Mixture models for photometric redshifts

Determining photometric redshifts to high accuracy is paramount measure distances in wide-field cosmological experiments. With only information at hand, photo-zs are prone systematic uncertainties the intervening extinction and unknown underlying spectral-energy distribution of different astrophysical sources. Here, we aim resolve these model degeneracies obtain a clear separation between intrinsic physical properties sources extrinsic systematics. We estimates full photo-z probability distributions, their uncertainties. perform probabilistic determination using Mixture Density Networks (MDN). The training data-set composed optical ($griz$) point-spread-function magnitudes measurements from SDSS-DR15, WISE midinfrared ($3.4 \mu$m $4.6 \mu$m) magnitudes. use Infinite Gaussian models classify objects our as stars, galaxies or quasars, determine number MDN components achieve optimal performance. fraction that correctly split into main classes 94%. Our method improves bias redshift estimation (i.e. mean $\Delta z$ = (zp - zs)/(1 + zs)) by one order magnitude compared SDSS photo-z, decreases $3 \sigma$ outliers 3rms$(\Delta z) < \Delta z$). relative, root-mean-square uncertainty resulting down 1.7% for low-redshift (zs $<$ 0.5). have demonstrated feasibility machine-learning based methods produce distributions with performance competitive state-of-the art techniques. can be applied surveys where vary significantly across sky sparse spectroscopic calibration samples.