[daip] Re: IMAGR multiresolution question

[Eric Greisen]

Samuel Conner writes:
 > Dear Dr. Greisen,
 > 
 >    Has consideration ever been given to a Clean component type "Rectangle"?
 > 
 >    I have noticed in some simple experiments on the cleaning of
slightly resolved sources with the standard point clean component
model that I get a predictable pattern of errors in the input Data
divided by Clean model. This pattern is identical to the pattern I get
when I create a fake UV data set with a uniform rectangular model and
then clean it with point source CC components --- even if the
rectangular model is the size of a single image cell. The identity of
these patterns suggests a common cause in the difference between the
flat run of amplitude with UV of a point component and the "sinc"
function which a uniform rectangle would produce
 > 
 >    Would it be legitimate to think of the underlying sky brightness
distribution as being composed of two type of components:  points and
uniform filled cells --- delta functions and "box functions"?  If that
were valid, it might be possible to get away with just 2 resolutions.
This would relieve the would-be multi-resolver of having to figure out
which resolution sizes to use and what weighting to assign the
different resolution fields --- if one wanted multiresolution, one
would simply turn on the 2nd component type, which would be sized to
match the image cell size. I imagine that the relative weight in the
decision of which field to clean would be dependent on the SNR, but
that might be a simpler decision than the current method.
 > 
 >    I looked into hacking my local installation to add this feature,
but I don't understand the OOP architecture and will assuredly break
anything I touch.

------------------

The fundamental failure in cleaning extended objects somes from the
bias against equal components at adjacent cells.  The bias arises from
the fact that the beam is several cells across and therefore that the
distinction between adjacent cells would require data from longer
baselines than those actually available.  I am not suprised that a
careful examination will show the effects of this.  We see it often in
comparing the DDT/Y2K Clean images from different computers which use
identical data and software but different compilers.  The difference
image is a "bed-of-nails" reflecting the source structure.

I am reluctant to use source models that have sharp edges in the image
plane because they then go on forever in the UV plane.  The
multi-resolution algorithm is intended to locate scales that differ on
the order of a factor of 10 in area.  I usually find what the 0 width
scale gives as a beam and then set the other beams about a factor of 3
in diameter from it.  Then I look at the peak brightness of the
uncleaned images to determine how to set the bias against the larger
beams so as to "level the playing field" with IMAGRPRM(11).  A decent
guess for FGAUS and the multi-recolution schem works well.  Some folks
have also had good results with the SDI algorithm which directly
fights the bias against adjacent cells.

Eric Greisen