[daip] Re: Multi-Resolution Clean in AIPS

[Eric Greisen]

Greg Ball writes:

 > I would like to clarify exactly what the IMAGR MR clean is doing and how 
 > it is supposed to work.  My understanding is the following:
 > 
 > For each field, images are made using a series of beams, beam (i) being 
 > the "point-source" beam convolved with a Gaussian of FWHM W(i).  For W=0 
 > the image is cleaned as usual using a delta function and the point-source 
 > beam; for W>0 the clean components are Gaussians, FWHM W(i).  Each clean 
 > component is subtracted from ALL the images.

         Actually from the UV data but it has that effect.

 > 
 > Cleaning an extended source with delta functions produces well-known 
 > "bed-of-nails" type artifacts, but cleaning an unresolved source with a 
 > Gaussian produces a "bullseye" artifact.  So at each step we have to clean 
 > the image which corresponds to the real size scale of the brightest 
 > emission.  The effect of the Gaussian convolved beams is that a feature of 
 > some finite size will be increasingly bright, in units of J/beam, for 
 > larger beams while the beam is smaller than the source, but will then stay 
 > about the same brightness in Jy/beam as the beam gets still larger.  So we 
 > want to clean the image where the peak brightness reaches roughly its 
 > limiting value.

    Actually it gets it wrong all the time - no extended source is
round - so we depend on Clean being iterative and working to undo its
mistakes.  What we try for is to keep the early mistakes small since
they are hard to undo in a reasonable time.

 > 
 > In practice, what we do is weight the peak brightness for each image by 
 > some power of the beam area, controlled by IMAGRPRM(11).  This ensures 
 > that the peak-brightness-versus-W(i) curve turns over, and hopefully, that 
 > the turn over point is the size scale we are after.  We must have 0 < 
 > IMAGRPRM(11) < 1 so that a) the weights decrease as W increases, and b) 
 > the initial increasing part of the curve, which should be about linear at 
 > first, is not "squashed".  (I won't go into interpretation of the other 
 > IMAGRPARM values since you recommend leaving them zero.)
 > 
 > So much for my understanding of the algorithm.

   That is pretty mush right.

 > 
 > > The most important thing is to be experimental with it - watch it on
 > > the TV with DOTV true throughout.
 > 
 > I'm trying this, but it isn't clear to me what sort of intervention is 
 > appropriate.  Should try to put clean boxes around different 
 > features at different resolutions as cleaning proceeds, to direct the 
 > process (rather then just ensuring my original box contains all the 
 > emission)?  Forcing clean to pick a particular field to clean next?
 > Changing IMAGRPRM(11) with TELL to get different behavior?

     I use BOXFILE = OBOXFILE so that when I restart it remembers the
windows I have set interactively.  I do not use TELL but I do
occasionally use the TV to force an image or to remake all images.
Otherwise, it makes choices on images that don't actually represent
the current state of the uv data until after it chooses them - which
occasionally forces it to try again.

The extended source parts mostly cut down on the bed of nails errors
and somewhat fill in the negative bowls that surround objects when
there are too few short spacings.

Eric Greisen


 > 
 > Thanks for any suggestions.
 > 
 > Greg Ball
 > 
 > 
 > --
 > Gregory H. Ball             Harvard-Smithsonian Center for Astrophysics
 > Graduate Student            60 Garden St MS #10, Cambridge MA 02138, USA
 > Astronomy Department        Email: gball at cfa.harvard.edu
 > Harvard University          Office: +1 617 496-4946