Think Outside the Color Box: Probabilistic Target Selection and the SDSS-XDQSO Quasar Targeting Catalog

TL;DR

The paper presents XDQSO, a probabilistic quasar-targeting method that uses extreme-deconvolution to model star and quasar flux-space densities, explicitly convolving with photometric uncertainties to produce robust quasar probabilities for SDSS sources. By binning in $i$-band magnitude and separating relative fluxes from total flux, the authors build a per-bin, multi-class density model that yields $P(\text{quasar})$ and $P(\text{star})$ for every object, enabling uniform and efficient targeting down to $i \approx 22$ mag. The resulting SDSS-XDQSO catalog contains 160,904,060 objects with class probabilities across low-, mid-, and high-redshift quasar classes, and demonstrates ~50% efficiency for mid-redshift quasar targeting in the BOSS CORE sample, outperforming several flux-based methods. The method is general, extensible to additional bands and variability, and accompanied by public code, representing a significant advance for constructing large, uniform quasar samples in current and future surveys such as Pan-STARRS and LSST.

Abstract

We present the SDSS-XDQSO quasar targeting catalog for efficient flux-based quasar target selection down to the faint limit of the Sloan Digital Sky Survey (SDSS) catalog, even at medium redshifts (2.5 <~ z <~ 3) where the stellar contamination is significant. We build models of the distributions of stars and quasars in flux space down to the flux limit by applying the extreme-deconvolution method to estimate the underlying density. We convolve this density with the flux uncertainties when evaluating the probability that an object is a quasar. This approach results in a targeting algorithm that is more principled, more efficient, and faster than other similar methods. We apply the algorithm to derive low-redshift (z < 2.2), medium-redshift (2.2 <= z <= 3.5), and high-redshift (z > 3.5) quasar probabilities for all 160,904,060 point sources with dereddened i-band magnitude between 17.75 and 22.45 mag in the 14,555 deg^2 of imaging from SDSS Data Release 8. The catalog can be used to define a uniformly selected and efficient low- or medium-redshift quasar survey, such as that needed for the SDSS-III's Baryon Oscillation Spectroscopic Survey project. We show that the XDQSO technique performs as well as the current best photometric quasar-selection technique at low redshift, and outperforms all other flux-based methods for selecting the medium-redshift quasars of our primary interest. We make code to reproduce the XDQSO quasar target selection publicly available.

Figures (15)

• Figure 1: Number counts $p(f_i | i \in \mathrm{class})$ for the various target classes. These have the expected property that higher redshift quasars are fainter since they are more distant.
• Figure 2: Flux-flux diagrams for a bin in $i$-band magnitude from the single-epoch 'star' catalog. The first column shows the single-epoch data, the second column shows a sampling from the extreme-deconvolution fit to the single-epoch relative fluxes with the errors from the co-added data added, and the third column shows the co-added data. The relevant goodness-of-fit comparison is between the second and third column. The grayscale is linear in the density and the contours contain 68, 95, and 99 percent of the distribution. Samples falling outside the outermost contour are individually shown. This shows that the XD technique recovers the underlying error-deconvolved distribution given noisy training data.
• Figure 3: Same as \ref{['fig:singlezexfit']} but the color-color diagrams.
• Figure 4: Flux-flux and color-color diagrams for a bin in $i$-band magnitude from the quasar ($2.2 \leq z \leq 3.5$) catalog. The first column shows a sampling from the extreme-deconvolution fit with the errors from the quasar data added and the second column shows the quasar data resampled according to the quasar luminosity function as described in Section \ref{['sec:qsodata']}. The third and fourth columns show the same as the first and second columns, but for colors.
• Figure 5: Flux-flux and color-color diagrams for a bin in $i$-band magnitude from the stellar training catalog of co-added, non-variable SDSS Stripe-82 point-source data. The first column shows a sampling from the extreme-deconvolution fit with the errors from the stellar data added and the second column shows the stellar training data. The third and fourth columns show the same as the first and second columns, but for colors.