TFR: log transform & relative frequency normalization

Dear Mike and group members,

I’m doing TFR analysis using morlet wavelets in the alpha band (8:1:12 Hz, 5-cycle wavelet) in 1-sec epochs (plus zero padding). I would really appreciate to know what you think on the following questions with regards to the treatment of the power values after the time-frequency calculation:


  1. Is log10 transform beneficial for removal of the 1/f?
  2. Is it best to do log10(alpha_power) or log10(alpha_power / baseline)?
  3. Is it ok to apply log10 on a time point by time point basis (or better to first average over a time window of interest)?


  1. Is this a good way to normalize: power at a time point at a freq point / mean power at that time point over all freq points
  2. Is it ok to do this normalization time point by time point or better to average over a time window first?

Finally, in which order is it best to apply log10 and relative frequency normalization (which one first)?

Thank you very much in advance for your time and help,

Best wishes,

Hi Ioanna. For your first set of questions?

  1. log10 transform is a good idea for several reasons. Removing 1/f is one of them. But if you are only extracting 8-12 Hz activity, there isn’t any meaningful 1/f addition to worry about. So in your case, I would say you could do baseline normalization if that makes sense if your task, but don’t worry about the 1/f.
  2. Normalizing by baseline power is great, if you have a suitable baseline.
  3. Yes! That’s the normal way to do it.

As for relative frequencies:

  1. People do this sometimes with resting-state data and with neurofeedback training. I’m not personally a fan of this approach, because the resulting values actually mix multiple different data features (across the different frequency bands). But it can be useful in some situation.
  2. Also yes :wink: