9
$\begingroup$

Is it possible to have the PDF of the difference of two iid r.v.'s look like a rectangle (instead of, say, the triangle we get if the r.v.'s are taken from the uniform distribution).

i.e. is it possible for the PDF f of j-k (for two iid r.v.'s taken from some distribution) to have f(x) = 0.5 for all -1 < x < 1?

There are no restrictions on the distribution we take j and k from except that the min is -1 and the max is 1.

After some experimentation, I'm thinking this might be impossible.

$\endgroup$
  • $\begingroup$ The difference of two uniform distributions is a triangular distribution, so if you ask if it is possible to get uniform of a difference of i.i.d. uniforms, then the answer is not. $\endgroup$ – Tim May 15 '18 at 6:48
  • $\begingroup$ Same Q asked here: math.stackexchange.com/questions/2048939/… so far without answers! $\endgroup$ – kjetil b halvorsen May 15 '18 at 7:48
  • $\begingroup$ It would indeed seem difficult to avoid realizations outside $[-1,1]$ when both $j$ and $k$ have probability mass close to these endpoints. $\endgroup$ – Christoph Hanck May 15 '18 at 9:54
  • 2
    $\begingroup$ It is not possible. To my recollection this is (in slightly different form) already answered somewhere on site. I'll see if I can locate it $\endgroup$ – Glen_b Jun 19 '18 at 10:07
  • 1
    $\begingroup$ @Glen_b You might be recalling stats.stackexchange.com/questions/125360/…. It's not quite a duplicate, though, because a difference $X-Y$ of iid variables, although expressible as a sum $X+(-Y),$ could involve a sum of variables with non-identical distributions. I believe a trivial modification of my solution will address this difference; Silverfish's solution looks like it applies directly with almost no modification, but first one has to remove a lot of extraneous material to see that. $\endgroup$ – whuber Jun 19 '18 at 13:37
9
$\begingroup$

Theorem: There is no distribution $\text{Dist}$ for which $A-B \sim \text{U}(-1,1)$ when $A, B \sim \text{IID Dist}$.


Proof: Consider two random variables $A, B \sim \text{IID Dist}$ with common characteristic function $\varphi$. Denoting their difference by $D=A-B$. The characteristic function of the difference is:

$$\begin{equation} \begin{aligned} \varphi_D(t) = \mathbb{E}(\exp(t D)) &= \mathbb{E}(\exp(t (A-B))) \\[6pt] &= \mathbb{E}(\exp(t A)) \mathbb{E}(\exp(-t B)) \\[6pt] &= \varphi(t) \varphi(-t) \\[6pt] &= \varphi(t) \overline{\varphi(t)} \\[6pt] &= |\varphi(t)|^2. \\[6pt] \end{aligned} \end{equation}$$

(The fourth line of this working follows from the fact that the characteristic function is Hermitian.) Now, taking $D \sim \text{U}(-1,1)$ gives a specific form for $\varphi_D$, which is:

$$\begin{equation} \begin{aligned} \varphi_D(t) = \mathbb{E}(\exp(itD)) &= \int \limits_{\mathbb{R}} \exp(itr) f_D(r) dr \\[6pt] &= \frac{1}{2} \int \limits_{-1}^1 \exp(itr) dr \\[6pt] &= \frac{1}{2} \Bigg[ \frac{\exp(itr)}{it} \Bigg]_{r=-1}^{r=1} \\[6pt] &= \frac{1}{2} \frac{\exp(it)-\exp(-it)}{it} \\[6pt] &= \frac{1}{2} \frac{(\cos(t) + i \sin(t)) - (\cos(-t) + i \sin(-t))}{it} \\[6pt] &= \frac{1}{2} \frac{(\cos(t) + i \sin(t)) - (\cos(t) - i \sin(t))}{it} \\[6pt] &= \frac{1}{2} \frac{2i \sin(t)}{it} \\[6pt] &= \frac{\sin(t)}{t} = \text{sinc}(t). \\[6pt] \end{aligned} \end{equation}$$

where the latter is the (unnormalised) sinc function. Hence, to meet the requirements for $\text{Dist}$, we require a characteristic function $\varphi$ with squared-norm given by:

$$|\varphi(t)|^2 = \varphi_D(t) = \text{sinc}(t).$$

The left-hand-side of this equation is a squared norm and is therefore non-negative, whereas the right-hand-side is a function that is negative in various places. Hence, there is no solution to this equation, and so there is no characteristic function satisfying the requirements for the distribution. (Hat-tip to Fabian for pointing this out in a related question on Mathematics.SE.) Hence, there is no distribution with the requirements of the theorem. $\blacksquare$

$\endgroup$
3
$\begingroup$

This is an electrical engineer's take on the matter, with a viewpoint that is more suitable for dsp.SE rather than stats.SE, but no matter.

Suppose that $X$ and $Y$ are continuous random variables with common pdf $f(x)$. Then, if $Z$ denotes $X-Y$, we have that $$f_Z(z) = \int_{-\infty}^\infty f(x)f(x+z) \ \mathrm dx.$$ The Cauchy-Schwarz inequality tells us that $f_Z(z)$ has a maximum at $z=0$. In fact, since $f_Z$ is actually the "autocorrelation" function of $f$ regarded as a "signal", it must have a unique maximum at $z=0$ and thus $Z$ cannot be uniformly distributed as is desired. Alternatively, if $f_Z$ were indeed a uniform density (remember that it is also an autocorrelation function), then the "power spectral density" of $f_Z$ (regarded as a signal) would be a sinc function, and thus not a nonnegative function as all power spectral densities must be. Ergo, the assumption that $f_Z$ is a uniform density leads to a contradiction and so the assumption must be false.

The claim that $f_Z \sim \mathcal U[-1,1]$ is obviously invalid when the common distribution of $X$ and $Y$ contains atoms since in such a case the distribution of $Z$ will also contain atoms. I suspect that the restriction that $X$ and $Y$ have a pdf can be removed and a purely measure-theoretic proof constructed for the general case when $X$ and $Y$ don't necessarily enjoy a pdf (but their difference does).

$\endgroup$
  • $\begingroup$ Part of that doesn't seem right to me. The characteristic function of the $\text{U}(-1,1)$ distribution is the $\text{sinc}$ function, so clearly that kind of Fourier transform is allowable. Your logic seems to me to lead to prove too much - it appears to prove not only that $Z$ cannot be uniform, but that the uniform distribution cannot exist at all. Have I misunderstood? $\endgroup$ – Ben Jun 19 '18 at 23:58
  • 1
    $\begingroup$ Whether or not the characteristic function of $\mathcal U[-1,1]$ exists is not the the issue; it does exist. The pdf of $Z$ is an autocorrelation function. Well, the power spectral density of any autocorrelation function must be a nonnegative function. So, the assumption that $f_Z \sim \mathcal U[-1,1]$ leads to a power spectral density which is a sinc function (that takes on both positive and negative values). Since this is not a valid power spectral density (remember that $f_Z$ is an autocorrelation function also), the assumption that $f_Z \sim \mathcal U[-1,1]$ must be false. $\endgroup$ – Dilip Sarwate Jun 20 '18 at 14:21

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.