Measure the noise away from the event
Take two seconds of data, subtract their mean, taper both ends, then Fourier-transform the samples. Squared Fourier amplitudes tell us how much variation occurs at each frequency. The power spectral density expresses that power per hertz. One short stretch has plenty of scatter, so slide the two-second window along nearby off-event data and average the estimates. Welch averaging in the LIGO–Virgo guide, §4.1
For GW150914, this exercise uses 25 overlapping segments from the first 14 and last 13 seconds of each 32-second file. In the published data-quality channels, all 32 seconds have data and pass the basic compact-binary quality mask. The windows purposely leave a gap around the known chirp. GWOSC technical details
Turn power into an amplitude weight
Write the one-sided noise density as Sn(f). Its square root is an amplitude spectral density. The white transform is proportional to
The scale convention of the discrete transform applies to every row the same way. Restrict it to the exercise band between 35 and 300 hertz, then transform back to get a trace in time. Outside the band the signal used by the comparison is zero. Neighboring samples remain correlated in the filtered time series.
Worked comparison
Suppose the noise power at one frequency is nine times the power at another. Equal physical strain amplitudes are then weighted in the ratio one to three after whitening. In a product of two whitened traces that becomes one to nine, which is the noise-power weight used in the matched filter.
Exercise. Why does the first 14 seconds provide more information about ordinary noise than just the sample at the event peak?
Answer. The peak mixes possible signal with noise. The off-event windows give many independent or partially overlapping views of ordinary variation without adding the strongest chirp cycles to the estimate.