NRFG, day one

Today was the start of the second annual Not Ready for Gaia workshop, which brings together the Galactic dynamics groups of Binney (Oxford) and Rix. We are in beautiful Oxford, staying at Merton College. I came as part of Rix's party. Many interesting ideas swung by; here are a few highlights for me:

Binney and Posti (Oxford) talked about equilibrium models for the Milky Way disk and halo. Binney claimed that any reasonable axisymmetric models will be well approximated by integrable axisymmetric models and argued strongly for using integrable models, at least as a starting point. Posti gave some examples.

Evans (Cambridge) argued that we should be using far more flexible models for gravitational potentials, arguing for expansions around simple models. He is spending some time finding simple models that support easy-to-calculate orthogonal bases for perturbations. Both he and Binney emphasized that a good expansion is one in which not too many terms are required for good representation of the data. That's a tall order!

I got into a fight with Evans, Belokurov (Cambridge), and Koposov (Cambridge) about their fitting of the Sagittarius stream. They constrain the stream using a few summary statistics of the data, not the full data. I claimed that their model was bad because it couldn't possibly go through all the data! However, by the end of the day they more-or-less convinced me that maybe the model does go close to the data. Still awaiting a very simple-to-make plot!

Schoenrich (Oxford) and Sanders (Oxford) both discussed the extended distribution function in phase space coordinates plus chemical and age coordinates. An argument broke out during Schoenrich's talk about whether you can "deconvolve" the radial migration process to find out the original birth radius of anything, and whether or not that would be useful. I enjoyed that! Rix pointed to evidence from Mitshang and collaborators that chemical tagging might not work; I have comments on that work which I hope to write up soon.

At the end of the day, Schoenrich showed what might be extremely good metallicity determinations using photometry. He can demonstrate precision, but not yet accuracy. Very interesting to watch; he is building on early work by Ivezic back in the day.


Dr Wu and Dr Duffell

Ronin Wu (CEA and Tokyo) came through NYU today, and told us about Herschel spectroscopy of M83. She finds that the spatially resolved and integrated spectra of star-forming galaxies seem to be saying slightly different things. She also made us think that the nucleus of M83 might be like a little mini-ULIRG.

At the end of the day, Paul Duffell (NYU) defended his PhD thesis. He had a huge crowd for the seminar and it was a great show, just as his thesis was a great read. He has moved supersonic numerical fluid simulation into some new regimes, with adaptive, moving meshes. His most exciting results (to me) are on the phenomenology of binary black holes, and (separately) gap-opening in protoplanetary disks. In both cases, he can make novel predictions for new observations. Gruzinov (NYU) and I encouraged him to make those predictions! Congratulations to Dr Paul Duffell on a great piece of work and welcome to the community of scholars (as it were).


Spergel and BICEP2

David Spergel (Princeton) came into town and talked about CMB cosmology. He spent a good fraction of his talk throwing doubt on the new, hot BICEP2 / Keck result on B-mode polarization on large scales. He is concerned that the experiments don't seem to satisfy the standard null tests. He laid out a set of requests, which are easy to satisfy. He noted that six new experiments can confirm the result if it is real, so we will know extremely soon. One amusing thing about Spergel's talk was the repeated point (obvious, but often overlooked) that because all CMB experiments are observing the same, single sky, they ought to agree to better than one-sigma, especially on large scales where cosmic variance dominates.


language parsing, photometric redshift templates

Brendan O'Connor (CMU) gave a talk late morning about natural language models for data science in the foreign policy and politics domains. He showed nice results based on subject–verb–object parsing of news stories. He also looked at some twitter data, showing an analysis of the emergence and propagation of the new twitter "words" (or acronyms) "idk" and "af". After the talk, at lunch, I complainined that natural language processing does not do a good job of understanding sentences and doesn't even give probabilistic results when uncertain. Yann LeCun (NYU) opined that understanding sentences is "AI complete", which is a phrase I need to use more often.

In the afternoon, Gabe Brammer (STScI) appeared and we talked about photometric redshifts. He is tweaking the model spectra using the data, and I suggested we go further in that direction. We gave each other homework on the subject.


precision and accuracy, post-aperture photometry

At Columbia today I met up with Bonaca and Price-Whelan and Johnston and Küpper. Bonaca has nearly finished a paper that asks what might be going wrong when we fit the Milky Way potential with a smooth, time-invariant, analytic model, which does not include all the messiness of substructure, mass accretion, and radial gradients of triaxiality. We discussed the issue that there is no really principled way to distinguish these different kinds of messiness; they are all just slightly different aspects of the same fundamental messiness, and they overlap substantially.

Price-Whelan has nearly finished a paper that looks at how we can make inferences about the Milky Way using small numbers of extremely well-measured stars in tidal streams. Bonaca's paper is about accuracy, and Price-Whelan's is about precision.

Late in the day I finished (or, really, started) testing (on real Kepler data no less) my method for making a generalization of aperture photometry that is much more precise. It works extremely well! I now want to combine it (unfortunately) with data-driven focal-plane calibration, because I have a (possibly wrong) intuition that it will shine even brighter relative to normal aperture photometry as the calibration gets better.


WISE parallaxes, binary AGN accretion

The WISE mission took two passes over the sky. So it is nearly pessimal for parallax determination; parallaxes and proper motions are all mixed up. That said, at each pass there are multiple epochs, so maybe there is a glimmer of hope. Today, in a heroic move, Lang coded up simultaneous parallax and proper motion determination in the un-coadded WISE individual frames, a la this old paper. As expected, parallaxes are badly determined. However, we can use MCMC to explore the full degeneracy between parallax and proper motion conditioned on the WISE data, combine external information, get ready for NEOWISE (or whatever comes next), and apply any prior knowledge. I don't know if it counts as research to just cheer while someone else codes?

At lunch, Brian Farris (NYU) gave a nice talk about simulations of accretion onto merging black holes. He proposes that we might be able to see late-stage (0.1 pc ish) merging black-holes by looking for periodicities in the accretion luminosity. Oddly, the periodicity is not at the orbital frequency but instead at some lower frequency that has to do with how the disk reacts to the torques from the orbiting pair of holes.



Today Dan Formeman-Mackey, Tim Morton (Princeton), and I measured the rate at which Sun-like stars host Earth-like planets! Foreman-Mackey did all the heavy lifting—it was a frantic hack day of sorts—but the work was based on the incredible data sent to us by Erik Petigura (Berkeley). Our measurement of "eta-Earth" involved building a period and radius distribution model which, when multiplied by survey completeness and transit probability, does a good job of modeling Petigura's scatter plot of data points.

IMHO, the correct definition of eta-Earth is the number of planets per star per natural-logarithmic interval in some pair of quantities, which could be radius and period, or which could be mass and insolation, evaluated at Earth's properties. That is, it involves an extrapolation or interpolation of any distribution function (measured from the observational data) to the location of Earth. From Petigura's data, and given assumptions (some listed yesterday), we get a number like five percent, plus-or-minus a percent or two. Importantly, this is five percent per natural-logarithmic interval of radius and per natural-logarithmic interval of period, so it should be compared to other papers and press releases with that clear definition in mind. Also, of course, it is conditioned on some pretty severe assumptions.

This blog post does not properly convey my Stoked-ness.

[Note added later: We found a bug in the code; the rate is probably a factor of two larger even!]


planet radius distribution

Erik Petigura (Berkeley) and collaborators found a turnover in the planet-radius distribution at small radii in this high-impact paper. They found this (more-or-less) by weighting their data samples with inverse selection probabilities. These kind of reweighted-data estimators are often unbiased but always high variance: They put the largest weights on the most marginal data. Fortunately, in the beautiful new world of exoplanets, sharing is the norm, and Petigura generously shared all of his data with Tim Morton (Princeton), Foreman-Mackey, and me. Awesome!

Today Foreman-Mackey and I pair-coded a forward model of the Petigura et al planet sample, using parameterized distribution functions, the completeness calculated by the original team, and a model for transit probability. With assumptions of independent planets (okay assumption), stationary distributions as a function of host star properties (bad assumption), negligible uncertainties (bad assumption) and separable period–radius distribution (bad assumption), it is possible to write down a fully justified likelihood function and turn the inference crank. That is, there are no real methodological freedoms. That's cool! We built and turned that crank today, also employing an iPython notebook (my first time). We got some preliminary results, which require some work to check. The thing I am excited about is that our assumptions are essentially the same as those of Petigura et al but our method is both simple and (conditionally) optimal.


Stripe 82, day three

After a one-day hiatus, it was back to the Stripe 82 meeting in Princeton. Highlights today were too many to mention, but here are a few of my favorites:

Ivezic (UW) showed incredible tests of Stripe 82 calibration, including evidence for residuals as a function of pixel position, and camera column. He can even show that the filter curve shapes must be different for the different camera columns! He showed some incredible Hess diagrams showing the Sagittarius stream in different cuts; he can even measure its metallicity with photometry alone.

Smith (Shanghai) showed some of the Koposov (2013) time-domain catalog in S82 that replaces—and is more precise than—the Bramich (2008) catalog. He showed amazing overdensities of halo stars in velocity space; just incredible.

MacLeod (USNA) talked about quasar variability, comparing structure functions and damped random walks. She finds a huge range in amplitude and time-scale of the best-fit damped random walks but it was not clear what fraction of the variance in that distribution comes from noise (uncertainties) and what comes from a true underlying variance. That is a key question I would like to work on with Hernitshek (MPIA) and Mykytyn. She finds that variability decreases strongly with intrinsic luminosity, which is of great interest to Fadely and me.

Lang finished the meeting with a beautiful, pedagogical talk about how you optimally detect point sources in multi-epoch, multi-band imaging. He defined "optimal" to mean "saturating the Cramér–Rao bound. I love that! The results are beautiful, simple, and useful. Paper soon, I hope.


all projects move one tiny step forward

I worked on code to make figures to go in the captions of the Sloan Atlas. I discussed reincarnating The Thresher with Patel and Federica Bianco (NYU), who wants to use it on her Lucky Imaging data. Vakili and I asked Foreman-Mackey to teach Vakili how to access and use SDSS Stripe 82 data on stars, to build a prior PDF for the point-spread function. I wrote text on my ideas for improving aperture photometry. I got so engrossed in the latter writing that I missed my subway stop on my way up to Columbia and ended up having to walk 17 blocks back downtown for #NYCastroML.


Stripe 82, day one

Today was the first day of the Stripe 82 meeting in Princeton. I went down for most of the day. Many great things happened at the meeting; this is just a spotty scan of things that I saw in the time I was around:

Richards (Drexel) gave an impressive overview of all the data sets that overlap the SDSS Stripe 82 area. There are four or five significant data sets in every general bandpass, like x-ray or radio or mid-infrared. It is pretty impressive. Richards emphasized the need for ways to represent and visualize the mutual coverage and the relationships among bands. Astronomers have not solved this problem yet.

Hernitshek (MPIA) showed that she can reverberation-map a quasar using just Stripe 82 photometry and a single SDSS spectrum. She gets black-hole masses that track but are larger than the standard (Kaspi 2000) spectroscopic values.

McGreer (Arizona) showed nice results on the redshift-five quasar luminosity function, leveraging spectroscopy with imaging. Richards shouted from the audience that the (substantial) differences between McGreer et al and Richards et al is resolved simply: Just ignore the older Richards et al results!

Tucker (FNAL) made an appeal to figure out photometric calibration issues in the Stripe. It is effectively the calibration standard for many surveys, so we better get it right. There are some issues in comparison with DES and PanSTARRS.

Sako (Penn) gave a beautiful talk about SDSS-II supernovae. He emphasized that the first year of any time-domain survey is a year filled with false alarms and prodigious alerts, most of which turn out to be uninteresting kinds of variable stars. (Or perhaps interesting, depending on your perspective.) If you want to have clean, manageable samples of supernovae, you need to get your variable catalogs under control. That takes time! This has implications for DES and LSST.

Gronwall (PSU) reviewed the design and plans for HETDEX. This project is insane! And I mean that in the best possible way. They are going to do so much science that is not on their short list of primary survey objectives. Can't wait.



I spent the day in Toronto, where I gave the Astronomy Colloquium. Ray Carlberg mentioned (perhaps critically) that my talk was a bit more technical than usual for this venue. Oh well! In the morning I spoke with David Law (Dunlap) about how the MANGA project intends to go from spectral pixel data to imaging spectroscopic intensity voxel. They plan to go through a "spectral extraction" step which is very sensible and practical, but probably at least somewhat lossy. We discussed what could be better.

In the afternoon I spoke with Peter Martin (CITA) and Amir Hajian (CITA), both of whom are working with multi-wavelength data of very varied angular resolution, although in very different contexts. I argued for forward modeling (duh!) for both, but also sparse (compressed sensing) dictionary methods for Hajian, whose CMB-contaminating point sources are very distinct in support and spectrum from the CMB.

In the evening, I had dinner with the graduate students. We talked about realism, and the possible reactions to the issue that all models are (or must be) wrong.


expand your sample or go deeper?

Bolton, Lang, and many others showed up today for a two-day target-selection hack-fest at NYU, hosted by Jeremy Tinker (NYU). I spent as much time as I could distracting them away from their goals. In particular, I got Bolton to recap his very nice results on whether you should expand a sample to more objects or go deeper on the objects you already have. His basic result (quantitative result, with not very many assumptions) is that astronomers tend to go too deep, when they should be expanding their samples. At least when you have a well-defined quantity that you want to measure (the population mean and variance of quantity X, for example), you usually do better by getting more minimal data on a larger sample of objects. It is not the way to discover or explore, but it is the way to measure. I encouraged him to publish this. It is relevant to many things we work on.


Patrick Huggins

Today I spent much of the morning writing remarks for the memorial that took place this afternoon for Patrick Huggins (NYU), who died at the beginning of this year after a challenging illness. Although what I said in the end had almost nothing to do with what I wrote, here's what I wrote. Huggins was enormously important to me as a mentor, as a colleague, and as a friend. He is and will be very sorely missed, by me and so many others.


why bayes?

At the reading group #NYCastroML, I said a few words about chapter 5 of the book, in which Bayesian approaches are introduced. I emphasized the dimensional way of looking at probability density, which is the approach to debugging probability expressions I advocate here. I also said that the main reason to be a Bayesian is that it gives you the ability to marginalize away nuisance parameters. In most other respects, Bayes doesn't give you that much capability, and many people who think "Bayes!" just want an estimator in the end anyway. As I occasionally say here, my view is that when you present your observational results, they should not be in the form of posterior pdfs, they should be in the form of likelihood functions, possibly with the nuisance parameters marginalized out.

After the reading group I discussed with Price-Whelan and Johnston the scope and content of the nearly-finished stream-fitting paper. Very late in the day I spoke with Alekh Agarwal (Microsoft Research) about generalizing distributed computing methods beyond map–reduce (and the like). He was pessimistic that anything structurally different would gain more than it cost.


quantum gravity, causality, planets per star

Two seminars today: Gia Dvali (NYU and LMU) gave the lunchtime brown-bag talk, about the possible interaction of quantum gravity with the axion and neutrino sectors. The idea is that the "strong CP problem" implies that some invariant of the gluon field is so small that the interaction between gluons and the graviton field could actually dominate or be dynamically significant. It was only moderately incomprehensible!

Late in the day, Jennifer Hill (NYU) spoke about causal inference in the social sciences. She talked about the Rubin formulation of causality, and the problems of having observational (rather than randomized) data. She argued the point that almost all important questions in science are causal questions, so we have to face this! Her methods involve fitting incredibly flexible and complex models to the parts of the problem she doesn't care about to reveal the residual correlations that she does care about.

Early in the day I spoke with Tim Morton (Princeton) about inferring planets-per-star rate statistics from data. The methods in the literature seem highly biased (and naive, as they involve "transforming" rather than modeling the data). We talked about next steps.


continuity of exoplanets

In a day of talks—I had to leave early a talk by Ruth Angus (Oxford) on stellar ages from Kepler to see a talk in Computer Science by Alekh Agarwal (Microsoft) on distributed and clever machine-learning algorithms and engineering—Bekki Dawson (Berkeley) showed us results on the statistics of exoplanet populations and tests of planetary migration scenarios. She showed that the continuity of tidal circularization models (conservative exoplanet flow, in some sense) makes a prediction for the distribution of planets in the period–eccentricity plane, and that the prediction is falsified strongly by Kepler. There is not yet any good model for the formation and migration of exoplanets that explains the main features of the data, but there are many possible effects, and it is possible that all of them are acting at some level. Her talk suggested scores of other projects that could and should be done. On a side note, she showed convincingly that you can measure eccentricities just with Kepler data alone, and that there are strong asymmetries that make it much more likely that you will see a faster-than-circular transit than a slower-than-circular transit when the transiting-planet orbit is eccentric. She also showed some transit timing work by our own Foreman-Mackey.


seeing the whole Universe in a single galaxy

In the morning, Andrea Maccio (MPIA) gave a nice talk about making star-formation and AGN feedback in numerical simulations of galaxy formation much more realistic. He is very negative about the possibility that AGN can stop star formation: AGN emit jets, which punch through the central part of the ISM and don't really heat all of the necessary volume. He also showed that the the dark-energy equation of state can affect galaxy evolution, not because the dark energy has any direct effect on how structures form, but because it changes the timing of gravitational collapse and star-formation episodes. That got the audience all in a tizzy: Can we infer the dark-energy equation of state from the radial distribution of stars in a galaxy? The answer is no: It looks like this effect is strongly degenerate with feedback parameters, but it is super-intriguing.

Late in the day, Foreman-Mackey and I checked in with Dawson about the in-transit noise. She has some systems that show a very strong effect of higher noise in transit than out. We suggested improvements to the statistical tests, and Dawson will try to move to smaller and smaller planets tomorrow.


excess noise in-transit?

Bekki Dawson (Berkeley) showed up for a few days, in which she will work with Foreman-Mackey and Angus and me and also give our astro seminar. We discussed her observation (not new) that the photometric noise in a star light curve is often higher in transit than out of transit. This is explained by there being strong surface features on the star that get selectively blocked by the planet but affect the out-of-transit lightcurve only in an integrated way. Dawson's concern is that this excess noise is not being incorporated in completeness and sensitivity tests with the Kepler data; that is, we might be being over-optimistic about our small-planet samples. We made a plan to test this, and, if necessary, make more realistic artificial planet injections for better completeness and sensitivity studies.


data hacking

This morning was a "hack session" in the Josh-Peek-organized reading group on astrostatistics up at Columbia called #NYCastroML. I helped Kelle Cruz (CUNY) and others build a mixture-of-Gaussians model of the WISE point-source catalog. Well, we just worked on single-Gaussian models, but we are getting ready to do multiple Gaussians. We got a long way, although, as usual, most of the session was really "data munging" and not "data analysis". That's not uncommon: Getting the data into a consistent, useful, checked state is often the hardest part of the project. And, as far as I know, there is no "theory" of this part of data analysis. It just is.