Hi all;

My question is how we can fit a mixture of Gaussian to a distribution, not to the data directly. Let's assume that we have got only the distribution, we don't have the original data. Then, if we assume the histogram (or pdf) is a mixture of Gaussians, how do we estimate the parameters of this model? I found several software (.m files) dealing with GMMs (Gaussian mixture models), but all of them fit a GMM to the original data, not to the distribution itself. Does anybody know how to solve this problem? Thanks a lot for your help;

Ali

> My question is how we can fit a mixture of Gaussian to a distribution, not 
> to the data directly. Let's assume that we have got only the distribution, 
> we don't have the original data. Then, if we assume the histogram (or pdf) 
> is a mixture of Gaussians, how do we estimate the parameters of this 
> model? I found several software (.m files) dealing with GMMs (Gaussian 
> mixture models), but all of them fit a GMM to the original data, not to 
> the distribution itself. Does anybody know how to solve this problem? 
> Thanks a lot for your help;

Ali, it would be better if you could get the original data, but here's an 
illustration of fitting a function that is a scaled gaussian mixture density 
to some (x,y) data having that form.

-- Tom

% define function that computes this density
f = @(p,x) p(1)*(p(2)   *normpdf(x,p(3),abs(p(4))) + ...
                (1-p(2))*normpdf(x,p(5),abs(p(6))));

% define some parameters
%     scale  proportion  mu1 sig1    mu2 sig2
p0 = [ 10        0.6      25  10     60   20];

% generate normal mixture density plus noise
x = linspace(0,100)';
y = f(p0,x) + 0.01*randn(size(x));
plot(x,y,'bx')

% fit to data and overlay on plot
%   hint: start with larger sigma values to help iterations succeed
p = nlinfit(x,y,f,[20 .5 30 30 70 30])
line(x,f(p,x),'color','r')