BEAMS with Bias Corrections (BBC) in Cosmology

• BEAMS with Bias Corrections (BBC) is a statistical methodology that integrates selection bias correction and probabilistic mixture modeling to construct unbiased Hubble diagrams from photometric supernova samples.
• It leverages forward simulations, such as BiasCor, to correct for Malmquist and selection biases in light-curve parameters of Type Ia SNe, thereby enhancing the accuracy of cosmological parameter estimation.
• BBC extends the original BEAMS framework by incorporating host-galaxy correlations and multi-dimensional bias modeling, making it essential for modern SN cosmology analyses.

BEAMS with Bias Corrections (BBC) is a statistical and computational methodology designed to infer unbiased cosmological parameters and construct Hubble diagrams from contaminated, photometrically classified samples such as Type Ia supernovae (SNe Ia), while correcting for survey selection effects and non-Ia contamination. BBC generalizes the original BEAMS framework by integrating rigorous bias corrections, probabilistic population separation, and forward model systematics, and has become the backbone of modern SN cosmology analyses, including the Dark Energy Survey and Pan-STARRS light-curve cosmology (Kessler et al., 2016, Popovic et al., 2021, Kessler et al., 2023).

1. Theoretical Foundation: BEAMS and Its Extensions

The Bayesian Estimation Applied to Multiple Species (BEAMS) likelihood is the foundation for BBC. BEAMS models data as a mixture of distinct populations (e.g., Ia and core-collapse SNe), using event-wise type probabilities $P_i$:

$$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$

where $L_i^{(\mathrm{Ia})}$ and $L_i^{(\mathrm{CC})}$ are the likelihoods under each species, and $P_i$ is the classifier-assigned Type Ia SN probability (Kessler et al., 2023, Kessler et al., 2016). The standard BEAMS likelihood assumes $P_i$ are calibrated and independent of the data; this is inadequate in practical surveys with selection-induced biases and feature/probability correlations (Newling et al., 2011). BBC extends BEAMS by

• Explicitly modeling correlations between type probabilities and measured features, $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$.
• Incorporating classifier biases via selection-bias reweighting.
• Propagating systematic and selection effects into likelihood and posterior estimation.

The full posterior is

$$P(\theta|D,P) \propto P(\theta) \prod_i \Big[ P_i f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,A) + (1-P_i) f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,B) \Big]$$

where $\theta$ parameterizes both cosmology and population/systematics (Newling et al., 2011).

2. Bias Corrections and BiasCor Simulations

Photometric SNe samples incur Malmquist and selection-driven biases in light-curve fitted parameters $(m_B, x_1, c)$. BBC corrects these by generating large-scale forward simulations ("BiasCor") of the survey, evaluating the average bias in each observed parameter as a function of $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$0:

$$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$1

The bias-corrected parameters are:

$$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$2

yielding a bias-corrected distance modulus (modified Tripp estimator):

$$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$3

Alternatively, the three corrections are aggregated into a single $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$4, usually tabulated/interpolated on a 5D grid $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$5 (Kessler et al., 2016, Kessler et al., 2023).

Key to the statistical rigor, BBC models the contaminant likelihood $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$6 directly from simulations, incorporating the classifier's cross-classification performance. BBC also corrects for incorrectly estimated type probabilities by exploiting "debiasing" steps, using all variables that influence spectroscopic confirmation, following a weight-function approach if classifier features and selection drivers differ (Newling et al., 2011).

3. BEAMS Likelihood Implementation and Estimation Procedures

For each event, the combined likelihood is

$$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$7

where $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$8 incorporates an overall scaling $$L_i = P_i \, L_i^{(\mathrm{Ia})} + (1-P_i)\, L_i^{(\mathrm{CC})}$$9 to absorb classification normalization errors, and the likelihoods are:

$L_i^{(\mathrm{Ia})}$0

with total variance $L_i^{(\mathrm{Ia})}$1.

Core-collapse (CC) likelihoods are constructed from large-scale Monte Carlo simulations, building a 2D $L_i^{(\mathrm{Ia})}$2 map, normalized in $L_i^{(\mathrm{Ia})}$3 for each $L_i^{(\mathrm{Ia})}$4.

The maximized posterior yields both the Tripp nuisance parameters $L_i^{(\mathrm{Ia})}$5, the zero-point $L_i^{(\mathrm{Ia})}$6, the per-bin Hubble diagram offsets $L_i^{(\mathrm{Ia})}$7, and the intrinsic scatter $L_i^{(\mathrm{Ia})}$8 (Kessler et al., 2016).

4. Advanced Bias Modeling: Host-Galaxy Correlations and Multi-Dimensional Extensions

Systematic host-galaxy correlations (notably the mass step, $L_i^{(\mathrm{Ia})}$9) are crucial for cosmological accuracy. BBC7D extends the framework by introducing host-galaxy stellar mass dependence into the bias-correction grid and Tripp-like parameterization:

• Underlying SN Ia light-curve property distributions $L_i^{(\mathrm{CC})}$0 and $L_i^{(\mathrm{CC})}$1 are empirically modeled in host-mass bins using migration matrices, allowing forward modeling of host-mass–dependent selection and intrinsic population effects.
• The mass step is parameterized as a smooth sigmoid:

$L_i^{(\mathrm{CC})}$2

with $L_i^{(\mathrm{CC})}$3 amplitude, $L_i^{(\mathrm{CC})}$4 pivot log-mass, and $L_i^{(\mathrm{CC})}$5 step width, and propagated through the bias correction grid (Popovic et al., 2021).

• BBC7D adds two grid dimensions: a magnitude shift parameter $L_i^{(\mathrm{CC})}$6 and $L_i^{(\mathrm{CC})}$7, enabling interpolation for any trial $L_i^{(\mathrm{CC})}$8 and explicit host property dependence.
• The empirical results show that BBC7D reduces $L_i^{(\mathrm{CC})}$9-bias by a factor of $P_i$05 and $P_i$1-bias by $P_i$22 over the BBC5D formalism and recovers input $P_i$3 to within $P_i$4 mmag and $P_i$5 to sub-percent accuracy (Popovic et al., 2021).

BBC also enables handling of other host-property dependencies and is validated for population and systematics modeling, including the use of the BS20 dust-based intrinsic scatter model.

5. Construction of the Hubble Diagram: Binned, Unbinned, and Re-binned Approaches

BBC produces both binned and unbinned Hubble diagrams (HDs):

• Binned HD: The default output is a set of redshift bins with fitted offsets $P_i$6; these are cosmology-independent and serve as a lossless summary for subsequent cosmological fits.
• Unbinned HD: Recent advances demonstrate that unbinned HDs further reduce systematic uncertainties by $P_i$77% in total $P_i$8 precision ($P_i$920% reduction in systematics), particularly in the presence of self-calibration (Kessler et al., 2023). For large samples, computational constraints motivate “rebinned” HDs in multidimensional bins of $P_i$0, striking an optimal balance of information retention and tractability.

The total covariance matrix $P_i$1 is constructed using repeated analyses under systematic shifts, propagating all key sources (calibration, SALT2 retraining, filter curves, etc.) (Kessler et al., 2023).

6. Simulation Validation, Performance, and Best Practices

Large-scale DES-like simulations confirm BBC's performance:

Configuration	Recovered $P_i$2	Recovered $P_i$3	$P_i$4-bias	Post-NN contamination
Nominal CC rate	$P_i$5	$P_i$6	$P_i$7 (COH)	$P_i$8
3$P_i$9 nominal CC rate	$f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$0	$f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$1	$f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$2 (C11)	$f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$3

BBC enables $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$4 and $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$5 recovery to $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$6 and $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$7-bias at the $f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$8 level on DES-scale samples ($f_{D_i|\Theta,P_i,T_i}(d_i|\theta,p_i,\tau)$9,000 SNe) (Kessler et al., 2016). Inclusion of host-galaxy-dependent population models reduces mass-step and $$P(\theta|D,P) \propto P(\theta) \prod_i \Big[ P_i f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,A) + (1-P_i) f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,B) \Big]$$0-biases below 0.004 mag and 0.01, respectively (Popovic et al., 2021).

Recommended practices include:

• Recording and using all features driving follow-up to allow probability debiasing.
• Calibrating classifier probabilities against the unconfirmed SN population.
• Applying BBC likelihoods with full type-probability dependence and data-probability correlation modeling.
• For large survey data, adopting rebinned HDs in $$P(\theta|D,P) \propto P(\theta) \prod_i \Big[ P_i f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,A) + (1-P_i) f_{D_i|\theta,P_i,T_i}(d_i|\theta,P_i,B) \Big]$$1 (Kessler et al., 2023, Newling et al., 2011).

7. Generalizations Beyond Supernovae and Related Contexts

"BBC" as BEAMS with Bias Corrections is specific to SN cosmology, but the formalism is transferable to any context where multiple-population data, systematic contamination, or selection biases dominate inference. For instance, CMB beam-focused "BBC" methods correct temperature-to-polarization leakage by creating unbiased estimators and map reconstructions using comparable principles of instrument model correction and full joint error propagation (Wallis et al., 2014, Svalheim et al., 2022).

In all contexts, the signature methodological features are:

• Explicit forward-modeling of population- and measurement-dependent biases.
• Probabilistic mixture modeling with per-object weights, rigorous uncertainty propagation, and correction for measurement-induced selection.
• Monte Carlo–driven or spurious-map–based approaches for instrument model systematics, integrated within fully Bayesian pipelines.

BBC has thus become a methodological standard for next-generation cosmological analyses where sub-percent systematics must be achieved in contaminated or incomplete datasets.