CosmoFit.Union3Likelihood

class CosmoFit.Union3Likelihood(cosmology, version='union3', marginalize_offset=True)

Bases: BaseLikelihood, AnalyticOffsetMixin

Union3 (Rubin et al. 2023): 2087 supernovae compressed into 22 binned distance moduli with a full 22x22 covariance.

The third of the three modern SN Ia compilations, alongside PantheonLikelihood and DESSN5YRLikelihood. It matters here because the DESI dark-energy results are argued with all three, and they do not agree with each other on how far the data sits from a cosmological constant – DES-SN5YR pulls hardest, Pantheon+ least, Union3 in between. A library that can only fit one of them cannot reproduce that comparison, which is the actual state of the evidence.

What makes Union3 different is not the sample so much as the analysis: UNITY1.5 is a Bayesian hierarchical model that fits light-curve standardization, host-galaxy mass dependence, selection effects and outliers jointly with cosmology, and marginalizes them internally. The released product is therefore already a binned distance-modulus vector – there is no per-supernova catalogue to re-standardize, and no stretch/colour nuisance parameters left for a downstream fit to vary.

As with both other SN samples, the overall zero point is degenerate with H0 and is marginalized analytically (see AnalyticOffsetMixin) rather than fit – which, for a 22-point dataset, matters more than it does for a 1600-point one.

Parameters:

• cosmology – Cosmology model instance (LCDM, CPL, …).

• version (str, optional) – Dataset version.

• marginalize_offset (bool, optional) – If True (default), the constant distance-modulus offset is marginalized over analytically. If False, cosmology.MB is added to the model as an explicit nuisance parameter instead.

Warning

Do not combine "union3" with "pantheon" or "des_sn5yr" in the same fit. Union3 shares a large majority of its supernovae with Pantheon+ (both compile essentially the same literature samples), and its high-redshift half overlaps the DES sample; treating any two of the three as independent double-counts most of the data. This is the same rule the library already applies between Pantheon+ and DES-SN5YR – one SN Ia compilation per fit.

References

Rubin et al. (2023), “Union Through UNITY: Cosmology with 2,000 SNe Using a Unified Bayesian Framework”, arXiv:2311.12098 (ApJ, accepted).

__init__(cosmology, version='union3', marginalize_offset=True)

Parameters:

• version (str)

• marginalize_offset (bool)

Methods

__init__(cosmology[, version, ...])
best_fit_offset()	Best-fit additive offset that analytic marginalization would assign, given the current cosmology.
chi2()	Chi-square statistic.
log_likelihood()	Return log-likelihood.
marginalized_chi2()	Chi-square with the offset analytically marginalized out.
model()	Predicted binned distance modulus.
predictions()	Alias for model().
residuals()	Data minus model residuals.
summary()	Return a summary of the likelihood evaluation.

Attributes

n_data	Number of data points.
name_and_size	Name together with the number of data points.
observable