Validating the Galaxy and Quasar Catalog-Level Blinding Scheme for the DESI 2024 analysis

TL;DR

This work addresses the risk of experimenter bias in precision cosmology by implementing and validating a catalog-level blinding scheme for DESI DR1 focused on BAO, RSD, and PNG observables. The authors develop AP-like and RSD blinding shifts in redshift space and apply scale-dependent weights to mimic PNG, constraining blinding parameters to prevent spurious distortions while preserving testability. Validation on 25 AbacusSummit mocks and blinded real data shows that the blinding introduces controlled shifts consistent with the underlying cosmology and that BAO and ShapeFit analyses remain robust under blinding, with posteriors recoverable after unblinding. The methodology, including a second layer of blinding for sanity checks, demonstrates that DR1 analyses can proceed unbiasedly, setting a precedent for future blind analyses in DESI and other surveys. These results enhance the credibility and reliability of cosmological inferences drawn from large-scale structure data by ensuring that analysis decisions are not inadvertently guided by prior expectations.

Abstract

In the era of precision cosmology, ensuring the integrity of data analysis through blinding techniques is paramount -- a challenge particularly relevant for the Dark Energy Spectroscopic Instrument (DESI). DESI represents a monumental effort to map the cosmic web, with the goal to measure the redshifts of tens of millions of galaxies and quasars. Given the data volume and the impact of the findings, the potential for confirmation bias poses a significant challenge. To address this, we implement and validate a comprehensive blind analysis strategy for DESI Data Release 1 (DR1), tailored to the specific observables DESI is most sensitive to: Baryonic Acoustic Oscillations (BAO), Redshift-Space Distortion (RSD) and primordial non-Gaussianities (PNG). We carry out the blinding at the catalog level, implementing shifts in the redshifts of the observed galaxies to blind for BAO and RSD signals and weights to blind for PNG through a scale-dependent bias. We validate the blinding technique on mocks, as well as on data by applying a second blinding layer to perform a battery of sanity checks. We find that the blinding strategy alters the data vector in a controlled way such that the BAO and RSD analysis choices do not need any modification before and after unblinding. The successful validation of the blinding strategy paves the way for the unblinded DESI DR1 analysis, alongside future blind analyses with DESI and other surveys.

Figures (19)

• Figure 1: Parameter space of interest for $(w_0,w_a)$ under the DESI DR1 blinding scheme. The white region represents the parameter region that allows for changes in $\alpha_\parallel$ and $\alpha_\perp$ of less than 3% with respect to a fiducial chosen value of 1 in the redshift range $0.4 < z < 2.1$. The black points are 8 random selections used to blind our mock catalogs, which we use to validate our methodology.
• Figure 2: Comparison of blinded and unblinded mocks for multipoles $\ell = 0$, $\ell = 2$, and $\ell = 4$, for the correlation function (left column) and power spectrum (right column). The curves show the mean is across 25 AbacusSummit catalogs which are blinded with the same blinding parameters.
• Figure 3: Pre-reconstruction anisotropic BAO fits using the correlation function (top) and the power spectrum (bottom) for LRG samples for the first redshift bin (each row) from 16 different blinded mock catalogs with $(w_0,w_a)$ choices identified by indices 1-8 and two $f_\mathrm{{NL}}$ values by blue and orange, respectively. The top two subplots in each panel plot $\Gamma_i$, defined as the ratio of measured vs expected ratios of the $i$th parameter from each sim vs a reference sim (identified with black marker-edge); here $i$ = $\alpha_{\mathrm{iso}}$, $\alpha_{\mathrm{AP}}$, where measured values are from the analysis pipeline while expected ones are from the theoretical connection with the respective $(w_0,w_a)$; error bars capture the measurement uncertainties while propagating the errors. This statistic allows comparing all the sims against a reference sim. The bottom subplot in each panel displays the reduced $\chi^2$ values, with shaded areas representing $1\sigma$ and $2\sigma$ regions; the $\sigma$-limits are obtained as the standard deviation from the mean of the $\chi^2$ distribution of the 16 $\chi^2$ values. This confirms the consistency and reliability of BAO measurements under various blinding shifts given the small variations.
• Figure 4: Post-reconstruction anisotropic BAO fits for LRG samples for the first redshift bin (each row) following the structure in \ref{['fig:pre_recon_bao_fits_LRG']}. Here, too, we see that while our $\Gamma$ statistic varies around the expected value of unity, the reduced $\chi^2$ indicates good fits.
• Figure 5: ShapeFit fits using LRG samples for the first redshift bin (each column) from 16 different blinded mock catalogs. Various details here are the same as in \ref{['fig:pre_recon_bao_fits_LRG']}, except that $i$ = $\alpha_{\mathrm{iso}}$, $\alpha_{\mathrm{AP}}$, $df$, $m$ in $\Gamma_i$ while $\tilde{\Gamma}_i$ is the same as $\Gamma_i$ but comparing differences as opposed to ratios between measured and expected (since expected is 0). As for BAO fits, we see that the ratios (differences) are close to 1 (0) and the $\chi^2$ variations are within 1-2$\sigma$, demonstrating the robustness of the fits.