[Eeglablist] Critical pitfall of spectral power analysis?

Wed Sep 24 13:34:45 PDT 2025

Hi Mate,

You're right, sorry, I did that too fast without double-checking everything. Fig. 2 in its current form is confusing. I mixed a true alpha ERD with broadband offset/slope changes, so the GLM intercept doesn’t cleanly reflect the oscillatory decrease. With the rough log–log fits I used for offset and slope (and with baseline alpha power as a predictor), collinearity means variance can bleed into the intercept, making it look like an increase instead of the intended ERD.

So this demo doesn't show that the GLM “recovers” the ground truth yet. It just illustrates (1) conventional dB scaling warps the apparent alpha change when broadband shifts differ between baseline and post, and (2) GLM offers a framework to explicitly partition variance between oscillatory and aperiodic components rather than committing to subtract/divide a fit. The diagnostics in Fig. 3 show how offset and exponent covariates explain away much of the broadband bias, even if the intercept isn’t done perfectly yet.

I fully agree, next step should be to simulate a fixed alpha change across conditions with different broadband backgrounds and check whether the GLM intercept remains stable (and how it differs with dB). that would better demonstrate the added value (the GLM should give a consistent intercept while dB scaling would not). I also plan to rerun with IRASA- or FOOOF-derived parameters instead of quick log–log fits, which should reduce the artifacts we see here.

I'll update the page as soon as possible and will get back to you and Makoto out privately to move this into a more formal collaboration and spare the list from the high number of emails.
Thanks,

Cedric

On Tuesday, September 23rd, 2025 at 3:53 PM, Gyurkovics, Mate <mategy at illinois.edu> wrote:

> Hi both,
>
> thanks for the explanations. I checked out the github pages too, although I have not had time to look at the code in too much detail for either one of them.
>
> First and foremost, I wanted to say that I would absolutely be interested in a collaboration on either one of these (or both of these) projects, as they are quite close to my interests. We can also take this discussion off of eeglablist to a) speed it up and b) declutter everyone else's inbox.
>
> Makoto, I would also be happy to take you up on your offer to have an Zoom discussion about this. I do get the very nice animation you shared but I'm not sure all the pieces have fallen into place for me yet. Does "high-frequency input" in your point 3 simply refer to spikes coming in in rapid succession from the presynaptic neuron?
>
> If yes, it seems to me that it is the freq of the input that is a key thing here - high frequency input will be less susceptible to the low-pass filtering effect (because the return current will be closer to the location of the input, right?), whereas slower input (more spaced out spikes?) will lead to membrane potential changes that are more affected by low-pass filtering (i.e., purple line in the simulation). Cedric's summary of the idea makes more intuitive sense but it seemed like your explanation involved some extra steps.
>
> My initial reading of your post was simply that faster fluctuations (e.g., the blue lines in the simulation, which look like AMPA spikes) lead to higher freq power whereas slower fluctuations (e.g., the purple lines looking more like GABA) will lead to more low freq power, which is basically what Gao et al. are saying as well, just without a reference to the distance of the return current from the input.
>
> Another important related question for me, as I'm interested in dynamic changes in 1/f shape, is that how would this model explain for instance post-stimulus changes in 1/f shape.
>
> Cedric: thanks for sharing your Github page too. I think my brain is a bit tired at the moment because I'm not sure I'm clearly seeing the stated benefits here. Fig. 1 does not seem to capture the problem - both difference spectra show a dip in alpha and a broadband increase (albeit the shape of the broadband increase is quite different across the two bottom plots). Fig. 2 then seems to show an increase in GLM-adjusted alpha power, which is the exact opposite of the ground truth. So I'm guessing I'm misinterpreting what Fig. 2 is meant to be showing?
>
> In any case, the problem that you are trying to tackle here seems to be a bit different than what we were interested in. Specifically, what if you had the same post-stimulus oscillatory change added on top of different levels of background activity (so e.g., an alpha increase of the same magnitude but added to spectra with a higher offset in one condition and a lower offset in another). Would your method recover an alpha change, regardless of whether you are putting in linear or dB units (!) that is the same across these conditions, or would there be an illusory difference between conditions in one of these cases (e.g., with dB units).
>
> Thanks,
> Mate
>
> ---------------------------------------------------------------
>
> Feladó: eeglablist <eeglablist-bounces at sccn.ucsd.edu>, meghatalmazó: Cedric Cannard via eeglablist <eeglablist at sccn.ucsd.edu>
> Elküldve: 2025. szeptember 22., hétfő 19:55
> Címzett: EEGLAB List <eeglablist at sccn.ucsd.edu>
> Tárgy: Re: [Eeglablist] Critical pitfall of spectral power analysis?
>
> Hi Mate,
>
> Here is the link again: https://urldefense.com/v3/__https://github.com/amisepa/eeg_glm_aperiodic_covariate__;!!Mih3wA!F2cZ4zA9VpAtE6GYKa23GcQU9JRsRySAXn4LlZSLQG7qO-lZHwrkGd0wYTHpGwGnaiSUkSlRROgoTE4ULCd-usxHoQ$
> And yes, that was a direct attempt to provide a solution to the problem pointed out in your paper.
>
> I simulated spectra with both narrowband (alpha) and aperiodic (offset, slope) changes. Instead of subtracting an aperiodic fit, I regressed the post–baseline change in power on estimated aperiodic parameters and baseline power. The intercept then captures the residual oscillatory change after accounting for broadband variance, while the predictors show how offset, slope, and baseline power relate to the effect. This way the model can handle both additive and multiplicative contributions, rather than just subtracting a curve as in the traditional baseline removal.
>
> For speed I used a simple log–log regression to estimate offset and slope, but the goal is to redo this with IRASA- or FOOOF-derived parameters. IRASA, as you note, leaves it to the researcher whether to subtract or divide the fractal estimate; the GLM will avoid that decision by using these parameters as covariates and letting the model partition the variance.
>
> I ran the simulation in both log and linear power space to show how results differ depending on the scaling. The outputs compare the GLM intercept with conventional dB scaling, so you can see when the classic approach misestimates the true change. I’m hoping to develop this into a more formal paper and would welcome feedback or collaboration.
>
> Makoto, impressive simulation!
>
> Do I understand correctly:
>
> -
>
> In the passive scenario, the neuron gets inputs that don’t reach firing threshold, so no spikes occur. Instead, excitatory and inhibitory inputs both create small return currents in the dendrites. Because dendrites act like leaky cables, fast fluctuations die out more quickly with distance, while slower ones survive. Together with the naturally slower time course of inhibitory signals, this makes the spectrum tilt toward slow activity, giving a steeper 1/f slope.
>
> -
>
> In the active scenario, excitatory inputs are strong enough to trigger spikes. Spikes send large currents through the whole neuron, which add broadband, high-frequency power on top of the slower background from passive inputs.
>
> This fits with the E/I balance idea (Gao et al.): ESPSs contribute faster activity, ISPSs contribute slower activity, dendritic filtering emphasizes the slower parts even more, and when depolarizations occur, the spikes add extra broadband high-frequency power. So in sum, this filtering comes from the biophysics of neurons themselves, not from volume conduction in the skull, which doesn’t filter by frequency.
>
> Mate, I love your idea of comparing MEG and EEG (since EEG is more sensitive to extracellular return currents and MEG to intracellular axial currents flowing along the dendrites). We could see the different slope/offset patterns between the two directly.
>
> Cedric Cannard
>
> On Saturday, September 20th, 2025 at 7:53 AM, Gyurkovics, Mate <mategy at illinois.edu> wrote:
>
>> Dear all,
>>
>> sorry, I was a disengaged from emails for a couple of days. It sounds like in the meantime there are now at least two very interesting strands of simulations being run, one on comparing the E:I balance idea and the dendrite-length idea of 1/f generation, and another looking at the removal of the aperiodic component from the full spectrum. I think both of these are exciting and I'm happy to contribute in any way I can.
>>
>> @Makoto: As for the first, regarding the generative model of 1/f, I can easily see both of these mechanisms contributing to surfcae-level aperiodic components. I have not had the chance to look at your simulations yet, Makoto, but I'll be curious to see how you model these effects, and specifically how they might contribute to the signal measured on the outside of the skull. I think that is potentially the shakiest part of the E:I balance idea (i.e., that the framework works well for intracranial data, but not necessarily for scal-derived). To be fair, it was not (primarily) formulated to explain M/EEG data anyway.
>>
>> A propos, M/EEG data, I have some simultaneous MEG/EEG recordings, so we could easily compare 1/f dynamics across modalities. I wonder if the two ideas you describe make differential predictions for these two modalities with different sensitivity profiles for sources of different orientations.
>>
>> Finally, I'm not sure I follow why you refer to the E:I balance stuff as active dendritic filtering? My understanding is that the differential effects of IPSPs and EPSPs on spectral shape are derived from the different temporal decay profiles of excitatory and inhibitory signals, rather than speficially on the involvement of the dendrites. But I don't want to misrepresent the core idea here.
>>
>> @Cedric: As for Cedric's work, that is much closer to what we are discussing in Gyurkovics et al., 2021, the paper that started this whole discussion. It's still not clear to me how a GLM approach would account for both additive and multiplicative effects (we would essentially just be looking at the residuals after partialling out aperiodic parameters, which is tantamount to subtraction of the predicted signal, so the real question would still be whether we are entering logged values or not, right?). However, I'll be happy to see your simulations. It's also not clear to me what the entered predictors would be if you do not "layer it" on top of IRASA or fooof or some other method that gives you aperiodic parameters. What else would you be regressing out then, if not these estimates of aperiodic contribution?
>>
>> As a sidenote, if my memory is correct, IRASA itself is also agnostic about HOW you separate the aperiodic and periodic parts. It gives you an estimate of the former (in linear space, I think), and then you can just do whatever you want with that, subtract that from the spectrum, or divide the spectrum by that, etc. So it's still up to you as a researcher to make this decision (and we argue this is a key decision to make).
>>
>> As for the derivative, I agree, I do not see this working without massive consequences elsewhere. It should attentuate lower frequencies and thus flatten the spectrum, but I'm not sure it provides a principled way of separating aperiodic and periodic components?
>>
>> Thanks again for the exciting discussions,
>> Mate
>>
>> ---------------------------------------------------------------
>>
>> Feladó: eeglablist <eeglablist-bounces at sccn.ucsd.edu>, meghatalmazó: Cedric Cannard via eeglablist <eeglablist at sccn.ucsd.edu>
>> Elküldve: 2025. szeptember 19., péntek 22:20
>> Címzett: EEGLAB List <eeglablist at sccn.ucsd.edu>
>> Tárgy: Re: [Eeglablist] Critical pitfall of spectral power analysis?
>>
>> Dear Jinwon,
>>
>> Interesting idea. My understanding is that the temporal derivative steepens high-frequency content and reduces the dominance of low frequencies, thereby minimizing the 1/f slope (also called prewhitening). It does reduce 1/f scaling, but at a cost: oscillatory peaks can be distorted (e.g., attenuated in alpha relative to higher bands), and the signal’s statistical properties are altered in ways that complicate interpretation. It is thus a blunt tool compared to model-based separation (like FOOOF) or regression-based covariates (GLM). It can be useful in specific contexts such as highlighting sharp transients or ripples, as in Cox et al. (2017), but it is not a general solution for oscillation/aperiodic separation in EEG.
>>
>> I’ve been reading about IRASA (Wen & Liu, 2016): https://urldefense.com/v3/__https://pmc.ncbi.nlm.nih.gov/articles/PMC4706469/__;!!Mih3wA!GV4DGiGOw26Mc54x3I3afDKbEcfCSXJf-y1fFCQe3Wh5yyBMpSeMIhyZqV6k96k6h_WunX3lvkHwcnnzBaHelJDr6Q$
>>
>> This method separates fractal (aperiodic) from oscillatory components through irregular resampling. Its strength is that it does not assume a parametric form, making the decomposition data-driven and robust across datasets. The limitations, however, are (if I understand correctly):
>> - if oscillatory peaks are very strong, broad, or overlapping, separating them cleanly is difficult—resampling shifts peaks, but disentangling overlaps can remain messy.
>> - IRASA also presumes that the fractal component is scale-free; if the aperiodic background departs from a pure power law (e.g., knees, plateaus), the isolation can be imperfect.
>> - High-frequency noise (such as EMG or amplifier noise) can further interfere, particularly when spectral flattening overlaps the analysis band.
>> - because IRASA requires relatively long data segments and repeated resamplings, it sacrifices temporal specificity compared to methods designed for short windows or transient events.
>>
>> Alternatively, FOOOF offers a complementary parametric strategy, explicitly fitting the aperiodic component (offset and slope) and modeling oscillatory peaks on top. This provides clean, interpretable parameters, which is useful for group comparisons or linking spectral features to physiology. But the approach depends heavily on model quality: noisy spectra, overlapping peaks, or deviations from a pure power-law background can bias fits, and overfitting or misfitting are ongoing concerns.
>>
>> The GLM approach I proposed would complement, not replace, these methods. By treating aperiodic parameters (offset, slope, or IRASA-derived estimates) as covariates, it adjusts oscillatory estimates statistically instead of subtracting background activity. This allows additive and multiplicative contributions to be tested explicitly, and reduces the risk of conflating broadband shifts with narrowband effects. In practice, GLM can be applied directly to raw or dB spectra, or layered on top of IRASA or FOOOF outputs to provide a more rigorous inference framework.
>>
>> As a next step, I plan to extend the simulations to compare these approaches side by side (derivative, IRASA, FOOOF, and GLM) against known ground-truth oscillatory changes, and to evaluate how GLM behaves when combined with each.
>>
>> Cedric
>>
>> On Thursday, September 18th, 2025 at 7:17 PM, 장진원 via eeglablist <eeglablist at sccn.ucsd.edu> wrote:
>>
>>> Dear all,
>>>
>>> Thank you for sharing all the interesting ideas in this topic. I just
>>> wonder whether utilizing the derivative of the EEG time series to minimize
>>> 1/f scaling (Cox et al., 2017) is more or less effective than other methods
>>> like FOOOF or GLM proposed by
>>>
>>> Cedric
>>>
>>>
>>> .
>>>
>>> Best Regards,
>>> Jinwon Chang
>>>
>>> 2025년 9월 18일 (목) 오후 11:29, Makoto Miyakoshi via eeglablist <
>>> eeglablist at sccn.ucsd.edu>님이 작성:
>>>
>>> > Hello Mate and all,
>>> >
>>> > I became curious about the 1/f^n issue, so I spent a whole day today
>>> > investigating this problem.
>>> > Following Cedric's obsession, I uploaded simulation code and results to
>>> > Github, YouTube, and SCCN Wiki page.
>>> >
>>> > https://urldefense.com/v3/__https://github.com/MakotoMiyakoshi/cableTheoryDemo/tree/main__;!!Mih3wA!CORpSZkbn8mntguT3SylRvIRIqrj-8phD2Bm9zBE0rNFZZa8DMJcfa0mXv2pG2xDJfbgRregAtjdLfj53TBo4g-7dEc$
>>> >
>>> > https://urldefense.com/v3/__https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline*Where_does_power_distribution_come_from.3F_.28For_510.2C000_page_views.2C_Added_on_09.2F17.2F2025.29__;Iw!!DZ3fjg!7laPJVE-aD5lzcXtJK-VV5uu70cV0ZVdaJNWbqM88Mg2JboVI71dPCsxVYi76q1L_edm6Elk8xph9xCUtNLkbhvIYOc$
>>> >
>>> > https://urldefense.com/v3/__https://www.youtube.com/watch?v=SitmGp8LYtY__;!!Mih3wA!CORpSZkbn8mntguT3SylRvIRIqrj-8phD2Bm9zBE0rNFZZa8DMJcfa0mXv2pG2xDJfbgRregAtjdLfj53TBosE7VLyc$
>>> >
>>> > https://urldefense.com/v3/__https://www.youtube.com/watch?v=iaRQsaU1_2s__;!!Mih3wA!CORpSZkbn8mntguT3SylRvIRIqrj-8phD2Bm9zBE0rNFZZa8DMJcfa0mXv2pG2xDJfbgRregAtjdLfj53TBoDhtKxC8$
>>> >
>>> > I think the mechanism you are referring to, Makoto, is what's usually just
>>> > described as "dendritic filtering", is it not? I'm aware of this process
>>> > plus the low-pass filtering that happens as a consequence of spatial
>>> > summation.
>>> >
>>> > Yes, a dendritic filtering. But to be honest I was not aware of the fact
>>> > that there were actually two scenarios, one is passive (like EFB) and the
>>> > other is active (like Gao et al.)
>>> > I once asked Paul whether the subthreshold passive cable theory was
>>> > sufficient to explain the generative model of EEG signals. He said the
>>> > subthreshold model was 'good enough'. As I read Reimann et al. (2013) this
>>> > time, I found that the difference between passive vs. active cable is
>>> > present but probably not very critical after all. As we can easily imagine,
>>> > the active model with action potentials adds more power at higher
>>> > frequencies (in FOOOF terminology, 'flatter aperiodic') If I were an
>>> > American, I would say duh--action potentials are spikes. But Reimann and
>>> > colleagues also investigated the impact on laminer CSD patterns, which was
>>> > good to confirm.
>>> >
>>> > As my simulation confirms, we can make testable hypotheses on modulation of
>>> > 1/f-ness.
>>> > For example, selective engagement of pyramidal neurons whose soma are at L5
>>> > should show more low-pass filter effect simply because their dendrites are
>>> > longer than those that reside at L3.
>>> > Also, any preferential engagement of neural populations that are known to
>>> > use non-myelinated axons should produce more low-pass filter effects etc
>>> > etc..
>>> >
>>> > I want to see a quantitative comparison between the dendritic length model
>>> > within the framework of the passive cable theory and AMPA/GABA_A model
>>> > within the framework of Gao's action potential-based model. Is the latter
>>> > so prominently larger than the former so that we can safely forget it? It's
>>> > an empirical question.
>>> >
>>> > Makoto
>>> >
>>> > On Sun, Aug 31, 2025 at 6:30 PM Gyurkovics, Mate mategy at illinois.edu
>>> > wrote:
>>> >
>>> > > Thanks again for all the interesting points. I'm certianly learning a lot
>>> > > on the physics side - and also about the PSD in different animals, I was
>>> > > genuinely unaware of all this, but sounds super interesting.
>>> > >
>>> > > I think the mechanism you are referring to, Makoto, is what's usually
>>> > > just
>>> > > described as "dendritic filtering", is it not? I'm aware of this process
>>> > > plus the low-pass filtering that happens as a consequence of spatial
>>> > > summation. These are the two main ones I was thinking of, so you are
>>> > > right,
>>> > > it was very imprecise on my part to talk about the low-pass filtering
>>> > > properties of the tissue.
>>> > >
>>> > > As demonstrated in a classical study by Lopes da Silva and van
>>> > > Leeuwen (1977), alpha oscillation is generated within the cortex which is
>>> > > only 4-5 mm thick.
>>> > >
>>> > > Just out of curiosity, what point was this sentence supporting?
>>> > >
>>> > > I'm very happy to learn the conceptual distinction between trivial and
>>> > > non-trivial contributions to the changes of 1/f power distribution. Thank
>>> > > you Mate. Your works are impressive.
>>> > >
>>> > > Very nice of you to say this, Makoto - I would, however, also like to
>>> > > stress that this particular distinction just reflects how I personally
>>> > > think about this problem (i.e., the contribution of the ERP to the
>>> > > spectrum), I'm sure reasonable people could disagree. (Although this is
>>> > > basically the logic we published in the Journal of Neuroscience paper
>>> > > linked above.)
>>> > >
>>> > > @Eugen - you raise many interesting points. (I certainly agree that in
>>> > > scalp recordings at leat, oscillatory activity is sparse, and most
>>> > > robustly
>>> > > occurs when the brain is not engaged with a task, in the form of alpha
>>> > > activity, but that is about as bold as I can be here.)
>>> > >
>>> > > is sinusoidal (with sharp peaks in the spectrum) activity and broadband
>>> > > (not necessarily 1/f) activity generated by different mechanisms?
>>> > >
>>> > > I think, to put it very simply, this is one of the fundamental questions
>>> > > underlying this discussion here. If they are separate AND fairly
>>> > > independent, that is when the conclusions of our paper hold. If they are
>>> > > separate but interact, our conclusions will sometimes hold, other times
>>> > > maybe not. If they reflect the same underlying mechanism, our conclusions
>>> > > would likely rarely be a concern. I personally think that they likely
>>> > > reflect mechanisms that are separate at least to some extent - plus
>>> > > 1/f-like broadban activity likely reflects several, not just one,
>>> > > generative mechanisms as highlighted by Makoto and others here too.
>>> > >
>>> > > Moreover, the function 1/f may simply be the result of the Fourier
>>> > > transform of a single excitatory or inhibitory postsynaptic potential.
>>> > >
>>> > > As far as I recall, this point is covered in Gao et al. (2017), the paper
>>> > > that first linked 1/f-like features to excitation/inhibition balance.
>>> > >
>>> > > Thanks,
>>> > > Mate
>>> > >
>>> > > ------------------------------
>>> > > Feladó: Евгений Машеров emasherov at yandex.ru
>>> > > Elküldve: 2025. augusztus 30., szombat 9:07
>>> > > Címzett: Gyurkovics, Mate mategy at illinois.edu
>>> > > Másolatot kap: EEGLAB List eeglablist at sccn.ucsd.edu; Wirsing,
>>> > > Karlton
>>> > > kwirsing at vt.edu; Cedric Cannard ccannard at protonmail.com; 장진원 <
>>> > > jinwon06292 at gmail.com>; Makoto Miyakoshi mmiyakoshi at ucsd.edu
>>> > > Tárgy: Re: [Eeglablist] Critical pitfall of spectral power analysis?
>>> > >
>>> > > As a (naive and insufficiently substantiated hypothesis) — is sinusoidal
>>> > > (with sharp peaks in the spectrum) activity and broadband (not
>>> > > necessarily
>>> > > 1/f) activity generated by different mechanisms? Moreover, sinusoidal
>>> > > activity is a manifestation not of action, but of inaction of the brain.
>>> > > Physiological (alpha rhythm when closing the eyes, possibly also mu
>>> > > rhythm
>>> > > in the absence of proprioceptive signals, sleep spindles) or pathological
>>> > > (alpha coma and low-frequency sinusoids in the delta or theta ranges). It
>>> > > can be associated with the regulation of the level of constant potential
>>> > > and, in general, with metabolic processes carried out by the integral
>>> > > regulator generating oscillations (but we see oscillations directly only
>>> > > as
>>> > > an idle rhythm). Broadband activity seems to be directly associated with
>>> > > the functioning of individual neurons. Moreover, the function 1/f may
>>> > > simply be the result of the Fourier transform of a single excitatory or
>>> > > inhibitory postsynaptic potential. These mechanisms are interconnected,
>>> > > but
>>> > > different. Perhaps the mathematical apparatus for their study should also
>>> > > be different.
>>> > >
>>> > > Your truly
>>> > >
>>> > > Eugen Masherov,
>>> > > Burdenko Neurosurgery Institute
>>> > >
>>> > > > Thanks again everyone, for these very interesting points.
>>> > > >
>>> > > > Just to add to something that was said recently - yes, 1/f (or rather,
>>> > > > 1/f^x) features are quite ubiquitous, I think practically any time series
>>> > > > with some amount of autocorrelation will have a similar shape:
>>> >
>>> > https://urldefense.com/v3/__https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(14)00085-0__;!!Mih3wA!E0DB3LnNc6JHuXWc3N5M5iAUeusaX8LTN3dqbu9ZUeyzmDOwf3GVazliJU65YVxC6mC4VLMvHBOlkHKz4AeGoXXj$
>>> >
>>> > > - its ubiquity is covered nicely in this lovely paper, as far as I
>>> > > remember.
>>> > >
>>> > > > I also get most of Makoto's points about how just the location of the
>>> > > > neuronal inputs, either in terms of proximity to the soma or in terms of
>>> > > > cortical layers, will affect the strength of the low-pass filtering, and
>>> > > > thus the shape of the 1/f scaling. This is super interesting, and this
>>> > > > and
>>> > > > dendritic filtering are certainly discussed in the literature to some
>>> > > > extent. I am a bit more sceptical whether such subtle differences could
>>> > > > contribute to 1/f changes in scalp recordings, but Makoto suggests they
>>> > > > could and I trust his expertise.
>>> > > >
>>> > > > If you perform an ERP task, it would change 1/f power distribution, not
>>> > > > surprisingly, because task-triggered cortico-cortical and
>>> > > > thalamo-cortical
>>> > > > inputs are recruited.
>>> > > >
>>> > > > This is a very interesting point. In our 2021 paper linked above, we
>>> > > > also make the point that 1/f shape should change in an event-related
>>> > > > design, but for a more trivial reason: ERPs are non-oscillatory (in the
>>> > > > simple sense that they are transient bursts that do not repeat with a
>>> > > > clear
>>> > > > period), and will thus have a 1/f shape in the frequency domain (indeed,
>>> > > > they do, there are some figures in the paper). Thus, 1/f scaling will
>>> > > > change after an event trivially because there are well-known
>>> > > > non-band-limited changes happening in the EEG (the ERPs). We tried to
>>> > > > correct for the contribution of the ERPs and still found post-stimulus
>>> > > > 1/f
>>> > > > changes that we consider non-trivial (a steepening to be specific).
>>> > > > These,
>>> > > > then, could be explained by the mechanism that Makoto suggests (which we
>>> > > > did not consider in the paper, as it seemed maybe a bit small-scale to
>>> > > > explain scalp-derived effects) and/or by Gao et al.'s
>>> > > > excitation-inhibition
>>> > > > balance idea (this is the framework we used in the paper). It certainly
>>> > > > cannot be explained by the general low-pass filtering properties of the
>>> > > > tissue or similar more or less fixed variables, as those should not
>>> > > > change
>>> > > > so rapidly.
>>> > > >
>>> > > > I share much of your scepticism about oscillatory mechanisms (in scalp
>>> > > > recordings), Makoto, but if we take the most typical generative
>>> > > > mechanisms
>>> > > > assigned to these phenomena (interplay of pyramidal cells and
>>> > > > interneurons), they seem like they could potentially interact with these
>>> > > > other mechanisms described above, or be fairly independent.
>>> > > >
>>> > > > So we've got this really complex picture, where there could be
>>> > > > oscillations going on (maybe in alpha only), there could be
>>> > > > (independent?)
>>> > > > 1/f dynamics happening for multiple reasons, e.g., because of the
>>> > > > location
>>> > > > and/or the nature (E vs. I) of neuronal inputs changing, and there could
>>> > > > be
>>> > > > ERPs happening too, which might partly be phase-locked oscillations, and
>>> > > > could also be related to where the neuronal inputs are located, so they
>>> > > > "straddle" these different mechanisms quite a bit, probably. Not too sure
>>> > > > about the ERPs to be honest.
>>> > > >
>>> > > > Two more minor points:
>>> > > >
>>> > > > I can't put up with the fuzziness of how the term 'oscillation' is used
>>> > > > in the field now. Is a try-phasic burst, such as a classical
>>> > > > event-related
>>> > > > N1-P1 waveform, an oscillation?
>>> > > >
>>> > > > I agree completely that it is very unclear what constitutes an
>>> > > > oscillation - basically, how many cycles are enough for something to be
>>> > > > considered an oscillation, and how do we show that those cycles come from
>>> > > > the same generative mechanism, and not just multiple successive events
>>> > > > happening. This is less of a question for longer, more stable
>>> > > > oscillations,
>>> > > > e.g., alpha at rest.
>>> > > >
>>> > > > And as for Michael's question: my limited experience with this topic
>>> > > > would certainly suggest that 1/f dynamics (for whatever reason) could
>>> > > > change very rapidly, and often in a systematic fashion (e.g., predictably
>>> > > > after a stimulus). They also do seem to change on much slower time scales
>>> > > > as well, e.g., across the lifespan.
>>> > > >
>>> > > > Thanks,
>>> > > > Mate
>>> >
>>> > _______________________________________________
>>> > To unsubscribe, send an empty email to
>>> > eeglablist-unsubscribe at sccn.ucsd.edu or visit
>>> > https://urldefense.com/v3/__https://sccn.ucsd.edu/mailman/listinfo/eeglablist__;!!DZ3fjg!7laPJVE-aD5lzcXtJK-VV5uu70cV0ZVdaJNWbqM88Mg2JboVI71dPCsxVYi76q1L_edm6Elk8xph9xCUtNLkOtPs8sM$ .
>>>
>>> _______________________________________________
>>> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu or visit https://urldefense.com/v3/__https://sccn.ucsd.edu/mailman/listinfo/eeglablist__;!!DZ3fjg!7laPJVE-aD5lzcXtJK-VV5uu70cV0ZVdaJNWbqM88Mg2JboVI71dPCsxVYi76q1L_edm6Elk8xph9xCUtNLkOtPs8sM$ .
>> _______________________________________________
>> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu or visit https://urldefense.com/v3/__https://sccn.ucsd.edu/mailman/listinfo/eeglablist__;!!DZ3fjg!7laPJVE-aD5lzcXtJK-VV5uu70cV0ZVdaJNWbqM88Mg2JboVI71dPCsxVYi76q1L_edm6Elk8xph9xCUtNLkOtPs8sM$ .
> _______________________________________________
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu or visit https://urldefense.com/v3/__https://sccn.ucsd.edu/mailman/listinfo/eeglablist__;!!DZ3fjg!4jHy9ba8iFazM9S1DL6VY23rPG1NIYCz5wQ68gqhAuIEJB7gb-BnOzNjXY9HVzOzdUY6GHACtaxmRdEd0zdJeZdToRQ$ .