[evlatests] RQ resetting comments

[Rick Perley]

    I was unable to attend Bryan's meeting yesterday, dealing with RQ 
adjustments.  Based on some feedback from those in attendance, I think 
there may be some misunderstanding remaining on the issue of RQ 
settings.   I write below (following some discussion with Ken to ensure 
my understanding is correct), some comments on the need, or not, for 
adjusting RQ settings.

     1) The RQ adjustment attempt to optimize the digital power within 
each spectral window.  Optimize here means to maximize the SNR, 
following the prescriptions found in TMS (for example).  In practical 
terms, this results in the 4 bits (16 levels) of the output requantizers 
being nearly fully populated, so that a small fraction of the noise 
samples fall outside the 16-level range. These over and under range 
values are not ignored -- they are counted as maximum or minimum, 
leading to underestimates of the total power.

     2) Without SPW dependent RQ adjustments, a single value is used for 
all spectral windows.  Spectral windows at the edge of the passband 
normally have lower than average power, while some in the middle will be 
higher.

     3) Because the optimum SNR setting fully populates the 16 levels, 
it is far better to be low in power than high.   Each bit in the sampler 
is a factor of four in power -- losing 1, or even 2, bits is hardly 
critical.  However, being too 'hot' by a factor of 4 in power is bad, 
and will lead to large correlation errors, due to overflow.  (This is 
easily seen in my P-band tests, using Cygnus A). Being too low is not 
entirely without consequence either -- the 'van vleck' correction will 
change depending on how many sampler levels are actually used.

     4) So, by how much does the SPW power vary?  There are three cases 
to consider:

     a) Variation between sources or sky positions.  The strongest 
isolated sources -- Cas A and  Cygnus A, each cause the total power to 
rise by about a factor of up to 5 at P, L, and S bands.  This is 
slightly more than one bit -- and hence of modest concern.  (It is why, 
when I observe these sources as part of a run, I set up the system 
attenuators and RQ on Cygnus A, arguing that being low on the colder 
objects is better than being too high on Cygnus).    More significant is 
the variation in sky background power (temperature). At low frequencies, 
the sky dominates system power.  The variation in power easily exceeds 
an order of magnitude between the galactic center and the coldest sky.  
Factors of two between 'ordinary' sources, or between calibrators, will 
be standard.  This is of significant concern.

     b) Variation across the passband.  To judge this, I looked at some 
real data.  Note that examination of the BP table is not sufficient to 
judge this issue -- the SNR can vary dramatically across the bandpass 
(especially at P and L bands),and the BP solution sees only the 
correlated power.  I used the digital power (PSum) monitor for this (as, 
unfortunately, the autocorrelations had been removed by my processing).
           At LBand, the maximum variation I could find, between SPW 8 
and 16, was about a factor of 5.  Most antennas had the range within 
about a factor of two.  All these ratios are for the same source -- this 
does not include the on-off Cygnus A problem -- which boosts the maximum 
range to about a factor of nearly 100.  Thus, if you're not observing 
Cas or Cygnus, the current arrangement of not applying SPW-dependent RQ 
corrections should be safe.
           But at P-band, the situation is much worse.  The *typical* 
range between SPWs for a single antenna was a factor of 10.  The extreme 
between antennas was a factor of 100 -- for the same cold sky.   At this 
band, (and at 4-band) we really must employ RQ corrections for every 
separate observation.

     c) Variation in the BP, due to strong RFI.  A SPW with strong RFI 
will certainly overfill the 4-bit RQ window.   For some SPWs, notably 
SPW3 in S-band, it does more than overfill the 16 levels -- it can 
overflow the accumulators of the correlator, causing complete rubbish in 
the spectrum on ~ 50% of the scans.  Implementing the RQ correction 
(scan by scan here is recommended, since the RFI is so dynamically 
variant, due to the sidelobes) *should* help both overflows (but I'm not 
promising it will make everything better).

     5) Finally -- a note on my experience with utilizing RQ 
corrections.  For the P-band holography, we employed a single RQ setting 
at the beginning.  This works fine, as we were not changing source or 
tuning.  (Note that any change to the correlator will 'lose' the RQ 
settings).
     For my calibrator model/flux density work, RQ resetting was 
employed for *every* scan at P, L, and S bands.  It was arranged that 
this be done 20 seconds after the first antenna reached the source.  
(There's apparently a python command for this).  The 20 seconds was to 
allow the system to setup, primarily to move the subreflector.  This 
procedure worked beautifully well.



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listmgr.nrao.edu/pipermail/evlatests/attachments/20150916/8b30d019/attachment.html>