The number of successes in n attempts: binomial distribution and normal approximation

18 min · 2 drills

In the textbook: Ch. 5.5 Die Binomialverteilung

Exactly k, at most k, at least k: that is how the number of successes in n independent trials is asked about.

How many drivers are not wearing a seat belt, how many invited people come: each person here is a separate attempt with the same probability of success. For a dozen attempts a formula gives the answer; for hundreds the bell curve of the normal distribution takes its place.

In the evening 6 guests come into a small café, and each orders a dessert with probability 0.3, independently of the others.

What is the probability that exactly two of them order a dessert?

The first thought: two order, so 0.3⋅0.3=0.09. But the other four order nothing, and that has to be multiplied in too. And the two can be any two of the six: there are 15 such pairs, each a case of its own.

One such case, say the first and second guests take dessert and the others do not: 0.32⋅0.74=0.09⋅0.2401=0.021609. All 15 pairs have the same chance, so exactly two order dessert with probability 15⋅0.021609≈0.3241, about one evening in three.

Check yourself. A task at exam level: an office has 7 printers, and each breaks down on a given day with probability 0.15, independently of the others. What is the probability that exactly one printer breaks down in a day? Round to four decimal places.

The basics: the bell curve, standardising and the table Φ

While there are few attempts, the number of successes is computed with a formula. With hundreds, three things from the article “Normal distribution: probability as area” are needed. Here they are on the parcels from there: the delivery time is normally distributed, with mean μ=48 hours and σ=8 hours.

  • Area: for a continuous quantity the probability is an area under the curve. The whole area is 1; a single point has no area.
  • Standardising: z=x−μσ says how many standard deviations x lies from the mean. For 58 hours z=58−488=1.25.
  • Table Φ (“Tabelle 1” on the formula sheet): Φ(z) is the area to the left of z, and 1−Φ(z) is what remains on the right. A share of 1−Φ(1.25)=1−0.894=0.106 of the parcels takes longer than 58 hours.

Delivery time is continuous. The number of successes, by contrast, is a whole number: 0, 1, 2 and so on, and each value has its own probability. That whole number is where we start. The bell curve comes back when there are hundreds of attempts.

The number of successes in several independent attempts: exactly, at most, at least

The café from the start is exactly such a task: 6 guests, each with the same chance 0.3. In the exam drivers take the place of the guests, and there are already 13 trials:

Exam-style task

Bei einer Verkehrskontrolle ist jeder kontrollierte Autofahrer unabhängig von den anderen mit Wahrscheinlichkeit 11 % nicht angeschnallt. Die Polizei hält 13 Autos an. Wie groß ist die Wahrscheinlichkeit, dass genau 2 der Fahrer nicht angeschnallt sind? Runden Sie auf vier Nachkommastellen.

  1. a)0.1187
  2. b)0.3405
  3. c)0.4610
  4. d)0.4270
  5. e)0.2619

There are n identical independent attempts (13 drivers), each with the same probability of “success” p=0.11, and the question is about the number of successes. Here the quantity is whole again, 0, 1, …, 13, and no area is needed.

Where the formula comes from. Take one order: drivers 1 and 2 wear no seat belt, the other 11 do. The attempts are independent, so the probability of this order is a product: 0.112⋅0.8911. But any two of the 13 can be the ones without a belt, and each such order has the same probability. How many ways there are to choose 2 of 13 is the binomial coefficient (132)=13⋅122=78 from the article “Probability as a share: how many cases fit”. Hence the probability of exactly k successes out of n

P(X=k)=(nk)pk(1−p)n−k

Here X is the number of successes, n the number of attempts, p the probability of success in one attempt and k the value of X asked about. On the sheet the number of successes is called Sn: there it reads P(Sn=k). It also has E(Sn)=n⋅p and V(Sn)=n⋅p⋅(1−p).

For the café from the start: n=6 guests, p=0.3, k=2. P(X=2)=(62)⋅0.32⋅0.74=15⋅0.09⋅0.2401=0.3241: exactly two order dessert on about a third of evenings. The first idea, 0.09, forgot both the four without dessert and the 15 pairs.

Without seat belt, of 13

n = 13, p = 0.11, k = 2

Binomial distribution for n = 13 and p = 0.11; the bar k = 2 marked0123456
Step 1 of 3Fix n and p

n=13 cars, a “success” is a driver without a seat belt, p=0.11, 1−p=0.89, k=2. In the picture each k from 0 to 13 has a bar of height P(X=k).

Why would the answer without the factor 78 be exactly 78 times too small? Hint: how many different pairs of drivers could have been the ones without a belt.

At most and at least. “At most k” is the sum of the bars from 0 to k, “at least k” from k to n. It is convenient to compute the side with fewer bars: “at least one” is 1−P(X=0)=1−(1−p)n, and “more than one” is 1−P(X=0)−P(X=1).

Each of the 10 orders of an online shop arrives damaged independently with probability 5 %.

With what probability are more than one damaged?

Binomial distribution for n = 10 and p = 0.05; bars 2 to 10 marked, 0 and 1 outlined210103456789

P(2 ≤ X ≤ 10) ≈0.0861

Each bar is the probability of exactly that number of damaged orders out of 10. Compare the marked bars from 2 to 10 with the outlined 0 and 1: together they make 1, and the two outlined ones are easier to compute.

“More than one” is X≥2, nine bars. The opposite is shorter: X=0 or X=1.

P(X=0)=0.9510=0.59874, P(X=1)=10⋅0.05⋅0.959=0.31512.

1−0.59874−0.31512=0.08614≈0.0861.

Each of 6 emails lands in spam independently with probability 0.08. With what probability does at least one land in spam? The opposite: none, 0.926=0.6064. So at least one lands with 1−0.6064=0.3936. This is the same “at least once” formula as in the article on independence.

Your turn1 of 2

Each of 12 flights is delayed independently with probability 0.1. With what probability is exactly one flight delayed? Round to four decimal places.

Traps. The wrong options of the exam task and of tasks of this kind:

  • Neighbouring k, option (a), 0.1187: this is P(X=3). Check: in the formula the exponent of p is the k of the question.
  • “At least” instead of “exactly”, option (d), 0.4270: this is P(X≥2). “Exactly 2” is one bar.
  • Without the binomial coefficient: 0.112⋅0.8911, one order instead of all 78.
  • p and 1−p swapped: 0.892⋅0.1111. Success is what is being counted.
  • Complement forgotten: for “at least one”, 0.8913 is written, the probability of “none”.
  • Only one term: for “at most 1”, only P(X=1) or only P(X=0) is taken.
  • Wrong n: for example, 12 trees over 3 years are n=36 attempts, not 12.
  • Linear: 13⋅0.11 is the mean number of successes n⋅p, not a probability.
  • “More than 1” and “at least 1”: for a whole number “more than 1” is X≥2, “at least 1” is X≥1.

On the formula sheet: (nk), P(Sn=k)=(nk)pk(1−p)n−k, μ=E(Sn)=n⋅p and σ2=n⋅p⋅(1−p). The sheet writes Sn, the task usually X.

Many attempts or terms: normal approximation for a count and a sum

With 13 drivers the bars can still be added one by one. With hundreds of people and a question about “more than”, you would have to add hundreds of bars, and this is where the bell from the basics comes back:

Exam-style task

Mehr als 35 Fälle mit Fieber am nächsten Tag kann die Gesundheitsstelle einer Universität nicht betreuen. Bei einer Impfaktion lassen sich 800 Studierende impfen; 95 % vertragen die Impfung ohne Beschwerden, die übrigen bekommen leichtes Fieber, jede Person unabhängig von den anderen. Mit welcher Wahrscheinlichkeit gibt es mehr Fieberfälle, als die Stelle betreuen kann?

  1. a)0.791
  2. b)0.761
  3. c)0.763
  4. d)0.548
  5. e)0.767

The number of people is large, 800, and the question is “more than”. With the binomial formula you would have to add 765 bars, one for each number of fever cases from 36 to 800. For large n the bars of the binomial distribution settle under a bell with the same mean and the same spread, and the area is taken from the table. The approximation is good when n⋅p⋅(1−p)>9, here 800⋅0.05⋅0.95=38.

Continuity correction. The bar of the number 36 covers the stretch from 35.5 to 36.5 on the axis. For a whole number, “more than 35” means “at least 36” and starts at the left edge of this bar, so the area under the bell is taken from 35.5 on. “At most k”, on the other hand, includes the whole bar k and ends at its right edge, at k+0.5. On the sheet this is the line P(Sn≤k)≈Φ(k+0.5−μσ).

Fever cases, of 800

μ = 40, σ ≈ 6.16

Binomial distribution for n = 800 and p = 0.05 with the bell curve over it; area from 35.5 on4018243036485460

Excerpt of table 1: row −0.7, columns 0.02, 0.03 and 0.04; Φ(−0.73) marked0.020.030.04−0.70.2360.2330.230
Step 1 of 3μ and σ of the count

What is counted is the fever cases. Each person gets a fever with probability p=1−0.95=0.05, and n=800. So μ=n⋅p=40, σ2=n⋅p⋅(1−p)=38 and σ=6.164.

Why is the centre more likely overwhelmed than not, although 95 % tolerate the vaccine without complaints? Hint: compare 35 with the mean.

A sum of many independent terms. The same bell fits any sum S=X1+⋯+Xn of many independent identical quantities: the number of successes is such a sum too, of zeros and ones (the block “Why is that?” above). Another example is the service time of n customers. From the article on expected value and variance: if one term has mean μ1 and standard deviation σ1, then E(S)=n⋅μ1, V(S)=n⋅σ12, so σ(S)=σ1⋅n, not n⋅σ1. There is no correction here: time is continuous.

A cashier at a supermarket till serves a customer in 4 minutes on average, σ1=1.5 minutes, and customers are independent.

With what probability do 36 customers take more than 150 minutes?

n=36 customers; the sum of their times is S.

E(S)=36⋅4=144 minutes, σ(S)=1.5⋅36=9 minutes.

z=150−1449=0.67, Φ(0.67)=0.749, to the right 1−0.749=0.251.

Try it: at least 190 by card. Each of 300 customers of a shop pays by card independently with probability 0.6. With what probability do at least 190 pay by card? μ=300⋅0.6=180, σ=300⋅0.6⋅0.4=72=8.485. “At least 190” starts at the left edge of the bar, at 189.5. Standardise and read table 1.

Your turn2 of 2

Of 300 visitors to a website, each buys independently with probability 0.1. With what probability are there more than 36 buyers? Use the normal approximation.

Traps. The wrong options of the exam task and of tasks of this kind:

  • Without correction, option (a), 0.791: bound 35 instead of 35.5, z=−0.81.
  • Correction on the wrong side: bound 36.5, z=−0.57, 1−0.284=0.716. “More than 35” includes the bar 36 whole.
  • Counting with 95 %: what is counted is the fever cases, so p=0.05; the 95 % in the task is the complementary share.
  • Complement taken: 0.233, the share of cases in which the centre can look after everyone.
  • n⋅σ1 instead of σ1⋅n: for a sum the spread grows with the square root of the number of terms.
  • Spread without 1−p, option (b), 0.761: σ=800⋅0.05=6.325, z=−0.71. The sheet has σ2=n⋅p⋅(1−p).
  • Variance instead of σ, option (d), 0.548: dividing by 38 instead of 6.164, z=−0.12.
  • Wrong number of terms: 36 customers are 36 service times. The gaps between the arrivals of the first and the 36th customer are 35, which is a different task.

On the formula sheet: μ=E(Sn)=n⋅p, σ2=n⋅p⋅(1−p), P(Sn≤k)≈Φ(k+0.5−μσ) and Tabelle 1.

At a glance: the number of successes and the bell curve

The topic in a minute

  • Binomial distribution: n independent attempts with probability of success p, the number of successes k is asked:
P(X=k)=(nk)pk(1−p)n−k
  • Mean and spread of the number of successes:
μ=n⋅p,σ2=n⋅p⋅(1−p)
  • At most and at least: a sum of bars; compute the side with fewer bars. “At least one” is 1−(1−p)n.
  • Approximation: for a large n take the bell curve with the correction ±0.5, “at most k” up to k+0.5, “at least k” from k−0.5. It fits when
n⋅p⋅(1−p)>9
  • Sum of n independent terms: mean n⋅μ1, spread σ1⋅n, no correction.
  • Binomial distribution

    n identical independent attempts, the number of successes asked: exactly, at most, at least.
    P(X=k)=(nk)pk(1−p)n−k

    Trap: forgetting the binomial coefficient.

  • Normal approximation

    A large n for a number of successes, or a sum of many independent quantities.
    P(Sn≤k)≈Φ(k+0.5−μσ)

    Trap: forgetting the correction 0.5 or shifting it the wrong way.

The chapter's tasks look alike, and one or two words give the method away; they are collected in the article “Which method? Probability”.

Now practise it

The number of successes in n attempts: binomial distribution and normal approximation

Trainer · Probability

Practise
Next: Which method? Probability
0/2

Was this article helpful?