[evlatests] Images from WIDAR, C band

[Vivek Dhawan]

Data of 1st September, ~80 minutes @ C band, 4 subband, 4 stokes.
There was 3C286 (7.5Jy, offset by 1'), a blank field, and 2 more
I did not examine in detail.

The summary is that post-processing is limiting the dynamic range,
but I do not yet have a simple recipe to get the best images.

Details:

There were a few 'zero dropouts' - Martin knows about this.

Straightforward reduction on 3C286 (see end) gets ~5,000 dynamic
range in a 4MHz BW, i.e. 1.4mJy off-source rms in RR & LL images,
where 0.4 mJy is expected. V images get down to 0.8 mJy.

Making images with averaged channels in a subband does not reduce
the noise by sqrt(chans).

Subbands were processed separately and they should have independent
noise.  Indeed, averaging a chunk of channels across different sub-
bands _does_ improve the noise, so somehow my simple processing has
caused adjacent channels in a subband to get co-related.

With the same processing, the noise on the blank field behaves well,
keeps going down with wider BW, no special effort. The 'I' image
using 4x120 MHz, 16 minutes of data, has 60 uJy rms, about 2 times
the theoretical.

To try and improve the result on the strong source, individual 4MHz
chunks were split off, self-calibrated & imaged separately, and images
co-added. The noise improved as expected.  I think the default option
on BPASS (which I used) is not optimal. Unfortunately, no other option
I tried could improve matters sufficiently. I hoped that proper use
of Bpass would be equivalent to the split-and-selfcalibrate, but I have
not found the right recipe.


(*) Minimal processing is as follows:

1. Fringe fit over a few minutes, apply to all data, to set delays.
   This should be all that is needed. However, in this case, since the
   target was observed off center, the delay vs. time had to be fitted
   to track the (slowly) varying geometric delay.
2. bandpass over a few minutes, apply to all data.
3. Self-cal (few sec interval) to set flux and zero the phase.
4. Split - separate sources, multichannel data.
5. Image - average as many channels as desired.

Note the LCP data show a range of ~300ns in delays, the RCP ~50ns.
This leads to higher closure errors on the LCP, but FRINGE and SPLIT
respectively solve for the delays and correct for the closure error,
so finally LCP images are no worse than RCP after calibration.

Frequent BPASS calibration did little to improve the results.

Lastly, I did no polarization calibration, and 3C286 is ~10% polarized,
so I am a bit surprised that split-and-calibrate is able to get within
a factor 2 of the floor. I expected closure errors from polarization
but perhaps they do factor by antenna for a short ~1hr observation.

-------------------