[daip] aips imagr

[Eric Greisen]

Cao Hongmin wrote:
> Hi Eric,
> 
> "difmap" was used to do the "clean" test with no clean windows and no 
> self-calibrations (AIPS version i used is 31DEC11).
> I had thought the image looks like that is because there might be many 
> side-lobes having been cleaned, which causes severe image aliasing effects.
> 
> I am ambiguous with origin of the image noise (rms) (?)
> It seems the numerical error dominates the contributions of the rms, but 
> how about the measurement errors (of amplitude, phase, etc.) that we can 
> not corrected.
> 
> So the only gain about the data average is that we get a smaller dataset 
> than before, which data processing software can readily handle. (?)
> 
> The (heavily) averaged data are loaded into difmap for imaging, and the 
> "uvweight 0,-2" is used. Suppose the longer baselines are down weighted 
> properly and use the 2000 Km (just European-base antennas are included) 
> to estimate the FOV, then the resulting FOV will be about 1.5". we did 
> detect a weak source (S/N > 7sigma) ~ 1" away from the phase centre of 
> correlation. I feel that it's kind of expedient.
> As noted in the help document, "boxfile" is used when there are at least 
> 64 fields, if so, it seems that it still could not alleviate the large 
> workload.

I do not know the details of difmap's imaging code.  I would not expect 
AIPS' IMAGR to blow up as you described.  But we do double precision 
computation
inside the "AP" which does most of the Cleaning and Fourier transforms and
corrections for the FT of the gridding function.  This last helps remove 
the worst of the aliasing.

True imaging rms comes from the uncertainties in the visibility data. 
There can be contributions to the apparent rms from receiver noise, but 
also from calibration error and sidelobes of the dirty beam.  Numerical 
error should not be important although with very large numbers of 
samples (10^9 or so), we had issues in the corners of images when they 
were computed with single precision.
The correction in the corners for the FT of the gridding function 
exceeds 10^6..

Averaging is really intended to make data sets more manageable and 
difmap was not designed for the really large data sets we now have.

BOXFILE is REQUIRED for NFIELD > 64.  It may be used with NFIELD=1  very p
profitably.  Having found this source, you might wish to use 1 facet 
centered on its found location with data that are less averaged and 
tapered.  You would have to use AIPS imaging I suspect.

31DEC11 is probably recent enough (it is called OLD here), I am 
developing 31DEC13 at present.

Eric Greisen