We are often interested in the state an arriving customer observes in a queue. If arrivals are Poisson, then the average arrival sees the system in its equilibrium state.
This may seem counterintuitive. When a customer arrives, we might expect the system to be slightly emptier in anticipation of the arrival. However, this is not the case for Poisson processes: they are uniformly distributed over time, and arrivals after time are independent of what happened before
.
Theorem (PASTA). If is a positive recurrent continuous-time Markov chain with stationary distribution
, and
is a Poisson process (which may or may not induce transitions in
), then
Proof.
Let denote the transition times of
. Define transitions by pairs of states
. Then both
and
are Markov chains.
We have
Here is the proportion of time spent in state
, while
weights this by the rate of leaving
. The normalizing constant
is the long-run rate of transitions of the process.
We can now compute the stationary distribution of the transition process:
Thus, by ergodicity, the long-run proportion of transitions from to
is
Let be a set of “arrival transitions”. Then
Suppose transitions in occur as a Poisson process. In particular, assume that for each state
there is a unique transition
with rate
. Then
Thus, the average arrival sees state with probability
.