[widar-wg] R-R test

[Michael Rupen]

We have been looking into the Stokes V problem intensively this week.
Stokes V is RR-LL; there are some indications that there is a problem
with the LL correlation, which Bob will discuss on Tuesday.  As a partial
test of the path RR and LL take through the correlator, we configured
the correlator to tune two filters to the same frequency setting,
and send the resulting two (hopefully identical) subbands to two different
baseline board pairs for correlation.  The BlB pairs, and their
configurations, were chosen to match those used in the recent Stokes V
tests which suggested the LL problem.

   - The most obvious difference is that we are routing RCP signals for
     both subbands in this case, vs. one RCP and one LCP for the true
     Stokes V experiment.  This means that, instead of using both data
     paths for a single wafer out of the crossbar board, we are using
     single data paths for two wafers.  The CTRL signals which go along
     with the wafer come from the RCP Station Board I believe (right Bruce?),
     so this could be an interesting difference.  Less plausibly interesting
     is that the crossbars in the recirc chips are set differently for the two
     cases.

For this experiment we expect the two sets of RR correlations to be
absolutely (bit perfect) identical.

We (meaning Emmanuel :) did not check this explicitly in the lag frames,
but took the data into post-processing (AIPS) to check the full data path.

Indeed, most of the baselines showed no detectable difference.
Some baselines showed differences at the level of a part in a million or so.
These are the baselines whose correlations had to be "flipped" by the CBE:
e.g., a correlation 7-2 has to be flipped to be 2-7 before being written
out.

This is plausibly explained (as suggested by Ken) by the use of
single-precision floating point in the flipping process, which has an
accuracy of order a part in a million.  Admittedly the input signals are
supposed to be absolutely identical, and one would have hoped that the
flipping process behaved identically with identical inputs.

In any case, no differences were seen with anywhere near the amplitude needed
to explain the Stokes V problem.

           -- Michael (reporting Emmanuel's data reduction results)