1: \documentclass[10pt]{article}
2: \setlength{\baselineskip}{14pt}
3: \setlength{\textheight}{21.5cm}
4: \setlength{\textwidth}{16cm}
5: \setlength{\oddsidemargin}{0cm}
6: \setlength{\evensidemargin}{0cm}
7:
8: \usepackage{amsmath,amsthm,amssymb,amsbsy,amsopn}
9: \usepackage{graphicx}
10: \def\Var{\mathop{\rm Var}}
11: \def\Cov{\mathop{\rm Cov}}
12: \def\thefootnote{\fnsymbol{footnote}}
13:
14: \newtheorem{Def}{Definition}[section]
15: \newtheorem{Thm}[Def]{Theorem}
16: \newtheorem{Lem}[Def]{Lemma}
17: \newtheorem{Prop}[Def]{Proposition}
18: \newtheorem{Coro}[Def]{Corollary}
19: \newtheorem{Rem}[Def]{Remark}
20: \newtheorem{Exam}[Def]{Example}
21:
22: \theoremstyle{definition}
23:
24: \newcounter{enum2}
25: \newenvironment{enumerate2}{%
26: \begin{list}%
27: {%
28: % \arabic{enum2}.\ \,% 見出し記号／直後の空白を調節
29: }%
30: {%
31: \usecounter{enum2}
32: \setlength{\itemindent}{0em}% ここは 0 に固定
33: \setlength{\leftmargin}{3em}% 左のインデント
34: \setlength{\rightmargin}{0em}% 右のインデント
35: \setlength{\labelsep}{1em}% 黒丸と説明文の間
36: \setlength{\labelwidth}{3em}% ラベルの幅
37: \setlength{\itemsep}{0em}% 項目ごとの改行幅
38: \setlength{\parsep}{0em}% 段落での改行幅
39: \setlength{\listparindent}{0em}% 段落での一字下り
40: }
41: }{%
42: \end{list}%
43: }
44:
45: \def\theequation{\thesection.\arabic{equation}}
46: \makeatletter
47: \@addtoreset{equation}{section}
48: \makeatother
49:
50:
51: \title{On Layered Stable Processes}
52: \author{C. Houdr\'e\footnote{Laboratoire d'Analyse et de Math\'ematiques
53: Appliqu\'ees, CNRS UMR 8050, Universit\'e Paris XII, 94010 Cr\'eteil
54: Cedex, France, and School of Mathematics, Georgia Institute of Technology,
55: Atlanta, GA, 30332-0160, USA, houdre@math.gatech.edu}
56: ~and R. Kawai\footnote{Quantitative Research Department, Daiwa Securities SMBC
57: Co.Ltd., 1-14-5, Eitai, Koto-ku, Tokyo, 135-0034, Japan, reiichiro.kawai@daiwasmbc.co.jp}}
58: \date{March 28, 2005}
59:
60: \begin{document}
61: \maketitle
62:
63: \begin{abstract}
64:
65: {\it Layered stable} (multivariate) distributions and processes are
66: defined and studied.
67: A layered stable process combines stable trends of two different
68: indices, one of them possibly Gaussian.
69: More precisely, in short time, it is close to a stable process while, in long
70: time, it approximates another stable (possibly Gaussian) process.
71: We also investigate the absolute continuity of a layered stable process with
72: respect to its short time limiting stable process.
73: A series representation of layered stable processes is derived, giving
74: insights into both the structure of the sample paths and of the short
75: and long time behaviors.
76: This series is further used for sample paths simulation.
77: \end{abstract}
78:
79: \footnotetext{
80: \noindent {\it Keywords:} L\'evy processes, stable distributions and
81: processes, layered stable distributions and processes.
82: }
83: \footnotetext{{\it AMS Subject Classification (2000):} 60G52, 60G51, 60E07, 60F05.}
84:
85: \section{Introduction and preliminaries}
86:
87: Stable processes form one of the simplest class of L\'evy processes without
88: Gaussian component.
89: They have been thoroughly studied by many authors and have
90: been used in several fields of applications, such as statistical
91: physics, queueing theory, mathematical finance.
92: One of their major attractions is the scaling property induced by the
93: structure of the corresponding L\'evy measure.
94: Sato~\cite{sato} and Samorodnitsky and Taqqu~\cite{samorodnitsky taqqu}
95: contain many basic facts on stable distributions and processes.
96: Recent generalizations of stable processes can also be found, for
97: example, in Barndorff-Nielsen and Shepard~\cite{barndorff-nielsen
98: shepard} and in Rosi\'nski~\cite{super rosinski}.
99: These new classes are also of great interest in applications and have
100: moreover motivated our study.
101: %In this paper, we will introduce and study a generalization; {\it
102: %layered stable processes}.
103:
104: In the present paper, we introduce and study further generalizations which
105: we call {\it layered stable} distributions and processes.
106: They are defined in terms of the structure of their L\'evy measure
107: whose radial component behaves asymptotically as an inverse polynomial
108: of different orders near zero and at infinity.
109: %They are constructed by dividing the radial part of the stable L\'evy
110: %measure into two layers associated with distinct stability indices.
111: {\it The inner and outer (stability) indices} correspond respectively to
112: these orders of polynomial decay.
113: This simple layering leads to the following properties:
114: The outer index determines the moment properties (Proposition
115: \ref{moment proposition}), while the variational properties depend on
116: the inner index (Proposition \ref{variation proposition}).
117: On the other hand, the inner and outer indices also correspond
118: to short and long time behavior of the sample paths.
119: In short time, a layered stable process behaves like a stable
120: process with the corresponding inner index (Theorem \ref{short of eS}).
121: The long time behavior has two modes depending on the outer index.
122: When the outer index is strictly smaller than two, a layered stable
123: process is close to a stable process with this index, while
124: behaving like a Brownian motion if the outer index is strictly greater
125: than two (Theorem \ref{lS long-time theorem}).
126: In relation to the short time behavior, we investigate the mutual absolute
127: continuity of a layered stable process and of its short time limiting
128: stable process (Theorem \ref{absolute continuity theorem}).
129: A shot noise series representation reveals the nature of layering and
130: also gives direct insights into the properties of layered stable processes.
131: We present typical sample paths of a layered stable process, which are
132: simulated via the series representation for various combinations of stability
133: indices in order to cover all the types of short and long time behavior.
134:
135: Let us begin with some general notations which will be used throughout
136: the text.
137: $\mathbb{R}^d$ is the $d$-dimensional Euclidean space with the norm
138: $\|\cdot\|$, $\mathbb{R}_0^d:=\mathbb{R}^d\setminus \{0\},$
139: $\mathcal{B}(\mathbb{R}^d_0)$ is the Borel $\sigma$-field of
140: $\mathbb{R}^d_0$, and $S^{d-1}:=\{x\in \mathbb{R}^d:\|x\|=1\}$.
141: $A'$ is the transpose of the matrix $A$, while $\|\cdot\|_{{\rm o}}$ is
142: the operator norm of the linear transformation $A\in \mathbb{R}^{d\times
143: d}$, i.e., $\|A\|_{{\rm o}}=\sup_{\|x\|\le 1}\|Ax\|.$
144: $f(x) \sim g(x)$ indicates that $f(x)/g(x)\to 1$, as $x\to x_0\in
145: [-\infty,\infty]$, while $f(x)\asymp g(x)$ is used to mean that there
146: exist two positive constants $c_1$ and $c_2$ such that $c_1g(x)\le
147: f(x)\le c_2g(x)$, for all $x$ in an approximate set.
148: $\mathcal{L}(X)$ is the law of the random vector $X$, while
149: $\stackrel{\mathcal{L}}{=}$ and $\stackrel{\mathcal{L}}{\to}$
150: denote, respectively, equality and convergence in distribution, or of
151: the finite dimensional distributions when random processes are considered.
152: $\stackrel{d}{\to}$ is used for the weak convergence of
153: random processes in the space $\mathbb{D}([0,\infty),\mathbb{R}^d)$ of
154: c\`adl\`ag functions from $[0,\infty)$ into $\mathbb{R}^d$ equipped with
155: the Skorohod topology.
156: $\stackrel{v}{\to}$ denotes convergence in the vague topology.
157: For any $r>0$, $T_r$ is a transformation of measures on $\mathbb{R}^d$
158: given, for any positive measure $\rho$, by $(T_r\rho)(B)=\rho (r^{-1}B),$
159: $B\in\mathcal{B}(\mathbb{R}^d).$
160: $\mathbb{P}|_{\mathcal{F}_t}$ is the restriction of a
161: probability measure $\mathbb{P}$ to the $\sigma$-field $\mathcal{F}_t,$
162: while $\Delta X_t$ denotes the jump of $X$ at time $t,$ that is, $\Delta
163: X_t:=X_t-X_{t-}.$
164: Finally, and throughout, all the multivariate or matricial integrals
165: are defined componentwise.
166:
167: Recall that an infinitely divisible probability measure $\mu$ on
168: $\mathbb{R}^d$, without Gaussian component, is called {\it stable} if
169: its L\'evy measure is given by
170: \[
171: \nu (B)=\int_{S^{d-1}}\sigma (d\xi
172: )\int_0^{\infty}{\bf 1}_B(r\xi)\frac{dr}{r^{1+\alpha}},\quad B\in
173: {\mathcal B}(\mathbb{R}_0^d),
174: \]
175: where $\alpha \in (0,2)$ is the stability index and where $\sigma$ is a
176: finite positive measure on $S^{d-1}$.
177: It is well known that the characteristic function of $\mu$ is given by
178: \begin{eqnarray}
179: \widehat{\mu}(y)&=&\exp\left[i\langle y,\eta\rangle +\int_{\mathbb{R}_0^d}
180: (e^{i\langle y,z\rangle}-1-i\langle y,z\rangle {\bf 1}_{\{\|z\|\le 1\}}(z))
181: \nu_{\alpha}(dz)\right] \label{stable cf}\\
182: &=&
183: \begin{cases}
184: \exp\left[i\langle y,\tau_{\alpha} \rangle-c_{\alpha}\int_{S^{d-1}}|\langle y,\xi\rangle
185: |^{\alpha}\left(1-i\tan\frac{\pi \alpha}{2}{\rm sgn}\langle y,\xi
186: \rangle\right)\sigma (d\xi)\right],&{\rm if}~\alpha \ne 1,\\
187: \exp\left[i\langle y,\tau_1\rangle-c_1\int_{S^{d-1}}\left(|\langle y,\xi\rangle|+i\frac{2}{\pi}\langle y,\xi\rangle \ln |\langle y,\xi\rangle
188: |\right)\sigma (d\xi)\right],&{\rm if}~\alpha = 1,
189: \end{cases}\nonumber
190: \end{eqnarray}
191: for some $\eta\in \mathbb{R}^d$, and where $c_{\alpha}=|\Gamma
192: (-\alpha)\cos \frac{\pi \alpha}{2}|$ when $\alpha \ne 1$ while
193: $c_1=\pi/2$, with moreover $\tau_{\alpha}=\eta
194: -\frac{1}{1-\alpha}\int_{S^{d-1}}\xi\sigma (d\xi)$ when $\alpha \ne 1$
195: and $\tau_1=\eta -(1-\gamma)\int_{S^{d-1}}\xi \sigma (d\xi)$, $\gamma
196: (=0.5772...)$ being the Euler constant.
197: A L\'evy process $\{X_t:t\ge 0\}$ such that $\mathcal{L}(X_1)\sim \mu$
198: is called a {\it stable process}.
199: Stable processes enjoy the {\it selfsimilarity} property, i.e.,
200: for any $a>0$,
201: \[
202: \{X_{at}:t\ge 0\}\stackrel{\mathcal{L}}{=}\{a^{1/\alpha}X_t+bt:t\ge 0\},
203: \]
204: for some $b\in \mathbb{R}^d$.
205: Next, we recall a shot noise series representation of stable
206: processes on a fixed finite horizon $[0,T]$, $T>0$.
207: %It was first introduced by LePage \cite{lepage1} and
208: Related results can be found, for example, in Theorem 1.4.5 of
209: Samorodnitsky and Taqqu~\cite{samorodnitsky taqqu}.
210: The centering constants given below are obtained in Proposition 5.5 of
211: Rosi\'nski~\cite{super rosinski}.
212:
213: \begin{Lem}\label{alpha series}
214: Let $T>0$.
215: Let $\{T_i\}_{i\ge 1}$ be a sequence of iid uniform random variables on
216: $[0,T]$, let $\{\Gamma_i\}_{i\ge 1}$ be an arrival times of a standard Poisson
217: process, and let $\{V_i\}_{i\ge 1}$ a sequence of iid random vectors in
218: $S^{d-1}$ with common distribution $\sigma (d\xi)/\sigma (S^{d-1})$.
219: Also let
220: \begin{equation*}
221: z_0=
222: \begin{cases}
223: 0,&if~\alpha \in (0,1),\\
224: \int_{S^{d-1}}\xi\sigma (d\xi)/\sigma (S^{d-1}) ,&if~\alpha \in [1,2),
225: \end{cases}
226: \end{equation*}
227: and
228: \begin{equation*}
229: b_T=
230: \begin{cases}
231: 0,&if~\alpha \in (0,1),\\
232: \sigma (S^{d-1})T(\gamma +\ln (\sigma (S^{d-1})T)),&if~\alpha =1,\\
233: \left(\frac{\alpha}{\sigma(S^{d-1})T}\right)^{-1/\alpha}\zeta (1/\alpha),&if~\alpha \in (1,2),
234: \end{cases}
235: \end{equation*}
236: where $\zeta$ denotes the Riemann zeta function.
237: Then, the stochastic process
238: \[
239: \left\{\sum_{i=1}^{\infty}\left[\left(\frac{\alpha
240: \Gamma_i}{\sigma (S^{d-1})T}\right)^{-1/\alpha}V_i\,{\bf 1}(T_i\le
241: t)-\left(\frac{\alpha i}{\sigma (S^{d-1})T}\right)^{-1/\alpha}
242: z_0\frac{t}{T}\right]+b_Tz_0\frac{t}{T}:t\in [0,T]\right\},
243: \]
244: converges almost surely uniformly in $t$ to an $\alpha$-stable process
245: $\{X_t:t\in [0,T]\}$ satisfying $\mathbb{E}[e^{i\langle
246: y,X_T\rangle}]=\widehat{\mu}(y)^T$, where $\widehat{\mu}$ given by
247: (\ref{stable cf}) with
248: \[
249: \eta=
250: \begin{cases}
251: \frac{1}{1-\alpha}\int_{S^{d-1}}\xi \sigma (d\xi),&{\rm if}
252: ~\alpha \ne 1,\\
253: 0,&{\rm if} ~\alpha =1.
254: \end{cases}
255: \]
256: \end{Lem}
257:
258: \section{Definition and basic properties}
259:
260: We first define a layered stable multivariate distribution by precising
261: the structure of its L\'evy measure in polar coordinates.
262:
263: \begin{Def}
264: Let $\mu$ be an infinitely divisible probability measure on
265: $\mathbb{R}^d$ and without Gaussian component.
266: Then, $\mu$ is called layered stable if its L\'evy measure on
267: $\mathbb{R}_0^d$ is given by
268: \begin{equation}
269: \nu(B)=\int_{S^{d-1}}\sigma
270: (d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)q(r,\xi)dr,\quad B\in
271: \mathcal{B}(\mathbb{R}_0^d),\label{def of eS}
272: \end{equation}
273: where $\sigma$ is a finite positive measure on $S^{d-1}$, and
274: $q$ is a measurable function from $(0,\infty)\times S^{d-1}$ to
275: $(0,\infty)$ such that for each $\xi\in S^{d-1}$,
276: \begin{equation}\label{short q}
277: q(r,\xi)\sim c_1(\xi)r^{-\alpha -1},\quad as~r\to 0,
278: \end{equation}
279: and
280: \begin{equation}\label{long q}
281: q(r,\xi)\sim c_2(\xi)r^{-\beta -1},\quad as~r\to\infty,
282: \end{equation}
283: where $c_1$ and $c_2$ are integrable (with respect to $\sigma$)
284: functions on $S^{d-1}$, and where $(\alpha ,\beta)\in (0,2)\times
285: (0,\infty )$.
286: \end{Def}
287:
288: $q(\cdot ,\cdot )$ is called {\it the $q$-function} of $\mu$, or of its
289: L\'evy measure $\nu$.
290: Clearly, $\nu$ is well defined as a
291: L\'evy measure since it behaves like an $\alpha$-stable L\'evy measure
292: near the origin while decaying like a $\beta$-Pareto density when
293: sufficiently far away from the origin.
294: $\alpha$ and $\beta$ are respectively called {\it the inner and
295: outer (stability) indices} of $\mu$, or of $\nu$.
296:
297: {\it For convenience, we henceforth use the notations $\sigma_1$
298: and $\sigma_2$ for the finite positive measures on $S^{d-1}$ defined
299: respectively by
300: \begin{equation}\label{def of sigma1}
301: \sigma_1(B):=\int_Bc_1(\xi)\sigma (d\xi),\quad B\in \mathcal{B}(S^{d-1}),
302: \end{equation}
303: and
304: \begin{equation}\label{def of sigma2}
305: \sigma_2(B):=\int_Bc_2(\xi)\sigma (d\xi),\quad B\in \mathcal{B}(S^{d-1}),
306: \end{equation}
307: while $\nu^{\alpha}_{\sigma}$ is used for the positive measure on
308: $\mathbb{R}_0^d$ given by
309: \begin{equation}\label{def of stable levy measure}
310: \nu^{\alpha}_{\sigma}(B):=\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}{\bf
311: 1}_B(r\xi)\frac{dr}{r^{\alpha+1}},\quad B\in \mathcal{B}(\mathbb{R}_0^d),
312: \end{equation}
313: where $\alpha\in (0,\infty)$ and where $\sigma$ is a finite positive
314: measure on $S^{d-1}$.}
315: Note that if $\alpha \in (0,2)$, $\nu_{\sigma}^{\alpha}$ is simply an
316: $\alpha$-stable L\'evy measure, while not well defined as a L\'evy
317: measure when $\alpha \ge 2$.
318:
319: \begin{Exam}\label{special remark}{\rm
320: The following layered stable L\'evy measure is simple, yet interesting:
321: \begin{eqnarray}
322: \nu (B)&=&\int_B{\bf 1}_{\{\|z\|\le
323: 1\}}(z)\nu_{\sigma}^{\alpha}(dz)+\int_B{\bf 1}_{\{\|z\|>
324: 1\}}(z)\nu_{\sigma}^{\beta}(dz)\nonumber \\
325: &=&\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}{\bf
326: 1}_B(r\xi)\frac{dr}{r^{\alpha+1}{\bf 1}_{(0,1]}(r)+r^{\beta+1}{\bf
327: 1}_{(1,\infty)}(r)},\quad B\in
328: \mathcal{B}(\mathbb{R}_0^d).\label{original lS}
329: \end{eqnarray}
330: The corresponding $q$-function is given by
331: \[
332: q(r,\xi)=\sigma(S^{d-1})^{-1}(r^{-\alpha-1}{\bf 1}_{(0,1]}(r)+r^{-\beta-1}{\bf
333: 1}_{(1,\infty)}(r)),\quad \xi\in S^{d-1},
334: \]
335: which is independent of $\xi$.
336: The measure $\nu$ consists of two disjoint domains of stability,
337: and this construction results in two layers for the radial
338: component associated with each respective stability index.
339: The name ``{\it layered stable}'' originates from this special
340: structure.
341: %To better illustrate our results in what follows, we will use this
342: % special L\'evy measure as an example.
343:
344: Recall that an infinitely divisible probability measure $\mu$ on
345: $\mathbb{R}^d$ is said to be of class $L_0$, or
346: selfdecomposable if for any $b>1$, there exists a probability measure
347: $\varrho_b$ such that
348: $\widehat{\mu}(z)=\widehat{\mu}(b^{-1}z)\widehat{\varrho_b}(z).$
349: Equivalently, the L\'evy measure of $\mu$ has the form
350: \[
351: \int_{S^{d-1}}\sigma(d\xi)\int_0^{\infty}{\bf
352: 1}_B(r\xi)k_{\xi}(r)\frac{dr}{r},\quad B\in \mathcal{B}(\mathbb{R}_0^d),
353: \]
354: where $\sigma$ is a finite positive measure on $S^{d-1}$ and where
355: $k_{\xi}(r)$ is a nonnegative function measurable in $\xi \in S^{d-1}$ and
356: decreasing in $r>0.$
357: Clearly, the L\'evy measure (\ref{original lS}) induces a
358: selfdecomposable measure.
359: Moreover, the classes $L_m$, $m=1,2,\ldots ,$ are defined recursively as
360: follows; $\mu\in L_m$ if for every $b>1$, there exists $\varrho_b\in
361: L_{m-1}$ such that
362: $\widehat{\mu}(z)=\widehat{\mu}(b^{-1}z)\widehat{\varrho_b}(z).$
363: Clearly, $L_0\supset L_1 \supset L_2 \supset \cdots$.
364: Let $h_{\xi}(u):=k_{\xi}(e^{-u})$, be the so-called {\it $h$-function} of
365: $\mu$, or of its L\'evy measure.
366: Then, alternatively, $\mu \in L_0$ is shown to be in $L_m$ if and only if
367: $h_{\xi}(u)\in C^{m-1}$ and $h^{(j)}\ge 0$, for $j=0,1,\ldots ,m-1$.
368: % recursively as follows;
369: %$\mu \in L_m$ if and only if for every $b>1$ there exists $\varrho_b\in
370: % L_{m-1}$ such that
371: % $\widehat{\mu}(z)=\widehat{\mu}(b^{-1}z)\widehat{\varrho}_b(z).$
372: (See Sato \cite{sato class L} for more details.)
373: The $h$-function of the L\'evy measure (\ref{original lS}) is given by
374: \[
375: h_{\xi}(u)=e^{\alpha u}{\bf 1}_{(0,\infty)}(u)+e^{\beta u}{\bf 1}_{(-\infty,0]}(u),
376: \]
377: which is in $C^0$ but not in $C^1$.
378: Therefore, the infinitely divisible probability measure induced by
379: (\ref{original lS}) is in $L_1$, but not in $L_2$.
380: }\end{Exam}
381:
382: The following result asserts that a layered stable distribution has the
383: same probability tail behavior as $\beta$-Pareto distributions, or
384: $\beta$-stable distributions if $\beta\in (0,2)$.
385:
386: \begin{Prop}\label{moment proposition}{\rm (Moments)}
387: Let $\mu$ be a layered stable distribution with L\'evy measure $\nu$
388: given by (\ref{def of eS}) and let $\sigma_2$ be the measure (\ref{def
389: of sigma2}).
390: If $\sigma_2(S^{d-1})\ne 0$, then
391: \[
392: \int_{\mathbb{R}^d}\|x\|^{p}\mu (dx)
393: \begin{cases}
394: <\infty, & p\in (0,\beta),\\
395: =\infty, & p\in [\beta,\infty).
396: \end{cases}
397: \]
398: Moreover, $\int_{\mathbb{R}^d}\|x\|^p\mu (dx)<\infty$, $p\ge \beta$ and
399: $\int_{\mathbb{R}^d}e^{\theta \|x\|}\mu (dx)<\infty$, $\theta >0$ if
400: and only if $\sigma_2(S^{d-1})=0.$
401: \end{Prop}
402:
403: \begin{proof}
404: By Theorem 25.3 of Sato \cite{sato}, it is enough to show that the
405: restriction of $\nu$ to the set $\{z\in \mathbb{R}_0^d:\|z\|>1\}$ has
406: the corresponding moment properties.
407:
408: First, assume $\sigma_2(S^{d-1})\ne 0$.
409: Observe that $\int_{\|z\|>1}\|z\|^p \nu(dz)=
410: \int_{S^{d-1}}\sigma(d\xi)\int_1^{\infty}r^pq(r,\xi)dr$, and then by
411: (\ref{long q}), the right hand side is bounded from above and below by
412: constant multiples of
413: $\sigma_2(S^{d-1})\int_1^{\infty}r^p\frac{dr}{r^{\beta +1}}$ if $p\in
414: (0,\beta)$, while it is otherwise clearly infinite.
415:
416: Next, assume $\sigma_2(S^{d-1})=0$ and let $p\in [\beta,\infty)$.
417: Then, there exists $M>0$ such that
418: $\int_{\|z\|>1}\|z\|^p\nu (dz)\asymp\int_{S^{d-1}}\sigma(d\xi)\int_1^Mr^pq(r,\xi)dr$ and
419: $\int_{\|z\|>1}e^{\theta \|z\|}\nu (dz)\asymp
420: \int_{S^{d-1}}\sigma(d\xi)\int_1^Me^{\theta r}q(r,\xi)dr.$
421: Conversely, if $\sigma_2(S^{d-1})\ne 0$ and $p\in [\beta,\infty)$,
422: then $\int_{\|z\|>1}\|z\|^p\nu (dz)=+\infty$ as already shown and, again
423: by (\ref{long q}), $\int_{\|z\|>1}e^{\theta \|z\|}\nu
424: (dz)=\int_{S^{d-1}}\sigma (d\xi)\int_1^{\infty}e^{\theta
425: r}q(r,\xi)dr=+\infty.$
426: \end{proof}
427:
428: Let us define the associated L\'evy processes.
429:
430: \begin{Def}
431: A L\'evy process, without Gaussian component, is called layered stable
432: if its L\'evy measure is given by (\ref{def of eS}).
433: \end{Def}
434:
435: Henceforth, $\{X^{LS}_t:t\ge 0\}$ denotes a layered stable process
436: in $\mathbb{R}^d$.
437: Its characteristic function at time $1$ is given by
438: \begin{equation}\label{def via cf}
439: \mathbb{E}[e^{i\langle y,X^{LS}_1\rangle}]=\exp\left[i\langle
440: y,\eta\rangle +\int_{\mathbb{R}_0^d}
441: (e^{i\langle y,z\rangle }-1-i\langle y,z\rangle
442: {\bf 1}_{\{\|z\|\le 1\}}(z))\nu(dz)\right],
443: \end{equation}
444: where $\nu$ is the L\'evy measure given by (\ref{def of eS}) and
445: $\eta \in\mathbb{R}^d$.
446: For convenience of notation, we write $\{X^{LS}_t:t\ge 0\}\sim
447: LS_{\alpha,\beta}(\sigma,q;\eta)$ when (\ref{def via cf}) holds.
448: Similarly, for $\alpha\in (0,2)$, $\{X^{(\alpha)}_t:t\ge 0\}$ denotes an
449: $\alpha$-stable L\'evy process.
450: Its characteristic function at time $1$ is given by
451: \begin{equation}\label{cf of stable for lS}
452: \mathbb{E}[e^{i\langle y,X^{(\alpha)}_1\rangle}]=
453: \begin{cases}
454: \exp\left[i\langle y,\eta\rangle+\int_{\mathbb{R}_0^d}(e^{i\langle
455: y,z\rangle}-1)\nu^{\alpha}_{\sigma}(dz)\right],&if~\alpha\in (0,1),\\
456: \exp\left[i\langle y,\eta\rangle+\int_{\mathbb{R}_0^d}(e^{i\langle
457: y,z\rangle}-1-i\langle y,z\rangle
458: {\bf 1}_{\{\|z\|\le 1\}}(z))\nu^1_{\sigma}(dz)\right],&if~\alpha =1,\\
459: \exp\left[i\langle y,\eta\rangle+\int_{\mathbb{R}_0^d}(e^{i\langle
460: y,z\rangle}-1-i\langle y,z\rangle
461: )\nu^{\alpha}_{\sigma}(dz)\right],&if~\alpha\in (1,2),
462: \end{cases}
463: \end{equation}
464: where $\nu_{\sigma}^{\alpha}$ is given by (\ref{def of stable levy measure}),
465: and we write $\{X^{(\alpha)}_t:t\ge 0\}\sim S_{\alpha}(\sigma;\eta)$ when
466: (\ref{cf of stable for lS}) holds.
467:
468: \vspace{1em}
469: A layered stable process shares the variational
470: properties of a stable process with inner index $\alpha$.
471:
472: \begin{Prop}\label{variation proposition}{\rm ($p$-th variation)}
473: Let $X:=\{X^{LS}_t:t\ge 0\}\sim LS_{\alpha,\beta}(\sigma,q;\eta)$.
474:
475: \noindent (i) If $\sigma_1(S^{d-1})>0,$ then $X$ is a.s. of finite variation on
476: every interval of positive length if and only if $\alpha \in (0,1).$
477:
478: \noindent (ii) If $\sigma_1(S^{d-1})>0$, $(\alpha,\beta) \in [1,2)\times
479: (1,\infty)$ and $\eta =-\int_{S^{d-1}}\xi\sigma
480: (d\xi)\int_1^{\infty}rq(r,\xi)dr,$ then $X$ is a.s. of finite $p$-th
481: variation on every interval of positive length if and only if $p> \alpha$.
482:
483: \noindent (iii) If $\sigma_1(S^{d-1})=0,$ then it is a.s. of finite
484: variation on every interval of positive length.
485: \end{Prop}
486:
487: \begin{proof}
488: (i) Recall that the radial component of the layered stable L\'evy measure
489: near the origin behaves like the one of an $\alpha$-stable L\'evy measure.
490: The first claim then follows immediately from Theorem 3 of Gikhman and
491: Skorokhod \cite{gikhman skorokhod}.
492:
493: (ii) Since $X$ is now centered, Th\'eor\`eme III b of Bretagnolle
494: \cite{bretagnolle} directly applies.
495:
496: (iii) Letting $\nu$ be the L\'evy measure of $X$, there exists
497: $\epsilon\in (0,1)$ such that $\nu (\{z\in \mathbb{R}_0^d:\|z\|\le
498: \epsilon\})< \infty$ and so $\int_{\|z\|\le 1}\|z\|^p\nu (dz)<\infty$,
499: $p\ge 1.$
500: As in (i), the result follows from Theorem 3 of Gikhman and Skorokhod
501: \cite{gikhman skorokhod}.
502: \end{proof}
503:
504: Let us now consider a series representation for a general layered
505: stable process $\{X^{LS}_t:t\ge 0\}\sim LS_{\alpha,\beta}(\sigma,q;0).$
506: Fix $T>0$.
507: Let $\{T_i\}_{i\ge 1}$ be a sequence of iid
508: uniform random variables on $[0,T]$, let $\{\Gamma_i\}_{i\ge 1}$ be
509: Poisson arrivals with rate $1$, and let $\{V_i\}_{i\ge 1}$ be a
510: sequence of iid random vectors in $S^{d-1}$ with common
511: distribution $\sigma (d\xi)/\sigma (S^{d-1})$.
512: Assume moreover that the random sequences $\{T_i\}_{i\ge 1}$,
513: $\{\Gamma_i\}_{i\ge 1}$, and $\{V_i\}_{i\ge 1}$ are all mutually
514: independent.
515: Also, let
516: \[
517: \overleftarrow{q}(u,\xi):=\inf\{r>0:q([r,\infty),\xi)<u\},
518: \]
519: and let $\{b_i\}_{i\ge 1}$ be a sequence of constants given by
520: \[
521: b_i=\int_{i-1}^i\mathbb{E}[\overleftarrow{q}(s/T,V_1)V_1{\bf
522: 1}(\overleftarrow{q}(s/T,V_1)\le 1)]ds,
523: \]
524: Then, by Theorem 5.1 of Rosi\'nski \cite{rosinski2} with the help of
525: the LePage's method \cite{lepage}, the stochastic process
526: \begin{equation}\label{original series}
527: \left\{\sum_{i=1}^{\infty}\left[\overleftarrow{q}(\Gamma_i/T,V_i)V_i{\bf
528: 1}(T_i\le t)-b_i\frac{t}{T}\right]:t\in [0,T]\right\},
529: \end{equation}
530: converges almost surely uniformly in $t$ to a L\'evy process whose
531: marginal law at time $1$ is $LS_{\alpha,\beta}(\sigma,q;0).$
532:
533: \begin{Exam}{\rm
534: The L\'evy measure (\ref{original lS}) leads to a very illustrative
535: series representation.
536: Indeed,
537: \[
538: \overleftarrow{q}(r,\xi)=\left(\frac{\beta r}{\sigma
539: (S^{d-1})}\right)^{-1/\beta}{\bf
540: 1}_{(0,\sigma(S^{d-1})/\beta]}(r)+\left(\frac{\alpha r}{\sigma
541: (S^{d-1})}+1-\frac{\alpha}{\beta}\right)^{-1/\alpha}{\bf 1}_{(\sigma
542: (S^{d-1})/\beta,\infty)}(r),
543: \]
544: and so the stochastic process
545: \begin{eqnarray}
546: &&\Bigg\{\sum_{i=1}^{\infty}
547: \Bigg[\Bigg(\left(\frac{\beta\Gamma_i}{\sigma
548: (S^{d-1})T}\right)^{-1/\beta}{\bf 1}_{(0,\sigma
549: (S^{d-1})T/\beta]}(\Gamma_i)\label{special series}\\
550: &&~+\left(\frac{\alpha\Gamma_i}{\sigma
551: (S^{d-1})T}+1-\frac{\alpha}{\beta}\right)^{-1/\alpha}{\bf 1}_{(\sigma
552: (S^{d-1})T/\beta,\infty)}(\Gamma_i)\Bigg)V_i{\bf 1}(T_i\le t)
553: -b_iz_0\frac{t}{T}\Bigg]:t\in [0,T]\Bigg\},\nonumber
554: \end{eqnarray}
555: where
556: \[
557: b_i=\left(\frac{\beta}{\sigma(S^{d-1})T}\right)^{-1/\beta}\frac{(i\land
558: \sigma (S^{d-1})T/\beta)^{1-1/\beta}-((i-1)\land \sigma (S^{d-1})T/\beta)^{1-1/\beta}}{1-1/\beta},
559: \]
560: converges almost surely uniformly in $t$ to a L\'evy process whose
561: marginal law at time $1$ is $LS_{\alpha,\beta}(\sigma,q;0)$, with
562: $z_0=\int_{S^{d-1}}\xi\sigma(d\xi)/\sigma(S^{d-1})$.
563: This series representation directly reveals the nature of layering; all
564: jumps with absolute size greater than $1$ are due to the $\beta$-stable
565: shot noise series $\left(\frac{\beta \Gamma_i}{\sigma
566: (S^{d-1})}\right)^{-1/\beta}V_i$, while smaller jumps come from
567: $\left(\frac{\alpha \Gamma_i}{\sigma
568: (S^{d-1})}+1-\frac{\alpha}{\beta}\right)^{-1/\alpha}V_i$, which
569: resembles $\alpha$-stable jumps.
570: }\end{Exam}
571:
572: \section{Short and long time behavior}
573: We now present one of the two main results of this
574: section by giving the short time behavior of a layered stable process.
575: The results of this section were motivated by Section 3 of Rosi\'nski
576: \cite{super rosinski}, where stable behavior is obtained (for tempered stable processes)
577: in short time while Gaussian convergence is obtained in long time.
578: Here, in addition, we also obtain a further level of stable (non--Gaussian)
579: convergence in long time.
580: Recall that $\sigma_1$ and $\sigma_2$ are the finite positive measures
581: respectively given in (\ref{def of sigma1}) and (\ref{def of sigma2}),
582: and that for any $r>0,$ $T_r$ transforms the positive measure $\rho$, via
583: $(T_r\rho)(B)=\rho (r^{-1}B),$ $B\in \mathcal{B}(\mathbb{R}^d).$
584: For convenience, we will use the notation
585: $\nu_{\sigma,q}^{\alpha,\beta}$ for the L\'evy measure of
586: a layered stable process $LS_{\alpha,\beta}(\sigma,q;\eta)$ throughout
587: this section.
588:
589: \begin{Thm}\label{short of eS} \underline{\it Short time behavior}:
590: Let $\{X^{LS}_t:t\ge 0\}\sim LS_{\alpha,\beta}(\sigma,q;0)$, let
591: \[
592: \eta_{\alpha,\beta}=
593: \begin{cases}
594: \int_{S^{d-1}}\xi\sigma (d\xi)\int_0^1r q(r,\xi)dr,&if~\alpha\in (0,1),\\
595: -\int_{S^{d-1}}\xi\sigma (d\xi)\int_1^{\infty}r
596: q(r,\xi)dr,&if~(\alpha,\beta)\in (1,2)\times(1,\infty),\\
597: 0,&otherwise,
598: \end{cases}
599: \]
600: and let
601: \[
602: b_{\alpha,\beta}=
603: \begin{cases}
604: \frac{1}{\alpha -1}\int_{S^{d-1}}\xi\sigma_1 (d\xi),&if~(\alpha,\beta) \in
605: (1,2)\times (0,1],\\
606: 0,&otherwise.
607: \end{cases}
608: \]
609: Then,
610: \[
611: \{h^{-1/\alpha}(X^{LS}_{ht}+ht\eta_{\alpha,\beta})-t b_{\alpha,\beta}
612: :t\ge 0\}\stackrel{d}{\to}\{X^{(\alpha)}_t:t\ge 0\},\quad as ~h\to 0,
613: \]
614: where $\{X^{(\alpha)}_t:t\ge 0\}\sim S_{\alpha}(\sigma_1;0)$.
615: \end{Thm}
616:
617: \begin{proof}
618: Since a layered stable process is a L\'evy process, by a theorem of
619: Skorohod (see Theorem 15.17 of Kallenberg \cite{kallenberg}), it
620: suffices to show the weak convergence of its marginals at time $1$.
621: To this end, we will show the proper convergence of the generating
622: triplet of the infinitely divisible law, following Theorem 15.14 of
623: Kallenberg \cite{kallenberg}.
624:
625: For the convergence of the L\'evy measure, we need to show that as $h\to
626: 0$,
627: \[
628: h(T_{h^{-1/\alpha}}\nu^{\alpha,\beta}_{\sigma,q})\stackrel{v}{\to}
629: \nu^{\alpha}_{\sigma_1}, %\label{short condition 1}
630: \]
631: or equivalently,
632: \[
633: \lim_{h\to 0}\int_{\mathbb{R}_0^d}f(z)h(T_{h^{-1/\alpha}}\nu^{\alpha
634: ,\beta}_{\sigma,q})(dz)=\int_{\mathbb{R}_0^d}f(z)\nu^{\alpha}_{\sigma_1}(dz),
635: \]
636: for all bounded continuous function $f:\mathbb{R}^d_0\to \mathbb{R}$
637: vanishing in a neighborhood of the origin.
638: Letting $f$ be such a function with $|f|\le C<\infty$ and $f(z)\equiv 0$
639: on $\{z\in \mathbb{R}^d_0:\|z\|\le \epsilon\}$, for some $\epsilon>0$,
640: we get by (\ref{short q}),
641: \begin{eqnarray*}
642: \int_{\mathbb{R}_0^d}f(z)h(T_{h^{-1/\alpha}}\nu^{\alpha,\beta}_{\sigma,q})
643: (dz)&=&\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(h^{-1/\alpha}r\xi)h
644: q(r,\xi)dr\\
645: &=&\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(r\xi)h^{1+1/\alpha}
646: q(h^{1/\alpha}r,\xi)dr\\
647: &\to&\int_{S^{d-1}}c_1(\xi)\sigma(d\xi)\int_0^{\infty}f(r\xi)\frac{dr}{r^{\alpha+1}},
648: \end{eqnarray*}
649: as $h\to 0$, where the last convergence holds true since for $h\in (0,1)$,
650: \[
651: \left|\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(h^{-1/\alpha}r\xi)h
652: q(r,\xi)dr\right|\le C \left|\int_{S^{d-1}}\sigma (d\xi)\int_{\epsilon}^{\infty}q(r,\xi)dr\right|<\infty.
653: \]
654:
655: For the convergence of the Gaussian component, we need to show that for
656: each $\kappa >0,$
657: \[
658: \int_{\|z\|\le \kappa}zz'h(T_{h^{-1/\alpha}}\nu^{\alpha,\beta}_{\sigma,q})(dz)
659: \to \int_{\|z\|\le \kappa}zz'\nu^{\alpha}_{\sigma_1}(dz),%\label{short condition 2}
660: \]
661: as $h\to 0$.
662: Again, by (\ref{short q}),
663: \begin{eqnarray*}
664: \int_{\|z\|\le
665: \kappa}zz'h(T_{h^{-1/\alpha}}\nu^{\alpha,\beta}_{\sigma,q})(dz)
666: &=&\int_{S^{d-1}}\xi\xi'\sigma
667: (d\xi)\int_0^{h^{1/\alpha}\kappa}r^2h^{1-2/\alpha}q(r,\xi)dr\\
668: &=&\int_{S^{d-1}}\xi\xi'\sigma(d\xi)\int_0^{\kappa}r^2h^{1+1/\alpha}
669: q(h^{1/\alpha}r,\xi)dr\\
670: &\to&
671: \int_{S^{d-1}}\xi\xi'\sigma_1(d\xi)\int_0^{\kappa}r^2\frac{dr}{r^{\alpha+1}}\\
672: &=&\int_{\|z\|\le\kappa}zz'\nu^{\alpha}_{\sigma_1}(dz),
673: \end{eqnarray*}
674: where the passage to the limit is justified since, for $h\in (0,1),$
675: \[
676: \left\|\int_{S^{d-1}}\xi\xi'\sigma (d\xi)\int_0^{h^{1/\alpha}\kappa}r^2
677: h^{1-2/\alpha}q(r,\xi)dr\right\|_{{\rm o}}\le \left\|\int_{S^{d-1}}\xi\xi'
678: \sigma (d\xi)\int_0^{\kappa}r^2q(r,\xi)dr\right\|_{{\rm o}} <\infty.
679: \]
680:
681: For the convergence of the drift part, assume first that
682: $(\alpha,\beta) \notin (1,2)\times (0,1].$
683: For a $\sigma$-finite positive measure $\nu$ on $\mathbb{R}_0^d$, let
684: \begin{equation}\label{def of constant c}
685: C_{\alpha}(\nu):=
686: \begin{cases}
687: \int_{\|z\|\le 1}z\nu (dz),&{\rm if}~\alpha \in (0,1),\\
688: 0,&{\rm if}~\alpha =1,\\
689: -\int_{\|z\|>1}z\nu (dz),&{\rm if}~\alpha \in (1,2).
690: \end{cases}
691: \end{equation}
692: Clearly, $\eta_{\alpha,\beta}=C_{\alpha}(\nu_{\sigma,q}^{\alpha,\beta})$
693: and we then show that as $h\to 0$,
694: \[
695: C_{\alpha}(h(T_{h^{-1/\alpha}}\nu^{\alpha,\beta}_{\sigma,q}))-
696: \int_{\kappa <\|z\|\le 1}zh(T_{h^{-1/\alpha}}\nu^{\alpha,
697: \beta}_{\sigma,q})(dz) \to C_{\alpha}(\nu^{\alpha}_{\sigma_1})-
698: \int_{\kappa <\|z\|\le 1}z\nu^{\alpha}_{\sigma_1}(dz),
699: \]
700: for each $\kappa >0.$
701: Letting
702: \[
703: B=
704: \begin{cases}
705: \{z\in \mathbb{R}_0^d:\|z\|\le \kappa\},&{\rm if}~\alpha \in (0,1),\\
706: \{z\in \mathbb{R}_0^d:\kappa <\|z\|\le 1\},&{\rm if}~\alpha =1,\\
707: \{z\in \mathbb{R}_0^d:\|z\|>\kappa\},&{\rm if}~(\alpha,\beta)\in
708: (1,2)\times (1,\infty),\\
709: \end{cases}
710: \]
711: we have as $h\to 0$,
712: \begin{eqnarray*}
713: \int_{\mathbb{R}_0^d}{\bf 1}_B(z)zh(T_{h^{-1/\alpha}}\nu_{\sigma,q}^{\alpha,\beta})(dz)
714: &=&\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
715: 1}_B(h^{-1/\alpha}r\xi)rh^{1-1/\alpha} q(r,\xi)dr\\
716: &=&\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
717: 1}_B(r\xi)rh^{1+1/\alpha}q(h^{1/\alpha}r,\xi)dr\\
718: &\to&\int_{S^{d-1}}\xi\sigma_1(d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)r
719: \frac{dr}{r^{\alpha+1}},
720: \end{eqnarray*}
721: where the convergence holds true since for $h\in (0,1)$, and with the
722: help of (\ref{short q}),
723: \[
724: \left\|\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
725: 1}_B(r\xi)rh^{1+1/\alpha}q(h^{1/\alpha}r,\xi)dr\right\|\asymp
726: \left\|\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
727: 1}_B(r\xi)rq(r,\xi)dr\right\|<\infty.
728: \]
729: Finally, assume $(\alpha,\beta) \in (1,2)\times (0,1].$
730: Then, as $h\to 0$,
731: \[
732: -b_{\alpha,\beta}-\int_{\kappa<\|z\|\le 1}zh(T_{h^{-1/\alpha}}\nu_{\sigma,q}^{\alpha,\beta})(dz)\to-\int_{\|z\|>\kappa}z\nu_{\sigma_1}^{\alpha}(dz),
733: \]
734: for each $\kappa >0$, where the convergence holds true as before.
735: This completes the proof.
736: \end{proof}
737:
738: Our next result is also important.
739: Unlike in short time, the long time behavior of a layered stable
740: process depends on its outer stability index $\beta$.
741: This behavior is akin to a $\beta$-stable process if $\beta \in (0,2)$,
742: while akin to a Brownian motion whenever $\beta \in (2,\infty ).$
743:
744: \begin{Thm}\label{lS long-time theorem} \underline{\it Long time
745: behavior}: Let $\{X^{LS}_t:t\ge 0\}\sim
746: LS_{\alpha,\beta}(\sigma ,q;0).$
747:
748: \noindent (i) Let $\beta \in (0,2),$ let
749: \[
750: \eta_{\alpha,\beta}=
751: \begin{cases}
752: \int_{S^{d-1}}\xi\sigma (d\xi)\int_0^1rq(r,\xi)dr,&if~(\alpha,\beta)\in
753: (0,1)\times (0,1),\\
754: -\int_{S^{d-1}}\xi\sigma (d\xi)\int_1^{\infty}rq(r,\xi)dr,&if~\beta\in
755: (1,2),\\
756: 0,&otherwise,
757: \end{cases}
758: \]
759: and let
760: \[
761: b_{\alpha,\beta}=
762: \begin{cases}
763: \frac{1}{1-\beta}\int_{S^{d-1}}\xi\sigma_2(d\xi),&if~(\alpha,\beta)\in
764: [1,2)\times (0,1),\\
765: 0,&otherwise.
766: \end{cases}
767: \]
768: Then,
769: \[
770: \{h^{-1/\beta}(X^{LS}_{ht}+ht\eta_{\alpha,\beta})+tb_{\alpha,\beta}:t\ge 0\}\stackrel{d}{\to}
771: \{X^{(\beta)}_t:t\ge 0\},\quad as ~h\to \infty ,%\label{eS long 1}
772: \]
773: where $\{X^{(\beta)}_t:t\ge 0\}\sim S_{\beta}(\sigma_2;0).$
774:
775: \noindent (ii) Let $\beta \in (2,\infty )$ and let
776: \begin{equation}\label{def of constant a}
777: \eta=-\int_{S^{d-1}}\xi\sigma(d\xi)\int_1^{\infty}rq(r,\xi)dr.
778: \end{equation}
779: Then,
780: \begin{equation}
781: \{h^{-1/2}(X^{LS}_{ht}+ht\eta):t\ge 0\}\stackrel{d}{\to}
782: \{W_t:t\ge 0\},\quad as ~h\to \infty ,\label{eS long 2}
783: \end{equation}
784: where $\{W_t:t\ge 0\}$ is a centered Brownian motion with covariance
785: matrix $\int_{\mathbb{R}_0^d}zz'\nu_{\sigma,q}^{\alpha,\beta}(dz).$
786: \end{Thm}
787:
788: \begin{proof}
789: The claim (i) can be proved as (i) in Theorem \ref{short of eS}.
790: For the convergence of the L\'evy measure, we will show that
791: \[
792: \lim_{h\to
793: \infty}\int_{\mathbb{R}_0^d}f(z)h(T_{h^{-1/\beta}}\nu^{\alpha
794: ,\beta}_{\sigma,q})(dz)=\int_{\mathbb{R}_0^d}f(z)\nu^{\beta}_{\sigma_2}(dz),
795: \]
796: for all bounded continuous function $f:\mathbb{R}^d_0\to
797: \mathbb{R}$ vanishing in a neighborhood of the origin.
798: Letting $f$ be such a function with $|f|\le C<\infty$ and $f(z)\equiv 0$ on
799: $\{z\in \mathbb{R}^d_0:\|z\|\le \epsilon\}$, for some $\epsilon >0$, we
800: get by (\ref{long q}),
801: \begin{eqnarray*}
802: \int_{\mathbb{R}_0^d}f(z)h(T_{h^{-1/\beta}}\nu^{\alpha,\beta}_{\sigma,q})
803: (dz)&=&\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(h^{-1/\beta}r\xi)h
804: q(r,\xi)dr\\
805: &=&\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(r\xi)h^{1+1/\beta}
806: q(h^{1/\beta}r,\xi)dr\\
807: &\to&\int_{S^{d-1}}c_2(\xi)\sigma(d\xi)\int_0^{\infty}f(r\xi)\frac{dr}{r^{\beta+1}},
808: \end{eqnarray*}
809: as $h\to \infty$, where the last convergence holds true because of
810: (\ref{long q}) and since for sufficiently large $h>0$,
811: \begin{eqnarray*}
812: \left|\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}f(h^{-1/\beta}r\xi)h
813: q(r,\xi)dr\right|&=&\left|\int_{S^{d-1}}\sigma (d\xi)\int_{h^{1/\beta}
814: \epsilon}^{\infty}f(h^{-1/\beta}r\xi)hq(r,\xi)dr\right|\\
815: &\le&hC\int_{S^{d-1}}\sigma(d\xi)\int_{h^{1/\beta}\epsilon}^{\infty}q(r,\xi)dr\\&\asymp&hC
816: \sigma_2
817: (S^{d-1})\int_{h^{1/\beta}\epsilon}^{\infty}\frac{dr}{r^{\beta+1}}\\
818: &=&C\sigma_2(S^{d-1})\frac{\epsilon^{-\beta}}{\beta}<\infty.
819: \end{eqnarray*}
820: For the convergence of the Gaussian component, we have as $h\to \infty$
821: and for each $\kappa >0,$
822: \begin{eqnarray*}
823: \int_{\|z\|\le
824: \kappa}zz'h(T_{h^{-1/\beta}}\nu_{\sigma,q}^{\alpha,\beta})(dz)&=&\int_{S^{d-1}}\xi\xi'\sigma
825: (d\xi)\int_0^{\kappa}r^2h^{1+1/\beta}q(h^{1/\beta}r,\xi)dr\\
826: &\to&\int_{S^{d-1}}\xi\xi'\sigma_2
827: (d\xi)\int_0^{\kappa}r^2\frac{dr}{r^{\beta+1}}\\
828: &=&\int_{\|z\|\le \kappa}zz'\nu_{\sigma_2}^{\beta}(dz),
829: \end{eqnarray*}
830: where the passage to the limit is justified next.
831: Let $h\in (\kappa^{-\beta},+\infty)$ and write
832: \begin{eqnarray*}
833: \left\|\int_{\|z\|\le \kappa}zz'h(T_{h^{-1/\beta}}\nu_{\sigma,q}^{
834: \alpha,\beta})(dz)\right\|_{{\rm o}}&\le &
835: \left\|\int_{S^{d-1}}\xi\xi'\sigma(d\xi)\int_0^1h^{1-2/\beta}
836: r^2q(r,\xi)dr\right\|_{{\rm o}}\\
837: &&+\left\|\int_{S^{d-1}}\xi\xi'\sigma(d\xi)\int_1^{h^{1/\beta}\kappa}h^{1-2/\beta}
838: r^2q(r,\xi)dr\right\|_{{\rm o}}.
839: \end{eqnarray*}
840: The first term of the right hand side above is clearly bounded by
841: $\kappa^{2-\beta} \|\int_{S^{d-1}}\xi\xi'\sigma(d\xi)\int_0^1
842: r^2q(r,\xi)dr\|_{{\rm o}}$, while the second term is also bounded since
843: for $h\in (\kappa^{-\beta},+\infty),$
844: \begin{eqnarray*}
845: \left\|\int_{S^{d-1}}\xi\xi'\sigma(d\xi)\int_1^{h^{1/\beta}\kappa}h^{1-2/\beta}
846: r^2q(r,\xi)dr\right\|_{{\rm
847: o}}&\asymp&h^{1-2/\beta}\int_1^{h^{1/\beta}\kappa}r^2\frac{dr}{r^{\beta+1}}
848: \left\|\int_{S^{d-1}}\xi\xi'\sigma_2(d\xi)\right\|_{{\rm o}}\\
849: &=&\frac{\kappa^{2-\beta}-h^{1-2/\beta}}{2-\beta}\left\|\int_{S^{d-1}}
850: \xi\xi'\sigma_2(d\xi)\right\|_{{\rm o}}<\infty.
851: \end{eqnarray*}
852:
853: Finally, we study the convergence of the drift part.
854: Assume first that $(\alpha,\beta)\notin [1,2)\times (0,1).$
855: Let $C_{\beta}(\nu)$ be the constant defined as in (\ref{def of constant
856: c}) but depending on $\beta$ and $\nu$.
857: Clearly, $\eta_{\alpha,\beta}=C_{\beta}(\nu_{\sigma,q}^{\alpha,\beta}).$
858: We will then show that as $h\to \infty$,
859: \[
860: C_{\beta}(h(T_{h^{-1/\beta}}\nu^{\alpha,\beta}_{\sigma,q}))-\int_{\kappa
861: <\|z\|\le 1}zh(T_{h^{-1/\beta}}\nu^{\alpha,\beta}_{\sigma,q})(dz)\to
862: C_{\beta}(\nu^{\beta}_{\sigma_2})-\int_{\kappa <\|z\|\le
863: 1}z\nu^{\beta}_{\sigma_2}(dz),
864: \]
865: for each $\kappa >0$.
866: Letting
867: \[
868: B=
869: \begin{cases}
870: \{z\in \mathbb{R}_0^d:\|z\|\le \kappa\},&{\rm if}~\beta \in (0,1),\\
871: \{z\in \mathbb{R}_0^d:\kappa <\|z\|\le 1\},&{\rm if}~\beta =1,\\
872: \{z\in \mathbb{R}_0^d:\|z\|>\kappa\},&{\rm if}~\beta\in (1,2),\\
873: \end{cases}
874: \]
875: we have by (\ref{long q}) that
876: \begin{eqnarray*}
877: \int_{\mathbb{R}_0^d}{\bf 1}_B(z)zh(T_{h^{-1/\beta}}\nu_{\sigma,q}^{\alpha,\beta})(dz)
878: &=&\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
879: 1}_B(h^{-1/\beta}r\xi)rh^{1-1/\beta} q(r,\xi)dr\\
880: &=&\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf
881: 1}_B(r\xi)rh^{1+1/\beta}q(h^{1/\beta}r,\xi)dr\\
882: &\to&\int_{S^{d-1}}\xi\sigma_1(d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)r
883: \frac{dr}{r^{\beta+1}},
884: \end{eqnarray*}
885: as $h\to \infty$, where the convergence holds true since for $h\in
886: (1,\infty),$ and with the help of (\ref{long q}),
887: \[
888: \left\|\int_{S^{d-1}}\xi\sigma (d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)r
889: h^{1+1/\beta}q(h^{1/\beta}r,\xi)dr\right\|\asymp \left\|\int_{S^{d-1}}
890: \xi\sigma (d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)rq(r,\xi)dr\right\|<\infty.
891: \]
892: Next, let $(\alpha,\beta)\in [1,2)\times (0,1)$.
893: Then, observe that for each $\kappa >0$, and as $h\to \infty$,
894: \[
895: -b_{\alpha,\beta}-\int_{\kappa<\|z\|\le 1}zh(T_{h^{-1/\alpha}}\nu_{\sigma,q}^{\alpha,\beta})(dz)\to-\int_{\|z\|>\kappa}z\nu_{\sigma_1}^{\alpha}(dz),
896: \]
897: where the convergence holds true as before.
898: This completes the proof of (i).
899:
900: (ii) The random vector $h^{-1/2}X_{h}^{LS}$ is infinitely divisible
901: with generating triplet
902: \[
903: \left(-\int_{\|z\|\ge 1}z
904: h(T_{h^{-1/2}}\nu^{\alpha,\beta}_{\sigma,q})(dz),0,h(T_{h^{-1/2}}\nu_{\sigma,q}^{\alpha,\beta})\right).
905: \]
906: Letting $f$ be a bounded continuous function from $\mathbb{R}_0^d$ to
907: $\mathbb{R}$ such that $|f|\le C<\infty$ and $f(z)\equiv 0$ on $\{z\in
908: \mathbb{R}^d:\|z\|\le \epsilon\}$, for some $\epsilon >0$,
909: the L\'evy measure $h(T_{h^{-1/2}}\nu_{\sigma,q}^{\alpha,\beta})$
910: converges vaguely to zero as $h\to \infty$ since for sufficiently large
911: $h>0$,
912: \begin{eqnarray}
913: \nonumber \left|\int_{\mathbb{R}_0^d}f(z)h(T_{h^{-1/2}}\nu^{\alpha
914: ,\beta}_{\sigma,q})(dz)\right|&=& \left|\int_{S^{d-1}}\sigma
915: (d\xi)\int_{h^{1/2}\epsilon}^{\infty}f(h^{-1/2}r\xi)hq(r,\xi)dr\right|\\
916: \nonumber &\le&hC\int_{S^{d-1}}\sigma(d\xi)\int_{h^{1/2}\epsilon}^{\infty}
917: hq(r,\xi)dr\\
918: \nonumber &\asymp&hC\int_{S^{d-1}}c_2(\xi)\sigma(d\xi)\int_{h^{1/2}\epsilon}^{\infty}
919: \frac{dr}{r^{\beta+1}}\\
920: \label{long levy measure convergence} &=&h^{1-\beta/2}C\sigma_2(S^{d-1})\frac{\epsilon^{-\beta}}{\beta}\to 0,
921: \end{eqnarray}
922: as $h\to \infty$.
923: For the convergence of the Gaussian component, we have as $h\to +\infty$
924: and for each $\kappa >0$,
925: \begin{equation}\label{gaussian long equation}
926: \int_{\|z\|\le
927: \kappa}zz'h(T_{h^{-1/2}}\nu^{\alpha,\beta}_{\sigma,q})(dz)=
928: \int_{\|z\|\le h^{1/2}\kappa}zz'\nu_{\sigma,q}^{\alpha,\beta}(dz)\to \int_{\mathbb{R}_0^d}zz'\nu^{\alpha,\beta}_{\sigma,q}(dz),
929: \end{equation}
930: which is clearly well defined since $\int_{\mathbb{R}_0^d}\|z\|^2\nu^{\alpha,\beta}_{\sigma,q}(dz)<\infty$.
931: Finally, for sufficiently large $h>0$,
932: \begin{eqnarray}\nonumber
933: \left\|\int_{\|z\|>\kappa}zh(T_{h^{-1/2}}\nu^{\alpha,\beta}_{\sigma,q})
934: (dz)\right\|&=&h^{1/2}\left\|\int_{S^{d-1}}\xi\sigma
935: (d\xi)\int_{h^{1/2}\kappa}^{\infty}rq(r,\xi)dr\right\|\\
936: \nonumber &\asymp&h^{1/2}\int_{h^{1/2}\kappa}^{\infty}r\frac{dr}
937: {r^{\beta+1}}\left\|\int_{S^{d-1}}\xi\sigma_2(d\xi)\right\|\\
938: \label{long shift convergence}&=&h^{1-\beta/2}\frac{\kappa^{1-\beta}}
939: {\beta-1}\left\|\int_{S^{d-1}}\xi\sigma_2(d\xi)\right\|,
940: \end{eqnarray}
941: As $h\to \infty,$ (\ref{long shift convergence}) converges to zero and
942: this concludes the proof of (ii).
943: \end{proof}
944:
945: For $\beta =2$, layered stable processes do not seem to possess
946: any nice long time behavior, and this can be seen from the
947: improper convergence of the L\'evy measure, i.e., as $h\to \infty$,
948: $h(T_{h^{-1/2}}\nu^{\alpha ,2}_{\sigma,q})$ converges vaguely to
949: \[
950: \int_{S^{d-1}}\sigma_2(d\xi)\int_0^{\infty}{\bf 1}_B(r\xi)\frac{dr}{r^{2+1}},\quad
951: B\in \mathcal{B}(\mathbb{R}_0^d),
952: \]
953: which is not well defined as a L\'evy measure.
954: However, additional assumptions on $\sigma_2$ lead to the weak
955: convergence towards a Brownian motion as $\beta$ approaches to 2.
956:
957: \begin{Prop}\label{special Gaussian convergence}
958: Let $\{X^{LS}_t:t\ge 0\}\sim LS_{\alpha,\beta}(\sigma ,q;0)$ in $\mathbb{R}^d$.
959:
960: \noindent (i) Let $\beta \in (1,2)$ and let $\eta=-\int_{S^{d-1}}\xi
961: \sigma (d\xi)\int_1^{\infty}rq(r,\xi)dr$.
962: If $\sigma_2$ is uniform on $S^{d-1}$ such that
963: $\sigma_2(S^{d-1})=d(2-\beta)$, then
964: \[
965: \{h^{-1/\beta}(X^{LS}_{ht}+ht\eta):t\ge 0\}\stackrel{d}{\to}
966: \{W_t:t\ge 0\},\quad as~h\to \infty,~\beta \uparrow 2,
967: \]
968: where $\{W_t:t\ge 0\}$ is a $d$-dimensional (centered) standard Brownian
969: motion.
970: (The limit is taken over $h\to \infty$ first.)
971:
972: \noindent (ii) Let $\beta \in (2,\infty )$ and let $\eta$ be the constant
973: (\ref{def of constant a}).
974: If $\sigma_2$ is symmetric such that $\sigma_2(S^{d-1})=\beta -2,$ then
975: \[
976: \{h^{-1/2}(X_{ht}^{LS}+ht\eta):t\ge 0\}\stackrel{d}{\to}\{W_t:t\ge 0\},\quad
977: as~h\to \infty,~\beta\downarrow 2,
978: \]
979: where $\{W_t:t\ge 0\}$ is a centered Brownian motion with covariance
980: matrix $\int_{\mathbb{R}_0^d}zz'\nu_{\sigma,q}^{\alpha,2}(dz).$
981: (The limit can be taken either over $h\to \infty$ or over $\beta
982: \downarrow 2$ first.)
983: \end{Prop}
984:
985: \begin{proof}
986: (i) By Theorem \ref{short of eS} (i),
987: $h^{-1/\beta}(X^{LS}_h+h\eta)\stackrel{\mathcal{L}}{\to}X^{(\beta)}_1$, as
988: $h\to \infty$, where $\{X^{(\beta)}_t:t\ge 0\}\sim S_{\beta}(\sigma_2;0).$
989: Then, by E.18.7-18.8 of Sato \cite{sato}, we get
990: $\mathbb{E}[e^{i\langle
991: y,X^{(\beta)}_1\rangle}]=\exp[-c_{\beta,d}\|y\|^{\beta}],$ where
992: \[
993: c_{\beta,d}=\frac{\Gamma(d/2)\Gamma ((2-\beta)/2)}{2^{\beta}\beta\Gamma((\beta+d)/2)}\sigma_2(S^{d-1}).
994: \]
995: Taking $\beta\uparrow 2$ and since $\Gamma (x+1)=x\Gamma (x),$ $x>0$, we
996: get the result.
997:
998: (ii) In view of (\ref{long levy measure convergence}), we get
999: \[
1000: h^{1-\beta/2}C\sigma_2(S^{d-1})\frac{\epsilon^{-\beta}}{\beta}=
1001: h^{1-\beta/2}C(\beta -2)\frac{\epsilon^{-\beta}}{\beta}\to 0,
1002: \]
1003: as $\beta \downarrow 2$, which shows that the L\'evy measure
1004: $h(T_{h^{-1/2}}\nu_{\sigma,q}^{\alpha,\beta})$ converges vaguely to zero.
1005: Moreover, in view of (\ref{long shift convergence}), for sufficiently
1006: large $h>0$,
1007: \[
1008: h^{1-\beta/2}\frac{\kappa^{1-\beta}}{\beta-1}\left\|\int_{S^{d-1}}\xi\sigma_2(d\xi)\right\|=0,
1009: \]
1010: by the symmetry of $\sigma_2.$
1011: In view of (\ref{gaussian long equation}), it remains to show that
1012: $\int_{\mathbb{R}_0^d}\|z\|^2\nu_{\sigma,q}^{\alpha,2}(dz)<\infty.$
1013: Observe that
1014: \begin{eqnarray*}
1015: \int_{\mathbb{R}_0^d}\|z\|^2\nu_{\sigma,q}^{\alpha,\beta}(dz)=\int_{S^{d-1}}\sigma(d\xi)\int_0^1r^2q(r,\xi)dr+\int_{S^{d-1}}\sigma(d\xi)\int_1^{\infty}r^2q(r,\xi)dr,
1016: \end{eqnarray*}
1017: and the first term of the right hand side above is clearly uniformly
1018: bounded in $\beta\in [2,\infty)$, and so is the second term, since for
1019: every $\beta \in [2,\infty),$
1020: \[
1021: \int_{S^{d-1}}\sigma(d\xi)\int_1^{\infty}r^2q(r,\xi)dr\asymp
1022: \int_{S^{d-1}}c_2(\xi)\sigma (d\xi)\int_1^{\infty}r^2\frac{dr}{r^{\beta
1023: +1}}=\frac{\sigma_2(S^{d-1})}{\beta-2}=1.
1024: \]
1025: This concludes the proof in view of (\ref{eS long 2}).
1026: \end{proof}
1027:
1028: \begin{Rem}{\rm
1029: The short time behavior (Theorem \ref{short of eS}) and the
1030: (non-Gaussian) long time behavior (Theorem \ref{lS long-time theorem}
1031: (i)) can also be inferred from the series representation (\ref{original
1032: series}).
1033: For simplicity, consider the symmetric case.
1034: Letting $X_t:=\sum_{i=1}^{\infty}\overleftarrow{q}(\Gamma_i/T,V_i)V_i{\bf
1035: 1}(T_i\le t),$ we have
1036: \[
1037: h^{-1/\alpha}X_{ht}=\sum_{i=1}^{\infty}h^{-1/\alpha}\overleftarrow{q}
1038: (\Gamma_i/(hT),V_i)V_i{\bf 1}(hT_i\le ht),
1039: \]
1040: and so for each $u>0$ and each $\xi\in S^{d-1}$ such that
1041: $c_1(\xi)\in [0,\infty)$, bounded convergence gives
1042: \begin{eqnarray*}
1043: h^{-1/\alpha}\overleftarrow{q}(h^{-1}u,\xi)&=&h^{-1/\alpha}\inf
1044: \left\{r>0:\int_r^{\infty}q(s,\xi)ds<h^{-1}u\right\}\\
1045: &=&\inf
1046: \left\{r>0:\int_r^{\infty}h^{1+1/\alpha}q(h^{1/\alpha}s,\xi)ds<u\right\}\\
1047: &\to&\inf
1048: \left\{r>0:c_1(\xi)\int_r^{\infty}s^{-\alpha-1}ds<u\right\}=
1049: \left(\frac{\alpha u}{c_1(\xi)}\right)^{-1/\alpha},
1050: \end{eqnarray*}
1051: as $h\to 0$, which is indeed an $\alpha$-stable shot noise.
1052: The (non-Gaussian) long time behavior can be inferred just similarly.
1053: }\end{Rem}
1054:
1055: \section{Absolute continuity with respect to short time limiting stable process}
1056: Two L\'evy processes, which are mutually absolutely continuous, share any
1057: almost sure local behavior.
1058: The next theorem confirms this fact in relation with the short
1059: time behavior result of Theorem \ref{short of eS}.
1060: Indeed, given any layered stable process with respect to some
1061: probability measure, one can find a probability measure under which the
1062: layered stable process is identical in law to its short time limiting
1063: stable process.
1064: This result should be compared with Section 4 of Rosi\'nski
1065: \cite{super rosinski}.
1066:
1067: Recall that $c_1$ and $c_2$ are integrable (with respect
1068: to $\sigma$) functions on $S^{d-1}$ appearing in (\ref{short q}) and
1069: (\ref{long q}), while $\sigma_1$ and $\sigma_2$ are the finite positive
1070: measures (\ref{def of sigma1}) and (\ref{def of sigma2}), respectively.
1071: As before, we use the notation $\nu_{\sigma ,q}^{\alpha,\beta}$ for the
1072: L\'evy measure of a layered stable process $X:=\{X_t:t\ge 0\}\sim
1073: LS_{\alpha,\beta}(\sigma,q;\eta)$, while $\nu_{\sigma}^{\alpha}$ is the
1074: measure (\ref{def of stable levy measure}).
1075:
1076: \begin{Thm}\label{absolute continuity theorem}
1077: Let $\mathbb{P}$, $\mathbb{Q}$ and $\mathbb{T}$ be probability measures
1078: on $(\Omega ,\mathcal{F})$ such that under $\mathbb{P}$ the canonical
1079: process $\{X_t:t\ge 0\}$ is a L\'evy process in $\mathbb{R}^d$ with
1080: $\mathcal{L}(X_1)\sim LS_{\alpha,\beta}(\sigma,q;k_0)$, while under
1081: $\mathbb{Q}$ it is a L\'evy process with $\mathcal{L}(X_1)\sim
1082: S_{\alpha}(\sigma_1;k_1)$.
1083: Moreover, when $\beta \in (0,2)$ and under $\mathbb{T}$, $\{X_t:t\ge
1084: 0\}$ is a L\'evy process with $\mathcal{L}(X_1)\sim
1085: S_{\beta}(\sigma_2;\eta)$, for some $\eta\in \mathbb{R}^d$.
1086: Then,
1087:
1088: \noindent (i) $\mathbb{P}|_{\mathcal{F}_t}$ and $\mathbb{Q}|_{\mathcal{F}_t}$ are
1089: mutually absolutely continuous for every $t >0$ if and only if
1090: \[
1091: k_0-k_1=
1092: \begin{cases}
1093: \int_{S^{d-1}}\xi\sigma(d\xi)\int_0^1rq(r,\xi)dr,&\alpha \in (0,1),\\
1094: \int_{S^{d-1}}\xi\sigma(d\xi)\int_0^1r(q(r,\xi)-c_1(\xi)r^{-\alpha-1})dr,&\alpha
1095: =1,\\
1096: \frac{1}{\alpha
1097: -1}\int_{S^{d-1}}\xi \sigma_1(d\xi)+\int_{S^{d-1}}\xi\sigma(d\xi)\int_0^1r(q(r,\xi)-c_1(\xi)r^{-\alpha-1})dr,&\alpha \in (1,2).
1098: \end{cases}
1099: \]
1100:
1101: \noindent (ii) If $\alpha \ne \beta$, then for any choice of $\eta\in \mathbb{R}^d$,
1102: $\mathbb{P}|_{\mathcal{F}_t}$ and $\mathbb{T}|_{\mathcal{F}_t}$ are
1103: singular for all $t>0$.
1104:
1105: \noindent (iii) For each $t>0$,
1106: \[
1107: \frac{d\mathbb{Q}}{d\mathbb{P}}|_{\mathcal{F}_t}=e^{U_t},
1108: \]
1109: where $\{U_t:t\ge 0\}$ is a L\'evy process defined on $(\Omega
1110: ,\mathcal{F},\mathbb{P})$ by
1111: \begin{eqnarray}
1112: U_t&:=&\lim_{\epsilon\downarrow 0}\sum_{\{s\in (0,t]:\|\Delta X_s\|>\epsilon\}}\Bigg[
1113: \ln \left(\frac{q(\|\Delta X_s\|,\Delta X_s/\|\Delta X_s\|)}{c_1(\Delta
1114: X_s/\|\Delta X_s\|)\|\Delta X_s\|^{-\alpha-1}}\right)\nonumber \\
1115: &&\qquad \qquad \qquad \qquad -t (\nu_{\sigma,q}^{\alpha,\beta}-
1116: \nu_{\sigma_1}^{\alpha})(\{z\in\mathbb{R}_0^d:\|z\|>\epsilon\})\Bigg].
1117: \label{levy process u}
1118: \end{eqnarray}
1119: In the above right hand side, the convergence holds $\mathbb{P}$-a.s. uniformly
1120: in $t$ on every interval of positive length.
1121: %The L\'evy measure $\nu$ of $\mathcal{L}(U_1)$ is concentrated on $(-\infty,0)$
1122: % when $\alpha <\beta$ and is given by
1123: %\[
1124: % \nu (-\infty ,y)=\alpha^{-1}\sigma
1125: % (S^{d-1})e^{\frac{\alpha}{\alpha-\beta}y},\quad y<0,
1126: %\]
1127: %while when $\alpha >\beta$ it is concentrated on $(0,\infty)$ and is given by
1128: %\[
1129: % \nu (y,\infty)=\alpha^{-1}\sigma
1130: % (S^{d-1})e^{\frac{\alpha}{\alpha-\beta}y},\quad y>0.
1131: %\]
1132: \end{Thm}
1133:
1134: \begin{proof}
1135: \noindent (i) By Theorem 33.1 and Remark 33.3 of Sato~\cite{sato}, it is necessary
1136: and sufficient to show that the following three conditions hold;
1137: \begin{gather}
1138: \int_{\{z:|\varphi (z)|\le
1139: 1\}}\varphi(z)^2\nu_{\sigma_1}^{\alpha}(dz)<\infty, \label{absolute
1140: continuity condition 1}\\
1141: \int_{\{z:\varphi (z)>1\}}e^{\varphi
1142: (z)}\nu_{\sigma_1}^{\alpha}(dz)<\infty, \label{absolute continuity
1143: condition 2}\\
1144: \int_{\{z:\varphi
1145: (z)<-1\}}\nu_{\sigma_1}^{\alpha}(dz)<\infty,\label{absolute
1146: continuity condition 3}
1147: \end{gather}
1148: where the function $\varphi:\mathbb{R}_0^d\to \mathbb{R}$ is defined by
1149: $(d\nu_{\sigma,q}^{\alpha
1150: ,\beta}/d\nu_{\sigma_1}^{\alpha})(z)=e^{\varphi (z)},$ that is,
1151: \[
1152: \varphi (z)=\ln\left(\frac{q(\|z\|,z/\|z\|)}{c_1(z/\|z\|)\|z\|^{-\alpha-1}}
1153: \right),\quad z\in \mathbb{R}_0^d.
1154: \]
1155: Now, observe that
1156: \begin{equation}\label{absolute continuity domain 1}
1157: \lim_{\|z\|\to 0}\varphi (z)=\lim_{\|z\|\to 0}\ln
1158: \left(\frac{c_1(z/\|z\|)\|z\|^{-\alpha-1}}
1159: {c_1(z/\|z\|)\|z\|^{-\alpha-1}}\right)=0,
1160: \end{equation}
1161: and that as $\|z\|\to\infty$,
1162: \begin{equation}\label{absolute continuity domain 2}
1163: \varphi (z)\sim \ln \left(\frac{c_2(z/\|z\|)\|z\|^{-\beta-1}}
1164: {c_1(z/\|z\|)\|z\|^{-\alpha-1}}\right)=
1165: \ln \left(\frac{c_2(z/\|z\|)}{c_1(z/\|z\|)}\right)+(\alpha -\beta)\ln \|z\|\to
1166: \begin{cases}
1167: -\infty,&{\rm if}~\alpha <\beta,\\
1168: +\infty,&{\rm if}~\alpha >\beta.
1169: \end{cases}
1170: \end{equation}
1171: The conditions (\ref{absolute continuity condition 1}) and
1172: (\ref{absolute continuity condition 3}) are thus immediately
1173: satisfied, respectively, by (\ref{absolute continuity domain 1}) and
1174: (\ref{absolute continuity domain 2}) with $\alpha <\beta$.
1175: In view of (\ref{absolute continuity domain 2}) with $\alpha >\beta$,
1176: the condition (\ref{absolute continuity condition 2}) is satisfied since
1177: $\int_{\{z:\varphi (z)>1\}}e^{\varphi (z)}\nu_{\sigma_1}^{\alpha}(dz)$
1178: is bounded from above and below by constant multiples of $\int_{\|z\|>1}\frac{q(\|z\|,z/\|z\|)}{c_1(z/\|z\|)\|z\|^{-\alpha-1}}\nu_{\sigma_1}^{\alpha}(dz)
1179: =\nu_{\sigma,q}^{\alpha,\beta}(\{z\in\mathbb{R}_0^d:\|z\|>1\})$.
1180: When $\alpha=\beta \in (0,2)$, we have, by (\ref{absolute continuity
1181: domain 1}) and (\ref{absolute continuity domain 2}),
1182: \[
1183: \begin{cases}
1184: \lim_{\|z\|\to 0}\varphi (z)=0,&\\
1185: \lim_{\|z\|\to \infty}\varphi (z)=\lim_{\|z\|\to \infty}\ln
1186: \left(\frac{c_2(z/\|z\|)}{c_1(z/\|z\|)}\right)<\infty .&
1187: \end{cases}
1188: \]
1189: The condition (\ref{absolute continuity condition 1}) is then satisfied
1190: since $\int_{\{z:|\varphi (z)|\le
1191: 1\}}\varphi(z)^2\nu_{\sigma_1}^{\alpha}(dz)$ is bounded from above and
1192: below by constant multiples of
1193: $\int_{\|z\|>1}\varphi(z)^2\nu_{\sigma_1}^{\alpha}(dz)$, which is
1194: further bounded by $C \nu_{\sigma_1}^{\alpha}(\{z\in
1195: \mathbb{R}_0^d:\|z\|>1\})$ for some constant $C$.
1196: The conditions (\ref{absolute continuity condition 2}) and
1197: (\ref{absolute continuity condition 3}) are also satisfied since the domains
1198: $\{z\in \mathbb{R}_0^d:\varphi (z)>1\}$ and $\{z\in
1199: \mathbb{R}_0^d:\varphi(z)<-1\}$ are contained in some compact sets of
1200: $\mathbb{R}_0^d$.
1201:
1202: (ii) It suffices to show that either one of the following two
1203: conditions always fails;
1204: \begin{gather}
1205: \int_{\{z:\psi (z)>1\}}e^{\psi (z)}\nu_{\sigma_2}^{\beta}(dz)<\infty,
1206: \label{absolute continuity condition 5}\\
1207: \int_{\{z:\psi (z)<-1\}}\nu_{\sigma_2}^{\beta}(dz)<\infty,\label{absolute
1208: continuity condition 6}
1209: \end{gather}
1210: where the function $\psi:S^{d-1}\to \mathbb{R}$ is defined by
1211: $(d\nu_{\sigma,q}^{\alpha ,\beta}/d\nu_{\sigma_2}^{\beta})(z)=e^{\psi
1212: (z)},$ that is,
1213: \[
1214: \psi(z)=\ln \left(\frac{q(\|z\|,z/\|z\|)}
1215: {c_2(z/\|z\|)\|z\|^{-\beta-1}}\right),\quad z\in \mathbb{R}_0^d.
1216: \]
1217: As in the proof of (i), observe that
1218: \[
1219: \lim_{\|z\|\to \infty}\psi (z)=\lim_{\|z\|\to \infty}\ln \left(
1220: \frac{c_2(z/\|z\|)\|z\|^{-\alpha-1}}{c_2(z/\|z\|)\|z\|^{-\alpha-1}}\right)=0,
1221: \]
1222: and that as $\|z\|\to 0$,
1223: \[
1224: \psi (z)\sim \ln \left(
1225: \frac{c_1(z/\|z\|)\|z\|^{-\alpha-1}}{c_2(z/\|z\|)\|z\|^{-\beta-1}}\right)=
1226: \ln \left(\frac{c_1(z/\|z\|)}{c_2(z/\|z\|)}\right)+(\beta -\alpha)\ln \|z\|\to
1227: \begin{cases}
1228: +\infty,&{\rm if}~\alpha >\beta,\\
1229: -\infty,&{\rm if}~\alpha <\beta.
1230: \end{cases}
1231: \]
1232: Therefore, the condition (\ref{absolute continuity condition 5}) fails when
1233: $\alpha >\beta$ since
1234: \[
1235: \int_{\{z:\psi (z)>1\}}e^{\psi
1236: (z)}\nu_{\sigma_2}^{\beta}(dz)=\nu_{\alpha,q}^{\alpha,\beta}
1237: (\{z\in\mathbb{R}_0^d:\varphi (z)>1\})=+\infty,
1238: \]
1239: while (\ref{absolute continuity condition 6}) fails when $\alpha <\beta$ since
1240: $\nu_{\sigma_2}^{\beta}(\{z\in\mathbb{R}_0^d:\psi (z)<-1\})=+\infty.$
1241:
1242: (iii) This is a direct consequence of (i) with Theorem 33.2 of
1243: Sato~\cite{sato}.
1244: \end{proof}
1245:
1246: \begin{Rem}\label{absolute example}{\rm
1247: As in Remark \ref{special remark}, let
1248: \[
1249: q(r,\xi)=\sigma(S^{d-1})^{-1}(r^{-\alpha-1}{\bf
1250: 1}_{(0,1]}(r)+r^{-\beta-1}{\bf 1}_{(1,\infty)}(r)), ~\xi\in S^{d-1}.
1251: \]
1252: Then, the L\'evy process $\{U_t:t\ge 0\}$ given in (\ref{levy process
1253: u}) becomes
1254: \[
1255: U_t=(\alpha-\beta)\sum_{\{s\in (0,t]:\|\Delta X_s\|>1\}}\ln (\|\Delta
1256: X_s\|)-t\left(\frac{1}{\beta}-\frac{1}{\alpha}\right)\sigma (S^{d-1}).
1257: \]
1258: Intuitively speaking, $(d\mathbb{Q}/d\mathbb{P})|_{\mathcal{F}_t}$ replaces all
1259: $\beta$-stable jumps of a layered stable process up to time $t$ (i.e.,
1260: jumps with absolute size greater than $1$) by the corresponding
1261: $\alpha$-stable jumps without changing direction.
1262: Moreover, when $\alpha <\beta$, the L\'evy measure $\nu$ of $\mathcal{L}(U_1)$
1263: is concentrated on $(-\infty,0)$ and is given by
1264: \[
1265: \nu (-\infty,y)=\alpha^{-1}\sigma
1266: (S^{d-1})e^{\frac{\alpha}{\alpha-\beta}y},\quad y<0,
1267: \]
1268: while when $\alpha >\beta$, it is concentrated on $(0,\infty)$ and is
1269: given by
1270: \[
1271: \nu (y,\infty)=\alpha^{-1}\sigma
1272: (S^{d-1})e^{\frac{\alpha}{\alpha-\beta}y},\quad y>0.
1273: \]
1274: Let us next restate the absolute continuity result
1275: (Theorem \ref{absolute continuity theorem}) based on the fact that
1276: a series representation generates sample paths of a L\'evy process
1277: directly by generating every single jump.
1278: For simplicity, we consider the symmetric case.
1279: Let $\{Y_t:t\ge 0\}$ be an $\alpha$-stable process with
1280: $\mathcal{L}(Y_1)\sim S_{\alpha}(\sigma;k_1)$.
1281: By Lemma \ref{alpha series}, there exists a version of
1282: $\{Y_t:t\in [0,T]\}$ given by
1283: \[
1284: Y'_t=\sum_{i=1}^{\infty}\left(\frac{\alpha \Gamma_i}{\sigma
1285: (S^{d-1})T}\right)^{-1/\alpha}V_i{\bf 1}(T_i\le t)+k_1t.
1286: \]
1287: Also, let $\{X_t:t\ge 0\}$ be a layered stable process with
1288: $\mathcal{L}(X_1)\sim LS_{\alpha,\beta}(\sigma,q;k_0)$.
1289: In view of the series representation (\ref{special series}), there
1290: exists a version of $\{X_t:t\in [0,T]\}$ given by
1291: \begin{eqnarray*}
1292: && X'_t=\sum_{i=1}^{\infty}\Bigg[\left(\frac{\beta \Gamma_i}{\sigma
1293: (S^{d-1})T}\right)^{-1/\beta}{\bf 1}_{(0,\sigma
1294: (S^{d-1})T/\beta)}(\Gamma_i)\\
1295: &&\qquad \qquad \qquad +\left(\frac{\alpha \Gamma_i}{\sigma
1296: (S^{d-1})T}+1-\frac{\alpha}{\beta}\right)^{-1/\alpha}{\bf 1}_{(\sigma
1297: (S^{d-1})T/\beta,\infty)}(\Gamma_i)\Bigg]V_i{\bf 1}(T_i\le t)+k_0t,
1298: \end{eqnarray*}
1299: where all the random sequences are the same as those appearing in
1300: $\{Y'_t:t\in [0,T]\}$ above.
1301: By Theorem \ref{absolute continuity theorem}, they are
1302: mutually absolutely continuous if and only if
1303: \[
1304: k_0-k_1=
1305: \begin{cases}
1306: \frac{1}{\alpha-1}\int_{S^{d-1}}\xi\sigma_1(d\xi),&{\rm if}~\alpha \in
1307: (0,1)\cup (1,2),\\
1308: 0,&{\rm if}~\alpha =1.
1309: \end{cases}
1310: \]
1311: We infer that the L\'evy process $\{U_t:t\in [0,T]\}$ in the
1312: Radon-Nykodym derivative of Theorem \ref{absolute continuity theorem}
1313: (iii), that is,
1314: \[
1315: \frac{d\mathbb{Q}}{d\mathbb{P}}|_{\mathcal{F}_t}=e^{U_t},
1316: \]
1317: has a version given by
1318: \[
1319: U'_t=-\frac{\alpha -\beta}{\alpha}\sum_{i=1}^{\infty}\ln
1320: \left(\frac{\alpha \Gamma_i}{\sigma (S^{d-1})T}\right){\bf 1}_{(0,\sigma
1321: (S^{d-1})T/\alpha]}(\Gamma_i){\bf 1}(T_i\le
1322: t)-t\left(\frac{1}{\beta}-\frac{1}{\alpha}\right)\sigma (S^{d-1}).
1323: \]
1324: As a direct consequence, we have
1325: \[
1326: \mathbb{P}(X\in B)=\mathbb{E}_{\mathbb{P}}[e^{U'_T}{\bf 1}_B(Y')], \quad B\in \mathcal{B}(\mathbb{D}([0,T],\mathbb{R}^d)).
1327: \]
1328: Moreover, in view of Theorem 33.2 of Sato~\cite{sato},
1329: \[
1330: \frac{d\mathbb{P}}{d\mathbb{Q}}|_{\mathcal{F}_t}=e^{-U_t},
1331: \]
1332: and so we can derive a version of $\{U_t:t\in [0,T]\}$ in terms of the
1333: jumps of the layered stable process as follows;
1334: \[
1335: U''_t=-\frac{\alpha -\beta}{\beta}\sum_{i=1}^{\infty}\ln
1336: \left(\frac{\beta \Gamma_i}{\sigma (S^{d-1})T}\right){\bf 1}_{(0,\sigma
1337: (S^{d-1})T/\beta]}(\Gamma_i){\bf 1}(T_i\le
1338: t)-t\left(\frac{1}{\beta}-\frac{1}{\alpha}\right)\sigma (S^{d-1}).
1339: \]
1340: Similarly, we have
1341: \[
1342: \mathbb{Q}(Y\in B)=\mathbb{E}_{\mathbb{Q}}[e^{-U''_T}{\bf 1}_B(X')], \quad B\in \mathcal{B}(\mathbb{D}([0,T],\mathbb{R}^d)).
1343: \]
1344: }\end{Rem}
1345:
1346:
1347: \section{Concluding remarks}
1348:
1349: \noindent $\bullet$ The weak convergence towards a Brownian motion,
1350: proved in Proposition \ref{special Gaussian convergence} (i), is
1351: interesting in the sense that a stable process with uniformly dependent
1352: components converges in law to standard Brownian motion, i.e., with
1353: independent components.
1354: It is also interesting to see how a stable process with independent
1355: components converges towards a Brownian motion.
1356: To this end, for $i=1,\ldots,d$, let $a_{i}\in [0,\infty)$, let
1357: \[
1358: b_{i+}:=(0,\ldots,0,+1,0,\ldots,0),\quad b_{i-}:=(0,\ldots,0,-1,0,\ldots,0),
1359: \]
1360: where $+1$ and $-1$ are located in the $i$-th component, and finally set
1361: \[
1362: \sigma
1363: (d\xi):=\sum_{i=1}^d\frac{2-\alpha}{2}a_i(\delta_{b_{i+}}(d\xi)+
1364: \delta_{b_{i-}}(d\xi)),\quad \xi\in S^{d-1},
1365: \]
1366: where $\delta$ is the Dirac measure.
1367: Clearly, $\sigma$ is a symmetric finite positive measure on $S^{d-1}$.
1368: Also, let $\{X_t^{(\alpha)}:t\ge 0\}\sim S_{\alpha}(\sigma;0).$
1369: Then, if $y_i$ is the $i$-th component of $y$, we have, using
1370: $\Gamma (x+1)=x\Gamma (x),$ $x>0,$
1371: \begin{eqnarray*}
1372: \mathbb{E}[e^{i\langle y,X_1^{(\alpha)}\rangle}]&=&
1373: \exp\left[-\frac{\Gamma(1/2)\Gamma((2-\alpha)/2)}{2^{\alpha}\alpha\Gamma((1+\alpha)/2)}\int_{S^{d-1}}|\langle y,\xi\rangle
1374: |^{\alpha}\sigma(d\xi)\right]\\
1375: &=&\exp\left[-\frac{1}{2}\sum_{i=1}^d\frac{\Gamma(1/2)\Gamma(1+(2-\alpha)/2)}{2^{\alpha-2}\alpha\Gamma((1+\alpha)/2)}a_i|y_i|^{\alpha}\right]\\
1376: &\to&\exp\left[-\frac{1}{2}\sum_{i=1}^da_i|y_i|^2\right],\quad {\rm
1377: as}~\alpha \uparrow 2.
1378: \end{eqnarray*}
1379: Therefore, we get $\{X_t^{(\alpha)}:t\ge 0\}\stackrel{d}{\to}\{W_t:t\ge
1380: 0\}$ as $\alpha \uparrow 2$, where $\{W_t:t\ge 0\}$ is a Brownian motion
1381: with covariance matrix
1382: \[
1383: \begin{pmatrix}
1384: a_1 & 0 & \ldots & 0 \\
1385: 0 & a_2 & \ldots & 0 \\
1386: \vdots & \vdots & \ddots & \vdots \\
1387: 0 & 0 & \ldots & a_d \\
1388: \end{pmatrix}
1389: .
1390: \]
1391:
1392: \vspace{1em}
1393: \noindent $\bullet$ By making use of the absolute continuity of L\'evy measures, we can
1394: derive two more forms of series representations for a layered stable
1395: process induced by the L\'evy measure (\ref{original lS}), with $\alpha
1396: <\beta$.
1397: With the notations of Theorem \ref{absolute continuity theorem}, we get
1398: for $z\in \mathbb{R}_0^d,$
1399: \[
1400: \frac{d\nu_{\sigma,q}^{\alpha,\beta}}{d\nu_{\sigma}^{\alpha}}(z)={\bf
1401: 1}_{(0,1]}(\|z\|)+\|z\|^{\alpha-\beta}{\bf 1}_{(1,\infty)}(\|z\|) \le 1,
1402: \]
1403: and
1404: \[
1405: \frac{d\nu_{\sigma,q}^{\alpha,\beta}}{d\nu_{\sigma}^{\beta}}(z)=
1406: \|z\|^{\beta-\alpha}{\bf 1}_{(0,1]}(\|z\|)+{\bf 1}_{(1,\infty)}(\|z\|)\le 1.
1407: \]
1408: Then, by the rejection method of Rosi\'nski \cite{rosinski2}, the summands
1409: $\{\overleftarrow{q}(\Gamma_i/T,V_i)V_i\}_{i\ge 1}$ in (\ref{special series})
1410: can be respectively replaced by
1411: \[
1412: \left\{\left(\frac{\alpha
1413: \Gamma_i}{\sigma(S^{d-1})T}\right)^{-1/\alpha}{\bf 1}\left(
1414: \frac{d\nu_{\sigma,q}^{\alpha,\beta}}{d\nu_{\sigma}^{\alpha}}
1415: \left(\left(\frac{\alpha\Gamma_i}{\sigma(S^{d-1})T}\right)^{-1/\alpha}
1416: V_i\right)\ge U_i\right)V_i\right\}_{i\ge 1},
1417: \]
1418: and
1419: \[
1420: \left\{\left(\frac{\beta \Gamma_i}{\sigma(S^{d-1})T}\right)^{-1/\beta}
1421: {\bf 1}\left(\frac{d\nu_{\sigma,q}^{\alpha,\beta}}{d\nu_{\sigma}^{\beta}}
1422: \left(\left(\frac{\beta\Gamma_i}{\sigma(S^{d-1})T}\right)^{-1/\beta}V_i\right)\ge
1423: U_i\right)V_i\right\}_{i\ge 1},
1424: \]
1425: where $\{U_i\}_{i\ge 1}$ is a sequence of iid uniform random variables
1426: on $[0,1]$, independent of all the other random sequences.
1427:
1428: \vspace{1em}
1429: \noindent $\bullet$ In similarity to the work presented in \cite{ftsm},
1430: it is possible to define a notion of fractional layered stable motion (fLSm).
1431: Then, as in \cite{ftsm}, fLSm will, in short time, be close to
1432: fractional stable motion (with inner index $\alpha$) while in long time
1433: it is close to either fractional Brownian motion (if $\beta >2$) or to
1434: fractional stable motion (with index $\beta <2$).
1435:
1436: \vspace{1em}
1437: \noindent $\bullet$ Let us observe some sample paths of a layered
1438: stable process, generated via the series representation (\ref{special
1439: series}).
1440: By Theorem \ref{short of eS} and \ref{lS long-time theorem}, the entire
1441: situation is exhausted by the following three cases;
1442: \begin{enumerate2}
1443: \item[(i)] $\alpha <\beta <2,$
1444:
1445: \item[(ii)] $\beta \in (2,\infty),$
1446:
1447: \item[(iii)] $\alpha >\beta$ with $\beta \in (0,2)$.
1448: \end{enumerate2}
1449:
1450: Figure \ref{eS path} corresponds to the case (i) and typical sample
1451: paths of a symmetric layered stable process with
1452: $(\alpha,\beta)=(1.3,1.9)$ are drawn in short, regular, and long time
1453: span settings.
1454: For better comparison, we also drew its corresponding $1.3$-stable and
1455: $1.9$-stable processes.
1456: All these sample paths are generated via the series representation
1457: (\ref{special series}) for a layered stable process, or the one given in
1458: Lemma \ref{alpha series} for stable processes.
1459: Three sample paths within each figure are generated on a common
1460: probability space in the sense that a common set of random sequences
1461: $\{\Gamma_i\}_{i\ge 1}$, $\{V_i\}_{i\ge 1}$ and $\{T_i\}_{i\ge 1}$ are used.
1462: The desired short and long time behaviors are apparent.
1463: In the top figure, the layered stable process and its short time
1464: limiting stable process are almost indistinguishable in a graphical
1465: sense (of course, not in a probabilistic sense).
1466:
1467: For the case (ii), we drew in Figure \ref{eS path 2}
1468: typical sample paths of a symmetric layered stable process with
1469: $(\alpha,\beta)=(1.1,2.5)$, along with its corresponding $1.1$-stable
1470: process and a Brownian motion with a suitable variance.
1471: The layered stable process and the $1.1$-stable process are
1472: generated dependently as before, while the Brownian motion is
1473: independent of the others.
1474: As expected, the long time Gaussian type behavior (Theorem \ref{lS
1475: long-time theorem} (ii)) is clearly apparent.
1476: These stable type short time and Gaussian type long time behaviors have
1477: long been considered to be very appealing in applications.
1478: Such a study for asset price modeling will be presented elsewhere
1479: \cite{finance paper}.
1480:
1481: Finally, for the case (iii), we give in Figure \ref{eS path 3} typical
1482: sample paths of a symmetric layered stable process with
1483: $(\alpha,\beta)=(1.9,1.3)$, along with its corresponding $1.9$-stable
1484: and $1.3$-stable processes.
1485: Unlike the sample path behaviors observed in Figure \ref{eS path}, the
1486: path of the layered stable processes behaves more continuously (like a
1487: $1.9$-stable) in short time, while more discontinuously in long time (like
1488: a $1.3$-stable).
1489: In the short time figure, the layered stable and the $1.9$-stable are
1490: graphically indistinguishable.
1491:
1492: \begin{figure}
1493: \begin{center}
1494: \begin{tabular}{c}
1495: \resizebox{100mm}{58mm}{\includegraphics{eS_short.eps}} \\
1496: $t\in [0,0.03]$ \\
1497: \resizebox{100mm}{58mm}{\includegraphics{eS_regular.eps}} \\
1498: $t\in [0,3]$\\
1499: \resizebox{100mm}{58mm}{\includegraphics{eS_long.eps}} \\
1500: $t\in [0,100]$
1501: \end{tabular}
1502: \caption{Typical sample paths of layered stable process (---) with
1503: $(\alpha ,\beta )=(1.3,1.9)$, $1.3$-stable process ($\cdot \cdot \cdot$), and $1.9$-stable process (-$\cdot$-)}
1504: \label{eS path}
1505: \end{center}
1506: \end{figure}
1507:
1508: \begin{figure}
1509: \begin{center}
1510: \begin{tabular}{cc}
1511: \resizebox{100mm}{58mm}{\includegraphics{lS_short.eps}} \\
1512: $t\in [0,0.03]$\\
1513: \resizebox{100mm}{58mm}{\includegraphics{lS_regular.eps}} \\
1514: $t\in [0,3]$\\
1515: \resizebox{100mm}{58mm}{\includegraphics{lS_long.eps}} \\
1516: $t\in [0,100]$
1517: \end{tabular}
1518: \caption{Typical sample paths of layered stable process (---) with
1519: $(\alpha ,\beta )=(1.1,2.5)$, $1.1$-stable process ($\cdot \cdot
1520: \cdot$), and a Brownian motion (-$\cdot$-)}
1521: \label{eS path 2}
1522: \end{center}
1523: \end{figure}
1524:
1525: \begin{figure}
1526: \begin{center}
1527: \begin{tabular}{c}
1528: \resizebox{100mm}{58mm}{\includegraphics{lS_short2.eps}} \\
1529: $t\in [0,0.03]$ \\
1530: \resizebox{100mm}{58mm}{\includegraphics{lS_regular2.eps}} \\
1531: $t\in [0,3]$\\
1532: \resizebox{100mm}{58mm}{\includegraphics{lS_long2.eps}} \\
1533: $t\in [0,100]$
1534: \end{tabular}
1535: \caption{Typical sample paths of layered stable process (---) with
1536: $(\alpha ,\beta )=(1.9,1.3)$, $1.9$-stable process (-$\cdot$-), and
1537: $1.3$-stable process ($\cdot \cdot \cdot$)}
1538: \label{eS path 3}
1539: \end{center}
1540: \end{figure}
1541:
1542: \vspace{1em}
1543: \noindent $\bullet$ To finish this study, we briefly introduce another
1544: generalization of stable processes.
1545: Again, let $\mu$ be an infinitely divisible probability measure on
1546: $\mathbb{R}^d$ and without Gaussian component.
1547: Then, $\mu$ is {\it mixed stable} if its L\'evy measure is given by
1548: \begin{equation}\label{mixed stable levy measure}
1549: \nu(B)=\int_{(0,2)}\int_{S^{d-1}}\sigma (d\xi)\int_0^{\infty}{\bf
1550: 1}_B(r\xi)\frac{dr}{r^{\alpha+1}}\varphi (d\alpha),\quad B\in \mathcal{B}(\mathbb{R}_0^d),
1551: \end{equation}
1552: where $\varphi$ is a probability measure on $(0,2)$ such that
1553: \[
1554: \int_{(0,2)}\frac{1}{\alpha (2-\alpha)}\varphi (d\alpha)<\infty.
1555: \]
1556: Its characteristic function is given by
1557: \begin{eqnarray*}
1558: \widehat{\mu}(y)&=&\exp\Bigg[i\langle y,\eta\rangle -\int_{(0,2)}c_{\alpha}\int_{S^{d-1}}|\langle
1559: y,\xi\rangle|^{\alpha}(1-i\tan \frac{\pi \alpha}{2}{\rm sgn}\langle
1560: y,\xi\rangle )\sigma (d\xi)\varphi (d\alpha)\\
1561: &&\qquad \qquad \qquad \qquad \qquad -\varphi (\{1\})c_1\int_{S^{d-1}}(|\langle y,\xi\rangle
1562: |+i\frac{2}{\pi}\langle y,\xi\rangle \ln |\langle y,\xi\rangle| )\sigma (d\xi)\Bigg],
1563: \end{eqnarray*}
1564: for some $\eta \in \mathbb{R}^d$, and where $c_{\alpha}=|\Gamma
1565: (-\alpha)\cos \frac{\pi \alpha}{2}|$ when $\alpha \ne 1$ while $c_1=\pi/2.$
1566: Recall that in Example \ref{special remark} we defined the classes
1567: $L_m$, $m=0,1,\ldots$
1568: Let also $L_{\infty}:=\cap_{m=0}^{\infty}L_m$.
1569: It is proved in Sato \cite{sato class L} that an infinitely divisible
1570: probability measure without Gaussian component is in $L_{\infty}$ if and
1571: only if its L\'evy measure has the form (\ref{mixed stable levy
1572: measure}).
1573: We have seen in Example \ref{special remark} that an infinitely
1574: divisible probability measure $\mu$ is in $L_0$ if and only if the
1575: L\'evy measure of $\mu$ has the form
1576: \[
1577: \int_{S^{d-1}}\sigma(d\xi)\int_0^{\infty}{\bf
1578: 1}_B(r\xi)k_{\xi}(r)\frac{dr}{r},\quad B\in \mathcal{B}(\mathbb{R}_0^d),
1579: \]
1580: where $\sigma$ is a finite positive measure on $S^{d-1}$ and where
1581: $k_{\xi}(r)$ is a nonnegative function measurable in $\xi \in S^{d-1}$ and
1582: decreasing in $r>0.$
1583: Recently, Barndorff-Nielsen et al.\cite{bnms} defined the class
1584: $T$ by further requiring that the function $k_{\xi}(r)$ be completely
1585: monotone in $r$ for $\sigma$-a.e. $\xi$.
1586: Mixed stable distributions are indeed in the class $T$ since
1587: $\int_{(0,2)}r^{-\alpha}\varphi (d\alpha)$ is completely monotone.
1588:
1589: Finally, note that the associated L\'evy process that we call a {\it
1590: mixed stable process} possesses an interesting series representation.
1591: For simplicity, assume that $\sigma$ in (\ref{mixed stable levy
1592: measure}) is symmetric.
1593: Let $\{\Gamma_i\}_{i\ge 1},$ $\{T_i\}_{i\ge 1}$ and $\{V_i\}_{i\ge 1}$
1594: be random sequences defined as before.
1595: In addition, let $\{\alpha_i\}_{i\ge 1}$ be a sequence of iid random
1596: variables with common distribution $\varphi$.
1597: Assume moreover that all these random sequences are mutually independent.
1598: Then, with the help of the generalized shot noise method of Rosi\'nski
1599: \cite{rosinski2}, it can be shown that the stochastic process
1600: \[
1601: \left\{\sum_{i=1}^{\infty}\left(\frac{\alpha_i \Gamma_i}{\sigma
1602: (S^{d-1})T}\right)^{-1/\alpha_i}V_i{\bf 1}(T_i\le t):t\in [0,T]\right\},
1603: \]
1604: converges almost surely uniformly in $t$ to a mixed stable process whose
1605: marginal law at time $1$ is mixed stable with the L\'evy measure
1606: (\ref{mixed stable levy measure}).
1607: Comparing this result with the series representation of a stable process
1608: given in Lemma \ref{alpha series}, a mixed stable process can be thought
1609: of as a stable process with each of its jumps obeying a randomly chosen
1610: stability index.
1611:
1612: \begin{thebibliography}{999}
1613:
1614: \bibitem{bnms}
1615: Barndorff-Nielsen, O.E., Maejima, M., Sato, K. (2004)
1616: Some classes of multivariate infinitely divisible distributions
1617: admitting stochastic integral representation,
1618: {\it Preprint}.
1619:
1620: \bibitem{barndorff-nielsen shepard}
1621: Barndorff-Nielsen, O.E., Shepard, N. (2002)
1622: Normal modified stable processes,
1623: {\it Theory Probab. Math. Statist.} {\bf 65}, 1-20.
1624:
1625: \bibitem{bretagnolle}
1626: Bretagnolle, J. (1972)
1627: $p$-variation de fonctions al\'eatoires,
1628: In: {\it S\'eminaire de Probabilit\'es VI}, Lect. Notes in Math.
1629: {\bf 258}, Springer, 51-71.
1630:
1631: \bibitem{gikhman skorokhod}
1632: Gikhman, I.I., Skorokhod, A.V. (1969)
1633: {\it Introduction to the theory of random processes,}
1634: W.B. Saunders.
1635:
1636: \bibitem{ftsm}
1637: Houdr\'e, C., Kawai, R. (2005)
1638: On fractional tempered stable motion, {\it Preprint}.
1639:
1640: \bibitem{finance paper}
1641: Houdr\'e, C., Kawai, R. (2005)
1642: An empirical study on the time dependence structure of assets price
1643: dynamics, {\it In preparation}.
1644:
1645: \bibitem{kallenberg}
1646: Kallenberg, O. (2001)
1647: {\it Foundations of Modern Probability} (2nd ed.),
1648: Springer.
1649:
1650: \bibitem{lepage}
1651: LePage, R. (1980)
1652: Multidimensional infinitely divisible variables and processes II,
1653: {\it Lecture Notes Math.} {\bf 860}, Springer-Verlag, 279-284.
1654:
1655: \bibitem{rosinski2}
1656: Rosi\'nski, J. (2001)
1657: Series representations of L\'evy processes from the perspective of
1658: point processes,
1659: In: {\it L\'evy Processes - Theory and Applications,}
1660: Eds. Barndorff-Nielsen, O.-E., Mikosch, T., Resnick, S.I.,
1661: Birkh\"auser, 401-415.
1662:
1663: \bibitem{super rosinski}
1664: Rosi\'nski, J. (2004)
1665: Tempering stable processes,
1666: {\it Preprint}.
1667:
1668: \bibitem{samorodnitsky taqqu}
1669: Samorodnitsky, G. Taqqu, M.S. (1994)
1670: {\it Stable non-Gaussian random processes,}
1671: Chapman \& Hall, New York.
1672:
1673: \bibitem{sato class L}
1674: Sato, K. (1980)
1675: Class $L$ of multivariate distributions and its subclasses,
1676: {\it J. Multivariate Analysis} {\bf 10}, 207-232.
1677:
1678: \bibitem{sato}
1679: Sato, K. (1999)
1680: {\it L\'evy processes and infinitely divisible distributions,}
1681: Cambridge University Press.
1682:
1683: \end{thebibliography}
1684:
1685: \end{document}
1686:
1687: