[daip] help with UVLSF, please

[Eric Greisen]

Without even looking at the plots, I think I can answer to some
extent.  Higher order fits are really not justified by the noise in
single records of vis data.  In particular, the high order fits are
not constrained by data in those channels excluded from the fit.  If
there are too few channels for the S/N and the higher order then
things will get out of hand in the unconstrained regions of the
spectrum.  In particular, you will see bad results in scalar averaging
in which the fact that the phase also got out of hand in those regions
does not protect you from bad signals.  vector averaging with phases
wandering wildly will reduce the large amplitudes that appear.

UVLSF with high order will only work if there is really a significant
continuum signal and you can use e.g at least half the channels in the
fit.  In particular, the lower numbered channels need to be line
free.

In other words, this change to UVLSF is not magic although it works
well in a limited set of circumstances.

Eric Greisen