Abstract
We prove under general conditions that a trimmed subordinator satisfies a self-standardized central limit theorem (SSCLT). Our basic tool is a powerful distributional approximation result of Zaitsev (Probab Theory Relat Fields 74:535–566, 1987). Among other results, we obtain as special cases of our subordinator result the recent SSCLTs of Ipsen et al. (Stoch Process Appl 130:2228–2249, 2020) for trimmed subordinators and a trimmed subordinator analog of a central limit theorem of Csörgő et al. (Probab Theory Relat Fields 72:1–16, 1986) for intermediate trimmed sums in the domain of attraction of a stable law. We then use our methods to prove a similar theorem for general Lévy processes.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
We shall begin by stating our results for trimmed subordinators. Special cases of our main result for subordinators, Theorem 1 below, have already been proved by Ipsen, Maller and Resnick (IMR) [6] , using classical methods. See, in particular, their Theorem 4.1. Our approach is based on a powerful distributional approximation result of Zaitsev [11], which we shall see in Sect. 5 extends to general trimmed Lévy processes. We shall first establish some basic notation.
Let \(V_{t}\), \(t\ge 0\), be a subordinator with Lévy measure \(\varLambda \) on \( {\mathbb {R}}^{+}=\left( 0,\infty \right) \) and drift 0. Define the tail function \({\overline{\varLambda }}(x)=\varLambda ((x,\infty ))\), for \(x>0\), and for \(u>0\) let
where \(\sup \varnothing :=0\).
Remark 1
For later use, we observe that we always have
Notice that (2) is formally true if \({\overline{\varLambda }}(0+)=c>0,\) since in this case for all \(u>c\), \(\sup \{x:{\overline{\varLambda }} (x)>u\}=\varnothing \) and we define \(\sup \varnothing :=0\), and thus \( \varphi (u)=0\) for \(u>c\). The limit (2) also holds whenever
To see this, assume (3) and choose any sequence \(x_{n}\searrow 0\) such that \(u_{n}:={\overline{\varLambda }}(x_{n})>0\) for \(n\ge 1.\) Clearly, \(u_{n} \rightarrow \infty \) as \(n\rightarrow \infty \). By the definition (1), the fact that \({\overline{\varLambda }}\) is nonincreasing on \(\left( 0,\infty \right) \) and \(x_{n}\notin \{x:{\overline{\varLambda }}(x)>u_{n}\}\) necessarily \(\varphi (u_{n})\le x_{n}\), and thus since \(\varphi \) is nonincreasing, (2) holds. Furthermore, when (3) holds,
To verify this, choose \(0<y_{n+1}<y_{n}\) such that \(y_{n}\searrow 0\), as \(n\rightarrow \infty \), and \(v_{n+1}={\overline{\varLambda }}(y_{n+1})>v_{n} ={\overline{\varLambda }}(y_{n})\) for \(n\ge 1.\) Therefore, \(y_{n+1}\in \{x:\overline{\varLambda }(x)>v_{n}\}\) and hence \(\varphi (v_{n})\ge y_{n+1}>0\) for all \(n\ge 1\). Since \(v_{n}\nearrow \infty \), we have (4).
Recall that the Lévy measure of a subordinator satisfies
The subordinator \(V_{t}\), \(t\ge 0\), has Laplace transform
where
which can be written after a change of variable
For any \(t>0\) denote the ordered jump sequence \(m_{t}^{\left( 1\right) }\ge m_{t}^{\left( 2\right) }\ge \cdots \) of \(V_{t}\) on the interval \(\left[ 0,t \right] \). Let \(\omega _{1},\omega _{2},\ldots \) be i.i.d. exponential random variables with parameter 1 and for each \(n\ge 1\) let \(\varGamma _{n}=\omega _{1}+\cdots +\omega _{n}\). It is well known that for each \(t>0\)
See, for instance, equation (1.3) in IMR [6] and the references therein. It can also be inferred from a general representation for subordinators due to Rosiński [9].
Set \(V_{t}^{\left( 0\right) }:=V_{t}\) and for any integer \(k\ge 1\) consider the trimmed subordinator
which on account of (8) says for any integer \(k\ge 1\) and \(t>0\)
Set for any \(y>0\)
We see by Remark 1 that (3) implies that
Throughout these notes, Z, \(Z_{1},Z_{2}\) denote standard normal random variables. Here is our self-standardized central limit theorem (SSCLT) for trimmed subordinators. In Examples 4 and 5 we show that our theorem implies Theorem 4.1 and Remark 4.1 of IMR [6], who treat the case when \( t_{n}=t\) is fixed and \(k_{n}\rightarrow \infty \).
Theorem 1
Assume that \({\overline{\varLambda }}(0+)=\infty \). For any sequence of positive integers \(\left\{ k_{n}\right\} _{n\ge 1}\) converging to infinity and sequence of positive constants \( \left\{ t_{n}\right\} _{n\ge 1}\) satisfying
we have uniformly in x, as \(n\rightarrow \infty \),
which implies as \(n\rightarrow \infty \)
Corollary 1
Assume that \(V_{t}\), \(t\ge 0\), is a subordinator with drift 0, whose Lévy tail function \({\overline{\varLambda }}\) is regularly varying at zero with index \(-\alpha \), where \( 0<\alpha <1\). For any sequence of positive integers \(\left\{ k_{n}\right\} _{n\ge 1}\) converging to infinity and sequence of positive constants \(\left\{ t_{n}\right\} _{n\ge 1}\) satisfying \( k_{n}/t_{n}\rightarrow \infty \), we have as \(n\rightarrow \infty \),
Remark 2
Notice that whenever
\(\varGamma _{k_{n}}/t_{n}\overset{\mathrm {P}}{\rightarrow }\infty \) and \( k_{n}\rightarrow \infty \), then
and thus (12) holds. In particular, (16) is satisfied whenever \(\varphi \) is regularly varying at infinity with index \(-1/\alpha \) with \(0<\alpha <2\).
Using the change of variables formula: For \(p\ge 1\), whenever the integrals exist, for \(r>0\),
(for (17), see p. 301 of Brémaud [3]) one readily sees that (16) is fulfilled whenever the Feller class at zero condition holds (e.g., Maller and Mason [8]):
(For more details, refer to Example 2.)
Remark 3
Corollary 1 implies part of Theorem 9.1 of IMR [6], namely, whenever for \(0<\alpha <1\),
then for each fixed \(t>0\), as \(n\rightarrow \infty \),
The first part of their Theorem 9.1 can be shown to be equivalent to (19).
Remark 4
The analog of Corollary 1 for a sequence of i.i.d. positive random variables \(\xi _{1},\xi _{2}\dots \) in the domain of attraction of a stable law of index \(0<\alpha <2\) says that as \( n\rightarrow \infty \),
where for each \(n\ge 2\), \(\xi _{n}^{\left( 1\right) }\ge \dots \ge \xi _{n}^{\left( n\right) }\) denote the order statistics of \(\xi _{1},\dots ,\xi _{n}\), \(\left\{ r_{n}\right\} _{n\ge 1}\) is a sequence of positive integers \(1\le r_{n}\le n\) satisfying \(r_{n}\rightarrow \infty \) and \( r_{n}/n\rightarrow 0\) as \(n\rightarrow \infty \), and \(c\left( r_{n}/n\right) \) and \(a\left( r_{n}/n\right) \) are appropriate centering and norming constants. For details refer to S. Csörgő, Horváth and Mason [4]. The proof of our Corollary 1 borrows ideas from the proof of their Theorem 1.
2 Preliminaries for Proofs
In this section, we collect some facts that are needed in our proofs. Lemmas 1 and 2 are elementary; however, for completeness we indicate proofs.
2.1 A Useful Special Case of a Result of Zaitsev [11]
We shall be making use of the following special case of Theorem 1.2 of Zaitsev [11]. which in this paper we shall call the Zaitsev Fact.
Fact (Zaitsev [11]) Let Y be an infinitely divisible mean 0 and variance 1 random variable with Lévy measure \(\varLambda \) and Z be a standard normal random variable. Assume that the support of \(\varLambda \) is contained in a closed ball with center 0 of radius \(\tau >0\), then for universal positive constants \(C_{1}\) and \(C_{2}\) for any \(\lambda >0\)
where
with \(B^{\lambda }=\left\{ y\in {\mathbb {R}}\text {:}\inf _{x\in B}\left| x-y\right| <\lambda \right\} \) for \(B\in {\mathcal {B}}\), the Borel sets of \({\mathbb {R}}\).
Notice that under the conditions of the Zaitsev Fact for all x, \(\lambda >0\) and \(\varepsilon >\varPi \left( Y,Z;\lambda \right) \), with \(\varepsilon >0\),
and
and thus
In particular, the Zaitsev Fact says that for all \(x\in {\mathbb {R}}\) and \(\lambda >0\),
2.2 Moments of a Positive Random Variable
Given \(t>0\), let \(X_{t}\) be a positive random variable with Laplace transform
where \(\varPhi \) is the Laplace exponent
and \(\varphi \) a nonincreasing positive function on \(\left( 0,\infty \right) \) such that \(\varphi \left( u\right) \rightarrow 0\) as \(u\rightarrow \infty \) . Assume that
which implies \(\varPhi \left( \lambda \right) <\infty \) for all \(\lambda >0\) and \(\varPhi \left( \lambda \right) \) twice differentiable on \(\left( 0,\infty \right) .\) Differentiating \(\varPsi _{X_{t}}\left( \lambda \right) \) with respect to \(\lambda \) twice and evaluating \(\varPsi _{X_{t}}^{\prime }\left( 0+\right) \) and \(\varPsi _{X}^{^{\prime \prime }}\left( 0+\right) \), we get the following moments:
Lemma 1
Under the above assumptions,
2.3 An Asymptotic Independence Result
We shall need the following elementary asymptotic independence result.
Lemma 2
Let \(\left( X_{n},Y_{n}\right) _{n\ge 1}\) be a sequence of pairs of real-valued random variables on the same probability space, and for each \(n\ge 1\) let \(\phi _{n}\) be a measurable function. Suppose that for distribution functions F and G for all continuity points x of F and y of G
then
Proof
Notice that
which by (20) converges to zero. \(\square \)
3 Proof of Subordinator Results
3.1 Proof of Theorem 1
For each \(t>0\) and \(y>0\), consider the random variable
with \(\left( \varGamma _{i}^{\prime }\right) _{i\ge 1}\) \(\overset{\mathrm {D}}{= }\) \(\left( \varGamma _{i}\right) _{i\ge 1},\) which has Laplace transform
where \(\varPhi _{t,y}\left( \lambda \right) \) is the Laplace exponent,
Introducing the Lévy measure \(\varLambda _{t,y}\) defined on \(\left( 0,\infty \right) \) by the tail function
we see that
and thus
Clearly, \(T\left( t,y\right) \) is an infinitely divisible random variable and the support of \(\varLambda _{y/t}\) is contained in \(\left[ 0,\varphi (y/t) \right] \). Applying Lemma 1, one finds that
and
Note that (3) implies (11) and thus for all \(y>0\), \( \sigma ^{2}\left( \frac{y}{t}\right) >0\). For each \(t>0\) and \(y>0\), consider the standardized version of \(T\left( t,y\right) \)
We can write
Now \(S\left( t,y\right) \) is an infinitely divisible random with
whose Lévy measure has support contained in \(\left[ 0,\varphi (y/t)/\left( \sqrt{t}\sigma \left( \frac{y}{t}\right) \right) \right] \). Applying the Zaitsev Fact to the infinitely divisible random variable \(S\left( t,y\right) \), we get for any \(t>0,\) \(y>0\) and \(\lambda >0\) and for universal positive constants \(C_{1} \) and \(C_{2}\)
This implies that whenever \(\left\{ t_{n}\right\} _{n\ge 1}\) is a sequence of positive constants and \(Y_{k_{n}}\) is a sequence of positive random variables such that each \(Y_{k_{n}}\) is independent of \( \left( \varGamma _{i}^{\prime }\right) _{i\ge 1}\) and
then uniformly in x
and thus we have
By choosing \(Y_{k_{n}}=\varGamma _{k_{n}}\) and independent of \( \left( \varGamma _{i}^{\prime }\right) _{i\ge 1},\) with \(\left( \varGamma _{i}^{\prime }\right) _{i\ge 1}\overset{\mathrm {D}}{=}\) \(\left( \varGamma _{i}\right) _{i\ge 1}\), we get by (10) that
Keeping (12) in mind, (13) and (14) follow from (23) and (24), respectively. \(\square \)
3.2 Proof of Corollary 1
The proof will be a consequence of Theorem 1 and Lemma 2. Note that \(V_{t}\) has Laplace transform
of the form given by (6). Since \({\overline{\varLambda }}\) is assumed to be regularly varying at 0 with index \(-\alpha \), \(0<\alpha <1\), the \( \varphi \) in (7) is regularly varying at \(\infty \) with index \( -1/\alpha \) and thus for \(x>0\),
where \(L\left( x\right) \) is slowly varying at infinity. This implies that as \(z\rightarrow \infty ,\)
and
where \(a_{\alpha }=\alpha /\left( 1-\alpha \right) \) and \(b_{\alpha }^{2}=\alpha /\left( 2-\alpha \right) \).
With this notation, we can write
which equals
Claim 1
As \(n\rightarrow \infty \),
Proof
This follows from the fact that \(\varGamma _{k_{n}}/k_{n}\overset{ \mathrm {P}}{\rightarrow }1\), \(k_{n}/t_{n}\rightarrow \infty \) and \(\sigma \left( z\right) \) is regularly varying at \(\infty \) with index \(-1/\alpha +1/2.\) \(\square \)
Claim 2
As \(n\rightarrow \infty \),
Proof
This is a consequence of \(k_{n}/t_{n}\rightarrow \infty \) combined with (25) and (27), which together say
and
\(\square \)
Claim 3
As \(n\rightarrow \infty \),
Proof
Since
for any \(0<\varepsilon <1\) there exists a \(c>0\) such that
for all large enough n. When \(\varGamma _{k_{n}}\in \left[ k_{n}-c\sqrt{k_{n}} ,k_{n}-c\sqrt{k_{n}}\right] \),
Now for any \(\lambda >1\), for all large enough n
which converges to
Since \(\lambda >1\) can be made arbitrarily close to 1 and \(\varepsilon >0\) can be chosen arbitrarily close to 0, we see using Claim 2 that Claim 3 is true. \(\square \)
Putting everything together, keeping (29) in mind, we conclude that as \( n\rightarrow \infty \),
Choose \(Y_{k_{n}}=\varGamma _{k_{n}}\) and independent of \(\left( \varGamma _{i}^{\prime }\right) _{i\ge 1}\) \(\overset{\mathrm {D}}{=}\left( \varGamma _{i}\right) _{i\ge 1}\). We get by Remark 2 that (12) holds, which implies (13). Thus, by (13) and Lemma 2, for independent standard normal random variables \(Z_{1}\) and \(Z_{2}\), as \( n\rightarrow \infty \)
Noting that \(\sigma \left( \frac{Y_{k_{n}}}{t_{n}}\right) /\sigma \left( \frac{ k_{n}}{t_{n}}\right) \overset{\mathrm {P}}{\rightarrow }1\) and \(Z_{1}+\sqrt{ \frac{2-\alpha }{\alpha }}Z_{2}\overset{\mathrm {D}}{=}\sqrt{\frac{2}{\alpha }} Z \), we get as \(n\rightarrow \infty \),
which since
gives (15). \(\square \)
4 Examples of Theorem 1
In the following examples, we always assume that (3) holds.
Example 1
There always exist \(k_{n}\rightarrow \infty \) and \(t_{n}\rightarrow \infty \) such that (12) holds. For example for any \(k_{n}\rightarrow \infty \), let \(t_{n}=\rho k_{n}\) for some \(\rho >0. \) Since \(\varGamma _{k_{n}}/k_{n}\overset{\mathrm {P}}{\rightarrow }1\), \( \varGamma _{k_{n}}/t_{n}\overset{\mathrm {P}}{\rightarrow }1/\rho \), which implies that
and thus (12) holds and hence by Theorem 1, we conclude ( 13) and (14).
Example 2
Assume the Feller class at zero condition (18). Noting that \({\overline{\varLambda }}(\varphi (y)-)\ge y\), we get from (18) that
which says
This implies that
Therefore, as in Remark 2, we see that if \(\varGamma _{k_{n}}/t_{n}\overset{ \mathrm {P}}{\rightarrow }\infty \) and \(k_{n}\rightarrow \infty \), then (12) holds and thus by Theorem 1, we infer (13) and (14).
Example 3
Let
Clearly, \(\varphi (u)=\exp \left( -u\right) \), \(0<u<\infty \), and for \(0<x<1\)
which \(\nearrow \infty \), as \(x\searrow 0\). Thus, the Feller class at zero condition does not hold. However, the domain of attraction to normal at infinity condition holds (e.g., Doney and Maller [5] and Maller and Mason [7]), since for all \(x\ge 1\)
In this example for all \(y>0\) and \(t>0\),
Thus, for any sequence of positive integers \(k_{n}\rightarrow \infty \) and sequence of positive constants \(t_{n}\rightarrow \infty \)
which says that (12) is satisfied and hence by Theorem 1, ( 13) and (14) hold.
Next we show that as a special case of Theorem 1, we get Theorem 4.1 and Remark 4.1 of IMR [6], who consider the case when \(t_{n}=t\) is fixed and \(k_{n}\rightarrow \infty .\) Their Theorem 4.1 and Remark 4.1 say that whenever there exist constants \(a_{n}\) and \(b_{n}\) such that for a nondegenerate random variable \(\varDelta \)
then for all \(t>0\) the following self-standardized trimmed central limit theorem (CLT) holds
Remark 5
We should note that in the statements of Theorem 4.1 and Remark 4.1 of IMR [6], “ \(\mu \)” should be “\(t\mu \)” , and, in equation (4.2), “ \(\lim _{r\rightarrow \infty }\) ” should be removed and “\(=\varPhi \left( x\right) \), \(x\in {\mathbb {R}}\).” should be replaced by “\(\Rightarrow \varPhi \left( x\right) \), \(x\in {\mathbb {R}}\), as \(r\rightarrow \infty .\)”
IMR [6] have shown in their Theorem 2.1 that for (31) to hold it is necessary and sufficient that there exist functions \(a\left( r\right) \) and \(b\left( r\right) \) of \(r>0\) such that whenever \(a\left( r\right) x+b\left( r\right) >0\)
where \(h\left( x\right) \in {\mathbb {R}}\) is a nondecreasing function having the form for some \(\gamma \le 0\),
In which case \(P\left\{ \varDelta \le x\right\} =P\left\{ Z\le h\left( x\right) \right\} \).
The next two examples show that whenever (31) holds and hence (32) with \(h\left( x\right) \) as in (33) is satisfied, then special cases of condition (12) are fulfilled. Example 4 treats the case when \(\gamma <0\) in (33), and Example 5 considers the case when \( \gamma =0\) in (33).
Example 4 [The case \(\gamma <0\) in (33)] From Proposition 4.1 of IMR [6], we get that whenever (31) holds and we have (33) for some \(\gamma <0\) then
and \({\overline{\varLambda }}(x)\) is slowly varying at 0. Since \(\varphi \left( z\right) \searrow 0\) as \(z\nearrow \infty \), this implies that as y/t converges to \(\infty \),
which by (3), for each fixed \(t>0\), converges to infinity as \( y\rightarrow \infty \). We readily see then that (12) is satisfied, whenever \(k_{n}\rightarrow \infty \) and \(t_{n}=t>0\), fixed, as \(n\rightarrow \infty \), and thus by Theorem 1, (13) and (14) hold. Notice that a Lévy measure that satisfies (34) is not in the Feller class at zero.
Example 5 [The case \(\gamma =0\) in (33)] Using the notation from Proposition 4.2 of IMR [6], set
Proposition 4.2 of IMR [6] says when \(\gamma =0\) in (33) that for a function \(\pi _{2}\)
which from (4.13) in IMR [6] satisfies
This implies that as y/t converges to \(\infty \) and ty is bounded away from 0, then
Thus, if \(\varGamma _{k_{n}}/t_{n}\overset{\mathrm {P}}{\rightarrow }\infty \) and for some \(\varepsilon >0\), \({\mathbb {P}}\left\{ t_{n}\varGamma _{k_{n}}>\varepsilon \right\} \rightarrow 1\), then (12) is fulfilled and hence by Theorem 1, (13) and (14) hold. In particular, this is satisfied when \(k_{n}\rightarrow \infty \) and \( t_{n}=t>0\), fixed, as \(n\rightarrow \infty \).
5 A SSCLT for a Trimmed Lévy Process
Before we can talk about a SSCLT for a trimmed Lévy process, we must first establish a pointwise representation for the Lévy process that we shall consider, as well as some necessary notation and auxiliary results needed to define what we mean by a trimmed Lévy process and to prove a SSCLT for it.
5.1 A Pointwise Representation for the Lévy Process
Let \((\varOmega ,{\mathcal {F}},{\mathbb {P}})\) be a probability space carrying a real-valued Lévy process \((X_{t})_{t\ge 0}\), with \(X_{0}=0\) and canonical triplet \((\gamma ,\sigma ^{2},\varLambda )\), where \(\gamma \in {\mathbb {R}}\), \(\sigma ^{2}\ge 0\), and \(\varLambda \) is a Lévy measure, that is a nonnegative measure on \({\mathbb {R}}\) satisfying
For \(x>0\), put
with corresponding Lévy measures \(\varLambda _{+}\) and \(\varLambda _{-}\) on \( {\mathbb {R}}^{+}=\left( 0,\infty \right) \) and set
We assume always that
For \(u>0\) let
By Remark 1, we have
The process \((X_{t})_{t\ge 0}\) has the representation (e.g., Bertoin [2] and Sato [10])
with
where for \(0<\varepsilon <1\)
and \(\left( Z_{t}\right) _{t\ge 0}\) is a standard Wiener process independent of \(\left( X_{t}^{\left( 1\right) }\right) _{t\ge 0}\) and \( \left( X_{t}^{\left( 2\right) }\right) _{t\ge 0}\). (As usual \(\varDelta X_{s} =X_{s}-X_{s-}.)\) The limit in (39) is defined as in pages 14–15 of Bertoin [2].
Decomposing further, we get
with
where for \(0<\varepsilon <1\)
and
For any \(t>0\), denote the ordered positive jump sequence
of \(X_{t}\) on the interval \(\left[ 0,t\right] \) and let
denote the corresponding ordered negative jump sequence of \(X_{t}\). Note that the positive and negative jumps are independent. With this notation, we can write
and
Let \(\left( \varGamma ^{+}\right) _{i\ge 1}\) \(\overset{\mathrm {D}}{=}\) \(\left( \varGamma _{i}^{-}\right) _{i\ge 1}\overset{\mathrm {D}}{=}\left( \varGamma _{i}\right) _{i\ge 1}\), with \(\left( \varGamma _{i}^{+}\right) _{i\ge 1}\) and \(\left( \varGamma _{i}^{-}\right) _{i\ge 1}\) independent. It turns out that by the same arguments that lead to (8), for each \(t>0\)
and
Let \({\widehat{X}}_{t}^{\left( 1,\pm \right) }\) and \({\widehat{X}}_{t}^{\left( 2,\pm \right) }\) be defined as \(X_{t}^{\left( 1,\pm \right) }\) and \( X_{t}^{\left( 2,\pm \right) }\) with \(m_{t}^{\left( i,\pm \right) }\) replaced by \(\pm \varphi _{\pm }\left( \frac{\varGamma _{i}^{\pm }}{t}\right) \). We see then by (40) that for each fixed \(t\ge 0\)
where \(\left( Z_{t}\right) _{t\ge 0}\) is a Wiener process independent of \( \left( \varGamma _{i}^{+}\right) _{i\ge 1}\) and \(\left( \varGamma _{i}^{-}\right) _{i\ge 1}\).
Our aim is to show that for a trimmed version \({\widetilde{T}} _{t_{n}}^{(k_{n},\ell _{n})}\) of \({\widehat{X}}_{t_{n}}\) defined for suitable sequences of positive integers \(\left( k_{n}\right) _{n\ge 1}\) and \(\left( \ell _{n}\right) _{n\ge 1}\) and positive constants \(\left( t_{n}\right) _{n\ge 1}\) that under appropriate regularity conditions there exist centering and norming functions \(A_{n}\left( \cdot ,\cdot \right) \) and \( B_{n}\left( \cdot ,\cdot \right) \) such that uniformly in \(x\in {\mathbb {R}}\), as \(n\rightarrow \infty \),
which implies
Statement (45) is what we call a SSCLT for a trimmed Lévy process. In order to define \({\widetilde{T}}_{t_{n}}^{(k_{n},\ell _{n})}\), specify the centering and norming functions \(A_{n}\left( \cdot ,\cdot \right) \) and \( B_{n}\left( \cdot ,\cdot \right) \), and state and prove our versions of (44) and (45) given in Theorem 2 in Sect. 5.6, we must first introduce some notation and preliminary results, which we shall do in the next four subsections.
5.2 A Useful Spectrally Positive Lévy Process
Let \(\left( P_{t}\right) _{t\ge 0}\), be a nondegenerate spectrally positive Lévy process without a normal component and having zero drift with infinitely divisible characteristic function
where
and \(\pi \) is a Lévy measure on \({\mathbb {R}}^{+}\) with \(\int _{\left( 0,\infty \right) }(x^{2}\wedge 1)\pi (\text {d}x)\) finite. Such a process has no negative jumps. Again we shall assume
As above for \(u>0\) let \(\varphi _{\pi }(u)=\sup \{x:{\overline{\pi }}(x)>u\}\). Applying Remark 1, we see that (46) implies
(Often in the definition of a spectrally positive Lévy process it is assumed that it is not a subordinator. See Abdel-Hameed [1].)
The process \(\left( P_{t}\right) _{t\ge 0}\) has the representation
where \(P_{t}^{\left( 1\right) }=\)
and
The processes \(\left( P_{t}^{\left( 1\right) }\right) _{t\ge 0}\) and \( \left( P_{t}^{\left( 2\right) }\right) _{t\ge 0}\) are independent Lévy processes. Observe that for any \(t>0\), we can write
with \({\widehat{P}}_{t}^{\left( 1\right) }=\)
where \(\left\{ \varGamma _{i}\right\} _{i\ge 1}\) is as above. Also write
For each \(t>0\), we have
The random variable \({\widehat{P}}_{t}^{\left( 1\right) }\) has characteristic function
where
5.3 A Useful Infinitely Divisible Random Variable
For each \(t>0\) and \(y>0\) with \(\left( \varGamma _{i}^{\prime }\right) _{i\ge 1} \) \(\overset{\mathrm {D}}{=}\) \(\left( \varGamma _{i}\right) _{i\ge 1}\) , consider the random variable
where for \(0<\varepsilon <1\)
with
and
Also let
Introduce the rate 1 Poisson process
We can write (53) as
Consider the Lévy measure \(\pi _{y/t}\) defined on \(\left( 0,\infty \right) \) by the tail function
Note that for all \(u>0\)
For future reference, we record that \({\widehat{P}}_{t}^{\left( 1\right) }\left( y\right) \) has characteristic function
where
By an examination of (61) and (62), we see that \({\widehat{P}} _{t}^{\left( 1\right) }\left( y\right) \) is an infinitely divisible random variable. Clearly, from (59), we get
and
where the fact that \(\sigma _{\pi }^{2}\left( y/t\right) >0\) follows from ( 46) implies (47).
5.4 Application of the Above Constructions
For any fixed \(y>0\) and \(t>0\), consider the tail functions defined for \(x>0\) by
and
Let \(N_{1}\) and \(N_{2}\) be two independent rate 1 Poisson processes on \( \left( 0,\infty \right) \) with jumps \(\varGamma _{i}^{\left( 1\right) }\), \( i\ge 1\), and \(\varGamma _{i}^{\left( 2\right) }\), \(i\ge 1,\) respectively. Now for \(t>0\) and \(y_{1}>0\) let \({\widehat{X}}_{t}^{\left( 1,+\right) }\left( y_{1}\right) \) and \({\widehat{X}}_{t}^{\left( 2,+\right) }\left( y_{1}\right) \) be constructed as \({\widehat{P}}_{t}^{\left( 1\right) }\left( y_{1}\right) \) and \({\widehat{P}}_{t}^{\left( 2\right) }\left( y_{1}\right) \) using the Poisson process \(N_{1}\) and the Lévy measure \(\varLambda _{+}\) with inverse \(\varphi _{+}\). In the same way for \(t>0\) and \(y_{2}>0\), construct \(\widehat{X }_{t}^{\left( 1,-\right) }\left( y_{2}\right) \) and \({\widehat{X}}_{t}^{\left( 2,-\right) }\left( y_{2}\right) \) using the Poisson process \(N_{2}\) and the L évy measure \(\varLambda _{-}\) with inverse \(\varphi _{-}\). One finds that \( {\widehat{X}}_{t}^{\left( 1,+\right) }\left( y_{1}\right) \) and \({\widehat{X}} _{t}^{\left( 1,-\right) }\left( y_{2}\right) \) are independent infinitely divisible random variables with Lévy measures defined via the above tail functions \(\varLambda _{y_{1}/t,+}\) and \(\varLambda _{y_{2}/t,-}\), respectively, whose supports are contained in \(\left[ 0,\varphi _{+}(y_{1}/t)\right] \) and \(\left[ 0,\varphi _{-}(y_{2}/t)\right] \), respectively. Moreover,
and by (63),
and
For \(t>0\), \(y_{1}>0\) and \(y_{2}>0\), consider the random variable
where \(\sigma \ge 0\) and \(\left( Z_{t}\right) _{t\ge 0}\) is a standard Brownian motion independent of the variables \({\widehat{X}}_{t}^{\left( 1,+\right) }\left( y_{1}\right) \) and \({\widehat{X}}_{t}^{\left( 1,-\right) }\left( y_{2}\right) \). Set
where by (64) and (65), \(\sigma ^{2}\left( t,y_{1},y_{2}\right) >0.\)
A basic step toward extending Theorem 1 from subordinators to general Lévy processes is the following result: For each \(t>0\), \(y_{1}>0 \) and \(y_{2}>0\) consider the standardized version of \({\widehat{Y}} _{t}^{\left( 1\right) }\left( y_{1},y_{2}\right) \) given by
The random variable \(S^{(1)}\left( t,y_{1},y_{2}\right) \) is infinitely divisible with
whose Lévy measure has support contained in
Since the random variable \(S^{(1)}\left( t,y_{1},y_{2}\right) \) is infinitely divisible, we can apply the Zaitsev Fact to get for \(t>0\) , \(y_{1}>0\), \(y_{2}>0\) and \(\lambda >0\) and for universal positive constants \(C_{1}\) and \(C_{2}\)
where \(\varphi \left( t,y_{1},y_{2}\right) =\max \left\{ \varphi _{+}(y_{1}/t),\varphi _{-}(y_{2}/t)\right\} \).
5.5 Definition of Trimmed Lévy Process
Set for \(0<\varepsilon <1\), \(t>0\) and \(y>0\)
Let \(\left( Z_{t}\right) _{t\ge 0}\), \(\left( \varGamma _{i}^{+}\right) _{i\ge 1}\) and \(\left( \varGamma _{i}^{-}\right) _{i\ge 1}\) be as in (43). We shall consider for sequences of positive constants \(t_{n}\) and positive integers \(k_{n}\) and \(\ell _{n}\) trimmed versions of the Lévy process \( X_{t}\) at \(t_{n}\), namely \({\widehat{X}}_{t_{n}}\), given by
where \({\widetilde{T}}_{t_{n}}^{(k_{n},+)}=\)
and \({\widetilde{T}}_{t_{n}}^{(\ell _{n},-)}=\)
Notice that by construction, \(Z_{t_{n}},{\widetilde{T}}_{t_{n}}^{(k_{n},+)}\) and \({\widetilde{T}}_{t_{n}}^{(\ell _{n},-)}\) are independent.
5.6 Our SSCLT for a Trimmed Lévy Process
Armed with the notation and auxiliary results established in the previous subsections, we now state and prove our SSCLT for the trimmed Lévy process defined in (69). We note in passing that assumption (37) can be relaxed a bit; however, the present version of our SSCLT and its proof suffices to reveal the main ideas.
Theorem 2
Assume that (37) holds. For any two sequences of positive integers \(\left\{ k_{n}\right\} _{n\ge 1}\) and \(\left\{ \ell _{n}\right\} _{n\ge 1}\) converging to infinity and sequence of positive constants \(\left\{ t_{n}\right\} _{n\ge 1}\) satisfying
and
we have uniformly in x, as \(n\rightarrow \infty \)
which implies as \(n\rightarrow \infty \)
A simple example Before we prove Theorem 2, we shall give a simple example. Let \((X_{t})_{t\ge 0}\) be a Lévy process with canonical triplet \((0,0,\varLambda )\). Recall the notation (35). Assume that \(\varLambda _{+}=\varLambda _{-}\) and \(\varLambda _{+}\) is regularly varying at zero with index \(-\alpha \), where \(0<\alpha <2\). This implies that \(\varphi _{+}=\varphi _{-}\) is regularly varying at \(\infty \) with index \(-1/\alpha \) and thus for \(x>0\),
where \(L\left( x\right) \) is slowly varying at infinity.
Applying (64), we see that
which by (74), as \(y_{1}/t\rightarrow \infty \),
where \(b_{\alpha }^{2}=\alpha /\left( 2-\alpha \right) \). In the same way, we get as \(y_{2}/t\rightarrow \infty \)
Note that in this example \(\sigma ^{2}=0\), so that
Assuming \(k_{n}\rightarrow \infty \) and \(k_{n}/t_{n}\rightarrow \infty \), we get that
and thus
This implies that
and
from which we readily infer that
Thus, by Theorem 2 we have uniformly in x, as \(n\rightarrow \infty \)
By (77) we can replace the random norming in (78) by a deterministic norming to get uniformly in x, as \(n\rightarrow \infty \)
Proof of Theorem 2
Consider two sequences of random variables \((Y_{1,k_{n}}) _{n\ge 1}\), independent of \((\varGamma _{i}^{(1) }) _{i\ge 1}\), and \(( Y_{2,\ell _{n}}) _{n\ge 1}\), independent of \((\varGamma _{i}^{(2) }) _{i\ge 1}\), and independent of each other. Assume that \( t_{n}>0\), \(k_{n}>0\) and \(\ell _{n}>0\) are such that
then by applying (67) we get uniformly in x, as \(n\rightarrow \infty \),
For \(t>0\), \(y_{1}>0\) and \(y_{2}>0\), set
Further, let
and
We see that, if addition to (79), we assume that \(t_{n}>0\), \( k_{n}>0\) and \(\ell _{n}>0\) are such that
then by (38)
which implies that
This gives
which in combination with (80) implies that uniformly in x
Let \(\left( Y_{1,k_{n}},Y_{2,\ell _{n}}\right) =\left( \varGamma _{k_{n}}^{+},\varGamma _{\ell _{n}}^{-}\right) ,\) and be independent of \( \left( \varGamma _{i}^{\left( 1\right) }\right) _{i\ge 1}\) and \(\left( \varGamma _{i}^{\left( 2\right) }\right) _{i\ge 1}\). We see that
Combining (83) with (84), we get (72) and (73 ). \(\square \)
References
Abdel-Hameed, M.: Lévy Processes and Their Applications in Reliability and Storage. SpringerBriefs in Statistics. Springer, Berlin (2014)
Bertoin, J.: Lévy Processes. Volume 121 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge (1996)
Brémaud, P.: Point Processes and Queues. Martingale Dynamics. Springer, Berlin (1981)
Csörgő, S., Horváth, L., Mason, D.M.: What portion of the sample makes a partial sum asymptotically stable or normal? Probab. Theory Relat. Fields 72, 1–16 (1986)
Doney, R.A., Maller, R.A.: Stability and attraction to normality for Lévy processes at zero and infinity. J. Theor. Probab. 15, 751–792 (2002)
Ipsen, Y., Maller, R., Resnick, S.: Trimmed Lévy processes and their extremal components. Stoch. Process. Appl. 130, 2228–2249 (2020)
Maller, R.A., Mason, D.M.: Stochastic Compactness of L évy Processes. High Dimensional Probability V: the Luminy Volume, vol. 5, pp. 239–257. Inst. Math. Stat. (IMS) Collect., Beachwood (2009)
Maller, R.A., Mason, D.M.: Small-time compactness and convergence behavior of deterministically and self-normalized Lévy processes. Trans. Am. Math. Soc. 362, 2205–2248 (2010)
Rosiński, J.: Series representations of Lévy processes from the perspective of point processes. In: Bertoin, J. (ed.) Lévy Processes, pp. 401–415. Birkhäuser, Boston (2001)
Sato, K.: Lévy Processes and Infinitely Divisible Distributions. Cambridge Univ. Press, Cambridge (2005)
Zaĭtsev, A.Yu.: On the Gaussian approximation of convolutions under multidimensional analogues of S. N. Bernstein’s inequality conditions. Probab. Theory Relat. Fields 74, 535–566 (1987)
Acknowledgements
The author thanks the two referees for pointing out a number of misprints and minor oversights, as well as for suggestions that improved the presentation of the results.
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
Mason, D.M. Self-Standardized Central Limit Theorems for Trimmed Lévy Processes. J Theor Probab 34, 2117–2144 (2021). https://doi.org/10.1007/s10959-020-01021-0
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-020-01021-0