The last discrete family in the chapter and the only one that does not count successes among trials. A Poisson variable counts occurrences in a fixed interval of time or space, when those events happen with a known average rate and independently of the time since the last one. It is written X follows P of mu, where mu is the mean number of occurrences per interval, and mu scales in proportion when the interval is lengthened or shortened. Its single parameter is both its mean and its variance, so the standard deviation is the square root of mu, and the family doubles as an approximation to the binomial when the number of trials is large and the success probability small.
Subject: Statistics · 69 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · Chapter 4 — Discrete Random Variables
Counting events in an interval of time or space, from one number: the average rate
Objectives
The binomial, geometric and hypergeometric distributions all count something about a set of trials: how many successes, or how long until the first. This section meets a count with no trials in it at all. Calls arrive, loaves go on a shelf, misspellings turn up on a page, and the question is simply how many land in a stretch of time or space.
The whole section rests on one number, μ, the average count in the interval you care about. Getting μ right is most of the work: the formula is short, but it faithfully turns a wrong μ into a wrong answer that looks perfectly reasonable. Watch for that as you go.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 250-254 — the section this deck covers
Warm-up
Section 4.3 gave you the binomial: n independent trials, each a success with probability p. Recall its mean and its formula now, on a case with many trials and a small chance of success on each.
Discussion prompt
A machine fills 20 bottles, and each one independently has a 0.05 chance of being underfilled. Find the expected number of underfilled bottles, and the probability that none of the 20 is underfilled.
Keep two numbers in mind: a mean of 1 and a chance of 0.3585 for zero. In Part 5 you will meet the Poisson with mean 1 and find that it gives 0.3679 for zero, close to the binomial answer. That closeness is not a coincidence; it is the second characteristic of this section.
Notice also what you needed: a definite number of trials, 20. The first half of this section is about counts where no such number exists, and it is worth asking yourself as you read each example whether you could say what n would be.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.3, p. 237 — the binomial mean np
Section
Part 1
Concept
Leah receives about six telephone calls between 8 and 10 in the morning. On any particular morning the number of calls is a count: it could be 4, or 8, or on a quiet day none at all. Here are three simulated mornings with that average rate.
Figure (svg): Three horizontal timelines from 8 a.m. to 10 a.m., one per morning, with a dot for each phone call. The mornings have 6, 8, 6 calls. A yellow band marks 9:00 to 9:15 and says how many calls fell inside it on each morning.
The three mornings bring 6, 8, 6 calls. None of them has to bring exactly six. What stays fixed is the average rate, six per two hours; what varies is the count in any one interval.
Poisson random variable — X = the number of occurrences of an event in a fixed interval of time or space, when the events happen at a known average rate and independently of the time since the last event.
Look at the yellow window, 9:00 to 9:15. The count inside it also varies, and it is usually 0 or 1. A shorter interval has a smaller average count, and Part 2 is about exactly that. For now the point is the kind of question: not how many successes in so many trials, but how many events in this stretch.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 250 — the first characteristic, and X = the number of occurrences in the interval of interest
Concept
The book's first characteristic is a single long sentence. Unpack it and it makes three separate demands, and a count is a Poisson only if all three hold.
The third requirement is the subtle one. It says the process has no memory. After an hour with no calls, Leah is not owed a call; after two calls in a minute, she is not due a rest. The chance of a call in the next minute is the same either way.
\[ \text{rate per minute} = \frac{6 \text{ calls}}{120 \text{ minutes}} = 0.05 \text{ in every minute} \]
That constant rate, the same in every minute of the morning, is what lets you rescale the average onto any interval you like. If calls were far more likely at 8:05 than at 9:40, no single rate would describe the morning and the whole method would fail.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 250 — the first characteristic
Picture it
Independence from the time since the last event is easiest to see by comparing processes that have it with processes that do not. Each timeline below has about ten events in an hour.
Figure (svg): Three one-hour timelines. The top one has 15 dots scattered irregularly, some close together and some far apart. The middle one has ten dots exactly six minutes apart. The bottom one has ten dots in three tight bunches.
The top line is a Poisson process. Its gaps are irregular: some events almost touch, and there are long quiet stretches. That is not a sign of anything going wrong; it is exactly what memoryless events look like. Genuine randomness is lumpier than people expect.
The middle line is a timetable. Each event tells you precisely when the next one comes, which is the opposite of independence. The bottom line is clustered, like aftershocks after an earthquake or customers arriving by the busload, where one event makes more events likely soon after. Neither is Poisson, even though both have the same average rate as the top line.
So a rate on its own is not enough. Before using this section's formula, ask whether an event makes the next one more or less likely. If it does, the Poisson model will get the spread of the counts wrong.
Concept
The interval does not have to be time. The book's first illustration is an editor counting misspelled words, which occur on average five times in 100 pages. The interval is the 100 pages.
Figure (svg): Three grids of 100 small squares, one square per page. A few squares in each grid are red, marking a page with a misspelled word: 5 red pages in the first grid, 8 in the second and 3 in the third.
The figure is an illustration of three books with that average rate. The counts differ, 5, 8 and 3, for the same reason Leah's mornings differ: the rate fixes the average, not the count. Distance, area and volume work the same way: flaws per metre of cable, weeds per square metre of lawn, bacteria per millilitre of water.
\[ \text{rate} = \frac{5 \text{ misspellings}}{100 \text{ pages}} \]
Whatever the interval, write the rate as a fraction with its units showing, count over length of interval. That habit will pay for itself in Part 2, where nearly every problem asks about an interval of a different size from the one the rate was quoted for.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 250 — the book editor's misspelled words
Concept
A binomial count stops at n: twenty bottles cannot contain twenty-one underfilled ones. A Poisson count has no such ceiling. Leah could, in principle, receive any number of calls in fifteen minutes, however unlikely a large number is.
\[ \text{binomial: } x = 0, 1, 2, \ldots, n \qquad \text{Poisson: } x = 0, 1, 2, 3, \ldots \]
The three dots at the end of the Poisson list are not an abbreviation for a last value you forgot to write; there is no last value. The probabilities get tiny very quickly, but every one of them is positive.
This has a practical consequence you will meet in Part 3. A question about more than some number asks for the probability of infinitely many values. You cannot add infinitely many terms by hand, so you will take the finite part and subtract it from one.
\[ P(x > 2) = 1 - P(x \le 2) \]
The missing ceiling is also a quick recognition test. If you can name the largest possible count, you are almost certainly looking at trials and one of the earlier families. If the count could keep going, think Poisson.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 251-252 — x = 0, 1, 2, 3, …
Sorting
For each count, ask two questions. Is there a fixed number of trials? If not, is there an interval with a steady average rate of independent events?
Sort into buckets
Sort each count into the family that models it.
The two in the last bucket are the ones worth thinking about. A bus timetable has a perfectly good average rate, five per hour, but the buses are anything but independent: when one has just come, the next is twelve minutes away. The tour coach does the opposite and delivers people in bunches. A rate alone does not make a Poisson.
And notice how the two binomial items announce themselves with a number of trials, 20 or 10. None of the Poisson items has one: there is no sensible answer to how many chances did the road have to contain a pothole.
The potholes and the misspellings are intervals of space, and they sort exactly like the emails, an interval of time. If you hesitated over them, it was probably because the word interval sounds like time. The model only cares that the stretch is fixed and the rate along it is steady.
Worked example
A news reporter says uh, on average, two times per broadcast. What is the probability that the reporter says uh more than two times in a broadcast? The book asks you to set the problem up in four parts before computing anything, and that setup is the skill.
Name the interval of interest
Why: the question is about one broadcast, and so is the rate, so no rescaling is needed.
\[ \text{interval} = \text{one broadcast} \]
Read off the average for that interval
Why: the rate is quoted per broadcast already.
\[ \mu = 2 \]
Define X in words and list its values
Why: a count of uhs, with no upper limit.
\[ X = \text{number of uhs in one broadcast}, \quad x = 0, 1, 2, 3, \ldots \]
Why is this Poisson? You are counting occurrences in an interval, a broadcast, at a known average rate, and there is no fixed number of trials: the reporter does not get a set number of chances to say uh.
Write the probability question
Why: more than two means three or more, which is strictly greater than 2.
\[ P(x > 2) \]
You will compute this in Part 3, once you have the formula; it comes to about 0.3233. For now the four lines above are the whole job, and every Poisson problem in the section starts with the same four.
Check the event against the words
Why: more than two uhs includes 3, 4, 5 and so on and excludes 2 itself, which is exactly the set of x strictly greater than 2.
\[ \{3, 4, 5, \ldots\} = \{x : x > 2\} \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Example 4.28
Worked example
A bank expects to receive six bad checks per day, on average. What is the probability of the bank getting fewer than five bad checks on any given day? Write the probability question mathematically.
Name the interval
Why: the question is about any given day, and the rate is per day.
\[ \text{interval} = \text{one day}, \quad \mu = 6 \]
Define X
Why: the number of bad checks the bank receives in one day.
\[ X \sim P(6), \quad x = 0, 1, 2, \ldots \]
Now translate fewer than five. A count is a whole number, so fewer than five means four or fewer. The values are 0, 1, 2, 3 and 4, and there are exactly five of them.
Write the event
Why: strictly less than 5.
\[ P(x < 5) \]
Rewrite it with at most, for computing
Why: for a whole-number count, below 5 is the same as at most 4.
\[ P(x < 5) = P(x \le 4) \]
That second form matters because tables, software and calculators give cumulative probabilities in the form at most k. Turning every question into at most or one minus at most is the translation step that decides whether the answer is right.
Check by listing the values
Why: the list must contain 4 and must not contain 5.
\[ \{0, 1, 2, 3, 4\} \text{ is five values, all below } 5 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Example 4.27
Trap
Setting up Example 4.27, a first attempt writes:
\[ P(x < 5) = P(x \le 5) \]
and computes it by adding the probabilities for 0 through 5.
\[ 0.2851 + 0.1606 = 0.4457 \]
Wrong. Five checks is not fewer than five.
For a count, the words decide which endpoint is in. Fewer than five stops at four. The extra term is the probability of exactly five, and with μ equal to 6 it is one of the largest bars.
\[ P(x < 5) = P(x \le 4) \]
\[ \quad = 0.2851 \]
Including 5 nearly doubles the answer. Say the list of values out loud before you write any symbol, and match the phrases: fewer than and below exclude the number; at most and no more than include it.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Example 4.27
Step zero
An emergency room gets an average of five patients per hour. A doctor wants the probability that it gets more than five patients in an hour. Before any computation, the book asks for the reason this is a Poisson distribution.
Discussion prompt
Give the reason the number of patients per hour is a Poisson count, name the interval and μ, and write the probability question.
The value 0.3840 was computed with node; you will be able to reproduce it after Part 3. What matters here is the reasoning, and in particular the phrase no fixed number of trials. There is no count of opportunities for a patient to arrive, which rules out the binomial straight away.
Is the independence assumption believable? Mostly. It fails after a bus crash, when many patients arrive together, just like the tour coach on the sorting slide. A good answer states the assumption instead of hiding it.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 251-252 — Try It 4.28
Prediction
Each situation below involves counting arrivals. Only one of them violates the requirement that events occur at a known, steady average rate across the interval.
Predict first
Which count is a poor fit for one Poisson distribution over the whole interval?
Correct: customers entering a café from 7 a.m. to 3 p.m., with a big rush at 8 a.m. and at noon
Why: The café's rate changes through the day, so no single average rate describes the whole eight hours, and rescaling one rate to an hour would be wrong. The other three have a steady rate. For the café, a separate Poisson with its own rate for each hour could still work.
The fix for the café is instructive. The model fails over the whole day, but within a quiet mid-morning hour the rate may be steady enough. Choosing an interval over which the rate really is constant is part of setting up the problem.
Radioactive decay is the textbook case of a genuine Poisson process: each atom decays independently of every other, with no memory. Meteors and a steady typist's errors are good approximations. None of them has a rush hour.
Section
Part 2
Concept
In Example 4.26, twelve loaves of bread are put on a bakery shelf in a half-hour period, on average. The question is about five minutes. The rate is quoted for one interval and the question asks about another, which is the usual situation.
Figure (svg): A horizontal bar for a 30-minute period, split by dashed lines into six 5-minute windows, with 12 orange dots for the loaves put on the shelf at scattered times. Under each window is the number of loaves that fell in it.
Cut the half hour into six five-minute windows. Because the rate is steady, each window gets the same share of the twelve on average.
\[ \mu_{5} = \frac{5}{30}(12) = 2 \text{ loaves} \]
The fraction five over thirty is the share of the stated interval that the new interval covers. Multiply the stated average by that share. Equivalently, turn the rate into loaves per minute and multiply by the number of minutes.
\[ \frac{12 \text{ loaves}}{30 \text{ min}} \times 5 \text{ min} = 0.4 \times 5 = 2 \text{ loaves} \]
Both routes give 2, and both are the same idea: μ is the average count, and an interval one sixth as long holds one sixth as much on average. The counts in the actual windows vary, as the figure shows; μ is their long-run average.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 250 — Example 4.26
Notation
Every rescaling in the section is this one line. Read it as: the rate, times the length of the interval you are asked about.
Annotate
On: \( \mu = \frac{\text{stated count}}{\text{stated length}} \times \text{new length} \)
The fraction on the left is a rate, count per unit of length. The units check tells you whether you have it the right way up: calls per minute times minutes leaves calls. If you end up with minutes per call times minutes, you have divided where you should have multiplied.
And there is a free direction check. Before computing, say whether the new interval is longer or shorter than the stated one. The answer must move the same way. That single sentence catches the most common error in the section.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 250-252 — Examples 4.26 and 4.29
Worked example
Leah receives about six telephone calls between 8 a.m. and 10 a.m. The question is about the next 15 minutes, so that is the interval of interest.
Define X for the interval asked about
Why: calls in 15 minutes, not in two hours.
\[ X = \text{calls Leah receives in 15 minutes}, \quad x = 0, 1, 2, \ldots \]
Put both intervals in the same units
Why: two hours is 120 minutes.
\[ \text{stated: } 6 \text{ calls per } 120 \text{ min} \]
The book counts intervals instead: two hours contain eight fifteen-minute intervals, so each gets one eighth of the six calls. The rate route below is the same calculation written with units.
Form the rate
Why: calls per minute.
\[ \frac{6}{120} = 0.05 \text{ calls per minute} \]
Multiply by the new length
Why: fifteen minutes.
\[ \mu = 0.05 \times 15 = 0.75 \]
Match the book's route
Why: one eighth of the two-hour average.
\[ \mu = \frac{1}{8}(6) = 0.75 \]
So X follows a Poisson distribution with μ equal to 0.75. Leah gets, on average, three quarters of a call per quarter hour; the average is not a possible count, and it does not need to be.
Check by scaling back up
Why: eight quarter-hours at 0.75 each must give back the two-hour average of 6, and 15 minutes is shorter than two hours, so μ had to fall below 6.
\[ 8 \times 0.75 = 6 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — Example 4.29
Fill the middle
The average number of fish caught in an hour is eight. The interval of interest is 15 minutes. Fill in the share of the hour and the average for 15 minutes.
Fill in the blanks
\mu = \frac602} \times 8 = ___
Why: An hour is 60 minutes, so 15 minutes is one quarter of it, and one quarter of eight fish is two fish. The new interval is shorter, so the average had to go down from eight.
The blank in the denominator is where the units decide the answer. Writing the hour as 1 instead of 60 would give 120 fish in a quarter of an hour, which the direction check rejects at once: a shorter interval cannot hold more fish on average.
Notice that the fishing μ and the bakery μ are both 2, from completely different rates and intervals. Once μ is found, the problem forgets where it came from. From here on, a quarter hour of fishing and five minutes of bread are the same distribution.
That is the sense in which μ is the only parameter. Everything about the situation, the rate, the units, the length of the interval, is summarised in the single number you just wrote.
One more thing to notice: the rate per minute is the same in both forms of the problem, eight over sixty fish per minute. Only the count changes with the interval. If a question ever seems to ask you to change the rate, reread it; it almost certainly wants a new μ from the same rate.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Try It 4.26
Estimation
A coffee kiosk serves 20 customers per hour on average, steadily through the morning. The question is about a 10-minute interval. Decide the direction first, then the size.
Predict first
What is μ for a 10-minute interval?
Correct: about 3.3, smaller than 20
Why: Ten minutes is one sixth of an hour, so the average is one sixth of twenty, about 3.33 customers. The rate does not change, but the interval is shorter, so the average count in it is smaller. Dividing 20 by 10 treats the minutes as if they were hours.
\[ \mu = \frac{10}{60} \times 20 = \frac{20}{6} \approx 3.33 \]
Three of the options come from real slips. Multiplying by 6 instead of dividing gives 120 and fails the direction check. Dividing by 10 mixes minutes with hours. Keeping 20 confuses the rate, which is the same in every interval, with the average count, which grows with the interval.
An estimate like this takes a few seconds and needs no formula. Ten minutes is a sixth of an hour; a sixth of twenty is a bit over three. Make that estimate every time, and a slip in the arithmetic will announce itself.
Worked example
On average, a clothing store gets 120 customers per day. Assume the store is open 12 hours each day. The question is about the first four hours.
The trap is in the word day. It is tempting to divide by 24, but the customers only arrive while the store is open. The rate has to be spread over the hours in which events can happen.
Rate per open hour
Why: 120 customers over the 12 hours the doors are open.
\[ \frac{120}{12} = 10 \text{ customers per hour} \]
Multiply by the interval of interest
Why: the first four hours.
\[ \mu = 10 \times 4 = 40 \]
State the distribution
Why: customers in the first four hours.
\[ X \sim P(40) \]
The exercise then asks for the probability of exactly 35 customers in those four hours. With μ this large, the formula needs software or logarithms, as you will see in Part 4; the answer is about 0.0485.
\[ P(x = 35) = \frac{40^{35} e^{-40}}{35!} \approx 0.0485 \]
Using 24 hours would have given a μ of 20 instead of 40, and a probability computed from it would have looked just as believable. The model assumes a steady rate during opening hours; the book's own assumption that the store is open 12 hours is what makes the rate usable.
Check by scaling back to a day
Why: three four-hour blocks make the 12-hour day, and three times 40 is the stated 120.
\[ 3 \times 40 = 120 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 practice, p. 270 — Exercises 56-59
Trap
For Leah's 15 minutes, a first attempt writes the rate upside down:
\[ \frac{120 \text{ min}}{6 \text{ calls}} = 20 \]
\[ \mu = 20 \times 15 = 300 \]
Another attempt divides the wrong way round:
\[ \mu = 6 \div \tfrac{15}{120} = 48 \]
Wrong both times.
Twenty is minutes per call, the average gap between calls, not a count. And a quarter hour cannot hold 48 or 300 calls on average when two whole hours hold only 6.
\[ \mu = \frac{6}{120} \times 15 \]
\[ \quad = 0.75 \]
Two checks catch both errors instantly. The direction check: fifteen minutes is shorter than two hours, so μ must be smaller than 6. The units check: calls per minute times minutes leaves calls; minutes per call times minutes leaves nonsense.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — Example 4.29
Section
Part 3
Concept
Once μ is on the right interval, the probability of exactly x occurrences depends on μ alone. The book gives the formula with Example 4.29.
\[ P(x) = \frac{\mu^{x} e^{-\mu}}{x!}, \qquad x = 0, 1, 2, \ldots \]
Two pieces may be new. The number e is a constant, about 2.71828, and your calculator has an e key. The exclamation mark is a factorial: x! means x times every whole number below it, down to 1, and zero factorial is defined to be 1.
\[ 3! = 3 \times 2 \times 1 = 6, \qquad 0! = 1 \]
Start with the simplest case, zero events. Anything to the power 0 is 1 and zero factorial is 1, so the formula collapses to a single term.
\[ P(0) = \frac{\mu^{0} e^{-\mu}}{0!} = e^{-\mu} \]
For Leah's quarter hour, that is e to the minus 0.75, about 0.4724: nearly half of all quarter hours bring no call. The larger μ, the smaller the chance of a quiet interval, and this one expression is the reason. The TI calculators call μ lambda and compute these values with poissonpdf and poissoncdf, but every value in this deck is also shown as arithmetic.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — the formula, with Example 4.29
Notation
Four ingredients, and each has a job. Read the formula aloud: μ to the x, times e to the minus μ, over x factorial.
Annotate
On: \( P(x) = \frac{\mu^{x}\, e^{-\mu}}{x!} \)
The contest between μ to the x on top and x factorial underneath is what shapes the distribution. While x is below μ, each step up multiplies the top by more than it multiplies the bottom, and the bars rise. Once x passes μ, the factorial wins and the bars fall.
The factor e to the minus μ does no shaping at all; it only scales. It is there so that the infinite list of probabilities adds up to exactly one, which is a fact you can check numerically, as the next part does.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — P(x) = μ to the x e to the minus μ over x factorial
Concept
Divide one probability by the one before it. Almost everything cancels: one extra factor of μ on top, one extra factor of x underneath.
\[ \frac{P(x)}{P(x-1)} = \frac{\mu^{x} / x!}{\mu^{x-1} / (x-1)!} = \frac{\mu}{x} \]
So you can build the whole distribution from P(0) by repeated multiplication. For the bakery, with μ equal to 2, start at e to the minus 2 and multiply by 2 over 1, then 2 over 2, then 2 over 3.
Figure (svg): A probability histogram for mu equals 2 with bars at 0 to 7. Curved orange arrows join each bar to the next, labelled times 2 over 1, times 2 over 2, times 2 over 3, times 2 over 4 and times 2 over 5. The bar at 3 is yellow, height 0.1804.
\[ 0.1353 \xrightarrow{\times 2} 0.2707 \xrightarrow{\times 1} 0.2707 \xrightarrow{\times 2/3} 0.1804 \]
The multiplier explains the shape of every Poisson distribution. It is bigger than 1 while x is less than μ, so the bars climb; it is less than 1 once x exceeds μ, so they fall. When μ is a whole number, the multiplier at x equal to μ is exactly 1, and two bars tie for the peak, as 1 and 2 do here.
This also gives you a fast, reliable way to compute several consecutive probabilities by hand, which is exactly what a cumulative question needs.
Worked example
Part 2 found μ for five minutes at the bakery: two loaves. The question asks for the probability that exactly three loaves are put on the shelf in five minutes. The book stops once it has written P of x equal to 3; here is the finish.
State the distribution
Why: the five-minute average from Part 2.
\[ X \sim P(2), \quad \text{find } P(x = 3) \]
Substitute into the formula
Why: μ is 2 and x is 3.
\[ P(3) = \frac{2^{3} e^{-2}}{3!} \]
Evaluate the pieces
Why: two cubed, e to the minus 2, and three factorial.
\[ 2^{3} = 8, \quad e^{-2} \approx 0.135335, \quad 3! = 6 \]
Multiply the top
Why: eight times e to the minus 2.
\[ 8 \times 0.135335 = 1.08268 \]
Divide by the factorial
Why: over six.
\[ P(3) = \frac{1.08268}{6} \approx 0.1804 \]
In context: about 18 percent of five-minute periods see exactly three loaves go on the shelf. Three is a little above the average of two, and it is a common outcome, but less common than two.
Check with the ratio chain
Why: start from e to the minus 2 and multiply by 2 over 1, 2 over 2 and 2 over 3; the product of the multipliers is four thirds, and it lands on the same value.
\[ 0.135335 \times 2 \times 1 \times \tfrac{2}{3} \approx 0.1804 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 250 — Example 4.26; the value 0.1804 is supplied here
Picture it
Here is Leah's distribution for a quarter hour, the book's Figure 4.6, drawn from the formula with μ equal to 0.75. The question in Example 4.29 is the probability of more than one call.
Figure (svg): A probability histogram of the number of calls in 15 minutes when mu is 0.75: bars at 0 to 5 of heights 0.472, 0.354, 0.133, 0.033, 0.006 and 0.001. The bars at 0 and 1 are blue; the bars from 2 on are yellow and labelled more than one call, total 0.1734.
With μ below 1, the tallest bar is at zero and every bar after it is smaller: the ratio μ over x is 0.75 at the first step and shrinks from there. This is the strongly skewed shape a small μ always gives.
More than one call is the yellow part, and it has no end. It includes 2, 3, 4 and every count beyond. The blue part, zero or one call, is just two bars. So add the two bars you can see, and take the total away from one.
\[ P(x > 1) = 1 - \left[P(0) + P(1)\right] \]
Every more-than or at-least question in this section is solved by this picture. The finite side is added; the infinite side is whatever is left of the total probability of one.
Compare the heights with the book's figure: 0.472, 0.354, 0.133, 0.033, and a bar labelled etc. The book's etc. is honest. The bars at 5, 6, 7 and beyond are too short to see, but they are there, and together with the visible yellow bars they make up the 0.1734.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 253 — Figure 4.6
Worked example
With μ equal to 0.75 from the first half, find the probability that Leah receives more than one call in the next 15 minutes.
Turn the tail into a complement
Why: more than one is the opposite of at most one.
\[ P(x > 1) = 1 - P(x \le 1) \]
Spell out at most one
Why: the two counts 0 and 1.
\[ P(x \le 1) = P(0) + P(1) \]
Compute P(0)
Why: zero events: e to the minus μ.
\[ P(0) = \frac{0.75^{0} e^{-0.75}}{0!} = e^{-0.75} \approx 0.4724 \]
Compute P(1)
Why: one step up the ratio chain, times μ over 1.
\[ P(1) = \frac{0.75^{1} e^{-0.75}}{1!} = 0.75(0.4724) \approx 0.3543 \]
Add them
Why: the probability of at most one call.
\[ P(x \le 1) = 0.4724 + 0.3543 = 0.8267 \]
Subtract from one
Why: the tail.
\[ P(x > 1) = 1 - 0.8267 = 0.1733 \]
The book's text says to subtract the probability less than 1, but what it actually subtracts, correctly, is P(0) plus P(1), the probability of at most 1. Also, adding the rounded terms gives 0.1733; unrounded, the terms are 0.47237 and 0.35427, and the tail is 0.17336, which the calculator's poissoncdf reports as 0.1734. In context: in about 17 percent of quarter hours, Leah gets more than one call.
Check by adding the tail directly
Why: the first four tail bars, for 2 to 5 calls, already come to 0.1733 and the rest are tiny, so the complement and the direct sum agree.
\[ 0.1329 + 0.0332 + 0.0062 + 0.0009 = 0.1732 \approx 0.1734 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 252-253 — Example 4.29
Concept
The complement rule, one minus the other side, is a convenience for most distributions. For the Poisson it is a necessity, and the reason is the missing upper limit from Part 1.
\[ P(x > 1) = P(2) + P(3) + P(4) + \cdots \]
That sum never ends. You could add terms until they are too small to matter, as the check on the last slide did, but you would have to judge when to stop. The complement replaces an endless sum with a short one, and it is exact.
It relies on one fact: all the Poisson probabilities, over every count from zero upwards, add to exactly one. Here is that fact checked numerically, adding the first sixty terms for three different μ values in node.
| μ | P(0) + P(1) + … + P(60) |
|---|---|
| 0.75 | 1.0000 |
| 2 | 1.0000 |
| 6 | 1.0000 |
So the rule for every question in this section is: whatever side of the event is finite, add it; if the event itself is the infinite side, add the other side and subtract from one.
\[ P(x \ge k) = 1 - P(x \le k - 1) \]
Getting that k minus 1 right is where people lose marks, which is the next slide.
Error analysis
Four students computed the probability of more than one call in Leah's quarter hour, with μ equal to 0.75. One is right. Read each line and find the fault in the other three.
Annotate
On: \( 1 - P(0) = 0.5276 \quad 1 - P(1) = 0.6457 \quad 1 - [P(0) + P(1)] = 0.1734 \)
Three of the four mistakes come from the same place: deciding which bars are in the event before writing any symbols. Draw the bars, colour the ones the words describe, and the complement writes itself.
The first attempt deserves a second look, because the book's own sentence invites it. On p. 252 the text says subtract the probability less than 1, but less than 1 means only zero calls. The calculation printed just below it does the right thing, subtracting zero and one calls. When words and working disagree, check against the picture.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — the complement in Example 4.29
Matching
Every Poisson question arrives in words. Match each phrase, for a count X, to the calculation that answers it. The rule of thumb: add the finite side.
Match the pairs
Why: Fewer than 8 stops at 7. More than 2 starts at 3, so subtract everything up to 2. At least 2 starts at 2, so subtract only 0 and 1. At most 160 includes 160 and is already a finite sum from 0. Exactly 175 is a single bar.
The pair that catches people is more than 2 against at least 2. They differ by a single bar, the one at 2, and that bar is often one of the largest. The subtraction for at least 2 stops at 1; for more than 2 it stops at 2.
Fewer than 8 is Try It 4.27: an electronics store expecting ten returns a day wants the chance of fewer than eight. In symbols it is the probability of at most 7, which with μ equal to 10 comes to about 0.2202.
Say the first value included and the last value included, out loud, before choosing. That habit turns every one of these into a list you can check.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 251, 253 — Try It 4.27, Example 4.30 and Try It 4.30
Worked example
Back to the bank of Example 4.27, which expects six bad checks a day. Part 1 set up the question; now compute it.
The event
Why: fewer than five is at most four.
\[ P(x < 5) = P(x \le 4) = P(0) + P(1) + P(2) + P(3) + P(4) \]
Start the chain at P(0)
Why: e to the minus 6.
\[ P(0) = e^{-6} \approx 0.00248 \]
Multiply up by 6 over x
Why: 6 over 1, then 6 over 2, 6 over 3, 6 over 4.
\[ 0.00248 \to 0.01487 \to 0.04462 \to 0.08924 \to 0.13385 \]
Add the five terms
Why: the cumulative probability.
\[ P(x \le 4) \approx 0.2851 \]
Figure (svg): A probability histogram of bad checks per day for mu equals 6, bars at 0 to 14. The bars at 0 to 4 are green; the bar at 5 is red and labelled not fewer than 5; the rest are grey. The peak is shared by 5 and 6.
In context: on about 29 percent of days the bank receives fewer than five bad checks. With an average of six a day, four or fewer is well below average but hardly rare.
Check with the formula for one term
Why: compute P(4) directly and compare with the chain's last value.
\[ P(4) = \frac{6^{4} e^{-6}}{4!} = \frac{1296(0.00248)}{24} \approx 0.1339 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Example 4.27; the value 0.2851 is computed here
Fill the middle
The reporter says uh twice per broadcast on average, so μ is 2, and the question from Part 1 was the probability of more than two. The three bars you need are the ones already built on the ratio-chain slide. Fill in the missing bar and the answer.
Fill in the blanks
P(x > 2) = 1 - (0.1353 + 0.2707 + 0.2707) = 0.3233
Why: With a mean of two, the bars for one and for two are equal, both about 0.2707, because the multiplier from one to two is two over two. The three bars add to 0.6767, and one minus that leaves 0.3233.
In context: in about a third of broadcasts the reporter says uh three or more times. That is not a rare event; more than the average is common when the average is small, because the count can only fall two below it but can rise without limit.
If you got 0.5940 you subtracted only the first two bars, which answers at least two. If you got 0.1804 you found the single bar at three and stopped, which answers exactly three.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Example 4.28; the value 0.3233 is computed here
Worked example
A customer service centre receives about ten emails every half hour. What is the probability that it receives more than four emails in the next six minutes?
Rescale onto six minutes
Why: six minutes is one fifth of a half hour.
\[ \mu = \frac{6}{30} \times 10 = 2 \]
Write the event as a complement
Why: more than four is the opposite of at most four.
\[ P(x > 4) = 1 - P(x \le 4) \]
List the five terms
Why: the chain for μ equal to 2, from the bakery slide.
\[ 0.1353, \; 0.2707, \; 0.2707, \; 0.1804, \; 0.0902 \]
Add them
Why: at most four emails.
\[ P(x \le 4) = 0.9473 \]
Subtract from one
Why: the tail.
\[ P(x > 4) = 1 - 0.9473 = 0.0527 \]
In context: about 5 percent of six-minute stretches bring more than four emails. With an average of two, a burst of five or more is unusual but happens a few times a day, since a working day holds dozens of six-minute stretches.
Check the direct tail
Why: the next three bars, for 5, 6 and 7 emails, already come to almost all of it.
\[ 0.0361 + 0.0120 + 0.0034 = 0.0515 \approx 0.0527 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 253 — Try It 4.29
Worked example
Text message users receive or send an average of 41.5 text messages per day. How many is that per hour, what is the probability of exactly two in an hour, and of more than two?
Rescale onto one hour
Why: an hour is one twenty-fourth of a day.
\[ \mu = \frac{41.5}{24} \approx 1.7292 \]
Exactly two, from the formula
Why: μ squared times e to the minus μ, over two factorial.
\[ P(2) = \frac{1.7292^{2} e^{-1.7292}}{2!} \]
Evaluate
Why: 1.7292 squared is 2.9901 and e to the minus 1.7292 is 0.17743.
\[ P(2) = \frac{2.9901 \times 0.17743}{2} \approx 0.2653 \]
The two bars below
Why: P(0) is e to the minus μ; P(1) is μ times that.
\[ P(0) \approx 0.1774, \quad P(1) \approx 0.3068 \]
At most two, then the complement
Why: add the three bars and subtract from one.
\[ P(x > 2) = 1 - (0.1774 + 0.3068 + 0.2653) = 0.2505 \]
Figure (svg): A probability histogram of texts per hour for mu equals 1.7292, bars at 0 to 7. The bars at 0 and 1 are grey, the bar at 2 is yellow with height 0.2653, and the bars from 3 on are orange, with total 0.2505.
In context: a typical user sends or receives about 1.7 texts an hour, sees exactly two in about a quarter of hours, and more than two in another quarter. The book assumes the texts are spread evenly over all 24 hours, sleep included, which is the steady-rate assumption doing a lot of work.
Check the rescaling and the total
Why: 24 hours at 1.7292 gives back the daily 41.5, and the four pieces of the picture add to one.
\[ 24(1.7292) \approx 41.5, \quad 0.1774 + 0.3068 + 0.2653 + 0.2505 = 1 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 254 — Example 4.31
Tweak it
The chance of no events at all has the simplest formula of all, e to the minus μ, and μ is the rate times the length of the interval. The graph shows the chance of a quiet interval against its length in minutes, for a rate you choose.
Parameter explorer
Set the rate to 0.05 calls per minute, Leah's rate, and read the chance of no call in 15 minutes. Then find how long an interval gives an even chance of no call at all, and what happens to that length when you double the rate.
\[ \text{rate} = {r} \text{ per minute}, \quad P(\text{no call in } t \text{ minutes}) = e^{-\text{rate} \times t} \]
\[ P(0) = e^{-0.05 \times 15} = e^{-0.75} \approx 0.4724 \]
The curve never reaches zero: however long the interval, a quiet stretch remains possible, just increasingly unlikely. An hour with no call happens with probability e to the minus 3, about 0.05, one morning-hour in twenty.
The even-chance length solves e to the minus 0.05t equal to one half, which is about 13.9 minutes. Double the rate and every length on the curve halves, because only the product of rate and length matters. That is the rescaling idea of Part 2, seen as a picture.
Section
Part 4
Concept
The book gives the notation X ~ P(μ) and one extra fact: the standard deviation is the square root of μ. The variance, the square of the standard deviation, is therefore μ itself.
\[ \text{mean} = \mu, \qquad \sigma^{2} = \mu, \qquad \sigma = \sqrt{\mu} \]
You can check this with the expected value table of Section 4.2, on the bakery distribution with μ equal to 2. The table below is computed from the formula; the rows beyond 8 carry less than 0.0003 of the probability.
| x | P(x) | x P(x) | x² P(x) |
|---|---|---|---|
| 0 | 0.1353 | 0.0000 | 0.0000 |
| 1 | 0.2707 | 0.2707 | 0.2707 |
| 2 | 0.2707 | 0.5413 | 1.0827 |
| 3 | 0.1804 | 0.5413 | 1.6240 |
| 4 | 0.0902 | 0.3609 | 1.4436 |
| 5 | 0.0361 | 0.1804 | 0.9022 |
| 6 | 0.0120 | 0.0722 | 0.4331 |
| 7 | 0.0034 | 0.0241 | 0.1684 |
| 8 | 0.0009 | 0.0069 | 0.0550 |
\[ \sum x\,P(x) \approx 2.00, \qquad \sum x^{2} P(x) \approx 6.00 \]
\[ \sigma^{2} = 6 - 2^{2} = 2 = \mu \]
That is why a Poisson is fully specified by a single number. Once you know the average count, you know the spread as well. Compare the binomial, which needs n and p, and whose variance npq is smaller than its mean.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 252 — X ~ P(μ) and σ = √μ
Worked example
According to an email management company, a user gets on average 147 emails per day, and X, the number received in a day, follows P(147). Find the probability of exactly 160, of at most 160, and the standard deviation.
Write the exact probability
Why: μ is 147, x is 160.
\[ P(160) = \frac{147^{160} e^{-147}}{160!} \]
Every piece of that fraction is astronomically large or small; 160 factorial has 285 digits. A calculator's poissonpdf gets around this internally. You can too, by working with natural logarithms, which turn the powers and the factorial into sums.
Take logs
Why: 160 times ln 147, minus 147, minus ln 160 factorial.
\[ \ln P = 160(4.99043) - 147 - 655.4849 \]
Evaluate
Why: 798.4692 minus 147 minus 655.4849.
\[ \ln P \approx -4.0156 \]
Undo the log
Why: e to that power.
\[ P(160) = e^{-4.0156} \approx 0.0180 \]
At most 160
Why: the sum of 161 terms, from software, poissoncdf of 147 and 160.
\[ P(x \le 160) \approx 0.8666 \]
The standard deviation
Why: the square root of μ.
\[ \sigma = \sqrt{147} \approx 12.1244 \]
Figure (svg): A probability histogram of emails per day for mu equals 147, bars from 105 to 190 forming a nearly symmetric hump peaking near 147 at about 0.033. A blue band marks 147 plus or minus 12.12. The bar at 160 is yellow; bars up to 160 are blue and bars above are grey.
Why logs? A logarithm turns a product into a sum and a power into a multiple, so 147 to the 160th becomes 160 times ln 147, a number you can write down. The log of 160 factorial is the sum of the logs of 1 to 160, which software adds in a blink. This is exactly what poissonpdf does inside.
In context: a user gets 147 emails a day on average, typically within about 12 of that. A day of exactly 160 happens less than 2 percent of the time, as any single exact count does when μ is large, but a day of 160 or fewer happens about 87 percent of the time.
Check against the picture
Why: 160 is 13 above the mean, about 1.07 standard deviations, so most of a hump-shaped distribution should lie below it, and 0.8666 does.
\[ \frac{160 - 147}{12.12} \approx 1.07 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 253 — Example 4.30
Picture it
Three Poisson distributions drawn from the formula: μ equal to 0.75, like Leah's quarter hour; 5.5, like the law office coming up; and 20. Each has a yellow marker at μ and a blue band one standard deviation either side.
Figure (svg): Three probability histograms side by side. Mu equals 0.75: tallest bar at 0, falling fast, strongly skewed right. Mu equals 5.5: a hump peaking at 5 with a longer right tail. Mu equals 20: a nearly symmetric bell from about 8 to 34. Each has a yellow marker at mu and a blue band one standard deviation either side.
At μ equal to 0.75 the distribution is jammed against zero: counts cannot go below zero, but they can run to the right. At 5.5 it is a hump with a noticeably longer right tail. At 20 it is nearly symmetric and bell-shaped, the shape Chapter 6 will call normal.
Look at the widths. The blue bands grow as μ grows, from about 0.87 to 2.35 to 4.47, but far more slowly than μ. Multiply μ by about 27 and the standard deviation grows only by about 5, its square root. Relative to the average, a larger μ means a tighter distribution.
\[ \frac{\sigma}{\mu} = \frac{\sqrt{\mu}}{\mu} = \frac{1}{\sqrt{\mu}} \]
That ratio is why a large average count is more predictable in relative terms. A day with 147 emails rarely strays by more than a few percent; a quarter hour with 0.75 calls routinely doubles or zeroes its average.
Tweak it
The graph draws the Poisson probabilities for the μ you choose, each bar as a flat step of width one centred on its count. Slide μ from small to large.
Parameter explorer
Start at mu equal to 0.75 and slide up slowly. Find where the tallest bar leaves zero, what happens at a whole-number mu such as 3, and how the height of the tallest bar changes as mu grows past 10.
\[ \mu = {m}, \quad P(x) = \frac{\mu^{x} e^{-\mu}}{x!} \]
Below μ of 1, the multiplier μ over x is less than 1 from the very first step, so zero is always the most likely count. Just past 1, the first step multiplies by more than 1 and the peak jumps to 1.
At a whole number such as 3, the multiplier at x equal to 3 is exactly 3 over 3, so the bars at 2 and 3 are equal and tie for the peak. In general the peak sits at the whole-number part of μ.
As μ grows, the peak gets lower. The total is always one, and the probability is spread over more and more counts, so no single count can be very likely. Exactly 160 emails, at under 2 percent, was a case of this.
Prediction
A law office's switchboard gets an average of 5.5 calls during the noon hour on Mondays. Use the ratio μ over x before you compute anything.
Predict first
Which count of calls is the single most likely?
Correct: 5
Why: Going from 4 to 5 multiplies by 5.5 over 5, which is more than one, so the bar rises. Going from 5 to 6 multiplies by 5.5 over 6, which is less than one, so it falls. The peak is at 5, with probability about 0.1714, just above 0.1571 at 6.
\[ P(5) \approx 0.1714, \qquad P(6) \approx 0.1571, \qquad P(4) \approx 0.1558 \]
The value 5.5 is the mean, but it is not a possible count, so it cannot be the most likely one. Six is tempting because it is the whole number nearest 5.5 from above; the ratio rule says otherwise, because the step from 5 to 6 multiplies by less than one.
The three central bars are close in height. With μ this size, the distribution is fairly flat near its middle: 4, 5 and 6 calls are almost equally likely, and together they account for fewer than half of noon hours.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, p. 279 — Exercise 117
Worked example
People aged 14 to 17 send an average of 187 text messages each day, so X, the number sent in a day, follows P(187). Find the probability of exactly 175, of at most 150, and the standard deviation. (The book's wording switches between people and girls aged 14 to 17; the model is the same.)
The standard deviation first
Why: it tells you what counts are ordinary.
\[ \sigma = \sqrt{187} \approx 13.67 \]
Exactly 175, by logs
Why: 175 ln 187, minus 187, minus ln 175 factorial.
\[ P(175) = \frac{187^{175} e^{-187}}{175!} \approx 0.0203 \]
How far below the mean is 150?
Why: in standard deviations.
\[ \frac{150 - 187}{13.67} \approx -2.71 \]
At most 150, from software
Why: the sum of 151 terms.
\[ P(x \le 150) \approx 0.0030 \]
In context: a teenager sending 175 texts in a day is entirely typical, less than one standard deviation below the mean, though any one exact count is unlikely. Sending 150 or fewer happens on only about 3 days in 1,000; if you saw it, you would suspect a phone left at home, or that the Poisson model with this μ does not describe that person.
The standard deviation is what makes these judgements possible without any tables. Anything within one σ of μ is ordinary; beyond two σ it is unusual; 2.7 σ below is rare.
Check the variance equals the mean
Why: square the standard deviation.
\[ 13.67^{2} \approx 186.9 \approx 187 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 253 — Try It 4.30
Worked example
A Minneapolis law office's switchboard gets an average of 5.5 incoming calls during the noon hour on Mondays. The staff can handle up to six calls in an hour. Let X be the number of calls received at noon.
Mean and standard deviation
Why: one parameter gives both.
\[ \mu = 5.5, \quad \sigma = \sqrt{5.5} \approx 2.35 \]
At most six: start the chain
Why: P(0) is e to the minus 5.5.
\[ P(0) = e^{-5.5} \approx 0.0041 \]
Multiply up by 5.5 over x
Why: for x from 1 to 6.
\[ 0.0041, \, 0.0225, \, 0.0618, \, 0.1133, \, 0.1558, \, 0.1714, \, 0.1571 \]
Add the seven terms
Why: the chance the staff can cope.
\[ P(x \le 6) \approx 0.6860 \]
More than eight
Why: two more terms, 0.1234 and 0.0849, then the complement.
\[ P(x > 8) = 1 - 0.8944 = 0.1056 \]
Figure (svg): A probability histogram of calls in the noon hour for mu equals 5.5, bars at 0 to 14. Bars 0 to 6 are green, labelled staff can cope, total 0.6860. Bars 7 and 8 are yellow. Bars 9 and above are red, labelled more than eight, total 0.1056.
In context: exactly six calls happens about 16 percent of Mondays, which means the staff is working at full stretch. More seriously, the chance of more than six is 1 minus 0.6860, about 0.31. Nearly one Monday noon in three overwhelms the current staff, and one in ten brings more than eight calls. That is an argument for another person at noon.
Check the last term of the chain from the formula
Why: compute P(6) directly; it must equal the chain's 0.1571, or an earlier multiplication went wrong.
\[ P(6) = \frac{5.5^{6} e^{-5.5}}{6!} = \frac{27680.64 \times 0.0040868}{720} \approx 0.1571 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, p. 279 — Exercise 117
Trap
For Example 4.30, a first attempt at the spread writes:
\[ \sigma = \mu = 147 \]
and concludes that 160 emails is only a tiny fraction of a standard deviation from the mean, and so is barely different from 147.
Wrong. μ is the variance, not the standard deviation.
The Poisson fact is that the variance equals the mean. The standard deviation is its square root, and that is in the same units as the counts.
\[ \sigma^{2} = 147 \]
\[ \sigma = \sqrt{147} \approx 12.12 \]
A standard deviation of 147 would mean a typical day strayed by 147 emails from an average of 147, so zero emails would be ordinary. With 12.12, a day more than 25 or so from 147 is unusual. Always ask whether the spread you found makes sense against the mean.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 253 — Example 4.30: σ = √147 ≈ 12.1244
Estimation
Atlanta's Hartsfield-Jackson airport, one of the busiest in the world, averages 2,700 arrivals and departures each day. The book asks how many per hour, and for the probability of exactly 100 and of at most 100 in one hour. Estimate before computing.
\[ \mu = \frac{2700}{24} = 112.5, \qquad \sigma = \sqrt{112.5} \approx 10.61 \]
Predict first
Roughly how likely is an hour with at most 100 arrivals and departures?
Correct: about 0.13
Why: One hundred is 12.5 below the mean of 112.5, a little over one standard deviation of about 10.6. For a hump-shaped distribution, the part more than about one standard deviation below the mean is a bit under a fifth, so the answer should be around one in eight. Software gives 0.1279.
\[ P(x = 100) \approx 0.0194, \qquad P(x \le 100) \approx 0.1279 \]
The estimate works because the distribution with μ equal to 112.5 is almost symmetric, like the right-hand panel of the shapes slide. You locate 100 in standard deviations, about 1.18 below the mean, and read the answer from the shape.
One honest caveat: an airport's traffic is not steady around the clock; there are few flights at 3 a.m. and banks of departures at peak hours. The average of 112.5 per hour is really an average over very different hours, which is exactly the café problem from Part 1.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 254 — Try It 4.31
Comparison
The standard deviation grows with μ, but more slowly. Fill in the blanks to see how the ratio of σ to μ behaves across the examples of this section.
Comparison matrix
| example | μ | σ = √μ | σ ÷ μ |
|---|---|---|---|
| Leah, 15 minutes | 0.75 | 0.87 | 1.15 |
| law office, noon hour | 5.5 | 2.35 | 0.43 |
| clothing store, 4 hours | 40 | 6.32 | 0.16 |
| emails, one day | 147 | 12.12 | 0.08 |
Read down the last column. With μ below 1, the spread is bigger than the average itself: a quarter hour with 0 or 2 calls is routine. With μ at 147, the spread is only about 8 percent of the average.
This is the square root at work. The ratio is one over the square root of μ, so quadrupling μ only halves the relative spread. It is also why larger intervals give more predictable counts: a store knows its daily total far more precisely, in relative terms, than its total for any one minute.
The same pattern will reappear in Chapter 7, when the spread of an average shrinks with the square root of the sample size.
Real world
Because a Poisson's variance must equal its mean, real count data carry a built-in test of the model. Suppose a city records the number of burglaries per week in one district over two years, and finds a mean of 4 and a variance of 9. These numbers are an illustration, not from the book.
Discussion prompt
Explain why the variance of real count data can be used to test whether a Poisson model fits. What does a variance of 9 against a mean of 4 suggest about how the burglaries happen?
Statisticians call this overdispersion, and it is the most common way real counts fail the Poisson model. The clustered timeline in Part 1 is the picture of it: the same average, but bunches and gaps too large for independent events.
The opposite, a variance well below the mean, points to events that are more regular than random, like the bus timetable. Either way, comparing two numbers you can compute from any data set tells you whether the one-number model deserves your trust.
Section
Part 5
Concept
The second characteristic of a Poisson experiment is a use rather than a description: the Poisson may be used to approximate the binomial when the probability of success is small and the number of trials is large. The book's first examples of small and large are 0.01 and 1,000.
The idea is that a binomial with a huge n and a tiny p looks, from the outside, like a count of rare events in an interval. Twenty thousand bottles, each with a tiny chance of a fault, behave like faults arriving at a steady rate along the production line.
\[ X \sim B(n, p) \quad \approx \quad X \sim P(\mu), \qquad \mu = np \]
The Poisson is matched to the binomial by giving it the same mean, n times p. The chapter review states the conditions under which the match is good.
\[ n \ge 20 \quad \text{and} \quad p \le 0.05 \]
On p. 254 the book words these as n greater than 20 and p less than 0.05; the chapter review on p. 263 uses at least and at most. The boundary does not matter in practice; what matters is that p is small. Recall the warm-up: 20 bottles with p equal to 0.05 gave 0.3585 for zero faults, and the Poisson with μ equal to 1 gives e to the minus 1, which is 0.3679.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 250, 254, 263 — the second characteristic and the chapter review
Concept
The two characteristics are really one idea seen from both ends. Take Leah's quarter hour, with μ equal to 0.75, and cut it into n tiny slices, so short that two calls in one slice almost never happen. Each slice is then a trial: it holds a call or it does not.
\[ p = \frac{0.75}{n} \text{ per slice}, \qquad np = 0.75 \]
A quiet quarter hour means every slice is empty, which is a binomial probability of zero successes. Here it is for more and more slices, computed in node, against the Poisson value e to the minus 0.75.
| slices n | p = 0.75/n | binomial P(0) = (1 − p)ⁿ | binomial P(1) |
|---|---|---|---|
| 10 | 0.075 | 0.4586 | 0.3718 |
| 100 | 0.0075 | 0.4710 | 0.3559 |
| 1000 | 0.00075 | 0.4722 | 0.3544 |
| Poisson | 0.4724 | 0.3543 |
As the slices get finer, the binomial answers settle onto the Poisson ones. With ten slices the probability of a quiet quarter hour is off in the second decimal place; with a thousand it agrees to three. The same thing happens for one call, and for every other count.
That is why the Poisson approximates the binomial, and it is also where the formula's e to the minus μ comes from: it is the limit of the chance that every one of a huge number of tiny trials fails. You do not need the limit to use the formula, but it explains why a Poisson needs no n. The number of slices is arbitrary; only their total expected count, μ, survives.
Concept
Both distributions have the same mean, np, by construction. Compare their variances. The binomial's variance is npq; the Poisson's is its mean, np. Their ratio is just q.
\[ \frac{\text{binomial variance}}{\text{Poisson variance}} = \frac{npq}{np} = q = 1 - p \]
Figure (svg): A graph of the ratio of the binomial variance to the Poisson variance, q equals 1 minus p, for p from 0 to 0.5. It falls in a straight line from 1 to 0.5. The strip from p equals 0 to 0.05 is shaded; dots mark p equals 0.0102, 0.03 and 0.4 with ratios 0.99, 0.97 and 0.6.
When p is small, q is close to 1 and the two spreads agree. At the seismic example's p of 0.0102 the ratio is 0.99; for the bulbs of Exercise 119, 0.97. Once p reaches 0.4 the Poisson's variance is two thirds larger than the binomial's, so its standard deviation is about 29 percent too big, and no value of n can repair that, because n cancels out of the ratio.
So the condition on p is the binding one. The condition on n makes sure the mean np is a sensible size and that the individual probabilities are close, but a large n with a large p is no excuse to use a Poisson. Part 5's counterexample shows what goes wrong.
Worked example
The probability of low seismic activity in Alaska over a 48-hour period was reported as about 1.02 percent. Over the next 200 days, find the probability of low seismic activity on exactly ten days, with the binomial and with the Poisson. Are they close?
Check the conditions
Why: 200 trials, and a small chance on each.
\[ n = 200 \ge 20, \quad p = 0.0102 \le 0.05 \]
The binomial term
Why: choose 10 of 200 days, times the probabilities.
\[ P(10) = \binom{200}{10}(0.0102)^{10}(0.9898)^{190} \]
Evaluate its three pieces
Why: each computed in node.
\[ (2.2451 \times 10^{16})(1.2190 \times 10^{-20})(0.14257) \]
Multiply
Why: the binomial probability.
\[ P_{B}(10) \approx 0.000039 \]
The Poisson mean
Why: n times p.
\[ \mu = 200(0.0102) = 2.04 \]
The Poisson term
Why: 2.04 to the tenth is 1248.25; e to the minus 2.04 is 0.13003; ten factorial is 3,628,800.
\[ P_{P}(10) = \frac{1248.25 \times 0.13003}{3628800} \approx 0.000045 \]
Figure (svg): Paired bars for x equals 0 to 7: blue bars for the binomial with n equals 200 and p equals 0.0102, orange bars for the Poisson with mu equals 2.04. At every x the two bars are almost exactly the same height, peaking at about 0.27 at x equals 2.
Notice the Poisson line needs no combination at all. The binomial's choose 10 of 200 is about 22 thousand million million, and multiplying it by a power of 0.0102 that small is exactly the kind of arithmetic that goes wrong by hand. The Poisson packs all of it into μ equal to 2.04.
The two answers, 0.000039 and 0.000045, differ by about 15 percent of their size, but both say the same thing in context: ten days of low activity out of 200 is essentially impossible when the average is about two. The figure shows the agreement where the probability actually is, near the mean, is excellent.
Check the spreads agree
Why: the binomial standard deviation, root npq, against the Poisson's root np.
\[ \sqrt{200(0.0102)(0.9898)} \approx 1.421, \quad \sqrt{2.04} \approx 1.428 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 254 — Example 4.32
Worked example
The probability of seismic activity for the next 48 hours in the Kuril Islands was reported at about 1.43 percent. For the next 100 days, find the probability of that activity on five of them, with both distributions. (The book calls it moderate activity in one sentence and low activity in the next; either way it is one event with p equal to 0.0143.)
Check the conditions
Why: 100 trials, small p.
\[ n = 100, \quad p = 0.0143, \quad \mu = np = 1.43 \]
The binomial term
Why: choose 5 of 100, times the probabilities.
\[ \binom{100}{5}(0.0143)^{5}(0.9857)^{95} \]
Evaluate
Why: 75,287,520 times 5.9797 times ten to the minus 10, times 0.25454.
\[ P_{B}(5) \approx 0.01146 \]
The Poisson term
Why: 1.43 to the fifth is 5.9797; e to the minus 1.43 is 0.23931; five factorial is 120.
\[ P_{P}(5) = \frac{5.9797 \times 0.23931}{120} \approx 0.01192 \]
In context: activity on five of the next 100 days is unlikely, a little over 1 percent by either method, since the average is under one and a half days. The two answers agree to within about 4 percent of their size, closer than in the Alaska example because five is nearer the mean here than ten was there.
Why is the agreement better here? The Poisson's larger variance shows up mostly in the tails, where it puts a little too much probability. Ten days out of 200 was deep in the tail of a distribution centred near 2; five days out of 100 is less far out from a mean of 1.43, so the tail excess matters less.
Notice how much easier the Poisson line is. The binomial needs a combination in the tens of millions and a power to the 95th; the Poisson needs one small power, one e and one factorial. Before calculators, that was the whole point of the approximation.
Check the ratio of the answers
Why: the Poisson value divided by the binomial value should be close to one.
\[ \frac{0.01192}{0.01146} \approx 1.04 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 254 — Try It 4.32
Worked example
A manufacturer knows that 3 percent of its bulbs are defective. Find the probability that a string of 100 lights contains at most four defective bulbs, using both the binomial and the Poisson.
Check the conditions and match the mean
Why: n is 100, p is 0.03.
\[ n = 100 \ge 20, \quad p = 0.03 \le 0.05, \quad \mu = 3 \]
Poisson terms for 0 to 4
Why: e to the minus 3, then multiply by 3 over x.
\[ 0.0498, \; 0.1494, \; 0.2240, \; 0.2240, \; 0.1680 \]
Add them
Why: the Poisson answer.
\[ P_{P}(x \le 4) \approx 0.8153 \]
Binomial terms for 0 to 4
Why: each computed in node from the binomial formula.
\[ 0.0476, \; 0.1471, \; 0.2252, \; 0.2275, \; 0.1706 \]
Add them
Why: the exact answer.
\[ P_{B}(x \le 4) \approx 0.8179 \]
In context: about 82 percent of strings have four or fewer faulty bulbs, so nearly one string in five has five or more. The two methods agree to about a quarter of a percentage point, closer than either single term, because small errors in the individual terms partly cancel in the sum.
Look at the terms side by side: the Poisson puts slightly more probability at 0 and at the far tail, and slightly less near the middle. That is its larger variance, np against npq, showing up bar by bar.
Check the difference
Why: subtract; it should be small next to either answer.
\[ 0.8179 - 0.8153 = 0.0026 \]
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, p. 280 — Exercise 119
Counterexample
Someone argues: the binomial B(20, 0.4) has n equal to 20, which is large enough, so a Poisson with the same mean, 8, will approximate it. Here are the two distributions bar by bar, both computed from their formulas.
Figure (svg): Paired bars for x equals 0 to 18: blue for the binomial with n equals 20 and p equals 0.4, orange for the Poisson with mu equals 8. The blue bars form a tall narrow hump around 8, reaching about 0.18; the orange bars are lower and wider, spreading further to both sides.
Discussion prompt
Use the figure and the variances to explain why the approximation fails here, and say which condition is broken.
The value of n is not the problem. Make n a million with p still 0.4 and the ratio of the variances is still q, 0.6. The Poisson describes rare events, and a success that happens forty percent of the time is not rare.
Contrast the seismic figure, where the bars matched to two decimal places. The difference is entirely in p: 0.0102 there, 0.4 here.
Discrimination
For each binomial, decide whether the Poisson approximation is appropriate by the book's conditions, n at least 20 and p at most 0.05.
Sort into buckets
Sort each binomial.
The two items that catch people are the thousand voters and the ten parts. A thousand trials sounds decisive, but with p equal to 0.2 the variance ratio q is 0.8 and the Poisson's spread is badly off. Ten parts has a wonderfully small p but too few trials for the conditions; with only ten trials the binomial is easy to compute directly anyway.
The first three are Exercises 123, 124 and 125 of the homework. The book asks for a Poisson in each, and each passes both conditions.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, p. 280 — Exercises 123-125
Comparison
Section 4.6 closes the chapter's list of named distributions. Fill in the blanks to set the Poisson against the three families that count trials.
Comparison matrix
| family | X counts | values of X | parameters | mean |
|---|---|---|---|---|
| binomial | successes in n trials | 0, 1, …, n | n, p | np |
| geometric | trials until the first success | 1, 2, 3, … | p | 1/p |
| hypergeometric | group-of-interest items in a sample, without replacement | 0 up to n or r | r, b, n | nr/(r + b) |
| Poisson | occurrences in an interval | 0, 1, 2, … | μ | μ |
Read the columns. Three families count something about trials and need at least one probability of success. The Poisson has no trials and no p; its single parameter is the average count itself.
Look at the mean column too. Every other family's mean is built from its parameters by a formula: np, one over p, nr over r plus b. The Poisson's mean is not computed from anything; it is the parameter. That is why getting μ from the stated rate is the whole of the Poisson setup.
Two families have no upper limit on X: the geometric, because you might wait any number of trials, and the Poisson, because events keep coming. Those are the two where a more-than question always needs a complement.
Deciding which family applies is the first move in every problem of this chapter. Ask: are there trials? Fixed in number? With replacement? Or is it a count in a stretch of time or space?
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 review, p. 263 — the formula review for Sections 4.3-4.6
Section
Part 6
Pattern
Every problem in this section ran through the same five moves. The formula is only the fourth.
Figure (svg): A flow diagram of five boxes. Name the interval and define X. Rescale the rate to get mu. Translate the words into an event on x. Exact values from the formula, or a complement for anything unbounded. Interpret in context with sigma equals root mu.
\[ \mu = \text{rate} \times \text{length}, \quad P(x) = \frac{\mu^{x} e^{-\mu}}{x!}, \quad \sigma = \sqrt{\mu} \]
Notice how little of the list is about arithmetic. The Poisson formula is short, and the chain of multiplications by μ over x makes several terms quick to produce. The judgement is in deciding what to count, over what stretch, and which counts the words include.
The first three moves are where the errors live, and each has a free check: the independence question, the direction and units checks, and saying the first and last count aloud. When the conditions n at least 20 and p at most 0.05 hold, a binomial can be sent through the same five moves with μ equal to np.
Ranking
An electronics store expects ten returns per day on average. The manager wants the probability of fewer than eight returns on a given day. Put the moves in the order you would carry them out.
Put in order
Order the steps for the returns problem.
Why: The interval and X come first, because they decide which rate is relevant. The mean depends on the interval. The event must be pinned down before any computing, or you do not know which terms to add. The interpretation comes last and is what the manager asked for.
Two steps are forced into place. The rescaling cannot happen before X names its interval, and the terms cannot be chosen before the words are translated. Here the rescaling is trivial, the rate is already per day, but you still have to notice that before moving on.
The translation step is where Try It 4.27 is usually lost: fewer than eight stops at seven, so there are eight terms, 0 to 7, not nine.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, p. 251 — Try It 4.27
Check
One rate, a different interval. Do the direction check before you pick.
Check your understanding
A help desk logs 18 tickets per 8-hour shift, at a steady rate. What is μ for a 20-minute interval?
Answer: A
Why: The rate is 18 over 480 minutes, which is 0.0375 per minute, and 20 minutes of that is 0.75. Equivalently, 2.25 per hour times one third of an hour. The interval is much shorter than a shift, so the answer must be much smaller than 18.
Every wrong answer here is a real, common slip, and all three are caught by one sentence: twenty minutes is a small part of a shift, so μ must be a small part of 18.
Writing the rate with units, 18 tickets per 480 minutes, and then multiplying by 20 minutes, makes the units cancel to leave tickets. That is the most reliable route, because it works the same way whether the new interval is shorter or longer.
The trap in the six option is subtle: twenty minutes is a third of something, but of an hour, not of the shift. Converting everything to minutes first, 480 and 20, removes that ambiguity entirely.
And notice this help desk has the same μ as Leah's quarter hour. The two problems are now the same distribution, so a question about its probabilities has the same answers as hers.
Check
Now the translation step. Decide which bars are in the event before you compute.
Check your understanding
A fertile cat produces an average of three litters per year, with X ~ P(3). What is the probability of at least two litters in one year?
Answer: A
Why: At least two means two or more, an unbounded set, so subtract the other side: zero and one litters, with probabilities about 0.0498 and 0.1494. One minus their sum, 0.1991, is about 0.8009.
This is Exercise 122 of the homework. The chain gives the bars quickly: e to the minus 3 is about 0.0498, and times 3 gives 0.1494 for one litter.
The wrong options each differ from the right one by a single bar. At least two and more than two differ by the bar at two, 0.2240, which is one of the two largest bars when μ is 3. That is why the translation step deserves as much care as the arithmetic.
In context: about four fertile cats in five have two or more litters in a year.
A quick sanity check is available before any arithmetic. With an average of three, two or more should be the usual outcome, so the answer should be well above one half. That alone rules out the exact-two option and makes the more-than-two option suspicious.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, p. 280 — Exercise 122
Check
Finally, the approximation and the standard deviation together.
Check your understanding
X ~ B(500, 0.004). Using the Poisson approximation, what are the mean and the standard deviation of X?
Answer: A
Why: The conditions hold, with 500 trials and a success probability well under 0.05. The approximating mean is n times p, which is 2, and a Poisson's standard deviation is the square root of its mean, about 1.41. The exact binomial value is about 1.411, almost identical.
The binomial answer, the square root of n times p times q, is the square root of 1.992, about 1.411. The Poisson's 1.414 is a close match because q is 0.996, almost exactly one. That closeness of q to one is the whole reason the approximation works.
If you chose σ equal to 2, revisit the trap in Part 4. The mean and the variance share a value; the standard deviation does not.
In context: this binomial counts about two successes in 500 trials on average, typically give or take about one and a half.
If you chose μ equal to 0.004, you used the chance for one trial as the count for all of them. Every Poisson approximation starts by multiplying: the average number of successes in all n trials is what becomes μ.
Explain it to yourself
A binomial needs two parameters, n and p; a Poisson needs only one. Explaining why is a good test of whether the section has landed.
Discussion prompt
In a few sentences, explain why a Poisson distribution is completely determined by μ, and why the value of μ depends on the interval you choose. Use Leah's calls as your example.
A strong answer separates the rate, which belongs to the process and never changes, from μ, which belongs to the interval and does. Most errors in this section come from blurring those two.
A good extension: say what a second parameter would have to describe if the model needed one. It would have to be a spread separate from the mean, and that is exactly what real data with too much variance demand, as the burglary example showed.
Exit ticket
Homework Exercise 118. The maternity ward at a hospital in Manila is one of the busiest in the world, with an average of 60 births per day. Let X be the number of births in an hour.
Discussion prompt
Find μ and σ for one hour. Then find the probability of exactly three births in an hour, at most three, and more than five, and say in a sentence whether more than five births in an hour is unusual.
The rescaling is the first mark: a day is 24 hours, and births happen around the clock, so μ is 60 over 24. The second is the translation: more than five is one minus the sum from 0 to 5, six terms, with P(4) about 0.1336 and P(5) about 0.0668 added to the four shown.
The last sentence earns the rest. Four percent of hours is rare for one hour, but a day has 24 of them, so the ward should expect a busy hour most days. Both halves need the number to back them.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6 homework, pp. 279-280 — Exercise 118
Recap
A Poisson variable counts occurrences in a fixed interval of time or space, when the events happen at a known average rate and independently of the time since the last one. It has no fixed number of trials and no upper limit.
| idea | formula | read it as |
|---|---|---|
| notation | X ~ P(μ) | a Poisson count with average μ in the interval of interest |
| rescaling | μ = rate × length | a shorter interval, a proportionally smaller μ |
| exact value | P(x) = μˣ e^(−μ) / x! | the chance of exactly x occurrences |
| ratio rule | P(x) = P(x − 1) × μ/x | bars rise while x < μ, fall after |
| complement | P(x > k) = 1 − P(x ≤ k) | the only way to handle the infinite tail |
| spread | σ² = μ, σ = √μ | one number fixes centre and spread |
| approximation | B(n, p) ≈ P(np) | when n ≥ 20 and p ≤ 0.05 |
The work is mostly in the setup: name the interval, rescale the rate onto it, and translate the words into a set of counts before computing anything. The formula then does the rest, and σ, the square root of μ, tells you whether an answer describes an ordinary interval or an unusual one.
The Introductory Business Statistics text covers the same distribution in its own §4.4, with business contexts, if you want more practice in a different setting. Chapter 5 turns from counts to continuous measurements, and Chapter 6 shows the bell shape you saw emerging at large μ in its own right.
OpenStax Introductory Statistics 2e, §4.6 Poisson Distribution §4.6, pp. 250-254 — the whole section
OpenStax Introductory Business Statistics 2e, §4.4 Poisson Distribution §4.4 Poisson Distribution — the same distribution in business contexts
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.