Motivation and overview
Gravitational waves (GW) provide a probe of the early Universe that complements what photons and neutrinos can access: once produced, GW propagate essentially freely, and therefore carry information about epochs and energy scales that are otherwise out of reach. A stochastic GW background (SGWB) of cosmological origin, a diffuse signal made up of many independent, unresolved contributions from across cosmological history, which is potentially accessible to space and ground-based interferometers, and can thus reveal physics at energy scales inaccessible to current colliders [1].
First-order phase transitions (FOPT) associated with spontaneous symmetry breaking in the early Universe are a particularly well-motivated source of a SGWB. As the temperature decreases, the temperature-dependent effective potential of a scalar field ϕ \phi ϕ develops a minimum at a non-zero field value v v v . The minimum at ϕ = 0 \phi= 0 ϕ = 0 becomes a local minimum (false vacuum with unbroken symmetry), and the global minimum is located at ϕ = v \phi= v ϕ = v (true vacuum with spontaneously broken symmetry). In a first-order phase transition, the transition from the false to the true vacuum proceeds via the nucleation of bubbles, inside of which ϕ = v \phi= v ϕ = v . Subsequently, these bubbles expand and combine until the symmetry is broken in the entire Universe [2]. The resulting signal is characterised by a small set of thermodynamical parameters: the transition temperature, strength, duration and bubble wall velocity, all of which we introduce in § 1.1 .
Several mechanisms contribute to the GW signal. Collisions of expanding bubble walls in the scalar field sector produce a characteristic broken-power-law spectrum determined by the typical bubble size at collision. Early studies based on the envelope approximation have been refined by fully relativistic lattice simulations, which confirm the qualitative features of these spectra while highlighting deviations due to nonlinear field dynamics and overlapping bubble walls [3 , 4 ] . Although this contribution is typically short-lived, it can dominate in strongly supercooled transitions, where the Universe is vacuum-dominated during the transition and the GW amplitude is boosted [3 ] . In thermal phase transitions, in contrast, a significant fraction of the released vacuum energy is transferred to the surrounding plasma. After bubble collisions, this energy sources long-lived sound waves in the form of expanding acoustic shells; numerical simulations with relativistic hydrodynamics show that these coherent bulk flows persist for a Hubble time or longer and source gravitational radiation efficiently [5 ] . The acoustic contribution is therefore often the dominant GW source in realistic thermal FOPT, as supported by both numerical and semi-analytical studies [6 , 7 ] . At later times, the breakdown of coherent sound waves may lead to magnetohydrodynamic (MHD) turbulence, an additional, typically subdominant source that generates a broad spectrum whose detailed shape depends on the turbulence model and its decay properties [8 , 9 , 10 ] . Although subject to larger theoretical uncertainties, it is relevant for a complete characterisation of the signal.
The LIGO-Virgo-KAGRA (LVK) collaboration searches for an isotropic SGWB by cross-correlating the data of its detectors. No statistically significant detection has been reported so far, and the resulting upper limits on the SGWB amplitude in the frequency band from roughly 20 to 100 Hz are starting to constrain parts of the parameter space of strongly supercooled FOPT [11 , 12 , 13 , 14 ] . Interpreting these limits in terms of FOPT parameters requires templates for the GW spectrum [15 , 16 ] , a model of the astrophysical foreground from unresolved compact binary coalescences, and a choice of priors for the fit parameters. The earlier recasts of the O1–O3 results [12 , 13 ] relied on the scale-invariant Jeffreys prior only, and the recent release of data from the first part of the fourth observing run (O4a) [14 , 17 ] makes it timely to revisit this procedure.
A central aim of this work is to address a critical gap in the Bayesian inference of constraints on SGWB, the systematic assessment of robustness under prior change. Robust Bayesian inference should be approximately prior-independent to allow for meaningful conclusions, making a comparison of multiple priors essential in order to investigate to which extent the limits inferred from GW data are determined by the likelihood and to which extent they are influenced by assumptions about the prior probability distribution. This question is especially relevant for searches for a SGWB from a FOPT, where the available data do not yet provide a significant detection and where the parameter space spans several orders of magnitude. In such a situation, apparently quantitative upper limits or preferred regions potentially reflect the prior volume as much as the information contained in the data.
We examine three complementary approaches to constraining GW parameters from LVK data. First, a geometric approach extracts limits on the parameters appearing directly in the templates for the GW spectrum, most importantly the peak amplitude and frequency as well as the amplitude of the astrophysical foreground from unresolved mergers. The second approach constrains the characteristic parameters of a FOPT such as the nucleation temperature and the strength of the phase transition. Third, a model-based approach places bounds directly on the Lagrangian parameters of a generic Abelian Higgs model in a dark sector, from which we compute the FOPT parameters (using dimensional reduction to accurately determine the effective scalar potential at finite temperature) and the resulting GW spectrum.
For each approach, we systematically test at least two different priors, including the scale-invariant Jeffreys prior, a prior informed by the detector sensitivity, and priors motivated by the expected distribution of FOPT signals. We then compare the resulting posterior distributions and quantify their degree of agreement. This thorough prior sensitivity analysis demonstrates the profound impact of prior choice and establishes a methodology we advocate for broader adoption across the community.
On the technical side, we also improve the description of the peak of the GW spectrum stemming from sound waves and MHD turbulence, which is needed to relate the geometrical parameters (peak amplitude and frequency) to the break frequencies and amplitudes of the double-broken-power-law (DBPL) template for the spectrum. The details are given in § 1.1.1 .
The remainder of this paper is organised as follows. § 1 reviews the phenomenology of FOPT and collects the spectral templates used throughout the analysis: the broken-power-law (BPL) description of bubble collisions, the DBPL description of sound waves and MHD turbulence including the refined peak relations discussed above, and the astrophysical foreground from compact binary coalescence, from which the cosmological signal must be separated. § 1.2 sets up the Bayesian inference framework by defining the likelihood for the LVK cross-correlation data and introducing as well as motivating the prior choices compared in this work. § 1.2.6 introduces a minimal dark Abelian Higgs model as a concrete particle-physics realisation of a supercooled FOPT, whose parameters are sampled directly. In § 2 , we present our results for the three complementary approaches outlined above: constraints on the geometrical parameters of the templates, constraints on the underlying thermodynamical parameters of the FOPT, and constraints on the fundamental parameters of the Abelian Higgs model. Throughout, the robustness of the constraints under a change of prior is quantified using the diagnostics collected in App. B. We conclude in § 2.3 . Technical details are collected in Appendices § 2.5 –E.
Gravitational wave spectra
Characteristic parameters of first-order phase transitions
A cosmological FOPT is characterised by the following thermodynamical parameters.
The strength parameter α \alpha α , defined as the ratio of latent heat to radiation energy density,
α = 1 ρ r a d [ Δ ρ v a c − T 4 ∂ Δ ρ v a c ∂ T ] T = T n , \alpha= \frac{1}{\rho_{\rm rad}}\left[\Delta\rho_{\rm vac} - \frac{T}{4}\frac{\partial\Delta\rho_{\rm vac}}{\partial T}\right]_{T=T_n}, α = ρ rad 1 [ Δ ρ vac − 4 T ∂ T ∂ Δ ρ vac ] T = T n , where Δ ρ v a c \Delta\rho_{\rm vac} Δ ρ vac is the released vacuum energy density.
For a supercooled FOPT, the nucleation temperature is much lower than the intrinsic symmetry breaking scale, T n ≪ v T_n \ll v T n ≪ v . In the extreme limit α ≫ 1 \alpha\gg1 α ≫ 1 , the false vacuum energy is much larger than the radiation density. Thus, the Universe is vacuum-dominated during the phase transition, which can significantly boost GW amplitudes [3 ] . For the non-experts, we summarise in Table I the differences between supercooled and non-supercooled transitions.
Table I. Qualitative comparison between non-supercooled and supercooled cosmological first-order phase transitions. Supercooling delays bubble nucleation to temperatures well below the symmetry breaking scale v v v , enhancing the vacuum energy release and modifying both bubble dynamics and the resulting stochastic gravitational wave background [7 , 18 , 19 ] .
Spectrum templates for first-order phase transitions
The GW spectrum is represented as
Ω G W ( f ) = 1 ρ c d ρ G W d ln f , (1) \Omega_{\rm GW}(f) = \frac{1}{\rho_c}\frac{d\rho_{\rm GW}}{d\ln f}, \tag*{(1)} Ω GW ( f ) = ρ c 1 d ln f d ρ GW , (1)
where f f f is the frequency, ρ G W \rho_{\rm GW} ρ GW is the GW energy density, and ρ c \rho_c ρ c is the critical density of the Universe. It is difficult to accurately calculate the relations between Ω G W \Omega_{\rm GW} Ω GW and the above thermodynamical parameters. Therefore, bounds from experiments on these parameters cannot be obtained directly.
In practice, the GW spectrum is hence modelled in terms of a template F ( f , f p , q ) \mathcal{F}(f, f_p, q) F ( f , f p , q ) as
Ω G W ( f ) = Ω p F ( f , f p , q ) , (2) \Omega_{\mathrm{GW}}(f) = \Omega_p \mathcal{F}(f, f_p, q), \tag*{(2)} Ω GW ( f ) = Ω p F ( f , f p , q ) , (2)
where Ω p \Omega_p Ω p and f p f_p f p are the peak amplitude and the peak frequency, respectively. The template is obtained from simulations of a given GW source and also depends on a set of additional parameters q q q appropriate for that source. Together with Ω p \Omega_p Ω p and f p f_p f p , they are referred to as geometrical parameters . The data directly constrain these parameters. With the help of relations between geometrical and thermodynamical parameters obtained by analytical calculations and simulations, these constraints on the former parameters can then be matched onto the latter. The SGWB resulting from bubble collisions is commonly modelled by a broken-power-law (BPL) spectrum, whereas spectra originating from sound waves and turbulence have been described by double broken power laws (DBPL) in recent years [15 , 16 ] .
In order to derive constraints from current LVK data, we adopt the approach proposed in Refs. [12 , 13 ] , which is basically testing for the hypothesis that the data contain a signal from a phase transition, using the templates from Ref. [15 ] for the GW spectra, which we will describe in the following.
Bubble collisions
For GW production from bubble collisions and highly relativistic fluid shells, which dominate for FOPT with vacuum domination, i.e., α ≫ 1 \alpha\gg1 α ≫ 1 , we use a BPL,
Ω G W B P L ( f ) = Ω b ( f f b ) n 1 [ 1 2 + 1 2 ( f f b ) a 1 ] n 2 − n 1 a 1 , (3) \Omega_{\mathrm{GW}}^{\mathrm{BPL}}(f) = \Omega_b \left(\frac{f}{f_b}\right)^{n_1} \left[\frac{1}{2} + \frac{1}{2}\left(\frac{f}{f_b}\right)^{a_1}\right]^{\frac{n_2-n_1}{a_1}}, \tag*{(3)} Ω GW BPL ( f ) = Ω b ( f b f ) n 1 [ 2 1 + 2 1 ( f b f ) a 1 ] a 1 n 2 − n 1 , (3)
where f b f_b f b is the break frequency. From the asymptotic behaviour
Ω G W B P L ∝ { f n 1 , f ≪ f b , f n 2 , f ≫ f b , (4) \Omega_{\mathrm{GW}}^{\mathrm{BPL}} \propto\begin{cases} f^{n_1}, & f \ll f_b, \\ f^{n_2}, & f \gg f_b, \end{cases} \tag*{(4)} Ω GW BPL ∝ { f n 1 , f n 2 , f ≪ f b , f ≫ f b , (4)
we see that the spectral slopes n 1 > 0 n_1 > 0 n 1 > 0 and n 2 < 0 n_2 < 0 n 2 < 0 determine the rise at low frequencies and the fall-off at large frequencies, respectively. The smoothness parameter a 1 a_1 a 1 governs the width of the transition region. In our analysis, we fix these parameters to the values given in the first line of Table XXIII , which were determined by simulations [20 ] .
Table XXIII. Spectral slopes and smoothness parameters used in Eqs. (3) and (8) .
The amplitude and frequency of the peak of the spectrum, denoted by Ω p \Omega_p Ω p and f p f_p f p , respectively, are different from Ω b \Omega_b Ω b and f b f_b f b in general and given by Eqs. (75) , (76) . They are related to the parameters characterising the FOPT by Eqs. (77) –(80) , which can be summarised as
h 2 Ω p ≃ 8.2 × 10 − 7 ( 100 g ∗ ( T ∗ ) ) 1 / 3 ( α 1 + α ) 2 ( H ∗ β ) 2 (5) h^2\Omega_p \simeq8.2 \times10^{-7} \left(\frac{100}{g_*(T_*)}\right)^{1/3} \left(\frac{\alpha}{1+\alpha}\right)^2 \left(\frac{H_*}{\beta}\right)^2 \tag*{(5)} h 2 Ω p ≃ 8.2 × 1 0 − 7 ( g ∗ ( T ∗ ) 100 ) 1/3 ( 1 + α α ) 2 ( β H ∗ ) 2 (5)
f p ≃ 1.8 × 10 − 6 H z ( g ∗ ( T ∗ ) 100 ) 1 / 6 ( T ∗ 100 G e V ) ( β H ∗ ) (6) f_p \simeq1.8 \times10^{-6}\,\mathrm{Hz}\left(\frac{g_*(T_*)}{100}\right)^{1/6}\left(\frac{T_*}{100\,\mathrm{GeV}}\right)\left(\frac{\beta}{H_*}\right) \tag*{(6)} f p ≃ 1.8 × 1 0 − 6 Hz ( 100 g ∗ ( T ∗ ) ) 1/6 ( 100 GeV T ∗ ) ( H ∗ β ) (6)
where g ∗ g_* g ∗ is the effective number of degrees of freedom in the thermal bath.
Sound waves
Sound waves and MHD turbulence dominate GW production for α < 1 \alpha< 1 α < 1 . They produce a spectrum that we model as a DBPL,
Ω G W D B P L ( f ) = Ω 2 S ( f ) S ( f 2 ) (7) \Omega_{\mathrm{GW}}^{\mathrm{DBPL}}(f) = \Omega_2 \frac{S(f)}{S(f_2)} \tag*{(7)} Ω GW DBPL ( f ) = Ω 2 S ( f 2 ) S ( f ) (7)
with the shape function
S ( f ) = N ( f f 1 ) n 1 [ 1 + ( f f 1 ) a 1 ] − n 1 − n 2 a 1 [ 1 + ( f f 2 ) a 2 ] − n 2 − n 3 a 2 (8) S(f) = \mathcal{N}\left(\frac{f}{f_1}\right)^{n_1}\left[1+\left(\frac{f}{f_1}\right)^{a_1}\right]^{-\frac{n_1-n_2}{a_1}}\left[1+\left(\frac{f}{f_2}\right)^{a_2}\right]^{-\frac{n_2-n_3}{a_2}} \tag*{(8)} S ( f ) = N ( f 1 f ) n 1 [ 1 + ( f 1 f ) a 1 ] − a 1 n 1 − n 2 [ 1 + ( f 2 f ) a 2 ] − a 2 n 2 − n 3 (8)
where the normalisation factor N \mathcal{N} N ensures that ∫ S ( f ) d log f = 1 \int S(f)\,\mathrm{d}\log f = 1 ∫ S ( f ) d log f = 1 . The break frequencies are f 1 < f 2 f_1 < f_2 f 1 < f 2 , and Ω 2 \Omega_2 Ω 2 is the GW density at the second break. Again, we consider the asymptotic behaviour,
Ω G W D B P L ∝ { f n 1 , f ≪ f 1 , f n 2 , f 1 ≪ f ≪ f 2 , f n 3 , f ≫ f 2 . (9) \Omega_{\mathrm{GW}}^{\mathrm{DBPL}} \propto\begin{cases} f^{n_1}, & f \ll f_1, \\ f^{n_2}, & f_1 \ll f \ll f_2, \\ f^{n_3}, & f \gg f_2. \end{cases} \tag*{(9)} Ω GW DBPL ∝ ⎩ ⎨ ⎧ f n 1 , f n 2 , f n 3 , f ≪ f 1 , f 1 ≪ f ≪ f 2 , f ≫ f 2 . (9)
Thus, the increase at frequencies below the peak is governed by the positive parameters n 1 n_1 n 1 and n 2 n_2 n 2 , while the decrease at higher frequencies is determined by n 3 < 0 n_3 < 0 n 3 < 0 .
For sound waves, we use the values given in the second line of Table XXIII for these parameters as well as the smoothness parameters a 1 a_1 a 1 and a 2 a_2 a 2 [21 ] . The other parameters of the DBPL spectrum are related to the thermodynamical parameters via Eqs. (82) –(88) , which yield
f 2 f 1 ≃ 2.5 Δ w − 1 (10) \frac{f_2}{f_1} \simeq2.5\,\Delta_w^{-1} \tag*{(10)} f 1 f 2 ≃ 2.5 Δ w − 1 (10)
Compared to the BPL case, the bubble wall velocity v w v_w v w becomes relevant, entering via the parameter
Δ w = v s h e l l max ( v w , c s ) (11) \Delta_w = \frac{v_{\mathrm{shell}}}{\max(v_w,c_s)} \tag*{(11)} Δ w = max ( v w , c s ) v shell (11)
that is related to the sound shell thickness, where v s h e l l = ∣ v w − c s ∣ v_{\mathrm{shell}} = \lvert v_w-c_s\rvert v shell = ∣ v w − c s ∣ is the shell velocity and c s = 1 / 3 c_s = 1/\sqrt{3} c s = 1/ 3 is the sound speed.
The peak frequency f p f_p f p is close but not equal to the second break frequency f 2 f_2 f 2 . For f p , f 2 ≫ f 1 f_p, f_2 \gg f_1 f p , f 2 ≫ f 1 , the peak frequency and amplitude can be approximated by relations analogous to Eqs. (75) , (76) [15 ] . For the general case without a strong hierarchy between the break frequencies, we will now derive precise relations that are used in our numerical analysis.
The peak frequency is obtained by maximizing S ( f ) S(f) S ( f ) , or equivalently ln S ( f ) \ln S(f) ln S ( f ) , leading to
n 1 + ( n 2 − n 1 ) ( f p f 1 ) a 1 1 + ( f p f 1 ) a 1 + ( n 3 − n 2 ) ( f p f 2 ) a 2 1 + ( f p f 2 ) a 2 = 0. (12) n_{1}+(n_{2}-n_{1})\frac{\left(\frac{f_{p}}{f_{1}}\right)^{a_{1}}}{1+\left(\frac{f_{p}}{f_{1}}\right)^{a_{1}}}+(n_{3}-n_{2})\frac{\left(\frac{f_{p}}{f_{2}}\right)^{a_{2}}}{1+\left(\frac{f_{p}}{f_{2}}\right)^{a_{2}}}=0. \tag*{(12)} n 1 + ( n 2 − n 1 ) 1 + ( f 1 f p ) a 1 ( f 1 f p ) a 1 + ( n 3 − n 2 ) 1 + ( f 2 f p ) a 2 ( f 2 f p ) a 2 = 0. (12)
Introducing
z = f p f 2 , (13) z=\frac{f_{p}}{f_{2}}, \tag*{(13)} z = f 2 f p , (13)
the peak equation (12) becomes
n 1 + ( n 2 − n 1 ) ( f 2 f 1 ) a 1 z a 1 1 + ( f 2 f 1 ) a 1 z a 1 + ( n 3 − n 2 ) z a 2 1 + z a 2 = 0. (14) n_{1}+(n_{2}-n_{1})\frac{\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}z^{a_{1}}}{1+\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}z^{a_{1}}}+(n_{3}-n_{2})\frac{z^{a_{2}}}{1+z^{a_{2}}}=0. \tag*{(14)} n 1 + ( n 2 − n 1 ) 1 + ( f 1 f 2 ) a 1 z a 1 ( f 1 f 2 ) a 1 z a 1 + ( n 3 − n 2 ) 1 + z a 2 z a 2 = 0. (14)
This equation is transcendental and, in general, must be solved numerically. However, in the limit f 2 / f 1 → ∞ f_{2}/f_{1}\to\infty f 2 / f 1 → ∞ , the first transition has already been completed before the second one is reached, and thus Eq. (14) reduces to
n 2 + ( n 3 − n 2 ) z a 2 1 + z a 2 = 0 , (15) n_{2}+(n_{3}-n_{2})\frac{z^{a_{2}}}{1+z^{a_{2}}}=0, \tag*{(15)} n 2 + ( n 3 − n 2 ) 1 + z a 2 z a 2 = 0 , (15)
whose solution is
z 0 = ( − n 2 n 3 ) 1 / a 2 (16) z_{0}=\left(-\frac{n_{2}}{n_{3}}\right)^{1/a_{2}} \tag*{(16)} z 0 = ( − n 3 n 2 ) 1/ a 2 (16)
or equivalently
f p ( 0 ) = f 2 ( − n 2 n 3 ) 1 / a 2 , (17) f_{p}^{(0)}=f_{2}\left(-\frac{n_{2}}{n_{3}}\right)^{1/a_{2}}, \tag*{(17)} f p ( 0 ) = f 2 ( − n 3 n 2 ) 1/ a 2 , (17)
which is the approximation used in Ref. [15 ] . For finite values of f 2 / f 1 f_{2}/f_{1} f 2 / f 1 , the first transition is not fully saturated. Rather than replacing the first transition factor in Eq. (14) by unity, one may evaluate it at the zeroth-order solution,
R ≡ ( f 2 f 1 ) a 1 z 0 a 1 1 + ( f 2 f 1 ) a 1 z 0 a 1 = ( f 2 f 1 ) a 1 ( − n 2 n 3 ) a 1 / a 2 1 + ( f 2 f 1 ) a 1 ( − n 2 n 3 ) a 1 / a 2 , (18) R\equiv\frac{\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}z_{0}^{a_{1}}}{1+\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}z_{0}^{a_{1}}}=\frac{\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}\left(-\frac{n_{2}}{n_{3}}\right)^{a_{1}/a_{2}}}{1+\left(\frac{f_{2}}{f_{1}}\right)^{a_{1}}\left(-\frac{n_{2}}{n_{3}}\right)^{a_{1}/a_{2}}}, \tag*{(18)} R ≡ 1 + ( f 1 f 2 ) a 1 z 0 a 1 ( f 1 f 2 ) a 1 z 0 a 1 = 1 + ( f 1 f 2 ) a 1 ( − n 3 n 2 ) a 1 / a 2 ( f 1 f 2 ) a 1 ( − n 3 n 2 ) a 1 / a 2 , (18)
and plug it into Eq. (14) to obtain the more accurate approximation
n 1 + ( n 2 − n 1 ) R + ( n 3 − n 2 ) z a 2 1 + z a 2 ≃ 0. (19) n_{1}+(n_{2}-n_{1})R+(n_{3}-n_{2})\frac{z^{a_{2}}}{1+z^{a_{2}}}\simeq0. \tag*{(19)} n 1 + ( n 2 − n 1 ) R + ( n 3 − n 2 ) 1 + z a 2 z a 2 ≃ 0. (19)
This is solved by
z a 2 ≃ − n 1 + ( n 1 − n 2 ) R n 3 + n 1 − n 2 + ( n 2 − n 1 ) R , (20) z^{a_{2}}\simeq\frac{-n_{1}+(n_{1}-n_{2})R}{n_{3}+n_{1}-n_{2}+(n_{2}-n_{1})R}, \tag*{(20)} z a 2 ≃ n 3 + n 1 − n 2 + ( n 2 − n 1 ) R − n 1 + ( n 1 − n 2 ) R , (20)
and therefore
f p ≃ f 2 [ − n 1 + ( n 1 − n 2 ) R n 3 + n 1 − n 2 + ( n 2 − n 1 ) R ] 1 / a 2 . (21) f_{p}\simeq f_{2}\left[\frac{-n_{1}+(n_{1}-n_{2})R}{n_{3}+n_{1}-n_{2}+(n_{2}-n_{1})R}\right]^{1/a_{2}}. \tag*{(21)} f p ≃ f 2 [ n 3 + n 1 − n 2 + ( n 2 − n 1 ) R − n 1 + ( n 1 − n 2 ) R ] 1/ a 2 . (21)
This expression interpolates smoothly between the asymptotic solution (17) for f 2 / f 1 → ∞ f_{2}/f_{1} \to\infty f 2 / f 1 → ∞ (R → 1 R \to1 R → 1 ) and the regime of moderate hierarchy, where the finite value of R − 1 R - 1 R − 1 accounts for the incomplete saturation of the first spectral transition.
Having determined the peak frequency, we can find the peak amplitude Ω p \Omega_{\mathrm{p}} Ω p and its relation to the break amplitude Ω 2 \Omega_{2} Ω 2 by considering Eq. (7) at f = f p f = f_{\mathrm{p}} f = f p ,
Ω p = Ω G W D B P L ( f p ) = Ω 2 S ( f p ) S ( f 2 ) . (22) \Omega_{\mathrm{p}} = \Omega_{\mathrm{GW}}^{\mathrm{DBPL}}(f_{\mathrm{p}}) = \Omega_{2}\frac{S(f_{\mathrm{p}})}{S(f_{2})}. \tag*{(22)} Ω p = Ω GW DBPL ( f p ) = Ω 2 S ( f 2 ) S ( f p ) . (22)
Conveniently, this relation is independent of the normalisation factor N \mathcal{N} N .
Magnetohydrodynamic turbulence
As mentioned, the shape of the GW spectrum produced by MHD turbulence is modelled by a DBPL as well, but now the shape parameters are given by the third line of Table XXIII [15 ] , and the relations with the thermodynamical parameters are given by Eqs. (102) –(106) . They determine the ratio of the break frequencies as
f 2 f 1 ≃ 5.1 Ω s − 1 / 2 , (23) \frac{f_{2}}{f_{1}} \simeq5.1 \Omega_{\mathrm{s}}^{-1/2}, \tag*{(23)} f 1 f 2 ≃ 5.1 Ω s − 1/2 , (23)
where the energy in the GW source stemming from turbulence, Ω s \Omega_{\mathrm{s}} Ω s , appears as a new parameter. It is proportional to the parameter ϵ \epsilon ϵ introduced in § 1.1 , see Eq. (102) . As MHD turbulence yields a subdominant contribution to the GW signal, we include it only in the analysis in terms of thermodynamical parameters.
Astrophysical foregrounds
Unresolved compact binary coalescences (CBC) produce a well-known astrophysical foreground that must be modelled or subtracted when searching for cosmological SGWB. Following previous analyses [11 , 12 , 13 , 14 ] , we assume a power-law spectrum at frequencies relevant for LVK observations,
Ω G W C B C ( f ) = Ω r e f ( f f r e f ) 2 3 , (24) \Omega_{\mathrm{GW}}^{\mathrm{CBC}}(f) = \Omega_{\mathrm{ref}}\left(\frac{f}{f_{\mathrm{ref}}}\right)^{\frac{2}{3}}, \tag*{(24)} Ω GW CBC ( f ) = Ω ref ( f ref f ) 3 2 , (24)
where f r e f f_{\mathrm{ref}} f ref is a reference frequency typically chosen within the frequency range probed by the experiment, conventionally f r e f = 25 f_{\mathrm{ref}} = 25 f ref = 25 Hz, and serves to normalise the spectrum. The amplitude of the SGWB from CBC at this frequency is denoted by Ω r e f \Omega_{\mathrm{ref}} Ω ref .
Other astrophysical contributions like rotating neutron stars, magnetars, and core collapse supernovae can be modelled with specific energy spectrum integrals over cosmic source rates (see, e.g., [11 , 22 , 23 , 24 ] and references therein) but are not included here, as they are assumed to be less important than CBC.
Bayesian inference and LVK stochastic searches
Current limits from LVK stochastic searches
The LVK detectors searched for isotropic and anisotropic stochastic backgrounds using data from observing runs O1–O4a. Employing the cross-correlation between the different detectors to distinguish an SGWB from noise, they reported upper limits on Ω G W ( f ) \Omega_{\mathrm{GW}}(f) Ω GW ( f ) for a power-law spectrum with a range of spectral indices [11 ] . The combined O3 result (including O1 and O2) has been recast into constraints on geometrical and thermodynamical parameters associated with FOPT by assuming templates for Ω G W ( f ) \Omega_{\rm GW}(f) Ω GW ( f ) [12 , 13 ] .
The LVK collaborations have since extended this search to the first part of the fourth observing run (O4a), combining the new data with the full O1–O3 dataset [14 , 17 ] . As in the earlier runs, the O4a cross-correlation analysis finds no statistically significant isotropic stochastic background. For a CBC-like spectrum with spectral index 2 / 3 2/3 2/3 , the search constrains Ω G W C B C ( 25 H z ) ≤ 2.0 × 10 − 9 \Omega_{\rm GW}^{\rm CBC}(25\,\mathrm{Hz}) \leq2.0 \times10^{-9} Ω GW CBC ( 25 Hz ) ≤ 2.0 × 1 0 − 9 at 95% credibility, a factor of 1.7 improvement over the corresponding O3 limit, while a scale-invariant background is constrained to Ω G W ( 25 H z ) ≤ 2.8 × 10 − 9 \Omega_{\rm GW}(25\,\mathrm{Hz}) \leq2.8 \times10^{-9} Ω GW ( 25 Hz ) ≤ 2.8 × 1 0 − 9 , a factor of 2.1 improvement.
In the following, we recast both the older dataset and the latest one into constraints on FOPT parameters by assuming the templates for Ω G W ( f ) \Omega_{\rm GW}(f) Ω GW ( f ) described above. We will refer to the O1–O3 data as “O3” and to the most comprehensive available dataset from runs O1–O4a as “O4a”.
Likelihood model
We adopt a simplified Gaussian likelihood for cross-correlation estimators C I J ( f ) C_{IJ}(f) C I J ( f ) between detectors I I I and J J J [25 ] ,
L ( θ ) = p ( { C I J ( f ) } ∣ θ ) ∝ exp [ − 1 2 ∑ I < J ∑ f [ C I J ( f ) − Γ I J ( f ) Ω G W ( f ; θ ) ] 2 σ I J 2 ( f ) ] . (25) \mathcal{L}(\theta) = p(\{C_{IJ}(f)\}|\theta) \propto\exp\left[-\frac{1}{2}\sum_{I<J}\sum_f \frac{\left[C_{IJ}(f)-\Gamma_{IJ}(f)\Omega_{\rm GW}(f;\theta)\right]^2}{\sigma_{IJ}^2(f)}\right]. \tag*{(25)} L ( θ ) = p ({ C I J ( f )} ∣ θ ) ∝ exp − 2 1 I < J ∑ f ∑ σ I J 2 ( f ) [ C I J ( f ) − Γ I J ( f ) Ω GW ( f ; θ ) ] 2 . (25)
where Γ I J ( f ) \Gamma_{IJ}(f) Γ I J ( f ) is the overlap reduction function{2 } and σ I J 2 ( f ) \sigma_{IJ}^2(f) σ I J 2 ( f ) the variance for the estimator at frequency bin f f f . This form of the likelihood is commonly used when interpreting LVK cross-correlation results (see, e.g., [11 , 13 ] ).
Based on this likelihood, we perform a Bayesian analysis to determine the allowed regions for the fit parameters θ \theta θ , for which we consider three different sets that will be described in the following sections. To sample the parameter space, we use nested sampling [26 , 27 ] as implemented in the dynesty package [28 , 29 ] . Each run starts with a static nested sampling stage, which maps out the overall structure of the likelihood, and then switches to dynamic nested sampling [30 ] , which reallocates the sampling effort to the regions that contribute most to the posterior. We chose this combination because it gives us a well-sampled posterior and the Bayesian evidence from a single run, without having to fix the sampling resolution in advance.
Nested-sampling convergence and evidence stability
To assess the numerical stability of the nested-sampling inference, we compare independent dynesty runs performed with different sampling resolutions. For a fixed likelihood L ( d ∣ θ ) \mathcal{L}(d \mid\theta) L ( d ∣ θ ) and prior π ( θ ) \pi(\theta) π ( θ ) , the Bayesian evidence is
Z = ∫ L ( d ∣ θ ) π ( θ ) d θ . (26) Z = \int\mathcal{L}(d \mid\theta)\,\pi(\theta)\,\mathrm{d}\theta. \tag*{(26)} Z = ∫ L ( d ∣ θ ) π ( θ ) d θ . (26)
Each run provides an estimate Log Z i \operatorname{Log} Z_i Log Z i and an associated numerical uncertainty σ i ≡ σ Log Z i \sigma_i \equiv\sigma_{\operatorname{Log} Z_i} σ i ≡ σ Log Z i . For two independently initialised runs a a a and b b b , we define the evidence-consistency statistic
R Z ( a , b ) ≡ ∣ Log Z a − Log Z b ∣ σ a 2 + σ b 2 . (27) R_Z^{(a,b)} \equiv\frac{\left|\operatorname{Log} Z_a-\operatorname{Log} Z_b\right|}{\sqrt{\sigma_a^2+\sigma_b^2}}. \tag*{(27)} R Z ( a , b ) ≡ σ a 2 + σ b 2 ∣ Log Z a − Log Z b ∣ . (27)
We adopt R Z ( a , b ) ≤ 1 R_Z^{(a,b)} \le1 R Z ( a , b ) ≤ 1 as a practical one-standard-deviation consistency check. This criterion tests whether the difference between the reported evidence estimates is compatible with their combined numerical uncertainties; it is not, by itself, a proof of sampler convergence. In particular, a small R Z R_Z R Z is insufficient when both estimates have large uncertainties. We therefore also inspect the stability of the marginalised posterior distributions and their credible regions as the sampling effort is increased.
The comparison is carried out only between runs with the same dataset, likelihood normalisation, physical model, and prior family and bounds; random seeds and sampling resolution may differ. For dynamic nested sampling, the number of initial live points is one measure of resolution, while the subsequent allocation of samples and stopping settings also affect precision. Evidence comparisons between different prior choices are not interpreted as convergence tests, since the evidence itself depends on the prior. We use this diagnostic to assess numerical stability, separately from the prior-sensitivity measures considered below.
Priors and parameter estimation for geometrical parameters
Bayesian analyses rely on priors to infer posterior probabilities for the parameters that are to be constrained by data. Consequently, in order to obtain statistically robust results from such an analysis it is indispensable to understand and justify the choice of priors and to assess how it affects the output [31 , 32 ] . In physics applications, physical intuition or measurements should guide the choice of appropriate models and parameter priors. Sensible priors are chosen to span physically motivated ranges (e.g., the region of detector sensitivity) and to follow a physically motivated distribution. However, in many cases, it is precisely the lack of well-justified priors that is a challenging part of the analysis (see, e.g., Ref. [33 ] for a discussion in the context of cosmology).
Previous analyses constraining the parameters of FOPT from LVK data [12 , 13 ] used only the Jeffreys prior [34 ] 3 without considering alternatives. Although this is a conservative prior, Bayesian analysis relies on testing different priors, and for this reason, we look for additional well-motivated priors tailored to the LVK data. In the remainder of this section, we justify the priors that we use for the analysis of the density of GW fitted in terms of Eq. (2) for the different templates described in § 1.1.1 .
For the analysis in terms of geometrical parameters, we assume the GW spectrum potentially contained in the data to be the sum of an astrophysical contribution of the form [24 ] and a contribution from a FOPT, described either by a BPL (dominant bubble collisions, Eq. (3) ) or a DBPL (dominant sound waves, Eq. (7) ). Consequently, the geometrical fit parameters characterising the GW spectrum are
Ω r e f \Omega_{\mathrm{ref}} Ω ref , the amplitude of the astrophysical foreground at the reference frequency f r e f = 25 H z f_{\mathrm{ref}} = 25\,\mathrm{Hz} f ref = 25 Hz ,
f p f_p f p , the peak frequency of the SGWB produced by the FOPT,
Ω p \Omega_p Ω p , the corresponding peak amplitude Ω G W B P L ( f p ) \Omega_{\mathrm{GW}}^{\mathrm{BPL}}(f_p) Ω GW BPL ( f p ) or Ω G W D B P L ( f p ) \Omega_{\mathrm{GW}}^{\mathrm{DBPL}}(f_p) Ω GW DBPL ( f p ) , and
Δ w \Delta_w Δ w from Eq. (11) in the case of dominant sound waves; this parameter depends on the bubble wall velocity and determines the position of the first frequency break in the DBPL template.
Jeffreys prior
For strictly positive scale parameters (e.g., cross sections, masses, or amplitudes), the log-uniform (improper Jeffreys) prior takes the form
π ( θ ) ∝ 1 θ [ Log θ max − Log θ min ] ≡ LogU ( θ min , θ max ) , θ ∈ ( θ min , θ max ) . (28) \pi(\theta) \propto\frac{1}{\theta[\operatorname{Log}\theta_{\max} - \operatorname{Log}\theta_{\min}]} \equiv\operatorname{LogU}(\theta_{\min},\theta_{\max}), \quad\theta\in(\theta_{\min},\theta_{\max}). \tag*{(28)} π ( θ ) ∝ θ [ Log θ m a x − Log θ m i n ] 1 ≡ LogU ( θ m i n , θ m a x ) , θ ∈ ( θ m i n , θ m a x ) . (28)
This prior expresses ignorance about the order of magnitude of θ \theta θ rather than its absolute value. It is invariant under rescaling of units, but it is improper without bounds, so one must choose finite limits θ min \theta_{\min} θ m i n and θ max \theta_{\max} θ m a x ; the choice of these limits affects evidence calculations.4 We can use this prior for all fit parameters. For some parameters, such as Δ w \Delta w Δ w , we use its linear counterpart (uniform prior), which is obtained by replacing Log θ → θ \operatorname{Log}\theta\to\theta Log θ → θ and denoted as U ( θ min , θ max ) \operatorname{U}(\theta_{\min},\theta_{\max}) U ( θ m i n , θ m a x ) .
Detector-sensitivity-based prior
When external measurements, theoretical expectations, or calibration data provide information about plausible values of a parameter, a Gaussian (normal) prior is a convenient choice:
π ( θ ) = 1 2 π σ exp [ − ( θ − μ ) 2 2 σ 2 ] ≡ N ( μ , σ ) . (29) \pi(\theta) = \frac{1}{\sqrt{2\pi}\sigma}\exp\left[-\frac{(\theta-\mu)^2}{2\sigma^2}\right] \equiv\operatorname{N}(\mu,\sigma). \tag*{(29)} π ( θ ) = 2 π σ 1 exp [ − 2 σ 2 ( θ − μ ) 2 ] ≡ N ( μ , σ ) . (29)
This prior is useful for encoding existing constraints (e.g., a previously measured particle mass or calibration offset). Choosing a large standard deviation σ \sigma σ yields a weakly informative prior; conversely, a small σ \sigma σ encodes strong prior belief. When physical parameters must remain in a limited range, one often uses a truncated normal.
The Jeffreys prior gives an implicit equal treatment (on a logarithmic scale) over the entire parameter range. Taking into consideration the detector sensitivity, this is not well-justified, since it gives equal probability to parameter regions where sensitivity is poor and to regions of peak sensitivity. For example, in the context of LVK data analyses, using a Jeffreys prior π ( f p / H z ) = LogU ( 10 − 2 , 10 3 ) \pi(f_p/\mathrm{Hz}) = \operatorname{LogU}(10^{-2},10^3) π ( f p / Hz ) = LogU ( 1 0 − 2 , 1 0 3 ) for the peak frequency [13 ] can be considered naive, given that the detectors have negligible sensitivity below about 10 Hz, where, on a logarithmic scale, three of the five decades spanned by this prior (that is, 60% of its probability) are located. This prior also ignores the detector noise, which varies by orders of magnitude across the frequency band. Last but not least, the Hanford-Livingston correlation coefficient varies significantly with frequency; this is not reflected in the Jeffreys prior, but is taken into account in the released public data and encoded in Eq. (25) .
Therefore, we consider instead a detector-sensitivity-based prior, which is a log-normal prior for f p f_p f p centered at the LVK peak sensitivity f s e n s ≃ 40 H z f_{\mathrm{sens}} \simeq40\,\mathrm{Hz} f sens ≃ 40 Hz ,
π s e n s ( f p ) ∝ 1 f p exp [ − ( Log 10 ( f p / f s e n s ) ) 2 2 σ L o g 2 ] ≡ LogN ( Log 10 ( f s e n s / H z ) , σ L o g ) (30) \pi_{\mathrm{sens}}(f_p) \propto\frac{1}{f_p}\exp\left[-\frac{\left(\operatorname{Log}_{10}(f_p/f_{\mathrm{sens}})\right)^2}{2\sigma_{\mathrm{Log}}^2}\right] \equiv\operatorname{LogN}\left(\operatorname{Log}_{10}(f_{\mathrm{sens}}/\mathrm{Hz}),\sigma_{\mathrm{Log}}\right) \tag*{(30)} π sens ( f p ) ∝ f p 1 exp [ − 2 σ Log 2 ( Log 10 ( f p / f sens ) ) 2 ] ≡ LogN ( Log 10 ( f sens / Hz ) , σ Log ) (30)
with σ L o g = 0.8 \sigma_{\mathrm{Log}} = 0.8 σ Log = 0.8 (corresponding to the 1 1 1 -σ \sigma σ bandwidth in Log 10 \operatorname{Log}_{10} Log 10 units). Equivalently, the prior of the logarithm of the peak frequency is given by a Gaussian distribution,
π s e n s ( Log 10 ( f p / H z ) ) = N ( Log 10 40 , 0.8 ) , (31) \pi_{\mathrm{sens}}\left(\operatorname{Log}_{10}(f_p/\mathrm{Hz})\right) = \operatorname{N}\left(\operatorname{Log}_{10}40,0.8\right), \tag*{(31)} π sens ( Log 10 ( f p / Hz ) ) = N ( Log 10 40 , 0.8 ) , (31)
truncated to the frequency interval 10 H z ≤ f p ≤ 1000 H z 10\,\mathrm{Hz} \le f_p \le1000\,\mathrm{Hz} 10 Hz ≤ f p ≤ 1000 Hz where LVK is sensitive.
We will use this prior for the peak frequency, and also in the two prior sets we describe in the following. In those sets, we vary the priors on the SGWB amplitudes in addition.
Model-based prior
Although analyses typically adopt log-uniform or uniform priors for the peak energy density Ω p \Omega_{\mathrm{p}} Ω p of a FOPT signal [11 , 12 , 13 , 14 ] , the distribution of GW amplitudes in realistic particle physics models need not be log-uniform. For example, Ref. [37 ] maps the parameters of the phase transition in a real-singlet extension of the Standard Model onto the peak amplitude and frequency of the stochastic signal, and shows which region of this observable space the model can populate. Investigations of extended scalar sectors such as that of two-Higgs-doublet models showed that viable parameter space points produce a wide range of Ω p \Omega_{\mathrm{p}} Ω p values but not an extremely steep distribution [38 ] . Therefore, we adopt the prior
π ( Ω p ) ∝ Ω p − 0.7 (32) \pi(\Omega_{\mathrm{p}}) \propto\Omega_{\mathrm{p}}^{-0.7} \tag*{(32)} π ( Ω p ) ∝ Ω p − 0.7 (32)
as a model-inspired prior reflecting the distribution of GW amplitudes that emerge in realistic FOPT scenarios with signals strong enough to be tested, in contrast to purely scale-invariant choices. This prior biases the signal amplitude toward smaller values, consistent with both theoretical expectations and the absence of a detection, but includes a fairly heavy tail for strong signals.
In addition, we consider improved estimates of the CBC population based on detections of CBC events and LVK O3 constraints [11 ] . As a consequence, in the model-based prior set we narrow the range of the log-uniform prior for the amplitude of the astrophysical foreground.
Mixed prior
In order to further illustrate the potential consequences of a prior change, we use a mixed prior where we retain the broad Jeffreys prior for the astrophysical contribution Ω r e f \Omega_{\mathrm{ref}} Ω ref but change the prior for the amplitude of the FOPT contribution to a steep power law and a very small lower limit,
π ( Ω p ) ∝ Ω p − 3 , Ω p ∈ ( 10 − 14 , 10 − 6 ) . (33) \pi(\Omega_{\mathrm{p}}) \propto\Omega_{\mathrm{p}}^{-3}, \qquad\Omega_{\mathrm{p}} \in(10^{-14},10^{-6}). \tag*{(33)} π ( Ω p ) ∝ Ω p − 3 , Ω p ∈ ( 1 0 − 14 , 1 0 − 6 ) . (33)
This steep distribution is meant as a deliberately extreme counterpart to the model-based prior rather than as the result of a scan of a specific model. It represents a population dominated by weak, sound-wave-dominated transitions, i.e., many weak and few strong ones. For such transitions the amplitude is strongly suppressed; for instance, Eq. (84) gives h 2 Ω p h^2\Omega_{\mathrm{p}} h 2 Ω p of order 10 − 13 10^{-13} 1 0 − 13 for K ∼ H ∗ R ∗ ∼ 10 − 2 K \sim H_*R_* \sim10^{-2} K ∼ H ∗ R ∗ ∼ 1 0 − 2 , which explains the very small lower limit of the range. This prior thus tests the sensitivity of the posteriors in the direction opposite to that of the model-based prior.
In Table II , we present an overview of the combinations of priors that we use for constraining the geometrical parameters of the templates presented in § 1.1.1 by the O3 data. For the parameter Δ w \Delta_w Δ w , we do not have a preferred range from theory and do not expect tight constraints from the data either; therefore, we simply adopt a uniform prior in all cases.
Table II. Prior specification for the O3 data analysis in terms of geometrical parameters. Ω r e f \Omega_{\rm ref} Ω ref is associated with the astrophysical foreground, while Ω p \Omega_p Ω p and f p f_p f p parameterise the signal from a cosmological phase transition. Log-uniform (Jeffreys), log-normal, and uniform priors are denoted by LogU, LogN, and U, respectively. Here “U”, corresponds to a uniform prior in the corresponding range.
Prior modifications for the O4a analysis
The LVK O4a isotropic SGWB analysis searches for an SGWB using power-law and other spectral models, different from the BPL or DBPL in our analysis. Consequently, there is no unique set of “official O4a priors” for the parameters Ω p \Omega_{\mathrm{p}} Ω p and f p f_{\mathrm{p}} f p . The LVK analysis finds no statistically significant evidence for an isotropic SGWB and provides substantially improved constraints on the amplitude compared with previous observing runs [14 ] . Jeffreys prior To employ these data to constrain the geometrical BPL or DBPL parameters, we adopt a reasonable broad choice for the Jeffreys prior that extends to smaller amplitudes compared to the O3 analysis and is given in the second column of Table III . For the amplitude Ω r e f \Omega_{\rm ref} Ω ref of the astrophysical foreground, we use the same prior as the LVK analysis. The lower limit of 10 − 13 10^{-13} 1 0 − 13 , also adopted for Ω p \Omega_p Ω p , should be understood as a deliberately broad prior rather than as an experimental sensitivity threshold. It allows the posterior to extend well below the region where the LVK detectors are sensitive, thereby avoiding an artificial truncation. Conversely, the conservative upper limits of 10 − 6 10^{-6} 1 0 − 6 and 10 − 5 10^{-5} 1 0 − 5 , respectively, avoid an artificial upper cutoff close to the expected astrophysical contribution.
Table III. Prior specification for the O4a analysis in terms of geometrical parameters. The notation is the same as in Table II .
For the peak frequency, we retain the same broad range as in the O3 analysis. In the spirit of the Jeffreys prior, it is not restricted to the most sensitive part of the LVK band.
Detector-sensitivity-based prior For the O4a analysis, we center the sensitivity-based prior for f p f_p f p on the frequency f s e n s = 25 f_{\rm sens}=25 f sens = 25 Hz, following the LVK stochastic background analysis of the first part of O4. This differs from the 40 Hz frequency used in our prior choice for O3. The new choice is particularly appropriate for the O4a analysis because the LVK results are reported in terms of Ω G W ( 25 H z ) \Omega_{\rm GW}(25\,\mathrm{Hz}) Ω GW ( 25 Hz ) .
Model-based prior The O4a results provide an updated astrophysical reference scale: combining the stochastic-background analysis with the GWTC-4 population inference, the expected total CBC amplitude is
Ω C B C ( 25 H z ) = 0. 9 − 0.5 + 1.1 × 10 − 9 (34) \Omega_{\rm CBC}(25\,\mathrm{Hz})=0.9^{+1.1}_{-0.5}\times10^{-9} \tag*{(34)} Ω CBC ( 25 Hz ) = 0. 9 − 0.5 + 1.1 × 1 0 − 9 (34)
at 90% credibility, with binary black holes providing the dominant contribution [14 ] . We use this astrophysical prediction as motivation for the model-based prior, which centers the range for Ω r e f \Omega_{\rm ref} Ω ref on the central value of 9 × 10 − 10 9\times10^{-10} 9 × 1 0 − 10 , rather than interpreting it as an LVK-reported prior.
The shallow power-law prior for Ω p \Omega_{\mathrm{p}} Ω p is almost the same as for the O3 analysis, with only the lower limit reduced by an order of magnitude.
Mixed prior Here we retain the steep power law for Ω p \Omega_{\mathrm{p}} Ω p with the same range as in the O3 analysis. The priors for Ω r e f \Omega_{\mathrm{ref}} Ω ref and f p f_{\mathrm{p}} f p are adjusted as for the Jeffreys and the detector-sensitivity-based prior, respectively.
Priors and parameter estimation for thermodynamical parameters
For the analysis in terms of thermodynamical parameters, we follow the same approach as for the geometrical parameters in principle. In this case we define two prior sets, where the first choice is again the Jeffreys prior and the other one takes into account the experimental sensitivity. For this second choice, which we call “conditional prior”, we start from the conditions
20 H z < f p < 1000 H z , (35) 20\,\mathrm{Hz}<f_{\mathrm{p}}<1000\,\mathrm{Hz}, \tag*{(35)} 20 Hz < f p < 1000 Hz , (35)
Ω p h 2 > 1 × 10 − 9 (36) \Omega_{\mathrm{p}}h^2>1\times10^{-9} \tag*{(36)} Ω p h 2 > 1 × 1 0 − 9 (36)
for the O4a analysis, in order to restrict the sampled region to signals that the data can actually test. The first constraint keeps only points whose peak frequency lies in the band in which the LVK cross-correlation data are sensitive; for peaks outside this band the likelihood is to a good approximation independent of the phase transition parameters, so such points would only fill the posterior with prior volume. The requirement on Ω p \Omega_{\mathrm{p}} Ω p ensures that we avoid sampling points whose peak amplitude is far below the sensitivity of the respective data set. The threshold in Eq. (36) is of the order of the upper limit on the amplitude of an SGWB at 25 Hz reported by LVK, cf. § 1.2.1 . For the O3 data, this threshold is raised to 6 × 10 − 9 6\times10^{-9} 6 × 1 0 − 9 .
BPL spectrum Assuming a cosmological SGWB dominantly produced by bubble collisions, whose spectrum is thus described by the BPL of Eq. (3) , we use the relevant thermodynamical parameters T ∗ T_\ast T ∗ , α \alpha α , and β / H ∗ \beta/H_\ast β / H ∗ defined in § 1.1 as fit parameters. For simplicity, we set g ∗ ( T ∗ ) = 100 g_\ast(T_\ast)=100 g ∗ ( T ∗ ) = 100 . In addition, the amplitude Ω r e f \Omega_{\mathrm{ref}} Ω ref of the astrophysical foreground is used as before. We present the two prior choices for these parameters in Table IV .
Table IV. Prior specification for the O4a data analysis in terms of the thermodynamical parameters α \alpha α , β / H ∗ \beta/H_\ast β / H ∗ , and T ∗ T_\ast T ∗ described in § 1.1 for a BPL spectrum produced by bubble collisions. As before, Ω r e f \Omega_{\mathrm{ref}} Ω ref parameterises the astrophysical foreground.
The lower edge of the conditional prior on α \alpha α follows from Eq. (5) in combination with the amplitude requirement (36) and the condition for the phase transition to complete, β / H ∗ > 1 \beta/H_\ast>1 β / H ∗ > 1 . For a given value of α \alpha α , Eq. (5) also allows to translate the amplitude requirement into an upper bound on β / H ∗ \beta/H_\ast β / H ∗ , which is the upper prior edge for this parameter. Finally, for each value of β / H ∗ \beta/H_\ast β / H ∗ within its prior range, the frequency constraint (35) yields the prior range for T ∗ T_\ast T ∗ via Eq. (6) .
DBPL spectrum In the analysis constraining an SGWB from weaker FOPT that is characterised by the DBPL spectrum (7) , we take into account both production mechanisms, i.e., the dominant sound waves and the subdominant MHD turbulence. Hence, we have two additional parameters, the bubble wall velocity v w v_w v w and the turbulent energy fraction ϵ \epsilon ϵ . However, in this case, the relations (11) and (D7–D14) between the thermodynamical parameters and the parameters of the DBPL spectrum are fairly complicated. In particular, Eqs. (11) and (D7) as well as the discussion about the efficiency factor κ ( α , v w ) \kappa(\alpha,v_w) κ ( α , v w ) in App. D.§ 1.1.1 show that sampling over α \alpha α , β / H ∗ \beta/H_\ast β / H ∗ , and v w v_w v w would require distinguishing several cases depending on the size of the wall velocity relative to the sound and Jouguet velocities (or adding yet another parameter to sample). Because of that, it is more convenient to trade these fit parameters for suitable combinations. This leads to the following fit parameters for the thermodynamical DBPL analysis in addition to Ω r e f \Omega_{\rm ref} Ω ref :
the kinetic energy fraction K K K defined in Eq. (86) , which is related to the strength of the phase transition,
the characteristic bubble separation H ∗ R ∗ H_*R_* H ∗ R ∗ , see Eq. (81) ,
the nucleation temperature T ∗ T_* T ∗ in GeV,
Δ w \Delta_w Δ w from Eq. (11) parameterising the width of the sound-shell feature, and
the energy fraction of turbulence, ϵ \epsilon ϵ .
This set of thermodynamical fit parameters is very similar to the one in [15 ] , except that we replaced v w v_w v w by Δ w \Delta_w Δ w .
This sampling is particularly important for the combination of sound waves and MHD turbulence, because both sources are controlled by the same two quantities, K K K and H ∗ R ∗ H_*R_* H ∗ R ∗ : the sound-wave break frequencies and amplitude depend on K K K , H ∗ R ∗ H_*R_* H ∗ R ∗ , and Δ w \Delta_w Δ w , see Eqs. (82) –(84) , whereas the turbulent source is characterised by Ω s = ϵ K \Omega_s=\epsilon K Ω s = ϵK and the same H ∗ R ∗ H_*R_* H ∗ R ∗ , see Eqs. (102) –(105) . The combined spectrum is thus a function of ( K , H ∗ R ∗ , Δ w , ϵ ) (K,H_*R_*,\Delta_w,\epsilon) ( K , H ∗ R ∗ , Δ w , ϵ ) and, through H ∗ , 0 H_{*,0} H ∗ , 0 , of T ∗ T_* T ∗ only, whereas a parametrisation in terms of ( α , β / H ∗ , v w , ϵ ) (\alpha,\beta/H_*,v_w,\epsilon) ( α , β / H ∗ , v w , ϵ ) would require the case distinctions discussed above.
The corresponding priors are specified in Table V . Here we apply the frequency and amplitude requirements {(35, 36)} on the sound-wave contribution, since the MHD contribution is subdominant and hence does not affect the peak of the overall spectrum appreciably. In order to determine the edges of the conditional priors, we start by noting that Eq. (81) yields the conservative upper limit H ∗ R ∗ < ( 8 π ) 1 / 3 H_*R_*<(8\pi)^{1/3} H ∗ R ∗ < ( 8 π ) 1/3 , since v w < 1 v_w<1 v w < 1 and β / H ∗ > 1 \beta/H_*>1 β / H ∗ > 1 . Next, we use Eqs. (22) and (84) –(88) to express the peak amplitude in terms of thermodynamical parameters. Introducing the abbreviation
Table V. Prior specification for the O4a data analysis in terms of thermodynamical parameters for a DBPL spectrum. Instead of the strength and duration parameters α \alpha α and β \beta β as well as the wall velocity, we use the combinations K K K , H ∗ R ∗ H_*R_* H ∗ R ∗ , and Δ w \Delta_w Δ w that appear in the relations determining the spectrum. As before, Ω r e f \Omega_{\mathrm{ref}} Ω ref parameterises the astrophysical foreground. The quantities c p ( Δ w ) c_p(\Delta_w) c p ( Δ w ) , H ∗ R ∗ min H_*R_*^{\min} H ∗ R ∗ m i n , and z z z are given in Eq. (38) , Eq. (39) , and Eq. (20) , respectively.
x ≡ f 2 f 1 ≃ 2.5 Δ w , (37) x\equiv\frac{f_2}{f_1}\simeq\frac{2.5}{\Delta_w}, \tag*{(37)} x ≡ f 1 f 2 ≃ Δ w 2.5 , (37)
see Eq. (10) , and plugging in Eq. (74) for g ∗ ( T ∗ ) = 100 g_*(T_*)=100 g ∗ ( T ∗ ) = 100 , we find
Ω p h h 2 = 1.8 × 10 − 6 1 π ( 2 + 2 x 1 + x 2 ) S ( f p ) S ( f 2 ) K 2 H ∗ R ∗ min ( 1 , 2 H ∗ R ∗ 3 K ) ≡ c p ( Δ w ) min ( K 2 H ∗ R ∗ , 2 ( H ∗ R ∗ ) 2 3 K 3 / 2 ) , (38) \begin{aligned}
\Omega_{\rm ph}h^2=1.8\times10^{-6}\frac{1}{\pi}\left(\sqrt{2}+\frac{2x}{1+x^2}\right)\frac{S(f_p)}{S(f_2)}K^2H_*R_*\min\left(1,\frac{2H_*R_*}{\sqrt{3}K}\right) \\
&\equiv c_p(\Delta_w)\min\left(K^2H_*R_*,\frac{2(H_*R_*)^2}{\sqrt{3}}K^{3/2}\right), \tag*{(38)}
\end{aligned} Ω ph h 2 = 1.8 × 1 0 − 6 π 1 ( 2 + 1 + x 2 2 x ) S ( f 2 ) S ( f p ) K 2 H ∗ R ∗ min ( 1 , 3 K 2 H ∗ R ∗ ) ≡ c p ( Δ w ) min ( K 2 H ∗ R ∗ , 3 2 ( H ∗ R ∗ ) 2 K 3/2 ) , (38)
where the ratio S ( f p ) / S ( f 2 ) S(f_p)/S(f_2) S ( f p ) / S ( f 2 ) depends only on x x x and z = f p / f 2 z=f_p/f_2 z = f p / f 2 determined by Eq. (20) , which implies that the quantity c p c_p c p is a function of only Δ w \Delta_w Δ w . As Ω p h h 2 \Omega_{\rm ph}h^2 Ω ph h 2 is the minimum of two expressions that both increase with H ∗ R ∗ H_*R_* H ∗ R ∗ , the amplitude requirement {(36)} is guaranteed to be satisfied if H ∗ R ∗ H_*R_* H ∗ R ∗ is sufficiently large for the Ω p h h 2 \Omega_{\rm ph}h^2 Ω ph h 2 values resulting from both expressions to exceed the threshold. A conservative lower limit is therefore
H ∗ R ∗ > H ∗ R ∗ min ≡ max [ 10 − 9 c p ( Δ w ) K 2 , ( 3 × 10 − 9 2 c p ( Δ w ) ) 1 / 2 ] K − 3 / 4 . (39) H_*R_*>H_*R_*^{\min}\equiv\max\left[\frac{10^{-9}}{c_p(\Delta_w)K^2},\left(\frac{\sqrt{3}\times10^{-9}}{2c_p(\Delta_w)}\right)^{1/2}\right]K^{-3/4}. \tag*{(39)} H ∗ R ∗ > H ∗ R ∗ m i n ≡ max c p ( Δ w ) K 2 1 0 − 9 , ( 2 c p ( Δ w ) 3 × 1 0 − 9 ) 1/2 K − 3/4 . (39)
This condition can only be satisfied simultaneously with H ∗ R ∗ < ( 8 π ) 1 / 3 H_*R_*<(8\pi)^{1/3} H ∗ R ∗ < ( 8 π ) 1/3 if K K K is not too small, resulting in the lower prior edge for K K K ,
K > ( 10 − 9 ( 8 π ) 1 / 3 c p ( Δ w ) ) 1 / 2 . (40) K>\left(\frac{10^{-9}}{(8\pi)^{1/3}c_p(\Delta_w)}\right)^{1/2}. \tag*{(40)} K > ( ( 8 π ) 1/3 c p ( Δ w ) 1 0 − 9 ) 1/2 . (40)
The upper edge is K < 0.6 K < 0.6 K < 0.6 according to Eq. (86) , since the efficiency factor κ \kappa κ in that equation cannot exceed 1.
Finally, we employ the frequency constraint to derive the prior range for T ∗ T_* T ∗ for given values of H ∗ R ∗ H_*R_* H ∗ R ∗ and Δ w \Delta_w Δ w . Using Eqs. (73) and (D9) yields
f p = z f 2 = 1.65 × 10 − 5 H z z 2 Δ w T ∗ H ∗ R ∗ 100 G e V . (41) f_p = z f_2 = 1.65 \times10^{-5}\,\mathrm{Hz}\,\frac{z}{2\Delta_w}\frac{T_*}{H_*R_*\,100\,\mathrm{GeV}}. \tag*{(41)} f p = z f 2 = 1.65 × 1 0 − 5 Hz 2 Δ w z H ∗ R ∗ 100 GeV T ∗ . (41)
Solving for T ∗ T_* T ∗ and applying the constraint (35) on f p f_p f p , we find
2.4 × 10 8 Δ w z H ∗ R ∗ < T ∗ / G e V < 1.2 × 10 10 Δ w z H ∗ R ∗ . (42) 2.4 \times10^8 \frac{\Delta_w}{z}H_*R_* < T_*/\mathrm{GeV} < 1.2 \times10^{10}\frac{\Delta_w}{z}H_*R_*. \tag*{(42)} 2.4 × 1 0 8 z Δ w H ∗ R ∗ < T ∗ / GeV < 1.2 × 1 0 10 z Δ w H ∗ R ∗ . (42)
Direct sampling for U ( 1 ) U(1) U ( 1 ) extensions of the Standard Model
Minimal U ( 1 ) U(1) U ( 1 ) gauge extensions of the Standard Model (SM) are among the simplest and best-motivated classes of physics beyond the SM: adding a single new Abelian factor, spontaneously broken by a complex scalar that gives mass to the associated gauge boson Z ′ Z' Z ′ , is renormalizable and calculable, and its phenomenology is fixed almost entirely by the choice of which SM (or hidden-sector) charges the new symmetry couples to. This single construction therefore covers a remarkably broad range of well-studied scenarios. Gauged B − L B-L B − L (baryon minus lepton number) is one of the simplest anomaly-free extensions of the SM once right-handed neutrinos are included, and is directly tied to the origin of (Majorana) neutrino masses [39 , 40 , 41 ] ; its breaking has been studied as an explicit SGWB source in its own right [42 ] , at a scale set by the same right-handed neutrino Majorana mass that also controls the light-neutrino spectrum, so that a GW detection would probe neutrino-mass physics directly. Likewise, gauged L α − L β L_\alpha-L_\beta L α − L β models (α , β ∈ { e , μ , τ } \alpha,\beta\in\{e,\mu,\tau\} α , β ∈ { e , μ , τ } ) are anomaly-free with the SM fermion content alone. At the opposite extreme, secluded “dark-sector” U ( 1 ) U(1) U ( 1 ) groups, under which no SM field is directly charged, communicating with the visible sector only through kinetic mixing with hypercharge, underlie the broad class of dark-photon models [43 , 44 ] . In each of these realizations, the spontaneous breaking of the new U ( 1 ) U(1) U ( 1 ) symmetry can proceed via a FOPT that sources an SGWB whose parameters ( α , β / H ∗ , T ∗ , v w , … ) (\alpha,\beta/H_*,T_*,v_w,\ldots) ( α , β / H ∗ , T ∗ , v w , … ) of § 2.2 can be computed directly from the Lagrangian parameters of the model (gauge coupling, quartic coupling, VEV) rather than treated as free phenomenological quantities to be scanned independently. This is the central appeal of the direct-sampling approach pursued in this section: it maps an LVK bound directly onto the mass and coupling of a hypothetical new gauge boson, in a form that can be confronted, model by model, with collider, precision-electroweak, and fifth-force constraints on the same particle a level of interpretability that the geometrical and template-based analyses of § 2.1 and § 2.2 do not by themselves provide. Badger et al. [13 ] used exactly this logic for the gauged B − L B-L B − L case, converting LVK O1–O3 posterior/upper-limit results on a phenomenological Ω G W ( f ) \Omega_{\mathrm{GW}}(f) Ω GW ( f ) model into excluded regions of ( m Z ′ , g B − L ) (m_{Z^\prime},g_{B-L}) ( m Z ′ , g B − L ) parameter space; the pipeline we apply below follows the same underlying logic, with the FOPT parameters computed from first principles as described in App. E rather than mapped through the original envelope/sound-wave fits of that analysis.
As a concrete worked example of this generic pipeline, we specialize in this paper to the case of a U ( 1 ) X U(1)_X U ( 1 ) X gauge symmetry living entirely within a secluded dark sector, i.e. one under which no SM field is charged and which communicates with the SM only thermally rather than through any direct portal coupling. We choose this realization for two reasons. First, it is arguably the simplest UV-complete setting in which the direct-sampling pipeline can be demonstrated cleanly: the dark sector’s own dynamics the phase transition and the resulting GW signal decouple entirely from the details of any portal coupling to the SM, which we are therefore free to leave unspecified. Second, among the U ( 1 ) U(1) U ( 1 ) extensions listed above it has the fewest free Lagrangian parameters (just the gauge coupling g g g , the quartic coupling λ \lambda λ , and the VEV v v v ), which keeps the sampling priors, and hence the prior-robustness study that is the focus of this paper, most transparent.
We stress, however, that the pipeline itself is generic: it applies unchanged to any of the more elaborate, SM-charged U ( 1 ) U(1) U ( 1 ) extensions above, such as B − L B-L B − L , or to kinetically-mixed dark photon scenarios. The only difference in those cases is the additional model-building input required to specify how the new sector couples to the SM (charge assignments or portal/kinetic-mixing couplings) and, correspondingly, the additional experimental constraints (collider, beam-dump, precision-electroweak, or fifth-force bounds) that then apply alongside the GW limits derived here.
Here, we consider a generic Abelian Higgs model in a dark sector with a complex scalar Φ \Phi Φ charged under a gauge group U ( 1 ) X U(1)_X U ( 1 ) X and the corresponding gauge boson X μ X_\mu X μ . The Lagrangian is
L = ( D μ Φ ) † ( D μ Φ ) − 1 4 X μ ν X μ ν − V ( Φ ) , (43) \mathcal{L}=(D_\mu\Phi)^\dagger(D^\mu\Phi)-\frac{1}{4}X_{\mu\nu}X^{\mu\nu}-V(\Phi), \tag*{(43)} L = ( D μ Φ ) † ( D μ Φ ) − 4 1 X μν X μν − V ( Φ ) , (43)
where
D μ Φ = ( ∂ μ + i g X μ ) Φ , (44) D_\mu\Phi=(\partial_\mu+igX_\mu)\Phi, \tag*{(44)} D μ Φ = ( ∂ μ + i g X μ ) Φ , (44)
X μ ν = ∂ μ X ν − ∂ ν X μ , (45) X_{\mu\nu}=\partial_\mu X_\nu-\partial_\nu X_\mu, \tag*{(45)} X μν = ∂ μ X ν − ∂ ν X μ , (45)
V ( Φ ) = − μ 2 Φ † Φ + λ ( Φ † Φ ) 2 = λ ( Φ † Φ − v 2 2 ) 2 + c o n s t . (46) V(\Phi)=-\mu^2\Phi^\dagger\Phi+\lambda(\Phi^\dagger\Phi)^2
=\lambda\left(\Phi^\dagger\Phi-\frac{v^2}{2}\right)^2+\mathrm{const.} \tag*{(46)} V ( Φ ) = − μ 2 Φ † Φ + λ ( Φ † Φ ) 2 = λ ( Φ † Φ − 2 v 2 ) 2 + const. (46)
For convenience we have chosen the U ( 1 ) X U(1)_X U ( 1 ) X charge equal to one for the scalar. We assume μ 2 > 0 \mu^2>0 μ 2 > 0 , such that the U ( 1 ) X U(1)_X U ( 1 ) X symmetry is broken spontaneously by a vacuum expectation value v = μ / λ v=\mu/\sqrt{\lambda} v = μ / λ , which yields masses m h = 2 λ v m_h=\sqrt{2\lambda}v m h = 2 λ v for the dark Higgs and m X = g v m_X=gv m X = g v for the gauge boson at tree level.
We assume a coupling between the extra sector and the SM particles (visible sector) that is sufficiently strong to keep both sectors in thermal equilibrium with a common temperature but sufficiently weak to justify neglecting the impact of SM particles on the extra Higgs effective potential and the phase transition. This assumption is needed here precisely because the U ( 1 ) X U(1)_X U ( 1 ) X we sample is secluded: it is automatically satisfied, without any additional input, whenever the new gauge boson couples directly to SM fermions, as in the gauged B − L B-L B − L or L μ − L τ L_\mu-L_\tau L μ − L τ realizations discussed above, since in those cases the new sector shares a common thermal bath with the SM by construction.
To determine whether spontaneous symmetry breaking is associated with a FOPT and what its thermodynamical parameters are, we compute the finite-temperature effective potential of the dark Higgs using dimensional reduction with next-to-leading order (NLO) matching. For each sampled point, we determine the phase structure from the effective potential and calculate the thermal bounce action S 3 ( T ) S_3(T) S 3 ( T ) . The transition temperature is obtained from the percolation condition. We calculate α \alpha α from the effective potential difference and its temperature derivative (see § 1.1 ) and β / H ∗ \beta/H_* β / H ∗ from the temperature derivative of S 3 ( T ) / T S_3(T)/T S 3 ( T ) / T at the transition temperature. The wall velocity is subsequently evaluated in the local thermal equilibrium (LTE) approximation (which we validate in § 2.2.2.1 ) using WallGo [45 ] . Finally, we obtain the GW spectrum by passing the computed thermodynamical parameters to the DBPL template for sound waves and MHD turbulence, fixing ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 for the latter. Further details are presented in App. E.
As we see from Eqs. (43) –(46) , the Abelian Higgs model is fully described by the parameters g g g , λ \lambda λ , and v v v . We sample these together with the astrophysical foreground amplitude Ω r e f \Omega_{\mathrm{ref}} Ω ref , comparing uniform and log-uniform priors with identical ranges for the model parameters, which are shown in Table VI . In both cases, Ω r e f \Omega_{\mathrm{ref}} Ω ref has the same log-uniform prior used previously, L o g U ( 10 − 10 , 10 − 7 ) \mathrm{LogU}(10^{-10},10^{-7}) LogU ( 1 0 − 10 , 1 0 − 7 ) for O3 and L o g U ( 10 − 13 , 10 − 6 ) \mathrm{LogU}(10^{-13},10^{-6}) LogU ( 1 0 − 13 , 1 0 − 6 ) for O4a.
Table VI. Prior specification for the data analysis in terms of Abelian Higgs model parameters.
Results
Results based on different priors for the geometrical parameters
In this section, we present the results in terms of the geometrical parameters Ω r e f \Omega_{\mathrm{ref}} Ω ref , Ω p \Omega_p Ω p , f p / H z f_p/\mathrm{Hz} f p / Hz , and Δ w \Delta_w Δ w , cf. § 1.2.4 . These parameters characterise the overall amplitude and location of the phase-transition signal in frequency space, and therefore provide the most direct way of assessing how much information is supplied by the data themselves.
Sound waves modelled as a double broken power law
For the sound-wave DBPL model we retain the full O3 and O4a analyses side by side, since both runs are sufficiently informative to be compared quantitatively. We are particularly interested in the dependence of the inferred SGWB on the choice of prior, considering the four different prior prescriptions Jeffreys, Detector-Sensitivity-Based, Model-Based, and Mixed, as defined in Table II and Table III . The same Bayesian inference procedure is applied independently to the O3 and O4a stochastic data. The resulting posterior distributions are shown in Figure 1 and Figure 2 for the O3 and O4a datasets, respectively.
Figure 1. Posterior distributions obtained from the public O3 stochastic data for the double broken power law modelling the contribution from sound waves, using the Jeffreys, Detector-Sensitivity-Based, Model-Based, and Mixed prior, respectively. The prior definitions are given in Table II .
Figure 2. Posterior distributions obtained from the O4a stochastic data for the double broken power law modelling the contribution from sound waves, using the Jeffreys, Detector-Sensitivity-Based, Model-Based, and Mixed prior, respectively. The prior definitions are given in Table III .
We concentrate the first part of the discussion on the peak amplitude and frequency, since we varied both prior shapes and ranges for these parameters. For the remaining parameters Ω r e f \Omega_{\mathrm{ref}} Ω ref and Δ w \Delta_w Δ w , we used the same shape for all priors and only varied the prior range in the case of Ω r e f \Omega_{\mathrm{ref}} Ω ref . Hence, the posteriors for these parameters are not expected to vary significantly, except that the extension of the Ω r e f \Omega_{\mathrm{ref}} Ω ref posterior to small values is given by the (varying) lower prior edge in the absence of a detection. This expectation is vindicated by the corresponding parts of Figure 1 and Figure 2 . However, the deduced upper limit on Ω r e f \Omega_{\mathrm{ref}} Ω ref does depend on the prior range [11 ] ; we will return to this issue in § 2.1.1.3 .
In order to quantify the prior-dependence of the results for Ω p \Omega_p Ω p and f p f_p f p , we consider the robustness measures Maximin Robustness Ratio ρ max \rho_{\max} ρ m a x , posterior mean ratio ρ m e a n \rho_{\mathrm{mean}} ρ mean , credible-region overlap ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap , and Kullback-Leibler divergence D K L D_{\mathrm{KL}} D KL . They are defined in App. B, and the values we obtain for the present analysis are given in Table VII and Table VIII . Based on these measures, we classify parameters as ROBUST, MODERATE, or NOT ROBUST under prior change, as specified in Table XXII and Eq. (67) . We also show the 68% credible posterior intervals in Table IX .
Table VII. Maximin Robustness Ratio ρ max \rho_{\max} ρ m a x (Eq. (66) ) for the geometrical parameters of the sound-wave DBPL model, O3 and O4a datasets.
Table VIII. Posterior mean ratio ρ m e a n = max π ⟨ θ ⟩ π / min π ⟨ θ ⟩ π \rho_{\mathrm{mean}} = \max_{\pi}\langle\theta\rangle_{\pi}/\min_{\pi}\langle\theta\rangle_{\pi} ρ mean = max π ⟨ θ ⟩ π / min π ⟨ θ ⟩ π , the ratio of the largest to the smallest posterior mean over the 4 priors (independent of their order), credible-region overlap and Kullback-Leibler divergence (worst case over the pairs of priors) for the geometrical parameters of the sound-wave DBPL model, O3 and O4a datasets. The robustness status is obtained from ρ max \rho_{\max} ρ m a x , ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap and D K L D_{\mathrm{KL}} D KL : ROBUST if at least two of them are ROBUST, MODERATE if one is, NOT ROBUST if none is.
Table IX. 68% credible intervals for the geometrical parameters of the sound-wave DBPL model under the four priors of Tabs. II and III.
Table XXII. Criteria for robustness for ρ max \rho_{\max} ρ m a x and ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap .
O3 stochastic data Figure 1 shows the posterior distributions obtained with the O3 stochastic data. The corner plots display the one-dimensional marginalised posteriors (diagonal panels) together with the two-dimensional correlations among the geometrical parameters. The latter show the 68% and 95% credible regions. The marginalized distributions of both Ω p \Omega_p Ω p and f p f_p f p change noticeably when moving between the different prior prescriptions, while their two-dimensional distributions with the other parameters also show significant changes in shape and correlation structure. This behaviour is expected when the data do not contain sufficient information to localise a peaked cosmological contribution independently of the prior volume. In such a situation, the posterior should not be interpreted as providing a prior-independent measurement of the peak amplitude or peak frequency.
The prior-dependence of the inferred peak frequency is particularly evident in Figure 1 . According to Table IX , the 68% credible intervals extend from [ 3.3 × 10 − 2 , 18.2 ] [3.3 \times10^{-2}, 18.2] [ 3.3 × 1 0 − 2 , 18.2 ] Hz for the Jeffreys prior, the narrowest of the four – the reverse of what one might naively expect from a scale-invariant prior –, to [ 17.9 , 459.1 ] [17.9, 459.1] [ 17.9 , 459.1 ] Hz for the model-based prior, the widest, giving a maximum-to-minimum width ratio ρ max = 24.2 \rho_{\max} = 24.2 ρ m a x = 24.2 in Table VII . The distributions also show a noticeable displacement, leading to a vanishing credible-region overlap and D K L = 0.9 D_{\mathrm{KL}} = 0.9 D KL = 0.9 . Nevertheless, the different posterior regions become substantially more compatible when the 95% credible regions are considered. This distinction is important: overlap at a high credibility level does not imply that the corresponding parameter estimate is insensitive to the prior. Rather, it indicates that the available data still permit a broad range of peak frequencies.
The posterior intervals for Ω p \Omega_p Ω p span substantially different regions of parameter space. The narrowest interval is obtained for the mixed prior, [ 1.1 × 10 − 14 , 2.5 × 10 − 14 ] [1.1 \times10^{-14}, 2.5 \times10^{-14}] [ 1.1 × 1 0 − 14 , 2.5 × 1 0 − 14 ] , and the widest for the Jeffreys prior, [ 3.3 × 10 − 9 , 6.0 × 10 − 6 ] [3.3 \times10^{-9}, 6.0 \times10^{-6}] [ 3.3 × 1 0 − 9 , 6.0 × 1 0 − 6 ] , corresponding to ρ max = 4.2 × 10 8 \rho_{\max} = 4.2 \times10^{8} ρ m a x = 4.2 × 1 0 8 and zero overlap. The posterior mean under the mixed prior lies more than eight orders of magnitude below that of the Jeffreys prior (ρ m e a n = 2.8 × 10 8 \rho_{\mathrm{mean}} = 2.8 \times10^{8} ρ mean = 2.8 × 1 0 8 ). Also the Kullback-Leibler divergence reaches a very large value of D K L = 26.8 D_{\mathrm{KL}} = 26.8 D KL = 26.8 . Thus, the NOT ROBUST label in Table VIII reflects a genuine disagreement in both location and width, not merely a difference in precision.
All in all, Table IX provides a useful quantitative summary of the main message of Figure 1 . In particular, the absence of a statistically significant detection means that the posterior information on Ω p \Omega_p Ω p and f p f_p f p remains strongly influenced by the volume of parameter space assigned by the prior.
O4a stochastic data Figure 2 presents the corresponding analysis using the O4a stochastic data. The improved sensitivity of the dataset generally produces more localized posterior distributions. However, a narrower posterior does not automatically imply a more prior-independent inference. In other words, it is important to distinguish between statistical precision and statistical robustness.
The cosmological peak amplitude and peak frequency are still not well-determined by the data, and their posterior distributions remain sensitive to the adopted prior, as demonstrated by the robustness measures listed in the O4a blocks of Table VII and Table VIII . Most of these measures show an improvement compared to their O3 counterparts. For example, ρ max \rho_{\max} ρ m a x for f p f_p f p falls significantly from 24.2 to 3.1. The relative ρ max \rho_{\max} ρ m a x change for Ω p \Omega_p Ω p is even larger, yet at huge absolute values. We do not find any case where the status of a parameter changes, i.e., it remains NOT ROBUST for both parameters according to all measures. The Kullback-Leibler divergence for Ω p \Omega_p Ω p even increases very slightly.
According to Table IX , it is the same priors as before that yield the broadest and narrowest credible intervals. The Jeffreys and the model-based priors yield the broadest intervals Ω p ∈ [ 1.2 × 10 − 12 , 8.1 × 10 − 8 ] \Omega_p \in[1.2 \times10^{-12}, 8.1 \times10^{-8}] Ω p ∈ [ 1.2 × 1 0 − 12 , 8.1 × 1 0 − 8 ] and f p ∈ [ 16.4 , 390.0 ] f_p \in[16.4, 390.0] f p ∈ [ 16.4 , 390.0 ] Hz, respectively. The narrowest interval for Ω p \Omega_p Ω p , [ 1.1 × 10 − 14 , 2.5 × 10 − 14 ] [1.1 \times10^{-14}, 2.5 \times10^{-14}] [ 1.1 × 1 0 − 14 , 2.5 × 1 0 − 14 ] , is obtained from the mixed prior, while the Jeffreys prior gives the narrowest interval for f p f_p f p , [ 4.7 × 10 − 2 , 118.9 ] [4.7 \times10^{-2}, 118.9] [ 4.7 × 1 0 − 2 , 118.9 ] Hz; the latter is considerably broadened compared to O3, though, resulting in a non-zero ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap .
Taken together, Figure 1 and Figure 2 , and Table VII –Table IX demonstrate that increased detector sensitivity and prior robustness are distinct concepts. The O4a dataset generally reduces the volume of the allowed parameter space. The peak amplitude and peak frequency remain sensitive to the prior even when their posterior distributions become narrower. Therefore, a narrower credible interval should not by itself be interpreted as evidence for a more robust constraint.
Bubble collisions modelled as a broken power law
We now turn to the BPL model describing GW from bubble collisions. In this case we present the O4a results in full, including the robustness diagnostics of App. B for Ω p \Omega_p Ω p and f p f_p f p , and only briefly comment on the O3 results, as the two datasets lead to the same conclusions and the O4a data are more constraining. Figure 3 shows the posterior distributions for the geometrical parameters governing the bubble collision contribution. The four prior prescriptions are the same as for the sound-wave contribution (except for Δ w \Delta_w Δ w , which is not needed here), providing an important cross-check because the BPL parameterization differs from the DBPL description of the sound-wave contribution.
Figure 3. Posterior distributions obtained from the O4a stochastic data for the simple broken power law modelling the contribution from bubble collisions, using the Jeffreys, Detector-Sensitivity-Based, Model-Based, and Mixed prior, respectively. The different priors are defined in Table III .
Table X summarises the robustness diagnostics for the O4a dataset, and Table XI gives the corresponding 68% credible intervals. As for the DBPL analysis, the improved data sensitivity generally leads to more localized posterior regions compared to O3. The peak amplitude is strongly prior-dependent with ρ max = 3.1 × 10 6 \rho_{\max} = 3.1 \times10^{6} ρ m a x = 3.1 × 1 0 6 and ρ o v e r l a p = 0 \rho_{\mathrm{overlap}} = 0 ρ overlap = 0 (NOT ROBUST). The Maximin Robustness Ratio for the peak frequency improves markedly in absolute terms relative to O3, falling from ρ max = 19.8 \rho_{\max} = 19.8 ρ m a x = 19.8 to 2.9 2.9 2.9 . Also the credible-region overlap improves from ρ o v e r l a p = 3.3 × 10 − 3 \rho_{\mathrm{overlap}} = 3.3 \times10^{-3} ρ overlap = 3.3 × 1 0 − 3 to 0.3 0.3 0.3 . However, neither improvement is sufficient to change the classification and it remains NOT ROBUST. The one-dimensional posterior for the peak frequency and its correlations with the amplitude parameters vary visibly between the four choices, and the differences are especially apparent when comparing the broad Jeffreys posterior with the more concentrated Model-Based and Mixed posteriors.
Table X. Maximin Robustness Ratio, posterior mean ratio, credible-region overlap, and Kullback-Leibler divergence for the geometrical parameters of the bubble-collision BPL model, O4a dataset.
Table XI. 68% credible intervals for the geometrical parameters of the bubble-collision BPL model under the four priors of Table III .
The comparison between the BPL results here and the DBPL results in § 2.1.1 is useful for separating two sources of uncertainty. Some differences between the posterior distributions arise from the parameterisation used to describe the GW spectrum, while others arise directly from the adopted prior. The persistence of qualitatively similar prior sensitivity across the two phenomenological descriptions – and across both the O3 and O4a datasets – indicates that this effect is not restricted to a particular choice of BPL template.
Implications for the interpretation of Bayesian limits
The results presented in this section have important consequences for the interpretation of Bayesian upper limits and credible regions. When the likelihood is sufficiently informative, reasonable prior choices lead to similar posterior distributions and the resulting limits can be regarded as approximately data dominated. Conversely, when the likelihood provides only weak information about a parameter, the posterior can retain a significant memory of the prior.
This distinction is clearly visible in the present analysis. The peak parameters Ω p \Omega_p Ω p and, most notably, f p f_p f p show large variations. Therefore, they cannot be assigned a single apparently precise constraint without also specifying the prior assumptions that produced it.
We consequently regard the comparison between priors as an integral part of the inference rather than as an auxiliary consistency check. A constraint is considered more reliable when its posterior support remains substantially unchanged under reasonable variations of the prior. When different priors produce substantially different credible regions, the appropriate conclusion is not that one of the priors is incorrect, but rather that the available data are insufficient to determine that parameter independently of prior assumptions.
This point is particularly important for searches for cosmological SGWB in the absence of a significant detection. In such a regime, quoting a single Bayesian limit without testing its stability against reasonable prior choices can give an incomplete picture of the information supplied by the data . Fitting the geometrical parameters directly has the advantage that it is not necessary to specify how the underlying phase-transition dynamics determine the peak amplitude and frequency of the GW signal. However, this approach still relies on an assumed spectral template and on priors for its parameters. It therefore does not, by itself, guarantee prior-independent limits on a cosmological SGWB. The results of this section therefore motivate the more quantitative robustness measures defined in App. B and applied throughout, where the overlap and relative-width variations of the posterior distributions are used to distinguish robust constraints from parameters that remain prior dominated.
Upper limit on the astrophysical foreground
As mentioned earlier, the posterior distributions of the astrophysical foreground amplitude Ω r e f \Omega_{\rm ref} Ω ref do not vary significantly under the considered priors for this parameter, which all have the same shape but different ranges. In this section, we discuss the effect of the prior range on the 95% CL upper limit on Ω r e f \Omega_{\rm ref} Ω ref . We restrict ourselves to the Jeffreys prior, for which we have already used different ranges in the O3 and O4a analyses, see Table II and Table III . In addition, we consider different combinations of datasets, prior ranges, and GW spectra as well as fit parameters for the cosmological signal. The results are shown in Table XII .
Table XII. 95% CL upper limit on the amplitude Ω r e f \Omega_{\mathrm{ref}} Ω ref of the astrophysical foreground for the Jeffreys prior. The fit parameters are either the geometrical ones of § 1.2.4 or the thermodynamical ones of § 1.2.5 . In both cases the GW spectrum of the phase transition is a BPL for bubble collisions and a DBPL for sound waves, where the thermodynamical analysis also includes MHD turbulence. The fourth column gives the prior range for Ω r e f \Omega_{\mathrm{ref}} Ω ref . The default ranges are those of Table II for O3 and Table III for O4a; for comparison, we also give the limits for O3 data with the O4a range and for O4a data with the O3 range.
We observe that our results are well compatible with those of previous studies in the cases where the same dataset and prior range is used. Specifically, for the O3 data, the limits using the range ( 10 − 13 , 10 − 7 ) (10^{-13},10^{-7}) ( 1 0 − 13 , 1 0 − 7 ) agree with the results of Ref. [13 ] , which used the same range. However, they differ from the limit Ω r e f < 3.4 × 10 − 9 \Omega_{\rm ref}<3.4\times10^{-9} Ω ref < 3.4 × 1 0 − 9 reported by the LVK collaboration [11 ] by almost a factor of 2. That LVK limit is based on a prior range with a much smaller lower edge of 10 − 13 10^{-13} 1 0 − 13 .5 If we perform our analysis with the same lower prior edge, the derived upper limit agrees well with the LVK result.
A dependence of the upper limit on the prior range can be attributed to the fact that the likelihood becomes nearly flat below the detector sensitivity. Thus, decreasing the lower prior bound increases the posterior probability assigned to this region. Consequently, the 95% upper credible limit can be pulled towards smaller values.
Interestingly, we do not observe such a dependence using the O4a dataset. All our analyses yield a limit that is virtually identical to the value of 2.0 × 10 − 9 2.0\times10^{-9} 2.0 × 1 0 − 9 reported by LVK [14 ] , independently of whether we use the same lower edge of 10 − 13 10^{-13} 1 0 − 13 as LVK or the larger value of 10 − 10 10^{-10} 1 0 − 10 from the O3 prior set.
Thus, we find that the dependence of a limit on the prior range can be very different for different datasets. Presumably, it can only be determined by repeating the analysis with different prior ranges.
Results based on different priors for the thermodynamical parameters
The analysis in the previous section demonstrated that the inferred geometrical parameters of the stochastic GW spectrum exhibit a significant dependence on the prior. However, the geometrical parameters are not the fundamental quantities describing the underlying phase transition. In order to assess how the prior dependence propagates to the physical properties of the transition, we next perform the inference directly in terms of the thermodynamical phase-transition parameters. As in § 2.1 , we present the O4a analysis in full for both templates considered below.
In contrast to the geometrical analysis, sampling in the thermodynamical parameter space involves a larger number of parameters and therefore exposes additional degeneracies. Several combinations of thermodynamical parameters can produce very similar GW spectra. Consequently, a constraint on one thermodynamic quantity should not be interpreted independently of the allowed ranges for the other parameters. This makes the dependence on the adopted prior especially relevant when interpreting posterior constraints on the underlying phase transition.
Bubble-collision template
For the bubble-collision contribution, we consider the parameters { α , β / H ∗ , T ∗ } \{\alpha,\beta/H_{\ast},T_{\ast}\} { α , β / H ∗ , T ∗ } , see § 1.1 . The priors are specified in Table IV . Figure 4 shows the posterior distributions obtained from the O4a stochastic data. The two corner plots correspond to the Jeffreys and the Conditional prior, respectively. The corresponding robustness diagnostics are collected in Table XIII and Table XIV .
Figure 4. Posterior distributions for the thermodynamical parameters determining the GW spectrum from bubble collisions, using the O4a stochastic data and two different priors, Jeffreys and Conditional.
Table XIII. Maximin Robustness Ratio, posterior mean ratio, credible-region overlap, and Kullback-Leibler divergence for the thermodynamical parameters of the bubble-collision template, O4a dataset.
Table XIV. 68% credible intervals for the thermodynamical parameters of the bubble-collision template, O4a dataset.
The comparison between the Jeffreys and Conditional priors provides a direct test of the robustness of the thermodynamical inference. The transition strength α \alpha α is ROBUST with ρ max = 1.2 \rho_{\max}=1.2 ρ m a x = 1.2 , ρ o v e r l a p = 0.8 \rho_{\mathrm{overlap}}=0.8 ρ overlap = 0.8 , and D K L = 9.1 × 10 − 2 D_{\mathrm{KL}}=9.1 \times10^{-2} D KL = 9.1 × 1 0 − 2 . This classification is driven by the upper edge of its 68% credible interval in Table XIV , which is similar for the two priors (1.3 for Jeffreys and 1.6 for Conditional), whereas the lower edge still moves with the prior, from 3.8 × 10 − 3 3.8 \times10^{-3} 3.8 × 1 0 − 3 (Jeffreys) to 5.9 × 10 − 2 5.9 \times10^{-2} 5.9 × 1 0 − 2 (Conditional). In contrast, β / H ∗ \beta/H_* β / H ∗ and T ∗ T_* T ∗ are both NOT ROBUST with respect to all measures. For these two parameters the posterior distributions obtained with the two priors overlap only marginally or not at all, and both the extent and the location of the allowed regions change between the two prior choices. This is particularly important for parameters for which the likelihood does not strongly localize the posterior. In such cases, the posterior distribution retains a significant contribution from the prior volume.
The temperature T ∗ T_* T ∗ is also strongly correlated with the other transition parameters and the least robust quantity due to ρ max = 1.8 × 10 3 \rho_{\max}=1.8 \times10^3 ρ m a x = 1.8 × 1 0 3 and ρ o v e r l a p = 0 \rho_{\mathrm{overlap}}=0 ρ overlap = 0 , although D K L = 2.4 D_{\mathrm{KL}}=2.4 D KL = 2.4 is a bit smaller than the value of 6.5 obtained for β / H ∗ \beta/H_* β / H ∗ . Its 68% credible intervals do not overlap at all: the one for the Jeffreys prior, [858.9, 1.0 × 10 7 1.0 \times10^7 1.0 × 1 0 7 ] GeV, lies entirely below the one for the Conditional prior, [1.0 × 10 9 1.0 \times10^9 1.0 × 1 0 9 , 1.9 × 10 10 1.9 \times10^{10} 1.9 × 1 0 10 ] GeV. This behaviour is expected from the relation (D5) between the characteristic frequency of the GW signal, the redshifted Hubble scale at the transition (which is proportional to T ∗ T_* T ∗ ), and the inverse duration β / H ∗ \beta/H_* β / H ∗ . A change in T ∗ T_* T ∗ can be compensated, over part of the parameter space, by a corresponding change in β / H ∗ \beta/H_* β / H ∗ , for which ρ max = 320.2 \rho_{\max}=320.2 ρ m a x = 320.2 and ρ o v e r l a p = 1.5 × 10 − 3 \rho_{\mathrm{overlap}}=1.5 \times10^{-3} ρ overlap = 1.5 × 1 0 − 3 . Going from the Jeffreys to the Conditional prior moves T ∗ T_* T ∗ to higher values and β / H ∗ \beta/H_* β / H ∗ to lower ones, from [4.8, 2.3 × 10 3 2.3 \times10^3 2.3 × 1 0 3 ] to [1.2, 8.3]. The resulting degeneracy is visible in the two-dimensional posterior distributions.
Similarly, the transition strength α \alpha α and the inverse duration β / H ∗ \beta/H_* β / H ∗ are not independently determined by the stochastic data. The amplitude of the bubble-collision contribution depends on the amount of energy released by the transition and on the characteristic duration of the transition. As a consequence, different combinations of α \alpha α and β / H ∗ \beta/H_* β / H ∗ can lead to comparable GW amplitudes.
Figure 4 illustrates an important difference between constraining a phenomenological spectrum and constraining the underlying phase transition. Even when the geometrical parameters of the spectrum appear reasonably localised, the inverse mapping to thermodynamical parameters can introduce substantial degeneracies. The resulting posterior sensitivity to the prior must consequently be taken into account before interpreting the inferred values of α \alpha α , β / H ∗ \beta/H_* β / H ∗ , or T ∗ T_* T ∗ as quantitative measurements of the phase transition.
Sound waves and MHD turbulence
For the combined sound-wave and MHD contribution, it is convenient to use the variables { K , H ∗ R ∗ , T ∗ , Δ w , ε } \{K, H_{*}R_{*}, T_{*}, \Delta_{w}, \varepsilon\} { K , H ∗ R ∗ , T ∗ , Δ w , ε } , as discussed in § 1.2.5 . The priors for these parameters are specified in Table V . Figure 5 shows the results of the corresponding analysis based on the O4a dataset. The increased dimensionality of this parameter space leads to a richer correlation structure than in the bubble-collision analysis. In particular, the posterior distributions reflect that the same amplitude of the GW signal can be obtained through different combinations of the bulk kinetic-energy fraction K K K , the characteristic bubble separation H ∗ R ∗ H_{*}R_{*} H ∗ R ∗ , the sound-shell parameter Δ w \Delta_{w} Δ w , and the turbulence fraction ε \varepsilon ε .
Figure 5. Posterior distributions for the thermodynamical parameters determining the GW spectrum from a combination of sound waves and MHD turbulence, modelled by a double broken power law. The results stem from the O4a stochastic data and two different priors, Jeffreys and Conditional, defined in Table V .
The quantitative robustness diagnostics under a change from Jeffreys to Conditional prior are collected in Table XV and Table XVI . As in § 2.1.1 , we do not present the robustness measures for
Table XV. Robustness measures for the thermodynamical parameters of the sound-wave + MHD template, O4a dataset.
Table XVI. 68% credible intervals for the thermodynamical parameters of the sound-wave + MHD template, O4a dataset.
Ω r e f \Omega_{\mathrm{ref}} Ω ref and Δ w \Delta_{w} Δ w , whose priors we did not vary. For the same reason, the measures for the turbulent energy fraction ϵ \epsilon ϵ are omitted.
The bulk kinetic-energy fraction K K K has a fairly large Kullback-Leibler divergence D K L = 0.9 D_{\mathrm{KL}} = 0.9 D KL = 0.9 but is nevertheless classified as ROBUST due to the remaining diagnostics ρ max = 1.2 \rho_{\max} = 1.2 ρ m a x = 1.2 , ρ o v e r l a p = 0.9 \rho_{\mathrm{overlap}} = 0.9 ρ overlap = 0.9 , and ρ m e a n = 1.0 \rho_{\mathrm{mean}} = 1.0 ρ mean = 1.0 . The lower edge of its 68% credible interval moves from 4.3 × 10 − 4 4.3 \times10^{-4} 4.3 × 1 0 − 4 (Jeffreys) to 2.4 × 10 − 2 2.4 \times10^{-2} 2.4 × 1 0 − 2 (Conditional), while the upper edge, 0.2, is the same for both priors. The mean bubble separation H ∗ R ∗ H_{\ast}R_{\ast} H ∗ R ∗ is NOT ROBUST (ρ max = 3.3 \rho_{\max} = 3.3 ρ m a x = 3.3 , ρ o v e r l a p = 0.2 \rho_{\mathrm{overlap}} = 0.2 ρ overlap = 0.2 , D K L = 0.9 D_{\mathrm{KL}} = 0.9 D KL = 0.9 ), with 68% credible intervals [ 1.3 × 10 − 3 , 0.7 ] [1.3 \times10^{-3}, 0.7] [ 1.3 × 1 0 − 3 , 0.7 ] (Jeffreys) and [ 0.2 , 2.4 ] [0.2, 2.4] [ 0.2 , 2.4 ] (Conditional). The corresponding degeneracies are visible in the two-dimensional posterior distributions of Figure 5 .
The transition temperature T ∗ T_{\ast} T ∗ again appears in correlations with the parameters controlling the characteristic frequency and amplitude of the signal, and it is by far the least robust parameter of this analysis with ρ max = 567.1 \rho_{\max} = 567.1 ρ m a x = 567.1 , ρ o v e r l a p = 0 \rho_{\mathrm{overlap}} = 0 ρ overlap = 0 and D K L = 2.4 D_{\mathrm{KL}} = 2.4 D KL = 2.4 . Its 68% credible intervals, [ 886.5 , 1.0 × 10 7 ] [886.5, 1.0 \times10^{7}] [ 886.5 , 1.0 × 1 0 7 ] GeV for the Jeffreys prior and [ 1.6 × 10 8 , 5.9 × 10 9 ] [1.6 \times10^{8}, 5.9 \times10^{9}] [ 1.6 × 1 0 8 , 5.9 × 1 0 9 ] GeV for the Conditional prior, do not overlap. This reflects the fact that the redshifted Hubble scale at the transition enters directly into the mapping between the phase-transition parameters and the GW spectrum.
Comparison of the O3 and O4a thermodynamical analyses
Figures 4 and 5 provide two complementary descriptions of the O4a constraints. The former considers the bubble-collision contribution, while the latter includes the sound-wave and MHD contributions. Despite the different spectral models and parameterizations, both analyses illustrate the same general feature: the absence of a statistically significant cosmological stochastic background implies that the likelihood does not completely localize the underlying phase-transition parameters, for either run.
This is particularly relevant for the parameters α \alpha α , β / H ∗ \beta/H_{\ast} β / H ∗ , T ∗ T_{\ast} T ∗ , K K K , and H ∗ R ∗ H_{\ast}R_{\ast} H ∗ R ∗ . These quantities enter the GW spectrum through combinations rather than independently, and therefore broad degeneracy directions can remain even when the GW amplitude itself is constrained. A narrow posterior in one direction of parameter space should therefore not be interpreted as a model-independent determination of the corresponding physical quantity.
The comparison of the Jeffreys and Conditional posteriors provides a conservative diagnostic of this effect. Parameters for which the two posterior regions overlap substantially can be regarded as comparatively stable with respect to the choice of prior, whereas parameters for which the posterior regions are displaced or have substantially different widths should be regarded as prior sensitive.
As emphasised in § 2.2.1 and § 2.2.1.1 , moving from O3 to O4a does not uniformly improve the robustness classification of a parameter, even though it typically narrows the individual posteriors: β / H ∗ \beta/H_{\ast} β / H ∗ and H ∗ R ∗ H_{\ast}R_{\ast} H ∗ R ∗ remain firmly NOT ROBUST in both runs despite a ∼ 25 \sim25 ∼ 25 –50 % 50\% 50% reduction in ρ max \rho_{\max} ρ m a x , while T ∗ T_{\ast} T ∗ is, if anything, marginally less robust in O4a than in O3 in both templates. Increased statistical sensitivity is therefore not itself a test of prior independence: a reduction in ρ max \rho_{\max} ρ m a x or in credible-interval width reflects the likelihood becoming more informative, while whether the posterior actually becomes prior independent is a separate question that must be answered with the overlap- and divergence-based diagnostics of App. B, applied to each run individually.
This distinction is important because the transformation from geometrical to thermodynamical parameters is not innocuous from a statistical perspective. A prior that is relatively uninformative in one set of variables need not remain uninformative after a nonlinear transformation. Thus, prior robustness should ideally be tested in the physical parameter space in which the final constraints are quoted, rather than inferred only from the apparent stability of the geometrical posterior, or assumed to improve automatically as more sensitive data become available.
Overall, the thermodynamic analysis reinforces the conclusion obtained from the geometrical parameter study: Bayesian limits on an SGWB should be interpreted together with an explicit, run-by-run assessment of prior dependence. This is particularly important when translating the absence of a detected signal into constraints on fundamental properties of a FOPT.
Results for the Abelian Higgs model
The phenomenological analysis presented in the previous sections constrains the parameters describing the GW spectrum and, subsequently, the thermodynamical properties of the phase transition. We now apply the same inference framework directly to a microscopic realisation of a FOPT, the minimal Abelian Higgs model with gauge group U ( 1 ) X U(1)_X U ( 1 ) X described in § 1.2.6 . This setup provides a simple microscopic framework that determines the phase-transition properties and the resulting GW spectrum by the underlying model parameters. Thus, the constraints from SGWB searches can be propagated to fundamental particle-physics parameters, allowing us to investigate the dependence of the inferred particle-physics parameters on the choice of prior.
We analyse the O4a stochastic data by sampling θ = { Ω r e f , g , λ , v } \theta= \{\Omega_{\mathrm{ref}}, g, \lambda, v\} θ = { Ω ref , g , λ , v } . We compare uniform and log-uniform priors for g g g , λ \lambda λ , and v v v , which are given in Tab. VI and referred to as linear and log prior, respectively, in the following. In Figure 6 , we present the resulting marginalised posterior distributions and the correlations between the sampled parameters for the two priors. Also in this analysis, the distributions of the model parameters change with the prior choice, with the log-uniform prior shifting the posterior support towards smaller values, particularly for λ \lambda λ and v v v . In contrast, the posterior for Ω r e f \Omega_{\mathrm{ref}} Ω ref remains similar between the two analyses, for which its prior is unchanged.
Figure 6. Results for the direct sampling of fundamental parameters of the U ( 1 ) U(1) U ( 1 ) model for the O4a dataset. The plot on the left shows the results for a linear prior and the one on the right those for a logarithmic prior. We have used the LTE approximation in WallGo for this case.
Since these results rely on the LTE approximation for the wall velocity, we first compare this prescription with the full collision-based WallGo [45 ] calculations at selected parameter points. We then examine the 68% credible intervals under the two priors and quantify their differences using the robustness diagnostics introduced previously.
Representative full wall-dynamics cross-checks
The model-level posterior figures in Figure 6 are based on the LTE wall-velocity prescription. To assess its behaviour in selected regions of the NLO dimensionally reduced parameter space, we performed independent, off-equilibrium WallGo calculations with collision integrals for nine representative points drawn from successful LTE scan evaluations. The calculations were carried out at the archived transition temperature T ∗ T_{*} T ∗ for each point, without adjusting the temperature to seek a particular wall solution. The native solver classifications and velocities are collected in Table XVII .
Table XVII. Representative comparisons between the LTE prediction and full collision-based WallGo calculations. N N N denotes the momentum grid size. A dash in v w f u l l v_{w}^{\mathrm{full}} v w full indicates that no stationary velocity was returned; RUNAWAY is the native solver classification, not a measured terminal speed. Point E is a near-Jouguet candidate: its reported velocity equals the Jouguet boundary at both resolutions and must not be treated as an independently converged root. Other single-grid stationary solutions are provisional with respect to grid convergence.
For the four representative subluminal LTE solutions A, B, D and G, the relative differences between the full and LTE velocities are approximately 0.079%, 0.072%, 0.104% and 0.069%, respectively. An independent stationary reference point, checked at N = 5 N = 5 N = 5 , 7 and 9, likewise shows a stable full-wall velocity near 0.5883. These comparisons indicate small off-equilibrium
caption: Central 68% credible intervals of the Abelian Higgs model parameters under the linear and the log prior.
:::
corrections to the wall speed in the tested stationary regime, but are not a systematic error estimate for the entire posterior.
Point E is qualitatively different: its LTE prescription returns v w L T E = 1 v_w^{\mathrm{LTE}} = 1 v w LTE = 1 , whereas the full solver returns the DEFLAGRATION classification with v w f u l l = 0.588903 v_w^{\mathrm{full}} = 0.588903 v w full = 0.588903 at both N = 5 N = 5 N = 5 and N = 7 N = 7 N = 7 . This reported value coincides exactly with the Jouguet velocity v J v_J v J . At N = 7 N = 7 N = 7 the wall pressure changes sign close to v J v_J v J . However, the identical returned values should not be interpreted as proof of velocity convergence: the solution lies at a hydrodynamic branch boundary and may reflect the numerical stopping criterion. We therefore regard E as a near-Jouguet stationary candidate, not as a precision determination of a terminal wall velocity.
By contrast, C, F, H and I are classified RUNAWAY in both the deflagration/hybrid and detonation searches. For C this classification persists from N = 5 N = 5 N = 5 to N = 7 N = 7 N = 7 , with negative wall pressures across the sampled detonation velocities. For F, H and I only the N = 5 N = 5 N = 5 result has been checked. These classifications are conditional on the included particle content, collision treatment, and approximations of the thermal effective description. In particular, the collision integrals here use leading-order thermal mass inputs, while the equilibrium potential is obtained by NLO dimensional reduction. A native RUNAWAY classification should not be identified with a stationary solution at v w = 1 v_w = 1 v w = 1 or a complete prediction for the division of released energy between plasma motion and scalar-field gradients.
The validation therefore supports the LTE velocity approximation for the tested subluminal stationary examples, but also shows that v w L T E = 1 v_w^{\mathrm{LTE}} = 1 v w LTE = 1 alone does not uniquely identify the outcome of the off-equilibrium wall calculation. We consequently retain the LTE-based posterior analysis as the principal result and present these tests as limited wall-dynamics benchmarks; a complete collision-based reanalysis of the GW spectra is beyond the scope of the present study.
Robustness results
The most direct way to see the effect of the prior is to put the 68% credible intervals side by side (Tab. XVIII), because the different credible regions show up immediately. For all parameters except Ω r e f \Omega_{\mathrm{ref}} Ω ref , whose interval is the same under both priors, the log prior moves the whole interval towards smaller values, as expected from a prior that assigns comparatively more weight to the small-value region than a flat one. The shift, however, is far from a rigid translation. For g g g the lower edge falls by a factor of about 1.5 and the upper edge by a factor of about 1.3; for λ \lambda λ the lower edge falls by about a factor of 2 (3.1 × 10 − 3 → \times10^{-3} \rightarrow × 1 0 − 3 → 1.5 × 10 − 3 \times10^{-3} × 1 0 − 3 ) while the upper edge drops by only a factor of 1.2 (8.7 × 10 − 3 → \times10^{-3} \rightarrow × 1 0 − 3 → 7.2 × 10 − 3 \times10^{-3} × 1 0 − 3 ); for v / G e V v/\mathrm{GeV} v / GeV the lower edge falls by a factor of about 50 (1.6 × 10 7 → \times10^{7} \rightarrow × 1 0 7 → 3.0 × 10 5 \times10^{5} × 1 0 5 ) and the upper edge by a factor of 2.6 (8.4 × 10 7 → \times10^{7} \rightarrow × 1 0 7 → 3.2 × 10 7 \times10^{7} × 1 0 7 ). The lower tails are therefore more sensitive to the prior than the upper ones, strikingly so for v v v , which suggests that the likelihood bounds these parameters mainly from above, whereas the small-value side is largely shaped by the prior.
Judging by the width in Table XIX alone, we can see that no interval changes in size by more than a factor 2.1. The couplings g g g and λ \lambda λ are essentially unaffected (ρ max = 1.2 \rho_{\max} = 1.2 ρ m a x = 1.2 and 1.0, respectively), and only v v v exceeds the reference value of 2 for a ROBUST parameter (ρ max = 2.1 \rho_{\max} = 2.1 ρ m a x = 2.1 ), its interval being about twice as wide under the linear prior as under the log prior.
Table XIX. Maximin Robustness Ratio ρ max \rho_{\max} ρ m a x of the 68% credible intervals of the Abelian model, together with the largest and smallest interval width over the two priors. Only v / G e V v/\mathrm{GeV} v / GeV exceeds 2, the reference used in Table XXII ; see App. B for the general definition of ρ max \rho_{\max} ρ m a x (Eq. (66) ), of which this table is the two-prior, direct-sampling analogue.
This diagnostic has a blind spot, though: it measures the size of an interval and is completely insensitive to its position . The self-coupling λ \lambda λ is the clearest example, since its two intervals have practically identical widths (5.6 × 10 − 3 5.6 \times10^{-3} 5.6 × 1 0 − 3 under both priors) and yet sit at visibly different places on the axis, see Tab. XVIII. A small ρ max \rho_{\max} ρ m a x is thus necessary for robustness, but by no means sufficient, which is why the position-sensitive diagnostics below are indispensable.
The posterior means in Table XX tell the same story as the interval edges, but with a sharper contrast. For g g g and λ \lambda λ the mean is lowered by roughly 30% when going from the linear to the log prior (ρ m e a n = 1.4 \rho_{\mathrm{mean}} = 1.4 ρ mean = 1.4 , Δ r e l = 0.3 \Delta_{\mathrm{rel}} = 0.3 Δ rel = 0.3 and 0.4, respectively), which puts both in the MODERATE band between ROBUST and NOT ROBUST. For v / G e V v/\mathrm{GeV} v / GeV , by contrast, the mean is reduced by about 70% (ρ m e a n = 3.5 \rho_{\mathrm{mean}} = 3.5 ρ mean = 3.5 , Δ r e l = 1.1 \Delta_{\mathrm{rel}} = 1.1 Δ rel = 1.1 ), more than twice the level Δ r e l = 0.5 \Delta_{\mathrm{rel}} = 0.5 Δ rel = 0.5 beyond which we declare a parameter NOT ROBUST.
Table XX. Posterior-mean ratio, credible-interval overlap and symmetrised Kullback-Leibler divergence between the linear and the log prior for the Abelian model. The last column is the final status obtained by combining all three diagnostics.
Table XX also adds two diagnostics that look at the posteriors beyond their mean: the interval overlap and the Kullback-Leibler divergence, the latter being sensitive to the full shape of the posteriors. The robustness status deduced from these measures coincides with that from ρ m e a n \rho_{\mathrm{mean}} ρ mean .
Table XXI collects everything in one place. The couplings g g g and λ \lambda λ each pass only one test, the Maximin Robustness Ratio, while v / G e V v/\mathrm{GeV} v / GeV passes none. For g g g the Kullback-Leibler divergence, D K L = 0.24 D_{\mathrm{KL}} = 0.24 D KL = 0.24 (rounded to 0.2 in the table), lies just above the value of 0.2 below which it counts as passed. The parameters therefore fall into a hierarchy of prior dependence, in which g g g and λ \lambda λ are comparably affected by the prior, with g g g marginally less so in all three position-sensitive diagnostics (overlap, mean shift and divergence), while v / G e V v/\mathrm{GeV} v / GeV is the parameter that is largely determined by the prior.
Table XXI. Summary of all robustness diagnostics for the Abelian Higgs model (O4a data). The final status counts how many of the four diagnostics ρ max \rho_{\max} ρ m a x , ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap , Δ r e l = 2 ∣ ⟨ θ ⟩ l i n e a r − ⟨ θ ⟩ l o g ∣ / ( ⟨ θ ⟩ l i n e a r + ⟨ θ ⟩ l o g ) \Delta_{\mathrm{rel}} = 2 \left| \langle\theta\rangle_{\mathrm{linear}} - \langle\theta\rangle_{\mathrm{log}} \right| / \left(\langle\theta\rangle_{\mathrm{linear}} + \langle\theta\rangle_{\mathrm{log}}\right) Δ rel = 2 ∣ ⟨ θ ⟩ linear − ⟨ θ ⟩ log ∣ / ( ⟨ θ ⟩ linear + ⟨ θ ⟩ log ) , and D K L D_{\mathrm{KL}} D KL satisfy the robustness criteria discussed in App. B.
The label MODERATE for g g g and λ \lambda λ rests on a single passed test, the Maximin Robustness Ratio, which, as discussed above, cannot see shifts. The three position-sensitive diagnostics place both parameters in the intermediate band: none of them is ROBUST, but none is in the NOT ROBUST regime either, which is precisely the situation that the label MODERATE is meant to describe. For v / G e V v/\mathrm{GeV} v / GeV , instead, all four diagnostics flag a strong dependence on the prior. However, as we can see from the analysis of § 2.1 and § 2.2 , the robustness tests are much more favourable in this case.
The O4a data thus constrain the Abelian model to a certain extent. The couplings g g g and λ \lambda λ are moderately prior-dependent : their posterior means shift by about 30% between the two priors, which is visible but below the level at which conclusions would change. The parameter v / G e V v/\mathrm{GeV} v / GeV is prior-dependent , with the lower tail of its posterior driven largely by the prior (the lower edge of the interval moves by a factor of about 50) while the upper edge is comparatively stable (a factor of 2.6); the same pattern, much weaker, is seen for λ \lambda λ .
In practice we therefore recommend to state the prior explicitly whenever results on g g g , λ \lambda λ and, in particular, v / G e V v/\mathrm{GeV} v / GeV are quoted, and to report them together with the corresponding result under the alternative prior rather than as a single interval; claims that rest on the upper edges of the intervals are the least exposed to this ambiguity. Since all diagnostics are marginal, a prior dependence of the joint posterior that is not visible in any single parameter cannot be excluded by the present analysis, and the quantitative classification above is meant as a guide to the relative sensitivity of the parameters rather than as a sharp statistical test.
Conclusions and future directions
We have re-examined the Bayesian inference of cosmological FOPT parameters from the LVK stochastic searches in the observing runs O1–O3 (“O3”) and O1–O4a (“O4a”), with the dependence of the results on the prior as our central concern. The inference was set up at three levels: in terms of the geometrical parameters of the spectral templates (Ω r e f \Omega_{\mathrm{ref}} Ω ref , Ω p \Omega_{\mathrm{p}} Ω p , f p f_{\mathrm{p}} f p and, for sound waves, Δ w \Delta_{\mathrm{w}} Δ w ), in terms of the thermodynamical parameters of the transition (α \alpha α , β / H ∗ \beta/H_* β / H ∗ , T ∗ T_* T ∗ for bubble collisions; K K K , H ∗ R ∗ H_*R_* H ∗ R ∗ , Δ w \Delta_{\mathrm{w}} Δ w , T ∗ T_* T ∗ , ϵ \epsilon ϵ for sound waves plus MHD turbulence), and in terms of the Lagrangian parameters of a dark Abelian Higgs model. At each level we compared at least two priors, including the Jeffreys prior, priors informed by the detector sensitivity, and priors motivated by the expected distribution of FOPT signals, which we specified in Table II –Table III for the geometrical parameters and in Table IV –Table V for the thermodynamical parameters. The agreement between the resulting posteriors was quantified by the Maximin Robustness Ratio ρ max \rho_{\max} ρ m a x , the posterior-mean ratio ρ m e a n \rho_{\mathrm{mean}} ρ mean , the credible-region overlap ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap and the Kullback-Leibler divergence D K L D_{\mathrm{KL}} D KL , which are defined in App. B.
On the technical side, we derived a refined approximation for the peak frequency of the double broken power law, Eq. (21) , and the corresponding peak amplitude, Eq. (22) , which is needed to relate these parameters to the break frequencies and amplitudes of the template. The new approximation reproduces the exact peak position to about 2% or better for the sound-wave template, compared to deviations of up to 10% for the zeroth-order result.
Geometrical parameters. No cosmological SGWB is detected. The 95% upper limits on the amplitude of the astrophysical foreground for the Jeffreys prior are Ω r e f < 6 × 10 − 9 \Omega_{\mathrm{ref}} < 6 \times10^{-9} Ω ref < 6 × 1 0 − 9 for O3 and Ω r e f < 2 × 10 − 9 \Omega_{\mathrm{ref}} < 2 \times10^{-9} Ω ref < 2 × 1 0 − 9 for O4a, as reported in Table XII . They are essentially independent of the template used for the FOPT component. However, we emphasise that, for a log-uniform prior, such limits also depend on the prior range assumed for Ω r e f \Omega_{\mathrm{ref}} Ω ref . For the O3 dataset, this causes a variation of the limit by a factor of almost 2, whereas we find a negligible variation for the O4a data.
The parameters of the cosmological signal are not robust with respect to change of prior. For the peak amplitude we find ρ max > 10 6 \rho_{\max} > 10^{6} ρ m a x > 1 0 6 . The posteriors do not overlap and their means differ by orders of magnitude between the priors. The peak frequency improves when going from O3 to O4a data but not sufficiently to reach prior-independence. For example, ρ max \rho_{\max} ρ m a x drops from 24.2 to 3.1 for sound waves. The credible-region overlap increases only to ρ o v e r l a p = 0.3 \rho_{\mathrm{overlap}} = 0.3 ρ overlap = 0.3 for both sound waves and bubble collisions. The Jeffreys posterior, although the narrowest, extends to frequencies of order (0.01–0.1) Hz where the detectors are insensitive, i.e., it is set by the prior to a large extent. This is expected in the absence of a detection: the likelihood does not localise a peak independently of the prior volume.
Thermodynamical parameters. Mapping the data onto the characteristic parameters of the phase transition allows for an interpretation in terms of parameters related to the physical GW source but suffers from degeneracies, since different combinations of thermodynamical parameters can produce very similar GW spectra. For GW produced by bubble-collisions and the O4a dataset, the phase transition strength α \alpha α turns out to be robustly constrained with, for example, a small Kullback-Leibler divergence D K L ≃ 0.1 D_{\mathrm{KL}} \simeq0.1 D KL ≃ 0.1 . The inverse duration β / H ∗ \beta/H_{*} β / H ∗ and the transition temperature T ∗ T_{*} T ∗ are strongly prior-dependent and affected by the degeneracy between T ∗ T_{*} T ∗ and β / H ∗ \beta/H_{*} β / H ∗ in the peak frequency and between α \alpha α and β / H ∗ \beta/H_{*} β / H ∗ in the peak amplitude. Their posterior distributions obtained with the two priors considered overlap only marginally or not at all, and both the extent and the location of the allowed regions change drastically.
For sound waves plus MHD turbulence, we use different thermodynamical parameters that simplify the analysis and avoid case distinctions. Also in this case, the parameter determining the phase transition strength, K K K , is robustly constrained by the O4a data, while no firm conclusions are possible for the mean bubble separation H ∗ R ∗ H_{*}R_{*} H ∗ R ∗ and the temperature T ∗ T_{*} T ∗ due to a strong prior-dependence of their posteriors.
Generally, parameters that directly determine the dominant amplitude of the spectrum are constrained most efficiently, whereas parameters tied to the peak frequency, subdominant contributions, or degenerate combinations remain sensitive to the prior, independently of the LVK run.
Abelian Higgs model. Specifying a benchmark particle physics model in which we computed the phase transition parameters allowed us to map the SGWB search data onto the microphysical parameters (gauge and quartic scalar coupling, vacuum expectation value) appearing in the Lagrangian. We found that the data allow to draw limited conclusions about the couplings, which are moderately prior-dependent, while the posterior for the vacuum expectation value, which sets the energy scale of the phase transition and thus the peak frequency of the GW spectrum, is not robust under a prior change.
Sensitivity versus prior-independence. One important conclusion is that smaller statistical uncertainty does not equate to less prior-dependence. The increased sensitivity obtained from adding the data from observing run O4a narrows most posteriors, but the robustness measures for most fit parameters remain in the NOT ROBUST regime. Thus, prior-independence is a separate property that has to be assessed run by run, for example, using the overlap- and divergence-based diagnostics we discussed here. We therefore advocate treating the comparison of priors as an integral part of the inference, in particular for searches without a significant detection, where a single Bayesian limit or credible region can give an incomplete picture of the information supplied by the data.
Moreover, even for parameters such as the astrophysical foreground amplitude, which are robustly constrained by the data, an upper limit can depend on the prior range used. As a consequence, limits from Bayesian analyses should always be presented together with the employed prior, specifying both its shape and its range.
Direct sampling versus two-step reconstruction. Another, complementary approach is to reconstruct thermodynamical parameters from the geometrical parameters, a technique considered by the LISA Cosmology Working Group [18 ] . Direct sampling from the thermodynamical parameter space represents the conceptually straightforward approach, which we have used. However, this method suffers from significant computational overhead in resolving parameter degeneracies and from the mismatch between parameter space dimensions, since there are 3 (bubble collisions) to 5 (sound waves and turbulence) thermodynamical parameters vs. 2 to 3 geometrical parameters.
The other method corresponds to a two-step reconstruction: (i) constraining the geometrical parameters peak amplitude, peak frequency, and (for DBPL templates) knee parameter, and (ii) converting the reconstructed geometrical parameters to thermodynamical parameters via transformation equations, applying the transformation to each point in the produced geometrical ensemble and taking into account Jacobian factors. This method is preferred for practical implementation due to the degeneracy-free sampling (geometrical parameters exhibit no significant degeneracies, unlike their thermodynamical counterparts), computational efficiency (running geometrical parameter inference first substantially reduces computational cost) and finally dimension matching (the parameter space dimension mismatch inherent in direct thermodynamical sampling is avoided). The two-step approach technically employs different prior assumptions compared to direct sampling due to the variable transformation. However, the practical impact is expected to be negligible because the employed priors are flat, the parameter relationships in logarithmic space are predominantly linear, and any prior transformation effects are minimal. As the main scope of this paper is to confront the validity of sampling with only one prior, we leave an implementation of this approach for future work.
Outlook. LVK stochastic searches already constrain portions of the parameter space for extremely strong supercooled FOPT, though theoretical uncertainties remain significant: the conversion from microphysics to Ω G W ( f ) \Omega_{\mathrm{GW}}(f) Ω GW ( f ) involves assumptions on the bubble wall velocity, the partition of energy between scalar gradients and bulk motion, the lifetime of the sound-wave source, and reheating and dilution effects, and foreground subtraction and non-Gaussian detector noise can bias the inference if not carefully controlled. Improved templates, joint astrophysical/cosmological inference, and upcoming detectors (LISA for lower frequencies, ET/CE for higher sensitivity) will dramatically improve the reach. Careful modelling – including sound-wave lifetime, bubble dynamics, and reheating – is necessary to reliably map GW constraints to particle physics, and the prior-sensitivity analysis advocated here should accompany any such mapping.
Acknowledgements
We thank Jorinde van der Vis, Philipp Schicho, Kylar Greene, and Thomas Konstandin for helpful discussions. The work of L.V.S. is supported by the National Research Foundation of Korea (NRF), grant no. RS-2023-00273508. J.K. was supported by the NRF’s Brain Pool program under grant no. RS-2023-00283129 and the KIAS visitor program. AI tools (Claude, Anthropic) were used for editing tables and formatting in the text; the authors take full responsibility for the content of this paper.
Calculation of the 95% upper limit on Ω r e f \Omega_{\mathrm{ref}} Ω ref
The 95% credible upper limits on Ω r e f \Omega_{\mathrm{ref}} Ω ref quoted in this paper (e.g., in Table XII ) are computed from the output of the nested sampling runs with dynesty. A nested sampling run does not return samples drawn from the posterior but a list of N N N points { x i } i = 1 N \{x_i\}_{i=1}^{N} { x i } i = 1 N (here the values of Ω r e f \Omega_{\mathrm{ref}} Ω ref ), each with an importance weight
w i = L i Δ X i Z , ∑ i = 1 N w i = 1 , (47) w_i = \frac{\mathcal{L}_i \Delta X_i}{\mathcal{Z}}, \qquad\sum_{i=1}^{N} w_i = 1, \tag*{(47)} w i = Z L i Δ X i , i = 1 ∑ N w i = 1 , (47)
where L i \mathcal{L}_i L i is the likelihood at the point, Δ X i \Delta X_i Δ X i the prior-volume shell associated with it, and Z \mathcal{Z} Z the evidence; in dynesty, w i = exp ( log w t i − log Z N ) w_i = \exp(\log\mathrm{wt}_i - \log Z_N) w i = exp ( log wt i − log Z N ) . Posterior statistics must therefore be computed with these weights.
Weighted percentile
Let x ( 1 ) ≤ ⋯ ≤ x ( N ) x_{(1)} \le\cdots\le x_{(N)} x ( 1 ) ≤ ⋯ ≤ x ( N ) denote the sorted points, w ( i ) w_{(i)} w ( i ) the correspondingly ordered weights, and
C ~ k = ∑ i < k w ( i ) ∑ i = 1 N w ( i ) (48) \widetilde{C}_k = \frac{\sum_{i<k} w_{(i)}}{\sum_{i=1}^{N} w_{(i)}} \tag*{(48)} C k = ∑ i = 1 N w ( i ) ∑ i < k w ( i ) (48)
the normalised cumulative weight. The 95% credible upper limit x 95 x_{95} x 95 is the smallest value for which the posterior probability of Ω r e f ≤ x 95 \Omega_{\rm ref} \le x_{95} Ω ref ≤ x 95 reaches 0.95,
∑ { i ∣ x i ≤ x 95 } w i ≥ 0.95. (49) \sum_{\{i \mid x_i \le x_{95}\}} w_i \ge0.95. \tag*{(49)} { i ∣ x i ≤ x 95 } ∑ w i ≥ 0.95. (49)
This is evaluated as follows:
Remove non-finite values by applying the same mask to the points and the weights.
Sort the points in ascending order and apply the same ordering to the weights.
Compute the cumulative sum of the sorted weights and normalise it to [ 0 , 1 ] [0, 1] [ 0 , 1 ] .
Find the smallest index k ∗ k^* k ∗ such that C ~ k ∗ ≥ 0.95 \widetilde{C}_{k^*} \ge0.95 C k ∗ ≥ 0.95 .
The upper limit is x 95 = x ( k ∗ ) x_{95} = x_{(k^*)} x 95 = x ( k ∗ ) .
The result is deterministic: for a given nested sampling run, it is reproduced exactly, in contrast to the resampling approach. In this sense, the limit is taken at a sample point rather than interpolated between points. For the large numbers of live points used, the spacing of neighbouring points is small and the difference from an interpolated quantile is negligible. The remaining uncertainty is the statistical error of the nested sampling run itself, which depends on the number of live points. It can be assessed by repeating a run with different random seeds; it is not included in the values quoted in the tables. The upper limit is a quantile of the posterior. It therefore depends on the prior and for the log-uniform prior used for Ω r e f \Omega_{\rm ref} Ω ref also on the range of the prior (in particular the lower edge, since the likelihood is flat for amplitudes far below the sensitivity). Limits must thus always be quoted together with the prior specified in Table II –Table III and Table V .
Measures of robustness
Let θ \theta θ denote the model parameters and let π ( θ ) \pi(\theta) π ( θ ) be a prior distribution. Given data D D D , the posterior is, [46 , 47 ] ,
p ( θ ∣ D , π ) ∝ L ( D ∣ θ ) π ( θ ) . (50) p(\theta\mid D,\pi) \propto\mathcal{L}(D \mid\theta)\pi(\theta). \tag*{(50)} p ( θ ∣ D , π ) ∝ L ( D ∣ θ ) π ( θ ) . (50)
For two priors, π 1 \pi_1 π 1 and π 2 \pi_2 π 2 , we obtain
p 1 ( θ ∣ D ) = p ( θ ∣ D , π 1 ) , p 2 ( θ ∣ D ) = p ( θ ∣ D , π 2 ) . (51) p_1(\theta\mid D) = p(\theta\mid D,\pi_1), \qquad p_2(\theta\mid D) = p(\theta\mid D,\pi_2). \tag*{(51)} p 1 ( θ ∣ D ) = p ( θ ∣ D , π 1 ) , p 2 ( θ ∣ D ) = p ( θ ∣ D , π 2 ) . (51)
The purpose of the robustness analysis is not to decide which prior is “correct”, but to determine whether the inferred constraints are stable when the prior is changed [48 , 49 , 50 ] . Posterior samples are stored as { θ ( k ) } k = 1 N \{\theta^{(k)}\}_{k=1}^{N} { θ ( k ) } k = 1 N , where each row is a posterior sample and each column corresponds to one parameter. Below we present the most well-known statistical measures [47 , 49 , 50 ] that can be implemented. In § 2.1 and § 2.2 we applied these measures to the geometrical and thermodynamic parameters, respectively, while in § 2.2.2 the measures were extended to the fundamental parameters of the Abelian U ( 1 ) U(1) U ( 1 ) model. All robustness quantities that we present here are statements about the posterior obtained with each prior and computed with the importance weights from Eq. (47) taken into account. They are built from the credible intervals, the posterior means, and the posterior densities.
Note that the corner plots in § 2 show the posterior distributions on logarithmic scales, whereas the credible intervals, posterior means, and resulting robustness measures are given in terms of the fit parameters themselves (i.e., on a linear scale). An alternative approach would be to consider the credible intervals etc. for the logarithms of the parameters, which would naturally result in different values for the robustness measures.
68% credible intervals
For a parameter θ i \theta_i θ i , the analysis uses the central equal-tail credible interval [47 , 51 ] . For credibility level α \alpha α ,
θ i , l o w = Q ( 1 − α ) / 2 ( θ i ) , θ i , h i g h = Q ( 1 + α ) / 2 ( θ i ) . (52) \theta_{i,\mathrm{low}} = Q_{(1-\alpha)/2}(\theta_i), \qquad\theta_{i,\mathrm{high}} = Q_{(1+\alpha)/2}(\theta_i). \tag*{(52)} θ i , low = Q ( 1 − α ) /2 ( θ i ) , θ i , high = Q ( 1 + α ) /2 ( θ i ) . (52)
At 68% credibility [52 ] ,
C 68 ( θ i ) = [ Q 0.16 ( θ i ) , Q 0.84 ( θ i ) ] . (53) C_{68}(\theta_i) = [Q_{0.16}(\theta_i), Q_{0.84}(\theta_i)]. \tag*{(53)} C 68 ( θ i ) = [ Q 0.16 ( θ i ) , Q 0.84 ( θ i )] . (53)
For every parameter and every prior, we record
W i = θ i , h i g h − θ i , l o w , (54) W_i = \theta_{i,\mathrm{high}} - \theta_{i,\mathrm{low}}, \tag*{(54)} W i = θ i , high − θ i , low , (54)
θ ˉ i = 1 N ∑ k = 1 N θ i ( k ) , (55) \bar{\theta}_i = \frac{1}{N}\sum_{k=1}^{N}\theta_i^{(k)}, \tag*{(55)} θ ˉ i = N 1 k = 1 ∑ N θ i ( k ) , (55)
θ i , m e d = Q 0.5 ( θ i ) , (56) \theta_{i,\mathrm{med}} = Q_{0.5}(\theta_i), \tag*{(56)} θ i , med = Q 0.5 ( θ i ) , (56)
as well as the posterior standard deviation.
Credible-region overlap ratio
For two credible regions, define [53 , 54 ]
ρ o v e r l a p = ∣ C α ( π 1 ) ∩ C α ( π 2 ) ∣ ∣ C α ( π 1 ) ∪ C α ( π 2 ) ∣ . (57) \rho_{\mathrm{overlap}} = \frac{\lvert C_{\alpha}(\pi_1) \cap C_{\alpha}(\pi_2)\rvert}{\lvert C_{\alpha}(\pi_1) \cup C_{\alpha}(\pi_2)\rvert}. \tag*{(57)} ρ overlap = ∣ C α ( π 1 ) ∪ C α ( π 2 )∣ ∣ C α ( π 1 ) ∩ C α ( π 2 )∣ . (57)
For the parameter-by-parameter implementation, the regions are intervals, C 1 = [ L 1 , U 1 ] C_1 = [L_1, U_1] C 1 = [ L 1 , U 1 ] , C 2 = [ L 2 , U 2 ] C_2 = [L_2, U_2] C 2 = [ L 2 , U 2 ] . Their intersection has length
I = max [ 0 , min ( U 1 , U 2 ) − max ( L 1 , L 2 ) ] , (58) I = \max[0, \min(U_1, U_2) - \max(L_1, L_2)], \tag*{(58)} I = max [ 0 , min ( U 1 , U 2 ) − max ( L 1 , L 2 )] , (58)
and their union has length
U = max ( U 1 , U 2 ) − min ( L 1 , L 2 ) . (59) U = \max(U_1, U_2) - \min(L_1, L_2). \tag*{(59)} U = max ( U 1 , U 2 ) − min ( L 1 , L 2 ) . (59)
Hence
ρ o v e r l a p = I U . (60) \rho_{\mathrm{overlap}} = \frac{I}{U}. \tag*{(60)} ρ overlap = U I . (60)
The ratio satisfies 0 ≤ ρ o v e r l a p ≤ 1 0 \le\rho_{\mathrm{overlap}} \le1 0 ≤ ρ overlap ≤ 1 . The limiting cases are
ρ o v e r l a p = 1 ⇒ identical credible intervals , (61) \rho_{\mathrm{overlap}} = 1 \quad\Rightarrow\quad\text{identical credible intervals}, \tag*{(61)} ρ overlap = 1 ⇒ identical credible intervals , (61)
ρ o v e r l a p = 0 ⇒ non-overlapping credible intervals . (62) \rho_{\mathrm{overlap}} = 0 \quad\Rightarrow\quad\text{non-overlapping credible intervals}. \tag*{(62)} ρ overlap = 0 ⇒ non-overlapping credible intervals . (62)
Comparison of several priors
For priors π 1 , π 2 , … , π M \pi_1,\pi_2,\ldots,\pi_M π 1 , π 2 , … , π M , the overlap ratio is evaluated for every unique pair ( π a , π b ) (\pi_a,\pi_b) ( π a , π b ) , a < b a < b a < b . For each parameter θ i \theta_i θ i this gives the pairwise overlap
ρ o v e r l a p a b ( θ i ) ≡ ρ o v e r l a p ( θ i ; π a , π b ) = ∣ C 68 ( π a ) ∩ C 68 ( π b ) ∣ ∣ C 68 ( π a ) ∪ C 68 ( π b ) ∣ , (63) \rho_{\mathrm{overlap}}^{ab}(\theta_i) \equiv\rho_{\mathrm{overlap}}(\theta_i;\pi_a,\pi_b) = \frac{\left|C_{68}(\pi_a) \cap C_{68}(\pi_b)\right|}{\left|C_{68}(\pi_a) \cup C_{68}(\pi_b)\right|}, \tag*{(63)} ρ overlap ab ( θ i ) ≡ ρ overlap ( θ i ; π a , π b ) = ∣ C 68 ( π a ) ∪ C 68 ( π b ) ∣ ∣ C 68 ( π a ) ∩ C 68 ( π b ) ∣ , (63)
which is simply Eq. (57) evaluated at the 68% credibility level for the specific pair ( π a , π b ) (\pi_a,\pi_b) ( π a , π b ) . Values close to unity indicate that the inferred parameter region is stable under the change of prior, whereas small values indicate that the posterior support changes substantially. Equivalently, for more than two priors one may construct a symmetric overlap matrix
R a b ( i ) = ρ o v e r l a p ( θ i ; π a , π b ) , (64) R_{ab}^{(i)} = \rho_{\mathrm{overlap}}(\theta_i;\pi_a,\pi_b), \tag*{(64)} R ab ( i ) = ρ overlap ( θ i ; π a , π b ) , (64)
with
R a a ( i ) = 1 , R a b ( i ) = R b a ( i ) . (65) R_{aa}^{(i)} = 1,\qquad R_{ab}^{(i)} = R_{ba}^{(i)}. \tag*{(65)} R aa ( i ) = 1 , R ab ( i ) = R ba ( i ) . (65)
This pairwise approach avoids privileging one prior as a reference. In § 2.2 and § 2.2.2 only two priors are compared at a time (Jeffreys versus Conditional [34 , 55 , 56 ] , or Jeffreys versus the model’s Conditional prior), so Eq. (63) reduces to a single number per parameter; in § 2.1 , where four priors (Jeffreys, Detector-Based, Model-Based, Mixed) are compared, the tabulated ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap is the minimum of R a b ( i ) R_{ab}^{(i)} R ab ( i ) over all six pairs, so that a parameter is only reported as robust if every pair of priors agrees.
Maximin Robustness Ratio
The overlap ratio probes agreement between the locations and extents of the credible intervals. A complementary diagnostic,
ρ max = max π W i ( π ) min π W i ( π ) , (66) \rho_{\max} = \frac{\max_{\pi} W_i(\pi)}{\min_{\pi} W_i(\pi)}, \tag*{(66)} ρ m a x = min π W i ( π ) max π W i ( π ) , (66)
measures changes in their widths (Eq. (54) ) [49 ] . Ratios > 2.0 > 2.0 > 2.0 indicate substantial prior-dependent uncertainty.
These diagnostics are complementary. Two priors can produce similarly located intervals but substantially different widths, or intervals of similar width that are shifted relative to one another. This is why, throughout § 2 , an improvement in ρ max \rho_{\max} ρ m a x between the O3 and O4a runs (i.e., a narrower posterior under the more sensitive dataset) is never by itself taken as evidence of improved prior independence: only a joint improvement in ρ max \rho_{\max} ρ m a x and ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap (and, where available, in the Kullback-Leibler divergence [57 ] introduced below) supports that conclusion.
For the adopted diagnostic classification,
ρ o v e r l a p > 0.7 ⇒ ROBUST , 0.4 < ρ o v e r l a p ≤ 0.7 ⇒ MODERATE , ρ o v e r l a p ≤ 0.4 ⇒ NOT ROBUST . (67) \begin{aligned}
\rho_{\mathrm{overlap}} &> 0.7 &&\Rightarrow\text{ROBUST},\\
0.4 < \rho_{\mathrm{overlap}} &\le0.7 &&\Rightarrow\text{MODERATE},\\
\rho_{\mathrm{overlap}} &\le0.4 &&\Rightarrow\text{NOT ROBUST}. \tag*{(67)}
\end{aligned} ρ overlap 0.4 < ρ overlap ρ overlap > 0.7 ≤ 0.7 ≤ 0.4 ⇒ ROBUST , ⇒ MODERATE , ⇒ NOT ROBUST . (67)
These thresholds are diagnostic criteria for the analysis rather than universal statistical thresholds. The implemented overlap ratio is a one-dimensional, parameter-by-parameter measure. It compares the marginal 68% credible intervals of an individual parameter under two priors. It should therefore not be interpreted as the volume overlap of the full multidimensional 68% posterior regions. Such a quantity would require constructing the corresponding joint credible sets and evaluating their multidimensional volumes.
The one-dimensional measure is nevertheless useful for reporting the robustness of individual parameter constraints: it is symmetric, transparent, easy to reproduce, and directly connected to the posterior intervals that are reported in the analysis.
The prior-robustness analysis combines two complementary pieces of information. First, the 68% credible interval describes the constraint obtained under each prior, Eq. (53) . Second, the overlap ratio Eq. (63) quantifies the common part of two such constraints. Together with the interval-width ratio ρ max \rho_{\max} ρ m a x of Eq. (66) , these quantities provide a compact assessment of whether the posterior constraints are stable against reasonable changes of prior. Robust inference is indicated by substantial overlap between credible intervals and by limited changes in their widths.
Posterior mean discrepancy and Kullback-Leibler divergence
In order to complement the width-based quantities ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap and ρ max \rho_{\max} ρ m a x , we consider several further diagnostics. The relative discrepancy between posterior means under two priors π 1 \pi_1 π 1 and π 2 \pi_2 π 2 is
Δ r e l ( θ i ) = ∣ 1 − E [ θ i ∣ D , π 1 ] E [ θ i ∣ D , π 2 ] ∣ . (68) \Delta_{\mathrm{rel}}(\theta_i) = \left|1-\frac{\mathbb{E}[\theta_i \mid D,\pi_1]}{\mathbb{E}[\theta_i \mid D,\pi_2]}\right|. \tag*{(68)} Δ rel ( θ i ) = 1 − E [ θ i ∣ D , π 2 ] E [ θ i ∣ D , π 1 ] . (68)
It probes the stability of the posterior location independently of its width, and can therefore disagree with ρ max \rho_{\max} ρ m a x for a given parameter (see § 2.2.2.2 ).
For an arbitrary number of priors, ρ m e a n \rho_{\mathrm{mean}} ρ mean is the ratio of the largest to the smallest posterior mean
ρ m e a n = max π E [ θ i ∣ D , π ] / min π E [ θ i ∣ D , π ] ≥ 1. (69) \rho_{\mathrm{mean}} = \max_{\pi}\mathbb{E}[\theta_i\mid D,\pi] \,/\, \min_{\pi}\mathbb{E}[\theta_i\mid D,\pi] \ge1. \tag*{(69)} ρ mean = π max E [ θ i ∣ D , π ] / π min E [ θ i ∣ D , π ] ≥ 1. (69)
It does not depend on the order of the priors.
Finally, the symmetrised Kullback-Leibler divergence between two posteriors [55 , 57 , 58 , 59 ] ,
D K L s y m ( π 1 , π 2 ) = 1 2 [ K L ( p 1 ∥ p 2 ) + K L ( p 2 ∥ p 1 ) ] , (70) D_{\mathrm{KL}}^{\mathrm{sym}}(\pi_1,\pi_2) = \frac{1}{2}\left[\mathrm{KL}(p_1\Vert p_2)+\mathrm{KL}(p_2\Vert p_1)\right], \tag*{(70)} D KL sym ( π 1 , π 2 ) = 2 1 [ KL ( p 1 ∥ p 2 ) + KL ( p 2 ∥ p 1 ) ] , (70)
is a full-posterior-shape diagnostic that is sensitive to differences in location, width, and higher moments simultaneously (see Eq. (70) ). As with ρ o v e r l a p \rho_{\mathrm{overlap}} ρ overlap and ρ max \rho_{\max} ρ m a x , neither Δ r e l \Delta_{\mathrm{rel}} Δ rel nor D K L s y m D_{\mathrm{KL}}^{\mathrm{sym}} D KL sym should be used in isolation: a combined classification, as in Table XXI , requires all four diagnostics together.
General redshifting relations
The present-day frequency and energy density of GW produced in a FOPT are related to their values at production by [15 ]
f = a ∗ a 0 f ∗ = H ∗ , 0 f ∗ H ∗ , (71) f = \frac{a_*}{a_0}f_* = H_{*,0}\frac{f_*}{H_*}, \tag*{(71)} f = a 0 a ∗ f ∗ = H ∗ , 0 H ∗ f ∗ , (71)
h 2 Ω G W ( f ) = h 2 F G W , 0 Ω ∗ ( f ∗ ) . (72) h^2\Omega_{\mathrm{GW}}(f) = h^2F_{\mathrm{GW},0}\,\Omega_*(f_*). \tag*{(72)} h 2 Ω GW ( f ) = h 2 F GW , 0 Ω ∗ ( f ∗ ) . (72)
The redshifted Hubble rate and energy-density factor are
H ∗ , 0 ≃ 1.65 × 10 − 5 H z ( g ∗ ( T ∗ ) 100 ) 1 / 6 ( T ∗ 100 G e V ) (73) H_{\ast,0} \simeq1.65 \times10^{-5}\,\mathrm{Hz}\left(\frac{g_{\ast}(T_{\ast})}{100}\right)^{1/6}\left(\frac{T_{\ast}}{100\,\mathrm{GeV}}\right) \tag*{(73)} H ∗ , 0 ≃ 1.65 × 1 0 − 5 Hz ( 100 g ∗ ( T ∗ ) ) 1/6 ( 100 GeV T ∗ ) (73)
h 2 F G W , 0 = h 2 ( a ∗ a 0 ) 4 ( H ∗ H 0 ) 2 ≃ 1.64 × 10 − 5 ( 100 g ∗ ( T ∗ ) ) 1 / 3 . (74) h^2 F_{\mathrm{GW},0}=h^2\left(\frac{a_{\ast}}{a_0}\right)^4\left(\frac{H_{\ast}}{H_0}\right)^2 \simeq1.64 \times10^{-5}\left(\frac{100}{g_{\ast}(T_{\ast})}\right)^{1/3}. \tag*{(74)} h 2 F GW , 0 = h 2 ( a 0 a ∗ ) 4 ( H 0 H ∗ ) 2 ≃ 1.64 × 1 0 − 5 ( g ∗ ( T ∗ ) 100 ) 1/3 . (74)
Details of spectrum templates
Here we collect all details regarding the GW templates from [15 ] that have not been given in the main text.
Template I: Broken power law
The BPL template is given by Eq. (3) . The peak frequency and peak amplitude are
f p = f b ( n 1 n 2 ) 1 / a 1 , (75) f_{\mathrm{p}}=f_{\mathrm{b}}\left(\frac{n_1}{n_2}\right)^{1/a_1}, \tag*{(75)} f p = f b ( n 2 n 1 ) 1/ a 1 , (75)
Ω p = Ω b [ 1 2 ( − n 2 n 1 ) n 1 n 1 − n 2 + 1 2 ( − n 1 n 2 ) − n 2 n 1 − n 2 ] n 1 − n 2 a 1 . (76) \Omega_{\mathrm{p}}=\Omega_{\mathrm{b}}\left[\frac{1}{2}\left(-\frac{n_2}{n_1}\right)^{\frac{n_1}{n_1-n_2}}+\frac{1}{2}\left(-\frac{n_1}{n_2}\right)^{\frac{-n_2}{n_1-n_2}}\right]^{\frac{n_1-n_2}{a_1}}. \tag*{(76)} Ω p = Ω b [ 2 1 ( − n 1 n 2 ) n 1 − n 2 n 1 + 2 1 ( − n 2 n 1 ) n 1 − n 2 − n 2 ] a 1 n 1 − n 2 . (76)
For strong transitions, modelled after bubble collisions, the relevant energy fraction is
K ~ = α 1 + α . (77) \widetilde{K}=\frac{\alpha}{1+\alpha}. \tag*{(77)} K = 1 + α α . (77)
The peak amplitude and frequency today read
h 2 Ω p = h 2 F G W , 0 A s t r K ~ 2 ( H ∗ β ) 2 , (78) h^2\Omega_{\mathrm{p}}=h^2F_{\mathrm{GW},0}A_{\mathrm{str}}\widetilde{K}^2\left(\frac{H_{\ast}}{\beta}\right)^2, \tag*{(78)} h 2 Ω p = h 2 F GW , 0 A str K 2 ( β H ∗ ) 2 , (78)
f p ≃ 0.11 H ∗ , 0 β H ∗ , (79) f_{\mathrm{p}}\simeq0.11H_{\ast,0}\frac{\beta}{H_{\ast}}, \tag*{(79)} f p ≃ 0.11 H ∗ , 0 H ∗ β , (79)
where
A s t r ≃ 0.05 , (80) A_{\mathrm{str}}\simeq0.05, \tag*{(80)} A str ≃ 0.05 , (80)
and the rest of the parameters, corresponding to the redshifting parameters, are defined in App. § 2.7 .
Template II: Double broken power law
The double broken power law (DBPL) spectrum is given by Eqs. (7) , (8) . Depending on the production mechanism, the relevant parameters and relations between them vary. In all cases, the mean bubble separation, R ∗ R_{\ast} R ∗ at the end of the PT is relevant and estimated through [7 ]
R ∗ H ∗ = ( 8 π ) 1 / 3 max ( v w , c s ) H ∗ β , (81) R_{\ast}H_{\ast}=(8\pi)^{1/3}\max(v_{\mathrm{w}},c_{\mathrm{s}})\frac{H_{\ast}}{\beta}, \tag*{(81)} R ∗ H ∗ = ( 8 π ) 1/3 max ( v w , c s ) β H ∗ , (81)
where c s = 1 / 3 c_{\mathrm{s}}=1/\sqrt{3} c s = 1/ 3 is the sound speed. Note that we write R ∗ H ∗ R_{\ast}H_{\ast} R ∗ H ∗ as this combination appears explicitly in various fits.
Sound-wave contribution
The frequency breaks are
f 1 ≃ 0.2 H ∗ , 0 ( H ∗ R ∗ ) − 1 , (82) f_1 \simeq0.2 H_{*,0}(H_*R_*)^{-1}, \tag*{(82)} f 1 ≃ 0.2 H ∗ , 0 ( H ∗ R ∗ ) − 1 , (82)
f 2 ≃ 0.5 H ∗ , 0 Δ w − 1 ( H ∗ R ∗ ) − 1 , (83) f_2 \simeq0.5 H_{*,0}\Delta_w^{-1}(H_*R_*)^{-1}, \tag*{(83)} f 2 ≃ 0.5 H ∗ , 0 Δ w − 1 ( H ∗ R ∗ ) − 1 , (83)
where the bubble wall velocity v w v_w v w enters via the parameter Δ w \Delta_w Δ w defined in Eq. (11) . We use
h 2 Ω i n t = h 2 F G W , 0 A s w K 2 ( H ∗ τ s w ) ( H ∗ R ∗ ) . (84) h^2\Omega_{\mathrm{int}}=h^2F_{\mathrm{GW},0}A_{\mathrm{sw}}K^2(H_*\tau_{\mathrm{sw}})(H_*R_*). \tag*{(84)} h 2 Ω int = h 2 F GW , 0 A sw K 2 ( H ∗ τ sw ) ( H ∗ R ∗ ) . (84)
for the integrated GW energy density Ω i n t = ∫ Ω G W B P L ( f ) d log f \Omega_{\mathrm{int}}=\int\Omega_{\mathrm{GW}}^{\mathrm{BPL}}(f)\,\mathrm{d}\log f Ω int = ∫ Ω GW BPL ( f ) d log f , where [7 , 21 ]
A s w ≃ 0.11 , (85) A_{\mathrm{sw}}\simeq0.11, \tag*{(85)} A sw ≃ 0.11 , (85)
K ≃ 0.6 κ ( α , v w ) α 1 + α , (86) K\simeq0.6\kappa(\alpha,v_w)\frac{\alpha}{1+\alpha}, \tag*{(86)} K ≃ 0.6 κ ( α , v w ) 1 + α α , (86)
H ∗ τ s w ≃ min ( 1 , 2 H ∗ R ∗ 3 K ) . (87) H_*\tau_{\mathrm{sw}}\simeq\min\left(1,\frac{2H_*R_*}{\sqrt{3K}}\right). \tag*{(87)} H ∗ τ sw ≃ min ( 1 , 3 K 2 H ∗ R ∗ ) . (87)
The integrated energy density is related to the parameter Ω 2 \Omega_2 Ω 2 in Eq. (7) by [15 ]
Ω 2 = 1 π ( 2 + 2 f 2 / f 1 1 + f 2 2 / f 1 2 ) Ω i n t . (88) \Omega_2=\frac{1}{\pi}\left(\sqrt{2}+\frac{2f_2/f_1}{1+f_2^2/f_1^2}\right)\Omega_{\mathrm{int}}. \tag*{(88)} Ω 2 = π 1 ( 2 + 1 + f 2 2 / f 1 2 2 f 2 / f 1 ) Ω int . (88)
The stochastic GW signal produced by sound waves depends on the fraction of the released vacuum energy that is converted into bulk kinetic energy of the plasma. This fraction is parameterized by the hydrodynamic efficiency factor κ ( α , v w ) \kappa(\alpha,v_w) κ ( α , v w ) in Eq. (86) . The factor 0.6 reflects the correction due to the quantity K / 0.6 K/0.6 K /0.6 representing the contribution from a single bubble. Unlike the early Jouguet approximation, which assumes that the wall propagates at the Jouguet velocity, v w = v J ( α ) v_w=v_J(\alpha) v w = v J ( α ) , modern treatments regard the wall velocity as an independent hydrodynamic variable determined by microscopic friction. Consequently, κ \kappa κ must be evaluated throughout the entire hydrodynamic parameter space. This treatment is suitable for parameter inference and Bayesian analyses in which the wall velocity is sampled independently. The Jouguet velocity is given by
v J ( α ) = α ( 2 3 + α ) + 1 3 1 + α . (89) v_J(\alpha)=\frac{\sqrt{\alpha\left(\frac{2}{3}+\alpha\right)+\frac{1}{3}}}{1+\alpha}. \tag*{(89)} v J ( α ) = 1 + α α ( 3 2 + α ) + 3 1 . (89)
Following the fitting formulae of Ref. [19 ] , calibrated to numerical hydrodynamic solutions, we define the auxiliary quantities
κ A = 6.9 α 1.36 − 0.037 α + α , (90) \kappa_A=\frac{6.9\alpha}{1.36-0.037\sqrt{\alpha}+\alpha}, \tag*{(90)} κ A = 1.36 − 0.037 α + α 6.9 α , (90)
κ B = α 2 / 5 0.017 + ( 0.997 + α ) 2 / 5 , (91) \kappa_B=\frac{\alpha^{2/5}}{0.017+(0.997+\alpha)^{2/5}}, \tag*{(91)} κ B = 0.017 + ( 0.997 + α ) 2/5 α 2/5 , (91)
κ C = α 0.135 + 0.98 + α , (92) \kappa_C=\frac{\sqrt{\alpha}}{0.135+\sqrt{0.98+\alpha}}, \tag*{(92)} κ C = 0.135 + 0.98 + α α , (92)
κ D = α 0.73 + 0.083 α + α , (93) \kappa_D=\frac{\alpha}{0.73+0.083\sqrt{\alpha}+\alpha}, \tag*{(93)} κ D = 0.73 + 0.083 α + α α , (93)
together with
δ κ = − 0.9 ln ( α 1 + α ) . (94) \delta_\kappa=-0.9\ln\left(\frac{\sqrt{\alpha}}{1+\sqrt{\alpha}}\right). \tag*{(94)} δ κ = − 0.9 ln ( 1 + α α ) . (94)
The efficiency factor is then evaluated continuously over the three hydrodynamic regimes. For subsonic deflagrations, we have
v w < c s , (95) v_w < c_s, \tag*{(95)} v w < c s , (95)
and the efficiency factor is
κ = c s 11 / 5 κ A κ B ( c s 11 / 5 − v w 11 / 5 ) κ B + v w c s 6 / 5 κ A . (96) \kappa= \frac{c_s^{11/5}\kappa_A\kappa_B}{\left(c_s^{11/5}-v_w^{11/5}\right)\kappa_B+v_wc_s^{6/5}\kappa_A}. \tag*{(96)} κ = ( c s 11/5 − v w 11/5 ) κ B + v w c s 6/5 κ A c s 11/5 κ A κ B . (96)
For supersonic deflagrations,
c s ≤ v w < v J , (97) c_s \le v_w < v_J, \tag*{(97)} c s ≤ v w < v J , (97)
we consider the hybrid solutions, for which we introduce
x = v w − c s v J − c s , (98) x = \frac{v_w-c_s}{v_J-c_s}, \tag*{(98)} x = v J − c s v w − c s , (98)
and interpolate according to
κ = κ B + ( v w − c s ) δ κ + x 3 [ κ C − κ B − ( v J − c s ) δ κ ] . (99) \kappa= \kappa_B+(v_w-c_s)\delta\kappa+x^3\left[\kappa_C-\kappa_B-(v_J-c_s)\delta\kappa\right]. \tag*{(99)} κ = κ B + ( v w − c s ) δ κ + x 3 [ κ C − κ B − ( v J − c s ) δ κ ] . (99)
This interpolation provides a smooth transition between the deflagration and detonation branches and avoids introducing artificial discontinuities near the Jouguet point. For detonations,
v w ≥ v J , (100) v_w \ge v_J, \tag*{(100)} v w ≥ v J , (100)
and the efficiency factor becomes
κ = ( v J − 1 ) 3 v J 5 / 2 v w − 5 / 2 κ C κ D [ ( v J − 1 ) 3 − ( v w − 1 ) 3 ] v J 5 / 2 κ C + ( v w − 1 ) 3 κ D . (101) \kappa= \frac{(v_J-1)^3v_J^{5/2}v_w^{-5/2}\kappa_C\kappa_D}{\left[(v_J-1)^3-(v_w-1)^3\right]v_J^{5/2}\kappa_C+(v_w-1)^3\kappa_D}. \tag*{(101)} κ = [ ( v J − 1 ) 3 − ( v w − 1 ) 3 ] v J 5/2 κ C + ( v w − 1 ) 3 κ D ( v J − 1 ) 3 v J 5/2 v w − 5/2 κ C κ D . (101)
Magnetohydrodynamic turbulence
MHD turbulence is expected to be subdominant but not necessarily negligible. In this case, the spectrum parameters listed in Table XXIII were taken from Ref. [15 ] , where they were determined based on the simulations in Ref. [60 ] . The turbulent energy fraction is
Ω s = ϵ κ , (102) \Omega_s=\epsilon\kappa, \tag*{(102)} Ω s = ϵ κ , (102)
where ϵ \epsilon ϵ represents an unknown fraction of overall kinetic energy in bulk motion that is converted to MHD turbulence. The break frequencies are then given by
f 1 = 3 Ω s 2 N H ∗ , 0 ( H ∗ R ∗ ) − 1 , (103) f_1=\frac{\sqrt{3\Omega_s}}{2\mathcal{N}}H_{*,0}(H_*R_*)^{-1}, \tag*{(103)} f 1 = 2 N 3 Ω s H ∗ , 0 ( H ∗ R ∗ ) − 1 , (103)
f 2 ≃ 2.2 H ∗ , 0 ( H ∗ R ∗ ) − 1 , (104) f_2\simeq2.2H_{*,0}(H_*R_*)^{-1}, \tag*{(104)} f 2 ≃ 2.2 H ∗ , 0 ( H ∗ R ∗ ) − 1 , (104)
where N ≃ 2 \mathcal{N}\simeq2 N ≃ 2 [61 ] .
The amplitude at the second break is
h 2 Ω 2 = h 2 F G W , 0 A M H D Ω s 2 ( H ∗ R ∗ ) 2 , (105) h^2\Omega_2=h^2F_{\mathrm{GW},0}A_{\mathrm{MHD}}\Omega_s^2(H_*R_*)^2, \tag*{(105)} h 2 Ω 2 = h 2 F GW , 0 A MHD Ω s 2 ( H ∗ R ∗ ) 2 , (105)
where
A M H D ≃ 4.37 × 10 − 3 . (106) A_{\mathrm{MHD}}\simeq4.37\times10^{-3}. \tag*{(106)} A MHD ≃ 4.37 × 1 0 − 3 . (106)
Numerical procedure for the Abelian Higgs phase transition
In this Appendix we briefly summarize the numerical procedure used to compute the thermodynamic and dynamical quantities entering the first-order phase transition analysis. The calculation is based on the dimensionally reduced Abelian Higgs model and combines three ingredients: the finite-temperature effective theory, the thermal nucleation calculation, and the wall-velocity calculation with WallGo [45 ] .
Four-dimensional theory
The four-dimensional Abelian Higgs model is defined in Eqs. (43) –(46) of § 1.2.6 , in terms of a complex scalar Φ \Phi Φ of unit U ( 1 ) X U(1)_X U ( 1 ) X charge, a gauge field X μ X_\mu X μ , the gauge coupling g g g , and the scalar-potential parameters μ 2 \mu^2 μ 2 and λ \lambda λ . For the dimensional-reduction formulae in this appendix, we use the equivalent notation A μ ≡ − X μ A_\mu\equiv-X_\mu A μ ≡ − X μ and m 2 ≡ μ 2 m^2 \equiv\mu^2 m 2 ≡ μ 2 . Consequently, the covariant derivative is D μ = ∂ μ − i g A μ = ∂ μ + i g X μ D_\mu= \partial_\mu- igA_\mu= \partial_\mu+ igX_\mu D μ = ∂ μ − i g A μ = ∂ μ + i g X μ . The minus sign in the A μ A_\mu A μ convention is therefore just a field redefinition; the value and sign of the gauge coupling g g g are unchanged. We use A μ A_\mu A μ and m 2 m^2 m 2 consistently in the remainder of this appendix.
To calculate the effective potential, we introduce the real scalar background ϕ \phi ϕ by writing Φ b g = ϕ / 2 \Phi_{\rm bg} = \phi/\sqrt{2} Φ bg = ϕ / 2 ; the background phase can be removed by a U ( 1 ) X U(1)_X U ( 1 ) X transformation. Fluctuations about this background are written as Φ = ( ϕ + h + i G ) / 2 \Phi= (\phi+ h + iG)/\sqrt{2} Φ = ( ϕ + h + i G ) / 2 , where h h h is the real (Higgs) mode and G G G is the Goldstone mode. We retain G G G explicitly because the calculation is performed in Landau gauge rather than in unitary gauge. Thus, Φ \Phi Φ is the full complex scalar field and ϕ \phi ϕ its canonically normalised real background value.
In terms of ϕ \phi ϕ , the potential and squared vacuum expectation value at zero-temperature are
V 4 D ( ϕ ) = − 1 2 m 2 ϕ 2 + λ 4 ϕ 4 , v 2 = m 2 λ = μ 2 λ . (107) V_{4D}(\phi) = -\frac{1}{2}m^2\phi^2 + \frac{\lambda}{4}\phi^4,\qquad v^2 = \frac{m^2}{\lambda} = \frac{\mu^2}{\lambda}. \tag*{(107)} V 4 D ( ϕ ) = − 2 1 m 2 ϕ 2 + 4 λ ϕ 4 , v 2 = λ m 2 = λ μ 2 . (107)
The sign convention for the quadratic term differs from that of Ref. [62 ] ; here m 2 = μ 2 > 0 m^2 = \mu^2 > 0 m 2 = μ 2 > 0 is the symmetry-breaking mass parameter introduced in Eq. (46) . The field-dependent masses of the Higgs mode, Goldstone mode and gauge boson in Landau gauge are
m h 2 ( ϕ ) = − m 2 + 3 λ ϕ 2 , m G 2 ( ϕ ) = − m 2 + λ ϕ 2 , m A 2 ( ϕ ) = g 2 ϕ 2 . (108) m_h^2(\phi) = -m^2 + 3\lambda\phi^2,\qquad m_G^2(\phi) = -m^2 + \lambda\phi^2,\qquad m_A^2(\phi) = g^2\phi^2. \tag*{(108)} m h 2 ( ϕ ) = − m 2 + 3 λ ϕ 2 , m G 2 ( ϕ ) = − m 2 + λ ϕ 2 , m A 2 ( ϕ ) = g 2 ϕ 2 . (108)
At ϕ = v \phi= v ϕ = v , these reproduce m h 2 = 2 λ v 2 m_h^2 = 2\lambda v^2 m h 2 = 2 λ v 2 and m A 2 = g 2 v 2 m_A^2 = g^2v^2 m A 2 = g 2 v 2 , as in § 1.2.6 .
Dimensional reduction and two-loop thermal masses
At high temperature the non-zero Matsubara modes, with masses of order π T \pi T π T , are integrated out [63 , 64 ] . We follow the standard matching procedure [62 , 65 , 66 ] , assuming the power counting λ , g 2 ≪ 1 \lambda, g^2 \ll1 λ , g 2 ≪ 1 , m 2 ∼ g 2 T 2 m^2 \sim g^2T^2 m 2 ∼ g 2 T 2 . Scalar masses are matched at two loops and couplings at one loop. The resulting three-dimensional theory is
L 3 D = 1 4 F i j F i j + ( D i Φ ) † ( D i Φ ) − m 3 2 Φ † Φ + λ 3 ( Φ † Φ ) 2 + 1 2 ( ∂ i A 0 ) 2 + 1 2 m D 2 A 0 2 + 1 2 h 3 A 0 2 Φ † Φ + … (109) \mathcal{L}_{3D} = \frac{1}{4}F_{ij}F_{ij} + (D_i\Phi)^\dagger(D_i\Phi) - m_3^2\Phi^\dagger\Phi+ \lambda_3(\Phi^\dagger\Phi)^2 + \frac{1}{2}(\partial_iA_0)^2 + \frac{1}{2}m_D^2A_0^2 + \frac{1}{2}h_3 A_0^2\Phi^\dagger\Phi+ \ldots \tag*{(109)} L 3 D = 4 1 F ij F ij + ( D i Φ ) † ( D i Φ ) − m 3 2 Φ † Φ + λ 3 ( Φ † Φ ) 2 + 2 1 ( ∂ i A 0 ) 2 + 2 1 m D 2 A 0 2 + 2 1 h 3 A 0 2 Φ † Φ + … (109)
where D i = ∂ i − i g 3 A i D_i = \partial_i - \mathrm{i}g_3 A_i D i = ∂ i − i g 3 A i , A 0 A_0 A 0 is the temporal gauge component, which acquires the Debye mass m D m_D m D , and the ellipsis denotes the A 0 4 A_0^4 A 0 4 coupling and higher-dimensional operators, which are neglected. The sign of the mass term is chosen such that m 3 2 m_3^2 m 3 2 has the same sign convention as m 2 m^2 m 2 in Eq. (107) . At leading order h 3 = 2 g 2 T h_3 = 2g^2T h 3 = 2 g 2 T , so that 1 2 h 3 A 0 2 Φ † Φ = g 2 T A 0 2 Φ † Φ \frac{1}{2}h_3 A_0^2\Phi^\dagger\Phi= g^2T A_0^2\Phi^\dagger\Phi 2 1 h 3 A 0 2 Φ † Φ = g 2 T A 0 2 Φ † Φ .
We define
L b ≡ 2 ln μ ˉ e γ E 4 π T , (110) L_b \equiv2\ln\frac{\bar{\mu}e^{\gamma_E}}{4\pi T}, \tag*{(110)} L b ≡ 2 ln 4 π T μ ˉ e γ E , (110)
where μ ˉ \bar{\mu} μ ˉ is the M S ‾ \overline{\mathrm{MS}} MS renormalisation scale and m 2 m^2 m 2 , λ \lambda λ , g 2 g^2 g 2 are M S ‾ \overline{\mathrm{MS}} MS parameters at μ ˉ \bar{\mu} μ ˉ . In Landau gauge we find
g 3 2 = g 2 T [ 1 − g 2 48 π 2 L b ] , (111) g_3^2 = g^2T\left[1-\frac{g^2}{48\pi^2}L_b\right], \tag*{(111)} g 3 2 = g 2 T [ 1 − 48 π 2 g 2 L b ] , (111)
λ 3 = T { λ + 1 16 π 2 [ ( 6 g 2 λ − 10 λ 2 − 3 g 4 ) L b + 2 g 4 ] } , (112) \lambda_3 = T\left\{\lambda+\frac{1}{16\pi^2}\left[(6g^2\lambda-10\lambda^2-3g^4)L_b+2g^4\right]\right\}, \tag*{(112)} λ 3 = T { λ + 16 π 2 1 [ ( 6 g 2 λ − 10 λ 2 − 3 g 4 ) L b + 2 g 4 ] } , (112)
m D 2 = g 2 T 2 3 + g 2 T 2 144 π 2 [ ( 7 − L b ) g 2 + 12 λ ] − g 2 m 2 4 π 2 , (113) m_D^2 = \frac{g^2T^2}{3}+\frac{g^2T^2}{144\pi^2}\left[(7-L_b)g^2+12\lambda\right]-\frac{g^2m^2}{4\pi^2}, \tag*{(113)} m D 2 = 3 g 2 T 2 + 144 π 2 g 2 T 2 [ ( 7 − L b ) g 2 + 12 λ ] − 4 π 2 g 2 m 2 , (113)
h 3 = g 2 T [ 2 + λ π 2 − g 2 ( L b − 4 ) 24 π 2 ] . (114) h_3 = g^2T\left[2+\frac{\lambda}{\pi^2}-\frac{g^2(L_b-4)}{24\pi^2}\right]. \tag*{(114)} h 3 = g 2 T [ 2 + π 2 λ − 24 π 2 g 2 ( L b − 4 ) ] . (114)
The scalar mass parameter is determined at two loops,
m 3 2 = m 2 [ 1 + ( 3 g 2 16 π 2 − λ 4 π 2 ) L b ] − ( g 2 4 + λ 3 ) T 2 − T 2 576 π 2 { 24 λ 2 ( L b − 6 γ E + 72 ln A ) − 24 g 2 λ ( 3 L b − 1 − 6 γ E + 72 ln A ) + g 4 ( 69 L b − 8 − 108 γ E + 1296 ln A ) } − 1 16 π 2 [ 4 g 3 4 + 1 2 h 3 2 − 8 g 3 2 λ 3 + 8 λ 3 2 ] ln μ ˉ 3 μ ˉ , (115) \begin{aligned}
m_3^2 ={}& m^2\left[1+\left(\frac{3g^2}{16\pi^2}-\frac{\lambda}{4\pi^2}\right)L_b\right]-\left(\frac{g^2}{4}+\frac{\lambda}{3}\right)T^2\\
&-\frac{T^2}{576\pi^2}\left\{24\lambda^2(L_b-6\gamma_E+72\ln\mathcal{A})-24g^2\lambda(3L_b-1-6\gamma_E+72\ln\mathcal{A})\right.\\
&\left.\qquad\qquad\qquad\quad+g^4(69L_b-8-108\gamma_E+1296\ln\mathcal{A})\right\}\\
&-\frac{1}{16\pi^2}\left[4g_3^4+\frac{1}{2}h_3^2-8g_3^2\lambda_3+8\lambda_3^2\right]\ln\frac{\bar{\mu}_3}{\bar{\mu}}, \tag*{(115)}
\end{aligned} m 3 2 = m 2 [ 1 + ( 16 π 2 3 g 2 − 4 π 2 λ ) L b ] − ( 4 g 2 + 3 λ ) T 2 − 576 π 2 T 2 { 24 λ 2 ( L b − 6 γ E + 72 ln A ) − 24 g 2 λ ( 3 L b − 1 − 6 γ E + 72 ln A ) + g 4 ( 69 L b − 8 − 108 γ E + 1296 ln A ) } − 16 π 2 1 [ 4 g 3 4 + 2 1 h 3 2 − 8 g 3 2 λ 3 + 8 λ 3 2 ] ln μ ˉ μ ˉ 3 , (115)
where A ≃ 1.2824 \mathcal{A}\simeq1.2824 A ≃ 1.2824 is the Glaisher–Kinkelin constant (ln A ≃ 0.24875 \ln\mathcal{A}\simeq0.24875 ln A ≃ 0.24875 ) and μ ˉ 3 \bar{\mu}_3 μ ˉ 3 is the renormalisation scale of the three-dimensional theory. The last line of Eq. (115) is the two-loop running of m 3 2 m_3^2 m 3 2 inside the 3D theory. It is the only parameter that runs in 3D, with
μ ˉ 3 d m 3 2 d μ ˉ 3 = − 1 16 π 2 [ 4 g 3 4 + 1 2 h 3 2 − 8 g 3 2 λ 3 + 8 λ 3 2 ] ≃ − T 2 16 π 2 ( 6 g 4 − 8 g 2 λ + 8 λ 2 ) , (116) \bar{\mu}_3\frac{\mathrm{d}m_3^2}{\mathrm{d}\bar{\mu}_3}
=-\frac{1}{16\pi^2}\left[4g_3^4+\frac{1}{2}h_3^2-8g_3^2\lambda_3+8\lambda_3^2\right]
\simeq-\frac{T^2}{16\pi^2}\left(6g^4-8g^2\lambda+8\lambda^2\right), \tag*{(116)} μ ˉ 3 d μ ˉ 3 d m 3 2 = − 16 π 2 1 [ 4 g 3 4 + 2 1 h 3 2 − 8 g 3 2 λ 3 + 8 λ 3 2 ] ≃ − 16 π 2 T 2 ( 6 g 4 − 8 g 2 λ + 8 λ 2 ) , (116)
where the last step holds to the order considered. The expressions (E5)–(E9) are independent of μ ˉ \bar{\mu} μ ˉ at the order considered, as follows from the one-loop β \beta β -functions
μ d g 2 d μ = g 4 24 π 2 , μ d λ d μ = 20 λ 2 − 12 g 2 λ + 6 g 4 16 π 2 , μ d m 2 d μ = ( 8 λ − 6 g 2 ) m 2 16 π 2 . (117) \mu\frac{\mathrm{d}g^2}{\mathrm{d}\mu}=\frac{g^4}{24\pi^2},\qquad
\mu\frac{\mathrm{d}\lambda}{\mathrm{d}\mu}=\frac{20\lambda^2-12g^2\lambda+6g^4}{16\pi^2},\qquad
\mu\frac{\mathrm{d}m^2}{\mathrm{d}\mu}=\frac{(8\lambda-6g^2)m^2}{16\pi^2}. \tag*{(117)} μ d μ d g 2 = 24 π 2 g 4 , μ d μ d λ = 16 π 2 20 λ 2 − 12 g 2 λ + 6 g 4 , μ d μ d m 2 = 16 π 2 ( 8 λ − 6 g 2 ) m 2 . (117)
Choosing μ ˉ = 4 π e − γ E T \bar{\mu}=4\pi e^{-\gamma_E}T μ ˉ = 4 π e − γ E T removes L b L_b L b . At leading order the matching reduces to g 3 2 = g 2 T g_3^2=g^2T g 3 2 = g 2 T , λ 3 = λ T \lambda_3=\lambda T λ 3 = λ T , m D 2 = g 2 T 2 / 3 m_D^2=g^2T^2/3 m D 2 = g 2 T 2 /3 and m 3 2 = m 2 − ( λ / 3 + g 2 / 4 ) T 2 m_3^2=m^2-(\lambda/3+g^2/4)T^2 m 3 2 = m 2 − ( λ /3 + g 2 /4 ) T 2 .
The field A 0 A_0 A 0 has mass m D ∼ g T m_D\sim gT m D ∼ g T and is integrated out in the next step of the hierarchy. Expanding its one-loop contribution − 1 12 π ( m D 2 + h 3 Φ † Φ ) 3 / 2 -\frac{1}{12\pi}(m_D^2+h_3\Phi^\dagger\Phi)^{3/2} − 12 π 1 ( m D 2 + h 3 Φ † Φ ) 3/2 to quadratic order in Φ † Φ \Phi^\dagger\Phi Φ † Φ shifts the parameters of the remaining theory by
m ˉ 3 2 = m 3 2 + h 3 m D 8 π , λ ˉ 3 = λ 3 − h 3 2 32 π m D , (118) \bar{m}_3^2=m_3^2+\frac{h_3m_D}{8\pi},\qquad
\bar{\lambda}_3=\lambda_3-\frac{h_3^2}{32\pi m_D}, \tag*{(118)} m ˉ 3 2 = m 3 2 + 8 π h 3 m D , λ ˉ 3 = λ 3 − 32 π m D h 3 2 , (118)
which reduce to m ˉ 3 2 = m 3 2 + g 3 2 m D / ( 4 π ) \bar{m}_3^2=m_3^2+g_3^2m_D/(4\pi) m ˉ 3 2 = m 3 2 + g 3 2 m D / ( 4 π ) and λ ˉ 3 = λ 3 − g 3 4 / ( 8 π m D ) \bar{\lambda}_3=\lambda_3-g_3^4/(8\pi m_D) λ ˉ 3 = λ 3 − g 3 4 / ( 8 π m D ) for h 3 = 2 g 3 2 h_3=2g_3^2 h 3 = 2 g 3 2 . The three-dimensional fields are related to the four-dimensional ones by ϕ 3 2 = ϕ 2 ( 1 + δ Z ) / T \phi_3^2=\phi^2(1+\delta Z)/T ϕ 3 2 = ϕ 2 ( 1 + δ Z ) / T with δ Z = − 3 g 2 16 π 2 L b \delta Z=-\frac{3g^2}{16\pi^2}L_b δ Z = − 16 π 2 3 g 2 L b .
Effective potential
The phase-transition dynamics is controlled by the effective potential, which we write in four-dimensional normalisation as V e f f ( ϕ , T ) = T V 3 D ( ϕ 3 ) V_{\mathrm{eff}}(\phi,T)=T V_{3\mathrm{D}}(\phi_3) V eff ( ϕ , T ) = T V 3 D ( ϕ 3 ) with ϕ = T ϕ 3 \phi=\sqrt{T}\phi_3 ϕ = T ϕ 3 . At leading order in the three-dimensional theory,
V 3 D ( ϕ 3 ) = − 1 2 m 3 2 ϕ 3 2 + 1 4 λ 3 ϕ 3 4 − ( g 3 2 ϕ 3 2 ) 3 / 2 6 π − 1 12 π ( m D 2 + 1 2 h 3 ϕ 3 2 ) 3 / 2 , (119) V_{3\mathrm{D}}(\phi_3)=-\frac{1}{2}m_3^2\phi_3^2+\frac{1}{4}\lambda_3\phi_3^4-\frac{(g_3^2\phi_3^2)^{3/2}}{6\pi}-\frac{1}{12\pi}\left(m_D^2+\frac{1}{2}h_3\phi_3^2\right)^{3/2}, \tag*{(119)} V 3 D ( ϕ 3 ) = − 2 1 m 3 2 ϕ 3 2 + 4 1 λ 3 ϕ 3 4 − 6 π ( g 3 2 ϕ 3 2 ) 3/2 − 12 π 1 ( m D 2 + 2 1 h 3 ϕ 3 2 ) 3/2 , (119)
that is,
V e f f ( ϕ , T ) = − 1 2 m 3 2 ϕ 2 + λ 3 4 T ϕ 4 − g 3 3 6 π T ϕ 3 − T 12 π ( m D 2 + h 3 ϕ 2 2 T ) 3 / 2 , (120) V_{\mathrm{eff}}(\phi,T)=-\frac{1}{2}m_3^2\phi^2+\frac{\lambda_3}{4T}\phi^4-\frac{g_3^3}{6\pi\sqrt{T}}\phi^3-\frac{T}{12\pi}\left(m_D^2+\frac{h_3\phi^2}{2T}\right)^{3/2}, \tag*{(120)} V eff ( ϕ , T ) = − 2 1 m 3 2 ϕ 2 + 4 T λ 3 ϕ 4 − 6 π T g 3 3 ϕ 3 − 12 π T ( m D 2 + 2 T h 3 ϕ 2 ) 3/2 , (120)
where m 3 2 m_3^2 m 3 2 , λ 3 \lambda_3 λ 3 , g 3 2 g_3^2 g 3 2 , m D 2 m_D^2 m D 2 , and h 3 h_3 h 3 are given in Eqs. (111) –(115) . The cubic term arises from the two transverse polarisations of the three-dimensional gauge field, each contributing − ( g 3 ϕ 3 ) 3 / ( 12 π ) -(g_3\phi_3)^3/(12\pi) − ( g 3 ϕ 3 ) 3 / ( 12 π ) , and the last term from the temporal component A 0 A_0 A 0 , whose mass m D 2 + 1 2 h 3 ϕ 3 2 m_D^2+\frac{1}{2}h_3\phi_3^2 m D 2 + 2 1 h 3 ϕ 3 2 depends on the background field. At leading order, V e f f ⊃ − g 3 3 T 6 π ϕ 3 − g 3 3 T 12 π ( ϕ 2 + T 2 / 3 ) 3 / 2 V_{\mathrm{eff}}\supset-\frac{g_3^3T}{6\pi}\phi^3-\frac{g_3^3T}{12\pi}(\phi^2+T^2/3)^{3/2} V eff ⊃ − 6 π g 3 3 T ϕ 3 − 12 π g 3 3 T ( ϕ 2 + T 2 /3 ) 3/2 . Scalar loops are not included, which is justified for λ ≪ g 2 \lambda\ll g^2 λ ≪ g 2 .
For ϕ 2 ≪ T 2 / 3 \phi^2\ll T^2/3 ϕ 2 ≪ T 2 /3 the A 0 A_0 A 0 term may be expanded and absorbed into shifted parameters, Eq. (118) , which yields the polynomial form V e f f = A ϕ 2 + B ϕ 3 + C ϕ 4 V_{\mathrm{eff}}=A\phi^2+B\phi^3+C\phi^4 V eff = A ϕ 2 + B ϕ 3 + C ϕ 4 with A = − 1 2 m 3 2 A=-\frac{1}{2}m_3^2 A = − 2 1 m 3 2 , B = − g 3 3 / ( 6 π T ) B=-g_3^3/(6\pi\sqrt{T}) B = − g 3 3 / ( 6 π T ) and C = λ 3 / ( 4 T ) C=\lambda_3/(4T) C = λ 3 / ( 4 T ) . This expansion fails in the supersoft regime (λ ≲ 0.07 g 3 \lambda\lesssim0.07\,g^3 λ ≲ 0.07 g 3 ) and for ϕ ≳ T / 3 \phi\gtrsim T/\sqrt{3} ϕ ≳ T / 3 , so we do not use it for the results in this paper; the critical temperature is instead obtained by solving V e f f ( ϕ t r u e , T c ) = V e f f ( 0 , T c ) V_{\mathrm{eff}}(\phi_{\mathrm{true}},T_c)=V_{\mathrm{eff}}(0,T_c) V eff ( ϕ true , T c ) = V eff ( 0 , T c ) numerically from Eq. (120) .
Critical and spinodal temperatures
We normalise the potential of Eq. (120) to vanish in the symmetric phase,
V ~ ( ϕ , T ) ≡ V e f f ( ϕ , T ) − V e f f ( 0 , T ) , (121) \widetilde{V}(\phi,T)\equiv V_{\mathrm{eff}}(\phi,T)-V_{\mathrm{eff}}(0,T), \tag*{(121)} V ( ϕ , T ) ≡ V eff ( ϕ , T ) − V eff ( 0 , T ) , (121)
so that ϕ f a l s e = 0 \phi_{\mathrm{false}}=0 ϕ false = 0 at all temperatures. Its derivative factorises, ∂ ϕ V ~ = ϕ F ( ϕ , T ) \partial_\phi\widetilde{V}=\phi F(\phi,T) ∂ ϕ V = ϕF ( ϕ , T ) , with
F ( ϕ , T ) = − m 3 2 + λ 3 T ϕ 2 − g 3 3 2 π T ϕ − h 3 8 π m D 2 + h 3 ϕ 2 2 T , (122) F(\phi,T)=-m_3^2+\frac{\lambda_3}{T}\phi^2-\frac{g_3^3}{2\pi\sqrt{T}}\phi-\frac{h_3}{8\pi}\sqrt{m_D^2+\frac{h_3\phi^2}{2T}}, \tag*{(122)} F ( ϕ , T ) = − m 3 2 + T λ 3 ϕ 2 − 2 π T g 3 3 ϕ − 8 π h 3 m D 2 + 2 T h 3 ϕ 2 , (122)
so that the non-trivial extrema are the positive roots of F ( ϕ , T ) = 0 F(\phi,T)=0 F ( ϕ , T ) = 0 . When two positive roots exist, the larger one is the broken minimum ϕ t r u e \phi_{\mathrm{true}} ϕ true and the smaller one is the top of the barrier, ϕ b a r \phi_{\mathrm{bar}} ϕ bar .
Critical temperature. The critical temperature T c T_c T c is defined by degenerate minima, i.e. by the simultaneous solution for ( ϕ t r u e , T c ) (\phi_{\mathrm{true}},T_c) ( ϕ true , T c ) of
V ~ ( ϕ t r u e , T c ) = 0 , F ( ϕ t r u e , T c ) = 0 , ϕ t r u e > 0. (123) \widetilde{V}(\phi_{\mathrm{true}},T_c)=0,\qquad F(\phi_{\mathrm{true}},T_c)=0,\qquad\phi_{\mathrm{true}}>0. \tag*{(123)} V ( ϕ true , T c ) = 0 , F ( ϕ true , T c ) = 0 , ϕ true > 0. (123)
Spinodal temperatures. The symmetric phase ceases to be metastable when ∂ ϕ 2 V ~ ( 0 , T ) = − F ( 0 , T ) = 0 \partial_\phi^2\widetilde{V}(0,T)=-F(0,T)=0 ∂ ϕ 2 V ( 0 , T ) = − F ( 0 , T ) = 0 , which is equivalent to
m ‾ 3 2 ( T s p l o w ) = 0 , m ‾ 3 2 ≡ m 3 2 + h 3 m D 8 π , (124) \overline{m}_3^2(T_{\mathrm{sp}}^{\mathrm{low}})=0,\qquad\overline{m}_3^2\equiv m_3^2+\frac{h_3m_D}{8\pi}, \tag*{(124)} m 3 2 ( T sp low ) = 0 , m 3 2 ≡ m 3 2 + 8 π h 3 m D , (124)
with m ‾ 3 2 \overline{m}_3^2 m 3 2 as in Eq. (118) . The upper spinodal, at which the broken minimum first appears, is where ϕ t r u e \phi_{\mathrm{true}} ϕ true and ϕ b a r \phi_{\mathrm{bar}} ϕ bar merge, i.e. ∂ ϕ V ~ = ∂ ϕ 2 V ~ = 0 \partial_\phi\widetilde{V}=\partial_\phi^2\widetilde{V}=0 ∂ ϕ V = ∂ ϕ 2 V = 0 at some ϕ s p > 0 \phi_{\mathrm{sp}}>0 ϕ sp > 0 , which is equivalent to
F ( ϕ s p , T s p u p ) = 0 , ∂ ϕ F ( ϕ s p , T s p u p ) = 0 , (125) F(\phi_{\mathrm{sp}},T_{\mathrm{sp}}^{\mathrm{up}})=0,\qquad\partial_\phi F(\phi_{\mathrm{sp}},T_{\mathrm{sp}}^{\mathrm{up}})=0, \tag*{(125)} F ( ϕ sp , T sp up ) = 0 , ∂ ϕ F ( ϕ sp , T sp up ) = 0 , (125)
with
∂ ϕ F = 2 λ 3 T ϕ − g 3 3 2 π T − h 3 2 ϕ 16 π T m D 2 + h 3 ϕ 2 / ( 2 T ) (126) \partial_{\phi} F = \frac{2\lambda_{3}}{T}\phi- \frac{g_{3}^{3}}{2\pi\sqrt{T}} - \frac{h_{3}^{2}\phi}{16\pi T\sqrt{m_{D}^{2}+h_{3}\phi^{2}/(2T)}} \tag*{(126)} ∂ ϕ F = T 2 λ 3 ϕ − 2 π T g 3 3 − 16 π T m D 2 + h 3 ϕ 2 / ( 2 T ) h 3 2 ϕ (126)
These temperatures satisfy T s p l o w < T c < T s p u p T_{\mathrm{sp}}^{\mathrm{low}} < T_{c} < T_{\mathrm{sp}}^{\mathrm{up}} T sp low < T c < T sp up . Expanding the square root in Eq. (122) to first order in h 3 ϕ 2 / ( 2 T m D 2 ) h_{3}\phi^{2}/(2Tm_{D}^{2}) h 3 ϕ 2 / ( 2 T m D 2 ) reproduces the polynomial potential A ϕ 2 + B ϕ 3 + C ϕ 4 A\phi^{2}+B\phi^{3}+C\phi^{4} A ϕ 2 + B ϕ 3 + C ϕ 4 , for which the conditions reduce to the closed forms A = B 2 / ( 4 C ) A=B^{2}/(4C) A = B 2 / ( 4 C ) , ϕ t r u e = − B / ( 2 C ) \phi_{\mathrm{true}}=-B/(2C) ϕ true = − B / ( 2 C ) and 9 B 2 = 32 A C 9B^{2}=32AC 9 B 2 = 32 A C . These are reliable only for ϕ 2 ≪ T 2 / 3 \phi^{2}\ll T^{2}/3 ϕ 2 ≪ T 2 /3 and are not used here.
The quantities recorded at the transition are
ϕ f a l s e = 0 , ϕ t r u e , ϕ b a r , Δ V = V ~ ( ϕ t r u e , T ) , (127) \phi_{\mathrm{false}}=0,\qquad\phi_{\mathrm{true}},\qquad\phi_{\mathrm{bar}},\qquad\Delta V=\widetilde{V}(\phi_{\mathrm{true}},T), \tag*{(127)} ϕ false = 0 , ϕ true , ϕ bar , Δ V = V ( ϕ true , T ) , (127)
together with m 3 2 m_{3}^{2} m 3 2 , λ 3 \lambda_{3} λ 3 , g 3 2 g_{3}^{2} g 3 2 , m D 2 m_{D}^{2} m D 2 , h 3 h_{3} h 3 in place of A ( T ) A(T) A ( T ) , B ( T ) B(T) B ( T ) , C ( T ) C(T) C ( T ) .
Scan strategy
For efficiency, the parameter scan is organised in stages. A fast analytic filter first checks that the effective potential permits the required first-order transition structure. The nucleation/percolation calculation is then carried out for points passing this filter, and the successful points are mapped to the GW spectrum using the LTE wall-velocity treatment. This is the calculation used in the Bayesian scans shown in the main text.
The debugging-stage output table contains
g , λ , y , n f , q f , m 2 , v , T , ϕ f a l s e , ϕ t r u e , Δ V , A ( T ) , B ( T ) , C ( T ) , T c , T s p i n o d a l , v w , L w , v w L T E , (128) \begin{aligned}
&g,\ \lambda,\ y,\ n_f,\ q_f,\ m^{2},\ v,\ T,\ \phi_{\mathrm{false}},\ \phi_{\mathrm{true}},\ \Delta V,\\
&A(T),\ B(T),\ C(T),\ T_c,\ T_{\mathrm{spinodal}},\ v_w,\ L_w,\ v_w^{\mathrm{LTE}}, \tag*{(128)}
\end{aligned} g , λ , y , n f , q f , m 2 , v , T , ϕ false , ϕ true , Δ V , A ( T ) , B ( T ) , C ( T ) , T c , T spinodal , v w , L w , v w LTE , (128)
together with the collision-cache directory and solver status. This provides a compact record of the thermodynamic, nucleation, and wall-dynamics information before connecting the calculation to the full FOPT and GW analysis pipeline.
To examine the validity of the LTE wall treatment, a small separate set of successful transition points is passed to the full collision-based WallGo solver. For these validation points we additionally retain v w f u l l v_w^{\mathrm{full}} v w full when a stationary solution exists, v w L T E v_w^{\mathrm{LTE}} v w LTE , the native solver classification, the wall width when returned, the collision-cache path, and the momentum grid size. The full collision-based solver is not used at each likelihood evaluation of the production scan.
Wall velocity calculation
The wall velocity in the production likelihood is obtained within the local-thermal-equilibrium (LTE) treatment of WallGo, using the phase thermodynamics and hydrodynamic matching from the NLO dimensionally reduced effective potential. This avoids computing the off-equilibrium collision integrals at every sampled point. An LTE limiting result v w L T E = 1 v_w^{\mathrm{LTE}}=1 v w LTE = 1 should be understood as the outcome of this approximation, not as an independent full collision-based demonstration of runaway propagation.
For the separate validation points, we solve the full wall problem with off-equilibrium perturbations. Schematically, the scalar equation of motion in the wall frame may be written as
d 2 ϕ d z 2 − ∂ V e f f ∂ ϕ − F f r i c [ ϕ , v w ] = 0 , (129) \frac{d^{2}\phi}{dz^{2}}-\frac{\partial V_{\mathrm{eff}}}{\partial\phi}-\mathcal{F}_{\mathrm{fric}}[\phi,v_w]=0, \tag*{(129)} d z 2 d 2 ϕ − ∂ ϕ ∂ V eff − F fric [ ϕ , v w ] = 0 , (129)
where F f r i c \mathcal{F}_{\mathrm{fric}} F fric represents plasma friction arising from deviations from LTE. In WallGo these perturbations are obtained by solving linearised Boltzmann equations for the species included in the collision model. The matrix elements are generated using WallGoMatrix, with collision integrals computed by WallGoCollision and cached for reuse. The collision calculation uses leading-order thermal-mass inputs; these should be distinguished from the NLO effective potential used for the equilibrium phase thermodynamics.
Each validation point is evaluated at its archived transition temperature T ∗ T_\ast T ∗ , without temperature-offset retries. We check the native solver status rather than treating any numerical return as a stationary solution. The stationary branches are separated from cases classified RUNAWAY, for which the code does not return a stationary wall velocity. Grid-size tests are reported where available, including N = 5 N=5 N = 5 and 7 7 7 for points C and E and N = 5 N=5 N = 5 , 7 7 7 and 9 9 9 for a separate stationary reference point. The candidate at E is located at the Jouguet branch boundary and is therefore not assigned a precision-converged terminal velocity. These comparisons are used only as a limited validation of the LTE treatment and do not replace the LTE calculation in the nested-sampling likelihood.
Numerical convergence of the Abelian nested-sampling analysis
We assess the numerical stability of the U ( 1 ) X U(1)_X U ( 1 ) X posterior using the evidence-consistency statistic R Z R_Z R Z defined in Eq. (27) . For the O4a dataset, comparisons are performed separately for the linear and logarithmic priors, holding fixed the likelihood, physical parameters and all prior bounds while varying the nested-sampling resolution. For each pair of independently initialised runs, we calculate ∣ Δ Log Z ∣ |\Delta\operatorname{Log} Z| ∣Δ Log Z ∣ , the combined uncertainty σ c o m b = σ a 2 + σ b 2 \sigma_{\mathrm{comb}} = \sqrt{\sigma_a^2+\sigma_b^2} σ comb = σ a 2 + σ b 2 , and the ratio R Z = ∣ Δ Log Z ∣ / σ c o m b R_Z = |\Delta\operatorname{Log} Z|/\sigma_{\mathrm{comb}} R Z = ∣Δ Log Z ∣/ σ comb .
A value R Z ≤ 1 R_Z \le1 R Z ≤ 1 indicates consistency of the two evidence estimates within their quoted combined numerical uncertainty. We interpret this result together with the precision of the individual evidence estimates and the stability of the one-dimensional posterior distributions and credible regions. Thus, numerical convergence within a given prior is assessed independently of the changes induced by replacing the linear prior with a logarithmic one; the latter are the subject of the prior-robustness analysis in the following subsections.
The nominal evidence-consistency statistic R Z = ∣ Log Z 20000 − Log Z 200 ∣ / σ 20000 2 + σ 200 2 R_Z = |\operatorname{Log} Z_{20000} - \operatorname{Log} Z_{200}|/\sqrt{\sigma_{20000}^2+\sigma_{200}^2} R Z = ∣ Log Z 20000 − Log Z 200 ∣/ σ 20000 2 + σ 200 2 is 0.283 0.283 0.283 for the linear prior and 0.285 0.285 0.285 for the logarithmic prior, both below the adopted threshold R Z = 1 R_Z=1 R Z = 1 . This indicates stability of the estimated evidence when the sampling resolution increases, for each prior separately. The saved samples reveal that the 200-live-point runs share their initial pseudo-random samples with the corresponding 20 000-live-point runs; thus this is a resolution-stability diagnostic rather than an independent-realisation test. Agreement of the evidences alone does not establish convergence of all marginal posterior distributions.
Table XXIV. Resolution-stability check of the log Bayesian evidence for the Abelian U ( 1 ) X U(1)_X U ( 1 ) X NLO-3D LTE analysis with the combined O3+O4a dataset. Within each row the model, likelihood and prior bounds are identical; the initial number of live points changes from 200 to 20 000. The uncertainties are those reported by dynesty.
References