exoSAMSI day one

Today was the first day of the SAMSI workshop on statistics and computation for Kepler data, organized by Ford (Florida) and a few others (including, absolutely minimally, me). The workshop is a long one, because the goal is to get astronomers and statisticians together to solve some real problems. In the first day, far more happened than I can possibly blog responsibly about; here are some personal highlights:

The day started with Barclay (Ames) explaining the mission and its data, with Ford enforcing strict communicate-with-the-statisticians speaking style! Payne (CfA) talked a bit about long-period candidates, where in a single Kepler source only one transit is seen. What can we conclude? A lot, I think; both the depth and duration bring important information; that's something I want to keep thinking about in the next few weeks. Morton (Caltech) talked by phone about false positives and (very detailed) efforts to remove them by probabilistic modeling. He said that the right thing to do is compute the Bayes Evidence, something I agree with in principle, but disagree with in practice. I disagree because such computations are very prior-dependent and very hard; if you don't know your priors extremely well, then you can get substantially wrong answers after a very large amount of work. Besides, decisions fall outside the realm of Bayes, as my loyal reader has heard me say more than once. Don't get me started about long-term future discounted free cash flow!

Carter (CfA) blew us away with the awesome physics of three-body (and more) transiting systems. There are huge transit timing variations, exoplanets around binary systems, and more, all of which is insane to model and hard to find (because they generally aren't strictly periodic; sometimes they aren't even close). He is living the dream: There is an n-body integration inside his likelihood function. That's something that we tried at CampHogg in the past, and one of the things that got us thinking about methods for doing MCMC that minimize likelihood calls per independent sample.

Loredo (Cornell), Baines (Davis), and I talked about hierarchical models. Loredo was one of the first (maybe the first) to use them in astronomy, and remains one of the clearest thinkers in the business about measuring stuff in astrophysics. I am glad he is here. Baines showed that you can enormously speed up inference in hierarchical models if you transform variables. The challenging thing is that you need to change variables across layers in the model. Then he showed how to benefit maximally from multiple parameterizations with a modified sampling method. This is probably extremely relevant to our (future) projects.