Arxiv:2106.15290V1 [Gr-Qc] 29 Jun 2021 the Aim of This Work Is to Account Fully for the Contri- Diﬀerent Types of Object Following the Same Geodesic

Constraints on Equivalence Principle Violation from Gamma Ray Bursts

D. J. Bartlett,1, ∗ D. Bergsdal,2, 3 H. Desmond,1 P. G. Ferreira,1 and J. Jasche3 1Astrophysics, University of Oxford, Denys Wilkinson Building, Keble Road, Oxford, OX1 3RH, UK 2Particle and Astroparticle Physics, Department of Physics, KTH Royal Institute of Technology, SE 106 91, Stockholm, Sweden 3The Oskar Klein Centre, Department of Physics, Stockholm University, AlbaNova University Centre, SE 106 91 Stockholm, Sweden Theories of gravity that obey the Weak Equivalence Principle have the same Parametrised Post- Newtonian parameter γ for all particles at all energies. The large Shapiro time delays of extragalactic sources allow us to put tight constraints on differences in γ between photons of different frequencies from spectral lag data, since a non-zero ∆γ would result in a frequency-dependent arrival time. The majority of previous constraints have assumed that the Shapiro time delay is dominated by a few local massive objects, although this is a poor approximation for distant sources. In this work we consider the cosmological context of these sources by developing a source-by-source, Monte Carlo-based forward model for the Shapiro time delays by combining constrained realisations of the local density field using the BORG algorithm with unconstrained large-scale modes. Propagating uncertainties in the density field reconstruction and marginalising over an empirical model describing other contributions to the time delay, we use spectral lag data of Gamma Ray Bursts from the BATSE satellite to constrain ∆γ < 3.4 × 10−15 at 1σ confidence between photon energies of 25 keV and 325 keV.

I. INTRODUCTION they experience time delays identically. This can be achieved either when the source emits light at varying A basic feature of metric theories of gravity is that frequencies which can be independently measured [6], or objects on either timelike or null geodesics experience when it emits other types of energy in addition to light, apparent time delays due to motion through regions of such as gravitational waves [7]. Within the PPN frame- varying spacetime curvature, caused by the difference work, this constrains the difference between their γ fac- between proper and observer time induced by a grav- tors. itational field. This effect, first derived by Shapiro in In general metric theories the time delay is propor- 1964 [1], has now been accurately measured in the Solar tional to the integral of the fluctuation in Newtonian po- System: the best constraint on the fractional deviation tential φ along the line of sight to the source; within the of the time delay from the prediction of General Rel- PPN framework it has a prefactor 1 + γ where γ = 1 ativity (GR) is currently at the 10−5 level, using data in GR. Predicting time delays is therefore equivalent to from the radio link with the Cassini spacecraft [2]. In determining δφ. In typical analyses, δφ φ is modelled the Parametrised Post-Newtonian (PPN) framework the as arising from one or a few isolated sources' near the line time delay is governed by the light-bending parameter γ of sight that are believed to be predominantly respon- [3]. For sources at greater distance, however, the lack of sible for sourcing the potential. However, as φ 1/r knowledge of the time of emission of a signal precludes one should expect it to be sensitive to the distribution∼ of direct measurement of the delay. distant sources, and indeed the contribution to φ from a Nevertheless, the time delay effect has found use in spherical shell at distance r from a test point diverges as testing various aspects of the theory of gravity and the r . This casts doubt on the multiple point masses standard model of cosmology. This is done by comparing approximation→ ∞ [8, 9]. time delays, either between nearby geodesics or between arXiv:2106.15290v1 [gr-qc] 29 Jun 2021 The aim of this work is to account fully for the contri- different types of object following the same geodesic. The butions to the time delay from all mass in the non-linear former may be achieved by comparing the time of recep- cosmological density field. We derive the contribution tion of photons originating from a common source but from local structures using constrained density fields gen- traversing different paths due to gravitational lensing. erated by the BORG algorithm [10–14] and combine this When the source is a time-varying quasar and the lens with an unconstrained contribution from distant sources a single massive elliptical galaxy or cluster, this method to produce a Monte-Carlo based source-by-source for- has been used to constrain the Hubble parameter, H0, ward model for the expected Shapiro time delay. on which time delay distances depend [4, 5]. Conversely, We apply our method to high-energy astrophysical comparing time delays between different objects travel- sources that have previously been used to test the ling along the same geodesic allows the Weak Equivalence Equivalence Principle for photons of varying wavelength. Principle (WEP) to be tested by investigating whether Marginalising over uncertainties in the density field reconstruction and parameters describing non-Equivalence- Principle-violating contributions, we compare our pre- ∗ [email protected] dictions to the observations via a Markov Chain Monte 2

Carlo (MCMC) algorithm. We find our constraints are which can be parametrised as a difference in the PPN pa- 40 times tighter than literature results, illustrating the rameter γ between two wavelengths. Thus by comparing benefit∼ of using complete mass distributions when study- the measured time delays with the line-of-sight integral ing non-local relativistic effects such as time delays in a in Equation 2, one can put constraints on how γ changes cosmological setting. with photon frequency, and thus on any violation of the In Section II we discuss time delays in the context of WEP. metric theories of gravity, including the importance of long wavelength modes in calculating the Shapiro time delay. We describe the Gamma Ray Burst (GRB) obser- B. Time delay angular power spectrum vations in Section III and detail our inference methods in Section IV. The results are presented in Section V. The angular power spectrum of the gravitational time A discussion of systematic uncertainties, future applica- delay is [8] tions of our methods and a comparison to the literature Z Z rs 2 are presented in Section VI. We conclude in Section VII. 2 2 C` = dk k Pφ (k) dr D (r) j` (kr) , (4) π 0

II. TIME DELAY DIFFERENCES IN METRIC such that the mean-squared time delay is THEORIES OF GRAVITY X 2` + 1 t2 = C , (5) grav 4π ` A. Equivalence principle violation `

where Pφ is the power spectrum of the potential. For Consider a perturbed FRW-like metric in the Newto- k 1/rs, the spherical Bessel function can be expanded nian gauge, as 2 2 2 2 √π ds = (1 + 2γδφ) dt + a (t) (1 2δφ) dr , (1) j (x) = x` + x2+` , (6) − − ` `+1 3 2 Γ ` + 2 O where we have introduced the PPN parameter γ to allow for deviations from General Relativity (where γ = 1). so that For a massless particle, the gravitational time delay to 4 √π Z dk C = [T (k)]2 A kns−1k2` a source at distance rs is given by (1 + γ) tgrav, where ` 3 `+1 3 φ s π 2 Γ ` + 2 k (7) Z rs Z 2 ! ` 2+2` tgrav = dr δφ0(r)D(r), (2) dr D (r) r + k , − 0 O where D(r) is the linear growth factor, δφ is the poten- 0 where Tφ is the potential transfer function and we assume tial fluctuation evaluated using the present-day matter a nearly scale-invariant primordial power spectrum field and we have used the weak field limit δφ 1. For P ns−1 a set of point masses φ = i GMi/ri, but in a general R (k) = Ask . (8) density field δφ must be found by solving the Poisson P equation On super-Hubble scales, Tφ (k) constant, so the smallest power of k in the integral has∼ an exponent n 2+2`. s − 2δφ(r) = 4πGδρ(r), (3) For ` 1, this is ns, which, given that ns 0.97 ∇ [15], means≥ that, for≥ ` 1, the integral does not∼ di- ≥ where δρ is the total matter density fluctuation. This is verge at small k. However for ` = 0, the exponent is most readily solved in Fourier space. Thus determining ns 2 < 1, and so the integral diverges. We therefore − − time delays amounts to mapping out the three dimen- see that the monopole diverges due to the contribution sional density field to at least the redshift of the source. from large scales. This is not to say that there is an We describe how we do this in Section IV A. infinite time delay for a given universe, but that its vari- There exist several formulations of the Equivalence ance across all possible universes is infinite. Thus when Principle, of varying strength. The WEP, our focus here, we forward model the time delays in Section IV, we will states that all freely-falling test objects follow the same only be able to predict angular fluctuations about the trajectories given the same initial conditions, irrespec- mean Shapiro time delay at a given redshift, and not its tive of their composition or structure. This requires that absolute value. all objects from a given source experience the same time delay, regardless of their composition or energy. Multiple astrophysical sources have been detected with C. Other contributions to the time delay precise timings of photons of different wavelengths. One contribution to the difference in the time of reception of We suppose that the time delay between frequencies νi these photons is a difference in gravitational time delay, and νj for a source at location r can be written as the 3 sum of the following independent terms The physical mechanisms which result in the GRB spectral lag are unknown; the search for the nature of −2 −2 ∆tij (r) = ∆γijtgrav + A (r) ν ν + Bij. (9) spectral lag is an ongoing research topic where most i − j of the focus lies on investigating possible effects at the Besides the Equivalence Principle-violating term, our source [see e.g. 20–26]. It has for example been shown time delay contains two other contributions. The first that the effects of spectral lag can be recreated from of these (containing A (r)) describes the dispersion due simple source models utilising rapid bulk acceleration on to electrons. The second (Bij) represents the combina- relativistic jet shells [26]. Therefore, when modelling the tion of an intrinsic time delay at the source and the in- intrinsic contribution to these time delays in Section IV E strument response, which we assume is independent of we will have to rely on empirical models, as opposed to observed angle and redshift. the ideal case where we can calibrate our noise model The mean of A depends on the temporal evolution of with simulations [27]. the comoving electron density,n ¯e,c, as

2 Z z(r) 0 ¯ e 1 dz 0 0 IV. METHODS A (r) = 0 (1 + z )n ¯e,c (z ) , (10) 2πme 4π0 0 H (z ) In this section we use the large-scale structure infor- and fluctuations about this depend on fluctuations in the mation described in Section IV A to produce a source- electron density and other relativistic effects [16]. The by-source probabilistic forward model for the expected comoving electron density can be modelled as Shapiro time delay difference for a given ∆γij. Combin- 2 ing this with an empirical model describing other contri- 3H0 Ωb xe (z) (1 + xH (z)) n¯e,c (z) = . (11) butions to the measured time delays (noise) outlined in 8πGmp 2 Section II, we calculate the likelihood function and constrain ∆γij and the parameters describing the noise via where xe is the free electron fraction, xH is the hydrogen a Markov Chain Monte Carlo (MCMC) algorithm. The mass fraction, Ωb is the baryon density fraction, and mp parameters which are fixed in this section are summarised is the proton mass. Assuming xe = 1 and xH = 0.75, we in Table I. find the contribution from the electron plasma to be

"Z z(r) 0 # −2 −2 dz 0 A. Bayesian large scale structure inference A (r) νi νj = 0 (1 + z ) − 0 E (z ) × " # (12) E −2 E −2 This work builds upon previous results of applying the i j 5.6 10−17 s, Bayesian Origin Reconstruction from Galaxies (BORG) keV − keV × × algorithm to the data of the SDSS-III/BOSS galaxy compilation [see e.g. 10–14]. where E (z) H (z) /H . Since in this work we will be ≡ 0 The BORG algorithm is a fully probabilistic infer- considering Gamma Ray Bursts, we will find that this ence method aimed at reconstructing matter fields from contribution is negligible for the probed frequencies, and galaxy observations. This algorithm incorporates a phys- thus we will neglect it in our analysis. If we were to ical model for gravitational structure formation, enabling consider radio bursts we would need to consider this term, inference of the three-dimensional density field and the as we would expect contributions of (1 s). O corresponding initial conditions at an earlier epoch from present observations. Specifically the algorithm explores a large-scale struc- III. OBSERVATIONAL DATA ture posterior distribution consisting of a Gaussian prior for the initial density field at an initial cosmic scale fac- The spectral lag data we use is a sample of the BATSE tor of a = 10−3, linked to a Poissonian model of galaxy detections of GRBs catalogued by Hakkila et al. [17]. We formation at a scale factor a = 1 via second order La- make use of a set of sources compiled by Yu et al. [18], grangian perturbation theory [for details see 11]. The where the four energy bins considered are sensitive to the model accurately describes one-, two- and three-point ranges Ch1: 25-60 keV, Ch2: 60-110 keV, Ch3: 110-325 functions and represents very well higher-order statistics, keV and Ch4: >325 keV. These four bins result in up to as was calculated by e.g. [28–32]. Thus BORG naturally 6 time delay pairs per source, ∆tij, where i, j label the accounts for the filamentary structure of the cosmic web channels used. Without loss of generality we define i > j. typically associated with higher-order statistics induced In the cases where no ∆tij is recorded for a source due by nonlinear gravitational structure formation processes. to low signal-to-noise, we ignore that particular pair but The posterior distribution also accounts for the system- still consider the others. These sources also have pseudo- atic and stochastic uncertainties encountered in cosmo- redshifts calculated using the spectral peak energy-peak logical surveys, including survey geometries, selection ef- luminosity relation [19]. fects and noise. 4

TABLE I. Parameters used to constrain the Equivalence Principle-violating contribution to the time delay. Above the horizontal line are the parameters used to forward model the time delay and below are the parameters passed to multinest in the MCMC analysis. In the ﬁnal column we give the value chosen for each parameter, although we show in Section VI that our results are unchanged for reasonable alternative values. Parameter Description Value −1 Lbox / h Mpc Side length of box used to reconstruct local density ﬁeld. 4000

Nbox Number of grid points per side of box used to reconstruct local density ﬁeld. 256 −1 −5 k0 / Mpc Minimum wavenumber used to calculate long wavelength contribution. 10

`max Maximum multipole used to compute time delay from angular power spectrum. 2000

Nside Resolution of HEALPix map used to calculate the monopole. 64

Nmon Number of redshifts between z = 0.1 and 2 used to calculate the monopole. 20

Nz,R Number of redshift points used to calculate the second term in Equation 18. 512

Nz,L Number of redshift points used to calculate long-wavelength time delay contributions. 1024

Nk,L Number of wavenumber points used to calculate unconstrained time delay contributions. 512 4 NMC Number of Monte Carlo runs to get the distribution of time delays for the template signal. 10

Nbin Number of redshift bins to determine redshift evolution of noise model. 5 n live points Number of live points used in multinest sampling. 800 importance nested sampling Whether to use importance nested sampling with multinest. True multimodal Whether to allow mode separation in multinest. True evidence tolerance Evidence tolerance for multinest. 0.5 sampling efficiency Sampling eﬃciency for multinest. 0.8 const efficiency mode Whether to use constant eﬃciency mode in multinest. False

Applying the BORG algorithm to the SDSS-III/BOSS time delay into two parts galaxy sample [33, 34], three-dimensional matter density fields have been inferred on a cubic Cartesian grid of tgrav (r) = t¯grav (r) + δtgrav (r) , (13) side length of 4000 h−1Mpc consisting of 2563 equidistant ¯ voxels. This results in a grid resolution of 15 h−1Mpc. where tgrav (r) is the mean Shapiro time delay across all The inference assumes a standard ΛCDM∼ model with angles at a given r (i.e. the monopole), and δtgrav gives the fluctuation about this. These fluctuations can be the following set of cosmological parameters: Ωm = decomposed into the sum of three terms 0.2889, ΩΛ = 0.7111, Ωb = 0.048597, h = 0.6774, σ8 = 0.8159, n = 0.9667 [15]. s δt (r) = δt(L) (r) + δt(R) (r) + δt(S) (r) (14) We also consider the 2M++ particle-mesh reconstruc- grav grav grav grav tion of BORG [14, 35], which again has 2563 equidistant corresponding to the long wavelength (L; k < kmin), re- voxels, but a smaller side length of 677.7 Mpc/h and thus solved (R; kmin k < kmax) and short wavelength (S; a finer spatial resolution.1 In Section IV B we compare ≤ k kmax) contributions respectively. The ‘resolved’ con- how well we can determine the Shapiro time delay using tribution≥ can be determined using the inferred matter the SDSS-III/BOSS vs 2M++ reconstructions. fields from BORG. To determine the resolution and box size required to accurately reconstruct the time delay fluctuation, we B. Calculating the Shapiro time delay compare the results of calculating the angular power spectrum with a finite resolution and box size to the As discussed in Section II B, the variance of the mean continuous, infinite volume case. Obviously the latter Shapiro time delay at a given redshift across all possible is impossible in practice, but we approximate this limit −5 −1 universes diverges. We can rephrase our problem so that by choosing a sufficiently small minimum (10 Mpc ) −1 this is not an issue. Let us decompose the gravitational and large maximum (100 Mpc ) k. To approximate the finite volume result, we consider a box length Lbox with Nbox grid cells along each side, so we have minimum and maximum (non-zero) k of

1 A set of 100 dark matter-only simulations based on the 2M++ 2π π (Nbox 1) reconstruction, dubbed CSiBORG [36], is also available. This kmin = , kmax = − . (15) could be used to access smaller scales because the initial condi- Lbox Lbox tions are augmented with white noise. We find in Section IV B that these scales are unimportant however, so we do not consider Note that we do not include a factor of √3 in kmax as the these simulations further. sphere in k-space of radius kmax does not fully fit inside 5 the first Brillouin zone if this is included. To mimic using the monopole at some intermediate redshift we linearly this box, we top-hat filter the potential power spectrum, interpolate between the sampled redshifts. The ensem- allowing modes between kmin and kmax. We calculate ble mean of the resulting time delay fluctuation map is two quantities: (i) the C`’s themselves, and (ii) the RMS plotted in Figure 2. One can see that the typical fluctua- fluctation in the time delay using Equation 5, where we tion from the resolved contribution is larger than 1011 s ∼ find `max = 2000 to be more than sufficient (halving this and thus, given that the measured time delays are typi- to `max = 1000 does not change the final result). cally (0.1 s), one would expect ∆γij to be smaller than In Figure 1 we compute the C ’s for a source at redshift 10−O12. | | ` ∼ z = 0.10 using the Core Cosmology Library [37]. We If the source is outside the volume, we now add an un- compare the results from the continuous case with the constrained contribution to the fluctuation drawn from a −1 SDSS-III/BOSS (Lbox = 4000h Mpc, Nbox = 256) and Gaussian of width given by Equation 5, but only using −1 2M++ (Lbox = 677.7h Mpc, Nbox = 256) reconstruc- k modes accessible to BORG (Equation 15) and inte- tions. We note that the maximum distance a source can grating between rb and rs using Nz,R = 512 intermediate be from the observer in the latter two cases is Lbox√3/2, points. This corresponds to fluctuations due to the sec- which is 5.17 Gpc for the SDSS-III/BOSS and 876 Mpc ond term in Equation 18. We find `max = 2000 to be for the 2M++ reconstruction, corresponding to redshifts sufficient for this calculation. of 1.9 and 0.21 respectively. Thus our test source at To marginalise over the uncertainties in the pseudo- z = 0.10 is within both boxes. redshifts, angular position and unconstrained regions, we The continuous case gives the same result as SDSS- use simulation-based Bayesian forward modelling to cre- 5 III/BOSS to 4 parts in 10 as the C`’s only disagree at ate predictions from the statistical models. In particular, ∼ ` 1 while the result is dominated by low `. The lack of we take N = 104 Monte Carlo draws from the input MC power at low k results in the 2M++ box underestimating distributions to build the likelihood. For each iteration, the low ` contribution, and thus the RMS time delay. we draw an angular location from Gaussian distributions The points of disagreement are as expected, since for this characterised by the positional uncertainty given in [18] redshift, using the Limber approximation [38] ` kr, we and a redshift from a two-sided Gaussian using the upper ∼ find that kmin and kmax correspond to multipoles and lower uncertainties determined by [19]. We then ran- domly choose a density field from the converged part of SDSS−III/BOSS − ` 0.46, `SDSS III/BOSS 59 (16) the BORG MCMC chain to calculate the time delay for min ∼ max ∼ `2M++ 2.7, `2M++ 346, (17) a source at this position. From Equation 9, we see that min ∼ max ∼ the gravitational contribution to the time delay between which is approximately where we see the results diverg- any two frequencies is proportional to ∆γij. We there- ing from the continuous case. Since the total time delay fore only need to run this Monte Carlo procedure once is dominated by low `, we conclude that a large box size per source to construct a template signal with ∆γij = 1 is far more important than a fine resolution, justifying since this can be trivially reintroduced as a scaling factor our choice of the SDSS-III/BOSS over 2M++ reconstruc- later. tion. For simplicity, we will henceforth refer to the SDSS- III/BOSS BORG reconstruction simply as ‘BORG’. Nonetheless, for sources at higher redshift, the density C. Monte Carlo modelling field not contained within the inferred volume contributes significantly to the gravitational time delay. For sources Now that we have NMC samples per source of our prob- outside of the BORG volume, Equation 2 can be split (R) into a contribution from the observer to the edge of the abilistic model for δtgrav, we must convert these samples into a distribution to use in our likelihood analysis. box, rb, and then from rb to the source: We model the samples as a Gaussian mixture model Z rb Z rs (GMM) [41], where the likelihood function for some r r tgrav = dr δφ0( )D(r) dr δφ0( )D(r). (18) source s is − 0 − rb

(α) To convert the first term (or Equation 2 for sources inside X ws δt(R) ∆γ = 1 = the box) to a fluctuation, we must subtract the mean fluc- s grav ij (α) L | √2πτs × tuation at the upper limit of the integral. We do this by α 2 2  (R) (α)  (19) computing the time delays on HEALPix [39, 40] maps δtgrav λs with Nside = 64 at Nmon = 20 logarithmically-spaced exp  −  , − (α)2  redshifts between z = 0.1 and z = 2. We then compute 2τs the monopoles of these maps with healpy. To determine where X w(α) = 1, w(α) 0, (20) 2 http://healpix.sf.net s s ≥ α 6

z = 0.10 z = 0.10 0.0 1023

0.2 1020 − 2 ` s 0.4 / 17

/C − ` 10 ` C C π +1 4 ` ∆ 0.6 2 1014 − Continuous Continuous 0.8 SDSS-III/BOSS − SDSS-III/BOSS 1011 2M++ 2M++ 1.0 − 100 101 102 103 100 101 102 103 ` `

FIG. 1. The predicted Shapiro time delay fluctuation angular power spectrum for a source at z = 0.10 calculated with the Core Cosmology Library. For the SDSS-III/BOSS and 2M++ calculations, we only include k modes within the range sampled in the simulated volume due to the finite box length and resolution. The right hand panel shows the fractional difference between these and the continuous case. Using SDSS-III/BOSS gives practically the same result for the time delay as the continuous case, whereas 2M++ underestimates this by ∼ 32 per cent. The result is driven by the smallest `, making large box size far more important than high spatial resolution.

maximum likelihood estimate ˆ. Independent Gaussians are obtained for each source.L To account for a diﬀerent ∆γij, we must transform the means and widths of the Gaussians in the GMM as z = 0.1

λ˜(α) = ∆γ λ(α), τ˜(α) = ∆γ τ (α). (22) sij ij s sij | ij| s

-5.00 5.00 (R) 11 δtgrav / 10 s D. Adding larger scale modes

FIG. 2. Mollweide projection in equatorial coordinates of the Thus far we have only calculated the ‘resolved’ contri- ensemble mean of the time delay ﬂuctuations at z = 0.1 from 11 bution to the Shapiro time delay ﬂuctuations. Since the resolved wavelengths. The typical scale is ∼ 10 s which, since the majority of observed time delays are O(0.1 s), in- large-k contributions are negligible, we do not consider < −12 these further. To incorporate the long wavelength modes, dicates that constraints at least as tight as |∆γij | ∼ 10 should be possible. we evaluate Equation 5 for each source, where we inte- −5 −1 grate between k0 = 10 Mpc and kmin (Equation 15) using Nk,L = 512 intermediate points. We integrate up to and the sum runs over the Gaussians. The number of the three-sigma redshift in the pseduo-redshift distribu- Gaussians is chosen to minimise the Bayesian Informa- tion with Nz,L = 1024 points and again use `max = 2000. tion Criterion (BIC) For source s this gives the width of the distribution of long-wavelength contributions, ξs, which we assume to BIC = log 2 log ˆ, (21) be Gaussian distributed since the density should also be K N − L Gaussian distributed on large scales. We must convolve for model parameters, = N data points, and this long-wavelength distribution with our GMM to get K N MC 7

(0) the total likelihood of δtgrav, each bin have the sameµ ˜ij (z). Since we have now de- coupled ∆γij from the monopole term, our results are (α) X ws driven by the angular variation in the time delay and not (δt ) = Ls grav r × the absolute value. α (α)2 ˜2 2π τ˜sij + ξsij Finally, we must incorporate the errors on the measured ∆t . If these were independent of ∆t then this  2  (23) ij ij ˜(α) (β) δtgrav λs could be incorporated into σ , however longer lags  −  { ij } exp   , tend to have larger errors [17]. We thus use the quoted − 2 τ˜(α)2 + ξ˜2 sij sij measurement errors for each source and time delay pair, εsij, so that the likelihood for an observed time delay is where (α) (β) ˜ X ws ωij ξsij ∆γijξs. (24) (∆t ) = ≡ Ls ij r × αβ (α)2 ˜2 (β)2 2 2π τ˜sij + ξsij + σij + εsij

E. Modelling the noise  2  ˜(α) (β) ∆tij λsij µ˜ij (zs) exp  − −  . Given that we are neglecting the contribution from the  (α) (β)  −2 τ˜ 2 + ξ˜2 + σ 2 + ε2 electron plasma in Equation 9, to determine the likeli- sij sij ij sij hood of an observed time delay, we must ﬁnally convolve (30) Eq. 23 with the likelihoods for the intrinsic contribution and monopole terms. We assume that for each pair of F. Likelihood model frequencies the distribution of Bij can be written as the sum of NG Gaussians, Treating each source as independent, the likelihood for  2  (β) (β) Bij µ our dataset is then X ωij ij (B ) = exp  −  . (25) D ij (β)  2  Y L √2πσ − (β) ij ( θ) = s (∆tij) , (31) β ij 2σij L D| L s where (0) (β) (β) (β) where θ ∆γij, µ˜ij (z) , ∆µij , σij , ωij . Finally, X ≡ { } ω(β) = 1, ω(β) 0, (26) given some prior on θ, P (θ), we use Bayes’ theorem to ij ij ≥ β obtain the posterior

ij ( θ) P (θ) and β 0, 1,...,NG 1 . Without loss of generality, ij (θ ) = L D| , (32) ∈ { − } (β) (β+1) P |D ( ) we define the Gaussians such that ω ω . Zij D ij ≥ ij For non-zero ∆γij, by binning the sources by redshift, where the evidence, ij( ), is the constant probability one can constrain t0 for each bin and a universal set of pa- of the data for any θZ. D (β) (β) (β) rameters describing B , ω , µ , σ . However we We treat each pair of frequencies as independent and ij { ij ij ij } will find that ∆γij is consistent with zero, which leaves t0 separately derive posteriors on ∆γij and the noise pa- completely unconstrained and thus a different approach rameters using pymultinest [42–44] with the settings is required. This is unfortunate since t0 is the only term given in Table I. The priors for all inferred parameters, that couples the different frequency pairs, so now we must given in Table II, are found to be much broader than the consider each pair separately. posteriors for Gaussians with non-zero weights. (β) We therefore do not infer µij , but rather We run our inference with NG in the range 1–5 (in- { } clusive) to determine the posterior distributions for each (0) (0) noise model. To determine the appropriate NG, we com- µ˜ij (z) µij + ∆γijt0 (z) , (27) ≡ pare the models by calculating the BIC (Equation 21); and ∆µ(β) , where the best-fitting model minimises this statistic. Since we { ij } now have access to the full posterior, as suggested by ˜ ∆µ(β) µ(β) µ(0), (28) Handley and Lemos [45] we set = dij, where the ij ≡ ij − ij Bayesian model dimensionality (BMD)K is defined to be such that Z 2 d˜ij ij (θ ) (β) (0) (β) ij (θ ) log P |D dθ, (33) µ˜ij (z) =µ ˜ij (z) + ∆µij . (29) 2 ≡ P |D P (θ)

We bin our sources into Nbin = 5 linearly spaced red- and is computed using the anesthetic software package shift bins between the minimum and maximum pseudo- [46]. Although not as “Bayesian” as comparing the evi- redshifts in the sample and assume that all sources within dence, we prefer this statistic due to its insensitivity to 8

TABLE II. Model parameters describing the predicted signal TABLE III. 1σ constraints on ∆γij for the different time delay and the empirical noise model. All priors are uniform in the pairs. range given. 15 (i, j) Constraint on ∆γij × 10 Parameter Prior (2, 1) 0.66 −13 −13 ∆γij −10 , 10 (3, 1) 1.09 (0) µ˜ij (z) / s [−10, 10] (3, 2) 0.73 (β) ∆µij / s [−10, 10] (4, 1) 3.39 (β) σij / s [0, 10] (4, 2) 1.58 (β) P (β) (β) (β+1) ωij [0, 1], β ωij = 1, ωij ≥ ωij (4, 3) 0.78 the prior. We have deliberately set our priors very wide, in Figure 6 and indeed find that for all pairs any variation (0) making ratios of ij difficult to interpret. Z in µ˜ij is comparable to its uncertainty, indicating that our{ assumption} is reasonable. Furthermore, the time delay pair giving the weakest constraint ((i, j) = (4, 1)) has V. RESULTS noise parameters most suggestive of redshift evolution, as would be expected. In Figure 3 we plot the BIC as a function of number of Gaussian components in the noise model, NG, where we set the number of model parameters equal to the BMD. VI. DISCUSSION We find that the change in BIC is large when we use non- optimal NG, indicating that the best NG is unambiguous. A. Systematic uncertainties The same trend is found when we fix ∆γij = 0 (i.e. fit the observations with just the noise model). For N = 5, G Our probabilistic forward model is designed to prop- we find that the weight of the lowest-weighted Gaussian, (4) (4) agate uncertainties in the source localisation and den- ωij , is extremely small: ωij < 0.03 at 68% confidence sity field reconstruction, which we marginalise over via for all frequency pairs. It is unsurprising then that the 4 a MCMC algorithm. We use NMC = 10 Monte Carlo BIC increases for this case, since we are adding three new samples per source to estimate the likelihood, but we but negligible parameters to the noise model. Similar also check that this is sufficiently large to fully sample results are found if we compare models using the Bayes the distributions we wish to marginalise over. Running factor; the only discrepancy is for (i, j) = (4, 1) where the 3 the inference using NMC = 5 10 yields identical con- BIC prefers NG = 3 whereas the Bayes factor suggests × straints on ∆γij and the noise parameters as the fiducial that NG = 4 is optimal. The constraints are similar for 4 case of NMC = 10 , indicating that the number of sam- both models. ples is appropriate. For the optimal NG we plot the marginalised one- In order to measure the fluctuations in the Shapiro dimensional posteriors of ∆γij in Figure 4, and in Fig- time delay using BORG, we had to subtract the ure 5 we show the corner plot for the time delay pair with monopole at a given redshift. This involved sampling the weakest constraints on ∆γij. Nmon = 20 logarithmically spaced redshifts, where for For all time delay pairs we find ∆γ to be consistent ij each redshift we computed the monopole using a Nside = with zero. The 1σ constraints (defined to be half the 64 HEALPix map. In Figure 7 we plot the ensemble difference between the 84th and 16th percentile) are tab- mean of the inferred monopole, t0, to subtract from the ulated in Table III. The tabulated results use the best-fit BORG contribution as a function of redshift, z, and see N , however we find that these change by no more than G that t0 is a smoothly varying function of z, indicating 50% across the range N = 3 5. G − that Nmon = 20 should be sufficient. We find that if We find that our weakest constraint is for the time we were to use small values of Nside we would calcu- delay between the highest and lowest energy channels late the wrong monopole at high redshift, but that the ((i, j) = (4, 1)), where we find ∆γ < 3.4 10−15 at 1σ 41 values quickly converge with increasing Nside. The max- confidence. This is unsurprising because each× pulse of a imum fractional difference between the calculated t0 at GRB’s lightcurve is known to peak first at higher energy −3 Nside = 32 and Nside = 64 is 3 10 , indicating that due to spectral evolution [47–50], although the reasons our map’s resolution is sufficient.× Moreover, we run the for this are not fully understood [20–26]. This results in inference again with Nside = 32 and Nmon = 10 and find the majority of observed time delays obeying ∆t > 0 −15 −15 ij 1σ constraints on ∆γ41 of 3.0 10 and 3.2 10 for i > j, so one expects the largest time delay for this respectively, suggesting that our× constraints are× robust pair and hence the weakest constraint. to these choices. Given the form of our noise model, if ∆γij = 0 then one In Section IV B, at each Monte Carlo iteration we drew expects µ˜(0) to be redshift independent. We check this a redshift from a two-sided Gaussian. One may be con- { ij } 9

0 103 (i, j) = (2, 1) (i, j) = (3, 1) 100 − (i, j) = (3, 2) (i, j) = (4, 1) 102 (i, j) = (4, 2) ))

ij (i, j) = (4, 3) Z

101

max ( − / ∆BIC 1 10 ij (i, j) = (2, 1) Z

(i, j) = (3, 1) log ( 102 (i, j) = (3, 2) − (i, j) = (4, 1) 100 (i, j) = (4, 2) (i, j) = (4, 3) 0 103 1 2 3 4 5 − 1 2 3 4 5 NG NG

(a) (b)

FIG. 3. Noise model comparison, including the Equivalence Principle-violating term, using (a) the BIC (Equation 21) with the number of parameters of our model equal to the BMD (Equation 33), and (b) the Bayes factor. The best-ﬁtting model should minimise the BIC and maximise the Bayes factor. For all cases except (i, j) = (4, 1) we see that NG = 4 is the optimal noise model using either statistic. For (i, j) = (4, 1) the BIC suggests NG = 3 is optimal, whereas the Bayes factor prefers NG = 4. We adopt NG = 3 for this frequency pair, but ﬁnd similar constraints on ∆γ41 for both NG = 3 and NG = 4.

model, binning our sources into Nbin = 5 linearly spaced (i, j) = (2, 1) redshift bins. Repeating the inference with Nbin = 1 or (i, j) = (3, 1) −15 (i, j) = (3, 2) Nbin = 10 changes our constraint to γ41 < 3.0 10 and γ < 4.3 10−15 respectively. This is as one× would (i, j) = (4, 1) 41 × (i, j) = (4, 2) expect; the constraint tightens as we decrease the number (i, j) = (4, 3) of parameters in the noise model and weakens as this increases. However this change is only 0.1 dex, indicating that our constraint is relatively insensitive to the choice of Nbin.

B. Comparison with the literature

The majority of previous attempts to constrain viola- 8 0 8 − 15 ∆γ 10− tions of the WEP via ∆γij with GRBs [18, 51–54], FRBs × [6, 55–58], Supernovae [59, 60], Gravitational Waves [61– 69], Blazar Flares [70–74] or Pulsars [75–78] have as- FIG. 4. Posteriors on ∆γij for the different time delay pairs, sumed that the gravitational potential is dominated by marginalised over the noise parameters. These constraints use the contribution from the Milky Way and/or other mas- the noise models which minimise the BIC (Equation 21). sive objects such as the Laniakea supercluster. This has two major shortcomings for distant sources: firstly the gravitational potential in Equation 2 should be a fluctu- cerned that this is not the correct distribution. To test ation about the cosmological mean (and thus can take the effect of the redshift uncertainty, we repeat the anal- either sign, unlike in the multiple-source approximation ysis but fix the redshift to the mean value, i.e. we as- where it is strictly additive), and secondly the long range sume zero redshift error. We find the constraints change behaviour of the gravitational potential means that we by 1.5% compared to our fiducial method, indicating cannot neglect the large-scale distribution of mass [8, 9]. ∼ that our constraints are not dominated by redshift un- These studies claim that only including a few sources certainty, and thus that the exact redshift distribution is underestimates the Shapiro time delay, making the con- not important. straints obtained conservative [79]. However, since the Finally, to account for the monopole of the Shapiro potential can take either sign, this reasoning is incorrect. time delay (which we are unable to predict or constrain Furthermore, the modelling of other contributions to the by itself), we introduce a redshift dependence in our noise time delay in these studies are relatively simplistic com- 10

5 6 1 0 1 0 ×

0.09 s / ) 0 0.06 z ( ) 0 ( 0.03 0.12 s / )

1 0.08 z ( ) 0 ( 0.04 s

/ 0.08 ) 2 z ( ) 0

( 0.00 s

/ 0.06 ) 3 z ( )

0 0.00 ( s

/ 0.08 ) 4 z ( ) 0.00 0 (

s 0.6 / ) 1 ( 0.3 s

/ 2 ) 2 (

0.4 ) 1 ( 0.3 )

2 0.2 (

0.1

0.06 s / ) 0 ( 0.03

0.75 s /

) 0.50 1 (

0.25

3.0 s / ) 2 ( 1.5

6 0 6 0.03 0.06 0.09 0.04 0.08 0.12 0.00 0.08 0.00 0.06 0.00 0.08 0.3 0.6 1 2 0.3 0.4 0.1 0.2 0.03 0.06 0.25 0.50 0.75 1.5 3.0

15 (0) (0) (0) (0) (0) (1) (2) (1) (2) (0) (1) (2) × 10 (z0) / s (z1) / s (z2) / s (z3) / s (z4) / s / s / s / s / s / s

FIG. 5. Constraints on ∆γij and the parameters describing other contributions to the time delay for (i, j) = (4, 1). The ij indices have been suppressed for clarity since all parameters are only for this pair. The contours show the 1 and 2σ confidence intervals. pared to our analysis; for example Yu et al. [18] assume clude the cosmological contribution to the time delays −13 all noise can be modelled as a single Gaussian, which we yielded constraints weaker than ∆γij < 10 [8, 80]. have shown to be a poor approximation. Not only have Using the same sources as us and the point-mass method- −13 we considered more sophisticated noise models, but we ology, Yu et al. [18] find ∆γ41 < 1.3 10 , which is a have demonstrated that our constraints are robust to the factor 40 weaker than our constraints.× As a consis- choice of model. tency check,∼ we re-run our inference for (i, j) = (4, 1) but (incorrectly) assume that the gravitational potential As well as being more robust than previous work, our is dominated by the Laniakea supercluster (of mass M = constraints are also stronger. The first attempts to in- 11

−13 straint is ∆γ41 < 1.1 10 , similar to that of [18]. One 0.08 (i, j) = (2, 1) might have expected× that increasing the number of Gaus- (i, j) = (3, 1) sian components describing the noise would weaken the (i, j) = (3, 2) constraint, not strengthen them. We ﬁnd that for NG = 1 0.06 (i, j) = (4, 1) (i, j) = (4, 2) the constraint is dominated by the few sources with the (i, j) = (4, 3) largest time delays. This is because the WEP-violating

s term must account for the wide tails in the measured / 0.04 time delays, upweighting larger values of ∆γ . When al- (0) ij ˜

µ | | lowing NG > 1 the broader Gaussians capture the tails instead, favouring smaller ∆γ . This is preferable be- 0.02 haviour because the constraint| | on ∆γ should come from the angular correlation of the measured time delays with

0.00 those predicted by BORG, rather than from the width of the measured time delay distribution itself. The long Shapiro time delays for extragalactic sources 0 1 2 3 4 result in tight constraints on ∆γij, however one can only Redshift bin measure the time delay differences and not the absolute time delays; we constrain ∆γij and not γi. Therefore, FIG. 6. Evolution of the means of the highest-weighted Gaus- although our constraints may appear tighter than So- (0) sian in the noise model,µ ˜ij , with redshift, where it is as- lar System measurements from the Cassini spacecraft sumed that all sources in the same redshift bin have the same [2] or Very Long Baseline Interferometry [81, 82] of (0) −5 −4 µ˜ij . Any variation with redshift is comparable to the uncer- γ 1 = (2.1 2.3) 10 and ( 0.8 1.2) 10 re- (0) − ± × − ± × tainty onµ ˜ij , as one would expect for a redshift-independent spectively, these have the advantage of constraining the intrinsic time delay contribution, since ∆γij is consistent with PPN parameter itself and thus can differentiate between zero. different theories which obey the Equivalence Principle. Nonetheless, our constraints indicate that the Equiva- lence Principle should be obeyed to within 3.4 10−15 for photons in the energy range 25 325 keV. × 10 −

9 C. Further applications 8

s 7 12 In this work we have used the BORG reconstruction

10 of the SDSS-III/BOSS galaxy compilation to predict the / 6 0

t Shapiro time delay to GRBs. The mean pseudo-redshifts 5 of the sources compiled by Yu et al. [18] is 2, and thus we had to include unconstrained contributions∼ to 4 the calculation between the edge of the constrained vol- 3 ume and the sources (Equation 18). This, coupled with the importance of long wavelength modes (Section IV B),

1 0 suggests that our constraints could be improved by the 10− 10 z next generation of galaxy surveys, such as with Euclid [83] or the Rubin Observatory [84], which will be sensitive up to z 2. To estimate the potential improvement, FIG. 7. Evolution of the ensemble mean of the monopole we run the end-to-end∼ inference assuming that we can ac- of Shapiro time delays, t0, with redshift, z. The monopole curately reconstruct the density field up to z = 1.26 and is found to vary smoothly with redshift, suggesting that our to twice this redshift by generating Gaussian Random interpolation procedure is reasonable. The ensemble mean is Fields. We find that the constraint improves by a factor positive due to local massive structures. 30% with this extra information. By combining the SDSS-III/BOSS∼ and 2M++ reconstructions one could also reduce the uncertainty of the low-redshift part of 1017M , at a distance d = 79 Mpc, with RA=10h32m the calculation, since 2M++ provides better constraints and Dec=-46 ◦000) as is done by Yu et al. [18]. Also mim- on the local density field. icking Yu et al. [18], we now assume that the redshifts A violation of the Equivalence Principle is not the only and angular positions of the sources have no uncertainty. phenomenon which could lead to a delay between arrival −14 We find constraints of ∆γ41 < 1.1 10 , which are times of photons of different energy. If the photon veloc- tighter than [18]. This is due to our× use of the optimal ity is energy-dependent, then photons which travel more noise model with NG = 3; if we use NG = 1 then our con- slowly will arrive later. If the photon has a non-zero rest 12 mass, mγ , then the speed of propagation increases with determine whether the large scale structure information increasing energy from the BORG algorithm could afford a detection of r this phenomenon. m2 1 m2 v = 1 γ 1 γ , (34) − E2 ≈ − 2 E2 VII. CONCLUSIONS (c 1), whereas in a quantum gravity scenario [85] one would≡ expect a different scaling with energy The PPN parameter γ is the same for all particles at E all energies if the WEP is obeyed. If there is a difference v 1 ξ , (35) ∼ − EQG in γ, ∆γij, between two photon frequencies, νi and νj, then there will be a spectral lag proportional to ∆γij and where ξ = 1 and E is the quantum gravity energy the Shapiro time delay. Most previous attempts to cal- ± QG scale, presumably around the Planck scale. In these culate the Shapiro delay in order to constrain ∆γij have cases, the Shapiro term is no longer dominant when com- assumed that the gravitational potential is dominated by puting the time delay, but the expected time delay for the only a few local sources and incorrectly argue that this massive photon is produces conservative constraints. In this work we constructed a source-by-source, Monte Z z dz0 ∆tij = ∆vij , (36) Carlo-based forward model for the Shapiro time delay 0 0 2 0 H (z ) (1 + z ) from Gamma Ray Bursts detected by the BATSE satellite. We work in a cosmological context by combining and for the quantum gravity case [86] the constrained local density field determined using the Z z 0 BORG algorithm with unconstrained, long-wavelength (1 + z ) 0 ∆tij = ∆vij 0 dz , (37) modes. Propagating uncertainties in the density field 0 H (z ) reconstruction via Monte Carlo sampling and marginalising over an empirical model characterising other con- where ∆vij is the difference in photon velocity between tributions to the time delay, we derive constraints on observed frequencies νi and νj. Applying a similar methodology to that developed here would allow one to ∆γij between the four energy channels, and for all pairs −15 constrain on the photon mass and quantum gravity en- find constraints at least as tight as ∆γij < 3.4 10 at 1σ confidence. These constraints are a factor× 30 ergy scale, and we will do so in future work. ∼ Strong lensing of distant objects smears their light into times tighter than previous results that use a cosmological model and 40 times tighter than if one neglects the an Einstein ring. The paths of photons observed across ∼ the ring trace out two cones that intersect at the lens cosmological contribution. plane. The relative time delays of these photons there- Our modelling is applicable to alternative multi- fore contain information not only on the distribution of messenger probes of the WEP, although these may re- mass in the lens that sources a large part of the poten- quire different models for the other contributions to the tial, but also on the path length itself which the time time delays. For example, for FRBs one would need ac- curate maps of the electron density of the universe to delay is proportional to. This length is a function of H0, enabling this fundamental cosmological parameter to be forward model the contribution of scattering from the constrained by measuring the time delays across the ring electron plasma, which depends on the integrated elec- [4, 5]. As in this work, the modelling is potentially sen- tron density (and is thus direction dependent) and scales −2 sitive to mass further away from the geodesics that the as ν . Furthermore, cosmological calculations of the photons follow than either the lens itself or the few ad- Shapiro time delay can determine the accuracy of lens ditional massive objects modelled in [4]. Provided one modelling when considering time delays across the Ein- has sufficient spatial resolution to determine time delay stein ring of a strongly lensed source or find use in anal- differences across the Einstein ring, by using the time de- ysis of the cosmic microwave background. Identifying lay maps from constrained density fields one can test the potential systematics in these probes of the Universe is validity of the external source assumptions. vitally important, especially in the context of the Hubble In this work we have only considered the time delay tension [89]. differences between different frequencies from a given source. The only cosmological regime in which the time delay itself may be directly measurable is the cosmic ACKNOWLEDGMENTS microwave background, where different delays between different regions of the sky imply varying times of re- We thank David Alonso for useful discussions and Guil- combination, and correspondingly varying temperature hem Lavaux for performing the SDSS-III/BOSS BORG of the blackbody radiation [87]. Standard autocorre- inference. DJB is supported by STFC and Oriel Col- lation techniques cannot currently detect such a delay, lege, Oxford. HD is supported by St John’s College, Ox- however cross-correlation with other fields appears to be ford. PGF is supported by the STFC. HD and PGF ac- a promising avenue [88]. We leave it to future work to knowledge financial support from ERC Grant No 693024 13 and the Beecroft Trust. JJ acknowledges support by the (https://www.aquila-consortium.org/). Swedish Research Council (VR) under the project 2020- 05143 – “Deciphering the Dynamics of Cosmic Struc- Some of the results in this paper have been derived ture”. This work was done within the Aquila Consortium using the healpy and HEALPix packages.

[1] I. I. Shapiro, Phys. Rev. Lett. 13, 789 (1964). [34] G. Lavaux, J. Jasche, and F. Leclercq, arXiv e-prints , [2] B. Bertotti, L. Iess, and P. Tortora, Nature 425, 374 arXiv:1909.06396 (2019). (2003). [35] J. Jasche and G. Lavaux, Astron. Astrophys. 625, A64 [3] C. M. Will and J. Nordtvedt, Kenneth, ApJ 177, 757 (2019). (1972). [36] D. J. Bartlett, H. Desmond, and P. G. Ferreira, [4] K. C. Wong et al., MNRAS 498, 1420 (2020). Phys. Rev. D 103, 023523 (2021). [5] S. Birrer et al., A&A 643, A165 (2020). [37] N. E. Chisari et al., ApJS 242, 2 (2019). [6] Z.-X. Luo, B. Zhang, J.-J. Wei, and X.-F. Wu, Journal [38] P. Lemos, A. Challinor, and G. Efstathiou, Journal of of High Energy Astrophysics 9, 35 (2016). Cosmology and Astro-Particle Physics 2017, 014 (2017). [7] LIGO Scientific Collaboration and Virgo Collaboration, [39] A. Zonca, L. Singer, D. Lenz, M. Reinecke, C. Rosset, Phys. Rev. Lett. 119, 161101 (2017). E. Hivon, and K. Gorski, Journal of Open Source Soft- [8] A. Nusser, ApJ 821, L2 (2016). ware 4, 1298 (2019). [9] O. Minazzoli, N. K. Johnson-McDaniel, and M. Sakel- [40] K. M. Górski,E. Hivon, A. J. Banday, B. D. Wandelt, lariadou, Phys. Rev. D 100, 104047 (2019). F. K. Hansen, M. Reinecke, and M. Bartelmann, ApJ [10] J. Jasche and B. D. Wandelt, MNRAS 425, 1042 (2012). 622, 759 (2005). [11] J. Jasche and B. D. Wandelt, MNRAS 432, 894 (2013). [41] F. Pedregosa et al., Journal of Machine Learning Re- [12] J. Jasche, F. S. Kitaura, B. D. Wandelt, and T. A. search 12, 2825 (2011). Enßlin, MNRAS 406, 60 (2010). [42] F. Feroz and M. P. Hobson, MNRAS 384, 449 (2008). [13] J. Jasche, F. Leclercq, and B. D. Wandelt, J. Cosmology [43] F. Feroz, M. P. Hobson, and M. Bridges, MNRAS 398, Astropart. Phys. 2015, 036 (2015). 1601 (2009). [14] G. Lavaux and J. Jasche, MNRAS 455, 3169 (2016). [44] F. Feroz, M. P. Hobson, E. Cameron, and A. N. Pettitt, [15] Planck Collaboration, A&A 641, A1 (2020). The Open Journal of Astrophysics 2, 10 (2019). [16] D. Alonso, arXiv e-prints , arXiv:2103.14016 (2021). [45] W. Handley and P. Lemos, Phys. Rev. D 100, 023512 [17] J. Hakkila, T. W. Giblin, K. C. Young, S. P. Fuller, C. D. (2019). Peters, C. Nolan, S. M. Sonnett, D. J. Haglin, and R. J. [46] W. Handley, The Journal of Open Source Software 4, Roiger, ApJS 169, 62 (2007). 1414 (2019). [18] H. Yu, S.-Q. Xi, and F.-Y. Wang, ApJ 860, 173 (2018). [47] D. Kocevski and E. Liang, ApJ 594, 385 (2003). [19] D. Yonetoku, T. Murakami, T. Nakamura, R. Yamazaki, [48] B. E. Schaefer, ApJ 602, 306 (2004). A. K. Inoue, and K. Ioka, ApJ 609, 935 (2004). [49] F. Ryde, A&A 429, 869 (2005). [20] R.-F. Shen, L.-M. Song, and Z. Li, MNRAS 362, 59 [50] J. Hakkila and R. D. Preece, ApJ 740, 104 (2011). (2005). [51] C. Sivaram, Bulletin of the Astronomical Society of India [21] R. J. Lu, Y. P. Qin, Z. B. Zhang, and T. F. Yi, MNRAS 27, 627 (1999). 367, 275 (2006). [52] H. Gao, X.-F. Wu, and P. Mészáros,ApJ 810, 121 [22] F. Daigne and R. Mochkovitch, MNRAS 342, 587 (2003). (2015). [23] Z. Y. Peng, Y. Yin, X. W. Bi, Y. Y. Bao, and L. Ma, [53] Y. Sang, H.-N. Lin, and Z. Chang, MNRAS 460, 2282 Astronomische Nachrichten 332, 92 (2011). (2016). [24] S.-S. Du, D.-B. Lin, R.-J. Lu, R.-Q. Li, Y.-Y. Gan, [54] J.-J. Wei, X.-F. Wu, H. Gao, and P. Mészáros,Journal of J. Ren, W. Xiang-Gao, and E.-W. Liang, ApJ 882, 115 Cosmology and Astroparticle Physics 2016, 031 (2016). (2019). [55] J.-J. Wei, H. Gao, X.-F. Wu, and P. Mészáros, [25] R.-J. Lu, Y.-F. Liang, D.-B. Lin, J. Lü,X.-G. Wang, H.- Phys. Rev. Lett. 115, 261101 (2015). J. Lü,H.-B. Liu, E.-W. Liang, and B. Zhang, ApJ 865, [56] S. J. Tingay and D. L. Kaplan, ApJ 820, L31 (2016). 153 (2018). [57] N. Xing, H. Gao, J.-J. Wei, Z. Li, W. Wang, B. Zhang, [26] Z. L. Uhm and B. Zhang, ApJ 825, 97 (2016). X.-F. Wu, and P. Mészáros,ApJ 882, L13 (2019). [27] D. J. Bartlett, H. Desmond, and P. G. Ferreira, Phys. [58] D. Wang, Z. Li, and J. Zhang, Physics of the Dark Uni- Rev. D 103, 123502 (2021). verse 29, 100571 (2020). [28] F. Moutarde, J. Alimi, F. R. Bouchet, R. Pellat, and [59] M. J. Longo, Phys. Rev. Lett. 60, 173 (1988). A. Ramani, ApJ 382, 377 (1991). [60] L. M. Krauss and S. Tremaine, Phys. Rev. Lett. 60, 176 [29] T. Buchert, A. L. Melott, and A. G. Weiss, A&A 288, (1988). 349 (1994). [61] X.-F. Wu, H. Gao, J.-J. Wei, P. Mészáros,B. Zhang, Z.- [30] F. R. Bouchet, S. Colombi, E. Hivon, and R. Juszkiewicz, G. Dai, S.-N. Zhang, and Z.-H. Zhu, Phys. Rev. D 94, A&A 296, 575 (1995). 024061 (2016). [31] R. Scoccimarro, ApJ 544, 597 (2000). [62] E. O. Kahya and S. Desai, Physics Letters B 756, 265 [32] F. Leclercq, J. Jasche, H. Gil-Mar´ın, and B. Wandelt, J. (2016). Cosmology Astropart. Phys. 11, 048 (2013). [63] I. LIGO Scientific Collaboration and Virgo Collabora- [33] D. J. Eisenstein et al., AJ 142, 72 (2011). tion, Fermi Gamma-ray Burst Monitor, ApJ 848, L13 14

(2017). (2016). [64] M. Liu, Z. Zhao, X. You, J. Lu, and L. Xu, Physics [76] Y. Zhang and B. Gong, ApJ 837, 134 (2017). Letters B 770, 8 (2017). [77] S. Desai and E. Kahya, European Physical Journal C 78, [65] J.-J. Wei, B.-B. Zhang, X.-F. Wu, H. Gao, P. Mészáros, 86 (2018). B. Zhang, Z.-G. Dai, S.-N. Zhang, and Z.-H. Zhu, J. [78] C. Leung, B. Hu, S. Harris, A. Brown, J. Gallicchio, and Cosmology Astropart. Phys. 2017, 035 (2017). H. Nguyen, ApJ 861, 66 (2018). [66] I. M. Shoemaker and K. Murase, Phys. Rev. D 97, 083013 [79] J.-J. Wei and X.-F. Wu, Frontiers of Physics 16, 44300 (2018). (2021). [67] S. Boran, S. Desai, E. O. Kahya, and R. P. Woodard, [80] H. Wang, F.-W. Zhang, Y.-Z. Wang, Z.-Q. Shen, Y.-F. Phys. Rev. D 97, 041501 (2018). Liang, X. Li, N.-H. Liao, Z.-P. Jin, Q. Yuan, Y.-C. Zou, [68] L. Yao, Z. Zhao, Y. Han, J. Wang, T. Liu, and M. Liu, Y.-Z. Fan, and D.-M. Wei, ApJ 851, L18 (2017). ApJ 900, 31 (2020). [81] S. B. Lambert and C. Le Poncin-Lafitte, A&A 499, 331 [69] S.-C. Yang, W.-B. Han, and G. Wang, MNRAS 499, (2009). L53 (2020). [82] S. B. Lambert and C. Le Poncin-Lafitte, A&A 529, A70 [70] Z.-Y. Wang, R.-Y. Liu, and X.-Y. Wang, Phys. Rev. (2011). Lett. 116, 151101 (2016). [83] R. Laureijs et al., arXiv e-prints , arXiv:1110.3193 (2011). [71] J.-J. Wei, J.-S. Wang, H. Gao, and X.-F. Wu, ApJ 818, [84] LSST Science Collaboration, arXiv e-prints , L2 (2016). arXiv:0912.0201 (2009). [72] S. Boran, S. Desai, and E. O. Kahya, European Physical [85] G. Amelino-Camelia, J. Ellis, N. E. Mavromatos, D. V. Journal C 79, 185 (2019). Nanopoulos, and S. Sarkar, Nature 393, 763 (1998). [73] R. Laha, Phys. Rev. D 100, 103002 (2019). [86] U. Jacob and T. Piran, J. Cosmology Astropart. Phys. [74] J.-J. Wei, B.-B. Zhang, L. Shao, H. Gao, Y. Li, Q.-Q. 2008, 031 (2008). Yin, X.-F. Wu, X.-Y. Wang, B. Zhang, and Z.-G. Dai, [87] W. Hu and A. Cooray, Phys. Rev. D 63, 023504 (2000). Journal of High Energy Astrophysics 22, 1 (2019). [88] P. Li, S. Dodelson, and W. Hu, Phys. Rev. D 100, 043502 [75] Y.-P. Yang and B. Zhang, Phys. Rev. D 94, 101501 (2019). [89] E. D. Valentino et al., (2020), arXiv:2008.11284.