# Expected time spent in a node between visits to another node in an asymmetric random walk on integers

Recall that the first passage time to state $$i$$ is the random variable $$T_i$$ defined by $$\begin{equation*} T_i(\omega)=\mathrm{inf}\{n\geq1:X_n(\omega)=i\} \end{equation*}$$ where inf$$\emptyset=\infty$$. Let $$(X_n)_{n\geq0}$$ be a simple random walk on $$\mathbb{Z}$$ with $$p_{i,i-1}=q. Find $$\begin{equation*} \gamma^0_1=\mathbb{E}_0(\sum_{n=0}^{T_0-1}1_{X_n=1}). \end{equation*}$$

My efforts:

$$\begin{equation*} \gamma^0_1=\sum^\infty_{n=0}\mathrm{Pr}(X_n=1,n+1\leq T_0). \end{equation*}$$ $$\begin{equation*} \mathrm{Pr}(X_{2k}=1,2k+1\leq T_0)=0. \end{equation*}$$ $$\begin{equation*} \mathrm{Pr}(X_{2k+1}=1,2k+2\leq T_0)=c_kp^{k+1}q^k. \end{equation*}$$

We only need to determine $$c_k$$, which is a complicated counting problem. I know $$c_0=1,c_1=1,c_2=2,c_3=4$$. We only need to consider the walk to the right hand side of 0. In the case of $$X_{2k+1}=1,2k+2\leq T_0$$, the farthest place that the walker can reach is $$k+1$$. Of course the walker does not necessarily reach $$k+1$$. It can just repeat between some nodes, e.g., $$k-1$$ and $$k$$.

This is a transient random walk. So we cannot use the stationary distribution to calculate this.

The generating function method of calculating number of steps returning to 0 starting from 1 doesn't help either. Number of steps taken to return to 0 and number of visits to 1 are two very different things, although you must pass 1 so as to return to 0.

There seems to be no way to avoid counting. So my problem is: what is a smart way to count $$c_k$$?

$$\begin{equation*} \gamma^0_0=1. \end{equation*}$$ $$\begin{equation*} \gamma^0_i=q\gamma^0_{i+1}+p\gamma^0_{i-1},i\neq0. \end{equation*}$$ Thus $$\begin{equation*} \gamma^0_i=A+(1-A)(\frac{p}{q})^i. \end{equation*}$$ Consider the case of $$i>0$$. To visit $$i+1$$ and return to 0, the walker must visit $$i$$ at least twice. $$\gamma^0_i$$ is a non-increasing function of $$i$$. But $$(p/q)^i$$ is an increasing function. The best we can do is setting $$A=1$$ so that $$\gamma^0_i=1$$ for all $$i>0$$.
Consider the case to the left of 0. Still let $$i>0$$. We have $$\begin{equation*} \gamma^0_{-i}=A+(1-A)(\frac{q}{p})^i. \end{equation*}$$ lim$$_{i\rightarrow\infty}\gamma^0_{-i}=0$$, so $$A=0$$ and $$\gamma^0_{-i}=(q/p)^i$$ for all $$i>0$$.