Good data covariance matrix

[emarca]

Dear MNE users,

I am working on real-time analysis, and I am trying to understand how many epochs are enough to estimate a “good enough” data covariance matrix for source reconstruction.
Every buffer epoch is 1000 samples. I am using a running window with a varying number of epochs, and I’m visualizing the data covariance matrix at each step. My question is: how can I determine the point at which my data covariance matrix is sufficiently reliable for my analysis?

e.g.:
data covariance matrix obtained using one 1000 samples epoch:

data covariance matrix obtained after accumulating 10 epochs:

Thank you!

[alexrockhill]

It depends on how noisy your data is and how much trial to trial variation there is. One strategy might be to set a convergence threshold such that when the covariance matrix stops changing by more than 1% for instance for some number of trials, maybe 3 or 5, that you have a justification for stopping.