Asymptotic Ruin Probabilities for a Renewal Risk Model with a

Random Number of Delayed Claims

Jinzhu Li
School of Mathematical Science and LPMC
Nankai University, Tianjin 300071, P.R. China
E-mail: [email protected]

May 19, 2022

Abstract
In this paper, we consider a non-standard renewal risk model, in which each main claim may
produce a random number of delayed claims. Such a model takes full account of lasting impacts
of severe claims and hence may avoid the underestimation of an insurer’s operation risk within
a relatively long period. We obtain asymptotic expansions for the …nite-time ruin probability,
given that the claim size distributions belong to the general subexponential class. When the
claim size distributions are restricted to the smaller class of extended regular variation, we push
our study forward to the in…nite-time ruin probability.
Keywords: asymptotics; renewal risk model; ruin probability; delayed claim; subexponen-
tiality; extended regular variation
Mathematics Subject Classi…cation: Primary 62P05; Secondary 62E20, 91B30

1 Introduction
In the real world of insurance industry, almost all severe accidents and natural disasters may trigger
two types of claims. The …rst type is made up of the claims that need to be settled promptly and
we prefer to call them main claims, including direct property losses and immediate death. The
second type consists of the so-called delayed claims, such as claims caused by subsequent medical
treatment and death, that may gradually occur in a period of time after their corresponding main
claims. Taking a tra¢ c accident as an example, a damaged car will give rise to an auto insurance
claim (the main claim) immediately, while the injured driver and passengers or their insurance
bene…ciaries may also claim medical reimbursements or death compensations (the delayed claims)
some time later. Thus, ignoring the follow-up e¤ects of the main claims may lead to a serious
underestimation of an insurer’s operation risk within a relatively long period.
Actually, in the last two decades more and more researchers in the …eld of risk theory have
started to pay close attention to such lasting impacts of severe claims. In the meantime, quite
a few risk models with delayed claims taken into account have been proposed and investigated.
See Waters and Papatriandafylou (1985), Yuen and Guo (2001), Yuen et al. (2005) and Xiao
and Guo (2007) for studies on the scenario where both main and delayed claims are modeled
by random variables with light tails. On the other hand, empirical data have proved that most
insurance claims, especially the severe ones, possess the heavy-tail feature; see, e.g., Embrechts
et al. (1997) and McNeil et al. (2005). Moreover, an insurer is always required to hold a huge
number of initial assets to guarantee its solvency under certain regulatory frameworks, say, the
popular Solvency II Framework. Hence, the related studies now are increasingly turning their focus
to seeking asymptotic results for important actuarial quantities, including ruin probabilities, as the

1
claims are heavy-tailed and the insurer’s initial assets is large. We refer the reader to Li (2013),
Fu et al. (2015), Fu and Li (2016), Gao et al. (2019) and Yang and Li (2019) for some recent
contributions on this direction.
However, it is worth noting that all the existing related works, including the aforementioned
ones, assumed that each main claim will de…nitely produce only one delayed claim. Clearly, this
unrealistic assumption is made solely for mathematical tractability rather than practical relevance.
Recalling the example of the tra¢ c accident mentioned above, it is entirely possible that a main
claim may produce more than one delayed claims. Actually, we can not predict the exact number
of delayed claims derived from a main claim in a certain period of future, and hence it is more
reasonable to describe such a number as a random variable taking values in the set of non-negative
integers.
Based on the above facts, in this paper we consider a non-standard renewal risk model, in
which each main claim may produce a random number of delayed claims. Speci…cally speaking,
for each positive integer i, we assume that an insurer’s ith main claim with size Xi will occur at
time i and may produce Mi delayed claims. Each of the delayed claims has size Yij and occurs
at i + Dij for 1 j Mi , where Dij denotes the delay time of the jth delayed claim caused by
the ith main claim. Let fMi ; i 1g be a sequence of non-negative, integer-valued and identically
distributed random variables with generic random variable M . Let fXi ; i 1g, fYij ; i 1; j 1g
and fDij ; i 1; j 1g be three sequences of non-negative, independent and identically distributed
(i.i.d.) random variables with generic random variables X, Y , D and common distributions F , G,
H, respectively. Assume that the arrival times of the main claims 1 , 2 , . . . constitute a renewal
sequence such that the inter-arrival times 1 , 2 1, 3 2 , . . . are i.i.d. and the corresponding
counting process fN (t); t 0g is a renewal process with …nite renewal function
1
X
(t) = EN (t) = P( i t) :
i=1
Throughout the paper, we assume that fXi ; i 1g, fMi ; i 1g, fYij ; i 1; j 1g, fDij ; i 1; j 1g
and fN (t); t 0g are mutually independent. Moreover, denote by non-negative constants x, c and
r the initial assets, the premium rate and the force of interest, respectively. Then, the insurer’s
surplus process can be expressed as
Z t N (t)
X N (t) Mi
XX
R(t) = xert + c er(t s) ds Xi er(t i ) Yij er(t i Dij ) 1f i +Dij tg ; t 0; (1.1)
0 i=1 i=1 j=1

where 1f g is the indictor function. For any …xed time T > 0, the …nite-time ruin probability of
risk model (1.1) can be formulated as

(x; T ) = P inf R(t) < 0 R(0) = x ;

0 t T

and hence the corresponding in…nite-time ruin probability is

lim (x; T ) =: (x) :
T !1
Following the latest trend in research, in this paper we aim to seek asymptotic expansions for
(x; T ) as x ! 1, given that the distributions of the main and delayed claims (i.e., F and G)
belong to the general subexponential class, which covers almost all commonly-used distributions
with heavy tails. Furthermore, when F and G are restricted to the smaller class of extended regular
variation, we will push our study forward to obtain asymptotic results for (x).
The rest of this paper consists of three sections. In Section 2 we state our main results after
introducing necessary preliminaries. In Section 3 we prepare some useful lemmas for later use. In
Section 4 we prove our main results.

2
2 Preliminaries and Main Results
Hereafter, all limit relationships hold as x ! 1 unless stated otherwise. For two positive functions
f and g, we write f (x) . g (x) or g(x) & f (x) if lim sup f (x) =g (x) 1, write f (x) g (x)
if lim f (x) =g (x) = 1, write f (x) = o (g (x)) if lim f (x) =g (x) = 0 and write f (x) g (x) if
0 < lim inf f (x) =g (x) < lim sup f (x) =g (x) < 1. As usual, for two real numbers a and b, we
write a ^ b = minfa; bg and a _ b = maxfa; bg. Particularly, we write a+ = a _ 0. For two real-
valued random variables 1 and 2 , we say that 1 is stochastically not greater than 2 , denoted by
1 st 2 , if P ( 1 > x) P ( 2 > x) for every x 2 ( 1; 1).
A distribution V on [0; 1) is said to belong to the subexponential class, denoted by V 2 S, if
V (x) = 1 V (x) > 0 for all x 0 and the relation

V n (x)
lim =n
V (x)

holds for all (or, equivalently, for some) n 2, where V n is the n-fold convolution of V with itself.
Further, when V is supported on ( 1; 1), we still say it belongs to the class S if V (x)1fx 0g
does. It is known that if V 2 S then V 2 L, which stands for the class of long-tailed distributions
characterized by the property

V (x + z)
lim = 1; z 2 ( 1; 1):
V (x)

An important subclass of the class S is the class ERV of distributions with extended regularly
varying tails. By de…nition, V belongs to the class ERV if there are some 0 < < 1 such
that
V (xz) V (xz)
z lim inf lim sup z ; z 2 (0; 1): (2.1)
V (x) V (x)
We usually signify the regularity property shown in (2.1) as V 2 ERV( ; ), so that ERV is the
union of all ERV( ; ) over the range 0 < < 1. In particular, when = , the class
ERV( ; ) coincides with the famous class R of distributions with regularly varying tails.
Thus, if V 2 R for some 0 < < 1 then

V (xz)
lim =z ; z > 0: (2.2)
V (x)

We refer the reader to Bingham et al. (1987) and Embrechts et al. (1997) for comprehensive
introductions on the above-mentioned distribution classes.
For a distribution V 2 S, the well-known Kesten’s bound says that, for every " > 0, there is
some C > 0 such that the inequality

V n (x) C(1 + ")n V (x) (2.3)

holds for all n 1 and x 0. The proof of (2.3) can be found in Athreya and Ney (1972) or
Lemma 1.3.5(c) of Embrechts et al. (1997). For a distribution V 2 ERV( ; ), by Proposition
2.2.3 of Bingham et al. (1987) we know that, for every " > 0, there are some x0 > 0 and C > 1
such that the inequalities

1 +" " V (xz) +" "

z ^z C z _z (2.4)
C V (x)

3
hold whenever x > x0 and xz > x0 . Actually, by Lemma 1(ii) of Li (2018), only the condition
x > x0 su¢ ces to establish the second inequality in (2.4) for the upper bound part. It can be
derived from (2.4) that, if V 2 ERV( ; ), then for every p > we have
p
x = o(V (x)): (2.5)

Now, it is time to state our …rst theorem regarding the …nite-time ruin probability.

Theorem 2.1. Consider risk model (1.1), in which fMi ; i 1g is a sequence of i.i.d. random
variables with Ee M < 1 for some > 0. Let T > 0 be any …xed time such that P ( 1 T ) > 0.
(i) If F 2 S, G 2 S and G(x) F (x), then
Z T Z T Z T t
rt
(x; T ) F (xe ) (dt) + EM G(xer(s+t) )H(ds) (dt): (2.6)
0 0 0

(ii) If F 2 S and G(x) = o F (x) , then

Z T
(x; T ) F (xert ) (dt): (2.7)
0

Next, we turn to asymptotic expansions for the in…nite-time ruin probability. To this purpose,
the claim size distributions have to be restricted to the class ERV.

Theorem 2.2. Consider risk model (1.1), in which r > 0 and fMi ; i 1g is a sequence of i.i.d.
random variables with Ee M < 1 for some > 0.
(i) If F 2ERV and G 2ERV, then
Z 1 Z 1Z 1
rt
(x) F (xe ) (dt) + EM G(xer(s+t) )H(ds) (dt): (2.8)
0 0 0

(ii) If F 2ERV and G(x) = o F (x) , then

Z 1
(x) F (xert ) (dt): (2.9)
0

It is shown from Theorems 2.1(ii) and 2.2(ii) that, under the conditions of the theorems, the
asymptotic ruin probabilities are insensitive to the delayed claims whenever the common tail of the
delayed claims is negligible compared to that of the main claims. This fact is obviously consistent
with our intuition.
Furthermore, if we restrict our discussion within the scope of regular variation, then (2.8) and
(2.9) in Theorem 2.2 can be re…ned to much more elegant forms, via using (2.2), (2.4) and the
dominated convergence theorem. We summarize such results to establish the following corollary.

Corollary 2.1. Let the other conditions of Theorem 2.2 hold.

(i) If F 2 R 1 and G 2 R 2 , then

Ee 1 r 1 Ee 2 rD Ee 2 r 1
(x) r
F (x) + EM G (x) : (2.10)
1 Ee 1 1 1 Ee 2 r 1

(ii) If F 2 R and G(x) = o F (x) , then

Ee r 1
(x) r
F (x) : (2.11)
1 Ee 1

4
 Another interesting scenario of model (1.1) is the one in which the main claims may produce
an identical number of delayed claims and such a number is described by the random variable M .
In this case, through the arguments similar to (but simpler than) the ones shown below in the
proofs of Theorems 2.1 and 2.2, we can still obtain relations (2.6)–(2.11) under the corresponding
conditions, given that a stronger moment assumption regarding M and N (T ) is satis…ed when
considering the …nite-time ruin probability. Hence, we just list such results in the following two
theorems and omit their derivations.
Theorem 2.3. Consider risk model (1.1), in which Mi = M for each i 1. Let T > 0 be any
…xed time such that P ( 1 T ) > 0. Assume that Ee M N (T ) < 1 for some > 0. Then, relations
(2.6) and (2.7) hold under conditions (i) and (ii) of Theorem 2.1, respectively.
Theorem 2.4. Consider risk model (1.1), in which r > 0 and Mi = M for each i 1 with
Ee M < 1 for some > 0. Then, relations (2.8) and (2.9) hold under conditions (i) and (ii) of
Theorem 2.2, respectively. Also, relations (2.10) and (2.11) hold under conditions (i) and (ii) of
Corollary 2.1, respectively.
We remark that, when M degenerates to a non-negative integer (or, in other words, follows a
one-point distribution), our main results mentioned above are applicable for lots of well-studied
models, including the ones without delayed claims (i.e., M = 0) and the ones in which each main
claim may produce exactly one delayed claim (i.e., M = 1). Additionally, in case M is a constant,
the moment assumptions regarding M and N (T ) in Theorems 2.1–2.4 (including Corollary 2.1) are
redundant, since fN (t); t 0g is a renewal process and hence the moment generating function of
N (T ) always exists in a neighborhood of 0.

3 Lemmas
We begin with some results regarding the asymptotic tail behavior of randomly weighted sums of
subexponential random variables. The following Lemma 3.1(i) was given by Lemma 3.2 of Yang
and Li (2019) and is a natural extension to Theorem 1 of Tang and Yuan (2014), while Lemma
3.1(ii) can be proved by the arguments similar to those used in the proof of Lemma 3.2 of Yang
and Li (2019) combined with Theorem 3 of Tang and Yuan (2014) (see also Chen and Yuen (2009)
or Zhang et al. (2009)) and Theorem 3.5(v) of Cline and Samorodnitsky (1994).
Lemma 3.1. Let Z1 , . . . , Zn+m be n + m independent real-valued random variables with distribu-
tions V1 , . . . , Vn+m . Let 1 , . . . , n+m be n + m non-negative random variables, where 1 , . . . , n
are non-degenerate at 0. Assume that fZ1 ; : : : ; Zn+m g and f 1 ; : : : ; n+m g are independent.
(i) Let Vi 2 S, V i (x) V 1 (x) and V j (x) = o V 1 (x) for each 1 i n and n + 1 j n + m.
Assume that 1 , . . . , n are bounded above and for each n + 1 j n + m there is some 1 ij n
such that j st ij . Then,
n+m
! n
X X
P Z
k k > x P ( i Zi > x) : (3.1)
k=1 i=1

(ii) Assume that V1 2 ERV( 1 ; 1 ), . . . , Vn 2 ERV( n; n ) and for each nW+ 1 j n+m
there is some 1 ij n such that V j (x) = o V ij (x) . If there is some p > ni=1 i such that
E pk < 1 for each 1 k n + m, then (3.1) holds.
The next lemma is due to Theorem 3.3 of Chen and Yuen (2009), and it is concerned with the
in…nite sum of randomly weighted random variables with extended regularly varying tails. See also
Wang and Tang (2006) and Zhang et al. (2009) for related discussions.

5
Lemma 3.2. Let fZi ; i 1g be a sequence of i.i.d. non-negative random variables with common
distribution V 2 ERV( ; ). Let f i ; i 1g be a sequence of non-negative random variables
independent of fZi ; i 1g. If there is some 0 < < ^ (1 )1f0< <1g + 1f 1g such that
1
X 1
X 1
1
+ ( + )_1
E i
( + )_1
+ E i < 1; (3.2)
i=1 i=1

then !
1
X 1
X
P i Zi >x P ( i Zi > x) :
i=1 i=1

The following lemma will play a role in the proof of Theorem 2.2(ii).

Lemma 3.3. Let fZi ; i 1g be a sequence of i.i.d. real-valued random variables with common dis-
tribution V satisfying V (x) = o U (x) for some distribution U 2 ERV( ; ). Let f i ; i 1g be
a sequence of non-negative random variables independent of fZi ; i 1g and satisfying the moment
condition given by (3.2). Then,
1
!
X
P i Zi > x = o U (x) :
i=1

Proof. Denote by Z the generic random variable of fZi ; i 1g and by V+ the distribution of Z + .
For every 0 < " < 1, since V+ (x) V (x) = o U (x) , it holds for large x that V+ (x) "U (x).
Introduce a new non-negative random variable Z with the survival function as

V (x) := max "U (x) ; V+ (x) V+ (x); x 2 ( 1; 1):

Clearly, we have Z + st Z and V (x) "U (x), which implies that V 2 ERV( ; ). Let Z1 ,
Z2 , . . . be i.i.d. copies of Z and let them be independent of all other sources of randomness. Thus,
1
! 1
! 1
!
X X X
+
P i Zi > x P i Zi > x P i Zi > x ;
i=1 i=1 i=1

where in the last step

P we also used the continuityPof probability and the monotonicity of the
event sequences f ni=1 i Zi+ > xg; n 1 and ff ni=1 i Zi > xg; n 1g. Since (3.2) holds for
f i ; i 1g, applying Lemma 3.2 leads to
1
! 1
X X
P Z
i i > x . P ( i Zi > x) : (3.3)
i=1 i=1

Let =( + ) < < 1. For each i 1, we further write

Z x
V (x=u)
P ( i Zi > x) U (x=u) P ( i 2 du) + P ( i > x ):
0 U (x=u)

Recalling (2.4) and V (x) "U (x), there are some x0 > 0 and C > 1 such that the relation
Z x
V (x=u) +
U (x=u) P ( i 2 du) "C E i +E i U (x)
0 U (x=u)

6
holds for all x > x0 . Here, it is worth noting that x0 is unrelated to i and that C is unrelated to
both i and ". Plugging these estimates into (3.3) and using Chebyshev’s inequality yield that
1
! 1 1
X X X
P i Zi > x . "C E i +E i +
U (x) + E i+ x ( + )
i=1 i=1 i=1
1
X +
"C E i +E i U (x) ;
i=1

where the last step follows from (2.5). Then, noting (3.2) and the arbitrariness of " completes the
proof.

The following lemma is extracted from the proof of Theorem 2.1 of Hao and Tang (2008), and
will be used in deriving the asymptotic expansions for the …nite-time ruin probability.
Lemma 3.4. Let fZi ; i 1g be a sequence of i.i.d. non-negative random variables with common
distribution V 2 S, and let fN (t); t 0g be a renewal counting process with jump times 1 , 2 , . . .
and renewal function (t). Assume that fZi ; i 1g and fN (t); t 0g are independent. Then, for
any r 0 and T > 0 such that P ( 1 T ) > 0, we have
P
1 P
n
r
P (Zi e i > x; N (T ) = n)
n=m+1 i=1
lim sup RT = 0:
m!1 x 0 V (xert ) (dt)
0

The next two lemmas compose the crucial ingredients of the proofs of Theorems 2.1 and 2.2. For
the conciseness in writing, we denote hereafter by (x; T ), o (x; T ), (x), and o (x) the right-hand
sides of (2.6)–(2.9), i.e.,
Z T Z T Z T t
(x; T ) := F (xert ) (dt) + EM G(xer(s+t) )H(ds) (dt); (3.4)
0 0 0
Z T
o (x; T ) := F (xert ) (dt); (3.5)
0
Z 1 Z 1Z 1
(x) := F (xert ) (dt) + EM G(xer(s+t) )H(ds) (dt); (3.6)
0 0 0
and Z 1
o (x) := F (xert ) (dt): (3.7)
0
Lemma 3.5. Let the conditions of Theorem 2.1 hold. We have
0 0 1 1 8
N (T )
X Mi
X < (x; T ); under (i)
P @ @ Xi e r i
+ Yij e r( i +Dij ) A
1f i +Dij T g > xA ;
:
i=1 j=1 o (x; T ); under (ii)

where (x; T ) and o (x; T ) are given by (3.4) and (3.5).

Proof. We only give in detail the proof under condition (i), because the proof under condition (ii)
can be parallelly established with some slight modi…cations, which will be speci…ed at the end of
this proof.

7
 Under condition (i), arbitrarily choose some large integer N and write
0 0 1 1
N (T ) Mi
X X
P@ @Xi e r i + Yij e r( i +Dij ) 1f i +Dij T g A > xA
i=1 j=1

N 1
! 0 n 0 Mi
1 1
X X X X
= + P@ @Xi e r i
+ Yij e r( i +Dij )
1f i +Dij Tg
A > x; N (T ) = nA
n=1 n=N +1 i=1 j=1

=: I1 (x; T ) + I2 (x; T ): (3.8)

We …rst deal with I1 (x; T ). For any set A of non-negative integers, we denote a vector (ai ; i 2 A)
_ W _
as aA and write aA = i2A ai . For each 1 n N , according to whether M f1;:::;ng N or not,
we further split I1 (x; T ) into two parts denoted by I11 (x; T ) and I12 (x; T ). It holds that

I11 (x; T )
0 0 1 1
N
X X n
X mi
X
= P@ @Xi e r i
+ Yij e r( i +Dij )
1f i +Dij Tg
A > x; N (T ) = nA
n=1 mf1;:::;ng 2f0;:::;N gn i=1 j=1

P Mf1;:::;ng = mf1;:::;ng
0 1
N
X X n
X X mi
n X
= P@ Xi e r i
1fN (T )=ng + Yij e r( i +Dij )
1f i +Dij T g 1fN (T )=ng > xA
n=1 mf1;:::;ng 2f0;:::;N gn i=1 i=1 j=1

P Mf1;:::;ng = mf1;:::;ng :

Now, consider e r i 1fN (T )=ng and e r( i +Dij ) 1f i +Di T g 1fN (T )=ng with 1 i n and 1 j mi
as n + m1 + + mn random weights. Then, applying Lemma 3.1(i) yields that

I11 (x; T )
N
X X n
X
r
P Xi e i
1fN (T )=ng > x P Mf1;:::;ng = mf1;:::;ng
n=1 mf1;:::;ng 2f0;:::;N gn i=1
N
X X X mi
n X
r( i +Dij )
+ P Yij e 1f i +Dij T g 1fN (T )=ng > x P Mf1;:::;ng = mf1;:::;ng
n=1 mf1;:::;ng 2f0;:::;N gn i=1 j=1
N X
X n
_
r
= P Xe i
1fN (T )=ng > x P M f1;:::;ng N
n=1 i=1
XN X n N
X _
r( i +D)
+ P Ye 1f i +D T g 1fN (T )=ng >x mi P(M = mi ) P M f1;:::;ngnfig N :
n=1 i=1 mi =0
(3.9)

8
Hence,
1 X
X n
I11 (x; T ) . P Xe r i
1fN (T )=ng > x
n=1 i=1
1 X
X n
r( i +D)
+EM P Ye 1f i +D T g 1fN (T )=ng >x
n=1 i=1
1
X 1
X
r r( i +D)
= P Xe i
1f i Tg > x + EM P Ye 1f i +D Tg >x
i=1 i=1
= (x; T ): (3.10)
By EM < 1 and Lemma 3.4, for every positive but small enough ", all 1 n N and all
1 i n, we can choose N large enough such that
1
X
_ _
P M f1;:::;ngnfig > N P M f1;:::;ng > N N P (M > N ) mP(M = m) < ";
m=N +1
1
X n
X
r
P Xe i
1fN (T )=ng > x < " (x; T );
n=N +1 i=1
and
1
X n
X 1
X n
X
P Ye r( i +D)
1f i +D T g 1fN (T )=ng >x P Ye r i
1fN (T )=ng > x . "w (x; T );
n=N +1 i=1 n=N +1 i=1

where w is some constant such that lim sup G(x)=F (x) < w. Plugging these estimates into (3.9)
leads to
I11 (x; T )
1 X n
!
X
& P Xe r i
1fN (T )=ng > x " (x; T ) (1 ")
n=1 i=1
1 X n
!
X
r( i +D)
+ P Ye 1f i +D T g 1fN (T )=ng >x "w (x; T ) (EM ") (1 ")
n=1 i=1
(1 ") (1 " (1 + w + wEM )) (x; T )
=: (1 "K1 ) (x; T ): (3.11)
Combining (3.10) with (3.11) gives
(1 "K1 ) (x; T ) . I11 (x; T ) . (x; T ): (3.12)
For I12 (x; T ), it holds that
I12 (x; T )
0 0 1 1
1 X
X n n
X Mi
X
P@ @X i e r i
+ Yij e r( i +Dij )
1f i +Dij Tg
A > x; N (T ) = n; Mk > N A
n=1 k=1 i=1 j=1
n
1 X 1 n+m1X
+ +mn
!
X X X
P (Xi + Yi ) > xer 1 ; N (T ) = n
n=1 k=1 mk =N +1 mf1;:::;ngnfkg 2f0;1;:::gn 1 i=1

P Mf1;:::;ng = mf1;:::;ng ;

9
where fYi ; i 1g is a sequence of i.i.d. random variables with G as the common distribution
and is independent of all other sources of randomness. Lemma 3.1(i) also tells us that, under the
conditions of Theorem 2.1,
F G(x) F (x) + G(x) . (1 + w) F (x);
and F G 2 S due to the de…nition of the class S. By conditioning on the value of 1 and then
using Kesten’s bound (2.3) with the above facts, we can derive that, for every > 0 and some
constant C > 0,
I12 (x; T )
1 X
X n X1 X Z T
C(1 + )n+m1 + +mn
F G(xert )
n=1 k=1 mk =N +1 mf1;:::;ngnfkg 2f0;1;:::gn 1 0

P(N (T t) = n 1)P ( 1 2 dt) P Mf1;:::;ng = mf1;:::;ng

. C (1 + w) E (1 + )M 1fM >N g
1
X Z T
n 1
n E (1 + )M (1 + )n F (xert )P (N (T t) = n 1) (dt)
n=1 0
N (T )
C(1 + w)(1 + )E (1 + )M 1fM >N g E (N (T ) + 1) (1 + )E (1 + )M (x; T ):

Since Ee M < 1 for some > 0, we can choose small enough and N large enough such that
E (1 + )M 1fM >N g < ":
0
Further, note that Ee N (T ) < 1 for some 0 > 0 since fN (t); t 0g is a renewal process. By
Cauchy-Schwarz inequality, the can be chosen small enough such that
N (T )
E (N (T ) + 1) (1 + )E (1 + )M < 1:

Hence,
N (T )
I12 (x; T ) . "C (1 + w) (1 + ) E (N (T ) + 1) (1 + )E (1 + )M (x; T )

=: "K2 (x; T );
which, together with (3.12), implies that
(1 "K1 ) (x; T ) . I1 (x; T ) . (1 + "K2 ) (x; T ): (3.13)
Next, we turn to I2 (x; T ). By the arguments similar to those used in handling I12 (x; T ), we have
I2 (x; T )
1 n+m1X
+ +mn
!
X X
P (Xi + Yi ) > xer 1 ; N (T ) = n
n=N +1 mf1;:::;ng 2f0;1;:::gn i=1

P Mf1;:::;ng = mf1;:::;ng
1
X X Z T
C(1 + )n+m1 + +mn
F G(xert )P(N (T t) = n 1)P ( 1 2 dt)
n=N +1 mf1;:::;ng 2f0;1;:::gn 0

P Mf1;:::;ng = mf1;:::;ng
Z T
N (T )+1
. C (1 + w) E (1 + ) E (1 + )M 1fN (T ) Ng F (xert ) (dt):
0

10
 0 0
Since Ee N (T ) < 1 for some > 0, we can choose small enough and N large enough such that
N (T )+1
E (1 + ) E (1 + )M 1fN (T ) Ng < ":

Thus,
I2 (x; T ) . "C (1 + w) (x; T ): (3.14)
Plugging (3.13) and (3.14) into (3.8) and noting the arbitrariness of " complete the proof under
condition (i)
Moreover, under condition (ii), noting that the weight e r( i +Dij ) 1f i +Di T g 1fN (T )=ng of Yij is
obviously not greater than the weight e r i 1fN (T )=ng of Xi , we can still apply Lemma 3.1(i) to get
(3.9) without the terms involving Y and then use the arguments similar to (but simpler than) the
ones shown above to complete the proof.

Lemma 3.6. Let the conditions of Theorem 2.2 hold. We have

0 0 1 1 8
X 1 Mi
X < (x); under (i)
P @ @Xi e r i
+ Yij e r( i +Dij ) A
>xA ; (3.15)
:
i=1 j=1 o (x); under (ii)
where (x) and o (x) are given by (3.6) and (3.7).

Proof. As before, we show the details of the proof under condition (i), and then only highlight the
di¤erent arguments that will be used in the proof under condition (ii).
Under condition (i), we rewrite the probability on the left-hand side of (3.15) as
0 0 1 1 !
1
X 1
X X1
P@ @X i + Yij e rDij
1fMi jg A e r i
> xA =: P (Xi + i ) e r i
>x :
i=1 j=1 i=1

Clearly, f i ; i 1g is a sequence of i.i.d. random variables. Denote by the generic random variable
of f i ; i 1g. Note that Ee M < 1 for some > 0, which implies that P (M j) converges to 0
at least exponentially fast as j ! 1. Thus, for each i 1, every p > 0 and every q > 0, it holds
that
X1 1
X
p 1=q
E e rDij
1fMi jg (P (M j))1=q < 1:
j=1 j=1

Then, applying Lemma 3.2 to i leads to

1
X
rDij rD
P( i > x) P Yij e 1fMi jg > x = EM P Y e >x :
j=1

It follows from Lemma 1 of Chen and Ng (2007) that the distributions of Y e rD and hence of
belong to the class ERV. Since X and are independent and both have the distributions from the
class ERV, using Lemma 3.1(ii) yields that
rD
P (X + > x) F (x) + P ( > x) F (x) + EM P Y e >x ; (3.16)

which implies that the distribution of X + belongs to the class ERV by de…nition. It is easy to
see that fe r i ; i 1g satis…es the moment conditions required in Lemma 3.2. Hence, applying
Lemma 3.2 again gives that
1
! 1
X X
r i
P (Xi + i ) e >x P (Xi + i ) e r i > x (x);
i=1 i=1

11
where the last step is obtained by conditioning on the value of each i and then using (3.16). This
completes the proof under condition (i).
To establish the proof under condition (ii), we just need to go along the same lines of the above
proof, but note that P ( > x) = o(F (x)) due to Lemma 3.3. Then, applying Lemma 3.1(ii) may
give us (3.16) with a re…ned form as P (X + > x) F (x), which su¢ ces to verify the assertion
under condition (ii).

4 Proof of Main Results

Proof of Theorem 2.1. Firstly, let condition (i) hold. Recalling (2.6) and (3.4), we just need to
prove that (x; T ) (x; T ). It is clear that
rt
(x; T ) = P e R (t) < 0 for some 0 < t T R(0) = x :

On the one hand,

0 0 1 1
N (T ) Mi
X X
(x; T ) P@ @Xi e r i
+ Yij e r( i +Dij )
1f i +Dij Tg
A > xA (x; T );
i=1 j=1

where the last step follows from Lemma 3.5. On the other hand, observing the form of (x; T ), it is
PN (T ) P i
easy to check via Lemma 3.5 that the distribution of i=1 Xi e r i + M j=1 Yij e
r( i +Dij ) 1
f i +Dij Tg
is long-tailed for every …xed T under the conditions of Theorem 2.1. Hence, we have
0 0 1 1
N (T )
X Mi
X Z T
(x; T ) P@ @Xi e r i + Yij e r( i +Dij ) 1f i +Dij T g A > x + c e rs dsA
i=1 j=1 0
0 0 1 1
N (T ) Mi
X X
P@ @Xi e r i
+ Yij e r( i +Dij )
1f i +Dij Tg
A > xA
i=1 j=1

(x; T ):

Combining the above two estimates completes the proof under condition (i). Mimicking the above
proof under condition (i), we can prove the assertion under condition (ii).
Proof of Theorem 2.2. Under condition (i), we need to prove that (x) (x). On the one hand,

(x) = P e rt R (t) < 0 for some 0 < t < 1 R(0) = x

0 0 1 1
1
X XMi
P@ @Xi e r i + Yij e r( i +Dij ) 1f i +Dij T g A > xA
i=1 j=1

(x);

where the last step is due to Lemma 3.6. On the other hand, for every " > 0, we have
0 0 1 1
X1 Mi
X c
(x) P@ @X i e r i + Yij e r( i +Dij ) 1f i +Dij T g A > x + A
r
i=1 j=1
c
x+
r
& ((1 + ") x):

12
Since F 2 ERV, G 2 ERV and " may be arbitrarily small, it holds that

(x) & (x);

which completes the proof under condition (i). Finally, using the same arguments as those shown
in the proof under condition (i), we can prove the assertion under condition (ii).

Acknowledgments. The author is very grateful to the two anonymous reviewers for their thorough
reading of the paper and constructive suggestions. This work was supported by the National Natural
Science Foundation of China (Grant Numbers: 11871289, 11931018 and 11911530091).

References
[1] Athreya, K. B.; Ney, P. E. Branching Processes. Springer-Verlag, New York-Heidelberg, 1972.

[2] Bingham, N. H.; Goldie, C. M.; Teugels, J. L. Regular Variation. Cambridge University Press,
Cambridge, 1987.

[3] Chen, Y.; Ng, K. W. The ruin probability of the renewal model with constant interest force and
negatively dependent heavy-tailed claims. Insurance: Mathematics and Economics 40 (2007),
no. 3, 415–423.

[4] Chen, Y.; Yuen, K. C. Sums of pairwise quasi-asymptotically independent random variables
with consistent variation. Stochastic Models 25 (2009), no. 1, 76–89.

[5] Cline, D. B. H.; Samorodnitsky, G. Subexponentiality of the product of independent random

variables. Stochastic Processes and their Applications 49 (1994), no. 1, 75–98.

[6] Embrechts, P.; Klüuppelberg, C.; Mikosch, T. Modelling Extremal Events for Insurance and
Finance. Springer, New York, 1997.

[7] Fu, K.; Li, H. Asymptotic ruin probability of a renewal risk model with dependent by-claims
and stochastic returns. Journal of Computational and Applied Mathematics 306 (2016), 154–
165.

[8] Fu, K.; Qiu, Y.; Wang, A. Estimates for the ruin probability of a time-dependent renewal risk
model with dependent by-claims. Applied Mathematics-A Journal of Chinese Universities 30
(2015), no. 3, 347–360.

[9] Gao, Q.; Zhuang, J.; Huang, Z. Asymptotics for a delay-claim risk model with di¤usion,
dependence structures and constant force of interest. Journal of Computational and Applied
Mathematics 353 (2019), 219–231.

[10] Hao, X.; Tang, Q. A uniform asymptotic estimate for discounted aggregate claims with subex-
ponential tails. Insurance: Mathematics and Economics 43 (2008), no. 1, 116–120.

[11] Li, J. On pairwise quasi-asymptotically independent random variables and their applications.
Statistics and Probability Letters 83 (2013), 2081–2087.

[12] Li, J. On the joint tail behavior of randomly weighted sums of heavy-tailed random variables.
Journal of Multivariate Analysis 164 (2018), 40–53.

[13] McNeil, A. J.; Frey, R.; Embrechts, P. Quantitative Risk Management. Concepts, Techniques
and Tools. Princeton University Press, Princeton, NJ, 2005.

13
[14] Tang, Q.; Yuan, Z. Randomly weighted sums of subexponential random variables with appli-
cation to capital allocation. Extremes 17 (2014), no. 3, 467–493.

[15] Wang, D.; Tang, Q. Tail probabilities of randomly weighted sums of random variables with
dominated variation. Stochastic Models 22 (2006), no. 2, 253–272.

[16] Waters, H. R.; Papatriandafylou, A. Ruin probabilities allowing for delay in claims settlement.
Insurance: Mathematics and Economics 4 (1985), no. 2, 113–122.

[17] Xiao, Y.; Guo, J. The compound binomial risk model with time-correlated claims. Insurance:
Mathematics and Economics 41 (2007), no. 1, 124–133.

[18] Yang, H.; Li, J. On asymptotic …nite-time ruin probability of a renewal risk model with
subexponential main claims and delayed claims. Statistics and Probability Letters 149 (2019),
153–159.

[19] Yuen, K. C.; Guo, J. Ruin probabilities for time-correlated claims in the compound binomial
model. Insurance: Mathematics and Economics 29 (2001), no. 1, 47–57.

[20] Yuen, K. C.; Guo, J.; Ng, K. W. On ultimate ruin in a delayed-claims risk model. Journal of
Applied Probability 42 (2005), no. 1, 163–174.

[21] Zhang, Y.; Shen, X.; Weng, C. Approximation of the tail probability of randomly weighted
sums and applications. Stochastic Processes and their Applications 119 (2009), no. 2, 655–675.

14