I know some methods to obtain the probability distribution of functions on a random variable:

CDF method: If $X$ is a random variable and $Y=f(X)$, then computing the cumulative distribution function $F(Y)=P(Y\leq y)=P(f(X)\leq y)=\int_{x=x_0}^{f^{-1}(y)} f(x)dx $, and $F'(y)=f_Y(y)$

Product of 2 probability distributions: X and Y are RV with $f_X(x)$, $f_Y(y)$ PDF, Z=XY, $f_Z(z) = \int_{-\infty}^{\infty}f_X(x) f_Y(z/x)\frac{1}{|x|} dx$

Sum of 2 probability distributions: X and Y are RV with $f_X(x)$, $f_Y(y)$ PDF,$Z=X+Y$, $f(z)=f_X*f_Y$ is the convolution of the two PDF.

I'm trying to obtain the probability density function of the function on 3 random variables $\theta$, $G_1$ and $G_2$:


$\theta$ uniformly distributed in $\theta\in[0,2\pi]$. $G_1$ and $G_2$ are independent Gaussians. I think that I can not apply the multiplication rule at both sides of the substraction and then the convolution of both just because of $\cos(\theta)$ and $\sin(\theta)$ are 2 functions on the same random variable, they are correlated...



2 Answers 2


One way is to use conditional distributions.

Assuming $G_1 \sim N(0,1), G_2 \sim N(0,1) , \theta \sim Uniform(0,2\pi)$ and that all variables are independent we can use the useful rule, that $$\cos(\theta)G_1 - \sin(\theta)G_2 \: | \: \theta =x \sim \cos(x)G_1-\sin(x)G_2$$ Now for any $x\in[0,2\pi]$ we have by properties of the normal distribution, that $$\cos(x)G_1 - \sin(x)G_2 \sim N(0,\cos^2(x)+\sin^2(x)) = N(0,1).$$ But this means that the conditional distribution of $\cos(\theta)G_1 - \sin(\theta)G_2$ given $\theta$ does not depend on the value of $\theta$ and we can thereby conclude, that $$\cos(\theta)G_1-\sin(\theta)G_2 \sim N(0,1)$$ and also that $\cos(\theta)G_1-\sin(\theta)G_2$ is independent of $\theta$.

Edit: More generally, if we are able to find the conditional pdf $f_{Y|X}(y \: | \: x)$, where $X$ is a $k$-dimensional vector, then we can find the joint pdf as $f_{X,Y}(x_1,x_2,...,x_k,y)=f_{Y|X}(y|x_1,x_2,\dots,x_k)f_X(x_1,\dots,x_k)$ and from here we can find the marginal pdf of $Y$ as $$f_Y(y) = \int_{\mathbb{R}^k} f_{X,Y}(x_1,x_2,...,x_k,y) \: d(x_1,x_2,\dots , x_k)$$

Now suppose $Y= \frac{\cos(\theta_1)+\cos(\theta_2)}{2}G_1 - \frac{\sin(\theta_1)+\sin(\theta_2)}{2}G_2$ and $X=(\theta_1,\theta_2)$. We can again use the useful rule to conclude, that \begin{align*} Y|X=(x_1,x_2) &\sim N(0,(\frac{\cos(x_1)+\cos(x_2)}{2})^2 + (\frac{\sin(x_1)+\sin(x_2)}{2})^2 ) \\ &= N(0,\cos^2(\frac{x_1-x_2}{2})) \end{align*} This distribution is degenerate if $\cos^2(\frac{x_1-x_2}{2})=0$, but if $X$ is assumed continuous, then we can ignore this case, since it will happen with probability $0$. From here you can compute \begin{align*} f_Y(y) &= \int_{-\infty}^\infty \int_{-\infty}^\infty f_{Y|X}(y \: | \: x_1,x_2)f_{\theta_1,\theta_2}(x_1,x_2) \: dx_1 dx_2 \\ &= \int_{0}^{2\pi} \int_{0}^{2\pi} \frac{1}{\sqrt{2\pi}\cos(\frac{x_1-x_2}{2})}\exp(-\frac{y^2}{2\cos^2(\frac{x_1-x_2}{2})}) f_{\theta_1,\theta_2}(x_1,x_2) dx_1 dx_2 \end{align*} assuming that angles are in $(0,2\pi)$.


Thanks Kristensen!

Damn, maybe I shown the worst example, I understood that you reduced the problem to a linear combination of Normal distributions due to the relations $\sin(x)^2 + \cos(x)^2 = 1$, $aN(0,1) = N(0,a^2)$ and $N(0,b)+N(0,c) = N(0,b+c)$.

What if I have some distribution in $\theta$ ($\theta_1$, $\theta_2$) which not reduces the problem in this way?

For instance:

$\frac{\cos(\theta_1) + \cos(\theta_2)}{2}G_1 - \frac{\sin(\theta_1) + \sin(\theta_2)}{2}G_2$

Doing simulations I know that the PDF must have a shape similar to $e^{-|x|}$ but I'm not able to obtain the pdf.

Thanks a lot!

  • $\begingroup$ I made an edit to the answer that handles the more general case. I do not believe that it was a bad example. Simple examples often improve understanding much more than complicated examples. $\endgroup$ May 17, 2020 at 14:52

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.