THE SDSS-III BARYON OSCILLATION SPECTROSCOPIC SURVEY: QUASAR TARGET SELECTION FOR DATA RELEASE NINE

The current cosmic microwave background (CMB) data are in excellent agreement with the theoretical predictions of a flat cosmological model with cold dark matter which is dominated by dark energy with an equation of state parameter w = −1 (ΛCDM; Komatsu et al. 2011; Larson et al. 2011). Acoustic peaks in the CMB anisotropy power spectrum are generated by cosmological perturbations exciting sound waves in the relativistic plasma of the early universe (Sunyaev & Zeldovich 1970; Peebles & Yu 1970; Bond & Efstathiou 1984, 1987; Holtzman 1989; Meiksin et al. 1999). The scale of these peaks, which is set by the sound horizon at last scattering (Eisenstein & Hu 1998; Blake & Glazebrook 2003; Seo & Eisenstein 2003), can be used as a cosmological standard ruler. These baryon acoustic oscillations (BAOs) are present in the distribution of matter at late times as well and were first measured in the large-scale distribution of galaxies by Eisenstein et al. (2005) and Cole et al. (2005).

BAO should also be present in the distribution of neutral hydrogen gas in the intergalactic medium, and thus should be observable in the Lyα forest (LyαF) absorption spectra of distant quasars (White 2003; McDonald & Eisenstein 2007; Slosar et al. 2009; Norman et al. 2009; Barenboim et al. 2010; White et al. 2010; McQuinn & White 2011). Measurements of BAO in the LyαF would provide the first measurements of cosmic expansion and the angular diameter distance at redshift z > 2 (other than the CMB itself), a regime not constrained by current data, thus giving important constraints on, and tests of, the standard cosmological model.

The Sloan Digital Sky Survey (York et al. 2000) is now in its third phase (SDSS-III; Eisenstein et al. 2011) and is carrying out a combination of four interleaved surveys that will continue until the summer of 2014. One of those surveys, the Baryon Oscillation Spectroscopic Survey (BOSS²⁷), commenced operations in late 2009 and is using essentially all the dark time for SDSS-III. The key goal of BOSS is to measure the absolute cosmic distance scale and expansion rate to an accuracy of a few percent from the signature of BAO in the distribution of galaxies and neutral hydrogen (Schlegel et al. 2007, 2009). This will be achieved by measuring spectroscopic redshifts for ≈1.5 million luminous red galaxies and, simultaneously, the LyαF toward ≈150,000 high-redshift quasars.²⁸ Both samples aim to constrain the equation of state of dark energy by measuring the angular diameter distance, d_A, and the Hubble parameter, H(z), at z = 0.3, 0.6, and ∼2.5. In addition to the cosmology goals, the unprecedented data set of z  ∼ 2.5 quasars will enable tests of black hole growth, wind, and feedback models and provide insights into the links between galaxy formation, evolution, and luminous active galactic nucleus (AGN) activity. Using data from the original SDSS quasar survey will also allow studies of spectroscopic variability. BOSS uses the same 2.5 m Sloan Foundation telescope (Gunn et al. 2006) that was used in SDSS-I/II, but since BOSS will observe fainter targets, the fiber-fed spectrographs have been significantly upgraded. These upgrades include: new CCDs with improved blue and red response; 1000 2'' instead of 640 3'' optical diameter fibers; higher throughput gratings over a spectral range of 3600–10000 Å at a resolution of about 2000, and improved optics.

Quasars have colors distinct from those of the much more numerous stars in the five-color photometry of the SDSS (Fan 1999). Unobscured quasars have very blue continua, without any breaks redward of the Lyα emission line, and so can be distinguished from hot stars which have a strong Balmer break in the u − g, g − r color–color diagram (Figure 1). In particular, at z < 2.2, quasars have a UV excess (as measured by u − g) that distinguishes them from most stars and they lie well away from the stellar locus at most higher redshifts (but see below). SDSS-I/II targeted quasars for spectroscopy (Richards et al. 2002) by selecting point sources which lie far from the locus of stars in color–color space (and all extended sources with a strong UV excess), as well as point sources with radio emission from the Faint Radio Sources at Twenty cm (FIRST) survey (Becker et al. 1995). The majority of the more than 100,000 quasars spectroscopically observed by SDSS (Schneider et al. 2010) were targeted in this way.

Zoom In Zoom Out Reset image size

The LyαF enters the sensitive range of the BOSS spectrographs at z > 2.2, and the number density of quasars falls dramatically at z > 3 (Osmer 1982; Schmidt et al. 1995; Richards et al. 2006), so BOSS quasar target selection (QTS) is designed to focus on the range 2.2 < z < 3.5. However, at z ∼2.7, SDSS quasar colors are very similar to those of A stars and blue horizontal branch (BHB) stars (Fan 1999), thus the optimal quasars for studying the LyαF are the most difficult ones for BOSS to target. Indeed, the SDSS-I/II QTS algorithm deliberately sparse-sampled objects in the region of color space where z = 2.7 quasars should lie (Richards et al. 2002), in an attempt to minimize the contamination by stars.

The BOSS survey requirements are for spectroscopy of 15 or more z > 2.2 quasars deg⁻² (150,000 quasars over the BOSS footprint of 10,000 deg²; Eisenstein et al. 2011). Combining calculations from McDonald & Eisenstein (2007) and McQuinn & White (2011) with the luminosity function given by Jiang et al. (2006), we find that targeting to a magnitude of g < 22 with perfect completeness will provide a surface density of 20 z > 2.2 quasars deg⁻². This magnitude limit is approaching the detection limit of SDSS photometry (Abazajian et al. 2004), meaning that photometric errors will significantly broaden the stellar locus (Figure 1) and star–galaxy separation will be a factor. Contamination at both the bright and the faint end of the BOSS target range is mainly from metal-poor halo A and F stars, faint lower redshift (z  ∼ 0.8) quasars, and compact galaxies. To put these requirements into perspective, the final quasar catalog from the original SDSS-I/II quasar survey (Schneider et al. 2010) contained 17,582 z > 2.2 objects over 9380 deg², while the 2dF-SDSS LRG And QSO (2SLAQ) survey (Croom et al. 2009), which observed to g < 21.85 and concentrated on UV-excess objects, contained 1110 such quasars selected over 192 deg². The original 2dF QSO redshift survey (2QZ; Croom et al. 2004) focused on the redshift range z < 2.1.

These challenges required a new approach to QTS. The first year of the BOSS survey ("Year One;" 2009 September through 2010 July) was devoted in part to refining our algorithms for selecting these objects. The resulting sample of quasars at z > 2.2 is comparable in size to the SDSS high-redshift quasar sample and of course reaches much fainter magnitudes with much higher surface density. Thus the new sample itself represents the best test of our selection algorithms, and we modified those algorithms multiple times through the year. Year One included roughly three months of commissioning of the upgraded BOSS spectrographs and instrument control software as well as a steady ramp-up to full efficiency operations, so it includes well under 20% of the anticipated final sample for the five-year BOSS survey. As of 2011 April, BOSS is on track to complete its intended 10,000 deg² of spectroscopic survey area assuming historical weather patterns and continuation of the current observing efficiency.

Motivated by the first science investigations based on Year One data (e.g., Slosar et al. 2011), this paper presents the methods and performance of the QTS during this year. In what follows,"Year Two" will refer both to the spectroscopic observations carried out during BOSS' second year, 2010 August to 2011 July, and the results of the QTS presented in this paper over the entire 10,000 deg² BOSS footprint; the distinction should be clear from context. Data from spectroscopic observations in Years One and Two will be included in SDSS Data Release Nine (DR9²⁹). The final SDSS-III QTS algorithm will appear in a separate paper.

Background quasars have no causal influence on structure in the LyαF at the BAO scale.³⁰ Hence, the sample of quasars we use for LyαF cosmological studies may be quite heterogeneous, with the only consequence that the window function of the survey will depend on the distribution of the quasars for which we have spectra. Since the precision of the BAO measurement improves rapidly with the surface density of quasars (at fixed spectroscopic signal-to-noise ratio (S/N)), we have implemented a target selection scheme in BOSS that can maximize the number of quasars found at z > 2.2 in any area of the sky, taking advantage of any available information (e.g., auxiliary data). In Year One, we explored a variety of methods, settling on our final target selection algorithms late in the year.

At the same time, in order to use the quasars themselves for statistical studies (such as luminosity functions or clustering analyses), we must also produce a uniformly selected sample, which we refer to hereafter as CORE (Section 3.1). However, we changed the definition of the CORE sample several times over Year One, as we tested various algorithms. Therefore, our fully uniform quasar sample will not include data from this first year of the survey. However, statistical studies (luminosity functions, clustering, and so forth) can utilize all five years of BOSS data by including moderate incompleteness corrections for Year One selection relative to the final CORE algorithm (see Section 6). We describe the evolution of our algorithms in detail in this paper, concluding with a description of the method we finally adopted. We give the target selection for both Years One and Two, and thus for the DR9, and analyze our performance from spectra obtained in Year One. By the end of Year Two, QTS had been performed over the whole 10,000 deg² BOSS footprint. Data from Year One were gathered over 880 deg²; see Table 1.

This paper is organized as follows. In Section 2, we describe the SDSS photometry on which the target selection algorithms are most heavily based. Section 3 describes our methods for selecting quasars (Richards et al. 2009b; Yèche et al. 2010; Kirkpatrick et al. 2011; Bovy et al. 2011a). These four papers suggest different, but complementary, methods, and we have used a union of these techniques in different combinations through the survey. In Section 4, we describe the implementation of these targeting methods through the first year. In Section 5, we report on the global properties of the resulting sample, including high-z quasar targeting efficiency, from the data gathered during the first year of BOSS, and we compare the effectiveness of the various methods. In Section 6, we discuss the production of a statistical quasar sample. We conclude in Section 7 and suggest improvements to BOSS QTS for the remainder (Years Three, Four, and Five) of the survey. Appendix A tabulates the logical cuts used on the input imaging data. Appendix B gives more detail about Year One target selection, while Appendix C describes a pre-BOSS pilot survey using the MMT. Appendix D characterizes the redshift completeness of our spectroscopic data.

We assume a cosmological model throughout with Ω_b = 0.046, Ω_m = 0.228, Ω_Λ = 0.725 (Komatsu et al. 2011). All optical magnitudes are quoted in, and based upon, the SDSS approximation to the AB zero-point system (Oke & Gunn 1983; Adelman-McCarthy et al. 2006), while all near-infrared (NIR) magnitudes are based on the Vega system. Throughout the paper, "magnitude" refers to SDSS point-spread function (PSF) magnitudes (Stoughton et al. 2002).

BOSS uses the same imaging data as that of the original SDSS-I/II survey, with an extension in the South Galactic Cap (SGC). These data were gathered using a dedicated 2.5 m wide-field telescope (Gunn et al. 2006) to collect light for a camera with 30 2k × 2k CCDs (Gunn et al. 1998) over five broad bands—ugriz (Fukugita et al. 1996); this camera has imaged 14,555 unique deg² of the sky, including 7500 deg² in the North Galactic Cap (NGC) and 3100 deg² in the SGC (Aihara et al. 2011). The imaging data were taken on dark photometric nights of good seeing (Hogg et al. 2001), and objects were detected and their properties were measured (Lupton et al. 2001; Stoughton et al. 2002) and calibrated photometrically (Smith et al. 2002; Ivezić et al. 2004; Tucker et al. 2006; Padmanabhan et al. 2008) and astrometrically (Pier et al. 2003).

Padmanabhan et al. (2008) present an algorithm which uses overlaps between SDSS imaging scans to photometrically calibrate the SDSS imaging data. BOSS target selection uses data calibrated using this algorithm from the SDSS Data Release Eight (DR8) database (Section 3.3; Aihara et al. 2011). The 25 wide stripe along the celestial equator in the Southern Galactic Cap, commonly referred to as "Stripe 82," was imaged multiple times, with up to 80 epochs at each point along the stripe spanning a 10 year baseline (Abazajian et al. 2009). In Section 4, we will discuss how the commissioning phase of BOSS used co-added catalogs in SDSS Stripe 82, generated by averaging the photometric measurements from ∼20 individual repeat scans; the details are discussed in Appendix A.6 and in Kirkpatrick et al. (2011).

Roughly 50% of the SDSS footprint has been imaged more than once (Aihara et al. 2011); combining the photometric measurements in these overlap regions reduces the flux errors.

Using the imaging data, BOSS quasar target candidates are selected for spectroscopic observation based on their PSF fluxes and colors in SDSS bands. Fluxes that are used for QTS are corrected for Galactic dust extinction according to the maps of Schlegel et al. (1998). All objects classified as point-like (OBJC_TYPE = 6) and are brighter than g = 22 or r = 21.85 are passed to the various QTS algorithms. The joint magnitude limit was imposed due to concerns of the LyαF moving into the g band at z ≈ 2.3 resulting in suppressed flux at redshifts greater than this. In practice, almost all our targets satisfy both these conditions. Throughout this paper, magnitudes use the asinh scale at low flux levels, as described by Lupton et al. (1999).

During processing of the imaging data by the SDSS photometric pipeline, a number of photometric flags are set for each detected object (Stoughton et al. 2002). These are generated by the SDSS photometric pipeline (Lupton et al. 2001), the Resolve algorithm (Aihara et al. 2011), and by photometric calibration (Padmanabhan et al. 2008). Some of these flags indicate problems with the de-blending of close pairs of objects. Other flags are set due to poor or unreliable photometry, e.g., if an object was saturated due to a bright star's diffraction spike or an object was too close to the edge of a frame. If these flags are ignored, they can lead to artifacts in the imaging data being selected as quasar targets. Details of these flags are given in Stoughton et al. (2002) and have been updated in DR8 (Aihara et al. 2011).

There are four distinct sets of quasars targeted by BOSS: targets selected by a uniform method, targets selected in a non-uniform way, matches to previously known z > 2.2 quasars, and matches to objects in the FIRST survey. We refer to these subsets of targets in this paper as CORE, BONUS, KNOWN, and FIRST, respectively.

Each of our targeting algorithms has different imaging flag cuts, as well as different flux limits imposed. We refer to these criteria collectively as "logic cuts." All such cuts are applied using single-epoch data with one exception: color cuts made on FIRST targets use co-added, multi-epoch data wherever these are available. FIRST objects are thus not considered to be part of the CORE statistical sample, unless they independently meet the CORE selection criteria. The logic cuts are described in detail in Appendix A.

The methods, data, and logic flag cuts for BOSS QTS are summarized in Figure 2. During Year One, we carried out QTS and designed spectroscopic plates on areas of ∼100–300 deg² at a time. We refer to these areas, within which all the algorithms used in QTS are uniform, as "chunks." Once QTS was more settled in Year Two, the areas of chunks could be, and sometimes were, more than 1000 deg². For guidance in the following discussion, Chunks 1–9, inclusive, constitute Year One, and Chunks 10–19, Year Two. Stripe 82 was targeted twice with different targeting algorithms: once in Year One (Chunk 1) and once in Year Two (Chunk 11).

Zoom In Zoom Out Reset image size

If an object satisfies the selection criteria of one or more of our methods outlined below, bits in the BOSS_TARGET1 target flag are set. Table 2 gives the flag name, the bit value, and the short description of the different target selection flags.

As discussed in the introduction, we wish to define a CORE sample that is uniformly selected over the BOSS footprint for statistical studies of quasars, such as measurements of the luminosity function and the clustering of quasars. While these goals do not drive our technical requirements, the survey we have designed to measure the BAO signal will also provide an unprecedented spectroscopic data set for studies of quasars themselves. Thus, design choices that are roughly neutral with regard to cost and impact on the cosmology goals are guided by these additional science considerations.

This is the motivation for dividing our quasar targets into two broad classes. Since the one (imaging) data set that we have over the entire BOSS footprint is the SDSS single-epoch photometry (including the new coverage in the SGC; Aihara et al. 2011), we define quasar CORE targets as a sample of 20 targets deg⁻², which are selected only from this single-epoch imaging data, using a uniform algorithm. As we shall see, the efficiency of the CORE sample is near our goal of 50% (i.e., ∼10 out of 20 CORE targets deg⁻² are z > 2.2 quasars). The CORE sample is designed to have a well understood, uniform, and reproducible selection function.

In contrast, the "BONUS" sample is selected using as many methods and additional data as deemed necessary to achieve our desired quasar density. The BONUS sample has a target density of 20 deg⁻². The number of BONUS targets added in each region of sky is adjusted to assure that the total density of targets, CORE + BONUS, is uniform across the sky, as we will show in Section 4.5 below. However, as we detail below, the number of BONUS targets was extended up to 60 targets deg⁻² (and then 40 targets deg⁻²), during the BOSS commissioning and early science phases, for a total (CORE+BONUS) of 80 (and then 60) targets deg⁻². The efficiency of BONUS selection is generally lower than that of CORE, despite the use of multiple algorithms and auxiliary data, simply because the relatively "easy" targets have already been picked by CORE and are therefore not included in BONUS.

Prior to BOSS, there was no extant survey that successfully targeted z > 2.2 quasars to the depth and surface density and with the efficiency we needed. The first year of BOSS spectroscopy was therefore largely a commissioning year for QTS, during which we gathered the quasar sample needed to test our various algorithms. In particular, it was only at the end of the year that we settled on the final CORE and BONUS algorithms. Thus, the nominal CORE-selected objects from the first year are not a uniformly selected sample. Section 6 describes the completeness of the final CORE sample in Year One spectroscopy.

Through this first year, we worked on and refined a variety of algorithms for BOSS target selection, as it was not clear from the outset that any single method could meet our scientific goals. These methods include the following.

• 1.  
  The Non-parametric Bayesian Classification and Kernel Density Estimator (KDE; Richards et al. 2004, 2009b), which measures the densities of quasars and stars in color–color space from training sets. Richards et al. (2009b) showed that this was able to identify quasars at 2.2 < z < 3.5 from SDSS photometry with an efficiency of 46.4% ± 5.8%, down to a magnitude limit of i = 21.3, approximately ∼0.5 mag brighter than the BOSS limit.
• 2.  
  A likelihood approach (Kirkpatrick et al. 2011), which determines the likelihood that each object is a quasar, given its photometry and models for the stellar and quasar loci.
• 3.  
  A neural network (NN) approach from Yèche et al. (2010), which takes as input the SDSS photometry and errors.
• 4.  
  A variant of the likelihood approach, which accounts for the observational errors more properly when determining the stellar locus, called "Extreme Deconvolution" (XD; Bovy et al. 2011c). Bovy et al. (2011a) present full details on how the XD method can be used to describe a probabilistic QTS technique called "XDQSO" which uses density estimation in flux space to assign quasar probabilities to all SDSS point sources. XDQSO was not used in Year One target selection, but it did become the CORE method in Year Two.

Each of the methods described above has one, or more, key parameters; these are summarized in Table 3, and Table 2 gives the associated bitwise target flags. We now describe each of these methods in turn, leaving the details for the cited papers. We also introduce a variant of the NN, the "combined neural network" (a.k.a. the NN-Combinator), which incorporates information from all the methods and produces the BONUS sample. We also describe several ancillary methods of selection, including objects associated with FIRST radio sources (Section 3.6) and repeat observations of previously known z > 2.2 quasars (Section 3.7).

Table 3. Key Parameters for the Various Methods and the Variable Name in the Output Target Files

Method	Key	Variable Name	References
	Parameter(s)	in Target Files
Likelihood	$\mathcal {P}$	LIKE_RATIO	Kirkpatrick et al. (2011)
KDE	KDEprob	KDE_Prob	Richards et al. (2009b)
	χ2star	chi2_star	Hennawi et al. (2010)
Neural Network	yNN	NN_XNN	Yèche et al. (2010)
	zpNN	NN_ZNN_phot	Yèche et al. (2010)
XDQSO	P(XDQSO)	QSOED_PROB	Bovy et al. (2011a)
Combined-NN	NN value	NN_VALUE	This paper

Download table as:  ASCIITypeset image

Gray & Moore (2003), Gray & Riegel (2006), and Riegel et al. (2008) describe the KDE classification scheme. Richards et al. (2004) and Richards et al. (2009b) have applied it to the SDSS imaging data to produce photometric quasar catalogs with ≈10⁶ quasars. The principles of the KDE are as follows. A sample of objects of known classification (stars and quasars) serves as a training set, from which the smoothed distributions of quasar and star probability as a function of color are constructed. This allows one to compute the probability that any object of interest from the test set is a star, "KDE star density," or quasar, "KDE quasar density" (e.g., Figure 8 in Richards et al. 2009b). Based on these probabilities, we define the "KDE probability" (see Figure 2 and Table 3) as

$$\begin{equation} {\rm KDE}_{{\rm Prob}} = \frac{ {\rm KDE\, quasar\, density}}{{\rm KDE\, quasar\, density + KDE\, star\, density}}, \end{equation} \tag{ 1 }$$

which can be used to decide whether a given object should be targeted as a quasar. As described in Section 3.5 of Richards et al. (2009b), for our purposes, we define the quasar density just for those objects with 2.2 < z < 3.5; all other quasars are put into the "star" category.

Richards et al. (2004, 2009b) actually define two KDEs, split at g = 21, with separate color loci (different "trainings") for the bright and faint estimations. This approach crudely accounts for the very different photometric errors of the two sets, given that the KDE method, as implemented, does not take errors explicitly into account.

Roughly 45% of objects in the KDE catalog of Richards et al. (2009b) in the "mid-z" range (i.e., the redshift range of interest to BOSS) are not stars (Table 4, Richards et al. 2009b), based on an analysis of the classification efficiency using clustering (e.g., Myers et al. 2006). In the absence of significant contamination by galaxies at the faint end of the KDE catalog, the KDE algorithm is thus about 45% efficient at the Richards et al. (2009b) target density of 18.6 mid-z quasars deg⁻².

We need a higher efficiency for BOSS, so we have applied an additional cut beyond that of the Richards et al. papers to improve the efficiency of the KDE method. This cut is based on the χ²_star statistic introduced by Hennawi et al. (2010), which quantifies how far a given object is from the stellar locus:

$$\begin{equation} \chi ^{2}_{{\rm star}} = \sum _{m=ugriz} \frac{ \big[ f^{m}_{\rm data} - A f^{m}_{\rm model} \big]^{2}}{ \big[\sigma ^{m}_{\rm data}\big]^{2} + A^{2}\big[\sigma ^{m}_{\rm model}\big]^{2} }, \end{equation} \tag{ 2 }$$

where f is the flux in each of the five SDSS bands (m = ugriz) for the data and for the model, σ^m_data is the flux error in each band, σ^m_model is the model uncertainty in each band, and A is a normalization. Following Hennawi et al. (2010), the stellar locus is defined by a set of ≈14,000 stars with accurate photometry from SDSS spectroscopic plates, on which all point sources were targeted above a flux limit of i < 19.1 regardless of color (Adelman-McCarthy et al. 2006). The minimum distance to the stellar locus, χ²_star, can then be computed by minimizing the value χ²(A, g − i), where A is the normalization constant relating the data to a model, f^m_data = Af^m_model, and g − i is the color chosen as a proxy for stellar temperature. The distribution of the minimum distance to the stellar locus, i.e., range of χ²_star, is shown in Figure 3 of Hennawi et al. (2010). The crucial strength that the χ²_star cut adds to our KDE selection is the rejection of objects that have colors consistent with those of quasars, but have flux errors that make them consistent with the stellar locus as well.

The key parameters (Figure 2) for the KDE method are the minimum thresholds for selection in both KDE_prob and χ²_star. Early in Year One, CORE objects were selected solely by the KDE algorithm (Section 4); at that time, we applied a limit χ²_star ⩾ 7. Later, when KDE was no longer the CORE algorithm, we relaxed this criterion to χ²_star ⩾ 3. Objects selected by the KDE method have the QSO_KDE target flag set.

Full details of the Likelihood method, including an in-depth analysis of its performance, are presented in Kirkpatrick et al. (2011). We summarize it briefly here.

Like KDE, the Likelihood method starts with a sample of known quasars and a sample of "Everything Else" (EE in what follows), i.e., stars and galaxies, with ugriz photometry and errors. One defines likelihoods that a given object with fluxes f^m and errors σ^m (m = ugriz) is drawn from the quasar or EE catalog by summing a χ²-like statistic over the full training set:

$$\begin{equation} {\cal L}_{\rm quasar} = \sum _i \prod _m \sqrt{1 \over 2\,\pi \big(\sigma ^m_i\big)^2} \exp \left(- {\left[f^m - {\rm quasar}^m_i\right]^2 \over 2\,(\sigma ^m)^2}\right) \end{equation} \tag{ 3 }$$

$$\begin{equation} {\cal L}_{{\rm EE}} = \sum _i \prod _m \sqrt{1 \over 2\,\pi \big(\sigma ^m_i\big)^2} \exp \left(- {\left[f^m - {\rm EE}^m_i\right]^2 \over 2\,(\sigma ^m)^2}\right). \end{equation} \tag{ 4 }$$

The sums are over all objects i in the training set. By restricting the sum to those training-set quasars in a specific redshift range, one can define an equivalent likelihood that the object in question is in this redshift range; in Year One, this was done by summing over those quasars with z > 2.2. Given these likelihoods, one defines a probability that the object is a quasar to be targeted (compare with Equation (1)):

$$\begin{equation} \mathcal {P} =\frac{ {\cal L}_{\rm quasar}(z>2.2) / A_{\rm quasar}}{ {\cal L}_{{\rm EE}}/ A_{{\rm EE}} + {\cal L}_{\rm quasar}({\rm all}\ z)/ A_{\rm quasar}}, \end{equation} \tag{ 5 }$$

where the A normalize for the possibly different effective solid angles of the quasar and EE training sets. In the denominator, the likelihood sum is over quasars at all redshifts, not just those at z > 2.2.

Like the KDE method above, this method makes use of the varying densities of objects in color space and includes a χ² selection. Note that it correctly utilizes the flux errors in determining whether a given object belongs to the quasar or EE class. Potential quasar targets can be ranked by their probability $\mathcal {P}$. We define a threshold $\mathcal {P} \ge 0.234$; for $\mathcal {P}$ above this value, we target all objects as quasars. The Likelihood method was chosen as the CORE algorithm near the end of Year One (Section 4.2). Objects selected by the Likelihood method have the QSO_LIKE target flag set.

We use an artificial NN at two stages of the selection process. Full details of this algorithm may be found in Yèche et al. (2010). As in the previous methods, we define training sets of known quasars, and objects that are not quasars.

For the first stage, we use the NN with 10 inputs for each object (the SDSS g-band magnitude, the five SDSS magnitude errors, and the four SDSS colors). The training set for non-quasars is a set of ∼30,000 SDSS point sources from SDSS DR7 (Abazajian et al. 2009), selected over the magnitude range 18.0 < g < 22.0 and with Galactic latitude b ≈ 45° to average the effects of Galactic extinction. The training set for quasars consisted of spectroscopically confirmed quasars from the 2QZ (Croom et al. 2004), 2SLAQ (Croom et al. 2009), and the SDSS (Schneider et al. 2010) quasar catalogs.

The NN developed for target selection has four layers of "neurons" (see Figure 3 of Yèche et al. 2010). The fourth layer only has one neuron, providing a single output parameter, y_NN. The quantity y_NN quantifies the probability that an input object is a quasar, although since y_NN can be greater than 1, it is not a probability in the formal sense. A photometric redshift estimate, z_pNN, is also generated (see Section 5 of Yèche et al. 2010), with a cut placed on this photometric redshift estimate, z_pNN > 2.1. Objects selected by the NN method have the QSO_NN target flag set.

XD (Bovy et al. 2011c) is a method to describe the underlying distribution function of a series of points in parameter space (e.g., quasars in color space), by modeling that distribution as a sum of Gaussians convolved with measurement errors. Bovy et al. (2011a) apply XD to the problem of QTS, using flux data from the SDSS DR8. The so-called XDQSO method is conceptually similar to the Likelihood method, but explicitly models the non-uniform errors of the training set from which the quasars and stellar/EE loci are derived. Indeed, the Likelihood method effectively double counts the errors of the training set, since the observed distribution of fluxes from which the Likelihood training set is built is the true underlying distribution convolved with the uncertainty distribution. XD avoids this double-counting by deconvolving the underlying distribution of the training set.

XDQSO constructs a model of the distribution of the fluxes of stars and quasars in different redshift ranges based on training samples of known stars and quasars. XDQSO then builds a model of the relative-flux distribution as a mixture of 20 Gaussian components and fits this model to the training data, taking the heteroscedastic nature of the SDSS flux uncertainties fully into account. The XD model for the relative-flux distribution is fit in narrow bins in i-band magnitude and combined with an apparent-magnitude dependent prior based on star counts in Stripe 82 and the Hopkins et al. (2007) quasar luminosity function. The probability for an object to be a mid-redshift quasar (2.2 < z < 3.5) is given by the ratio between the number density of mid-redshift quasars and that of stars plus all quasars at the object's fluxes (in the spirit of Equation (5)). The probability that a given object is a mid-z quasar is then

$$\begin{eqnarray} &&P(\mathrm{QSO}_{\mathrm{midz}} | \lbrace f^m\rbrace) \propto \nonumber\\ &&\quad P(\lbrace f^m{/}f^i\rbrace | \mathrm{QSO}_{\mathrm{midz}})\, P(f^i | \mathrm{QSO}_{\mathrm{midz}})\,P(\mathrm{QSO}_{\mathrm{midz}})\,, \end{eqnarray} \tag{ 6 }$$

where m indexes the fluxes and fⁱ is the SDSS i-band flux. The first factor on the right is given by the XD model for the relative-flux (i.e., color) distribution of quasars, while the second and third factors are obtained from the quasar luminosity function. The underlying relative-flux distribution is convolved with the object's flux uncertainties before evaluation. The expressions for stars and high/low-redshift quasars are similar. Probabilities are normalized assuming that these classes exhaust the possibilities (P(QSO_midz) + P(QSO_hiloz) + P(star) = 1). Objects are ranked on their mid-redshift quasar probability for targeting.

Since XDQSO target selection properly takes the flux uncertainties into account both in the training and the evaluation stage, it can be trained and evaluated on data of low S/N. It can also incorporate data from surveys other than SDSS in a straightforward way, as we describe for near-infrared and ultraviolet surveys below. The performance of XDQSO, using Stripe 82 data, is given in Bovy et al. (2011a) and its performance in Year Two will be described in a future paper. The catalog of SDSS objects selected by XDQSO is available through the SDSS-III DR8 Science Archive Server.³¹

The XDQSO method was not used during Year One, but we then set, and fixed, XDQSO as CORE for Year Two and the remainder of BOSS. In Section 6, we detail how to replicate the CORE selection using XDQSO for the BOSS quasars. Objects selected by the XDQSO method have the QSO_CORE_MAIN, and sometimes the QSO_CORE_ED, target flag set (see Section 6).

3.5.1. The UKIRT Infrared Deep Sky Survey

3.5.2. GALEX: The Far- and Near-UV

As in the SDSS-I/II quasar survey, objects that are detected in the FIRST radio survey (Becker et al. 1995) are also incorporated in target selection. Radio stars are rare, thus most radio sources with faint, unresolved optical counterparts are quasars. Optical stellar objects with g ⩽ 22.00 or r ⩽ 21.85 which have FIRST counterparts within 1'' are considered as potential quasar targets, irrespective of the radio morphology.

In the early BOSS commissioning data (Section 4), we simply selected all such radio matches. This approach targeted a substantial number of quasars with z < 2.2, and thus we placed an additional color cut, (u − g) > 0.4, to exclude UV-excess sources at lower redshift (Figure 3). Thus the QSO_FIRST flag designates objects with (u − g) > 0.4 that matched a FIRST source. Bluer FIRST sources are not rejected outright, but are required to pass one of the regular optical color selections to be selected. Section 4 describes when in Year One this (u  −  g) > 0.4 cut was implemented.

Zoom In Zoom Out Reset image size

The density of z > 2.2 quasars known before BOSS started was ∼2 objects deg⁻². Given the superior throughput of the BOSS spectrographs over those of SDSS-I/II, we decided to re-observe these objects for improved LyαF clustering signal. Moreover, this allows vital checks of survey quality and uniformity, and the data can be used to study the spectroscopic variability of quasars. We thus target previously known spectroscopically confirmed z > 2.15 quasars from the literature. We include such objects as targets if they match a point source in the target imaging to within 15 or if they match a point source in the target imaging to within 2'' and match the magnitude of that object to within 0.5.

The catalogs of previously known quasars we use include the SDSS DR7 quasar catalog (Schneider et al. 2010), the 2SLAQ quasar catalog (Croom et al. 2009), the 2QZ survey (Croom et al. 2004), the AAT-UKIDSS-SDSS (AUS) survey (S. M. Croom et al. 2012, in preparation), and the MMT-BOSS pilot survey (Appendix C).

To compare and check our moderate resolution spectra of generally fainter quasars to those taken by 10 m class telescopes using high-resolution spectrographs (e.g., KECK-HIRES and VLT-UVES), we also mined the data archives (the NED,³³ the Keck Observatory Archive³⁴, and the ESO Science Archive Facility³⁵) and added those quasars with z > 2.15 that were not included from the above catalogs.

The full sample of known quasars contains ∼18,000 z > 2.15 objects. We assign those objects in the BOSS footprint the QSO_KNOWN_MIDZ flag and give them highest targeting priority in tiling (Blanton et al. 2003).

We also veto previously known low-redshift (z < 2.15) quasars identified from the SDSS-I/II, 2QZ, 2SLAQ, and MMT surveys, labeling them with the QSO_KNOWN_LOHIZ target flag and never assigning them spectroscopic fibers.³⁶ We are confident that we are not inadvertently rejecting any real z > 2.2 quasars, since the vast majority of these objects were visually inspected and identified in the SDSS, 2QZ, and MMT surveys (Schneider et al. 2010; Croom et al. 2005). A veto of objects with known stellar spectra, again from the SDSS-I/II, 2QZ, 2SLAQ, and MMT surveys, was not implemented until Chunk 5, because we were not initially confident that shallower surveys, at their faint end, would have sufficient S/N to correctly identify stars and that our initial matching procedures were not discarding some quasars of utility to BOSS.

Combining results from several of the methods described above in target selection requires a method to merge the (overlapping) ranked lists from these methods into a single ranked catalog. The challenge is shown in Figure 4, which shows the surface density of the union of those objects selected by the KDE, Likelihood, and NN methods with no further refinement, to yield an average target density of ∼60 targets deg⁻². The tidal stream of the Sagittarius dwarf spheroidal galaxy (Ibata et al. 1995; Belokurov et al. 2006) is quite striking in this figure, spanning 180° < α < 240° and 0° < δ < +15°. The Sagittarius stream is so prominent since its constituent stars are point-like in SDSS imaging, faint (g < 23 and r < 22), and have a (g − r) < 0.4 color, satisfying criteria similar to that of our high-z QTS. The target density in Figure 4 varies from 35 to 70 deg⁻².

Zoom In Zoom Out Reset image size

3.8.1. Tuning and Ranking

3.8.2. NN-Combinator

As the above makes clear, BOSS quasar selection has been through a complex series of changes during its first two years. Here we recall the reasons for this complexity and summarize the main points of this history.

BOSS QTS is complex because

• 1.  
  for the survey's defining science goal, measurement of BAO in the LyαF, the primary requirement is a high surface density of quasars in the relevant redshift range, not simplicity or homogeneity of selection;
• 2.  
  selection of quasars in the desired redshift range from single-epoch SDSS imaging is difficult because of proximity to the stellar locus and substantial photometric errors near the magnitude limit for BOSS selection; and
• 3.  
  pre-BOSS quasar samples provided inadequate training sets in our desired magnitude and redshift range, so the quasars we discovered in this first year allowed us to refine our algorithms as the year proceeded.

Roughly speaking, the effective survey volume for measurement of LyαF clustering is quadratic in the number of quasars, so even modest gains in efficiency have a significant science impact.

As discussed in Section 3.1, the goal of CORE selection is to provide a homogeneously selected sample suitable for quasar science. Ideally, we would have frozen the CORE algorithm at the very beginning of BOSS, but the higher imperative of maximizing efficiency has led us to alter CORE as our algorithms improved. We started by using KDE+χ² as the CORE algorithm but switched to Likelihood based on its greater flexibility and simplicity. Finally, we switched from Likelihood to XDQSO based on its better performance (at a level of ∼1 additional high-z quasar deg⁻²). The chunk-by-chunk history of these changes is given in Section 4 below. We intend to maintain a fixed CORE algorithm for Years 2–5 of the survey, and for many purposes we anticipate that completeness corrections will allow use of Year One data in statistical studies of the quasar population (see Section 6).

Beyond CORE, we use whatever combinations of data and methods can maximize our targeting efficiency, including known quasars, FIRST candidates, and the BONUS sample. Because the methods described in Sections 3.2–3.5 have complementary strengths, we draw on all of them in creating the BONUS sample. We have tried different methods of forming a combined BONUS list during the first year, and we have now settled on the NN-Combinator (Section 3.8.2) as our primary tool for doing so. The individual methods feeding into the NN-Combinator use co-added SDSS photometry where it is available in overlap regions, in contrast to CORE, which relies on single-epoch photometry to ensure uniformity. Auxiliary data such as UKIDSS and GALEX photometry are fed into the XDQSO selection, which in turn is fed into the NN-Combinator.

The first area that we targeted and observed for BOSS was SDSS Stripe 82, along the celestial equator in the Southern Galactic Cap. The target field covered 3170 ⩽ α_J2000 ⩽ 450, −125 ⩽ δ_J2000 ⩽ 125, for a total area of 220 deg² (smaller than the ∼300 deg² imaging coverage on the Stripe).

The KDE method, based on single-run data (and with a cut at χ²_star ⩾ 7.0), was used as the CORE (QSO_CORE) selection for Chunk 1. The KDE method was one of the techniques used for BONUS (QSO_BONUS) with targets chosen using the co-added data described in Section 2 and given the flag QSO_KDE_COADD. Co-added data were not used in later chunks, thus the QSO_KDE_COADD flag was used only for Chunk 1. In Chunk 1, with the benefit of co-added data, the quasar and stellar loci were better defined than in the standard one-epoch SDSS data. Hence there was far more overlap between the samples of sources targeted by all of the methods, freeing fibers to be placed on lower-priority KDE targets. As most of these lower-priority targets proved to be stars, the overall efficiency of selection of the KDE method is thus quite low in Chunk 1.

The density of the co-added KDE targets was tuned on a second χ²_star value calculated from co-added data, to obtain the required total of 80 targets deg⁻² across all methods. This second χ²_star parameter is dependent on right ascension, but is always >4.0.

At tiling, all quasar targets were given priority over all other BOSS targets (such as galaxies) for Chunks 1 and 2. The tiling priorities for all the chunks are given in Appendix B (Table 13). About 40 deg² of Chunk 1 was observed following the end of hardware commissioning, i.e., after MJD 55176. Stripe 82 was re-observed in 2010 fall as part of Chunk 11 (Section 4.2).

For Chunk 2 in the NGC (Figure 5), the targeting algorithms were similar, but not identical, to Chunk 1, as co-added photometry was not available. The surface density of known quasars was lower than in Chunk 1 (since Stripe 82 contains more extensive spectroscopy from prior surveys), and with no UKIDSS coverage, there were no KX-selected targets. Unresolved optical objects that had a match to any FIRST source (Section 3.6) were included and given the target flag QSO_FIRST (bit 18). The CORE method remained the KDE. As in Chunk 1, the KDE was tuned on the χ²_star parameter to obtain a total of 80 targets deg⁻² over all methods. In Chunk 2, flux errors are larger than in Chunk 1, due to the use of single-epoch data. Thus, the stellar locus is expanded and there is far less overlap between the targets chosen by various methods. The target density of QSO_BONUS sources is thus approximately halved in Chunk 2.

At this stage, the list of quasars with high-resolution spectroscopy (Section 3.7) was added to the database of known quasars, although few lie within the boundaries of Chunk 2.

We already had our initial spectroscopic results in hand from ∼20 plates from Chunks 1 and 2 when we identified targets in Chunk 3, and we used these results to refine our algorithms. In particular, we rejected FIRST sources with (u − g) < 0.4, greatly decreasing contamination from z < 2.2 quasars, but decreasing the number of FIRST z > 2.2 objects by only 10%. The resulting FIRST target density drops to ∼1–2 deg⁻², 40% of which turn out to be 2.2 < z < 3.5 quasars (see Figure 3).

In the first two chunks, we found that only 1 new bright (i ⩽ 17.7) z > 2.2 quasar had been discovered from 486 bright targets. Thus, a bright limit of i > 17.8 was set to reduce stellar contamination at the bright end. Due to the proximity of Chunk 3 to the Milky Way, we also imposed a Galactic latitude cut of b > 25°.

There was a change in the target density and methodology from those in Chunks 1 and 2. For Chunks 3–6, the ranking method described in Section 3.8.1 was adopted, allowing us to combine Likelihood, KDE, and NN for CORE at 20 targets deg⁻². All remaining targets, to a total density of 60 deg⁻², were designated as BONUS. To monitor the CORE and BONUS changes, two new target flags, QSO_CORE_MAIN (flag bit 40) and QSO_BONUS_MAIN (flag bit 41)³⁷ were introduced. Finally, for Chunks 3–6, to provide a more uniform galaxy sample, galaxy targets were given precedence over quasar targets in tiling (see Appendix B and Table 13).

Chunks 7–9 were the first chunks targeted at the nominal survey target density of 40 quasar targets deg⁻². The area covered by Chunk 9 in the SGC was not in the original SDSS survey, and target selection was done from the DR8 imaging (Aihara et al. 2011), a region of sky where there were no previously known z > 2.2 quasars in our catalog (Section 3.7). This change led to a lower efficiency (see Section 5).

For Chunks 7–9, based on the tests described in Section 5.5, we set the CORE method to Likelihood, while BONUS targets were selected using the NN-Combinator (Section 3.8.2). In addition, previously known stars from SDSS or 2dF spectroscopy were now vetoed. The NN photometric redshift threshold was relaxed slightly, from z_pNN > 2.1 to 2.0. Chunk 9 was the last chunk to be observed in Year One, and thus the last data included in the spectroscopic sample presented in this paper. Target selection for Chunk 10 was performed in the first year of BOSS, but the Chunk 10 plates were not observed until the second year, after the summer 2010 shutdown.

Chunk 11, the re-observation of Stripe 82, was observed at the start of the second year of BOSS observations. As described in detail by Palanque-Delabrouille et al. (2011), a variability-based quasar selection was performed for Chunk 11, and for this special case, there was no differentiating between CORE and BONUS targets. The variability-based selection led to a significantly higher high-z quasar density than elsewhere in the survey, 24 z > 2.15 quasars deg⁻² (Palanque-Delabrouille et al. 2011), as we describe further in Section 5.

The XDQSO method was introduced in Chunk 12 to test its efficiency. There is substantial overlap between the target list of XDQSO and Likelihood, and including the highest ranked 20 targets deg⁻² from each yielded a total of 25 targets deg⁻². We thus defined CORE to be the union of all these targets. The NN-Combinator was retained as the method for BONUS.

The Chunk 12 and 13 spectroscopic results demonstrated the superiority of XDQSO for the core algorithm. Therefore, from Chunk 14 onward, and for the rest of BOSS, the XDQSO algorithm, and that alone, was set to be CORE. This led to a gain of ∼1 high-z quasar deg⁻² in the CORE.

Various further improvements were implemented in BONUS. For Chunk 14, a change in fiber collision prioritization (see Appendix B) led to a gain of ∼1 quasar deg⁻². In Chunk 15 we began a policy of re-observing previously known quasars in plate overlap regions, leading to a spectroscopic S/N gain of ∼15% per quasar. In Chunk 16, we incorporated UKIDSS photometry into the training of XDQSO as an input to the NN-Combinator. This led to a gain of 2–3 high-z quasars deg⁻² where UKIDSS data were available. Overlap between adjacent imaging scans allowed improved photometry for objects observed more than once (Section 2), leading to a gain of 0.3–0.5 quasars deg⁻² in Chunk 16. Also in Chunk 16, since the CORE selection was now set as XDQSO, there was no longer any requirement to use the QSO_LIKE, QSO_CORE_ ED, or QSO_CORE_LIKE target flags. Likelihood probabilities ($\mathcal {P}$) were still calculated, but these were directly consumed by the NN-Combinator, without having the QSO_LIKE target flag set. Hence the drop to ≈0 QSO_LIKE targets deg⁻² is shown in Table 5. In Chunk 17, an optical-only trained version of the XDQSO (essentially what is used for CORE) was also used as an input to the NN-Combinator used for BONUS, with a gain of ∼0.5 quasars deg⁻².

BOSS spectroscopic plates are designed by giving priority first to BOSS galaxy and quasar targets, followed by objects in various ancillary programs (Section 2 of Eisenstein et al. 2011). If additional fibers are available, we assign them to previously known 1.8 < z < 2.15 quasars; these are labeled as SUPPZ in Figure 2 and are flagged as QSO_KNOWN_SUPPZ in Table 2. These objects were observed only in Chunk 17. Reobserving these objects allows a measurement of the spectral structure from metal lines along the line of sight and spectral artifacts that may contaminate Lyα structure measurements (McDonald et al. 2006). Chunks 18 and 19 had the same algorithm for selection as Chunk 17, but did not include QSO_KNOWN_SUPPZ targets.

The sky distribution of the BOSS quasar targets is shown in Figures 6–9. In Figures 6 and 7, we show the surface densities of BOSS quasar targets for the NGC and the SGC, respectively, as selected by the CORE method (XDQSO) for DR9. In Figures 8 and 9, we show the surface densities of BOSS quasar targets for the NGC and the SGC, respectively, as selected by the CORE (XDQSO), BONUS (NN-Combinator), and FIRST methods, as well as the inclusion of all previously known z > 2.2 quasars. The CORE sample is designed to produce a mean surface density of 20 targets deg⁻², and although it is reasonably uniform, the density of targets ranges from 10 to 30 targets deg⁻² over the footprint of the survey. The largest variations are associated with Galactic structure, with excesses visible at low Galactic latitudes and in the Sagittarius stream. The BONUS sample adds enough targets in each area of sky to give a much more uniform 40 targets deg⁻².

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Table 6 summarizes the results from the first year of BOSS quasar observations. BOSS quasar targets are those which have one of the target bit flags listed in Table 2 set. There were 54,909 spectra of objects targeted as quasars, of which 52,238 were unique objects. These were observed over a footprint of 878 deg², giving a mean surface density of 63.8 targets deg⁻².

Of the 54,909 (52,238 unique) spectra, 35,305 (33,556) had high-quality redshifts, as designated by the "zWarning" flag of the spectroscopic pipeline (Adelman-McCarthy et al. 2008; Aihara et al. 2011). From visual inspection of the data, the zWarning flag is reliable at the 90%–95% level for the quasar target spectra; very few of the objects flagged as having high-quality redshifts (i.e., zWarning = 0) are incorrect. We present the performance of zWarning as a function of magnitude and S/N in Appendix D; most objects with zWarning ≠ 0 are faint objects with low S/N spectra. Given the faint magnitude limit of BOSS, it is not surprising that many of the targets that are not quasars lack the clearly identified spectral features required to assign a high-confidence redshift. We will present a detailed examination of the performance of the reduction pipeline, the zWarning flag, and the findings from the visual inspection of the data when we publish the BOSS Quasar DR9 Catalog in a separate paper.

Of the 33,556 unique objects with high-quality redshifts, 11,149 are stars, while 13,580 have z > 2.20. The remaining 8827 objects are mostly quasars at z  ∼ 0.8 and ∼1.6, and low-z compact galaxies; see Figure 11. Of the 13,580 high-redshift objects, 2317 had the QSO_KNOWN_MIDZ flag set; thus the first year of BOSS observations resulted in the spectroscopic confirmation of 11,263 new z > 2.2 quasars. A full breakdown of the number of objects associated with each target flag, the number of good (zWarning = 0) redshifts, and the number of z > 2.2 quasars obtained is given in Table 7.

Zoom In Zoom Out Reset image size

Figure 11 shows the redshift distribution of BOSS quasars from the first year and compares it with that from the SDSS DR7 quasar sample (Schneider et al. 2010) and the 2SLAQ survey (Croom et al. 2009). This plot is very similar, but not identical, to that shown in the SDSS-III overview paper of Eisenstein et al. (2011). Of course, the DR7 sample is selected over the full SDSS-II imaging area, approximately 9380 deg², while the BOSS Year One data come from observations of 880 deg². Already BOSS has slightly more quasars in the z = 2.2–2.8 range, while at higher redshifts the DR7 sample remains larger.

Degeneracies in the color–redshift relation of quasars led to the selection of low-z quasars in BOSS. The quasars at z ∼ 0.8 have Mg ii λ2800 at the same wavelength as Lyα at redshift z ∼ 3.1, giving these objects similar broadband colors, while the large number of objects at z ∼ 1.6 is due to the confusion between λ1549 C iv and Lyα at z ≈ 2.3. We shall come back to this feature when comparing the performance of the NN, KDE, and Likelihood methods in Section 5.4. The tail of objects at z ≳ 3.5 includes a significant contribution from re-observations of previously known quasars.

Figures 12 and 13 present our key results, the efficiency of the current target selection algorithms. For these tests, we have constructed a control sample of targets on Stripe 82, where our spectroscopy is more complete than anywhere else on the sky, albeit still not perfect. Here we include data from Year Two from Chunk 11, where Stripe 82 was retargeted using a variability selection for quasars (Palanque-Delabrouille et al. 2011). Stripe 82 also has high completeness because quasars are selected from co-added photometry, with much smaller photometric errors.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For Figure 12, we select the quasar targets in our normal way from single-epoch data, with the first 20 targets deg⁻² selected by the XDQSO CORE algorithm. Targets are ranked in order of probability, and the plot shows the number of z > 2.2 quasars deg⁻² versus the number of targets deg⁻², with the slope of the curve indicating the efficiency of selection. The CORE algorithm selects 10.7 z > 2.2 quasars deg⁻² from its 20 targets. We then show the average contribution of KNOWN and FIRST quasars, totaling 1.6 high-z quasars deg⁻². This increment assumes a surface density of 0.9 known high-z quasars deg⁻² (and 0.7 deg⁻² from FIRST), which is consistent with our Year One data (see Table 7) but lower than the surface density of known pre-BOSS high-z quasars on Stripe 82, which is unusually well studied. Finally, we add the BONUS targets from the NN-Combinator, again in rank order. At 40 targets deg⁻², we are just above the minimum BOSS goal, with a mean density of 15.4 z > 2.2 quasars deg⁻². Stripe 82 samples a wide range of Galactic latitude and thus stellar density; we therefore anticipate that this test should be representative of selection efficiency averaged over the full BOSS survey region. We also found from observations of early chunks that adding additional fibers beyond the nominal 40 deg⁻² led to only very minimal gains in yield.

Figure 13 shows the impact of adding UKIDSS and GALEX data to single-epoch SDSS photometry. For this test we use the XDQSO algorithm alone, since this is where these auxiliary data sets currently enter our selection procedures, and we extend the efficiency curves up to 80 targets deg⁻². At 40 targets deg⁻², the efficiency for XDQSO with single-epoch SDSS imaging alone is 15.0 z > 2.2 quasars deg⁻². Adding GALEX data improves the efficiency to 16.2 deg⁻², adding UKIDSS improves it to 17.3 deg⁻², and adding both improves it to 18.6 deg⁻². Thus, both of these data sets can significantly enhance the efficiency of BOSS QTS in regions where they are available. Stripe 82 has medium-deep ("MIS") GALEX data, and the improvement with shallower ("AIS") coverage will be smaller, but our tests indicate that GALEX addition will still improve the selection.

Figure 14 shows the redshift distribution of all known quasars on Stripe 82 as a function of redshift, as well as those selected by the single-epoch SDSS algorithms illustrated in Figure 12 above. The ratio of the two measures the completeness of BOSS single-epoch quasar selection relative to known quasars in this well-studied region, ranging from 40% to 70% over our critical redshift range 2.2 < z < 3.5. Of course, this remains a lower limit to the true completeness at the BOSS magnitude limit, though in the 2.2 < z < 3.5 redshift range we anticipate that the BOSS Stripe 82 sample selected from co-added photometry and variability has high completeness (Palanque-Delabrouille et al. 2011).

Zoom In Zoom Out Reset image size

Figure 15 shows examples of BOSS spectra of quasar targets from the Year One data. From top to bottom: a z > 5 quasar found by the Likelihood method (and not selected by any other method); a newly discovered z = 2.6 quasar at a typical S/N; a z = 3.5 quasar selected only by the KX method; a re-observed BAL quasar showing spectroscopic variability over 3377 days in the observed frame; a star at our typical S/N; and a z = 1.5 quasar with our typical S/N.

Zoom In Zoom Out Reset image size

Figure 16 shows the distribution of quasar targets that are spectroscopically confirmed as either stars or z > 2.2 quasars from the BOSS first-year data in the (u–g) versus g color–magnitude plane. The distribution of stars at the bright end, g < 18, and the lack of bright z > 2.2 quasars led us to impose the bright i = 17.8 limit. Objects fainter than g = 22 are brighter than our r-band limit of 21.85 mag.

Zoom In Zoom Out Reset image size

Figure 17 shows the SDSS (u − g), (g − r), (r − i), and (i − z) colors as a function of redshift for the BOSS Year One data. Also shown are the mean color in redshift bins (thin solid line) and the model of Bovy et al. (2011b; thick colored line). This model is systematically bluer than the data at low redshift; BOSS target selection systematically excludes UV-excess quasars, and thus those low-redshift quasars that happen to enter the sample are redder than the average quasar. The trends with redshift are due to various emission lines moving in and out of the SDSS broadband filters, and the onset of the LyαF and Lyman-limit systems (e.g., Fan 1999, Richards et al. 2002, 2003, Hennawi et al. 2010, Bovy et al. 2011a and Peth et al. 2011, but see also Prochaska et al. 2009 and Worseck & Prochaska 2011). I. D. McGreer et al. (2012, in preparation) will present a detailed analysis of this diagram and its implications for our completeness.

Zoom In Zoom Out Reset image size

Figure 18 shows the SDSS color–color diagrams for the first year BOSS quasars, for all quasars with good (zWarning = 0) redshifts above z = 2.2. This figure illustrates the redshift dependence of quasar colors as the Lyα emission line moves from the g band to the r band at z ≈ 3.5. Quasars with 2.2 < z < 3.5 lie in the range −0.3 < (g − r) < 0.6, while objects with z > 3.5 generally have (g − r) > 0.8.

Zoom In Zoom Out Reset image size

Figure 19 shows the distribution of objects in the redshift–luminosity ("L–z") plane for three recent large quasar surveys: SDSS (black points), 2SLAQ (cyan), and BOSS (red). There are ≈105, 000 objects in the SDSS DR7 catalog and ≈9000g ⩽ 21.85 low-redshift quasars from the 2SLAQ Survey (Croom et al. 2009). We calculate the absolute i-band magnitudes, M_i, using the observed i-band PSF magnitudes and the k-corrections given in Table 4 of Richards et al. (2006). The three surveys together cover the L–z plane well, with a dynamic range in luminosity of ≈4 mag at any given redshift up to z ∼ 3.5. This coverage will be vital for calculating the evolution of the faint end of the quasar luminosity function and placing strong constraints on the luminosity dependence of quasar clustering.

Zoom In Zoom Out Reset image size

Because of the BOSS hardware commissioning in fall 2009, only 37.4 deg² (out of a possible 220 deg²) were observed in Chunk 1 under survey-quality conditions after MJD 55169. Thus, Stripe 82 was re-targeted, re-tiled, and re-observed for Year Two as Chunk 11 (Palanque-Delabrouille et al. 2011). However, the non-survey quality data from prior to MJD 55169 were visually inspected during the very early part of the survey and used to inform subsequent QTS decisions.

In Chunks 1–6, the QTS algorithm was generous, allocating 60–80 targets deg⁻². Chunk 7 was the first time we ran the BOSS QTS at the nominal 40 targets deg⁻². Of the 4506 unique targets in this chunk, 1595 (35%) are classified as z > 2.20 objects with zWarning = 0 (Table 6). Although this does not reach the BOSS efficiency goal of 50%, there are several reasons that this number can be considered a lower bound. First, Chunk 7 is in the region of sky known to have a high density of faint stellar sources, due to the presence of the tidal stream of the Sagittarius dwarf spheroidal galaxy (see Ibata et al. 1995; Ibata et al. 1997; Belokurov et al. 2006, and our Figure 4). Second, visual inspections of the spectra identified 0.5–1 more high-z quasars per square degree than the pipeline, and while not all of these might be suitable for LyαF analyses (e.g., due to BALs which cause the pipeline to fail), there should be a net gain upon production of the final BOSS quasar catalogs. Finally, and potentially most importantly, we know that our target selection methods and algorithms have continued to improve, with the incorporation of XDQSO and ancillary data such as UKIDSS and GALEX (see also the discussion on a variability based QTS in Section 7).

In this context, the performance of QTS in Chunks 8 and 9, with only 11.5 and 9.5 z > 2.2 quasars deg⁻², respectively, was disappointing. Chunk 8 lies at relatively low Galactic latitudes and is affected by stellar contamination. Chunk 9 is in a region of sky where there was neither previously known quasars nor FIRST radio coverage. We continued to observe the rest of Chunks 8 and 9 in Year Two.

The original motivation for the implementation of multiple target selection algorithms was the lack of evidence prior to BOSS observations that a single method could select z > 2.2 quasars down to g ≈ 22 with our required efficiency. With the Year One data now in hand, we can compare the effectiveness of our different methods. However, due to the continually changing nature of the BOSS QTS over this year, where different methods were used as CORE and BONUS, this comparison will be generally qualitative in nature. The interested reader is referred to the discussions in Bovy et al. (2011a) for further comparisons.

As an aid for our discussions, we give a condensed version of Table 7 in Table 8, where we list the number of targets from this first year, broken down by the three key selection methods. Again, given the non-uniform selection over this year, this table is provided as an informational guide only; it should not be used as a direct comparison between methods.

The redshift distributions for objects with reliable redshifts selected by our three main methods (NN, KDE, and Likelihood) are given in Figure 20. Again, because of the non-uniform manner in which these methods were applied during Year One, this plot should not be interpreted as a quantitative comparison between the methods. There is substantial overlap between the methods; many objects are selected by more than one technique. The three histograms have similar shapes over the range 2.2 < z < 3.5. While NN avoids being confused by z ∼ 1.5 objects and KDE avoids objects at z > 3.5, all three methods select a substantial number of objects at z ∼ 0.8.

Zoom In Zoom Out Reset image size

Figures 21, 22, and 23 show the color–color and the color–magnitude distributions of z > 2.2 quasars selected by the Likelihood, NN, and KDE methods, respectively. The figures show in orange and black the ratio of numbers of objects selected by each method to the total number of Year One quasars, at each point in color space. This ratio is normalized to the global ratio of targets from Column 4 of Table 8; thus a point in color space with a value >100% is one where the method in question outperforms the total selection on average. The difference between the three methods is clear in the (u − g) versus (g − r) diagrams. The contours for the Likelihood method are fairly flat away from the stellar locus. NN performs well at (u − g) ∼ 0.6, (g − r) ∼ 0, and in those regions of color–color space corresponding to higher-redshift quasars, but does more poorly elsewhere. KDE selects objects only over a very narrow range in (g − r). From the (g − r) versus i-band color–magnitude diagram (bottom right panels of the figures), we see that the Likelihood method was more efficient at selecting fainter, i ≳ 21.0 quasars, while the NN tends to select the brighter i ≲ 20.0 objects at all (g − r) colors.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

These trends can be understood given the methodology of these algorithms. The Likelihood method downweights objects close to the stellar locus as the denominator of Equation (5) gets large, which is why Likelihood selects few objects there. Otherwise, the Likelihood method traces the overall BOSS Year One sample in color–color and color–magnitude space. The Likelihood method did not place any cuts on photometric redshift and hence samples the high redshift distribution of the BOSS data well, especially at (g − r) ≳ 1 (corresponding to redshift z > 3.5). We refer the interested reader to Kirkpatrick et al. (2011) for full details of the Likelihood performance.

At the crux of an artificial NN is the sample of objects used to train it (see Yèche et al. 2010 and references therein, and Section 3.4). The training set for the NN we have used was based on the SDSS quasar catalog and the 2SLAQ surveys, and did not use data from the MMT pilot survey (Appendix C) or the AUS survey. Thus, this training set was geared toward brighter quasars (i < 20.2), giving rise to the tendency for NN to select the brighter quasars.

The KDE training set included only 2.2 < z < 3.5 quasars, and thus the redshift histogram drops to zero at z = 3.5 (Figure 20). This is related to the fact that KDE quasars inhabit a much narrower range of the (g − r) versus (r − i) color–color plane than the other two methods. In summary, Figures 21–23 reflect the relative strengths and trainings of these methods; ultimately, the three methods complemented each other well.

After spectroscopy from the first few chunks had been analyzed, it became clear that the survey would have to decide on a single method for the CORE and that we would have to restrict ourselves to the nominal target density of 40 targets deg⁻². Thus, we designed a test to decide which combinations of methods gave the best yields for the CORE and BONUS selections.

The "Blind Test Area" is a region of sky of ∼1000 deg² in the NGC at high declination (δ > +40°) and high Galactic latitudes, shown by the white line in Figure 4. This area is used for tuning the threshold of each method to a particular target density. The resulting thresholds were then applied to existing data to determine the selection efficiency.

Table 9 summarizes these tests. This table gives the surface density of 2.2 < z < 3.5 quasars from early (Chunks 1–3) BOSS spectroscopic data that would be recovered by various methods at various thresholds of their key parameters when they are tuned to yield a surface density of 20 or 40 deg⁻² in the blind survey region. The effectiveness of each quasar spectrum for LyαF studies depends on its redshift (and thus the spectral coverage of the forest) and its brightness (and thus the S/N of the spectrum). This "value" is quantified by a score of each quasar, motivated by the checks performed in McDonald & Eisenstein (2007); summing this over the expected quasars per square degree gives the numbers in Table 9. These scores do not include contributions from quasars outside the redshift range 2.2 < z < 3.5. "Weighted Likelihood" was an adaption of the Likelihood method to maximize this score, as discussed in detail by Kirkpatrick et al. (2011).

We also tried selecting quasars using a simple color region isolating the region where z ∼ 2.7 quasars are found, akin to the mid-z box used by Richards et al. (2002), but this did not deliver an efficiency close to our requirements.

Although Table 9 shows that the KDE method returns the most z > 2.2 quasars (9.45 deg⁻²) at the CORE target density of 20 deg⁻², after much deliberation, we selected the Likelihood method as CORE for the latter stages of Year One, since it is a simpler algorithm to understand and explain, it has a more uniform spatial selection, and is easier to reproduce. Further tests showed that using the NN in its "Combinator" mode for BONUS would yield the highest number of high-z quasars overall. The difference when weighting by the LyαF score was too small to motivate us to include it; see the discussion in McQuinn & White (2011).

However, tests of the Year One data with the XDQSO method (Bovy et al. 2011a) showed it selected about 1 z > 2.2 quasar deg⁻² more than Likelihood. Thus in Chunks 12 and 13 (Section 4.3) the union of Likelihood and XDQSO was treated as CORE, allowing us to test them directly against one another (Bovy et al. 2011a). In Chunks 12 and 13, 2426 out of 4710 XDQSO targets had spectra with zWarning = 0 and 2.2 < z < 3.5, for an efficiency of 52%, while Likelihood obtained 2296 quasars from 5086 targets, for a 45% efficiency. This result is our motivation for declaring XDQSO to be CORE for the rest of the BOSS quasar survey.

The photometric pipeline sets a series of flag bits for each detected object which identify problems with the processing of the SDSS photometry, ranging from the presence of bad columns to issues with deblending (Stoughton et al. 2002). These are particularly useful in recognizing when the photometry might be poor and therefore color selection of targets unreliable. The detailed meaning of the specific flag bits in what follows is described in Stoughton et al. (2002) and the SDSS-III Web site;³⁹ the logic behind these flag combinations is given in Richards et al. (2002).

Note that unlike the latter paper, we did not calculate and apply the flag checks on each band separately, and just use the flags associated with the union of the detections in the five SDSS bands. While this could cause us to reject some genuine quasars, checks on Stripe 82 (where the flag checking on the co-added data was significantly less strict; see below) showed only a statistically insignificant 1% difference in the number of quasars identified.

We first define a combination of flag bits that denotes whether the source in question was adversely affected by interpolation across bad pixels, bad columns, or bleed trails:

INTERP_PROBLEMS = (PSF_FLUX_INTERP && (gerr > 0.2∥rerr > 0.2∥ierr > 0.2)) ∥ BAD_COUNTS_ERROR ∥ (INTERP_CENTER && CR),

a combination that identifies objects in which the deblending of overlapping images may be questionable:

DEBLEND_PROBLEMS = PEAKCENTER ∥ NOTCHECKED ∥ (DEBLEND_NOPEAK && (gerr > 0.2∥rerr > 0.2∥ierr > 0.2))

and a combination which identifies objects with detectable proper motion between the exposures in the different SDSS filters (asteroids):

MOVED = DEBLENDED_AS_MOVING && (rowv/rowverr)² + (colv/colverr)² > 3².

Here, the symbols (&&, ∥, !) have their standard meanings from Boolean logic. The quantities rerr, gerr, and ierr are the quoted uncertainties in the PSF photometry in g, r, and i, respectively, rowv and colv are the measured proper motion along the rows and columns of the CCD, and rowverr and colverr are their errors.

A source is considered to have clean photometry if it satisfies the following:

GOOD = BINNED1 && !BRIGHT && !SATURATED && !EDGE && !BLENDED && !NODEBLEND && !NOPROFILE && !INTERP_PROBLEMS && !DEBLEND_PROBLEMS && !MOVED.

The GMAG_BITMASK records whether a target satisfies the magnitude limits required to be targeted as a quasar. Magnitude cuts are made on PSF magnitudes measured by the SDSS, corrected for Schlegel et al. (1998) Galactic extinction. These limits are encoded in a flag:

GMAG_BITMASK = $(g \le 22 \parallel r \le 21.85) \&\& i \ge 17.8$.

This includes a cut at the bright end, reflecting the fact that bright z > 2.2 quasars are extraordinarily rare. We also define a variant of this flag:

GMAG_BITMASK_NOB = (g ⩽ 22∥r ⩽ 21.85),

to be used when no bright cut is required—such as when retargeting known quasars or FIRST objects.

The DR8 paper (Aihara et al. 2011) describes the algorithm used to define the primary detection of a given object, if it lies in the ∼50% of the SDSS footprint covered by more than one scan. The RESOLVE_BITMASK records whether a source is a primary target in the SDSS photometry.

The BOUNDS_BITMASK records whether a source is within the footprint of the SDSS imaging, which is useful for keeping track of data from the ancillary surveys (FIRST, UKIDSS, GALEX) used in the target selection.

We saw in Section 3.6 that we could limit the number of z < 2.2 sources targeted by FIRST with u − g color cut. Thus we define:

FIRST_COLOR_BITMASK = (u − g > 0.4).

The single-epoch photometry used for co-addition on Stripe 82 is first vetted by a series of quality cuts. All fluxes used in the co-addition are limited by the following conditions.

• 1.  
  They must be primary, i.e., RESOLVE_BITMASK must be true.
• 2.  
  They must be observed under photometric conditions (an important issue from Stripe 82, as it was repeatedly observed under non-photometric conditions as part of the SDSS Supernova Survey; see Frieman et al. 2008);
• 3.  
  They must have a positive estimated inverse flux variance (zero values are indicative of problems with the data);
• 4.  
  They must pass various flag cuts:(!DEBLEND_TOO_MANY_PEAKS && !SATUR && !BADSKY && !SATUR_CENTER && !INTERP_CENTER && !DEBLEND_NOPEAK && !PSF_FLUX_INTERP).

Prior to the commencement of BOSS spectroscopy, we carried out spectroscopy of quasar candidates selected from co-added photometry in SDSS Stripe 82 to increase the number of faint quasars available in the BOSS redshift range for testing and training of BOSS targeting algorithms. Candidate quasars for these observations were selected in two ways: first, using very inclusive cuts in the (χ²_phot, χ²_star) plane, where these χ² statistics are as defined in Hennawi et al. (2010), and second, using the methods outlined in Richards et al. (2009b, 2009a). These observations were intended to include as large a sample of z > 2.2 quasars as possible, but do not represent a statistically well-defined sample, so we do not describe their selection in greater detail.

Observations of these candidates were carried out in queue mode between 2008 September and 2009 January using the Hectospec multi-fiber spectrograph (Fabricant et al. 2005) on the 6.5m Multiple Mirror Telescope (MMT). The data were reduced using Juan Cabanela's ESPECROAD⁴⁰ pipeline, an external version of the SAO SPECROAD pipeline (Mink et al. 2007). Quasars were identified by eye, and redshifts were measured using IRAF.

The MMT program was conducted before the release of the SDSS DR7 quasar catalog. In addition, BOSS targets all confirmed quasars from the MMT program for re-observation (Section 3.7). Thus, most of the MMT observations have been superseded by subsequent SDSS DR7 or BOSS spectroscopy at better resolution, wavelength coverage, and S/N. In Tables 14 and 15, we provide positions, PSF photometry (as observed, uncorrected for Galactic extinction), and redshifts for confirmed quasars from the MMT survey. Objects that are not flagged Primary in the CAS are listed separately. Over 99% of quasars that were observed a second time have redshifts in agreement (to Δz < 0.05) between the MMT survey and the SDSS/BOSS pipelines.