“Always write down the probability of everything” another fav, page 61

Pretty sure I heard this somewhere, the analogy has been coming to mind more and more as AI slop pollutes primary source information

Another benefit of HMC should be joint modeling of disparate noisy sparse datasets. RV time series come to mind, you already showed photometry. Spectopolarimetry is tricky but possibly an opportunity: could leverage symmetry of multiple repeated projected disks with some physics informed modes.

Wait— there’s some similarity of the math here that should apply to the CMB, but only in reverse and inside out. In the CMB we directly sample the 3D heatmap, and then apply FTs to get correlation statistics then physics. Spherical harmonics decomposition, mapped through Autodiff is useful there.

Yes exactly, probabilistic reconstructions, so you can make trustworthy statements of confidence of any given reconstruction. For example answering questions on latitudinal distribution of spots or spot evolution, or ensembles of systems. So many directions to go with this.

Yeah certainly no urgency since the lack of targets. Awesome work team!!

It would collapse some of the symmetric modes and such. Of course probably not any real world analogy system exists, so mostly an academic exercise… but still!

Would love to see the information gain from hypothetical transits or eclipses of sufficiently resolvable bodies: basically you can do some amount of information stacking from the fact that you know/constrain the orbit.

I did have autodiffable FFTs if I recall, which blew my mind at the time, still sort of does. Your performance must be incredible with closed form and autodiff.

Closest analogy I can think of is Ian Czekala’s work on ALMA image reconstruction, but did not have the spherical harmonics component since it represented potentially arbitrary scene reconstruction. I think that was in Julia and then PyTorch if I recall correctly.