[daip] Re: SDI clean

[Eric Greisen]

Michael Nord writes:

 > I am trying to understand SDI clean and how it is executed in AIPS.
 > imageprm(4) allows me to turn it on and off using the TV, but if I am not
 > using the TV, it will bounce back and forth between BCG and SDI clean when

    You should always use the TV when experimenting with new
algorithms.  Unless you like publishing erroneous "science" of
course.

 > some criterion I am not sure about is met.  Can you explain this
 > criterion?  It is in the explain file, but I wasn't able to understand it.

     I tried hard there so I do not klnow if I can do better here.  In
the Clean boxes of the residual images we compute a histogram of data
values.  At the beginning of the Clean, the histogram will have some
rather high intensity values occupied by only a few pixels in the
Clean boxes.  If one is willing to load 20000 pixels to the "AP" for
the next round of Cleaning, one looks down from the peak of the
histogram (highest intensity) and counts pixels until, going one more
level in the histogram, will get more than 20000 pixels.  This then
sets the cutoff for loading into the AP - a familiar message tells you
loading all pixels above x mJy as a result of this.

If you have a source with a lot of extended emission at about the same
level, then, as one cleans, the peak intensity in any Clean box goes
down and eventually gets near the intensity of the extended emission.
In other words, the histogram becomes rather flat with lots of pixels
close to the pixel with the higest intensity.  This is when the SDI
method probably surpasses the Clark Clean.  IMAGRPRM(4) controls two
things:
1. > 0 => SDI is possible.  Then the ratio of the peak in the
  histogram to the level to load into the AP is reported at each major
  iteration and the task will slip into SDI when that ratio gets
  right.  If the edge effects of SDI make the next cycle be above that
  level it will go back into Clark and so on.  This is actually a
  rather good way to deal with the edge problems.
2. When max intensity in histogram divided by the level in the
  histogram that the Clark Clean would use as a cutoff becomes
      < (1 + GAIN * IMAGRPRM(4))
  the SDI celan begins.  In other words, for GAIN=0.1 and
  IMAGRPRM(4)=2, SDI begins when the ratio is less than 1.2.  Values
  of iMAGRPRM up to about 10 or so are probably reasonable.

 > 
 > Can SDI clean just be switched on (ie, don't allow regular clean EVER
 > without having to set that in the TV)?

     SDI attempts to slice the top off a plateau.  To do this it needs
to know the contribution to the slicing that the adjacent pixels will
make.  If a pixel is surrounded with pixels which will all be sliced
at the same time, (and the beam is > 1 pixel in size), then each of
the 8 neighbors will take a bit away from that pixel.  But pixels near
an edge will not have 8 neighbors, but a smaller number.  If a single
SDI gain were used, then these edge pixels would have - say 4
neighbors subtraction only and then end up rather higher than the
pixels in the center of the plateau.  This makes small-scale
structures (that are not physical) around the edge and make a mess of
the next major cycle.

     IMAGR has a tactic to reduce this affect.  To make the tactic
really work however would probably make SDI go 4 or more times slower
than it already does.  So it is a compromise that helps but does not
cure the problem.

     Clean the points with Clark.

 > 
 > I am imaging the galactic center at 330 MHz A array, so I have a LARGE
 > field with lots of interesting diffuse stuff and lots of point sources.

    "field" ? singular?  You should be using many facets to cover this
area.  Run SETFC with IMSIZE and CELLSIZE zero for advice about those
parameters.  Then pick values and rerun SETFC to give you the facets
that are required to avoid major problems in 3D.

 > >From experimentation, I know that the SDI clean does a LOT better job
 > cleaning the bright, diffuse emission around SAG A.  If I just use SDI
 > clean, what is the trade off?  I get better cleaning on diffuse sources,
 > but worse on point sources?
 > 
 > If SDI seems to work so well on diffuse sources, why even use the old
 > style clean ever?  The original SDI paper seems to indicate that SDI clean
 > is pretty much better than the old clean in every category.
 > 

    The authors of any new algorithm want you to think it surpasses
sliced bread as the best invention of the century.  It does not.  It
is useful for diffuse objects but develops problems at the edges of
those.

Eric Greisen