Normal distribution: probability as area

31 min · 3 drills

In the textbook: Ch. 5.4 Die Normalverteilung

How do you find the probability for a quantity that can take any value, such as a time or a mass? The answer is in the table on the formula sheet, once you know how to read it.

Time, mass and length take any values, not only whole numbers. For such quantities the probability is computed as an area under a curve, most often under the bell of the normal distribution.

A courier service delivers a parcel in 48 hours on average, some faster, some slower. The delivery time X is normally distributed with a standard deviation of 8 hours.

What share of parcels takes longer than 58 hours?

Counting cases one by one, like the faces of a die, does not work: the time can be 58.3 hours, 58.31 hours, and so on without end. A different way of computing probability is needed.

That method is explained below; the result is this. 58 hours lie 10/8=1.25 standard deviations above the mean. For 1.25 the table on the formula sheet gives 0.894: that share of parcels arrives sooner. The remaining 1−0.894=0.106 take longer than 58 hours, about one parcel in ten.

Check yourself. A task at exam level: the weight of a loaf is normally distributed, μ=500 g, σ=10 g. What share of loaves is lighter than 480 g? From the table: Φ(−2.00)=0.023, Φ(2.00)=0.977.

Continuous quantity: probability as area

Write down the times of 1000 deliveries and sort them into bars one hour wide: how many parcels took 40 to 41 hours, 41 to 42, and so on. The height of a bar is the share of parcels in that hour, and all bars together give 1. The share “longer than 58 hours” is the sum of the bars to the right of 58. Make the bars narrower and their tops merge into a smooth curve. Then the share to the right of 58 is the area under the curve to the right of 58:

Delivery time, hours
Bell curve of delivery time with mean 48 hours; the area to the right of 58 hours is marked583240485664

P(X ≥ 58) ≈0.106

This is how probability is computed for any continuous quantity (German “stetig”), a quantity that can take any value in an interval: time, weight, length.

  • The whole area under the curve is 1, like the sum of all the bars.
  • The height of the curve is called the density (German: Dichte). It is not a probability: only an area gives a probability.
  • The probability of landing between a and b is the area between a and b.
  • A single point has no area: P(X=58)=0. So it does not matter whether the task says < or ≤.

The delivery curve has the shape of a bell: it is symmetric around the mean μ, and its width is set by the standard deviation σ. Such a distribution is called normal (German “normalverteilt”). About 68 % of the values lie within μ±σ, here between 40 and 56 hours.

Standardising: z = (x − μ)/σ

The area under the bell cannot be computed with school formulas; it is taken from a table. But there are infinitely many bells, with any μ and any σ. There is one table, for the bell with μ=0 and σ=1, called the standard normal distribution. Every bell is reduced to it by one number:

z=x−μσ

z says how many standard deviations x lies from the mean. For 58 hours: z=58−488=1.25, one and a quarter σ above the mean. Subtracting μ shifts the bell to zero, dividing by σ squeezes it to width 1. The shares of area do not change.

On the formula sheet this is the line Z=X−μσ, and Φ(z) (“phi”) is the area of the standard bell to the left of z, the line Φ(z)=P(Z≤z).

The daily revenue with μ=460 GE (“Geldeinheiten”, monetary units: just money) and σ=30 GE shows it. In the picture X is the revenue, F the share of area up to 490 GE, and Φ(1) the same share of the standard bell left of z=1:

Reading the table Φ

After standardising, every bound becomes a z, and Φ(z) is no longer calculated but looked up in the table. On the formula sheet, table 1 (Tabelle 1) gives Φ(z). The row holds z with one decimal, from −2.9 to 2.9, the column the second decimal, from 0.00 to 0.09. Where they cross stand three decimals of Φ(z). For z=1.25: row 1.2, column 0.05.

Excerpt of table 1: rows 1.2 and −1.2, columns 0.04, 0.05 and 0.06; Φ(1.25) and Φ(−1.25) marked0.040.050.061.2−1.20.8930.8940.8960.1070.1060.104

The answer to the question from the start: to the left of 58 hours lies Φ(1.25)=0.894, to the right 1−0.894=0.106. About 10.6 % of parcels take longer than 58 hours.

Negative z. The table has rows with a minus. Row −1.2 and column 0.05 give Φ(−1.25)=0.106: the second decimal goes the same way, −1.2−0.05=−1.25. Check with the symmetry of the bell: to the left of −1.25 lies as much as to the right of 1.25, so Φ(−z)=1−Φ(z); this line is on the sheet too.

Table 2 (Tabelle 2) is read the other way round: for a share α to the left it gives z with four decimals, the sheet line Φ(z)=α⇒z=Φ−1(α). For example, α=0.93 gives z=1.4758, and α=0.10 gives z=−1.2816. It follows table 1 on the formula sheet.

A share under a normal distribution: above, below or between bounds

With the parcels everything came down to one tail, to the right of 58 hours. In the first exam task there are two tails, and the bound hides in the words about the deviation from the mean:

Exam-style task

Eine Abfüllanlage füllt Müslipackungen; das Füllgewicht ist normalverteilt mit Erwartungswert 400 g und Standardabweichung 4 g. Eine Packung wird aussortiert, wenn ihr Gewicht um mehr als 7 g vom Erwartungswert abweicht. Welcher Anteil der Packungen wird aussortiert?

  1. a)0.073
  2. b)0.066
  3. c)0.080
  4. d)0.076
  5. e)0.040

It says “normally distributed”, μ and σ are given with a bound, and a share is asked. “Deviates by more than 7 g” is two sides at once: lighter than 393 g and heavier than 407 g. The bell is symmetric, so the tails are equal, and it is enough to find one.

Pack weight, g

σ = 4 g given, μ = 400 g

Bell curve of the pack weight with mean 400 g; two tails marked, left of 393 g and right of 407 g392396400404408

Excerpt of table 1: row 1.7, columns 0.04, 0.05 and 0.06; Φ(1.75) marked0.040.050.061.70.9590.9600.961
Step 1 of 4Find σ

The standard deviation is given, σ=4 g. Had the variance been given, you would first take the square root.

Why are there two tails, although the task has only one number 7? Hint: a pack is rejected for being underweight and for being overweight.

Every question about an area is built from Φ:

Questionarea
less than xΦ(z)
more than x1−Φ(z)
between a and bΦ(zb)−Φ(za)
deviation of more than c from μ2⋅(1−Φ(c/σ))
deviation of at most c2⋅Φ(c/σ)−1

The weight of a coffee pack on the packing line is normally distributed, μ=1000 g, σ=5 g.

What share of packs weighs between 992 and 1006 g?

za=992−10005=−1.6 and zb=1006−10005=1.2.

Row 1.2, column 0.00: Φ(1.2)=0.885. Row −1.6, column 0.00: Φ(−1.6)=0.055.

0.885−0.055=0.830: everything left of 1006 g minus what lies left of 992 g.

From the table: under 24 minutes. The commute takes 30 minutes on average, σ=4 minutes, normally distributed. With what probability does it take less than 24 minutes? z=24−304=−1.5. Look up Φ(−1.5) in row −1.5, column 0.00.

Your turn1 of 3

The weight of a sugar pack is normally distributed, μ=750 g, variance σ2=25 g². With what probability does a pack deviate from 750 g by more than 9 g? The table gives Φ(1.80)=0.964.

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

  • Only one side, option (e), 0.040: one tail is computed, 1−0.960. “Deviates by more than” is underweight and overweight.
  • Two sides where one is asked: “heavier than 407 g” is only the right tail, without the factor 2.
  • Variance instead of σ: z=7/16 when σ2=16 is given. In z you divide by the standard deviation.
  • Opposite side taken: Φ(z) for “more”, 1−Φ(z) for “less”. Check: a bound above the mean and the question “more” give an answer below 0.5.
  • Φ instead of the area: Φ(z) itself is given as the answer, without moving to the region asked.
  • Φ(z)−0.5: that is only the area from the mean to the bound. It does not answer “more” or “less”.

On the formula sheet: Z=X−μσ, Φ(z)=P(Z≤z) (Tabelle 1) and Φ(−z)=1−Φ(z).

The bound for a share: which value only a few exceed

So far the bound was given and the share was asked. The second exam task goes the other way, from the share to the bound, with table 2 from the section “Reading the table Φ”:

Exam-style task

Die Wartezeit in der Hotline eines Mobilfunkanbieters ist normalverteilt mit Erwartungswert 340 Sekunden und Standardabweichung 78 Sekunden. Welche Wartezeit wird nur von 7 % der Anrufer überschritten? Runden Sie auf zwei Nachkommastellen.

  1. a)358.63
  2. b)523.83
  3. c)224.89
  4. d)455.11
  5. e)115.11

Here the question runs the other way: a share is given and the bound is asked. That is finding the area under the bell, backwards. Table 2 gives z for the share to the left, and z=x−μσ turns into x=μ+z⋅σ.

Waiting time, s

right of g: 7 %, left: 93 %

Bell curve of the waiting time with mean 340 s; unknown bound g with 7 percent to its rightg = ?18426234041849693 %7 %

Excerpt of table 2: shares 0.92, 0.93 and 0.94 and their z; α = 0.93 markedΦ⁻¹(α)0.920.930.941.40511.47581.5548
Step 1 of 3The share below

“Exceeded by only 7 %” means 0.07 to the right of the bound g and 1−0.07=0.93 to the left. Table 2 works with the share to the left.

Why is z positive here, but negative for “only 7 % wait less”? Hint: on which side of the mean do the 7 % lie.

The commute time is normally distributed, μ=40 minutes, σ=5 minutes.

Which time do 10 % of trips not reach, that is, 10 % of trips are shorter?

10 % of trips are shorter, so the share to the left of the bound is α=0.10.

α=0.10 gives z=−1.2816; the bound lies below the mean.

x=40+(−1.2816)⋅5=40−6.408=33.59 minutes.

A machine that fills bags has a spread of σ=8 g fixed by its design, while the mean is set by the adjustment.

To what mean μ should it be set so that at most 2 % of the bags weigh less than 500 g?

To the left of 500 g lies the share 0.02: z=−2.0537.

500−μ8=−2.0537, so μ=500+2.0537⋅8=516.43 g.

Solve: the bound of class “Extra”. The mass of apples is normally distributed, μ=180 g, σ=20 g. The heaviest 15 % go into the class “Extra”. From which mass does it start? To the left of the bound lies 1−0.15=0.85, table 2 gives z=1.0364. Convert back to grams.

Your turn2 of 3

The lifetime of a battery is normally distributed, μ=12 years, σ=1.5 years. Which lifetime is exceeded by only 15 % of the batteries?

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

  • Wrong side, option (c), 224.89: 340−115.11. “Exceeded by only 7 %” is a bound above the mean, z>0.
  • Mean forgotten, option (e), 115.11: this is only z⋅σ, the distance from the mean. The bound is μ+z⋅σ.
  • Tables mixed up: 0.93 taken as z and looked up in table 1, Φ(0.93)=0.824, and this number put in as z: 340+0.824⋅78=404.27. Check: table 1 turns a z into a share, table 2 turns a share into a z, and the task gives a share.
  • Variance instead of σ: if σ2 is given, take the square root first.
  • Share instead of z: 340+0.93⋅78. The share is first turned into z with table 2.

On the formula sheet: Φ(z)=α⇒z=Φ−1(α) (Tabelle 2) and Z=X−μσ, which gives x=μ+z⋅σ.

The time to failure from the distribution function

The bell fits when the values crowd around the mean. Time until failure looks different: most failures come right at the start. Instead of the Φ table the question gives a formula with the number e:

Exam-style task

Ein Webserver läuft bereits seit 50 Tagen ohne Fehler. Die fehlerfreie Laufzeit T (in Tagen) erfüllt P(T>t)=e−0.02 t. Wie wahrscheinlich ist es, dass er noch mindestens 47 weitere Tage fehlerfrei läuft? Runden Sie auf vier Nachkommastellen.

  1. a)0.0078
  2. b)0.1437
  3. c)0.8735
  4. d)0.9400
  5. e)0.3906

The data are the time until an event (a server fault, a firm closing, a call ending) and a function with the number e. That is either the “distribution function” F(t) or, as in the task above, P(T>t) directly. Unlike the number W of the attempt with the first success, time is not whole: a firm can close after 1.3 years. So the probabilities are not bars but a function.

We start with the simpler case, where F is given. A start-up centre describes the time T in years until a newly founded firm closes. The distribution function (German “Verteilungsfunktion”) F(t)=P(T≤t) tells what share of firms closes no later than time t. This is the same idea as Φ(z) for the bell: the area to the left of the bound, only the curve is different. Everything else follows from it:

  • P(T>t)=1−F(t): the complement, at time t the firm is still running;
  • P(a<T≤b)=F(b)−F(a): closed by time b, but not by time a.

The number e≈2.718 (the ex key on the calculator). Tasks have e−0.5t: at t=0 it is 1, and as t grows the value falls towards 0. This describes the share of firms still running: all at first, then fewer and fewer. For example, e−0.5=0.6065. If F(t)=1−e−λt, where λ (“lambda”) is the number in the exponent, here 0.5, then

P(T>t)=1−F(t)=e−λt

Such a distribution is called exponential (German “exponentialverteilt”). What λ means: the share of failures in a short time: of those still working, roughly the share λh fails within a short time h. With λ=0.5 per year, in one month, h=112 of a year, about 0.5⋅112≈0.04 close, that is 4 % of the firms still running.

For the start-up centre, F(t)=1−e−0.5t. What share of the firms stays in business longer than 1 year?

The height of the curve above year t is the share of firms closed by then, F(t) = 1 − e to the minus 0.5 t; at t = 1 compare the height F(1) with the rest up to one, that is P(T > 1)01234561yearsF(t) = P(T ≤ t)
Step 1 of 4Read the distribution function

F(t)=P(T≤t) is the share of firms closed no later than t years. The curve rises from 0 towards 1.

Why do “longer than 1 year” and “at least 1 year” give the same number here, but not for the attempt number W of the method “First success on the k-th try: throws until the first hit”? Hint: can the time be exactly 1.000… years, and with what probability.

A single moment has no probability: closing between 1 and 1+h years has probability F(1+h)−F(1), and that goes to 0 as h gets smaller and smaller. So for a time it does not matter whether the task says < or ≤.

Sometimes the task gives not F but S(t)=P(T>t), the share still running. Then P(T≤t)=1−S(t) and P(a<T≤b)=S(a)−S(b).

A battery runs longer than t years with probability S(t)=e−0.25t.

With what probability does it fail between the 2nd and the 4th year? And after how many years has half of the batteries failed?

“Between the 2nd and the 4th year” means 2<T≤4: running after 2 years, but not after 4.

S(2)−S(4)=e−0.5−e−1=0.60653−0.36788=0.23865≈0.2387.

S(t)=0.5: e−0.25t=0.5. The logarithm brings the exponent down: ln⁡x is the power you must raise e to in order to get x, and ec is e to the power c, so ln⁡(ec)=c. So −0.25t=ln⁡0.5=−0.6931 and t=2.77 years. This time is called the median.

In general e−λt=0.5 gives −λt=ln⁡0.5. Since 0.5=2−1, the rule ln⁡(an)=n⋅ln⁡a gives ln⁡0.5=−ln⁡2, and the median is t=ln⁡2λ. Here 0.69310.25=2.77.

This is the task from the sheet above. What is given is S(t)=P(T>t)=e−0.02t directly, the share of servers still running without a fault after t days. The new element is the condition: this server has already been running for 50 days.

“At least 47 more days” means the whole fault-free time reaches past 50+47=97 days, T>97. Already known is T>50. So what is wanted is the share of servers that make 97 days among those that made 50 days. This is written P(T>97∣T>50); the vertical bar reads “given that”.

Whoever runs 97 days has certainly made the 50 days. So the share wanted is S(97)S(50)=e−0.02⋅97e−0.02⋅50=0.14370.3679.

Dividing powers with the same base subtracts the exponents: e−1.94/e−1=e−0.94=0.3906. Option (e).

What is left is e−0.02⋅47, the same number as for a new server that is to last 47 days. This holds for every P(T>t)=e−λt, because e−λ(s+t)/e−λs=e−λt, and is called the memoryless property: a device that has been running for a while has the same chances for the future as a new one. You then do not need the elapsed time from the task.

The lifetime T of a device in years has F(t)=1−e−0.25t. With what probability does the device last at least 4 years? “At least 4 years” is T≥4, the opposite of “fails earlier”, so P(T≥4)=1−F(4)=e−0.25⋅4=e−1=0.3679. 0.6321 is F(4), the share that failed earlier.

Your turn3 of 3

The length T of a customer call in minutes has the distribution function F(t)=1−e−0.4t for t≥0. With what probability does a call last between 1 and 3 minutes?

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

  • Counting the elapsed time, option (b), 0.1437: this is S(97)=e−0.02⋅97, the chance that a new server lasts 97 days. But this server already has the first 50 days behind it. Check: a server that is already running cannot have a smaller chance than a new one has for the whole stretch.
  • Height of the curve instead of probability, option (a), 0.0078: this is 0.02⋅e−0.02⋅47, as if 1−e−0.02t had been differentiated and t=47 put in. That is the density, the height of a curve at one point, and one point has no probability.
  • λt as a probability, option (d), 0.9400: 0.02⋅47=0.94 is the exponent, not the answer. λt belongs in the exponent; the probability is e−0.94.
  • Complement, in tasks with F: in the example with the firms, F(1)=0.3935 is the share that closes in the first year, while the question asked for those that stay longer. Check: F is always “no later than”; longer is 1−F.
  • Lower bound forgotten, in “between” tasks: F(b) is taken instead of F(b)−F(a). Check: with a=1, b=3, calls under one minute do not belong to the answer.

The line P(T>t)=e−λt is not on the formula sheet. You get it from the given F(t) through the complement P(A′)=1−P(A).

At a glance: area, tables and three methods

The topic in a minute

  • Continuous quantity: probability is an area under the curve; a point has no area.
  • Standardising: z=(x−μ)/σ, Φ(z) is the area to the left of z (Tabelle 1, with negative rows too). Trap: dividing by the variance.
  • Quantile: share to the left, z from Tabelle 2, x=μ+z⋅σ. Trap: the wrong side.
  • Lifetime: F(t)=P(T≤t), longer than t is 1−F(t)=e−λt, between a and b is F(b)−F(a). Trap: taking F(t) for “longer”.
  • Which area under the bell curve

    “Normally distributed”, given μ and σ (or σ²) and a bound, with a share asked.
    z=x−μσ,P(X>x)=1−Φ(z)

    Trap: taking one tail when a deviation in both directions is asked.

  • Quantiles

    A share is given (“exceeded by only 7 %”), the bound or the mean is asked.
    x=μ+Φ−1(α)⋅σ

    Trap: taking the share to the right instead of to the left.

  • Lifetime with a continuous distribution

    The time until failure is given by F(t)=1−e−λt or by the share S(t) still working.
    P(T>t)=1−F(t)=e−λt

    Trap: taking F(t) for “longer”.

The number of successes in several attempts is often approximated by the bell curve too. When that works and how to compute without it is in “The number of successes in n attempts: binomial distribution and normal approximation”.

Cards to memorise

Tap a card to see the formula. All topic cards

Formula

Normal approximation of a count: from which boundary do you calculate for "at least k" and for "more than k"?

Formula

A value is exceeded by the share q. Which quantile is it?

Now practise it

Normal distribution: probability as area

Trainer · Probability

Practise
Next: The number of successes in n attempts: binomial distribution and normal approximation
0/3

Was this article helpful?