THE RAVE CATALOG OF STELLAR ELEMENTAL ABUNDANCES: FIRST DATA RELEASE

The chemical pipeline utilizes RAVE reduced spectra and their stellar parameters, such as effective temperature ($T_{\rm{eff}\normalfont }^{\rm {RAVE}}$), gravity (log g^RAVE), and metallicity ([m/H]^RAVE), returned by the RAVE pipeline (described in Siebert et al. 2011).²⁰ The pipeline also uses an EW library based on the RAVE line list. The RAVE line list contains 604 known absorption lines identified in the RAVE wavelength range (Section 2.2). The EW library contains the EWs of absorption lines for every point on the T_eff, log g, and [m/H] grid (Section 2.3), which accounts for the variation of the EW with the stellar parameters. Variations of EW with elemental abundances are then further calculated for each point of this grid, for five different abundance levels in the range [ − 0.4, 0.4] with respect to the value of metallicity [m/H].

The pipeline contains some auxiliary codes to (re)normalize the spectrum and detect spectral defects like bad normalization, cosmic rays, and other unwanted features.

The pipeline algorithm can be outlined as follows.

• 1.  
  Upload the normalized, RV corrected, and wavelength calibrated spectrum and the estimated stellar parameters $T_{\rm{eff}\normalfont }^{\rm {RAVE}}$, log g^RAVE, [m/H]^RAVE.
• 2.  
  Upload the RAVE line list from the EW library and the corresponding EWs for the stellar parameters at the five different abundance levels.
• 3.  
  Extract a shorter line list of those lines which, at the estimated stellar parameters, have large enough EWs to be visible above the noise.
• 4.  
  Fit the strong lines and correct the continuum.
• 5.  
  Construct the COGs of the lines by fitting a polynomial function through the five EW abundance points.
• 6.  
  Create the model by assuming a Gaussian profile for each line and summing these profiles together.
• 7.  
  Vary the chemical elemental abundances to obtain different models by changing the EWs of the lines according to their COG.
• 8.  
  Minimize the χ² between the models and the observed spectrum to find the best-matching model.

In the following we give further details pertaining to these steps.

The RAVE line list contains 604 absorption lines identified in spectra of the Sun and Arcturus. The lines correspond to the element species N i, O i, Mg i, Al i, Si i, S i, Ca i, Ti i, Ti ii, Cr i, Fe i, Fe ii, Co i, Ni i, Zr i, and to the CN molecule. In order to get precise chemical elemental abundances from the EWs, we first need reliable atomic parameters for the lines. A critical parameter is the oscillator strength (gf, often expressed as logarithm log gf). If precise laboratory measurements are missing for the log gf values then they are obtained through an inverse spectral analysis to obtain "astrophysical log gfs." This is the case for the RAVE wavelength region, where the Vienna Atomic Line Database (VALD; Kupka et al. 1999) reports that only 11 Fe i lines have laboratory oscillator strength measurements. The Sun's and Arcturus's spectra were synthesized and the log gf values for each line were obtained using an automated procedure which varies the log gfs so as to match both spectra simultaneously. For the synthesis we adopted solar abundances as given by Grevesse & Sauval (1998). They will be the zero point of the abundances of this work. The details of the procedure are outlined in Appendix A.

The pipeline makes use of a shorter working line list which contains the lines strong enough to be visible above the noise. The lines deemed visible are those satisfying

$$\begin{displaymath} {\rm EW}(\mbox{m\AA})> 2\cdot \frac{1}{{\rm STN}}\cdot d\cdot 1000, \end{displaymath}$$

where STN is the signal-to-noise ratio²¹ and d is the spectral dispersion (for RAVE spectra d = 0.4 Å pixel⁻¹). Other lines can blend with these visible lines, contributing to their flux absorption and changing their apparent EW. We therefore added to the working line list all those lines having EW > 1 mÅ and being closer than 0.6 Å (1.2 Å is the FWHM of a typical RAVE spectrum) to the visible lines.

In order to compute the line COGs we built an EW library with the RAVE line list for 30, 640 atmosphere models. The EWs were computed using the LTE line-analysis software MOOG (Sneden 1973) and the ATLAS9 model atmosphere grid (Castelli & Kurucz 2003). As the spacing of the ATLAS9 grid is different to our requirements, we linearly interpolated the ATLAS9 grid to produce our own grid, covering the T_eff range [3600, 7600] K in steps of 100 K, gravity range [0.0, 5.0] dex in steps of 0.2 dex, and metallicities in the range [−2.5, +0.5] dex in steps of 0.1 dex. We did not compute the EWs for atmospheres with T_eff > 5100 K and log g < 1.0 (hot supergiants), because MOOG does not produce results for these parameters and furthermore there are no such stars in our catalog. For every atmosphere model we computed the EWs of the lines for five abundance levels with respect to iron: [X/Fe] = −0.4, −0.2, 0.0, 0.2, 0.4 dex assuming [Fe/H] = [m/H]. The whole EW library consists of 145, 080 files. To obtain EWs for stellar parameters between the grid points of the EW library we linearly interpolate the closest points on the grid.

The EW computation requires a value of the microturbulence ξ for each atmosphere model. For high-resolution data ξ is typically determined by measuring the EWs of Fe i lines and changing ξ until the iron abundances inferred from strong and weak lines agree. As we do not measure individual EWs for our spectra we cannot use this procedure, but instead rely on relations giving ξ as a function of the stellar parameters. Such relations have been derived by Edvardsson et al. (1993), Reddy et al. (2003), and Allende Prieto et al. (2004), where ξ is given as a function of T_eff and log g. Unfortunately, these results cover only specific regions of parameter space (e.g., hot dwarfs or cold giants). Thus, we derived our own relation covering a wide a range of T_eff and log g. To do so we made use of literature results from high-resolution spectroscopy that report both stellar parameters and ξ of their stellar samples. We collected data for 712 giants and dwarfs spanning a wide range in T_eff and log g from Luck & Heiter (2006, 2007), Bensby et al. (2005), Fuhrmann (1998), Fulbright et al. (2006), and Allende Prieto et al. (2004). Figure 1 displays the coverage in the T_eff–log g plane for the sample. A third-degree polynomial fit was used to obtain the microturbulence dependence on gravity and effective temperature. Appendix B gives the coefficients of this polynomial fit, ξ_poly.

Zoom In Zoom Out Reset image size

Figure 2 compares the Allende Prieto et al. (2004) law with ξ_poly. Uncertainties of σ_ξ = 0.32 km s⁻¹ in our polynomial law translate approximately into ∼0.04 dex elemental abundance uncertainties for dwarfs stars (as estimated by Reddy et al. 2003 and Mishenina et al. 2004).

Zoom In Zoom Out Reset image size

Before measuring EWs, the spectrum's continuum needs to be re-normalized and its strong lines fitted. While the spectrum is continuum normalized by the RAVE pipeline, we have found that this normalization is not rigorous enough for elemental abundance estimation: the RAVE pipeline employs a third-order spline function to fit the continuum, leaving behind fringing effects on scales of less than 50 Å wide seen in some RAVE-normalized spectra. The chemical pipeline performs a new normalization to remove such unwanted features, fitting the continuum and the strong Ca ii and H i Paschen lines. These strong lines are then excluded from the measurement process because they are difficult to fit properly. The pipeline then considers as "continuum" the fit to the classical continuum plus the strong lines, and by comparison with this level the metallic lines are measured.

Before applying the continuum correction we estimate a preliminary metallicity, which we call [m/H]^est to distinguish it from the final metallicity [m/H]^chem. This preliminary metallicity is required because the intensity of the absorption lines must be known in order to subtract them. When they are subtracted the continuum level can then be seen. However, in order to measure the intensity of a line one must know where the continuum lies. The continuum level is therefore determined iteratively, starting with the normalization of the RAVE pipeline and measuring the line intensities using the chemical elemental abundances estimation subroutine (see Section 2.6), with the difference that all the abundances vary together as one variable [X/H] = [m/H]^est. Wavelength intervals ∼20 Å wide centered on the strong lines are avoided as the large wings of these lines can affect the results of the χ² determination of the metallicity. The best-match [m/H]^est is then used to synthesize the metallic lines, which are then subtracted before the strong lines are fit.

Fitting the continuum is performed in four steps, summarized below and illustrated in Figure 3 from top to bottom.

Zoom In Zoom Out Reset image size

• 1.  
  Perform a preliminary metallicity and subtract the metallic lines (gray line in panel (a), Figure 3).
• 2.  
  Fit the strong lines with a Lorentzian profile for Ca ii and a Gaussian profile for H i and subtract them (gray line in panel (b), Figure 3).
• 3.  
  Estimate the continuum profile by box-car smoothing what remains (gray line in panel (c), Figure 3).
• 4.  
  Add the strong lines and the continuum together to obtain the new "continuum" (gray line in panel (d), Figure 3).

The chemical pipeline uses the parameters $T_{\rm{eff}\normalfont }^{\rm {RAVE}}$ and log g^RAVE for the elemental abundance estimation. The observed spectrum is fit by a model spectrum first obtained by subtracting the flux absorbed by the metallic lines to that derived above in Section 2.5, i.e., the strong lines + continuum fit (gray line in panel (d), Figure 3). The absorption lines are assumed to have Gaussian profiles, and as the instrumental profile is dominant with respect to the line profile, we use the same FWHM for all lines. However, this FWHM is varied and optimized during the χ² minimization process.

The pipeline computes the COG of the lines by using the EWs at five levels of abundances ([X/Fe] = −0.4, −0.2, 0.0, 0.2, 0.4 dex, assuming [Fe/H] = [m/H]^RAVE) from the EW library. For every line, the five EW points are fit with a third-order polynomial function, which serves as the COG. However, these polynomial functions represent only part of the COGs and will diverge if extrapolated too far beyond the five levels in the EW library. To avoid such divergences we limit accepted abundance results to those in the range −0.6 ⩽ [X/Fe] ⩽+0.6.²²

Using the COGs the pipeline then creates the model spectrum, using [X/H] = [m/H]^RAVE as a first guess for the metallicity. It computes the χ² between the observed spectrum and the model and, through a minimization process, changes the elemental abundances [X/H] until the best match (minimum χ²) with the observed spectrum is reached. The minimization process is performed with 15 variables: 13 elemental abundances, one molecule, and the instrumental FWHM of the lines.

Figure 4 shows the best-match model spectrum of the Sun and compares it with the observed one. The three spectra on the top show how the model spectrum is built up: the spectra of three elements (Fe, Si, Mg) at parameters T_eff = 5861 K and log g = 4.54 are generated according to the estimated elemental abundances and added together to construct the solar spectrum.

Zoom In Zoom Out Reset image size

This methodology has a drawback: the computed COG of any absorption line neglects the opacity of the neighboring lines, leading to an overestimation of the EW of blended lines and so an underestimation of the elemental abundances. This systematic originates from the EW library created by using the MOOG's driver "ewfind," which computes the expected EW of the lines as if they were isolated. In order to minimize this systematic error we apply a correction coefficient to reduce the EWs of lines which are physically blended. The corrected EW is given by

$$\begin{equation} {\rm EW}_{\rm{corr}}={\rm EW} \cdot {\rm coeff} \cdot {\rm cont}, \end{equation} \tag{ 1 }$$

where the coefficient

$$\begin{displaymath} {\rm coeff}=1-\sum _i^{{\rm neighbor}<0.2\, {\AA}} {\rm EW}_i/2.50/{\rm dispersion} \end{displaymath}$$

is computed by considering all the neighboring lines within 0.2 Å of the line of interest. The dispersion is expressed in Å pixel⁻¹ and cont is the value of the continuum (gray line in panel (d) of Figure 3) at the central wavelength of the line. The multiplication of the EW by cont in Equation (1) corrects for the effects of strong lines (such as Ca ii) if the line of interest lies within their large wings. This correction reduces the systematic error in the abundances from ∼ − 0.15 dex to ∼ − 0.1 dex or less for most elements (see details in the following sections, where quality checks are outlined).

Roughly 25% of RAVE spectra are affected by defects like fringing or cosmic rays that cannot be removed by continuum correction (they usually affect regions smaller than 100 Å). In order to determine the locus and size of the defect we define the flux residual between the observed and the best-match model for the ith pixel with

$$\begin{displaymath} r(i)=f_{{\rm model}}(i)-f_{{\rm obs}}(i), \end{displaymath}$$

where f is the flux of the spectra, and we use the following algorithm (also used in Siebert et al. 2011).

• 1.  
  Consider the interval I_j = [j − 10, j + 10] centered on the pixel j. Compute

  $$\begin{displaymath} \tilde{\chi }^2(j) =\frac{1}{\max (I_j)-\min (I_j)}\cdot \sum _{i=\min (I_j)}^{\max (I_j)}\left(\frac{r(i)}{\sigma }\right)^2, \end{displaymath}$$

  $$\begin{displaymath} \psi (j) =\frac{1}{\max (I_j)-\min (I_j)}\sum _{i=\min (I_j)}^{\max (I_j)} r(i), \end{displaymath}$$

  where $\tilde{\chi }^2(j)$ is the reduced χ² and ψ(j) is the estimation of the area between the observed and model spectra. σ is the inverse of the signal-to-noise (STN) ratio.
• 2.  
  If $\tilde{\chi }^2(j)>2$ and ψ(j) > 2 · σ then the pixels in the interval I_J are labeled as a defect.

The process is repeated for all the pixels of the spectrum, resulting in an array of 0 and 1, with 1 label indicating non-defective pixels. The fraction of pixels that are non-defective is then denoted with the frac parameter (i.e., the higher the frac parameter, the better the quality of the spectrum). We deem spectra with frac < 0.7 as overly affected by fringing/cosmic ray defects and exclude these spectra from our analysis.

We ran the pipeline on a sample of 1353 synthetic spectra to which three intensity levels of artificial noise were added, testing the accuracy of the results at S/N =100, 40, and 20. To make the sample as realistic as possible, the spectra have been synthesized with distributions of T_eff and log g from a mock sample of RAVE observations (M. Williams 2009, private communication) created by using the Besançon model (Robin et al. 2003). The T_eff versus log g distribution of the sample is shown in Figure 5.

Zoom In Zoom Out Reset image size

Each spectrum of the sample had the elemental abundances of one of the stars from the Soubiran & Girard (2005) catalog, ensuring realistic distribution of the chemical abundances. As our line list has only eight elements in common with the Soubiran & Girard catalog, we assigned the elemental abundance [X/H] = [Fe/H] to the elements Cr, Co, Ni, Zr. We also assigned [X/H] = [α/H] to C, N, O, Mg, Al, Si, S, Ca, Ti when the measurement of one of these elements was missing. The spectra were synthesized using the code MOOG at resolution 0.01 Å pixel⁻¹ and degraded to RAVE resolution (0.4 Å pixel⁻¹, 1.4 Å FWHM).

The first test was performed by adopting the stellar parameters T_eff, log g, and [m/H] of the synthetic spectrum in order to evaluate the accuracy of the elemental abundance measurements of the pipeline alone. The results are illustrated in Figure 6. For the second test we added uncertainties to the stellar parameters so as to simulate real RAVE spectra. Following the typical RAVE errors, we added a random Gaussian error with standard deviation 300 K in T_eff, 0.5 dex in log g, and 0.3 dex in [m/H] to the correct parameters and ran the pipeline adopting such "wrong" parameters. For the sake of brevity and clarity we present only the results for the seven most reliable elements (see also Figure 7).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

3.1.1. Results Assuming No Error in the Stellar Parameters

Results at S/N = 100. With no errors added to the stellar parameters the residuals between measured and expected elemental abundances [X/H] have a systematic error of ≃ − 0.10 dex and standard deviation σ ≃ 0.10 dex or smaller (see Figure 6), with a slight variation from element to element. The relative enhancement [X/Fe] shows a small (≃ − 0.10 dex) systematic error at high [X/Fe].

Results at S/N = 40. The elemental abundance residuals have σ ≃ 0.10–0.20 dex depending on the element and show the same systematics seen with S/N = 100, but slightly less pronounced.

Results at S/N = 20. At this S/N the pipeline can still estimate abundances of Fe, Al, and Si and the relative enhancement to iron [X/Fe], but with σ ≃ 0.2–0.3 dex. Other elements like Mg and Ti exhibit larger systematic errors. An α-element abundance estimated as the average of Mg and Si still yields reliable results with an error of σ ≃ 0.2 dex. Elements such as Ca and Ni cannot be reliably measured because the lines are too weak to be detected.

We found correlations between stellar parameters and elemental chemical abundances obtained. In particular, the abundances correlate with T_eff (Figure 8). This is very likely due to continuum correction, which can appear lower than the real level in spectra crowded of lines, as in cold giants stars.

Zoom In Zoom Out Reset image size

3.1.2. Results Including Errors in the Stellar Parameters

Results at S/N = 100. When errors are added to the stellar parameters the elemental abundances [X/H] are affected to the level of σ ≃ 0.15–0.20 dex for elements like Mg, Al, Si, Ca, Fe, Ni. Titanium abundances show clear systematic errors, being overestimated at high abundance and underestimated at low abundance. Enhancements relative to Fe, [X/Fe], have errors σ < 0.2 dex and little sign of systematic errors for most of the elements.

Results at S/N = 40. The elemental abundance errors are in the range σ ≃ 0.15–0.30 dex depending on the element. The relative elemental abundances [X/Fe] are reliable with errors σ ≃ 0.2 dex.

Results at S/N = 20. At this S/N the pipeline can still estimate abundances of Fe, Al, and Si and the relative enhancement with respect to iron [X/Fe] even with an error of σ ≃ 0.2–0.3 dex. Other elements like Mg and Ti show systematic errors. An α-element abundance computed by averaging Mg and Si yields reliable results with an error of σ ≃ 0.2 dex. Elements such as Ca and Ni cannot be reliably measured because the lines are too weak to be detected.

As could be expected, there is a correlation between errors in stellar parameters and errors in the elemental abundance estimates. Figure 9 highlights these correlations, showing that most of the elemental abundance errors are correlated with T_eff errors. In general, an overestimate in T_eff corresponds to an overestimate in abundance, since most of the lines have smaller EW at higher T_eff. Only weak correlations with log g and [m/H] errors are evident.

Zoom In Zoom Out Reset image size

The initial RAVE input catalog had no stars whose chemical elemental abundances were known to high accuracy to compare with the results of our pipeline. To create a comparison data set of real stars, we therefore observed 104 stars chosen from Soubiran & Girard (2005, hereafter SG05), a collection of high-precision elemental abundance measurements from the literature. Since these are bright stars, most of the RAVE spectra have S/N > 100 and therefore the accuracy in our abundance estimates is representative of the high S/N case. Ruchti et al. (2010, hereafter R10) have also re-observed 243 RAVE stars at high-resolution and measured stellar parameters and chemical elemental abundances. The RAVE spectra of these stars have S/N between 30 and 90 and therefore we can use them to test the accuracy of our procedure for intermediate S/N.

We have seven elements in common with SG05 and five elements with R10 (Fe, Mg, Si, Ca, and Ti). However, for the latter we could only compare four elements because the weak Ca i lines were not strong enough in the RAVE spectra to be measured by the pipeline. As with synthetic spectra, we ran a first test by adopting the high-resolution stellar parameters to evaluate the performance of the pipeline alone and then ran a second test by using the RAVE stellar parameters, which have larger uncertainties relative to the high-resolution data.

The results of the first test are shown in Figure 10 where the gray dots and black "+'s" correspond to SG05 and R10, respectively. When the high-resolution stellar parameters are adopted the chemical pipeline's elemental abundances agree with the SG05 and R10 results to within ∼0.15 dex on average. The smaller dispersion of the SG05 stars around the one-to-one correspondence line compared to R10 is due to the higher S/N of the RAVE spectra of these stars. All elements are typically underestimated by ∼0.1 dex, an offset that is roughly constant across the entire abundance range [−2.0, +0.5] dex. The α-enhancement (Figure 11) is estimated well for all but [α/Fe] > 0.4 dex, where there is a small systematic bias of ∼0.1 dex.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

When the RAVE stellar parameters are adopted (the "second test," Figure 12) the dispersion in our elemental abundances compared to the published values increases to ∼0.2–0.3 dex on average, depending on the element. In Table 1 we report the mean and standard deviation after grouping the sample in dwarf and giant stars separately. The larger dispersion at low abundances ([X/H] <−1.0 dex) is very likely due to the fact that estimating precise stellar parameters on such spectra is challenging due to the few and weak visible absorption features (excluding for the Ca ii triplet, which is not measured).

Zoom In Zoom Out Reset image size

The enhancement estimates ([X/Fe], see Figure 13) are accurate to within ∼0.3 dex. At high enhancements, abundances are systematically underestimated by ∼0.1–0.2 dex, depending on the element. However, as we are comparing our measurements with other authors' (high resolution) measurements, the variance of the residuals is the quadratic sum of the variances of our and the high-resolution results. This means that the standard deviations reported in Figures 12 and 13 represent a conservative estimate of the expected RAVE errors.

Zoom In Zoom Out Reset image size

In order to take into account the effect of binary stars, RAVE observes some stars (∼5% of the whole sample) more than once. In this chemical catalog, 964 stars have been observed twice, 90 thrice, and 67 four times. Multiple observations are also useful to estimate the internal error by comparing the results obtained from different spectra of the same object. A fair comparison would require that the repeated observations of an object have the same S/N. Unfortunately, due to variable atmospheric conditions the spectra often have different S/N values. We therefore averaged the STNs of the spectra for each object and computed the corresponding standard deviations of the elemental abundances. In this way we could assess the magnitude of internal errors in our procedure as a function of STN. In Figure 14 we plot the results for 12, 504 RAVE stars having multiple observations, where we use the RAVE internal data release as it has a substantially larger number of stars. For all the elements the 1σ confidence line (gray solid line) lies below 0.2 dex and hovers around ∼0.1 dex for STN > 80.

Zoom In Zoom Out Reset image size

In Figure 15 we compare the distributions of [m/H]^RAVE, [m/H]^chem, and [Fe/H]^chem for 37819 RAVE spectra with STN > 20. The metallicity [m/H]^chem is inferred from the chemical elemental abundances with the equation given by Salaris et al. (1993)

$$\begin{equation} [{\rm m/H}]^{{\rm chem}}=\rm{[Fe/H]}+ \log (0.638\times 10^{\rm {[\alpha /Fe]}}+0.362), \end{equation} \tag{ 2 }$$

where the α-enhancement is computed as

$$\begin{equation} \rm{[\alpha /Fe]}=\frac{\rm{[Mg/H]+[Si/H]}}{2}-\rm{[Fe/H]} \end{equation} \tag{ 3 }$$

and the elemental abundances [Mg/H], [Si/H], and [Fe/H] come from the chemical pipeline.

[m/H]^RAVE is ∼0.1 dex lower than [m/H]^chem, which is to be expected as the non-calibrated RAVE metallicity is underestimated by ∼0.15 dex with respect to the reference stars used in Siebert et al. (2011). The shape of the [m/H]^RAVE distribution is fairly similar to the [m/H]^chem distribution for dwarf stars but different for giants. Moreover, [m/H]^RAVE seems to better match [Fe/H]^chem than [m/H]^chem, particularly for giants. This could be due to α-enhancement: giant stars have a higher proportion of thick disk stars (α enhanced) than do dwarf stars, which are mostly thin disk (not α enhanced). Although used during the stellar parameters estimation process, the RAVE pipeline is unable to measure α-enhancement (Zwitter et al. 2008). Therefore, it is possible that particularly for α enhanced stars [m/H]^RAVE better represents [Fe/H]. Apart from the discussed shift of 0.1 dex in metallicity, there seems to be fair agreement between the [m/H]^RAVE and [m/H]^chem distributions.

Zoom In Zoom Out Reset image size

Our elemental abundance measurements are indirect measurements in the sense that they are inferred from the comparison between the intensity of lines seen in real spectra and their intensity predicted by stellar atmosphere models. Since the models are different for different stellar parameters, a question arises regarding whether all the models yield elemental abundances which refer to the same zero point, i.e., the origin of the internal elemental abundance scale. The same question concerns whether this zero point refers to the same zero point of the real spectra, i.e., between the internal and external scales.

The latter question has a prompt answer: we do not know the external scale, because real stellar atmospheres have never been directly probed and all elemental abundance measurements refer to models. Therefore, we can only check the consistency of the internal scale. This can be performed by comparing the measured elemental abundances of a sample of synthetic spectra at different stellar parameters, as done in Figure 6.

In this plot, at S/N = 100, the residuals between measured and expected elemental abundances are on average zero for any gravity and metallicity; the points align along a straight line with slope roughly equal to one. This means that the measured differences in elemental abundance at any metallicity regime are the same (constant offset), i.e., they refer to the same zero point. However, the offset from element to element may vary, showing that the offsets are due to the measurement process and not due to the metallicity of the atmosphere models.

On the other hand, there is a clear correlation between residuals and T_eff: the higher the T_eff, the higher the measured elemental abundance. This systematic may be due to continuum correction: cooler stars have spectra that are crowded with absorption lines and so the continuum appears lower than it is. Additionally, the varying behavior of the elements in Figure 6 can be explained by considering that their lines lie in different regions of the spectrum, i.e., on the wings of a Ca ii line or in a region relatively free of lines. This can translate into systematic, wavelength-localized shifts of the estimated continuum as a function of T_eff.

Although this systematic effect appears tractable and correctable, we did not apply any correction to the data because such a systematic is not visible on tests with real spectra. When the RAVE elemental abundances are compared with the expected elemental abundances of standard stars as function of T_eff  no clear trend is visible (Figure 16), except that RAVE elemental abundances have a slight systematic bias to lower values. A reasonable explanation is that the elemental abundance measurements given by SG05 and R10 suffer from the same systematic error, due to uncertainties in estimating the correct continuum level in regions crowded with absorption lines or affected by the wings of strong, broad lines.

Zoom In Zoom Out Reset image size

We conclude that the RAVE chemical pipeline can determine chemical elemental abundances with a systematic error that may be as large as ∼0.1 dex as a function of T_eff. Since this systematic looks quite linear with T_eff (see Figures 9 and 16) to reduce this error to ±0.05 dex we suggest analyzing separately samples of stars with T_eff > 5500 K (mostly dwarfs) from samples with T_eff < 5500 K (mostly giants).

The spectra have been selected from the DR3 RAVE database using the following constraints.

• 1.  
  Effective temperature 4000 ⩽ T_eff(K) ⩽ 7000. This is the temperature range within which the RAVE line list has been calibrated. At lower temperature the spectra are characterized by molecular lines other than CN (CH, TiO, and other molecules), while at higher temperatures lines of ionized atoms appear, and both are not included in the line list.
• 2.  
  Signal-to-noise STN > 20. For STN < 20 the absorption lines are strongly affected by noise and the stellar parameters and chemical elemental abundances are not reliable.
• 3.  
  Rotational velocity V_rot < 50 km s⁻¹. At higher rotational velocity the lines have an FWHM larger than that due to the spectral resolution (FWHM ≃ 1.2 Å) and they cannot be precisely measured. Moreover, any spectra showing larger lines might be a double-lined spectrum, which also must be avoided.

Despite these selection criteria, some spectra exhibit emission or are affected by bad continuum normalization. For such spectra the stellar parameters and chemical elemental abundances are not reliable and it is advisable to reject them. Constraints on the parameters χ² and frac help identify these spectra. For statistical studies we suggest rejecting spectra with χ² > 2000 and frac < 0.7, which removes 739 out of 37,848 spectra of the catalog.

The distributions of the stellar parameters and STN values are presented in Figure 17. The quantities T_eff and log g are given by the RAVE data archive whereas STN and [Fe/H] are estimated by the chemical pipeline. The distributions in T_eff and log g show two peaks corresponding to giants at lower temperature and dwarfs, which are mostly at higher temperatures. The iron abundance distribution peaks at [Fe/H] ≃ −0.1 dex for dwarfs and [Fe/H] ≃ −0.5 dex for giants. The latter are in average more metal-poor because they lie further from the Galactic plane and have therefore a larger fraction of thick disk stars.²³

Zoom In Zoom Out Reset image size

We encourage the user to pay particular attention to selection effects introduced by the stellar parameters and STN. Together they affect the total number of abundance estimations as well as their accuracy. Absorption lines can be undetectable because of the low [m/H] of the star, because the too-high (or too-low) T_eff does not populate enough of its electronic levels, or because the spectrum has a too-low STN. This is illustrated in Figures 18 and 19, where the number of spectra having [X/H] estimation diminishes with STN and [m/H]. In general, only metal-rich stars have elemental abundance estimations at any STN whereas metal-poor stars have estimations only if their spectra have high STN.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We now discuss and summarize the reliability of the chemical abundances for individual elements in light of the previous discussion. First, we make some general remarks about the measured elemental abundances and their errors.

• 1.  
  The accuracy of the chemical abundances depends on several variables. In particular if T_eff is erroneously estimated, the abundances are affected to different degree for different elements. If there are systematics deviations in T_eff there will be systematic deviations in [X/H] as well.
• 2.  
  In general [X/H] are underestimated, and the magnitude of the bias is a function of T_eff: stars with T_eff < 5000 K yield on average abundances that are ∼0.1 dex lower than stars with >5000 K, whose abundance errors average to zero. This effect increases the errors when stars with different T_eff are analyzed. On the other hand, relative abundances [X/Fe] are nearly unaffected because the trend is similar for all elements.
• 3.  
  The errors given below refer to the expected errors for two intervals of STN at [m/H] ∼ 0.0 dex. As already discussed in the previous section, errors can increase for lower STN and lower [m/H].

Magnesium yields reliable results on synthetic and real spectra. At any T_eff we expect an abundance error σ_Mg ⩽ 0.15 dex for STN ⩾ 40 and ∼0.25 dex for 20 ⩾ STN ⩾ 40.

Aluminum abundances are obtained from only two physically isolated lines (which are instrumentally blended at RAVE resolution). Despite the instrumental blending, the lines are strong and give accurate estimates in tests with synthetic and real spectra that shows no systematic offset. For STN ⩾ 40 we expect abundance errors σ_Al ⩽ 0.2 dex and ∼0.3 dex for 20 ⩾ STN ⩾ 40.

Silicon is the most reliable element together with Fe. We expect an abundance error σ_Si ⩽ 0.15 dex for STN ⩾ 40 and ∼0.20 dex for 20 ⩾ STN ⩾ 40.

Calcium abundances are obtained from only five weak Ca i lines. Ca i is better measured at higher metallicity and T_eff < 5000 K. Estimated errors are σ_Ca ∼0.25 dex for STN ⩾ 40 and ∼0.4 dex for 20 ⩾ STN ⩾ 40.

Titanium gives reliable abundances at high STN. At STN ∼ 40 its abundance is reliable for T_eff < 5000 K and underestimated for higher temperatures. The correlation with T_eff errors is particularly strong (T_eff underestimation generates [Ti/H] underestimation and vice versa) leading to larger errors. We expect an abundance error of σ_Ti ∼ 0.2 dex at STN ⩾ 40 and ∼ 0.3 dex for 20 ⩾ STN ⩾ 40. We recommend separately analyzing stars with T_eff lower and higher than 5000 K.

Iron is the most reliable element together with Si. It can be accurately measured on spectra with any T_eff in the range we consider. We expect an abundance error of σ_Fe ∼ 0.1 dex at STN ⩾ 40 and ∼0.2 dex at 20 ⩾ STN ⩾ 40.

Nickel has six weak lines in the RAVE wavelength range that are visible only at T_eff < 5000 K. In synthetic spectra with STN = 100 the pipeline yields errors comparable to those of other elements but biased to lower values by ∼0.2 dex. It is not measurable for [m/H] <−0.6 dex. Tests on real spectra are inconclusive because they were all performed for stars with T_eff > 5000 K. We expect uncertainties of σ_Ni ∼ 0.2 dex for STN ⩾ 40 and ∼0.3 dex for 20 ⩾ STN ⩾ 40. Because of the small number and the weakness of its lines, it is advisable to use Ni values with care.

The overall metallicity [m/H]^chem is inferred from the chemical elemental abundances with Equation (2) (Salaris et al. 1993), with the abundances [Mg/H], [Si/H], and [Fe/H] delivered by the chemical pipeline. The metallicity [m/H]^chem is reliable and we expect errors of σ_[m/H] ∼ 0.1 dex for STN ⩾ 40 and ∼0.2 dex for 20 ⩾ STN ⩾ 40.

The α-enhancement [α/Fe] has been computed as the average of [Mg/Fe] and [Si/Fe] (Equation (3)). In the range 20 ⩾ STN ⩾ 40 it is advisable to use it instead of the single α element values because it is more reliable. We estimate errors of σ_α ∼ 0.1 dex for STN ⩾ 40 and ∼0.2 dex for 20 ⩾ STN ⩾ 40.

The RAVE chemical catalog contains data for 37,848 spectra of 36,561 stars and it is provided as an ASCII table of 37,848 lines. It contains chemical abundances for the elements Mg, Al, Si, Ca, Ti, Fe, and Ni, RAVE stellar parameters, signal-to-noise STN, object name, and other parameters for quality checks as explained in Table 2. The distributions of the relative chemical abundances are shown in Figure 20. The catalog will be accessible online at http://www.rave-survey.org and via the Strasbourg Astronomical Data Center (CDS) service.

Zoom In Zoom Out Reset image size

The RAVE line list consists of 604 absorption lines (excluding the H i and Ca ii strong lines) identified by looking at the high S/N Sun and Arcturus spectra (Hinkle et al. 2000) and verified by cross checking the line identification of Moore et al. (1966) and Hinkle et al. (2000). Some previously unidentified lines have been added in order to match features that are clearly visible in spectra of the Sun and Arcturus.

The lines have been selected from Kurucz data (Kurucz 1995) and we have also adopted Kurucz's wavelengths and excitation potentials χ. Following Chen et al. (2000) we also adopted the enhancement factor E_γ for the Unsöld approximation because it is commonly accepted that the line broadening due to Van der Waals interaction obtained from Unsöld's approximation is too weak and needs to be corrected. For the CN molecule we used the dissociation energy D = 7.75 eV, as suggested by Sauval & Grevesse (1994).

We performed the inverse spectral analysis by synthesizing solar and Arcturus's spectra with MOOG (a standard LTE line analysis and synthesis code by Sneden 1973) and changing the log gfs values until a good match with both observed spectra is reached. We show here that matching these two spectra (or more) allows us to improve the log gfs of blended lines, which otherwise would be impossible to disentangle.

The EW of a measured absorption line is related to the log gf-value through the curve of growth equation (see Gray 2005, formula 16.4)

$$\begin{equation} \log \frac{{\rm EW}}{\lambda }=\log C+\log A+ \log (gf\cdot \lambda)-\theta \chi -\log \kappa _{\nu }, \end{equation} \tag{ A1 }$$

where C and θ are functions²⁴ of the temperature T, and κ_ν is the opacity at the frequency ν. When all the atmosphere model parameters are fixed, the curve of growth equation can be written as a function, f, of log gf and elemental abundance A

$$\begin{equation} {\rm EW}=f(\log gf, A). \end{equation} \tag{ A2 }$$

This allows us to disentangle two or more blended lines: let EW^x_blend be the EW of a blended feature of the star x composed of n lines of m different elements (n ⩾ m) and having EW₁, EW₂,..., EW_n unknown EWs. Then we can write the system of equations

$$\begin{eqnarray} {\rm EW}_{{\rm blend}}^{1} & = & \sum ^{n} f(\log gf_{n},A_{n})\\ \nonumber {\rm EW}_{{\rm blend}}^{2} & = & \sum ^{n} f(\log gf_{n},A_{n})\\ \nonumber \vdots & & \vdots \\ \nonumber {\rm EW}_{{\rm blend}}^{N} & = & \sum ^{n} f(\log gf_{n},A_{n}),\\ &&\nonumber \end{eqnarray} \tag{ A3 }$$

where A_n represents the abundance of the element the line belongs to.

Since the physical conditions and elemental abundance A of each stars are in general different, the values EW^x_blend are different as well. The equation system A3 is solvable when the (N) number of equations is greater than the unknown n+m. The existence and uniqueness of the solution is assured even for lines where the weak-line approximation does not hold, because the first derivative of the curve of growth is always positive.

This means that it is always possible to determine the log gfs of a blended feature and the elemental abundances by using a large enough number of spectra of different stars. In our case we have used two spectra: the Sun and Arcturus.

The following routine is valid when the stellar parameters and elemental abundances of both stars are known. We also consider later the case where elemental abundances are not known. To synthesize the solar spectrum we assumed an effective temperature T_eff = 5777 K, gravity log g = 4.44, microturbulence V_t = 1.0 km s⁻¹, and metallicity [m/H] = 0.00. For Arcturus the atmospheric parameters are taken from Earle et al. (2005) (T_eff = 4340 K, log g = 1.93, V_t = 1.87 km s⁻¹, [m/H] = −0.55 dex).

The synthetic spectra have been obtained assuming solar chemical abundances by Grevesse & Sauval (1998, hereafter GS98) and adopting the grid ATLAS9 model atmospheres of Castelli & Kurucz (2003). The log gfs correction is performed using the following procedure.

• 1.  
  Synthesize the solar and Arcturus spectra from our line list.
• 2.  
  Select the first line of the list.
• 3.  
  Integrate the absorbed flux of the observed and synthesized spectra over an interval spanning 0.4 Å around the center of the line (roughly 1 FWHM, which takes into account most of the absorbed flux) and calculate the normalized EW residual (NEWR), defined as follows

  $$\begin{equation} {\rm NEWR}=\frac{(F_{{\rm synt}}-F_{{\rm obs}})}{F_{{\rm obs}}}, \end{equation} \tag{ A4 }$$

  where F_synt and F_obs are the flux of the synthetic and the observed spectrum, respectively. For the Sun and Arcturus we write (NEWR_Sun) and (NEWR_Arc).
• 4.  
  If NEWR_Sun < −0.05 then add +0.01 to the log gf, else if NEWR_Sun > 0.05 then add −0.01 to the log gf (i.e., match the Sun first). Go to step 7.
• 5.  
  If |NEWR_Sun| < 0.05 and NEWR_Arc < −0.05 then add −0.01 to the log gf, else if |NEWR_Sun| < 0.05 and NEWR_Arc > 0.05 then add +0.01 to the log gf (e.g., if the Sun matches then try to match Arcturus). Go to step 7.
• 6.  
  If |NEWR_Sun| and |NEWR_Arc| are both <0.05 then a valid log gf has been reached.
• 7.  
  Select the next line from the line list and return to step 3.

The routine is iterated until no more improvement is seen. For about 10 lines a good match could not be reached. Several reasons may be behind this: bad continuum correction, lines located over the Ca ii lines where the fit is inaccurate, misidentified lines, or unrecognized blends. For these lines, we did a visual check and manually corrected the log gf values. The correction is good enough to derive gf values that yield abundance errors per measured line smaller than 0.2 dex in average for both the Sun and Arcturus.

Since the abundance of Arcturus is not known for all elements, we decided to assume an initial abundance for all the elements ([X/H] = [m/H]) and find the right abundances with the following procedure.

• 1.  
  Synthesize the Sun's and Arcturus's spectra.
• 2.  
  Correct the log gf-values by running the above correction routine until no more improvements are possible.
• 3.  
  Check the distribution of the NEWRs for each element separately: if a positive/negative offset is present, decrease/increase its abundance.
• 4.  
  Synthesize the solar and Arcturus's spectra with the new log gfs and abundances.
• 5.  
  Return to step 2.

This algorithm is repeated until there are no more improvements. The results of this procedure are the new log gfs and Arcturus's chemical abundances (listed in Table 3).

A.2.1. Multiplets Treatment

In the RAVE/Gaia wavelength range there are features that belong to atomic multiplets. An interesting characteristic of these multiplets is that they share the same excitation potential χ and have small differences in wavelength. In this case, obtaining a proper log gf for the individual lines is problematic.

We work around this problem in the following way: consider the two Al i lines at λ₁ = 8773.896 Å and λ₂ = 8773.897 Å that share the same excitation potential χ = 4.021947 eV. The two lines have unknown equivalent widths EW₁ and EW₂ but the sum

$$\begin{equation} {\rm EW}_{{\rm Tot}}={\rm EW}_1+{\rm EW}_2, \end{equation} \tag{ A5 }$$

is measured and well known. We can write the curve of growth as

$$\begin{equation} \frac{{\rm EW}}{\lambda }=\frac{C\cdot A_{Al} \cdot (gf)\cdot \lambda }{10^{\theta \chi } \cdot \kappa _{\nu }}, \end{equation} \tag{ A6 }$$

where A_Al is the aluminum abundance and the other quantities are as in Equation (A1). Substituting Equation (A6) into Equation (A5) and adopting the approximation λ₁ ≃ λ₂ = λ we obtain

$$\begin{equation} \frac{{\rm EW}_{{\rm Tot}}}{\lambda }=\frac{{\rm EW}_1+{\rm EW}_2}{\lambda }=\frac{C \cdot A_{Al} \cdot (gf_1+gf_2) \cdot \lambda }{10^{\theta \chi } \cdot \kappa _{\nu }}. \end{equation} \tag{ A7 }$$

Using the logarithmic form, this equation then becomes

$$\begin{equation} \log \frac{{\rm EW}_{{\rm Tot}}}{\lambda }=\log C+\log A+ \log ((gf_1+gf_2) \cdot \lambda)-\theta \chi -\log \kappa _{\nu }, \end{equation} \tag{ A8 }$$

which suggests that an appropriate functional form for a "dummy" variable log gf_d would be log (gf₁ + gf₂). Even if this does not allow us to disentangle the two log gfs, we can consider the multiplet as if it were one single line and correct the log gf_fic; it will be a "dummy" value but it results in giving the correct abundance of Al. This correction has also been applied to three Mg i multiplets, centered at wavelengths 8717.814, 8736.020, and 8773.896 Å.

In order to test the accuracy of this procedure, we have measured by hand the EWs of 100 atomic lines in the solar spectrum. Figure 21 presents the difference in abundances obtained by using the Kurucz gf values (open points) and the corrected ones (solid points). Our correction reduced the standard deviations of the abundance residuals from σ = 0.63 dex to σ = 0.15 dex and their averages from +0.16 to −0.05.

Zoom In Zoom Out Reset image size

We also compared our results to the log gf by Kirby et al. (2008), who corrected the gf values by applying a similar method, i.e., by synthesizing the solar and Arcturus's spectra and comparing the synthetic spectra with the observed ones. A comparison between our and their results is shown in Figure 22 for the 89 corrected lines in common: on average our log gf are lower by −0.05 dex. This is very likely due to different solar elemental abundances adopted: Kirby adopts the Anders & Grevesse (1989) abundances which are on average higher than Grevesse & Sauval (1998), which we adopt.

Zoom In Zoom Out Reset image size

One more test involves comparing our results for Arcturus's elemental abundances with abundances quoted in the literature. This is done in Table 3 where our elemental abundances are compared with those of Luck & Heiter (2005), Thévenin (1998, hereafter T98), and Fulbright et al. (2007, hereafter F07). From the consistency of these tests, we are confident that our computed log gf values are accurate and that our procedure yields robust and reliable elemental abundance estimates. In Table 4 we outline the final RAVE line list with the obtained log gf.