164 THE BINOMIAL AND THE POISSON DISTRIBUTIONS [VI.

TABLE 7
EXAMPLE ( e) : COUNTS OF BACTERIA

k 0 1 2 3 4 5 6 7 X-
Level
----
Observed Nk 5 19 26 26 21 13 8 97
Poisson theor. 6.1 18.0 26.7 26.4 19.6 11.7 9.5
----
Observed Nk 26 40 38 17 7 66
Poisson theor. 27.5 42.2 32.5 16.7 9.1
----
Observed Nv 59 86 49 30 20 26
Poisson theor. 55.6 82.2 60.8 30.0 15.4
----
Observed Nk 83 134 135 101 40 16 7 63
Poisson theor. 75.0 144.5 139.4 89.7 43.3 16.7 7.4

Observed Nk 8 16 18 15 9 7 97
Poisson theor. 6.8 16.2 19.2 15.1 9.0 6.7
----
Observed Nk 7 11 11 11 7 8 53
Poisson theor. 3.9 10.4 13.7 12.0 7.9 7.1

Observed Nk 3 7 14 21 20 19 7 9 85
Poisson theor. 2.1 8.2 15.8 20.2 19.5 15 9.6 9.6
--
Observed Nk 60 80 45 16 9 78
Too'
th",,.,. .. t::.'") £ '7C 0 AC 0 10 C '7 '1
~-.~ ~.-o "TJ . V ~V.J • .J

The last entry in each row includes the figures for higher classes and should be
labeled "k" or more."

representative of an important practical application of the Poisson dis-

tribution to spatial distributions of random points. ~

8. WAITING TIMES. THE NEGATIVE

BINOMIAL DISTRIBUTION

Consider a succession of n Bernoulli trials ancllet us inquire how long

it will take for the rth success to turn up. Here r is a fixed positive integer.
The total number of Successes in n trials may, of course, fall short of r,
but the probability that the rth success occurs at the trial number v < n
VI.S] WAITING TIMES. THE NEGATIVE BINOMIAL DISTRIBUTION 165

is clearly independent of n and depends only on 'V, r, and p. Since

necessarily 'V >r, it is preferable to write 'V = k + r. The probability
that the rth success occurs at the trial number r + k (where k = 0, 1, ... )
will be denoted by !(k; r, p). It equals the probability that exactly k
failures precede the rth success. This event occurs if, and only if, among
the r + k - 1 trials there are exactly k failures and the following, or
(r+k)th, trial results in success; the corresponding probabilities are
(r+k -l \
\ J pr lqlc and p, whence
k

(S.l)

Rewriting the binomial coefficient in accordance with 11,(12.4), we find

the alternative form

(S.2) k = 0,1,2, ....

Suppose now that Bernoulli trials are continued as long as necessary for
r successes to turn up. A typical sample point is represented by a sequence
containing an arbitrary number, k, of letters F and exactly r letters S,
the sequence terminating by an S; the probability of such a point is, by
definition, prqk. We must ask, however, whether it is possible that the
trials never end, that is, whether an infinite sequence of trials may produce
OCJ

fewer than r successes. Now "2.f(k; r,p) is the probability that the rth
k=O
success occurs after finitely many trials; accordingly, the possibility of an
infinite sequence with fewer than r successes can be discounted if, and
only if,
OCJ

(S.3) "2. f(k; r, p) = 1.

k=O

This is so because by the binomial theorem

(S.4) ."2.
OCJ

k=O
(-r)
k
(_q)k = (l_q)-r = p-r.

Multiplying (S.4) by pr we get (S.3).

In our waiting time problem r is necessarily a positive integer, but the
quantity defined by either (S.1) or (S.2) is non-negative and (S.3) holds for
any positive r. For arbitrary fixed real r > and ° °
< p < 1 the
sequence {f(k; r, p)} is called a negative binomial distribution. It occurs
in many applications (and we have encountered it in problem 24 of V, as
166 THE BINOMIAL AND THE POISSON DISTRIBUTIONS [VI.S

TABLE 8
THE PROBABILITIES (8.5) IN THE MATCH Box PROBLEM

r Ur Dr r Ur Ur

0 0.079589 0.079589 15 0.023 171 0.917941

1 0.079589 0.159 178 16 0.019 081 0.937022
2 0.078 785 0.237963 17 0015447 0952469
3 0.077 177 0.315 140 18 0.012283 0.964752
4 0.074790 0.389 931 19 0.009587 0.974338

5 0.071 674 0.461 605 20 0.007338 0.981 676

6 0067902 0529 506 21 0.005504 0.987 180
7 0.063 568 0.593 073 22 0.004041 0.991 220
8 0.058 783 0.651 855 23 0.002901 0.944 121
9 0.053 671 0.705 527 24 0.002034 0.996 155

10 0.048 363 0.753 890 25 0.001 392 0.997547

11 0.042989 0.796879 26 0.000928 0.998475
12 0.037676 0.834555 27 0.000602 0.999077
13 0.032538 0.867094 28 0.000379 0.999456
14 0.027676 0.894770 29 0.000232 0.999688

Ur is the probability that, at the moment for the first time a match box is found
empty, the other contains exactly r matches, assuming that initially each box
contained 50 matches. Ur = U o + UI + ... + u/ is the corresponding prob-
ability of having not more than r matches.

the limiting form of the Polya distribution). When r is a positive integer,

{f(k; r,p)} may be interpreted as the probability distribution for the
waiting time to the rth success; as such it is also called the Pascal distribu-
tion. For r = 1 it reduces to the geometric distribution {pqk}.
Examples. (a) The problem of Banach's match boxes. Is A certain
math€matician always carries one match box in his right pocket and one in
his left. When he wants a match, he selects a pocket at random, the suc-
cessive choices thus constituting Bernoulli trials with p = t. Suppose that
initially each box contained exactly N matches and consider the moment
when, for the first time, our mathematician discovers that a box is empty.

18 This example was inspired by a humorous reference to Banach's smoking habits

made by H. Steinhaus in an address honoring Banach. It became unexpectedly popular

in the literature and for this reason I leave the name unchang~d. References to Banach's
Oeuvres completes are, of course, spurious.
VI.9] THE MULTINOMIAL DISTRIBUTION 167

At that moment the other box may contain 0, 1, 2, ... , N matches, and
we denote the corresponding probabilities by U r • Let us identify "success"
with choice of the left pocket. The left pocket will be found empty at a
moment when the right pocket contains exactly r matches if, and only if,
exactly N - r failures precede the (N + l)st success. The probability of
this event is f(N-r; N+l, t). The same argument applies to the right
pocket and therefore the required probability is

(8.5)

Numerical values for the case N = 50 are given in table 8. (Cf. problems
21, and 22, and problem 11 of IX,9).
(b) Generalization: Table tennis. The nature of the preceding problem
becomes clearer when one attributes different probabilities to the two
boxes. For a change we interpret this variant differently. Suppose that
Peter and Paul playa game which may be treated as a sequence of
Bernoulli trials in which the probabilities p and q serve as measures for
the players' skill. In ordinary table tennis the player who first accumulates
21 individual victories wins the whole game. For comparison with the
preceding example we consider the general situation where 2v + 1 indi-
vidual successes are required. The game lasts at least 2v + 1 and at most
4v + 1 trials. Denote by ar the probability that Peter wins at the trial
number 4v + 1 - r. ThIS event occurs If, and only If, III the first 4v - r
trials Peter has scored 2v successes and thereafter wins the (2v + 1)st
trial. Thus

(8.6)

In our game ao + ... + a2N is the probability that Peter wins. The
probability that the game ends exactly at the trial number 4v + 1 - r
is given by ar + br , where br is defined by (8.6) with p and q inter-
changed.
If vie put 211 }1 and p q ~,the probabilities ar I b r reduce
to the probabilities U r of the preceding example. ~

9. THE MULTINOMIAL DISTRIBUTION

The binomial distribution can easily be generalized to the case of n

repeated independent trials where each trial can have one of several
outcomes. Denote the possible outcomes of each trial by E 1 , ••• ,Er , and
suppose that the probability of the realization of Ei in each trial is
"