I got started on various late writing projects today. I wrote about generalizing The Cannon to deal with spectroscopic data that are taken with a variable-width (or variable shape) effective line-spread function in the spectrograph. This could be caused by the instrument or by rotation or convection in the star. This is trivial to do at test time: You just convolve the spectral model before comparison to the data. But at training time this is difficult. We either have to run the convolution backwards, or build a model in the deconvolved space. Related to this, I had a conversation at lunch with Alex Barnett and Leslie Greengard about when you can invert a convolution operator. The answer varies from “never” to “over some range of spatial scales” to “it depends on the signal-to-noise and scale”. Thinking about all this on behalf of Andy Casey.
I also wrote (well typeset equations, anyway) for my project to build the minimal, interpretable, data-driven nucleosynthesis model. This is one of my answers to Hans-Walter Rix's question: Imagine you had 100,000 stars with 15 measured chemical abundances (and we have this!); what do you do with them?