[daip] CALIB

[Eric Greisen]

Lynn D. Matthews wrote:
> CALIB in 31DEC08 seems to be behaving oddly in the way it handles the
> requested SOLINT.
> 
> Here are the results of some tests I've done on two different 
> single-source data sets (10s and 1s dump times, respectively):
> 
> AIPS    Record	SOLINT	# Solutions   SN table 
> vers.   Length                        delta t 
> ------------------------------------------------
> 31DEC07  10s     20s      1100        20s 
> 31DEC08  10s     20s       910        25s
> 
> 31DEC07  10s     30s      718         30s 
> 31DEC08  10s     30s      576         40s
> 
> 31DEC07  1s      1s       10750       1s 
> 31DEC08  1s      1s        7824       1.5s
> 
> 31DEC07  1s      2s        7199       2s
> 31DEC08  1s      2s        4800       3s
> 
> 31DEC07  1s      3s        4800       3s
> 31DEC08  1s      3s        4800       3s

I changed the averaging in 31DEC08 in an attempt to make it work better 
- which it will in some circumstances.  But averaging is a black art in 
AIPS in part because the times are not kept in double precision.  Note 
that if we did make that change (which would be VERY VERY nasty to do) 
then we would have to delete FITTP from the system which would rule out 
carrying data to any package but OBIT.  The lack of double precision 
means that 1sec may be 0.90 to 1.1 sec if the time is not large and can 
be entirely lost if the time is several days.  I am actually surprised 
that things seem so regular.  I did find something I did not like in 
CALIB and have changed it.  A datum was included in the average if its 
time was <= start datum time + "solint".  That would mean with exact 
times a 1s data and solint 1s that you would get 2 samples in each 
integration.  I have changed that to <= start time + "solint" - 0.001s
but all this is subject to the vagaries of the time recording and 
storage accuracies.

To avoid these vagaries some, the old versions used to expand the 
intervals by a small amount (5s in the old days) but that led to strange 
behavior now that we have NX tables all the time.  So I now try to do 
things exactly and use the start time of the first sample in an 
integration rather than some attempt at a regular grid of times.  If you 
really want 1 sample per integration say, with 1s data, SOLINT = 0.9s in 
the new versions.

Eric Greisen