How the Corr matix is saved?

Francois · November 2, 2016, 7:19am

Hi Junpeng,

In the processes “Correlation 1xN”, “Correlation NxN” and “Correlation AxB”, the measure that is used is the same.
In the NxN case, it just repeats it for all the pairs of signals, there is no covariance of the other signals involved.
The results are the same as the non-diagonal values returned by the corrcoef function with only two signals in input.

The full code of the function and some additional explanations are available here:

github.com

brainstorm-tools/brainstorm3/blob/master/toolbox/connectivity/bst_corrn.m

function [R, pValues] = bst_corrn(X, Y, RemoveMean)
% BST_CORRN: Calculates the same correlation coefficients as Matlab function corrcoef (+/- rounding errors), but in a vectorized way
%            Equivalent to bst_correlation with nDelay=1 and maxDelay=0 
%
% INPUTS:
%    - X: [Nx,Nt], Nx signals varying in time
%    - Y: [Ny,Nt], Ny signals varying in time
%    - RemoveMean: If 1, removes the average of the signal before calculating the correlation
%                  If 0, computes a scalar product instead of a correlation
%
% NOTE: The rounding errors
%    Corrcoef computes the correlation coefficients based on the variance values computed with cov(),
%    instead of a direct sum of the squared values (sum(Xc.^2,2)).
%    Hence it uses a corrected algorithm for the computation of the variance, that is not sensible to
%    the rounding errors for large number of time samples. We do not divide the values by the number 
%    of samples here, so if the two signals are the same range of dynamics, those rounding errors
%    should not be a problem, even for a very large number of time samples.

% @=============================================================================
% This function is part of the Brainstorm software:

This file has been truncated. show original

Cheers,
Francois