Sunteți pe pagina 1din 3

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-34, NO.

3, 1985 AUGUST

209

Chi-Square Probabilities are


Poisson Probabilities in Disguise
Andrew Gorski, Member IEEE
Consolidated Edison Co., New York

underlying distribution is Poisson, or when the Poisson is


used as an approximation of the binomial distribution.

Key Words-Chi-square distribution, Poisson distribution, s-Confidence level.

poif(r; ,u) -csqfc(2,u; 2(r + 1))

(2)

poifc(r; As)

(3)

Reader Aids-

Purpose: Tutorial
Special math needed for explanations: Elementary probability
Special math needed to use results: Same
Results useful to: Reliability and quality assurance engineers

Abstract-Since chi-square probabilities and Poisson probabilities

are different forms of the same mathematical function, it is reasonable to

use the chi-square tables for obtaining Poisson probabilities (for example,
the Pearson & Hartley, Biometrica Tables).

The exact relationships between the two cumulative


distributions are:

=csqf(21t; 2r)

Notation

poif(; A), poifc(; A)

Poisson Cdf, Sf with

csqf(,; v), csqfc(a; v)

chi-square Cdf, Sf with degrees-

mean,

of-freedom, v
number of failures

The usual chi-square table is superior to the usual


Poisson tables in that the selected probability levels are
shown in the heading of the table. This saves a lot of time
and
dispenses with the frequent need for interpolation.
INTRODUCTION
Many Poisson tables have the mean ,s shown in the heading:
According to E. S. Pearson & H. 0. Hartley [1, p 9]. the individual terms are in the left-hand column, and the
It is "well known" that the chi-square integral and the cumulative distribution function in the right-hand column.
cumulative sum of terms of the Poisson distribution are
different forms of the same mathematical function.
EXAMPLE: ACCEPTANCE SAMPLING PLAN
If this is so why write a paper about it? The answer is
in the illusiveness of the term "well known" - well Step 1. The design of a single sampling plan (rectangular
known to whom? My guess is that less than 0.1 o of stopping boundary) involves choosing several parameters.
reliability and quality assurance engineers know it even The three used in this example are:
though many statisticians do. In my 23 years of industry
onsumer risk.
experience I never met an engineer who knew it and yet I
1.
thethe
c3
11,
2.
c
+
number: Accept for c or fewer
worked on projects such as Minuteman I and Apollo where failures, reject for rejection
c + 1 or more failures.
the intent was to bring on board the best people available.
3. X2, limiting failure-rate level at which the accepThis subject is also treated in Gorski & Epstein [2].
tance probability is : and the rejection probability is 1 -A.

Step 2. Find the mean, ,u, in the Poisson tables such that
DISCUSSION
the rejection probability is 1 -f for c + 1 or more failures,
Traditionally, the chi-square table is used by reliabili- or equivalently, the acceptance probability is f3 for c or
ty engineers to calculate a statistical confidence interval for fewer failures. Once the mean, ,u, is found, then the totalconstant failure rate, X, from a test that is run for a total- test-time, T, is found from
test time, T, with r failures (r is the random variable):
X1-a/2 (2r)
Xa/2
(4)
2,u = 2X2T, T = /X2
Xi --/2
X,!/2 (2r + 22)v
(1) Explanation of Step 2
2T___(2r)_ <2rT
2T
2T
The probability of acceptance is the sum of Poisson
It is also used in a similar fashion to calculate the operating terms up to and including c:
characteristic of an acceptance test with a rectangular stopping boundary (the test stops when a given total-test-time Pr{Accept} = poif(c; it)
(Sa)
is reached, or a given number of failures has occurred
whichever comes first).
= csqfc(2,u; 2c + 2)
(5b)
What follows is an example showing the use of the
chi-square table for designing sampling plans when the from (2).
0018-9529/85/0800-0209$01 .00(C 1985 IEEE

IEEE TRANSACTIONS ON RELIABILITY, VOL. R-34, NO. 3, 1985 AUGUST

210

TABLE 1
Survivor Function (Sf) of the Chi-Square Distribution
Cumulative Distribution Function (Cdf) of the Poisson Distribution

2C or
2C +2

2)

-2npor2TA

Q-0.99

0.95

0.90

0.80

0.60

0,20

0.10

0.05

0.01

2
4
6
8
10

0.020
0.297
0.872
1.646
2.558

0.103
0.711
1.635
2.733
3.940

0.211
1.064
2.204
3.490
4.865

0.446
1.649
3.070
4.594
6.179

1.386
3.367
6.348
7.344
9.342

3.219
6.989
8.568
11.030
13A42

4.605
7.779
10.645
13.362
15.987

5.991
9.488
12.592
15.507
18,307

9.210
13.277
16.812
20.090
23.209

12
14
16
18
20

3.571
4.660
5.812
7.015
8.260

5.226
6.571
7.962
9.390

7.807
9.467

18.649

14.578

11.340
13.339
15.338
17.338
19.337

15.821
18.161
20.465

10.851

6.304
7.790
9.312
10.865
12A43

22
24
26
28
30

9.542
10.856
12.198
13.565
14.954

12.338
13.848
15.379
16.928
18.493

14.041
16.659

17.292
18.939
20.599

16.314
18.062
19.820
21.588
23.364

32
34
36
38
40

16.36
17.79
19.23
20.69
22.16

20.07

22.27

23.95
25.64
27.34
29.05

42
44
46
48
60

23.7
25.1
26.7
28.2
29.7

33.1
34.8

52
54
56
58
60

31.2
32.8
34.3
35.9
37.5

62

39.1
40.6
42.2
43.8
45.4

48.3
50.0
61.7

47.1
48.7
50.3

63.5
55.2
66.9

64
66
68

70
72
74
76
78
80

51.9
53.5
-2.3263

26.038

21.064
23.542
25.989
28.412

21.020
23.685
26.296
28.869

31.410

26.217
29.141
32.000
34.805
37.566

21.337
23.337
25.336
27.336
29.336

27.301
29.553
31.795
34.027
36.250

30.813
33.196
35.563
37.916
40.256

33.924
36.415
38.885
41.337
43.773

40.289
42.980
45.642
48.278
50.892

25.15
26.94
28.73
30.54
32.34

31.34
33.34
35.34
37.34
39.34

38.47
40.68
42.88
45.08
47.27

42.58
44.90
47.21
49.51
51.81

46.19
48.60
51.00
53.38
55.76

53.49
56.06
58.62
61.16
63.69

30.8
32.5
34.2
35.9
37.7

34.2
36.0
37.8
39.6
41.4

41.3
43,3
46.3

49.5

54.1

58.1
60.5
62.8
65.2

66.2
68.7

36.4
38.1
39.8
41.5
43.2

39.4
41.2
42.9
44.7

43.3
45.1
47.0
48.8
50.6

44.9
46.6

48.2

62.5
64.3

21.66
23.27

24.88
26.51
28.1
29.8

31.4

58.7

60.4
-1.6449

. .. ..

46.5

50.0

11.152
12.857

53.5

56.2
58.0

57.1

61.8
63.6

51.8

55.3

59.9

47.3

68.2

51.3
53.6
55.3

62.5

61.3

71.1
73.3
76A
77.6
79.7

59.3
63.3
65.3
67.3

69.3

64.3
-1.2816

-0.8418

0.0000

___.__I

64.7

66.8
69.0

67.3
692

65.5

60.3

67.3

62.5

60.7

51.6
63.8
56.0

49.3

71.3
73.3
75.3
77.3
79.3

58.9

22.760

81.9

56.4

71.2
73.7
76.2

58.6
60.9
63.2

67.5

65A
67.7
69.9
72.2
74.4

69.8
72.2
74.5
76.8
79.1

88.4

76.6
78.9
81.1
83.3

81.4
83.7
86.0
88.3
90.5

90.8
93.2
95.6
98.0
100.4

92.8

102.8
105.2
107.6

85.5

78.6
81.1
83.5
86.0

86.1
88.3

GOA9

87.7
90.0
92.2
94.4
96.6

101.9

112.3

+0.8418

+1.2816

+1.6449

+2.3263

84.0

95.1

97.4

99.6

110.0

GORSKI: CHI-SQUARE PROBABILITIES ARE POISSON PROBABILITIES IN DISGUISE

211

When X = X2, then Pr{Accept} = ,B from the problem


The discrimination ratio (at a producer risk of 5Wo and
statement.
a consumer risk of 20%) is X2/X1 = l/lAt1 = 8.558/1.635
From (3)-(5) it is clear that the degrees-of-freedom, v, = 5.23 which is not very good.
The failure rate for an acceptance probability of 5% is
needed for Poisson calculations is always an even number.
To simplify matters, and to make it harder to misread the obtained 2by using an Sf = 5% for degrees-of-freedom =
table, table 1 has been constructed using only an even 6. The x = 12.592 and corresponding X = 12.592/(2 x
number for v. Table 1 also goes up to 80 degrees-of- 143 000 hours) = 4.41 Wo/1000 hr.
The discrimination ratio (at producer and consumer
freedom, rather than the usual 30.
risks of 5% each) is 12.592/1.635 = 7.70, a large number
indeed.
NUMERICAL ILLUSTRATION
The discriminating ability is determined solely by the
Let f3 =20%
value of c. Not all test-plans are economically feasible. The
one for this example requires a total-test-time of 143 000
X2 = 3%/1000 hours
hours. If 43 units were put on test, a calendar time of
c= 2
10 000 hours would be needed (if the test were passed);
c + 1 =3.
that is over one solid year of testing, and not very feasible.
The table could be used to find a compromise between
Apply (5) at the consumer risk point (3 & X2).
a good operating characteristic and the management objectives in running the test.
0.20 = poif(c; X2T) = csqfc(2X2T; 2c + 2)
= csqfc(2X2T; 6)
(6)
Table 1 gives the Sf (viz, csqfc) for the chi-square distribuREFERENCES
tion. The value2 of x = 2X2T for 6 degrees-of-freedom and [1] E. S. Pearson, H. 0. Hartley, Biometrica Tables for Statisticians,
Sf = 0.20 is x = 8.558.
vol 1, Cambridge University Press, 1962, pp 9, 122.
[2] A. Gorski, B. Epstein, "Limitations of plans designed to
2X2T = 8.558
demonstrate minimum life with high confidence", 9-th Nat'l. Symp.
X2T = 4.279
Reliability and Quality Control, 1963, pp 415-419.
4.279
T=
= 143 000 hours
(7)
3%o/(1000 hours)
AUTHOR
What is Xi (an acceptable failure rate) such that the Andrew Gorski; 239 East 79 Street, Apt. 11 C; New York, New York
producer risk, a, is 5%o? that is, the probability of accep- 10021 USA.
ting the lot is 95%o?
Andrew Gorski (M'63, M'81) was born in Poland on 1921 March 4.
The degrees-of-freedom is still 6, (6 = 2c + 2), but During World War II he was a heavy bomber pilot in England. He came
the Sf is 95%Mo. The value of x2 = 2XI T is 1.635; and T is to the United States in 1949 and graduated in Economics from
Washington State University in 1952. Graduate work in Statistics and In143 000 hours.
dustrial Management were completed at Los Angeles State College in
2X1T = 1.635

1963. He worked in Aerospace from 1959 to 1973. In 1973 he joined Consolidated Edison Co. of New York as a senior reliability engineer.

XI = 1.635/(2T)
= 0.573%o/1000 hours

(8)

Manuscript TR82-100 received 1982 September 18; revised 1983 April 26;
revised 1984 December.

On Some Common Interests Among Reliability,

Inventory, and Queuing

(continuedfrom page 208)

Dr. Douglas R. Miller; Department of Operations Research; George


Washington University; Washington, DC 20052 USA.
Douglas R. Miller is professor, Department of Operations Research,
School of Engineering and Applied Science, George Washington University. He is a member of ASA, ORSA and the IEEE Computer Society.
His research interests include: simulation and computational analysis of
queues, inventories, and reliability models; inference and estimation
methods for fault-tolerant systems; and software reliability.

Richard M. Soland is professor of operations research at George


Washington University's School of Engineering and Applied Science, and
previously taught at the University of Texas at Austin and Ecole
Polytechnique de Montreal. His areas of current research interest are
discrete optimization and multiple criteria decision making. Author or coauthor of 37 papers that have appeared in various professional journals,
he is a senior member of IEEE and IIE, and a member of ORSA, TIMS,
CORS, and MPS.

Dr. Richard M. Soland; Department of Operations Research; George


Washington University; Washington, DC 20052 USA.***

Manuscript TR84-105 receivedl1984 August 27; revisedl1985 May 23.

S-ar putea să vă placă și