I would have a proof of the following fact; but it's a bit clunky, and am wondering if one can get a more elegant one (and/or improve the constants). I couldn't find this anywhere, and searching properties of the Skellam distribution didn't help much either.

Let $X\sim\operatorname{Poi}(\lambda)$ and $Y\sim\operatorname{Poi}(\mu)$ be two independent random variables. Then $$ \max(|\lambda-\mu|, \frac{1}{40}\min(\sqrt{\lambda+\mu}, \lambda+\mu)) \leq \mathbb E[|X-Y|] \leq\min(|\lambda-\mu|+\sqrt{\lambda+\mu}, \lambda+\mu) $$

Here is the current proof I came up with:

The upper bound follows from Cauchy–Schwarz, as $\mathbb E[(X-Y)^2] = (\lambda-\mu)^2 + \lambda+\mu$; and by the triangle inequality, as $\mathbb{E}[|X-Y|]\leq \mathbb{E}[X+Y]$. The lower bound $|\lambda-\mu|$ is a direct consequence of Jensen's inequality; we now proceed to establish the second term.

  • Consider first the case $\lambda+\mu\geq 1$. We will use Paley–Zygmund, noting that for any $c\in(0,1)$ we have $$ \mathbb{E}[|X-Y|] \geq c\sqrt{\mu+\lambda}\cdot\mathbb{P}\{|X-Y|\geq c\sqrt{\mu+\lambda} \} $$ and $$ \begin{align*} \mathbb{P}\{|X-Y|&\geq c\sqrt{\mu+\lambda}\} \\ &= \mathbb{P}\{(X-Y)^2\geq c^2(\mu+\lambda)\} \\ &\geq \mathbb{P}\{(X-Y)^2\geq c^2\mathbb{E}[(X-Y)^2]\}\\ &\geq (1-c^2)^2\frac{\mathbb{E}[(X-Y)^2]^2}{\mathbb{E}[(X-Y)^4]}\\ &= \frac{ (1-c^2)^2(\lambda+\mu+(\lambda-\mu)^2)^2}{\lambda+\mu+6(\lambda^2+\mu^2)+6(\lambda-\mu)^2(\lambda+\mu)+(\lambda-\mu)^2+(\lambda-\mu)^4}\\ &\overset{(\dagger)}{\geq} (1-c^2)^2 \frac{(\lambda+\mu)^2+(\lambda-\mu)^4}{(\lambda+\mu)+10(\lambda+\mu)^2+3(\lambda-\mu)^4}\\ &\overset{(\ddagger)}{\geq} \frac{(1-c^2)^2}{11} \end{align*} $$ where $(\dagger)$ is using the AM-GM inequality and $(a-b)^2\leq (a+b)^2$; and $(\ddagger)$ relies on $\lambda+\mu \leq (\lambda+\mu)^2$, which holds since $\lambda+\mu\geq 1$. Taking $c=1/\sqrt{5}$ concludes the proof of this case, as then $\frac{c(1-c^2)^2}{11} > \frac{1}{40}$.

  • If $\lambda+\mu \leq 1$, we will use properties of the modified Bessel function of the first kind $I_0$ (namely, that it is nondecreasing on $[0,\infty)$, with $e^{-x}I_0(x) \geq 1-\frac{x}{2}$ for $x\in[0,1]$) to conclude, as $$ \begin{align} \mathbb{E}[|X-Y|] &\geq 1 - \mathbb{P}\{|X-Y|=0\} \\ &= 1- e^{-(\lambda+\mu)} I_0(2\sqrt{\lambda\mu}) \\ &\geq 1- e^{-(\lambda+\mu)} I_0(\lambda+\mu) \geq \frac{\lambda+\mu}{2} \end{align}$$ where the first inequality is the AM-GM inequality.


$\newcommand{\la}{\lambda}$Let us show that \begin{equation*} E|Z|\ge J(1)\min[c,\sqrt c\,], \tag{1} \end{equation*} where \begin{equation*} Z:=X-Y,\quad c:=\la+\mu, \end{equation*} \begin{equation*} J(x):=\frac2\pi\,\int_0^\infty\frac{1-\exp\{-x\, (1-\cos t)\}}{t^2}\,dt. \end{equation*} Mathematica's command NIntegrate[] produces \begin{equation*} J(1)=0.67390\dots. \end{equation*} So, the constant factor $J(1)$ is an about 27-times improvement of the corresponding coefficient $\frac1{40}$ in the OP. Moreover, the constant factor $J(1)$ is the best possible.

Inequality (1) is based on the Zolotarev identity \begin{equation*} E|Z|=\frac2\pi\,\int_0^\infty\frac{1-\Re Ee^{itZ}}{t^2}\,dt; \end{equation*} see e.g. formula (3.26) in this paper or formula (48) in the corresponding arXiv preprint. In our case, \begin{equation*} \Re Ee^{itZ}=\exp\{-c(1-\cos t)\}\cos((\la-\mu)\sin t)\le\exp\{-c(1-\cos t)\}; \end{equation*} Moreover, the latter inequality will turn into the equality when $\la=\mu$.

So, \begin{equation*} E|Z|\ge J(c), \tag{2} \end{equation*} and the latter inequality will turn into the equality when $\la=\mu$.

Note that, for each real $z>0$, \begin{equation*} \frac{1-e^{-zx}}x=\int_0^z e^{-xu}\,du \end{equation*} is decreasing in $x$. So, $J(x)/x$ is decreasing in real $x>0$. Moreover, as shown in this answer, $J(x)/\sqrt x$ is increasing in real $x>0$. So, $J(c)\ge cJ(1)$ if $c\in(0,1]$ and $J(c)\ge\sqrt c\,J(1)$ if $c\in[1,\infty)$.

Thus, (1) follows from (2). Moreover, (1) turns into the equality when $\la=\mu=1/2$ (so that $c=1$).

  • 2
    $\begingroup$ That's great, thanks! From numerical experiments (in particular, for the case $\lambda=\mu$), it looks like the lower bound on that term could be improved to some value $\approx 0.67$, and the upper bound (for the constant in front of the $\lambda+\mu$ term, i.e., large values) as well... any idea how achievable that could be? $\endgroup$
    – Clement C.
    Jun 30 '21 at 8:10
  • $\begingroup$ @ClementC. : Now we have the optimal constant, $\approx0.67$. $\endgroup$ Jul 1 '21 at 9:54
  • $\begingroup$ Wonderful! ${}{}{}{}{}$ $\endgroup$
    – Clement C.
    Jul 1 '21 at 12:18

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.