Level 4 Offline Trigger Software

The purpose of Level 4 software filtering is to reject backgrounds just before full event reconstruction in the very beginning of the DST production chain. At the same time, high efficiencies for signal events are also required. The signal events include not only $B\bar{B}$ events but also other physics related events such as or $\tau$ events, and events necessary for detector calibration etc. such as di-muon events. Most of backgrounds are related to beam activities by both electrons and positrons. Fig. shows $dr$ (top) and $dz$ (bottom) distributions obtained by a fast tracker, [92], which has been developed for the use in Level 4. The open histograms are track-by-track basis distributions, and the hatched distributions are for tracks closest to the coordinate origin in each event. It can be clearly seen that many tracks are created at the beam pipe and that a large fraction of tracks are generated far away from the origin in the $z$ direction caused by beam interactions with residual gases in the beam pipe. Therefore, the basic strategy to reject backgrounds is to select events with tracks originating from the interaction point, IP.

Figure: $dr$ (top) and $dz$ (bottom) distributions. The open histograms are for track-by-track basis and the hatched histograms show $dr$ or $dz$ distributions closest to the origin in each event.

There are low multiplicity (in terms of charged particles) events even in $B\bar{B}$ decays such as . In order to save these events, we keep events with large energy deposits in the electromagnetic calorimeter, ECL, redundantly. The Level 4 algorithm consists of 4 stages of event selections as follows:

• Selection by Level 1 trigger information:
  Because of demands from the sub-detector groups, events with some specific trigger bits are saved without any further processing.
  
• Energy measured by ECL:
  We require the ECL energy reconstructed by to be greater than 4 GeV. Concurrently a veto logic using Level 1 information is applied, which is formed as a coincidence of ECL and KLM hits, to suppress background events induced by cosmic rays.
  
• Selection of events with a track coming from IP:
  By reconstructing charged tracks by , events with at least one ``good track'' are selected, where a ``good track'' is defined as a track with greater than 300 MeV/c, less than 1.0 cm, and less than 4.0 cm. This corresponds to the application of cuts on the hatched distributions in Fig. .
  
• Level 4 monitoring:
  For monitoring of the performance of Level 4, 1% of events which fail our selection criteria are saved and passed to further full reconstruction.

In order to reduce CPU usage, events satisfying our requirement in each stage are selected immediately and submitted to the full reconstruction chain without further processing. Table  summarizes the Level 4 efficiency measured for various real data samples.

Table: Fraction of events which satisfy the requirements of Level 4 in various samples of real data.

Event type	Fraction passing Level 4
All triggered events	26.7%
Hadronic events (loose selection)	98.1%
Hadronic events (tight selection)	99.8%
Muon pair events	98.7%
Tau pair events	97.6%
Two photon candidates	92.9%

It can be seen that the data are significantly reduced while keeping high efficiencies for the signals. The efficiency for the high purity hadronic sample is very close to 100%. The efficiency for two specific decay modes are also tested using MC data. First is a so-called gold plated mode, where the and decay generically and the decay of the other $B$ is also generic. Second is where again the daughter particles and the other $B$ decay generically. Out of 1000 generated events each, the efficiency of Level 4 is 100% for both decay modes. For the remaining events, the purity is 76.5% where the purity is defined as the number of events used in either physics analysis or calibration jobs divided by the number of events passing Level 4. This implies that only 6% more events can be suppressed even if we achieve 100% of purity ( ). Therefore, the reduction rate by Level 4 is close to the optimal value. The further drastic suppression requires changes in the definition of signals, i.e. tighter cuts in the event classification. Figure  shows the run dependence of the fraction of events which pass the Level 4 requirements.

Figure: Fraction of events which pass the Level 4 requirements as a function of run number (the full range is about 6 months). The macro structure is due to changes in the accelerator parameters.

There are some steps and bunch structures. Such structures are caused by the changes in the accelerator parameters. Except for the macro structures, the Level 4 performance has been reasonably stable.