HLT description

Maintained by Eric van Herwijnen

Last update: December 20, 2011









Text and some figures extracted from the HLT section 3.2 of the LHCb Upgrade LOI




General structure




·        Strategy



·        Extra tracks



·        Decision

·        Examples



HLT2: exclusive and inclusive selections

HLT monitoring






General structure

The High Level Trigger (HLT) is the second (and last) level of trigger of LHCb, running on events passing the L0 trigger. It consists of a C++ application (called “Moore”) that will run in several copies on every CPU of the Event Filter Farm (EFF), which consists of approximatively 1000 multi-core computing nodes. Currently there are 26110 copies of Moore running in the EFF. The HLT application has access to all data in one event.

The HLT is divided in two steps: HLT1 and HLT2.

HLT1 reduces the rate from the 1.1 MHz output of L0 to ~50 kHz. HLT1 reconstructs particles in the VELO and determines the position of the primary vertices (PV) in the event. To limit the CPU consumption, a selection of VELO tracks is made based on their smallest impact parameter (IP) to any PV, and their quality. For these selected VELO tracks their track-segment in the T-stations are sought to determine their momentum (p), so-called forward tracking. HLT1 selects events with at least one track which satisfies minimum requirements in IP, p, pT and track quality. (Transverse is defined perpendicular to the beam-pipe, in the plane formed by the track and the beam-pipe.) It reduces the rate to a sufficiently low level to allow forward tracking of all VELO tracks.

HLT1 should reduce the rate to a sufficiently low level to allow the full pattern recognition on the remaining events.

HLT2 searches for secondary vertices, and applies decay length and mass cuts to reduce the rate to the level at which the events can be written to storage and processed offline. Currently this is ~3 kHz. It first performs a complete pattern recognition to find all particle tracks in the event, using VELO tracks as seeds. Then, a set of different selections are applied. Some of them are inclusive, aiming for generic B decays or for resonances like J/Psi(->mm) or D*(->pKp), and some of them are exclusive, aiming to provide the highest possible efficiency on specific B decay channels.

The schematic of the overall LHCb trigger scheme is shown in the figure below:


Go to top





A small fraction of the events in LHCb have large detector occupancies, especially in the OT. Some of these events take even seconds to reconstruct in HLT1, and are rejected before any reconstruction to keep the average processing time ~25 ms per event. Events are rejected if the OT occupancy is larger than 20%. For the remaining events the reconstruction strategy is determined by the following considerations:


·         All B meson decays studied at LHCb contain at least two charged tracks in their final state;

·         B mesons are heavy, and their average momentum in the LHCb acceptance is ~100 GeV/c, so their decay products will have a large momentum (p) and transverse momentum (pT) compared with light-quark hadrons originating from the PV;

·         The average decay length of B mesons produced in the LHCb acceptance is ~1 cm so that their decay products will have a large impact parameter (IP) with respect to their PV.

·         The VELO reconstruction is fast enough to allow a full 3D pattern recognition and PV finding to be performed for all events entering the HLT.


Go to top


Velo reconstruction

This last point is illustrated in the figure:



The timing of the VELO 3D pattern recognition as a function of the off-line reconstructed

number of PVs in the event. The entry at 4 PVs contains all events with more than 3 PV.


The timing increases linearly with the number of PV in the events, and is low enough to allow sufficient time for the subsequent reconstruction and event selection. The event reconstruction therefore begins with the VELO pattern recognition and PV finding. Because of timing constraints, the momentum can only be determined for a limited number of VELO tracks. It is necessary to select those tracks which are most likely to come from a B decay. Three selection criteria are used:


1.      The IP of the VELO track with respect to the closest PV

2.      The number of hits assigned to the VELO track

3.      The difference between the number of hits assigned to the VELO track and the number of hits expected given the track direction and the first measured point on the track (missed hits)


The following figures compare the distributions of the latter two criteria for tracks from a minimum bias event and for the highest pT B daughter track in Bs φφ decays.




The number of hits on the VELO track for minimum bias (dashed red) and the highest pT offline-selected Bs φφ daughter (solid blue).

The difference between the expected and observed number of hits on a VELO track for minimum bias (dashed red) and the highest pT offline-selected Bs φφ daughter (solid blue). Note the logarithmic scale.

It can be seen that both the number of hits on a VELO track and the number of missed hits are good discriminants, especially since they can be applied before any forward reconstruction, thus saving CPU time by having to consider fewer VELO tracks. Requiring |IP| > 125 μm, and number of VELO hits and missing hits > 9 and < 3 respectively, is very efficient at selecting the highest pT daughter from a Bs φφ decay. The figure below shows the number of VELO tracks per event for which the momentum needs to be determined.


Number of VELO tracks per event which enter the forward reconstruction to have their momentum determined.

Go to top



The VELO tracks thus selected are extrapolated to the tracking stations using the Pat-Forward algorithm. Imposing a minimum momentum and transverse momentum (p, pT) in the forward tracking significantly reduces the search windows which have to be opened in the IT and OT tracking stations and consequently reduces the required CPU power. This is illustrated in the figure below (left), which shows the reconstruction time per event of the forward tracking as a function of the minimum (p, pT) cutoff imposed. The figure below (right) shows the transverse-momentum distributions of Bs φφ decay products which have been selected by the VELO IP and quality cuts.



The CPU time of the forward tracking as a function of the minimum pT imposed. The corresponding minimum p is always taken as ten times the minimum pT.

Distribution of the largest transverse momentum per event of Bs φφ decay products which have been selected by the VELO IP and quality cuts.

Hence, imposing a minimum (p, pT) of (12, 1.2) GeV/c results in a negligible loss of signal, while the CPU time per event is acceptable. The final track selection proceeds in two stages. First (p, pT) cuts are used in order to reduce the rate. The remaining tracks are fitted using a Kalman filter with outlier removal, in order to obtain an offline-quality value for the track χ2 as well as an offline-quality covariance matrix at the first state of the track, allowing a cut on the IP significance squared (IP χ2). The number of tracks which have to be fitted is low enough by this stage that the contribution of the track fit to the overall timing is negligible. The track χ2 is a powerful tool for ghost rejection in the trigger; however, being particularly sensitive to the detector performance, it needs to be verified with real data. The figure below (left) shows the distribution of the online track χ2/ndf for minimum bias events recorded in 2010 surviving to this stage of the trigger, as well as the highest momentum daughter from offline selected real data D+ h+h+h decays. The data demonstrate that LHCb achieved an excellent reconstruction performance with very low online track χ2 values for genuine signal tracks. The track χ2 and IP χ2 are therefore used to achieve the final required rate reduction. The figure below (right) shows the fraction of events which are triggered by at least one ghost track as a function of the OT occupancy, imposing a track χ2/ndf < 3 cut for minimum bias events surviving to this stage of the trigger. On average 15% of the events accepted by HLT1 are triggered due to a ghost track.



The track χ2/ndf of forward reconstructed tracks in the trigger in real data minimum bias events (dashed red) and of the highest momentum daughter from offline selected real data D+ h+h+hdecays (solid blue).

Fraction of HLT1 accepted events which are triggered by at least one ghost track as a function of the OT occupancy, imposing a track χ2/ndf < 3 cut.


Go to top


Muons and radiative decays

For decays involving muons an additional cut is available: the muon ID algorithm  which is in itself a powerful ghost and background rejection tool. This fact is exploited in a parallel trigger line which is run only on those events passing the L0 muon trigger, and hence most of the above cuts can be relaxed for muon candidates. For radiative B decays, e.g. Bs φγ, the requirements of the offline background rejection impose a tight LLT electromagnetic cut of 2.4 GeV, so that L0 electromagnetic triggers make up only a small fraction of the total L0 rate. Therefore HLT1 contains another parallel line only running on L0 electromagnetic triggered events for which the (p, pT) cuts are significantly reduced. In addition, as this trigger line searches for lower momentum tracks, the track χ2 cut is loosened.

Go to top




HLT2 is mainly based on three inclusive trigger lines, so-called topological lines. These lines in principle cover all B decays with a displaced vertex, and with at least two charged particles in the final state. In addition HLT2 contains trigger lines which exploit the presence of muons, and a few lines which aim at exclusively reconstructing golden B-decay modes. The topological lines are designed to have:


·         high efficiency for any B decay with at least two charged daughters, due to the inclusive nature of the trigger lines;

·         an excellent timing performance and background rejection, due to the small number of trigger lines;

·         excellent data-mining properties due to their inclusive nature;

·         trigger redundancy for golden modes, for which a few more exclusive selections will be deployed.


In an inclusive trigger cuts must be avoided on quantities such as the mass of the B candidate or how well the direction of its momentum agrees with the direction defined by the primary and secondary vertices. To trigger efficiently on B decays with long-lived resonances (such as D mesons), tight cuts on the quality of the vertices must also be avoided. Instead, quantities that preserve the inclusiveness of the trigger while also providing large background rejection factors are used. All VELO tracks are extrapolated to the tracking stations to have their momentum measured, imposing a minimum (p, pT) of (5000, 500) MeV/c in the PatForward algorithm to save CPU time. To reduce the background rate due to ghosts, all tracks are required to have a track χ2/ndf value less than 5. To reduce the background rate due to prompt particles, all tracks are required to have an IP χ2 value greater than 16. Due to the inclusive nature of the HLT2 topological lines, this does not mean that all of the B daughters need to satisfy these criteria. The trigger is designed to allow for the omission of one or more daughters when forming the trigger candidate. Processing time is saved in the HLT2 topological lines by simply assigning each input particle a kaon mass.


The multibody candidates are built as follows: two input particles are combined to form a two-body object; another input particle is added to the two body object to form a three-body object, and so on. An n-body candidate is thus formed by combining an (n1)-body candidate and a particle, not by directly combining n particles. The importance of this distinction is in how the distance of closest approach (DOCA) cuts are made. When a 2-body object is built, a DOCA< 0.15 mm cut is imposed for the object to either become a 2-body candidate or to become the seed for a 3-body candidate. When a 3-body object is made by combining a 2-body object and another particle, another DOCA< 0.15 mm cut is imposed for the object

to either become a 3-body candidate or input to a 4-body candidate. This DOCA is of the 2-body object and the additional particle, not the maximum DOCA of the three particles. This greatly enhances the efficiency of the HLT2 topological lines on B DX decays. A similar procedure is followed when making 4-body candidates from 3-body objects and an additional particle. All n-body candidates that pass these DOCA cuts are then filtered using a number of other selection criteria. If a trigger candidate only contains a subset of the daughter particles, then the mass of the candidate will be less than the mass of the B. Thus, any cuts on the mass would need to be very loose if the trigger is to be inclusive. Instead a cut is made on the

corrected mass obtained as follows:




where pTmiss is the missing momentum transverse to the direction of flight, obtained using the primary and secondary vertices, of the trigger candidate. The quantity mcorr would be the mass of the parent if a massless particle was omitted from the trigger candidate, i.e. it is the minimum correction to the trigger-candidate mass if any daughters are missing. The figure below demonstrates the performance of mcorr.








B candidate masses from B K*μμ decays: (a) HLT2 2-body topological trigger candidates; (b) HLT2 3-body topological trigger candidates; (c) HLT2 4-body topological trigger candidates. In each plot, both the measured mass of the n = 2, 3, 4 particles used in the trigger candidate (shaded) and the corrected mass obtained using the equation above the figure (unshaded) are shown.


For cases where there are missing daughters, the mcorr distributions are fairly narrow and peak near the B mass. When the trigger candidate is formed from all of the daughters, the mcorr distributions are slightly wider and shifted upwards by a small amount as compared with the mass distributions, as expected. Thus, the performance of mcorr is ideal for an inclusive trigger line. The HLT2 topological lines require 4 < mcorr < 7 GeV/c2. The HLT2 topological lines further reduce the background retention rate by requiring the pT of the hardest daughter be greater than 1.5 GeV/c and also that the sum of the daughter pT values be greater than 4.0, 4.25 and 4.5 GeV/c for the 2-body, 3-body and 4-body lines, respectively. To further reduce the background rate from candidates with ghost tracks the HLT2 topological lines require that at least one daughter particle has a track χ2/ndf < 3. The trigger candidate’s flight-distance significance value is required to be greater than 8, and its vertex must be downstream of the closest PV. The sum of the daughter IP χ2 values should be greater than 100, 150 and 200 for the 2-body, 3-body and 4-body lines, respectively. One of the larger background contributions to the HLT2 topological lines comes from prompt D mesons, which is suppressed by requiring that all (n 1)-body objects used by an n-body line either have a mass greater than 2.5 GeV/c2 or that they have an IP χ2 > 16. A complete list of the cuts used in all three of the HLT2 topological lines is given in the table below.




Selection Criteria

all input particle transverse momenta

all input particle momenta

all input particle track χ2/ndf

all input particle IP χ2

pminT > 500 MeV/c

pmin > 5 GeV/c

χ2/ndf < 5

IP χ2 > 16

B candidate corrected mass

largest daughter transverse momentum

best daughter track χ2/ndf

sum of daughter transverse momenta

sum of daughter IP χ2

n-body DOCA

B candidate signed flight distance χ2

prompt D veto

4 < mcorr < 7 GeV/c2

pmaxT > 1.5 GeV/c

χ2/ndf < 3

SpT > 4.0, 4.25, 4.5 GeV/c (2, 3, 4-body)

IP χ2 > 100, 150, 200 (2, 3, 4-body)

DOCA< 0.15 mm

FD χ2 > 64

mnbody > 2.5 GeV/c2 or 2, 3-body IP χ2 > 16


In addition to the topological lines, HLT2 contains a set of lines which exploit tracks which have been identified as muons. Dimuon candidates are formed and, depending on their mass, cuts are applied on the flight distance and pT of the dimuon candidate. Single muon candidates are accepted either requiring large pT, or a combination of IP χ2 and pT cuts. 

Go to top

HLT monitoring


Each HLT1 alley and HLT2 selection produces summary information which is written to storage for the accepted events. This summary contains the information of all tracks and vertexes which triggered the event.

It is foreseen to reserve a significant fraction of the output bandwidth for triggers on semi-leptonic B-decays, hence a sample in which the trigger did not bias the decay of the accompanying B-hadron. The summary information is used to check if an event would have triggered, even if the B decay of interest would not have participated in the trigger. It therefore allows studying the trigger performance. The summary information is also sufficient to guarantee that during the analysis the trigger source of an individual event is known.

To assure that during off-line analysis the trigger conditions are known, the combination of trigger algorithms with their selection parameters will be assigned a unique key, the Trigger Configuration Key (TCK). All trigger configurations with their associated TCK are pre-loaded in the EFF before a fill. To change from one trigger configuration to another one, for example to follow the decaying luminosity in a fill, the operator will select a new TCK. This TCK is attached by the Time and Fast Control system (TFC) to each event, and it steers the configuration of the algorithms on the EFF and allows full traceability of the used configuration.

Go to top