Vous êtes sur la page 1sur 10

Gamma Distribution and Gamma Approximation

Xiaoming Zeng a Fuhua (Frank) Cheng b


a Xiamen

University, Xiamen 361005, China


xmzeng@jingxian.xmu.edu.cn

b University

of Kentucky, Lexington, Kentucky 40506-0046, USA


cheng@cs.uky.edu

Abstract
The Gamma distribution and related approximation properties of this distribution to certain of classes
of functions are discussed. Two asymptotic estimate formulas are given.
Keywords: Gamma distribution, Gamma approximation, locally bounded functions, Lebesgue-Stieltjes
integral, probabilistic methods

Introduction and Main Results

Probability distributions and related approximation properties have played a central role in several
areas of computer aided geometric design (CAGD) and numerical analysis. For instance, the
(Bernstein) basis functions of Bezier curves and surfaces are taken from a binomial distribution,
and the basis functions of B-spline curves and surfaces model a simple stochastic process [1,
2]. Properties of these curves and surfaces are closely related to the approximation features
of the corresponding basis functions. Therefore studying probability distributions and related
approximation properties is important both in theory and applications.
The aim of this paper is to do a study in that direction. In this paper we shall consider the
following Gamma random variable with probability density function:
( x1
u
exp(u),
if u 0
(x)
(1)
p (u) =
0,
if u < 0
where x is a fixed number in (0, ) and (x) is the Gamma function of x defined as follows:
Z
ux1 eu du,
(x) =
0

and study related approximation process based on the above Gamma distribution. We first prove
the following lemma.

Lemma 1. Let {i }
i=1 be a sequence of independent random variables with the same Gamma
n
P
distribution as , and let n =
i . Then the probability distribution of the random variable n
i=1

is

1
P (n y) =
(nx)

unx1 eu du.

(2)

P roof . Let the distribution functions of the random variables i be Fi (x), i = 1, 2, 3, ..., and
let the distribution function of n be Fn (x). We have
Z y
Z y x1
u
Fi (y) =
pi (u)du =
eu du
(x)

0
and
Fn (y) = (F1 F2 ... Fn )(y),
where

(F1 F2 )(y) =

F2 (y u)dF1 (u),

and (F1 F2 ... Fn )(y) is defined recursively. By convolution of probability distributions we


obtain
Z y
1
unx1 eu du.
P (n y) = Fn (y) =
(nx) 0
Hence, Lemma 1 is proved.
In this paper we consider the following Gamma approximation process Gn :
Z
Z +
nnx

Gn (f, x) =
f (u)unx1 enu du
f (t/n)dFn (t) =
(nx) 0

(3)

where f belongs to the class of function B or the class of function DB , which are defined
respectively by
B = {f | f is bounded on every finite subinterval of [0, ),
and |f (t)| M et ,
and

(M > 0; 0; t )}.

DB = {f | f (x) f (0) =

h(t)dt; x 0; h is bounded on every finite


0

subinterval of [0, ), and |f (t)| M et ,

(M > 0; 0; t )}.

For a function f B , we introduce the following metric form:


x (f, ) =

sup

|f (t) f (x)|,

t[x,x+]

where x [0, ) is fixed, 0.


It is clear that
x (f, ) is non-decreasing with respect to ;
lim0 x (f, ) = 0, if f is continuous at the point x;
2

If f is of bounded variation on [a, b], and

Wb

a (f )

x (f, )

denotes the total variation of f on [a, b], then


x+
_

(f )

For further properties of x (f, ), refers to Zeng and Cheng [8].


The main results of this paper are as follows:
Theorem 1. Let f B . If f (x+) and f (x) exist at a fixed point x (0, ), then for
n > 3, we have


f (x) x + 4 X
Gn (f, x) f (x+) + f (x) + f (x+)

(g
,
x/

k) + O(n1 ), (4)
x
x

2
nx k=1
3 2nx
where

f (u) f (x+), x < u < ;


0,
u = x;
gx (u) =

f (u) f (x), 0 u < x.

(5)

We point out that Theorem 1 subsumes the case of approximation of functions of bounded
variation. From Theorem 1 we get immediately
Corollary 1. Let f be a function of bounded variation on every subinterval of [0, ) and let
f (t) = O(et ) for some 0 as t . then for x (0, ) and n > 3, we have


x+4X
f
(x+)
+
f
(x)
f
(x+)

f
(x)
Gn (f, x)

(g
,
x/
k) + O(n1 )
x
x

2
nx k=1
3 2nx

n x+x/
x+4X _

nx

k=1 xx/ k

(gx ) + O(n1 ),

(6)

Corollary 2. Under the conditions of Theorem 1, if x (gx , ) = o(), then


Gn (f, x) =

f (x+) + f (x) f (x+) f (x)

+ o(n1/2 ).
2
3 2nx

(7)

Theorem 2. Let f be a function in DB and let f (t) M et for some M > 0 and 0 as
t . If h(x+) and h(x) exist at a fixed point x (0, ), then for n > 3 we have

[ n]
4x + 2 X

x
||x5/2 + 7M (x + 1)3/2 e3x

Gn (f, x) f (x)
,

(
,
x/k)
+
x
x

2n
n k=1
x3 n3/2

(8)

where = h(x+) h(x), and

h(t) h(x+), x < t < ;


0,
t = x;
x (t) =

h(t) h(x), 0 t < x.

(9)

Remark . If f is a function with derivative of bounded variation, then f DB . Thus the


approximation of functions with derivatives of bounded variation is a special case of Theorem 2.

Auxiliary Results

To prove Theorem 1 and Theorem 2 we need some auxiliary results.


Lemma 2. For a fixed x (0, ), we have
Gn (1, x) = 1,
(nx + k 1)(nx + k 2)...(nx + 1)nx
, k>1
nk
P roof . Direct computations give Gn (1, x) = 1, and
Z
nnx

k
Gn (u , x) =
unx+k1 enu du
(nx) 0
Z
nnx
=
tnx+k1 et dt
nx+k
n
(nx) 0
(nx + k)
=
nk (nx)
(nx + k 1)(nx + k 2)...(nx + 1)nx
=
.
nk
Gn (uk , x) =

(10)

Lemma 3. For x (0, ) there hold


x
;
n
3(x + 1)2
Gn ((u x)4 , x)
;
n2
49(x + 1)3
Gn ((u x)6 , x)
;
n3
for n > 3.
Gn (e2u , x) e6x ,
Gn ((u x)2 , x) =

P roof . By Lemma 2 and direct computations we get


x
Gn ((t x)2 , x) = ;
n
3x2 6x
3(x + 1)2
Gn ((t x)4 , x) = 2 + 3
;
n
n
n2
15n2 x3 + 130nx2 + 120x
49(x + 1)3
Gn ((t x)6 , x) =

;
n5
n3
4

(11)
(12)
(13)
(14)

t
, we have
n 2
Z
nnx
2u
Gn (e , x) =
e2u unx1 enu du
(nx) 0
Z
nnx
=
tnx1 et dt
(n 2)nx (nx) 0

nx
n
=
n 2
6x
e .

In addition, if n > 3, putting u =

Lemma 4. [3, Chapter 2] Let {k }


k=1 be a sequence of independent and identically distributed
random variables with the expectation E, the variance E( E)2 = 2 > 0, E( E)4 <
n
P

, and let Fn stand for the distribution function of


(k E)/ n. If Fn is not a lattice
k=1

distribution, then the following equation holds for all t (, +)


Z t
E( )3
1
1
2
2
(1 t2 ) et /2 + O(n1 ).
eu /2 du =
Fn (t)
3
6 n
2
2

Lemma 5. Let

nnx
Kn (x, u) =
(nx)

If 0 v u < x, then
Kn (x, u)

(15)

v nx1 env dv.

x
.
n(x u)2

(16)

P roof . Let be a Gamma random variable with probability density function defined as in (1).
Let {i }
i=1 be a sequence of independent random variables with the same Gamma distribution
as . Then by direct computation we have
E() = x,

E( E)2 = 2 = x,

E( E)3 = 2x, E( E)4 = 3x2 + 6x < .


Let n =

n
P

i . By the addition operation of random variables, we obtain

i=1

E(n ) = nx,

E(n En )2 = nx

through simple computation. Thus by Chebyshev inequality it follows that


Kn (x, u) = P (|n nx| nx nu)

x
.
n(x u)2

(17)
(18)

Proof of Theorem 1

Let f satisfy the conditions of Theorem 1, then f can be expressed as

f (x+) + f (x)
f (x+) f (x)
f (x+) + f (x)
f (u) =
+ gx (u) +
sgn(u x) + x (u) f (x)
,
2
2
2
(19)

1, u = x
where gx (u) is defined in (5), sgn(u) is sign function and x (u) =
.
0, u 6= x
Obviously,
(20)
Gn (x , x) = 0.
Let {i }
i=1 be a sequence of independent random variables with the same Gamma distribution
and their probability density functions are
( x1
u
exp(u), if u 0
(x)
pi (u) =
0,
if u < 0.
where x (0, ) is a parameter. Let n =
n
P

n
P
i=1

i and Fn stand for the distribution function of

(i Ei )/ n. Then by Lemma 1, the probability distribution of the random variable n is

i=1

1
P (n y) =
(nx)
Thus
Gn (sgn(u

nnx
x), x) =
(nx)

unx1 eu du.

nx1 nu

nnx
du
(nx)

unx1 enu du

= 1 2P (n nx) = 1 2Fn (0).

(21)

By Lemma 4, (17), (18) and straightforward computation, we have


1

2Fn (0)

2
2E(1 a1 )3 1

+ O(n1 ) =
+ O(n1 ).
=
3
6 n
2
3 2nx

It follows from (19)(22) that

f
(x+)

f
(x)
f
(x+)
+
f
(x)
Gn (f, x)
|Gn (gx , x)| + O(n1 ).

2
3 2nx

(22)

(23)

We need to estimate |Gn (gx , x)|. Let


nnx
Kn (x, u) =
(nx)

v nx1 env dv.

Then by Lebesgue-Stieltjes integral representation, we have


Z

gx (u)du Kn (x, u)
Gn (gx , x) =
0

(24)

Decompose the integral of (24) into four parts, as


Z
gx (u)du Kn (x, u) = 41,n (gx ) + 42,n (gx ) + 43,n (gx ) + 44,n (gx ),
0

where
Z

xx/ n

41,n (gx ) =

gx (u)du Kn (x, u),

42,n (gx ) =

Z
43,n (gx ) =

x+x/ n

xx/ n

2x

x+x/ n

gx (u)du Kn (x, u),

gx (u)du Kn (x, u),

44,n (gx ) =

gx (u)du Kn (x, u).


2x

We will evaluate 41,n (gx ), 42,n (gx ), 43,n (gx ) and 44,n (gx ) separately. First, for 42,n (gx ), note
that gx (x) = 0. Hence
Z

x+x/ n

|42,n (gx )|

|gx (u) gx (x)|du Kn (x, u) x (gx , x/ n).

(25)

xx/ n

To estimate |41,n (gx )|, note that x (gx , ) is non-decreasing with respect to , thus it follows
that
Z
Z

xx/n

xx/ n

|41,n (gx )| =
gx (u)du Kn (x, u)
x (gx , x u)du Kn (x, u).
0

Using integration by parts with y = x x/ n, we have


Z y
Z y
x (gx , x u)du Kn (x, u) x (gx , x y)Kn (x, y) +
Kn (x, u)du (x (gx , x u)). (26)
0

From (26) and Lemma 5 we get


x
+
|41,n (gx )| x (gx , x y)
n(x y)2

x
du (x (gx , x u)).
n(x u)2

(27)

Using integration by parts once again, from (26), (27) it follows that
1
2x
|41,n (gx )|
x (gx , x) +
nx
n

Putting u = x x/ t for the last integral we get


Z
0

xx/ n

xx/ n

x (gx , x u)
1
du =
3
(x u)
nx

Consequently
1
|41,n (gx )|
nx

Z
x (gx , x) +

x (gx , x u)
dt.
(x u)3

x (gx , x/ t)dt.

x (gx , x/ t)dt .

(28)

Using a similar method to estimate |43,n (gx )|, we get

Z n

1
x (gx , x/ t)dt .
x (gx , x) +
|43,n (gx )|
nx
1
7

(29)

Finally, by assumption we know that gx (t) M (et ) as t . Using Holder inequality and
Lemma 3, we have, for n 3,
R
|44,n (gx )| M R 2x eu du Kn (x, u)

M
(u x)2 eu du Kn (x, u)
x2 R
0
1/2 R 2u
1/2
(30)

(u x)4 du Kn (x, u)
e du Kn (x, u)
M
x2
0
0
3x
2M (x+1)e
.
n
Equations (25) and (28)(30) lead to
|Gn (gx , x)| |41,n (gx )| + |42,n (gx )| + |43,n (gx )| + |44,n (gx )|

Rn

2
x (gx , x/ n) + nx
x (gx , x) + 1 x (gx , x/ u)du +
n

P
3x
x+4

(g
,
x/
k) + 2M (x+1)e
.
x
x
n
n

2M (x+1)e3x
n

(31)

k=1

Theorem 1 now follows from (23) and (31).

Proof of Theorem 2

By direct computation we find that


Gn (f, x) f (x) =

h(x+) h(x)
Gn (|u x|, x) An,x (x ) + Bn,x (x ) + Cn,x (x ),
2

where

An,x (x ) =

x (u)du dt Kn (x, t),


0

2x

Bn,x (x ) =

x (u)du dt Kn (x, t),


Z

x (u)du dt Kn (x, t),

Cn,x (x ) =
2x

Integration by parts derives


Z

x (u)du dt Kn (x, t)
x Z x
Z x

Kn (x, t)x (t)dt


x (u)duKn (x, t) +
=
0
t
0
Z
!

An,x (x ) =

xx/ n

+
0

xx/ n

Kn (x, v)x (v)dv

Note that Kn (x, v) 1 and x (x) = 0, it follows that

[ n]
X

x
2x
x

Kn (x, nv)x (v)dv x x ,


x (x , x/k).

n k=1
n
n
xx/ n

(32)

On the other hand, by Lemma 5 and using change of variable v = x x/u, we have

xx/n
x xx/ n x (x , x v)

Kn (x, v)x (v)dv


dv

n 0
(x v)2
1
=
n

Thus, it follows that

[ n]

1X
x (x , x/u)du
x (x , x/k).
n k=1

|An,x (x )|

[ n]
2x + 1 X

x (x , x/k).

(33)

2x + 1 X
|Bn,x (x )|
x (x , x/k).
n k=1

(34)

k=1

A similar evalutation gives

[ n]

Next we estimate Cn,x (x ), by the assumption that f (t) M et (M > 0, 0), and using
Lemma 3 we have
Z +
et dt Kn (x, t)
|Cn,x (x )| M
Z2x
M +

(t x)3 et dt Kn (x, t)
x3 2x
1/2
1/2 Z +
Z +
M
2t
6
e dt Kn (x, t)

(t x) dt Kn (x, t)
x3
0
0

7M (x + 1)3/2 e3x
.
x3 n3/2

(35)

Finally, we estimate the first order absolute moment of the Gamma approximation process:
x|, x). By direct calculation, we have
Z
nnx

Gn (|u x|, x) =
|u x|unx1 enu du
(nx) 0

Z x
Z
nnx
nx1 nu
nx1 nu
=
(x u)u
e du +
(u x)u
e du
(nx)
0
x
Z x
2nnx
=
(x u)unx1 enu du
(nx) 0
Z nx
Z nx
2x
2
nx1 t
=
t
e dt
tnx et dt.
(nx) 0
n(nx) 0

Gn (|u

Note that

nx

Z
nx t

nx nx

t e dt = (nx) e

nx

+ nx

tnx1 et dt,

thus
Gn (|u x|, x) =
9

2(nx)nx
.
nenx (nx)

(36)

Now using Stirlings formula (cf. [4]):

(z + 1) = 2z(z/e)z ecz ,
from (36) we have

(12z + 1)1 < cz < (12z)1 ,

r
2x
2x
Gn (|u x|, x) =
(1 ecx ),
n
n
and a simple calculation derives
r
1
2x
1
3/2
(1 ecx ) 3/2 .
n
18 xn
12 xn

(37)

(38)

Theorem 2 now follows from (32)(38) combining with some simple calculations.

Acknowledgement
The first author was supported by National and Fujian provincial Science Foundation of China.
The second author is supported by USA NSF under grants DMI-0422126 and DMS-0310645.

References
[1]

W. Dahmen and C. Micchelli, Statistical encounters with B-spline, AMS Series on Contemporary
Mathematics, 59 (1986), 17-48.

[2]

R. N. Goldman, Urn models and B-splines, Constructive Approximation, 4 (1988), 265-288.

[3]

W. Feller, An Introduction to Probability Theory and Its Applications, John Wiley & Sons, Inc.
New York, London, Toronto, 1971.

[4]

V. Namias, A simple derivation of Stirings asymptotic series, Amer. Math. Monthly, 93 (1986),
25-29.

[5]

E. Omey, Operators of probabilistic type, Theory Prob. Appl., 41 (1996), 219-225.

[6]

A. N. Shiryayev, Probability , Springer-Verlag, New York. 1984.

[7]

C. Zhang and F. Cheng, Triangular patch modeling using combination method, Computer Aided
Geometric Design, 19 (2002), 645-662.

[8]

X. M. Zeng and F. Cheng, On the rates of approximation of Bernstein type operators, J. Approx.
Theory 109 (2001), 242-256.

10

Vous aimerez peut-être aussi