arXiv:1104.3576v1 [astro-ph.GA] 18 Apr 2011
The RAdial Velocity Experiment (RAVE): Third Data Release
A. Siebert1, M. E. K. Williams2, A. Siviero2,3, W. Reid4, C. Boeche2, M. Steinmetz2, J.
Fulbright5, U. Munari3, T. Zwitter6,7, F. G. Watson8, R. F. G. Wyse5, R. S. de Jong2, H.
Enke2, B. Anguiano2, D. Burton8,9, C. J. P. Cass8, K. Fiegert8, M. Hartley8, A. Ritter4,
K. S. Russel8, M. Stupar8, O. Bienaym´ e1, K. C. Freeman9, G. Gilmore10, E. K. Grebel11,
A. Helmi12, J. F. Navarro13, J. Binney14, J. Bland-Hawthorn15, R. Campbell16, B.
Famaey1, O. Gerhard17, B. K. Gibson18, G. Matijeviˇ c6, Q. A. Parker4,8, G. M. Seabroke19,
S. Sharma15, M. C. Smith20,21, E. Wylie-de Boer9
– 2 –
We present the third data release of the RAdial Velocity Experiment (RAVE)
which is the first milestone of the RAVE project, releasing the full pilot survey.
1Observatoire astronomique de Strasbourg, Universit´ e de Strasbourg, CNRS, UMR 7550, 11 rue de
l’universit´ e, 67000, Strasbourg, France
2Leibniz-Institut f¨ ur Astrophysik Potsdam (AIP), An der Sternwarte 16, D-14482, Potsdam, Germany
3INAF Osservatorio Astronomico di Padova, 36012 Asiago (VI), Italy
4Department of Physics and Astronomy, Faculty of Sciences, Macquarie University, NSW 2109, Sydney,
5Johns Hopkins University, Departement of Physics and Astronomy, 366 Bloomberg center, 3400 N.
Charles St., Baltimore, MD 21218, USA
6University of Ljubljana, Faculty of Mathematics and Physics, Jadranska 19, 1000 Ljubljana, Slovenia
7Center of excellence SPACE-SI, Aˇ skerˇ ceva cesta 12, 1000 Ljubljana, Slovenia
8Australian Astronomical Observatory, P.O. box 296, Epping, NSW 1710, Australia
9Research School of Astronomy and Astrophysics, Australian National University, Cotter Rd., ACT,
10Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge, CB3 OHA, UK
11Astronomisches Rechen-Institut, Zentrum f¨ ur Astronomie der Universit¨ at Heidelberg, M¨ onchhofstr. 12-
14, D-69120, Heidelberg, Germany
12Kapteyn Astronomical Institut, University of Groningen, Landleven 12, 9747 AD, Groningen, The
13Department of Physics and Astronomy, University of Victoria, P.O. box 3055, Victoria,BC V8W 3P6,
14Rudolf Peierls Center for Theoretical Physics, University of Oxford, 1 Keeble Road, Oxford, OX1 3NP,
15Sydney Institute for Astronomy, School of Physics A28, University of Sydney, NSW 2006, Australia
16Western Kentucky University, Bowling Green, Kentucky, USA
17Max-Planck-Institut f¨ ur extraterrestrische Physick, Giessenbachstrasse, D-85748, Garching, Germany
18Jeremiah Horrocks Institute, University of Central Lancashire, Preston, PR1 2HE, UK
19Mullard Space Science Laboratory, University College London, Holmbury St Mary, Dorking, RH5 6NT,
20Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing, China
21National Astronomical Observatories, Chinese Academy of Sciences, Beijing, China
– 3 –
The catalog contains 83072 radial velocity measurements for 77461 stars in the
southern celestial hemisphere, as well as stellar parameters for 39833 stars. This
paper describes the content of the new release, the new processing pipeline, as well
as an updated calibration for the metallicity based upon the observation of addi-
tional standard stars. Spectra will be made available in a future release. The data
release can be accessed via the RAVE webpage: http://www.rave-survey.org.
Subject headings: catalogs, surveys, stars: fundamental parameters
A detailed understanding of the Milky Way, from its formation and subsequent evolu-
tion, to its present-day structural characteristics, remains key to understanding the cosmic
processes that shape galaxies. To achieve such a goal, one needs access to multi-dimensional
phase space information, rather than restricted (projected) properties - for example, the
three components of the positions and the three components of the velocity vectors for a
given sample of stars. Until a decade ago, only the position on the sky and the proper
motion vector was known for most of the local stars. Thanks to ESA’s Hipparcos satel-
lite (Perryman et al. 1997), the distance to more than 100000 stars within a few hundred
parsecs has been measured, allowing one to recover precise positions in the local volume (a
sphere roughly 100 pc in radius centered on the Sun). However, the 6thdimension of the
phase space was still missing until recently, when Nordstr¨ om et al. (2004) and Famaey et al.
(2005) released radial velocities for subsamples of respectively 14000 dwarfs and 6000 giants
from the Hipparcos catalog.
In recent years, with the availability of multi-object spectrometers mounted on large
field-of-view telescopes, two projects aiming at measuring the missing dimension have been
initiated: RAVE and SEGUE, the Sloan Extension for Galactic Understanding and Explo-
ration. SEGUE uses the Sloan Digital Sky Survey (SDSS) instrumentation and acquired
spectra for 240000 faint stars, 14 < g < 20.3, in 212 regions sampling three quarters of the
sky. The moderate resolution spectrograph (R∼ 1800) combined with coverage of a large
spectral domain (λλ = 3900 − 9000˚ A) allows one to reach a radial velocity accuracy of
σRV∼ 4kms−1at g ∼ 18 and 15kms−1at g = 20 as well as an estimate of stellar atmospheric
parameters. The SEGUE catalog was released as part of the SDSS-DR7 and is described
in Yanny et al. (2009). Altogether, the SDSS-I and II projects provide spectra for about
490000 stars in the Milky Way. As of January 2011, the SDSS Data Release 8 marks the first
release of the SDSS-III survey (Eisenstein et al. 2011). This release (SDSS-III collaboration
2011) provides 135040 more spectra from the SEGUE-2 survey targeting stars in the Milky
– 4 –
RAVE commenced observations in 2003 and has thus far released two catalogs : DR1
in 2006 and DR2 in 2008 (Steinmetz et al. 2006; Zwitter et al. 2008), hereafter Papers I
and II, respectively. The survey targets bright stars compared to SEGUE, 9 < I < 12,
in the southern celestial hemisphere, making the two surveys complementary. The RAVE
catalogs contain respectively 25000 and 50000 measurements of radial velocities plus stellar
parameter estimates for about half the catalog for DR2. RAVE uses the 6dF facility on
the Anglo-Australian Observatory’s Schmidt telescope in Siding Spring, Australia. This
instrument allows one to collect up to 150 spectra simultaneously at an effective resolution
of R = 7500 in a 385˚ A wide spectral interval around the near-infrared calcium triplet
(λλ8410 − 8795˚ A). The CaII triplet being a strong feature, RAVE can measure radial
velocities with a median precision of about 2kms−1.
RAVE is designed to study the signatures of hierarchical galaxy formation in the Milky
Way and more specifically the origin of phase space structures in the disk and inner Galac-
tic halo. Within this framework, Williams et al. (2011) discovered the Aquarius stream,
while Seabroke et al. (2008) studied the net vertical flux of stars at the solar radius and
showed that no dense streams with an orbit perpendicular to the Galactic plane exist in
the solar neighborhood, supporting the revised orbit of the Sagittarius dwarf galaxy by
Fellhauer et al. (2006). On the other hand, Klement et al. (2008) looked directly at stellar
streams in DR1 within 500 pc of the Sun and identified a stream candidate on an extreme
radial orbit (the KFR08 stream), in addition to three previously known phase space struc-
tures (see also Kiss et al. 2011, for an analysis of known moving groups). A later analysis
of the DR2 catalog by the same authors, using the newly available stellar atmospheric pa-
rameters in the catalog, revised their detection of the KFR08 stream, the stream being now
only marginally detected (Klement et al. 2011).
If RAVE is designed to look at cosmological signatures in the Milky Way, it is also well-
suited to address more general questions. For example, Smith et al. (2007) used the high
velocity stars in the RAVE catalog to revise the local escape speed, refining the estimate
of the total mass of the Milky Way. Co¸ skunoˇ glu et al. (2011) used RAVE to revise the
motion of the Sun with respect to the LSR, while Siebert et al. (2008) measured the tilt
of the velocity ellipsoid at 1 kpc below the Galactic plane. Veltz et al. (2008) combined
RAVE, UCAC2, and 2MASS data towards the Galactic poles to revisit the thin-thick disk
decomposition and Munari et al. (2008) used RAVE spectra to confirm the existence of the
λ8648˚ A diffuse interstellar band and its correlation with extinction.
RAVE, being a randomly-selected, magnitude-limited survey, possesses content repre-
sentative of the Milky Way for the specific magnitude interval, in addition to peculiar and
– 5 –
rare objects within the same interval. Together, this makes RAVE a particularly useful cat-
alog to study the origin of the Milky Way’s stellar populations. For example, Ruchti et al.
(2010) studied the elemental abundances of a sample of metal-poor stars from RAVE to
show that direct accretion of stars from dwarf galaxies probably did not play a major role
in the formation of the thick disk, a finding corroborated by the study of the eccentricity
distribution of a thick disc sample from RAVE (Wilson et al. 2011). Also, Matijeviˇ c et al.
(2010) used RAVE to study double lined binaries using RAVE spectra while Fulbright et al.
(2010) used RAVE to detect very metal poor stars in the Milky Way. It also happens that
bright objects from nearby Local Group galaxies are observed; Munari et al. (2009), for
example, identified eight luminous blue variables from the Large Magellanic Cloud in the
So far RAVE has released only radial velocities and stellar atmospheric parameters. To
really gain access to the full 6D phase space, the distance to the stars remains a missing, yet
important, parameter, unless one focuses on a particular class of stars, such as red clump
stars (see for examples Veltz et al. 2008; Siebert et al. 2008). Combining the photometric
magnitude from 2MASS and RAVE stellar atmospheric parameters, Breddels et al. (2010)
derived the 6D coordinates for 16,000 stars from the RAVE DR2, allowing a detailed in-
vestigation of the structure of the Milky Way. This effort of providing distances for RAVE
targets was later improved by Zwitter et al. (2010), taking advantage of stellar evolution
constraints, and by Burnett et al. (2011), by using the Bayesian approach described in
Burnett & Binney (2010). The distance estimates have been used by Siebert et al. (2011)
to detect non-axisymmetric motions in the Galactic disk. These works will be extended to
DR3, distributed in a separate catalog, and will provide a unique sample to study the details
of the formation of the Galaxy. Moreover, for the bright part of the RAVE sample, the
signal-to-noise ratio per pixel allows one to estimate fairly accurate elemental abundances
from the RAVE spectra. This catalog containing of order 104stars (Boeche et al., in prep)
will provide a unique opportunity to combine dynamical and chemical analyses to understand
In this paper we present the 3rddata release of the RAVE project, releasing the radial
velocity data and stellar atmospheric parameters of the pilot survey program that were
collected during the first three years of operation, therefore DR3 includes the data collected
for DR1 and DR2. The spectra are not part of this release. These data were processed using
a new version of the processing pipeline. This paper follows the first and second data releases
described in Papers I and II. The pilot survey release is the last release relying on the original
input catalog, based on the Tycho-2 and SuperCosmos surveys. Subsequent RAVE releases
will be based on targets selected from the DENIS survey I-band. The paper is organized as
follows: Section 2 presents the new version of the processing pipeline which calculates the
– 6 –
radial velocities and estimates the stellar atmospheric parameters. Section 3 presents the
validation of the new data, as well as the updated calibration relation for metallicity, while
Section 4 describes the DR3 catalog.
2. A revised pipeline for stellar parameters
In Papers I and II we described in detail the processing pipeline used to compute the
radial velocities and the stellar atmospheric parameters, making use of a best-matched tem-
plate to measure the radial velocities and set the atmospheric parameters reported in the
catalog. This pipeline performs adequately for well-behaved spectra, permitting the mea-
surement of precise radial velocities, and we showed in Paper II that the stellar atmospheric
parameters Teff, logg, and [m/H] can be estimated. However, to compare the RAVE [m/H] to
high resolution measurements [M/H]1, a calibration relation must be used. Also, in the case
where a RAVE spectrum suffers from (small) defects, the stellar atmospheric parameters are
less well-constrained. We therefore set out to improve the pipeline, while still maintaining its
underlying computational techniques. This section reviews the modifications of the RAVE
pipeline, which is otherwise fully described in Paper II.
The RAVE pipeline for DR1 and DR2 relied on the Munari et al. (2005) synthetic
spectra library based on ATLAS 9 model atmospheres. This library contains spectra with
three different values for the micro-turbulence µ of 1, 2, and 4kms−1. However, the library is
well-populated only for the µ = 2kms−1value, about 3000 spectra having µ = 1 or 4kms−1,
compared to ∼ 55000 having µ = 2kms−1.
For this new data release (DR3), new synthetic spectra for intermediate metallicities
were added in order to provide a more realistic spacing towards the densest region of the
observed parameter space and so remove biases towards low metallicity. The new grid has
[m/H] = −2.5, −2.0, −1.5, −1.0, −0.8, −0.6, −0.4, −0.2, 0.0, 0.2, 0.4, and 0.5dex.
We also restricted the library to µ = 2kms−1, discarding all other micro-turbulence
values. This does not impact the quality of the measured stellar parameters as, at our
S/N level and resolution, we are unable to constrain the micro-turbulence, and the pipeline
1Throughout this paper, [m/H] refers to the metallicity obtained using the RAVE pipeline while [M/H]
refers to metallicity obtained using detailed analyses of high resolution echelle spectra.
– 7 –
usually converges on the most common micro-turbulence value in the library (µ = 2kms−1).
Furthermore, since the nominal resolution of the 6dF instrument does not allow us to
measure precisely the rotational velocity of the star, we chose to restrict the Vrotdimension,
removing six of the lower Vrotvalues (0, 2, 5, 15, 20, and 40kms−1), retaining only the 10,
30, 50kms−1, and higher, velocities.
Removing one dimension of the parameter space and reducing the rotational velocity
dimension helps to stabilize the solution and allows us to lower the number of neighboring
spectra used for the fit. We lower this number from 300 to 150. As for Paper II, the Laplace
multipliers for the penalisation terms were set using Monte Carlo simulations. We increased
the Laplace multiplier handling the penalisation on the sum of the weights, which constrains
the level of the continuum to unity for continuum normalized spectra, as a 0.3% offset was
not uncommon in the previous pipeline.
To date, the processing pipeline used S/N estimates as described in Paper I. However,
this S/N estimate tends to underestimate the true S/N and is less dependent on the true
noise than it is on the weather conditions or spectrum defects, such as fringing (see Paper
II). In Paper II, a new S/N estimate, S2N, was presented based on the best fit template
but was not used by the pipeline as it was an a posteriori estimate. We showed that S2N is
closer to the true S/N.
Because of the new continuum correction procedure (see Section 2.3), the S/N must be
computed correctly before the continuum correction is applied. Therefore, it must be known
prior to the processing. We thus developed an algorithm to measure the S/N of a spectrum
in which no flux information is used. This new S/N estimate, STN, is obtained using the
observed spectrum (no continuum normalization applied) as follows:
1 Smooth the observed spectrum s(i), with i the pixel index, to produce a smoothed
spectrum f(i). This smoothing is done with a smoothing box three pixels long.
2 Compute the residual vector R(i) = f(i) − s(i) and its rms σ.
3 Remove from s pixels that diverge from f by more than 2σ.
4 Smooth the clipped spectrum as above to form a new smoothed spectrum f and repeat
the clipping process until convergence.
– 8 –
5 Compute the local standard deviation σl(i) using pixels i − 1, i and i + 1.
6 Compute STN= median(s(i)/σl(i))/1.62.
The factor of 1.62 is set using numerical realizations of a Poisson noise. As shown in the
left panel of Figure 1, S/N and STN are on a 1:1 relation. However, in a real spectrum,
instrument noise also contributes to the residuals and we expect an additional normalization
factor. The S2N value as computed in Paper II follows closely the true S/N. Hence, to assess
the validity of the STN measurement, we compared it to the S2N in Paper II (Figure 1 right
panel). A correction factor of 0.58 for S2N is found to produce a 1:1 relation between the
two measurements, a correction that we apply in the pipeline.
0 50 100 150
0 50 100 150
3635 noisy synthetic spectra
3752 RAVE spectra
Fig. 1.— Comparison of the various signal-to-noise estimates. Left panel: signal-to-noise
STN compared to the original RAVE S/N. Right panel: comparison of the scaled STN to
S2N, the signal-to-noise estimator constructed for DR2.
In the low S/N regime (S/N < 10), the metallic lines are no longer visible. In this case,
[m/H] measurements converge to the highest allowed value ([m/H] = +0.5 dex) which gives
the lowest possible χ2value, i.e. the algorithm fits the noise. In this regime, the stellar
– 9 –
parameters are not reliable and are therefore not published. In the intermediate regime
10 < S/N < 50, a correlation between [m/H] and S/N is observed in the RAVE data.
While some of the above correlation is understood and arises from the change of the
underlying stellar content as one moves further away from the plane and the S/N simulta-
neously decreases2, some part of this correlation arises from to the continuum normalization
failing to recover the proper continuum level. The former pipeline uses the IRAF continuum
task with asymmetric rejection parameters (1.5σ for the low rejection level and 3.0σ for the
high rejection level). While these parameters are well-suited for the high S/N regime (> 60),
at low S/N they tend to produce an estimated continuum that is too high. This is due to
the routine considering the spikes below the continuum as spectral lines when, in fact, they
are mainly due to noise.
We ameliorate this problem by using a low rejection value that is a function of S/N. This
rejection level must be close to 1.5 for high S/N spectra and larger for low S/N. Numerical
tests indicate that using the following formula
lowrej= 1.5 + 0.2exp
with σSTN = 16, from the top left panel of Figure 2, reduces significantly the continuum
normalization problem. The top panels in Figure 2 show the mean residual between the
observed continuum-normalized spectra and best fit template as a function of S/N, before
and after the change in the low rejection level, while the bottom panels present the resulting
distributions of [m/H] as a function of S/N.
The new continuum normalization reduces significantly the correlation between metal-
licity and S/N, while no trend in the residual as a function of S/N remains. This indicates
that the new continuum normalization algorithm performs adequately, although a weak cor-
relation is still seen in the metallicity versus S/N (∼ 0.1dex per 100 in S/N).
2.4.Masking bad pixels
Approximately 20% of RAVE spectra suffer from defects such as fringing or residual
cosmic rays, which cannot be removed by the automatic procedure we use to reduce our
data. While residual cosmic rays do not affect the determination of the stellar atmospheric
2The exposure time being fixed, a lower S/N indicates a fainter magnitude.
– 10 –
Fig. 2.— Top panels: average residuals best fit template−observed spectra for 4,684 RAVE
spectra as a function of S/N. Bottom panels: [m/H] distributions as a function of S/N. The
left columns are for the previous version of the continuum normalization algorithm while the
right column includes the low rejection level being a function of S/N. The gain from the new
continuum normalization is clear from these figures: the correlation between metallicity and
S/N is strongly reduced, while the residuals do not show any correlation with S/N. The thick
black line represents the STN limit below which atmospheric parameters are not published
in the RAVE catalog.
– 11 –
parameters (these are similar to emission lines, which are not taken into account in the tem-
plate library), fringing results in poor local continuum normalization, leading to inaccurate
Regions strongly affected by fringing are difficult to detect prior to the processing, but
we can make use of the best fit template to estimate whether a spectrum suffers from such
a continuum distortion and therefore whether the atmospheric parameter determination is
likely to be in error.
To estimate the fraction of a spectrum contaminated by continuum distortions, we
compute the reduced χ2(i) along the spectrum in a box 21 pixels wide centered on the
pixel i. We then also compute the mean difference S(i) between the best-fit template and
the observed spectrum in the same box. If χ2(i) > 2 and S(i) > 2/STN, a systematic
difference between the template and the observed spectrum exists. The corresponding region
of the spectrum is then flagged as a defect. The fraction of good pixels in each spectrum
is then recorded and given in the RAVE catalog (see MaskFlag in Table 12). From visual
inspection, we find that when the number of bad pixels is larger than 30% then the spectrum
is problematic and the stellar parameters should be treated with caution. Figure 3 shows
different examples of real RAVE spectra where a significant fraction of the spectrum is
marked as defect.
2.5. Improving the zero-point correction
As explained in previous papers (e.g., Paper I), thermal instabilities in the spectrograph
room induce zero point shifts of the wavelength solution that depend on the position along
the CCD (e.g., fiber number). This results in instabilities of the radial velocity zero-point.
To correct the final radial velocities for this effect, the processing pipeline uses available
sky lines in the RAVE window and fits a low-order polynomial (3rdorder) to the relation
between sky radial velocity and fiber number. This 3rdorder polynomial defines the mean
trend of zero point offsets and provides the zero point correction as a function of fiber
number3. However, in some cases, a low-order polynomial is not the best solution and a
3The zero-point correction could in principle be obtained directly from the radial velocity of the sky lines.
However the radial velocity measured from the sky lines suffers from significant errors while the trend of
the zero-point offset with respect to the fiber number due to thermal changes is expected to be a smooth
function of fiber number. Therefore, using a smooth function to recover the mean trend is better suited to
correct for zero point offsets. Tests have shown that using a 3rdorder polynomial provides in most cases the
best solution (see paper I).
– 12 –
Fig. 3.— Example of five RAVE spectra with regions marked as problematic by the MASK
code. The regions marked in grey are recognized as suffering from poor continuum normal-
ization. If more than 30% of the spectrum is marked by the code, the observation is flagged
as problematic by the pipeline. The normalized fluxes are in arbitrary units and a vertical
offset is added between the spectra for clarity.
– 13 –
constant shift should be used instead. In former releases, these cases were corrected by hand
in the catalog. In this release, we introduced a new zero-point correction routine to the
processing pipeline that is able to select which correction should be applied, automatically.
The zero-point correction now computes both the cubic correction, using the 3rdorder
polynomial, and the constant correction. It then computes the mean and standard deviation
between the measured sky radial velocities and the corrections for the entire field and for three
regions in fiber number that are contiguous on the CCD (fibers 1−50, 51−100, 101−150).
For each region, the cubic fit is used unless any of these four conditions apply:
- there are less than two sky fibers in that region, to avoid under-constrained fits,
- the mean in that region for the constant correction is better than the corresponding
mean for the cubic fit,
- the standard deviation for the cubic correction is greater than 5kms−1, which is the
case for noisy data,
- the maximum difference between the constant correction and the cubic correction is
larger than 7kms−1.
We tested the new procedure, together with other options, against pairs of repeat ob-
servations. The results are presented in Table 1. They show clearly that the new procedure
performs better than the previous version in terms of dispersion, while the mean difference
is unchanged. While the constant term correction appears better in this table, the left panel
in Fig. 4 shows that the distribution of the residuals is less peaked than for the cubic correc-
tion. In addition, the mean-square-error, defined as MSE = E[(RV − RVfit)2], shows a net
decrease with the new fitting procedure compared to a constant shift. This indicates that
for the general case, a constant correction for the entire field will result in a larger dispersion
and hence a larger zero-point offset residual. This gives us confidence in the use of the new
3.Calibration and validation
3.1. Radial velocity
3.1.1.Internal error distribution
RAVE obtains its radial velocity from a standard cross-correlation routine. For each
radial velocity measurement the associated error, eRV, gives the internal error due to the
– 14 –
Mean residual between fit and sky RVs (km/s)
MSE between fit and sky RVs (km/s)^2
Fig. 4.— Left: mean residual between the fit and the sky radial velocity for three different
fitting functions. A constant shift (black histogram), the cubic fit used in DR1 and DR2
(red histogram), and the new fitting procedure (blue histogram). Right: associated mean-
Table 1: Radial velocity difference between pairs of repeat observations using different zero-
point correction solutions. The old correction is a combination of cubic fit and corrections
applied by hand. The number of pairs used is 25,172.
Methodµ (kms−1)σ (kms−1)
No correction 0.382.74
Old correction 0.232.49
New correction 0.232.22
95% ( kms−1)
– 15 –
fitting procedure. Figure 5 presents the distribution of eRV per 0.2kms−1bin for the data
new to each RAVE release. While first year data are of lower quality due to the second-
order contamination of our spectra, second and third year data are of equal quality with
a mode at 0.8kms−1, a median radial velocity error of 1.2kms−1, and 95% of the sample
having internal errors better than 5kms−1. Comparing these values to the old version of
the pipeline used for DR1 and DR2 (see Table 2 and Fig. 9 of Paper II), the new pipeline
marginally improves the internal accuracy with a gain of ∼ 0.1kms−1for the mode and the
median radial velocity error.
0 2 4 6 8 10
0 2 4 6 8 10
Number per 0.2 km/s bin
Fig. 5.— Distribution of the radial velocity error (eRV) in the 3rddata release. Top: number
of stars with eRV in 0.2kms−1bins for first-year data (dash-dotted line), second-year data
(dashed line), and third-year data (full line). Bottom: cumulative distribution of the eRV.
The dotted lines mark respectively 50, 68 and 95% of the samples.
The aforementioned error values represent the contribution of the internal errors to the
RAVE error budget. External errors are also present and are partially due to the zero-point
correction which corrects only a mean trend, not including the fiber-to-fiber variations. The
contribution of the external errors is obtained using external datasets and is discussed in
– 16 –
Our internal error budget is the sum of (i) the error associated with the evaluation of
the maximum of the Tonry-Davis correlation function, and (ii) the contribution from the
zero-point error. The first contribution is given by the pipeline (§ 3.1.1). The magnitude of
the second term can be obtained from the analysis of the re-observed targets as, for a given
star whose apparent magnitude is fixed, the radial velocity is constant (if the star is not a
binary) and the internal errors are the main source of uncertainties.
We therefore use the re-observed stars in the RAVE DR3 catalog, selecting only stars
observed during the second and third year, as they share the same global properties in terms
of observing conditions. Data from the first year of observing are discarded, as they suffer
from second-order contamination which renders the internal error inhomogeneous and can
therefore bias our estimate. We also removed from the sample stars that were observed on
purpose to calibrate our stellar atmospheric parameters, as these are specific bright targets
with high S/N that do not share the random selection function nor the standard observational
protocol of the RAVE catalogue.
The cumulative distribution of the radial velocity difference is presented in the left panel
of Fig. 6 where the solid line represents the full sample of re-observed targets and the dashed
line the sample restricted to individual measurements differing by less than 3σ in a pair.
Since our sample is contaminated by spectroscopic binaries, this selection is compulsory if
one wants to address the error distribution for normal stars but is only a crude approximation
when trying to remove all the binaries in the sample. Applying this cut rejects 6% of the
sample, a value clearly below the expected contamination (see below). Therefore, the errors
estimated from the repeat observations are likely to overestimate the true errors. With this
limitation in mind, from Fig. 6, focusing on the dashed line, one can conclude that 68.2% of
the sample has an error below 2.2kms−1while ∼ 93% of the sample lies below the 5kms−1
To estimate the contribution from the zero-point errors to the total internal error budget,
we computed the distribution of the normalized radial velocity difference, the relative differ-
ence in radial velocity between two observations divided by the square root of the quadratic
sum of the errors on radial velocity. If our measurements were affected only by the random
errors for (i), then the distribution of this normalized radial velocity difference would follow a
Gaussian distribution of zero mean and unit standard deviation. An additional contribution
to the error budget due to a random zero-point error would broaden the distribution and
hence enhance the dispersion of the resulting distribution. The result of this test is presented
in the right panel of Figure 6, where we fitted the sum of two Gaussians to the observed
– 17 –
0 5 10 15
|∆ RV|/sqrt(2) (km/s)
−5 0 5
Number per 0.25 σ bin
Fig. 6.— Left: Cumulative fraction of the radial velocity difference for re-observed RAVE
targets in the Third Data Release. The solid line corresponds to the full sample, and the
dashed line relates to the sample restricted to pairs whose individual measurement differ
by less than 3σ (hence rejecting the spectroscopic binaries with the largest radial velocity
difference). The horizontal lines indicate 50, 68.2, and 95% of the sample. The grey lines
are the expected distributions of the radial velocity difference for Gaussian errors of 1, 2, 3,
4, and 5kms−1from inside out. Right: distribution of the radial velocity difference ∆RV in
units of σ for re-observed targets. The blue line corresponds to our best-fit double Gaussian
model to the distribution. The red dashed lines show the respective contribution of each
– 18 –
The dominant Gaussian distribution corresponds to stars stable in radial velocity. The
width of the associated Gaussian function is 0.83σ, narrower than a normal distribution,
indicating that the internal errors quoted in the catalog are likely overestimated. Our quoted
internal error can therefore be assumed to be an upper bound on the true internal errors,
including the contribution of the zero-point error.
Spectroscopic binary contamination: subsidiarily, the broad Gaussian comprises
spectra with defects (or where the zero-point solution could have diverged) as well as the
contribution from spectroscopic binaries. The fraction of spectra with defects is small in
this sample, as the catalog has been cleaned of fields where the zero-point solution did
not converge. Hence, the relative weight of the two Gaussian functions gives an estimate,
in reality an upper limit, of the contamination level by spectroscopic binaries with radial
velocity variation between observations larger than 1σ in the RAVE catalog. Our best-fit
solution gives a relative contribution for this second population of 26% which allows us to
conclude that the fraction of spectroscopic binaries with radial velocity variations larger
than 2kms−1in the RAVE catalog is less than or equal to 26%. A more detailed analysis of
repeated observations based on 20000 RAVE stars by Matijeviˇ c et al. (2011) gives a lower
limit of 10-15% of the RAVE sample being affected by binarity (see also Matijeviˇ c et al.
2010). However, the time span between repeat observations being biased towards short
periods (days to weeks), long period variations are not detected. The previous estimates
do not take into account this population and a more detailed analysis will be required to
estimate the contribution of long period variables to our survey.
3.1.3. Validation using external datasets
Our external datasets (or, ‘reference’ datasets) comprise data from the Geneva-Copenhagen
Survey (Nordstr¨ om et al. 2004, hereafter GCS), Elodie and Sophie high resolution obser-
vations from the Observatoire de Haute Provence, Asiago echelle observations, and spectra
obtained with the ANU 2.3m facility in Siding Spring. The targeted stars are chosen to cover
the possible range of signal-to-noise conditions and stellar atmospheric conditions. Figure 7
presents the distributions of the reference stars as a function of signal-to-noise S2N, Tefflogg
and [m/H] compared the the RAVE DR3 distributions. While for [m/H] the distribution
ressembles the distribution of the data release, the distribution of logg shows a lack of giant
stars that translates to a reduced peak at temperature below 5000 K compared to the full
DR3 sample. This is due to the GCS sample, our primary source of reference stars, that
contains F and G dwarfs and no giants. For the S2N distribution, we chose to sample almost
uniformly the RAVE S2N interval, top left panel of Fig. 7, which enables us to verify that
– 19 –
signal-to-noise does not impact the quality of our radial velocities (see below).
50 100 150
4000 6000 8000
0 2 4
−2 −1 0
Fig. 7.— Histograms of the distribution of the reference sample (dash-dotted histograms)
and the RAVE DR3 sample (full lines) as a function of signal-to-noise, Teff, logg and RAVE
[m/H]. The dash-dotted histograms are multiplied by a factor 50 to enhance their visibility.
A comparison of the radial velocities obtained by RAVE and the external datasets is
presented in Figure 8, while the detailed values for the comparison for each sample can be
found in Table 2.
With the new version of the pipeline, we find no significant difference for the mean radial
velocity difference compared to DR2. The values for the mean difference and its dispersion
are consistent between these two releases. From the right panel of Figure 8 one sees that
– 20 –
−100 0 100
−5 0 5
Number per 0.25 bin
Fig. 8.— Comparison of RAVE radial velocities to external sources. Left : RVRAVE vs.
RVextfor all the different sources: GCS (red circles), ANU 2.3m (green triangles), Elodie
(blue squares), Sophie (yellow crosses), and Asiago echelle spectra (magenta diamonds). The
black downwards triangles are stars identified as binaries. Right: distribution of the radial
velocity differences divided by the associated errors. The red curve is a Gaussian distribution
with zero mean and σ = 1.
– 21 –
Table 2: Global properties of the comparison of RAVE radial velocities to external datasets
for stars observed during the second and third year of the program. ∆RV is defined as
∆RV = RVext−RVRAVE. The mean deviations and standard deviations are computed using
a sigma clipping algorithm. The second column gives the number of data points used to
compute the mean and σ while the numbers in parenthesis are the total number of stars in
the sample (N1) and the number of unique objects (N2). The last two lines are obtained
after correcting each dataset for the mean deviation.
All but GCS142(201,142)
mean deviation corrected
All but GCS127(201,142)
– 22 –
the distribution of the radial velocity difference divided by the internal errors is wider than
a normal distribution : its dispersion is 1.37σ. We can then estimate the upper limit to the
external error contribution as σext ≤ 0.9kms−1. This is an upper limit as the zero-point
errors of the other sources of radial velocity also contribute to the measured σextand are
The dependency of the radial velocity difference on signal-to-noise ratio is weak, as can
be seen from Figure 9 (top left panel). The mean difference is consistent with no offset, at
all S2N levels. There is a slight tendency for an increase in dispersion at low S2N, but the
dispersion values remain very well-behaved (σ ∼ 1.2kms−1at S2N>100 and σ ∼ 2.0kms−1
for S2N<40). In addition, no strong variation with logg, Teff, or [m/H] is seen, indicating
that our radial velocity solution is stable as a function of stellar type.
3.2.Stellar atmospheric parameters
During the second and third years of its program, RAVE observed 2266 stars more than
once; 1917 stars were observed twice, 256 were observed three times, and 93 were observed
four times. 1391 of these stars have more than one measurement of stellar parameters.
We use these re-observations to estimate the stability and error budget for our estimated
stellar atmospheric parameters. These parameters are the parameters from the synthetic
template spectrum used to compute the final radial velocity. This template is constructed
using a penalized chi-square algorithm where the template spectrum is a weighted sum of
the synthetic spectra of the library of Munari et al. (2005). The weights of the best-match
are obtained by minimization of a χ2plus additional constraints (weights must be positive
and smoothly distributed in the atmospheric parameters space). The algorithm is described
in Paper II.
3.2.1.Internal stability from repeat observations
As a first step, we estimate the stability from the difference in the measured parameters
using, for a given star, the spectrum with the highest S/N as the reference measurement.
The distribution of the stellar parameter differences ∆P, where P may stand for any of the
stellar atmospheric parameters considered, is shown in Figure 10 while Figure 11 presents
the distributions for dwarfs and giants stars respectively. The red curves in each panel are
Gaussian functions whose parameters (mean and standard deviation) are obtained using
an iterative sigma-clipping algorithm. The corresponding mean and standard deviation for
– 23 –
0 50 100 150
4000 5000 6000 7000 8000
0 2 4
Fig. 9.— Radial velocity difference between the RAVE observations and the external sources
as a function of the signal-to-noise ratio S2N (top left), effective temperature (top right),
logg (bottom left), and [m/H] (bottom right) of the RAVE observation. The symbols follow
Figure 8 while the full and dashed thick lines represent the mean and dispersion about the
mean of the radial velocity difference per interval of 10 in S2N, 500 K in Teff, 0.5 dex in logg,
or 0.25 dex in [m/H].
– 24 –
each parameter are reported in Table 3. For all parameters, the mode of the distributions
is consistent with zero, indicating good stability of our atmospheric parameter measure-
ments. The average internal error for the atmospheric parameters can be estimated from
the standard deviation. For Teff one obtains 200 K and 0.3 dex for logg, while the [m/H]
and [α/Fe] distributions show a dispersion of 0.2 and 0.1 dex respectively. These values
must be regarded as underestimates of the true errors as they do not include external errors
such as the inadequacy of the template library in representing real spectra or variations in
the abundances of the chemical species with respect to the solar abundances (using but one
value of the α-enhancement).
In Fig. 10 the distributions of Teff, [m/H] and [α/Fe] are relatively symmetric although
not Gaussian. The distribution of logg is less symmetric and that of Vrotis very skew. Since
our reference measurements are the spectra with the highest S/N, symmetry indicates that
there is no strong bias in the atmospheric parameter estimation as one reduces the signal-
to-noise ratio: a systematic effect with the S/N would imply that as one lowers the S/N the
measured parameters would be either higher or lower than the reference value.
For Vrot, a systematic effect is likely. As one lowers the S/N, the wings of the spectral
lines become more affected by the noise, making the lines appear narrower, hence mimicking
a lower Vrot. The same effect applies to logg.
Table 3: Standard RAVE errors on stellar atmospheric parameters from repeat observations
for the full sample of re-observed stars. The mean and standard deviations are computed
using an iterative sigma-clipping algorithm and ∆P = Pref− Pstar.
Internal errors on the atmospheric parameters depend on the physical condition of the
star, logg being better constrained for giants and Teff for cool stars. The internal errors,
as defined in Paper II, depend mostly on the algorithm used and the grid spacing of the
synthetic spectra for these two parameters. Neither has been modified in the new version of
the pipeline. Hence, the internal errors for the different parameters remain unchanged and
upper limits for these errors are presented in Fig. 19 in Paper II. However, using re-observed
RAVE stars, one is able to refine this estimate based on the scatter of the atmospheric
parameter measurements in various Teff and logg intervals. These refined estimates are
– 25 –
−1000 −500 0 500 1000
Number per 50 K bin
−2−1 0 1 2
Number per 0.1 dex bin
−1 0 1
Number per 0.05 dex bin
Number per 0.02 dex bin
−50 0 50
Number per 2 km/s bin
Fig. 10.— Distributions of the difference in the measured stellar atmospheric parameters in
re-observed targets. The spectrum with highest S/N for a given star is used as reference.
The red lines in the different panels correspond to a Gaussian function whose parameters
(mean and dispersion) are obtained using an iterative sigma-clipping algorithm (see Table 3).
– 26 –
−1000 −500 0 500
Number per 50 K bin
−2−1 0 1 2
Number per 0.1 dex bin
−1 0 1
Number per 0.05 dex bin
Number per 0.02 dex bin
−40−20 0 20 40
Number per 2 km/s bin
Fig. 11.— Same as Fig. 10 but for the sub-samples of dwarf stars (top curves) and giant
stars (bottom curves). The samples are selected according to logg using the separating line
logg = 3.5 dex. The histograms for dwarf stars are shifted upwards by 100 counts per bin
– 27 –
presented in Table 4 where a smooth-averaging procedure is used to compute the dispersion
at a given grid point. Only grid points with three or more repeated observations are given
in the table.
Table 4: Dispersion in Teff (K), logg (dex) and [m/H] (dex) as a function of Teffand logg.
The dispersions are computed by smooth-averaging sigmas in individual grid points. Only
grid points where three or more repeated objects are present are quoted.
4000(logg)0.07 0.160.19 0.21
([m/H])0.06 0.08 0.080.09
4500 (logg) 0.12 0.180.20
([m/H]) 0.070.09 0.08
– 28 –
3.2.2.Effect of the correlations between atmospheric parameters
In Paper II, we showed that the method we use to estimate the stellar atmospheric
parameters introduces correlations in the errors of the recovered parameters. Here, we use
the re-observations of standard RAVE program stars to estimate the amplitude of these
correlations. The results of these tests are presented in Figure 12 where the contours in each
panel contain 30, 50, 70, and 90% of the total sample. Looking at the different panels, a
clear correlation is observed between the deviations in Teff, logg ,and [m/H] while deviations
in [α/Fe] are only correlated with deviations in [m/H]. Vroton the other hand does not show
any correlation, regardless of the atmospheric parameter considered. Since the correlation
between logg and [m/H] is broader than between logg and Teff, it is likely that errors on Teff
are the primary source of errors, and that these errors propagate to the other atmospheric
These correlations indicate that the true [M/H] will be a function of all the parameters,
except for Vrot. The correlation with logg being weaker than that with Teffand [m/H], the
true calibration relation might be independent of logg or at least, we expect logg to play a
secondary role in the estimation of the true [M/H]. This will be studied more deeply in the
3.2.3.Comparison to external data
In the previous paragraphs, we checked the consistency of the RAVE atmospheric-
parameter solutions and the correlations that exist between these parameters. The consis-
tency of the atmospheric parameters is satisfactory given our medium resolution (R ∼ 7500)
and our small wavelength interval. The dispersions around the reference values are ∼ 200 K
for Teff, 0.3 dex for logg, 0.2 dex for [m/H], and 0.1 dex for [α/Fe], with no significant
The next step is to compare our measured atmospheric parameters with independent
measurements. As for DR2, RAVE stars are generally too faint to have been observed in
other studies from the literature. We therefore used custom RAVE observations of bright
stars from the literature4as well as high-resolution observations of bright RAVE targets to
construct our calibration sample. This sample comprises four different sources of atmospheric
4These stars are not part of the original input catalog but are added to the observing queue to permit
the validation of the RAVE atmospheric parameters.
– 29 –
∆Teff(K) ∆logg (dex)∆[m/H] (dex) ∆[α/Fe] (dex)
−1000 −500 0 500 1000
−2 −1 0 1 2
−1 0 1
−1000 −500 0 500 1000
−2 −1 0 1 2
−1 0 1
−1000 −500 0 500 1000
−2 −1 0 1 2
−1000 −500 0 500 1000
Fig. 12.— Correlation between the stellar atmospheric parameters based on re-observed
RAVE targets. The contours contain 30, 50, 70, and 90% of the data respectively.
correlation between the error in two parameters indicates that a systematic error in one
parameter influences the result in the other.
– 30 –
- RAVE observations of Soubiran & Girard (2005) stars,
- Asiago echelle observations of RAVE targets (R ∼ 20000),
- AAT 3.9m UCLES echelle observations of RAVE targets,
- APO ARC echelle observations of RAVE targets (R ∼ 35000).
The last three sources of calibration data make the bright RAVE targets sample and were
all reduced and processed within the RAVE collaboration using the same technique and are
therefore merged in the following and referred to as “echelle data”. We follow a standard
analysis procedure using Castelli ODFNEW atmosphere models. The gf values for iron lines
are taken from three different sources
- the list from Fulbright (2000) for metal poor stars based
- a list of differential loggf from Acturus (Fulbright et al. 2006) best suited for metal
- a list of differential loggf from the Sun best suited for dwarf stars.
The three line lists give reasonable agreement (∆Teff< 50K and ∆[Fe/H] < 0.1dex) in the
parameter boundary regions. The alpha- and heavy-element line list is basedon Fulbright
(2000) for metal-poor stars and Fulbright et al. (2007) for metal-rich stars. Teffvalues are
obtained using the excitation balance, forcing the distribution of logǫ(Fe)5vs. excitation
potential for individual FeI lines to have a flat slope. logg is obtained via the ionisation
balance, forcing the logǫ(Fe) values derived from FeI and FeII lines to agree. Both methods
are fully independent from the technique used by the RAVE pipeline to estimate atmospheric
parameters from medium-resolution spectra.
For the RAVE observation of stars studied in the literature, we chosed to build our
sample upon the Soubiran & Girard (2005) catalog.
measurements from the literature paying a particular attention at reducing the systematics
between the various studies. It makes this catalog particularly suited for calibration purposes.
This catalog contains abundances
Table 5 summarizes the content of each sample while Figure 13 presents the distribution
in logg and Teffof stars in the calibration sample. The GCS also provides photometric Teff
measurements but as for DR2, we choose not to include photometric Teffin our analysis.
5ǫ(X) is the ratio of the number density of atoms of element X to the number density of hydrogen atoms.
– 31 –
4000 6000 8000
log g (dex)
Fig. 13.— Location of the reference stars in the (Teff,logg) plane. Squares are echelle data,
the dashed line representing our separation between dwarfs (open symbols) and giants (grey
symbols) for the calibration relation. Crosses are stars in Soubiran & Girard (2005).
Table 5: Samples used to calibrate the RAVE atmospheric parameters. The echelle sample
covers the data obtained using UCLES, ARC, and Asiago spectrographs and were processed
and analysed consistently.
Soubiran & Girard 102107
(1): Soubiran & Girard (2005) do not report metallicity [M/H], so
their values are derived from a weighted sum of the quoted element
abundances of Fe, O, Na, Mg, Al, Si, Ca, Ti, and Ni, assuming the
solar abundance ratio from Anders & Grevesse (1989).
– 32 –
In the following, we separate the analysis of Teffand logg from [M/H], the latter requiring
a specific calibration.
• Teffand logg :
Table 6 presents the results of the comparison of the RAVE pipeline outputs with the
reference datasets. Since outliers are present, we use a standard iterative (sigma-clipping)
procedure to estimate the mean offset and standard deviation for each atmospheric pa-
rameter. The new version of the pipeline shows a slight tendency to overestimate Teff by
∼ 50 − 60 K compared to the previous version, with an increase of the standard deviation
from 188 K to 250 K. For logg the results are consistent between the two versions of the
pipeline. We note here that the reference samples used for the new release have increased
considerably, with the number of Soubiran & Girard (2005) stars increasing by a factor of
two and the number of echelle observations by a factor of four.
To further validate our atmospheric parameters, we compare the offset between the
reference atmospheric parameters with the RAVE values. This is presented in Figure 14 for
Teff(top panels) and logg (bottom panels) as a function of reference Teff(left), logg (middle),
and [m/H] (right). The crosses indicate the data discarded by the iterative procedure as being
For Teff, no correlation is observed either as a function of Teffor [m/H]. Considering the
echelle data alone (open squares) a tendency for Teff to be overestimated as logg increases
is observed, producing the −85K offset reported in Table 6. However, at low logg the
discrepancy vanishes. This tendency is not seen for the Soubiran & Girard (2005) stars.
Since this effect is not systematic, it leads us to conclude that the apparent trend in Teff
with logg is not due to the RAVE data but instead due to the different methods used to
derive this parameter in the other works.
For logg, no trend is observed with Teff. However a trend with logg seems to be present,
Table 6: Mean offset and standard deviation for Teffand logg between the reference datasets
and RAVE DR3 values. Ntotis the total number of observations in the reference datasets,
and Nrej,Teffand Nrej,loggare the number of observations rejected by the iterative procedure
for estimating the mean difference and dispersion for Teffand logg respectively.
Echelle227−85 ± 14
Soubiran & Girard107−63 ± 26
All334−72 ± 14
−0.12 ± 0.03
−0.05 ± 0.03
−0.10 ± 0.02
– 33 –
such that the RAVE logg is slightly overestimated at the low end (by ∼0.5 dex). In addition,
a tendency to overestimate logg at low metallicities is seen, amounting to the same order.
Because this effect is limited to the very low logg end of the distribution (logg < 1), which
is not highly populated in the RAVE catalog, this leads to the conclusion that our logg
determination are reliable within our quoted uncertainties.
• [M/H] :
As stated in Paper II, the metallicity indicator obtained by the RAVE pipeline is, due
to our medium resolution and limited signal-to-noise ratio, a mixture of the real metallicity,
alpha enhancement, and possibly rotational velocity. To obtain an unbiased estimator, we
rely on a calibration relation set using a sample of stars with known atmospheric parameters.
Paper II presented a first calibration relation using an iterative fitting procedure of the
[M/H] = c0+ c1.[m/H] + c2.[α/Fe] + c3.logg .
The coefficients of this relation were obtained based on a sample of 45 APO, 24 Asiago,
49 Soubiran & Girard (2005), and 12 M67 cluster member stars. With the larger number of
reference stars available for this release and due to the new version of the processing pipeline,
modified to increase the reliability of the atmospheric parameters, we recompute and extend
the calibration relation.However, we now restrict the analysis to the reference sample
consisting of echelle data. This sample was selected to evenly cover the (logg,Teff) plane
of the RAVE survey and was processed using the same technique and reduction algorithm,
therefore providing an homogeneous set of reference data. Also, with the knowledge gained
from the analysis of the correlation between parameters, the proposed calibration relation
now takes the form
[M/H] = c0+ c1.[m/H] + c2.[α/Fe] + c3.Teff
5040+ c4.logg + c5.STN, (2)
where we added Teff to the calibration relation due to the strong correlation observed in
Fig. 12 and discussed in Section 3.2.2. S/N is also included as one expects an impact of
the noise at the low S/N regime where the pipeline may mistake noise spikes for enhanced
metallicity. Since Teffseems to be the primary source of error for [m/H], we computed four
calibration relations for the various cases with and without S/N or logg. As for the DR2
calibration, we see no evidence for higher order terms and therefore restrict our search for
the best calibration to first order (linear) relations.
The coefficients for the calibration relations are obtained by minimizing the difference
– 34 –
4000 6000 8000
0 2 4
−3 −2 −1 0 1
4000 6000 8000
0 2 4
−3 −2−1 0 1
Fig. 14.— Difference between the atmospheric parameters of the reference datasets and
of the RAVE DR3 parameters as a function of the reference Teff, logg and [M/H] for Teff
(top), and logg (bottom). Circles stand for stars in Soubiran & Girard (2005) while squares
denote echelle data. Grey symbols represent the giants, open symbols mark the location of
the dwarfs. Crosses indicate data rejected by the iterative procedure used for Table 6.
– 35 –
between the calibrated [M/H] and the reference [M/H] using an iterative procedure to reject
outliers. The resulting calibration relations are summarized in Table 7 where Ntot is the
total number of observations used to compute the calibration relation. A blank value in a
column indicates that the calibration relation does not include the corresponding parameter.
The residuals between the calibrated [M/H] and the reference [M/H] as a function of the
reference [M/H] are presented in Figure 15 where the top panels present the raw output of the
DR3 pipeline (panel marked original) and the residuals obtained using the DR2 calibration
relation on the DR3 atmospheric parameters values. The following four panels are for the
different calibration relations considered here. Finally, Table 8 presents the mean offset and
standard deviation computed from the residuals in the different cases.
From Figure 15, it is clear that applying the DR2 calibration to the DR3 pipeline outputs
is not satisfactory and produces a bias at low metallicity. This behavior is expected as the
pipeline has been modified to produce a better agreement to the metallicity distribution
which, for DR2 showed a reduced tail at the low metallicity end. As the correlation between
the parameters is significant (see § 3.2.2) and because the calibration relation is built upon
the output parameters (with a large contribution from [m/H] which is modified compared to
the DR2 pipeline), one therefore expects the DR2 calibration relation not to hold for the DR3
parameters. Ideally, the DR3 parameters would not need a calibration relation. However the
raw output of the DR3 pipeline still suffers from a small systematic effect, underestimating
the true metallicity by ∼0.1 dex with some systematic dependency on Teff.
Applying the calibration relations proposed, the RAVE metallicties agree with the
echelle values (see Table 8). However, as can be seen from Fig. 15, a systematic trend is
observed for dwarfs at high metallicity, where the difference between RAVE and the echelle
value reaches 0.4 dex for the highest metallicity stars. At low metallicity, the dispersion is
significantly reduced and when applying any of the calibration relations, the two determi-
nations agree well. Adding logg or S/N to the calibration relation does not improve the
residuals significantly. For logg this is understood as it is the atmospheric parameter with
the largest uncertainty. Hence its dispersion prevents it from having significant weight in the
calibration relation, even though we know the error on this parameter is strongly correlated
to errors in [M/H] (see Section 3.2.2). For S/N, the situation is less clear but part of its
low weight in the calibration relation is linked to the fact that, in order to observe RAVE
targets at high resolution, we selected targets in the bright part of the catalog to ensure
enough S/N in the spectra to allow precise measurements of the atmospheric parameters.
Hence, the region of the S/N space where this parameter plays an important role (S/N < 20)
is not properly sampled, lowering its weight on the calibration relation whereas above this
threshold, no correlation with S/N is observed.
– 36 –
Table 7: Coefficients in the calibration relation for the RAVE metallicities using different sets of parameters for the fit.
Ntotis the total number of data points used to derive the calibration, ciare the coefficients from Eq. 2. The first line
presents the output of the new RAVE pipeline while the second line presents the results obtained when one applies
the calibration relation of Paper II. The following lines are the calibration relations obtained using the new pipeline
DR3 no S/N no logg 2230.578± 0.098
DR3 with S/N2170.587± 0.091
DR3 with logg2230.518± 0.127
DR3 with S/N and logg2220.429± 0.132
DR3 no S/N no logg890.612± 0.236
DR3 with S/N750.706± 0.199
DR3 with logg 82−0.174 ± 0.222
DR3 with S/N and logg81−0.170 ± 0.217
DR3 no S/N no logg1270.763± 0.197
DR3 with S/N 1190.399± 0.178
DR3 with logg1270.354± 0.287
DR3 with S/N and logg 1270.239± 0.297
-1.095 ± 0.022
1.106 ± 0.024
1.111 ± 0.031
1.101 ± 0.032
1.081 ± 0.045
1.250 ± 0.055
1.061 ± 0.047
1.063 ± 0.047
0.232 ± 0.038
0.219 ± 0.037
1.094 ± 0.027
1.087 ± 0.027
1.162 ± 0.044
1.154 ± 0.045
– 37 –
−3−2 −1 0 1
DR2 calibration all
−3−2 −1 0 1
New calibration all
no logg no SNR
−3−2−1 0 1
[M/H]ref (dex)[M/H]ref (dex)
New calibration all
−3−2−1 0 1
New calibration all
−3−2−1 0 1
New calibration all
with logg +SNR
−3−2 −1 0 1
Fig. 15.— Difference between the reference [M/H] and RAVE [M/H] using the different
calibration relations as a function of reference [M/H]. The crosses indicate the observations
rejected from the fit by the iterative procedure.
– 38 –
New calibration split
no logg no SNR
−3−2−1 0 1
New calibration separate
−3−2−1 0 1
New calibration separate
−3−2 −1 0 1
New calibration separate
with logg +SNR
−3 −2−1 0 1
Fig. 16.— Same as Fig. 15 but using separate calibration relations for dwarfs and giants.
– 39 –
Table 8: General properties of the different calibration relations presented in Tab. 7. ∆[M/H]
is the mean difference [M/H]ref−[M/H]correctedand σ[M/H]is the dispersion. Nrejis the number
of observations rejected by the iterative procedure as outliers. For each calibration relation,
we also provide separate statistics for dwarfs and giants obtained using the calibration rela-
tions derived specifically for each sample.
DR3 no SNR no logg
DR3 with SNR
DR3 with logg
DR3 with SNR and logg
– 40 –
Finally, to improve on the situation for the dwarfs, we split the sample between dwarfs
and giants (see Fig. 13 for the criterion used) and applied the same procedure to each
sub-population. The result of these calibration relations is presented in Fig. 16, the basic
statistics being reported in Table 8 for each calibration relation.
Using separate calibration relations for dwarfs and giants does help improve the disper-
sion for dwarfs, but, as we can see in Fig. 16, the calibration relation is unable to remove the
bias at high [M/H], the most discrepant stars being rejected by the fit. Only a mild improve-
ment is obtained. Separating the dwarfs from the giants changes the calibrated metallicity
for these stars by only 0.02 dex, or 0.05 dex if one also uses logg in the calibration.
The DR3 release of the RAVE catalog contains 83072 radial velocity measurements
for 77461 individual stars. Atmospheric parameters are provided for 41672 spectra (39833
stars). These data were acquired over 257 observing nights, spanning the time interval
April 11th2003 to March 12th2006, and 976 fields. The data new to this release cover the
time interval March 31st2005 to March 12th2006 where 32477 new spectra were collected.
The total coverage of the pilot survey is then 11500 square degrees. Figure 17 plots the
general pattern of (heliocentric) radial velocities, where the dipole distribution is due to a
combination of asymmetric drift and the Solar motion with respect to the Local Standard
The DR3 release is split into two catalogs: Catalog A and Catalog B. The first catalog
contains the higher signal-to-noise data, which yields reliable values for the stellar param-
eters, and includes both radial velocities and stellar parameters (temperature, gravity and
metallicity). The second catalog contains the lower signal-to-noise data and does not include
stellar parameters. The criterion for dividing between the two catalogues was based on the
STN values, where available, with a threshold value of STN = 20 between Catalogs A and
B. Table 9 summarizes the catalogues, where we see that 70% of the data are in Catalog A.
The DR3 release can be queried or retrieved from the Vizier database at the CDS, as
well as from the RAVE collaboration website (www.rave-survey.org). Table 12 describes
its column entries, where the same format is used for both catalogs for ease-of-use even
though the stellar parameter columns are NULL in Catalog B. Catalog A contains the
measured stellar parameters from the RAVE pipeline and includes also the inferred value of
the α-enhancement. As explained in the DR2 paper, this is provided strictly for calibration
purposes only and cannot be used to infer the α-enhancement of individual objects.
– 41 –
Fig. 17.— Aitoff projection in Galactic coordinates of RAVE 3rdData Release fields. The
yellow line represents the celestial equator and the background is from Axel Mellinger’s
Table 9.The two DR3 catalogs
Selection criteriaResults included
Catalog A 5727220 < STN or 20 <
6 ≤ STN < 20 or
6 ≤ SNRatio < 20
Radial velocities, stellar parameters
Catalog B 25800Radial velocities
– 42 –
Following Paper II, in Fig. 18 we plot the location of all spectra on the temperature-
gravity-metallicity wedge for different slices in Galactic latitude. The main-sequence and
giant-star groups (particularly the red-clump branch) are clearly visible, with their relative
frequency and metallicity distribution varying with latitude. For the hotter stars (Teff >
9000K) there is significant discretization in logg. This is caused by the combination of
a degeneracy in metallicity for these Paschen-line dominated spectra and a smaller range
in possible logg, which leads to the penalization algorithm having a tendency to converge
on the same solution. Figure 19 plots histograms of the parameters for different latitudes.
The fraction of main-sequence stars increases with the distance from the Galactic plane (see
Paper 2 for a discussion). The metallicity distribution function becomes more metal poor
for the higher-latitude fields as well. Also the shift of the temperature distribution towards
higher temperature turn-off stars with decreasing Galactic latitude is clearly visible.
60 < |b|
40 < |b| < 60
20 < |b| < 40
T eff [K]
|b| < 20
60 < |b|
40 < |b| < 60
20 < |b| < 40
−2.0 −1.5 −1.0 −0.5 0.0 0.5
|b| < 20
60 < |b|
40 < |b| < 60
T eff [K]
20 < |b| < 40
−2.0 −1.5 −1.0 −0.5 0.0 0.5
|b| < 20
Fig. 18.— The temperature-gravity-metallicity plane for different wedges in Galactic lati-
– 43 –
−3.0 −2.0 −1.0 0.0
60 < |b|
40 < |b| < 60
20 < |b| < 40
|b| < 20
T eff [K]
Fig. 19.— Temperature, gravity, and metallicity histograms for spectra with published stellar
parameters. Histograms for individual Galactic latitude bands are plotted separately with
the key given in the top panel. Spectra with |b| ≤ 20◦include calibration fields.
– 44 –
As in the previous releases, DR3 includes cross-identifications with optical and near-IR
catalogs (USNO-B: B1, R1, B2, R2; DENIS: I, J, K; 2MASS: J, H, K). The nearest-
neighbor criterion was used for matching and we provide the distance to the nearest neighbor
and a quality flag on the reliability of the match. Table 10 shows the completeness and flag
statistics for the two catalogs, where we see that Catalog A’s coverage and quality are slightly
better than those of Catalog B. This is because Catalog A is dominated by lower-magnitude
objects while Catalog B contains mainly the higher-magnitude objects. For both, however,
nearly all stars were successfully matched with the 2MASS and USNO-B catalogs. About
3/4 of the stars lie in the sky area covered by the DENIS catalog.
Our wavelength range is best represented by the I filter. As discussed in detail in
Paper II, there are some problems with a fraction of the DENIS I magnitudes, particular
for IDENIS < 10, due to saturation effects. Following the methodology of DR2, we com-
pare the DENIS magnitudes against an approximate one calculated from 2MASS J and K
(see Equation 24 in Paper 2). Fig. 20 compares the DENIS and the “jury-rigged” 2MASS
I−magnitudes for all stars in the current data release. We see that the two magnitudes
agree for the majority of objects, but a significant fraction have large errors: 10% have
|(IDENIS− I2MASS)| > 0.2, with differences of up to 4 magnitudes. It was proposed in Pa-
per II that IDENISmagnitudes should be avoided when the condition
− 0.2 < (IDENIS− J2MASS) − (J2MASS− K2MASS) < 0.6,(3)
is not met. In Figure 20 we differentiate between the stars that do and do not satisfy this
condition, where we see how it selects out the problematic IDENISmagnitudes.
As in DR2, the proper motions are sourced from the PPMX, Tycho-2, SSS and UCAC2
catalogs. As described in Paper II, the most accurate available proper motion is chosen for
each object. Table 11 summarizes for both Catalogs A and B the proper-motion sources and
the average and 90thpercentile errors. The quality of the proper motions is slightly worse
for Catalog B, because the fainter objects in this catalog include a higher proportion of the
more distant objects in the survey.
– 45 –
I DENIS − I 2MASS
Fig. 20.— Difference between DENIS and jury-rigged 2MASS I magnitudes as a function of
IDENIS. The blue points are one satisfying equation 3, while the larger red points are ones
that do not.
– 46 –
This third data release of the RAVE survey reports 83072 radial-velocity measurements
for 77461 stars, covering more than 11500 square degrees in the southern hemisphere. The
sample is randomly selected in the magnitude interval 9 ≤ I ≤ 12. This release also provides
stellar atmospheric parameters for 41672 spectra representing 39833 individual stars.
Since DR2, we modified the RAVE processing pipeline to account better for defects in the
observed spectra due to bad pixels, fringing, or locally inaccurate continuum normalization.
The main driver of the modification was to improve on the known limitation of our estimates
of stellar atmospheric parameters. Also, the algorithm to correct for the zero-point offset
has been revised, enabling a better control of our radial-velocity accuracy.
The accuracy for the radial velocities is marginally improved with the new pipeline,
the distribution of internal errors in the radial velocities has mode 0.8kms−1and median
1.2kms−1, and 95% of the sample having an internal error better than 5kms−1, which is
the primary objective of RAVE. Comparing our radial velocities to independent estimates
based on 373 measurements from five data sources we find no evidence for a bias in our
radial velocities, our mean radial velocity error being ∼ 2kms−1.
A significant effort has been spent in improving the quality and validation of our stellar
atmospheric parameters with respect to the external errors and biases. The internal errors
due to the method and the sampling of synthetic spectra grid remain unchanged and are
presented in Paper II. The new calibration sample consists of 362 stars from four different
sources (either custom observations or literature) and cover the full HR diagram. Comparing
our measured parameters to these reference measurements, we find a good agreement for Teff
and logg with a mean offset and dispersion of (-63, 250) K for Teff and (-0.1, 0.43) dex
for logg, which are consistent with DR2. The [m/H] distribution is improved but the true
metallicity [M/H] remains a combination of [m/H], [α/Fe], and Teff. Taking logg or S/N into
the calibration of [M/H] only marginally improves the situation and the simplest calibration
relation is preferred.
This data release is the last one based on the pilot-survey input catalog. Further re-
leases will be based on an input catalog built upon DENIS-I magnitudes. This catalog,
supplemented by the catalog of distances, make this release an unprecedented tool to study
the Milky Way.
Acknowledgments Funding for RAVE has been provided by: the Anglo-Australian
Observatory; the Leibniz-Institut f¨ ur Astrophysik Potsdam (AIP); the Australian National
University; the Australian Research Council; the French National Research Agency; the
– 47 –
German Research foundation; the Istituto Nazionale di Astrofisica at Padova; The Johns
Hopkins University; the National Science Foundation of the USA (AST-0908326); the W.M.
Keck foundation; the Macquarie University; the Netherlands Research School for Astronomy;
the Natural Sciences and Engineering Research Council of Canada; the Slovenian Research
Agency; the Swiss National Science Foundation; the Science & Technology Facilities Council
of the UK; Opticon; Strasbourg Observatory; and the Universities of Groningen, Heidelberg
and Sydney.The RAVE web site is at http://www.rave-survey.org.
Research Council has provided financial support through ERC-StG 240271(Galactica).
Anders, E., & Grevesse, N., 1989, Geochimi. Cosmochim. Acta, 53, 197
Breddels, M. A., et al., 2010, A&A, 511,90
Burnett, B. & Binney, J., 2010,MNRAS, 407, 339
Burnett, B., et al., 2010, A&A, in prep.
Co¸ skunoˇ glu, B., et al., 2011, MNRAS, 412, 1237
Eisenstein, D., et al., 2011, AJ, submitted (arXiv:1101.1529)
Famaey, B., et al., 2005, A&A, 430, 165
Fellhauer, M., et al., 2006, ApJ, 651, 167
Fulbright, J.P., 2000, AJ, 120, 1841
Fulbright, J.P., McWilliam, A., Rich, R.M., 2006, ApJ, 636, 821
Fulbright, J.P., McWilliam, A., Rich, R.M., 2007, ApJ, 661, 1152
Fulbright, J.P., et al., 2010, ApJL, 724, 104
Kiss, L.L., et al., 2011, MNRAS, 411, 117
Klement, R., et al., 2008, ApJ, 685, 261
Klement, R.J., Bailer-Jones, C.A.L., Fuchs, B., Rix, H.W., Smith, K.W., 2011, ApJ, 726,
Matijeviˇ c, G., et al., 2010, AJ, 140, 184
– 48 –
Matijeviˇ c, G., et al., 2011, AJ, submitted
Munari, U., et al., 2005, A&A, 442, 1127
Munari, U., et al., 2008, A&A, 488, 969
Munari, U., et al., 2009, A&A, 503, 511
Nordstr¨ om, B., et al., 2004, A&A,418,989
Perryman, M.A.C., et al., 1997, ESASP, 1200
Ruchti, G.R., et al., 2010, ApJ, 721L, 92
SDSS-III collaboration, 2011, ApJS, in press (arXiv:1101.1559)
Seabroke, G., et al., 2008, MNRAS, 384, 11
Soubiran, C., & Girard, P., 2005, A&A, 438, 139
Siebert, A., et al., 2008, MNRAS, 391, 793
Siebert, A., et al., 2011, MNRAS, 412, 2026
Smith, M. C., et al., 2007, MNRAS, 379, 755
Steinmetz, M., et al., 2006, AJ, 132, 1645 (Paper I)
Veltz, L., et al., 2008, A&A, 480,753
Williams, M.E.K., et al., 2011, ApJ, 728, 102
Wilson, M.L., et al., 2011, MNRAS, in press (arXiv:1009.2052)
Yanny, B., et al., 2009, AJ,137,4377
Zwitter, T., et al., 2008, AJ, 136, 421 (Paper II)
Zwitter, T. et al., 2010, A&A, 522A, 54
This preprint was prepared with the AAS LATEX macros v5.2.
– 49 –
Table 12 describes the contents of individual columns of the Third Data Release catalog.
The catalog is accessible online at www.rave-survey.org and via the Strasbourg astronom-
ical Data Center (CDS) services.
– 50 – Download full-text
Table 10.Number and fraction of RAVE database entries with a counterpart in the
Catalog nameNumber of
% of entries
% with quality flag