数学杂志  2014, Vol. 34 Issue (1): 58-64   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
HUA Zhi-qiang
YANG Shao-hua
CHEN Li-ying
ASYMPTOTIC LOWER BOUNDS OF LARGE DEVIATION FOR SUMS OF UPPER TAIL ASYMPTOTIC INDEPENDENT RANDOM VARIABLES
HUA Zhi-qiang1, YANG Shao-hua2, CHEN Li-ying1    
1. School of Mathematics and Computational Science, Fuyang Normal College, Fuyang 236037, China;
2. College of Mathematics, Inner Monglia Uiversity for the Nationalities, Tongliao 028043, China
Abstract: This paper investigates the asymptotic lower bounds of large deviation for random variable sums of the upper tail asymptotic independent random variables with long tailed in a multirisk model. By using the classic method of large deviation, we obtain some expressions of random and nonrandom sums, which extend the corresponding independent and identically distributed results.
Key words: large deviation     upper tail asymptotic independence     multi-risk model    
上尾渐近独立随机变量和的大偏差的渐近下界
华志强1, 杨少华2, 陈丽莹1    
1. 阜阳师范学院数学与计算科学学院, 安徽 阜阳 236037;
2. 内蒙古民族大学数学学院, 内蒙古 通辽 028043
摘要:本文研究了多元风险模型中服从长尾分布的带上尾渐近独立的随机变量和的大偏差渐近下界.利用大偏差的经典求法, 得到了随机变量的非随机和和随机和的大偏差表达式, 推广了独立同分布情形下的相关结论.
关键词大偏差    上尾渐近独立性    多元风险模型    
1 Introduction

Motivated by the recent work of [1], we investigate the tail probability of the non-random sums in a multi-risk model

$\begin{equation} \sum\limits^{k}_{i=1}\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}. \end{equation}$ (1.1)

Here $\{X_{ij},j\geq 1\}^{k}_{i=1}$ are the sequences of upper tail asymptotic independent random variables with long tailed, and $\{C_{ij},j\geq 1\}^{k}_{i=1}$ be another nonnegative real sequences. The corresponding random sums of (1.1) is

$\begin{equation} \sum\limits^{k}_{i=1}\sum\limits^{N_{i}(t)}_{j=1}C_{ij}X_{ij}, \end{equation}$ (1.2)

where $\{N_{i}(t)\}$ be non-negative integer-value process with $\lambda_{i}(t)=N_{i}(t)$, while $\{N_{i}(t), i=1,2,\cdots, k\}$ and $\{X_{ij},j\geq 1\}^{k}_{i=1}$ are mutually independent.

Since asymptotic behavior of precise large deviations for non-random sums and random sums of random variables has important theoretical significance and extensive applications, it attracts much attention and there appears to be a lot of research literature. For recent works of this aspect, we refer the reader to [1-13]. Among these papers, [11] studied the asymptotic lower bounds of precise large deviations for sums of nonnegative and independent and identically distributed random variable sequence $\{X_{j},j\geq 1\}$. [9] extended the results of [11] to nonnegative and negatively dependent random variables. Also, [1] extended those of [11] to a multi-risk model. We will extend and improve their results to the upper tail asymptotic independent structure.

At the end of this section, we introduce some corresponding notations and concepts of this paper required. Denoted by $F(x)=P(X\leq x)$, $\overline{F}(x)=1-F(x)$, and $\lfloor x \rfloor $ is the integer part of $x$. We use the following notations for two positive functions $a_{1}(x)$ and $a_{2}(x)$

$\begin{equation} a_{1}(x)\gtrsim a_{2}(x) {\rm if} \liminf\limits_{x\rightarrow\infty}\frac{a_{1}(x)}{a_{2}(x)}\geq 1; a_{1}(x)\sim a_{2}(x) {\rm if} \lim\limits_{x\rightarrow\infty}\frac{a_{1}(x)}{a_{2}(x)}=1. \end{equation}$

Definition 1.1 [11] we say that a distribution $F$ on $(-\infty, +\infty)$ belongs to the long-tailed distribution class, denoted by$F\in \mathcal{L}$, if for any $y\in(-\infty,+\infty)$,

$\begin{equation} \overline{F}(x+y)\sim \overline{F}(x). \end{equation}$

Remark It is known that the long-tailed distribution class is one of the most important heavy-tailed distribution classes, where we say $X$ (or its distribution $F$) is heavy tailed if it has no exponential moments.Also, one can see that, a distribution $F\in \mathcal{L}$ if and only if there exists a function $h(\cdot):[0,\infty)\mapsto [0,\infty)$ such that $h(x)\rightarrow \infty$, $\lim\limits_{x\rightarrow\infty}\frac{h(x)}{x}=0$ and

$\begin{equation} \overline{F}(x+y)\sim \overline{F}(x) \end{equation}$ (1.3)

holds uniformly for all $|y|\leq h(x)$.

Definition 1.2 [12] we say that random variable sequence $\{X_{j},j\geq 1\}$ is upper tail asymptotic independent (UTAI), if all natural numbers $i\neq j$,

$ \lim\limits_{\min\{x_{i},x_{j}\}\rightarrow\infty}P\left(X_{i}>x_{i}|X_{j}>x_{j}\right)=0. $

Remark For research on this structure, we refer the reader to [12], which presented some examples to illustrate that this structure is wider than the other dependent structures.

2 Asymptotic Lower Bounds for Nonrandom Sums

Theorem 2.1 For $i=1,2,\cdots ,k$, let $\{X_{ij},j\geq 1\}$ be a sequence of UTAI and nonnegative random variables with common distribution $F_{i}\in \mathcal{L}$, and let $\{n_{i}\}$ be a positive integer sequence. We assume that $\{X_{ij},j\geq 1\}_{i=1}^{k}$ are mutually independent. Then, for any fixed $0< C_{ij}<\infty$, $i=1,2,\cdots ,k$, $j\geq 1$,

$\begin{equation} P(\sum\limits^{k}_{i=1}\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>x)\gtrsim \sum\limits^{k}_{i=1}\sum\limits^{n_{i}}_{j=1}P(C_{ij}X_{ij}>x)\\ \end{equation}$ (2.1)

holds $\mbox{as}\ x\rightarrow\infty$.

Proof We use induction to prove $(2.1)$.

(I) When $k=1$, for an $h(x)$ satisfying $(1.3)$, we have

$\begin{eqnarray} &&P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>x)\geq P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>x, \bigcup^{n_{1}}_{j=1}(C_{1l}X_{1l}>x+h(x)))\nonumber\\ &\geq& \sum\limits^{n_{1}}_{l=1}P(C_{1l}X_{1l}>x+h(x))-\sum\limits_{1\leq l< j\leq n_{1}}P(C_{1j}X_{1j}>x+h(x), C_{1l}X_{1l}>x+h(x))\nonumber\\ &=&\sum\limits^{n_{1}}_{l=1}P(C_{1l}X_{1l}>x+h(x))\nonumber\\ &&-\sum\limits_{1\leq l< j\leq n_{1}}P(C_{1j}X_{1j}>x+h(x)|C_{1l}X_{1l}>x+h(x))P(C_{1l}X_{1l}>x+h(x))\nonumber\\ &=&K_{1}-K_{2}. \end{eqnarray}$ (2.2)

By $F_{1}\in \mathcal{L}$, we find

$\begin{eqnarray} \liminf\limits_{x\rightarrow\infty}\frac{K_{1}}{P(\sum\limits^{n_{1}}\limits_{j=1}C_{1j}X_{1j}>x)}= \liminf\limits_{x\rightarrow\infty}\frac{P(\sum\limits^{n_{1}}\limits_{j=1}X_{1j}>\frac{x}{C_{1j}}+ \frac{h(x)}{C_{1j}})}{P(\sum\limits^{n_{1}}\limits_{j=1}C_{1j}X_{1j}>x)}\geq 1. \end{eqnarray}$ (2.3)

For $K_{2}$, along with UTAI property, we have that

$\begin{eqnarray} &&\limsup\limits_{x\rightarrow\infty}\frac{K_{2}}{P\left(\sum\limits^{n_{1}}\limits_{j=1}C_{1j}X_{1j}>x\right)}\nonumber\\ &&= \limsup\limits_{x\rightarrow\infty}\left[\sum\limits_{1\leq l< j\leq n_{1}}P(C_{1j}X_{1j}>x+h(x)|C_{1l}X_{1l}>x+h(x)) \frac{P(C_{1l}X_{1l}>x+h(x))}{P(\sum\limits^{n_{1}}\limits_{j=1}C_{1j}X_{1j}>x)}\right]\nonumber\\ &&=0. \end{eqnarray}$ (2.4)

Hence, combining (2.2)-(2.4) leads to (2.1) when $k=1$, that is

$\begin{eqnarray} P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>x)\gtrsim \sum\limits^{n_{1}}_{j=1}P(C_{1j}X_{1j}>x). \end{eqnarray}$ (2.5)

(II) For the case in which $k=2$, there is

$\begin{eqnarray*} &&P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}+\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>x)\nonumber\\ &\geq &P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>x+h(x),\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>-h(x))\nonumber \\ &&+P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>-h(x),\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>x+h(x))\nonumber\end{eqnarray*}$
$\begin{eqnarray*} &=&P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>x+h(x))P(\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>-h(x))\nonumber\\ &&+P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}>-h(x))P(\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>x+h(x)).\nonumber\\ \end{eqnarray*}$

For any $0<\delta <1$, by (2.5) and $F_{i}\in \mathcal{L}$, $i=1,2$, for sufficiently large $x$, we get

$\begin{eqnarray} P(\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>x+h(x))\geq (1-\delta)[\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>x], i=1,2. \end{eqnarray}$ (2.6)

Since for $i=1,2,$ $\{X_{ij},j\geq 1\}$ are nonnegative, for any fixed $0< C_{ij}<\infty$, $j=1,2,\cdots, n_{i} $, we have

$\begin{eqnarray} P(\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>-h(x))>(1-\delta), i=1,2. \end{eqnarray}$ (2.7)

Then, using (2.6)-(2.7), we obtain

$\begin{eqnarray*} P(\sum\limits^{n_{1}}_{j=1}C_{1j}X_{1j}+\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>x) \geq (1-\delta)^{2}P[\sum\limits^{n_{1}}_{j=1}P(C_{1j}X_{1j}>x)+P(\sum\limits^{n_{2}}_{j=1}C_{2j}X_{2j}>x)]. \end{eqnarray*}$

Therefore, letting $\delta \downarrow 0$, we obtain (2.1) when the case of $k=2$.

(III) Now suppose that (2.1) holds for $k-1$. As for $k$, using a similar argument to that in (II), for any $0<\delta <1$ and any fixed $C_{ij}$, $i=1,2,\cdots , k$, $j\geq 1$, and when $x$ is sufficiently large, we have

$\begin{eqnarray*} &&P(\sum\limits^{k}_{i=1}\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>x)\nonumber\\ &\geq& P(\sum\limits^{k-1}_{i=1}\sum\limits^{n_{i}}_{j=1}C_{ij}X_{ij}>x+h(x))+P(\sum\limits^{n_{k}}_{j=1}C_{kj}X_{kj}>-h(x))\nonumber\\ &\geq &(1-\delta)^{2}[\sum\limits^{k-1}_{i=1}\sum\limits^{n_{i}}_{j=1}P(C_{ij}X_{ij}>x+h(x))]+(1-\delta)^{2}\sum\limits^{n_{k}}_{j=1}P(C_{kj}X_{kj}>-h(x))\nonumber\\ &=&(1-\delta)^{2}\sum\limits^{k}_{i=1}\sum\limits^{n_{i}}_{j=1}P(C_{ij}X_{ij}>x+h(x)).\nonumber \end{eqnarray*}$

Letting $\delta \downarrow 0$, we obtain the desired result, and the proof Theorem 2.1 is now complete.

3 Asymptotic Lower Bounds for Random Sums

Theorem 3.1 For $i=1,2,\cdots ,k$, let$\{X_{ij},j\geq 1\}$ be a sequence of UTAI and nonnegative random variables with common distribution $F_{i}\in \mathcal{L}$, and let $\{N_{i}(t)\}$ be non-negative integer-value process with $\lambda_{i}(t)=N_{i}(t)$. We assume that $\{X_{ij},j\geq 1\}_{i=1}^{k}$ and $\{N_{i}(t), i=1,2,\cdots ,k\}$ are mutually independent and that $\{N_{i}(t), i=1,2,\cdots ,k\}$ satisfies

$ {\rm Assumption I}: \frac{N_{i}(t)}{\lambda_{i}(t)}\rightarrow_{p}1 {\rm as}t\rightarrow \infty. $

Then, for any fixed $0< C_{ij}<\infty$, $i=1,2,\cdots ,k$, $j\geq 1$,

$\begin{equation} P(\sum\limits^{k}_{i=1}\sum\limits^{N_{i}(t)}_{j=1}C_{ij}X_{ij}>x)\gtrsim \sum\limits^{k}_{i=1}\sum\limits^{N_{i}(t)}_{j=1}P(C_{ij}X_{ij}>x) \end{equation}$ (3.1)

holds $\mbox{as}\ x\rightarrow\infty$.

Proof Again by induction, as in the proof of Theorem 3.1, it is sufficient to show that (3.1) holds for $k=1,2$.

(I) Taking $k=1$, for any $0<\delta <1$ and any fixed $C_{1j}$, $j\geq 1$, and for sufficiently large $t$,

$\begin{eqnarray*} &&P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>x)=\sum\limits^{\infty}_{n=1}P(\sum\limits^{n}_{j=1}C_{1j}X_{1j}>x)P(N_{1}(t)=n) \\ &\geq& \sum\limits_{\lfloor (1-\delta)\lambda_{1}(t)\rfloor \leq k\leq \lfloor (1+\delta)\lambda_{1}(t)\rfloor}P(\sum\limits^{n}_{j=1}C_{1j}X_{1j}>x)P(N_{1}(t)=n)\\ &\geq& P(\sum\limits^{\lfloor(1-\delta)\lambda_{1}(t)\rfloor}_{j=1}C_{1j}X_{1j}>x)P(|\frac{N_{1}(t)}{\lambda_{1}(t)}-1|<\delta )\\ &\geq& (1-\delta)^{2}\sum\limits^{\lfloor(1-\delta)\lambda_{1}(t)\rfloor}_{j=1}P(C_{1j}X_{1j}>x), \end{eqnarray*}$

where the last inequality holds due to Assumption I and (2.5). Letting $\delta \downarrow 0$, Theorem 3.1 holds for $k=1$, that is

$\begin{equation} P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>x)\gtrsim \sum\limits^{N_{1}(t)}_{j=1}P(C_{1j}X_{1j}>x). \end{equation}$ (3.2)

(II) When $k=2$, we get

$\begin{eqnarray*} &&P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}+\sum\limits^{N_{2}(t)}_{j=1}C_{2j}X_{2j}>x)\nonumber\\ &\geq& P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>x+h(x),\sum\limits^{N_{2}(t)}_{j=1}C_{2j}X_{2j}>-h(x))\nonumber\\ &&+ P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>-h(x),\sum\limits^{N_{2}(t)}_{j=1}C_{2j}X_{2j}>x+h(x))\nonumber\end{eqnarray*}$
$\begin{eqnarray*} &=&P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>x+h(x))P(\sum\limits^{N_{2}(t)}_{j=1}C_{2j}X_{2j}>-h(x))\nonumber\\ &&+P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}>-h(x))P(\sum\limits^{N_{2}(t)}_{j=1}C_{2j}X_{2j}>x+h(x)). \end{eqnarray*}$

By (3.2) and $F_{i}\in \mathcal{L}(i=1,2)$, for any $0<\delta <1$, for sufficiently large $t$, we arrive

$\begin{eqnarray} P(\sum\limits^{N_{i}(t)}_{j=1}C_{ij}X_{ij}>x+h(x))\geq (1-\delta)[P(\sum\limits^{N_{i}(t)}_{j=1}C_{ij}X_{ij}>x)], i=1,2. \end{eqnarray}$ (3.3)

Since for $i=1,2,$ $\{X_{ij},j\geq 1\}$ are nonnegative, and for any fixed $0< C_{ij}<\infty$, $j\geq1$, we have

$\begin{eqnarray} P(\sum\limits^{N_{i}(t)}_{j=1}C_{ij}X_{ij}>-h(x))>(1-\delta), i=1,2. \end{eqnarray}$ (3.4)

Then, using (3.3)-(3.4), we obtain

$\begin{eqnarray*} &&P(\sum\limits^{N_{1}(t)}_{j=1}C_{1j}X_{1j}+\sum\limits^{N_{2}(t)}_{j=1}C_{1j}X_{1j}>x)\nonumber\\ &\geq& (1-\delta)^{2}P[\sum\limits^{N_{1}(t)}_{j=1}P(C_{1j}X_{1j}>x)+P(\sum\limits^{N_{2}(t)}_{j=1}C_{1j}X_{1j}>x)]. \end{eqnarray*}$

Therefore, letting $\delta \downarrow 0$, we obtain (3.1) for $k=2$. The proof Theorem 3.1 is completed.

References
[1] Lu Dawei. Lower bounds of large deviation for sums of long-tailed claims in a multi-risk model[J]. Statist. Prob. Lett, 2012, 82: 1242–1250. DOI:10.1016/j.spl.2012.03.020
[2] Klüppelberg C, Mikosch T. Large deviations of heavy-tailed random sums with applications in insurance and finance[J]. J. Appl. Prob, 1997, 34: 293–308. DOI:10.1017/S0021900200100956
[3] Li Kewen, Hu Yijun. An estimate of large deviations for heavy-tailed random sums of independent random variables[J]. J. of Math. (PRC), 2002, 22(2): 132–139.
[4] Liu Yan, Hu Yijun. Large deviations for a heavy-tailed stationary sequence[J]. J. of Math. (PRC), 2003, 23(1): 11–18.
[5] Kai W Ng, Tang Qihe, Yan Jiaan, Yang Hailiang. Precise large deviations for sums of random variables with consistently varying tails[J]. J. Appl. Probab., 2004, 41: 93–107. DOI:10.1017/S0021900200014066
[6] Tang Qihe. Insensitivity to negative dependence of the asymptotic behavior of precise large deviations[J]. Electron. J. Probab., 2006, 11(4): 107–120.
[7] Wang Shijie, Wang Wensheng. Precise large deviations for sums of random variables with consistently varying tails in multi-risk models[J]. J. Appl. Probab., 2007, 44: 889–900. DOI:10.1017/S0021900200003612
[8] Liu Li. Precise large deviations for dependent random variables with heavy tails[J]. Statist. Prob. Lett., 2009, 79: 1290–1298. DOI:10.1016/j.spl.2009.02.001
[9] Konstantinides D G, Loukissas F. Precise large deviations for sums of negatively dependent random variables with common long-tailed distributions[J]. Comm. Statist. Theory Methods, 2011, 40: 3663–3671. DOI:10.1080/03610926.2011.581186
[10] Wang Yuebao, Cheng Dongya. Basic renewal theorems for random walks with widely dependent increments[J]. J. Math. Anal. Appl., 2011, 384: 597–606. DOI:10.1016/j.jmaa.2011.06.010
[11] Loukissas F. Precise large deviations for long-tailed distributions[J]. J. Theoret. Probab., 2012, 25(4): 913–924. DOI:10.1007/s10959-011-0367-2
[12] Liu Xijun, Gao Qingwu, Wang Yuebao. A note on a dependent risk model with constant interest rate[J]. Statist. Prob. Lett., 2012, 82: 707–712. DOI:10.1016/j.spl.2011.12.016
[13] Ch en, Yiqing, Yuen Kam C. Precise large deviations of aggregate claims in a size-dependent renewal risk model[J]. Insurance: Mathematics and Economics, 2012, 51(2): 457–461. DOI:10.1016/j.insmatheco.2012.06.010