Changes for S5 Pass2 BlockNormal Processing
IntroductionFollowing the completion and documentation of the initial S5 FirstYear analysis (see web page S5 FirstYear Pass1 Summary), we have made changes in the processing to remedy deficiencies in that analysis. These changes are presented below.
Event Generation ChangesThe biggest change will be in the event generation (BlockNormal). We are changing the event significance metric to improve detection sensitivity for longer-duration bursts. We have also found that modifying the change-point statistics removes many artifacts and reduces CPU time to 10% of previous.
Event Significance MetricFollowing a study of event significance metrics (see web page Event Threshold Tests), we are changing from `relative excess power' (PR) to `excess power probability' (&LambdaE). This gives a factor of 2 improvement in sensitivity for longer-duration (> 20ms) bursts.
Change-Point StatisticsFollowing studies of the change-point statistic (see web pages Change Point Tests Dec07, Change Point Tests Jan08), we have modified the &rho2 statistic such that it now acheives the expected Poisson duration distributions and takes 10-15% of the CPU time.
Support for strain-only MDCsWe have modified our BlockNormal pipeline (which uses DARM_ERR) so that it can convert the MDC time-series from strain to DARM_ERR, using the same routines as the BurstMDC production software. This required splitting the monolithic `BNETG' routine into three parts for time-series preparation (bnetgprep), per-band event processing (bnetgproc), and post-production clean-up (bnetgpost).
Coincidence ChangesWe have also made changes in event clustering and coincidence to improved detection sensitivity of longer-duration bursts. To improve sensitivity on wide-bandwidth white-noise bursts, we are changing to merging clusters between frequency bands prior to coincidence tests. To improve sensitivity on long-duration lower-frequency Sine-Gaussians, we will be increasing the width of the coincidence windows.
Full-Bandwidth CoincidenceOur sensitivity for bursts that had different h+ and hx waveforms was quite poor in the initial processing (see web page S5A WNB2 MDC results). We incorporated a change to better cluster per-band events in the Pass1 production. However, to improve sensitivity on such bursts that spanned several frequency bands, we found that a change to merge per-IFO events across frequency bands prior to coincidence (instead of after) would be of benefit. (see web page InterBand Clustering Study). To be successful, we will have to lower the per-IFO event rate.
Wider Coincidence WindowsWe also found that our detection sensitivity on very-long-duration bursts (such as Q=100 SineGaussians) could be improved by increasing the coincidience window width (see web page Coinc Window Tests). The exact value awaits a re-analysis with the new event signficance metric, but should be about 50-70ms. We note that another pipeline (incoherent WaveBurst) was using 100ms windows for S5 FirstYear.
CorrPower ChangesIn the Pass1 processing, the bandwidth for triggers accepted by CorrPower was limited to 1600Hz due to anti-aliasing filters. However, we found that changing the low-pass filtering to use the anti-aliasing in the MATLAB `resample' function allowed the bandwidth to be expanded to the Nyquist at 2048Hz (see web page S5 HSG6 MDC New CorrPower Results).
We may also add notch filters for the out-of-band spectral features (60Hz, low-lying calibration lines) to improve the whitening filters.