Abstract
The Baryon Oscillation Spectroscopic Survey (BOSS), part of SDSS-III, aims to map the expansion history of the universe by observing the "Lyman-alpha forest" in the spectra of 150,000 high-redshift quasars. This paper details the complex statistical and machine-learning machinery—including Neural Networks and Extreme Deconvolution—used to identify these rare targets from millions of similar-looking stars in imaging data.
The Challenge: Finding Needles in a Cosmic Haystack
To measure Baryon Acoustic Oscillations (BAO) at high redshifts ($z > 2$), astronomers require a high surface density of quasars to act as "backlights." As their light travels toward Earth, it passes through intergalactic hydrogen clouds, creating a "forest" of absorption lines known as the Lyman-alpha forest. This forest provides a 3D map of the large-scale structure of the early universe.
The primary challenge is one of selection. Quasars are significantly rarer than stars. At a redshift of $z \approx 2.7$, the colors of quasars in standard optical filters ($ugriz$) overlap almost perfectly with common A-type stars and blue horizontal branch stars. To succeed, the BOSS team needed to develop a target selection algorithm that was both efficient (minimizing wasted spectroscopic fibers on stars) and complete (ensuring enough quasars for a dense map).
Advanced Selection Methods
The authors moved beyond traditional color-cut boxes, implementing sophisticated probabilistic and machine-learning approaches to maximize yield:
- KDE (Kernel Density Estimation): A non-parametric method estimating the density of stars and quasars in multi-dimensional color space to assign probabilities to candidates.
- Likelihood Methods: Utilizing $\chi^2$-like statistics to compare candidate fluxes against training sets of known objects.
- Artificial Neural Networks (ANN): A multi-layer perceptron that classifies objects based on magnitudes and associated errors.
- XDQSO (Extreme Deconvolution): Serving as the "CORE" algorithm, XDQSO models the underlying distribution of objects as a sum of Gaussians while explicitly accounting for heteroscedastic measurement errors. This is particularly effective for faint targets near the detection limit.
- NN-Combinator: A "meta-classifier" that integrates outputs from various algorithms to produce a final prioritized target list.
Key Insights: Breaking the $z=2.7$ Barrier
A critical finding of the paper is the difficulty of identifying quasars at $z \approx 2.7$. These scientifically valuable objects reside exactly on the "stellar locus" in color space. Furthermore, standard color cuts fail at faint magnitudes because photometric noise "smears" the distributions, making stars look like quasars and vice versa.
To manage this, the survey split targets into two categories:
- CORE Sample: Selected via a single, fixed algorithm (XDQSO) to ensure a uniform sample for statistical studies of quasar evolution.
- BONUS Sample: Utilized every available technique, including multi-wavelength data from GALEX (UV), UKIDSS (Near-IR), and FIRST (Radio), to maximize the density of backlights for BAO measurements.
Results and Future Directions
In its first year alone, BOSS confirmed 11,263 new quasars at $z > 2.2$, nearly doubling the global sample known at the time. The algorithms achieved a surface density of approximately 15 high-$z$ quasars per square degree from a budget of 40 targets per square degree, meeting the technical requirements for the BAO experiment. XDQSO proved superior, selecting roughly one additional high-$z$ quasar per square degree compared to previous methods.
Despite these advances, limitations remain. Approximately 25-30% of targets are still stars, especially at the faint limit ($g \approx 22$), and a completeness "dip" persists around $z \approx 2.7$. Future work may incorporate variability data—measuring how an object's brightness changes over time—to further distinguish flickering quasars from steady-burning stars.
Summary updated