Binomial Distribution (Binomialverteilung)

40 min · 15 drills

The tasks in this topic describe a process that repeats: items are inspected, customers approached, answers guessed. Below it stand five claims about the number of successes, its probability, the expected value, or whether the model fits at all. There is little to compute; what you have to decide is which computation a claim calls for and where it swaps that computation for a convenient wrong shortcut.

The formula sheet (German: Formelsammlung) contains only the binomial coefficient (nk)=n!k!(n−k)!, neither the formula for k successes nor the expected value and the variance. If the task text does not print the formula, you assemble it from the product rule and this coefficient; the expected value and the variance you bring along from memory.

A Bernoulli chain requires four things at once

Bernoulli trial

A Bernoulli trial is a random experiment with exactly two outcomes. One outcome is called a success (German: Treffer). The other is called a failure, and the task decides which is which: a success need not be anything pleasant, a defective item can be one too. The probability of a success is called p; the probability of a failure is then 1−p.

If an experiment has more than two outcomes, you merge them into two groups. With a die, “5 or 6 spots” becomes the success with p=26=13, and the other four faces form the failure. Every outcome must belong to exactly one of the two groups.

Bernoulli chain

A Bernoulli chain (German: Bernoulli-Kette) of length n consists of n Bernoulli trials for which three further conditions hold: the number n is fixed before the start, the success probability p is the same in every trial, and the trials do not influence each other. The random variable X counts the successes and can take the values 0,1,…,n.

What is fixed in advance is the number of trials, not the number of successes. Each of the four conditions can fail on its own, and in the tasks each of them appears in disguise.

ConditionHow it fails
two outcomesnever for good: several outcomes can be merged into success and failure
fixed n"until the first success arrives": then the number of trials is itself random
same pa machine wears out, an apprentice learns with every item
independencea group confers before each person votes; items are drawn without replacement from a small stock

Your turn1 of 15

A pupil answers 10 questions independently of each other: each of the first five correctly with probability 0.9, each of the last five with 0.4. May you compute P(X=7) for the number X of correct answers with the binomial formula for n=10 and the average p=0.65?

Drawing without replacement (German: ohne Zurücklegen) breaks independence: every item drawn changes the stock and with it the probability for the next one. A box holds 10 tickets, 5 of them winners, and 2 tickets are drawn. Drawing no winner has, without replacement, the probability

510⋅49=29≈0.2222

With replacement it would be 0.52=0.25. Without replacement every losing ticket drawn makes the next one rarer, and the binomial distribution would compute this wrongly.

If instead 10 000 tickets with 5 000 winners lie in the drum, one losing ticket drawn barely changes the share: 5 00010 000⋅4 9999 999≈0.249975 instead of 0.25. Hence the rule of thumb: if the sample is small compared with the stock, you may compute without replacement as if items were replaced. A survey of 80 randomly chosen people from a whole country is therefore a Bernoulli chain, a poll of 12 of the 30 members of a club is not.

Your turn2 of 15

Which situation can be described as a Bernoulli chain?

The formula is built from the product rule and the binomial coefficient

A hotel has 7 reservations for one night. Each is cancelled independently of the others with probability 0.3, and X counts the cancellations. We want P(X=2).

  1. One particular sequence. That exactly the first two reservations are cancelled and the other five are not has, by the product rule for independent events, the probability 0.3⋅0.3⋅0.75=0.32⋅0.75=0.0151263.
  2. Every other sequence is equally likely. If the two cancellations sit at other positions, say the third and the sixth, the same factors appear in a different arrangement: 0.3 twice and 0.7 five times. The product does not change.
  3. Count the sequences. An sequence with two cancellations is fixed as soon as you decide which 2 of the 7 positions are cancelled. There are (72)=7!2!⋅5!=7⋅62=21 ways to do that, and this is the formula from the formula sheet.

The different sequences exclude each other. Their probabilities are therefore added, and 21 equal terms give

P(X=2)=21⋅0.32⋅0.75≈0.3177

Replace 7 by n, 2 by k and 0.3 by p, and the general formula appears.

Binomial distribution

If X counts the successes in a Bernoulli chain of length n with success probability p, then X is called binomially distributed. For k=0,1,…,n

P(X=k)=(nk)⋅pk⋅(1−p)n−k

Each factor has a job, and that lets you check a printed expression quickly. (nk) counts the sequences, pk stands for the k successes and (1−p)n−k for the remaining failures. The two exponents always add up to n; an expression such as (72)⋅0.32⋅0.74 can therefore be recognised as wrong from the exponent alone.

Some values need no calculator: (n0)=(nn)=1, because there is only one sequence without successes and only one made entirely of successes, and (n1)=n, because a single success can sit at any of the n positions. In general (nk)=(nn−k): choosing k positions for the successes means choosing the n−k positions for the failures.

Your turn3 of 15

The expression P(X=3)=(93)⋅0.353⋅0.656 contains a binomial coefficient. What does (93) stand for?

For the hotel all eight values can be computed this way. The table shows them rounded to four places.

kP(X=k)
00.0824
10.2471
20.3177
30.2269
40.0972
50.0250
60.0036
70.0002

The sum is 1, up to rounding in the last place, because some number of cancellations is certain to occur. If a task claims that a few values already add up to 1, check whether every value from 0 to n is really included; if one with positive probability is missing, the sum stays below 1.

Your turn4 of 15

For a Bernoulli chain with n=4 the following are known: P(X=0)=0.4096, P(X=1)=0.4096, P(X=2)=0.1536 and P(X=3)=0.0256. What is P(X=4)?

No success, at least, at most: special cases and the complement

P(X=0)=(1−p)nP(X=n)=pnP(X≥1)=1−(1−p)n

For X=0 every single trial must be a failure, and there is only one sequence for that; the product rule is enough. The same holds for X=n with nothing but successes. For the hotel, P(X=0)=0.77≈0.0824 and P(X=7)=0.37≈0.0002.

Complementary event

The complementary event (German: Gegenereignis) of an event A occurs exactly when A does not occur. Its probability is 1−P(A).

“No success” has exactly one complement, “at least one success” (German: mindestens ein Treffer): either none of the trials is a success, or at least one is. For the hotel, P(X≥1)=1−0.77≈0.9176. “Exactly one success” is not the complement: together with “no success” it has only 0.0824+0.2471≈0.329, and the rest belongs to two or more successes.

The most convenient wrong computation for “at least one success” is n⋅p. For the hotel it would give 7⋅0.3=2.1, more than 1, and no probability can be that. Even where n⋅p does not exceed 1, the computation is wrong: for six throws of a die, 6⋅16=1 would mean a six is certain, while the true value is 1−(56)6≈0.6651. The product n⋅p is the expected value.

Throws nP(at least one six)
10.1667
20.3056
30.4213
40.5177
60.6651
120.8878

Twice as many throws therefore do not bring twice the probability, and it never reaches 1, because (56)n never becomes zero. If a task asks from how many trials on “at least one success” exceeds a bound, try n one after another: the die exceeds 12 for the first time at n=4, because (56)3≈0.5787 is still above 12 and (56)4≈0.4823 is already below.

The formula for X=0 can also be read backwards. If for n=3 it is known that P(X=0)=0.343, then 1−p=0.3433=0.7 and hence p=0.3; no other p gives this value. If the chain is made twice as long, P(X=0) is not halved but squared: 0.3432=0.117649.

Your turn5 of 15

A commuter takes a train on four days; on each day, independently of the others, it is late with probability 0.15. How likely is at least one delay?

The words of the task translate into ranges of k, and at the boundary a single word decides whether the boundary value belongs to the range.

WordingEventComputation
exactly k (genau)X=kone value of the formula
at most k (höchstens)X≤kP(X=0)+P(X=1)+⋯+P(X=k)
fewer than k (weniger als)X≤k−1as above, but without P(X=k)
at least k (mindestens)X≥k1−P(X≤k−1) or summed directly up to n
more than k (mehr als)X≥k+11−P(X≤k)

For the hotel, “at most two cancellations” covers the cases k=0, 1 and 2. They exclude each other, so their probabilities are added: 0.0824+0.2471+0.3177≈0.647. Multiplying would be wrong here, because the product rule applies to events that occur together, and there are no two different numbers of cancellations in the same night.

“At least three cancellations” is its complement and has 1−0.647≈0.353. The detour through the complement pays off whenever it needs fewer terms; done directly, there would be five here. If k lies close to n, the direct way is shorter: for 10 coin tosses, “at least nine heads” has only the terms k=9 and k=10, that is (109)+(1010)210=111 024.

Your turn6 of 15

In a Bernoulli chain of length n=12, the value P(X≤3) is known. Which event has the probability 1−P(X≤3)?

The distribution is symmetric only when p = 0.5

The binomial coefficient is symmetric, (nk)=(nn−k). It does not follow that k and n−k successes are equally likely, because the powers differ. For the hotel, two and five cancellations each have 21 sequences, yet P(X=2)≈0.3177 and P(X=5)=21⋅0.35⋅0.72≈0.0250. Likewise “all cancelled” and “none cancelled”, at 0.0002 and 0.0824, are rare to very different degrees.

Only at p=0.5 do p and 1−p coincide. Then every sequence has the same probability 12n, and P(X=k) is the share of sequences with k successes. In ten coin tosses, three heads and seven heads each have (103)=(107)=120 of the 1 024 sequences. Five heads has 2521 024≈0.2461, not 12: half successes is the most frequent single case, but not a case with probability 12.

What always holds is a correspondence between two points of view. k successes at success probability p are the same event as n−k failures, and a failure has probability 1−p. Exactly two cancellations mean exactly five reservations that stand, both at 0.3177. It is wrong, however, to swap p and 1−p in the formula and leave k as it is: (72)⋅0.72⋅0.35 is not P(X=2) but P(X=5)≈0.0250.

Your turn7 of 15

For which Bernoulli chain does P(X=3)=P(X=6) hold?

The most likely value lies near the expected value but is a whole number

In the hotel table k=2 has the largest probability, while n⋅p=2.1. The most likely value lies near n⋅p but is a whole number, because X takes only whole numbers; a claim that it is 2.1 is wrong from its form alone. Nor does it have to lie at half of n. The expected value is not rounded for this: E(X)=2.1 remains the mean, which need not be a whole number, and the most likely value comes from the calculation further down, not from rounding n⋅p.

Most likely value

The most likely value of a distribution is the k for which P(X=k) is largest. There can be two such values side by side if their probabilities are exactly equal.

Whether the distribution rises or falls from k to k+1 is shown by the quotient of two neighbours: if it is greater than 1, the distribution rises; if smaller, it falls; if exactly 1, the two neighbours are equally likely. Almost all the powers cancel in it:

P(X=k+1)P(X=k)=n−kk+1⋅p1−p

For the hotel the quotient from k=1 to k=2 is 62⋅37=1814, so greater than 1, and from k=2 to k=3 it is 53⋅37=1521, so smaller: the distribution rises up to k=2 and falls after it.

The quotient yields a shortcut: the most likely value is (n+1)⋅p, rounded down to a whole number. If (n+1)⋅p is itself a whole number, that value and the one below it share the maximum. For the hotel 8⋅0.3=2.4, so k=2; for n=7 and p=0.25 it is 8⋅0.25=2, so k=1 and k=2.

Your turn8 of 15

A Bernoulli chain has length n=11 and success probability p=13. Which value of X is the most likely?

The expected value is an average over many repetitions

Expected value

The expected value (German: Erwartungswert) of a binomially distributed number of successes states how many successes come out on average when the whole chain is repeated very often:

E(X)=n⋅p

The result is plausible: in the long run the share p of the n trials becomes a success. For the hotel E(X)=7⋅0.3=2.1. In a single night, though, 0, 1, 2 or more reservations are cancelled, never 2.1. The expected value need not be a whole number, and for a single chain it promises nothing, neither exactly nor at least that many successes. Nor is it a probability: if n⋅p comes out as a whole number m, that does not make the case X=m certain, nor does it give it the probability p.

Because E(X) grows in the same ratio as n and as p, the expected value doubles when n is doubled, and it halves when p is halved.

If each cancellation costs the hotel 40 euros, the expected loss per night is 2.1⋅40=84 euros. An insurance that replaces every cancellation loss therefore pays off for the hotel on average only if it costs less than 84 euros per night. 7⋅40=280 euros would be wrong: that is the loss if every reservation were cancelled.

Your turn9 of 15

A customer service team calls back 9 customers on every working day. Each of them, independently of the others, picks up with probability 0.7. The team lead first works out how many unanswered callbacks are to be expected on one day and then scales that up to 20 working days. Which line gives the expected number of unanswered callbacks correctly?

Variance and standard deviation measure the spread

Variance

The variance of the number of successes is V(X)=n⋅p⋅(1−p). It measures how strongly X fluctuates around the expected value from chain to chain, in squared units.

Standard deviation

Taking the square root of the variance gives the standard deviation σ=V(X). It has the same unit as X and gives the typical deviation from the expected value.

V(X)=n⋅p⋅(1−p)σ=n⋅p⋅(1−p)

For the hotel V(X)=7⋅0.3⋅0.7=1.47 and σ=1.47≈1.21 cancellations. Variance and standard deviation are two different numbers; they are equal only when the variance is 0 or 1.

How strongly p drives the spread is shown by the factor p⋅(1−p).

p0.10.30.50.70.9
p⋅(1−p)0.090.210.250.210.09

For the same n the variance is largest at p=0.5, and p and 1−p give the same variance.

If n is doubled, the variance doubles, but the standard deviation grows only by the factor 2≈1.41. With 14 reservations instead of 7, V(X)=2.94 and σ=2.94≈1.71. The expected value doubles, the typical deviation grows more slowly, and relative to the expected number the chain thus becomes more predictable.

Your turn10 of 15

A café has 40 glasses in use on an evening. Each breaks independently of the others with probability 0.12, and each broken glass costs 25 euros. What cost is to be expected per evening?

Your turn11 of 15

In a Bernoulli chain n is quadrupled with the same p. How do the expected value and the standard deviation of the number of successes change?

Your turn12 of 15

Chain A has n=100 trials with p=0.5, chain B has n=400 trials with p=0.1. Which claim about the standard deviation of the number of successes is correct?

An acceptance rule is a question about at most c defects

Acceptance rule

An acceptance rule (German: Annahmeregel) states that n items from a delivery are inspected and the delivery is accepted if at most c of them are defective. If each item is defective independently with probability p, the acceptance probability is P(X≤c).

A wholesaler takes 9 items from every delivery and accepts it as long as at most two of them are defective. The acceptance probability is made up of three cases:

P(X≤2)=(1−p)9+9⋅p⋅(1−p)8+36⋅p2⋅(1−p)7
Defect rate pSample of 9Sample of 18
0.050.99160.9419
0.150.85910.4797
0.300.46280.0600

Three things can be read off. The acceptance probability depends on n, p and the limit c at once. With the same limit a larger sample makes the rule stricter, because among more inspected items three defective ones are easier to find. And it is neither 1−p nor the share of admitted values: k=0, 1 and 2 are three of the ten possible values, but 310 has nothing to do with 0.9916.

Your turn13 of 15

An acceptance rule tests 10 items and accepts the delivery if at most one of them is defective. Each item is defective independently with probability 0.1. Which expression is the probability that the delivery is rejected?

Five claims about a Bernoulli chain

Each of the five claims is a decision of its own. Check the model first, then translate the wording into the k that are meant, take the cheap cases P(X=0), P(X=n) and P(X≥1) first, and check in the printed expression whether the exponents add up to n. A complete example. An online shop sends 10 customers one voucher each, every customer redeems it independently of the others with probability 0.4, and X counts the redemptions.

  • With 0.610≈0.0060 not a single voucher is redeemed. Correct. All ten customers must decline, and there is only one sequence for that.
  • At least one redemption occurs with 10⋅0.4. Wrong. The product is 4 and for that reason alone cannot be a probability; the correct value is 1−0.610≈0.9940.
  • Four and six redemptions are equally likely, because (104)=(106). Wrong. The numbers of sequences are equal, the powers are not: P(X=4)≈0.2508 and P(X=6)≈0.1115. The two values would be equal only at p=0.5.
  • Exactly four redemptions are the most likely. Correct. (n+1)⋅p=4.4 rounds down to 4, and comparison with the neighbours confirms it: P(X=3)≈0.2150 and P(X=5)≈0.2007 both lie below 0.2508.
  • The standard deviation is 2.4. Wrong. 10⋅0.4⋅0.6=2.4 is the variance; the standard deviation is 2.4≈1.55.

Your turn14 of 15

From a box of 20 light bulbs, 4 of them defective, an inspector takes 5 without replacement; X counts the defective ones. Which claim is correct?

Your turn15 of 15

A Bernoulli chain has n=6 and p=0.2. Which claim is correct?

Key takeaways

  • Bernoulli chain. Two outcomes, fixed n, same p, independent trials; X counts the successes.
  • The formula. It is not on the formula sheet, but (nk) is; the rest is the product rule for a single sequence.
  • Special cases. P(X=0)=(1−p)n, P(X=n)=pn, P(X≥1)=1−(1−p)n. Compute on the side with fewer terms: summed directly or via the complement.
  • Key figures. E(X)=n⋅p is an average over many chains, V(X)=n⋅p⋅(1−p), σ=V(X); the most likely value is a whole number near n⋅p.
P(X=k)=(nk)⋅pk⋅(1−p)n−k

Now practise it

Binomial Distribution (Binomialverteilung)

Trainer · Mathematics

Practise
0/15

Was this article helpful?