[daip] [!4593]: aips - SPECR

[Michael Bietenholz]

Michael Bietenholz updated #4593
--------------------------------

SPECR 
------

           Ticket ID: 4593
                 URL: https://help.nrao.edu/staff/index.php?/Tickets/Ticket/View/4593
           Full Name: Michael Bietenholz
               Email: mbieten at yorku.ca
             Creator: User
          Department: AIPS Data Processing
       Staff (Owner): -- Unassigned --
                Type: Issue
              Status: Open
            Priority: Default
                 SLA: NRAO E2E
      Template Group: Default
             Created: 26 February 2014 08:16 PM
             Updated: 26 February 2014 08:16 PM
                 Due: 28 February 2014 08:17 PM (2d 0h 0m)
      Resolution Due: 06 March 2014 08:17 PM (8d 0h 0m)



SPECR behaves a little strangely when interpolating spectra.  Admittedly this is in the regime that one is warned against in the HELP file of
"Some care must be taken by the user to avoid totally over-sampling the spectra".

Attached is a screen cap showing before/after plot showing visibility amplitude vs frequency.  The original data was made by UVSUB and consists of a 1-Jy point source model, so all visibility amplitudes are just identically 1.0, and are shown in plot by the larger yellow crosses (in this fashion of plotting, all the visibilities for each channel lie exactly on top of one another, so there is only point per channel on the plot). The smaller green crosses are after running SPECR to re-grid three original channels (12 MHz each) to 30 new channels (1.2 MHz each).    One might expect that the interpolated spectrum would just be constant at 1.0 Jy, as is the input spectrum, but it is not really so.  In particular, at the high end the output SPECR seems to be extrapolating down to 0 which is likely not what users might want.  Even between the existing channels, the interpolated values go to just over 10% *higher* than the input values.

The average value over all channels is not exactly preserved either - after SPECR in this instance, the average flux density over all the channels was 0.954, so about 4.5% lower than the input value.

I imagine SPECR is convolving with some kernel (Gaussian?) in frequency space and then resampling.    However, the kernel should probably be chosen (or some normalization done) so that a constant input produces a constant output between input channels. 

Also,  when it is extrapolating it would probably be more sensible for it to just copy the value from the last channel.  A solution to this would of course be to prevent SPECR from extrapolating at all, but  I would request that SPECR *retain* its current ability to extrapolate in frequency.
I have often used to to produce multi-channel data sets from single-channel ones for Monte-Carlo simulations.  After blowing up the channels in this way, one can adding noise to the individual channels, and winds up with  compact version having many noise realizations of the same data set, that can, for example be IMAGRed in one go.  The same could also be achieved using loops and many files, but the above route seems more convenient.

------------------------------------------------------
Staff CP:  https://help.nrao.edu/staff
-------------- next part --------------
A non-text attachment was scrubbed...
Name: SPECR_uvplt.png
Type: image/png
Size: 11957 bytes
Desc: not available
URL: <http://listmgr.nrao.edu/pipermail/daip/attachments/20140226/0dd49f18/attachment.png>