2
$\begingroup$

I have this Mixture Distribution of two Normal Distributions:

 mix[p_, μ1_, μ2_, σ_] := 
 MixtureDistribution[{p, 
 1 - p}, {NormalDistribution[μ1, σ], 
 NormalDistribution[μ2, σ]}];

Say $x_1, x_2$ are samples from this mixture. Set $y=x_1+x_2$, so that $y$ is the sum of two samples from the mixture. I now generate 100 $y's$ (with arbitrary p, mu and sigma, each is a sum of two $x$:

mixdat = RandomVariate[mix[0.75, 0.5, -1.5, 0.2], 1000];
mixdatSum = 
Total[RandomVariate[mix[0.75, 0.5, -1.5, 0.2], 2]] & /@ Range[100];

mixdatSum contains these 100 values. Now given the data in mixdatSum I want to estimate the Distribution Parameters. So a natural way to solve this would be:

FindDistributionParameters[mixdatSum, 
TransformedDistribution[x1 + x2, 
Thread[Distributed[{x1, x2}, 
MixtureDistribution[{p, 
  1 - p}, {NormalDistribution[μ1, σ], 
  NormalDistribution[μ2, σ]}]]]]]

Problem is, the sum of a mixture does not follow the mixture distribution, so this does not work and I get lots of errors.

However I want to treat the sum of a mixture AS IF IT WERE following the same mixture, for approximation purposes.

Is there a way to do this?

$\endgroup$
4
$\begingroup$

Since the PDF can be computed in closed form you might have some luck with ProbabilityDistribution and some half-way reasonable starting values.

Generate the data...

mix[p_, m1_, m2_, s_] := MixtureDistribution[{p, 1 - p}, {NormalDistribution[m1, s], 
    NormalDistribution[m2, s]}];

mixdatSum = Plus @@ RandomVariate[mix[0.75, 0.5, -1.5, 0.2], {2, 100}];

Use your TransformedDistribution to get the PDF.

sumdist = TransformedDistribution[x1 + x2, Thread[Distributed[{x1, x2}, 
     MixtureDistribution[{p,  1 - p}, {NormalDistribution[m1, s], 
       NormalDistribution[m2, s]}]]]];

pdf[z_] = FullSimplify[PDF[sumdist, z], DistributionParameterAssumptions[sumdist]]

Create a ProbabilityDistribution from your PDF with the proper assumptions set.

pdist = ProbabilityDistribution[pdf[z], {z, -Infinity, Infinity}, 
  Assumptions -> DistributionParameterAssumptions[sumdist]]

With some starting values in the right ballpark you can get good estimates of the parameters.

FindDistributionParameters[mixdatSum, pdist, {{m1, 1}, {m2, -1}, {s, 1}, {p, .5}}]

(*{m1 -> 0.514195, m2 -> -1.53428, s -> 0.226609, p -> 0.719953}
$\endgroup$
  • $\begingroup$ thanks, this is looking great. Is there a way to get this to work with a Mixture Distribution of two LogNormal Distributions? There is no closed form there. $\endgroup$ – spore234 Jul 28 '14 at 21:05
  • $\begingroup$ To my knowledge you won't have any luck with ProbabilityDistribution if you don't have a closed form for one of the distribution functions. $\endgroup$ – Andy Ross Jul 29 '14 at 2:21
  • $\begingroup$ According to a paper I read the sum of two Lognormals is approximately Lognormal. Can I use this to define my own ProbabilityDensity for the sum of two Lognormals? $\endgroup$ – spore234 Jul 29 '14 at 5:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.