4
$\begingroup$

I am trying to find different methods of showing that there exists no random variable $V$ that can additively transform standard normal random variable into Laplace random variable.

Formally, we would like to show that there is no random variable $V$ independent of $N$(standard normal) such that \begin{align} W=V+N, \end{align} where $W$ is zero mean and with Laplace parameter $b\ge 1$. We want to show this is impossible for all $b\ge 1$.

Proof 1 (Via Characteristic Functions): I have a proof via characteristic functions that goes as follows \begin{align} \phi_W(t)=\phi_V(t) \phi_N(t) \rightarrow \phi_V(t)=\frac{\phi_N(t)}{\phi_W(t)} \end{align} now we know that $\phi_N(t)=e^{-\frac{t^2}{2}}$ and $\phi_W(t)=\frac{1}{1+b^2 t^2}$, so \begin{align} \phi_V(t)= \frac{e^{\frac{t^2}{2}}}{1+b^2 t^2} \end{align} Clear this can not be a characteristic function since there are values of $t$ for which $\phi_V(t) >1$.

Proof 2 (Via analyticity after convolution) This proof was suggested by D.Thomine (see comments below). Since convolution 'improves' regularity the pdf of $W$ must be analytic, however, the pdf of Lapalace distribution is not analytic at zero.

My question: What would be some other methods of showing this?

I was also thinking of the following approach. We know that \begin{align} c_1 e^{-|w|/b} &= E[ c_2 e^{-|X+w|^2/2} ] \end{align}

Can we show that \begin{align} E[ c_2 e^{-|X+w|^2/2} ] \le c_3 e^{-w^2/2} \end{align} this would lead to a contraditiction since $e^{-w^2/2}$ decays faster than $e^{-|w|/b}$.

This question is related to something I asked here .

Looking forward to your solutions. Thank you.

$\endgroup$
6
  • 3
    $\begingroup$ Another way of seeing the Fourier transform approach : when adding two independent random variables, you convole (?) the underlying measures on $\mathbb{R}$. Convolution improves regularity. Hence, $N+V$ has an analytic density for any $V$ independent of $N$. The distribution density of a Laplace random variable is merely continuous. $\endgroup$ – D. Thomine Dec 9 '16 at 22:57
  • $\begingroup$ Great answer. This is a nice approach. Big (+1). Please put this as an answer. $\endgroup$ – Boby Dec 9 '16 at 23:09
  • $\begingroup$ @D.Thomine When you say that it is analytic you mean that its characteristic function can be written as a power series? $\endgroup$ – Boby Dec 9 '16 at 23:18
  • 1
    $\begingroup$ Yes, this comes from some measure-theoretical / complex version of the "differentiation under integral" kind of results. The only difficulty is locating a good reference ; I guess that this does the trick : nieuwarchief.nl/serie5/pdf/naw5-2001-02-1-032.pdf. $\endgroup$ – D. Thomine Dec 9 '16 at 23:22
  • $\begingroup$ @D.Thomine Thanks. Got it. Please put as an answer if you want. $\endgroup$ – Boby Dec 9 '16 at 23:33
2
$\begingroup$

What you use to solve this problem is that the Fourier transform of a standard Gaussian decays much faster than the Fourier transform of $e^{-|x|}$. If we handwave a little, this means that the Gaussian is much smoother than $e^{-|x|}$, which is something we can exploit directly.

The main idea is classical : adding an independent real-valued random variable is equivalent, at the distribution level, with convolution the densities, and convolution improves smoothness. This is often used when we want to deal with smooth distribution; we just need to add some random noise to smooth any initial distribution.

Proposition

Let $N$ be a standard normal random variable. Let $V$ be a real-valued random variable, independent from $N$. Then the density of the distribution of $N+V$ has an entire version.

Proof

Let $\mu$ be the distribution of $V$. Then $(\mathbb{R}, \mathcal{B}, \mu)$ is a measure space.

For $M >0$, let $G_M := \{z \in \mathbb{C}: \ |Im(z)| <M\}$. Then $G$ is open, and for $z =: x+iy \in G$,

$$\left| \frac{1}{\sqrt{2 \pi}}e^{-\frac{z^2}{2}} \right| = \frac{1}{\sqrt{2 \pi}}e^{\frac{y^2-x^2}{2}} \leq \frac{1}{\sqrt{2 \pi}}e^{\frac{M^2}{2}}.$$

For $z \in G_M$ and $\omega \in \mathbb{R}$, let $f(z,\omega) := \frac{1}{\sqrt{2 \pi}}e^{-\frac{(z-\omega)^2}{2}}$. Then $f$ is bounded, holomorphic in the first variable, and measurable in the second. Hence, using complex differentiation under the integral, the function :

$$F: z \mapsto \int_{\mathbb{R}} \frac{1}{\sqrt{2 \pi}}e^{-\frac{(z-\omega)^2}{2}} d \mu (\omega)$$

is holomorphic on $G_M$. Since this is true for all $M>0$, the function $F$ is holomorphic on $\mathbb{C}$, and thus, entire. But $F$ is exactly the distribution of $N+V$.

~~~~~

All we need to conclude is the observation that the density of a Laplace random variable is never smooth at $0$.

$\endgroup$

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.