Noise in Q and U, and MAD vs theoretical about rm-tools HOT 3 CLOSED

Hi @AlecThomson:

These results are pretty in-line with similar simulations I've done. It's definitely known/expected that uncleaned FDFs will report higher measured noise, as a result of sidelobes of the real signal. After many discussions with Bryan about this, we've essentially concluded that the measured noise in dirty FDFs should never be used. The theoretical noise seems to be well behaved and accurate, assuming the channel uncertainties are accurate (as they are in these sims).

I'm not sure why the estimators differ so much on the pure-noise FDFs. I was under the impression that the MADFM at least was reliable if there was no signal. The RMS I don't particularly expect to be well behaved when computed using the absolute value/PI, because of polarization bias. My policy is to never use/trust the RMS under any circumstance, and I've been debating just flat out removing it from RM-Tools entirely.

I think it is expected that the theoretical noise should be appropriate for Stokes Q and U as well as PI (in the high S:N regime). I've seen/done some derivations that essentially show that the noise in PI is equal to the noise in Q/U once you're at a decent S:N (>~5?) so that polarization bias becomes less of a problem.

Running the noise estimators in Q and U rather than PI is probably a good idea, I think (in my head, I half-thought they already were). The RMS at the very least should be better behaved.

The only really surprising thing for me is the difference between the MAD and the theoretical noise for the pure noise case. Is it perhaps not normalizing the MAD to the Gaussian 1-sigma equivalent? There's a conversion factor of 0.6-something, and so I wonder if that would make the difference.

To summarize my own conclusions on the subject from similar work:

• we can't reliably extract noise values from dirty FDFs. Even the windowing done by measure_FDF_parms isn't enough (although it helps).
• theoretical noise works pretty well in sims, where we know exactly the actual noise amplitude. Harder to know for sure if we can estimate the channel noise accurately on real data, but Max Fine's work on injecting simulated sources into POSSUM data were pretty encouraging.
• Q/U and PI errors are approximately interchangeable, as long as your estimator doesn't assume zero mean (i.e., as RMS does).

Cheers,
Cameron

from rm-tools.

Adding a note here after a discussion with Dave McConnell: there's probably an inconsistency in how we're defining the noise: the theoretical noise is, I think, nominally for Q or U (of the FDF), while the empirical noise is calculated as the MADFM of the PI (of the wings of the spectrum, which would nominally be noise-like for Cleaned spectra.

I should probably fix that to be consistent -- it would be easy to modify the MAD and RMS calculations to operate on Q and U. I'll add this to my list of bugs to fix before finishing the paper...

from rm-tools.

This should be fixed now. I removed the RMS noise estimate before, because it was easily biased high by outliers (like actual signal). The corrected-MAD noise estimate was flawed: it calculated the noise of polarized intensity, which has very non-Gaussian statistics, so it ended up biased low. I've changed it to calculate the MADFM of the real and imaginary components together. This reproduces the correct noise amplitude on simulations, and is reflective of the actual error in PI.

This produces the correct results on both signal-free simulations and Faraday-simple RM-cleaned simulations. Dirty FDFs still result in noise estimates that are biased high (albeit not by too much, in most cases), but that's inescapable -- for dirty FDFs, the theoretical noise may generally be more reliable.

from rm-tools.