function [coeff, score, latent, tsquare] = princomp(x,econFlag)
%PRINCOMP Principal Components Analysis.
%   COEFF = PRINCOMP(X) performs principal components analysis on the N-by-P
%   data matrix X, and returns the principal component coefficients, also
%   known as loadings.  Rows of X correspond to observations, columns to
%   variables.  COEFF is a P-by-P matrix, each column containing coefficients
%   for one principal component.  The columns are in order of decreasing
%   component variance.
%
%   PRINCOMP centers X by subtracting off column means, but does not
%   rescale the columns of X.  To perform PCA with standardized variables,
%   i.e., based on correlations, use PRINCOMP(ZSCORE(X)).  To perform PCA
%   directly on a covariance or correlation matrix, use PCACOV.
%
%   [COEFF, SCORE] = PRINCOMP(X) returns the principal component scores,
%   i.e., the representation of X in the principal component space.  Rows
%   of SCORE correspond to observations, columns to components.
%
%   [COEFF, SCORE, LATENT] = PRINCOMP(X) returns the principal component
%   variances, i.e., the eigenvalues of the covariance matrix of X, in
%   LATENT.
%
%   [COEFF, SCORE, LATENT, TSQUARED] = PRINCOMP(X) returns Hotelling's
%   T-squared statistic for each observation in X.
%
%   When N <= P, SCORE(:,N:P) and LATENT(N:P) are necessarily zero, and the
%   columns of COEFF(:,N:P) define directions that are orthogonal to X.
%
%   [...] = PRINCOMP(X,'econ') returns only the elements of LATENT that are
%   not necessarily zero, i.e., when N <= P, only the first N-1, and the
%   corresponding columns of COEFF and SCORE.  This can be significantly
%   faster when P >> N.
%
%   See also BARTTEST, BIPLOT, CANONCORR, FACTORAN, PCACOV, PCARES, ROTATEFACTORS.

%   References:
%     [1] Jackson, J.E., A User's Guide to Principal Components,
%         Wiley, 1988.
%     [2] Jolliffe, I.T. Principal Component Analysis, 2nd ed.,
%         Springer, 2002.
%     [3] Krzanowski, W.J., Principles of Multivariate Analysis,
%         Oxford University Press, 1988.
%     [4] Seber, G.A.F., Multivariate Observations, Wiley, 1984.

%   Copyright 1993-2009 The MathWorks, Inc.
%   $Revision: 2.9.2.11 $  $Date: 2009/01/30 14:50:55 $

% When X has more variables than observations, the default behavior is to
% return all the pc's, even those that have zero variance.  When econFlag
% is 'econ', those will not be returned.
if nargin < 2, econFlag = 0; end

[n,p] = size(x);
if isempty(x)
    pOrZero = ~isequal(econFlag, 'econ') * p;
    coeff = zeros(p,pOrZero); coeff(1:p+1:end) = 1;
    score = zeros(n,pOrZero);
    latent = zeros(pOrZero,1);
    tsquare = zeros(n,1);
    return
end

% Center X by subtracting off column means
x0 = bsxfun(@minus,x,mean(x,1));
   
if nargout < 2
    % The principal component coefficients are the eigenvectors of
    % S = X0'*X0./(n-1), but computed using SVD.
  if n >= p && (isequal(econFlag,0) || isequal(econFlag,'econ'))
        [coeff,ignore] = eig(x0'*x0);
        coeff = fliplr(coeff);
    else
        [ignore,ignore,coeff] = svd(x0,econFlag);
    end
    % When econFlag is 'econ', only (n-1) components should be returned.
    % See comment below.
    if (n <= p) && isequal(econFlag, 'econ')
        coeff(:,n) = [];
    end
else
    r = min(n-1,p); % max possible rank of X0
   
    % The principal component coefficients are the eigenvectors of
    % S = X0'*X0./(n-1), but computed using SVD.
    [U,sigma,coeff] = svd(x0,econFlag); % put in 1/sqrt(n-1) later

    % Project X0 onto the principal component axes to get the scores.
    if n == 1 % sigma might have only 1 row
        sigma = sigma(1);
    else
        sigma = diag(sigma);
    end
    score = bsxfun(@times,U,sigma'); % == x0*coeff
    sigma = sigma ./ sqrt(n-1);

    % When X has at least as many variables as observations, eigenvalues
    % n:p of S are exactly zero.
    if n <= p
        % When econFlag is 'econ', nothing corresponding to the zero
        % eigenvalues should be returned.  svd(,'econ') won't have
        % returned anything corresponding to components (n+1):p, so we
        % just have to cut off the n-th component.
        if isequal(econFlag, 'econ')
            sigma(n,:) = []; % make sure this shrinks as a column
            coeff(:,n) = [];
            score(:,n) = [];

        % Otherwise, set those eigenvalues and the corresponding scores to
        % exactly zero.  svd(,0) won't have returned columns of U
        % corresponding to components (n+1):p, need to fill those out.
        else
            sigma(n:p,1) = 0; % make sure this extends as a column
            score(:,n:p) = 0;
        end
    end

    % The variances of the pc's are the eigenvalues of S = X0'*X0./(n-1).
    latent = sigma.^2;

    % Hotelling's T-squared statistic is the sum of squares of the
    % standardized scores, i.e., Mahalanobis distances.  When X appears to
    % have column rank < r, ignore components that are orthogonal to the
    % data.
    if nargout == 4
        if n > 1
            q = sum(sigma > max(n,p).*eps(sigma(1)));
            if q < r
                warning('stats:princomp:colRankDefX', ...
                        ['Columns of X are linearly dependent to within machine precision.\n' ...
                         'Using only the first %d components to compute TSQUARED.'],q);
            end
        else
            q = 0;
        end
        tsquare = (n-1) .* sum(U(:,1:q).^2,2); % == sum((score*diag(1./sigma)).^2,2)
    end
end