Homec4science
Diffusion iCAPs public (master) Browse

iCAPs/packages/spm12/external/fieldtrip/connectivity/private/sfactorization_wilson3x3.m
8c8accbbb214master

Last Change
Raw File

sfactorization_wilson3x3.m
View Options

function [H, Z, S, psi] = sfactorization_wilson3x3(S,freq,Niterations,tol,cmbindx,fb,init,checkflag,stabilityfix)
% SFACTORIZATION_WILSON3X3 performs triplet-wise non-parametric spectral factorization on
% cross-spectra, based on Wilson's algorithm.
%
% Usage : [H, Z, psi] = sfactorization_wilson(S,freq);
%
% Inputs : S (1-sided, 3D-spectral matrix in the form of Channel x Channel x frequency)
% : freq (a vector of frequencies) at which S is given.
%
% Outputs: H (transfer function)
% : Z (noise covariance)
% : S (cross-spectral density 1-sided)
% : psi (left spectral factor)
%
% This function is an implemention of Wilson's algorithm (Eq. 3.1)
% for spectral matrix factorization.
%
% Ref: G.T. Wilson,"The Factorization of Matricial Spectral Densities,"
% SIAM J. Appl. Math.23,420-426(1972).
% Written by M. Dhamala & G. Rangarajan, UF, Aug 3-4, 2006.
% Email addresses: mdhamala@bme.ufl.edu, rangaraj@math.iisc.ernet.in
% Copyright (C) 2009-2017, Jan-Mathijs Schoffelen
%
% This file is part of FieldTrip, see http://www.fieldtriptoolbox.org
% for the documentation and details.
%
% FieldTrip is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% FieldTrip is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with FieldTrip. If not, see <http://www.gnu.org/licenses/>.
%
% $Id$
if nargin<9, stabilityfix = false; end
if nargin<8, checkflag = true; end
if nargin<7, init = 'chol'; end
if nargin<6, fb = 'none'; end
if nargin<5
ft_error('FieldTrip:connectivity:sfactorization_wilson3x3', 'when requesting multiple triplet-wise spectral decomposition, ''cmbindx'' needs to be specified');
end
if nargin<4, tol = 1e-8; end
if nargin<3, Niterations = 1000; end;
dfreq = round(diff(freq)*1e5)./1e5; % allow for some numeric issues
if ~all(dfreq==dfreq(1))
ft_error('FieldTrip:connectivity:sfactorization_wilson3x3', 'frequency axis is not evenly spaced');
end
if freq(1)~=0
ft_warning('FieldTrip:connectivity:sfactorization_wilson3x3', 'when performing non-parametric spectral factorization, the frequency axis should ideally start at 0, zero padding the spectral density');
dfreq = mean(dfreq);
npad = freq(1)./dfreq;
% update the freq axis and keep track of the frequency bins that are
% expected in the output
selfreq = (1:numel(freq)) + npad;
freq = [(0:(npad-1))./dfreq freq];
S = cat(3, zeros(size(S,1), size(S,1), npad), S);
else
selfreq = 1:numel(freq);
end
% ensure input S is double (mex-files don't work with single)
S = double(S);
% check whether the last frequency bin is strictly real-valued.
% if that's the case, then it is assumed to be the Nyquist frequency
% and the two-sided spectral density will have an even number of
% frequency bins. if not, in order to preserve hermitian symmetry,
% the number of frequency bins needs to be odd.
Send = S(:,:,end);
N = numel(freq);
m = size(cmbindx,1);
if all(imag(Send(:))<abs(trace(Send)./size(Send,1)*1e-9))
hasnyq = true;
N2 = 2*(N-1);
else
hasnyq = false;
N2 = 2*(N-1)+1;
end
% preallocate memory for the identity matrix
I = repmat(eye(3),[1 1 m N2]); % Defining 3 x 3 identity matrix
% %Step 1: Forming 2-sided spectral densities for ifft routine in matlab
% Sarr = zeros(3,3,m,N2) + 1i.*zeros(3,3,m,N2);
% for c = 1:m
% Stmp = S(cmbindx(c,:),cmbindx(c,:),:);
%
% % the input cross-spectral density is assumed to be weighted with a
% % factor of 2 in all non-DC and Nyquist bins, therefore weight the
% % DC-bin with a factor of 2 to get a correct two-sided representation
% Sarr(:,:,c,1) = Stmp(:,:,1).*2;
%
% for f_ind = 2:N
% Sarr(:,:,c, f_ind) = Stmp(:,:,f_ind);
% Sarr(:,:,c,(N2+2)-f_ind) = Stmp(:,:,f_ind).';
% end
% end
% Sarr2 = Sarr;
% preallocate memory for the 2-sided spectral density
Sarr = zeros(3,3,N2,m) + 1i.*zeros(3,3,N2,m);
for c = 1:m
Sarr(:,:,1:N,c) = S(cmbindx(c,:),cmbindx(c,:),:);
end
if hasnyq
N1 = N;
else
N1 = N + 1; % the highest frequency needs to be represented twice, for symmetry purposes
end
Sarr(:,:,N1:N2,:) = flip(Sarr(:,:,2:N,:),3);
Sarr(2,1,N1:N2,:) = conj(Sarr(2,1,N1:N2,:));
Sarr(1,2,N1:N2,:) = conj(Sarr(1,2,N1:N2,:));
Sarr(3,1,N1:N2,:) = conj(Sarr(3,1,N1:N2,:));
Sarr(1,3,N1:N2,:) = conj(Sarr(1,3,N1:N2,:));
Sarr(3,2,N1:N2,:) = conj(Sarr(3,2,N1:N2,:));
Sarr(2,3,N1:N2,:) = conj(Sarr(2,3,N1:N2,:));
Sarr = permute(Sarr, [1 2 4 3]);
Sarr(:,:,:,1) = Sarr(:,:,:,1).*2; % weight the DC-bin
% the input cross-spectral density is assumed to be weighted with a
% factor of 2 in all non-DC and Nyquist bins, therefore weight the
% Nyquist bin with a factor of 2 to get a correct two-sided representation
if hasnyq
Sarr(:,:,:,N) = Sarr(:,:,:,N).*2;
end
%Step 2: Computing covariance matrices
gam = real(reshape(ifft(reshape(Sarr, [9*m N2]), [], 2),[3 3 m N2]));
%Step 3: Initializing for iterations
gam0 = gam(:,:,:,1);
h = complex(zeros(size(gam0)));
for k = 1:m
switch init
case 'chol'
[tmp, dum] = chol(gam0(:,:,k));
if dum
ft_warning('initialization with ''chol'' for iterations did not work well, using arbitrary starting condition');
tmp = rand(3,3); %arbitrary initial condition
tmp = triu(tmp);
end
case 'rand'
tmp = rand(3,3); %arbitrary initial condition
tmp = triu(tmp);
otherwise
ft_error('initialization method should be eithe ''chol'' or ''rand''');
end
h(:,:,k) = tmp;
%h(:,:,k) = chol(gam0(:,:,k));
end
psi = repmat(h, [1 1 1 N2]);
%Step 4: Iterating to get spectral factors
ft_progress('init', fb, 'computing spectral factorization');
for iter = 1:Niterations
ft_progress(iter./Niterations, 'computing iteration %d/%d\n', iter, Niterations);
invpsi = inv3x3(psi);
g = sandwich3x3(invpsi, Sarr) + I;
gp = PlusOperator3x3(g,m,N,stabilityfix); %gp constitutes positive and half of zero lags
psi_old = psi;
psi = mtimes3x3(psi, gp);
if checkflag
psierr = abs(psi-psi_old)./abs(psi);
psierrf = mean(psierr(:));
if(psierrf<tol)
fprintf('reaching convergence at iteration %d\n',iter);
break;
end % checking convergence
end
end
ft_progress('close');
%Step 5: Getting covariance matrix from spectral factors
gamtmp = reshape(real(ifft(transpose(reshape(psi, [9*m N2]))))', [3 3 m N2]);
%Step 6: Getting noise covariance & transfer function (see Example pp. 424)
%
A0 = gamtmp(:,:,:,1);
A0inv = inv3x3(A0);
Z = zeros(3,3,m);
for k = 1:m
%Z = A0*A0.'*fs; %Noise covariance matrix
Z(:,:,k) = A0(:,:,k)*A0(:,:,k).'; %Noise covariance matrix not multiplied by sampling frequency
%FIXME check this; at least not multiplying it removes the need to correct later on
%this also makes it more equivalent to the noisecov estimated by biosig's mvar-function
end
% H = complex(zeros(3,3,m,N));
% S = complex(zeros(3,3,m,N));
% for k = 1:N
% for kk = 1:m
% H(:,:,kk,k) = psi(:,:,kk,k)*A0inv(:,:,kk); % Transfer function
% S(:,:,kk,k) = psi(:,:,kk,k)*psi(:,:,kk,k)'; % Cross-spectral density
% end
% end
H = mtimes3x3(psi,A0inv(:,:,:,ones(1,size(psi,4))));
S = mtimes3x3(psi,ctranspose3x3(psi));
siz = [size(H) 1 1];
H = reshape(H, [9*siz(3) siz(4:end)]);
siz = [size(S) 1 1];
S = reshape(S, [9*siz(3) siz(4:end)]);
siz = [size(Z) 1 1];
Z = reshape(Z, [9*siz(3) siz(4:end)]);
siz = [size(psi) 1 1];
psi = reshape(psi, [9*siz(3) siz(4:end)]);
%if numel(selfreq)~=numel(freq)
% return only the frequency bins that were in the input
H = H(:,selfreq,:,:);
S = S(:,selfreq,:,:);
psi = psi(:,selfreq,:,:);
%end
%---------------------------------------------------------------------
function gp = PlusOperator3x3(g,ncmb,nfreq, stabilityfix)
% This function is for [ ]+operation:
% to take the positive lags & half of the zero lag and reconstitute
% M. Dhamala, UF, August 2006
g = transpose(reshape(g, 9*ncmb, []));
gam = ifft(g);
% taking only the positive lags and half of the zero lag
gamp = gam;
beta0 = 0.5*gam(1,:);
%for k = 1:ncmb
% gamp(1,(k-1)*4+1:k*4) = reshape(triu(reshape(beta0(1,(k-1)*4+1:k*4),[2 2])),[1 4]);
%end
beta0(2:9:9*ncmb) = 0;
beta0(3:9:9*ncmb) = 0;
beta0(6:9:9*ncmb) = 0;
gamp(1,:) = beta0;
gamp(nfreq+1:end,:) = 0;
% smooth with a window, only for the long latency boundary: this is a
% stabilityfix proposed by Martin Vinck
if stabilityfix
w = tukeywin(nfreq*2, 0.5);
gamp(1:nfreq,:) = gamp(1:nfreq,:).*repmat(w(nfreq+1:end),[1 nchan^2]);
else
% nothing to be done here
end
% reconstituting
gp = fft(gamp);
gp = reshape(transpose(gp), [3 3 ncmb numel(gp)/(9*ncmb)]);
c4science · Help