A.Vicer for the Virgo Collaboration

1
A.Viceré for the Virgo Collaboration Amaldi 6th
conference June 23rd, 2005
2
Scope of the talk

• Brief description of the Virgo pipelines
• Restricted to codes based on matched filtering
• Not underlining much online/offline differences.
• Application to Virgo C5 commissioning data
• Quick description of the run conditions
• Trigger generation
• Veto application
• Future perspectives

3
VERY Simplified Virgo analysis scheme

• The h-reconstruction process h Rec starts from
  raw data
• Uses frame data from the data acquisition or from
  data files
• May include some noise removal mostly control
  noise.
• A data conditioning process prepares the data for
  the analysis
• Essentially provides down sampled streams
• Also provide streams with line removed
• Different processes receive frames, add new
  channels or events.
• Mbta, Merlino are just two examples
• Results (events) are fused and saved to disk

4
The Virgo Flat search, a.k.a. Merlino

• Distributed signal processing concept
• Receives data from online-offline
• Performs further data pre-conditioning
• Runs detection codes
• Collects, clusterizes events in time

5
The Virgo Multi-band Pipeline MBTA

• Initialization
• Estimate spectral density
• Grid of full frequency band (VIRTUAL) templates
• Grids of (REAL) templates for each frequency band
• Processing
• Run synchronously each grid of REAL templates on
  data
• Data chunk twice the longest REAL template
• Check if any REAL template triggers
• Recombine associated VIRTUAL templates
• Coherent sum of real templates outputs
• Obtain VIRTUAL templates triggers
• Cluster the triggers

6
On-line versus off-line ?

• Virgo choice to get ready for continuous,
  in-time operation
• Not a necessity now but a reasonable goal for
  the future
• No difference in design
• Analysis codes designed as pipelines
• Receive frame data from data acquisition or from
  disk
• Send events to a trigger manager, or just save
  them
• Limited requests on the computing environment
• A simple Beowulf cluster for both Merlino and
  MBTA
• Possible to re-run (play back) off-line the
  full on-line analysis, from the raw data down to
  the events
• Implications in statistical assessment
• Conflicts with the playground data concept
• Choice don't preclude the long term goal in
  design

7
Test the C5 run

• 5 days run
• Major portion without power recycling, shorter
  one in recycled mode.
• About 55 hours analyzed, spanning 3.5 days
  without PR.
• Include hardware injections
• Some analysis performed on-line
• preliminary h-reconstruction, test of the Merlino
  analysis pipeline
• All the analysis re-done off-line

8
Hardware injections

• Performed sending pre-shaped voltages to mirror
  coils
• Takes into account electro-mechanical response
• Normalized according to the sensitivity measured
  before the run
• Inspiral signals at nominal SNR 7 and 14.
• SNR 7 corresponds to 17 kPc for an optimally
  oriented source
• Injection frequency about 1 event every 10 minutes

9
Data pre-processing

• Recombined ITF data selected for analysis
• C5 sensitivity was better in recycled mode
• But h-reconstruction was still preliminary.
• Science mode segments 55 hours of good data
• 5 hour period the so called quiet(1) period
  free from HW injections
• h-reconstruction
• Tested on-line
• providing synchronous stream of preliminarily
  calibrated data
• streams at 20kHz and 4kHz for bursts and inspiral
  analyses.
• Incorporating some noise removal
• part of the control noises
• 50Hz and harmonics
• Data quality
• Provides an online flag incorporating information
  on the lock status and on the actions performed
  on the instrument.

10
Reconstruction vs Sensitivity

• h(t) data accurate
• correspond well to the sensitivity estimated in
  the frequency domain
• Reconstruction includes some noise removal
• Control noise at low frequencies, 50Hz harmonics,
  and calibration lines

11
Inspiral Horizon distance

• For NS-NS et BH-BH (1.4 and 10 Solar Mass)

12
Analysis outlook

• MBTA analysis on the entire run
• 65 templates around 1.4, 1.4 injected events
• Two bands, 60 185 2000 Hz
• Keep triggers with SNR 6
• MBTA analysis on quiet(1) period
• Spanning 1.0, 5.0 M range
• About 2000 templates
• Merlino analysis on quiet(1) period
• Spanning 0.9, 10.0 M range, with 6677 templates
• Keep triggers with SNR 6
• Merlino analysis on the entire run
• Spanning 1.0, 5.0 M range, with 3693 templates
• Keep triggers with SNR 6.5
• Vetoes
• Using Allen's??2 in Merlino (15 frequency bands)
• Testing Ochsen-Shawhan time-domain veto on MBTA
  and Merlino events

13
MBTA SNR distribution

• Using only science mode segments
• Exclude first, last minute of each segment
• Detector actions vetoed yet SNR extends up to
  200
• Some injection events much stronger than expected
• Large SNR tail fully associated with hardware
  injections?

14
High power period

• During part of the run, coil drivers switched to
  a lock acquisition mode
• More power to the coils
• Wrong calibration of the injected events! About
  26 times stronger
• Transition detected by the data analysis crew!
• Explain large SNR of part of the injected events
• Explains also an increase of control noise

15
MBTA and hardware injections (1)
triggers associated with CB injections
triggers associated with burst injections

• Tight association based on on event ending time
• 138 CB out of 200, and 48 bursts out of 640
• Some events in lists were not actually injected
• All the large SNR events are associated with
  hardware injections
• Still, the background extends up to SNR 50

16
MBTA and hardware injections (2)
triggers associated with CB injections
triggers associated with burst injections

• Try using a looser association constraint
• Associate burst events if present within the
  duration of the CB template
• Further cleaning of the distribution
• SNR of false alarm distribution extends up to 25

17
MBTA and SNR association

• SNR more or less OK when in low noise mode
• expect clusters at 7 and 14
• SNR also understood in high power mode
• Should be around 180 and 360 because of increased
  coil power
• Was a factor 0.5 lower because of the noise
  floor variation!

18
The quiet(1) period

• First night of the run
• 5 hours of very quiet science mode data
• No hardware injections yet
• Distribution cleaner, SNRmax
• The rest of the run was more interesting for
  testing vetoes

19
Stability Merlino template number evolution

• During C5 the number of templates varies as much
  as 50
• By comparison, during C4 the variation was at
  most 25
• An indication of greater variability of the shape
  of the noise spectral density (not just of its
  level)

20
MBTA and time-domain veto

• The Shawhan-Ochsen veto looks at the behaviour of
  the correlator around the peak.
• Can be tuned to veto events displaying too many
  side peaks
• First plot a false alarm event
• Second plot a true hardware injections, vetoed!
  Why?
• Injection performed during the high power period
• Coil driver not in science mode state --
  different electrical TF
• Signal injected with a significant distortion,
  detected by the veto.
• We will confirm this explanation when considering
  the ?2

21
Effectiveness of time-domain veto

• Light colors indicate vetoed events
• Some of the louder false events. including
  bursts, are rejected
• Some CB events are killed too
• Partly, in the high-power region because of
  signal distortion
• Elsewhere, because of residual injection errors?
  Need more study

22
Merlino ?2 analysis (high-power period)

• Allen's ?2 method
• C(k) k-th band correlator
• Even signal distribution over the p15 bands
• At given SNR value, r2 smaller for true signals
• Average proportional to SNR2 (locus of the
  injected events in picture)

• A clear separation of background and events is
  apparent
• The time immediately before unlocks and after
  relocks was not excluded! The r2 seems to be
  sufficient for the purpose.

23
Merlino ?2 on a distorted signal

• The plot shows the deviation, in each frequency
  band, of the signal content from the distribution
  expected.
• Example chosen during the high-noise period.
• The filtering compensates for the signal
  distortion by maximizing the template parameters
  to recover a maximum of SNR.
• The price is a deviation at low and high
  frequency.

24
Merlino ?2 during the quiet(1) period

• Green line first cut
• Kills events at very large ?2

• Blue line SNR2
• Locus of the events, for the kind of injections
  (1.4-1.4 NS pairs)
• But, no injections during this quiet period
• A cut somewhere above the blue line would
  exclude all events
• A real, robust cut would have to consider more
  general injections

25
Conclusions

• Virgo runs two partially independent CB analysis
  pipelines
• Both are complete to the point of producing
  events
• Merlino incorporates a 15 band ?2 veto,
  constituting a significant fraction of the
  computational costs.
• MBTA incorporates a built-in N band ?2 veto, not
  fully exploited.
• More work has to be done on vetoes
• The ?2 threshold needs more simulation work to be
  tuned.
• Time domain vetoes on the correlation behaviour
  are being tested and appear promising.
• C5 analysis particularly useful for veto
  development
• The high-power period was a nuisance but also an
  opportunity
• The ?2 veto was confirmed to be a useful method
  for Virgo.
• The time-domain veto could be tried on some real
  cases.