3
$\begingroup$

I need to choose the probability distribution $f_Y(y)$ of a random variable $Y$ such that the variable $Z=Y\cos(X)$ is normally distributed with zero mean, i.e. $$f_Z(z)=\frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{z^2}{2\sigma^2}}$$ for the independent and uniform distributed variable $X$, i.e. $$f_X(x)=\frac{1}{2\pi}$$ for $x\in[0,2\pi]$.


I started looking for the probability distribution of $K=\cos(X)$ and found $f_K(k)=\frac{1}{\pi\sqrt{1-k^2}}$ for $k\in[-1,1]$. Next, I wrote down the product distribution for $Y$ and $K$ (http://en.wikipedia.org/wiki/Product_distribution)

$$f_Z(z)=\int_{-\infty}^\infty{f_Y(x)f_K\left(\frac{z}{x}\right)\frac{1}{|x|}dx}$$

and put in what I have:

$$\int_{|z|<|x|}{f_Y(x)\frac{1}{\pi\sqrt{1-z^2/x^2}}\frac{1}{|x|}dx}\stackrel{!}{=}\frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{z^2}{2\sigma^2}}$$

But now I´m stuck. How to find the $f_Y$ that fulfills this equation?


Response to answers

Great, $Y=\sqrt{-2 \ln(U)}$ with $U\sim\mathcal{U}[0,1]$ does the job for $Z\sim\mathcal{N}_{0,1}$ and thus $Y=\sqrt{-2\sigma ^2 \ln(U)}$ leads to $Z\sim\mathcal{N}_{0,\sigma}$, so:

$$f_Y(y)=\frac{y}{\sigma ^2}e^{-\frac{y^2}{2\sigma ^2}}$$


Extension to nonzero mean

Now how would one go about extending this to allow for nonzero means, i.e. $Z´=Y´\cos(X)\sim\mathcal{N}_{\mu,\sigma}$?

We can write $Z´=Z+\mu$ and thus $Y´=Y+\mu/\cos(X)$ and find for $Q=\mu/\cos(X)$ $f_Q(q)=\frac{|\mu|}{q^2\sqrt{1-\mu^2/q^2}}$ for $|q|>|\mu|$. Then, because Y and Q are independent, the probability distribution for $Y´$ should be the convolution $$f_{Y´}(y´)=\int_{-\infty}^\infty{f_Y(y´-\tau)f_Q(\tau)d\tau}$$ correct? And now i´m stuck on this :-)

$f_{Y´}(y´)$ should look like the blue curve here: Mathematica code for $f_{Y'}(y)$

$\endgroup$
3
$\begingroup$

Suppose $Y$ is exponentially distributed with expected value $\sigma$ so that $\Pr(Y>y)=e^{-y/\sigma}$, and let \begin{align} Z_1 & = \sqrt{-2\log_e Y}\,\cos X, \\ Z_2 & =\sqrt{-2\log_e Y}\,\sin X. \end{align} Then each of $Z_1,Z_2$ is normally distributed with expected value $0$ and variance $\sigma^2$ and $Z_1,Z_2$ are indepedent.

See this Wikipedia article about the Box–Muller transform.

Postscript in response to comments: If $Z_1,Z_2\sim\text{i.i.d. }N(0,1)$ then their joint density is $$ f_{(Z_1,Z_2)} (z_1,z_2) = \text{constant}\cdot e^{-z_1^2/2} e^{-z_2^2/2} = \text{constant} \cdot e^{-(z_2^2+z_2^2)/2}. $$ Since this density is constant on circles defined by $z_a^2+z_2^2=\text{radius}^2$, the angle that the random point makes with the positive $z_1$-axis is uniformly distributed between $0$ and $2\pi$. Hence its coordinates are the cosine and sine of a uniformly distributed random angle. Then the remaining question is that of the distribution of the distance of the random point from the origin.

$\endgroup$
  • $\begingroup$ This is assuming one knows the answer and checking it, not finding the answer. $\endgroup$ – Did Sep 22 '14 at 14:26
  • $\begingroup$ @Did : Perhaps, but what "one knows" from standard courses in my experience would include the fact that if $Z_1,Z_2\sim\text{i.i.d. }N(0,1)$ then $Z_1^2+Z_2^2\sim\chi_2^2$, and that is an exponential distribution with expected value $2$, and that the bivariate normal distribution with covariance $0$ has a circular symmetry (hence $(\cos X,\sin X)$, where $X$ is uniformly distrtibuted between $0$ and $2\pi$). ${}\qquad{}$ $\endgroup$ – Michael Hardy Sep 22 '14 at 14:31
  • $\begingroup$ Not judging from what the OP shows (since, for once, we have an OP who actually shows something...). $\endgroup$ – Did Sep 22 '14 at 14:49
0
$\begingroup$

We have: $$\begin{eqnarray*}\mathbb{P}\left[Y=\frac{Z}{\cos(X)}\in[0,\eta]\right]&=&\frac{1}{\sqrt{2}\,\sigma\,\pi^{3/2}}\int_{0}^{\eta}\int_{-1}^{1}\frac{e^{-\frac{(ux)^2}{2\sigma^2}}}{\sqrt{1-x^2}}\,dx\,du\\&=&\frac{1}{\sigma\sqrt{2\pi}}\int_{0}^{\eta}e^{-\frac{u^2}{4\sigma^2}}\,I_0\left(\frac{u^2}{4\sigma^2}\right)\,du\end{eqnarray*}$$ where $I_0$ is a modified Bessel function of the first kind, giving:

$$ f_Y(y) = \frac{1}{\sigma\sqrt{2\pi}} e^{-\frac{y^2}{4\sigma^2}}\sum_{m=0}^{+\infty}\frac{1}{m!^2}\left(\frac{y^4}{64\sigma^4}\right)^m.$$

The pdf of $Y$ in a neighbourhood of zero can be very well approximated by the pdf of a Cauchy distribution.

$\endgroup$
  • $\begingroup$ Seems to me you're making this too complicated. Doesn't the standard Box-Muller transformation do it? $\endgroup$ – Michael Hardy Sep 22 '14 at 14:17
  • $\begingroup$ Box-Muller or Mellin, the pdf of $Y$ is still the product of a Gaussian and a Bessel function. $\endgroup$ – Jack D'Aurizio Sep 22 '14 at 14:19

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.