Densely targeting emission line dominated objects with NIRSPEC MSA

Notes on an obvious/crazy idea that has fermented in various people's brains,
and recently re-emerged from conversations involving Marijn Franx,
Daniel Eisenstein & HWR; these thoughts were floated at the
May 2016 NIRCAM-NIRSPEC meeting, with basically positive reception to the group there.


Conjecture: for emission-line dominated objects it is possible/sensible to open
many more shutters than the conservative "no overlapping spectra" targeting might suggest.
[For the time being, this pertains to NIRSPEC R=1.000 or higher resolution.]

Starting facts/assumption:
-- the spectra of most faint, high-z (z>~5-6) galaxies are emission line dominated.
-- the potentially most interesting ("PopIII") galaxies are emission line dominated
-- the # of targets that have potentially detectable emission lines in 10^(4-5) sec NIRSPEC
   exposures is far larger than the slit real estate budget, assuming no-spectral-overlap.
   [galaxies with broad-band magnitudes ~30 may well have strong, hence detectable lines...]
-- for emission line dominated spectra of (very) faint object, only a tiny portion of the spectral
   range therefore contains "significant" pixel.

Consequences:
-- the vast majority of detector pixels contains no "interesting" signal
-- many faint high-z emission line galaxy candidates will go un-targetted in
   any one deep NIRSPEC MSA setting.

Proposed remedy (to be taken as a though experiment, first):
open ALL shutters at focal plain locations that plausibly (according to NIRCAM photometry)
have high-z, presumably strong emission line targets.
[spectral overlap and chip gaps be damned for now.]
Let's take N~5 as a mental strwaman-plan.

[Nomenclature: dispersion runs along 'columns', slit runs along 'rows']
Let's presume that means we would have N shutters open in any one column (on N galaxies).
Advantages: N times more targets
Disadvantages:
  a)  N x higher sky background
  b)  "confusion of N overlapping spectra"

Addressing the disadvantages:
 on  a) How does the monochromatic surface brightness of emission lines (say, 10Msun/yr at z~6)
  compare to the background? I.e. are the cores of strong emission lines above/below the background.
  NB 1: compared to 'slitless' spectroscopy the background
  is still N/365 lower than slitless. [365 == # of shutters in dispersion direction]
  NB 2: the ensemble of open slits will inform us about the background

 on b) lines are narrow and sparse; if the continuum is negligible; the spectral signatures
 don't overlap. In principle there is some wavelength degeneracy
 (which line came through which slit); but this should be manageable, as long as there are
 no shutters open in the same column and adjacent/nearby rows.
on b): information on the continuum will be severely degraded;
 --> for emission-line dominated objects, there is little information in the continuum anyway;
 photometry will help

One specific approach is to target all objects that were done on R=100 mode, in R=1000, with disregard to overlap problems. The R=100 mode should break many of the degeneracy issues.
[Thanks to Chris W. for suggesting this.]

Basis of this: could one get Brant's and Christina's mock -data catalog, including their photo-z estimates?

Target assignment priorities for JWST MSA (NIRSPEC) Part I

after discussion in Victoria, May 2016, some notes on my thoughts
on the slit assignment from NIRSPEC MSA:

Let's presume, the plate solution is perfectly known and there are no failed shutters,
and we are interested in a 3-dither (no nods) mask design.

Let's presume we have a set of targets, with foremost attributes:
alpha,delta, size,flux,"scientific value", wavelength-range of highest interest (presuming we know z).

Let's presume, we want no spectral overlap; and we can neglect the issue of spectra
falling off the chip

The question is: what is the best combination of
a) telescope pointing (field center & orient)
b) target list

For a) we have in practice we have only 2 DOF, presuming the orient is given.
presumably these are fixed by the positions of very few "high-value" targets.
So a) and b) basically decouple.

Conjecture: in the limit that spectra cover the entire detector, the matter is simple:
in each slit position, one finds the "best available" object.
Complications foremost arise in defining which object is "best":
we need to define a merit function between
-- how intrinsically valuable is the target
-- how off-center (w.r.t. the micro shutter) should its centroid be;
   let's quantify this by a single number  log(S/N)-log(S/N_best),
  where S/N_best is defined as the S/N (given t_exp) we could get for
  a perfectly centered source with best sky subtraction.

Things to pre-compute:

Jan Rybitzki's Chempy projects

Just to commit to memory here is a discussion draft of Jan's papers
[from HWR - JR conversations; and DWH input]

Paper II or I :
  -- write-up of the basic model chempy (with thesis advisor Andreas Just).
  -- science bit: given a set of yields, how much can the abundances of
      a single star constrain the chempy parameters: the SFR, high-mass IMF slope,
      the SN-delay, the feed-back-mass-loading, the fraction of WD's that go SN Ia, and
     the gas inflow rate.
    [corollary: is having the age of a star helpful, if the star is not very old]
 -- implementation: take the Sun and Arcturus abundances and ages, and
    construct a chempy parameter pdf triangle plot.
-- consider taking the 'cosmic abundance' instead of Arcturus
-- consider different yield tables
-- discussion:
    explain why this fails..


Paper I or II:
   -- are the APOGEE data good enough to tell us which (published) yields tables are "best"
  -- Melissa will canonize the Hawkins et al accurate APOGEE abundances/ages to the
     RC sample of APOGEE; we then presume that abundance zero-point systematics
     are a sub-dominant error source
  -- Jan will try all 9 yield table (3x) combination to match the APOGEE RC sample
     (effectively marginalizing over the chempy params) and ask which fits best --> make Hogg happy
    [How much of this could be in paper I]
CHANGE of scope: use the ~30 abundance standards of Jofre et al 2016, to fix the field tables...
  that stays Paper III

Aside: can we ask what the set of [X/H] zero-point shifts in APOGEE can be, that would make
Arcturus, the Sun, and the cosmic standard likely??


Paper III:
   -- goal: the (varied?) chemical prehistories of all stars in the APOGEE sample
  -- use ensemble fit to tweak yield tables (see Paper II); we then assume both the
      yields and the abundance zero points to be "correct" (i.e. we won't marginalize)
  -- chempy has four parameters that may plausibly vary from star-to-star:
     the SFR, high-mass IMF slope (?), the feed-back-mass-loading, and
     the gas inflow rate.
  -- construct the pdf of these parameters for every single star in APOGEE;
     also exploit the ages for the stars we have..
  -- this enables:
      ** did the IMF vary as a function of time, of FeH?
      ** does the inferred mass-loading, or the inflow correlate with other properties
        (age, FeH, etc..)

CHANGE of scope: apply all of this to the ~30 abundance standards of Jofre et al 2016

Paper IV (Hogg)

Binaries/rotation/dredge-up ages of giants

Had a conversation with Selma de Mink:
her claim: on the MS, stars more massive than 2Mo rotate a lot
faster than power-mass stars (because of magnitude field stars).
The question is whether some of that rotation survives recognizeably
into the giant phase.

Does Apogee have a rotation parameter? If so, does that correlate with mass?

Exploring Apogee's abundances space with chemical evolution models

Starting point

Jan Rybizki arrived Feb 1, 2016 and brought with him his one-zone chemical evolution model. The overall plan is to explore what we can learn about a) the Milky Way (its chemical evolution and the yields of the stars), b) the Apogee abundances, and c) Jan's one-zone model and its limitations.
.. and then learn it. Here's a set of HWR notes after talking with Jan.

Jan overplayed his (fiducial) model predictions to the Apogee abundances as-is. It looks like in the plots below, basically a mismatch... which is an OK starting point.

Towards understanding the data-model (mis-)match

The reasons why data and model may disagree are manyfold ...disagreement is good, it teaches us something new. 

Model parameters

The most immediate advantage of Jan's models is that it can 'fit data' via MCMC,

which makes the variation of many, even all model parameters feasible. The model parameters fall into two categories:

Galaxy History: This entails the SFR, the inflow and outflow terms form the box, and the IMF; this is a 'handful of parameters'

Stellar physics: this entails essentially the yields; these are in some sense parameters, but presumably the space of all possible yields is not spanned by a set of continuous parameters. Hence, see below.

Data Calibration

It is well known that the Apogee ASCAP pipeline, at least for some elements has systematic offsets that are far in excess of the typical error bars. One simple way to explore the role of possible offsets is to make a basic abundance offset [X/Fe]_0 a fitting parameter, presumably with some prior on it (to avoid complete 'runaway').

Yields

It would be good to think about how one can turn the options on yields into something that is parameterizable, and therefore fit-able.

Applicability of a 1-zone Model

The model as is, makes a 'unique prediction' for [X/H] or [X/Fe] = f(age). Clearly, the data show a 'spread' in their abundance patterns, well beyond their errors. In a galaxy where radial migration must be prevalent, a 1-zone model cannot be correct.

In the longer run, we should explore in which regime it is a useful approximation: this could be done by restricting the observations to a limited range in [Fe/H], if [Fe/H] has been a good birth-radius predictor (for stars younger than 8Gyrs). Or we could fit a superposition of 1-zone models.

Next Steps

What portion of [X/H]-age space can be reached by varying the parameters of a one-zone model of a given yield-table?

How to go about it? Perhaps by simply defining the fill range of plausible model parameters, (galaxy evolution and IMF, not yields)  and then sampling model predictions uniformly.

It would be good to have an approach, to figure out how degenerate (or not), variations in the IMF and of the SFH are (if one has Apogee-type data).

What is the best one-zone fit to the Apogee data under the (untenable) assumption that all yields are right, and the data have no systematic offsets?

Apparently for historical reasons it is still not 'easy' in Jan's model to hold an arbitrary number of parameters (including none) fixed, fitting the rest via emcee. That piece of infrastructure should probably be put in place as one of the first steps.

What is the best one-zone fit to the Apogee data under the (untenable) assumption that all yields are right, and the data may have systematic offsets?

Apparently for historical reasons it is still not 'easy' in Jan's model to hold an arbitrary number of parameters (including none) fixed, fitting the rest via emcee. That piece of infrastructure should probably be put in place as one of the first steps.

Can we learn about the yields?

Conceptually the next step would be to have a few yield-knobs to tune, and let them loose in the data fitting.

Sample selection:

We need to decide what good sample cuts are, to make the one-zone model a sensible framework. Options are:

-- spatial cut (solar radius)

-- abundance cut (FeH==birth radius) at solar FeH +- 0.x dex

-- no cuts

Thoughts on BigData/InferenceSummerSchool/KTS

Two basic issues:
1)  where does the disconnect between the astronomers and the ML/CS folks actually occur?
2)  are the any women of any seniority on the astro/data-mining world

On 1)
HWR's suspicion is that astronomers almost always deal with data that are noisy,
 and heteroscedastic at that.
Therefore, astronomers know how to write down likelihood, go Bayes, but: and then??
CS/ML folks have amazing tools to classify, but these tools almost all fail ungracefully in the
"very noisy" regime. I.e. are not good at simply ignoring differences in object labels that can sensibly
attributed to only noise, not to any inherent difference of the objects.

What's a good definition of noisy here:
Let's presume any object has many data points (e.g. pixels in the spectrum of a star);
there is the regime where the data variance at any one pixel due to noise is comparable
to the (to be classified) ensemble variance (noiseless) in that 'pixel'. What are good ML tools
in that regime.

What to do about it
a) try to spell this out clearly, and ask Hogg, DFM, ZI, BernhardSchöllkopf etc. for insights?
b) initiate some culture-gap bridging exercise with the HITS group, to see how useful that is..


On 2)
The Tel Aviv Big Data conference (Dec. 15), the search for lecturers for an IMPRS summer school,
showed the paucity of "obvious" female scientists to serve as lecturers or tutors.
What to do?
-- find out whether this is just a consequence of HWR's ignorance? Ask Hogg, Schöllkopf, Ivezic, Bailer-Jones, DFM for names...
-- initiate discussion with Reutter to see whether there might be any interest in initiating
   an "award", amounting to a 3-year fellowship, for women in data science.
  To be taken as a post-doc, or a repeat summer fellowship.

Tel Aviv notes

Post Tel Aviv

-- M-stars WD's with Dani

-- Cannon residual into Nikos Gianniotis' machinery

-- write-up for Chao Liu's PD.