betaFeedback

WorksPhysics and Astronomy

Seeing through the fog: astrophysical priors and the evidence for new physics in pulsar timing arrays

Abstract

Pulsar timing arrays (PTAs) have found evidence for a gravitational wave background at nHz frequencies. Inspiralling supermassive black hole binaries (SMBHBs) likely contribute to at least part of the signal, but there may be additional cosmological sources from new physics in the early universe. Disentangling the two is complicated by the prohibitively slow modelling of SMBHB spectra based on realistic population models. Here we introduce binaryfog, a surrogate of up-to-date population models also used by the NANOGrav collaboration, that returns the statistical distribution of bent spectra O(106)\mathcal{O}(10^6) faster than direct computation. We use binaryfog to profile over the six astrophysical parameters in fits of four extensively discussed new-physics models: a dark sector phase transition, cosmic superstrings, inflationary gravitational waves, and domain walls. Compared to the widely adopted power-law description of the SMBHB background, the preference for a phase transition drops from 2.8σ2.8σ to 1.5σ1.5σ. The other three models follow a similar pattern. We also recommend updated astrophysical priors based on galaxy surveys published since 2014, which for current PTA data still only give a marginal preference of 1.7σ1.7σ for a phase transition. An interpretation of the signal in terms of SMBHBs alone favours binaries that merge within about 2 2\,Gyr after their host galaxies have merged, requiring an efficient mechanism to overcome the final parsec of separation. If binaries need longer than about 6.5 6.5\,Gyr to merge, the preference for a phase transition exceeds 3σ3σ under our recommended updated prior set. binaryfog can help resolve these questions, as it allows users to study the impact of updated astrophysical priors and a realistic SMBHB population model on new-physics searches with PTAs.

Introduction

Pulsar timing arrays (PTAs) have reported evidence for a stochastic gravitational wave (GW) background at nanohertz frequencies, with the characteristic angular correlations predicted by Hellings and Downs [1, 2, 3, 4, 5, 6]. A natural explanation consists in the population of inspiralling supermassive black hole binaries (SMBHBs) left behind by galaxy mergers [7]. Fitting the observed GW background amplitude requires black hole masses or merger rates at the high end of what galaxy surveys support [8, 9, 10, 11, 12]. A recent determination of the massive end of the galaxy stellar mass function has been reported to remove much of this discrepancy [13], but a fit that includes the environmental hardening of the binaries finds it insufficient [14]. The measured spectrum is also flatter than the hc∝f−4/3h_\mathrm{c} \propto f^{-4/3} of binaries driven by gravitational wave emission alone [15]; part of the excess may come from the priors on the intrinsic red noise of the pulsars [16], and it is unsettled how binaries traverse the final parsec [17, 18, 19, 20].

A cosmological source, such as a phase transition in a dark sector, cosmic strings, domain walls, or blue-tilted inflationary tensor modes, could also explain (part of) the signal [21, 22, 23, 24, 25, 26, 27, 28, 29]. Fits of such sources so far have left the binaries out [23, 24, 26, 27, 30, 31, 32, 33, 34], represented them by a power law with amplitude AA and spectral index γ\gamma, under a fixed Gaussian prior based on population models in which the binaries evolve by gravitational wave emission alone [8, 21, 28, 29, 35, 36, 37] or with γ=13/3\gamma= 13/3 fixed [22, 38], or described them by population parameters fitted to the timing data without priors from galaxy surveys [25]. We refer to the very common second treatment as the (A,γ)(A, \gamma) power law. It discards much of the information the underlying population models contain, with independent parameters describing the galaxy stellar mass function, the relation between black hole and bulge mass, and the hardening of the binaries. The background in a given frequency bin, furthermore, is dominated by a few loud sources, so the prediction is a skewed distribution of spectra over realisations of the population [39, 40, 41, 42]. It is debated whether a pulsar timing array on its own can separate SMBHBs from cosmological backgrounds [43, 44, 45, 46].

Here we ask how strongly the data prefer – or constrain – a cosmological contribution once the binaries are modelled as an actual population, and how this preference depends on the astrophysical priors. To this end we construct binaryfog, a surrogate built from a library of holodeck models [8, 47] that returns this distribution in every frequency bin for any point of the six-dimensional astrophysical parameter space in 0.2 milliseconds instead of three minutes. We use it to profile over those astrophysical parameters in fits of four often discussed scenarios of new physics that may contribute to the GW background – a dark sector phase transition, cosmic superstrings, inflationary gravitational waves, and domain walls – to the NANOGrav 15 yr free spectrum re-analysed in 2026 [48]. We generally find that the evidence for a phase transition decreases significantly when using the full population model rather than a simple power law. We point out that this conclusion depends crucially on the adopted priors for the parameters of the population model, and we discuss in detail the impact of up-to-date astrophysically motivated priors on new-physics interpretations in the context of the four scenarios mentioned above.

After introducing the population model and the surrogate in § 2 and the likelihood in § 3, we test the surrogate and compare the priors in § 4, before turning to our main focus and results, namely the new-physics fits and their dependence on the free spectrum and on the priors in § 5. We conclude in § 6. The appendices collect the new-physics spectra (appendix § 8), the priors and their sources (appendix § 9), and a quick-start guide to downloading and using binaryfog (appendix § 10).

A surrogate for GW backgrounds from supermassive black hole binaries

The standard treatment of the binary background in searches for new physics, the (A,γ)(A,\gamma) power law [21], takes the characteristic strain to be hc(f)=A(f/fyr)(3−γ)/2h_{\mathrm{c}}(f)=A\left(f/f_{\mathrm{yr}}\right)^{(3-\gamma)/2}, or

h2ΩGWbin(f)=2π2fyr23H1002A2(ffyr)5−γ(1)h^2\Omega_{\mathrm{GW}}^{\mathrm{bin}}(f)=\frac{2\pi^2 f_{\mathrm{yr}}^2}{3H_{100}^2}A^2\left(\frac{f}{f_{\mathrm{yr}}}\right)^{5-\gamma} \tag*{(1)}

with fyr=1/yrf_{\mathrm{yr}}=1/\mathrm{yr} and H100=100 km s−1 Mpc−1H_{100}=100\ \mathrm{km}\ \mathrm{s}^{-1}\ \mathrm{Mpc}^{-1}. The prior on (log⁡10A,γ)(\log_{10} A,\gamma) is a two-dimensional Gaussian centred at log⁡10A=−15.61\log_{10} A=-15.61 and γ=4.71\gamma=4.71, with standard deviations 0.53 and 0.35 and covariance −0.0026-0.0026; it derives from power-law fits to population models in which the binaries evolve by gravitational wave emission alone [8, 21]. In these models a binary is taken to emit at the same redshift as its galaxy merger. The inspiral is thus effectively treated as instantaneous, while in reality it would take several billion years without environmental effects. To quote the NANOGrav collaboration [8], this model therefore is not self-consistent, as GW-only evolution is unable to bring binaries to the PTA band within a Hubble time, but still kept for comparison because it tends to produce the highest GW background amplitudes. This motivates us to consider, instead, a self-consistent population model (and its impact on new-physics searches).

Supermassive black hole binary population synthesis

We work with the phenomenological semi-analytic models of holodeck, the population synthesis code used for NANOGrav’s own astrophysical interpretation [8, 47], in the parametrisation of Chen, Sesana, and Conselice [49]. All our libraries are built with holodeck in the first version to include the correction of the erratum [48], see appendix C for further details. In this parametrisation, the differential number of binaries dN\mathrm{d}N emitting in a logarithmic orbital frequency interval dln⁡fp\mathrm{d}\ln f_{\mathrm{p}} follows from the comoving number density nn and the time a binary spends there as [8]

d4Ndlog⁡10M dq dz dln⁡fp=d3ndlog⁡10M dq dzdVcdzdtdln⁡fpdzdt(2)\frac{\mathrm{d}^4N}{\mathrm{d}\log_{10}M\,\mathrm{d}q\,\mathrm{d}z\,\mathrm{d}\ln f_{\mathrm{p}}} = \frac{\mathrm{d}^3n}{\mathrm{d}\log_{10}M\,\mathrm{d}q\,\mathrm{d}z} \frac{\mathrm{d}V_{\mathrm{c}}}{\mathrm{d}z} \frac{\mathrm{d}t}{\mathrm{d}\ln f_{\mathrm{p}}} \frac{\mathrm{d}z}{\mathrm{d}t} \tag*{(2)}

with MM the total binary mass, q≤1q\leq1 the mass ratio between the two black holes, zz the redshift, and VcV_{\mathrm{c}} the comoving volume. The orbital frequency fpf_{\mathrm{p}} and the time tt are measured in the binary rest frame, so that dt/dln⁡fp\mathrm{d}t/\mathrm{d}\ln f_{\mathrm{p}} is the time the binary spends per logarithmic interval of orbital frequency, its residence time. A circular binary emits gravitational waves at twice its orbital frequency, observed today at f=2fp/(1+z)f=2f_{\mathrm{p}}/(1+z). The first factor on the right-hand side of eq. (2) is the comoving number density of binaries per unit mass, mass ratio, and redshift. It follows from the galaxy merger rate, which is determined by a galaxy stellar mass function, a galaxy pair fraction, and a galaxy merger timescale; a relation between the black hole mass and the stellar bulge mass thus turns the galaxy mergers into mergers of black hole pairs.

The inspiral of each pair can last over a billion years, i.e. cosmological timescales. holodeck evaluates the number density at the redshift of the galaxy merger, integrates the evolution of the binary (described below) from there, and evaluates the residence time and the cosmological factors at the later redshift zz at which the binary reaches the orbital frequency fpf_{\mathrm{p}}. Binaries that would reach fpf_{\mathrm{p}} only in the future hence do not contribute. For more details of the galaxy merger modelling we refer to ref. [8].

The evolution of the binary through environmental effects is the least certain ingredient of the model. After their host galaxies merge, the black hole pairs have to harden from kiloparsec separations to the sub-parsec ones at which gravitational wave emission takes over. In the PTA band binaries of 10810^{8} to 1010 M⊙10^{10}\ M_{\odot} masses are 0.004 to 0.1 pc apart, and they coalesce at 0.4 to 40 μHz\mu\mathrm{Hz} in their rest frame, between the PTA band and the LISA band [50]. In spherical galaxies, stars orbiting close to the binary (within the so-called loss cone) are at some point ejected faster than two-body relaxation refills it. Hardening therefore stalls at parsec-scale separations, which is known as the final parsec problem [17, 18, 51]. Proposed solutions include stellar scattering in non-spherical or merging galaxies [19, 20], torques from gas [52], three-body interactions [53, 54], and the dynamical friction exerted by spikes of self-interacting dark matter [55]. Population synthesis codes do not resolve the problem but parametrise its outcome [10]. In this work we use a broken power law for the rate at which the binary separation rr shrinks [8, 56],

drdt∣env=−A(rrchar)1−νinner(1+rrchar)νinner−νouter(3)\left.\frac{\mathrm{d}r}{\mathrm{d}t}\right|_{\mathrm{env}}=-A\left(\frac{r}{r_{\mathrm{char}}}\right)^{1-\nu_{\mathrm{inner}}}\left(1+\frac{r}{r_{\mathrm{char}}}\right)^{\nu_{\mathrm{inner}}-\nu_{\mathrm{outer}}} \tag*{(3)}

with tt again the time in the rest frame of the binary. The time spent per logarithmic interval of separation hence scales as rνinnerr^{\nu_{\mathrm{inner}}} well inside the break separation rcharr_{\mathrm{char}} and as rνouterr^{\nu_{\mathrm{outer}}} well outside it. The normalisation AA is fixed by requiring that the total time from an initial separation of 10 kpc to the innermost stable circular orbit equals a target lifetime τf\tau_{\mathrm{f}}. The hardening of the binary by gravitational wave emission is added to this rate and takes over at small separations.

As a default, we use NANOGrav’s fiducial Phenom model for the galaxy merger rates, see eqs. (14)–(18) and table B1 of ref. [8]. Its galaxy stellar mass function

Ψ(m⋆,z)=ln⁡(10) Ψ0(m⋆MΨ)αΨexp⁡(−m⋆MΨ)(4)\Psi(m_{\star},z)=\ln(10)\,\Psi_{0}\left(\frac{m_{\star}}{M_{\Psi}}\right)^{\alpha_{\Psi}}\exp\left(-\frac{m_{\star}}{M_{\Psi}}\right) \tag*{(4)}

is a single Schechter function, providing the number of galaxies per comoving Mpc3\mathrm{Mpc}^{3} per decade of stellar mass m⋆m_{\star}, the total mass of the stars in a galaxy, at redshift zz. It follows a power law of slope αΨ=1+αΨ,0+αΨ,zz\alpha_{\Psi}=1+\alpha_{\Psi,0}+\alpha_{\Psi,z}z at low mass and is cut off exponentially above a characteristic mass modelled as log⁡10(MΨ/M⊙)=mΨ,0+mΨ,zz\log_{10}(M_{\Psi}/M_{\odot})=m_{\Psi,0}+m_{\Psi,z}z. Its normalisation changes exponentially with redshift, log⁡10(Ψ0/Mpc−3)=ψ0+ψzz\log_{10}(\Psi_{0}/\mathrm{Mpc}^{-3})=\psi_{0}+\psi_{z}z. Only ψ0\psi_{0} and mΨ,0m_{\Psi,0} are free parameters, while ψz=−0.60\psi_{z}=-0.60, mΨ,z=+0.11m_{\Psi,z}=+0.11, αΨ,0=−1.21\alpha_{\Psi,0}=-1.21 and αΨ,z=−0.03\alpha_{\Psi,z}=-0.03 are fixed at the values used in the NANOGrav analysis (table B1 of ref. [8]), see appendix § 9.§ 1 for further discussion. The mass M∙M_{\bullet} of each black hole of the binary follows the mass of the galactic bulge, Mbulge=0.615m⋆M_{\mathrm{bulge}}=0.615m_{\star}, with the bulge fraction 0.615 taken from measurements in massive galaxies [57, 58], and a log-normal scatter ϵμ\epsilon_{\mu},

log⁡10(M∙/M⊙)=μ+αμlog⁡10(Mbulge/1011 M⊙)+N(0,ϵμ)(5)\log_{10}(M_{\bullet}/M_{\odot})=\mu+\alpha_{\mu}\log_{10}(M_{\mathrm{bulge}}/10^{11}\ M_{\odot})+\mathcal{N}(0,\epsilon_{\mu}) \tag*{(5)}

where μ\mu is a dimensionless black hole mass normalisation and the slope is fixed to αμ=1.10\alpha_{\mu}=1.10. Black hole masses correlate with properties of their host galaxies, such as the bulge mass, and these correlations are read as evidence that black holes and galaxies grow together [59]. As recommended by the NANOGrav analysis, we vary six parameters: the normalisation ψ0\psi_{0} and the characteristic mass mΨ,0m_{\Psi,0} of the galaxy stellar mass function, the normalisation μ\mu and the intrinsic scatter ϵμ\epsilon_{\mu} of the black hole to bulge mass relation, the binary lifetime τf\tau_{\mathrm{f}} in Gyr, and the inner hardening index νinner\nu_{\mathrm{inner}}. We also follow ref. [8] in fixing the outer index to

νouter=+2.5\nu_{\mathrm{outer}}=+2.5, as motivated by dynamical friction against the stars and dark matter of the merged host galaxy [8, 60]. The break separation is held fixed at rchar=100 pcr_{\mathrm{char}}=100\,\mathrm{pc}.1

We consider four sets of priors on these parameters, as summarised in Table 1. Flat priors represent a completely uninformed choice. The NANOGrav priors [8] are Gaussian in ψ0\psi_0, mψ,0m_{\psi,0}, μ\mu, and ϵμ\epsilon_\mu and uniform in τf\tau_f and νinner\nu_{\mathrm{inner}}. Their centres follow from fits to galaxy survey data and black hole mass measurements [59, 62, 63], but their widths are up to four times larger than the uncertainties of these fits. The Tomczak–Kormendy priors take the centres and uncertainties of these fits, and the survey-average priors the centres and the spread of new measurements published since then. § 4.2 further motivates these specific prior choices, and appendix § 9 traces them back to their origin in the literature. When not referring to a specific prior choice in the following, results refer to the NANOGrav priors.

parameterflatNANOGravTomczak–Kormendysurvey average
ψ0\psi_0U(−3.5,−1.5)U(-3.5,-1.5)N(−2.56,0.40)\mathcal{N}(-2.56,0.40)N(−2.61,0.10)\mathcal{N}(-2.61,0.10)N(−2.20,0.15)\mathcal{N}(-2.20,0.15)
mψ,0m_{\psi,0}U(10.5,12.5)U(10.5,12.5)N(10.90,0.40)\mathcal{N}(10.90,0.40)N(10.96,0.11)\mathcal{N}(10.96,0.11)N(10.82,0.11)\mathcal{N}(10.82,0.11)
μ\muU(7.6,9.0)U(7.6,9.0)N(8.60,0.20)\mathcal{N}(8.60,0.20)N(8.69,0.05)\mathcal{N}(8.69,0.05)N(8.80,0.10)\mathcal{N}(8.80,0.10)
ϵμ\epsilon_\muU(0,0.9)U(0,0.9)N(0.32,0.15)\mathcal{N}(0.32,0.15)N(0.28,0.05)\mathcal{N}(0.28,0.05)N(0.40,0.05)\mathcal{N}(0.40,0.05)
τf/Gyr\tau_f/\mathrm{Gyr}U(0.1,11)U(0.1,11)U(0.1,11)U(0.1,11)U(0.1,11)U(0.1,11)U(0.1,11)U(0.1,11)
νinner\nu_{\mathrm{inner}}U(−1.5,0)U(-1.5,0)U(−1.5,0)U(-1.5,0)U(−1.5,0)U(-1.5,0)U(−1.5,0)U(-1.5,0)

Table 1. The six parameters of the population model and the four prior sets that we adopt in this work. UU denotes a uniform prior between its two arguments, and N\mathcal{N} indicates a Gaussian with stated central value and width. ψ0\psi_0 is stated in units of log⁡10 Mpc−3\log_{10}\,\mathrm{Mpc}^{-3}, mψ,0m_{\psi,0} and μ\mu in units of log⁡10 M⊙\log_{10}\,\mathrm{M}_{\odot}, ϵμ\epsilon_\mu in dex. The NANOGrav priors are those of the holodeck library of ref. [8]; the Tomczak–Kormendy priors take the centres and uncertainties of the fits they are based on, and the survey-average priors the centres and spread of measurements published since 2014, see also appendix § 9.

Realisation scatter

The characteristic strain in a redshifted frequency bin is the sum over the population in eq. (2) weighted by the squared strain hs2h_s^2 of each binary, in the form of Phinney [15] and of Sesana, Vecchio, and Colacino [64]. The population herein is characterised by its expected number of binaries in each bin of mass, mass ratio, redshift, and frequency. The number actually realised in a given universe is modelled as following a Poisson distribution, so the characteristic strain squared in an observed frequency bin Δf\Delta f is

hc2(f)=∑M,q,zP(d4Ndlog⁡10M dq dz dln⁡fpΔlog⁡10M Δq Δz Δln⁡f)hs2(fp)Δln⁡f,(6)h_{\mathrm{c}}^2(f)=\sum_{M,q,z}\mathcal{P}\left(\frac{\mathrm{d}^4N}{\mathrm{d}\log_{10}M\,\mathrm{d}q\,\mathrm{d}z\,\mathrm{d}\ln f_{\mathrm{p}}}\Delta\log_{10}M\,\Delta q\,\Delta z\,\Delta\ln f\right)\frac{h_{\mathrm{s}}^2(f_{\mathrm{p}})}{\Delta\ln f}, \tag*{(6)}

where P(λ)\mathcal{P}(\lambda) is a Poisson variate with mean λ\lambda and hsh_s is the sky- and polarisation-averaged strain of a single binary, which decreases with its comoving distance dc(z)d_c(z) as hs2∝dc−2h_s^2\propto d_c^{-2} [8]. Keeping the inclination of each binary, which holodeck averages over, would broaden the distribution by about ten per cent when many sources share a bin [39]. Averaging eq. (6) over many draws gives a smooth expected strain that, however, no individual realisation follows, since near the measured amplitude a few loud binaries typically carry the signal.

The spread between realisations grows with frequency. We quantify the realisation scatter in each frequency bin as half the difference between the sixteenth and eighty-fourth percentiles of the predicted logarithmic strain squared throughout this work. For our fiducial astrophysical model, this half-width increases from 0.03 dex in the lowest frequency bin to 0.32 dex in the fourteenth bin. By comparison, the corresponding half-widths of the NANOGrav free-spectrum posteriors are 0.40 and 0.81 dex. Thus, the scatter between different realisations of the SMBHB population is smaller than the observational uncertainty in both bins. We therefore expect the realisation scatter alone to move the favoured new-physics parameter regions little, given the current data.

For the purpose of a likelihood we further need the distribution of log⁡10hc2\log_{10} h_c^2 in each bin. We adopt a one-parameter warp of a normally distributed variable, which we find (cf. § 4.3) to describe the distribution of the holodeck realisations better than a log-normal or a map-Airy distribution:

log⁡10hc2(fi)=mi+σi(ξ+cξ2),ξ∼N(0,1),(7)\log_{10} h_c^2(f_i) = m_i + \sigma_i\left(\xi+ c\xi^2\right), \qquad\xi\sim\mathcal{N}(0,1), \tag*{(7)}

with mim_i the median of log⁡10hc2\log_{10} h_c^2 in bin ii, σi\sigma_i its half-width between the sixteenth and eighty-fourth percentiles, and cc a skew parameter. The skew is taken to be the same in every frequency bin of a population model but depends on the model, i.e. on the six astrophysical parameters, through the map spelled out right below in § 2.3.2 The warp in eq. (7) is held flat below ξ=−1/(2c)\xi=-1/(2c), where it turns over, which sets a floor on how faint a realisation can be.

A surrogate for the spectral distribution

Evaluating holodeck with 2000 realisations takes three minutes on one core. The NANOGrav analysis therefore interpolates their library with Gaussian processes for the median and the standard deviation of log⁡10hc2\log_{10} h_c^2 per bin [8, 65], as do fits of new physics affecting the binary evolution [66, 67]. Alternative approaches have used similar neural-network emulators [68, 69], a normalising flow over five bins that keeps the correlations between them [70], and a direct evaluation of the semi-analytic models [71]. For realistic full parameter scans, however, we need the shape of the distribution in each bin and an evaluation in a fraction of a millisecond for efficient sampling, and hence fit polynomial maps instead.

The median spectrum is described by a transition from environment-driven to GW-driven hardening, in the phenomenological turnover form used in pulsar timing analyses [72]:

hc(f)=A(ffyr)−2/3[1+(fbendf)κ]−1/2,(8)h_c(f) = A\left(\frac{f}{f_{\rm yr}}\right)^{-2/3}\left[1+\left(\frac{f_{\rm bend}}{f}\right)^\kappa\right]^{-1/2}, \tag*{(8)}

with fyr=1 yr−1f_{\rm yr}=1\,\mathrm{yr}^{-1}, an amplitude AA, a bend frequency fbendf_{\rm bend} and a fixed sharpness κ=1.8\kappa=1.8, first used on measured data in ref. [73].3

The surrogate maps the six astrophysical parameters θ=(ψ0,mψ,0,μ,ϵμ,τf,νinner)\theta= (\psi_0, m_{\psi,0}, \mu, \epsilon_\mu, \tau_f, \nu_{\mathrm{inner}}) onto the three quantities needed by the likelihood. Each is a polynomial in u(θ)u(\theta), which maps the ranges of the flat priors linearly onto the hypercube [−1,1]6[-1,1]^6:

log⁡10hc∣i2(θ)=2log⁡10hc(fi;A(θ),fbend(θ))+∑∣n∣≤4wn,imedun(9)\log_{10} h_{c|i}^2(\theta) = 2\log_{10} h_c(f_i; A(\theta), f_{\mathrm{bend}}(\theta)) + \sum_{|n|\leq4} w^{\mathrm{med}}_{n,i}u^n \tag*{(9)}
log⁡10σi(θ)=∑∣n∣≤4wn,iσun,c(θ)=∑∣n∣≤2wncun.(10)\log_{10}\sigma_i(\theta) = \sum_{|n|\leq4} w^\sigma_{n,i}u^n,\qquad c(\theta) = \sum_{|n|\leq2} w^c_n u^n. \tag*{(10)}

Here n=(n1,…,n6)n=(n_1,\ldots,n_6) runs over the multi-indices of total degree ∣n∣=n1+⋯+n6≤4|n|=n_1+\cdots+n_6\leq4, un=∏k=16uknku^n=\prod_{k=1}^6u_k^{n_k}, and hch_c in the first term is the bend of eq. (8). Rather than fitting AA and fbendf_{\mathrm{bend}} directly, which are poorly determined for spectra whose turnover falls outside the simulated range, we fit the value and the slope of log⁡10hc2\log_{10}h_c^2 at a reference frequency of 5 nHz and invert eq. (8) for them in closed form at fixed κ\kappa. The total size of the four maps is 54 kB: two polynomials of degree four for the reference value and slope, one per bin for the median correction and for the width, and one of degree two for the skew, which varies little through the parameter space. The library holds 2000 astrophysical models θ\theta, drawn by Latin hypercube sampling over the flat prior ranges, each realised 2000 times. Models whose spectrum at 5 nHz lies more than three decades in hc2h_c^2 below the measured one are not used. Three quarters of the remaining models are used for the fit and one quarter is held out for testing.

The pulsar timing likelihoods and parameter inference

Free-spectrum likelihood

We use the free-spectrum posteriors distributed through ceffyl [74] and PTArcade [75] instead of the timing residuals. They are marginalised over the pulsar noise and dispersion measure parameters, which makes the likelihood about a thousand times faster than in enterprise [76]. We restrict ourselves to the Nf=14N_f=14 lowest frequency bins, as recommended by NANOGrav [77], since the higher bins carry almost no information about a signal. In each bin ii the free spectrum gives the posterior of ρi\rho_i, the square root of the timing-residual power, without assuming a spectral shape. A stochastic background h2ΩGWh^2\Omega_{\mathrm{GW}} can be translated to the timing residual power and the characteristic strain hch_c through

h2ΩGW(f)=8π4H1002ρ2f5T=2π23H1002f2hc2(f),(11)h^2\Omega_{\mathrm{GW}}(f)=\frac{8\pi^4}{H_{100}^2}\rho^2 f^5T=\frac{2\pi^2}{3H_{100}^2}f^2h_c^2(f), \tag*{(11)}

where TT is the observation time span and H100≡100 km s−1 Mpc−1H_{100}\equiv100\ \mathrm{km\ s^{-1}\ Mpc^{-1}}, so that the likelihood of a deterministic spectrum is calculated from the product over bins of the measured density evaluated at the predicted ρi\rho_i. We follow ceffyl’s conventions at the edges of the stored kernel density grids: below the grid, the probability density is held at its value at the lower grid boundary, since sufficiently weak signals are indistinguishable from noise. Above the grid, the model is rejected because the predicted gravitational-wave amplitude exceeds the observational upper limit.

Marginalisation over realisations, profile likelihood, and statistical preference

The binaries do not contribute as a deterministic prediction for the spectrum but through the distribution discussed in § 2.2. As a result the log-likelihood of a new-physics spectrum Ωnp\Omega_{\mathrm{np}}, depending on the free parameters ϑ\vartheta and the model nuisance parameters ϑnuis\vartheta_{\mathrm{nuis}}, together with a population model θ\theta consists of the sum over the NfN_{\mathrm{f}} Fourier frequencies of their log-likelihoods. Each Fourier bin’s contribution to the log-likelihood is an integral over the binary contribution’s distribution in that bin, hence

ln⁡L(ϑ,ϑnuis,θ)=∑i=1Nfln⁡∫dξ φ(ξ) p(ρi(Ωnp(ϑ,ϑnuis)+Bi(ξ∣θ))∣data).(12)\ln\mathcal{L}(\vartheta,\vartheta_{\mathrm{nuis}},\theta)=\sum_{i=1}^{N_{\mathrm{f}}}\ln\int\mathrm{d}\xi\,\varphi(\xi)\,p\left(\rho_i\left(\Omega_{\mathrm{np}}(\vartheta,\vartheta_{\mathrm{nuis}})+B_i(\xi\mid\theta)\right)\mid\mathrm{data}\right). \tag*{(12)}

Here φ\varphi is the standard normal density, and Bi(ξ∣θ)B_i(\xi\mid\theta) the binary contribution in bin ii of the realisation labelled by ξ\xi through eq. (7), with the median, width and skew supplied by the surrogate, and p(ρi)p(\rho_i) is the posterior distribution of the free-spectrum analysis. In practice, we evaluate the integral as a sum over a fixed grid of 101 nodes in ξ\xi between −4-4 and 44, weighted by φ\varphi. In the limit of vanishing width of the binary background eq. (12) reduces to the usual likelihood of a summed spectrum of two contributions.

For the inference of the 1σ1\sigma and 2σ2\sigma-preferred regions of the new-physics parameter space we use a profile likelihood and Wilks’ theorem: for a given parameter (combination) ϑ\vartheta which we want to constrain, all astrophysical nuisance parameters θ\theta and model nuisance parameters ϑnuis\vartheta_{\mathrm{nuis}} are maximised (profiled) over:

ln⁡Lprof(ϑ)=max⁡ϑnuis,θ[ln⁡L(ϑ,ϑnuis,θ)+ln⁡π(θ)].(13)\ln\mathcal{L}_{\mathrm{prof}}(\vartheta)=\max_{\vartheta_{\mathrm{nuis}},\theta}\left[\ln\mathcal{L}(\vartheta,\vartheta_{\mathrm{nuis}},\theta)+\ln\pi(\theta)\right]. \tag*{(13)}

We treat the new-physics model parameters as having flat priors over their ranges stated in appendix § 8. For the SMBHB background parameters θ\theta, we use instead the priors listed in Table 1. The 1σ1\sigma and 2σ2\sigma regions, which would contain 68.3% and 95.4% of the probability of a Gaussian distribution, then span the parameter space with ln⁡Lprof\ln\mathcal{L}_{\mathrm{prof}} up to 1.15 and 3.09 below the best-fit likelihood ln⁡Lprofmax⁡\ln\mathcal{L}_{\mathrm{prof}}^{\max} for two scanned model parameters within ϑ\vartheta, and up to 0.50 and 2.00 for one model parameter [78], respectively. To quantify whether new physics is preferred, we use the profile likelihood ratio (or gain) GG of the best fit over the same fit without it,

G=ln⁡Lprofmax⁡−max⁡θ[ln⁡L(θ)∣Ωnp=0+ln⁡π(θ)].(14)G=\ln\mathcal{L}_{\mathrm{prof}}^{\max}-\max_{\theta}\left[\left.\ln\mathcal{L}(\theta)\right|_{\Omega_{\mathrm{np}}=0}+\ln\pi(\theta)\right]. \tag*{(14)}

To express GG as a nominal significance, we use the test statistic Λ=2G\Lambda=2G and adopt the approximate conversion p=12P(χk2≥Λ)p=\frac{1}{2}P(\chi_k^2\geq\Lambda), where χk2\chi_k^2 denotes a chi-square distribution with kk degrees of freedom, and kk is the number of free parameters of the new-physics model. This prescription has also been used in gamma-ray source searches [79]. We quote the corresponding one-sided Gaussian significance Z=Φ−1(1−p)Z=\Phi^{-1}(1-p), where Φ\Phi is the cumulative standard normal distribution. The conversion is approximate because the spectral-shape parameters are not identifiable when the new-physics contribution vanishes, so the exact null distribution need not follow this prescription [80, 81]. Monte Carlo simulations can be used to go beyond this approximation, see for instance ref. [82], which we leave for future studies. The quoted significances should therefore be regarded as nominal. A significance of 1, 2, and 3σ3\sigma then requires G=0.50G=0.50, 2.00, and 4.50 for k=1k=1, 1.15, 3.09, and 5.91 for k=2k=2, and 1.76, 4.01, and 7.08 for k=3k=3. A contour in a plane uses k=2k=2, and a significance derived from GG uses the number kk of free parameters of the new-physics model, including those profiled over. A model with more parameters thus has to gain more in GG to reach the same significance, which makes the significances of different models comparable. The astrophysical parameters do not enter kk, since both terms of eq. (14) contain them.

The (A,γ)(A, \gamma) power law can be profiled as well via eq. (13), with its Gaussian prior as a penalty, so that both treatments of the binaries differ only in the family of background spectra and in the prior on it. This prior matters in its own right, since part of GG comes from the prior penalty when the astrophysical parameters move away from the values galaxy surveys prefer. In § 5.4 we will get back to this, and discuss these two contributions separately.

Noise modelling and available data sets

The free-spectrum posteriors p(ρi)p(\rho_i) depend on the modelling of the intrinsic pulsar noise, the dispersion measure variations and the correlations between pulsars. We use the public posteriors of the NANOGrav 15 yr data set in their most recent version, 2.0.0 of the data record [83], re-run after a correction of the sampler [48, 84]. This release also includes the spectrum with Hellings–Downs (HD) correlations fitted together with a monopole (MP), as e.g. resulting from clock errors, a dipole (DP), like from errors of the solar system ephemeris, and a common uncorrelated red noise (CURN). The NANOGrav analysis adopts this HD+MP+DP+CURN spectrum as the most conservative choice for their astrophysical interpretation [8, 48]. Let us remark that, at the time of writing, PTArcade [85] still downloads the spectra released in 2023 together with NANOGrav’s first analyses of the 15 yr data set, from before the correction.4

We use the CURN spectrum of the 2026 release as a baseline for our analysis, which assumes the same red noise process in every pulsar without correlations between pulsars, with deterministic dispersion measure offsets. The spectrum rests on the auto-correlations of the pulsars, so CURN recovers it without the suppression of the Hellings–Downs overlap reduction function [74, 75]. NANOGrav’s new-physics search used the HD spectrum [21]. In § 5.3, we will explore in detail the impact of these choices on new-physics interpretations of the PTA data.

The binary background, its priors, and validation of the surrogate

Before we use the surrogate for new-physics searches in § 5, we take a closer look at how the astrophysical parameters individually shape the spectrum (§ 4.1) and how this overall translates to the range of spectra compatible with the priors considered in this work (§ 4.2). We then perform a number of validation tests for the surrogate, by comparing the warp of eq. (7) with other distributions used in the literature (§ 4.3) and by reproducing NANOGrav’s fit of the population model (§ 4.4).

Anatomy of the binary background and comparison to a power law

The first three columns of Figure 1 show how varying each of the six parameters individually changes the median spectrum, when holding the other parameters fixed at their central values. We also include, for comparison, our baseline free spectrum inferred by NANOGrav.

Figure 1. The median predicted SMBHB spectra that result when individually varying each of the six parameters of the population model (left) and, alternatively, the two parameters of the (A,γ)(A,\gamma) power law (right) across the indicated range (the 2σ2\sigma range of the NANOGrav priors for the four Gaussian parameters, and the full range of the uniform priors for τf\tau_{\mathrm{f}} and νinner\nu_{\mathrm{inner}}, cf. Table 1). We restrict ourselves to mψ,0>10.5m_{\psi,0}>10.5 here since the 2σ2\sigma range of the NANOGrav prior reaches below the lower limit of the uniform prior used to generate the model library. The five values in each panel are equally spaced, dark blue for the smallest and yellow for the largest, with the other parameters at the centres of the NANOGrav priors; for the power law the range is the centre of its Gaussian prior ±2σ\pm2\sigma. For comparison, we also show our baseline free spectrum, the CURN spectrum of the 2026 NANOGrav 15 yr release [48].

The parameters ψ0\psi_0, mψ,0m_{\psi,0}, μ\mu, and ϵμ\epsilon_\mu mostly set the amplitude, but mψ,0m_{\psi,0} and ϵμ\epsilon_\mu also steepen the spectrum at high frequencies. Across the ranges shown here the characteristic mass mψ,0m_{\psi,0} moves hc2h_{\mathrm{c}}^2 at 2 nHz by more than two orders of magnitude, ψ0\psi_0 and μ\mu by about one and a half, and ϵμ\epsilon_\mu by one, while the slope in the lowest five bins changes by less than one unit. These four parameters are clearly strongly degenerate and hence need galaxy survey priors to be disentangled. The binary lifetime τf\tau_{\mathrm{f}} and the inner hardening index νinner\nu_{\mathrm{inner}} instead bend the spectrum at low frequencies: across their priors they change the slope in the lowest five bins by more than one and a half units [8], and with it hc2h_{\mathrm{c}}^2 at 2 nHz by one and a half to two orders of magnitude. A good fit to the data requires a short lifetime and a shallow inner index.

The final column of Figure 1 shows, instead, how varying the parameters of the much simpler, commonly used power-law description affects the predicted GW background. Clearly, this simpler prescription can imitate the amplitude parameters of the population model, but not the bend induced by the hardening parameters. The full model therefore has in particular more freedom to fit the relatively high second frequency bin, implying less need for an additional cosmological contribution.

Galaxy survey priors

Figure 2 shows the background that the three Gaussian prior sets, cf. Table 1, predict before any fit to the timing data. The NANOGrav and Tomczak–Kormendy priors predict a median spectrum below the measured one, but the survey-average priors come noticeably closer to it. The width of the prediction comes mostly from the uniform priors on the two hardening parameters, except under the NANOGrav priors, whose wide Gaussian priors of the other parameters contribute a similar amount.

Figure 2. The gravitational wave background of the binaries predicted under the NANOGrav (left), Tomczak–Kormendy (middle) and survey-average priors (right) of Table 1, before any fit to the timing data, from binaryfog with 40000 draws from each prior set and one realisation of the background per draw. Solid lines show the median, shaded bands the 1σ1\sigma range over the prior and the realisation scatter. The dashed lines show the 1σ1\sigma range due to the remaining four astrophysical parameters and realisation scatter when τf\tau_f and νinner\nu_{\mathrm{inner}} are held fixed. This range is averaged over their uniform priors and centred on the median of the full distribution. For the Tomczak–Kormendy and survey-average priors these two hardening parameters set most of the width of the band. The violins are the NANOGrav 15 yr CURN free spectrum. Under all three prior sets the median prediction lies below the measured spectrum, in the lowest bin by 0.8 dex in h2ΩGWh^2\Omega_{\mathrm{GW}} for the NANOGrav and Tomczak–Kormendy priors and by 0.2 dex for the survey-average priors.

The NANOGrav priors are centred on fits of the galaxy stellar mass function to the data of Tomczak et al. [62] and on the black hole to bulge mass relations of Kormendy and Ho [59] and of McConnell and Ma [63], but their widths are about four times the uncertainties of these fits, without a stated justification (appendix § 9.§ 1). Since these widths strongly affect the predicted amplitude of the binary background, and hence the preference for new physics, the Tomczak–Kormendy priors take the centres and uncertainties of the fits themselves (appendix § 9.§ 2). We do not claim that the astrophysical parameters are known this precisely: fits to different galaxy catalogues differ by more than these uncertainties [86, 87], and so do the bulge-mass fits of Kormendy and Ho and of McConnell and Ma. Whether such narrow priors are compatible with the measured amplitude is debated [9, 13, 87, 88, 89]. They basically illustrate what the data say about new physics if the astrophysical inputs are as precise as the measurements they are based on. The survey-average priors take the mean and the spread of six determinations of the galaxy stellar mass function published since 2014, and of black hole mass relations brought to a common initial mass function (appendix B.3). Their widths are similar to those of the Tomczak–Kormendy priors, but their centres predict a louder background, by a factor of about two in hch_{\mathrm{c}} (Figure 2).

Distribution of the realisation scatter and accuracy of the surrogate

The distribution of eq. (6) is asymmetric in log⁡10hc2\log_{10} h_{\mathrm{c}}^2, with a heavy tail towards loud signals. A finite number of sources steepens the spectrum and scatters it [64, 90], and ignoring this scatter biases the inferred amplitude [91]. When many sources share a bin the distribution approaches a universal limit [92], which Ali-Haïmoud identified as a reflected map-Airy distribution [39]. In the opposite limit of finite source counts the t-process of refs. [93, 94] offers a good fit, as it is skewed towards large amplitudes. The distribution has been computed for various population models [95], its loud tail traced to the chance of a nearby source [96], its higher moments derived with a low-redshift regularisation [40, 97], and wave interference added [41]. Simplified versions of the integral over realisations in eq. (12) are used in refs. [10, 25, 66, 67, 93, 98, 99, 100].

Here we compare the warp with two other distributions, each matched to the median and the width of the realisations in each bin (the latter defined here as the half-distance between the sixteenth and eighty-fourth percentiles [92]): the log-normal, common in the literature [21, 43, 44, 45], and the map-Airy distribution [39]. Against 2000 realisations of each of 150 holodeck models held out of the fit of § 2.3, the Kolmogorov–Smirnov distance averaged over models and bins is 0.044 for the warp, 0.073 for the log-normal, and 0.094 for the map-Airy distribution. This ordering is opposite to the one of ref. [39], where the map-Airy law fits better than a log-normal distribution at every amplitude.5

In order to resolve the tails of the distributions well, we generate 5000 realisations at the best fit of the binaries alone to the CURN spectrum under the NANOGrav priors. In Figure 3, we show the resulting histograms together with the three analytical parametrisations described above. The width in log⁡10hc2\log_{10} h_{\mathrm{c}}^2 grows from 0.09 in the first bin to 0.35 in the fourteenth, and the distribution becomes increasingly skewed towards loud realisations. In the fourteenth bin, the mean sits a factor of 2.5 above the median [64]. In that bin the Kolmogorov–Smirnov distance, most sensitive to the bulk, is 0.048 for the warp, 0.076 for the log-normal, and 0.110 for the map-Airy law. The 1-Wasserstein distance, the mean distance between quantiles and sensitive to the tails, is 0.016, 0.079, and 0.130 in log⁡10hc2\log_{10} h_{\mathrm{c}}^2. The warp gets the loud tail right, and the other two are too little skewed.

Figure 3. Top panel: The distribution of log⁡10hc2\log_{10} h_{\mathrm{c}}^{2} over 5000 realisations of the holodeck model at the best fit of the binaries alone (NANOGrav 15 yr common uncorrelated red noise spectrum, NANOGrav priors), in the lowest, the eighth, and the highest of the fourteen bins (grey), against a log-normal (purple, dotted), the warp of eq. (7) with the skew c=0.22c = 0.22 of the surrogate at this point (orange, solid) and the map-Airy distribution (dark blue, dash-dotted), each matched to the same median and width. The spike of the warp is the probability it piles up at its floor (§ 2.2). Middle panel: The posterior of the amplitude of the bin in the NANOGrav CURN free spectrum, as a function of the binary amplitude. Bottom panel: The product of the two top panels, the integrand of eq. (12), normalised such that the area under the histogram equals one. The annotated numbers denote Δln⁡L\Delta\ln L, the difference of the logarithm of the area under the histogram and the respective parametrisation of the same colour, corresponding to the difference in the likelihood when using the respective distributions. The warp of eq. (7) is the only one which consistently reproduces the likelihood of the realisations on the (sub-)per cent level across the bins.

The warp is flat below ξ=−1/(2c)\xi=-1/(2c) and misdescribes the faintest realisations. Replacing this cutoff by a constant floor moves the best-fit ln⁡L\ln L for a phase transition by 3×10−63 \times10^{-6}, since such faint realisations lie far below the measured spectrum. Figure 3 shows where the shape of the adopted parametrisation matters in practice. In the lowest and the highest bin, the posterior of the free spectrum hardly varies across the binary distribution, and all three parametrisations result in the same log-likelihood for this bin to within 0.03. In the eighth bin the posterior rises steeply towards louder signals and the integrand is carried by the loud tail. The warp follows the realisations there (Δln⁡L=+0.009\Delta\ln L = +0.009), while the log-normal and the map-Airy law put too little probability into the tail (Δln⁡L=−0.20\Delta\ln L = -0.20 and −0.21-0.21). Summed over all fourteen bins, the warp overestimates ln⁡L\ln L of the binaries alone by 0.13, and the log-normal and the map-Airy law underestimate it by 0.07 and 0.20. At the best fit with a phase transition (see § 5.1) all three agree with the realisations to 0.002.

On models held out of the fit the surrogate reproduces the median spectrum to 0.048 in log⁡10hc2\log_{10} h_{\mathrm{c}}^{2}, the width to four per cent and the skew to 0.05, which is small compared to the width of the measured free spectrum in every bin. The error on the held-out models stops decreasing once about half of the library is used, and splitting the realisations of each model into halves changes the median, the width, and the skew two to eight times less than these errors. Its errors shift ln⁡L\ln L of a held-out model by 0.20 through the median, 0.05 through the width, and 0.03 through the skew, in the median over models, and by −1.2-1.2 to +1.2+1.2 in total for the middle two thirds of the models. While this sounds like a dramatic bias on the following analyses, we remark that a contour is drawn relative to the maximum of its own scan, and the gain GG is a difference of two maxima. A shift of ln⁡L\ln L common to all points of a scan therefore cancels in both, and only the part of the error that varies across a scan can move a contour or a preference. We tested this directly: with the background taken from holodeck instead of the surrogate, the profile likelihood of the phase transition changes by 0.003 inside the 1σ1\sigma region. The shape of the distribution therefore mostly matters in those bins where the measured posterior changes across the spread of the realisations.

Validation against the NANOGrav posterior

Figure 9 in refs. [8, 48] shows NANOGrav’s posterior of the six parameters θ\theta from the five lowest bins of the HD+MP+DP+CURN spectrum, under flat and under NANOGrav priors, with Gaussian processes for the median and the standard deviation of log⁡10hc2\log_{10} h_c^2 in each bin. We repeat this fit with our surrogate and the warp of § 2.2, on the 2026 re-run of the same spectrum [83]. The two pipelines share no code except holodeck. We sample with emcee [101], with 64 walkers for 30000 steps, discarding the first 5000 and thinning by ten, which leaves 160000 samples per prior set, more than a hundred autocorrelation times.

Figure 4 compares our posteriors with the published contours, extracted from the vector graphics of the published figure. We observe that under the NANOGrav priors the two sets of contours lie on top of each other in every panel. For flat priors the published contours are noisier, since they are drawn from an unsmoothed histogram, and we identify visible differences in particular in the row of the inner hardening index νinner\nu_{\mathrm{inner}}, where our posterior prefers shallower values. To quantify these differences, for every point of each published contour we take the distance to the closest point of ours, with each axis scaled to the flat prior range, and the median over points and panels. Under the NANOGrav priors this distance has a median of 1.3% at 1σ1\sigma (1.6% at 2σ2\sigma), about the width of the lines, and every median agrees with the published one to within 0.17 posterior standard deviations. Under flat priors it is 6.4% and 6.7%, mostly because the published contours are derived from an unsmoothed histogram. Drawn the same way, ours give 2.2% and 3.2%, against 2.0% and 1.8% between two halves of our own chain. The excess over this noise comes from the five panels involving νinner\nu_{\mathrm{inner}}, which differ by 3.1% at 1σ1\sigma while the other ten differ by 2.0%. The remaining difference we associate dominantly with a shift of the inner hardening index by a quarter of a posterior standard deviation, with a median of −0.66-0.66 against −0.75-0.75.6 Despite these slight discrepancies in the νinner\nu_{\mathrm{inner}} posterior distribution, we consider the surrogate validated.

Figure 4. Posteriors for the six astrophysical parameters obtained with binaryfog (thick) against those extracted from the vector graphics of figure 9 of refs. [8, 48] (thin), for flat priors over the ranges in table 1 (blue) and the NANOGrav priors (orange). Our fits in this figure and the published ones both use the 2026 HD+MP+DP+CURN free spectrum. The published contours derive from an unsmoothed histogram of 34 bins; pieces spanning fewer than three bins are not drawn here. Contours enclose the 1, 2, and 3σ3\sigma credible regions; the diagonal shows the one-dimensional marginalised posteriors.

Impact of the astrophysical background on new-physics searches

In this section, we turn to the main focus of this work, namely the impact of the SMBHB modelling on new-physics searches with PTAs. We stress that using the surrogate developed within binaryfog is paramount for this purpose: the surrogate evaluates the distribution of the binary background in 0.18 milliseconds on one core, against the 180 seconds holodeck needs for the same prediction. A profile likelihood over the six astrophysical parameters and the parameters of a new-physics model needs of order 10610^6 spectra per scan, i.e. a few minutes with the surrogate and about six CPU years with holodeck. In other words, the surrogate profiles over θ\theta in the time holodeck needs for the realisations of a single θ\theta. We start by identifying which parameter regions of a dark sector phase transition are preferred by the data in § 5.1, by treating the astrophysical parameters as nuisance parameters, and then perform the same exercise for cosmic superstrings, inflationary gravitational waves and domain walls in § 5.2. We compare this to results obtained based on other versions of the free spectrum in § 5.3, and to variations in astrophysical priors in § 5.4.

Dark sector phase transition

A first-order phase transition in a dark sector is among the best-motivated early universe sources of a nanohertz background: a hidden sector that breaks a symmetry at temperatures of MeV to GeV puts the peak of its spectrum in the pulsar timing band, and such transitions have been fitted to the PTA signal since the first hints of evidence for a GW background were announced [21, 22, 23, 27, 29, 37]. We parametrise the phase transition by its strength α\alpha, its inverse duration β/H\beta/H, and the reheating temperature TrehT_{\rm reh}, with the sound-wave spectrum of ref. [29]. Appendix § 8 describes the further modelling choices and the spectral template that we used.

The left panel of Figure 5 shows the profile likelihood in the plane of α\alpha and β/H\beta/H, with TrehT_{\rm reh} profiled over and evaluated without binaries, with the (A,γ)(A,\gamma) power law and with binaryfog, respectively. Above α≃10\alpha\simeq10 the phase transition happens practically in vacuum and its spectrum no longer depends on α\alpha, so the 1σ1\sigma region becomes a horizontal slab. Its range in β/H\beta/H is 22 to 76 without binaries, 20 to 76 with the power law and 12 to 81 with binaryfog, the widening mostly happening as a result of the binary lifetime and the inner hardening index, which bend the binary spectrum at low frequencies. There is a similar broadening of the 1σ1\sigma band at smaller values of α\alpha. Slow phase transitions in that region are therefore far less constrained than a power-law treatment suggests – which translates to less severe tuning requirements for a weakly coupled dark sector to produce such transitions [29]. The 2σ2\sigma region widens even further: with binaryfog, it covers the whole plane outside the hatched region, which is excluded from cosmology because the gravitational waves of the phase transition alone would add more than ΔNeff=0.22\Delta N_{\mathrm{eff}} = 0.22 to the radiation density [102]. The gain G=2.9G = 2.9 is below the 2σ2\sigma threshold of 3.09, and at every point the reheating temperature can move the signal out of the PTA frequency band.

Figure 5. Left panel: The allowed region in the plane of phase transition strength α\alpha and inverse duration β/H\beta/H, with TrehT_{\rm reh} being profiled over. With the binaries treated as a population with realisation scatter under the NANOGrav priors (orange), the preferred regions are significantly larger than with the commonly adopted (A,γ)(A,\gamma) power law [21] (dashed blue) as well as with no binaries included in the fit (dotted blue); in all cases, darker (lighter) colours refer to 1σ1\sigma (2σ2\sigma) confidence regions. The symbols mark the best fit of each treatment, whose spectrum is shown on the right; the maxima lie on a ridge along which ln⁡L\ln L changes by less than 10−410^{-4}. The binaryfog 2σ2\sigma region covers the whole panel outside the hatched region, in which the gravitational waves of the phase transition alone add more than 0.220.22 to NeffN_{\rm eff} [102]. Right panel: The best-fit spectra for a phase transition alone, a phase transition combined with the (A,γ)(A,\gamma) power-law background, and a phase transition combined with the binaryfog population model, compared with the NANOGrav 15 yr CURN spectrum. Also shown is the best fit of the binaries alone (green), with the band containing 90% of its realisations.

In the right panel of Figure 5, we show the three best-fit spectra. They are almost indistinguishable, since the phase transition dominates the signal in all cases. At the binaryfog best fit the four Gaussian parameters sit at the centres of the NANOGrav priors and τf\tau_{\mathrm{f}} at the lower end of its range (see § 5.4). The maximum likelihood is reached along a ridge from (α,β/H)=(55,46)(\alpha, \beta/H) = (55, 46) with Treh=21T_{\mathrm{reh}} = 21 MeV to weaker and slower phase transitions at higher reheating temperatures, along which ln⁡L\ln L changes by less than 10−410^{-4}. The fit of the binaries alone (green), on the other hand, pulls the galaxy stellar mass function away from the centres of the NANOGrav priors and misses the maximum of the posterior in the first bin.

For slow phase transitions, with β/H≲10\beta/H \lesssim10, the peak of the induced GW spectrum moves to about 100 nHz (at Treh≃1T_{\mathrm{reh}} \simeq1 GeV). The binaries then contribute mostly to the low-frequency part of the observed spectrum, in particular the lowest bin at 2 nHz, and the steeper spectrum of the phase transition fits the upper bins. The binaries alone cannot fit the apparent turnover at 3 to 6 nHz (green curve, see also ref. [103]), which is why a steeper additional contribution is preferred. This holds in particular for the simple power-law description, but to a somewhat lesser degree for the bent spectra expected in population models. Both fits place the binary lifetime τf\tau_{\mathrm{f}} at the lower end of its range, 0.1 Gyr. In § 5.4 we discuss how the preference for a phase transition depends on the astrophysical priors and population models.

Both the favoured region and the preference depend on the spectral template that is used for the phase transition. The best fits lie along a ridge that reaches α≫1\alpha\gg1, while the sound-wave template of appendix § 8 rests on simulations of phase transitions with α\alpha up to about 0.1 [104]. If the signal is to be interpreted in terms of a strong phase transition, the relevant template is instead the one of ref. [105], from simulations of strongly supercooled phase transitions that include the expansion of the universe during the phase transition. For dissipative shells and α→∞\alpha\to\infty it no longer depends on α\alpha, so the question of how strongly the data prefer a strong phase transition has a quantitative answer that does not rest on a template outside its range. Scanning again β/H\beta/H with TrehT_{\mathrm{reh}} profiled over for this template, the preference for a strong phase transition is 1.9σ1.9\sigma under the NANOGrav priors and 3.2σ3.2\sigma with the (A,γ)(A,\gamma) power law. Its 1σ1\sigma interval spans the region from β/H=5.8\beta/H = 5.8 to 14.9 and from 6.2 to 14.3, respectively.

Cosmic superstrings, inflationary gravitational waves and domain walls

We repeat the study of phase transitions for the three further cosmological sources that we consider in this work (see appendix § 8). We chose them as representatives of the many new-physics models that have been discussed in the literature in this context, because their spectra feature three very distinct shapes. The spectrum of cosmic superstring loops is close to a power law across the entire frequency band. Inflationary gravitational waves follow a power law of free blue tilt, ΩGW∝fnt\Omega_{\mathrm{GW}} \propto f^{n_{\mathrm{t}}}, which turns over to fnt−2f^{n_{\mathrm{t}}-2} above the frequency of the mode that re-entered the horizon at reheating. Domain walls annihilating into a dark sector give a peaked spectrum, rising as f3f^3 and falling as f−bf^{-b} with b≤1b \le1 above a peak that can lie inside the band. For the strings we scan the tension GμG\mu and profile over the intercommutation probability PP. For inflation we scan the tensor index ntn_{\mathrm{t}} and profile over the tensor-to-scalar ratio rr and the reheating temperature TrhT_{\mathrm{rh}}, since rr and ntn_{\mathrm{t}} are strongly degenerate in the band (cf. figure 5 of ref. [21]). For the domain walls we scan the annihilation temperature T∗T_\ast and profile over bb, the sharpness cc of the peak and ΔNeff\Delta N_{\mathrm{eff}}, which measures the energy in the walls. The six astrophysical parameters are profiled over as before.

Figure 6 shows the gravitational wave signals predicted by the three models. Since the cosmic string spectrum stays close to a power law, the (A,γ)(A,\gamma) power law imitates it about as well as the full population model. The 2σ2\sigma interval of log⁡10Gμ\log_{10}G\mu nevertheless grows by 75% when profiling with binaryfog, from 1.2 to 2.1 decades, almost entirely at the low-tension end, which moves down by a factor of 6 as the binaries carry more of the signal. This lower end rests on a small margin, since the gain G=2.04G = 2.04 lies only marginally above the one-parameter threshold of 2.00, and most of the widening disappears when PP is allowed below 10−310^{-3}. At the best fit point we find an intercommutation probability of P=1.3×10−3P = 1.3 \times10^{-3}, close to the lower end of its prior, where the signal is strongest. In the figure, we still show results based on this prior, for better comparability with ref. [21], but note that the preferred parameter space opens up when decreasing the lower prior boundary of PP, allowing arbitrarily small values of GμG\mu at 2σ2\sigma.

Figure 6. Comparison of the impact of a population-based astrophysical background instead of a power law on the preferred parameter regions for three distinct new-physics models. The solid curves in the top panels indicate the total signal for a given model parameter as marked by a triangle in the lower panels; their dashed and dotted counterparts indicate the astrophysical and the new-physics contributions, with the shaded bands holding 90% of the realisations the respective population would produce. Magenta curves belong to the best fit, purple (green) curves to the lower (upper) end of the 2σ2\sigma-favoured region in parameter space; these are the ends and the best fit of the binaryfog interval. In the lower panels, we plot the profile likelihood ratio, with the dotted horizontal lines indicating the 1σ1\sigma and 2σ2\sigma thresholds. The shaded bands represent the 2σ2\sigma intervals resulting from the NANOGrav 15 yr common uncorrelated red noise spectrum, when using the power-law (light blue) or binaryfog prescription (orange) for the SMBHB contribution, respectively.

For inflationary gravitational waves, the tensor index is bounded at 2σ2\sigma between nt=1.32n_{\mathrm{t}} = 1.32 and 4.42 with the power law and between 1.29 and 5.20 with binaryfog, a widening of the interval by 26% at the steep end. Such a highly blue-tilted inflationary spectrum contributes only in the upper bins and leaves the bend at low frequencies to the binaries. At the best fit, nt=2.5n_{\mathrm{t}} = 2.5 and Trh=0.34T_{\mathrm{rh}} = 0.34 GeV, and the binary contribution is negligible. Within the 2σ2\sigma region the tensor-to-scalar ratio falls as ntn_{\mathrm{t}} grows, from r≈10−4r \approx10^{-4} at the lower end to r≈10−35r \approx10^{-35} at its upper end.

The annihilation temperature T∗T_\ast of the domain walls is affected most among the new-physics scenarios we consider here: with the power law it is pinned to a window a tenth of its prior range wide, while with binaryfog it remains unconstrained at 2σ2\sigma. The domain wall spectrum turns over inside the frequency band in a way similar to a population with efficient environmental hardening. Since its amplitude prior is limited by ΔNeff≤0.39\Delta N_{\mathrm{eff}} \le0.39 the two contributions thus become interchangeable beyond the 1σ1\sigma band. At the best fit point, ΔNeff=0.39\Delta N_{\mathrm{eff}} = 0.39 and b=0.50b = 0.50 sit at the ends of their prior ranges, where the signal is largest and falls most slowly above the peak, with c=2.47c = 2.47. Lowering the upper bound from ΔNeff≤0.39\Delta N_{\mathrm{eff}} \le0.39 (as used by NANOGrav) to ΔNeff≤0.22\Delta N_{\mathrm{eff}} \le0.22 (the same bound as we apply to the phase transition), domain walls are no longer preferred; concretely, we find G=1.6G = 1.6 for the power law and 0.6 for binaryfog in that case.

In comparison, the preference with the power law is 2.9σ2.9\sigma for cosmic strings, 2.8σ2.8\sigma for inflation, and 2.1σ2.1\sigma for domain walls, which drops to 1.5σ1.5\sigma, 1.5σ1.5\sigma, and 0.7σ0.7\sigma with binaryfog. We observe that the preference drops by a similar amount for every model – which is directly related to the main phenomenological property of the SMBHB population spectrum stressed several times above, namely the flexibility of bending the spectrum compared to a simple power law. Phase transition and inflationary contributions, in particular, generally reach the same gain since they both fit an excess (on top of the binaries) in the upper bins.

Let us conclude by mentioning that, for the (A,γ)(A,\gamma) power law, our 1σ1\sigma intervals for the three models shown in Figure 6 are very similar to NANOGrav’s 1σ1\sigma credible intervals [21]: for cosmic superstrings we find log⁡10Gμ\log_{10} G\mu from −12.09-12.09 to −11.52-11.52 against their −12.09-12.09 to −11.49-11.49, for inflation, with rr as the scanned parameter as in ref. [21], log⁡10r\log_{10} r from −23.9-23.9 to −8.2-8.2 against −23.0-23.0 to −7.2-7.2, and for domain walls log⁡10(T∗/GeV)\log_{10}(T_\ast/\mathrm{GeV}) from −1.09-1.09 to −0.64-0.64 against −1.27-1.27 to −0.67-0.67. Differences of this size are expected between a profile likelihood analysis and Bayesian credible intervals, and between our spectrum and the 2023 Hellings–Downs spectrum they used, which predates the correction of ref. [48].

Comparison of different NANOGrav free spectra

The results presented above are all based on the same free spectrum, CURN of the 2026 release with deterministic dispersion measure offsets. Here we repeat the phase transition fit with five additional versions, and summarise the results in Figure 7: the CURN and HD spectra of the 2024 PTArcade record [85] and of the 2026 release [83], all with deterministic dispersion measure offsets, the 2026 HD+MP+DP+CURN spectrum, the most conservative of the available noise models, and an HD spectrum from NANOGrav’s 2023 record [106] in which the dispersion measure variations are modelled as a Gaussian process to capture the many effects that may delay the radio pulses of a given pulsar. Overall, the release date and the correlation pattern are found to matter less than the noise model and the set of components.

Figure 7. Six versions of the NANOGrav 15 yr free spectrum, each fitted with the binaries as the (A,γ)(A,\gamma) power law (blue) and as the binaryfog background (orange). Left panel: The band in β/H\beta/H for a strong phase transition, in the limit α→∞\alpha\to\infty, solid for 1σ1\sigma and pale for 2σ2\sigma. An arrow marks a band that continues past the end of the scanned range. Right panel: The preference for a phase transition, i.e. the gain GG of eq. (14) converted as described in § 3.2. The shaded row is the default spectrum used in the rest of this paper.

The choice between HD and CURN only marginally affects the preference and the band in β/H\beta/H in either release, given the current state of the data. In every version the preference with binaryfog is about 1σ1\sigma lower than with the power law, and for HD+MP+DP+CURN only an extremely mild preference of 0.6σ0.6\sigma remains. With the Gaussian-process model of the dispersion measure variations the phase transition has to be weaker, at larger β/H\beta/H, and is less preferred.7 With binaryfog the 2σ2\sigma band in β/H\beta/H is comparable across the 2026 products.

The preference therefore depends on the modelling of the dispersion measure variations as strongly as on the binary background.

Let us mention that the 2023 posteriors carry a numerical artefact for amplitudes far below the measured signal: their density grids run down to log⁡10(ρi/s)=−15.1\log_{10}(\rho_i/\mathrm{s})=-15.1, but the Markov chains never visited values below about −9-9, resulting in a numerical floor with isolated spikes up to 31 in ln⁡L\ln L above it. The tails of the distribution thus cannot be used far from the measured values, as already pointed out in refs. [27, 29]. The 2026 posteriors, in contrast, only keep the range that is actually populated by the chains.

All the posteriors explored here were obtained with NANOGrav’s log-uniform priors on the amplitude and the slope of the intrinsic red noise of each pulsar. Dai et al. [16] showed that, since the auto-correlations of a pulsar only measure the sum of the background and its intrinsic red noise, the prior volume grows with the background amplitude and pulls the inferred amplitude up: at a fixed slope of 13/313/3 by 0.06 to 0.2 in log⁡10A\log_{10} A compared with other choices of the noise priors, and with a free slope by 0.3 in log⁡10A\log_{10} A. With those other pulsar-intrinsic red noise priors the Bayes factors for new-physics sources against binaries fall to order unity. Their free spectra are not public. To estimate the size of the effect on our results, we lowered log⁡10ρi\log_{10}\rho_i of the common uncorrelated red noise spectrum by either 0.1 (or 0.2) in all bins. The preference for a phase transition then falls from 2.8σ2.8\sigma to 2.7σ2.7\sigma (or 2.6σ2.6\sigma) with the (A,γ)(A,\gamma) power law, and from 1.5σ1.5\sigma to 1.4σ1.4\sigma (or 1.2σ1.2\sigma) under the NANOGrav priors. Under the Tomczak–Kormendy priors it drops from 3.3σ3.3\sigma to 2.7σ2.7\sigma (or 2.2σ2.2\sigma), and from 1.7σ1.7\sigma to 1.3σ1.3\sigma (or 1.0σ1.0\sigma) under the survey-average priors (§ 5.4). The best fit with a phase transition is unchanged by the shift, since the transition takes up the difference, while the fit of the binaries alone improves, and with it the tension between the binaries and the astrophysical priors. The steeper spectrum that the alternative noise priors also give is not included in this estimate; it would further favour the binaries.

Dependence on the astrophysical priors

The preferences for a phase transition stated in sections 5.1 and 5.2 depend strongly on the astrophysical priors. With flat priors as the most extreme example, 99% of the scanned plane of α\alpha and β/H\beta/H would lie inside the 1σ1\sigma region and the preference for a phase transition drop to 0.2σ0.2\sigma, since the binaries can then be tuned to reproduce the measured spectrum. For the three Gaussian prior sets that do take into account the information contained in galaxy surveys, Table 2 summarises the preference for all four new-physics models, with the (A,γ)(A,\gamma) power law and with binaryfog, on the three free spectra of § 3.3. With CURN and HD free spectra the Tomczak–Kormendy priors raise the preference for new physics slightly above the respective power-law values and well above the respective values under the NANOGrav priors, while the 2σ2\sigma intervals of the model parameters change only slightly. Under our recommended survey-average priors, the preferences drop again, to values similar to those found when adopting the NANOGrav priors. Within an HD+MP+DP+CURN fit the preferences drop to between 0.2σ0.2\sigma and 1.6σ1.6\sigma, because the additional degrees of freedom in the overlap reduction function broaden the uncertainty of the measured spectrum sufficiently for even a faint binary background to fit the data.

modelfree spectrum(A,γ)(A,\gamma) power lawbinaryfog NANOGravbinaryfog Tomczak–Kormendybinaryfog survey average
phase transition (k=3)(k=3)CURN2.8σ2.8\sigma1.5σ1.5\sigma3.3σ3.3\sigma1.7σ1.7\sigma
HD2.8σ2.8\sigma1.3σ1.3\sigma2.9σ2.9\sigma1.5σ1.5\sigma
HD+MP+DP+CURN1.4σ1.4\sigma0.6σ0.6\sigma1.2σ1.2\sigma0.4σ0.4\sigma
cosmic superstrings (k=2)(k=2)CURN2.9σ2.9\sigma1.5σ1.5\sigma3.4σ3.4\sigma1.8σ1.8\sigma
HD2.9σ2.9\sigma1.3σ1.3\sigma3.0σ3.0\sigma1.5σ1.5\sigma
HD+MP+DP+CURN1.6σ1.6\sigma0.8σ0.8\sigma1.4σ1.4\sigma0.5σ0.5\sigma
inflation (k=3)(k=3)CURN2.8σ2.8\sigma1.5σ1.5\sigma3.3σ3.3\sigma1.7σ1.7\sigma
HD2.8σ2.8\sigma1.4σ1.4\sigma2.9σ2.9\sigma1.5σ1.5\sigma
HD+MP+DP+CURN1.4σ1.4\sigma0.6σ0.6\sigma1.2σ1.2\sigma0.4σ0.4\sigma
domain walls (k=4)(k=4)CURN2.1σ2.1\sigma0.7σ0.7\sigma2.6σ2.6\sigma1.1σ1.1\sigma
HD2.3σ2.3\sigma0.7σ0.7\sigma2.4σ2.4\sigma1.0σ1.0\sigma
HD+MP+DP+CURN1.1σ1.1\sigma0.4σ0.4\sigma0.9σ0.9\sigma0.2σ0.2\sigma

Table 2. Preference for each new-physics model over the binaries alone, expressed as one-sided significance from the gain GG in eq. (14); kk is the number of free parameters of the model. The binaries are modelled by the (A,γ)(A,\gamma) power law under its Gaussian prior, or by binaryfog under the NANOGrav, the Tomczak–Kormendy, or the survey-average priors of Table 1. The three free spectra are those of the 2026 NANOGrav 15 yr release with deterministic dispersion measure offsets: the common uncorrelated red noise (CURN) used as reference choice throughout this paper, the Hellings–Downs (HD) spectrum, and the Hellings–Downs spectrum fitted together with a monopole (MP), a dipole (DP), and CURN.

The preferences stated in Table 2 are hence a result of two effects: the prior penalty in fits with only the binaries, and the improved goodness-of-fit when considering both contributions. Under the NANOGrav priors the fit without a phase transition signal puts the binary lifetime at 0.1 Gyr, the fastest hardening of NANOGrav’s library [8], and moves the normalisation and the characteristic mass of the galaxy stellar mass function by 1.3 and 1.0 standard deviations from their prior centres. Of the gain G=2.9G=2.9 in the combined fit, a gain of 1.6 results from the improved goodness-of-fit and the remaining 1.3 come from the prior penalty, against 3.5 and 3.0 when using the (A,γ)(A,\gamma) power law. Under the Tomczak–Kormendy priors the fit of the binaries alone moves mψ,0m_{\psi,0} by 1.9, ψ0\psi_0 by 1.2 and μ\mu by 0.8 standard deviations, its ln⁡L\ln L falls from −1.38-1.38 to −6.51-6.51, and the gain rises to G=8.0G=8.0 (3.3σ3.3\sigma), of which 5.0 come from the data and 3.0 from the prior penalty: these priors predict too faint a background, which any additional power, astrophysical or cosmological, would relieve. The survey-average priors predict a louder background (see Figure 2), which the binaries alone reach close to the prior centres, so that the penalty disappears despite the narrow widths and all preferences stay below 2σ2\sigma.

In all of the considered prior sets of Table 1, the hardening parameters τf\tau_{\mathrm{f}} and νinner\nu_{\mathrm{inner}} are modelled by uniform distributions. Figure 8 shows the preference for a phase transition with τf\tau_{\mathrm{f}} and νinner\nu_{\mathrm{inner}} fixed and the other four astrophysical parameters profiled over. In each of the panels, representing the three different prior choices, the preference is dominated by how well the binaries alone can fit the data. The data prefer fast hardening in order to match the required signal amplitude, τf=0.1\tau_{\mathrm{f}}=0.1, 0.7, and 0.2 Gyr at the best fit points (cf. the posterior of τf\tau_{\mathrm{f}} in Figure 4), and moderate turnovers of the GW spectrum due to the hardening process having a mild slope νinner=−0.45\nu_{\mathrm{inner}}=-0.45, −1.0-1.0, and −0.75-0.75 (see also Figure 1). At the centres of the NANOGrav priors the background at τf=11\tau_{\mathrm{f}}=11 Gyr is more than an order of magnitude fainter than at 1 Gyr. Hence, while a lower bound on τf\tau_{\mathrm{f}} of 1 Gyr leaves the preference almost unchanged (1.6σ1.6\sigma, 3.3σ3.3\sigma and 1.8σ1.8\sigma), raising the lower bound to 3 and 6 Gyr leads to an increased preference of 1.8σ1.8\sigma and 2.3σ2.3\sigma, respectively, under the NANOGrav priors. Under our recommended survey-average priors, the preference for a phase transition exceeds the 2σ2\sigma and 3σ3\sigma thresholds for hardening timescales larger than 2.0 Gyr and 6.5 Gyr, respectively.

Figure 8. Preference for a dark sector phase transition over the binaries alone on the CURN spectrum, at fixed binary lifetime τf\tau_f and inner hardening index νinner\nu_{\mathrm{inner}}, with the other four astrophysical parameters profiled under the NANOGrav (left), Tomczak–Kormendy (middle) and survey-average (right) priors of Table 1. The colour gives the local one-sided significance from the gain GG of eq. (14) in steps of 0.5σ0.5\sigma, and the black lines mark 22, 33, 44 and 5σ5\sigma. The stars mark the best fit of the binaries alone with all six parameters free, where the preference is 1.5σ1.5\sigma, 3.3σ3.3\sigma and 1.7σ1.7\sigma, respectively.

The above discussion illustrates that the preference for a new-physics contribution to the PTA data depends as much on where the astrophysical priors are centred as on their widths; a more robust future determination of such a signal component will thus require a significant gain in both accuracy and precision from galaxy surveys and merger simulations. The other conclusion is that prior boundary effects on the hardening parameters have a particularly significant impact on the preference for new physics. This motivates further studies as to how SMBHBs overcome the final parsec [55, 66], and in particular how small timescales τf\tau_f can realistically be achieved.

Discussion of assumptions and caveats

One of the basic assumptions in the population model used here is that the binaries’ orbits are circular. Eccentricity shifts power to higher harmonics and hardens binaries faster, which suppresses the signal at the bottom of the band, as environmental hardening does, and in some models enhances it nearer the top [107]. Raidal et al. [108] find the two effects to be near-degenerate in the spectrum, although the degeneracy is broken by the spectral fluctuations and the correlations between frequency bins, which our free-spectrum likelihood drops. They obtain a mean eccentricity of 0.83±0.050.83 \pm0.05 at 2 nHz for evolution by gravitational wave emission alone, while with environmental effects circular binaries remain only mildly disfavoured. The eccentricity distribution is degenerate with the binary mass function rather than with the stellar density of the hosts, since the turnover frequency depends only weakly on the latter [109]. Our library already spans turnovers from below to above the PTA frequency band, but strongly eccentric binaries with fast hardening and large merger rates could give flat and loud spectra outside it. Including them would need a change to holodeck.

Another choice is that we treat the background as stochastic. We thus do not separate the loudest binaries as resolvable sources, which a joint search can do [110] and which biases a spectral fit otherwise [111]. The same loud binaries make the background anisotropic [45, 112]. The realisation statistics themselves cannot be measured by a pulsar timing array [113], so they have to come from a population model. Given the lack of evidence for continuous wave signals [114] and only modest chances of observing other signs of discrete sources at the current quality of PTA data [45], we hence do not expect any relevant biases from treating the GW background as stochastic.

Like holodeck, we moreover assume that every binary emits at an exact multiple of the inverse observation time [115], neglect the correlations of the realisations between bins, and use one parametrisation of the population. Others exist, for instance a double Schechter mass function with an Illustris merger rate [86, 116], used with an evolving black hole mass relation in ref. [14]. While the amount of additional small-mass galaxies as described by a double Schechter mass function only has a minor effect on the PTA signal (cf. Figure 9 in the appendix), an evolving black hole mass relation may affect our results by increasing the signal source counts at large redshift, and hence lead to somewhat stronger binary signals. As nHz GWs themselves are currently one of the most precise probes of this relation [14, 86], the latter caveat should rather be understood as a source of uncertainty, similar to the unknown details of binary hardening, which requires more detailed studies.

Figure 9. Left panel: The six galaxy stellar mass functions from refs. [86, 134, 135, 136, 137, 138], multiplied by (m⋆/1011M⊙)1.83(m_\star/10^{11}M_\odot)^{1.83}, the contribution of each decade of stellar mass to hc2h_c^2, in the redshift bin of each survey that contains z=0.5z = 0.5 (z=0z = 0 for GAMA DR4), and the single Schechter function of eq. (4) at z=0.5z = 0.5 at the centres of the three Gaussian prior sets (dashed). The area under each curve between the grey bands is the hc2h_c^2-weighted number density nGWn_{\rm GW}. Right panel: The values of ψ0\psi_0 and mψ,0m_{\psi,0} fitted to each survey (coloured symbols), with the 1σ1\sigma (thick) and 2σ2\sigma (thin) ellipses of the three prior sets, drawn without correlation as they enter the fits. The 2σ2\sigma ellipse of the NANOGrav priors extends beyond the panel. Along each grey line nGWn_{\rm GW} is constant (as indicated by the labels, the lines are 0.2 dex apart).

The free-spectrum likelihood treats the frequency bins as independent, in the same way as NANOGrav’s astrophysical analysis does. Normalising-flow likelihoods go beyond this [70, 117], which mostly matters for eccentric populations [108]. Taking into account these correlations is beyond the scope of this work. In fact, it does not seem required at the current quality level of pulsar timing data, since the bins of the free spectrum appear to be close to uncorrelated, which allowed the factorised likelihood of ceffyl in the first place [74]. Within a given bin, eq. (12) treats the timing residuals as Gaussian for each realisation and averages over the realisations. This is accurate for a single pulsar [41], and for many pulsars it requires the effective number of sources in a bin to be large compared to a third of the number of pulsars [39]. For the 67 pulsars of NANOGrav and the source counts of § 4.3 this holds only in the three lowest bins, so the treatment of the higher bins is the least controlled step of our likelihood (just as of NANOGrav’s).

Conclusions

Supermassive black hole binaries (SMBHBs) are the expected source of the nHz gravitational wave background, but there are several well-motivated new-physics scenarios that would add, possibly subdominant, cosmological contributions. In this work we have explored how the modelling of the astrophysical SMBHB background affects any conclusions about those new-physics contributions. We also presented a new tool binaryfog that we developed specifically for this purpose, allowing us to compute the astrophysical background based on state-of-the-art models with unprecedented speed, without compromising on accuracy.

We stress that runtime was a bottleneck in previous analyses searching for new physics in pulsar timing array data, which therefore largely left out the SMBHB contribution entirely or roughly approximated it by a simple phenomenological power law. The actual prediction from population models, however, is a distribution of bent spectra that depends on at least six uncertain astrophysical parameters. binaryfog returns the median, width, and skew of the resulting distribution in each frequency bin in 0.2 milliseconds, a million times faster than holodeck, which previous analyses had to rely on. It reproduces held-out holodeck models to 0.048 in log⁡10hc2\log_{10} h_{\mathrm{c}}^{2}, its warped Gaussian describes the realisations better than the log-normal or map-Airy distributions that have previously been promoted in this context, and it fully reproduces NANOGrav’s posterior of the population parameters. It is this speed-up that makes it possible to profile over the six astrophysical parameters in any new-physics fit.

We find that under the same priors as in the new-physics analysis by NANOGrav [21], the more realistic description of the binary population adopted here lowers the preference for a dark sector phase transition to 1.5σ1.5\sigma, compared to 2.8σ2.8\sigma when assuming a simple (A,γ)(A,\gamma) power law. The preference for signal contributions from cosmic superstrings, inflationary gravitational waves, or domain walls drops in a very similar way.

We have discussed that the astrophysical parameters which most affect how well the SMBHBs can fit the measured spectrum are the binary lifetime and the inner hardening index, which bends the spectrum at low frequencies. Changing the astrophysical priors, in particular for these parameters, therefore strongly affects the numbers quoted in the previous paragraph. The NANOGrav priors are up to four times wider than the fits they are based on, and when instead using the widths of those fits the preference for a phase transition rises to 3.3σ3.3\sigma. We also presented updated astrophysical priors in this work, based on newer measurements published since 2014 (for a detailed discussion, see appendix § 9.§ 3). With these priors, which appear to be closer to representing the current state of knowledge, the preference for a phase transition drops to a very mild 1.7σ1.7\sigma. Lowering the spectrum by 0.1 to 0.2 in log⁡10ρ\log_{10}\rho (the square root of the timing-residual power), as the intrinsic pulsar noise priors may require [16], further decreases the preference to 1.0σ1.0\sigma. This clearly demonstrates that any preference for new physics from pulsar timing arrays must be quoted along with the astrophysical priors (and the free spectrum it rests on).

In assessing these findings, it is important to realise that in a fit of the binaries alone the data favour binaries that merge within about 2 Gyr after their host galaxies have merged. Including a phase transition in the fit, on the other hand, the data hardly constrain the binary lifetime or the inner hardening index (§ 5.4). If binaries turn out to require longer than about 6.5 Gyr to merge, the preferred binary background becomes fainter and the preference for a phase transition exceeds 3σ3\sigma under our recommended updated prior set. Whether the gravitational wave background is explained by the binaries alone or with an additional new-physics contribution thus remains tightly coupled to a resolution of the final parsec problem.

In the future, better determinations of the galaxy stellar mass function, of the black hole to bulge mass relation, and of how binaries harden will be decisive in settling whether the data describing the nHz gravitational wave background prefer a cosmological contribution. We have designed binaryfog to help with this task. It takes as input any new-physics spectrum as a Python function, along with any choice of astrophysical priors, and profiles over the binary population in minutes on a laptop, so that global fits of new-physics models to pulsar timing data can incorporate a realistic astrophysical background. binaryfog is straightforward to use, cf. appendix § 10, and available on GitHub [118].

Acknowledgments

We thank William Lamb, Kayhan Gültekin and David Wright for discussions of holodeck, PTArcade, and ceffyl, and Shreyas Tiruvaskar for sharing his version of holodeck with us. We are grateful to the NANOGrav collaboration for releasing the model files, the freespectrum data products, and the corrected versions of both. CT gratefully acknowledges support from the Alexander von Humboldt Foundation, through a Feodor Lynen Research Fellowship, as well as from the Spanish National Grant PID2025-171343NB-C55 and from Generalitat Valenciana through the grant CIPROM/22/69. TB is supported through a FRIPRO grant of the Norwegian Research Council (project ID 353561 ‘DarkTurns’), and thanks the University of Canterbury for hospitality and support in the form of a Visiting Erskine Fellowship. CG acknowledges support from the Marsden Fund Council from Government funding administered by the Royal Society Te Apārangi (grants MFP-UOA2131 and MFP-UOA2531). The computations for this work were performed on the local SOM clusters at the Instituto de Física Corpuscular (IFIC). The authors used Claude Opus 5.5 for assistance with the Python code. We have checked and take full responsibility for the results.

Gravitational wave spectra of the new-physics models

We study four different cosmological contributions to the nanohertz signal. This appendix briefly summarises the spectral shape for each of them, and how it is determined. For the inflationary, domain wall, and cosmic superstring contributions, we use the implementations and prior ranges provided in the online repository [77] belonging to the NANOGrav 15 yr search for new-physics signals [21], and for the phase transition interpretation we use the TransitionListener [119] implementation (see also ref. [29]). We refer to these works for further information on the individual contributions and an in-depth discussion of the underlying model assumptions.

Dark sector phase transition

A first-order phase transition in a dark sector proceeds by the nucleation of bubbles, which then expand and source fluid motion in the dark plasma. Here we assume the rather generic case of bubbles which reach a terminal velocity (e.g. due to the presence of gauge bosons obtaining a mass in the transition, resulting in velocity-dependent friction terms [120]), so sound waves dominate the gravitational wave production. We keep that contribution alone and set the efficiencies of wall collisions and of turbulence to zero. The template is based on simulations [104, 121], used exactly as in ref. [29],

h2ΩGWPT(f)=Rh2AswK2YswRH∗S~(f),h^2\Omega_{\mathrm{GW}}^{\mathrm{PT}}(f)=\mathcal{R}h^2 A_{\mathrm{sw}} K^2 \mathcal{Y}_{\mathrm{sw}} RH_* \widetilde{S}(f),
S~(f)=N(ff2)3[1+(ff1)2]−1[1+(ff2)4]−1,\widetilde{S}(f)=N\left(\frac{f}{f_2}\right)^3\left[1+\left(\frac{f}{f_1}\right)^2\right]^{-1}\left[1+\left(\frac{f}{f_2}\right)^4\right]^{-1},
N=22π[1+f22f12+2f2f1].(15)N=\frac{2\sqrt{2}}{\pi}\left[1+\frac{f_2^2}{f_1^2}+\sqrt{2}\frac{f_2}{f_1}\right]. \tag*{(15)}

a doubly broken power law rising as f3f^3 below f1f_1, as ff between the breaks, and falling as f−3f^{-3} above f2f_2. The two break frequencies are given by

f1≃0.2a∗H∗RH∗,f2≃0.5Δw−1a∗H∗RH∗,Δw=∣vw−cs∣max⁡(vw,cs),(16)f_1\simeq0.2\frac{a_*H_*}{RH_*},\qquad f_2\simeq0.5\Delta_{\mathrm{w}}^{-1}\frac{a_*H_*}{RH_*},\qquad\Delta_{\mathrm{w}}=\frac{\lvert v_{\mathrm{w}}-c_{\mathrm{s}}\rvert}{\max(v_{\mathrm{w}},c_{\mathrm{s}})}, \tag*{(16)}

where RH∗RH_* is the mean bubble separation at percolation in units of the Hubble radius and a∗H∗a_*H_* the Hubble rate at percolation redshifted to today. Three further factors set the amplitude: the normalisation Asw≃0.11A_{\mathrm{sw}}\simeq0.11 measured in lattice simulations, the fraction KK of the released energy that ends up in bulk fluid motion, and the factor Ysw\mathcal{Y}_{\mathrm{sw}} by which a finite sound wave lifetime suppresses the signal,

K=0.6κsw(α,vw)1+αα,Ysw=min⁡(1,τshH∗),τshH∗≃RH∗3K/4,(17)K=0.6\frac{\kappa_{\mathrm{sw}}(\alpha,v_{\mathrm{w}})}{1+\alpha}\alpha,\qquad\mathcal{Y}_{\mathrm{sw}}=\min(1,\tau_{\mathrm{sh}}H_*),\qquad\tau_{\mathrm{sh}}H_*\simeq\frac{RH_*}{\sqrt{3K/4}}, \tag*{(17)}

with the efficiency factor κsw\kappa_{\mathrm{sw}} taken from ref. [122] and the factor 0.6 found in the Higgsless simulations of ref. [121]. The amplitude of the spectrum is redshifted by a factor

Rh2=Ωγh2(h0hreh)4/3grehgγ,(18)\mathcal{R}h^2=\Omega_\gamma h^2\left(\frac{h_0}{h_{\mathrm{reh}}}\right)^{4/3}\frac{g_{\mathrm{reh}}}{g_\gamma}, \tag*{(18)}

with Ωγh2=2.473×10−5\Omega_\gamma h^2=2.473\times10^{-5}, gγ=2g_\gamma=2, and h0=3.93h_0=3.93, where hh counts entropy degrees of freedom and gg energy ones. We take vw=1v_{\mathrm{w}}=1, cs=1/3c_{\mathrm{s}}=1/\sqrt{3} and two light dark sector degrees of freedom beyond the Standard Model bath, so that greh=gSM(Treh)+2g_{\mathrm{reh}}=g_{\mathrm{SM}}(T_{\mathrm{reh}})+2 and likewise for hrehh_{\mathrm{reh}}. The three parameters scanned in § 5.1 are the transition strength α\alpha, the inverse duration β/H\beta/H, and the reheating temperature TrehT_{\mathrm{reh}}, which fixes a∗H∗a_*H_* and hence where the spectral breaks appear today. We profile the reheating temperature between 1 MeV and 10 GeV, where the lower bound is motivated by bounds from big bang nucleosynthesis [123]. The inverse duration enters through the bubble separation,

βH=(8π1−Pf(T∗))1/3max⁡(cs,vw)RH∗,(19)\frac{\beta}{H}=\left(\frac{8\pi}{1-P_{\mathrm{f}}(T_*)}\right)^{1/3}\frac{\max(c_{\mathrm{s}},v_{\mathrm{w}})}{RH_*}, \tag*{(19)}

evaluated at the percolation criterion 1−Pf=0.291-P_{\mathrm{f}}=0.29, with PfP_{\mathrm{f}} the fraction of space still in the false vacuum. We keep the factor 0.29 due to the argument presented in ref. [119], noting that it is often omitted in the literature, see for instance ref. [104].

Dark sector domain walls

A broken discrete symmetry leaves a network of domain walls whose energy density grows relative to the surrounding radiation bath until the network annihilates at a temperature T∗T_*. The bath the network annihilates into is here assumed to be a secluded dark sector, instead of the photon bath. We take the model file dw_ds of ref. [77] unchanged, where

h2ΩGWdw(f)=h2πϵgw960αdw2g∗(heqh∗)4/3T04Mpl2H02S(ff0),withαdw=ΔNeffh∗4/313.6g∗,h^2\Omega_{\mathrm{GW}}^{\mathrm{dw}}(f)=h^2\frac{\pi\epsilon_{\mathrm{gw}}}{960}\alpha_{\mathrm{dw}}^2g_*\left(\frac{h_{\mathrm{eq}}}{h_*}\right)^{4/3}\frac{T_0^4}{M_{\mathrm{pl}}^2H_0^2}S\left(\frac{f}{f_0}\right),\qquad\text{with}\quad\alpha_{\mathrm{dw}}=\frac{\Delta N_{\mathrm{eff}}h_*^{4/3}}{13.6g_*},
S(x)=(3+b)c(bx−3/c+3xb/c)c,f0=(heqh∗)1/3(π2g∗90)1/2T0T∗Mpl.(20)S(x)=\frac{(3+b)^c}{\left(bx^{-3/c}+3x^{b/c}\right)^c},\qquad f_0=\left(\frac{h_{\mathrm{eq}}}{h_*}\right)^{1/3}\left(\frac{\pi^2g_*}{90}\right)^{1/2}\frac{T_0T_*}{M_{\mathrm{pl}}}. \tag*{(20)}

Here ϵgw=0.7\epsilon_{\mathrm{gw}}=0.7 is the efficiency with which the annihilating network radiates, as measured in the simulations of ref. [124], αdw\alpha_{\mathrm{dw}} the fraction of the energy density stored in the walls at annihilation, ΔNeff\Delta N_{\mathrm{eff}} the relativistic degrees of freedom that the decay products contribute, g∗g_* and h∗h_* the energy and entropy degrees of freedom at T∗T_*, heqh_{\mathrm{eq}} the latter at matter-radiation equality, T0T_0 the present photon temperature, MplM_{\mathrm{pl}} the reduced Planck mass and H0H_0 the Hubble rate today. The spectrum rises as f3f^3 below f0f_0 and falls as f−bf^{-b} above it, with cc controlling how sharply it turns over. The four free model parameters are −4<log⁡10(T∗/GeV)<4-4<\log_{10}(T_*/\mathrm{GeV})<4, −3<log⁡10ΔNeff<log⁡100.39-3<\log_{10}\Delta N_{\mathrm{eff}}<\log_{10}0.39, 0.5<b<10.5<b<1 and 0.3<c<30.3<c<3, in the parametrisation of ref. [125]. We keep T∗T_* as a scanned parameter and profile over the three others. The prior on ΔNeff\Delta N_{\mathrm{eff}} stops at 0.39, the bound from nucleosynthesis alone adopted in ref. [125]. It is weaker than the joint bound 0.22 from nucleosynthesis and the cosmic microwave background (CMB) [102] that we apply to the phase transition, because the CMB bound holds only if the decay products of the walls remain relativistic until recombination.

Cosmic superstrings

Loops chopped off a cosmic string network oscillate and radiate gravitational waves. The resulting stochastic background can be computed via the sum over loop harmonics integrated over cosmic history. We take the model file super of ref. [77], which computes the signal from field-theory cosmic superstrings and then normalises it to the intercommutation probability PP [126], such that it corresponds to the superstring prediction [21],

h2ΩGWsuper(f)=h2PΩGWcs(f),whereΩGWcs(f)=8π3H02(Gμ)2∑k=1kmax⁡PkIk(f),h^2\Omega_{\mathrm{GW}}^{\mathrm{super}}(f)=\frac{h^2}{P}\Omega_{\mathrm{GW}}^{\mathrm{cs}}(f),\qquad\text{where}\quad\Omega_{\mathrm{GW}}^{\mathrm{cs}}(f)=\frac{8\pi}{3H_0^2}(G\mu)^2\sum_{k=1}^{k_{\max}}P_k\mathcal{I}_k(f),
Pk=Γζ(q)1kq,Ik(f)=2kf∫tinit0dt(a(t)a(t0))5nℓ(2kfa(t)a(t0),t).(21)P_k=\frac{\Gamma}{\zeta(q)}\frac{1}{k^q},\qquad\mathcal{I}_k(f)=\frac{2k}{f}\int_{t_{\mathrm{ini}}}^{t_0}\mathrm{d}t\left(\frac{a(t)}{a(t_0)}\right)^5 n_\ell\left(\frac{2k}{f}\frac{a(t)}{a(t_0)},t\right). \tag*{(21)}

The string tension GμG\mu is given in units of Newton’s constant, PkP_k is the power radiated into the kk-th harmonic of a loop, normalised such that ∑kPk=Γ=50\sum_kP_k=\Gamma=50 as found in simulations [127], and q=4/3q=4/3 for emission dominated by cusps. The loop number density nℓ(ℓ,t)n_\ell(\ell,t) follows from the velocity-dependent one-scale model with the loop-chopping parameter and loop size fixed to their simulated values [128], and the sum over harmonics is truncated at kmax⁡=105k_{\max}=10^5. The two parameters are −14<log⁡10Gμ<−6-14<\log_{10}G\mu<-6 and −3<log⁡10P<0-3<\log_{10}P<0, and we profile over PP.

Inflationary gravitational waves

Inflationary tensor fluctuations are generally expected to follow, to leading approximation, a power-law shape PT\mathcal{P}_T in frequency. Converted to our units and redshifted through the cosmic expansion history, the spectrum (igw of [77]) reads

h2ΩGWigw(f)=h2Ωrad24g(f)g0(h0h(f))4/3PT(f)T(f),PT(f)=rAs(ffcmb)nt,h^2\Omega_{\rm GW}^{\rm igw}(f)=h^2\frac{\Omega_{\rm rad}}{24}\frac{g(f)}{g_0}\left(\frac{h_0}{h(f)}\right)^{4/3}\mathcal{P}_{T}(f)T(f),\qquad\mathcal{P}_{T}(f)=rA_{\rm s}\left(\frac{f}{f_{\rm cmb}}\right)^{n_{\rm t}},
T(f)=[1−0.22(ffrh)3/2+0.65(ffrh)2]−1,(22)T(f)=\left[1-0.22\left(\frac{f}{f_{\rm rh}}\right)^{3/2}+0.65\left(\frac{f}{f_{\rm rh}}\right)^2\right]^{-1}, \tag*{(22)}

with transfer function T(f)T(f) as fitted in refs. [129, 130]. Here Ωrad\Omega_{\rm rad} is the present radiation density, AsA_{\rm s} the measured scalar amplitude at the CMB pivot scale fcmbf_{\rm cmb}, and frhf_{\rm rh} the frequency of a mode that was of horizon size at reheating, fixed by the reheating temperature TrhT_{\rm rh}. The degrees of freedom functions gg and hh are evaluated at the temperature at which the corresponding mode re-entered the Hubble horizon. Modes that re-entered the horizon before reheating are suppressed by the transfer function, which falls as f−2f^{-2} above frhf_{\rm rh}. The three free parameters are the tensor-to-scalar ratio −40<log⁡10r<0-40 < \log_{10} r < 0, the (tensorial) spectral tilt 0<nt<60 < n_{\rm t} < 6 and the reheating temperature −3<log⁡10(Trh/GeV)<3-3 < \log_{10}(T_{\rm rh}/{\rm GeV}) < 3. A blue tilt, nt>0n_{\rm t} > 0, is needed for the signal to reach the nanohertz band at all, and the three parameters are strongly degenerate, which is why we scan ntn_{\rm t} and profile over rr and TrhT_{\rm rh} (§ 5.2).

Priors of the population model

Table 1 in § 2.1 lists the six parameters of the population model with their uniform prior ranges and the three Gaussian prior sets. This appendix traces the three Gaussian prior sets – excluding the completely uninformed, flat one – to the measurements they are based on.

NANOGrav priors

The priors on the galaxy stellar mass function used in the NANOGrav analysis [8] are stated as being based on fits to the data of Tomczak et al. [62] (table B1 of ref. [8], footnotes a and d), without explicit reference as to how their widths were chosen. A comment in the parameter space of the holodeck library indicates that they derive from fits to Tomczak et al. “with 4x stdev values”.8 That notebook fits straight lines in redshift to the parameters of the single Schechter fits, eq. (4), of the total galaxy sample in table 3 of ref. [62] (page 12) and finds ψ0=−2.57±0.11\psi_0=-2.57\pm0.11 and mψ,0=10.94±0.11m_{\psi,0}=10.94\pm0.11, with slopes ψz=−0.59\psi_z=-0.59 and mψ,z=+0.12m_{\psi,z}=+0.12 and the low-mass slope αψ,0=−1.21\alpha_{\psi,0}=-1.21, close to the fixed values of § 2.1. We reproduced these fits. Two of the rows the notebook takes as the total sample, at 1.5<z<2.01.5<z<2.0 and 2.0<z<2.52.0<z<2.5, erroneously belong to the star-forming sample of the same table; with the rows of the total sample the fit gives ψ0=−2.61±0.10\psi_0=-2.61\pm0.10 and mψ,0=10.96±0.11m_{\psi,0}=10.96\pm0.11. The NANOGrav priors N(−2.56,0.40)\mathcal{N}(-2.56,0.40) and N(10.90,0.40)\mathcal{N}(10.90,0.40) hence appear to have about four times the width of the actual fits.

The NANOGrav priors on the normalisation and scatter of the black hole to bulge mass relation, μ=8.60±0.20\mu=8.60\pm0.20 and ϵμ=0.32±0.15\epsilon_\mu=0.32\pm0.15, lie between two measurements. Kormendy and Ho [59] find M∙/109M⊙=0.49−0.05+0.06(Mbulge/1011M⊙)1.16±0.08M_{\bullet}/10^9M_{\odot}=0.49^{+0.06}_{-0.05}(M_{\rm bulge}/10^{11}M_{\odot})^{1.16\pm0.08} with an intrinsic scatter of 0.29 dex (their eq. 10, page 571 of the journal version), i.e. μ=8.69±0.05\mu=8.69\pm0.05. They quote no uncertainty on that scatter. For the relations with the infrared bulge luminosity and with the velocity dispersion they find ϵμ=0.31±0.02\epsilon_\mu= 0.31 \pm0.02 and 0.29±0.030.29 \pm0.03 dex (page 566). McConnell and Ma [63] find μ=8.46±0.08\mu= 8.46 \pm0.08 with slope 1.05 and scatter ϵμ=0.34\epsilon_\mu= 0.34 dex (their table 2). The MM–σ\sigma and MM–LL relations of Gültekin et al. [131], cited by NANOGrav as well, which relate the black hole mass to the bulge velocity dispersion and luminosity, respectively, and do not involve the bulge mass, give scatters of ϵμ=0.44±0.06\epsilon_\mu= 0.44 \pm0.06 and ϵμ=0.38±0.09\epsilon_\mu= 0.38 \pm0.09 dex. The width 0.20 on μ\mu thus appears to span the difference between the two bulge-mass fits, and the width 0.15 on ϵμ\epsilon_\mu the spread of the quoted scatters (although this is not the reasoning provided in ref. [8]). holodeck also contains a narrower prior set based on Kormendy and Ho, with μ=8.69±0.05\mu= 8.69 \pm0.05, slope 1.17±0.081.17 \pm0.08 and ϵμ=0.28±0.05\epsilon_\mu= 0.28 \pm0.05, where a comment notes that no uncertainty on the scatter is given, and 0.28 dex is the scatter that Kormendy and Ho find for the relation with the velocity dispersion (their eq. 5). Also the widths of the NANOGrav priors on μ\mu and ϵμ\epsilon_\mu are hence about four times those of this narrower set. While we could not identify a clear justification of this discrepancy, this difference has a strong effect on the predicted amplitude of the binary background and hence on the preference for new physics (as demonstrated in the main text).

Tomczak–Kormendy priors

The Tomczak–Kormendy priors take the fits described in appendix § 9.§ 1 themselves. For the galaxy stellar mass function they take the fit to the total sample of Tomczak et al. [62], ψ0∼N(−2.61,0.10)\psi_0 \sim\mathcal{N}(-2.61, 0.10) and mψ,0∼N(10.96,0.11)m_{\psi,0} \sim\mathcal{N}(10.96, 0.11), for the black hole to bulge mass relation the narrower holodeck set, μ∼N(8.69,0.05)\mu\sim\mathcal{N}(8.69, 0.05) and ϵμ∼N(0.28,0.05)\epsilon_\mu\sim\mathcal{N}(0.28, 0.05), and for τf\tau_f and νinner\nu_{\mathrm{inner}} the same uniform priors as before (table 1). Like the NANOGrav priors, they combine the stellar masses of Tomczak et al. with the relation of Kormendy and Ho, whose bulge masses are calibrated to masses inferred from the orbital velocities of stars and are therefore larger by a factor of 1.34 for the same galaxy (see section 6.6.1 in ref. [59]). Inserting the stellar masses of Tomczak et al. into this relation underestimates the actual black hole masses by about 0.15 dex, just like a smaller μ\mu would, and reduces the predicted h2ΩGWh^2\Omega_{\mathrm{GW}} by about a factor of 1.7.

Survey-average priors

What we refer to as survey-average priors throughout our analysis is based on measurements of the galaxy stellar mass function and of the black hole to bulge mass relation published since 2014. Since there is no individual reference which fits these recent measurements to the parametrisation of § 2.1, we perform here a simple average of the centres and spreads of such measurements, which is sufficient for our purpose of illustrating the effect of the priors on the preference for new physics. The following discussion is not meant to be exhaustive, but to give a representative overview of the recent literature.

Galaxy surveys publish the mass function in redshift bins as fits of a single Schechter function in the form of eq. (4), or of a double Schechter function, the sum of two Schechter functions with the same characteristic mass and different low-mass slopes, which also follows the rise in the number of galaxies below about 1010M⊙10^{10} M_\odot. [62, 132] Above the characteristic mass the mass function falls off exponentially. In surveys, there is a statistical bias known as the Eddington bias [133]: more galaxies scatter up from the many lower masses than down from the few higher mass ones, and the measured mass function lies above the true one at the massive end. A survey corrects for it by convolving its model of the mass function with the error on the stellar mass of a single galaxy, 0.1 dex in Wright et al. [134] and 0.15 dex in McLeod et al. [135], before comparing it with the counts, so that the published parameters aim to describe the true mass function.

We use the six available determinations that are corrected in this way, all of the total galaxy population, star-forming and quiescent [86, 134, 135, 136, 137, 138]. Davidzon et al. (COSMOS2015, table 1) and Weaver et al. (COSMOS2020, table C.1) use deep near-infrared imaging of the two square degrees of the COSMOS field and reach z≃6z \simeq6 and 7.5. McLeod et al. (table 5) combine ground-based near-infrared surveys with imaging of the Hubble Space Telescope. Leja et al. (appendix B) use galaxies of the COSMOS and 3D-HST fields and give a mass function continuous in redshift for 0.2<z<30.2 < z < 3. Their masses come from star formation histories without a fixed functional form, and their number densities are higher than those of the other surveys (their section 6.1). Wright et al. (table A3) combine the GAMA, G10-COSMOS and 3D-HST surveys and also give a mass function continuous in redshift. Driver et al. (GAMA DR4, table 7) give the local mass function at z<0.1z < 0.1 from spectroscopic redshifts over 250 square degrees. The first three share galaxies of the COSMOS field and are not fully independent. The mass functions of Tomczak et al., on which the NANOGrav priors rest, are not corrected for the Eddington bias and are not among the six.

The binary background weights galaxies by their mass. For binaries that shrink by gravitational wave emission, the characteristic strain squared, hc2h_{\mathrm{c}}^2, at a given frequency is proportional to the number density of binaries times their chirp mass to the power 5/35/3 [15], which Liepold and Ma use to translate the local mass function into an amplitude of the background [13]. In the population model that we use the galaxy pair fraction and the galaxy merger time do not depend on the stellar mass (table B1 of ref. [8]), the mass of a black hole grows as M∙∝Mbulge1.10∝m⋆1.10M_{\bullet} \propto M_{\mathrm{bulge}}^{1.10} \propto m_{\star}^{1.10}, and at a fixed mass ratio the chirp mass is proportional to M∙M_{\bullet}. The galaxies of one decade of stellar mass therefore contribute to hc2h_{\mathrm{c}}^2 in proportion to Ψ m⋆5/3×1.10=Ψ m⋆1.83\Psi\,m_{\star}^{5/3 \times1.10} = \Psi\,m_{\star}^{1.83}. At the prior centres this product peaks at m⋆≃1011.1 M⊙m_{\star} \simeq10^{11.1}\,M_{\odot}, and about 90% of its integral comes from 1010.510^{10.5} to 1012 M⊙10^{12}\,M_{\odot} (Figure 9, left).

We fit the single Schechter function of eq. (4) to each survey. Its low-mass slope and redshift evolution are fixed at the values of section 2.1, so only ψ0\psi_0 and mψ,0m_{\psi,0} are free. With these two parameters it cannot follow the double Schechter curves of the surveys at all masses, and the fit has to choose where to match them. We minimise the squared difference in log⁡10Ψ\log_{10}\Psi between the two on a grid of masses over 10.5<log⁡10(m⋆/M⊙)<1210.5 < \log_{10}(m_{\star}/M_{\odot}) < 12, in all redshift bins of a survey with z≤3z \leq3 at once, and weight each mass bin by its contribution to hc2h_{\mathrm{c}}^2, Ψ m⋆1.83\Psi\,m_{\star}^{1.83}. The fit then reproduces the published curves to about 0.1 dex where this weight peaks. The mean and standard deviation of the six fits give ψ0=−2.20±0.15\psi_0 = -2.20 \pm0.15 and mψ,0=10.82±0.11m_{\psi,0} = 10.82 \pm0.11. We use this result to motivate our choice of survey-average priors in Table 1, as visualised in the right panel of Figure 9. This spread exceeds the statistical error of each survey, 0.02 to 0.1 dex on the normalisations (e.g. COSMOS2020 table C.1), and comes mainly from the modelling of the stellar populations (ref. [86], section 6.1).

The background does not depend on ψ0\psi_0 and mψ,0m_{\psi,0} separately but on how many massive galaxies they predict. We measure this by the hc2h_{\mathrm{c}}^2-weighted number density of galaxies,

nGW=∫10.512Ψ(m⋆,z=0.5)(m⋆1011 M⊙)1.83 dlog⁡10(m⋆/M⊙),(23)n_{\mathrm{GW}} = \int_{10.5}^{12} \Psi(m_{\star}, z=0.5) \left(\frac{m_{\star}}{10^{11}\,M_{\odot}}\right)^{1.83} \,\mathrm{d}\log_{10}(m_{\star}/M_{\odot}), \tag*{(23)}

evaluated at z=0.5z = 0.5, between the redshifts at which the contributions to hc2h_{\mathrm{c}}^2 peak in the models of Liepold and Ma, z≃0.3z \simeq0.3, and of NANOGrav, z≃1z \simeq1 [8, 13]. In the right panel of Figure 9, we show how the surveys spread along the lines of constant nGWn_{\mathrm{GW}}, since a higher normalisation of the mass function is compensated by a lower characteristic mass. They agree on log⁡10nGW\log_{10} n_{\rm GW} to 0.15 dex but fix ψ0\psi_0 and mψ,0m_{\psi,0} only to 0.15 and 0.11, with a correlation coefficient of −0.77-0.77 that the Gaussian priors neglect. The survey-average centre lies 0.20 dex above the NANOGrav centre in log⁡10nGW\log_{10} n_{\rm GW}. The NANOGrav centre is a fit of straight lines in redshift to the Schechter parameters of their table 3 (appendix § 9.§ 1). Our weighted fit of the same tables gives log⁡10nGW\log_{10} n_{\rm GW} between 0.12 and 0.14 dex above the NANOGrav centre, so most of the 0.20 dex come from the fit, and the newer surveys add 0.06 to 0.08 dex.

For the relation between black hole and bulge mass we centre the prior between two fits whose bulge masses are on the stellar-population scale of the galaxy surveys rather than on a dynamical one. Ref. [139] finds μ=8.77\mu= 8.77 (table 5), and eq. 10 of ref. [59] corresponds to μ=8.84\mu= 8.84 without the 0.126 dex by which its bulge masses were raised to a dynamical scale (before eq. 8). We take their mean, μ=8.80\mu= 8.80. For the width we take 0.10, corresponding to the spread of four fits: adding to those of refs. [59, 139], we take the fits of refs. [63, 140], whose bulge masses we lower by the same 0.126 dex. The intrinsic scatters ϵμ=0.40\epsilon_\mu= 0.40 dex [139] and ϵμ=0.42±0.04\epsilon_\mu= 0.42 \pm0.04 dex [140] motivate a prior of ϵμ=0.40±0.05\epsilon_\mu= 0.40 \pm0.05. In short, the survey-average priors lie in between the other two sets in how tightly they constrain μ\mu: the width of 0.10 on μ\mu lies between the 0.20 of the NANOGrav priors and the 0.05 of the Tomczak–Kormendy priors. On the stellar-population mass scale of the galaxy surveys, however, black holes are heavier at a given bulge mass, so that the centre of the prior on μ\mu lies 0.20 dex above that of the NANOGrav priors and 0.11 dex above that of the Tomczak–Kormendy priors. Together with the larger number of massive galaxies in the newer mass functions, our updated survey-average priors predict the loudest SMBHB background of all prior sets considered here.

The binaryfog code

The machinery used to derive the results of this paper is released as binaryfog [118] together with this paper and can be found at

https://github.com/tasicarl/binaryfog

The libraries of this first binaryfog release are built with holodeck at the current version 1.7.1 [47], including the corrections described in ref. [48]. The holodeck libraries, which took 130 CPU hours to generate, and the scripts that produce every figure and number in this paper accompany the publication as an archived release with a digital object identifier https://doi.org/10.5281/zenodo.24294394. binaryfog downloads the measured free spectra [83, 85, 106] from their public records on first use.

In the code, a new-physics model is supplied as a simple Python function which takes a frequency array as input and returns h2ΩGWh^2\Omega_{\mathrm{GW}} evaluated at these frequencies. Nothing about the package is specific to the models studied here, any of the publicly available model files of the NANOGrav new-physics search [77] can be handed to it unchanged, and any freespectrum product in the same format can replace the measured spectra. A typical scan (see examples/peaked_source.py in the repository) looks like

import numpy as np
import binaryfog as bf

def peaked(f_Hz, log10_amp, log10_fpeak):
    x = np.asarray(f_Hz) / 10.0 ** log10_fpeak
    return 10.0 ** log10_amp * x ** 3 / (1.0 + x ** 7)


data = bf.FreeSpectrum.load("ng15_cp_v2")
fpeak = np.linspace(-8.5, -7.0, 16)
like = bf.Likelihood(data, bf.Background())
scan = bf.ProfileScan(like, peaked, scan={"log10_fpeak": fpeak},
                      nuisance={"log10_amp": (-12, -5, "linear")},
                      astro_prior="nanograv", restarts=3)


table = scan.run(jobs=8)
inside = bf.region(table, 2, ndim=1)

and returns the profiled likelihood ln⁡Lprof(fpeak)\ln L_{\mathrm{prof}}(f_{\mathrm{peak}}) together with the astrophysical parameters that achieved it, in this case for a simple broken power law h2ΩGW(f)=A(f/fpeak)3/(1+(f/fpeak)7)h^2\Omega_{\mathrm{GW}}(f)=A(f/f_{\mathrm{peak}})^3/(1+(f/f_{\mathrm{peak}})^7) in the pandas table. From this, the 2σ2\sigma region is obtained in the last line of the above code. The function bf.preference(table, like) computes the fit without new physics and returns the gain GG of eq. (14). Further code documentation can be found online, at the repository address stated above. We encourage users to get in touch with us9 regarding any questions or issues related to the code.

References

  1. [1]R. W. Hellings and G. S. Downs, Upper limits on the isotropic gravitational radiation background from pulsar timing analysis, Astrophys. J. Lett. 265 (1983) L39–L42.
  2. [2]NANOGrav Collaboration, G. Agazie et al., The NANOGrav 15 yr Data Set: Evidence for a Gravitational-wave Background, Astrophys. J. Lett. 951 (2023), no. 1 L8, [2306.16213].
  3. [3]EPTA, InPTA Collaboration, J. Antoniadis et al., The second data release from the European Pulsar Timing Array - III. Search for gravitational wave signals, Astron. Astrophys. 678 (2023) A50, [2306.16214].
  4. [4]D. J. Reardon et al., Search for an Isotropic Gravitational-wave Background with the Parkes Pulsar Timing Array, Astrophys. J. Lett. 951 (2023), no. 1 L6, [2306.16215].DOI
  5. [5]H. Xu et al., Searching for the Nano-Hertz Stochastic Gravitational Wave Background with the Chinese Pulsar Timing Array Data Release I, Res. Astron. Astrophys. 23 (2023), no. 7 075024, [2306.16216].DOI
  6. [6]M. T. Miles et al., The MeerKAT Pulsar Timing Array: the first search for gravitational waves with the MeerKAT radio telescope, Mon. Not. Roy. Astron. Soc. 536 (2024), no. 2 1489–1500, [2412.01153].DOI
  7. [7]A. Sesana and D. G. Figueroa, Nanohertz Gravitational Waves, Ann. Rev. Astron. Astrophys. 64 (2026), no. 1 537–586, [2512.18822].arxiv.org/abs/2512.18822
  8. [8]NANOGrav Collaboration, G. Agazie et al., The NANOGrav 15 yr Data Set: Constraints on Supermassive Black Hole Binaries from the Gravitational-wave Background, Astrophys. J. Lett. 952 (2023), no. 2 L37, [2306.16220]. [Erratum: Astrophys.J.Lett. 1006, L67 (2026)].
  9. [9]G. Sato-Polito, M. Zaldarriaga, and E. Quataert, Where are the supermassive black holes measured by PTAs?, Phys. Rev. D 110 (2024), no. 6 063020, [2312.06756].
  10. [10]J. Ellis, M. Fairbairn, G. Hütsi, J. Raidal, J. Urrutia, et al., Gravitational waves from supermassive black hole binaries in light of the NANOGrav 15-year data, Phys. Rev. D 109 (2024), no. 2 L021302, [2306.17021].DOI
  11. [11]N. Chen, T. Di Matteo, Y. Zhou, L. Z. Kelley, L. Blecha, et al., The Gravitational-wave Background from Massive Black Holes in the ASTRID Simulation, Astrophys. J. Lett. 991 (2025), no. 1 L19, [2502.01024].arxiv.org/abs/2502.01024
  12. [12]C. M. F. Mingarelli, Energetic Ceilings and Benchmarks of Astrophysical Gravitational-Wave Backgrounds, 2601.18859.
  13. [13]E. R. Liepold and C.-P. Ma, Big Galaxies and Big Black Holes: The Massive Ends of the Local Stellar and Black Hole Mass Functions and the Implications for Nanohertz Gravitational Waves, Astrophys. J. Lett. 971 (2024), no. 2 L29, [2407.14595]. [Erratum: Astrophys.J.Lett. 1004, L20 (2026)].arxiv.org/abs/2407.14595
  14. [14]C. Matt et al., Inferring M_BH–M_bulge Evolution from the Gravitational-wave Background, Astrophys. J. 997 (2026), no. 2 188, [2508.18126].DOI
  15. [15]E. S. Phinney, A Practical theorem on gravitational wave backgrounds, astro-ph/0108028.DOI
  16. [16]B. Dai, M. Zaldarriaga, and Y. Ali-Haïmoud, Prior volume effect inflates the gravitational-wave background in pulsar timing arrays, 2610.28739.arxiv.org/abs/2610.28739
  17. [17]M. C. Begelman, R. D. Blandford, and M. J. Rees, Massive black hole binaries in active galactic nuclei, Nature 287 (1980) 307–309.DOI
  18. [18]M. Milosavljevic and D. Merritt, Long-term evolution of massive black hole binaries, Astrophys. J. 596 (2003) 860, [astro-ph/0212459].DOI
  19. [19]F. M. Khan, A. Just, and D. Merritt, Efficient Merger of Binary Supermassive Black Holes in Merging Galaxies, Astrophys. J. 732 (2011) 89, [1103.0272].DOI
  20. [20]E. Vasiliev, F. Antonini, and D. Merritt, The Final-parsec Problem in the Collisionless Limit, Astrophys. J. 810 (2015), no. 1 49, [1505.05480].DOI
  21. [21]NANOGrav Collaboration, A. Afzal et al., The NANOGrav 15 yr Data Set: Search for Signals from New Physics, Astrophys. J. Lett. 951 (2023), no. 1 L11, [2306.16219]. [Erratum: Astrophys.J.Lett. 971, L27 (2024)].
  22. [22]EPTA, InPTA Collaboration, J. Antoniadis et al., The second data release from the European Pulsar Timing Array - IV. Implications for massive black holes, dark matter, and the early Universe, Astron. Astrophys. 685 (2024) A94, [2306.16227].
  23. [23]E. Madge, E. Morgante, C. Puchades-Ibáñez, N. Ramberg, W. Ratzinger, et al., Primordial gravitational waves in the nano-Hertz regime and PTA data – towards solving the GW inverse problem, JHEP 10 (2023) 171, [2306.14856].DOI
  24. [24]D. G. Figueroa, M. Pieroni, A. Ricciardone, and P. Simakachorn, Cosmological Background Interpretation of Pulsar Timing Array Data, Phys. Rev. Lett. 132 (2024), no. 17 171002, [2307.02399].DOI
  25. [25]J. Ellis, M. Fairbairn, G. Franciolini, G. Hütsi, A. Iovino, et al., What is the source of the PTA GW signal?, Phys. Rev. D 109 (2024), no. 2 023522, [2308.08546].arxiv.org/abs/2308.08546
  26. [26]M. Wang, Z. Zhang, and H. Xu, Discriminating Between Models of the Nanohertz Gravitational-Wave Background with Pulsar Timing Arrays, 2510.22713.arxiv.org/abs/2510.22713
  27. [27]T. Bringmann, P. F. Depta, T. Konstandin, K. Schmidt-Hoberg, and C. Tasillo, Does NANOGrav observe a dark sector phase transition?, JCAP 11 (2023) 053, [2306.09411].arxiv.org/abs/2306.09411
  28. [28]S. Balan, T. Bringmann, F. Kahlhoefer, J. Matuszak, and C. Tasillo, Sub-GeV dark matter and nano-Hertz gravitational waves from a classically conformal dark sector, JCAP 08 (2025) 062, [2502.19478].DOI
  29. [29]T. Bringmann, T. Konstandin, J. Matuszak, K. Schmidt-Hoberg, and C. Tasillo, Tuning the violins: dark sector phase transition models for the PTA signal, 2602.09092.arxiv.org/abs/2602.09092
  30. [30]J. Ellis, M. Lewicki, C. Lin, and V. Vaskonen, Cosmic superstrings revisited in light of NANOGrav 15-year data, Phys. Rev. D 108 (2023), no. 10 103511, [2306.17147].arxiv.org/abs/2306.17147
  31. [31]S. Vagnozzi, Inflationary interpretation of the stochastic gravitational wave background signal detected by pulsar timing array experiments, JHEAp 39 (2023) 81–98, [2306.16912].DOI
  32. [32]G. Franciolini, A. Iovino, Jr., V. Vaskonen, and H. Veermäe, Recent Gravitational Wave Observation by Pulsar Timing Arrays and Primordial Black Holes: The Importance of Non-Gaussianities, Phys. Rev. Lett. 131 (2023), no. 20 201401, [2306.17149].DOI
  33. [33]L. Bian, S. Ge, J. Shu, B. Wang, X.-Y. Yang, et al., Gravitational wave sources for pulsar timing arrays, Phys. Rev. D 109 (2024), no. 10 L101301, [2307.02376].arxiv.org/abs/2307.02376
  34. [34]Y.-M. Wu, Z.-C. Chen, and Q.-G. Huang, Cosmological interpretation of the stochastic signal in pulsar timing arrays, Sci. China Phys. Mech. Astron. 67 (2024), no. 4 240412, [2307.03141].DOI
  35. [35]Y. Gouttenoire and E. Vitagliano, Domain wall interpretation of the PTA signal confronting black hole overproduction, Phys. Rev. D 110 (2024), no. 6 L061306, [2306.17841].arxiv.org/abs/2306.17841
  36. [36]N. Bhaumik, R. K. Jain, and M. Lewicki, Ultralow mass primordial black holes in the early Universe can explain the pulsar timing array signal, Phys. Rev. D 108 (2023), no. 12 123532, [2308.07912].DOI
  37. [37]M. W. Winkler and K. Freese, Origin of the stochastic gravitational wave background: First-order phase transition versus black hole mergers, Phys. Rev. D 111 (2025), no. 8 083509, [2401.13729].arxiv.org/abs/2401.13729
  38. [38]Y. Gouttenoire, First-Order Phase Transition Interpretation of Pulsar Timing Array Signal Is Consistent with Solar-Mass Black Holes, Phys. Rev. Lett. 131 (2023), no. 17 171404, [2307.04239].DOI
  39. [39]Y. Ali-Haïmoud, A practical theorem on gravitational-wave background statistics, 2604.19701.
  40. [40]W. G. Lamb and S. R. Taylor, Spectral Variance in a Stochastic Gravitational-wave Background from a Binary Population, Astrophys. J. Lett. 971 (2024), no. 1 L10, [2407.06270].DOI
  41. [41]X. Xue, Z. Pan, and L. Dai, Non-Gaussian statistics of nanohertz stochastic gravitational waves, Phys. Rev. D 111 (2025), no. 4 043022, [2409.19516].arxiv.org/abs/2409.19516
  42. [42]NANOGrav Collaboration, G. Agazie et al., The NANOGrav 15 yr Data Set: Looking for Signs of Discreteness in the Gravitational-wave Background, Astrophys. J. 978 (2025), no. 1 31, [2404.07020].DOI
  43. [43]C. J. Moore and A. Vecchio, Ultra-low-frequency gravitational waves from cosmological and astrophysical processes, Nature Astron. 5 (2021), no. 12 1268–1274, [2104.15130].DOI
  44. [44]B. Bécsy, N. J. Cornish, and L. Z. Kelley, Exploring Realistic Nanohertz Gravitational-wave Backgrounds, Astrophys. J. 941 (2022), no. 2 119, [2207.01607].DOI
  45. [45]T. Konstandin, A.-M. Lemke, A. Mitridate, and E. Perboni, The impact of cosmic variance on PTAs anisotropy searches, JCAP 04 (2025) 059, [2408.07741].arxiv.org/abs/2408.07741
  46. [46]E. C. Gardiner et al., Constraints on Supermassive Black Hole Binaries from the Lack of Resolvable Sources in the NANOGrav 15-yr Dataset, 2609.32306.
  47. [47]The NANOGrav Collaboration, holodeck: massive black hole binary population synthesis, 2026. software, version 1.7.1; libraries built at commit 44fe308, https://github.com/nanograv/holodeck.
  48. [48]NANOGrav Collaboration, G. Agazie et al., Erratum: The NANOGrav 15 yr Data Set: Constraints on Supermassive Black Hole Binaries from the Gravitational-wave Background, Astrophys. J. Lett. 1006 (2026), no. 2 L67.
  49. [49]S. Chen, A. Sesana, and C. J. Conselice, Constraining astrophysical observables of Galaxy and Supermassive Black Hole Binary Mergers using Pulsar Timing Arrays, Mon. Not. Roy. Astron. Soc. 488 (2019), no. 1 401–418, [1810.04184].DOI
  50. [50]LISA Collaboration, P. Amaro-Seoane et al., Laser Interferometer Space Antenna, 1702.00786.
  51. [51]D. Merritt and M. Milosavljević, Massive black hole binary evolution, Living Rev. Rel. 8 (2005) 8, [astro-ph/0410364].DOI
  52. [52]P. J. Armitage and P. Natarajan, Accretion during the merger of supermassive black holes, Astrophys. J. Lett. 567 (2002) L9–L12, [astro-ph/0201318].DOI
  53. [53]T. Ryu, R. Perna, Z. Haiman, J. P. Ostriker, and N. C. Stone, Interactions between multiple supermassive black holes in galactic nuclei: a solution to the final parsec problem, Mon. Not. Roy. Astron. Soc. 473 (2018), no. 3 3410–3433, [1709.06501].DOI
  54. [54]M. Bonetti, A. Sesana, E. Barausse, and F. Haardt, Post-Newtonian evolution of massive black hole triplets in galactic nuclei – III. A robust lower limit to the nHz stochastic background of gravitational waves, Mon. Not. Roy. Astron. Soc. 477 (2018), no. 2 2599–2612, [1709.06095].
  55. [55]G. Alonso-Álvarez, J. M. Cline, and C. Dewar, Self-Interacting Dark Matter Solves the Final Parsec Problem of Supermassive Black Hole Mergers, Phys. Rev. Lett. 133 (2024), no. 2 021401, [2401.14450].arxiv.org/abs/2401.14450
  56. [56]A. Sesana and F. M. Khan, Scattering experiments meet N-body – I. A practical recipe for the evolution of massive black hole binaries in stellar environments, Mon. Not. Roy. Astron. Soc. 454 (2015), no. 1 L66–L70, [1505.02062].DOI
  57. [57]P. Lang et al., Bulge Growth and Quenching since z = 2.5 in CANDELS/3D-HST, Astrophys. J. 788 (2014) 11, [1402.0866].DOI
  58. [58]A. F. L. Bluck, J. T. Mendel, S. L. Ellison, J. Moreno, L. Simard, et al., Bulge mass is king: The dominant role of the bulge in determining the fraction of passive galaxies in the Sloan Digital Sky Survey, Mon. Not. Roy. Astron. Soc. 441 (2014), no. 1 599–629, [1403.5269].DOI
  59. [59]P. Kormendy and L. C. Ho, Coevolution (Or Not) of Supermassive Black Holes and Host Galaxies, Ann. Rev. Astron. Astrophys. 51 (2013) 511–653, [1304.7762].
  60. [60]L. Z. Kelley, L. Blecha, and L. Hernquist, Massive Black Hole Binary Mergers in Dynamical Galactic Environments, Mon. Not. Roy. Astron. Soc. 464 (2017), no. 3 3131–3157, [1606.01900].DOI
  61. [61]E. C. Gardiner, B. Bécsy, L. Z. Kelley, and N. J. Cornish, Characterizing Continuous Gravitational Waves from Supermassive Black Hole Binaries in Realistic Pulsar Timing Array Data, Astrophys. J. 988 (2025), no. 2 222, [2502.16016].arxiv.org/abs/2502.16016
  62. [62]A. R. Tomczak et al., Galaxy Stellar Mass Functions from ZFOURGE/CANDELS: An Excess of Low-mass Galaxies since z = 2 and the Rapid Buildup of Quiescent Galaxies, Astrophys. J. 783 (2014) 85, [1309.5972].DOI
  63. [63]N. J. McConnell and C.-P. Ma, Revisiting the Scaling Relations of Black Hole Masses and Host Galaxy Properties, Astrophys. J. 764 (2013) 184, [1211.2816].DOI
  64. [64]A. Sesana, A. Vecchio, and C. N. Colacino, The stochastic gravitational-wave background from massive black hole binary systems: implications for observations with Pulsar Timing Arrays, Mon. Not. Roy. Astron. Soc. 390 (2008) 192, [0804.4476].DOI
  65. [65]S. R. Taylor, J. Simon, and L. Sampson, Constraints On The Dynamical Environments Of Supermassive Black-hole Binaries Using Pulsar-timing Arrays, Phys. Rev. Lett. 118 (2017), no. 18 181102, [1612.02817].DOI
  66. [66]S. Tiruvaskar and C. Gordon, Self-interacting dark-matter spikes and the final-parsec problem: Bayesian constraints from the NANOGrav 15-year gravitational-wave background, Phys. Rev. D 113 (2026), no. 4 043501, [2506.18153].DOI
  67. [67]S. Tiruvaskar, R. Boey, R. Easther, and C. Gordon, Ultralight dark matter constraints from nano-Hertz gravitational waves, Phys. Rev. D 113 (2026), no. 6 063541, [2512.15292].DOI
  68. [68]M. Bonetti, A. Franchini, B. G. Galuzzi, and A. Sesana, Neural networks unveiling the properties of gravitational wave background from supermassive black hole binaries, Astron. Astrophys. 687 (2024) A42, [2311.04276].DOI
  69. [69]S. Tiruvaskar and C. Gordon, Replacing Gaussian Processes with Neural Networks in Pulsar Timing Array Inference of the Gravitational-Wave Background, 2604.04340.
  70. [70]N. Laal et al., Deep Neural Emulation of the Supermassive Black Hole Binary Population, Astrophys. J. 982 (2025), no. 1 55, [2411.10519].arxiv.org/abs/2411.10519
  71. [71]K. Lemoto, B. Goncharov, G. Sato-Polito, and X. Bi, Bridging the Population Synthesis of Supermassive Binary Black Holes and the Gravitational Wave Background, 2609.07703.arxiv.org/abs/2609.07703
  72. [72]L. Sampson, N. J. Cornish, and S. T. McWilliams, Constraining the Solution to the Last Parsec Problem with Pulsar Timing, Phys. Rev. D 91 (2015), no. 8 084055, [1503.02662].DOI
  73. [73]NANOGrav Collaboration, Z. Arzoumanian et al., The NANOGrav Nine-year Data Set: Limits on the Isotropic Stochastic Gravitational Wave Background, Astrophys. J. 821 (2016), no. 1 13, [1508.03024].
  74. [74]W. G. Lamb, S. R. Taylor, and R. van Haasteren, Rapid refitting techniques for Bayesian spectral characterization of the gravitational wave background using pulsar timing arrays, Phys. Rev. D 108 (2023), no. 10 103019, [2303.15442].DOI
  75. [75]A. Mitridate, D. Wright, R. von Eckardstein, T. Schröder, J. Nay, et al., PTArcade, 2306.16377.arxiv.org/abs/2306.16377
  76. [76]J. A. Ellis, M. Vallisneri, S. R. Taylor, and P. T. Baker, ENTERPRISE: Enhanced Numerical Toolbox Enabling a Robust PulaR Inference SuitE, 2020. software, https://github.com/nanograv/enterprise.DOI
  77. [77]The NANOGrav Collaboration, The NANOGrav 15-year search for new physics signals – model files, 2023.
  78. [78]S. S. Wilks, The Large-Sample Distribution of the Likelihood Ratio for Testing Composite Hypotheses, Annals Math. Statist. 9 (1938), no. 1 60–62.DOI
  79. [79]Fermi-LAT Collaboration, A. A. Abdo et al., Fermi Large Area Telescope First Source Catalog, Astrophys. J. Suppl. 188 (2010) 405–436, [1002.2280].
  80. [80]E. Gross and O. Vitells, Trial factors for the look elsewhere effect in high energy physics, Eur. Phys. J. C 70 (2010) 525–530, [1005.1891].DOI
  81. [81]R. Protassov, D. A. van Dyk, A. Connors, V. L. Kashyap, and A. Siemiginowska, Statistics: handle with care, detecting multiple model components with the likelihood ratio test, Astrophys. J. 571 (2002) 545, [astro-ph/0201547].DOI
  82. [82]CTAO Collaboration, S. Abe et al., Dark matter line searches with the Cherenkov Telescope Array, JCAP 07 (2024) 047, [2403.04857].
  83. [83]The NANOGrav Collaboration, KDE representations of the gravitational wave background free spectra present in the NANOGrav 15-year dataset, version 2.0.0, 2026.
  84. [84]J. Ellis and R. van Haasteren, jellis18/PTMCMCSampler: Official Release, 2017.DOI
  85. [85]A. Mitridate and D. Wright, PTArcade – data, 2024.DOI
  86. [86]J. Leja, J. S. Speagle, B. D. Johnson, C. Conroy, P. van Dokkum, et al., A New Census of the 0.2 < z < 3.0 Universe. I. The Stellar Mass Function, Astrophys. J. 893 (2020) 111, [1910.04168].DOI
  87. [87]G. Sato-Polito and M. Zaldarriaga, Uncertainties in the supermassive black hole abundance and implications for the GW background, Phys. Rev. D 114 (2026), no. 6 063052, [2509.08041].arxiv.org/abs/2509.08041
  88. [88]A. Lapi, F. Shankar, M. Bosi, D. Roberts, H. Fu, et al., Semi-empirical framework of supermassive black hole evolution: highlighting a possible tension between demographics and gravitational wave background, JCAP 02 (2026) 001, [2507.15436].arxiv.org/abs/2507.15436
  89. [89]J. Ellis, M. Fairbairn, G. Hütsi, J. Urrutia, V. Vaskonen, et al., Consistency of JWST black hole observations with NANOGrav gravitational wave measurements, Astron. Astrophys. 691 (2024) A270, [2403.19650].DOI
  90. [90]E. Roebber, G. Holder, D. E. Holz, and M. Warren, Cosmic variance in the nanohertz gravitational wave background, Astrophys. J. 819 (2016) 163, [1508.07336].DOI
  91. [91]V. Ravi, J. S. B. Wyithe, G. Hobbs, R. M. Shannon, R. N. Manchester, et al., Does a ‘stochastic’ background of gravitational waves exist in the pulsar timing band?, Astrophys. J. 761 (2012) 84, [1210.3854].
  92. [92]J. Ellis, M. Fairbairn, G. Hütsi, M. Raidal, J. Urrutia, et al., Prospects for future binary black hole gravitational wave studies in light of PTA measurements, Astron. Astrophys. 676 (2023) A38, [2301.13854].DOI
  93. [93]S. C. Sardesai, J. Simon, and S. J. Vigeland, Characterizing the Nanohertz Gravitational Wave Background Using a t-process Power Spectral Density, Astrophys. J 976 (2024), no. 2 212, [2408.10139].arxiv.org/abs/2408.10139
  94. [94]B. E. Moreschi, K. A. Gersbach, G. Shaifullah, S. R. Taylor, and A. Sesana, Modeling the nanohertz gravitational wave background with the t-process: Improved frequentist analysis of gravitational-wave power, 2610.08906.arxiv.org/abs/2610.08906
  95. [95]G. Sato-Polito and M. Zaldarriaga, Distribution of the gravitational-wave background from supermassive black holes, Phys. Rev. D 111 (2025), no. 2 023043, [2406.17010].
  96. [96]J. Raidal, J. Urrutia, V. Vaskonen, and H. Veermäe, Heavy tailed non-Gaussianity of the supermassive black hole gravitational wave background, Phys. Rev. D 114 (2026), no. 4 043550, [2604.08506].arxiv.org/abs/2604.08506
  97. [97]H. Hisamatsu and K. Kyutoku, Higher-order statistics of the stochastic gravitational wave background from supermassive black hole binaries, Phys. Rev. D 114 (2026), no. 6 063041, [2605.17983].DOI
  98. [98]J. Yu, Z. Pan, X. Xue, and L. Dai, Population statistics of nanohertz gravitational wave sources, 2607.07477.arxiv.org/abs/2607.07477
  99. [99]M. Falxa, Approximating the statistics of a gravitational wave background, 2609.07686.arxiv.org/abs/2609.07686
  100. [100]M. Falxa and A. Sesana, Modeling non-Gaussianities in pulsar timing array data analysis using Gaussian mixture models, Phys. Rev. D 113 (2026), no. 4 043047, [2508.08365].arxiv.org/abs/2508.08365
  101. [101]D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman, emcee: The MCMC Hammer, Publ. Astron. Soc. Pac. 125 (2013) 306–312, [1202.3665].DOI
  102. [102]T.-H. Yeh, J. Shelton, K. A. Olive, and B. D. Fields, Probing physics beyond the standard model: limits from BBN and the CMB independently and combined, JCAP 10 (2022) 046, [2207.3133].DOI
  103. [103]NANOGrav Collaboration, G. Agazie et al., The NANOGrav 15 yr Data Set: Piecewise Power-law Reconstruction of the Gravitational-wave Background, Astrophys. J. Lett. 1004 (2026), no. 2 L40, [2601.09481].
  104. [104]LISA Cosmology Working Group Collaboration, C. Caprini, R. Jinno, M. Lewicki, E. Madge, M. Merchand, et al., Gravitational waves from first-order phase transitions in LISA: reconstruction pipeline and physics interpretation, JCAP 10 (2024) 020, [2403.03723].DOI
  105. [105]M. Lewicki and V. Vaskonen, Impact of cosmic expansion on gravitational wave spectra from strongly supercooled first-order phase transitions, Phys. Dark Univ. 53 (2026) 102425, [2511.15687].arxiv.org/abs/2511.15687
  106. [106]The NANOGrav Collaboration, KDE representations of the gravitational wave background free spectra present in the NANOGrav 15-year dataset, version 1.0.0, 2023.
  107. [107]L. Z. Kelley, L. Blecha, L. Hernquist, A. Sesana, and S. R. Taylor, The Gravitational Wave Background from Massive Black Hole Binaries in Illustris: spectral features and time to detection with pulsar timing arrays, Mon. Not. Roy. Astron. Soc. 471 (2017), no. 4 4508–4526, [1702.02180].DOI
  108. [108]J. Raidal, J. Urrutia, V. Vaskonen, and H. Veermäe, Eccentricity effects on the supermassive black hole gravitational wave background, Astron. Astrophys. 691 (2024) A212, [2406.05125].DOI
  109. [109]S. Chen, A. Sesana, and W. Del Pozzo, Efficient computation of the gravitational wave spectrum emitted by eccentric massive black hole binaries in stellar environments, Mon. Not. Roy. Astron. Soc. 470 (2017), no. 2 1738–1749, [1612.00455].DOI
  110. [110]B. Goncharov, G. Sato-Polito, X. Bi, and M. Zaldarriaga, A Joint Optimal Search for Gravitational Waves from Resolved and Unresolved Supermassive Binary Black Holes with Pulsar Timing Arrays, 2606.18241.arxiv.org/abs/2606.18241
  111. [111]B. Bécsy et al., How to Detect an Astrophysical Nanohertz Gravitational Wave Background, Astrophys. J. 959 (2023), no. 1 9, [2309.04443].DOI
  112. [112]E. C. Gardiner, L. Z. Kelley, A.-M. Lemke, and A. Mitridate, Beyond the Background: Gravitational-wave Anisotropy and Continuous Waves from Supermassive Black Hole Binaries, Astrophys. J. 965 (2024), no. 2 164, [2309.07227].arxiv.org/abs/2309.07227
  113. [113]C. Cecchini, J. El Gammal, G. Franciolini, and M. Pieroni, Are PTA measurements sensitive to gravitational wave non-Gaussianities?, 2605.05157.arxiv.org/abs/2605.05157
  114. [114]NANOGrav Collaboration, G. Agazie et al., The NANOGrav 15 yr Data Set: Bayesian Limits on Gravitational Waves from Individual Supermassive Black Hole Binaries, Astrophys. J. Lett. 951 (2023), no. 2 L50, [2306.16222].
  115. [115]W. G. Lamb, J. M. Wachter, A. Mitridate, S. C. Sardesai, B. Bécsy, et al., Finite populations and finite time: The non-Gaussianity of a gravitational wave background, Phys. Rev. D 113 (2026), no. 12 123065, [2511.09659].DOI
  116. [116]V. Rodriguez-Gomez et al., The merger rate of galaxies in the Illustris Simulation: a comparison with observations and semi-empirical models, Mon. Not. Roy. Astron. Soc. 449 (2015), no. 1 49–64, [1502.01339].DOI
  117. [117]D. C. Wright, A. D. Johnson, J. S. Hazboun, and W. G. Lamb, Fast and Flow-rious: Gravitational Wave Background Bayesian Model Comparison using Normalizing Flows and Nested Sampling, 2609.38122.arxiv.org/abs/2609.38122
  118. [118]C. Tasillo, T. Bringmann, and C. Gordon, binaryfog: the supermassive black hole binary background as a nuisance in pulsar timing array searches, 2026. software, version 0.1.0, https://github.com/tasicarl/binaryfog.
  119. [119]J. Matuszak and C. Tasillo, TransitionListener v2.0 – Robust gravitational wave predictions for cosmological phase transitions, 2605.15259.arxiv.org/abs/2605.15259
  120. [120]Y. Gouttenoire, R. Jinno, and F. Sala, Friction pressure on relativistic bubble walls, JHEP 05 (2022) 004, [2112.07686].DOI
  121. [121]R. Jinno, T. Konstandin, H. Rubira, and I. Stomberg, Higgsless simulations of cosmological phase transitions and gravitational waves, JCAP 02 (2023) 011, [2209.04369].DOI
  122. [122]J. R. Espinosa, T. Konstandin, J. M. No, and G. Servant, Energy Budget of Cosmological First-order Phase Transitions, JCAP 06 (2010) 028, [1004.4187].DOI
  123. [123]P. F. de Salas, M. Lattanzi, G. Mangano, G. Miele, S. Pastor, et al., Bounds on very low reheating scenarios after Planck, Phys. Rev. D 92 (2015), no. 12 123534, [1511.00672].DOI
  124. [124]T. Hiramatsu, M. Kawasaki, and K. Saikawa, On the estimation of gravitational wave spectrum from cosmic domain walls, JCAP 02 (2014) 031, [1309.5001].DOI
  125. [125]R. Z. Ferreira, A. Notari, O. Pujolas, and F. Rompineve, Gravitational waves from domain walls in Pulsar Timing Array datasets, JCAP 02 (2023) 001, [2204.04228].DOI
  126. [126]M. G. Jackson, N. T. Jones, and J. Polchinski, Collisions of cosmic F and D-strings, JHEP 10 (2005) 013, [hep-th/0405229].DOI
  127. [127]J. J. Blanco-Pillado and K. D. Olum, Stochastic gravitational wave background from smoothed cosmic string loops, Phys. Rev. D 96 (2017), no. 10 104046, [1709.02693].DOI
  128. [128]J. J. Blanco-Pillado, K. D. Olum, and B. Shlaer, The number of cosmic string loops, Phys. Rev. D 89 (2014), no. 2 023512, [1309.6637].DOI
  129. [129]S. Kuroyanagi, T. Takahashi, and S. Yokoyama, Blue-tilted Tensor Spectrum and Thermal History of the Universe, JCAP 02 (2015) 003, [1407.4785].DOI
  130. [130]S. Kuroyanagi, T. Takahashi, and S. Yokoyama, Blue-tilted inflationary tensor spectrum and reheating in the light of NANOG­rav results, JCAP 01 (2021) 071, [2011.03323].DOI
  131. [131]K. Gültekin et al., The M-sigma and M-L Relations in Galactic Bulges and Determinations of their Intrinsic Scatter, Astrophys. J. 698 (2009) 198–221, [0903.4897].DOI
  132. [132]I. K. Baldry et al., Galaxy And Mass Assembly (GAMA): The galaxy stellar mass function at z < 0.06, Mon. Not. Roy. Astron. Soc. 421 (2012) 621–634, [1111.5707].DOI
  133. [133]A. S. Eddington, On a formula for correcting statistics for the effects of a known error of observation, Mon. Not. Roy. Astron. Soc. 73 (1913) 359–360.DOI
  134. [134]A. H. Wright, S. P. Driver, and A. S. G. Robotham, GAMA/G10-COSMOS/3D-HST: Evolution of the galaxy stellar mass function over 12.5 Gyr, Mon. Not. Roy. Astron. Soc. 480 (2018), no. 3 3491–3502, [1808.01754].DOI
  135. [135]D. J. McLeod, R. J. McLure, J. S. Dunlop, F. Cullen, A. C. Carnall, et al., The evolution of the galaxy stellar-mass function over the last 12 billion years from a combination of ground-based and HST surveys, Mon. Not. Roy. Astron. Soc. 503 (2021), no. 3 4413–4435, [2009.03176].DOI
  136. [136]I. Davidzon et al., The COSMOS2015 galaxy stellar mass function - Thirteen billion years of stellar mass assembly in ten snapshots, Astron. Astrophys. 605 (2017) A70, [1701.02734].DOI
  137. [137]J. R. Weaver, I. Davidzon, S. Toft, O. Ilbert, H. J. McCracken, et al., COSMOS2020: The galaxy stellar mass function. The assembly and star formation cessation of galaxies at 0.2 < z ≤ 7.5, Astron. Astrophys. 677 (2023) A184, [2212.02512].DOI
  138. [138]S. P. Driver et al., Galaxy And Mass Assembly (GAMA): Data Release 4 and the z < 0.1 total and z < 0.08 morphological galaxy stellar mass functions, Mon. Not. Roy. Astron. Soc. 513 (2022), no. 1 439–467, [2203.08539].DOI
  139. [139]N. Sahu, A. W. Graham, and B. L. Davis, Black Hole Mass Scaling Relations for Early-type Galaxies. I. M_BH–M_*,sph and M_BH–M_*,gal, Astrophys. J. 876 (2019), no. 2 155, [1903.04738].DOI
  140. [140]R. P. Saglia, M. Opitsch, P. Erwin, J. Thomas, A. Beifiori, et al., The SINFONI Black Hole Survey: The Black Hole Fundamental Plane Revisited and the Paths of (Co)evolution of Supermassive Black Holes and Bulges, Astrophys. J. 818 (2016), no. 1 47, [1601.00974].DOI

Paper details

vlnc:W_26mvgbz9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude Opus 5.5 (bare)

View statement
Looking for related papers…
Contents