Testing Likelihood

I've been given the task to figure out why the likelihood method is missing QSOs that other methods are selecting.

I'm doing this by looking at QSOs from the 3PC (3%) catalog (see Feb 3rd email from Nic Ross). This file has the positions, redshifts, and targeting information for the objects.

I can get the psf_fluxes by doing a spherematch with the QSOs from the 3PC with the collated summery files (spAll-v5_4_5.fits) which have additional information for these objects. Nick gave me a trimmed file which I put on reimann: /home/jessica/boss/Single_objects_53839.dat which doesn't have duplicate entries for the same objects.

Below is a color-color diagram of the QSO catalog (green) and the everything else catalog (red). These are the catalogs we are using for calculating the likelihood:


Below is a color-color diagram of spectroscopically confirmed quasars which were targeted by the likelihood method (cyan) and missed by the likelihood method (magenta) on top of the normal star template (red) and quasar template (green):


As you would expect, we are missing objects which fall closer to the stellar locus. We can see this better when I plot the green QSO catalog with bigger points:


Looking at the straight psffluxes (which is what we calculate the likelihoods on). Below are plots of log-log plots of the psfflux for various color bands. The cyan points and quasars that the likelihood method targeted. The magenta points are quasars that the likelihood method didn't target (missed). This helps show regions in psfflux space where we do not have QSOs in our catalog which exist in the targeting data set.

u vs g

Below shows the above plot on top of the "everything" (red)

and "quasar" (green) catalogs used in the likelihood method.
u vs r

u vs i

u vs z
g vs r
g vs i

g vs z

r vs i
r vs z

i vz z

I think we should think about including some of these quasars in the inputs for the monte carlo so that we better cover these color regions.

I'm going to also plot where the mis-targeted stars are on these diagrams too.

More Reconstruction Binning

Below shows the evolution the reconstruction of binning on a run with 200,000 photometric and spectroscopic objects with 20 spectroscopic redshift bins. The number after "reconstruction" in the legend is the number of reconstruction bins. The best reconstruction is with 40 bins which is again 2X the number of spectroscopic redshift bins. I am running another run with more redshift bins to see how that effects things. I am going to look deeper into the reconstruction code to see if there is something hard-wired in that Alexia forgot to tell me about. Some reason why it only reconstructs well with 2X the redshift bins.


Looking at binning close up around the best fit (40). You can see that even around the best fit the reconstruction behavior is erratic:

Reconstruction Binning

As I discussed in a previous post, the number of bins I use in the reconstruction seems to have a large effect on the shape of the reconstruction. I want to get to the bottom of this. Here are reconstructions of a small run with several different values of npbins. Alexia, I'd be very interested if you have a similar issue when you did the reconstruction:

5 Reconstruction Bins

10 Reconstruction Bins


15 Reconstruction Bins

20 Reconstruction Bins (best fit)


25 Reconstruction Bins

30 Reconstruction Bins

The above run had 10 redshift bins in the spectroscopic data, so I am puzzled by the fact that the best fit is 20 bins because you would think you wouldn't be able to reconstruct as well with finer binning than we have redshift bins. It is also puzzling that the reconstruction gets worse with smaller bins, because you would think that the less binning the better the reconstruction would work.

To see how this binning effect changes with the number of spectroscopic redshift bins, I am running the reconstruction with more (20) redshift bins. Stay tuned!