[evlatests] Still more on spectral line noise

[George Moellenbrock]

Rick,

> Note that for all observations, the data were
> loaded as correlation coefficients, and for all datasets, the correction
> required to convert these to flux density were the same.  In other
> words, the apparent loss of sensitivity seen in the noisier maps and
> wider histograms for the 12.5 MHz data were *not* seen in the
> correlation coefficients.  How can this be?

By "not seen", you must mean that the correlation coefficients are not 
_reduced_ by this effect, such as an increase in pure Tsys would do.  I 
thought this was already established by the claim that fractional 
polarizations come out ok for mode(s) where the cross-hands have the 
~correct noise.

I presume your SEFDs are the just histogram widths corrected for bandwidth 
(so you can uniformly compare different BWs)?  This, combined with 
globally consistent calibration factors establishes that this is not a 
pure (pre-corr) system noise effect.  Since it is presumably not the 
bandwidth (bandpasses look ok?), it must therefore be that the effective 
integration time is actually only some fraction of what it is supposed to 
be.  Your SEFDs are revealing not just Tsys(Jy), but rather 
Tsys(Jy)/sqrt(dt), where dt is the effective integration time.

Remember, the Tsys(K or Jy) sets the scale of both the noise and the 
correlation coefficient (and thus the calibration factors you require to 
reach Jy), but the integration time (and bandwidth) influence only the 
scale of the noise.  I think the cross-hands in mode PA are the most 
interesting fact available.  They show that the correct full sensitivity 
is indeed somehow available for 12.5 bandwidth.  If these cross-hands were 
also noiser, it would have been much harder to notice this problem. They 
are the key to understanding it, I think.

I would confirm that the fractional pol comes out right in mode PA for a 
known calibrator (e.g., 3C286), to confirm the above assertions.  I think 
the pol tests earlier in the year probably already establish this.

Then I would concentrate on understanding how most 12.5 MHz bandwidth 
spectral line correlations might somehow be throwing away ~75% of the 
integration time (or duplicating 25% of the bitstream by 4X?)  I don't 
know how the h/w & s/w work.  Is something like this possible?

-George