0
$\begingroup$

I am working with the following TransformedDistribution:

Dist = TransformedDistribution[2 v S1 + 2 v S2 - v c, {S1 \[Distributed] BinomialDistribution[1/2 (c + t), p], S2 \[Distributed] BinomialDistribution[1/2 (c - t), 1 - p]}]

This random variable has neat expressions for mean, variance, skewness, and kurtosis:

(-1 + 2 p) t v

-4 c (-1 + p) p v^2

((1 - 2 p) t v)/(c Sqrt[-c (-1 + p) p v^2])

3 - 6/c + 1/(c p - c p^2)

What I am trying to do is compare this random variable, when $t=0$, $c = \frac{1}{v^2}$, $v$ approaches $0$, and $c$ thus approaches infinity, with one drawn from the Normal distribution. Under these restrictions, $\mu$ = 0, $\sigma^2 = 4 (1 - p) p$, $\lambda = 3$, and $\kappa = 0$. Thus, it appears we are dealing with the Normal distribution, but that is not necessarily the case (see:link).

Improve this question
$\endgroup$
6
  • $\begingroup$ The distribution is never exactly normal. So what distance metric would characterize an important difference and what size of difference would be important? I ask because "how close is close" is a subject matter issue rather than a statistical issue. $\endgroup$
    – JimB
    Dec 29, 2018 at 22:49
  • $\begingroup$ @JimB That is a good question. Suppose I wanted to just inspect visually. For example, if I also set p = 0.5 then variance becomes 1 and a comparison to the Standard Normal distribution would seem appropriate. Could a comparison be setup using Manipulate? $\endgroup$
    – user120911
    Dec 29, 2018 at 22:53
  • $\begingroup$ That's the "I'll know it when I see it" approach. It's used all the time. (But it's not necessarily a consistent approach.) My point is that there's no universally accepted rule. $\endgroup$
    – JimB
    Dec 29, 2018 at 22:58
  • $\begingroup$ The answer from @user120911 below shows that there is an equivalence in distribution in the limit. Do you also need to know "how close" you are for small values of $v$? $\endgroup$
    – JimB
    Dec 30, 2018 at 18:54
  • $\begingroup$ @JimB That is interesting, but unless you already have that worked out, I would not request it. $\endgroup$
    – user120911
    Dec 30, 2018 at 19:21

1 Answer 1

Reset to default
2
$\begingroup$ 
Dist = FullSimplify[TransformedDistribution[2 v S1 + 2 v S2 - v c, {S1 \[Distributed] BinomialDistribution[1/2 (c + t), p], S2 \[Distributed] BinomialDistribution[1/2 (c - t), 1 - p]}]]

The moment generating function of the Dist is

MomentGeneratingFunction[Dist, x]

E^(-c v x) (-E^(2 v x) (-1 + p) + p)^((c - t)/2) (1 + (-1 + E^(2 v x)) p)^((c + t)/2)

Now let us consider the situation where c = 1/v^2, v-> 0, and t = 0. Under these restrictions, the moment generating function of the Dist is

Limit[With[{c = 1/v^2, t = 0},E^(-c v x) (-E^(2 v x) (-1 + p) + p)^((c - t)/2) (1 + (-1 + E^(2 v x)) p)^((c + t)/2)], v -> 0]

which simplifies to

E^(-2 (-1 + p) p x^2)

In comparison, if we examine the Normal Distribution with variance given by -4(-1 + p)p (i.e., the variance of Dist under the stated restrictions), then we notice the moment generating function is

MomentGeneratingFunction[NormalDistribution[0, Sqrt[-4 (-1 + p) p ]],x]

E^(-2 (-1 + p) p x^2)

which we notice is equal to the above.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Did you obtain the mgf from Mathematica commands? If so, would you consider adding those to your answer? $\endgroup$
    – JimB
    Dec 30, 2018 at 18:50
  • $\begingroup$ Excellent! Thank you. $\endgroup$
    – JimB
    Dec 30, 2018 at 20:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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