Treating observational systematics for LSS in photometric surveys

Treating observational systematics for LSS in photometric surveys X Meeting on Fundamental Cosmology Seville, 18-10-2024 Martín Rodríguez Monroy IFT, Madrid, Spain The Dark Energy Survey 2 DES survey evolution 3 Sevilla-Noarbe et al. 2021 Rodríguez-Monroy et al. 2021 DES-Y3 results: galaxy clustering 4 DES Collaboration et al. 2021 DES-Y3 cosmology and Y6 BAO results 5 DES Collaboration: Abbott et al., 2024 Stage IV surveys: LSST and Euclid 6 Bianco et al., 2021 Stage IV surveys: LSST and Euclid Using the current baseline cadence, LSST ten-year survey will take more than five million exposures, collecting over 50 petabytes of raw image data to produce a deep, time-dependent, movie of about 20,000 square degrees of sky. 7 Euclid Wide Survey (EWS) coverage and colour-coded yearly progress. Blue borders = 16000 deg2 region of interest that contains the 13416 deg2 observed sky of the EWS. Euclid. I. Overview of the Euclid mission Image credit: Elisa Chisari and Nacho Sevilla 8 LSST and Euclid: cosmology forecasts Euclid. I. Overview of the Euclid mission LSST-DESC Euclid Pushing the limits of statistical error! 9 LSST and Euclid: cosmology forecasts Euclid. I. Overview of the Euclid mission LSST-DESC Euclid Pushing the limits of statistical error! Image credit: Elisa Chisari and Nacho Sevilla 16 ● Stacking of images ⇒ need summary statistic (weighted mean, min., max., variance….) ● Several photometric bands ○ DES: griz ○ LSST: ugrizY ○ Euclid: YE, JE, HE Pearson’s coeff. ● Several astrophysical foregrounds: stars, dust extinction, HI…. ● Many maps are (highly) correlated → dimensionality reduction: ○ PCA ○ SOMs ● Risks = under / overcorrection ●Must use data-driven selections Observational systematics: template selection 16 DES-Y3 contamination templates (i-band) Mitigation methods for observational systematics 17 18 Modeling of observational systematic contamination: δg obs = f(s) + δg true , so δg true is the residual from the fit ●What is the form of f(s)? ○ 1-dimensional regression ■ Iterative process ○ Multi-dimensional regression ○ More complex method δg syst 1 e.g. PSF syst 2 e.g. depth Healpix pixels s = template of contamination Observational systematics: modeling 19 Modeling of observational systematic contamination: δg obs = f(s) + δg true , so δg true is the residual from the fit δg syst 1 e.g. PSF syst 2 e.g. depth Healpix pixels s = template of contamination 19 Observational systematics: modeling True density field Net systematic contaminant Observed density field Set of systematic maps Weaverdyck & Huterer 2007.14499 ⨉ δg obs = f(s) + δg true δg true Preliminary cleaning: ● Identify and mask out extreme regions (pixels) Use available SP maps to model δg obs and mitigate contamination. How? ● Act at the observable / estimator level (e.g. w(θ)) ● Modify randoms used by estimator (e.g. Landy-Szalay) ●Act at the map level (e.g. δg field) → weight map Observed on the ground Back to truth 20 Correcting for observational systematics DES correction methods: ENet Elastic Net (ENet) regularization = LASSO + ridge regressions (Zou & Hastie 2005) ● Multilinear fit in N-dim space (N SP maps) ●All SP maps are associated a contamination amplitude, αi ● Avoid over correction: ○ ○LASSO (L1) term: penalises non-zero αi ⇒ favours scarcity of explanatory variables ○ ○Ridge (L2) term: penalises correlated variables ○ ● ENet ⇒ minimize loss function: ● Estimate λ1 and λ2 with cross-validation on data subsets LASSO Ridge OLS 21 DES correction methods: ISD Apply the weight map to our galaxy sample 23 Identify most significant SP map 4 Iterative systematics decontamination (ISD) in a nutshell: ●Fix a threshold for 1D contamination ● Iterative process: ○5 Re-evaluate significance of SPs until process converges ○1 Define 1D significance by evaluating against log-normal mocks 22 23 ●SOM systematics ○ Reduces dimensionality ○ Considers all maps at once ○ Non-linear parameterization ●Neural net(s) ○ First steps in the use of AI for systematics ○ Problem reframing for exploiting AI capabilities → solutions can vary from neural networks to evolutionary algorithms ●Alternative regularisation method ●Random-level method ○ Recovering and updating old method based on depth (DARTH-systematics) Ongoing work at IFT: new methods for LSST-DESC / Euclid on DES Methods validation Validation on simulations and data: ●Methods performance / configuration ●Completeness of SP map set ●Methodological blind spots ●Systematic contribution to covariance: ○Over / undercorrection ○Difference between methods ●Impact on cosmology ●Cross-correlation with external tracers Rodríguez-Monroy, Weaverdyck+. 2105.13540 24 Summary 25 Methods validation Validation on simulations and data: ●Methods performance / configuration ●Completeness of SP map set ●Methodological blind spots Rodríguez-Monroy, Weaverdyck+. 2105.13540 ●Different method assumptions → different corrections 32 Additional robustness tests Similar results for MagLim 33 Correlations with LSS 34 Correlations with LSS 35