Spectral Sirens in Gravitational-Wave Cosmology

• Spectral sirens are a gravitational-wave method that infers redshift from distinct features in the compact binary mass spectrum, bypassing the need for electromagnetic observations.
• They use hierarchical Bayesian analysis to jointly extract cosmological parameters, such as H₀, and source-population hyperparameters from merger catalogs.
• Advancements in flexible population modelling and scalability to third-generation detectors enhance precision and mitigate biases in spectral sirens inference.

to=arxiv_search.search 经彩票_markdown üpjureness=0.47 code='{"query":"all:\"spectral sirens\" OR ti:\"spectral sirens\" gravitational-wave cosmology H0", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"}' to=arxiv_search.search 万亚chive=markdown code='{"query":"ti:\"spectral sirens\" OR abs:\"spectral sirens\" gravitational wave cosmology", "max_results": 10, "sort_by": "relevance", "sort_order": "descending"}' to=arxiv_search.search tungaanut_markdown code='{"query":"\"spectral sirens\" gravitational-wave standard sirens Hubble constant", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"}' Spectral-sirens inference is a gravitational-wave cosmology methodology that uses gravitational-wave data alone to infer cosmological parameters by exploiting structure in the source-frame distribution of compact-binary masses. In contrast to bright standard sirens, which require an electromagnetic counterpart, and dark sirens, which require statistical association with galaxy catalogs, spectral sirens use the redshifted-mass relation $m_{\rm det} = (1+z)m_{\rm src}$ together with the luminosity-distance relation to extract redshift information from peaks, breaks, cutoffs, and gaps in the compact-object mass spectrum (Hernandez et al., 3 Sep 2025). In its modern form, the method is a hierarchical population analysis in which cosmology and source-population hyperparameters are inferred jointly from catalogs of compact-binary mergers, especially binary black holes, with the full mass distribution functioning as a set of internal spectral markers (Ezquiaga et al., 2022).

1. Conceptual basis and relation to other siren methods

The defining premise of spectral sirens is that detector-frame masses are redshifted relative to source-frame masses. If the source population contains robust mass features, then those features appear in the observed catalog at detector-frame locations that “slide” with redshift and therefore with cosmology. Across many events, the joint distribution of detector-frame masses and luminosity distances statistically constrains the distance–redshift relation and hence parameters such as $H_0$ (Hernandez et al., 3 Sep 2025).

This places spectral sirens between the two standard reference cases of gravitational-wave cosmology. Bright sirens use an electromagnetic counterpart to identify the host galaxy and measure the source redshift directly. Dark sirens infer redshift statistically by cross-correlating the gravitational-wave localization volume with galaxy catalogs. Spectral sirens require neither an electromagnetic counterpart nor galaxy catalogs; they instead rely on astrophysical structure in the compact-binary population itself (Hernandez et al., 3 Sep 2025).

The modern formulation was broadened by the argument that the entire mass spectrum, rather than a single feature, should be used as a cosmological probe. In that picture, the mass spectrum of compact binaries introduces 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 gap, and the minimum neutron-star mass (Ezquiaga et al., 2022). This full-spectrum perspective is important because multiple spectral anchors can break degeneracies between cosmological evolution and astrophysical mass evolution unless all features shift coherently in a way that mimics the Hubble diagram (Ezquiaga et al., 2022).

A broader antecedent is the luminosity-distance–only standard-siren framework, which used a prior on the detected redshift distribution constructed from merger-rate models and detector selection to infer cosmology without redshift measurements. In spectral-siren language, that is a rate-only limit of the broader GW-only inference problem (Ding et al., 2018).

2. Hierarchical Bayesian formulation

Most spectral-siren analyses adopt a flat $\Lambda$CDM cosmology in which

$$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$

The detector-to-source mapping is then

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

These two relations couple the mass population to cosmology, because any trial cosmology changes the inferred redshift and therefore the source-frame mass assigned to each event (Hernandez et al., 3 Sep 2025).

The standard inference problem is hierarchical. For multiple events $\{d_i\}$, one jointly infers cosmological parameters and population hyperparameters through a population likelihood with selection effects. One common form is

$$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$

where $\theta$ denotes source parameters, $p(\theta \mid z; \phi)$ is the population model, $p(z \mid H_0)$ encodes the cosmology-dependent redshift prior, and $H_0$0 is the detectable fraction (Hernandez et al., 3 Sep 2025). Equivalent Poisson point-process forms are also used, with the expected number of detections $H_0$1 carrying the same selection normalization (Tagliazucchi et al., 6 Jan 2026).

At population level, the source-frame density is typically written as

$H_0$2

with $H_0$3 a parametric merger-rate evolution such as $H_0$4 or a Madau-like form (Hernandez et al., 3 Sep 2025). Selection enters through injection campaigns that estimate $H_0$5 and hence $H_0$6 for the actual detector sensitivity (Hernandez et al., 3 Sep 2025).

In practice, catalog analyses operate on event-level posterior samples released by LVK and reweight them by the ratio of the trial population prior to the original sampling prior, together with the Jacobian implied by the cosmology-dependent distance–redshift mapping. This importance-sampling strategy propagates measurement uncertainties in masses, spins, and luminosity distance directly into the hyperposterior (Hernandez et al., 3 Sep 2025).

3. Population models and the identification of spectral markers

The earliest widely used spectral-siren models were heuristic parametric descriptions of the binary black hole mass function. These include the Powerlaw + Peak model, with a truncated power law plus a Gaussian excess, and the Broken Powerlaw + 2 Peaks model, with a break mass, two slopes, and two Gaussian peaks (Hernandez et al., 3 Sep 2025). Such models encode the idea that stellar-evolution channels generate a declining black-hole mass function with additional structure near the onset of the pair-instability gap.

As catalogs grew, richer models were introduced to reduce misspecification risk. A Gaussian-process model places a Gaussian-process prior on a smoothly varying black-hole mass function and allows localized features and rate suppression to be learned from the data rather than prescribed. In GWTC-4.0 analyses, this model recovers 10–40 $H_0$7 structure, suppression above $H_0$8 consistent with the onset of the pair-instability supernova mass gap, and a high-mass subpopulation at $H_0$9–$\Lambda$0 (Hernandez et al., 3 Sep 2025).

Semiparametric approaches go further by making the feature locations themselves data-adaptive. A recent B-spline model writes the primary-mass distribution as a smooth deformation of a truncated power law, with cubic B-spline coefficients regularized by Gaussian priors. Its adaptive-knot procedure identifies candidate knots from peaks in the derivative of the log of the mean observed source-frame primary-mass distribution, repeats this over $\Lambda$1 draws, and clusters the candidates with a Gaussian Mixture Model whose number of clusters is set by the Bayesian Information Criterion. In GWTC-4.0 this resolves three peaks at roughly 10, 18, and $\Lambda$2 (Tagliazucchi et al., 6 Jan 2026).

Two additional extensions address feature blurring. Multi-spectral sirens model the population as a mixture of subpopulations, with Dirichlet mixture fractions and joint mass–spin components, so that otherwise blurred mass edges reappear within each component. In GWTC-3, a two-component model separated a low-spin first-generation component from a higher-spin hierarchical component and sharpened the upper edge of the first subpopulation and the lower edge of the second (Li et al., 2024). Spinning spectral sirens use the fact that dimensionless spins are not redshifted to anchor a source-frame mass scale through a mass–spin correlation. In that framework, a transition mass $\Lambda$3 separates a narrow low-$\Lambda$4 regime from a broader high-spin regime, and the detector-frame transition track $\Lambda$5 itself becomes a cosmological ruler (Tong et al., 15 Feb 2025).

A separate parametric direction introduces mixtures of tapered power laws with very steep slopes. The 3sPL and 4sPL families were designed to capture narrow low-mass peaks, dips near $\Lambda$6, the $\Lambda$7 peak, and possible high-mass structure, while an alternative sPL+2G model combines a tapered power law with two truncated Gaussians (Bertheas et al., 6 Mar 2026). Across these developments, a common lesson is that cosmological performance depends directly on how faithfully the mass model resolves all informative structure.

4. Results from current gravitational-wave catalogs

Current catalog analyses show broad consistency across models, but they also show that richer or sharper mass structure tends to tighten the inferred $\Lambda$8 posterior. In GWTC-4.0, a 152-event binary-black-hole sample excluding GW231123 produced BBH-only constraints of $\Lambda$9 for Powerlaw + Peak, $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$0 for Broken Powerlaw + 2 Peaks, and $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$1 for the Gaussian Process model; combining the same posteriors with GW170817 gave $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$2, $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$3, and $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$4 respectively (Hernandez et al., 3 Sep 2025).

Study or model	Data	Reported result
Gaussian Process + GW170817	GWTC-4.0, 152 BBHs + GW170817	$$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$5
pls-dd-14, $$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$6	137 GWTC-4.0 BBHs	$$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$7
4sPL	GWTC-4.0, 150 BBHs	$$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$8
TwoSpin multi-spectral sirens	GWTC-3, 42 BBHs	$$D_L(z; H_0, \Omega_m) = (1+z)\frac{c}{H_0}\int_0^z \frac{dz'}{\sqrt{\Omega_m(1+z')^3 + (1-\Omega_m)}}.$$9
Spin-only ruler	GWTC-3, 69 BBHs	$$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$0

The semiparametric B-spline analysis is notable because the evidence-favored configuration, pls-dd-14 with $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$1, was strongly preferred over the two-peak parametric baseline with a Bayes factor of 226, and a more flexible prior with $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$2 yielded $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$3 (Tagliazucchi et al., 6 Jan 2026). The sharper-feature 4sPL analysis is notable because it reached 23% precision, $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$4, described as a $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$5 improvement over the corresponding LVK binary-black-hole-only analysis (Bertheas et al., 6 Mar 2026).

On GWTC-3, subpopulation separation already improved precision. The two-component “TwoSpin” multi-spectral-siren model gave $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$6, about 19% tighter than a single-population PowerLaw+Peak analysis in the same framework, and the combination with GW170817 yielded $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$7 for a uniform prior and $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$8 for a log-uniform prior (Li et al., 2024). Spin-only spectral information is currently weaker but already informative: a mass–spin transition model on GWTC-3 gave $$m_{\rm src} = \frac{m_{\rm det}}{1+z}, \qquad m_{\rm det} = (1+z)m_{\rm src}.$$9, while combining spin information with other mass features gave $\{d_i\}$0 (Tong et al., 15 Feb 2025).

5. Systematics, model risk, and robustness

The central controversy in spectral-siren inference is not whether the hierarchy can be written down, but whether the assumed population model captures the true mass distribution with sufficient fidelity. A systematic study using astrophysically motivated binary-black-hole populations found that, with 2000 detected mergers, spectral-siren measurements of $\{d_i\}$1 can be biased up to $\{d_i\}$2 if heuristic mass models fail to capture redshift evolution or unforeseen structure (Pierra et al., 2023).

Several explicit failure modes were demonstrated. If catalogs generated from a Power Law plus Peak or Multi-Peak population are analyzed with a Broken Power Law, the true $\{d_i\}$3 is excluded at $\{d_i\}$4 credible level. If the Gaussian peak drifts linearly with redshift, a shift of $\{d_i\}$5 excludes the true $\{d_i\}$6 at $\{d_i\}$7 credible level. In the more complex A03 astrophysical catalog, inference with PLP yielded $\{d_i\}$8, MLTP gave $\{d_i\}$9, and only the broad BPL posterior remained consistent within $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$0 credible interval; randomizing the mass–redshift pairing largely removed the bias and returned PLP to $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$1 and MLTP to $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$2 (Pierra et al., 2023).

Current catalog papers respond to this risk in three ways. First, they use more flexible models: Gaussian processes, semiparametric B-splines, and steep tapered-power-law mixtures all explicitly aim to reduce misspecification (Hernandez et al., 3 Sep 2025). Second, they perform model checks and sensitivity studies. Recent GWTC-4.0 analyses report broad consistency across parametric and non-parametric models, propagate waveform and calibration uncertainties through the LVK per-event posteriors, incorporate detection efficiencies via GWTC-4.0 injections, and verify that excluding GW231123 has negligible impact on $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$3 (Hernandez et al., 3 Sep 2025). Third, they introduce internal cross-checks. Spinning spectral sirens show that a misspecified fixed-peak mass model can bias the mass-only channel high, while the spin-only channel remains nearly unchanged; in a 300-event simulation with an injected evolving peak, the misspecified mass model still gave $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$4, nearly identical to the correct-model result $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$5 (Tong et al., 15 Feb 2025).

A plausible implication is that future precision will depend less on raw event counts than on whether population models can accommodate mass–spin–redshift complexity without erasing the very features that carry cosmological information.

6. Third-generation scalability and unified dark-siren frameworks

Third-generation detectors change the problem from one of proof of concept to one of scalable inference. A blinded mock data challenge for the Einstein Telescope tested three public pipelines—ICAROGW, CHIMERA, and pymcpop-gw—on simulated ET catalogs containing the best $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$6 binary black hole mergers observable in one year. All three pipelines recovered consistent cosmological and population parameters (Tagliazucchi et al., 19 Feb 2026).

In the SNR $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$7 catalog, containing $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$8 high-S/N events, the challenge measured $$p(H_0,\phi \mid \{d_i\}) \propto p(H_0)p(\phi)\prod_i \int dz\, d\theta\, \frac{p(d_i \mid \theta, z)\, p(\theta \mid z; \phi)\, p(z \mid H_0)}{\alpha(H_0,\phi)},$$9 at $\theta$0 with 2.4% precision and achieved a mean precision on $\theta$1 of 2.8% across $\theta$2, corresponding to joint constraints of $\theta$3 on $\theta$4 and $\theta$5 on $\theta$6 (Tagliazucchi et al., 19 Feb 2026). The event-level information content was highly structured: low-distance sources near population features drive the constraining power on all cosmological parameters, while higher-distance events primarily constrain $\theta$7 (Tagliazucchi et al., 19 Feb 2026). Computationally, GPU acceleration brought ET-scale analyses into a manageable regime, and the study concluded that on a single modern GPU, $\theta$8 events with $\theta$9 posterior samples and a few million detected injections can be analyzed in approximately one to two weeks (Tagliazucchi et al., 19 Feb 2026).

The same era motivates hybrid formulations that combine spectral and dark-siren information in a single likelihood. A unified harmonic framework treats gravitational-wave sources and galaxies as Gaussian random fields in tomographic shells, with spectral-siren information setting the radial weights through the detected redshift distribution and angular power spectra encoding the clustering and distance–redshift relation. In forecasts for a 2 Einstein Telescope + 1 Cosmic Explorer setup over two years, the cross-correlation part alone measures $p(\theta \mid z; \phi)$0 and $p(\theta \mid z; \phi)$1 to 1% and 5% precision, while the full 3×2pt analysis reaches $p(\theta \mid z; \phi)$2 and $p(\theta \mid z; \phi)$3 (Cheng et al., 13 Mar 2026).

Longer-range forecasts already anticipated this shift. Using the full mass distribution, one study argued that second-generation detectors could achieve better than 10% precision on $p(\theta \mid z; \phi)$4 at $p(\theta \mid z; \phi)$5 within a year, while third-generation detectors could reach $p(\theta \mid z; \phi)$6 at $p(\theta \mid z; \phi)$7 within one month; in that picture, the pair-instability gap dominates current 2G inference, whereas the lower mass gap becomes most powerful in the 3G era (Ezquiaga et al., 2022).

7. Terminological note

In current arXiv usage, “spectral sirens” refers to the gravitational-wave cosmology program summarized above. This should be distinguished from work on SIRENs as sinusoidal implicit neural representations, where papers study rapid prediction of SIREN encoding error from image features and network hyperparameters (Vonderfecht et al., 2024) or propose target-aware initialization schemes such as WINNER to mitigate a “spectral bottleneck” in sinusoidal representation networks (Chandravamsi et al., 16 Sep 2025). Those studies concern implicit neural representations rather than gravitational-wave standard sirens.

Within gravitational-wave astronomy, however, spectral-sirens inference has become a distinct and rapidly developing program: a GW-only route to cosmology in which the source population itself provides the redshift information. Its current limitations are dominated by population modeling, redshift evolution, and selection-function control; its principal opportunity is that larger catalogs sharpen spectral markers and make internal consistency tests increasingly powerful.