[alma-config] CLEAN and difference results for B arrays

[Mark Holdaway]

On Thu, 7 Dec 2000, Stephane Guilloteau wrote:

> In cases like Mars which are dominated by some obvious extended structure,
> there is
> a good general solution which has not been implemented in any package to my
> knowledge.
> I used it a couple of times on my own data and that gives fairly good
> results.
> 
> The trick is to remove from the UV data the Fourier component of an original
> model (in the
> case of Mars, a uniform disk would do it, but the model can be completely
> arbitrary). You then
> perform a normal deconvolution of the difference UV data. The last step is
> to add the original
> model to the Clean components found in this deconvolution, convolve it with
> a Clean beam,
> and add the deconvolution residual.

AIPS++ uses this as its standard scheme: all deconvolution algorithms
acting on the (u,v) data have an option of reading in a starting
model image; if it exists, it is FT'ed and subtracted from the
visibilities, and then the residual visibilities are imaged, and
the model image updated.

> Since the Fourier Transform is a linear operation, that works whatever the
> model you take. Best results
> are obviously obtained when the model leaves as little differences as
> possible.
> 
> What I say here can also be applied to MEM deconvolution, of course,
> although in that case you would
> probably prefer to remove a model which is some underestimate of the true
> emission, to take advantage
> of the positivity constraint, implying that you would still have to use a
> default image.
>

There is maximum emptiness, which doesn't have a positivity constraint.
Or, in AIPS++, we have an incremental entropy: the entropy and its
gradients are calculated from the cumulative model (initial plus
whatever else you have added to it in the various major cycles of
deconvolution), so the incremental image can be negative in several
places without adversely affecting the algorithms performance.


	-Mark