AtlasMetTriggerAnalysis

Introduction

The goal of this page is to list the histograms which are needed to study the performance and the efficiency of the ATLAS missing E_T trigger (henceforth, MET trigger) and to outline the operations to be done on them.

Streams

We assume that one can select the streams corresponding to the events passing:

• any (L1 or HLT) egamma, tau, jet, MET or SumET selection (Calo stream);
• zero bias triggers (ZeroBias stream);
• minimum bias triggers (MinBias stream);
• muon triggers (Muon stream);
• random triggers (Rndm stream).

Binning

The binning is 3 GeV/bin for MET, MEx, MEy, MEz and 6 GeV/bin for SumET, with a bin centered on zero [i.e. (-1.5 GeV to 1.5 GeV for MET, -3 GeV to 3 GeV for SumET)], when linear scale is used.

The binning for log-scale quantities is 0.15/bin = 1.413X/bin (factor of ~2 each two bins). For MEx, MEy, MEz one defines 41 bins with xmin=-3.075, xmax=3.075 (limits are +/- 1.189 TeV). For MET, SumET and SumE one defines 34 bin with xmin=-1.125 (75 MeV) and xmax=3.975 (6.683 TeV).

Offline reference

In AtlasP1HLT 15.5.6.10 there is no hadronic calibration at EF, which is at EM scale exactly as L1Calo and L2. In addition, the muon correction at HLT is done by taking the muon features computed by muFast (both at L2 and EF). Finally, SumET is not corrected for the muon p_T. This is the configuration for the initial data taking, until a peak luminosity of 1e32 cm^-2 s^-1 is reached. Later, hadronic correction at L1 and HLT should be applied and the muon correction at EF may be refined.

Dependently on the muon correction and on the hadronic calibration, different offline algorithms need to be used to compare against. On the long term, when the offline hadronic calibration is validated, the reference quantities are:

• MET_Final - MET_Cryo for trigger quantities corrected for the muons
• MET_Final - MET_Cryo - MET_Muon for the trigger quantities not corrected for muons

where a vector (scalar) sum is actually used for MET (SumET).

For the initial data taking, simpler algorithms will be used. The calorimeter only contribution to MET and SumET is best estimated by MET_Topo (EM scale sum over all topological clusters), which is the reference quantity, though MET_Base (sum over all cells after a 2-sided 2-sigma noise cut) may also be useful.

L1 items and primary HLT chains

For the first data taking, the following L1 signatures are implemented:

• XE10, XE15, XE20, XE25, XE30, XE35, XE40, XE50
• TE10, TE30, TE100, TE180

and the primary chain is the lowest unprescaled chain in the following list, at any luminosity:

• xe20_noMu has (L1, L2, EF) thresholds set to (10, 12, 20) GeV
• xe25_medium_noMu = (10, 12, 25)
• xe25_noMu = (10, 15, 25)
• xe25_tight_noMu = (10, 17, 25)
• xe30_loose_noMu = (10, 12, 30)
• xe30_medium_noMu = (10, 15, 30)
• xe30_noMu = (15, 20, 30)

Corresponding chains with muon correction are defined but disabled by default in the initial trigger menu.

Clean-up cuts

To ensure that the comparison against the reference quantity is carried on over a sample which is relatively free from mismeasurements, one may apply the following clean-up cuts:

1. MBTS timing should be consistent with collision
2. LAr timing consistent with collision
3. jet clean-up
   • bla bla
4. track clean-up
   • bla bla?

Comparison before/after some change

We want to look for the effects on the MET trigger due to changes in the configuration or in the hardware (including problems) which are well isolated (i.e. we can isolate runs before and after such change). This is supposed to be a quick check.

We assume that a list of relevant (run, luminosity block) pairs define the dataset B (= before the change) and another list defines the dataset A (= after the change).

The following sections list the histograms to be filled during the loop over events and explain how they will be used in the later analysis to make useful plots.

L1 MET and SumET distributions

Histograms: L1 MET and SumET distributions separately filled for datasets B and A for Calo, MinBias and Rndm streams.

In addition to overlay the distributions obtained before and after the change, one will also compare the corresponding rates as function of the threshold by taking the normalized integral from x to infinity (complement of the cumulative function).

L2 MET, SumET distributions for events passing L1

Histograms: L2 MET and SumET distributions separately filled for datasets B and A for events in the Calo stream which passed each L1 XE and TE signature.

In addition to overlay the distributions obtained before and after the change, one will also compare the corresponding rates as function of the threshold by taking the normalized integral from x to infinity (complement of the cumulative function).

EF MET, SumET distributions for events passing L2

Histograms: EF MET and SumET distributions separately filled for datasets B and A for events in the Calo stream which passed each L2 XE and TE chain.

In addition to overlay the distributions obtained before and after the change, one will also compare the corresponding rates as function of the threshold by taking the normalized integral from x to infinity (complement of the cumulative function).

Reference MET, SumET distributions for events passing the trigger

Histograms: Offline MET and SumET distributions separately filled for datasets B and A for events in the Calo stream which passed each XE and TE chain.

Overlay the distributions obtained before and after the change. In addition, compare the cumulative distribution functions (normalized integral from - infinity to x), interpreted as the integral trigger "efficiency" for each threshold on the reference distribution. [Because no clean-up cut is applied, this "efficiency" includes signal and noise events. Turn-on curves (see below) are used to estimate the selection efficiency separately for each definition of signal and background.]

Comparison with offline and simulation

Histograms: 2D histograms with the reference quantity (offline or MC) on the horizontal axis and the trigger (L1, L2, EF) quantity on the vertical axis, separately filled for datasets B and A for all events in the Calo and MinBias streams.

The same histograms are filled again for the events in the Calo stream which survive clean-up cuts.

Turn-on curves

Histograms: Starting from the Calo stream, remove all events which do not satisfy the clean-up cuts and fill MET and SumET histograms with the reference (offline or MC) quantity for all events and after each of the following cuts:

1. L1 XE signature
2. L1 TE signature
3. L2 XE chain
4. L2 TE chain
5. EF XE chain
6. EF TE chain

The turn-on curves are the fits of the bin-wise divisions between the distribution filled after the cut and the distribution before the cut.

Histograms: If the statistics allow for this, repeat this procedure starting from the events selected by "orthogonal" triggers (used to fill the distributions before the cut) and fill the distributions after each of the cuts listed above.

Performance and efficiency study

The trigger behaviour can be characterized by a study of the L1, L2, EF resolution as function of the total calorimetric activity in the event [the assumption being that the calo dominates the resolution] and of its efficiency in selecting "signal" events while rejecting most of the "background" events.

Care must be taken to properly and precisely define what "signal" and "background" are, expecially for the efficiency study.

All histograms defined above are useful. In addition, the following sections list few more ones.

Resolution

In all official ATLAS papers, the MET resolution is computed in different SumET intervals. For this reason, this is the baseline check also for the trigger case. However, SumE is probably even better to quantify the MET resolution.

Two important comments shall be done. First, even though one speaks about the "MET resolution", actually one fits the distribution of (trigger - reference) MEx and MEy because MET is not Gaussian distributed around the reference value.

Second, the choice of the bin size in SumET or SumE is important. A uniform binning in linear scale has the drawback of having a very poor statistics at higher values. A better choice would be a bin width increasing at the same pace as the resolution (as done for the detector book for the MET trigger) but this may still be too slow. Another choice is a uniform binning in log-scale, which is likely better.

Histograms: 2D histograms with the reference SumET or SumE on the horizontal axis, with increasing bin size, and the difference (trigger - reference) MEx or MEy on the vertical axis, in linear scale.

The binning on the horizontal axis is chosen such that one has uniform spacing on the log-scale. The log-scale bin width is 0.15 (see above). The binning on the vertical scale is the same as for the MEx, MEy distributions.

Efficiency

The histograms are defined as mentioned above, and filled before and after each selection (single item at L1, chain at HLT). The problem is the definition of the initial sample.

What matters at the trigger level is to achieve the maximum efficiency for the "signal" compatible with the rate, which is dominated by "background". A way to do so is to look at the ratio between the two efficiencies, keeping in mind that the rate is the most important constraint.

To make an "unbiased" efficiency measurement, care must be taken to select the initial sample with a statistically uncorrelated trigger. At present, three possible selections have been examined:

1. zero bias trigger (so far, random trigger). It seems that requiring MBTS_1_1 does not change the result, hence also "minimum bias" triggers are good. No statistics for high thresholds;
2. electron trigger. The goal is to select W decays. For the lowest XE thresholds the statistics might be too low. For the lowest TE chains there is some bias, because electrons are triggered by calorimetry and contribute to SumET;
3. muon trigger. The goal is again to select W decays. Trigger and reference quantities must be computed without muon correction for this method to work. Very low L1 efficiency is expected for events in which the muon and the neutrino are back to back (no muon correction is possible at L1).

-- DiegoCasadei - 26-Apr-2010