PHYSICAL REVIEW D 108, 063008 (2023)

Exploring the capabilities of Gibbs sampling in pulsar timing arrays

Nima Laal ,1,* William G. Lamb ,2 Joseph D. Romano ,3 Xavier Siemens,1
Stephen R. Taylor,2 and Rutger van Haasteren 4
1
Department of Physics, Oregon State University, 1500 SW Jefferson Way, Corvallis, Oregon 97331, USA
2
Department of Physics and Astronomy, Vanderbilt University,
2301 Vanderbilt Place, Nashville, Tennessee 37235, USA
3
Department of Physics, Texas Tech University, Box 41051, Lubbock, Texas 79409, USA
4
Max-Planck-Institut für Gravitationsphysik (Albert-Einstein-Institut),
Callinstrasse 38, D-30167, Hannover, Germany

(Received 23 May 2023; accepted 17 August 2023; published 7 September 2023)

We explore the use of Gibbs sampling in estimating the noise properties of individual pulsars and
illustrate its effectiveness using the NANOGrav 11-year dataset. We find that Gibbs sampling noise
modeling (GM) is more efficient than the current standard Bayesian techniques (SM) for single pulsar
analyses by yielding model parameter posteriors with average effective-sample-size ratio (GM/SM) of
6 across all parameters and pulsars. Furthermore, the output of GM contains posteriors for the Fourier
coefficients that can be used to characterize the underlying red noise process of any pulsar’s timing
residuals, which are absent in current implementations of SM. Through simulations, we demonstrate the
potential for such coefficients to measure the spatial cross-correlations between pulsar pairs produced by a
gravitational wave background.

DOI: 10.1103/PhysRevD.108.063008

I. INTRODUCTION span of the data and linearly with the number of pulsars in
the array. In this regime, increasing the number of pulsars is
Pulsar timing arrays (PTAs) [1,2] are low-frequency
the best way to maximize PTA sensitivity to the GWB.
gravitational-wave (GW) detectors that use high-precision
Currently, the International Pulsar Timing Array (IPTA)
measurements of the times-of-arrival (TOAs) of pulses monitors 65 millisecond pulsars with 27 of such pulsars
produced by an array of millisecond pulsars (MSPs). MSPs observed for more than 10 years [5]. For this reason, in each
have ultra-stable spin periods on the order of milliseconds, new release of a PTA dataset the number of pulsars used in
and if their TOAs are measured to sufficient accuracy using GWB detection analyses is expected to grow, which in turn
large and sensitive radio telescopes, they can be used as makes the computational cost of noise modeling and
cosmic clocks spread throughout our galaxy. Accurate parameter estimation increase significantly. This poses a
models are constructed to predict the time at which each significant challenge for Bayesian inference as typical
pulse is expected to arrive, and small deviations from the searches for a GWB involve working with a very large
expected TOAs caused by GWs can be detected by parameter space making the use of computationally effi-
searching for quadrupolar spatial correlations in those cient algorithms a necessity.
deviations between pulsars in the PTA [3]. The standard Bayesian techniques for single and multi-
In recent years, multiple PTA searches for an isotropic pulsar noise modeling often result in a joint probability
stochastic gravitational wave background (GWB) have distribution for all of the model parameters (see Sec. II D).
uncovered a common red noise process [4–7]. This process Despite the flexibility that this approach offers in choosing
was recently observed to posses a quadrupolar correlation and implementing various noise models, the computational
signature matching the predictions of Einstein’s general cost of parameter estimation using Markov Chain
theory of relativity with various levels of significance [8–11]. Monte Carlo (MCMC) simulations becomes prohibitive
The sensitivity of PTAs to a GWB depends primarily on quickly. For instance, in the case of single-pulsar analyses,
the number of pulsars in the array [12]. This is due to the the number of parameters required to describe a pulsar’s
fact that, at late times, the lowest frequencies in PTA noise may well exceed forty (see Sec. III). This problem is
datasets become GW-dominated, and the significance of more severe for the case of multipulsar analyses as even the
the cross-correlations grows with the square root of the time simplest noise models require a number of parameters that
is larger than twice the number of pulsars in the PTA.
Hence, more computationally efficient data analysis tech-
*
[email protected] niques are critical for the future of PTA analyses.

2470-0010=2023=108(6)=063008(14) 063008-1 © 2023 American Physical Society

NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

To mitigate these problems, there have been numerous pðρ; b; njrÞ ∝ pðrjb; nÞpðaÞpðϵÞpðρÞpðnÞ; ð2Þ
efforts toward the development of more efficient Bayesian
GWB detection techniques to analyze PTA datasets, such where,
as those presented in [13–19]. In particular, the work of van
Haasteren and Vallisneri [14] provides an outline for single- exp f− 12 ½ðr − TbÞT N −1 ðr − TbÞg
pðrjb; nÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð3Þ
pulsar noise analyses in which Gibbs sampling can be used det fð2πÞNg
to characterize the red noise component of each pulsar’s
timing residuals. In this paper, we explore the capabilities exp f− 12 ½aT φ−1 ag
of the Gibbs sampling method in single-pulsar noise pðaÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð4Þ
det fð2πÞφg
analyses by applying it on the NANOGrav 11-year dataset
20]] as well as simulated datasets. We show that the Gibbs Yk
1
sampling method is well suited for PTA single-pulsar pðρÞ ¼ ; ð5Þ
analyses and results in probability distribution functions ρ
s¼1 s
for all model parameters in a significantly shorter timescale
compared to those obtained via the standard MCMC for
methods. Furthermore, we show, via simulated datasets,
that the Fourier coefficients that result from the Gibbs φ ¼ haaT i; ð6Þ
sampling procedure can be used to identify the shape of the
underlying spatially-correlated signal in a PTA dataset. B ¼ hbbT i; ð7Þ
The paper is structured as follows. In Sec. II, we review
and simplify the methods presented in van Haasteren and and ρ denoting the collective set fρ1 ; ρ2 ; …; ρk g whose
Vallisneri [14] to outline the Gibbs sampling method and its elements are used to parametrize a pulsar’s power-spectral-
accompanying noise modeling. Furthermore, in order to density, frequency-bin by frequency-bin,2 and describe the
use the output of Gibbs sampling in a subsequent multi- variance of the Fourier coefficients. Additionally, a log-
pulsars analysis, and inspired by Anholm et al. [21], we uniform (conjugate) prior pðρk Þ ¼ 1=ρk is considered as
introduce our version of a frequency domain optimal seen in Eq. (5).
statistic which follows from the PTA multipulsar likelihood Moreover, we have assumed an unbounded improper
function. In Sec. III, we employ the outlined method in prior for the linear timing model parameters and have set
order to analyze the NANOGrav 11 year dataset and hϵϵT i ¼ diagf∞g. Such choices for the linear timing
compare the results to those obtained by the standard model parameters are typical of PTA noise analyses due
Bayesian PTA detection techniques. Finally, in Sec. IV, we to the lack of physically motivated priors for all of the
analyze PTA simulated datasets to reveal the potential of timing model parameters and are acceptable as long as the
the Gibbs sampling technique in searches for a common data is informative with respect to such parameters. Hence,
correlated signal across an array of pulsars. we can write
 
−1
0 0
II. METHODS B ¼ : ð8Þ
0 φ−1
We begin our review of the Gibbs sampling method [14]
by writing a simple model for a pulsar’s post-fit timing To proceed with Gibbs sampling, the posterior for each
residuals, r, in terms of a set of Fourier coefficients a, of the model parameters needs to be cast into a conditional
Fourier design matrix F, linear timing model parameters ϵ, probability distribution form where each model parameter
timing design matrix M, and white noise w1: is conditioned upon the other model parameters and the
timing residuals. In the following two subsections, we
r ¼ Mϵ þ Fa þ w derive such conditional probabilities for parameters b
and ρ.
¼ Tb þ w; ð1Þ
A. Conditional probability of coefficients
where bT ¼ ½ϵ; a and T ¼ ½M; F. Assuming Gaussian For the coefficients b, the conditional probability can be
white noise, parametrized by the set of parameters n with found by rewriting the full posterior [i.e., the product of
prior pðnÞ, the above model allows for the construction of Eqs. (3)–(5)] while ignoring all factors not depending on b
posterior probability density functions following Bayes’
theorem: 2
Note that the total number of frequency-bins is k, but there are
two Fourier coefficients per each frequency-bin. Both acos s and
2
s have the same variance parametrized by ρs . This is reflected
1
Refer to Table I and Sec. A for more details on the definitions asin
of the quantities used throughout this paper. in Eq. (4).

063008-2
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

coefficients explicitly. In other words, all model parameters the analytic form for the dependence of ρ on the Fourier
are treated as constants and only the b coefficients are coefficients a, the lower and the upper bounds of the inverse-
allowed to vary: gamma distribution extending to zero and infinity would lead
to astrophysically and statistically incorrect assumptions as
1
ln pðbjρ; r; nÞ ≃ − ½ðr − TbÞT N −1 ðr − TbÞ þ bT B−1 b such bounds need to be finite and constrained to avoid the
2 implicit use of improper priors in the modeling of red noise
1 processes. Thus, a truncated version of the derived inverse-
¼ − ½bT ðT T N −1 T þ B−1 Þb
2 gamma distribution needs to be considered. In Sec. B, we
1 show how to obtain such a truncated distribution.
− ½−2bT T T N −1 r: ð9Þ
2
The above equation suggests that the b-dependence of the C. Conditional probability of white noise parameters
probability pðbjρ; r; nÞ is Gaussian. Using the maximum In contrast to b and ρ, the white noise parameters cannot
a posteriori estimate of b found by maximizing Eq. (9) as be written in terms of standard statistical distributions. This
an estimate of the mean of the Gaussian, one can write the is mainly due to the dependence of the white noise
conditional probability distribution of the b coefficients, parameters to various radio telescope receivers (i.e., each
Eq. (9), in the form backend of each radio telescope needs its own white noise
parameters). Solving the full-likelihood for the white noise
exp f− 12 ðμ̂ − bÞT Σðμ̂ − bÞg parameters, collectively denoted by n, results in
pðbjρ; r; nÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð10Þ
det fð2πÞΣ−1 g
1X
p
where,3 ln pðnjρ; b; rÞ ¼ − fðr − TbÞT N −1 ðr − TbÞg
2 i¼1
Σ ¼ T T N −1 T þ B−1 ; ð11Þ 1X
p
− ln ðdet f2πNgÞ; ð14Þ
μ̂ ¼ Σ−1 T T N −1 r: ð12Þ 2 i¼1

where the sum is over the TOAs. Since Eq. (14) cannot be
B. Conditional probability of red noise simplified further in any useful way, we have no choice but
power-spectral-density to utilize a non-Gibbs MCMC procedure to sample the
posterior.
Similar to the b coefficients, the conditional probability
of the ρ parameters can be found by taking advantage of the
full posterior and ignoring all the factors not depending on D. Standard method of single-pulsar analyses
ρ explicitly. Additionally, we make the observation such The standard method of single-pulsar analyses involves
that the relevant probability distributions can be factorized an analytical marginalization of the product of Eqs. (3)
over frequency-bins: and (4) over the coefficients b. The result is
Y
k  
1 1
pðρja; r; nÞ ¼ pðρs jas ; r; nÞ pðrjρÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − rT C−1 r ; ð15Þ
s¼1 det fð2πÞCg 2
Y   
k
1 1 as · as
¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − C ¼ N þ TBT T ; ð16Þ
s¼1 ρs ð2πÞρs
2 2 ρs
Y  as ·as  C−1 ¼ N −1 − N −1 TΣ−1 T T N −1 ; ð17Þ
k
1 2
∝ 2
exp −
s¼1
ρs ρ s where in the last line, we have used the Woodbury identity:
Yk  
as · as
¼ InvGamma α ¼ 1; β ¼ : ð13Þ ðX þ UYVÞ−1 ¼ X−1 − X−1 UðY −1 þ VX−1 UÞ−1 VX−1 ;
s¼1
2
ð18Þ
In the above, the dot-product denotes the sum of the square of
the cosine and sine Fourier coefficients for each frequency- and Σ is defined in Eq. (11). The dependence of Eq. (15) on
bin that is as · as ¼ ðacos 2 sin 2
s Þ þ ðas Þ . Furthermore, despite the red noise parameters ρk is through the elements of the
matrix Σ−1 . Once Eq. (15) is multiplied by the appropriate
3
The definition of Σ in Eq. (11) is chosen so that this paper’s Σ priors of the model parameters, the resulting joint proba-
represents the same quantity as the Σ defined in the PTA GWB bility distribution of pðρjrÞ is ready to be given to a non-
detection literature. Gibbs MCMC algorithm for parameter estimation.

063008-3
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

E. Gibbs sampling
Gibbs sampling [22] is a MCMC algorithm designed to
take advantage of the conditional probability distributions
of all model parameters in order to perform parameter
estimation. It is often used in statistical inferences where a
joint probability distribution of all parameters is difficult to
sample, yet each model parameter’s probability distribution
can be written in terms of the rest of the parameters and the
data. Gibbs sampling allows for random draws from the
conditional probability distributions of model parameters
whose analytic functional form must be found prior to the
start of the sampling process as we have done for the case of FIG. 1. A schematic representation of the first three steps of the
single-pulsar noise analyses by deriving Eqs. (10) and (13). outlined Gibbs sampling procedure. The first step of the sampling
Due to the existence of analytic forms for the probabilities, process (blue) starts by guesses of the ρ and the white noise
the concept of rejection of random states, an integral part of parameters and results in an estimate of the coefficients b
the other MCMC algorithms, does not belong to the Gibbs following Equation (10) using the previously guessed values.
sampling as all draws are considered accepted. Nevertheless, The second (red) and the third (green) steps of the sampling
Gibbs sampling is still a MCMC algorithm as it possesses continue the sequence by estimating the next remaining model
features such as no long-term-memory and the need for burn- parameter given the most recent estimates of the other two
in of the final Markovian chain. We will outline a step-by- parameters using the conditional probability distributions of
Eqs. (10), (13), and (14).
step implementation of Gibbs sampling for a single-pulsar
noise analysis in the remaining part of this section.
Knowing the conditional probabilities of our model section, the output of Gibbs sampling provides enough
parameters, ρ, b, and n, it is simple to implement Gibbs information to perform multipulsar analyses aiming at
sampling in the following way: detecting a GWB.
(i) Step 1: Make initial guesses of ρ and n denoted by ρ0 Using only the Fourier coefficients a, one can construct a
and n0 . factorized likelihood in the frequency domain in the
(ii) Step 2: Using Eq. (10), find an estimate of b0 given ρ0 following way:
and n0 .
n o
(iii) Step 3: To start the first iteration, find an estimate of ρ1 exp − 12 ðaTk;I Φ−1 ks as;J Þ
given b0 and n0 using Eq. (13). pðAjΦÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð19Þ
(iv) Step 4: Continuing the first iteration, find an estimate detfð2πÞΦg
of n1 given b0 and ρ1 with a very short MCMC
procedure sampling Eq. (14). n o
(v) Step 5: To end the first iteration, find an estimate of b1 Y exp − 12 ðaTs;I Φ−1 a
ss s;J Þ
¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð20Þ
given ρ1 and n1 using Eq. (10). s detfð2πÞΦss g
Figure 1 provides an illustration of the explained
procedure. The above steps can be repeated until all the for
model parameters reach satisfactory convergence. Due to
the analytical draws of the ρ and the b coefficients, 2 3
Φ11 0 … 0
convergence will be reached quickly compared to the fully
6 7
6 0 Φ22 … 0 7
non-Gibbs MCMC algorithms. This is one of the most 7
desirable features of Gibbs sampling as the overall run-time Φ¼6
6 .. .. .. .. 7; ð21Þ
4 . . . . 7
of the PTA single-pulsar noise analyses will be reduced 5
significantly. 0 0 … Φkk

2 3
F. Frequency domain multipulsar likelihood φk;1 Γk;12 ρ2k;g … Γk;1m ρ2k;g
The outlined Gibbs sampling procedure is an efficient 6 7
6 Γk;21 ρ2k;g φk;2 … Γk;2m ρ2k;g 7
Bayesian scheme capable of estimating each pulsar’s 6 7
Φkk ¼6 .. .. .. .. 7; ð22Þ
power-spectral-density as well as the Fourier coefficients 6 7
4 . . . . 5
required to describe the total red noise (i.e., GWB plus
spatially-uncorrelated intrinsic red noise process) compo- Γk;m1 ρ2k;g Γk;m2 ρ2k;g … φk;m
nent of the timing residuals. However, the information
required in characterizing a GWB requires subsequent where A denotes the collection of Fourier coefficients,
multipulsar analyses. As will be demonstrated in this across all pulsars and all frequencies, (i.e., A ¼ fak;m g),

063008-4
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

ρk;g parametrizes the common power-spectral-density of the and aJ . The trivial option is to draw randomly from the
GWB (indicated by the subscript g) observed across the multivariate probability distribution of each pulsar’s a (the
entire pulsar array at frequency k, and ΓIJ represents output of Gibbs sampling) and obtain the cross product of
the functional form of the cross correlations (e.g., such random draws for each pulsar pair. Another option is
Hellings and Downs curve). to construct posteriors of the mean, μ̂, following Eq. (10),
One can use this equation to derive an optimal estimator and draw randomly from such posteriors. Similar to the
of the signal-to-noise analogous to those presented in previous case, the cross product of the random draws can be
Anholm et al. [21] and Chamberlin et al. [23]. We leave used in Eq. (23) and with the difference that the normali-
the details of the derivation to our future project [24] where zation factor in the denominator of Eq. (23) should be
we explore the use of the Fourier coefficients in GWB recalculated (see Appendix C for more details). Lastly, for
characterization in great detail. Here, we simply report the the choice of φI , we use the total red noise power PI.
results in the form of the optimal estimators of the cross- As a final note, it is important to recognize the limitations
correlations λIJ and their uncertainty σ IJ : of the presented technique as well as the optimal statistic in
general. In practice, optimal statistic results in biased
P P̂ estimates of the GWB amplitude and the signal-to-noise
s as;I · as;J φs;I φg s;J
λIJ ¼ P ; ð23Þ ratio if one does not have separate estimates for the
P̂2g
s φs;I φs;J spatially-uncorrelated as well as the common red noise
power. In other words, if one uses the red noise power
X −1 estimates from the single-pulsar analyses instead of
P̂2g 2
σ IJ ¼ : ð24Þ obtaining separate estimates for a common red noise signal
s
φs;I φs;J and intrinsic red noise signal, one cannot characterize a
common correlated signal correctly. This has been explored
Without a need for a detailed derivation, Eqs. (23) and (24) in depth in [25].
can be understood by following a very simple rational. The
numerator is the weighted product of as;I · as;J . The weights
III. ANALYSIS OF THE NANOGrav
associated with such product, 1=φI and 1=φJ , have the role of
11 YEAR DATASET
suppressing the contributions from pulsars whose total non-
GWB noise power is substantial (i.e., dominant spatially- To test the capabilities of the outlined single-pulsar data
uncorrelated intrinsic red noise). Moreover, the choice for the analysis technique, we analyze the NANOGrav 11 year
normalization in the denominator ensures that the estimated dataset [20] using Gibbs sampling. The results are then
correlations would yield A2g ΓIJ if averaged over many compared to the ones obtained via standard Bayesian
realizations of GWB as is shown in Sec. C. modeling detection routine used by the NANOGrav col-
Additionally, estimates of the amplitude, the uncertainty laboration in their most recent work [4]. To ensure the
of the estimated amplitude, and the signal-to-noise-ratio fairness of the convergence comparisons, we allow each
(SNR) can be made from Eqs. (23) and (24) by minimizing technique to sample the dataset for two hours for each
a weighted-chi-squared statistic of the form pulsar. After the two hours time-limit, we compare the
posteriors’ effective-sample-size (ESS) and rank-normal-
X ðλIJ − A2g ΓIJ Þ2 ized-split R-hat (r̂) values using the diagnostic tools
χ2 ¼ ; ð25Þ provided by Kumar et al. [26].
IJ
σ 2IJ

with respect to A2g which results in A. Details of the Bayesian modeling

The Gibbs sampling implementation used for the 11 year
P P P̂
IJ;I≠J s as;I · as;J ΓIJ φs;I φg J;s dataset models the data as outlined in Sec. II. This Bayesian
Â2g ¼ P P ; ð26Þ modeling together with Gibbs sampling is referred to as
2 P̂2
s ΓIJ φI;s φs;J
g
IJ;I≠J Gibbs Method (GM) from hereon. Moreover, the compet-
ing method of analyzing the NANOGrav 11 year dataset
X X −1
P̂2g 2 follows the standard single-pulsar analyses currently imple-
σg ¼ Γ2 ; ð27Þ mented in the most recent GWB searches [8–10] and
s IJ φI;s φJ;s
IJ;I≠J explained in Sec. II D. The PTMCMC sampling package
[27] as well as the structure of the Bayesian modeling
Â2g accompanying this sampling is referred to as Standard
SNR ¼ : ð28Þ
σg Method (SM) from hereon.
For both SM and GM, we have allowed each pulsar’s set
When estimating the optimal correlations using Eq. (23), of red noise parameters, ρ, to follow a 30 frequency free-
one has a few options to select from for the choice of aI power-spectral-density model with frequencies ranging

063008-5
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

FIG. 2. A comparison of posteriors for all model parameters of PSR J1713 þ 0747 for the NANOGrav 11 year dataset obtained via
GM (blue) and SM (red). The posteriors on the left column belong to the red noise model parameters, collectively referred to as ρ,
whereas the posteriors on the right column belong to the white noise parameters EFAC, EQUAD, and ECORR. There is one ρk
parameter for each frequency (k ¼ 30 frequencies in total) and three white noise parameters for each receiver (8 receivers in total). To
obtain the plots via GM, 30 steps of a Metropolis Hasting algorithm within each step of Gibbs sampling has been implemented for the
white noise parameters. The above plots show a great level of consistency in extracting the posteriors between the two methods.

from 1=T obs to 30=T obs in which T obs denotes the obser- ECORR½s ∼ log-Uniformð−8.5; −5Þ; ð31Þ
vational baseline of each considered pulsar. The choice of
prior for the model parameters are listed below. For each EFAC ∼ Uniformð0.01; 10Þ; ð32Þ
pulsar, the white noise parameters are per receiver/backend
system while the ρ parameters are per frequency: for [s] denoting the unit of the quantities, which is seconds.

ρ½s ∼ log-Uniformð−9; −4Þ; ð29Þ B. Comparison of posteriors

For the sake of brevity, out of the thirty four pulsars of
EQUAD½s ∼ log-Uniformð−8.5; −5Þ; ð30Þ the NANOGrav 11 year dataset, we have chosen to feature

063008-6
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

FIG. 3. A histogram showcasing the distribution of the Hel- FIG. 4. A histogram showcasing the distribution of the Hel-
linger distance values for the log10 ρ posteriors (blue) and the linger distance values between the log10 ρ posteriors obtained
white noise parameters (orange) obtained by comparing the using GM with two different number of MCMC steps (30 and 5
outputs of GM and SM. The histogram contains the Hellinger steps) for each step of Gibbs sampling for the white noise
distances of model parameters across all frequencies and pulsars. parameters. The figure is made by combining the Hellinger
As evident by the distribution, GM and SM result in sufficiently distance values across all of the pulsars and all of the frequencies.
similar distributions with a few exceptions whose inconsistencies As evident by the distribution, choosing a much lower number of
can be attributed to the differences in the level of convergence of MCMC steps for each step of Gibbs sampling for the white noise
posteriors resulting from GM and SM even though we have parameters does not change the shape of the target log10 ρ
allowed sufficient time for SM to converge (i.e., more than two posteriors significantly. This effect can be attributed to the
hours). GM posteriors follow the general shape of SM posteriors knowledge of GM about the analytical shape of the log10 ρ
but are significantly more converged. parameters prior to the start of the sampling.

C. The effect of using different number

a GM vs SM posterior comparison plot for only PSR of MCMC steps in GM
J1713 þ 0747 as this pulsar has the longest observational To obtain the white noise posteriors of Fig. 2, 30 steps of
baseline as well as the largest number of TOAs making it a Metropolis Hasting algorithm for each step of Gibbs
the most computationally expensive pulsar to analyze. As sampling has been implemented. The choice for the number
shown in Fig. 2, the two techniques yield consistent of MCMC steps for each step of the Gibbs sampling
posteriors for both the red noise and the white noise model depends on factors such as the number of TOAs, one’s
parameters for PSR J1713 þ 0747 showcasing the robust- threshold and preferred measure of convergence for the
ness and the capability of GM to be implemented on real posteriors as well as the efficiency of the type of MCMC
PTA datasets. The same consistency is also observed in all algorithm used in the white noise parameter estimation.
the remaining thirty three pulsars. For a quantification of However, the red noise parameters’ posteriors are not
the degree of consistency between the two sets of poste- overly sensitive to this choice as the target distributions
riors, refer to Fig. 3 which highlights the differences in the for ρk parameters are all analytically determined prior to the
output of GM and SM in the form of a histogram of start of sampling. To test the sensitivity of the red noise
Hellinger distance [28]4 values across all pulsars. With the parameters to the choice for the number of MCMC steps for
exception of a few white noise parameters, the Hellinger each step of Gibbs sampling, we have applied GM on all of
distances are concentrated between 0 and 0.2 indicating an the NANOGrav 11 year pulsars using only 5 steps of
adequate degree of consistency between the GM and the MCMC. As shown in Fig. 4, the estimated Hellinger
SM posteriors. We attribute the higher Hellinger distance distance values between the two sets of posteriors of
values of some model parameters (especially the white log10 ρ parameters are sufficiency low suggesting a weak
noise parameters) to the differences in the level of con- degree of correlation between the red noise parameters’
vergence of the posteriors as GM is more successful at posteriors to the white noise parameters’ if analyzed via
yielding converged posteriors than SM. Refer to Sec. III D GM. Nevertheless, our current implementation of GM is
for a more detailed discussion. adequately optimized to handle large number of MCMC
steps without much of a sacrifice in the overall run-time of a
single-pulsar analysis.
4
Hellinger distance is a measure of similarity between two D. Comparison of convergence levels
probability distributions ranging from 0 (identical distributions)
to 1 (disagreeing distributions). For two discrete probability Despite resulting in consistent posteriors, SM and GM
distributions p and q, the Hellinger distance H is defined as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P pffiffiffiffiffi pffiffiffiffi 2ffi
differ significantly in their state of convergence of the
1ffiffi
H¼ 2 p
i ð pi − qi Þ , where i ranges over the binned model parameters, especially those pertaining to the effec-
quantities of interest whose probability distribution is described tive-sample-size (ESS). Figure 5 shows the spread of the
by p and q. ratio of ESS values (GM divided by SM) across all of the

063008-7
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

the correlation content of the dataset using the concepts

discussed in Sec. II F as the 11 year dataset lacks a common
correlated signal across pulsar pairs [20]. For studying the
correlations, we will dedicate future projects to the analysis
of the NANOGrav 15 year [8] and the upcoming IPTA’s
DR3 datasets. Meanwhile, to explore the capability of the
Fourier coefficients a in characterizing a common spatially
correlated signal, we make use of simulated PTA datasets.

FIG. 5. A scatter-plot showcasing the differences in the spread A. Details of the simulations
of the effective-sample-size (ESS) values for the log10 ρ and the We have chosen two types of simulated datasets, referred
white noise parameters expressed in the form of the ratio of GMs’
to as SIM0 and SIM1, with 300 realizations for each type,
ESS over SMs’ ESS (blue circles). For each pulsar, there is one
log10 ρk for each frequency (30 frequencies in total) and three to analyze in order to explore the capability of the Fourier
white noise parameters for each receiver. Across all of the pulsars, coefficients a to characterize a common correlated signal.
GM is more capable at yielding posteriors with significantly The two simulated datasets are identical in every aspect
higher ESS levels given the two hour time limit. Considering all except the content of their spatially-uncorrelated intrinsic
model parameters, the average ESS ratio is 6. The values of ESS red noise: for SIM0, the log of the amplitude of the
are found using the functionalities provided in Kumar et al. [26]. spatially-uncorrelated intrinsic red noise of each pulsar is
randomly chosen from a uniform distribution between
10−16 and 10−14 while for SIM1 this range is between
10−14 and 10−13 . For both datasets’ pulsars, the spectral
index of the spatially-uncorrelated intrinsic red noise
follows a uniform distribution with lower and upper bounds
of 0 and 7 respectively. Additionally, each dataset has 90
pulsars uniformly scattered across the sky timed for
20 years with random timing cadences between 14 to
30 days. Furthermore, each dataset contains 10 microsec-
onds of white Gaussian noise for each pulsar as well as a
unique realization of a GWB with amplitude of Ag ¼
FIG. 6. A scatter-plot showcasing the differences in the spread 2 × 10−15 and spectral index of γ g ¼ 13=3. Lastly, to
of rank-normalized-split-R-hat, r̂, values obtained from the employ GM on each dataset, we keep the white noise
log10 ρ and the white noise parameters analyzed by GM and parameters constant and use the same range of frequency-
SM. Each blue circle represents a single model parameter and the bins for all pulsars which is f1=20 yrs; 2=20 yrs; 3=20 yrs;
figure is obtained by combining the r̂ − 1 values of all model 4=20 yrs; 5=20 yrsg.
parameters for all pulsars and across all frequencies. As evident It is worth mentioning that our intention is not about
by the figure, GM is more capable at resulting in posteriors with a
simulating realistic datasets and analyzing it with GM. We
lower r̂ level given the two hour time limit. The r̂ values are
estimated by dividing each Markov chain into two sub-chains and
have already shown the capability of GM in single-pulsar
applying the rank-normalized-split-R-hat test [26] on it. analyses of real datasets. Our intention is to highlight what
the Fourier coefficients can potentially reveal about an
existing GWB signal, hence the reason behind our choices
model parameters for every pulsar. As evident by Fig. 5, a for the specific parameters of the two simulated datasets.
significant majority of each pulsar’s model parameters have Nonetheless, we have introduced very high levels of
higher ESS values when analyzed using GM as compared to spatially-uncorrelated intrinsic red noise in the SIM1
SM. The average ESS ratio across all parameters and pulsars dataset (higher than what is observed in the real PTA
is 6. Figure 5 proves our claim about the high efficiency of datasets) as dealing with such processes is an extremely
GM. Additionally, the same observation can be made about challenging part of GWB searches using PTAs whose
the rank-normalized-split R-hat (r̂) values calculated for both impact on the correlation recovery using the Fourier
GM and SM posteriors for each pulsar. Figure 6 points coefficients is nontrivial.
toward the higher state of convergence of a significant
majority of the model parameters that were analyzed by GM. B. Reconstruction of red noise signal
using Fourier coefficients
IV. SIMULATIONS
The a coefficients are capable of reconstructing the red
Despite the successful implementation of GM on the component of the timing residulas as suggested by
NANOGrav 11 year dataset, we have not tried to analyze Equation (1). The reconstructed signal is pre-fit and

063008-8
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

FIG. 9. Two histograms comparing the distributions of the

FIG. 7. A comparison between postfit time series recovered common correlated signal between SIM0 (blue) and
reconstruction using the Fourier coefficients obtained from SIM1 (red) dataset using the method provided in Sec. II F. The
GM (blue), the injected red noise time series (red), and the total blue and the red vertical lines indicate the 16th and the 84th
residuals (green) for one of SIM0’s pulsars. The reconstructed percentiles (dotted lines) as well as the mean (dashed line) of each
residuals are made by considering the entire posterior probability distribution. Each distribution is obtained by combining the
distribution of the recovered Fourier coefficients. As evident by estimates of the amplitude [Eq. (26)] of the cross-correlated
the figure, the reconstructed post-fit red noise signal matches the signal over 300 realizations. The injected GWB signal is
underlying red noise signal closely. indicated with a vertical dashed gray line. The figure suggests
that the Fourier coefficients contain the right amount of infor-
white-noise-free. Once the reconstructed signal obtained by mation about the amplitude of the cross-correlated signal in the
case of SIM0. In the case of SIM1, due to the existence of
Fa is fitted for the timing model parameters, it mirrors the
significantly higher non-GWB red noise power, the recovered
underlying total post-fit red noise signal in the dataset GWB amplitude is more scattered.
closely. Figure 7 highlights this case for one of SIM0’s
pulsars. As suggested by the figure, the Fourier coefficients
are capable of reconstructing the underlying red noise
process of the total timing residuals. This fact allows the pulsar [see Eq. (10)]. Furthermore, since our goal is to
Fourier coefficients to be adequate replacement for the showcase the potential of the Fourier coefficients in
timing residuals in the frequency domain with the added revealing information about the GWB signal rather than
benefit that one no longer needs to take into account a white outlining a complete and practical pipeline capable of fully
noise process or be concerned with the complications of the characterizing a GWB signal, the weights φI are set to the
timing model parameters when using the a coefficients in a total red noise power that was used to generate the
subsequent analysis. In fact, the effects of the timing model simulated datasets.
parameters and the white noise levels are implicit in the The shape of the correlation recovery is depicted in Fig. 8
posteriors for the Fourier coefficients obtained via GM. for both simulated datasets. This shape is obtained by
dividing the pulsar pairs of each realization into 15 different
angular separation bins such that all bins have 267 pulsars
C. Searching for correlations using Fourier coefficients pairs in them. Additionally, the average, 16th, and 84th
To characterize the GWB signal in each of the realiza- percentiles (over the 300 realizations) of the correlations for
tions of SIM0 and SIM1, we use Eqs. (23) and (24) with μ̂I each angular separation bin is computed and indicated in
as the quantity representing the Fourier coefficients of each Fig. 8. Furthermore, the histogram of the estimated

FIG. 8. A plot depicting the reconstruction of the Hellings and Downs correlation (gray dashed curve) using GM’s estimates of the
Fourier coefficients obtained for both SIM0 (blue circles) and SIM1 (red stars) datasets. The reconstructions are the average over 300
realizations of both datasets. The error-bar of each point indicates the range between the 16th and the 84th percentiles over the 300
realizations. Remarkably, the recovery of the shape of the correlations is not affected significantly by the introduction of extremely high
levels of spatially-uncorrelated red noise to each pulsar in SIM1.

063008-9
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

using the output of GM (e.g., a power-law fit to the free-

spectrum model) [19]. Combined with the fitting utilities
provided by Lamb et al. [19], GM can become a powerful
and efficient tool for use in the future PTA GWB detection
analyses.

A. Software
The GM code takes advantage of the functionalities
FIG. 10. Two histograms comparing the distributions of the provided by ENTERPRISE [29] and ENTERPRISE-exten-
signal-to-noise ratio (SNR) between SIM0 (blue) and SIM1 (red) sions [30], and PTMCMC sampler [27]. The package ArviZ
dataset using the method provided in Sec. II F. The blue and the [26] has been used for diagnosing MCMC chains. Python
red vertical lines indicate the 16th and the 84th percentiles (dotted packages Matplotlib [31] and plotly [32] have been used for
lines) as well as the mean (dashed line) of each distribution. Each generating the figures in this paper.
distribution is obtained by combining the estimates of the SNR
(Equation (28) of the cross-correlated signal over 300 realiza-
tions. As expected, SIM1 dataset exhibits a lower SNR due to ACKNOWLEDGMENTS
containing a significantly higher non-GWB red noise power than
the GWB power. We thank our colleagues in NANOGrav for fruitful
discussions and feedback during the development of this
technique. We thank Justin A. Ellis for his early work on
this subject and the early version of the Gibbs method’s
amplitude Âg and signal-to-noise ratio of all the 300
code. The work of N. L., W. G. L, J. D. R, X. S., and S. R. T
realizations of each dataset are stacked on top of each other
was supported by the NANOGrav NSF Physics Frontier
(i.e., no averaging is performed) and presented in Figs. 9 and
Center awards No. 2020265 and No. 1430284. N. L. and
10 respectively. The impact of introducing extreme levels of
X. S acknowledges the support from the George and
intrinsic spatially-uncorrelated red noise to the dataset
Hannah Bolinger Memorial Fund, as well as the Larry
manifests itself in the form of lowering the signal-to-noise
W. Martin and Joyce B. O’Neill Endowed Fellowship in the
ratio and more scattered amplitude recovery. However, the
College of Science at Oregon State University. S. R. T
shape of the correlations recovery remains remarkably close
acknowledges support from NSF AST-2007993, and an
to the Hellings and Downs curve over many realizations.
NSF CAREER No. 2146016. This work was conducted in
part using the resources of the Advanced Computing
V. DISCUSSION AND FUTURE WORK Center for Research and Education (ACCRE) at
Vanderbilt University, Nashville, TN. This work was
In this paper, we have shown that the Gibbs method (GM)
performed in part at Aspen Center for Physics, which is
is an efficient single pulsar Bayesian noise analysis technique
supported by National Science Foundation Grant No. PHY-
capable of producing posteriors for the single-pulsar free-
2210452. J. D. R. acknowledges support from start-up
power-spectral-density and the white noise model parameters
funds from Texas Tech University.
with convergence properties that are superior to those
obtained using standard Bayesian methods (SM). GM is a
robust and computationally efficient alternative to SM for APPENDIX A: GWB DETECTION
future PTA noise analyses. Additionally, we have shown that TERMINOLOGY
the Fourier coefficients resulting directly from GM contain Most of the PTA noise analysis concepts have been
adequate information about the shape the cross-correlations developed over many years and scattered over many papers
signal through the use of simulations. In effect, GM produces [12–14,21,23,33–37]. To help readers better understand the
the frequency domain representation of each pulsar’s red methods used in this paper, we define the necessary PTA
noise signal, free of white noise and timing model param- noise analysis quantities and concepts in this section.
eters, hence providing all the necessary information to start Additionally, refer to Table I for a short description of
performing subsequent GWB detection analyses exclusively the mathematical symbols used throughout this paper.
in the frequency domain.
GM results in raw information in the frequency domain 1. Basis matrices and their coefficients
which may need to be processed further depending on the
needs of the subsequent analyses. For instance, the astro- To model the contribution of any red noise process to the
physical interpretation of a pulsar’s red noise signal will timing model residuals of a given pulsar, rRed , we employ a
require a more constrained model of the power spectral Fourier basis matrix and a vector of coefficients such that
density than the free-spectrum model which could be
achieved by fitting for the parameters of such a model rRed ¼ Fa; ðA1Þ

063008-10
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

TABLE I. A table listing the symbols most commonly used throughout this paper and a short description of what
they represent. Refer to Sec. A for more details on the definition of some of the quantities.

Symbol Description
T obs Observational baseline
t Time
f Frequency
I, J Pulsar indices
k, s Indices for the frequency bins
m Index for the number of pulsars in the array
p Number of TOAs for a given pulsar
q Number of timing model parameters
r Timing residual
F Fourier design matrix
M Timing model design matrix
T Combination of F and M such that T ¼ ½M; F
N White noise covariance matrix
B Covariance matrix for the linear timing model parameters and the Fourier coefficients (i.e., hbbT i)
Γ Hellings and Downs cross correlation matrix
A Collection of fourier coefficients across pulsars and frequencies fak;m g,
φ Single-pulsar red process covariance matrix
Φ Multi-pulsar red process covariance matrix
a Fourier sin-cos coefficients
λ Estimated cross correlations
ϵ Linear timing model parameters
b Combination of a and ϵ such that bT ¼ ½ϵ; a
w White noise time series
n Collection of all white noise parameters
ρ Free-spectrum parameter used in describing power-spectral-density (ρ2 ¼ haaT i)
A Amplitude of a red noise process
P Power-spectral-density of a red noise process
P̂ Spectral shape of a red noise process obtained by P=A2

0 1
sin ð2πf 1 t1 Þ cos ð2πf 1 t1 Þ  sin ð2πf k t1 Þ cos ð2πf k t1 Þ
B C
B sin ð2πf 1 t2 Þ cos ð2πf 1 t2 Þ  sin ð2πf k t2 Þcos ð2πf k t2 Þ C
B C
F¼B .. .. .. .. .. C; ðA2Þ
B . . . . . C
@ A
sin ð2πf 1 tp Þ cos ð2πf 1 tp Þ    sin ð2πf k tp Þ cos ð2πf k tp Þ

aT ¼ ðasin
1 ; a1 ; …; ak ; ak Þ;
cos sin cos
ðA3Þ 0 1
1 t1 t21 
B C
B 1 t2 t22 C
for tp denoting the last measured TOA, f k denoting the kth B C
M¼B. . .. C: ðA5Þ
considered frequency-bin, and asin and acos referring to the B .. .. . C
@ A
coefficients of sin and cos elements of the F matrix
respectively. 1 tp t2p 
To model the contribution of any linear timing model
parameter to the timing residuals, rT , we use a basis matrix While the first three columns of the design matrix models
known as the timing-design-matrix such that the quadratic spin down of all millisecond pulsars, the
unspecified columns of the matrix are populated with
various timing model contributions specific to each pulsar.
rT ¼ Mϵ; ðA4Þ Moreover, it is often convenient to project the residuals

063008-11
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

onto a subspace orthogonal to the timing model parameters, Pðf k Þ ¼ T obs ρ2k : ðA12Þ
or in other words, to create fitted timing residuals. The so-
called G matrix is a useful matrix obtained via singular-
value-decomposition of the design-matrix constructed to 3. Covariance matrices
perform the fitting: The white noise covariance matrix N plays a key role in
posterior probability calculation of all model parameters.
M ¼ USV T This matrix is modeled as
Gxy ¼ U xy ; ðA6Þ
N ¼ diagðσ w1 ; …; σ wp Þ: ðA13Þ
where x ranges from 1 to p (the number of TOAs) while y
ranges from q to p for q being the total number of the linear Note that the introduction of ECORR white noise param-
timing model parameters. eter will complicate this picture. See chapter 7 of Taylor
To model the contribution of the white noise to the [35] for more details. Furthermore, the red process covari-
timing model residuals, rw , we consider a m × m identity ance matrix is obtained via the discretized form of the
matrix as the basis with the coefficients n such that Wiener-Khinchin theorem

rw ¼ w; ðA7Þ hrred ðti Þrred ðtj Þi ¼ ½FφFT ij ; ðA14Þ

wi ∼ Normal ðmean ¼ 0; scale ¼ σ wi Þ; ðA8Þ φ ¼ diagfPred ðf 1 Þ; Pred ðf 1 Þ;

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi …; Pred ðf k Þ; Pred ðf k Þg; ðA15Þ
σ wi ¼ ef i σ 2i þ eq2i ; ðA9Þ
where Pred is the one-sided PSD of a red noise process and
for σ i being the TOA error of observation i, and ef and eq the diagonal matrix φ is the matrix representation of
being the usual EFAC and EQUAD parameters [20]. Note that PSD.
that the Gaussianity of the white noise is an assumption
included in our all of our models. APPENDIX B: TRUNCATED INVERSE-GAMMA
DISTRIBUTION
2. Noise power-spectral-density modeling To obtain a truncated inverse-gamma distribution, we
In this paper, we only consider one-sided power-spectral- take advantage of inverse-transform sampling method.
densities (PSD). Most commonly for PTA noise analyses, However, first, we need to find a normalization factor,
the PSD is expressed in two ways: Norm, for the truncated inverse-gamma distribution defined
(i) Power law: assuming the PSD to follow a simple between the lower bound ρmin and the upper bound ρmax :
power-law relation with amplitude A and spectral
index γ as well as a reference frequency f ref across ðak · ak Þ
all frequency-bins βk ¼ ; ðB1Þ
2
 3−γ Z   −1
A2 f ρmax β β
PðfÞ ¼ ; ðA10Þ Norm ¼ dρk 2k exp − k
12π 2 f 3 f ref ρmin ρk ρk
βk
PðfÞ ¼ : ðB2Þ
P̂ðfÞ ¼ : ðA11Þ exp − ρmax − exp − ρβmink
βk
A2

The quantity P̂ describes the shape of the spectrum and Note that the above process can be repeated for all
is used in Sec. IV C. frequency bins. Equation (B2) allows for calculation of
(ii) Free-spectrum: allowing the PSD to have independent the cumulative distribution function (CDF), which in turn
amplitude in each frequency-bin with normalization can be used to find a distribution for ρk given a uniform
constant T obs equal to a fixed observation time. The random number U defined between 0 and 1 based on
observation time can either be the baseline of each inverse-transform sampling method. This yields the follow-
pulsar or the baseline of the total PTA experiment. ing as the target distribution for the ρk parameters:

063008-12
EXPLORING THE CAPABILITIES OF GIBBS SAMPLING IN … PHYS. REV. D 108, 063008 (2023)

β
pðρk jak ; r; nÞ ¼ − n h k io : ðB3Þ
ln exp − ρβmink Uð0; 1Þ exp − ρβmax
k
− exp − ρβmink

APPENDIX C: DERIVATION OF THE

NORMALIZATION FACTOR IN EQ. (23) X ðΓIJ A2g P̂g ÞP̂g
¼ ðC6Þ
The choice of normalization in the denominator of k
φI φJ
Eq. (23) enforces the condition that the estimated cross
correlations must yield GWB amplitude if averaged over
many realizations as is shown below: X P̂2g
ptop ¼ ΓIJ A2g ; ðC7Þ
λIJ ¼ ; ðC1Þ k
φI φJ
Norm
X aI aJ
ptop ¼ P̂ ; ðC2Þ P P̂2
φI φJ g which makes Norm ¼ k φI φg J consequently.
k
Furthermore, when the quantity μ̂ of Eq. (12) is used in
X aI aJ estimating the correlations following Eq. (23), the nor-
hptop i ¼ P̂ ðC3Þ
φI φJ g malization need to be reestimated since haI aTJ i ≠ hμ̂I μ̂TJ i for
k
an average over many GWB realizations. The new nor-
X aI aJ malization factor is found to be
¼ P̂ ðC4Þ
k
φI φJ g

X haI aJ iP̂g hμ̂I μ̂TJ i ¼ hΣI FTI D−1 T T −1 T

I FI aI aJ F J ðDJ Þ FJ ΣJ i
T

¼ ðC5Þ ¼ ΣI FTI D−1 −1 T

I FhaI aJ iFJ ðDJ Þ FJ ΣJ : ðC8Þ
T T T
k
φI φJ

[1] M. V. Sazhin, Opportunities for detecting ultralong gravi- [8] G. Agazie et al. (NANOGrav Collaboration), The NANO-
tational waves, Sov. Phys. J. 22, 36 (1978), https://ui.adsabs Grav 15 yr data set: Evidence for a gravitational-wave
.harvard.edu/abs/1978SvA....22...36S/abstract. background, Astrophys. J. Lett. 951, L8 (2023).
[2] S. Detweiler, Pulsar timing measurements and the search for [9] J. Antoniadis et al., The second data release from the
gravitational waves, Astrophys. J. Lett. 234, 1100 (1979). European pulsar timing array III. Search for gravitational
[3] R. W. Hellings and G. S. Downs, Upper limits on the wave signals, arXiv:2306.16214.
isotropic gravitational radiation background from pulsar [10] D. J. Reardon et al., Search for an isotropic gravitational-
timing analysis, Astrophys. J. Lett. 265, L39 (1983). wave background with the parkes pulsar timing array,
[4] Z. Arzoumanian et al. (The NANOGrav Collaboration), The Astrophys. J. Lett. 951, L6 (2023).
nanograv 12.5 yr dataset: Search for an isotropic stochastic [11] H. Xu et al., Searching for the nano-hertz stochastic
gravitational-wave background, Astrophys. J. Lett. 905, L34 gravitational wave background with the Chinese pulsar
(2020). timing array data release I, Res. Astron. Astrophys. 23,
[5] J. Antoniadis et al., The international pulsar timing array 075024 (2023).
second data release: Search for an isotropic gravitational [12] X. Siemens, J. Ellis, F. Jenet, and J. D. Romano, The
wave background, Mon. Not. R. Astron. Soc. 510, 4873 stochastic background: Scaling laws and time to detection
(2022). for pulsar timing arrays, Classical Quantum Gravity 30,
[6] S. Chen et al., Common-red-signal analysis with 24-yr high- 224015 (2013).
precision timing of the European Pulsar Timing Array: [13] L. Lentati, P. Alexander, M. P. Hobson, S. Taylor, J. Gair,
Inferences in the stochastic gravitational-wave background S. T. Balan, and R. van Haasteren, Hyper-efficient model-
search, Mon. Not. R. Astron. Soc. 508, 4970 (2021). independent Bayesian method for the analysis of pulsar
[7] B. Goncharov et al., On the evidence for a common- timing data, Phys. Rev. D 87, 104021 (2013).
spectrum process in the search for the nanohertz gravita- [14] R. van Haasteren and M. Vallisneri, New advances in the
tional-wave background with the parkes pulsar timing array, Gaussian-process approach to pulsar-timing data analysis,
Astrophys. J. Lett. 917, L19 (2021). Phys. Rev. D 90, 104012 (2014).

063008-13
NIMA LAAL et al. PHYS. REV. D 108, 063008 (2023)

[15] G. E. Freedman, A. D. Johnson, R. van Haasteren, and S. J. background in pulsar timing arrays, Phys. Rev. D 98,
Vigeland, Efficient gravitational wave searches with pulsar 044003 (2018).
timing arrays using Hamiltonian Monte Carlo, Phys. Rev. D [26] R. Kumar, C. Carroll, A. Hartikainen, and O. Martin, ArviZ a
107, 043013 (2023). unified library for exploratory analysis of Bayesian models
[16] S. Hourihane, P. Meyers, A. Johnson, K. Chatziioannou, and in python, J. Open Source Softwaare 4, 1143 (2019).
M. Vallisneri, Accurate characterization of the stochastic [27] J. Ellis and R. van Haasteren, jellis18/ptmcmcsampler: Official
gravitational-wave background with pulsar timing arrays by release (2017), https://doi.org/10.5281/zenodo.1037579.
likelihood reweighting, Phys. Rev. D 107, 084045 (2023). [28] E. Hellinger, Neue begründung der theorie quadratischer
[17] J. Sun, P. T. Baker, A. D. Johnson, D. R. Madison, and X. formen von unendlichvielen veränderlichen., J. Reine
Siemens, Implementation of an efficient Bayesian search for Angew. Math. 1909, 210 (1909).
gravitational wave bursts with memory in pulsar timing [29] J. A. Ellis, M. Vallisneri, S. R. Taylor, and P. T. Baker,
array data, Astrophys. J. 951, 121 (2023). Enterprise: Enhanced numerical toolbox enabling a robust
[18] S. R. Taylor, J. Simon, L. Schult, N. Pol, and W. G. Lamb, A pulsar inference suite, Zenodo (2023), https://doi.org/10
parallelized Bayesian approach to accelerated gravitational- .5281/zenodo.4059815.
wave background characterization, Phys. Rev. D 105, [30] S. R. Taylor, P. T. Baker, J. S. Hazboun, J. Simon, and S. J.
084049 (2022). Vigeland, enterprise-extensions (2021), v2.4.3, https://
[19] W. G. Lamb, S. R. Taylor, and R. van Haasteren, The need github.com/nanograv/enterprise_extensions.
for speed: Rapid refitting techniques for Bayesian spectral [31] J. D. Hunter, Matplotlib: A 2d graphics environment, Comput.
characterization of the gravitational wave background using Sci. Eng. 9, 90 (2007).
PTAs, arXiv:2303.15442. [32] P. T. Inc., Collaborative data science (2015), https://plotly
[20] J. S. Hazboun et al., The NANOGrav 11 yr data set: .com/python/.
Evolution of gravitational-wave background statistics, As- [33] Z. Arzoumanian et al. (NANOGrav Collaboration), The
trophys. J. 890, 108 (2020). NANOGrav nine-year data set: Limits on the isotropic
[21] M. Anholm, S. Ballmer, J. D. E. Creighton, L. R. Price, and stochastic gravitational wave background, Astrophys. J.
X. Siemens, Optimal strategies for gravitational wave Lett. 821, 13 (2016).
stochastic background searches in pulsar timing data, Phys. [34] J. S. Hazboun, J. D. Romano, and T. L. Smith, Realistic
Rev. D 79, 084030 (2009). sensitivity curves for pulsar timing arrays, Phys. Rev. D 100,
[22] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. 104028 (2019).
Teller, and E. Teller, Equation of state calculations by fast [35] S. R. Taylor, The nanohertz gravitational wave astronomer,
computing machines, J. Chem. Phys. 21, 1087 (1953). arXiv:2105.13270.
[23] S. J. Chamberlin, J. D. E. Creighton, X. Siemens, P. Demorest, [36] D. J. Reardon et al., The gravitational-wave background
J. Ellis, L. R. Price, and J. D. Romano, Time-domain imple- null hypothesis: Characterizing noise in millisecond pulsar
mentation of the optimal cross-correlation statistic for sto- arrival times with the parkes pulsar timing array, Astrophys.
chastic gravitational-wave background searches in pulsar J. Lett. 951, L7 (2023).
timing data, Phys. Rev. D 91, 044048 (2015). [37] B. Goncharov, D. J. Reardon, R. M. Shannon, X.-J. Zhu, E.
[24] R. Case, N. Laal, J. D. Romano, and X. Siemens (to be Thrane, M. Bailes, N. D. R. Bhat, S. Dai, G. Hobbs, M. Kerr,
published). R. N. Manchester, S. Osłowski, A. Parthasarathy, C. J. Russell,
[25] S. J. Vigeland, K. Islo, S. R. Taylor, and J. A. Ellis, Noise- R. Spiewak, N. Thyagarajan, and J. B. Wang, Identifying and
marginalized optimal statistic: A robust hybrid frequentist- mitigating noise sources in precision pulsar timing data sets,
Bayesian statistic for the stochastic gravitational-wave Mon. Not. R. Astron. Soc. 502, 478 (2020).

063008-14