coherent WaveBurst Review Meeting 11 June 2009 09:00 Pacific / 12:00 Eastern
Agenda: Thursday 11 Jun 12:00 Eastern
- Discussion of how to deal with calibration uncertainties [wiki]
Dial-in number: 1 800 704 9896 International dial-in: +1 404 920 6472 Conference code: 5374 2349 #
AttendanceVirginia Re, Marco Drago, Francesco Salemi, Patrick Sutton (minutes), Giovanni Prodi
- [ wiki ] RMS Gauss plot for SG1451Q9: black curve (set 0) is the curve including the calibratio uncertainties. Sets 1-9 are for specific fixed miscalibrations.
- FS: RMS stated from the calibration group is not variability of calibration but statement of uncertainty in the calibration; i.e., calibration fixed but unknown. This is an uncertainty in the level of knowledge. PJS: Systematic rather than statistical.
- Decision from burst F2F in Orsay: Show typical curve and error bars.
- Have 3X and 4X results ready. Need to discuss 2 issues:
- RMS error bars are over-estimates due to the way the 9 miscalibrations were selected.
- There is a problem with one set and the 2756Hz waveform, as shown in the last plot on the RMS page. The random number drawn from the Gaussian distribution for set 3 has a very bad effect on the penalty cut -- the efficiency asymptotes to 0.2! FS: Due to large relative difference between H1 and H2 for this miscalibration.
- Method of choosing miscalibrations: FS: Miscalibrations chosen as 3 different draws for three subsets of data. Subsets turn out to be non-homogeneous. SK: Comparing 3 different random draws applied to 3 different data sets. Could also be that systematic difference between data sets affects RMS. PJS: Is procedure to draw one value for each detector "i" from N(mu_i,sigma_i), test this one value on each of three sets of data, then repeat twice more? Yes.
- GP: Actually 9 draws, not 3. Also, bad draw gives factor of 2 amplitude error difference between H1 and H2. Gaby says this large an error is not plausible. Calibration people suggest calibration uncertainty is not symmetric but more one-sided.
- FS: Thinks sets 1-3 are on same period of data, ditto 4-6, 7-9. Thinks perfomance is correlated to period.
- GP (interpreting Gaby): Dominant contribution to calibration uncertainty is open-loop gain at high frequency. Due to (a) a small statistical error in measurement and (b) a systematic failure of the model of the OLG (25% at 5kHz -- dominant). Interpretation: (i) failure in actuation stage, in which case it should be ignored, or (ii) an error in the sensing of mass and therefore should be included in h(t). In latter case Gaby thinks uncertainty is somewhere in [0,>0], where the error means the actual sensitivity is better than what we think. Our estimates of effects of calibration uncertainties are therefore probably conservative (overestimates).
- FS: Data divided into 3 periods of approximately 160 segments each. Sonce non-homogeneous, scatter across these 3 overestimates RMS (since not averaging over all 3 periods, as we do in computing efficiencies for the search).
- MD: Suggest moderating effect of inhomogeneity by alternating use of draws 1,2,3 by injection rather than by long block of segments. Gives much smaller RMS using same draws. Example: Third plot on RMS Gaus page ("fix123b"). 2756 Hz waveform still has bad RMS but the others are much smaller.
- Next steps: FS: Repeat MDC using fresh draws from a more realistic model for H1-H2. Gives 9 more draws, plus drop 1 or 2 least physical draws from the present set. Should be enough -- takes a few days on the cluster.
- FS: Recall from burst F2F that the 1-sigma is not to be used in upper limit construction. Now find error bars much smaller than before. Expect 15% increase in h50 for the lowest frequency SGs.
- LFS: 11% net shift decided for bursts. HFS: Might find 6% or 15% shift for lowest-frequency sine-Gaussian; i.e., quoted error could be smaller than 11% used in LFS for same waveforms (in overlap region).