From jwf@ipac.caltech.eduTue Mar 25 09:38:34 1997
Date: Thu, 13 Mar 1997 17:33:54 -0800
From: jwf@ipac.caltech.edu
To: 2mass@ipac.caltech.edu
Cc: chas@ipac.caltech.edu, sstrom@ipac.caltech.edu,
    stiening@ipac.caltech.edu
Subject: WG Mtg #117 Minutes

           IPAC 2MASS Working Group Meeting #117 Minutes
                             3/11/97

Attendees: R. Beck, T. Chester, R. Cutri, T. Evans, J. Fowler,
           L. Fullmer, T. Jarrett, G. Kopan, B. Light,
           S. Wheelock, J. White

AGENDA

1.) Single-Frame Extraction Position Statistics
2.) Telescope/Camera Status
3.) Effects Due to Band-Dependent Plate Scale
4.) Source Merging Splinter Session Report


DISCUSSION


1.)  Single-Frame Extraction Position Statistics

    G. Kopan described a study he has been working on regarding position
statistics for the single-frame point-source extractions. He has found some
significant dependences of position errors on seeing, especially in the
differences between Read1 sources and Read2-Read1 sources. The latter
differences appear to be largely due to "freezing" of the seeing over the
short integration time for Read1, with coherent seeing displacements over the
focal plane. This causes a significant correlated shift between the positions
of the two read types. Investigations including several scans with poor-to-bad
seeing have indicated that the "pfrac" parameter is probably usable as a
position-uncertainty model parameter, possibly in a form as simple as making
the position sigma a linear function of the point-spread function's FWHM, with
the latter being a TBD function of pfrac. Care is needed in defining the
position uncertainty, however, since simply assigning a seeing-dependent sigma
to Read1 sources would not be appropriate both to dispersions in Read1
positions about mean position (i.e., averaged over six apparitions in scan
coordinates) and dispersions in position discrepancies between the two read
types (because of the highly correlated Read1 position shifts due to freezing
the seeing). Thus the use to which the uncertainties will be put must be
clearly agreed upon before an exact model can be established. Some discussion
points that arose were as follows.

     A.) Since dense scans involve clipping FREXAS input to PFPREP, there can
     be more Read1 sources input than Read2-Read1 sources, hence Read1 could
     dominate the frame-offset solutions; this should not be allowed.

     B.) If inverse-variance weighting is used in the frame-offset solution,
     then giving the Read1 sources position uncertainties including the
     component due to freezing the seeing might suffice as a way to prevent
     the Read1 sources from corrupting the solution.

     C.) Saturated Read2-Read1 sources should not be allowed to contribute
     any position information in PFPREP or POSFRM.


2.) Telescope/Camera Status

    R. Cutri reported that first light was achieved last week at the northern
observatory. Optical alignment is on hold while polar alignment problems are
being worked out. Pointing reproducibility over short time intervals appears
to be good (about 5 arcsec at some high confidence level), based on tests
involving slewing away from a given position and returning to it. The dates
for receipt of three-channel data at IPAC and start of survey are uncertain.


3.) Effects Due to Band-Dependent Plate Scale

    Possible software impacts due to band-dependent plate scale were
discussed. The intersection of all three arrays mapped onto the sky is
expected to be such that excess area of some arrays may be as large as a pixel
or more. Some sources detected in or near this region that should be
multi-band will be missing some bands, and the different edge tolerances in
the detection and extraction phases may yield anomalies. In addition, plate
scale differences will make it very difficult to achieve optimal scan
dithering in all three bands.
    The fracturing of multi-band sources near the cross-scan edges will
probably have to be fixed during catalog preparation time, since it appears
inadvisable to limit detection activity there, and in any case, multiple
records of sources in the scan overlap regions will require sorting out, as in
the Faint Source Survey. Fractured sources will have been scanned better on
adjacent scans, except at coverage gaps. The point was made that one large gap
was preferable to an equal area distributed over many small gaps, because the
latter would involve more edge artifacts.
    It was suggested that a flag indicating the proximity of a source to the
edge of the scan might be desirable. The tentative decision was that the
U-Scan X coordinate would serve this purpose; if its absolute value is more
than about 250, then the source came within an arcsec or so of the edge. The
exact value will depend on more detailed information regarding plate scale and
array alignment.


4.) Source Merging Splinter Session Report

    After the previous meeting (two weeks ago), a splinter meeting concerned
with source merging problems was held. This splinter meeting continued the
following week and replaced the regular meeting. The issues involved and
decisions made in that splinter session will be reported herein.

     Issues

          A.) Read1 sources occasionally appear replicated or fractured into
          multiple sources.

          B.) When confronted with too many Read1 sources around Read2-Read1
          sources, BANDMERGE often joins the two read types (e.g.,
          Read2-Read1 J detection with a Read1 H detection, although a
          Read2-Read1 H detection is also available).

          C.) Sometimes BANDMERGE joins a detection with good photometry
          with a detection that has no photometric information (i.e., both
          the PSF magnitude and the aperture magnitude are 99.999).

    Issue A was found to result from the PFPREP search box being too small.
PFPREP computes an initial estimate of the frame offsets using only J
Read2-Read1 sources; for this a search window half-width of 3 arcsec in
preliminary U-Scan coordinates is used. After the improved U-Scan coordinates
are available, the other bands and read types are merged; for this the
half-width has been 1.5 arcsec. A typical fractured Read1 source was traced to
this point and found to have components separated by as much as 2.55 arcsec;
the problem occurred because the "seed" for merging (the first detection in
frame order) was the farthest from the mean position, leaving other pieces
outside the window to form their own merged source later. Examination revealed
that the full-width separation of 2.55 arcsec between Read1 pieces was not
unreasonable and corresponded to about a 5-sigma event, which can be expected
to occur because of the large number of sources that must be processed.
Furthermore, this is just over one pixel in size, and the chances of producing
two distinct Read1 detections this close is negligible, so opening up the
window to 3 arcsec appears reasonable. This was the decision for resolving
this issue.
    Issue B involves the fact that BANDMERGE uses a position chi-square test
to judge the acceptability of different-band detections for merging. This is
probably about as good a test as possible when the candidates have position
errors drawn from approximately the same Gaussian random population. Since the
Read1 and Read2-Read1 position errors apparently constitute different
populations, the type of test may need to be changed. A goodness-of-match
parameter such as the one used in the IRAS point-source processing was
suggested; this involves the cross-correlation of the two position density
functions evaluated at the observed separation. This has the advantage that,
all else equal, it is biased against measurements with larger uncertainties,
since the corresponding density functions are broader, hence lower, hence more
inclined to fall below an acceptance threshold. It has the disadvantage that
thresholds cannot be set purely on the basis of the acceptable fraction of
real cases to sacrifice in the attempt to avoid false matches, unless the
actual distributions of position errors are known in advance. For IRAS, Monte
Carlo simulations were used to set thresholds. But for IRAS, the density
functions were highly non-Gaussian, which is one reason the particular method
was used instead of something like chi-square tests.
    This dilemma was alleviated somewhat by the realization that the Read1
position uncertainties have been artificially high during all the tests so
far. This is a result of the temporary substitute POSFRM, which was inserted
solely to pass semi-realistic data downstream so that the rest of the pipeline
could be tested. This version of POSFRM (-0.9) did not attempt to reduce the
Read1 uncertainties as it averaged their single-frame positions; when this is
fixed, the Read1 uncertainties will not be so different from the Read2-Read1
uncertainties (this may change when the impact of Agenda item 1 above is fully
absorbed), and the Read1 position error model has been preliminary anyway. For
now, the decisions made by BANDMERGE with reduced-uncertainty Read1 input will
be studied, and no fundamental change to the decision algorithm will be made
unless it becomes clear that it is still necessary. Improved Read1 merging in
PFPREP should also alleviate this problem.
    For the remaining cases of merging the two read types, and for the sake of
Issue C, some augmentation of the BANDMERGE goodness-of-fit parameter in cases
of diagnosed confusion were considered. The method adopted is as follows. Once
at least two candidates pass the position test for merging with a seed, the
chi-square parameter will be expanded by adding what is most easily described
as a dot product of two vectors. The components of the first type of vector
are the MAPCOR flag elements describing merge/purge/persistence and a yes/no
component describing whether the detection has a non-99.999 magnitude. Each
detection being considered for merging has such a vector, and the difference
vector will be computed. Each component of the difference vector will be made
positive, and then a dot product with the second type of vector will be taken.
The second type of vector has components expressing the weight of the
corresponding components of the first type of vector. The dot product will be
added to the position chi-square, and the smallest result will determine the
candidate to be merged with the seed.
    This much will be tested. If it works acceptably, other more complicated
measures will not be taken; for the record, these involved the following
suggestions.

     A.) The search window in PFPREP could be a function of the read-type
     combination, or a function of the position uncertainties encountered, or
     a function of the seeing, or a function of the detection magnitudes, or
     some combination of these (the position uncertainties themselves could
     be functions of seeing in addition to signal/noise ratio; see Agenda
     item 1 above).

     B.) Another pass could be added to PFPREP in which mean merge-group
     positions serve as seeds (this would avoid the probem of selecting the
     worst outlier of a group as a seed; it would be rather expensive and
     risky to implement, however).

     C.) A new program could be inserted between POSFRM and PFPOST to massage
     the merge-group file before Read1 sources are taken from it to pass on
     to MAPCOR; this could fix fractured groups.

     D.) The order in which PFPREP processes band/read-type could be changed;
     currently, after the preliminary J Read2 frame-offset computation, the
     processing is simply J/1, J/2, H/1, H/2, K/1, K/2. It woud probably be
     better to use all the Read2-Read1 detections as seeds before going to
     Read1 seeds.