stop fully masking negative fluxes

[megbedell]

Currently if any fluxes in the HARPS data are reported as negative, we set the ivars to zero.. but this has the effect of leaving no data in the deep cores of super strong lines, which makes the template go to continuum in the middle of the line (bc regularization). We should change this to preserve a little information in the line cores. This is relevant especially for doing science with the stellar template!

[mzechmeister]

Negative data points can occur simply at zero flux due to readout noise.
I flag significant negative flux values.
https://github.com/mzechmeister/serval/blob/73e1fd843b196ab75f5b95c3652519b7ed90c2d3/src/read_spec.py#L513
In contrast to that, negative template fluxes appear unphysically and should be avoided. Probably, it can not occur in your model, since you work in log-space. But I'm curios how you transform the negative data.

[megbedell]

Thanks, this is useful info! Currently I take every pixel that has unnormalized flux count < 1 and set it to 1, along with setting the inverse variance on that point to zero so that it's essentially masked out:

wobble/wobble/data.py

Line 122 in 041cd58

def mask_low_pixels(self, min_flux = 1., padding = 2, min_snr = 5.):

I also mask out pixels on either side of the "bad" one, although that may not be the correct thing to do if it's read noise causing the negative fluxes. In the cases I've seen it seems more like the background sky subtraction (which is done in the HARPS pipeline spectral extraction) is overly aggressive.
I do this setting to 1 step because transforming the flux to log space won't work otherwise - so as you say we don't have any issues with unphysical negative fluxes in the template, but the problem is sort of transferred to come earlier in the process when logging the data!