Using fitgmdist on a Histogram

I have a Histogram in which I want to find the peaks of a mixture of gaussians (see attached .fig file).

The Histogram shows 4 clear peaks of which I need to know the centres.

However, I can not get fitgmdist to properly fit gaussians. It does perform some kind of fit, but returns mu and Sigma with strange, impossible values!

I have also attached a MWE for reproducibilty.

I believe I am confused as to what the first argument to fitgmdist has to be.

This is a one dimensional problem, and Mathworks examples all deal with higher dimensional problems which I have difficulty to reduce.

Answer 1

[Image Analyst]

Fit the count data with fitnlm. I attach a demo for two Gaussians. Make the obvious adaptations for 4 Gaussians.

[Douglas Bock]

Many thanks Image Analyst!

I have previously considered hardcoding the GMM function to fit the data.

However, I may have anything between 1 and 10 peaks for which I need the gaussians. This would require me to hardcode 10 different GMM, which, although doable, would make the fitgmdist function redundant. I want to exploit that one of the arguements for fitgmdist is the number of peaks to be detected...

[Image Analyst]

Another option is to use the clean algorithm. The strategy astronomers use in finding stars is the "Clean" algorithm. Take largest peak, fit it, subtract from the data to get new data, and then iterate on all the smaller remaining peaks until you've gotten to peaks so small thatyou want to stop. I suppose it works okay empirically but I gotta think there's a fancier, mathematically accurate way. Though things are harder in 2-D - it's not just a straightforward extension of the 1-D case.

Try this link

Answer 2

[Alfredo Dinho]

Hello my friend! You just have to transpose the matrix ... see: I ran here and it worked perfectly ..

fitData = fitData + p(n)*normpdf(binCenters,mu(n),sig(n))';