I tested many versions of the Log Likelihood code, and the most reliable was using the TDF as the timing probability function and only using it for hits with < 100 MIP. All but two of the showers in 2006/2007 will reconstruct.
With shower 02-03-07, the largest hit is 25 MIP. Removing it allows the shower to reconstruct. However, just putting a cut on the timing probability for hits > 20 MIP does not allow it to reconstruct. So there is some additional problem besides the fact that large hits seem to be a problem with P_TDF.
I made maps of the -Log(Likelihood) for a sample of recent showers. The angle, time, and energy variables are set to the best value found by the chi-square fit. Then x and y are varied and the likelihood computed and plotted.
Showers that pass:
03-18-07b: Large map, Small map
03-17-07: Large map, Small map
03-05-07a: Large map, Small map
02-27-07: Large map, Small map
02-03-07_v2: Large map, Small map
Showers that fail:
03-18-07a: Large map, Small map
03-05-07b: Large map, Small map
02-03-07: Large map, Small map
I don't see any obvious difference between the showers that fail and the showers that pass.
It is interesting to note that the only sites visible on the maps are sites that were hit. That means that there is almost no contribution from unhit sites (ie, no penalty for expecting a large hit and not getting one). Changing this might keep showers from going off the map.
