2
$\begingroup$

I have read a paper that says that the following is exponentially distributed

$$ Y= \bigl| \sum_{i=1}^n \gamma_i^{-\frac{1}{2}} h_i \bigl|^2$$

where $\gamma_i$ are non-negative constants and

$$h_i \sim \mathcal{CN} (0,1)\,,$$

where $\mathcal{CN}(\mu,\Gamma)$ denotes a circularly symmetric complex normal distribution.

The authors claim that the mean of this exponential distributed $Y$ is $$\sum_{i=1}^n{\gamma_i}^{-1}$$.

Does anyone know why?

My thoughts are the following

1- The sum of Gaussian is also Gaussian

2- Magnitude of Gaussian is Rayleigh distributed

3- Taking the square of Rayleigh is exponential

But how do we get that the mean ?

My second part of this question is

What happens if $$h_{i} \sim \text{Nakagami } m $$

$\endgroup$
  • $\begingroup$ @GLen_b complex normal $\endgroup$ – Tyrone Nov 6 '14 at 21:58
  • $\begingroup$ I am actually looking for two things, 1- Why is the mean $\sum_{i=1}^n \gamma^{-1}$ for the case of Normal Gaussian 2- What happens , what is the distribution of $Y$ if I redefine the $h_i$ to be Nakagami iid parameter m. $\endgroup$ – Tyrone Nov 6 '14 at 22:26
  • $\begingroup$ It says in the wiki page, it is usually denoted as... same way I write it above $\endgroup$ – Tyrone Nov 6 '14 at 22:29
  • $\begingroup$ this is circullary symmetric are you comfortable now? $\endgroup$ – Tyrone Nov 6 '14 at 22:44
  • $\begingroup$ Oh, I forgot to check .. are the $h_i$ meant to be independent of each other? $\endgroup$ – Glen_b Nov 6 '14 at 23:21
1
$\begingroup$

Just addressing the first part for now:

Your line of thinking about why it's exponential is essentially along the right lines, after small modifications to get some details correct (you'll need independence in your step 2 for example, to invoke the Rayleigh).

For constant $c_i$, I think we have that $c_i h_i \sim \mathcal{CN}(0,c_i^2I)$ (where here $\mathcal{CN}$ with two arguments refers to the circularly symmetric complex normal).

This along with linearity of expectation explains where the $\sum \gamma_i^{-1}$ comes from.


Answer to Q from comments:

how does the scale parameter $γ^{−1}_i$ change the distribution of $|h_i|^2$?

If $h_i^2 \sim \text{Gamma}(\alpha, \beta)$ (where here I mean the shape-scale parameterization, not the shape-rate parameterization), then

$γ^{−1}_i\,h_i^2\sim\text{Gamma}(\alpha, γ^{−1}_i\beta)$.

(If you want the shape-rate parameterization, you multiply the rate parameter by $\gamma_i$)

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I think our reasoning is correct then, do you happen to know what happens if those $h_i$ are Nakagami distrbuted ? $\endgroup$ – Tyrone Nov 6 '14 at 23:40
  • $\begingroup$ There's a reason I said "Just addressing the first part for now". I've heard of it but I need to go look it up. $\endgroup$ – Glen_b Nov 6 '14 at 23:44
  • $\begingroup$ I am not sure there is a closed form of the sum of Nakagami, I looked myself. I think convolution is the way to go if one assumes the summand are independent. $\endgroup$ – Tyrone Nov 6 '14 at 23:49
  • $\begingroup$ It appears to be a scaled chi distribution. Okay -- I also don't think linear combinations of such variables are going to have a simple closed form. $\endgroup$ – Glen_b Nov 7 '14 at 0:33
  • $\begingroup$ I think for the Nakagami case, we can find the distribution of $\sum_{i=1}^n \gamma_i^{-1}|h_i|^2$ instead rather than $|\sum_{i=1}^n\gamma_i^{-\frac{1}{2}}h_{i}|^2$ $\endgroup$ – Tyrone Nov 7 '14 at 0:47
1
$\begingroup$

A standard circularly symmetric complex normal random variable is of the form $Z = X + iY$ where $X$ and $Y$ are independent zero-mean normal random variables with variance $\frac 12$. Here $i = \sqrt{-1}$. We write $Z \sim \mathcal{CN}(0,1)$ and $X, Y \sim \mathcal{N}(0,\frac 12)$. Note that $$E[|Z|^2] = E[X^2+Y^2]= \frac 12 + \frac 12 =1.$$ Note also that $|Z|^2$ is an exponential random variable with parameter (and hence mean) equal to $1$. From this, it is easy to deduce that $\sigma Z \sim \mathcal{CN}(0,\sigma^2)$ and that $|\sigma Z|^2$ is an exponential random variable with mean $\sigma^2$ and hence parameter $\frac{1}{\sigma^2}$.

Thus, since the $H_k$ are independent $\mathcal{CN}(0,1)$ random variables, meaning that $H_k = X_k + i Y_k$ where $X_k$ and $Y_k$ are independent $\mathcal N(0,\frac 12)$ random variables, $\gamma_k^{-1/2}H_k$ is a $\mathcal{CN}(0,\gamma_k^{-1})$ random variable, and so $$Z = \sum_{k=1}^N \gamma_k^{-1/2}H_k = \sum_{k=1}^N\gamma_k^{-1/2}(X_k+iY_k) = \bar{X} + i\bar{Y} \sim \mathcal{CN}\left(0,\sum_k \gamma_k^{-1}\right).$$ It follows that $|Z|^2 = \displaystyle \left|\sum_{k=1}^N \gamma_k^{-1/2}H_k\right|$ is an exponential random variable with mean $\sum_k \gamma_k^{-1}$.

About the only claim that does not follow from simple applications of general rules is that if $X$ and $Y$ are independent zero-mean normal random variables with common variance $\frac 12\sigma^2$, then $X^2+Y^2$ has an exponential distribution. Note that the usual method changing to polar coordinates gives $$P\{X^2+Y^2 > z\} =\int_{\sqrt{z}}^\infty \int_0^{2\pi} \frac{1}{\pi\sigma^2} r\exp(-r^2/\sigma^2) \,\mathrm d\theta \, \mathrm dr = \exp(-z/\sigma^2)$$ showing that $X^2+Y^2$ has an exponential distribution with parameter $\sigma^{-2}$ and hence mean $\sigma^2$.

| cite | improve this answer | |
$\endgroup$

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.