Äîêóìåíò âçÿò èç êýøà ïîèñêîâîé ìàøèíû. Àäðåñ îðèãèíàëüíîãî äîêóìåíòà : http://www.stecf.org/conferences/adass/adassVII/reprints/decuyperjp.ps.gz Äàòà èçìåíåíèÿ: Mon Jun 12 18:51:45 2006 Äàòà èíäåêñèðîâàíèÿ: Tue Oct 2 04:13:15 2012 Êîäèðîâêà:

Astronomical Data Analysis Software and Systems VII
ASP Conference Series, Vol. 145, 1998
R. Albrecht, R. N. Hook and H. A. Bushouse, e
Ö Copyright 1998 Astronomical Society of the Pacific. All rights reserved.
ds.
On the Need for Input Data Control in Pipeline
Reductions
J.­P. De Cuyper and H. Hensberge
Royal Observatory of Belgium, Ringlaan 3, B­1180 Brussel, Belgium
Abstract.
An analysis of series of flat­field exposures obtained in echelle mode
adds support to the view that the performance of an instrument should
be monitored closely and described quantitatively. This simplifies enor­
mously the data reduction procedure and allows the user to interpret
high­quality data correctly.
1. Introduction
Nowadays it is not unusual for the knowledge about an instrument acquired by
the design and construction team to be only partially transmitted to those de­
veloping the data reduction software or operating the instrument; the feedback
to the observer is still smaller. This may result in confusion about basic instru­
mental parameters, the precision of the set­up, the amount and distribution of
calibration frames which are required and finally the assumptions on which the
data reduction procedure should rely. The data set then does not deliver the
precision aimed at during observing, or observing time was spent ine#ciently.
The concept of pipeline reductions and on­line controlled operation of the in­
strument and of the observational procedure o#er unprecedented possibilities for
delivering data with known error characteristics, provided that the instrumental
design, the set­up, the calibration and the data reduction procedure are tuned
to each other at a consistent level of accuracy. Our experience with echelle spec­
troscopy indicates that, as a visiting astronomer, it is impossible to collect the
information needed to obtain this goal. Moreover, it is not at all evident to what
extent forthcoming pipelines will include realistic error estimates (in addition to
random noise estimates).
2. An Example: Series of flat­fields
The results shown here refer to two adjacent series of eight consecutively taken
flat­fields (tungsten lamp exposures). The telescope was in the zenith and track­
ing was o#. The analysis method applied is insensitive to global changes in
intensity and to small changes in location, width or shape of the cross­order
profile (these changes are below 0.5% in intensity and within 0.02 pix for the
other quantities in both series). For each pair of frames, a parameter d indicat­
ing the lack of similarity in the shape of the blaze profile was computed (Figure
312

Input Data Control in Pipeline Reductions 313
Figure 1. Histogram of the dissimilarity parameter d for all combi­
nations of 2 out of the 16 flat­fields. d 2 scales with the variance due to
long­range (# 100 pixels) deformations in ratios of extracted flat­field
orders, except for an o#set due to the contribution of random noise.
The value for pure random noise is indicated
1). It turns out that the excess of d over its value calculated from the case of
random noise can be modelled as a ``distance'' between frames (Figure 2).
The lack of repeatability in subsequent frames is due to instabilities that
grow and disappear, rather than to a slow, continuous change with time. Con­
secutively taken frames di#er more than the ones separated by an intermediate
frame, and the global pattern of changes repeats in the two independent series
(Figure 2). Relative deformations in the shape of the blaze profile appear over
several spectral orders in the same set of rows of the detector (see e.g., Figure 3
near x # 300).
It is not our intention to show that things can sometimes go very wrong,
but merely that the accuracy is generally limited by systematic errors. The
example shown above does not refer to an exceptional malfunctioning, but to
a common situation. Notice that it is not uncommon to detect stronger e#ects
when comparing exposures taken with longer time delays and/or in di#erent
telescope positions. The detectability of such systematic e#ects sets a natural
limit on the precision of order merging (since the intensity ratio of the wavelength
overlap region of consecutive orders is a#ected), on the level up to which faint,
shallow spectral features can be trusted and on the precision of the continuum
placement.
3. Data Reduction Strategy
Experience with echelle spectroscopy confirms that the previous example is not
an isolated case of bias. Systematic errors are detectable in almost all types
of frames and they influence directly the applicability and the accuracy of the

314 De Cuyper and Hensberge
Figure 2. The dissimilarity of flat­fields represented as an Euclidian
distance between frames situated in a plane (par1, par2). The num­
bering indicates the order of the exposures. Notice the usually large
distance between consecutive frames, and the similarity between the
series 1­8 and 9­16
data reduction algorithms. Depending on the ratio of systematic to random
noise, algorithms that are based on the dominance of random noise (such as
the detection of radiation events, the application of optimal extraction and any
reduction step involving non­robust least­squares fitting) may need refinements.
Rather than commenting on specific sources of bias, we like to outline a
procedure, interconnecting the di#erent phases from instrument development
and testing to data reduction software development, that in our opinion would
permit the user to evaluate properly the quality of the final data with regard to
the particular aspects of interest to his/her specific purpose:
. Characterize detector and instrument using the most stable mode of oper­
ation. Specify the maximum precision that will be supported in the set­up,
calibration and data reduction flow.
. Specify the target set­up accuracies (consistent with the data reduction
requirements) and check during operation whether the calibration and sci­
ence frames fall within these specifications. Specify how frequently exten­
sive stability checks are needed.
. Use robust techniques. Exploit knowledge about the instrument and envi­
ronment, at least in a di#erential if not in an absolute way. A data frame

Input Data Control in Pipeline Reductions 315
Figure 3. Relative deformation of the blaze profile expressed as the
intensity ratio of the sum of frames #11 and #13 to the sum of #10 and
#12. This ratio is shown along 12 consecutive orders, shifted vertically
for clarity (the dashed lines indicate the relative intensity level 1.0 for
each order). Instrument: CASPEC, detector: ESO CCD #32
with its specific calibration frames should not be treated independently
from the global observing run. Several factors vary systematically with
quantities that are known or measurable and do not require ill­determined
free parameters.
. Quantify the extent to which assumptions made during the data reduction
are invalid and archive them with the final data.
. Use the experience gained during the reduction process to improve at con­
venient times (e.g., when important instrumental interventions are un­
avoidable) the observing and data reduction strategy, and, ultimately, the
development of new instruments.
Acknowledgments. This research was carried out in the framework of
the project `IUAP P4/05' financed by the Belgian Federal Scientific Services
(DWTC/SSTC). We thank W. Verschueren (RUCA, University of Antwerp),
who obtained the calibration spectra discussed in this paper, and H. Van Diest
for help with the data handling. This work is based on observations obtained at
the European Southern Observatory (ESO), La Silla, Chile.