# Source reconstruction of power spectral density values

**URL:** <https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810>\
**Category:** Support & Discussions\
**Created:** [April 28, 2023, 8:41am UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810 "2023-04-28T08:41:19Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![sophie](https://yyz2.discourse-cdn.com/free1/user_avatar/mne.discourse.group/sophie/32/3270_2.png) [@sophie](https://mne.discourse.group/u/sophie)\
**Post date:** [April 28, 2023, 8:41am UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810/1 "2023-04-28T08:41:19Z")

</div>

Hello forum!  
I would like to obtain the source image for a topography of power values at one selected frequency extracted from a PSD.  
The background is that we have a very specific pipeline to extract oscillatory peaks in low frequency bands, which requires first removing the 1/f slope (we use the IRASA technique), and then selecting a narrow peak frequency.

My bold attempt was to take the topography of power values (102) magnetometers, put them into a EvokedArray and source-project with dSPM. I run into scaling problems and overflow errors, which I assume come from the fact that my noise-covariance computed from an empty room recording has a scale of femtotesla, but the power values range roughly between -2 and 2.

I am aware that there are other alternatives, like source-projecting the data covariance, or computing the PSD in source space, but my goal is to visualize this particular power topography.

My questions:

- Is my bold approach wrong for reasons other than the scaling?
- If the approach can be defended, how can I overcome the scaling problem? I tried to just rescale the power values before applying the inverse. I do get results, but this feels very hacky.

Very grateful for any insights!  
Sophie

---

<div class="post-metadata">

**Author:** ![agramfort](https://avatars.discourse-cdn.com/v4/letter/a/7c8e57/32.png) [@agramfort](https://mne.discourse.group/u/agramfort)\
**Post date:** [April 28, 2023, 10:50am UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810/2 "2023-04-28T10:50:46Z")

</div>

you cannot do this. You need polarity information to have dipole locations. With the complex valued spectrum it could work but not with only power magnitude.

Alex

---

<div class="post-metadata">

**Author:** ![sophie](https://yyz2.discourse-cdn.com/free1/user_avatar/mne.discourse.group/sophie/32/3270_2.png) [@sophie](https://mne.discourse.group/u/sophie)\
**Post date:** [April 28, 2023, 11:25am UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810/3 "2023-04-28T11:25:57Z")

</div>

Thanks Alex, you confirmed my doubts.

---

<div class="post-metadata">

**Author:** ![sophie](https://yyz2.discourse-cdn.com/free1/user_avatar/mne.discourse.group/sophie/32/3270_2.png) [@sophie](https://mne.discourse.group/u/sophie)\
**Post date:** [April 28, 2023, 11:50am UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810/4 "2023-04-28T11:50:42Z")

</div>

I want to keep the 1/f removal, which is achieved by resampling the time-domain data at non-integer numbers, computing the PSD and averaging to remove narrow peaks (IRASA).  
Performing this in source space will be computationally heavy.  
Any suggestion on how to proceed?

---

<div class="post-metadata">

**Author:** ![agramfort](https://avatars.discourse-cdn.com/v4/letter/a/7c8e57/32.png) [@agramfort](https://mne.discourse.group/u/agramfort)\
**Post date:** [May 3, 2023, 12:45pm UTC](https://mne.discourse.group/t/source-reconstruction-of-power-spectral-density-values/6810/5 "2023-05-03T12:45:39Z")

</div>

I don’t know this method in details but reading what you say maybe source localize using DICS or dSPM in spectral domain after the  
“resampling time-domain data at non-integer numbers”?

Alex