Next Article in Journal
Regularity for Quasi-Linear p-Laplacian Type Non-Homogeneous Equations in the Heisenberg Group
Next Article in Special Issue
Testing Multivariate Normality Based on F-Representative Points
Previous Article in Journal
OTNet: A Small Object Detection Algorithm for Video Inspired by Avian Visual System
Previous Article in Special Issue
Representative Points from a Mixture of Two Normal Distributions
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Local Closure under Infinitely Divisible Distribution Roots and Esscher Transform

1
School of Mathematics and Statistics, Changshu Institute of Technology, Suzhou 215000, China
2
School of Mathematical Sciences, Soochow University, Suzhou 215006, China
*
Author to whom correspondence should be addressed.
These authors contributed equally to this work.
Mathematics 2022, 10(21), 4128; https://doi.org/10.3390/math10214128
Submission received: 6 October 2022 / Revised: 30 October 2022 / Accepted: 1 November 2022 / Published: 5 November 2022
(This article belongs to the Special Issue Distribution Theory and Application)

Abstract

:
In this paper, we show that the local distribution class L l o c OS l o c is not closed under infinitely divisible distribution roots, i.e., there is an infinitely divisible distribution which belongs to the class, while the corresponding Lévy distribution does not. Conversely, we give a condition, under which, if an infinitely divisible distribution belongs to the class L l o c OS l o c , then so does the Lévy distribution. Furthermore, we find some sufficient conditions that are more concise and intuitive. Using different methods, we also give a corresponding result for another local distribution class, which is larger than the above class. To prove the above results, we study the local closure under random convolution roots. In particular, we obtain a result on the local closure under the convolution root. In these studies, the Esscher transform of distribution plays a key role, which clarifies the relationship between these local distribution classes and related global distribution classes.

1. Preliminary

In this paper, we study the closure under infinitely divisible distribution (I.I.D.) roots for some local distribution classes, also known simply as the local closure under I.I.D. roots. In other words, we discuss the following problem, if an I.I.D. belongs to a local distribution class, does its corresponding Lévy distribution also belong to this class? These results are closely related to some local distribution classes and Esscher transform of distributions. Thus, in order to better illustrate the main results of this paper, we first introduce the above concepts and their basic properties in this section.
Throughout the paper, unless stated otherwise, all limits are taken as x tends to infinity; for two positive functions f and g, f ( x ) g ( x ) means lim sup f ( x ) / g ( x ) = 1 , f ( x ) g ( x ) means 0 < lim inf f ( x ) / g ( x ) lim sup g ( x ) / f ( x ) < , f ( x ) = o g ( x ) means lim f ( x ) / g ( x ) = 0 ; for a distribution V, let V ¯ = 1 V be the tail distribution of V, V * k be the k-fold convolution of V with itself for all integers k 2 , V * 1 = V and V * 0 be the distribution degenerate at zero; and all distributions are supported on [ 0 , ) .

1.1. Infinitely Divisible Distribution

Let H be an I.D.D. with the Laplace transform
0 e λ y H ( d y ) = exp a λ 0 ( 1 e λ y ) υ ( d y ) ,
where a 0 is a constant, and υ is a Borel measure on ( 0 , ) with the properties μ = υ ( 1 , ) < and 0 min { 1 , y 2 } ν ( d y ) < . Let
F ( x ) = υ ( 0 , x ] 1 { x > 1 } / μ , x ( , )
be the Lévy distribution generated by the measure υ . The distribution H admits the representation H = H 1 H 2 , which is reserved for the convolution of two distributions H 1 and H 2 satisfying
H ¯ 1 ( x ) = O ( e β x ) for each β > 0
and
H 2 ( x ) = e μ k = 0 F * k ( x ) μ k / k ! , x ( , ) .
See, for example, pages 450 and 571 of Feller [1], Embrechts et al. [2] and Chapter 4 of Sato [3].
One of the research topics of I.D.D is the closure under I.D.D. roots for all types of distribution classes. More precisely, we say that a certain distribution class is closed under I.D.D. roots, if an I.D.D. belongs to the class, then its Lévy distribution also belongs to the same one; otherwise, we say that the class is not closed under the I.D.D. roots.
This paper mainly studies the closure of some local distribution classes under the I.D.D. roots, known simply as the local closure under the I.D.D. roots.

1.2. Related Distribution Classes

In this paper, for each 0 < T , we denote
V ( x + Δ T ) = V ( x , x + T ] = V ¯ ( x ) V ¯ ( x + T ) and V ( x + Δ ) = V ¯ ( x ) , x 0 .
For each distribution V and 0 < T < , we set that there is a x 0 = x 0 ( V , T ) 0 such that V ( x + Δ T ) > 0 , x x 0 .
We say that a distribution V belongs to the distribution class L l o c , if for each 0 < T ,
V ( x t + Δ T ) V ( x + Δ T ) for each t > 0 .
We say that a distribution V belongs to the distribution class S l o c , if V belongs to the class L l o c and for each 0 < T ,
V * 2 ( x + Δ T ) 2 V ( x + Δ T ) .
See, for example, Borokov and Borokov [4].
The classes L l o c and S l o c are included in two new distribution classes OS l o c and OL l o c defined by the following conditions that, for each 0 < T ,
C Δ T * ( V , t ) = lim sup V ( x t + Δ T ) / V ( x + Δ T ) < for each t > 0 ;
and for each 0 < T ,
C Δ T * ( V ) = lim sup V * 2 ( x + Δ T ) / V ( x + Δ T ) < ,
respectively.
In the definitions of the above-mentioned local distribution classes, if “for each 0 < T ” is replaced by “for some 0 < T ”, then these classes are successively called local long-tailed distribution class, local subexponential distribution class, O-local long-tailed distribution class and O-local subexponential distribution class, denoted by L Δ T , S Δ T , OL Δ T with indicator C Δ T * ( V , t ) for each 0 < t < and OS Δ T with indicator C Δ T * ( V ) , respectively. The classes L Δ T and S Δ T for some 0 < T were introduced by Asmussen et al. [5]. The class OS Δ T for some 0 < T originates from the work of Wang et al. [6]. Clearly, the inclusion relations L l o c L Δ T and S l o c S Δ T for each 0 < T are proper.
For research on the local distribution classes, in addition to the above-mentioned references, please refer to Wang et al. [7], Wang et al. [8], Denisov et al. [9], Yang et al. [10], Watanabe [11], etc.
In particular, when T = , we get the corresponding global distribution classes L , S , OL with indicator
C * ( V , t ) = C Δ * ( V , t ) = lim sup V ¯ ( x t ) / V ¯ ( x ) < for each t > 0
and OS with indicator
C * ( V ) = C Δ * ( V ) = lim sup V * 2 ¯ ( x ) / V ¯ ( x ) < ,
respectively. The classes L and S were introduced by Chistyakov [12], and the classes OL and OS come from Shimula and Watanabe [13] and Klüppelberg [14], respectively.
Further, Lemma 2 of Chistyakov [12] shows S L . This inclusion relation is proper, see Section 3 of Embrechts and Goldie [15], and so on. However, with respect to the prerequisite that V L Δ T is necessary in the definition of the class S Δ T for some 0 < T < , see Propositions 3.1 and 3.2 of Chen et al. [16]. Another proper inclusion relation OS OL is given by Proposition 2.1 of Shimura and Watanabe [13].
Clearly, the class OL contains the heavy-tailed distribution classes 0 < T L Δ T and the class OS contains the heavy-tailed distribution classes 0 < T S Δ T . Here, a distribution V is called the heavy-tailed distribution, if M ( V , α ) = 0 e α y V ( d y ) = for each α > 0 ; otherwise, it is called the light-tailed distribution. Furthermore, for some γ > 0 , the following light-tailed distribution class L ( γ ) is a subclass of OL , another light-tailed distribution class S ( γ ) is a subclass of OS . Both classes were introduced by Chover et al. [17,18].
A distribution V belongs to the distribution class L ( γ ) for some γ > 0 , if
V ¯ ( x t ) V ¯ ( x ) e γ t for each t > 0 .
A distribution V belongs to the distribution class S ( γ ) for some γ > 0 , if V L ( γ ) , M ( V , γ ) = 0 e γ y V ( d y ) < and
V * 2 ¯ ( x ) 2 M ( V , γ ) V ¯ ( x ) .
Clearly, here M ( V , γ ) 1 . In addition, the prerequisite that V L ( γ ) for some γ > 0 also is necessary in the definition of the class S ( γ ) , because the distribution here is closely related to its local distribution. In fact, if we define two distribution classes L Δ T ( γ ) and S Δ T ( γ ) for some 0 < γ , T < , then we can easily find that L Δ T ( γ ) = L ( γ ) and S Δ T ( γ ) = S ( γ ) .
In the definition of the class L ( γ ) , if V is a lattice, then x and t should be restricted to values of the lattice span of V, see Bertoin and Doney [19].
In addition, we might also set L = L ( 0 ) and S = S ( 0 ) .
There are many research results on the distribution classes mentioned above, see Foss et al. [20], Wang [21] and the references therein.

1.3. Esscher Transform

Now, we use the Esscher transform to show the relationship between some heavy-tailed local distribution and the corresponding light-tailed global distribution.
For any distribution V and γ 0 , by M ( V , γ ) min { 1 , e γ x V ( x ) } , x 0 , we know that M ( V , γ ) > 0 . Further, if M ( V , γ ) < , then we define a distribution V γ such that
V γ ( x ) = 0 x e γ y V ( d y ) 1 [ 0 , ) ( x ) / M ( V , γ ) , x ( , ) ,
which is called the Esscher transform (or the exponential tilting) of V. Clearly, for γ > 0 , we have
0 < M ( V , γ ) < 1 , V = ( V γ ) γ and M ( V , γ ) M ( V γ , γ ) = 1 ,
and for all k 1 ,
( V * k ) γ = ( V γ ) * k = V γ * k , ( V * k ) γ = ( V γ ) * k = V γ * k and M ( V * k , γ ) = M k ( V , γ ) ,
see Teugels [22], Veraverbeke [23] and Embrechts and Goldie [24] for technical details.
Further, for some 0 < T < and γ > 0 , Definitions 1.1 and 1.2 of Wang and Wang [25] define four global distribution classes as follows:
TL Δ T ( γ ) = V : M ( V , γ ) < and V γ L Δ T ,
TS Δ T ( γ ) = V : M ( V , γ ) < and V γ S Δ T ;
TL l o c ( γ ) = V : M ( V , γ ) < and V γ L l o c
and
TS l o c ( γ ) = V : M ( V , γ ) < and V γ S l o c .
The following proposition reveals the important role of the Esscher transform for the study of local distribution classes, see Propositions 2.1 and 2.2 of Wang and Wang [25]. On the contrary, this result also shows that some local distribution classes give new vitality to the Esscher transform.
Proposition 1.
( i ) For some 0 < T < and γ > 0 , a distribution V L Δ T (or S Δ T ) V γ TL Δ T ( γ ) (or TS Δ T ( γ ) ) .
( i i ) A distribution V L l o c (or S l o c ) V γ L ( γ ) (or S ( γ ) ) , that is TL l o c ( γ ) = L ( γ ) ( or TS l o c ( γ ) = S ( γ ) ) . Furthermore, each of them implies that, for each 0 < T <
V ( x + Δ T ) γ T e γ x V γ ¯ ( x ) / M ( V γ , γ ) = M ( V , γ ) γ T e γ x V γ ¯ ( x ) .
More results of the Esscher transform can be found in the above references and the others therein.
The paper is organized as follows. In Section 2, we present the main results for Theorems 1–3 related to local closure under I.I.D. roots. In Section 3, we prove the above results. To this end, we study the local closure under random convolution roots. Then in Section 4, we show that the condition (10) of Theorem 3 can be replaced by a more concise and intuitive condition (11). Finally, in Section 5, we briefly introduce some applications of the obtained results and further research problems. As an application of Theorem 2, we give a positive result on the local closure under the convolution root, which represents the local version of common Embrechts and Goldie conjecture.

2. Main Results

Before giving the main results of this paper, we recall some existing results on closure under I.I.D. roots.
For the global distribution classes, the class S ( γ ) is closed under I.D.D. roots, see Embrechts et al. [2] for the case γ = 0 , Sgibnev [26], Pakes [27] and Watanabe [28] for the case γ > 0 . Recently, Cui et al. [29] proved that the class L ( γ ) OS for some γ 0 is closed under the roots with some restrictive condition.
However, for some global distribution classes without special restrictions, there were some negative results, i.e., there exists an I.D.D. H belonging to some class, while its Lévy distribution F does not belong to the same class; see Theorem 1.1 (iii) of Shimura and Watanabe [13] for the class OS , Theorem 1.2 (3) of Xu et al. [30] for the class L OS and L OS and Theorem 1.1 of Xu et al. [31] for the class L ( γ ) OS with some γ > 0 .
As previously mentioned, this paper mainly studies the closure of some local distribution classes under the I.D.D. roots. Clearly, if a distribution V L Δ T for some 0 < T < , then V ( x + Δ T ) = o V ( x ) . Therefore, the study of local distribution cannot be replaced by that of global distribution.
One of the difficulties in the study of local distributions is the loss of their almost monotonic decreasing property. Corollary 3.1 of Jiang et al. [32] shows that some local distributions in the class S l o c and the class L l o c S l o c are not even close to decreasing. Therefore, the study of local distribution is definitely more challenging than that of global distribution. Furthermore, we find hardly any existing results regarding local closure under I.I.D. roots.
Now, we first give a negative conclusion for the class L l o c OS l o c .
Theorem 1.
The class L l o c OS l o c is not closed under I.D.D. roots.
Next, we give two positive conclusions for the class L l o c OS l o c and TL Δ T 0 ( γ ) OS Δ T 0 with some 0 < γ , T 0 < , respectively.
Theorem 2.
Let H be an I.D.D. with the Lévy distribution F. Assume that H L l o c OS l o c , and for all k 1 ,
lim inf F γ * k ¯ ( x t ) / F γ * k ¯ ( x ) e γ t f o r e a c h t > 0 .
Then the following two conclusions hold.
( i ) H 2 L l o c OS l o c and H 2 ( x + Δ T ) H ( x + Δ T ) for each 0 < T .
( i i ) There exists an integer l 0 1 such that F * n L l o c OS l o c for all n l 0 and F * n L l o c OS l o c for all 1 n l 0 1 . In particular, if F OS l o c , then F * n L l o c OS l o c for all n 1 .
Remark 1. ( i ) According to Corollary 1.1 of Cui et al. [29], the condition (8) can be implied by some more concise and convenient conditions that
F γ OL , lim F γ ¯ ( x ) C * ( F γ , x ) = 0 a n d ( 8 ) h o l d s f o r k = 1 .
Therefore, all conclusions of Theorem 2 hold under the conditions (9) and H L l o c OS l o c . Some related examples can be found in Corollary 1.2 and Example 4.1 of Cui et al. [29].
( i i ) In the proof of Theorem 1, we can find that there exists an I.D.D. H with Lévy distribution F such that l 0 = 2 . This fact shows that there are many distributions F that satisfy condition (8), but which do not belong to the class L ( γ ) .
Clearly, the local distribution class L Δ T 0 OS Δ T 0 for some 0 < T 0 is larger than the class L l o c OS l o c . Therefore, it is natural to investigate the corresponding result for the former. To this end, we first consider its corresponding light-tailed global distribution class TL Δ T 0 ( γ ) OS Δ T 0 for some 0 < γ , T 0 < , which is larger than the class L ( γ ) OS . We will find that the research method of the following result is different from that of Theorem 2.
Theorem 3.
Let H be an I.D.D. with the Lévy distribution F. For some 0 < γ , T 0 < , assume that H TL Δ T 0 ( γ ) OS Δ T 0 and for all k 1 ,
lim inf F γ * k ( x t + Δ T 0 ) / F γ * k ( x + Δ T 0 ) 1 f o r e a c h t > 0 .
Then the following two conclusions hold.
( i ) H 2 TL Δ T 0 ( γ ) OS Δ T 0 and H 2 ( x + Δ T 0 ) H ( x + Δ T 0 ) .
( i i ) There is an integer l 0 1 such that F * n TL Δ T 0 ( γ ) OS Δ T 0 for all n l 0 and F * n TL Δ T 0 ( γ ) OS Δ T 0 for all 1 n l 0 1 . In particular, if F OS Δ T 0 , then F * n TL Δ T 0 ( γ ) OS Δ T 0 for all n 1 .
Remark 2.
The condition (10) can also be replaced by the following more concise and convenient conditions:
F γ OL Δ T 0 , lim F γ ( x + Δ T 0 ) C Δ T 0 * ( F γ , x ) = 0 a n d ( 10 ) h o l d s f o r k = 1 .
See Theorem 6 with V = F γ below.

3. The Proofs of Theorems 1–3

3.1. Proof of Theorem 1

Let F ( 0 ) be a heavy-tailed distribution such that
F ( 0 ) ¯ ( x ) = 1 ( , a 0 ) ( x ) + C n = 0 i = n 1 a i α x a n a n α + 1 1 [ a n , 2 a n ) ( x ) + i = n + 1 1 a i α 1 [ 2 a n , a n + 1 ) ( x )
with the density f ( 0 ) ( x ) = C n = 0 a n α 1 1 [ a n , 2 a n ) ( x ) for all x, where
α 3 2 , 5 + 1 2 , a n = a r n for r = 1 + 1 α , some a > 8 α and all n 1 , and C = n = 0 a n α 1 .
Let F 1 ( 0 ) be the class comprising the above distributions F ( 0 ) defined by (12). Further, for some γ > 0 and distribution F ( 0 ) F 1 ( 0 ) , define the light-tailed distribution F ( γ ) in the form
F ( γ ) ¯ ( x ) = 1 ( , 0 ) ( x ) + e γ x F ( 0 ) ¯ ( x ) 1 [ 0 , ) ( x )
with its density f ( γ ) for all x. Then we can construct a new distribution class
F 1 ( γ ) = { F ( γ ) defined by ( 13 ) : F ( 0 ) F 1 ( 0 ) } .
See the proof of Theorem 1 of Xu et al. [31].
Let H = H 1 H 2 is an I.D.D. with Lévy distribution F ( γ ) F 1 ( γ ) for some γ > 0 . Then Proposition 1 and Theorem 1 of Xu et al. [31] show that, H , H 2 and F ( γ ) * k for all k 2 belong to the class L ( γ ) OS S ( γ ) , while F ( γ ) with M ( F ( γ ) , γ ) < belongs to the class OL L ( γ ) OS .
Because M ( F ( γ ) , γ ) < , M ( H , γ ) < , then H γ = H 1 , γ H 2 , γ , as the Esscher transform of H, is defined and is I.D.D. with Lévy distribution F ( γ ) , γ . To reveal the properties of H γ and F ( γ ) , γ , we need the following result.
Lemma 1.
For some 0 < γ , T < , V γ OS Δ T V OS Δ T . Thus, V γ OS l o c V OS l o c . Further, if V γ L ( γ ) , then V γ OS V γ OS l o c . Therefore,
V γ L ( γ ) OS V L l o c OS l o c .
Proof. 
We now prove the first conclusion. From (2.4) of Wang and Wang [25], we have
V γ ( x + Δ T ) = M ( V γ , γ ) e γ x V ( x + Δ T ) γ 0 T e γ y V ( x + y , x + T ] d y , x 0 .
Further, we obtain the following inequality,
e γ T V ( x + Δ T ) e γ x V γ ( x + Δ T ) / M ( V γ , γ ) V ( x + Δ T ) , x 0 .
If V OS Δ T , then according to Radon–Nikodym Theorem, by (15) and (4), we have
V γ * 2 ( x + Δ T ) = 0 x V γ ( x y + Δ T ) V γ ( d y ) + x x + T V γ ( 0 , x y + T ] V γ ( d y ) e γ x M ( V γ , γ ) 0 x V ( x y + Δ T ) V ( d y ) / M ( V , γ ) + V γ ( x + Δ T ) e γ x M ( V γ , γ ) V * 2 ( x + Δ T ) / M ( V , γ ) + V γ ( x + Δ T ) 2 C Δ T * ( V ) M ( V γ , γ ) e γ x V ( x + Δ T ) / M ( V , γ ) + V γ ( x + Δ T ) 2 C Δ T * ( V ) M ( V γ , γ ) e γ T / M ( V , γ ) + 1 V γ ( x + Δ T ) for large enough x > 0 ,
that is V γ OS Δ T . Conversely, if V γ OS Δ T , then we also get V OS Δ T by the same approach.
The second conclusion comes from the arbitrariness of T.
If V γ L ( γ ) , then V γ * 2 L ( γ ) . Thus, for each 0 < T , by
V γ * k ( x + Δ T ) ( 1 e γ T ) V γ * k ¯ ( x ) , k = 1 , 2 ,
the third conclusion holds.
Proposition 1 and the third conclusion imply the final conclusion. □
Now, we continue to prove the theorem. According to Lemma 1 and Proposition 1, by H L ( γ ) OS S ( γ ) and F ( γ ) OL L ( γ ) OS , we know that H γ L l o c OS l o c S l o c , while F ( γ ) , γ OL l o c L l o c OS l o c . Therefore, the class L l o c OS l o c is not closed under I.D.D. roots.

3.2. Proof of Theorem 2

To prove this theorem, we give two preliminary results. Firstly, we consider the closure under random convolution roots for the distribution class L l o c OS l o c . Clearly, this result and the following Theorem 5 not only play a key role in the proof of Theorems 2 and 3, but also have their own independent value.
Let V be a distribution and let τ be a nonnegative integer-valued random variable with masses p k = P ( τ = k ) for all nonnegative integers k satisfying k = 0 p k = 1 . Denoted by V * τ is the random convolution or compound convolution generated by V and τ , i.e.,
V * τ = k = 0 p k V * k .
Let m = sup { k : p k > 0 } . In this paper, we consider the following two cases:
Case 1 : p k > 0 for all k 1 ; Case 2 : 1 m < and p k > 0 for all 1 k m .
Theorem 4.
Assume that for any 0 < ε < 1 and some 0 < T 0 < , there exists an integer n 0 = n 0 ( V , ε , τ , T 0 ) 1 such that
k = n 0 + 1 p k V * ( k 1 ) ( x + Δ T 0 ) ε V * τ ( x + Δ T 0 ) , x 0 ,
and for each k 1 in Case 1 or 1 k m in Case 2,
lim inf V γ * k ¯ ( x t ) / V γ * k ¯ ( x ) e γ t f o r e a c h t > 0 .
If V * τ L l o c OS l o c , then for the above two cases, there exists an integer l 0 1 in Case 1 or 1 l 0 m in Case 2 such that V * n L l o c OS l o c for all l 0 n < m and V * n L l o c OS l o c for all 1 n l 0 1 . In particular, if V OS l o c , then V * n L l o c OS l o c for all n 1 in Case 1 or 1 n m in Case 2.
Proof. 
We first prove the theorem for Case 1 that m = .
In Lemma 1, we replace V with V * τ . Then by V * τ L l o c OS l o c , we know that
( V * τ ) γ L ( γ ) OS .
In addition,
0 < M ( V * τ , γ ) = k = 0 p k M k ( V , γ ) = E M τ ( V , γ ) < 1
and
( V * τ ) γ ¯ ( x ) = k = 1 p k M k ( V , γ ) M ( V * τ , γ ) V γ * k ¯ ( x ) = k = 1 q k V γ * k ¯ ( x ) = ( V γ ) * σ ¯ ( x ) , x 0 ,
where σ is a random variable such that P ( σ = k ) = q k for all nonnegative integers k satisfying k = 0 q k = 1 .
For any 0 < ε < 1 , we denote ε 0 = ε e γ T 0 . By (15), (18) and (16) replaced ε with ε 0 , according to Fubini Theorem, for the corresponding n 0 = n 0 ( V , ε 0 , τ , T 0 ) large enough, we have
k = n 0 + 1 q k V γ * ( k 1 ) ¯ ( x ) = k = n 0 + 1 q k m = 0 V γ * ( k 1 ) ( x + m T 0 + Δ T 0 ) k = n 0 + 1 p k M k 1 ( V , γ ) M k 1 ( V γ , γ ) m = 0 e γ ( x + m T 0 ) V * ( k 1 ) ( x + m T 0 + Δ T 0 ) / M ( V * τ , γ ) = k = n 0 + 1 p k m = 0 e γ ( x + m T 0 ) V * ( k 1 ) ( x + m T 0 + Δ T 0 ) / M ( V * τ , γ ) = m = 0 e γ ( x + m T 0 ) k = n 0 + 1 p k V * ( k 1 ) ( x + m T 0 + Δ T 0 ) / M ( V * τ , γ ) ε 0 m = 0 e γ ( x + m T 0 ) V * τ ( x + m T 0 + Δ T 0 ) / M ( V * τ , γ ) ε 0 e γ T 0 m = 0 ( V * τ ) γ ( x + m T 0 + Δ T 0 ) = ε ( V * τ ) γ ¯ ( x ) = ε ( V γ ) * σ ¯ ( x ) , x 0 .
Since ( V * τ ) γ L ( γ ) OS , according to Theorem 2.1 with γ > 0 of Cui et al. [29], by (19) and (17), we have V γ * n L ( γ ) OS for all n n 0 . Thus, according to Lemma 1, V * n L l o c OS l o c for all n n 0 .
Let l 0 = min { n : V * n L l o c OS l o c } . Then 1 l 0 n 0 . According to Lemma 1, by V * l 0 L l o c OS l o c , we know that V γ * l 0 L ( γ ) OS . Furthermore, according to Theorem 3 of Embrechts and Goldie [15] and Proposition 2.6 of Shimura and Watanabe [13], we have V γ * n L ( γ ) OS for all n l 0 . Therefore, V * n L l o c OS l o c for all n l 0 after using Lemma 1 again.
Similarly, we can prove V * n L l o c OS l o c for all 1 n l 0 1 .
In particular, if V OS l o c , then V γ OS . Thus, according to Theorem 2.1 with γ > 0 of Cui et al. [29], we have V γ L ( γ ) , which implies V L l o c . Therefore, l 0 = 1 , that is V * n L l o c OS l o c for all n 1 .
Next, we prove the theorem for the Case 2 that 1 m < and p m > 0 .
Because
( V * τ ) γ ¯ ( x ) q m V γ * m ¯ ( x ) q m V γ * k ¯ ( x ) , 1 k m 1 ,
( V * τ ) γ ¯ ( x ) V γ * m ¯ ( x ) . Then by ( V * τ ) γ OS , we immediately get V γ * m OS . Consequently, there is an integer l 0 = min { 1 n n 0 : V γ * n OS } such that 1 l 0 m and V γ * l 0 OS . According to Proposition 2.6 of Shimura and Watanabe [13], V γ * n OS and ( V * τ ) γ ¯ ( x ) V γ * n ¯ ( x ) for all l 0 n m . Thus, for each n l 0 , there is a constant D n = D n ( V , τ ) > 0 such that
lim sup ( V * τ ) γ ¯ ( x ) / V γ * n ¯ ( x ) = D n < .
Further, we prove V γ * n L ( γ ) for each l 0 n m . Since ( V * τ ) γ L ( γ ) , for any 0 < ε < 1 and each t > 0 , there is a constant x 1 = x 1 ( V γ , τ , ε , t ) > t such that, for all x > x 1 ,
ε ( V * τ ) γ ¯ ( x ) ( V * τ ) γ ¯ ( x t ) e γ t ( V * τ ) γ ¯ ( x ) = 1 k n m + k = n p k V γ * k ¯ ( x t ) e γ t V γ * k ¯ ( x ) ε e γ t 1 k n n 0 p k V γ * k ¯ ( x ) + p n V γ * n ¯ ( x t ) e γ t V γ * n ¯ ( x ) p n V γ * n ¯ ( x t ) e γ t V γ * n ¯ ( x ) ε e γ t ( V * τ ) γ ¯ ( x ) ,
which implies that for all x > x 1 ,
V γ * n ¯ ( x t ) e γ t V γ * n ¯ ( x ) + ( 1 + e γ t ) ε ( V * τ ) γ ¯ ( x ) / p n .
Hence,
lim sup V γ * n ¯ ( x t ) / V γ * n ¯ ( x ) e γ t + ( 1 + 2 e γ t ) ε D n / p n .
Clearly, the fixed integer n is independent of ε . Thus, combined with the arbitrariness of ε , (20) and (17) lead to V γ * n L ( γ ) .
In particular, if V γ OS , then by the same method, we can get V γ * n L ( O ) OS for all 1 n m . □
Secondly, we consider the closure under convolution roots for the distribution class L l o c OS l o c .
Lemma 2.
Let G 1 be a distribution, G 2 = V * τ as above and G = G 1 * G 2 . Assume that for any 0 < ε < 1 , there exists an integer n 0 = n 0 ( V , ε , τ ) 1 such that
k = n 0 + 1 p k V γ * ( k 1 ) ¯ ( x ) ε ( V * τ ) γ ¯ ( x ) , x 0 .
Further, suppose that (17) is satisfied for all k 1 and
G 1 , γ ¯ ( x ) = o G 2 , γ ¯ ( x )
If G L l o c OS l o c , then
G 2 L l o c OS l o c and G 2 ( x + Δ T 0 ) G ( x + Δ T 0 ) .
Proof. 
According to Lemma 1, by G L l o c OS l o c , we know that
G γ = G 1 , γ * G 2 , γ L ( γ ) OS .
Thus, according to Lemma 3.1 of Cui et al. [29], by (17) for all k 1 , (21) for any given 0 < ε < 1 and (22), we have
G 2 , γ L ( γ ) OS and G γ ¯ ( x ) M ( G 1 , γ , γ ) G 2 , γ ¯ ( x ) .
Therefore, according to Lemma 1 and Proposition 1, by (7), we can prove the lemma. □
Now, we prove Theorem 2.
( i ) Firstly, we prove
H 1 , γ ¯ ( x ) = o H 2 , γ ¯ ( x ) .
To the end, we denote
H γ ¯ ( x ) = H γ ¯ ( ln e x ) = H γ ¯ ( ln y ) = f γ ( y ) .
According to Lemma 1, by H L l o c OS l o c , we have H γ L ( γ ) OS . Thus, f γ ( · ) is a regular variation function with index γ , which implies
e β x H γ ¯ ( x ) for each β > γ .
By H 1 ¯ ( x ) = O e β x for each β > 0 , we have
e β x H 1 , γ ¯ ( x ) e β x H 1 ¯ ( x ) / M ( H 1 , γ ) 0 for each β > 0 .
For i = 1 , 2 , let X i be a random variable with distribution H i , γ . Then
H γ ¯ ( x ) = H 1 , γ * H 2 , γ ¯ ( x ) P max { X 1 , X 2 } > x / 2 H 1 , γ ¯ ( x / 2 ) + H 2 , γ ¯ ( x / 2 ) .
Thus, by (24) and (25), we know that
e β x H 2 , γ ¯ ( x ) = e ( 2 1 β ) 2 x H 2 , γ ¯ ( 2 x / 2 ) for each β > 2 γ .
Combining with (25) and (26), we know that (23) holds.
Secondly, by (18), according to Proposition 6.1 of Watanabe and Yamamuro [33], we have
q k = p k M k ( F γ , γ ) / M ( H 2 , γ ) = e μ μ k M k ( F γ , γ ) / M ( H 2 , γ ) k ! for all k 0 .
Thus, for any 0 < ε < 1 , there exists an integer n 0 = n 0 ( F γ , H 2 , γ , ε ) 1 such that
k = n 0 + 1 q k F γ * ( k 1 ) ¯ ( x ) ε ( F * τ ) γ ¯ ( x ) = ε H 2 , γ ¯ ( x ) , x 0 .
Finally, according to Lemma 2 replaced G i with H i , i = 1 , 2 , combining with (8), (23) and (27), by H L l o c OS l o c , we know that
H 2 L l o c OS l o c and H 2 ( x + Δ T 0 ) H ( x + Δ T 0 ) .
( i i ) In Theorem 4, we take V = F and V * τ = H 2 . Clearly, (16) holds for each distribution V = F , any 0 < ε < 1 and some n 0 1 , see, for example, Watanabe and Yamamuro [33]. Further, according to Theorem 4, by H 2 L l o c OS l o c , combined with (8) and (16), we obtain all the conclusions.

3.3. Proof of Theorem 3

In order to prove the theorem, we need the following two results. The first result is the local version of the half of Lemma 2.1 of Cui et al. [29].
Lemma 3.
Let V * τ be a random convolution defined as above.
( i ) If p k 1 p k > 0 for all k 2 , then the following proposition ( B ) implies the proposition ( A ) for some 0 < T < .
( A ) For any 0 < ε < 1 , there exists an integer n 0 = n 0 ( V , τ , ε , T ) 1 such that
k = n 0 + 1 p k V * k ( x + Δ T ) ε V * τ ( x + Δ T ) , x 0 .
( B ) For any 0 < ε < 1 , there exists an integer n 0 = n 0 ( V , τ , ε , T ) 1 such that (16) holds.
( i i ) If V * τ OS Δ T for some 0 < T < with p 1 > 0 , then the proposition ( B ) implies the proposition ( A ) replaced x 0 by x x 1 for some x 1 x 0 .
Remark 3. ( i ) In particular, if τ obeys a Poisson distribution, then for any 0 < ε < 1 , (16) holds for some n 0 1 . Further, because p k 1 p k > 0 for all k 2 , (28) holds for the same ε and n 0 .
( i i ) The condition 0 < p k p k 1 for all k 2 can be relaxed to the condition that 0 < p k C p k 1 for some C > 0 and all k 2 .
Proof. 
( i ) If (16) holds, then by p k 1 p k > 0 for all k 2 , we know that for any n 0 1 ,
k = n 0 + 1 p k V * k ( x + Δ T ) k = n 0 + 1 p k 1 V * k ( x + Δ T ) , x 0 .
Therefore, (28) is implied by (16).
( i i ) Clearly, we only need to prove the lemma for Case 1. Because V * τ OS Δ T for some 0 < T < , there exists a constant x 1 = x 1 ( V , τ , T ) x 0 such that
D * ( V * τ , T ) = sup x x 1 ( V * τ ) * 2 ( x + Δ T ) / V * τ ( x + Δ T ) < .
For any 0 < ε < 1 , we take
ε 0 = p 1 ε / 1 + D * ( V * τ , T ) ,
then 0 < ε 0 < 1 .
For the above ε 0 , according to proposition ( B ) , by m = , there exists an integer n 0 = n 0 ( V , τ , ε 0 , T ) 1 such that 0 < a n 0 = k = n 0 + 1 p k < ε 0 and (16) holds, in which ε is replaced with ε 0 . Then k = n 0 + 1 p k V * ( k 1 ) / a n 0 can be considered as a distribution. Therefore, by p 1 > 0 and V * τ ( x + Δ T ) p 1 V ( x + Δ T ) for all x 0 , we have
k = n 0 + 1 p k V * k ( x + Δ T ) = a n 0 V * k = n 0 + 1 p k V * ( k 1 ) / a n 0 ( x + Δ T ) a n 0 0 x k = n 0 + 1 p k V * ( k 1 ) / a n 0 ( x y + Δ T ) V ( d y ) + a n 0 V ( x + Δ T ) ε 0 0 x V * τ ( x y + Δ T ) V * τ ( d y ) + V * τ ( x + Δ T ) / p 1 ε 0 V * 2 τ ( x + Δ T ) + V * τ ( x + Δ T ) / p 1 ε 0 1 + D * ( V * τ , T ) V * τ ( x + Δ T ) / p 1 = ε V * τ ( x + Δ T ) , x x 1 ,
that is (28) holds for any 0 < ε < 1 , all x x 1 and some n 0 1 . □
Theorem 5.
Assume that V * τ TL Δ T 0 ( γ ) OS Δ T 0 for some 0 < γ , T 0 < with p k > 0 for all k 1 in Case 1 or 1 k m in Case 2. If for any 0 < ε < 1 , there exists an integer n 0 = n 0 ( V , τ , ε , T 0 ) 1 such that (16) holds, and for each the above k,
lim inf V γ * k ( x t + Δ T 0 ) / V γ * k ( x + Δ T 0 ) 1 f o r e a c h t > 0 ,
then there exists an integer l 0 1 in Case 1 or 1 l 0 m in Case 2 such that V * n TL Δ T 0 OS Δ T 0 for all n l 0 in Case 1 or l 0 n m in Case 2 and V * n TL Δ T 0 OS Δ T 0 for all 1 n l 0 1 . In particular, if V OS Δ T 0 , then V * n TL Δ T 0 OS Δ T 0 for all n 1 in Case 1 or 1 n m in Case 2.
Proof. 
For case 1, we first prove V * n OS Δ T 0 for all n n 0 , where n 0 fixed in (16). Because V * τ TL Δ T 0 ( γ ) , M ( V * τ , γ ) < . Thus, M ( V * n , γ ) < for all n 1 . By (14), it holds that,
V * θ ( x + Δ T 0 ) = M ( V * θ , γ ) e γ x ( V * θ ) γ ( x + Δ T 0 ) γ 0 T 0 ( V * θ ) γ ( x + y , x + T 0 ] e γ y d y , x 0 ,
where θ = k for each k 1 or θ = τ . Thus, similar to (15), we have
e γ T 0 ( V * θ ) γ ( x + Δ T 0 ) e γ x V * θ ( x + Δ T 0 ) / M ( V * θ , γ ) ( V * θ ) γ ( x + Δ T 0 ) , x 0 .
When θ = τ , just as (18), we denote
( V * τ ) γ ( x + Δ T 0 ) = k = 1 q k V γ * k ( x + Δ T 0 ) = ( V γ ) * σ ( x + Δ T 0 ) , x 0 ,
where q k = p k M k ( V , γ ) / M ( V * τ , γ ) , k 1 . According to Proposition 1 and Lemma 1, since V * τ TL Δ T 0 ( γ ) OS Δ T 0 , so ( V * τ ) γ L Δ T 0 OS Δ T 0 . Furthermore, according to Lemma 3, by (31), (28) with x 1 x 0 and (32), for 0 < ε < 1 and n 0 in (16), we have
k = 1 n 0 q k V γ * k ( x + Δ T 0 ) e γ x k = 1 n 0 p k V * k ( x + Δ T 0 ) / M ( V * τ , γ ) ( 1 ε ) e γ x V * τ ( x + Δ T 0 ) / M ( V * τ , γ ) ( 1 ε ) e γ T 0 ( V * τ ) γ ( x + Δ T 0 ) , x x 1 .
Further, for each n 1 , using Fatou lemma, by (29), we have
lim inf V γ * ( n + 1 ) ( x + Δ T 0 ) V γ * n ( x + Δ T 0 ) 0 lim inf V γ * n ( x y + Δ T 0 ) V γ * n ( x + Δ T 0 ) 1 [ 0 , x ] ( y ) V γ ( d y ) 1 .
Combining with (33), (34) and ( V * τ ) γ OS Δ T 0 , we know that
V γ * n ( x + Δ T 0 ) ( V * τ ) γ ( x + Δ T 0 ) and V γ * n OS Δ T 0 for all n n 0 .
Using Lemma 1 again, by (31), we have
V * n ( x + Δ T 0 ) V * τ ( x + Δ T 0 ) .
Therefore, by V * τ OS Δ T 0 , we know that V * n OS Δ T 0 for all n n 0 .
Next, we prove that V * n TL Δ T 0 ( γ ) for each n n 0 . According to Lemma 3 ( i i ) , by (31), (16) and (32), for the above 0 < ε < 1 , n 0 , x 1 x 0 and each n n 0 , there exists an integer m 0 = m 0 ( F , τ , ε , T 0 , γ ) n such that
k = m 0 + 1 q k V γ * k ( x + Δ T 0 ) e γ ( T 0 + x ) k = m 0 + 1 p k V * k ( x + Δ T 0 ) / M ( V * τ , γ ) e γ ( T 0 + x ) ε e γ T 0 V * τ ( x + Δ T 0 ) / M ( V * τ , γ ) ε ( V * τ ) γ ( x + Δ T 0 ) = ε ( V γ ) * σ ( x + Δ T 0 ) for all x x 1 .
Further, by ( V * τ ) γ L Δ T 0 OS Δ T 0 , (29) and (36), for each t > 0 , there exists a constant x 2 = x 2 ( V , τ , ε , t , m 0 , γ ) x 1 such that, for all x x 2 ,
ε ( V * τ ) γ ( x + Δ T 0 ) ( V * τ ) γ ( x t + Δ T 0 ) ( V * τ ) γ ( x + Δ T 0 ) = 1 k n m 0 + k = n + k m 0 + 1 q k V γ * k ( x t + Δ T 0 ) V γ * k ( x + Δ T 0 ) ε 1 k m 0 q k V γ * k ( x + Δ T 0 ) + q n V γ * n ( x t + Δ T 0 ) V γ * n ( x + Δ T 0 ) ε ( V * τ ) γ ( x + Δ T 0 ) ,
which implies that
V γ * n ( x t + Δ T 0 ) V γ * n ( x + Δ T 0 ) + 3 ε ( V * τ ) γ ( x + Δ T 0 ) / q n , x x 2 .
Hence, by ( V * τ ) γ ( x + Δ T 0 ) V γ * n ( x + Δ T 0 ) and the arbitrariness of ε , we can get
lim sup V γ * n ( x t + Δ T 0 ) / V γ * n ( x + Δ T 0 ) 1 .
Combined with (29) and (37), V γ * n L Δ T 0 . Therefore, V * n TL Δ T 0 ( γ ) , for all n n 0 .
Similar to the proof of Theorem 4, the theorem can be proved.
For Case 2, by (34), we have V γ * m ( x + Δ T 0 ) ( V * τ ) γ ( x + Δ T 0 ) . Then, it is easy to get that V γ * m OS Δ T 0 .
Next, we prove that V * m TL Δ T 0 ( γ ) . For any 0 < ε < 1 and x large enough, by (29) for 1 k m , we can get
ε ( V * τ ) γ ( x + Δ T 0 ) ( V * τ ) γ ( x t + Δ T 0 ) ( V * τ ) γ ( x + Δ T 0 ) = 1 k < m + k = m q k V γ * k ( x t + Δ T 0 ) V γ * k ( x + Δ T 0 ) ε 1 k < m q k V γ * k ( x + Δ T 0 ) + q m V γ * m ( x t + Δ T 0 ) V γ * m ( x + Δ T 0 ) for each t > 0 .
After the same simplification, we have
V γ * m ( x t + Δ T 0 ) V γ * m ( x + Δ T 0 ) + 2 ε ( V * τ ) γ ( x + Δ T 0 ) / q m for each t > 0 .
Hence, we can obtain the same conclusion as (37) for n = m which implies V * m TL Δ T 0 ( γ ) .
We omit the proof of the remaining conclusion, which is similar to that of Theorem 4. □
Now, we prove Theorem 3.
( i ) Firstly, we prove that
H 1 , γ ( x + Δ T 0 ) = o H 2 , γ ( x + Δ T 0 ) .
Its proof is slightly more difficult than that of (23). For this, we denote
H γ ( x + Δ T 0 ) = H γ ( ln e x + Δ T 0 ) = H γ ( ln y + Δ T 0 ) = f γ ( y ) , x 0 .
According to Lemma 1, by H TL Δ T 0 ( γ ) OS Δ T 0 , we have H γ L Δ T 0 OS Δ T 0 . Thus, f γ ( · ) is a regular variation function with index 0, which implies
e β x H γ ( x + Δ T 0 ) for each β > 0 .
By H 1 ¯ ( x ) = O e β x for each β > 0 and (15) with V = H 1 , γ and T = T 0 , we have
e β x H 1 , γ ( x + Δ T 0 ) e γ T 0 M 1 ( H 1 , γ ) e ( β + γ ) x H 1 ( x + Δ T 0 ) 0 for each β > 0 .
Then by (39) and (40), we know that
H 1 , γ ( x + Δ T 0 ) = o H γ ( x + Δ T 0 ) .
Furthermore, by (10), for each pair m , k 1 , we have
lim inf F γ * k ( x j T 0 + Δ T 0 ) / F γ * k ( x + Δ T 0 ) 1 for each 1 j m .
In addition, there exists an integer n 1 = n 1 ( H 1 , γ , T 0 ) large enough such that H 1 , γ ( 0 , n 1 T 0 ] > 0 . Then by (41) and H γ L Δ T 0 OS Δ T 0 , for any
0 < ε < H 1 , γ ( 0 , n 1 T 0 ] / 2 ( n 1 + 1 ) C Δ T 0 * ( H γ ) ,
there exists an integer m 1 = m 1 ( H 1 , H 2 , ε , T 0 , γ ) and a constant x 3 x 2 such that
H 1 , γ ( x + Δ T 0 ) < ε H γ ( x + Δ T 0 ) , x m 1 T 0 ,
j = 0 n 1 H γ * 2 ( x + j T 0 + Δ T 0 ) 2 ( n 1 + 1 ) C Δ T 0 * ( H γ ) H γ ( x + Δ T 0 ) , x m 1 T 0
and
j = 0 m 1 F γ * n 0 ( x j T 0 + Δ T 0 ) 2 ( m 1 + 1 ) F γ * n 0 ( x m 1 T 0 + Δ T 0 ) , x x 3 ,
where the final inequality stems from (42) with k = n 0 in (35) and (16). In addition, by (35) with V = F , we know that for the above m 1 and n 0 , there are
0 < C 1 = C 1 ( H 2 , γ , T 0 , m 1 , n 0 ) < C 2 = C 2 ( H 2 , γ , T 0 , m 1 , n 0 ) <
and x 4 = x 4 ( H 2 , γ , T 0 , m 1 , n 0 ) x 3 such that, for all 1 j m 1 ,
C 1 H 2 , γ ( x j T 0 + Δ T 0 ) F γ * n 0 ( x j T 0 + Δ T 0 ) C 2 H 2 , γ ( x j T 0 + Δ T 0 ) , x x 4 .
For i = 1 , 2 , let X i be a random variable with distribution H i , γ . Assume that X 1 is independent of X 2 and ( X 1 * , X 2 * ) is an independent copy of ( X 1 , X 2 ) . Further, denote A 0 = { X 1 + X 2 x + Δ T 0 } for all x 0 . We then divide H γ ( x + Δ T 0 ) = P ( A 0 ) as follows:
P ( A 0 ) = P ( A 0 , 0 X 2 x m 1 T 0 ) + P ( A 0 , x m 1 T 0 < X 2 x + T 0 ) = P 1 ( x ) + P 2 ( x ) , x 0 .
For P 1 ( x ) , by (43), (44) and H γ L Δ T 0 OS Δ T 0 , we have
P 1 ( x ) = 0 x m 1 T 0 H 1 , γ ( x y + Δ T 0 ) H 2 , γ ( d y ) ε 0 x m 1 T 0 H γ ( x y + Δ T 0 ) H 2 , γ ( d y ) = ε P ( X 1 * + X 2 * + X 2 x + Δ T 0 , 0 X 2 x m 1 T 0 ) ε P ( x < X 1 * + X 2 * + X 2 x + T 0 , 0 X 1 n 2 T 0 ) / H 1 , γ ( 0 , n 2 T 0 ] ε P ( x < X 1 + X 2 + X 1 * + X 2 * x + T 0 + n 2 T 0 ) / H 1 , γ ( 0 , n 2 T 0 ] = ε j = 0 n 2 H γ * 2 ( x + j T 0 + Δ T 0 ) / H 1 , γ ( 0 , n 2 T 0 ] 2 ε ( n 2 + 1 ) C Δ T 0 * ( H γ ) H γ ( x + Δ T 0 ) / H 1 , γ ( 0 , n 2 T 0 ] , x m 1 T 0 .
For P 2 ( x ) , by (45) and (46), we have
P 2 ( x ) P ( x x 0 < X 2 x + T 0 ) = j = 0 m 1 H 2 , γ ( x j T 0 + Δ T 0 ) j = 0 m 1 F γ * n 0 ( x j T 0 + Δ T 0 ) / C 1 2 ( m 1 + 1 ) F γ * n 0 ( x m 1 T 0 + Δ T 0 ) / C 1 2 C 2 ( m 1 + 1 ) H 2 , γ ( x m 1 T 0 + Δ T 0 ) / C 1 , x x 4 .
Combined with (47), (48) and (49), we have
1 2 ε ( n 2 + 1 ) C Δ T 0 * ( H γ ) H 1 , γ ( 0 , n 2 T 0 ] H γ ( x + Δ T 0 ) 2 C 2 ( m 1 + 1 ) C 1 H 2 , γ ( x m 1 T 0 + Δ T 0 )
for x max { m 1 T 0 , x 4 } . Furthermore, by (50), (39) and 2 ε ( n 2 + 1 ) C * ( H γ ) / H 1 , γ ( 0 , n 2 T 0 ] < 1 , we know that
e β ( x m 1 T 0 ) H 2 , γ ( x m 1 T 0 + Δ T 0 ) = e β m 1 T 0 e β x H 2 , γ ( x m 1 T 0 + Δ T 0 ) for each β > 0 ,
that is
e β x H 2 , γ ( x + Δ T 0 ) for each β > 0 .
Then by (15), (40) and (51), it holds that
H 1 , γ ( x + Δ T 0 ) / H 2 , γ ( x + Δ T 0 ) = e β x H 1 , γ ( x + Δ T 0 ) / e β x H 2 , γ ( x + Δ T 0 ) 0 for each β > 0 ,
and thus (38) holds.
Secondly, by H γ L Δ T 0 OS Δ T 0 and (50), we know that
H 2 , γ ( x + Δ T 0 ) H γ ( x + Δ T 0 ) and H 2 , γ OS Δ T 0 .
Finally, we prove that H 2 , γ L Δ T 0 . On one hand, for any 0 < ε < 1 / 2 , take n 0 in (33) and (16) with V γ = F γ , by (32) and (10), according to Lemma 3 ( i ) , for each t > 0 , there is a constant x 5 = x 5 ( F , ε , t , γ ) x 4 such that
H 2 , γ ( x t + Δ T 0 ) k = 1 n 0 q k F γ * k ( x t + Δ T 0 ) ( 1 ε ) k = 1 n 0 q k F γ * k ( x + Δ T 0 ) ( 1 ε ) k = 1 q k F γ * k ( x + Δ T 0 ) k = n 0 + 1 q k F γ * k ( x + Δ T 0 ) ( 1 2 ε ) H 2 , γ ( x + Δ T 0 ) , x x 5 .
On the other hand, for any 0 < ε < 1 , each t > 0 and n 0 in (33) with V γ = F γ , by (36) and (10) for all k 1 , there is a constant x 6 = x 6 ( F , ε , t , γ ) x 5 such that, when x x 6 ,
H 2 , γ ( x + Δ T 0 ) H 2 , γ ( x t + Δ T 0 ) H 2 , γ ( x t + Δ T 0 ) k = 1 n 0 F γ * k ( x + Δ T 0 ) F γ * k ( x t + Δ T 0 ) 1 + ε ε ( n 0 + 1 ) .
Combining (52) and (53), with the arbitrariness of ε , we know that H 2 , γ L Δ T 0 . Then ( i ) holds by Lemma 1 and Proposition 1.
( i i ) In Theorem 5, we take V = F , G = H , G 1 = H 1 and G 2 = H 2 . Because p k = e μ μ k / k ! k 0 , according to Remark 3 ( i ) , (16) holds for each distribution V, thus for F. Therefore, since H 2 TL Δ T 0 ( γ ) OS Δ T 0 , according to Theorem 5, by (16) for F and (10), we obtain all the results.

4. On the Condition (10)

In this section, we give some concise and convenient conditions to replace condition (10), see the following Theorem 6. To this end, we require three lemmas.
Lemma 4.
If a distribution V OL Δ T for some 0 < T < satisfying
lim inf V ( x t + Δ T ) / V ( x + Δ T ) 1 f o r e a c h t > 0 ,
then
V ( x t + Δ T ) V ( x + Δ T ) V ( x + t + Δ T ) f o r e a c h t > 0
and
V ( x + Δ T 1 ) = O V ( x + Δ T ) f o r e a c h p a i r 0 < T 1 T < .
Proof. 
Firstly, by (54) and V OL Δ T , we know that, for each t > 0 ,
V ( x + Δ T ) V ( x t + Δ T ) C Δ T ( V , t ) V ( x + Δ T ) ,
that is V ( x t + Δ T ) V ( x + Δ T ) . Thus
V ( x + Δ T ) = V ( x + t t + Δ T ) V ( x + t + Δ T ) .
Therefore, (55) holds.
Secondly, for each 0 < T 1 T < , there exists an integer m 1 such that ( m 1 ) T < T 1 m T . Further, by V OL Δ T and V ( x + Δ T ) V ( x + t + Δ T ) for each t > 0 , we have
V ( x + Δ T 1 ) V ( x + Δ m T ) = k = 0 m 1 V ( x + k T + Δ T ) k = 0 m 1 C Δ T * ( V , k T ) V ( x + Δ T ) ,
that is (56) holds. □
Lemma 5.
For i = 1 , 2 , let V i be a distribution such that V i OL Δ T for some 0 < T < and
lim inf V i ( x t + Δ T ) / V i ( x + Δ T ) 1 f o r e a c h t > 0 .
( i ) T h e n V i ( x + Δ T ) = O V 1 V 2 ( x + Δ T ) , i = 1 , 2 .
( i i ) If
lim C Δ T ( V 1 , x ) V 2 ( x + Δ T ) = 0 ,
then V 1 V 2 OL Δ T and
C Δ T * ( V 1 * V 2 , t ) max { C Δ T * ( V 1 , t ) , C Δ T * ( V 2 , t ) } f o r e a c h t 0 .
Proof. 
( i ) For any 0 < A < , according to Fatou lemma, by (57), we have
lim inf V 1 * V 2 ( x + Δ T ) V i ( x + Δ T ) 0 A lim inf V i ( x y + Δ T ) V i ( x + Δ T ) V j ( d y ) V j ( [ 0 , A ] ) 1 , as A ,
for all 1 i j 2 , that is (58) holds.
( i i ) In order to prove (60), we perform some preparatory work.
For each t > 0 , any 0 < ε < 1 and i = 1 , 2 , by V i OL Δ T , there exists x i = x i ( V i , ε , T , t ) > 0 such that
V i ( x t + Δ T ) ( 1 + ε ) C Δ T * ( V i , t ) V i ( x + Δ T ) for all x x i .
For the above ε , by (59), there exists x 3 = x 3 ( V i , ε , T ) > 0 such that when x x 3 ,
C Δ T * ( V 1 , x ) V 2 ( x + Δ T ) < ε .
For above t > 0 , by (56), (61), (62) and (58), we know that there is x 0 = x 0 ( V 1 , V 2 , ε , T , t ) max { x 1 , x 2 , x 3 } , 0 < K i = K i ( V 1 , V 2 , T ) < , i = 1 , 2 and m = m ( V 1 , V 2 , T , ε , t ) 2 such that, when x m x 0 ,
V 1 ( x x 0 t T + Δ 2 T ) V 2 ( x 0 + Δ T ) K 1 V 1 ( x x 0 t T + Δ T ) V 2 ( x 0 + Δ T ) K 1 ( 1 + ε ) 3 C Δ T * ( V 1 , T ) C Δ T * ( V 1 , t ) C Δ T * ( V 1 , x 0 ) V 1 ( x + Δ T ) V 2 ( x 0 + Δ T ) K 1 K 2 ( 1 + ε ) 3 C Δ T * ( V 1 , T ) C Δ T * ( V 1 , t ) C Δ T * ( V 1 , x 0 ) V 2 ( x 0 + Δ T ) V 1 V 2 ( x + Δ T ) < ε ( 1 + ε ) 3 K V 1 V 2 ( x + Δ T ) ,
where K = K 1 K 2 C Δ T * ( V 1 , T ) C Δ T * ( V 1 , t ) . In addition, let X and Y be the two random variables with corresponding distributions V 1 and V 2 . Suppose that X is independent of Y. Denote
A t = { X + Y x t + Δ T } for t 0 .
In the following, we deal with V 1 V 2 ( x t + Δ T ) in two cases where T t < and 0 < t < T . For T t < , by (61) and x m x 0 + T , we have
V 1 V 2 ( x t + Δ T ) = P ( A t , 0 X x t x 0 ) + P ( A t , x t x 0 < X x t + T ) P ( A t , 0 X x t x 0 ) + P ( A t , 0 < Y T + x 0 ) = 0 x t x 0 V 2 ( x t y + Δ T ) V 1 ( d y ) + 0 T + x 0 V 1 ( x t y + Δ T ) V 2 ( d y ) ( 1 + ε ) ( C Δ T * ( V 2 , t ) 0 x t x 0 V 2 ( x y + Δ T ) V 1 ( d y ) + C Δ T * ( V 1 , t ) 0 T + x 0 V 1 ( x y + Δ T ) V 2 ( d y ) ) ( 1 + ε ) max { C Δ T * ( V 1 , t ) , C Δ T * ( V 2 , t ) } V 1 * V 2 ( x + Δ T ) .
For 0 < t < T , we give a segmentation for V 1 V 2 ( x t + Δ T ) which is different from (64) as follows. Further, by (61), (63) and x m x 0 + T , we have
V 1 V 2 ( x t + Δ T ) P ( A t , 0 X x t x 0 ) + P ( A t , 0 < Y x 0 ) + P ( A t , x 0 < Y T + x 0 ) 0 x t x 0 V 2 ( x t y + Δ T ) V 1 ( d y ) + 0 x 0 V 1 ( x t y + Δ T ) V 2 ( d y ) + V 1 ( x x 0 t T + Δ 2 T ) V 2 ( x 0 + Δ T ) ( 1 + ε ) C Δ T * ( V 2 , t ) 0 x t x 0 V 2 ( x y + Δ T ) V 1 ( d y ) + C Δ T * ( V 1 , t ) 0 x 0 V 1 ( x y + Δ T ) V 2 ( d y ) + ε ( 1 + ε ) 2 K V 1 V 2 ( x + Δ T ) ( 1 + ε ) max { C Δ T * ( V 1 , t ) , C Δ T * ( V 2 , t ) } + ε ( 1 + ε ) 3 K V 1 V 2 ( x + Δ T ) .
Therefore, V 1 V 2 OL Δ T and (60) holds by (64), (65) and the arbitrariness of ε . □
Lemma 6.
Let V 1 and V 2 be the two distributions belonging to the class OL Δ T for some 0 < T < . If conditions (57) and (59) are satisfied, then for each t > 0 ,
lim inf V 1 V 2 ( x t + Δ T ) / V 1 V 2 ( x + Δ T ) 1 .
Proof. 
In order to prove (66), we carry out some preparatory work. For each t > 0 , by (57) and V 2 OL Δ T , we know that, for any 0 < ε < 1 , there exists x 1 = x 1 ( V 1 , V 2 , ε , T , t ) > 0 such that, when x x 1 ,
( 1 ε ) V i ( x + Δ T ) V i ( x t + Δ T ) ( 1 + ε ) C Δ T * ( V i , t ) V i ( x + Δ T ) , i = 1 , 2 .
Furthermore, according to Lemma 5, there exists x 2 = x 2 ( V 1 , V 2 , T , t ) > 0 and C > 0 such that, when x x 2 ,
V 1 ( x + Δ T ) C V 1 V 2 ( x + Δ T ) and V 1 ( or 2 ) x + Δ t ( or Δ t + T ) C V 1 ( x + Δ T ) .
Further, we denote x 0 = max { x 1 , x 2 } + t + T and A t = { X + Y x t + Δ T } , where X and Y are two random variables defined in Lemma 5.
Now, we prove (66) for each t > 0 . When x x 0 , by (67), we have
V 1 V 2 ( x t + Δ T ) = P ( A t , 0 X x t x 0 ) + P ( A t , x t x 0 < X x t + T ) = 0 x t x 0 V 2 ( x t y + Δ T ) V 1 ( d y ) + P ( A t , x t x 0 < X x t + T ) ( 1 ε ) 0 x t x 0 V 2 ( x y + Δ T ) V 1 ( d y ) + P ( A t , x t x 0 < X x t + T ) ( 1 ε ) ( V 1 * V 2 ( x + Δ T ) x t x 0 x x 0 V 2 ( x y + Δ T ) V 1 ( d y ) P ( A 0 , x x 0 < X x + T ) + P ( A t , x t x 0 < X x t + T ) ) = ( 1 ε ) V 1 V 2 ( x + Δ T ) P 1 ( x ) P 2 ( x ) + P 3 ( x ) .
Firstly, we estimate P 1 ( x ) . When x t x 0 < y x x 0 ,
V 2 ( x y + Δ T ) P ( x 0 < Y x 0 + t + T ) .
Then, by (68), (67) and (59), we know that
P 1 ( x ) V 2 ( x 0 + Δ t + T ) V 1 ( x x 0 t + Δ t ) C 2 V 2 ( x 0 + Δ T ) V 1 ( x x 0 t + Δ T ) ( 1 + ε ) 2 C 2 V 1 ( x + Δ T ) C Δ T * ( V 1 , t ) C Δ T * ( V 1 , x 0 ) V 2 ( x 0 + Δ T ) ( 1 + ε ) 2 C 3 V 1 V 2 ( x + Δ T ) C Δ T * ( V 1 , t ) C Δ T * ( V 1 , x 0 ) V 2 ( x 0 + Δ T ) = o V 1 V 2 ( x + Δ T ) as x 0 .
Secondly, we estimate P 3 ( x ) P 2 ( x ) .
P 3 ( x ) P 2 ( x ) = P ( A t , x t x 0 < X x t + T , 0 < Y < x 0 + T ) P ( A 0 , x x 0 < X x + T , 0 Y < x 0 + T ) = 0 x 0 V 1 ( x t y + Δ T ) V 1 ( x y + Δ T ) V 2 ( d y ) + x 0 x 0 + T P ( X x t y + Δ T , x t x 0 < X x t + T ) V 2 ( d y ) x 0 x 0 + T P ( X x y + Δ T , x x 0 < X x + T ) V 2 ( d y ) = P 11 ( x ) + P 12 ( x ) P 13 ( x ) .
By (67), we have
P 11 ( x ) V 1 V 2 ( x + Δ T ) ε V 1 * V 2 ( x + Δ T ) 0 x 0 V 1 ( x y + Δ T ) V 2 ( d y ) ε .
Using the proof method of (70), we can get that
P 12 ( x ) = x 0 x 0 + T P ( x t x 0 < X x t y + T ) V 2 ( d y ) V 1 ( x t x 0 + Δ T ) V 2 ( x 0 + Δ T ) = o V 1 V 2 ( x + Δ T ) , a s x 0 .
Similarly, we have
P 13 ( x ) = x 0 x 0 + T P ( x x 0 < X x y + T ) V 2 ( d y ) V 1 ( x x 0 + Δ T ) V 2 ( x 0 + Δ T ) = o V 1 * V 2 ( x + Δ T ) , a s x 0 .
Combining with (69)–(74), we know that (66) holds. □
Theorem 6.
Suppose that V OL Δ T for some 0 < T < . If conditions (57) and (59) are satisfied for V 1 = V 2 = V , then for all k 2 , V * k OL Δ T ,
C Δ T * ( V * k , t ) C Δ T * ( V , t ) for each t 0 .
and
lim inf V * k ( x t + Δ T ) / V * k ( x + Δ T ) 1 for each t 0 .
Proof. 
We use mathematical induction to prove the result.
Clearly, (75) and (76) hold for k = 1 . Assume that V * k OL Δ T , (75) and (76) hold for some k 2 . Set V 1 = V * k and V 2 = V in Theorem 6. By (59) and (75), we have V * ( k + 1 ) = V 1 V 2 OL Δ T and (76) holds for k + 1 . Thus, (75) holds for k + 1 , too. □
In particular, we take V = F γ and T = T 0 in (76), then we obtain (10) in Theorem 3 of this paper.

5. Conclusions and Future Work

In this paper, we prove that the class L l o c OS l o c , in addition to TL Δ T 0 ( γ ) OS Δ T 0 for some 0 < γ , T 0 < are not closed under the I.I.D. root. However, by adding certain conditions, the two classes become closed under the I.I.D. root. At the same time, we also provide the corresponding results under the random convolution roots.
In this section, we briefly introduce the theoretical significance and application value of the above results reported herein, in addition to some unresolved problems.

5.1. Theoretical Significance and Application Value

In complex practice, F is often in a “black box”, that is, it is unknown or partially unknown. For example, in Theorem 2, we only know that F has property (8) or (9), but we do not know whether it has property F * k L l o c OS l o c for some k 1 . Furthermore, the properties of H, as the external expression of F, can be estimated by some statistical methods. Therefore, it is of great theoretical significance and application value to use known H to estimate unknown F. This presents the research purpose of this paper.
In the following, we provide some specific examples to illustrate applications of the research findings herein.
Firstly, it is well known that the distribution of components of the Lévy process is I.I.D. Therefore, research on I.I.D. H is beneficial to the Lévy process.
Secondly, in the Cramér–Lundeberg risk model, the distributions F, H 2 and H 1 satisfying the conditions (2) and (3) can be regarded as the distributions of the claim, the total claim amount and the perturbation to the total claim amount, respectively, see SubSection 1.3.3 of Embrechts et al. [34]. If the disturbed distribution of total claim amount H = H 1 H 2 is an I.I.D. and H L l o c OS l o c , then according to Theorem 2, we have H 2 L l o c OS l o c and H 2 ( x + Δ T ) H ( x + Δ T ) for each 0 < T under condition (8) or (9). Interestingly, F does not have to belong to class L l o c OS l o c , but F * k belongs to that class for all k 2 , see Theorems 1 and 2 mentioned in this paper.
There are many similar examples, such as H 2 which is the distribution of proportional reinsurance or the claim in Poisson model, see Example 5.2 (i) of Klüppelberg and Mikosch [35] and the main theorems of Veraverbeke [36].
Therefore, the results of this paper undoubtedly play an important role in risk theory and other fields.
Finally, the results of this paper can offer a more complete and profound answer to the famous Embrechts–Goldie conjecture, see Section 5.2 below for details.

5.2. On the Embrechts–Goldie Conjecture

Let X be a distribution class, and let V be a distribution. If V * 2 X implies V X , then we say that the class X is closed under convolution roots. Clearly, the closure under I.D.D roots is the natural extension of the closure under convolution roots for some distribution class.
Theorem 2 of Embrechts et al. [2] shows that the class S is closed under convolution roots. The same conclusion also holds for the class S ( γ ) for some γ > 0 if the distribution V L ( γ ) , see Theorem 2.10 of Embrechts and Goldie [24]. Therefore, Embrechts and Goldie [15,24] put forward a famous conjecture:
If V * k L ( γ ) for some ( even for all ) k 2 and γ 0 , then V L ( γ ) .
Many positive or negative conclusions related to the conjecture are then proposed. Some positive results can be found in Theorem 1.2 of Watanabe [11] for the class S ( γ ) for some γ > 0 , Theorem 6 of Xu et al. [31] for the classes L ( γ ) and L ( γ ) OS . Of course, these outcomes are valid under certain restrictive conditions.
The following references provide us with the negative results.
Theorem 1.1 of Watanabe [11] shows that the class S ( γ ) for some γ > 0 is not closed under the convolution roots in general.
Earlier, Shimura and Watanabe [37] showed that there is a distribution V such that V * 2 L ( γ ) OS for some γ 0 , while V OL γ 0 L ( γ ) OS and V ¯ ( x ) = o V * 2 ¯ ( x ) .
Further, Theorem 1.1 of Xu et al. [31] points out that there is a distribution V OL γ 0 L ( γ ) OS and V ¯ ( x ) o V * 2 ¯ ( x ) such that V * 2 L ( γ ) OS S ( γ ) for each γ > 0 .
For γ = 0 , Theorem 2.2 (1) of Xu et al. [30] shows that there is a distribution V such that V OL L OS and V ¯ ( x ) o V * 2 ¯ ( x ) , while V * k L ( γ ) OS S for all k 2 . Then, Proposition 2.2 of Xu et al. [30] points out that there are two distributions V 1 and V 2 such that V 1 , V 2 OL , while V 1 * k L OS S and V 2 * k L OS for all k 2 .
This result reveals a surprising phenomenon that, although the properties of a distribution V are very poor, its convolution, and even its random convolution and the corresponding I.I.D., bear good properties.
Therefore, the Embrechts–Goldie conjecture has been denied for the class L ( γ ) and its subclasses S ( γ ) S , ( L ( γ ) OS ) S ( γ ) and L ( γ ) OS for each γ 0 , where the corresponding distribution V OL γ 0 L ( γ ) OS , and even V OL .
In this subsection, we mainly focus on the local closure under the convolution root.
For negative conclusions, Corollary 1.1 of Watanabe [11] shows that the classes S l o c , L l o c , S Δ T and L Δ T for some 0 < T < are not closed under convolution roots. Further, Theorem 1 of the paper and its proof show that the class L l o c OS l o c is not closed either.
In addition, Theorem 1.1 and Corollary 1.1 of Watanabe and Yamamuro [38] and Theorem 1.1 and Corollary of Watanabe [39] obtain some results corresponding to Corollary 1.1 of Watanabe [11] for the subexponential density classes and the subexponential lattice distribution classes, respectively. Clearly, the lattice distribution is a special local distribution, and the density is closely related to its local distribution.
As positive conclusions, Theorem 2.1 of Watanabe [39] shows that the subexponential lattice distribution classes are closed under convolution roots with a condition. However, other positive conclusions about the local closure in non-lattice cases are rare.
In this paper, according to Theorem 6 of Xu et al. [31], Proposition 1 and Lemma 1 of the paper, using the Esscher transform, we give a corresponding positive result for the class L l o c OS l o c and omit the proof details.
Theorem 7.
Let V be a distribution, and let γ and T be two positive and finite constants.
( i ) Assume that V OS l o c and
lim inf V γ ¯ ( x t ) / V γ ¯ ( x ) e γ t f o r e a c h t > 0
or
V γ ¯ ( x ) = o V γ * 2 ¯ ( x ) .
If V * 2 L l o c , then V L l o c .
( i i ) Assume that V L l o c with the mean μ V < , the condition (77) is satisfied and
C * ( V γ * 2 ) < 6 M ( V γ * 2 , γ ) .
If V * 2 L l o c OS l o c , then V L l o c OS l o c .
Using the Esscher transform, by (15), we can replace the condition (78) with a more immediate condition.
Proposition 2.
If V * 2 L l o c , then (78) is implied by the following condition:
V ( x + Δ T ) = o V * 2 ( x + Δ T ) .

5.3. Some Unresolved Problems

Clearly, for other local distribution classes, such as the class L l o c OS l o c and the class TL Δ T 0 ( γ ) OS Δ T 0 for some 0 < γ , T 0 < , the following corresponding questions arise:
Are they closed under the I.I.D. root? If not, under what conditions are they closed under the I.I.D. root?
Perhaps we can first solve the corresponding problem of the global distribution class L ( γ ) OS with some γ > 0 . In addition, the existing results, apart from Proposition 2.1 of Xu et al. [30], often assume that F OL . Then, if F OL , what will we get?
Further, if F does not belong to the class L l o c OS l o c , L l o c OS l o c or TL Δ T 0 ( γ ) OS Δ T 0 for some 0 < γ , T 0 < , what kind of F can make F * k for all k l 0 and some l 0 2 , H 2 and H belong to the same class? Even if F OL l o c , what will we get?
In our opinion, these questions are both interesting and difficult to solve. The theory will become more complete following the provision of solutions to these questions.

Author Contributions

Conceptualization, Z.C. and Y.W.; methodology, Y.W.; formal analysis, Z.C.; writing—original draft preparation, H.X.; writing—review and editing, Y.W. and Z.C.; funding acquisition, Z.C. and Y.W. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Natural Science Foundation of China grant number 11071182.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Not applicable.

Acknowledgments

We would like to thank the Editor and Reviewers for their valuable comments and suggestions that assisted in improving our manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Feller, W. An Introduction to Probability Theory and Its Applications; Wiley: Hoboken, NJ, USA, 1971. [Google Scholar]
  2. Embrechts, P.; Goldie, C.M.; Veraverbeke, N. Subexponentiality and infinite divisibility. Z. Wahrscheinlichkeitstheorie Verw. Gibiet. 1979, 49, 335–347. [Google Scholar] [CrossRef]
  3. Sato, K. Lévy processes and infinitely divisible distributions. In Cambridge Studies in Advanced Mathematics; Cambridge University Press: Cambridge, UK, 1999. [Google Scholar]
  4. Borovkov, A.A.; Borovkov, K.A. Asymptotic Analysis of Random Walks; Cambridge University Press: Cambridge, UK, 2008. [Google Scholar]
  5. Asmussen, S.; Foss, S.; Korshunov, D. Asymptotics for sums of random variables with local subexponential behavior. J. Theor. Probab. 2003, 16, 489–518. [Google Scholar] [CrossRef]
  6. Wang, Y.; Xu, H.; Cheng, D.; Yu, C. The local asymptotic estimation for the supremum of a random walk. Stat. Pap. 2018, 59, 99–126. [Google Scholar] [CrossRef] [Green Version]
  7. Wang, Y.; Cheng, D.; Wang, K. The Closure of Local subexponential distribution class under convolution roots with applications. J. Appl. Probab. 2005, 42, 1194–1203. [Google Scholar] [CrossRef]
  8. Wang, Y.; Yang, Y.; Wang, K.; Cheng, D. Some new equivalent conditions on asymptotics and local asymptotics for random sums and their applications. Insur. Math. Econ. 2007, 40, 256–266. [Google Scholar] [CrossRef]
  9. Denisov, D.; Dieker, A.B.; Shneer, V. Large deviations for random walks under subexponentiality: The big-jump domain. Ann. Probab. 2008, 36, 1946–1991. [Google Scholar] [CrossRef]
  10. Yang, Y.; Leipus, R.; Siaulys, J. Local presice deviations for sums of random variables with O-reqularly varying densities. Stat. Probab. Lett. 2010, 80, 1559–1567. [Google Scholar] [CrossRef]
  11. Watanabe, T. The Wiener condition and the conjectures of Embrechts and Goldie. Ann. Probab. 2019, 47, 1221–1239. [Google Scholar] [CrossRef] [Green Version]
  12. Chistyakov, V.P. A theorem on sums of independent positive random variables and its application to branching processes. Theory Probab. Its Appl. 1964, 9, 640–648. [Google Scholar] [CrossRef]
  13. Shimura, T.; Watanabe, T. Infinite divisibility and generalised subexponentiality. Bernoulli 2005, 11, 445–469. [Google Scholar] [CrossRef]
  14. Klüppelberg, C. Asymptotic ordering of distribution functions and convolution semigroups. Semigroup Forum 1990, 40, 77–92. [Google Scholar] [CrossRef]
  15. Embrechts, P.; Goldie, C.M. On closure and factorization properties of subexponential tails. J. Aust. Math. Soc. (Ser. A) 1980, 29, 243–256. [Google Scholar] [CrossRef] [Green Version]
  16. Chen, W.; Yu, C.; Wang, Y. Some discussions on the local distribution classes. Stat. Probab. Lett. 2013, 83, 1654–1661. [Google Scholar] [CrossRef]
  17. Chover, J.; Ney, P.; Wainger, S. Functions of probability measures. J. d’Anal. MathéMatique 1973, 26, 255–302. [Google Scholar] [CrossRef]
  18. Chover, J.; Ney, P.; Wainger, S. Degeneracy properties of subcritical branching processes. Ann. Probab. 1973, 1, 663–673. [Google Scholar] [CrossRef]
  19. Bertoin, J.; Doney, R.A. Some asymptotic results for transient random walks. Adv. Appl. Probab. 1996, 28, 207–226. [Google Scholar] [CrossRef]
  20. Foss, S.; Korshunov, D.; Zachary, S. An Introduction to Heavy-tailed and Subexponential Distributions, 2nd ed.; Springer: Berlin/Heidelberg, Germany, 2013. [Google Scholar]
  21. Wang, Y. The Criterion of A Big Jump: Theory and Application of Heavy tailed Distribution; Science Press: Beijing, China, 2022. (In Chinese) [Google Scholar]
  22. Teugels, J.L. The class of subexponential distributions. Ann. Probab. 1975, 3, 1000–1011. [Google Scholar] [CrossRef]
  23. Veraverbeke, N. Asymptotic behavior of Wiener-Hopf factors of a random walk. Stoch. Process. Their Appl. 1977, 5, 27–37. [Google Scholar] [CrossRef] [Green Version]
  24. Embrechts, P.; Goldie, C.M. On convolution tails. Stoch. Process. Their Appl. 1982, 13, 263–278. [Google Scholar] [CrossRef] [Green Version]
  25. Wang, Y.; Wang, K. Random walks with non-convolution equivalent increments and their applications. J. Math. Anal. Appl. 2011, 374, 88–105. [Google Scholar] [CrossRef]
  26. Sgibnev, M.S. Asymptotics of infinite divisibility on R. Sib. Math. J. 1990, 31, 115–119. [Google Scholar] [CrossRef]
  27. Pakes, A.G. Convolution equivalence and infinite divisibility. J. Appl. Probab. 2004, 41, 407–424. [Google Scholar] [CrossRef]
  28. Watanabe, T. Convolution equivalence and distributions of random sums. Probab. Theory Relat. Fields 2008, 142, 367–397. [Google Scholar] [CrossRef]
  29. Cui, Z.; Wang, Y.; Xu, H. Some positive conclusions related to the Embrechts-Goldie conjecture. Sib. Math. J. 2022, 63, 216–231. [Google Scholar] [CrossRef]
  30. Xu, H.; Foss, S.; Wang, Y. Convolution and convolution-root properties of long-tailed distributions. Extremes 2015, 18, 605–628. [Google Scholar] [CrossRef] [Green Version]
  31. Xu, H.; Wang, Y.; Cheng, D.; Yu, C. On the closure under infinitely divisible distribution roots. Lith. Math. J. 2022, 62, 259–287. [Google Scholar] [CrossRef]
  32. Jiang, T.; Wang, Y.; Cui, Z.; Chen, Y. On the almost decrease of a subexponential density. Stat. Probab. Lett. 2019, 153, 71–79. [Google Scholar] [CrossRef] [Green Version]
  33. Watanabe, T.; Yamamuro, K. Ratio of the tail of an infinitely divisible distribution on the line to that of its Lévy measure. Electron. J. Probab. 2010, 15, 44–74. [Google Scholar] [CrossRef]
  34. Embrechts, P.; Klüppelberg, C.; Mikosch, T. Modelling Extremal Events for Insurance and Finance; Springer: Berlin/Heidelberg, Germany, 1997. [Google Scholar]
  35. Klüppelberg, C.; Mikosch, T. Large deviations of heavy-tailed random sums with applications in insurance and finance. J. Appl. Probab. 1997, 34, 293–308. [Google Scholar] [CrossRef]
  36. Veraverbeke, N. Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Insur. Math. Econ. 1993, 13, 57–62. [Google Scholar] [CrossRef]
  37. Shimura, T.; Watanabe, T. On the convolution roots in the convolution-equivalent class. Inst. Stat. Math. Coop. Res. Rep. 2005, 175, 1–15. [Google Scholar]
  38. Watanabe, T.; Yamamuro, K. Two non-closure properties on the Class of subexponential densities. J. Theor. Probab. 2017, 30, 1059–1075. [Google Scholar] [CrossRef] [Green Version]
  39. Watanabe, T. Embrechts-Goldie’s problem on the class of lattice convolution equivalent distributions. J. Theor. Probab. 2021. [CrossRef]
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Cui, Z.; Wang, Y.; Xu, H. Local Closure under Infinitely Divisible Distribution Roots and Esscher Transform. Mathematics 2022, 10, 4128. https://doi.org/10.3390/math10214128

AMA Style

Cui Z, Wang Y, Xu H. Local Closure under Infinitely Divisible Distribution Roots and Esscher Transform. Mathematics. 2022; 10(21):4128. https://doi.org/10.3390/math10214128

Chicago/Turabian Style

Cui, Zhaolei, Yuebao Wang, and Hui Xu. 2022. "Local Closure under Infinitely Divisible Distribution Roots and Esscher Transform" Mathematics 10, no. 21: 4128. https://doi.org/10.3390/math10214128

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop