Bright Star Spread Model Issues

-Eli Rykoff.

As was noted in A Modest Proposal for Preliminary Star/Galaxy Separation, in the SVA1 imaging there is a turn-up in spread_model at the bright end:

These are all stars correctly identified as such by, eg, CLASS_STAR, but appear to be significantly fatter than the psf model. What's up with that?

More Clues

A few things that I've checked (no plots since I'm lazy):

  1. The turn-up is not significantly present in single epoch imaging. There is a bit of a turn-up in some of the (deeper?) single epoch images, but it is small enough that at the brightest end SPREAD_MODEL < 0.002. This may be a manifestation of the brighter/fatter relation. But what we're seeing in the coadds is a factor of several stronger.
  2. This is not obviously caused by astrometric errors. I ran a test field with improved relative astrometry and found the same turn up.
  3. The turn-up is not present in PSF homogenized coadd catalogs produced by the Munich group.
  4. Joe Mohr has pointed out that the extreme outliers (SPREAD_MODEL_I > 0.003) may have a geometric structure. One SPT cluster in particular has some apparent structure SPT-CL0500-5116 (see SVA1 stars with SPREAD_MODEL_I > 0.002).

Joe has suggested that these outlier misclassified stars are caused by discontinuities in the coadd PSF model. I am not sure why this should affect the brightest stars only ... plus if you look at the spread_model plot above it's clear that the turn-up affects all the bright stars, with some scatter.

Where are the outliers?

This is a map of the number of SE images that went into the DES0500-5116 i-band coadd, which is the tile containing SPT-CL0500-5116. Red has 2 images, green 7, and blue >10. White squares are the SPREAD_MODEL_I>0.002 outliers, magenta diamonds are all non-saturated stars with i<16.

It's clear that these misclassified stars are preferentially in the deeper regions, rather than in the chip gaps which would be predicted from the discontinuous PSF hypothesis. Furthermore, only a very small fraction of the larger outliers are in the shallower fields where the psf is more likely to have abrupt changes. On the other hand, this is consistent with the finding that there is very little turn-up on the spread_model locus in SE images.

As another test of this hypothesis, I made a subregion coadd centered at 75.2, -51.55, getting a contiguous region of relatively uniform depth. If the spread_model turn-up is caused by discontinuous psfs at the chip gaps, then this should do a better job of measuring the psf.

The stars in the sub-image are overplotted in red. They have the same turn-up! (Though with lower statistics...)

Although I would believe that if you stack 100+ images the average psf will look very round and regular due to reversion to the mean, it's possible that the issue is that with 10+ images the non-homogenized coadd psf is so hopeless complex that psfex can't handle it at all. But I would think this would show up in other ways as well.

In any event, the chip gaps and edges do not appear to be correlated with the star misclassification, which is good.

But the mystery remains.

To add to the mistery, I have attached a couple of plots from a tile from Y1CX which were ran with and without PSF homogeneization. The turn-up is visible in both, so maybe this has nothing to do with PSF homogeneization. It seems then that Munich is doing something different that avoids the turn-up.


if I look at the same in the DC6B experiment we did with this, the PSF homogeneization does make a difference, so maybe it's just the lack of statistics that conceals the effect in the above plots, or that it is much more pronounced in non-PSF homogeneized images (NOTE: mislabeled axes). In this case (plot not included), I also see an important correlation with MAG_PSF residual.