paper_cw_mcmc.tex 59.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
\documentclass[aps, prd, twocolumn, superscriptaddress, floatfix, showpacs, nofootinbib, longbibliography]{revtex4-1}

% Fixes Adbobe error (131)
\pdfminorversion=4

% Ams math
\usepackage{amsmath}
\usepackage{amssymb}

% Packages for setting figures
\usepackage[]{graphicx}
\usepackage{subfig}

% Colors
\usepackage{color}

% Bibliography
\usepackage{natbib}

\usepackage{placeins}

\usepackage{multirow}

\usepackage{hhline}

% get hyperlinks
%\usepackage{hyperref}

% Tables
\newcommand{\specialcell}[2][c]{%
  \begin{tabular}[#1]{@{}c@{}}#2\end{tabular}}

\input{definitions.tex}

% For editing purposes: remove before submition
\usepackage[normalem]{ulem}	%% only added for 'strikeout' \sout
\usepackage[usenames,dvipsnames]{xcolor}

\newcommand{\dcc}{LIGO-{\color{red}}}

%% ---------- editing/commenting macros: make sure all a cleared at the end! ----------
\newcommand{\mygreen}{\color{green!50!black}}
\newcommand{\addtext}[1]{\textcolor{green!50!black}{#1}}
\newcommand{\meta}[1]{\addtext{#1}}
\newcommand{\CHECK}[1]{\textcolor{red}{#1}}
\newcommand{\strike}[1]{\textcolor{red}{\sout{#1}}}
\newcommand{\comment}[1]{\textcolor{red}{[#1]}}
\newcommand{\replace}[2]{\strike{#1}\addtext{$\rightarrow$ #2}}
%% ---------- end: editing/commenting macros ----------------------------------------

\begin{document}

53
\title{MCMC follow-up methods for continuous gravitational wave candidates}
54
55
56
57
58
59
60
61
62
63
64
65
66
67

    \author{G. Ashton}
    \email[E-mail: ]{gregory.ashton@ligo.org}
    \affiliation{Max Planck Institut f{\"u}r Gravitationsphysik
                 (Albert Einstein Institut) and Leibniz Universit\"at Hannover,
                 30161 Hannover, Germany}
    \author{R. Prix}
    \affiliation{Max Planck Institut f{\"u}r Gravitationsphysik
                 (Albert Einstein Institut) and Leibniz Universit\"at Hannover,
                 30161 Hannover, Germany}

\date{\today}

\begin{abstract}
68

69
We detail methods to follow-up potential CW signals (as identified by
70
71
72
73
74
75
76
77
78
79
80
wide-parameter space semi-coherent searches) leveraging MCMC optimisation of
the $\mathcal{F}$-statistic. Such a framework provides a unique advantage when
used during the `zoom' (in which the coherence time is increased aiming to
localise the fully-coherent candidate) in that several candidates can be
effeciently followed up simultaneously. We describe an automated method to
define the number of zoom stages and verify such methods work in a Monte-Carlo
study. More, MCMC optimisation naturally produces parameter estimation for the
final fully-coherent candidate. Finally, we show that with minor modifications
the follow-up may allow for the CW waveform to be transient or undergo
glitches; this may allow the discovery of signals which would otherwise go
underdetected.
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100

\end{abstract}

\pacs{04.80.Nn, 97.60.Jd, 04.30.Db}
\input{git_tag.tex}
\date{\commitDATE; \commitIDshort-\commitSTATUS, \dcc}

\maketitle


\section{Introduction}

A possible target for the advanced gravitational wave detector network of LIGO
and Virgo are long-lived periodic sources called continuous waves (CWs).
Rapidly rotating nonaxisymmetric neutron stars are potentially capable of
producing detectable CWs which last much longer than typical observation spans.
There exists three well known sources of the nonaxisymmetry: `mountains',
precession, and r-mode oscillations; each of which make a prediction for the
scaling between $\nu$, the NS spin frequency and $f$, the gravitational wave
frequency. In any case, observing neutron stars through their gravitational
101
wave emission would provide a unique astrophysical insight and has hence
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
motivated numerous searches.

As shown by \citet{jks1998}, the gravitational wave signal from a
nonaxisymmetric source produces a strain in the detector $h(t, \A, \blambda)$;
where $\A{=}\{h_0, \cos\iota, \psi, \phi_0\}$ is a vector of the four
\emph{amplitude-parameters} (expressed in `physical coordinates') and
$\blambda$ is a vector of the \emph{Doppler-parameters} consisting of the
sky-location, frequency $f$, and any spindown terms required by the search.

CW searches typically use a fully-coherent matched-filtering methods whereby a
template (the signal model at some specific set of parameters) is convolved
against the data resulting in a detection statistic. Since the signal
parameters are unknown, it is usual to perform this matched filtering over a
grid of points.  Three search categories can be identified: \emph{targeted}
searches for a signal from a known electromagnetic pulsar where the Doppler
parameters are considered `known'; \emph{directed} searches in which the
location is known, but not the frequency and spin-down (i.e.  searching for the
neutron star in a supernova remnant which does not have a known pulsar); and
\emph{all-sky} searches where none of the parameters are considered known.
Searching over more parameters amounts to an increase in the dimension of the
search space. Since the density of grid points required to resolve a signal
scales inversely with total coherence time $\Tcoh$ (the span of data used in
a fully-coherent matched filter), wide-parameter searches (such as the all-sky)
with many search dimensions over long durations of data are computationally
demanding.

At a fixed computing cost, it has been shown (see for example \citep{brady1998,
prix2012}) that a semi-coherent search is more sensitive for unknown signals
than a fully-coherent search. While the exact details of how the search works
depends on the implementation, semi-coherent search work by splitting the total
observation span $\Tspan$ into $\Nseg$ segments (each lasting for $\Tcoh$) and
in each segment computes the fully-coherent detection statistic; the
semi-coherent detection statistic is then computed by some combination of all
segments summed at the same point in parameter space. Fundamentally, this gain
in sensitivity is because the width of a peak in the detection statistic due to
137
a signal is inversely proportional to the coherence time: shorter coherence times
138
139
140
141
142
143
144
145
146
147
148
make the peak wider and hence the a lower density of templates. This idea was
first proposed by \citet{brady2000} along with the first implementation, the
`Stack-slide' search. Since then, several modifications such as the
`Hough-transform' \citep{krishnan2004, astone2014}, and the `Powerflux' method
(first described in \citet{allyskyS42008}) have been proposed, implemented and
applied to gravitational wave searches.

Wide parameter space searches produce a list of candidates with an associated
detection statistic which passes some threshold. In order to verify these
candidates, they are subjected to a \emph{followed-up}: a process of increasing
the coherence time, eventually aiming to calculate a fully-coherent detection
149
statistic over the maximal span of data. In essence, the semi-coherent search
150
151
152
153
is powerful as it spreads the significance of a candidate over a wider area of
parameter space, so a follow-up attempts to reverse this process and recover
the maximum significance and tightly constrain the candidate parameters. The
original hierarchical follow-up of \citet{brady2000} proposed a two-stage method
154
(an initial semi-coherent  stage followed directly by a fully-coherent search.
155
156
157
158
159
160
161
162
163
164
165
166
167
168
However, it was shown in a numerical study by \citet{cutler2005} that allowing
an arbitrary number of semi-coherent stages before the final fully-coherent
stage can significantly improve the efficiency: ultimately they concluded that
three semi-coherent stages provide the best trade-off between sensitivity and
computational cost.

The first implementation of a two-stage follow-up was given by
\citet{shaltev2013} and used the Mesh Adaptive Direct Search algorithm for
optimisation. At the time of publication, this method was limited to two stages
and could not handle binary parameters \comment{(I think these limitations have
now been removed, but I can't find a publication)}, however these are practical
limitations which can \comment{(have?)} be overcome.
\comment{Add something on multiple modes?}

169
170
In this paper, we propose an alternative hierarchical follow-up procedure using
Markov-Chain Monte-Carlo (MCMC) as the optimisation tool. In terms of the
171
semi-coherent to follow-up procedure, an MCMC tool is advantages due to it's
172
173
ability to trace the evolution multiple modes simultaneously through the
follow-up procedure and allow the optimisation to decide between them without
174
arbitrary cuts. In addition, MCMC methods also provide two further
175
176
177
advantages: they can calculate directly calculate Bayes factors, the significance
of a candidate and because they are `grid-less' one can arbitrarily vary the
waveform model without requiring an understanding of the underlying topology.
178
We will exploit this latter property to propose an additional step in the
179
follow-up procedure which allows for the CW signal to be either a transient-CW
180
181
182
183
184
185
(a periodic signal lasting $\mathcal{O}(\textrm{hours-weeks})$) or to undergo
glitches (as seen in pulsars).

We begin in Section~\ref{sec_hypothesis_testing} with a review of search
methods from a Bayesian perspective. Then in
Section~\ref{sec_MCMC_and_the_F_statistic} we introduce the MCMC optimisation
186
procedure and give details of our particular implementation. In
187
188
189
190
191
192
193
194
195
196
197
198
199
Section~\ref{sec_follow_up} we will illustrate applications of the method and
provide a prescription for choosing the setup. In Sections~\ref{sec_transients}
and \ref{sec_glitches} we demonstrate how searches can be performed for either
transient or glitches CWs before we finally conclude in
Section~\ref{sec_conclusion}.

\section{Hypothesis testing}
\label{sec_hypothesis_testing}

\subsection{Bayes factors}
Given some data $x$ and a set of background assumptions $I$, we formulate
two hypotheses: $\Hn$, the data contains solely Gaussian noise and $\Hs$, the
data contains an additive mixture of noise and a signal $h(t; \A, \blambda)$.
200
In order to make a quantitative comparison, we use Bayes theorem in the usual
201
202
way to write the odds as
\begin{equation}
203
O_{\rm S/N} \equiv \frac{P(\Hs| x, I)}{P(\Hn| x, I)} =
204
205
206
207
208
\Bsn(x| I) \frac{P(\Hs| I)}{P(\Hn | I)},
\end{equation}
where the second factor is the prior odds while the first factor is the
\emph{Bayes factor}:
\begin{equation}
209
\Bsn(x| I) = \frac{P(x| \Hs, I)}{P(x| \Hn, I)}.
210
211
212
\end{equation}
Typically, we set the prior odds to unity such that it is the Bayes factor
which determines our confidence in the signal hypothesis. In this work we will
213
therefore discuss the Bayes factor with the implied assumption this is
214
215
equivalent to the odds, unless we have a good reason to change the prior odds.

216
We can rewrite the Bayes factor in terms of the two sets of signal parameters
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
as
\begin{equation}
\Bsn(x| I) = \frac{P(x, \A, \blambda|\Hs, I)}
{P(\A| \Hs, I)P(\blambda| \Hs, I)P(x| \Hn, I)}.
\end{equation}
Marginalising over the two sets of parameters we find that
\begin{equation}
\Bsn(x| I)= \iint
\mathcal{L}(x; \A, \blambda)
P(\A| \Hs, I)P(\blambda| \Hs, I)
d\blambda d\A
\label{eqn_full_bayes}
\end{equation}
where
\begin{equation}
\mathcal{L}(x; \A, \blambda) \equiv \frac{P(x |\A, \blambda, \Hs, I)}{P(x| \Hn, I)},
\label{eqn_likelihood}
\end{equation}
is the \emph{likelihood-ratio}.

At this point, we can appreciate the problems of searching for unknown signals:
238
one has four amplitude parameters and several Doppler parameters (three plus
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
the number of spin-down and binary parameters) over which this integral must be
performed. If a single signal exists in the data, this corresponds to a single
peak in the likelihood-ratio, but at an unknown location. Therefore, one must
must first search for peaks (candidates), and then subsequently analyse their
significance. If one has prior knowledge this can be used; for example in
targeted searches for gravitational waves from known pulsars emitting CWs from
a mountain the Doppler parameters are considered known collapsing their
integral to a single evaluation.

\subsection{The $\F$-statistic}

For directed and all-sky searches, a common method introduced by
\citet{jks1998} to reduce the parameter space is the maximum-likelihood
approach. In this approach (often referred to as `frequentist'), one starts by
defining the likelihood-ratio, Equation~\eqref{eqn_likelihood}, which in this
context is a \emph{matched-filtering} amplitude. Then, analytically maximising
this likelihood-ratio with respect to the four amplitude parameters results
(c.f.~\citet{prix2009}) in a maximised log-likelihood given by $\F(x|
\blambda)$: the so-called $\F$-statistic. Picking a particular set of Doppler
parameters $\blambda$ (the template) one can then compute a detection statistic
259
(typically $2\F$ is used) which can be used to quantify the significance of the
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
template. Usually this is done by calculating a corresponding false alarm rate,
the probability of seeing such a detection statistic in Gaussian noise.

Calculations of the significance are possible due to the properties of the
detection statistic: in Gaussian noise, it can be shown \citep{jks1998,
cutlershutz2005} that $\widetilde{2\F}$ follows a chi-squared distribution with
4 degrees of freedom. In the presence of a signal, $\widetilde{2\F}$ is
still chi-squared with 4 degrees of freedom, but has a non-centrality parameter
$\tilde{\rho}^{2}$ such that its expectation value is
\begin{equation}
\textrm{E}[\widetilde{2\F}(x; \blambda)] = 4 + \tilde{\rho}(x; \blambda)^2.
\label{eqn_twoF_expectation}
\end{equation}
The non-centrality parameter in this context is the SNR of the matched-filter
given by
\begin{equation}
\rho^{2} = (h | h) \propto \frac{h_0^2}{\Sn}\Tcoh \mathcal{N}
\end{equation}
where $(h|h)$ is the inner product of the signal with itself (see for example
\citet{prix2009}), $\Sn$ is a (frequency-dependent) measure of the noise in
the detector and $\mathcal{N}$ is the number of detectors.


283
284
285
286
287
\subsection{Using the $\F$-statistic to compute a Bayes factor} At first, it
appears that the $\F$-statistic is independent of the Bayesian framework since
it was first derived directly from the likelihood. However, it was shown by
\citet{prix2009} that if we marginalise over the four amplitude parameters of
Equation~\eqref{eqn_full_bayes}, choosing a prior $\Pic$ such that
288
\begin{equation}
289
P(\A| \Hs, \Pic, I) \equiv \left\{
290
291
292
293
294
295
296
297
298
\begin{array}{ll}
C & \textrm{ for } \ho < \homax \\
0 & \textrm{ otherwise}
\end{array}
\right..
\end{equation}
then the integral, when $\homax \gg 1$, is a Gaussian integral and can be
computed analytically as
\begin{align}
299
\Bsn(x| \Pic, \blambda, I) & \equiv
300
301
302
\int
\mathcal{L}(x ;\A, \blambda)
P(\A| \Hs, I) d\A
303
\label{eqn_B_S_N}
304
305
306
307
308
309
310
311
312
313
314
315
316
\\
& = \frac{C (2\pi)^{2} e^{\F(x| \blambda)}}
{\sqrt{\textrm{det} \mathcal{M}}},
\end{align}
where $C$ is a normalisation constant, $\textrm{det}\mathcal{M}$ is an antenna
pattern factor dependent on the sky-position and observation period, and
$\mathcal{F}$ is the frequentist log-likelihood of \citet{jks1998}. This result
demonstrates that the $\mathcal{F}$-statistic is proportional to the log-Bayes
factors when calculated with a uniform prior on the amplitude parameters and
fixed Doppler parameters.

As such, we can define the Bayes-factor of Equation~\eqref{eqn_full_bayes} as
\begin{equation}
317
318
\Bsn(x| \Pic, I) = \int
\Bsn(x| \Pic, \blambda, I) P(\blambda| \Hs, I)
319
320
321
322
 d\blambda,
\end{equation}
or neglecting the constants
\begin{equation}
323
\Bsn(x| \Pic, I) \propto \int
324
325
326
327
328
329
330
331
e^{\F(x| \blambda)} P(\blambda| \Hs, I)
 d\blambda.
\label{eqn_bayes_over_F}
\end{equation}

Formulating the significance of a CW candidate in this way is pragmatic in that
there exists a wealth of well-tested tools \citep{lalsuite} capable of
computing the $\mathcal{F}$-statistic for CW signals, transient-CWs, and CW
332
signals from binary systems; these can be leveraged to compute
333
Equation~\eqref{eqn_bayes_over_F}, or adding in the constant
334
$\Bsn(x| \Pic)$ itself. The disadvantage to this method is that
335
336
337
338
339
340
341
we are forced to use the prior $\Pic$, which was shown by \citet{prix2009} to
be unphysical.

\section{MCMC and the $\mathcal{F}$-statistic}
\label{sec_MCMC_and_the_F_statistic}

The MCMC class of optimisation tools are formulated to solve the problem of
342
inferring the posterior distribution of some general model parameters $\theta$
343
344
345
346
347
348
given given some data $x$ for some hypothesis $\H$. Namely, Bayes rule
\begin{equation}
P(\theta| x, \H, I) \propto P(x| \theta, \H, I)P(\theta| \H, I),
\label{eqn_bayes_for_theta}
\end{equation}
is used to evaluate proposed jumps from one point in parameter to other points;
349
jumps which increase this probably are accepted with some probability. The
350
351
algorithm, proceeding in this way, is highly efficient at resolving peaks in
high-dimension parameter spaces.
352
353
354
355

At this point, we note the equivalence of Equation~\eqref{eqn_bayes_for_theta}
to the integrand of Equation~\eqref{eqn_bayes_over_F}:
\begin{equation}
356
357
P(\blambda | x, \Pic, \Hs, I)
%=\Bsn(x| \Pic, \blambda) P(\blambda| \Hs I).
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
\propto e^{\F(x| \blambda)} P(\blambda| \Hs I),
\label{eqn_lambda_posterior}
\end{equation}
where $e^{\F}$ is the likelihood.
In this work, we will focus on using MCMC methods to sample this, the
posterior distribution of the Doppler parameters and moreover compute the
final Bayes factor.

\subsection{The \texttt{emcee} sampler}

In this work we will use the \texttt{emcee} ensemble sampler
\citep{foreman-mackay2013}, an implementation of the affine-invariant ensemble
sampler proposed by \citet{goodman2010}. This choice addresses a key issue with
the use of MCMC sampler, namely the choice of \emph{proposal distribution}. At
each step of the MCMC algorithm, the sampler generates from some distribution
373
374
(known as the proposal-distribution) a jump in parameter space. Usually, this
proposal distribution must be `tuned' so that the MCMC sampler efficiently
375
376
377
walks the parameter space without either jumping too far off the peak, or
taking such small steps that it takes a long period of time to traverse the
peak. The \texttt{emcee} sampler addresses this by using an ensemble, a large
378
379
380
381
382
383
384
385
386
387
number ${\sim}100$ parallel \emph{walkers}, in which the proposal for each
walker is generated from the current distribution of the other walkers.
Moreover, by applying an an affine transformation, the efficiency of the
algorithm is not diminished when the parameter space is highly anisotropic. As
such, this sampler requires little in the way of tuning: a single proposal
scale and the number of steps to take.

\subsection{Parallel tempering: sampling multimodal posteriors}
Beyond the standard ensemble sampler, we will also use one further
modification, the parallel-tempered ensemble sampler. A parallel
388
389
tempered MCMC simulation, first proposed by \citet{swendsen1986}, runs
$\Ntemps$ simulations in parallel with the likelihood in the $i^{\rm th}$
390
391
392
parallel simulation is raised to a power of $1/T_{i}$ where $T_i$ is referred
to as the temperature. As such, Equation~\eqref{eqn_lambda_posterior} is
written as
393
\begin{equation}
394
395
P(\blambda | T_i, x, \Pic, \Hs, I)
%=\Bsn(x| \Pic, \blambda)^{T_i} P(\blambda| \Hs I).
396
397
\propto (e^{\F(x| \blambda)})^{T_i} P(\blambda| \Hs I).
\end{equation}
398
Setting $T_0=1$ with $T_i > T_0 \; \forall \; i > 1$, such that the lowest
399
400
temperature recovers Equation~\eqref{eqn_lambda_posterior} while for higher
temperatures the likelihood is broadened (for a Gaussian likelihood, the
401
402
standard deviation is larger by a factor of $\sqrt{T_i}$). Periodically, the
algorithm swaps the position of the walkers between the different
403
404
405
temperatures. This allows the $T_0$ chain (from which we draw samples of the
posterior) to efficiently sample from multi-modal posteriors. This introduces
two additional tuning parameters, the number and range of the set of
406
temperatures $\{T_i\}$, we will discuss their significance when relevant.
407

408
\subsection{Parallel tempering: estimating the Bayes factor}
409
410
In addition, parallel-tempering also offers a robust method to estimate the
Bayes factor of Equation~\eqref{eqn_bayes_over_F}. If we define
411
$\beta\equiv1/T$, the inverse temperature and $Z(\beta)\equiv \Bsn(x| \Pi_{\rm
412
413
414
415
416
c}, I)$, then as noted by \citet{goggans2004} for the general case, we may
write
\begin{align}
\frac{1}{Z} \frac{\partial Z}{\partial \beta}=
\frac{
417
418
419
\int \Bsn(x| \Pic, \blambda)^{\beta}
\log(\Bsn(x| \Pic, \blambda))P(\blambda| I)
d\blambda
420
421
}
{
422
423
\int \Bsn(x| \Pic, \blambda)^{\beta})P(\blambda| I)
d\blambda
424
425
426
427
428
}
\end{align}
The right-hand-side expresses the average of the log-likelihood at $\beta$. As
such, we have
\begin{align}
Gregory Ashton's avatar
Gregory Ashton committed
429
\frac{\partial \log Z}{\partial \beta} =
430
431
432
433
434
435
436
437
\langle \log(\Bsn(x| \Pic, \blambda) \rangle_{\beta}
\end{align}
The log-likelihood are a calculated during the MCMC sampling. As such, one
can numerically integrate to get the Bayes factor, i.e.
\begin{align}
\log \Bsn(x| \Pic, I) = \log Z = \int_{0}^{1}
\langle \log(\Bsn(x| \Pic, \blambda) \rangle_{\beta} d\beta.
\end{align}
438
In practice, we use a simple numerical quadrature over a finite ladder of
439
$\beta_i$ with the smallest chosen such that choosing a smaller value does not
440
441
change the result beyond other numerical uncertainties. Typically, getting
accurate results for the Bayes factor requires a substantially larger number of
442
443
temperatures than are required for efficiently sampling multi-modal
distributions.  Therefore, it is recommended that one uses a small number of
444
445
446
447
temperatures during the search stage, and subsequently a larger number of
temperatures (suitably initialised close to the target peak) when estimating
the Bayes factor.

Gregory Ashton's avatar
Gregory Ashton committed
448
\subsection{Evaluating the search space}
449
450
451
452

In general, MCMC samplers are highly effective in generating samples of the
posterior in multi-dimensional parameter spaces. However, they will perform
poorly if the posterior has multiple small maxima with only a small number of
Gregory Ashton's avatar
Gregory Ashton committed
453
454
455
456
457
458
459
460
461
462
463
464
large maxima which occupy a small fraction of the prior volume. Later on, in
Section~\ref{sec_topology} we will show that for a signal in Gaussian noise our
log-likelihood $\F$ will have a single peak due to the signal which is
surrounded by other peaks particular to given noise realisation. The relative
size of the these peaks to the signal peak depends on the SNR.  Howeever,
regardless of the strengt of the signal peak, if the width of the signal peak
is small compared to the prior volume, the sampler will get `stuck' on the
local maxima and be inefficient at finding the global maxima.  This problem is
exacerbated in higher-dimensional search spaces where the volume fraction of
the signal scales with the exponent of the number of dimensions.

\subsubsection{The metric}
465
466
467

In a traditional CW search which uses a grid of templates (also known as a
template bank), the spacings of the grid are chosen such that the loss of
Gregory Ashton's avatar
Gregory Ashton committed
468
469
470
471
472
473
474
475
signal to noise ratio (SNR) is bounded to less than $\mu$, the
\emph{template-bank mismatch}. The optimal choice of grid then consists of
minimising the computing cost while respecting this minimum template-bank
mismatch or vice-verse (for examples see \citet{pletsch2010, prix2012,
wette2013, wette2015}). We will now discuss how the work on setting up these
grids can be applied to the problem of determining whether the setup is
appropriate for an MCMC method: i.e. given the prior volume do we expect a
signal to occupy a non-negligible volume?
476
477
478

For a fully-coherent $\F$-statistic search on data containing Gaussian noise
and a signal with Doppler parameters $\blambdaSignal$, the template-bank
Gregory Ashton's avatar
Gregory Ashton committed
479
mismatch at the grid point $\blambda_{l}$ we define as
480
\begin{align}
Gregory Ashton's avatar
Gregory Ashton committed
481
\mutilde(\blambdaSignal, \blambda_{l}) \equiv 1 -
482
483
484
485
486
\frac{\tilde{\rho}(\blambda_l;\blambdaSignal)^{2}}
{\tilde{\rho}(\blambdaSignal; \blambdaSignal)^{2}},
\end{align}
where $\tilde{\rho}(\blambda_l; \blambdaSignal)$ is the non-centrality
parameter (c.f. Equation~\ref{eqn_twoF_expectation}) at $\blambda_l$, given
Gregory Ashton's avatar
Gregory Ashton committed
487
488
489
490
491
492
493
that the signal is at $\blambdaSignal$; as such $\tilde{\rho}(\blambdaSignal;
\blambdaSignal)$ is the perfectly-matched non-centrality parameter.  For a
fully-coherent search, this non-centrality parameter is equivalent to
fully-coherent matched-filter signal to noise ratio (SNR). However, as noted by
\citet{leaci2015}, this is true for the fully-coherent case only.  Therefore,
we choose to use the non-centrality parameter which easily generalises to the
semi-coherent case.
494
495
496
497

To make analytic calculations of the mismatch possible, as first shown by
\citet{brady1998}, the mismatch can be approximated by
\begin{equation}
Gregory Ashton's avatar
Gregory Ashton committed
498
499
\mutilde(\blambda, \Delta\blambda) \approx
\tilde{g}_{ij}\Delta\lambda^{i}\Delta\lambda^{j}
500
+ \mathcal{O}\left(\Delta\blambda^{3}\right)
Gregory Ashton's avatar
Gregory Ashton committed
501
\label{eqn_mutilde_expansion}
502
\end{equation}
Gregory Ashton's avatar
Gregory Ashton committed
503
504
where we use index notation with an implicit summation over repeated indices.
Here, $\tilde{g}_{ij}^{\phi}$ is the \emph{phase-metric} given by
505
\begin{align}
Gregory Ashton's avatar
Gregory Ashton committed
506
507
508
509
\tilde{g}_{ij}(\blambda) =
\langle
\partial_{\Delta\lambda^{i}}\phi
\partial_{\Delta\lambda^{j}}\phi
510
511
\rangle
-
Gregory Ashton's avatar
Gregory Ashton committed
512
513
\langle
\partial_{\Delta\lambda^{i}}\phi
514
515
\rangle
\langle
Gregory Ashton's avatar
Gregory Ashton committed
516
\partial_{\Delta\lambda^{j}}\phi
517
518
519
520
521
522
\rangle,
\label{eqn_metric}
\end{align}
where $\langle \cdot \rangle$ denotes the time-average over $\Tcoh$ and
$\phi(t; \blambda)$ is the phase evolution of the source. The phase metric is
in fact an approximation of the full metric which includes modulations of the
Gregory Ashton's avatar
Gregory Ashton committed
523
524
525
amplitude parameters $\A$. It was shown by \citet{prix2007metric} that
when using data spans longer than a day and data from multiple detectors, the
difference between phase metric and full metric is negligible.
526

527
The phase metric, Equation~\eqref{eqn_metric} provides the necessary tool to
528
measure distances in the Doppler parameter space in units of mismatch. To
Gregory Ashton's avatar
Gregory Ashton committed
529
calculate components, we define the phase evolution
530
531
532
533
of the source as \citep{wette2015}
\begin{align}
\phi(t; \blambda) \approx 2\pi\left(
\sum_{s=0}^{\smax} f^{(s)}\frac{(t-\tref)^{s+1}}{(s+1)!}
Gregory Ashton's avatar
Gregory Ashton committed
534
+ \frac{r(t)\cdot\mathbf{n}(\alpha, \delta)}{c} \fmax\right),
535
536
537
\label{eqn_phi}
\end{align}
where $\mathbf{n}(\alpha, \delta)$ is the fixed position of the source with
Gregory Ashton's avatar
Gregory Ashton committed
538
539
540
541
respect to the solar system barycenter (with coordinates $\alpha$ and $\delta$,
the right ascension and declination), $f^{(s)}\equiv d^{s}\phi/dt^s$, and
$\fmax$ a constant approximating $f(t)$, chosen conservatively to be the
maximum frequency over the data span \citep{wette2013}.
542
543
544

The frequency and spin-down components of the metric can be easily calculated
due to their linearity in Equation~\eqref{eqn_phi} and for the special case in
Gregory Ashton's avatar
Gregory Ashton committed
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
which $\tref$ is in the middle of the data span, the phase-evolution metric
with $\smax{=}1$ (i.e. the frequency and spin-down components) are given by
\begin{equation}
\tilde{g}^{\textrm{PE}}_{ij} \equiv
\left[
\begin{array}{cc}
\frac{\pi^{2}\Tcoh^{2}}{3} & 0 \\
0 & \frac{4\pi^{2}\Tcoh^{4}}{45}
\end{array}
\right]
\textrm{ for } i, j \in [f, \dot{f}].
\end{equation}.

Accurately approximating the sky components of the metric is non-trivial, but
was accomplished by \citet{wette2013} for the fully-coherent case. In
\citet{wette2015} it was shown how to calculate the equivalent semi-coherent
metric $\hat{g}_{ij}(\blambda, \Nseg)$; in the following, we
will work with this generalised semi-coherent method with the understanding
that $\hat{g}_{ij}(\blambda, \Nseg{=}1)=
\tilde{g}_{ij}(\blambda)$.

\subsubsection{The metric volume}
567
To understand the volume of parameter space which a true signal would occupy,
Gregory Ashton's avatar
Gregory Ashton committed
568
we can make use of the \emph{metric-volume}, defined as \citep{prix2007}
569
\begin{align}
Gregory Ashton's avatar
Gregory Ashton committed
570
571
572
\mathcal{V} = \int
\sqrt{\textrm{det}\hat{g}_{ij}(\blambda, \Nseg)} d\blambda \approx
\sqrt{\textrm{det}\hat{g}_{ij}(\blambda, \Nseg)} \Delta\blambda
573
574
\end{align}
where in the second step we assume a constant coefficient metric. Here, $\Delta
Gregory Ashton's avatar
Gregory Ashton committed
575
\blambda$ is the volume element which is given by
576
\begin{equation}
Gregory Ashton's avatar
Gregory Ashton committed
577
\Delta\lambda = \frac{\cos(\delta_0)\Delta\delta\Delta\alpha}{2}
578
579
580
%\frac{1}{2}\sin\delta\Delta\delta\Delta\alpha
\prod_{s=0}^{\smax} \Delta f^{(s)},
\end{equation}
Gregory Ashton's avatar
Gregory Ashton committed
581
582
583
584
585
where the numerator of the prefactor is the solid angle of the sky-patch at a
declination of $\delta_0$, the factor of $1/2$ comes from converting the
normalised determinant (which is computed over the whole sky) to the solid angle
of the directed search, and $\Delta f^{(s)}$ is the extent of the frequency and
spin-down range(s) searched.
586
587
588
589
590

The metric volume $\V$ is the approximate number of templates required to cover
the the given Doppler parameter volume at a fixed mismatch of $\approx 1$. As
such, its inverse gives the approximate (order of magnitude) volume fraction of
the search volume which would be occupied by a signal. This can therefore be
591
592
used as a proxy for determining if an MCMC search will operate in a regime where
it is efficient (i.e. where the a signal occupies a reasonable fraction of the
593
594
595
596
597
598
599
600
601
search volume).

The volume $\V$ combines the search volume from all search dimensions. However,
let us know discuss how we can delve deeper to understand how each dimension
contributes to the total product. This is done by noticing that the metric has
a particular block form:
\begin{align}
g_{ij} = \left[
\begin{array}{cc}
Gregory Ashton's avatar
Gregory Ashton committed
602
603
g^{\textrm{Sky}} & 0 \\
0 & g^{\textrm{PE}}.
604
605
606
607
608
609
610
611
612
613
614
\end{array}
\right]
\end{align}
where $g^{\rm Sky}$ is the $2\times2$ sky-metric, while $g^{\rm PE}$ is the
$(\smax{+}1){\times}(\smax{+}1)$ phase-evolution metric.
As such, the volume can be decomposed as
\begin{align}
\mathcal{V} & =
\sqrt{\textrm{det}g^{\rm Sky}}\frac{\Delta\Omega}{2} \times
\sqrt{\textrm{det}g^{\rm PE}}\prod_{s=0}^{\smax}\Delta f^{(s)} \\
& = \Vsky \times \Vpe.
615
\label{eqn_metric_volume}
616
617
618
619
620
621
622
623
624
\end{align}
Moreover, if $\tref$ is in the middle of the observation span, the diagonal
nature of $g^{\rm PE}$ means that one can further identify
\begin{align}
\Vpe = \prod_{s=0}^{\smax}\sqrt{g^{\rm PE}_{ss}} \Delta f^{(s)}
= \prod_{s=0}^{\smax}\Vpe^{(s)}
\end{align}
This decomposition may be useful in setting up MCMC searches.

Gregory Ashton's avatar
Gregory Ashton committed
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
\subsection{The topology of the likelihood}
\label{sec_topology}

We intend to use the $\F$-statistic as our log-likelihood in MCMC simulations,
but before continuing, it is worthwhile to acquaint ourselves with the typical
behavior of the log-likelihood by considering a specific example.

As shown in Equation~\eqref{eqn_twoF_expectation}, the expectation of
$\widetilde{2\F}$ is 4 in Gaussian noise alone, but proportional to the square
of the non-centrality parameter (or SNR) in the presence of a signal. To
illustrate this, let us consider $\widetilde{2\F}$ as a function of $f$ (the
template frequency) if there exists a signal in the data with frequency $f_0$.

We will fix all the other Doppler parameters so that they are perfectly
matched.  Such an example can be calculated analytically, taking the
matched-filtering amplitude (Equation~(11) of \citep{prix2005}) with
$\Delta\Phi(t) = 2\pi(f - f_0) t$, the expectation of $\widetilde{2\F}$ as a
function of the template frequency $f$ is given by
\begin{equation}
\textrm{E}[\widetilde{2\F}](f) = 4 +
(\textrm{E}[\widetilde{2\F_0}] -4)\textrm{sinc}^{2}(\pi(f-f_0)\Tcoh))
\label{eqn_grid_prediction}
\end{equation}
where $\textrm{E}[\widetilde{2\F_0}]$ is the expected $\widetilde{2\F}$ for
a perfectly matched signal (when $f=f_0$).

In Figure~\ref{fig_grid_frequency} we compare the analytic prediction of
Equation~\eqref{eqn_grid_prediction} with the value computed numerically from
simulating a signal in Gaussian noise. This is done over a frequency range of
$\sim20\Delta f$ where $\Delta f$ is the frequency difference between a signal
and the template with $\tilde{\mu}=1$ as calculated from
Equation~\eqref{eqn_mutilde_expansion}. As expected, close to the signal
frequency $f_0$ the detection statistic peaks with a a few local secondary
maxima. Away from this frequency, in Gaussian noise, we see many local maxima
centered around the expected value of 4.
\begin{figure}[htb]
\centering \includegraphics[width=0.45\textwidth]{grided_frequency_search_1D}
\caption{Comparison of the analytic prediction of
Equation~\eqref{eqn_grid_prediction} (in red) with the value computed
numerically from simulating a signal with frequency $f_0$ in Gaussian noise (in
black). $\Delta f$ is the frequency difference corresponding to a
metric-mismatch of unity.}
\label{fig_grid_frequency}
\end{figure}

670
671
672
673
674
675
676
677
678
679
\subsection{Example: signal in noise}

In order to familiarise the reader with the features of an MCMC search, we will
now describe a simple directed search (over $f$ and $\dot{f}$) for a simulated
signal in Gaussian noise. The signal will have a frequency of $30$~Hz and a
spin-down of $-1{\times}10^{-10}$~Hz/s, all other Doppler parameters are
`known' and so are irrelevant. Moreover, the signal has an amplitude
$h_0=10^{-24}$~Hz$^{-1/2}$ while the Gaussian noise has
$\Sn=10^{-23}$~Hz$^{-1/2}$ such that the signal has a depth of 10.

680
First, we must define a prior for each search parameter Typically, we recommend
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
either a uniform prior bounding the area of interest, or a normal distribution
centered on the target and with some well defined width. However, to ensure
that the MCMC simulation has a reasonable chance at finding a peak, one should
consider the corresponding metric-volume given in
Equation~\eqref{eqn_metric_volume}. For this example, we will use a uniform
prior with a frequency range of $\Delta f = 10^{-7}$~Hz and a spin-down range
of $\Delta \fdot=10^{-13}$~Hz/s both centered on the simulated signal frequency
and spin-down rate. We set the reference time to coincide with the middle of
the data span, therefore the metric volume can be decomposed into the frequency
contribution and spin-down contribution:
frequency,
\begin{align}
\Vpe^{(0)} = \frac{(\pi\Tcoh\Delta f)^2}{3} \approx 2.46
\end{align}
and
\begin{align}
\Vpe^{(1)} = \frac{4(\pi\Delta \fdot)^2\Tcoh^{4}}{45} \approx 48.9
\end{align}
such that $\V\approx120$ (note that $\Vsky$ does not contribute since we do
not search over the sky parameters). This metric volume indicates that the
signal will occupy about 1\% of the prior volume, therefore the MCMC is
expected to work. Alternative priors will need careful thought about how to
703
translate them into a metric volume: for example using a Gaussian one could use
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
the standard deviation as a proxy for the allowed search region.

In addition to defining the prior, one must also consider how to
\emph{initialise} the walkers. If the prior genuinely represents the stated
prior knowledge, the usual solution is to initialise the walkers from the
prior: that is the starting position is drawn from the prior. However,
instances do occur when one would like to initialise the walkers from a
different distribution. For example, if one only needs to estimate the evidence
(given a particular prior), but is aware from previous searches that the only
significant peak lies in a small area of parameter space, one could initialise
the walkers in a small cluster close to that area. In this example, we
initialise the walkers from the prior such that they have the chance to explore
the entire prior volume.

Having defined the prior, the final setup step is to define the number of
\emph{burn-in} and \emph{production} steps the sampler should take and the
number of walkers; this is a tuning parameter of the MCMC algorithm. The number
of walkers should be typically a few hundred, the greater the number the more
samples will be taken resulting in improved posterior estimates. The burn-in
steps refers to an initial set of steps which are discarded as they are taken
724
whilst the walkers converge. After they have converged the steps are known as
725
726
727
728
729
730
731
732
733
production steps since they are used to produce posterior estimates and
calculate the marginal likelihood.

Using these choices, the simulation is run. To illustrate the full MCMC
process, in Figure~\ref{fig_MCMC_simple_example} we plot the progress of all
the individual walkers (each represented by an individual line) as a function
of the total number of steps. The red portion of steps are burn-in and hence
discarded, from this plot we see why: the walkers are initialised from the
uniform prior and initially spend some time exploring the whole parameter space
734
before converging. The fact that they converge to a single unique point is due
735
to the strength of the signal (substantially elevating the likelihood about
736
that of Gaussian fluctuations) and the tight prior which was quantified through the
737
738
metric volume $\V$. The production samples, colored black, are only taken once
the sampler has converged - these can be used to generate posterior plots.
739
740
\begin{figure}[htb]
\centering
Gregory Ashton's avatar
Gregory Ashton committed
741
\includegraphics[width=0.45\textwidth]{fully_coherent_search_using_MCMC_walkers}
742
743
744
745
746
747
748
749
\caption{The progress of the MCMC simulation for a simulated signal in Gaussian
noise, searching over the frequency and spin-down. The upper two panels show
the position of all walkers as a function of the number of steps for the
frequency and spin-down; when they are colored red the samples are discarded as
burn-in (the first 100 steps), while when they are colored black they are used
as production samples. The bottom panel shows the distribution of
$\widetilde{2\F}$ taken from the production samples.}
\label{fig_MCMC_simple_example}
750
751
752
753
754
755
756
757
\end{figure}

\section{Follow-up}
\label{sec_follow_up}

Incoherent detection statistics trade significance (the height of the peak) for
sensitivity (the width of the peak). We will now discuss the advantages of
using an MCMC sampler to follow-up a candidate found incoherently, increasing
758
759
the coherence time until finally estimating its parameters and significance
fully-coherently. We begin by rewriting Equation~\eqref{eqn_lambda_posterior},
760
761
762
the posterior distribution of the Doppler parameters, with the explicit
dependence on the coherence time $\Tcoh$:
\begin{equation}
763
764
P(\blambda | \Tcoh, x, \Pic, \Hs, I)
%=\Bsn(x| \Tcoh, \Pic, \blambda) P(\blambda| \Hs I).
765
766
767
\propto e^{\hat{\F}(x| \Tcoh, \blambda)} P(\blambda| \Hs I).
\end{equation}

768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
Introducing the coherent time $\Tcoh$ as a variable provides a free parameter
which adjusts width of signal peaks in the likelihood. Therefore, a natural way
to perform a follow-up is to start the MCMC simulations with a short coherence
time (such that the signal peak occupies a substantial fraction of the prior
volume) and then subsequently incrementally increasing this coherence time in a
controlled manner, aiming to allow the MCMC walkers to converge to the new
likelihood before again increasing the coherence time. Ultimately, this
coherence time will be increased until $\Tcoh = \Tspan$. If this is done in
$\Nstages$ discreet \emph{stages}, this introduces a further set of tuning
parameters, namely the ladder of coherence times $\Tcoh^{i}$, where $i \in [0,
\Nstages]$.

In some ways, this bears a resemblance to `simulated annealing', a method in
which the likelihood is raised to a power (the inverse temperature) and
subsequently `cooled'. The important difference being that the semi-coherent
likelihood is wider at short coherence times, rather than flatter as in the
case of high-temperature simulated annealing stages. For a discussion and
examples of using simulated annealing in the context of CW searches see
\citet{veitch2007}.

Of course in practice, we can't arbitrarily vary $\Tcoh^i$, but rather the
number of segments at each stage $\Nseg^{i}\equiv \Tspan/\Tcoh^{i} \in
\mathbb{N}$.  Ideally, the ladder of segment should be chosen to ensure that
the metric volume at the $i^{th}$ stage $\V_i \equiv \V(\Nseg^i)$ is a constant
792
793
794
795
796
fraction of the volume at adjacent stages. That is we define
\begin{equation}
\mathcal{R} \equiv \frac{\V_i}{\V_{i+1}},
\end{equation}
where $\mathcal{R} \ge 1$ as $\Tcoh^{i+1} > \Tcoh^{i}$.
797

798
799
800
801
802
803
804
For the MCMC simulations to be succesful, this initial bounding
box, given the segment setup which produced the candidate, must be small
compared to the width of the signal (at that segment setup). If we start out
follow-up with the same search setup used in the source search, i.e.  we set
$\Nseg^0$ equal to the number of segments used in the input search, then for
the MCMC simulation to work, we require that $\V(\Nseg^{0}) \sim \mathcal{O}(100)$.

805
806
807
808
809
810
811
812
813
814
815
816
817
818
Given a fixed prior on the Doppler parameters and a fixed span of data, the
metric volume $\V_{\Nstages}$ for $\Tcoh^{\Nstages} = \Tspan$ is fixed, or in
terms of the number of segments, $\Nseg^{\Nstages} = 1$.
We can therefore define a minimisation problem:
\begin{align}
\min_{\Nseg^{i+1} \in \mathbb{N}}
\left| \log\mathcal{R} + \log\V(\Nseg^{i+1}) - \log\V(\Nseg^i)\right|
\end{align}
which can be solved numerically to a specified tolerance. Due to the difficulty
in solving such a minimisation with the integer constrain on $\Nseg^{i+1}$ we
simply solve it as a real scalar, and then round to the nearest integer. We now
have a method to generate a ladder of $\Nseg^{i}$ which keep the ratio of
volume fractions fixed. Starting with $\Nseg^{\Nstages}$ = 1, we generate
$\Nseg^{\Nstages-1}$ such that $\V^{\Nstages-1} < \V^{\Nstages}$ and
819
subsequently iterate.  Finally we must define $\V^{\rm min}$ as the stopping
820
821
822
823
824
825
826
827
828
829
830
831
832
833
criterion: a metric volume such that the initial stage will find a signal. This
stopping criterion itself will set $\Nstages$; alternatively one could set
$\Nstages$.

We have now defined a method to automate the task of choosing the ladder of
coherence times. This requires only two tuning parameters, the ratio between
stages $\mathcal{R}$ and the minimum metric volume $\V^{\min}$.

\subsection{Example}

We now provide an illustrative example of the follow-up method. We consider a
directed search over the sky position and frequency in 100 days of data from a
single detector, with $\sqrt{\Sn}=10^{-23}$~Hz$^{-1/2}$ (at the fiducial
frequency of the signal). The simulated signal has an amplitude
834
$h_0=2\times10^{-25}$ such that the signal has a depth of $\sqrt{\Sn}/h_0=50$
835
836
837
in the noise.

First, we must define the setup for the run. Using $\mathcal{R}=10$ and
838
$\V^{\rm min}=100$ our optimisation procedure is run and proposes the setup
839
layed out in Table~\ref{tab_follow_up_run_setup}. In addition, we show the
840
number of steps taken at each stage.
841
\begin{table}[htb]
842
843
\caption{The search setup used in Figure~\ref{fig_follow_up}, generated with
$\mathcal{R}=10$ and $\V^{\rm min}=100$.}
844
\label{tab_follow_up_run_setup}
845
\input{follow_up_run_setup}
846
847
\end{table}

848
The choice of $\mathcal{R}$ and $\V^{\rm min}$ is a compromise between the
849
850
851
852
853
total computing time and the ability to ensure a candidate will be identified.
From experimentation, we find that $\V^{\rm min}$ values of 100 or so are
sufficient to ensure that any peaks are sufficiently broad during the
initial stage. For $\mathcal{R}$ value much larger than $10^{3}$ or so where
found to result in the MCMC simulations `loosing' the peaks between stages, we
854
conservatively opt for 10 here, but values as large as 100 where also successful.
855
856

In Figure~\ref{fig_follow_up} we show the progress of the MCMC sampler during
857
the follow-up.  As expected from Table~\ref{tab_follow_up_run_setup}, during
858
859
860
the initial stage the signal peak is broad with respect to the size of the
prior volume, therefore the MCMC simulation quickly converges to it. Subsequently,
each time the number of segments is reduced, the peak narrows and the samplers
861
similarly converge to this new solution. At times it can appear to be inconsistent,
862
863
however this is due to the changing way that the Gaussian noise adds to the signal.
Eventually, the walkers all converge to the true signal.
864
\begin{figure}[htb]
Gregory Ashton's avatar
Gregory Ashton committed
865
\centering
866
\includegraphics[width=0.5\textwidth]{follow_up_walkers}
867
\caption{In the top three panels we show the progress of the 500 parallel
868
869
walkers (see Figure~\ref{fig_MCMC_simple_example} for a description) during the
MCMC simulation for each of the search parameters, frequency $f$,
870
871
right-ascension $\alpha$ and declination $\delta$. Each vertical dashed line indicates the start of a new stage of the search, the parameters for all stages
are listed in Table~\ref{tab_follow_up_run_setup}.}
872
873
874
\label{fig_follow_up}
\end{figure}

875
The key advantage to note here is that all walkers successfully converged to the
876
877
878
879
signal peak, which occupies $\sim 10^{-6}$ of the initial volume. While it is
possible for this to occur during an ordinary MCMC simulation (with $\Tcoh$
fixed at $\Tspan$), it would take substantially longer to converge as the
chains explore the other `noise peaks' in the data.
880

Gregory Ashton's avatar
Gregory Ashton committed
881
882
\section{Monte Carlo studies}

883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
In order to understand how well the MCMC follow-up method works, we will test
its ability to succesfully identify simulated signals in Gaussian. This will be
done in a Monte Carlo study, with independent random realisations of the
Guassian noise, amplitude, and Doppler parameters in suitable ranges. Such a
method is analagous to the studies performed in \citet{shaltev2013}, except
that we present results as a function of the fixed injected signal depth,
rather than the squared SNR.

In particular we will generate a number of 100-day data sets containing
independent realisations of Gaussian noise and a simulated CW signal. We choose
the parameters of the signal in such a way to model the candidates generated
from directed and all-sky searches by drawing the signal parameters from
appropriate distributions. However, we do not draw $h_0$ randomly, but instead
run the MC study at a number of selected values chosen such that given the
fixed $\sqrt{S_n}=2\times10^{3}$, the signals are injected with a depth
$\mathcal{D} \in [100, 400]$.  To simulate an isotropic distribution of
sources, we draw the remaining amplitude parameters for each signal uniformly
from $\phi \in [0, 2\pi]$, $\psi \in [-\pi/4, \pi/4]$, and $\cos\iota \in [-1,
1]$.

To provide a reference, we will compare the MC follow-up study against the
expected maximum theoretical detection probability for an infinitly dense
fully-coherent search of data containing isotropically-distributed signals as
calculated by \citet{wette2012}. Note however that we will parameterise with
respect to the signal depth (i.e. using Equation~(3.8) of \citet{wette2012} to
relate the averaged-SNR to the depth). The probability is maximal in the
sense that signals are `lost' during the follow-up due simply to the fact that
they are not sufficiently strong to rise above the noise.
911
912


Gregory Ashton's avatar
Gregory Ashton committed
913
\subsection{Follow-up candidates from a directed search}
914
\label{sec_directed_follow_up}
Gregory Ashton's avatar
Gregory Ashton committed
915

916
917
918
In a directed search, the sky location parameters $\alpha$ and $\delta$ are
fixed - in our study we fix them on the location of the Crab pulsar, but this
choice is arbitrary and holds no particular significance. The ouput of a
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
semi-coherent gridded directed search would provide the grid-point with the
highest detection statistic, and some box bounding the candidate given some
uncertainty. We will assume in this section that given the search setup
$\Nseg^{0}$ of the input search, the bounding box is sufficiently small to
ensure that $\V(\Nseg^0)\sim \mathcal{O}(100)$. This is not a limiting
assumption as any search can (quite cheaply) increase the density of grid
points around any interesting candidates in order to better constrain their
uncertainty.

The behaviour of the follow-up is independent of the exact frequency and
spin-down values used (this is not true for an all-sky follow-up as discussed
in Section~\ref{sec_all_sky_follow_up}). As such, we can, without loss of
generality define our Monte-Carlo follow-up in the following way. First, we
select an arbitrary frequency and spindown of $f_0=30$~Hz and
$\dot{f}_0=10^{-10}$Hz/s and a surrounding uncertainty box $\Delta f$ and
$\Delta \dot{f}$ chosen such that the uncertainty in frequency and spindown are
roughly equivalent in terms of mismatch. Then, we pick a candidate uniformly
from within this uncertainty box; this choice reflects the fact that the grid
is chosen such that the probability distribution of candidate signals is
uniform.



Having generated the data given the prescription above, we proceed to perform a
hierarchical MCMC follow-up. Given the data span and initial bounding box, we
compute the optimal heirarchical setup, the details of which are given in
Table~\ref{tab_directed_MC_follow_up}, this setup was used for all MC
simulations. This table also lists the number of steps, which in this case we
chose to be 10 - quite a small number of steps for a typical MCMC simulation.
948

Gregory Ashton's avatar
Gregory Ashton committed
949
950
\begin{table}[htb]
\caption{Run-setup for the directed follow-up Monte-Carlo study, generated with
951
$\mathcal{R}=10$ and $\Nseg^0=20$.}
Gregory Ashton's avatar
Gregory Ashton committed
952
953
954
955
\label{tab_directed_MC_follow_up}
\input{directed_setup_run_setup}
\end{table}

956
957
958
959
960
961
962
963
964
965
This process yeilds a maximum detection statistic $\widetilde{2\F}^{\rm max}$.
The signal is considered `detected' if $\widetilde{2\F}^{\rm max} >
\widetilde{2\F}^{\rm th}$, where we set the threshold at $2\F^{\rm th}=60$,
corresponding to a p-value of \comment{Finish section}.  In
Figure~\ref{fig_directed_MC_follow_up} we plot the the fraction of the MC
simulations which where recovered and compare this against the theoretical
maximum, given the threshold. The figure demonstrates that the recovery power
of the MCMC follow-up shows only a small margin of loss compared to the
theoretical maximum. \comment{Probably best to run with more steps and see if
this doesn't remove the loss}
Gregory Ashton's avatar
Gregory Ashton committed
966
967
968
969
970
971
972
973
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\textwidth]{directed_recovery}
\caption{Recovery fraction for the directed follow-up. The Monte-Carlo results
come from random draws searches using the setup described in
Table~\ref{tab_directed_MC_follow_up}.}
\label{fig_directed_MC_follow_up}
\end{figure}
974

975

976
\subsection{Follow-up candidates from an all-sky search}
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
\label{sec_all_sky_follow_up}

We now test the follow-up method when applied to candidates from an all-sky
search which, by definition, have uncertain sky-position parameters $\alpha$
and $\delta$. Searching over these parameters, in addition to the frequency
and spin-down not only increases the parameter space volume that needs to be
searched, but also adds difficulty due to correlations between the sky-position
and spin-down \comment{Does some of Karl's work discuss this?}.

To replicate the condition of candidates from an all-sky search, we draw the
candidate positions isotropically from the unit sphere (i.e. $\alpha$ uniform
on $[0, 2\pi]$ and $\delta = \cos^{-1}(2u{-}1){- }\pi/2$ where $u$ is uniform
on $[0, 1]$). We then place an uncertainty box containing the candidates with a
width $\Delta\alpha=\Delta\delta=0.05$; this box is not centered on the
candidate, but chosen in such a way that the location of the candidate has a
uniform probability distrubution within the box. This neglects the non-uniform
variation in $\delta$ over the sky-pacth, but given the small area should not
cause any significant bias. The frequency, spin-down, and amplitude parameters
are chosen in the same way as for the directed search
(Section~\ref{sec_directed_follow_up}). 

Producing \CHECK{1000} indepedant MC simulations we the perform a follow-up on
each using the setup given in Table~\ref{tab_allsky_MC_follow_up}. The
resulting recovery fraction as a function of the injected signal depth is given