Edge effects and bandpassing at the post-process

[saleem-muhammed]

Historically, DeepClean noise prediction has been bandpassed to the target band to avoid any additional noise content being added outside the range of interest. A huge fraction of the affected edge is caused by this step. In the earlier version, the deepclean-prod, this has been proved to be a necessary step. However, in deepcleanv2, there are differences in the way the noise prediction is made and also the order in which the normalization and bandpass are applied in the preprocessing.

1. Noise prediction is bandpassed before being subtracted off the original strain

• the plot shows the difference between two DeepClean CLEANED strain both between the same start and end time with a duration of 3 s.
• Strain 1 is cropped from a 128 s long cleaned segment where the 3 s lies at the middle (62 s to 65 s) such that it is not on an edge.
• Strain 2 is a 3 s long cleaned segment such that there are edges on this.
• The difference appears to be non-zero except the 1 s (or slightly longer) in the middle

The difference in the raw predictions (before bandpass and descaling) from the same two analysis are shown below:

Here, the affected edge is less than 0.02 s which means that the bandpass enhances the edge effects.

Zoomed plots shown below:

Remove bandpassing step from the post-processing, what happens?

The difference in the two noise predictions is now zero for a longer duration as we saw for the raw prediction in the previous plot.

Does the asd ratio differ by avoiding the bandpass step?

However, the frequencies outside the target band is affected if we do not apply the bandpass

Though I initially thought we can avoid bandpass in the postprocessing, it seems a necessary step that we can not avoid!