What is the absolute expected value of the sum of a distribution made up of Bernoulli and Normal distributions?

121 Views Asked by At

$$X \sim \mathcal N(\mu, \sigma^2)$$ $$Y \sim Bern(p)$$

$$Z = XY$$

I then have that the pdf of $Z$ is:

$$f(z) = (p-1)\delta(z) + p g(z)$$

where $g(z)$ is the pdf of the normal distribution.

I then need to find the sum of $Q$ of these distributions and finally have to find the expected value of the absolute value of the final resulting distribution.

I do not know where to go from here. I have tried doing it with MGFs but I hit a brick wall with divergent integrals. Any help please?

EDIT:

I forgot to add that $\frac{Q}{2}$ of these distributions will have the normal with $-\mu$ and the other half with $\mu$. I have tried doing this with a Monte Carlo simulation and then with a similar method to what Henry suggested in the comments.

My simulation has the following results:

Q mean stdev
2 3.406746 3.146736
4 5.156385 4.041445
6 6.393295 4.865093
8 7.391076 5.586568
10 8.283423 6.254462
12 9.075857 6.846168
14 9.788690 7.383987
16 10.471054 7.905363
18 11.102388 8.384783
20 11.715745 8.837827

And the equation I got to with the various combinations of folded normal distributions is, where I have used that $p = 0.5$ (if anyone has a more generic solution even better):

FURTHER EDIT

I got it for $p = 0.5$, I had made a silly mistake at the start. Thank you all.

2

There are 2 best solutions below

10
On

Denote $f(a)=E(|Z-a|)$ if $Z\sim N(0,1)$ Denoting $\Phi(a)=\Pr(Z<a)$ we get $$f(a)=2a\Phi (a)-a+\sqrt{\frac{2}{\pi}}e^{-a^2/2}.$$ Now let $X_1,\ldots,X_n$ and $Y_1,\ldots, Y_n$ independent such that $X_i$ is Bernoulli of mean $p$ and $Y_i=\mu+\sigma Z_i$ with $Z_i\sim N(0,1). $ Note that $$|Y_1+\cdots+Y_k|=|k\mu+\sigma (Z_1+\cdots+Z_k)|\sim \sqrt{k}\sigma\left|Z_1+\frac{\sqrt{k}}{\sigma }\mu\right|$$ since $Z_1+\cdots+Z_k\sim N(0,k)$ and therefore $E(|Y_1+\cdots+Y_k|)=\sqrt{k}\sigma f(-\frac{\sqrt{k}}{\sigma }\mu)=g_k.$ To conclude$$E(|X_1Y_1+\cdots+X_nY_n|)=\sum_{k=0}^nC^k_np^k(1-p)^{n-k}g_k.$$

7
On

Let $X_i, X_j'\sim Be(p)$, and $Z_i, Z'_j\sim N(0,1)$ and $Y_i=\mu+\sigma Z_i,$ $Y'_j=-\mu+\sigma Z_j.$ Then $$E(|X_1Y_1+\cdots+X_nY_n+X'_1Y'_1+\cdots+X'_nY'_n|)$$$$=\sum_{k=0,r=0}^nC^k_n C^r_np^{k+r}(1-p)^{2n-p-r}E(|(k-r)\mu+\sigma \sqrt{k+r} Z_1|)$$$$=\sum_{k=0,r=0}^nC^k_n C^r_np^{k+r}(1-p)^{2n-p-r} \sqrt{k+r}\sigma f(\frac{r-k}{\sigma \sqrt{k+r}}\mu)$$