Back to the main page.

Bug 833 - minimumnormestimate should work with lambda in an interpretable range

Reported 2011-07-19 14:48:00 +0200
Modified 2012-03-21 15:42:55 +0100
Product: FieldTrip
Component: inverse
Version: unspecified
Hardware: All
Operating System: All
Importance: P1 enhancement
Assigned to: Jan-Mathijs Schoffelen
Depends on:
See also:

Jan-Mathijs Schoffelen - 2011-07-19 14:48:27 +0200

-Right now, the influence of the lambda is determined by the magnitude of the noise covariance matrix, and the magnitude of the gram-matrix (including the source priors). -Check how this is done in the MNE-suite Implement something similar

Jan-Mathijs Schoffelen - 2012-02-06 16:48:27 +0100

MNE-suite allows for the spatial filter to be computed based on prewhitened leadfields, where the leadfields are prewhitened with an estimate of C^-1/2 where C is the noise covariance estimate. In a next step, the source covariance estimate is tailored such that the trace of the A*R*A' matrix is the same as the trace of I, where A is the prewhitened leadfield matrix and R the source covariance matrix. As a consequence of the prewhitening, the lambda parameter can take 'normal' values. According to the documentation can be taken as ~ 1/SNR I implement this in minimumnormestimate as: cfg.mne.prewhiten = 'yes';

Jan-Mathijs Schoffelen - 2012-02-07 16:17:13 +0100


Jan-Mathijs Schoffelen - 2012-02-10 09:02:46 +0100

Hi Lilla, I added an option cfg.mne.scalesourcecov, which now by default still is switched off. When you specify cfg.mne.scalesourcecov = 'yes', one can use 'meaningful' values for lambda. Let's discuss during FT-meeting whether this will be made default.