Bias Gaussianization Correction (BGc) Overview

• BGc is a statistical method that removes the lognormal bias induced by Gaussian uncertainties in extragalactic distance surveys, ensuring unbiased inference.
• It employs local median recentering and rescaling to transform distance and velocity measurements into a Gaussian distribution, enhancing power spectrum analysis.
• Validated with simulated catalogs and galaxy clustering data, BGc achieves consistent Hubble constant recovery and accurate large-scale structure mapping.

Bias Gaussianization Correction (BGc) is a statistical methodology developed to eliminate the systematic lognormal bias arising from Gaussian uncertainties in distance modulus measurements in extragalactic distance surveys. It has been applied to both peculiar velocity surveys such as Cosmicflows-3 (CF3) and to galaxy clustering analyses using Gaussianization of density fields to remove scale-dependent bias and enhance the fidelity of power spectrum measurements (Hoffman et al., 2021, 1511.02034).

1. Statistical Basis and Origin of Lognormal Bias

In distance indicators based on relations such as Tully–Fisher, Fundamental Plane, Cepheids, and SNe Ia, the observed distance modulus is expressed as

$\mu_{\rm obs} = \mu + \sigma_\mu \epsilon$

where $\epsilon \sim \mathcal{N}(0,1)$. The conversion to linear luminosity distance produces

$D = 10^{(\mu_{\rm obs} - 25)/5}$

rendering $D$ lognormally distributed at fixed true distance $d$:

$$P(D|d) = \frac{1}{\sqrt{2\pi} \tilde\sigma_\mu D}\; \exp\left[ -\frac{(\ln(D/d))^2}{2\tilde{\sigma}_\mu^2} \right],\quad\tilde{\sigma}_\mu \equiv \frac{\sigma_\mu}{5/\ln 10}$$

This causes the conditional mean and median of $D$ at fixed $d$ to differ,

$$\langle D | d\rangle = d\,e^{\tilde{\sigma}_\mu^2/2}, \quad \mathrm{Median}[D|d] = d$$

and, crucially, peculiar velocities $V = cz - H_0 D$ inherit the non-Gaussian skewness. This lognormal bias introduces spurious patterns (e.g., artificial flows) if left untreated (Hoffman et al., 2021).

2. BGc Algorithmic Framework

The BGc method exploits the invariance of the median under lognormal transformation. By recentering individual observed values relative to local medians and rescaling by the error parameters, it Gaussianizes the error distribution, thereby correcting the bias.

Workflow Steps

1. Data Grouping: Galaxies are grouped by redshift distance $\epsilon \sim \mathcal{N}(0,1)$0 to define homogeneous bins for local statistics. CF3 uses $\epsilon \sim \mathcal{N}(0,1)$1 Mpc/h to ensure error-dominated regime.
2. Local Median Computation: For each target, the $\epsilon \sim \mathcal{N}(0,1)$2 nearest neighbors in $\epsilon \sim \mathcal{N}(0,1)$3 are selected (typically $\epsilon \sim \mathcal{N}(0,1)$4–$\epsilon \sim \mathcal{N}(0,1)$5), and the medians $\epsilon \sim \mathcal{N}(0,1)$6 and $\epsilon \sim \mathcal{N}(0,1)$7 are computed.
3. Gaussianization Transform: The error variable is

$\epsilon \sim \mathcal{N}(0,1)$8

A Gaussianized estimator for distance:

$\epsilon \sim \mathcal{N}(0,1)$9

(setting $D = 10^{(\mu_{\rm obs} - 25)/5}$0). The velocity estimator:

$D = 10^{(\mu_{\rm obs} - 25)/5}$1

with $D = 10^{(\mu_{\rm obs} - 25)/5}$2. Special treatment is applied for targets outside the fiducial redshift range.

BGc thus produces locally unbiased and Gaussianized estimates for both distance and peculiar velocity at each spatial location (Hoffman et al., 2021).

3. Validation on Simulated Catalogs

BGc performance was assessed using mock CF3 catalogs constructed from MultiDark-2 $D = 10^{(\mu_{\rm obs} - 25)/5}$3CDM simulations. Simulated true distances and velocities were perturbed by CF3-like Gaussian errors, and the BGc pipeline was applied:

• Residuals: Raw distances and velocities showed large biases with means deviating from zero; after BGc, both means and medians of the residuals are $D = 10^{(\mu_{\rm obs} - 25)/5}$4 across $D = 10^{(\mu_{\rm obs} - 25)/5}$5–$D = 10^{(\mu_{\rm obs} - 25)/5}$6 Mpc/h.
• V-D Correlation: BGc symmetrized the $D = 10^{(\mu_{\rm obs} - 25)/5}$7 vs $D = 10^{(\mu_{\rm obs} - 25)/5}$8 scatter, eliminating spurious infall/outflow.
• Hubble Constant Recovery: Fits to $D = 10^{(\mu_{\rm obs} - 25)/5}$9 in $D$0 Mpc/h revealed
  • $D$1 km/s/Mpc
  • Error budget is dominated by cosmic variance

Averaging over $D$2 mock catalogs confirmed the statistical robustness of BGc for unbiased Hubble constant estimation (Hoffman et al., 2021).

4. Impact on Wiener-Filter Reconstruction

BGc enables optimal reconstruction of the large-scale velocity and density field using Wiener Filtering (WF). The WF estimator uses the BGc-corrected peculiar velocities as input, producing fields free from the systematic monopole/dipole errors induced by lognormal bias:

• WF Comparisons: Velocity maps from raw (biased) and BGc-corrected mocks showed that only the BGc/WF reconstruction recovered the full amplitude and structure of the simulated velocity field.
• Error Analysis: The statistical error from BGc is negligible compared to cosmic variance in WF reconstructions at large radii ($D$3 Mpc/h).

This correction is essential for unbiased cosmographical mapping from peculiar velocity surveys (Hoffman et al., 2021).

5. Application to Galaxy Clustering and Power Spectrum Estimation

In the context of galaxy clustering statistics, BGc is operationalized as a Gaussianization transform on the density field $D$4. By mapping the empirical cumulative distribution function $D$5 of overdensity values to a Gaussian via

$D$6

the one-point PDF is rendered Gaussian, removing local, monotonic bias and improving the agreement between galaxy, dark matter, and linear power spectra at quasi-linear scales.

• Power Spectrum Results: In real space, Gaussianized red and blue galaxy $D$7 agree with $D$8 within $D$9 to $d$0–$d$1 h/Mpc (see Table below).
• Redshift Space: Small-scale velocity dispersions (“fingers of god”) degrade the agreement, but after FoG correction, Gaussianized spectra again match the underlying matter spectrum up to $d$2 h/Mpc.
• Comparison with Clipping: Clipping offers similar $d$3-reach but requires ad-hoc threshold tuning, whereas Gaussianization is threshold-free but amplifies shot noise.

Method	$d$4 (P_Gal/P_DM$d$51)	Comments
BGc (real space)	$d$6–$d$7 h/Mpc	Red/blue tracers unified; 10% agreement
Clipping	$d$8–$d$9 h/Mpc	Threshold-dependent; better shot-noise control

BGc thus unifies two-point galaxy statistics and enables power spectrum analyses deeper into quasi-nonlinear regimes (1511.02034).

6. Practical Application and Limitations

Applied to CF3, BGc produces unbiased Hubble constant measurements:

$$P(D|d) = \frac{1}{\sqrt{2\pi} \tilde\sigma_\mu D}\; \exp\left[ -\frac{(\ln(D/d))^2}{2\tilde{\sigma}_\mu^2} \right],\quad\tilde{\sigma}_\mu \equiv \frac{\sigma_\mu}{5/\ln 10}$$0

in agreement with previous estimates but with explicit decomposition of statistical and cosmic variance.

Notably, BGc only corrects statistical lognormal bias due to measurement errors in distance moduli. Zero-point calibration systematics, small-scale velocity effects, and model-dependent assumptions (particularly in the clustering/statistics context) are not addressed. For joint analyses, BGc has been found consistent with more complex Bayesian MCMC techniques, with detailed cross-comparisons ongoing (Hoffman et al., 2021).

7. Implications and Extensions

BGc, by transforming biased (lognormal) observables to unbiased, Gaussianized forms via local median recentering and rescaling, enables high-fidelity cosmological inference from both distance/redshift surveys and galaxy clustering data. Its main strengths are parameter-free correction (except for error model inputs), local operation (no large-scale smoothing, no external priors), and demonstrated performance at the one-point and two-point statistical level.

A plausible implication is that BGc or related Gaussianization methods could be extended to a wider class of astronomical observables where lognormal observational bias and scale-dependent systematics hamper direct inference. However, control of additional systematic uncertainties and detailed treatment of non-local bias remain domains for further research (Hoffman et al., 2021, 1511.02034).