Uniqueness of Compound Poisson Distribution

characteristic-functionspoisson distributionprobabilityprobability distributionsstochastic-processes

Let $X$ be a random variable such that $X =^d \sum_{i=1}^N Y_i$ where $N$ is Poisson and $(Y_i)$ is a sequence of IID random variables independent of $N$. Then $X$ is said to have a compound Poisson distribution. Now if $X_1 = \sum_{i=1}^N Y_i$ and $X_2 = \sum_{i=1}^M Z_i$ are compound Poisson and $X_1 =^d X_2$, where $N \sim \mathrm{Pois}(\lambda_N)$, $M \sim \mathrm{Pois}(\lambda_M)$ with $P(Y_i=0) = P(Z_i=0)=0$, how can I show that $\lambda_N = \lambda_M$ and $Y_i =^d Z_i$?

The characteristic function of $X_1$ and $X_2$ are $$\phi_1(t) = \exp[-\lambda_N(1-\phi_Y(t))], \quad \phi_2(t) = \exp[-\lambda_M(1-\phi_Z(t))]$$ respectively. We know that $\phi_1 = \phi_2$. How can I show that $\lambda_N = \lambda_M$ and $\phi_Y = \phi_Z$?

Best Answer

You just need to note that since $-\ln \phi_1(t)=-\ln\phi_2(t)$ you obtain

$$\lambda_N(1-\phi_Y(t))=\lambda_M(1-\phi_Z(t))$$

Inverse-Fourier transforming this equality gives a relation between the pdfs of $Y,Z$, $f_Y(x), f_Z(x)$:

$$(\lambda_N-\lambda_M)\delta(x)=\lambda_Nf_Y(x)-\lambda_M f_Z(x)$$

Integrate this to obtain a relationship between the cdf's

$$(\lambda_N-\lambda_M)\theta(x)=\lambda_NF_Y(x)-\lambda_MF_Z(x)$$

Since there is no point mass at the origin we have that $F_{Y,Z}(0^-)=F_{Y,Z}(0^+)$, so taking left and right limits around $x=0$ forces $\lambda_N=\lambda_M$, as it is easy to see. The equality of probability distributions follows trivially thereafter.

EDIT: In case the pdf of $Y,Z$ are not defined it is still possible to prove uniqueness. First, note that any random variable has a CDF, by definition. From well known theorems, it is seen that

  1. Any single variable function that obeys the CDF axioms is right continuous everywhere on the real line (ref: Kun Il Park, Fundamentals of Probability and Stochastic Processes with Applications to Communications, p.79)

  2. Any single variable function that is right continuous everywhere can be shown to be discontinuous only on a countable subset $B$ of the reals, see here

In light of these results, the following inversion formula gives the unique CDF assigned to a characteristic function

$$2F_X(x)-1=\frac{1}{\pi}\int_0^\infty\frac{e^{ixt}\phi_X(-t)-e^{-ixt}\phi_X(t)}{it}dt$$

and the integral on the RHS converges on $x\in B^c$.

Assuming $\phi_1(t)=\phi_2(t)$ yields the family of solutions

\begin{equation}\lambda_N-\lambda_M+2n\pi i=\lambda_N \phi_Y(t)-\lambda_M\phi_Z(t),~~ n\in \mathbb{Z}\tag{1}\end{equation}

Forming the linear combination $(it)^{-1}(e^{itx}(1)(-t)-e^{-itx}(1)(t))$ and integrating yields after some rearrangement the relation

$$\lambda_N F_Y(x)-\lambda_M F_Z(x)=(\lambda_N-\lambda_M)\theta(x)+n\pi i ~\text{sgn}(x)~~, x\in B^c$$

Obviously, since the LHS is purely real, this forces $n=0$. From this point on, one can use the argument presented pre-edit to conclude that $\lambda_N=\lambda_M~,~ F_Y(x)=F_Z(x) ~, ~x\in B^c$, and the right continuity to conclude that the jumps are also fixed to be the same, which implies that $F_Y(x)=F_Z(x) ~~,~ x\in \mathbb{R}$.

Related Question