[evla-sw-discuss] Telcal -> antsol -> cpslss

[Barry Clark]

At the heart of Telcal is the antsol routine, which solves for each
antenna's amplitude and phase.  There are questions in my mind about
what this routine should do in the Widar case.

1.  Should it operate independently on each subband, or attempt a
global solution for the whole baseband?
You certainly get a better signal to noise, and thus can work on
fainter sources by solving for the whole baseband.  I tend to think
it is not worth the effort of implementing this rather complicated
algorithm.

2.  Should it simply sum all channels for a given baseline and subband,
or should it simultaneously use the channelization to search for a delay
error while solving for gains?
The answer to this one depends on both the a priori accuracy to which
we maintain our instrumental delays and on what we will use the antsol
solutions for.  If, as now, we are fairly sloppy about maintaining
instrumental delays, at sort of the 1-2 ns level, and if we want
accurate instrumental responses saved for both tracking the instrument's
behavior over the long term and for tracking source fluxes, we might
need a multichannel antsol.  A delay error of a quarter turn across the
band (one ns error in a 256 MHz subband, for instance) corresponds to
about a 10% error in the amplitude.

So I thought I would look into generalizing antsol to a multichannel
situation.  It is not clear to me if we should use it, because of the
substantially higher computation required, but it might be worth looking
at.

The routine at the bottom of the current antsol is called cpslss.  It
takes a set of complex values, one per baseline, and solves for the 
response of each antenna, in both amplitude and phase.  This routine
has an amusing history.  It was originally written in Sail by Larry
D'Addario for the Dec 10 data reduction system.  I (or I think it was I)
translated it into Fortran for the Modcomp version of telcal, and Ken
Sowinski translated it into C for the interim telcal.  In the process,
almost all comments have been lost, if they were ever there in the first
place.  And it is a rather obscure program.  It is an iterative
solution, because the problem becomes non-linear in the low signal-to-
noise case.

Anyway, just to see how it looks, I propose to generalize cpslss to the
multichannel case.  The only obvious problem with doing so is that the
convergence criterion used in the single value version makes no sense in
the multichannel case.  I will, of course, write it in Java, making the
fourth language for the algorithm.