[daip] [!7296]: AIPS - OMFIT and strange chi2 values

[Michael Bietenholz]

Michael Bietenholz updated #7296
--------------------------------

OMFIT and strange chi2 values
-----------------------------

           Ticket ID: 7296
                 URL: https://help.nrao.edu/staff/index.php?/Tickets/Ticket/View/7296
           Full Name: Michael Bietenholz
               Email: mbieten at yorku.ca
             Creator: User
          Department: AIPS Data Reduction
       Staff (Owner): -- Unassigned --
                Type: Issue
              Status: Open
            Priority: Default
                 SLA: NRAO E2E
      Template Group: Default
             Created: 22 September 2015 08:35 PM
             Updated: 22 September 2015 08:35 PM
                 Due: 24 September 2015 08:35 PM (2d 0h 0m)
      Resolution Due: 30 September 2015 08:35 PM (8d 0h 0m)



I think something is wrong with the Chi2 values that OMFIT reports.  The "Ketan Chi2" values may be better.

Using the test data set attached, which should just be a 10 mJy point source at 0,0mas with some noise (UVMOD),
along with the also attached simpldot.omfit INFILE and  

DEFAULT OMFIT
[getn attached data]
INFILES = [attached simpledot.omfit file]
PRTLEV 4; NITER 300


I get the following (31DEC15; binary; MNJ today)

postfit rms=0.07116784; DOF=18287; Ketan chi2=92.621122; Real chi2=36574.000

So far so good, the actual fit values seem reasonable (flux density 9.36 mJy).  However, the fact that "real" Chi2 is *exactly*  2*DOF is suspicious. 

However, if I now force the model to have a fixed 1.0 Jy, ie a very poor fit to the data, I get

postfit rms= 0.70413390; DOF=18288 ; Ketan chi2=9067.2735; Real chi2=36576.000

So now the postfit rms is up by ~10x; consistent with a much worse fit;  Ketan's chi2 is up by about 100x, as might be expected from the 10x higher postfit rms, however, the "real" chi2 almost unchanged, and still exactly 2*DOF.

Moreover, if I set NOISE(1)=1, I would expect the same result as for NOISE(1)=0, since that is supposed to be the
default, however, if I do that (flux density is free again), I get 

postfit rms=0.07116784; DOF=18287; Ketan chi2=92.621122; Real chi2=185.24224

So Ketan's chi2 doesn't change (as expected), but the "real" chi2 changes by a lot.  For other non-zero values
 of NOISE(1), the "real" chi2 does indeed scale as NOISE(1)**2, as expected.  It seems to be using NOISE(1)= postfit rms if NOISE(1)=0.   I think factor of two comes in because it fits real and imaginary separately.

What is even stranger though, is that for a SN1993J data set, it does *not* seem to default to NOISE(1)=post-fit rms.


------------------------------------------------------
Staff CP:  https://help.nrao.edu/staff
-------------- next part --------------
A non-text attachment was scrubbed...
Name: testx.fits
Type: image/fits
Size: 987840 bytes
Desc: not available
URL: <http://listmgr.nrao.edu/pipermail/daip/attachments/20150922/6e04c155/attachment-0001.fits>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: simpledot.omfit
Type: application/octet-stream
Size: 284 bytes
Desc: not available
URL: <http://listmgr.nrao.edu/pipermail/daip/attachments/20150922/6e04c155/attachment-0001.obj>