Why are there so many starforming galaxies in the quench sample?

by mlpeck

This question has yet to be answered satisfactorily, and I think it's one question that must be addressed. I have a partial answer that may take several posts to explain. This one will just show the scope of the problem.

By my count there are 541 (47% of the total) quench sample (subset 2) galaxies in the pure starforming part of the BPT diagram, vs. 484 in the control sample. The quench sample star formers fall squarely on the local "star forming main sequence" as seen in the following graph where I've plotted the SFR-M* relationship for the QS overlaid on contours of the relationship for all control sample galaxies. The straight line is an estimate of the local relationship from Elbaz et al. (2007). All plotted values are from the DR8+ version of the MPA pipeline.

I've only found one significant difference between quench star forming galaxies and controls: their broad band colors are redder, and that's because they are dustier as measured by Balmer decrement based color excesses. The graph below compares the CDFs of quench and control starforming galaxies:

The median color excess for QS starformers is E(B-V) ≈ 0.43 ± 0.02 mag. vs. 0.32 ± 0.01 for controls. Not a huge difference, but statistically significant at least.

Posted March 11, 2014 2:36 PM
by mlpeck

Here is a partial explanation of what I think we are seeing. In a study of distant galaxy clusters Dressler et al. (1999) devised a classification scheme for the spectroscopic properties of galaxies that was based primarily on the line strengths of Hδ and [O II] 3727-3729. I've copied and pasted a schematic diagram of the system from their paper:

Their k+a (and a+k) class is familiar: these are just galaxies with strong Hδ absorption and weak or absent emission features, which are generally agreed to be post starburst systems. It's their e(a) class that I want to focus on. These are galaxies with strong Hδ absorption along with emission lines.

The most popular interpretation of e(a) galaxies is that they are dusty starbursts with the OB stars that ionize the gas hidden by dust while a slightly older population has migrated away from their "birth cocoons," thus producing the strong Balmer absorption features (Poggianti & Wu 2000, Balogh et al. 2005, Poggianti et al. 2009).

It apparently is still an open question whether e(a) galaxies are the precursors of PSGs, but Balogh+ concluded based on near IR imaging and photometry that most e(a) galaxies cannot evolve into k+a's.

So how many e(a) galaxies are in the quench sample? This depends sensitively on the thresholds chosen for strong Hδ and weak emission, but here is the breakdown by BPT class with a threshold of 3Å EW for H&delta_A and -5Å EW for both Hα and [O II] as the dividing line for weak emission:
```
              bpt.q
kac.q          no em.  SF Comp. AGN LINER Weak em.
  e(a)              0 386   141  42    31       25
  e(a) weak em      0  12    17   6    13       10
  k+a               3   0    14  11    45       50
  other             2 143    75  23    33       67
```
So, over 54% of our sample are e(a) galaxies of which 61% are starforming (as indicated by line ratios), and conversely over 71% of the starforming galaxies are e(a) galaxies. Only a minority (basically the 123 k+a systems) would pass traditional filters for PSGs.

So this is a bit of a problem for this method of sample selection: it is apparently insensitive to the presence of a hidden population of hot stars. On the other hand it may be a strength of this method that it doesn't discriminate against the presence of emission lines. In a detailed study of the post starburst galaxy NGC 1266 Alatalo et al. (2014) concluded

Galaxies with K+A spectra and shock-like ionized gas line ratios may
comprise an important, overlooked segment of the post-starburst
population, containing exactly those objects in which the AGN is
actively expelling the star-forming material.

If there's an actual discovery of an overlooked class of quenched galaxies in the quench sample it most likely lies among those with non-starforming emission line properties.

Posted March 12, 2014 3:41 PM
by JeanTate in response to mlpeck's comment.

Very cool thread, mlpeck! 😄

By my count there are 541 (47% of the total) quench sample (subset 2) galaxies in the pure starforming part of the BPT diagram, vs. 484 in the control sample.

Just to confirm, these are from "the 1149" (for QS), and a subset of 1149 QC objects from "the 1196" which have redshifts between 0.02 and 0.10 AND which have estimated z-band absolute mags brighter than -21.0, right?

In trying to identify potentially problematic objects, I've found two kinds which may produce otherwise unrecognized systematic biases:
1. for an object to be placed on the BPT diagram we are using, its SDSS spectrum needs to reliably record flux in two small wavelength windows which include the two pairs of redshifted (emission) lines. If, for any reason, one or both windows contains no - or incomplete - SDSS spectral data, that object cannot be plotted (or not, if any lines are not in emission, or do not have sufficient S/N). As in the case of AGS00002sv (see below). The effect of these 'bad spectra' objects on your analyses is hard to judge without a robust estimate of the numbers of such objects; I hope to have such an estimate by early next week.
2. Objects whose integrated fluxes in those two small windows is dominated by noise will not be placed on our BPT diagrams, as the S/N of the emission lines will be too low (extreme 'Green Pea' like [OIII] objects excluded; there are no such GPs in either QS or QC catalogs!). If the distribution of 'small windows fluxes' is sufficiently similar - QS vs QC - a systematic effect produced by this bias would wash out. However, I suspect it's not; in particular, QS objects are far more concentrated* than their QC counterparts, so the fraction with 'noisy spectra in those two windows' is likely to be considerably greater among QC objects than among QS ones ... which then introduces a systematic bias. Does such a bias actually exist? I don't know (yet).
Here's the DR9 PNG summary spectrum of AGS00002sv:

And here's AGS00004k0, an example of a QC object with a very noisy spectrum, and which has a shallow integrated radial surface brightness profile (as judged by eye; there are fairly easily obtained quantitative measures):

*in the sense that the integrated radial surface brightness falls faster

Posted March 13, 2014 3:22 PM
by mlpeck in response to JeanTate's comment.

Just to confirm, these are from "the 1149" (for QS), and a subset of
1149 QC objects from "the 1196" which have redshifts between 0.02 and
0.10 AND which have estimated z-band absolute mags brighter than -21.0, right?

Yes.

The effect of these 'bad spectra' objects on your analyses is hard to
judge without a robust estimate of the numbers of such objects

Well, the exact number of starforming objects doesn't matter. There are a lot, and that needs explaining. If we knew what was in those missing bits of spectra we could only add to the number of starforming objects. If, contrary to reality, we had higher S/N spectra objects might shift around in the BPT diagram but there's no particular reason to expect the distributions to change significantly.

The control sample hardly matters at all in this discussion.

I mentioned this in a footnote elsewhere: I decided Brinchmann's "Low S/N AGN and SF" groups weren't adding much information. For one thing there's only one "Low S/N AGN" in the entire quench+control samples, and decreasing the detection threshold to 2σ for "Low S/N SF" seems inconsistent. So, to replace that group I added a classification for galaxies with one or more missing or undetected BPT diagnostic lines but with a 3σ detection in any one of [O II], [O III] 5007, Hα, [N II] 6584. I called this class "Weak emission", but it will include spectra where some lines are masked -- usually when this happens the masked region is around Hα. Perhaps I should have called this group "some emission" instead.

With that addition there are 5 unclassifiable galaxies in the quench sample of 1149 and 152 "Weak emission". I've checked the 5 -- 2 of them have obvious emission lines but entries of 0 for all fluxes in galSpecLine, 1 has Hα masked but weak or absent emission lines elsewhere, and 2 more have no missing data and no apparent emission and so are truly passively evolving.

Just for fun I queried CasJobs for objects meeting traditional K+A criteria: Hδ_A > 5 Å, Hα > -5, [O II] > -5 (note a negative equivalent width indicates emission in the MPA pipeline). The query returned 1612 objects at all redshifts (so far I've found 2 erroneous redshifts). Out of those 582 have 0.02 ≤ z ≤ 0.10 -- I haven't done any magnitude cuts yet. Here is a breakdown of BPT class for this sample:
```
  no em.       SF    Comp.      AGN    LINER Weak em. 
      25       14       65       28      127      323 
```
This distribution looks more like the k+a subset a couple posts up. Only about 2% of the sample is in the starforming part of the BPT diagram, and only 40% is classifiable at all with over half of those LINERs. This is more or less expected since I specifically selected against strong Hα & [O II] emission.

Posted March 13, 2014 11:09 PM
by mlpeck

I checked the 152 spectra in the "Weak emission" class and by my count there were 6 with the area around Hα masked. I didn't watch Hβ.

This is kind of neat. I created stacked spectra for the quench sample by adding the individual spectra weighted by the inverse variances of the fluxes, and separated by BPT class. These were smoothed to 200 km/sec and are scaled arbitrarily with values offset for each group:

You can see a clear progression of properties going from the starforming group at bottom to "Weak emission" at top. To me the weak emission group looks like the very low excitation continuation of LINERs.

Posted March 14, 2014 9:19 PM
by JeanTate in response to mlpeck's comment.

If we knew what was in those missing bits of spectra we could only add to the number of starforming objects.

True, but if there were far more 'missed SF objects' among "the 1196 QC" than among the QS counterparts ... Not that this is likely; I estimate the total number - of objects with 'bad spectra pixels' affecting at least one of the four redshifted lines - at ~1±0.5 (in dex units; i.e. more than 3 but less than 30), in each.

If, contrary to reality, we had higher S/N spectra objects might shift around in the BPT diagram but there's no particular reason to expect the distributions to change significantly.

And those who've spent a decade or more working the astronomical data would surely have developed instincts on this that very likely could be trusted. Me? I'm a total noob, from Missouri (the "Show Me" state; but of course I'm not!), so I'd prefer to nail it down.

The control sample hardly matters at all in this discussion.

I sorta/kinda worked that out, but I still don't really understand why

For one thing there's only one "Low S/N AGN" in the entire quench+control samples,

There is? Which object is it?

and decreasing the detection threshold to 2σ for "Low S/N SF" seems inconsistent.

I agree ... if only because it surely introduces additional unknown biases (a bright object with weak lines and faint object with strong lines would/could both end up in this class, yet they are quite different!).

BTW, I haven't been explicitly noting them, but there are at least as many objects with 'masked' [OII] or Hδ lines as any of the traditional four (possibly more; 'masking' in the UV region - blueward of the 4000A cliff - seems at least as common as for Hα/[N II], although the missing chunks are generally narrower).

And of course, I haven't yet fully absorbed the details of your analyses ... 😦

Posted March 15, 2014 2:23 PM
by JeanTate in response to mlpeck's comment.

That is very cool! 😄

Did you do it using R (I'll bet you did)?

Posted March 15, 2014 2:24 PM
by mlpeck in response to JeanTate's comment.

Did you do it using R (I'll bet you did)?

Mostly R. I wrote a C program using cfitsio routines to extract the spectra and some limited metadata from the SDSS spectrum files. That part could probably be dispensed with since there's an R package for reading and writing FITS files, however I haven't tried it.

Posted March 16, 2014 2:26 PM
by mlpeck in response to JeanTate's comment.
```
The control sample hardly matters at all in this discussion.
```
I sorta/kinda worked that out, but I still don't really understand why
I'll try to write more later and address your other points, but for now:

The control sample hardly matters because you could randomly pick any redshift limited sample of spectra from SDSS and you'd get about the same proportion falling in the starforming area of the BPT diagram. And conversely if you make a selection based on traditional criteria for "K+A" like spectra you'll get a very small proportion in the starforming area. I will post the SQL query I used to get a K+A sample later if you're interested.

The maybe more important point is that there's no reasonable amount of cherry picking that can turn "a lot" of starforming galaxies in the quench sample into "not a lot," and the question of why there are a lot will need to be discussed in any paper that gets submitted (I hope).

Posted March 16, 2014 2:48 PM
by mlpeck in response to JeanTate's comment.
```
For one thing there's only one "Low S/N AGN" in the entire quench+control samples,
```
There is? Which object is it?
AGS004e3 is the only one I found. That's in v2 of subset 2 of the control sample. If you check the spectrum [N II] is relatively obvious, Hα is a barely significant detection in the MPA pipeline, [O III] is present, and Hβ is undetected.

Posted March 17, 2014 3:54 PM
by JeanTate in response to mlpeck's comment.

Thanks, that helps a lot.

At a sufficiently high level (and in my own words), if QS objects are 'post-quenched', then there should be very few in the 'star-forming' part of the BPT diagram. Why? Because - at some level - 'quenched' means 'no more star-formation'.

From the objects I've looked at so far - among "the 1149" and "the 1196" (0.02 < z < 0.10 AND abs_z brighter than -20.0) - there are a handful of ppos with respect to 'galaxy', and placement on a BPT diagram.

For one thing, there's the 'aperture effect' we've discussed before (the part of the galaxy sampled by the spectroscopic fiber not being representative of the galaxy as a whole); but I haven't really looked into this yet.

At one stage in a true merger, both nuclei are still clearly distinct, and one or two disks can still be seen, sorta. In such objects, is there one galaxy, or two? Perhaps one can easily be said to be 'post-quenched', but the other not? In these systems, what can we say if we have a spectrum of only one piece (which may not even be a nucleus; knots of intense star-formation may have been chosen as the spectroscopic target)?

Then there are the galaxies where the fiber did not cover/include the nucleus: most such are outside our redshift range, but there are some - both QS and QC - which are within it.

If only because the total number of such ppos is relatively small, it is highly unlikely to make a difference to the high-level result ('there are rather too many star-forming QS objects'), but will affect it when we 'do the numbers'.

Posted March 17, 2014 4:20 PM
by mlpeck in response to JeanTate's comment.

At one stage in a true merger, both nuclei are still clearly distinct,
and one or two disks can still be seen, sorta. In such objects, is
there one galaxy, or two? Perhaps one can easily be said to be
'post-quenched', but the other not? In these systems, what can we say
if we have a spectrum of only one piece (which may not even be a
nucleus; knots of intense star-formation may have been chosen as the
spectroscopic target)?

We have an incomplete, and probably statistically unrepresentative collection of spectroscopically confirmed mergers. Having a look at the spectra of the neighbors as well as the quench sample objects to see what can be learned about their evolutionary state seems like an interesting project, but it's outside the scope of this investigation I think. I may do it anyway.

Posted March 18, 2014 3:09 PM