Review of background methods

From the previous talk, Preshower, it seems that the peak could be sitting on a non uniform background

To investigate this further I have tried several different background mixing methods, first I shall go over how these were developed and then I will give a review of the current state of this investigation

Background methods and their pros and cons

All methods

All methods must form a "mimic" pair with one track passing the L0 trigger cut and the other track passing the L2 trigger cut (see last weeks talk).

Current method

Currently we pair + with + and - with - in the event we wish to find Upsion signal (+ with -) in the same event

Pros:

Easy to code, easy to normalise

Mixing across different events preseves any correlations

Cons:

  1. Limited and fixed statistics prone to fluctuations

  2. Could be physical correlations between ++ and -- tracks in the event that will deviate from a pure background assumption

Mixing with other events at random

This was the simpelest solution and the point where I started to develope the code:

The boxes represent events, each containing + and - tracks

The box on the left hand side is our current event, + and - tracks here are paired with + and - tracks from other events

The arrows show where the tracks from one event are paired with the tracks from another

Pros:

  1. Unlimited background statistics, just pick as many as you want

  2. Events are true background, tracks can't be correlated over events (depending on shape this could also be a "con")

  3. Increased statistics for +- tracks

Cons:

  1. No vertex or multiplicity correlation between events

  2. Events picked at random ramp up computing time to collosal level

Conclusion

  1. The multiplicity and vertex problem was fixed by adding a dynamic array, a first pass though the data uses these arrays to bins the vertex and multiplicity of each event. Then when the code starts paring tracks between events, it use these bins as a reference to mix events with a similare multiplicity and vertex

  2. Only the first primary vertex in events were mixed with the first primary vertex from different events as this dramatically simplifed the code

  3. The computing time problem came from the large input file size, the TTree is much larger then the memory and only seems to store part of the tree in memory at one time If you pick wildly different events the postions in the TTree are likley still on disk and not in memory. This slows the computing time down by up to orders of 1000. This was fixed by tagging the multiplicity array and vertex array with the location in the TTree of these events, then picking points in the bin that were close to the current event in memory

Binned Multiplicity and vertex with array "speed fix" (the next iteration will perform a centrality selection)

  1. Multiplicity binned 10 bins from 0-1000 tracks per event
  2. Vertex binned in 8 10 cm. bins -40cm to 40cm.

Pros: all the above

Cons: probably a pain to normalise, currently using the ratio of the background integral to the signal integral but a track by track normalisation would be better (next iteration will remove the peak region, i.e. normailise from 6-8 GeV and 11-15 GeV)

Very preliminary results:

Current background PID + MIMIC trigger cuts

Background from mixing between events PID + MIMIC trigger cuts

Background from event by event (red) scaled to 8 GeV bin of same event paring (Black)

Conclusion

Note: I rewrote a lot of code so I am checking this step by step, normalisations will need to be looked for a start as I used just an integral for this discussion, but I think it is 95% certain that the BG mixing will improve the signal, considering these peliminary results