Spectral-Siren Cosmology

• Spectral-siren cosmology is a method that deduces cosmological parameters by mapping observed detector-frame masses into a coherent source-frame population via statistical redshift calibration.
• It employs hierarchical Bayesian inference and flexible, non-parametric mass models to jointly constrain the mass distribution and cosmology from gravitational-wave catalogs.
• This approach enables precision measurement of parameters like the Hubble constant and matter density without relying on electromagnetic counterparts or host-galaxy identification.

Spectral-siren cosmology is a gravitational-wave population method in which cosmological parameters are inferred from the joint distribution of luminosity distances and redshifted compact-object masses, without requiring electromagnetic counterparts or host-galaxy identification. Its defining relation is the redshifting of source-frame masses into the detector frame,

$m_{\rm det}=(1+z)m_{\rm src},$

together with the cosmological distance–redshift map $d_L=d_L(z;H_0,\Omega_m,\ldots)$. In this construction, features in the source-frame mass spectrum act as statistical redshift calibrators: the cosmology preferred by the data is the one under which the observed detector-frame catalog maps into a coherent source-frame population (Ezquiaga et al., 2022, Pierra, 2024, Farah et al., 2024).

1. Conceptual basis

Spectral sirens extend the standard-siren idea to the case in which no event-by-event redshift is available. In a bright siren, the redshift comes from an electromagnetic counterpart and host-galaxy identification. In a dark siren, it is supplied statistically by a galaxy catalog. In a spectral siren, it is supplied by the compact-binary population itself: detector-frame masses and luminosity distances are measured from the waveform, and the source-frame mass spectrum provides the rest-frame reference against which redshift is inferred (Farah et al., 2024).

The original formulation emphasized that the entire compact-object mass spectrum can function as a cosmological probe, not merely a single feature. In that picture, current gravitational-wave data already motivate at least five independent mass features: the upper and lower edges of the pair-instability supernova gap, the upper and lower edges of the neutron-star–black-hole lower mass gap, and the minimum neutron-star mass (Ezquiaga et al., 2022). The same work argued that the dominant spectral anchor depends on detector generation: for second-generation detectors the lower edge of the pair-instability gap dominates, whereas in the third-generation era the lower mass gap becomes more powerful because low-mass binaries become much more abundant in the detected sample (Ezquiaga et al., 2022).

The method is intrinsically population-level. A single event does not break the mass–redshift degeneracy. A catalog does so only if the source-frame distribution contains reproducible structure shared across events. This is why the method is especially developed for binary black holes in current catalogs: BBHs dominate the event rate, and several analyses now find that their mass spectrum is structured rather than featureless (Hernandez et al., 2024, Hernandez et al., 3 Sep 2025).

2. Hierarchical inference and cosmological mapping

Spectral-siren analyses are formulated as hierarchical Bayesian population inference with selection effects. A common structure is an inhomogeneous Poisson-process likelihood for a catalog of detections,

$$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$

with the corresponding expected number of detections

$$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$

Here $\theta$ denotes intrinsic binary parameters, $\Lambda$ collects population and cosmological hyperparameters, and the factor $(1+z)^{-1}$ converts source-frame time to detector-frame time (Pierra, 2024).

The source-population rate is typically factorized as

$$\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s} = R_0\,\psi(z;\Lambda)\,p_{\rm pop}(\vec m_{\rm s}|\Lambda)\,\frac{dV_c}{dz},$$

or in closely related forms that allow additional population dimensions such as mass ratio or spins (Pierra, 2024). Cosmology enters twice: through the event-level conversion between luminosity distance and redshift, and through the comoving-volume element $dV_c/dz$ (Pierra, 2024).

For flat $\Lambda$CDM, the luminosity distance is

$d_L=d_L(z;H_0,\Omega_m,\ldots)$0

and several spectral-siren studies either infer $d_L=d_L(z;H_0,\Omega_m,\ldots)$1 alone with $d_L=d_L(z;H_0,\Omega_m,\ldots)$2 fixed or fit $d_L=d_L(z;H_0,\Omega_m,\ldots)$3 jointly (Pierra, 2024, Hernandez et al., 2024, Hernandez et al., 3 Sep 2025). In current BBH analyses, the operational observable pair is therefore $d_L=d_L(z;H_0,\Omega_m,\ldots)$4, mapped through a trial cosmology into $d_L=d_L(z;H_0,\Omega_m,\ldots)$5, and scored against a common source-frame population model (Farah et al., 2024).

This formalism also makes clear why spectral and catalog-based dark sirens are not fundamentally separate likelihood classes. A unified treatment based on galaxy number density writes the merger rate per galaxy and the source population within the same hierarchical model, so that galaxy catalogs and source-frame mass spectra become two components of a single redshift-inference problem rather than competing frameworks (Mastrogiovanni et al., 2023).

3. Population models and spectral anchors

The central modeling object in spectral-siren cosmology is the source-frame mass distribution. Early and still widely used phenomenological models include the broken power law (BPL), the power law plus peak (PLP), and multipeak extensions in which Gaussian components are added to a power-law continuum to represent pile-ups, cutoffs, or hierarchical-merger populations (Pierra, 2024). These models are attractive because their sharp features can act as strong statistical anchors, and correctly specified injections confirm that sharper spectral structure yields tighter $d_L=d_L(z;H_0,\Omega_m,\ldots)$6 posteriors than smoother models (Pierra, 2024).

A major development has been the move toward flexible non-parametric or semi-parametric descriptions. One line of work places a Gaussian-process prior directly on the primary-mass distribution, with a Matérn kernel and penalized-complexity priors on the kernel length scale and variance, so that the mass spectrum is inferred jointly with cosmology without assuming a specific parametric morphology in advance (Farah et al., 2024). A related real-data analysis uses a binned Gaussian process with an exponentiated-quadratic kernel over source-frame mass and redshift bins, allowing the BBH source-frame mass spectrum and redshift evolution to be learned simultaneously from GWTC-3 (Hernandez et al., 2024). These approaches are motivated by the same principle: spectral sirens need a population model flexible enough to let the data reveal the relevant source-frame features.

Other extensions target structure that is not well captured by one-dimensional primary-mass fits. Multi-spectral sirens decompose the BBH population into subpopulations, using spins and semiparametric PowerLawSpline mass functions so that features blurred in the total population become sharper within individual components (Li et al., 2024). More recent analyses argue that the secondary-mass distribution and pairing function are themselves cosmological ingredients: a flexible paired-mass model for GWTC-4.0 identifies peaks near $d_L=d_L(z;H_0,\Omega_m,\ldots)$7 and $d_L=d_L(z;H_0,\Omega_m,\ldots)$8 together with mass-dependent pairing transitions near $d_L=d_L(z;H_0,\Omega_m,\ldots)$9 and $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$0–$$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$1, and reports improved $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$2 constraints relative to the LVK Fullpop-4.0 analysis (Li et al., 12 May 2026). Closely related work argues that heavy black holes also matter because an additional mass scale at $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$3 sharpens both spectral- and dark-siren constraints (Pierra et al., 6 Jan 2026).

A different calibration strategy uses Normalizing Flows trained on synthetic astrophysical catalogs. In that construction, the learned object is a joint source-frame density $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$4, including redshift evolution and formation-channel structure. Applied to GWTC-4.0 with a B-POP-calibrated isolated-plus-dynamical mixture, this approach yields $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$5 at 68.3% credibility, while explicitly tying the cosmological inference to a calibrated astrophysical prior (Scarpa et al., 16 Mar 2026). The common methodological message is that the source-population model is not an auxiliary nuisance term; it is the redshift calibrator itself.

4. Model misspecification, redshift evolution, and methodological controversy

The most important methodological warning in the literature is that spectral-siren precision can become misleading when the source-population model is wrong. A controlled simulation study showed that standard phenomenological BBH mass models—BPL, PLP, and multipeak variants—can produce significantly biased $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$6 estimates if the true source-frame mass spectrum contains unmodeled structure or evolves with redshift, with the bias reaching roughly $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$7 for $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$8 detections (Pierra, 2024). In one synthetic astrophysical population, misspecified analyses returned $$\mathcal{L}(\{x\}|\Lambda) \propto e^{-N_{\rm exp}(\Lambda)} \prod_i T_{\rm obs}\int d\theta\,dz\; \mathcal{L}_{\rm GW}(x_i|\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda),$$9 or $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$0 for an injected $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$1, because the assumed models underrepresented systems between roughly $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$2 and $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$3 and therefore pushed those events to higher inferred redshift (Pierra, 2024).

That same study isolated two failure modes. The first is unmodeled spectral structure: if the true source-frame distribution contains localized overdensities or broad support not representable by the chosen parameterization, the cosmological fit absorbs the mismatch. The second is unmodeled redshift evolution of mass-spectrum features. A simple demonstration lets the Gaussian peak in a PLP model drift linearly with redshift,

$$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$4

and then analyzes the catalog with a redshift-independent model. The resulting $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$5 bias is approximately linear in the amount of peak evolution (Pierra, 2024). A control experiment in which masses and redshifts are randomly shuffled removes most of the bias, showing that the main culprit is the mass–redshift correlation rather than static mismatch in the one-dimensional mass histogram (Pierra, 2024).

This controversy also clarifies the meaning of “astrophysics-free” spectral sirens. Flexible non-parametric analyses explicitly argue that one need not know the correct parametric form or physical origin of every mass-spectrum feature in advance, but they do not claim that the inference is assumption-free. The method still assumes a common underlying source population, smoothness conditions implied by the GP prior, a parametric cosmology, and sufficiently mild redshift evolution that population drift does not exactly mimic cosmological redshifting (Farah et al., 2024). The phrase therefore means freedom from detailed parametric astrophysical prior knowledge, not freedom from population assumptions altogether (Farah et al., 2024).

The original full-spectrum argument provides the conceptual resolution. Using the full mass distribution rather than a single edge breaks many apparent degeneracies, because cosmology shifts all source-frame features coherently through $$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$6, whereas generic astrophysical evolution deforms different parts of the spectrum differently. The degeneracy fails only in the case of an “astrophysical conspiracy” that shifts all features simultaneously following the Hubble-diagram evolution (Ezquiaga et al., 2022). This suggests that robustness grows with the number of genuinely informative source-frame landmarks, provided the model can represent them.

5. Current catalog analyses and published constraints

Current spectral-siren constraints are already based on real gravitational-wave catalogs, but the results remain visibly model dependent. A non-parametric binned-GP analysis of 69 GWTC-3 BBHs, combined with the bright siren GW170817/NGC 4993, reported

$$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$7

at 68% credibility, around a factor of 1.4 better than GW170817 alone (Hernandez et al., 2024). A multi-spectral analysis of GWTC-3 using two subpopulations in mass and spin found

$$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$8

from BBHs alone, and

$$N_{\rm exp}(\Lambda)=T_{\rm obs}\int d\theta\,dz\; P_{\rm det}(\theta,z,\Lambda)\,\frac{1}{1+z}\,\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s}(\Lambda).$$9

or

$\theta$0

when combined with GW170817, depending on the $\theta$1 prior (Li et al., 2024).

GWTC-4.0 has produced a wider range of published spectral-siren measurements because the larger catalog permits more ambitious population models. An analysis of 152 BBHs with parametric and non-parametric mass functions reported broad consistency across models, with the tightest result coming from the Gaussian-process model combined with GW170817,

$\theta$2

corresponding to a 10% measurement (Hernandez et al., 3 Sep 2025). A B-POP-calibrated Normalizing-Flow analysis of 153 GWTC-4.0 events reported

$\theta$3

together with a dynamical-channel fraction $\theta$4, but the authors explicitly characterize this result as model dependent because the detailed source distributions are fixed by the trained astrophysical catalogs (Scarpa et al., 16 Mar 2026).

Other GWTC-4.0 analyses isolate specific aspects of the mass model. A flexible joint model for component masses and pairing, emphasizing secondary-mass structure, found

$\theta$5

from 142 CBCs alone and

$\theta$6

when combined with GW170817 (Li et al., 12 May 2026). A model designed to test heavy-black-hole structure obtained

$\theta$7

with spectral sirens and

$\theta$8

with dark sirens, attributing the improvement to a new mass scale at $\theta$9 (Pierra et al., 6 Jan 2026). A BBH-only GWTC-4.0 reanalysis using sharper mixtures of tapered power laws reported

$\Lambda$0

at 68% confidence and described this as a 23% measurement and roughly a 50% improvement over the corresponding LVK BBH-only analysis (Bertheas et al., 6 Mar 2026).

Taken together, these results establish that spectral-siren cosmology is already empirically operative on current catalogs, but they also show that different source-population assumptions can shift both the central value and the width of the inferred $\Lambda$1 posterior. That spread is consistent with the literature’s general conclusion that cosmological inference and mass-spectrum inference cannot be cleanly separated (Pierra, 2024).

6. Hybrid methods, spectroscopy, and the third-generation regime

Spectral sirens increasingly appear as one limit of a broader population-based siren program. A unified hierarchical treatment based on galaxy density shows that the spectral-siren method and the galaxy-catalog dark-siren method are “two sides of the same coin”: when the catalog is empty or effectively uninformative, the likelihood reduces to a spectral-siren analysis; when the catalog is highly informative, the redshift prior becomes galaxy dominated (Mastrogiovanni et al., 2023). More recent work using CHIMERA makes the same point operationally by modeling catalog incompleteness: as completeness decreases, the correlation between $\Lambda$2 and the BBH mass-scale parameter strengthens, and the inference transitions toward the spectral-siren regime (Borghi et al., 22 Sep 2025).

This hybrid viewpoint has made spectroscopy a central infrastructure issue. A white paper on future standard sirens argues that spectroscopic redshifts with

$\Lambda$3

allow 100 well-localized BBH events at O5-like sensitivity to deliver percent-level $\Lambda$4, whereas photometric redshifts with

$\Lambda$5

degrade the same sample to a 9% $\Lambda$6 measurement (Borghi et al., 20 Dec 2025). In a related simulation framework with spectroscopic catalogs, complete catalogs yield $\Lambda$7 precisions of 1.6%, 1.3%, and 0.9% for constant, linear, and quadratic stellar-mass host weighting, and about 2% precision remains achievable even when only 50% of potential hosts are present within the gravitational-wave horizon (Borghi et al., 22 Sep 2025). Another dark-siren robustness study finds that 100 BBH detections with complete or volume-limited catalogs and corrected stellar-mass weighting can reach approximately 3% precision on $\Lambda$8, while equal host weighting degrades the precision to approximately 6% (Alfradique et al., 24 Mar 2025). These are not spectral-siren measurements in the strict GW-only sense, but they quantify the regime in which catalog incompleteness forces the inference back toward mass-spectrum information.

The third-generation outlook is correspondingly twofold: more events and stronger sensitivity to population systematics. The original full-spectrum forecast argued that second-generation detectors could achieve better than 10% precision on $\Lambda$9 at $(1+z)^{-1}$0 within a year, while third-generation detectors could reach $(1+z)^{-1}$1 at $(1+z)^{-1}$2 within one month (Ezquiaga et al., 2022). A blinded mock data challenge for ET-era spectral sirens has now validated three public pipelines—icarogw, chimera, and pymcpop-gw—on a catalog of $(1+z)^{-1}$3 high-S/N BBHs, finding consistent recovery of cosmological and population parameters and reporting a 2.4% measurement of $(1+z)^{-1}$4 at $(1+z)^{-1}$5, a mean precision of 2.8% across $(1+z)^{-1}$6, and joint constraints of $(1+z)^{-1}$7 on $(1+z)^{-1}$8 and $(1+z)^{-1}$9 on $$\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s} = R_0\,\psi(z;\Lambda)\,p_{\rm pop}(\vec m_{\rm s}|\Lambda)\,\frac{dV_c}{dz},$$0 (Tagliazucchi et al., 19 Feb 2026). That study also identifies the dominant carriers of cosmological information: low-distance sources near population features drive the constraining power on all cosmological parameters, while higher-distance events contribute mainly to $$\frac{dN_{\rm CBC}}{d\theta\,dz\,dt_s} = R_0\,\psi(z;\Lambda)\,p_{\rm pop}(\vec m_{\rm s}|\Lambda)\,\frac{dV_c}{dz},$$1 (Tagliazucchi et al., 19 Feb 2026).

The resulting picture is technically consistent across the literature. Spectral-siren cosmology is a GW-only route to cosmological inference whose power derives from structured source-frame populations, whose vulnerability derives from population-model misspecification, and whose future precision depends simultaneously on detector yield, population-model flexibility, and—when galaxy information is incorporated—the quality and completeness of spectroscopic host catalogs.