[evlatests] Exercises in Imaging ...

[Rick Perley]

    The X-band database that Michael provided me yesterday is useful to 
checking the array sensitivity, and exploring some issues with snapshot 
high dynamic range imaging. 

    To recap, the observation is 14 minutes long, at X-band, with two 
subbands and dual polarization.  Channel resolution is 500 kHz.  Each 
subband is 128 MHz wide.  Time averaging is 0.1 seconds.  This fine 
resolution in time/frequency allows precise editing for outliers.   The 
two sources observed were 3C286 and a blank piece of sky 2 degrees away 
in declination.  After flagging the travel time, we end up with 5.3 
minutes on 3C286, and 3.4 minutes on the blank field. 

    To replicate, more or less, what we might actually do in the future 
with D-configuration data, I decimated the data down to 2 seconds 
averaging, with 4 MHz channel resolution.  Flagging and delays were 
applied prior to the averaging. 

    Calibration was then done in the usual way.  A bandpass was solved 
for (once for each of the four short observations of 3C286), and 
amplitudes and phases computed using the central channels.  Two point 
interpolation was used to calibrate.  The CL table was generated to 
match the times of the integration. 

    The imaging results of the blank field match the expected noise 
levels exactly.  The single-baseline histogram shows an rms of 80 mJy 
(about right for the known SEFD, time and frequency averaging).  The 
noise on the image (using robust weighting) declines exactly as expected 
for the number of channels, polarizations, and subbands  utilized.  All 
is well. 

    But the results for 3C286 are (unsurprisingly) not so good.   To 
simplify:

    a) For a single central channel and RR polarization, the rms noise 
in the image is as expected:  0.89 mJy.

    b) As the number of RR channels utilized is increased (but always 
centered on mid-band), the rms noise first declines, then rises:  0.50 
mJy for 4 channels, 0.39 mJy for 16 channels, but 0.52 for the full BW 
(30 channels). 

    c) Combining more polarizations (RR and LL) and subbands makes the 
images worse!  (This is largely because there is something wrong with 
the LL side, but I can't figure out what it is). 

    An easy suspect for the failure for the noise to decline with 
sqrt(Nchan) is the rocking phase due to the delay 'clunks'.  This will 
affect the edge channels the most.  An image of single channels near the 
edge of the bandpass easily confirms this -- channel 10 (RR) gives 1.15 
mJy noise.  Channel 2 gives 1.62 mJy.  (Compare to 0.89 for the center 
channel, which is not affected by the clunks). 

    So I then unleashed BPASS, to make a bandpass solution for *every* 2 
second integration and each 4 MHz channel.  This worked!  But the 
resulting images are still less good than hoped for:

    a) All single channel images (whether from the edge or the center) 
now give about the same noise:  0.91 mJy for the center channel, 0.99 
for channel 2.  So far, so good.

    b) Multi channel images are notably improved -- for the full BW 
image the rms noise is 0.35 mJy, rather than 0.52 mJy for the averaged 
BP approach.  However, this is still not as good as it should be -- I 
expect 0.16 mJy. 

    To see if the failure to reach expected noise is due to the detailed 
BPASS, I split the data into 32 single-source subsets (each with 2 
second averaging and 4 MHz bandwidth), then combined them all together 
with DBCONs.  A single point-source calibration was done, and the 
results imaged.  The noise went down to 0.28 mJy (from 0.35), but still 
a factor of two too high. 

    I then tried to run BLCAL to compute closure offsets.  The resulting 
solutions are disturbingly large (up to a couple of percent), but I 
suspect are largely driven by noise.  Unfortunately, due to an AIPS 
issue, I cannot apply these solutions to the data.  (I'll consult with 
Eric on this). 

    This is about as far as I can get with this dataset.