%DKL Kullback-Leibler divergence
%
%     D=DKL(DATA,PROTO)
%
% INPUT
%   DATA   a dataset
%   PROTO  the dataset with prototypes (representation set)
% OUTPUT
%   D      dataset with a distance matrix
%
% DESCRIPTION
% The Kullback-Leibler divergence (pp. 151, Pattern Recognition,
% Theodoridis, p. 176 in 2nd edition)
%

% $Id: dkl2.m,v 1.2 2005/03/02 13:39:45 serguei Exp $

function d=dkl2(data,proto)

fprintf(' dkl2:');

if size(data,2)~=size(proto,2)
  error('both datasets must have the same feature sizes');
end
sc=size(data,1); % training sample size
pc=size(proto,1); % test sample size
d=zeros(sc,pc);

labdata=getlab(data);
labproto=getlab(proto);
data2=data; data=+data;
proto2=proto; proto=+proto;


% normalization: we consider a histogram being a set of probability
% estimates
data=data./repmat(sum(data,2),[1,size(data,2)]);
proto=proto./repmat(sum(proto,2),[1,size(proto,2)]);

step=round(pc/10);
for i=1:pc
  t=repmat(proto(i,:),[sc,1]);

  w = warning;
  warning('off');
  l = log(t./data);
  warning(w);  

  l(~isfinite(l)) = 0;
  d(:,i)=sum((t-data).*l ,2);

  if mod(i,step)==0, fprintf('.'); end
end

d=setdat(data2,d);
d=setfeatlab(d,labproto);

return