[LMSout, blms, Rsq, error1]=LMTSregor(y, X, max_fits, max_points)

function [LMSout, blms, Rsq, error1]=LMTSregor(y, X, max_fits, max_points)
% % Syntax;
% % [LMSout,blms,Rsq]=LMTSregor(y, X, max_fits, max_points);
% %
% % Least Median Trimmed Squares Through the Origin
% %
% % This program is a modification of LMSregor.  It has been modified to
% % trim the input data sets adn trim the number of combinations of
% % line fits that are processed.  The trimming allows the program to
% % accomodate large data sets.
% %
% % This program performs the Least Median Trimmed Squares Robust
% % Regression thorugh the origin for simple or multiple columns of
% % data and outputs the regression parameters.
% %
% % Breakdown has been observed to occur at 50%; however, the breakdown
% % point is not known for all problems.
% %
% % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% %
% % Input Variable Description
% %
% % y is the column vector of the dependent variable.
% %
% % X is the matrix of the independent variable. If it is one dimensional,
% %     then it should be a column vector.  If X is an empty matrix, then
% %     X is assumed to be a column of integers starting from 0.
% %
% % max_fits is the number of best fit pairs of data. The default value
% %     does not exceed 1000.  The maximum value is 10000.
% %
% % max_points is the number of data points for curve fitting.
% %     The maximum value is 100000.  The default value is 100000.
% %
% % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% %
% % Output Variable Description
% %
% % LMSout is the LMTS estimated values vector.
% %
% % blms is the LMTS [intercept slopes] vector.
% %
% % Rsq is the R-squared error estimate of the fit.
% %
% % error1 is 1 if there is an error otherwise it is 0.
% %
% % **********************************************************************
% %
% % This program is originally the work of
% %
% % Alexandros Leontitsis
% % Institute of Mathematics and Statistics
% % University of Kent at Canterbury
% % Canterbury
% % Kent, CT2 7NF
% % U.K.
% %
% % University e-mail: [e-mail] (until December 2002)
% % Lifetime e-mail: [e-mail]
% % Homepage: http://www.geocities.com/CapeCanaveral/Lab/1421
% %
% % Sep 3, 2001.
% %
% % **********************************************************************
% %
% % Reference:
% % Rousseeuw PJ, Leroy AM (1987):
% % Robust regression and outlier detection. Wiley.
% %
% % **********************************************************************
% %
% Example='';
% % Establish an exact solution (xe, ye)
%
% xe=1/100*(1:10000)';
% ye=10*xe;
%
% % Create a noisy data set with an outlier (X, y)
%
% X=1/100*randn(size(xe))+xe;
% y=10*(1/100*randn(size(xe))+xe);
%
% % Perform the robust median trimmed squares linear regression through the
% % origin.
% max_fits=100;
% max_points=500;
%
% % Outlier data points form a line with a different slope
% % randomly select 49% of the data points to be outliers
% [ndraw]=rand_int(1, length(xe), 0.40*length(X), 1, 1);
% y(ndraw)=50./100.*randn(size(ndraw))+50*ndraw;
% X(ndraw)=1./100.*randn(size(ndraw))+1./100*ndraw;
%
% [LMSout,blms,Rsq]=LMTSregor(y, X, max_fits, max_points);
% % plot the robust solution
% xr=xe;
% yr=polyval([blms(1) 0], xr);
% % plot the typical regression solution
% xp=xr;
% p=polyfit(X, y, 1);
% yp=polyval(p, xp);
%
% figure(1); plot(X, y, 'linestyle', 'none', 'marker', '.', 'markersize', 3, 'markeredgecolor', 'k');
% hold on; plot(xe, ye, 'g', 'linewidth', 1);
% plot(xr, yr, 'r', 'linewidth', 1);
% plot(xp, yp, 'b', 'linewidth', 1);
% legend({'Scattered Data', 'Exact Solution', 'Robust Solution', 'Regular Regression'});
% xlim([1 100]);
% ylim([1 100000]);
% title({'60% of the data is near the x-axis', '40% of the data is near the y-axis'}, 'fontsize', 20);
% xlabel('x-axis', 'fontsize', 18);
% ylabel('y-axis', 'fontsize', 18);
% set(gca, 'fontsize', 14);
%
% % ***********************************************************
% %
% % This program was modified by Edward L. Zechmann
% %
% %      date  1 February 2008  updated comments
% %                             added rand_int code to randomly select
% %                             data points
% %
% %  modified 11 February 2008  trimmed the input data arrays
% %                             updated comments
% %
% %  modified 14 February 2008  trimmed the input data arrays
% %                             updated comments.
% %                             Improved the error handling and default
% %                             values.
% %
% % ***********************************************************
% %
% % Feel free to modify this code.
% %
% %

% set the flag to null
% set the error to no error
flag=0;
error1=0;

if nargin < 1 || isempty(y)
    warning('Not enough input arguments.  Return empty array.');
    flag=1;
    error1=1;
    n=1;
    y=1;
else
    % y must be a column vector
    y=y(:);
    % n is the length of the data set
    n=length(y);
end

if nargin < 2 || isempty(X)
    % if X is omitted give it the values 1:n
    X=(1:n)';
else
    % X must be a 2-dimensional matrix
    [mx, nx]=size(X);
    if nx > mx
        X=X';
    end

    if ndims(X) > 2
        warning('Invalid data set X.  Return empty array.');
        flag=1;
        error1=1;
    end

    if n~=size(X,1)
        warning('The rows of X and y must have the same length');
        flag=1;
        error1=1;
    end
end

if isequal(flag, 1)
    LMSout=[];
    blms=[];
    Rsq=[];
else

    LMSout=1;
    blms=1;
    Rsq=1;
    error1=1;
    
    pp=size(X,2);

    % If not input, set the maximum number of fits
    if nargin < 3 || isempty(max_fits)
        % default value of max_fits is 1000
        max_fits=min([1000, floor(n/pp)]);
    end

    % make sure that max_fits does not exceed 10000
    max_fits=min( [max_fits, floor(n/pp), 10000]);

    % If max_points is not an input, set the maximum number of points
    % for the input arrays X and y to a reasonable value.
    if nargin < 4
        max_points=max([min([length(y), 100000]), max_fits*pp]);
    end

    if max_points < max_fits
        max_points=max_fits;
    end


    % Program Modified Here
    % input data is trimmed
    % best fit combinations are trimmed
    [C, y, X, n, p]=LMS_trim(y, X, max_fits, max_points, 2);

    % The "half" of the data points
    h=floor(n/2)+floor((p+1)/2);

    rmin=Inf;



    for i=1:size(C,1)
        
        %A=[ones(p, 1) X(C(i,:), :)];
        %b=y(C(i,:));
        
        for j=1:p
            A(j,:)=X(C(i,j),:);
            b(j,1)=y(C(i,j));
        end
        
        if rank(A')==p

            % Calculate the slopes
            c=inv(A'*A)*A'*b;

            % There is no intercept, so the estimation is straightforward
            est=X*c;
            r=y-est;
            r2=r.^2;
            r2=sort(r2);
            rlms=r2(h);
            if rlms<rmin
                rmin=rlms;
                blms=c;
                LMSout=est;
                % Chapter 2, eq. 3.12
                Rsq=1-(median(abs(r))/median(abs(y)))^2;
            end

        end
    end

end