3
$\begingroup$

The problem

Let there be two independent random variables given. Let's name them $X$ and $Y$. Suppose the variables have both normal distribution $N(\mu, \lambda^2).$

Let's define third random variable $Z = X + Y$. The task is to find the density function of $Z$.

My attempt

I know that the density function can be easily calculated using the distribution function (I would need to differentiate it). That's why I started with finding the distribution. $$F_Z(t) = Pr(Z \le t) = Pr(Y + X \le t) = Pr(Y \le t - X) = \int \limits_{-\infty}^{\infty}Pr(Y \le t-x)f_X(x) dx$$ The equation above leads us to the following one: $$ \int \limits_{-\infty}^{\infty} \int \limits_{-\infty}^{t-x} f_Y(y)f_X(x) \mbox{d} y\mbox{d}x \tag{1}.$$ Using the knowledge about $f_X$ and $f_Y$ we can write $(1)$ in this form: $$\theta\int \limits_{-\infty}^{\infty} \int \limits_{-\infty}^{t-x}e^{-\frac{(y- \mu)^2}{\lambda^2}} e^{-\frac{(x- \mu)^2}{\lambda^2}} \mbox{d} y\mbox{d}x,$$ where $\theta = \frac{1}{\sqrt{2 \lambda^2}}.$
I wonder if my attempt was correct because the integral is quite horrible. Is it possible to calculate it in a "nicer" way?


Is $(1)$ equivalent to this expression $(f_X*f_Y)(x)$?
If yes I would get the following integral: $$\int \limits_{\mathbb{R}} e^{-\frac{(y- \mu)^2}{\lambda^2}} e^{-\frac{(x- y - \mu)^2}{\lambda^2}} \mbox{d}y.$$
Is there a method of calculating such things?

$\endgroup$
  • 2
    $\begingroup$ en.wikipedia.org/wiki/… $\endgroup$ – saulspatz Jun 4 '18 at 18:50
  • $\begingroup$ @Hendrra In general, sum of random variables $X$ and $Y$ with densities $f$ and $g$ has density being convolution of $f$ and $g$, you are correct $\endgroup$ – Jakobian Jun 4 '18 at 19:04
  • $\begingroup$ @Adam Thank you. What about my first approach? $\endgroup$ – Hendrra Jun 4 '18 at 19:33
1
$\begingroup$

From your expression, $F_Z(t) = \int_\mathbb{R}P(Y \leq t-x)f_X(x)dx$, differentiate to obtain the density (you may need to justify why you can differentiate under the integral),

$$ f_Z(t) = \int_\mathbb{R}f_Y(t-x)f_X(x)dx = \int_\mathbb{R}\frac{1}{\sqrt{2\pi}}e^{-(t-x-\mu)^2/2\lambda^2}\frac{1}{\sqrt{2\pi}}e^{-(x-\mu)^2/2\lambda^2}dx$$

which we recognize as $(f_Y*f_X) (x)$. To actually evaluate the integral, expand the exponent terms inside the integrand,

$$-\frac{1}{2\lambda^2}\left[(t-x-\mu)^2 + (x-\mu)^2\right] = -\frac{1}{2\lambda^2}(t^2-2\mu t + 2\mu^2 + \underbrace{2x^2-2tx}_{(*)})$$

Complete the square on the $(*)$ term, $2x^2 - 2tx = 2\left(x-\frac{t}{2}\right)^2-\frac{t^2}{2}$. Putting things together,

$$-\frac{1}{2\lambda^2}(t^2-2\mu t + 2\mu^2 + 2x^2-2tx) = -\frac{1}{4\lambda^2}(t^2-4\mu t + 4\mu^2)-\frac{1}{\lambda^2}(x-t/2)^2$$

So, we can pull out the $t$ terms from the integrand to be left with,

$$ f_Z(t) = \frac{1}{\sqrt{4\pi\lambda^2}}e^{-(t-2\mu)^2/4\lambda^2}\underbrace{\int_\mathbb{R}\frac{1}{\sqrt{\pi\lambda^2}}e^{-(x-t/2)^2/\lambda^2}dx}_{=1} = \frac{1}{\sqrt{4\pi\lambda^2}}e^{-(t-2\mu)^2/4\lambda^2}$$

The integral above equals $1$ because the integrand is the density function of a $N(t/2,\lambda^2/2)$ distribution.

$\endgroup$
  • $\begingroup$ At the end, did you intend $N(2 \mu,2 \lambda^2)$ for the mean and variance rather than $t/2, \lambda^2 /{2}$? $\endgroup$ – Henry Jun 5 '18 at 16:31
  • $\begingroup$ Sorry, I was referring to the density function inside the final integral (the one that equals 1). The distribution of $Z$ is definitely $N(2\mu,2\lambda^2)$. $\endgroup$ – Flowsnake Jun 5 '18 at 16:33
  • $\begingroup$ So was I. I think a density $f_Z(t)= \frac{1}{\sqrt{4\pi\lambda^2}}e^{-(t-2\mu)^2/4\lambda^2}$ suggests a normal distribution for $Z$ with a mean of $2 \mu$ and a variance of $2 \lambda^2$ $\endgroup$ – Henry Jun 5 '18 at 16:37
  • $\begingroup$ The density function I was referring to was, $\frac{1}{\sqrt{\pi\lambda^2}}e^{-(x-t/2)^2/\lambda^2}$, which is the density of a $N(t/2,\lambda^2/2)$. $\endgroup$ – Flowsnake Jun 5 '18 at 16:49
-1
$\begingroup$

The problem can also be solved by using the following property: If $X\sim N(\mu_1,\sigma_1^2) $ and $Y \sim N(\mu_2,\sigma_2^2)$ and are independent, then $aX + bY \sim N(a\mu_1 + b\mu_2,a^2\sigma_1^2 +b^2\sigma_2^2)$ and thus in your case $Z \sim N(2\mu,2\lambda^2)$.

Edit: I just noticed the comment by @saulspatz above who already pointed this out.

$\endgroup$

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.