[evlatests] 12.5 MHz Spectral Line -- Very puzzling results

[Rick Perley]

    Tests done last April seemed to demonstrate that the extra noise 
seen in 12.5 MHz spectral line modes (with the single exception of the 
cross-hand correlations in mode PA) was removed when 'self-test' was 
turned off. 

    Subsequently, Bryan arranged for this to automatically occur 
whenever 12.5 MHz BW spectral line observations were made.  A brief test 
made earlier this week however failed to demonstrate the expected 
absence of the additional noise -- the parallel hand correlations in 
mode 'PA' were about 50% noisier than the cross-hand correlations. 

    Ken thus recovered the files I used in the April tests, and re-ran 
this experiment last night.  In summary, the files observed 3C147 and a 
nearby blank field at L-band in continuum, and in modes PA, PB and '4' 
for bandwidths 25, 12.5, and 6.25 MHz.   The file was run twice, once 
with self-test on, once with self-test off. 
    Analysis is by examination of the (calibrated) noise histograms of 
the real and imaginary parts of the visibility when on the noise 
source..   The AIPS task UVHGM is now set up to fit a gaussian to the 
distribution -- this proves to be very useful. 

    I reduced that data this afternoon, and determined that the basic 
characteristics that I have reported before are still in effect, namely:
       * all correlations when in spectral line mode with BW  = 12.5 MHz 
-- with the exception of the cross-hand correlations in mode PA -- are 
too noisy by a factor of two to three. 
       *  all observations at 50, 25 and 6.25 MHz gave histograms 
exactly at the expected widths. 
       *  turning off self-test decreased the noise -- *** but did not 
return the histogram widths to the correct values***.  More intriguing 
(or perhaps disturbing) is that the histograms of the affected data are 
distinctly non-gaussian -- there are large extended 'tails' on both 
sides (positive and negative) of the distribution.  These are evident in 
the data taken with self-test on, and are very much more evident in 
those histograms taken with self-test off.   By contrast, the histograms 
of the unaffected data (non-12.5 MHz, and cross-hands in mode PA) are 
beautifully gaussian. 

    Some numbers to show the scale of the effect:

    In modes PA, PB, and 4, with BW = 12.5 MHz, the derived channel 
widths are 781 kHz.  We used 1.66 seconds averaging.  With self-test on, 
we find the following histogram widths (dispersion):
      Mode          Parallel Hand            Cross Hand
----------------------------------------------------------------
         PA                  0.98 Jy                    0.43
         PB                  0.92                         0.88
           4                   1.00                           NA
----------------------------------------------------------------
    Note that 0.43 Jy is the corrrect answer, as determined by 
observations in the continuum, and in the 25 and 6.25 MHz BW observations. 

    When self-test is turned off,  we get the following
    Mode            Parallel Hand              Cross Hand
------------------------------------------------------------------
       PA                    0.62                          0.42
       PB                    0.49                          0.48
         4                     0.60                           NA
----------------------------------------------------------------

    Note that turning self-test off always reduces the noise, but in no 
case does the noise go to the correct value.  Also, *all* of the noise 
histograms are distinctly non-gaussian, and all in the same way, with 
long, tails of excess for both positive and negative values.

    These results seem to be contrary to what was claimed in the April 
tests -- taken with the same file on the same source.  I recovered these 
data, and recalibrated the 12.5 MHz PA mode data.  What I claimed in 
April is again seen -- turning off self-test restored the histogram 
widths to near the right values.  But not fully -- and I see something 
that I overlooked then -- the histograms are distinctly non-gaussian 
(with the same extended wings) both with self-test on, and self-test 
off, for the parallel hand correlations.  As always, the cross-hand 
correlations were fine -- both in width and in shape. 

    Ken assures me the system has not been changed between April and 
now.  So -- how can the system give strikingly different results for 
these 12.5 MHz modes?