It took me a while to get in to LPNHE; I had no response from Seb Bongard on chat and he wasn’t in his office when I arrived. I set up shop at a desk in an open lounge area, and eventually Pierre Antilogus, an old SNfactory colleague, ran across me and handed me off to Nicolas Regnault, our local contact for the SkyMapper Hubble diagram work. Everyone was surprised to see me; the weak link in my plan was communicating only with Seb. Lesson learned: next time I must email more than one person, and email them far enough in advance that they can plan to be there — and close enough to the date that everyone is prepared.
Nevertheless, my French colleagues very graciously dropped at least some of the things they were doing to make sure I had a desk, wireless access, lunch and so forth. Seb arrived after a few minutes with Nicolas, and we all sat in the common area drinking espresso with Pierre Astier and Julien Guy.
After the initial shock of discovering I was visiting, everyone wanted to know how SkyMapper was doing. I told them as much as I knew: that an intervention to fix the vibrations limiting the telescope’s image quality was due for sometime around… now. There is some concern about engineering manpower (or is it now “resourcing” in admin-speak?) for the ongoing SkyMapper commissioning, but we’re hiring a new project manager soon, so hopefully we’ll be able to accelerate progress substantially when that person takes over.
Seb and I then talked about Gaussian processes (see this free e-book for a good technical introduction), which are becoming increasingly popular in SNfactory, currently being used in at least three different ongoing research projects. So we went over some of the mathematical foundations of the method and discussed its advantages over, say, spline fitting. My experience is that GPs beat data smoothing and spline fitting because it lends itself well to Bayesian inference; you get the whole posterior to sample, instead of just a best fit, and you can sample from the posterior to build confidence regions (which you can’t do with straight smoothing) and/or to run simple Monte Carlo analyses on the results of the regression. (For example, you can calculate the same absorption line measurements on a curve selected from the posterior of a GP regression fit to a supernova spectrum as you would on the spectrum itself, and from the spread of derived values you get more realistic uncertainties than you would from normal propagation of errors.) Still, it’s easy to abuse GPs, and you have to make sure that your prior assumptions, like the chosen form of the GP covariance function, are plausible and well-suited to the problem you’re trying to solve.