Specutils template matching needs to handle resampling without returning nan

[eblur]

I am working on a tutorial for fitting an SDSS quasar spectrum with a template and solving for redshift: https://drive.google.com/file/d/1cOPKSH3ED_0fokcf07yPj2t2mqYwtnrP/view?usp=sharing

(Astropy Tutorials Issue #441)

Due to the nature of the template, the model has to be resampled onto the observed spectral axis in order to calculate chi^2 (see the first three lines of the _chi_square_for_templates). If the template axis doesn't span the observed axis over the full range of redshift values provided, then the resampling function returns 'nan', which breaks everything.

The function needs a mask to ignore regions that can't be resampled, or at least return a value other 'nan' for regions that are out-of-bounds of the template.

[nmearl]

@eblur This case should already be supported with the extrapolation_treatment argument in the resamplers. By default it is nan_fill, but you can use zero_fill instead. There's an example in the docs as well.

[eblur]

The template_comparison library doesn't provide any arguments or keywords for passing that information to the resampler. Thus, in order to apply this library to a very simple example (a spectrum downloaded directly to the SDSS archive), the user would have to re-write much of this library.

So, I see two possible ways forward: (1) Make zero_fill extrapolation the default for the samplers returned in the _resampler function, or (2) Provide a resample_fill (or similarly named) keyword in the template_match and template_redshift functions that allow the user to refine the template sampling method.

I suggest (2) because I can think of some examples where (1) could cause problems.

Also template_redshift should have resample_method keyword, like the template_match function.

[dhomeier]

Speaking of which, is there an imperative reason resample does not also offer some basic extrapolation option like const_fill (i.e. instead of setting the off_edges to zero, to the last valid flux value off the region)? Of course the simple solution would be quite sensitive to noise, so a better implementation should average over some region close to the edge, and at that point the options (and possibly also the temptation to abuse by extrapolating too far) may quickly get out of control. Still it seems for points say within one original step width this could often be a better approximation than zero.

[camipacifici]

Proposed solution: do not fill with zeros/nans, do not extrapolate. Use only the wavelength range that both the observation and the template(s) cover. Issue a warning that the template(s) does(do) not cover a large enough wavelength rage and the spectra are truncated.

[rosteen]

I implemented @camipacifici's proposal in #1121, closing this.