行政院國家科學委員會專題研究計畫 成果報告
有限個數的 i.i.d. Weibull 分布隨機變數之和的機率分布
函數的近似值
計畫類別: 個別型計畫 計畫編號: NSC93-2118-M-009-018- 執行期間: 93 年 08 月 01 日至 94 年 07 月 31 日 執行單位: 國立交通大學統計學研究所 計畫主持人: 彭南夫 報告類型: 精簡報告 處理方式: 本計畫可公開查詢中 華 民 國 94 年 10 月 31 日
有限個數的 i.i.d. 韋伯隨機變數之和的累積機率函數近似解
關鍵字 : 韋伯隨機變數 , 累積機率函數
An Approximation of the Distribution Function of the Sum of Finite
i.i.d. Weibull Random Variables
Keywords : Weibull random variables, Cumulative Distribution
function
1.INTRODUTION
The main purpose in this discussion is to find the distribution function of the sum
of independent and identical distributed Weibull random variables. The Weibull
distribution is a two-parameter distribution which by adjusting a scale parameter
(denoted by c) and a shape parameter (denoted by β ) we can obtain a variety of shapes to fit experimental data. Thus, this distribution is highly adaptable and widely
used in practice. Without loss of generality, we can assume the scale parameter to be
one. In addition, the Weibull distribution is widely used in reliability modeling since
other distributions such as exponential (c=1), Rayleigh (c=2), and normal (if a
suitable value for the shape parameter is chosen) are special cases of the Weibull
distribution. In the version of hazard-rate function, when c>1, the hazard rate is a
monotonically increasing function with no upper bound that describes the wear-out
region, and when c=1, the hazard rate becomes constant (constant failure-rate region),
failure-rate region). This enables the Weibull distribution to describe the failure rate of
many failure data in many kinds of application.
Unlike the renewal processes with the exponential inter arrivals, those with the
Weibull interarrivals do not possess elegant properties. It is well known that by the
nature of the renewal processes, the distribution of the process is related to the
distribution of a finite sum of the independent and identically distributed inter arrivals,
i.e. suppose that inter arrivals X1,X2,...,Xn are a random sample of Weibull(c,β ) distributed with probability density function
β β c x c e x c x f = −1 − ) ( . (1)
The probability of interest is
P(Sn ≤ where t) Sn = X1+...+Xn, (2) which has no simple formula to evaluate. The major result of this report is to present a
simple, elegant and relatively accurate form of an approximation.
2. METHODOLOGY
In this section, we discuss the approximation of Fn(t)=P(Sn ≤t), which could be seen as the convolution of n functions. When n is large, the evaluation of n-fold
convolution includes n-dimensional integration, thus the evaluation becomes difficult.
Even using the numerical analysis, the results of evaluation are obtained pointwisely.
That is, when fixed one value of t, we can numerically evaluate Fn(t), but just only one value in every evaluation. The numerical computation of the inverse Laplace
transform [8] still directs to the answer pointwisely. Another approach is to use
Edgeworth Expansion [3 and 7], but the shortage is that we have to includ numerical
would not behave well relatively. Here we use three quantiles to obtain a simple
approximation method, and the evaluation of approximation is obtained
functionwisely, not pointwisely. Our approach describes as the following:
Suppose X is a random variable from i Weibull( c, 1)distribution. Via the transform of variable Yi = Xic, we obtain that Y is a random variable from i Exponential(1) distribution with i=1,...,n. Thus, we expect to use the distribution function of
n Y
Y1+...+ to approximate the distribution function of X1+ ...+Xn. Since the two distribution functions are continuously differentiable functions, for every t≥0, we expect to find some continuously differentiable function w(t), such that
P(X1+...+Xn ≤t)=P(Y1+...+Yn ≤w(t)), (3) with 0w(0)= . Thus, we consider the form of third degree polynomial
) (t
w =αt3c +γt2c +τtc. (Note that we can consider the form αt2c +γtc, but the approximation isn’t well relatively.) Three unknown parameters above can be
determined by three quantiles of X1+ ...+Xn and Y1 + ...+Yn. First, let t satisfy p
P(X1 +...+Xn ≤tp)= p, (4) and θ satisfy p
P(Y1+...+Yn ≤θp)= p, (5) Thus, t and p θ are the p-quantile of p X1 + ...+Xn and Y1+ ...+Yn, respectively. When p 0.1, 0.5, and 0.9, we can apply = αtp3c+γ tp2c +τtpc =θp to construct three linear functions. Thus, three unknown parameters can be evaluated through the three
Lemma 1:
Suppose Y1,Y2,...,Yn are a random sample from Exponential(1) distributed. Then
∑
− = − − = ≤ + + 1 0 1 ... ) 1 / ! ( n i i n e i Y Y P θ θθ . (6) Proof: First let∑
= = n i i n Y S 1 *, there is a counting process {N(θ)=sup(n:Sn* ≤θ),θ ≥0} which is a renewal process, conclude that
Sn* ≤θ ⇔ N(θ)≥n. (7) As inter arrivals are a random sample from Exponential(1) distributed, we know that )N(θ is a random variable from Poisson(θ) distributed. Thus
P(Y1+...+Yn ≤θ)=P(Sn* ≤θ)=P(N(θ)≥n)
∑
− = − − = < − = 1 0 ! / 1 ) ) ( ( 1 n i i i e n N P θ θθ (8) ■Via lemma 1, the value of θ can be obtained by numerical method (We use the p Newton-Raphson method). Different to θ which follows lemma1, we must p simulate the value of t . We interpret the procedure of the approach as the follows: p
(1) Through Newton-Raphson method we can obtain the solutions of the equation
p / i! θ e n i i p θP = −
∑
− = − 1 01 , where p 0.1, 0.5, 0.9, i.e. = θ0.1 ,θ0.5,and θ0.9.
(2) Determine the numbers of simulate data N =107, let T1,T2,...,TN be a random sample from Fn(x)=P(X1 +...+Xn ≤x) , and order them become to
) ( ) 2 ( ) 1 ( ,T ,...,TN
where p 0.1, 0.5, 0.9. = (3) Via the three linear equations
1 0 1 0 2 1 0 3 1 0 ˆ ˆ ˆ c . . c . c . γ t τ t θ t α + + = , 2 05 05 5 0 3 5 0 ˆ ˆ ˆ c . . c . c . γ t τ t θ t α + + = , (9) 9 0 9 0 2 9 0 3 9 0 ˆ ˆ ˆ c . . c . c . γ t τ t θ t α + + = .
Three of unknown parameters α γ,, τ can be determined.
(4) Under fixed n and shape parameter c, the estimated form of w(t) is wˆ(t)=αt3c +γ t2c +τtc
. (10)
(5) Thus the approximation of the distribution function of interest is
∑
− = − − = ≤ + + ≅ ≤ + + 1 0 ) ( ˆ 1 1 ... ) ( ... ˆ( )) 1 ˆ( ) / ! ( n i i t w n n t P Y Y wt e w t i X X P . (11)We choose the number of simulate data is N =107, and T( Np) is the p-th sample quantile. From [6] we can know that T( Np) is asymptotically following the
) ) ( ) 1 ( , ( 2 N t f p p t N p n p −
distributed, the error is proportional to 2 7
10
1 −
=
N , thus T( Np)
would close to the true p-th quantile. Equation (11) is the most important result of this
dissertation and offers a simple approximation via the correlation between the
Exponential and the Weibull distributed. Moreover, the approximation approach will
reduce a lot of time relative to simulation. In the next section we will discuss the
accuracy of the approach.
In order to appraise the accuracy of our approximation, the criterion that be used is the
relative error. We denote the relative error is P
P P− ˆ
, where p is the true probability and ))Pˆ =P(Y1+...+Yn ≤wˆ(T(Np) is the approximation of p. Total cases that we have done are as n=2, 3,..., 10, c=0.1, 0.2,..., 0.9 and c=2, 3,..., 9. In table 3.1, we list the relative error table includes only cases of n=2, 5and9, c = 0.5 and 5, the others have the analogous result and tendency. Fixed n and c, in every block of
table 3.1 includes the relative errors of probabilities 0.05, 0.1, …0.95. Further, we list
the same cases of three of parameters in table A.1 of Appendix A.
We can discover that when c=0.5, the maximum relative error is 0.14%; and c=5, the
maximum relative error is 1.4%, noted that happened as p=0.05, concludes that the
absolute error isn’t very large. Hence, we believe that Equation (11) brings a nice
accuracy of approximation. [Table 3.1] Fixed as n=2 p c 0.5 5 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.000726445242646 -0.000000002616051 0.000119820278476 -0.000194908815901 0.000138893117596 -0.000011260430451 0.000043543016760 0.000186524455173 0.000024819173472 -0.000000000567722 -0.000148473532176 -0.000059519398334 -0.010161527424567 -0.000000000000001 0.003828824693817 0.005143137693705 0.005324085016400 0.005005107033746 0.004334074875617 0.002847272084006 0.001416814541063 0.000000000000000 -0.001696971491006 -0.003003892633153
0.65 0.7 0.75 0.8 0.85 0.9 0.95 -0.000040240442348 -0.000105114490566 -0.000172923165788 -0.000064298286200 0.000007023458372 -0.000000001380996 0.000165231376151 -0.004332428288305 -0.005246866270862 -0.005520852102506 -0.005102997297659 -0.003342411510635 0.000000000000000 0.004920445151018 Fixed as n=5 p c 0.5 5 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 -0.000028135251537 -0.000000007634631 0.000008069914397 0.000509988574720 0.000340433284307 0.000628625741176 0.000642236022542 0.000125130856831 0.000196852924741 -0.000000001023538 0.000325840009388 0.000290632220714 0.000252098046966 0.000277480011670 0.000254161667295 0.000199371124861 0.000174674060825 -0.000000001927334 -0.000021837121247 -0.014064661989290 0.000000000000002 0.003784904432964 0.005825262499093 0.006454284943475 0.005655125010712 0.004201874492865 0.002810794471418 0.001418848371266 0.000000000000002 -0.001174124992608 -0.002266358442895 -0.002949416858121 -0.003413516974510 -0.003491304489550 -0.003091183240431 -0.001950997481086 0.000000000000001 0.002886342734995 Fixed as n=9 p c 0.5 5 0.05 0.001436234621018 -0.010512189365780
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 -0.000000012003200 0.000557973789439 0.000524827975903 0.000174871687116 0.000275117740350 0.000340236827250 0.000406740368423 0.000277355403994 -0.000000001422286 0.000113761481000 -0.000128186125497 -0.000099682351945 0.000002286246061 0.000047694801632 -0.000065649121020 0.000010740192212 -0.000000002419210 0.000015724643715 0.000000000000005 0.004208106793534 0.004523974335599 0.004347404044753 0.003550649249438 0.002914461327970 0.001891905365702 0.000952014957686 -0.000000000000003 -0.000865972087613 -0.001469193796314 -0.002062921597401 -0.002299894298618 -0.002261558612072 -0.001909441671796 -0.001183948276717 -0.000000000000003 0.001799676396769
APPENDIX A
[Table A.1]Fixed n and c, there are three parameters in every block, where α,γ and τ are in the upper, middle and down side, respectively.
n c 0.5 5 2 0.000280751644006 -0.010114574919690 1.255212150831596 0.000004835157895 -0.000513436509664 0.087023676809991 3 -0.000395335224320 -0.012406564768126 1.464308967719061 0.000000025577837 -0.000019005934388 0.018978874641112 4 -0.000224864735582 -0.016947170359548 1.650435209948320 0.000000000555125 -0.000001658936240 0.006292385415443
5 -0.000493569188358 -0.018315452950023 1.816576635488150 0.000000000026227 -0.000000236200742 0.002643130317068 6 -0.000726355289725 -0.018704570924749 1.967622246201590 0.000000000002166 -0.000000047912910 0.001296831369222 7 -0.000497097072477 -0.022139525917010 2.112546353444382 0.000000000000270 -0.000000012598155 0.000709991929183 8 -0.000863384052326 -0.020185034372005 2.239924284448030 0.000000000000042 -0.000000003807886 0.000419152653162 9 -0.000835110482110 -0.021605902558135 2.367244177097372 0.000000000000008 -0.000000001360679 0.000263824260089 10 -0.000897769786661 -0.021741736990287 2.484964323969154 0.000000000000002 -0.000000000536868 0.000174175974127
REFERENCES
1. Bickel, P. J. & Doksum, K. A. (2001). Mathematical Statistics, Basic Ideas and Selected Topics. 2nd edition. Volume1. New Jersey: Prentice-Hall.
2. Elsayed, A. (1996). Reliability Engineering. MA: Addison-Wesley.
3. Hall, P. (1992). The Bootstrap and Edgeworth Expansion. New York: Springer-Verlag.
4. Ross, S. M. (1996). Stochastic Processes. 2nd edition. New York: Wiley. 5. Ross, S. M. (1997). Simulation. 2nd edition. New York: Academic Press.
6. Serfling, R. J. (1980). Approximation Theorems of Mathematical Statistics. New York: J. Wiley & sons.
7. Stuart, A. & Ord, J. K. (1987). Kendall’s Advanced Theory of Statistics. 5th ed., Volume 1, Distribution Theory. London: Charles Griffin & Co.
8. Trivedi, K. S. (2002). Probability and Statistics with Reliability, Queuing and Computer Science Applications. 2nd edition. New York: Wiley.