Welcome to Randal Douc's wiki

You've loaded an old revision of the document! If you save it, you will create a new version with this data.

====== Statement ======

<WRAP center round box 80%>
Let $(U_i)$ be iid  Rademacher random variables, i.e. $U_i=1$ or $-1$ with probability $1/2$ and set $S_i=\sum_{j=1}^iU_j$ the associated partial sum. Define $\Delta=\inf\set{t>0}{S_t=0}$. Show that $S_n$ returns to 0 with probability one. What is the law of $\Delta$? 
</WRAP>

====== Proof ======
Obviously, $\PP(\Delta=2k+1)=0$. It will be useful in some parts of the proof to note that, by symmetry, $\PP(\Delta=2k, S_1>0)=\PP(\Delta=2k, S_1<0)$, so that $\PP(\Delta=2k)=2\PP(\Delta=2k, S_1>0)$
Define 
\begin{align*}
&\alpha_k=\PP(\Delta=2k, S_1>0)\\
&\beta_k=\PP(S_{2k}=0, S_i \geq 0, \forall i \in [1:2k-1])
\end{align*}
Note first that $\alpha_1=1/4$. Moreover, for all $k>1$,
\begin{align*}
\alpha_k&=\PP(\Delta=2k, S_1>0)\\
&=\PP(U_1=1,U_{2k}=-1, U_2+\ldots+U_i\geq 0\ \forall i \in [2:2k-2], U_2+\ldots+U_{2k-1}=0)\\
&= \beta_{k-1}/4
\end{align*}
And since $\sum_{k=1}^{\infty} \alpha_k=\PP(\Delta \in 2\nset^*,S_1>0)<\infty$, we deduce that $\sum_{k=1}^{\infty}\beta_k<\infty$. This allows to define $A(z)=\sum_{k=1}^{\infty} \alpha_k z^k$ and $B(z)=\sum_{k=1}^{\infty} \beta_k z^k$ for all $|z| \leq 1$ and the previous identities implies
$$
A(z)=\frac{z}{4}+\frac{z B(z)}{4}
$$
Moreover, by the first entrance decomposition,
\begin{align*}
\beta_k&=\PP(S_{2k}=0, S_i \geq 0, \forall i \in [1:2k-1])\\
&=\sum_{\ell=1}^{k} \PP(\Delta=2\ell, S_{2k}-S_{2\ell}=0, S_i-\underbrace{S_{2\ell}}_{=0}\geq 0,\ \forall i \in [2\ell+1:2k-1])\\
&=\alpha_k+\sum_{\ell=1}^{k-1} \alpha_\ell \beta_{k-\ell}
\end{align*}
Multiplying by $z^k$ and summing over $k \in \nset^*$ yields $B(z)=A(z)+A(z)B(z)$. Finally,
$$
\frac{A(z)-z/4}{z/4}=A(z)\lr{1+\frac{A(z)-z/4}{z/4}}
$$
This is equivalent to $0=A^2(z)-A(z)+\frac{z}{4}$ so that $A(z)=\frac{1\pm \sqrt{1-z}}{2}$. Only one of the roots has an expansion with only non-negative coefficients, that is, $A(z)=\frac{1-\sqrt{1-z}}{2}$. This implies that 
$$
\PP(\Delta<\infty)=\sum_{k=1}^\infty \underbrace{\PP(\Delta=2k)}_{2\PP(\Delta=2k,S_1>0)}= 2 A(1)=1
$$
that is, with probability 1, this random walk returns to 0.

Moroever, expanding $\frac{1-\sqrt{1-z}}{2}$ yields: $A(z)=\sum_{k=1}^{\infty} \begin{pmatrix}2k-2\\k-1\end{pmatrix}
\frac{1}{k 4^k} z^k$. Therefore, by symmetry, $\PP(\Delta=2k)=2\PP(\Delta=2k,S_1>0)=2\begin{pmatrix}2k-2\\k-1\end{pmatrix}\frac{1}{k 4^k}$.

====== Comments ======

Maybe there exists a more simple way to find the expression of $\alpha_k$... A particular trajectory has a probability $(1/2)^{2k}=1/4^k$ to occur. It remains to count the number of trajectories. Since $U_1=1$ and $U_{2k}=-1$, it remains to count the number of trajectories of size $2k-2$ such that $S_i\geq 0$ for all intermediate time $i$ and $S_{2k-2}=0$. Without the constraint on the positivity of $S_i$ at each intermediate times, it would have been as if we take $k-1$ among $2k-2$ terms but in that case we count all the trajectories (even those that cross the $X$-axis), that is $\begin{pmatrix}2k-2\\ k-1\end{pmatrix}$. Comparing with the expression we finally obtained, we have find a justification for dividing this number by $k$. An idea would be to split them into well chosen equivalent classes of size $k$ with only one representant in each class with positive trajectories but I can't find this equivalence relation...

Please fill all the letters into the box to prove you're human. T I N Y᠎ C Please keep this field empty:

Edit summary

Welcome to Randal Douc's wiki

User Tools

Site Tools

Page Tools