The chapter's second named density and the first that is not flat. The exponential distribution is concerned with the amount of time until some specific event occurs, and its shape is a declining curve: there are fewer large values and more small ones. It has a single parameter, the decay parameter m, which is the reciprocal of the mean, and unusually its standard deviation equals its mean. Probabilities come from a closed form rather than a table, with the area to the left given by one minus e to the minus m x and the area to the right by e to the minus m x. Its defining property is memorylessness: the probability of waiting a further t, given that you have already waited r, equals the unconditional probability of waiting t. The lesson closes on the equivalence between exponential gaps between events and Poisson counts of events.
Subject: Statistics · 70 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · Chapter 5 — Continuous Random Variables
Waiting times: many short, a few long, and a curve with no memory
Objectives
Section 5.2's uniform distribution was flat: every value in its range was equally likely. Most waiting times are nothing like that. Short waits are common and long waits are rare, and the curve that describes them falls steadily from its highest point at zero. This section is about that curve.
One number, m, controls everything, so most of the work is getting m right and deciding which side of a point the question is asking about. The formulas themselves are short; the traps are in reading the question, and you will meet each one on its own slide.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 302-312 — the section this deck covers
Warm-up
Section 4.6 gave you the Poisson distribution for counts of events in an interval. Recall its formula on the simplest possible question, the chance of no events at all.
Discussion prompt
Leah's calls arrive at an average of 0.75 per quarter hour, independently. What is the probability that she gets no calls at all in a particular quarter hour?
Now read the same fact another way. No call in the next quarter hour means the wait for the next call is longer than a quarter hour. So you have just computed a probability about a waiting time, and it came out as e to the power of minus something.
Keep the number 0.4724 in mind. It will turn up again in Part 4, in a question about a post office that looks entirely different, and Part 5 will explain why the two are the same calculation.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §4.6, pp. 250-254 — the Poisson formula
Section
Part 1
Concept
The book's examples of exponential variables are all amounts of time until something happens: until an earthquake, until a car battery dies, until a customer is served. Here are 200 simulated visits to a postal clerk whose visits average four minutes.
Figure (svg): A histogram of 200 simulated service times in bins two minutes wide, from 0 to 24 minutes. The first bar, 0 to 2 minutes, is the tallest with 88 customers, and each bar after it is shorter. A smooth falling curve, 0.25 times e to the minus 0.25 x, runs along the tops of the bars.
Look at the shape. The tallest bar is the first one: 88 of the 200 visits are over within two minutes. Every bar after it is shorter, and only 10 visits run to twelve minutes or more. The simulated visits average 3.98 minutes, close to the four the clerk averages in the long run.
The book puts this in one sentence: there are fewer large values and more small values. The same shape shows up for the amount a shopper spends on one supermarket trip, the length of long-distance business calls, and even the loose change in your pocket. Few people carry a lot; many carry a little.
exponential random variable — A continuous random variable, usually an amount of time until an event, whose density is highest at zero and falls steadily, so that small values are common and large values are rare.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 302 — the opening examples
Concept
The uniform distribution needed two numbers, its two endpoints. The exponential needs only one, called the decay parameter and written m. It comes straight from the mean.
\[ m = \frac{1}{\mu} \]
So a clerk averaging four minutes has m equal to one quarter, and a computer part lasting ten years on average has m equal to one tenth. The units of m are per minute, or per year: it is a rate. A large m means events come quickly and the waits are short.
\[ X \sim \text{Exp}(m), \qquad f(x) = m\,e^{-mx}, \quad x \ge 0 \]
The density starts at height m, because e to the zero is one, and every increase in x multiplies the height by the same factor. That steady shrinking is what decay means here.
\[ \sigma = \mu \]
The last line is unusual and worth remembering: the standard deviation equals the mean. A four-minute average comes with a four-minute standard deviation, a very wide spread, which is the long right tail showing up in a single number.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Example 5.7, the notation and the density
Notation
Every symbol in the density has a job. Hover over each piece before moving on, because the next twenty slides use all of them.
Read the formula as a sentence: the density at x is the starting height m, shrunk by the factor e to the minus m for every unit of x. That is all an exponential is.
Annotate
On: \( X \sim \text{Exp}(m), \quad f(x) = m\,e^{-mx}, \quad m = \frac{1}{\mu} \)
A good habit: whenever you write the distribution, say its mean out loud as well. Exp of 0.25 is a mean of four; Exp of 0.1 is a mean of ten. Saying both catches the commonest slip in the section, putting the mean where m belongs.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — X ~ Exp(m) and f(x) = m e^(−mx)
Worked example
Let X be the amount of time, in minutes, a postal clerk spends with a customer. The time is exponential with an average of four minutes. Find m and σ, write the distribution and the density, and find the height of the curve at five minutes.
Everything in this example follows from one given number, the mean. The plan is to turn the mean into m, then let m fill in the rest: the spread, the name of the distribution and the formula for the curve. Watch how often the same 0.25 appears.
Read off the mean
Why: time is measured, so X is continuous, and the average is given.
\[ \mu = 4 \text{ minutes} \]
Take the reciprocal for m
Why: the decay parameter is one over the mean.
\[ m = \frac{1}{4} = 0.25 \text{ per minute} \]
The standard deviation equals the mean
Why: a property of every exponential.
\[ \sigma = \mu = 4 \text{ minutes} \]
Write the distribution and density
Why: name it by m, and put m in both places.
\[ X \sim \text{Exp}(0.25), \qquad f(x) = 0.25\,e^{-0.25x} \]
Substitute x = 5 into the density
Why: multiply the exponent first.
\[ f(5) = 0.25\,e^{(-0.25)(5)} = 0.25\,e^{-1.25} \]
Evaluate
Why: e to the minus 1.25 is 0.2865.
\[ f(5) = 0.25(0.2865) \approx 0.072 \]
Figure (svg): The curve f of x equals 0.25 e to the minus 0.25 x, for x from 0 to 20 minutes. A dot at x equals 0 marks the height 0.25, labelled f of 0 equals m. A dot at x equals 5 marks the height 0.072.
The value 0.072 is a height, not a probability. For a continuous variable the probability of exactly five minutes is zero; probabilities are areas, and Part 2 is about finding them.
Check the height at zero
Why: the book's own check: at x = 0 the density must equal m.
\[ f(0) = 0.25\,e^{0} = 0.25(1) = 0.25 = m \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Example 5.7 and Figure 5.22
Fill the middle
Spouses shopping for anniversary cards spend an exponential amount of time, with an average of eight minutes. Fill in the decay parameter and the density. Then look at the figure to see what doubling the mean does to the curve.
Fill in the blanks
m = \frac0.1250.125 = 0.125, \qquad f(x) = ___\,e^___}x}
Why: One over eight is 0.125, and that one number goes in front of the exponential and in the exponent. The distribution is written X ~ Exp(0.125), and its standard deviation is 8 minutes, the same as the mean.
Figure (svg): Two exponential curves on the same axes for x from 0 to 30 minutes. The blue curve, m equals 0.25, starts at height 0.25 and falls fast. The orange curve, m equals 0.125, starts at half that height and falls slowly. Dashed lines mark the means, 4 and 8.
Compare the two curves. The card shoppers' curve starts at 0.125, half the clerk's height, and falls half as fast. Both enclose an area of one, so the lower curve has to stretch further to the right to make up the difference, and that stretch is the longer mean.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Try It 5.7
Tweak it
Here is the density with m on a slider. Before you move it, predict what a larger m does to the height at zero and to how quickly the curve reaches the axis.
Parameter explorer
Slide m from 0.1 to 1. Read the height where the curve meets the vertical axis, and find roughly where the curve has fallen to a third of that height. How do both depend on m?
\[ f(x) = {m}\,e^{-{m}x}, \quad \mu = 1/{m} \]
What to notice: the starting height and the mean move in opposite directions. At m equal to 1 the curve starts at 1 and has dropped to about 0.37 by x equal to 1. At m equal to 0.1 it starts at 0.1 and takes until x equal to 10 to fall by the same fraction.
That is the whole content of m equal to one over μ, seen as a picture. Whatever m is, the curve has fallen to about 37 percent of its starting height at the mean. That fixed fraction is one over e, and it will reappear as a probability in Part 2.
Trap
For the clerk, with a mean of four minutes, a first attempt writes:
\[ X \sim \text{Exp}(4) \]
\[ f(x) = 4\,e^{-4x} \]
Wrong. That curve has mean one quarter of a minute.
The parameter is the reciprocal of the mean. A curve starting at height 4 would be over almost at once: its visits would average fifteen seconds.
\[ m = \frac{1}{4} = 0.25 \]
\[ f(x) = 0.25\,e^{-0.25x} \]
The quick check is the start height: it must be small when the waits are long. If the mean is bigger than one and your curve starts above one, you have swapped m and μ.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Example 5.7
Sorting
For each quantity, ask whether small values are common and large ones rare, with the density highest right at zero. If the most likely values sit somewhere in the middle, or every value in a range is equally likely, it is not exponential.
Sort into buckets
Sort each quantity by the model that fits it best.
The bus is the instructive one. A timetable gives a flat wait, as in Section 5.2. If instead buses turned up at random, independently, at the same average rate, the wait would be exponential. The difference is whether an arrival is scheduled or random.
The change in your pocket is the surprising member of the list. It is not a time at all, yet it has the same shape: lots of people carry a few coins and very few carry a lot. The model cares about the shape, not about whether the variable is measured in minutes.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 302 — the book's list of exponential examples
Prediction
Two exponential waiting times have decay parameters 0.1 and 2, both per minute. Commit to an answer before you reason it through.
Think about what m measures before you choose. The decay parameter is a rate: how many events per minute, or how fast the curve falls. Ask yourself what a fast rate means for how long you wait.
Predict first
Which of the two has the longer average wait?
Correct: The one with m equal to 0.1
Why: The mean is one over m. One over 0.1 is 10 minutes and one over 2 is half a minute, so the smaller decay parameter goes with the much longer wait. A large m means rapid decay: the curve is used up almost at once.
\[ \mu = \frac{1}{0.1} = 10 \text{ minutes}, \qquad \mu = \frac{1}{2} = 0.5 \text{ minutes} \]
The instinct that a bigger parameter means a bigger answer is exactly backwards here, which is why this question catches people. Think of m as a rate: two events a minute arrive far faster than one every ten minutes, so the wait for the next is far shorter.
Keep the picture from the slider in mind too: a large m makes a tall curve that hits the axis almost at once, and a curve that is finished quickly describes waits that are over quickly.
Section
Part 2
Concept
A probability for a continuous variable is an area under the density. For the uniform the areas were rectangles. The exponential's areas are under a curve, and yet they have a short formula, so no table is needed.
\[ P(X < x) = 1 - e^{-mx} \]
This is the cumulative distribution function: it gives the area to the left of x. For the clerk and five minutes, substitute m equal to 0.25 and x equal to 5.
\[ P(X < 5) = 1 - e^{(-0.25)(5)} = 1 - e^{-1.25} \]
\[ P(X < 5) = 1 - 0.2865 = 0.7135 \]
Figure (svg): The clerk's curve with the region under it from 0 to 5 shaded blue and labelled 0.7135. The unshaded tail to the right of 5 is labelled 0.2865.
About 71 percent of visits are over within five minutes. Notice how that fits the histogram in Part 1, where the first few bars held most of the customers. Whether the inequality is strict or not makes no difference: for a continuous variable, one single point has no area.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Example 5.8 and the cumulative distribution function
Intuition
The book states the formula rather than deriving it, because deriving it takes calculus. You can still test it with nothing more than rectangles, which is what the area under a curve means anyway.
Figure (svg): The clerk's curve from 0 to 10 minutes with eight thin rectangles, each half a minute wide, packed under it from 0 to 4. Each rectangle's height is the curve's height at its middle. Their total area is 0.6317, beside the formula's 0.6321.
Cut the region from 0 to 4 into eight strips, each half a minute wide, and give each strip the curve's height at its middle. The strips' areas are width times height:
\[ 0.5(0.117) + 0.5(0.104) + 0.5(0.091) + \cdots + 0.5(0.049) \]
\[ \approx 0.6317 \]
The formula says the area to the left of 4 is one minus e to the minus 1:
\[ 1 - e^{(-0.25)(4)} = 1 - e^{-1} = 0.6321 \]
Eight crude strips agree with the formula to the third decimal place, and thinner strips would close the gap entirely. So the formula is not magic; it is the exact total of infinitely many thin strips. It also passes the other test every area formula must pass: as x grows without bound, e to the minus m x shrinks to zero and the area approaches one.
Concept
Questions about more than a value ask for the area to the right. The total area is one, so subtract the left area from one and simplify.
\[ P(X > x) = 1 - P(X < x) \]
\[ P(X > x) = 1 - \left(1 - e^{-mx}\right) \]
\[ P(X > x) = e^{-mx} \]
The two ones cancel and the right tail is a single exponential. That is the reverse of the discrete distributions, where more than questions needed a complement and extra work. Here the right tail is the easy one.
Try it at the mean itself, x equal to one over m. The exponent becomes minus one whatever m is:
\[ P(X > \mu) = e^{-m \cdot (1/m)} = e^{-1} \approx 0.3679 \]
Only about 37 percent of values exceed the mean, for every exponential. That is the long tail again: a few very large values pull the mean up, so most values fall below it.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 306 — P(X > x) = 1 − (1 − e^(−mx)) = e^(−mx)
Notation
These two lines do all the probability work in the section. Read them together, because choosing the wrong one is the section's most common error.
Annotate
On: \( P(X < x) = 1 - e^{-mx} \qquad P(X > x) = e^{-mx} \)
An interval needs no third formula. The area between a and b is the left area at b minus the left area at a, and the ones cancel again, leaving the right tail at a minus the right tail at b. The next two worked examples use both forms.
\[ P(a < X < b) = e^{-ma} - e^{-mb} \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 303-307 — Examples 5.8 and 5.9
Worked example
Using Example 5.7's clerk, find the probability that a clerk spends four to five minutes with a randomly selected customer.
An interval is a strip under the curve. The formula gives left areas, so the plan is to find the left area up to five, the left area up to four, and take the difference: the strip is what you gain by moving the right edge from four to five.
\[ X \sim \text{Exp}(0.25), \qquad P(4 < X < 5) = ? \]
Left area at 5
Why: one minus e to the minus 0.25 times 5.
\[ P(X < 5) = 1 - e^{-1.25} = 0.7135 \]
Left area at 4
Why: one minus e to the minus 0.25 times 4.
\[ P(X < 4) = 1 - e^{-1} = 0.6321 \]
Subtract the smaller from the larger
Why: the strip is what the longer interval has and the shorter one lacks.
\[ P(4 < X < 5) = 0.7135 - 0.6321 = 0.0814 \]
Figure (svg): The clerk's curve with a thin vertical strip under it between x equals 4 and x equals 5 shaded yellow, labelled 0.0814. Everything to the left of 4 is shaded faint blue.
In context: about 8 percent of customers spend between four and five minutes with the clerk. The strip is one minute wide, yet holds only 8 percent, because by four minutes the curve has already fallen to less than half its starting height.
Check with the right-tail form
Why: the ones cancel, so the difference of the two right tails must give the same strip. Exactly it is 0.08137, which rounds to the book's 0.0814.
\[ e^{-1} - e^{-1.25} = 0.36788 - 0.28650 = 0.08137 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 303-304 — Example 5.8(a) and Figure 5.23
Worked example
On average a certain computer part lasts ten years, and its lifetime is exponential. What is the probability that a part lasts more than seven years?
The question says more than, so this is a right tail, the one area with a single-term formula. The only preparation is m, and the only decision is the side; get those two right and the rest is one press of the e key.
Define X and find m
Why: X is the lifetime in years; m is one over the mean.
\[ \mu = 10, \quad m = \frac{1}{10} = 0.1 \]
Choose the form
Why: more than seven is a right tail, so use the single exponential.
\[ P(X > 7) = e^{-mx} \]
Substitute
Why: multiply m by 7 in the exponent.
\[ P(X > 7) = e^{(-0.1)(7)} = e^{-0.7} \]
Evaluate
Why: a calculator, or the e to the x key.
\[ P(X > 7) \approx 0.4966 \]
Figure (svg): The curve 0.1 e to the minus 0.1 x for x from 0 to 40 years. The region to the right of x equals 7 is shaded yellow and labelled 0.4966; the region from 0 to 7 is faint blue, labelled 0.5034.
In context: just under half of these parts last more than seven years. That may feel low when the average is ten, but the average is inflated by the few parts that last twenty or thirty years; the typical part fails sooner than the average suggests.
Check against the complement
Why: the left area and the right tail must add to one.
\[ P(X < 7) = 1 - e^{-0.7} = 0.5034, \quad 0.5034 + 0.4966 = 1 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 305-306 — Example 5.9(a) and Figure 5.25
Worked example
For the same computer part, what is the probability that it lasts between nine and eleven years?
The same method as the clerk's strip, now on a curve that decays ten times more slowly. Before computing, predict the size: the strip is two years wide on a curve whose height near ten is under 0.04, so expect something under 0.08.
\[ X \sim \text{Exp}(0.1), \qquad P(9 < X < 11) = ? \]
Left area at 11
Why: the larger endpoint first.
\[ P(X < 11) = 1 - e^{(-0.1)(11)} = 1 - e^{-1.1} = 0.6671 \]
Left area at 9
Why: the smaller endpoint.
\[ P(X < 9) = 1 - e^{(-0.1)(9)} = 1 - e^{-0.9} = 0.5934 \]
Subtract
Why: larger minus smaller.
\[ P(9 < X < 11) = 0.6671 - 0.5934 = 0.0737 \]
Figure (svg): The same computer-part curve with the strip between 9 and 11 years shaded yellow and labelled 0.0737. A dashed line marks the mean at 10.
In context: only about 7 percent of parts fail within a year either side of the ten-year average. For a bell-shaped variable the region around the mean holds a lot of probability; for an exponential the mean is not where the values crowd. They crowd near zero.
Check with the right tails
Why: the book's calculator line computes e to the minus 0.9 minus e to the minus 1.1 directly.
\[ e^{-0.9} - e^{-1.1} = 0.40657 - 0.33287 = 0.07370 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 307 — Example 5.9(d) and Figure 5.27
Worked example
The length of a phone call, in minutes, is exponential with decay parameter one twelfth. Someone arrives at a public telephone just before you. Find m, μ and σ, and the probability that you have to wait more than five minutes.
This time the book hands you m instead of the mean, so the first job is to run the reciprocal the other way. Then translate the story: you wait for as long as the other person's call lasts, so your wait is their call length.
Read off m
Why: this time the decay parameter is given, not the mean.
\[ m = \frac{1}{12} \]
Invert it for the mean
Why: the mean is one over m.
\[ \mu = \frac{1}{1/12} = 12 \text{ minutes} \]
The standard deviation
Why: equal to the mean.
\[ \sigma = 12 \text{ minutes} \]
Your wait is their call
Why: you wait more than five minutes exactly when their call lasts more than five: a right tail.
\[ P(X > 5) = e^{-(1/12)(5)} = e^{-5/12} \]
Evaluate
Why: five twelfths is about 0.4167.
\[ P(X > 5) = e^{-0.4167} \approx 0.6592 \]
Figure (svg): The curve one twelfth times e to the minus x over 12, for x from 0 to 48 minutes. The region beyond x equals 5 is shaded yellow and labelled 0.6592.
In context: about two times in three you will wait more than five minutes. That fits a mean call of twelve minutes, but notice it is well below the three quarters or so you might guess, because many calls are short.
Check with the left area
Why: the calls under five minutes must make up the rest.
\[ 1 - e^{-5/12} = 0.3408, \quad 0.3408 + 0.6592 = 1 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 307 — Example 5.10
Fill the middle
The distance, in miles, that people are willing to commute to work is exponential with decay parameter one twentieth. Fill in the mean, the standard deviation and the chance a person is willing to commute more than 25 miles.
Fill in the blanks
\mu = 20, \quad \sigma = 20, \quad P(X > 25) = e^0.2865 = ___
Why: The mean is one over one twentieth, which is 20 miles, and the standard deviation equals it. More than 25 miles is a right tail, e to the minus 25 over 20, which is e to the minus 1.25, about 0.2865.
In context: a little over a quarter of people would commute more than 25 miles. Notice the answer, 0.2865, also appeared in Part 2 as the clerk's tail beyond five minutes. That is no coincidence: in both cases the exponent is minus 1.25, because 25 over 20 equals 5 over 4.
That observation is useful: an exponential probability depends only on x divided by the mean. Five minutes on a four-minute clerk and 25 miles on a 20-mile commute are both one and a quarter means, so they carry the same tail.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 307-308 — Try It 5.10
Picture it
The commuting answer matched the clerk's because both questions sat one and a quarter means out. Here is why that always works. On the left are the right tails of three exponentials from this section, drawn against x in their own units.
Figure (svg): Two panels. Left: three falling curves, the right tails e to the minus x over the mean, for means 4, 10 and 20, plotted against x from 0 to 60; they fall at very different speeds. Right: the same three curves plotted against x divided by the mean, from 0 to 3; all three lie on one curve, with the value 0.2865 marked at 1.25.
On the left the three curves look like different animals: the clerk's tail is gone in twenty minutes, the commute's takes a hundred miles. On the right the horizontal axis measures x as a multiple of the mean, and the three curves fall on top of one another exactly.
\[ P(X > x) = e^{-mx} = e^{-x/\mu} \]
The algebra says the same: with m equal to one over μ, the exponent is minus x over μ, which depends only on how many means out you are. So one picture, e to the minus u, answers every exponential question once you convert to multiples of the mean.
That gives you three anchors worth memorising. At one mean the tail is 0.37; at two means it is 0.14; at three means it is 0.05. A waiting time beyond three times its mean happens about one time in twenty, whatever the process.
Use the anchors to check any answer before you trust it. If a question asks about a wait of two means and your answer is 0.6, something has gone wrong, most likely the side or the reciprocal.
Estimation
A light bulb's lifetime is exponential with a mean of 1,000 hours. Before computing anything, estimate.
Predict first
Roughly what fraction of bulbs last longer than their mean of 1,000 hours?
Correct: About a third
Why: The right tail at the mean is e to the minus one, about 0.37, for every exponential. The long right tail pulls the mean above most of the values, so only about 37 percent of bulbs outlast it.
\[ P(X > \mu) = e^{-1} \approx 0.3679 \]
The instinct says a half, because for a symmetric distribution the mean is in the middle. The exponential is not symmetric. A few bulbs that last five or six thousand hours drag the mean far to the right, so the typical bulb falls short of it.
This single fact is a useful sanity check on every tail you compute in this section: a point below the mean should have a right tail above 0.37, and a point beyond the mean should have one below it.
Error analysis
A student answers Example 5.9(a), the computer part with mean ten years, like this. The final number is wrong, and the error is easy to miss because the two answers are so close.
Annotate
On: \( P(X > 7) = 1 - e^{-(0.1)(7)} = 1 - 0.4966 = 0.5034 \)
The check that would have caught it is the one from the estimation slide. Seven years is below the mean of ten, so the right tail must be above 0.37; both answers pass that. Only the sketch settles it: shade the region beyond seven and ask whether it looks like slightly less than half the area, which it does.
Near the median, the two sides are nearly equal and a side error changes the answer by very little. Far from the median it is catastrophic: the chance a part lasts more than thirty years is 0.0498, and the wrong form gives 0.9502.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 305-306 — Example 5.9(a)
Discrimination
For each question about an exponential waiting time, decide which formula gives the answer in one line. Translate the words first: within, at most and fewer than are left areas; more than and at least are right tails.
Sort into buckets
Sort each question by the form that answers it directly.
The car question is the one to slow down on. Does not pass for at least 15 seconds is a double negative that means the gap is 15 seconds or longer, a right tail. It is Try It 5.11(d), and you will compute it in Part 5.
For a continuous variable at least and more than give the same area, since one single point carries no probability. That was not true for the Poisson counts in Section 4.6, where fewer than five and at most five differed by a whole bar.
Trap
Asked for the chance that a clerk's visit lasts five minutes, a first attempt uses the density:
\[ P(X = 5) = f(5) \]
\[ = 0.25\,e^{-1.25} = 0.072 \]
Wrong. A height is not an area.
For a continuous variable, the probability of one exact value is zero: a single point has no width. The density's height only says how thickly the values are packed near that point.
\[ P(X = 5) = 0 \]
A sensible question names an interval, such as between four and five minutes, and the answer is the area of that strip:
\[ P(4 < X < 5) = 0.0814 \]
The giveaway is that a height can exceed one: the police-call curve in Part 5 starts at height 4. No probability can.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 303 — Example 5.7: 0.072 is the height of the curve
Section
Part 3
Concept
A percentile question gives you an area and asks for the point. The kth percentile, written k, has the given fraction p of the area to its left. Set the left-area formula equal to p and solve for k, one move at a time.
\[ 1 - e^{-mk} = p \]
Subtract one from both sides and multiply by minus one, to get the exponential on its own:
\[ e^{-mk} = 1 - p \]
Now undo the exponential. The natural logarithm is the inverse of e to the power, so taking ln of both sides brings the exponent down:
\[ -mk = \ln(1 - p) \]
Finally divide by minus m:
\[ k = \frac{\ln(1 - p)}{-m} \]
This is the book's percentile formula. The minus sign in the denominator is not a mistake: ln of a number between 0 and 1 is negative, so dividing by a negative gives a positive time, as it must.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 304-305 — the percentile formula k = ln(1 − p)/(−m)
Worked example
For the postal clerk, with m equal to 0.25, half of all customers are finished within how long? That is the 50th percentile, the median.
This runs the left-area formula backwards: you know the area, one half, and want the point. Each line undoes one layer of the formula, the subtraction first, then the exponential, then the multiplication by m.
Set the left area equal to one half
Why: k is the median.
\[ 1 - e^{-0.25k} = 0.50 \]
Isolate the exponential
Why: subtract from one.
\[ e^{-0.25k} = 1 - 0.50 = 0.50 \]
Take natural logarithms
Why: ln undoes e.
\[ -0.25k = \ln(0.50) = -0.6931 \]
Divide by minus 0.25
Why: the negatives cancel.
\[ k = \frac{-0.6931}{-0.25} = 2.7726 \approx 2.8 \text{ minutes} \]
Figure (svg): The clerk's curve with the area from 0 to 2.77 shaded blue and labelled 0.50, and the rest shaded orange and labelled 0.50. A yellow dashed line marks the median at 2.77 and a pink dashed line marks the mean at 4.
In context: half of all customers are finished within about 2.8 minutes. The book's part (c) then asks which is larger, the mean or the median. The mean is 4 minutes and the median 2.8, so the mean is larger: the long tail pulls the balance point to the right of the halfway point.
Check by substituting back
Why: the left area at the median must be one half.
\[ 1 - e^{(-0.25)(2.7726)} = 1 - e^{-0.6931} = 1 - 0.5000 = 0.5000 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 304-305 — Example 5.8(b)-(c) and Figure 5.24
Concept
Put p equal to one half in the percentile formula, without any particular m, and something general drops out.
\[ k_{50} = \frac{\ln(0.5)}{-m} = \frac{\ln 2}{m} \]
Since one over m is the mean, this says the median is ln 2 times the mean:
\[ \text{median} = (\ln 2)\,\mu \approx 0.6931\,\mu \]
For the clerk that is 0.6931 times 4, which is 2.77 minutes, the answer on the last slide. For the computer part it is 6.93 years. Every exponential has its median at about 69 percent of its mean.
Why is the mean the larger one? The mean is a balance point, and a few very long waits sit far out on the right, each pulling hard on the balance. The median only counts: it neither knows nor cares how far out the long waits are. So a long right tail always pushes the mean past the median, and this is the clearest example of that you will meet.
It also explains the fact from Part 2 that only 37 percent of values exceed the mean. The mean sits beyond the median, so fewer than half the values can exceed it.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 305 — Example 5.8(c): the mean is larger
Worked example
For the computer part with mean ten years, eighty percent of parts last at most how long? Find the 80th percentile, k.
At most k means an area of 0.80 to the left of k. This time use the percentile formula directly rather than re-deriving it; the risk is feeding it the wrong number, so write the one minus p out in full before simplifying.
Set up the left area
Why: at most k, with 0.80 to the left.
\[ P(X < k) = 0.80 \]
Use the percentile formula
Why: p is 0.80 and m is 0.1.
\[ k = \frac{\ln(1 - 0.80)}{-0.1} \]
Simplify inside the logarithm
Why: one minus 0.80 is the right tail.
\[ k = \frac{\ln(0.20)}{-0.1} \]
Evaluate the logarithm
Why: ln 0.2 is about minus 1.6094.
\[ k = \frac{-1.6094}{-0.1} = 16.094 \]
Figure (svg): The computer-part curve with the area from 0 to 16.09 years shaded blue and labelled 0.80, and the tail beyond it shaded orange and labelled 0.20.
In context: eighty percent of these parts last at most about 16.1 years, and one in five lasts longer. Compare this with the median of 6.93 years. Going from the 50th to the 80th percentile more than doubles the time, because the curve thins out so slowly on the right.
Check with the right tail
Why: the area beyond k must be 0.20.
\[ P(X > 16.094) = e^{-1.6094} = 0.2000 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 306 — Example 5.9(c) and Figure 5.26
Trap
For the 80th percentile of the computer part, a first attempt writes:
\[ k = \frac{\ln(0.80)}{-0.1} \]
\[ k = 2.23 \text{ years} \]
Wrong. That is the 20th percentile.
The formula takes the logarithm of the right tail, one minus p, because it came from solving e to the minus m k equals one minus p.
\[ k = \frac{\ln(1 - 0.80)}{-0.1} = 16.09 \]
A sense check catches it: an 80th percentile must lie above the median, 6.93 years. An answer of 2.23 is below the median, so it cannot be right. Check any percentile against the median before you report it.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 306 — Example 5.9(c)
Concept
The book's Example 5.9(b) asks a different kind of question about the computer part: on average, how long would five parts last if each is installed the moment the last one fails? No curve is needed, only the mean.
\[ 5 \times 10 = 50 \text{ years} \]
Averages add. The expected total of five lifetimes is the sum of five expected lifetimes, and each is ten years. Here are four simulated runs of five parts each.
Figure (svg): Four horizontal runs, each a chain of five coloured boxes, one box per computer part, its length the part's simulated lifetime. The runs end at 21, 52, 35, 35 years. A dashed yellow line marks the average total, 50 years.
The four runs lasted 21, 52, 35, 35 years. None is exactly 50, and they spread widely, because each part's own standard deviation is ten years, as large as its mean. What the rule promises is the long-run average of the total, not any one total.
Be careful not to stretch this into a rule about probabilities. The total of five exponential lifetimes is not itself exponential, and the chance that five parts last more than 50 years is not e to the minus 5. The means add; the curve does not carry over. Examples 5.11(b) and Try It 5.11 use the same rule for three customers and seven cars.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 306 — Example 5.9(b)
Worked example
The number of days ahead that travellers buy their airline tickets is exponential with a mean of 15 days. Find the probability that a traveller buys a ticket fewer than ten days in advance, and how many days half of all travellers wait.
Two questions of different kinds: the first gives a point and asks for an area, the second gives an area and asks for a point. Label which is which before you start, because each runs the formula in the opposite direction.
Find m
Why: one over the mean.
\[ m = \frac{1}{15} \approx 0.0667 \text{ per day} \]
Fewer than ten days is a left area
Why: substitute x = 10.
\[ P(X < 10) = 1 - e^{-10/15} = 1 - e^{-0.6667} \]
Evaluate
Why: e to the minus two thirds is 0.5134.
\[ P(X < 10) = 1 - 0.5134 = 0.4866 \]
The median
Why: ln 2 over m, or ln 2 times the mean.
\[ k_{50} = 15\ln 2 = 15(0.6931) = 10.40 \text{ days} \]
In context: just under half of travellers buy within ten days of flying, and half buy within about ten and a half days. The two answers agree with each other, as they should: ten days is just short of the median, so its left area is just short of one half.
Check the median
Why: the left area there must be exactly one half.
\[ 1 - e^{-10.40/15} = 1 - e^{-0.6931} = 0.5000 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 305 — Try It 5.8
Worked example
Running shoes used every day last an average of 18 months, exponentially distributed. Find the chance a pair lasts more than 15 months, how long six pairs used one after the other last on average, and how long eighty percent of pairs last at most.
A right tail, with m = 1/18
Why: more than 15 months.
\[ P(X > 15) = e^{-15/18} = e^{-0.8333} = 0.4346 \]
Means add for pairs used one after another
Why: the book's Example 5.9(b) does the same with five parts and 50 years.
\[ 6 \times 18 = 108 \text{ months} \]
The 80th percentile
Why: the formula with p = 0.80 and m = 1/18.
\[ k = \frac{\ln(0.20)}{-1/18} = 18(1.6094) = 28.97 \text{ months} \]
In context: about 43 percent of pairs last more than 15 months; six pairs in a row average nine years between them; and eighty percent of pairs are worn out within about 29 months. Only the second answer uses the mean alone, because an average of a total is the total of the averages.
Notice the six pairs will not last exactly 108 months. Each pair varies a great deal, with a standard deviation of 18 months, and the total varies too; 108 is only the long-run average of the total.
Check the percentile against the median
Why: the 80th must sit above the median, 18 ln 2, and it does.
\[ k_{50} = 18(0.6931) = 12.48 < 28.97 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 306-307 — Example 5.9(b) and Try It 5.9
Ranking
For the computer part, with mean ten years, four landmarks have come up so far. Put them in order from the smallest to the largest, using the formulas rather than intuition.
Put in order
Order these values for X ~ Exp(0.1), from smallest to largest.
Why: Percentiles increase with the area to their left. The mean sits at about the 63rd percentile, since the left area there is one minus e to the minus one, about 0.63, so it falls between the median and the 80th percentile.
\[ P(X < \mu) = 1 - e^{-1} \approx 0.6321 \]
The surprise is where the mean lands. For a symmetric distribution it would be at the 50th percentile; for every exponential it is at about the 63rd. That is the same fact seen three times now: the mean exceeds the median, only 37 percent of values exceed the mean, and the mean sits at the 63rd percentile.
The gaps are uneven too. From the 25th percentile to the median is about four years; from the median to the 80th percentile is about nine. The right tail stretches.
Edge cases
The percentile formula has a denominator that is fixed and a numerator, ln of one minus p, that changes with p. Push p towards one and watch what happens to k.
Parameter explorer
The curve plots the percentile k against the fraction p, from 0 to 0.99. Slide m. What happens to k as p approaches 1, and what does m do to the whole curve?
\[ k = \frac{\ln(1 - p)}{-{m}} \]
What to notice: the curve is nearly straight for small p and then shoots upward. With m equal to 0.1, the 90th percentile is 23 years, the 99th is 46 years, and the 99.9th would be 69 years. Every extra nine multiplies the tail's rarity by ten but adds the same 23 years.
This is the boundary the exponential never reaches. The uniform of Section 5.2 had a 100th percentile, its upper endpoint. The exponential has none: however long you name, some small fraction of waits are longer. That is the long right tail, and it is why product warranties are priced with care.
Section
Part 4
Concept
Go back to the store of Example 5.11, where the time between customers is exponential with a mean of two minutes, so m equals 0.5. (The book's passage on p. 310 says Example 5.7 here, but the two-minute store is Example 5.11; Example 5.7 is the clerk.)
Suppose five minutes have passed since the last customer, an unusually long gap. It seems natural that a customer must be due, so one is more likely to arrive in the next minute. For the exponential distribution that is false. The extra time you wait does not depend on how long you have already waited.
\[ P(X > r + t \mid X > r) = P(X > t) \quad \text{for all } r \ge 0,\; t \ge 0 \]
Read it slowly. On the left: the chance of waiting more than t further, given that r has already passed. On the right: the chance of waiting more than t from a fresh start. They are equal. The past wait, r, has vanished from the answer.
memoryless property — Knowing that a waiting time has already exceeded r does not change the probabilities for the time still to come. An exponential wait is always as good as new.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — memorylessness of the exponential distribution
Concept
You can prove the property with the conditional probability rule from Chapter 3 and the right-tail formula. The event of waiting more than r plus t is inside the event of waiting more than r, so their intersection is just the first event.
\[ P(X > r + t \mid X > r) = \frac{P(X > r + t)}{P(X > r)} \]
Now write both tails as single exponentials:
\[ = \frac{e^{-m(r + t)}}{e^{-mr}} \]
Split the exponent in the numerator: e to the minus m r times e to the minus m t. The e to the minus m r cancels with the denominator, and that cancellation is the whole property.
\[ = \frac{e^{-mr}\,e^{-mt}}{e^{-mr}} = e^{-mt} = P(X > t) \]
Test it on the store with r equal to 5 and t equal to 1. The two tails are e to the minus 3 and e to the minus 2.5:
\[ \frac{e^{-3}}{e^{-2.5}} = \frac{0.0498}{0.0821} = 0.6065 = e^{-0.5} \]
So the chance that more than one further minute passes is 0.6065, exactly the chance of waiting more than one minute right after a customer arrives. The book gets the same number from the formula directly.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — P(X > 5 + 1 | X > 5) = P(X > 1) = e^(−0.5) ≈ 0.6065
Picture it
The algebra has a picture. Take the store's density and look only at the part beyond five minutes, the part still possible once five minutes have passed.
Figure (svg): Two panels. Left: the curve 0.5 e to the minus 0.5 x for the store's gaps, with the tail beyond 5 minutes shaded yellow, holding 0.0821 of the area. Right: that tail slid back to start at zero and scaled up by 1 over 0.0821, drawn as a dashed yellow curve lying exactly on top of the original curve.
On the left, the yellow tail holds 0.0821 of the area. On the right, it has been slid back so that five minutes becomes zero, and stretched upward by one over 0.0821 so that it has area one again, which is what conditioning does. The stretched tail lies exactly on the original curve.
That is the geometric meaning of memorylessness: every tail of an exponential is a scaled copy of the whole curve. No other continuous distribution has this property. Compare it with a bell curve, whose tail beyond some point is a thin sliver that looks nothing like the bell.
Notation
The formula is short but each symbol matters. Most errors come from misreading which time is r and which is t.
Here r is always something the question tells you has already happened, and t is always what the question asks about. If the question hands you a total time, you must subtract r yourself.
Annotate
On: \( P(X > r + t \mid X > r) = P(X > t) \)
In words: an old exponential waiting time behaves like a new one. The book puts it for the computer part: an old part is not any more likely to break down at any particular time than a brand new part. It stays as good as new until it suddenly breaks.
\[ P(X > 17 \mid X > 10) = P(X > 7) = e^{-0.7} = 0.4966 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — the memoryless property and the computer part
Worked example
The time a postal clerk spends with a customer is exponential with a mean of four minutes. A customer has already spent four minutes with the clerk. What is the probability they spend at least three more minutes?
The phrase already spent four minutes is the signal for memorylessness. Your job is to write the question as a conditional probability, identify the time already spent and the further time, and then let the past drop out.
Find m
Why: one over the mean.
\[ m = \frac{1}{4} = 0.25, \quad X \sim \text{Exp}(0.25) \]
Translate into a conditional probability
Why: four minutes spent, at least three more: a total of more than seven.
\[ P(X > 7 \mid X > 4) \]
Apply memorylessness
Why: r is 4 and t is 3.
\[ P(X > 7 \mid X > 4) = P(X > 3) \]
A fresh right tail
Why: e to the minus m t.
\[ P(X > 3) = e^{-0.25(3)} = e^{-0.75} \approx 0.4724 \]
Figure (svg): The clerk's curve for x from 0 to 20 minutes. The region beyond 4 is shaded faint blue, with area 0.3679, and inside it the region beyond 7 is shaded yellow, with area 0.1738. The ratio of the two is 0.4724.
In context: a customer who has already been at the counter for four minutes has about a 47 percent chance of being there at least three more. That is the same chance as a brand new customer staying more than three minutes. It is also the warm-up's answer, e to the minus 0.75, for Leah's quiet quarter hour; Part 5 explains why exponential waits and Poisson counts share their numbers.
Check with the conditional probability rule
Why: divide the two tails directly, without using the property.
\[ \frac{P(X > 7)}{P(X > 4)} = \frac{e^{-1.75}}{e^{-1}} = \frac{0.1738}{0.3679} = 0.4724 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — Example 5.12 and Figure 5.31
Worked example
A light bulb's lifetime is exponential with a mean of eight years. A bulb has already lasted 12 years. Find the probability that it lasts a total of over 19 years.
The word total is the key. Memorylessness needs the further time, t, not the total, so you will subtract the twelve years already survived before taking the tail.
Find m
Why: one over eight.
\[ m = \frac{1}{8} = 0.125 \]
Write the conditional probability
Why: the question gives a total, 19, and the time already spent, 12.
\[ P(X > 19 \mid X > 12) \]
Subtract to find t
Why: the further time is the total minus the time already spent.
\[ t = 19 - 12 = 7 \text{ years} \]
Apply memorylessness
Why: a fresh tail at seven years.
\[ P(X > 19 \mid X > 12) = P(X > 7) = e^{-7/8} \]
Evaluate
Why: seven eighths is 0.875.
\[ e^{-0.875} \approx 0.4169 \]
In context: a twelve-year-old bulb has about a 42 percent chance of reaching nineteen, exactly the chance a new bulb has of lasting seven years. The trap in this question is the word total. Using 19 instead of 7 would give e to the minus 19 over 8, only 0.0930, which is the chance a new bulb lasts nineteen years, a different question.
Check with the conditional probability rule
Why: the ratio of the two tails.
\[ \frac{e^{-19/8}}{e^{-12/8}} = \frac{0.0930}{0.2231} = 0.4169 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 311 — Try It 5.12
Trap
After five minutes with no customer at the two-minute store:
\[ P(\text{arrival in next minute}) \]
\[ \;> P(X < 1) = 0.3935? \]
Wrong. A customer is not due.
If the gaps really are exponential, the elapsed five minutes carry no information. The chance of an arrival in the next minute is the same as right after a customer came in.
\[ P(X < 6 \mid X > 5) = P(X < 1) \]
\[ = 1 - e^{-0.5} = 0.3935 \]
The feeling that a long wait makes an event due is the gambler's fallacy. It is right for scheduled events, like a bus on a timetable, and wrong for independent random ones. Which kind you have is a question about the situation, not about the maths.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — the store after five minutes
Counterexample
Section 5.2's conditioning made a probability change. Here is a clean case. A bus comes every 15 minutes on a strict timetable and you arrive at a random moment, so your wait is uniform from 0 to 15 minutes.
Figure (svg): Two panels. Left: a flat density of height one fifteenth from 0 to 15 minutes, the wait for a bus, with the part beyond 5 shaded, area two thirds. Right: after waiting 5 minutes, the remaining wait is flat from 0 to 10 at height one tenth, and the part beyond 5 more minutes is shaded, area one half.
Discussion prompt
Show that the uniform wait is not memoryless: compare the chance of waiting more than 5 minutes from a fresh start with the chance of waiting more than 5 further minutes after 5 have already passed.
This is exactly the intuition the store trap warned against, and here it is correct. The bus runs on a schedule, so the time already waited tells you something. An exponential process has no schedule, so it does not. The same question has opposite answers depending on how the events are generated.
Real world
Memorylessness is a strong assumption, and whether it holds is a fact about the world, not about the formula. The book says the exponential is often used for the lifetimes of electrical and mechanical devices. Test that claim.
Discussion prompt
For each of these, say whether an exponential lifetime with its memoryless property is a reasonable model, and why: a fuse or electronic component that fails from a random power surge; a car tyre; a person's lifetime.
The rule of thumb: the exponential suits failures caused by outside shocks that strike at random, and fails for anything that wears out or ages. Engineers check this with data; if the failure rate among old units is higher than among new ones, the exponential will be too optimistic about old equipment.
The book's own Example 5.11(f) asks the same kind of question about customers, and you will see it in Part 5.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 310 — the longevity of an electrical or mechanical device
Section
Part 5
Concept
Calls arrive at a police station at an average of four per minute, independently. You can ask two kinds of question about the same stream of calls.
Figure (svg): A four-minute timeline with 21 orange dots for calls arriving at a police station. Double-headed yellow arrows above the line mark three of the gaps between calls. Below, each minute is labelled with its count of calls: 3, 8, 5, 5.
How long until the next call? That is a waiting time, a continuous variable, and it is exponential. How many calls in a minute? That is a count, a whole number, and it is Poisson. In this simulated stretch the four minutes brought 3, 8, 5, 5 calls, averaging close to four, while the gaps between them were mostly short with a few long ones.
The book's statement: if the times between events are exponential with mean μ and independent of each other, then the number of events per unit time is Poisson with mean λ equal to one over μ. It also works the other way round.
\[ \text{gap} \sim \text{Exp}(m) \quad \Longleftrightarrow \quad \text{count per unit time} \sim \text{Poisson}(\lambda), \quad \lambda = m = \frac{1}{\mu} \]
Notice the decay parameter m and the Poisson mean λ are the same number. Four calls per minute is λ; the gaps have mean a quarter of a minute, so m is four per minute. The decay parameter is the event rate.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 311 — the relationship between the Poisson and the exponential
Concept
The link is one sentence. The wait for the next event is longer than t exactly when no event happens in the next t units of time. Those are the same event described two ways, so they have the same probability.
\[ P(\text{gap} > t) = P(\text{no events in an interval of length } t) \]
The count in an interval of length t is Poisson with mean λ times t. The Poisson chance of zero events is the formula with x equal to zero:
\[ P(\text{no events}) = \frac{(\lambda t)^{0} e^{-\lambda t}}{0!} = e^{-\lambda t} \]
And that is precisely the exponential right tail with m equal to λ:
\[ P(\text{gap} > t) = e^{-mt} = e^{-\lambda t} \]
This is why the warm-up and Example 5.12 gave the same number. Leah's quiet quarter hour had λt equal to 0.75; the clerk's three extra minutes had m t equal to 0.25 times 3, also 0.75. Both were e to the minus 0.75, 0.4724. Every exponential right tail is a Poisson probability of zero, and memorylessness is just the Poisson's independence of disjoint intervals seen from the waiting side.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 311 — the Poisson formula P(X = k) and the converse
Worked example
An average of 30 customers per hour arrive at a store, and the time between arrivals is exponential. Find the average gap, the average time for three customers to arrive after opening, and the chance that the next customer arrives within one minute.
The book states a rate, customers per hour, and the exponential wants a mean gap in minutes. So the first move is a conversion, and it is worth doing with the units written in, because every later part inherits whatever you get here.
(a) Convert the rate to an average gap
Why: 30 customers in 60 minutes.
\[ \mu = \frac{60 \text{ minutes}}{30 \text{ customers}} = 2 \text{ minutes} \]
(b) Means add
Why: three gaps of two minutes on average.
\[ 3 \times 2 = 6 \text{ minutes} \]
(c) Find m
Why: one over the mean gap, or the rate per minute.
\[ m = \frac{1}{2} = 0.5, \quad X \sim \text{Exp}(0.5) \]
Within one minute is a left area
Why: substitute x = 1.
\[ P(X < 1) = 1 - e^{(-0.5)(1)} = 1 - 0.6065 = 0.3935 \]
In context: after any customer arrives, there is about a 39 percent chance the next one comes within a minute. Part (b)'s six minutes is an average only; on any particular morning the first three customers might take two minutes or fifteen.
Check m against the stated rate
Why: 30 per hour is 0.5 per minute, the same number as m, as the Poisson link says it must be.
\[ \frac{30}{60} = 0.5 = m \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 308 — Example 5.11(a)-(c) and Figure 5.28
Worked example
For the same store, with X exponential with m equal to 0.5, find the chance that more than five minutes pass before the next customer, and the time within which seventy percent of customers arrive after the previous one.
Part (d) is a right tail and part (e) is a percentile, the two tools of Parts 2 and 3, now in a Poisson setting. Nothing new is needed; what matters is keeping m at 0.5 per minute and the times in minutes.
(d) More than five minutes is a right tail
Why: the single exponential.
\[ P(X > 5) = e^{(-0.5)(5)} = e^{-2.5} \approx 0.0821 \]
(e) Set the left area equal to 0.70
Why: k is the 70th percentile.
\[ 1 - e^{-0.5k} = 0.70 \]
Isolate the exponential
Why: subtract from one.
\[ e^{-0.5k} = 0.30 \]
Take logarithms and divide
Why: ln 0.30 is about minus 1.2040.
\[ k = \frac{\ln(0.30)}{-0.5} = \frac{-1.2040}{-0.5} = 2.41 \text{ minutes} \]
Figure (svg): The curve 0.5 e to the minus 0.5 x for gaps between customers, x from 0 to 10 minutes. The area from 0 to 2.41 is shaded blue and labelled 0.70; the tail beyond is labelled 0.30.
In context: a five-minute gap happens after only about 8 percent of customers, and seventy percent of customers arrive within about two and a half minutes of the one before. The five-minute gap from part (d) is the long wait the memorylessness slide started from.
Check the percentile
Why: the right tail beyond 2.41 minutes must be 0.30.
\[ e^{(-0.5)(2.4079)} = e^{-1.2040} = 0.3000 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 308-309 — Example 5.11(d)-(e) and Figures 5.29-5.30
Step zero
Before trusting any of the store's numbers, the book asks whether the model fits at all. This is the step that is easiest to skip and most valuable to practise.
Discussion prompt
Name the assumptions behind modelling the store's gaps as exponential, and say which of them is doubtful for real shoppers.
These are the book's own two objections. Neither makes the model useless: over a short stretch with a steady rate, and counting a group as one arrival, it can be very good. But the probability of a five-minute gap, 0.0821, is only as good as the assumption that the rate does not change across those five minutes.
The same two checks, one at a time and a steady rate, are the Poisson conditions from Section 4.6, the same assumptions seen from the other side.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 309 — Example 5.11(f)
Worked example
Cars pass on a stretch of highway at an average of five per minute, with exponential gaps. Find the average gap in seconds, how long on average for another seven cars to pass, and the chances that the next car passes within 20 seconds and that it does not pass for at least 15 seconds.
The rate is per minute but the questions are in seconds. Choose one unit and stay in it. Seconds is the natural choice here, since three of the four answers are asked for in seconds.
Convert the rate to seconds
Why: five cars in 60 seconds.
\[ \mu = \frac{60}{5} = 12 \text{ seconds}, \quad m = \frac{1}{12} \text{ per second} \]
Seven more cars
Why: means add.
\[ 7 \times 12 = 84 \text{ seconds} \]
Within 20 seconds is a left area
Why: 20 over 12 is five thirds.
\[ P(X < 20) = 1 - e^{-20/12} = 1 - 0.1889 = 0.8111 \]
At least 15 seconds is a right tail
Why: 15 over 12 is 1.25.
\[ P(X \ge 15) = e^{-15/12} = e^{-1.25} = 0.2865 \]
In context: after any car, the next passes within 20 seconds about 81 percent of the time, and a quiet gap of at least 15 seconds happens about 29 percent of the time. Working in seconds throughout keeps the units consistent; mixing a per-minute m with a time in seconds is the classic slip here.
Check with minutes instead
Why: m is 5 per minute and 15 seconds is a quarter of a minute: the same exponent.
\[ e^{-5(0.25)} = e^{-1.25} = 0.2865 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 309 — Try It 5.11
Worked example
Calls come in to a police station at an average of four per minute, with exponential gaps. We count only arrivals, ignoring time spent on the phone. Find the average time between calls, and the chance the next call comes in less than ten seconds.
Parts (a) and (b) are about the time between calls, so they live on the exponential side. Keep an eye on units: the rate is per minute, the question is in seconds, and the conversion is where this example is usually lost.
(a) Average gap
Why: four calls a minute, so a quarter of a minute apart.
\[ \mu = \frac{1}{4} = 0.25 \text{ minutes} = 15 \text{ seconds} \]
(b) Find m
Why: one over 0.25.
\[ m = \frac{1}{0.25} = 4, \quad T \sim \text{Exp}(4) \]
Convert ten seconds to minutes
Why: m is per minute, so t must be in minutes.
\[ 10 \text{ seconds} = \frac{10}{60} = \frac{1}{6} \text{ minute} \]
Less than one sixth of a minute is a left area
Why: substitute.
\[ P\left(T < \tfrac16\right) = 1 - e^{(-4)(1/6)} = 1 - e^{-0.6667} \]
Evaluate
Why: e to the minus two thirds is 0.5134.
\[ = 1 - 0.5134 = 0.4866 \]
Figure (svg): The curve 4 e to the minus 4 t for the gap between police calls, t from 0 to 1.2 minutes. The area from 0 to one sixth of a minute is shaded blue and labelled 0.4866.
In context: almost half the time, the next call arrives within ten seconds. The curve starts at height 4, well above one, which is a reminder that the density's height is not a probability.
Check with the Poisson side
Why: no call in one sixth of a minute has Poisson mean 4 times one sixth, and its chance must be the right tail.
\[ P(\text{no calls}) = e^{-4/6} = 0.5134 = 1 - 0.4866 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 311-312 — Example 5.13(a)-(b) and Figure 5.32
Worked example
Now count calls. Let X be the number of calls in one minute. Find the probability of exactly five calls in a minute, and of fewer than five.
Now the question is how many, not how long, so switch to the Poisson side. The connection says the count per minute is Poisson with mean equal to the rate, four. From there it is Section 4.6's arithmetic.
The count is Poisson
Why: mean four calls per minute, from the connection.
\[ X \sim \text{Poisson}(4) \]
(c) Exactly five
Why: the Poisson formula with x = 5.
\[ P(X = 5) = \frac{4^{5} e^{-4}}{5!} = \frac{1024(0.01832)}{120} \approx 0.1563 \]
(d) Fewer than five means at most four
Why: X is a whole number.
\[ P(X < 5) = P(X \le 4) \]
The five terms
Why: each from the formula, or by multiplying up by 4 over x.
\[ 0.0183 + 0.0733 + 0.1465 + 0.1954 + 0.1954 \]
Add them
Why: the book's calculator gives the same.
\[ P(X \le 4) = 0.6288 \]
Figure (svg): A probability histogram of calls in one minute for a Poisson distribution with mean 4, bars at 0 to 12. Bars 0 to 4 are blue, totalling 0.6288; the bar at 5 is yellow with height 0.1563; the rest are grey.
In context: exactly five calls happen in about 16 percent of minutes, and fewer than five in about 63 percent. The translation from fewer than five to at most four is the same one Section 4.6 insisted on, and here it is not optional, because a count, unlike a time, can land exactly on five.
Check the sum with the complement
Why: the probability of five or more must make up the rest.
\[ 1 - 0.6288 = 0.3712 = P(X \ge 5) \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 312 — Example 5.13(c)-(d)
Worked example
Find the probability that more than 40 calls occur in an eight-minute period.
The interval has changed from one minute to eight, and a Poisson mean belongs to its interval. So before anything else, rescale: four calls a minute becomes a mean of thirty-two calls in eight minutes.
Define Y for the new interval
Why: the count in eight minutes, not one.
\[ Y = \text{calls in eight minutes} \]
Rescale the mean
Why: four per minute for eight minutes.
\[ \lambda = 8 \times 4 = 32, \quad Y \sim \text{Poisson}(32) \]
More than 40 needs the complement
Why: the values above 40 never end.
\[ P(Y > 40) = 1 - P(Y \le 40) \]
The cumulative sum
Why: 41 terms, 0 to 40, added in node; the calculator's poissoncdf gives the same.
\[ P(Y \le 40) = 0.92934 \]
Subtract
Why: the tail.
\[ P(Y > 40) = 1 - 0.92934 = 0.0707 \]
Figure (svg): A probability histogram for a Poisson distribution with mean 32, bars from 16 to 50 forming a nearly symmetric hump peaking at about 0.07 near 32. The bars from 41 on are red and labelled total 0.0707; the rest are blue.
In context: about 7 percent of eight-minute stretches bring more than forty calls, a busy but not rare spell. The book prints 0.9294 for the sum and then 0.0707 for the tail; the exact sum is 0.92934, which rounds to 0.9293, and 0.0707 is the correct tail.
Check the size of the answer
Why: the standard deviation is the square root of 32, and 40 is about 1.4 of them above the mean, where a tail of several percent is expected.
\[ \sigma = \sqrt{32} = 5.66, \quad \frac{40 - 32}{5.66} \approx 1.41 \]
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 312 — Example 5.13(e)
Trap
For more than 40 calls in eight minutes, a first attempt keeps the one-minute distribution:
\[ X \sim \text{Poisson}(4) \]
\[ P(X > 40) \approx 0 \]
Wrong. Four is the mean for one minute.
The Poisson mean belongs to the interval. For eight minutes it is eight times the rate, and the answer changes from essentially impossible to a busy spell that happens about one time in fourteen.
\[ Y \sim \text{Poisson}(32) \]
\[ P(Y > 40) = 0.0707 \]
The exponential side does not need rescaling, because m is a rate. The Poisson side always does, because λ is a count for a stated interval. Name the interval before you write the distribution.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 312 — Example 5.13(e)
Matching
Every question about a Poisson process is either about a length of time or about a number of events. Match each question about the police station, four calls per minute, to the distribution that answers it directly.
Match the pairs
Why: Within a time is a waiting-time question, and more than a time with no call is its right tail, which is also a Poisson zero. Exactly three and at least ten are counts, and the second needs the mean rescaled to two minutes, eight calls.
The no-call question belongs to both sides at once, which is the whole connection in one example. A 30-second silence is a gap longer than half a minute, e to the minus 4 times one half; it is also zero calls in half a minute from a Poisson with mean 2, e to the minus 2. Same number, 0.1353.
When a question can be asked either way, pick the side with less arithmetic. A long silence is one exponential; the same thing as a count is one Poisson term. At least ten calls, though, is only a count question, and it needs ten terms and a complement.
Comparison
Sections 4.6, 5.2 and 5.3 give three models that are easy to mix up. Fill in the blanks.
Comparison matrix
| model | the variable | values | parameter | mean | standard deviation |
|---|---|---|---|---|---|
| uniform U(a, b) | a wait on a schedule | a to b | a, b | (a + b)/2 | √((b − a)²/12) |
| exponential Exp(m) | a wait for a random event | 0 to infinity | m | 1/m | 1/m |
| Poisson P(λ) | a count | 0, 1, 2, … | λ | λ | √λ |
Read the table by rows. The uniform and the exponential are both waiting times, but one is scheduled and one is random; that difference is the memoryless property. The exponential and the Poisson describe the same random process, one by its gaps and one by its counts.
Now read the last two columns. For the exponential, the standard deviation equals the mean. For the Poisson, the variance equals the mean. The two facts look alike and are easy to swap: σ equals μ for the exponential, σ equals the square root of λ for the Poisson.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 311 — and the formula reviews for §4.6 and §5.2
Section
Part 6
Pattern
Every problem in this section ran through the same short sequence. The formulas are the easy part; the decisions come first.
Use the diagram as a checklist. The first branch, time or count, decides everything below it, and it is the branch most often skipped. The questions in Part 5 changed sides twice within one example, so ask it again for every part of a problem.
Figure (svg): A flow diagram. First, is the question about a waiting time or a count? A count goes to a Poisson box with mean lambda times t. A waiting time goes to a box that finds m as one over the mean or as the rate, then branches three ways: left area one minus e to the minus m x, right tail e to the minus m x, percentile k equals ln of one minus p over minus m. A last box says interpret in context.
\[ m = \frac{1}{\mu}, \quad P(X < x) = 1 - e^{-mx}, \quad P(X > x) = e^{-mx}, \quad k = \frac{\ln(1 - p)}{-m} \]
Most wrong answers in this section are right calculations of the wrong quantity: the mean used as m, the left area reported for a right-tail question, a total time used where a further time was meant, or a per-minute mean applied to a longer interval. Each has a one-line check, and the pattern above puts each check where it belongs.
Check
Decide the side before you pick an option.
Start by finding m, and then say aloud whether more than 60 months is a left area or a right tail. Only then look at the options.
Check your understanding
A battery's lifetime is exponential with a mean of 40 months. What is the probability that a battery lasts more than 60 months?
Answer: A
Why: The decay parameter is one over 40, so the exponent is minus 60 over 40, which is minus 1.5. More than 60 months is a right tail, the single exponential, about 0.2231. Sixty months is past the mean, so the answer should be below 0.37, and it is.
In context: a little over one battery in five is still working at five years. Each wrong option is a real slip: the wrong side, the mean in place of m, and forgetting the exponential altogether.
Notice that you never needed m as a decimal. Writing the exponent as minus 60 over 40, the time divided by the mean, is the quickest route and the hardest to get the wrong way up.
The quick checks at work: the answer must lie between 0 and 1, which removes one option at once; and a point beyond the mean must have a tail under 0.37, which removes the left-area option.
Check
Find t before anything else.
Write the question as a conditional probability first. The time already survived is r; the total minus r is the further time t.
Check your understanding
A fuse's lifetime is exponential with a mean of 5 years. A fuse has already worked for 3 years. What is the probability it works for a total of more than 8 years?
Answer: A
Why: Given it has lasted 3 years, lasting a total of more than 8 means lasting more than 5 further years. By memorylessness that is a fresh right tail at 5, e to the minus 5 over 5, which is e to the minus one, about 0.3679.
Check the ratio method directly: the tail at 8, e to the minus 1.6, divided by the tail at 3, e to the minus 0.6, is e to the minus 1, the same answer. When the two methods agree you know t was found correctly.
In context: a fuse that has already survived three years has about a 37 percent chance of surviving five more, exactly the chance a new one has.
Whether that is believable depends on why fuses fail. If they fail from random surges, it is; if they degrade with age, the exponential flatters the old fuse, and the real chance is lower.
Check
Decide which side of the connection the question is on.
A silence of twenty minutes can be read as a long gap or as zero emails; either way works if the rate and the interval match.
Check your understanding
Emails arrive at an average of 6 per hour, independently. What is the probability that no email arrives in the next 20 minutes?
Answer: A
Why: Twenty minutes is a third of an hour, so the count has Poisson mean 6 times one third, which is 2, and the chance of zero is e to the minus 2. Equivalently, the gap is exponential with m equal to 6 per hour, and more than a third of an hour is e to the minus 2.
Both sides give e to the minus 2 because a silence of twenty minutes and zero arrivals in twenty minutes are the same event. If your two methods disagree, one of them has the mean or the rate the wrong way up.
In context: about one twenty-minute stretch in seven passes with no email at all.
If you had been asked for exactly one email in twenty minutes instead, only the count side would do: Poisson with mean 2, two times e to the minus 2, about 0.27. Silence is the one question both sides answer equally easily.
Explain it to yourself
The fact that an exponential's mean exceeds its median came up several times. Explaining it in your own words is a good test of whether the shape has landed.
Discussion prompt
For the postal clerk, the median visit is 2.8 minutes and the mean is 4. Explain in two or three sentences why the mean is larger, and what that means for the claim that the average customer takes four minutes.
A strong answer names the long right tail and the difference between counting and adding. A very strong one adds the number: the median is always ln 2, about 0.69, times the mean, whatever the value of m.
The same reasoning explains why reported average incomes or house prices sit above what a typical person earns or pays: those distributions have long right tails too.
Exit ticket
Try It 5.13. Automobile accidents in a small city follow a Poisson distribution with an average of three per week.
Discussion prompt
Find the probability of at most two accidents in a given week, and the probability that at least two weeks pass between one accident and the next. Say in a sentence what each answer means.
The first part is a count, so it is Poisson with three terms. The second is a waiting time, so switch sides: the rate, three per week, is the decay parameter, and at least two weeks is a right tail. The same answer comes from the count side as the Poisson chance of zero accidents in two weeks, with mean 6.
If you got e to the minus two thirds for the second part, you used the mean gap, one third of a week, where the rate belongs. The start-height check from Part 1 catches it: a gap averaging a third of a week needs a steep curve, m equal to 3.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, p. 312 — Try It 5.13
Recap
The exponential distribution models the time until a random event: small values are common and large ones rare. It has one parameter, the decay parameter m, which is one over the mean and also the rate of the events.
| idea | formula | read it as |
|---|---|---|
| notation | X ~ Exp(m), m = 1/μ | named by the rate, not the mean |
| density | f(x) = m e^(−mx), x ≥ 0 | starts at height m and decays |
| spread | σ = μ | a standard deviation as big as the mean |
| left area | P(X < x) = 1 − e^(−mx) | waits shorter than x |
| right tail | P(X > x) = e^(−mx) | waits longer than x; at the mean, 0.37 |
| percentile | k = ln(1 − p)/(−m) | median = (ln 2)μ ≈ 0.69μ, below the mean |
| memoryless | P(X > r + t | X > r) = P(X > t) | an old wait is as good as new |
| Poisson link | gaps Exp(λ) ⇔ counts Poisson(λt) | no event in t = a gap longer than t |
The work is mostly in the setup: decide whether the question is about a time or a count, get m the right way up, and choose the side of the curve before substituting. Then check the answer against two anchors, the median at 69 percent of the mean and the tail of 0.37 at the mean itself.
The Introductory Business Statistics text covers the same distribution in its own §5.3, with business contexts, if you want more practice in a different setting. Chapter 6 turns to the normal distribution, the symmetric bell whose mean and median coincide.
OpenStax Introductory Statistics 2e, §5.3 The Exponential Distribution §5.3, pp. 302-312 — the whole section
OpenStax Introductory Business Statistics 2e, §5.3 The Exponential Distribution §5.3 The Exponential Distribution — the same distribution in business contexts
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.