# compute\_covariance

**URL:** <https://mne.discourse.group/t/compute-covariance/935>\
**Category:** Mailing List Archive (read-only)\
**Tags:** list-archive\
**Created:** [December 9, 2015, 10:00pm UTC](https://mne.discourse.group/t/compute-covariance/935 "2015-12-09T22:00:16Z")\
**Posts on this page:** 7\
**Page:** 1

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 9, 2015, 10:00pm UTC](https://mne.discourse.group/t/compute-covariance/935/1 "2015-12-09T22:00:16Z")

</div>

Hi,  
This is probably a dummy question but to assemble the inverse operator in  
the older versions of mne\_python, one had to first compute the noise  
covariance then regularize it manually by assigning mag grad and eeg  
regularisation factors. the defaults are 0.1 here  
\<[http://martinos.org/mne/stable/generated/mne.cov.regularize.html?highlight=regularize#mne.cov.regularize](http://martinos.org/mne/stable/generated/mne.cov.regularize.html?highlight=regularize#mne.cov.regularize)\>  
.  
At the moment the automated paradigm computes the noise covariance and  
regularizes in the same framework by cross-validation. the defaults of  
regularisation factors for empirical approach and fixed diagonal are much  
less than the old ones (0.01, 0.01 and 0.0 for grad, mag and eeg  
respectively).

I had a look at the new cov.py code to see if can see some scalings that  
could explain why the new regularisation uses smaller defaults but seems  
that both the old and new versions use the same "regularize" function for  
empirical and fixed diagonal.

I was wondering if you could please let me know where am I making a  
mistake? and if I'm not making a mistake then why the regularisation  
factors are so different in the old and new versions?

Many thanks,  
Rezvan  
-------------- next part --------------  
An HTML attachment was scrubbed...  
URL: [http://mail.nmr.mgh.harvard.edu/pipermail/mne\_analysis/attachments/20151209/89296bdf/attachment.html](http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20151209/89296bdf/attachment.html)

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 9:22am UTC](https://mne.discourse.group/t/compute-covariance/935/2 "2015-12-10T09:22:44Z")

</div>

hi,

> This is probably a dummy question but to assemble the inverse operator in  
> the older versions of mne\_python, one had to first compute the noise  
> covariance then regularize it manually by assigning mag grad and eeg  
> regularisation factors. the defaults are 0.1 here.

correct. You can still do this. Although when working on this with  
Denis we realized  
that the 0.1 value is typically too high. This leads to smaller  
whitened data and for example weaker dSPM/sLORETA values.

> At the moment the automated paradigm computes the noise covariance and  
> regularizes in the same framework by cross-validation. the defaults of  
> regularisation factors for empirical approach and fixed diagonal are much  
> less than the old ones (0.01, 0.01 and 0.0 for grad, mag and eeg  
> respectively).

when doing cross-validation the parameters are optimized for your data.

you can use the evoked.plot\_white method to check your whitening quality.

> I had a look at the new cov.py code to see if can see some scalings that  
> could explain why the new regularisation uses smaller defaults but seems  
> that both the old and new versions use the same "regularize" function for  
> empirical and fixed diagonal.
> 
> I was wondering if you could please let me know where am I making a mistake?  
> and if I'm not making a mistake then why the regularisation factors are so  
> different in the old and new versions?

check with evoked.plot\_white the quality of your noise covariances  
(old and new ones).

you can share the plots if you need feedback.

HTH  
Alex

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 9:58am UTC](https://mne.discourse.group/t/compute-covariance/935/3 "2015-12-10T09:58:44Z")

</div>

Sorry for coming late, and hi Rezvan,

> hi,
> 
> \> This is probably a dummy question but to assemble the inverse operator in  
> \> the older versions of mne\_python, one had to first compute the noise  
> \> covariance then regularize it manually by assigning mag grad and eeg  
> \> regularisation factors. the defaults are 0.1 here.
> 
> correct. You can still do this. Although when working on this with  
> Denis we realized  
> that the 0.1 value is typically too high. This leads to smaller  
> whitened data and for example weaker dSPM/sLORETA values.
> 
> \> At the moment the automated paradigm computes the noise covariance and  
> \> regularizes in the same framework by cross-validation. the defaults of  
> \> regularisation factors for empirical approach and fixed diagonal are much  
> \> less than the old ones (0.01, 0.01 and 0.0 for grad, mag and eeg  
> \> respectively).
> 
> when doing cross-validation the parameters are optimized for your data.
> 
> you can use the evoked.plot\_white method to check your whitening quality.

The whole idea is that here we cannot know in advance what is the best. You  
improve your chance of getting closer to the optimal solution by using  
different regularization strategies. I would recommend you to also use  
'shrunk', it will in fact learn use a grid search to determine the best  
regularization values and then is compared against other solutions. So far  
with 'diagonal\_fixed' no tuning is involved, same for 'empricial'. A  
minimal and fast way of optimizing reguralization would be ['empricial',  
'diagonal\_fixed', 'shrunk'] where you pass your guess for diagonal\_fixed to  
the method\_params, e.g.

method\_params=dict(diagonal\_fixed=dict(grad=0.01, mag=0.01, eeg=0.0))

This way you can compare no regularization, your guess and the grid search.  
The function will return you the covariance estimator that fits best unseen  
data. If you set `return_estimators` to True you will get a list ordered by  
fit. You can then pass this to evoked.plot\_white and inspect model  
violations for your estimators as Alex pointed out.

> \> I had a look at the new cov.py code to see if can see some scalings that  
> \> could explain why the new regularisation uses smaller defaults but seems  
> \> that both the old and new versions use the same "regularize" function for  
> \> empirical and fixed diagonal.  
> \>  
> \> I was wondering if you could please let me know where am I making a  
> mistake?

no mistake

> \> and if I'm not making a mistake then why the regularisation factors are  
> so  
> \> different in the old and new versions?

I hope it's clearer now. These defaults have never been very informed, they  
are just-so guesses that yield ok results on data we have seen. They are  
not authoritative by any means.

Cheers,  
Denis

> check with evoked.plot\_white the quality of your noise covariances  
> (old and new ones).
> 
> you can share the plots if you need feedback.
> 
> HTH  
> Alex  
> \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_  
> Mne\_analysis mailing list  
> Mne\_analysis at nmr.mgh.harvard.edu  
> [Mne\_analysis Info Page](https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis)
> 
> The information in this e-mail is intended only for the person to whom it  
> is  
> addressed. If you believe this e-mail was sent to you in error and the  
> e-mail  
> contains patient information, please contact the Partners Compliance  
> HelpLine at  
> [MyComplianceReport.com: Compliance and Ethics Reporting](http://www.partners.org/complianceline) . If the e-mail was sent to you in  
> error  
> but does not contain patient information, please contact the sender and  
> properly  
> dispose of the e-mail.

-------------- next part --------------  
An HTML attachment was scrubbed...  
URL: [http://mail.nmr.mgh.harvard.edu/pipermail/mne\_analysis/attachments/20151210/5440e023/attachment.html](http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20151210/5440e023/attachment.html)

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 10:39am UTC](https://mne.discourse.group/t/compute-covariance/935/4 "2015-12-10T10:39:15Z")

</div>

Thanks a lot Alex and Denis, very helpful.  
Denis, yes, I was using the three options as you suggested but then it  
occurred to me that my old guessed regularisation is somehow performing  
better for my data (I'm using plain MNE for source estimation), then  
checked the code and saw that the defaults for fixed diagonal is different  
and thought this might be the reason.  
I guess smaller values are usually better but will check the whitened data  
and returned estimators to see why it may not be the case for my data.

Many thanks,  
Rezvan

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 1:01pm UTC](https://mne.discourse.group/t/compute-covariance/935/5 "2015-12-10T13:01:18Z")

</div>

Hi Rezvan, I'm not sure I brought across the point entirely.

By using the `method_params` in the case of diagonal\_fixed you can use your  
preferred regularization parameters-- see little example in my email-- and  
skip the defaults. It would be equivalent to calling the regularize  
function with the old parameters, just that you can expose the resulting  
covariance estimation to quantitative comparisons.  
If you then use other estimators by passing e.g. method=['empirical',  
'diagonal\_fixed', 'shrunk'] a cross validation will tell you which one best  
captures the statistics of new data. As the negative log-likelihood of the  
covariance is used for that it is a direct quantitative comparison that in  
principle should receive more respect than any visual comparison. It may  
then turn out that your preferred guess is ok on some subjects but not for  
others.

I hope it's getting clearer,  
Cheers, Denis

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 1:33pm UTC](https://mne.discourse.group/t/compute-covariance/935/6 "2015-12-10T13:33:09Z")

</div>

Sorry Denis, my email was unclear, your point was clear and I got it 🙂  
I meant before you add the new regularisation scheme to the mne python I  
had used regularisation parameters of 0.1, then you introduced the new  
method and I used it with the new defaults. And noticed the old defaults  
were probably working better, but thought to email first to make sure that  
you have not used some sort of scaling in the new function that I have  
missed and would make using regularisation factors of around 0.1 with the  
new function inappropriate. Your and Alex's explanations indeed made things  
clear, thanks a lot.

---

<div class="post-metadata">

**Author:** ![system](https://global.discourse-cdn.com/free1/uploads/mne/original/1X/85cc6bd2b69cb698a166dc6d880fb550510d0144.jpeg) [@system](https://mne.discourse.group/u/system)\
**Post date:** [December 10, 2015, 1:36pm UTC](https://mne.discourse.group/t/compute-covariance/935/7 "2015-12-10T13:36:08Z")

</div>

Ok I believe you 🙂  
Just to be safe use the comparison functions that you get for free. You can  
then compare your guess against nothing VS auto-reg.  
Happy MNE-ing.

Denis