Message Board

Dear readers, authors and reviewers,you can add a message on this page. We will reply to you as soon as possible!

2019 Volume 41 Issue 5
Article Contents

Shuang HU, Zuo-xiang PENG. A Pickands-Type Estimator Based on the Block Method[J]. Journal of Southwest University Natural Science Edition, 2019, 41(5): 53-58. doi: 10.13718/j.cnki.xdzk.2019.05.009
Citation: Shuang HU, Zuo-xiang PENG. A Pickands-Type Estimator Based on the Block Method[J]. Journal of Southwest University Natural Science Edition, 2019, 41(5): 53-58. doi: 10.13718/j.cnki.xdzk.2019.05.009

A Pickands-Type Estimator Based on the Block Method

More Information
  • Corresponding author: Zuo-xiang PENG
  • Received Date: 23/05/2018
    Available Online: 20/05/2019
  • MSC: O211

  • Based on the block method proposed by Davydov et al., we propose in this paper a new Pickands-type estimator. The asymptotic properties of the new estimator, such as its consistency and asymptotic normality, are derived under some regular conditions.
  • 加载中
  • [1] PICKANDS J. Statistical Inference Using Extreme Order Statistics[J]. Annals of Statistics, 1975, 3(1):119-131.

    Google Scholar

    [2] DEKKERS A L M, HAAN L D. On the Estimation of the Extreme-Value Index and Large Quantile Estimation[J]. Annals of Statistics, 1989, 17(4):1795-1832. doi: 10.1214/aos/1176347396

    CrossRef Google Scholar

    [3] 彭作祥. Pickands型估计的推广[J].数学学报, 1997(5):759-762.

    Google Scholar

    [4] PENG Z X, NADARAJAH S. The Pickands' Estimator of the Negative Extreme-Value Index[J]. Acta Scicentiarum Naturalum Universitis Pekinesis, 2001, 37(1):12-19.

    Google Scholar

    [5] DAVYDOV Y, PAULAUSKAS V, RAČKAUSKAS A. More on P-Stable Convex Sets in Banach Spaces[J]. Journal of Theoretical Probability, 2000, 13(1):39-64.

    Google Scholar

    [6] PAULAUSKAS V. A New Estimator for a Tail Index[J]. Acta Applicandae Mathematica, 2003, 79(1-2):55-67.

    Google Scholar

    [7] QI Y C. On the Tail Index of a Heavy Tailed Distribution[J]. Annals of the Institute of Statistical Mathematics, 2010, 62(2):277-298.

    Google Scholar

    [8] VAIČIULIS M. Local-Maximum-Based Tail Index Estimator[J]. Lithuanian Mathematical Journal, 2014, 54(4):503-526. doi: 10.1007/s10986-014-9260-x

    CrossRef Google Scholar

    [9] PAULAUSKAS V, VAIČIULIS M. Comparison of the Several Parameterized Estimators for the Positive Extreme Value Index[J]. Journal of Statistical Computation & Simulation, 2016, 87(7):1342-1362.

    Google Scholar

    [10] 马跃, 彭作祥.广义误差-帕累托分布及其在保险中的应用[J].西南大学学报(自然科学版), 2017, 39(1):99-102.

    Google Scholar

    [11] HAAN L D. Slow Variation and Characterization of Domains of Attraction[J]. Statistical Extremes and Applications, 1984:31-48.

    Google Scholar

    [12] MICHAILIDIS G, STOEV S. Extreme Value Theory:An Introduction[J]. Technometrics, 2007, 49(4):491-492.

    Google Scholar

    [13] RESNICK S I. A Probability Path[M]. Boston:Birkhäuser, 1999.

    Google Scholar

  • 加载中
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Article Metrics

Article views(1066) PDF downloads(143) Cited by(0)

Access History

Other Articles By Authors

A Pickands-Type Estimator Based on the Block Method

    Corresponding author: Zuo-xiang PENG

Abstract: Based on the block method proposed by Davydov et al., we propose in this paper a new Pickands-type estimator. The asymptotic properties of the new estimator, such as its consistency and asymptotic normality, are derived under some regular conditions.

  • 设{Xnn≥1}为独立同分布随机变量序列,分布函数为F(x),X1,nX2,n≤…≤Xnn为其次序统计量.若存在常数an>0,bn∈$\mathbb{R}$使得对于非退化分布函数Gγ(x),有

    其中:${G_\gamma }\left(x \right) = {{\rm{e}}^{ - {{\left({1 + \gamma x} \right)}^{ - \frac{1}{\gamma }}}}}$,1+γx>0,γ∈$\mathbb{R}$(γ=0时,Gγ(x)=e-e-x).此时称F属于极值吸引场Gγ,即FD(Gγ),γ被称为极值指数.

    当分布F未知时,对极值指数γ∈$\mathbb{R}$,文献[1]提出如下的Pickands型估计量:

    文献[2]讨论了其相合性和渐近正态性,在此基础上文献[3-4]对Pickands型估计量进行了推广.

    Pickands估计量使用了较少的样本信息,计算简便,且具有位置和尺度不变性.文献[5]提出块方法,将样本分为若干块,利用每块中最大的和次大的样本的比率构造估计量.文献[6]从理论和模拟两方面说明了该估计量的良好性质.关于极值指数估计量及其应用的更多研究,见文献[7-10].

    本文将使用块方法构造新的Pickands型估计量.将样本X1X2,…,Xn分成knV1V2,…,Vkn,使得每一块包含$m = m(n) = \left[{\frac{n}{{{k_n}}}} \right]$个样本([x]表示取整数部分),即Vj=X(j-1)m+1,…,Xjm,1≤jkn.令X1,m(j)X2,m(j)≤…≤Xmm(j)表示第jm个样本的次序统计量,定义块Pickands型估计量为:

    其中ms=s(n)∈$\mathbb{N}$+且满足n→∞时,

    在上述条件下,主要讨论kn(n)→∞和kn(n)≡k(常数)两种情况下$\hat \gamma _n^\mathit{Q}$的相合性和渐近正态性.

1.   相合性与渐近正态性
  • 令$U = {\left({\frac{1}{{1 - F}}} \right)^←}$为$\frac{1}{{1 - F}}$的广义逆. (1)式成立当且仅当存在辅助函数a(t)>0使得

    x>0局部一致成立,其中$R(t, x): = \frac{{U(tx) - U(t)}}{{a(t)}}$,${D_\gamma }(x) = \frac{{{x^\gamma } - 1}}{\gamma }$(γ=0时,Dγ(x)=log x).由文献[11]易知

    对于xy>0,y≠1局部一致成立.

    {Enn≥1}是独立同分布序列,均服从标准指数分布. {Ynn≥1}是分布函数$F(x) = 1 - \frac{1}{x}(x1)$的独立同分布序列.E1,m(j)E2,m(j)≤…≤Emm(j)Y1,m(j)Y2,m(j)≤…≤Ymm(j)分别是两组序列第j组样本的次序统计量.易知对j=1,2,…,kn

    对于kn→∞和knk两种情况,$\hat \gamma _n^\mathit{Q}$具有相同的相合性性质:

    定理1(弱收敛性)  若(1)和(3)式成立,则$\hat \gamma _n^Q\mathop \to \limits^P \gamma (n \to \infty)$.

    定理2(强收敛性)  若(1)和(3)式成立,且$\frac{s}{{\log \log m}} \to \infty $S,则$\hat \gamma _n^Q\mathop \to \limits^{{\rm{a}}.{\rm{s}}.} \gamma (n \to \infty)$.

    为进一步探究$\hat \gamma _n^Q$的渐近分布,假设存在辅助函数A(t)→0(t→∞)且无限远处恒正或恒负,使得当t→∞时,

    对于x>0局部一致成立,其中

    U是二阶正规变换函数.由文献[12]推论2.3.6和注记B.3.8可知,对于任意εδ>0,存在t0=t0(εδ)使得当txt0时,

    定理3  假设(3)和(7)式成立.若n→∞时,kn→∞,$\frac{{{k_n}}}{{s(n)}} \to 0$,$\sqrt {{k_n}s} A\left({\frac{n}{{2s{k_n}}}} \right) \to {\lambda _1} \in {\mathbb{R}}$,则$\sqrt {{k_n}s} \left({\hat \gamma _n^\varrho - \gamma } \right)\mathop \to \limits^d N\left({\mu {\lambda _1}, {\sigma ^2}} \right)$,其中

    定理4  假设(3)与(7)式成立.若knk<∞且$\sqrt s A\left({\frac{n}{{2ks}}} \right) \to {\lambda _2} \in {\mathbb{R}}$,则$\sqrt s \left({\hat \gamma _n^Q - \gamma } \right)\mathop { \to N\left({\mu {\lambda _2}, \frac{{{\sigma ^2}}}{k}} \right)}\limits$.其中μσ2同定理3.

    注1  若kn=1,此时$\hat \gamma _n^Q$就是Pickands估计量(2)式,进一步在定理4中令λ2=0,可得到与文献[2]相同的渐近性质.

2.   定理的证明
  • 定理1的证明  设(1)和(3)式成立,对于j=1,2,…,kn,$\frac{{Y_{m - s + 1, m}^{(j)}}}{{Y_{m - 2s + 1, m}^{(j)}}}\frac{d}{ = }{{\rm{e}}^{\left({E_{m - s + 1, m}^{(j)} - E_{m - 2s + 1, m}^{(j)}} \right.}}\mathop = \limits^d Y_{s, 2s - 1}^{(j)}\mathop \to \limits^P 2$成立(见文献[2]之推论2.1)且E|Ys,2s-1(j)|<∞,由(5)和(6)式知,n→∞时

    定理证毕.

    定理2的证明  若s满足定理2的条件,则

    相似于定理1的证明,定理2得证.

    定理3的证明  若(3)和(7)式成立,定义

    n→∞时,由文献[2]的定理2.1知,

    由于(4)式局部一致成立,利用泰勒展式、(6)式和Smirnov引理[12],有

    根据γρ的不同,应分类对(9)式各部分进行讨论,在此只证明γ+ρ≠0,ρ<0且γ≠0的情况,其他情况类似可证.由文献[2]之推论2.1,将$ {{\text{e}}^{\gamma \left( {E_{m - s + 1,m}^{(j)} - E_{m - 2s + 1,m}^{(j)}} \right)}}$在log 2处泰勒展开有

    其中Rn(j)=Em-s+1,m(j)-Em-2s+1,m(j)-log 2,Qn(j)=Em-2s+1,m(j)-Em-4s+1,m(j)-log 2,j=1,2,…,kn.

    由文献[2]定理2.3的证明易知,(10)式中Rn(j)Qn(j)(j=1,2,…,kn)相互独立,且由

    记${R_n}: = \sqrt {\frac{s}{{{k_n}}}} \sum\limits_{j = 1}^{{k_n}} {\sum\limits_{l = s}^{2, - 1} {\frac{{E_l^{(j)} - 1}}{l}} }$的特征函数为

    由泰勒展式知,

    同理,${Q_n}: = \sqrt {\frac{s}{{{k_n}}}} \sum\limits_{j = 1}^{{k_n}} {\sum\limits_{l = {2_s}}^{4s - 1} {\frac{{E_l^{(j)} - 1}}{l}} } $的特征函数fnQn(t)具有如下性质:

    由于Rn(j)Qn(j)(j=1,2,…,kn)相互独立,2γRn-Qn的特征函数fn(t)具有如下性质:

    注意到

    所以n→∞时,

    由于$\sqrt {{k_n}s} \left( {\sum\limits_{l = s}^{2s - 1} {\frac{1}{l}} - \log 2} \right) \to 0$,由Slutsky定理[13]知,

    对于ρ<0且γ+ρ≠0,因为n→∞时$Y_{s, 2s - 1}^{(j)}\mathop \to \limits^P 2\left({j = 1, 2, \cdots, {k_n}} \right)$,有

    由(8)式可知,对于足够大的tε(tx)有界且局部一致收敛到0,所以

    设$\sqrt {{k_n}s} A\left({\frac{m}{{2s}}} \right) \to {\lambda _1}$,根据(9),(11)-(13)式可得当n→∞时

    定理证毕.

    定理4的证明类似定理3,此处省略.

Reference (13)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return