Code covered by the BSD License

# Nipals algorithm for Principle Component Analysis

### Qiaonan Duan (view profile)

10 Jun 2013 (Updated )

This function is written largely based on nipals function from R chemometrics package.

nipals(X,a,it,tol )
function [ T,P,pcvar ] = nipals(X,a,it,tol )
%Nipals algorithm for Principle Component Analysis
%
% Input:
%   X: a matrix with rows obseravations and columns variables.
%   a: number of components desired in the output.
%   it: maximum number of iterations allowed to find one component. default 1000.
%   tol: tolerance value to assese convergence. default 1e-4
%
% Output:
%   T: scores or transformed coordinates.
%   pcvar: variance captured in each output component.
%
% If X is m*n matrix, then T is m*a, P is n*a
%
% This function is written largely based on nipals function from R chemometrics
% package. It was frustrating that Matlab did not has a robust nipals
% function so I wrote one for my project and share here.
%
% Author: Qiaonan Duan, 6/7/2013, MSSM.
%
% Line 48 (6/10/2013): prec = thnew-th; --> prec = (thnew-th)'*(thnew-th);
%
if nargin == 2
it = 1000;
tol = 1e-4;
elseif nargin == 3
tol = 1e-4;
end

[obsCount,varCount] = size(X);
Xh = X - repmat( mean(X,1), obsCount, 1 );
T = zeros(obsCount,a);
P = zeros(varCount,a);
pcvar = zeros(varCount,1);
varTotal = sum(var(Xh));
currVar = varTotal;
nr = 0;

for h = 1:a
th = Xh(:,1);
ende = false;

while(~ende)
nr = nr+1;
ph = Xh'*th/(th'*th);
ph = ph/norm(ph);
thnew = Xh*ph/(ph'*ph);
prec = (thnew-th)'*(thnew-th);
th = thnew;

if prec <= tol^2
ende = true;
elseif it <= nr
ende = true;
disp('Iteration stops without convergence')
end
end

Xh = Xh-th*ph';
T(:,h) = th;
P(:,h) = ph;
oldVar = currVar;
currVar = sum(var(Xh));
pcvar(h) = ( oldVar - currVar )/varTotal;
nr = 0;

end