Parallel Photon Studies

In working toward a prompt photon cross section measurement, we are using two analyses in parallel.  The first is a two dimensional fit using parameters identified in the earlier OPAL and ZEUS experiments.  The latter is an LDA running on most of the available raw variables.

What was missing from our earlier analyses was a realistic MC sample.  Mike Betancourt provided code to associate appropriate weights with each gamma candidate, and after this only minor changes are necessary to the existing 2D and LDA fitting algorithms.

In the following plots, 'Background' refers to any particle that produces a gamma candidate in the tree that is not marked as a Prompt Photon in the pythia record.

Currently, the scale of these plots is arbitrary, but it can be adjusted for a specific luminosity without too much trouble.

figure 1:  Weighted Prompt Photon Spectrum (blue) and Background Spectrum (red).
Combined Spectrum

In order to avoid biasing our results, we've split our MC dataset into two parts.  The first part is only used to construct our algorithms, the second part is treated as real data (save, of course, when checking the fidelity of our algorithm as below).

For the 2D fit, the training sample is divided into pt bins (candidate pt), and those are further divided in two:  prompt photons and not prompt photons. 
-- With each population we plot f_max (ratio of energy in the high tower to the total candidate energy) vs sigma_cluster (energy weighted square distance from the high tower). 
--These two plots are smoothed by drawing a new histogram where each data point in the previous is replaced by a gaussian.  The widths of the gaussians is one of the parameters we have to tune, but is based on the RMS in each direction of the original histograms.
--The resulting smoothed histograms are used as basis functions for a two parameter fit of the as-yet unused test sample:
   data=x*photons+(y-x)*nonphotons
   x is the photon population, and y is the total population, fixed to be the integral of the data histogram.  (This could be expressed as a single variable fit, but it would require either making the total population a global variable, or normalizing the data histogram, neither of which we wanted to do.)
--From the best fit, we can quickly extract x.  This is then plotted bin-by-bin against the actual values which we can calculate from the test sample. Errors are not yet propagated.

figure 2: Results of the first 2d fit on the barrel.  (Black points are the actual inegrated weight in each bin, Red points are the weights found by the fit).
First Fit Results

Obviously, this is just a first attempt.  More careful tuning of the gaussian widths, and more MC data in the higher pt bins will both help to improve this fit.