Expositor: Mariano Dominguez (IATE-OAC-UNC and FRC-ULS)
Fecha: viernes 18 de septiembre, 11:30 hs.
Resumen: Stage-IV dark energy experiments—such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), Euclid, and the Roman Space Telescope—require unprecedented control over spatial completeness and photometric redshift p(z) uncertainties. Traditional analysis pipelines that compress galaxy positions into scalar point estimates lose critical probabilistic information, diluting spatial clustering measurements and introducing systematic biases. In this talk, we present an end-to-end computational framework designed to manage petabyte-scale astronomical catalogs, construct survey-scale spatial selection functions, estimate robust photometric redshift distributions, and measure PDF-aware spatial correlation functions.
First, we introduce HATS (Hierarchical Adaptive Tiling Scheme) and LSDB, which provide adaptive HEALPix spatial partitioning and a Dask-parallelized query engine for executing cross-matches and time-series analytics across multi-terabyte datasets. Building upon this schema, Skykatana constructs survey-scale boolean and bit-packed HEALSparse spatial masks to excise bright-star halos, diffraction spikes, and observational artifacts while operating under tight memory constraints. We further highlight its floatmask extension, which replaces rigid binary exclusions with continuous probabilistic weights to maximize retained sky area for small-scale clustering.
Second, to estimate photometric redshifts across the 16 observational configurations of the LSST-DESC Joint Challenge, we present Pontifex. Pontifex combines physical template likelihoods (BPZ, LePHARE) and machine learning regressors (FlexZBoost, GPz, PZFlow) via a Mixture-of-Experts (MoE) gating engine optimized through Particle Swarm Optimization (PSO). Prior to inference, FeatureGuard sanitizes corrupted flux variances and extreme outliers. To eliminate catastrophic Balmer/Lyman break degeneracies in crowded or blended fields, Pontifex selectively engages the NUGUNDAM module, which executes an iterative spatial Expectation-Maximization (EM) calibration against reference spectroscopic samples to damp unphysical PDF modes.
Finally, we detail νGundam, a high-performance C++/Fortran and Python framework designed to compute two-point spatial correlation functions $w_p(r_p, \pi)$ directly from full per-object redshift probability density functions. We evaluate four interchangeable pair-counting modes: exact empirical-PDF integration (ePDF), Gaussian Mixture Model compression (GMM), Monte Carlo sampling (MC), and quantile node compression (16quant). We demonstrate that radial uncertainties primarily redistribute clustering signal along the line of sight ($\pi$) and across projected bins ($r_p$). While exact grid integration provides an error-free benchmark, 16quant (1.3% RMS residual, ~28× speedup) and MC sampling (3.6% RMS residual, ~98× speedup) accurately recover truth-redshift correlation shapes while dramatically accelerating pair-counting workflows.
Together, this unified toolset provides a scalable, mathematically rigorous foundation for precision large-scale structure analyses in upcoming wide-field surveys.
