Independence: when probabilities multiply

33 min · 5 drills

In the textbook: Ch. 5.2 Bedingte Wahrscheinlichkeiten

A copy shop has two printers. The first breaks down on a given day with probability 0.1, the second with probability 0.2. The printers are different and stand in different rooms; a breakdown of one has no effect on the other. Printing is impossible only if both are broken.

With what probability can the copy shop not print on a random day?

Adding, 0.1+0.2=0.3, does not help here: it counts the days when at least one is broken, and it counts the days when both are broken twice. Multiplying gives the answer, and only because the printers do not affect each other. Systems of several parts and the tasks on the number of attempts until the first success rest on this multiplying too.

Both printers are broken on the same day with probability 0.1⋅0.2=0.02. On average the copy shop cannot print on two days out of a hundred.

Check yourself. A task at exam level: a warehouse has two smoke detectors; they go off independently of each other, the first with probability 0.9, the second with probability 0.8. With what probability does at least one go off in a fire?

Independent events: knowing about one does not change the chance of the other

Picture 1000 working days of the copy shop. The events:

  • A: printer 1 is broken;
  • B: printer 2 is broken;
  • A′ and B′, the complements: printer 1 works, printer 2 works.

Printer 1 breaks down on about 100 of 1000 days, a share of 0.1. What happens to printer 2 on these 100 days? It does not care whether the first one is broken: it breaks down on the same share of days as always, 0.2. So on 20 of these 100 days both are broken. On the other 900 days printer 2 also breaks down at a share of 0.2, that is 180 days. On the remaining days printer 2 works: 100−20=80 and 900−180=720. This gives a two-way table from the article “Conditional probability: the share within a group”:

Table of the two printers for 1000 days: rows printer 1 broken and working, columns printer 2 broken and working2 broken2 works1 broken1 works2080180720totaltotal1009002008001000

Both are broken on 20 of 1000 days, the probability P(A∩B)=0.02 (what the intersection “and” of events means). This answers the copy shop's question. See where the 20 comes from: it is 0.2 of 100, that is 0.2⋅0.1⋅1000. As shares of all days:

P(A∩B)=0.1⋅0.2=0.02

The key is the phrase “printer 2 does not care”. Look at the share of days on which printer 2 is broken:

  • in the row “1 broken”: 20/100=0.2;
  • in the row “1 works”: 180/900=0.2;
  • among all days: 200/1000=0.2.

The condition “printer 1 is broken” does not change the share of B:

P(B∣A)=P(B)

Such events are called independent (German “stochastisch unabhängig”). The article on conditional probability had the multiplication rule P(A∩B)=P(A)⋅P(B∣A). Put P(B∣A)=P(B) into it and you get the rule for independent events:

P(A∩B)=P(A)⋅P(B)

How to check whether events are independent. Compare the intersection with the product:

  • equal: the events are independent;
  • not equal: dependent, the condition changes the share.

In the online shop from the article on conditional probability, 0.25 of all orders are returned, 0.40 are paid by invoice, and 0.16 are both returned and paid by invoice. The product 0.25⋅0.40=0.10 is not 0.16, so returning and paying by invoice are dependent: among invoice orders returns are more frequent, 0.40 against 0.25 among all.

Complements and “at least one”

Not only the cell “both broken” is a product of shares. In the printer table for 1000 days from the previous section, every cell is the product of its margins. Both work on 720 days: 0.9⋅0.8=0.72. This is no coincidence. If printer 2 does not care whether the first one is broken, it also does not care whether the first one works. So all four pairs of events are independent:

  • A and B, A′ and B′;
  • A and B′, A′ and B.

For example, for “both work”:

P(A′∩B′)=P(A′)⋅P(B′)=(1−0.1)⋅(1−0.2)=0.72

This gives a short way to “at least one”. “At least one printer is broken” is all days except those on which both work. The complement (the formula sheet line P(A′)=1−P(A)) gives

P(at least one broken)=1−0.72=0.28

The addition rule from the last article gives the same: 0.1+0.2−0.02=0.28. The way through “none” is handier when there are more than two parts. For three, five or n independent parts, “none broke down” is the product of the probabilities that they work, one factor per part. The three methods below rest on this: systems, at least once and the first success.

Independent and disjoint: not the same thing

Only independent events may be multiplied. They are often confused with disjoint ones, which is quite a different property. Disjoint events cannot happen together: “printer 1 is broken” and “printer 1 works” on the same day. For them P(A∩B)=0, more in the section “Overlapping or not: disjoint events”. Independent events, on the contrary, do happen together, and exactly with probability P(A)⋅P(B): both printers are broken on 20 days, not on 0.

  • Disjoint: intersection 0, “or” is the sum, P(A∪B)=P(A)+P(B).
  • Independent: the intersection is the product, “or” is P(A)+P(B)−P(A)⋅P(B).
  • Two disjoint events with positive probabilities are never independent: if one happens, the other is impossible, and the condition changes its share very strongly.

The answer 0.1+0.2=0.3 from the start treats the printers exactly as disjoint, as if they never broke down on the same day.

Independent events: “and”, “or” and an unknown share

With the printers both shares were known, and all that was left was to multiply. In the first exam task one share is unknown, but “at least one” is given, and the multiplication turns into an equation:

Exam-style task

Eine Prüfung besteht aus einem Test und einer Hausarbeit, die unabhängig voneinander bestanden werden. Den Test bestehen 62 % der Teilnehmer. Mindestens eines von beiden bestehen Teilnehmer mit der Wahrscheinlichkeit 0.8062. Mit welcher Wahrscheinlichkeit besteht ein Teilnehmer die Hausarbeit?

  1. a)0.61
  2. b)0.19
  3. c)0.49
  4. d)0.68
  5. e)0.30

The words “independently of each other” and the given “at least one of the two” show the way. This is a four-field table like the printer one in the first section, only with one margin unknown: the share of those who pass the term paper. Call the unknown share x. The cell “passed nothing” lies outside “at least one”, and for independent events it too is the product of its margins. That cell gives x.

margins: pass 0.62, fail 0.38, term paper x

Table of test and term paper in shares: every cell is the product of the margins, the cell passed nothing gives the equation for xpaper passedpaper failedtotaltest passedtest failedtotal0.620.38x1 − x1
Step 1 of 4Margins of the table

0.62 pass the test, 1−0.62=0.38 fail it. Call the share of those who pass the term paper x, the rest 1−x.

Why does the cell “test failed, paper passed” hold 0.38⋅0.49 and not simply 0.49? Hint: how many people are in the row “test failed”.

An online shop watches two events: A “the customer uses a coupon”, B “the order is large”. Among large orders 30 % use a coupon; among orders with a coupon 50 % are large. The events are independent.

What share of all orders is large and with a coupon?

The condition does not change the share: P(A∣B)=P(A) and P(B∣A)=P(B).

P(A)=0.3, P(B)=0.5.

P(A∩B)=0.3⋅0.5=0.15. Check: among orders with a coupon, 0.15/0.3=0.5 are large, as stated.

Independent or not: cake and card. A café has 200 guests in a day. 80 take cake, 125 pay by card, 50 do both. Are “cake” and “card” independent? Shares: P(cake)=80/200=0.4, P(card)=125/200=0.625, P(both)=50/200=0.25. Compare the intersection with the product.

Your turn1 of 5

The events A and B are independent, P(A)=0.4 and P(B)=0.35. What is P(A∪B)?

Traps. The wrong options of the exam task come about like this:

  • Disjoint instead of independent, option (b), 0.19: 0.8062−0.62=0.1862, as if the addition rule had no intersection. That would hold if nobody passed both. Check: for independent events the intersection is 0.62⋅x, not 0.
  • Wrong margin, option (e), 0.30: 0.1862/0.62. 0.1862 is the cell “test failed, term paper passed”: only those who failed the test land in it, who are 0.38, and the division is by 0.38.
  • Union instead of intersection, in tasks of this kind: “both” is asked, and P(A)+P(B)−P(A)⋅P(B) is computed. Check: “and” is never larger than the smaller of the two probabilities.
  • Product instead of quotient, in tasks of this kind: the intersection and P(A) are given, the share within A is asked, and the intersection is multiplied by P(A). A share within a group is a division: P(B∣A)=P(A∩B)/P(A).

On the formula sheet: P(A∪B)=P(A)+P(B)−P(A∩B), P(A′)=1−P(A) and P(A∣B)=P(A∩B)P(B). The rule P(A∩B)=P(A)⋅P(B) for independent events is not on the sheet; you remember it.

Does a system of independent parts work

The test and the term paper were two separate events. In the next exam task independent parts work as one system: two scanners check the same attachment:

Exam-style task

Eine Firma prüft E-Mail-Anhänge mit zwei unabhängig arbeitenden Virenscannern. Scanner 1 erkennt Schadsoftware mit 79 % Sicherheit, Scanner 2 mit 60 % Sicherheit; 2 von 1000 Anhängen enthalten Schadsoftware. Bei einem schädlichen Anhang schlug mindestens einer der beiden Scanner an. Wie wahrscheinlich ist es, dass Scanner 2 angeschlagen hat?

  1. a)0.9295
  2. b)0.8624
  3. c)0.8194
  4. d)0.6550
  5. e)0.9622

Two parts work “independently”, and it is known that “at least one of the two” raised an alarm. That is the situation of a system where one of two parts is enough. The question, though, is not how likely the system raises an alarm but a share under this condition: of all cases where at least one scanner fires, how many belong to scanner 2? The figure “2 in 1000 attachments” is not needed for this, because the question already speaks of a harmful attachment. The task is solved in example 3 of this section; first come the cases of a system made of several parts.

At most one part fails. A waterworks runs three identical pumps, each of which fails within a year with probability 22 %, independently of the others. Operation continues if at most one pump fails. With what probability is that the case in a year? As with the printers, you go through the cells, only here there are three parts, and the table becomes a list of cases, one row per case. The probability of a row is a product, one factor per pump.

At most one: none or exactly one

Cases for three pumps: no failure and exactly one failure, each row the product of three probabilities; two sums belowpump 1pump 2pump 3probabilityno failure1 fails2 fails3 failsexactly one, togetherat most one0.780.780.78?0.220.780.78?0.780.220.78?0.780.780.22???
Step 1 of 4Write out the cases

“At most one pump fails” is two kinds of cases: no failure or exactly one failure. A pump works with probability 1−0.22=0.78.

Why does “exactly one failure” have three terms and not one? Hint: the task does not say which pump fails.

If all n parts have the same failure probability f, the steps fold into one line:

P(at most one failure)=(1−f)n+n⋅f⋅(1−f)n−1

For the pumps: 0.783+3⋅0.22⋅0.782=0.8761. If the parts have different probabilities, you do not use the formula but write one row per case.

The parts of a system do not always fail equally often: in a sorting centre three conveyors fail independently, with probabilities 10 %, 20 % and 5 %. Sorting runs if at most one has failed.

With what probability does it run?

No failure or exactly one. The conveyors work with probabilities 0.9, 0.8 and 0.95.

0.9⋅0.8⋅0.95=0.684.

One row per conveyor: 0.1⋅0.8⋅0.95+0.9⋅0.2⋅0.95+0.9⋅0.8⋅0.05=0.076+0.171+0.036=0.283.

0.684+0.283=0.967.

In series and in parallel. Sometimes a system needs all of its parts, sometimes one of several is enough:

  • all needed (in series): the system works only if all work, which is a product; a pump 0.9 and a filter 0.95 give 0.9⋅0.95=0.855;
  • one is enough (in parallel): the system fails only if all fail; two spare pumps at 0.8 each give 1−0.2⋅0.2=0.96;
  • a mixed layout is computed in parts: first the parallel group, then the product along the chain. Below: pump A, then spare pumps B and C in parallel, drawn as B∥C, and filter D.
Pump A, then two pumps B and C in parallel, then filter D: the chain works only if every link works, the pair B and C if at least one doesA0.9B ∥ C = 0.96B0.8C0.8D0.95

0.9 · 0.96 · 0.95 =0.8208

Here the group of B and C works with probability 0.96, and the whole chain with 0.9⋅0.96⋅0.95=0.8208.

This is the task from the sheet above. For a harmful attachment scanner 1 fires with probability 0.79 and scanner 2 with 0.6, independently of each other. At least one fired. How likely is it that scanner 2 was among them?

“At least one fires” is the parallel pair: it fails only if both stay silent. Scanner 1 stays silent with 1−0.79=0.21, scanner 2 with 1−0.6=0.4, both together with 0.21⋅0.4=0.084. So at least one fires with 1−0.084=0.916.

What you need is “scanner 2 fires” under the condition “at least one fires”. If scanner 2 fires, then at least one has fired anyway. The intersection of the two events is therefore simply “scanner 2 fires”, with probability 0.6.

A share under a condition is the intersection divided by the condition, as in P(A∣B)=P(A∩B)P(B) on the formula sheet: 0.60.916=0.65502≈0.6550, option (d). The number is larger than 0.6, because the news “at least one fired” rules out the case in which both stay silent. Putting scanner 1's rate in the numerator, 0.790.916=0.8624, answers the question about the other scanner, option (b).

A server room has four identical servers; each fails in a month independently with probability 0.1. The service runs if at most one has failed. No failure: 0.94=0.6561. Exactly one failure: 4⋅0.1⋅0.93=0.2916. Together 0.6561+0.2916=0.9477. 0.6561+0.0729=0.729 comes out when one term is taken for “exactly one” without the factor 4.

Your turn2 of 5

A cold store has five cooling units; each fails independently with probability 10 %. The cooling is enough if at most one has failed. With what probability is that so? Round to four decimal places.

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

  • The other rate in the numerator, option (b), 0.8624: 0.790.916 is the probability for scanner 1. Check: the numerator holds the part the question asks about.
  • Ignoring the condition: simply answering 0.6. That would be right if nothing were known about the scanners; but “at least one fired” is information, and it raises the chance.
  • Only “no failure”: for the pumps 0.783=0.4746, and the cases with one failure are lost. “At most one” includes exactly one.
  • Wrong bound: “at most two” is computed instead of “at most one”. Check: reread how many failures the system tolerates, and write out the cases before calculating.
  • Condition multiplied: “pump 1 has already failed; with what probability does pump 2 fail?” The failure of pump 1 has already happened, and under independence the answer is simply 0.22, not 0.22⋅0.22=0.0484.
  • Complement: the answer 1−0.8761=0.1239, the probability that the system does not run.

This method is not on the formula sheet: only P(A′)=1−P(A), to go from failing to working.

First success on the k-th try: throws until the first hit

In the system the parts worked at the same time. In the next task the same trial repeats over time, and the question is when the first success comes:

Exam-style task

Bei einer Tombola gewinnt jedes Los unabhängig mit der Wahrscheinlichkeit 0.1. Eine Besucherin zieht so lange Lose, bis sie zum ersten Mal gewinnt; W ist die Nummer des Gewinnloses. Bestimmen Sie die Wahrscheinlichkeit, dass sie weniger als 4 Lose braucht, also P(W<4). Runden Sie auf vier Nachkommastellen.

  1. a)0.3439
  2. b)0.2710
  3. c)0.2973
  4. d)0.2439
  5. e)0.7290

One and the same attempt is repeated “independently”, and the text says “until she wins for the first time”. W is the number of the first successful attempt. These are the same parts of a system, only one after another in time: attempts instead of pumps.

The key is one thought. The first win comes after the third ticket, W>3, so the first three tickets all lost. If a ticket wins with probability p, then losing, the complement, has probability q=1−p. The attempts are independent, so for any number of tickets k

P(W>k)=qk,P(W≤k)=1−qk
Each bar is the probability that the first win falls on exactly this ticket; the marked bars 1, 2 and 3 together give the answer123456780.050.1number of the first winning ticketwin: p = 0.1, loss: q = 0.9
Step 1 of 4Find p and q

Success is a win: p=0.1, loss q=1−0.1=0.9. The bar above number k is the probability that the first win falls exactly on ticket k.

Each next bar is the previous one times q: to win exactly with the k-th ticket you first need k−1 losses, P(W=k)=qk−1⋅p. Such a W has a geometric distribution (German “geometrische Verteilung”).

Before calculating, translate the wording into a bound for W:

Wordingeventcalculation
at most k attempts, up to the k-th inclusiveW≤k1−qk
fewer than k attemptsW≤k−11−qk−1
more than k attemptsW>kqk
at least k attemptsW>k−1qk−1
exactly on the k-thW=kqk−1⋅p

At a hotline each call gets through with probability 0.2. With what probability do you get through for the first time no earlier than the 2nd and no later than the 4th call? That is 2≤W≤4, so “no later than the 4th” minus “right on the 1st”. No later than the 4th: 1−0.84=0.5904. On the 1st: 1−0.8=0.2. The difference is 0.5904−0.2=0.3904. Check with the bars: 0.8⋅0.2+0.82⋅0.2+0.83⋅0.2=0.16+0.128+0.1024=0.3904.

Your turn3 of 5

A basketball player scores with each shot from half court independently with probability 0.25. With what probability does she need fewer than 5 shots until the first hit?

Traps. The wrong options of the exam task:

  • Strict bound read as non-strict, option (a), 0.3439: 1−0.94 answers “at most 4 tickets”. “Fewer than 4” for a whole number is W≤3.
  • Complement mixed up, option (e), 0.729: this is 0.93, “the first three lost”, that is P(W>3). The last step, subtracting from 1, is missing.
  • Exactly instead of at most, in tasks of this kind: 0.92⋅0.1=0.081 is the win exactly with the third ticket. “Fewer than 4” includes the 1st and the 2nd.
  • n times p, in “at least once” tasks: 10⋅0.07=0.7 is the mean number of successes. For large n it can exceed 1, a probability never can.
  • Rounding n down: at n=14 the probability does not yet reach 0.9. The smallest n is always rounded up.

1 − (1 − p)ⁿ: at least once in n tries

W≤k means “won at least once in the first k attempts”. So the same formula answers the question “at least once in n attempts”: 1−(1−p)n.

In a sales team a call ends in a sale with probability 0.15, independently of the others.

With what probability does at least one of 6 calls end in a sale?

p=0.15, n=6, failure q=0.85.

“At least one sale” is the opposite of “none”: all 6 calls without a sale.

0.856=0.3771.

1−0.3771=0.6229. The number 6⋅0.15=0.9 is the mean number of sales in 6 calls, not a probability.

At least one winner in a prize draw. The prize draws of the article “Probability as a share: how many cases fit” are also asked this way: with what probability does at least one of N participants win? This is the same method. One participant wins with probability p1: one divided by the number of selections. If 3 balls are drawn from 10, there are (103)=120 selections and p1=1120. A participant loses with probability 1−p1. The participants choose independently of each other, so all N lose with probability (1−p1)N:

P(at least one winner)=1−(1−p1)N

360 people guess independently of each other. Nobody wins with probability (119120)360=0.0492. So at least one wins with probability 1−0.0492=0.9508. The expected number of winners here is 360⋅1120=3, and still nobody wins with probability 0.0492.

The smallest number of attempts: the logarithm

Sometimes the question is how many calls are needed for at least one sale with probability at least 0.9. The unknown n sits in the exponent: 1−0.85n≥0.9. The logarithm ln⁡ brings the exponent down (the ln key on the calculator). You need one property: ln⁡(an)=n⋅ln⁡a, the exponent becomes a factor.

  1. Rearrange: 0.85n≤1−0.9=0.1.
  2. Take ln⁡ of both sides: n⋅ln⁡0.85≤ln⁡0.1.
  3. On the calculator ln⁡0.85=−0.1625 and ln⁡0.1=−2.3026. The logarithm of a number below 1 is negative.
  4. Dividing by a negative number flips the inequality sign: n≥−2.3026−0.1625=14.17.
  5. n is whole, so round up: n=15. Check: 0.8514=0.1028 is still above 0.1, 0.8515=0.0874 already below.

Your turn4 of 5

Each part is defective with probability 0.07, independently of the others. An inspector takes 10 parts. With what probability is at least one of them defective? Round to four decimal places.

Your turn5 of 5

In a number game 3 of 15 numbers are drawn, and only a guess with exactly these numbers wins. 455 people guess purely at random, independently of each other. What is the probability that at least one guess wins?

The formula sheet only has P(A′)=1−P(A). The lines P(W≤k)=1−(1−p)k and 1−(1−p)n are not on the sheet; you derive them from “all lost”.

At a glance: product, “none” and three methods

The topic in a minute

  • Independent events: the condition does not change the share, P(B∣A)=P(B), so P(A∩B)=P(A)⋅P(B). Trap: confusing them with disjoint events, whose intersection is 0.
  • “At least one”: 1− “none”, and “none” for independent parts is the product of the probabilities that they work. Trap: multiplying p by the number of attempts.
  • First success: P(W>k)=qk, P(W≤k)=1−qk; “fewer than k” is W≤k−1.
  • Independent events

    “Independently of each other” and a question about “and”, “at least one” or an unknown share. Every cell of the table is the product of its margins.
    P(A∩B)=P(A)⋅P(B)

    Trap: treating the events as disjoint and losing the intersection.

  • Systems of independent parts

    Parts fail independently; the system tolerates at most one failure, or needs all parts, or needs only one.
    (1−f)n+nf(1−f)n−1

    Trap: computing only “no failure”.

  • At least once and the first success

    The same attempt is repeated independently: “at least once in n attempts” or “until it works for the first time”.
    P(W≤k)=1−(1−p)k

    Trap: reading “fewer than k” as W≤k.

Daily revenue depends on chance too, but it is a number, not an event, and it has an average and a spread: “Expected value and variance: mean and spread”.

Now practise it

Independence: when probabilities multiply

Trainer · Probability

Practise
Next: Expected value and variance: mean and spread
0/5

Was this article helpful?