Why the time to empty is a different question
A server starts work with one customer present. New customers may arrive before the server finishes; the people they bring into the system may in turn be followed by more arrivals. How long will the server remain continuously busy? An average arrival rate alone does not answer this question: the service clock and the distributional assumptions matter as well.
We connect the Poisson and birth–death models already studied to this first-passage question.
Model and the first return to zero
Definition 3.7.1
(One server with exponential service)
Let arrivals form a Poisson process of rate , and let successive service requirements be independent exponential random variables of rate , independent of arrivals. There is one server, unlimited waiting room and first-come, first-served service. The server works whenever anyone is present. Write for the number in the system, including any customer in service.
When , the busy period is , with if zero is never reached. Define the load .
Exponential memorylessness makes the number in service and waiting sufficient to determine the distribution of the next event. From state , arrivals occur at rate and departures at rate ; at zero the only transition is an arrival. These are bounded total rates, so the jump construction is non-explosive. With a general service distribution the elapsed service time can matter, and alone need not be Markov.
Theorem 3.7.2
(Busy-period transform and eventual emptying)
For , define . Then
Moreover,
The expectation is extended: a positive probability of makes it infinite.
Proof
(First-step equation and root selection)
Let be the first event time. Competing independent exponential clocks give ; the event is a departure with probability and an arrival with probability , independently of . After a departure the busy period is over. After an arrival the count is two. To reach zero from two it must first reach one, then zero. By spatial homogeneity above zero and the strong Markov property at that first passage, the time from two to zero has the same extended distribution as a sum of two independent copies of . For this distributional construction, an independent second copy can be sampled even when the first is infinite; it does not describe a passage that actually begins on that path. Hence, using ,
The quadratic has two positive roots. Its larger root exceeds one: the quadratic expression is positive at zero and negative at one because . Since , only the displayed smaller root is possible.
As , monotone convergence gives , yielding the stated minimum. If , differentiation of the explicit root shows ; monotone convergence of then identifies the finite mean. If , the right derivative is infinite, so the same monotone argument gives infinite mean even though emptying is certain. If , the probability of never emptying is positive, which already makes the extended mean infinite.
Proof X-ray
Why solving the quadratic is not enough
The proof has two stages, and the second is where the care lies.
- The first-step equation uses competing exponential clocks for the first event, and spatial homogeneity above zero with the strong Markov property to write the passage from two to zero as two independent copies of . The convention lets the equation hold even when has positive probability.
- At the quadratic has roots and ; both lie in when arrivals are faster, so an algebraic root need not be the probability of the actual event. The key move is to work at , where and for , so exactly one root lies in .
- Monotone convergence as then identifies and, through , the mean.
The argument closes by keeping three cases apart: an infinite busy period, a finite one with infinite mean, and a finite one with finite mean.
Starting with customers, the process must cross in sequence. The same passage argument gives transform and probability of eventual emptying. When , the mean is . These conclusions concern an initially finite .
A numerical busy period
Worked example 3.7.3
(One arrival per hour, two services per hour)
Take and , in inverse hours. Then , the busy period ends almost surely, and its mean is one hour. Its transform is
Differentiating the quadratic twice gives hours squared and hours squared. The mean is twice the mean service time : arrivals made during service extend the period. With three customers initially, the mean time to empty is three hours, not one hour.
The second moment in that example comes from a general formula, which shows how the mean alone can hide congestion.
Worked example 3.7.4
(Equal mean busy periods, very different variability)
Differentiate in : . When , put and to get . Differentiating once more and setting ,
For this gives , as above. For the mean busy period is again hour, but and . The expected number served in a busy period rises from to .
At critical load the same transform still determines the behavior, but the derivative at zero is infinite.
Worked example 3.7.5
(Critical load: certain emptying with infinite mean)
Take . Then and
so the busy period ends almost surely. For the mean, increases to as , and
By monotone convergence . Since , the stationary theorem below also gives no stationary probability.
A service has mean , and the expected number of arrivals during it is . Regard those arrivals as its children, the arrivals during their services as the next generation, and so on. Their expected generation sizes are : condition on the finite previous generation and use independent service lengths and Poisson increments. If , monotone convergence gives expected total customers served . In the example this is two. This is an expectation of a random family size; it does not assert that every busy period has exactly two services.
Concept lensStructural
The number served is a family size
Count customers instead of time. During one exponential service of rate , the number of arrivals is with probability , independently across services. The number served in a busy period is therefore the total family size of a branching process with this geometric offspring law, and satisfies , that is,
a quadratic of the same form as the busy-period equation. The mean offspring number is , which is why governs both the family size and the busy period.
Equilibrium is a separate question
Theorem 3.7.6
(Stationary number in the system)
The queue-length chain has a stationary probability distribution precisely when . In that case it is unique and
Proof
(Balance equations and their summability)
For the displayed probabilities sum to one and satisfy for each . Thus incoming and outgoing flow balance at every state, including zero. The bounded-rate generator and its uniformization show that this probability is invariant for the continuous-time semigroup.
Conversely, any stationary probability satisfies the balance equation at zero, . Its equation at state one then gives , and induction gives . If , all entries vanish; otherwise this sequence is summable only for . Normalization fixes , proving both necessity and uniqueness. Differentiating the geometric series yields the mean.
Common mistake
(Confusing the busy fraction with the chance of emptying)
The empty fraction under equilibrium is ; the busy fraction is . These are long-run state probabilities under the stationary law, not the probability that a busy period started now ever ends. In the numerical example the stationary mean number is one, whereas a newly started busy period has mean duration one hour. Their numerical equality is accidental and their units differ.
Common mistake
(Using the subcritical mean formula at heavy load)
The formula was derived only for . For it is negative and cannot be a mean, and for it is undefined; in both cases the extended mean is infinite.
The two regimes at and above critical load show what goes wrong if eventual emptying is taken as the definition of stability.
Counterexample mode
Certain return need not have a finite mean
At , the busy period ends with probability one, but its mean is infinite and no stationary probability exists. Calling the queue “stable” merely because it eventually empties would conceal both facts. At , it empties with probability only when started from one; the other half of paths never reach zero. The same first-step equation gives both regimes only after the correct solution is selected.
Practice with solutions
Quick check
(Suppose and per hour. Find the emptying probability and the mean busy period starting from two customers.)
Starting from two, the queue must make two downward passages.
Solution · Answer
Since , emptying is certain. The two consecutive downward passages have mean each, giving one hour in total. Multiplying both rates by a common positive constant leaves the emptying probability unchanged but divides each passage time by that constant.
Quick check
(Let and . Find the probability that the queue ever empties from three customers, and determine the unconditional mean busy period from one.)
Each downward passage succeeds with probability .
Solution · Answer
A downward passage succeeds with probability , so three successive passages give . From one, nonemptying has probability , so the extended mean is infinite. The number is negative and cannot be a mean; the hypothesis of the subcritical formula fails.
Exercise 3.7.7
(For , find .)
Core
Solution(3)
- 1
Condition the transform on emptying
Here , and .
- 2
Differentiate the root at zero
Key step
, so .
Because: At , when .
- 3
Divide by the emptying probability
. It is finite although the unconditional extended mean is infinite; the event of never emptying still cannot be ignored in an unconditional calculation.
Quick check
(In the example, find the stationary probability of at least two customers and the expected number served in a busy period started by one arrival.)
Use the geometric stationary law and the generation count.
Solution · Answer
The geometric stationary tail is . The expected total served during the busy period is , by generation counting. One is a probability at a stationary observation time; the other counts customers over an entire random interval.
Scope of this unit
The analysis uses independent Poisson arrivals, independent exponential services and a single continuously working server. General-service queues, waiting-time distributions, finite-capacity systems and arrival-time sampling require further arguments.