1: 

2: %\documentstyle[twoside,11pt,thm]{article}

3: \documentclass[11pt]{article} 

4: \oddsidemargin0.2in 

5: \evensidemargin0.2in 

6: \topmargin-0.25in

7: \textwidth6.3in

8: \textheight8.75in

9: 

10: \input epsf

11: \input amssym.def

12: \input amssym.tex

13: 

14: \usepackage{graphicx}

15: \usepackage{caption2}

16: %\usepackage{latexsym}

17: \usepackage{thm}

18: 

19: \newcommand{\BB}{\Bbb B}

20: \newcommand{\beq}{\begin{equation}}

21: %\newcommand{\Box}{Q.E.D.}

22: \newcommand{\CAC}{{\cal C}}

23: \newcommand{\CAH}{{\cal H}}

24: \newcommand{\CAN}{{\cal N}}

25: \newcommand{\diam}{\mbox{diam}}

26: \newcommand{\eeq}{\end{equation}}

27: \newcommand{\gnaive}{\widetilde{1/\gamma}_{\eta}}

28: \newcommand{\gols}{\widehat{1/\gamma}_{\eta}}

29: \newcommand{\gr}{{\rm\mbox{gr}}}

30: \newcommand{\hnaive}{\widetilde{1/\upsilon}_{\eta}}

31: \newcommand{\hols}{\widehat{1/\upsilon}_{\eta}}

32: \newcommand{\naive}{\widetilde{\upsilon}_{\eta}}

33: \newcommand{\ols}{\widehat{\upsilon}_{\eta}}

34: \newcommand{\osc}{{\rm\mbox{Osc}}}

35: \newcommand{\prodi}{\mathop{{\lower3pt\hbox{\epsfxsize=7pt\epsfbox{pi.ps}}}}}

36: \newcommand{\proof}{\noindent{\bf Proof.}}

37: \newcommand{\RR}{\Bbb R}

38: \newcommand{\sas}{S\alpha S}

39: 

40: \newcommand{\blankbox}[2]{%

41:  \parbox{\columnwidth}{\centering

42:  \setlength{\fboxsep}{0pt}%

43:  \fbox{\raisebox{0pt}[#2]{\hspace{#1}}}%

44:  }%

45: }

46: 

47: 

48: \begin{document}

49: 

50: \title{Estimating the $p$-variation index of a sample function:\\

51: An application to financial data set}

52: %\vskip 1pc

53: \author{Rimas Norvai\v sa\thanks{This research was partially supported 

54: by NSERC Canada grant number 203232-98 at York University and by

55: Lithuanian State Science and Studies Foundation Grant K014.}\\

56: Vilnius

57: \and Donna Mary Salopek\thanks{This research was partially supported 

58: by NSERC Canada grant number 203232-98 at York University}\\

59: Toronto}

60: \date{\today}

61: 

62: %\maketitle

63: 

64: \noindent

65: {\bf\LARGE Estimating the $p$-variation index of a sample function:\\

66: An application to financial data set}

67: 

68: \vskip 2pc

69: \noindent

70: RIMAS NORVAI\v SA\footnote{This research was partially supported 

71: by NSERC Canada grant number 203232-98 at York University and by

72: Lithuanian State Science and Studies Foundation Grant K014.}

73: \hfill norvaisa@ktl.mii.lt

74: 

75: \noindent

76: {\em\small Institute of Mathematics and Informatics, Akademijos 4, 

77: Vilnius 2600, Lithuania}

78: 

79: \vskip 1pc

80: \noindent

81: DONNA MARY SALOPEK\footnote{This research was partially supported 

82: by NSERC Canada grant number 203232-98 at York University.}

83: \hfill dsalopek@mathstat.yorku.ca

84: 

85: \noindent

86: {\em\small York University, Department of Mathematics and Statistics,

87: 4700 Keele Street,} 

88: 

89: \noindent

90: {\em\small Toronto, Ontario M3J 1P3 Canada}

91: 

92: \vskip 1pc

93: \noindent

94: {\small {\bf Abstract.}

95: In this paper we apply a real analysis approach to test continuous

96: time stochastic models of financial mathematics.

97: Specifically, fractal dimension estimation methods are applied to

98: statistical analysis of continuous time stochastic processes.

99: To estimate a roughness of a sample function we modify a box-counting 

100: method typically used in estimating fractal dimension of a graph

101: of a function. 

102: Here the roughness of a function $f$ is defined as the infimum of numbers

103: $p>0$ such that $f$ has bounded $p$-variation, which we call the 

104: $p$-variation index of $f$.

105: The method is also tested on estimating the exponent $\alpha\in [1,2]$ of a 

106: simulated symmetric $\alpha$-stable process, and on estimating the Hurst 

107: exponent $H\in (0,1)$ of a simulated fractional Brownian motion.}

108: 

109: \vskip 1pc

110: \noindent

111: {\small {\bf Keywords:} estimation, $p$-variation index, box-counting index, 

112: financial data analysis}

113: 

114: \vskip 1pc

115: \noindent

116: {\small {\bf AMS 1991 Subject Classification:} 90A20, 62M09, 60G17  }

117: 

118: \section{Introduction: the $p$-variation index and financial modeling}

119: 

120: 

121: In financial mathematics, a simplest continuous time  

122: model assumes that a stock price, or other financial asset,

123: is a stochastic process $P=\{P(t)\colon\,0\leq t\leq T\}$ satisfying

124: the relation 

125: \beq\label{BSM-model}

126: P(t)=1+(I)\int_0^tP\,dB,\qquad 0\leq t\leq T,

127: \eeq

128: which is the It\^o integral equation with respect to a standard Brownian 

129: motion $B=\{B(t)\colon\,t\geq 0\}$.

130: Equation (\ref{BSM-model}) is usually written in the form of

131: a stochastic differential equation, or simply by giving its solution

132: $P(t)=\exp\{B(t)-t/2\}$, $0\leq t\leq T$.

133: In the financial literature this is known as the Black-Scholes-Merton 

134: stock price model.

135: Its pertinence is backed-up by the assumption that increments of the log 

136: transform of a stock price are independent and normally distributed, which

137: is known as the strong form of Random Walk Hypothesis.

138: In agreement with a relaxed form or an alternative form of the Random Walk 

139: Hypothesis, the Brownian motion $B$ in equation (\ref{BSM-model}) can be 

140: replaced by a more general stochastic process $X$, 

141: and the linear It\^o integral equation

142: (\ref{BSM-model}) by a different integral equation with respect to $X$.

143: In this paper, the stochastic process $X$ is called the {\em return process},

144: and the unique solution of an integral equation with respect to $X$ 

145: is called the {\em stock price process} $P$. 

146: 

147: The mainstream econometric analysis of continuous time financial models is

148: to test different hypotheses about an integral equation describing a stock 

149: price process $P$, or to test various parameters of a distribution of a return 

150: process $X$ (see e.g.\ Section 9.3 in Campbell, Lo and MacKinlay, 1997).

151: In this paper we attempt to test the degree of roughness of a

152: return process $X$.

153: The legitimacy of such an endeavor is based on the fact that

154: the support of a distribution of a stochastic processes is

155: a particular class of functions.

156: That is,  a suitable class of functions contains almost all

157: sample functions  of a stochastic process.

158: More specifically, if $X$ is a regular enough stochastic process

159: defined on a given probability space $(\Omega,{\cal F},\Pr)$, then for a 

160: suitable class of functions $F$, a {\em sample function} 

161: $X(\omega)=\{X(t,\omega)\colon\, 0\leq t\leq T\}$ belongs to $F$ 

162: for almost all $\omega\in\Omega$.

163: Often $F$ can be taken as a proper subspace of the space of all 

164: continuous functions on $[0,T]$, or a proper subspace of the (Skorohod) 

165: space of all regulated and right continuous functions on $[0,T]$.

166: For example, the support of the distribution of a standard Brownian

167: motion is a subset of a class of functions having the order

168: of H\"older continuity strictly bigger than $1/2$.

169: However, the H\"older continuity is not applicable to characterize 

170: sample functions of a L\'evy stochastic process without a Gaussian component,

171: because almost every sample function of such a process is discontinuous. 

172: A simple example of a L\'evy process is a symmetric $\alpha$-stable

173: stochastic process, or a $\sas$ stochastic process, $X_{\alpha}$ with 

174: the exponent $\alpha\in (0,2]$.

175: The case $\alpha =2$ gives the only Gaussian component, a standard Brownian 

176: motion, that is, $X_2=B$.

177: In general, if $X_{\alpha}$ is a $\sas$ stochastic process with 

178: $\alpha\in (0,2]$,  then almost all sample functions of $X_{\alpha}$ 

179: have bounded $p$-variation for each $p>\alpha$, and have infinite

180: $p$-variation if $0<p\leq\alpha$.

181: This fact was known since the time P.\ L\'evy first introduced 

182: this process around the beginning of thirties.

183: 

184: To recall the property of boundedness of $p$-variation, let

185: $f$ be a real-valued function on an interval $[0,T]$.

186: For a number $0<p<\infty$, let 

187: $$

188: s_p(f;\kappa):=\sum_{i=1}^n|f(t_i)-f(t_{i-1})|^p, 

189: $$

190: where $\kappa=\{t_i\colon\,i=0,\dots,n\}$ is a partition of $[0,T]$,

191: that is $0=t_0<t_1<\cdots <t_n=T$.

192: The {\em $p$-variation} of $f$ is defined by

193: $$

194: v_p(f):=v_p(f;[0,T]):=\sup\big\{s_p(f;\kappa)\colon\,

195: \kappa\,\,\mbox{is a partition of}\,\, [0,T]\big\}.

196: $$ 

197: The function $f$ has bounded $p$-variation on $[0,T]$ if $v_p(f)<\infty$.

198: By H\"older's inequality it follows that if $v_p(f)<\infty$ and

199: $q>p$, then $v_q(f)<\infty$.

200: The number

201: $$

202: \upsilon (f;[0,T]):=\inf\big\{p>0\colon\,v_p(f;[0,T])<\infty\big\}

203: =\sup\big\{p>0\colon\,v_p(f;[0,T])=\infty\big\}

204: $$ 

205: is called the {\em $p$-variation index} of $f$.

206: For a regular enough stochastic process $X$, $\upsilon (X)(\omega):=

207: \upsilon (X(\cdot,\omega))$, $\omega\in\Omega$, is a random variable

208: which we call the $p$-variation index of $X$.

209: In fact, for all stochastic processes discussed in this paper, their

210: $p$-variation indices are known to be constants and we seek

211: to estimate these constants.   

212: 

213: Now for a $\sas$ stochastic process

214: $X_{\alpha}$, we can restate its sample regularity by saying that 

215: its $p$-variation index $\upsilon (X_{\alpha})=\alpha$ almost surely.

216: This is a special case of the following more general fact.

217: 

218: \begin{exmp}\label{exmp1}

219: Let $X$ be a homogeneous L\'evy stochastic process with the L\'evy

220: measure $\nu$, which is a $\sigma$-finite Borel measure on $\RR\setminus \{0\}$

221: such that

222: $$

223: \int_{\RR\setminus\{0\}}\min\{1, |x|^2\}\,\nu (dx)<+\infty.

224: $$

225: The {\em Blumenthal-Getoor index} $\beta_X$ of $X$ is defined by

226: $$

227: \beta_X :=\inf\big\{ \alpha>0\colon\,\int_{\RR\setminus\{0\}}\min\{1,

228: |x|^{\alpha}\}\,\nu (dx)<+\infty\big\}.

229: $$

230: Note that $0<\beta_X\leq 2$.

231: If $X$ has no Gaussian part, then for any $0<T<\infty$,

232: $$ 

233: \upsilon(X;[0,T])=\beta_X,\qquad\mbox{almost surely.} 

234: $$

235: This follows from Theorems 4.1 and 4.2 of Blumenthal and Getoor (1961),

236: and from Theorem 2 of Monroe (1972).

237: \end{exmp}

238: 

239: A stock price model having as a return a L\'evy process without

240: a Gaussian part is a common alternative to the Black-Scholes-Merton model.

241: Another popular alternative is a fractional Brownian

242: motion $B_H$ with the Hurst exponent $H\in (0,1)$, where

243: $B_H$ with $H=1/2$ is a standard Brownian motion. 

244: The $p$-variation index of a fractional Brownian motion

245: $\upsilon (B_H;[0,T])=1/H$ almost surely.

246: This is the special case of the following fact.

247: 

248: \begin{exmp}\label{exmp2}

249: Let $X=\{X(t)\colon\,t\geq 0\}$ be a Gaussian stochastic process

250: with stationary increments and continuous in quadratic mean.

251: Let $\sigma_X$ be the incremental variance of $X$ defined by

252: $\sigma_X(u)^2:=E[(X(t+u)-X(t))^2]$ for $t, u\geq 0$.

253: Let

254: $$

255: \gamma_{\ast}:=\inf\big\{\gamma >0\colon\,\lim_{u\downarrow 0}

256: \frac{u^{\gamma}}{\sigma_X(u)}=0\big\}\quad\mbox{and}\quad

257: \gamma^{\ast}:=\sup\big\{\gamma >0\colon\,\lim_{u\downarrow 0}

258: \frac{u^{\gamma}}{\sigma_X(u)}=+\infty\big\}.

259: $$

260: Then $0\leq\gamma_{\ast}\leq\gamma^{\ast}\leq +\infty$.

261: If $\gamma_{\ast}=\gamma^{\ast}$, then we say that $X$ has

262: an {\em Orey index} $\gamma_X:=\gamma_{\ast}=\gamma^{\ast}$.

263: Furthermore, if $X$ has an Orey index $\gamma_X\in (0,1)$, then for any

264: $0<T<+\infty$,

265: \beq\label{var-of-Orey}

266: \upsilon (X;[0,T])=1/\gamma_X,\qquad\mbox{almost surely.}

267: \eeq

268: This follows from the fact that almost all sample functions of

269: $X$ obey a uniform H\"older condition with exponent $\gamma

270: <\gamma_X$ (see Section 9.4 of Cramer and Leadbetter, 1967)

271: and from the inequality of Berman (1969) connecting the

272: $p$-variation with the Fourier transform of local times of $X$.

273: Relation (\ref{var-of-Orey}) also follows from the characterization

274: of the $p$-variation index for arbitrarily Gaussian processes due to

275: Jain and Monrad (1983).

276: \end{exmp}

277: 

278: From the point of view of a statistical time series analysis,

279: estimation of the $p$-variation index in the above two examples offer 

280: a new perspective to analyzing financial data sets.

281: For instance, a symmetric $\alpha$-stable process and a fractional 

282: Brownian motion with the Hurst exponent $H$, both have the same $p$-variation

283: index in the case $\alpha=1/H\in (1,2)$. 

284: However the latter has exponentially small tails, while the former 

285: has not even the second moment.

286: These two examples are extensions of the Black-Scholes-Merton model

287: (that is when $\alpha=1/H=2$) into two different directions.

288: The Orey index, and so the $p$-variation index by relation (\ref{var-of-Orey}),

289: have already been estimated in the paper Norvai\v sa and Salopek (2000).

290: They used two estimators based on the result of Gladyshev (1961).  

291: The estimators of the present paper can be applied under much less

292: restrictive hypotheses about stock price returns, and helps to reconcile

293: the two divergent directions of theoretical analyses of financial markets. 

294: 

295: \section{The oscillation $\eta$-summing index and related estimators}

296: 

297: In this section we describe a method of estimating the $p$-variation

298: index of a function based on existence of the metric entropy index

299: (or the box-counting dimension) of its graph.

300: Let $f$ be a real-valued function defined on an interval $[0,T]$.

301: Let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing

302: positive integers.

303: With $\eta$ one can associate a sequence $\{\lambda (m)\colon\,m\geq 1\}$

304: of partitions $\lambda (m)=\{iT/N_m\colon\,i=0,\dots,N_m\}$ of $[0,T]$ into 

305: subintervals $\Delta_{i,m}:=[(i-1)T/N_m,iT/N_m]$, $i=1,\dots,N_m$, 

306: all having the same length $T/N_m$.

307: For each $m\geq 1$, let

308: \beq\label{Q(f)}

309: Q(f;\lambda (m))=\sum_{i=1}^{N_m}\osc (f;\Delta_{i,m}),

310: \eeq

311: where for a subset $A\subset [0,T]$,

312: $$

313: \osc (f;A):=\sup\big\{|f(t)-f(s)|\colon\,s, t\in A\big\}

314: =\sup_{t\in A}f(t)-\inf_{s\in A}f(s).

315: $$

316: The sequence $Q_{\eta}(f):=\{Q(f;\lambda (m))\colon\,m\geq 1\}$ will be called

317: the {\em oscillation $\eta$-summing} sequence.

318: For a bounded non-constant function $f$ on $[0,T]$, and any sequence $\eta$ 

319: as above, let

320: $$

321: \delta_{\eta}^{-}(f):=

322: \liminf_{m\to\infty}\frac{\log Q(f;\lambda (m))/N_m}{\log (1/N_m)}

323: \quad\mbox{and}\quad\delta_{\eta}^{+}(f)

324: :=\limsup_{m\to\infty}\frac{\log Q(f;\lambda (m))/N_m}{\log (1/N_m)}.

325: $$

326: Then we have

327: \beq\label{Qin[1,2]}

328: 0\leq\delta_{\eta}^{-}(f)\leq\delta_{\eta}^{+}(f)\leq 1.

329: \eeq

330: The lower bound follows from the bound $Q(f;\lambda (m))/N_m\leq 

331: \osc (f;[0,T])<\infty$, which is valid for each $m\geq 1$.

332: The upper bound holds because $Q(f;\lambda (m))\geq C/2>0$ for 

333: all sufficiently large $m\geq 1$.

334: Indeed, if $f$ is continuous, then $C=v_1(f;[0,T])$. 

335: Otherwise $f$ has a jump at some $t\in [0,T]$, so that $C$ can be taken to be

336: a saltus at $t$ if $t\not\in\lambda (m)$ for all sufficiently large $m$,

337: or $C$ can be taken to be a one-sided non-zero saltus at $t$ if 

338: $t\in\lambda (m)$ for infinitely many $m$. 

339: Instead of relation (\ref{Qin[1,2]}), a sharp one-sided bound is given by

340: Lemma \ref{bc-pv} in Appendix A.

341: 

342: If $f$ has bounded variation, then for each $\eta$,

343: \beq\label{bv}

344: \delta_{\eta}^{-}(f)=\delta_{\eta}^{+}(f)=1.

345: \eeq

346: Indeed, since $Q(f;\lambda (m))\leq v_1(f;[0,T])<\infty$,

347: we have $\log Q(f;\lambda (m))/N_m\geq \log v_1(f;[0,T])/N_m$

348: for all sufficiently large $m$.

349: Thus $\delta_{\eta}^{-}(f)\geq 1$, and equalities in (\ref{bv}) follow from 

350: relation (\ref{Qin[1,2]}).

351: 

352: 

353: \begin{defn}\label{OS-index}

354: Let $f$ be a non-constant real-valued function on $[0,T]$,

355: and let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing

356: positive integers.

357: If $\delta_{\eta}^{-}(f)=\delta_{\eta}^{+}(f)$, then we say that $f$ has

358: the {\em oscillation $\eta$-summing index} $\delta_{\eta}(f)$ 

359: and is defined by

360: \beq\label{1OS-index}

361: \delta_{\eta}(f):=\delta_{\eta}^{-}(f)=\delta_{\eta}^{+}(f)

362: =\lim_{m\to\infty}\frac{\log Q(f;\lambda (m))/N_m}{\log (1/N_m)}. 

363: \eeq

364: \end{defn}

365: 

366: Next we give a sufficient condition for exitence of the oscillation

367: $\eta$-summing index for any $\eta$.

368: Let $E$ be a nonempty bounded subset in a plane $\RR^2$, and let 

369: $N(E;\epsilon)$, $\epsilon >0$, be the minimum number of closed balls of 

370: diameter $\epsilon$  required to cover $E$.

371: The {\em lower} and {\em upper metric entropy indices} of the set $E$ are 

372: defined respectively by

373: $$

374: \Delta^{-}(E):=\liminf_{\epsilon\downarrow 0}\frac{\log N(E;\epsilon)}

375: {\log (1/\epsilon)}\quad\mbox{and}\quad

376: \Delta^{+}(E):=\limsup_{\epsilon\downarrow 0}\frac{\log N(E;\epsilon)}

377: {\log (1/\epsilon)}.

378: $$

379: If $\Delta^{-}(E)=\Delta^{+}(E)$, then the common value denoted by

380: $\Delta (E)$ is called the {\em metric entropy index} of the set $E$.

381: In the actual calculations of the metric entropy index, it is often simpler

382: to replace closed balls by squares (boxes) of a grid (cf.\ Lemma \ref{boxes}

383: below).

384: Therefore in fractal analysis, $\Delta (E)$ is also known as 

385: the box-counting dimension.

386: 

387: The proof of the following theorem is given in Appendix A.

388: 

389: \begin{thm}\label{main}

390: Let $f$ be a regulated non-constant function on $[0,T]$ with the $p$-variation

391: index $\upsilon (f)$.

392: If the metric entropy index of the graph $\gr (f)$  of $f$ is defined

393: and

394: \beq\label{2main}

395: \Delta (\gr (f))=2-1/(1\vee\upsilon (f)), 

396: \eeq

397: then for any sequence $\eta$, $f$ has the oscillation $\eta$-summing index 

398: \beq\label{index=var}

399: \delta_{\eta}(f)=1/(1\vee\upsilon (f)).

400: \eeq 

401: \end{thm}

402: 

403: Essentially, relation (\ref{2main}) is the  lower bound condition

404: on the metric entropy index because the following always holds. 

405: 

406: \begin{prop}\label{3main}

407: For a regulated function $f$ on $[0,T]$,

408: $\Delta^{+} (\gr (f))\leq 2-1/(1\vee\upsilon (f))$.

409: \end{prop}

410: 

411: The proof is similar to the proof of Theorem \ref{main}

412: and is also given in Appendix A.

413: 

414: The oscillation $\eta$-summing index is a slightly modified concept

415: of a real box-counting method introduced by Carter, Cawley and Mauldin (1988).

416: Independently, Dubuc, Quiniou, Roques-Carmes, Tricot and Zucker 

417: (1989) arrived at essentially the same notion, but called it

418: the variation method.

419: Both papers applied the new index to estimate 

420: the fractal dimension of several continuous functions whose dimension

421: is known, and found the new algorithm superior over several other

422: fractal dimension estimation methods (see also Section 6.2 in 

423: Cutler, 1993, for further discussion on this).

424: 

425: 

426: \paragraph*{Oscillation $\eta$-summing estimators.}

427: Let $f$ be a real-valued function defined on $[0,1]$, and let 

428: $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing

429: positive integers.

430: Let $\{u_1,\dots,u_N\}\subset [0,1]$ be a set of points such that

431: for some integer $M$,

432: \beq\label{1ose}

433: \cup_{m=1}^M\lambda (m)=\{u_1,\dots,u_N\},\qquad

434: \mbox{where $\lambda (m):=\{i/N_m\colon\,i=0,\dots,N_m\}$.}

435: \eeq

436: Given a finite set of values $\{f(u_1),\dots,f(u_N)\}$,

437: we want to estimate the $p$-variation

438: index $\upsilon (f)$.

439: To achieve this, for each $m\in\{1,\dots,M\}$, let

440: \beq\label{Q(m)}

441: Q(m):=\sum_{i=1}^{N_m}\Big [\max_{u_k\in\Delta_{i,m}}\big\{ f(u_k)\big\}

442: -\min_{u_k\in\Delta_{i,m}}\big\{ f(u_k)\big\}\Big ],

443: \eeq

444: where $\Delta_{i,m}=[(i-1)/N_m,i/N_m]$.

445: For large enough $M$, the finite set $\{Q(m)\colon\,m=1,\dots,M\}$

446: may be considered as an approximation to the oscillation $\eta$-summing 

447: sequence $Q_{\eta}(f)$ defined by (\ref{Q(f)}).

448: For $m=1,\dots,M$, let 

449: \beq\label{r_m}

450: r(m):=\frac{\log_2 1/N_m}{\log_2 Q(m)/N_m}

451: =\frac{\log_2 N_m}{\log_2 N_m/Q(m)}.

452: \eeq

453: Relation (\ref{index=var}) suggests that the set $\{r(m)\colon\,

454: m=1,\dots,M\}$ may be used to estimate the $p$-variation index

455: $\upsilon (f)$.

456: 

457: \begin{defn}\label{ose}

458: Let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing 

459: positive integers, let $\{u_1,\dots,u_N\}\subset [0,1]$ be such that 

460: (\ref{1ose}) holds for some integer $M$, and let $\{f(u_1),\dots,f(u_N)\}$ 

461: be a set of known values of a real-valued function $f$ on $[0,1]$.

462: We will say that $\naive (f):=r(M)$ is the {\em naive oscillation

463: $\eta$-summing estimator} of $1\vee \upsilon (f)$ based on

464: $\eta_M:=\{N_m\colon\,1\leq m\leq M\}$.

465: Letting $x_m:=\log_2 (N_m/Q(m))$ and $\bar x:=M^{-1}\sum_{m=1}^Mx_m$,

466: we will say that

467: $$

468: \ols (f):=\frac{\sum_{m=1}^M(x_m-\bar x)\log_2 N_m}{\sum_{m=1}^M

469: (x_m-\bar x)^2}

470: $$ 

471: is the {\em OLS oscillation $\eta$-summing estimator} of $1\vee\upsilon (f)$

472: based on $\eta_M:=\{N_m\colon\,1\leq m\leq M\}$.

473: The estimators $\naive$ and $\ols$ will be called the OS estimators, and

474: the estimation either by $\naive$ or by $\ols$ will be called the OS 

475: estimation.

476: \end{defn}

477: 

478: Relation (\ref{index=var}) alone, if it holds for a function $f$ and 

479: a sequence $\eta$, does not imply that the two estimators will converge 

480: to $\upsilon (f)$ as $N\to\infty$, and so as $M\to\infty$ by relation

481: (\ref{1ose}).

482: If $\upsilon (f)<\infty$, then $f$ is a regulated function on $[0,1]$,

483: that is, there exist the limits $f(t+):=\lim_{u\downarrow t}f(u)$ for each 

484: $t\in [0,1)$ and $f(s-):=\lim_{u\uparrow s}f(u)$ for each $s\in (0,1]$.

485: Assuming that $f$ is regulated and either right- or left-continuous,

486: then $\osc (f;A)$ is the same as $\osc (f;A\cap U)$ for a countable

487: and dense subset $U\subset [0,1]$ and any subset $A\subset [0,1]$.

488: For such a function $f$, one can show that the naive estimator $\naive (f)$

489: will approach $\upsilon (f)$ as the set $\{u_1,\dots,u_N\}$ will 

490: increase to $\cup_m\lambda (m)$.

491: For sample functions of a stationary Gaussian stochastic process $X$,

492: Hall and Wood (1993) showed that the two estimators corresponding to

493: the reciprocal of relation (\ref{r_m}) converge to $1/(1\vee\upsilon (X))$, 

494: and they also calculated asymptotic bias and variance.

495: 

496: \paragraph*{Oscillation $\eta$-summing index of stochastic processes.}

497: Here we show that the conditions of Theorem \ref{main} hold for almost

498: all sample functions of several important classes of stochastic processes.

499: To this aim we use known results on Hausdorff-Besicovitch

500: dimension of graphs of sample functions.

501: Let $E\subset\RR^2$ be a bounded set, and let $\diam (A)$ denote the

502: diameter of a set $A\subset\RR^2$.

503: An $\epsilon$-covering of $E$ is a countable collection $\{E_k\colon\,

504: k\geq 1\}$ of sets such that $E\subset\cup_kE_k$ and $\sup_k\diam (E_k)

505: \leq\epsilon$.

506: For $s>0$, the Hausdorff $s$-measure of $E$ is defined by

507: $$

508: \CAH^s(E):=\lim_{\epsilon\downarrow 0}\inf\Big\{\sum_{k\geq 1}

509: \big (\diam (E_k)\big )^s\colon\, \{E_k\colon\,k\geq 1\}\,\,

510: \mbox{is an $\epsilon$-covering of $E$}\,\,\Big\}.

511: $$

512: Given $E\subset\RR^2$, the function $s\mapsto\CAH^s (E)$ is nonincreasing.

513: In fact, there is a critical value $s_c$ such that $\CAH^s (E)

514: =\infty$ for $s<s_c$ and $\CAH^s (E)=0$ for $s>s_c$.

515: This critical value $s_c$ is called the 

516: {\em Hausdorff-Besicovitch dimension} and is denoted by $\dim_{HB}(E)$. 

517: That is,

518: $$

519: \dim_{HB}(E):=\inf\big\{s>0\colon\,\CAH^s (E)=0\}

520: =\sup\{s>0\colon\,\CAH^s(E)=\infty\}.

521: $$

522: A relation between the lower metric entropy index $\Delta^{-}(E)$ of $E$

523: and the Hausdorff-Besicovitch dimension of $E$ is given by the following

524: result.

525: 

526: \begin{lem}\label{HB-BC}

527: For a bounded subset $E\subset\RR^2$, $\dim_{HB}(E)\leq\Delta^{-}(E)$.

528: \end{lem}

529: 

530: \begin{proof}

531: Let $E\subset\RR^2$ be bounded, and let $s>\Delta^{-}(E)$.

532: Since 

533: $$

534: \Delta^{-}(E)=\sup\big\{\alpha >0\colon\,\lim_{\epsilon\downarrow 0}

535: N(E;\epsilon)\epsilon^{\alpha}=+\infty\big\},

536: $$

537: we have that $\lim_{\epsilon\downarrow 0}N(E;\epsilon)\epsilon^{s}=0$.

538: Thus, for each $\epsilon >0$, there exists an  $\epsilon$-covering of $E$ by 

539: $N (E;\epsilon)$ balls of equal diameter $\epsilon$.

540: Hence

541: $$

542: \inf\Big\{\sum_{k\geq 1}\big (\diam (E_k)\big )^{s}\colon\, 

543: \{E_k\colon\,k\geq 1\}\,\,\mbox{is an $\epsilon$-covering of $E$}\,\,\Big\}

544: \leq N (E;\epsilon)\epsilon^{s},

545: $$

546: which gives $\CAH^{s}(E)\leq\liminf_{\epsilon\downarrow 0}

547: N (E;\epsilon)\epsilon^{s}=0$.

548: Thus $\dim_{HB}(E)\leq s$, and so $\dim_{HB}(E)\leq\Delta^{-}(E)$,

549: proving the lemma.

550: \qed\end{proof}

551: 

552: Let $\gr (X)$ be the graph of a regulated sample function of a stochastic 

553: process $X$.

554: By Proposition \ref{3main} and by the preceding lemma, we have

555: \beq\label{4main}

556: \dim_{HB}(\gr (X))\leq\Delta^{-}(\gr (X))\leq\Delta^{+}(\gr (X))

557: \leq 2-1/(1\vee\upsilon (X)).

558: \eeq

559: In fact, the left side is equal to the right side almost surely

560: for several classes of stochastic processes.

561: For example, let $X_{\alpha}$ be a symmetric $\alpha$-stable process for

562: some $0<\alpha\leq 2$.

563: Then by Theorem B of Blumenthal and Getoor (1962), for

564: almost all sample functions of $X_{\alpha}$, we have

565: $$

566: \dim_{HB}(\gr (X_{\alpha}))=2-\frac{1}{1\vee\alpha}

567: =2-\frac{1}{1\vee\upsilon (X_{\alpha})}.

568: $$

569: For another example,

570: let $X$ be a stochastic process of Example \ref{exmp2} having

571: the Orey index $\gamma_X\in (0,1)$.

572: Then by Theorem 1 of Orey (1970), for almost all sample functions of

573: $X$, we have

574: $$

575: \dim_{HB}(\gr (X))=2-\gamma_X=2-\frac{1}{\upsilon (X)}.

576: $$

577: For these stochastic processes, relation (\ref{4main}) yields that assumption

578: (\ref{2main}) of Theorem \ref{main} is satisfied for almost every

579: sample function, and so we have the following result. 

580: 

581: \begin{cor}\label{HBdim=var}

582: Let $X=\{X(t)\colon\,0\leq t\leq 1\}$ be a stochastic process, and let

583: $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing positive 

584: integers.

585: The relation

586: $$

587: \delta_{\eta}(X)\equiv

588: \lim_{m\to\infty}\frac{\log Q(X;\lambda (m))/N_m}{\log (1/N_m)}

589: =\frac{1}{1\vee\upsilon (X)}

590: $$

591: holds for almost all sample functions of $X$ provided 

592: either $(a)$ or $(b)$ holds, where

593: \begin{enumerate}

594: \item[$(a)$] $X$ is a symmetric $\alpha$-stable process for some

595: $\alpha\in (0,2]${\rm ;}

596: \item[$(b)$] $X$ is a mean zero Gaussian stochastic process with

597: stationary increments continuous in quadratic mean and such that

598: the Orey index $\gamma_X$ exists.

599: \end{enumerate}

600: \end{cor}

601: 

602: Similar results for more general processes other than $(a)$ and $(b)$

603: of Corollary \ref{HBdim=var} can be respectively found in

604: Pruitt and Taylor (1969, Section 8) and in K\^ono (1986). 

605: 

606: \section{Simulated symmetric $\alpha$-stable process}\label{simul:sas}

607: 

608: In this section we carry out a simulation study of small-sample

609: properties of the OS estimators from Definition \ref{ose}. 

610: To this aim, we simulate a symmetric $\alpha$-stable process for

611: several values of the exponent $\alpha$, which is equal to

612: its $p$-variation index.

613: Using repeated samples we calculate the bias, the standard deviation and 

614: the mean square error for the two estimators.

615: 

616: \paragraph*{Simulating $\sas$ process.}

617: Let $X_{\alpha}=\{X_{\alpha}(t)\colon\,t\geq 0\}$ be a symmetric

618: $\alpha$-stable stochastic process with the exponent $\alpha\in (0,2]$.

619: As stated in the introduction, the $p$-variation index of $X_{\alpha}$

620: is given by $\upsilon (X_{\alpha})=\alpha$ almost surely.

621: Since the OS estimators do not capture the values

622: of the $p$-variation index below $1$, we restrict our study to

623: estimating the exponent $\alpha \in [1,2]$.

624: To simulate a sample function of a $\sas$ process $X_{\alpha}$,

625: we generate a set $\{\xi_i\colon\,i=1,\dots,n\}$ of symmetric 

626: $\alpha$-stable pseudo-random variables and use the central limit theorem 

627: to get an approximation $\widetilde{X}_{\alpha}$ of $X_{\alpha}$, where

628: $$

629: \widetilde{X}_{\alpha}(t)=\frac{1}{n^{1/\alpha}}\sum_{i=1}^{[nt]} \xi_i,

630: \qquad 0\leq t\leq 1,

631: $$

632: and $[r]$ denotes the integer part of $r$.

633: By the central limit theorem, the distribution of 

634: $\widetilde{X}_{\alpha}$ on the Skorohod space $D[0,1]$ converges

635: weakly to the distribution of $X_{\alpha}$ as $n\to\infty$.

636: We take $n=2^{14}$.

637: To generate a symmetric $\alpha$-stable random variable $\xi$, we use the

638: results of Chambers et al. (1976) (see also Section 4.6

639: in Zolotarev 1986).

640: That is,  in the sense of equality in distribution, we have

641: $$

642: \xi=\frac{\sin\alpha U}{(\cos U)^{1/\alpha}}\cdot

643: \Big (\frac{\cos (U-\alpha U)}{E}\Big )^{(1-\alpha )/\alpha}

644: \quad\mbox{if $\alpha\not =1$,}\qquad\mbox{and}\qquad \xi =\tan U\quad

645: \mbox{if $\alpha =1$,}

646: $$

647: where the random variable $U$ has the uniform distribution on 

648: $[-\pi/2,\pi/2]$, and the random variable $E$, which is independent of $U$, 

649: has the standard exponential distribution.

650: All calculations are done using the computing system {\em Mathematica}.

651: 

652: \begin{table}[tb]

653: \caption{Properties of $100$ samples of estimates $\naive (X_{\alpha})$

654: based on $\{2^m\colon\,1\leq m\leq 14\}$}

655: \label{simul1}

656: \medskip

657: \centering

658: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|}\hline

659: $\alpha$& 1.0 & 1.1 & 1.2 & 1.3 & 1.4 & 1.5 & 1.6 & 1.7 & 1.8 & 1.9 & 2.0\\ 

660: \hline\hline 

661: \rule[0mm]{0mm}{4mm} $\overline{\alpha}$ & 1.278  & 1.346 & 1.406 & 1.473

662:         & 1.556  & 1.632 & 1.715  & 1.798 & 1.883 & 1.967 & 2.051 \\ \hline

663: \rule[0mm]{0mm}{4mm} $\overline{\alpha}-\alpha$ & .2783 & .2464 & .2064 & .1727

664:         & .1563 & .1322  & .1149  & .0984 & .0832 & .0667 & .0510\\ \hline

665: \rule[0mm]{0mm}{4mm} $SD$ & .0867 & .0864 & .0542 & .0285 & .0264

666:         & .0147 & .0089 & .0063 & .0059 & .0040 & .0025 \\ \hline

667: \rule[0mm]{0mm}{4mm} $MSE$ & .0849 & .0681 & .0455 & .0306 & .0251        

668:         & .0177 & .0133 & .0097 & .0070 & .0045 & .0026 \\ \hline

669: \end{tabular}}

670: \end{table}

671: 

672: 

673: \begin{table}[tb]

674: \caption{Properties of $100$ samples of estimates $\ols (X_{\alpha})$

675: based on $\{2^m\colon\,1\leq m\leq 14\}$}

676: \label{simul2}

677: \medskip

678: \centering

679: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|}\hline

680: $\alpha$& 1.0 & 1.1 & 1.2 & 1.3 & 1.4 & 1.5 & 1.6 & 1.7 & 1.8 & 1.9 & 2.0\\ 

681: \hline\hline 

682: \rule[0mm]{0mm}{4mm} $\overline{\alpha}$ & 1.158  & 1.197 & 1.269 & 1.337

683:       & 1.383  & 1.464 & 1.529  & 1.598 & 1.654 & 1.711 & 1.767 \\ \hline

684: \rule[0mm]{0mm}{4mm} $\overline{\alpha}-\alpha$ & .1582 & .0974 & .0693 & .0375

685:       & -.0172 & -.0362  & -.0712  & -.1022 & -.1459 & -.1887 & -.2329\\ \hline

686: \rule[0mm]{0mm}{4mm} $SD$ & .0726 & .0785 & .0902 & .0804 & .0964

687:       & .0794 & .0791 & .0786 & .0691 & .0555 & .0435 \\ \hline

688: \rule[0mm]{0mm}{4mm} $MSE$ & .0302 & .0156 & .0129 & .0078 & .0095

689:       & .0076 & .0113 & .0166 & .0260 & .0387 & .0560 \\ \hline

690: \end{tabular}}

691: \end{table}

692: 

693: \paragraph*{OS estimators.}

694: Let $\eta=\{2^m\colon\,m\geq 1\}$.

695: We simulate a sample function $\widetilde{X}_{\alpha}$ at $2^{14}+1$ equally 

696: spaced points  $\{u_1,\dots,u_N\}=\{i2^{-14}\colon\,i=0,\dots,2^{14}\}$.

697: For each $1\leq m\leq 14$, let

698: $$

699: Q(m):=\sum_{i=1}^{2^m}\Big [

700: \max_{u_k\in\Delta_{i,m}}\big\{\widetilde{X}_{\alpha}(u_k)\big\}

701: -\min_{u_k\in\Delta_{i,m}}\big\{\widetilde{X}_{\alpha}(u_k)\big\}\Big ],

702: $$

703: where $\Delta_{i,m}:=[(i-1)2^{-m},i2^{-m}]$. 

704: Then define $r(m)$, $m=1,\dots,14$, by relation (\ref{r_m}) with $N_m=2^m$.

705: We use the OS estimators based on $\eta_{14}=\{2^m\colon\,1\leq m\leq 14\}$.  

706: Thus by Definition \ref{ose}, the naive oscillation $\eta$-summing estimator 

707: $\naive (X_{\alpha})=r(14)$, and

708: the OLS oscillation $\eta$-summing estimator

709: $$

710: \ols (X_{\alpha})=\frac{\sum_{m=1}^{14}(x_m-\bar x)m}{

711: \sum_{m=1}^{14}(x_m-\bar x)^2},

712: $$

713: where $x_m=\log_2(2^m/Q(m))$ and $\bar x=14^{-1}\sum_{m=1}^{14}x_m$.

714: 

715: 

716: \paragraph*{Monte-Carlo study.}

717: For $11$ different values of $\alpha\in\{1.0,1.1,\dots,2.0\}$, 

718: we simulate a vector 

719: $\{\widetilde{X}_{\alpha}(k/2^M): k=0,\dots,2^M\}$ of values

720: of a sample function of $X_{\alpha}$ with $M = 14$.

721: By the preceding paragraph, we have two estimators $\widehat{\alpha}$ of 

722: $\alpha$: the naive oscillation $\eta$-summing estimator $\naive (X_{\alpha})$,

723: and  the OLS oscillation $\eta$-summing estimator $\ols (X_{\alpha})$,

724: both based on $\{2^m\colon\,1\leq m\leq 14\}$.   

725: We repeat this procedure $K=100$ times to obtain the estimates 

726: $\widehat{\alpha}_1,\dots \widehat{\alpha}_K$ of $\alpha$ 

727: for each of the two cases. 

728: Then we calculate:

729: \begin{itemize}

730: \item The estimated expected value $\overline{\alpha}

731:       :=\big (\sum_{i=1}^K\widehat{\alpha}_i\big )/K$;

732: \item The bias $\overline{\alpha}-\alpha$;

733: \item The estimated standard deviation  $SD :=\sqrt{\sum_{i=1}^K

734:       (\widehat{\alpha}_i-\overline{\alpha})^2/(K-1)}$;

735: \item The estimated mean square error $MSE :=

736: \sum_{i=1}^K(\widehat{\alpha}_i-\alpha)^2/K$.

737: \end{itemize}

738: The estimation results are presented in Tables \ref{simul1} and \ref{simul2}.  

739: Next is a qualitative description of the performance of the two OS

740: estimators.

741: 

742: \begin{description}

743: \item[Bias] The estimation results show a different behavior of the bias

744: for the two estimators.

745: The naive estimator $\naive(X_{\alpha})$ display monotonically decreasing 

746: positive bias when $\alpha$ values increase from 1 to 2.

747: While the bias of the OLS estimator $\ols (X_{\alpha})$ monotonically 

748: decrease from a positive bias for $\alpha < 1.4$ to a negative bias 

749: for $\alpha\geq 1.4$, and the minimal absolute bias is achieved

750: when $\alpha =1.4$.

751: \end{description}

752: \noindent

753: \begin{description}

754: \item[SD] The estimated standard deviation is also different for

755: the two estimators.

756: SD values for the naive estimator $\naive (X_{\alpha})$ monotonically

757: decrease when $\alpha$ values increase from 1 to 2, and are quite

758: small when $\alpha$ is close to $2$.

759: While SD values for the OLS estimator $\ols (X_{\alpha})$ remain similar

760: and somewhat larger than for the naive estimator; only a little 

761: improvement one can notice when $\alpha$ is close to 2.

762: \end{description}

763: \noindent

764: \begin{description}

765: \item[MSE] The estimated mean square error remain different for

766: the two estimators.

767: MSE values for the naive estimator $\naive (X_{\alpha})$ decrease steady

768: when $\alpha$ values increase from 1 to 2, 

769: while MSE values for the OLS estimator $\naive (X_{\alpha})$ is smallest 

770: when $\alpha =1.5$,

771: and are increasing for all other values of $\alpha$.

772: \end{description}

773: \noindent

774: In conclusion the results show a distinction between the two OS estimators:

775: the OLS estimator $\naive (X_{\alpha})$ display better

776: performance for values $1.4\leq\alpha\leq 1.6$,

777: while the naive estimator $\naive (X_{\alpha})$

778: behaves best for $\alpha$ values close to  $2$.

779: This Monte-Carlo study was extended to sample functions based on

780: a larger number of points, 

781: i.e.\ $\{\widetilde{X}_{\alpha}(k/2^M): k=0,\dots,2^M\}$ with 

782: $M\in\{15,16,17\}$.

783: The results from the increased sample size show the same qualitative 

784: behavior as before, but with increased accuracy 

785: (see Norvai\v sa and Salopek, 2000b).

786: 

787: To the best of our knowledge, the two OS estimators provide the first

788: attempt to estimate the exponent $\alpha$ of a $\sas$ process from

789: a sample function.

790: Recently Crovella and Taqqu (1999) introduced a new method to estimate 

791: the exponent $\alpha$ of a $\sas$ random variable.

792: 

793: \section{Simulated fractional Brownian motion}\label{simul:fbm}

794: 

795: Here we perform a Monte Carlo study like in the preceding section,  

796: with a fractional Brownian motion.

797: A fractional Brownian motion $\{B_H(t)\colon\,t\geq 0\}$ with the Hurst 

798: exponent $H\in (0,1)$ is a Gaussian stochastic process with stationary 

799: increments having the covariance function

800: $$

801: E\big [B_H(t)B_H(s)\big ]=\frac{1}{2}\Big [t^{2H}+s^{2H}-|t-s|^{2H}\Big ],

802: \qquad\mbox{for $t,s\geq 0$}

803: $$

804: and $B_H(0)=0$ almost surely.

805: Since its incremental variance $\sigma_{B_H}(u)=u^H$, the Orey index

806: $\gamma_{B_H}$ exists and is equal to the Hurst exponent $H$

807: (cf. Example \ref{exmp2}). 

808: Thus by relation (\ref{var-of-Orey}), $B_H$ has the $p$-variation index 

809: $\upsilon (B_H)=1/H$ almost surely.

810: To simulate a fractional Brownian motion we use the program of

811: Maeder (1995) written in {\em Mathematica}.

812: 

813: \begin{table}[tb]

814: \caption{Properties of $100$ samples of estimates $\naive (B_{H})$

815: based on $\{2^m\colon\,1\leq m\leq 14\}$}

816: \label{simul5}

817: \medskip

818: \centering

819: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|}\hline

820: $h=1/H$& 1.2 & 1.4 & 1.6 & 1.8 & 2.0 & 2.2 & 2.4 & 2.6 & 2.8 & 3.0 \\ 

821: \hline\hline 

822: \rule[0mm]{0mm}{4mm} $\overline{h}$ & 1.132  & 1.304 & 1.474 & 1.640

823:      & 1.802  & 1.961 & 2.117  & 2.270 & 2.420 & 2.566 \\ \hline

824: \rule[0mm]{0mm}{4mm} $\overline{h}-h$ & -.0683 & -.0956 & -.1262 & -.1603

825:      & -.1979 & -.2390  & -.2826  & -.3296 & -.3804 & -.4339 \\ \hline

826: \rule[0mm]{0mm}{4mm} $SD$ & .0039 & .0018 & .0013 & .0020 & .0021

827:         & .0027 & .0026 & .0035 & .0042 & .0042 \\ \hline

828: \rule[0mm]{0mm}{4mm} $MSE$ & .0047 & .0092 & .0159 & .0257 & .0392        

829:         & .0571 & .0799 & .1086 & .1447 & .1883 \\ \hline

830: \end{tabular}}

831: \end{table}

832: 

833: 

834: 

835: 

836: \begin{table}[tb]

837: \caption{Properties of $100$ samples of estimates $\ols (B_{H})$

838: based on $\{2^m\colon\,1\leq m\leq 14\}$}

839: \label{simul6}

840: \medskip

841: \centering

842: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|}\hline

843: $h=1/H$& 1.2 & 1.4 & 1.6 & 1.8 & 2.0 & 2.2 & 2.4 & 2.6 & 2.8 & 3.0 \\ 

844: \hline\hline 

845: \rule[0mm]{0mm}{4mm} $\overline{h}$ & 1.186  & 1.340 & 1.481 & 1.614

846:     & 1.712  & 1.821 & 1.891  & 1.971 & 2.043 & 2.095 \\ \hline

847: \rule[0mm]{0mm}{4mm} $\overline{h}-h$ & -.0143 & -.0598 & -.1185 & -.1864

848:     & -.2879 & -.3795  & -.5089  & -.6294 & -.7572 & -.9053 \\ \hline

849: \rule[0mm]{0mm}{4mm} $SD$ & .0490 & .0466 & .0499 & .0469 & .0465

850:         & .0474 & .0399 & .0447 & .0494 & .0471 \\ \hline

851: \rule[0mm]{0mm}{4mm} $MSE$ & .0026 & .0057 & .0165 & .0369 & .0850        

852:         & .1462 & .2606 & .3982 & .5757 & .8217 \\ \hline

853: \end{tabular}}

854: \end{table}

855: 

856: 

857: \paragraph{OS and G estimation.}

858: In this section we apply four estimators of the $p$-variation index.

859: As before, the two OS estimators from Definition \ref{ose} will be used 

860: to estimate $h:=1/H$.

861: Moreover, we invoke the two estimators of the Orey index

862: introduced in Norvai\v sa and Salopek (2000), and which will be called

863: {\em G estimation}, which is short for the Gladyshev estimation.

864: More specifically, let $\eta=\{2^m\colon\,m\geq 1\}$.

865: We simulate a sample function $\widetilde{B}_H$ at $2^{14}+1$ equally

866: spaced points $\{u_1,\dots,u_N\}=\{i2^{-14}\colon\,i=0,\dots,2^{14}\}$.

867: For each $1\leq m\leq 14$, let

868: $$

869: Q(m):=\sum_{i=1}^{2^m}\Big [

870: \max_{u_k\in\Delta_{i,m}}\big\{\widetilde{B}_H(u_k)\big\}

871: -\min_{u_k\in\Delta_{i,m}}\big\{\widetilde{B}_H(u_k)\big\}\Big ],

872: $$

873: where $\Delta_{i,m}:=[(i-1)2^{-m},i2^{-m}]$. 

874: Then define $r(m)$, $m=1,\dots,14$, by (\ref{r_m}) with $N_m=2^m$.

875: Therefore one can use estimators based on $\eta_{14}=\{2^m\colon\,

876: 1\leq m\leq 14\}$.  

877: Thus by Definition \ref{ose}, the naive oscillation $\eta$-summing estimator 

878: $\naive (B_H)=r(14)$, and the OLS oscillation $\eta$-summing estimator 

879: $$

880: \ols (B_H)=\frac{\sum_{m=1}^{14}(x_m-\bar x)m}{

881: \sum_{m=1}^{14}(x_m-\bar x)^2},

882: $$

883: where $x_m=\log_2(2^m/Q(m))$ and $\bar x=14^{-1}\sum_{m=1}^{14}x_m$.

884: 

885: To recall the G estimation, again let $\eta=\{2^m\colon\,m\geq 1\}$, and

886: let $\widetilde{B}_H$ be a sample function given by its values at

887: $2^{14}+1$ equally spaced points $\{i2^{-14}\colon\,i=0,\dots,2^{14}\}$.

888: For each $1\leq m\leq 14$, let

889: $$

890: s_2(m):=\sum_{i=1}^{2^m}\Big [\widetilde{B}_H(i/N_m)

891: - \widetilde{B}_H((i-1)/N_m)\Big ]^2\quad\mbox{and}\quad

892: r(m):=\frac{\log 2^{-m}}{\log \sqrt{s_2(m)2^{-m}}}.

893: $$

894: The {\em naive Gladyshev estimator} of the $p$-variation index

895: is defined by $\gnaive (B_H)=r(14)$, and  the OLS Gladyshev estimator 

896: is defined by

897: $$

898: \gols (B_H)=\frac{\sum_{m=1}^{14}(x_m-\bar x)m}{

899: \sum_{m=1}^{14}(x_m-\bar x)^2},

900: $$

901: where $x_m=\log_2\sqrt{2^m/s_2(m)}$ and $\bar x=14^{-1}\sum_{m=1}^{14}x_m$.

902: 

903: 

904: 

905: \begin{table}[tb]

906: \caption{Properties of $100$ samples of estimates $\gnaive (B_{H})$

907: based on $\{2^m\colon\,1\leq m\leq 14\}$}

908: \label{simul7}

909: \medskip

910: \centering

911: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|}\hline

912: $h=1/H$& 1.2 & 1.4 & 1.6 & 1.8 & 2.0 & 2.2 & 2.4 & 2.6 & 2.8 & 3.0 \\ 

913: \hline\hline 

914: \rule[0mm]{0mm}{4mm} $\overline{h}$ & 1.164  & 1.348 & 1.530 & 1.709

915:     & 1.887  & 2.062 & 2.236  & 2.407 & 2.576 & 2.742 \\ \hline

916: \rule[0mm]{0mm}{4mm} $\overline{h}-h$ & -.0364 & -.0524 & -.0704 & -.0907

917:     & -.1133 & -.1383  & -.1645  & -.1928 & -.2245 & -.2580 \\ \hline

918: \rule[0mm]{0mm}{4mm} $SD$ & .0039 & .0020 & .0013 & .0023 & .0022

919:         & .0031 & .0030 & .0039 & .0044 & .0048 \\ \hline

920: \rule[0mm]{0mm}{4mm} $MSE$ & .0013 & .0027 & .0050 & .0082 & .0128        

921:         & .0191 & .0271 & .0372 & .0504 & .0666 \\ \hline

922: \end{tabular}}

923: \end{table}

924: 

925: 

926: 

927: 

928: \begin{table}[tb]

929: \caption{Properties of $100$ samples of estimates $\gols (B_{H})$

930: based on $\{2^m\colon\,1\leq m\leq 14\}$}

931: \label{simul8}

932: \medskip

933: \centering

934: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|}\hline

935: $h=1/H$& 1.2 & 1.4 & 1.6 & 1.8 & 2.0 & 2.2 & 2.4 & 2.6 & 2.8 & 3.0 \\ 

936: \hline\hline 

937: \rule[0mm]{0mm}{4mm} $\overline{h}$ & 1.236  & 1.429 & 1.628 & 1.818

938:      & 1.987  & 2.184 & 2.325  & 2.483 & 2.645 & 2.790 \\ \hline

939: \rule[0mm]{0mm}{4mm} $\overline{h}-h$ & .0363 & .0292 & .0281 & .0185

940:      & -.0131 & -.0161  & -.0751  & -.1171 & -.1548 & -.2100 \\ \hline

941: \rule[0mm]{0mm}{4mm} $SD$ & .0705 & .0870 & .1209 & .1241 & .1418

942:         & .1689 & .1425 & .2459 & .2650 & .2343 \\ \hline

943: \rule[0mm]{0mm}{4mm} $MSE$ & .0062 & .0084 & .0153 & .0156 & .0201        

944:         & .0285 & .0258 & .0736 & .0935 & .0985 \\ \hline

945: \end{tabular}}

946: \end{table}

947: 

948: 

949: 

950: 

951: 

952: \paragraph{Monte-Carlo study.}

953: For $10$ different values of $H\in\{0.83\approx 1.2^{-1},0.71\approx 

954: 1.4^{-1},\dots,0.33\approx 3.0^{-1}\}$,

955: we simulate the vector $\{ B_H(i2^{-m}): i=0,\dots,2^m\}$ with $m = 14$,

956: and calculate the four estimators.

957: This gives us four different estimates $\widehat{h}$ of $h=1/H$.  

958: We repeat this procedure $K=100$ times to obtain the estimates 

959: $\widehat{h}_1,\dots \widehat{h}_K$ of $h$ for each of the four cases. 

960: Then we calculate:

961: \begin{itemize}

962: \item The estimated expected value $\overline{h}

963:       :=\big (\sum_{i=1}^K\widehat{h}_i\big )/K$;

964: \item The bias $\overline{h}-h$;

965: \item The estimated standard deviation  $SD :=\sqrt{\sum_{i=1}^K

966:       (\widehat{h}_i-\overline{h})^2/(K-1)}$;

967: \item The estimated mean square error $MSE :=

968: \sum_{i=1}^K(\widehat{h}_i-h)^2/K$.

969: \end{itemize}

970: The estimation results are presented in Tables \ref{simul5} - \ref{simul8}.  

971: First one can compare the OS estimation results of $\alpha\in [1,2]$ 

972: (Tables \ref{simul1} and \ref{simul2}) for a $\sas$ process, 

973: and the OS estimation results of

974: $h=1/H\in (1,2]$ (columns $h=1.2,1.4,1.6,1.8,2.0$ of Tables 

975: \ref{simul5} and \ref{simul6}) for a fractional Brownian motion with

976: the Hurst exponent $H$.

977: The accuracy of the naive OS estimator for the two processes is similar.

978: There is only some differences in the character of monotonicity along

979: different values of parameters.  

980: The same holds for the OLS OS estimator for the two processes.

981: 

982: Now if we look at columns $h\in\{2.0,2.2,\dots,3.0\}$ of Tables \ref{simul5} 

983: and \ref{simul6}, it is clear that the OS estimates of $h$ 

984: are very poor in this case.

985: This is so since estimation errors appear in the denominator of

986: the relation (\ref{r_m}).

987: For example, if $H$ and $\widehat{H}$ both are small, then the left side of

988: the relation

989: \beq\label{difference}

990: \frac{1}{H}-\frac{1}{\widehat{H}}=(\widehat{H}-H)\frac{1}{H\widehat{H}}

991: \eeq

992: can be relatively large as compared to $\widehat{H}-H$.

993: The same remark applies to the G estimates in Tables \ref{simul7} and 

994: \ref{simul8}. 

995: To show that this is so we applied the OS estimator to evaluate the Hurst

996: exponent directly using the relations (\ref{var-of-Orey}) and 

997: (\ref{index=var}).

998: This means that in our earlier estimation formulas we need just

999: to interchange the numerator and the denominator in (\ref{r_m}).

1000: The estimation results are presented in Tables \ref{simul3} and 

1001: \ref{simul4}, where $\widehat{H}$, the bias, SD and MSE are defined as

1002: before with $h$ replaced by $H$.

1003: In Tables \ref{simul5} - \ref{simul4}, the estimation results for

1004: $H=1/2$, and so for $h=2$, are all based on the same set of $100$ simulated

1005: sample functions, which can be used to verify the effect of the above 

1006: relation \ref{difference}.

1007: Also, the results of the OS estimation of $H$ in Tables 

1008: \ref{simul3} and \ref{simul4} can be compared with the results of the

1009: G estimation of $H$ in Tables 1 and 3 of Norvai\v sa and Salopek (2000).

1010: The naive estimators corresponding to the OS and G estimations

1011: show very similar properties.

1012: As far as the OLS estimators concern, for small vaules of $H$,

1013: the bias and MSE of the OS estimation are larger than the bias 

1014: and MSE of G estimation.

1015: However, for the same values of $H$, the standard deviation of

1016: the OS estimation is smaller than the standard deviation of the

1017: G estimation. 

1018: In sum the two estimation methods OS and G show similar results

1019: when applied to estimate the Hurst exponent of a fractional

1020: Brownian motion.

1021: 

1022: 

1023: 

1024: 

1025: \begin{table}[tb]

1026: \caption{Properties of $100$ samples of estimates $\hnaive (B_{H})$

1027: based on $\{2^m\colon\,1\leq m\leq 14\}$}

1028: \label{simul3}

1029: \medskip

1030: \centering

1031: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|}\hline

1032: $H$& 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9  \\ 

1033: \hline\hline 

1034: \rule[0mm]{0mm}{4mm} $\overline{H}$ & .1578  & .2572 & .3566 & .4559

1035:      & .5549  & .6539 & .7526  & .8511 & .9496  \\ \hline

1036: \rule[0mm]{0mm}{4mm} $\overline{H}-H$ & .0578 & .0572 & .0566 & .0559

1037:      &  .0549 & .0539  & .0526  & .0511 & .0496  \\ \hline

1038: \rule[0mm]{0mm}{4mm} $SD$ & .0006 & .0007 & .0006 & .0007 & .0006

1039:         & .0007 & .0010 & .0020 & .0101 \\ \hline

1040: \rule[0mm]{0mm}{4mm} $MSE$ & .0033 & .0033 & .0032 & .0231 & .0030        

1041:         & .0029 & .0027 & .0026 & .0026  \\ \hline

1042: \end{tabular}}

1043: \end{table}

1044: 

1045: 

1046: 

1047: 

1048: \begin{table}[tb]

1049: \caption{Properties of $100$ samples of estimates $\hols (B_{H})$

1050: based on $\{2^m\colon\,1\leq m\leq 14\}$}

1051: \label{simul4}

1052: \medskip

1053: \centering

1054: \small{\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|}\hline

1055: $H$& 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9  \\ 

1056: \hline\hline 

1057: \rule[0mm]{0mm}{4mm} $\overline{H}$ & .3471  & .3932 & .4481 & .5109

1058:     & .5802  & .6508 & .7320  & .8146 & .8962 \\ \hline

1059: \rule[0mm]{0mm}{4mm} $\overline{H}-H$ & .2471 & .1932 & .1481 & .1109

1060:     & .0802 & .0508  & .0320  & .0146 & -.0038 \\ \hline

1061: \rule[0mm]{0mm}{4mm} $SD$ & .0069 & .0081 & .0113 & .0141 & .0175

1062:         & .0235 & .0276 & .0364 & .0449 \\ \hline

1063: \rule[0mm]{0mm}{4mm} $MSE$ & .0611 & .0374 & .0221 & .0125 & .0067        

1064:         & .0031 & .0018 & .0015 & .0020 \\ \hline

1065: \end{tabular}}

1066: \end{table}

1067: 

1068: 

1069: 

1070: \section{Financial data analysis}

1071: 

1072: In this section, we analyze the financial data set provided by 

1073: Olsen \& Associates.

1074: It is the high-frequency data set HFDF96, which consists of $25$ different

1075: foreign exchange spot rates, $4$ spot metal rates, and $2$ series of

1076: stock indices.

1077: This data set was recorded from $1$ Jan 1996 GMT to 31 Dec 1996 GMT.

1078: Each set has 17568 entries recorded at half hour intervals.

1079: The same financial data set was studied in Norvai\v sa and Salopek (2000) 

1080: using the two estimators of the Orey index based on the result

1081: of Gladyshev (1961).

1082: 

1083: \paragraph*{Returns.}

1084: First notice that returns in continuous time and 

1085: discrete time financial models are treated slightly differently.

1086: A return in a discrete time model is a function 

1087: $\widehat{R}$ defined on a lattice $t\in\{0,1,\dots,T\}$ with values 

1088: being a suitable transform of a pair $\{P(t-1),P(t)\}$, where $P$ is a 

1089: stock price process.

1090: A return in a continuous time model is a function $R$ defined

1091: on $[0,T]$ so that

1092: $R(t)-R(t-1)=\widehat{R}(t)$ and $R(0)=0$ for all $t\in \{1,\dots,T\}$.

1093: This gives a $1-1$ correspondence between continuous time returns

1094: used in this paper and the usual discrete time returns

1095: (see Section 2.1 in Norvai\v sa, 2000a, for further details).

1096: 

1097: Given a historical data set $\{d_0,\dots,d_K\}$ of values of a financial 

1098: asset, let $P$ be a function defined on $[0,1]$ with values 

1099: $d_k$ at $u_k:=k/K$ for $k=0,\dots,K$.

1100: Usually in econometric literature, returns are log transforms of

1101: the price.

1102: Thus in continuous time models, this corresponds to assuming that

1103: $P(t)=P(0)\exp\{X(t)\}$, $0\leq t\leq 1$, for some return process $X$

1104: which is to be analyzed.

1105: Alternatively, the price process $P$ can be a solution of an integral

1106: equation, which is not a simple exponential.

1107: If a stochastic process $X$ has the quadratic variation along the sequence

1108: of partitions $\lambda=\{\lambda (m)\colon\,m\geq 1\}$ defined by 

1109: (\ref{1ose}), and the price process $P$ is a solution of a linear integral 

1110: equation with respect to $X$ 

1111: (such as the Black-Scholes-Merton model (\ref{BSM-model}), 

1112: with $B$ replaced by $X$), then $P$ also has the quadratic variation 

1113: along the sequence $\lambda$ and

1114: the process $X$ can be recovered by the relation 

1115: $$

1116: X(t)=R_{net}(P)(t):=\lim_{m\to\infty}\sum_{i=1}^{N_m}\big [P(t\wedge

1117: i/N_m)-P(t\wedge(i-1)/N_m)\big ]/P(t\wedge (i-1)/N_m),

1118: $$

1119: for $0\leq t\leq 1$.

1120: Here the quadratic variation is understood in the sense of F\"olmer (1981),

1121: which is further developed in Norvai\v sa (2000b).

1122: The above process $X$ will be called the {\em net return} of $P$,

1123: which is analogous to discrete time simple net returns.

1124: The {\em log return} $X$ of the price process $P$ is defined by 

1125: $X(t)=R_{log}(P)(t):=\log [P(t)/P(0)]$, $0\leq t\leq 1$. 

1126: The difference between the two returns is

1127: \beq\label{dif_returns}

1128: R_{net}(P)(t)-R_{log}(P)(t)=\frac{1}{2}\int_0^t\frac{d[P]^c}{P^2}

1129: -\sum_{(0,t]}\Big [\log\frac{P}{P_{-}}-\frac{\Delta^{-}P}

1130: {P_{-}}\Big ]-\sum_{[0,t)}\Big [\log\frac{P_{+}}{P}-\frac{\Delta^{+}P}{P}

1131: \Big ]

1132: \eeq

1133: for $0\leq t\leq 1$, where $[P]^c$ is a continuous part of the quadratic 

1134: variation of $P$.

1135: Since $P$ has the quadratic variation along the sequence $\lambda$, 

1136: the difference  $R_{net}(P)-R_{log}(P)$ given by (\ref{dif_returns}) has 

1137: bounded variation. 

1138: Thus the $p$-variation indices of the two returns are equal provided

1139: both are not less than $1$.

1140: A discussion of the difference $R_{net}(P)-R_{log}(P)$ when $P$ is the

1141: geometric Brownian motion or $P$ is a model for USD/JPY exchange rates 

1142: can be found on pages 362 and 366 of Norvai\v sa and Salopek (2000).

1143: \begin{figure}

1144: \onelinecaptionsfalse

1145: \unitlength1cm

1146: \centering

1147: \begin{minipage}[b]{8.0cm}

1148:  \centering

1149:  \includegraphics[height=5.0cm, width=6.cm]{ergr1.eps}

1150: \end{minipage}%

1151: \hfill

1152: \begin{minipage}[b]{8.0cm}

1153:  \centering

1154:  \includegraphics[height=5.0cm, width=6.cm]{ergr3.eps}

1155: \end{minipage}\\ 

1156: \begin{minipage}[t]{8.0cm}

1157: \captionstyle{flushleft}

1158: \setcaptionwidth{8.cm}

1159: \caption{Naive estimators: $\naive=1.27$ and $\gnaive=1.357$ 

1160: based on $\{2^m\colon\,m=1,\dots,14\}$.}

1161: \label{usdjpy1}

1162: \end{minipage}% 

1163: \hfill

1164: \begin{minipage}[t]{8.0cm}

1165: \setcaptionwidth{8.cm}

1166: \caption{Naive estimators: $\naive=1.108$ and $\gnaive=1.025$

1167: based on $\{m\colon\,m=1,\dots,240\}$.}

1168: \label{usdjpy2}

1169: \end{minipage}%

1170: \end{figure}

1171: \begin{figure}

1172: \onelinecaptionsfalse

1173: \unitlength1cm

1174: \centering

1175: \begin{minipage}[b]{8.0cm}

1176:  \centering

1177:  \includegraphics[height=5.0cm, width=6.cm]{ergr2.eps}

1178: \end{minipage}%

1179: \hfill

1180: \begin{minipage}[b]{8.0cm}

1181:  \centering

1182:  \includegraphics[height=5.0cm, width=6.cm]{ergr4.eps}

1183: \end{minipage}\\ 

1184: \begin{minipage}[t]{8.0cm}

1185: \captionstyle{flushleft}

1186: \setcaptionwidth{8.cm}

1187: \caption{OLS estimators: $\ols=1.814$ and $\gols=2.266$

1188: based on $\{2^m\colon\,m=1,\dots,14\}$.}

1189: \label{usdjpy3}

1190: \end{minipage}% 

1191: \hfill

1192: \begin{minipage}[t]{8.0cm}

1193: \setcaptionwidth{8.cm}

1194: \caption{OLS estimators: $\ols =1.939$ and $\gols=2.049$

1195: based on $\{m\colon\,m=1,\dots,240\}$.}

1196: \label{usdjpy4}

1197: \end{minipage}%

1198: \end{figure}

1199: \paragraph*{OS and G estimation.}

1200: To estimate the $p$-variation index of a return, the oscillation

1201: $\eta$-summing estimators from Definition \ref{ose} will be used

1202: when the sequence $\eta=\{N_m\colon\,m\geq 1\}$ is given by

1203: $N_m=2^m$ and $N_m=m$ for integers $m\geq 1$.

1204: A comparison will be made with the results of Norvai\v sa and Salopek (2000)

1205: by using their estimators based on the result of Gladyshev (1961).

1206: As before, the oscillation $\eta$-summing estimation is called the

1207: {\em OS estimation}, and the estimation as in Norvai\v sa and 

1208: Salopek (2000) is called the {\em G estimation}.  

1209: 

1210: 

1211: Given a price process $P$ at points $u_k=k/K$, $k=0,\dots,K$,

1212: find the maximal integer $M$ such that

1213: \beq\label{1est}

1214: \cup_{m=1}^M\lambda (m)\subset\{u_0,\dots,u_K\},\qquad

1215: \mbox{where $\lambda (m)=\{i/N_m\colon\,i=0,\dots,N_m\}$.}

1216: \eeq

1217: Notice that in the case $N_m=m$, the sequence $\{\lambda (m)\colon\,

1218: m\geq 1\}$ is {\em not} nested.

1219: Since for HFDF96 data set $K=17568$, $M=14$ when $N_m=2^m$ and 

1220: $M=240$ when $N_m=m$.

1221: 

1222: 

1223: Let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing 

1224: positive integers, and let $M$ be such that relation (\ref{1est}) holds.

1225: In the case $X=R_{net}(P)$ or $X=R_{log}(P)$,

1226: for each $1\leq m\leq M$, let

1227: $$

1228: Q(m):=Q(X;\lambda (m))=\sum_{i=1}^{N_m}\Big [

1229: \max_{u_k\in\Delta_{i,m}}\big\{X(u_k)\big\}

1230: -\min_{u_k\in\Delta_{i,m}}\big\{X(u_k)\big\}\Big ],

1231: $$

1232: where $\Delta_{i,m}:=[(i-1)/N_m,i/N_m]$. Also for each $1\leq m\leq M$, let

1233: \beq\label{2est}

1234: q(m):=\log_2(N_m/Q(m))\quad\mbox{and}\quad

1235: r(m):=\frac{\log (1/N_m)}{\log Q(m)/N_m}

1236: =\frac{\log_2 N_m}{q(m)}.

1237: \eeq

1238: By Definition \ref{ose}, the naive oscillation $\eta$-summing estimator 

1239: $\naive:=\naive (X)=r(M)$.

1240: The OLS oscillation $\eta$-summing estimator $\ols$ is defined by

1241: \beq\label{3est}

1242: \ols:=\ols (X)=\frac{\sum_{m=1}^M(x_m-\bar x)

1243: \log_2 N_m}{\sum_{m=1}^M(x_m-\bar x)^2},

1244: \eeq

1245: where $x_m=q(m)=\log_2(N_m/Q(m))$ and $\bar x=M^{-1}\sum_{m=1}^Mx_m$.

1246: 

1247: 

1248: Turning to the G estimation, again let $\eta=\{N_m\colon\,m\geq 1\}$ 

1249: be a sequence of strictly increasing positive integers, and let $M$

1250: be such that relation (\ref{1est}) holds.

1251: In the case $X=R_{net}(P)$ or $X=R_{log}(P)$,

1252: for each $1\leq m\leq M$, let

1253: $$

1254: s_2(m):=\sum_{i=1}^{N_m}\Big [X(i/N_m)- X((i-1)/N_m)\Big ]^2.

1255: $$

1256: In relations (\ref{2est}) and (\ref{3est}), for each $1\leq m\leq M$, 

1257: replacing $N_m/Q(m)$ by $\sqrt{N_m/s_2(m)}$, let

1258: $$

1259: q(m):=\log_2\sqrt{N_m/s_2(m)}\quad\mbox{and}\quad

1260: r(m):=\frac{\log (1/N_m)}{\log \sqrt{s_2(m)/N_m}}

1261: =\frac{\log_2 N_m}{q(m)}.

1262: $$

1263: Then the {\em naive Gladyshev estimator} of the $p$-variation index

1264: is $\gnaive :=\gnaive (X)=r(M)$.

1265: The OLS Gladyshev estimator $\gols$ is defined by

1266: $$

1267: \gols:=\gols (X)=\frac{\sum_{m=1}^M(x_m-\bar x)

1268: \log_2 N_m}{\sum_{m=1}^M(x_m-\bar x)^2},

1269: $$

1270: where $x_m=q(m)=\log_2\sqrt{N_m/s_2(m)}$ and $\bar x=M^{-1}\sum_{m=1}^Mx_m$.

1271: 

1272: 

1273: 

1274: \begin{table}[tb]

1275: \noindent

1276: \caption{\label{summary} OS and G estimation of returns of exchanges rates

1277: based on $\{2^m\colon\,1\leq m\leq 14\}$.} 

1278: \medskip

1279: \centering

1280: \begin{tabular}{|l|c|c|c|c|c|c|c|c|}\hline

1281: Currency &\multicolumn{4}{c|}{net-returns}

1282:          &\multicolumn{4}{c|}{log-returns}\\ \cline{2-9}

1283:     &\rule[0mm]{0mm}{5mm}$\naive$&$\ols$&$\gnaive$&$\gols$&$\naive$&$\ols$&

1284: $\gnaive$&$\gols$ \\  [0.5ex] \hline\hline

1285: AUD/USD & 1.2452 & 1.7617 & 1.3563 & 2.2436 

1286:         & 1.2452 & 1.7664 & 1.3563 & 2.2663\\ \hline

1287: CAD/USD & 1.1493 & 1.8362 & 1.2362 & 2.7382 

1288:         & 1.1493 & 1.8357 & 1.2362 & 2.7422 \\ \hline \hline

1289: DEM/ESP & 1.1504 & 1.7197 & 1.4043 & 2.5684

1290:         & 1.1504 & 1.7274 & 1.4042 & 3.4243 \\ \hline 

1291: DEM/FIM & 1.2124 & 1.8537 & 1.3566 & 2.4970

1292:         & 1.2124 & 1.8571 & 1.3566 & 2.5047 \\ \hline 

1293: DEM/ITL & 1.2420 & 1.7619 & 1.3673 & 2.0673

1294:         & 1.2420 & 1.7552 & 1.3673 & 2.0496 \\ \hline 

1295: DEM/JPY & 1.2671 & 1.8045 & 1.3661 & 2.4714

1296:         & 1.2671 & 1.8034 & 1.3662 & 2.4770 \\ \hline 

1297: DEM/SEK & 1.2483 & 1.7601 & 1.3767 & 2.2612

1298:         & 1.2483 & 1.7558 & 1.3767 & 2.2359 \\ \hline \hline

1299: GBP/DEM & 1.2325 & 1.6459 & 1.3380 & 1.9566

1300:         & 1.2325 & 1.6493 & 1.3380 & 1.9678 \\ \hline 

1301: GBP/USD & 1.2398 & 1.7708 & 1.3334 & 2.1373

1302:         & 1.2398 & 1.7725 & 1.3334 & 2.1466 \\ \hline \hline

1303: USD/BEF & 1.2615 & 1.7083 & 1.4745 & 2.4220

1304:         & 1.2615 & 1.7176 & 1.4745 & 2.5022 \\ \hline 

1305: USD/CHF & 1.2796 & 1.7000 & 1.3902 & 2.0817

1306:         & 1.2796 & 1.7044 & 1.3902 & 2.1001 \\ \hline 

1307: USD/DEM & 1.2432 & 1.7646 & 1.3416 & 2.2111

1308:         & 1.2432 & 1.7673 & 1.3416 & 2.2294 \\ \hline

1309: USD/DKK & 1.2896 & 1.5574 & 1.7142 & 2.2990

1310:         & 1.2894 & 1.6096 & 1.7102 & 2.6473 \\ \hline 

1311: USD/ESP & 1.3574 & 1.9885 & 1.5331 & 2.5367

1312:         & 1.3574 & 2.0155 & 1.5331 & 2.7526 \\ \hline 

1313: USD/FIM & 1.3208 & 1.8583 & 1.4520 & 2.2860

1314:         & 1.3208 & 1.8682 & 1.4520 & 2.3226 \\ \hline 

1315: USD/FRF & 1.2467 & 1.7718 & 1.3643 & 2.2542

1316:         & 1.2467 & 1.7766 & 1.3643 & 2.2811 \\ \hline 

1317: USD/ITL & 1.2977 & 2.1020 & 1.4115 & 3.0934

1318:         & 1.2977 & 2.0863 & 1.4115 & 2.9658 \\ \hline 

1319: USD/NLG & 1.2591 & 1.7570 & 1.3815 & 2.2647

1320:         & 1.2591 & 1.7622 & 1.3816 & 2.2937\\ \hline 

1321: USD/SEK & 1.3263 & 1.9542 & 1.4453 & 2.5304

1322:         & 1.3263 & 1.9606 & 1.4453 & 2.5575 \\ \hline 

1323: USD/XEU & 1.2450 & 1.8503 & 1.3594 & 2.5097

1324:         & 1.2450 & 1.8539 & 1.3594 & 2.5423 \\ \hline \hline

1325: USD/JPY & 1.2703 & 1.8104 & 1.3573 & 2.2434

1326:         & 1.2703 & 1.8143 & 1.3573 & 2.2662 \\ \hline 

1327: USD/MYR & 1.1121 & 1.6715 & 1.3957 & 2.9137

1328:         & 1.1121 & 1.6905 & 1.3956 & 3.0156 \\ \hline 

1329: USD/SGD & 1.1377 & 1.8821 & 1.2584 & 3.2629 

1330:         & 1.1377 & 1.8794 & 1.2584 & 3.2128 \\ \hline 

1331: USD/ZAR & 1.2377 & 1.5460 & 1.4400 & 1.9238

1332:         & 1.2377 & 1.5516 & 1.4401 & 1.9449\\ \hline 

1333: \end{tabular}

1334: \end{table}

1335: 

1336: 

1337: %\clearpage

1338: 

1339: 

1340: 

1341: 

1342: \begin{table}[tb]

1343: \noindent

1344: \caption{\label{metalsum}  OS and G estimation of returns of metal rates

1345: based on $\{2^m\colon\,1\leq m\leq 14\}$.}

1346: \medskip

1347: \centering

1348: \begin{tabular}{|l|c|c|c|c|c|c|c|c|}\hline

1349: Metal   &\multicolumn{4}{c|}{net-returns}

1350:         &\multicolumn{4}{c|}{log-returns}\\ \cline{2-9}

1351:    &\rule[0mm]{0mm}{5mm} $\naive$&$\ols$&$\gnaive$&$\gols$

1352:    &$\naive$&$\ols$&$\gnaive$&$\gols$ \\  [0.5ex] \hline\hline

1353: Gold    & 1.2190 & 1.7249 & 1.3488 & 2.2808 

1354:         & 1.2190 & 1.7216 & 1.3489 & 2.2700 \\ \hline 

1355: Silver  & 1.4097 & 1.8482 & 1.6227 & 2.6066 

1356:         & 1.4097 & 1.8263 & 1.6229 & 2.5728 \\ \hline \hline

1357: Palladium & 1.3371 & 1.6650 & 1.5481 & 2.3535

1358:           & 1.3371 & 1.6603 & 1.5480 & 2.3390 \\ \hline 

1359: Platinum  & 1.2358 & 1.6472 & 1.4134 & 2.2933

1360:           & 1.2358 & 1.6433 & 1.4135 & 2.2790 \\ \hline 

1361: \end{tabular}

1362: \end{table}

1363: 

1364: \begin{table}[tb]

1365: \noindent

1366: \caption{\label{indexsum}  OS and G estimation of returns of stock indices

1367: based on $\{2^m\colon\,1\leq m\leq 14\}$.}

1368: \medskip

1369: \centering

1370: \begin{tabular}{|l|c|c|c|c|c|c|c|c|}\hline

1371: Index &\multicolumn{4}{c|}{net-returns}

1372:       &\multicolumn{4}{c|}{log-returns}\\ \cline{2-9}

1373:   &\rule[0mm]{0mm}{5mm} $\naive$&$\ols$&$\gnaive$&$\gols$ 

1374:   &$\naive$&$\ols$&$\gnaive$&$\gols$\\  [0.5ex] \hline\hline

1375: SP 500     & 1.1685 & 1.3723 & 1.3726 & 1.9085  

1376:            & 1.1685 & 1.3734 & 1.3727 & 1.9162 \\ \hline

1377: DOW JONES  & 1.1806 & 1.3909 & 1.3937 & 1.9035  

1378:            & 1.1806 & 1.3928 & 1.3938 & 1.9133 \\ \hline 

1379: \end{tabular}

1380: \end{table}

1381: 

1382: 

1383: \paragraph*{Estimation results.}

1384: Estimation results for the data set HFDF96 are given by Tables \ref{summary},

1385: \ref{metalsum} and \ref{indexsum}.

1386: More specifically, the tables contain the estimated $p$-variation indices

1387: for the two returns of the bid price associated with the nearest

1388: prior datum.

1389: We picked the log-returns of USD/JPY exchange rates to illustrate 

1390: by Figures \ref{usdjpy1} - \ref{usdjpy4} a difference between the OS and

1391: G estimation results in more detail. 

1392: Figures \ref{usdjpy2} and \ref{usdjpy4} show the estimation results based

1393: on a sequence $\eta=\{N_m=m\colon\,m\geq 1\}$ truncated at $M=240$.

1394: The associated sequence of partitions $\{\lambda (m)\colon\,m\geq 1\}$ in 

1395: this case is not nested, and technical calculations in this case are more 

1396: complex.

1397: Figures \ref{usdjpy1} and \ref{usdjpy3} show the estimation results

1398: based on dyadic partitions, which are used for the rest of results.

1399: Because the conditions of Theorem \ref{main} are more general

1400: as compared to the conditions of the main result of Gladyshev (1961),

1401: the results of the OS estimation are more reliable than the results

1402: of G estimation.

1403: This is also seen from the Figures \ref{usdjpy2} and \ref{usdjpy4}.

1404: The columns $\naive$ and $\ols$ of Tables \ref{summary}, \ref{metalsum} 

1405: and \ref{indexsum} suggest that estimated $p$-variation indices of the 

1406: returns of the financial data are more likely to belong to the interval 

1407: $(1,2)$. 

1408: These columns are essentially the same for net-returns as well as

1409: for log-returns, which would be in agreement if a stock price has the 

1410: quadratic variation along a nested sequence of dyadic partitions 

1411: and its $p$-variation index is not less than 1.

1412: Due to the pattern exhibit by the results of Section \ref{simul:fbm}

1413: when estimating $1/H$, it is unlikely that a fractional Brownian

1414: motion $B_H$ would give a satisfactory fit to the financial data for some 

1415: values of $H$.

1416: However, as seen in Figures \ref{usdjpy1} - \ref{usdjpy4}, a more general 

1417: process from Example \ref{exmp2} might be used to model the HFDF96 data sets.

1418: Especially this concerns a modelling of stock indices 

1419: (see Figure \ref{indexsum}).

1420: Recall that the G estimation is reliable  when applied to sample functions 

1421: of stochastic processes having a suitable relationship between a sample 

1422: function behavior and an asymptotic behavior of the incremental variance.

1423: The estimation results do not reject the hypothesis that returns 

1424: may be modeled by a L\'evy process. 

1425: To be more specific about a degree of data fitting, we need a theoretical

1426: asymptotic analysis of both, the OS and G estimations, which is not

1427: available at this writing. 

1428: 

1429: 

1430: \paragraph*{Related results and techniques.}

1431: The OS and G estimators are based on properties of a function similar

1432: to a kind of self-similarity property with respect to shrinking 

1433: time intervals, sometimes refered to as a fractal or empirical scaling law.

1434: It is natural that a high-frequency data have already been used to

1435: detect such laws if exist.   

1436: The work of M\"uller et.\ al.\ (1995) addresses this question and provide

1437: some preliminary findings in their analysis of a high-frequency FX data

1438: collected from raw data vendors such as Reuters, Knight-Rider and Telerate. 

1439: Another related work of Mandelbrot (1997) have already been discussed

1440: in Section 4.3 of Norvai\v sa and Salopek (2000).

1441: 

1442: \section{Discussion}

1443: 

1444: Parameter estimation of a financial model is a typical econometric 

1445: analysis task.

1446: What is atypical in the preceding analysis is a generality of

1447: the underlying financial model.

1448: This model applies far beyond of limits imposed by the semimartingale theory, 

1449: and its outline can be found in Bick and Willinger (1994), and 

1450: Norvai\v sa (2000a).

1451: The $p$-variation index considered as a parameter of a model is defined

1452: for any function.

1453: Its estimate provides a grade for each concrete continuous time

1454: model of a price process $P$ governed by an exponential, or by a linear 

1455: integral equation having the indefinite integral

1456: \beq\label{prod-int}

1457: P(t)=\lim_{\kappa}\prod_{i=1}^n\big (1+X(t_i\wedge t)-X(t_{i-1}\wedge t)),

1458: \qquad 0<t\leq 1,

1459: \eeq

1460: as its unique solution, where the limit is understood either in the sense of

1461: refinements of partitions $\kappa=\{t_i\colon\,i=0,\dots,n\}$ of $[0,1]$,

1462: or in the more general sense along a fixed sequence of nested partitions.

1463: The linear It\^o stochastic integral equation with respect to

1464: a semimartingale $X$ is one such example of a continuous time model.

1465: A further generality of the underlying financial model could be achieved 

1466: once the net-returns are modified so as to reverse a solution of a 

1467: non-linear integral equation.

1468: 

1469: We stress the importance of the notion of a return because it

1470: provides a two direction link between theory and practice.  

1471: A financial model without the notion of a return is just an exercise

1472: in theory building.

1473: As we noted earlier, the results of the preceding section

1474: show that the estimated $p$-variation indices for the calculated 

1475: net-returns and log-returns are almost the same.

1476: Here we discuss the difference between the net- and log-returns

1477: $R_{net}(P)-R_{log}(P)$ given by (\ref{dif_returns}) for a simulated 

1478: price process $P$.

1479: Suppose that $P$ is the Dolean exponential ${\cal E}(X_{\alpha})$

1480: of a symmetric $\alpha$-stable process $X_{\alpha}$:

1481: \beq\label{dolean}

1482: {\cal E}(X_{\alpha})(t)

1483: :=\exp\big\{X_{\alpha}(t)-X_{\alpha}(0)\big\}\prod_{(0,t]}

1484: (1+\Delta^{-}X_{\alpha})e^{-\Delta^{-}X_{\alpha}},\qquad

1485: 0<t\leq 1.

1486: \eeq   

1487: That is, $P$ is defined by (\ref{prod-int}) with $X$ replaced by $X_{\alpha}$. 

1488: Then the continuous part of the quadratic variation $[P]^c\equiv 0$, and

1489: so the simulation gives the remaining sums in (\ref{dif_returns})

1490: as shown in Figure \ref{stable_returns}.

1491: 

1492: \begin{figure}

1493: \onelinecaptionsfalse

1494: \unitlength1cm

1495: \centering

1496: \begin{minipage}[b]{7.0cm}

1497:  \centering

1498:  \includegraphics[height=6.7cm, width=6.7cm]{dol5.eps}

1499: \end{minipage}%

1500: \hfill

1501: \begin{minipage}[b]{7.0cm}

1502:  \centering

1503:  \includegraphics[height=6.7cm, width=6.7cm]{dol6.eps}

1504: \end{minipage}\\ 

1505: %\begin{minipage}[t]{7.0cm}

1506: %\captionstyle{flushleft}

1507: %\setcaptionwidth{7.0cm}

1508: \caption{Four trajectories of $R_{net}(P)-R_{log}(P)$ (left) defined by

1509: (\ref{dif_returns}) corresponding to the Dolean exponential 

1510: $P={\cal E}(X_{\alpha})$ (right) given by (\ref{dolean}), where

1511: $X_{\alpha}$ is a S$\alpha$S process with $\alpha=1.7$.

1512: Vertical lines join jumps.}

1513: \label{stable_returns}

1514: %\end{minipage}% 

1515: %\hfill

1516: %\begin{minipage}[t]{7.0cm}

1517: %\setcaptionwidth{7.0cm}

1518: %\caption{OLS estimators: $\ols=1.814$ and $\gols=2.266$

1519: %where $\eta=\{N_m=2^m\colon\,m=1,\dots,14\}$.}

1520: %\label{usdjpy4}

1521: %\end{minipage}%

1522: \end{figure}

1523: 

1524: 

1525: To simulate the Dolean exponential ${\cal E}(X_{\alpha})$ we use its 

1526: representation as the indefinite product integral (\ref{prod-int}), 

1527: proved by Dudley and Norvai\v sa (1999, Corollary 5.23 in Part II). 

1528: Theoretically, the Dolean exponential composed with the net-return

1529: gives $X_{\alpha}=R_{net}(P)$.

1530: Thus the maximal discrepancy $d:=\sup_t| 

1531: \widetilde{X}_{\alpha}(t)-\widetilde{R}_{net}(\widetilde{P})(t)|$ between 

1532: simulated versions of the two sides should be small if simulation is 

1533: accurate enough.

1534: Indeed, we get that the discrepency $d=3.1\times 10^{-14}$ holds

1535: uniformly for all trajectories in the left Figure \ref{stable_returns}. 

1536: Notice that one of the four trajectories breaks because its Dolean

1537: exponential jumps to the negative side, and so its logarithm is

1538: undefined.

1539: 

1540: 

1541: \section*{Appendix A}

1542: 

1543: This section contains proofs of Theorem \ref{main} 

1544: and Proposition \ref{3main}.

1545: The proofs consist of combining a few relations

1546: between the metric entropy index, the oscillation $\eta$-summing

1547: index and the $p$-variation index, and will be discussed first.

1548: 

1549: As noted earlier, in the actual calculations of the metric entropy 

1550: it is often simpler to replace closed balls by squares (boxes), 

1551: which leads to the box-counting dimension (see Section 6.1 in 

1552: Cutler, 1993).

1553: First we modify the box-counting method for sets which arise when

1554: connecting a graph of a discontinuous function.

1555: Throughout this section, $f$ is a regulated function defined on a

1556: closed interval $[0,T]$.

1557: This means that for each $t\in (0,T]$, there exists a limit 

1558: $f(t-):=\lim_{u\uparrow t}f(u)$, and for each $t\in [0,T)$, there exists 

1559: a limit $f(t+):=\lim_{u\downarrow t}f(u)$.

1560: For $t\in (0,T]$, let $V_{-}(f(t))$ be the interval connecting the points 

1561: $(t,f(t-))$ and $(t,f(t))$, and for $t\in [0,T)$, let $V_{+}(f(t))$ be the 

1562: interval connecting the points $(t,f(t))$ and $(t,f(t+))$.

1563: On a plane $\RR^2$ with the Cartesian coordinates $(t,x)$, for each

1564: $0\leq u <v\leq T$, let

1565: \beq\label{G_f}

1566: G_f([u,v]):=(u,V_{+}(f(u)))\bigcup\Big (\bigcup\limits_{u< t <v}(t,V_{+}(f(t))

1567: )\bigcup\, (t,V_{-}(f(t)))\Big )\bigcup\, (v,V_{-}(f(v)))\subset\RR^2,

1568: \eeq

1569: where $(t,A(t)):=\{(t,x)\colon\,x\in A(t)\}$.

1570: Let $G_f:=G_f([0,T])$, and let $\gr (f)$ be the graph of the function $f$.

1571: It is clear that $\gr (f)\subset G_f$, and the equality holds between the

1572: two sets if and only if $f$ is continuous on $[0,T]$.

1573: For $\epsilon >0$, a {\em grid} $\CAC_{\epsilon}$ of side length $\epsilon$ 

1574: covering a bounded set $E$, is called a collection of disjoint squares

1575: of equal side length $\epsilon$ whose union is a square containing $E$.

1576: We can also assume that each grid $\CAC_{\epsilon}$ intersects the

1577: origin of the Cartesian coordinates, and its squares have sides parallel

1578: to the coordinates axes.

1579: The usual box-counting method counts all squares of the grid $\CAC_{\epsilon}$

1580: which have nonempty intersection with $E$.

1581: We need a special counting rule when a vertical segment of the set

1582: $E=G_f$ is a part of a vertical line of a grid $\CAC_{\epsilon}$.

1583: Let $\{t_i\colon\,i=0,\dots,N_{\epsilon}\}$ be the partition of $[0,T]$ 

1584: induced by intersecting a grid $\CAC_{\epsilon}$ with the $t$-axes, 

1585: and for each $i=1,\dots,N_{\epsilon}$, let $n_i(\epsilon)$ be the number 

1586: of squares of the grid $\CAC_{\epsilon}$ contained in the strip 

1587: $[t_{i-1},t_i]\times\RR$ which have a nonempty intersection with 

1588: $G_f([t_{i-1},t_i])$.

1589: For each $\epsilon >0$, we then define 

1590: $$

1591: M(G_f;\epsilon):=\sum_{i=1}^{N_{\epsilon}}n_i(\epsilon).

1592: $$

1593: This definition avoids a double covering of certain parts of vertical 

1594: segments, and it reduces to the usual definition if $f$ is continuous.

1595: The present extension allows us to apply Theorem \ref{main} to sample

1596: functions of L\'evy processes, which may have left-side discontinuities.

1597: 

1598: Next statement shows the usefulness of the box-counting method.

1599: 

1600: \begin{lem}\label{boxes}

1601: Let $f$ be a regulated function on $[0,T]$. 

1602: For lower and upper metric entropy indices, we have

1603: $$

1604: \Delta^{-}(G_f)=\liminf_{\epsilon\downarrow 0}\frac{\log M(G_f;\epsilon)}

1605: {\log (1/\epsilon)}\quad\mbox{and}\quad

1606: \Delta^{+}(G_f)=\limsup_{\epsilon\downarrow 0}\frac{\log M(G_f;\epsilon)}

1607: {\log (1/\epsilon)}.

1608: $$

1609: \end{lem}

1610: 

1611: \begin{proof}

1612: For $\epsilon >0$, let $\CAC_{\epsilon}$ be a grid of side length

1613: $\epsilon$ covering $G_f$.

1614: Every square of side $\epsilon$ is included in the ball of diameter

1615: $\epsilon\sqrt{2}$, which, in turn, is included in at most $9$

1616: squares of the grid $\CAC_{\epsilon}$.

1617: Thus

1618: $$

1619: N(G_f;\epsilon\sqrt{2})\leq M(G_f;\epsilon)\leq 9N(G_f;\epsilon\sqrt{2}),

1620: $$

1621: proving the lemma.

1622: \qed\end{proof}

1623: 

1624: Next we look at replacing a limit as $\epsilon\downarrow 0$ by a limit

1625: along a countable sequence.

1626: Let $\{\epsilon_m\colon\,m\geq 1\}$ be a sequence of positive real 

1627: numbers strictly decreasing to $0$, and for each $m\geq 1$, 

1628: let $\Delta_m:=(\epsilon_{m+1},\epsilon_m]$.

1629: For a family $\{A_{\epsilon}\colon\,\epsilon >0\}$ of numbers,  we have

1630: $$

1631: \limsup_{\epsilon\downarrow 0}A_{\epsilon}=\limsup_{m\to\infty}

1632: \sup_{\epsilon\in\Delta_m}A_{\epsilon}\geq\limsup_{m\to\infty}

1633: A_{\epsilon_m}

1634: $$

1635: and

1636: $$

1637: \liminf_{\epsilon\downarrow 0}A_{\epsilon}=\liminf_{m\to\infty}

1638: \inf_{\epsilon\in\Delta_m}A_{\epsilon}\leq\liminf_{m\to\infty}

1639: A_{\epsilon_m}.

1640: $$

1641: Thus by the preceding lemma, we have that

1642: \beq\label{2subsequences}

1643: \Delta^{-}(G_f)\leq\liminf_{m\to\infty}\frac{\log M(G_f;\epsilon_m)}

1644: {\log (1/\epsilon_m)}\quad\mbox{and}\quad

1645: \limsup_{m\to\infty}\frac{\log M(G_f;\epsilon_m)}{\log (1/\epsilon_m)}

1646: \leq \Delta^{+}(G_f).

1647: \eeq

1648: 

1649: 

1650: \begin{lem}\label{subsequences}

1651: Let $f$ be a regulated function on $[0,T]$, and let $\{\epsilon_m

1652: \colon\,m\geq 1\}$ be a sequence of positive real numbers strictly

1653: decreasing to $0$ so that

1654: \beq\label{1subsequences}

1655: \lim_{m\to\infty}\frac{\log \epsilon_m}{\log\epsilon_{m+1}}=1.

1656: \eeq

1657: Then

1658: \beq\label{3subsequences}

1659: \left\{ \begin{array}{l}

1660: \liminf_{\epsilon\downarrow 0}\\

1661: \limsup_{\epsilon\uparrow 0}

1662: \end{array} \right\}

1663: \frac{\log M (G_f;\epsilon)}{\log (1/\epsilon)}

1664: =\left\{ \begin{array}{l}

1665: \liminf_{m\to\infty}\\

1666: \limsup_{m\to\infty}

1667: \end{array} \right\}

1668: \frac{\log M (G_f;\epsilon_m)}{\log (1/\epsilon_m)}.

1669: \eeq

1670: \end{lem}

1671: 

1672: \begin{proof}

1673: It is enough to prove the reverse inequalities in relations

1674: (\ref{2subsequences}). 

1675: For $0<\epsilon\leq\epsilon_1$, let $m\geq 1$ be such that 

1676: $\epsilon\in\Delta_m=(\epsilon_{m+1},\epsilon_m]$.

1677: Since a square with a side length $\epsilon$ is contained in at most

1678: four squares with a side length $\epsilon_m$, we have

1679: $ M(G_f;\epsilon_m)\leq 4 M (G_f;\epsilon)$.

1680: Similarly, $M(G_f;\epsilon)\leq 4 M (G_f;\epsilon_{m+1})$.

1681: Thus for each $\epsilon\in\Delta_m$,

1682: $$

1683: \frac{\log \epsilon_m}{\log\epsilon_{m+1}}\Big (\frac{\log M(G_f;\epsilon_m)

1684: -\log 4}{\log 1/\epsilon_m}\Big )\leq 

1685: \frac{\log M(G_f;\epsilon)}{\log (1/\epsilon )}

1686: \leq\frac{\log \epsilon_{m+1}}{\log\epsilon_{m}}\Big (\frac{\log M

1687: (G_f;\epsilon_{m+1})+\log 4}{\log 1/\epsilon_{m+1}}\Big ).

1688: $$

1689: The conclusion now follows by the assumption (\ref{1subsequences}).

1690: \qed\end{proof}

1691: 

1692: Now recall relation (\ref{Q(f)}) defining the oscillation $\eta$-summing 

1693: sequence $\{Q(f;\lambda (m))\colon\,m\geq 1\}$.

1694: 

1695: \begin{lem}\label{CCM}

1696: Let $f$ be a regulated non-constant function on $[0,T]$.

1697: For any sequence $\{N_m\colon\,m\geq 1\}$ of strictly increasing positive 

1698: integers, we have

1699: \beq\label{1CCM}

1700: \left\{\begin{array}{l}

1701: \liminf_{m\to\infty} \\ \limsup_{m\to\infty}

1702: \end{array}\right\}

1703: \frac{\log M (G_f;T/N_m)}{\log N_m} 

1704: =\left\{\begin{array}{l}

1705: \liminf_{m\to\infty} \\ \limsup_{m\to\infty}

1706: \end{array}\right\}

1707: \frac{\log N_m Q (f;\lambda (m))}{\log N_m}.

1708: \eeq

1709: \end{lem}

1710: 

1711: \begin{proof}

1712: Let $\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing positive 

1713: integers.

1714: For each $m\geq 1$, let $A_m$ and $B_m$ be the terms under the limit

1715: signs in equality (\ref{1CCM}) in the given order.

1716: It is enough to prove that

1717: \beq\label{3CCM}

1718: \liminf_mA_m\leq \liminf_mB_m,\qquad

1719: \limsup_m A_m\leq\limsup_mB_m

1720: \eeq

1721: and 

1722: \beq\label{4CCM}

1723: \liminf_mB_m\leq \liminf_mA_m,\qquad

1724: \limsup_m B_m\leq\limsup_mA_m.

1725: \eeq

1726: Suppose that the set $G_f$ is on a plane with the Cartesian coordinates 

1727: $(t,x)$.

1728: For each integer $m\geq 1$, let $\CAC_m$ be the grid of side length $T/N_m$ 

1729: intersecting the origin and with sides parallel to the coordinate axes.

1730: Thus $\CAC_m$ intersects with the $t$-axes at each point of the partition

1731: $\lambda (m)=\{t_i^m=iT/N_m\colon\,i=0,\dots,N_m\}$.

1732: Let $m\geq 1$.

1733: For $i=1,\dots,N_m$, let $n_i=n_i(m)$ be the number of squares of the grid

1734: $\CAC_m$ contained in the strip $\Delta_{i,m}\times\RR$ and

1735: covering the set $G_f(\Delta_{i,m})$ defined by relation (\ref{G_f}),

1736: where $\Delta_{i,m}=[t_{i-1}^m,t_i^m]$.

1737: By the definition of $G_f(\Delta_{i,m})$, 

1738: $$

1739: (N_m/T)\osc (f;\Delta_{i,m})\leq n_i\leq 2+(N_m/T)\osc (f;\Delta_{i,m}),

1740: $$ 

1741: for each $i=1,\dots,N_m$.

1742: Summing over all indices $i$, it follows that the inequalities

1743: \beq\label{2CCM}

1744: M(G_f;T/N_m)-2N_m\leq (N_m/T)Q(f;\lambda (m))\leq M(G_f;T/N_m)

1745: \eeq

1746: hold for each $m\geq 1$.

1747: By the second inequality in display (\ref{2CCM}), we have that 

1748: inequalities (\ref{4CCM}). 

1749: To prove inequalities (\ref{3CCM}), first suppose that $\liminf_mB_m=1$.

1750: Then by relation (\ref{Qin[1,2]}), $\limsup_mB_m=1$, and inequalities 

1751: (\ref{3CCM}) follow because $\liminf_mA_m\geq 1$.

1752: Now suppose that $\liminf_mB_m>1$.

1753: In that case, inequalities (\ref{3CCM}) follow from

1754: the first inequality in display (\ref{2CCM}) and from the following two 

1755: relations:

1756: $$

1757: \liminf_{m\to\infty}\frac{\log a_m}{\log b_m}

1758: =\sup\big\{\alpha >0\colon\,\lim_{m\to\infty}b_m^{-\alpha}a_m=+\infty\big\}

1759: =\sup\big\{\alpha >0\colon\,\inf_{m\geq 1}b_m^{-\alpha}a_m>0\big\}

1760: $$

1761: and

1762: \beq\label{s-limsup}

1763: \limsup_{m\to\infty}\frac{\log a_m}{\log b_m}

1764: =\inf\big\{\alpha >0\colon\,\lim_{m\to\infty}b_m^{-\alpha}a_m=0\big\}

1765: =\inf\big\{\alpha >0\colon\,\sup_{m\geq 1}b_m^{-\alpha}a_m<+\infty\big\},

1766: \eeq

1767: valid for any two sequences $\{a_m\colon\,m\geq 1\}$ and $\{b_m\colon\,

1768: m\geq 1\}$ of positive numbers such that $\lim_{m\to\infty}b_m=+\infty$.

1769: \qed\end{proof}

1770: 

1771: Combining Lemmas \ref{boxes}, \ref{subsequences} and \ref{CCM}, 

1772: it follows that the following statement holds.

1773: 

1774: \begin{cor}

1775: Let $f$ be a regulated non-constant function on $[0,T]$, and let 

1776: $\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing positive 

1777: integers such that

1778: \beq\label{4subsequences}

1779: \lim_{m\to\infty}\frac{\log N_m}{\log N_{m+1}}=1.

1780: \eeq

1781: Then

1782: $$

1783: \Delta (G_f)=\lim_{\epsilon\downarrow 0}\frac{\log M(G_f;\epsilon)}{\log

1784: (1/\epsilon )}=\lim_{m\to\infty}\frac{\log N_mQ(f;\lambda (m))}{\log N_m},

1785: $$

1786: provided that at least one of the three limits exists and is finite.

1787: \end{cor}

1788: 

1789: Next is the final step in a chain of inequalities used to

1790: prove Theorem \ref{main}.

1791: 

1792: \begin{lem}\label{bc-pv}

1793: Let $f$ be a regulated function on $[0,T]$, and let 

1794: $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing 

1795: positive integers.

1796: Then

1797: \beq\label{1bc-pv}

1798: \Big (2-\delta_{\eta}^{+}(f)=\Big )\qquad

1799: \limsup_{m\to\infty}\frac{\log N_mQ(f;\lambda (m))}{\log N_m}\leq 

1800: 2-\frac{1}{1\vee\upsilon (f)}.

1801: \eeq

1802: \end{lem}

1803: 

1804: \begin{proof}

1805: If $\upsilon (f)<1$ then $f$ has bounded $1$-variation, and so

1806: (\ref{1bc-pv}) holds by relation (\ref{bv}).

1807: If $\upsilon (f)=+\infty$, then (\ref{1bc-pv}) holds because

1808: $\delta_{\eta}^{+}(f)\geq 0$ by relation (\ref{Qin[1,2]}).

1809: Thus one can assume that $1\leq \upsilon (f)<+\infty$.

1810: Let $v_p(f;[0,T])<\infty$ for some $1\leq p<\infty$.

1811: We claim that for each integer $m\geq 1$,

1812: \beq\label{3bc-pv}

1813: N_m^{\frac{1}{p}-1}Q(f;\lambda (m))\leq v_p(f;[0,T])^{1/p}.

1814: \eeq

1815: Indeed, if $p=1$ then $Q(f;\lambda (m))\leq v_1(f;[0,T])$.

1816: Suppose that $p>1$ and $m\geq 1$.

1817: Since for each $i=1,\dots,N_m$, $\osc (f;\Delta_{i,m})\leq v_p(f;

1818: \Delta_{i,m})^{1/p}$, by H\"older's inequality, we have

1819: \begin{eqnarray*}

1820: Q (f;\lambda (m))&\leq& \sum_{i=1}^{N_m}v_p(f;\Delta_{i,m})^{1/p}

1821: \leq N_m^{1-1/p}\Big (\sum_{i=1}^{N_m}v_p(f;\Delta_{i,m})\Big )^{1/p}\\[2mm]

1822: \mbox{by subadditivity of $v_p$}\quad &\leq & N_m^{1-1/p}v_p(f;[0,T])^{1/p},

1823: \end{eqnarray*}

1824: proving relation (\ref{3bc-pv}).

1825: If $p>\upsilon (f)\geq 1$, then by (\ref{3bc-pv}), 

1826: $$

1827: \sup_{m\geq 1}N_m^{-(2-1/p)}\big [N_mQ(f;\lambda (m))\big ]

1828: \leq v_p(f;[0,T])^{1/p}<+\infty.

1829: $$

1830: Thus relation (\ref{1bc-pv}) follows by the relation (\ref{s-limsup}), 

1831: proving the lemma.

1832: \qed\end{proof}

1833: 

1834: Now we are ready to complete the proofs.

1835: 

1836: \medskip

1837: \noindent

1838: {\bf Proof of Theorem \ref{main}.}

1839: Let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing

1840: positive integers.

1841: We have to prove that $\delta_{\eta}(f)=1/(1\vee\upsilon (f))$.

1842: By Definition \ref{OS-index}, this will be done once we will

1843: show that

1844: \beq\label{1main}

1845: \lim_{m\to\infty}\frac{\log N_m Q(f;\lambda (m))}{\log N_m}

1846: =2-\frac{1}{1\vee\upsilon (f)}. 

1847: \eeq

1848: Since the graph $\gr (f)\subset G_f$, we have

1849: \begin{eqnarray*}

1850: \Delta^{-}(\gr (f)) &\leq & \Delta^{-}(G_f)=\liminf_{

1851: \epsilon\downarrow 0}\frac{\log M(G_f;\epsilon)}{\log (1/\epsilon)}

1852: \quad\stackrel{\mbox{by (\ref{2subsequences})}}{\leq}\quad

1853: \liminf_{m\to\infty}\frac{\log M(G_f;1/N_m)}{\log N_m}\\

1854: \mbox{by (\ref{1CCM})}\quad

1855: &=&\liminf_{m\to\infty}\frac{\log N_mQ(f;\lambda (m))}{\log N_m}

1856: \leq \limsup_{m\to\infty}\frac{\log N_mQ(f;\lambda (m))}{\log N_m}\\

1857: \mbox{by (\ref{1bc-pv})}\quad

1858: &\leq& 2-\frac{1}{1\vee\upsilon (f)}.

1859: \end{eqnarray*}

1860: Since the left and right sides are equal by assumption (\ref{2main}),

1861: relation (\ref{1main}) holds, proving Theorem \ref{main}.

1862: \qed

1863: 

1864: 

1865: \medskip

1866: \noindent

1867: {\bf Proof of Proposition \ref{3main}.}

1868: If $f$ is constant then the conclusion clearly holds,

1869: and so we can assume that $f$ is non-constant.

1870: Let $\eta=\{N_m\colon\,m\geq 1\}$ be a sequence of strictly increasing

1871: positive integers such that relation (\ref{4subsequences}) holds.

1872: Since the graph $\gr (f)\subset G_f$, we have

1873: \begin{eqnarray*}

1874: \Delta^{+}(\gr (f)) & =&\limsup_{\epsilon\downarrow 0}

1875: \frac{\log M(G_f;\epsilon)}{\log (1/\epsilon)}

1876: \quad\stackrel{\mbox{by (\ref{3subsequences})}}{=}\quad

1877: \limsup_{m\to\infty}\frac{\log M(G_f;1/N_m)}{\log N_m}\\

1878: \mbox{by (\ref{1CCM})}\quad

1879: &=&\limsup_{m\to\infty}\frac{\log N_mQ(f;\lambda (m))}{\log N_m}

1880: \quad\stackrel{\mbox{by (\ref{1bc-pv})}}{\leq}\quad

1881: 2-\frac{1}{1\vee\upsilon (f)},

1882: \end{eqnarray*}

1883: proving Proposition \ref{3main}.

1884: \qed

1885: 

1886: %\end{appendix}

1887: 

1888: \vspace*{0.1truein}

1889: %\addcontentsline{toc}{chapter}{\protect\numberline{}{Bibliography}}

1890: 

1891: \section*{References}

1892: 

1893: \parindent=2pc\def\hang#1{\vbox{\hsize\the\hsize\noindent%

1894:      \hangindent\the\parindent#1\par}\smallskip}%

1895: 

1896: 

1897: \hang{S.\ M.\ Berman,

1898: ``Harmonic analysis of local times and sample functions of

1899: Gaussian processes,''

1900: {\em Trans.\ Amer.\ Math.\ Soc.} vol.\ 143 pp.\ 269-281, 1969.}

1901: 

1902: \hang{A.\ Bick and W.\ Willinger,

1903: ``Dynamic spanning without probabilities,''

1904: {\em Stoch.\ Proc.\ Appl.} vol.\ 50 pp.\ 349-374, 1994.} 

1905: 

1906: \hang{R.\ M.\ Blumenthal and R.\ K.\ Getoor,

1907: ``Sample functions of stochastic processes with stationary independent

1908: increments,''

1909: {\em J.\ Math.\ and Mech.} vol.\ 10 pp.\ 493-516, 1961.}

1910: 

1911: \hang{R.\ M.\ Blumenthal and R.\ K.\ Getoor,

1912: ``The dimension of the set of zeros and the graph of a symmetric

1913: stable process,'' 

1914: {\em Illinois J.\ Math.} vol.\ 6 pp.\ 308-316, 1962.}

1915: 

1916: \hang{J.\ Y.\ Campbell, A.\ W.\ Lo and A.\ C.\ MacKinlay,

1917: {\em The econometrics of financial markets},

1918: Princeton University Press: New Jersey, 1997.}

1919: 

1920: \hang{P.\ H.\ Carter, R.\ Cawley and R.\ D.\ Mauldin,

1921: ``Mathematics of dimension measurement for graphs of functions,''

1922: in {\em Fractal Aspects of Materials: Disordered Systems},

1923: Eds. D.\ A.\ Weitz, L.\ Sander and B.\ Mandelbrot,

1924: Material Research Society, US, 1988, pp.\ 183-186. }

1925: 

1926: \hang{J.\ M.\ Chambers, C.\ L.\ Mallows and B.\ W.\ Stuck,

1927: ``A method for simulating stable random variables,''

1928: {\em J.\ Amer.\ Statist.\ Assoc.} vol.\ 71 no. 354 pp.\ 340-344, 1976.}

1929: 

1930: \hang{H.\ Cramer and M.\ R.\ Leadbetter,

1931: {\em Stationary and related stochastic processes},

1932: Wiley: New York, 1967.}

1933: 

1934: \hang{M.\ E.\ Crovella and M.\ S.\ Taqqu,

1935: ``Estimating the heavy tail index from scaling properties,''

1936: {\em Meth.\ and Comp.\ in Appl.\ Probab.} vol.\ 1 pp.\ 55-79, 1999.}

1937: 

1938: \hang{C.\ D.\ Cutler,

1939: ``A review of the theory and estimation of fractal dimension,''

1940: in {\em Dimension Estimation and Models}, Ed.\ H.\ Tong.,

1941: World Scientific, Singapore, 1993, pp.\ 1-107.}

1942: 

1943: \hang{B.\ Dubuc, J.\ F.\ Quiniou, C.\ Roques-Carmes, C.\ Tricot

1944: and S.\ W.\ Zucker,

1945: ``Evaluating the fractal dimension of profiles,''

1946: {\em Physical Review A} vol.\ 39 pp.\ 1500-1512, 1989.}

1947: 

1948: \hang{R.\ M.\ Dudley and R.\ Norvai\v sa,

1949: {\em Differentiability of Six Operators on Nonsmooth Functions and

1950: p-Variation},

1951: Lect.\ Notes in Math., vol.\  1703,

1952: Springer: Berlin, 1999}

1953: 

1954: \hang{H.\ F\"olmer,

1955: ``Calcul d'It\^o sans probabilit\'es,''

1956: in {\em S\'eminaire de Probabilit\'es} XV, Eds.\ J.\ Az\'ema and 

1957: M.\ Yor, Lect.\ Notes Math.\ vol.\ 850, Springer, 1981, pp.\ 143-150.} 

1958: 

1959: 

1960: \hang{E.\ G.\ Gladyshev,

1961: ``A new limit theorem for stochastic processes with Gaussian increments,''

1962: {\em Theor.\ Probability Appl.} vol.\ 6 pp.\ 52-61, 1961.}

1963: 

1964: \hang{P.\ Hall and A.\ Wood,

1965: ``On the performance of box-counting estimators of fractal dimension,''

1966: {\em Biometrika}, vol.\ 80 pp.\ 256-252, 1993.}

1967: 

1968: \hang{N.\ C.\ Jain and D.\ Monrad,

1969: ``Gaussian measures in $B_p$,''

1970: {\em Ann.\ Probab.} vol.\ 11 pp.\ 46-57, 1983.}

1971: 

1972: \hang{N.\ K\^ono,

1973: ``Hausdorff dimension of sample paths for self-similar processes,''

1974: in {\em Dependence in Probability and Statistics}, Eds.,

1975: E.\ Eberlein and M.\ S.\ Taqqu,

1976: Birkh\"auser, Boston, 1986, pp.\ 109-117.}

1977: 

1978: \hang{R.\ E.\ Maeder,

1979: ``Fractional Brownian motion,''

1980: {\em Mathematica J} vol.\ 6 pp.\ 38-48, 1995.}

1981: 

1982: \hang{B.\ B.\ Mandelbrot,

1983: {\em Fractals and Scaling in Finance: Discontinuity, Concentration,

1984: Risk}, Selecta volume E,

1985: Springer: New York, 1997.}

1986: 

1987: \hang{I.\ Monroe,

1988: ``On the $\gamma$-variation of processes with stationary independent

1989: increments,''

1990: {\em Ann.\ Math.\ Statist.} vol.\ 43 pp.\ 1213-1220, 1972.}

1991: 

1992: \hang{U.\ A.\ M\"uller, M.\ M.\ Dacorogna, R.\ D.\ Dav\'e, O.\ V.\ Pictet,

1993: R.\ B.\ Olsen and J.\ R.\ Ward,

1994: ``Fractals and intrinsic time - a chalenge to econometricians,''

1995: Internal document UAM.1993-08-16, Olsen \& Associates, 

1996: Switzerland, 1995. http://www.olsen.ch}

1997: 

1998: \hang{R.\ Norvai\v sa,

1999: ``Modelling of stock price changes: A real analysis approach,''

2000: {\em Finance and Stochastics} vol.\ 4 pp.\ 343-369, 2000a.}

2001: 

2002: \hang{R.\ Norvai\v sa,

2003: ``Quadratic variation, $p$-variation and integration with applications

2004: to stock price modelling,'' 2000b. (in preparation)}

2005: 

2006: \hang{R.\ Norvai\v sa and D.\ M.\ Salopek,

2007: ``Estimating the Orey index of a Gaussian stochastic process with

2008: stationary increments: An application to financial data set,''

2009: in {\em Stochastic Models}, Proc.\ Int.\ Conf., Ottawa, Canada

2010: June 10-13, 1998, Eds.\ L.\ G.\ Gorostiza and B.\ G.\ Ivanoff.

2011: Canadian Math.\ Soc., Conference Proceedings, vol.\ 26 2000, pp.\ 353-374.}

2012: 

2013: \hang{R.\ Norvai\v sa and D.\ M.\ Salopek,

2014: ``Supplement to the present paper, containing codes in Mathematica,''

2015: 2000b. (available from the authors upon request)}

2016: 

2017: \begin{sloppypar}

2018: \hang{S.\ Orey,

2019: ``Gaussian sample functions and the Hausdorff dimension of level crossings,''

2020: {\em Z.\ Wahrsch.\ verw.\ Gebiete} vol.\ 15 pp.\ 249-256, 1970.}

2021: \end{sloppypar}

2022: 

2023: \hang{W.\ E.\ Pruit and S.\ J.\ Taylor,

2024: ``Sample path properties of processes with stable components,''

2025: {\em Z.\ Wahrsch.\ verw.\ Geb.} vol.\ 12 pp.\ 267-289, 1969.}

2026: 

2027: %\hang{G.\ Samorodnitsky and M.\ Taqqu,}

2028: 

2029: %\hang{C.\ Tricot,

2030: %{\em Curves and Fractal Dimension}.

2031: %Springer, New York, 1995.}

2032: 

2033: \hang{V.\ M.\ Zolotarev,

2034: {\em One-dimensional Stable Distributions},

2035: Translations of Mathematical Monographs, vol.\ 65,

2036: American Mathematical Society, 1986.}

2037: 

2038: \end{document}

2039: 

2040: 

2041: \begin{thebibliography}{99}

2042: 

2043: \bibitem{SMB69}

2044: {\sc S.\ M.\ Berman},

2045: Harmonic analysis of local times and sample functions of

2046: Gaussian processes.

2047: {\em Trans.\ Amer.\ Math.\ Soc.}, {\bf 143} (1969), 269-281.

2048: 

2049: \bibitem{BandG61}

2050: {\sc R.\ M.\ Blumenthal and R.\ K.\ Getoor},

2051: Sample functions of stochastic processes with stationary independent

2052: increments.

2053: {\em J.\ Math.\ and Mech.}, {\bf 10} (1961), 493-516.

2054: 

2055: \bibitem{BandG62}

2056: {\sc R.\ M.\ Blumenthal and R.\ K.\ Getoor},

2057: The dimension of the set of zeros and the graph of a symmetric

2058: stable process. 

2059: {\em Illinois J.\ Math.}, {\bf 6} (1962), 308-316.

2060: 

2061: \bibitem{CLM97}

2062: {\sc J.\ Y.\ Campbell, A.\ W.\ Lo and A.\ C.\ MacKinlay},

2063: {\em The econometrics of financial markets.}

2064: Princeton University Press, New Jersey, 1997.

2065: 

2066: \bibitem{CCM88}

2067: {\sc P.\ H.\ Carter, R.\ Cawley and R.\ D.\ Mauldin},

2068: Mathematics of dimension measurement for graphs of functions.

2069: In: {\em Fractal Aspects of Materials: Disordered Systems},

2070: Eds. D.\ A.\ Weitz, L.\ Sander and B.\ Mandelbrot, pp. 183-186.

2071: Material Research Society, US, 1988. 

2072: 

2073: \bibitem{CMS76}

2074: {\sc J.\ M.\ Chambers, C.\ L.\ Mallows and B.\ W.\ Stuck},

2075: A method for simulating stable random variables.

2076: {\em J.\ Amer.\ Statist.\ Assoc.}, {\bf 71} (1976), no. 354, 340-344.

2077: 

2078: \bibitem{CandL67}

2079: {\sc H.\ Cramer and M.\ R.\ Leadbetter},

2080: {\em Stationary and related stochastic processes.}

2081: Wiley, New York, 1967.

2082: 

2083: \bibitem{CDC93}

2084: {\sc C.\ D.\ Cutler},

2085: A review of the theory and estimation of fractal dimension.

2086: In: {\em Dimension Estimation and Models}, Ed.\ H.\ Tong.,

2087: pp.\ 1-107.

2088: World Scientific, Singapore, 1993.

2089: 

2090: \bibitem{DQRTZ89}

2091: {\sc B.\ Dubuc, J.\ F.\ Quiniou, C.\ Roques-Carmes, C.\ Tricot

2092: and S.\ W.\ Zucker},

2093: Evaluating the fractal dimension of profiles.

2094: {\em Physical Review A}, {\bf 39} (1989), 1500-1512.

2095: 

2096: \hang{R.\ M.\ Dudley and R.\ Norvai\v sa,

2097: {\em Differentiability of Six Operators on Nonsmooth Functions and

2098: p-Variation},

2099: Lect.\ Notes in Math., Vol.\ {\bf 1703}.

2100: Springer: Berlin, 1999}

2101: 

2102: \bibitem{EGG61}

2103: {\sc E.\ G.\ Gladyshev},

2104: A new limit theorem for stochastic processes with Gaussian increments.

2105: {\em Theor.\ Probability Appl.}, {\bf 6} (1961), 52-61.

2106: 

2107: \bibitem{HandW93}

2108: {\sc P.\ Hall and A.\ Wood},

2109: On the performance of box-counting estimators of fractal dimension.

2110: {\em Biometrika}, {\bf 80} (1993), 256-252.

2111: 

2112: \bibitem{JandM83}

2113: {\sc N.\ C.\ Jain and D.\ Monrad},

2114: Gaussian measures in $B_p$.

2115: {\em Ann.\ Probab.}, {\bf 11} (1983), 46-57. 

2116: 

2117: \bibitem{IM72}

2118: {\sc I.\ Monroe},

2119: On the $\gamma$-variation of processes with stationary independent

2120: increments.

2121: {\em Ann.\ Math.\ Statist.}, {\bf 43} (1972), 1213-1220.

2122: 

2123: \bibitem{NandS00}

2124: {\sc R.\ Norvai\v sa and D.\ M.\ Salopek},

2125: Estimating the Orey index of a Gaussian stochastic process with

2126: stationary increments: An application to financial data set.

2127: In: {........}, pp.\ 353-374.

2128: Canadian Math.\ Soc., Conference Proceedings, {\bf 26}, 2000.

2129: 

2130: \bibitem{SO70}

2131: {\sc S.\ Orey},

2132: Gaussian sample functions and the Hausdorff dimension of level crossings.

2133: {\em Z.\ Wahrsch. verw. Gebiete}, {\bf 15} (1970), 249-256.

2134: 

2135: %\bibitem{SandT}

2136: %{\sc G.\ Samorodnitsky and M.\ Taqqu},

2137: 

2138: 

2139: %\bibitem{CT95}

2140: %{\sc C.\ Tricot},

2141: %{\em Curves and Fractal Dimension}.

2142: %Springer, New York, 1995.

2143: 

2144: 

2145: 

2146: \bibitem{VMZ86}

2147: {\sc V.\ M.\ Zolotarev},

2148: {\em One-dimensional Stable Distributions}.

2149: AMS, USA, 1986.

2150: 

2151: 

2152: 

2153: \end{thebibliography}

2154: \end{document}

2155: