BOSS Quasar Target Selection

The paper by Ross et al. 2012 describes the BOSS quasar target selection in detail. Here we summarize some of the main points from that paper.

The principal scientific goal of the BOSS quasar sample is to map the large-scale structure traced by the Ly-α forest. This enters the wavelength coverage of the BOSS spectrographs for redshifts greater than 2.15, thus (unlike quasar target selection in SDSS-I/II), the BOSS quasar sample is designed to be sensitive to quasars only for 2.15 < z < 3.5. This is challenging, because the quasar locus crosses that of main sequence stars at z ~ 2.7. Moreover, while SDSS-I/II targeting was limited to objects with i < 19.1 (for UV-excess objects) and i < 20.2, BOSS quasar targeting pushes much fainter, g < 22.0, or r < 21.85, to give a significantly higher surface density of targets than in SDSS; this is close enough to the magnitude limit of SDSS photometry that the photometric errors significantly broaden the stellar locus. The SDSS-I/II quasar target selection algorithm defined the stellar locus as a one-dimensional structure in SDSS color-color space; crudely speaking, quasar targets were defined to be those objects sufficiently far from this locus. In BOSS, we took a different approach, and used a number of algorithms to identify candidates:

  • Kernel Density Estimation (KDE; Richards et al. 2009, ApJS, 180, 67) assigns each pixel in multi-dimensional color-color space an expected density of stars and quasars, allowing a probability that any given object is a quasar to be estimated.
  • Likelihoods (Kirkpatrick et al. 2011, ApJ, 743, 125) that a given object is a quasar are calculated by summing its Gaussian distance from each object in a training set of known quasars and stars in color space.
  • A Neural Network (Yèche et al. 2010, A&A, 523, 14) is used to compare the five-band photometry and errors of any given object with that of a large training set of quasars and stars to estimate a probability that a given object is a high-redshift quasar.
  • Extreme Deconvolution, or XDQSO (Bovy et al. 2011, ApJ, 729, 141), describes the stellar and quasar loci as a sum of Gaussians convolved with photometric measurement errors; this allows the likelihood that any given object is a z > 2.2 quasars to be calculated.
  • Radio selection: sufficiently red objects (u-g > 0.4) matched to sources in the FIRST radio survey were targeted.
  • Previously known quasars (including those discovered by SDSS I/II) with z > 2.15 are targeted by BOSS, to get high S/N observations of the Ly-α forest.

Note that many objects are selected by more than one of these algorithms! To meet our science goals, we must obtain spectra for at least 15 z > 2.2 quasars/deg2. For Ly-α forest studies, we do not need a sample of quasars with well-defined selection function, but for many other quasar science goals (such as clustering and luminosity function measurements), we do need a uniformly selected sample. We therefore define a CORE sample of 20 targets per square degree (of which about half are true quasars upon inspection), identified with the XDQSO algorithm, which is held fixed throughout the BOSS survey. (Note that the algorithm associated with CORE varied through the first year of the survey, as described in Ross et al.). We identify another 20 targets per square degree, from combining the results of all other target selection algorithms, and refer to these as the BONUS sample. This is not required to be uniformly selected, and includes additional information (e.g., from near-infrared photometry from the UKIDSS Large Area Survey) where available.