# How to fit a multi modal distribution using a weighted sum of PDFs?

27 views (last 30 days)
Maria Amr on 2 Mar 2021
Commented: Maria Amr on 6 Apr 2021
I am new to matlab and I know my question is rudimentary. I really appreciated if you help me. I have a data-set (attached) shows multi modal distributions and I want to make a fit using a weighted sum of PDFs. How may I do that?
I have applied the Kernel distribution but I am not sure is right.
clc;
clear all;
close all;
%.....................................................................
rr=data(:,1);%gr/cm3
[x11,y11]=hist(rr,36);
hist(rr,36);
hold on
[f,x11] = ksdensity(rr,'Bandwidth',0.0028);
plot(x11,f,'-r','LineWidth',2)
##### 1 CommentShowHide None
Maria Amr on 2 Mar 2021
Bjorn Gustavsson Perfect! Really appreciated!

Bjorn Gustavsson on 2 Mar 2021
Edited: Bjorn Gustavsson on 2 Mar 2021
You don't have enough samples to confidently claim you have a multimodal distribution. If you simply try this with fitting exponential distributions to your data you'll see that they work reasonably OK:
PARHAT = expfit(abs(rr));
PARHATp = expfit(rr(rr>=0));
PARHATm = expfit(-rr(rr<=0));
hist(NUM,40)
hold on
x = linspace(-0.1,0.1,1001);
plot(x,30*exp(-abs(x)/PARHAT),'c')
plot(x,30*exp(-abs(x)/PARHATp),'r')
plot(x,30*exp(-abs(x)/PARHATm),'m')
There are indications that there might be a multimodal distribution, but if you do fit for a multimodal distribution you will probably find that the parameter uncertainty will be very large. First you need to gather more observations (hopefully this will be possible without too large costs in time and resources).
HTH
Maria Amr on 6 Apr 2021
Bjorn Gustavsson, You have already provided me a great answer that is working very well. Now I need to compare the previous result with a case with a few modes. You have already mentioned the data is not enough to claim a multimodal distribution. I truly appreciate you to direct me how I can show a multimodal dataset assuming the samples are enough. I means I want to show a data set like the attached figure (the figure is not mine and I only use it to express my question). The dataset is the same already attached above.Thank you in advance!

### More Answers (1)

Tom Lane on 2 Mar 2021
I'm glad Bjorn provided an answer that works for you. For future reference, there is a function for fitting mixtures of normal distributions:
Also, there is an example that fits a mixture of two normals, but it can be adapted to fix mixtures of any distributions:
Tom Lane on 3 Mar 2021
@Maria Amr, I don't know what h() is in your code, but if it is a variable then you are indexing into it using values like -0.1.