THE RAVE SURVEY: RICH IN VERY METAL-POOR STARS^*

The RAVE (Steinmetz et al. 2006; Zwitter et al. 2008) is a spectroscopic survey of apparently bright stars, I ≲ 13, with some 350,000 stars observed to date. RAVE uses the six-degree-field (6dF) multi-object, fiber-fed spectrograph on the UK Schmidt telescope. Spectra of ∼100 targets are obtained simultaneously. The dominant target population belongs to the thin disk, but includes many thick disk and halo giant stars. The spectra are centered on the infrared calcium triplet, cover the wavelength range region λλ8410–8795 and have resolving power R ∼ 7500. This is significantly higher than that which is typically employed in surveys to find metal-poor candidates for later confirmation—for example, the Hamburg/ESO survey (HES; Christlieb et al. 2004) is based on objective-prism spectra with R ∼ 400 at the Ca ii K line at 3934 Å. We show here that RAVE has the resolution to determine the metal-poor nature of a star directly from the survey spectra, without the need for follow-up confirmation, providing ideal targets for later detailed analysis.

The RAVE spectral analysis pipeline provides estimates for each star of the line-of-sight velocity and the stellar parameters (T_eff, log g, [m/H], and vsin i) through χ² fits to a grid of synthetic spectra.¹⁷ RAVE pipeline metallicities ([m/H]_RAVE) have uncertainty of ∼0.2 dex, derived from an external calibration of high-quality spectra of predominantly metal-rich stars, which dominate the sample (Zwitter et al. 2008). The grid of synthetic spectra utilized extends to [m/H] = −2.5, requiring additional analysis below this limit: we provide that here.

The RAVE survey is designed primarily to obtain excellent line-of-sight velocities. This has been achieved, with accuracy and precision of a few km s⁻¹ for the vast majority of the stars (Steinmetz et al. 2006). A large subset of the spectra has sufficiently high signal-to-noise ratio (S/N) for the RAVE pipeline also to provide good quality estimates of the values of stellar parameters (Zwitter et al. 2008). At the start of the present investigation, the RAVE (internal) database contained over 147,000 spectra from which stellar parameters were derived. How do we identify VMP stars among these? Stars with true metallicity near or below the lower limit of the synthetic spectra in the RAVE pipeline (−2.5 dex) will be assigned metallicities close to that lower limit. A metallicity cut in the RAVE database of, say, 3σ above the lower bound, [m/H]_RAVE ⩽ −1.8, should then isolate probable VMP stars. Adopting this cut and isolating only normal stars gives a sample of 622 potential metal-poor stars.

Furthermore, extremely metal-poor stars will be erroneously assigned to significantly higher metallicities by the RAVE pipeline. The RAVE spectrum of a typical VMP star is nearly featureless, with only the Ca triplet lines being measurable, as shown in Figure 1. There exists a degeneracy between the stellar parameters and metallicity, in that the strengths of the Ca triplet lines can be equal for both a lower-metallicity cooler giant and a higher-metallicity hotter turnoff star. A temperature difference of 100 K corresponds to a [m/H]_RAVE difference of about 0.1 dex, assuming the parameters of the star still lie along an appropriately metal-poor old evolutionary track. Hence, stars for which the RAVE T_eff value is significantly hotter than the value that is derived by other indicators are likely to be stars whose real [m/H] values are significantly lower than [m/H]_RAVE. This expectation is confirmed by identification in the RAVE database of several known VMP stars, incorrectly allocated [m/H]_RAVE > − 1.8, including CD −38°245 with [Fe/H] =−4.2 (Cayrel et al. 2004) and CS 29502-092 with [Fe/H] = −2.76 (Aoki et al. 2007). As anticipated, known VMP stars with anomalously high [m/H]_RAVE values also have anomalously high estimated stellar effective temperatures from the RAVE pipeline.

Zoom In Zoom Out Reset image size

Temperature anomalies may be identified using photometric temperatures. Photometry from the Two Micron All Sky Survey (2MASS; Cutri et al. 2003) is available for the RAVE sample, so that an effective temperature, T_phot, may be calculated using the (de-reddened) color–T_phot relationship of Alonso et al. (1999), with due care to use the two-color relations to check for cool binary companions. The mean value of E(B − V) for our candidate metal-poor stars is 0.11 mag (median 0.06 mag). We assume conservatively that the stars of interest are beyond the reddening layer, since their estimated distances based on the stellar parameters are typically several hundred parsecs to a few kiloparsecs from the Sun.

These T_phot values allow us to adopt the following modified criterion to define a candidate VMP star:

$$\begin{eqnarray*} &&{\rm [m/H]_{RAVE}} - 0.1(({T_{\rm RAVE}} - {T_{\rm phot}})/(100\, {\rm K})) < -1.8. \end{eqnarray*}$$

Application of this selection criterion to the RAVE database of non-variable stars within the appropriate temperature range to be halo turnoff/red giant stars (7000 K <T_RAVE< 3800 K) yields 320 stars in addition to those selected by the straight cut [m/H]_RAVE ⩽ −1.8. The total number of stars we selected for further analysis of their RAVE spectrum, based on the above criteria, is then 942.

For each candidate VMP star, we measure the equivalent width (EW) of the Ca triplet and Fe i lines from the RAVE spectrum.¹⁸ These EW values are then the input for an abundance analysis using the MOOG program (Sneden 1973), assuming LTE and one-dimensional, plane-parallel Kurucz model atmospheres.¹⁹ We adopted the NLTE corrections of Starkenburg et al. (2010) for all our stars (in the end significant only for stars with [Fe/H] <−2.5).

The estimates of stellar gravity from the RAVE pipeline analysis identify most VMP candidates as giants: 86% of these stars have RAVE log g <3, with the mean RAVE log g value being 1.0 ± 0.9. Stellar gravity values from the RAVE pipeline are accurate to better than 0.5 dex for the typical metal-rich stars in the RAVE database (Zwitter et al. 2008). For metal-poor stars, the lower-metallicity cutoff of the synthetic spectrum grid biases the pipeline analysis to overestimate, systematically, both T_eff and log g values, so the actual fraction of true giants in the VMP sample should be even higher. Indeed, there were only four stars whose location on a reduced proper-motion diagram indicated that they were dwarfs. For these we adopt log g = 4.5. The stellar surface gravity for the remaining stars was obtained by fitting to an old (12 Gyr) Yonsei-Yale isochrone (Demarque et al. 2004) with a metallicity matching the current metallicity estimate of the star (the procedure is iterated). The gravity value derived from the isochrone fitting was adopted in the calculation of the model atmosphere (while the Yonsi-Yale isochrones do not extend beyond the red giant branch, <0.1 dex error in final abundance results should a star be actually on the asymptotic giant branch). Our analysis was iterated until the derived (from CaT) metallicity did not differ by more than 0.1 dex from the value used in calculating the model atmosphere and isochrone fit. We were able to derive Ca abundances for 771 stars out of our 942 star initial selection from RAVE: these 771 stars form our VMP candidate sample.

The next step is to provide an external calibration of our abundance scale, derived from high signal-to-noise echelle spectra.

We obtained echelle spectroscopic data for 112 candidate metal-poor stars selected from the RAVE database using a preliminary version of the criteria described above. The observations were carried out with several telescope/spectrograph combinations, including Magellan-Clay/MIKE, APO-3.5m/ARCES, AAT/UCLES, Max Planck-2.2m/FEROS, CFHT/Espandons, and VLT/UVES. Full details of the data acquisition and reduction are deferred to our Letter on the derived elemental abundances (J. P. Fulbright et al. 2011, in preparation). Briefly, all spectrographs delivered a resolving power greater than 30,000 and, with the exception of the UCLES setup, the wavelength region covered from below 4000 Å to beyond 8000 Å, albeit with some coverage gaps. The effective wavelength range for the UCLES spectra is 4460–7260 Å. In each case the data were reduced using standard methods for echelle data, utilizing pipeline programs when available. The typical S/N level of the spectra is greater than 100 pixel⁻¹ and often exceeds 200 pixel⁻¹.

The abundance analysis utilized Kurucz stellar atmospheres and the MOOG program, with now the metallicity of the stellar atmosphere for each star set equal to the value of [Fe ii/H] derived from the analysis. The value of the stellar effective temperature was derived using the excitation temperature method based on Fe i lines. The mean difference of T_ex − T_phot is −66 ± 138 K. The value of the stellar gravity was again taken to be that value corresponding to the derived T_eff on an old (12 Gyr) Yonsi-Yale isochrone of the appropriate metallicity. The microturbulent velocity was set to the value that minimized the slope of the relationship between the iron abundance derived from Fe i lines and the value of the reduced EW of each line. The Fe i lines measured in the RAVE spectra are usually too weak in true VMP stars to give reliable results. The iron abundances derived from these lines were useful, however, in rejecting from the analysis those (few) stars for which the Ca triplet lines appeared weak, but the Fe i lines were strong. Further inspection of the spectra of these stars revealed that the Ca lines are likely weakened by emission cores.

The iron abundances from the echelle data provide an immediate check on our selection criteria. The result is encouraging: of the 92 stars with echelle data for which we predicted [Fe/H] <−2 from the RAVE data, 87 (95%) indeed have [Fe/H] values below −2.

Calibration of the RAVE EW abundances is then achieved through least-squares fitting between the echelle-based iron abundance, [Fe/H]_Hi-Res, and the EW-based calcium abundance from the RAVE spectra, [Ca/H]_RAVE. The correlation between these two quantities is shown in Figure 2 and the relationship is

$$\begin{eqnarray*} &&{\rm [Fe/H]_{Hi-Res}} = 0.93 {\rm [Ca/H]_{RAVE}} -1.33, \end{eqnarray*}$$

with a correlation coefficient of 0.82 and standard deviation of the residuals of 0.25 dex.

Zoom In Zoom Out Reset image size

Given that our candidate sample had an upper metallicity cutoff determined by uncertain quantities, significant incompleteness at the high-metallicity end is expected. We check this using published catalogs of candidate VMP stars and a second RAVE study.

A SIMBAD-aided literature search showed that only 47 of our 612 very metal-poor stars ([Fe/H] <−2 dex) have previously been proposed as being metal-poor, from low-resolution spectroscopic analyses, including 21 from Frebel et al. (2006) and 11 from Bond (1980). Ten of the 138 stars for which we derive −2< [Fe/H] <−1 also have entries in SIMBAD identifying them as VMP candidates.

The most uniform comparison is with the HES catalog of over 20,000 candidate VMP stars (Christlieb et al. 2008). Cross-identification with the RAVE database (at that time with 200,000 entries) yielded 473 matches, with HES metallicity estimates for 296 stars, all with [Fe/H]_HES ⩽ −2. One hundred nine of these stars are included in our RAVE VMP sample. The RAVE pipeline gave [m/H]_RAVE < −1 for 76 of the remaining 187 stars. We re-analyzed these RAVE spectra as above, obtaining [Fe/H] <−1 for only 22 stars, with just 4 with [Fe/H] <−2, of which the most metal-poor has [Fe/H] = −2.26. These stars are therefore all very close to the VMP threshold.

The high-resolution sample of Ruchti et al. (2010) was selected from the same RAVE catalog as this study and includes 20 stars for which those authors derive [Fe/H] <−2 from their echelle spectra but which were not selected by our VMP criteria. We re-analyzed the RAVE spectra of these stars following the procedures developed above and found good agreement: [Fe/H] <−2.0 for 15 stars, and all 20 stars having [Fe/H] <−1.7. The lowest iron abundance of this group is −2.41 dex, within 2σ of our calibration. Approximately half of these stars have [m/H]_RAVE between −1.5 and −1.8, the remainder having RAVE spectra of low quality. As expected, measuring errors in our parent RAVE sample lead to increasing incompleteness in our final sample, at above [Fe/H] ∼−2.5.

At very low metallicities, we test for true VMP stars observed by RAVE, but excluded from our sample, by cross-matching the RAVE database with published catalogs of confirmed VMP stars. There are 253 RAVE stars in the HERES project (Barklem et al. 2005), which provides high-resolution follow-up for HES catalog stars. There are only two candidate VMP matches: J112243.4-020936 (RAVE) is HE1120-0153 (HERES) and J132244.1-135531 (RAVE) is HE1320-1339 (HERES). Our VMP star criteria identified both, and our calibration gave each an iron abundance of [Fe/H] = −2.6. The HERES values are [Fe/H] = −2.77 and −2.78, respectively, in excellent agreement. No confirmed very metal-poor star has been excluded from our RAVE sample.

There are three stars in our RAVE sample at or below −4 dex. These are C0022448-172429 for which we derive, from this calibration, [Fe/H] = −4.0 dex, CD −38°245 for which we derive [Fe/H] = −4.2 dex, and a third star, with a RAVE spectrum at our minimum acceptable RAVE S/N cutoff, which our calibration gives [Fe/H] = −4.0 dex. Our echelle-based result for C0022448-172429 is [Fe/H]_Hi-Res = −4.02, and for CD −38°245 is [Fe/H]_Hi-Res = −4.20. CD −38°245 was previously confirmed from echelle data to have [Fe/H] = −4.2 by Cayrel et al. (2004). Our RAVE-based calibration is indeed providing true iron abundances to ∼0.25 dex accuracy, even at the lowest known metallicities.

Our analysis is the first confirmation of the "ultra metal-poor" nature, [Fe/H] <−4 dex, of C0022448-172429. The low-resolution, objective-prism HES spectrum for C0022448-172429 (Christlieb et al. 2008, where this star is identified as HE0020-1741) provided estimates of −3.0 and −3.3, depending on which methods those authors used. Christlieb et al. further estimate that C0022448-172429 is carbon-rich, with [C/Fe] =+1.0. Our echelle analysis confirms the carbon-rich nature, with our value being [C/Fe] = +0.9, and we find that this star has an extremely high nitrogen abundance, [N/Fe] = +3.2. Thus, this star joins the select group of carbon-enhanced metal-poor (CEMP) stars (see Norris et al. 2010 for a recent discussion of this class of object).

Although this makes our selection function not internally consistent, adding the identified extra VMP stars after re-analysis of their RAVE spectra (15 from Ruchti et al., plus 4 from the HES cross-check) to our main sample of 612 VMP stars, results in the dashed histogram in Figure 3. This contains 631 stars with [Fe/H] ⩽−2 dex, i.e., "very metal-poor," and two stars with [Fe/H] <−4, i.e., "ultra metal-poor" and one at ∼−4 dex. Five hundred sixty seven of these VMP stars were not known to be VMP previously, and one of the two ultra metal-poor stars is a new confirmation. The remaining star at ∼−4 dex has no previous published abundance determination and due to its marginal quality RAVE spectrum remains as a candidate, awaiting scheduled high-resolution follow-up; if confirmed this would be a new discovery.