Daniel@0: function [pc_vec]=pca_kpm(features,N, method);
Daniel@0: % PCA_KPM Compute top N principal components using eigs or svd.
Daniel@0: % [pc_vec]=pca_kpm(features,N) 
Daniel@0: %
Daniel@0: % features(:,i) is the i'th example - each COLUMN is an observation
Daniel@0: % pc_vec(:,j) is the j'th basis function onto which you should project the data
Daniel@0: % using pc_vec' * features
Daniel@0: 
Daniel@0: [d ncases] = size(features);
Daniel@0: fm=features-repmat(mean(features,2), 1, ncases);
Daniel@0: 
Daniel@0: 
Daniel@0: if method==1 % d*d < d*ncases
Daniel@0:   fprintf('pca_kpm eigs\n');
Daniel@0:   options.disp = 0;
Daniel@0:   C = cov(fm'); % d x d matrix
Daniel@0:   [pc_vec, evals] = eigs(C, N, 'LM', options);
Daniel@0: else 
Daniel@0:   % [U,D,V] = SVD(fm), U(:,i)=evec of fm fm', V(:,i) = evec of fm' fm
Daniel@0:   fprintf('pca_kpm svds\n');
Daniel@0:   [U,D,V] = svds(fm', N);
Daniel@0:   pc_vec = V;
Daniel@0: end
Daniel@0: 
Daniel@0: if 0
Daniel@0: X = randn(5,3);
Daniel@0: X = X-repmat(mean(X),5,1);
Daniel@0: C = X'*X;
Daniel@0: C2=cov(X)
Daniel@0: [U,D,V]=svd(X);
Daniel@0: [V2,D2]=eig(X)
Daniel@0: end