coherent WaveBurst Review Meeting Monday 22 Jan 2009 09:00 Pacific / 12:00 Eastern
Agenda: Thursday 22 Jan 2009 09:00 Pacific / 12:00 Eastern
- Code review: high-frequency search
Dial-in number: 1 800 704 9896 International dial-in: +1 404 920 6472 Conference code: 5374 2349 #
AttendancePatrick, Michele (minutes), Giovanni, Virgina, Gabriele, Romain.
HFS - Discussion of "Processing" Overview
(Section 3 on the CodeReview wiki page.)
Michele: Do you have a linear reduction of the memory usage
with the re-sampling?
Gabriele: Yes. Needed to reduce frequency range to avoid exceeding memory limit of CIT nodes.
Patrick: why 6400Hz?
Gabriele: it is a requirement of the wavelet library, which only works for certain discrete frequencies.
Patrick: what are the entries in:
resample rsm(12800, 0 ,0, 6400); // RESAMPLE
Virginia: upper frequency frequency offset starting frequency Nyquist frequency.
Michele: since you have a new version of the whitening filter,
did you check that it zero phase. meaning it does not introduces delays?
Gabriele: the changes are minimal but we checked it.
Michele: could prepare a paragraph do document the checks?
Patrick: asks to summarize the different wavelet decomposition levels in the low and high frequency searches. Giovanni pointed out that these are explained in this link from the code review page.
Patrick: Why did you choose to use a new linear predictor filter?
Virgina: Done for original version of analysis (where the low-frequency cutoff was about 1000Hz) to have a better smoothing especially for the ~5000Hz line. Did not revisit after changing to 0-6400Hz band.
Patrick: Sounds fine.
Gabriele directed us to the DownConversionEffect wiki page, which displays the results of tests of the resampling.
Gabriele: the error in the re-sampling is orders of magnitude
smaller than the noise.
Action item (reviewers): Examine resampling code and tests page more carefully offline.
Action item (Gabriele): Patrick would like to see impulse repsonse of the resampling. This FFT procedure is effectively IIR -- want to make sure effective filter length is not too long.
Virginia: Described the different configurations and parameter files.
Patrick: Are you planning to do the upper limit for H1H2 ?
Patrick: You do not separate background into tuning and UL sets for H1H2. You need a background for tuning (unless it is very small).
Giovanni: For other configurations we separated the triggers used for the tuning and those used for simulations. We did not do it for H1H2.
Patrick: That means you cannot assess the significance of the zero lag events?
Giovanni: Yes, but can't do that with time-lag background even if we reserved some for this purpose instead of using it all for tuning. Correlated zero-lag noise makes significance estimation from H1-H2 time lags unreliable.
Patrick: To gain a better estimate of the background you could use H1H2 data when another estimator is on.
Giovanni: Unfortunately the values of the background would be different for the two cases so it is not reliable. Tests show statistically significant difference between backgrounds estimated using H1-H2 only time and H1-H2 triggers in H1-H2+(other detector) time. Periodicity in effect about 1 month -- too long for us to treat it as ergodic over the year-2 search.
Virginia describes quickly the Simulations. The first set (Q9 SGs) has been analysed, and also some of the set2/3 injections. We'll cover it in more detail next week.