UV Star Formation Rates in the Local Universe

Abstract

We measure star formation rates (SFRs) of 50,000 optically selected galaxies in the local universe (z ≈ 0.1)—from gas-rich dwarfs to massive ellipticals. We obtain dust-corrected SFRs by fitting the GALEX (ultraviolet) and SDSS photometry to a library of dust-attenuated population synthesis models. For star-forming galaxies, our UV-based SFRs compare remarkably well with those from SDSS-measured emission lines (Hα). Deviations from perfect agreement are shown to be due to differences in the dust attenuation estimates. In contrast to Hα measurements, UV provides reliable SFRs for galaxies with weak Hα, and where Hα is contaminated with AGN emission (1/2 of the sample). Using full-SED SFRs, we calibrate a simple prescription that uses GALEX far- and near-UV magnitudes to produce dust-corrected SFRs for normal star-forming galaxies. The specific SFR is considered as a function of stellar mass for (1) star-forming galaxies with no AGNs, (2) those hosting an AGN, and (3) galaxies without Hα emission. We find that the three have distinct star formation histories, with AGNs lying intermediate between the star-forming and the quiescent galaxies. Star-forming galaxies without an AGN lie on a relatively narrow linear sequence. Remarkably, galaxies hosting a strong AGN appear to represent the massive continuation of this sequence. On the other hand, weak AGNs, while also massive, have lower SFRs, sometimes extending to the realm of quiescent galaxies. We propose an evolutionary sequence for massive galaxies that smoothly connects normal star-forming galaxies to quiescent galaxies via strong and weak AGNs. We confirm that some galaxies with no Hα show signs of star formation in the UV. We derive a cosmic star formation density at z = 0.1 with significantly smaller total error than previous measurements.

Full text

Page 1

arXiv:0704.3611v2 [astro-ph] 7 May 2007
2007 APR 20 — APJS SPECIAL GALEX ISSUE
Preprint typeset using LATEX style emulateapj v. 10/09/06
UV STAR FORMATION RATES IN THE LOCAL UNIVERSE
SAMIR SALIM1,*, R. MICHAEL RICH1, STÉPHANE CHARLOT2, JARLE BRINCHMANN3, BENJAMIN D. JOHNSON4, DAVID
SCHIMINOVICH4, MARK SEIBERT5, RYAN MALLERY1, TIMOTHY M. HECKMAN6, KARL FORSTER7, PETER G. FRIEDMAN7, D.
CHRISTOPHER MARTIN7, PATRICK MORRISSEY7, SUSAN G. NEFF8, TODD SMALL7, TED K. WYDER7, LUCIANA BIANCHI9, JOSE
DONAS10, YOUNG-WOOK LEE11, BARRY F. MADORE5, BRUNO MILLIARD10, ALEX S. SZALAY6, BARRY Y. WELSH12, SUKYOUNG K.
YI11
2007 Apr 20 — ApJS Special GALEX Issue
ABSTRACT
We measure star formation rates (SFRs) of ≈ 50,000 optically-selected galaxies in the local universe
(z ≈ 0.1), spanning a range from gas-rich dwarfs to massive ellipticals. We obtain dust-corrected SFRs by
fitting the GALEX (ultraviolet) and SDSS (optical) photometry to a library of population synthesis models that
include dust attenuation. For star-forming galaxies, our UV-based SFRs compare remarkably well with those
derived from SDSS-measured emission lines (primarily Hα). Systematic deviations from perfect agreement
between these two methods is shown to be due to differences in the dust attenuation estimates. In contrast to
measurements based on Hα, UV provides reliable SFRs for galaxies with weak or no Hα emission, and where
Hα is contaminated with an emission from an AGN (1/2 of the sample). We use full-SED SFRs to calibrate
a simple prescription that uses GALEX far-UV magnitude and the UV slope to produce good dust-corrected
SFRs for normal star-forming galaxies. The specific SFR (SFR normalized by stellar mass) is considered as
a function of stellar mass for (1) star-forming galaxies with no AGN, (2) those hosting an AGN, and for (3)
galaxies without Hα emission (the latter two groups forming the bulk of the optical red sequence). We find
that the three have distinct star formation histories, with AGN lying intermediate between the star-forming
and the quiescent galaxies. Normal star forming galaxies (without an AGN) lie on a relatively narrow linear
sequence. Remarkably, galaxies hosting a strong AGN appear to represent the massive continuation of this
sequence. On the other hand, weak AGN, while also massive, have lower SFR, sometimes extending to the
realm of quiescent galaxies. We proposean evolutionarysequence for massive galaxies that smoothly connects
normal star-forming galaxies to quiescent (red sequence) galaxies via strong and weak AGN. We confirm that
some galaxies with no Hα emission show signs of star formation in the UV. We derive a UV-based cosmic star
formation density at z = 0.1 with significantly smaller total error than previous measurements.
Subject headings: galaxies: evolution—galaxies: fundamental parameters— ultraviolet: galaxies—surveys—
galaxies: active
1. INTRODUCTION
Current studies of galaxies are characterized by two ma-
jor features: availability of large samples of objects (galaxy
surveys),and the utilization of the multiwavelengthapproach.
Such studies extend from the nearby galaxies to those close
1Department of Physics and Astronomy, University of California, Los
Angeles, CA 90095
⋆Current address: NOAO, 950 North Cherry Ave., Tucson, AZ 85719,
samir@noao.edu
2Institut d’Astrophysique de Paris, CNRS, 98 bis boulevard Arago, F-
75014 Paris, France
3Centro de Astrofísica da Universidade do Porto, Rua das Estrelas 4150-
762 Porto, Portugal
4Department of Astronomy, Columbia University, New York, NY 10027
5Observatories of the Carnegie Institution of Washington, 813 Santa Bar-
bara St., Pasadena, CA 91101
6Department of Physics and Astronomy, The Johns Hopkins University,
Homewood Campus, Baltimore, MD 21218
7California Institute of Technology, MC 405-47, 1200 East California
Boulevard, Pasadena, CA 91125
8Laboratory for Astronomy and Solar Physics, NASA Goddard Space
Flight Center, Greenbelt, MD 20771
9Center for Astrophysical Sciences, The Johns Hopkins University, 3400
N. Charles St., Baltimore, MD 21218
10Laboratoire d’Astrophysique de Marseille, BP 8, Traverse du Siphon,
13376 Marseille Cedex 12, France
11Center for Space Astrophysics, Yonsei University, Seoul 120-749, Ko-
rea
12Space Sciences Laboratory, University of California at Berkeley, 601
Campbell Hall, Berkeley, CA 94720
to the epoch of the formation of the first galaxies. A rate
at which a galaxy forms stars is one of the more important
propertiesin studyinggalaxy evolution. The multiwavelength
approach allows us to employ a suite of star formation (SF)
indicators—from X-rays to radio wavelengths. Major effort
in recent years was made to provide reliable calibrations for
different SF estimators and to understand their differences,
advantages and limitations. The most straightforward evalua-
tion is achieved by comparing two or more star formation in-
dicatorsforthesame set ofobjects. While newgalaxysurveys
provide large statistical samples with which one can attempt
such studies, the sample of galaxies for which more than a
single star formation indicator can be applied is not necessar-
ily large. Moreover, various SF indicators are often applied
for different redshift regimes. Global cosmic star formation
history is therefore the result of studies that employ different
SF indicators.
Among the most frequently used star formation indica-
tors are the UV continuum (usually at λ < 2000 Å), nebu-
lar recombination lines (primarily Hα, but also [OII]), far-
IR dust emission, and the synchrotron radio continuum at 21
cm (Kennicutt 1983). A comprehensive review of most of
these methods was presented in Kennicutt (1998), together
with simple formulae for the conversionof the true flux into a
star formation rate (SFR). Recently, the use of other SF indi-
cators has been explored, such as the X-ray continuum (e.g.,
David et al. 1992; Kilgard et al. 2002), or the luminosity of

Page 2

2SALIM ET AL.
PAH features in the mid-IR (Roussel et al. 2001). Star for-
mation indicators are only as good as the assumptions that
connect a certain observed luminosity to the actual star for-
mation rate, and therefore all are sensitive to various system-
atic uncertainties. One factor has proved more frustrating to
account for than the others—the effect of dust obscuration on
the UV and Hα star formation rates. In contrast to these two
dust-sensitiveindicators,bolometricIR luminosityandthe ra-
dioluminosityare oftenconsideredto be “true”SF indicators.
While this is certainly an oversimplification, such perception
is bolstered by the very tight correlation of the IR and the ra-
dio luminosities, at least as observed for the normal galaxies
in the local universe (de Jong et al. 1985). However, even if
they were perfect, IR or radio methods cannot make UV or
Hα methods obsolete for many practical reasons.
There have been a number of studies that compared the
UV and Hα star formation rates—either as the observed lu-
minosities, or by applying various schemes to correct for the
dust attenuation.13These comparisons were often carried out
with respect to some dust-free SF indicator. Hopkins et al.
(2001) used a sample of several hundred objects with Hα and
UV measurements (with the U-band photometry serving as a
proxy for the UV), and compared them to the far-IR measure-
ments compiledby Cram et al. (1998). They find that in order
to reconcile Hα and far-IR luminosities one needs to apply a
dust attenuation that is not fixed, but rather depends on the
SFR itself. Actually, this dependence is a consequence of the
relationship between the dust attenuation and the far-IR lu-
minosity found by Wang & Heckman (1996). Hopkins et al.
(2001) use an extinction curve to extrapolate the dust attenu-
ations obtained for Hα into the UV regime, but find that such
correctionfails to bring UV luminosities to agreewith the far-
IR, i.e., the simple application of the same attenuation mech-
anism to both the Hα and the UV emission does not appear to
be correct. Bell & Kennicutt (2001) used actual UV observa-
tions of 50 nearby galaxies in two ultraviolet bands obtained
with the Ultraviolet Imaging Telescope (UIT), and compared
themtoHα luminositiesfromthenarrow-bandimaging. They
determine UV attenuation for 13 galaxies using the correla-
tion with the UV slope (Calzetti et al. 1994), and measure Hα
attenuation for 21 objects using the thermal radio continuum
fluxes. They find that both can reach high levels even for nor-
mal galaxies (∼ 4 mag for Hα, and ≥ 5 mag for far-UV), and
lend support to previous notions that the attenuation increases
with SFR. Sullivan et al. (2000, 2001) confirmed that a bet-
ter agreement between the UV observations from the FOCA
balloon mission (Milliard et al. 1992), and the fiber Hα spec-
tra is achieved when attenuation corrections are taken to be
luminosity-dependent. They also suggest that one perhaps
cannot use simple extinction curve scalings to convert Hα
attenuations into UV attenuations. The breakthrough in re-
solving this problem came with the introduction of the two-
component dust attenuation model of Charlot & Fall (2000).
This model was indeed motivated by the need to produce a
consistent model for dust attenuation affecting Hα and UV
continuum photons. It postulates the existence of short-lived
(10 Myr) giant molecular clouds that affect photons produc-
ing the Hα line. On the other hand, the attenuation of the
UV continuum, having timescales longer than the lifetime of
13In this paper we will use the term “attenuation”, rather than the more
commonly used term “extinction”, thus emphasizing the complex processes
of absorption and scattering in a galaxy, rather than the dimming of light
along a single line of sight.
giant molecularclouds, is predominantlyproducedby the dif-
fuseISM (afterthe molecularcloudshavedispersed),at levels
that are typically 3 times lower (for a given wavelength) than
those in the molecular clouds. In this paper we will be using
the Charlot & Fall (2000) model, thus testing it for the first
time on a large scale.
In addition to systematic trends, we should mention that
some previous studies were finding that the measurement
errors are smaller than the observed scatter. Sullivan et al.
(2000, 2001); Iglesias-Páramo et al. (2004) offer the explana-
tion for this scatter as arising from the differing timescales
over which Hα and UV SFRs are sensitive, so that UV could
in some cases (especially in low-mass galaxies) detect a short
starburst that is no longer observable in Hα.
Sometimes, the systematics are present in the observations
themselves. Rosa-González et al.(2002)showedthateventhe
relatively reliable determination of Hα attenuation can be af-
fected by the systematics when not correcting for the under-
lyingstellar absorptionin Balmer emission lines. Systematics
can also arise with galaxy samples selected at different wave-
lengths (Buat et al. 2002).
The main obstacle in obtaining UV measurements for a
large number of galaxies in the local universe is that they
need to be made outside of Earth’s atmosphere. For this rea-
son in 2003 NASA launched the Galaxy Evolution Explorer
(GALEX, Martin et al. 2005). GALEX is currently conduct-
ing the first ever survey of the UV sky. The imaging (in two
UV bands) is executed in several modes—from a shallow all-
sky survey, to the ultra-deep fields. In this paper we are us-
ing measurements obtained in the medium-deep survey that
is designed to image regions of the sky covered by the Sloan
Digital Sky Survey (SDSS). Thus we obtain a large sample
of galaxies with both UV and optical photometry, as well as
spectroscopic redshifts. GALEX and SDSS data, and the re-
sulting sample, are described in §§2 and 3. Dust-corrected
star formation rates (and some other physical properties such
as the stellar mass) are obtained by comparing the observed
colorsto stellar populationsynthesis modelsto whichthe dust
attenuation has been applied (§4). While we use full UV to
optical SED, the SFRs are essentially constrained by the UV
(Salim et al. 2005). The large size of our sample permits a
robust statistical analysis. We compare our UV-based SFRs
to the results of the major study of Brinchmann et al. (2004),
who use SDSS spectra to derive Hα-based SFRs for ∼ 105
galaxies (§5). They employ an intricate scheme to correct for
the fiber aperture effects. Thus, our study also serves as a
check on the reliability of their methodology. Finally, we dis-
cuss star formation histories of different classes of galaxies
(§7) and derive a UV-based estimate of the global star forma-
tion density at z = 0.1 (§8).
2. DATA
2.1. GALEX data
Technical aspects of GALEX telescope, detectors and data
products are presented in Morrissey et al. (2005, 2007). Here
we give a summary of relevant information. GALEX surveys
the sky in either the imaging or the grism spectroscopy mode.
It simultaneously produces a far-UV (FUV) and a near-UV
(NUV) image having a circular field of view of 1.25◦diame-
ter. FUV and NUV filters are broadband,with effective wave-
lengths of 1528 Å and 2271 Å, respectively. We will denote
the magnitudes measured in these photometric bands as FUV
and NUV. A single field of view imaged by GALEX is called
a tile. A tile can consist of one or more visits, i.e., individual

Page 3

UV SFRs IN THE LOCAL UNIVERSE3
exposures. GALEX surveys the sky in several imaging modes,
which differ in the exposure time per tile. In this study we
use the Medium Imaging Survey (MIS), which is designed to
maximizethecoverageoftheskythatisincludedintheSDSS.
Typical exposure times in MIS are 1500 s, yielding limiting
magnitudes of FUV = NUV = 22.7 mag (AB system is used
throughout). GALEX cannot point in the vicinity of bright
sources (usually stars), which inevitably produces some gaps
in the coverage.
GALEX data used in this paper come from the MIS por-
tion of the Internal data release 1.1 (IR1.1), which is an ex-
panded version of the first public GALEX release (GR1) of
MIS (in both IR1.1 and GR1 the same pipeline, version 4, is
used to reduce the data and produce source catalogs). Dataset
consists of 705 GALEX tiles. Because of the anomaly with
the FUV detector, 98 tiles lack FUV images. The 705 tiles
cover 797 sq.deg. of the sky. Each GALEX field is restricted
to 0.6◦radius, close to the maximum field of view.14Because
of the overlap between GALEX tiles, the total unique area is
741 sq.deg. Source catalog for each tile was produced using
the SExtractor software (Bertin & Arnouts 1996).
InourstudyweuseFUVandNUVfluxesmeasuredinKron
elliptical apertures. We recalculate flux errors because they
were incorrect in the pipeline version 4 reductions (thus also
affecting GR1, cf. Trammell et al. 2007). Kron magnitudes
should represent a good measure of a galaxy’s total flux. If
both the FUV and the NUV detections are present, we use
the aperture (size, shape and position) derived from the NUV
image to measure the FUV light. We go from NUV to FUV
because we are generally more sensitive in the NUV. We add
zero-pointcalibrationerrorsof0.052and0.026magnitudesto
FUV and NUV respectivelyto accountfor systematic inaccu-
racies. The calibration errors were estimated by analyzing the
repeat imaging of a calibration stellar object, and confirmed
by a large number of repeat observations of the same field
(Morrissey et al. 2007). If a UV detection is present in only
one band, we still measure the formal flux and its error in the
other band (using the aperture defined in the detected band).
In some cases, the optical source is not detected at all by
GALEX, in which case we note the sky background at the po-
sition of the SDSS source and compute the flux error. In this
calculation we use galaxy sizes that have been derived (us-
ing a calibration based on objects detected by GALEX) from
SDSS Petrosian radii in r-band. Finally, in some cases the
FUV cannot be used because the FUV image does not exist
(15% of all galaxies).
2.2. SDSS data
In addition to the ultraviolet data from GALEX, we
use optical data from SDSS Data Release 4 (DR4,
Adelman-McCarthy et al. 2006).
band broadband photometry (ugriz bands), and the spec-
troscopic follow-up of most galaxies with r ≤ 17.77 (main
galaxy spectroscopic survey, Strauss et al. 2002). In addition
to the official data products from the SDSS collaboration,
we use value-added galaxy catalogs produced by MPA/JHU
SDSS team. These catalogs include reprocessed SDSS spec-
troscopic data, and some derived galaxy physical parameters
SDSS is providing five-
14Many other studies with GALEX restrict analysis to 0.5 or 0.55◦radius
field of view. This is because artifacts are more common near the detector
edge, and the PSF becomes distorted. Since in this study we match SDSS
objects to GALEX sources, the chances of a match with an artifact are small.
Also, while the PSF at the edges is distorted, the total GALEX flux is not
affected (see also §4.3).
that are based on the spectroscopic data.
SDSS catalog lists magnitudes measured in several differ-
ent ways. We use MODELMAG magnitudes, which are the
measurements of choice for relative fluxes (i.e., provide sta-
ble colors), while still capturing most of the total light. We
apply slight adjustments (−0.04 and 0.02 mag) to u and z
magnitudes respectively to bring them closer to the actual
AB system (Abazajian et al. 2004). We convert SDSS mag-
nitudes and errors into fluxes using the transformations given
in Scranton et al. (2005). In addition to catalog photometric
errors, we add 0.01 mag of uncorrelated calibration error to
each of the bands (see §4.3), and, specifically for the u band,
we add an additionalcolor-dependenterror due to the red leak
(σu,RL=0.0865(r−i)2+0.0679(r−i),basedonAbazajian et al.
2004). In rare cases an SDSS magnitude would be missing
and such band has to be excluded from the analysis. We fur-
ther exclude from analysis individual SDSS magnitudes that
are nominally fainter than ugriz = 25. These are invariably
spurious,regardlessofthelistedphotometryerror. Finally,we
excludemagnitudeswith errorslargerthan10×thetypicaler-
rorinthatband(wefindtypicalerrorstobe0.07,0.009,0.007,
0.007 and 0.017 mag for u, g, r, i, and z bands respectively).
While large, we find these errors to be significantly underes-
timated. To summarize, except in cases listed above when we
exclude certain individual flux points, our input data consists
of7-bandphotometryandspectroscopicredshiftsfromSDSS.
2.3. GALEX/SDSS matched catalog
Of 741 sq.deg. of GALEX unique imaging, 645 sq.deg.
overlaps with SDSS DR4 spectroscopic area, thus defining
the solid angle of our sample.15We estimate this area by
counting the SDSS galaxies that: (1) have spectra, (2) have
a dereddened magnitude 14.5 < rPetro< 17.5 (the faint end
is taken to be comfortably brighter than the spatially variable
spectroscopic limit) and (3) lie at redshifts 0.005 < z < 0.22.
We count such galaxies both in our survey, and in the full
DR4, whose spectroscopic area is known. The ratio of the
two counts gives us the size of our survey area. This estimate
should be good to within 1 sq.deg. There are 67,883 objects
from the SDSS DR4 spectroscopic survey (not restricted to
the main galaxy survey) that lie in this area, and are spectro-
scopically classified as galaxies. For each of these objects we
search for a match in the GALEX source catalog (which al-
ready combines FUV and NUV detections) within 4′′. Our
analysis of SDSS point sources with a match in GALEX in-
dicates that GALEX positions have a random error of 0.′′8 in
either R.A. or declination (becoming somewhat larger at the
edgesofthefield). Inaddition,thereareoveralltile-to-tileoff-
sets between GALEX and SDSS coordinatesystems of several
tenths of an arcsecond. In any case, astrometric uncertainties
are significantly smaller than our matching radius (see also
Trammell et al. 2007). A genuine match can be missed if the
centroid of the optical light is offset by more than 4′′com-
pared to the centroid of UV light. We expect such cases to be
quite rare, since at the mean redshift of the sample (z = 0.104)
this offset would translate into a 7 kpc separation. A problem
with matching in general is that what is considered to be a
single object in one catalog can be resolved into multiple ob-
jects in another catalog, whether they are indeed separate ob-
jects (blending),oractually belongto the same system (shred-
15Matching with the current SDSS data release DR5 (which uses the same
pipeline as DR4) would not significantly increase the overlap, since current
GALEX pointings mostly follow the footprint of SDSS DR1 and DR2.

Page 4

4 SALIM ET AL.
ding). Thisproblemis morepronouncedwhencombiningcat-
alogs with differentresolutions, as is the case here (4 to 5′′for
GALEX vs. 1.′′4 for SDSS). However, in our particular sam-
ple this problemis not severe. SDSS galaxies with spectra are
relatively bright objects, and if more than one GALEX object
is found within the search radius our procedure was to take
the one that is positionally the closest. Since we are dealing
with high-latitude fields where the density of foregroundstars
is not that high, this simple matching procedure produces re-
liable results. Besides, we do have mechanisms of identifying
the majority of incorrect matches at the later stage, when we
perform an SED fitting to the combined GALEX/SDSS pho-
tometry (see §4.3). Since we combine photometry from two
differentcatalogs there is a concern of photometriczero-point
mismatch. We explore such a possibility in §5.4. In our final
matched catalog we eliminate duplicate GALEX observations
(stemming from overlaps or repeat observations) by keeping
those that lies closer to the center of GALEX field of view.
3. THE SAMPLE
3.1. Sample selection
We initially define our sample by applying magnitude and
redshift cuts to galaxies with SDSS spectroscopy (note that
objects spectroscopically classified as QSOs are excluded).
We require the dereddened Petrosian magnitude to be in the
14.5 < rPetro≤ 17.77 range, where the faint limit is the nom-
inal limit of the SDSS main galaxy spectroscopy survey (see
also §4.4), while the bright limit is chosen to avoid objects
with saturated SDSS photometry. We require redshifts to lie
within 0.005 < z ≤ 0.22 range. The lower limit is chosen to
eliminate galaxiesthat may deviatefromthe Hubbleflow, i.e.,
whose redshift distances could be unreliable. Redshift dis-
tribution in SDSS main galaxy sample peaks sharply around
z = 0.1 with few galaxies beyond our upper redshift limit.
Thesemagnitudeandredshiftlimitsareidenticaltothoseused
in Brinchmann et al. (2004), to which we will be comparing
many of our results. We will refer to the above redshift range
as representing the “local universe”.
There are 49,346 galaxies that meet the conditions defin-
ing our initial sample. Good quality SED fitting (see §4.2)
is obtained for 98% of them (48,295 galaxies). In the rest of
the paper we will use only this latter sample and refer to it as
“all” galaxies. Note that since we retain objects regardless of
whether they were detected by GALEX (as long as they fall
within the area covered by GALEX), our sample is only opti-
cally (r-band) selected.
Throughoutthe paper we will use Ωm= 0.3, ΩΛ= 0.7, H0=
70kms−1Mpc−1(i.e., h70= 1) cosmology.
3.2. Emission line diagnostics and sample classification
This work focuses on the physical properties of galaxies,
such as their star formation rates (SFRs) and stellar masses.
Before we start analyzing the sample based on these derived
properties, we would like to characterize it in terms of its ob-
servable quantities.
Throughoutthis paperwewill relyonopticalemissionlines
to classify galaxies in our sample. This classification, based
on the BPT diagram (Baldwin et al. 1981), plots the flux ra-
tio of [OIII]λ5007Å and Hβ lines against the flux ratio of
[NII]λ6584Å and Hα lines. In this paper we fully adopt
the BPT classification of Brinchmann et al. (2004) (hereafter
B04), which is based on emission-line strengths corrected for
the underlying stellar absorption (see their Fig. 1). The posi-
tion of a galaxy in the BPT diagram indicates the nature of its
ionizing source. The majority of galaxies in the diagram fall
within one of the two branches. One branchlies mostly above
the maximum line ratios expected from star formation. These
galaxies must have some ionizing source in addition to young
stars. Most researchers attribute this emission to an active
galactic nucleus (AGN). Specifically, the emission is associ-
ated with a narrow-line AGN (a LINER or a Seyfert 2), also
known as Type 2 AGN (Kauffmann et al. 2003c). Note that
any Type 1 AGN spectrum (QSO or Seyfert 1) would have
broad lines, and would be classified by the SDSS pipeline
as a QSO and thus not included in this sample. Following
Kauffmann et al. (2003c), B04 classify galaxies that lie on
the lower portion of the AGN branch, but with line ratios al-
lowed by star formation, as “composite” (star forming/AGN)
galaxies. For galaxies that lie on the star forming branch we
expect very little emission line flux to come from an AGN
(Kauffmann et al. 2003c). B04 required a S/N ratio in all four
lines to be > 3 in order to apply the BPT diagram classifi-
cation. However, in some cases a secure AGN classification
is possible even when only [NII]6584 and Hα have S/N> 3,
i.e., in cases when this ratio is larger then the one allowed
by SF. B04 calls this class a “low-S/N AGN (LINER)”. Fol-
lowing B04, we will study this class together with the AGN.
B04 also introduce the category of “low-S/N SF” galaxies.
These are the galaxies that do not fall in any of the previous
categories because their lines have low S/N, yet they have an
Hα detection with S/N> 2. While their lines are not strong
enough for secure placement on the BPT diagram, if we nev-
ertheless do so, we find that many object lie in the high-mass
end part of the SF sequence, as well as in the lower portion of
the AGN branch. Therefore, these objects represent a hetero-
geneous class. Finally, there are galaxies without detectable
lines, thus precluding the classification in the above scheme.
B04 call this group “unclassifiable”. We will call them “No
Hα” class. We find that in ∼ 2% of galaxies in this class the
Hα non-detection is due to some artifact in the spectrum or
line-extraction pipeline. We exclude these galaxies from this
class (but not from the whole sample). The fraction of galax-
ies in our sample belonging to different classes is as follows:
Star forming (SF) – 27%, Low S/N SF – 19%, Composite
(SF/AGN) – 8%, AGN – 12% and No Hα – 33%.
3.3. UV-to-optical color-magnitude diagram
Color-magnitude diagram (CMD) is a powerful tool in as-
sessing the basic properties of a sample of galaxies. Histori-
cally, the study of CMDs was preceded by the studies of a re-
lationship between color and morphology. Basic morpholog-
ical segregation into disk-like spiral galaxies and spheroidal
elliptical galaxies was established by Hubble (1926). After-
ward (e.g., Hubble 1936) it was realized that spiral (late type)
galaxies have bluer colors than the ellipticals (early types).
Optical CMDs were constructed for cluster early type galax-
ies (e.g., Sandage 1972), where they were found to form a
narrow sequence (the so called red sequence).
ral galaxies also displayed the color-magnitude relationship
(Chester & Roberts 1964), albeit with larger scatter. The bi-
modal nature of the field galaxy CMD became much more
apparent with recent large scale surveys, in particular SDSS
(Baldry et al. 2004). However, unlike the traditional optical
CMDs, a CMD in which the color is constructed from an ul-
traviolet and an optical magnitude has a particular diagnostic
power. By virtueofcontrastingthe current(orrecent)star for-
mation as indicated by the UV light (modulo attenuation) to
Field spi-

Page 5

UV SFRs IN THE LOCAL UNIVERSE5
FIG. 1.— UV to optical color-magnitude diagrams (CMDs). Upper left panel shows a greyscale scatter plot of all galaxies in our sample with a near-UV
detection. (The shade of gray is directly correlated to the number of points contained in a given “pixel”.) Note the pronounced bimodality of the blue and the red
sequences, and their large separation. Dashed line represents the completeness limit at the mean redshift of our sample. The remaining panels show CMDs of
different galaxy classes (SF – star forming, Comp – star-forming/AGN composite, AGN – Type 2 AGN, SF (low S/N) – star-forming with weak Hα, and No Hα),
as determined from the position in the BPT diagram. The outer contour encompasses 90% of the entire sample is plotted for reference. While SF galaxies mostly
lie in the blue sequence, and those with no Hα in the red, most galaxies in between the two sequences are AGN or AGN/SF composites. Absolute magnitude is
given in z = 0 r-band, and the color is K-corrected to z = 0 rest-frame, as indicated by superscript 0.
the total past star formation as indicated by the optical light,
the UV to optical color is a good proxy of a galaxy’s SF his-
tory (e.g., Salim et al. 2005). On the other axis, one plots ab-
solute optical magnitude, which is related (modulo variations
in the optical mass-to-light ratio) to galaxy’s current stellar
mass. We present the observed CMD of our sample in Figure
1. In this and many subsequent figures, the individual data
points have been converted into a greyscale density scatter
plot, in such a way that the shade of gray is proportional to
the number of objects occupying a given “pixel”. Such repre-
sentation is desirable when dealing with large samples where
it is easy to saturate a traditional scatter plot. Unless noted
otherwise, the figures are constructed from raw, unweighted
counts. We will include volume corrections later, when ap-
propriate. Both the color and the absolute magnitude have
been K-corrected to z = 0 rest-frame bands (§4.4). The up-
per left panel shows the CMD of all galaxies with an NUV
detection (85% of the total sample; see Table 1 for a break-
down of UV detection rates per class). A striking feature of
a UV to optical CMD is the pronounced bimodality of blue
and red galaxies. Blue galaxies form a well-defined sequence
extending to faint luminosities. The red sequence is some-
what more narrow than the blue sequence (note that this be-
comes evident only after K corrections have been applied),
and extends to intrinsically more luminous galaxies than the
blue sequence. This is related to well-known fact that the
most massive elliptical galaxies are more massive than the
most massive spirals (Holmberg 1965). The two sequences
are separated by ∼ 3 mag. If the two sequences are modeled
as gaussians, one finds that there is an excess of galaxies in
the gap (Wyder et al. 2007). This is not the case in classical
TABLE 1
AVERAGE ERRORS OF “UV” AND “Hα”-BASED STAR FORMATION
RATESa
Class Number UV detectedb
“UV”“Hα”
?σ(log SFR)??σ(log SFR)?
Allc
SF
SF (low S/N)
Comp
AGN
No Hα
48295
12901
9060
3966
5827
16159
86%
99%
93%
96%
90%
68%
0.38
0.20
0.30
0.28
0.41
0.60
0.43
0.29
0.39
0.40
0.49
0.54
aUV-based SFRs are averaged over 100 Myr.bEither FUV or NUV detection.
Of objects detected in FUV, 98% are also detected in NUV.cIncludes 382
objects for which classification was not possible, see §3.2.
optical CMDs (Baldry et al. 2004). We will refer to the gap
region and its population as the “green valley” (Martin et al.
2007). A detailed quantitative study of the GALEX UV to
optical CMD is presented in Wyder et al. (2007).
In the subsequent panels in Figure 1 we display the CMDs
of various classes of galaxies as defined in §3.2. Each panel
has the greyscale normalized to the number of galaxies in the
given class. For better reference with respect to the full sam-
ple, in each subsequent panel we repeat the contour contain-
ing 90% of all galaxies. Not surprisingly, the star-forming
galaxies (SF) occupy the blue portion of the CMD. Part of the
width of the sequence is due to the intrinsic dust attenuation.
The CMD of composite galaxies (showing both signatures
of SF and AGN) are shown in the upper right panel. Their
NUV −r colors are offset to the red compared to those of the
“pure” SF galaxies. Galaxies with narrow line AGN (lower

Page 6

6 SALIM ET AL.
left panel) occupy the regions of the red sequence and of
the green valley (Martin et al. 2007; Kauffmann et al. 2007).
Moreover, most galaxies with intermediate colors are either
AGN or composites. Star formation histories of AGN will be
discussed more extensively in §7.3. Next, we have low-S/N
SF galaxies (middle lower panel), which indeed mostly lie on
the blue, star-forming sequence (preferentially its more lumi-
nous part). However, there is a tail of red galaxies which are
probably contaminated by AGN as observed in §3.2. Finally,
we have galaxies with no detectable Hα (lower right panel).
As expected,these galaxiesformthe bulk ofthe red sequence.
However, there is a tail of galaxies of this class extending to-
wards the blue colors. Since the spectral classification used
here is nuclear (central 3′′) it is possible that some of these
galaxies are classified as not having Hα because of the domi-
nant bulge, while the relatively faint star forming disk is giv-
ing a galaxy an overall blue color. However, as we discuss in
§7.4, this is not the case for most of them. The CMDs pre-
sented in Figure 1 require a detection in NUV. The remaining
15% of our sample with no NUV detection falls mostly in the
red sequence (as evidenced from their u−r colors). These
galaxies are too faint to be detected in the UV.
4. OBTAINING GALAXY PROPERTIES WITH SED FITTING
4.1. SEDs of model galaxies
Spectral energy distribution (SED) fitting is becoming
a widely used technique for deriving galaxy properties.
It was pioneered in works of Searle et al. (1973) and
Larson & Tinsley (1978). To first order it consists of com-
paring the observed SED to a set of model or template SEDs,
and searching for the best match. It is assumed that the phys-
ical properties of model or template galaxies are known, and
that one can use this knowledgeto deduce the propertiesof an
observed galaxy.
In this study, we compare the observed SED to a large
number of model SEDs constructed from Bruzual & Charlot
(2003) population synthesis code. To construct the models
we first choose input model properties by randomly select-
ing values from prior distributions defined in the following
way. Formation of a galaxy is uniformly distributed between
0.1 Gyr and the age of the universe. Global metallicity is
uniformly distributed between 0.1 and 2 Z⊙. Star formation
histories are not single stellar populations, but the combi-
nation of an exponentially declining continuous star forma-
tion, SFR(t) ∝ exp(−γt), with 0 ≤ γ ≤ 1Gyr−1uniformly dis-
tributed over that range, and the random starbursts superim-
posed on the continuousSF. Bursts are parametrizedto have a
duration uniformly distributed in the 30–300 Myr range, with
a strength such that the mass produced in a burst is between
0.03 and 4 times the mass producedin the continuousSF over
the present lifetime of a galaxy (distributed uniformly in log).
Such parametrizationconforms with most observationalstud-
ies. Finally, bursts are randomly produced so that the proba-
bility that a givengalaxyhas undergoneat least one burst over
a 2 Gyr period is 50%. The above parameters define how the
populationsynthesis for each model will be carried out. Next,
each of the above model SEDs is subjected to dust attenua-
tion parametrized according to the two-component model of
Charlot & Fall (2000). For the V-band optical depth we ran-
domly chose a value from a distribution 0 < τV< 6 (it peaks
at τV= 1.2 and has most values in the 0 < τV< 2 range). The
prior distribution for τVis empirical, and comes from Balmer
decrements in SDSS spectra. The choice of τVprior distribu-
tion will be discussed in §5.4. Fraction of the optical depth
that affects stellar populations older than 10 Myr (i.e., most
of the UV continuum flux) is given by factor µ, with values
ranging from 0.1 to 1, peaking around µ = 0.3. Altogether,
we produce 100,000 model spectra spanning a range of SF
histories, metallicities and dust attenuations. Note that since
we pick input parameters randomly, we do not call our set of
modelsa grid,whichwouldsuggestsa set ofpointswithequal
spacing in parameter space.
Finally, we convolve the resulting model spectra with
GALEX and SDSS bandpasses at five redshifts equally spaced
in the [0.05, 0.25] interval, producing the libraries of model
broad-band photometry. In each library we keep only mod-
els that have an age smaller than the age of the universe at
that redshift. This effectively reduces the number of models
from 95,000 at z = 0.05 to 78,000 at z = 0.25. We also add
effects of the interveningintergalactic extinction according to
Madau et al. (1996). Our final libraries list model photom-
etry as well as a number of galaxy parameters associated to
that model (such as the SFR averaged overseveral timescales,
stellar masses, dust attenuation parameters, etc.)
4.2. SED fitting
In our study, the observedSED is constructed from GALEX
and SDSS broadband photometric fluxes (the broad-band
SED). For an observed galaxy at some redshift we select the
model library with the closest redshift. We step through a li-
braryonemodelat the time. Modelfluxpointswill havesome
arbitrary zero point, and in order to see how well an observed
flux compares to the model flux we first need to find a factor
a that minimizes the χ2between the observed and the model
points. In other words, for model i, we need to minimize the
following expression:
χ2
i=
?
X
?Fobs,X−aiFmodi,X
σ(Fobs,X)
?2
.
(1)
Here, the summation is over X = (FUV,NUV,u,g,r,i,z), the
observed flux points are Fobs, and their errors σ(Fobs). Flux
points of model i are Fmodi. By taking the derivative of Equa-
tion 1 with respect to a, and equating it with zero, we find the
scale factor a∗that best matches the observed and the model
flux points:
a∗
i=
?
?
X
Fobs,XFmodi,X
σ2(Fobs,X)
?
X
Fmodi,X
σ(Fobs,X)
?2.
(2)
Because we include non-detections that have meaningless
fluxes (small or even negative values), we use the observed
fluxes without correcting for Galactic reddening, and instead
apply the reddening to model fluxes, which are always pos-
itive. This allows us to treat non-detections like any other
flux point with a known flux error. Reddening corrections for
SDSS magnitudes come from SDSS database, and those for
GALEX from relations in Wyder et al. (2007) (Their Eq. 2 for
NUV).
The goodness of the fit between an observed galaxy and
modeli is then obtainedby insertinga∗
tion 1. Note that only a single parameter is being fit—the
scale factor a. We emphasize that the word “parameter”when
fitting is discussed should not be confused with (in princi-
ple) arbitrary number of galaxy parameters (i.e., properties)
that correspond to each model SED. Therefore,since we have
seven photometric points, there are six degrees of freedom.
iin place of aiin Equa-

Page 7

UV SFRs IN THE LOCAL UNIVERSE7
For a given galaxy we evaluate χ2
lined above. Now, each of these χ2
the relative weight wifor the given model i (and therefore the
weight of the galaxy parameters associated with that model)
as wi= exp(−χ2
function(PDF)foreverygalaxyparameterofinterestbycom-
pounding weights at the corresponding parameter value. We
repeat the procedureforeach of the ∼105models in the given
redshift library. We normalize the final PDFs and note the pa-
rameter values corresponding to the 2.5, 16, 50 (median), 84
and 97.5 percentiles of the cumulative PDF. We use the av-
erage of the PDF as our nominal estimate of the parameter
value. Since in the general case the PDF is not symmetric,
medians, averages and modes will differ. For most relevant
parameters in this study the differences between the median
and the average are small. We do not consider using a mode,
because due to the discreteness of the model parameter space,
the mode can be stochastically offset from the bulk of the
PDF, and is sensitive to PDF binning. For similar reasons
we do not use a single model with the best χ2(i.e., the maxi-
mum likelihood model) as representativeof parametervalues,
but we do keep track of the best χ2values in order to evaluate
the overall quality of the fitting for a given galaxy (§4.3). Ad-
ditionally, for certain parameters we preserve more detailed
information on the shape of the PDF (in the case of the spe-
cific SFR and stellar mass, we actually keep their mutual two
dimensional PDFs). In certain cases we will discuss the er-
rors of the derived parameters. We use 1/4 of the 2.5-97.5
percentile range as a proxy for what would have been a 1σ
error in the gaussian distribution. We will refer to these quan-
tities as “formal” errors.
We prefer the above-described Bayesian approach (e.g.,
Leonard & Hsu 2003) to often-used maximum likelihood,
since it allows one to transparently test the dependence on
the prior, i.e., how well the data constrain certain parame-
ters given that the shape of the resulting PDF can to some
extentdependon the distribution of the givenparameterin the
model library (especially when the observational constraints
are poor). Another advantage is that while the maximum-
likelihood model might indeed correspond to a high proba-
bility density, in the case of extensive models such as ours, it
might actually correspond to a negligible probability "mass",
and might therefore lead to misleading parameter estimation
(see for instance the discussion in MacKay 2003). Note also
that our technique properly accounts for the degeneracies in-
herent in some galaxy properties. Imagine that we have two
galaxy parameters whose values can be picked in such a way
as to produce identical model SEDs. In such case our fitting
will give equal probabilities to all of the degenerate models,
and the resulting PDFs for each of these two parameters will
thereforereflecttheentirerangethattheseparameterscantake
(i.e., there will be nothing to constrain the PDFs). If we then
characterize the errors of those parameters using the width of
the PDF, we will obtain very large values, i.e., we will know
that a certain parameter is poorly constrained. This, in effect,
is in contrast to fitting using a single best-fit model, which
would pick one of the degenerate models as the best-fitting
(and the correspondingparameter value), without giving a re-
searcher an idea that there is an underlying degeneracy mak-
ing the obtained value to be quite uncertain. Finally, we want
to emphasize that the widths of the PDFs (and also their typ-
ical values) have a meaning only if the observational errors
have been estimated reasonablycorrectly,and if the overall fit
ifor each model as out-
ivalues will determine
i/2). We then build a probability distribution
(which can be well characterized by the χ2of the best-fitting
model) is reasonably good (§4.3).
4.3. Quality of a fit
For our initial sample of 49,346 galaxies we obtain model
parameters in the way described in §4.2. For each galaxy we
also note the χ2of the single best-fitting model, χ2
value serves as an indicator of the overall quality of the fitting
for a given galaxy. Since we include galaxies close to the
edgeof GALEX field of view wherewe knowthat image PSFs
get distorted, we would first like to check if this affects the
photometry, and therefore the quality of SED fitting. Thus
we look at the average χ2
GALEXimagecenter. Wefindnocorrelationwiththedistance
or degradation close to the edge.
In the ideal case, χ2
distribution. The average of such distribution should equal
the degrees of freedom (in our case 6). Before looking at
the χ2distribution we notice that χ2
cally higherfor blue galaxiescomparedto the red. We believe
that this difference arises from the presence of emission lines
in blue galaxies that effectively increase the discrepancy be-
tweenthe observedSED andthe model(line-free)continuum.
Therefore, we start by comparing the observed and the theo-
retical χ2distribution of red galaxies. Initially, the observed
distribution was shifted towards the values that were too low.
This was indicativeof photometryerrorsbeingoverestimated.
We arrive at the χ2distribution that was well-matched to the
theoretical one when we adopt 0.01 mag calibration errors in
each of the 5 SDSS bands.16While now the bulk of the red
galaxies follows a χ2-like distribution, there is also present a
tail of high χ2
distribution. One can imagine that the main reason for the
presence of χ2outliers is that the data do not reflect the actual
SEDforsomereason. We will investigatepossiblecausesmo-
mentarily. The high-χ2tail of red galaxies contains some 2%
of objects. We decide to exclude these galaxies from further
analysis (although we do account for them statistically when
makingany kind of space density calculation). Assuming that
the reasons driving extreme χ2values are color independent,
we also clip 2% of the blue-galaxy χ2
leaves us with 48,295 galaxies—our final sample.
While removingthe outliers is necessary to have results not
be biased by unreliable values, it is instructive to understand
what causes some χ2to assume such high values. We thus
examine 22 galaxies with extreme χ2
evenlyin the −1<NUV −r <7 range (thus samplinggalaxies
with various SF histories). In 10 cases we find obvious prob-
lems with cataloged photometry. This is most often the case
with SDSS (9 objects), where the visual inspection of broad-
band SEDs and comparison with spectra indicates that some
flux points are large outliers. Next, in 7 cases we are deal-
ing with either close galaxies being blended in GALEX (4), or
with SDSS shredding galaxies into smaller photometric ob-
jects (2). For one object it is unclear whether we have blend-
ing or shredding. Further, in one case an object shows a clear
QSO spectrum, i.e., the continuum is affected by the AGN,
yet it is spectroscopically (mis)classified as a galaxy. Finally,
best. This
bestas a function of distance from the
bestvalues would form a theoretical χ2
bestvalues are systemati-
bestvalues, in numbers well above any expected
bestdistribution tail. This
bestvalues, distributed
16The “official” SDSS calibration errors are given as 2% for r band, and
between 2% and 3% for each of the four colors. However, the individual
magnitude errors (which figure in Equation 1) are correlated, which is why
we find them to be smaller than 2%. Our results stay practically the same
even if errors are increased, but the χ2distribution then deviates from the
expected one.

Page 8

8 SALIM ET AL.
in four cases it appears that the broad-band SEDs (and spec-
tra) are genuinely unusual. While their number is too small to
affect the present study, this type of outliers deserves a spe-
cial scrutiny. It needs to be explored if there exists any set of
model input parameters (SFH, dust attenuation, metallicity)
beyond our already wide range, which can produce a model
SEDmatchingtheseobservations. Forourfullsamplewealso
expectontheorderof10caseswhereaforegroundstarofsim-
ilar brightness is superimposedwith a galaxy, and falls within
the fiber, thus producing a “composite” galaxy/star spectrum
which would fail the fitting.
Since our model libraries are calculated at some fixed red-
shifts, there is some error associated with the library redshift
not perfectly matching the galaxy redshift. We thus contrast
the χ2
set (close to 0.025) to those with no offset. We find a system-
atic increase in χ2
overall χ2distribution.
Finally, we want to assess the effects of the “resolution” of
our model parameters, i.e, we can ask how extensive should a
model library be that the resolution is not an issue. To that ex-
tent,weartificiallydegradeourmodellibrariesbyrunningfull
SED fitting using every 2nd, every 4th and every 8th model
SED. We then compare average χ2
Overall, the quality of the fits are surprisingly stable. The run
with every 2nd model produces only a 4% increase in χ2
values, while the run using every 16th model is 16% worse
compared to the full library. This implies that increasing the
extent of the library (for the given range of input parameters)
would not bring significant improvements. This, of course, is
not a general conclusion, but rather states that with the pre-
cision of our data the current number of models is adequate,
i.e.,havinga finerresolutionofinputmodelparameterswould
create model SEDs that are degeneratefrom the observational
point of view. In addition to this, the Bruzual & Charlot
(2003) models themselves are not likely to be of such an ac-
curacy to warrant more extensive model library.
bestvalues of galaxies having the maximum redshift off-
bestof 10%, which is too low to affect the
bestvalues from each run.
best
4.4. K corrections and volume corrections
For deriving the galaxy properties from the SED fitting, we
are not required to know the K corrections. This is because
for a galaxy at a given redshift we compare the observed
magnitudes to model magnitudes with bandpasses that have
been shifted to the matching galaxy’s redshift (i.e., the red-
shift of the model library). Therefore, the “K corrections”
are present in the SED fitting implicitly. However, for some
applications, such as constructing a CMD (§3.3), or obtain-
ing Vmaxweights (§§7,8), we need to know the K corrections
explicitly. The process of SED modeling allows us in prin-
ciple to derive the K corrections alongside galaxy parame-
ters. It even has the advantage of allowing one to construct
a PDF for any K-corrected magnitude, which thus yields an
estimate of a K correction error. We find this approach to
be especially useful in K correcting higher redshift (z ∼ 1)
samples (Salim et al., in prep.) In the current case, given
that our redshift range is covered by only 5 model libraries,
we choose to calculate the K corrections using the publicly
available code KCORRECT V4_1_417(Blanton & Roweis
2007). The code allows GALEX magnitudes to be used to-
gether with SDSS magnitudes to constrain the SED fit from
which K corrections are derived. We obtain K corrections us-
ing the same combined GALEX/SDSS photometry used for
17Available from http://cosmo.nyu.edu/blanton/kcorrect.
the SED fitting, except that < 3σ UV detections are not used.
We derive K corrections without a priori assuming an evolu-
tion of the luminosity function. We will explicitly correct for
evolution where appropriate.
In cases when we require a volume-corrected sample, we
assign a weight to each galaxy according to its Vmaxvalue—
the volume in which a galaxy would be visible taking into
account redshift and apparent magnitude limits, and the solid
angle of a survey. While our faint limit (r = 17.77) is also
the nominal limit of the SDSS main spectroscopic survey, we
takeintoaccountthattheactualspectroscopicmagnitudelimit
varies from one spectroscopic plate to another (i.e., that it can
be brighter than r = 17.77). Thus, for each object we mod-
ify the volume by the spectroscopic completeness (usually
around 0.9). We take spectroscopic limits and completeness
values from NYU-VAGC catalog (Blanton et al. 2005). We
caution that in order to calculate Vmaxcorrectly, it has to be
done iteratively, because the K correction depends on redshift
limits that determineVmax.
5. COMPARISON OF “UV” AND “Hα” STAR FORMATION RATES
5.1. “UV” SFR estimates
We derive our SFRs from photometric constraints that ex-
tend from the UV to the z-band. However, it is the UV lu-
minosity that constrains the SFR the most. Therefore, in the
remainder of the paper we will refer to SFRs derived from
the SED fitting as “UV”, where quotation marks stand to re-
mind us that it is not only UV information that contributed to
these SFR estimates. In our model SED libraries we report
“current” star formation rates averaged over several time in-
tervals, most notably over the most recent 10 Myr, 100 Myr
and 1 Gyr. By comparing the formal errors of SFR estimates
(derivedfrom95 percentilerangeof each galaxy’s SFR PDF),
we find that the best-constrained SFR for the overall sample
is the one averaged over the last 100 Myr (Table 1), which
is the timescale for UV bright stars. For the star-forming
class specifically, we find that the SFRs have slightly smaller
formal errors over 1 Gyr timescales than over 100 Myr. On
the other hand, SFR estimates on timescale of 10 Myr for SF
galaxies (a timescale that matches Hα SFR) have drastically
larger formal errors, as expected since we do not have infor-
mation to constrain them. In the remainder of the paper we
will use SFR estimates averaged over 100 Myr, which repre-
sent a good compromise between the quality and a timescale
that is not too long compared to that of the Hα SFRs.
5.2. “Hα” SFR estimates
B04 have developeda methodof derivingSFRs fromSDSS
spectra that is primarily based on the intensity of an Hα
line. This represents the largest currently available sam-
ple of precise Hα-based SFRs.
from SDSS spectra is far from straightforward.
purposes of this paper we briefly describe the procedures
employed in B04. The essence of the B04 approach was
to model SDSS spectra by first removing the absorption
line spectrum using a combination of burst models from
Bruzual & Charlot (2003), and then to model the emission
lines using the Charlot & Longhetti (2001) models which
combinetheBruzual & Charlot (1993) galaxyevolutionmod-
els with the CLOUDY photoionization code (Ferland 1996).
Also, the same two-component dust attenuation prescription
of Charlot & Fall (2000) was used by B04 to model attenua-
tionofthe emissionlines, andbyus tomodelbroad-bandcon-
tinuum attenuation. While conventionally the attenuation of
Obtaining accurate SFRs
For the

Page 9

UV SFRs IN THE LOCAL UNIVERSE9
Hα flux is determined by comparing the observed Hα to Hβ
ratio (Balmer decrement) to the theoretical one, the attenua-
tion estimate in B04 is constrained using many emission lines
(although it is dominated by Hα/Hβ). Using a suite of mod-
els, B04 applya Bayesianapproachto producethe probability
distributions for each of the four parameters (gas metallicity,
ionization parameter, dust attenuation, and dust-to-metal ra-
tio). B04 thus simultaneouslyproduce attenuationestimate as
well as the attenuation-corrected SFR within SDSS 3′′fiber.
SFR within the fiber is constrained by many emission lines,
with the greatest weight carried by Hα. This is one reason
why we will denote B04 SFR and attenuation estimates by
“Hα”, with quotation marks again serving to indicate that not
only Hα was involved in those estimates. The above proce-
dure is directly applicable only to galaxies for which the ion-
izing source is predominantly star formation, and not, for ex-
ample, an AGN activity. B04 therefore calculate fiber SFRs
directly from emission lines only for galaxies classified in the
BPTdiagramasstar formingandlow-S/Nstarforming(§3.2).
For other classes of galaxies (AGN, composite, galaxies with
no Hα) for which either there is no emission line detection or
the lines are contaminated by a non-SF ionizing source, B04
usetherelationbetweentheD4000spectralindexandthespe-
cific SFR (SFR normalized by stellar mass) that has been cal-
ibrated using the star-forming galaxies. Therefore, for these
galaxies, emission-line SFRs are used indirectly. This is an-
other reason why we denote B04 SFRs as “Hα” with quota-
tion marks. We will return to this calibration in §5.3. The
above procedure gives only the SFRs within the 3′′aperture
of SDSS spectroscopic fibers, i.e, the fiber SFRs. In order to
obtaintotalSFRs, an aperturecorrectionis required. Thetotal
SFR is the sum of the SFR within the fiber and outside of it.
To estimate SFR outside of the fiber, galaxies are first divided
into a color-color grid, based on their fiber colors. Then, for
each color cell, a distributionof i band-luminositynormalized
SFRs (SFR/Li) is constructed by adding up all galaxies of SF
or low-S/N SF class that fall within the givencolorcell. Then,
the color of a galaxy outside of fiber is used to select the ap-
propriate color cell, and the SFR/Liin that cell is used with
the i-band luminosity to derive the SFR outside of fiber. This
is addedto previouslyfoundSFR withinthe fiberto obtainthe
total SFR. In other words, the aperture correction procedure
is based on two assumptions. First, that SFRs within fibers
have the same dependency on color as the SFR outside of it,
andsecond,that this dependency,calibratedwith star-forming
galaxies, applies to other galaxy classes. The aperture correc-
tion factors range from close to one (no correction) to around
a hundred-fold. On average they are 0.9 dex for the entire
sample and 0.6 dex for the star-forming class.
5.3. Comparison of “UV” and “Hα” SFRs for all galaxies
The comparison between SF indicators serves the obvious
purpose of providing a better understanding of each indica-
tor, but also ensures the mutual check on the reliability of the
techniques used to produce the SFR estimates—in this case,
the SFRs based on UV (this study) and on Hα (from B04).
We begin by comparing the formal errors of the two meth-
ods. Note that B04 used a sample based on the expanded
SDSS DR1. Here, when we refer to B04 SFRs we mean the
B04calculationsappliedtoSDSSDR4, andavailableas apart
of MPA/JHU SDSS value-added catalog. In Table 1 we show
average formal errors in SFR for the entire sample, and for
each galaxy class. We see that the formal errors of the two SF
estimates happen to be comparable.
FIG. 2.— Comparison of dust-corrected SFRs derived by Brinchmann et al.
(2004) (B04, “Hα”) and the dust-corrected SFRs from this study (“UV”).
Comparison is given for all galaxies in the sample. Our (“UV”) SFRs come
from the modeling of the broad-band SED (two GALEX UV bands and five
SDSS optical bands), and span almost 5 orders of magnitude. While most
galaxies compare reasonably well, those with low “UV” SFRsare quite offset
from the equality line. The error bar represents average formal errors (from
Bayesian fitting) of the two estimates. Both estimates are given for (Chabrier
2003) IMF.
FIG. 3.— Comparison of B04 (“Hα”) and “UV” dust-corrected SFRs for
low-S/N star-forming, SF/AGN composite, AGN, and galaxies without Hα
detection. Note that except for low-S/N SF class, the B04 estimate comes
from relations calibrated using the “Hα” SFRs of galaxies classified as star
forming. In any class the discrepancies are particularly large when the “UV”
estimate of SFR is low. Error bars represents average errors in each class.
To facilitate the comparison,we first convertB04 SFRs that
were calculated for Kroupa (2001) IMF to Chabrier (2003)
IMF used in this study (both with 0.1-100 M⊙limits). From
Bruzual & Charlot (2003) models we find the conversion fac-
tor of 1.06 (in the sense that Kroupa IMF SFRs are slightly
higher). In Figure 2 we plot B04 “Hα” SFR estimates against
the “UV” for all galaxies in the sample (throughout the pa-
per SFR always means dust-corrected SFR). Note that the
“UV” SFRs exhibit 5 orders of magnitude of a dynamic

Page 10

10SALIM ET AL.
range—fromall but negligiblerates to those approaching100
M⊙yr−1. We notice that a large number of galaxies (those
with higher “UV” SFRs) has comparable SFRs. However,
there is a plume of galaxies for which B04 SFRs are up to
2 orders of magnitude higher than the “UV” ones. This sys-
tematic discrepancy cannot be explained by random errors of
the two methods. Recall that B04 do not derive SFRs di-
rectly for classes other than star-forming (and low S/N star-
forming),but instead use relationscalibratedon SF class sam-
ple. Thus, in Figure 3 we break the comparison into galaxy
classes (omitting for the moment the SF class). For the low-
S/N star-forming class (upper left panel), for which B04 SFR
estimates do come from the emission lines, the comparison
is generally good for the majority of galaxies (apart from the
overall offset that makes “Hα” SFRs higher). For the other
three classes in Figure 3, B04 SFRs rely on calibrations made
using the SF galaxies. Composite (SF/AGN) galaxies com-
pare relativelywell, but the comparisonbecomes significantly
worse for AGN, where we can see both the significant offset
for galaxies with higher SFRs, and the prominent plume for
those with lower “UV” SFRs. Finally, for galaxies for which
no Hα detections were possible, the discrepancy is very large
foralmost all galaxies. It is this class that contributesthe most
to the plume seen in the comparison of all galaxies (Figure
2). Given that the galaxies without Hα detections belong al-
most exclusivelyin the red sequence (Figure 1), the low SFRs
rates estimated by the SED fitting (“UV”) appear more realis-
tic (see also §7.1). Also, such high SFRs as estimated by B04
would result in a high fraction of No Hα galaxies to have a
detectable Hα emission, which is obviously not the case. In
general, it appears that the largest discrepancies between the
“Hα” and “UV” rates occur for galaxies with low SFRs, es-
pecially when B04 estimate SFRs based on calibrations that
employed SF galaxies.
B04 SFRs are the sum of SFRs within and outside of fiber.
To trace the source of the SFR discrepancies, we revisit the
calibrations used to derive B04 SFRs for non-SF classes.
First, within the fiber, B04 base the SFRs for non-SF classes
on the relation between the specific SFR (in the fiber) and the
D4000 index calibrated using the SF and low-S/N SF galax-
ies. We show this calibration in both panels of Figure 4 as the
dashed line (this can be compared to their Figure 11). Then,
B04 use D4000 to determine the specific SFR (and therefore
the SFR itself) of galaxies of other classes. In the upper panel
of Figure 4, we compare B04 relationship to our (“UV”) rela-
tionshipsforthestarforming(boldsolidline)andthelow-S/N
star-forming (thin solid line) galaxies. Note that in principle
we cannot compare the B04 specific SFR inside of fiber to
the total specific SFRs obtained from the UV, so the compar-
ison is more for illustrative purposes. At D4000 < 1.7 either
SF or low-S/N SF lines from “UV” compares well to B04.
However, notice that there are no galaxies classified as star
forming beyond D4000 = 1.9. Therefore, from that point on,
B04 relationship is based only on low-S/N SF class which
are most likely just AGN contaminants (see §3.2). In con-
trast to B04, the “UV” relationship for low S/N SF continues
to decline as D4000 increases. In the lower panel of Figure
4 we show the same B04 calibration against “UV” rates for
the AGN (thin solid line) and galaxies with no Hα detections
(bold solid line). In these cases, the relationship for these two
types as indicated by the “UV” is entirely different from that
of the star forming galaxies, even at lower D4000 values. For
clarity we omit showing the composite class. Altogether, this
leads toa conclusionthat the relationshipbetweenthespecific
FIG. 4.— Relationships between SFR/M∗(the specific SFR) and the D4000
spectral index. In the upper panel we compare the relationship from B04
(constructed from SF and low-S/N SF galaxies combined ) to our “UV” rela-
tionships for the SF and the low-S/N SF class, separately. The B04 relation-
ship comes from measurements within the fiber, while others are total, so this
comparison is primary an illustration. B04 use such relationship to determine
fiber SFRs of non-SF classes of galaxies. Note that beyond D4000 = 1.9, the
B04 calibration is actually based on low S/N spectra which may be contami-
nated by AGN. In the lower panel the same B04 relationship is compared to
our “UV” rates for the AGN, and for galaxies with no Hα detection. The two
follow different relationships, with AGN and “No Hα” classes having much
lower specific SFRs than the SF class. The non-unique mapping between the
specific SFR and the D4000 index for various galaxy groups, and the fact that
the correlation cannot be established well using the emission lines indicates
that fiber SFRs in B04 may be systematically affected.
FIG. 5.— Relationship between the luminosity-normalized SFR and op-
tical colors. B04 use luminosity-normalized SFRs binned in g−r vs. r −i
color-color grid to estimate SFR outside of fiber. They build their relation-
ships using SF galaxies, but apply them to other types as well. Here, based on
“UV” measurements, we show that different classes of galaxies follow dif-
ferent relationships, and that applying the one from SF galaxies would lead
to systematic differences, which cause most of the SFR discrepancies seen in
Figure 3.

Page 11

UV SFRs IN THE LOCAL UNIVERSE11
SFR andthe D4000indexis not uniquefordifferentclasses of
galaxies. We tested if this is an artifact of the fact that D4000
comes from the central 3′′, while the specific SFR for “UV”
is integral. We find that the relationships are qualitatively the
same for z < 0.07 and for z > 0.12 samples for which D4000
probes differentphysical sizes. Instead, the most likely expla-
nation is that since D4000 and specific SFR are sensitive to
SF activity on different timescales, the galaxies with differing
SF histories will have different relationships. To much lesser
extent, some of the differences could be due to metallicity
(Poggianti & Barbaro 1997).
Given the level of aperture corrections (a factor of several),
the analysis given above does not provide the explanation
for the large discrepancies we see when comparing the to-
tal SFRs. Since the SFR outside of fiber dominates in B04
total SFR, we will now focus on the calibration B04 use to
determine it. As outlined in §5.2 B04 aperture correction re-
lies on the calibration of the luminosity-normalized SFR as a
function of g−r and r−i color. Using our UV-based SFRs
we can test one of the two assumptions behind B04 aper-
ture corrections—whetherthe SFR/L calibration againstcolor
holds for different classes of galaxies. In Figure 5 we plot
SFR/L against two colors that B04 use to determine SFRs
outside of fiber. The bold line represents the relation from
galaxies classified as star-forming. This is the basis for B04
calibration which they apply to other classes as well. The
dashed and the thin solid lines show the relations for AGN
and galaxies with no Hα. We see (left panel) that as the g−r
color increases, the discrepancybetween SF and other classes
rises, and reaches some 0.8 dex for AGN and 1.3 dex for No
Hα. Similar levels of difference are present against r−i (right
panel). The level of these differences matches the discrepan-
cies (plumes) in the comparison of B04 and “UV” SFRs in
Figure 3. We can interpret these differences as reflecting the
fact that the colors used for B04 calibration trace old popula-
tions, while the current SFR traces young populations, and it
is not surprising that the two will differ for different classes
of galaxies. Also note that the discrepancies of specific SFRs
of different classes of galaxies are qualitatively similar in the
case of D4000 and in the case of g−r color, which is not sur-
prising since they are both sensitive to population age on the
similar timescales, and are similarly not affected by the dust.
To conclude, UV-derived SFR is subject to fewer limitations,
soit canbeappliedtoa morediversetypesofnormalgalaxies.
5.4. Comparison of “UV” and “Hα” SFRs for star-forming
galaxies
To avoid the problems indicated in the previous section, we
need to compare the SFR estimates from the “Hα” and the
“UV” methods in galaxies where Hα is well-detected (S/N
> 3) and arises predominantly from star formation, i.e., to
galaxies classified as star-forming (SF). In Figure 6 we show
the B04 “Hα” star formation rates plotted against our “UV”.
We see that the comparison, spanning some three orders of
magnitude in SFR, is exceptionally good. Formal error bars
are comparable, with the “UV” being somewhat smaller (see
Table 1). The scatter (standard deviation of the difference)
of the two measurements is 0.50 dex. When 3σ outliers are
excluded, the scatter is reduced to 0.36 dex. This is very well
matched by the sum (in quadrature) of the formal errors of
the two methods (0.35dex), confirmingthat the two measures
are predominantly independent. There is an average offset
betweenthe two SFRs of only 0.06dex in the sense that “Hα”
SFR is higher (which reduces to 0.02 dex, i.e., 5%, when 3σ
FIG. 6.— Comparison of B04 (“Hα”) to our “UV” SFRs (both indepen-
dently dust-corrected) for galaxies classified as star-forming. The two com-
pare very well on one-to-one basis. Also, the scatter is compatible with each
measurement’s errors.
FIG. 7.— The difference of B04 (“Hα”) and our “UV” SFRs, for galaxies
classified as star-forming, as a function of aperture correction applied to B04
SFRs. There is no apparent trend, suggesting that the aperture corrections
derived by B04 are quite robust. Running average (in 0.05 dex bins) is shown
as a thick line, and the ±1σ range as dashed lines.
outliers are excluded).
However, Figure 6 alone can potentially hide some system-
atic trends between the two SFR estimates. Where can the
differences in the SFR estimates arise from? Deriving SFRs
for either methods incorporates several steps: (a) obtaining
full “Hα” or “UV” observed luminosity (which in the case
of “Hα” involves aperture corrections), (b) correcting the ob-
served luminosity for dust attenuation, and (c) converting the
luminosity into a SFR. While both methods perform these
threesteps simultaneously,we can still designtests thatwould
differentiate between. Given in some cases the large aperture
correction applied by B04, we first plot in Figure 7 the dif-
ference in SFR estimates (SFR residuals) with respect to the
level of aperture correction applied to “Hα” rates. We see
no correlation. We also check, but do not find a correlation
between the residuals and the apparent sizes of galaxies (not
shown). Those are good indications that the aperture correc-
tions of B04 are quite reliable for galaxies classified as star-
forming. Finally, we find that the SFR residuals exist when
plotted against the stellar mass of a galaxy (shown in Figure
8). We have a change of SFR residual of 0.38 dex over the

Page 12

12SALIM ET AL.
FIG. 8.— The difference of B04 (“Hα”) and our “UV” SFRs, for galaxies
classified as star-forming, as a function of galaxy stellar mass. There is a
clear trend with respect to mass, leading to a difference of SFRs of 0.38 dex
(a factor of 2.4) over the 8.5 < logM∗< 11 range. Running average (in 0.2
dex bins) is shown as a thick line, and the ±1σ range as dashed lines.
8.5<log M∗<11range,with 1σ dispersionof0.33–0.41dex
aroundthe runningmean. To double check if the residuals are
in any way connected with B04 aperture corrections, we limit
the sample to farther, and therefore on average smaller galax-
ies (z > 0.12). We still find that the residuals correlate with
stellar mass.
Before rulingout step (a) as being responsiblefor the mass-
dependentresiduals, we performtwo additional checks. First,
imagine there is an offset in zero points between GALEX and
SDSS magnitudes (after all, they come from different surveys
and are not measured in identical apertures). E.g., if GALEX
fluxes were systematically overestimated, this could cause
“UV” SFRs to be overestimated as well, possibly in such a
way to preferentially boost blue, low-mass galaxies (as the
trend in Figure 8 would suggest). To test this, we perform full
SED-fitting runs in which we make FUV and NUV fainter
by 0.1, 0.2 and 0.4 mag. We first check the quality of these
new fits by comparing the distribution of χ2
nominal (no magnitude offset) run. Runs with 0.2 and 0.4
mag offsets produceevidentlyinferiorfits, already suggesting
that any potential offset cannot be that large. The run with 0.1
magoffset, however,appearsas goodas the nominalrun,with
an even slightly smaller average χ2
the “UV” SFRs produced with offset GALEX magnitudes to
“Hα” SFRs, we find that while slightly flattening the slope of
the residuals with respect to mass, the trend is not eliminated.
In fact, the scatter of the residuals around the mean is larger
when modified “UV” SFRs are used instead of nominal. We
also look at the quality of fits with GALEX magnitudes offset
in the opposite direction (making them brighter), but such fits
are evidently inferior.
The second test concerns SDSS magnitudes. Namely, B04
use CMODEL SDSS magnitudes to transform fiber SFRs
into total SFRs, while we use MODEL SDSS magnitudes
when performing SED fitting.18Therefore, we perform an-
other SED fitting using CMODEL magnitudes instead. First,
bestvalues with the
best. However, comparing
18SDSS MODEL magnitudes (defined as either the exponential or
de Vaucouleurs magnitude, depending on which profile better describes a
galaxy) are preferred magnitudes for SED fitting as they preserve relative
fluxes (colors) better. CMODEL magnitudes (defined as a composite of ex-
ponential and de Vaucouleurs magnitude) should provide a good measure of
a total galaxy light.
FIG. 9.— Comparison of B04 (“Hα”) to our (“UV”) estimates of V-band
dust opacity (attenuation), for the SF galaxy class. The formal error of both
estimates (average errors shown) is large, leading to a large scatter. However,
there is an agreement between the two in the general sense. B04 τV is to first
orderconstrained bythe Hα/Hβ ratio (Balmer decrement), while ourestimate
is predominately constrained by the UV slope. Both estimates were made in
accordance with the Charlot & Fall (2000) two-component dust attenuation
model.
FIG. 10.— The difference of B04 (“Hα”) and our “UV” estimates of V-
band dust opacity as a function of galaxy stellar mass(for SFgalaxies). There
is a trend with respect to mass, leading to a 0.5 difference over the 8.5 <
log M∗< 11 range. Running average (in 0.2 dex bins) is shown as a thick
line, and the ±1σ range as dashed lines.
we notice that the quality of fitting with CMODEL magni-
tudesis noticeablyinferior,stressing theirinadequacyfor pro-
ducing reliable color estimates. Also, the use of CMODEL
magnitudes does not remove the trend of SFR residuals, and
the scatter of the residuals becomes larger than in the nominal
run, especially at higher masses. With this test we exhaust the
possibilities that discrepancies arise in step (a) above.
We now move onto step (b), i.e., correcting the observed
flux for dust attenuation. Note that both B04 and this study
use Charlot & Fall (2000) prescription for dust attenuation,
the aim of which was to produce consistent treatment for Hα,
and UV continuum attenuation. Is the presence of the SFR
residuals an indication that this dust attenuation model does
not produce fully consistent answers? In our SED fitting we
keep track of estimates on τV—the dust opacity in rest-frame
V band. We can thus compare our τV values with those ob-
tained by B04 (which we denote τV(“UV”) and τV(“Hα”), re-
spectively). While we use the same model to constrain at-

Page 13

UV SFRs IN THE LOCAL UNIVERSE13
FIG. 11.— The difference of B04 (“Hα”) and the modified “UV” SFRs as a
function of galaxy stellar mass. We modify one of the two SFRs (in this case
the “UV” SFR) by the difference in B04 and our attenuation estimate. After
this modification, the trend of the residuals with respect to mass is eliminated,
and the overall difference is close to zero. Therefore, we find that emission
lines and the continuum essentially produce the same SFR estimates. Run-
ning average (in 0.2 dex bins) is shown as a thick line, and the ±1σ range as
dashed lines.
tenuations, B04 obtain them from the emission lines (to first
order from the Balmer decrement), while we obtain them
from the broad-band SED (to first order from the UV spec-
tral slope, β, or equivalently, the UV color). Also, B04 mea-
surements are restricted to the fraction of the galaxy inside
of the fiber, which can in principle produce some systematic
differences. Direct one-to-one comparison, shown in Figure
9, shows a rough agreement, albeit with a large scatter due
to relatively large errors of both estimates (average formal er-
rors shown as the error bar). Looking instead at the distribu-
tions of the two τV estimates (not shown), we find them to
be very similar on the whole, with the τV(“UV”) slightly off-
set towards the larger values with respect to τV(“Hα”). Since
we find SFR residuals to correlate with the mass, it is more
instructive to check whether such a trend exists in the dif-
ference of τV estimates from “Hα” and “UV”. In Figure 10
we plot τV residuals against the stellar mass. Despite the
large scatter, we find a systematic trend. The gradient is 0.47
(equivalent to 0.52 mag) over the 8.5 < log M∗< 11 range.
For lower masses, “UV” τVattenuation is higher than “Hα”,
which implies that the observed UV flux is corrected more
than τV(“Hα”) would require, resulting in “UV” SFRs be-
ing higher than SFR(“Hα”). The situation is reversed at high
masses. This trend matches the sense of the trend of SFR
residuals.
In order to establish whether the difference in τVestimates
is the dominant cause of the trend of SFR residuals, we will
try to modify one of the two SFR estimates (e.g., the “UV”
SFR) based on the difference in τV values. Note that this
is an approximate technique, since our SED SFRs are not
“corrected” for dust by the application of a single number,
but rather by a complex application of Charlot & Fall (2000)
prescription on many different populations that constitute the
model SED. Nevertheless, for the purposes of this test we
would assume that the attenuation in the FUV drives the cor-
rection of the “UV” SFR estimate. We therefore modify it
with:
SFR(“UV′′)mod= SFR(“UV′′)×100.4∆AFUV,
where ∆AFUVis the difference between the attenuation in the
(3)
rest-frame FUV as implied by τV(“Hα”) and that determined
fromthe SED fitting. We lookat the parametersfromthe SED
fitting to calibrate the relationship between the attenuation in
FUV and theV-band opacity, and find the following relation-
ship:
∆AFUV= µ(5.4−0.84τV)∆τV,
(4)
where ∆τV = τV(“Hα′′)−τV(“UV′′), while τV and µ come
from the SED fitting (i.e., they are “UV” measurements). The
parameter µ is the coefficient in the Charlot & Fall (2000)
model that determines the fraction of the total opacity af-
fecting the diffuse ISM. We modify SFR(“UV”) according to
Equations 3 and 4, and in Figure 11 re-plot the difference in
SFRs against the mass. The trend of the SFR residuals with
mass has disappeared, and the difference stays within ±10%
across the full mass range. The dispersion around the running
mean is 20% larger than with the non-modified SFR resid-
uals, which is to be expected due to the approximate nature
of the modification applied. Therefore, we conclude that the
“UV” continuum and the “Hα” in essence produce identical
answers for the star formation rate, the apparent difference
stemming from different estimates of the attenuation that the
two methods provide. With this in mind, we can also rule out
other possibilities for the cause of the original differences in
the SFRs, such as the reliability of converting the luminosity
into a SFR (point (c) above). Also, these results show that,
contrary to some notions, Bruzual & Charlot (2003) models
cannot be too much off in the UV.
Although we can now account for the SFR differences, we
would still like to understand the origin of attenuation differ-
ences that cause them. Related to this, we should try to de-
termine which attenuation estimate (emission-line or the UV
continuum) is more accurate. At this point we need to em-
phasize that our SED fitting has a non-flat prior distribution
of τV(described in §4.1), while B04 uses a flat prior, with at-
tenuation taking values in the 0.01< τV< 4 range. While the
non-flat prior used in our study is a reasonable assumption,
and the one that helps constrain the parameter space to physi-
cally realistic values, we would like to check whether the dif-
ferences in priors induce any systematic effects. To that effect
we first perform an additional SED fitting run, with the model
libraries created with a flat attenuation prior, taking values in
the 0 < τV< 3 range. At lower (τV< 0.7) values, we find the
distributionoftheresultingτVvaluestobesimilartotheorigi-
naldistribution,while at highervaluesit displaces the original
peak, located at τV≈ 1.3, to a broader distribution peaking at
τV≈ 2.0. The latter is the obvious consequence of assign-
ing equal probability to models with high attenuations. In any
case, we end up with a τVdistribution that differs more with
respect to the B04 τV(“Hα”) than did the original τV(“UV”).
Nevertheless, we proceed and check the relationship between
the new SFR residuals and the stellar mass. The trend ob-
servedin theoriginalrelationshipis still present, althoughit is
somewhat weaker at low masses. Anyhow, it appears that we
cannot force the two attenuation estimates to reach an exact
agreement by a simple modification of the prior distribution
used in SED fitting.
Next we will try to get a sense of which τVestimate is more
realistic. Since the direct comparison with an external in-
dependent measurement is not readily available, we will try
to evaluate the differences using internal relations. For star-
forming galaxies we expect attenuation to be correlated with
the stellar mass. In Figure 12 we show this relation for our

Page 14

14SALIM ET AL.
FIG. 12.— Attenuation-mass relationship for star-forming galaxies. The
two attenuation estimates, one from our SED fitting (“UV”, upper panel),
and the other from B04 (“Hα”, lower panel) are plotted against the stellar
mass. The B04 estimate (principally from Hα/Hβ ratio) produces a tighter
relationship with a greater span of attenuation values, indicating that it is
probably more accurate.
“UV” estimates (upperpanel)andB04 “Hα” estimates (lower
panel). It is apparent that the latter defines a more tight rela-
tionship. Also, τV(“Hα”) spans a larger range of values than
τV(“UV”) (1.4 vs. 0.8 over 8.5 < log M∗< 11). Finally„ the
dispersion around a running mean is some 20% smaller for
τV(“Hα”). This all indicates that B04 τVestimates are prob-
ably more accurate. While the two τV estimates agree well
on the whole (Figure 9), the “UV” estimate appears more
poorlyconstrained,whichmaycauseit to moreoftentakeval-
ues away from the extrema, suppressing the span in relation-
ship with the mass. As a result, the attenuation estimates for
low-mass galaxies will be overestimated, and for high-mass
galaxies underestimated. It is difficult to directly test this ex-
planation. Namely, if we restrict the analysis only to galaxies
with small formal errors in τV(“UV”) (or alternatively small
errors in FUV −NUV), we at the same time bias the sam-
ple to smaller values of τV(“UV”). Alternative explanation is
that the Charlot & Fall (2000) ∝ λ−0.7extinction curve devi-
ates from the true extinction curve in a way that would be
mass-dependent(orsome quantityrelated to mass, suchas the
metallicity). A possibility that the extinction curve is mass-
dependent is raised in Johnson et al. (2007). We would be
able to acquire further insight by comparing both attenua-
tion estimates to some external measure, such as the IR ex-
cess (IRX). Johnson et al. (2007) have recently obtained IRX
attenuations from Spitzer MIPS and GALEX observations of
SDSS galaxies, therefore such sample can be used to include
UV (i.e., SED) derived attenuations. Also, the SED-based at-
tenuation estimates can possibly be improved by adding the
near-IR data, such as the JHKsphotometryfrom 2MASS sur-
vey (Skrutskie et al. 2006), which would improve attenuation
estimates by placing stronger constraints on the stellar metal-
licity.
FIG. 13.— Relationship between the attenuation in the GALEX far-UV
and the UV slope. The FUV attenuation estimate comes from our SED fit-
ting, and the FUV −NUV rest-frame color serves as an indicator of the UV
SED slope. Data and the solid line linear fit is for galaxies classified as star
forming. The dashed line is a fit to blue (rest frame NUV −r < 4) galaxies.
The fits reproduce the trends in general, however, there is a significant scat-
ter. Meurer et al. (1999) relation for starburst galaxies is plotted as the dotted
line (M99). It is obviously not appropriate for our sample of mostly normal
star-forming galaxies.
6. OBTAINING THE UV SFRS WITHOUT SED MODELING
In this work we are using sophisticated SED modeling to
obtain UV-based star formation rates. In many applications
such a detailed approach is not practical, or not even possi-
ble. In such cases one would like to obtain a reasonably good
estimate of a dust-corrected star formation by applying some
simple transformations to the UV photometry. While such
methods have been used in many previous studies, here we
will use the results of the detailed SED analysis to calibrate
such simplified models, with a special emphasis on users of
UV data obtained with GALEX. Since for the foreseeable fu-
tureGALEX will remainthe onlyfacility capableofobserving
a large number of galaxies in the UV, it is not without justifi-
cation to treat its filters as defining some standard photomet-
ric bands in the UV domain. Researchers who study high-z
galaxies in rest-frame UV can calibrate their blueshifted fil-
ters against GALEX FUV and NUV response curves.19
Obtaininga UV-basedSFR consists of K-correctingthe UV
magnitudes, estimating the dust attenuation of the FUV flux,
andconvertingthe dust-correctedFUV luminosityintoaSFR.
We assume that the user will correcttheir data forthe Galactic
extinction, and then apply some standard K-correction pro-
cedure to obtain rest-frame FUV and NUV magnitudes. Of
course, obtaining reliable K corrections requires optical or
near-IR photometry. To calibrate the relations in this section,
we use KCORRECT V4_1_4 (§4.4).
We start by using the results of the full SED fitting to cali-
brate the well-known correlation between the attenuation and
the UV spectral slope (e.g., Calzetti et al. 1994). This is of-
ten refereed to as IRX-β relation, although strictly speaking
IRX indicates IR excess, which is correlated with UV attenu-
ation. IRX-β relationship for normal galaxies, such as those
in our sample, has been previously studied with GALEX data
bySeibert et al. (2005); Cortese et al. (2006); Gil de Paz et al.
(2007); Boissier et al. (2007); Panuzzo et al. (2007). In Fig-
ure 13 we show FUV attenuations (AFUV) obtained from our
19GALEX filter response curves can be obtained from
http://galexgi.gsfc.nasa.gov/tools/Resolution_Response.

Page 15

UV SFRs IN THE LOCAL UNIVERSE15
SED modeling, for galaxies classified as star-forming, plot-
ted against the rest-frame UV color, which is linearly corre-
lated to the UV spectral slope. We see that the majority of
objects lies along the ridge. This confirms that there exists
an “IRX-β” relationship for normal galaxies, not just for star-
burst galaxies as usually assumed. To quantify the relation-
ship for our sample, and thus allow AFUV to be determined
from GALEX observations of normal galaxies, we fit a linear
function to running medians (with value from each 0.05 mag
bin weighted equally). Using the medians is necessary in or-
der to avoid the fit to be affected by numerous outliers. Also,
we find that past some red UV color the attenuation does not
seem to rise, so we adopt a constant value. In any case, AFUV
exhibits a large scatter for such red colors.
AFUV=
?3.320(FUV −NUV)+0.22,
3.37,
0(FUV −NUV) < 0.95
0(FUV −NUV) ≥ 0.95,
(5)
where AFUV is in magnitudes, and superscript 0 designates
rest-frame colors. This relation is plotted in Figure 13 as a
solid line. In many applications, especially at higher red-
shift, spectroscopic classification may not be available. In
thosecases onecanselect star-forminggalaxiesbasedontheir
color. For blue-sequence galaxies (0(NUV −r) < 4, and with-
out applying any class selection) one should use a slightly
modified relation:
AFUV=
?2.990(FUV −NUV)+0.27,
2.96,
0(FUV −NUV) < 0.90
0(FUV −NUV) ≥ 0.90,
(6)
shown in Figure 13 as a dashed line. Note that these rela-
tionships are optimized for GALEX bandpasses and for nor-
mal star-forming galaxies. We overplot Meurer et al. (1999)
relation as a dotted line (we used Seibert et al. 2005 transfor-
mation to obtain FUV −NUV from the UV spectral slope).
Apparently, this relation, constructed from a sample of star-
burst galaxies, is well above the majority of galaxies in our
sample. Seibert et al. (2005) and Cortese et al. (2006), using
smaller samples (RC3 and cluster galaxies, respectively) also
find that Meurer et al. (1999) relation overpredicts FUV at-
tenuation. Seibert et al. (2005) derive a relation that exactly
bisects our SF fit and the Meurer et al. (1999) relation. Also
note that these relations pertain to the optically selected sam-
ple(sincepracticallyallSFgalaxiesaredetectedinUVbands,
there is no additional UV selection). In general, samples se-
lected at different wavelengths will have different IRX-β re-
lations, as demonstrated by Buat et al. (2005), using GALEX
NUV and IRAS 60µm selected samples.
For a source with known redshift, the dust-corrected, rest-
frame FUV is converted into a FUV luminosity. The final
step requires converting the luminosity into SFR. Following
B04 notation, we define:
η0
FUV= L0
FUV/SFR(“UV′′),
(7)
to be the inverse conversion factor between a dust-corrected
rest-frame FUV luminosity (in erg s−1Hz−1) and the SED-
derived SFR. The conversion factors comes from the stellar
population modeling that was used to perform the SED fit-
ting. Like the equivalent conversionfactor for Hα luminosity,
η0
our sample (the metallicity of which is on average 0.8Z⊙),
the median conversion factor is:
FUVis sensitive to metallicity, albeit more weakly so. For
FIG. 14.— Comparison of the SFRs “predicted” using a set of simple trans-
formations, to SFRs derived using the full SED fitting. The predicted SFR
is obtained using the transformations that we calibrate from the SED SFRs.
They are applied to K-corrected FUV and NUV magnitudes (see §6). The
comparison is good for the majority of galaxies (greyscale scatter plot). The
dashed lines show 16 and 84 percentile values of the conditional distribution
at each SFR(“UV”), from which we see that the predicted SFRs deviate from
SFR(“UV”) at low and high SFRs.
logη0
FUV= 28.165.
(8)
The above factor is given for Chabrier (2003) IMF, which
is used in this paper. The often-used conversion factor given
by Kennicutt (1998) for the Salpeter (1955) IMF is:
SFR(M⊙yr−1) = 1.4×10−28Lν(erg−1s−1Hz−1).
We find using Bruzual & Charlot (2003) models that the ap-
propriate transformation factor between UV-derived SFRs
that assume Chabrier and Salpeter IMFs is 1.58 (both with
0.1-100M⊙limits)20. Thus our “empirical” conversionfactor
(Equation 8) for Salpeter IMF becomes:
(9)
SFR(M⊙yr−1) = 1.08×10−28L0
FUV(erg−1s−1Hz−1).
(10)
This implies that the Kennicutt (1998) conversion factor
(Equation 9) is 30% higher. We verify that the effect of the
differencein bandpasses (Kennicutt1998factor was givenfor
the 1500–2800 Å range, while the GALEX FUV filter spans
1300–1800 Å) is completely negligible. The actual reasons
for the differenceare two-fold. First, even when obtainingthe
model conversion factor using Bruzual & Charlot (2003) and
the same assumptionsas givenby Kennicutt (1998) (i.e., solar
metallicity and a constant star formationhistory), we still find
Kennicutt (1998) conversionto be 15% higher (i.e., the corre-
sponding η is lower than one produced by Bruzual & Charlot
2003). Further, we have that our sample has average metallic-
ity somewhat lower than the solar. This accounts for another
5% difference. Finally, the remaining 10% difference stems
from the fact that our sample has a variety of star formation
histories, giving η that is on average higher than that for con-
stant SF history. To summarize, for optically selected sam-
ples similar to ours, we suggest using “empirical” conversion
given in Equation 8 or 10.
Finally,we combinethe aboveingredientsand“predict”the
SFR from K-corrected FUV and NUV magnitudes. In Figure
20Note that the conversion between the IMFs is not generally the same for
Hα-derived SFR, UV SFR, or stellar mass.