Proof Check: Galton Watson Process converges a.s. to $0$

295 Views Asked by At

Let $Z$ be the simple Galton-Watson Process $Z$ such that $Z_{n}=\sum\limits_{k = 1}^{Z_{n-1}}\xi_{n,i}$ where $(\xi_{n,i})_{n,i\in \mathbb N}$ are IID taking on values in $\mathbb N_{0}$ and $\mathbb E[\xi_{1,1}]=1$ and $P(\xi_{1,1} = 0) > 0$. Note also that $Z_{0}: =1$

Show that $Z_{n} \xrightarrow{n \to \infty} 0$ almost surely $(*)$.

My idea:

I have already checked that this is a martingale so I will not go into those details. By a version of the martingale convergence theorem, we obtain $Z_{n} \xrightarrow{n \to \infty} Z_{\infty}$ almost surely.

Now in order to prove $(*)$, I introduce a stopping time $\tau:=\inf\{n \in \mathbb N: Z_{n} = 0\}$

Clearly, on $\{\tau < \infty\}$ we obtain $Z_{\infty}=0$ almost surely since for $n$ large enough, i.e. $\tau(\omega)\leq n$, we obtain $Z_{m}(\omega)=0$ for all $m \geq n$.

But then consider the event $\{ \tau = \infty\}$,

If $\tau = \infty$, this implies that $Z_{n}\geq 1$ for all $n \in \mathbb N$.

Note that $ \{ Z_{n}\geq 1 \operatorname{ for all }n \}\subseteq \{ \exists i_{n} \leq Z_{n-1} \operatorname{ such that } \xi_{n,i_{n}} > 0 \operatorname{ for all }n \}\subseteq \{\xi_{n,i_{n}} > 0 \operatorname{ for all }n\}\subseteq \{\operatorname{for any n} \exists i :\xi_{n,i} > 0 \}$.

From the independence of $(\xi_{n,i})_{n,i\in \mathbb N}$, we obtain

$P(\operatorname{for any n} \exists i :\xi_{n,i} > 0 \operatorname{ for all }n)=\lim\limits_{n \to \infty}(1-P(\xi_{1,1}=0))^{n}=0$

Hence $P(\tau = \infty)=0$ and thus

$Z_{\infty} = Z_{\infty}\chi_{\tau < \infty} + Z_{\infty} \chi_{\tau = \infty}=Z_{\infty}\chi_{\tau < \infty}= 0$ almost surely.

The key part of my part were the chain of inclusions. Is my proof ok?

1

There are 1 best solutions below

2
On

This proof is not correct (the statement, though, is correct). The key issue is when you say "from the independence of...": you have chosen the index $i_n$ to depend on $\xi_{n,i_n}$ and so the law of $\xi_{n,i_n}$ has in fact changed. Something in the vein of what you're describing is possible but you'd have to instead union bound over all possible indices $i_n$ for each $n$.

The easiest proof of the fact you seek is through generating functions: if $\phi(s) = \sum_{k} P(\xi = k) s^k$, show that the probability $q$ of death is a fixed point of $\phi$. Use the fact that $1 = \phi'(1)$, $\phi(0) > 0$ and $\phi$ is convex to deduce that the only fixed point in $[0,1]$ is $1$.