[daip] vtess comments

[Eric Greisen]

Sara Beck writes:

 > I am using VTESS for the first time, on an extra-galactic source, and ran
 > into some puzzling behaviour at the short wavelengths (U and K band).  at
 > these wavelengths the source (seen in conventional CLEAN maps) is small
 > and weak, like a few mJy.  the vtess program would act sensibly only if
 > it was given an estimate of the total flux (FLUX<0 in the inputs) and if
 > the field mapped was limited to 6 arcseconds. otherwise it would hang up
 > somewhere, get into a state after a few iterations where the step and
 > gradient were 0 and stay there, and the final flux would be huge (1.45 Jy
 > instead of 3 mJy) and the map total garbage, broad bright streaks all
 > over.
 > 
 > also, the final fluxes for VTESS maps are way off, like a factor 2-3
 > higher, than the CLEAN maps at U and K. (these are the maps that behaved,
 > not the weird ones).
 > 
 > i get the impression from the explain vtes material in aips that the
 > quantitative results are not to be trusted unless the S/N is really high,
 > which it isn't here. so can i just ignore the high fluxes in the vtess
 > results? or do i figure that they weren't really good fits even though
 > they converged?

     Convergence is difficult to judge in a space with a great many
local minima.  We have not trusted VTESS even in good S/N cases until
Tim taught me the trick of FLUX < 0 as an advisory flux.  Also he
recommens FLUX = -flux/10 where flux is the real flux.

     VTESS minimizes over the entire area and so can be very wrong
over small areas while being quite content with the results it gets.
It does not represent point-like sources well at all.

 > 
 > Note that at L, C, an X on the same source Vtess worked very sensibly and
 > the results are very consistent with the regular maps except they look a
 > little less blotchy/mottled, which is what we wanted from Vtess.

Have you tried the multi-resolution option in IMAGR?  It also is
designed to get rid of the point-source Clean mottling and is somewhat
closer to a linear process than VTESS.

You are right to distruct the fluxes and even the images from VTESS.
The total flux should be estimated in part from the visibility data.
How big a flux do you see on a UVPLT?  (Of course, that is the real
flux plus 1 sigma or more)

Eric Greisen