UNIVERSITY OF SOUTHAMPTON

FACULTY OF SOCIAL, HUMAN AND MATHEMATICAL SCIENCES

Gravity Group

Timing variations in neutron stars:

models, inference, and their implications
for gravitational waves

Supervisors:
Author:
Dr. David Ian Jones
Gregory Ashton
Dr. Reinhard Prix

Thesis for the degree of Doctor of Philosophy

July 21, 2016

 UNIVERSITY OF SOUTHAMPTON

ABSTRACT

FACULTY OF SOCIAL, HUMAN AND MATHEMATICAL SCIENCES

Gravity Group

Doctor of Philosophy

TIMING VARIATIONS IN NEUTRON STARS: MODELS, INFERENCE, AND

THEIR IMPLICATIONS FOR GRAVITATIONAL WAVES
by Gregory Ashton

Timing variations in pulsars, low frequency ubiquitous structure known as timing noise
and sudden increases in the rotational frequency which we call glitches, provide a means
to study neutron stars. Since the first observations, many models have been proposed,
yet no definitive explanation has arisen.

In this thesis, we aim to improve this situation by developing models of timing noise. We
focus chiefly on precession models which explain periodic modulation seen in radio pulsar
data. Developing models and testing them provides an opportunity to infer the elemental
properties of neutron stars: evidence for long period precession has implications for the
superfluid component predicted by models used to explain glitches. However, often more
than one model can qualitatively explain the data, therefore we need a method to decide
which model best fits the data. This is precisely the case for PSR B1828-11 which has
been used as evidence for both precession and so-called magnetospheric switching. We
address this confusion by applying the tools of probability theory to develop a Bayesian
model comparison and find that the evidence is in favour of precession.

In the second part of this thesis, we will discuss the implications of timing variations
for the detection of continuous gravitational waves from neutron stars. To search for
these signals, matched filtering methods are used which require a template, a guess for
what the signal ‘looks like’. Timing variations, as seen in the electromagnetic signal,
may also exist in the gravitational wave signal. If detected, these could provide an
invaluable source of information about neutron stars. However, if not included in the
template, they may mean that the gravitational wave signal is not detected in the first
place. We investigate this issue for both timing noise and glitches, using electromagnetic
observations to predict for what types of gravitational wave searches this may be an issue.
We find that while timing noise is unlikely to be an issue for current gravitational wave
searches, glitches may cause a significant problem in all-sky searches for gravitational
waves from neutron stars.
One man’s ‘noise’ is
another man’s ‘signal’.
EDWIN THOMPSON JAYNES
Contents

Declaration of Authorship xi

Acknowledgements xiii

1 Introduction 1
1.1 Observation of pulsars and their identification with neutron stars . . . . . 2
1.2 Pulsar timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Categorising neutron stars . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 The physics of rotation powered pulsars . . . . . . . . . . . . . . . . . . . 7
1.5 Radio pulsar population statistics . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Neutron stars and gravitational waves . . . . . . . . . . . . . . . . . . . . 11
1.7 Plan of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Timing variations 15
2.1 Glitches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Timing noise: observations . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Timing noise: interpretations . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Random walk models . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Free precession . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.3 Magnetospheric switching . . . . . . . . . . . . . . . . . . . . . . . 25
Appendix 2.A Toy model of stochastic resonance: particle in a potential . . . 33

3 Action of the electromagnetic torque on a precessing neutron star 35

3.1 Defining the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Spherical star . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 Biaxial neutron star with no anomalous torque . . . . . . . . . . . . . . . 39
3.3.1 Neutron star A in region τS  τP . . . . . . . . . . . . . . . . . . . 41
3.3.2 Neutron star B in region τS ∼ τP . . . . . . . . . . . . . . . . . . . 46
3.3.3 Neutron star C in region τS  τP . . . . . . . . . . . . . . . . . . . 46
3.4 Biaxial neutron star including the anomalous torque . . . . . . . . . . . . 47
3.4.1 The effective body frame . . . . . . . . . . . . . . . . . . . . . . . 47
3.4.2 Phase space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4.3 Neutron star A in region τS > τA > τP . . . . . . . . . . . . . . . . 51
3.4.4 Neutron star B in region τS > τP > τA . . . . . . . . . . . . . . . . 53
3.4.5 Neutron star C in region τP > τS > τA . . . . . . . . . . . . . . . . 55
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Appendix 3.A Considerations for the timescales . . . . . . . . . . . . . . . . . 56

vii
viii CONTENTS

4 Modelling observations of precessing pulsars 59

4.1 Rotating into the inertial frame . . . . . . . . . . . . . . . . . . . . . . . . 60
4.1.1 Euler rotation matrices . . . . . . . . . . . . . . . . . . . . . . . . 60
4.1.2 Evolution of the Euler angles . . . . . . . . . . . . . . . . . . . . . 61
4.1.3 Initial conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2 Evolving the Euler angles . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.1 Torque free biaxial body . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.2 Torqued biaxial body . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3 Precessing pulsars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.1 The reference plane . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.2 Dynamics of the magnetic dipole . . . . . . . . . . . . . . . . . . . 68
4.3.3 Understanding the dynamics of the magnetic dipole . . . . . . . . 70
4.3.4 The effective wobble angle . . . . . . . . . . . . . . . . . . . . . . . 72
4.4 Observable features: the phase residual . . . . . . . . . . . . . . . . . . . . 73
4.4.1 Effect of precession on the phase residual . . . . . . . . . . . . . . 74
4.4.2 Effect of torqued precession on the phase residual: electromagnetic
amplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Observable features: the spin-down rate . . . . . . . . . . . . . . . . . . . 77
4.5.1 Derivation of the precession spin-down rate . . . . . . . . . . . . . 78
4.5.2 Simulations of the precession spin-down rate . . . . . . . . . . . . 80
4.6 Observable features: the pulse profile . . . . . . . . . . . . . . . . . . . . . 82
4.6.1 Variations in the pulse intensity . . . . . . . . . . . . . . . . . . . . 83
4.6.2 Variations in the beam-width . . . . . . . . . . . . . . . . . . . . . 84
4.7 Application: switching and precession . . . . . . . . . . . . . . . . . . . . 87
4.7.1 Switching in the spin-down torque only . . . . . . . . . . . . . . . 90
4.7.2 Switching in the spin-down and anomalous torque . . . . . . . . . 91
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

5 Comparing models of the periodic variations in spin-down and beam-

width for PSR B1828-11 95
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Bayesian Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.1 The odds-ratio and posterior probabilities . . . . . . . . . . . . . . 98
5.2.2 Signals in noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2.3 Choosing prior distributions . . . . . . . . . . . . . . . . . . . . . . 100
5.3 Defining and fitting the models . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3.1 Noise-only model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.3.2 Switching model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.3.3 Precession model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4 Estimating the odds-ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4.1 Thermodynamic integration . . . . . . . . . . . . . . . . . . . . . . 125
5.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4.3 Effect of the choice of prior . . . . . . . . . . . . . . . . . . . . . . 127
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Appendix 5.A Procedure for MCMC parameter estimation . . . . . . . . . . . 130
Appendix 5.B Implications for the unobserved beam . . . . . . . . . . . . . . 131
CONTENTS ix

6 Continuous gravitational waves: calculating the mismatch 133

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.2 Introduction to the mismatch . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.2.1 Defining the mismatch . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.2.2 Interpreting the mismatch . . . . . . . . . . . . . . . . . . . . . . . 135
6.2.3 Taylor expansion signals and templates . . . . . . . . . . . . . . . 136
6.2.4 The metric-mismatch approximation for fully-coherent searches . . 137
6.2.5 The mismatch for semi-coherent searches . . . . . . . . . . . . . . 137
6.3 Exact mismatch from irregularities in the phase . . . . . . . . . . . . . . . 139
6.3.1 Two subdomains with a phase discontinuity . . . . . . . . . . . . . 139
6.3.2 N subdomains with phase discontinuities . . . . . . . . . . . . . . . 140
6.3.3 Oscillating phase deviations . . . . . . . . . . . . . . . . . . . . . . 142
6.4 Generalising the metric-mismatch approximation to arbitrary signals . . . 143
6.4.1 Piecewise Taylor expansion . . . . . . . . . . . . . . . . . . . . . . 143
6.4.2 The generalised metric-mismatch approximation for fully-coherent
searches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.4.3 Explicit calculation of the metric . . . . . . . . . . . . . . . . . . . 146

7 Glitches in continuous gravitational waves 149

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.2 Continuous gravitational-wave searches . . . . . . . . . . . . . . . . . . . . 151
7.3 Statistical properties of the observed glitch database . . . . . . . . . . . . 152
7.3.1 Glitch magnitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.3.2 Overview of the population of glitches . . . . . . . . . . . . . . . . 154
7.3.3 Extrapolating: glitch magnitudes . . . . . . . . . . . . . . . . . . . 156
7.3.4 Extrapolating: average glitch rate . . . . . . . . . . . . . . . . . . 158
7.4 Calculating the mismatch due to a single glitch . . . . . . . . . . . . . . . 161
7.4.1 A single glitch in a fully-coherent search . . . . . . . . . . . . . . . 163
7.4.2 A single glitch in a semi-coherent search . . . . . . . . . . . . . . . 165
7.5 Predicting the mismatch and rate of glitches in gravitational wave searches168
7.5.1 Fully-coherent searches . . . . . . . . . . . . . . . . . . . . . . . . 169
7.5.2 Semi-coherent searches . . . . . . . . . . . . . . . . . . . . . . . . . 170
7.5.3 The follow-up stage . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.5.4 Including the recovery from glitches . . . . . . . . . . . . . . . . . 173
7.5.5 Application to past and future searches . . . . . . . . . . . . . . . 174
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Appendix 7.A Bayesian model comparison: test of mixture models . . . . . . . 176
Appendix 7.B Linear regression in log-space . . . . . . . . . . . . . . . . . . . 179
Appendix 7.C Understanding the uncertainty in the predictions . . . . . . . . 179

8 Timing noise in continuous gravitational waves: a numerical study 183

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.2 Timing noise as described by the Crab ephemeris . . . . . . . . . . . . . . 186
8.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.4.1 The effect of timing noise on narrow-band searches . . . . . . . . . 189
8.4.2 Results relevant to recent narrow-band searches . . . . . . . . . . . 190
x CONTENTS

8.4.3 Minimum mismatch as a function of the observation epoch . . . . 191

8.4.4 Averaged minimum mismatch as a function of the observation
duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

9 Timing noise in continuous gravitational waves: random walk models197

9.1 Defining a random walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.1.1 Initial definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.1.2 Writing the parameter offsets in terms of normal distributions . . . 200
9.2 Random walk models: a simple treatment . . . . . . . . . . . . . . . . . . 201
9.2.1 Taking the expectation . . . . . . . . . . . . . . . . . . . . . . . . 202
9.2.2 Verifying the results . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.3 Random walk models: minimising the mismatch . . . . . . . . . . . . . . 203
9.3.1 Random walk in the phase . . . . . . . . . . . . . . . . . . . . . . 204
9.3.2 Random walk in the frequency . . . . . . . . . . . . . . . . . . . . 206
9.3.3 Verifying the results . . . . . . . . . . . . . . . . . . . . . . . . . . 207
9.4 Understanding the random walk model . . . . . . . . . . . . . . . . . . . . 207
9.5 Application to the Crab pulsar . . . . . . . . . . . . . . . . . . . . . . . . 210
9.5.1 Distribution of jumps in the Crab ephemeris . . . . . . . . . . . . 211
9.5.2 Predicting the mismatch in the Crab . . . . . . . . . . . . . . . . . 213
9.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Appendix 9.A Summation identities . . . . . . . . . . . . . . . . . . . . . . . . 215
Appendix 9.B Least-squares minimisation of a random walk . . . . . . . . . . 216
9.B.1 Least squares fitting of a polynomial . . . . . . . . . . . . . . . . . 216
9.B.2 Least squares fitting a polynomial to a random walk . . . . . . . . 217
9.B.3 Zeroth order fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
9.B.4 First order fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
9.B.5 Second order fitting . . . . . . . . . . . . . . . . . . . . . . . . . . 221
9.B.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

10 Conclusion and outlook 223

 Declaration of Authorship

I, Gregory Ashton, declare that the thesis entitled Timing variations in neutron stars:
models, inference, and their implications for gravitational waves and the work presented
in the thesis are both my own, and have been generated by me as the result of my own
original research. I confirm that:

• this work was done wholly or mainly while in candidature for a research degree at
this University;

• where any part of this thesis has previously been submitted for a degree or any
other qualification at this University or any other institution, this has been clearly
stated;

• where I have consulted the published work of others, this is always clearly at-
tributed;

• where I have quoted from the work of others, the source is always given. With the
exception of such quotations, this thesis is entirely my own work;

• I have acknowledged all main sources of help;

• where the thesis is based on work done by myself jointly with others, I have made
clear exactly what was done by others and what I have contributed myself;

• parts of this work have been published as: [24] and [25].

Signed:.......................................................................................................................

Date:..........................................................................................................................

xi
 Acknowledgements

I would like to thank Ian Jones, my supervisor in Southampton, for his tireless efforts to
guide my research in a meaningful direction, his clear and honest evaluation, but most
importantly for being a friendly and enthusiastic mentor. Reinhard Prix, my supervisor
in Hannover, also deserves a hearty thanks for his enduring commitment, many useful
insights, and for introducing me to Bayesian probability theory, without which much of
this thesis would not exist. It has been a pleasure to work with both Ian and Reinhard.
Thanks also to my advisor, Nils Andersson, for his useful input early on and his continued
humour throughout such as his brilliant suggestion to title the work ‘A waste of timing
noise’. To my examiners, Graham Woan and Wynn Ho, I extend my thanks for taking
the time to provide detailed and insightful comments which I thoroughly appreciate.

For the work presented in Chapter 5, I am grateful to Will Farr, Danai Antonopoulou,
and Ben Stappers for valuable discussions and comments, Dan Foreman-Mackay for the
software used in generating posterior probability distributions [61], and Andrew Lyne
[111] for generously sharing the data for PSR B1828-11. For the work presented in
Chapter 7, I kindly thank Christobal Espinoza for maintaining the glitch catalogue and
for helping to extract the required data. I would also like to express my gratitude to
Matthew Pitkin whose advice on MCMC simulation software was of great help.

All of my friends in the Maths departments deserve my praise both for their moral
support and making Southampton an enjoyable place to live and work. To my final year
companions, Vanessa Graber, Marta Colleoni, Liana Kontogeorgaki, and Yafet Sanchez,
I have thoroughly enjoyed the adventure we had together and I wish you all the best
of luck in your future pursuits. I particularly would like to acknowledge Yafet for his
willingness to question everything and reminding me that while life may be difficult,
consider the alternative. Matthew Cobain, Marc Scott and Andrew Meadowcroft, I
thank for their kind help in editing this thesis, many hours of badminton, many enjoyable
discussions, and many gin & tonics.

My heartfelt thanks to Emily Lawrance for her solid support, wonderful cooking, and
for keeping me sane; to Christine Ashton for her everlasting enthusiasm and words of
wisdom; and finally to the man who inspired me to such heights, Mark Ashton.

xiii
Chapter 1

Introduction

Neutron stars were first postulated by Landau as ‘dense stars which look like giant atomic
nuclei’ [171], even before the discovery of the neutron by Chadwick [38]. However, it
was Baade and Zwicky [26] who made the explicit prediction of a neutron star whilst
trying to explain the energy released in observed supernova.

In a main-sequence star, the nuclear fusion of hydrogen atoms into helium provides
outward pressure balancing the star in an equilibrium configuration with the inward
pressure of the star’s self-gravity. Eventually the star depletes its reserves of hydrogen
and can no longer maintain equilibrium. If the star has an initial mass greater than
∼ 8M , then it may undergo a core-collapse supernova during which some of the mass is
ejected, but the rest falls in creating a new compact object. In this object, temperatures
and pressure rapidly soar and the electrons and protons undergo inverse beta decay
combining to form neutrons and neutrinos:

e− + p → n + ν. (1.1)

If the compact object’s mass is less than the maximum mass of a stable neutron star
[130], then, once the pressure reaches nuclear densities of ∼ 2.3 × 1014 g/cm3 , neutron
degeneracy pressure can halt the collapse in a new compact stable equilibrium config-
uration which we call a neutron star. The exact value of the maximum mass depends
on the equation of state of matter at high densities, but typical values range from 1.5
to 3 M [32]. For remnants with larger masses, this is not possible and the object will
collapse to form a black-hole; the detail of exactly what the critical mass a neutron star
can sustain is sensitive to the equation of state of matter under these conditions.

Our knowledge of neutron stars is founded on observations made by electromagnetic

astronomy. This has revealed a wealth of different neutron stars and their phenomena
which we will introduce in the following sections. Many of these observed phenomena
have well defined models which allow us to infer properties of neutron stars and their
environments. However, our knowledge of neutron stars is far from complete: current
interpretations of the observations can be contradictory or have features not explained
by any known physical models. Improvements in electromagnetic astronomy will bring

1
2 Chapter 1 Introduction

to light a greater number of new neutron stars and improve the resolution of those
currently observed; it is hoped that this will help us to better understand them.

There are two other methods we can utilise to learn more about neutron stars: improved
modelling of current observations and by observing them from their gravitational wave
emissions. In this thesis, we will study how one of the observed phenomena, so-called
‘timing variations’, can help us to learn more from current observations and also test
whether it may hinder the current search for gravitational waves from neutron stars.

In this introduction, we will acquaint the reader with the current observations of neutron
stars and introduce some basic physics.

1.1 Observation of pulsars and their identification with

neutron stars

After the conception of neutron stars as stable compact objects there was thought to be
little chance of observing them. They are many orders of magnitude smaller than other
celestial objects and soon after their formation in rare supernovae events (van den Bergh
[160] predicts just 3 supernovae per century in the Milky Way Galaxy) they are rapidly
cooled by the emitted neutrinos making their thermal emission difficult to detect.

In 1968 a bright periodic electromagnetic (EM) signal, now known as PSR B1919+21,
was identified by Hewish et al. [78] during a high time-resolution survey for interplanetary
scintillation. The source was measured with a radio frequency of 81.5 MHz and pulsed
with a period of ∼ 1.377 s; this led to the name pulsars to refer to such sources.
Following this, several other similar objects where discovered. A unifying feature of
all pulsars is the clock-like stability of the pulsations - something not rivalled by any
other astrophysical phenomenon. This stability indicates that the source must be a
collimated beam fixed to a rotating body such that, as it rotates, the beam sweeps out
like a lighthouse; in this way the pulsation period is exactly the rotation period of the
body. Alternative models such as emission due to accretion from a binary companion
could never reach the stability’s seen in pulsars.

For any rotating body to remain gravitationally bound, its rotation frequency is con-
strained by the requirement that the centrifugal acceleration at the equator be less than
the gravitational acceleration. The rotation frequency of the observed pulsars ruled out
all known astrophysical bodies except the two most compact objects, neutron stars and
black holes, since all other bodies would not be gravitationally bound at these frequen-
cies. Isolated black holes are unable to support an electromagnetic emission mechanism.
This left only neutron stars as candidates.

The explicit identification of pulsars with neutron stars came from Gold [68], however,
Pacini [131] had already predicted the basic ingredients of the model before the observa-
tion. They suggested that a rapidly rotating neutron star with a strong dipolar magnetic
field would stream radiation out along the magnetic axis. If this axis is misaligned from
the rotation axis, then the beams are swept out like a lighthouse. Beams passing over
Chapter 1 Introduction 3

the earth are observed as periodic pulses at the rotation frequency of the star. Such a
model predicts that the electromagnetic radiation should exert a torque and slow-down
the rotation. This slowdown was subsequently measured in the Crab pulsar, which was
discovered at the centre of the Crab nebulae, a supernova remnant, agreeing with the
prediction of Baade and Zwicky [26].

These early detections gave birth to a new field of astronomy: pulsar astronomy. Since
then, researchers in the field have detected over 2000 radio pulsars, measured thermal
emission from a handful of nearby neutron stars with typical temperatures of 105 K
to 106 K [133], found neutron stars in accreting binary systems with companions, and
identified many other ways to observe neutron stars. We will discuss some of these
which are relevant to this thesis in Section 1.3, but first in Section 1.2 we describe the
techniques used by the field to identify and ‘time’ radio pulsars.

1.2 Pulsar timing

Pulsars can be observed by measuring the variation in amplitude of the radio waves
from a particular sky location. A single observation consists of measuring the amplitude
over a time period of approximately 30 minutes or so, which, for pulsars with spin
period ∼ 1 s, means recording up to several thousand individual pulsations.

The shape of individual pulses can vary substantially during a single observation; to
demonstrate this, in Figure 1.1, successive pulses from PSR B1919+21, the first discov-
ered pulsar, are vertically stacked and aligned. Each pulsation lasts for a small fraction
of the pulse period. As an example, the pulses in PSR B1919+21, shown in Figure 1.1,
have typical widths of 0.031 s, but the pulse period is 1.337 s; note that the stacked plot
is truncated to show only the pulsation itself. Later, in Figure 1.6, we will show this
to be a common characteristic for the normal radio pulsar population by looking at the
duty-cycle, the ratio of the pulse duration to the pulse period.

In order to understand the gross features of a pulsar, astronomers average over the
hundreds to thousands of pulses observed during a single observation to create a single
integrated pulse profile. This is done by sampling the radio signal at fixed time intervals
then ‘folding’ all the samples at the pulse period (for a complete review see Chapter 15
of Lyne and Graham-Smith [110]). In contrast to the individual pulses which, as shown
in Figure 1.1, can be highly variable, the integrated pulse profile is highly stable between
independent observations over timescales of years.

The integrated pulse profile not only gives a stable picture of what the pulsations look
like on average, but it also provides a highly accurate measurement of the time of arrival
(TOA) of a single pulse during the observation. It is this TOA which can be used to ‘time’
a pulsar. To do this, regular observations of a pulsar must be made every few months or
so, with each observation resulting in a precise TOA measurement. Having obtained a
series of TOAs, pulsar astronomers generate a timing model which attempts to exactly
count each and every pulse. Between any two observations there may be several million
pulses so the timing model needs to account for any mechanisms which may produce
4 Chapter 1 Introduction

Figure 1.1: The radio amplitude of successive pulses from PSR B1919+21
stacked vertically, figure reproduced from Mitton [126], originally produced by
Craft [50].

variations in the TOAs. The process is standardised by the software package TEMPO2
developed by Hobbs et al. [80], of which we will now describe the essential features.

The TOA of a pulse at the detector on Earth depends on many factors, such as: the time
at which the beam was directed by the source towards the Earth; the relative motions
of the source and detector; and any mechanisms effecting the signal during its transit.
In this thesis, we will be concerned only with the time at which pulses are generated
(when the source beams towards the earth) which is governed by the timing properties
of the star itself. As described by Edwards et al. [57] these can be modelled by a Taylor
expansion in the phase at time t, given by

X ν (n−1)
φ(t) = φ0 + 2π (t − tref )n , (1.2)
n!
n≥1

where {ν (n) } is the nth time derivative of the spin frequency of the rotating body, φ0
is the initial phase, and tref is an arbitrary reference time. This expansion is usually
truncated at n = 3, the second order spin-down rate ν̈. The timing model then includes
corrections to this to model the relative motion of the source and detector, intergalactic
transit, and other effects; these are described in full in Edwards et al. [57].

Between any two TOAs, if the timing model is correct, an integer number of rotations
must have occurred; this allows the use of the deviation of φ(tjT OA ) from an integer
as a test statistic. The timing model minimises the root-mean-square (RMS) of these
deviations with respect to the timing model parameters, for example the frequency and
Chapter 1 Introduction 5

frequency derivatives. The output of applying a particular timing model (choice of

corrections) to a set of data is then the best-fit of these parameters and an estimate
of their associated errors. The corrections applied in a timing model provide a method
to investigate pulsar physics: for example in some pulsars an orbital correction must
be applied which models the periodic motion of the star due to an orbital companion.
Using this technique, Wolszczan and Frail [170] discovered the first exoplanets orbiting
the pulsar PSR B1257+12.

A minimisation of the timing model parameters will converge regardless of whether of not
the model itself is appropriate. To qualitatively check if the fitted model described the
data, pulsar astronomers refer to the timing residual, which is the difference between the
TOA, as given by the timing model, and the actual TOA. The timing residual provides a
mechanism to evaluate the timing model: for example a periodic variation in the timing
residual with period 365 days may indicate the correction of the Earth’s orbit about the
Sun may be incorrect. If the timing model is correct, the residual data points should be
Gaussianly distributed around zero. A timing model is described as phase-connected if it
is accurate enough to track the pulsar to within a single rotation. For most pulsars this
is the case and a single set of coefficients can track the spin-down over periods greater
than a year.

However, for all pulsars the timing residual contains ‘structure’ known as timing varia-
tions which cannot be associated with any known correction. These variations are the
focus of this work and we will describe the details further in Section 2. In the next
section, we will describe the variety of known pulsars which have been timed using this
method.

1.3 Categorising neutron stars

The timing properties, and other features measured by the timing model, for over 2000
pulsars observed can be accessed via the Australia Telescope National Facility (ATNF)
pulsar catalogue [117]. We can categorise the population by their measured values of
period P and the period derivative Ṗ . This is done by plotting them in a so-called P − Ṗ
diagram, as shown in Figure 1.2. The various categories to which each pulsar can be
assigned have been marked in this plot and we now discuss their features.

The majority of pulsars (referred to as the ‘normal’ pulsars) are found isolated without a
binary companion and have typical periods of P = 10−2 − 101 s. These can be described
as rotation powered pulsars, since the electromagnetic (EM) radiation is powered by the
loss of rotational energy. As described later in Section 1.4, estimates can be made of their
characteristic age, τage , and surface magnetic field strength, B0 , based on a dipole spin-
down model. Constant lines of these quantities are plotted in Figure 1.2. Of the normal
pulsars, we can choose to identify the young pulsars as those for which τage < 105 yrs.
Some of these, such as the Crab and Vela pulsars, can be directly associated with their
supernova remnant from which they were formed [95].
6 Chapter 1 Introduction

10−8 B
s =1
Normal pulsars 0 16
−9 G
10 Young pulsars
Binary pulsars
10−10 MSP B
s =1
Binary MSP 0 15
G
10−11 Magnetars

10−12 1 yrs
0 B
=1 s =1
e
Period derivative [s/s]

τ ag 0 14
−13 G
10

10−14 3 yrs
0 B
=1 s =1
e
τ ag 0 13
−15 G
10

10−16 5 yrs
0 B
=1 s =1
e
τ ag 0 12
−17 G
10

10−18 7 yrs
0 B
=1 s =1
e
τ ag 0 11
G
10−19

10−20 9 yrs
0 B
=1 s =1
e
τ ag 0 10
−21 G
10
10−4 10−3 10−2 10−1 100 101 102
Period [s]

Figure 1.2: Period - Period derivative diagram using data taken from the ATNF
pulsar catalogue [117]. Dashed lines show inferred characteristic ages and sur-
face magnetic fields as given by Eqn. (1.10) and Eqn. (1.13) respectively.

A second smaller population of isolated rotation powered pulsars exists with P < 10−1 s.
These are known as the millisecond pulsars (MSPs). This special class of pulsars are
believed to start their life as normal pulsars, but are then spun-up through accretion
from a normal star. In support of this hypothesis, the majority of MSPs in Figure 1.2
have a binary companion [169]. Additionally, we see so-called low-mass X-ray binary
systems (LMXBs) which are systems where a neutron star in a binary accretes matter
from its companion; the infalling matter releases gravitational potential energy in the
form of X-rays (see for example pg. 73 of Lewin et al. [105]). It is thought that these
LMXBs are the progenitors of the MSPs; recent results of ‘transitional systems’ (see for
example [20]) which switch between the two, appear to confirm this.
Chapter 1 Introduction 7

We include one final class of neutron stars, magnetars, thought to have large magnetic
fields of B & 5 × 1013 G. These are in fact observed from two channels which we will now
describe. Some pulsars are observed to emit X-ray radiation; usually this is powered by
the accretion of matter from a binary companion, but this mechanism does not apply
to isolated stars. Subsequently, isolated stars observed in the X-ray band where named
anomalous X-ray pulsars (AXPs). It was shown by Duncan and Thompson [56] that
AXPs are magnetars where the emission is powered by the decay of the strong magnetic
field. At the same time, astronomers found a class of objects emitting irregular bursts
of γ-rays or X-rays which they named the soft γ-ray repeaters (SGRs). As discussed
in Kouveliotou et al. [96] these are now understood to be magnetars which undergo
rearrangement of their magnetic fields. The two individual observations where unified
by observation of X-ray bursts from AXPs by Gavriil et al. [63]. In Figure 1.2 we label
observations from both these sources as magnetars.

1.4 The physics of rotation powered pulsars

For rotation powered pulsars, the normal population in Figure 1.2, we can infer a sub-
stantial amount about their physics by applying a simple model to their observed timing
properties. In this section, we will introduce such a model and acquaint the reader with
methods to infer the spin-down ages and magnetic fields, important quantities in under-
standing neutron stars.

Let us model the star as described by Pacini [131] and Gold [68] and illustrated in
Figure 1.3: a rapidly rotating body with a magnetic dipole fixed in the crust at an angle
α to the rotation axis.
Rotation axis

Dipole axis

Figure 1.3: An illustration of the dipole spin-down model. The dipole and some
of the closed field lines are fixed at an angle α to the rotation axis. As the
body rotates, radiation is emitted along both ends of the dipole axis producing
a torque on the body.
8 Chapter 1 Introduction

From §67 of Landau and Lifshitz [101], the total radiation from a dipole rotated at
angular frequency Ω can be shown to be given by

2 Ω4 2
Υ= d , (1.3)
3 c3 0
where d0 is the projection of the dipole moment on the plane perpendicular to the
axis of rotation [131]. We note that we could alternatively parameterise with the pulse
period P = 2π Ω . Following the arguments in §10.5 of Shapiro and Teukolsky [154], the
magnitude of the magnetic dipole moment for a star with radius R and surface magnetic
field strength B0 is B0 R3 /2. Including the projection onto the plane perpendicular to
the rotation axis, the total radiation is then

1 Ω4 2 6 2
Υ= B R sin α. (1.4)
6 c3 0

The rotational energy of a body spinning at Ω with a moment of inertia I0 is given by

1
E = I0 Ω2 . (1.5)
2
Differentiating this expression with respect to time gives the loss of rotational energy,
Ė = I0 ΩΩ̇ where Ω̇ is the angular spin-down rate which can be related to the changing
pulse period by Ṗ = −2π ΩΩ̇2 . Assuming that all the energy is lost to the rotation of
the dipole, hence the name rotation powered pulsars, we can equate Ė = −Υ. We
then rearrange to give a power-law relation between the spin-down rate and the spin-
frequency:

B02 R6 sin2 α 3
Ω̇ = − Ω . (1.6)
6c3 I0

This power-law dependence is a model specific version of a more general phenomenolog-

ical power-law braking model
Ω̇ = −kΩn . (1.7)

Generalising in this way suggests a powerful method to determine the type of braking for
a given pulsar. Specifically, differentiating Eqn. (1.7) and rearranging it can be shown
that
Ω̈Ω
n= . (1.8)
Ω̇2
Therefore, if Ω̈ can be measured, then n can be determined, and hence used to infer
the type of braking. For example, measuring n = 3 would indicate the pulsar braking
is dominated by losses due to the magnetic dipole, in contrast, it can be shown that
gravitational wave braking would produce n = 5 [154, pg.284]. Unfortunately, in reality,
pulsars do not constrain this value. Work by Biryukov et al. [29] found (see Figure 1.4)
that younger pulsars tend to have braking indices of the correct order of magnitude.
However, beyond τch ≈ 105 years the absolute value of the braking index rapidly grows,
reaching values as large as 106 for the oldest pulsar. In addition, an almost equal number
of pulsars have positive and negative values of the braking index.
Chapter 1 Introduction 9

Figure 1.4: The inferred observed braking index nobs against the characteristic
age τch = τage values for 1337 ordinary rotation powered pulsar, figure repro-
duced from Biryukov et al. [29].

To infer the age of the pulsar, Eqn. (1.7) can be integrated between the initial values
(t = 0, Ω = Ωi ) and the observed value (Ωo ) to give
 
1 Ωo Ωn−1
t= 1 − on−1 . (1.9)
(1 − n) Ω̇o Ωi

Typically, we make the assumption that all pulsars, regardless of their measured braking
index, are dominated by EM braking such that n = 3. Then additionally assuming that
Ωi  Ωo we can approximate to a characteristic age

Ωo P
τage = − = . (1.10)
2Ω̇o 2Ṗ

To infer the approximate surface magnetic field strength, we first note that in the EM
dipole braking model:

B02 R6 sin2 α
k= . (1.11)
6c3 I0

Then rearranging Eqn. (1.6) we can estimate the surface magnetic field strength at the
poles by
  21 !1   21 p
6c3 I0 −Ω̇ 2 1 6c3 I0
B0 = = P Ṗ (1.12)
R6 sin2 α Ω3 2π R6 sin2 α

In general we do not know the inclination angle α, but we can evaluate a minimum
magnetic field strength by setting α = π/2. In CGS units, for a canonical pulsar with
R = 106 cm, Ip 0 = 10
45 g cm2 , we can approximate the magnetic field strength as

B0 = 6.4 × 1019 P Ṗ Gauss. It is conventional, however, to quote the magnetic field at

the equator Bs which differs by a factor of 2 (see Lyne and Graham-Smith [110] pg. 71)
10 Chapter 1 Introduction

such that p
Bs = 3.2 × 1019 P Ṗ Gauss. (1.13)

In this section, we have introduced some of the simple results that can be obtained by
modelling the time evolution of pulsars with a power law. Many advancements can be
made on this model, such the existence of a magnetosphere predicted by Goldreich and
Julian [70], but in practice this model is consistent with most pulsar observations and
provides a useful way to categorise them via their spin-down age and magnetic field. We
will frequently refer back to this model as it is a useful platform from which to begin
understanding neutron stars.

1.5 Radio pulsar population statistics

Radio pulsars make up the majority of the observed neutron star population and will
be the focus of discussion in this thesis. In this section, we will provide some simple
empirical population statistics for the normal radio pulsar population: we ignore the
millisecond population since they are disjoint from the normal population and have
a distinct history, but include the young pulsars. More detailed Monte Carlo based
population synthesis studies have been performed by Faucher-Giguere and Kaspi [60]
and Popov et al. [138] in which substantive models were made which allow inferences
to be performed for the underlying population distribution of all neutron stars. In this
section, we provide only summaries of the observations and will discuss their results when
relevant. All data in this section is taken from the ATNF pulsar catalogue Manchester
et al. [117] and it should be stated that in each case the observed property is an average
over all observations made for each pulsar.

For each observable property of the population of neutron stars (such as the frequency),
we will present the data as a histogram choosing an appropriate binning size in each
instance. In order to make simple inferences about the population, we will also calculate
the mean and standard-deviation. We will test normality using the Jones et al. [92]
implementation of the d’Agostino [51] test. This results in a p-value, which, if less than
0.05, rejects the hypothesis that the data is normal with 95% confidence. We will give
this p-value in the legend for each observable property and show the normal distribution
with the calculated mean and standard-deviation.

In Figure 1.5 we present the data for the three timing properties measured directly
from the pulsar timing models. For normal radio pulsars the pulsation frequency, ν,
can always be accurately measured provided at least one observation has been made.
Several precise observations of a pulsar must be made in order to measure the higher
order derivatives of the frequency. As a result, the pulsar catalogue contains missing
information and the number of data points for ν̇ and ν̈ is smaller than the total observed
number of pulsars: the exact numbers are given in the caption. By eye, the histograms
are clustered and appear to be approximately normal. However, for all three properties,
the normal hypothesis is rejected; we note that the level of rejection is dependent on the
number of data points. This rejection is not surprising given the complicated physics
Chapter 1 Introduction 11

mean=0.216 Norm. fit mean=14 Norm. fit mean=26 Norm. fit

std. dev.=0.361 p-val.≈ 10-11 std. dev.=1.11 p-val.≈ 10-6 std. dev.=1.46 p-val.≈ 10-5
histogram histogram histogram

Normalised count

−1 0 1 2 −20.0 −17.5 −15.0 −12.5 −10.0 −30 −27 −24 −21 −18
log10 (ν/s−1 ) log10 (ν̇/s−2 ) log10 (ν̈/s−3 )

Figure 1.5: The distribution in log-space of the frequency ν and the first two
frequency derivatives ν̇ and ν̈ for normal radio pulsars in the ATNF pulsar
catalogue. Appropriate bin sizes were selected for each quantity. The population
sizes are 1942, 1686, 339 for ν, ν̇, ν̈ respectively.

which governs these systems. The detailed population synthesis studies by Faucher-
Giguere and Kaspi [60] and Popov et al. [138] are able to relate the observed features to
the underlying physics and find similar results for the period and period derivative.

In Figure 1.6 we present some other interesting quantities held in the ATNF catalogue.
Firstly, in the left-hand panel we plot the characteristic age as defined in Eqn. (1.10).
Then, in the middle panel we give a measure of the pulsar beam-width W10 . Specifically,
W10 is the width of the integrated pulse profile (in seconds) at 10% of the integrated
pulse profile maximum. In the right-hand panel we plot W10 ν, i.e. the product of the
beam-width and frequency for each pulsar. This gives information about the effective
duty-cycle: the ratio between the pulse duration and period. Notably, the majority
of pulsars have duty-cycles substantially less than a 0.5 indicating that the pulses are
short compared to the period. In this instance, the normal hypothesis is rejected for
the beam-width and duty-cycle, but accepted for the spin-down age. It would be an
interesting exercise to investigate this further.

These results discussed in this section provide an overview of the observed radio pulsar
population. It must be remembered that these observed pulsars are a sample from what
may be a much larger population. These summaries are intended only to give a brief
overview of the observations.

1.6 Neutron stars and gravitational waves

Gravitational waves (GWs) were first predicted by Albert Einstein in 1916 [58] when he
found that the linearised weak-field equations of his General Theory of Relativity had
transverse wave solutions. Much like the generation of electromagnetic waves requires
12 Chapter 1 Introduction

mean=6.76 Norm. fit mean=-1.45 Norm. fit mean=-1.25 Norm. fit

std. dev.=0.997 p-val.= 0.35 std. dev.=0.344 p-val.≈ 10-6 std. dev.=0.303 p-val.≈ 10-21
histogram histogram histogram
Normalised count

2 4 6 8 10 12 −2.4 −1.6 −0.8 0.0 −2.4 −1.6 −0.8 0.0

log10 (τage /yrs) log10 (W10 /s) log10 (W10 ν)

Figure 1.6: The distribution in log-space of the characteristic age τage , the W10
measure of the beam-width, and the effective duty-cycle W10 ν for normal radio
pulsars in the ATNF pulsar catalogue. Appropriate bin sizes were selected for
each quantity. The population sizes are 1942, 915, and 915 for τage , W10 , and
W10 ν.

the acceleration of electrical charges, GWs are generated by any source with a time-
varying mass quadrupole moment and can be understood as ‘ripples’ or spatial strains
in the spacetime itself which travel at the speed of light.

Gravitational waves were first directly detected by the LIGO collaboration [12]. They
observed a signal consistent with the inspiral and merger event of two ∼ 30M black-
holes over approximately 0.2 s. To detect such signals, LIGO uses a laser interferometer
to measures the relative change in length between two orthogonal arms. In particular, if
L is the length of either arm without a signal and a gravitational wave passes through,
the detector measures the strain
δLx − δLy
h(t) = (1.14)
L
where δLx and δLy are the time-varying stretching and squeezing of the two arms caused
by the gravity wave. For the observed binary black-hole merger the peak strain in the
detector was ∼ 10−21 .

Prior to this detection, indirect evidence for the existence of gravitational waves was
found by observing the orbital periods of compact binary systems. Such systems have
a time-varying quadrupole moment and emit gravitational waves, which radiate energy
away from the system causing, a decay of the orbital period. In 1975, Hulse & Taylor dis-
covered a binary neutron star system where one of the stars, PSR B1913+16, was visible
as a pulsar [83]. Due to the powerful techniques of pulsar timing, subsequent analysis
by Taylor and Weisberg [159] was able to verify that the orbital decay matched exactly
the predictions of General Relativity. Since this observation, more double neutron star
system have been discovered, including a system, PSR J0737-3039A/B, discovered by
Burgay et al. [37], where both neutron stars are seen as pulsars. This so-called double
pulsar system tests the agreement with General Relativity at the 0.05% level [98].
Chapter 1 Introduction 13

Neutron stars observed as pulsars are often referred to as ‘cosmic clocks’ for the regularity
of their pulsations. The most stable pulsars are the radio millisecond pulsars (MSPs),
which, due to their stability, many workers in the field utilise in an attempt to search
for GWs via a pulsar timing array [79]: this searches for correlated signatures in the
TOAs from a network of well-timed MSPs. Such a detector is sensitive to a stochastic
background of gravitational waves by measuring the so-called Hellings & Downs curve
[77], or to the mergers of super-massive black hole binary systems [103].

Isolated neutron stars themselves are potential sources of gravitational waves through
one of three mechanisms. If the star has a rotation axis misaligned with its symmetry
axis then it will undergo precession: a ‘wobble‘ of the star which has a time-varying
quadrupole moment. This will produce GWs at the rotation frequency and twice the
rotation frequency, but the small amplitudes of possible sources and questions over
how long lived they might be make this an unlikely candidate for LIGO [91]. If the
neutron star is subject to non-axisymmetric instabilities, such as the r-mode instability
in newborn and rapidly accreting neutron stars [19], then these too can produce GWs
(for a review see Andersson [17]). Finally, if the star possesses a non-axisymmetric
distortion, , also known as a ‘mountain’, it will produce a continuous gravitational
wave at twice its rotation frequency with a strain amplitude proportional to . The
LIGO detectors have already been used to search for signals from known neutron stars
and, by not observing any radiation, are able to place upper limits on  (see for example
Abbott et al. [11], Abadie et al. [9]).

All three of these detection mechanisms are potential sources of the first detection of
gravitational waves from neutron stars and realising this would provide a unique op-
portunity to learn about neutron stars. But is it feasible? A statistical argument can
be made for the ‘loudest expected signal from unknown isolated neutron stars’. This
argument is given in Abbott et al. [10], although the origin can be dated back to Bland-
ford (1984) as attributed by Kip Thorne in Hawking and Israel [75]. Essentially, one
makes the assumption that the population of 105 neutron stars predicted to exist in
our galaxy by stellar evolution models are all born with a high spin-down rate and
subsequently spin-down principally due to the emission of gravitational waves. With
additional assumptions that the stars are born randomly throughout the Galactic disk
with a constant birthrate the populations are transformed into a population of neutron
star strains. Then it is shown that there is a 50% chance a source exists with a strain
amplitude

h0 ∼ 4 × 10−24 , (1.15)

which is close to ‘detectable’ by LIGO, although the exact details depend on the source
frequency and duration. While this is a purely statistical argument, and changing any
of the assumptions tends to decrease this signal strain [140], the rewards for detection
in terms of astrophysics are sufficient to motivate further research.
14 Chapter 1 Introduction

1.7 Plan of the thesis

Following this introductory chapter, we will introduce timing variations in pulsars:

glitches and timing noise. This will familiarise the reader with the observed phenomena
and describe the current state of modelling. We will provide some original work on
simple ways in which the models could be tested.

In the next four chapters we evaluate models of timing noise in the face of current obser-
vations and attempt to constrain the models. In Chapter 3 we explore how precession,
a potential ingredient to explain timing noise, can be described, when viewed from the
frame rotating with the star. Following this, Chapter 4 looks at how precession will
manifest in the observations made by pulsar astronomers. In Chapter 5 we perform a
rigorous quantitative model comparison between precession and the leading alternative,
magnetospheric switching, for describing timing variations seen in PSR B1828-11.

In the final three chapters we approach another important aspect of timing variations
for neutron stars: the effect they may have on our ability to detect gravitational waves
from neutron stars. In Chapter 6 we introduce the methods and formalisms used by
gravitational wave astronomers before analysing the effect of glitches on gravitational
wave searches in Chapter 7. For the effect of timing noise on gravitational wave searches,
we perform a numerical study on data from the Crab pulsar in Chapter 8 and then model
the effect of different timing noise interpretations in Chapter 9. Finally, we will conclude
in Chapter 10.
Chapter 2

Timing variations

Timing variations are, broadly speaking, any time when the usual Taylor expansion
in the phase Eqn. (1.2) (typically up to the second-order spin-down rate ν̈) does not
accurately describe the phase evolution of the pulsations. Two distinct types of vari-
ations exist: the sporadic event-like glitches seen in some pulsars and the ubiquitous
timing noise present, at some level, in all pulsars. In this chapter, we will discuss the
observations of these variations and the models proposed to understand them.

2.1 Glitches

In addition to the regular spin-down of isolated radio pulsars due to magnetic braking,
some pulsars undergo anomalies in their timing solutions known as glitches. These are
sudden rapid increases∗ in the pulsation frequency which were first observed in the Crab
[34, 147] and Vela pulsars [144, 145]. Pulsar timing methods model this as a permanent
increase in the phase, frequency, and first frequency derivative in addition to a frequency
increment that subsequently decays exponentially to zero. To model this, for each glitch
pulsar astronomers add on an additional term to Eqn. (1.2) which is Eqn. (121) of
Edwards et al. [57]† :
    
∆ν̇ 2 t − tg
φg = H(t − tg ) ∆φ + ∆ν(t − tg ) + (t − tg ) + 1 − exp − ∆νt τ ,
2 τ
(2.1)

where H(t) is the Heaviside step function. The first three terms are the permanent
increase in phase, frequency, and spin-down, while the last term gives the transient in-
crease in the frequency ∆νt which decays exponentially with a timescale τ . To illustrate
this, in Figure 2.1 we show the spin-frequency model of a glitch including a permanent
increase in frequency ∆ν and a component ∆νt which is ‘recovered’.
∗
A single ‘anti-glitch’ in an isolated magnetar has also been reported by Archibald et al. [21] in
which the pulsation frequency spontaneously decreased. The implications of this remain unclear, but its
existence was further confirmed by Hu et al. [82].
†
Note that there is a typographical error in Eqn. (121) of Edwards et al. [57]: the final t − tg should
in fact be τ

15
16 Chapter 2 Timing variations

Figure 2.1: Illustration of the glitch model fitted by pulsar astronomers.

In effect, pulsar astronomers fit separate Taylor expansions either side of the glitch. This
is a good model when the rise-time of the glitch, during which the frequency increases,
is short compared to the duration between observations. This evolution of the frequency
during a glitch has yet to be observed, but high time resolution monitoring of the Vela
pulsar placed an upper limit of 40 s for the rise-time between the original and the new
period [55]. Since we cannot resolve the glitch itself, Eqn. (2.1) is appropriate and used
for all known glitches.

A comprehensive review of glitches was carried out by Espinoza et al. [59]; to illustrate
a typical glitch, in Figure 2.2 we reproduce data from this review on a glitch in the Crab
pulsar.

Over 165 of the ∼2000 observed pulsars have been seen to glitch, often multiple times.
Typical values of the instantaneous frequency change range from 10−9 Hz to 10−4 Hz For
some pulsars this is accompanied by a change, with either sign, in the spin-down rate
∆ν̇ with absolute magnitudes between 10−19 Hz/s to 10−12 Hz/s. The glitch recovery
parameter is defined as

∆νt
Q= . (2.2)
∆ν + ∆νt

Most glitches are not resolved with sufficient detail to determine this recovery parameter
accurately. A review of those pulsar glitches with measured values of Q was conducted
by Lyne et al. [114]; they found that in glitches from 18 pulsars, Q correlates with
|ν̇| reaching values as large as ∼ 0.9 for the youngest pulsar with the highest absolute
spin-down rate, the Crab pulsar.

Many pulsars have been observed to glitch several times, Melatos et al. [123] considered
the waiting times between glitches and concluded that in most glitching pulsars the
glitches happen randomly with waiting times consistent with a Poisson process, except in
PSR J0537-6910 and PSR B0833-45 (the Vela pulsar) which displayed quasi-periodicity
in the waiting times.
Chapter 2 Timing variations 17

Figure 2.2: A glitch in PSR B0531+21, the Crab pulsar. It occurred around
MJD‡ 53067 and had a permanent fractional frequency jump of ∆ν/ν = 5.33 ±
0.05×10−9 . (a) The timing residuals relative to a slowdown model with two
frequency derivatives when fitting data only up to the glitch date. (b) Timing
residuals after fitting all data in the plot; note that the glitch feature is still
visible. Both these panels have the same y-scale, covering 500 ms. (c) Frequency
residuals, obtained by subtracting the main slope given by an average ν̇. (d)
The behaviour of ν̇ through the glitch. This figure and caption are adapted
from Figure (1) of Espinoza et al. [59].

In Chapter 7, we perform our own investigation into the population statistics of glitches
with an aim to understand their implication for gravitational wave searches. We find,
in agreement with Espinoza et al. [59] and references therein, that the distribution of
glitch magnitudes has multiple modes which suggests that glitches may come from more
than one mechanism. We go on to apply a statistical model and determine empirically
the properties of the underlying source populations.

Glitches provide a unique opportunity to investigate the physics of neutron stars. The
two models able to explain some of the observations are known as the superfluid unpin-
ning model and the starquake model.
‡
The modified Julian date (MJD) calender counts days from an epoch of midnight on the 17th of
November 1858.
18 Chapter 2 Timing variations

In the superfluid unpinning model proposed by Anderson and Itoh [16], the star contains
a superfluid component in which the angular momentum is stored in an array of vortices
which are ‘pinned’ to the crust. The magnetic dipole, rigidly fixed to the crust, exerts
a torque on the crust gradually spinning it down. The superfluid component cannot
decrease its angular momentum without reducing the number of vortices per unit area,
so does not spin down at the same rate. A lag in frequency between the superfluid
component and rest of the crust develops until the forces are sufficiently large to cause
an avalanche of unpinning events rapidly transferring the stored angular momentum in
the superfluid component to the crust. An observer measures the frequency and spin-
down rate from the rate of pulsations. Since these pulsations originate from the EM
dipole which is frozen into the crust of the star, when this unpinning occurs, we see a
rapid increase in the frequency.

The second model, starquakes, follows from the observation that a rapidly spinning fluid
body has an oblate ‘rest shape’ with a bulge about its equator due to the centrifugal
force. The crust of a star spinning at some frequency will solidify as the star cools, with
a corresponding oblateness which we call the reference oblateness. Subsequently, as the
star spins down, it will have a different rest oblateness due to its decreased frequency, but
the crust will retain a memory of the earlier reference oblateness at which it solidified.
This will cause strains in the crust which eventually cause a starquake relieving the
strain, resetting the reference rest shape, and producing glitch like features. This model
was first proposed by Ruderman [148] and later built upon by Baym and Pines [28].

Both of these models have support in the literature and have been developed significantly
to explain the variety of observed glitches. However, there are observations which cause
difficulties for both models: glitches seen in the Vela pulsar are too large and too often
to be consistent with a starquake model [39], while the unpinning model requires a
superfluid component which is at odds with observation of precession (we discuss this
further in Chapter 5). In this thesis, we will not use glitches as a tool for inferring
neutron star physics, but any predictions we do make must be compatible with what
has already been learnt from glitches.

2.2 Timing noise: observations

Timing noise refers to small-scale structure in the timing residual which cannot be
attributed to any other source and hence cannot be modelled and included in the timing
model. The presence of timing noise indicates that we do not have a complete picture
of the neutron star: there is unmodelled physics.

Characterising timing noise is a difficult task: the exact form it takes will depend on
the order of Taylor expansion used to fit the timing parameters. Typically, pulsar
astronomers truncate at ν̈, but fitting to higher orders is possible and will tend to
decrease the ‘level’ of the resulting structure understood as timing noise. Of course
using a sufficiently large number of terms in the Taylor expansion, eventually one will
fit out all the structure. However, this does not provide any additional insight into the
Chapter 2 Timing variations 19

cause of timing noise. In this thesis, we will define timing noise as the remainder having
fitted and subtracted a second order Taylor expansion. In the literature, a second order
fit is most commonly used (see for example Hobbs et al. [81]), but examples exist of
fitting and removing higher order Taylor expansions (e.g. PSRs B0919-06, B1540-06,
and B1828-11 in Lyne et al. [111]).

To illustrate timing noise and how it can depend on the order of Taylor expansion used,
in Figure 2.3 we show the phase residual remaining after fitting and removing a 3rd ,
4th , and 5th order Taylor expansion to the Crab pulsar. In all three instances we see a
quasi-periodic structure reaching residuals up to half a cycle; higher order have lower
residuals, but the form of the structure remains consistent between orders.

0.6
3rd order
0.3

0.0

−0.3

−0.6
Phase residual (cycles)

0.6
4th order
0.3

0.0

−0.3

−0.6
0.6
5th order
0.3

0.0

−0.3

−0.6
45000 45500 46000 46500 47000
Modified Julian Date

Figure 2.3: A phase residual demonstrating the structure which is named timing
noise. This is generated from data on the Crab pulsar (see Section 8.2 for
details).

Several methods exist in the literature to quantify the strength of timing noise such
as the ∆8 value introduced by Arzoumanian et al. [23], the generalisation of the Allan
variance [119], the covariance function of the residuals [42], and fitting for timing noise as
part of the pulsar timing model [104]. The most comprehensive and recent analysis was
performed by Hobbs et al. [81] who considered 366 pulsars over timescales & 10 years.
We summarise their conclusions here:
20 Chapter 2 Timing variations

1. Timing noise is widespread in pulsars

2. Timing noise is inversely correlated to the characteristic age as defined in Eqn. (1.9)

3. The structures seen in the timing residual vary with data span: as more data is
collected, more quasi-periodic features are observed.

4. The dominant contribution to timing noise for young pulsars with τage < 105 years
can be explained as being caused by the recovery from previous glitches.

5. A handful of pulsars exhibit significant periodicity while quasi-periodicities are

observed in many pulsars

These general features give a broad picture, but there is great variation in the form
of timing noise between pulsars; this is illustrated by the variety of timing residuals
reported in Hobbs et al. [81]. To understand the variety of observation, in the next
section we will discuss some of the models for timing noise that exist in the literature
and which observations they are able to explain.

2.3 Timing noise: interpretations

The underlying mechanism which causes timing noise is not understood. Since the first
discussions in Boynton et al. [33] multiple models have been proposed which are able to
describe some of the features. However, the variety of ways timing noise manifests and
the uncertainty in the mechanisms at work have made it difficult for any conclusive state-
ments to be made about the models. A complete understanding of timing noise must not
only explain the observed variations, but also remain consistent with our understanding
of neutron stars derived from other observations such as glitches. A complicating fac-
tor in understanding timing noise is that the observed features have timescales similar
to the duration that we have been able to observe pulsars; it is therefore possible, as
noted by Hobbs et al. [81], that observations which look like a random walk over a short
timescale, may in fact be periodic or something else entirely over longer timescales. Ob-
servations of new features prompt new models for timing noise; as a result the timing
noise interpretations have evolved with the observations. In this section, we will present
an overview of these interpretations and the evidence which supports them.

2.3.1 Random walk models

Timing noise was first quantified and interpreted by Boynton et al. [33] in the context
of the Crab pulsar as a Poisson like random walk in one of the phase, frequency, or
spin-down. The pulsar spins down according to the power law spin-down of Eqn. (1.7)
except that at random times the pulse phase, frequency, or spin-down will undergo a
sudden step; these are referred to as phase noise (PN), frequency noise (FN) and spin-
down noise (SN). The waiting times between steps are Poisson distributed with a rate
λ. The number of events in such a model is considered to be large such that in a
Chapter 2 Timing variations 21

typical observation time Tobs we observe the superposition of many events. Following
the method of Cordes [43], the strength of the three types of timing noise can be defined
as

SPN = λhδϕ2 i, (2.3)

2
SFN = λhδν i, (2.4)
SSN = λhδ ν̇ 2 i, (2.5)

where hδϕ2 i, hδν 2 i, and hδ ν̇ 2 i are the mean-square values of the steps in phase, frequency,
and spin-down rate. These strength parameters form the basis of much of the analysis of
random walk models. They are determined by separating the observed TOA into blocks
of data of duration T and fitting a timing model to each block. In each block, one can
then calculate the phase residual ∆ϕ(t). Time averaging the squared phase residual for
each block of duration T and then averaging over all the blocks of data, we define
 Z T 
1
σϕ2 (T ) = 2
∆ϕ(t) dt , (2.6)
T 0

which is an estimate of the variance of the phase residual over a time span T . Cordes
[43] then showed that this is related the strengths of each noise component by
 S T

 2 ,
PN

2 SFN T 3
σϕ (T ) = 12 ,
(2.7)

 SSN T 5
120 .

Combining this with an estimate of σϕ2 from the data using Eqn. (2.6), these relations
give the strength of each noise component for blocks of duration T .

Given a sufficient amount of data, one can then calculate the noise strengths for different
values of T . If the strengths vary with the T , the random walk component can be
rejected since this is not predicted by the model. However, if the strength for a given
noise component is robust to changes in T , this provides evidence that the random walk
is in fact occurring. However, this consistency is a necessary, but not sufficient condition;
it can’t be ruled out that another mechanism mimics this behaviour. Furthermore, the
situation can be complicated by mixing of noise processes as shown in Groth [73] and
Cordes [43].

Using a similar process, Boynton et al. [33] categorised timing noise in the Crab pulsar as
frequency-like. They found that over a 5 year period the noise process was stationary and
consistent with the frequency noise hypothesis. No deterministic process could account
for the timing residuals strengthening their conviction that some random process was
taking place. The Crab was subsequently analysed by Groth [73] and Cordes [43]; the
analysis was extended to other pulsars by Helfand et al. [76] and Cordes and Helfand
[48].

The model of timing noise as a Poisson random walk is a purely empirical statistical one.
It is however backed up for a rich variety of possible substantive physical models. A
key feature of any physical random walk model is that it must be able to produce both
22 Chapter 2 Timing variations

increases and decreases in the relevant parameter. For this reason, it is unlikely that the
mechanism responsible for a random walk timing noise is the same as that proposed to
explain glitches. Since it is assumed that the number of events is large so that we can
take a statistical average, this does not test if the process is indeed discrete, or if it is
continuous.

The first substantive physical model to explain the random walk was proposed by Boyn-
ton et al. [33]. The noise process consisted of the accretion of small lumps of matter
onto the star from the interstellar medium. Lumps of matter fall randomly onto the
surface of the star causing either a spin-up or spin-down through the transfer of angu-
lar momentum. After this, many models were proposed such as small starquakes and
the random pinning and unpinning of vortex lines; these were reviewed by Cordes and
Greenstein [47] and evaluated against observational constraints. Of these, only three
mechanisms where found to be consistent with observations: crust breaking by vortex
pinning, a response to heat pulses, and luminosity related torque fluctuations. Since
this review, new random walk mechanisms have been proposed such as variations in the
magnetospheric gap size [40], the interference by debris entering the magnetosphere [49],
and the accumulation of multiple micro-glitches [84]. It would be a useful exercise to
compare both the old and new mechanisms against the current observational catalogue
since improvements in the quality and duration of data and a larger number of sources
may better constrain some of these physical models.

The first measurement of apparent individual timing noise events was made by Cordes
and Downs [46] who identified ∼ 20 step-like events in both frequency and spin-down
which could not be explained by a glitch. By considering 24 pulsars over a period
of ∼ 13 years, the authors concluded that the timing noise seen in the data could not
be explained solely by an idealised random walk processes in either the phase, or its
derivatives. They suggested instead that most of the activity is due to a mixture of
events in the phase, frequency and/or frequency derivative.

A recent observational review of timing noise was performed by Hobbs et al. [81] for
366 pulsars. The authors do not use the observations to constrain random walk models.
Instead, they state that timing noise cannot be explained by a simple random walk
model and when observed on sufficiently long timescales, the residuals which may before
have looked like a random walk, contained quasi-periodic features. It is impossible to
argue that this is not the case for pulsars which only display random-walk features
over current observation periods, since the quasi-periodicity may have periods longer
than the observation. This effect is apparent for the Crab pulsar which was found by
Boynton et al. [33] to be consistent with frequency-like noise, but in Figure 2.3 we see
quasiperiodic features. On the basis of Figure 2.3, we agree that it is plausible that
there are features of the Crab timing that require more than a random walk. However,
if one simply wants to characterise timing noise over short periods of data, a random
walk provides a good empirical description.
Chapter 2 Timing variations 23

2.3.2 Free precession

A mechanism which could quite naturally produce strictly periodic variations in the
observable features of a pulsar is free precession. This occurs in any non-spherical rigid
body for which the angular momentum is not aligned with a principal axis of the moment
of inertia. Such a circumstance could arise given the chaotic birth of neutron stars.
However, we must be clear that the timing noise induced by precession alone would be
strictly deterministic; this is something which we do not typically observe except in a
handful of pulsars, one of which is considered in Chapter 5. We will now consider the
mechanics of free precession.

In the simplest case, we take a biaxial body, rotating about an axis Ω, with a moment
of inertia given by
   
I0 0 0 I0 0 0
   
I =  0 I0 0  =  0 I0 0 , (2.8)
0 0 I0 + ∆I 0 0 I0 (1 + I )

where I0 is the total moment of inertia of the star involved in precession and I =
∆I/I0 is the measure of oblateness or prolateness which participates in precession. This
asymmetry ∆I, refers only to the portion of the deformation that follows the precession
of the star, i.e. that part source by strain, not the part sourced by the centrifugal
deformation Jones and Andersson [90]. If the body is free from torques, then in the
rotating frame of the body, Euler’s equations of motion [100] are given by

I Ω̇ + Ω × (IΩ) = 0. (2.9)

This is a system of three coupled ordinary differential equations (ODEs). Writing the
components of the spin-vector as Ω = [Ωx , Ωy , Ωz ], we have

Ω̇x = −I Ωy Ωz , Ω̇y = I Ωx Ωz , Ω̇z = 0 (2.10)

We can find a solution by first setting Ωz = const. We are then left with a set of two
coupled ODEs, solving these with appropriate initial conditions the solutions take the
form

Ωx = Ω0 sin(a0 ) sin (Ω0 cos(a0 )I t) , (2.11)

Ωy = Ω0 sin(a0 ) cos (Ω0 cos(a0 )I t) , (2.12)
Ωz = Ω0 cos(a0 ), (2.13)

where a0 is the angle between the spin-vector and the body frame z axis, and Ω0 is the
magnitude of the spin-vector.

We observe that the spin axis of the body will trace out a cone about the z principal
axis of the moment of inertia with a period of

2π
τP = . (2.14)
Ωz I
24 Chapter 2 Timing variations

The half-angle of the cone is set by the initial conditions and will not evolve. This is
the motion of free precession and is illustrated in Figure 2.4.

Figure 2.4: Illustration of free precession for a simple biaxial body. The spin
axis Ω traces out a cone about the angular momentum vector J.

It was shown by Ruderman [148] that neutron stars should have a solid crystalline crust
which is capable of supporting shear stresses such that there may be differences in the
components of associated moment of inertia tensor. Precession, as a candidate to explain
timing noise fluctuations, was first discussed by Ruderman [149]. He found that the free
precession period was, for an ellipticity of I ∼ 10−9 , able to explain periodic fluctuations
in the Crab pulsar. Precession is one of the few periodic mechanism that could act on
neutron stars with the periods observed in typical timing residuals.

However, the superfluid unpinning interpretation for glitches poses a problem for sus-
tained free precession as an interpretation of timing noise. If correct, then the interior
of a neutron star must contain a superfluid component pinned to the crust. It was
shown by Shaham [152] that for perfect pinning the deformation relevant to precession
is modified to give

∆I Ipsf
I = + , (2.15)
I0 I0

where Ipsf is the moment of inertia of the superfluid pinned to the crust. To explain the
magnitude of observed large glitches this second term is typically found to be ∼ 0.01
(see for example Table (1) of Andersson et al. [18]). Inserting this into Eqn. (2.14),
we see that the free precession period must be approximately 100 times the spin-period.
Therefore, the existence of a perfectly pinned superfluid precludes long-period precession
where the precession period is millions of times longer than the spin-period.

Despite the inconsistency with the superfluid pinning model for glitches, evidence was
presented by Stairs et al. [156] of free precession in PSR B1828-11. They found the
phase residuals and variations in the pulse profile could be accounted for by precession.
This pulsar has been studied by others since and in Chapter 5 we will discuss this pulsar
and the precession interpretation in more detail.
Chapter 2 Timing variations 25

Figure 2.5: Figure taken from Kramer et al. [97] showing the switched spin-down
of pulsar PSR B1931+24

One resolution to this comes from relaxing the assumption that the superfluid pinning
is perfect. Sedrakian et al. [150] found that imperfect pinning of the superfluid allows
long-lived precession with damping. It is therefore possible, as noted by Cordes [44]
that free precession may be excited by torque fluctuations that counter the damping
process; in turn, the precession can drive torque fluctuations. However, there remains
some uncertainty as to the details of this model, so whether precession can coexist with
a superfluid core remains an open question. Resolving this, and determining exactly
how the models are incompatible is an important task which we hope to make some
headway on in this thesis by studying the case for precession in Chapter 5.

2.3.3 Magnetospheric switching

Recently a new model has been proposed by Lyne et al. [111] to explain the observation of
quasi-periodicity in timing noise structure. This work was motivated by the observations
of Kramer et al. [97] that the pulses from PSR B1931+24 were intermittent: the pulsar
acts as a normal pulsar for ∼ 10 days and then switches off, being undetectable for
∼ 25 days, before switching on again. This behaviour can be understood as the pulsar
switching between two states. Analysing the spin-down rate between the on and off
states, they determined the spin-down rate ν̇ was ∼ 50% faster in the on state. The
figure illustrating this is reproduced in Figure 2.5. The upper panel, Figure 2.5A, shows
the evolution of the rotational frequency over a 160 day period encompassing several
switching events. A line shows the long-term averaged spin-down of the pulsar while the
dots show individual measurements made during the on state. During these on states,
the gradient of the reduction in frequency is increased: the spin-down rate is faster. It is
thought that measurements of the frequency in the off state would produce a line with
26 Chapter 2 Timing variations

decreased spin-down connecting the dots. Plotted in Figure 2.5B is the timing residual;
this shows significant quasi-periodic modulations correlated with the switching.

PSR B1931+24 evidently switches suddenly and periodically between two distinct states.
Kramer et al. [97] realised that the neutron star magnetosphere is the only mechanism
which could produce such sharp changes with correlated changes in the spin-down rate
and beam. It makes sense that magnetospheric state which does not produce EM emis-
sion may have a lower spin-down rate while the state which does produce radiation has
a higher spin-down rate.

Motivated by this observation, Lyne et al. [111] tested a set of other pulsars and pre-
sented a study of 17 pulsars for which they found evidence for ‘two-state magnetospheric
switching’. Unlike in PSR B1931+24, these pulsars are not intermittent, but are seen
to continuously pulse. The authors measured changes in the spin-down rate ν̇ of each
pulsar by using what we will call in this thesis the observer-method. This consists of
fitting a Taylor expansion to short segments of data of duration T . In each segment,
the fitted coefficient ν̇ is taken as a measurement of the spin-down rate at the mid-point
of the segment. Repeating this process in a sliding window at intervals T /4 throughout
the whole data set gives the evolution of the spin-down rate over the entire observation.
Over a ∼ 20 year period they found a variety of smooth periodic fluctuations with typi-
cal periods of years. In Figure 2.6 we reproduce their original plot showing the periodic
variations in spin-down rates.

The method used to calculate the spin-down in Figure 2.6 required time averaging;
the baseline used T is given below the pulsar name on the right-hand side of the plot,
typically T ∼ 100 days. Lyne et al. [111] argue that, if the pulsar is in fact switching
between magnetospheric states, then its spin-down rate will be periodically switching
between at least two values. In the event that the switching occurs over timescale
comparable to the time averaging baseline, the resulting time averaged spin-down rate
will be smoothed out. Therefore, the smooth periodic variations in Figure 2.6 could be
produced by the spin-down rate switching between two (or more) well defined values.

If the switching was magnetospheric in origin, then Lyne et al. [111] realised that changes
in the spin-down rate should correlate with changes in the beam-shape. They found
that for 6 pulsars the pulse width did indeed correlate with changes in the spin-down
rate, although for some pulsars this was an anti-correlation rather than a correlation.
However, like the spin-down rate these pulse shape variations are also smooth and subject
to the same time averaging process.

To test if the mechanism causing these variations is smooth or instantaneous (i.e. switch-
ing) the authors needed to look at an observed property not subject to the time-averaging
process. This is possible by looking at individual measurements of the pulse width which
are taken during each short ∼ 30 minute observation. In this way, they are time averaged
over a duration much shorter than the modulation period and could, in principal, resolve
individual switching events. They were able to do this for two pulsars, PSRs B1822-09
and B1828-11, and argue that the individual beam-width measurements, by eye, can be
interpreted as taking one of two values with few observations seen in between. They
took this as evidence that the pulsars were indeed undergoing sudden, instantaneous,
Chapter 2 Timing variations 27

Figure 2.6: Figure adapted from Lyne et al. [111] showing the spin-down rate
of 17 pulsars over a ∼ 20 year period.

and periodic switching between two states. We note that the stronger of these two
candidates, PSR B1828-11, was previously discussed in the context of free precession
in Section 2.3.2. Lyne et al. [111] argue that the fast state changes seem to rule out
free precession as the origin of oscillatory behaviour observed in timing residuals, in
particular for PSR B1828-11.

Since this initial study, evidence for switching in two others pulsars have been reported
by Perera et al. [134] and Perera et al. [135] with additions added to the model in terms
of the detail of the switching. However, so far the magnetospheric switching hypothesis
is missing an underlying physical model which both causes the switching between states
and provides the clock regulating the period.

Jones [89] argues that the dismissal of precession is premature since the modulation
period of the switching has yet to be explained. Instead, the author raises the idea
that precession and magnetospheric switching are not mutually exclusive, but preces-
sion may in fact cause the switching. Pulsars are most probably born in a randomly
distributed magnetospheric state. At least some may therefore exist under a delicate
balance between two states. Precession may be capable of periodically varying the sta-
tistical probability of existing in one state or the other. Sharp changes would be caused
by an ‘avalanche effect’ as the particle energies reach a threshold. This provides the
28 Chapter 2 Timing variations

timescale for switching along with the ability for the switching to be quasi-periodic
since the precession only biases the probability.

Another idea proposed by Cordes [45] interpreted two state switching as evidence for
a system in a state of ‘stochastic resonance’. This occurs in systems in which, under
certain conditions, a weak periodic forcing function is amplified by stochastic noise.
To explain this phenomenon in Appendix 2.A we present a toy model of stochastic
resonance for a particle in a well. The switching could therefore be the result of any
periodic modulation, such as precession, coupled to random fluctuations. This would
quite naturally explain the stability of states, the timescales over which they occur, and
the fact that it is observed in only some pulsars.

As already mentioned, determining the cause of periodic modulations has important

implications for neutron star physics. Later on in Chapter 5 we will investigate this issue,
particularly in the context of PSR B1828-11, using a quantitative model comparison.
In the remainder of this section we will build a model for magnetospheric switching
incorporating all the ideas of Lyne et al. [111] and Perera et al. [134]; this model forms the
basis of the model compared against precession in Chapter 5. Finally, in Section 2.3.3.3
we show how this simple empirical model of switching makes a testable prediction which
could be checked by pulsar astronomers using current data.

2.3.3.1 Simple empirical model

In the supplementary material to Lyne et al. [111], the authors presented results from a
simple empirical switching model to demonstrate the resulting phase residuals. In this
section, we will similarly develop a simple model and show the effect of switching on the
time-averaged spin-down rate and phase residuals.

We model a pulsar as spinning down in the usual way, except that its magnetosphere
periodically and suddenly switches between two different states which we label A and B.
The star spends a time tA and tB in each state such that tA + tB gives the total period
of switching. We then associate spin-down rates, ν̇A and ν̇B to each of the states such
that the underlying spin-down rate is a square wave oscillating between these values.
Since we do not concern ourselves in this description with the exact periods, only the
gross features, let us define the ratio of time spent state in state A and B as R = tB /tA .

This is a purely deterministic model and, having generated the spin-down values, we can
integrate twice to get the phase. From the phase, we can use the observer-method to
calculate the time-averaged spin-down rate hν̇i, or we can fit a polynomial to the entire
phase evolution and remove it to get a phase residual. This integration to the phase and
then applying the observer-method models the data collection mechanism so that our
resulting spin-down rate predictions are smooth.

In Figure 2.7 we show the underlying spin-down rate ν̇, the time-averaged hν̇i, and phase
residual (in cycles) for some typical values. For the time-averaged spin-down rate we
carefully chose an averaging time similar to the switching period. For the final phase
Chapter 2 Timing variations 29

residual we fit and remove a Taylor expansion up to ν̈ and divide the residual difference
by the period in order to get the phase residual in cycles.

R = 2.0, D = 0
ν̇2

ν̇
ν̇1

ν̇2
hν̇i

ν̇1
Phase residual

0.6
0.3
[cycles]

0.0
−0.3
−0.6
0 tA tA + tB 2(tA + tB ) 3(tA + tB ) 4(tA + tB )
time

Figure 2.7: A deterministic realisation of the Lyne switched spin-down model.

The resulting structure in the time-averaged spin-down rate and the phase resid-
uals are strictly periodic. This shows the underlying spin-down rate in the top
panel, the time-averaged spin-down rate (calculated using the observer-method)
in the second panel, and the phase residual in the bottom panel.

The phase residuals in Figure 2.7 are strictly periodic which is inconsistent with the
observed variations in the majority of pulsars [81]. To develop this, Lyne et al. [111]
added a probabilistic ’dither’ D in the waiting time between switches. Now we have
periods tiA and tiB which are Gaussian distributed with a mean of tA and tB and a
standard deviation DtA and DtB . In Figure 2.8 we repeat the results of Figure 2.7
with D = 0.3. This dither can be used to understand why the residuals and averaged
spin-down rate are quasi-periodic rather than strictly periodic. However, this model is
missing a substantive physical explanation, namely what causes the switching and why
it is quasiperiodic.

2.3.3.2 Simple empirical model: four time periods

The time averaged spin-down rates in Figure 2.7 and Figure 2.8 result from having two
spin-down rates and two durations which the system spends in those states. These results
can be contrasted with the spin-down rates of PSR B1828-11 seen in Figure 2.6. With
only the ingredients described by Lyne et al. [111], it is not possible to consistently
replicate the double peaked spin-down rate variations seen for this pulsar. A similar
problem was faced by Perera et al. [134] for PSR B0919+06, in response the authors
devised the following solution. Rather than introduce a third spin-down state, they
simply required that the system has four periods instead of two as in the original Lyne
et al. [111] description; that is, we have four times tA , tB , tC , and tD which sum up to
30 Chapter 2 Timing variations

R = 2.0, D = 0.3
ν̇2

ν̇
ν̇1

ν̇2
hν̇i

ν̇1
Phase residual

1.0
0.5
[cycles]

0.0
−0.5
−1.0
−1.5
0 tA tA + tB 2(tA + tB ) 3(tA + tB ) 4(tA + tB )
time

Figure 2.8: A realisation of the Lyne model with a random element producing
the observed quasi-period structure. For a description of the three panels see
Figure 2.7.

the total period. The system then switches between the two states four times during a
single cycle. To illustrate this, in Figure 2.9 we show a single cycle in which the four
periods are 100, 100, 50, and 100 days.

tA tB tC tD
ν̇2

ν̇
ν̇1 hν̇i

0 50 100 150 200 250 300 350

time [days]

Figure 2.9: The spin-down rate ν̇ and its time-average hν̇i for a single cycle of
the Perera et al. [134] switching model in which the system switches between the
two states four times during a single cycle. For the time-averaged spin-down
rate, we have used a baseline of 100 days, which is longer than the shortest
period of tC =50 days; for this period we highlight the true spin-down rate by
a red dot and the maximum time-average spin-down rate by a blue dot.

The Perera et al. [134] switching model naturally produces the doubly peaked time-
averaged spin-down rate seen in PSR B1828-11 when the time-averaging baseline is
longer than one of the switching periods. To explain more fully, when the time averaging
Chapter 2 Timing variations 31

baseline is longer than a single period, then the time-average over that section of the
data will always include some amount of the other spin-down rate. This can be seen
in Figure 2.9 for the third period which has duration 50 days, while the time averaging
baseline is 100 days. In Figure 2.10, we give the spin-down rate and phase residual for
this model to illustrate the variations in the residuals.

ν̇2
ν̇
ν̇1

ν̇2
hν̇i

ν̇1
Phase residual

0.08
[cycles]

0.00
−0.08
−0.16
0 ttot 2ttot 3ttot
time

Figure 2.10: Illustration of the Perera et al. [134] extension to the Lyne et al.
[111] model in which the system switches between the two states four times
per cycle, giving four independent switching periods. In this figure, ttot =
tA + tB + tC + tD . For a description of the three panels see Figure 2.7.

In Section 5.3.2 we will consider this switching model further in the context of PSR B1828-
11 and develop a complete model for the spin-down rate and beam-width modulations.

2.3.3.3 A simple test of the switching hypothesis

PSR B1828-11, which has a distinct doubly peaked spin-down, is one of two pulsars for
which the authors of Lyne et al. [111] applied their magnetospheric switching hypothesis
too. Since that result, several other pulsars have been found with a similar double-
peaked structure [134, 135]. In the previous section we introduced a basic empirical
model to explain this double-peak in the context of magnetospheric switching. We will
now describe a simple method to test this switching hypothesis which so far, to our
knowledge, has not been attempted, but could be applied to current observational data.

If the double peaked spin-down rates of PSRs B1828-11, B0919+06, or any other pulsar
is due to the model proposed by Perera et al. [134], then the maximum spin-down rate of
the lower of the two peaks is a function of the time-averaging process and not the pulsar
itself. Specifically, if the time-averaging was shorter than the shortest period there would
be no second lower peak: both peaks would have the same value, ν̇. Therefore, if one
had access to the raw data used to produce the time-averaged spin-down rate, one could
32 Chapter 2 Timing variations

repeat the time-averaging process varying the time-averaging baseline which we denote
by T . Then we could imagine measuring the maximum spin-down rate of the two peaks
in the time-averaged spin-down rate and taking their ratio

ν̇max of lower peak

R= . (2.16)
ν̇max of larger peak

Then, if the Perera switching model is to be believed, R should depend on T .

To demonstrate this, we can simulate the result numerically. We will do this by varying
the time averaging baseline for the simulation in Figure 2.9 and calculating R. The
numerical results are plotted in Figure 2.11. This clearly shows that for T ≤ 50 days,

1.1

1.0

0.9

R 0.8

0.7

0.6

0.5
0 20 40 60 80 100 120 140 160
T the time averaging baseline [days]

Figure 2.11: Demonstration of the variations in R, defined in Eqn. (2.16), in

response to changing the time averaging baseline, T .

when the time averaging baseline is shorter than the shortest switching duration tC (see
Figure 2.10), R = 1 because the two peaks are of equal size. For T > 50 days, R
decreases until T = 150 days, above this the time-averaging is of a similar length to the
total period of switching and so we can’t resolve any of the features.

We have shown that variations in the ratio of the two peaks in spin-down rate would
provide clear evidence in support of the Perera et al. [134] model. Of course in this
example we have considered a simple situation where we have known a priori which peak
is being underestimated due to the observer-method of calculating the time averaged
spin-down rate. In real data, the situation may be more complicated: it could be that
it is the trough which is overestimated instead of the peak being underestimated; or
more than one switching duration may be shorter than the time averaging baseline.
Nevertheless, the changing ratio of the two peaks (or troughs) may still prove a simple
way to test the Perera model.
Chapter 2 Timing variations 33

Appendix 2.A Toy model of stochastic resonance: particle

in a potential

Here we present a simple toy model of stochastic resonance. This is a statistical phe-
nomena occurring when a weak periodic forcing function is amplified by noise (see Jung
[94] for a full treatment). The application to neutron stars and the problem of timing
noise was first discussed by Cordes [45]. Here, we simply aim to describe the essential
features of stochastic resonance (not its application to neutron stars).

We will consider a 1-dimensional space with a particle at a position x. The particle is

subject to some potential and acted upon by a forcing function F (t). In general, x could
be any state variable of the system.

First consider the static case of a particle in a potential U (x) given by:

x4 x2
U (x) = − . (2.17)
4 2
This potential is characterised by two wells at x = ±1, a maximum exists between them
at the origin. The particle in one of the wells sees a potential barrier ∆U corresponding
to the height of the maximum above its position.

Assume the particle is acted upon by a random forcing function F (t) which is modelled
as a Gaussian white noise with standard-deviation σ. Depending on the magnitude of
σ with respect to the potential, the motion of the particle admits two distinct cases:

1. σ  ∆U in which case the particle remains inside whichever well it initially

starts in and does not escape.

2. σ  ∆U in this case the particle will not see the individual wells only the larger
one.

The motion of the particle obeys the following equation of motion:

dx ∂V (x, t)
=− + F (t). (2.18)
dt ∂x
The motion of the particle has two components, the deterministic effect of the potential
and random fluctuations.

We now modify the potential to be acted on by a weak periodic function; this introduced
a third possible type of behaviour. Writing the time dependent potential as

x4 x2
V (x, t) = − + εx cos(ω0 t). (2.19)
4 2
Inserting this potential into the equations of motion:

dx
= x − x3 + F (t) + ε cos(ω0 t). (2.20)
dt
34 Chapter 2 Timing variations

Solving this numerically we fix ε = 0.001, ω0 = 2π 10 and choose three values of σ which
illustrate typical behaviours of the solution The first and last runs replicate the behaviour

Figure 2.12: Three solutions to Eqn. (2.20) changing the random forcing func-
tions strength σ. The left and right panels show the non-resonant solutions
for the particle position: either the forcing function is weak compared to the
potential and so the particle remains in well in which it begins; or the forcing
function is much stronger than the potential and so the particle freely moves
about the two wells. The middle panel illustrates the special case of stochastic
resonance whereby the periodic fluctuations of the potential allow quasi-periodic
variations in the particle’s position between the two wells.

expected for a static well, either the particle is confined to the well it starts in, or the
random noise is too strong and the individual wells are not observed. The middle case
displays strong stochastic resonance: the solution displays a switching between bi-stable
states but does not strictly follow the period of the forcing function. The important
point here is that the forcing function may be weak, but provided it is periodic or at
least quasi-periodic the signal is amplified by the random noise such that it may be
visible in data sets where it would typically be considered lost.
Chapter 3

Action of the electromagnetic

torque on a precessing neutron
star

The rotation of an isolated neutron star is gradually slowed by torques. For most stars,
the dominant torque is thought to be the electromagnetic torque due to an inclined mag-
netic dipole [132, 74]. In Section 2.3.2, we introduced free precession, the most general
motion for a non-axisymmetric body. In later chapters we will be considering precession
as a potential candidate to explain long term periodicities seen in pulsar timing residu-
als. Therefore, it is appropriate to first understand how precession manifests under the
action of an electromagnetic torque and discuss the alignment of various axes of the star
during its lifetime (as first investigated for a biaxial star by Goldreich [69]).

The full form of the electromagnetic torque was first written down by Deutsch [54]
where it was found there were two distinct components: the spin-down component and
the anomalous component (defined in Section 3.1). The subject of precession under the
electromagnetic torque has already been studied, but many authors neglect the anoma-
lous component in order to simplify the analytic calculations. In this chapter, we will
build a numerical model solving the Euler rigid-body equations with the Deutsch torque.
This will allow us to compare solutions in which the anomalous torque is ‘switched off’
with solutions to the full torque; in doing so we hope to better understand the im-
portance of the anomalous torque and verify that in cases where it has been neglected
(see for example [69]) that such an assumption was appropriate. A similar numerical
model was studied by Melatos [122], although the authors did not switch on and off the
anomalous component. Furthermore, the authors found solutions which they labelled as
‘persistent precession’: in Section 3.4 we will confirm that such solutions exist, but give
a physical explanation from which we conclude that precession is not persistent.

The rest of this chapter is organised as follows: in Section 3.1 we will define the Euler
rigid-body equations and the Deutsch [54] torque. In Section 3.2, we will discuss the
simple case of a spherical star, then in Section 3.3 and Section 3.4 we will discuss a
biaxial star with and without the anomalous part of the torque; in these sections we

35
36 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

will give an intuitive geometric way to understand the alignment predictions made by
Goldreich [69]. Finally, we will conclude the chapter in Section 3.5.

3.1 Defining the model

Euler’s rigid-body equations describe the motion of a rigid-body in a rotating reference

frame. To derive these, we begin with Euler’s second law for the angular momentum J
in the inertial frame  
dJ
= T, (3.1)
dt in
where T is the applied torque. The angular momentum is the product of the moment of
inertia tensor and the spin-vector, J = IΩ; in the inertial frame both of these quantities
vary in time. We can simplify by transforming to a frame fixed in the rotating frame
of the rigid-body and aligning the axes with the principal axes of I. In such a frame,
the moment of inertia tensor is constant and diagonal. The transformation to a rotating
reference frame requires a modification of the time derivative in Euler’s second law (see
for example §36 of Landau and Lifshitz [100] for a detailed discussion). The result is
Euler’s rigid-body equations in the rotating frame

dJ
+ Ω × J = T. (3.2)
dt
We will consider a biaxial body with the moment of inertia tensor defined in Eqn. (2.8)
which has principal moments given by

Ixx = Iyy = I0 Izz = I0 (1 + I ). (3.3)

For neutron stars, their extreme gravity ensures that I  1. For a uniform density
incompressible star, we take a typical moment of inertia to be

2
I0 = M R2 ≈ 1045 g cm2 M1.4 R62 , (3.4)
5

where M1.4 = 1.4M and R6 = 106 cm are typical values.

In Cartesian coordinates, Eqn. (3.2) gives us three coupled ODEs for Ωx , Ωy and Ωz . In
the torque-free case, analytic solutions can be found as we demonstrated in Section 2.3.2.

Neutron stars are acted on by torques from both their electromagnetic (EM) dipole and,
if I 6= 0, they may also be acted on by torques from the emission of gravitational waves.
For now, we will consider only the former effect which probably dominates for isolated
neutron stars observed as pulsars. To model this EM torque, we imagine a dipole with
magnitude m pointing along a unit vector m̂ frozen into the star. In the rotating frame,
the body is axially symmetric about the ẑ axis; therefore we are free to fix the dipole
in the x̂ − ẑ plane at an angle χ to the ẑ axis without loss of generality. Often in the
literature the effect of such an EM torque is modelled by a vacuum dipole torque when
modelling neutron stars, but Deutsch [54] demonstrated that in a real finite volume (as
opposed to a point-like) star, the dipole torque has two components: the usual spin-down
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 37

component and a second so-called anomalous component. We present the full torque
here in the form found in Goldreich [69],

2R
T = I0 A ω 2 (Ω × m̂) × m̂ + A I0 (Ω · m̂)(Ω × m̂), (3.5)
3c
where ω = |Ω| and A is an effective magnetic ‘deformation’ as defined by Glampedakis
and Jones [66]; it is of the order of the ratio of the magnetostatic energy to the rest-mass
energy of the star and is given by
 2
m2 B0
A = ≈ 10−7 , (3.6)
I0 Rc2 1015 G

where in the second step we have expressed it in terms of the surface magnetic dipole
field strength. Alternatively, it may be useful to solve for the surface magnetic field
strength
 1/2
4A I0 c2
B0 = . (3.7)
R5

The first term in the Deutsch torque is often referred to as the spin-down, or braking
torque. As the name suggests, it is responsible for the power law retardation of spin fre-
quency and has an associated timescale τS . The second term is known as the anomalous
torque which acts on a timescale τA . We can then define three timescales (including the
free precession timescale as derived in Eqn. (2.14)) which relate to this system, these
are initially given by

2π
τP = , (3.8)
I ω0 cos(a0 )
2π
τA = , (3.9)
A ω0
3c 1
τS = . (3.10)
2R A ω02

The spin-down age defined here, τS , is equivalent to the original definition of τage in
Eqn. (1.10) up to a factor of 2/ sin2 (α) where α is the angle between the spin-vector and
the magnetic dipole. Since we are considering cases where this angle evolves, we will
parameterise by τS in this chapter.

The ordering of these timescales characterises the type of solutions found to Eqn. (3.2)
with the torque of Eqn. (3.5). There are six potential orderings, however by considering
the ratio τA /τS ∼ v/c, where v is the equatorial velocity, we see that τA  τS reducing
the number of possible orderings to three.

We now have the basic components of a neutron star: a biaxial rigid-body spun down
by a dipole torque. To help familiarise the reader with the notation, in Table 3.1 we list
the spherical coordinates (the magnitude, polar angle, and azimuth) of the spin-vector
and magnetic dipole and also show their projection onto the x̂ − ẑ plane in Figure 3.1.
We work in a system of coordinates where the polar angle is that made with the ẑ axis
38 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

and the azimuthal angle is the angle between the projection of the vector onto the x̂ − ŷ
plane and the x̂ axis.

Magnitude Polar angle Azimuthal angle

Spin-vector Ω ω a ϕ
Magnetic dipole m m χ 0

Table 3.1: Spherical coordinates of the spin-vector and magnetic dipole.

ẑ
Ω

m̂
a χ

x̂

Figure 3.1: A sketch of the x̂ − ẑ plane, the magnetic dipole lies solely in this
plane such that its azimuthal component is always zero. Only the projection
into this plane of the spin-vector is shown. In general it has an azimuthal
component given by ϕ.

In the following sections we will develop our intuition for the different types of solutions
to this model. These results are calculated by exact numerical evolution of Eqn. (3.2)
with the torque defined in Eqn. (3.5). The simulations model a neutron star with a
magnetic field of B0 ≈ 1013 G and a radius R = 106 cm. We must also provide an
initial condition for the angular frequency ω. This is not the spin frequency at the birth
of the neutron star, but the frequency at the start of the observation; typical values
are of the order ∼ 10 rad/s. However, at this frequency the numerical solutions are
computationally expensive. To reduce the time required to produce a solution we will
work with an initial frequency of ω = 104 rad/s. This does not change the form of
solution, only the timescales over which they evolve.

The rest of this chapter is organised as follows: in Section 3.2 we will begin by discussing
a spherical star. Since we particularly aim to understand the role of the anomalous
torque, in Chapter 3.3 we will investigate a biaxial star acted on by only the spin-down
component of the torque. In Chapter 3.4 we then repeat the investigation of a biaxial
star, but this time include the anomalous torque and compare the results. In doing this
we hope to better understand the role of the anomalous torque on precession.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 39

3.2 Spherical star

We begin by considering the simplest case: a spherical star with I = 0 under the in-
fluence of the spin-down torque only. This model was first considered by Davis and
Goldstein [52] and separately by Michel and Goldwire [125]; both sets of authors discov-
ered that torque causes the spin axis to align with the magnetic axis on the spin-down
timescale.

To observe this behaviour, we set χ = 30◦ with the spin-vector initially at a = 50◦ and
A = 5 × 10−11 . Evolving the model numerically, we plot the spherical coordinates of the
spin-vector in Figure 3.2. This shows a brief spin-down which halts once the spin-vector
aligns with the magnetic dipole a → χ = 30◦ . The azimuthal angle ϕ does not vary;
this is expected since precession, a monotonic increase in ϕ, does not occur for spherical
bodies. This result agrees with the analytic calculations: the alignment occurs on a few
times the spin-down time scale τS = 9 × 106 s.

Figure 3.2: Plot of the spherical coordinates of the spin-vector for a spherical
star. Note that the magnetic dipole is inclined at χ = 30◦ to the ẑ axis.

3.3 Biaxial neutron star with no anomalous torque

We will now consider a biaxial star with I 6= 0. But to first build some understanding
we will remove the anomalous part of the torque. This is done so that later on when
we consider solutions in the presence of the anomalous torque, we understand which
effects are due to the anomalous part and which to the spin-down part. Neglecting the
anomalous torque reduces the number of timescales to just two; the spin-down and the
precession timescale. The solutions are categorised by the ordering of these timescales;
40 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

it is shown in appendix 3.A that these orderings remain valid for the duration of the
spin-down. The ordering of the two timescales gives three different types of solutions
corresponding to three cases:

Regions A: τS > τP Region B: τS ∼ τP Region C: τS < τP (3.11)

Since all solutions in each region display the same characteristic behaviour, we will
illustrate this by picking three ‘neutron stars’ and presenting results for each of these.
Each of these pulsars, defined by the parameters listed in Table 3.2 will help us to
understand the types of solutions for each ordering of the timescales.

Star I A B0 [Gauss] τS [s] τP [s]

A 1.0×10−9 5.0×10−11 1.3×1013 9.0×106 9.8×105
B 4.0×10−11 5.0×10−11 1.3×1013 9.0×106 2.4×107
C 1.0×10−15 5.0×10−11 1.3×1013 9.0×106 9.8×1011

Table 3.2: Table of I and A and the inferred magnetic field and initial
timescales for the three simulated stars. In calculating the magnetic fields and
timescales, we have used a canonical value of R = 106 cm for the radius, the
initial value of the spin frequency, ω0 = 104 rads/s and the initial value of
a0 = 50◦ . As ω and a evolve, the timescales of each star will also evolve from
these initial values.

The phase space of solutions is plotted in Figure 3.3, parameterised by the elastic de-
formation I and the surface magnetic field

2m
B0 = . (3.12)
R3
This illustrates how the phase space is sliced into separate regions by the ordering of
the timescales.

Without the anomalous torque, the ordering of the timescales defines two distinct regions
separated by τS = τP . In using an abnormally large value for the initial angular spin
frequency (ω0 = 104 [rads/s]), we have skewed this phase space - that is, it cannot be
compared with the real pulsar population. Using a more realistic value, the character of
solutions remains the same - only the timescales vary.

The elastic deformation I may take either a positive or a negative value; these corre-
spond to either an oblate or prolate mass distribution. In the following work, we use only
positive values of I . It will be noted when this is important, but otherwise solutions
are invariant to the sign of I . Finally, we must choose the angle between the magnetic
dipole and the elastic deformations χ. We choose two values (30◦ and 75◦ ) to demon-
strate important cases discussed in the next section. Initial conditions are thought to
have some bearing on the solutions [see 122]. In the current work, we will start with the
spin-vector having an inclination angle of a0 = 50◦ and azimuth of ϕ = 0◦ .
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 41

Figure 3.3: Phase space of solutions categorised by the ordering of the two
timescales as a function of the elastic deformation I and the surface magnetic
fields B0 .

3.3.1 Neutron star A in region τS  τP

Neutron stars in this region are described by the work of Goldreich [69]. In addressing
the shortcomings of the analytic spherical solutions [52, 125], Goldreich considered a
neutron star with a solid crust capable of supporting elastic strains. In order to find
an analytic solution, Goldreich assumed that the precession timescale was significantly
shorter than the spin-down timescale τS  τP . Averaging over equations (3.2) with this
assumption yields
D dω E   
2R 3
= − a ω 3 sin2 χ + sin2 a 1 − sin2 χ , (3.13)
dt 3c 2
D da E  
2R 2 3 2
= − a ω sin a cos a 1 − sin χ . (3.14)
dt 3c 2

Goldreich realised that the second equation tells us whether the polar angle a will either
grow or decay on the spin-down timescale depending on whether χ was greater or less
than a critical value χcr ≈ 55◦ .

We first present results for all the spherical coordinates during the spin-down in Fig-
ure 3.4. For the angle χ, we use two values, 30◦ and 75◦ , and find that a tends to either
0 or π/2 on the spin-down timescale, in agreement with Goldreich [69]. During this
alignment, the spin magnitude decays and ϕ monotonically increases corresponding to
the spin-vector precessing about the ẑ axis.

In Figure 3.5, we investigate the differences between solutions to Goldreich’s averaged

equations (3.14) and the exact solution without averaging. This shows that both agree
on the long term behaviour occurring on the spin-down timescale. However, in the exact
solutions we see additional oscillations at the faster precession timescale; this precession
42 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

(a) χ = 30◦ < χcr (b) χ = 75◦ > χcr

Figure 3.4: Plot of the spherical coordinates of the spin-vector for neutron
star A without the anomalous torque component. The two choices of χ allow
us to confirm Goldreich’s dependence of the alignment of the spin-vector on χ:
eventually we see that a → 0 for χ = 30◦ while a → 90◦ for χ = 75◦ .

(a) χ = 30◦ < χcr (b) χ = 75◦ > χcr

Figure 3.5: Plot of the angle a comparing the numerical solution of Eqn. (3.2)
and the solution to the averaged equation given in Eqn. (3.14) for neutron star
A without the anomalous torque component.

timescale is longer than the initial value defined in Table 3.2 due to the spin-down of
the star. Moreover, since the two values of χ produce opposing evolutions in a, the
precession timescale is also different in the two solutions being shorter for χ = 30◦
than for χ = 75◦ . In contrast, the averaged equations do not include these short term
oscillations because Goldreich averaged over the precessional modulations.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 43

A geometric interpretation of the critical value χcr

The significance of the critical value χcr ∼ 55◦ has not been discussed in the literature.
We will now develop a novel geometric argument, based on the intuitive ideas in §37
of Landau and Lifshitz [100], which provides a way to interpret this value as a natural
consequence of the conservation of energy and angular momentum.

For a biaxial mass distribution as described above, over short timescales compared to
the spin-down, the conservation of energy requires that

Jx2 Jy2 Jz2 J2 Jy2 Jz2

1 = + + = x + + , (3.15)
2Ixx E 2Iyy E 2Izz E 2I0 E 2I0 E 2I0 (1 + I )E

where Ji are the components of the angular momentum. For the conservation of mo-
mentum, we have that

J 2 = Jx2 + Jy2 + Jz2 . (3.16)

Both of these conservation equations have a geometric interpretation in the Cartesian

momentum space: Eqn. (3.15) describes a biaxial ellipsoid with semi-axes given by
√ √ p
2I0 E, 2I0 E and 2I0 (1 + I )E centred on the origin, while Eqn. (3.16) describes a
sphere of radius J also centred on the origin. The ellipsoid is oblate if I > 0 while it
is prolate if I < 0. To keep the following discussion simple, we will assume that I > 0
and then we will discuss this choice at the end of this section.

Physical solutions require the conservation of both the energy and momentum and there-
fore possible solutions lie on the intersection of the sphere and ellipsoid. In order that
the two intersect and hence both conservation laws are satisfied, the radius of the sphere
must be bound by
2EI0 < J 2 < 2EI0 (1 + I ). (3.17)

If the energy and momentum of the system are fixed, the intersection of the sphere
(describing the conservation of momentum) and ellipsoid (describing the conservation
of energy) always describes a circle about the ẑ axis. To illustrate this, in Figure 3.6 we
plot the intersection for three different sizes of the sphere holding the ellipsoid fixed. The
intersection describes exactly the solutions which satisfy both the conservation equations
and therefore the set of all physical solutions.

Over times much shorter than the spin-down timescale, the energy and momentum of
our biaxial star acted on by the spin-down torque are fixed. As we have seen above,
in order to satisfy the conservation equations this means that the angular momentum
vector Ji has a class of possible solution consisting of a ring about the momentum-space
ẑ. Since Ji = Iij ω j , the star is biaxial, and moment of inertia tensor is static. Solutions
for the spin-vector ω j must also exist as points on a related circle. This related circle is
exactly the precession of ω about the ẑ axis.

Both the energy and the momentum decay in time due to the spin-down. As a result
both the sphere and ellipsoid will shrink, but not necessarily at the same rate. If we
imagine observing the ellipsoid such that it appears to be fixed, we may see the sphere
44 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

x y

Figure 3.6: Intersections of the ellipsoid and sphere defined in Eqn. (3.15) and
Eqn. (3.16) respectively for three different values of M at fixed E. The spheres
of radius M are not visible, only the blue line shows where they would intersect
the fixed ellipsoid.

either shrink or grow with respect to the ellipsoid. We can parameterise the relative
rate of shrinking by considering the quantity

J2
A(t) = . (3.18)
2EI0

We could have chosen any one of the ellipsoid’s semi-axes since they are all proportional
up to factors of I . The evolution of the spin-vector is related to the rate of change
of A. This quantity determines whether the sphere shrinks or grows with respect to
the ellipsoid. It is worth taking a moment to understand the different cases which are
parameterised by the sign of Ȧ:

1. Ȧ > 0 The sphere grows with respect to the ellipsoid. In this case, the intersection
circles will ‘close up’ around the ẑ axis. This agrees with the solution for χ = 30◦
given in Figure 3.5(a). The angle between ω and the ẑ axis tends to zero while ϕ
monotonically increases. This describes the spin-vectors following a circle about
the ẑ axis, which gradually decreases in radius.

2. Ȧ < 0 The sphere shrinks with respect to the ellipsoid. In this case the intersection
√
circles will increase in radius until J → 2EI0 . This describes exactly the solution
in Figure 3.5(b): that is a → 90◦ while ϕ monotonically increases.

The sign of Ȧ tells us about the long-term evolution of the system. To determine it, we
begin by differentiating Eqn. (3.18)
 
1 d 2 d 2
Ȧ = (J )2EI0 − (2EI0 )J , (3.19)
(2EI0 )2 dt dt
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 45

simplifying and writing in index notation such that J 2 = J i Ji , we have that

1  i˙ i

Ȧ = 2J Ji E − ĖJ J i . (3.20)
2E 2 I0

Now the energy is purely kinetic such that E = 12 Iij ω i ω j , and therefore Ė = T i ωi . Then,
noting that J˙i = T i , we have

1  i j k i j

Ȧ = J T I
i jk ω ω − T ωi J Jj . (3.21)
E2

For a diagonalised matrix, Iij has only 3 non zero components and therefore Ijk ω j ω k =
(Iω)j ω j . In addition, we can write J i = (Iω)i and so we find

1 i
Ȧ = T ((Iω)i ωj − ωi (Iω)j ) (Iω)j , (3.22)
E2
finally simplifying to  
1 i j i j
Ȧ = T ω ω
i j (Iω) I i − I j . (3.23)
E2
Expanding the summation on the right hand side, all terms for which i = j will be zero.
Working with a biaxial mass distribution where I1 = I2 , then the only non-zero terms
are given by
 
T i ωi ωj (Iω)j I ii − I jj =T1 ω1 I0 (1 − i )ω32 (I0 − I0 (1 + I ))
+ T2 ω2 I0 (1 − i )ω32 (I0 − I0 (1 + I ))
+ T3 ω3 (I0 (1 + I ) − I0 )(I0 ω12 + I0 ω22 ), (3.24)


=I02 I ω3 T3 (ω12 + ω22 ) − ω3 (1 + I )(T1 ω1 + T2 ω2 ) . (3.25)

Working up to 1st order in I and transforming to spherical polar coordinates, we then

insert the spin-down torque and neglect the anomalous torque. This yields
 
2 2 3 2R
ȦE =I0 I ω A ω cos a [sin χ cos χ sin a cos ϕ − sin2 χ cos a] sin2 a
3
3c


− cos a([sin χ cos χ sin a cos ϕ − sin2 χ cos a] cos ϕ sin a − sin2 a sin2 ϕ) . (3.26)

Now we work in the limit τP  τS where ϕ varies on the precession timescale, which is
much faster than the evolution in either a or ω. Therefore, we can average over a free
precession period and drop the constant coefficients:
 
2 2 1 2 2 2


hȦi ∝ cos a − sin a cos a sin χ + sin a cos a cos χ + sin a cos a (3.27)
2
 
2 2 3 2
∝ cos a sin a 1 − sin χ . (3.28)
2

Finally, putting all the coefficients back in, we have an expression for the averaged rate
of change of A under the same assumptions made by Goldreich [69]
 
1 2 6 2R 2 2 3 2
hȦi = I I ω A cos a sin a 1 − sin χ . (3.29)
hEi2 0 3c 2
46 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

This result suggests that only two factors determine the sign of hȦi: the sign of I
and the term 1 − 32 sin2 χ. However, for the argument which led to this conclusion, we
assumed that I > 0. Repeating the argument with I < 0 (starting from Eqn. (3.17)
which would need to be rearranged to reflect this change), we find that Eqn. (3.29) is
multiplied by an overall minus sign. Therefore, the only factor which determines the
sign of hȦi is in fact 1 − 23 sin2 χ, from which we find the critical value χcr ≈ 55◦ ; this is
in agreement with the findings of Goldreich [69]. Moreover, by calculating the condition
in this way, we have found a new interpretation for the critical value as a tipping point
between the rates of change of energy and angular momentum.

3.3.2 Neutron star B in region τS ∼ τP

Unlike the previous examples, there is no comparison in the literature which considers
τP ∼ τS while neglecting the anomalous torque. This is primarily because the similarity
of the timescales means simplifying assumptions cannot be made. The results have not
been included here as they are exactly alike with the results of neutron star A, except
that the precession, the alignment of the spin-vector, and the spin-down all happen on
the same timescale.

3.3.3 Neutron star C in region τS  τP

In neutron star C, the timescales take the ordering τP  τS . The limit of this region
is the spherical star (I → 0) discussed in section 3.2; as such we should expect the
results to be similar. Solving the equations of motion for neutron star C, the spherical
coordinates of the spin axis are given in Figure 3.7. Both choices of χ yield a similar
result; the polar angle a tends to χ on the spin-down timescale as anticipated by the
likeness to the spherical star. In contrast, the azimuthal angle ϕ appears to begin
increasing monotonically before decreasing to a constant value. It is thought that this
is due to the small but non-vanishing magnitude of I . In the perfectly spherical limit,
we observe no such increase in ϕ. Therefore, the small deformation offsets the steady
state solution of the spin axis from the magnetic dipole.

For an intuitive understanding consider that the torque acts to align the spin-vector
with the magnetic dipole; when this occurs the torque vanishes. Since the mass is not
spherical, non-precessing solutions exist when the spin-vector aligns with the principal
axes of the moment of inertia tensor x̂, ŷ and ẑ. So when both effects act, the steady
state solution is an intermediary point between the two, weighted by the relative strength
of each effect. In this case, the torque dominates and so the spin-vector lies close to, but
not exactly aligned with, the magnetic dipole. During this time, the star spins down
but upon alignment the spin frequency asymptotically approaches a constant non-zero
value.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 47

(a) χ = 30◦ < χcr (b) χ = 75◦ > χcr

Figure 3.7: Plot of the spherical coordinates of the spin-vector for neutron star C
without the anomalous torque component.

3.4 Biaxial neutron star including the anomalous torque

In the previous section we considered solutions to Eqn. (3.2) under the action of the
torque in Eqn. (3.5) without the anomalous contribution. In this section, we will include
this contribution and understand the effect it has on the character of solutions. However,
before doing so, we will first consider the appropriate reference frame in which to view
the system.

3.4.1 The effective body frame

Working with a biaxial body and the Deutsch torque, [122] discovered non-trivial so-
lutions where the spin axis aligns with a non-zero angle between the magnetic dipole
and the principal axis of the moment of inertia tensor. This was interpreted by Andrew
Melatos as evidence for persistent precession. We show in Section 3.4.4 that in fact the
spin-vector is aligning with another rotating reference frame that is held at a fixed angle
to the original rotating reference frame. This reference frame, which we will call the ef-
fective body frame, is a direct result of the anomalous torque contribution in Eqn. (3.5);
we will now show how this comes about from a rearrangement of the equations of motion
as first noted by Glampedakis and Jones [66].
48 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

Writing Eqn. (3.2) with the Deutsch torque, we then manipulate it to give

J i,t + ijk Ωj Jk = Nspin-down

i
+ A I0 Ωa m̂a ijk Ωj m̂k , (3.30)
⇒ J i,t + ijk Ωj (Jk − A I0 Ωa m̂a m̂k ) = Nspin-down
i
, (3.31)
⇒ J i,t ijk
+ a
Ωj (Ika − A I0 m̂a m̂k ) Ω = i
Nspin-down . (3.32)

By arranging the equation this way and making use of Ji = Iij Ωj , we can write an
effective moment of inertia tensor given by

0
Ijk = Ijk − I0 A m̂j m̂k , (3.33)
 
I0 (1 − A sin2 χ) 0 −I0 A sin χ cos χ
 
= 0 I0 0 . (3.34)

−I0 A sin χ cos χ 0 I0 (1 + I − A cos2 χ)

Note that we will use the prime to denote the coordinates in the effective body frame.

We can calculate the eigenvalues of the effective moment of inertia tensor, which are
given by
 q 
I0 2 2
λ2 = I0 , λ± = 2 + I − A ± A + I − 2A I cos(2χ) . (3.35)
2

If we diagonalise this effective moment of inertia tensor, these eigenvalues are the diag-
onal entries, and the associated eigenvectors are the principal axes of the effective body
frame. The eigenvalues always take the ordering

λ+ > I0 > λ− . (3.36)

Defining the axes of the effective body frame by ei , it is natural to associate e2 with the
ŷ of the body frame axes such that they are always parallel. Since the other axes must
be orthonormal, the transformation must consist of a rotation in the x − z plane by an
angle β. We are free to set the e3 axis to always take the largest eigenvalue and then
define β as the angle made by e3 with the ẑ axis. β is then given by
 q 
   −  cos(2χ) − 2A + 2I − 2A I cos(2χ)
e3,1 I A
β = arctan = arctan   . (3.37)
e3,3 2A sin χ cos χ

In Figure 3.8, we provide a schematic of the various angles and relating the effective
body frame to the body frame axes.

Understanding the effective body frame The effective body frame is a simple
rotation of the axis defined by the moment of inertia tensor to accommodate the effects
of the anomalous torque. To understand when this becomes significant, we plot β as
a function of |A /I | in Figure 3.9. When |A  I |, such that the anomalous torque
effects are negligible, we recover the usual moment of inertia tensor. That is, β = 0 or
β = π/2, dependent on the sign of I . This is a consequence of choosing e3 to take the
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 49

e3 ẑ
Ω
β
a0 m̂
a χ
e1
x̂

Figure 3.8: Schematic of the two sets of axes for an arbitrary β which in this is
negative. Note that β is defined to be positive for a right hand rotation about
the ŷ axis which is defined to be into the page.

largest eigenvalue. In the opposing limit |A  I |, in which the magnetic deformation
dominates, the sign of the elastic deformation no longer splits the solutions and in both
cases a tends to χ − 90◦ . This means the effective body frame always aligns the e1 axis
(which has the smallest eigenvalue) with the magnetic dipole.

We will show in the following sections that, when the anomalous torque is included, it is
the effective body frame to which the system aligns on the spin-down timescale (much as
it is the body frame to which the system aligns when the anomalous torque contribution
was neglected in Section 3.3).

For the three neutron stars (A, B, and C as first used in Section 3.3) representative of
different orderings of the timescales, all three have a non-zero β values. To familiarise
the reader with the magnitudes in the different regions, in Table 3.3 we list the relevant
timescales and β for each of the three neutron stars.

Star I A B0 [Gauss] τS [s] τA [s] τP [s] β(χ = 30◦ ) β(χ = 75◦ )

A 1.0×10−9 5.0×10−11 1.3×1013 9.0×106 1.3×107 9.8×105 -1.3◦ -0.69◦
B 4.0×10−11 5.0×10−11 1.3×1013 9.0×106 1.3×107 2.4×107 -35.0◦ -8.4◦
C 1.0×10−15 5.0×10−11 1.3×1013 9.0×106 1.3×107 9.8×1011 -60.0◦ -15.0◦

Table 3.3: Table of relevant values for the selected points updated from Table 3.2
to include the value of β, as defined in Eqn. (3.37).

In the following, we will present all results in the effective body frame, rather than the
usual rotating frame. The solutions are first calculated by numerical evolution of the
Euler rigid-body equations with the full torque Eqn. (3.5) in the usual rotating frame.
Then, in the post-processing stage, we apply the effective body frame transformation (a
rotation by the angle β).
50 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

Figure 3.9: Plot of β as a function of the ratio |A /I | for a prolate, I < 0,
and oblate, I > 0, mass distribution with χ = 30◦ . In the limit A  I
for both the oblate and prolate cases, the effective body frame aligns with the
magnetic dipole. In the opposing limit the magnetic dipole has little effect and
the effective body frame aligns with the principal axis of the moment of inertia.
There is a smooth transition between these two extremes when A ∼ I . The
splitting into oblate and prolate solutions results from choosing to associate e3
with the largest eigenvalue.

3.4.2 Phase space

The introduction of the anomalous torque means we must also consider the anoma-
lous torque timescale as given in Eqn. (3.9). Three timescales can take six orderings.
However, the spin-down and anomalous timescale obey the following relation

3c 2π
τS = τA . (3.38)
2R ω0

The coefficient 3c/2Rω0 effectively measures the importance of special relativity at the
surface of the star. Since we expect c  Rω0 , this means that only three physical
orderings exist:

1. Region A: τS > τA > τP

2. Region B: τS > τP > τA

3. Region C: τP > τS > τA

We can therefore update the phase space of possible solutions with the additional anoma-
lous timescale; the result is given in Figure 3.10.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 51

Figure 3.10: Phase space diagram including the anomalous torque contributions.
This differs from Figure 3.3 in that there is an additional grey band marking
‘region B’ where τS > τP > τA .

3.4.3 Neutron star A in region τS > τA > τP

We expect the alignment of the spin axis to agree with the findings in Section 3.3.1
and study the changes in behaviour due to the introduction of the anomalous torque.
Working in the effective body frame rotates the solutions (for star A) by a few degrees
as shown in Table 3.3. Plotted in Figure 3.11 are the spherical coordinates of the spin-
vector in the effective body. This can be compared with the result in Figure 3.4, which
is the same system without the anomalous torque component.

The final state alignment agrees well with results neglecting the anomalous torque for
both values of χ. For χ = 30◦ , the intermediate behaviour appears similar to those shown
in Section 3.3.1. For χ = 75◦ , we have a large discontinuity in the solution suggesting
a significant difference between solutions when including the anomalous torque. To
understand the cause of this, we return to the geometric argument made in Section 3.3.1.

A geometric interpretation Working in the effective body frame for which the
effective moment of inertia tensor is triaxial with principal moments of inertia given by
Eqn. (3.35), the conservation of energy equation is

J12 J2 J32
1= + 2 + , (3.39)
2λ− E 2I0 E 2λ+ E

while the conservation of momentum remains as in Eqn. (3.16), that is

J 2 = J12 + J22 + J32 . (3.40)

52 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

(a) χ = 30◦ < χcr (b) χ = 75◦ > χcr

Figure 3.11: Plot of the spherical coordinates of the spin-vector for neutron
star A. We note that for the χ = 75◦ solution, eventually a0 → 90◦ as the
oscillations damp.

In analogue to the geometric argument of Section 3.3.1, the first equation, the conser-
p
vation of momentum, describes a triaxial ellipsoid with semi axes given by 2λ− E,
√ p
2I0 E, and 2λ+ E, while the conservation of energy still describes a sphere of ra-
dius J. In order to satisfy both conservation equations, the sphere and ellipsoid must
intersect bounding the radius of the sphere by

2Eλ− < J 2 < 2Eλ+ . (3.41)

In the case of a biaxial moment of inertia, we find that the solution for which both
conservation laws are satisfied corresponded to circles around the ẑ axis. Now that we
have a triaxial moment of inertia tensor (due to the addition of the anomalous torque),
the solutions correspond to the intersection of the triaxial ellipsoid with the sphere, for
which there are three distinct types of intersections as drawn in Figure 3.12.

Each of these types of intersection can be understood through the orderings of the radius
of the sphere and the sizes of the three semi-axes of the ellipsoid:

(a) 2Eλ− < J 2 < 2EI0 . In this region, curves form two sets of complete loops around
the e1 axis. As J 2 → 2Eλ− the intersection loops will close up about the e1 axis.

(b) J 2 = 2EI0 . This is a special case in which the intersection forms two closed ellipses

(c) 2EI0 > J 2 > 2Eλ+ . In this region, curves form two sets of complete loops around
the e3 axis. As J 2 → 2Eλ+ the intersection loops will close up about the e3 axis.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 53

e3 e3 e3

e1 e2 e1 e2 e1 e2

(a) 2Eλ− <J 2 <2EI0 (b) J 2 =2EI0 (c) 2EI0 <J 2 <2λ+

Figure 3.12: Intersections of the ellipsoid and sphere defined in Eqn. (3.39) and
Eqn. (3.40) respectively for three different values of M at fixed E.

Now that we have understood the possible types of solution via the intersections of
the sphere and ellipsoid, we can understand the cause of the unusual behaviour in
Figure 3.11(b). First, we note that all simulations begin in the same initial state. As
previously observed, changing χ through the critical value changes the sign of Ȧ and
hence whether the sphere will grow or shrink with respect to the ellipsoid. So for χ = 30◦ ,
the sphere grows with respect to the ellipsoid (Ȧ > 0) ending up with the loops closing
up about the e3 axis. For χ = 75◦ the sphere shrinks with respect to the ellipsoid
(Ȧ < 0) ending up with the loops closing up about the e1 axis. The two end states are
then the (a) and (c) pictures in Figure 3.12.

If they both begin from the same initial configuration, then one of them must pass
through the special J 2 = 2EI0 state corresponding to Figure 3.12(b). When this hap-
pens, we will observe a change in the axis about which the solution precesses.

Both sets of solution start with the solution precessing about the e3 axis (ϕ monoton-
ically increasing) corresponding to Figure 3.12(c). In Figure 3.11(b), the discontinuity
is in fact exactly the point when the solution changes the axis of precession.

To better understand this, we plot the data from the χ = 75◦ simulation (Figure 3.11(b))
in Figure 3.13. Firstly in (a), we project the spherical coordinates onto the unit sphere
and choose three short sections of data. This shows that at early times (in blue) the
solution precesses about the e3 axis. During the discontinuity (in red) the solution shows
similarities with the J 2 = 2EI0 case from Figure 3.12. Finally, at late times, in black,
the solution precesses about the e1 axis. In Figure 3.13(b) we also plot the solution in
3D during the apparent discontinuity. This again shows that the solution goes through
a change of the axis about which it precesses.

3.4.4 Neutron star B in region τS > τP > τA

Solutions for neutron star B take on a new significance having included the anomalous
torque – we can now make a comparison with the work of Melatos [122]. In this work,
54 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

(a) (b)

Figure 3.13: Numerical solutions for neutron star A with χ = 75◦ plotted in
3D showing the point where the spin-vector changes the axis about which it
precesses. Figure (a) shows the evolution of the spin-vector on the unit sphere
at three distinct time steps: blue marks the time before the discontinuity, red
is during the discontinuity, and black is some time after. Figure (b) shows the
complicated behaviour of the spin-vector during the discontinuity, the changing
colour is used to illustrate the evolution with time, progressing from green to
blue.

Melatos found that when the precession timescale and anomalous torque timescales are
comparable, the solutions involve persistent precession. Specifically, this is defined as
the polar angle a tending to a constant, non-zero value with the nutation amplitude
(oscillations in a) either decaying or remaining constant. Because the spin-vector has
not aligned with a principal axis of the moment of inertia, this was interpreted as the
spin-vector undergoing persistent precession.

We simulate neutron star B with the properties listed in Table 3.3 and a magnetic
inclination angle χ = 75◦ . In Figure 3.14(a) and Figure 3.14(b) we plot the spherical
components of the spin-vector in the body frame and effective body frame respectively.
Figure 3.14(a), the components in the body frame, confirms the finding by Melatos that
a 6= 90◦ at the end of the simulation.

However, transforming to the effective body frame in Figure 3.14(b), we find that
the spin-vector aligns with the principal axis of the effective moment of inertia ten-
sor (a0 → π/2). Therefore, the persistent precession angle found by Melatos is precisely
the angle β. The spin-vector has aligned with the principal axis of the effective moment
of inertia tensor associated with the smallest eigenvalue. Since both the polar angle a0
and azimuthal angle ϕ0 are constant (this is also true in the body frame), we argue that
the angular velocity is constant and this is not persistent precession: the spin-vector
always aligns with the effective body frame and in doing so the precession of the body
is damped.
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 55

(a) In the body frame axis (b) In the effective body frame axis

Figure 3.14: Plot of the spherical coordinates of the spin-vector for neutron
star B. Crucially, at late times the polar angles a and a0 differ by a factor 8.4◦ ,
the value of β in Table 3.3.

3.4.5 Neutron star C in region τP > τS > τA

In Figure 3.15, for all values of χ, the spin-vector is found to align with the effective
body frame e1 axis. In this limit where A  I , from Figure 3.9, we see that the e1
axis aligns with the magnetic dipole. This agrees with the results for a spherical star
in which I → 0. The addition of the anomalous torque introduces oscillations on the
anomalous torque timescale.

3.5 Conclusion

In this chapter, we have studied a numerical model comprising a rigid body supporting a
biaxial strain and acted on by the Deutsch [54] dipole torque. In agreement with Melatos
[122], we confirm the analytic solutions of Goldreich [69]: when τP  τS the inclination
of the spin-vector a either damps or grows on the spin-down timescale depending on
whether χ is smaller or larger than χcr ≈ 55◦ . Moreover, we provide a new interpretation
of this critical value as a tipping point between the rates of change of energy and angular
momentum.

When τP  τS we find that the spin-vector aligns with the magnetic dipole; this also
agrees with the numerical and analytic solutions found by Melatos [122]. When including
the anomalous torque, in the regime where τP ∼ τS , we confirm the existence of so-called
persistent precession solutions found by Melatos [122] in which the spin-vector does not
align with either the principal axis of the moment of inertia tensor or the magnetic
56 Chapter 3 Action of the electromagnetic torque on a precessing neutron star

(a) χ = 30◦ < χcr (b) χ = 75◦ > χcr

Figure 3.15: Plot of the spherical coordinates of the spin-vector for neutron
star C.

dipole. However, we provide a new insight into this by showing that for all orderings
of the three timescales, the spin-vector aligns with the principal axes of the effective
moment of inertia tensor (defined in Section 3.4.1). In this sense, there is no persistent
precession.

By studying solutions with and without the anomalous torque, we conclude that the
anomalous torque is not important for realistic systems since it amounts to nothing
more than a small shift in the axes about which that spin-vector precesses. For this
reason, the solutions found by Goldreich [69] is the most relevant for real pulsars and it
will be this that we use in the rest of this thesis.

Appendix 3.A Considerations for the timescales

We have used the ordering of the three timescales at ω0 = ω(t = 0) to categorise the
results. However, clearly the magnitude of the spin-vector will decrease and the different
dependencies on the spin-vector could cause a ‘crossing’ of the timescales producing
unexpected results. Writing the timescales as

2π
τP (t) = , (3.42)
I ω(t)
2π
τA (t) = , (3.43)
A ω(t)
3c 2π
τS (t) = , (3.44)
2R A ω(t)2
Chapter 3 Action of the electromagnetic torque on a precessing neutron star 57

only τP could ‘cross’ with the other two timescales since the spin-down and anomalous
timescales obey τS > τA provided that

3c
ω(t) > . (3.45)
2R
This condition is satisfied by setting the initial rotational spin frequency at less than
ω0 ∼ 104 [rad/s], that is the star should not break special relativity. At later times, the
spin frequency will decay and so this condition is still satisfied.

For the anomalous torque and precession timescale crossings, two cases exists: either the
deformations are such that the spin-down timescale is initially larger than the precession
timescale (region A and B) or, as in region C, the precession timescale is larger than the
spin-down timescale. Let us consider the first case:

3c I
τS > τP ⇒ ω(t) < . (3.46)
2R A

In this particular state, I > A and so, as in the previous case, while the spin-vector
decays this inequality is always satisfied and the orderings remain the same. In the other
case we have
3c I
τP > τS ⇒ ω(t) > . (3.47)
2R A
3c I
In this case, it is possible for the timescales to cross at ωcr = 2R A . For pulsar C,
this corresponds to a rotational spin frequency of 0.9 [rad/s], which is four orders of
magnitude smaller than the initial frequency. For this reason, we can rule out this
crossing of the timescales as an important factor in calculations.
Chapter 4

Modelling observations of
precessing pulsars

In Section 3, we investigated the role of precession under an EM torque in the rotating

frame of the star. From this we learnt that the Goldreich [69] analytic solutions, which
ignore the anomalous component of the Deutsch [54] torque, are the relevant solutions
for the precession of a torqued pulsar. We will now model the effect of precession on
the ‘observable features’ of a pulsar, by which we mean the phase residual, spin-down
rate, and pulse profile. This will require us to view the pulsar in the inertial frame of
an observer and model the data collection methods of pulsar astronomy as part of a
predictive model of pulsar precession.

Modelling observations of precessing pulsars has a history dating back to Ruderman

[149] and Chiuderi and Occhionero [41] when they were attempting to explain a small
amplitude ‘wobble’ in the frequency of the Crab pulsar [147, 146]. Since then, new
observations have prompted many workers in the field to develop predictive models and
in the course of this chapter we will discuss their contributions when relevant. The novel
material presented here is in numerically solving the Euler rigid-body equations under
the full Deutsch [54] torque in addition to solving for the Euler angles which allows us to
make direct predictions of the observable features measured by observers. Moreover, we
will also model the data collection mechanisms such as the observer-method to measure
the evolution of the spin-down rate. We will compare this numerical model against
the analytic results from the literature, in particular we will use the work of Jones and
Andersson [90] as it is the most complete work on the subject. By comparing with a
numerical model, we can gain confidence in the use of analytic solutions. For example, in
Chapter 5 we will make use of the analytic solution for the spin-down rate of a precessing
pulsar which we will derive in Section 4.5.1; this derivation neglects the anomalous torque
discussed in Section 3, but we can be confident in its predictions because in Section 4.5.2
we compare the analytic approximation against numerical solutions calculated with the
full torque.

This chapter is organised in the following way. In Section 4.1 we will introduce the idea of
the Euler rotation angles and how to formulate a predictive model for precessing pulsars

59
60 Chapter 4 Modelling observations of precessing pulsars

in the inertial frame. In Section 4.2 we compare some typical numerical solutions to
analytic solutions in the torque free case and then illustrate the effect of the full Deutsch
[54] torque on the evolution of the Euler angles. Before discussing the observable features
of pulsars, we need some intuition to understand the results, so in Section 4.3 we give a
detailed account of the motion of the magnetic dipole of a freely precessing pulsar. The
observable features of a precessing pulsar are discussed in Section 4.4, Section 4.5, and
Section 4.6 for the timing residuals, spin-down rate, and shape of pulsations respectively.
Finally, in Section 4.7 we give some preliminary findings on a hybrid model of precession
and magnetospheric switching before concluding in Section 4.8.

4.1 Rotating into the inertial frame

4.1.1 Euler rotation matrices

The Euler rigid-body equation, introduced in Chapter 3, is defined in the rotating frame
of the star. To discuss results in the inertial frame of an observer, we need to transform
the solutions of Euler’s rigid-body equations into the inertial frame. An efficient way to
do this is to determine the three Euler angles which transform the rotating frame axes,
denoted by (x0 , y 0 , z 0 ), to the inertial frame axis, denoted by (x, y, z). In particular, we
define three rotation matrices
 
cos ψ sin ψ 0
 
B =  − sin ψ cos ψ 0  , (4.1)
0 0 1
 
1 0 0
 
C =  0 cos θ sin θ  , (4.2)
0 − sin θ cos θ
 
cos φ sin φ 0
 
D =  − sin φ cos φ 0  . (4.3)
0 0 1

Then their product defines the Euler angle rotation matrix

R = BCD. (4.4)

In index notation the components are

 
cosψ cosφ−cosθ sinφsinψ cosψ sinφ+cosθ cosφsinψ sinψ sinθ
 
Rab =  −sinψ cosφ−cosθ sinφcosψ −sinψ sinφ+cosθ cosφcosψ cosψ sinθ  . (4.5)
sinθ sinφ −sinθ cosφ cosθ

Here we are using the (φ, θ, ψ) Euler angle parameterisation as described in §35 of Landau
and Lifshitz [100]; a diagram of how these angles are defined is given in Figure 4.1. This
matrix, given values for the Euler angles, transforms from the rotating frame to the fixed
inertial system of coordinates. So for Arot
b , a vector defined in the inertial frame, in the
Chapter 4 Modelling observations of precessing pulsars 61

Figure 4.1: The Euler rotation angles between the rotating frame (x0 , y 0 ,0 z 0 ) and
the inertial frame (x, y, z). Image courtesy of Brits [36].

rotating frame the vector has components given by

Ain b rot
a = R a Ab . (4.6)

4.1.2 Evolution of the Euler angles

The Euler angles themselves will evolve with time. To calculate this, we express the
components of the spin-vector Ωa along the moving axes (x0 , y 0 , z 0 ). As shown by Landau
and Lifshitz, this results in a set of three ODEs which are coupled both to the other
Euler angles and the components of the spin-vector. To have solutions both for the
motion of the spin-vector in the rotating frame and the Euler angles we need to solve
all six ODEs together. To be specific, the three Euler rigid-body equations for a biaxial
body with a moment of inertia as defined in Eqn. (2.8) are given by

Tx
Ω̇x = − I Ωy Ωz , (4.7)
I0
Ty
Ω̇y = + I Ωx Ωz , (4.8)
I0
Tz
Ω̇z = . (4.9)
I0 (1 + I )

The rearranged set of Euler angle ODEs from Equation (35.1) of Landau and Lifshitz
[100] are given by

Ωx sin ψ + Ωy cos ψ
φ̇ = , (4.10)
sin θ
θ̇ = Ωx cos ψ − Ωy sin ψ, (4.11)
ψ̇ = Ωz − φ̇ cos θ, (4.12)
62 Chapter 4 Modelling observations of precessing pulsars

where I = ∆I/I0 is the deformation of star. This set of six coupled ODEs can be solved
numerically using a time stepper; we will use the rkf45 stepper provided by GSL [72].
Solutions give the components of the spin-vector in the rotating frame and the evolution
of the Euler angles which can be used to transform rotating frame quantities into the
inertial frame. Unlike the results of Section 3, numerical solutions to these ODEs require
the fast spin frequency to be resolved and hence require greater computing time.

4.1.3 Initial conditions

Solving the rigid-body equations, as in Section 3, we set Ωa (t = 0) to lie in the x0 − z 0

plane at an angle a0 to the z 0 axis. This gives a set of three initial conditions

Ωx = ω0 sin(a0 ), Ωy = 0, Ωz = ω0 cos(a0 ), (4.13)

where ω0 is the initial magnitude of the spin-vector, recalling that we define ω = |Ωa |.

For the Euler angle equations, Eqn. (4.11) to Eqn. (4.10), the initial conditions need to
be chosen carefully so that the result can be meaningfully interpreted. In particular, we
need to be sure we understand how the inertial frame is orientated. Let us note that
the angular momentum in the two frames are related by

Jain = Rb a Jbrot . (4.14)

where Rb a is defined in Eqn. (4.5).

We have already set the initial condition on Ωa in Eqn. (4.13). Therefore the initial
angular momentum in the rotating frame is given by

Jarot (t = 0) = I ba Ωb (t = 0). (4.15)

If we set an initial condition on the angular momentum in the inertial frame Jain , then
Eqn. (4.5) uniquely defines the initial Euler angles. We choose to set the initial angular
momentum in the inertial frame to lie along the inertial z axis such that

Jain (t = 0) = |Jain |ẑ. (4.16)

The magnitude of the angular momentum is

q
b
|Ja | = |I a Ωb | = ω0 I0 sin2 a0 + cos2 a0 (1 + I )2 . (4.17)

Using the inverse of Eqn. (4.14) and substituting in Eqn. (4.15) and Eqn. (4.16) we have
   
sin a0 q sin ψ0 sin θ0
   
 0  = sin2 a0 + cos2 a0 (1 + I )2  cos ψ0 sin θ0  . (4.18)
(1 + I ) cos a0 cos θ0

This gives us three equations for two unknowns. Our choice to set Jain along the z axis
leaves the initial value of φ as a free variable. We set φ(t = 0) = 0 without loss of
Chapter 4 Modelling observations of precessing pulsars 63

generality. Rearranging the third component of Eqn. (4.18) yields

" #
(1 + I ) cos a0
θ0 = arccos p 2 . (4.19)
sin a0 + (1 + I )2 cos2 a0

In the limit I  1, it can be seen that θ0 ≈ a0 . For ψ0 , we rearrange the first component
of Eqn. (4.18) to give

1 sin a0
sin ψ0 = p . (4.20)
sin θ0 sin a0 + (1 + I )2 cos2 a0
2

√
To simplify the first factor, we use the identity sin(arccos(x)) = 1 − x2 along with
Eqn. (4.19) giving
 1/2
sin2 a0
sin θ0 = . (4.21)
sin2 a0 + (1 + I )2 cos2 a0
Inserting this into Eqn. (4.20) and rearranging we find that

sin a0
sin ψ0 =
1/2 (4.22)
sin2 a0
sin a0
= (4.23)
| sin a0 |
= sign(a0 ), (4.24)

where by sign(x) we mean the sign of x. Finally, the initial condition is given by
π
ψ0 = sign(a0 ) . (4.25)
2
We can check the sanity of this result by inserting it into the second component of
Eqn. (4.18) and finding that it balances the left hand side.

In this section, we have defined the appropriate initial conditions for the system: in
addition to Eqn. (4.13) defining the initial spin-vector, we set the angular momentum in
the inertial frame to lie along the z axis and fix φ(t = 0) = 0 without loss of generality.
While the initial conditions on the spin-vector are arbitrary, if the initial Euler angle
are not carefully defined then we do not have a meaningful interpretation for how the
inertial frame is orientated.

4.2 Evolving the Euler angles

Having defined the coupled ODEs in Eqn. (4.7) to Eqn. (4.12) and appropriate initial
conditions, we can now calculate the evolution of the spin-vector and Euler angles. The
ultimate aim of this section is to go on to simulate the observable features. However,
let us begin by considering the evolution of the simulation parameters (the components
of Ωa and the three Euler angles). We will consider first a torque free simulation in
Section 4.2.1 and then add in the full torque in Section 4.2.2.
64 Chapter 4 Modelling observations of precessing pulsars

4.2.1 Torque free biaxial body

The set of ODEs given in Eqn. (4.7) to Eqn. (4.12) for a biaxial body free of torques
have simple analytic solutions [100] given by

θ(t) = θ0 ≈ a0 , (4.26)
φ(t) = φ̇t + φ0 = φ̇t, (4.27)
∆I π
ψ(t) = ψ̇t + ψ0 = − φ̇t + sign(a0 ) , (4.28)
I0 2

where in the second equalities we have inserted the initial conditions. Solution to the
rigid-body equations were derived in Eqn. (2.11) to Eqn. (2.13), transforming these into
the spherical coordinates of the spin-vector we have

|Ωa | = ω0 , (4.29)
a = a0 , (4.30)
ϕ = I ω0 cos(a0 )t, (4.31)

where ϕ is the azimuthal angle of the spin-vector with respect to the x0 − z 0 plane.
This set of analytic solutions gives us a method to verify our numerical solver with the
appropriate initial conditions.

In Figure 4.2 we present the numerical solutions for some arbitrary values of the simu-
lation parameters along with the analytic prediction of Eqn. (4.26) to Eqn. (4.31). This
demonstrates ‘almost’ perfect agreement between the two: the components of Ωa behave
as expected; φ monotonically increases at the spin frequency; ψ decreases at the slower
precession frequency; the polar angle θ should remain constant during this simulation,
however, we find it varies fractionally by ∼ 10−11 . This error is caused by the finite
numerical precision when performing the subtraction in Eqn. (4.11).

4.2.2 Torqued biaxial body

To see the difference introducing the EM torque of Eqn. (3.5) makes to the solutions, in
Figure 4.3 we repeat the simulation plotted in Figure 4.2 with both the anomalous and
spin-down components of the EM torque. The most striking contrast is the ‘wobble’
in both θ and a; on closer inspection one also finds this wobble in the other angles,
while the magnitude of Ωa both wobbles and undergoes a secular spin-down. To help
illustrate this we have additionally plotted the derivatives φ̇ and ψ̇ which were constant
in the torque free case. These are calculated numerically, the numerical errors produce
the observed noise.

These solutions show how, using numerical simulations, we can easily simulate the effect
of the full Deutsch [54] torque. In the rest of this chapter, we will discuss how to model
the features of a star observed by pulsar astronomers. This will allow us to change the
physics of the star, i.e. by modifying the torque, and directly simulate the effect this has
on the observable properties.
Chapter 4 Modelling observations of precessing pulsars 65

ω [rad/s] 10
8
6
4
2
0

90
a [deg]

75
60
45
30
15
0
×103
2.5
ϕ [deg]

2.0
1.5
1.0
0.5
0.0
0 τp 2τp 3τp 4τp 5τp 6τp 7τp
time

(a) Spherical coordinates of the spin-vector in the rotating

frame

15.6
θ [deg]

15.3
15.0
14.7
14.4
×106
3.0
2.5
φ [deg]

2.0
1.5
1.0
0.5

×103
0.0
ψ [deg]

−0.6
−1.2
−1.8
−2.4
−3.0
0 τp 2τp 3τp 4τp 5τp 6τp 7τp
time

(b) Euler angles

Figure 4.2: Solutions to the ODEs defined in Eqn. (4.7) to Eqn. (4.12) for
a torque free biaxial star with a deformation of I = 10−3 . The red dashed
line is the analytic calculation found by Jones and Andersson [90] as given in
Eqn. (4.26) to Eqn. (4.31).
66 Chapter 4 Modelling observations of precessing pulsars

10.01
10.00

ω [rad/s]
9.99
9.98
9.97
9.96
9.95
9.94

90
a [deg]

75
60
45
30
15
0
×103
3.0
ϕ [deg]

2.5
2.0
1.5
1.0
0.5
0.0
0 τp 2τp 3τp 4τp 5τp 6τp 7τp
time

(a) Spherical coordinates of the spin-vector in the rotating

frame

24
22
θ [deg]

20
18
16
14
×106
3.0
2.5
φ [deg]

2.0
1.5
1.0
0.5

286.8
φ̇ [deg/s]

286.4
286.0
285.6
285.2
×103
0.0
ψ [deg]

−0.8
−1.6
−2.4
−3.2

−0.24
ψ̇ [deg/s]

−0.32
−0.40
−0.48

0 τp 2τp 3τp 4τp 5τp 6τp 7τp

time

(b) Euler angles and selected derivatives

Figure 4.3: Solutions to the ODEs defined in Eqn. (4.7) to Eqn. (4.12) including
the torque defined in Eqn. (3.5) for a biaxial body with A = I /2 and I =
1×10−3 ; the ratio of the precession timescale to the spin-down age is ≈ 7×10−4 .
Note the x-axis is in units of τP , the initial precession period.
Chapter 4 Modelling observations of precessing pulsars 67

4.3 Precessing pulsars

Before we move to modelling the observable features such as phase-residuals, we need to

build our intuition for precession. To do this, we will introduce the notion of a reference
plane from which we can decompose the motion of vectors into rotations about cones.

4.3.1 The reference plane

Considering a biaxial body, let us define n̂a as the deformation vector (a unit vector
pointing along the body’s symmetry axis). Then the moment of inertia tensor can be
written compactly as

I ab = I0 δ ab + ∆I n̂a n̂b , (4.32)

where δ ab is the Kronecka delta. The principal moments are given by I0 = Ixx = Iyy
and Izz = I0 + ∆I such that the body is biaxial with n̂a lying along the ẑ axis. As such,
the angular momentum is given by

Ja = I0 Ωa + ∆IΩz n̂a . (4.33)

As found by Pines and Shaham [136], this shows that the three vectors Ja , Ωa , and n̂a
are coplanar in the so-called reference plane as shown in Figure 4.4. In this figure, we

J Ω
n
a
^
ϑ
ϑ

Figure 4.4: The reference plane containing the spin-vector Ω, the angular mo-
mentum vector J and the deformation vector n.

have defined the angle between angular momentum and n̂a as θ, conventionally referred
to as the wobble-angle of precession. We have also given a, the angle between the ẑ-axis
and the spin-vector as defined in Section 3.

Following Jones and Andersson [90], let us now decompose the angular velocity into the
Euler angle components along the angular momentum and symmetry axis (these are the
same Euler angles as defined in Section 4.1.1)

Ωa = φ̇Jˆa + ψ̇n̂a , (4.34)

68 Chapter 4 Modelling observations of precessing pulsars

where Jˆa is a unit vector pointing along the angular momentum. Then, substituting
this into Eqn. (4.33) we find that
 
Ja = I0 φ̇Jˆa + I0 ψ̇ + ∆IΩz n̂a . (4.35)

Comparing components, we have that

|Ja | = I0 φ̇, (4.36)

and
∆I
ψ̇ = − Ωz . (4.37)
I0

To relate the rates of change of the two angles, we can use Eqn. (4.12), giving

∆I ∆I
ψ̇ = − φ̇ cos θ ≈ − φ̇ = −I φ̇ (4.38)
I0 + ∆I I0

where we assumed ∆I  I0 and θ  1 in the last step.

We can relate both of these frequencies to the angular frequency ω = |Ωa |. First, from
the cosine rule we have that

|Ωa |2 = φ̇2 + ψ̇ 2 − 2φ̇ψ̇ cos θ. (4.39)

and so rearranging and neglecting terms of order 2I , we have that

ω
φ̇ ≈ √ . (4.40)
1 − 2I cos θ

Since I  1, we know that φ̇ is approximately the fast spin frequency of the star ω.
Then from Eqn. (4.38), we see that ψ̇  ω is a much slower frequency, the precession
frequency.

We will now discuss the effect of free precession on the motion of m̂a , the magnetic
dipole. We will do this initially by studying the evolution of the dipole in the inertial
frame via the Euler angle rotations and then we will return to this reference frame
picture in Section 4.3.3 to provide a deeper intuitive understanding.

4.3.2 Dynamics of the magnetic dipole

Numerical solutions to Eqn. (4.7) to Eqn. (4.12) allow us to calculate the motion of
any quantity in the inertial frame from which the neutron star is observed. This can
be used, for example, to calculate the motion of the spin-vector as seen by an observer
for any arbitrary torque. However, pulsar astronomers observe the pulsar through the
pulsations of EM emission. If this emission is collinear with the dipole, then it points
along the unit vector of the magnetic dipole m̂a . Therefore, we are particular interested
in the motion of m̂a in the inertial frame. This system was first studied in the follow
Chapter 4 Modelling observations of precessing pulsars 69

way by Bisnovatyi-Kogan et al. [30], but here we follow the treatment by Jones and
Andersson [90].

In the rotating frame, we set m̂a to lie at an angle χ to the z 0 axis with unit vector
[sin(χ), 0, cos(χ)] without loss of generality. Using Eqn. (4.5), the Euler rotation matrix,
we can transform to the inertial frame; the components of the magnetic dipole in the
inertial frame are
 
cos φ cos ψ sin χ − sin φ cos θ sin ψ sin χ + sin φ sin θ cos χ
 
m̂a =  sin φ cos ψ sin χ + cos φ cos θ sin ψ sin χ − cos φ sin θ cos χ  . (4.41)
sin θ sin ψ sin χ + cos θ cos χ

We define two angles Φ and Θ which describe the polar and azimuthal angles of m̂a in
the inertial frame. From Eqn. (4.41) the azimuthal angle is given by
    
m̂y π 1 cos ψ tan χ
Φ = arctan = φ − + arctan , (4.42)
m̂x 2 cos θ tan θ − sin ψ tan χ

while the polar angle is

Θ = arccos(m̂z ) = arccos(sin θ sin ψ sin χ + cos θ cos χ). (4.43)

Eqn. (4.42) and Eqn. (4.43) are exact results and we will use them later in this chapter
to model the observable features such as the phase residuals and spin-down rate. Specif-
ically, given a numerical solution to our system of coupled ODEs, i.e. the evolution of
the spin-vector and three Euler angles, we can use these two equations to describe the
evolution of the magnetic dipole orientation in the inertial frame. Before we do this, let
us now build some feeling for the dynamics of the magnetic dipole in the inertial frame
in the torque free case.

Using our numerical solution, we simulate a star without any EM torque. The resulting
solutions for the Euler angles can then be substituted into Eqn. (4.43) to give the evo-
lution of Θ. This is done for three choices of θ and χ in Figure 4.5. From this figure, we
see that the polar angle is modulated at the precession period, τP . In the next section,
we will provide a deeper understanding of the size of modulations and the significance
in the choice of θ and χ.

The observed spin frequency of a pulsar is the rate with which observers measure the
observed pulsations. Usually we think of this as exactly equivalent to the rotation
frequency ω/2π. However, for a precessing star the observed frequency is not given
by ω/2π due to the additional motions of precession. Instead, the observer would fit a
timing model to the TOAs as the pulse passes through the plane containing the observer
and the angular momentum vector, the phase of which is Φ. So the frequency measured
would be given by the first derivative of Φ:
 
sin χ ψ̇(cos θ sin χ − sin ψ sin θ cos χ) + θ̇ cos ψ(cos θ cos χ − sin ψ sin θ sin χ)
Φ̇ = φ̇ + .
(sin θ cos χ − cos θ sin ψ sin χ)2 + cos2 ψ sin2 χ
(4.44)
70 Chapter 4 Modelling observations of precessing pulsars

5θ
χ = θ/2
χ=θ
4θ χ = 4θ
magnetic dipole Θ
Polar angle of the

3θ

2θ

3θ/2

θ/2

0
0 τp 2τp 3τp 4τp 5τp
time

Figure 4.5: The effect of free precession on the polar angle of the dipole m̂a as
given by Eqn. (4.43) for three choices of χ and θ.

This is the instantaneous electromagnetic frequency; an observer will measure the time
averaged value of Φ̇ as the ‘spin frequency’ of the star.

In Figure 4.6 we plot the frequency modulations for the three choices of θ and χ used in
Figure 4.5. Again, we see modulation at the precession period; in the following section,
we will provide a deeper understanding of these modulations and our choices of θ and χ.

4.3.3 Understanding the dynamics of the magnetic dipole

We will now return to the reference plane picture discussed in Section 4.3.1. In particular,
we provide a detailed description of the motion of m̂a as seen by the observer and use
this to explain the modulations seen in Figure 4.5 and Figure 4.6.

Firstly, let us state our assumption that EM radiation is emitted along the dipole axis m̂a
such that it is this axis from which an observer sees pulsations. From the decomposition
of the spin-vector in Eqn. (4.34), the motion of the magnetic dipole, a vector fixed
in the rotating frame, when viewed in the inertial frame, can be understood as the
superposition of two rotations. If we hold φ fixed, then m̂a rotates in a cone of half-
angle χ about the symmetry axis n̂a at the slow precession angular frequency ψ̇ ≈ −I ω
and we call this the precession cone. On the other hand, if we hold ψ fixed, then n̂a
rotates in a cone of half-angle θ about Ja at the angular frequency φ̇ ≈ ω. Still holding
ψ fixed, since m̂a is necessarily always at a fixed angle χ to n̂a , it too will be swung in a
cone of half-angle Θ (as given in Eqn. (4.43)) about the angular momentum vector with
angular frequency φ̇; let us define this to be the dipole cone. Now, ψ evolves at a slower
Chapter 4 Modelling observations of precessing pulsars 71

χ = θ/2

Instantaneous electromagnetic
χ=θ
χ = 4θ

frequency Φ̇ φ̇

φ̇ + ψ̇

0 τp 2τp 3τp 4τp 5τp

time

Figure 4.6: The effect of free precession on the instantaneous electromagnetic

frequency given by Eqn. (4.44) for three choices of χ and θ.

rate to φ by a factor I , moreover, the negative relation between them in Eqn. (4.38),
indicates that the precession cone counter-rotates with respect to the dipole cone. If we
consider the effect of both rotations together and their effect on the dipole cone, we see
that the half-angle of the dipole cone Θ is modulated by precession (see Figure 4.5): as
m̂a rotates about the dipole cone, the cone gradually opens and closes at the precession
period. The rate at which m̂a rotates about the dipole cone is given by Φ̇, as given by
Eqn. (4.44). Similarly this is also modulated by precession as shown in Figure 4.6.

The resulting motion of the dipole cone is best understood by considering three choices
of χ and θ: χ < θ, the special case χ = θ, and χ > θ. We illustrate the projection of the
precession cone swept out by m̂a about n̂a (in red) and the cone swept out by n̂a about
Ja onto the reference plane for the three orderings of χ and θ in Figure 4.7. Note that
the dipole cone is not shown here.

Let us now describe the evolution of the dipole cone for the three cases in Figure 4.7.
We will discuss in particular the choices of χ and θ which were used in Figure 4.5 and
Figure 4.6 for the variations in polar angle and instantaneous electromagnetic frequency
respectively.

• The χ < θ case (χ = θ/2): the precession cone is narrow and does not extend
over the angular momentum vector. The polar angle Θ of the dipole cone oscil-
lates periodically between θ + χ and θ − χ during a precession cycle, as shown in
Figure 4.5. The spin frequency Φ̇, as shown in Figure 4.6, has an average value of
φ̇ and oscillates about this value. Comparing with the Θ variations demonstrates
these oscillations are locked in phase with the rotation of m̂a in the precession
cone. Recalling that the precession cone counter-rotates with respect to the spin
72 Chapter 4 Modelling observations of precessing pulsars

φ̇ φ̇
ψ̇ φ̇
J J ψ̇
ω̂ z n̂ ω̂ z J ψ̇
n̂ ω̂ z n̂
χ m̂
θ θ χ χ −θ
m̂ θ χ

m̂

(a) χ < θ (b) χ = θ (c) χ > θ

Figure 4.7: Diagrams depicting the projections of the cone swept out by m̂a
about n̂a (in red) and the cone swept out by n̂a about Ja (in blue) onto the ref-
erence plane for torque free precession. We show the three orderings of χ and θ.

cone, at θ + χ the precession cone motion acts in the opposing direction to the
spin cone. This causes a reduction in the spin frequency away from the average.
By contrast, at θ − χ the precession cone motion acts in the same direction as the
spin cone producing an increase in the spin frequency above the average.

• The χ = θ case: in this special case the angular momentum vector sits exactly on
the side of the precession cone. This suggests that at certain precessional phases
m̂a can align exactly with the angular momentum. When this happens the spin
frequency tends to zero manifesting as sharp dips in the spin frequency; at the
same time the polar angle tends to zero.

• The χ > θ case (χ = 4θ): The precession cone now extends over the angular
momentum vector. This means it always acts to reduce the spin frequency; as a
result the spin frequency has an average value of φ̇ + ψ̇. The polar angle can vary
between θ + χ and χ − θ. For χ close to θ the deviations away from the average
are large while as χ increases the deviations get smaller as the half angle of the
dipole cone increases.

4.3.4 The effective wobble angle

The angle θ (see Figure 4.4) is referred to as the wobble angle since, in torque-free
precession, its magnitude determines the ‘amount’ of precession. Taking Eqn. (4.33) at
an instant when Ωy = 0 without loss of generality it can be shown that

I0 1
tan θ = tan(a) = tan(θ + θ̂). (4.45)
I0 + ∆I 1 + I

Using the angle sum identity for tan(θ + θ̂) and expanding about I = 0 it can be shown
[90] that

θ̂ ≈ I sin θ cos θ. (4.46)

Chapter 4 Modelling observations of precessing pulsars 73

Then for θ  1, we have that a ≈ θ + O(I ).

When including the EM torque the wobble angle does not have the intuitive interpre-
tation of the ‘amount’ of precession. In particular, the anomalous part of the torque
defined in Eqn. (3.5) can be understood to create an effective rotating frame as shown
in Section 3.4.1. This means that solutions exist which appear to undergo ‘persistent
precession’ where θ ≈ a 6= 0 such that the body remains misaligned from the principal
axes of its moment of inertia tensor. However, we demonstrated that in fact the body
has aligned with the principal axes of it effective moment of inertia tensor.

Let us then define an effective wobble angle

θ̃ = θ − β, (4.47)

which describes the amount of precession under the full electromagnetic torque. That
is, we rotate the usual wobble angle by β defined in Eqn. (3.37). In the limit where
β → 0 we recover the usual wobble angle referred to by Jones and Andersson [90] θ = θ̃.
But, if the effects of the anomalous torque are important (see Chapter 3 for examples
of when this is the case), then the effective wobble angle θ̃ will measure the amount of
precession (i.e. its magnitude determines the amplitude of precession). This definition
will be used later on when setting up solutions which initially have a minimal amount
of precession.

4.4 Observable features: the phase residual

The principal observational quantity reported on for a pulsar is the timing residual.
This is the difference between the measured TOA of a pulse and a timing model of the
pulsar as discussed in Section 1.2. The phase modulation of pulsars due to precession
was first modelled in response to periodic variations observed in the frequency of the
Crab pulsar [149, 41]. Precession as a candidate for periodic variations in Hercules
X-1 was studied by Bisnovatyi-Kogan et al. [30] and Bisnovatyj-Kogan and Kahabka
[31]. Since then, there has been extensive work in the literature: Nelson et al. [128]
calculated phase residuals both analytically and numerically for freely precessing stars;
torqued precession was considered for a general torque by Jones [93] and Cordes [44] and
then for the Deutsch [54] torque by Melatos [121, 122]. In this section, we will calculate
residuals by calculating the pulsar phase from our numerical model and then fitting and
removing a Taylor expansion. This means that we capture the data collection methods
present in any real data. Note that timing residuals and phase residuals are equivalent
up to factors of the pulse period. We will discuss this once we have defined the phase
residual.

The azimuthal angle Φ given in Eqn. (4.42) gives us the phase of the magnetic dipole. For
the observer, the pulsation occurs when the dipole passes through the plane containing
the observer and the angular momentum vector. We can always reorient the observer
such that the pulsation occurs every time Φ is a multiple of 2π. In this way, Φ is also the
phase of the observed pulsations. To generate Φ we define the star’s properties i.e. the
74 Chapter 4 Modelling observations of precessing pulsars

magnetic field and initial angles, then numerically evolve Eqn. (4.7) to Eqn. (4.12). The
resulting time series of the spin-vector components and Euler angles are then substituted
into Eqn. (4.42) to generate the exact phase.

Following the methods used by observers, we define our ‘timing model’ as a Taylor
expansion of the phase about some reference time tref up to f¨:
!
f˙ f¨
Φfit (t; tref , f, f˙, f¨) = 2π f (t − tref ) + (t − tref ) + (t − tref ) ,
2 3
(4.48)
2! 3!

with the timing properties (f , f˙, f¨) as free parameters. Higher order terms can be
included, as discussed in Section 2.2, but here we truncate at f¨. We do not include
an initial phase since we can arbitrarily define our reference time to coincide with a
pulsation such that the initial phase is zero. Unlike pulsar astronomers, we do not need
to worry about other corrections such as the motion of the Earth since Φ is given in the
inertial frame.

The phase residual is then the difference between the exact phase and the fitted phase

∆Φ(t) = Φ(t) − Φfit (t; tref , f, f˙, f¨). (4.49)

We then use a least-squares fitting method to minimise the root mean squared error of
the phase residual. The fitted coefficients {f, f˙, f¨} constitute our best-fit phase model:
the best fit to the data Φ(t) of a power law spin-down model. It is worth noting that a
residual depends on which section of data was used in the fit.

In this work, we will report only on phase residuals. However, we note that the phase
residual can also be re-scaled to give the timing residual by calculating the residual as
a fraction of a cycle then multiplying by the period:

∆Φ(t)
∆T = P. (4.50)
2π

In the following subsections, we will compare analytic results for the phase residuals due
to precession with our numerical results found by fitting and removing a polynomial. As
shown in Section 4.3.3 the form of solution will depend on whether χ < θ or χ > θ. Of
these two choices, Jones and Andersson [90] concluded that ‘the wobble angle of rapidly
rotating stars are limited to small values for the finite crustal breaking strain’; following
this reasoning we will limit our study to the χ > θ case although the numerical code is
capable of finding solutions for either.

4.4.1 Effect of precession on the phase residual

The effect of free precession on phase residuals was first considered by Nelson et al. [128];
since the modulations are due to the geometric effect of precession, they are referred to as
the geometric effect. To calculate a phase residual, Jones and Andersson [90] subtracted
the secular phase evolution from Eqn. (4.43) and, found that the phase residual when
Chapter 4 Modelling observations of precessing pulsars 75

χ > θ is given by  π
∆Φ49 (t) = −θ cot χ cos ψ̇t + , (4.51)
2
where the subscript here refers to the equation number of Jones and Andersson [90]. For
a precessing star ψ̇ is given by Eqn. (4.38) and the initial value of the cosine argument
is ψ(t = 0) = π/2 as derived in Section 4.1.3.

Equation (4.51) is calculated in the absence of any EM torque. Nevertheless, it is still

appropriate when a torque is applied provided that the geometric effect is stronger that
any other (this is discussed in the next few sections). As such, we begin by simulating
a star with the properties listed in Table 4.1. This uses typical values for the angular
spin frequency, magnetic field, and a small, but non-zero angle a. The resulting phase
residual, in cycles, is given in Figure 4.8. This figure shows strong agreement between
the numerical result and the analytic prediction of Eqn. (4.51).

Simulation parameters
ω0 = 10 rad/s
B0 = 1013 G
χ = 50◦
a0 = 2◦
AEM = 4.3×10−6

Table 4.1: Simulation parameters used for the phase residual plotted in Fig-
ure 4.8. Note that AEM is the electromagnetic amplification factor defined later
in Eqn. (4.54).

×10−3
6
Numerical
∆Φ49
4
Phase residual [cycles]

−2

−4

−6
0 τp 2τp 3τp 4τp 5τp
time

Figure 4.8: The simulated phase residual in cycles for a simulated star with
the properties described in Table 4.1. We also plot the corresponding analytic
prediction of Eqn. (4.51) which is the geometric effect of precession on the phase
residual.
76 Chapter 4 Modelling observations of precessing pulsars

4.4.2 Effect of torqued precession on the phase residual: electromag-

netic amplification

The geometric effects of precession can be amplified by the EM torque [44]. Using a
simple description based on a vacuum point-dipole spin-down torque, and calculating
the departure from a non-precessing power-law spin-down, the amplified phase residuals
are given by
 
1  τP  τP
∆Φ63 = − θ cot χ sin(ψ̇t + π/2), (4.52)
π P τage

where τage = |Φ̇/Φ̈|, note this is equivalent up to a factor 2 with the original defini-
tion in Eqn. (1.10); here, we will use the definition from Jones and Andersson [90] for
consistency. This result is derived in Jones and Andersson [90] for the magnitude of
modulations. Here we keep the exact time evolution behaviour intact. We will de-
rive this expression later on in Section 4.5.1. Notably, we can write the magnitude of
modulations in term of Eqn. (4.51) as
 
1  τP  τP
|∆Φ63 | = |∆Φ49 | (4.53)
π P τage

The two ratios of timescales define an ‘amplification factor’

τ   τ 
P P
AEM = . (4.54)
P τage

The amplification, as first noted by Cordes [44], increases the magnitude of phase resid-
uals for young pulsars with short periods. It is worth noting however, that there is a
relative phase difference between Eqn. (4.51) and Eqn. (4.52).

To verify this amplification, we simulate a star using the properties in Table 4.2 for
which we use a larger angular spin frequency (this aids in speeding up the computation)
and a larger magnetic field than the properties listed in Table 4.1: this results in an
amplification factor which is greater than unity. The resulting phase residual is plotted
in Figure 4.9 along with the predictions of Eqn. (4.52) and Eqn. (4.51). This clearly
demonstrates that the amplified residuals agree with the results calculated numerically
while the geometric effect is both out of phase and has a smaller amplitude by a factor
AEM /π, where the amplification factor is given in Table 4.2.

Simulation parameters
ω0 = 104 rad/s
B0 = 1014 G
χ = 50◦
a0 = 2◦
AEM = 43

Table 4.2: Simulation parameters used for the phase residual plotted in Fig-
ure 4.9 and the spin-down rate plotted in Figure 4.10.
Chapter 4 Modelling observations of precessing pulsars 77

0.08
Numerical
∆Φ49
0.06
∆Φ63

0.04

Phase residual [cycles]

0.02

0.00

−0.02

−0.04

−0.06

−0.08
0 τp 2τp 3τp 4τp 5τp
time

Figure 4.9: The simulated phase residual in cycles for a simulated star with
the properties described in Table 4.2. We also plot the corresponding analytic
prediction for the geometric effect of precession given by Eqn. (4.51) and the
EM amplification of precession given by Eqn. (4.52) which is the appropriate
prediction for this simulation.

4.5 Observable features: the spin-down rate

The derivatives of the phase (the frequency, spin-down rate, etc.) display similar periodic
departures from their secular evolutions; in the same way as for the phase residual, these
periodic modulations can be amplified by the EM torque. In Figure 4.6 we have already
shown the modulations of the frequency due to free precession. However, long term
modulations of the frequency (on top of the usual secular spin-down), are not reported
in the literature and so we will not consider them further. The second observable feature
that we will consider is the spin-down rate for which modulations are reported in the
literature.

PSR B1828-11 is a normal radio pulsar which displays strong periodic modulation in
its spin-down rate. This has been modelled analytically by several authors [156, 90,
108, 14] who all found that the timing residuals and spin-down rate modulations were
consistent with EM-dominated precession. We can confirm this by noting that the
inferred precession period for this pulsar is τP ≈ 500 days, the period is P ≈ 0.405 s,
and the spin-down age is τage ≈ P/Ṗ ≈ 2.1 × 105 yrs (values taken from the ATNF
catalogue Manchester et al. [117]), therefore the EM amplification factor defined in
Eqn. (4.54) is

AEM ≈ 700. (4.55)

78 Chapter 4 Modelling observations of precessing pulsars

In this section, we will derive an analytic model for the spin-down rate of a precessing
pulsar dominated by the EM torque; this model will be used in Chapter 5 to test the
precession model for PSR B1828-11. Then, we will compare this analytic model with a
spin-down rate calculated numerically. We will not consider the geometrically dominated
spin-down rate modulations. Further details on that subject can be found in Jones and
Andersson [90].

4.5.1 Derivation of the precession spin-down rate

Let us now derive the spin-down rate for a precessing pulsar under a vacuum point-
dipole spin-down torque. Equivalent derivations can be found in Link and Epstein [108]
and Jones and Andersson [90]; in this derivation we will start with a generalisation of
vacuum point-dipole torque to allow for a braking index n 6= 3, but retain the angular
dependence. The following provides the complete details of a calculation we will use
later in Chapter 5 and which was published in Ashton et al. [25].

Our generalisation of the vacuum point-dipole spin-down torque can be written as

Φ̈ = −k Φ̇n sin2 Θ, (4.56)

where k is a positive constant. From Eqn. (4.43) we have that

cos Θ = sin θ sin ψ(t) sin χ + cos θ cos χ. (4.57)

Rearranging and expanding about θ = 0 and keeping terms up to O(θ2 ), we find

  
2 2 2 2 sin2 χ 1
sin Θ = sin χ + θ cos χ − − 2θ sin χ cos χ sin ψ(t) + θ2 sin2 χ cos(2ψ(t)),
2 2
(4.58)

where the first term, in square brackets, is a constant while the second two terms provide
the first and second harmonic modulations in sin2 Θ.

In order to find approximate solutions to Eqn. (4.56), we begin by defining the time-
averaged (constant) value of sin2 Θ as
 
2 2 2 2 sin2 χ
sin Θ0 = sin χ + θ cos χ − , (4.59)
2

then solving Eqn. (4.56) with this constant value we get

" # −1
n−1
1
Φ̇(t) = (n−1)
+ (n − 1)k sin2 Θ0 t , (4.60)
Φ̇0

where Φ̇0 = Φ̇(0). Now the spin-down timescale is

|Φ̇0 | 1
τage = ≈ , (4.61)
|Φ̈0 | k|Φ̇0 |(n−1) sin2 Θ0
Chapter 4 Modelling observations of precessing pulsars 79

again we note that this differs by a factor of 2 to the definition in Eqn. (1.10), we use
this definition here for consistency with Jones and Andersson [90]. Rearranging gives

1
k= . (4.62)
˙
τage |Φ0 |(n−1) sin2 Θ0

Then Eqn. (4.60), our zeroth-order solution to the spin-down (under a constant Θ), can
be written as
  −1
t n−1
Φ̇(t) = Φ̇0 1 + (n − 1) . (4.63)
τage

Now, we substitute Eqn. (4.63) back into Eqn. (4.56) along with the expanded, but
complete variation in sin2 Θ, given by Eqn. (4.58), giving
  −n
−1 t n−1
Φ̈(t) = 2
Φ̇n0 1 + (n − 1)
τage |Φ̇0 |(n−1) sin Θ0 τage
  (4.64)
2 1 2 2
× sin Θ0 − 2θ sin χ cos χ sin ψ(t) + θ sin χ cos [2ψ(t)] .
2

In principle, this is complete and constitutes an approximation of the spin-down under

precession. To simplify this expression, we expand the second bracket with t/τage  1,
then
  
1 t 1 1 2 2
Φ̈(t) = Φ̇0 −1 + n + 2θ sin χ cos χ sin ψ(t) − θ sin χ cos(2ψ(t))
τage τage sin2 Θ0 2
 2   !
t t
+O ,θ
τage τage
(4.65)

Next we expand 1/ sin2 Θ0 , from Eqn. (4.59), in θ:

  −1
2 2 2 sin2 χ 1
sin χ + θ cos χ − ≈ 2 + O(θ2 ). (4.66)
2 sin χ

Note that we only to expand to this order since this term is multiplied by a factor of θ
and we are neglecting O(θ3 ) terms. Then simplifying we have
  
1 t 1 2
Φ̈(t) = Φ̇0 −1 + n + 2θ cot χ sin ψ(t) − θ cos(2ψ(t)) . (4.67)
τage τage 2

Finally, converting to spin-frequencies and periods

 
1 t 1 2
ν̇(t) = −1 + n + 2θ cot χ sin ψ(t) − θ cos(2ψ(t)) (4.68)
P τage τage 2
80 Chapter 4 Modelling observations of precessing pulsars

We have not made any assumptions on ψ(t) in this derivation. However, since we are
interested in cases where τP  τage we can make an assumption that

ψ(t) = ψ̇t + ψ0 . (4.69)

From Figure 4.3 (where τP /τage ≈ 7 × 10−4 ) we see that the torque causes modulations
in ψ̇. However, in all the cases where we apply this formulae, these modulations will be
negligible.

The O(θ) modulation term in this expression can be integrated twice to derive Eqn. (4.52);
this reflects that this spin-down rate prediction includes the amplification of the EM
torque. When AEM < 1, Eqn. (4.68) is not suitable since the geometric variations
(derivatives of Eqn. (4.51)) will dominate. We will not consider the geometric domi-
nated spin-down rates in this work, although our simulation code was tested against
known analytic results and found to agree.

4.5.2 Simulations of the precession spin-down rate

In this subsection, we will verify Eqn. (4.68) against our numerical model. In Chapter 3,
we concluded that the anomalous torque, which is not included in the derivation of
Eqn. (4.68), was not important for realistic pulsars. By verifying that this equation
agrees with numerical solutions (calculated with the anomalous torque) we can confirm
this conclusion and build confidence in the analytic solution which will be used later in
Chapter 5.

To calculate spin-down rates numerically, we could calculate the pulse phase (Eqn. (4.42))
and then numerically differentiate to get Φ̈. However, when studying the long-term mod-
ulations in the spin-down rate, observers (see for example Lyne et al. [111], Perera et al.
[134]) use what we will refer to as the ‘observer-method’: a second order Taylor expan-
sion is fit to short sections of data of length T and the resulting coefficient ν̇ is recorded
at the mid-point of the section of data; repeating this process every ∼ T /4 in a ‘sliding-
window’ builds a picture of how the spin-down varies with time. We will mimic this data
collection process by fitting Taylor expansions to short sections of the simulated phase
Φ. This has the benefit that we include any potential peculiarities of the data collection
mechanism in our simulation. We choose T such that it is a fraction of the precession
period over which we expect quantities to be modulated. This is consistent with the
observer-method where T is chosen in order to resolve the observed modulations.

To verify that Eqn. (4.68) accurately models the spin-down rate evolution of precessing
pulsars in an EM amplification regime, we simulate a star with the properties given in
Table 4.2 (the same set of properties used for the simulated phase residual in Figure 4.9).
Most significantly, AEM  1, which ensures that the EM torque amplifies the spin-down
rate modulations. We use an initial spin frequency ω0 much larger than typical values,
this is to allow the numerical solutions to evolve with reasonable computational power;
for more physical values the solutions are qualitatively unchanged.
Chapter 4 Modelling observations of precessing pulsars 81

The numerical spin-down rate variations, calculated using the observer-method are plot-
ted along with the analytic predictions of Eqn.(4.68) in Figure 4.10. This figure demon-
strates good agreement: small variations result from the application of the observer-
method, most notably the numerical results under and over estimates the minimum and
maximums due to the time averaging.

×10−3
−5.2
Numerical result
Spin-down rate ν̇

−5.4 Analytic

−5.6

−5.8

−6.0

−6.2
τp 2τp 3τp 4τp 5τp
time

Figure 4.10: Spin-down rate for simulation parameters listed in Table 4.2 com-
pared against the corresponding analytic prediction of Eqn.(4.68).

Figure 4.10 captures the essential features of the spin-down rate of a precessing pulsar
amplified by an EM torque. However, there is a special case which produces the ‘double-
peaked’ spin-down rate which is observed in PSR B1828-11 [111] (see also Chapter 5)
and PSR 0919+06 [134]. This double-peak arises naturally in the precession model when
θ  1 and χ is close to π/2. This can be seen directly from the O(θ2 ) term in Eqn. (4.68):
when χ ∼ π/2, we get a second harmonic at τP /2. To illustrate this, we will repeat the
simulation of Figure 4.10, but set χ = 85◦ ; the full set of simulation parameters are
listed in Table 4.3. The spin-down rate for this ‘almost orthogonal’ dipole simulation is
plotted in the top panel of Figure 4.11 and demonstrates the distinctive double peak.

Simulation parameters
ω0 = 104 rad/s
B0 = 1014 G
χ = 85◦
a0 = 10◦
AEM = 75

Table 4.3: Simulation parameters for the spin-down rate plotted in Figure 4.11.

We can understand this double-peaked feature by recalling that the dipole-cone (intro-
duced in Section 4.3.3) is the cone swept out by the magnetic dipole, m̂a , at the fast
spin frequency; the half-angle this cone makes with the angular momentum vector is Θ
(see Eqn. (4.43)). In the lower panel of Figure 4.11 we plot this polar angle for the same
simulation. Notably the occurrence of the double-peak in the spin-down rate coincides
with times when Θ > 90◦ . What we are seeing is the magnetic dipole entering the
lower hemisphere of the star. By considering the form of the Deutsch [54] torque, we see
that if the torque is maximal when the dipole is perpendicular to the spin-vector. Since
θ  1 and therefore close to the angular momentum vector, we have a maximum in the
82 Chapter 4 Modelling observations of precessing pulsars

×10−3
−9.0
−9.1 Numerical result

Spin-down rate ν̇
Analytic
−9.2
−9.3
−9.4
−9.5
−9.6
−9.7
−9.8
−9.9

100
magnetic dipole Θ
Polar angle of the

95
90
[degs]

85
80
75
70
τp 2τp 3τp 4τp 5τp
time

Figure 4.11: Top panel: spin-down rate for simulation parameters listed in
Table 4.3 compared against the corresponding analytic prediction of Eqn.(4.68).
This simulation differs from Fig. 4.10 in that we chose χ such that the magnetic
dipole lies close to the rotation equator. Bottom panel: the variation in the
polar angle of the magnetic dipole, Θ, for the simulation.

torque and hence spin-down rate when Θ ≈ 90◦ . As Θ increases past 90◦ and the dipole
enters the lower hemisphere, the absolute value decreases again causing the distinctive
double-peak in the spin-down rate.

4.6 Observable features: the pulse profile

So far we have considered observable features which are calculable from the timing
properties of the star: the rate at which pulses occur. However, pulsar astronomers
also report on the shape of the pulsation by averaging over many pulses to form an
integrated pulse profile. Long term variability in pulse profiles is observed in pulsars.
The correlated changes in shape for PSR B1828-11 allowed Stairs et al. [156] to rule out
a planetary precession hypothesis. More examples of this variability can be found in
Lyne et al. [111].

To quantify the changing shape of pulsations from PSR B1828-11, astronomers calculate
the shape parameter S. This is defined by first noting that PSR B1828-11 alternatives
between two ‘modes’, one narrow and one broad, then defining S as the fraction of the
mean pulse shape attributed to the narrow component (for more details see Stairs et al.
[157]). The effect of precession on S for a variety of beam geometries has been considered
by Akgün et al. [14]. However, this shape parameter is limiting as it requires there to
be two components. A more general method was introduced by Lyne et al. [111] and
Chapter 4 Modelling observations of precessing pulsars 83

defines the beam-width as the width of the pulsation at some fraction of the maximum
observed intensity.

In the context of precession, the pulse profile is sensitive to both the geometry of the
beam itself and the angle made between the beam and the observer. Let us define an
observer fixed in the inertial frame such that they maintain a constant angle ι with
the angular momentum of the star Ja . For this observer, a pulse can be defined as the
moment the dipole cuts through the plane containing them and the angular momentum
vector. At this moment, the angle between them and the dipole will be determined by
both ι ∈ [0, π] and Θ ∈ [0, π]. In particular, the angle between the observer and the
beam is given by
∆Θ = Θ − ι. (4.70)

Since ι is fixed, variations in ∆Θ come solely from variations in the polar angle Θ. We
studied these variations in Figure 4.5 and found that Θ has an average value of the
larger of θ or χ, then oscillates about this on the precession period with a magnitude
given by the smaller value of θ or χ.

In this section, we aim to numerically simulate variations in the beam-width due to

precession. To do this, we first model variations in the pulse intensity and then show
how to model the determination of the beam-width. The process described here, will be
developed later in Chapter 5 to build a predictive analytic model for the beam-width of
PSR B1828-11.

4.6.1 Variations in the pulse intensity

The intensity of radiation received by an observer will depend on the orientation of the
magnetic dipole with respect to the observer and the beam geometry. It may presumably
be maximal when pointing directly at the observer and fall off as the angle between the
two grows. For each rotation of the star, the intensity will peak when the beam cuts
the plane containing the observer and the angular momentum vector; at this instant
the angle between the observer and the beam is given by Eqn. (4.70). If the star is
precessing, then the periodic pulses of intensity due to the azimuthal rotation of the star
will be modulated by the slower variations in Θ seen in Figure 4.5.

To model this, we take an observer’s azimuth and polar angle to be (Φobs , ι) and then
assume the beam geometry follows a Gaussian profile with a single conal emission. This
could later be adapted, for example to include conal emission as modelled by Akgün
et al. [14]. For such a model of the beam geometry, the intensity of pulsations will vary
with the angular separation of the vector from the centre of the star to the observer and
the magnetic dipole vector. By considering the intersection of these vectors with the
unit sphere, the angular separation can be shown to be

∆d = cos−1 (cos(Θ) cos(ι) + sin(Θ) sin(ι) cos(Φ − Φobs )) (4.71)

84 Chapter 4 Modelling observations of precessing pulsars

Then, taking a Gaussian beam geometry, the intensity of the pulse will be given by
 
∆d2 (Θ, Φ, ι, Φobs )
I(Θ, Φ, ι, Φobs , ρ) = I0 exp − (4.72)
2ρ2

where I0 is the maximum intensity and ρ is a measure of the width of the Gaussian
intensity.

Given a value for I0 and ρ, we can use a numerical solution to the governing ODEs to
simulate this pulse intensity exactly. This is done in Figure 4.12 for a system where
the variations in Θ occur on a timescale not much longer than the pulse period. This
nonphysical simulation is intended to show the modulation of the individual pulsations
due to precession. We can also predict the maximum pulse intensity at any given instant

0.9
observered intensity
0.8 maximum intensity
Normalised Intensity

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0 0.5τp 1τp
time

Figure 4.12: Amplitude variation using a 2D Gaussian emission.

by setting Φ = Φobs . Then simplifying, we find that

 
−(Θ − ι)2
Imax (Θ, Φ, ι, ρ) = I0 exp , (4.73)
2ρ2

this is also shown in Figure 4.12.

4.6.2 Variations in the beam-width

It is unlikely that the absolute variations in intensity seen in Figure 4.12 will ever be
unambiguously visible in nature. This is because in real observations the intensity will
also be subject to variations in the amount of scintillation from the interstellar medium.
Therefore, pulsar astronomers do not typically report on intensities themselves, but
characterise the pulsation by their beam-width. This is the width of the pulse at some
percentage p of the observed maximum intensity. Note that this is not the maximum
intensity that the beam produces, I0 , but the maximum at that instant in time Imax ,
as given by Eqn. (4.73)
Chapter 4 Modelling observations of precessing pulsars 85

To calculate the beam-width, we first note that in Eqn. (4.72), Θ varies on the slow
precession timescale, while Φ varies on the rapid spin timescale: we are looking to
measure the variations with respect to the slow precession timescale. The pulse width is
measured by the time spent above a percentage p of the maximum pulse intensity, this
can be defined as an inequality
p
I(Φ, Θ, ι, Φobs , ρ) > Imax . (4.74)
100
Substituting in Eqn. (4.72) and Eqn. (4.73) then rearranging yields
r h i
p Imax
cos −2ρ2 ln 100 I0 − cos(Θ) cos(ι)
cos(Φ − Φobs ) > (4.75)
sin(Θ) sin(ι)

where we note that Imax is not a constant, but will evolve with the polar angle Θ as
given by Eqn. (4.43).

Let’s consider a single rotation with the magnetic dipole starting and ending in the
antipodal point to the observer’s position. During this rotation, Φ − Φobs increases
linearly between −π and π, and so the left hand side of the inequality is a simple cosine
function as illustrated in Figure 4.13. Since we expect Θ to vary slowly compared to the
rotation period we can, over a single pulsation, think of Θ as a constant; then the whole
right hand side of Eqn. (4.75) is a constant. In Figure 4.13, we illustrate this constant
along with the evolution of the left hand side. Then we also illustrate ∆Φ, the fraction
of the rotation period for which the cosine is less than the constant, i.e. Eqn. (4.75) is
satisfied.

const.
cos(Φ − Φobs )

∆Φ

Φobs − π Φobs Φobs + π

Dipole azimuth Φ

Figure 4.13: Illustration of cos(Φ − Φobs ) during a single rotation. The constant
value represents the right hand side of this equation. The width ∆Φ indicates
the angular period during which Eqn. (4.75) is satisfied.
86 Chapter 4 Modelling observations of precessing pulsars

The angular fraction at which the inequality is satisfied is given by ∆Φ. To convert this
into the beam-width reported by pulsar astronomers we multiply by the pulse period.
So from ∆Φ, we can calculate the beam-width as

∆Φ
Wp = P (4.76)
2π
where P is the spin period.

The angular width ∆Φ when the inequality is satisfied is given by

 r h i 
2 p Imax
 cos −2ρ ln 100 I0 − cos(Θ) cos(ι) 
∆Φ = 2 cos−1 

,
 (4.77)
sin(Θ) sin(ι)

and so the beam-width is

 r h i 
2 p Amax
 cos −2ρ ln 100 A0 − cos(Θ) cos(ι) 
1
Wp = cos−1 

.
 (4.78)
π Φ̇ sin(Θ) sin(ι)

Finally, we can also simplify this results using Eqn. (4.73) to give
 q  
2 − 2ρ2 ln p

1  cos (Θ − ι) 100 − cos(Θ) cos(ι) 
Wp = cos−1  . (4.79)
π Φ̇ sin(Θ) sin(ι)

To demonstrate an example of the beam-width modulation, in Figure 4.14 we plot Θ

calculated from a numerical simulation with the simulation parameter listed in Table 4.4
and W10 as calculated from Eqn. (4.79) with ι = 82◦ and ρ = 0.3.

Simulation parameters
ω0 = 1 rad/s
B0 = 0G
χ = 80◦
a0 = 6◦
AEM = 0

Table 4.4: Simulation parameters for the beam-width modulations plotted in

Figure 4.14.

This shows that there is periodic modulation of the beam-width, with an interesting
two-peak structure. This structure can be understood by realising that the polar angle
of the dipole Θ ‘passes over’ the observer such that for some of the precessional phase
the observer views the beam from below, and other from above. This leads to the two-
peaked structure due to the symmetry of the emission geometry. This is a special case
which depends on the choices of θ, χ, and ι: more generally the beam-width has a single
period oscillatory behaviour.
Chapter 4 Modelling observations of precessing pulsars 87

θ+χ

magnetic dipole Θ
Polar angle of the
ι
χ

|θ − χ|

0.2080

0.2075

0.2070
W10

0.2065

0.2060

0.2055
0 τp 2τp
time

Figure 4.14: Simulation results for the polar angle Θ and W10 , a measure of the
observed beam-width as calculated from Eqn. (4.79). The simulation parameters
for this are listed in Table 4.4 and we additionally set ι = 82◦ and ρ = 0.3 when
calculating W10 .

This numerical approach to modelling the beam-width allows us to simulate the effects of
precession and probe how this manifests in our observations. We will use the methods
discussed in this section later in Chapter 4 when developing and fitting a precession
model for the observed beam-width of PSR B1828-11.

4.7 Application: switching and precession

Recently, some workers in the field [111, 134] have suggested that quasi-periodic structure
observed in some pulsar timing residuals is a result of magnetospheric torque-switching
events as described in Section 2.3.3. In such models, the magnetospheric torque, and
hence spin-down rate periodically switches between two distinct values and these changes
correlate with changes in the beam-width. These models however are lacking a key
feature: the clock which provides the periodicity. It has been suggested [89] that it
may in fact be precession which provides this clock. Ultimately, the numerical model
developed in this section could study this effect, for example by implementing a hybrid
model in which propensity for a magnetospheric switch to occur is related to the angle
Θ. In this way, the observed switching would undergo stochastic resonance as suggested
by Cordes [45] and discussed in Appendix 2.A.

In this section, we will present novel, but preliminary results on a simplified hybrid
magnetospheric switching - precession model. Ultimately we would like to realise the
ideas discussed above in a full numerical model. However, for now we will simply consider
88 Chapter 4 Modelling observations of precessing pulsars

the effect of a single magnetospheric torque switching event. Our numerical code is able
to model the full hybrid model, but we must first understand the basic physics. In
particular, we will simulate a star in which at tswitch = Tobs /2 the EM torque suddenly
decreases. This is a qualitatively new idea which has not yet been discussed in the
literature.

In the EM dipole spin-down model, the torque has two distinct components: the reg-
ular spin-down component and the anomalous component. This latter term does not
contribute to the spin-down, but as discussed in Section 3.4.1 will modify the axis of
precession. Magnetospheric switching models are based on evidence that the spin-down
rate is switching, therefore the torque switching must occur in the spin-down compo-
nent. However, it is unclear if it will also occur in the anomalous component. To answer
this, one would need a detailed model of how the magnetosphere reconfigures during the
switching. We will not do this here, but instead investigate a simple phenomenological
switching event in which we modify Eqn. (3.5) as follows

T = (1 − SS H(t − tswitch ))TS + (1 − SA H(t − tswitch ))TA (4.80)

where the subscripts label the spin-down and anomalous components, S is the strength
of switching, and H(t) is the Heaviside step function. In this model we can control which
components are switched by choosing SS and SA appropriately.

In our model the strength of the EM torque is parameterised by A , related to the surface
magnetic field strength by

R5 2
A = B . (4.81)
4I0 c2 0

Rearranging Eqn. (1.12) we can then write the spin-down rate as

B02 R6 sin2 (α)ω03

ω̇0 = −
6I0 c3
(4.82)
2RA sin2 (α)ω03
=− ,
3c
where α is the angle between the spin-vector and magnetic dipole. Since we expect
these to be misaligned in order to observe pulsations, we can take sin2 α ≈ 1, then our
spin-down rate is approximately

Rω03
ν̇ = − A . (4.83)
3πc

From this we can equate the switching in the spin-down rate SS directly to that measured
from pulsar observations. That is, from Eqn. (4.80) we have that

A → 0A = (1 − SS )A . (4.84)

Chapter 4 Modelling observations of precessing pulsars 89

Then the spin-down rate also changes as

ν̇ → ν̇ 0 = (1 − SS )ν̇. (4.85)

during a switching event.

4.7.0.1 Minimal precession initial state

In the following section we will investigate the effects of a single switch, but first we
choose to define an initial ‘minimal precession’ from which to begin our simulations.
This is done to mimic the state of real pulsars, which in general are not found to
be precessing. Moreover, it helps to discern the effects of the switching from that of
precession.

Precession will not occur when the spin-vector is aligned with the principal axis of the
effective moment of inertia tensor (defined in Sec. 3.4.1). The angle between these is
approximately the effective wobble angle defined in Eqn. (4.47). For minimal precession
we should therefore set this wobble angle to zero. In all simulations, we consider a
biaxial body with the full torque given by (4.80). From our previous discussion on the
wobble angle we can minimise the precession initially by setting the initial polar angle
of the spin-vector in the rotating frame to lie along the effective body-frame axis. In the
presence of the anomalous torque this is done with

a0 = β(I , A , χ), (4.86)

where β is defined in Eqn. (3.37).

In Figure 4.15 we illustrate the phase residual of our simulation in the absence of a
switching event; the selected simulation parameters are listed in Table 4.5. We use a large
magnetic field so that the effect of switching can be easily identified, a typical angular
spin frequency, and an initial polar angle which is exactly the angle β, as calculated
using Eqn. (3.37).

Simulation parameters
ω0 = 100 rad/s
B0 = 1015 G
χ = 50◦
a0 = -0.78◦
AEM = 23

Table 4.5: Simulation properties used for Figure 4.15, Figure 4.16, and Fig-
ure 4.17.

Notably this minimal precession solution does show some precession with phase resid-
uals ∼ 10−5 . This is because the spin-down torque produces a wobble in the angular
momentum vector and as a result the wobble angle is not truly zero. For all the phase
residuals calculated in this section, we will fit and remove a Taylor expansion up to and
including ν̈.
90 Chapter 4 Modelling observations of precessing pulsars

×10−5
4
3

Phase residual [cycles]

2
1
0
−1
−2
−3
−4
0 τp 2τp 3τp 4τp 5τp 6τp 7τp 8τp 9τp 10τp 11τp
time

Figure 4.15: The phase residual for a minimal precession simulation with no
switching event; the simulation properties are listed in Table 4.5. In this case,
we have fitted and removed a Taylor expansion up to and including ν̈. Note
that, due to inherent errors in the fitting process, the residual exhibits an ap-
parent periodicity which matches its span (or some integer number of periods
fits into the span, in this case 2). This was shown to be the case by varying the
observation span and observing a corresponding shift in the period. As such it
is a non-physical effect that can be ignored.

We will now set up simulations of this ‘minimal precession‘ star, and then manually
switch the torque. We choose a star where the EM torque amplification is important.
In the following sections, we will consider first switching in the spin-down torque, and
then also in the anomalous torque. For all simulations considered here we will use the
properties listed in Table 4.5.

4.7.1 Switching in the spin-down torque only

We now consider manually switching the spin-down torque halfway though the simula-
tion, with no switch occurring in the anomalous component. That is we set

Tobs
tswitch = , SS = 0.4, SA = 0.0, (4.87)
2
such that halfway though the simulation the spin-down torque is reduced by a fraction
0.4 while the anomalous torque remains unaffected.

In Figure 4.16 we plot the phase residuals from this simulation. In the top plot is
the residual as calculated over the entire observation period. We find a single periodic
variation with a period of the observation time Tobs : due to sudden change in spin-down
rate the Taylor expansion fit is not finding a suitable fit. The precession features which
existed in the initial minimal precession are swamped by the larger variations due to
the switch. To study this simulation further, in the lower plot we plot two residuals:
the first is calculated in the region [0, tswitch ] and the second in [tswitch , Tobs ]. Because
the switch does not occur in either of these periods we can resolve the free precession
during each period and note that the precession modulation is in fact smaller after the
switch. This reduction in the size of modulations is because the precession is due to the
Chapter 4 Modelling observations of precessing pulsars 91

Phase residual [cycles]

3
2
1
0
−1
−2
−3
0 τp 2τp 3τp 4τp 5τp 6τp 7τp 8τp 9τp 10τp 11τp
time
×10−5
Phase residual [cycles]

4
3
2
1
0
−1
−2
−3
−4
0 τp 2τp 3τp 4τp 5τp 6τp 7τp 8τp 9τp 10τp 11τp
time

Figure 4.16: Phase residuals for a simulation with a single switch in the spin-
down torque with the parameters given in Eqn. (4.87) and the simulation prop-
erties listed in Table 4.5. In the top plot we show the residual calculated using
the whole observation time. The bottom plot is the separate residuals calculated
in the region before and after the switch.

spin-down torque wobbling the angular moment vector and being amplified by the EM
torque. After the switch the spin-down torque, and hence AEM , is reduced by a factor
SS and therefore the size of the modulation is similarly reduced.

4.7.2 Switching in the spin-down and anomalous torque

We now consider manually switching both the spin-down and anomalous torque halfway
though the simulation. That is we set

Tobs
tswitch = , SS = SA = 0.01 (4.88)
2

In a similar fashion to Figure 4.16, we first show the total residual in the top plot of
Figure 4.17, and then the individual residuals in the lower plot. Again, the overall
phase residual exhibits periodic modulation with a period of the observation time Tobs
resulting from the failing of the fitting algorithm to find an appropriate fit. In contrast
to the spin-down only switching, the amount of precession when considered before and
after the glitch now increases.

To understand this, recall that we begin with a minimal precession state, where θ = β
and the precession results from effect of the spin-down torque. After the switch, we have
changed the size of the anomalous torque and hence we have modified the effective body
92 Chapter 4 Modelling observations of precessing pulsars

frame and the angle β. This means that after the switch the star is no longer in a minimal
precession configuration. This generates a significantly larger wobble angle producing
a significant increase in the phase residuals fitted in the post-switch period. The effect
is not observable when fitting to the entire simulation period since the switching event
remains dominant. In the next section we will calculate this new wobble angle.
Phase residual [cycles]

0.08
0.06
0.04
0.02
0.00
−0.02
−0.04
−0.06
−0.08
0 τp 2τp 3τp 4τp 5τp 6τp 7τp 8τp 9τp 10τp 11τp
time
×10−4
Phase residual [cycles]

1.5
1.0
0.5
0.0
−0.5
−1.0
−1.5
0 τp 2τp 3τp 4τp 5τp 6τp 7τp 8τp 9τp 10τp 11τp
time

Figure 4.17: Phase residuals for a simulation with a single switch in the spin-
down and anomalous torque with the parameters given in Eqn. (4.88) and the
simulation parameters listed in Table 4.5. In the top plot we show the residual
calculated using the whole observation time. The bottom plot is the separate
residuals calculated in the region before and after the switch.

4.7.2.1 Calculating the new wobble angle after a switch

We now estimate the change in wobble angle after switching a fraction of the anoma-
lous torque. This will be useful in making testable predictions for anomalous torque
switching.

The two-state switching changes the value of A according to Eqn. (4.84), which in turn
redefined the effective rotating frame as defined in Section 3.4.1. A non-precessing star
at an angle β(I , A , χ) will, after an anomalous torque switch by a fractional amount
SA , no longer be aligned with the rotating frame axis. This is because the effective
rotating frame will have shifted to β 0 = β(I , SA A , χ). As a result, we should expect the
previously non-precessing star to begin precessing after a torque switching event. The
exact size of the new precession wobble angle will depend on the phase of precession (if
it was precessing) at the instant before the switching event. The exact change in the
wobble angle will depend on the precessional phase at the time of the glitch. We can
Chapter 4 Modelling observations of precessing pulsars 93

approximate it in a crude way by considering

∆β(I , A , χ, SA ) ∼ |β − β 0 |. (4.89)

The expression for ∆β is not easily amenable to analytic calculation, but can easily be
explored graphically. In Figure 4.18 we plot Eqn. (4.89) by calculating β and β 0 values
using Eqn. (3.37); this is done for several choices of SA to show the typical variations.
This illustrates that the new precession angle after a switch can be as much as a few
degrees although it tends to zero in the limit I  A and I  A .

12
SA = 1
5
∆β = |β 0 − β| [degrees]

10 SA = 1
2
SA = 4
5
8

0
10−2 10−1 100 101 102
A
I

Figure 4.18: Illustrating the magnitude of the precession angle after switching
due to the new rotation of the effective rotating frame. We plot the half-angle
(∆β) of the precession cone as a function of the ratio A /I . Typically we expect
real stars to have A < I .

A simple application of this work is to apply our findings to PSR B1828-11, which was
interpreted by Lyne et al. [111] as undergoing switching with a fractional change in the
spin-down rate given by

∆ν̇
= 0.007. (4.90)
ν̇
Manipulating Eqn. (4.85), we see that

∆ν̇
|SS | = . (4.91)
ν̇
Using data from the ATNF catalogue [117], PSR B1828-11 has a frequency ν = 2.47 Hz,
a spin-down rate ν̇ = −3.65 × 10−13 Hz/s. Rearranging Eqn. (4.83) we then have
c
B1828−11
A = ν̇ ≈ 3 × 10−11 . (4.92)
8πRν 3

Now in this interpretation, I is unconstrained for PSR B1828-11 since the periodic mod-
ulations are assumed to be resulting from switching and not precession. Nevertheless,
94 Chapter 4 Modelling observations of precessing pulsars

we numerically maximise ∆β over I to find

∆βmax (A = 2.89 × 10−11 , I = 2.87 × 10−11 ) = 0.05 degs. (4.93)

This is the maximum new wobble angle which must occur every time PSR B1828-11
undergoes a switching event, if the switch also occurs in the anomalous torque. This
therefore provides a method to probe how the switching mechanism works by looking
at timing residuals in the post-switch timing data to check for any signs of precession.

4.8 Conclusions

In this chapter, we have developed a method to simulate observable properties of neutron

stars by evolving the governing system of ODEs. These equations, given in Eqn. (4.7) to
Eqn. (4.12), contain both the components of the spin-vector in the rotating frame and
the Euler rotation angles required to transform from the rotating frame to the inertial
frame. This step is important since it is in this reference frame that observers view the
effects of precession on the pulsar’s observation features.

We developed an intuitive model for how the magnetic dipole, along which EM radiation
is emitted, moves in the inertial frame. Comparing with analytic models where possible,
we showed how to calculate from the motion of the dipole the evolution of the star’s tim-
ing properties (the phase, frequency, and spin-down rate) and how to calculate features
of the beam shape such as the beam-width. In doing this, we derived the spin-down rate
for a precessing pulsar (see Section 4.5.1), which will be used in Chapter 5, and verified
it against numerical solutions.

Finally, we used the phase residuals to investigate the effect of a simple torque switching
model in which the components of the torque instantaneously change. This was done
to model the magnetospheric switching which is proposed by Lyne et al. [111] as an
explanation of the periodic modulation of PSR B1828-11. We present preliminary results
showing the complicated interactions which occur in single switching events; in the future
we would like to extend this to having the precession determine when the switching
occurs. Our numerical model is ideally suited to this task as it captures the complicated
feedback between precession and the changing torque. Furthermore, we would eventually
like to link the switching to the precessional phase in a probabilistic way as proposed
by Jones [89]. In this way, precession provides the clock to the switching. By virtue of
being probabilistic, such a system will display stochastic resonance as first described by
Cordes [45].
Chapter 5

Comparing models of the periodic

variations in spin-down and
beam-width for PSR B1828-11

In this thesis, we endeavour to understand the cause of timing variations in pulsars

with the aim being that in doing so, we can understand more about the physics of
neutron stars. In Chapter 2 we discussed a variety of models to explain timing noise
including two, precession and magnetospheric switching, which apply to pulsars with
periodic modulations. Such pulsars are particularly interesting given the conclusion
of Hobbs et al. [81] that, as more data is collected for pulsars, more quasi-periodic
features are observed. Therefore, it is hoped that understanding what causes the periodic
modulations may eventually lead to a more complete understanding of timing noise
across the population.

In Chapter 3 and Chapter 4, we investigated precession and its effects on the observed
features of pulsars. One pulsar, PSR B1828-11, provides strong evidence for precession
[156], but has also been put forth as a candidate for magnetospheric switching [111].
There is no decisive feature in the data on PSR B1828-11 which rules out either model.
Therefore, in this chapter we will apply a Bayesian model comparison to determine
quantitatively, which model best fits the data. In doing this, we will fully develop both
the precession model and the switching model into complete predictive models for the
spin-down rate and beam-width for which we have observational data. This chapter
contains material reprinted with permission from Ashton et al. [25] as follows: Ashton,
G.; Jones, D. I.; Prix, R. Volume 458, Issue 1, p.881-899 (2016). Copyright (2016) by
the Monthly Notices of the Royal Astronomical Society.

5.1 Introduction

The pulsar PSR B1828-11 demonstrates periodic variability in its pulse timing and beam
shape at harmonically related periods of 250, 500, and 1000 days. The modulations in

95
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
96 for PSR B1828-11

the timing was first taken as evidence that the pulsar is orbited by a system of planets
by Bailes et al. [27]. A more complete analysis by Stairs et al. [156] concluded that the
corresponding changes in the beam-shape would require at least two of the planets to
interact with the magnetosphere, which does not seem credible. Instead, the authors
proposed that the correlation between timing data and beam-shape suggested the pulsar
was undergoing free precession. If true, such a claim would require rethinking of the
vortex-pinning model used to explain the pulsar glitches since the pinning should lead to
much shorter modulation period than observed [152], and fast damping of the modulation
[107].

The idea of precession for PSR B1828-11 has been studied extensively in the literature:
Jones and Andersson [90] derived the observable modulations due to precession and
noted that the electromagnetic spin-down torque will amplify these modulations (see
Section 4.4.2 for more details). Link and Epstein [108] fitted a torqued-precession model
to the spin-down and beam-shape, followed by Akgün et al. [14] where a variety of shapes
and the form of the spin-down torque were tested. All of these authors agree that pre-
cession is a credible candidate to explain the observed periodic variations. Furthermore,
as we found in Section 4.5.2, to explain the double-peaked spin-down modulations, the
so-called wobble angle must be small while the angle between the symmetry axis of the
biaxial moment of inertia tensor and the magnetic dipole must be close to π/2.

More recently Arzamasskiy et al. [22] updated the previous estimates (based on a vacuum
approximation) to a plasma filled magnetosphere. They also find that the magnetic
dipole and spin-vector must be close to orthogonal, but solutions could exist where it
is the wobble angle which is close to π/2 while the magnetic dipole lies close to the
angular momentum vector; we will not consider such a model here, but note it is a valid
alternative which deserves testing.

The distinctive spin-down of PSR B1828-11 was analysed by Seymour and Lorimer [151]
for evidence of chaotic behaviour. They found evidence that PSR B1828-11 was subject
to three dynamic equations with the spin-down rate being one governing variable. This
further motivates the precession model since it results from applying Euler’s three rigid-
body equations to a non-spherical body [100].

The precession hypothesis was challenged by Lyne et al. [111] when reanalysing the data.
They noted that measuring the spin-down and beam-shape with any accuracy requires
time averaging over periods ∼ 100 days, which smooths out any behaviour acting on
this time-scale. Motivated by the intermittent pulsar PSR B1931+24, they put forward
the phenomenological hypothesis that instead the magnetosphere is undergoing periodic
switching between (at least) two metastable states. Such switching would result in
correlated changes in the beam-width and spin-down rate. They returned to the data
and instead of studying a time-averaged beam-shape-parameter as done by Stairs et al.
[156], they instead considered the beam-width at 10% of the observed maximum W10 .
This quantity is time-averaged, but only for each observation lasting ∼ 1hr. This makes
W10 insensitive to any changes which occur on timescales shorter than an hour. If
the meta-stable states last longer than this, W10 will be able to resolve the switching.
The relevant data was kindly supplied to us courtesy of Andrew Lyne [111], and is
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 97

reproduced in Figure 5.1. From these observations, the authors concluded that the
individual measurements of W10 for PSR B1828-11 did in fact appear to switch between
distinct high and low values, as opposed to a smooth modulation between the values,
with this switching coinciding with the periodic changes in the spin-down. On this basis,
they interpret the modulations of PSR B1828-11 as evidence it is undergoing periodic
switching between two magnetospheric states. When studying another pulsar which
also displays double-peaked spin-down modulations, Perera et al. [134] extended the
switching model, as discussed in Section 5.3.2, to be capable of producing the double-
peak spin-down rate; it is this modification of the switching model which we will be
comparing with precession.

−364.0
ν̇ [10−15 Hz s−1 ]

−364.5 A
−365.0
−365.5
−366.0
−366.5
14
12 B
W10 [ms]

10
8
6

50000 51000 52000 53000 54000

MJD

Figure 5.1: Observed data for PSR B1828-11 spanning from MJD 49710 to
MJD 54980. In panel A we reproduce the spin-down rate with error-bars and
in panel B the beam-width W10 (for which no error bars were available). All
data courtesy of Lyne et al. [111].

In our view, it is not immediately clear by eye whether the data presented in Figure 5.1 is
sufficient to rule out or even favour either of the precession or switching interpretations.
For this reason, we develop a framework in which to evaluate models built from these
concepts and argue their merits quantitatively using a Bayesian model comparison. We
note that a distinction must be made between a conceptual idea, such as precession, and
a particular predictive model built from it. As we will see, each concept can generate
multiple models, and furthermore we could imagine using a combination of precession
and switching, with the precession acting as the ‘clock’ that modulates the probability
of the magnetosphere being in one state or the other, an idea developed by Jones [89].
The models considered here cover the precession and switching interpretations, but we
do not claim the models to be the ‘best’ that these hypotheses could produce.

We note that there also exists high-resolution observations of the shape parameter of
PSR B1828-11 which are presented in Stairs et al. [157] and Lyne [109]. The shape pa-
rameter is defined to be the proportion of time spent in one of two modes as determined
from the integrated pulse profile. These observations show that PSR B1828-11 appears
to ‘switch’ between the modes on short timescales (less than ∼ 8 hrs). This, it seems,
provides evidence for any model where the switching occurs on short timescales, but is
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
98 for PSR B1828-11

biased by a longer timescale of ∼ 500 days. The effect of precession on the shape param-
eter was investigated by Akgün et al. [14] who demonstrated how a ‘patchy’ emmission
region could produce such observations under a pure precession hypothesis. Alterna-
tively, precession could in fact be the clock biasing the switching as in the hybrid model
of Jones [89]. For the switching model, this observation also complicates the story since
it requires stability over two quite different timescales. In this chapter, we will focus
only on the long timescale variations shown in Fig. 5.1 and how these can be interpreted
under a deterministic switching model or precession; in effect assuming that the short
timescale variations are unrelated to the models under investigation. In the future, we
would like to include these data sets and see how this changes the picture.

The rest of the chapter is organised as follows: in Section 5.2 we will describe the
framework to fit and evaluate a given model, in Section 5.3 we will define and fit several
predictive models from the conceptual ideas, and then in Section 5.4 we shall tabulate
the results of the model comparison. Finally, the results are discussed in Section 5.5.

5.2 Bayesian Methodology

We now introduce a general methodology to compare and evaluate models for this form
of data. The technique is well practised in this and other fields and so in this section
we intend only to give a brief overview; for a more complete introduction to this subject
see [155, 86, 64].

5.2.1 The odds-ratio and posterior probabilities

There are two issues that we wish to address. Firstly, given two models, how can one
say which is preferred, and by what margin? Secondly, assuming a given model, what
can be said of the probability distribution of the parameters that appear in that model?

We can address the first issue by making use of Bayes theorem for the probability of
model Mi given some data:

P (Mi )
P (Mi |data) = P (data|Mi ) . (5.1)
P (data)

The quantity P (data|Mi ) is known as the marginal likelihood of model Mi given the
data.

In general, we cannot compute the probability given in equation (5.1) because we do

not have an exhaustive set of models to calculate P (data). However, we can compare
two models, say A and B, by calculation of their odds-ratio:

P (MA |data) P (data|MA ) P (MA )

O= = . (5.2)
P (MB |data) P (data|MB ) P (MB )

In the rightmost expression, the first factor is the ratio of the marginal likelihoods (also
known as the Bayes factor ) which we will discuss shortly, while the final factor reflects
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 99

our prior belief in the two models. If no strong preference exists for one over the other,
we may take a non-informative approach and set this equal to unity. We will follow this
approach in what follows below.

We need to find a way of computing the marginal likelihoods, P (data|Mi ). To this end,
consider a single model Mi with model parameters θ, and define P (data|θ, Mi ) as the
likelihood function and P (θ|Mi ) as the prior distribution for the model parameters. We
can then perform the necessary calculations by making use of
Z
P (data|Mi ) = P (data|θ, Mi )P (θ|Mi )dθ. (5.3)

The likelihood function can also be used to explore the second issue of interest, by
calculating the joint-probability distribution for the model parameters, also known as
the posterior probability distribution:

P (data|θ, Mi )P (θ|Mi )
P (θ|data, Mi ) = . (5.4)
P (data|Mi )

Note that the marginal likelihood P (data|Mi ) described above plays the role of a nor-
malising factor in this equation.

In general, the integrand of Eqn. (5.3) makes analytic, or even simple numeric integration
difficult or impossible. This is the case for the probability model that we will use and so
instead we must turn to sophisticated numerical methods. For this study we use Markov-
Chain Monte-Carlo (MCMC) techniques which simulate the joint-posterior distribution
for the model parameters up to the normalising constant

P (θ|data, Mi ) ∝ P (data|θ, Mi )P (θ|Mi ). (5.5)

In particular we will use the Foreman-Mackey et al. [62] implementation of the affine-
invariant MCMC sampler [71] to approximate the posterior density of the model param-
eters. Further details of our MCMC calculations can be found in Appendix 5.A.

Once we are satisfied that we have a good approximation for the joint-posterior density
of the model parameters we discuss how to recover the normalising constant to calculate
the odds-ratio in Section 5.4.

5.2.2 Signals in noise

We now need to build a statistical model to relate physical models for the spin-down
and beam-width to the data observed in Figure 5.1. To do this we will turn to a method
widely used to search for deterministic signals in noise.

We assume our observed data y obs is a sum of a stationary zero-mean Gaussian noise
process n(t, σ) (here σ is the standard deviation of the noise process) and a signal model
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
100 for PSR B1828-11

f (t|Mj , θ) (where θ is a vector of the model parameters) such that

y obs (ti |Mj , θ, σ) = f (ti |Mj , θ) + n(ti , σ). (5.6)

Given a particular signal model, subtracting the model from the data should, if the
model and model parameters are correct, leave behind a Gaussian distributed residual -
the noise. That is
y obs (ti |Mj , θ, σ) − f (ti |Mj , θ) ∼ N (0, σ). (5.7)

The data, for either the spin-down or beam-width, consists of N observations (yiobs , ti ).
For a single one of these observations, the probability distribution given the model and
model parameters is
( )
obs 1 − (f (ti |Mj , θ) − yi )2
P (yi |Mj , θ, σ) = √ exp . (5.8)
σ 2π 2σ 2

The likelihood is the product of the N probabilities

N
Y
obs
P (y |Mj , θ, σ) = P (yiobs |Mj , θ, σ), (5.9)
i=1

where yobs denotes the vector of all the observed data.

In Section 5.3 we will define the physical models, f (t|Mj , θ), for the precession and
switching interpretations; for now we recognise that once defined, we may calculate the
likelihood of the data under the model using Eqn. (5.9).

5.2.3 Choosing prior distributions

In the previous section we have developed the likelihood function P (data|θ, Mj ) for
any arbitrary model producing a deterministic signal f (ti |Mj , θ) in noise. To compare
between particular models, using Eqn. (5.2), we must compute the marginal likelihood
as defined in Eqn. (5.3) which requires a prior distribution P (θ|Mj ).

The choice of prior distribution is important in a model comparison since it can poten-
tially have a large impact on the resulting odds-ratio. We want to use astrophysically
informed priors wherever possible, or suitable uninformative (but proper) priors oth-
erwise. However, the switching model presents a particular challenge in this respect,
as its switching parameters (cf. Section 5.3.2) are ad-hoc and purely phenomenological,
and were initially informed by the same data we are trying to test the models on. It is
therefore important to avoid potential circularity in properly assessing the prior volume
of its parameter space, which affects the relevant ‘Occam factor’ for this model (e.g. see
MacKay [116]).

To resolve this, we will make use of the availability of two different and independent
data sets: the spin-down and the beam-width data. First, we will perform parameter
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 101

estimation using the spin-down data with astrophysical priors where possible and uni-
form priors based on crude estimates from the data otherwise. For the model parameters
common to both the spin-down and beam-width models, we will use the posterior distri-
butions from the spin-down data as prior distributions for the beam-width model. For
the remaining beam-width parameters which are not common to both the spin-down and
beam-width models we will use astrophysically-motivated priors. In this way we can do
model comparisons based on the beam-width data using proper, physically motivated
priors. In addition, this enforces consistency between the beam-width and spin-down
solutions: for example constraining the two to be in phase.

An obvious alternative is to do the reverse and use the beam-width data to determine pri-
ors for the spin-down data. However, for both models, we found difficulties in obtaining
good quality posteriors when conditioning on the beam-width data with uniform priors
based on crude estimates. Specifically, we found the posteriors to be non-Gaussian and
multimodal. To deal with this we would need to use a more sophisticated methodology
than that discussed in Appendix 5.A. By contrast this is not the case when conditioning
on the spin-down data first (results presented in Section 5.3). This is expected since,
even by eye, we see that the spin-down data contains an easily visible ‘signal’, while the
beam-width data is relatively ‘noisy’. For this work we are primarily interested in laying
out the framework to perform model comparisons and either method should suffice and
give the same solution. For now then, we will use the more straight-forward method of
using the spin-down data to set priors for the beam-width.

5.3 Defining and fitting the models

In this section we will take each conceptual idea (precession or switching) and define a
predictive signal model f (t|Mj , θ). Each concept may motivate multiple signal models:
already we have seen the extension to the original Lyne et al. [111] switching model by
Perera et al. [134]. In this work we do not aim to exhaust all known models and are well
aware that more models exist that have not yet been considered.

For each concept, we will first discuss the theoretical model, then discuss the choice of
priors and finally the resulting posterior and posterior-predictive checks. For both these
concepts we build models for both the spin-down and beam-width using the former to
inform the priors for the latter as described in Section 5.2.3. Model comparisons will
be made on the beam-width data only. In addition to these two concepts, we will also
consider a noise-only model for the beam-width data.

It is worth stating that by using the signals-in-noise statistical model, we do not make
any assumptions on the cause of the noise other than requiring it to be stationary
and Gaussian (cf. Jaynes [86]). Given the uncertain physics of neutron stars and the
measurement of pulses, it seems likely the noise will contain contributions both from
the neutron star itself, and from the measurement process, with the former dominating.
We will add a subscript to the noise component σ[ν̇,W10 ] to distinguish between the two
data sources.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
102 for PSR B1828-11

5.3.1 Noise-only model

5.3.1.1 Defining the noise-only beam-width model

Before evaluating the precession and switching hypothesis, let us first consider a noise-
only model. This will introduce some generic concepts and provide a benchmark against
which to test other models. The noise-only model asserts that the beam-width data
(as seen in panel B of Figure 5.1) does not contain any periodic modulation, but is the
result of noise about a fixed beam-width: the signal model f (t) = W10 is a constant.

We will not consider the spin-down data under such a hypothesis since it is the beam-
width data alone that we will use to make model comparisons and it is clear by eye that
such a model is incorrect.

5.3.1.2 Fitting the model to the beam-width data

For the noise-only model we have two parameters which require a prior: the constant
beam-width W10 and the noise σW10 . For the beam-width we will set a prior using
astrophysical data on the period P from the ATNF database (available at www.atnf.
csiro.au/people/pulsar/psrcat, for a description see [117]). This value, P ATNF =
0.405043321630±1.2×10−11 s, provides a strict upper bound on W10 , although typically
integrated pulse profiles only occupy between 2% and 10% of the period [112]. Therefore,
we will use a uniform prior on [0, 0.1P ATNF ] for hW10 i. The choice of 10% adds a degree
of ambiguity into the model comparison since varying it will change the odds-ratio; we
investigate this in Section 5.4.3.

For the noise parameter σW10 we will use a prior Unif(0, 5) ms based on a crude estimate
from the data. We must be careful here as by doing this we are in a sense using the data
twice, but this will not introduce bias into the model comparison provided the same
prior is applied for all beam-width models.

The MCMC simulations converge quickly to a normal distribution as shown in Figure 5.2.
Of note is the mode of σW10 ∼ 2 ms; this is the Gaussian noise required to explain the
variations in W10 about a fixed mean. For other models, we hope to explain some of
the variations with periodic modulation and the rest with Gaussian noise. So for these
models we should expect σW10 < 2 ms.

In Figure 5.3 we plot the maximum posterior estimate (MPE) of the signal alongside the
data, i.e. the model prediction when the parameters are set equal to the peak values of
the posterior probability distributions. This figure demonstrates that, for the noise-only
model, the observed W10 has a mean value of approximately 8 ms, and all the variations
about this mean are due to the noise. In the following section we will develop models
where at least some the variation is explained by periodic modulations.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 103

Figure 5.2: The estimated marginal posterior probability distributions for the
noise-only model parameters of the beam-width data.

5.3.2 Switching model

The switching idea is phenomenological and we will build the model based on the mod-
ification of Lyne et al. [111] by Perera et al. [134]: that is we assume the magnetosphere
switches between two meta-stable states twice during a single period (the motivation for
this was discussed in Section 2.3.3 and we will further motivate it in Section 5.3.2.1).
For this work we will assume the switching to be deterministic, although improvements
could be made by allowing the switching time to dither, or probabilistic variations in the
switching states themselves; see Lyne et al. [111] for some exploration of such ideas. This
fully deterministic model captures the primary features without explaining the underly-
ing physics, for example the cause of the switching. Both Jones [89] and Cordes [45] have
worked to improve the physical motivations for the switching and provide a consistent
picture. Nevertheless, in this work we choose to use the simple phenomenological model
as a basis, which can be improved upon in future work.

5.3.2.1 Defining the spin-down rate model

The model proposed by Lyne et al. [111] poses two states for the magnetosphere which
we will label as S1 and S2 . Then associated to each of these states is a corresponding
spin-down rate ν̇1 and ν̇2 . The smoothly varying spin-down that we observe is a result
of the time-averaging process required to measure the spin-down rate. Lyne et al. [111]
suggested a square-wave-like switching with a duty rate measuring the fraction of time
spent in one state compared to the other. They also proposed a dither in the switching
period which will obscure the periodicity, and may give rise to low-frequency structure;
we will not consider the dither in this work, but will investigate it in future work. While
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
104 for PSR B1828-11

Figure 5.3: Posterior predictive check of the fit of the noise-only model posterior
distribution to the data: the solid black line is the maximum posterior estimate
(MPE), i.e. the model prediction when the parameters are set equal to the
values corresponding to the peaks of the posterior probability distributions.
The shaded region indicates the MPE of the 1σW10 noise about the beam-width
model. Black dots are the original data.

studying PSR B0919+06, which also demonstrates a double-peaked spin-down rate like
PSR B1828-11, the authors of Perera et al. [134] realised that a (deterministic) switching
model which flips once per cycle is incapable of explaining the double-peak observed in
the spin-down rate (in particular that one peak is systematically smaller than the other).
In order to explain this double-peaked structure, they propose that the mode-changes
responsible for switching in the spin-down rate must be doubly-periodic: that is the
spin-down rate changes state twice during a single cycle. Other modifications, such as
introducing a third magnetospheric state, are possible, but in this work we will apply
the Perera et al. [134] switching model to PSR B1828-11.

We now discuss the particular formulation of this model used in this analysis, firstly
defining the underlying spin-down model and then the time-averaging process. To aid
in this discussion we plot both the underlying spin-down model and time-average in
Figure 5.4 and gradually introduce each feature.

We begin by defining t̃ = (t − tref ) + φ0 T mod (T ) where φ0 ∈ [0, 1] is an arbitrary

phase offset and tref is a reference time. For all models in this study we will set
tref = MJD 49621 to coincide with the epoch at which the ATNF database Manch-
ester and Lyne [118] records measurements for PSR B1828-11. Then the function which
generates the switching is
(
ν̇1 if 0 < t̃ < t1 or t2 < t̃ < t3
ν̇u (t) = , (5.10)
ν̇2 if t1 < t̃ < t2 or t3 < t̃ < T

where the subscript ‘u’ denotes that this is the underlying spin-down model.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 105

T
tA tB tC tD
ν̇1

ν̇2

0 t1 t2 t3 T
t̃

Figure 5.4: Schematic of the doubly-periodic spin-down rate model proposed

by Perera et al. [134]. The solid line is the underlying spin-down evolution
while the dashed line indicates the measured time-averaged quantity. In this
instance, the time-average window is longer than tC , but shorter than the other
three durations.

There are multiple ways to parameterise the switching times in the model. For the data
analysis, we have chosen to parameterise by the total cycle duration T and three of the
segment durations tA , tB and tC .

This model is subject to label-switching degeneracy in the choice of ν̇1 and ν̇2 and
also between the various timescales and initial phase φ0 . This degeneracy may cause
difficulties in the MCMC search algorithm, and we therefore fix this gauge freedom by
specifying that |ν̇1 | < |ν̇2 |, tA ≥ tC , and we require that tA refers to a segment where
ν̇ = ν̇1 .

Based on a cursory inspection of the observed PSR B1828-11 spin-down rate (see Fig-
ure 5.1 panel A), it is clear that the secular second-order spin-down rate is non-zero. To
model this we will include a constant ν̈ in the underlying spin-down model

ν̇(t) = ν̇u (t) + ν̈(t − tref ) . (5.11)

This gives the intrinsic spin-down rate of the pulsar which would be observed if mea-
surements could be taken without time-averaging

To simulate the observed spin-down rate, we could time-average Eqn. (5.11) directly. In-
stead, we choose to mimic the data collection process responsible for the time-averaging.
Let us first discuss the data collection process as described by Lyne et al. [111] (this has
already been described in other parts of this thesis as the observer-method of calculating
the spin-down rate). Observers start with the time-of-arrival of pulsations, which is a
measure of the pulsar rotational phase. Taking a 100 day window of data, starting at
the earliest observation, a second-order Taylor expansion in the phase is fitted to the
data yielding a measurement of ν̇. Then the processes is repeated, sliding the window
in intervals of 25 days over the whole data set. The measured ν̇ values at the centre of
each window gives a time-averaged spin-down rate.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
106 for PSR B1828-11

To mimic this data collection process, we first integrate Eqn. (5.11) twice to generate the
phase and then repeat the above process. When integrating, we can ignore the arbitrary
phase and frequency offsets since we discard them when calculating the spin-down rate.
The resulting spin-down rates constitutes our signal model which is the time-average of
Eqn. (5.11). A schematic representation of the sort of spin-down that is then found is
given by the dotted curve in Figure 5.4. Clearly, the time-averaged spin-down is much
smoother than the underlying spin-down.

It is worth taking a moment to realise that the relation of the time-average spin-down
to the underlying model ν̇ depends on both the segment durations and the length of
the time average (tave = 100 days). For the ith segment, if the duration ti > tave then
the time-averaged spin-down will ‘saturate’ and have a flat spot as in segment A of the
illustration in Figure 5.4. On the other hand, if ti < tave then the maximum spin-down
rate in this segment will be a weighted sum of the two underlying spin-down rates as in
segment C of the illustration. The weighting is determined by the amount of time the
underlying spin-down rate spends in each state during the time-average window.

5.3.2.2 Parameter estimation for the spin-down

In Table 5.1 we list the selected priors. For ν̈ we define ν̈ ATNF = (8.75±0.09)×10−25 s−3
(the value from the ATNF catalogue) and use a normal prior with this mean and standard
deviation. In the tables we show the difference with respect to this value. For the
remaining parameters we select uniform priors using crude estimates of the data in panel
A of Figure 5.1. As previously mentioned, this means we are using the data twice: once
in setting up the priors and once for the fitting. Therefore, it would be inappropriate to
use the results in a model comparison and this is not our intention: we want to use the
posterior distribution as a prior for the beam-width parameter estimation.

Parameter Distribution Units

ν̇1 Unif(-3.66 × 10-13 , -3.64 × 10-13 ) s-2
ν̇2 Unif(-3.67 × 10-13 , -3.66 × 10-13 ) s-2
ν̈-ν̈ ATNF N (0, 9.0 × 10-27 ) s-3
T Unif(450, 550) days
tA Unif(0, 250) days
tB Unif(0, 250) days
tC Unif(0, 250) days
φ0 Unif(0, 1)
σν̇ Unif(0, 1 × 10-15 ) s-2

Table 5.1: Prior distributions for the spin-down switching model.

For the spin-down data under the switching model, the MCMC simulations converge
quickly to unimodal and approximately Gaussian distributions. The distributions are
plotted in Figure 5.5A, and we summarise the results by their mean and standard-
deviation in Table 5.2. In the case of parameters with uniform priors, this indicates
that the data is informative and a well defined ‘best-fit’ has been selected. For ν̈ the
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 107

Figure 5.5: A: The estimated marginal posterior probability distribution for

the Switching spin-down model parameters. B: Checking the fit of the model
using the maximum posterior values to the data; see Figure 5.3 for a complete
description.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
108 for PSR B1828-11

Parameter Mean ± s.d. Units

ν̇1 -3.6489 × 10-13 ±6.33 × 10-17 s-2
ν̇2 -3.6635 × 10-13 ±4.44 × 10-17 s-2
ν̈-ν̈ ATNF -3.1051 × 10-28 ±9.0 × 10-27 s-3
T 485.52±0.8649 days
tA 157.75±7.6587 days
tB 159.71±11.7798 days
tC 15.1379±4.3925 days
φ0 0.5278±0.0143
σν̇ 4.0932 × 10-16 ±1.84 × 10-17 s-2

Table 5.2: Posterior estimates for the spin-down switching model.

posterior has not departed significantly from the (informative) prior meaning that the
data agrees with the prior.

To check that our fit is sensible, we plot the observed spin-down data in Figure 5.5B
alongside the maximum posterior estimate for the signal. The relative size of the noise
component informs us how well the model fits the data: if σν̇ is of a similar size to the
variations in spin-down rate, then the model does poorly and we require a large noise
component. In this case the noise-component is smaller than the variations in spin-down
rate and the signal model explains most of the variations in the data.

Comparing the maximum posterior values of the four segment times to the baseline on
which we time-average (fixed at 100 days) can give an insight into how the model has
best fit the data. If we take the posterior of tC , we find it has a mean value of ∼ 15 days
which is significantly shorter than the baseline on which we time-average. For the other
three segments, their durations are longer than this baseline. The reason that the fit in
Figure 5.5B has one maxima smaller than the other, is because the segment duration for
that segment, tC , is shorter than the time-average baseline. This is expected and was
precisely the motivation for using the model proposed by Perera et al. [134]; a switching
model split into only two segments could not produce this feature.

5.3.2.3 Defining the beam-width model

For the beam-width, we assume that changes in the spin-down rate directly correlate to
changes in this beam-width through changes in the beam geometry. Since we require
the switching to be doubly-periodic for the spin-down to make sense, so we must require
the beam-width to be doubly periodic. That is we define the beam-width model to be
(
W1 if 0 < t̃ < t1 or t2 < t̃ < t3
W (t) . (5.12)
W2 if t1 < t̃ < t2 or t3 < t̃ < T

Lyne et al. [111] noted that the larger beam-widths tended to correlate with the lower
(absolute) spin-down rate (ν̇1 in our model). We could fix this by requiring that W1 > W2
(recalling that we set |ν̇1 | < |ν̇2 |), but instead we will not implement such a constraint
and allow the data to decide. As with the spin-down, to break the degeneracy in the
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 109

times we will require again that tA ≥ tC . We will not assume any secular changes in the
beam-width for simplicity.

5.3.2.4 Parameter estimation for the beam-width

Having obtained a sensible fit to the spin-down data, we use the resulting posteriors (as
summarised in Table 5.2) to inform our priors for the beam-width data. We can do this
only for those parameters common to both the beam-width and spin-down predictions
of the switching model: namely the four timescales, and the phase-offset. We would like
to relate the spin-down rates ν̇1 and ν̇2 to the beam-widths. However, the underlying
physics is not understood, and so instead we will take a naive approach and set a prior
on the beam-widths from astrophysical data.

For the beam-widths, W1 and W2 , we will use the same prior as defined in Section 5.3.1.2
for the noise-only model: namely a uniform prior on [0, 0.1P ATNF ] covering 10% of the
spin-period. Using such a prior introduces some ambiguity into the model comparison
as the result could be ‘tuned’ by varying the fraction f of the spin-period used to set
the uniform prior limits (here f = 0.1). This issue is addressed in Section 5.4.3 where
we find all sensible choices of f lead to the same overall conclusion.

The final parameter which requires a prior distribution is the noise-component: as de-
scribed in Section 5.3.1 we apply a prior to σW10 using a crude estimate from the data;
this is tabulated along with the other priors in Table 5.3.

Parameter Distribution Units

W1 Unif(0, 40.5000) ms
W2 Unif(0, 40.5000) ms
T∗ N (485.5, 0.8649) days
tA ∗ N (158.0, 7.6587) days
tB ∗ N (160.0, 11.7798) days
tC ∗ N (15.1379, 4.3925) days
φ0 ∗ N (0.5278, 0.0143)
σW10 Unif(0, 5) ms

Table 5.3: Prior distributions for the beam-width switching model. Parameters
for which the prior is taken from spin-down posteriors are labelled by ∗ .

We plot the posterior estimate in Figure 5.6A which demonstrates non-Gaussianity and
multimodal features in the segment times and the phase. This indicates the existence of
multiple solutions which could explain the data. We note that the noise component σW10
has a mode at 1.6 ms which is less than the 2ms required in the noise-only model. This
indicates that some of the variability is being explained by the signal model. In Table 5.4
we summarise the posterior. We find that the posterior modes satisfy W1 > W2 : larger
beam-widths are associated with the smaller absolute spin-down rates as found by Lyne
et al. [111].

Again we check the predictive power of our estimated posterior by plotting the MPE
alongside the data in Figure 5.6B. The fit to the data is not as good as the spin-down
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
110 for PSR B1828-11

Figure 5.6: A: The estimated marginal posterior probability distribution for

the Switching beam-width model parameters. B: Checking the fit of the model
using the maximum posterior values to the data; see Figure 5.3 for a complete
description.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 111

Parameter Mean ± s.d. Units

W1 9.5166±0.0956 ms
W2 7.2327±0.0830 ms
T 485.04±0.7286 days
tA 150.98±4.3909 days
tB 155.29±3.0598 days
tC 16.4134±4.5771 days
φ0 0.5409±8.38 × 10-3
σW10 1.5964±0.0427 ms

Table 5.4: Posterior estimates for the beam-width switching model.

fit: by eye it is clear that most data points lie away from the signal model requiring a
greater (relative) level of noise.

5.3.3 Precession model

We will now define the precession model and its predictions for the expected signal in
the spin-down and beam-width data.

Classical free precession refers to the rotation of a rigid non-spherical body when there
is a misalignment between its spin axis and angular momentum vector. For this work we
will consider a biaxial star, acted upon by an electromagnetic torque as discussed in the
next section. We will work with the angles defined in Jones and Andersson [90]: that is
the star emits its EM radiation beam along the magnetic dipole m which makes an angle
χ with the symmetry axis of the moment of inertia, and θ is the so-called wobble angle
made between the symmetry axis and the angular momentum vector. We will consider
the small-wobble angle regime where θ  1 since this is thought to be the most physical
solution for PSR B1828-11. Finally, we define P as the rotation period, and Ṗ its time
derivative, where the small variations due to precession have been averaged over, and

P
τage ≡ . (5.13)
Ṗ
as a characteristic spin-down age.

5.3.3.1 Defining the spin-down rate model

Observers infer the spin-down rate by measuring the arrival times of pulsations. For
a freely precessing star the spin-down rate is periodically modulated on a time-scale
known as the free precession period, which we will denote as τP . This result is referred
to as the geometric modulation [90] since it is a geometric effect. Under the action of
a torque the geometric effect persists, but an additional electromagnetic effect enters
owing to torque variations [44]. The authors of Jones and Andersson [90] and Link
and Epstein [108] studied both effects in the presence of a vacuum dipole torque [52]
and agreed that the electromagnetic contributions dominate for PSR B1828-11: we will
therefore neglect the geometric effect. The precession model for PSR B1828-11 has been
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
112 for PSR B1828-11

developed by Akgün et al. [14] where a non-vacuum dipole torque was considered and
additionally by Arzamasskiy et al. [22] where the effect of a plasma-filled magnetosphere
was investigated. All of these are potential areas of improvement, but in this work we
will restrict our focus to the simplest specification capable of explaining the observations.

For this model comparison, we will use the spin-down rate under an EM torque derived
in Section 4.5.1 and verified numerically in Section 4.5.2. This derivation uses a gener-
alisation of the vacuum dipole torque to allow for a braking index n 6= 3, but retains the
angular dependence; this ansatz may be written as

ν̇ = −kν n sin2 Θ, (5.14)

where k is a positive constant, and Θ is the polar angle between the dipole and the
angular momentum vector as calculated in Eqn. (52) of Jones and Andersson [90]. Then
we calculate secular solutions in which Θ takes its fixed, time-averaged value. We
denote by θ the angle between the symmetry axis of the moment of inertia and the
angular momentum vector, and denote by χ the angle between the symmetry axis and
the magnetic dipole m. We can then combine the secular solution with an expansion of
sin2 Θ in the small θ limit, to give a spin-down rate
  
1 1 θ
ν̇(t) = −1 + n (t − tref ) + θ 2 cot χ sin (ψ(t)) − cos (2ψ(t)) , (5.15)
τage P τage 2

where
t − tref
ψ(t) = 2π + ψ0 . (5.16)
τP
This is Eqn. (4.68) derived in Section 4.5.1 with the transformation t → t − tref where, as
in the switching model, tref = MJD 49621. The first two terms are the secular spin-down
rate and its first derivative. The term in the square brackets is the modulation and can
be found from appropriate manipulation of Eqn. (58) and (73) in Jones and Andersson
[90], or Eqn. (20) in Link and Epstein [108], aside from a factor of χ in the harmonic
term which we believe to be a misprint.

For χ < π/2 the spin-down rate modulations are sinusoidal. When χ ≈ π/2 (such
that the star is nearly an orthogonal rotator), we will see a strong harmonic at twice
the precession frequency. It is precisely this behaviour which is able to explain the
doubly-peaked spin-down rate for PSR B1828-11.

5.3.3.2 Parameter estimation for the spin-down

In Table 5.5 we list the priors selected for the spin-down precession model. For τP and
σν̇ we use a prior based on a crude estimate from the data. For the spin-down age,
braking-index, and pulse period we use a normal prior taking the mean and standard
deviation from PSR B1828-11 measurements reported in the ATNF catalogue: the values
are listed in Table 5.6. In the analysis, we present the posteriors of the difference to
the means of these values for convenience. Finally for ψ0 we give the full domain of
possible values. Since our derivation of the signal models in Section 5.3.3 assumed the
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 113

Parameter Distribution Units

ATNF
τage -τage N (0, 0.3169) yrs
ATNF
n-n N (0, 0.1700)
P -P ATNF N (0, 1.2 × 10-11 ) s
τP Unif(450, 550) days
θ Unif(0, 0.1) rad
χ Unif(2π/5, π/2) rad
ψ0 Unif(0, 2π) rad
σν̇ Unif(0, 1 × 10-15 ) s-2

Table 5.5: Prior distributions for the spin-down precession model.

P ATNF 0.405043321630 ±1.2 × 10−11 s

nATNF 16.08 ± 0.17
ATNF
τage 213827.91 ±0.32 yrs

Table 5.6: Measured and inferred values of the precession spin-down model
parameters from the ATNF pulsar catalogue [117]. These are given at epoch
MJD 49621.

small wobble-angle regime θ  χ and χ ∼ π/2, we similarly restrict their uniform priors
to the relevant range.

Running the MCMC simulations we plot the resulting posterior in Figure 5.7 and provide
a summary in Table 5.7. For all parameters the posterior distribution is Gaussian: in
the case of parameters which we gave a uniform prior, this indicates that the spin-down
data is informative. For the parameters using an informative prior (from the ATNF
catalogue), the posterior and prior are similar. This indicates the data agrees with the
prior, but does not significantly improve our estimates.

Parameter Mean ± s.d. Units

ATNF
τage -τage 7.461 × 10-3 ±0.3159 yrs
ATNF
n-n 0.0199±0.1701
P -P ATNF 1.0436 × 10-14 ±1.19 × 10-11 s
τP 485.56±0.8188 days
θ 0.0490±0.0020 rad
χ 1.5517±0.0013 rad
ψ0 3.8709±0.0697 rad
σν̇ 4.0423 × 10-16 ±1.81 × 10-17 s-2

Table 5.7: Posterior estimates for the spin-down precession model. For the
secular spin-down quantities, we report the posterior difference with respect to
the values as listed in Table 5.6.

In Figure 5.7B we check the fit of the posterior for the spin-down data. The spin-down
model fits to the data points well with only a small amount of noise required to explain
the data.

The posterior distributions conditioned on the spin-down data (as summarised in Ta-
ble 5.7) can be compared with the values reported in Table 2 of Link and Epstein [108].
When comparing, it should be noted that we are considering a longer stretch of data
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
114 for PSR B1828-11

Figure 5.7: A: The estimated marginal posterior probability distribution for the
precession spin-down model parameters. For the secular spin-down quantities,
we show the difference with respect to the values as listed in Table 5.6. B:
Checking the fit of the model using the maximum posterior values to the data;
see Figure 5.3 for a complete description.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 115

which includes most, but not all of the period studied by Link and Epstein [108]. For
the two angles χ and θ, the fractional difference is 0.001 and 0.14 respectively while the
precession periods differ by a fractional amount 0.05. Clearly the solution found here is
similar to that found by Link and Epstein [108].

5.3.3.3 Defining the beam-width model: Gaussian intensity

Modulation of the observed beam due to precession is a purely geometric effect. We have
already investigated the beam-width for a Gaussian intensity using our numerical model
in Section 4.6.2; in this subsection we will develop the corresponding analytic model
of the beam-width under a Gaussian intensity. This is an original development of the
precession model, although Akgün et al. [14] have considered the equivalent calculation
for the so-called shape parameter S. Ultimately, in agreement with Akgün et al. [14],
we will find that a Gaussian intensity profile will not produce the required variations
in W10 . Instead, in Section 5.3.3.5 we will develop a modified-Gaussian intensity which
can explain the variations.

Fixing the beam-axis to coincide with the magnetic dipole m and following Jones and
Andersson [90], we define Θ and Φ as the polar and azimuthal angles of m with respect
to a fixed Cartesian coordinate system with z along the angular momentum vector J.
The observer is fixed in the Cartesian coordinate system with a polar angle ι to J, and
azimuth Φobs . The slow precessional motion of the spin-vector causes modulation in the
angle Θ:

Θ(t) = cos−1 (sin θ sin χ sin ψ(t) + cos θ cos χ) , (5.17)

which, in the θ  1 limit is approximately

Θ(t) ≈ χ − θ sin ψ(t). (5.18)

Taking the plane containing the angular momentum vector and the observer, in Fig-
ure 5.8 we demonstrate the range of motion of m over a precessional cycle by a grey
shaded region. The region has a mean polar value of χ and a range of 2θ.

2θ
Θ ΔΘ
ι m

Figure 5.8: Illustration of the angles as the beam-axis m cuts the plane con-
taining the observer and the angular momentum J. The grey shaded region
indicates the extent to which m varies over a precessional cycle.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
116 for PSR B1828-11

For an observer fixed at an angle ι to J we define an impact parameter:

∆Θ(t) = Θ(t) − ι, (5.19)

which will vary in time with the precession period τP . This impact parameter determines
how the observer’s line-of-sight cuts the emission beam; if Θ(t) varies due to changes in
∆Θ(t), then the observer will measure the beam to vary on the slow precession time-
scale. To help visualise the setup, in Figure 5.9 we plot the unit sphere with points
corresponding to the beam-axis m and the observer. For each of these we have added
lines of latitude and longitude. Then we see that ∆Θ is the difference between the lines
of latitude and we can also define ∆Φ(t) = Φ(t) − Φobs as the difference in the lines of
longitude.

m
∆Φ
∆Θ

Obs.

Figure 5.9: The angular position of the observer and the magnetic dipole m
on the unit sphere centred on the star; ∆Θ and ∆Φ are then the polar and
azimuthal angles between them.

The analysis by Stairs et al. [156] characterised the beam by a shape parameter. Link
and Epstein [108] used an expansion in ∆Θ to model the beam-width and hence the
shape-parameter. This allowed them to use their fit to the timing-data to infer the
beam-geometry which they found to be hour-glass shaped; see their Figure 5 for a
schematic illustration. The authors of Lyne et al. [111] did not use a shape-parameter
as it requires time-averaging over a longer base-line, something they wish to avoid in
order to be able to observe the switching. Instead, they considered the beam-width at
10% of the maximum, W10 , which is measured on a shorter time-baseline (∼ 1 hr). If
we want to use the beam-width to make a model comparison, we will require a model
for W10 that is not informed by the data.

The integrated pulse profile of PSR B1828-11 (Figure 4 of Lyne et al. [111]) shows a single
peak, often described as core emission [112]. Since we do not have a detailed model of
the emission mechanism, we will now consider the most rudimentary and natural beam
geometry which fits this: a circularly symmetric (about the beam-axis) intensity which
falls-of with a Gaussian function. In the following, we will repeat some of the work of
Sec. 4.6.2 for completeness. Specifically, let us define ∆d as the central angle between
the observer’s line-of-sight and the beam (this is the spherical distance between the two
points marked in Figure 5.9), then the intensity is
 
∆d(t)2
I(t) = I0 exp − . (5.20)
2ρ2
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 117

Here I0 is the intensity when observed directly along the dipole and ρ measures the
angular width of the beam. From the spherical law of cosines, for an observer located
at (Φobs , ι), we have

∆d(t) = cos−1 [cos Θ(t) cos ι + sin Θ(t) sin ι cos |∆Φ(t)|] . (5.21)

The observer will see a maximum pulse intensity at ∆Φ = 0, given by

!
(Θ(t) − ι)2
Imax = I0 exp − . (5.22)
2ρ2

Now let us recognise that Θ varies on the slow precession time-scale, while Φ varies on
the rapid spin time-scale: a single pulse consists of Φ varying between Φobs − π and
Φobs + π. So over a single pulse, we can treat Θ as a constant. The pulse width W10 ,
as measured by observers, is the duration for which the pulse intensity is greater than
10% of the peak observed intensity. For a single pulse, we can define this duration as
the period for which the inequality

1
I > Imax , (5.23)
10
is satisfied. Substituting equations (5.20) and (5.22) into (5.23) and rearranging we find
that
cos Ψ(t) − cos Θ(t) cos ι
cos(|∆Φ(t)|) > , (5.24)
sin Θ(t) sin ι
where
q
Ψ(t) = (Θ(t) − ι)2 + 2ρ2 ln (10). (5.25)

Since we treat Θ as a constant over a single pulsation, we can also treat the whole
right-hand-side of the inequality as a constant during each pulse.

Now consider a single rotation with the magnetic dipole starting and ending in the an-
tipodal point to the observer’s position such that ∆Φ(t) increases between −π and π
during this rotation. Then inequality (5.24) measures the fraction of the pulse corre-
sponding to the beam-width measurement. In terms of the rotation, we define δΦ as the
angular width for which the inequality is satisfied and calculate it to be
 
−1 cos Ψ(t) − cos Θ cos ι
δΦ(t) = 2 cos . (5.26)
sin Θ sin ι

Then the beam-width is

δΦ(t)
W10 (t) = P , (5.27)
2π
from which we arrive at
 
P cos Ψ(t) − cos Θ(t) cos ι
W10 (t) = cos−1 . (5.28)
π sin Θ(t) sin ι
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
118 for PSR B1828-11

In order for the observer to measure the width at 10% of the maximum, the beam
intensity must of course drop below this value before increasing again. In reality, we
typically observe pulse durations lasting for small fractions of the period, especially when
they are close to orthogonal rotators [112].

To set a prior on ρ we consider a special case in which the polar angle of the beam
and the observer are at the equator (Θ = ι = π/2). From our spin-down analysis, we
know the first of these conditions is true for PSR B1828-11 since χ is close to π/2. The
second condition is based on the assumption that the observer would not see a tightly
pulsed beam if they are not close to the polar angle of the beam. In this special instance,
inserting Eqn. (5.25) into Eqn. (5.28), the beam-width is

 P√
W10  = 2 ln 10ρ. (5.29)
Θ=ι=π/2 π

To set a prior on ρ, we can equate this with the beam-widths used in the switching
model, for which we set a uniform prior from 0 to 0.1P ATNF . To make an even-handed
comparison we will therefore set a uniform prior on ρ from 0 to
π
√ ≈ 0.15, (5.30)
10 2 ln 10
so that, for this special case, the prior range of ρ corresponds exactly to the prior range
of the beam-widths in the switching model. This prior range will change, but not by
orders of magnitude when considering a system close to, but not exactly at, this special
case. Therefore, this prior assures that the model comparison does not introduce any
significant bias into the model comparison.

5.3.3.4 Parameter estimation for the Gaussian beam-width model

We are in a position to fit the Gaussian beam model to the observed W10 values. In
Table 5.8 we list the priors taken from the spin-down fit along with three additional
priors. For ι we choose a uniform prior in cos ι on [−1, 1], this corresponds to allowing
ι to range from [0, π] (the observer could be in either hemisphere); for ρ we apply the
prior from Eqn. (5.30); and for σW10 we use a crude estimate based on the data (again
we use the same prior for all three models).

Fitting Eqn. (5.28) to the data we discover that the Gaussian beam model is a poor fit to
the data. In Figure 5.10B the MPE shows that while the model is able to fit the averaged
beam-width, it cannot simultaneously fit the amplitude of periodic modulations.

The posterior distribution (as seen in Figure 5.10A) is Gaussian for all of the parameters
except cos ι for which it concentrates the probability at ι ≈ 0: the observer looks almost
down the angular momentum vector. Since χ ≈ π/2 and θ  1, for each pulsation the
beam must therefore sweep out a cone with such a large opening angle it is close to a
plane orthogonal to the rotation vector. Meanwhile, the rotation vector is nearly parallel
to the angular momentum, since θ  1. As a result, the beam remains approximately
orthogonal to the observer for the entirety of each pulsation. We find this result difficult
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 119

Figure 5.10: A: The estimated marginal posterior probability distribution for

the Gaussian spin-down model parameters. B: Checking the fit of the model
using the maximum posterior values to the data; see Figure 5.3 for a complete
description.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
120 for PSR B1828-11

Parameter Distribution Units

τP ∗ N (485.6, 0.8188) days
P -P ATNF ∗ N (1.04 × 10-14 , 1.19 × 10-11 ) s
θ∗ N (0.0490, 0.0020) rad
χ∗ N (1.5517, 0.0013) rad
ψ0 ∗ N (3.8709, 0.0697) rad
ρ Unif(0, 0.1500) rad
cos(ι) Unif(-1, 1)
σW10 Unif(0, 5) ms

Table 5.8: Prior distributions for the beam-width Gaussian precession model.
Parameters for which the prior is taken from spin-down posteriors are labelled
by ∗ .

to believe on the grounds that the observer would not see tightly collimated pulsed
emission. For this reason, we conclude that the Gaussian beam-intensity fails to fit the
data because the best-fit is unphysical. In retrospect, this result is not surprising since a
Gaussian beam intensity is known to have a beam-width (as measured by W10 ) which is
independent of the impact angle as discussed by Akgün et al. [14]. This is a direct result
of measuring the beam-width with respect to the observed maximum and the self-similar
nature of the Gaussian intensity under changes in the impact parameter.

5.3.3.5 Refining the beam-width model: modified-Gaussian intensity

As we have demonstrated that the Gaussian beam is unable to explain both the observed
variations and average beam-width, we must now consider how it could be varied in a
natural way which does explain the data. One suggestion from Akgün et al. [14] is to
impose a sharper cut-off, or introduce a conal component in addition to the Gaussian
core emission. We will follow a slightly different path below, one which represents a less
drastic modification of the beam profile.

The beam intensity described by Eqn. (5.20) is circularly symmetric about the beam
axis as viewed on the surface of the sphere. In the context of the hollow-beam model
[143], Narayan and Vivekanand [127] found that pulsar beams can be elongated with
the ratio of major to minor axis being ∼ 3 for typical pulsars. PSR B1828-11 does not
fit into the hollow-beam model (having only a single core component), but nevertheless
if the conal emission can be non-circular a generalisation of our core intensity would be
to allow for an elliptical beam.

To consider non-symmetric geometries, let us take the planar limit of Eqn. (5.21) by
applying small angle approximations in ∆d, ∆Θ and ∆Φ:

∆d(t)2 = ∆Θ(t)2 + sin Θ(t) sin ι∆Φ(t)2 . (5.31)

This corresponds to setting the observer close to m in Figure 5.9.

Obviously ∆Φ ranges over [0, 2π] in each rotation, but when ∆Φ is not small, the
intensity vanishes rapidly due to the Gaussian beam shape Eqn. (5.22). Therefore,
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 121

Eqn. (5.31) is a good approximation for the separation when the beam is pointing near
to the observer, while away from this it is a poor approximation, but the intensity is
negligible and so the differences are inconsequential.

We can now allow for an elliptical beam geometry by postulating the beam intensity to
be
 
−∆Θ(t)2 (sin Θ(t) sin ι∆Φ(t))2
I(t) = I0 exp − − . (5.32)
2ρ21 2ρ22

Then to calculate the beam-width, we first find the maximum:

 
−∆Θ(t)2
Imax = I0 exp − . (5.33)
2ρ21

Solving for the beam-width we find

√
P 2 ln 10ρ2
W10 (t) = , (5.34)
π sin Θ(t) sin ι

which is independent of ρ1 , the latitudinal standard-deviation. The extra degree of

freedom introduced in Eqn. (5.32) is irrelevant to the beam-width measure because W10
is defined by the ratio of the intensity to that at the observed peak Imax .

This loss of a degree of freedom means that Eqn. (5.34) is an equivalent to an expansion
of Eqn. (5.28) in the planar limit (i.e. the non-circular nature introduced by Eqn. (5.32)
does not manifest in the prediction for W10 ) and so will suffer the same problems if
fitted to the data. To further generalise our intensity model we will therefore modify
the beam-geometry by allowing a varying degree of non-circularity. This is done by
expanding the longitudinal standard deviation as

ρ2 (t) = ρ02 + ρ002 ∆Θ(t)2 . (5.35)

Note that we have neglected to include a linear term here, forcing the geometry to be
longitudinally symmetric about the beam-axis. Preliminary studies began by fitting a
linear term only (this giving a modulation at the frequency 1/τP ), but it was found
that including a second-order term (which provides modulation at both 1/τP and 2/τP )
gave a better fit. Including both terms, we found that the data was unable to provide
inference on both ρ02 and ρ002 due to degeneracy. In light of this, we drop the first term,
but keep the second, which we feel is the simplest model which is able to fit the data.

Solving for the beam-width (i.e. with Eqn. (5.35) substituted into Eqn. (5.32)) we obtain
a signal model
s
P 2 ln 10

W10 (t) = ρ02 + ρ002 ∆Θ(t)2 , (5.36)
π sin Θ(t) sin ι

which we will refer to as the modified-Gaussian precession beam-width model.

 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
122 for PSR B1828-11

5.3.3.6 Parameter estimation for the modified-Gaussian precession beam-

width

For Eqn. (5.36), we give the relevant prior distributions in Table 5.9. As in the previous
Gaussian model, we let ι range over [0, π]; for ρ02 , we apply the prior on intensity widths
as given by Eqn. (5.30); and for ρ002 we will use a normal prior with zero mean favouring
a Gaussian intensity. The standard-deviation of this prior can have a measurable impact
on the inference: if it is too small then the degree of freedom introduced by Eqn. (5.35)
is effectively removed. Instead, we want to make it significantly larger than the (a priori
unknown) posterior value of ρ002 : this generates a so-called non-informative prior. To set
the prior standard-deviation then, we need to provide a rough scale for what value ρ002
should have. To do this we will define our prior expectation such that

ρ2 (∆Θ = ρ02 ) ∼ 2ρ02 , (5.37)

which is to say we expect ρ2 to increase by no more than a factor of order unity over
angular distances of the beam-width comparable to ρ02 (the beam-width when the ob-
server cuts directly through the beam-axis). This amounts to assuming that the beam
does not depart very far from circularity. Plugging this into Eqn. (5.35), we get

1
ρ002 ∼ . (5.38)
ρ02

From this, we use the upper limit from the uniform prior on ρ02 (as calculated in
Eqn. (5.30)), to set the standard-deviation for ρ002 at 1/0.15 ≈ 7. We also tested dif-
ferent choices of ρ002 and found that the posteriors and odds-ratios were robust to the
choice, provided the standard-deviation did not exclude the posterior value reported in
Table 5.10.

Parameter Distribution Units

τP ∗ N (485.6, 0.8188) days
P -P ATNF ∗ N (1.04 × 10-14 , 1.19 × 10-11 ) s
θ∗ N (0.0490, 0.0020) rad
χ∗ N (1.5517, 0.0013) rad
ψ0 ∗ N (3.8709, 0.0697) rad
ρ02 Unif(0, 0.1464) rad
ρ002 N (0, 6.8308) rad-2
cos(ι) Unif(-1, 1)
σW10 Unif(0, 5) ms

Table 5.9: Prior distributions for the beam-width modified-Gaussian precession

model. Parameters for which the prior is taken from spin-down posteriors are
labelled by ∗ .

The MCMC simulations converge quickly to a Gaussian distribution as shown in Fig-

ure 5.11A and the posterior is summarised in Table 5.10. The model parameters com-
mon to the spin-down model do not vary significantly from the spin-down posterior:
this indicates the two models are consistent. We find that ι is close to π/2 as expected,
ρ2 is sufficiently small indicating a narrow pulse beam, but ρ002 has departed from its
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 123

Figure 5.11: A: The estimated marginal posterior probability distribution for

the modified-Gaussian precession beam-width model parameters. B: Check-
ing the fit of the model using the maximum posterior values to the data; see
Figure 5.3 for a complete description.

prior mean of zero. This confirms that our generalisation of the Gaussian intensity,
Eqn. (5.35), is important in fitting the data.

In Figure 5.11B we perform the posterior predictive check plotting the MPE alongside
the data. This demonstrates that the best-fit puts χ within θ of ι such that during
the precessional cycle the beam-axis passes twice through observer’s location. This
corresponds to the grey region in Figure 5.8 intersecting the observer’s line-of-sight.
When this happens, the modulation of the beam-width picks up a second harmonic at
twice the precession frequency. The minima in Figure 5.11B corresponds to the point
in the precessional phase when the beam-axis point directly down the observer’s line-of-
sight.
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
124 for PSR B1828-11

Parameter Mean ± s.d. Units

τP 484.87±0.4706 days
P -P ATNF -1.7719 × 10-13 ±1.19 × 10-11 s
θ 0.0490±0.0020 rad
χ 1.5517±0.0013 rad
ψ0 3.9701±0.0403 rad
ρ02 0.0245±0.0004 rad
ρ002 3.4421±0.3878 rad-2
cos(ι) 7.9326 × 10-3 ±1.9 × 10-3
σW10 1.5833±0.0422 ms

Table 5.10: Posterior estimates for the beam-width modified-Gaussian preces-

sion model.

5.3.3.7 Recreating the beam-geometry

Since we have defined a beam-intensity in Eqn. (5.32) we can recreate the beam-geometry
and pulse-shape from our MPE values. The data we have does not provide information
about the latitudinal beam-shape parameter ρ1 ; therefore we consider that there are
a family of beam-geometries parameterised by ρ1 = λρ02 where λ is an arbitrary scale
parameter and ρ02 is the MPE value.

In Figure 5.12 we pick four illustrative values for λ and plot the resulting beam-geometry
as contour lines at fixed fractions of the maximum beam intensity (which occurs at
the origin). This demonstrates that the beam-geometry has an hour-glass shape in
agreement with Link and Epstein [108], although this becomes weaker with smaller
values for λ.

In Figure 5.12, a pulse corresponds to a horizontal cut through the intensity at fixed
∆Θ. Our posterior distribution, Figure 5.11A, also provides information on how the
observations cut through this beam-geometry. Under the precession hypothesis, the ob-
server has a time-averaged ∆Θ of χ − ι: this has been plotted as a horizontal dashed line
in Figure 5.12. Precession modulates ∆Θ about this average value by ±θ; the observer’s
line-of-sight through the beam therefore varies by 2θ ≈ 0.1 rad over a precessional cycle.
We have plotted a grey shaded region in Figure 5.12 to show the extent, χ − ι ± θ, over
which ∆Θ varies during a precessional cycle.

We stress here that the contour lines cannot be used directly to measure the beam-width
W10 . This is because W10 is defined as the width at 10% of the peak intensity for that
observed pulse and not the maximum intensity of the beam. The peak intensity for an
observed pulse (a horizontal slice) is the intensity at ∆Φ = 0 and it is with respect to
this, which W10 is measured.

By construction, the four beam geometries in Figure 5.12 all produce the same W10
behaviour as observed in Figure 5.11B. The reason for this is that we have lost informa-
tion on the total intensity by using W10 ; other measurements of the beam-width could
potentially yield more information and better constrain the beam geometry.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 125

Figure 5.12: Recreating the beam geometry from the MPE of the modified-
Gaussian precession beam-width model parameters for four different values of
λ. Thick black lines indicate contour lines of the intensity function at fractions
of the maximum intensity. The hatched area indicates the region of horizontal
cuts (pulses) sampled by the observer: this has a mean, χ − ι, close to zero
(marked by a dashed line) and varies by ±θ about this mean.

Fixing λ = 1 we can also consider the variations in the pulse profile. In Figure 5.13 we
plot the normalised intensity for three values of ∆Θ corresponding to the mean, and
edges of the grey region in Figure 5.12. This figure shows that the narrow beam-widths
occur when ∆Θ is small, which, since χ is close to π/2 coincide with the larger (absolute)
spin-down rates. This agrees with the findings of Lyne et al. [111] and this figure can
be directly compared with panel C in Figure 3 of that work.

5.4 Estimating the odds-ratio

5.4.1 Thermodynamic integration

Having checked that our MCMC simulations are a reasonable approximation to the pos-
terior distribution we now calculate the marginal likelihood for each model and then
their odds-ratio. To calculate the marginal likelihood we will use thermodynamic inte-
gration. This requires running N parallel MCMC simulations and raising the likelihood
to a power 1/T where T is the ‘temperature’ of the chain. This method was originally
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
126 for PSR B1828-11

0.10

Normalised intensity
∆Θ = χ − ι + θ
0.08 ∆Θ = χ − ι
∆Θ = χ − ι − θ
0.06

0.04

0.02

0.00
−0.10 −0.05 0.00 0.05 0.10
Pulse phase

Figure 5.13: Recreating the pulse profiles for three particular slices through the
beam using a fixed value of λ = 1.

proposed by Swendsen and Wang [158] to improve the efficiency of MCMC simulations
for multimodal distributions. In this work we use this method not to help with the
efficiency of the simulations∗ , but instead so that we can apply the method prescribed
by Goggans and Chi [67] to estimate the evidence as follows.

First we define the inverse temperature β = 1/T , then let the marginal likelihood as a
function of β be Z
Z(β) = P (data|θ)β P (θ)dθ. (5.39)

When β = 1, this gives exactly the marginal likelihood first defined in Eqn. (5.3). After
some manipulation we see that
R
1 ∂Z ln(P (data|θ)P (data|θ)β P (θ)dθ
= R . (5.40)
Z ∂β P (data|θ)β P (θ)dθ

From this, we note that the right-hand-side is an average of the log-likelihood at β and
so
∂
(ln(Z)) = hln(P (data|θ))iβ . (5.41)
∂β
Using the likelihoods calculated in the MCMC simulations, we numerically integrate the
averaged log-likelihood over β which yields an estimation of the marginal likelihood. To
be confident that the estimate is correct, we ensure that we use a sufficient number of
temperatures and that they cover the region of interest.

5.4.2 Results

Applying the thermodynamic integration technique to all the models, we estimate the
evidence for each model. Taking the ratio of the evidences gives us the Bayes factor and
since we set the ratio of the prior on the models to unity, the Bayes factor is exactly the
odds ratio (see Eqn. (5.2)).
∗
All the posteriors are either unimodal or multimodal with little separation between the modes.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 127

We present the log10 odds-ratio between the models in Table 5.11. A positive value
indicates that the data prefers model A over model B. Note that the error here is an
estimate of the systematic error due to the choice of β values [see 62, for details].

Model A Model B log10 (odds-ratio)

switching noise-only 57.4 ± 0.5
precession∗ noise-only 60.1 ± 0.5
precession∗ switching 2.7 ± 0.5

Table 5.11: Tabulated log-odds-ratios for all models. ∗ By the precession model
here we mean the precession with a modified Gaussian beam model as discussed
in Section 5.3.3.5.

This table allows quantitative discrimination amongst the models. The first two rows
compare the switching and modified-Gaussian precession models against the noise-only
model with the periodic modulating models being strongly preferred in both cases. Then
in the last row we present the log-odds-ratio between the modified-Gaussian precession
and switching model which shows that the data prefers the precession Modified Gaus-
sian model by a factor 102.7 . Using the interpretation of Jeffreys [87], the strength of
this evidence can be interpreted as ‘decisive’ in favour of this precession model. For
completeness, we also mention that the odds-ratio for the non-modified Gaussian model
(which failed to fit the data in a physically meaningful way) against the noise-only model
was 3.1 ± 0.6.

5.4.3 Effect of the choice of prior

For both beam-widths in the switching model we used uniform priors on [0, f P ATNF ]
with f = 0.1 and these were transformed to also provide a fair prior on ρ02 and ρ002 .
This choice of f was taken from the upper limit quoted in Lyne and Manchester [112]
for typical values of the pulse width. Nevertheless, changing f can have a measurable
impact on the odds-ratio and so we will now study this effect.

To begin, we rewrite Eqn. (5.3), the marginal likelihood, by factoring out the N param-
eters which have a uniform prior
Z
1
P (data|Mi ) = QN P (data|θ, Mi )P (θ ∗ |Mi )dθ, (5.42)
i (bi − a i )

where by P (θ ∗ |Mi ) we mean the probability distribution of all remaining parameters

which are not factored out, and [ai , bi ] is the range for the ith uniform parameter. For the
switching beam-width, the prefactor of this integral (factoring out the prior on W1 and
W2 ) is (f P ATNF )−2 : varying f directly impacts the evidence for the switching model.
For the precession model, we cannot factor the dependence on f in the same way, as we
use a central normal prior on ρ002 . We set the standard-deviation of this prior by applying
Eqn. (5.38) so that it is inversely proportional to f . If both the prior on ρ02 and ρ002 had
been proportional to f we would have an exact cancellation in the odds-ratio and hence
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
128 for PSR B1828-11

no dependence on f . This is not the case and due to our prior on ρ002 the odds-ratio will
depend on f .

To test the dependence, in Figure 5.14 we plot the log odds-ratio as a function of log10 (f )
(note that f = 0.1 corresponds to the result in Table 5.11). There are several features
to understand. First, for f . 0.024 the odds-ratio rapidly grows, favouring precession;
this is because for such small values of f , the beam-width switching prior excludes the
values of W1 required to fit the data. As a result the switching solutions are unnaturally
disadvantaged compared to the precession solutions. Such odds-ratios do not fairly
compare the models.

For 0.024 . f . 0.3 the log-odds-ratio is approximately linear growing from 1.44 when
f = 0.024 to 3.21 when f = 0.3. In this region the solutions for both models are sup-
ported by the prior in that it does not exclude or disfavour the posterior value. The
variation in the odds-ratio results from changes in the prior volume of the switching
model, the evidence for the precession model is constant in this region. Small f values
maximally constrains the prior volume for the switching model (without excluding poste-
rior values) and hence give the greatest weight of evidence to switching and the smallest
odds-ratios. For larger f values the log of the prior volume grows linearly resulting in
the observed growth.

For f & 0.3 our choice of standard-deviation for ρ002 starts to disfavour the posterior value
because it is inversely proportional to f . As a result the evidence for the precession model
decreases faster than the loss of evidence for the switching model leading to the observed
drop in the odds-ratio. In this case it is the precession solutions which are unnaturally
disadvantaged by our choice of prior and so, as in the f . 0.024 case, we do not consider
such odds-ratios as a fair comparison of the models.

4.0
3.5
log10 (odds-ratio)

3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.024 0.1 0.3 1.0
f

Figure 5.14: Dependence of the odds-ratio with f , the fraction of P ATNF used to
constrain the beam-width priors. The vertical line marks the choice of f = 0.1
used in our model comparison based on the upper limit given by Lyne and
Manchester [112].

In summary the log-odds-ratio and hence our conclusion is robust to reasonable varia-
tions in f from 0.05 to 0.5.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 129

5.5 Discussion

In this chapter, we are using a data set (provided by Lyne et al. [111]) on the spin-
down and beam-width of PSR B1828-11 to compare models for the observed periodic
variations. The two concepts under consideration are free precession and magnetospheric
switching. In order to be quantitative, we built signal models for the beam-width and
spin-down from these conceptual ideas. Using the spin-down data to create proper,
physically motivated priors for the beam-width parameters, we then perform a Bayesian
model comparison between the models asking ‘which model does the beam-width data
support?’. For the models considered here, the data most strongly supports a precession
model with a modified-Gaussian beam geometry allowing for an elliptical beam where
the ellipticity has a latitudinal dependence.

To be clear, this does not rule out the switching interpretation since we have not tested
an exhaustive set of models — we can only compare between particular models. As an
example we could imagine modifying the switching model such that either the switching
times, or the magnetospheric states are probabilistic (or a combination of the two).
Further we believe there is good grounds to develop models combining the precession
and switching interpretation like those discussed in Jones [89] and the preliminary results
of Section 4.7.

In addition to the data considered here, a number of high-time-resolution observations

of PSR B1828-11 were performed by the Parkes telescope, as discussed in Stairs et al.
[157] and Lyne [109]. This data set shows interesting variability in beam-width on short
timescales of O(100) pulses. While the qualitative ‘noisiness’ of the beam-width data
is already apparent from the current data-set (e.g. see Figure 5.1), such high-time-
resolution data could be very interesting to include in a more detailed future model
comparison.

The process of fitting the models to the data and performing posterior predictive checks
also provides a mechanism to evaluate the models. For both spin-down models the
maximum posterior plots with the data (Figs. 5.5B and 5.7B) revealed a systematic
failure to fit the second (slightly lower) minima. This suggests new ingredients could be
introduced to both models to explain this.

The posteriors for the precession model indicate that PSR B1828-11 is a near-orthogonal
rotator and we observe it from close to the equatorial plane. If this is the case, and the
two beams of the pulsar are symmetric about the origin, then we expect to see the
second beam as an interpulse. Indeed, we discuss further in Appendix 5.B how during
the precessional cycle we should expect the intensity of this second beam to dominate at
certain phases. Since no such interpulse is reported, either the second beam is weaker,
or the beams must have a kink of greater than 4.6◦ (see Appendix 5.B).

In this chapter, we have developed the framework to evaluate models for the variations
observed in PSR B1828-11. This is not intended as an exhaustive review of all models,
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
130 for PSR B1828-11

but rather a discussion on the intricacies that arise such as setting up proper and well-
motivated priors. This lays the groundwork for a more exhaustive test of all available
models and can also be extended by including other data sources.

Appendix 5.A Procedure for MCMC parameter estima-

tion

The procedure used to simulate the posterior distribution can determine the quality of
the estimation. Therefore, we will now set out an algorithmic method to ensure that
our results are reproducible.

To estimate the posterior given a signal model and prior, we run two MCMC simula-
tions: an initialisation and production. In the following the term walker refers to single
chains in the MCMC simulation. The Foreman-Mackey et al. [62] implementation runs
a number of these in parallel for each simulation.

• For the initialisation run, we draw samples from the prior distribution to set the
initial parameters for each walker. The simulation therefore has the chance to
explore the entire parameter space. After a sufficient number of steps, the walkers
will converge to the local maxima in the log-likelihood. By visually inspecting the
data we determine the nature of the local maxima: in all cases a single maxima
dominated such that, given a sufficient length of simulation, we expect all walkers
to converge to this maxima. Alternatively we could have found multiple similarly
strong maxima, in this case further analysis would be required. This was not found
to be the case for any of the models in this analysis.

• For the second step we set the initial state of 100 walkers by uniformly dispersing
them in a small range about the maximum-likelihood found in the previous step.
The simulation proceeds from this initial state and we divide the resulting samples
equally into two: discarding the first half as a so-called ‘burn-in’. We retain the
second half as the production data used to estimate the posterior. The burn-in
removes any memory of the artificial initialisation of the walkers at the start of
this step.

Having run an MCMC simulation we check that the chains have properly converged
(ifor a discussion on this see Chapter 10 of Gelman et al. [64]). The MCMC simulations
provide an estimate of the posterior densities for the model parameters. We will also
perform ‘posterior predictive checks’ to ensure the posterior is a suitable fit to the data,
i.e. we compare the data to the model prediction when the model parameters are set to
the values corresponding to the peaks of the posterior probability distributions.
Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
for PSR B1828-11 131

Appendix 5.B Implications for the unobserved beam

The precession model developed here assumes the observer only ever sees one pole of
the beam-axis, but in the canonical model we often imagine there is also emission from
an opposite magnetic pole. In several pulsars this can be seen as an interpulse, 180◦ out
of phase from the main pulse [112, 115]; these pulsars are generally found to be close
to orthogonal rotators† . No such interpulse is reported for PSR B1828-11. We will now
discuss the implications of this given the precession interpretation.

Let us imagine a scenario where the observer is in the northern (magnetic) hemisphere
and label the beam protruding into their hemisphere (which they will see with the greater
intensity due to the smaller angular separation) at the start of the thought experiment
as the north pole. Then, when Θ < π/2, the north and south pole make angles Θ and
π − Θ respectively with the fixed angular momentum vector. Now we see that if, during
the course of the precessional cycle, Θ > π/2 then the south pole will protrude into
the northern hemisphere and the north pole into the southern hemisphere. Provided
both poles are identical, but regardless of the details of the beam-geometry, at this time
we must expect the observer to see the south pole at greater intensity than the north
pole. An example of this is shown in Figure 5.15, but note that the observer will see the
greatest intensity from the south-pole half a rotation after this instance.

Our posterior distributions inform us that, if the precession interpretation is correct, we

are in exactly this situation: Θ ranges from 85.8◦ to 92.3◦ over a precessional cycle‡ so
we should see the interpulse.

This is readily explained if the south pole is substantially weaker in intensity, or by the
one-pole interpretation Manchester and Lyne [118]. Alternatively, it could be that the
two beams are not diametrically opposed but are latitudinally ‘kinked’. In the later case
we can put a lower bound on the kink angle by requiring that the polar angle of the
south pole is always greater than that of the north (see Figure 5.15). From our MPE
this gives a lower bound of 4.6◦ for the polar kink angle. This latitudinal kink can be
compared with the longitudinal kink of interpulses observed in other pulsars: often these
are not found at exactly 180◦ , but can deviate by 10’s of degrees (see the separation of
interpulses for double-pole interpulses in Table 1 of [115]). Allowing for such kinks in
both beams is a possible extension to the precessional model.

†
The use of ‘interpulse’ here strictly refers to seeing the opposite beam of the pulsar, and not cases
where the pulsar is almost aligned and interpulses are thought to come from the same beam
‡
Numbers generated from the maximum posterior estimates of χ and θ using the spin-down data.
The estimated error for both values is ±0.08◦
 Chapter 5 Comparing models of the periodic variations in spin-down and beam-width
132 for PSR B1828-11

ι
Θ

Figure 5.15: Extension to Figure 5.8 adding in the south pole (demarked by a
white triangle and solid line) for an instance in which the north pole is in the
southern hemisphere. The dotted line with a white arrowhead is the ‘kinked’
beam. For clarity, the north and south poles are shown at times that differ by
half a rotation period.
Chapter 6

Continuous gravitational waves:

calculating the mismatch

Timing variations, in the form of glitches or timing noise, may pose a serious risk to
efforts to discover continuous GWs from neutron stars. This is because searches rely on
matched filtering, which searches the data using a template; if the template and signal
don’t match, due to timing variations in the signal, the template may be unable to
detect the signal. In this chapter, we will give a general introduction to the detection
methods used in GW data analysis and then develop our tools which will used in later
chapters to quantify the significance of timing variations for continuous GW searches.
The details of estimating the risk posed by glitches to GW searches will be discussed in
Chapter 7 and then in Chapter 8 and Chapter 9 we will investigate the role of timing
noise for GW searches.

6.1 Introduction

Rotating neutron stars capable of supporting non-axisymmetric mass distributions will

emit continuous gravitational waves∗ (GWs) due to their time-varying quadrupole mo-
ments. These may be detectable by ground-based detectors. The emitted signals can
persist for much longer than the duration of a typical search, but are weak in amplitude,
making them difficult to detect in the noise of the detector.

To find a signal, continuous GW searches use matched filtering techniques such as the
F-statistic [85], which compare the output of the detector with a template. These
techniques are powerful, provided that the signal and template remain coherent for the
duration of the observation. If the signal can be perfectly matched by a template, then
the signal to noise ratio, ρ, used to quantify the detection likelihood, scales as ρ2 ∝ Tcoh
∗
Note that in general ‘continuous waves’ can refer to any quasi-monochromatic long-lasting
gravitational-wave signals, such as emitted by binaries of white dwarfs, neutron stars or black holes,
which would be detectable by LISA or pulsar timing arrays. Here we refer to continuous GWs exclu-
sively in the context of spinning nonaxisymmetric neutron stars as relevant to ground-based detectors.

133
134 Chapter 6 Continuous gravitational waves: calculating the mismatch

(e.g. see [140]) where Tcoh is the coherent observation time. This suggests searching over
longer observations increases the chances of making a detection.

The templates must model the monotonic spin-down of the source due to the electro-
magnetic (EM) and gravitational torque; this is done by Taylor expanding the phase
(see Eqn. (1.2)) about some reference time tref .

Note that all times refer to the solar system barycentre and we assume the timing model
has already correctly accounted for the dispersion measure, proper motion and other
parameters as discussed in Edwards et al. [57]. Pulsar astronomers fit this model to
observed time of arrivals (TOAs). If the best fit model is accurate enough to track
the pulsar to within a single rotation, the resulting timing solution is described as
phase-connected. Often such solutions are capable of tracking the pulsar over durations
greater than a year. For gravitational-wave searches, this level of accuracy motivates
the use of the same Taylor expansion phase models to account for the spin-down. Pulsar
astronomers measure the frequency ν and higher order coefficients describing the rotation
of the pulsar itself. In this work we will denote the GW frequency and its derivatives
by f, f˙, etc. The relation between the GW frequency and rotation frequency depends on
the emission mechanism; in this work we will consider only searches for emission from
non-axisymmetric neutron stars such that f = 2ν (see for example §16.6 of Shapiro and
Teukolsky [154]).

The rest of this chapter is organised as follows: in Section 6.2 we will define the mis-
match which quantifies the loss of signal to noise ratio for imperfectly matched signals.
In Section 6.3 we will familiarise the reader with methods to calculate the mismatch
when the phase of the GW is not a smooth Taylor expansion and describe why this
method of calculating the mismatch becomes intractable for complicated signals. Then,
in Section 6.4 we will define a new approach to calculating the mismatch by splitting
the signal into separate subdomains.

6.2 Introduction to the mismatch

While Taylor expansion models are on average reliable enough to track the spin-down,
pulsars do show timing variations either in the form of glitches, occasional sudden in-
creases in the rotation frequency, or continuous low-frequency variations known as timing
noise (see Chapter 2 for an overview). To quantify the effect these variations will cause,
we will use the mismatch, which is the loss of signal to noise ratio due to an imperfect
match between the template and signal. The mismatch is non-zero whenever the signal
and template are not perfectly matched. Even if the signal does not contain any timing
variations, a mismatch can occur when the template, chosen from a finite grid of points
in parameter space, does not share exactly the same parameters as the signal. This
problem has been studied in detail since finite computing resources limit the number of
templates that can be used in a search, ensuring that any prospective search must have
some level of mismatch. In this section, we introduce the tools that have been developed
to understand this problem. We will return to the problem of timing variations later on
in this chapter.
Chapter 6 Continuous gravitational waves: calculating the mismatch 135

6.2.1 Defining the mismatch

We begin by assuming that the gravitational-wave detector strain data contains a peri-
odic GW signal with phase Φs (t; λαs ), where λαs is a vector of the signal parameters. A
fully-coherent search consists of applying a matched filtering algorithm over a coherence
time Tcoh to search the data using a template; let us then denote the phase evolution of
the template as Φt (t; λαt ), such that λαt is a vector of the template parameters.

Defining the phase-difference between the signal and the template as

∆Φ(t; λαt , λαs ) = Φs (t; λαs ) − Φt (t; λαt ), (6.1)

then, following the work of Prix and Itoh [141] and neglecting amplitude modulations
in the signal, we can define the matched filtering amplitude
Z
1
X= ei∆Φ dt, (6.2)
Tcoh Tcoh

This amplitude will have a global maximum of |X| = 1 for a perfectly matched signal
and will decrease for increasing phase mismatches. In this way, if we define ρ̃ as the
signal to noise ratio (SNR) measured in a fully-coherent search, then X defines the loss
of SNR incurred due to the non-zero phase difference ∆Φ as compared to a perfectly
matched signal with SNR ρ̃pm such that

ρ̃ = |X|ρ̃pm (6.3)

A simple, dimensionless measure of the loss of SNR is found by rearranging and defining
the fully-coherent mismatch as

µ̃(λαs , λαt ) = 1 − |X(λαs , λαt )|2 , (6.4)

such that
ρ̃2pm − ρ̃2
µ̃ = . (6.5)
ρ̃2pm

6.2.2 Interpreting the mismatch

In this work we will quantify the effect of glitches by the mismatch. However, it may be
useful to interpret a mismatch in the following way. For a perfectly matched continuous
GW signal it can be shown [85] that the SNR scales as

h0 p
ρpm ∝ √ Tcoh N , (6.6)
Sn

where h0 is the strain-amplitude for an optimally oriented source with respect to the
detector, Sn , the noise measured in the detector and N is the number of detectors.
136 Chapter 6 Continuous gravitational waves: calculating the mismatch

Combining this equation with Eqn. (6.5), the SNR of a signal with fully-coherent mis-
match µ̃ scales as
p h0 p
ρ ∝ 1 − µ̃ √ Tcoh N , (6.7)
Sn
√
which is to say it reduces the sensitivity by a factor 1 − µ̃.

6.2.3 Taylor expansion signals and templates

So far we have not yet defined a phase evolution model for either the signal or the
template. The signal phase Φs (t; λαs ) will depend on the GW production mechanism,
but assuming the signal is generated by the canonical non-axisymmetric distortion, the
GW signal will be produced with a frequency twice that of the rotational frequency and
this frequency will have a spin-down rate twice the rotational spin-down rate. From the
success of pulsar astronomy, we know that the phase evolution of the rotation can be
modelled by a Taylor expansion. For now, let us truncate at the first order in frequency
derivative, but we could use higher order terms if required. Therefore, we can begin by
modelling the GW phase evolution using a Taylor expansion
" #
f˙s
Φs (t, λαs ) = φs + 2π fs (t − tref ) + (t − tref )2 , (6.8)
2

such that λαs = [φs , fs , f˙s ]. In principle, we can include higher order terms, but for this
work we truncate at the first derivative of the frequency.

For blind searches, where we have no hints about the phase evolution, typical continuous
GW searches assume that the GW signal is a Taylor expansion and search for it using a
Taylor expansion template. Defining λαt = [φt , ft , f˙t ] as the Taylor expansion template
parameters, Eqn. (6.1), the phase difference, is
" #
∆ f˙
∆Φ(t; λαs , λαt ) = ∆φ + 2π ∆f (t − tref ) + (t − tref )2 , (6.9)
2

where ∆λα = λαs − λαt is the difference between the signal and the template which we
refer to as the parameter space offset. Note that by choosing any two of λαs , λαt , and
∆λα we have a choice of three equivalent parameterisations. When the parameter offset
∆λα vanishes, the matched filtering amplitude tends to unity and hence the mismatch
tends to zero. In the other extreme, for a suitably large parameter space offsets, the
mismatch approaches zero and so the signal is completely lost.

Substituting Eqn. (6.9) into Eqn. (6.4) gives the mismatch between a smooth Taylor
expansion signal and template. In this work we will calculate the mismatch for a signal
that is not a smooth Taylor expansion. This can be done in this way for some special
cases. For example, taking ∆λα = [∆φ, 0, 0], the mismatch is found to be zero: this
reflects the fact that fully-coherent matched filtering is insensitive to any arbitrary phase
offset between the signal and template. In more complicated cases, the integrals become
intractable. To aid in calculation, we will use the so-called metric approximation. In
Chapter 6 Continuous gravitational waves: calculating the mismatch 137

the next section we will introduce it in the context of smooth Taylor expansion signals.
Later on in Section 6.4 we will describe how we extend it to arbitrary signals.

6.2.4 The metric-mismatch approximation for fully-coherent searches

The fully-coherent mismatch Eqn. (6.4) has a local minimum of zero at ∆λα = 0.
Expanding about this minimum up to the leading order term, Brady et al. [35] approx-
imated the mismatch by
µ̃(λαt , ∆λα ) ≈ gαβ ∆λα ∆λβ , (6.10)

where gαβ is the parameter space metric given by


1 α

α 
gαβ = ∂α ∂β µ̃(λt , ∆λ ) . (6.11)
2 ∆λα =0

Note that we define ∂α ≡ ∂∆λα . The metric is a function of both the total coherence
time Tcoh over which the matched filter is performed and the reference time at which the
Taylor expansions are defined, but not the signal itself. Partially evaluating the metric,
we find that
Z Tcoh
1
gαβ = ∂α ∆Φ∂β ∆Φdt
Tcoh 0
Z Tcoh Z Tcoh (6.12)
1
− 2 ∂α ∆Φdt ∂β ∆Φdt.
Tcoh 0 0

This metric formulation provides a method to measure the metric-mismatch between a

signal and template that are both Taylor expansions. Brady et al. [35] proposed this
method with the aim of picking the spacing of templates in parameter space such that
the maximum allowable mismatch would not rise above a pre-defined threshold. In this
work, we will instead use this metric-mismatch approximation to calculate mismatches
between signals and Taylor expansion templates.

It is worth commenting that the full mismatch, as calculated from Eqn. (6.4), is bounded
by [0, 1]. In contrast, the approximate metric-mismatch is bounded by [0, ∞). This is
because the expansion of the mismatch in Eqn. (6.10) was taken about ∆λα = 0 and so
for sufficiently large parameter space offsets the expansion breaks down. In this case,
the metric-mismatch approximation non-linearly overestimates the true mismatch. A
metric-mismatch above one, while losing the interpretation as a direct loss of SNR, still
corresponds to a large true mismatch, and hence a significant loss of SNR.

6.2.5 The mismatch for semi-coherent searches

Fully-coherent searches are computationally demanding, so much so that it is infeasible

to perform fully-coherent searches for wide-parameter searches such as the all-sky search.
Instead, semi-coherent searches are used which require far fewer templates and result in
138 Chapter 6 Continuous gravitational waves: calculating the mismatch

more sensitive searches at a fixed computing cost [140]. There are numerous implemen-
tations of semi-coherent searches, for example the E@H search uses the Hough-transform
method [99]; in this section we will discuss the generic case.

A semi-coherent search divides an observation time Tobs into Nseg segments of duration
Tcoh . Each segment is searched fully coherently with a resulting SNR ρ̃2j where j labels
the segment. The semi-coherent search then recombines these segments by summing all
segments at the same point in parameter space λαt to give a new detection statistic

Nseg
X
ρ̂2 (λαt , ∆λα ) = ρ̃2j (λαt , ∆λα ), (6.13)
j

where the ‘hat’ denotes that it is a semi-coherent quantity.

For each fully-coherent segment, we can rearrange Eqn. (6.5) to give

ρ̃2j (λαt , ∆λα ) = ρ̃2pm (1 − µ̃j (λαt , ∆λα )). (6.14)

Here we take ρ̃pm to be the same for all segments such that it does not carry a j-
index; in this way we neglect variations in the signal such as the signal amplitude or
the motion of the Earth and also require all segments to be of equal duration. For the
semi-coherent detection statistic we may define ρ̂2pm as the squared SNR in the absence
of any mismatch. Then the semi-coherent mismatch µ̂ can be implicitly defined as

ρ̂2 (λαt , ∆λα ) = ρ̂2pm (1 − µ̂(λαt , ∆λα )), (6.15)

where µ̂ is the semi-coherent mismatch. When there is no mismatch, the sum of the ρ̃2pm
is equal the semi-coherent ρ̂2pm such that

ρ̂2pm = Nseg ρ̃2pm . (6.16)

Inserting all these expressions into Eqn. (6.13) we see that the semi-coherent mismatch
is an average over the individual full-coherent segment mismatches:
Nseg
1 X
µ̂(λαt , ∆λα ) = µ̃j (λαt , ∆λα ). (6.17)
Nseg
j=1

With this expression, we can first calculate the mismatch for each fully-coherent segment,
and then calculate the mismatch for a semi-coherent search. However, we must be careful
to sum all the mismatches at the same point in parameter space.

In this section we have introduced: the fully-coherent mismatch, used to quantify the
loss of SNR in fully-coherent GW searches; the metric-mismatch approximation, a useful
way to calculate mismatches that avoids the complicated integration; and the semi-
coherent mismatch, quantifying the loss of SNR in semi-coherent GW searches. The
mismatch is a useful way to quantify how well a search performs given a particular GW
signal. However, the tools discussed in this section to calculate this mismatch apply
only to smooth Taylor expansion signals and templates. Signals which contain timing
Chapter 6 Continuous gravitational waves: calculating the mismatch 139

variations are by definition not smooth Taylor expansions. Therefore, we need a method
to calculate the mismatch for arbitrary signals. In the next section, we will discuss some
special cases of timing variations where the mismatch can be calculated exactly. Then
in the final section, we will introduce a new tool which can be used to calculate the
mismatch for arbitrary signals.

6.3 Exact mismatch from irregularities in the phase

The F-statistic considered in Brady et al. [35] analytically minimises over the phase.
Therefore, if the signal and template can both be modelled by a single smooth Taylor ex-
pansion, but with a finite phase-offset, the mismatch is zero: the mismatch is insensitive
to any overall phase difference between the signal and template. However, a mismatch
will occur if the overall phase between the signal and template changes during an obser-
vation. In this section, we will investigate three scenarios where this occurs. While not
all these scenarios are physically relevant to real astrophysical systems, this introduces
some concepts in a simple setting before we tackle the more difficult real astrophysical
systems.

6.3.1 Two subdomains with a phase discontinuity

We begin with a simple system in which the signal undergoes an instantaneous ‘jump’ in
its GW phase halfway through the observation. We can model this signal by a piecewise
Taylor expansion with two subdomains. Both subdomains are of equal duration and
follow a smooth spin-down, except that there is a phase discontinuity at their interface;
we illustrate this setup in Figure 6.1. Parameterising by the offset with respect to the

template
signal

∆φ2
Phase

∆φ1

time

Figure 6.1: Illustration of the signal and template defined in equation (6.18)
140 Chapter 6 Continuous gravitational waves: calculating the mismatch

template parameters λ0 for some arbitrary template and phase jump, we write the phase
deviations in the two subdomains as
(
∆φ1 0 < t < T /2
∆Φ(t) = . (6.18)
∆φ2 T /2 < t < T

To compute the matched filtering amplitude given in Eqn. (6.2), we can factorise the
integral using the additivity of integration on intervals into two integrations
Z T
1
X = ei∆Φ(t) dt (6.19)
T 0
Z Z !
T /2 T
1 i∆φ1 i∆φ2
= e dt + e dt (6.20)
T 0 T /2
1  i∆φ1 
= e + ei∆φ2 . (6.21)
2
Because our choice in splitting up the integral exactly matches the subdomains defined
in Eqn. (6.18) and in each subdomain ∆Φ(t) is constant, we can compute the integrals.
The absolute square value of the matched filtering amplitude is then

1  i∆φ1  
|X|2 = e + ei∆φ2 e−i∆φ1 + e−i∆φ2 (6.22)
4
1 
= 2 + ei(∆φ1 −∆φ2 ) + e−i(∆φ1 −∆φ2 ) (6.23)
4
1
= (1 + cos(∆φ1 − ∆φ2 )) . (6.24)
2
Finally, the fully-coherent mismatch can be calculated from Eqn. (6.4) to be

1
µ̃ = (1 − cos(∆φ1 − ∆φ2 )) . (6.25)
2

From this result, we learn that it is not the phase difference with respect to each of the
Taylor expansions (∆φ1 or ∆φ2 ) that is important, but the total phase jump at their
interface. For ∆φ1 = ∆φ2 the mismatch vanishes, this recovering the case of a single
Taylor expansion signal with an arbitrary overall phase offset, which always has zero
mismatch. This result can be used to calculate the mismatch due to a glitch, if the
glitch is purely in the phase. In reality, we know that glitches are discontinuities in the
frequency and spin-down rate, this will be discussed later in Chapter 7, but this simple
calculation provides a stepping stone to understanding the more complicated complete
result.

6.3.2 N subdomains with phase discontinuities

We can further generalise Eqn. (6.25) by letting the signal be comprised of N equal
duration subdomains with a phase discontinuity ∆φi for the ith subdomain. Then, the
Chapter 6 Continuous gravitational waves: calculating the mismatch 141

matched filtering amplitude can be written

Z Z Z !
t1 t2 tN
1
X = ei∆φ1 dt + ei∆φ2 dt + · · · + ei∆φN dt (6.26)
T 0 t1 tN −1
 
1 T i∆φ1 T T
= e + ei∆φ2 + · · · + ei∆φn (6.27)
T N N N
N
1 X i∆φi
= e . (6.28)
N
i=1

Squaring the matched filtering amplitude and simplifying

N
! N 
1 X X
|X|2 = ei∆φi  e−i∆φj  (6.29)
N2
i=1 j=1
 
XN X N
1
= ei∆φi  e−i∆φj  (6.30)
N2
i=1 j=1
 
N N
1 X i∆φi 
e−i∆φi +
X 
−i∆φj 
= e  e  (6.31)
N2
i=1 j=1
j6=i
 
N
X N X
X N
1 

= 1+ ei(∆φi −∆φj )  (6.32)
N2  
i=1 i=1 j=1
j6=i
 
N X
X N
1 N +

= cos(∆φi − ∆φj ) + i sin(∆φi − ∆φj )
. (6.33)
N2 
i=1 j=1
j6=i

In this final summation, for each pair (i, j) the corresponding pair (j, i) will exist in the
sum. This leads to a cancellation of the imaginary part and a doubling of the real part
 
N N
1
 1 XX 
2
|X| =  + 2 cos(∆φi − ∆φj )
. (6.34)
N N
i=1 j=1
j6=i

Then the fully-coherent mismatch is given by:

N N
1 1 XX
µ̃ = 1 − − 2 cos(∆φi − ∆φj ). (6.35)
N N
i=1 j=1
j6=i

This result could be used to estimate the mismatch due to a random walk model of
timing noise where the random walk was purely in the phase (see Section 2.3.1 for
example). In Chapter 9 we will discuss random walks in GW signals in more detail and
also calculate the mismatch for random walks in the frequency and spin-down rate.
142 Chapter 6 Continuous gravitational waves: calculating the mismatch

6.3.3 Oscillating phase deviations

Timing residuals, or equivalently the phase offsets between Taylor expansions and the
real signals, are often found to have a quasi-periodic structure [81]. It is reasonable
to ask if, because the residuals oscillate about the origin, the loss of detection will be
cancelled out. In this section we consider such an oscillating phase deviation and show
that, over many oscillations, the mismatch is non-zero.

We model a small sinusoidal phase offset by a trigonometric function

∆Φ(t) = ε cos(ωt), (6.36)

where ε is the amplitude of variations and ω is the angular frequency of the oscillations.
If we assume that the variations are small ε  1, we can Taylor expand the exponential
phase offset in the matched filtering amplitude
Z
1 2 i3

X= 1 + i cos (ωt) − cos2 (ωt) − cos3 (ωt)dt + O 4 . (6.37)
T T 2 6

Inserting this into Eqn. (6.4) and computing the fully-coherent mismatch we find that
 
2 1 sin (2T ω) sin2 (T ω)

µ̃ = ε + + + O 4 . (6.38)
2 4T ω T 2ω2

In the limit T → ∞, this tends to ε2 /2. The mismatch oscillates around this value over
observation times similar to the oscillation period; this is shown in Figure 6.2.

ε2
2
Mismatch

0 π 2π 3π 4π
ωT

Figure 6.2: Plot demonstrating the behaviour of Eqn. (6.38), the mismatch for
an oscillating phase offset.

This simple model demonstrates that if the signal and template are on average coherent,
but the phase residual between them oscillates about the mean, then over several cycles
Chapter 6 Continuous gravitational waves: calculating the mismatch 143

the mismatch will approach a constant value, which depends on the maximum amplitude
of the residuals. For a quasi-periodic signal, we expect the same overall picture will
emerge with the size of the mismatch dependent on the mean average size of oscillations
in the residual.

6.4 Generalising the metric-mismatch approximation to

arbitrary signals

In the previous section, we discussed some special cases where the mismatch can be
calculated exactly for a signal which has a phase which can’t be described by a single
Taylor expansion. For more general signals, where the frequency and spin-down rate
may vary, the equations soon become unwieldy. In Section 6.2.4 we introduced the
metric-mismatch, a tool to estimate the loss of SNR for Taylor expansion signals and
templates in fully-coherent searches. In this section, we will extend the fully-coherent
metric-mismatch to handle cases where the signal is not a Taylor expansion, but can
be approximated by a piecewise Taylor expansion; this will allow us to handle any
arbitrary phase evolution of the signal. By piecewise Taylor expansion we mean breaking
any arbitrary signal into Nsd subdomains each of which we describe the signal by a
single Taylor expansion; the resulting collection of all individual Taylor expansions is a
piecewise Taylor expansion. We will distinguish between a subdomain, which describes
a part of the piecewise function and a segment, which is the division of an observation
period for a semi-coherent search as discussed in Section 6.2.5.

This generalised metric-mismatch will be used in Chapter 7 to calculate the mismatch

due to glitches and in Chapter 9 to calculate the mismatch due to a random walk model
of timing noise.

6.4.1 Piecewise Taylor expansion

Before we get into the details of formulating a general metric-mismatch, it is worth

stating some details about piecewise Taylor expansions. Allowing the signal to be piece-
wise amounts to placing a second index on the signal parameters, λαa s , which labels the
subdomain of the piecewise function. Note that we discriminate between Greek indices,
labelling the parameter components, and Roman indices which label the subdomain.
For the ath subdomain of the piecewise function, we define taref as the reference time of
the Taylor expansion in that subdomain. These can be defined arbitrarily, although it
is usual to either define them relative to each subdomain, or at a fixed value for all the
subdomains. The piecewise signal function in the ath subdomain is then
!
f˙a
s
Φs (t; {λαa a a a
s }) = φs + 2π fs (t − tref ) + (t − taref )2 , (6.39)
2
144 Chapter 6 Continuous gravitational waves: calculating the mismatch

when ta < t ≤ ta+1 and otherwise undefined: in future this condition will be assumed.
By {λαas } we indicate the set of all signal parameters and by implication their reference
times.

Since we will be calculating the phase difference between the signal and template it is
convenient to similarly make the template a piecewise function with parameters λαa t
such that
!
αa a a a f˙ta a 2
Φt (t; {λt }) = φ + 2π ft (t − tref ) + (t − tref ) . (6.40)
2

Then we may write the parameter space difference as ∆λαa = λαa αa

s − λt , where both
the signal and template must refer to the same reference time.

Since we have set up the template phase in Eqn. (6.40) as a piecewise Taylor expansion,
the template can be any arbitrary function described by a piecewise Taylor expansion.
However, all current and planned searches use a single global Taylor expansion template.
To model this, we must introduce consistency relations which ensure that the local
subdomains all lie along a single global template with parameters [φt , ft , f˙t ] defined at
tref . These consistency relations may be written as

f˙t2 a
φat = φt + ft (taref − tref ) + (t − tref )2 ,
2 ref
(6.41)
ft a = ft + f˙t (taref − tref ),
f˙ta = f˙t .

Note the subtle distinction between tref , the global reference time and taref the reference
time for the ath subdomain.

Finally, we can generalise Eqn. (6.9) for the phase difference to the piecewise Taylor
expansion description of the signal and template:

∆Φa (t) = Φs (t; λαa αa

s ) − Φt (t; λt ) (6.42)
!
∆f˙a
= ∆φa + 2π ∆f a (t − taref ) + (t − taref )2 . (6.43)
2

6.4.2 The generalised metric-mismatch approximation for fully-coherent

searches

We will now generalise the metric-mismatch approximation first introduced by Brady

et al. [35] to the case where the signal and template are piecewise Taylor expansions.
This calculation follows that given in Section 6.2.4 with the addition of an index labelling
the subdomains.

Using a piecewise Taylor expansion to describe the signal and template, the discrete
nature of the phase offset ∆Φa allows us to partition the matched filtering amplitude
due to the additivity of integration on intervals. The matched filtering amplitude, given
Chapter 6 Continuous gravitational waves: calculating the mismatch 145

the set of template parameters and parameter offsets {λαa αa

t , ∆λ }, is then
Z Tcoh
1
X({λαa αa
t , ∆λ }) = ei∆Φ(t) dt (6.44)
Tcoh 0
N
Xsd Z
1 c
= ei∆Φ dt, (6.45)
Tcoh c tc

where the integration is taken over the bounds of the cth subdomain. The mismatch, in
this general formulation, is then defined by

µ̃({λαa αa αa αa
t , ∆λ }) = 1 − |X({λt , ∆λ })|
2
(6.46)

The mismatch has a local minimum of zero when {∆λαa } = 0. Expanding in powers of
∆λαa , the leading order term is

1
µ̃({λαa αa αa αa αa βb
t , ∆λ }) = ∂αa ∂βb µ̃({λt , ∆λ })∆λ ∆λ , (6.47)
2
where ∂αa ≡ ∂∆λαa . The metric in this formalism is identified as

1 αa αa 

gαβab = ∂αa ∂βb µ̃({λt , ∆λ }) . (6.48)
2 {∆λαa }=0

We can then define the fully-coherent generalised metric-mismatch as

µ̃({λαa αa αa βb
t , ∆λ }) = gαβab ∆λ ∆λ . (6.49)

On the right hand side we sum over the repeated indices. The mismatch is a scalar value
quantifying the loss of signal to noise due to the set of parameter offsets {∆λαa }.

Partially evaluating the metric we have

Z
1 X
gαβab = ∂αa ∆Φc ∂βb ∆Φc dt
Tcoh c tc
Z XZ (6.50)
1 X c 0
− 2 ∂αa ∆Φ dt ∂βc0 ∆Φc dt.
Tcoh tc 0 tc0
b b

Expanding the summation over the a index, we note that,

 α
∂αa ∆Φc = 1, 2π(t − tar ), π(t − tar )2 δac , (6.51)

and so we can simplify Eqn. (6.50) by dropping the terms which vanish from the sum-
mation
Z
1
gαβab = δab ∂αa ∆Φa ∂βa ∆Φa dt
Tcoh ta
Z Z (6.52)
1 a
− 2 ∂αa ∆Φ dt ∂βb ∆Φb dt.
Tcoh ta tb

This expression, given a choice of decomposition of the observation time T into Nsd (not
146 Chapter 6 Continuous gravitational waves: calculating the mismatch

necessarily equal) time subdomains labelled ta , will result in a rank 4 tensor. Then, given
a set of {∆λαa } which together with the decomposition of the time into subdomains
defines the signal and template evolution, the fully-coherent metric-mismatch can be
calculated from Eqn. (6.49).

6.4.3 Explicit calculation of the metric

In this final subsection, we will derive two explicit calculation of the metric given choices
for the reference times. These will be used later on in this thesis.

6.4.3.1 Reference times in the middle of the subdomains

If there are Nsd subdomains with the ath subdomain of duration ∆Ta then we can bound
the integration over each subdomain by ta , ta + ∆Ta . We are free to choose the reference
time in any way we like; for this example, we define the reference time in each subdomain
to be halfway through, that is taref = ta + ∆Ta /2. Using the following identity
Z    
ta +∆Ta
∆Ta n (1 + (−1)n ) ∆Ta n+1
t − (ta + ) dt ≡ , (6.53)
ta 2 n+1 2

we have that
 π∆Ta3
αβ
Z ta +∆Ta ∆Ta 0 12
 π 2 ∆Ta3 
∂α ∆Φa ∂β ∆Φa dt =  0 3 0  , (6.54)
ta π∆Ta3 π 2 ∆Ta5
12 0 80

and Z  α
ta +∆Ta
a π∆Ta3
∂α ∆Φ dt = ∆T, 0, . (6.55)
ta 12
which can be inserted into Eqn. (6.52) to calculate the metric.

The rank 4 metric gαβab is degenerate in the subdomain indices a and b having only two
distinct terms. We can distinguish these by either a = b or a 6= b. This allows to write
the metric gαβab compactly as
  2 αβ
δab π∆Ta3 π∆Ta4
δab ∆T
Tcoh −
a ∆Ta
0 − 2
 Tcoh Tcoh 12 12Tcoh 
 2 3 
gαβij =  0 δab π3T∆T a
0  . (6.56)
 coh  2 
δab π∆Ta3 π∆Ta4 δab π 2 ∆Ta5 π∆Ta 3
Tcoh 12 − 2
12Tcoh
0 Tcoh 80 − 12Tcoh
Chapter 6 Continuous gravitational waves: calculating the mismatch 147

6.4.3.2 Reference times at the start of each subdomain

If instead we set the reference time to be at the start of each subdomain then, following
the method used to derive Eqn. (6.56), the metric can be written compactly as

2
α,β
δij N −1 −N −2 π∆T δij N −1 −N −2  ∆T3 π  δij N −1 −N −2 

−1 −1 −2 
gαβij = π∆T δij N −1 −N −2 π 2 ∆T 2 δij 4N3 −N −2 π 2 ∆T 3 δij N2 − N3  .

    
∆T 2 π −1 −2 2 3 N −1 N −2 2 4 N −1 N −2
3 δij N −N π ∆T δij 2 − 3 π ∆T 5 − 9
(6.57)
Chapter 7

Glitches in continuous
gravitational waves

Continuous gravitational waves from neutron stars could provide an invaluable resource
to learn about their interior physics. One common detection method involves matched
filtering a modelled template against the noisy gravitational wave data to find unknown
signals. As introduced in Chapter 6, this method suffers a mismatch (a loss of signal
to noise ratio) if the unknown signal deviates from the template. One significant way
this may happen is if the neutron star undergoes a glitch, a sudden rapid increase in
the rotation frequency as introduced in Section 2.1, a phenomenon seen in the timing
of many radio pulsars. While the mechanism which causes pulsars to glitch is not fully
understood, it is likely that any continuous gravitational waves emitted by the star will
also be affected by the glitch since both are intimately related to the neutron star crust.

In this chapter, we use information on the rate and size of pulsar glitches, as deduced
from the observed population of glitching radio pulsars, to estimate the potential mis-
match introduced when searching for gravitational waves from neutron stars whose rota-
tional timing is unknown. We intend to publish this chapter, along with the generalised
mismatch defined in Sec. 6.4, in due course.

7.1 Introduction

Electromagnetic (EM) observations of pulsar glitches have long been one of the most
fruitful sources of insight into neutron star physics. Glitches are characterised by a
sudden increase in the rotation frequency, often accompanied by a jump in the frequency
derivative and an exponential recovery of some fraction of the initial frequency jump.
The events happen rapidly and are sufficiently disruptive that pulsar timing models
often loose phase coherence over the event.

Two leading models exist to explain glitches. In the superfluid pinning model, some
portion of the interior superfluid is pinned, and does not participate in the smooth
torque-driven spin-down of the rest of the crust (where ‘crust’ refers to the actual crust,
149
150 Chapter 7 Glitches in continuous gravitational waves

plus whatever other parts of the star that are strongly coupled to it). After some
period, the crust will therefore have developed a frequency lag compared to the pinned
superfluid. A glitch occurs when the two components recouple, transferring angular
momentum from the pinned superfluid to the crust and producing a spin-up of the crust
[16, 15]. Alternatively glitches could be caused by crust cracking as the crust readjusts
to a minimum energy configuration brought about by the gradual decay of the spin-down
rate [28]. It is also possible that glitches result from a combination of these two models;
evidence for this was found by Melatos et al. [123]. In either case it seems reasonable to
assume that both the crust and the core will be involved.

Rotating isolated neutron stars can produce continuous gravitational-wave (GW) emis-
sion from non-axisymmetric distortions, colloquially also known as ‘mountains’. These
require the mountain to be supported by either elastic stresses in the crust or magnetic
fields. In this model, the star emits a monochromatic GW at a frequency fs which is
twice the rotation frequency fs , i.e, fs = 2ν.

It is unclear exactly how a glitch, as observed in the EM channel, will manifest in the
GW channel. Observing a glitch in both the EM and GW channels would provide a
unique opportunity to investigate the neutron star interior, but to do this we must first
detect continuous GW emission.

Estimates for the intrinsic gravitational wave strain amplitude h0 for canonical models
of GW emissions (see for example Abbott et al. [11]) suggest they are extremely weak
compared to the noise level of advanced detectors [5]. As a result, in order to detect a
signal, significant effort has been put into advanced data analysis methods, which may
be capable of identifying the putative signals. Many of these methods rely on matched
filtering in which a template is correlated with the data with the hope of detecting the
presence of the unknown signal similar to the template. The power of these methods
is that the signal-to-noise ratio (SNR) scales as the square-root of the observation time
[140]. Due to the longevity of GW signals, this allows the weak signal to be discerned
from the noise.

These methods are powerful, but harbour a vulnerability in any instance where the
template, a Taylor expansion in the phase usually up to second order, does not match
the signal. For GW signals from non-axisymmetric distortions of neutron stars, we can
expect that discrepancies from such a template may manifest in one of two ways. Firstly,
we know from radio pulsar timing that the spin-down of a pulsar differs from a Taylor
expansion due to timing noise. We will consider the effect of timing noise in Chap-
ter 8 and Chapter 9; in this chapter we will address the second potential manifestation:
glitches.

There are two distinct questions to answer in the case of glitches: ‘how probable is
it that a glitch will occur during our GW observation?’, and ‘if a glitch does occur,
what effect will it have on our ability to discover the GW signal?’. To answer these
questions, we first use known radio pulsar glitch statistics to estimate the size and rate
of radio pulsar glitches for the parameter spaces considered in continuous GW searches.
Then, we assume that the EM and GW channels are locked in phase, such that a glitch
observed in radio pulsations also exist in the GW signal at twice the phase, frequency
Chapter 7 Glitches in continuous gravitational waves 151

and spin-down rate. Having obtained an estimate for the magnitude and rate of glitches
for potential continuous GW sources, we quantify the effect such glitches will have on
current continuous GW detection methods by calculating the mismatch. We will do
this using the generalised metric-mismatch introduced in Section 6.4. Specifically, we
model a glitch as a piecewise Taylor expansion with a discontinuity at the glitch; we do
not model the exponential recovery observed in some glitches, but we will discuss the
significance this may have in Section 7.5.4. Ultimately, the goal of this chapter is to
estimate the risk faced by current and ongoing continuous GW searches to glitches in
their target population.

The remainder of this chapter is organised in the following way. We begin in Section 7.2,
with a description of current GW searches and how glitches may effect them. Then, in
Section 7.3 we investigate the statistical properties of the observed radio pulsar glitches.
In Section 7.4 we calculate, given the properties of a single glitch, the corresponding
mismatch for a semi-coherent and fully-coherent search. To give quantitative results
relevant to real GW searches, in Section 7.5 we use these calculations to transform the
predicted glitch sizes into predicted mismatches and rates for GW searches.

7.2 Continuous gravitational-wave searches

Searches which target a particular source making use of the EM emission (for example
the Abbott et al. [13] targeted search for the Crab pulsar) are able to handle the epoch
of a glitch, either by avoiding searches over the glitch, or allowing for a jump in the
timing solution at that point [13]. By this merit, such searches have a very low risk
of being disrupted by a glitch coupled to the EM channel, provided the GW channel
follows closely the phase evolution of the EM channel.

In contrast, blind GW searches which, by definition, search for signals without an EM

counterpart do not have any such prior knowledge. This category of searches includes
both directed searches where a small patch of sky is searched in which a neutron star
is believed to exist (e.g. the Aasi et al. [6] search for continuous GWs from supernova
remnants), and all-sky searches; in both instances a band of frequencies and frequency-
derivatives are usually searched since they are inherently unknown. Typically, these
searches match filter against smooth templates built from a Taylor expansion in the
phase; as such the templates do not include glitches. If a neutron star, emitting de-
tectable levels of GW emission, undergoes a glitch in the GW channel, then the matched
filtering method may fail because the template is a poor match to the real, glitching
signal.

All-sky and targeted blind searches are most notably at risk of glitch disruption and
so it is these which we fill focus on in this chapter. Continuous GW searches ideally
employ a fully-coherent search which consists of matched filtering the template against
the data over a coherence time Tcoh . However, such a search is computationally inten-
sive and so GW searches typically revert to using a semi-coherent method in which the
total observation time Tspan is divided into Nseg segments each of duration Tcoh . Each of
152 Chapter 7 Glitches in continuous gravitational waves

these is then fully-coherently searched and then recombined to give a semi-coherent mea-
surement which is insensitive to phase jumps between segments. This method provides
more sensitive searches at fixed computing cost [142]. Typically, a semi-coherent search
is performed first, then interesting candidates are followed-up, reducing the number of
segments, aiming to finally detect the signal with a fully-coherent search; see Shaltev
and Prix [153] for a two-stage follow-up procedure.

A variety of targeted and all-sky searches have already been completed, with more
ongoing and planned for the advanced detector era. In Table 7.1 we list the search
parameters used in some recent searches.

fs f˙s Tcoh Tspan Ref

[Hz] [nHz/s] [hrs] [days]
S5 E@H all-sky 50, 1190 -2, 0.1 25 694 [3]
S5 E@H galactic center 78, 496 -71, 0 11.5 302 [1]
S5 all-sky 50, 1000 -0.89, 0 0.5 365 [2]
VSR low-freq. all-sky 20, 128 -10, 0.15 2.3 185 [7]
S5 supernova remnant (Cas A) 91, 573 -60.5, -1.6 N/A 8.4 [6]

Table 7.1: Summary of some recent all-sky and directed searches. All numbers
are approximate, see the references for exact ranges. Note that S5 refers to
data from the fifth LIGO science run, similarly VSR refers to data from the
second and fourth Virgo science runs. For the S5 supernova remnant search we
present parameters for the Cassiopeia A search only and note that this was a
fully-coherent search with no semi-coherent stage.

To make practical use of the results obtained in this chapter, we will make predictions for
the glitch rate and magnitude in the context of typical GW searches. As an illustrative
example, we will use the search parameters that were used in LIGO’s Einstein@Home
all-sky search for periodic gravitational waves in the fifth science run (S5) data [3],
hereafter referred to as the E@H search. This search used the semi-coherent Hough-
transform method [99] and is typical in terms of the timing parameters, although in
future searches it is expected that the number of segments will be decreased due to
improvements in computing power. We find in Section 7.4 that the effect of a glitch is
independent of the sign of the frequency derivative, and so for convenience we truncate
the GW spin-down parameter space to f˙s = [−2, 0] × 10−9 Hz s−1 .

7.3 Statistical properties of the observed glitch database

In this section, we will discuss the properties of glitches in the observed radio pulsar
population using the glitch catalogue maintained by Espinoza et al. [59] and available
at www.jb.man.ac.uk/pulsar/glitches.html. Our goal is to make a statement about
how often glitches occur and their magnitudes for the types of neutron star which may
be emitting GWs in the parameter space of typical GW searches. This task is made
difficult since many searches look for young, rapidly rotating stars for which we only
have a small sample of observations. Therefore, we must moderately extrapolate to
apply our results to the observed radio pulsar population.
Chapter 7 Glitches in continuous gravitational waves 153

Radio pulsar timing methods detect glitches by fitting a piecewise Taylor expansion in
the phase on either side of the event, with a modelled jump in between (see Edwards
et al. [57] or Section 2.1 for a detailed discussion). The glitch catalogue reports 472
events from 165 pulsars (as of the 27th of June 2016); for each of these events a value is
reported for the frequency jump δν and frequency derivative δ ν̇, if it can be measured.
We cross-reference the glitch catalogue with the ATNF Manchester et al. [117] pulsar
catalogue available at www.atnf.csiro.au/people/pulsar/psrcat/ in order to get the
glitching pulsar’s timing properties.

Of the 472 listed glitches, we find 15 with no ATNF cross-reference, 1 with δν < 0, and
4 with no measured ν̇ in the ATNF catalogue; these pulsars are removed from our data
set. Additionally, we find 54 glitches which have either no measured δ ν̇, or a measured
value consistent with 0. These will be included where possible.

7.3.1 Glitch magnitudes

Espinoza et al. [59] argued that the glitch catalogue contains glitches from two distinct
sub-populations of pulsars. There is the main population with δν magnitudes ranging
from 10−9 to 10−5 Hz, and a second, less numerous population with larger magnitudes of
δν. The smaller population are described as ‘Vela-like’ in that the pulsars that undergo
these glitches have similar characteristic ages and magnetic fields to the Vela pulsar
(B0833-45). We reproduce the evidence for this finding in Figure 7.1 where we plot
the histogram of all observed δν values. This demonstrates the bimodality found by
Espinoza et al. [59].

To check that the bimodality is not an artefact of the histogram bin sizes we estimate the
probability density function using a Gaussian kernel density estimate (KDE). Specifically
we use the Jones et al. [92] implementation. This is also plotted in Figure 7.1 and
demonstrates two distinct peaks, although the smaller peak could be interpreted as two
modes close together.

By eye, it is clear that there are at least two modes to the histogram. However, it is
also possible that there may be more modes with less distinct differences in their mean.
We investigate this issue in Appendix 7.A by applying a Bayesian model comparison
to Gaussian mixture models (see Chapter 18 of Gelman et al. [64] for an introduction
to mixture models) varying the number of components and also allowing for a skew
as described in O’Hagan and Leonard [129]. We find that all models with 2 or more
components fit the data decisively better than a single component. Marginal gains are
found by allowing the models to be skewed and have 4 or more components, but no
single model is outstanding amongst the others. For this reason we choose to use a
2-component model with skew; this provides a good empirical description of the data
and is pragmatic in that we limit the number of components to 2 for interpretability.
We note that this description is empirical and we do not intend to make any substantive
claim relating the two components to the two mechanisms discussed in Chapter 2.1.

Having obtained a fit to the magnitude of the glitch in frequency using a 2-component
skewed-Gaussian mixture model, we use this fit to label each data point as originating
154 Chapter 7 Glitches in continuous gravitational waves

from one of the 2 skewed-Gaussian components . Specifically, to each data point we

assign the label based on the maximum probabilities of each of the two components,
given the maximum posterior model parameters derived in the fitting process.

The resulting mixture components and individual distributions are plotted in Figure 7.1
and in Table 7.2 we provide the resulting mean, standard deviation, weights and skewness
of the two components in log-space. This method identifies the two subpopulations in a
manner consistent with the observations by Espinoza et al. [59] and notably the Vela-like
component suffers a significant skew.

Mean Std. Dev. Weight Skew

Normal -8.413 1.590 0.701 1.084
Vela-like -4.407 0.534 0.299 -9.660

Table 7.2: Log-normal mean, standard deviation and weight for the components
of the mixture model as fitted in Figure 7.1.

Figure 7.1: The distribution of glitch magnitudes δν observed in the glitch

catalogue. This is given as both a binned histogram and a Gaussian KDE,
as discussed in the text. The coloured lines mark the two components of the
skewed Gaussian mixture model fitted to model the bimodality.

In Figure 7.2, we plot histograms for δν and δ ν̇ along with the raw data in a scatter
plot. We have separated the data into the individual sub-populations, as labelled by the
Gaussian mixture model, and colour coded to aid the eye. Several pulsars of interest are
picked out using coloured halos. It is of general interest that not all the Vela glitches are
categorised by our method as Vela-like; this can be seen by looking at the distribution
of Vela glitches in Figure 7.2.

7.3.2 Overview of the population of glitches

To give an overview of all observed glitches in the context of the whole population of
observed radio pulsars listed in the ATNF catalogue, in Figure 7.3 we plot two copies of
the familiar ν-ν̇ diagram. In panel A, for each pulsar which has been observed to glitch
Chapter 7 Glitches in continuous gravitational waves 155

Figure 7.2: Glitch magnitudes as provided by the glitch-database [59]. This

shows a scatter plot of all pairs of δν and δ ν̇ where the colouring depends
on the labelling given by the mixture model. Purple circles are the points
categorised as ‘normal glitches’, while green circles are the points from the
‘Vela-like’ population. Histograms for both glitch magnitudes are also given for
each sub-population. Coloured halos highlight glitches from interesting pulsars.

we add a coloured circle. Those pulsars which glitched multiple times are marked by
a larger coloured circle with the area proportional to the number of glitches. In panel
B, we again mark each pulsar that has been observed to glitch with a coloured circle,
but here the area of the coloured halo marks the pulsar’s average glitch magnitude.
For both plots, different colours have been used to partition the Vela-like and normal
glitches (note that some pulsars display glitches from both populations) and a shaded
box marks the parameter space searched by the S5 E@H search. Finally, dashed lines
mark isoclines of constant characteristic age as defined by τage = |ν/ν̇|.

While the bulk of observed glitching pulsars are from the main pulsar population, the
fraction of young pulsars (τage < 105 yrs) which glitch is proportionally higher than
in the normal population. Vela-like glitches occur predominantly in the young pulsars
with none seen in pulsars with τage > 107 yrs. It is also noticeable that younger pulsars
display a greater number of glitches. Note that, since we have not observed all pulsars for
the same duration, one cannot infer the relative glitch rate from the number of glitches
alone.
156 Chapter 7 Glitches in continuous gravitational waves

Figure 7.3: Frequency-frequency derivative plot of all pulsars in the ATNF

catalogue [117]. A: A single coloured point marks pulsars which have been
observed to glitch; the area of the coloured halo is proportional to the number
of observed glitches from that pulsar. B: A single coloured point marks pulsars
which have been observed to glitch, the area of the coloured halo is proportional
to the average glitch magnitude from that pulsar. We have used purple for
‘normal’ glitches and green for ‘Vela-like‘ glitches, as defined by the skewed
Gaussian mixture model in Section 7.3.1. Note that, for the glitch magnitudes,
the relative scaling for the Vela-like and normal populations are not the same
since the Vela-like pulsars are significantly larger: the area representing the
normal glitch magnitudes are scaled 3 times larger than the Vela-like glitch
magnitudes. The grey shaded box marks the parameter space of typical GW
all-sky searches which cover a rotational frequency ν range of 10 − 600 Hz
(assuming they search for signals with fs = 2fs ).

For the normal-glitch population, Espinoza et al. [59] noted that pulsars with τage <
5 × 103 yr undergo small or medium sized glitches (δν < 105 Hz). It is postulated
that the higher temperatures in younger pulsars prevents the glitch mechanism working
effectively. This effect is consistent with Figure 7.3B: the pulsars with the largest average
glitch sizes have τage ∼ 105 yrs, while younger pulsars tend to exhibit smaller glitches
on average.

7.3.3 Extrapolating: glitch magnitudes

We would like to be able to predict the glitch magnitude for the unobserved pulsar
population targeted by GW searches. In particular, we need to extrapolate up to the
large values of ν̇ searched for in many all-sky GW searches, where few observed radio
pulsars exist.
Chapter 7 Glitches in continuous gravitational waves 157

It has previously been found by McKenna and Lyne [120], Lyne et al. [114], Wang
et al. [163], and Espinoza et al. [59] that the glitch activity (defined in the first of these
references) correlates well with |ν̇| and the characteristic age τage . We choose not to
combine the rate and magnitude information together into the activity, but estimate
both separately as these are of most direct relevance to GW searches.

We investigated correlations of the glitch magnitudes δν and δ ν̇ with the frequency,

frequency-derivative and characteristic age. In Table 7.3 we present the Pearson corre-
lation coefficient for each glitch magnitude (δν and δ ν̇). This is done for three groups: all
the data together, then individually for the normal population and the Vela-like popula-
tion. For the normal population, both glitch magnitudes most strongly correlates with

log10 |τage | log10 |ν| log10 |ν̇|

log10 |δν| -0.634 0.538 0.68
All
log10 |δ ν̇| -0.846 0.672 0.88
log10 |δν| -0.631 0.390 0.64
Normal
log10 |δ ν̇| -0.864 0.604 0.88
log10 |δν| 0.035 0.12 0.040
Vela-like
log10 |δ ν̇| -0.62 0.376 0.593

Table 7.3: The correlation coefficient between the glitch magnitudes and the
measured and inferred timing properties of the source pulsar. For each row, the
largest value is highlighted in bold.

spin-down rate ν̇, although we recognise that τage is only marginally worse. In contrast,
δν for the Vela-like population has a weak correlation with all predictor variables, but
δ ν̇ does correlate well showing the strongest correlation with the characteristic age. We
choose to use ν̇ as a predictor variable for both the normal and Vela-like populations.
For the latter the correlation coefficient suggests that τage may be a better predictor,
however ν̇ is only marginally worse and it makes it simpler to interpret later results if
the same predictor is used for both populations. In practise, our results will be robust
to either choice of the predictor variable.

In Figure 7.4A and Figure 7.4B we scatter-plot the glitch magnitudes against the spin-
down rate of the pulsar to demonstrate the correlation. For both plots we have added
coloured halos to label several interesting pulsars. These help to show that there can be
almost as much variation in the glitch magnitude of a single pulsar as from the entire
population.

Fitting a linear function in log-log space (see Appendix 7.B for details) our resulting
fitting formulae for the frequency jump due to each separate population is

hδνiNormal =10−0.89 |ν̇|0.55 10±0.93 ,

(7.1)
hδνiVela-like =10−4.62 |ν̇|0.01 10±0.28 .

and for the frequency-derivative jumps is

hδ ν̇iNormal =10−4.16 |ν̇|0.90 10±0.67 ,

(7.2)
hδ ν̇iVela-like =10−7.06 |ν̇|0.57 10±0.66 .
158 Chapter 7 Glitches in continuous gravitational waves

Figure 7.4: A: the magnitude of δν as a function of the source pulsar’s spin-

down rate. B: the absolute value of the magnitude of δ ν̇ against the measured
spin-down rate. The coloured lines and shaded bands are the best fits from
Eqn. (7.1) for δν, and Eqn. (7.2) for δ ν̇; the green lines mark the Vela-like fit
while the purple lines park the fit to the normal population. Vertical clustering
in the observed data points is the result of multiple glitches observed from a
single source. Coloured halos highlight glitches from some interesting pulsars.

Note that the last factor here provides an estimate of the variability about the linear
fit, while neglecting this term gives the mean. We plot both of these fitting formulae in
Figure 7.4; the estimate of the variability is indicated by a shaded band.

Taking these fitting formulae, in Figure 7.5 we plot the predicted glitch magnitudes (in
frequency and spin-down rate respectively) over the S5 E@H search parameters. We have
similarly transformed the variation estimate and plotted it as a shaded band. These fits
do not provide a precise statement about the magnitude of glitches, but are sufficient to
estimate the order-of-magnitude that we might expect.

7.3.4 Extrapolating: average glitch rate

In order to estimate the average rate of glitches, Espinoza et al. [59] grouped pulsars
by their spin-down rate ν̇, including pulsars which have not yet been observed to glitch.
From this grouping, the authors used the measured number of glitches Ng to calculate
Chapter 7 Glitches in continuous gravitational waves 159

Figure 7.5: Predicted glitch magnitude of δν, in the upper panels, and δ ν̇, in
the lower panels. These are plotted over the E@H parameter space using the
fitting formulae of Eqn. (7.1) and Eqn. (7.2). The shaded bands indicate the
estimated error neglecting correlations.

a mean glitch rate hṄg i. In Figure 10 of their work they show that, to a good approx-
imation, in log-space the mean glitch rate depends linearly on the spin-down rate; we
reproduce this in Figure 7.6 using the data from Table 4 of Espinoza et al. [59].

In order to extrapolate, we perform a linear regression to the data in log-space following

the method described in Appendix 7.B. This gives a prediction of the glitch rate as

hṄg i =10−3.00 |ν̇|0.47 10±0.31 s−1 , (7.3)

where ν̇ is measured in Hz/s. The exponent agrees with that found by the original
authors (they do not provide the prefactor).

We plot this average glitch rate in Figure 7.7A over the range of f˙s values considered
in the S5 E@H search. We also multiply this rate by the span of a typical search,
Tspan ≈ 1 yr, to obtain λ(Tspan =1yr), the expected number of glitches during the search;
this is also plotted in Figure 7.7A, on the right-hand axis.

Interpreting this average glitch rate requires a substantive physical model. Melatos et al.
[123] demonstrated that glitch waiting times are consistent with an avalanche process
transferring angular momentum from the core superfluid to the crust. Choosing 9 pulsars
which had glitched 5 times or more, they found that 7 of these where consistent with
a constant rate Poisson process such that each glitch event is statistically independent.
In the remaining two, J0537-6910 and B0833-45 (Vela), they find that a quasiperiodic
component coexists with the Poisson process and accounts for about 20% of the events.

Of the glitch catalogue, J0537-6910 accounts for 23 and B0833-45 (Vela) for 17 of the to-
tal 472 events. Assuming that 20% of these are due to the quasiperiodic component, this
160 Chapter 7 Glitches in continuous gravitational waves

Figure 7.6: Reproduction of Figure 10 from Espinoza et al. [59] given as both
log of the glitch rate per second (left axis) and per year (right axis). Black dots
are the original data points, the solid line and shaded region are our best-fit
line and a measure of the variability as given in Eqn. (7.3). A vertical dashed
line marks −2 × 10−9 , the largest absolute spin-down rate used in the S5 E@H
search.

is ∼ 1.7% of the total number of observed glitches. It is possible that other pulsars also
exhibit a quasiperiodic component, so the total fraction of glitches from a quasiperiodic
component should be greater that 1.7%. However, we may still claim that the majority
of glitches in the catalogue are due to a Poisson like process.

Assuming that all glitches used to estimate hṄg i are due to a Poisson process, we can
calculate the probability of one or more glitches occurring during a typical search lasting
1 yr as a function of f˙s . To do this, we take the estimated number of glitches during
a typical search λ, as given on the right axis of Figure 7.7A, and sum the Poisson
probability mass function from 1 to infinity
∞
X λNg e−λ
P (Ng ≥ 1; Tspan ) = , (7.4)
Ng !
Ng =1

where the dependence on Tspan is in λ = λ(Tspan ). Note that in practise we truncate the
summation at a finite level where the mass function is negligible.

In Figure 7.7B we plot this probability using λ from the upper plot. This suggests that
for almost all of the E@H spin-down rate parameter space, it is more probable to have
at least one glitch per year than to have none. This estimation will suffer bias from the
inclusion of the quasiperiodic components into the calculation of hṄg i. Nevertheless, we
hope it provides some quantitative measure of the probability of a glitch occurring.
Chapter 7 Glitches in continuous gravitational waves 161

Figure 7.7: A: the average glitch rate against the spin-down rate f˙s over the
range of values used in the S5 E@H search. On the right axis is the corresponding
average number of glitches for a search of duration 1 yr. B: the probability of
observing one or more glitches during a search lasting for 1 yr assuming a Poisson
distribution.

7.4 Calculating the mismatch due to a single glitch

In Section 7.3 we established the statistical properties of observed radio pulsar glitches
and provided fitting formulae for the magnitude and rate of glitches. In this section,
we will convert these results into a quantified effect on fully-coherent and semi-coherent
GW searches using the tools developed in Section 6.2 and Section 6.4.

For this work, we consider a glitch to consist of an instantaneous jump in the phase,
frequency, and frequency derivative of the GW signal; the magnitudes of these quantities
we will denote by δφ, δf, and δ f˙. Since we don’t yet understand the glitch mechanism
and what happens during the glitch, it’s unclear how meaningful it is to discuss a jump
in the GW phase. Nevertheless, modelling the signal with a piecewise Taylor expansion
naturally includes such a phase jump, therefore we will include it in our discussion. The
size of the jump in the GW frequency and frequency derivative can be estimated from
the observed jumps in Section 7.3 (assuming δ f˙ = 2δν etc.). Aside from these jumps
there may also be an exponential relaxation in the post-glitch dynamics and a rise-time
during which the glitch occurs; for now all such phenomena will be ignored, but we
return to this in Section 7.5.4.
162 Chapter 7 Glitches in continuous gravitational waves

Our glitch signal can be modelled by a piecewise Taylor-expansion with two subdomains.
We denote the time of the glitch as RTspan where R is a dimensionless fractional quantity
such that R ∈ [0, 1] and Tspan is the observation time of the search. Then labelling the
period before the glitch as A and after as B, the parameter offsets may be written
(
αa ∆λαA if t < RTcoh
∆λ (t) = , (7.5)
∆λαB if t > RTcoh

where ∆λαj = λαj αj

s − λt . We will set the reference times for each subdomain half-
way through the subdomain and also define a global reference time tref at the glitch,
RTcoh . To help orient the reader with these choices, we provide a schematic of the signal
frequency over the glitch in Figure 7.8.

Figure 7.8: Illustration of the references times and frequency jump over the
glitch. Note that the global reference time tref is set to coincide with the time
at which the glitch occurs; in this instant it is half-way through, but could in
general be at any time during the observation. The local reference times are set
halfway through each subdomain.

The jumps at the glitch itself may then be parameterised as

δφ = φB A
s (RTcoh ) − φs (RTcoh )
δf = fsB (RTcoh ) − fsA (RTcoh ) (7.6)
δ f˙ = f˙B − f˙A
s s

The time-dependence here indicates that one must account for the changes due to lower
order terms between the reference time, halfway through the subdomain, and the glitch.
Since we do not consider second-order spin-down terms, this is not required for the δ f˙
calculation.

In the following sections, we consider fully-coherent and semi-coherent GW searches for

this signal which contains a single glitch. In general these searches are performed over a
grid of Taylor expansion templates in the frequency and spin-down parameter space and
then, if some subset of the templates return a sufficiently high SNR, the template with
Chapter 7 Glitches in continuous gravitational waves 163

the smallest mismatch is taken as a candidate. This method minimises the mismatch
with respect to the search parameters ft and f˙t ; we will include this minimisation step
in the calculation.

7.4.1 A single glitch in a fully-coherent search

7.4.1.1 A fully-coherent search: analytic metric-mismatch

To calculate the fully-coherent mismatch, we first expand the summation in Eqn. (6.49)
over the two subdomains giving three terms

µ̃ =gαβAA ∆λαA ∆λβA + 2gαβAB ∆λαA ∆λβB + gαβBB ∆λαB ∆λβB . (7.7)

To calculate the metric components, we use Eqn. (6.56) with ∆T A = RTcoh and ∆T B =
(1 − R)Tcoh . We then insert the parameter offsets ∆λαa from Eqn. (7.5), and explicitly
write the mismatch in terms of the components of ∆λαa = [∆φa , ∆f a , ∆f˙a ]. After some
manipulation, the mismatch between the glitch signal and an arbitrary Taylor expansion
template is

π 2 Tcoh
2  
µ̃ =R(1 − R)(∆φA − ∆φB )2 + R3 (∆f A )2 + (1 − R)3 (∆f B )2
3
4 h
π 2 Tcoh i
+ R5 (9 − 5R)(∆f˙A )2 − 10R3 (1 − R)3 ∆f˙A ∆f˙B + (1 − R)5 (5R + 4)(∆f˙B )2
720
2 h
πTcoh i
+ R(1 − R)(∆φB − ∆φA )((1 − R)2 ∆f˙B − R2 ∆f˙A ) .
6
(7.8)

The first three terms are the independent contributions to the mismatch from the jumps
in phase, frequency, and spin-down; the last term is the mixture term between the phase
and the spin-down. This expression has a maximum when R = 1/2, and collapses to the
usual mismatch between two smooth Taylor expansions when R is 0 and 1. Furthermore,
we verified that when the magnitude of the glitch is zero, one recovers the usual fully-
coherent mismatch between two smooth Taylor expansions.

The mismatch in Eqn. (7.8) is a function of the individual signal parameter λαa s , and
αa
the template parameters λt . We are not interested in arbitrary choices of the template
parameters, but those which minimise the mismatch as would be found by searching over
a small area in parameter space and selecting the template with the smallest mismatch
as a detection candidate. We therefore analytically minimise Eqn. (7.8) with respect to
the template parameters ft and f˙t . We find that

3δφ
ft min =fsA + δf (1 − R)2 (2R + 1) + R(1 − R)
πTcoh
(7.9)
Tcoh  ˙A 
+ fs (1 − R) + δ f˙(1 − R)3 (1 + R) ,
2
164 Chapter 7 Glitches in continuous gravitational waves

and

f˙tmin =f˙sA + δ f˙(1 − R)3 (6R2 + 3R + 1)

30δf 2 30δφ (7.10)
+ R (1 − R)2 + 2 R(1 − R)(2R − 1),
Tcoh πTcoh

which are expressed at the global reference time tref .

Inserting these into Eqn. (7.8) and simplifying yields a minimum fully-coherent mismatch
of

µ̃ =R(1 − R)(4R(1 − R)(5R(1 − R) − 2) + 1)δφ2

π 2 Tcoh
2
+ R3 (1 − R)3 (4 − 15R(1 − R))δf 2
3
π 2 Tcoh
4
+ R5 (1 − R)5 δ f˙2
5 (7.11)
+ 2πTcoh R2 (1 − R)2 (1 − 2R)(5R(1 − R) − 1)δf δφ
+ π 2 T 3 R4 (1 − R)4 (2R − 1)δ f˙δf
coh
2
πTcoh 3
+ R (1 − R)3 (2 − 12R(1 − R))δφδ f˙.
3
An important distinction must be made here between ∆f a which is the frequency offset
between the signal and template in the ath subdomain, and δf the signal frequency jump
at the glitch. Notably, the minimum mismatch depends only on δφ, δf , and δ f˙; it is
independent of the overall phase, frequency, or spin-down of the signal.

Since the probability distribution of R should be uniform over the search duration, we
can average Eqn. (7.11) over R to get the expectation
2   π2T 4
3 2 Tcoh
hµ̃iR = δφ + π 2 δf 2 − πδφδ f˙ + coh ˙2
δf . (7.12)
70 630 13860

The mixture terms δf δφ and δf δ f˙ vanish in the averaging process.

7.4.1.2 Simple estimates

We now make some rough estimates based on Eqn. (7.12), the R-averaged mismatch for
a fully-coherent search. Firstly, let us consider a glitch which consists of a jump solely in
the phase δφ. For such a glitch, we can calculate the size of a phase-jump which would
produce a mismatch of µ̃ = 0.1
r  
70µ̃ µ̃
δφ = ≈ 1.5 rad (7.13)
3 0.1

For the frequency and spin-down rate jumps, a simple way to quantify the significance
of a given glitch is to ask ‘over what coherence time would a fully-coherent search
accumulate a mismatch of µ̃ = 0.1?’. We will consider each type of jump independently,
Chapter 7 Glitches in continuous gravitational waves 165

since we are only hoping to make order of magnitude estimates. For the frequency jump
we find that
√  1  −1
630µ̃ µ̃ 2 δf
Tcoh = ≈ 2.9 days , (7.14)
πδf 0.1 10−5 Hz

while for the spin-down rate jump

 1  1 !− 1
13860µ̃ 4 µ̃ 4 δ f˙ 2

Tcoh = ≈ 40 days . (7.15)

π 2 δ f˙2 0.1 10−12 s−2

We have parameterised using the largest jumps seen in the glitch-catalogue as can be
seen in Figure 7.2. This is a useful order-of-magnitude estimate and tells us that over
timescales comparable to current and future searches (at least the fully-coherent follow-
up) the mismatch can potentially rise above 0.1. We will investigate the predictions of
Eqn. (7.12) in depth later on in Section 7.5.

7.4.2 A single glitch in a semi-coherent search

Having investigated the mismatch for a glitch in a fully-coherent search, we now consider
the same glitch, but in a semi-coherent search as introduced in Section 6.2.5. The
important point to recall is that the semi-coherent segments must be summed along
the same point in parameter space. In practise, this means all values of the template
parameters λαt must be equal when expressed at the same reference time.

7.4.2.1 A semi-coherent search: analytic metric-mismatch

For a semi-coherent search the observation time Tspan is divided into Nseg equal length
segments of duration Tcoh . We then compute the fully-coherent mismatch in each seg-
ment and then sum the squared SNR along the template parameters to get the squared
semi-coherent SNR. To calculate the semi-coherent mismatch, we can make the simpli-
fying assumption that the glitch occurs exactly at the interface between two segments
such that RNseg is an integer, where R measures the fraction of the observation period
at which the glitch occurs. We will derive the mismatch under this assumption and then
test how and when it breaks down using numerical simulations.

Under this assumption, in each segment both the signal and the template are Taylor
expansions and so we can use the Brady et al. [35] formalism described in Section 6.2.4.
We do not need to use the generalised metric-mismatch developed in Section 6.4.

For each segment we distinguish between the local reference time tjref for the j th segment
and tref the global reference time, taken to be at the glitch RTspan . Then, in the j th
segment, the parameter offsets are calculated by transforming the global parameters to
166 Chapter 7 Glitches in continuous gravitational waves

the local offset:

(
j ft − fsA + (f˙t − f˙sA )(tjref − tref ) if j ≤ RNseg
∆f = , (7.16)
ft − fsB + (f˙t − f˙sB )(tjref − tref ) if j ≥ RNseg
(
f˙t − f˙sA if j ≤ RNseg
∆f˙j = , (7.17)
f˙t − f˙sB if j ≥ RNseg

where tjref = Tcoh (j − 21 ). We do not need to consider the phase jump at the glitch,
since the fully-coherent mismatch is insensitive to an overall phase-offset; if the glitch
did not occur at the interface between two segments then this would not be the case.
The template parameters ft and f˙t are the same in each segment; this reflects the fact
that the semi-coherent search sums the SNR along a smooth Taylor expansion template.

We calculate the metric for a fully-coherent search with the reference time half-way
through from Eqn. (6.12); then for the j th segment, we calculate the fully-coherent
mismatch to be

π 2 Tcoh
2 π 2 Tcoh
4
µ̃j = (∆f j )2 + (∆f˙j )2 (7.18)
3 180

Inserting this into Eqn. (6.17) we average over all the segments with the parameter
offsets given by Eqn. (7.16) and Eqn. (7.17), to calculate the semi-coherent mismatch µ̂;
for brevity we do not provide the result here.

Next, we minimise µ̂ with respect to the global ft and f˙t to select the minimum mismatch.
The minimising values are given by
2 (R(9 − 6R) − 4) + 4)
R(5Nseg
ft min =fsB + 2 −4
δf
5Nseg
(7.19)
2R(1 − R)
+ 2 −4
2
(5Nseg R(1 − R) − 1)δ f˙Tcoh ,
5Nseg

and
2 R(3 − 2R) + 4)
R(5Nseg
f˙tmin =f˙sB + 2 −4
δ f˙
5Nseg
(7.20)
R(1 − R)Nseg
+ 30 2 − 4)
δf.
Tcoh (5Nseg
Chapter 7 Glitches in continuous gravitational waves 167

Inserting these back into the mismatch µ̂, we calculate the minimised mismatch and
express it in terms of the glitch parameters δf and δ f˙
2 (3R(1 − R) − 1) + 4
5Nseg
µ̂ =R(R − 1) 2 − 12
π 2 Tcoh
2
δf 2
15Nseg
+ R(1 − R)π 2 T 4 δ f˙2
coh
4 R2 (1 − R)2 + N 2 (5R(R − 1) − 5) + 4
25Nseg (7.21)
seg
× 2 − 180
225Nseg
3 R2 (1 − R)2
5Nseg
+ 2 − 12
(2R − 1)π 2 Tcoh
3
δf δ f˙.
15Nseg

Assuming that the glitch should occur with the same probability at any point during
the observation, we average over R and take the limit of Nseg being large. This gives

π 2 Tcoh
2 π 2 Tcoh
2 T2
span ˙2
hµ̂iR = δf 2 + δf . (7.22)
45 1260
In the last step we substituted Nseg = Tspan /Tcoh .

7.4.2.2 Checking the validity of the assumption

Eqn. (7.21) relies on the assumption that the glitch occurs exactly at the interface be-
tween two semi-coherent segments. This assumption is valid in the limit for which
Nseg  1, but will be imprecise for small numbers of segments. To demonstrate
this, in Figure 7.9 we plot the fractional absolute difference between the R-averaged
metric-mismatch approximation hµ̂iR , as given in Eqn. (7.22), and the R-averaged ex-
act mismatch (hµ̂iR )exact which we calculate numerically. To test both of the terms in
Eqn. (7.22) we plot the residual difference for two glitches: one in the frequency and the
other in the spin-down rate, the exact values are given in the legend. This figure shows
that for Nseg > 10 the error due to our assumption that the glitch occurs exactly at the
interface between two segments of the semi-coherent search decreases. The error which
persists for Nseg >> 1 is due to the metric-mismatch approximation which is expected
and understood.

7.4.2.3 Simple estimates

Taking Eqn. (7.22), the R-averaged semi-coherent metric-mismatch, we can estimate the
maximum coherence time for a semi-coherent search allowing for a mismatch µ = 0.1.
For the jumps in frequency this gives
√  1/2  −1
45µ̂ µ̂ δf
Tcoh = ≈ 0.782 days , (7.23)
πδf 0.1 10−5 Hz
168 Chapter 7 Glitches in continuous gravitational waves

Figure 7.9: Fractional difference between the R-averaged metric-mismatch ap-

proximation hµ̂iR , as given in Eqn. (7.22), and the R-averaged exact mismatch
(hµ̂iR )exact . The exact mismatch is calculated numerically and does not suf-
fer from either the metric-mismatch approximation or the assumption that the
glitch occurs exactly at the interface between two segments of the semi-coherent
search.

while for the jumps in frequency-derivative this gives

√
1260µ̂
Tcoh =
πδ f˙Tspan
 1  −1 !−1 (7.24)
µ̂ 2 Tspan δ f˙
≈ 0.689 days
0.1 694 days 10−12 Hz

These values are less than the coherence time used in the S5 E@H search, which was
25 hrs. These simple estimates therefore suggest that for the largest observed glitches,
semi-coherent searches may suffer a non-negligible mismatch. We will investigate this
and the fully-coherent mismatch case more rigorously in the next section.

7.5 Predicting the mismatch and rate of glitches in gravi-

tational wave searches

In this section, we will use the fitting formulae of Section 7.3.1 to predict the magnitude
of glitches for a range of parameters typical of all-sky searches. We will use the tools
of Section 7.4 to transform the magnitudes into an estimate of the fully-coherent and
semi-coherent mismatch assuming a single glitch occurred during the search. However,
this mismatch does not give a complete picture of the risk, since we must also consider
the predicted rate of glitches. Specifically, in converting a predicted glitch magnitude
into an estimate for the fully-coherent and semi-coherent mismatch we will make use
of Eqn. (7.12) and Eqn. (7.22): the R-averaged mismatch assuming that a single glitch
occurs uniformly and at random during the observation time. We can therefore take
these results as a lower bound when there is likely to be one or more glitches. On the
Chapter 7 Glitches in continuous gravitational waves 169

other hand, in regions where the chances of a glitch are low, a large mismatch only
indicates that the signal would be lost in the rare event that a glitch had occurred. To
provide the reader with both pieces of this puzzle, we will present results on both the
expected mismatch due to a single glitch and the probability of one or more glitches
occurring. This will be done for each population of glitches separately. Let us define

λnormal (ν̇, Tspan ) = wnormal hṄg iTspan (7.25)

λVela-like (ν̇, Tspan ) = wVela-like hṄg iTspan , (7.26)

as the expected number of normal and Vela-like glitches where wnormal and wVela−like
are the weights of the two populations as given in Table 7.2 and hṄg i is the fitted
glitch rate as a function of ν̇ which is given by Eqn. (7.3). Notice that we have made a
prior specification here that the proportion of normal and Vela-like pulsars in the target
population is the same as in the observed population. There is some evidence that in
fact the proportion of Vela-like pulsars increases with ν̇; this could be modelled by a ν̇-
dependent weighting, however we will ignore this effect here. Eqn. (7.25) and Eqn. (7.26)
can be transformed into the probability of one glitch or more occurring during the search
by substitution into Eqn. (7.4). In the following discussion it is this probability which
we will report on.

We note that in reality, the true mismatch for fully coherent searches will always be
bounded by [0, 1], but the results of this section can exceed this since we are using
the metric-mismatch approximation. Similarly, for the semi-coherent searches, it is in
fact bounded by the minimum of R and 1 − R; this can be seen by realising that
if the glitch is sufficiently large, the minimum mismatch is achieved by fitting to the
semi-coherent segments on only one side of the glitch. Here, we report the simpler
results of naively applying Eqn. (7.12) and Eqn. (7.22). The results can therefore be
interpreted by realising that if the metric-mismatch is greater than 1, there will be
large true mismatch, although the exact value will depend on where that glitch occurred
during the observation span.

7.5.1 Fully-coherent searches

In this section, we present results on the expected mismatch and expected number of
glitches for fully-coherent searches. In Figure 7.10 we plot contours of fixed mismatch
as a function of the coherence time Tcoh and the GW spin-down rate f˙s . Alongside are
plotted the number of glitches as a function of the observation time and spin-down rate.

These figures illustrate that the predicted mismatch depends on the source population.
For the normal population, larger absolute spin-down rates are associated with larger
glitch magnitudes; as a result larger absolute spin-down rates are predicted to suffer more
severe mismatches. For the Vela-like population, the mismatch is largely independent of
ν̇, this is because the fitting formulae Eqn. (7.1) finds little variation in the glitch size
with the spin-down rate.
170 Chapter 7 Glitches in continuous gravitational waves

Figure 7.10: The estimated fully-coherent mismatch as a function of Tcoh and f˙s .
In the top panel is the prediction based on the fit to the normal population, in
the bottom panel is the prediction based on the fit to the Vela-like population.
Solid black lines provide contours of fixed mismatch. Red dashed lines have been
added to indicate the probability (as a percentage) of one or more normal glitch
occurring calculated by substitution of Eqn. (7.25) (for the normal population)
or Eqn. (7.26) (for the Vela-like populations) into Eqn. (7.4).

For long observation times the mismatch can be severe, however this is only a concern
if a glitch occurs. The expected number of glitches in both populations is, for most of
the parameter space, lower than 1, but not vanishingly so.

We note here that these plots only show the ‘best-fit’ contour lines predicted by our
fitting for glitch sizes. There is substantial variability in the population glitch sizes and
rate, this is reflected in large error bars in the fitting formulae derived in Section 7.3. In
Appendix 7.C we translate these estimates into uncertainties on the mismatch and rate
contour lines.

7.5.2 Semi-coherent searches

For the semi-coherent search we repeat the process of predicting the glitch magnitudes,
but then use Eqn. (7.22) to transform them into predictions for the semi-coherent mis-
match. A semi-coherent search has two fixed times: the total observation span Tspan
Chapter 7 Glitches in continuous gravitational waves 171

Figure 7.11: The estimated semi-coherent mismatch as a function of Tcoh and f˙s
for a fixed observation time Tspan = 100 days; see Figure 7.10 for a description
of the contour lines.

and the coherent segment length Tcoh . In Figure 7.11 and Figure 7.12 we present results
at fixed observation times of 100 days and 365 days and show contours of the expected
semi-coherent mismatch and number of glitches as a function of the spin-down rate and
coherence time.

For the semi-coherent mismatch, the contour lines tell much the same story as the fully-
coherent search. The normal population produces a lower overall level of mismatch with
strong f˙s dependence while the Vela-like population has larger mismatches with weak
dependence on f˙s .

As in Figure 7.10, red dashed lines have been added to indicate the probability of one or
more normal glitches occurring, this is calculated by substitution of Eqn. (7.25) (for the
normal population) or Eqn. (7.26) (for the Vela-like population) into Eqn. (7.4). Note
these are not a function of Tcoh since we multiply the glitch rate per unit time by the
fixed observation period Tspan .
172 Chapter 7 Glitches in continuous gravitational waves

Figure 7.12: The estimated semi-coherent mismatch as a function of Tcoh and f˙s
for a fixed observation time Tspan = 365 days; see Figure 7.10 for a description
of the contour lines.

7.5.3 The follow-up stage

A semi-coherent search begins by searching with Nseg segments to find candidates.

Following successful identification of such candidates, these are followed-up by semi-
coherent searches with fewer segments, zooming in to constrain the parameter space.
We will now show how a signal can be identified in the initial step, but subsequently lost
in the follow-up, even if the glitch is quite small. For this exercise, we will investigate
a signal that contains a single glitch with δf = 5 × 10−7 Hz: this is a fairly typical
normal glitch size when compared to the observed normal population (see for example
Figure 7.2).

In Figure 7.13 we plot the semi-coherent mismatch for this glitch as a function of the
number of segments at a fixed observation time; this is inversely proportional to the
coherence time which we also give on the upper axis. This is predicted analytically from
the R-averaged Eqn. (7.22). We also performed a Monte-Carlo numerical simulation
in which the mismatch was calculated exactly for a fixed glitch size, but an R chosen
uniformly throughout the observation period. The resulting mismatches are then his-
togrammed and the density is shaded on to Figure 7.13: this demonstrates the spread of
Chapter 7 Glitches in continuous gravitational waves 173

mismatches about the average value. Notably there is a tight band centred on the an-
alytic R-averaged prediction of Eqn. (7.22); there is a low-density of mismatches below
this band. These occur when R is close to 0 or 1.

Figure 7.13: An example of how the semi-coherent mismatch changes with

Nseg for a fixed glitch size. The red line is the prediction of the R-averaged
Eqn. (7.22). The density map is constructed by performing numerical Monte-
Carlo simulations over R and binning the mismatches into histograms.

This plot shows that for this glitch during the initial search at Nseg = 667, the number of
segments used in the E@H all-sky search, the mismatch would be negligible ∼ 10−3.5 . As
a result the signal would be classified as a candidate and subsequently followed-up. We
see then that, as the number of segments is decreased, the mismatch rapidly increases
and hence the candidate would be dismissed.

7.5.4 Including the recovery from glitches

In this work, we have used the glitch catalogue maintained by Espinoza et al. [59]
which provides δν, the instantaneous frequency increase at the glitch. In addition to
this, some glitches also undergo a short-term exponential relaxation of some fraction
of the total glitch magnitude over times of tens to hundreds of days. Lyne et al. [114]
P P
characterised the fractional recovery by Q = ∆fi /δν where ∆fi is the total glitch
magnitude summed over the number of relaxation components. For each component,
∆fi is recovered over an e-folding timescale τid . In most cases, only a single component
is identified.

This may have an important effect on our estimates since, if a large fraction of the
glitch is recovered in a timescale short compared to the observation time, we will over-
estimate the mismatch. On the other hand, the exponential relaxation itself could also
cause a mismatch which would tend to vanish when τid  Tspan , but be maximal when
174 Chapter 7 Glitches in continuous gravitational waves

τid ∼ Tspan . Ideally we would like to model the exponential relaxation in detail, but the
more practical approach that we will take is to assume that the relaxation time is zero,
such that the actual long-term glitch magnitude is (1 − Q)δν; the results in this work
can be considered to correspond to Q = 0. The effect of this is that, since µ ∝ δν 2 , all
mismatches can be rescaled by a factor (1 − Q)2 , given the size of Q.

To determine the appropriate size of Q that we may expect, we can use the analysis by
Lyne et al. [114]. In general, it is found that for older pulsars, with lower spin-down
rates, Q  1, but Q correlates well with log10 |ν̇|. Taking all measured Q values for the
Crab pulsar [114, 164, 162], which has the highest spin-down rate −0.38 nHz/s in the
analysis, we find an average and standard-deviation of 0.77±0.2 for Q and 9.45±6.7 days
for the relaxation timescale. We note that in three cases the same glitch was measured
with differing values of Q and relaxation time; this presumably reflects the fact that the
measurements are uncertain and depend on the data span used. In contrast, the Vela
pulsar, which has the second largest spin-down rate 0.016 nHz/s has an average Q of
0.08 for the glitches in Lyne et al. [114], with none exceeding 0.2.

It is unclear what Q we might expect for the target population of all-sky searches, but
we can make an estimate by taking Q = 0.8. Given this and assuming the relaxation
time is much shorter than the observation period, this would mean the mismatch labels
of the black contour lines in Figure 7.10, Figure 7.11, and Figure 7.12 would decrease by
a factor ∼ 0.04, ultimately meaning that longer coherence times would be more robust
to typical glitches. Nevertheless, even given this, many searches would still be at risk
to lost signals due to the presence of a glitch. These estimates are a ‘best-case scenario’
in that they assume τid = 0. A more detailed analysis is required to understand how
a recovery time which is comparable to the observation time changes these estimate.
However, it seems likely that it will tend to increase the level of mismatch as compared
to the simple rescaling argument given here and not decrease it further.

7.5.5 Application to past and future searches

In Table 7.4 we present the expected number of glitches and the metric-mismatch (nor-
mal and Vela like) for the searches listed in Table 7.1. This is done taking the worst-case
scenario of the largest absolute spin-down rate and hence largest glitches one might ex-
pect with no glitch recovery. Note that this is reporting the metric-mismatch which is
unbounded above: a large metric-mismatch (greater than one) indicates a large mis-
match.

This table gives a clear picture that for all searches in the fully-coherent follow-up, the
worst mismatches are greater than one. It is true that the probability of a glitch is
less than one, but not by a sufficient margin to consider them unaffected. However, the
mismatch in the semi-coherent stage can be sufficiently small to be immune to glitches,
provided the coherence times and observation times are small enough.
Chapter 7 Glitches in continuous gravitational waves 175

normal population Vela-like population

hN i hµ̃i hµ̂i hN i hµ̃i hµ̂i
S5 E@H all-sky 2.7 1.2 × 104 0.3 1.2 9.0 × 104 2.7
S5 E@H galactic center 6.2 2.6 × 105 7.2 2.6 5.0 × 104 1.5
S5 all-sky 0.98 260.0 9.8 × 10−6 0.42 2.0 × 104 9.2 × 10−4
VSR low-frequency all-sky 1.5 1.2 × 10 3.8 × 10−3
3
0.65 5.9 × 103 2.2 × 10−2
−2
S5 supernova remnant (Cas A) 0.16 3.0 − 6.8 × 10 12.0 −

Table 7.4: Predictions for the expected number of glitches and metric-mismatch
at the highest spin-down rates for the searches listed in Table 7.1; these are worst
cases since we have chosen the highest spin-down rates. We present results both
for the number of expected glitches hN i, the fully-coherent metric-mismatch hµ̃i,
and the semi-coherent metric-mismatch hµ̂i. Note that the supernova remnants
search had no semi-coherent stage and we give the estimate for the search for
Cassiopeia A only.

7.6 Conclusions

We have investigated the effects of glitches in the GW signal when searched for using
semi-coherent and fully-coherent matched filtering techniques.

In Section 7.3 we confirmed the observation by Espinoza et al. [59] that glitches can be
regarded as originating from two distinct populations named Vela-like and normal, with
the Vela-like undergoing larger glitches. We then separated the data according to their
predicted source population and found fitting formulae for the glitches magnitudes using
the spin-down rate ν̇ as a predictor variable. Separating the populations was necessary
to avoid overestimating the glitch magnitudes when extrapolating into the parameter
space of all-sky gravitational wave searches where few pulsars have been observed. We
then used fitting formulae based on data from Espinoza et al. [59] to investigate the rate
of glitches.

In Section 7.4 we calculated the metric-mismatch for a glitch with no exponential rela-
tion. This was done both for a fully-coherent and semi-coherent searches (Eqn. (7.12)
and Eqn. (7.22) respectively) assuming that a single glitch occurs during the search. For
each type of search, we provided some simple estimates of the mismatch for the largest
glitches seen in the glitch catalogue.

Finally, we transformed our fits for the glitch magnitudes in the EM channel into pre-
dictions for the continuous GW channel assuming that fs = 2ν. This prediction for
the glitch magnitude was then used to estimate the mismatch for typical search dura-
tions. This predicts that in the initial semi-coherent search, a single glitch will cause a
moderate level of mismatch if either the glitch is Vela-like, or normal with the neutron
star having a large absolute spin-down rate. If a candidate signal with a glitch does get
captured in the semi-coherent stage, we show that, if naively followed up by a single
fully-coherent search over the full observation time, it will most likely have a mismatch
greater than 10% unless it has a very low absolute spin-down rate. These calculations
assume that a single glitch occurred during the search. To complete the picture, we used
our fitting formulae for the expected number of glitches to show that for typical search
durations there is a reasonable chance that a glitch will occur.
176 Chapter 7 Glitches in continuous gravitational waves

The mismatch estimates potentially overestimate the effect by ignoring the exponential
recovery from glitches. If these have short recovery timescales compared with typical
observation times, this could reduce the mismatch by a factor of (1 − Q)2 , where Q is
the glitch recovery parameter. It’s unclear exactly what value Q should take for the
target population, but we feel this is unlikely, even in the most optimistic scenarios, to
nullify the risk.

The levels of mismatch measured here are of concern to both future and past all-sky
searches for continuous GWs from neutron stars. If the effect of glitches is ignored,
detectable signals could easily by missed due to the presence of a glitch. In a fully-
coherent search the presence of a glitch can easily be determined either by including it
as a search parameter, or by considering different sections of data. A glitch has a weaker
effect on a semi-coherent search and so it is possible the signal will be identified as a
candidate, but subsequently lost in the follow-up. We therefore recommend modifying
follow-up procedures by introducing a greater number of steps. By studying how the
mismatch increases with the coherence time candidates when a glitch can be identified
and followed up using a search template which includes a glitch.

Appendix 7.A Bayesian model comparison: test of mix-

ture models

It seems clear by eye that the histogrammed magnitudes of the frequency change in
a glitch, log10 |δν|, as shown in Fig. 7.1, exhibit at least two distinct modes, which
suggests that it is generated by more than one mechanism. To model this, we will use
a Gaussian mixture model (GMM) [64] with N components. This model assumes that
the measured data is taken from a population with N sub-populations, each having
a Gaussian distribution with separate mean, variance and weight (µi , σi2 , λi ) where
P
i ∈ [1, N ]; note that N 1 λi = 1. Furthermore, we can also allow each of the components
to be skewed with a dimensionless skew parameter αi which can be either positive
or negative determining the direction of the skew, or 0, for which there is no skew.
Following O’Hagan and Leonard [129] then the probability density function of the ith
skewed Gaussian component is
Z x
f (x; µi , σi , αi ) = 2N (x; µi , σi ) N (αi x; µi , σi )dx, (7.27)
− inf

where N denotes the Gaussian distribution.

Let yi be the measured values of log10 |δν| and ϑ be the collection of all model parameters
{µi , σi , αi , λi }. Then the probability density for a GMM with N components is

N
X
P (yi |model, ϑ) = λi f (yi ; µi , σi , αi ). (7.28)
i=1
Chapter 7 Glitches in continuous gravitational waves 177

To compare different choices of N , we will perform a Bayesian model comparison (see

§4.2 of Jaynes [86] for an introduction) between each of the mixture models and the
simplest hypothesis, a mixture model with N=1.

For each model parameter we must specify a prior. We list these in Eqn. (7.29) having
defined hyi, |y|, and std(y) as the average, range, and standard-deviation of the data.

P (µi ) = Unif(hyi − |y|, hyi + |y|),

P (σi ) = Half-Cauchy (0, std(y)) ,
(7.29)
P (λi ) = Unif(0, 1),
P (αi ) = N (0, 10 × std(y)).

For the mean µi we use a uniform prior over a range of values containing all data points.
For the standard-deviation σi , we will use a Half-Cauchy distribution with zero-mean
as suggested by Gelman et al. [65]. A large standard-deviation, as compared to the
standard deviation of the data itself, provides a weakly informative prior. Instead, we
use a standard deviation of std(y)
2 to favour GMM components with small standard-
deviations as compared to the data. That is, our prior disfavours models in which any
of the components are wide and flat. The prior for λi is uniform on [0, 1] and for αi is
normally distributed with zero mean and a wide, weakly-informative standard deviation.
The choice of a zero mean favours non-skewed components. Note that, the non-skewed
models do not include αi as a model parameter and the GMM with N = 1 does not
include λi .

We use this choice of prior for the model parameters of each component in the GMM
with N components. In this way, models with larger values of N have a larger ‘prior
volume’ and hence there is a natural Occam-factor favouring the simpler models with
fewer components; this prevents over-fitting.

We will present results for the Bayes factor between a GMM with N components and
the simplest model, a GMM with N = 1 components. This is computed by
R
P (model|{yi }) P ({yi }|N GMM)P (θ)dθ
=R θ . (7.30)
P (N = 1|{yi }) ϑ P ({yi }|N=1 GMM)P (ϑ)dϑ

We use the emcee [62] MCMC algorithm to sample from the posterior and thermody-
namic integration to estimate the evidence integrals [67]. In Table 7.5 we provide the
log10 of the Bayes factor for several possible models. The Bayes factor between any two
of the models given in Table 7.5 can be calculated from their difference.

This table clearly shows that the data is decisive: a Gaussian mixture model with N ≥ 2
fits the data a great deal better than the simple N = 1 GMM. This is unsurprising given
the distinct multimodal nature of the data. However, the differences between the other
models is more subtle. No single model distinguishes itself by a decisive odds-ratio
compared to its neighbouring models. We have checked that these results are robust to
small changes in the prior specification.
178 Chapter 7 Glitches in continuous gravitational waves

 
P (model|d)
model log10 P (N=1 GMM|d)
2-components 39.12 ± 0.19
2-components (skewed) 41.60 ± 0.21
3-components 42.70 ± 0.23
4-components 44.27 ± 0.24
5-components 44.18 ± 0.22
6-components 43.21 ± 0.22
7-components 42.26 ± 0.22

Table 7.5: Table of the Bayes-Factor for all models considered in this study
compared to the simplest N=1 GMM. The error is an estimate of the numerical
error in the thermodynamic integration.

To help illustrate the differences between some of these models, in Figure 7.14 we plot
the probability density for the maximum posterior model parameters found for each
model. It is clear from these plots that the N = 2 model is not perfect, especially in

Figure 7.14: The distribution of glitch sizes in frequency along with the predic-
tions for the components of several GMM. The Bayes-factor for these models
can be found in Table 7.5.

predicting the number of glitches found in between the two primary subpopulations,
around log10 |δν/Hz| = −5.5; by comparison the N > 2 models and the N = 2 model
which allows for skewness can explain these points and this is reflected in the Bayes
factor.

From this analysis it is difficult to decide which model best fits the data. However, what
is clear is that simply modelling the data as a GMM with two components with a skew
provides a reasonable empirical model. For this reason, in our analysis of the glitch
population we will use this model and not any of the models with a greater number of
components.

It is important to realise that this comparison is purely empirical, in that the result was
not conditioned on a substantive physical model. It would be interesting to include such
modelling, this may provide some insight into the appropriateness of the mixture model
and the number of components.
Chapter 7 Glitches in continuous gravitational waves 179

Appendix 7.B Linear regression in log-space

In Section 7.3 we perform several linear regressions in log-space in order to calculate

power-law fits. This assumes that the observed values log(yi ) depend on the predictor
values log(xi ) as

log(yi ) = m log(xi ) + c + i (7.31)

where the i are independent and identically central normally distributed variables with
a standard-deviation σ. In this way, m and c are the linear fit free variables, while σ is
a measure of the variability in the observation about this linear fit.

We use a Bayesian linear regression in which we estimate the posterior distributions

of all three parameters using a Markov chain Monte Carlo algorithm; for the prior
distributions we use non-informative priors and test that these do not induce any bias.
In all cases we find the resulting posteriors to be Gaussian and so can take their mean
values to get best-fit parameters. The advantage of this method compared to a simple
least-squares linear regression is that we also estimate hσi, the variation about the linear
fit. The linear fit can therefore be written as

y(x) = hmix + hci ± hσi (7.32)

We can then rearrange this equation to give the corresponding power law fit in linear
space

y(x) = 10hci xhmi 10±hσi (7.33)

where the last term gives the variability about the mean. Hence neglecting this term
gives the mean.

This is an inherently problematic approach since many functions besides a power law
can appear linear in a log-log plot and the assumption of Gaussian error itself may not
be valid. Nevertheless, we will still apply this approach since we need only order-of-
magnitude estimates and can always check our predictions; we must be clear that the
power-law fit gives a good empirical fit, but is not intended to signify any substantive
underlying model.

Appendix 7.C Understanding the uncertainty in the pre-

dictions

As with all prediction, our estimates carry uncertainties both in the process itself, and
in the spread of the observed population. In calculating the expected mismatch, the
biggest source of errors come from the variability in the size of glitches in both δν and
δ ν̇, as shown by the shaded bands plotted in Figure 7.4. Also, calculating the expected
number of glitches we have uncertainty from the variability as shown by the shaded
bands in Figure 7.6. To understand how these uncertainties may change our belief in
180 Chapter 7 Glitches in continuous gravitational waves

Figure 7.15: A reproduction of the mismatch contours in Figure 7.10 with a

reduced number of contours, but showing the variation in predicted mismatch
due to the 1σ uncertainty in the predicted distributions as given by the shaded
bands in Figure 7.4. The two uncertainty bands overlap, giving rise to the
cross-hatched region.

the conclusions, in Figure 7.15 we have repeated the analysis that led to Figure 7.10,
with a reduced number of contour lines, and propagated the measure of uncertainty
in the glitch sizes. Specifically, we fill contour lines between the uncertainties in the
mismatch, given by the error in Eqn. (7.1) and Eqn. (7.2). In Figure 7.16 we repeat
the exercise showing the uncertainty on the number of glitches as given by the error in
Eqn. (7.3).

These figures illustrate that our uncertainty in exactly where the contour lines sit is
large. Nevertheless, even in the best case scenarios large portions of the parameter
space remain at risk.
Chapter 7 Glitches in continuous gravitational waves 181

Figure 7.16: A reproduction of the number of glitches in Figure 7.10 with a

reduced number of contours, but showing the variation in the expected number
of glitches due to the uncertainty in the glitch rate fit of Eqn. (7.3).
Chapter 8

Timing noise in continuous

gravitational waves: a numerical
study

In this chapter, we will investigate the effect of timing noise on targeted and narrow-band
coherent searches for continuous gravitational waves from pulsars. In particular, we will
use the Crab ephemeris data (introduced in Section 8.2) to simulate ‘noisy’ gravitational
wave signals and study how standard search tools perform when attempting to detect
the signals. This chapter contains material reprinted with permission from Ashton et al.
[24] as follows: Ashton, G.; Jones, D. I.; Prix, R. Physical Review D, Volume 91, Issue
6 (2015). Copyright (2015) by the American Physical Society.

8.1 Introduction

Timing noise is often represented by structure in the timing residual, which is the dif-
ference between the best fit Taylor expansion, typically up to second order in spin-down
f¨0 , and the observed phase. Timing noise refers specifically to deviations from Taylor
expansions that are intrinsic to the pulsar and not to systematic errors such as dispersion
in the interstellar medium. Hobbs et al. [81] conducted a wide ranging study on timing
noise across the pulsar population. They concluded, amongst other things, that timing
noise is ubiquitous and inversely correlated to the age of the pulsar. There already exists
measures used to quantify the strength of timing noise such as the ∆8 value introduced
by Arzoumanian et al. [23], the generalisation of the Allan variance [119], the covariance
function of the residuals [42], and the red spin noise parameter calculated when fitting
for timing noise as part of the pulsar timing model [104]. These do not convert directly
into the effect that timing noise may have on continuous GW searches for pulsars. To
quantify this, we need to measure the mismatch (as defined in Section 6.2) due to tim-
ing noise. This is closely related to the loss of signal to noise ratio due to the imperfect
matching between the template and signal (which we define explicitly in Section 8.3).

183
184 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

As discussed in Chapter 2.3, a variety of models exist to interpret timing noise, but
there is currently no consensus on a single mechanism. However, for the issue of timing
noise and continuous GW searches, we only need to consider the relation between the
components of the neutron star which produce the EM and GW signals. This was
investigated by Jones [88] who identified three possible scenarios. First, the two signals
are strongly coupled: the same timing noise will be observed in both. Second, the two
signals are loosely coupled: a similar, but different level of timing noise will be observed
in both. Third, timing noise exists only in the EM signal, there is no corresponding
variations in the GW signal. Of course these are really three cases from a full spectrum
of possibilities which could also include the pulsars GW signal being significantly more
noisy than its EM signal.

The significance of timing noise will vary between different types of continuous GW
searches; these can be divided into targeted, narrow-band, directed, and all-sky searches.
Targeted searches involve a single known pulsar where an estimate of the spin parame-
ters has been obtained from the EM signal. If we assume that the EM and GW signals
are strongly coupled, then we can use a single-template targeted search. Under this
assumption, when the level of timing noise in the EM signal is small, then a single
Taylor expansion is sufficient. If instead the level of timing noise is large, then the EM
data can be used to account for it; this is done by applying an adapted matched fil-
tering phase-model that closely follows the observed EM phase model [137]. If instead
we assume that the EM and GW signals are loosely coupled, then we should perform
a narrow-band search in a small area of parameter space. These narrow-band searches
aim to allow for small frequency offsets between the EM and GW signals, such as could
be caused by free precession, or a finite coupling time between the two components of
the neutron star [11]. Directed searches look for non-pulsing neutron stars predicted by
other means such as at the centre of the supernova remnant Cassiopeia A. An all-sky
search involves searching over the entire sky for unknown pulsars. For both directed and
all-sky searches the lack of EM data necessitates wide bands in the frequency and its
derivatives. For fully coherent matched filtering methods these searches can rapidly be-
come computationally prohibitive. To circumvent this, semi-coherent search techniques
are used that incoherently combine short fully-coherent sections of data [8]; these will
be less sensitive to timing noise. Nevertheless, semi-coherent searches ultimately need
to be followed up by targeted fully coherent searches, for which timing noise may be an
issue.

For the properties of the GW signal, the most general case is that it will exhibit some
timing noise, but it could be different to the timing noise observed in the EM signal. Until
a detection is made, we can only make assumptions about how the two are correlated.
To probe these assumptions, we will define two special cases corresponding to different
sorts of errors in a GW search:

• Special Case 1: Timing noise, exactly like that in the EM signal, exists in the
GW signal but is not included in the template. This will result in a loss of signal
to noise ratio for searches which assumed that timing noise was negligible. The
error potentially affects the narrow-band, directed, and all-sky searches since the
Chapter 8 Timing noise in continuous gravitational waves: a numerical study 185

level of timing noise is unknown. The single template targeted searches will not
be effected since they either check that the level of timing noise is negligible, or
correct for it using an adaptive phase model.

• Special Case 2: Timing noise is included in the template but does not exist in
the signal. This will result in a loss of signal to noise ratio for single-template
targeted searches that account for timing noise using an adapted phase model (for
example Abbott et al. [11]). Instead, these searches will now erroneously introduce
timing noise into the template while the signal will be a smooth Taylor expansion.

In this chapter, we will mimic narrow-band and single-template searches to directly

simulate special case 1. Specifically, we will inject a fake GW signal which contains a
realisation of timing noise, and recover it using templates based on a single global Tay-
lor series. This tests the scenarios in which the timing noise in the GW signal is either
exactly coupled to the EM signal, or they are at least similar. However, this also quanti-
fies special case 2 since the signal and template are interchangeable in matched filtering
methods. That is, timing noise in the signal, but not in the template is equivalent to
timing noise in the template, but not the signal.

While all known pulsars are potential GW sources, young pulsars are the most promising
due to their large spin-downs (see Abbott et al. [13] for a review). However, it was found
by Hobbs et al. [81] that the amount of timing noise is correlated with the spin-down
magnitude. This motivated us to study the effect of timing noise on GW searches for
neutron stars with large spin-downs.

The realisation of timing noise we will use to investigate timing noise in GW pulsar
sources is based on the young Crab pulsar. The Crab is a potentially detectable source
of gravitational waves due to its high spin-down rate and it has the highest spin-down
upper limit compared to the LIGO noise floor [11]. The EM signal from the Crab is
well documented (see Section 8.2) and contains exceptional levels of timing noise: it was
estimated by Jones [88] that such levels of timing noise in the GW signal may cause an
issue for current searches.

Several targeted searches have already been performed for GWs from the Crab pulsar. A
single-template search for GWs from the Crab pulsar was performed on data collected
during the LIGO S5 science run [11]. This search used the Crab ephemeris and an
adapted phase model to account for timing noise. In addition to this single-template
search, a narrow-band search for signals from the Crab was also performed by Abbott
et al. [11] on the S5 data. Another narrow-band search for the Crab was carried out
using data from the VIRGO VSR4 science run along with a search for the Vela pulsar
[5].

The structure of this chapter is as follows. In Section 8.2, we describe the observational
data available from the Crab ephemeris and discuss its relation to GW searches. In
Section 8.3, we describe the signal injection and recovery method. Results from this
method are presented in Section 8.4: we begin by considering the effect timing noise has
on narrow-band searches, then we consider the mismatch on stretches of data for which
narrow-band searches have been performed; we further investigate how the mismatch
186 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

depends upon epoch; and we finally examine how the mismatch depends on the duration
of observation. Finally, we summarise our results in Section 8.5.

8.2 Timing noise as described by the Crab ephemeris

The monthly Crab ephemeris [113] provides the phase evolution of the EM signal between
1982 and the present and can be found at http://www.jb.man.ac.uk/pulsar/crab.
html. It is unlike most timing data for pulsars where a timing model consists of the model
parameters (position, spin-down, etc.) given at a single reference time. For the Crab
ephemeris, each monthly update consists of the frequency and spin-down coefficients
along with a reference time coinciding with the time of arrival (TOA) of a pulse at
the solar system barycentre. The coefficients are calculated by least-squares fitting of a
Taylor expansion to the TOAs. The reference time for each month is chosen as the TOA
of the pulse closest to the mid-point; this is done to minimise the average phase error
of the local Taylor expansion. The period of a month is short enough such that these
coefficients and the Taylor expansion, Eqn. (1.2) track the rotational phase during the
month. The Crab ephemeris tracks the rotational frequency and frequency derivative of
the Crab pulsar (ν and ν̇). In this chapter, we will consider searches for gravitational
wave emission from non-axisymmetric neutron stars at twice this frequency f = 2ν;
hereafter all frequencies f and spin-downs f˙ refer to the pulsar’s GW emission.

The Crab ephemeris gives a distinct picture of the variations due to timing noise super-
imposed on the monotonic spin-down. To illustrate how this manifests itself, Figure 8.1
depicts the frequency evolution in two adjacent months. Notice that a discontinuity oc-
curs at the interface between months. Such discontinuities will occur in the spin-down,
frequency, and phase; timing noise can then be described by the magnitude of these
jumps. From the Crab ephemeris it can be shown that the distribution of jumps in
phase, frequency and spin-down appear to follow standard normal distributions. This
is consistent with timing noise models consisting of a large number of small unresolved
events accumulating over a month (e.g. the models considered by Cordes and Greenstein
[47]). The Crab ephemeris therefore provides an alternative way to discuss timing noise
(rather than in terms of structure in the timing residual), we investigate this further in
Section 9.5.1.

Timing noise is usually depicted by structure in the phase residuals calculated by re-
moving the best fit Taylor expansion to the phase from the real phase evolution. A
best fit Taylor expansion consists of a single set of coefficients f0 , f˙0 , and f¨0 valid over
the entire observation period. To make this distinct from the local Taylor expansions
describing the evolution in each month, this will be referred to as the global template.
In Figure 8.1 we see that if the discontinuity is non-zero, then it is impossible for any
global Taylor expansion template to exactly match the local signal in both months. The
phase residual, and hence timing noise, results from the inability to match a single global
template to the non-smooth signal.
Chapter 8 Timing noise in continuous gravitational waves: a numerical study 187

fi−1

∆f

fi

∆ti

ti ti+1

Figure 8.1: Illustration of the jumps between ‘local’ per-month signals in fre-
quency space, defining the frequency jump ∆f . This depiction amplifies the
order of magnitude of ∆f in order to highlight the timing noise: for the jumps
in the ephemeris, ∆f is several orders of magnitude smaller than the change in
frequency due to spin-down alone.

In this chapter, we aim to quantify the significance of timing noise in GW searches by

generating signals from the Crab ephemeris. This is an empirical description of timing
noise and so we make no assumptions on the underlying astrophysical model.

8.3 Method

We now describe the method to quantify the effect of timing noise on continuous GW
searches for signals from isolated pulsars. To be relevant to current pulsar continuous
GW search methods, we will base our method on the narrow-band searches of Abbott
et al. [11] and Aasi et al. [5]. The results can be interpreted as measuring the consequence
of special case 1 on narrow-band and single-template searches; that is we assume the
GW signal has a similar level of timing noise as the EM signal and search using global
Taylor expansion templates. For this study, a single-template search refers to a single
Taylor expansion template and not the adapted phase model proposed by Pitkin and
Woan [137].

We begin by generating a GW signal emulating timing noise using the Crab ephemeris.
This is done by stringing together month-long smooth Taylor expansion signals. Each
month uses the corresponding month from the Crab ephemeris for the Taylor expansion
coefficients. The ‘jumps’ at the interface between months constitutes the timing noise.
In more detail:
188 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

1. From the ephemeris select a period of data consisting of the reference times ti ,
frequency fi , and spin-down f˙i for each month i.

2. Generate the phase as a function of time from the data and then fit a global
Taylor expansion up to f¨ for the whole observation time. The fit results in a set of
interpolated coefficients [f0 , f˙0 , f¨0 ] at a global reference time halfway through the
data. These coefficients are used to centre the narrow-band search parameters.

3. We supplement the local monthly data {ti , fi , f˙i } with the fixed value of f¨0 cal-
culated in the previous step. The phase of the GW signal is always zero at each
monthly reference time ti of the ephemeris, which by construction coincides with
a pulse arrival time.

In this process we have assumed that a fixed value of f¨0 is sufficient. This can be
justified by considering the next term in the Taylor expansion and typical values of
...
f ∼ 10−30 Hz/s3 . Over typical search durations ∼ 1 year, this term contributes less
than a radian to the phase (which is less than typical contributions from timing noise
in the frequency and spin-down rate), and it can therefore be safely neglected.

We use the LALSuite [106] gravitational-wave analysis routines to generate data with a
fake GW signal; for these experiments we work without any simulated detector noise.
The standard tool to generate fake GW signals uses single Taylor expansion models.
Therefore, to include timing noise in the signal we do the following.

4. Into the data which we search for a signal, we inject each month-long Taylor
expansion generated from the Crab ephemeris lasting for only the duration of that
month. This method creates a fake GW signal, lasting several months, which
includes timing noise corresponding to the monthly ephemeris.

Once we have produced data, we then use LALSuite tools to recover the signal using the
F-statistic [85]. This is a matched filtering method in which the output of the detector
is compared to a signal template (see Prix [140] for more details).

Two types of searches are performed: a single template search at the interpolated co-
efficients [f0 , f˙0 , f¨0 ] and a narrow-band search in f and f˙ centred on the interpolated
coefficients. These searches were found to be sufficient to find the signal to within a
reasonable mismatch, so more sophisticated methods were not required.

The narrow-band consists of a grid of points in f and f˙. As found by Abbott et al. [11]
we find searching over f¨0 to be unnecessary for this experiment and so it is kept fixed
using the value found in step 2 above. The grid spacing is parameterised by m, the
one-dimensional maximal mismatch between two adjacent Taylor expansion templates.
From Aasi et al. [3] the corresponding grid spacing is given by
√ √
12m 720m
df = df˙ = 2 , (8.1)
πTobs πTobs

where Tobs labels the observation time.

Chapter 8 Timing noise in continuous gravitational waves: a numerical study 189

For the single-template and at each grid point in the narrow-band search, we measure the
squared SNR value ρ2 . In order to quantify the relative loss compared to the perfectly
phase-matched squared SNR ρ2s , we define the mismatch in the usual way (e.g. see Prix
[139]) as
ρ2 − ρ2
µ= s 2 . (8.2)
ρs
It is well known (e.g. see Prix [140]) that the SNR for a perfectly phase-matched signal is
independent of the signal phase evolution. Therefore, in the absence of timing noise the
measured value of ρ2 can reach the maximum value of ρ2s , and the mismatch therefore
vanishes in that template. In the presence of timing noise, even the best-matching
template will suffer some mismatch, and this effect will increase with the level of timing
noise.

In the single-template search, we measure a single mismatch value. The single-template

search can also be interpreted to quantify the error made in special case 2, when the
template is adapted to account for EM timing noise but none exists in the GW signal.
We can think of the narrow-band search as repeating the single-template search over a
grid of points; this allows us two degrees of freedom, corresponding to the frequency and
spin-down parameters, over which to minimise the mismatch. The grid point with the
minimum mismatch, which we denote by µmin , is the best candidate and will be used to
quantify the success of the search. Because the narrow-band can minimise the mismatch,
µmin must always be equal or smaller than the mismatch in the single-template search.

8.4 Results

8.4.1 The effect of timing noise on narrow-band searches

We begin by describing how timing noise degrades a narrow-band search. This is done
by comparing the result for a signal containing no timing noise with a signal generated
from the Crab ephemeris between MJD 45150 and 45668. This period holds no special
significance and is used simply to demonstrate the essential features of a signal containing
timing noise.

In Figure 8.2 we show the mismatch as a function of parameter space offset for (a) a
signal without timing noise, and (b) a signal containing timing noise. The signal without
timing noise is injected at the interpolated coefficients [f0 , f˙0 , f¨0 ]. Therefore, we find the
minimum mismatch with µmin = 0 at exactly the centre of the grid and the iso-mismatch
contours in the local neighbourhood around the origin are well described by ellipses (e.g.
see Prix [139]). For the signal with timing noise (b) we notice two distinctive effects: the
minimum achievable mismatch µmin is non-zero, and the iso-mismatch contours around
the best-match template are more irregular and less well described by ellipses. In the
following we will quantify the effect of timing noise by considering only the location and
value of the minimum mismatch grid point in the narrow-band search.
190 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

×10−14
1.0 1.0
0.9
0.8
0.5
0.7

f˙ − f˙0 [s−2 ]

Mismatch
0.6
0.0 0.5
0.4
0.3
−0.5
0.2
0.1
−1.0
0.0
−6 −4 −2 0 2 4 6
−1
f − f0 [s ] ×10−8

A: Signal without timing noise

×10−14
1.0 1.0
0.9
0.8
0.5
0.7
f˙ − f˙0 [s−2 ]

Mismatch
0.6
0.0 0.5
0.4
0.3
−0.5
0.2
0.1
−1.0
0.0
−6 −4 −2 0 2 4 6
−1
f − f0 [s ] ×10−8

B: Signal with timing noise

Figure 8.2: In Figure A we show the mismatch as a function of parameter space

for a signal without timing noise. The injected signal has parameters [f0 , f˙0 , f¨0 ].
As a result, the mismatch has a minimum at this point. This can be compared
with Figure B showing the mismatch from a signal including timing noise. The
signal is generated from the Crab ephemeris between MJD 45150 and 45668.

8.4.2 Results relevant to recent narrow-band searches

First we consider two particular periods of the Crab ephemeris corresponding to recent
narrow-band searches for the Crab: the LIGO S5 period [11] and the VIRGO VSR4
period [5]. The mismatch in the single-template and the minimum mismatch for the
narrow-band searches during both periods are listed in Table 8.1. For these periods
timing noise is found to produce a mismatch of ≈ 1%. As expected, the narrow-band
mismatch is smaller than the single-template search. The fractional difference between
the two searches is relatively small.
Chapter 8 Timing noise in continuous gravitational waves: a numerical study 191

Provided that the timing noise observed in the GW signal is at the same level (or less)
as that observed in the EM signal, this result signifies that the recent LIGO and VIRGO
narrow-band searches would not suffer significantly from the effects of timing noise.

In addition to producing a mismatch, timing noise may result in the best candidate
being found at some distance from the centre of the narrow-band search. However, we
find that the distance from the centre of the grid is small when compared to the grid
spacing used in actual narrow-band searches such as that used for the S5 and VSR4
searches. For the S5 period narrow-band search, we find that the minimum mismatch
was a fraction ∼ 0.01 of the grid spacing used in the Abbott et al. [11] search. At the
resolutions used in real narrow-band searches, the effects of timing noise on the location
of the minimum mismatch will not be evident.
Dates Single template Narrow band
MJD µ µmin
S5 53673 - 53977 0.00968 0.00933
VSR4 55681 - 55839 0.00659 0.00584

Table 8.1: Measurements of the mismatch during the S5 and VSR4 narrow-band
search periods.

Figure 8.3 shows the convergence of the measured best mismatch µmin for the narrow-
band search over the S5 period with the value of m. This demonstrates that the non-zero
values of µmin given in Table 8.1 are not the result of grid coarseness. For signals without
timing noise, the measured best mismatch µmin will have a minimum of ∼ m when the
putative signal is located halfway between grid points. In the limit of m → 0 we then
expect the measured mismatch to tend to zero. Instead, for a signal with timing noise
we observe a plateau after some initial reduction. This indicates that the grid is now
fully resolving the variations due to timing noise.

8.4.3 Minimum mismatch as a function of the observation epoch

We will now investigate how the best mismatch µmin varies as a function of the observa-
tion epoch. We only show the narrow-band search, as the results were found to be very
similar for the single-template search. The method consists of measuring the mismatch
µmin in a 6-month window, which is shifted in 1 month intervals over all the available
ephemeris data. The observation time of 6 months is chosen to be similar to typical GW
search durations. We are restricted to multiples of 1 month by the frequency of updates
to the Crab ephemeris.

Timing noise is not the only variability in the spin-down of pulsars - they can also
undergo sudden increases in rotation frequency known as glitches (for an introduction,
see Section 2.1). The Crab frequently glitches and these are catalogued by Espinoza
et al. [59] and available at http://www.jb.man.ac.uk/pulsar/glitches.html. The
mechanism which causes a glitch is not well understood and may involve unpredictable
variations in the GW signal. As a result, targeted GW searches either avoid periods
with known glitches [11], or allow for an arbitrary jump in gravitational wave phase
192 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

0.12

0.10

0.08

µmin
0.06

0.04

0.02

0.00
10−6 10−5 10−4 10−3 10−2 10−1
m

Figure 8.3: Measured best mismatch µmin as a function of grid spacing param-
eter m (see Eqn. (8.1)), for the Crab pulsar over the S5 period. This demon-
strates that µmin plateaus at a nonzero mismatch suggesting we are resolving a
mismatch due to timing noise instead of the effect of finite grid resolution.

at the time of the glitch [13]. For this chapter, we will ignore the complicating factor
introduced by glitches (which was considered in Chapter 7) and consider separately the
effect of timing noise. We do this by omitting windows which include glitches from the
search by using the aforementioned glitch catalogue.

We begin by searching in a small 40 × 40 grid in frequency and spin-down, with a fixed

grid space mismatch of m = 1 × 10−5 , and the grid spacing as defined in Eqn. (8.1). It
is possible that the minimum mismatch is found at the edge of the narrow-band grid;
such candidates are not true local minima in the mismatch. If this is the case, the search
is repeated with an increasingly larger grid size, but the same fixed grid spacing. This
process continues until we find a minimum mismatch which is not at the edge of the
grid.

Figure 8.4 shows the measured minimum mismatch in the narrow-band search for a
sliding 6-month window at the centre of the observation time. The mismatch due to
timing noise is the low level noise occurring in between glitches. Greater mismatches are
observed in the post-glitch periods; this is expected as the relaxation time after glitches
for the Crab is of the order 1 month [113]. We note the presence of an anomalous
period of large mismatch for all windows that include the ephemeris time MJD 55362.
The cause for this is unclear from the available data, but it may be caused either by
a measurement error or a small undetected glitch. In general, we find that the level of
mismatch due to timing noise is between µmin ∼ 10−3 − 10−2 for these 6-month searches.
Chapter 8 Timing noise in continuous gravitational waves: a numerical study 193

100

10−1

µmin
10−2

10−3

10−4
46000 48000 50000 52000 54000 56000
MJD

Figure 8.4: Minimum mismatch µmin found in 6-month sliding window searches
as a function of epoch at the centre of window. Vertical dashed lines indicate
glitch events as described by Espinoza et al. [59]. The solid vertical line indicates
the date MJD 55362, a period of anomalously large mismatch.

8.4.4 Averaged minimum mismatch as a function of the observation

duration

We can study the averaged behaviour of the mismatch µmin as a function of time by
varying the size of the sliding window in the previous section. This was done for both the
narrow-band and single-template searches; the mismatch from the narrow-band search
was found to be a fraction . 0.1 smaller on average than the single-template search. We
therefore will only present results from the narrow-band search. The shortest possible
window ∼ 6 months is restricted by the number of points needed to generate a fit to the
phase. Setting the upper limit at ∼ 17 months retains a statistically meaningful number
of points to average over. Having obtained the data from all sliding window sizes in this
range we want to analyse the average behaviour as a function of the observation time.
Before doing this we filter results in the following ways:

• We do not consider any windows that include or are bounded by glitch events

• Windows including the anomalous epoch MJD 55362 are omitted. We wish to
study the fluctuations due to timing noise, and this period is either an unidentified
glitch, or another highly unusual and unrepresentative form of timing noise

• While each entry of the ephemeris is on average valid over a whole month, some
months were truncated due to glitches. The sliding window, which works on a
fixed number of entries of the ephemeris will occasionally be shorter than average.
To ensure we are averaging over windows of a similar length we omit windows for
which the observation time differs from the average by 2 weeks.
194 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

In Figure 8.5 we plot the averaged minimum mismatch hµmin i as a function of observation
time. This indicates a growth of hµmin i with observation time resembling a power law.

0.30

0.25

0.20

0.15
hµmin i
0.10

0.05

0.00

−0.05
0 100 200 300 400 500 600
Observation time [Days]

Figure 8.5: Averaging the mismatch for sliding window searches and varying
the observation times. The points give the mean while the bars correspond to
one standard deviation.

To quantify the growth of the mismatch, we perform a least-squares fitting to a power

law. Fitting the expression:
 
Tobs n
hµmin ifit = κ , (8.3)
1 sec

we find the best fit parameters

κ = 1.5 ± 0.8 × 10−23 (8.4)

n = 2.88 ± 0.030. (8.5)

In Chapter 9 we will return to this and try to analyse if this fit can help to constrain
the source of timing noise.

For perfectly matched signals the squared SNR increases linearly [140] with observation
time. This suggests that longer observation times yield a greater likelihood of detection.
The power law fit with n > 1 implies that the average mismatch from Crab timing noise
grows faster than the squared SNR. Gains in SNR from longer observation time will
therefore eventually be outweighed by the increasing mismatch from timing noise. To
estimate when this may occur, we can rearrange Eqn. (8.2) to give

ρ2 = ρ2s (1 − µmin ) . (8.6)

Substituting the time dependencies for the perfectly matched SNR and the averaged
mismatch we have
n+1
ρ2 ∝ Tobs − κTobs . (8.7)
Chapter 8 Timing noise in continuous gravitational waves: a numerical study 195

Differentiating and solving for Tobs yields an expression for the observation time (in
seconds) beyond which the ρ2 value of a signal containing timing noise starts to decrease
 1/n
1
Tobs = . (8.8)
κ(1 + n)

For the fit values from Eqn. (8.5), this yields a critical observation time of Tobs ≈ 600 days
after which the mismatch exceeds hµmin i ≈ 0.25. In this case it is no longer true that
further increases in observation time will yield greater detectability.

Jones [88] estimated the maximum time the signal and template would remain coherent
given a random walk in frequency. A crude method used a phase residual of 1 rad
for the decoherence criteria. For the Crab, this estimates the decoherence time at 200
days. We can improve upon this result by setting a mismatch of 0.1 as the decoherence
criteria; using the fit to the averaged mismatch this gives us a decoherence time of
Tobs ≈ 400 days.

The growth of mismatch as a power law is suggestive of random walk timing noise models
for which, as shown in Section 2.3.1, the root-mean-square phase residual also grows as
a power law. The power to which the phase residual grows depends on the type of
random walk: [33, 47] classified the Crab as undergoing frequency-like noise. However,
to compare with Eqn. (8.5), such a scaling in the phase residual must first be converted
into a mismatch, which will depend on the search method, before a comparison can be
made. We will return to this in Chapter 9 where we directly calculate the mismatch due
to a random walk and update Figure 8.5 with this prediction.

8.5 Conclusions

We have used observational data on the Crab pulsar to characterise the possible ef-
fects of timing noise on coherent targeted single-template and narrow-band continuous
gravitational-wave searches for pulsars. This was done by generating fake signals based
on the Crab ephemeris data and searching for them using templates without timing
noise. Our analysis clarifies the impact for current searches; accordingly, our methods
mimic those used by Abbott et al. [11] and Aasi et al. [5].

Our principal result is summarised by Figure 8.5: when considering the average mis-
match as a function of observation time, we find that the averaged mismatch grows as
a power law. In addition to this, we found two interesting aspects when considering the
data without averaging over the epoch:

Firstly, for the S5 and VSR4 narrow-band searches, if the timing noise in the GW signal
from the Crab is at a similar level (or lower) to that in the EM signal, then we find
it will only have a small (≈ 1%) effect on the measured squared SNR of the putative
signal. We found the mismatch in single-template searches to be only fractionally larger
than the narrow-band searches. This also suggests phase-adapted searches would not be
significantly affected if the signal does not contain timing noise.
196 Chapter 8 Timing noise in continuous gravitational waves: a numerical study

Secondly, searching over all available Crab data with a 6-month window, we looked at
the mismatch as a function of observation epoch. Post glitch periods tend to admit
significant levels of mismatch; this is expected due to the exponential recovery from the
glitch. (We also discovered a period around MJD 55362 which has a large mismatch
and is not connected to a known glitch). The narrow-band and single-template searches
performed similarly in this and subsequent tests. Typically, the mismatch due to timing
noise for 6-month searches was found to be between 10−3 and 10−2 .

The scope of this work can be extended to directed and all-sky searches, which target
young rapidly spinning down stars which may emit the strongest GWs. These stars are
also known to exhibit the highest levels of timing noise and glitch frequently. Crucially
the lack of EM data means we cannot be certain a glitch does not occur during the
observation and we cannot account for timing noise in the signal. In future work we
would like to quantify both these effects and estimate safe upper limits for the search
durations.
Chapter 9

Timing noise in continuous

gravitational waves: random walk
models

It’s unclear at this time what causes timing noise, but from Chapter 8 we know that it
may pose a problem for continuous GW searches if the phase evolution is affected by
timing noise and this is not included in the search templates. In addition to searches for
isolated neutron stars, many searches [see for example 4, 102, 124] have been performed
for GWs from low-mass X-ray binary systems (LMXBs). These searches experience
similar difficulties due to a stronger form of timing noise known as ‘spin-wandering’.
This issue was investigated by Watts et al. [165] who quantified the effect for a spin-up
2
of ν̇s by defining a decoherence time Tdecoh , such that Tdecoh ν̇s = 1, after which the
Taylor expansion can no longer track the phase. They then estimated a worst-case
decoherence time by using the maximal spin-up rate due to the accretion torque and
found that for some sources such as the LMXB Scorpius X-1, the decoherence time can
be short as ∼ 1 week.

In this chapter, we will present some preliminary calculations and results related to mod-
elling timing noise in a continuous GW using a random walk model. Unlike the results
of Chapter 8, which used an empirical description of timing noise given by the Crab
ephemeris, the results derived here can be applied to any search in which it is thought
the signal may undergo a random walk. In this sense, it is equivalent to Chapter 7 in
that the ultimate aim is to estimate the risk faced by various searches by inferring from
the observed pulsar population. As given here, the task is incomplete and so we will
present our calculations and some preliminary results. We will study the effect in the
context of searches which assume a smooth Taylor expansion template; advanced meth-
ods such as the ‘loosely-coherent’ search introduced by Dergachev [53] could provide a
way to search for noisy signals without a loss of signal to noise ratio.

Recent observations by Hobbs et al. [81] suggest that a random walk model does not
capture the physics of timing noise in isolated radio pulsars and hence is not a useful way
to infer neutron star physics. Nevertheless, the random walk model remains a practical

197
198 Chapter 9 Timing noise in continuous gravitational waves: random walk models

empirical model; in this section then, we use it as such without requiring it to have any
deeper substantive meaning for what causes timing noise. In the same way, the random
walk model can also be applied to LMXBs where the amount of spin-wandering could
be inferred from fluctuations in the luminosity. From this, we would like in the future
to update the estimates by Watts et al. [165] to estimate the mismatch for searches for
continuous gravitational waves from LMXBs.

This chapter is organised in the following way. In Section 9.1 we will define the random
walk as a piecewise Taylor expansion in which the difference between the signal and
template is initially zero. We will then calculate the mismatch for this simple description
in Section 9.2. However, setting the difference between the signal and template to zero
initially does not minimise the mismatch, so in Section 9.3 we will discuss a pragmatic
method to estimate the minimised mismatch; for each predicted mismatch we verify our
calculations using Monte-Carlo like numerical simulations. In Section 9.4 we relate this
description of a random walk to the compound Poisson process random walk usually
discussed in the literature. Finally, in Section 9.5 we begin by discussing the data from
the Crab ephemeris (first introduced in Section 8.2) in the context of a random walk. We
then go on to apply the results found earlier in the chapter to predict how the mismatch
in the Crab depends on the observation time assuming that it undergoes a random walk
in the frequency; this is then compared to the results found in Chapter 8.

9.1 Defining a random walk

To calculate the fully-coherent mismatch, we will model the random walk as a zero-mean
Gaussian walk in the phase, frequency, and spin-down which jumps regularly at Nsd fixed
time intervals ∆T , such that the total observation time is Tobs = Nsd ∆T . This allows
us to write the signal as a piecewise Taylor expansion with Nsd subdomains. Choosing
fixed time intervals appears to introduce an additional timescale not usually present in
random walk models. However, as shown later in Section 9.4 this is consistent with a
large number of unresolved events which are measured over the fixed timescale.

9.1.1 Initial definitions

In this model, we choose the spin-down rate to be the highest order term which undergoes
a random walk. Recalling that the parameter space offset ∆f˙i (defined in Eqn. (6.9)) is
the difference between the signal and template in the ith subdomain, we may define the
jump in this difference between the i and i − 1 subdomains as df˙i , such that

∆f˙i − ∆f˙i−1 = df˙i ∼ N (0, σf2˙ ), (9.1)

where N denotes the normal distribution and we have defined σf2˙ as the standard-
deviation of the step sizes in the spin-down rate. Rearranging this gives an expression
Chapter 9 Timing noise in continuous gravitational waves: random walk models 199

for the offset in the ith subdomain. By induction we can also write down the i − 1 term

∆f˙i = df˙i + ∆f˙i−1 , (9.2)

∆f˙i−1 = df˙i−1 + ∆f˙i−2 . (9.3)

We set the initial difference between the signal and template as zero such that ∆f˙0 = 0,
that is we start each random walk from the origin (we will return to this point in
Section 9.3). Then as each step proceeds from the previous step, we have that

i
X
∆f˙i = df˙j . (9.4)
j=1

To illustrate this, in Figure 9.1 we plot an example of a random walk in the spin-down
as given by Eqn. (9.4).

∆f˙
R
∆f (t) = ∆f˙(t)dt
Spin-down rate

∆T
0

ti−7
ref ti−6
ref ti−5
ref ti−4
ref ti−3
ref ti−2
ref ti−1
ref tiref = ti
time

Figure 9.1: An example of a random walk in the spin-down rate, Eqn. (9.4).
The filled green blocks indicate the summation defined in Eqn. (9.5) required
to calculate the induced change in frequency at ti due to the random walk in
spin-down rate.

If we want to model a random walk in the phase, frequency, and spin-down rate concur-
rently, then we must consider the effect that a random walk in spin-down will have on
the frequency and phase. For example, if we increase the spin-down rate for a period of
time, then we would expect the frequency to decrease at a greater rate during this pe-
riod. In our discrete model, it is not possible to dynamically change the frequency during
a single subdomain. However, we can approximate this by updating the frequency in
the next subdomain with the induced frequency offset due to the spin-down in all the
previous subdomains. This must be done for the induced effect from the random walk in
spin-down rate on the phase and frequency, and for the induced effect from the random
200 Chapter 9 Timing noise in continuous gravitational waves: random walk models

walk in frequency on the phase. There is no induced effect for the spin-down rate: the
effect only propagates to lower order terms.

Because the random walk is discrete and modelled as constant in any given subdomain,
we can calculate the offset in the lower order terms from a Taylor expansion. The total
offset at the ith reference time is then given by the summation of the offset caused by
all higher order terms up to that reference time. We can choose the reference times
arbitrarily, but setting each to start at the beginning of the subdomain simplifies the
calculation. For the frequency offset induced by the spin-down, we have
i−1
X
∆fi = ∆f˙j ∆T. (9.5)
j=1

This can be thought of as the integration of the spin-down up to the ith reference time
and is illustrated by the green blocks in Figure 9.1.

We want to consider random walks in all three parameters so we now add in a random
walk in frequency. Each step is independent of the induced effect from the spin-down
and is given by dfi ∼ N (0, σf2 ). The two effects will sum linearly such that the frequency
offset is
Xi Xi−1
∆fi = dfj + ∆f˙j ∆T. (9.6)
j=1 j=1

By a similar process we can calculate the induced effect of the frequency and spin-down
on the phase. Including the random walk in the phase for which dφi ∼ N (0, σφ2 ), the
phase offset is given by
 
i
X Xi−1 i−1
X
1
∆φi = dφj + 2π  ∆fj ∆T + ∆f˙j ∆T 2  . (9.7)
2
j=1 j=1 j=1

9.1.2 Writing the parameter offsets in terms of normal distributions

Eqn (9.6) and Eqn. (9.7) give the difference between the signal and template as functions
of the random walks in higher order parameters. In order to calculate statistical values,
we now write these in terms of the normal distributions from which the random walks
are constructed. Substituting Eqn. (9.4) into Eqn. (9.6) and using the identity defined
in Eqn. (9.76) of Appendix 9.A, we have

i
X j
i−1 X
X
∆fi = dfj + df˙k ∆T, (9.8)
j=1 j=1 k=1
i
X i−1
X
= dfj + (i − j)df˙j ∆T. (9.9)
j=1 j=1
Chapter 9 Timing noise in continuous gravitational waves: random walk models 201

Similarly, substituting this equation into Eqn. (9.7) and using both Eqn. (9.76) and
Eqn. (9.80) from Appendix 9.A, we have
 
X i i−1
X i−1
X
1
∆φi = dφj + 2π  ∆fj ∆T + ∆f˙j ∆T 2 
2
j=1 j=1 j=1
 ! 
X i i−1 X
X j j−1
X i−1 X
X j
1
= dφj + 2π  dfk + (j − k)df˙k ∆T ∆T + ∆f˙k ∆T 2 
2
j=1 j=1 k=1 k=1 j=1 k=1
 
X i i−1
X X j−1
i−1 X i−1
X
1
= dφj + 2π  (i − j)dfj ∆T + (j − k)df˙k ∆T 2 + (i − j)∆f˙j ∆T 2 
2
j=1 j=1 j=1 k=1 j=1
 
X i i−1
X i−1
X
1
= dφj + 2π  (i − j)dfj ∆T + ((i − j) (i − j − 1)) + (i − j)) df˙j ∆T 2 
2
j=1 j=1 j=1
 
X i i−1
X i−1
X
1
= dφj + 2π  (i − j)dfj ∆T + (i − j)2 df˙j ∆T 2  . (9.10)
2
j=1 j=1 j=1

This result can be interpreted as the accumulated phase over a time i∆T due to a
random walk in the phase, frequency, and spin-down rate.

9.2 Random walk models: a simple treatment

We will now calculate the mismatch for a fully-coherent search given the random walk
in phase, frequency, and spin-down rate defined in the previous section.

Let us begin by expanding the metric-mismatch summation from Eqn. (6.49). Recall-
ing that Greek indices label the parameter components and Roman indices label the
subdomain, then writing the summations explicitly, we have

µ̃ = gαβij ∆λαi ∆λβj (9.11)

Nsd X
X Nsd
= gαβij ∆λαi ∆λβj (9.12)
i=1 j=1
Nsd
X Nsd X
X Nsd
αi βi
= gαβii ∆λ ∆λ + gαβij ∆λαi ∆λβj . (9.13)
i=1 i=1 j=1
j6=i

The summation has been intentionally split into terms for which the two subdomains
are the same and those for which they are different. We calculated the metric when the
reference time is at the beginning of each subdomain in Eqn. (6.57); by considering this
metric for the two cases, we can define two distinct components as
(
E
gαβ if i = j
gαβij = NE . (9.14)
gαβ if i 6= j
202 Chapter 9 Timing noise in continuous gravitational waves: random walk models

Finally, we can calculate the fully-coherent metric-mismatch from

Nsd
X Nsd X
X i−1
E
µ̃ = gαβ ∆λαi ∆λβi + 2 NE
gαβ ∆λαi ∆λβj . (9.15)
i=1 i=1 j=1

9.2.1 Taking the expectation

In Eqn. (9.4), Eqn. (9.9), and Eqn. (9.10) we have written the parameter space offsets
(which are to be used in calculating the mismatch) purely in terms of the random walk
distributions dφi , dfi , and df˙i . We can calculate the mismatch exactly given a set of
random walk jumps by inserting these into Eqn. (9.15). However, since we are dealing
with statistical quantities, we can instead infer the behaviour of the mismatch under the
random walk by taking an expectation.

Inserting Eqn. (9.4), Eqn. (9.9), Eqn. (9.10) in Eqn. (9.15) yields a number of terms
with all the permutations of two terms from [dφ, df, df˙]. Taking the expectation, all the
cross-correlated terms, such as dφi df˙, will have an expectation of zero since the steps
of the random walk are independent. The only non-vanishing terms are given by

E[dφi dφj ] = δij σφ2 , E[dfi dfj ] = δij σf2 , E[df˙i df˙j ] = δij σf2˙ , (9.16)

After some simplification we find that the mismatch is given by

   
Aφ 1 π 2 Af 3 2 1
E[µ̃] = Nsd − + 4Nsd + 5Nsd +
6 Nsd 30 Nsd
2   (9.17)
π Af˙ 5 3 2 94
+ 66Nsd − 21Nsd + 105Nsd + 217Nsd + 63 − ,
3780 Nsd

where
Aφ = σφ2 Af = σf2 ∆T 2 Af˙ = σf2˙ ∆T 4 , (9.18)

define three dimensionless ‘activity parameters’.

Recalling that Nsd = Tobs /∆T , Eqn. (9.17) makes predictions for the leading order
scaling of the three random walks with the observation period
3
Tobs 5
Tobs
Tobs
E[µ̃]PN ∼ σφ2 , E[µ̃]FN ∼ σf2 , E[µ̃]SN ∼ σf2˙ . (9.19)
∆T ∆T ∆T
We note here the exact relation to the scaling of the variance of the root-mean phase
residual as calculated by Cordes [43] and given in Eqn. (2.7). We will return to this
later in Sec. 9.4.

9.2.2 Verifying the results

We can observe the leading order scaling of Eqn. (9.19) directly and verify the predic-
tions made by Eqn. (9.17) by comparing with exact numerical results. That is, using the
Chapter 9 Timing noise in continuous gravitational waves: random walk models 203

LALSuite [106] signal injection and recovery tools developed in Section 8.3 of Chapter 8
we simulate signals undergoing a random walk and calculate the corresponding mismatch
(no minimisation step is done here; this is discussed in the next section). In particular,
we perform three Monte Carlo studies for a random walk in the phase, frequency, and
spin-down rate and in each case compare the simulated results with the analytic predic-
tion. The results are shown in Figure 9.2 and demonstrate good agreement between the
simulation means and the prediction of Eqn. (9.17).

0.20 0.30
Mismatch expectation with Mismatch expectation with
σφ = 0.1, σf = 0, σf˙ = 0 σφ = 0, σf = 1.0×10−9 ,σf˙ = 0
0.25
Signal injection results Signal injection results
0.15
Mean and Std Mean and Std
0.20

0.10
Mismatch

Mismatch
0.15

0.10
0.05

0.05
0.00
0.00

−0.05 −0.05
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
Observation time [months] Observation time [months]
(a) Random walk in phase (b) Random walk in frequency

1.0
Mismatch expectation with
σφ = 0, σf = 0, σf˙ = 1.0×10−16
0.8 Signal injection results
Mean and Std

0.6
Mismatch

0.4

0.2

0.0

−0.2
0 2 4 6 8 10 12 14 16 18
Observation time [months]
(c) Random walk in spin-down

Figure 9.2: A comparison of Monte Carlo numerical simulated mismatch with

the prediction of Eqn. (9.17) for a random walk in the phase, frequency, and
spin-down rate.

9.3 Random walk models: minimising the mismatch

In Section 9.1 we have defined a random walk model for which we subsequently calculated
the fully-coherent mismatch in Section 9.2. However, this is a special case in which
the random walk for each parameter offset (the difference between the signal and the
template) begins at the origin and then grows with time. It is the signal which undergoes
a random walk, so in this case we have set the template to exactly match the signal
at t = 0. However, one could imagine choosing a different template which would reduce
204 Chapter 9 Timing noise in continuous gravitational waves: random walk models

the overall mismatch; as such Eqn. (9.17) may overestimate the mismatch. The proper
procedure is to minimise the mismatch with respect to the template parameters λαt ,
which are implicitly in the calculation of Section 9.2, through the parameter space offset

∆λαi = λαi αi
s − λt , (9.20)

as first defined in Section 6.4.

Ideally, we would like to repeat the calculation leading to Eqn. (9.17), minimising the
mismatch with respect to the template parameters. However, this calculation is difficult
and so we have not yet attempted it. A practical alternative method which we will use
here is to begin with the random walk starting at the origin, as defined in Section 9.1,
and then fit a polynomial of degree k, minimising the root-mean-square residual between
the parameter space offset and the polynomial. We then subtract the best-fit polynomial
from the parameter space offset. This leaves us with a residual parameter space offset
for which we then compute the mismatch. We will verify that this captures the essential
features of minimising the mismatch by comparing with numerical simulations in which
the exact mismatch is minimised numerically.

In Appendix 9.B, we introduce the basic tools of least squares fitting and removing a
polynomial of degree k to a generic random walk. In the following sections, we will
calculate the minimised mismatch for random walks in the phase or frequency; we have
not yet calculated the corresponding result for mixtures or random walks in the spin-
down rate. In this process, we have a choice in the degree of polynomial to fit and remove.
Since most searches minimise the mismatch with respect to the template frequency ft
and spin-down rate f˙t , this is equivalent to fitting and removing a k = 2 polynomial to
the phase residual. In the following sections we will first calculate the analytic results
and then verify that these agree with exact numerical results in Section 9.3.3.

9.3.1 Random walk in the phase

We begin with the simplest case of a random walk in phase, for which we have
i
X
∆φi = N (0, σφ2 ). (9.21)
j=1

Then, as shown in Eqn. (9.94) of Appendix 9.B, we have that

E[∆φi ∆φj ] = σφ2 min(i, j). (9.22)

We define the residual difference between the signal and template after fitting and re-
(2)
moving the best-fit 2nd order polynomial, ŷi , as

(2)
∆(2) φi = ∆φi − ŷi . (9.23)
Chapter 9 Timing noise in continuous gravitational waves: random walk models 205

Note that the superscript ‘(2)’ indicates the degree of polynomial and by ∆(2) φi we mean
the residual difference between the signal and template after fitting and removing the
polynomial.

We set the difference between the signal and template in all other parameters to zero
such that the mismatch for a random walk in the residual phase is therefore

µ̃ = g00ij ∆(2) φi ∆(2) φj (9.24)

Nsd
X Nsd X
X i−1
E (2)
= g00 ∆ φi ∆(2) φi + 2 N E (2)
g00 ∆ φi ∆(2) φj . (9.25)
i=1 i=1 j=1

To calculate the expectation of the mismatch, we need to evaluate the expectation of

Nsd
! Nsd
!
X (2)
X (2)
∆(2) φi ∆(2) φj = ∆φi − Cik ∆φk ∆φj − Cjl ∆φl (9.26)
k=1 l=1
Nsd N
!
X (2)
Xsd
(2)
=∆φi ∆φj − Cik ∆φj ∆φk + Cjl ∆φi ∆φl
k=1 l=1
Nsd X
X Nsd
(2) (2)
+ Cik Cjl ∆φk ∆φl , (9.27)
k=1 l=1

(2)
where Cij is defined in Eqn. (9.111) and Eqn. (9.110) of Appendix 9.B and we have re-
placed the ∆x notation of the appendix with the time ∆T . Then taking the expectation
!
h i Nsd
X Nsd
X
(2) (2)
E ∆(2) φi ∆(2) φj = σφ2 min(i, j) − Cik min(j, k) + Cjl min(i, l)
k=1 l=1
Nsd X
Nsd
!
X (2) (2)
+ Cik Cjl min(k, l) . (9.28)
k=1 l=1

Using symbolic mathematics packages we calculate an analytic expression which is a

function of ∆T, i, j and Nsd . Inserting this into Eqn. (9.25) and simplifying we find that

Nsd
X h i Nsd X
X i−1 h i
E
E[µ̃] = g00 E ∆(2) φi ∆(2) φi + 2 NE
g00 E ∆(2) φi ∆(2) φj (9.29)
i=1 i=1 j=1
 
1 2 27
= σφ 3Nsd − . (9.30)
70 Nsd

This expression can be compared to Eqn. (9.17) ignoring the effect of the random walk
in the frequency and spin-down rate. Notably, we retain the same leading order scaling
of Nsd , but the overall coefficient is decreased; the same effect was found by Cordes [43]
for the root-mean-square phase residual after removing a polynomial.

Rearranging the expression in the bracket demonstrates the mismatch is negative or zero
for 1 ≥ Nsd ≥ 3. This is a reflection of the minimum number of points needed in order
to perform the quadratic fit. We will discuss this in more detail in Sec. 9.B.6 for the
simpler case of fitting and removing a polynomial from a generic random walk.
206 Chapter 9 Timing noise in continuous gravitational waves: random walk models

9.3.2 Random walk in the frequency

For a random walk in the frequency we have an added complexity caused by the frequency
offsets in the phase. For the frequency offset, we have
i
X
∆fi = N (0, σf2 ). (9.31)
j=1

Recalling that we set the reference time at the beginning of each subdomain, then as in
Section 9.1, the induced phase offset is
i−1
X
∆φi = 2π ∆fj ∆T (9.32)
j=1
j
i−1 X
X
= 2π∆T N (0, σf2 ) (9.33)
j=1 k=1
i
X
= 2π∆T (i − j)N (0, σf2 ). (9.34)
j=1

Note that we do not include a random walk in the phase here.

Then we calculate the expected values of combinations of the parameter space offsets
using Eqn. (9.94) from Appendix 9.B and the two summation identities Eqn. (9.76) and
Eqn. (9.80) from Appendix 9.A. This gives

E[∆fi ∆fj ] = σf2 min(i, j), (9.35)

min(i,j)
X
E[∆φi ∆fj ] = 2π∆T σf2 (i − k), (9.36)
k=1
min(i,j)
X
E[∆φi ∆φj ] = (2π∆T )2 σf2 (i − k)(j − k). (9.37)
k=1

In Eqn. (9.23), we defined the residual difference between the signal and template phase
after fitting and removing a second order polynomial. The second order polynomial
was chosen to model the effect of minimising over the template frequency and frequency
derivative. Let us now define
(1)
∆(1) fi = ∆fi − ŷi , (9.38)

as the residual difference between the signal and template frequency after fitting and
removing a first order polynomial. In this instance, the first order polynomial models
the effect of minimising over the template frequency and frequency derivative.

To calculate the mismatch, we expand Eqn. (6.49) summing over the residual frequency
offset ∆(1) fi (defined in Eqn. (9.38)) and the residual phase offset ∆(2) φi (given by
Chapter 9 Timing noise in continuous gravitational waves: random walk models 207

substituting Eqn. (9.34) into Eqn. (9.23)). This gives

Nsd 
X h i h i h i
E
E[µ̃] = g00 E ∆(2) φi ∆(2) φi + 2g01
E
E ∆(2) φi ∆(1) fi + g11
E
E ∆(2) fi ∆(1) fi
i=1
Nsd X
X i−1 h i h i
NE
+2 ( g00 E ∆(2) φi ∆(2) φj + g01
NE
E ∆(2) φj ∆(2) fi +
i=1 j=1
h i h i
NE
g10 E ∆(2) φi ∆(2) fj + g11
NE
E ∆(1) fi ∆(1) fj .
(9.39)

We calculate each of these expressions in a similar manner to Eqn. (9.28) replacing the
relevant expectations with those given in Eqn. (9.35) to Eqn. (9.37). This yields an
expected mismatch given by
 
π2 2 2 3 82
E[µ̃] = σ ∆T Nsd + 13Nsd + . (9.40)
630 f Nsd

This can be compared with the frequency noise term alone in Eqn. (9.17). We note that
the leading order power remains unchanged, but there is a reduction in the coefficient
and a difference in the second highest power. The reduction in the coefficient is expected
since we have minimised the mismatch; we do not have an intuitive explanation for the
change in the second highest power.

9.3.3 Verifying the results

We now verify Eqn. (9.40) and Eqn. (9.30) by comparing with Monte Carlo simulations as
we first did in Section 9.2.2. The numerical signals undergo a random walk as described
in Section 9.2. However, when searching for the signals, we now search over a grid of
points in ft and f˙t . Then we select the grid point with the minimum mismatch; this
minimises the mismatch over the frequency and spin-down. The results are plotted in
Figure 9.3 and demonstrate reasonable agreement between the analytic prediction and
the mean of the simulated mismatches.

9.4 Understanding the random walk model

In the previous two sections, we have used the random walk model defined in Section 9.1
which appears at first to be non-physical: the random walk ‘jumps’ every ∆T , but
what physics determines this length? Furthermore, we assumed that the jumps between
adjacent steps (see for example Eqn. (9.1)) were normally distributed with zero mean
and variance σφ2 , σf2 , and σf2˙ . But how physical is this assumption? In this section, we
will answer both these questions and show that such a random walk is equivalent to
the more natural ‘compound Poisson random walk process’ introduced in Section 2.3.1
which is used in the literature to model timing noise as a random walk. The methods
208 Chapter 9 Timing noise in continuous gravitational waves: random walk models

0.035 0.0020
Mismatch expectation with Mismatch expectation with
0.030 σφ = 0.1, σf = 0, σf˙ = 0 σφ = 0, σf = 1.0×10−9 ,σf˙ = 0
Signal injection results Signal injection results
0.025 Mean and Std 0.0015 Mean and Std

0.020
Mismatch

Mismatch
0.015
0.0010
0.010

0.005
0.0005
0.000

−0.005
0.0000
−0.010
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
Observation time [months] Observation time [months]
(a) Random walk in phase (b) Random walk in frequency

Figure 9.3: A comparison of the Monte Carlo numerical simulated mismatch

with the predictions of Eqn. (9.30) and Eqn. (9.40); this differs from Figure 9.2
in that the numerical mismatch is minimised by selecting the smallest mismatch
from a grid of points in ft and f˙t .

discussed here parallel the work of Groth [73] and Cordes [43], but are described in detail
as there are some subtleties to our analysis.

Consider a parameter ∆X(t) ∈ [∆φ, ∆ν, ∆ν̇] being the difference between the real noisy
signal and the search template. ∆X(t) then encodes the deviations due to timing noise
without concerning the secular spin-down. We will model timing noise by allowing
∆X(t) to undergo a compound Poisson process: a random walk where events are Poisson
distributed in time occurring with a rate λ, but the size the events are drawn from a
normal distribution with zero mean and a variance hδX 2 i, or more formally

N (t)
X
∆X(t) = δXi , (9.41)
i=1

where {N (t) : t ≥ 0} is a Poisson process with rate λ and {δXi : i ≥ 1} are independent
and identically distributed with δXi ∼ N (0, hδX 2 i). Note that hδX 2 i where X ∈ [φ, ν, ν̇]
is the variance of the random walk jumps.

If the event rate λ of the compound Poisson process is sufficiently large such that in a
time ∆T the number of events is large, it can be shown that, due to the central limit
theorem (see for example Weiss et al. [166]), ∆Xi (∆T ) is normally distributed. The
mean can be calculated using Wald’s equation [161] which gives

h∆X(∆T )i = 0, (9.42)

while the variance can be calculated using the law of total variance [166] from which we
find that

h∆X(∆T )2 i = λ∆T hδX 2 i. (9.43)

Chapter 9 Timing noise in continuous gravitational waves: random walk models 209

If we label the parameter offset measured for non-overlapping blocks of data ∆T as

∆Xi (∆T ) then {∆X1 (∆T ), ∆X2 (∆T ), . . . } is a sequence of independent and identi-
cally distributed random variables with a mean and variance given by Eqn. (9.42) and
Eqn. (9.43) respectively.

In the model defined in Section 9.1, we defined

dXi = ∆Xi − ∆Xi−1 , (9.44)

as the difference between two adjacent subdomains of the piecewise Taylor expansion.
Moreover, we assumed that dXi (∆T ) was normally distributed with zero mean and
a variance σX 2 . This assumption is valid if the underlying random walk process is a

compound Poisson process. To see this, we note that if both ∆Xi and ∆Xi−1 can be
described by the compound Poisson process in Eqn. (9.41) and the assumption of a large
number of events in each is valid, then
   
∆T ∆T
dXi (∆T ) = ∆Xi − ∆Xi−1 , (9.45)
2 2

is the difference between normally distributed random variables. The factors of a half
are due to splitting the period of dXi (∆T ) into two equal durations. It can be shown
(see for example Weisstein [168]) that the difference between two normal distributions,
i.e. dXi (∆T ), is also normally distributed with a mean which is the difference in the
means of ∆Xi (∆T ) and ∆Xi−1 (∆T ):

hdX(∆T )i = 0, (9.46)

and a variance which is the sum of the variance of ∆Xi (∆T ) and ∆Xi−1 (∆T ):

∆T
hdX(∆T )2 i = 2λ hδX 2 i = λ∆T hδX 2 i. (9.47)
2

The metric-mismatch calculations derived in the previous two sections are written as a
2 ∈ [σ 2 , σ 2 , σ 2 ] which is the variance of dX(∆T ). We can therefore relate
function of σX φ f f˙
the two by

λ∆T hδX 2 i = σX
2
. (9.48)

Rearranging this equation, we can equate λhδX 2 i with a GW random walk strength
parameter which we define as

σφ2
SGWPN := λhδφ2 i = , (9.49)
∆T
σf2
SGWFN := λhδf 2 i = , (9.50)
∆T
σf2˙
SGWSN := λhδ f˙2 i = . (9.51)
∆T
210 Chapter 9 Timing noise in continuous gravitational waves: random walk models

We distinguish between the strength of gravitational wave noise (SGWPN , SGWFN and
SGWSN ) and the usual strength of noise in the rotational parameters (SPN , SFN and SPN )
defined in Eqn. (2.3) to Eqn. (2.5). To relate these, we need to define the mechanism of
gravitational wave emission. In the canonical non-axisymmetric distortion mechanism,
f = 2ν. Therefore the underlying jumps in signals are related by

δφ = 2δϕ, (9.52)
δf = 2δν, (9.53)
δ f˙ = 2δ ν̇, (9.54)

where ϕ, ν and ν̇ are the rotational phase, frequency, and spin-down rate. The variances
of the GW parameters are therefore equal to the variances of the rotational parame-
ters multiplied by a factor of 4. Using the definition the rotation noise strengths from
Eqn. (2.3) to Eqn. (2.5), we see that

SGWPN = 4SPN , (9.55)

SGWFN = 4SFN , (9.56)
SGWSN = 4SSN . (9.57)

We can now write the expectation of the fully-coherent metric-mismatch due to a random
walk in terms of these strength parameters. First, for a random walk in the phase given
by Eqn. (9.30) we take just the leading order term and substitute in Eqn. (9.49). This
results in
3
E[µ̃] = SGWPN Tobs . (9.58)
70
Second, for a random walk in the frequency, we take the leading order terms from
Eqn. (9.40) and substitute using Eqn. (9.50). This gives

π2 3
E[µ̃] = SGWFN Tobs . (9.59)
630
From this we see how, when the random walk is considered as a compound Poisson
random walk, the mismatch does not depend on ∆T . This justifies the random walk
treatment that was used in the earlier sections.

9.5 Application to the Crab pulsar

In Section 8.2, we introduced the Crab ephemeris. The regular and independent mea-
surements of the frequency and spin-down rate in the Crab ephemeris provide a unique
view of timing noise as a ‘jump’ in the phase, frequency, and spin-down rate each month.
Specifically by a jump we mean the discontinuity at the interface between two months
as illustrated in Figure 8.1. In this section, we will interpret data from the ephemeris in
the context of a random walk model. Note that in this section, we discus the rotational
frequency and frequency derivatives, which we denote by ν and ν̇ which we will relate
Chapter 9 Timing noise in continuous gravitational waves: random walk models 211

to the gravitational wave frequency f using the non-axisymmetric emission model for
which f = 2ν.

9.5.1 Distribution of jumps in the Crab ephemeris

We begin with a purely empirical look at the distribution of jumps in the Crab ephemeris.
We will use the jumps in frequency in the next section to interpret the Crab ephemeris
in the context of a random walk.

From the Crab ephemeris, there are three distributions which we will calculate here: the
jumps in frequency, spin-down rate, and phase. It is not meaningful to simply look at
the difference in frequency (for example) between any two months without adjusting for
the secular spin-down. But what we really want is the difference which occurs at the
interface. To calculate this we define νi (t) as the frequency according to the ith month
as evaluated at time t. Then if ∆ti = ti+1 − ti , the frequency jump between months is

dνi = νi+1 (ti+1 − ∆ti /2) − νi (ti + ∆ti /2) , (9.60)

"   # "   #
∆ti ∆ti 2 ν̈i+1 ∆ti ∆ti 2 ν̈i
= νi+1 − ν̇i+1 + − νi + ν̇i + . (9.61)
2 2 2 2 2 2

The Crab ephemeris does not include the second order spin-down, but since the first
order spin-down is not constant we assume a constant average second order spin-down
given by
1 X ν̇i+1 − ν˙i
ν̈av = , (9.62)
N ti+1 − ti
i

where N is the number of data points in the ephemeris file. Inserting this into the
Taylor expansion, the second order terms cancel, leaving a frequency jump between
months given by
∆ti
dνi = (νi+1 − νi ) − (ν̇i+1 + ν̇i ) . (9.63)
2
Note, this is not the difference in frequency between monthly updates, but the jump
between months in frequency at the interface as illustrated in figure 8.1.

Calculating the result of Eqn. (9.63) for all the data points in the Crab ephemeris we
plot the estimated probability density using the Gaussian kernel density estimate (KDE)
method [92] in Figure 9.4A. Note that we have filtered out differences which occur over
known glitches as described by Espinoza et al. [59]. This is because we are interested in
the timing noise activity and not the effect of glitches themselves.

Moving now to the spin-down, we can calculate the jump between months in a similar
way

dν̇i = ν̇i+1 (ti+1 − ∆ti /2) − ν̇i (ti + ∆ti /2) (9.64)
= (ν̇i+1 − ν̇i ) − ν̈av ∆ti . (9.65)

For the spin-down rate, we found a population of large negative jumps in the post-glitch
periods and hence are probably not related to the timing noise activity that we are
212 Chapter 9 Timing noise in continuous gravitational waves: random walk models

Mean = 6.1×10−10 s−1 Mean= −1.7×10−15 s−2 Mean= 0.0012

Variance= 3.1×10−16 s−1 Variance= 1.8×10−28 s−2 Variance= 5.0×10−5

A B C
Normalised counts

−0.8 0.0 0.8 −6 −3 0 3 6 −0.04 −0.02 0.00 0.02

×10−7 ×10−14
dν [s−1 ] dν̇ [s−2 ] dN

Figure 9.4: KDEs for the ‘jumps’ in frequency, spin-down rate, and residual
number of rotations between adjacent per-month signals in the Crab ephemeris.
Adjacent signals over glitches are filtered along with large anomalous values of
dN once it was confirmed they occur within 200 days after a glitch.

interested in. As such in Figure 9.4B, we filter out these anomalous results to show the
KDE for data points likely to be related to timing noise.

For the phase, we can use that each reference time given in the ephemeris coincides with
a pulse. Therefore between two adjacent references times, the star has undergone an
integer number of rotations. We can evaluate how well the ephemeris performs here by
calculating the residual number of rotations between the timing model at a given step
and the pulse arrival time at the next step. The data in the ephemeris file does not
directly provide information on the phase evolution. It provides the independent phase
evolution in each month with the phase at the reference time being zero. To calculate
the full phase evolution between two reference times, we need to calculate the phase
difference between each reference time and the interface time between them. Take this
interface to be halfway between such that tmid = (ti + ti+1 )/2. Then the total number
of rotations between two reference times is
1
N= ((φi (tmid ) − φi (ti )) − (φii +1 (tmid ) − φi+1 (ti+1 ))) . (9.66)
2π
1
We set the phase at the reference time to zero, so this leaves N = 2π (φi (tmid ) − φi+1 (tmid ))
where the terms are explicitly given by
 
ν̇i 2 ν̈av 3
φi (tmid ) = 2π (tmid − ti )νi + (tmid − ti ) + (tmid − ti ) (9.67)
2! 3!
 
ν̇i+1 2 ν̈av 3
φi+1 (tmid ) = 2π (tmid − ti+1 )νi+1 + (tmid − ti+1 ) + (tmid − ti+1 ) . (9.68)
2! 3!

The total number of rotations N between months is a function of the length of a given
month and the spin-down parameters. Calculating this allows us to check how well
phase-connected lines of the ephemeris are. We quantify this by the residual number of
Chapter 9 Timing noise in continuous gravitational waves: random walk models 213

rotations, defined as
dN = N − round(N ), (9.69)

where by ‘round‘ we indicate rounding to the nearest integer number of rotations.

In Figure 9.4C we plot the Gaussian KDE of ∆N having filtered against known glitch
events. We found four jumps where 0.1 < ∆N < 1.0; again these were found to be
occur within ∼ 200 days of known glitches and so were removed to focus attention on
the timing noise activity.

Figure 9.4 shows that the distribution of jumps in the frequency, spin-down rate, and
residual number of rotations is centred on zero, as expected. The interesting part to
note here is the size of the standard deviations. These can potentially be used to test
timing noise models. In the next section, we interpret these in the context of a random
walk.

9.5.2 Predicting the mismatch in the Crab

Studies of the Crab pulsar consistently find that the random walk noise is best fit by
a frequency like noise process. In Table 9.1, we list the strengths reported from three
works on the issue; this is by no means a complete literature review, but shows the
typical values and spread.

SFN Hz2 /s
Boynton et al. [33] 0.9 × 10−22
Groth [73] 0.53 × 10−22
Cordes [43] 0.66 × 10−22

Table 9.1: Values for the strength of frequency noise in the Crab pulsar found
in the literature: all three authors demonstrated that this strength was robust
to changes in ∆T , the time which the data is divided into to calculate statistical
quantities.

We can also derive our own estimate for the strength of frequency like timing noise in
the Crab. To do this, we recall that from Eqn. (2.4) we have

SFN = λhδν 2 i. (9.70)

Then, as we did in Eqn. (9.50) for the gravitational wave frequency, we have that

hdν(∆T )2 i
SFN = . (9.71)
∆T

Substituting hdν(∆T )2 i = 3.1 × 10−16 s−1 (taken from Figure 9.4A) into Eqn. (9.71)
with ∆T = 30 days, this gives

SFN = 1.20 × 10−22 Hz2 /s. (9.72)

214 Chapter 9 Timing noise in continuous gravitational waves: random walk models

This strength is of the same order of magnitude as those listed in Table 9.1. We are not
able to test if this strength is invariant to changes in ∆T because the Crab ephemeris is
only calculated in ∆T = 1 month long intervals; we therefore cannot confirm that the
Crab undergoes frequency-like noise.

In Figure 8.5 we have already shown the dependence of the mismatch on the observation
time for the Crab; fitting a power law to this we found that the mismatch scaled as
 2.88
−23 Tobs
hµifit ≈ 1.5 × 10 . (9.73)
1 sec

We can compare this directly with Eqn. (9.59). The power, being close to 3, suggests
some similarity with the frequency like random walk in the mismatch. Now we can
additionally predict this dependence given a value for the strength of frequency noise.
This is done in Figure 9.5 for our estimated value of the strength parameter and using
that SGWFN = 4SFN (see Eqn. (9.56)). The prediction using the strength parameter

1.2
SFN = 1.20×10−22 Hz2 /s
1.0

0.8

0.6
hµmin i
0.4

0.2

0.0

−0.2
0 100 200 300 400 500 600
Observation time [Days]

Figure 9.5: This figure, updated from Figure 8.5, shows the dependence of
the minimum mismatch on the observation time for the Crab, along with the
prediction of Eqn. (9.59) for the our estimate of the frequency noise strength
parameter calculated in Eqn. (9.72).

calculated from the Crab ephemeris, as given in Eqn. (9.72), overestimates the depen-
dence of the mismatch calculated exactly from the data. It is unclear at the time what
causes this, but we have nevertheless demonstrated the essential elements of making
such predictions which may be useful in the future, if there is a clear case where the
timing noise is well modelled by a random walk.

9.6 Conclusion

In this Chapter, we have calculated the expectation of the fully-coherent mismatch when
searching for a GW signal which undergoes a random walk in one of the phase, frequency,
Chapter 9 Timing noise in continuous gravitational waves: random walk models 215

or spin-down rate. We did this first for a system in which the difference between the
signal and template was initially zero and then grew as a random walk. Since this is not
a minimised result, we then demonstrated how to minimise the mismatch with respect
to the template parameters ft and f˙t . The formulae derived in this section have been
verified against Monte-Carlo type simulations of the exact mismatch.

In Section 9.4, we related the random walk model to a more physical compound Poisson
random walk. Using this description, we formulated the two key results of this chapter:
Eqn. (9.58) and Eqn. (9.59). These are the expectation of the minimum metric-mismatch
due to random walk in the phase and frequency, respectively, as a function of the strength
of phase and frequency noise.

Following this, we developed our understanding of random walk models in the context of
the data from the Crab ephemeris. We showed that the frequency noise strength can be
measured from this data and has a value consistent with other results in the literature.
We then demonstrated the use of this derived prediction of the strength of frequency
noise and Eqn. (9.59) to predict the results found in Chapter 8 for the dependence of
the mismatch on the observation time. The prediction is not perfect, but does provide
a rough order of magnitude figure of the expected mismatch.

Appendix 9.A Summation identities

In this appendix, we derive two useful summation identities used in Sec. 9.2. First, we
have
c X
X b
Xa = (X1 ) + (X1 + X2 ) + . . . + (X1 + X2 + . . . + Xc−1 + Xc ) (9.74)
b=1 a=1
= cX1 + (c − 1)X2 + . . . + 2Xc−1 + Xc (9.75)
Xc
= (c + 1 − b)Xb , (9.76)
b=1
216 Chapter 9 Timing noise in continuous gravitational waves: random walk models

which is used in deriving Eqn. (9.9). Second, we have

j−1
i−1 X
X
(j − k)Xk =[0] + [X1 ] + [2X1 + X2 ] + [3X1 + 2X2 + X3 ] + . . .
j=1 k=1

+ [(i − 2) X1 + (i − 3) X2 + (i − 4) X3 + . . .
+3Xi−4 + 2Xi−3 + Xi−2 ] (9.77)
= (1 + 2 + 3 + . . . + (i − 4) + (i − 3) + (i − 2)) X1
+ (1 + 2 + 3 + . . . + (i − 4) + (i − 3)) X2
+ (1 + 2 + 3 + . . . + (i − 4)) X3 + . . .
+ (1 + 2 + 3)Xi−4 + (1 + 2)Xi−3 + Xi−2 (9.78)
i−2
X i−3
X 2
X 1
X
= kX1 + kX2 + . . . + kXi−3 + kXi−2 (9.79)
k=1 k=1 k=1 k=1
i−2 i−1−j
! i−2
X X 1X
= k Xj = (i − j)(i − j − 1)Xj , (9.80)
2
j=1 k=1 j=1

which is used in deriving Eqn. (9.10).

Appendix 9.B Least-squares minimisation of a random walk

In this appendix, we will describe the process of fitting and removing a polynomial from
N data points (xi , yi ) which undergoes a random walk. The polynomial will be fitted
using a least squares minimisation. The xi are the independent points at which yi (which
undergoes a random walk) is measured. We begin by defining the least-squares fitting
method then go on to calculate the residual for several different degrees of polynomial.
This introduces the method in a generic setting which is then applied in Section 9.3 to
calculate the minimised mismatch for a GW signal undergoing a random walk.

9.B.1 Least squares fitting of a polynomial

Given N data points xi , yi , we define the residual from a least-squares polynomial fit of
order k, as
(k) (k)
ri = yi − yi , (9.81)

where
(k)
yi = a0 + a1 xi + a2 x2i + · · · + ak xki , (9.82)

is a polynomial of degree k.

Then the residual which we want to minimise is

N 
X  N 
X  2
2 (k) 2
R = ri = yi − a0 + a1 xi + a2 x2i + · · · + ak xki . (9.83)
i=1 i=1
Chapter 9 Timing noise in continuous gravitational waves: random walk models 217

Partial differentiation with respect to the parameters ai , yields k simultaneous equations.

Writing these as a matrix and then solving for the best fit, ŷ ( k)i , it can be shown (see
for example Weisstein [167]) that
 
1 x1 x21 . . . xk1
 
(k)
−1 T  1 x2 x22 . . . xk2 
ŷi = X X X T
X yi where 
X= . . .. .. ..  (9.84)

 .. .. . . . 
1 xn x2n . . . xkn

Here X is an example of a Vandermonde matrix in which the terms follow a geometric

progression. It is useful to note that
 PN PN k 
N i=1 x i · · · xi
 PN PN 2
PNi=1 k+1 
 x
i=1 i i=1 ix · · · i=1 i x 
T
XX =   .. .. .. . (9.85)
.. 
 . . . . 
PN k PN k+1 PN 2k
i=1 xi i=1 xi ··· i=1 xi

Provided that the xi are suitably defined, then an analytic fit can be found for any k,
the difficulty lies in inverting the matrix.

9.B.2 Least squares fitting a polynomial to a random walk

We now take the xi , yi to be a Gaussian random walk beginning at the origin. To define
this, let dyi ∼ N (0, σ 2 ) be independent and identically distributed random variables for
which their sum generates the random walk:
i
X
yi = dyi . (9.86)
j=1

We also set each random walk event to occur according to xi = i∆x. Then the residual
after fitting and removing a k th order polynomial to the random walk yi , is
(k) (k)
−1
ri = yi − ŷi = yi − X X T X X T yi . (9.87)

This suggests the residual will be similar to the random walk, but modified by the least
squares fitting. To illustrate this, in Figure 9.6 we plot a simulated random walk along
with several fits.

9.B.3 Zeroth order fitting

We begin with the case of k = 0 in which X T = [1, 1, . . . 1] such that

−1 1
X XT X XT = JN (9.88)
N
218 Chapter 9 Timing noise in continuous gravitational waves: random walk models

k=0
k=1
k=2

0
y ∆y (k)0

Random walk
k=0
k=1
k=2

x x

Figure 9.6: Example of a random walk on the left along with three polynomial
fits of varying order. On the right is the corresponding residual after subtracting
these fits. A dotted line marks the origin in both plots.

where JN is the N × N matrix of ones. Inserting this into Eqn. (9.87), the residual from
a zeroth order fit is given by
N
(0) 1 X
ri = yi − yj . (9.89)
N
j=1

The zeroth order residual can be interpreted as the removing the average value hyi i from
the random walk: this was illustrated in Figure 9.6.

We can now take expectations to understand the behaviour of the residual when com-
pared to the original definition of the random walk in Eqn. (9.86). For example, consider
the mean square translation distance from the origin of a random walk after i steps. For
a normal random walk, this has the well known result

E[yi2 ] = iσ 2 . (9.90)

We can calculate the corresponding quantity of the k = 0 residual by first noting that
" i j
#
X X
E [yi yj ] = E dyk dyl (9.91)
k=1 l=1
i
XX j
= E [dyk dyl ] (9.92)
k=1 l=1
Xi X j
= δkl σ 2 (9.93)
k=1 l=1
2
= σ min(i, j), (9.94)
Chapter 9 Timing noise in continuous gravitational waves: random walk models 219

where δkl is the Kronecker delta. Then we have

  2 X
N XN X N
(0) 2
ri = yi2 − yi yk + N −2 yk yl (9.95)
N
k=1 k=1 l=1
i N
! N k N
!
X X X X X
= yi2 − 2N −1 yi yk + yi yk + N −2 yk yl + yk yl .
k=1 k=i+1 k=1 l=1 l=k+1
(9.96)

Taking the expectation we have

  ! !!
 2
i
X N
1 X X
N
Xk XN
(0) 2 2
E ri =σ i− k+ i + 2 l+ k (9.97)
N N
k=1 k=i+1 k=1 l=1 l=k+1
 
2 N 1 i2 i 1
=σ −i+ + − + . (9.98)
3 2 N N 6N

This result can be compared with Eqn. (9.90), the expectation of the squared value for
a random walk. In contrast, the expectation after i steps for the residual random walk
depends on the length of data N that was fitted. It can be shown the expectation has
a minimum at i = N/2.

To further understand the difference between the random walk and the residual random
walk, let us consider the sum of squares after N steps for the random walk
" N
# N
X X 1

E yi2 = iσ 2 = N 2 + N σ2. (9.99)
2
i=1 i=1

On the other hand, the sum of squares for the residual random walk is given by
" N 
#
X 2 1

(0)
E ri = N 2 − 1 σ2. (9.100)
6
i=1

Comparing equations (9.99) and (9.100) we note that, for the leading order term, the
coefficient is reduced, but the power remains the same.

9.B.4 First order fitting

We now consider a first order fitting for which

 
1 ∆x
 
(1)
−1  1 2∆x 
ŷi =X X X T
X yiT
with X=
 .. .. .
 (9.101)
 . . 
1 N ∆x
220 Chapter 9 Timing noise in continuous gravitational waves: random walk models

Inserting the definitions of xi we can write

" #

6
−1 1 4N + 2 − ∆x
XT X = 6 12 = C (1) . (9.102)
N (N − 1) − ∆x ∆x2 (N +1)

For convenience we have defined a symmetric matrix C (1) . We then proceed to define
another matrix
(1)
Cij := XC (1) X T (9.103)
 
1 ∆x
  " (1) (1)
#" #
 1 2∆x  C11 C12 1 1 ... 1
= .. .. 
 C (1) C (1) (9.104)
 . .  21 22
∆x 2∆x . . . N ∆x
1 N ∆x
   
2 3 ... N +1 1 2 ... N
 ..   .. 
(1) (1)  3 4 ... .   2 4 ... . 
= C11 JN + C12 ∆x   + C (1) ∆x2  
 ..  22  .. 
 .   . 
N + 1 ... ... 2N N ... ... N2
(9.105)

(1)
We can write ri as a summation by inferring the dependence of the ith row of each
matrix on the j th column
N
X
(1) (1) (1) (1) (1) (1)
ri = yi − Cij yj where Cij = C11 + C12 ∆x(i + j) + C22 ∆x2 ij (9.106)
j=1

We have now defined the first order residual. To understand that fitting an removing a
first order polynomial has, we compute the expectation of the square for the ith term
  
(1) 2 1
E ri = 2N 4 − 18N 3 i + 9N 3 + 78N 2 i2 − 78N 2 i
15N (N 2 − 1)
(9.107)
+ 14N 2 − 120N i3 + 180N i2 − 78N i


+9N + 60i4 − 120i3 + 78i2 − 18i + 2 ,

which can be compared with the classic result for a random walk given in Eqn. (9.90).
Alternatively, comparing with Eqn. (9.99), the expected sum of squares for the residual
random walk is "N #
X  (1) 2 1

E ri = N 2 − 4 σ2. (9.108)
15
i=1

As in the zeroth order fit, the power of the leading order term remains the same, but
the coefficient decreases.
Chapter 9 Timing noise in continuous gravitational waves: random walk models 221

9.B.5 Second order fitting

For the residual left after removing a quadratic, the argument proceeds in much the
same way with
N
X
(2) (2)
ri = yi − Cij yj , (9.109)
j=1

where
(2) (2) (2) (2) (2)
Cij =C11 + C22 ∆x2 ij + C33 ∆x4 i2 j 2 + C12 ∆x(i + j)
(2) (2)
+ C13 ∆x(i2 + j 2 ) + C23 ∆x3 (ij 2 + i2 j), (9.110)

and
 1 30

9N 2 + 9N + 6 − ∆x (36N + 18) ∆x2
1 − 1 (36N + 18) 12(2N +1)(8N +11) 
C (2) =  ∆x ∆x2 (N +1)(N +2)
− ∆x3180(N +2)  .
N (N − 1)(N − 2) 30
∆x2
− ∆x3180
(N +2)
180
∆x4 (N +1)(N +2)
(9.111)

The expression for the expected square value is too long to write out in full, but the
expected sum of squares for the residual random walk is
" N 
#
X 2 1

(2)
E ri = 3N 2 − 27 σ 2 , (9.112)
70
i=1

for which as in the case of the zeroth order and first order residuals, the power of the
leading order term remains unchanged, but the coefficient decreases.

9.B.6 Conclusions

We now have a method to calculate statistical quantities from the residual left over
after subtracting a k th order polynomial from a random walk. Considering the sum of
squares for a random walk and the residuals in equations (9.99), (9.100), (9.108), and
(9.112) we find that the leading order term retains the same power of N with increasing
k but the coefficient of this power gets smaller. This reflects the improved fitting with
the polynomial degrees. We also note that with each increase in the order of fit we get
a limit on N for which the sum of squares is positive. For zeroth order fitting this is
N > 1, for first order N > 2 and for second order N > 3. This is because in order to
perform a least squares fit, we need at least k + 1 points to fit.
Chapter 10

Conclusion and outlook

Timing variations in pulsars have long been used as a way to infer the elemental prop-
erties of neutron stars. However, while in isolated cases models for glitches and timing
noise have successfully explained the observations, we are some way from having a com-
plete and universal understanding of neutron stars. For timing noise, this may be due to
the varied ways in which it manifests, but in the comprehensive review by Hobbs et al.
[81] the authors suggest that over sufficiently long timescales (which may be much longer
than we are able to observe) timing residuals show quasi-periodic features. Moreover,
some pulsars (see Lyne et al. [111] for examples) have pronounced periodic modulations
in not only their timing residuals, but also the shape of pulsations, PSR B1828-11 is the
best example of this. With this in mind, in this thesis we investigated two mechanisms
of strictly periodic variations in detail: magnetospheric switching and precession.

We began in Chapter 3 by studying the action of the electromagnetic torque on pre-

cessing pulsars using numerical solutions to the Euler rigid-body equations coupled to
the Deutsch [54] electromagnetic torque. This allowed us to investigate the role of the
anomalous component of the torque and conclude that, for realistic neutron stars, this
can safely be ignored. We cleared up some confusion in the literature, by showing that
the ‘persistent precession’ solutions found by Melatos [122] were in fact not precessing,
but aligned with the principal axes of the effective body frame arising from the inclusion
of the anomalous torque.

In order to test models, they need to be predictive, in the sense that they model the
features of a neutron star observed by pulsar astronomers. To address this in the context
of precession, in Chapter 4 we developed numerical solutions which, together with solving
the Euler rigid-body equations, solve for the Euler rotation angles which transform from
the rotating frame to the inertial frame. This allowed us to directly model the phase
residual, spin-down rate, and pulse profile of a precessing pulsar. We compared the
numerical model against analytic solutions from the literature and against our derivation
of the spin-down rate which we used later in Chapter 5. We finished this chapter by
discussing some preliminary results from a hybrid model which couples precession to
magnetospheric torque switching. In the future, it would be interesting to explore this
further by understanding better how reconfiguring the magnetosphere will effect the

223
224 Chapter 10 Conclusion and outlook

Deutsch [54] torque and then testing different mechanisms to bias, in a probabilistic
way, the switching using the precessional cycle. The numerical model is perfectly suited
to this task as it captures the complicated feedback between the torque switching and
precession.

PSR B1828-11 stands out in the literature for the strong ∼ 500 day periodic modulations
observed in its timing properties. As a result, it has been used as evidence for both
precession [156] and periodic magnetospheric torque switching [111]. Jones [89] and
Cordes [45] suggested that these models may not be mutually exclusive, which prompted
our hybrid model discussed in Chapter 4. Neglecting these hybrid models, precession
and magnetospheric switching are mutually exclusive, but it is important to know which
is favoured since both have important implications for neutron star physics. To decide
this, in Chapter 5, we applied a Bayesian model comparison. We found an odds-ratio of
102.7±0.5 in favour of the precession model, a key result of this thesis which we published
in Ashton et al. [25]. This does not rule out the switching interpretation entirely as we
have not tested an exhaustive set of models, but it does provide a quantitative framework
to evaluate models. In this chapter we focus primarily on the methods used to ensure
that we make an unbiased comparison, although there is some development for the
modulations of the beam-width due to precession. In the future, we intend to use these
tools to test further modifications of the models. Furthermore, we would like to combine
the data from other pulsars with long period modulations, such as PSR 0919+06 [134],
and repeat the model comparison to see how this changes the odds-ratio.

Detecting gravitational waves from an isolated neutron star would provide a unique
opportunity to learn about them, especially if we additionally observed the object from
its electromagnetic output. However, signals which are subject to timing variations may
be difficult to detect if, as in the case of many current gravitational wave searches, we
use matched filtering templates which do not include timing variations. This issue has
not been properly tackled in the literature and so in the last chapters of this thesis we
begin the process of quantifying the risks posed by timing variations to efforts to detect
gravitational waves from neutron stars.

In Chapter 6 we set out the tools which we will use to calculate the mismatch, defined
to be the loss of signal to noise ratio. In doing so we defined a new approach, the
generalised metric-mismatch, which can calculate the mismatch for any arbitrary signal
by approximating it as a piecewise Taylor expansion.

Searches for gravitational waves from known pulsars can handle glitches in the signal
since they can see that they have occurred. In Chapter 7 we showed that for blind
searches, where we have no electromagnetic signal, glitches pose a substantial risk, espe-
cially since many searches target young rapidly spinning-down pulsars which we found
to be likely to have larger and more frequent glitches. Many of these searches use an
initial semi-coherent stage and then follow-up candidates with fully-coherent searches.
We showed how a signal can be identified as a candidate in the semi-coherent stage and
then subsequently lost in the follow-up. In the future, we would like to develop search
methods which are robust to glitches. These could then be applied to past data to ensure
that we have not already missed a continuous gravitational wave signal.
Chapter 10 Conclusion and outlook 225

Quantifying the effect of timing noise on gravitational wave searches is a more difficult
task due to the fact that we have no universal empirical description of what constitutes
timing noise; therefore, it is difficult to predict what timing noise may exist in gravi-
tational wave signals from isolated neutron stars. One way to approach this, discussed
in Chapter 8, is to use data on the frequency and spin-down rate evolution of the Crab
pulsar to generate ‘noisy gravitational wave signals’. We searched for these using stan-
dard matched filtering tools and calculated the minimum mismatch having searched in a
narrow band of frequency and spin-down rate. The conventional wisdom for continuous
wave searches contends that, since the SNR scales with the square-root of the observa-
tion time, provided we look for a sufficiently long time the signal will eventually become
detectable. However, by investigating how the minimum mismatch scaled with the ob-
2.88 . This result, published in
servation time for noisy signals, we found that hµ̃min i ∝ Tobs
Ashton et al. [24], means that for noisy signals the conventional wisdom does not hold.
There is in fact an observation time after which the SNR decreases with observation
time; for the Crab pulsar this was found to be ≈ 600 days.

In Chapter 9 we approached the issue of timing noise in continuous gravitational waves

from a different angle. Namely, by modelling timing noise as a random walk in the phase,
frequency, or spin-down rate: a phenomenological description dating back to Boynton
et al. [33]. Recently, this model has been disfavoured [81] as a substantive explanation
of timing noise, but it nevertheless provides a simple empirical description which is
consistent with the timing residuals of many pulsars. Moreover, while we apply it to
isolated neutron stars, it can also be applied to low-mass X-ray binary systems where
spin-wandering due to fluctuations in the torque can be modelled by a random walk. In
this chapter, we calculated the mismatch due to random walks and showed how to include
the minimisation step modelling a narrow-band search over the template frequency and
spin-down rate. We go on to use the Crab ephemeris to predict the strength of frequency
noise in the Crab pulsar and hence the dependence of the mismatch on observation time
found in Chapter 8. This will be useful in developing models that we would like to use
in the future to estimate the levels of mismatch for other searches which may be affected
by random walk models of timing noise.

In this thesis, we hope to have developed the understanding of timing noise in neutron
stars. Most significantly, by providing a framework with which to debate the merits of
models explaining periodic signals and quantify their comparison. We also hope that by
understanding the role of timing variations in continuous gravitational waves, we help
to guide efforts to detect these elusive signals in the future.
Bibliography

[1] Aasi, J., Abadie, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R., Acca-
dia, T., Acernese, F., Adams, C., Adams, T., et al. (2013a). Directed search for contin-
uous gravitational waves from the Galactic center. Physical Review D, 88(10):102002.

[2] Aasi, J., Abadie, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R.,
Accadia, T., Acernese, F., Adams, C., Adams, T., et al. (2014). Application of a
Hough search for continuous gravitational waves on data from the fifth LIGO science
run. Classical and Quantum Gravity, 31(8):085014.

[3] Aasi, J., Abadie, J., Abbott, B. P., Abbott, R., Abbott, T. D., Abernathy, M.,
Accadia, T., Acernese, F., Adams, C., Adams, T., et al. (2013b). Einstein@Home
all-sky search for periodic gravitational waves in LIGO S5 data. Physical Review D,
87(4):042001.

[4] Aasi, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R., Acernese, F.,
Ackley, K., Adams, C., Adams, T., Addesso, P., et al. (2015a). Directed search for
gravitational waves from Scorpius X-1 with initial LIGO data. Physical Review D,
91(6):062008.

[5] Aasi, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R., Acernese, F.,
Ackley, K., Adams, C., Adams, T., Addesso, P., et al. (2015b). Narrow-band search
of continuous gravitational-wave signals from Crab and Vela pulsars in Virgo VSR4
data. Physical Review D, 91(2):022004.

[6] Aasi, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R., Acernese, F.,
Ackley, K., Adams, C., Adams, T., Addesso, P., et al. (2015c). Searches for continuous
gravitational waves from nine young supernova remnants. The Astrophysical Journal,
813(1):39.

[7] Aasi, J., Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M. R., Acernese, F.,
Ackley, K., Adams, C., Adams, T., Addesso, P., et al. (2016). First low frequency all-
sky search for continuous gravitational wave signals. Physical Review D, 93(4):042007.

[8] Abadie, J., Abbott, B., Abbott, R., Abbott, T., Abernathy, M., Accadia, T., Acer-
nese, F., Adams, C., Adhikari, R., Affeldt, C., et al. (2012). All-sky search for periodic
gravitational waves in the full S5 LIGO data. Physical Review D, 85(2):022001.

[9] Abadie, J., Abbott, B. P., Abbott, R., Abernathy, M., Accadia, T., Acernese, F.,
Adams, C., Adhikari, R., Affeldt, C., Allen, B., et al. (2011). Beating the spin-down

227
228 BIBLIOGRAPHY

limit on gravitational wave emission from the Vela pulsar. The Astrophysical Journal,
737(2):93.

[10] Abbott, B., Abbott, R., Adhikari, R., Agresti, J., Ajith, P., Allen, B., Amin, R.,
Anderson, S. B., Anderson, W., Arain, M., et al. (2007). Searches for periodic grav-
itational waves from unknown isolated sources and Scorpius X-1: Results from the
second LIGO science run. Physical Review D, 76(8):082001.

[11] Abbott, B., Abbott, R., Adhikari, R., Ajith, P., Allen, B., Allen, G., Amin, R.,
Anderson, S. B., Anderson, W. G., Arain, M. A., et al. (2008). Beating the spin-down
limit on gravitational wave emission from the Crab pulsar. The Astrophysical Journal
Letters, 683(1):L45.

[12] Abbott, B. P., Abbott, R., Abbott, T., Abernathy, M., Acernese, F., Ackley,
K., Adams, C., Adams, T., Addesso, P., Adhikari, R. X., et al. (2016). Observa-
tion of gravitational waves from a binary black hole merger. Physical review letters,
116(6):061102.

[13] Abbott, B. P., Abbott, R., Acernese, F., Adhikari, R., Ajith, P., Allen, B., Allen,
G., Alshourbagy, M., Amin, R., Anderson, S., et al. (2010). Searches for gravitational
waves from known pulsars with science run 5 LIGO data. The Astrophysical Journal,
713(1):671.

[14] Akgün, T., Link, B., and Wasserman, I. (2006). Precession of the isolated neutron
star PSR B1828-11. Monthly Notices of the Royal Astronomical Society, 365:653–672.

[15] Alpar, M. A., Pines, D., Anderson, P. W., and Shaham, J. (1984). Vortex creep
and the internal temperature of neutron stars. i-general theory. The Astrophysical
Journal, 276:325–334.

[16] Anderson, P. W. and Itoh, N. (1975). Pulsar glitches and restlessness as a hard
superfluidity phenomenon. Nature.

[17] Andersson, N. (2003). Gravitational waves from instabilities in relativistic stars.

Classical and Quantum Gravity, 20(7):R105.

[18] Andersson, N., Glampedakis, K., Ho, W. C. G., and Espinoza, C. M. (2012). Pulsar
glitches: The crust is not enough. Physical Review Letters, 109:241103.

[19] Andersson, N. and Kokkotas, K. D. (2001). The r-mode instability in rotating

neutron stars. International Journal of Modern Physics D, 10(04):381–441.

[20] Archibald, A. M., Stairs, I. H., Ransom, S. M., Kaspi, V. M., Kondratiev, V. I.,
Lorimer, D. R., McLaughlin, M. A., Boyles, J., Hessels, J. W., Lynch, R., et al. (2009).
A radio pulsar/X-ray binary link. Science, 324(5933):1411–1414.

[21] Archibald, R. F., Kaspi, V. M., Ng, C.-Y., Gourgouliatos, K. N., Tsang, D., Scholz,
P., Beardmore, A. P., Gehrels, N., and Kennea, J. A. (2013). An anti-glitch in a
magnetar. Nature, 497(7451):591–593.
BIBLIOGRAPHY 229

[22] Arzamasskiy, L., Philippov, A., and Tchekhovskoy, A. (2015). Evolution of non-
spherical pulsars with plasma-filled magnetospheres. Monthly Notices of the Royal
Astronomical Society, 453:3540–3553.

[23] Arzoumanian, Z., Nice, D. J., Taylor, J. H., and Thorsett, S. E. (1994). Timing
behavior of 96 radio pulsars. The Astrophysical Journal, 422:671–680.

[24] Ashton, G., Jones, D. I., and Prix, R. (2015). Effect of timing noise on targeted
and narrow-band coherent searches for continuous gravitational waves from pulsars.
Physical Review D, 91(6):062009.

[25] Ashton, G., Jones, D. I., and Prix, R. (2016). Comparing models of the periodic
variations in spin-down and beamwidth for PSR B1828-11. Monthly Notices of the
Royal Astronomical Society, 458:881–899.

[26] Baade, W. and Zwicky, F. (1934). Cosmic Rays from Super-novae. Proceedings of
the National Academy of Science, 20:259–263.

[27] Bailes, M., Lyne, A. G., and Shemar, S. L. (1993). Limits on pulsar planetary
systems from the Jodrell Bank timing database. In Phillips, J. A., Thorsett, S. E.,
and Kulkarni, S. R., editors, Planets Around Pulsars, volume 36 of Astronomical
Society of the Pacific Conference Series, pages 19–30.

[28] Baym, G. and Pines, D. (1971). Neutron starquakes and pulsar speedup. Annals
of Physics, 66(2):816–835.

[29] Biryukov, A., Beskin, G., and Karpov, S. (2012). Monotonic and cyclic compo-
nents of radio pulsar spin-down. Monthly Notices of the Royal Astronomical Society,
420(1):103–117.

[30] Bisnovatyi-Kogan, G. S., Mersov, G. A., and Sheffer, E. K. (1990). Model of the
35-day cycle in the X-Ray Binary Hercules X-1. Soviet Astronomy, 34:44.

[31] Bisnovatyj-Kogan, G. and Kahabka, P. (1993). Period variations and phase residuals
in freely precessing stars. Astronomy & Astrophysics, 267:L43–L46.

[32] Bombaci, I. (1996). The maximum mass of a neutron star. Astronomy & Astro-
physics, 305:871.

[33] Boynton, P., Groth, E., Hutchinson, D., Nanos Jr, G., Partridge, R., and Wilkinson,
D. (1972). Optical timing of the Crab pulsar, NP 0532. The Astrophysical Journal,
175:217.

[34] Boynton, P. E., Groth, III, E. J., Partridge, R. B., and Wilkinson, D. T. (1969).
Apparent change in frequency of NP 0532. IAU Circ., 2179.

[35] Brady, P. R., Creighton, T., Cutler, C., and Schutz, B. F. (1998). Searching for
periodic sources with LIGO. Physical Review D, 57:2101–2116.

[36] Brits, L. (2010). Euler angles. https://commons.wikimedia.org/wiki/File:

Eulerangles-alternative.svg.
230 BIBLIOGRAPHY

[37] Burgay, M., D’Amico, N., Possenti, A., Manchester, R., Lyne, A., Joshi, B.,
McLaughlin, M., Kramer, M., Sarkissian, J., Camilo, F., et al. (2003). An increased
estimate of the merger rate of double neutron stars from observations of a highly
relativistic system. Nature, 426(6966):531–533.

[38] Chadwick, J. (1932). The existence of a neutron. Proceedings of the Royal Society
of London. Series A, 136(830):692–708.

[39] Chamel, N. and Haensel, P. (2008). Physics of neutron star crusts. Living Rev.
Relativity, 11(10).

[40] Cheng, K. (1987). Outer magnetospheric fluctuations and pulsar timing noise. The
Astrophysical Journal.

[41] Chiuderi, C. and Occhionero, F. (1970). Shape of the Crab Pulsar and its Period
Fluctuations. Nature, 226:337–338.

[42] Coles, W., Hobbs, G., Champion, D. J., Manchester, R. N., and Verbiest, J. P. W.
(2011). Pulsar timing analysis in the presence of correlated noise. Monthly Notices of
the Royal Astronomical Society, 418:561–570.

[43] Cordes, J. M. (1980). Pulsar timing. II-Analysis of random walk timing noise-
Application to the Crab pulsar. The Astrophysical Journal, 237:216–226.

[44] Cordes, J. M. (1993). The detectability of planetary companions to radio pulsars.

In Planets around pulsars, pages 43–60.

[45] Cordes, J. M. (2013). Pulsar State Switching from Markov Transitions and Stochas-
tic Resonance. ApJ, 775:47.

[46] Cordes, J. M. and Downs, G. S. (1985). JPL pulsar timing observations. III-Pulsar
rotation fluctuations. The Astrophysical Journal Supplement Series, 59:343–382.

[47] Cordes, J. M. and Greenstein, G. (1981). Pulsar timing. IV-Physical models for
timing noise processes. The Astrophysical Journal, 245:1060–1079.

[48] Cordes, J. M. and Helfand, D. J. (1980). Pulsar timing. iii-timing noise of 50 pulsars.
The Astrophysical Journal, 239:640–650.

[49] Cordes, J. M. and Shannon, R. M. (2008). Rocking the lighthouse: circumpulsar

asteroids and radio intermittency. The Astrophysical Journal, 682(2):1152.

[50] Craft, Jr., H. D. (1970). Radio Observations of the Pulse Profiles and Dispersion
Measures of Twelve Pulsars. PhD thesis, CORNELL UNIVERSITY.

[51] d’Agostino, R. B. (1971). An omnibus test of normality for moderate and large size
samples. Biometrika, 58(2):341–348.

[52] Davis, L. and Goldstein, M. (1970). Magnetic-dipole alignment in pulsars. The

Astrophysical Journal, 159.

[53] Dergachev, V. (2010). On blind searches for noise dominated signals: a loosely
coherent approach. Classical and Quantum Gravity, 27(20):205017.
BIBLIOGRAPHY 231

[54] Deutsch, A. (1955). The electromagnetic field of an idealized star in rigid rotation
in vacuo. Annales d’Astrophysique.

[55] Dodson, R. G., McCulloch, P. M., and Lewis, D. R. (2001). High time resolution
observations of the January 2000 glitch in the Vela pulsar. The Astrophysical Journal
Letters, 564(2):L85.

[56] Duncan, R. C. and Thompson, C. (1996). Magnetars. In High velocity neutron

stars and gamma- ray bursts, volume 366, pages 111–117. AIP Publishing.

[57] Edwards, R. T., Hobbs, G. B., and Manchester, R. N. (2006). TEMPO2, a new
pulsar timing package - II. The timing model and precision estimates. Monthly Notices
of the Royal Astronomical Society, 372:1549–1574.

[58] Einstein, A. (1916). Approximative integration of the field equations of gravitation.

Sitzungsber. Preuss. Akad. Wiss. Berlin (Math. Phys.), 688:1916.

[59] Espinoza, C. M., Lyne, A. G., Stappers, B. W., and Kramer, M. (2011). A study of
315 glitches in the rotation of 102 pulsars. Monthly Notices of the Royal Astronomical
Society, 414:1679–1704.

[60] Faucher-Giguere, C.-A. and Kaspi, V. M. (2006). Birth and evolution of isolated
radio pulsars. The Astrophysical Journal, 643(1):332.

[61] Foreman-Mackey, D. (2016). corner.py: Scatterplot matrices in Python. The Jour-

nal of Open Source Software, 24.

[62] Foreman-Mackey, D., Hogg, D. W., Lang, D., and Goodman, J. (2013). emcee: The
MCMC Hammer. Publications of the Astronomical Society of the Pacific, 125:306–312.

[63] Gavriil, F. P., Kaspi, V. M., and Woods, P. M. (2002). Magnetar-like X-ray bursts
from an anomalous X-ray pulsar. Nature, 419:142–144.

[64] Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin,
D. B. (2013). Bayesian Data Analysis. CRC press.

[65] Gelman, A. et al. (2006). Prior distributions for variance parameters in hierarchical
models (comment on article by Browne and Draper). Bayesian analysis, 1(3):515–534.

[66] Glampedakis, K. and Jones, D. I. (2010). Implications of magnetar non-precession.

Monthly Notices of the Royal Astronomical Society: Letters, 405(1):L6–L10.

[67] Goggans, P. M. and Chi, Y. (2004). Using thermodynamic integration to calcu-

late the posterior probability in bayesian model selection problems. AIP Conference
Proceedings, 707(1):59–66.

[68] Gold, T. (1968). Rotating Neutron Stars as the Origin of the Pulsating Radio
Sources. Nature, 218:731–732.

[69] Goldreich, P. (1970). Neutron star crusts and alignment of magnetic axes in pulsars.
The Astrophysical Journal.
232 BIBLIOGRAPHY

[70] Goldreich, P. and Julian, W. H. (1969). Pulsar electrodynamics. The Astrophysical

Journal.

[71] Goodman, J. and Weare, J. (2012). Ensemble samplers with affine invariance.
Comm. App. Math. Comp. Sci., 5.

[72] Gough, B. (2009). GNU scientific library reference manual. Network Theory Ltd.

[73] Groth, E. J. (1975). Timing of the Crab pulsar III. The slowing down and the nature
of the random process. The Astrophysical Journal Supplement Series, 29:453–465.

[74] Gunn, J. E. and Ostriker, J. P. (1969). Acceleration of high-energy cosmic rays by

pulsars. Physical Review Letters, 22(14):728.

[75] Hawking, S. W. and Israel, W. (1989). Three Hundred Years of Gravitation. Cam-
bridge University Press.

[76] Helfand, D. J., Taylor, J. H., Backus, P. R., and Cordes, J. M. (1980). Pulsar
timing. i-observations from 1970 to 1978. The Astrophysical Journal, 237:206–215.

[77] Hellings, R. W. and Downs, G. S. (1983). Upper limits on the isotropic gravita-
tional radiation background from pulsar timing analysis. The Astrophysical Journal,
265:L39–L42.

[78] Hewish, A., Bell, S. J., Pilkington, J. D. H., Scott, P. F., and Collins, R. A. (1968).
Observation of a Rapidly Pulsating Radio Source. Nature, 217:709–713.

[79] Hobbs, G., Archibald, A., Arzoumanian, Z., Backer, D., Bailes, M., Bhat, N. D. R.,
Burgay, M., Burke-Spolaor, S., Champion, D., Cognard, I., et al. (2010a). The inter-
national pulsar timing array project: using pulsars as a gravitational wave detector.
Classical and Quantum Gravity, 27(8):084013.

[80] Hobbs, G., Edwards, R., and Manchester, R. (2006). TEMPO2: a New Pulsar Tim-
ing Package. Chinese Journal of Astronomy and Astrophysics Supplement, 6(2):189–
192.

[81] Hobbs, G., Lyne, A. G., and Kramer, M. (2010b). An analysis of the timing
irregularities for 366 pulsars. Monthly Notices of the Royal Astronomical Society,
402(2):1027–1048.

[82] Hu, Y.-M., Pitkin, M., Heng, I. S., and Hendry, M. A. (2014). Glitch or anti-glitch:
a Bayesian view. The Astrophysical Journal Letters, 784(2):L41.

[83] Hulse, R. A. and Taylor, J. H. (1975). Discovery of a pulsar in a binary system.

The Astrophysical Journal, 195:L51–L53.

[84] Janssen, G. H. and Stappers, B. W. (2006). 30 glitches in slow pulsars. Astronomy

& Astrophysics, 457:611–618.

[85] Jaranowski, P., Królak, A., and Schutz, B. F. (1998). Data analysis of gravitational-
wave signals from spinning neutron stars: The signal and its detection. Physical
Review D, 58(6):063001.
BIBLIOGRAPHY 233

[86] Jaynes, E. T. (2003). Probability theory – the logic of science. Cambridge university
press.

[87] Jeffreys, H. (1998). The theory of probability. OUP Oxford.

[88] Jones, D. I. (2004). Is timing noise important in the gravitational wave detection
of neutron stars? Physical Review D, 70(4):1–9.

[89] Jones, D. I. (2012). Pulsar state switching, timing noise and free precession. Monthly
Notices of the Royal Astronomical Society, 420(3):2325–2338.

[90] Jones, D. I. and Andersson, N. (2001). Freely precessing neutron stars: model and
observations. Monthly Notices of the Royal Astronomical Society, 324(4):811–824.

[91] Jones, D. I. and Andersson, N. (2002). Gravitational waves from freely precessing
neutron stars. Monthly Notices of the Royal . . . .

[92] Jones, E., Oliphant, T., Peterson, P., et al. (2001). SciPy: Open source scientific
tools for Python. [Online; accessed 2015-11-26].

[93] Jones, P. B. (1988). Excitation of small-amplitude free precession in the Crab

pulsar. Monthly Notices of the Royal Astronomical Society, 235(2):545–550.

[94] Jung, P. (1991). Amplification of small signals via stochastic resonance. Physical
Review A, 44(12).

[95] Kaspi, V. M. (1996). Pulsar/Supernova Remnant Associations. In Johnston, S.,

Walker, M. A., and Bailes, M., editors, IAU Colloq. 160: Pulsars: Problems and
Progress, volume 105 of Astronomical Society of the Pacific Conference Series, page
375.

[96] Kouveliotou, C., Duncan, R. C., and Thompson, C. (2003). Magnetars. Scientific
American, 288(2):34–41.

[97] Kramer, M., Lyne, A. G., O’Brien, J. T., Jordan, C. A., and Lorimer, D. R. (2006a).
A Periodically Active Pulsar Giving Insight into Magnetospheric Physics. Science,
312:549–551.

[98] Kramer, M., Stairs, I. H., Manchester, R. N., McLaughlin, M. A., Lyne, A. G.,
Ferdman, R. D., Burgay, M., Lorimer, D. R., Possenti, A., D’Amico, N., Sarkissian,
J. M., Hobbs, G. B., Reynolds, J. E., Freire, P. C. C., and Camilo, F. (2006b). Tests
of General Relativity from Timing the Double Pulsar. Science, 314:97–102.

[99] Krishnan, B., Sintes, A. M., Papa, M. A., Schutz, B. F., Frasca, S., and Palomba, C.
(2004). Hough transform search for continuous gravitational waves. Physical Review
D, 70(8):082001.

[100] Landau, L. D. and Lifshitz, E. M. (1969). Mechanics, volume 1. Pergamon press,

second edition.

[101] Landau, L. D. and Lifshitz, E. M. (1971). The classical theory of fields, volume 2.
Pergamon press, third revised english edition.
234 BIBLIOGRAPHY

[102] Leaci, P. and Prix, R. (2015). Directed searches for continuous gravitational waves
from binary systems: Parameter-space metrics and optimal Scorpius X-1 sensitivity.
Physical Review D, 91(10):102003.

[103] Lee, K. J., Wex, N., Kramer, M., Stappers, B. W., Bassa, C. G., Janssen, G. H.,
Karuppusamy, R., and Smits, R. (2011). Gravitational wave astronomy of single
sources with a pulsar timing array. Monthly Notices of the Royal Astronomical Society,
414(4):3251–3264.

[104] Lentati, L., Alexander, P., Hobson, M. P., Feroz, F., van Haasteren, R., Lee, K. J.,
and Shannon, R. M. (2014). TEMPONEST: a Bayesian approach to pulsar timing
analysis. Monthly Notices of the Royal Astronomical Society, 437:3004–3023.

[105] Lewin, W. H., van Paradijs, J., and van den Heuvel, E. P. (1997). X-ray Binaries,
volume 26. Cambridge University Press.

[106] LIGO Scientific Collaboration (2014). LALSuite: FreeSoftware (GPL) Tools for
Data-Analysis.

[107] Link, B. (2003). Constraining Hadronic Superfluidity with Neutron Star Preces-
sion. Physical Review Letters, 91(10):101101.

[108] Link, B. and Epstein, R. I. (2001). Precession interpretation of the isolated pulsar
PSR B1828-11. The Astrophysical Journal, 556(1):392.

[109] Lyne, A. (2012). Timing noise and the long-term stability of pulsar profiles. Pro-
ceedings of the International Astronomical Union, 8(S291):183–188.

[110] Lyne, A. and Graham-Smith, F. (2012). Pulsar Astronomy. Cambridge University

Press, fourth edition.

[111] Lyne, A., Hobbs, G., Kramer, M., Stairs, I., and Stappers, B. (2010). Switched
magnetospheric regulation of pulsar spin-down. Science, 329(5990):408–412.

[112] Lyne, A. G. and Manchester, R. N. (1988). The shape of pulsar radio beams.
Monthly Notices of the Royal Astronomical Society, 234:477–508.

[113] Lyne, A. G., Pritchard, R. S., and Graham-Smith, F. (1993). Twenty-Three Years
of Crab Pulsar Rotational History. Monthly Notices of the Royal Astronomical Society,
265:1003.

[114] Lyne, A. G., Stairs, I. H., and Shemar, S. L. (2000). Periodicities in Rotation and
Pulse Shape in PSR B1828-11. In Kramer, M., Wex, N., and Wielebinski, R., editors,
IAU Colloq. 177: Pulsar Astronomy - 2000 and Beyond, volume 202 of Astronomical
Society of the Pacific Conference Series, page 93.

[115] Maciesiak, K., Gil, J., and Ribeiro, V. A. R. M. (2011). On the pulse-width
statistics in radio pulsars - I. Importance of the interpulse emission. Monthly Notices
of the Royal Astronomical Society, 414:1314–1328.

[116] MacKay, D. J. C. (2003). Information theory, inference and learning algorithms.

Cambridge university press.
BIBLIOGRAPHY 235

[117] Manchester, R. N., Hobbs, G. B., Teoh, A., and Hobbs, M. (2005). The Australia
Telescope National Facility Pulsar Catalogue. The Astronomical Journal, 129:1993–
2006.

[118] Manchester, R. N. and Lyne, A. G. (1977). Pulsar interpulses - Two poles or one.
Monthly Notices of the Royal Astronomical Society, 181:761–767.

[119] Matsakis, D. N., Taylor, J. H., and Eubanks, T. M. (1997). A statistic for describ-
ing pulsar and clock stabilities. Astronomy & Astrophysics, 326:924–928.

[120] McKenna, J. and Lyne, A. G. (1990). PSR1737 - 30 and period discontinuities in

young pulsars. Nature, 343:349.

[121] Melatos, A. (1999). Bumpy spin-down of anomalous X-ray pulsars: The link with
magnetars. The Astrophysical Journal Letters, 519(1):L77.

[122] Melatos, A. (2000). Radiative precession of an isolated neutron star. Monthly

Notices of the Royal Astronomical Society, 313(2):217–228.

[123] Melatos, A., Peralta, C., and Wyithe, J. S. B. (2008). Avalanche dynamics of
radio pulsar glitches. The Astrophysical Journal, 672(2):1103.

[124] Messenger, C., Bulten, H. J., Crowder, S. G., et al. (2014). Gravitational waves
from Sco X-1: A comparison of search methods. in preparation. (LIGO DCC-
P1400217).

[125] Michel, F. C. and Goldwire, Jr., H. C. (1970). Alignment of Oblique Rotators. In

Bulletin of the American Astronomical Society, volume 2 of Bulletin of the American
Astronomical Society, page 209.

[126] Mitton, S. (1977). The Cambridge encyclopaedia of astronomy. Jonathan Cape

Ltd.

[127] Narayan, R. and Vivekanand, M. (1983). Evidence for evolving elongated pulsar
beams. Astronomy & Astrophysics, 122:45–53.

[128] Nelson, R. W., Finn, L. S., and Wasserman, I. (1990). Trompe l’oeil’binary’pulsars.
The Astrophysical Journal, 348:226–231.

[129] O’Hagan, A. and Leonard, T. (1976). Bayes estimation subject to uncertainty

about parameter constraints. Biometrika, 63(1):201–203.

[130] Oppenheimer, J. R. and Volkoff, G. M. (1939). On massive neutron cores. Physical

Review, 55(4):374.

[131] Pacini, F. (1967). Energy Emission from a Neutron Star. Nature, 216:567–568.

[132] Pacini, F. (1968). Rotating neutron stars, pulsars and supernova remnants. Nature,
219:145–146.

[133] Pavlov, G. G. and Zavlin, V. E. (2003). Thermal radiation from cooling neutron
stars. In Texas in Tuscany. XXI Texas Symposium on Relativistic Astrophysics, pages
319–328.
236 BIBLIOGRAPHY

[134] Perera, B. B. P., Stappers, B. W., Weltevrede, P., Lyne, A. G., and Bassa, C. G.
(2015). Understanding the spin-down rate changes of PSR B0919+06. Monthly Notices
of the Royal Astronomical Society, 446:1380–1388.

[135] Perera, B. B. P., Stappers, B. W., Weltevrede, P., Lyne, A. G., and Rankin, J. M.
(2016). Correlated spin-down rates and radio emission in PSR B1859+07. Monthly
Notices of the Royal Astronomical Society, 455:1071–1078.

[136] Pines, D. and Shaham, J. (1972). The elastic energy and character of quakes in
solid stars and planets. Physics of the Earth and Planetary Interiors, 6(1):103–115.

[137] Pitkin, M. and Woan, G. (2004). Searching for gravitational waves from the Crab
pulsar-the problem of timing noise. Classical and Quantum Gravity, 21:843.

[138] Popov, S. B., Pons, J. A., Miralles, J. A., Boldin, P. A., and Posselt, B. (2010).
Population synthesis studies of isolated neutron stars with magnetic field decay.
Monthly Notices of the Royal Astronomical Society, 401(4):2675–2686.

[139] Prix, R. (2007). Search for continuous gravitational waves: metric of the multide-
tector F-statistic. Physical Review D, 75(2):023004.

[140] Prix, R. (2009). Gravitational Waves from Spinning Neutron Stars. In Becker,
W., editor, Astrophysics and Space Science Library, volume 357 of Astrophysics and
Space Science Library, page 651. https://dcc.ligo.org/LIGO-P060039/public.

[141] Prix, R. and Itoh, Y. (2005). Global parameter-space correlations of coherent

searches for continuous gravitational waves. Classical and Quantum Gravity, 22:1003.

[142] Prix, R. and Shaltev, M. (2012). Search for continuous gravitational waves: Op-
timal StackSlide method at fixed computing cost. Physical Review D, 85(8):084010.

[143] Radhakrishnan, V. and Cooke, D. J. (1969). Magnetic Poles and the Polarization
Structure of Pulsar Radiation. Astrophysical Letters, 3:225.

[144] Radhakrishnan, V. and Manchester, R. N. (1969). Detection of a Change of State

in the Pulsar PSR 0833-45. Nature, 222:228–229.

[145] Reichley, P. E. and Downs, G. S. (1969). Observed Decrease in the Periods of

Pulsar PSR 0833-45. Nature, 222:229–230.

[146] Richards, D. W., Pettengill, G. H., Counselman, C. C., and Rankin, J. (1969a).
Quasi-sinusoidal components in arrival time of pulsar NP 0532. IAU Circ., 2178.

[147] Richards, D. W., Pettengill, G. H., Roberts, J. A., Counselman, C. C., and Rankin,
J. (1969b). Np 0532. IAU Circ., 2181.

[148] Ruderman, M. (1969). Neutron Starquakes and Pulsar Periods. Nature, 223:597–
598.

[149] Ruderman, M. (1970). Pulsar wobble and neutron starquakes. Nature, 225:838–
839.
BIBLIOGRAPHY 237

[150] Sedrakian, A., Wasserman, I., and Cordes, J. M. (1999). Precession of isolated
neutron stars. I. Effects of imperfect pinning. The Astrophysical Journal, 524(1):341.

[151] Seymour, A. D. and Lorimer, D. R. (2013). Evidence for chaotic behaviour in

pulsar spin-down rates. Monthly Notices of the Royal Astronomical Society, 428:983–
998.

[152] Shaham, J. (1977). Free precession of neutron stars – role of possible vortex
pinning. The Astrophysical Journal, 214:251–260.

[153] Shaltev, M. and Prix, R. (2013). Fully coherent follow-up of continuous

gravitational-wave candidates. Physical Review D, 87(8):084057.

[154] Shapiro, S. L. and Teukolsky, S. A. (1983). Black Holes, White Dwarfs, and
Neutron Stars. John Wiley & Sons, Inc.

[155] Sivia, D. S. and Skilling, J. (1996). Data analysis: a Bayesian tutorial. Oxford
university press.

[156] Stairs, I., Lyne, A., and Shemar, S. (2000). Evidence for free precession in a pulsar.
Nature, 406(6795):484–6.

[157] Stairs, I. H., Athanasiadis, D., Kramer, M., and Lyne, A. G. (2003). High-
Resolution Observations of PSR B1828-11. In Bailes, M., Nice, D. J., and Thorsett,
S. E., editors, Radio Pulsars, volume 302 of Astronomical Society of the Pacific Con-
ference Series, page 249.

[158] Swendsen, R. H. and Wang, J. (1986). Replica Monte Carlo Simulation of Spin-
Glasses. Physical Review Letters, 57:2607–2609.

[159] Taylor, J. H. and Weisberg, J. M. (1982). A new test of general relativity-

Gravitational radiation and the binary pulsar PSR 1913+ 16. The Astrophysical
Journal, 253:908–920.

[160] van den Bergh, S. (1991). Galactic and extragalactic supernova rates. In Super-
novae, pages 711–719. Springer.

[161] Wald, A. (1944). On cumulative sums of random variables. The Annals of Math-
ematical Statistics, 15(3):283–296.

[162] Wang, J., Wang, N., Tong, H., and Yuan, J. (2012). Recent glitches detected in
the Crab pulsar. Astrophysics and Space Science, 340(2):307–315.

[163] Wang, N., Manchester, R. N., Pace, R. T., Bailes, M., Kaspi, V. M., Stappers,
B. W., and Lyne, A. G. (2000). Glitches in southern pulsars. Monthly Notices of the
Royal Astronomical Society, 317:843–860.

[164] Wang, N., Wu, X.-J., Manchester, R., Zhang, J., Lyne, A., and Yusup, A. (2001).
A large glitch in the Crab pulsar. Chinese Journal of Astronomy and Astrophysics,
1(3):195.
238 BIBLIOGRAPHY

[165] Watts, A. L., Krishnan, B., Bildsten, L., and Schutz, B. F. (2008). Detecting
gravitational wave emission from the known accreting neutron stars. Monthly Notices
of the Royal Astronomical Society, 389(2):839–868.

[166] Weiss, N. A., Holmes, P. T., and Hardy, M. (2006). A course in probability. Pearson
Addison Wesley Boston, Massachusetts, USA.

[167] Weisstein, E. W. (2014). Least squares fitting-polynomial. From

MathWorld-A Wolfram Web Resource. http://mathworld.wolfram.com/
LeastSquaresFittingPolynomial.html, last visited on 8/9/2014.

[168] Weisstein, E. W. (2016). Normal difference distribution. From

MathWorld-A Wolfram Web Resource. http://mathworld.wolfram.com/
NormalDifferenceDistribution.html, last visited on 15/04/2016.

[169] Wijnands, R. and Van der Klis, M. (1998). A millisecond pulsar in an X-ray binary
system. nature, 394(6691):344–346.

[170] Wolszczan, A. and Frail, D. A. (1992). A planetary system around the millisecond
pulsar psr 1257+12. Nature, 355(6356):145–147.

[171] Yakovlev, D. G., Haensel, P., Baym, G., and Pethick, C. (2013). Lev Landau and
the concept of neutron stars. Physics Uspekhi, 56:289–295.