add line locator (to identify f0s and bws for cleaning)

[avivajpeyi]

This is really neat repo, @farr ! Popped up when i was looking online for implementations for Lorentzian lines.

Id like to automate the process of determining the lines to be removed.

Following Gupta and Cornish '24, we can identify the locations of narrowband spectral lines using a running median of the periodogram.

Any frequency bins where the periodogram power exceeds the running median by a chosen threshold (e.g., a factor of 10) can be marked as narrowband lines.

These identified lines can then treated using your mcmc script :)

[avivajpeyi]

Examples of identified lines (f0 and bws):

[avivajpeyi]

Hmmm in some cases i guess the line-cleaner didnt work so well...

Im guessing the MCMC might not have converged or smth...

Any tips @farr, before i start looking at trace plots?

Frequency Range (Hz)	Strain Plot
26 – 29
35 – 38
40 – 42
59 – 61
119 – 121
127 – 129
179 – 181
255 – 257
299 – 304
331 – 333
499 – 510
598 – 600
604 – 608
991 – 1000

[farr]

Hi @avivajpeyi --- glad that you're looking into it! I have a few thoughts (before you go looking at trace plots):

• The line cleaner will not handle intervals where there are multiple lines---it explicitly assumes that the power spectrum of the narrow-band data it is presented with is a flat "continuum" plus exactly one Lorentzian profile. So I wouldn't expect it to do well in those cases (like, say, 35 to 38, where the lines aren't well separated, or 26 to 29).
• In other cases, it looks to me like your bandwidth is too narrow---the line cleaner will need to have a fair amount of "flat" PSD outside the line (i.e. after the line has fallen below the continuum by a sufficient amount) in order to get a good estimate of the background white-noise level (assumed white because the band is assumed to be narrow enough that the PSD doesn't evolve). I think that's the issue 171-181, and (to a lesser extent) 331-333 as well, and many other cases. As a rule of thumb, I would include at least as big a range of frequencies outside the line as the line's bandwidth itself.

Unfortunately, this will likely mean that at least some of the lines (e.g. 598-600) which aren't formally overlapping will still not be treatable with this simple model.