2

I have non-uniformly sampled data and I'm trying to look at its Power Spectral Density (PSD). For that, I looked at the Periodogram PSD estimate of an interpolated (with uniform sampling period) version of my data and compared it to the Lomb-Scargle PSD estimate of my original data.

enter image description here

enter image description here

I'm surprised to see that the resulting PSDs are different, especially for high frequencies, given that the interpolated data seems to follow quite well the original data. Now, I'm not sure which one should I trust!

I would appreciate it if anyone can explain why this difference and which one to use.

P.S: I found this related question For non-uniformly sampled data, how to decide whether interpolation or Lomb-Scargle periodogram is better? but no answers/comments were given.

Likely
  • 153
  • 1
  • 9
  • How did you interpolate? Is the underlying signal bandlimited? – Gillespie Jan 27 '23 at 05:09
  • @Gillespie, interpolated using 'interp1' matlab function with the linear method. I don't have a clear idea on the spectrum of the underlying signal, that's exactly why I'm trying to look at the PSD. – Likely Jan 27 '23 at 07:04
  • @Gillespie Also, this particular data I included is supposed to contain only noise. – Likely Jan 27 '23 at 07:14
  • Given that it's not bandlimited, there's no telling how aliased the signal was in the sampling process. And interpolation of nonuniformly sampled signals is complicated even when they are bandlimited: https://dsp.stackexchange.com/a/84232/55647. So I would say it's hard (impossible?) to tell which PSD is more accurate. – Gillespie Jan 27 '23 at 13:13
  • @Gillespie, Does this mean that we can't say, for sure, that Lomb-Scargle PSD gives a more accurate PSD, although it's supposed to be adequate for dealing with nonuniform sampling. – Likely Jan 27 '23 at 13:24
  • Non-zero aliasing isn't a dealbreaker for empirical methods. Your question is answerable. – OverLordGoldDragon Jan 27 '23 at 14:11
  • Also the linked question isn't "related" but a duplicate. You can set this question apart by including data/code. I'm for one unsure there's a meaningful difference in the plots since they are logscaled and that is subject to the transformation procedure, which for one isn't independent of input scaling, "input" being the output of PSD & Lomb-Scargle. I also see little high frequency content in the time-domain waveform, which justifies greater estimator uncertainty there. – OverLordGoldDragon Jan 27 '23 at 14:37

0 Answers0