11 ⊗ . . . ⊗ Frjj . We discuss in the next chapter the
potential advantages of using PNRDs which provide additional information, e.g. timing
information.
6.3.1 Dual-path states
To describe the dual-rail logical state of a partially-distinguishable photonic state we can

use a†b [ξe ]|0i as a basis for a single photon within one of two paths, labeled by a bit
b ∈ {0, 1}, with an internal state |ξe i from an orthonormal basis {|ξi i} of Hint . This
basis state can then be expressed as a dual-path state whose isomorphic Hilbert space is a
tensor product between the logical qubit and internal qudit spaces: a†b [ξe ]|0i = |bi ⊗ |ξe i.
Here the the internal state can be traced out to yield a dual-rail qubit. Similarly, the
dual-path encoding of a state of n photons can be obtained by mapping the corresponding
Fock space to the tensor product of logical and internal spaces, HF → HL ⊗ Hint , in an
analogous manner to the single-photon case.
If there is more than one or no photons within the two rails, then we describe all
such states by a third level |2i which is outside of the computational space. Such cases
150
will be considered as heralded losses as we will assume unit number-resolved detection

efficiency. We note that in order to correctly describe the final logical state in any scheme,
the tracing out procedure can only be performed upon photons which are detected or can
no longer interfere with any other photons.
6.3.2 Distinguishability models
It will often be helpful in what follows to consider photonic states with specific distin-
guishability structures. The two models that will be predominantly used are the orthog-
onal bad-bits model (OBBM) and the random source model (RSM).
1
In the OBBM we consider a set of photons with pure internal states |ψi i = V 4 |ξ0 i +
√ 1
(1 − V) 2 |ξi i with a common overlapping part |ξ0 i and a contribution from |ξi i which
is orthogonal to all other internal states, hξj |ξi i = 0 ∀ i 6= j. This ensures that the
HOM-visibility of any pair of photons is equal, Vij = |hψi |ψj i|2 = V ∀ i 6= j.
In the RSM we consider a photon source factory which produces independent and
identically distributed photon sources characterised by a set of variables x which produce
photons characterised by these variables %(x). Since the sources are independent, on a
single-shot basis it is possible to define a common average internal state for all the photons,
P P
% = x Pr(x)%(x) = j pj |ξj ihξj |. The average pairwise HOM-visibility of any pair of
sources is then given by V = Tr(%2 ) = j p2j . The RSM is particularly relevant when
P
using single-photon sources based on multiplexing a large number of non-deterministic

sources since each photon will originate from a different physical origin in each shot.
Under certain conditions, these models can become equivalent. Since each term in the
n-photon OBBM input state contains a different combination of internal states, they do
not interfere with each other and their coherences do not contribute to the logical state.
As a result, all logical quantities in the OBBM are equivalent to a model with the initial
N N √ √
state i %i = i V|ξ0 ihξ0 | + (1 − V)|ξi ihξi |.
This equivalent dephased model can then be understood as the high-dimensional limit
√ √
of a model consisting of i.i.d mixed states, % = V|ξ0 ihξ0 | + (1 − V) limd→∞ di |ξi ihξi |/d
P
such that the probability of any two bad photons being in the same internal state is
151
negligible. The OBBM is therefore equivalent to a specific case of the RSM described
√
above with one dominant eigenvalue p0 = V and a large number of vanishing ones
pi6=0 ∼ 0.
Finally, we note that all RSM inputs become equivalent to first-order in the limit
of high purity, p0 = 1 − . For n photon sources, the probability of one photon not
being in the dominant mode |ξ0 i is O() whereas the case where two photons are not in
the dominant mode will occur with probability O(2 ). To describe the n-photon state
to first-order in it is therefore sufficient to describe the single photon internal states as
% = (1−)|ξ0 ihξ0 |+|ξ1 ihξ1 |, since the distinguishable photon will not undergo interference
with any other photons whichever specific |ξk6=0 i it may be. This high purity limit will be
relevant when considering architectures that are approaching fault-tolerance.
6.4 Generating entanglement with

partially-distinguishable photons
In this section we provide explicit calculations of the logical state produced by some stan-
dard entanglement-generating linear optical gates as a function of the distinguishibility
structure of the input photons. The simplest case to consider is when one is allowed to
perform postselection. We consider the simple four path interferometer, shown in Fig
6.1a, which ideally generates a postselected Bell state of two indistinguishable photons.
The transfer matrix of this interferometer is
 
1 1 0 0
 
1 −1 
 
1  0 0
T =√   (6.4)
2
 0 0 1 1 

 
1 −1 0 0
152
and we consider an input of two photons in pure internal states in paths one and three
a†1 [ψ1 ]a†3 [ψ2 ]|0i. Evolving under the LON produces the output state
X X 1
Ti1 a†i [ψ1 ] Tj3 a†j [ψ2 ] |0i = a†1 [ψ1 ] + a†4 [ψ1 ] a†2 [ψ2 ] + a†3 [ψ2 ] |0i

i j
2
post. 1
−−→ √ a†1 [ψ1 ]a†3 [ψ2 ] + a†2 [ψ2 ]a†4 [ψ1 ] |0i

(6.5)
2
where we have applied the postselection to remove non-logical terms and renormalised on
the second line. By re-writing this postselected state in the logical-internal basis
1
√ |00i ⊗ |ψ1 ψ2 i + |11i ⊗ |ψ2 ψ1 i (6.6)
2
the logical postselected state can then be recovered by tracing out the additional degrees
of freedom, resulting in the postselected state
1h 2 i
ρpost
L = |00ih00| + |11ih11| + |hψ1 |ψ2 i| |00ih11| + |11ih00|
2
1 + V12 + + 1 − V12 − −
= |φ ihφ | + |φ ihφ |. (6.7)
2 2
Where V12 = |hψ1 |ψ2 i|2 is the HOM visibility of the two photons. This result can be readily
generalised to independent mixed input states in the additional DOF, A†1 [%1 ]◦A†3 [%2 ]◦|0ih0|,
by considering ensemble averages. In this case the state keeps the form of (6.7) with the
HOM visibility V12 = Tr(S(%1 ⊗ %2 )) = Tr(%1 %2 ) where S is the swap operator (see Sec.
3.4.2). We note that in this setting, dual-rail ‘entanglement’ is generated by photons with
any non-zero overlap in their internal states, and partial distinguishability simply results
in an effective dephasing channel applied to one qubit of the desired maximally-entangled
postselected state.
One observation we can make from this simple example is that the effect of partial-
distinguishability is to reduce the coherence of the logical state. We therefore see that
although photonic states are not as susceptible to traditional decoherence via the loss of
information into a thermal environment, the inability of detectors to resolve individual
optical modes necessarily means that information is lost and the logical states become
153
a) b)
c)
Figure 6.1: Linear optical networks for Bell state preparation and measurement. a) Post-
selected bell state generator. Generates a bell state with two single photons as input
and postselecting on logical states at the output. b) Heralded bell state generator. Gen-
erates a bell state with four single photons as input, heralding on measurement of two.
c) Probabilistic bell state analyser. Given a logical two qubit input, performs a projec-
tive measurement onto the |φ+ i and |φ− i bell states, otherwise gives a heralded failure
outcome.
mixed if their internal states are not identical.

In order to perform scalable computation with linear optics and single-photon detec-
tors, the entanglement generation must be heralded via the detection of additional ancilla
photons. The most efficient interferometer known for heralded generation of two-qubit
entangled states from single photons is the B8 interferometer discussed in Sec. 4.4.2 and
shown in Fig. 6.1b [179]. This circuit has the transfer matrix
 
2 2 0 0 0 0 0 0
 
0 0 2 −2 0 0 0 0
 
 
 
1 −1 1 1 1 1 −1 1
 
 
 
−1 −1 −1 −1
 
1  1 1 1 1 
B8 = √  . (6.8)
2 2
 −1 1 1 1 1 1 1 −1


 
−1 −1 −1 −1
 
 1 1 1 1 
 
 
 0 0 0 0 0 0 2 2 
 
0 0 0 0 2 −2 0 0
The network heralds the Bell state |φ+ i, up to permutations of the output paths, from
154
four photons conditional upon the measurement of two of them in a set of six possible
heralding patterns. We consider an input state of the form:
A†1 [%1 ] ◦ A†3 [%2 ] ◦ A†5 [%3 ] ◦ A†7 [%4 ] ◦ |0ih0|,
after evolution via B8 , projection onto one of the six heralding patterns and performing
any output path permutations required to produce the bell state |φ+ i, tracing out the
internal states and ignoring non-computational terms produces the logical state:
1
ρoL = ×
32 × po
 
1 + Vjk 0 0 Vij Vkl + Tr(%i %j %l %k )
 
1 − Vjl Vij Vkl − Tr(%i %j %k %l )
 
 0 0 
 
Vij Vkl − Tr(%i %l %k %j ) 1 − Vik
 
 0 0 
 
Vij Vkl + Tr(%i %k %l %j ) 0 0 1 + Vil
(6.9)
with po = [4 + (Vjk + Vil ) − (Vjl + Vik )]/32 and where o = 1, 2, 3 corresponds to the state
produced when an element from one of three different sets of measurement outcomes
(1,0,1,0) (0,1,0,1) (1,0,0,1) (0,1,1,0) (1,1,0,0) (0,0,1,1)
{F(3,4,5,6) , F(3,4,5,6) }, {F(3,4,5,6) , F(3,4,5,6) }, {F(3,4,5,6) , F(3,4,5,6) } is obtained respectively. For
o = 1, 2, 3 the state is specified by (i, j, k, l) = (1, 2, 3, 4); (2, 1, 3, 4); (1, 4, 3, 2) respectively
such that all ρoL are equivalent up to permutations of the input photons’ internal states.
We see that the logical state ρoL is no longer entirely determined by the pairwise
HOM-dip visibilities. Instead, it is a function of higher-order trace overlaps such as
Tr(%1 %2 %3 %4 ) corresponding to the expectation of permutation operators on the internal
space. However, it is important to understand in what ways the state is constrained
by just the dip visibilities, since this is the experimental quantity that is most readily
accessible and also determines the complexity of models used to simulate the effects of
distinguishability.
155
The state fidelity between ρoL and |φ+ i is:

1 1
F(ρoL , |φ+ i) = 1 + [Vil + Vjk ] + Vij Vkl + < [Tr(%i %j %l %j )]
32 × po 2
(6.10)
1 1 p
≤ 1 + (Vil + Vjk ) + Vij Vkl + Vij Vjl Vkl Vik ≡ Fub .
32 × po 2
where, to provide an upper bound in terms of visibilities, we have used |[Tr ( ni=1 %i )]| ≤
Q
√
V12 V23 . . . Vn1 as proven in appendix C.1. Pure states with entirely real positive inner
p
products hψi |ψj i = Vij saturate this bound. We note that different visibilities have
different roles in Fub . For example, the value of Vij affects Fub more than the value of Vjl
does. This suggests that in some cases the fidelity of the output may be optimised via a
prudent ordering of input photons. In the case that all visibilities are equal Vij = V, Fub
V
simplifies to Fub = (1 + V + 2V 2 )/4. This bound is saturated by the OBBM introduced
in Sec. 6.3.2.
Similarly, the purity can be upper bounded:
1 n
P(%oL ) = 4 + Vjk (Vjk + 2) + Vil (Vil + 2) + Vik (Vik − 2) + Vjl (Vjl − 2) + 4Vij2 Vkl 2
(32 × po )2
o
2 2
+ 2 |Tr(%i %j %l %k )| + 2 |Tr(%i %j %k %l )| + 4Vij Vkl <[Tr(%i %j %l %k ) − Tr(%i %j %k %l )]
1 n
≤ 4 + Vjk (Vjk + 2) + Vil (Vil + 2) + Vik (Vik − 2) + Vjl (Vjl − 2) + 4Vij2 Vkl 2
(32 × p )
o 2
q o
+ 2Vij Vjl Vkl Vik + 2Vij Vjk Vkl Vil + 4Vij Vkl Vij (1 − Vij )Vkl (1 − Vkl ) ≡ Pub .
(6.11)
For the proof of this upper bound see appendix C.2. This bound is not generally saturable
and Pub is not generally a monotonic function of the visibilities. However, in the region
Vij ≥ 0.8 ∀ i 6= j, Pub is monotonic. In the case that all the pairwise overlaps are equal
V
Pub becomes Pub = (1 + V 2 + V 3 + V 4 )/4.
Given a set of photons we can also calculate a lower bound on the average purity and
fidelity by considering the case corresponding to the RSM, where each photon shares a
common average internal state %. One could achieve this by for instance simply randomly
permuting the input photons on every trial.
156
0.2
0
0 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0.2 0.4 0.6 0.8 1
00 0.2
0.2 0.4
0.4 0.6
0.6 0.8 1
6. Linear Optical Quantum
0.99 Computing
0.99
0.995 with1 Partially-Distinguishable
0.995 1
Photons
1 1
0.995 1
99 0.995
0.995 11
1
a) 0.995 1 b) 1 1 0.995 1
1 1 1 1
V V V
Pub 1
1
1 Fub
0.995
= Fobb
0.995
2 2
0.8 V 0.8
Pobb
0.995
0.995
22
2 hFilb 3 3
0.6 hPilb 3
33
0.6
0.4 0.4
0.99 0.99
0.2 0.99
0.99 0.2
0 0.99 0 0.99
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
V V
Figure 6.2: Purity and Fidelity of logical Bell states produced with partially-
distinguishable photons. a) Purity of the output of the B8 circuit for differing distin-
V
guishability models as a function of the pairwise visibilities V. Pub is a generally un-
V V
saturable upper bound, hPlb i is the lower bound for average purity and Pobb is for the
OBBM. b) Fidelity of the output as a function of V. In this case the upper bound is
V V
saturated by the OBBM Fub = Fobb and the average lower bound hFlbV i is also shown. All
these lines are extremely similar, suggesting that the pairwise visibilities highly constrain
the resulting logical states in this circuit.
1
P(ρoL (%)) = 1 + V 2 + V 4 + Tr2 (%4 )

4
(V≥1/2) 1
(6.12)
≥ (5 − 4V + 6V 2 + 4V 3 + 5V 4 ) ≡ hPilb
16
1
√ √
Where the inequality uses the lower bound Tr(%n ) ≥ 2n
[(1 − 2V − 1)n + (1 + 2V − 1)n ]
which is valid and saturable when V > 1/2.
V V
Pub , Pobb and hPilb are plotted in Fig. 6.2a). The OBBM is also included since
it provides a saturable value for P which is not an average case. In the same way,
V
hFilb = (1 + 4V + 3V 2 )/8 can be found using the RSM. hFilb and Fub are plotted in
Fig. 6.2b.
Fig. 6.2 shows that the achievable purity and fidelities using B8 are heavily constrained
by the pairwise visibilities. This is likely to be a property of the network itself, since it is
known that some interferometers can result in detection events which are highly dependent
upon the internal state configuration [228]. This idea will also be explored in the following
chapter.
Using either the OBBM or RSM, the logical state ρL in Eq. (6.9) can be conveniently
157
represented by an ideal preparation of the Bell state |φ+ i subject to a Pauli error channel
applied to either qubit. In the case of the OBBM this single qubit error map is
1
E obb (ρ) = (1 + V + 2V 2 )ρ + (1 − V) XρX + Y ρY + (1 + V − 2V 2 )ZρZ

(6.13)
4
so that ρobb
L = (E
obb
⊗ 1)[]|φ+ ihφ+ |] = (1 ⊗ E obb )[]|φ+ ihφ+ |]. Similarly for the RSM,
1
E rs (ρ) = {1 + V + V 2 + tr(%4 )}ρ + {1 − V + V 2 − tr(%4 )}XρX
4 (6.14)
2 4 2 4

+ {1 − V − V + tr(% )}Y ρY + {1 + V − V − tr(% )}ZρZ .
Since both states are Bell-diagonal their entanglement properties can be straightfor-
wardly evaluated [229]. The concurrence of the states are given by the simple expressions
V−1 1
C(ρobb
L ) = Max 0, 2 + V
2
and C(ρrs 2 4
L ) = Max 0, 2 (V − 1 + V + tr(% )) . In contrast
to the postselected case, there is a minimum overlap condition to generate any entangle-
ment at all given by V > 21 .
The simple form of these error models can be explained by the symmetries of both
the transfer matrix T and the internal states. All four-photon trajectory amplitudes
Q
i Tσ(ri ),ci where c = (1, 3, 5, 7) and e.g. r = (1, 2, 3, 5, 7, 8), are invariant under ex-
changing columns (1, 3) and (5, 7) and rows (1, 2) and (7, 8) of T . This corresponds to
swapping internal states (%1 , %2 ) and (%3 , %4 ) as well as the output modes corresponding
to both qubits. This means that if the internal state has the permutational symmetry
(S ⊗ S)%int (S ⊗ S) = %int , then the resultant logical state will be invariant under the
logical X ⊗ X operation (the mode-swaps), (X ⊗ X)ρL (X ⊗ X) = ρL . It can also be
shown that the trajectories are invariant under the logical transformation Z ⊗ Z at the
output without any permuting of the input columns, therefore (Z ⊗ Z)ρL (Z ⊗ Z) = ρL .
Since (±XX, ±ZZ) are the stabilizer generators of the Bell states, any input with the
S ⊗ S internal state symmetry must result in a logical state that is a mixture of the four
Bell states. This explains why the resulting logical states are Bell-diagonal for both the
RSM and OBBM and the effective error maps are Pauli diagonal in Eqs. (6.13,6.14).
158
6.5 Entangled measurements with

partially-distinguishable photons
The other operations at the heart of LOQC are the measurements that probabilistically
fuse multiple small entangled states into a single larger one. Here we briefly describe
how to calculate the resulting reduced states and effective measurement operators when
partially-distinguishable photons are used in these gates.
A general linear optical measurement device is characterised by its LON transfer ma-
trix T , and a set of successful measurement patterns, {F}, registered by a collection of PN-
(p ,...,pj )
RDs. The POVM element associated with each pattern is E(pm ) = E(m11 ,...,m = ji Epmi i
N
j)
where Epm = U † (T )Fpm U(T ).

Consider an initial dual-path state of n photons ρ ∈ HL⊗n ⊗ Hint
⊗n
. The measurement
device is then performed upon the first d photons which encode logical qubits and include
any ancillary states. Since this input state will generally be non-separable between the
logical and internal spaces, it is not possible to describe an effective POVM acting on
the logical Hilbert space which is independant of the state in the internal Hilbert space.
However, an effective logical POVM can be described if the state on the internal Hilbert
space is fixed. In the following we will therefore consider an initial state of the form:
d+1...n
ρ = ρ1...n
L ⊗ %1...d
int ⊗ %int . (6.15)
In this case, the reduced state on the remaining systems is given by
1
ρd+1...n (~r) =Tr1..d Fpm U(T )ρU † (T )Fpm ]

(6.16)
N
1 d+1...n
= Tr1..d Epm (ρL1...n ⊗ %1..d

int )] ⊗ %int (6.17)
N
1
= TrL1...d Trint1...d Epm (1 ⊗ %12 1...n
⊗ %d+1...n

int )(ρL ⊗ 1) int (6.18)
N
1
= TrL1..d E%int ρ1...n ⊗ %d+1...n

L int (6.19)
N
= ρd+1...n
L ⊗ %d+1...n
int (6.20)
159

where E%int (~r) = Trint1...d Epm (1 ⊗ %1...d
int ) is the effective logical POVM element.
A simple example of a commonly used entangling measurement device is the two-

photon bell state analyser [118] (up to local unitaries this is also equivalent to the Type-
II Fusion operation described in Sec. 3.6.2). As can be seen from Fig. 6.1c, such a
measurement device is simply the inverse process of the postselected Bell state generator
(1,0,1,0)
in Fig. 6.1a, and the LON is given by T † (see Eq. (6.4)). The measurements E(1,2,3,4) and
(0,1,0,1)
E(1,2,3,4) ideally project onto the |φ+ i Bell state. Both of these detection patterns result
in the same effective logical operator
1 + V12 + + 1 − V12 − −
E%int = |φ ihφ | + |φ ihφ |, (6.21)
4 4
with V12 = Tr[S%int ]. Up to normalisation, this is the same as Eq. (6.7) which is unsur-
prising, as it is simply the time-reversed process of the postselected Bell state generator.
When %int = |ψ1 ihψ1 | ⊗ |ψ2 ihψ2 |, this reproduces a known result from Ref. [230].
6.6 Measurement-based resource generation

Having introduced the basic methods for obtaining the logical content of partially distin-
guishable photonic states we now turn our attention to understanding the implications of
partial-distinguishability in a large-scale LOQC architecture.
6.6.1 Overview and initial state
As touched on in Sec. 3.6.2, it was recently shown that LOQC can be performed ballisti-
cally [120, 121], i.e. without the need for a repeat-until-success strategy that would require
fast switching networks or quantum memories. These schemes use the Bell state measure-
ment circuits of [231, 232] in order to probabilistically fuse a supply of entangled 3-GHZ
states with a ‘boosted’ success probability of 75%. In the following we will be assuming an
all-optical implementation which begins with a large array of independent, deterministic
single photon sources. The best known approach to generating the 3-GHZ states with
single photons and linear optics is to multiplex many probabilistic GHZ generator circuits
160
such that with near unit probability at least one is successful. Once the fusion gates have
been implemented, the resulting global state is a three-dimensional graph state of a target
lattice with a fraction of missing bonds. Using ideas from percolation theory, since the
fusion succeeds with a probability far above the critical threshold of the diamond [120]
and pyrochlore [121] lattices it is guaranteed that there will be a spanning path through
the lattice providing a universal resource state. The computation can then take place via
a series of adaptive single qubit measurements upon this resource state.
Our calculations begin with the ansatz that in order to achieve fault tolerance the
photons in LOQC must be highly indistinguishable. If the photons are highly indistin-
guishable, the probability of finding any photon in the dominant internal mode |ξ0 i must
be close to one, p0 = 1 − . As described above, in this case the OBBM and RSM coincide
to first order and it is sufficient to consider the joint internal state ni %i = %⊗n where
N
% = (1 − )|ξ0 ihξ0 | + |ξ1 ihξ1 |. As such, the inital state of n single photons is
n †
ρ=[ i Ai (%)] ◦ |0ih0|
n
X
n n−1
= (1 − ) |Ψ0 ihΨ0 | + (1 − ) |Ψj ihΨj | + O(2 ) (6.22)
j=1
where |Ψ0 i = |ξ0 i⊗n is the fully indistinguishable part and |Ψj i = |ξ0 i⊗j−1 ⊗|ξ1 i⊗|ξ0 i⊗n−j
labels the position of the distinguishable photon. The parameter can be physically
understood in terms of the average pairwise visibility between any two sources V =
Tr(%2 ) = (1 − )2 + O(2 ) ⇒ + O(2 ) = (1 − V)/2.
Within this model, the probability of one or fewer photons being distinguishable de-
creases exponentially with an increasing number of photons. However we will consider
only the number of photons required to obtain an error rate on a single logical qubit, a
quantity which does not scale with the size of the computation. The schematic in Figure
6.3a shows the scheme for generating a single qubit in the full resource state lattice in
Ref [120] from 34 single photon sources. The generation of the resource state is analysed
in two steps; the production of 3-GHZ states from single photons followed by the fusion
operations.
161
a) b)
z
y
x
c) d)
Figure 6.3: Schematic of an all-optical implementation of ballistic LOQC from [120]. The
outputs of three GHZ generators (dark blue) are initially fused into a five qubit micro-
cluster by two boosted bell measurements and the local rotations depicted. This is then
fused with similar microclusters (light blue) via four further boosted fusion operations.
This process generates a single photon in the final lattice from 34 initial single photons.
b) A configuration of these qubits in the final diamond lattice in the case that all fusion
measurements are successful. c) Optical circuits. Left is the circuit from [233] which
generates a 3-GHZ state from 6 single photons with a probability of 1/32. Right is the
circuit from [231] which performs a bell-basis measurement using 4 single photon ancillas
with a success probability of 3/4.
6.6.2 GHZ generation
The circuit described in Ref. [233] is a generalisation of the heralded Bell state scheme B8
which ideally generates the three qubit linear graph state |G3 i from six indistinguishable
single photons with a probability 1/32 and is shown in path encoding in Figure 6.3c).
The evolved state of six photons from the input state (6.22) through this circuit can be
rewritten in the logical-internal basis. Three of these photons are then measured in one of
the successful heralding patterns which allows us to immediately trace out the additional
162
DOF information of these photons. The resulting state, prior to normalisation, of the
remaining three photons is then
1 6
ρghz = (1 − )6 |G3 ihG3 | ⊗ |Ψ0 ihΨ0 | + (1 − )5 [2ρdist + |2ih2|] + O(2 ) (6.23)
32 64
where |2ih2| collects all illogical terms (i.e. terms where at least one qubit is in the |2ih2|
logical state) and ρdist is the remaining (normalised) logical-internal state containing one
distinguishable photon. We therefore find that when there is a distinguishable photon,
although the overall heralding probability is increased (3/64), this is caused by leakage
error into non-computational states and the probability of heralding a valid logical state
is unchanged. Due to the symmetry of the interferometer and input state, ρdist is the
same for any detection pattern up to a conditional swapping of modes.
To analyse the logical part of the state, we represent it in the graph-state basis [234].
As described in Sec. 2.4.7 a graph-state, |Gi, can be defined by a set of n stabilizer
Q
generators Si ≡ Xi j∈N (i) Zj , where N (i) is the neighbourhood of vertex i in the corre-
sponding graph, such that Si |Gi = |Gi∀i. The graph state basis is an orthonormal basis
whose elements are the common eigenstates of these stabilizer generators. Each common
eigenstate |Gs i can be indexed by a bit-string s ∈ {0, 1}n where si = 1 corresponds to
performing a Z operation upon the ith qubit of the graph state |Gi. In the case of the
three qubit linear graph, this basis contains the ideal state |G(0,0,0) i = |G3 i and states
which can be obtained by acting Pauli unitaries on a single qubit of |G3 i.
After tracing out the internal Hilbert space, we find that the resultant logical state
Trint (ρdist ) is diagonal in the graph state basis and hence differs from the ideal state by
random applications of single qubit Pauli operators. Moreover, even without tracing out
the internal Hilbert space, ρdist effectively admits a decomposition with terms where the
location of the distinguishable particle is well defined and the logical state is diagonal in
the graph state basis,
163
3
!
X X X
ρdist = pjs |Gs ihGs | ⊗ |Ψj ihΨj | + Ri,j ⊗ |Ψi ihΨj |. (6.24)
j=0 s i6=j
Terms relating to coherence between different positions of the distinguishable particle

i.e. |Ψi ihΨj |, ∀i 6= j will not lead to observable effects in the type of device we are
studying. This is evident as these contributions do not lead to further interference after
the generation of the three-qubit resource state. The coefficients pjs are:
1 1
(p0000 , p0001 , ...) = (3, 1, 3, 1, 1, 1, 1, 1) , (p1000 , ...) = (1, 0, 0, 0, 1, 0, 0, 0) ,
24 12
1 1
(p2000 , ...) = (1, 0, 0, 0, 0, 1, 0, 0) , (p3000 , ...) = (1, 1, 0, 0, 0, 0, 0, 0) . (6.25)
12 12
Therefore given that one of the input photons is distinguishable, the total probability of
a logical error is Pr(error|dist) = 1 − 3j=0 pj000 = 5/8.
P
6.6.3 Fusion gates
We now consider the boosted Bell measurement approach from Ewert et al. [231] which
uses four single ancilla photons to perform a probabilistic Bell basis measurement on a two
photon logical state. The path-encoded circuit for this operation is shown in Fig. 6.3c.
Using ancilla photons in the initial state Eq. (6.22), the effective logical measurement
operators, as per Sec. 6.5, can be calculated for three input states; the ideal case where
all photons involved are indistinguishable, the case where one of the logical photons is
distinguishable and finally the case where one of the ancilla photons is distinguishable.
The scheme of Ref. [231] has a total of 572 detection patterns {Fi }, 500 associated
with successful Bell state projection and 72 associated with failure. Any detection events
in other patterns are detectable errors and will be treated as a heralded leakage. Ideally,
the logical measurement operators are given by Ei = Trint [Ei (1⊗|Ψ0 ihΨ0 |)] = hΨ0 |Ei |Ψ0 i.
Many of the Ei are proportional to each other so that when combining these equivalent
164
patterns the full measurement implements the logical POVM {E} with elements
1
E 1 = |ψ + ihψ + |, E 2 = |ψ − ihψ − |, E 3 = |φ+ ihφ+ |,
2 (6.26)
41 51 1
E = |φ− ihφ− |, E = |00ih00|, E 6 = |11ih11|.
2 2 2
The probability of a successful Bell state projection on a maximally mixed input is

P4 d
P6 d

d=1 Tr E /4 = 3/4 and failure d=5 Tr E /4 = 1/4.
In the case where either of the logical photons at the input is distinguishable, the
circuit implements the effective operators {hΨ3 |Ei |Ψ3 i}, {hΨ4 |Ei |Ψ4 i} which both result
in the measurement {EL } given by
1 + +
EL1 (φ+ ) = |φ ihφ | + |φ− ihφ− | ,

8
1 + +
EL2 (φ− ) = |φ ihφ | + |φ− ihφ− | ,

8
1 1
EL3 (ψ + ) = |ψ + ihψ + | + |ψ − ihψ − | + |01ih01|,
8 4
1 1
EL4 (ψ + ) = + + − −
|ψ ihψ | + |ψ ihψ | + |10ih10|,
8 4
1 1
EL5 (ψ − ) = − − + +
|ψ ihψ | + |ψ ihψ | + |01ih01|,
16 8
1 1
EL6 (ψ − ) = − − + +
|ψ ihψ | + |ψ ihψ | + |10ih10|,
16 8
1 1
EL7 (ψ − ) = − − + +

|ψ ihψ | + |ψ ihψ | + |01ih01|,
128 16
1 1
EL8 (ψ − ) = − − + +

|ψ ihψ | + |ψ ihψ | + |10ih10|,
128 16
11
EL9 (ψ − ) = − − + +

|ψ ihψ | + |ψ ihψ | ,
64
1
EL10 (00) = |00ih00|,
8
1
EL11 (11) = |11ih11|. (6.27)
8
where the bracket indicates the outcome corresponding to the measured patterns in the
ideal case. The probability of success is 5/8, failure is 1/16 and leakage is 5/16. When
a failure outcome is received, no logical error occurs. Using the same approach we can
find the measurement which is performed in the case where a randomly-chosen ancilla
165
photon is distinguishable. That is, we consider the internal state of the input state to
P
be %int = i∈{1,2,5,6} |Ψi ihΨi |/4 and calculate Trint [Ei (1 ⊗ %int )]. Collecting all equivalent
terms this measurement {EA } is given by
3 + + 1 1 + +
EA1 (φ+ ) = |φ ihφ | + |φ− ihφ− | + |ψ ihψ |,
32 32 8
3 + + 1
EA2 (φ+ ) = |φ ihφ | + |φ− ihφ− |,
32 32
3 − − 1 1 + +
EA3 (φ− ) = |φ ihφ | + |φ+ ihφ+ | + |ψ ihψ |,
16 16 8
5 + + 1
EA4 (ψ + ) = |ψ ihψ | + |00ih00|,
32 8
5 + + 1
EA5 (ψ + ) = |ψ ihψ | + |11ih11|,
32 8
3 + + 1
EA6 (ψ + ) = |ψ ihψ | + |00ih00|,
32 8
3 + + 1
EA7 (ψ + ) = |ψ ihψ | + |11ih11|,
32 8
1 − −
EA8 (ψ − ) = |ψ ihψ |,
2
3
EA9 (00) = |00ih00|,
16
3
EA10 (11) = |11ih11|.
16
and has a probability of success of 9/16, failure 3/32 and leakage 11/32. Measurement
patterns corresponding to EA8 perform an ideal projection. This is because |ψ − i detection
relies only on interference between the two logical photons, which in this case interfere
perfectly.
We note that a significant portion of the time, if we have a distinguishable photon
in the fusion measurement we are able to identify this by receiving an incorrect leakage
measurement pattern. Therefore this is a mechanism by which logical error is converted
into loss. It would be possible to go further by using the information about the specific
measurement pattern we detect. The post-fusion state is dependent upon the outcome of
the measurement. In the ideal case, local unitaries are conditionally applied such that the
final state remains a graph state with a joined bond (success) or no bond (failure) between
the neighbours of the fused qubits. When, instead, the measurement {EL } or {EA } is
166
implemented, the errors on the final state are conditional upon the specific measurement
pattern obtained, e.g. for patterns corresponding to the ideal outcome (φ+ ), the outcome
EA1 would result in a logical error 1/4 of the time, whereas outcome EA2 would result
in an error 5/8 of the time. This suggests that by considering only the EA1 outcome as
successful we can further suppress the deleterious effects of the distinguishability. Such
an approach is a trade-off in the ballistic scheme however, since a decrease in the fusion
success probability will push the scheme back towards the percolation threshold. The
precise nature of this trade-off is a subject for future work.
In the following we will imagine we treat all successful patterns equally and eval-
uate average error rates by considering the averaged post-fusion logical state, ρout =
1 i
P
N i Ki Tr1,2 [E ρin ]Ki , where Ki are the local correcting unitaries performed on the
remaining qubits. Since the pre-fusion states we are considering are diagonal in the
graph-state basis, this can be equivalently expressed as ρout = N1 Tr1,2 [ i K̃i E i K̃i ρin ] =
P
1
N
Tr1,2 [L̄ρin ], where now K̃i are unitaries performed on the measured qubits prior to
measurement (see appendix C.4 for details). This averaging process significantly sim-
plifies the above measurement operators in the case of successful detection patterns to
ĒL = 5(|φ+ ihφ+ |+|φ− ihφ− |)/4 and ĒA = (11|φ+ ihφ+ |+|φ− ihφ− |+3|ψ + ihψ + |+3|ψ − ihψ − |)/8.
6.7 Error rates on Lattice

Putting together the results from the previous section now allows us to calculate the full
output state on the lattice section displayed in Fig. 6.3b as a result of a single distin-
guishable photon and logical error rates on the full lattice. Combining the GHZ states
produced and Bell measurements performed with partially-distinguishable photons from
the previous sections, we can calculate the logical output states corresponding to all pos-
sible outcomes of the fusion gates involved (success or failure for each).
In order to maximise the connectedness of the post-fusion graph states, additional
local rotations are required before and after performing the circuit in 6.3 (see appendix
C.5 for details of these rotations). This results in a breaking of the symmetry of the final
noisy states between two inequivalent lattice sites. However, this effect is small and so
167
a) b) c)
Dist. Indis. Dist.
Indis. Dist. Ancilla Indis.
Dist. Qubit
Figure 6.4: Summary of LOQC with a distinguishable particle. a) the probability of logical
error, the probability of leakage and the probability of successful heralding of the GHZ
generator when using all indistinguishable photons (orange) and a single distinguishable
photon (blue). b) the same for the Boosted fusion gate with additionally the probability
of the failure outcome for indistinguishable photons(orange), a distinguishable logical
photon (blue) and a distinguishable ancilla photon (green). c) Same as a) for the local
patch of the final lattice generated via three GHZ states and six fusion gates, averaged
over all the measurement outcomes.
for simplicity we will just consider the average output state across all lattice sites.
All resulting states are graph-basis diagonal so that in each case the whole state
generation process is equivalent to the application of a Pauli error map parameterised by
, ΛG , applied to the ideally produced graph state |Gi.
These maps can be written in terms of the imperfect state generated by a single
distinguishable particle ρG as
1h i
ΛG (|GihG|) = (1 − 34)|GihG| + 34ηG ρG + O(2 ) (6.28)
N
where G is the graph state produced in the ideal case given the fusion outcomes, ηG is the
ratio of the average success probability of the fusions in the distinguishable case versus
the ideal case and N is a normalisation constant. The state ρG for the case where all
fusion gates are successful is shown in Fig. 6.5.
The simplest figure of merit that can be calculated from these maps is the average
probability of any logical error conditional upon having one distinguishable photon across
all outcome channels. This value is ∼ 0.54 and there is a leakage probability of 0.23
from the fusion gates which leads to the probability of error contributed from a single
lattice site error map scaling as ≈ 14 for small . Fig. 6.4 shows the probabilities of
168
(
Z Z
1 Z
584 + 264 + 29 + 29 + 29 +29
1362
Z Z
Z Z
+ 36 X + 12 Y + 31 Z + 31 Z + 31 Z + 31 Z
Z Z
(
Z Z Z Z
+ 68 + 68 + 52 Z + 52 Z
Z Z Z Z
Figure 6.5: Output state ρG given a single distinguishable particle within the tile shown
in Fig. 6.3. All addition signs refer to an incoherent mixture of the represented graph
states.
logical error, leakage and success/failure for the GHZ generator, boosted fusion and final
lattice generation. We see that the error per distinguishable photon is better for the
Boosted fusion, and therefore final lattice also, than the GHZ generator. This is due to
the conversion of errors into leakage within the measurement circuit.
So far we have only considered errors around one lattice site resulting from a single
bad photon, assuming all its neighbours to be ideal. If we imagine a perfect lattice formed
i
from all successful fusions, then we can apply the error map at each lattice site i ΛG ,
resulting in a global noisy state. From this full lattice error model we can then calculate
the logical error rates at each lattice site. Fig. 6.6 shows this error rate as a function of
the parameter (and 1 − V + O(2 )). Due to the first order error assumption in our input
state, we only investigate the regime in which this assumption is valid ( ≤ 10−3 ). The
probability of a logical error occurring on each qubit scales as ∼ 28 for small .
In the above calculation we have just considered errors at a single lattice site. Thresh-
olds for quantum error correcting codes (QECCs) are often numerically estimated by
assuming such i.i.d local weight-one error models and, as such, are only fully valid for
such models (e.g. for local depolarizing noise models, the surface code can tolerate ≈ 1%
logical error [66]). Although the physical errors considered here are i.i.d on each photon,
due to the use of the fusion gates to entangle non-neighbouring photons we find that
there is an appreciable level of correlated, higher weight errors. This is shown in Fig. 6.6
169
where the probability of logical error on neighbouring lattice sites, next-nearest neighbour
sites in the shown configuration (up to horizontal flip) and both nearest and next-nearest
neighbours (same) is shown. Also depicted are the corresponding error rates for lattice
sites that are separated by at least three bonds. We see that higher weight errors are
significantly more likely (scaling as rather than 2 or 3 ) at these nearby lattice sites
than an independent error model would predict.
Many QECCs are able to tolerate localised regions of noise. To assess the impact of
distinguishability on a QECC it important to consider the error maps which arise beyond
the first-order approximation which lead to higher weight errors spanning neighbouring
sites across the lattice. This can occur when the set of photons used to construct neigh-
bouring lattice sites both contain a bad photon. The resulting error map will not simply
given by ΛiG applied to these neighbouring sites since the two distinguishable particles
can interfere with each other during the secondary fusion stage resulting in their errors no
longer being independent. However, we note that when two photons do meet at a fusion
the maximum weight of the error caused by this event is no higher than when only one
distinguishable photon enters the fusion operation, corresponding to Pauli-Z errors on all
qubits in the neighbourhood of the fused qubits. Therefore, corrections due to photons
meeting at a fusion operation, for a given weight of error, will be of the order of 2 and
hence negligible in the 2 << regime (for completeness appendix C.7 provides an explicit
calculation of the error maps when two distinguishable photons meet at a fusion). Hence
we can say that the rate of an error with a given weight will be given by composing ΛiG s
independently up to higher orders of ε i.e. the probability of weight six errors given by
composing ΛiG is proportional to ε2 and will be correct up to O(ε3 ) such that Ξ ≈ i
i ΛG
where Ξ is the error map for the full lattice.

Finally, we note that our calculations only consider logical errors on the physical
resource state generated. As this is not a regular lattice however, it is expected that
logical qubits would be defined from large subsets of photons within the resource state, so
that they are in a suitable form for implementing standard techniques from measurement-
based quantum computing. This can be done using the renormalisation schemes of Ref.
[119] for example. However, the resulting error rate on the logical qubits may generally
170
1 V + O(✏2 )
10-6 10-5 10-4 10-3
10-3 Z
Z
10-4
Prob(error)
Z
Z
10-5
Z
10-6 Z
Z
Z
10-7 -7
10 10-6 10-5 10-4 10-3
✏
Figure 6.6: Logical error rates on lattice. Probabilities of logical error as a function of
and 1 − V + O(2 ). Blue, probability of logical error on any single lattice site. Orange
solid, probability of logical errors on both neighbouring lattice sites. Orange dashed,
probability of logical errors on both next-nearest neighbours in the configuration shown
(up to a 180◦ rotation). Orange dot-dashed, probability of logical error on both for any
other two lattice sites. Green, probability of logical error on three neighbouring sites in
the configuration shown (up to a 180◦ rotation). Green dot-dashed, probability of logical
errors on three lattice sites for any other set of lattice sites.
be higher, due to the combined errors on the constituent photons.
6.8 Discussion
In this chapter we have shown how to describe, evolve and measure computational states
of partially-distinguishable photons. This has allowed us to develop an error model for
a leading proposal for LOQC. Our approach could also be used to investigate errors in
other applications involving photonic cluster states such as all-optical quantum repeaters
[235, 236].
These calculations suggest that the requirements for single-photon sources to perform
fault-tolerant LOQC are indeed very stringent and challenging when considering the state-
of-the-art. On the other hand, it is to be expected that the per-photon error rate must be
low in LOQC given that tens of photons are consumed to produce a single lattice qubit.
In fact the error rate per distinguishable particle in the input state is lower for producing a
171
final lattice qubit than generating the GHZ state. As noted in section 6.6.3, this is because
using the information from detection patterns helps ‘filter’ distinguishability errors in
the fusion measurements. Although we showed in section 6.4 that such an effect is not
prevalent in the entanglement generation circuits considered in this scheme, it is possible
that using different interferometric configurations may allow for significant improvements
in the error rates. This effect will largely be the focus of the next chapter.
This work also suggests several other future directions. Can the specific features of
this error model be used to develop tailored error correction strategies for LOQC? Finally,
can more general results be obtained by developing a resource theory to understand the
conversion between the indistinguishability of a set of single photon sources and the
maximum achievable dual-rail entanglement?
Statement of Work
This chapter contains joint work with P. Birchall. In particular, the logical-internal rep-
resentation of photonic states, the RSM model for photon sources and the proof of the
bounds in Sec. 6.4 are due to P. Birchall. All calculations of imperfect state preparation,
measurements and logical error rates were performed by myself.
172
Chapter 7
Mitigating Distinguishability Errors
7.1 Introduction
Multiphoton quantum interference is the fundamental physical phenomenon underpinning
many photonic technologies from Boson Sampling [8] to device independent quantum se-
curity [237]. The production of highly indistinguishable photons is therefore crucial for
many applications. We have described both the difficulties in producing highly indis-
tinguishable photons and the errors they can result in. Much of the current work in
improving the indistinguishability of photon sources involves in-depth engineering and
material science considerations. However, an important observation made in the previous
chapter was that the logical errors introduced by partial-distinguishability could be, to
some extent, naturally filtered by conditioning upon specific detection patterns.
In this chapter we explore this concept further, proposing a series of schemes which
can help improve the interference of a set of non-ideal single photons without any direct
manipulation of their internal states. We first introduce a ‘filter’ based upon Hong-Ou-
Mandel interference and bosonic bunching which distills a single purer photon from a set of
noisy ones. We then consider how these ideas can help influence linear optical gate design
and finally show how high timing or spectral resolution in detectors may be exploited to
generate heralded entanglement, even from photons with distinguishable internal states.
174
7. Mitigating Distinguishability Errors
(a) % %0
%
‘0’ ‘1’
(b) % %0
..
.
. ‘0’ ‘p-1’
% ..
%
‘0’
%
‘0’
Figure 7.1: Hong-Ou-Mandel Filtering. a) Two photon case, where bunching of two pho-
tons in the state % is confirmed by a vacuum measurement followed by a heralded subtrac-
tion of one bunched photon resulting in a single photon state %0 . b) p-copy generalisation.
All photons are bunched in a single mode and p − 1 then subsequently subtracted.
7.2 Hong-Ou-Mandel Filtering

The underlying principle by which we observed a filtering of logical errors in the previous
chapter can be simply stated; the detection pattern of a set of photons at the exit of a linear
optical network can provide us with information regarding their degree of interference,
and therefore the distinguishability of their internal states. The simplest case in which
this phenomenon can be observed is for just two photons at a beamsplitter, a HOM setup.
P
Consider a photon in the mixed internal state % = i pi |ξi ihξi | incident at each port
of a beamsplitter. If the two photons are in the same mode |ξi i then they will interfere,
and bunch at the exit with a probability 1, whereas if they are in orthogonal modes then
they will not interfere and will bunch with a probability 1/2. Therefore, if we observe
the photons bunch, then via Bayes rule, they were more likely to be in the same mode.
Updating our description of the photon’s states then increases their purity. This process
can be thought of as filtering the internal states of the photons by conditioning on patterns
that are most consistent with ideal quantum interference.
The scheme that we consider is shown in Fig. 7.1a. By measuring vacuum at the
bottom port of the beamsplitter, we herald the bunching of the photons in the top port
175
in the state
1 X 2 †2 2
X † †

ρbunch = pi a1 [ξi ]|0ih0|a1 [ξi ] + pi pj a1 [ξi ]a1 [ξj ]|0ih0|a1 [ξi ]a1 [ξj ] (7.1)
N i i6=j
where N is a normalisation constant and {|ξi i} with i = 0, ..., m is a basis of internal

states. A single photon is then obtained by a heralded subtraction of a photon from this
mode. This can be achieved using another beamsplitter to split the photons and the
heralded detection of one photon in mode 2, described by the measurement operator F2
%0 = Tr2 F2 U(BS)ρbunch U † (BS)F2

(7.2)
1 X
= pi (pi + 1)a†1 [ξi ]|0ih0|a1 [ξi ] (7.3)
N i
X pi (pi + 1)
= |ξi ihξi | (7.4)
i
P +1
p2i is the purity of %. An eigenvalue pi is then increased if pi > P. The

P
where P = i
1
procedure works with a probability (P + 1)/8 .
When % has a single dominant eigenvalue p0 , two extremal cases can be identified;
when P is maximised and minimised given p0 . The purity is maximised by the state
%max = p0 |ξi ihξi | + (1 − p0 )|ξ1 ihξ1 |, where p00 = p0 (p0 + 1)/2[p0 (p0 − 1) + 1]. The purity
is minimised in the case that there are many additional eigenvectors, all with vanishing
eigenvalues. In this case p00 → p0 (p0 + 1)/(p20 + 1).
When % is almost pure, p0 = 1 − , then both cases tend to the relation p00 = 1 − /2 +
O(2 ), i.e. the probability of the photon being found in a ‘bad’ mode is halved by the
filter.
In principle, this process could be achieved essentially determinstically for high purity
inputs. When the input photons are already close to pure, they bunch almost all the time.
Photon subtraction can then be achieved with a probability approaching 1 in principle
via a repeat-until-success application of beamsplitters with reflectivity → 1 [238]. This
subtraction scheme can be applied adaptively to the states in the two modes after the
1
The scheme would still work with a reduced probability of success by replacing the beamsplitters
with non-1/2 reflectivities
176
initial beamsplitter is applied. Wherever the first photon is detected, the other mode can
then be fully measured. A vacuum detection then leads to a heralding of the remaining
bunched photon. In an integrated device, such an approach could be achieved via ring
resonators with tunable coupling to a single photon detector, e.g. the proposal in Ref.
[239].
The scheme readily generalises to the case in which we have p-copies of % (Fig. 7.1b).
Using p − 1 beamsplitters, all the photons can be heralded to have bunched into a single
mode resulting in the state
m
1 X O ni † ni
ρpbunch = p a [ξi ] |0ih0|a1 [ξi ]ni (7.5)
N n∈N i=0 i 1
p
P
where Np is the set of all m + 1 element tuples (n0 , n1 , ..., nm ) summing to p, i ni = p.
As before, these photons can then be split up. For convenience we will assume that all
√
photons are then antibunched into p modes via the transformation a†1 → j a†j / p
P
m
1 X O ni X † ni X ni
Uρpbunch U † = pi aj [ξi ] |0ih0| aj 0 [ξi ]
N n∈N i=0 j j0
p
m
(11...1) (11...1) 1 X X O ni †
F(12...p) Uρpbunch U † F(12...p) = p a [ξi ]|0ih0|aσ0 (j) [ξi ]. (7.6)
N n∈N σ,σ0 ∈S i=0 i σ(j)
p p
Tracing out all but the first mode then results in the reduced state
m
p−1

(11...1) (11...1) 1 X X Y ni
Tr23...p F(12...p) Uρpbunch U † F(12...p) a†1 [ξk ]|0ih0|a1 [ξk ]

= pi ni !
N n∈N k i=0 n0 , n1 , ..., nk − 1, ..., nm
p
X 1 XY m
0
% = pni i nk |ξk ihξk |
k
N n∈N i=0
p
X
= p0k |ξk ihξk |. (7.7)
k
Again, we can consider the two extremal purity cases for a fixed p0 . The expression for
177
1.0
0.9
0.8
p00
0.7 p=2
p=4
0.6 p = 10
0.5
0.5 0.6 0.7 0.8 0.9 1.0
p0
Figure 7.2: Improvement in dominant eigenvalue for HOM Filter. Dominant eigenvalue
of initial states p0 is plotted against that of the output photon p00 for p = 2, 4 and 10
initial copies of %. Dotted lines correspond to the maximum purity % given p0 , solid lines
to the minimum purity case.
the largest eigenvalue for the maximum purity case %max is

Pp n p−n
n=0 np0 (1 − p0 )
p00 = Pp n p−n + pp−n (1 − p )n )
(7.8)
n=0 n(p0 (1 − p0 ) 0 0
p
p+1

p0 (1 − p0 ) + p0 (p(2p0 − 1) + p0 − 1)
= (7.9)
p(2p0 − 1)(pp+1
0 − (1 − p0 )1+p )
Pd
and for the low purity case, with %min = p0 |ξ0 ihξ0 | + j=1 (1 − p0 )/d|ξj ihξj |,
Pp 0 p−n d+p−n−1
npn0 ( 1−p

n=0 d
) p−n
p00 = Pp d+p−n−1 Pp Pp−nk n0 1−p0 p−n0 d−1+p−n0 −nk −1
n 1−p0 p−n

n=0 np0 ( d ) p−n
+ d nk =0 nk n0 =0 p0 d p−n0 −nk
(7.10)
Pp n p−n
n=0 np0 (1 − p0 ) /(p − n)!
lim p00 = Pp n p−1 . (7.11)
p0 )p−n /(p − n)! + n=0 pn0 (1 − p0 )p−n /(p
P
n=0 np0 (1 − − 1 − n)!
d→∞
These expressions for p = 2, 4 and 10 are plotted in Fig. 7.2. In the high purity limit,
p0 = 1 − , both cases tend to p00 = 1 − /p + O(2 ).
178
a) b) c)
1/2 0.82
1 1/3 1 0.15
1/2 1/2 1/2 1/2

1
1/3 0.15
1/2 0.18
1
Figure 7.3: Linear optical networks for heralded bell generation. a) B8 b) B6 c)

B(0.18, 0.152) optimised for √
distinguishability
√ error tolerance. The phase shifter is a
√ √
π/2, beamsplitters are {{ η, 1 − η}, { 1 − η, − η}}.
7.3 The Distinguishability Tolerance of Bell State Gen-

erators
Although the HOM filtering scheme introduced in the previous section can in principle
produce arbitrarily indistinguishable photons, many photons must be consumed to achieve
this. Using the same principle that detection patterns which have a higher probability of
occurring with indistinguishable particles produce higher quality states, we can think of
the perhaps more practical approach of designing gates themselves which are naturally
more tolerant to partial-distinguishability errors. As a first example of this approach
we consider the two heralded Bell state generator circuits described in this thesis; B6 ,
introduced in Sec. 4.4.2, and B8 , considered in Sec. 6.4.
The two schemes are depicted together in Fig. 7.3a and b. One difference between
these circuits that can be noted is that in B8 there are no closed interferometers and for
the cases when heralding is successful there is no HOM-style interference (i.e. 2 photons
meeting at a beamsplitter and bunching) and only interference between different overall
trajectories of the photons occurs. In contrast, the B6 scheme is built around a central
closed interferometer and its operation is highly dependent upon HOM-style bunching.
B6 is in fact a special case of a two-parameter family of unitaries B(r1 , r2 ) which are
all Bell generator circuits with differing probabilities of success and error-tolerance. When
the qubits are defined by the first two and last two modes and photons are inputted in
modes 1,3,4 and 5, these transfer matrices are parameterised by
179
(a) C(⇢L ) (b) C(⇢L )

1.0 1.000
0.8 0.995
0.6 0.990
0.4 0.985
0.2 0.980
0.0
0.5 0.6 0.7 0.8 0.9 1.0 0.990 0.992 0.994 0.996 0.998 1.000
V V
(c) psuccess
0.20
max purity max purity
B8 B6
min purity min purity
0.15
max purity
B(0.18, 0.15)
0.10 min purity
0.05
0.00
0.5 0.6 0.7 0.8 0.9 1.0
V
Figure 7.4: Error tolerance of Bell generators. a) Concurrence of resulting logical state
produced from circuits B8 (black) B6 (red), and B(0.18,0.152) (blue) and input internal
states %. The two limiting cases which maximise (dotted) and minimise (solid) the purity
of % given p0 are shown, and intermediate cases shaded. b) Shows the same plot in the
linear region of high purity. c) Shows the probability of success in these schemes.
 √
√

1 − r1 r1 0 0 0 0
 √ √
 √ √ √
1−r
√
1−r

r r − 1 − r1 r2 √ 2 √ 2 0 0

 √ 1√ 2
 2 2

√ √ √ √ √ √
i √
 i r1 1−r2 √ r1 1−r2

− i 1−r√1 2 1−r2 1
− 12 − 2i 1−r√
1 1−r2

−

 √
2 2 2
r2 r2 √
2 2

B(r1 , r2 ) = 
 i√r1 √1−r2 √ √ √ √ √ √ √
i √

r1 1−r2
− i 1−r√1 2 1−r2 − 12 − i 1
− 1−r√1 2 1−r2 

− − √2

 √
2 2
r2 2 2
r2
√ √ √ √ √ √ 
 
1−r 1−r
− √2 2 − r1 r2 − 1 − r1 r2 

 0 0 √ 2
2
√ √
 
0 0 0 0 1 − r1 − r1
(7.12)
The probability of success in the ideal case is given by ps = 2r1 r2 (1 − r1 )(1 − r2 )2 and is
maximised by B6 , where r1 = 1/2, r2 = 1/3 and ps = 2/27.
Using the formalism developed in the previous chapter, we can evaluate the output
180
state of these circuits in the RS model where all the input photons are described with i.i.d
mixed states % in their additional degrees of freedom (see appendix D for the full state).
Fig. 7.4a,b plot the concurrence of the resulting Bell states generated via B8 , B6 (1/2, 1/3)
and a numerically optimised B(0.18, 0.152) as a function of the pairwise HOM visibilities.
We find that the B6 circuit is more error-tolerant than B8 , and can be further im-
proved by using B(0.18, 0.152). Fig. 7.4c shows the probability of success of these schemes
demonstrating a clear trade-off between efficiency and entanglement. Given the strict
logical error budget required to achieve fault-tolerance, such improvements, although not
dramatic, may still be worth trading for lower efficiencies in some contexts.
7.4 Generating entanglement from distinguishable pho-

tons
Finally, our entire discussion of partial-distinguishability thus far has assumed that the
single photon detectors cannot resolve the additional degrees of freedom. This assumption
is valid in many circumstances, when for instance the spectral response of the detector
is uniform over the photon’s bandwidth and their coherence length is shorter than the
detector’s temporal resolution. However, we now consider what can be achieved if we
instead use detectors which can resolve additional information.
We first note that obtaining this additional information alone does not solve the prob-
lem of distinguishability errors. As an example we can once again consider the Bell
state generator B8 . The state produced by this circuit with input photons with spectral
functions {ψi (ω)} in the logical-internal basis is
1h
|φ({ψi })i = |00i ⊗ |ψ1 ψ2 ψ3 ψ4 i + |ψ1 ψ3 ψ2 ψ4 i + |01i ⊗ |ψ1 ψ2 ψ4 ψ3 i − |ψ1 ψ4 ψ2 ψ3 i
16
i
+ |10i ⊗ |ψ2 ψ1 ψ3 ψ4 i − |ψ2 ψ3 ψ1 ψ4 i + |11i ⊗ |ψ2 ψ1 ψ4 ψ3 i + |ψ2 ψ4 ψ1 ψ3 i
(7.13)
In what follows, we will analyse the scenario in which the photons are produced with
181
Gaussian spectral functions
1 1/4 h −(ω − ω )2 i
i
ψi (ω) = 2
exp 2
(7.14)
2π∆ω 4∆ω
whose central frequencies are separated by much more than their bandwidth |ωi − ωj | >>
∆ω so that the condition hψi |ψj i ∼ 0 is met, i.e. the photons are all distinguishable.
Detectors which can measure the frequency of measured photons ω 0 to within a spec-
tral window of width Ω then produce a final state dependent upon overlaps of the
R ω0 +Ω/2
form A(ψi , ψj )|ω0 = ω0 −Ω/2 ψi (ω)ψj∗ (ω)dω. When Ω |ωi − ωj |, then A(ψi , ψj )|ω0 ∼
|hψi |ψj i|2 ∼ δij and the state tends to that calculated in Eq. (6.9) with all visibilities set
to zero, i.e. a maximally mixed state ρL = 1/4.
When we have high spectral resolution however, Ω |ωi − ωj |, then the four photons
can be distinguished by the measurement, A(ψi , ψj )|ω0 ∼ δij δωi ω0 , and the output is pro-
jected onto just one of the terms in Eq. (7.13). For instance, if the photons are detected
with frequencies (ω1 , ω2 , ω3 , ω4 ) respectively, then the logical state is collapsed to |00i.
The state in this case then remains pure, however, no dual rail entanglement has been
generated.
One may be tempted to conclude that the overlaps of the internal states entirely deter-
mine the amount of entanglement that can be generated from them via transformations
on the paths and detection. We now show that this is not always the case.
If we take the same set of photons, but now express their internal state in the temporal
domain via a Fourier transform, we find they have the form
2∆ω 2 1/4
exp − ∆ω 2 t2 exp iωi t

Ψi (t) = F[ψi (ω)](t) = (7.15)
π
where t is defined relative to a common emission time of all the photons. We see therefore
that all the photons have the same temporal envelope |Ψi (t)| = |Ψj (t)| = |Ψ(t)| ∀i, j. If
our detectors can detect photons within a timing window T , then the relevant overlaps
R t0 +T /2
which determine the output state become A(Ψi , Ψj )|t0 = t0 −T /2 Ψi (t)Ψ∗j (t)dt.
When T >> ∆t, where ∆t is the temporal width of the photons, we again find that all
overlaps A(Ψi , Ψj ) ≈ δij and the logical output state is again ρL = 1/4. However, when
182
T << ∆t, the overlaps become A(Ψi , Ψj )|t0 ∼ |Ψ(t0 )|2 exp[i(ωi − ωj )t0 ] and so interference
between the spectrally distinguishable photons can be observed. This is an example of
quantum erasure; the recovery of quantum interference by the removal of which-path (or
in this case which-frequency) information.
In the limit of infinite timing precision, and detecting the photons at times (t1 , t2 , t3 , t4 )
respectively, produces the pure state output
1 h i(t1 ω1 +t4 ω4 ) i(t2 ω2 +t3 ω3 )

|φ({ti })i = e {e + ei(t2 ω3 +t3 ω2 ) }|00i
N
+ei(t1 ω1 +t4 ω3 ) {ei(t2 ω2 +t3 ω4 ) − ei(t2 ω4 +t3 ω2 ) }|01i
+ei(t1 ω2 +t4 ω4 ) {ei(t2 ω1 +t3 ω3 ) − ei(t2 ω3 +t3 ω1 ) }|10i

i
+ei(t1 ω2 +t4 ω3 ) {ei(t2 ω1 +t3 ω4 ) + ei(t2 ω4 +t3 ω1 ) }|11i . (7.16)
Using the simple formula for the concurrence of a two-qubit pure state (Eq. (2.36)),
1
C(|φ({ti })i) = (7.17)
4 − 4 cos[(t3 − t2 )ω] + cos[2(t3 − t2 )ω]
for central frequencies given by ωk = ω0 + kω. Although the specific state produced is
dependent upon all detection times, entanglement is always generated and the amount
is determined just by the time difference t3 − t2 . When (t3 − t2 )ω = 2nπ, the state is
maximally entangled C = 1, whereas when (t3 − t2 )ω = (2n − 1)π the entanglement is
minimised, C = 1/9.
Taking into account a finite detection resolution, this entanglement is degraded as the
window T is increased, as shown in Fig. 7.5. We note that the same situation can be set
up conversely, such that temporally distinguishable photons can be interfered via high
precision spectral measurements.
The timing jitter of current single-photon detectors can be in the 10’s of picoseconds
[240], which may already make this approach practical for photons with lengths of 100’s of
picoseconds produced in current integrated experiments [241]. High resolution frequency
detection could be achieved via wavelength-division multiplexing into many spatial modes
183
(a) (b)
C(ρ) C(ρ)
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 (t3 -t2 )ω 0.0 T/ω

0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 7.5: Entanglement generation from spectrally distinguishable photons via temporal
resolution. (a) Concurrence of the resulting two qubit pure state as a function of the
timing difference of the ancilla photons (t3 − t2 )ω when perfectly time resolving detectors
are used to measure photons with central frequencies separated by ω. (b) Concurrence of
resulting two qubit state when all photons are detected simultaneously by detectors with
a temporal resolution of T .
which are each fed to a detector.
7.5 Discussion
As we have seen in previous chapters, photon distinguishability is a major source of logical
error in current quantum photonics experiments and is likely to be so for large-scale
photonic architectures as well. Producing highly indistinguishable photons from photon
sources is challenging without full control and manipulation of all degrees of freedom of
the photons. Here we have proposed a number of schemes which use just manipulations
of the paths and photon detection in order to mitigate the initial distinguishability of a
set of photons.
Using a simple circuit which bunches and then anti-bunches a set of photons, we can
distill a single purer photon from many mixed photons. We went on to show that certain
circuits which generate heralded entanglement from single photons can be more and less
error-tolerant in the prescence of partial-distinguishability errors. Finally we showed that
high resolution detectors can erase the distinguishability of photons in conjugate variable
degrees of freedom.
There are many open questions and avenues for future work that are suggested by
184
this work. It would be interesting to investigate the possibility of similar schemes to

the HOM filter approach which achieve higher success probabilties or do not require the
heralded subtraction stage. One can also consider the case when the input states are
known i.e. ni=1 %i instead of %⊗n . In this case it seems likely that the best circuit will
N
be one which uses this additional information (e.g. by the optimal pairing of photons at
beamsplitters).
Our discussion of error-tolerant gates naturally leads to a further investigation in which
new circuit designs are searched for by optimising the maximal entanglement given a set
of partially-distinguishable photons, rather than efficiency.
Finally, an experimental exploration of using high resolution detectors and long or
broadband photons in photonic information processing tasks such as LOQC or Boson
Sampling (where such a scheme has been proposed [242]) may reveal the practicality and
scalability such protocols.
Statement of Work
The idea of using quantum interference to filter distinguishability errors was conceived
by myself and P. Birchall. The HOM filter in 7.1a was conceived by P. Birchall and the
results for the many copy generalisation are due to myself. Using multi-pair emission in
sources to purify photons was conceived by P. Birchall and developed by myself. The
comparison between error tolerances of Bell generator circuits is due to myself. The
analysis of detector timing resolution in Bell state generation is due to myself.
185
Chapter 8
Conclusion
8.1 Key Results

In this thesis we have explored the opportunities and challenges associated with realising
practical photonic quantum technologies. Here I summarise the key results:
1. Prototyping quantum logic gates with universal linear optics.

Using the first universally reconfigurable linear optical processor we performed a
series of photonic logic gates. We performed process tomography on a collection of
single qubit gates and obtained the highest reported process fidelity for a postse-
lected CNOT demonstrating high performance across hundreds of circuit configura-
tions. We went on to present the first experimental heralded multiphoton quantum
gates in integrated optics. We implemented a KLM CNOT gate demonstrating
measurement-induced nonlinearity between photons. Introducing a new circuit for
Bell state generation, we performed a series of measurements to verify the produc-
tion of a heralded entangled state. Finally, we characterised the transfer matrices
of these gates to show that errors in these experiments are dominated by imperfect
sources and detectors.
2. Simulating the quantum dynamics of molecular vibrations with inte-

grated photonics.
We proposed reconfigurable integrated photonics as a platform for the quantum
186
8. Conclusion
simulation of molecular vibrational dynamics. We performed a wide array of ex-

periments to highlight the versatility and scope of this approach. A series of four
atom molecules were simulated and the coherent and incoherent energy transfer
properties of NMA were explored. Thermalisation and anharmonic evolution for
H2 O were simulated using additional optical modes and photons, and finally the de-
vice was used in an adaptive feedback loop to simulate the optimisation of selective
dissociation in ammonia.
3. An error model for linear optical quantum computing with partially-

distinguishable photons.
We showed how to evolve, measure and obtain the effective logical state from a
collection of partially-distinguishable photons in an optical quantum computer. We
used this formalism to obtain the logical states generated and measurements per-
formed by a series of standard linear optical gates for different distinguishability
models. Applying these techniques to a large-scale architecture we were then able
to calculate the error maps and logical error rates resulting from distinguishable
photons for the first time.
4. Schemes for improving the quantum interference of imperfect photon

sources.
We proposed several novel protocols for mitigating the errors introduced by photon
distinguishability without manipulating the photons’ internal states. We first in-
troduced the Hong-Ou-Mandel filter, where the heralded bunching of photons can
distill a single purer photon from many impure ones. We demonstrated that differ-
ent circuits which perform the same logical function can have different tolerances
to distinguishability errors and finally that entangled states can be heralded from a
set of completely distinguishable photons if high resolution detection is possible in
conjugate degrees of freedom.
187
8. Conclusion
8.2 Outlook
The development of universally reconfigurable linear optical devices has introduced a
new paradigm in photonic quantum information processing. New experiments no longer
require a lengthy cycle of design, fabrication and characterisation. With the further
integration of photon sources [146, 160] and single photon detectors [240, 243], along
with the increased component density and capabilities arising from the classical silicon
photonics industry [244, 245], large-scale, reconfigurable linear optical processors have
become a near-term prospect. The natural coherence and complexity of these devices
promise new possibilities for non-universal models of quantum computation, such as the
molecular simulation platform we have introduced here.
Many important questions remain unsettled for this route to supra-classical photonic
computation. Considering the apparent classical simulability limit for Boson sampling
[138, 139], such a quantum device would require the near-perfect interference of order
50 photons in a universal linear optical device with millions of highly precise, coherence-
preserving tunable optical components all with per-photon loss rates significantly better
than current experiments [138]. One may then legitimately question whether the con-
struction of such a device would not be more challenging than universal LOQC which can
be achieved with a constant-depth architecture, and interference only between photons
which are near-neighbours. Another more fundamental open question is whether it will
be possible to achieve a genuine quantum advantage for a ‘useful’ computational task
without fault-tolerance. All the leading approaches to demonstrating quantum computa-
tional advantage consist of sampling from distributions for which individual probabilties
cannot be accurately estimated in polytime. If we wish to gain useful information from
the physical simulations presented in Chapter 5, the output distribution will therefore re-
quire some structure or sparseness which may in turn be exploitable by classical methods
[246].
Given some of these difficulties, one potential avenue for future analog photonic com-
putation may move towards more general continuous variable states [128], measurements
[129] and the development of programmable nonlinear optical components. The pursuit
188
8. Conclusion
of classical nonlinear optical devices to solve computationally hard problems [247, 248]
also suggests a potential new shortcut to large-scale quantum photonics may be to push
existing complex classical optical systems incrementally into the quantum domain.
The rapid progress in integrated photonic platforms together with improvements in the
theoretical resource overheads [120, 121, 249, 250] has reignited the pursuit of universal
linear optical quantum computing. A full understanding of the relationship between
physical imperfections in photonic devices and logical error rates is therefore a crucial
priority. Comparing our calculations for partial-distinguishability to current technology
suggests that obtaining the far-below-threshold logical error rates desirable for practical
quantum computing may be difficult to achieve. This is particularly problematic for
the photonic architecture, where qubit loss and probabilistic gates already result in high
resource costs [251] versus matter-based platforms.
However, we have also shown that leveraging capabilities in circuitry and detection can
help mitigate these errors, reducing the hardware requirements on photon sources. More
generally, in contrast to matter systems coupled to thermal environments, the coherence
properties of optical states often allow for the conversion of logical errors into detectable
loss. The key question then becomes whether an architecture designed to optimally trade-
off errors and loss can result in a practical device blueprint.
The future success of both approaches to photonic quantum computation will therefore
rely on the combination of engineering expertise to drive component-level improvements
together with insights from quantum physics to develop the architectures and applications
which maximise their impact.
189
Bibliography
[1] A. Einstein, “Uber einen die erzeugung und verwandlung des lichtes betreffenden
heuristischen gesichtspunkt,” Annalen der Physik 322, 132–148 (1905).
[2] B. Hensen, H. Bernien, A. E. Dreau, A. Reiserer, N. Kalb, M. S. Blok, J. Ruitenberg,

R. F. L. Vermeulen, R. N. Schouten, C. Abellan, W. Amaya, V. Pruneri, M. W.
Mitchell, M. Markham, D. J. Twitchen, D. Elkouss, S. Wehner, T. H. Taminiau, and
R. Hanson, “Loophole-free bell inequality violation using electron spins separated
by 1.3 kilometres,” Nature 526, 682–686 (2015).
[3] Lynden K. Shalm, Evan Meyer-Scott, Bradley G. Christensen, Peter Bierhorst,

Michael A. Wayne, Martin J. Stevens, Thomas Gerrits, Scott Glancy, Deny R.
Hamel, Michael S. Allman, Kevin J. Coakley, Shellee D. Dyer, Carson Hodge,
Adriana E. Lita, Varun B. Verma, Camilla Lambrocco, Edward Tortorici, Alan L.
Migdall, Yanbao Zhang, Daniel R. Kumor, William H. Farr, Francesco Marsili,
Matthew D. Shaw, Jeffrey A. Stern, Carlos Abellán, Waldimar Amaya, Valerio
Pruneri, Thomas Jennewein, Morgan W. Mitchell, Paul G. Kwiat, Joshua C. Bien-
fang, Richard P. Mirin, Emanuel Knill, and Sae Woo Nam, “Strong loophole-free
test of local realism,” Phys. Rev. Lett. 115, 250402 (2015).
[4] Marissa Giustina, Marijn A. M. Versteegh, Sören Wengerowsky, Johannes Hand-

steiner, Armin Hochrainer, Kevin Phelan, Fabian Steinlechner, Johannes Kofler,
Jan-Åke Larsson, Carlos Abellán, Waldimar Amaya, Valerio Pruneri, Morgan W.
Mitchell, Jörn Beyer, Thomas Gerrits, Adriana E. Lita, Lynden K. Shalm, Sae Woo
Nam, Thomas Scheidl, Rupert Ursin, Bernhard Wittmann, and Anton Zeilinger,
“Significant-loophole-free test of bell’s theorem with entangled photons,” Phys. Rev.
Lett. 115, 250401 (2015).
[5] Alberto Politi, Martin J. Cryan, John G. Rarity, Siyuan Yu, and Jeremy L. O’Brien,
“Silica-on-silicon waveguide quantum circuits,” Science 320, 646–649 (2008).
[6] Jeremy L. O’Brien, Akira Furusawa, and Jelena Vuckovic, “Photonic quantum
technologies,” Nat Photon 3, 687–695 (2009).
[7] Damien Bonneau, Joshua W. Silverstone, and Mark G. Thompson, “Silicon quan-
tum photonics,” in Silicon Photonics III: Systems and Applications (Springer Berlin
Heidelberg, Berlin, Heidelberg, 2016) pp. 41–82.
190
BIBLIOGRAPHY
[8] Scott Aaronson and Alex Arkhipov, “The computational complexity of linear op-
tics,” in Proceedings of the Forty-third Annual ACM Symposium on Theory of Com-
puting, STOC ’11 (ACM, New York, NY, USA, 2011) pp. 333–342.
[9] Maximilian Schlosshauer, Johannes Kofler, and Anton Zeilinger, “A snapshot of

foundational attitudes toward quantum mechanics,” Stud. Hist. Phil. Mod. Phys.
44, 222–230 (2013).
[10] Christoph Sommer, “Another survey of foundational attitudes towards quantum

mechanics,” (2013), arXiv:1303.2719 .
[11] Travis Norsen and Sarah Nelson, “Yet another snapshot of foundational attitudes
toward quantum mechanics,” (2013), arXiv:1306.4646 .
[12] Hugh Everett, “‘relative state’ formulation of quantum mechanics,” Rev. Mod. Phys.
29, 454–462 (1957).
[13] G. C. Ghirardi, A. Rimini, and T. Weber, “Unified dynamics for microscopic and
macroscopic systems,” Phys. Rev. D 34, 470–491 (1986).
[14] David Bohm, “A suggested interpretation of the quantum theory in terms of ”hid-
den” variables. i,” Phys. Rev. 85, 166–179 (1952).
[15] Samson Abramsky and Bob Coecke, “A categorical semantics of quantum proto-
cols,” in Proceedings of the 19th Annual IEEE Symposium on Logic in Computer
Science: LICS 2004 (IEEE Computer Society Press, 2004) pp. 415–425.
[16] Christopher A. Fuchs, N. David Mermin, and Rudiger Schack, “An introduction to
qbism with an application to the locality of quantum mechanics,” American Journal
of Physics 82, 749–754 (2014).
[17] D. Hanneke, S. Fogwell, and G. Gabrielse, “New measurement of the electron

magnetic moment and the fine structure constant,” Phys. Rev. Lett. 100, 120801
(2008).
[18] David Mermin, “What’s wrong with this pillow?” Physics Today 42, 9 (1989).
[19] C. E. Shannon, “A mathematical theory of communication,” Bell System Technical

Journal 27, 379–423 (1948).
[20] Ryszard Horodecki, Pawel Horodecki, Michal Horodecki, and Karol Horodecki,
“Quantum entanglement,” Rev. Mod. Phys. 81, 865–942 (2009).
[21] Martin B. Plenio and Shashank Virmani, “An introduction to entanglement mea-
sures,” Quantum Info. Comput. 7, 1–51 (2007).
[22] Nicolas Brunner, Daniel Cavalcanti, Stefano Pironio, Valerio Scarani, and Stephanie
Wehner, “Bell nonlocality,” Rev. Mod. Phys. 86, 419–478 (2014).
191
BIBLIOGRAPHY
[23] A. Einstein, B. Podolsky, and N. Rosen, “Can quantum-mechanical description of

physical reality be considered complete?” Phys. Rev. 47, 777–780 (1935).
[24] Albert Einstein, “Physik und realitt,” Journal of the Franklin Institute 221, 313 –
347 (1936).
[25] J. S. Bell, “On the Einstein-Podolsky-Rosen paradox,” Physics 1, 195–200 (1964).
[26] John F. Clauser, Michael A. Horne, Abner Shimony, and Richard A. Holt, “Pro-
posed experiment to test local hidden-variable theories,” Phys. Rev. Lett. 23, 880–
884 (1969).
[27] Alain Aspect, Philippe Grangier, and Gérard Roger, “Experimental realization of
einstein-podolsky-rosen-bohm gedankenexperiment: A new violation of bell’s in-
equalities,” Phys. Rev. Lett. 49, 91–94 (1982).
[28] Reinhard F. Werner, “Quantum states with einstein-podolsky-rosen correlations

admitting a hidden-variable model,” Phys. Rev. A 40, 4277–4281 (1989).
[29] Antonio Acı́n, Richard Gill, and Nicolas Gisin, “Optimal bell tests do not require
maximally entangled states,” Phys. Rev. Lett. 95, 210402 (2005).
[30] Christopher Granade, Joshua Combes, and D G Cory, “Practical bayesian tomog-
raphy,” New Journal of Physics 18, 033024 (2016).
[31] Daniel F. V. James, Paul G. Kwiat, William J. Munro, and Andrew G. White,
“Measurement of qubits,” Phys. Rev. A 64, 052312 (2001).
[32] Stephen M. Barnett and Sarah Croke, “Quantum state discrimination,” Adv. Opt.
Photon. 1, 238–278 (2009).
[33] Michael A. Nielsen and Isaac L. Chuang, Quantum Computation and Quantum
Information: 10th Anniversary Edition, 10th ed. (Cambridge University Press, New
York, NY, USA, 2011).
[34] Alexei Gilchrist, Nathan K. Langford, and Michael A. Nielsen, “Distance measures
to compare real and ideal quantum processes,” Phys. Rev. A 71, 062310 (2005).
[35] D. H. Mahler, Lee A. Rozema, Ardavan Darabi, Christopher Ferrie, Robin Blume-
Kohout, and A. M. Steinberg, “Adaptive quantum state tomography improves
accuracy quadratically,” Phys. Rev. Lett. 111, 183601 (2013).
[36] C. A. Fuchs and J. van de Graaf, “Cryptographic distinguishability measures for

quantum-mechanical states,” IEEE Trans. Inf. Theor. 45, 1216–1227 (2006).
[37] Michael A Nielsen, “A simple formula for the average gate fidelity of a quantum
dynamical operation,” Physics Letters A 303, 249 – 252 (2002).
[38] E. Knill, D. Leibfried, R. Reichle, J. Britton, R. B. Blakestad, J. D. Jost, C. Langer,
192
BIBLIOGRAPHY
R. Ozeri, S. Seidelin, and D. J. Wineland, “Randomized benchmarking of quantum

gates,” Phys. Rev. A 77, 012307 (2008).
[39] A. Yu. Kitaev, A. H. Shen, and M. N. Vyalyi, Classical and Quantum Computation
(American Mathematical Society, Boston, MA, USA, 2002).
[40] John Watrous, “Semidefinite programs for completely bounded norms,” Theory of
Computing 5, 217–238 (2009).
[41] A. M. Turing, “On computable numbers, with an application to the entschei-

dungsproblem,” Proceedings of the London Mathematical Society s2-42, 230–265
(1937).
[42] Greg Aloupis, Erik D. Demaine, Alan Guo, and Giovanni Viglietta, “Classic nin-
tendo games are (computationally) hard,” in Fun with Algorithms: 7th International
Conference, FUN 2014, Lipari Island, Sicily, Italy, July 1-3, 2014. Proceedings,
edited by Alfredo Ferro, Fabrizio Luccio, and Peter Widmayer (Springer Interna-
tional Publishing, Cham, 2014) pp. 40–51.
[43] R.P. Feynman, “Simulating Physics with Computers,” International Journal of The-
oretical Physics 21, 467–488 (1982).
[44] D. Deutsch, “Quantum theory, the church-turing principle and the universal quan-
tum computer,” Proceedings of the Royal Society of London A: Mathematical, Phys-
ical and Engineering Sciences 400, 97–117 (1985).
[45] A. Y. Kitaev, “Quantum computations: algorithms and error correction,” Russ.

Math. Surv. 52, 1191 (1997).
[46] Seth Lloyd, “Almost any quantum logic gate is universal,” Phys. Rev. Lett. 75,
346–349 (1995).
[47] David Deutsch and Richard Jozsa, “Rapid solution of problems by quantum com-
putation,” Proceedings of the Royal Society of London A: Mathematical, Physical
and Engineering Sciences 439, 553–558 (1992).
[48] Ethan Bernstein and Umesh Vazirani, “Quantum complexity theory,” SIAM Journal
on Computing 26, 1411–1473 (1997).
[49] Daniel R. Simon, “On the power of quantum computation,” SIAM Journal on Com-
puting 26, 1474–1483 (1997).
[50] Peter W. Shor, “Polynomial-time algorithms for prime factorization and discrete
logarithms on a quantum computer,” SIAM Review 41, 303–332 (1999).
[51] R. L. Rivest, A. Shamir, and L. Adleman, “A method for obtaining digital signa-
tures and public-key cryptosystems,” Commun. ACM 21, 120–126 (1978).
[52] “Quantum algorithm zoo,” .
193
BIBLIOGRAPHY
[53] Lov K. Grover, “A fast quantum mechanical algorithm for database search,” in Pro-
ceedings of the Twenty-eighth Annual ACM Symposium on Theory of Computing,
STOC ’96 (ACM, New York, NY, USA, 1996) pp. 212–219.
[54] Seth Lloyd, “Universal quantum simulators,” Science 273, 1073–1078 (1996).
[55] Dave Wecker, Bela Bauer, Bryan K. Clark, Matthew B. Hastings, and Matthias
Troyer, “Gate-count estimates for performing quantum chemistry on small quantum
computers,” Phys. Rev. A 90, 022305 (2014).
[56] David Poulin, Matthew B. Hastings, Dave Wecker, Nathan Wiebe, Andrew C.
Doberty, and Matthias Troyer, “The trotter step size required for accurate quantum
simulation of quantum chemistry,” Quantum Info. Comput. 15, 361–384 (2015).
[57] Dominic W. Berry, Andrew M. Childs, Richard Cleve, Robin Kothari, and
Rolando D. Somma, “Simulating hamiltonian dynamics with a truncated taylor
series,” Phys. Rev. Lett. 114, 090502 (2015).
[58] “https://www.allaboutcircuits.com/textbook/digital/chpt-3/logic-signal-voltage-
levels/,” .
[59] W. K. Wootters and W. H. Zurek, “A single quantum cannot be cloned,” Nature

299, 802–803 (1982).
[60] Peter W. Shor, “Scheme for reducing decoherence in quantum computer memory,”
Phys. Rev. A 52, 2493–2496 (1995).
[61] D. Aharonov and M. Ben-Or, “Fault-tolerant quantum computation with constant

error,” in Proceedings of the Twenty-ninth Annual ACM Symposium on Theory of
Computing, STOC ’97 (ACM, New York, NY, USA, 1997) pp. 176–188.
[62] Emanuel Knill, Raymond Laflamme, and Wojciech H. Zurek, “Resilient quantum
computation: error models and thresholds,” Proceedings of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences 454, 365–384 (1998).
[63] Bryan Eastin and Emanuel Knill, “Restrictions on transversal encoded quantum
gate sets,” Phys. Rev. Lett. 102, 110502 (2009).
[64] Sergey Bravyi and Alexei Kitaev, “Universal quantum computation with ideal clif-
ford gates and noisy ancillas,” Phys. Rev. A 71, 022316 (2005).
[65] Panos Aliferis and John Preskill, “Fibonacci scheme for fault-tolerant quantum
computation,” Phys. Rev. A 79, 012332 (2009).
[66] R Raussendorf, J Harrington, and K Goyal, “Topological fault-tolerance in cluster

state quantum computation,” New Journal of Physics 9, 199 (2007).
[67] Clare Horsman, Austin G Fowler, Simon Devitt, and Rodney Van Meter, “Surface
code quantum computing by lattice surgery,” New Journal of Physics 14, 123011
194
BIBLIOGRAPHY
(2012).
[68] Austin G. Fowler, Matteo Mariantoni, John M. Martinis, and Andrew N. Cleland,
“Surface codes: Towards practical large-scale quantum computation,” Phys. Rev.
A 86, 032324 (2012).
[69] Daniel Gottesman, “The Heisenberg representation of quantum computers,” in

Group theoretical methods in physics. Proceedings, 22nd International Colloquium,
Group22, ICGTMP’98, Hobart, Australia, July 13-17, 1998 (1998) pp. 32–43,
arXiv:quant-ph/9807006 .
[70] Robert Raussendorf and Hans J. Briegel, “A one-way quantum computer,” Phys.
Rev. Lett. 86, 5188–5191 (2001).
[71] Charles H. Bennett, Gilles Brassard, Claude Crépeau, Richard Jozsa, Asher Peres,
and William K. Wootters, “Teleporting an unknown quantum state via dual classical
and einstein-podolsky-rosen channels,” Phys. Rev. Lett. 70, 1895–1899 (1993).
[72] Xiao-Song Ma, Thomas Herbst, Thomas Scheidl, Daqing Wang, Sebastian
Kropatschek, William Naylor, Bernhard Wittmann, Alexandra Mech, Johannes
Kofler, Elena Anisimova, Vadim Makarov, Thomas Jennewein, Rupert Ursin, and
Anton Zeilinger, “Quantum teleportation over 143 kilometres using active feed-
forward,” Nature 489, 269–273 (2012).
[73] Ji-Gang Ren, Ping Xu, Hai-Lin Yong, Liang Zhang, Sheng-Kai Liao, Juan Yin, Wei-
Yue Liu, Wen-Qi Cai, Meng Yang, Li Li, Kui-Xing Yang, Xuan Han, Yong-Qiang
Yao, Ji Li, Hai-Yan Wu, Song Wan, Lei Liu, Ding-Quan Liu, Yao-Wu Kuang,
Zhi-Ping He, Peng Shang, Cheng Guo, Ru-Hua Zheng, Kai Tian, Zhen-Cai Zhu,
Nai-Le Liu, Chao-Yang Lu, Rong Shu, Yu-Ao Chen, Cheng-Zhi Peng, Jian-Yu
Wang, and Jian-Wei Pan, “Ground-to-satellite quantum teleportation,” (2017),
arXiv:1707.00934 .
[74] Robert Raussendorf, Daniel E. Browne, and Hans J. Briegel, “Measurement-based

quantum computation on cluster states,” Phys. Rev. A 68, 022312 (2003).
[75] D. Gross and J. Eisert, “Novel schemes for measurement-based quantum computa-
tion,” Phys. Rev. Lett. 98, 220503 (2007).
[76] D. Gross, S. T. Flammia, and J. Eisert, “Most quantum states are too entangled
to be useful as computational resources,” Phys. Rev. Lett. 102, 190501 (2009).
[77] Michael J. Bremner, Caterina Mora, and Andreas Winter, “Are random pure states
useful for quantum computation?” Phys. Rev. Lett. 102, 190502 (2009).
[78] Edward Farhi, Jeffrey Goldstone, Sam Gutmann, and Michael Sipser, “Quantum
computation by adiabatic evolution,” (2000), arXiv:quant-ph/0001106 .
[79] Dorit Aharonov, Wim van Dam, Julia Kempe, Zeph Landau, Seth Lloyd, and
195
BIBLIOGRAPHY
Oded Regev, “Adiabatic quantum computation is equivalent to standard quantum

computation,” SIAM Journal on Computing 37, 166–194 (2007).
[80] E. Knill and R. Laflamme, “Power of one bit of quantum information,” Phys. Rev.
Lett. 81, 5672–5675 (1998).
[81] Animesh Datta, Anil Shaji, and Carlton M. Caves, “Quantum discord and the
power of one qubit,” Phys. Rev. Lett. 100, 050502 (2008).
[82] Aram W. Harrow and Ashley Montanaro, “Quantum computational supremacy,”

Nature 549, 203 EP – (2017).
[83] Michael J. Bremner, Ashley Montanaro, and Dan J. Shepherd, “Achieving quantum
supremacy with sparse and noisy commuting quantum computations,” Quantum 1,
8 (2017).
[84] Sergio Boixo, Sergei V. Isakov, Vadim N. Smelyanskiy, Ryan Babbush, Nan Ding,
Zhang Jiang, Michael J. Bremner, John M. Martinis, and Hartmut Neven, “Char-
acterizing quantum supremacy in near-term devices,” (2016), arXiv:1608.00263 .
[85] David P. DiVincenzo, “The physical implementation of quantum computation,”

Fortschritte der Physik 48, 771–783 (2000).
[86] Carlos A. Pérez-Delgado and Pieter Kok, “Quantum computers: Definition and
implementations,” Phys. Rev. A 83, 012303 (2011).
[87] Jonathan A. Jones, “Quantum computing with nmr,” Progress in Nuclear Magnetic
Resonance Spectroscopy 59, 91 – 120 (2011).
[88] T. P. Harty, D. T. C. Allcock, C. J. Ballance, L. Guidoni, H. A. Janacek, N. M.

Linke, D. N. Stacey, and D. M. Lucas, “High-fidelity preparation, gates, memory,
and readout of a trapped-ion quantum bit,” Phys. Rev. Lett. 113, 220501 (2014).
[89] Thomas Monz, Philipp Schindler, Julio T. Barreiro, Michael Chwalla, Daniel Nigg,
William A. Coish, Maximilian Harlander, Wolfgang Hänsel, Markus Hennrich, and
Rainer Blatt, “14-qubit entanglement: Creation and coherence,” Phys. Rev. Lett.
106, 130506 (2011).
[90] Naomi H. Nickerson, Ying Li, and Simon C. Benjamin, “Topological quantum
computing with a very noisy network and local error rates approaching one percent,”
Nat. Comms. 4, 1756 (2013).
[91] Bjoern Lekitsch, Sebastian Weidt, Austin G. Fowler, Klaus Mølmer, Simon J. De-
vitt, Christof Wunderlich, and Winfried K. Hensinger, “Blueprint for a microwave
trapped ion quantum computer,” Science Advances 3 (2017).
[92] R. Barends, J. Kelly, A. Megrant, A. Veitia, D. Sank, E. Jeffrey, T. C. White,

J. Mutus, A. G. Fowler, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth,
196
BIBLIOGRAPHY
C. Neill, P. O/’Malley, P. Roushan, A. Vainsencher, J. Wenner, A. N. Korotkov,

A. N. Cleland, and John M. Martinis, “Superconducting quantum circuits at the
surface code threshold for fault tolerance,” Nature 508, 500–503 (2014).
[93] R. Maurand, X. Jehl, D. Kotekar-Patil, A. Corna, H. Bohuslavskyi, R. Laviéville,

L. Hutin, S. Barraud, M. Vinet, M. Sanquer, and S. De Franceschi, “A cmos silicon
spin qubit,” Nat. Comms. 7, 13575 (2016).
[94] M. Veldhorst, C. H. Yang, J. C. C. Hwang, W. Huang, J. P. Dehollain, J. T.

Muhonen, S. Simmons, A. Laucht, F. E. Hudson, K. M. Itoh, A. Morello, and A. S.
Dzurak, “A two-qubit logic gate in silicon,” Nature 526, 410–414 (2015).
[95] Leonard Mandel and Emil Wolf, Optical coherence and quantum optics (Cambridge
university press, 1995).
[96] Pieter Kok and Brendon W. Lovett, Introduction to Optical Quantum Information
Processing (Cambridge University Press, 2010).
[97] C. Gerry and P. Knight, Introductory Quantum Optics (Cambridge University Press,
2004) p. 332.
[98] C. K. Hong, Z. Y. Ou, and L. Mandel, “Measurement of subpicosecond time inter-

vals between two photons by interference,” Phys. Rev. Lett. 59, 2044–2046 (1987).
[99] Sandu Popescu, “Knill-laflamme-milburn linear optics quantum computation as a

measurement-based computation,” Phys. Rev. Lett. 99, 250501 (2007).
[100] Stefan Scheel, “Permanents in linear optical networks,” (2004), arXiv:quant-

ph/0406127 .
[101] Howard M Wiseman, “How many principles does it take to change a light bulbinto
a laser?” Physica Scripta 91, 033001 (2016).
[102] J G Rarity, P R Tapster, and R Loudon, “Non-classical interference between inde-

pendent sources,” Journal of Optics B: Quantum and Semiclassical Optics 7, S171
(2005).
[103] Christian Weedbrook, Stefano Pirandola, Raúl Garcı́a-Patrón, Nicolas J. Cerf, Tim-
othy C. Ralph, Jeffrey H. Shapiro, and Seth Lloyd, “Gaussian quantum informa-
tion,” Rev. Mod. Phys. 84, 621–669 (2012).
[104] N. J. Cerf, C. Adami, and P. G. Kwiat, “Optical simulation of quantum logic,”

Phys. Rev. A 57, R1477–R1480 (1998).
[105] Hai-Woong Lee and Jaewan Kim, “Quantum teleportation and bell’s inequality
using single-particle entanglement,” Phys. Rev. A 63, 012305 (2000).
[106] Barry C. Sanders, Stephen D. Bartlett, Terry Rudolph, and Peter L. Knight,
“Photon-number superselection and the entangled coherent-state representation,”
197
BIBLIOGRAPHY
Phys. Rev. A 68, 042329 (2003).
[107] Isaac L. Chuang and Yoshihisa Yamamoto, “Simple quantum computer,” Phys. Rev.
A 52, 3489–3496 (1995).
[108] G. J. Milburn, “Quantum optical fredkin gate,” Phys. Rev. Lett. 62, 2124–2127
(1989).
[109] Nobuyuki Matsuda, Ryosuke Shimizu, Yasuyoshi Mitsumori, Hideo Kosaka, and
Keiichi Edamatsu, “Observation of optical-fibre kerr nonlinearity at the single-
photon level,” Nat Photon 3, 95–98 (2009).
[110] Jeffrey H. Shapiro, “Single-photon kerr nonlinearities do not help quantum compu-
tation,” Phys. Rev. A 73, 062305 (2006).
[111] A. Javadi, I. Söllner, M. Arcari, S. Lindskov Hansen, L. Midolo, S. Mahmoodian,

G Kiršanskė, T. Pregnolato, E. H. Lee, J. D. Song, S. Stobbe, and P. Lodahl,
“Single-photon non-linear optics with a quantum dot in a waveguide,” Nat. Comms.
6, 8655 (2015).
[112] Daniel J. Brod and Joshua Combes, “Passive cphase gate via cross-kerr nonlineari-
ties,” Phys. Rev. Lett. 117, 080502 (2016).
[113] Pieter Kok, W. J. Munro, Kae Nemoto, T. C. Ralph, Jonathan P. Dowling, and
G. J. Milburn, “Linear optical quantum computing with photonic qubits,” Rev.
Mod. Phys. 79, 135–174 (2007).
[114] E. Knill, R. Laflamme, and G. J. Milburn, “A scheme for efficient quantum com-
putation with linear optics,” Nature 409, 46–52 (2001).
[115] J. Eisert, “Optimizing linear optics quantum gates,” Phys. Rev. Lett. 95, 040502
(2005).
[116] Michael A. Nielsen, “Optical quantum computation using cluster states,” Phys. Rev.
Lett. 93, 040503 (2004).
[117] Daniel E. Browne and Terry Rudolph, “Resource-efficient linear optical quantum
computation,” Phys. Rev. Lett. 95, 010501 (2005).
[118] Markus Michler, Klaus Mattle, Harald Weinfurter, and Anton Zeilinger, “Interfer-
ometric bell-state analysis,” Phys. Rev. A 53, R1209–R1212 (1996).
[119] K. Kieling, T. Rudolph, and J. Eisert, “Percolation, renormalization, and quantum

computing with nondeterministic gates,” Phys. Rev. Lett. 99, 130501 (2007).
[120] Mercedes Gimeno-Segovia, Pete Shadbolt, Dan E. Browne, and Terry Rudolph,
“From three-photon greenberger-horne-zeilinger states to ballistic universal quan-
tum computation,” Phys. Rev. Lett. 115, 020502 (2015).
198
BIBLIOGRAPHY
[121] Hussain A. Zaidi, Chris Dawson, Peter van Loock, and Terry Rudolph, “Near-
deterministic creation of universal cluster states with probabilistic bell measure-
ments and three-qubit resource states,” Phys. Rev. A 91, 042301 (2015).
[122] J. Bermejo-Vega, D. Hangleiter, M. Schwarz, R. Raussendorf, and J. Eisert,

“Architectures for quantum simulation showing quantum supremacy,” (2017),
arXiv:1703.00466 .
[123] Seinosuke Toda, “Pp is as hard as the polynomial-time hierarchy,” SIAM J. Comput.
20, 865–877 (1991).
[124] L.G. Valiant, “The complexity of computing the permanent,” Theoretical Computer
Science 8, 189 – 201 (1979).
[125] Leonid Gurvits, “On the complexity of mixed discriminants and related problems,”
in Mathematical Foundations of Computer Science 2005: 30th International Sym-
posium, MFCS 2005, Gdansk, Poland, August 29–September 2, 2005. Proceedings
(Springer Berlin Heidelberg, Berlin, Heidelberg, 2005) pp. 447–458.
[126] Larry Stockmeyer, “The complexity of approximate counting,” in Proceedings of the

Fifteenth Annual ACM Symposium on Theory of Computing, STOC ’83 (1983) pp.
118–126.
[127] A. P. Lund, A. Laing, S. Rahimi-Keshari, T. Rudolph, J. L. O’Brien, and T. C.

Ralph, “Boson sampling from a gaussian state,” Phys. Rev. Lett. 113, 100502
(2014).
[128] Craig S. Hamilton, Regina Kruse, Linda Sansoni, Sonja Barkhofen, Christine Sil-
berhorn, and Igor Jex, “Gaussian boson sampling,” (2016), arXiv:1612.01199 .
[129] Levon Chakhmakhchyan and Nicolas J. Cerf, “Boson sampling with gaussian mea-
surements,” (2017), arXiv:1705.05299 .
[130] Scott Aaronson and Daniel J. Brod, “Bosonsampling with lost photons,” Phys. Rev.
A 93, 012335 (2016).
[131] J B Spring, B J Metcalf, P C Humphreys, W S Kolthammer, X-M Jin, M Barbieri,

A Datta, N Thomas-Peter, N K Langford, D Kundys, J C Gates, B J Smith, P G R
Smith, and I A Walmsley, “Boson Sampling on a Photonic Chip,” Science 339,
798–801 (2013).
[132] Matthew A. Broome, Alessandro Fedrizzi, Saleh Rahimi-Keshari, Justin Dove, Scott
Aaronson, Timothy C. Ralph, and Andrew G. White, “Photonic boson sampling
in a tunable circuit,” Science 339, 794–798 (2013).
[133] Max Tillmann, Borivoje Dakic, Rene Heilmann, Stefan Nolte, Alexander Szameit,
and Philip Walther, “Experimental boson sampling,” Nat Photon 7, 540–544 (2013).
199
BIBLIOGRAPHY
[134] Andrea Crespi, Roberto Osellame, Roberta Ramponi, Daniel J. Brod, Ernesto F.
Galvao, Nicolo Spagnolo, Chiara Vitelli, Enrico Maiorino, Paolo Mataloni, and
Fabio Sciarrino, “Integrated multimode interferometers with arbitrary designs for
photonic boson sampling,” Nat Photon 7, 545–549 (2013).
[135] Jacques Carolan, Jasmin D A Meinecke, Peter J Shadbolt, Nicholas J Russell, Nur
Ismail, Kerstin Wörhoff, Terry Rudolph, Mark G Thompson, Jeremy L O’Brien,
Jonathan C F Matthews, and Anthony Laing, “On the experimental verification of
quantum complexity in linear optics,” Nat. Photon. 8, 621 –626 (2014).
[136] Hui Wang, Yu He, Yu-Huai Li, Zu-En Su, Bo Li, He-Liang Huang, Xing Ding, Ming-
Cheng Chen, Chang Liu, Jian Qin, Jin-Peng Li, Yu-Ming He, Christian Schneider,
Martin Kamp, Cheng-Zhi Peng, Sven Höfling, Chao-Yang Lu, and Jian-Wei Pan,
“High-efficiency multiphoton boson sampling,” Nat Photon 11, 361–365 (2017).
[137] Ludovico Latmiral, Nicolò Spagnolo, and Fabio Sciarrino, “Towards quantum
supremacy with lossy scattershot boson sampling,” New Journal of Physics 18,
113008 (2016).
[138] Alex Neville, Chris Sparrow, Raphal Clifford, Eric Johnston, Patrick M. Birchall,
Ashley Montanaro, and Anthony Laing, “No imminent quantum supremacy by
boson sampling,” (2017), arXiv:1705.00686 .
[139] Peter Clifford and Raphal Clifford, “The classical complexity of boson sampling,”
(2017), arXiv:1706.01260 .
[140] Fabio DellAnno, Silvio De Siena, and Fabrizio Illuminati, “Multiphoton quantum
optics and quantum state engineering,” Physics Reports 428, 53 – 168 (2006).
[141] W. P. Grice and I. A. Walmsley, “Spectral information and distinguishability in

type-ii down-conversion with a broadband pump,” Phys. Rev. A 56, 1627–1634
(1997).
[142] A. L. Migdall, D. Branning, and S. Castelletto, “Tailoring single-photon and multi-

photon probabilities of a single-photon on-demand source,” Phys. Rev. A 66, 053805
(2002).
[143] Fumihiro Kaneda, Bradley G. Christensen, Jia Jun Wong, Hee Su Park, Kevin T.
McCusker, and Paul G. Kwiat, “Time-multiplexed heralded single-photon source,”
Optica 2, 1010–1013 (2015).
[144] Gabriel J. Mendoza, Raffaele Santagati, Jack Munns, Elizabeth Hemsley, Mateusz
Piekarek, Enrique Martı́n-López, Graham D. Marshall, Damien Bonneau, Mark G.
Thompson, and Jeremy L. O’Brien, “Active temporal and spatial multiplexing of
photons,” Optica 3, 127–132 (2016).
[145] Wolfgang Mauerer, Malte Avenhaus, Wolfram Helwig, and Christine Silberhorn,
“How colors influence numbers: Photon statistics of parametric down-conversion,”
200
BIBLIOGRAPHY
Phys. Rev. A 80, 053815 (2009).
[146] Silverstone J. W., Bonneau D., OhiraK., SuzukiN., YoshidaH., IizukaN., EzakiM.,
NatarajanC. M., TannerM. G., HadfieldR. H., ZwillerV., MarshallG. D., RarityJ.
G., O’BrienJ. L., and ThompsonM. G., “On-chip quantum interference between
silicon photon-pair sources,” Nat Photon 8, 104–108 (2014).
[147] B. Darquié, M. P. A. Jones, J. Dingjan, J. Beugnon, S. Bergamini, Y. Sortais,

G. Messin, A. Browaeys, and P. Grangier, “Controlled single-photon emission from
a single trapped two-level atom,” Science 309, 454–456 (2005).
[148] P. Maunz, D. L. Moehring, S. Olmschenk, K. C. Younge, D. N. Matsukevich, and

C. Monroe, “Quantum interference of photon pairs from two remote trapped atomic
ions,” Nat Phys 3, 538–541 (2007).
[149] B. Lounis and W. E. Moerner, “Single photons on demand from a single molecule
at room temperature,” Nature 407, 491–493 (2000).
[150] Hannes Bernien, Lilian Childress, Lucio Robledo, Matthew Markham, Daniel
Twitchen, and Ronald Hanson, “Two-photon quantum interference from separate
nitrogen vacancy centers in diamond,” Phys. Rev. Lett. 108, 043604 (2012).
[151] Sonia Buckley, Kelley Rivoire, and Jelena Vukovi, “Engineered quantum dot single-
photon sources,” Reports on Progress in Physics 75, 126503 (2012).
[152] Jake Iles-Smith, Dara P. S. McCutcheon, Jesper Mørk, and Ahsan Nazir, “Limits
to coherent scattering and photon coalescence from solid-state quantum emitters,”
Phys. Rev. B 95, 201305 (2017).
[153] Hui Wang, Z.-C. Duan, Y.-H. Li, Si Chen, J.-P. Li, Y.-M. He, M.-C. Chen, Yu He,
X. Ding, Cheng-Zhi Peng, Christian Schneider, Martin Kamp, Sven Höfling, Chao-
Yang Lu, and Jian-Wei Pan, “Near-transform-limited single photons from an effi-
cient solid-state quantum emitter,” Phys. Rev. Lett. 116, 213601 (2016).
[154] Xing Ding, Yu He, Z.-C. Duan, Niels Gregersen, M.-C. Chen, S. Unsleber, S. Maier,
Christian Schneider, Martin Kamp, Sven Höfling, Chao-Yang Lu, and Jian-Wei
Pan, “On-demand single photons with high extraction efficiency and near-unity
indistinguishability from a resonantly driven quantum dot in a micropillar,” Phys.
Rev. Lett. 116, 020401 (2016).
[155] G. Reithmaier, M. Kaniber, F. Flassig, S. Lichtmannecker, K. Mller, A. Andrejew,

J. Vukovi, R. Gross, and J. J. Finley, “On-chip generation, routing, and detection
of resonance fluorescence,” Nano Letters 15, 5208–5213 (2015).
[156] Anthony Laing, Alberto Peruzzo, Alberto Politi, Maria Rodas Verde, Matthaeus
Halder, Timothy C. Ralph, Mark G. Thompson, and Jeremy L. O’Brien, “High-
fidelity operation of quantum photonic circuits,” Applied Physics Letters 97, 211109
(2010).
201
BIBLIOGRAPHY
[157] P. J. Shadbolt, M. R. Verde, A. Peruzzo, A. Politi, A. Laing, M. Lobino, J. C. F.

Matthews, M. G. Thompson, and J. L. O’Brien, “Generating, manipulating and
measuring entanglement and mixture with a reconfigurable photonic circuit,” Nat
Photon 6, 45–49 (2012).
[158] Andrea Crespi, Roberta Ramponi, Roberto Osellame, Linda Sansoni, Irene Bon-
gioanni, Fabio Sciarrino, Giuseppe Vallone, and Paolo Mataloni, “Integrated pho-
tonic quantum gates for polarization qubits,” Nat. Comms. 2, 566 (2011).
[159] Benjamin J. Metcalf, Justin B. Spring, Peter C. Humphreys, Nicholas Thomas-

Peter, Marco Barbieri, W. Steven Kolthammer, Xian-Min Jin, Nathan K. Langford,
Dmytro Kundys, James C. Gates, Brian J. Smith, Peter G. R. Smith, and Ian A.
Walmsley, “Quantum teleportation on a photonic chip,” Nat Photon 8, 770–774
(2014).
[160] Justin B. Spring, Paolo L. Mennea, Benjamin J. Metcalf, Peter C. Humphreys,

James C. Gates, Helen L. Rogers, Christoph Söller, Brian J. Smith, W. Steven
Kolthammer, Peter G. R. Smith, and Ian A. Walmsley, “Chip-based array of near-
identical, pure, heralded single-photon sources,” Optica 4, 90–96 (2017).
[161] Julien Claudon, Joel Bleuse, Nitin Singh Malik, Maela Bazin, Perine Jaffrennou,
Niels Gregersen, Christophe Sauvan, Philippe Lalanne, and Jean-Michel Gerard,
“A highly efficient single-photon source based on a quantum dot in a photonic
nanowire,” Nat Photon 4, 174–177 (2010).
[162] Robert H. Hadfield, “Single-photon detectors for optical quantum information ap-
plications,” Nat Photon 3, 696–705 (2009).
[163] Adriana E. Lita, Aaron J. Miller, and Sae Woo Nam, “Counting near-infrared
single-photons with 95% efficiency,” Opt. Express 16, 3032–3040 (2008).
[164] Alex D. Semenov, Gregory N. Goltsman, and Alexander A. Korneev, “Quantum

detection by current carrying superconducting film,” Physica C: Superconductivity
351, 349 – 356 (2001).
[165] Jacques Carolan, Christopher Harrold, Chris Sparrow, Enrique Martı́n-López,

Nicholas J. Russell, Joshua W. Silverstone, Peter J. Shadbolt, Nobuyuki Mat-
suda, Manabu Oguma, Mikitaka Itoh, Graham D. Marshall, Mark G. Thompson,
Jonathan C. F. Matthews, Toshikazu Hashimoto, Jeremy L. O’Brien, and Anthony
Laing, “Universal linear optics,” Science 349, 711–716 (2015).
[166] Masato Tanida, Ryo Okamoto, and Shigeki Takeuchi, “Highly indistinguishable
heralded single-photon sources using parametric down conversion,” Opt. Express
20, 15275–15285 (2012).
[167] Peter Shadbolt, Complexity and Control in Quantum Photonics (Springer Interna-
tional Publishing, 2016).
202
BIBLIOGRAPHY
[168] Michael Reck, Anton Zeilinger, Herbert J. Bernstein, and Philip Bertani, “Exper-
imental realization of any discrete unitary operator,” Phys. Rev. Lett. 73, 58–61
(1994).
[169] A Hurwitz, “Über die Ereugung der Invarianten durh Integration,” Nachr. Akad.
Wiss. Gött. I. Math. Phys. Kl. , 71 (1897).
[170] William R. Clements, Peter C. Humphreys, Benjamin J. Metcalf, W. Steven

Kolthammer, and Ian A. Walmsley, “Optimal design for universal multiport in-
terferometers,” Optica 3, 1460–1465 (2016).
[171] Fulvio Flamini, Nicol Spagnolo, Niko Viggianiello, Andrea Crespi, Roberto Osel-
lame, and Fabio Sciarrino, “Benchmarking integrated photonic architectures,”
(2017), arXiv:1705.09211 .
[172] Brian Borchers, “CSDP, A C library for semidefinite programming,” Optim.

Method. Softw. 11, 613–623 (1999).
[173] Robin Blume-Kohout, “Optimal, reliable estimation of quantum states,” New Jour-
nal of Physics 12, 043034 (2010).
[174] Robin Blume-Kohout, “Hedged maximum likelihood quantum state estimation,”

Phys. Rev. Lett. 105, 200504 (2010).
[175] M. Ringbauer, C. J. Wood, K. Modi, A. Gilchrist, A. G. White, and A. Fedrizzi,

“Characterizing quantum dynamics with initial system-environment correlations,”
Phys. Rev. Lett. 114, 090402 (2015).
[176] T. C. Ralph, N. K. Langford, T. B. Bell, and A. G. White, “Linear optical

controlled-not gate in the coincidence basis,” Phys. Rev. A 65, 062324 (2002).
[177] J. L. O’Brien, G. J. Pryde, A. G. White, T. C. Ralph, and D. Branning, “Demon-

stration of an all-optical quantum controlled-not gate,” Nature 426, 264–267 (2003).
[178] T. C. Ralph, A. G. White, W. J. Munro, and G. J. Milburn, “Simple scheme for

efficient linear optics quantum gates,” Phys. Rev. A 65, 012314 (2001).
[179] Qiang Zhang, Xiao-Hui Bao, Chao-Yang Lu, Xiao-Qi Zhou, Tao Yang, Terry
Rudolph, and Jian-Wei Pan, “Demonstration of a scheme for the generation of
“event-ready” entangled photon pairs from a single-photon source,” Phys. Rev. A
77, 062316 (2008).
[180] Jaewoo Joo, Peter L. Knight, Jeremy L. O’Brien, and Terry Rudolph, “One-way
quantum computation with four-dimensional photonic qudits,” Phys. Rev. A 76,
052326 (2007).
[181] Barbara M. Terhal and Pawel Horodecki, “Schmidt number for density matrices,”
Phys. Rev. A 61, 040301 (2000).
203
BIBLIOGRAPHY
[182] Mirko Lobino, Dmitry Korystov, Connor Kupchak, Eden Figueroa, Barry C.
Sanders, and A. I. Lvovsky, “Complete characterization of quantum-optical pro-
cesses,” Science 322, 563–566 (2008).
[183] Saleh Rahimi-Keshari, Artur Scherer, Ady Mann, A T Rezakhani, A I Lvovsky,

and Barry C Sanders, “Quantum process tomography with coherent states,” New
Journal of Physics 13, 013006 (2011).
[184] Saleh Rahimi-Keshari, Matthew A. Broome, Robert Fickler, Alessandro Fedrizzi,

Timothy C. Ralph, and Andrew G. White, “Direct characterization of linear-optical
networks,” Opt. Express 21, 13450–13458 (2013).
[185] Anthony Laing and Jeremy L. O’Brien, “Super-stable tomography of any linear
optical device,” (2012), arXiv:1208.2868 .
[186] Richard Sinkhorn, “A relationship between arbitrary positive matrices and doubly
stochastic matrices,” Ann. Math. Statist. 35, 876–879 (1964).
[187] Richard Sinkhorn and Paul Knopp, “Concerning nonnegative matrices and doubly
stochastic matrices.” Pacific J. Math. 21, 343–348 (1967).
[188] Jacob Mower, Nicholas C. Harris, Gregory R. Steinbrecher, Yoav Lahini, and Dirk
Englund, “High-fidelity quantum state evolution in imperfect photonic integrated
circuits,” Phys. Rev. A 92, 032322 (2015).
[189] B. Ulmann, Analog Computing (Oldenbourg Wissenschaftsverlag, 2013).
[190] Constantin Brif, Raj Chakrabarti, and Herschel Rabitz, “Control of quantum phe-
nomena: past, present and future,” New Journal of Physics 12, 075008 (2010).
[191] Michael L. Hause, Y. Heidi Yoon, and F. Fleming Crim, “Vibrationally mediated
photodissociation of ammonia: The influence of nh stretching vibrations on passage
through conical intersections,” The Journal of Chemical Physics 125, 174309 (2006).
[192] Daan Brinks, Fernando D. Stefani, Florian Kulzer, Richard Hildner, Tim H.
Taminiau, Yuri Avlasevich, Klaus Mullen, and Niek F. van Hulst, “Visualizing
and controlling vibrational wave packets of single molecules,” Nature 465, 905–908
(2010).
[193] A.S. Alnaser, M. Kbel, R. Siemering, B. Bergues, Nora G Kling, K.J. Betsch,
Y. Deng, J. Schmidt, Z.A. Alahmed, A.M. Azzeer, J. Ullrich, I. Ben-Itzhak,
R. Moshammer, U. Kleineberg, F. Krausz, R. de Vivie-Riedle, and M.F. Kling,
“Subfemtosecond steering of hydrocarbon deprotonation through superposition of
vibrational modes,” Nat. Comms. 5 (2014), 10.1038/ncomms4800.
[194] Xin Tong, Alexander H. Winney, and Stefan Willitsch, “Sympathetic cooling of
molecular ions in selected rotational and vibrational states produced by threshold
photoionization,” Phys. Rev. Lett. 105, 143001 (2010).
204
BIBLIOGRAPHY
[195] M. McDonald, B. H. McGuyer, F. Apfelbeck, C. H. Lee, I. Majewska, R. Moszynski,

and T. Zelevinsky, “Photodissociation of ultracold diatomic strontium molecules
with quantum state control,” Nature 534, 122–126 (2016).
[196] Tyler L Cocker, Dominik Peller, Ping Yu, Jascha Repp, and Rupert Huber, “Track-
ing the ultrafast motion of a single molecule by femtosecond orbital imaging,” Na-
ture 539, 263–267 (2016).
[197] B. Wolter, M. G. Pullen, A.-T. Le, M. Baudisch, K. Doblhoff-Dier, A. Senftleben,

M. Hemmer, C. D. Schröter, J. Ullrich, T. Pfeifer, R. Moshammer, S. Gräfe, O. Ven-
drell, C. D. Lin, and J. Biegert, “Ultrafast electron diffraction imaging of bond
breaking in di-ionized acetylene,” Science 354, 308–312 (2016).
[198] M. Born and R. Oppenheimer, “Zur quantentheorie der molekeln,” Annalen der
Physik 389, 457–484 (1927).
[199] Daniel H. Friese, Christof Httig, and Jrg Komann, “Analytic molecular hessian cal-
culations for cc2 and mp2 combined with the resolution of identity approximation,”
Journal of Chemical Theory and Computation 9, 1469–1480 (2013).
[200] Christoph R. Jacob and Markus Reiher, “Localizing normal modes in large
molecules,” The Journal of Chemical Physics 130, 084106 (2009).
[201] D J Clouthier and D A Ramsay, “The spectroscopy of formaldehyde and thio-

formaldehyde,” Annual Review of Physical Chemistry 34, 31–58 (1983).
[202] Ove Christiansen, Henrik Koch, and Poul Jargensen, “The second-order approxi-
mate coupled cluster singles and doubles model cc2,” Chemical Physics Letters 243,
409 – 418 (1995).
[203] Florian Weigend and Reinhart Ahlrichs, “Balanced basis sets of split valence, triple
zeta valence and quadruple zeta valence quality for h to rn: Design and assessment
of accuracy,” Phys. Chem. Chem. Phys. 7, 3297–3305 (2005).
[204] M. Gruebele and P. G. Wolynes, “Vibrational energy flow and chemical reactions,”
Accounts of Chemical Research 37, 261–267 (2004).
[205] David M. Leitner, “Energy flow in proteins,” Annual Review of Physical Chemistry
59, 233–259 (2008).
[206] Dvira Segal and Bijay Kumar Agarwalla, “Vibrational heat transport in molecular
junctions,” Annual Review of Physical Chemistry 67, 185–209 (2016).
[207] Maja Kobus, Phuong H. Nguyen, and Gerhard Stock, “Coherent vibrational energy
transfer along a peptide helix,” The Journal of Chemical Physics 134, 124518 (2011).
[208] A. Arkhipov and G. Kuperberg, “The bosonic birthday paradox,” Geometry &
Topology Monographs 18, 1–7 (2012).
205
BIBLIOGRAPHY
[209] Sander Woutersen and Huib J. Bakker, “Resonant intermolecular transfer of vibra-
tional energy in liquid water,” Nature 402, 507–509 (1999).
[210] Jrg Lindner, Peter Vhringer, Maxim S. Pshenichnikov, Dan Cringus, Douwe A.
Wiersma, and Maxim Mostovoy, “Vibrational relaxation of pure liquid water,”
Chemical Physics Letters 421, 329 – 333 (2006).
[211] Krupa Ramasesha, Luigi De Marco, Aritra Mandal, and Andrei Tokmakoff, “Water
vibrations have strongly mixed intra- and intermolecular character,” Nat Chem 5,
935–940 (2013).
[212] Roger A Horn and Charles R Johnson, Topics in Matrix Analysis (Cambridge Uni-
versity Press, 1991).
[213] R. D. Amos, N. C. Handy, W. H. Green, D. Jayatilaka, A. Willetts, and P. Palmieri,

“Anharmonic vibrational properties of ch2f2 : A comparison of theory and experi-
ment,” The Journal of Chemical Physics 95, 8323–8336 (1991).
[214] Wataru Mizukami and David P. Tew, “A second-order multi-reference perturbation

method for molecular vibrations,” The Journal of Chemical Physics 139, 194108
(2013).
[215] D. Kraft, A Software Package for Sequential Quadratic Programming, Tech. Rep.
(DFVLR-FB 88-28, 1988).
[216] Eric Jones, Travis Oliphant, Pearu Peterson, et al., “SciPy: Open source scientific
tools for Python,” (2001–).
[217] Joonsuk Huh, Gian Giacomo Guerreschi, Borja Peropadre, Jarrod R. McClean, and
Alán Aspuru-Guzik, “Boson sampling for molecular vibronic spectra,” Nat Photon
9, 615–620 (2015).
[218] Andreas Khn and Christof Httig, “Analytic gradients for excited states in the
coupled-cluster model cc2 employing the resolution-of-the-identity approximation,”
The Journal of Chemical Physics 119, 5021–5036 (2003).
[219] John A. Nelder and Roger Mead, “A simplex method for function minimization,”
Computer Journal 7, 308–313 (1965).
[220] Alex Arkhipov, “Bosonsampling is robust against small errors in the network ma-
trix,” Phys. Rev. A 92, 062326 (2015).
[221] Nicholas J Russell, Levon Chakhmakhchyan, Jeremy L OBrien, and Anthony Laing,
“Direct dialling of haar random unitary matrices,” New Journal of Physics 19,
033007 (2017).
[222] C. M. Wilkes, X. Qiang, J. Wang, R. Santagati, S. Paesani, X. Zhou, D. A. B.

Miller, G. D. Marshall, M. G. Thompson, and J. L. O’Brien, “60db high-extinction
206
BIBLIOGRAPHY
auto-configured mach–zehnder interferometer,” Opt. Lett. 41, 5318–5321 (2016).
[223] Nicholas C. Harris, Gregory R. Steinbrecher, Mihika Prabhu, Yoav Lahini, Jacob
Mower, Darius Bunandar, Changchen Chen, Franco N. C. Wong, Tom Baehr-Jones,
Michael Hochberg, Seth Lloyd, and Dirk Englund, “Quantum transport simulations
in a programmable nanophotonic processor,” Nat Photon 11, 447–452 (2017).
[224] Michael Varnava, Daniel E. Browne, and Terry Rudolph, “Loss tolerance in one-
way quantum computation via counterfactual error correction,” Phys. Rev. Lett.
97, 120501 (2006).
[225] Sean D. Barrett and Thomas M. Stace, “Fault tolerant quantum computation with
very high threshold for loss errors,” Phys. Rev. Lett. 105, 200502 (2010).
[226] Terry Rudolph, “Why i am optimistic about the silicon-photonic route to quantum
computing,” APL Photonics 2, 030901 (2017).
[227] Jonathan Crickmore, Jonathan Frazer, Scott Shaw, and Pieter Kok, “Effect of
component variations on the gate fidelity in linear optical networks,” Phys. Rev. A
94, 022326 (2016).
[228] Adrian J. Menssen, Alex E. Jones, Benjamin J. Metcalf, Malte C. Tichy, Stefanie
Barz, W. Steven Kolthammer, and Ian A. Walmsley, “Distinguishability and many-
particle interference,” Phys. Rev. Lett. 118, 153603 (2017).
[229] Charles H. Bennett, David P. DiVincenzo, John A. Smolin, and William K. Woot-
ters, “Mixed-state entanglement and quantum error correction,” Phys. Rev. A 54,
3824–3851 (1996).
[230] Peter P. Rohde and Timothy C. Ralph, “Error models for mode mismatch in linear
optics quantum computing,” Phys. Rev. A 73, 062312 (2006).
[231] Fabian Ewert and Peter van Loock, “3/4-efficient bell measurement with passive
linear optics and unentangled ancillae,” Phys. Rev. Lett. 113, 140403 (2014).
[232] W. P. Grice, “Arbitrarily complete bell-state measurement using only linear optical
elements,” Phys. Rev. A 84, 042331 (2011).
[233] Michael Varnava, Daniel E. Browne, and Terry Rudolph, “How good must single
photon sources and detectors be for efficient linear optical quantum computation?”
Phys. Rev. Lett. 100, 060502 (2008).
[234] M. Hein, J. Eisert, and H. J. Briegel, “Multiparty entanglement in graph states,”

Phys. Rev. A 69, 062311 (2004).
[235] Koji Azuma, Kiyoshi Tamaki, and Hoi-Kwong Lo, “All-photonic quantum re-
peaters,” Nature Communications 6, 6787 (2015).
[236] Mihir Pant, Hari Krovi, Dirk Englund, and Saikat Guha, “Rate-distance tradeoff
207
BIBLIOGRAPHY
and resource costs for all-optical quantum repeaters,” Phys. Rev. A 95, 012304
(2017).
[237] Antonio Acı́n, Nicolas Brunner, Nicolas Gisin, Serge Massar, Stefano Pironio, and
Valerio Scarani, “Device-independent security of quantum cryptography against col-
lective attacks,” Phys. Rev. Lett. 98, 230501 (2007).
[238] Jaromı́r Fiurášek, Raúl Garcı́a-Patrón, and Nicolas J. Cerf, “Conditional generation
of arbitrary single-mode quantum states of light by repeated photon subtractions,”
Phys. Rev. A 72, 033822 (2005).
[239] Mikkel Heuck, Mihir Pant, and Dirk Englund, “Temporally and frequency multi-
plexed single photon source using quantum feedback control for scalable photonic
quantum technologies,” (2017), arXiv:1708.08875 .
[240] W. H. P. Pernice, C. Schuck, O. Minaeva, M. Li, G. N. Goltsman, A. V. Sergienko,

and H. X. Tang, “High-speed and high-efficiency travelling wave single-photon de-
tectors embedded in nanophotonic circuits,” Nat. Comms. 3, 1325 (2012).
[241] J. W. Silverstone, R. Santagati, D. Bonneau, M. J. Strain, M. Sorel, J. L. O’Brien,

and M. G. Thompson, “Qubit entanglement between ring-resonator photon-pair
sources on a silicon chip,” Nat. Comms. 6, 7948 (2015).
[242] Simon Laibacher and Vincenzo Tamma, “From the physics to the computational
complexity of multiboson correlation interference,” Phys. Rev. Lett. 115, 243605
(2015).
[243] Faraz Najafi, Jacob Mower, Nicholas C Harris, Francesco Bellei, Andrew Dane,
Catherine Lee, Xiaolong Hu, Prashanta Kharel, Francesco Marsili, Solomon Assefa,
et al., “On-chip detection of non-classical light by scalable integration of single-
photon detectors,” Nat. Comms. 6, 5873 (2015).
[244] K. Tanizawa, K. Suzuki, S. Suda, H. Matsuura, K. Ikeda, S. Namiki, and

H. Kawashima, “Silicon photonic 32 x 32 strictly-non-blocking blade switch and
its full path characterization,” in 2016 21st OptoElectronics and Communications
Conference (OECC) held jointly with 2016 International Conference on Photonics
in Switching (PS) (2016) pp. 1–3.
[245] Tae Joon Seok, Niels Quack, Sangyoon Han, Richard S. Muller, and Ming C.
Wu, “Large-scale broadband digital silicon photonic switches with vertical adiabatic
couplers,” Optica 3, 64–70 (2016).
[246] Martin Schwarz and Maarten Van den Nest, “Simulating quantum circuits with
sparse output distributions,” (2013), arXiv:1310.6749 .
[247] Takahiro Inagaki, Yoshitaka Haribara, Koji Igarashi, Tomohiro Sonobe, Shuhei
Tamate, Toshimori Honjo, Alireza Marandi, Peter L. McMahon, Takeshi Umeki,
Koji Enbutsu, Osamu Tadanaga, Hirokazu Takenouchi, Kazuyuki Aihara, Ken-ichi
208
BIBLIOGRAPHY
Kawarabayashi, Kyo Inoue, Shoko Utsunomiya, and Hiroki Takesue, “A coherent

ising machine for 2000-node optimization problems,” Science (2016), 10.1126/sci-
ence.aah4243.
[248] Peter L. McMahon, Alireza Marandi, Yoshitaka Haribara, Ryan Hamerly, Carsten
Langrock, Shuhei Tamate, Takahiro Inagaki, Hiroki Takesue, Shoko Utsunomiya,
Kazuyuki Aihara, Robert L. Byer, M. M. Fejer, Hideo Mabuchi, and Yoshihisa
Yamamoto, “A fully-programmable 100-spin coherent ising machine with all-to-all
connections,” Science (2016), 10.1126/science.aah5178.
[249] Mercedes Gimeno-Segovia, Hugo Cable, Gabriel J Mendoza, Pete Shadbolt,

Joshua W Silverstone, Jacques Carolan, Mark G Thompson, Jeremy L O’Brien,
and Terry Rudolph, “Relative multiplexing for minimising switching in linear-optical
quantum computing,” New Journal of Physics 19, 063013 (2017).
[250] Sam Morley-Short, Sara Bartolucci, Mercedes Gimeno-Segovia, Pete Shadbolt,

Hugo Cable, and Terry Rudolph, “Physical-depth architectural requirements for
generating universal photonic cluster states,” (2017), arXiv:1706.07325 .
[251] Ying Li, Peter C. Humphreys, Gabriel J. Mendoza, and Simon C. Benjamin, “Re-
source costs for fault-tolerant linear optical quantum computing,” Phys. Rev. X 5,
041007 (2015).
209
Appendix A
Universal Linear Optics
A.1 Transfer Matrix Reconstructions

Here we provide the transfer matrix reconstructions referred to in Chapter 3.
The circuit for the Bell state generator:
 
0.680 0.662 0.026 0.006 0.006 0.006
 
0.384 −0.400 + 0.063i 0.570 0.548 0.013 0.013
 
 
 
−0.385 + 0.047i −0.254 − 0.281i −0.275 + 0.306i
 
 0.384 0.416 0.388 
MB =  
−0.390 + 0.036i −0.309 + 0.293i −0.263 − 0.285i −0.413 −0.370 + 0.037i 
 
 0.393
 
−0.549 −0.075 − 0.428i −0.029 − 0.405i 
 
 0.012 0.0181 0.539
 
0.009 0.0125 0.017 0.012 0.684 −0.734 + 0.018i
(A.1)
The circuit for the KLM CNOT gate:
 
0.468 0.612 0.430 0.402 0.009 0.009
 
0.606 −0.478 0.015 0.035 0.624 0.022
 
 
 
−0.672 −0.394
 
 0.365 0.061 0.292 0.329 
MKLM =  
−0.686 + 0.150i −0.352 + 0.075i 0.315 + 0.217i 
 
 0.367 0.020 0.266
 
−0.4182 + 0.042i −0.466 + 0.068i
 
 0.010 0.605 0.461 0.048 
 
0.286 0.029 −0.174 + 0.114i −0.121 + 0.103i −0.323 + 0.040i −0.794 + 0.288i
(A.2)
210
A. Universal Linear Optics
The circuit for postselected CNOT gate:

 
0.559 0.773 0.029 0.000 0.006 0.006
 
0.783 −0.571 + 0.055i 0.022 0.000 0.008 0.011
 
 
 
 
 0.009 0.021 0.556 0.558 0.579 0.006 
Mpost =   (A.3)
−0.566 + 0.139i
 
 0.010 0.025 0.577 0.078 0.531 
 
−0.556 + 0.085i −0.533 + 0.105i 
 
 0.010 0.019 0.568 0.066
 
0.009 0.011 0.016 0.551 −0.566395 −0.587 − 0.111i
211
Appendix B
Simulating the Vibrational Quantum

Dynamics of Molecules with
Integrated Photonics
B.1 Molecule details

Here we include all additional computed vibrational frequencies and local basis transfor-
mations.
B.1.1 HFHF
The frequencies of the normal modes are calculated to be
{4068.29, 3981.51, 578.25, 464.46, 207.95, 148.93}cm−1 (B.1)
212
B. Simulating the Vibrational Quantum Dynamics of Molecules with Integrated
Photonics
and the transformation between the normal modes and localised modes is given by
 
0.768 0.147 0.281 0. −0.410 −0.377
 
 −0.178 0.920 0.265 0. 0.220 −0.046 
 
 
 0.064 −0.303 0.844
 
0. 0.357 0.254 
UL = 


 (B.2)
 0. 0. 0. 1. 0. 0. 
 
0.135 −0.351
 
 0.607 0. 0.462 0.525 
 
−0.076 0.145 0.123 0. −0.665 0.718
B.1.2 HNCO
{3694.87, 2260.22, 1241.85, 805.38, 605.69, 552.52}cm−1 (B.3)
 
0.826 0.116 −0.198 −0.438 0. 0.272
 
 −0.066 0.982 0.120 0.001 0. −0.129 
 
 
 0.157 −0.090 0.971 −0.085 0. 0.133 
 
UL = 
  (B.4)
 0.537 −0.037 0.039 0.667 0. −0.514 

 
 
 0. 0. 0. 0. 1. 0. 
 
0.028 0.111 −0.050 0.597 0. 0.792
B.1.3 N4
{1366.42, 1015.85, 947.85, 916.73, 412.35, 366.86}cm−1 (B.5)
213
Photonics
 
0.191 −0.356 −0.611 0.494 0. 0.469
 
 0.658 0.611 −0.356 −0.011 0. −0.258 
 
 
 0.658 −0.611 0.356 −0.011 0. −0.258 
 
UL = 
 
 (B.6)
 0.191 0.356 0.611 0.494 0. 0.469 
 
 
 0. 0. 0. 0. 1. 0. 
 
0.244 0. 0. −0.716 0. 0.654
B.1.4 P4
{608.25, 460.20, 460.20, 460.20, 366.27, 366.25}cm−1 (B.7)
 
0.5 −0.676 −0.096 −0.533 0. 0.
 
 0.5 0.727 0.074 −0.465 0. 0.
 

 
0.853 −0.522 
 
 0. 0. 0. 0.
UL = 
  (B.8)
0.059 −0.691 0.519

 0.5 0. 0. 
 
 −0.5 0.110 −0.713 −0.480
 
0. 0. 
 
0. 0. 0. 0. 0.522 0.853
B.1.5 SO3
{1334.97, 1334.55, 966.96, 497.57, 497.52, 469.50}cm−1 (B.9)
214
Photonics
 
0.608 −0.346 −0.577 −0.218 0.362 0.
 
0. 0.701 −0.579 0.417 0. 0. 
 

 
 
 0.608 0.346 0.577 0.218 0.362 0. 
UL = 
  (B.10)
−0.519 −0.012 0.855

 0. 0. 0. 
 
 −0.511
 
0. 0. 0. 0.859 0. 
 
0. 0. 0. 0. 0. 1.
B.1.6 NMA
The vibrational frequencies of NMA are calculated to be
{3589.68, 3143.77, 3113.38, 3111.68, 3060.23, 3024.31, 3000.1, 1808.96, 1572.7, 1484.02,
1481.06, 1469.21, 1452.61, 1427.76, 1391.94, 1279.31, 1170.4, 1150.05, 1120.42, 1053.5,
995.522, 876.075, 637.311, 628.305, 475.592, 436.109, 291.874, 176.452, 96.9364, 39.7879}cm−1
(B.11)
and the transformation used is UL ⊕ 124 where we select the subset
{1279.31, 1170.4, 1120.42, 995.522, 876.075, 628.305} (B.12)
to be the first six modes.

 
0.796 −0.007 −0.401 −0.345 0.050 −0.290
 
−0.115 0.794 0.141 −0.178 0.500 −0.233
 
 
 
−0.085
 
 0.518 0.807 0.139 0.156 0.172 
UL =   (B.13)
−0.312 0.214 −0.012
 
 0.211 0.147 0.889 
 
−0.190 −0.583 −0.031 0.704 −0.357
 
 0.006 
 
0.059 −0.011 −0.267 −0.197 0.427 0.839
215
Photonics
B.1.7 H2 O
The vibrational frequencies of the normal modes in the harmonic model are
{3914.92, 3787.59, 1627.01}cm−1
and the transformation between normal and local modes is

 
0.70711 −0.70709 −0.00502
 
UL = 0.70711 0.70709 (B.14)
 
0.00502 
 
0. −0.00710 0.99997
Now, restricting to the subspace of the asymmetric and symmetric stretch modes and
adjusting the energy levels for anharmonic potentials via perturbation theory results in
single excitation energy levels of
{3740.05, 3619.68}cm−1 .
With two excitations the {20, 02, 11} energy levels become
{7391.43, 7154.35, 7206.46}cm−1
and the normal to local transformation for this subspace is given by

 
0.70711 −0.70711 0.
 
UL = 0.70711 (B.15)
 
0.70711 0.
 
0. 0. 1.
B.1.8 NH3
The vibrational frequencies of the ground electronic state are given by
{3646.13, 3646.09, 3503.94, 1678.57, 1678.4, 1065.59}cm−1 (B.16)
216
Photonics
and the frequencies of the excited state are
{2590.27, 2452.95, 2451.97, 1052.62, 1050.93, 750.124}cm−1 . (B.17)
The transformation between the two sets of modes is given by

 
−0.085 −0.289 −0.216 0.090 0.913 −0.147
 
−0.058 −0.011 0.096 0.023 0.169 0.979 
 

 
0.935 −0.023 −0.341 −0.031 0.017
 
 0.087 
UGE =  (B.18)
0.313 −0.118 0.358 −0.072 
 
 0.131 0.859
 
0.308 −0.336 0.819 0.331 0.0694 −0.0856 
 

 
−0.052 0.256 −0.244 0.931 −0.071 0.014
and the transformation between normal and local coordinates in the excited state is given
by  
0.575 −0.542 −0.576 −0.058 0.200 0.
 
0.579 0.767 −0.185 0.195 −0.062 0.
 
 
 
−0.223 −0.163 −0.131 0.
 
 0.578 0.757 
UL =   (B.19)
−0.139
 
 0.012 0.212 0.857 0.448 0. 
 
−0.010 −0.444
 
 0.221 0.126 0.859 0. 
 
0. 0. 0. 0. 0. 1.
B.2 Additional data

Here we include plots displaying the full data from our experiments. All plots represent
the experimentally estimated probabilities for measuring patterns ordered top to bot-
tom as {|100000i, |010000i...} for one photon data, {|110000i, |101000i...} for two photon
data without number resolving, {|20000i, |020000i..., |110000i...} for two photon data with
number resolving, {|111000i, |110100i...} for three photon data and
{|22000i, |202000i..., |211000i..., |111100i} for four photon data. The mode ordering is
defined by the vibrational frequencies defined in the above Molecule details section.
217
Photonics
1.0 1.0 1.0 1.0
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
0 50 103 155 0 50 103 155

Time (fs) Time (fs)
1.0 1.0 1.0 1.0
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
0 50 103 155 0 50 103 155

Time (fs) Time (fs)
1.0 1.0 1.0 1.0
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
0 50 103 155 0 50 103 155

Time (fs) Time (fs)
Figure B.1: Four atom molecule data. Full data sets for |1i (black) and |1i|1i (blue) input
experiments for H2 CS, P4 , SO3 , HNCO, N4 and HFHF.
218
Photonics
1.0 1.0
0.8
0.8
0.6
0.6
0 88 176 265 353 442 530 618 707 795

0.4
0.4
Time (fs)
0.2
0.2
0
0
1.0
1.0 0.8
0.8 0.6
0.6 0.4
0.4 0.2
0.2 0
0 38 76 114 152 190

Time (fs)
3 11 19 27 35
Time (fs)
Figure B.2: H2 CS data. Full data sets for |1i (black), |1i|1i (blue), |1i|1i|1i (grey) and
|2i|2i (red) input experiments.
219
Photonics
1.0 1.0 1.0
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
63 143 222 302 381 461 541 620 700 779 26 79 132 185 238
Time (fs) Time (fs)
Figure B.3: NMA data. Full data sets for |1i (black), |1i|1i (blue) and |1i|1i|1i (grey)
input experiments.
Harmonic
1.0
26 79 132 185 238 291 344 397 450 503

0.8
Time (fs)
0.6
Anharmonic 0.4
0.2
26 79 132 185 238 291 344 397 450 503

Time (fs)
Figure B.4: H2 O data. Full data sets for |2i and |2i|1i input experiments for harmonic
and anharmonic. States are ordered top to bottom as {|20i, |11i, |02i}
220
Photonics
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iteration Iteration
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iteration Iteration
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 100 200 300 0 100 200 300
Iteration Iteration
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 100 200 300 400 0 100 200 300 400
Iteration Iteration
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iteration Iteration
1.0 1.0
0.8
0.5
Bunching prob
0.6
-Cost
0.0 0.4
0.2
-0.5
0.0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iteration Iteration
Figure B.5: Cost function (left) and probability of bunching in an NH stretch mode (right)
for a series of random initial states in the NH3 dissociation experiment.
221
Photonics
0.5 0.5
0.4 0.4
Prob
Prob
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.5 0.5
0.4 0.4
Prob
0.3 Prob 0.3

0.2 0.2
0.1 0.1
0.0 0.0
0.5 0.5
0.4 0.4
Prob
Prob
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.5 0.5
0.4 0.4
Prob
Prob
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.5 0.5
0.4 0.4
Prob
Prob
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0.5 0.5
0.4 0.4
Prob
Prob
0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
Figure B.6: Full distributions for initial (left column) and final (right column) states over
the five time steps for the dissociation experiments summarised in Fig. B.5. Red bars are
probability of bunching in an NH stretch mode, blue bars are remaining bunched states
and yellow are anti-bunched states.
222
Appendix C
Linear Optical Quantum Computing

with Partially-Distinguishable
Photons
C.1 Trace inequality proof

We note that for two positive matrices %i , %j :
X √ √
Tr(%i %j ) = [sk ( %i %j )]2 (C.1)
k
where sk (A) is the k th largest singular value of the matrix A. Proof: Due to the cyclic
√ √ √ √
property of the trace Tr(%i %j ) = Tr( %i %j %j %i ) and then using the singular value
√ √ √ √ √ √
decomposition of %i %j = U DV and that ( %i %j )† = %j %i (relies on positivity of
%’s), Tr(%i %j ) = Tr(U DV V † DU † ) = Tr(D2 ).
Using von-neumans trace inequality:
n
!
Y √ √ √ √ √ √ √
Tr %i = |Tr ( %1 %2 %2 %3 %3 . . . %n %1 )|
i=1 (C.2)
X √ √ √ √ √ √
≤ sk ( %1 %2 )sk ( %2 %3 ) . . . sk ( %n %1 )
k
223
C. Linear Optical Quantum Computing with Partially-Distinguishable Photons
Now we consider vectors v(i, j) with components given by the singular values [v(i, j)]k =
√ √ √ √ √ √ √ √
sk ( %i %j ) and a vector |Vi with components [V]k = sk ( %1 %2 )sk ( %2 %3 ) . . . sk ( %n−1 %n )
such that:
X √ √ √ √ √ √
sk ( %1 %2 )sk ( %2 %3 ) . . . sk ( %n %1 )
k
= V v(n, 1)
q
≤ hV|Vi v(n, 1) v(n, 1) (C.3)
q
≤ v(1, 2) v(1, 2) v(2, 3) v(2, 3) . . . v(n, 1) v(n, 1)
p
= Tr(%1 %2 )Tr(%2 %3 ) . . . Tr(%n %1 ).
Hence: !
n
Y p
Tr %i ≤ Tr(%1 %2 )Tr(%2 %3 ) . . . Tr(%n %1 ) (C.4)
i=1
C.2 Bound on real term

Aim is to upper and lower bound the expression:
< [Tr(%1 %2 %4 %3 ) − Tr(%1 %2 %3 %4 )] = < [Tr(%1 %2 [%4 , %3 ])] . (C.5)
First we note that [%4 , %3 ] is anti-hermitian. Next we will make use of the fact that: the
trace of the product of an anti-hermitian operator A and hermitian operator H, Tr(AH),
is imaginary. Proof:
Tr(AH) = Tr[(AH)† ]
= Tr[H † A† ]
(C.6)
= −Tr[HA]
= −Tr[AH]
where we have used the cyclic property and linearity of the trace. Similarly it can be
found that Tr(H1 H2 ) is real for two hermitian matrices H1 and H2 . Therefore we can
224
split %1 %2 into a hermitian part: 12 {%1 , %2 }, and an anti-hermitian part: 21 [%1 , %2 ] together
with the linearity of the trace this implies:
1
< {Tr (%1 %2 [%4 , %3 ])} = Tr([%1 , %2 ][%4 , %3 ]) . (C.7)
2
We can now bound the magnitude of this expression using the Hilbert-Schmidt inner
product (hA|Bi = Tr[A† B])
|Tr ([%1 , %2 ][%4 , %3 ])| = [%2 , %1 ] [%4 , %3 ]

(C.8)
≤ k[%2 , %1 ]kk[%4 , %3 ]k.
For pure states |ψihψ| and |φihφ| by decomposing |φi = α|ψi + β|ψ⊥ i using gram schmidt
we can compute [|ψihψ|, |φihφ|] = αβ̄|ψihψ⊥ | − ᾱβ|ψ⊥ ihψ| and therefore:
k[|ψihψ|, |φihφ|]k = 2|αβ|

p (C.9)
= 2 V (1 − V )
where V = Tr(|ψihψ||φihφ|) = |α2 |. Therefore, for two states %ψ =

P
i pi |ψi ihψi | and
P
%φ = i qi |φi ihφi |, using the convexity of the norm:
X
k[%ψ , %φ ]k ≤ pi qj k[|ψi ihψi |, |φj ihφj |]k
i,j
X q
=2 pi qj Vij (1 − Vi,j ) (C.10)
i,j
p
≤ 2 V(1 − V)
P
where V ≡ pi qj Vij = Tr(%ψ %φ ). The bound on the last line is due to Jenson’s
i,j
p
inequality since the function f (x) = x(1 − x) is concave for 0 ≤ x ≤ 1. Using equations
225
(C.7), (C.8) and (C.10) we observe:
p
− 2 V12 (1 − V12 )V34 (1 − V34 )
≤ < [Tr(%1 %2 [%4 , %3 ])] (C.11)

p
≤ 2 V12 (1 − V12 )V34 (1 − V34 )
where Vij = Tr(%i %j ).
C.3 Additional transfer matrices

The transfer matrix for the GHZ generator circuit in Fig. 6.3c is given by
 
2 2 0 0 0 0 0 0 0 0 0 0
 
0 0 2 −2 0 0 0 0 0 0 0 0
 
 
 
1 −1 1 1 1 1 −1 1 0 0 0 0
 
 
 
1 −1 1 1 −1 −1 1 −1 0 0 0 0
 
 
 
0 0 0 0 1 1 1 −1 1 1 −1 1
 
 
 
1 −1 −1 −1 −1
 
1  0 0 0 0 1 1 1 
UGHZ = √  √ √ √ √ 
2 2
 0 0 0 0 2 − 2 2 2 0 0 0 0


 √ √ √ √ 
− 2
 
 0 0 0 0 2 2 2 0 0 0 0 
 
−1 −1
 
 1 1 1 0 0 0 0 1 1 1 
 
−1 −1 −1 −1
 

 1 0 0 0 0 1 1 1 

 
 0 0 0 0 0 0 0 0 0 0 2 2 
 
0 0 0 0 0 0 0 0 2 −2 0 0
226
The transfer matrix for the boosted fusion circuit in Fig. 6.3d is given by
 
1 i i 0 −1 0 0 0
 
i 1 0 i 0 −1 0
0 
 

 
i −1 1 0 i 0 0 0 
 

 
−1
 
1 i 0 1 0 i 0 0 
UF =  
2
 0 0 i 0 1 0 i −1 

 
−1 i 
 
 0 0 0 i 0 1
 
−1
 
 0 0 0 i 0 1 i 
 
0 0 0 −1 0 i i 1
C.4 Equivalence of pre- and post-fusion corrections

Here we briefly explain why local post-fusion correcting Pauli unitaries can be replaced
with pre-fusion unitaries in the case of a graph-basis diagonal error model, leading to the
simplified averaged measurement operators provided at the end of Sec. 6.6.3.
The average sucessful post-fusion state for a logical measurement with operators la-
belled by i acting upon the first two qubits, and associated Pauli correction operator Ki
acting on the remaining qubits is
1 X
ρout = Ki Tr1,2 [E i ρin ]Ki (C.12)
N i
where N is a normalisation constant. Ideally, all of the measurement operators can be

decomposed as E i = K̃i E K̃i where K̃i are local Paulis acting on the measured qubits.
The corrections Ki are then chosen such that their combined operation with K̃i stabilises
the ideal pre-fusion graph state (K̃i ⊗ Ki )|Gi = |Gi∀i, ensuring that any successful
1
measurement outcome leads to the same post-fusion state, ρout = N
Tr1,2 [E|GihG|].
pj |Gj ihGj |, then
P
If the initial logical state is noisy but graph-basis diagonal, ρin = j
(K̃i ⊗ Ki )|Gj i = ±|Gj i∀i, j. (C.13)
227
Therefore for any measurement operators {E i },
1 X
Ki Tr1,2 E i ρin Ki

ρout = (C.14)
N i
1 X
Tr1,2 K̃i E i K̃i ρin

= (C.15)
N i
1 hX i
= Tr1,2 K̃i E i K̃i ρin (C.16)
N i
1
= Tr1,2 Ēρin (C.17)
N
C.5 Fusion Operations

A feature of the scheme introduced in [120] is that the connectivity of the graph state is
optimised when the Bell measurements fail. This requires performing a set of single qubit
rotations before and after each Bell measurement. The procedure is depicted in Fig. C.1
for both the fusion operations which form five-qubit microclusters and those which then
fuse these clusters together.
For the first case, shown in Fig. C.1a), we can understand its operation via tracking
the stabilizers of the state. The stabilizer generators of the initial state are given by
hX1 Z2 A, X2 Z1 , X3 Z4 Z5 , X4 Z3 B, X5 Z3 Ci (C.18)
where A,B and C represent Z on all other qubits connected to qubits 1,4 and 5 respec-
tively. Performing the Clifford rotations HRH to qubit 2 and HR to qubit 3, where H
√
is the Hadamard and R = (X − Y )/ 2, results in the updated generators
hX1 Y2 A, −X2 Z1 , Y3 Z4 Z5 , −X4 X3 B, −X5 X3 Ci. (C.19)
If the measurement is successful, then qubits 2 and 3 are projected onto a Bell state e.g.
|φ+ i (if a different Bell outcome is received then additional correcting Paulis are required)
which is stabilized by hX2 X3 , Z2 Z3 i. The updated generators (omitting qubits 2 and 3)
228
a)
Z
success
B
4
2 3
1
A HRH HR
5
C
failure
SZH
SZ
b)
S Z
success
Z
2 3
1 4
A HR B
failure
SZHS
Figure C.1: Fusion operations including local unitaries. a) Fusions within microclusters.
Shown are the local rotations required, before the Bell measurement circuit is used. If
fusion is successful, the two parts of the graph state are joined, if it is not successful
then the B and C sections are instead connected. Shown also are the local correcting
unitaries required to recover a graph state. b) The same shown for the fusions between
microclusters.)
then become
h−X1 Z4 Z5 A, Z1 X4 B, Z1 X5 Ci (C.20)
a Z1 rotation then results in the connected graph state desired.
229
If the measurement is unsuccessful it projects qubits 2 and 3 onto the state |00i (or
|11i) which is stabilized by hZ2 , Z3 i. The updated stabilizer generators in this case are
hY1 A, −Y4 Z5 B, −Z4Y 5Ci (C.21)
√
performing the local rotations SZ on qubit 1 and SZH on qubit 4, where S = Z, results
in
hX1 A, X4 Z5 B, X5 Z4Ci (C.22)
So we recover a graph state where a bond has been added between qubits 4 and 5 and
qubit 1 remains connected to its neighbours.
For the secondary fusion operations which connect microclusters, we can consider the
simple case shown in Fig. C.1b). The initial state can be represented as
hX1 Z2 A, X2 Z1 , X3 Z4 , X4 Z3 Bi (C.23)
where A and B are Z’s on all other qubits connected to qubits 1 and 4 respectively.
Performing the rotation HR on qubit 3 results in the pre-measurement state
hX1 Z2 A, X2 Z1 , Y3 Z4 , −X4 X3 Bi. (C.24)
Successful measurement results in the state given by
h−Y1 Z4 A, −X4 Z1 Bi (C.25)
where S1 Z4 can be used to produce a graph state with a bond between qubits 1 and 4.
Unsuccessful Bell measurement results in the state
hX1 A, −Y4 Bi (C.26)
where S4 recovers a graph state with no bond between qubits 1 and 4.
230
C.6 Additional measurement operators

First, to evaluate the resultant state when we perform fusion between a qubit from two
microclusters, where only one may contain a distinguishable photon, we sum the two cases
where one of the ancilla photons on one side of the bell measurement circuit is distin-
guishable. This measurement is given by the POVM with elements { 2j=1 hΨj |Ei |Ψj i},
P
{ 6j=5 hΨj |Ei |Ψj i} resulting in the measurement {EA0 }

P
1 1 1
EA1 0 (φ+ ) = (|φ+ ihφ+ | + |φ− ihφ− |) + |ψ + ihψ + |, EA2 0 (φ+ ) = |φ+ ihφ+ |
32 8 8
1 1
EA3 0 (φ+ ) = (|φ+ ihφ+ | + |φ− ihφ− |), EA4 0 (φ− ) = |φ− ihφ− |
32 8
1 1 1
EA5 0 (φ− ) = (|φ− ihφ− | + |φ+ ihφ+ |) + |ψ + ihψ + |, EA6 0 (ψ + ) = |ψ + ihψ + |
16 8 4
1 1 1 1
EA7 0 (ψ + ) = |ψ + ihψ + | + |00ih00|, EA8 0 (ψ + ) = |ψ + ihψ + | + |11ih11|,
8 4 8 4
1 0 3 0 3
EA9 0 (ψ − ) = |ψ − ihψ − |, E9A (00) = |00ih00|, A
E10 (11) = |11ih11|
2 16 16
and ĒA0 = ĒA = (11|φ+ ihφ+ | + |φ− ihφ− | + 3|ψ + ihψ + | + 3|ψ − ihψ − |)/8. So the average
post-fusion state is the same as when we sum over all the ancilla photons.
We now consider the cases where a bad photon from each of two microclusters meet
at a fusion measurement. There are three ways in which this can happen; the first case
is where two ancilla photons are bad (one on each side). This can be described by the
operators { 2i=1 6j=5 hΨi,j |Ei |Ψi,j i} and results in the measurement {EAA }
P P
1 1 + + 1
E1AA (φ+ ) = (|φ+ ihφ+ | + |φ− ihφ− |) + |ψ ihψ |, E2AA (φ+ ) = (|φ+ ihφ+ | + |φ− ihφ− |)
32 8 32
1 1 + + 1 1
E3AA (φ− ) = (|φ− ihφ− | + |φ+ ihφ+ |) + |ψ ihψ |, E4AA (ψ + ) = |ψ + ihψ + | + |00ih00|
16 8 8 4
1
E5AA (φ+ ) = |ψ − ihψ − |
2
the measurement is successful with a probability of 7/16. Ē AA = (7|φ+ ihφ+ | + |φ− ihφ− | +
231
3|ψ + ihψ + | + 3|ψ − ihψ − |)/8.

Next is the case where both logical photons are bad. This case involves interference
between the bad photons and therefore we must further specify the structure of the input
states %. We will consider the two extremal cases where the internal state of the ith input
photon %i is two-dimensional, %2i = (1−)|ξ0 ihξ0 |+|ξ1 ihξ1 |, hence the bad photons interfere
perfectly with each other and the high-dimensional case, equivalent to the OBBM, %∞
i =
(1 − )|ξ0 ihξ0 | + |ξi ihξi |, where the bad photons are orthogonal. These two cases can then
be evaluated via {Trint [Ei (1 ⊗ |ξ0 ihξ0 |⊗2 ⊗ |ξ1 ihξ1 |⊗2 ⊗ |ξ0 ihξ0 |⊗2 )]}, {Trint [Ei (1 ⊗ |ξ0 ihξ0 |⊗2 ⊗
|ξ3 ihξ3 | ⊗ |ξ4 ihξ4 | ⊗ |ξ0 ihξ0 |⊗2 )]}, resulting in the measurements {ELL,%2 }, {ELL,%∞ }
1 + 1 + + 2 − 1 − −
ELL,% 2 (φ ) = (|φ ihφ | + |φ− ihφ− |), ELL,% 2 (φ ) = (|φ ihφ | + |φ+ ihφ+ |)
4 4
3 + + + 4 − − −
ELL,%2 (ψ ) = |ψ ihψ |, ELL,% 2 (ψ ) = |ψ ihψ |
the measurement is successful with a probability of 3/4. ĒLL,%2 = (5|φ+ ihφ+ |+|φ− ihφ− |)/2.
1 + 1 + + 2 − 1 − −
ELL,% ∞ (φ ) = (|φ ihφ | + |φ− ihφ− |), ELL,% ∞ (φ ) = (|φ ihφ | + |φ+ ihφ+ |)
8 8
3 + 1 + + 4 − 1 − −
ELL,% ∞ (ψ ) = (|ψ ihψ | + |ψ − ihψ − |), ELL,% ∞ (ψ ) = (|ψ ihψ | + |ψ + ihψ + |)
2 2
the measurement is successful with a probability of 5/8. ĒLL,%∞ = Ē L = 5/4(|φ+ ihφ+ | +

|φ− ihφ− |).
Finally we have the case where one ancilla and one logical photon (from opposite
sides) are bad. Again, we can calculate the POVMs corresponding to the low and high
dimensional extremes {EAL,%2 }, {EAL,%∞ }
232
1 + 1
EAL,% 2 (φ ) = (|φ+ ihφ+ | + |φ− ihφ− |),
32
2 + 1
EAL,% 2 (φ ) = (|φ+ ihφ+ | + |ψ − ihψ − |),
32
3 + 1
EAL,% 2 (φ ) = (|φ+ ihφ+ | + 2|ψ + ihψ + | + 3|ψ − ihψ − |),
32
4 − 1
EAL,% 2 (φ ) = (|φ− ihφ− | + |φ+ ihφ+ |),
32
5 − 1
EAL,% 2 (φ ) = (|φ+ ihφ+ | + 2|ψ + ihψ + | + 3|ψ − ihψ − |),
32
6 − 1
EAL,% 2 (φ ) = (|φ− ihφ− | + |ψ + ihψ + |),
32
7 + 1 1
EAL,% 2 (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |01ih01|,
32 16
8 + 1 1
EAL,%2 (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |10ih10|,
32 16
9 + 1 1
EAL,%2 (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |00ih00|,
16 8
10 + 1 1
EAL,%2 (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |11ih11|,
16 8
11 − 1 3
EAL,%2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 3|01ih01|) + |00ih00|,
256 64
12 − 1 3
EAL,% 2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 3|10ih10|) + |00ih00|,
256 64
13 − 1 3
EAL,% 2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 3|01ih01|) + |11ih11|,
256 64
14 − 1 3
256 64
15 − 1 1
256 64
16 − 1 1
256 64
17 − 1 1
256 64
18 − 1 1
256 64
19 − 1
EAL,%2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + |),
16
20 − 1 1
EAL,%2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + |) + |01ih01|,
64 32
21 − 1 − − + + 1
EAL,%2 (ψ ) = (|ψ ihψ | + |ψ ihψ |) + |10ih10|,
64 32
22 − 1 − − + + 1
EAL,%2 (ψ ) = (|ψ ihψ | + |ψ ihψ |) + |00ih00|,
64 32
233
23 − 1 1
EAL,% 2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + |) + |11ih11|,
64 32
24 − 1 3
EAL,% 2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + |) + |00ih00|,
64 32
25 − 1 3
EAL,% 2 (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + |) + |11ih11|
64 32
the measurement is successful with a probability of 9/16. ĒAL,%2 = (9|φ+ ihφ+ | +

11|φ− ihφ− | + 9|ψ + ihψ + | + 7|ψ − ihψ − |)/16.
1 + 1
EAL,%∞ (φ ) = (|φ+ ihφ+ | + |φ− ihφ− |),
32
2 + 1
EAL,%∞ (φ ) = (|φ+ ihφ+ | + |φ− ihφ− | + |ψ + ihψ + | + |ψ − ihψ − |),
64
3 + 1
EAL,% ∞ (φ ) = (|φ+ ihφ+ | + |φ− ihφ− | + 3|ψ + ihψ + | + 3|ψ − ihψ − |),
64
4 − 1
EAL,% ∞ (φ ) = (|φ− ihφ− | + |φ+ ihφ+ |),
32
5 − 1
EAL,%∞ (φ ) = (|φ+ ihφ+ | + |φ− ihφ− | + 3|ψ + ihψ + | + |ψ − ihψ − |),
64
6 − 1
EAL,% ∞ (φ ) = (|φ+ ihφ+ | + |φ− ihφ− | + |ψ + ihψ + | + 3|ψ − ihψ − |),
64
7 + 1 1
EAL,% ∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |01ih01|,
32 16
8 + 1 1
EAL,%∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − |) + |10ih10|,
32 16
9 + 1
EAL,%∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − | + |10ih10| + 3|00ih00|),
64
10 + 1
EAL,%∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − | + |01ih01| + 3|11ih11|),
64
11 + 1 1
EAL,%∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − | + (|10ih10| + 5|11ih11|),
32 64
12 + 1 1
EAL,% ∞ (ψ ) = (|ψ + ihψ + | + |ψ − ihψ − | + (|01ih01| + 5|00ih00|),
32 64
13 − 1
EAL,% ∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 2|10ih10| + 9|00ih00|),
256
14 − 1
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 2|01ih01| + 9|11ih11|),
256
15 − 1
256
16 − 1
256
234
17 − 1
256
18 − 1
256
19 − 1
256
20 − 1
256
21 − 1
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 4|00ih00|),
64
22 − 1
EAL,% ∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 4|11ih11|),
64
23 − 1
EAL,% ∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 2|00ih00|),
64
24 − 1
EAL,% ∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 2|11ih11|),
64
25 − 15
EAL,% ∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 5|01ih01|),
512
26 − 15
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 5|10ih10|),
512
27 − 2
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 7|01ih01|),
512
28 − 2
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 7|10ih10|),
512
29 − 7
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + |01ih01|),
512
30 − 7
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + |10ih10|),
512
31 − 1
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 5|01ih01|),
512
32 − 1
EAL,%∞ (ψ ) = (|ψ − ihψ − | + |ψ + ihψ + | + 5|10ih10|).
512
the measurement is successful with a probability of 9/16. ĒAL,%∞ = (5|φ+ ihφ+ |+5|φ− ihφ− |+
4|ψ + ihψ + | + 4|ψ − ihψ − |)/8.
C.7 Second order fusion error maps

Here we investigate the difference between the state produced by putting a distinguishable
photon in each input state which produces a final lattice qubit and the case where we
simply compose the error map resulting from a single distinguishable photon at each lattice
235
site. The only place where these models differ when considering errors on neighbouring
lattice sites is when the two distinguishable photons meet in a fusion measurement.
Fig. C.2 shows the error maps Λ1G , Λ2G recovered from a model where the input state
Eq. (6.22) is applied to the three photons on each side of the fusion measurement. In
addition we then calculate the full error map in the case where we take second-order
errors into account, i.e. we use Eq. (6.22) applied to all six photons. To analyse this case
we require the extra measurement operators found in appendix C.6. We use the POVMs
ĒAL,ρ∞ and ĒLL,ρ∞ , corresponding to the OBB model as this will provide the largest
discrepancy in the two output maps. This second-order map can be compared to the
composition of the first-order model for each input qubit as shown in Fig. C.2. We find
that the probability of second-order logical error is only slightly modified. Furthermore,
as described in the main body, since both models produce the same error maps to O()
and higher order terms do not result in higher weight errors, these differences may be
reasonably neglected.
236
a)
((
Z Z
4
⇤1G = (1 ✏)3 + ✏(1 ✏)2 { +
1
+
1
+
1
}
3 2 4 4
Z Z
b)
((
Z Z
4 1 1 1
⇤2G = (1 ✏)3 + ✏(1 ✏)2 {
3
+
2
+
4
+
4
}
Z Z
c)
((
Z Z
⇤2G ⇤1G = (1 ✏)6 + ✏(1 ✏)5{
8
+
3
+
3
+
1
}
3 4 4 2
Z Z
Z Z
301 19 19 47
+ ✏2 (1 ✏)4{ + + + }
144 16 16 48
Z Z
d)
((
Z Z
8 3 3 1
⇤12
G = (1 ✏)6 + ✏(1 ✏)5{
3
+
4
+
4
+
2
}
Z Z
Z Z
53 31 31 11
+ ✏2 (1 ✏)4{ + + + }
24 24 24 8
Z Z
Figure C.2: Error maps for second order errors in fusion measurements. a) Right: the
effective error map on the remaining qubits when one of the upper three photons (dark
red) shown on left may be distinguishable. b) Right: the effective error map when one
of the lower three photons (dark red) shown on left may be distinguishable. c) The error
map obtained by composing those from a) and b). d) Right: the effective error map when
one of the upper three photons and one of the lower three photons shown on left may be
distinguishable.
237
Appendix D
Mitigating Distinguishability Errors
The logical state produced by the B6 Bell state generator with an RSM input, %⊗4 , is:
1
ρoL = ×
 216 √ 
(5 + 8Tr[%3 ] − 5V) − 3(1 + 2Tr[%3 ] − 3V √1 (5 + 18Tr[%3 ] − 14Tr[%4 ] − 11V + 2V 2 ) (−1 − 4Tr[%3 ] + 8Tr[%4 ] + 5V)
3
 √ √ 
− 3(1 + 2Tr[%3 ] − 3V 3(1 − V) (−1 − 4Tr[%3 ] + 2Tr[%4 ] + 5V − 2V 2 ) 3(1 + 2Tr[%3 ] − 3V)
 
 
 
√1 (5 1 −1
+ 18Tr[%3 ] − 14Tr[%4 ] − 11V + 2V 2 ) (−1 − 4Tr[%3 ] + 2Tr[%4 ] + 5V − 2V 2 ) + 48Tr[%3 ] − 36Tr[%4 ] − 41V + 12V 2 3 4 2 
(5 + 18Tr[% ] − 14Tr[% ] − 11V + 2V 

 3 3
(17 √
3
 √ −1

(−1 − 4Tr[%3 ] + 8Tr[%4 ] + 5V) 3(1 + 2Tr[%3 ] − 3V) √
3
(5 + 18Tr[%3 ] − 14Tr[%4 ] − 11V + 2V 2 (5 + 8Tr[%3 ] − 5V)
(D.1)
238

Sparrow C 2018 PHD Thesis

Uploaded by

Copyright:

Available Formats

Sparrow C 2018 PHD Thesis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sparrow C 2018 PHD Thesis

Uploaded by

Copyright:

Available Formats

Quantum Interference in Universal

Linear Optical Devices for Quantum

A thesis submitted in accordance with the requirements for the degree of

1. J. Carolan, C. Harrold, C. Sparrow, E. Martı́n-López, N. J. Russell, J. W. Sil-

2. A. Holleczek, O. Barter, A. Rubenok, J. Dilley, P. B.R. Nisbet-Jones, G. Langfahl-

3. A. Neville, C. Sparrow, R. Clifford, E. Johnston, P. Birchall, A. Montanaro, A.

4. C. Sparrow, E. Martı́n-López, N. Maraviglia, A. Neville, J. Carolan, C. Harrold, N.

5. C. Sparrow*, P. Birchall* et al., ‘Linear Optical Quantum Computing with Partially-

* authors contributed equally to this work

2 Quantum Information and Computation 20

2.3.4 Quantum tomography . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Universal Linear Optics 100

5 Simulating Vibrational Quantum Dynamics of Molecules with Integrated

6 Linear Optical Quantum Computing with Partially-Distinguishable Pho-

6.4 Generating entanglement with

7 Mitigating Distinguishability Errors 174

A Universal Linear Optics 210

B Simulating the Vibrational Quantum Dynamics of Molecules with Inte-

C Linear Optical Quantum Computing with Partially-Distinguishable Pho-

D Mitigating Distinguishability Errors 238

2.1 Bloch sphere representation of pure qubit state . . . . . . . . . . . . . . . 28

3.1 Hong-Ou-Mandel dip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4.1 The Reck et al. scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5.1 Simulating the vibrational dynamics of molecules in the harmonic approx-

5.3 Quantum energy transfer and dephasing in NMA . . . . . . . . . . . . . . 132

7.1 Hong-Ou-Mandel Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

B.1 Four atom molecule data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

C.1 Fusion operations including local unitaries . . . . . . . . . . . . . . . . . . 229

1.1 Thesis Outline

Quantum Information and

2.2 Quantum Theory

2.2.1 Physical states

A complete description of a physical state in quantum mechanics is given by a ray in a

Physically measurable properties, known as observables, are described by Hermitian op-

A|φk i = ak |φk i (2.2)

Evolution in quantum mechanics is described by a unitary operator, U −1 = U † . In

where Ĥ is the Hamiltonian, the observable corresponding to a system’s energy. When

assuming no additional time dependence of A.

The measurement of a physical observable is a process in which a classical observer learns

Pr(k|ψ) = hψ|Πk |ψi (2.5)

2.2.6 Operational quantum theory

Trace one: Tr[ρ] = 1

Positivity: hψ|ρ|ψi ≥ 0 ∀ |ψi ∈ H (2.7)

can be described by the partial trace over the space HB ,

Pr(k|ρ) = Tr[Ek ρ] (2.10)

The most general form of quantum evolution is via completely-positive trace-preserving

Complete positivity: [E ⊗ 1](ρ) ≥ 0 ∀ρ ∈ HA ⊗ HB (2.12)

up to a global unitary transformation. Any quantum channel can be described by coupling

2.3 Quantum Information

2.3.1 The Qubit

Since an arbitrary pure qubit state can be parameterised as

Z : {|0i, |1i} (2.18)

which form the vertices of an octahedron centred on the origin.

represent an arbitrary qubit state by a vector ~r = (rx , ry , rz ) where

P(ρ) = Tr(ρ2 ). (2.22)

and they form a basis for two qubit states.

and is therefore a separable, classically correlated state even though it is a mixture of

C(ρ) = max[0, λ1 − λ2 − λ3 − λ4 ] (2.33)

S = ha0 b0 i + ha0 b1 i + ha1 b0 i − ha1 b1 i (2.37)

5. C. Sparrow, P. Birchall et al., ‘Linear Optical Quantum Computing with Partially-

||E − Λ|| = max ||(1 ⊗ E)[ρ] − (1 ⊗ Λ)[ρ]||1 (2.55)