making inference faster, K-giants

Here at Camp Hogg we aim to provide a satisfying, high performance data-analysis experience. Today Andreas Berlind (Vanderbilt) swung by and told us about his need to sample a halo-occupation posterior PDF with only ten-ish parameters but a likelihood function that takes ten minutes to execute (because it ambitiously uses so much data). No problem! That's the kind of problem that emcee was built for. Foreman-Mackey and I described the algorithm to Berlind and gave him the task of providing a callable posterior PDF (or really likelihood times prior PDF) function.

In the afternoon, Foreman-Mackey explained how he made his Stripe 82 calibration model far faster and more robust, by switching to a model that is an approximation to what we want but has fast, analytic derivatives. Like with the censored data last week, an approximation to what you want that is much faster can be better in the long run. Also, I am not convinced that what we originally wrote down for our Stripe 82 recalibration isn't an approximation to Foreman-Mackey's new formulation; that is, this might be a conceptual improvement as well as a performance improvement.

In the morning, Hou nearly convinced Goodman and me that he has found a non-exoplanet-caused 17-day period in some K-giant radial velocity data. It isn't an exoplanet because it shows phase and amplitude variations with time. We are currently calling it a damped oscillator but I am wondering what it is going to turn out to be, really. Can K-giants have oscillations at such long periods? I wish I knew something about stars.


  1. The hammer is back. With ten minute likelihood calls and ~10 dimensions I would have no hesitation in recommending emcee or MultiNest over DNest. :)

  2. In addition to the period, it's probably also worth checking the mode lifetime (something related to the damping constant in Hou's model) for consistency with what's known about oscillations in giants. That's a lot these days because of Kepler.