[coeff,score,latent] = pca(x); % x为原始数据，行数为样本数，列数为特征数
x_bar = x - repmat(mean(x), size(x,1), 1); % 均值化
[u,s,d] = svd(cov(x)); % u === coeff, diag(s) === latent
xRot = x_bar * u; % xRot === score
PCA降维：
[coeff,score,latent] = pca(x);
x_pca = score(:, 1:20); % 降维后的数据，剩下20维
x_pca = x_bar * u(:, 1:20);
