San Francisco State University Physics and Astronomy Department
SFSU Pacific Oceanography Project
SFSU Private Area. WARNING: Material on this page is work in progress; results and conclusions may change without notice.
( research |PSM home |P&A Dept |SFSU |software )

Blue Whale Calls: Conclusions from exit-flag statistics.
Not yet evaluated for this analysis!

The "cheating fit" was introduced in this round of analysis. After the five fits of successively longer time series are carried out, a fit is carried out to the same time series as the fifth fit, but with an initial guess of 16.0 Hz and 0.045 Hz/sec, with the phi guess used in the first fit. If this fit converges and makes a noticible improvement on the chi-squared of fit five, its results are used and bit 5 (the sixth bit) is set.

This change has dramatically improved the success rate of the fitting process. Of 4016 events, 2291 converged through the first five fits (codes 30 or 31), and their fit was not improved by the cheating fit. Another 1205 events converged through the first five fits, byt the chi-squared was improved by the cheating fit; these were presumably events which found the wrong minumum of the chi-squared, and were saved by the cheating fit. And another 451 events failed to converge through the first five fits, and had a good cheating fit; these were presumably events which "got lost" in the fitting process. This leaves only 69 events which did not converge at all. We will need to look at f0 vs alpha plots to estimate how many of the 3947 convergent events are false triggers.

Central Events. This means near the default values f0 = 16 Hz, alpha = 0.045 . Essentially all of the events with a "cheating fit" are central. This is understandable, as the cheating fit starts at the central position.

Of the ~3500 events which converged, about 300 are not central. This may just be a tail of central events which escaped the cut.

"Good" (nchi2 < .7) events. The cut on chi-squared lost about 14% of the convergent-and-not-improved-by-cheating events. By contrast, about 40% of the events improved by cheating were lost. This must mean something - maybe they are events with a lower signal-to-noise ratio.

Good, central, convergent events. There are 3670 central convergent events (of 4016). After the nchi2 cut, there remain 2830. This is a loss of 23% due to the nchi2 cut. We will have to see if this is necessary.