I gave the Königstuhl Colloquium on data-driven models. With preparing for that and a long lunch, no other research today!
At MPIA Galaxy coffee, there were great talks about the Milky Way disk by Gail Zasowski (JHU) and Jonathan Bird (Vanderbilt). Zasowski showed results on the kinematics of the MW disk ISM based on diffuse interstellar bands. She used the three-d dust map from Schlafly et al to figure out the mean distance to each absorber, and sees a consistent story. Bird showed that the age–metallicity relationship in the disk (briefly, that older stars have higher velocity dispersion) is a product not just of disk heating but also of disk formation, if the disks form in the cosmological context as expected. Late in the day, Bird and I (with help from Rix) formulated a way to measure the age–metallicity relationship and its dependence on Galactocentric radius via a likelihood function (and therefore Bayes). We vowed to try to do this with the APOGEE data plus stellar ages from Ness's recent work with The Cannon.
At Milky Way group meeting, Price-Whelan talked about chaotic streams, Gail Zasowski (JHU) talked about the SDSS demographics survey, and Jonathan Bird (Vanderbilt) talked about the age–velocity relationship in the Milky Way Disk. This latter project started yesterday, with Bird asking Ness and me for the ages for red-clump stars, which Bovy had previously matched up to proper motion measurements. The age–velocity relation has only been precisely measured previously using ages (as opposed to chemical abundance proxies) in the Hipparcos sample (that is, very locally at the Solar position), so if our ages are good, this will become the best measurement of the disk heating mechanisms ever.
In the afternoon, Rix and I worked with Bird to turn his empirical velocity dispersion (measured as a function of age and Galactocentric radius) method into a probabilistic method with a likelihood function. With a few severe assumptions, we got a very simple form. The plan is to try to run some MCMC tomrrow. We encouraged Bird to spend an extra week in Heidelberg if he can!
OMG I am so stoked! I made plots of the ages of the APOGEE red-clump stars as a function of position in the Milky Way and got the image below (one plot of many). You can see the thin disk jump out but you can also see a flaring of the young disk at large radius that Rix and Bovy told Ness and me we should expect.
Jonathan Bird (Vanderbilt) crashed our party and asked if he could plot the age-metallicity relation that is implied by our red-clump-star ages. We said “yes” of course! Bird has ideas about the thickness of the disk being set not just by heating or scattering but also in part by the way disks form, which should be dynamically hotter at higher redshifts.
Ness and I are getting red-giant ages from APOGEE spectroscopy using The Cannon and a training set from Kepler. We worked on making plots that would help us understand where this age information is coming from. Options include: Chromospheric activity (which decreases with time in stars as the magnetic field decays), dredge-up of C, N, O from the nucleosynthetic core (which pollutes the surface abundances over time), trace element abundances (which might indicate birth place and time beyond the information in gross indicators like metallicity and alpha-enhancement), and non-LTE effects (which might be different in different stars since convection patterns and scale are a function of mass).
Tom Herbst (MPIA) showed me his fish-eye all-sky camera and data acquisition system, and we discussed science projects for it with Markus Pössel (MPIA). The whole system is on the roof, so its computer and controller and everything are isolated from all the building systems to protect the building and its IT infrastructure from lightning strikes!
Rix and I asked Yuan-Sen Ting (Harvard) to compute derivatives with respect to elements for the stellar models being used to analyze the APOGEE data. The idea is that we want to ask how well we can linearize the models around fiducial points and then model (and therefore make measurements from) the spectra.
Today was the last (and half) day of the Ringberg meeting. At breakfast, we discussed an argument James Binney (Oxford) had made the previous day about n-body simulations not being the answer to physics questions: His view is that simulations should just inspire theories that are continuous, large-scale, and manipulable with other tools. He gave some examples. I, in the end, disagreed with Binney's argument and his examples: There are many fields of science where all of the predictions come from small-scale (n-body or molecular or finite-element) simulations and all the “understanding” is pure story-time about those simulations. For example, structural and aerospace engineering is all about finite-element analysis. And climate and weather are all about simulations. And many of the beautiful, effective theories for solid-state physics and so on are not working very well as we push to new regimes; we have to go back to molecular-level simulations. All that said, I am sympathetic to Binney's point, which is that we don't always get a lot of understanding out of the crank-turning on n-body simulations.
Today's talks were very lively, with Carlberg (Toronto) and Küpper talking about stream substructure created by dark-matter substructure, and Bonaca talking about our work on streams in realistic (rather than analytic) potentials. Bonaca shows that our conlusions from analytic potential models can be very wrong, even in gross properties, but that things improve as we include more and more streams into the analysis. She got some lively questioning from Evans, who was not buying it. I'm with Bonaca, of course!
The drive home was a long discussion between Rix and me about all that had happened and what to do next. We are fired up about new approaches to stream searching, to use now on SDSS and PanSTARRS data, but also to have very ready for (what we are calling) Gaia DR2.
[As a result of crippling disorganization, blog posting has gone pear-shaped. I will be posting out-of-date posts with correct dates (and 23:59 time stamps), but possibly considerably after the fact and out of chronological order.]
Another great day at Ringberg Castle today. For me, the highlight was an unconference session called Gaia Zero-Day Exploit, which was about what we could do immediately upon release of Gaia data in its first and second releases. So many good ideas came up, and most of them are summarized in incredibly telegraphic form on the crowd-sourced meeting minutes (search down the document for "Shovel-ready"). Gaia DR2 is much more interesting than DR1 for Galactic dynamics. Some great ideas included: Find the alleged fountain of hyper-velocity stars coming from the Galactic Center and use them to infer the shape of the MW halo. And: Look at the vertical velocities in the disk and see the pattern speed of disk warping and precession (see mention below of Widrow). And: Look for direct evidence of the process of dynamical friction—that is, look for the stellar wakes. And, simply: Get the proper motions of all Milky Way companions and streams. The list is long. After the session, Rix and I felt like we should find some way to get some of these fleshed out and quasi-published in some community forum.
David Martinez-Delgado (Heidelberg) kicked off the day with his absolutely magnificent images of galaxies, in which he uses small telescopes and long exposure times to get amazingly sensitive images of tidal features. He compared similar images from his small telescopes to those taken with huge telescopes, and showed that smaller is better. There was some discussion afterwards of "why". Some of it is that his calibration requirements are high, and he spent more time on his own instruments (he may be the best flat-fielder in all the land). But some of it is that big telescopes have big cameras with many optical surfaces, which make it hard to go for large dynamic range in intensity.
Larry Widrow (Queen's) talked about galactic plane distortions and coined the term "galactoseismology".
After watching various talks in which streams were simulated, Price-Whelan and I noticed that for the first few orbits (even in chaotic potentials), the stream stars make a "bowtie" shape when very near apocenter. And the stars spend a lot of time out there. So perhaps we should be looking for these: There might be many more bowties than streams! We discussed.
Today was the half-day at the meeting (hike in the afternoon). The highlight for me was the talk by Carl Grillmair (IPAC), who talked about finding and tracing streams, observationally. Grillmair is the discoverer or co-discoverer of more than half of all the known Milky-Way streams. He showed us some of the visualizations that help him find them. He said that we should replace his by-eye work with automated methods, but (to my knowledge) no-one is close right now.
After Grillmair, Sesar (MPIA) showed a beautiful analysis of the new Ophiuchus stream, where he measures everything via probabilistic inference. Kallivayalil (Virginia) showed the evidence for the LMC proper motion measurement. I am a believer! She then showed noisier data from other Local-Group dwarfs and streams. Slater (Michigan) showed properties of the Milky Way's stellar halo at large radius. At large radius, the halo is consistent with being entirely disrupted satellites.
Today was a great day at the meeting; many good and lively talks and discussions. This is a healthy field! Here are some personal highlights for the day:
In his opening talk, Wyn Evans (Cambridge) gave credit to Koposov, Hogg, & Rix for fitting the GD-1 stream with a (trivial, wrong) orbit model (that is, not a physically correct stream model): The orbit model seems to return the correct potential. The authors of the paper refused this credit, since it appears to be luck rather than skill that obtained this coincidence. A few talks later, Bovy disagreed with Evans; he showed that as the potential is made more realistic, the stream model departs more from the orbit model.
Sanders (Cambridge), Bovy, Binney (Oxford), and McMillan (Lund) all talked about using actions and angles for modeling streams. The math is beautiful, but the computation is expensive. It is not clear to me that this is the right approach, long term. It also requires a strong constraint on the potential and the populated orbits: Many orbits are likely to be chaotic!
Continuing on that theme, Pearson (Columbia) and Price-Whelan coined the term "stream fanning" for streams on chaotic orbits, which show very different morphologies from those on regular orbits. Price-Whelan showed that the fanning happens earlier and more prominently than you would expect from standard chaos measures like Lyapunov time; he presented an argument in terms of frequency-space diffusion. In the question period it came up that if we take the long streams to be highlights of regular orbits, then this puts some kind of very complex (or even fractal) posterior on potential models. Crazy!
In the discussion between sessions, many great ideas came up. One is to use The Cannon to obtain more precise spectroscopy-informed stellar absolute magnitudes for red-giant stars. Another is to find the fountain of high-velocity stars above the bulge in Gaia DR2; we might be able to find A and F stars (most known high-velocity stars are B stars, I think).
I am at the Stellar Streams in the Local Universe meeting at Ringberg Castle. The meeting is small and workshop-like, so we are trying various experiments. One is to keep a running, completely public web document where everyone is encouraged to take notes on the talks and sessions. The first day was extremely lively. Here are some unfair and impressionistic thoughts:
Bovy opened the meeting, with his summary of the scientific subject of stellar streams in the Milky Way Halo, which range from dynamical modeling of the Milky Way to inferring the dark-matter and star-formation properties of low-mass galaxies.
I always think of the dynamical modeling, and rarely the galaxy evolution, so it was great to have Kathryn Johnston back up Bovy with the point that while we think of the Milky Way as being its own unique and bespoke object, there are enough collapsed, star-hosting overdensities disrupted in our halo that they probably constitute a relatively fair sample of small objects in the Universe. That is, we can do cosmology in our own halo (duh!). She also showed that the stream-to-shell ratio can be a cosmological test and speak to these questions, and she included a quantitative definition of "shell" and "stream"!
In the question period after a talk by Andrew Cooper (MPA) about simulations, he was asked whether in the outer halo the distribution of stars would follow the distribution of dark matter. This is a great question and idea, since both the stars and the dark-matter particles in those outer reaches are just test particles coming in. He said that they don't have an answer yet, but it looks likely that the stellar distribution is only a weakly biased version of the dark-matter distribution. That would be awesome.
Brendan Griffen (MIT) showed the Caterpillar suite of cosmological simulations of Milky-Way-like galaxies. This project is very simple, very well conceived, and producing very valuable tools.
Mike O'Neil and I worked on the real-space covariance function for the cosmic microwave background radiation. He summed the Legendre polynomials to convert the C_ell spectrum into an angle-space kernel or covariance matrix. He plotted the real-space covariance and it is very pretty; it shows the baryon acoustic feature and oscillates at huge angle. He then looked at some symmetries of the matrix: It can be expressed as a sparse sum of outer products. Because of this, it is possible that he can develop a fast method for matrix factorization, linear-algebra solve operations, and (log) determinant evaluations.
At Galaxy Coffee, Jonathan Stern (MPIA) showed an analysis of quasar momentum-driven winds in the context of AGN feedback for galaxy evolution. He finds that the quasar radiation is not sufficient to drive the wind. In the discussion of the point Joachim Bestenlehner (MPIA) pointed out that Wolf–Rayet stars can have momentum-driven winds that exceed the naive radiation pressure L/c by a factor of five-ish. It relates to the fact that there are multiple scatterings of the radiation. Andrea Macció's (MPIA but soon NYU-AD) showed some nice work on simulations of small galaxies that are infalling in to the halos of larger galaxies.
Mike O'Neil came down to HD for two days (from Frankfurt) and we spent part of the afternoon understanding how to run and interpret PICO, which computes CMB power spectra. We started towards converting its output (which is in "ell" space) into a real-space covariance function. When we left work, O'Neil was looking at recurrence relations for Legendre polynomials!
We took advantage of O'Neil's presence to have a discussion with various MPIAers about interpolating grids of models (so that you can have a model defined everywhere in parameter space, but only compute it in full at a sparse set of points). He had various extremely useful points of advice. However, his reaction to the problem of interpolating 1-D LTE spectral models was not “do this kind of interpolation” but rather “I bet we can make the models run much, much, much faster!” So we downloaded the original Kurucz model paper [warning: huge PDF file] and began exegesis.
At Milky Way group meeting, various projects were discussed. Marie Martig (MPIA) and Maria Bergemann (MPIA) are both looking at how we might extract stellar ages from stellar spectra. Martig is looking at the dredge-up of elements from the core into the stellar exterior. Bergemann is looking at how extremely weak chromospheric emission might adjust (minutely) the shapes of the Balmer lines. Both of these projects are directly relevant to what Ness and I are doing with The Cannon.
After lunch, the PSF coffee had two impressive presentations. In the first, Nestor Espinoza (PUC Santiago/Chile) showed very impressive transmission spectroscopy of transiting planets. He has some spectral features that are hard to explain, but he notes that they might be caused by some systematic problems with the pipelines. They are taking reasonably high resolution (few thousand) spectra, but binning them down for analysis; it is interesting to think about what could be done with non-trivial binnings of the spectra, that are customized to particular questions.
In the second presentation, Florian Rodler (MPIA) took time-domain spectra in the red side of the K band to find the forest of molecular CO lines in the inner planet in the Upsilon Andromeda multi-planet system. This system is amazingly complicated, with two super-Jupiters with huge mutual inclination, large eccentricity, and a (circular) hot Jupiter. It sure doesn't look like it could be stable! The system properties are known from a combination of radial-velocity monitoring (of the main star), astrometric monitoring (of the center of light), and direct detection (of the molecular lines from the hot Jupiter).
I spent the day debugging with various students. Eddie Schlafly (MPIA) and I helped Nina Hernitschek (MPIA) debug her Gaussian-process (plotting) code, I helped Kopytova debug her spectral fitting code, and Rix and I helped Anna Ho debug her LAMOST models. On the latter, I have a hypothesis that perhaps bad data (strange, corrupted, or physically odd stars) are distorting the data-driven model of The Cannon. In particular, if there are outliers, we will get bad results for the intrinsic scatter parameters in the model. This all came up because the goodness-of-fit measurements are coming in “too good”.
[Very few posts for the last few days, because I was on a mini-vacation and getting ready for the annual move to Heidelberg.]
My research day started in the garden with Rix, looking at Anna Ho's paper using The Cannon to transfer APOGEE-like stellar parameter labels onto stars with LAMOST spectra. We worked through the figures and made comments.
At the weekly PanSTARRS meeting, Nina Hernitschek (MPIA) showed beautiful samples of QSO candidates and RR Lyrae candidates found with time-domain analysis. Her method involves Gaussian-process modeling multi-band time-domain data, and then a supervised classification. The results are very convincing: For example, on a map of the sky with the MW dwarf galaxies and star clusters circled, there are clear overdensities of RRL stars in many of the circles. Also, the RRL stars clearly trace the MW halo, bulge, and disk. She also has incredible 3π maps of QSOs, which are amazingly isotropic near the Galactic poles, but also have many QSOs in (or near) the disk plane and even behind the Galactic center. I bet she has the best low-latitude QSO sample ever.
At lunch-time I had a conversation with Laura Inno (MPIA), Rix, Eddie Schlafly (MPIA), and others about Cepheid stars in the Milky Way disk. There should be tens of thousands but hundreds are known. Rix gave some very good reasons for finding them all: Cepheids are very young stars but they are cool (5000 K) not hot, so you can measure their abundances. They are also luminous, so you can see them at large distance and through the dust. We discussed how to find them with PanSTARRS and WISE data, perhaps starting from what Hernitschek has already done.
Today I got an unexpected treat when Melissa Ness showed me some new GALAH data from the HERMES instrument. The project is intending to take R 28000 spectra of about a million stars, and determine up to 30 element abundances, or something insane like that. Ness and I are looking at whether The Cannon could be used for basic parameter estimation and to assist in the long-run goals of the project. The data look absolutely beautiful and the first attempts with the data make it look like The Cannon will Just Work (tm). We are a bit concerned about continuum normalization, but aren't we always?
I spent the day at Extreme Precision Radial Velocities at Yale. It is a great meeting, because it is very focused on the instrumentation and code that underly radial-velocity planet search and characterization. Today was a stats-heavy day, with me, Eric Ford (PSU), and Tom Loredo (Cornell) leading off with pedagogical talks. I gave an entirely new (for me) talk about noise modeling, and it was followed by absolutely excellent questions (every question pointed out a talk slide I should have made). Loredo made a nice point, which is that statistics is not a method or tool, it is a language or framework for communicating about quantitative questions. I couldn't agree more!
At lunch, Ana Bonaca organized a gathering of probabilistic reasoners to discuss asteroseismology with Sarbani Basu (Yale). This gave us an opportunity to feel out some of the issues if we try to build a probabilistic model (a forward model of the time-domain data) to replace the standard practice of Fourier transformations (or periodograms or the like). That was productive and useful.
In the afternoon, one talk that particularly stood out was by Xavier Dumusque (CfA) about The Keplerian Fitting Challenge. He made fake radial-velocity data, filled with difficult but realistic noise sources, and challenged groups to find and characterize the injected signals. He did a great job describing the successes and failures of the different groups, and even awarded nice bottles of wine to the two top-performing teams. This project, like the GREAT projects for weak lensing, are important community-building and critical-review projects for difficult data-analysis challenges.
I spent my research time today preparing my slides for the Extreme Precision Radial Velocity meeting at Yale. I am talking about noise models, arguing for creating and taking advantage of great flexibility, and then controlling complexity with priors or regularization or hierarchical inference. I also want to give some ideas about how, technically, we do what we do.
I wrote text in the “Methods and data” section of the mass-and-age paper I am writing with Ness. I emphasized particularly the point that The Cannon is a probabilistic model: It is a likelihood function, which is optimized at training time, and then again at test time. The only difference between training and test is which parameters are varied. In the former, it is the spectral expectation and variance parameters, at fixed label values. In the latter, it is the label values, at fixed spectral expectation and variance parameter values. The cool thing about this likelihood formulation is that it makes it trivial to account for heteroskedastic noise variances and missing data (in both the training data and the test data).
It is the summer, and it feels like I am getting lots done! That said, I only got in a small amount of serious research time today. I spent most of it commenting on a draft manuscript from Price-Whelan on chaos in the halo of the Milky Way. Yes, dynamical chaos. This is the coming to fruition of a pretty old idea at CampHogg: What is the difference between stellar streams on regular orbits and those on chaotic orbits? It turns out that the differences are bigger than we expected!