4.4 Geometric Distribution

The binomial's first characteristic relaxed. A geometric experiment repeats a trial until the first success occurs and then stops, so the number of trials is the random variable rather than the number of successes, and it takes the values one, two, three and upward with no bound. The probability of a first success on trial x is q to the power x minus one times p, with no combination factor because exactly one sequence produces it. The mean is one over p, matching the intuition that a one-in-three chance takes about three tries. The distribution is also the only discrete one that is memoryless: previous failures carry no information about what comes next, which is a genuine property of independent trials and a poor model for anything that learns.

Subject: Statistics · 64 slides · symbolic lesson

Open the interactive version of this deck

What this lesson covers

The lesson, slide by slide

1. The Geometric Distribution

Title

Statistics · Chapter 4 — Discrete Random Variables

Keep trying until the first success: how long will it take?

2. What this lesson gives you

Objectives

Section 4.3 fixed the number of trials and counted successes. This section flips that around. You keep repeating a trial until the first success, then stop, and the random variable is how many trials that took. The question changes from how many successes to how long you wait, and that one change reshapes the formula, the mean and the picture.

Almost every calculation in this lesson is one multiplication chain: some failures, then one success. If you can draw that chain for a problem, you can do the problem. The skill that takes practice is reading the words carefully enough to know how long the chain is.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 242-247 — Section 4.4

3. Before anything new: the binomial's first characteristic

Warm-up

Section 4.3 listed three characteristics of a binomial experiment. The first of them is the one this lesson throws away, so bring it back to mind before reading on.

Discussion prompt

What is the first characteristic of a binomial experiment? Then decide: you throw darts at a bullseye until you hit it. Can you write down the number of trials before you start throwing, and so can the binomial formula describe this?

Keep the other two binomial characteristics in view too: independent trials, and the same success probability on every trial. Both survive unchanged into this section. Only the stopping rule changes, and with it what the random variable counts.

Also recall the multiplication rule for independent events from Chapter 3: the chance that several independent things all happen is the product of their chances. That rule is the whole engine of the geometric formula.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — a trial is repeated until a success occurs

4. The four characteristics

Section

Part 1

5. Repeat until the first success, then stop

Concept

Throw a dart at a bullseye. Miss, and you throw again. Keep going until you hit it, and then you stop. That is a geometric experiment: a string of failures, one success, and nothing after it.

Figure (svg): Six square tiles in a row numbered 1 to 6: five red tiles marked F and a green tile marked S on trial 6. After it the word STOP and four dashed empty tiles labelled never thrown. An arrow from the green tile points to the text X = 6.

The sequence always has this shape: a run of failures, one success, and nothing after it. The random variable is where the green tile landed.

Every run looks like this picture, however long it is. The book writes the six-throw run as failure, failure, failure, failure, failure, success, stop. The number worth recording is where the success landed, here trial 6.

\[ F,\ F,\ F,\ F,\ F,\ S \;\Longrightarrow\; x = 6 \]

Two words carry the definition: until and first. Until tells you the experiment has no planned length; first tells you it ends at the earliest success, so you never see a second one.

Could it take a hundred throws? Yes, if you are unlucky. A thousand? Very unlikely, but possible. In theory the number of trials has no ceiling at all, and that is the sharpest difference from a binomial experiment, where the number of trials was written down before anything happened.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — characteristic 1 and the dart example

6. The four characteristics, and the one that is new

Concept

The book lists four characteristics. Read them against the binomial's three and you will see that two are old friends.

  1. 1. A trial is repeated until a success occurs, then you stop.
  2. 2. The repeated trials are independent of each other.
  3. 3. The probability p of a success, and q of a failure, is the same on every trial.
  4. 4. The random variable X is the number of the trial on which the first success occurs.

\[ p + q = 1, \qquad q = 1 - p \]

Characteristics 2 and 3 are the binomial's independence and constant p, word for word. Characteristic 1 replaces the fixed n. Characteristic 4 changes what is counted: not successes, but trials.

The book's die makes the third characteristic concrete. The chance of rolling a three is one sixth, and it stays one sixth no matter how many rolls have gone before. The die has no memory of earlier misses, so the chance of a miss, five sixths, is the same on every roll too.

\[ p = \tfrac16, \quad q = \tfrac56 \quad \text{on every roll} \]

geometric experiment — A sequence of independent trials, each with the same probability of success, repeated until the first success occurs. The random variable is the number of the trial on which that first success happens.

As in the binomial, success means only the outcome the sequence is waiting for. In Example 4.17 the success is losing a game, because losing is what ends play. Pick the success by asking what makes you stop, never by asking what you would like to happen.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — the four characteristics

7. Reading X ~ G(p)

Notation

The book gives the geometric distribution a short name, just as it wrote the binomial as B of n and p. Here there is only one thing inside the brackets.

Annotate

On: \( X \sim G(p), \qquad x = 1,\ 2,\ 3,\ \ldots \)

  • The random variable, in words: the number of the trial on which the first success occurs.
  • Read it as is distributed as, or has the distribution.
  • Geometric. The name comes from the probabilities, which form a geometric sequence, each one a fixed multiple of the one before.
  • The only parameter: the probability of success on each trial. There is no n, because the number of trials is what X measures.
  • The values. At least one trial is needed, and there is no largest value.

Compare with the binomial, which needed two numbers, n and p. Once you know p here, you know everything: every probability, the mean and the standard deviation all come from p alone. That is why a geometric problem usually starts with the single sentence that finds p.

\[ X \sim B(n, p) \quad \text{versus} \quad X \sim G(p) \]

Notice also that zero is not in the list of values. You cannot have a first success on trial zero, because you have to try at least once. That small fact will matter when we meet the second way of counting in Part 5.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — Notation for the Geometric

8. The die: the first three on the fifth roll

Worked example

Roll a fair die until a three comes up. The book asks for the probability that the first three appears on the fifth roll. Before any formula, build the answer from the rolls themselves.

Name the success and find p

Why: a three is what ends the sequence; one face in six.

\[ p = \frac16, \qquad q = 1 - \frac16 = \frac56 \]

Say what the event means

Why: first three on roll 5 means rolls 1 to 4 were not threes and roll 5 was.

\[ F,\ F,\ F,\ F,\ S \]

Figure (svg): Five boxes labelled roll 1 to roll 5. The first four read not 3 in red with 5/6 beneath each; the fifth reads 3 in green with 1/6 beneath it. Multiplication signs join the fractions, and the product line reads 625/7776, about 0.0804.

One factor per roll. Independence is what lets you multiply, and a constant p is why every red factor is the same 5/6.

Multiply, by independence

Why: the rolls do not affect each other, so the chance of this exact sequence is the product of its five factors.

\[ P(x = 5) = \left(\tfrac56\right)\left(\tfrac56\right)\left(\tfrac56\right)\left(\tfrac56\right)\left(\tfrac16\right) \]

Collect the four equal factors

Why: four failures in a row.

\[ = \left(\tfrac56\right)^{4}\left(\tfrac16\right) \]

Evaluate the power

Why: five to the fourth over six to the fourth.

\[ \left(\tfrac56\right)^{4} = \frac{625}{1296} \]

Multiply by the success

Why: one more factor of one sixth.

\[ \frac{625}{1296} \cdot \frac16 = \frac{625}{7776} \approx 0.0804 \]

In context: about 8 times in 100, a player waiting for a three sees the first one on exactly the fifth roll. It sounds small, but no single value of X is likely here; the probability is spread over 1, 2, 3 and on without end.

Check against the book's decimal product

Why: multiplying the five decimals directly must give the same 0.0804 the book prints.

\[ (0.8333)^{4}(0.1667) \approx 0.0804 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — the die example

9. Binomial or geometric? Ask what is fixed

Sorting

For each question, decide whether the number of trials is fixed in advance and successes are counted, or trials continue until the first success. That single question separates the two distributions every time.

Sort into buckets

Place each question in the right group.

binomial: fixed number of trials, count successes
How many heads in 20 flips of a coin?; How many of 50 sampled adults prefer saving?; How many of a player's next 80 shots score?; How many of 200 sampled people will develop cancer?
geometric: repeat until the first success, count trials
How many flips until the first head?; How many stores do you call until one has the special ink?; How many reports are read until one blames ignored instructions?; How many people do you ask until one has developed cancer?
bin
Each of these names the number of trials before the experiment starts, twenty flips or fifty adults or eighty shots, and asks how many of them succeed.
geo
Each of these has no planned length. The word until marks the stopping rule, and the answer is how many trials it took to reach the first success.

The last two items use the same one-in-67 cancer risk, the book's Examples 4.15 and 4.21. The probability is identical; the question is not. Sampling 200 people and counting is binomial; asking people one at a time until the first yes is geometric.

So p alone never tells you which distribution you have. Look for a stated sample size, which points to binomial, or for the word until, which points to geometric. If a question has both, such as how many flips until the third head, it is neither of these two.

10. Example 4.19: a student who lives nearby

Worked example

You want a student at your college who lives within five miles of you. Of the 25,000 students, 55 percent do. You contact students at random until one says yes. What is the probability that you need to contact four people?

(a) Fill in the book's sentence for X

Why: what is counted, and what ends it.

\[ X = \text{the number of students you ask until one says yes} \]

(b) List the values

Why: at least one call; in principle no fixed ceiling.

\[ x = 1,\ 2,\ 3,\ \ldots \]

(c) Read off p and q

Why: a yes is the success.

\[ p = 0.55, \qquad q = 1 - 0.55 = 0.45 \]

(d) Write the probability question

Why: exactly four contacts.

\[ P(x = 4) \]

The book stops at the question, but it is worth finishing. Needing exactly four contacts means three noes and then a yes.

Write the sequence as a product

Why: three failures, then the success.

\[ P(x = 4) = (0.45)(0.45)(0.45)(0.55) \]

Evaluate the power

Why: computed in node.

\[ (0.45)^{3} = 0.091125 \]

Multiply by p

Why: the yes on the fourth call.

\[ 0.091125 \times 0.55 \approx 0.0501 \]

Figure (svg): A stem plot for p = 0.55 over x = 1 to 8 with heights 0.5500, 0.2475, 0.1114, 0.0501, 0.0226, 0.0101, 0.0046, 0.0021. The stem at x = 4 is yellow.

More than half the time the very first student says yes. The yellow stem, needing exactly four calls, is already down to about one chance in twenty.

In context: about one time in twenty you will need exactly four calls. The book's answer to (b) ends with the total number of students, since in reality you run out of people; with 25,000 students and random contacts, the chance of reaching that end is so small that the list 1, 2, 3 and on is the right model.

Check with the common ratio

Why: each stem is 0.45 times the last, so walking from P(1) should land on the same value.

\[ 0.55 \to 0.2475 \to 0.1114 \to 0.0501 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — Example 4.19

11. Why X has no largest value

Intuition

In a binomial experiment with n equal to 10, the count of successes cannot exceed 10. Nothing like that caps a geometric variable. For every whole number you name, there is a run of failures at least that long with a positive probability.

\[ P(\text{the first } k \text{ trials all fail}) = q^{k} > 0 \quad \text{for every } k \]

For the die, the chance that the first twenty rolls all miss the three is five sixths to the twentieth power, about 0.026. For fifty rolls it is about one in nine thousand. Tiny, but never zero, so the list of values never closes.

\[ \left(\tfrac56\right)^{20} \approx 0.026, \qquad \left(\tfrac56\right)^{50} \approx 0.00011 \]

That raises a fair worry. If there are infinitely many values, how can their probabilities add up to one? They can, because the stems shrink by the same factor q each time, and a sum of shrinking multiples like that settles on a finite total. In node, adding the first 500 stems for p equal to 0.35 gives 1 to fifteen decimal places.

\[ p + qp + q^{2}p + q^{3}p + \cdots = 1 \]

So the geometric distribution is a genuine probability distribution in the Section 4.1 sense, with every stem between zero and one and a total of exactly one. It is simply a table that never ends.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — X takes on the values 1, 2, 3, and could go on indefinitely

12. Try It 4.19: printer ink, before any number

Step zero

You need a store that carries a special printer ink. Of the stores that sell printer ink, 10 percent carry this one. You call stores at random until one has it.

Discussion prompt

Before computing anything, write X in words, list its values, and give p and q. Which single word in the problem tells you this is geometric rather than binomial?

The book's question is only p and q, and that is deliberate. Once p is on the page, every other question about this problem is one short calculation: the chance of needing exactly five calls, the chance of needing more than ten, or the average number of calls.

Notice which outcome is the success. It is a store that has the ink, the thing that makes you stop dialling, and here it is the rare outcome. A small p means long waits, and you will put a number on that in Part 4: on average ten calls.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — Try It 4.19

13. Trap: reading X as the number of successes

Trap

The trap

Asked what X can be in the ink problem, a common answer:

\[ x = 0 \text{ or } 1 \]

with the reasoning that you either find the ink or you do not.

Wrong. That describes whether you succeed, not when.

It is tempting after Section 4.3, where X always counted successes.

The fix

A geometric experiment always ends with exactly one success, so counting successes tells you nothing. X counts the trials it took:

\[ x = 1,\ 2,\ 3,\ \ldots \]

If you call five stores and the fifth has the ink, the answer is 5, not 1. Test your reading by saying the value in a sentence: X equals 5 means the first store with the ink was the fifth one called.

The number of successes is fixed at one. The number of trials is the random part.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — characteristic 4

14. When it is not geometric: drawing names from a box

Counterexample

Example 4.16 draws names without replacement from a committee of ten staff and six students. Suppose instead you draw names one at a time, without replacement, until you get a student.

Discussion prompt

Explain which geometric characteristic this breaks. Give the probability of a student on the first draw, and on the second draw after a staff name was drawn first.

The fix is the one the book uses for the accident reports in Example 4.18: put each item back before the next draw. With replacement, every draw faces the same 6 in 16, and the process is geometric again.

Example 4.19 sits in between. You do not call the same student twice, so strictly the draws are without replacement. But removing a few people from 25,000 changes p by a tiny amount, so treating p as constant is an excellent approximation. The problem is a small population, not the absence of replacement as such.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — Example 4.16

15. The formula

Section

Part 2

16. Build the formula from the tree

Concept

Draw the experiment as a tree. At every trial there are two branches: a success, which ends the experiment, or a failure, which sends you on to the next trial.

Figure (svg): A staircase tree. Along the top a red path runs through nodes labelled trial 1 to trial 4, each red step labelled F: q, ending in a dashed line and an ellipsis. From each node a green branch labelled S: p drops to a circle labelled x = 1, x = 2, x = 3, x = 4, and beneath them P = p, qp, q²p, q³p.

Only one path leads to each leaf: stay on the red line for x − 1 steps, then take the green branch down. Multiply along the path and you have the formula.

Each leaf is a value of X, and exactly one path reaches it. To reach x equals 4 you must fail three times along the top, then succeed.

\[ P(x = 4) = q \cdot q \cdot q \cdot p = q^{3}p \]

Nothing about 4 was special. For any value x, the path to that leaf has x minus one failures followed by one success, and independence lets you multiply along it.

\[ P(X = x) = \underbrace{q \cdot q \cdots q}_{x - 1 \text{ failures}} \cdot p \]

\[ P(X = x) = (1 - p)^{x - 1}p, \qquad x = 1,\ 2,\ 3,\ \ldots \]

Test the formula on the leaf you can check by eye. The first success on trial 1 needs no failures at all, and q to the power zero is 1, so the formula gives p. That is exactly the chance of succeeding at once.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — Case I: the geometric pdf

17. Decoding q to the x minus one, times p

Notation

Every symbol in the formula corresponds to a piece of the sequence of trials. Read it piece by piece.

Annotate

On: \( P(X = x) = (1 - p)^{x - 1}\,p \)

  • The chance that the first success comes on trial number x.
  • q, the chance of a failure on any one trial. It is the same on every trial by characteristic 3.
  • How many failures come first. Trial x is the success, so only the trials before it fail.
  • The single success that ends the sequence, on trial x itself.
  • There is no nCx factor, because only one ordering of the trials has its first success on trial x.

Read the formula aloud as a sentence: fail x minus one times, then succeed once. If you can say that sentence you can never get the exponent wrong, because the number of failures is always one fewer than the trial number of the success.

The book writes the same thing with q in place of one minus p. Both are fine, and in the worked examples it is usually quicker to compute q first, once, and then raise it to the needed power.

\[ P(X = x) = q^{x - 1}p \]

One more habit worth forming: before you evaluate, write the power out as the trials it stands for. For x equal to 4 that is F, F, F, S. The picture of three failures and a success is harder to get wrong than the symbol x minus one, and it is the same picture every time.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — the geometric pdf

18. Why there is no combination factor

Intuition

The binomial formula began with n choose x, the number of orders in which x successes can fall among n trials. The geometric formula has nothing of the kind. The picture below shows why.

Figure (svg): Left: five rows of five tiles, each row with one green S in a different position, headed binomial, one S in 5 trials. Right: the same five rows cut off at their S; the first four rows are grey and labelled stopped at 1, 2, 3, 4; only the bottom row, four F then S, is coloured and labelled X = 5.

The binomial asks how many successes, so every order counts and the combination factor counts them. The geometric stops at the first success, so four of the five orders never reach trial 5.

Put one success somewhere in five trials, and there are five places it could go. The binomial counts all of them, because it only asks how many successes, not where.

\[ \text{binomial: } \binom{5}{1}q^{4}p = 5q^{4}p \]

Now apply the geometric stopping rule to the same five rows. A row with its success on trial 2 stopped at trial 2; it never reached trial 5 at all. Of the five orders, only the one with all four failures first survives.

\[ \text{geometric: } 1 \cdot q^{4}p = q^{4}p \]

So the combination factor is not missing by accident. The stopping rule fixes the order completely: failures first, success last. There is exactly one way for the first success to be on trial x, and the count of ways is 1.

It follows that for the same p, the geometric chance of X equals 5 is one fifth of the binomial chance of one success in five trials. The two questions sound close and differ by a factor of five.

19. Example 4.17: play until you lose

Worked example

You play a game of chance, win or lose, until you lose. Your probability of losing is 0.57. What is the probability that it takes five games until you lose?

Choose the success

Why: losing is what ends play, so losing is the success, whatever you feel about it.

\[ p = 0.57, \qquad q = 1 - 0.57 = 0.43 \]

Define X and its values

Why: the book includes the losing game in the count.

\[ X = \text{games played until you lose}, \quad x = 1,\ 2,\ 3,\ \ldots \]

Write the question

Why: five games in all.

\[ P(x = 5) \]

Translate to a sequence

Why: four wins, then the loss.

\[ W,\ W,\ W,\ W,\ L \;\longrightarrow\; q^{4}p \]

Evaluate the power

Why: 0.43 to the fourth, in node.

\[ (0.43)^{4} = 0.03418801 \]

Multiply by p

Why: the loss on game 5.

\[ 0.03418801 \times 0.57 \approx 0.0195 \]

Figure (svg): A stem plot for p = 0.57 over x = 1 to 7 with heights 0.5700, 0.2451, 0.1054, 0.0453, 0.0195, 0.0084, 0.0036. The stem at x = 5 is yellow.

With a losing chance above one half, the game usually ends fast. Lasting until exactly the fifth game needs four wins in a row first, and the yellow stem is barely visible.

In context: about 2 times in 100 the game lasts exactly five rounds. A losing chance above one half makes long games rare, and the stem at 5 is tiny beside the stem at 1, where more than half of all sessions end.

Check by counting trials in the exponent

Why: four failures and one success make five games, matching x; the answer lies between zero and p, as every geometric stem after the first must.

\[ 4 + 1 = 5, \qquad 0 < 0.0195 < 0.57 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — Example 4.17

20. Try It 4.17: darts at the centre

Worked example

You throw darts until you hit the centre area. Your probability of hitting it is 0.17. Find the probability that it takes eight throws, and say what values X takes.

Define X and list its values

Why: throws until the first hit in the centre.

\[ X = \text{throws until the first centre hit}, \quad x = 1,\ 2,\ 3,\ \ldots \]

Find q

Why: a miss.

\[ q = 1 - 0.17 = 0.83 \]

Write the formula for eight throws

Why: seven misses, then the hit.

\[ P(x = 8) = (0.83)^{7}(0.17) \]

Evaluate the power

Why: in node.

\[ (0.83)^{7} \approx 0.27136 \]

Multiply by p

Why: the hit on throw eight.

\[ 0.27136 \times 0.17 \approx 0.0461 \]

Figure (svg): A stem plot for p = 0.17 over x = 1 to 16, falling slowly from 0.17 at x = 1. The stem at x = 8, height 0.0461, is yellow.

With a small p the stems shrink slowly, each keeping 83 percent of the one before, so eight throws is still a live possibility.

In context: about 1 time in 22 your first centre hit is exactly your eighth throw. Compare the shape with the game: with p small, the stems shrink slowly, so eight throws is far from absurd.

Notice that more than a quarter of the time you are still throwing after seven misses. That 0.27 from the power step is itself a probability, the chance that the first seven all miss. Part 3 turns that observation into a shortcut.

Check with a calculator-free cross-multiplication

Why: P(8) divided by P(7) must equal q, the common ratio.

\[ \frac{(0.83)^{7}(0.17)}{(0.83)^{6}(0.17)} = 0.83 = q \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — Try It 4.17

21. Which value of X is most likely?

Prediction

Take a geometric variable with p equal to 0.1, so the average wait is quite long. Commit to an answer before reading on.

Predict first

For a geometric distribution with p equal to 0.1, which single value of x has the largest probability?

  • x = 1
  • x = 10
  • x = 5
  • every value is equally likely

Correct: x = 1

Why: Each probability is q times the one before, and q is less than one, so the stems only ever shrink. The first stem, p itself, is always the tallest, whatever p is. The value 10 is the average wait, not the most likely one.

Most people pick 10, because a one-in-ten chance feels like it should take about ten tries. That feeling is about the average, and it is right about the average. But the most likely single outcome is always success on the very first try.

\[ P(1) = 0.1, \quad P(10) = (0.9)^{9}(0.1) \approx 0.0387 \]

How can both be true? Because no single value is likely here. The chance of exactly 1 is only 0.1; the rest of the probability is smeared thinly over a long tail, and it is that tail that pulls the average out to 10. Hold on to this: the mean of a geometric distribution sits well to the right of its peak.

22. The shape: every stem is q times the one before

Picture it

Put three geometric distributions side by side on the same vertical scale and the role of p is plain to see.

Figure (svg): Three stem plots side by side over x = 1 to 10 on the same vertical scale. Left, p = 0.6: a tall first stem of 0.6 and the rest collapsing almost to nothing by x = 5. Middle, p = 0.3: first stem 0.3, falling steadily. Right, p = 0.1: first stem 0.1, and the stems barely shrink across the plot.

Every panel starts at height p and multiplies by q at each step. A big p means a tall first stem and a fast collapse; a small p means a low, long tail.

Each panel starts at height p and then multiplies by q at every step. With p equal to 0.6, q is 0.4, and the stems fall by more than half each time; by x equals 5 there is almost nothing left. With p equal to 0.1, q is 0.9, and the stems lose only a tenth each step, so the tail stretches far past the edge of the plot.

\[ P(x + 1) = q \cdot P(x) \]

That one relation is the whole shape. It is why the distribution is called geometric: the probabilities form a geometric sequence with first term p and common ratio q, just like a sequence 0.6, 0.24, 0.096 in algebra.

\[ p = 0.6: \quad 0.6,\ 0.24,\ 0.096,\ 0.0384,\ \ldots \]

Every geometric distribution is skewed right, with its peak at 1. What changes with p is only how fast it decays. Keep this picture in mind when you compute: a probability far out in the tail should be small when p is large, and only modestly small when p is small.

23. Figure 4.4 and the common ratio

Picture it

The book's Figure 4.4 plots a geometric distribution with p equal to 0.02, a two percent chance that a computer component is defective, and points out that each bar is 0.98 times the bar before. That ratio is q.

Figure (svg): Stems for p = 0.02 at x = 1 to 9, drawn on a vertical axis that runs from 0.016 to 0.021 to magnify the steps. Heights 0.0200, 0.0196, 0.0192, 0.0188, 0.0184, 0.0181, 0.0177, 0.0174, 0.0170, each stem labelled. A note at the top right reads each stem = 0.98 times the one before.

Drawn from the formula, the first stem is p itself, 0.0200, and every later stem is 0.98 of its left neighbour. The axis is cut at 0.016 so that the two percent steps are visible at all.

Redrawn from the formula, the stems step down by two percent each time. The book prints the ratio from its own values: P(5) over P(4) is 0.018447 over 0.018823, which is 0.98.

\[ r = \frac{P(x = 5)}{P(x = 4)} = \frac{0.018447}{0.018823} \approx 0.98 = q \]

One correction. The book labels the first bar 0.0196, the same as the second, and blames rounding. It is not rounding: the first stem is P(1), which is p itself, 0.0200. The value 0.0196 is P(2). Every other bar in the figure matches the formula.

\[ P(1) = (0.98)^{0}(0.02) = 0.0200, \qquad P(2) = (0.98)(0.02) = 0.0196 \]

The ratio gives a fast way to walk along a distribution. Once you have any one stem, multiply by q to get the next. For small p the steps are gentle, which is why the book magnifies them; on an axis starting at zero the nine stems would look almost level.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 245-246 — Figure 4.4 and the common ratio

24. Figure 4.4: the seventh component by the common ratio

Worked example

The book says the probability that the sixth component tested is the first defective one is 0.018078, and uses the common ratio to get the seventh. Do the same, then check with the formula.

Confirm the sixth stem from the formula

Why: five good components, then the defect.

\[ P(x = 6) = (0.98)^{5}(0.02) \]

Evaluate

Why: in node.

\[ (0.98)^{5} = 0.903921, \quad \times 0.02 = 0.018078 \]

Step to the seventh by the ratio

Why: one more good component multiplies by q.

\[ P(x = 7) = 0.018078 \times 0.98 \]

Multiply

Why: one factor of 0.98.

\[ P(x = 7) \approx 0.017716 \]

Nothing new was computed: the ratio step is the formula's own structure. Going from x to x plus one adds one failure to the front of the sequence, and a failure costs a factor of q.

This is handy when a problem asks for several neighbouring probabilities. Compute the first by the formula, then walk along by multiplying by q, and use the formula once more at the end as a check.

Check with the formula directly

Why: six failures then a success must give the same value.

\[ (0.98)^{6}(0.02) = 0.885842 \times 0.02 \approx 0.017717 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 246 — the common ratio

25. Example 4.20: the first defect on the seventh component

Worked example

The probability of a defective computer component is 0.02. Components are selected at random. Find the probability that the first defect is caused by the seventh component tested.

Define X

Why: components tested until the first defect, including the defective one.

\[ X = \text{components tested until the first defect}, \quad X \sim G(0.02) \]

Write the question

Why: the seventh is the first defective one.

\[ P(x = 7) \]

Write the formula

Why: six good ones, then a defect.

\[ P(x = 7) = (0.98)^{6}(0.02) \]

Evaluate the power

Why: in node.

\[ (0.98)^{6} = 0.885842 \]

Multiply by p

Why: the defective seventh component.

\[ 0.885842 \times 0.02 \approx 0.0177 \]

In context: the seventh component is the first bad one about 1.8 times in 100. That matches the value you reached by the common ratio on the last slide.

The book does this on a TI calculator with geometpdf, entering 0.02 and 7. The arithmetic above is all that function does. The related function geometcdf gives the chance of at most 7, which you will get without a calculator in Part 3.

\[ \text{geometpdf}(0.02,\ 7) = (0.98)^{6}(0.02) \]

Check that the answer is plausible

Why: every stem after the first is below p and above zero, and this one is close to p because q is close to 1.

\[ 0 < 0.0177 < 0.02 = P(1) \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 246 — Example 4.20

26. Try It 4.20: the ninth steel rod

Fill the middle

The probability of a defective steel rod is 0.01. Rods are selected at random. Fill in the blanks to find the probability that the first defect occurs on the ninth rod.

Fill in the blanks

P(x = 9) = (0.99)^8} }(0.01) = 0.922745 \times 0.01 \approx 0.0092

Why: A good rod has probability 0.99. Eight good rods come before the defective ninth, so the power is eight, not nine. Eight factors of 0.99 give about 0.9227, and one factor of 0.01 gives about 0.0092.

The exponent is the step people miss. Count on your fingers: rods one to eight are good, rod nine is bad. That is eight failures, and the ninth trial is the success itself, which is the lone factor of 0.01 at the end.

In context: fewer than one time in a hundred is the ninth rod the first defective one. With p this small, every single value is unlikely; the probability is spread over hundreds of possible values, and Part 4 will show the average wait is 100 rods.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 247 — Try It 4.20

27. Two answers that look almost the same

Anomaly

Two students answer the ninth-rod question. One writes 0.0092, the other 0.0091. They differ by so little that it is tempting to call it rounding. It is not: one of them answered a different question.

Annotate

On: \( (0.99)^{9}(0.01) \approx 0.0091 \quad \text{vs} \quad (0.99)^{8}(0.01) \approx 0.0092 \)

  • Nine good rods and then a bad one. That is ten rods in all, so this is the chance the first defect is on the tenth rod.
  • Eight good rods and then a bad one: the ninth rod is the first defect, which is what was asked.
  • Consecutive stems differ by the factor q, and here q is 0.99, so neighbouring answers differ by only one percent.

With q near one, a wrong exponent hides inside the rounding. With a larger p it cannot hide. For the Example 4.19 students, where q is 0.45, the same slip turns 0.0501 into 0.0226, less than half the right answer.

\[ q = 0.45: \quad (0.45)^{3}(0.55) \approx 0.0501, \quad (0.45)^{4}(0.55) \approx 0.0226 \]

So never judge an exponent by whether the answer looks plausible. Count the factors instead: for the first success on trial x there must be exactly x of them, x minus one factors of q and one factor of p. The count is the check that works for every p.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 247 — Try It 4.20

28. Slide p and watch the stems

Tweak it

The slider sets p. The spikes are the geometric probabilities for x from 1 to 12, each drawn at its value of x.

Parameter explorer

Start at p equal to 0.3, then move p. Which stem is always tallest? What happens to the height of that stem, and to the tail, as p gets smaller?

\[ P(X = x) = (1 - {p})^{x - 1}({p}), \qquad \mu = \frac{1}{{p}} \]

  • p — from 0.05 to 0.9: p = success probability

Whatever p you choose, the spike at x equals 1 stands tallest, at height exactly p, and the others step down by the factor one minus p. Push p up toward 0.9 and nearly everything piles on the first spike; there is little doubt you succeed at once.

Pull p down toward 0.05 and the picture flattens: the first spike is short, and the twelfth is still almost as tall as the first. Most of the probability is off the right edge of the plot, beyond x equals 12. That is what a small p means in practice: long, unpredictable waits.

\[ p = 0.05: \quad P(X > 12) = 0.95^{12} \approx 0.54 \]

Watch the readout of the mean as you slide. At 0.3 it is about 3.3; at 0.05 it is 20. Part 4 explains why.

29. At least, more than: tails without adding forever

Section

Part 3

30. More than k trials means the first k all failed

Concept

Section 4.1 found the chance of a range of values by adding stems. For a geometric variable an upper range, such as more than 3 trials, has infinitely many stems. You cannot add them one at a time, and you do not need to.

Ask what more than three trials means in terms of the experiment. It means no success on trials 1, 2 or 3; that is, the first three trials all failed. What happens after trial 3 does not matter.

\[ X > 3 \iff \text{trials } 1, 2, 3 \text{ all fail} \]

Multiply the three failures

Why: independence again.

\[ P(X > 3) = q \cdot q \cdot q = q^{3} \]

\[ P(X > k) = q^{k} \]

That one line replaces an endless sum. The infinite tail of stems from 4 onward adds to exactly q cubed; the event just described is the same event, seen from the start of the sequence rather than the end.

The complement then gives the lower piece. At most k trials is the opposite of more than k, so its chance is one minus q to the k. This is what the TI function geometcdf computes.

\[ P(X \le k) = 1 - q^{k} \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — the alternative pair of formulas

31. Example 4.18(b): at least three reports

Worked example

A safety engineer believes 35 percent of plant accidents are caused by employees not following instructions. Reports are drawn at random, with replacement, until one shows that cause. What is the probability that the engineer must examine at least three reports?

Define X and p

Why: the success is a report blaming ignored instructions.

\[ X \sim G(0.35), \qquad q = 0.65 \]

Translate at least three

Why: three is included; the values run on forever.

\[ P(X \ge 3) = P(3) + P(4) + P(5) + \cdots \]

Say what the event means at the start

Why: needing at least three reports means reports 1 and 2 both failed.

\[ X \ge 3 \iff X > 2 \]

Use the tail rule with k equal to 2

Why: two failures in a row.

\[ P(X > 2) = q^{2} = (0.65)^{2} \]

Evaluate

Why: 0.65 times 0.65.

\[ (0.65)^{2} = 0.4225 \]

Figure (svg): A stem plot for p = 0.35 over x = 1 to 10. The stems at 1 and 2 (0.35 and 0.2275) are red; the stems from 3 upward are yellow. A label reads yellow total = 0.65² = 0.4225.

The yellow tail runs on forever, but you never add it up. It is exactly the event that the first two reports were failures, which is one product.

In context: in about 42 cases out of 100, the engineer reads at least three reports before finding one that blames ignored instructions. The book sets up this question but does not print the number.

The book suggests one minus P of fewer than 3, and that route works too. Fewer than three is just the two stems at 1 and 2, which gives the check below.

Check through the complement

Why: one minus the two red stems must equal q squared.

\[ 1 - (0.35 + 0.2275) = 1 - 0.5775 = 0.4225 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 243-244 — Example 4.18

32. Four phrases, four powers of q

Concept

Every tail question is one of four phrases. Each one decides how many trials must fail, and that number is the power of q.

Figure (svg): Four number lines over x = 1 to 8 with an ellipsis at the right. More than 3 fills 4 onward and is labelled q³. At least 3 fills 3 onward, labelled q². At most 3 fills 1, 2, 3, labelled 1 − q³. Fewer than 3 fills 1 and 2, labelled 1 − q². A dashed mark sits at 3 on every line.

Upper tails are powers of q; lower pieces are one minus a power of q. The power is the number of trials that must fail, and the dashed mark at 3 decides whether that is 3 or 2.
phrase (k = 3)valuestrials that must failprobability
more than 34, 5, 6, …the first 3q³
at least 33, 4, 5, …the first 2q²
at most 31, 2, 3complement of more than 31 − q³
fewer than 31, 2complement of at least 31 − q²

The boundary test from Section 4.1 still decides everything. Is 3 at least 3? Yes, so the event starts at 3, and only the first two trials must fail. Is 3 more than 3? No, so the event starts at 4, and the first three must fail.

\[ P(X \ge k) = q^{k - 1}, \qquad P(X > k) = q^{k} \]

A quick sanity check on any answer: at least 1 must have probability 1, since every experiment needs at least one trial. The formula gives q to the zero, which is 1. Good.

Notice that exactly one of each pair is a power of q and the other is one minus a power. Upper events, the ones that run off to infinity, are the easy ones here, because they only say that some early trials failed. Lower events are finite lists of stems, and the complement turns them back into an upper event. That is the reverse of the binomial, where the finite lists were the easy side.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — more than n and less than n

33. Match each question to its probability

Matching

A geometric variable has success probability p and failure probability q. Match each event to the expression that gives its probability.

Match the pairs

  • a. more than 5 trials
  • b. at least 5 trials
  • c. at most 5 trials
  • d. fewer than 5 trials
  • e. exactly 5 trials
  • p. q⁵
  • r. q⁴
  • s. 1 − q⁵
  • t. 1 − q⁴
  • u. q⁴p

Why: More than five means the first five all fail. At least five means only the first four fail. At most five is the complement of more than five, and fewer than five is the complement of at least five. Exactly five is four failures and then the success.

Two of the five use the same power of q, and one of them multiplies by p. That is the difference between exactly five, which pins down trial 5 as the success, and at least five, which only says trials 1 to 4 failed and leaves trial 5 open.

If you mismatched, go back to the boundary test. Say the number 5 out loud and ask whether it satisfies the phrase; then count how many trials before the event can start must have failed.

Put numbers in to see the sizes. With p equal to 0.35, more than five is 0.65 to the fifth, about 0.116, and at least five is 0.65 to the fourth, about 0.179. The gap between them, about 0.062, is exactly the stem at 5, which is 0.65 to the fourth times 0.35.

\[ 0.1785 - 0.1160 = 0.0625 = (0.65)^{4}(0.35) \]

34. Try It 4.18: at least ten exams

Worked example

An instructor feels that 15 percent of students get below a C on the final. Exams are drawn at random, with replacement, until one below a C turns up. Write the probability that at least ten exams must be examined, and evaluate it.

Define X and p

Why: an exam below a C is the success.

\[ X \sim G(0.15), \qquad q = 0.85 \]

State the question mathematically

Why: this is all the book asks.

\[ P(x \ge 10) \]

Translate into failures

Why: at least ten exams means the first nine were all C or above.

\[ P(x \ge 10) = q^{9} \]

Substitute q

Why: nine factors of 0.85.

\[ = (0.85)^{9} \]

Evaluate

Why: in node.

\[ \approx 0.2316 \]

In context: nearly a quarter of the time, the instructor reads at least ten exams before finding one below a C. That surprises people, since 15 percent does not sound rare. But nine passes in a row at 85 percent each is quite likely.

A common slip is to write q to the tenth. That is the chance of more than ten, 0.1969, and it drops the event where the tenth exam is the first one below a C.

Check by the complement

Why: at most nine is the opposite event, so the two must add to one.

\[ P(x \le 9) = 1 - 0.85^{9} = 0.7684, \quad 0.7684 + 0.2316 = 1 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — Try It 4.18

35. Error analysis: the book's fewer-than formula

Error analysis

The book's pair of tail formulas on page 244 contains a slip. Look at the second one, which is meant to answer: what is the probability it takes less than n trials for a success?

Annotate

On: \( P(x < n) = 1 - (1 - p)^{n} \quad \text{(as printed)} \)

  • This is the chance that the first n trials all fail, which is the event more than n.
  • So the right side is the complement of more than n, which is at most n, and it includes the value n itself.
  • Fewer than n stops at n minus one, so its complement is at least n, which needs only n minus one failures.
  • P of fewer than n equals one minus q to the power n minus one.

Test the printed formula with numbers from Example 4.18, where p is 0.35 and the question is fewer than 3. Fewer than 3 is the values 1 and 2, so add those two stems directly.

\[ P(x < 3) = 0.35 + 0.2275 = 0.5775 \]

The printed formula gives one minus 0.65 cubed, which is 0.7254. That is the chance of at most 3, one stem too many. The corrected formula gives one minus 0.65 squared, 0.5775, matching the direct sum.

\[ P(x < n) = 1 - (1 - p)^{n - 1}, \qquad P(x \le n) = 1 - (1 - p)^{n} \]

The lesson is the one from Section 4.1: always test the boundary value. A formula for fewer than must leave n out.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — the formula for less than n, corrected

36. Estimate: more than twenty tries at one in ten

Estimation

You call stores for the special ink, where one store in ten has it. Before you calculate, estimate the chance that you are still dialling after twenty calls.

Predict first

Roughly what is the probability that more than 20 calls are needed when p is 0.1?

  • about 0.001
  • about 0.01
  • about 0.12
  • about 0.5

Correct: about 0.12

Why: More than twenty calls means twenty failures in a row, 0.9 to the twentieth power. Ten misses in a row is about 0.35, and another ten multiplies by 0.35 again, giving about 0.12. Computed exactly in node, it is 0.1216.

The trick for estimating a power of q is to find a round power first. Ten factors of 0.9 are about 0.35, so twenty factors are about 0.35 times 0.35, a little over a tenth.

\[ P(X > 20) = 0.9^{20} = (0.9^{10})^{2} \approx 0.349^{2} \approx 0.12 \]

So roughly one shopper in eight makes more than twenty calls, twice the average. That long tail is the practical face of a small p, and it is why an average wait of ten calls is not a promise of ten calls.

Estimating first also protects you from a slipped exponent. If your calculator had shown 0.012, the answer to twenty misses in a row at 0.99 would be the wrong problem entirely, and the rough square of a third would have told you so at once.

37. How many calls to be 90 percent sure?

Concept

Planners rarely ask for the chance of one exact value. They ask how many tries to budget so that success is very likely. The at-most formula answers that directly: find the smallest n for which the chance of success within n tries reaches the level you want.

\[ P(X \le n) = 1 - q^{n} \ge 0.9 \iff q^{n} \le 0.1 \]

In words: you are 90 percent sure of success within n calls once the chance of n misses in a row has dropped to a tenth or less. For the ink, with q equal to 0.9, keep multiplying by 0.9 until you pass below 0.1.

Figure (svg): Dots for n = 1 to 36 showing the probability of finding the ink within n calls when p = 0.1, rising from 0.1 at n = 1 and levelling toward 1. A dashed horizontal line at 0.9 meets the dots between n = 21 (0.891, red) and n = 22 (0.902, yellow). A dashed vertical line at n = 10, the mean, meets the curve at 0.651.

The chance of success within n calls climbs fast at first and then crawls. It passes one half after 7 calls, is only 0.65 at the mean of 10, and needs 22 calls to pass 0.9.

Try 21 calls

Why: twenty-one misses in a row, computed in node.

\[ 0.9^{21} \approx 0.109 > 0.1 \]

Try 22 calls

Why: one more factor of 0.9.

\[ 0.9^{22} \approx 0.098 \le 0.1 \]

So you need to budget 22 calls to be 90 percent sure of finding the ink, more than twice the average of 10. At the mean itself you are only 65 percent sure; the curve in the picture climbs quickly at first and then crawls.

This is the most useful reading of a geometric model in practice. The mean tells you the long-run cost per success, but a person waiting once needs to know how bad a wait can plausibly get, and that is a statement about the tail.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — the probability that it takes n or fewer trials

38. Trap: at least three written as q cubed

Trap

The trap

For Example 4.18, needing at least three reports, a common line:

\[ P(X \ge 3) = q^{3} \]

\[ = 0.65^{3} \approx 0.2746 \]

Wrong. That is the chance of more than three.

It is tempting because the tail rule has k in the exponent, and 3 is the number in the question.

The fix

At least three includes X equal to 3, where the third report is the success. Only the first two must fail:

\[ P(X \ge 3) = q^{2} \]

\[ = 0.65^{2} = 0.4225 \]

Rewrite at least as more than first: at least 3 is the same event as more than 2. Then the power is the number after more than.

The wrong answer drops the case where report 3 is the first success.

\[ X \ge 3 \iff X > 2 \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 244 — at least translates to greater than or equal to

39. The mean and standard deviation

Section

Part 4

40. The mean is one over p: the average wait

Concept

If one store in ten carries the ink, how many calls do you expect to make? Almost everyone says ten, and the geometric mean agrees. Here is why.

Figure (svg): Top: a grid of 100 small squares in four rows of 25, with 19 yellow squares marking successes and the rest grey. Bottom: a dot plot of the 19 waits between successes over 1 to 12, with stacks of 4 at 1 and at 4, and a dashed yellow line at the average wait, about 5.05, beside the label average wait 96/19.

Nineteen successes in ninety-six trials means an average wait of about five, one over 0.2. The single most common wait is still the shortest one, a success on the very next trial.

This strip is 100 trials simulated with p equal to 0.2. Nineteen successes fell in the first 96 trials. Chop the strip at each success and you get nineteen waits, and their average is the total length divided by the number of pieces.

\[ \text{average wait} = \frac{96}{19} \approx 5.05 \]

In the long run, a long strip of N trials holds about Np successes, so it is chopped into about Np waits of total length N. Divide, and the N cancels.

\[ \mu = \frac{N}{Np} = \frac{1}{p} \]

So a one-in-five chance takes five tries on average, a one-in-ten chance ten, and a two-percent chance fifty. The book's formula says the same, and a long computation in node agrees: adding x times P(x) over the first 2000 values for p equal to 0.35 gives 2.857, which is one over 0.35.

Look at the dot plot, though. The most common wait in the simulation is the shortest, one trial, exactly as the stems predicted. The mean of 5 is pulled out by a few long waits, one of twelve and one of ten.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — the expected value of X, the mean for Case I

41. The standard deviation: nearly as big as the mean

Concept

The book gives the standard deviation in a form that looks unfamiliar. Simplify it and a cleaner form appears.

\[ \sigma = \sqrt{\left(\frac{1}{p}\right)\left(\frac{1}{p} - 1\right)} \]

Write the bracket over a common denominator

Why: one over p minus one is one minus p, over p.

\[ \frac{1}{p} - 1 = \frac{1 - p}{p} = \frac{q}{p} \]

Multiply the two brackets

Why: one over p times q over p.

\[ \sigma^{2} = \frac{1}{p} \cdot \frac{q}{p} = \frac{q}{p^{2}} \]

Take the square root

Why: p squared comes out as p.

\[ \sigma = \frac{\sqrt{q}}{p} \]

Compare it with the mean, one over p. The standard deviation is the mean times the square root of q. When p is small, q is close to one, its square root is close to one, and the standard deviation is almost as large as the mean.

\[ \frac{\sigma}{\mu} = \sqrt{q} \]

That is a statement about predictability. A waiting time with a small p is not just long on average; it is wildly variable. An average of 50 components comes with a standard deviation of about 49.5, so waits of 5 or of 150 are both unremarkable.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — the standard deviation of Case I

42. Example 4.18(a): how many reports on average?

Worked example

Return to the safety engineer, with p equal to 0.35. On average, how many reports would the engineer expect to look at until finding one caused by employees not following instructions? Find the standard deviation as well.

Write the mean formula

Why: the average number of trials, including the one that succeeds.

\[ \mu = \frac{1}{p} \]

Substitute p

Why: the 35 percent from the problem.

\[ \mu = \frac{1}{0.35} \]

Divide

Why: in node.

\[ \mu \approx 2.857 \approx 2.86 \text{ reports} \]

Write the standard deviation

Why: the simplified form.

\[ \sigma = \frac{\sqrt{q}}{p} = \frac{\sqrt{0.65}}{0.35} \]

Evaluate the root

Why: q is 0.65.

\[ \sqrt{0.65} \approx 0.8062 \]

Divide by p

Why: the same 0.35.

\[ \sigma \approx \frac{0.8062}{0.35} \approx 2.30 \text{ reports} \]

In context: over many searches, the engineer reads about 2.86 reports per search on average, give or take about 2.3. A mean that is not a whole number is fine; no single search reads 2.86 reports, but the long-run average can be anything.

Two notes on the book's version. It prints the mean as about 2.8; one over 0.35 is 2.857, so 2.86, or 2.9 to one place. And its solution rewords part (a) as a report caused by something other than failure to follow instructions, which is the opposite event and would have p equal to 0.65. The calculation, correctly, uses 0.35.

Check that the mean times p is one

Why: the long-run identity behind the formula.

\[ 2.857 \times 0.35 = 1.000 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — Example 4.18

43. Example 4.20, continued: components until the first defect

Worked example

With the defect probability 0.02 from Example 4.20, how many components do you expect to test until one is defective? Find the variance and the standard deviation too.

Mean

Why: one over p.

\[ \mu = \frac{1}{0.02} = 50 \text{ components} \]

Variance, in the book's form

Why: one over p, times one over p minus one.

\[ \sigma^{2} = \left(\frac{1}{0.02}\right)\left(\frac{1}{0.02} - 1\right) \]

Evaluate the brackets

Why: one over 0.02 is 50, and 50 minus 1 is 49.

\[ \sigma^{2} = (50)(49) = 2450 \]

Same thing in the simplified form

Why: q over p squared.

\[ \sigma^{2} = \frac{0.98}{0.0004} = 2450 \]

Standard deviation

Why: square root, in node.

\[ \sigma = \sqrt{2450} \approx 49.5 \text{ components} \]

Figure (svg): Thin stems for p = 0.02 over x = 1 to 200, falling from 0.02 toward 0.0004. Stems up to x = 50 are blue, later stems grey. A dashed yellow line at x = 35 is labelled median 35; a dashed orange line at x = 50 is labelled mean 50. A label reads 63.6 percent of the probability is at 50 or below.

The long right tail drags the mean out to 50, while half of all first defects have already turned up by the thirty-fifth component. The mean is a long-run average, not a typical wait.

Both forms of the variance give 2450, so use whichever you trust more. The book's form, one over p times one over p minus one, is quick with a round p like 0.02; the q over p squared form is quicker with an awkward p like 0.015.

In context: on average the fiftieth component is the first defective one, give or take about 49.5. The picture shows how lopsided that is. Half of all first defects appear by component 35, and about 64 percent by component 50; the mean is dragged right by rare, very long runs of good components.

Check the ratio of spread to mean

Why: it must equal the square root of q.

\[ \frac{49.5}{50} = 0.990, \quad \sqrt{0.98} = 0.990 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 246-247 — Example 4.20 and the variance

44. Is the mean a typical wait?

Prediction

For the components, the mean wait is 50. Before you look at the picture on the last slide again, commit to a guess.

Predict first

When p is 0.02 and the mean is 50, what is the probability that the first defect appears within the first 50 components?

  • exactly 0.5
  • about 0.36
  • about 0.64
  • about 0.98

Correct: about 0.64

Why: At most fifty is one minus the chance that all fifty are good, one minus 0.98 to the fiftieth power, which is about 0.64 when computed in node. The first defect comes before the mean more often than after it, because the distribution is skewed right.

It is tempting to think the mean splits the distribution in half, as it does for symmetric shapes. For a geometric distribution it never does. The stems are tallest on the left and the tail runs to the right forever, so the balance point, the mean, sits to the right of the halfway point.

\[ P(X \le 50) = 1 - 0.98^{50} \approx 0.636 \]

The halfway point is 35: that is where the chance of at most x first passes one half. So when someone quotes an average wait from a geometric model, most real waits will be shorter than it, and a few will be much longer.

The same lopsidedness shows at the peak. For the literacy survey below, with p equal to 0.12 and a mean of 8.33, asking just one woman is about two and a half times as likely as asking exactly eight. Say an average wait, never a typical or most likely wait.

\[ P(1) = 0.12, \qquad P(8) = 0.88^{7}(0.12) \approx 0.049 \]

45. Example 4.21: asking people about cancer

Worked example

The lifetime risk of developing cancer is about one in 67, or 1.5 percent. Let X be the number of people you ask until one says they have developed cancer, so X has a geometric distribution with p equal to 0.015.

(a) Ten people, the tenth saying yes

Why: nine noes, then a yes.

\[ P(x = 10) = (0.985)^{9}(0.015) \]

Evaluate

Why: in node.

\[ = 0.872823 \times 0.015 \approx 0.0131 \]

(b) Twenty people

Why: nineteen noes, then a yes.

\[ P(x = 20) = (0.985)^{19}(0.015) \approx 0.750392 \times 0.015 \approx 0.0113 \]

(c)(i) Mean

Why: one over p.

\[ \mu = \frac{1}{0.015} \approx 66.67 \text{ people} \]

(c)(ii) Standard deviation

Why: square root of q over p squared.

\[ \sigma = \sqrt{\frac{0.985}{0.015^{2}}} = \sqrt{4377.8} \approx 66.16 \text{ people} \]

Figure (svg): Thin stems for p = 0.015 over x = 1 to 250, starting at 0.015 and falling slowly to about 0.0004. The stems at 10 and 20 are yellow and labelled 0.0131 and 0.0113. A dashed line at x = 66.67 is labelled mean 66.67, and a note says one standard deviation, 66.16, either side of the mean runs from about 0 to 133.

With p this small every stem is short and the distribution is very long. The standard deviation is almost as big as the mean, so answers anywhere from a handful of people to well over a hundred are ordinary.

In context: on average you would ask about 67 people before the first one who has developed cancer, but with a standard deviation of 66 that average says little about any one search. The stems at 10 and 20 are almost the same height, because twenty-odd more failures at 98.5 percent barely dent the probability.

A wording note. Part (a) is printed as ask ten people before one says they have cancer. Read literally, before would mean ten noes and the yes on the eleventh person. The book's solution computes the chance that the tenth person is the first yes, and so does this slide.

Check against the book and the ratio

Why: the book prints 0.0131 and 0.0113, and their ratio must be q to the tenth.

\[ \frac{0.011256}{0.013092} = 0.860 = 0.985^{10} \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 247 — Example 4.21

46. Try It 4.21: literacy, one woman at a time

Worked example

The literacy rate for women in Afghanistan is given as 12 percent. Let X be the number of women you ask until one says she is literate.

(a) The distribution

Why: a literate woman is the success.

\[ X \sim G(0.12), \qquad q = 0.88 \]

(b) Five women, the fifth literate

Why: four failures, then the success.

\[ P(x = 5) = (0.88)^{4}(0.12) = 0.599695 \times 0.12 \approx 0.0720 \]

(c) Ten women

Why: nine failures, then the success.

\[ P(x = 10) = (0.88)^{9}(0.12) = 0.316478 \times 0.12 \approx 0.0380 \]

(d)(i) Mean

Why: one over p.

\[ \mu = \frac{1}{0.12} \approx 8.33 \text{ women} \]

(d)(ii) Standard deviation

Why: square root of q, over p.

\[ \sigma = \frac{\sqrt{0.88}}{0.12} = \frac{0.9381}{0.12} \approx 7.82 \text{ women} \]

Figure (svg): A stem plot for p = 0.12 over x = 1 to 20. Stems at 5 and 10 are yellow. A dashed line at 8.33 is labelled mean 8.33.

Between the two yellow stems the heights roughly halve: five more failures multiply by 0.88 five times, which is about 0.53.

In context: on average you would ask a little over eight women before the first one who is literate, with a standard deviation of nearly eight. As in the book, part (b) is read as the fifth woman being the first literate one.

Going from five to ten multiplied the probability by 0.88 five times, about 0.53. So the ten-woman answer is roughly half the five-woman one, which the two stems show.

Check with the ratio

Why: P(10) over P(5) must be q to the fifth.

\[ \frac{0.0380}{0.0720} \approx 0.528, \quad 0.88^{5} \approx 0.528 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 247 — Try It 4.21

47. Binomial beside geometric

Comparison

The two distributions share independent trials and a constant p, and differ in everything else. Fill in the blanks to pin the differences down.

Comparison matrix

binomial B(n, p)geometric G(p)
what is fixed in advancethe number of trials none success
what X countssuccessestrials
smallest value of X01
largest value of Xnnone
P(X = x)nCx · pˣ · qⁿ⁻ˣqˣ⁻¹ · p
meannp1/p
standard deviation√(npq)√q / p

Read the rows as a story. Fixing n makes the successes random, so the binomial's values run from 0 to n. Fixing one success makes the trials random, so the geometric's values start at 1 and never stop.

Even the means tell the same story from opposite ends. In n trials you expect np successes; to get one success you expect one over p trials. Both say that a success turns up once every one over p trials on average.

The spreads differ in an instructive way. The binomial's standard deviation grows only like the square root of n, so with many trials the count of successes is fairly predictable relative to its size. The geometric's standard deviation is nearly its whole mean when p is small, so a single waiting time is never predictable. That is the price of letting the experiment decide its own length.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 245 — like the binomial distribution, the geometric has a mean and standard deviation

48. Counting failures, memory, and modelling

Section

Part 5

49. Case II: counting only the failures

Concept

The book notes that X can be defined two ways. Case I, used so far, counts every trial including the success. Case II counts only the failures before the success. Same experiment, different number written down.

Figure (svg): Two stem plots for p = 0.35, one above the other, with identical stem heights 0.35, 0.2275, 0.1479 and so on. The top one's axis is labelled 1 to 7 and titled Case I, X = trials including the success. The bottom one's axis is labelled 0 to 6 and titled Case II, Y = failures before the success.

Nothing about the experiment changes, only what you count. So the picture is identical and the labels slide by one, which moves the mean by one and leaves the spread alone.

If the first success is on trial 3, Case I records 3 and Case II records 2. So the Case II variable is always one less, and its values start at zero, since you might succeed at once with no failures.

\[ Y = X - 1, \qquad y = 0,\ 1,\ 2,\ \ldots \]

\[ P(Y = y) = q^{y}p \]

The heights do not change, only their labels, so the mean moves down by exactly one and the spread does not move at all.

\[ \mu_{Y} = \frac{1}{p} - 1 = \frac{q}{p}, \qquad \sigma_{Y} = \frac{\sqrt{q}}{p} \]

The book gets two things wrong here. It prints the Case II standard deviation as the square root of q over p; sliding a distribution cannot change its spread, and the correct value is the same as Case I, root q over p. And on page 246 it gives the Case II mean for the components as one over p, 50; that is the Case I mean, and failures before the first defect average 49.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 245-246 — Case I and Case II

50. Case II: good components before the first defect

Worked example

With the defect probability 0.02 again, let Y be the number of good components tested before the first defective one. Find the chance of exactly six good ones first, and the mean and standard deviation of Y.

Write the Case II formula

Why: six failures, then the success, with the success not counted.

\[ P(Y = 6) = q^{6}p = (0.98)^{6}(0.02) \]

Evaluate

Why: the same product as Example 4.20.

\[ P(Y = 6) \approx 0.0177 \]

That is no coincidence. Six good components before the first defect is the same event as the first defect on component 7. Case II with y equal to 6 and Case I with x equal to 7 describe one sequence.

Mean of Y

Why: q over p, or one less than the Case I mean.

\[ \mu_{Y} = \frac{0.98}{0.02} = 49 \]

Standard deviation of Y

Why: unchanged by the shift.

\[ \sigma_{Y} = \frac{\sqrt{0.98}}{0.02} \approx 49.5 \]

The book's printed Case II standard deviation, the square root of q over p, would give the square root of 49, which is 7. A spread of 7 for waits that routinely run past 100 cannot be right, and it contradicts the book's own 49.5 for the same process.

Check that the two cases agree

Why: the Case II mean plus one must be the Case I mean, and the spreads must match.

\[ 49 + 1 = 50 = \mu_{X}, \qquad \sigma_{Y} = \sigma_{X} = 49.5 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 246 — Case II mean and variance

51. Which case does the wording ask for?

Discrimination

Decide for each question whether the count includes the successful trial, Case I, or counts only the failures before it, Case II. Listen for the words until and before.

Sort into buckets

Sort each question by the variable it describes.

Case I: trials including the success, values 1, 2, 3, …
How many rolls until the first three?; How many components are tested until one is defective?; How many games do you play, including the losing game?
Case II: failures before the success, values 0, 1, 2, …
How many rolls that are not a three come before the first three?; How many good components are tested before the first defect?; How many games do you win before you lose?
one
These count every trial up to and including the one that succeeds, so the smallest possible answer is 1 and the mean is one over p.
two
These count only the failures that come before the success, so the answer can be zero and the mean is q over p, one less.

Case I and Case II always come in pairs describing the same run of trials: a count of rolls until the first three sits next to a count of non-threes before it, and the two numbers differ by exactly one.

The test is whether zero is a sensible answer. You can win zero games before losing, so that question is Case II. You cannot play zero games and still lose, so a count of games played is Case I.

The book itself is loose with before in Examples 4.21 and Try It 4.21, where ask ten people before one says yes is solved as the tenth person saying yes. When a problem's wording and its intended case disagree, state which reading you use, as the worked examples here did.

52. Memorylessness: a run of failures changes nothing

Concept

The book's baseball player Jones gets a hit with probability 0.20 each time at bat, and has gone hitless in his last ten at-bats. What is the chance that his next hit comes on his third at-bat from now? The geometric answer ignores the ten failures completely.

Here is why. Given that the first ten at-bats failed, ask for the chance that the first hit comes on at-bat 13. Divide the joint probability by the probability of the condition.

\[ P(X = 13 \mid X > 10) = \frac{P(X = 13)}{P(X > 10)} \]

Substitute the formulas

Why: twelve failures then a hit, over ten failures.

\[ = \frac{q^{12}p}{q^{10}} \]

Cancel ten factors of q

Why: the slump divides out.

\[ = q^{2}p = P(X = 3) \]

Figure (svg): Left: stems for p = 0.2 over at-bats 1 to 20; stems 1 to 10 grey, stems 11 to 20 yellow and small. Right: the ten yellow stems, each divided by 0.8 to the tenth power and relabelled 1 to 10, standing exactly inside hollow cyan circles that mark the original probabilities 0.2, 0.16, 0.128 and so on.

Blow the leftover tail up so it totals one again and it is the original distribution, stem for stem. That is memorylessness: after any run of failures the future looks exactly like a fresh start.

The picture shows the same cancellation for every value at once. Take the tail left after ten failures, scale it back up to total one, and it is the original distribution, stem for stem.

\[ P(X = n + k \mid X \ge k + 1) = P(X = n) \]

This is the book's formal statement, with k the number of previous failures. The geometric is the only discrete distribution with this property. It is not magic; it is independence plus a constant p, which together mean each at-bat starts afresh.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — the geometric distribution is memoryless

53. Jones after ten hitless at-bats

Worked example

Jones gets on base with probability 0.20 per at-bat and has failed in his last ten. Find the probability that he gets a hit on his third at-bat from now, and the chance he goes hitless for ten more.

Apply memorylessness

Why: the ten earlier failures are irrelevant; count from now.

\[ P(\text{first hit on the 3rd at-bat from now}) = P(X = 3) \]

Use the formula

Why: two outs, then a hit.

\[ P(X = 3) = (0.8)^{2}(0.2) \]

Evaluate

Why: 0.8 squared is 0.64.

\[ = 0.64 \times 0.2 = 0.128 \]

Ten more hitless at-bats

Why: a fresh run of ten failures.

\[ P(X > 10) = 0.8^{10} \approx 0.107 \]

In context: there is a 12.8 percent chance that Jones's next hit comes on his third at-bat from now, exactly as if the slump had never happened. And about one time in nine he will go hitless for another ten, the same chance he had of the first slump.

That second number is worth pausing on. A ten at-bat slump happens about one time in nine for a player like Jones, purely by chance. It is not evidence that anything is wrong with him, and under this model it predicts nothing about what comes next.

Check with the conditional formula directly

Why: the unconditional chance of a first hit at at-bat 13, divided by the chance of ten failures, must give the same 0.128.

\[ \frac{0.8^{12}(0.2)}{0.8^{10}} = 0.8^{2}(0.2) = 0.128 \;\checkmark \]

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — the Jones example

54. Trap: he is due for a hit

Trap

The trap

After Jones's ten hitless at-bats, a common claim:

\[ P(\text{hit next}) > 0.20 \]

because a player who bats 0.20 must catch up to his average.

Wrong under the geometric model.

It is tempting because over a whole season his hits will be about a fifth of his at-bats.

The fix

With independent at-bats and a constant p, the next at-bat has the same chance as any other:

\[ P(\text{hit next}) = 0.20 \]

The season average is restored by the many future at-bats swamping the slump, not by the next ones compensating for it. Ten failures in several hundred at-bats barely move the long-run proportion.

If you believe he really is due, you believe the at-bats are not independent, and then the geometric model is the wrong tool.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 243 — Jones's probability of a hit begins anew each time

55. When the model is wrong: learning and wear

Real world

The book warns that a geometric model assumes the probability of success never changes, and so it cannot capture learning. That assumption is fine for dice and coins. Many real waiting problems break it.

Discussion prompt

For each situation, say whether a constant p is reasonable, and if not, whether the true chance of success rises or falls as the trials go on: a beginner throwing darts until the first bullseye; a sales caller phoning a random list until the first sale; an old machine run each day until its first breakdown.

Memorylessness is the test to apply. Ask: if I knew how many failures had already happened, would I change my estimate of the next trial? If the honest answer is yes, the trials remember, and the geometric model is at best a rough first guess.

This is not a flaw in the mathematics. The formula is exactly right whenever its four characteristics hold. The modelling question, whether they hold, is yours to answer before you compute.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, p. 242 — the geometric distribution cannot capture learning

56. Put the modelling steps in order

Ranking

Here are the steps for answering any geometric question, shuffled. Put them in the order you would carry them out.

Put in order

Order the steps, first to last.

  1. Check the stopping rule: repeat until the first success
  2. Check independence and a constant p
  3. Name the success and write p and q
  4. Define X in words and decide Case I or Case II
  5. Translate the question into failures and one success
  6. Compute, then state the answer in context

Why: The first two steps decide whether the geometric model applies at all. Only then does it make sense to name the success and its probability, fix what X counts, turn the words into a string of failures and a success, and finally compute and interpret.

Every worked example in this deck followed this order, even when a step took only a few words. Skipping the first two is how a without-replacement draw or a learning dart thrower gets forced into a formula that does not describe it.

The translation step is where most marks are lost. Before any arithmetic, say how many trials must fail and whether the last trial counted must be a success.

And the last step is not optional decoration. A bare 0.4225 answers nothing; the answer is that in about 42 cases out of 100 the engineer reads at least three reports. The sentence in context is where you show that you know what X counted and which event you computed.

57. Putting it together

Section

Part 6

58. Pattern: count the failures, then multiply

Pattern

Every geometric calculation in this lesson reduced to a short chain of failures, sometimes ending in a success. Here is the routine and the formulas it produces.

  1. Is it geometric? Repeat until the first success; independent trials; the same p every time.
  2. Define X in words; decide whether it counts all trials (Case I) or only failures (Case II).
  3. Exactly x trials: x − 1 failures then a success, so q to the x − 1, times p.
  4. Tails: more than k is q to the k; at least k is q to the k − 1; the lower pieces are their complements.
  5. Summaries: mean 1/p (Case I) or q/p (Case II), and standard deviation √q / p in both cases.
questionCase I formulaExample 4.18, p = 0.35
exactly xqˣ⁻¹ pP(2) = 0.2275
more than kqᵏP(X > 2) = 0.4225
at least kqᵏ⁻¹P(X ≥ 3) = 0.4225
at most k1 − qᵏP(X ≤ 2) = 0.5775
mean1/p2.86
standard deviation√q / p2.30

Two free checks catch most errors. Count the factors in any product: there must be exactly x of them for exactly x trials. And every answer must be a probability between zero and one, with P of 1 always the tallest stem.

Look down the last column of the table. The same q squared, 0.4225, answers two differently worded questions, more than 2 and at least 3, because they are the same event. And the at-most answer, 0.5775, is one minus it. Seeing those links is faster and safer than memorising four formulas separately.

If you remember only one thing, remember the sentence behind every line: fail, fail, fail, then succeed, and multiply.

59. Check: which situation is geometric?

Check

Run the four characteristics on each option before you choose. The trap options each fail exactly one.

Check your understanding

Which of these is a geometric experiment?

  • A. Flip a fair coin 12 times and count the heads
  • B. Draw cards without replacement from a 52-card deck until the first ace
  • C. Roll a fair die until the first six, and record the number of rolls (correct)
  • D. A new archer shoots until her first bullseye, improving with every arrow

Answer: C

Why: Rolling until the first six repeats a trial until a success, the rolls are independent, the chance of a six is one sixth every time, and the number of rolls is the variable. All four characteristics hold.

Why A tempts people
The number of flips is fixed at twelve and heads are counted, which is a binomial experiment, not a geometric one.
Why B tempts people
Without replacement the chance of an ace changes after every non-ace, so the trials are not independent and p is not constant.
Why D tempts people
An archer who improves has a rising chance of success, so p is not constant; the geometric model cannot capture learning.

Each wrong option breaks a different characteristic: a fixed number of trials, draws that change the odds, and a success rate that changes with practice. Naming which characteristic fails is the explanation a marker wants.

The card draw is the subtle one, because it does repeat until a success. Replacement is what would rescue it: shuffle each card back in and every draw faces the same 4 aces in 52. Without replacement, after 40 non-aces the four aces would be all that is left, and the next draw would be certain, which no geometric model allows.

The archer fails for a different reason. Her trials may well be independent, but her p is creeping upward, so no single p describes her.

60. Check: the first bullseye on the fourth dart

Check

A dart player hits the bullseye with probability 0.25 on each throw, independently. Write out the sequence before you pick.

Check your understanding

What is the probability that the first bullseye comes on the fourth throw?

  • A. (0.75)³(0.25) ≈ 0.1055 (correct)
  • B. (0.75)⁴(0.25) ≈ 0.0791
  • C. 4(0.75)³(0.25) ≈ 0.4219
  • D. (0.75)³ ≈ 0.4219

Answer: A

Why: The first bullseye on throw four means three misses and then a hit, so three factors of 0.75 and one of 0.25, which gives about 0.1055 when computed in node.

Why B tempts people
Four misses and then a hit is five throws; the exponent must be one less than the trial number.
Why C tempts people
The factor of four counts orders, as the binomial does. The geometric stopping rule allows only one order: misses first, hit last.
Why D tempts people
Three misses in a row is the chance of needing more than three throws, which includes every later first hit, not just the fourth.

Each wrong option is a correct answer to a different question: five throws, one hit in four binomial throws, and more than three throws. Writing M, M, M, H first makes the right product obvious.

Count the factors in your answer: four throws, four factors. That check alone rules out two of the wrong options.

Then ask whether the size is sensible. The first throw alone succeeds a quarter of the time, and each later stem is three quarters of the one before: 0.25, 0.1875, 0.1406, 0.1055. An answer above 0.25 cannot be right for any throw after the first.

61. Check: interpreting the mean

Check

One product in 40 from a production line is defective, independently, and items are tested until the first defect.

Check your understanding

Which statement about the number of items tested until the first defect is correct?

  • A. The mean is 40 and the most likely value is 1 (correct)
  • B. The mean is 40 and the most likely value is 40
  • C. The mean is 40, so half of all searches need more than 40 items
  • D. The mean is 39 and the standard deviation is about 6.2

Answer: A

Why: The mean is one over 0.025, which is 40, and every geometric distribution has its tallest stem at 1. Both parts of the statement hold.

Why B tempts people
The mean is not the peak. The stems shrink from the first one on, so one item is the single most likely value.
Why C tempts people
The distribution is skewed right, so fewer than half the searches exceed the mean: the chance of more than 40 is 0.975 to the fortieth, about 0.36.
Why D tempts people
Thirty-nine is the Case II mean, failures before the defect. The standard deviation is the square root of 0.975 over 0.025, about 39.5, not 6.2.

The option with a standard deviation of 6.2 uses the book's misprinted Case II formula, the square root of q over p. For a waiting time with p this small, the spread must be close to the mean.

The option that says half the searches exceed 40 treats the mean as a median. Only about 36 percent do; the long tail makes up the difference.

To get each summary right, write the three numbers side by side before judging any statement: p is 0.025, the mean is one over p, 40, and the standard deviation is the square root of 0.975 over 0.025, about 39.5. With those on the page, every option can be tested in a line.

\[ \mu = 40, \qquad \sigma = \frac{\sqrt{0.975}}{0.025} \approx 39.5 \]

62. Explain the missing combination factor

Explain it to yourself

The difference between the binomial and geometric formulas is one factor. Explaining why it is missing shows you understand both.

Discussion prompt

In two or three sentences, explain why the binomial formula for one success in five trials has a factor of 5, while the geometric formula for the first success on trial 5 has no such factor. Mention the stopping rule.

A strong answer names the orders. Five orders are equally likely; the binomial keeps all five, and the stopping rule discards four of them. That is the whole explanation.

If your answer only said the formulas are different, try drawing the five rows of tiles from Part 2 and crossing out the ones that stop early.

63. Exit ticket

Exit ticket

Ten percent of the stores that sell printer ink carry the special ink you need. You call stores at random until one has it.

Discussion prompt

Give the distribution of X. Find the probability that the fourth store is the first to have the ink, the probability that you need more than ten calls, and the mean and standard deviation of X. Finish with one sentence in context.

Check the powers: three misses before the fourth store, ten misses for more than ten calls. And the standard deviation sits just under the mean, as it must when p is small.

Every part of the lesson is in this one problem: recognising the stopping rule, the formula, a tail, and the summaries. If a step felt shaky, the matching Part is the one to revisit.

64. Recap

Recap

A geometric experiment repeats independent trials with a constant p until the first success, and X counts the trials. Every question reduces to a chain of failures, ending in a success when the question pins down the last trial.

ideain symbolsin words
distributionX ~ G(p), x = 1, 2, 3, …one parameter; no largest value
exactly xP(X = x) = qˣ⁻¹ px − 1 failures, then the success
tailsP(X > k) = qᵏthe first k trials all fail
meanμ = 1/paverage wait; the peak is always at 1
spreadσ = √q / pnearly as big as the mean when p is small
Case IIY = X − 1, μ = q/pfailures only; same spread
memorylessP(X = n + k | X > k) = P(X = n)past failures change nothing

\[ \text{Example 4.18: } \mu \approx 2.86, \quad P(X \ge 3) = 0.65^{2} = 0.4225 \]

The model is exactly right when its four characteristics hold, and misleading when trials learn or wear. The companion text, Introductory Business Statistics, treats the same distribution with the same two cases. Next, Section 4.5 drops independence altogether: sampling without replacement from two groups gives the hypergeometric distribution.

Three corrections to the printed section are worth carrying away: fewer than n is one minus q to the n minus one, the Case II standard deviation equals the Case I one, and the first bar of Figure 4.4 is p itself, 0.02.

OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution §4.4, pp. 242-247 — Section 4.4

OpenStax Introductory Business Statistics 2e, §4.3 Geometric Distribution §4.3 — Geometric Distribution

Sources

  1. OpenStax Introductory Statistics 2e, §4.4 Geometric Distribution — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0, pp. 241-248
  2. OpenStax Introductory Business Statistics 2e, §4.3 Geometric Distribution — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0

Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.

Book on Wyzant · Text (657) 465-8108