A birth-death process is a special form of a [[Markov Chain]] where transitions are restricted to moving up, moving down, or staying in the current state. For example, it can model a supermarket queue where the number of people increases by one (birth), decreases by one (death), or remains the same.
![[birth-death-process.png|center|400]]
Since a birth-death process is [[Convergence of Markov Chains#Recurrent & Transient States|recurrent]] and [[Convergence of Markov Chains#Periodicity|aperiodic]], it has well-defined steady-state probabilities when $n$ (the number of steps) becomes large.
![[birth-death-process-split.png|center|200]]
## Steady-State Probabilities
We can draw a random boundary in the middle of the Markov chain, and analyze the frequencies of crossing that boundary. If we cross e.g. 100 times from left to right, I have to cross 99, 100 or 101 times from right to left. So if $n$ is large we can say that:
$
\begin{align}
\pi_{i+1}*q_{i+1} &= \pi_i*p_i \\[4pt]
\pi_{i+1} &= \pi_i* \frac{p_i}{q_{i+1}}
\end{align}
$
where:
- $p_i$: Transition rate from state $i$ to state $(i+1)$, i.e. "birth".
- $q_i$: Transition rate from state $i$ to state $(i-1)$, i.e. "death".
Using this relation we can express any $\pi_j$ in terms of $\pi_0$ and the $\{p, q\}$ terms.
$ \pi_j=\pi_0* \prod_{i=0}^{j-1} \frac{p_i}{q_{i+1}}$
Since we know that all steady-state probabilities need to sum to $1$, we have an additional constraint.
$ \sum_j \pi_j=1$
Substituting $\pi_j$:
$
\begin{align}
\pi_0 +
\overbrace{\pi_0*\frac{p_0}{q_1}}^{\pi_1}+
\overbrace{\pi_0*\frac{p_0*p_1}{q_1*q_2}}^{\pi_2} + \dots &= 1 \\
\pi_0*\left(1+\frac{p_0}{q_1}+\frac{p_0*p_1}{q_1*q_2}+ \dots\right)&=1
\end{align}
$
## Special Cases of Birth-Death Processes
**Constant Rates:**
Let the "birth-rate" $p$ be constant across all $i$. Also the "death-rate" $q$ is constant across all $i$. We define $\rho$ as the ratio of these two constant rates.
$\rho=\frac{p}{q}$
The normalization condition simplifies to a [[Convergence of Series#Geometric Series|Geometric Series]].
$
\begin{align}
\pi_0*\left(1+\frac{p}{q}+\frac{p*p}{q*q}+ \dots\right)&=1 \\[4pt]
\pi_0*\Big(1+\rho+\rho^2+ \dots\Big)&=1
\end{align}
$
**Symmetric Random Walk:**
When $p=q$ we call it a symmetric random walk, where $\rho=1$ and all $\pi_i$ are equally likely.
$
\begin{align}
\pi_0*(1+\rho+\rho^2+ \dots+\rho^m)&=1\\
\pi_0*(1+1+1^2+ \dots+1^m)&=1
\end{align}
$
Since $1^n=1$, we are basically summing up ones for $(m+1)$ times.
$\pi_0=\frac{1}{1+m}$
By the recursive formula we see that all $\pi_i$ are equal in that scenario:
$ \pi_{i+1} = \pi_i* \frac{p}{q}, \quad \text{where}\: \frac{p}{q}=1$
**Geometric Series:**
When $p < q$ then it follows that $\rho < \vert 1 \vert$. For $m \to \infty$, we have a infinite geometric series that converges to a finite sum.
$
\begin{align}
\pi_0*(\overbrace{1+\rho+\rho^2+ \dots+\rho^m}^{\text{geometric}})&=1\\[6pt]
\pi_0* \sum_{i=1}^\infty \rho^i&=1 \\[4pt]
\pi_0* \frac{1}{1-\rho}&=1\\[6pt]
\pi_0&=1-\rho
\end{align}
$