Minutes of 2008-May-01 S5 QPipeline Review Teleconference
Shourov Chatterji, Jonah Kanner, Isabel Leonor, Dave Reitze
Minutes by Isabel Leonor.
Dave's list of questions for the review
- All of the links containing Q-pipeline reference material are broken. These need to be fixed!!!
- Code Review Isabelle and other code reviewers have reviewed and approved code. Code freeze put in place? Date? Any bugs found since then? Any re-review needed?
- Technical Documentation are the pipeline diagrams and other documentation at http://ldas-jobs.ligo.caltech.edu/~qonline/s5/1.0/documentation /technical/ up-to-date?
- Follow up on poorly localized bursts clustering studies have been done on the playground data, and do not seem to help the FAR. There was some discussion on doing a follow up on S5 data- For the future, the effects of clustering on different waveforms have to be studied in detail and Clustering can be revisited for S5, although it is not expected to give a big increase in sensitivity. I cant recall the arguments, might be worth going over this more.
- What is the limit of calibration uncertainty that Q-pipeline can handle? MDCs suggest that 10% is OK, but what about 15%, 20%? These are not unrealistic values based on our experience in LIGO.
- At some point, we talked about Q-pipeline running the S4 data set with L1 as a sanity check. How much effort (time, computing resources) would be needed? Would it be worth it?
- Playground data should plot number of events in addition to false alarm rate so that we can do the statistics ourselves.
- Time lags did you ever run a time lag of e or pi or the golden mean, or some other irrational number on the S5 production data?
- Were all the efficiencies and other tests done using linearly polarized injections?
Jonah's list of questions for the review
- How does the review committee deal with the 2nd year analysis? Several issues may change: Clustering, Omega pipeline (including scan over sky grid), Virgo, Significance remapping.... Does this review continue, is a "new" review begun (possibly with some of the same committee members), other scenarios....
- Do we understand the issues that had come up with the whitening filter? Do we understand how those issues were dealt with? Do we understand how H1H2 coherent/null streams are actually computed?
- Do we understand the "inflation factor" for null stream veto?
- Were there any bugs when opening the box? Right segments, DQ flags, network, non-clustered triggers, lag ?
- Is there an idea of how the H1H2 coherent/null stream analysis compares with an H1H2 coincidence analysis?
- Is there a "physical" explanation for the funny shape of the three IFO cut
- Is there still an interest in non-integer time lags?
The QPipeline functionality has now been superceded by the Omega Pipeline, and all new development work is focused on the Omega Pipeline. The Omega Pipeline top page is here.
S5 2nd year / VSR1 plans
The following is a brief outline of plans for the S5 2nd year / VSR1 search. It starts with the baseline similar to the 1st year analysis and lists extension in decreasing order of priority given the time constraints.
* Baseline approach similar to S5 1 year analysis - coherent analysis of H1H2 data - independent analysis of L1 and V1 data - time-frequency coincidence of H1H2 with L1 and V1 data - networks: H1H2, H1H2L1, H1H2V1, H1H2L1V1 (all networks including both H1 and H2, but not G1) - same signal space, 48 - 2048 Hz, Q from sqrt(11) to 100 (perhaps increase Q range to slightly to include high Q injections, since these have Q of 141 in QPipeline language) - use same 5 MDC sets from 1st year (HSG1, HSG5, HSG6, HGA2, HWNB2) - same set of scale of 23 scale factors from rerun of 1st year (since we may want to combine results?) - tuning on JW1 * QPipeline functionality superceded by Omega Pipeline. - plan to run Omega Pipeline - includes QPipeline code - easy to validate vs. QPipeline reference results - more people contributing to effort - should make things go faster * Definite extensions - working now and require little effort to include - minor correction to coherent and incoherent signal energy computation - clustering * Planned extensions - not available yet, but expected soon - can be treated as optional if time does not allow - coherent follow-up using all available detectors - likelihood based definition of effective significance * Possible extensions - lower priority - only pursue if time allows - networks not including H2, particularly H1L1V1
* We went through Dave's list of issues (refer to agenda above). The numbering below follows the same numbering of the list. 1) Shourov: Most links have been fixed. 2) Shourov: Code still frozen; no changes made. The only issue is that sometimes the calculation of incoherent energies results in negative values. This will (has been?) fixed for the S5 second year analysis. Isabel: Have been re-reviewing new code. The code looks good so far; only a few questions for Shourov about a few lines of code. But need more time (~week) to finish looking at some more code. About post-processing code, Isabel has not looked at this at all, and does not know if anyone besides Shourov has looked at it. Shourov: Leo has looked at it also. Also, some sanity checks have been done, i.e. comparison of output triggers with those of KleineWelle. 3) Shourov: Will clean up documentation. 4) Dave: Re clustering, is it worth doing this? Dave thinks answer is no. Shourov: No clustering for first year. Does not add orders of magnitude improvement. White noise bursts (WNB) can benefit slightly. Maybe implement for second year. Jonah: Did you implement clustering for playground data, though? Shourov: Not sure. There is a paper about clustering study written with Rubab Khan. Study showed noticeable improvement for non-localized bursts, i.e. WNB and inspirals. But only used single-detector data. There was order of magnitude reduction in false alarm rate (FAR) to get the same efficiency. High-Q sine-gaussians, e.g. Q=100, can also improve. These are on the table for second-year analysis. Jonah: Can you include link to paper in the documentation? Shourov: Shows us where to find copy of paper and points out some plots: http://ldas-jobs.ligo.caltech.edu/~shourov/clustering_paper.pdf Also, sample plots of clustering results can be seen here: http://ldas-jobs.ligo.caltech.edu/~shourov/omega/trunk/src/example/ Shourov: Clustering does not do as well for sine-gaussians. Jonah: Out of curiosity, if you see problems in first year analysis, what's being done to fix these for year two? 5) Dave: How do large calibration uncertainties affect efficiencies and things like that? Shourov: The energy cut chosen accounts for calibration uncertainty, i.e. 0.05 > 0.2^2 (assuming 20% uncertainty). Jonah: But this is for amplitude. What about H1H2 phase issue? Shourov: Not yet tested. Previously proposed including phase uncertainty in MDC simulations. This has been implemented by Keith Thorne, via h(t)->DARM->h(t) with error in alpha, beta. Shourov hopes to re-analyze these simulations before we can call first year analysis complete, especially at high-amplitude injections. Shourov: Another issue is DC offset in H1H2. Dave: According to Raab, there is a 1.5-Mpc hit on H1. Not just DC, some effect comes from high frequency. Shourov: But it's good to hear that photon calibration now agrees with standard calibration. Dave: Yes, to within ~5%. 6) Dave: Does not think rerunning on S4 data is worth doing, but put it down for discussion anyway. Shourov: S4 H1H2 was done, but not L1. S4 is small data set, but organization will still be needed. Jonah: What was comparison with WaveBurst? Shourov: Sergei plotted comparison. Dave: Circulated Sergei's email with comparison plot. QPipeline does better than WaveBurst at low Q. All: Comparison seems to make sense. Dave: So scratch item #6. 7) Dave: Plot number of events in addition to rates. Shourov: This can be done. Background has to be normalized so that it effectively has the same livetime as foreground. 8) Jonah: Use one or two irrational numbers for time lag as a sanity check, if it's easy to do. Shourov: Cheap for H1H2 vs. L1, more work for H1-H2. Will do H1H2 vs. L1. 9) Dave: Were injections limited to linearly polarized waves? Shourov: White noise burst injections had power in h+ and hx, so unpolarized. Gaussian, sine-gaussian injections were all linear. Shourov, Jonah, Dave: Need to ask burst group leaders to have simulations with non-linearly polarized waveforms as part of standard burst MDC set. * We then very quickly summarized the previous discussion of issues in Jonah's list. The items below are numbered the same way as the list in the agenda. Jonah recalls the following summary: 1) We (this particular review committee) will not deal with second year analysis. 2) Overlaps with improved documentation. 3) Address this in documentation also. 4) Postprocessing validation has to be done. 5) Not important. 6) Has been discussed already in this telecon. * We then listed action items which need to be done. This list appears below in "Action items" section. * We all agree to have a telecon next week.
- Fix broken documentation links
- Cleanup, update, and fix technical documentation
- Links to discussion of clustering
- MDC caliibration uncertainty study
- Plot statistics in terms of event number in addition to event rate
- Run H1H2 vs. L1 time lags at irrational time steps
- Improve documentation of whitening filter issues
- Document how vetoed area increases with null stream significance
- Complete code review
- Validation of post-processing code: compare with other codes?
- Request non-linearly polarized injections from the burst group