Up until the end of October, recoos had a bug which caused the u_pandel likelihood function to treat PMT white noise incorrectly.
Inside the u_pandel timing likeliehood function (found in reco_amanda.c), the log of the probability density is calculated using the pandel function. If the return value drops below the PMT noise floor, the noise probablity density (noise frequency) is used instead of what pandel returns [see Wiebusch, Muon Reconstruction with AMANDA, DESY-PROC-1999-01 pp 302-316]. This is done with the following line of code:
/* check for noise background */
log_p1 = (log_p1 <= reco_par.log_white_noise)
? log_p1 : reco_par.log_white_noise;
The bug was that reco_par.log_white_noise was set to twice the actual value. For example instead of a 10-6 ns-1 noise rate, recoos would use 10-12 ns-1.
This bug correction is significant in that an entire class of fakes are now eliminated without quality cuts. An example of this class of fakes is shown below.
Notice the two (red) early noise hits near the bottom of the detector. Prior to the bug fix, the probability of those hits being noise was so low that the algorithm would rather believe the red noise hits were real hits and the rest of the hits in the detector were from scattered light. Correcting for the bug, the algorithm now reconstructs this event as the downgooing purple fit.
In the following event there is just one early noise hit, but the the noise-corrected fit completely flips around:
The corrected code also eliminates a limited number of COG fakes:
Ty has run the noise-corrected verson of recoos on a sample of 58,000 data events which had passed a multiple iteration bayesian reconstruction (done with the bugged version of recoos). After running the noise-corrected bayesian fit alone, 54% of the 58,000 reconstruct downward! Or, applying cross-talk cleaning first reduces the sample of 58,000 to 9350, then applying the noise-corrected bayesian reconstruction reduces the set to about 5000.
This number is not suprising given that the expected number of noise hits with a PMT noise rate of 1000 Hz over a time window of 4500 ns is about 1.
I ran the noise-corrected bayesian reconstruction on a sample of 640 events from one of the final cut levels in one of Gary's previous analyses. 43% of the events in this sample now reconstruct as downward. The following is a zenith distribution of the 640 events as reconstructed with the noise-corrected bayesian fit:
We have visually scanned the events which were rejected. Most of them appear to belong to the early noise-hit class like those shown above.
Shown below is the difference in zenith angle between the noise-corrected bayesian fit and the bugged fit for the 57% of the 640 events that survive as upgoing. As you can see, the corrected fit is virtually the same as the bugged fit for those events which (presumably) have no noise hits or have a higher upgoing likelihood despite the possibility of noise.
Consistent with expectation, the corrected bayesian fit has little impact on previously reconstructed events with no apparent noise hits. However, the impact of the log white noise bug fix in recoos is very significant for background rejection. The noise-corrected bayesian fit reduces samples at high cut levels by about half. The class of early noise-hit fakes now appear to be eliminated without the imposition of further quality cuts.