Union3: SN Ia Compilation for Precision Cosmology

• Union3 is a modern Type Ia supernova compilation that unifies 2087 SNe Ia from 24 datasets using a hierarchical UNITY1.5 Bayesian framework.
• The methodology employs SALT3 light-curve fitting and incorporates detailed treatments of selection effects, outliers, and systematic uncertainties.
• The public Union3 product, often released as 22 binned distance moduli with a full covariance matrix, acts as a crucial high-redshift Hubble-diagram anchor.

Searching arXiv for the cited Union3 and related cosmology papers to ground the article in current literature. Union3 is a modern Type Ia supernova compilation designed for precision cosmology and analyzed with the UNITY Bayesian framework. In its canonical release, it comprises 2087 cosmologically useful SNe Ia from 24 datasets, broadly spanning $z\approx 0.01$–$2.26$, with light curves fit uniformly using SALT3 and with selection effects, standardization, outliers, peculiar velocities, and systematic uncertainties treated within UNITY1.5 (Rubin et al., 2023). In subsequent literature, Union3 is typically used not as a raw light-curve archive to be refit from scratch, but as a calibrated distance product—often in a compressed 22-bin form with a full covariance matrix—serving as a high-redshift Hubble-diagram anchor in late-time cosmology, dark-energy reconstruction, modified-gravity tests, and dataset-consistency studies (Mishra et al., 28 Mar 2026).

1. Compilation, scope, and survey composition

Union3 was assembled as an up-to-date “Union” compilation of Type Ia supernovae drawn from 24 datasets and placed on a common distance scale (Rubin et al., 2023). The contributing samples span nearby, intermediate-redshift, and high-redshift programs, including low-$z$ surveys such as LOSS, CSP, Foundation, CNIa0.02, and the CfA series; mid-$z$ surveys such as SDSS, SNLS, Pan-STARRS, and DES; and high-$z$ HST-based programs extending to $z\sim 2.3$ (Rubin et al., 2023). The sample is described in later analyses as containing 2087 SNe Ia from 24 surveys, with 1363 in common with PantheonPlus (Wang et al., 30 Jul 2025).

A central design feature of Union3 is cross-survey homogenization. The photometry is tied to CALSPEC spectrophotometric standards, passbands are harmonized through measured throughput curves or passband shifts, and zeropoints are updated to modern CALSPEC versions with wavelength-dependent uncertainties (Rubin et al., 2023). This structure is intended to make heterogeneous surveys usable within a single cosmological inference pipeline. A plausible implication is that Union3 is not merely a concatenation of existing SN samples, but a reprocessed compilation whose cosmological utility depends on common calibration conventions rather than on survey-level likelihoods being combined post hoc.

The redshift coverage reported across the literature depends on the data product being used. The full Union3 sample is described as spanning approximately $0.01$0.05\le z\le 2.26$ in downstream analyses, while public cosmology applications frequently use a binned release of 22 distance-modulus points covering the same broad range (Wang et al., 13 Apr 2026).

2. Standardization, SALT3, and UNITY1.5

All light curves in Union3 are fit with SALT3 over the full rest-frame optical wavelength range (Rubin et al., 2023). In the usual SALT-like notation, standardized SN distances are related to light-curve observables through a Tripp-type relation such as

$\mu = m_B - M + \alpha x_1 - \beta c,$

or, in some summaries,

$$\mu_{\rm obs} = m_B^* - M_B + \alpha x_1 - \beta c + \Delta_M + \Delta_B,$$

although many cosmological analyses that use Union3 do not fit these nuisance parameters themselves because they are already absorbed into the published Union3 products (Mishra et al., 28 Mar 2026).

The distinctive methodological element is UNITY1.5, described as a hierarchical Bayesian framework that models selection effects, standardization, outliers, unexplained dispersion, and systematic uncertainties simultaneously (Rubin et al., 2023). UNITY1.5 incorporates survey-dependent selection functions, population models for stretch and color, mixture-based outlier handling, peculiar-velocity covariance, and perturbative propagation of calibration systematics (Rubin et al., 2023). Several later papers summarize this operationally by stating that Union3 is “standardized and debiased with the UNITY1.5 Bayesian framework, which explicitly models selection biases, intrinsic dispersion, and systematic uncertainties” (Mishra et al., 28 Mar 2026).

This architecture matters because many downstream cosmology papers treat Union3 as a compilation-level likelihood rather than as a set of per-object SALT parameters to be re-estimated. In practice, this means that $2.26$0, $2.26$1, $2.26$2, host-mass corrections, and related hyperparameters are often taken as internal to the Union3/UNITY analysis rather than varied again in external MCMC runs. This suggests that Union3 is best understood as both a dataset and an inference framework product.

3. Public data products and likelihood construction

A recurrent feature of Union3 usage is the public availability of compressed, binned distance-modulus data. Multiple analyses state that the publicly released Union3 product consists of 22 binned $2.26$3 points together with a $2.26$4 covariance matrix, and that only the binned product is publicly available for some applications (Wang et al., 13 Apr 2026). In those uses, the standard SN likelihood takes the quadratic form

$2.26$5

with

$2.26$6

For flat FLRW backgrounds,

$2.26$7

These are the relations adopted explicitly in several Union3-based cosmology studies (Mishra et al., 28 Mar 2026).

The compressed format has methodological consequences. Flux averaging cannot be applied to Union3 because the released data are already pre-binned rather than object-level (Wang et al., 13 Apr 2026). Model-independent reconstructions based on derivatives of the SN Hubble diagram can inherit strong inter-bin correlations from the binned covariance; one analysis reports that the BLUE-derived $2.26$8 bins from Union3 show correlations up to $2.26$9 because the pre-binned inputs are themselves highly correlated (Wang et al., 13 Apr 2026). By contrast, the same compressed structure greatly simplifies cosmological sampling by removing the need to refit light-curve nuisance parameters.

A separate data-compression route is the “Spline-Interpolated Distance Moduli” model, in which Union3 distributes a Gaussian approximation for a 22-dimensional node posterior defined relative to a fiducial flat $z$0CDM baseline with $z$1 (Kim, 2024). That note emphasizes that the distributed object is a posterior in node space, not directly a likelihood in arbitrary cosmological parameter space, and that proper reuse requires dividing by the Union3 node prior and accounting for the Jacobian of the mapping to new model parameters (Kim, 2024). This clarified a technical misconception in third-party reuse of the Union3 spline posterior.

4. Role in cosmological inference

Union3 is used in the recent literature as one of the principal low-redshift geometric probes alongside BAO, cosmic chronometers, CMB distance priors, redshift-space distortions, and strong-lensing time delays. In model-dependent analyses, it is commonly combined with DESI BAO and Planck or late-time probes to constrain $z$2CDM, modified gravity, interacting dark energy, and other non-$z$3CDM scenarios (Giarè et al., 14 Feb 2025).

In parametric dark-energy fits, Union3 frequently favors $z$4 and $z$5 when combined with DESI and CMB. For example, in a CPL analysis surveying more than 35 dataset combinations, the combination CMB+DESI+Union3 yields

$z$6

with a reported $z$7 preference for CPL over $z$8CDM according to the paper’s $z$9-based significance metric (Giarè et al., 14 Feb 2025). In another analysis focused on late-time deceleration diagnostics, DESI BAO + CMB + Union3 gives

$z$0

together with

$z$1

which that work interprets as evidence that the cosmic acceleration has already peaked and is now declining within flat $z$2CDM (Wang et al., 30 Jul 2025).

Union3 also enters model-independent or weakly parametric reconstructions. A Gaussian-process reconstruction combining DESI DR1 BAO with Union3 finds that $z$3CDM is always included within $z$4, in contrast to stronger apparent deviations found with DES-5YR (Mukherjee et al., 2024). A non-parametric reconstruction of the dark-energy density ratio $z$5 with DESI DR2 + CMB + Union3 finds

$z$6

corresponding to an additional $z$7–$z$8 deviation at $z$9 and an approximately $z$0 deviation at $z$1, while still being consistent within $z$2 with the reconstructions from the other SN compilations overall (Wang et al., 13 Apr 2026).

These results do not all point in the same physical direction. Some papers find Union3 to strengthen evidence for dynamical dark energy or late-time departures from $z$3CDM, while others find it to be the SN compilation most consistent with $z$4CDM in nonparametric GP reconstructions (Mukherjee et al., 2024). A plausible implication is that Union3’s cosmological impact depends sensitively on the inference scheme, particularly on whether the analysis attributes inter-probe differences to genuine $z$5 evolution, to $z$6 inconsistencies, or to dataset-dependent systematics.

5. Consistency tests, dataset dependence, and interpretive tensions

Union3 occupies an important position in recent consistency debates because it often lies between Pantheon+/PantheonPlus and DESY5/DES-Dovekie in the strength of inferred departures from $z$7CDM. In the review of evolving dark energy constraints, DESI-based combinations with Union3 typically yield stronger dynamical-dark-energy hints than PantheonPlus but slightly weaker ones than DESY5 (Giarè et al., 14 Feb 2025). In “Probing departures from $z$8CDM by late-time datasets,” Union3 combined with CC + DESI DR2 pushes BA, JBP, CPL, Logarithmic, and GEDE models to roughly $z$9 departures, whereas DES-SN5Y reaches $z\sim 2.3$0–$z\sim 2.3$1 and Pantheon$z\sim 2.3$2 remains weaker (Chaudhary et al., 9 Oct 2025).

At the same time, calibration-independent tests using the Alcock–Paczynski variable $z\sim 2.3$3 portray Union3 as highly consistent with DESI DR2 BAO. In that framework,

$z\sim 2.3$4

on the BAO side, while on the SN side, in flat FLRW,

$z\sim 2.3$5

Because $z\sim 2.3$6 cancels in the BAO ratio and the SN absolute magnitude cancels in $z\sim 2.3$7, the test is independent of both $z\sim 2.3$8 and $z\sim 2.3$9. Using this construction, Union3 and Pantheon+ are found to have tension $0.01DES-Y5 sample showed $0.01Dinda et al., 24 Sep 2025, Dinda et al., 20 Jan 2026).

This coexistence of apparently significant dynamical-dark-energy fits and weak calibration-independent tension is a central interpretive issue. One model-independent study argues that the observed $0.01Wang et al., 13 Apr 2026). That paper notes that Union3 gives the highest SN-only flat-$0.01

$0.01

and argues that this preference helps explain its more pronounced $0.01Wang et al., 13 Apr 2026). This suggests that some of the Union3-driven dark-energy signal may reflect residual parameter inconsistencies between probes rather than uniquely constraining evidence for evolving dark energy.

6. Limitations, public-use caveats, and the Union3.1 update

Several limitations recur in the secondary literature. First, many cosmological analyses use only the 22-point binned Union3 product, not the full object-level sample, so information is compressed and certain procedures—especially flux averaging and internal sample splits—cannot be applied (Wang et al., 13 Apr 2026). Second, external studies usually rely on the Union3 collaboration’s internal treatment of calibration, host-mass corrections, peculiar velocities, bias corrections, and nuisance marginalization, rather than re-estimating these ingredients (Mishra et al., 28 Mar 2026). Third, the spline-posterior release requires careful treatment because the distributed Gaussian approximation is a posterior in node space with a non-flat induced prior in $0.01Kim, 2024).

A major update is Union3.1, which revises host-galaxy properties for approximately 2000 SNe using aperture-matched, homogeneously reduced photometry from the DESI Legacy Imaging Surveys and Prospector SED fitting (Hoyt et al., 27 Jan 2026). Union3.1 was motivated by the finding that host masses of $0.05\le z\le 2.26$0 SNe in Union3 had been overestimated on average, while the opposite was true for $0.05\le z\le 2.26$1 SNe in Pantheon+ (Hoyt et al., 27 Jan 2026). After correction, the two studies’ low-redshift mean distance moduli, previously more than $0.05\le z\le 2.26$2 mag discrepant, come into $0.05\le z\le 2.26$3 mag agreement (Hoyt et al., 27 Jan 2026).

The update modifies the UNITY analysis as well. In flat $0.05\le z\le 2.26$4CDM, Union3.1 gives

$0.05\le z\le 2.26$5

a $0.05\le z\le 2.26$6 shift from Union3 (Hoyt et al., 27 Jan 2026). When combined with BAO and CMB “exactly as done by DESI DR2,” Union3.1 yields

$0.05\le z\le 2.26$7

corresponding to $0.05\le z\le 2.26$8 evidence against a cosmological constant, down from $0.05\le z\le 2.26$9 for Union3 (Hoyt et al., 27 Jan 2026). The same paper reports that uncertainties on all standardization parameters shrink to $\mu = m_B - M + \alpha x_1 - \beta c,$0–$\mu = m_B - M + \alpha x_1 - \beta c,$1 their previous sizes (Hoyt et al., 27 Jan 2026).

Union3.1 therefore reframes Union3 not as a finished endpoint but as part of an evolving calibration program. The broader implication is that host-property inference, survey cross-calibration, and public data compression are not peripheral implementation details; they materially affect $\mu = m_B - M + \alpha x_1 - \beta c,$2, $\mu = m_B - M + \alpha x_1 - \beta c,$3, and $\mu = m_B - M + \alpha x_1 - \beta c,$4 in DESI-era combined analyses.