Can Exoplanet/Pymc3 interact with TTVFaster?

[asmasca]

Hi Dan, I'm Alejandro Suárez Mascareño, we spoke at the Oxford GPRV workshop.

Yesterday I started playing around with Exoplanet/Pymc3 following your recommendation. First of all I want to say I'm impressed about how efficient it is, and how convenient lots of the functions are. Thanks for pointing me here.

I tried to move a likelihood designed for Dynesty/Emcee into Exoplanet, that includes a call to the python interface of TTVFaster (1). I found out that TTVFaster really doesn't like the objects created by pymc3. Do you know if there's a way to make them interact nicely? Or an alternative that can be used?

I know I can always sample directly the times of transits (as in the tutorial), and then compute the masses later, but I wanted to sample the planetary parameters directly into the model to fit activity, transits and RV simultaneously.

(1)https://github.com/ericagol/TTVFaster/tree/master/Python

Thanks in advance

[dfm]

Unfortunately it wouldn't be trivial to interface TTVFaster and PyMC/exoplanet since a differentiable Theano/Aesara op would be necessary. That being said, I bet it wouldn't be too brutal to implement the TTVFaster equations directly in Theano.

[ericagol]

Yes, they are fairly straightforward. The most complicated part is computing the Laplace integrals.

[asmasca]

Ok, that sounds like a thing I might be tempted to try. Is the python code (ttvfaster.py) fully functional? or would it be better to continue using the C version (as called by _ttvfaster.pyx).