Abstract
Two hypotheses on the class \({\mathcal {L}}(\gamma )\) in the class \(\mathcal {OS}\cap \mathcal {ID}\) are discussed. Two weak hypotheses on the class \({\mathcal {L}}(\gamma )\) in the class \(\mathcal {OS}\cap \mathcal {ID}\) are proved. A necessary and sufficient condition in order that, for every \(t>0\), the t-th convolution power of a distribution in the class \(\mathcal {OS}\cap \mathcal {ID}\) belongs to the class \({\mathcal {L}}(\gamma )\) is given. Sufficient conditions are given for the validity of two hypotheses on the class \({\mathcal {L}}(\gamma )\).
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction and Results
In what follows, we denote by \({\mathbb {R}}\) the real line and by \({\mathbb {R}}_{+}\) the half line \([0,\infty )\). Denote by \({\mathbb {N}}\) the totality of positive integers and by \(a{\mathbb {N}}\) the set \(\{a,2a,3a,\dots \}\). The symbol \(\delta _a(\hbox {d}x)\) stands for the delta measure at \(a \in {\mathbb {R}}\). Let \(\eta \) and \(\rho \) be probability distributions on \({\mathbb {R}}\). We denote by \(\eta *\rho \) the convolution of \(\eta \) and \(\rho \) and by \(\rho ^{n*}\) n-th convolution power of \(\rho \) with the understanding that \(\rho ^{0*}(\hbox {d}x)=\delta _0(\hbox {d}x)\). Denote by \({{\bar{\xi }}}(x)\) the tail \(\xi ((x,\infty ))\) of a measure \(\xi \) on \({\mathbb {R}}\) for \(x \in {\mathbb {R}}\). Let \(\gamma \ge 0\). We define the \(\gamma \)-exponential moment \({{\widehat{\xi }}}(\gamma )\) as
If \({{\widehat{\xi }}}(\gamma ) < \infty ,\) we define the Fourier–Laplace transform \({{\widehat{\xi }}}(\gamma +iz)\) for \(z\in {\mathbb {R}}\) as
An integral \(\int _a^bg(x)\rho (\hbox {d}x)\) means \(\int _{a+}^{b+}g(x)\rho (\hbox {d}x)\). For positive functions \(f_1(x)\) and \(g_1(x)\) on \([A,\infty )\) for some \(A \in {\mathbb {R}}\), we define the relation \(f_1(x)\sim g_1(x)\) by \(\lim _{x \rightarrow \infty }f_1(x)/g_1(x)=1\) and the relation \(f_1(x)\asymp g_1(x)\) by
Let \(\gamma \ge 0\). A distribution \(\rho \) on \({\mathbb {R}}\) belongs to the class \({\mathcal {L}}(\gamma )\) if \({\overline{\rho }}(x)>0\) for all \(x >0\) and, for every \( a \in {\mathbb {R}}\),
A distribution \(\rho \) on \({\mathbb {R}}\) belongs to the class \({\mathcal {S}}(\gamma )\) if \(\rho \in {\mathcal {L}}(\gamma )\), \( {{\widehat{\rho }}}(\gamma ) < \infty \), and
A distribution \(\rho \) on \({\mathbb {R}}\) belongs to the class \(\mathcal {OL}\) if \({\overline{\rho }}(x)>0\) for \(x >0\) and, for all \(a \ge 0\),
A distribution \(\rho \) on \({\mathbb {R}}\) belongs to the class \(\mathcal {OS}\) if \({\overline{\rho }}(x)>0\) for all \(x >0\) and
Note that the class \(\mathcal {OS}\) is included in the class \(\mathcal {OL}\). A distribution \(\rho \) on \({\mathbb {R}}\) belongs to the class \({\mathcal {S}}_{\sharp }\) if \(\rho \in \mathcal {OS}\) and
The class \({\mathcal {S}}_{\sharp }\) includes \(\cup _{\gamma \ge 0}{\mathcal {S}}(\gamma )\), and it is closed under convolution powers. A finite measure \(\xi \) satisfies the Wiener condition if \({{\widehat{\xi }}}(iz)\ne 0\) for every \(z \in {\mathbb {R}}\). Denote by \({\mathcal {W}}\) the totality of finite measures on \({\mathbb {R}}\) satisfying the Wiener condition. We denote by \(\mathcal {ID}\) the class of all infinitely divisible distributions on \({\mathbb {R}}\). For \(\mu \in \mathcal {ID}\), denote by \(\nu \) its Lévy measure. Under the assumption that \({{\bar{\nu }}}(c)>0\) for every \( c >0,\) define \(\nu _1(\hbox {d}x):=1_{(1,\infty )}(x)\nu (\hbox {d}x)/{{\bar{\nu }}}(1).\) Let \(\mu \in \mathcal {ID}\). We define a compound Poisson distribution \(\mu _1\) with \(c={{\bar{\nu }}}(1)\) as
Denote by \(\mu ^{t*}\) the t-th convolution power of \(\mu \in \mathcal {ID}\) for \(t >0\). Note that \(\mu ^{t*}\) is the distribution of \(X_t\) for a certain Lévy process \(\{X_t\}\) on \({\mathbb {R}}\). Let \(\gamma \ge 0\). Define \(T(\mu ,\gamma )\) as
Since the class \({\mathcal {L}}(\gamma )\) is closed under convolutions by Theorem 3 of Embrechts and Goldie [2], \(T(\mu ,\gamma )\) is empty or an additive semigroup in \((0,\infty )\). We see from Lemma 2.2 that for \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), there are positive integers n such that \(\nu _1^{n*} \in \mathcal {OS}\). Let \(n_0\) be the positive integer defined by (2.1). Note that we do not yet know an example of \(\mu \in \mathcal {OS}\cap \mathcal {ID}\) such that \(n_0 \ge 3.\)
A class \({\mathcal {C}}\) of distributions is called closed under convolution roots if \(\rho ^{n*} \in {\mathcal {C}}\) for some \(n \in {\mathbb {N}}\) implies \(\rho \in {\mathcal {C}}\). We see from Shimura and Watanabe [11] that the class \(\mathcal {OS}\) is not closed under convolution roots, but from Watanabe and Yamamuro [15] that the class \(\mathcal {OS}\cap \mathcal {ID}\) is closed under convolution roots. Embrechts et al. [4] in the one-sided case and Watanabe [13] in the two-sided case proved that the class \({\mathcal {S}}(0)\) is closed under convolution roots, and Embrechts and Goldie [2] conjectured that the class \({\mathcal {L}}(\gamma )\) with \(\gamma \ge 0\) is closed under convolution roots, but Shimura and Watanabe [12] showed that the class \({\mathcal {L}}(\gamma )\) with \(\gamma \ge 0\) is not closed under convolution roots. Moreover, Watanabe and Yamamuro [16] proved that the class \({\mathcal {S}}_{ac}\) of all absolutely continuous distributions on \({\mathbb {R}}\) with subexponential densities is not closed under convolution roots. Embrechts and Goldie [3] conjectured that the class \({\mathcal {S}}(\gamma )\) with \(\gamma >0\) is closed under convolution roots. Watanabe [13] proved that \({\mathcal {S}}(\gamma )\cap \mathcal {ID}\) with \(\gamma \ge 0\) is closed under convolution roots, but Watanabe [14] showed that the class \({\mathcal {S}}(\gamma )\) with \(\gamma >0\) is not closed under convolution roots. We add the following. Klüppelberg [5] showed that the class \(\mathcal {OS}\) is closed under convolutions. The class \({\mathcal {S}}(\gamma )\) is closed under convolution powers for \(\gamma \ge 0\), but Leslie [7], for \(\gamma = 0\), and Klüppelberg and Villasenor [6], for \(\gamma > 0\), proved that the class \({\mathcal {S}}(\gamma )\) is not closed under convolutions.
We consider the following two hypotheses on the class \({\mathcal {L}}(\gamma )\) in the class \(\mathcal {OS}\cap \mathcal {ID}\):
Hypothesis I Let \(\gamma \ge 0\). For every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), if \(\mu ^{n*} \in {\mathcal {L}}(\gamma )\) for some \(n \in {\mathbb {N}}\), then \(\mu ^{(n+1)*} \in {\mathcal {L}}(\gamma )\).
Hypothesis II Let \(\gamma \ge 0\). For every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), if \(\mu ^{n*} \in {\mathcal {L}}(\gamma )\) for some \(n \in {\mathbb {N}}\), then \(\mu \in {\mathcal {L}}(\gamma )\).
We also consider the weak version of the above hypotheses:
Hypothesis I\(^\prime \) Let \(\gamma \ge 0\). For every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), if \(\mu ^{n*}, \mu ^{(n+1)*} \in {\mathcal {L}}(\gamma )\) for some \(n \in {\mathbb {N}}\), then \(\mu ^{(n+2)*} \in {\mathcal {L}}(\gamma )\).
Hypothesis II\(^\prime \) Let \(\gamma \ge 0\). For every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), if \(\mu ^{n*}, \mu ^{(n+1)*} \in {\mathcal {L}}(\gamma )\) for some \(n \in {\mathbb {N}}\), then \(\mu \in {\mathcal {L}}(\gamma )\).
Let \(\gamma \ge 0\). Define
and
Theorem 1.1
Let \(\gamma \ge 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). We have the following:
-
(i)
\(\mathcal {OS}\cap \mathcal {ID}={\mathcal {A}}(\gamma )\cup {\mathcal {B}}(\gamma )\cup {\mathcal {C}}(\gamma ).\) Thus, Hypotheses I\(^\prime \) and II\(^\prime \) are true.
-
(ii)
The relation \(\mu \in {\mathcal {A}}(\gamma )\) holds if and only if, for all \(a \ge 0\),
$$\begin{aligned} \lim _{x \rightarrow \infty }\frac{e^{-\gamma a}\overline{\nu _1}(x-a)-\overline{\nu _1}(x)}{\overline{\nu _1^{n_0*}}(x)} = 0. \end{aligned}$$(1.1)If \(\mu \in {\mathcal {A}}(\gamma )\), then \(\nu _1^{n*} \notin {\mathcal {L}}(\gamma )\cap \mathcal {OS}\) for \(1 \le n \le n_0-1\) and \(\nu _1^{n*} \in {\mathcal {L}}(\gamma )\cap \mathcal {OS}\) for \( n \ge n_0\).
Corollary 1.1
Let \(\gamma \ge 0\). Then the following are equivalent:
-
(1)
Hypothesis I is true.
-
(2)
Hypothesis II is true.
-
(3)
\({\mathcal {C}}(\gamma )\) is empty.
-
(4)
For every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\) it holds that, for every \(2t \in T(\mu ,\gamma )\) and for every \(a \ge 0\),
$$\begin{aligned} \limsup _{x \rightarrow \infty }\limsup _{\lambda \rightarrow \infty }\frac{|\int _x^{\lambda -x}(e^{-\gamma a}\overline{\mu _1^{t*}}(\lambda -a-u)-\overline{\mu _1^{t*}}(\lambda -u))\mu _1^{t*}(\hbox {d}u)|}{\overline{\mu _1^{t*}}(\lambda )}=0.\qquad \quad \end{aligned}$$(1.2)
Remark 1.1
Let \(\gamma = 0\). Then, \({\mathcal {C}}(0)\) is empty and Hypotheses I and II are true. The relation \(\mu \in {\mathcal {A}}(0)\) holds if and only if
If \(\mu \in {\mathcal {A}}(0)\), then \(\nu _1^{n*} \notin {\mathcal {L}}(0)\cap \mathcal {OS}\) for \(1 \le n \le n_0-1\) and \(\nu _1^{n*} \in {\mathcal {L}}(0)\cap \mathcal {OS}\) for \( n \ge n_0\). Xu et al. showed in Theorem 2.2 of [18] an example of \(\mu \in {\mathcal {A}}(0)\) with \(n_0=2\).
For \(\gamma > 0\), we cannot yet answer the question whether Hypotheses I and II are true. However, under some additional assumptions in terms of Lévy measure, we establish that \({\mathcal {C}}(\gamma )\) is empty.
Proposition 1.1
Let \(\gamma > 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). Suppose that, for every \(a \ge 0\),
Then, we have either \(T(\mu ,\gamma )= (0,\infty )\) or \(\emptyset \).
Remark 1.2
Cui et al. [1] proved a result analogous to the above proposition under a stronger assumption. Xu et al. showed in Theorem 1.1 of [19] an example of the case where \(T(\mu ,\gamma )\ne \emptyset \) in the above proposition.
Proposition 1.2
Let \(\gamma > 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). Suppose that \(\nu _1^{2*} \in {\mathcal {L}}(\gamma )\) and the real part of \({{\widehat{\nu }}}_1(\gamma +iz)\) is not zero for every \(z \in {\mathbb {R}}\). Then, either \(T(\mu ,\gamma )= (0,\infty )\) or \(\emptyset \).
Proposition 1.3
Let \(\gamma > 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). Suppose that there exists \(n_1 \in {\mathbb {N}}\) such that \(\nu _1^{n_1*} \in {\mathcal {S}}_{\sharp }\). Then, either \(T(\mu ,\gamma )= (0,\infty )\) or \(\emptyset \). The equality \(T(\mu ,\gamma )=(0,\infty )\) holds if and only if \(\nu _1 \in {\mathcal {S}}(\gamma )\).
Remark 1.3
Watanabe made in Theorem 1.1 of [14] a distribution \(\eta \in {\mathcal {S}}_{\sharp }\) such that \(\eta ^{n*} \in {\mathcal {S}}(\gamma )\) for every \(n \ge 2\) but \(\eta \notin {\mathcal {S}}(\gamma )\). Thus, taking this \(\eta \) as \(\nu _1 \), then Proposition 1.3 holds with \(T(\mu ,\gamma )= \emptyset \).
2 Preliminaries
In this section, we give several basic results as preliminaries. Pakes [8] proved the following.
Lemma 2.1
(Lemmas 2.1 and 2.5 of [8]) Let \(\mu \in \mathcal {ID}\). Then we have \(\mu \in {\mathcal {L}}(\gamma )\) if and only if \(\mu _1 \in {\mathcal {L}}(\gamma )\).
Watanabe and Yamamuro [15] proved the following.
Lemma 2.2
(Proposition 3.1 of [15]) Suppose that \(\mu \in \mathcal {ID}\). Then, we have \(\mu \in \mathcal {OS}\) if and only if there is \(n \in {\mathbb {N}}\) such that \(\nu _1^{n*} \in \mathcal {OS}\) and \(\overline{\mu _1^{t*}}(x) \asymp \overline{\nu _1^{n*}}(x)\) for any \(t >0\).
For \(\mu \in \mathcal {OS}\cap \mathcal {ID}\), define \(n_0 \in \mathbb N\) as
Lemma 2.3
Let \(\mu \in \mathcal {OS}\cap \mathcal {ID}\).
-
(i)
There exists \(C(a) >0\) such that, for all \( a \ge 0\) and all \( x >0\),
$$\begin{aligned} \overline{\nu _1^{n_0*} }(x-a)\le C(a) \overline{\nu _1^{n_0*} }(x). \end{aligned}$$ -
(ii)
There exists \(K >1\) such that, for all \(n \in {\mathbb {N}}\) and all \( x >0\),
$$\begin{aligned} \overline{\nu _1^{n*} }(x)\le K^n \overline{\nu _1^{n_0*} }(x). \end{aligned}$$
Proof
Assertion (i) is clear since \(\nu _1^{n_0*} \in \mathcal {OS}\subset \mathcal {OL}.\) We see from Proposition 2.4 of Shimura and Watanabe [11] that there exists \(K_1 >1\) such that, for all \(k \in {\mathbb {N}}\) and all \( x >0\),
Note that, for \(m \le n\),
Hence, we have, for \(0 \le j \le n_0-1\) and for all \(k \in \mathbb N\), with \(K=K_1^{2/n_0}>1\)
This inequality holds for \(k=0\) too. Thus, assertion (ii) is true. \(\square \)
Under the assumption that \(\zeta \in \mathcal {OS}\subset \mathcal {OL}\), we define the following. Let
Let \(\Lambda \) be the totality of increasing sequences \(\{\lambda _n\}_{n=1}^{\infty }\) with \(\lim _{n \rightarrow \infty } \lambda _n = \infty \) such that, for every \(x \in {\mathbb {R}}\), the following limit exists and is finite:
Define, for each sequence \(\{x_n\}_{n=1}^{\infty }\) with \(\lim _{n \rightarrow \infty } x_n = \infty \), \(T_n(y)\) as
Since \(\{T_n(y)\}_{n=1}^{\infty }\) is a sequence of increasing functions, uniformly bounded on every finite interval, by Helly’s selection principle, there exists an increasing subsequence \(\{\lambda _n\}\) of \(\{x_n\}\) with \(\lim _{n \rightarrow \infty } \lambda _n = \infty \) such that everywhere on \({\mathbb {R}}\) (2.2) holds. The limit function \(m(x;\{\lambda _n\})\) is increasing and is finite. That is, \(\{\lambda _n\} \in \Lambda \). It follows that, under the assumption that \(\zeta \in \mathcal {OS}\), there exists an increasing subsequence \(\{\lambda _n\} \in \Lambda \) of \(\{x_n\}\) for each sequence \(\{x_n\}_{n=1}^{\infty }\) with \(\lim _{n \rightarrow \infty } x_n = \infty \).
Lemma 2.4
Suppose that \(\zeta \in \mathcal {OS}\). Then, we have the following.
-
(i)
If \(\{\lambda _n\} \in \Lambda \), then \(\{\lambda _n-a\} \in \Lambda \) for every \(a \in {\mathbb {R}}\).
-
(ii)
For \(\{\lambda _n\} \in \Lambda \),
$$\begin{aligned} \int _{-\infty }^{\infty }m(x;\{\lambda _n\})\zeta (\hbox {d}x) < \infty \end{aligned}$$and
$$\begin{aligned} \lim _{a \rightarrow \infty }m(a;\{\lambda _n\}){{\bar{\zeta }}}(a)=0. \end{aligned}$$In particular, if \(\zeta \in \mathcal {OS}\cap {\mathcal {L}}(\gamma )\), then \(m(x;\{\lambda _n\})=e^{\gamma x}\) and \({{\widehat{\zeta }}}(\gamma ) < \infty .\)
Proof
We prove (i). Suppose that \(\{\lambda _n\} \in \Lambda . \) We have, for \(x,a \in {\mathbb {R}}\),
Thus, \(\{\lambda _n-a\} \in \Lambda . \) Next, we prove (ii). Let \(\rho \) be a distribution on \({\mathbb {R}}\). Note that, for \(x >2A,\)
We see from (2.3) that, for \(\{\lambda _n\} \in \Lambda \) and \(s >0\),
As \( s \rightarrow \infty \), we have
Since \(m(x;\{\lambda _n\})\) is increasing in x, we have
Hence, if \(\zeta \in \mathcal {OS}\cap {\mathcal {L}}(\gamma )\), then \(m(x;\{\lambda _n\})=e^{\gamma x}\) and \({{\widehat{\zeta }}}(\gamma ) < \infty .\) Thus, we have proved the lemma. \(\square \)
Pakes [8, 9] asserted and Watanabe [13] finally proved the following.
Lemma 2.5
(Theorem 1.1 of [13]) Let \(\gamma \ge 0\). Then \(\mu \in \mathcal {ID}\cap {\mathcal {S}}(\gamma )\) if and only if \(\nu _1 \in {\mathcal {S}}(\gamma )\).
Lemma 2.6
Let \(\gamma \ge 0\). Suppose that \(\rho \in {\mathcal {S}}_{\sharp }\).
-
(i)
If \({{\bar{\eta }}}(x) \asymp {{\bar{\rho }}}(x)\), then \(\eta \in {\mathcal {S}}_{\sharp }\).
-
(ii)
\(\rho \in {\mathcal {S}}(\gamma )\) if and only if \(\rho \in {\mathcal {L}}(\gamma )\).
Proof
Suppose that \(\rho \in {\mathcal {S}}_{\sharp }\). We prove (i). If \({{\bar{\eta }}}(x) \asymp {{\bar{\rho }}}(x)\), then there is \(C >0\) such that \({{\bar{\eta }}}(x)\le C {{\bar{\rho }}}(x)\) for \(x \in {\mathbb {R}}\). By using integration by parts in the second inequality, we obtain that
Thus, we see that
That is, \(\eta \in {\mathcal {S}}_{\sharp }\). Next we prove (ii). If \(\rho \in {\mathcal {S}}(\gamma )\), then clearly \(\rho \in {\mathcal {L}}(\gamma )\). Note that, for \(x >2A,\) (2.3) holds. If \(\rho \in {\mathcal {S}}_{\sharp }\cap {\mathcal {L}}(\gamma )\), then we have
Thus, we see that \(\rho \in {\mathcal {S}}(\gamma )\). \(\square \)
Watanabe [14] extended Wiener’s approximation theorem in [17] as follows.
Lemma 2.7
(Lemma 2.6 of Watanabe [14]) Let \(\xi \) be a finite measure on \({\mathbb {R}}\). The following are equivalent:
-
(1)
\(\xi \in {\mathcal {W}}.\)
-
(2)
If, for a bounded measurable function g(x) on \({\mathbb {R}}\),
$$\begin{aligned} \int _{-\infty }^{\infty }g(x-t)\xi (dt)=0\quad \hbox {for a.e.}\,\, \,x\in {\mathbb {R}}, \end{aligned}$$then \(g(x)=0\) for a.e. \(x\in {\mathbb {R}}.\)
3 Convolution Lemmas
In this section, we give important lemmas on convolutions.
Lemma 3.1
Let \(\gamma \ge 0\). Suppose that \(\zeta \in \mathcal {OS} \). For \(j=1,2\), let \(\rho _j\) be distributions on \({\mathbb {R}}_+\) satisfying
Let \(\{\lambda _n\} \in \Lambda . \)
-
(i)
Let \(\lambda _n > a+x\) and \(x >0\). We have, for every \(a \ge 0\),
$$\begin{aligned} e^{-\gamma a}\overline{\rho _1*\rho _2}(\lambda _n-a)-\overline{\rho _1*\rho _2}(\lambda _n)=:\sum _{j=1}^{4}I_j, \end{aligned}$$(3.2)where
$$\begin{aligned} I_1:= & {} -\int _{\lambda _n-a-x}^{\lambda _n-x}\overline{\rho _1}(\lambda _n-y)\rho _2(\hbox {d}y), \\ I_2:= & {} \overline{\rho _1}(x)(e^{-\gamma a}\overline{\rho _2}(\lambda _n-a-x)-\overline{\rho _2}(\lambda _n-x)), \\ I_3:= & {} \int _{0-}^{(\lambda _n-a-x)+}(e^{-\gamma a}\overline{\rho _1}(\lambda _n-a-y)-\overline{\rho _1}(\lambda _n-y))\rho _2(\hbox {d}y), \end{aligned}$$and
$$\begin{aligned} I_4:= & {} \int _{0-}^{x+}(e^{-\gamma a}\overline{\rho _2}(\lambda _n-a-y)-\overline{\rho _2}(\lambda _n-y))\rho _1(\hbox {d}y). \end{aligned}$$ -
(ii)
We have for \(j=1,2\)
$$\begin{aligned} \limsup _{x\rightarrow \infty }\limsup _{n \rightarrow \infty }\frac{|I_j|}{{{\bar{\zeta }}}(\lambda _n)}=0. \end{aligned}$$(3.3)
Proof
By using integration by parts, we have
and
Thus, assertion (i) is valid. We have by Lemma 2.4 for \(j=1,2\)
Lemma 3.2
Let \(\gamma \ge 0\). Suppose that \(\zeta \in \mathcal {OS} \). For \(j=1,2\), let \(\rho _j\) be distributions on \({\mathbb {R}}_+\) satisfying (3.1). Suppose further that for \(j=1,2\) and every \(a \ge 0\),
Then, for every \(a \ge 0\),
Proof. Let \(\{\lambda _n\} \in \Lambda . \) By the assumption for \(j=1\), there is \(\epsilon (x)>0 \) such that \(\epsilon (x) \rightarrow 0\) as \(x \rightarrow \infty \) and
for \(0 \le y \le \lambda _n-a-x.\) Thus, we have
As in the above argument, we have
Thus, by (3.2) and (3.3) of Lemma 3.1, we have proved (3.4). \(\square \)
Lemma 3.3
Let \(\gamma \ge 0\). Suppose that \(\zeta \in \mathcal {OS} \). For \(j=1,2\), let \(\rho _j\) be distributions on \({\mathbb {R}}_+\) satisfying (3.1). Suppose further that, for \(j=1,2,\) and for every \(a \ge 0\),
Then, we have, for every \(a \ge 0\),
Proof
Let \(\{\lambda _n\} \in \Lambda . \) Let \(\epsilon >0\) and \(a \ge 0\) be arbitrary, and let \( n \in {\mathbb {N}}\) and \(x \in (0,\lambda _n-a)\) be sufficiently large such that
for \(0 \le y \le \lambda _n-a-x\) and
for \(0 \le y \le x.\) By (3.2) and (3.3) of Lemma 3.1, we have only to prove that
We have
and
Thus, we see that
and
Since \(\epsilon >0\) is arbitrary, we established, for \(j =3,4\),
Thus, we have proved (3.6). \(\square \)
Lemma 3.4
Let \(\gamma \ge 0\). Suppose that \(\zeta \in \mathcal {OS}\cap {\mathcal {L}}(\gamma ) \). For \(j=1,2\), let \(\rho _j\) be distributions on \({\mathbb {R}}_+\) satisfying (3.1). Suppose further that, for every \(a \ge 0\),
and, for every \(a \ge 0\),
and that \(e^{\gamma x}\rho _1(\hbox {d}x) \in {\mathcal {W}}\). Then, we have, for every \(a \ge 0\),
Proof
Let \(\Lambda _2\) be the totality of increasing sequences \(\{\lambda _n\}_{n=1}^{\infty }\) with \(\lim _{n \rightarrow \infty } \lambda _n = \infty \) such that, for every \(x \in {\mathbb {R}}\), the following limit exists and is finite:
We have \(\Lambda _2\subset \Lambda \). As for \(\Lambda \), it follows that, under the assumption that \(\zeta \in \mathcal {OS}\) and \(\overline{\rho _2}(x) \le C_2{{\bar{\zeta }}}(x)\), there exists an increasing subsequence \(\{\lambda _n\} \in \Lambda _2 \) of \(\{x_n\}\) for each sequence \(\{x_n\}_{n=1}^{\infty }\) with \(\lim _{n \rightarrow \infty } x_n = \infty \). Let \(\{\lambda _n\} \in \Lambda _2. \) Recall from Lemma 2.4 that \(m(x;\{\lambda _n\}) =e^{\gamma x}\) and \({{\widehat{\zeta }}}(\gamma ) < \infty .\) As in the proof of Lemma 3.2, we have (3.5). We find that, for every \(a\in {\mathbb {R}}\),
Define \(M_2(y;\{\lambda _n\}):=e^{-\gamma y}m_2(y;\{\lambda _n\})\). Then \(M_2(y;\{\lambda _n\})\le C_2\) on \({\mathbb {R}}\). Note that
We see from (3.2), (3.3) of Lemma 3.1, (3.5), and (3.7) that, for every \(a\in {\mathbb {R}}\),
Thus, we obtain that, for every \(a, b \in {\mathbb {R}}\),
Since \(e^{\gamma y}\rho _1(\hbox {d}y) \in {\mathcal {W}}\), we find from Lemma 2.7 that, for every \(a\in {\mathbb {R}}\),
Since the function \(m_2(x;\{\lambda _n\})\) is increasing, the functions \(M_2(x+;\{\lambda _n\})\) and \(M_2(x-;\{\lambda _n\})\) exist for all \(x \in {\mathbb {R}}\). Taking \(b_n=b_n(a)\downarrow 0\) and \(b_n=b_n(a)\uparrow 0\), we have
As \(a\uparrow 0\) in the first equality, we see that
and hence, for every \(a\in {\mathbb {R}}\),
Since \(\{\lambda _n\} \in \Lambda _2 \) is arbitrary, we have (3.8). \(\square \)
Lemma 3.5
Let \(\gamma \ge 0\). Suppose that \(\zeta \in \mathcal {OS} \). For \(j=1,2\), let \(\rho _j\) be distributions on \({\mathbb {R}}_+\) satisfying (3.1). Suppose further that, for \(j=1,2,\) and for every \(a \ge 0\),
If we have, for every \(a \ge 0\),
then, for \(j=1,2,\) and for every \(a \ge 0\),
Proof
Suppose that, for some \(a >0\),
Then there is \(\{\lambda _n\} \in \Lambda \) such that, for some \(a >0\),
So there is \(\delta _1 >0\) such that, for some \(a > 0\),
Take \(y_0\) such that \(x>y_0 >\delta _1\) and \(\rho _1((y_0-\delta _1,y_0]) >0\). Let \(\lambda _n^{\prime }:=\lambda _n+y_0\) and \(a^{\prime }:=a+\delta _1\). Then we have
Let \(\lambda _n^{\prime } > a^{\prime }+x\) and \(x >0\). Define J as
Then we have as in assertion (i) of Lemma 3.1
where
and
For \(1\le j\le 3,\) let
and let
where
and
Then we have
As in the proof of Lemma 3.3, we see from the assumption and (3.9) that
This is a contradiction. Thus, we have, for every \(a \ge 0\),
By the analogous argument, we have for every \(a \ge 0\),
Thus, we have proved the lemma. \(\square \)
Lemma 3.6
Let \(\gamma \ge 0\). Let \(\rho \) be a distribution on \({\mathbb {R}}_+\). Suppose that \(\rho \in \mathcal {OS}\) and, for every \(a \ge 0\),
Then, for some positive integer \(n\ge 2\), \(\rho ^{n*} \in {\mathcal {L}}(\gamma )\) implies that \(\rho \in {\mathcal {L}}(\gamma )\).
Proof
Let \(\zeta :=\rho \). Then we see from Lemma 3.3 that, for every \(k \in {\mathbb {N}}\) and every \(a \ge 0\),
Thus, we find that \(\rho _1:=\rho \) and \(\rho _2:=\rho ^{(n-1)*}\) satisfy the assumptions of Lemma 3.5. Hence, we have by Lemma 3.5, for every \(a \ge 0\),
That is, \(\rho \in {\mathcal {L}}(\gamma )\). \(\square \)
Remark 3.1
For \(\gamma = 0\), the assumption (3.10) necessarily holds, but for \(\gamma > 0\), without the assumption (3.10) the lemma does not hold. For \(\gamma > 0\), Watanabe [14] made a distribution \(\eta \in \mathcal {OS}\) such that \(\eta ^{n*} \in {\mathcal {L}}(\gamma )\) for every \(n \ge 2\) but \(\eta \notin {\mathcal {L}}(\gamma )\).
4 Proof of Results
In this section, we prove the results stated in Sect. 1.
Lemma 4.1
Let \(\gamma \ge 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). If, for every \(a \ge 0\), (1.1) holds, then, for all \(n\in {\mathbb {N}}\) and every \(a \ge 0\),
and we have \(T(\mu ,\gamma )=(0,\infty )\).
Proof
By induction, we see from Lemma 3.2 that if (1.1) holds for every \(a \ge 0\), then, for all \(n\in {\mathbb {N}}\) and every \(a \ge 0\), we have (4.1). We have with \(c:={{\bar{\nu }}}(1)\), for \(t>0\),
Suppose that, for all \(n\in {\mathbb {N}}\) and every \(a \ge 0\), (4.1) holds. Let \(\epsilon > 0\) be arbitrary. By Lemma 2.3, we can choose sufficiently large \(N \in {\mathbb {N}}\) such that, for \(\epsilon > 0\),
We find from (4.1) that, for every \(a \ge 0\),
Thus, we see that, for every \(a \ge 0\) and for every \(t>0\),
Since \(\overline{\mu _1^{t*}}(x)\asymp \overline{\nu _1^{n_0*}}(x)\) for every \(t>0\), we have \(T(\mu ,\gamma )=(0,\infty )\). \(\square \)
Lemma 4.2
Let \(\gamma \ge 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). If 0 is a limit point of \(T(\mu ,\gamma )\), then, for every \(a \ge 0\), (1.1) holds.
Proof. Suppose that 0 is a limit point of \(T(\mu ,\gamma )\). Then, there exists a strictly decreasing sequence \(\{t_n\}_{n=1}^{\infty }\) in \(T(\mu ,\gamma )\) converging to 0 as \(n \rightarrow \infty \). We have with \(c:={{\bar{\nu }}}(1)\)
Since \(\{t_n\}_{n=1}^{\infty }\) in \(T(\mu ,\gamma )\) and \(\overline{\mu _1^{t_n*}}(x) \asymp \overline{\nu _1^{n_0*}}(x)\) from Lemma 2.2, we see that, for every \(a \ge 0\),
Thus, we obtain from Lemma 2.3 that, for every \(a \ge 0\),
Thus, we have (1.1) for every \(a \ge 0\). \(\Box \)
Lemma 4.3
Let \(\gamma \ge 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). If \(t_0, t_1 \in T(\mu ,\gamma )\) with \(t_1 >t_0\), then \(t_1-t_0 \in T(\mu ,\gamma )\). If \(T(\mu ,\gamma )\) has a limit point, then \(T(\mu ,\gamma ) =(0,\infty )\). If \(T(\mu ,\gamma )\) has the minimum \(a_0 >0\), then \(T(\mu ,\gamma ) =a_0{\mathbb {N}}\).
Proof
Suppose that \(t_0, t_1 \in T(\mu ,\gamma )\) with \(t_1 >t_0\). Let \(\zeta :=\rho _1:=\mu ^{t_0*}\) and \(\rho _2:=\mu ^{(t_1-t_0)*}\). The distribution \(e^{\gamma x}\rho _1(\hbox {d}x)/{{\widehat{\rho }}}_1(\gamma )\) is an exponentially tilted infinitely divisible distribution and hence itself is infinitely divisible, thus having a non-vanishing characteristic function. That is, \(e^{\gamma x}\rho _1(\hbox {d}x) \in {\mathcal {W}}\). See (iii) of Theorem 25.17 of Sato [10]. Thus, we see from Lemma 3.4 that \(\mu ^{(t_1-t_0)*}\in {\mathcal {L}}(\gamma )\). Thus, if \(T(\mu ,\gamma )\) has a limit point, then 0 is a limit point of \(T(\mu ,\gamma )\), and hence, by Lemmas 4.1 and 4.2, \(T(\mu ,\gamma ) =(0,\infty )\). If \(T(\mu ,\gamma )\) has the minimum \(a_0 >0\), then clearly \(a_0{\mathbb {N}}\subset T(\mu ,\gamma )\) and \( T(\mu ,\gamma ){\setminus } a_0{\mathbb {N}} =\emptyset \). \(\square \)
Proof of Theorem 1.1
Assertion (i) is clear from Lemmas 4.1, 4.2, and 4.3. The first part of assertion (ii) is due to Lemmas 4.1 and 4.2. Suppose that \(\mu \in {\mathcal {A}}(\gamma )\). If \(n < n_0,\) then \(\nu _1^{n*} \not \in \mathcal {OS}\) simply because of the definition of \(n_0\). If \(n \ge n_0\) and x is large, then \(\overline{\nu _1^{n*}}(x) \ge \overline{\nu _1^{n_0*}}(x)\), and hence, (4.1) implies that \(\nu _1^{n*} \in {\mathcal {L}}(\gamma )\). \(\square \)
Proof of Corollary 1.1
Suppose that \({\mathcal {C}}(\gamma )\) is not empty. Then there is the minimum \(a_0 >0\) in \(T(\mu ,\gamma )\) for \(\mu \in {\mathcal {C}}(\gamma )\). Since \(a_0>0\) is a period of \(T(\mu ,\gamma )\), for \(n=2\), \(\mu ^{a_0*}=(\mu ^{(a_0/n)*})^{n*} \in {\mathcal {L}}(\gamma )\) but \((\mu ^{(a_0/n)*})^{(n+1)*} \notin {\mathcal {L}}(\gamma )\) and \(\mu ^{(a_0/n)*} \notin {\mathcal {L}}(\gamma )\). Thus, Hypotheses I and II are not true. Suppose that \({\mathcal {C}}(\gamma )\) is empty. Then, obviously, Hypotheses I and II are true. Thus, (1), (2), and (3) are equivalent. We prove the equivalence of (3) and (4). Suppose that \({\mathcal {C}}(\gamma )\) is empty. Then for every \(\mu \in \mathcal {OS}\cap \mathcal {ID}\) it holds that, for every \(2t \in T(\mu ,\gamma )\), \(\mu _1^{t*}\in {\mathcal {L}}(\gamma )\), and hence, for all \(a \ge 0\), (1.2) holds. Conversely, suppose that \({\mathcal {C}}(\gamma )\) is not empty and, for \(a_0=2t \in T(\mu ,\gamma )\) with \(\mu \in {\mathcal {C}}(\gamma )\) and for all \(a \ge 0\), (1.2) holds. Letting \(\rho _1:=\rho _2:=\mu _1^{t*}\), \(\zeta :=\mu _1^{2t*}\), define \(\Lambda _2\) as in Lemma 3.4 and let \(\{\lambda _n\} \in \Lambda _2\subset \Lambda \). We have (3.3) by Lemma 3.1 for \(j=1,2\). We have \(I_3+I_4=2I_4+I_5\), where
We have by the assumption (1.2) for every \(a \ge 0\)
Define \(M_2(y;\{\lambda _n\}):=e^{-\gamma y}m_2(y;\{\lambda _n\})\). Thus, we find from (3.2), (3.3), and \(2t \in T(\mu ,\gamma )\) that, for every \(a \ge 0\),
The distribution \(e^{\gamma x}\rho _1(\hbox {d}x)/{{\widehat{\rho }}}_1(\gamma )\) is an exponentially tilted infinitely divisible distribution and hence itself is infinitely divisible, thus having a non-vanishing characteristic function. That is,
As in the proof of Lemma 3.4, we have \(\rho _2=\mu _1^{t*}\in {\mathcal {L}}(\gamma )\). This is a contradiction. Thus, (3) and (4) are equivalent. \(\square \)
Proof of Remark 1.1
Let \(\gamma =0\). Then we see from Lemma 3.6 that Hypothesis II is true. Thus, \({\mathcal {C}}(0)\) is empty, and hence, Remark 1.1 follows from Theorem 1.1. \(\square \)
Proof of Proposition 1.1
Let \(\gamma > 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). Suppose that (1.3) holds for every \(a \ge 0\). Let \(\zeta :=\nu _1^{n_0*}\). Then, by induction, we see from (1.3) and Lemma 3.3 that, for every \(n \in {\mathbb {N}}\) and every \(a \ge 0\),
Let \(\epsilon >0\) be arbitrary. Thus, letting \(N \in {\mathbb {N}}\) sufficiently large, we have, for every \(t >0\) and for every \(a \ge 0\),
Since \(\epsilon >0\) is arbitrary and \(\overline{\nu _1^{n_0*}}(x)\asymp \overline{\mu _1^{(t/n)*}}(x)\) for every \( n \in {\mathbb {N}}\), we obtain that \(\rho :=\mu _1^{(t/n)*}\) satisfies \(\rho \in \mathcal {OS}\) and (3.10) holds. Hence, we find from Lemma 3.6 that if \(t \in T(\mu ,\gamma )\), then \(t/n \in T(\mu ,\gamma )\) for every \( n \in {\mathbb {N}}\). Thus, by Lemmas 4.1 and 4.2, either \(T(\mu ,\gamma )= (0,\infty )\) or \(\emptyset \). \(\square \)
Proof of Proposition 1.2
Suppose that \(\nu _1^{2*} \in {\mathcal {L}}(\gamma )\) and the real part of \({{\widehat{\nu }}}_1(\gamma +iz)\) is not 0 for every \(z \in {\mathbb {R}}\). If \(t \in T(\mu ,\gamma )\), then
Define distributions \(\eta _1\) and \(\eta _2\) on \({\mathbb {R}}_+\) as
and
We see from Proposition 3.1 of Shimura and Watanabe [11] that \(\eta _j\in \mathcal {OS}\) and \(\overline{\eta _j}(x)\asymp \overline{\nu _1^{n_0*}}(x)\) for \(j=1,2\). Let \(\epsilon >0\) be arbitrary. We obtain from Lemma 2.3 that there is a positive integer \(N=N(a,\epsilon ,t)\) such that
Since \(\nu _1^{(2k)*}\in {\mathcal {L}}(\gamma )\) for every \(k \ge 0\), we have, for every \(a \ge 0\) and every \(t >0\),
Thus, with some \(C=C(t) >0\) we have, for every \(a \ge 0\) and every \(t >0\),
Since \(\epsilon >0\) is arbitrary, we have
Since
Let \(\zeta :=\rho _1:=\eta _2\) and \(\rho _2:=\nu _1.\) Then, by argument similar to the proof of (4.3),
Since the real part of \({{\widehat{\nu }}}_1(\gamma +iz)\) is not 0 for every \(z \in {\mathbb {R}}\),
for every \(z \in {\mathbb {R}}\), that is, \(e^{\gamma x}\rho _1(\hbox {d}x) \in {\mathcal {W}}\). Thus, we see from Lemma 3.4 that
Since \({{\bar{\zeta }}}(x)\asymp \overline{\nu _1^{n_0*}}(x)\), we see from Theorem 1.1 that \(T(\mu ,\gamma )=(0,\infty )\). Thus, we have proved the proposition. \(\square \)
Proof of Proposition 1.3
Let \(\gamma > 0\) and \(\mu \in \mathcal {OS}\cap \mathcal {ID}\). Suppose that \(\nu _1^{n_1*} \in {\mathcal {S}}_{\sharp }\). Since \(\overline{\mu ^{t*}}(x)\asymp \overline{\nu _1^{n_1*} }(x)\), we have \(\mu ^{t*}\in {\mathcal {S}}_{\sharp }\) for every \(t >0\). Thus, we see from Lemmas 2.5 and 2.6 that if \( T(\mu ,\gamma )\ne \emptyset \), then \(\nu _1 \in {\mathcal {S}}(\gamma )\) and hence \(T(\mu ,\gamma )=(0,\infty )\). That is, either \(T(\mu ,\gamma )= (0,\infty )\) or \(\emptyset \). Moreover, \(T(\mu ,\gamma )=(0,\infty )\) if and only if \(\nu _1 \in {\mathcal {S}}(\gamma )\). \(\square \)
References
Cui, Z., Wang, Y., Xu, H.: Some positive conclusions related to the Embrechts–Goldie conjecture. arXiv:1609.00912 (2016)
Embrechts, P., Goldie, C.M.: On closure and factorization properties of subexponential and related distributions. J. Aust. Math. Soc. A 29, 243–256 (1980)
Embrechts, P., Goldie, C.M.: On convolution tails. Stoch. Process. Appl. 13, 263–278 (1982)
Embrechts, P., Goldie, C.M., Veraverbeke, N.: Subexponentiality and infinite divisibility. Z. Wahrscheinlichkeitstheorie Verw. Gebiet. 49, 335–347 (1979)
Klüppelberg, C.: Asymptotic ordering of distribution functions on convolution semigroup. Semigroup Forum. 40, 77–92 (1990)
Klüppelberg, C., Villasenor, J.A.: The full solution of the convolution closure problem for convolution-equivalent distributions. J. Math. Anal. Appl. 160, 79–92 (1991)
Leslie, J.R.: On the non-closure under convolution of the subexponential family. J. Appl. Probab. 26, 58–66 (1989)
Pakes, A.G.: Convolution equivalence and infinite divisibility. J. Appl. Probab. 41, 407–424 (2004)
Pakes, A.G.: Convolution equivalence and infinite divisibility: corrections and corollaries. J. Appl. Probab. 44, 295–305 (2007)
Sato, K.: Lévy Processes and Infinitely Divisible Distributions. Cambridge Studies in Advanced Mathematics, vol. 68. Cambridge University Press, Cambridge (2013)
Shimura, T., Watanabe, T.: Infinite divisibility and generalized subexponentiality. Bernoulli 11, 445–469 (2005)
Shimura, T., Watanabe, T.: On the convolution roots in the convolution equivalent class. The Institute of Statistical Mathematics Cooperative Research Report 175, pp. 1–15 (2005)
Watanabe, T.: Convolution equivalence and distributions of random sums. Probab. Theory Relat. Fields 142, 367–397 (2008)
Watanabe, T.: The Wiener condition and the conjectures of Embrechts and Goldie. Ann. Probab. 47, 1221–1239 (2019)
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. 15, 44–74 (2010)
Watanabe, T., Yamamuro, K.: Two non-closure properties on the class of subexponential densities. J. Theor. Probab. 30, 1059–1075 (2017)
Wiener, N.: Tauberian theorems. Ann. Math. (2) 33, 1–100 (1932)
Xu, H., Foss, S., Wang, Y.: Convolution and convolution-root properties of longtailed distributions. Extremes 18, 605–628 (2015)
Xu, H., Wang, Y., Cheng, D., Yu, C.: On the closure under convolution roots related to an infinitely divisible distribution in the distribution class \({\cal{L}}(\gamma )\). arXiv:1512.01792 (2015)
Acknowledgements
The author is grateful to the referee for his careful reading the manuscript and helpful advice.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Watanabe, T. Two Hypotheses on the Exponential Class in the Class Of O-subexponential Infinitely Divisible Distributions. J Theor Probab 34, 852–873 (2021). https://doi.org/10.1007/s10959-019-00976-z
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-019-00976-z