MCMC kitchen sink

At our weekly MCMC meeting, Goodman reminded us of the Leslie Greengard (NYU) philosophy that any high-performing numerical method simply must be complicated. No simple ten-line algorithms can solve a large class of problems fast and in a wide range of cases. That's a bit of a cartoon, but you catch my meaning. Along those lines, we discussed heterogeneous approaches to MCMC, to work towards black-box solutions that would work without tuning or optimization or specialized initialization or long execution times on large classes of problems with challenging probability functions. In particular we looked at problems with our ensemble sampling methods, and ideas for methods that combine classification (of regions of parameter space) with sampling.


WFC3 calibration

Fadely and I had a great conversation with Kalirai (STScI) and Deustua (STScI) about WFC3 IR channel calibration, in preparation for our HST Archival proposal regarding self-calibration. We learned a lot about effects to think about but also the provenance of the current calibration information. It is hard to overstate the influence STScI and its calibration teams have had on astrophysics: The availability of calibrated data, free to everyone, has made it easy for anyone to be an observational astrophysicist!

I short-changed my reader yesterday with my brief description of my trip to Princeton. I learned from Dave Spiegel (IAS) something tremendous: At km (yes 1000 m) wavelengths, Earth is the most luminous body in the Solar System. More luminous than the Sun (or Jupiter or anything else). That's exceedingly interesting, don't you think?



I spent the day in Princeton, with Foreman-Mackey. We spoke with Spergel about WFIRST, Bakos about HAT, Tremaine about exoplanet distribution functions, and Turner and Spiegel about predicting and characterizing exoplanets. Before lunch, Guhathakurta (UCSC) talked about Andromeda and its substructure. At lunch, I talked about self-calibration. And in the late afternoon I gave my seminar, about probabilistic methods for measuring exoplanet populations. I also included a tiny bit of XDQSO, which was a good idea, since it really showed that hierarchical modeling works.



At lunch today, Blanton gave a very nice talk about the three spectroscopic projects that make up SDSS-IV. It reminded me once again how much potential there is for ground-breaking research with APOGEE spectra. Late in the day I worked on my slides for a seminar at Princeton tomorrow; in particular I worked on the problem of measuring the properties of a population of objects, no single member of which you detect significantly.


vacation week

I took a break for the week. I spent it in the wintry woods in an undisclosed location. I just did station-keeping email and a bit of writing each day. Apologies to all of you with deadlines! Will re-start research next week.


statistics for graduate students

I don't know whether this should really go on my teaching blog, but we (Blanton, Cranmer, Tinker, Modjaz, and I) have started an informal (open to anyone, well attended, join us) friday-morning statistics for graduate students class. (The reason I think this class should go here and not in my teaching blog is that in fact I think this kind of seminar class is really a research activity; that's what makes it such a good idea.) Today was Kyle Cranmer talking about the properties of estimators, the Cramér–Rao bound, the bias–variance trade-offs, and other theoretical matters. He made some comments at the end about the best estimators being optimized with respect to the statistician's utility, which is an idea right up my alley; I will return to those ideas in my lectures, which come later in the semester.


so many conversations!

Today was an all-talk day. I started off with Stumm and Foreman-Mackey for coffee, where we talked about various matters of only limited relation to astronomy. I chatted with yesterday's speaker Ruderman (Berkeley) about naturalness and the future of particle physics. Then Bob Lindner (Rutgers) arrived and we did an early lunch, at which I pitched improvements to radio astronomy image reconstruction methods. Lindner gave a mid-day talk on clusters observed in the radio and sub-millimeter to look at S-Z effect and star formation (and lensing). The afternoon was filled with our MCMC meeting, in which we battled Sven Kreiss about how they put error bars on parameters in the LHC ATLAS world. We learned something about the relationships between frequentist confidence intervals and the Bayesian equivalents in high dimensional parameter spaces. Awesome!


calibration, naturalness

Schiminovich and I met uptown to do some coding in preparation for testing the GALEX attitude model. I came back downtown for a great talk about technical naturalness in particle theory (especially SUSY) by Josh Ruderman (Berkeley). In the late afternoon, Fadely and I realized we are at the go/no-go decision on our HST self-calibration proposal.


transits and radial velocities

Foreman-Mackey and I pair-coded radial-velocity capabilities into his transit-modeling code. This may be monstrous mission creep! But it is so cool that we can handle all the data on the system simultaneously. We are also close to having astrometric modeling going at the same time, so we are even ready for Gaia. In my spare time today I worked on a demonstration that with proper inference, you can measure the properties of a population of objects, no single member of which you have ever detected (in the statistical significance sense of detected).


limb darkening

Today Foreman-Mackey showed (as we predicted) that our inferences about exoplanet sizes and orbital inclinations (relative to the line of sight) are modified as we loosen our prior beliefs on limb-darkening. That's a nice result for our first paper on Bart, Foreman-Mackey's rapid exoplanet transit inference system. In other news, I wrote words in my mixture-of-Gaussian paper (revision for referee) and listened to a very nice seminar by Scoccimarro (NYU) about velocity-space distortions and their use in testing metric theories of gravity.


automating high-resolution spectroscopy

I spent a great day at Utah, in the care of Adam Bolton. I gave my talk in the morning and then talked with the graduate students. In the afternoon, I visited two large group meetings, one the combined groups of Bolton and Kyle Dawson, the other the group of Inese Ivans. A lot of things were discussed in both meetings. In the Bolton–Dawson meeting, among other things, we discussed what to do with objects that have uncertain redshifts or redshifts based on a single emission line (which will be a problem for SDSS-IV). In the Ivans meeting, we discussed a wide range of problems related to high-resolution spectroscopy. Ivans and her group is finding that it is very hard to automate certain parts of spectroscopic analysis of stars. We discussed how these might be robustly automated. One problem—which also exists in cosmological projects that take spectra of quasar absorption lines—is to find the continuum relative to which absorption lines are measured. We came up with some ideas and I volunteered to work as a consultant in the future.


exoplanets are hard to find

I sent the title of this post to Utah as the title for my seminar there. Now, on the airplane I realize that I have nothing to say about this! Well, not nothing, but it is a stretch. In preparation for this talk (which is tomorrow) I worked on making these arguments: (a) Software is as important a component as hardware, and has to have contributions from extreme experts (just as, say, optics does); from now on you won't be able to write your software with part-time people (like students whose real work is something else) and in the evenings and weekends. (b) Hierarchical inference will make it possible for us to discover exoplanets that would be invisible without it.



I spent the day in and near AMNH, where Schiminovich and I tried to re-start all our GALEX projects. He proposed that we look at some of the very most recent data, where GALEX is pointing at high stellar density regions of the Milky Way. We were interrupted by Lang, who was finding issues in a SDSS-WISE match. He is making the match in preparation for (as an initialization for) a big fit of the WISE imaging with an SDSS-based model. There are missing objects from SDSS; we tracked (at least some of) them down to the exceedingly complicated definition of SDSS survey primary.



Other than two very pleasant conversations with Fergus and Foreman-Mackey, nothing got done today.


streams and inference

I spent my research time today getting ready a short note to share with Andreas Kuepper (Bonn) regarding likelihoods for comparing models of kinematic objects with observations of collections of particles (like stars) in a kinematic space. This is all related to fitting better models for cold stellar streams (and other structures) in the Milky Way halo. At the end of the day I sent him what I have, which I am very much hoping will be the skeleton of a future paper, perhaps also with Marla Geha (Yale).



Chris Martin (Caltech) completed his week at NYU with an astro seminar on GALEX, which has been amazingly successful and productive (especially per dollar). When GALEX came to its end of life, Caltech convinced NASA to transfer it, so Martin is not just the PI, he is the (effective) owner. He spent most of the time talking about the star-formation history of the Universe, but he also showed some nebulae that are fluorescent: They emit essentially only in the far-UV! That's cool: they are, in some sense, hotter than infinitely hot blackbodies. That's not really true, of course, but there aren't that many circumstances in which one can arrange such exceedingly non-thermal emission. As both my loyal readers know, Schiminovich and I are trying to extract the full photon stream from GALEX and calibrate it for time-domain science. If we succeed, we may end up making GALEX even more productive posthumously.

ps. Martin told us in the seminar that for about a million USD, you can buy the next year-ish of GALEX operations!