Pdf of Difference of Random Variables

  • I
  • Thread starter marcusl
  • Start date
  • #1
marcusl
Science Advisor
Gold Member
2,714
377

Main Question or Discussion Point

I want to find the probability density function (pdf) of the difference of two RV's,
[tex]p_{\Delta Y} = p_{(Y_1 - Y_2)},[/tex]where [itex]y = \sin \theta[/itex], and where [itex]\theta_1[/itex] and [itex]\theta_2[/itex] are random variables with the same uniform distribution [itex]p_{\theta}=\mathrm{rect}\left(\frac{\theta}{\pi}\right)[/itex]. This has support [itex]-\pi/2\leq\theta\leq \pi/2[/itex]. (Please let me know if I am misusing terminology, as Math is not my native language :oops:)

I can derive the distribution of y [tex]p_Y=\frac{\mathrm{rect}\left(\frac{y}{2}\right)}{\pi \sqrt{1-y^2}}[/tex]by starting from the uniform distribution. This is nonzero over [itex]|y| \leq 1[/itex]. I get bogged down with the pdf of Δy, however. A direct approach involving the convolution[tex]p_{\Delta Y}(\Delta y)=p(y)*p(y)=\int_{-\infty}^{\infty}p_Y(\Delta y-u)p_Y(u)du[/tex]results in the square root of a quartic in the denominator, which seems like a dead end.

I instead tried characteristic functions. From tables of Fourier transforms (FT's), I find that the CF of [itex]p_Y[/itex] is a Bessel function[tex]\varphi_Y(z)=\frac{1}{2\pi}J_0(z)[/tex][itex]p_{\Delta Y}[/itex] is then the inverse FT of the product of two of these CF's[tex]p_{\Delta Y}=F^{-1}(\varphi_{\Delta Y})=F^{-1}(\varphi_Y^2)=\frac{F^{-1}\left(J_0^2(z)\right)}{4\pi^2}[/tex]but I cannot find the inverse FT of the square of the Bessel function. Can anyone help me finish this off?
 

Answers and Replies

  • #2
andrewkirk
Science Advisor
Homework Helper
Insights Author
Gold Member
3,792
1,390
A direct approach involving the convolution[tex]p_{\Delta Y}(\Delta y)=p(y)*p(y)=\int_{-\infty}^{\infty}p_Y(\Delta y-u)p_Y(u)du[/tex]......
That doesn't look quite right to me. I think it should be:
$$
p_{\Delta Y}(\Delta y)=p(y)*p(y)=\int_{-1}^{1}p_Y(\Delta y+u)p_Y(u)du
$$
which is equal to:
$$
\int_{-1-\min(\Delta y,0)}^{1-\max(\Delta y, 0)}p_Y(\Delta y+u)p_Y(u)du
$$
I would be inclined to stop there and just use numerical integration. Expressing the result as a Bessel function may look more compact, but will still involve numerical integration to evaluate the Gamma function inside the Bessel sum.
 

Related Threads on Pdf of Difference of Random Variables

Replies
5
Views
1K
Replies
4
Views
6K
  • Last Post
Replies
7
Views
2K
Replies
3
Views
7K
Replies
3
Views
2K
Replies
7
Views
3K
Replies
3
Views
1K
Replies
7
Views
3K
  • Last Post
Replies
5
Views
2K
Top