Week 3: Little's Law
One equation, provable in a page, that lets you measure a queue you have no access to — and catches most people's intuition out.
8 March 2027
You are at the back of the airport security line.
Last week you could see the whole of a queue at once — six cars, one window, and the numbers came out of the model rather than out of you:
This one has no end you can see. No sign tells you the wait, and if one did you would be right not to believe it. Nobody is going to tell you anything.
You are about to work it out anyway, in about ninety seconds, using arithmetic you did when you were nine.
Look at what you can actually see
Two things are available to you from where you are standing, and neither requires anyone’s cooperation.
How many people are ahead of you. Not exactly — but you can count one switchback and multiply. Call it 240.
How fast the line is moving. Stand still, pick a fixed point, and count how many people cross it in two minutes. Say 80. That is 40 per minute.
That is the entire measurement. No access, no data request, no one to ask.
Before you scroll — how long is your wait?
Six minutes.
240 people ahead of you, leaving at 40 per minute, is 240 ÷ 40 = 6.
If that felt too easy, hold on to the feeling. The arithmetic really is that simple. What is not simple is knowing when you are allowed to do it, which is the rest of the lecture and most of your first assignment.
The law behind the arithmetic
L = λW
L is the average number in the system, λ the average arrival rate, W the average time each one spends there.Three quantities, and it is worth being pedantic about each, because almost every misuse of this law is a misuse of a definition rather than of the algebra.
L is a number of things, averaged over time. Not the number now — the number you would get by sampling the population of the system at random moments and taking the mean. People, cars, jobs, orders, dollars of inventory: the law does not care what they are, only that they are countable and they go in and come out.
λ — “lambda” — is a rate: things per unit time, measured as a long-run average. Crucially it is the throughput, the rate at which things actually flow through, which equals the arrival rate only when nothing is being turned away and nothing is piling up for ever.
W is a duration: the average time one thing spends inside, from the moment it crosses the boundary going in to the moment it crosses going out.
The units do the sanity-checking for you. People = (people per minute) × minutes. If your three numbers do not cancel like that, you have measured the wrong things.
What makes the law remarkable is not the algebra. It is the range of things it does not care about.
It holds for any arrival pattern, any service time distribution, any number of servers. And — this is the part people find hardest to believe — any discipline whatsoever. First come first served, shortest job first, random, priority, outright corrupt: Little’s Law does not notice. Everything Act II is about, the entire question of who goes first, is invisible to it.
Very few results in this field survive contact with a real system unmodified. This one does. It was in use as a rule of thumb for years before John Little published a general proof of it in 1961, and the proof turned out to require almost nothing: a boundary, a long enough run, and finite averages.
Why it is true
The proof is a page of accounting, and we will do it properly on the board. But the idea takes one picture, and once you have seen it you will not need the algebra.
Draw one horizontal bar per traveller: it starts when they join and ends when they clear the arch. The shaded region is person-minutes — the total amount of human time the hall has absorbed.
Now measure that region twice. Here is a hall that will let you do it both ways.
- L
- —
- λ
- —
- L ÷ λ
- —
- W
- —
- W90
- —
Count across it. Sweep from the back of the line to the front and tally. You are asking how many people are here right now, and the average of that answer is L. This is the cut into vertical strips, and it is the one you can do standing in a queue with no cooperation from anybody.
Follow one person through it. Start a clock when somebody joins, stop it when they clear the arch. You are asking how long did this one stay, and the average of that answer is W. This is the cut into horizontal strips.
Two different measurements, of two different things, by two different methods. Now read the bottom row of the scene: L ÷ λ and W agree to two decimal places, at every setting the slider reaches. They are not approximately equal. They are the same quantity — person-minutes — sliced two different ways.
Here is the accounting version, in four lines, so you can reconstruct it in an exam or an argument.
- Over a long window of length T, let N travellers pass all the way through.
- Add up how long each one stayed. Call the total A person-minutes. Then by definition the average stay is W = A / N.
- Now compute the same A the other way: at every instant, count the people present and integrate over the window. Averaged over T, that gives L = A / T.
- And the throughput is λ = N / T.
Divide the second by the third: L / λ = (A/T) ÷ (N/T) = A / N = W. The T cancels, the areas were always the same area, and the result falls out with no assumption anywhere about how anybody was ordered.
That is the whole theorem. It is not a fact about queues; it is a fact about rectangles, which is exactly why no assumption about arrivals or service or priority can break it.
Choosing the box is the whole game
Notice what step 1 quietly required: a boundary. Somewhere you can stand and say in and out. The law is silent about where you draw it, and that silence is where almost all of the interesting arguments live.
Draw three different boxes around the same airport and you get three different, equally correct answers.
Box A — the arch. In when you step onto the belt, out when you collect your tray. Two arches, thirty seconds each, so λ is about 4 a minute and L is about 2. W = 0.5 minutes. Security is extremely fast.
Box B — the hall. In when you join the tape, out when you collect your tray. λ is the same 4 a minute, L is the 240 you counted plus the 2 being screened. W = 60 minutes. Security is a disaster.
Box C — the journey. In when you leave the kerb, out when you sit down at the gate. Now λ is the same again, and L includes everyone walking the concourse, so W might be 85 minutes.
Nothing was measured wrongly. Three boxes, three numbers, all true, differing by a factor of 170 — and every one of them can be published as “the average wait” by somebody with a straight face.
The boundary is not a detail of the measurement. It is the claim.
The airport will quote you Box A. You are standing in Box B. Your flight cares about Box C.So the discipline is simple and almost never practised: state the box before you state the number. When you read somebody else’s waiting-time figure, your first question is not “is it accurate” but “what did they put inside the box, and what did they leave out”. Week 9 is an entire lecture on organisations that answered that question in their own favour and were not, technically, lying.
What it is good for
Four uses, in rising order of how much they matter.
Estimating a wait from outside. As above. Useful, slightly parlour-trick, and the fastest way to convince somebody that this field is worth an hour of their attention.
Sizing something you have not built. Rearranged, W = L/λ tells you how long people will wait given a capacity, before you commit to the capacity. Or, rearranged the other way, L = λW tells you how much floor space, how many chairs, how big a buffer. A call centre that will hold 40 calls at once and answers 3 a minute is committing to a 13-minute wait whether anyone has said so or not.
Catching a lie. This is the real one. If a service claims an average wait of 10 minutes, and you can see 500 people in the system and count 20 arrivals a minute, then the observable facts say 25 minutes. Somebody is measuring from a different starting line than you are — very often from the moment a ticket is called rather than the moment it is taken.
Nearly every published waiting-time statistic hides a definitional choice like that, and L = λW is the cheapest tool ever invented for finding it. We will use it as a lie detector properly in week 9.
Sanity-checking a promise. The strongest version: somebody tells you their plan will halve the wait without adding capacity and without turning anyone away. Then λ is unchanged, so L must halve. Ask them which half of the people currently in the system will not be there. If they cannot answer, the plan is arithmetic-free.
It is not really about queues
The law says nothing about lines, servers or people. It says: for any box with a boundary, in steady state, the amount of stuff inside equals how fast stuff flows through times how long stuff stays. Which means it applies to a startling range of things that nobody thinks of as queues.
A factory. Work-in-progress = throughput × cycle time. Everything on the floor, half-finished, is L. This version is old, it is the foundation of lean manufacturing, and it produces the counter-intuitive result that cutting the amount of unfinished work on the floor shortens how long each item takes, without anybody working faster.
A software team. Lead time = work in progress ÷ throughput. Same equation, same conclusion, and it is the entire mathematical content of a work-in-progress limit on a kanban board. A team with 40 open items completing 8 a week has a five-week lead time and cannot talk its way out of it.
A warehouse. Inventory turns are L and λ in a hat. Average stock ÷ annual sales = how long a unit sits on the shelf.
A support desk. 300 open tickets, closing 60 a day, means the average ticket lives five days. If the team promises 48-hour resolution while carrying 300 open tickets, the promise and the backlog are in direct contradiction, and no amount of effort reconciles them — one of the two numbers has to move.
A hospital. Beds occupied = admissions per day × average length of stay. This is why “reduce length of stay” and “increase bed numbers” are interchangeable at the level of arithmetic, and radically different at the level of cost.
Your support desk cuts average resolution time from 5 days to 2, and closes the same 60 tickets a day. What happened to the backlog?
It fell from 300 to 120, necessarily, and not as a side effect — as the same fact stated differently.
Which is why the two levers people argue about are one lever. A manager saying “we need to clear the backlog” and an engineer saying “we need to fix the slow cases” are, arithmetically, saying the same sentence to each other with different emphasis.
The useful follow-up: throughput did not change. So nobody worked harder, and the improvement came from the shape of the work — which is week 2 again, and you now have two weeks of reasons to look at variability before you look at capacity.
Three ways to get it wrong
Everything above assumed steady state: over your measurement window, what came in matched what went out. That assumption has three separate ways of failing, and all three will be waiting for you in the assignment.
The system was filling or emptying.
You watch a lunch rush from an empty room to a queue of forty. What does L = λW give you?
A confident, precise, wrong answer.
Nothing in the arithmetic warns you. You will get a number, it will have decimal places, and it will be meaningless — because the shaded region has not closed. People are still inside it when you stop counting, so the person-minutes you measured are not the person-minutes they will eventually spend.
The hall above is not lying to you about this, but only because it has been left running long enough to have forgotten how it started. Every figure in that readout is an average over a long run at one fixed setting. Move the slider and the hall takes a couple of minutes of its own time to mean it.
You measured the two numbers over different windows. L counted at 12:15 and λ averaged across 11:00–14:00 are facts about two different systems. The fix is dull and non-negotiable: one window, both quantities, stated on the page.
People left without being served. This one is subtle and it is everywhere. If a tenth of the queue gives up and walks out, they still occupied the system for a while, so they belong in L — but they never crossed the exit boundary, so whether they belong in λ depends on which box you drew. Count departures at the counter and you will under-count λ, and therefore over-estimate W, for the people who did stay. Count everyone who left by any means and you get a different, also-correct number that answers a different question. Week 11 is built on exactly this population, and the fact that they are the most informative people in the system and the least likely to be in anyone’s data.
This is not a footnote. It is most of the difficulty in your first assignment, due at the end of this week. Measuring a real queue turns out to be mostly the discipline of noticing when an assumption you were relying on quietly stopped being true — and the queue never tells you.
Where this goes
You can now measure any queue in the world from the outside.
Which raises the question the whole of Act I has been walking toward, and that this law is completely silent on:
that wait you just measured — what does it cost, and who is paying for it?
Teaching team
- Adaeze Nkemelu — convenor