Review Committee Meeting Monday 3 April 2006 09:00 PST / 12:00 EST
Minutes: Monday 3 April 2006 09:00 PST / 12:00 EST
Agenda and Contact Info
- Report on the status of the checks of timing reconstruction of the S5 untriggered analysis injections.
- Review of the GRB search S5 results [ HTML ].
- Orientation of Q-Pipeline documentation [ HTML ] and S5 technical documentation [ PDF ].
- Continue to go through SGR 1806-20 technical document [ PDF ].
AccuConference teleconferencing service: Phone: 1-800-704-9896, participant code: 038621# International callers ++1-404-920-6472 with same code
- S5 Untriggered Search Report.
- New links have been added to enote entry "WaveBurst modifications
between S4 and S5"
[ HTML ]. First item is on timing bias for
low frequency injections.
- Compare DARM_ERR timing reconstruction accuracy [ GIF ] vs. h(t) timing reconstruction accuracy [ GIF ].
- h(t) does not show any bias while DARM_ERR does. This suggests that the bias simply due to the fact that time reconstruction in DARM_ERR does not account for the time delay introduced by the transfer function.
- It is also possible that there is still an error in the construction of the DARM_ERR data ... could check simply by looking at response function to try to determine what the delay should be. Check what the expected bias is from response functions.
- Is this important for WaveBurst? Unlikely since (i) required accuracy in reconstructed time is ~100 ms while bias is only a few ms; (ii) TF tiles are around ~30 ms in duration and usual several tiles are involved in a cluster. However, any effect is simply an inefficiency in the search, and is accurately modelled by the MDCs (provided that there is no problem with MDC frames).
- Note: this issue may be present with q-pipeline too.
- Second item is sensitivity vs. frequency compared to noise curves.
[ GIF ]
- Can we predict the location of the hrss50 points? Jolien's very
rough estimate would be:
hrss50 ~ sqrt(5) * sqrt( exp(GS) ) * sqrt( S(f0) )
where S(f0) is noise power at the frequency of the SG, GS is the geometric significance threshold, assuming SNR^2=exp(GS), and the sqrt(5) is from angle averaging (though this is probably wrong). This would suggest the points should be a factor of ~10 above the noise floor for the HTA analysis with GS=3.05. They are roughly a factor of ~20 above the H2 noise floor. This simple estimate misses much subtlety about the analysis.
- Can we compare the HTA points with the LTA points (which have
GS=2.1). Would expect that the HTA points would be a factor of
sqrt( exp(GS_HTA - GS_LTA) ) ~ 1.6
above the points of the LTA. In fact, they are a factor of ~2.5 higher. Is the square-root wrong? Or is this simple analysis too simple? Need to check based on documentation.
- Erik: since we always want to be doing this check, it would be useful to have a document describing how to estimate the sensitivity of the search.
- Can we predict the location of the hrss50 points? Jolien's very rough estimate would be:
- New links have been added to enote entry "WaveBurst modifications between S4 and S5" [ HTML ]. First item is on timing bias for low frequency injections.
- S5 GRB Search Report [ HTML ]
- Automated extraction of GRB triggers from GCN. Isabel these every day.
- Most of the changes: AS_Q to DARM_ERR ; calculation of response function changed accordingly ; alpha and gamma are taken to be unity at this time. Other changes are very minor or not currently being used.
- Preliminary results [ HTML ].
- Limits seem to be in right order of magnitude. But how do we understand fluctuations? E.g., for H1-H2, sometimes the limits are much larger than at other times. Is this because (i) GRB is poorly oriented, (ii) instrument sensitivity was poor, or (iii) there was some glitch or something (perhaps a GRB!). E.g., GRB-060105 has a much higher H1-H2 upper limit than GRB-060223 (by a factor of 20). The max cc value is about the same. Corresponding probabilities are not too dissimilar. Fave LHO are not too dissimilar either: 0.308 for 060105 and 0.334 for 060223... see [ HTML ]. Expect then it is simply a matter of the noise floor at the time of the GRB. It would be helpful to have some measure of this quantity.
- Compare H1-L1 vs. H2-L1 upper limits when both are available. Here, angle factors are the same. For similar values of max cc (and corresponding factors) would expect that the H2-L1 upper limits should be a factor of 2 (?) higher than H1-L1 upper limits. But this does not seem to be so. E.g., GRB-060306. Is this something to do with using the normalized cc statistic rather than the unnormalized cc statistic? Need a way to estimate senstivity of search.
- Comparison of cc values for two different data conditioning methods. These are now posted on the data conditioning sanity check page [ HTML ] where it says: "The largest crosscorrelations found in the on-source segments of GRB 050223, 050306, 050318, and 050319, respectively, are: 0.353, 0.377, 0.365, 0.395 for the actual analysis, and 0.369, 0.333, 0.361, 0.385 for the conventional data conditioning (posted 3/31/06)." Here "real search" is Isabel's data conditioning scheme while the "conventional data conditioning" is using the Butterworth HP filter and the LPF. There does not seem to be any strong advantage of one scheme over the other based on this.
- Binomial statistic. OK for APS, but need to justify this choice. There was testing of various alternatives, but this has not been recorded anywhere. Need a document describing choice of statistical method. Something like producing a population of weak (not individually detectable) GRB signals and showing the effect on the cumulative distribution would help motivate the choice of statistics.
- Next week we'll discuss Q-Pipeline and SGR analyses.