1 Introduction
The goal of this paper is studying the precise large deviations for the aggregate claims where \(N_{t}\) is the number of claims in period t and \(\{X_{t,j}, j=1,2,\ldots, t=1,2,\ldots,n\}\) form an array of independent identically distributed (i.i.d.) claim-size random variables independent of \(N_{t}\) with distribution \(F_{X}=1-\overline{F}_{X}\). Here the claim-number process \(\{N_{t},t=1,2,\ldots\}\) is described by the Poisson first-order Autoregressive Conditional Heteroscedasticity (\(\operatorname{ARCH}(1)\)) process, which is defined as where \(a_{0} > 0\), \(0\leq a_{1}<1\), \(N_{0}\geq0\) is a deterministic integer, and \(\mathscr{F}_{t-1}=\sigma(N_{0}, N_{1}, \ldots, N_{t-1})\) is the σ-field generated by \(\{N_{0},N_{1}, \ldots, N_{t-1}\}\).
$$ S_{n}=\sum_{t=1}^{n} \sum_{j=1}^{N_{t}}X_{t,j} , $$
(1.1)
$$ \left \{ \textstyle\begin{array}{l} N_{t}|\mathscr{F}_{t-1}\sim \operatorname{Poisson}(\lambda_{t}), \\ \lambda_{t}=a_{0}+a_{1}N_{t-1}, \end{array}\displaystyle \right . $$
(1.2)
In the last few years, research on the time series models for count data has become a popular topic in the literature. Cossette et al. [1] used two integer-valued time series, namely the Poisson moving average (MA) and Poisson autoregressive (AR) processes, to model the claim frequency in the risk model. Li [2] proposed a discrete-time risk model with the claim number being an integer-valued ARCH (INARCH) process with Poisson deviates, namely the model (1.1) and derived some statistical properties and adjustment coefficient for the risk model.
Advertisement
The Poisson INARCH process was first considered by Rydberg and Shephard [3] and applied to finance to model the number of transactions taking place during a short time interval. In model (1.2), it is assumed that the conditional mean of the current claim number has a linear relationship with the previous values of observations. Streett [4] and Ferland et al. [5] point out that the process \(N_{t}\) in (1.2) is stationary if \(0\leq a_{1} < 1\). In particular, the expectation and variance of \(N_{t}\) are
$$\mathbb{E}N_{t}=\frac{a_{0}}{1-a_{1}} \quad \mbox{and} \quad \operatorname{Var} (N_{t})=\frac {\mathbb{E} N_{t}}{1-a_{1}^{2}}. $$
We study the precise large deviations for \(\{S_{n} , n \geq0\}\) in (1.1). We are only interested in the case of heavy-tailed claims. Heavy-tailed distributions belong to the core issues in actuarial science, because they are more in accordance with claims’ reality than light-tailed ones. A useful heavy-tailed class is the class \(\mathscr{C}\) of distribution functions with consistent variation (also called intermediate regular variation), characterized by the relations \(\overline{F}(x)>0\) for all \(x\geq0\) and Our main result is given below.
$$\lim_{y\uparrow1}\limsup_{x\rightarrow\infty}\frac {\overline{F}(xy)}{\overline{F}(x)}= \lim_{y\downarrow1}\liminf_{x\rightarrow\infty }\frac{\overline{F}(xy)}{\overline{F}(x)}=1. $$
Theorem 1.1
Consider the aggregate amount of claims (1.1), assume that
\(F_{X}(x)\in\mathscr{C}\), \(\mathbb{E}X=\mu\in(0,\infty)\), and
\(N_{0}\geq0\)
a deterministic integer. Then, for every fixed
\(\gamma>0\), uniformly for all
\(x\geq\gamma n\),
$$ \mathbb{P}\biggl(S_{n}-\frac{a_{0}}{1-a_{1}}n\mu >x\biggr) \sim\frac{na_{0}}{1-a_{1}} \overline{F}_{X}(x),\quad n\rightarrow\infty. $$
(1.3)
A precise large deviation is an important study task in applied probability, and it is usually used to quantitatively characterize the property of extremal events. As is well known, there is a vast amount of literature studying the asymptotic behavior of the large deviation of the risk models in the presence of heavy-tailed claim sizes. See, e.g., Klüppelberg and Mikosch [6], Ng et al. [7], Leipus and Šiaulys [8], Asmussen [9], Liu [10], Yang et al. [11], Chen and Yuen [12], among many others. On the other hand, the precise large deviations to the risk model with the claim number being a Poisson ARCH process has not been considered in the literature.
Advertisement
We now comment on the approaches used in this work. Our method is much more elementary and does not use the classical treatment in the area of precise large deviation. For convenience in application, we first show that the accumulated aggregate claim \(S_{n}\) in (1.1) has the same distribution with another random walk. Second, we establish the moderate deviation principle (MDP) for the partial sum \(\sum_{t=1}^{n} N_{t}\) generated by the Poisson \(\operatorname{ARCH}(1)\) process \({N_{t} }\) defined in (1.2). As a consequence of MDP, we are able to claim the genuine exponential decay for the probability that the sample average deviates away from its mathematical equilibrium value. This property is a crucial step of our proof of Theorem 1.1. Finally, we would like to point out that equation (1.3) agrees with existing ones in the literature. This indicates that the dependence structure of \(N_{t}\) defined by (1.2) does not affect the asymptotic behavior of the large deviations of \(\{S_{n}, n\geq0\}\).
2 Preliminaries
Throughout this paper, for two positive functions \(f(x)\) and \(g(x)\), we write For two positive bivariate functions \(f(x,n)\) and \(g(x,n)\), we say that the asymptotic relation \(f(x,n)\lesssim g(x,n)\) holds uniformly for x in a nonempty set \(\Delta_{n}\) if
$$\begin{aligned}& f(x)\sim g(x) \quad \mbox{if }\lim_{x\rightarrow\infty} f(x)/g(x)=1; \\& f(x)\lesssim g(x)\quad \mbox{if }\limsup_{x\rightarrow\infty} f(x)/g(x)\leqslant1\quad \mbox{and} \\& f(x)\gtrsim g(x)\quad \mbox{if } \liminf_{x\rightarrow\infty} f(x)/g(x)\geqslant1. \end{aligned}$$
$$\limsup_{n\rightarrow\infty}\sup_{x\in\Delta_{n}} \frac{f(x;n)}{g(x;n)} \leqslant1. $$
First, we show that the accumulated aggregate claim \(S_{n}\) in (1.1) has the same distribution with another random walk.
Lemma 2.1
Let
\(\{Y, Y_{j},j\geqslant1\}\)
be a sequence of i.i.d. non-negative random variables such that
Y
and
X
in (1.1) are identically distributed, then
where
\(\stackrel{d}{=}\)
denotes the identical distribution.
$$S_{n} \stackrel{d}{=}\sum_{j=1}^{N_{1}+N_{2}+\cdots+N_{n}}Y_{j}, $$
Proof
For any real r, the moment generating function of \(S_{n}\) is expressed as On the other hand, we have where \(m=n_{1}+n_{2}+\cdots+n_{n}\). Hence, by the uniqueness of the moment generating function, we know that \(S_{k}\) and \(\sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}\) have the same distribution. □
$$\begin{aligned} M_{s} =&E\bigl\{ \exp\{rS_{n}\}\bigr\} =E\Biggl\{ \exp\Biggl\{ r\sum_{i=1}^{n}\sum _{j=1}^{N_{i}}X_{i,j}\Biggr\} \Biggr\} \\ =&E\biggl\{ e^{r\sum _{j=1}^{N_{1}}X_{1,j}}e^{r\sum _{j=1}^{N_{2}}X_{2,j}}\cdots e^{r\sum _{j=1}^{N_{n}}X_{k,j}}\cdot\sum _{n_{1},n_{2},\ldots ,n_{n}}I_{(N_{1}=n_{1},N_{2}=n_{2},\ldots, N_{n}=n_{n})}\biggr\} \\ =&\sum_{n_{1},n_{2},\ldots ,n_{n}}E\bigl\{ e^{r\sum _{j=1}^{n_{1}}X_{1,j}}e^{r\sum _{j=1}^{n_{2}}X_{2,j}} \cdots e^{r\sum _{j=1}^{n_{n}}X_{n,j}}\cdot I_{(N_{1}=n_{1},N_{2}=n_{2},\ldots, N_{n}=n_{n})}\bigr\} \\ =&\sum_{n_{1},n_{2},\ldots ,n_{n}}\bigl(Ee^{rX} \bigr)^{\sum _{i=1}^{n}n_{i}}\cdot P\{ N_{1}=n_{1},N_{2}=n_{2}, \ldots, N_{n}=n_{n}\} \\ =&E\bigl\{ (M_{X})^{N_{1}+N_{2}+\cdots+ N_{n}}\bigr\} . \end{aligned}$$
$$\begin{aligned}& E\Biggl\{ \exp\Biggl\{ r\sum_{j=1}^{N_{1}+N_{2}+\cdots+N_{n}}Y_{j} \Biggr\} \Biggr\} \\& \quad = E\Biggl\{ \exp\Biggl\{ r\sum_{j=1}^{N_{1}+N_{2}+\cdots+N_{n}}Y_{j} \Biggr\} \cdot\sum_{m}I_{(N_{1}+N_{2}+\cdots +N_{n}=m)}\Biggr\} \\& \quad = \sum_{m}E\Biggl\{ \exp\Biggl\{ r\sum _{j=1}^{m}Y_{j}\Biggr\} \cdot I_{(N_{1}+N_{2}+\cdots +N_{n}=m)}\Biggr\} \\& \quad =\sum_{m}\bigl(Ee^{rY} \bigr)^{m}\cdot P\{N_{1}+N_{2}+\cdots +N_{n}=m\} =E\bigl\{ (M_{Y})^{N_{1}+N_{2}+\cdots+ N_{n}}\bigr\} , \end{aligned}$$
Next, the following lemma establishes the MDP for \(\{N_{t}, t\geq0\}\).
Lemma 2.2
Assume
\(\{N_{t}, t\geq1\}\)
defined by (1.2), and
\(N_{0}\ge0\)
be a deterministic integer. Let
\(b_{n}\)
be a sequence of positive numbers satisfying
\(b_{n}\rightarrow\infty\)
and
\(b_{n}/n\rightarrow0\), we have
for each closed set
\(H\subset \mathbb{R}\); and
for each open set
\(G\in\mathbb{R}\), where the rate function
\(I_{M}(\cdot)\)
is given as
$$ \limsup_{n\rightarrow\infty}\frac{1}{b_{n}}\log \mathbb{P} \Biggl\{ \frac{1}{\sqrt{nb_{n}}}\sum_{t=1}^{n} \biggl(N_{t}-{a_{0}\over 1-a_{1}} \biggr)\in H \Biggr\} \leqslant- \inf_{x\in H}I_{M}(x) $$
(2.1)
$$ \liminf_{n\rightarrow\infty}\frac{1}{b_{n}}\log \mathbb{P} \Biggl\{ \frac{1}{\sqrt{nb_{n}}}\sum_{t=1}^{n} \biggl(N_{t}-{a_{0}\over 1-a_{1}} \biggr)\in G \Biggr\} \geqslant- \inf_{x\in G}I_{M}(x) $$
(2.2)
$$I_{M}(x)={x^{2}\over 2\sigma^{2}} ,\quad x\in\mathbb{R}, \textit{where } \sigma^{2} =\mathbb{E}\bigl(\operatorname{Var}(N_{t}| \mathscr{F}_{t-1})\bigr)=\frac{a_{0}}{1-a_{1}}. $$
Proof
By the Gärtner-Ellis theorem (Theorem 2.3.6, p.44, Dembo and Zeitouni [13]), all we need to show is that
$$ \lim_{n\to\infty}{1\over b_{n}}\log\mathbb{E} \exp \Biggl\{ \beta \sqrt{b_{n}\over n} \sum_{t=1}^{n}(N_{t}- \mathbb{E}N_{1}) \Biggr\} ={1\over 2}\sigma^{2} \beta^{2},\quad \beta\in\mathbb{R}. $$
(2.3)
Let \(\beta\in\mathbb{R}\) be fixed but arbitrary and write where \(\theta_{n}=\beta\sqrt{\frac{b_{n}}{n}}\). Observe that for any \(t\ge1\),
$$l_{n}=a_{1}\bigl(e^{\theta_{n}}-1\bigr), $$
$$\begin{aligned}& \mathbb{E} \bigl[\exp \bigl\{ \theta_{n} (N_{t}-\mathbb {E}N_{1})-l_{n}N_{t-1} \bigr\} \vert{ \mathcal{F}}_{t-1} \bigr] \\& \quad =\exp \{-l_{n}N_{t-1}-\theta_{n} \mathbb{E}N_{1} \}\mathbb{E} \bigl[\exp\{\theta_{n} N_{t} \} \vert{\mathcal{F}}_{t-1} \bigr] \\& \quad =\exp \{-l_{n}N_{t-1}-\theta_{n}\mathbb{E} N_{1} \}\exp \bigl\{ (a_{0}+a_{1}N_{t-1}) \bigl(e^{\theta_{n}}-1\bigr) \bigr\} \\& \quad =\exp \biggl\{ \frac{a_{0}}{1-a_{1}}\bigl[(1-a_{1}) \bigl(e^{\theta_{n}}-1\bigr)-\theta_{n}\bigr] \biggr\} . \end{aligned}$$
Hence,
$$\begin{aligned}& \mathbb{E}\exp \Biggl\{ \sum_{t=1}^{n+1} \bigl\{ \theta_{n} (N_{t}-\mathbb{E} N_{1})-l_{n}N_{t-1} \bigr\} \Biggr\} \\& \quad = \biggl(\exp \biggl\{ \frac{a_{0}}{1-a_{1}}\bigl[(1-a_{1}) \bigl(e^{\theta_{n}}-1\bigr)-\theta _{n}\bigr] \biggr\} \biggr)^{n+1}. \end{aligned}$$
(2.4)
On the other hand,
$$\begin{aligned}& \mathbb{E}\exp \Biggl\{ \sum_{t=1}^{n+1} \bigl\{ \theta_{n} (N_{t}-\mathbb{E} N_{1})-l_{n}N_{t-1} \bigr\} \Biggr\} \\& \quad =\exp \bigl\{ -(n+1)l_{n}\mathbb{E}N_{1} \bigr\} \mathbb{E} \Biggl\{ \exp \bigl\{ \theta_{n}N_{n+1}-l_{n}N_{0} \bigr\} \exp \Biggl\{ (\theta_{n}-l_{n} )\sum _{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} \Biggr\} \\& \quad =\exp \bigl\{ -(n+1) l_{n}\mathbb{E}N_{1} \bigr\} \mathbb{E}\exp \Biggl\{ \theta_{n}N_{n+1}+ ( \theta_{n}-l_{n} )\sum_{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} . \end{aligned}$$
(2.5)
Combining (2.4) and (2.5) and by the definition of \(l_{n}\), By the Taylor expansion \(e^{\theta_{n}}=1+\theta_{n}+\frac{1}{2}\theta_{n}^{2}+o(\theta_{n}^{2})\), the right-hand side is asymptotically equivalent to Thus By the fact that \(\sup_{t\geq1}\mathbb{E}\exp\{\theta N_{t}\} <\infty\) (\(\forall\theta>0\)) (see Li [2]) and \(\theta_{n}\to0\), a standard argument of an exponential approximation by the Hölder inequality enables us to remove the term \(\theta_{n}N_{n+1}\) from the above equation. So we have
$$\begin{aligned}& \biggl(\exp \biggl\{ \frac{a_{0}}{1-a_{1}}\bigl[(1-a_{1}) \bigl(e^{\theta_{n}}-1\bigr)-\theta _{n}\bigr] \biggr\} \biggr)^{n+1} \bigl(\exp \bigl\{ a_{1}\bigl(e^{\theta_{n}}-1 \bigr)\mathbb{E} N_{1} \bigr\} \bigr)^{n+1} \\& \quad = \biggl(\exp \biggl\{ \frac{a_{0}}{1-a_{1}}\bigl[e^{\theta_{n}}-1- \theta_{n}\bigr] \biggr\} \biggr)^{n+1} \\& \quad = \mathbb{E}\exp \Biggl\{ \theta_{n}N_{n+1}+ ( \theta_{n}-l_{n} )\sum_{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} . \end{aligned}$$
$$\exp \biggl\{ {1\over 2}\sigma^{2} \beta^{2} b_{n} \biggr\} . $$
$$\lim_{n\to\infty}{1\over b_{n}}\log\mathbb{E}\exp \Biggl\{ \theta_{n}N_{n+1}+ (\theta_{n}-l_{n} )\sum _{t=1}^{n} (N_{t}- \mathbb{E}N_{1}) \Biggr\} ={1\over 2}\sigma^{2} \beta^{2}. $$
$$ \lim_{n\to\infty}{1\over b_{n}}\log\mathbb{E} \exp \Biggl\{ (\theta_{n}-l_{n} )\sum _{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} ={1\over 2}\sigma^{2} \beta^{2}. $$
(2.6)
By the Hölder inequality, therefore, where the second step follows from the fact that which can be proved by Jensen’s inequality.
$$\begin{aligned}& \mathbb{E}\exp \Biggl\{ (\theta_{n}-l_{n} )\sum _{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \\& \quad \le \Biggl(\mathbb{E}\exp \Biggl\{ \theta_{n}\sum _{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \Biggr)^{\theta_{n}-l_{n}\over \theta_{n}}\le\mathbb{E}\exp \Biggl\{ \theta_{n}\sum_{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} , \end{aligned}$$
$$\mathbb{E}\exp \Biggl\{ \theta_{n}\sum_{t=1}^{n} (N_{t}-\mathbb {E}N_{1}) \Biggr\} \ge1, $$
By the fact that \(\theta_{n}=\beta\sqrt{b_{n}\over n}\) and by (2.6), we obtain the lower bound
$$ \liminf_{n\to\infty}{1\over b_{n}}\log \mathbb{E}\exp \Biggl\{ \beta\sqrt{b_{n}\over n}\sum _{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \ge{1\over 2}\sigma^{2} \beta^{2}. $$
(2.7)
On the other hand, given a small number \(0<\delta<1\), \(\theta_{n}-l_{n} >(1-\delta)\theta_{n} =(1-\delta)\beta\sqrt{b_{n}\over n} \) as n is sufficiently large. By the Hölder inequality By (2.6), therefore, Since \(\beta\in\mathbb{R}\) can be arbitrary, replacing it by \((1-\delta)^{-1}\beta\) in the above leads to Letting \(\delta\to0^{+}\) on the right-hand side yields the desired upper bound, which, together with the lower bound (2.7), leads to (2.3). □
$$\begin{aligned}& \mathbb{E}\exp \Biggl\{ (1-\delta)\beta\sqrt{b_{n}\over n}\sum _{t=1}^{n} (N_{t}-\mathbb{E}N_{1}) \Biggr\} \\& \quad \le \Biggl(\mathbb{E}\exp \Biggl\{ (\theta_{n}-l_{n} ) \sum_{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \Biggr)^{(1-\delta)\theta_{n}\over \theta_{n}-l_{n}} \le \mathbb{E}\exp \Biggl\{ ( \theta_{n}-l_{n} )\sum_{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} . \end{aligned}$$
$$\limsup_{n\to\infty}{1\over b_{n}}\log \mathbb{E}\exp \Biggl\{ (1-\delta)\beta\sqrt{b_{n}\over n}\sum_{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \le{1\over 2} \sigma^{2} \beta^{2}. $$
$$\limsup_{n\to\infty}{1\over b_{n}}\log \mathbb{E}\exp \Biggl\{ \beta\sqrt{b_{n}\over n}\sum_{t=1}^{n} (N_{t}-\mathbb{E} N_{1}) \Biggr\} \le{1\over 2} \sigma^{2} \biggl({\beta\over 1-\delta} \biggr)^{2}. $$
As a consequence of Lemma 2.2, for every \(\eta>0\), considering the closed set \(H=\{x,|x|\geq\eta\}\) and \(b_{n}=\sqrt{n}\), we have where \(c_{\eta}=\frac{\eta^{2}}{2\sigma^{2}}>0\) is independent of n. This gives the genuine exponential decay for the probability that the sample average deviates from its expectation.
$$ \mathbb{P} \Biggl(\frac{1}{n}\Biggl\vert \sum _{t=1}^{n}(N_{t}-\mathbb {E}N_{1})\Biggr\vert \geq \eta \Biggr)\leq\mathbb{P} \Biggl( \frac{1}{n^{3/4}}\Biggl\vert \sum_{t=1}^{n}(N_{t}- \mathbb{E}N_{1})\Biggr\vert \geq\eta \Biggr) \le \exp \{-c_{\eta}\sqrt{n}\}, $$
(2.8)
The last lemma below is a restatement of Theorem 3.1 of Ng et al. [7].
Lemma 2.3
Let
\(\{Y, Y_{j}, j\geqslant1\}\)
be a sequence of i.i.d. non-negative random variables with common distribution function
\(F_{Y}\in\mathscr{C}\)
and finite expectation
μ, let
\(Q_{n}=\sum_{j=1}^{n}Y_{j}\). Then, for any fixed
\(\gamma>0\),
$$ P(Q_{n}-n\mu>y)\sim n \overline{F}_{Y}(y)\quad (n \rightarrow\infty) \textit{ uniformly for } y\geq\gamma n. $$
3 Proof of Theorem 1.1
Throughout this section, unless otherwise stated, every limit relation is understood as valid uniformly for all \(x\geq\gamma n\) as \(n\rightarrow\infty\). Trivially, equation (1.3) amounts to the conjunction of which will be proven separately in the following two subsections.
$$ \begin{aligned} &\mathbb{P}\biggl(S_{n}-\frac{a_{0}}{1-a_{1}}n \mu >x\biggr)\lesssim\frac{na_{0}}{1-a_{1}} \overline{F}_{X}(x)\quad \text{and} \\ &\mathbb{P}\biggl(S_{n}-\frac{a_{0}}{1-a_{1}}n\mu >x\biggr)\gtrsim \frac{na_{0}}{1-a_{1}} \overline{F}_{X}(x), \end{aligned} $$
(3.1)
3.1 Proof of the first relation (3.1)
Write \(\nu=\mathbb{E}N_{1}=\frac{a_{0}}{1-a_{1}}\). For arbitrarily fixed, but small η, \(0<\eta<1\), by Lemma 2.1, we derive where \([\cdot]\) denotes the integral part of ⋅, throughout this paper.
$$\begin{aligned}& \mathbb{P}(S_{n}-\nu n\mu>x) \\& \quad = \mathbb{P}\Biggl(\sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}- \nu n\mu >x\Biggr) \\& \quad = \mathbb{P}\Biggl\{ \sum_{t=1}^{n}N_{t}< n(\nu+\eta), \sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}- \nu n\mu>x\Biggr\} \\& \qquad {} +\mathbb{P}\Biggl\{ \sum_{t=1}^{n}N_{t} \geq n(\nu+\eta), \sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}- \nu n\mu>x\Biggr\} \\& \quad \leq \mathbb{P}\Biggl(\sum_{j=1}^{[n(\nu+\eta)]}Y_{j}- \nu n\mu>x\Biggr)+\mathbb{P}\Biggl\{ \sum_{t=1}^{n}N_{t} \geq n(\nu+\eta), \sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}- \nu n\mu>x\Biggr\} \\& \quad = \Delta_{1}+\Delta_{2}, \end{aligned}$$
(3.2)
From Lemma 2.3, we see that, for some small η such that \(\gamma-\eta\mu>0\),
$$\begin{aligned} \begin{aligned}[b] \Delta_{1}&= \mathbb{P}\Biggl(\sum _{j=1}^{[n(\nu+\eta)]}Y_{j}-\bigl[n(\nu+\eta)\bigr]\mu >x+n\nu\mu-\bigl[n(\nu+\eta)\bigr]\mu\Biggr) \\ &\sim \bigl[n(\nu+\eta)\bigr]\overline{F}_{Y}\bigl(x+n\nu\mu- \bigl[n(\nu+\eta)\bigr]\mu \bigr) \\ &\leq n(\nu+\eta)\overline{F}_{X}\bigl(x(1-\eta\mu/\gamma)\bigr). \end{aligned} \end{aligned}$$
(3.3)
As for the second term, By the assumption that \(F_{Y}=F_{X}\in\mathscr{C}\) there is a constant \(C>0\) independent of m, and for every \(p>J^{+}_{F}\) such that \(\overline{F}_{X} ({x\over m} )\leq Cm^{p}\overline{F}_{X}(x)\) for all m and sufficiently large x (see Lemma 2.2, Ng et al. [7]), where \(J^{+}_{F}\) is called the upper Matuszewska index. Hence We now claim that
$$\begin{aligned} \Delta_{2} =&\mathbb{P} \Biggl\{ \sum_{t=1}^{n}N_{t} \geqslant n(\nu+\eta), \sum_{j=1}^{N_{1}+\cdots+N_{n}}Y_{j}-n \nu\mu>x \Biggr\} \\ =&\sum_{m\geq n(\nu+\eta)}\mathbb{P} \Biggl\{ \sum _{t=1}^{n} N_{t}=m \Biggr\} \mathbb{P} \Biggl\{ \sum_{j=1}^{m}Y_{j}-n\nu \mu>x \Biggr\} \\ \leq&\sum_{m\geq n(\nu+\eta)}\mathbb{P} \Biggl\{ \sum _{t=1}^{n} N_{t}=m \Biggr\} \mathbb{P} \Biggl\{ \sum_{j=1}^{m}Y_{j}>x \Biggr\} \\ \leq&\sum_{m\geq n(\nu+\eta)}\mathbb{P} \Biggl\{ \sum _{t=1}^{n} N_{t}=m \Biggr\} \sum _{j=1}^{m}\mathbb{P} \biggl\{ Y_{j}> \frac{x}{m} \biggr\} \\ \leq&\sum_{m\geq n(\nu+\eta)}\mathbb{P} \Biggl\{ \sum_{t=1}^{n} N_{t}=m \Biggr\} m\overline{F}_{Y} \biggl({x\over m} \biggr). \end{aligned}$$
$$\begin{aligned} \Delta_{2} \leq& C\overline{F}_{X}(x)\sum _{m\geq n(\nu+\eta)}m^{p+1}\mathbb{P} \Biggl\{ \sum _{t=1}^{n} N_{t}=m \Biggr\} \\ \leq& C\overline{F}_{X}(x)\mathbb{E} \Biggl\{ \Biggl(\sum _{t=1}^{n} N_{t}\Biggr)^{p+1} \mathbb {1}_{\{\sum_{t=1}^{n} N_{t}\geq n(\nu+\eta)\}} \Biggr\} . \end{aligned}$$
(3.4)
$$ \lim_{n\rightarrow\infty}\mathbb{E} \Biggl\{ \Biggl(\sum _{t=1}^{n} N_{t} \Biggr)^{p+1} \mathbb {1}_{\{\sum_{t=1}^{n} N_{t}\geq n(\nu+\eta)\}} \Biggr\} =0. $$
(3.5)
Indeed, by the Cauchy-Schwarz inequality
$$\begin{aligned}& \mathbb{E} \Biggl\{ \Biggl(\sum_{t=1}^{n} N_{t}\Biggr)^{p+1} \mathbb {1}_{\{\sum_{t=1}^{n} N_{t}\geq n(\nu+\eta)\}} \Biggr\} \\& \quad \leq \Biggl\{ \mathbb{E}\Biggl(\sum_{t=1}^{n} N_{t}\Biggr)^{2(p+1)}\Biggr\} ^{1/2} \Biggl\{ \mathbb{P} {\Biggl(\sum_{t=1}^{n} N_{t}\geq n(\nu+\eta)\Biggr)}\Biggr\} ^{1/2}. \end{aligned}$$
Using (2.8), we have where \(c_{\eta}=\frac{\eta^{2}}{2\sigma^{2}}>0\) is independent of n. This together with fact that \(\mathbb{E}(\sum_{t=1}^{n} N_{t})^{2(p+1)}=O(n^{2(p+1)}) \) proves (3.5).
$$ \mathbb{P} {\Biggl(\sum_{t=1}^{n} N_{t}\geq n(\nu+\eta)\Biggr)}\leq\mathbb{P} \Biggl(\frac{1}{n} \Biggl\vert \sum_{t=1}^{n}(N_{t}- \nu)\Biggr\vert \geq\eta \Biggr) \le\exp\{ -c_{\eta}\sqrt{n}\}, $$
Hence, by (3.4) and (3.5), we have Substituting (3.3) and (3.6) into (3.2) yields By the arbitrariness of η and the condition \(F_{X}\in\mathscr{C}\), we obtain the first relation (3.1).
$$ \Delta_{2}\leq o\bigl(\overline{F}_{X}(x) \bigr)=o\bigl(n\nu\overline{F}_{X}(x)\bigr). $$
(3.6)
$$ \mathbb{P}(S_{n}-n\nu\mu>x)\lesssim n(\nu+\eta)\overline{F}_{X} \bigl(x(1-\eta\mu/\gamma)\bigr)+o\bigl(n\nu\overline{F}_{X}(x)\bigr). $$
3.2 Proof of the second relation (3.1)
Let \(0<\eta<1\) be arbitrarily fixed with η small. Consider the decomposition Applying (2.8) we have By Lemma 2.3, have Finally, substituting (3.8) and (3.9) into (3.7) yields By the arbitrariness of η and the condition \(F_{X}\in\mathscr{C}\), we obtain the second relation (3.1).
$$\begin{aligned} \mathbb{P}(S_{n}-n\nu\mu>x) =&\mathbb{P}\Biggl(\sum _{j=1}^{N_{1}+N_{2}+\cdots+N_{n}}Y_{j}-n\nu\mu>x\Biggr) \\ \geq&\mathbb{P}\Biggl(\sum_{t=1}^{n}N_{t}> n(\nu-\eta), \sum_{j=1}^{N_{1}+N_{2}+\cdots+N_{n}}Y_{j}-n \nu\mu>x\Biggr) \\ \geq&\mathbb{P}\Biggl(\sum_{t=1}^{n}N_{t}> n(\nu-\eta)\Biggr)\cdot \mathbb{P}\Biggl( \sum_{j=1}^{[n(\nu-\eta)]}Y_{j}-n \nu\mu>x\Biggr). \end{aligned}$$
(3.7)
$$ \mathbb{P}\Biggl\{ \sum_{t=1}^{n}N_{t}> n(\nu-\eta)\Biggr\} \geq \mathbb{P}\Biggl\{ \frac{1}{n}\Biggl\vert \sum _{t=1}^{n}(N_{t}-\nu)\Biggr\vert < \eta\Biggr\} \rightarrow1, \quad \mbox{as } n\rightarrow\infty. $$
(3.8)
$$\begin{aligned} \mathbb{P}\Biggl( \sum_{j=1}^{[n(\nu-\eta)]}Y_{j}-n \nu\mu>x\Biggr) \sim& \bigl[n(\nu-\eta)\bigr]\overline{F}_{Y}\bigl(x+n \nu\mu-\bigl[n(\nu-\eta)\bigr]\mu \bigr) \\ \geq& n(\nu-\eta)\frac{[n(\nu-\eta)]}{n(\nu-\eta)}\overline {F}_{X}\bigl(x(1+\eta \mu/\gamma)\bigr). \end{aligned}$$
(3.9)
$$\mathbb{P} (S_{n}-n\nu\mu>x )\gtrsim n(\nu-\eta)\overline{F}_{X} \bigl(x(1+\eta\mu/\gamma)\bigr). $$
Acknowledgements
This work is supported by National Natural Science Foundation of China (No. 11571138).
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Competing interests
The author declares that they have no competing interests.