A Statistical Distribution Function of

W i d e Applicability
BY W A L O D D I WEIBULL, 1 STOCKHOLM, S W E D E N

This paper discusses the applicability of statistics to a sary general condition this function has to satisfy is to be a posi-
wide field of problems. Examples of simple and complex tive, nondecreasing function, vanishing at a value xu, which is
distributions are given. not of necessity equal to zero.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/18/3/293/6747088/293_1.pdf by Bogazici Universitesi user on 08 November 2024

T h e most simple function satisfying this condition is

I F a variable X is attributed to the individuals of a population,

the distribution function ( d f ) of X, denoted F{x), may be
defined as the number of all individuals having an X < x,
divided b y the total number of individuals. This function also and thus we put
(3 —
Xtl
xjm

gives the probability P of choosing at random an individual

having a value of X equal to or less than x, and thus we have (j - Xu)"

F(x) = 1— e [5!
P(X < x) = F(x). [1]
The only merit of this df is to be found in the fact that it is the
Any distribution function may be written in the form simplest mathematical expression of the appropriate form, Equa-
tion [2 J, which satisfies the necessary general conditions. Experi-
F(x) = 1 • [2]
ence has shown that, in many cases, it fits the observations better
This seems to be a complication, but the advantage of this formal than other known distribution functions.
transformation depends on the relationship T h e objection has been stated that this distribution function
has no theoretical basis. But in so far as the author understands,
(1 — py = e- [3] there are—with very few exceptions—the same objections
against all other df, applied to real populations from natural or
The merits of this formula will be demonstrated on a simple
biological fields, at least in so far as the theoretical basis has any-
problem.
thing to do with the population in question. Furthermore, it is
Assume that we have a chain consisting of several links. If we
utterly hopeless to expect a theoretical basis for distribution
have found, b y testing, the probability of failure P at any load x
functions of random variables such as strength properties of ma-
applied to a "single" link, and if we want to find the probability
terials or of machine parts or particle sizes, the "particles" being
of failure P„ of a chain consisting of n links, we have to base our
fly ash, Cyrtoideae, or even adult males, born in the British Isles.
deductions upon the proposition that the chain as a whole has
failed, if any one of its parts has failed. Accordingly, the proba- It is believed that in such cases the only practicable way of
bility of nonfailure of the chain, (1 — P„), is equal to the progressing is to choose a simple function, test it empirically, and
probability of the simultaneous nonfailure of all the links. Thus stick to it as long as none better has been found. In accordance
we have (1 — P„) = (1 — P)". If then the df of a single link takes with this program the df Equation [5], has been applied n o t only
the form Equation [2], we obtain to populations, for which it was originally intended, but also to
populations from widely different fields, and, in many cases, with
P _ x _ [4] quite satisfactory results. T h e author has never been of the
opinion that this function is always valid. On the contrary, he
Equation [4] gives the appropriate mathematical expression very much doubts the sense of speaking of the " c o r r e c t " distri-
for the principle of the weakest link in the chain, or, more gen- bution function, just as there is no meaning in asking for the
erally, for the size effect on failures in solids. correct strength values of an S A E steel, depending as it does, not
The same method of reasoning may be applied to the large only on the material itself, but also upon the manufacturer and
group of problems, where the occurrence of an event in any part many other factors. In most cases, it is hoped that these factors
of an object may be said to have occurred in the object as a whole, will influence only the parameters. However, accidentally they
e.g., the phenomena of yield limits, statical or dynamical strengths, may even affect the function itself.
electrical insulation breakdowns, life of electric bulbs, or even
The purpose of this paper has been to illustrate with a few
death of man, as the probability of surviving depends on the
examples the experience that the df, Equation [5], may some-
probability of not having died from many different causes.
times render good service.
N o w we have to specify the function ip(x). The only neces- The number of examples has, by space, been limited to the
following:
1 Professor at the Royal Institute of Technology; Scientific Adviser

to AB Bofors; mailing address Bofors, Sweden. 1 Yield strength of a Bofors steel

Contributed by the Applied Mechanics Division for presentation,
by title, at the Annual Meeting, Atlantic City, N. J., November 25- 2 Size distribution of fly ash
3 0 , 1 9 5 1 , o f T H E A M E R I C A N S O C I E T Y OF M E C H A N I C A L ENGINEERS.
3 Fiber strength of Indian cotton
Discussion of this paper should be addressed to the Secretary,
ASME, 29 West 39th Street, New York, N. Y., and will be accepted 4 Length of Cyrtoideae
until October 10, 1951, for publication at a later date. Discussion 5 Fatigue life of a St-37 steel
received after the closing date will be returned.
NOTE: Statements and opinions advanced in papers are to be In the Appendix:
understood as individual expressions of their authors, and not those 6 Statures for adult males, born in the British Isles
of the Society. Manuscript received by ASME Applied Mechanics
Division on January 15, 1951. Paper No. 51—A-6. 7 Breadth of beans of Phaseolus Vulgaris
293

Copyright © 1951 by ASME

294 JOURNAL OF APPLIED MECHANICS SEPTEMBER, 1951

T h e correctness of fit has been checked b y applying the chi- S I Z E D I S T R I B U T I O N OF F L Y A S H

square method T h e observed values are taken f r o m J. M . Dalla Valle's work. 2
Of those populations, Nos. 1 - 3 are distributed in good agree- Fig. 2 gives the curve and T a b l e 2 the values. T h e parameters
ment with the df Equation [5 j, whereas the four remaining p o p u - are xu = 30fi, x0 = 128 fi, in = 2.288. W i t h o u t pooling, the d of f
lations have to be split up into t w o components, before such an are 12 — 3 = 9. Then x 2 = 8.44 gives a P = 0.49. If the classes
agreement is obtained. T h e first type will be called a " s i m p l e " 2 - 3 and 13-14 are pooled, the d of f are 7 and x 2 = 8.44 gives a
and the second type a " c o m p l e x " distribution. P = 0.29.
T h e fundamental question n o w arises, whether this splitting-
up is a purely formal operation, or whether it might unveil some TABLE 2 SIZE D I S T R I B U T I O N OF F L Y ASH
hidden real causes. I t may be said that any distribution may be (x - - particle diameter in 20 microns)

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/18/3/293/6747088/293_1.pdf by Bogazici Universitesi user on 08 November 2024

represented b y a sum of a sufficiently great number of simple Expected Observed
values values
distributions, just as any periodical function m a y be developed in X n n
a Fourier series. However, if the number of the components be 2 3 3
3 14 14
small and the number of observations sufficiently large, the like- 4 34 34
5 62 56
lihood of real causes seems to increase. In any case, it is very 6 92 85
easy to produce real complex distributions b y syntheses. 7 122 126
8 150 150
It seems obvious that the components of examples 4 and 5 are 9 172 175
10 188 188
due to real causes. I n examples 6 and 7 it is impossible t o 11 199 197
12 205 202
decide whether the division is a formal one or a real one, but the 13 209 208
fact itself m a y be a valuable stimulus to a closer examination of 14 211 211
the observed material.
T h e specific data for the examples follow.

Y I E L D STRENGTH OP A B O F O R S S T E E L

T h e observed values are obtained as routine tests of a Bofors

steel, the quality of which was chosen at random for purposes of
demonstration only. Fig. 1 gives the curve and T a b l e 1 the

TABLE 1 Y I E L D S T R E N G T H OF A BOFORS STEEL

(x = yield strength in 1.275 k g / m m 2 )
Expected Observed Normal
values values distribution
X n 71 il
1 32 10 10 8
2 33 36 33 28
3 34 84 81 71
4 35 150 161 141
5 36 224 224 225
0 37 291 289 301
7 38 340 336 351
8 39 369 369 376
9 40 383 383 386
10 42 389 389 388 FIG. 2 SIZE DISTRIBUTION OF F L Y A S H

Yield S t r e n g t h of a
Bofors S t e e l
A F I B E R STRENGTH OF INDIAN C O T T O N

T h e observed values are taken f r o m R . S. Koshal and A. J.

N-389
Turner. 3 Fig. 3 gives the curve, and T a b l e 3 the values. T h e
-.8-1- parameters are xu = 0.59 gram, x0 = 3.73 grams, m = 1.456.
If the classes 14 to 16 are pooled, the d of f are 13 — 3 = 10. Then
.+-1 - X 2 = 11.45 gives a P = 0.35.
*„-38.S7 kg/mm 2 T h e authors 3 have pointed out that the most striking feature
.0-1 x 0 » 7.74- kg/mm 2 about the frequency curve is its asymmetry, showing a well-
m- 2.9342 marked predominance of weak fibers. I t was f o u n d — t h e y s a y —
.6-2
X 2 - 5.40j i - 6 that the observation curve would be well fitted b y a theoretical
P . 0.49 curve of Pearson's T y p e 1, having the following equation

X log(x-x„) / \ 0.876T16 /

FIG. 1 YIELD STRENGTH OF A BOFORS STEEL

y = 599
• 3 ( i+ sy (>- 29.1947/

In this equation y represents the frequency of any strength x,

values, observed and calculated. T h e parameters are xu = 38.57 expressed in grams.
k g / m m 2 , .To = 7.74 k g / m m 2 , m = 2.934. W i t h o u t pooling, the T h e values computed f r o m this n o t very h a n d y equation are
degrees of freedom (d of f ) are 9 — 3 = 6. Then x 2 = 5.40 gives shown in the last column of T a b l e 3. T h e d of f are 13 — 5 = 8
P = 0.49. T h e agreement is thus very satisfactory. (as there are 5 parameters). Then x 2 = 14.43 gives a P = 0.07.
As a comparison, the values expected on the hypothesis of a
! "Micromeritics," by J. M . Dalla Valle, Pitman Publishing
normal distribution have been computed and are given in the
last column of T a b l e 1. If the classes 9 - 1 0 are pooled, the Corporation, New York, N. Y „ 1948, p. 57, Fig. 2.
3 "Studies in the Sampling of Cotton for the Determination of
d of f are 8 — 2 = 6. T h e n a x ' = 18.17 gives a P = 0.008, Fiber Properties," by R. S. Koshal and A. J. Turner, Journal of the
which is not satisfactory at all. Textile Institute Transactions, vol. 21, 1930, pp. 325-370.
 W E I B U L L — A S T A T I S T I C A L D I S T R I B U T I O N F U N C T I O N OF W I D E APPLICABILITY 295

TABLE 3 FIBER STRENGTH OF INDIAN COTTON In spite of the greater n u m b e r of parameters, the fit of this dis-
(x = tensile strength in grams) tribution function is n o t as close as that of the first one.
Expected Observed Pearson
values values Type 1
X n n n LENGTH OP CYRTOIDEAE
1 118 177 127
2 646 667 659 T h i s is the first example of a complex distribution. T h e o b -
3 1232 1219 1255 served values h a v e been obtained f r o m investigations b y D r .
4 1751 1729 1777
5 2161 2153 2184 Gustaf Arrhenius, on submarine cores f r o m the Swedish D e e p - S e a
6 2461 2465 2480
7 2667 2664 2683 Expedition W i t h Albatross. T h e measurements were m a d e b y
8 2802 2813 2816 Dr. C. Jungk, taking samples f r o m each 10 c m of the core,
9 2886 2887 2899
10 2937 2933 2949 corresponding t o an age interval of a b o u t 1 0 0 , 0 0 0 years. Some
11 2966 2962 2978
12 2982 2985 2994 fifty populations have been analyzed statistically. A b o u t 20 per

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/18/3/293/6747088/293_1.pdf by Bogazici Universitesi user on 08 November 2024

13 2991 2991 3003 cent of the populations showed a simple distribution, as exempli-
14 2996 2995 3007
15 2999 2999 3009 fied in a previous paper. 4 T h e remaining samples showed a t w o -
16 3000 3000 3010
c o m p o n e n t distribution.
Fig. 4 gives the curves and T a b l e 4 the values of one of the
complex populations. T h e undivided sample gives the curve
marked JVi + N 2 . I t is easy t o see that the distribution is a
-.4-0- complex one, and that it is necessary t o split up the population in
Fiber Strength of two parts. B y trial it was found that 86 of the individuals b e -
-.0-0 - Indian Cotton longed to c o m p o n e n t N o . 1, and 14 to c o m p o n e n t N o . 2.
N - 3000 T h e parameters are: C o m p o n e n t N o . 1: x„ = 3.75 .TO = 63.2
.6-1 - M, M = 2.097. P o o l i n g the classes 2 - 3 , 9 - 1 0 , and 11-13 gives
Xu 3 0.59 gc X' = 3 . 5 9 . T h e d of f are 7 — 3 = 4 , and P = 0 . 4 7 .
.2-1 x 0 • 3.73 gr C o m p o n e n t N o . 2 : xu = 1 2 2 . 0 )i, rr0 = 1 2 4 . 1 fi, m = 1.479.
m = 1.4561 T h e number of individuals is t o o small for the x 2 -test.
-.8-2 XJ 11.45, i-10
P 0.35 FATIGUE LIFE OF AN S T - 3 7 STEEL

T h e observed values are taken f r o m Muller-Stock. 6 T h e fre-

quency curve in Fig. 5 6 gives n o impression of a complex dis-
tribution, which, on the other hand, m a y easily be seen when
FIG. 3 F I B E R S T R E N G T H OF INDIAN COTTON
4 " A Statistical Analysis of the Size of Cyrtoideae in Albatross

Cores From the Bast Pacific Ocean," b y W . Weibull, Nature, vol.

TABLE 4 LENGTH OF CYRTOIDEAE 164, 1949, p. 1047.
6 " D e r Einfluss dauernd und unterbrochen wirkender, schwingender
(j- = length in microns)
Observed tlberbeanspruchung auf die Entwicklung des Dauerbruchs," by H .
. Expected values • values Muller-Stock, Miltcilungen Kohlc- und Eisenforschvng, (March, 1938),
X >n «2 111+ 2 ni+ 2 by measurements from his Fig. 13; reproduced in Fig. 5 of this paper.
1 10 1 1 0
2 20 5 5 5
3 30 13 13 12 TABLE 5 FATIGUE LIFE OF ST-37
4 40 23 23 24 (Rotating—beam test at ±32 kg/mm 2 )
5 50 35 35 38 Observed
6 60 47 47 45 AT -Expected values— values
7 70 58 58 58 10» m 71|+ j ni+ j
8 80 67 67 69 1 17.5 4.6 4.6 4.6
9 90 74 74 70 2 22.5 47.4 47.4 47.4
10 100 79 79 80 3 27.5 125.1 125.1 125.1
11 110 82 82 82 4 32.5 161.2 8.1 169.3 169.2
12 120 85 85 84 5 37.5 164.9 2S.0 192.9 192.7
13 130 86 86 86 6 42.5 165.0 41.9 206.9 207.3
14 140 86 '4 90 90 47.5 165.0 51.0 216.0 215.9
15 150 86 7 93 93 7 52.5
8 165.0 57.0 222.2 222.2
16 160 86 9 95 95 57.5 165.0 226.0 225.9
17 170 86 11 97 97 9 62.5 61.0
10 165.0 63.7 228.7 228.7
18 180 86 12 98 98 67.5 165.0 65.6 230.6 230.5
11
19 190 86 13 99 99 72.5 165.0 66.9 231.9 231.9
20 200 86 14 100 12 77.5
100 13 165.0 67.9 232.9 232.9
82.5 165.0 68.6 233.6 233.5
14 87.5 165.0
15 69.1 234.1 233.9
92.5 165.0 70.0 235.0 235.0
1G

2-0 Length of Cyrtoidea 100

Sample No.41.247
8-1— N- 100 B0

60

Component 1
\
\
x„= 3.75/t 40
m = 2.097
X2-- 3.59
20
i
P =
* 7-3
0.47

0
J
0,010 0,020 C030 0.040 0,050 0,060 0,070 0,080 0,090 0,100 x10"

FIG. 5 FREQUENCY CURVE OF FATIGUE LIFE OF ST-37 STEEL

FIG. 4 LENGTH OF CYRTOIDEAE (Number of specimens versus number of stress cyoles.)
 296 JOURNAL OF APPLIED MECHANICS SEPTEMBER, 1951

TABLE 6 S T A T U R E S F O R A D U L T M A L E S B O R N IN T H E B R I T I S H
using the plottings in Fig. 6. T h e parameters are: C o m p o n e n t ISLES
N o . 1: xu = 4.032, m = 5.956; C o m p o n e n t N o . 2: x u = 4.484, (x = height in inches)
Observed
m = 1.215. Table 5 shows the close agreement between the '— Expected values > values
observed and the calculated values. X TO ni H1+ i ni+ 2
57 2 2 2
58 6 6 6
59 20 20 20
60 56 56 61
61 143 143 144
62 333 333 313
63 702 702 707
64 1350 1350 1376
65 2351 2351 2366
66 3641 3641 3589
67 4917 4917 4918

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/18/3/293/6747088/293_1.pdf by Bogazici Universitesi user on 08 November 2024

68 5787 339 6126 6148
69 6134 1079 7213 7211
70 6197 1671 7868 7857
71 6200 2039 8239 8249
72 6200 2233 8433 8451
73 6200 2324 8524 8530
74 6200 2363 8563 8562
75 6200 2378 8578 8578
76 6200 2383 8583 8583
77 6200 2385 8585 8585

TABLE 7 PHASEOLUS VULGARIS

(Breadth of beans = 0.25 x + 6.70 m m )
Observe
j^xpeciea values values
X TO 112 ni+1 7U+ 1
1 32 32 32
2 130 130 135
I t m a y be pointed out that the frequency curve in Fig. 5 seems 3 400 400 374
t o be the result of a smoothing operation on the cumulative 4 1011 1011 998
5 2145 2145 2185
frequency curve. Accordingly, the sampling errors of the o b - 6 3832 3832 3835
7 5718 5718 5718
served values in Table 5 have been eliminated almost entirely 8 7140 486 7626 7648
(without affecting the function), which explains the really too 9 7761 1525 9286 9286
10 7800 2510 10400 10416
g o o d representation of the observed values. 11 7900 3229 11129 11153
12 7900 3671 11571 11580
T h e real causes of this splitting up in t w o components m a y be 13 7900 3908 11808 11801
found b y examining the frequency curve of the yield strength of 14 7900 4022 11922 11911
15 7900 4071 11971 11968
the same material, Fig. 7. I t is easy to see that the material, 16 7900 4091 11991 11992
17 7900 4098 11998 11998
probably not being killed, is composed of two different kinds. 18 7900 4100 12000 12000
If we suppose that all the specimens with a yield strength of less
than 25 k g / m m 2 belong to Component N o . 1, we obtain 14 speci-
mens o u t of 20, making 70 per cent. E x a c t l y the same propor-
tion has been found b y the statistical analysis, as 165/235 = 70
per cent.
T h e reason w h y this partition is so easily seen in Fig. 7 and
not at all in Fig. 5, depends, of course, upon the much larger
scatter in fatigue life than in yield strength.

Appendix
T h e foregoing statistical methods have been applied to many
problems outside the field of applied mechanics. I t may perhaps
be of interest to have examples of this kind, and for this reason, FIG. 7 FREQUENCY CURVE OF YIELD STRENGTH OF ST-37 STEEL
the following two are given with the tables only: (Number of specimens versus yield strength in k g / m m " . )

STATURES FOR ADULT M A L E S BORN IN THE BRITISH ISLES

If the classes 20-21 are pooled, x 2 = 5.11. T h e d of f are 8 —
T h e observed values are taken f r o m Y u l e and Kendall. 6 3 ' A = 4 ' / 2 , which gives a P = 0.35.
This distribution is classified b y the authors as being approxi-
mately of the symmetrical type, and there is no mention of its BREADTH OF BEANS OF PHASEOLUS VULGARIS

being composed of t w o parts. This is a classical example, quoted f r o m Charlier 7 to exemplify

B y trial it was found that the population had t o be split up the expansion in Edgeworth's series.
into t w o parts: N , = 6200 and N-> = 2385. T h e parameters are If the population is divided into t w o parts, Ni = 7900 and
as follows: = 4100, each of them maj r be very well fitted to a simple dis-
Component N o . 1: x u = 50.0 in., :io = 16.2 in., m = 9.6865. tribution function with the following parameters:
If the classes 1 - 2 and 14-15 are pooled, we get x 2 = 11.80. C o m p o n e n t No. 1: x u = — 3 . 0 ( = 5.95 m m ) , m = 6.2805.
As we have 7 parameters altogether, (one of them is the partition Without pooling we have the x 2 = 7.70, and the d of f 10 — 3 7 2
of the population), we take 3 ' / 2 to each of the components. T h e = 6 ' A giving a P = 0.29.
d of f are then 12 — 3 ' / 2 = 8 l / 2 , which gives a P = 0.20. Component No. 2: x u = + 7 . 2 ( = 8.50 m m ) , m = 1.6098.
Component N o . 2: xu = 67.4 in., x 0 = 2.3 in., m = 1.4662.
7 "Die Grundlagen der Mathematischen Statistik," by C. V. L.
6 " A n Introduction to the Theory of Statistics," by G. U. Yule Charlier, second edition, 1920, p. 73, quoted by Harald Cramfer in
and M . G. Kendall, eleventh edition, J. B. Lippincott Company, his book:"Mathematical Methods of Statistics," Princeton Univer-
Philadelphia. Pa., 1937, pp. 94 and 111. sity Press, Princeton, N. J., 1945, p. 440.
 W E I B U L L — A S T A T I S T I C A L D I S T R I B U T I O N F U N C T I O N OF W I D E APPLICABILITY 297

If the classes 17-18 are pooled, the value of X s = 4.50, and the normal distribution and asymptotic expansions f r o m it. The
d of f 9 — 3 7 2 = 5 7 2 give a P = 0.56. result was as follows:
I t m a y be of interest t o compare this result with those of
Charlier and Cramer. Normal distribution x 2 = 196.5 d of f 13 P < 0.001
First approximation x2 = 34.3 d of f 12 P < 0.001
Charlier says that, at the first look, the agreement with the Second approximation x 2 = 14.9 d of f 11 P = 0.19
normal distribution seems very satisfactory, but that a closer
examination shows a small negative skewness and a small posi- T h e agreement is satisfactory in the third case only, requiring
tive kurtosis. four terms of the series. This operation is certainly of a purely
Cramfir has calculated the values of x 2 ° n the hypotheses of formal character.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/18/3/293/6747088/293_1.pdf by Bogazici Universitesi user on 08 November 2024