function [bandwidth,density,xmesh,cdf]=kde(data,n,MIN,MAX)
% Reliable and extremely fast kernel density estimator for one-dimensional data;
%        Gaussian kernel is assumed and the bandwidth is chosen automatically;
%        Unlike many other implementations, this one is immune to problems
%        caused by multimodal densities with widely separated modes (see example). The
%        estimation does not deteriorate for multimodal densities, because we never assume
%        a parametric model for the data.
% INPUTS:
%     data    - a vector of data from which the density estimate is constructed;
%          n  - the number of mesh points used in the uniform discretization of the
%               interval [MIN, MAX]; n has to be a power of two; if n is not a power of two, then
%               n is rounded up to the next power of two, i.e., n is set to n=2^ceil(log2(n));
%               the default value of n is n=2^12;
%   MIN, MAX  - defines the interval [MIN,MAX] on which the density estimate is constructed;
%               the default values of MIN and MAX are:
%               MIN=min(data)-Range/10 and MAX=max(data)+Range/10, where Range=max(data)-min(data);
% OUTPUTS:
%   bandwidth - the optimal bandwidth (Gaussian kernel assumed);
%     density - column vector of length 'n' with the values of the density
%               estimate at the grid points;
%     xmesh   - the grid over which the density estimate is computed;
%             - If no output is requested, then the code automatically plots a graph of
%               the density estimate.
%        cdf  - column vector of length 'n' with the values of the cdf
%  Reference: 
% Kernel density estimation via diffusion
% Z. I. Botev, J. F. Grotowski, and D. P. Kroese (2010)
% Annals of Statistics, Volume 38, Number 5, pages 2916-2957. 

%
%  Example:
%           data=[randn(100,1);randn(100,1)*2+35 ;randn(100,1)+55];
%              kde(data,2^14,min(data)-5,max(data)+5);


%  Notes:   If you have a more reliable and accurate one-dimensional kernel density
%           estimation software, please email me at botev@maths.uq.edu.au


data=data(:); %make data a column vector
if nargin<2 % if n is not supplied switch to the default
    n=2^14;
end
n=2^ceil(log2(n)); % round up n to the next power of 2;
if nargin<4 %define the default  interval [MIN,MAX]
    minimum=min(data); maximum=max(data);
    Range=maximum-minimum;
    MIN=minimum-Range/10; MAX=maximum+Range/10;
end
% set up the grid over which the density estimate is computed;
R=MAX-MIN; dx=R/(n-1); xmesh=MIN+[0:dx:R]; N=length(unique(data));
%bin the data uniformly using the grid defined above;
initial_data=histc(data,xmesh)/N;  initial_data=initial_data/sum(initial_data);
a=dct1d(initial_data); % discrete cosine transform of initial data
% now compute the optimal bandwidth^2 using the referenced method
I=[1:n-1]'.^2; a2=(a(2:end)/2).^2;
% use  fzero to solve the equation t=zeta*gamma^[5](t)
try
    t_star=fzero(@(t)fixed_point(t,N,I,a2),[0,.1]);
catch
    t_star=.28*N^(-2/5);
end
% smooth the discrete cosine transform of initial data using t_star
a_t=a.*exp(-[0:n-1]'.^2*pi^2*t_star/2);
% now apply the inverse discrete cosine transform
if (nargout>1)|(nargout==0)
    density=idct1d(a_t)/R;
end
% take the rescaling of the data into account
bandwidth=sqrt(t_star)*R;
if nargout==0
    figure(1), plot(xmesh,density)
end
% for cdf estimation
if nargout>3
    f=2*pi^2*sum(I.*a2.*exp(-I*pi^2*t_star));
    t_cdf=(sqrt(pi)*f*N)^(-2/3);
    % now get values of cdf on grid points using IDCT and cumsum function
    a_cdf=a.*exp(-[0:n-1]'.^2*pi^2*t_cdf/2);
    cdf=cumsum(idct1d(a_cdf))*(dx/R);
    % take the rescaling into account if the bandwidth value is required
    bandwidth_cdf=sqrt(t_cdf)*R;
end

end
%################################################################
function  out=fixed_point(t,N,I,a2)
% this implements the function t-zeta*gamma^[l](t)
l=7;
f=2*pi^(2*l)*sum(I.^l.*a2.*exp(-I*pi^2*t));
for s=l-1:-1:2
    K0=prod([1:2:2*s-1])/sqrt(2*pi);  const=(1+(1/2)^(s+1/2))/3;
    time=(2*const*K0/N/f)^(2/(3+2*s));
    f=2*pi^(2*s)*sum(I.^s.*a2.*exp(-I*pi^2*time));
end
out=t-(2*N*sqrt(pi)*f)^(-2/5);
end


%##############################################################
function out = idct1d(data)

% computes the inverse discrete cosine transform
[nrows,ncols]=size(data);
% Compute weights
weights = nrows*exp(i*(0:nrows-1)*pi/(2*nrows)).';
% Compute x tilde using equation (5.93) in Jain
data = real(ifft(weights.*data));
% Re-order elements of each column according to equations (5.93) and
% (5.94) in Jain
out = zeros(nrows,1);
out(1:2:nrows) = data(1:nrows/2);
out(2:2:nrows) = data(nrows:-1:nrows/2+1);
%   Reference:
%      A. K. Jain, "Fundamentals of Digital Image
%      Processing", pp. 150-153.
end
%##############################################################

function data=dct1d(data)
% computes the discrete cosine transform of the column vector data
[nrows,ncols]= size(data);
% Compute weights to multiply DFT coefficients
weight = [1;2*(exp(-i*(1:nrows-1)*pi/(2*nrows))).'];
% Re-order the elements of the columns of x
data = [ data(1:2:end,:); data(end:-2:2,:) ];
% Multiply FFT by weights:
data= real(weight.* fft(data));
end