Äîêóìåíò âçÿò èç êýøà ïîèñêîâîé ìàøèíû. Àäðåñ îðèãèíàëüíîãî äîêóìåíòà : http://www.stecf.org/goods/spectroscopy/CDFS_Mastercat/szokoly04.ps Äàòà èçìåíåíèÿ: Thu Sep 16 11:25:45 2004 Äàòà èíäåêñèðîâàíèÿ: Tue Oct 2 01:04:54 2012 Êîäèðîâêà:

arXiv:astro­ph/0312324
v2
7
Jul
2004
The Chandra Deep Field South:
Optical Spectroscopy I. 1
G. P. Szokoly 1;2 , J. Bergeron 3 , G. Hasinger 1;2 , I. Lehmann 1 . L. Kewley 4;5 , V. Mainieri 6 , M.
Nonino 7 , P. Rosati 6 , R. Giacconi 8 , R. Gilli 9 , R. Gilmozzi 6 , C. Norman 4 , M. Romaniello 6 E.
Schreier 8;10 , P. Tozzi 7 , J. X. Wang 4 , W. Zheng 4 and A. Zirm 4
szgyula@mpe.mpg.de
ABSTRACT
We present the results of our spectroscopic follow-up program of the X-ray
sources detected in the 942 ks exposure of the Chandra Deep Field South (CDFS).
288 possible counterparts were observed at the VLT with the FORS1/FORS2
spectrographs for 251 of the 349 Chandra sources (including three additional
faint X-ray sources). Spectra and R-band images are shown for all the observed
sources and R K colours are given for most of them. Spectroscopic redshifts were
obtained for 168 X-ray sources, of which 137 have both reliable optical identi-
cation and redshift estimate (including 16 external identications). The R< 24
observed sample comprises 161 X-ray objects (181 optical counterparts) and 126
of them have unambiguous spectroscopic identication. There are two spikes in
the redshift distribution, predominantly populated by type-2 AGN but also type-
1 AGN and X-ray normal galaxies: that at z = 0:734 is fairly narrow (in redshift
space) and comprises two clusters/groups of galaxies centered on extended X-ray
1 Max-Planck-Institut fur extraterrestrische Physik, Giessenbachstrae, Garching, D-85748 Germany
2 Astrophysikalisches Institute Potsdam, An der Sternwarte 16, Potsdam, D-14482, Germany
3 Institut d'Astrophysique de Paris, 98bis, bd Arago, 75014 Paris, France
4 The Johns Hopkins University, Department of Physics and Astronomy, Baltimore, MD 21218, USA
5 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA
6 European Southern Observatory, Karl-Schwarzschild-Strasse 2, Garching, D-85748, Germany
7 Osservatorio Astronomico, Via G. Tiepolo 11, 34131 Trieste, Italy
8 Associated Universities, Inc. 1400 16th Stret, NW, Suite 730, Washington, DC 20036, USA
9 Osservatorio Astrosico di Arcetri, Largo E. Fermi 5, I-50125 Firenze, Italy
10 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA

{ 2 {
sources, the second one at z = 0:674 is broader and should trace a sheet-like struc-
ture. The type-1 and type-2 populations are clearly separated in X-ray/optical
diagnostics involving parameters sensitive to absorption/reddening: X-ray hard-
ness ratio (HR), optical/near-IR colour, soft X-ray ux and optical brightness.
Nevertheless, these two populations cover similar ranges of hard X-ray luminosity
and absolute K magnitude, thus trace similar levels of gravitational accretion.
Consequently, we introduce a new classication based solely on X-ray properties,
HR and X-ray luminosity, consistent with the unied AGN model. This X-
ray classication uncovers a large fraction of optically obscured, X-ray luminous
AGNs missed by the classical optical classication. We nd a similar number
of X-ray type-1 and type-2 QSOs (LX (0.5-10 keV)> 10 44 erg s 1 ) at z > 2 (13
sources with unambiguous spectroscopic identication); most X-ray type-1 QSOs
are bright, R. 24, whereas most X-ray type-2 QSOs have R& 24 which may ex-
plain the dierence with the CDFN results as few spectroscopic redshifts were
obtained for R> 24 CDFN X-ray counterparts. There are X-ray type-1 QSOs
down to z  0:5, but a strong decrease at z < 2 in the fraction of luminous X-ray
type-2 QSOs may indicate a cosmic evolution of the X-ray luminosity function
of the type-2 population. An X-ray spectral analysis is required to conrm this
possible evolution. The red colour of most X-ray type-2 AGN could be due to
dust associated with the X-ray absorbing material and/or a substantial contri-
bution of the host galaxy light. The latter can also be important for some redder
X-ray type-1 AGN. There is a large population of EROs (R K> 5) as X-ray
counterparts and their fraction strongly increases with decreasing optical ux,
up to 25% for the R 24 sample. They cover the whole range of X-ray hardness
ratios, comprise objects of various classes (in particular a high fraction of z & 1
X-ray absorbed AGNs, but also elliptical and starburst galaxies) and more than
half of them should be fairly bright X-ray sources (LX (0.5-10 keV)> 10 42 erg s 1 ).
Photometric redshifts will be necessary to derive the properties and evolution of
the X-ray selected EROs.
Subject headings: surveys | galaxies: active | cosmology: observations |
quasars: general, evolution | X-rays: galaxies: clusters | techniques: spectro-
scopic
1 Based on observations collected at the European Southern Observatory, Chile (ESO N o 66.A-0270(A)
and 67.A-0418(A)).

{ 3 {
1. INTRODUCTION
Deep X-ray surveys indicate that the cosmic X-ray background (XRB) is largely due
to accretion onto supermassive black holes, integrated over cosmic time. In the soft (0.5{2
keV) band more than 90% of the XRB ux has been resolved using 1.4 Msec observations
with ROSAT (Hasinger et al. 1998) and 1-2 Msec Chandra observations (Brandt et al.
2001a; Rosati et al. 2002; Brandt et al. 2002) and 100 ksec observations with XMM-
Newton (Hasinger et al. 2001). In the harder (2-10 keV) band a similar fraction of the
background has been resolved with the above Chandra and XMM-Newton surveys, reaching
source densities of about 4000 deg 2 . Surveys in the very hard (5-10 keV) band have been
pioneered using BeppoSAX, which resolved about 30% of the XRB (Fiore et al. 1999).
XMM-Newton and Chandra have now also resolved the majority (60-70%) of the very hard
X-ray background.
Optical follow-up programs with 8-10m telescopes have been completed for the ROSAT
deep surveys and nd predominantly Active Galactic Nuclei (AGN) as counterparts of the
faint X-ray source population (Schmidt et al. 1998; Zamorani et al. 1999; Lehmann et
al. 2001), mainly X-ray and optically unobscured AGN (type-1 Seyferts and QSOs) and a
smaller fraction of obscured AGN (type-2 Seyferts). The X-ray observations have so far been
about consistent with population synthesis models based on unied AGN schemes (Comastri
et al. 1995; Gilli et al. 2001), which explain the hard spectrum of the X-ray background by
a mixture of X-ray absorbed and unabsorbed AGN, folded with the corresponding luminosity
function and its cosmological evolution. According to these models, most AGN spectra are
heavily absorbed and about 80% of the light produced by accretion will be absorbed by
gas and dust which may reside in nuclear starburst regions that feed the AGN (Fabian et
al. 1998). However, these models are far from unique and contain a number of often
overlooked assumptions, so their predictive power remains limited until complete samples
of spectroscopically classied hard X-ray sources are available. In particular, they require
a substantial contribution of high-luminosity absorbed X-ray sources (type-2 QSOs), which
so far have only scarcely been detected. The cosmic history of obscuration and its potential
dependence on intrinsic source luminosity remain completely unknown. Gilli et al. (2001)
e.g. assumed a strong evolution of the absorbed/obscured fraction (ratio of type-2/type-1
AGN) from 4:1 in the local universe to much larger fractions (10:1) at high redshifts (see also
Fabian et al. 1998). The gas-to-dust ratio in high-redshift, high-luminosity AGN could be
completely dierent from the usually assumed Galactic value due to sputtering of the dust
particles in the strong radiation eld (Granato et al. 1997). There could thus be objects
which are heavily absorbed at X-rays and unobscured at optical wavelengths.
After having understood the basic contributions to the X-ray background, the general

{ 4 {
interest is now focussing on understanding the physical nature of these sources, the cosmolog-
ical evolution of their properties, and their role in models of galaxy evolution. We know that
basically every galaxy with a spheroidal component in the local universe has a supermassive
black hole in its centre (Gebhardt et al. 2000). The luminosity function of X-ray selected
AGN shows strong cosmological density evolution at redshifts up to 2, which goes hand in
hand with the cosmic star formation history (Miyaji et al. 2000). At the redshift peak
of optically selected QSOs, around z=2.5, the AGN space density is several hundred times
higher than locally, which is in line with the assumption that most galaxies have been active
in the past and that the feeding of their black holes is re ected in the X-ray background.
While the comoving space density of optically and radio-selected QSOs has been shown to
decline signicantly beyond a redshift of 2.5 (Schmidt et al. 1997; Fan et al. 2001; Shaver
et al. 1996), the statistical quality of X-ray selected high-redshift AGN samples still needs
to be improved (Miyaji et al. 2000). The new Chandra and XMM-Newton surveys are now
providing strong additional constraints.
Optical identications of the deepest Chandra and XMM-Newton elds are still in
progress, however, a mixture of obscured and unobscured AGN with an increasing fraction of
obscuration at lower ux levels seems to be the dominant population in these samples (Fiore
et al. 2000; Barger et al. 2001a; Tozzi et al. 2001; Rosati et al. 2002; Stern et al. 2002).
First examples of the long-sought class of high-redshift, radio-quiet, high-luminosity, heavily
obscured active galactic nuclei (type-2 QSO) have also been detected in deep Chandra elds
(Norman et al. 2002; Stern et al. 2002) and in the XMM-Newton deep survey in the
Lockman Hole eld (Hasinger 2002).
In this paper we report on our optical identication work in the Chandra Deep Field
South, which, thanks to the eôciency of the VLT, has progressed to the faintest magnitudes
among the deepest X-ray surveys.
2. THE CHANDRA DEEP FIELD SOUTH (CDFS)
The Chandra X-ray Observatory has performed deep X-ray surveys in a number of elds
with ever increasing exposure times (Mushotzky et al. 2000; Hornschemeier et al. 2000;
Giacconi et al. 2001; Tozzi et al. 2001; Brandt et al. 2001a) and has completed a 1 Msec
exposure in the Chandra Deep Field South, CDFS (Giacconi et al. 2002; Rosati et al. 2002)
and a 2 Msec exposure in the Hubble Deep Field North, HDF-N (Brandt et al. 2002). The
Megasecond dataset of the CDFS is the result of the coaddition of 11 individual Chandra
ACIS-I exposures with aimpoints only a few arcsec from each other. The nominal aim point
of the CDFS is  = 3 : 32 : 28:0, ô = 27 : 48 : 30 (J2000). This eld was selected in a

{ 5 {
patch of the southern sky characterized by a low galactic neutral hydrogen column density,
NH = 8  10 19 cm 2 , and a lack of bright stars (Rosati et al. 2002).
3. OPTICAL IDENTIFICATIONS IN THE CDFS
Our primary optical imaging was obtained using the FORS1 camera on the ANTU
(UT-1 at VLT) telescope. The R band mosaics cover 360 arcmin 2 to depths between 26
and 26.7 (Vega magnitudes). This data does not cover the full CDFS area and must be
supplemented with other observations (see Figure 14). The ESO Imaging Survey (EIS) has
covered this eld to moderate depths (5  limiting AB magnitudes of 26.0, 25.7, 26.4, 25.4,
25.5 and 24.7 in U 0 , U, B, V, R and I, respectively) in several bands (Arnouts et al. 2001;
Vandame et al. 2001). The EIS data has been obtained using the Wide Field Imager (WFI)
on the ESO-MPG 2.2 meter telescope at La Silla. The positioning of the X-ray sources is
better than 0.5 00 (Giacconi et al. 2002) and we readily identify likely optical counterparts in
85% of the cases.
Figure 1 shows the classical correlation between the R-band magnitude and the soft X-
ray ux of the CDFS sources. The objects are marked according to their classication (see
below). By comparison with the deepest ROSAT survey in the Lockman Hole (Lehmann
et al. 2001), the Chandra data extend the previous ROSAT range by a factor of about 40
in X-ray ux and to substantially fainter optical magnitudes. While the bulk of the type-1
AGN population still follows the general correlation along a constant fX =f opt line, the type-2
AGNs cluster at higher X-ray-to-optical ux ratios. There is also a population of normal
galaxies emerging at low uxes (thus discovered in the Chandra and XMM era).
To be consistent with the already published deep ROSAT catalogs (Lehmann et al.
2001), we used a modied version of the X-ray to optical ux ratios:
log 10 (f x =f o )  log 10 (f 0:5 2keV =fR )  log 10 (f 0:5 2keV ) + 0:4R + 5:71; (1)
where the ux is measured in erg cm 2 s 1 units in the 0.5-2 keV band and R is in Vega
magnitudes. The slight change in the normalization (Maccacaro et al. 1988) is motivated
by the signicantly narrower X-ray energy band used (the original denition was based on
the Einstein medium sensitivity survey band, 0.3-3.5 keV), which introduces a factor of 1.77
decrease in the ux for objects with a spectral energy index of 1 (classical type-1 AGN)
and the use of the R-band instead of V (here we assumed a V R color of 0.22, typical value
for galaxies).
To use this new X-ray to optical ux ratio denition for source classication, we also had
to convert the canonical ranges (Stocke et al. 1991) to our new system. The new ranges

{ 6 {
for dierent classes of objects are shown in Table 2. To calculate the new ranges of the
X-ray to optical ux ratios, we assumed typical X-ray spectra for each class and calculated
the shift in the X-ray ux due to the narrower energy band: a power law with a photon
index of =1-2.7 for AGN, a power law with a photon index of =1-2 for BL Lac objects, a
Raymond-Smith model with kT =2-7 keV, abundances of 0.1-0.6 and redshifts of z =0-1. For
stars and supernova remnants we used Raymond-Smith models with energies of kT =0.5-2
keV, for X-ray binaries powerlaw models with photon index =1-2. For galaxies, we adopted
a somewhat ad hoc shift of 0.1-0.3 in the logarithm of the ux due to the dierent energy
bands. This choice was motivated by examining dierent models for galaxies (warm and
hot plasma mixture, powerlaw like emission from X-ray binaries, typical supernova remnant
spectra, etc.).
For the shift in the optical ux (using R-band instead of the canonical V-band) we
assumed typical values for each class.
The resulting ranges of the X-ray-to-optical ux ratios are shown in Table 2. As can
be seen from the table, the new X-ray-to-optical ux ratio is not signicantly dierent
from the canonical one. The typical ranges are a bit wider, but this just a consequence of
converting the ranges instead of directly determining it from large surveys. With our new
normalization, we can use the original ranges (Stocke et al. 1991) to make an educated
guess on the galactic/extragalactic nature of objects.
4. TARGET SELECTION
Target selection was primarily based on our deep VLT/FORS imaging data (Giacconi
et al. 2002), reaching a depth of R  26:5. In regions not covered by this VLT/FORS deep
imaging, we used somewhat shallower VLT/FORS imaging in the R-band obtained as part
of the survey.
Possible optical counterparts of X-ray sources were selected based on the estimated
astrometry error of the X-ray object (for a relatively bright point source at zero o-axis
angle the astrometry rms error is  0: 00 5). We used the automatically generated optical
catalog, however, every object was visually inspected for deblending problems and artefacts.
The surface density of our X-ray objects is very well suited to MOS spectroscopy with
FORS/VLT. We could ll a large part of the masks with program objects and it was quite rare
that we had to choose between multiple optical counterpart candidates within the geometrical
constraints of the instrument. As a consequence, our target selection is nearly unbiased.
The only selection eect that should be considered was related to objects with multiple

{ 7 {
counterpart candidates. In these cases we usually selected the object in the appropriate
magnitude range for the particular mask, but in general we tried to revisit these objects {
unless the rst one turned out to be clearly the counterpart.
We also took advantage of the extremely high accuracy of the robotic masks: in some
cases, we recongured some of the slits between read-outs, without changing the telescope
pointing to observe many (brighter) optical counterparts. This way, the integration time on
bright objects could be shortened and we could use the remaining time on a dierent object,
while maintaining longer integration times for the faint ones.
During our last two runs (in November and December 2001) we were also using the
prefabricated masks (MXU mode { only available for FORS2), as opposed to movable robotic
slitlets (MOS mode). For our survey, the only important dierence between the two modes
is more freedom in the placement of slits in MXU mode. This improved our observing
eôciency in the later phase of the survey, where we concentrated on fainter objects (with a
higher surface density).
4.1. The Reliability of the Target Selection
The reliability of X-ray follow-up surveys using optical (or near infrared) spectroscopy
hinges on matching the X-ray source to the right optical object. This is primarily done
through astrometry. Just how reliable are these identications? Using deep galaxy number
counts (Metcalfe et al. 2001), we expect roughly 0.02 galaxies in every square arcsec area
that are brighter than R  26. Considering our best 3 astrometry error (1: 00 5), we expect
 0:15 eld galaxies to fall within our error circle { even in the best, zero o axis angle
case. In other words we expect one false candidate for every seventh X-ray object at R
< 26. Considering the roughly 250 X-ray sources we observed, we expect that for at least
35 of them, there will be a completely unrelated faint galaxy, even in an error circle of 1: 00 5.
Fortunately, the X-ray counterpart candidates typically have much brighter magnitudes (see
Figure 2). At these brighter magnitudes the probability of eld galaxy contamination is
much lower, so we should only worry about contamination for very faint (R=25-27) objects.
As our astrometric accuracy heavily depends on the signal-to-noise ratio of the object
(i.e. objects with low photon counts are centered with lower accuracy) and the o-axis angle
of the object (there is a signicant degradation of the PSF of increasing o-axis angles), the
total area covered by the sum of the error circles is quite large, around 3900 arcsec 2 . In
Figure 2, we show the magnitude distribution of our selected primary optical counterparts
and the expected magnitude distribution of random eld galaxies over this area, based on

{ 8 {
galaxy number counts (Metcalfe et al. 2001; Jones et al. 1991). Contamination by random
eld galaxies becomes a serious problem beyond R  24 and they start to dominate beyond
R  26, the practical limit of our imaging survey.
Therefore, extra caution is required in making sure that the right optical object is
identied as the counterpart. This is not always trivial as the optical spectra do not always
show clear signatures of active nuclei (AGNs). In some cases we had to observe every object in
the error circle. Fortunately this turns out to be feasible. At R  24 and fainter, deblending
is not a serious challenge (using both automated and visual tests). At brighter magnitudes,
where deblending would be near impossible (e.g. detecting a R  25 X-ray object in the
halo of a R  19 galaxy), the probability of eld galaxy contamination is negligible. Stellar
contamination is negligible at our high galactic latitude.
It is also important to point out that these estimates of eld galaxy contamination are
for the probability of nding an unrelated object in our X-ray error circle. We can also ask
a technical question: what is the probability of nding a eld object on a slit? Taking a
20 00  2 00 area (the typical slit length in FORS-1 is around 20 00 ), we expect to nd a R < 23
galaxy in 5% of the slits and we expect statistically a eld galaxy with magnitude R < 25
in every second slit. This means that one has to be extremely careful in the data reduction
and do a very careful book keeping in the process.
5. OBSERVATIONS AND DATA REDUCTION
Data were obtained during 11 nights in 2000 and 2001. A summary of the observations
is presented in Table 4. All observations were using the `150I' grism (150I+17 in FORS-1
and 150I+27 in FORS-2). These grisms provide a pixel scale (dispersed) of 280  A/mm, or
roughly 5.5  A/pixel. The nominal resolution of the conguration is R = =
=230, which
corresponds to roughly 20  A at 5600  A. The pixel scale of these instruments is 0.2 00 /pixel, so
there is no signicant degradation of the resolution due to the nite slit width.
In the initial phase of our survey, we exclusively used low resolution multiobject spec-
troscopy with varying integration time. This strategy maximizes the number of observed
objects and provides a (nearly) full spectral coverage for every exposure. This is clearly
a trade o, as we then get a signicantly lower S/N spectrum for the individual objects,
compared to higher resolution long-slit spectroscopy based on photometric redshifts, but the
latter technique was deemed to be prohibitevely expensive in observing time in the initial
phase of our project.
As our goal was to observe as many objects as possible, we used non standard order

{ 9 {
separation lters (either no lter, or the GG-375 lter, which cuts out light bluer than
3750  A). It was thus possible to cover a very wide spectral range in a single exposure (in
the standard conguration the order separation lter that cuts the light blueward of 5900  A,
thus the whole spectral range can only be covered in two exposures).
5.1. Data Reduction
Data were reduced by our own semi-automatic pipeline built on top of IRAF. In general
we followed standard procedures, but had to deviate slightly in several cases to accomodate
particularities of the FORS instrument and do a very rigorous book-keeping. In the following
sections, we enumerate these changes.
5.2. Bias, Overscan and Trim Correction
The FORS CCD's have in principle 4 read-out modes: high and low gain and one and
four amplier modes. To avoid serious complications, we only used the high gain/one am-
plier read-out mode for our spectroscopic observations. This decision resulted in a slightly
larger overhead, but this was deemed negligible considering our long integration times, com-
pared to the challenges posed by reducing a 4 amplier read-out mode spectroscopic obser-
vations, where we would have to calibrate the gain of each amplier very accurately (so we
do not introduce artical features in the spectra).
A suôcient number of full frame bias exposures were taken during each run (typically
around 20 per run). These were individually overscan corrected and trimmed. The resulting
(bias) frames were averaged with suspect pixels (too high or too low values) ltered out to
generate the master bias frame. In each case we veried that the bias frame does not change
signicantly from night to night within a run.
A slight complication was posed by our spectrophotometric standard observations.
These frames were also using one amplier/high gain, but (to save some time) only 500
rows were read out (centered on the standard star). Since ESO does not provide an un-
der/overscan region for windowed frames, we took a suôcient number (typically 10) of bias
frames in this conguration. Naturally (lacking under/overscan region) these frames were
not overscan corrected, nor trimmed. Instead, they were averaged to create a master bias
frame, which did include the articially introduced bias level. We checked the individual
frames and conrmed that the variation of this artical bias level is negligible for these very
high S/N frames.

{ 10 {
After creating the full and windowed bias frames, all object and calibration ( at and
arc) frames were overscan corrected and trimmed (except the windowed frames) and zero
subtracted.
At this point we applied a shift in the dispersion direction, based on the slit position,
to bring (very crudely; within 10 pixels or 50  A) the observations on a similar wavelength
scale. We also inserted gaps in the spatial direction between the neighbouring slits to reduce
the risk of contamination between slits. These two steps are purely practical, but make
bookkeeping signicantly easier.
5.3. Flatelding
In this processing step, we had to tackle three main issues:
The rst one is an inherent complication in the FORS instruments. Due to the mechan-
ical construction of the robotic slit masks and the location of the at-eld lamps, at-eld
exposures show higher ux levels in a few rows at the upper or lower edge of the slit. To
correct for this eect, there are two sets of at-eld lamps in the instrument. We took a
suôcient number of at-eld exposures using both sets of lamps. We generated merged ats
independently for each lamp set and generated the nal at-eld frame by taking the smaller
pixel value in the two frames. As the re ections from the two lamp sets do not overlap, this
feature can be fully removed.
The second issue is a consequence of our unusual observing strategy. In some cases
(due to geometric constraints imposed by the robotic slit masks) we could not target very
faint objects with a particular slit, but we had several bright candidates available. In these
cases, to maximize eôciency, we recongured these slits between readouts so that all bright
candidates were observed, while faint objects targeted with other slits were observed with a
longer integration time. Due to the extremely high mechanical stability of the FORS instru-
ments, this strategy is very safe. As the sensitivity variation between pixels is potentially
color dependent, we decided to generate at-eld frames for each mask. This may not be the
optimal strategy since for the slits that are in the same position in two masks, we could use
more exposures, thus to create a more accurate at-eld. This alternative strategy would be
too complex and the resulting data quality improvement is very marginal, consequently we
decided against it.
The last major issue is due to the extremely wide spectral coverage used. As our in-
tention was to correct only for the pixel-to-pixel sensitivity variations, we had to generate a
normalization image (a combination of the at-eld lamp spectrum and the overall quantum

{ 11 {
eôciency of the system as a function of wavelength and spatial position). For high resolution
(and smaller wavelength coverage) observations, this is often achieved by collapsing in the
spatial direction and tting a function in the dispersion direction. Unfortunately, this tech-
nique proved to be impractical for us. The main problem was that we were unable to nd
an ansatz function that could reproduce the very sharp cutos at both ends (due to either
the order separation lter or the natural cut-o of the CCD detector) without introducing
artical structure on intermediate scale. An additional complication was that the internal
at-eld lamps did not illuminate the slits homogenously { there is a slight gradient in the
spatial direction. Therefore, after a slight smoothing of the at-eld exposures, we created
the normalization image by a linear or (for very long slits) a second order polynomial t in
the spatial direction. Each at-eld exposure was divided by this normalization frame, thus
creating a `true' at-eld frame, which only contains pixel-to-pixel sensitivity variations. In
regions, where the signal was too low, the at-eld was articially set to one (to avoid the
introduction of too high photon noise).
After these steps, the individual, normalized at-eld frames were merged, eliminating
the eect of light re ection on the slit edges. Both science and wavelength calibration frames
for a given mask were divided by the resulting master at-eld frame.
5.4. Sky Subtraction
The sky background was estimated in each column by a linear t (for longer slits) or
just calculating the average (shorter slits) in each column of each slit, rejecting too high
pixels (i.e. the targeted object) and subtracting the result. It is important to note that
we did not correct for the very slight curvature of the dispersed spectra on the CCD in this
step. With the FORS instruments, this strategy works quite well (as opposed to LRIS on the
Keck telescope). Signicant sky residuals are only present around the very bright, narrow
sky lines { where sky subtration is doomed anyway due to pixel saturation.
This procedure works only for our typical faint objects. Extremely bright objects can
illuminate the whole slit, thus making correct sky subtraction impossible. Fortunately, in
those (very few) cases identication was still possible due to the extremely high object signal.
5.5. Fringe Removal
In some cases (especially in MXU masks), we could take advantage of our dithering
strategy to reduce further the eect of fringing and the sky residuals. As neither the fringe

{ 12 {
pattern nor the sky residuals are signicantly aected by the small (spatial) osets of the
telescope, we could, in some cases (with suôcient number of exposures in a given mask)
exclude (most of) the object signal and create a fringe/sky residual template for each slit.
Subtracting this from the frames resulted in an improved signal-to-noise ratio for the object
spectra. Depending on the seeing conditions and the dithering osets used, not all object
signal was perfectly removed, thus the extracted spectra signicantly underestimated the
real spectra. As our primary goal was object identication, not spectrophotometry, this was
an acceptable trade-o.
5.6. Coadding the Frames
After sky subtraction, all the slits were visually inspected to verify that the object is
indeed in the 'good' region of the slit. This step was necessary since the applied small spatial
osets between the science exposures can result in objects falling too close to the slit edge
(MOS blade corners are round, thus the slit is not usable there) or falling completely outside
the slit.
After this visual screening, the spatial oset between dierent exposures of the same
object was caculated based on the world coordinate system (WCS) information stored in the
frame headers. The individual exposures were coadded (including the rejection of suspicious
pixels or cosmic ray hits) after applying these