A service desk receives eight jobs per hour. If each job waits ten minutes and then takes five minutes to serve, Little’s Law gives about 1.33 jobs waiting but 2.00 jobs in the whole system. Both answers are correct. The difference is the boundary you measured.
The relationship L = λW is an accounting law about long-run averages, not a magic formula that predicts any waiting time. We will calculate waiting-room and whole-system counts, show how mismatched quantities give plausible-looking nonsense, and build a reusable measurement checklist.
Find the question you need
Choose a reading route or read all twelve chapters. Every teaching chapter remains open; the grouped index below is expandable.
Define the quantities: Chapters 1–2. Start this route.
Calculate a consistent system: Chapters 3–5. Start this route.
Correct the common mistake: Chapters 6–7. Start this route.
Audit the observation: Chapters 8–10. Start this route.
Practise and continue: Chapters 11–12. Start this route.
Full chapter index · Worked practice · Queueing Theory subject root
Chapter index
Define the quantities · Chapters 1–2
Calculate a consistent system · Chapters 3–5
Correct the common mistake · Chapters 6–7
Audit the observation · Chapters 8–10
Practise and continue · Chapters 11–12
In an appropriately stable system, Little’s Law states L = λW. L is the average number of jobs within a defined region, λ is the effective long-run flow of jobs through that region and W is the average time each job spends inside it. Eight jobs per hour multiplied by a quarter-hour per job gives two jobs.
The essential phrase is within a defined region. A waiting room, a service counter and the entire customer journey can have different counts and times. Identify what enters and exits before writing any number. Little’s Law does not by itself require exponential service times, Poisson arrivals or first-in-first-out service.
Define Region Q as the waiting room only. A job enters upon joining the queue and exits when service starts. Its average occupancy is Lq and mean time is Wq. Define Region S as the waiting room plus the service position. Its corresponding quantities are L and W.
If every admitted job is served exactly once and the flows balance in the long run, both regions share throughput λ, giving Lq = λWq and L = λW. Write W = Wq + E[S], where E[S] is mean service duration. These are distinct but compatible identities.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Assume a stationary illustrative service with λ = 8 completed jobs/hour, Wq = 10 minutes and E[S] = 5 minutes. Convert time carefully: waiting is 1/6 hour and total time is 1/4 hour.
Then Lq = 8 × (1/6) = 1.333 jobs, while L = 8 × (1/4) = 2.000 jobs. Average counts need not be whole numbers: instantaneous occupancy is integer, but its time average can be fractional.
An average of 2.000 jobs in the whole system does not mean two people are always there. Sometimes nobody is present, sometimes more than two. The equation does not specify their distribution.
| Count | Time measured | Average jobs |
|---|---|---|
| Waiting room Lq | 10 minutes | 8 × 10/60 = 1.333 |
| Entire system L | 15 minutes | 8 × 15/60 = 2.000 |
| In service | 5 minutes | 8 × 5/60 = 0.667 |
Contents · Previous chapter · Next chapter · Queueing Theory subject root
The difference L − Lq = 2.000 − 1.333 = 0.667 jobs describes mean occupancy in service. Independently, λE[S] = 8 × (5/60) = 2/3 jobs. The two calculations match because the regions partition the same residence time.
If there is precisely one continuously available work-conserving server, this 2/3 is also its long-run busy fraction. It is important that the arrival rate of eight jobs per hour is below the ideal service capacity of twelve per hour; setting arrival rate equal to capacity would undermine the intended stable single-server example.
In a multi-server installation, mean jobs in service can exceed one. Then λE[S] is not the utilisation of one individual server, and another definition is needed for capacity planning.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Imagine a graph N(t) showing how many jobs are inside the chosen region. The integral of N(t) over time measures the total job-hours accumulated in the region. The same area is obtained by summing each job’s time spent inside, with appropriate treatment of jobs present at the start or end.
Dividing job-hours by hours yields average occupancy; dividing total completed job-hours by the number of appropriately tracked jobs yields their mean residence time. Long-window flow accounting connects these quantities by L = λW.
For example, 24 completed visits at an average of 15 minutes give 6 job-hours. Over a suitable three-hour fully reconciled observation, that corresponds to two jobs on average. This is a constructed illustration, not a record from a real centre.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Suppose someone measures Lq = 1.333 and divides by λ = 8 jobs/hour. The answer is 0.1667 hour, ten minutes. That is a valid waiting time. If the report instead calls it the customer’s total time, it has silently omitted five minutes of service.
The reverse error divides the total occupancy L = 2.000 by ten minutes of queue-only waiting, obtaining an apparent rate of twelve jobs/hour. No rate increase occurred. The denominator changed its definition, so the equation no longer compared like with like.
Correct the labels before making operational decisions. Units can all work out while the underlying counted populations and boundaries do not.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Little’s Law tells us how one average relates to two others. It does not determine the 90th-percentile wait, the chance of waiting longer than an hour, the shape of the arrival process or the degree to which unusually long service jobs cause a queue.
Two services may share the same average arrival rate and average waiting time yet have very different individual waiting experiences. You need a fitted queueing model, distributional information or measured observations for those questions.
The related equal-mean service-time example shows why the second moment of service duration matters in an M/G/1 model, even though Little’s Law still relates each model’s compatible averages.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
If people leave a line before being served, the flow of admitted arrivals need not equal the rate of completed service over the same window. Measuring occupancy for everyone who waited, but throughput and durations only for successfully served people, can mix two different populations.
Choose what counts as an item: person, visit, ticket, case or transaction. One customer who opens three separate tickets creates three items for a ticket-based flow. Repeated entries must be counted consistently in λ, L and W.
A useful audit records arrivals, successful completions, abandonments, transfers and unfinished cases separately. The law can be applied to suitably defined flows, but inaccurate bookkeeping cannot be repaired by an elegant equation.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Suppose the queue begins the morning with 40 unfinished cases and ends with 70. New arrivals and completions have not balanced. Dividing this morning’s arrivals by elapsed hours and mixing that rate with the completed jobs’ residence times may fail to reproduce observed average occupancy.
A finite-window occupancy-time relation can be kept with explicit boundary terms, including jobs already present and still unfinished. Calling arbitrary short-window values steady-state averages is the unreliable step.
For practical reporting, state the measurement dates, starting backlog, ending backlog, treatment of incomplete items and what population supplied the mean duration. Stabilisation should be supported by data rather than assumed.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Suppose urgent cases enter at four per hour and spend an average of 0.10 hour inside; routine cases enter at six per hour and spend 0.20 hour. Their occupancies are 0.4 and 1.2 jobs respectively, so combined average occupancy is 1.6 jobs.
The total throughput is ten jobs per hour. Its throughput-weighted mean residence time is (4 × 0.10 + 6 × 0.20)/10 = 0.16 hour. Then L = 10 × 0.16 = 1.6. A naive unweighted mean of class times would be 0.15 hour and would not match.
This is why dashboards for hospitals, warehouses or help desks must name whether averages are job-weighted, time-weighted, class-specific or overall. A number without its averaging basis is easy to misuse.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Practice A: A stable processing region completes ten jobs per hour, with nine minutes average time inside. Its average occupancy is 10 × (9/60) = 1.5 jobs.
Practice B: In that same process, 4.2 minutes are spent waiting and 4.8 minutes in service. Mean waiting occupancy is 10 × (4.2/60) = 0.7 jobs. Mean service occupancy is 10 × (4.8/60) = 0.8 jobs. These add to 1.5.
Practice C: A supervisor reports 1.5 jobs waiting but a nine-minute total residence time. Can you calculate arrival rate from those alone? No. You need the mean queue-only time or the whole-system occupancy. State which missing quantity would fix the problem.
Contents · Previous chapter · Next chapter · Queueing Theory subject root
Before publishing any queue metric, write six items next to it: the entry event, exit event, population counted, flow rate denominator, observation window and treatment of unfinished work. Then choose either L = λW or Lq = λWq accordingly.
This is what makes the formula useful. If a calculation fails, it may reveal a definition problem, missing case, inconsistent duration or unstable window before it becomes a poor staffing decision.
Continue through The Gold Standard of Queueing Theory and the sibling Batch Arrivals, Event Rates and Job Counts, where the meaning of the word ‘arrival’ itself changes the outcome.
Sources and connected reading
Numerical scenarios are original teaching examples. Check the stated boundaries and long-run assumptions before applying these relationships to a real service.