PROBABILITY SPACE
Theorem 1. The function (1) has the following properties
F1. If x1x2 then F x1 dF x2 (i.е. F F x( ) is a monotonically nondecreasing function);
F2. lim ( ) 0
F x F x
f p f , lim ( ) 1
F x F x
f n f ;
F3. F x( ) is a right-continuous function
F x( 0) F x( ) ,x R , it has left limits at each point xR.86
Proof. The property F1 is the corollary of the following. As
@ @
1 , 1 , 2 2
x x
A f x f x A , then by the property of probability P A
x1 dP Ax2 . The property F2 is the corollary of the following. From xnp f it implies that f, xn
@
p , from ynn f it implies that f, yn@
nR. We also use a continuity from below (property Р3″) and continuity from above (property Р3′) of probability and monotonicity of the function ( )F x .It is not difficult to see that property F3 is also a consequence of the properties of continuity from below and from above. ז Definition 1. The function F F x( ) satisfying properties F1, F2, F3 is called the distribution function on the number line R.
Thus, by the Theorem 1, the distribution function F F x( ), defined by (1), corresponds to each probability function P in the space
R, ( )E R . It turns out that the converse is also true.Theorem 2. Let the function F F x( ) be the distribution function on the number scale R f f( , ).
Then on
R, ( )E R there is the only one probabilistic measure P such that for any interval f d fa b the following takes place: ,@
( ) ( )P a b F b F a . (2) Proof. By the theorem on the extension of the probability, for constructing a probability space
R,E R ,P, it suffices to specify the probability P on the algebraࣛ generated by intervals of the form
a,b@
(because ߪሺࣛሻ ൌ ߚሺܴሻ). But we know that any element A of algebra ࣛ can be written in the form of a finite sum of disjoint intervals of the form a,b@
:@
1
, ,
n
i i i i
i
A
¦
a b a b. (ai, bi may be infinite). Let’s by definition0 1
n
i i
i
P A
¦
ª¬F b F a º¼.Then the properties F1, F2 are satisfied, therefore the axioms Р1, Р2 are also fulfilled. Now it remains only to verify the countable additivity (or continuity) of ܲ on ࣛ.
Let ܤ א ࣛ, Bn1Bn, fB B n n
1
ࣛ.
87
Let’s show that P B0 n oP B0 , when ݊ ՜ λ (property of continuity «from below»).
Without loss of generality, we can assume that B consists of only one interval
a b,@
: B a b,@
. As BBn, then for any n there is an interval a bn, n@
, such that a bn, n@
Bn, and such that B a b,@
a bn, n@
. Let a semi-interval a bn, n@
be contained in Bn and be a maximal half-interval containing B. Furthemore, Bn, starting with some n, does not contain any half-intervals outside a bn, n@
(if c d,@
Bn for all n, then c d,@
B). Thus, from monotonicity of Bnand the fact that B a b,@
itfollows, that starting from some number, for all n, Bn
a bn, n@
, where ܽ ൌ ܽǡ ܾ ՝ ܾ. But by the property F3 ,0 n n n
P B F b F a oF b F a .
Thus, axiom P3 is also fulfilled. ז Thus, there exists a one-to-one correspondence between the probability functions
P defined on
R, ( )E R and the distribution functions F F x( ) that are defined on the number scale and satisfy the conditions F1, F2, F3.The probability measure P constructed by the distribution function F F x( )is usually called the Lebesgue-Stiltes probability measure. Particularly important is the case when
0 , 0,
( ) , 0 1,
1, 1
x
F x x x
x
° d d
®° !
¯
In this case, the probability measure corresponding to F(x) and denoted by λ is called the Lebesgue measure on the segment
> @
0,1 . It is clear that Oሺܽǡ ܾ༱ ൌ ܾ െ ܽ.In other words, the Lebesgue measure of the interval
a b,@
, as well as of any of the intervals a b, , a b,@
,>
a b, ,> @
a b, is just its length ܾ െ ܽ : a b,@
a b,>
a b,> @
a b, b aO O O O .
According to condition F3, the distribution function F F x( ) can have only points of discontinuity of the first kind. The value
( 0) ( 0) ( ) ( 0)
px F x F x F x F x
88
is called a jump of the distribution function ( )F x at the point x. A jump of the distribu- tion function is equal to zero at points of continuity and is strictly greater than zero at the points of discontinuity.
If for any H !0, the inequality (F x H) F x( H) 0! takes place then such a point x R is called a point of growth of the distribution function ( )F x .
Theorem 3. Any distribution function F F x( ) has at most a countable number of discontinuity points.
Proof. Denote by ( )C F (here and everywhere below) the set of points of con- tinuity of F F x( ), by C(F) – the complementary set (the set of points of discon- tinuity), and by Dk – the set of points of discontinuity, in which the jumps of the func- tion ( )F x lie in the half-open interval 1 ,1 , 1,2,...
1 k
k k
§ º
¨ »
© ¼ . So, by definition:
^ `
( ) : ( ) ( 0) , ( ) \ ( )
C F x R F x F x C F R C F ,
1 1
\ ( ) : ( ) ( 0) ,
k x 1
D x R C F F x F x p
k k
§ º½
® ¨© »¼¾
¯ ¿, k 1,2,... .
Then the number of elements of each set Dk is at most k, therefore the number of elements of the set
1
( ) \ ( ) k
k
C F R C F f DDkk (as a countable sum of not more than a countable number of sets) is at most countable. ז Remark 1. We obtain from the results of the theorem we have proved that the set ( )C F is everywhere dense on the real line R. But it turns out that C F can also be a countable everywhere dense set on R.
Example.
Let
^
r r1, ,...2`
be a set of rational numbers on a line (it is well known that this set is everywhere dense on R). To each point rk we put in correspondence a jumpk 2
k
pr . Then the set of points of discontinuity of a function
:
( ) k
k
k r x r
F x p
¦
d is theset C F
^
r r1, ,...2`
, and its closure is ª¬C F º ¼ R.The fact that the function defined in this way is a distribution function is verified directly.
A distribution function F x that changes its values only at points of a finite or countable set X
^
x x1, ,...2`
is called a discrete distribution function.If the interval
a bj, jº¼ contains only one point xjC F( ), then j, j ( )j ( )jP a b º ¼ F b F a . If ajnxj , bj pxj , then
89
^ `
( ) ( 0) ,^ `
1j j j j j
j
p P x F x F x
¦
P x .It is obvious that the introduced discrete distribution function F x can be rep- resented in the form
: j j
j x x
F x p
¦
d .The set of numbers p p1, 2,..., where pj P x
^ `
j , j 1j
¦
p , is called the dis- crete probability distribution.Examples of frequently occurring discrete probability distributions are given in Table 1 below.
Table 1 Discrete distributions
Distribution Name Probabilities pk Parameters
Discrete uniform k N
N1, 1,2,... . N 1,2,....
Bernoulli p1 p, p0 q. 0dpd1, q 1p.
Binomial Cnkpkqnk,k 0,1,2,...,n. 0d pd1,q 1p,n 1,2,...
Poisson , 0,1,2,...
!
k
e O Okk
. O!0.
Geometric qk1p , k 1,2,.... 0pd1, q 1p.
Negative binomial
(Pascal) Ckr11prqkr,k r,r1,... 0 pd1, q 1p. Hypergeometric ,k 0,1,...min(n,M).
C C C
Nn k
n M
k N
M
N, M, n are positive integers,
n N M
N ! , !
If there is an amnonnegative function ( )f x
f x t0 such that for all x R we can write down the distribution function F x as an integralx
F x f u du
f
³
. (3) Then such a distribution function is called an absolutely continuous distribution function.90
The integral (3), in the general case, must be understood in the sense of the Le- besgue integral ([9]), but for our purposes (in this textbook) this integral is sufficiently understood as an (improper) Riemann integral.
From the definition of the distribution functions we obtain that any nonnegative function ( )f x , which is integrable by Riemann integral and such that
( ) 1
f x dx
f f
³
determines by formula (3) a distribution function F F x( ).
In Table 2 we give examples of different distribution densities f x( ) that are especially important for probability theory and mathematical statistics, with indicating their names and parameters.
Table 2 Absolutely continuous distributions
Distribution Name Density function f(x) Parameters
Uniform on
> @
a,b b a adxdb1 ,
,
> @
ab x , ,0 .
b a R b
a, , .
Laplace V
V
a x
e
2
1 , fxf. a,VR, V !0.
Normal or Gaussian
f f
x e
a x
2 ,
1 2
2
2V
V
S . a,VR, V !0.
Gamma
0
1 , !
e x
xO Tx
O
O T
Г ,
0 ,
0 x .
, 0 , ,O T! T R
!0 O . Beta
௫ೝషభሺଵି௫ሻೞషభ
ሺǡ௦ሻ ǡ Ͳ ݔ ͳ,
Ͳǡ ݔ ב ሺͲǡͳሻ ݎ Ͳǡ ݏ Ͳ.
Cochi S[T2xa2]
T , fxf.
0 ,
, T!
T a R . Chi-square with n
freedom degrees, Fn2
0 , 2 2
1 2 1 2
2
t
¸¹
¨ ·
©
§
e x x
n
x n n
Г
,
0 ,
0 x .
,...
2 , 1
n .
Стьюдента с n степенями свободы, t
2 1 2
1 2 2 1 1
¸¸¹
·
¨¨©
§
¸¹
¨ ·
©
§
¸¹
¨ ·
©
§ n
n x n
n
n Г Г
S ,
f f
x .
,...
2 , 1
n .
91 F, F(m, n),
Fisher-Snedekor
2 2 1 2
2 1 2
n m m m
n mx x n B m
n m
¸¹
¨ ·
©§
¸¹
¨ ·
©
§
¸¹
¨ ·
©
§
,
, xt0,
0
0, x .
,...
2 , 1 ,m
n .
Exponential OeOx , xt0,
0 ,
0 x .
!0 O . Two-sided exponential OeOx
2 , fxf. O!0.
In Table 2, Г( ),x B x y( , ) are (respectively) the gamma function and the beta function:
1 0
( ) t x
Г x f
³
e t dt ,1 1 1
0
( ) ( )
( , ) (1 )
( )
x y Г x Г y
B x y t t dt
Г x y
³
.From Table 2, we note the following: if O 1, then gamma-distribution is expo- nential; if 1,
2 2
T O n (n is an integer), then gamma-distribution is Fn2 – distribution.
In the probability theory, beta-distribution with the parameters 1
p q 2 is cal- led the law of arcsine.
It turns out that there is a third type of distribution function – this is the so-called singular distribution function. This type is characterized by the fact that the distribution function F F x( ) is continuous, but the growth point forms a set of Lebesgue mea- sure zero.
Thus, a singular distribution function F x is continuous, but F xc 0 almost everywhere, and F f F f 1.
An example of a singular distribution function is the Cantor function.
Cantor function. This function is constructed as follows:
0, 0; 1, 1
F x xd F x xt .
For x
> @
0,1 define this function as follows. First, the segment> @
0,1 is divided into three identical parts:>
0,1/ 3 , 1/ 3, 2 / 3 , 2 / 3,1@ > @ > @
. On the middle segment, we assume F x 1/ 2. The remaining two extreme segments are again divided into three equal parts each and on the inner segments we assume that F x 1/ 4 and92
3 / 4
F x (respectively). Further, each of the remaining segments is again divided into three equal parts and on the inner segments we define F x as a constant which is equal to the arithmetic mean between the adjacent, already defined values of F x , etc. At points that do not belong to such internal segments, we define F x by conti- nuity. It is not difficult to see that the total length of the «internal» segments, on which
x
F is constant, is equal to
1 2 4 1 2 2 2 1 1
... 1 ... 1
3 9 27 3 3 3 3 1 2 / 3
§ § · ·
¨¨© ¨ ¸© ¹ ¸¸¹ ,
so that the function F x grows on a set of measure 0 (zero), but without jumps.
Remark 2. The question arises: is it possible to find the types of distribution